From 38d049382533c6662d815b08ca3395e96cca9f57 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sun, 4 Jan 2026 16:13:50 -0800 Subject: [PATCH 01/48] Fix case where upscale model wouldn't be moved to cpu. (#11633) --- comfy_extras/nodes_upscale_model.py | 24 +++++++++++++----------- 1 file changed, 13 insertions(+), 11 deletions(-) diff --git a/comfy_extras/nodes_upscale_model.py b/comfy_extras/nodes_upscale_model.py index 4d62b87be..ed587851c 100644 --- a/comfy_extras/nodes_upscale_model.py +++ b/comfy_extras/nodes_upscale_model.py @@ -78,18 +78,20 @@ class ImageUpscaleWithModel(io.ComfyNode): overlap = 32 oom = True - while oom: - try: - steps = in_img.shape[0] * comfy.utils.get_tiled_scale_steps(in_img.shape[3], in_img.shape[2], tile_x=tile, tile_y=tile, overlap=overlap) - pbar = comfy.utils.ProgressBar(steps) - s = comfy.utils.tiled_scale(in_img, lambda a: upscale_model(a), tile_x=tile, tile_y=tile, overlap=overlap, upscale_amount=upscale_model.scale, pbar=pbar) - oom = False - except model_management.OOM_EXCEPTION as e: - tile //= 2 - if tile < 128: - raise e + try: + while oom: + try: + steps = in_img.shape[0] * comfy.utils.get_tiled_scale_steps(in_img.shape[3], in_img.shape[2], tile_x=tile, tile_y=tile, overlap=overlap) + pbar = comfy.utils.ProgressBar(steps) + s = comfy.utils.tiled_scale(in_img, lambda a: upscale_model(a), tile_x=tile, tile_y=tile, overlap=overlap, upscale_amount=upscale_model.scale, pbar=pbar) + oom = False + except model_management.OOM_EXCEPTION as e: + tile //= 2 + if tile < 128: + raise e + finally: + upscale_model.to("cpu") - upscale_model.to("cpu") s = torch.clamp(s.movedim(-3,-1), min=0, max=1.0) return io.NodeOutput(s) From f2b002372b71cf0671a4cf1fa539e1c386d727e4 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sun, 4 Jan 2026 22:58:59 -0800 Subject: [PATCH 02/48] Support the LTXV 2 model. (#11632) --- comfy/latent_formats.py | 3 + comfy/ldm/lightricks/av_model.py | 837 ++++++++++++++++ comfy/ldm/lightricks/embeddings_connector.py | 305 ++++++ comfy/ldm/lightricks/latent_upsampler.py | 292 ++++++ comfy/ldm/lightricks/model.py | 715 +++++++++++--- comfy/ldm/lightricks/symmetric_patchifier.py | 87 +- comfy/ldm/lightricks/vae/audio_vae.py | 286 ++++++ .../vae/causal_audio_autoencoder.py | 909 ++++++++++++++++++ comfy/ldm/lightricks/vocoders/vocoder.py | 213 ++++ comfy/model_base.py | 57 +- comfy/model_detection.py | 2 +- comfy/sd.py | 9 +- comfy/supported_models.py | 17 +- comfy/text_encoders/llama.py | 79 ++ comfy/text_encoders/lt.py | 111 +++ comfy/utils.py | 2 +- comfy_extras/nodes_audio.py | 2 +- comfy_extras/nodes_hunyuan.py | 15 +- comfy_extras/nodes_lt.py | 188 +++- comfy_extras/nodes_lt_audio.py | 183 ++++ comfy_extras/nodes_lt_upsampler.py | 75 ++ nodes.py | 10 +- pyproject.toml | 2 +- 23 files changed, 4214 insertions(+), 185 deletions(-) create mode 100644 comfy/ldm/lightricks/av_model.py create mode 100644 comfy/ldm/lightricks/embeddings_connector.py create mode 100644 comfy/ldm/lightricks/latent_upsampler.py create mode 100644 comfy/ldm/lightricks/vae/audio_vae.py create mode 100644 comfy/ldm/lightricks/vae/causal_audio_autoencoder.py create mode 100644 comfy/ldm/lightricks/vocoders/vocoder.py create mode 100644 comfy_extras/nodes_lt_audio.py create mode 100644 comfy_extras/nodes_lt_upsampler.py diff --git a/comfy/latent_formats.py b/comfy/latent_formats.py index f1ca0151e..9bbe30b53 100644 --- a/comfy/latent_formats.py +++ b/comfy/latent_formats.py @@ -407,6 +407,9 @@ class LTXV(LatentFormat): self.latent_rgb_factors_bias = [-0.0571, -0.1657, -0.2512] +class LTXAV(LTXV): + pass + class HunyuanVideo(LatentFormat): latent_channels = 16 latent_dimensions = 3 diff --git a/comfy/ldm/lightricks/av_model.py b/comfy/ldm/lightricks/av_model.py new file mode 100644 index 000000000..759535501 --- /dev/null +++ b/comfy/ldm/lightricks/av_model.py @@ -0,0 +1,837 @@ +from typing import Tuple +import torch +import torch.nn as nn +from comfy.ldm.lightricks.model import ( + CrossAttention, + FeedForward, + AdaLayerNormSingle, + PixArtAlphaTextProjection, + LTXVModel, +) +from comfy.ldm.lightricks.symmetric_patchifier import AudioPatchifier +import comfy.ldm.common_dit + +class BasicAVTransformerBlock(nn.Module): + def __init__( + self, + v_dim, + a_dim, + v_heads, + a_heads, + vd_head, + ad_head, + v_context_dim=None, + a_context_dim=None, + attn_precision=None, + dtype=None, + device=None, + operations=None, + ): + super().__init__() + + self.attn_precision = attn_precision + + self.attn1 = CrossAttention( + query_dim=v_dim, + heads=v_heads, + dim_head=vd_head, + context_dim=None, + attn_precision=self.attn_precision, + dtype=dtype, + device=device, + operations=operations, + ) + self.audio_attn1 = CrossAttention( + query_dim=a_dim, + heads=a_heads, + dim_head=ad_head, + context_dim=None, + attn_precision=self.attn_precision, + dtype=dtype, + device=device, + operations=operations, + ) + + self.attn2 = CrossAttention( + query_dim=v_dim, + context_dim=v_context_dim, + heads=v_heads, + dim_head=vd_head, + attn_precision=self.attn_precision, + dtype=dtype, + device=device, + operations=operations, + ) + self.audio_attn2 = CrossAttention( + query_dim=a_dim, + context_dim=a_context_dim, + heads=a_heads, + dim_head=ad_head, + attn_precision=self.attn_precision, + dtype=dtype, + device=device, + operations=operations, + ) + + # Q: Video, K,V: Audio + self.audio_to_video_attn = CrossAttention( + query_dim=v_dim, + context_dim=a_dim, + heads=a_heads, + dim_head=ad_head, + attn_precision=self.attn_precision, + dtype=dtype, + device=device, + operations=operations, + ) + + # Q: Audio, K,V: Video + self.video_to_audio_attn = CrossAttention( + query_dim=a_dim, + context_dim=v_dim, + heads=a_heads, + dim_head=ad_head, + attn_precision=self.attn_precision, + dtype=dtype, + device=device, + operations=operations, + ) + + self.ff = FeedForward( + v_dim, dim_out=v_dim, glu=True, dtype=dtype, device=device, operations=operations + ) + self.audio_ff = FeedForward( + a_dim, dim_out=a_dim, glu=True, dtype=dtype, device=device, operations=operations + ) + + self.scale_shift_table = nn.Parameter(torch.empty(6, v_dim, device=device, dtype=dtype)) + self.audio_scale_shift_table = nn.Parameter( + torch.empty(6, a_dim, device=device, dtype=dtype) + ) + + self.scale_shift_table_a2v_ca_audio = nn.Parameter( + torch.empty(5, a_dim, device=device, dtype=dtype) + ) + self.scale_shift_table_a2v_ca_video = nn.Parameter( + torch.empty(5, v_dim, device=device, dtype=dtype) + ) + + def get_ada_values( + self, scale_shift_table: torch.Tensor, batch_size: int, timestep: torch.Tensor, indices: slice = slice(None, None) + ): + num_ada_params = scale_shift_table.shape[0] + + ada_values = ( + scale_shift_table[indices].unsqueeze(0).unsqueeze(0).to(device=timestep.device, dtype=timestep.dtype) + + timestep.reshape(batch_size, timestep.shape[1], num_ada_params, -1)[:, :, indices, :] + ).unbind(dim=2) + return ada_values + + def get_av_ca_ada_values( + self, + scale_shift_table: torch.Tensor, + batch_size: int, + scale_shift_timestep: torch.Tensor, + gate_timestep: torch.Tensor, + num_scale_shift_values: int = 4, + ): + scale_shift_ada_values = self.get_ada_values( + scale_shift_table[:num_scale_shift_values, :], + batch_size, + scale_shift_timestep, + ) + gate_ada_values = self.get_ada_values( + scale_shift_table[num_scale_shift_values:, :], + batch_size, + gate_timestep, + ) + + scale_shift_chunks = [t.squeeze(2) for t in scale_shift_ada_values] + gate_ada_values = [t.squeeze(2) for t in gate_ada_values] + + return (*scale_shift_chunks, *gate_ada_values) + + def forward( + self, + x: Tuple[torch.Tensor, torch.Tensor], + v_context=None, + a_context=None, + attention_mask=None, + v_timestep=None, + a_timestep=None, + v_pe=None, + a_pe=None, + v_cross_pe=None, + a_cross_pe=None, + v_cross_scale_shift_timestep=None, + a_cross_scale_shift_timestep=None, + v_cross_gate_timestep=None, + a_cross_gate_timestep=None, + transformer_options=None, + ) -> Tuple[torch.Tensor, torch.Tensor]: + run_vx = transformer_options.get("run_vx", True) + run_ax = transformer_options.get("run_ax", True) + + vx, ax = x + run_ax = run_ax and ax.numel() > 0 + run_a2v = run_vx and transformer_options.get("a2v_cross_attn", True) and ax.numel() > 0 + run_v2a = run_ax and transformer_options.get("v2a_cross_attn", True) + + if run_vx: + vshift_msa, vscale_msa, vgate_msa = ( + self.get_ada_values(self.scale_shift_table, vx.shape[0], v_timestep, slice(0, 3)) + ) + + norm_vx = comfy.ldm.common_dit.rms_norm(vx) * (1 + vscale_msa) + vshift_msa + vx += self.attn1(norm_vx, pe=v_pe, transformer_options=transformer_options) * vgate_msa + vx += self.attn2( + comfy.ldm.common_dit.rms_norm(vx), + context=v_context, + mask=attention_mask, + transformer_options=transformer_options, + ) + + del vshift_msa, vscale_msa, vgate_msa + + if run_ax: + ashift_msa, ascale_msa, agate_msa = ( + self.get_ada_values(self.audio_scale_shift_table, ax.shape[0], a_timestep, slice(0, 3)) + ) + + norm_ax = comfy.ldm.common_dit.rms_norm(ax) * (1 + ascale_msa) + ashift_msa + ax += ( + self.audio_attn1(norm_ax, pe=a_pe, transformer_options=transformer_options) + * agate_msa + ) + ax += self.audio_attn2( + comfy.ldm.common_dit.rms_norm(ax), + context=a_context, + mask=attention_mask, + transformer_options=transformer_options, + ) + + del ashift_msa, ascale_msa, agate_msa + + # Audio - Video cross attention. + if run_a2v or run_v2a: + # norm3 + vx_norm3 = comfy.ldm.common_dit.rms_norm(vx) + ax_norm3 = comfy.ldm.common_dit.rms_norm(ax) + + ( + scale_ca_audio_hidden_states_a2v, + shift_ca_audio_hidden_states_a2v, + scale_ca_audio_hidden_states_v2a, + shift_ca_audio_hidden_states_v2a, + gate_out_v2a, + ) = self.get_av_ca_ada_values( + self.scale_shift_table_a2v_ca_audio, + ax.shape[0], + a_cross_scale_shift_timestep, + a_cross_gate_timestep, + ) + + ( + scale_ca_video_hidden_states_a2v, + shift_ca_video_hidden_states_a2v, + scale_ca_video_hidden_states_v2a, + shift_ca_video_hidden_states_v2a, + gate_out_a2v, + ) = self.get_av_ca_ada_values( + self.scale_shift_table_a2v_ca_video, + vx.shape[0], + v_cross_scale_shift_timestep, + v_cross_gate_timestep, + ) + + if run_a2v: + vx_scaled = ( + vx_norm3 * (1 + scale_ca_video_hidden_states_a2v) + + shift_ca_video_hidden_states_a2v + ) + ax_scaled = ( + ax_norm3 * (1 + scale_ca_audio_hidden_states_a2v) + + shift_ca_audio_hidden_states_a2v + ) + vx += ( + self.audio_to_video_attn( + vx_scaled, + context=ax_scaled, + pe=v_cross_pe, + k_pe=a_cross_pe, + transformer_options=transformer_options, + ) + * gate_out_a2v + ) + + del gate_out_a2v + del scale_ca_video_hidden_states_a2v,\ + shift_ca_video_hidden_states_a2v,\ + scale_ca_audio_hidden_states_a2v,\ + shift_ca_audio_hidden_states_a2v,\ + + if run_v2a: + ax_scaled = ( + ax_norm3 * (1 + scale_ca_audio_hidden_states_v2a) + + shift_ca_audio_hidden_states_v2a + ) + vx_scaled = ( + vx_norm3 * (1 + scale_ca_video_hidden_states_v2a) + + shift_ca_video_hidden_states_v2a + ) + ax += ( + self.video_to_audio_attn( + ax_scaled, + context=vx_scaled, + pe=a_cross_pe, + k_pe=v_cross_pe, + transformer_options=transformer_options, + ) + * gate_out_v2a + ) + + del gate_out_v2a + del scale_ca_video_hidden_states_v2a,\ + shift_ca_video_hidden_states_v2a,\ + scale_ca_audio_hidden_states_v2a,\ + shift_ca_audio_hidden_states_v2a + + if run_vx: + vshift_mlp, vscale_mlp, vgate_mlp = ( + self.get_ada_values(self.scale_shift_table, vx.shape[0], v_timestep, slice(3, None)) + ) + + vx_scaled = comfy.ldm.common_dit.rms_norm(vx) * (1 + vscale_mlp) + vshift_mlp + vx += self.ff(vx_scaled) * vgate_mlp + del vshift_mlp, vscale_mlp, vgate_mlp + + if run_ax: + ashift_mlp, ascale_mlp, agate_mlp = ( + self.get_ada_values(self.audio_scale_shift_table, ax.shape[0], a_timestep, slice(3, None)) + ) + + ax_scaled = comfy.ldm.common_dit.rms_norm(ax) * (1 + ascale_mlp) + ashift_mlp + ax += self.audio_ff(ax_scaled) * agate_mlp + + del ashift_mlp, ascale_mlp, agate_mlp + + + return vx, ax + + +class LTXAVModel(LTXVModel): + """LTXAV model for audio-video generation.""" + + def __init__( + self, + in_channels=128, + audio_in_channels=128, + cross_attention_dim=4096, + audio_cross_attention_dim=2048, + attention_head_dim=128, + audio_attention_head_dim=64, + num_attention_heads=32, + audio_num_attention_heads=32, + caption_channels=3840, + num_layers=48, + positional_embedding_theta=10000.0, + positional_embedding_max_pos=[20, 2048, 2048], + audio_positional_embedding_max_pos=[20], + causal_temporal_positioning=False, + vae_scale_factors=(8, 32, 32), + use_middle_indices_grid=False, + timestep_scale_multiplier=1000.0, + av_ca_timestep_scale_multiplier=1.0, + dtype=None, + device=None, + operations=None, + **kwargs, + ): + # Store audio-specific parameters + self.audio_in_channels = audio_in_channels + self.audio_cross_attention_dim = audio_cross_attention_dim + self.audio_attention_head_dim = audio_attention_head_dim + self.audio_num_attention_heads = audio_num_attention_heads + self.audio_positional_embedding_max_pos = audio_positional_embedding_max_pos + + # Calculate audio dimensions + self.audio_inner_dim = audio_num_attention_heads * audio_attention_head_dim + self.audio_out_channels = audio_in_channels + + # Audio-specific constants + self.num_audio_channels = 8 + self.audio_frequency_bins = 16 + + self.av_ca_timestep_scale_multiplier = av_ca_timestep_scale_multiplier + + super().__init__( + in_channels=in_channels, + cross_attention_dim=cross_attention_dim, + attention_head_dim=attention_head_dim, + num_attention_heads=num_attention_heads, + caption_channels=caption_channels, + num_layers=num_layers, + positional_embedding_theta=positional_embedding_theta, + positional_embedding_max_pos=positional_embedding_max_pos, + causal_temporal_positioning=causal_temporal_positioning, + vae_scale_factors=vae_scale_factors, + use_middle_indices_grid=use_middle_indices_grid, + timestep_scale_multiplier=timestep_scale_multiplier, + dtype=dtype, + device=device, + operations=operations, + **kwargs, + ) + + def _init_model_components(self, device, dtype, **kwargs): + """Initialize LTXAV-specific components.""" + # Audio-specific projections + self.audio_patchify_proj = self.operations.Linear( + self.audio_in_channels, self.audio_inner_dim, bias=True, dtype=dtype, device=device + ) + + # Audio-specific AdaLN + self.audio_adaln_single = AdaLayerNormSingle( + self.audio_inner_dim, + use_additional_conditions=False, + dtype=dtype, + device=device, + operations=self.operations, + ) + + num_scale_shift_values = 4 + self.av_ca_video_scale_shift_adaln_single = AdaLayerNormSingle( + self.inner_dim, + use_additional_conditions=False, + embedding_coefficient=num_scale_shift_values, + dtype=dtype, + device=device, + operations=self.operations, + ) + self.av_ca_a2v_gate_adaln_single = AdaLayerNormSingle( + self.inner_dim, + use_additional_conditions=False, + embedding_coefficient=1, + dtype=dtype, + device=device, + operations=self.operations, + ) + self.av_ca_audio_scale_shift_adaln_single = AdaLayerNormSingle( + self.audio_inner_dim, + use_additional_conditions=False, + embedding_coefficient=num_scale_shift_values, + dtype=dtype, + device=device, + operations=self.operations, + ) + self.av_ca_v2a_gate_adaln_single = AdaLayerNormSingle( + self.audio_inner_dim, + use_additional_conditions=False, + embedding_coefficient=1, + dtype=dtype, + device=device, + operations=self.operations, + ) + + # Audio caption projection + self.audio_caption_projection = PixArtAlphaTextProjection( + in_features=self.caption_channels, + hidden_size=self.audio_inner_dim, + dtype=dtype, + device=device, + operations=self.operations, + ) + + def _init_transformer_blocks(self, device, dtype, **kwargs): + """Initialize transformer blocks for LTXAV.""" + self.transformer_blocks = nn.ModuleList( + [ + BasicAVTransformerBlock( + v_dim=self.inner_dim, + a_dim=self.audio_inner_dim, + v_heads=self.num_attention_heads, + a_heads=self.audio_num_attention_heads, + vd_head=self.attention_head_dim, + ad_head=self.audio_attention_head_dim, + v_context_dim=self.cross_attention_dim, + a_context_dim=self.audio_cross_attention_dim, + dtype=dtype, + device=device, + operations=self.operations, + ) + for _ in range(self.num_layers) + ] + ) + + def _init_output_components(self, device, dtype): + """Initialize output components for LTXAV.""" + # Video output components + super()._init_output_components(device, dtype) + # Audio output components + self.audio_scale_shift_table = nn.Parameter( + torch.empty(2, self.audio_inner_dim, dtype=dtype, device=device) + ) + self.audio_norm_out = self.operations.LayerNorm( + self.audio_inner_dim, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device + ) + self.audio_proj_out = self.operations.Linear( + self.audio_inner_dim, self.audio_out_channels, dtype=dtype, device=device + ) + self.a_patchifier = AudioPatchifier(1, start_end=True) + + def separate_audio_and_video_latents(self, x, audio_length): + """Separate audio and video latents from combined input.""" + # vx = x[:, : self.in_channels] + # ax = x[:, self.in_channels :] + # + # ax = ax.reshape(ax.shape[0], -1) + # ax = ax[:, : audio_length * self.num_audio_channels * self.audio_frequency_bins] + # + # ax = ax.reshape( + # ax.shape[0], self.num_audio_channels, audio_length, self.audio_frequency_bins + # ) + + vx = x[0] + ax = x[1] if len(x) > 1 else torch.zeros( + (vx.shape[0], self.num_audio_channels, 0, self.audio_frequency_bins), + device=vx.device, dtype=vx.dtype + ) + return vx, ax + + def recombine_audio_and_video_latents(self, vx, ax, target_shape=None): + if ax.numel() == 0: + return vx + else: + return [vx, ax] + """Recombine audio and video latents for output.""" + # if ax.device != vx.device or ax.dtype != vx.dtype: + # logging.warning("Audio and video latents are on different devices or dtypes.") + # ax = ax.to(device=vx.device, dtype=vx.dtype) + # logging.warning(f"Audio audio latent moved to device: {ax.device}, dtype: {ax.dtype}") + # + # ax = ax.reshape(ax.shape[0], -1) + # # pad to f x h x w of the video latents + # divisor = vx.shape[-1] * vx.shape[-2] * vx.shape[-3] + # if target_shape is None: + # repetitions = math.ceil(ax.shape[-1] / divisor) + # else: + # repetitions = target_shape[1] - vx.shape[1] + # padded_len = repetitions * divisor + # ax = F.pad(ax, (0, padded_len - ax.shape[-1])) + # ax = ax.reshape(ax.shape[0], -1, vx.shape[-3], vx.shape[-2], vx.shape[-1]) + # return torch.cat([vx, ax], dim=1) + + def _process_input(self, x, keyframe_idxs, denoise_mask, **kwargs): + """Process input for LTXAV - separate audio and video, then patchify.""" + audio_length = kwargs.get("audio_length", 0) + # Separate audio and video latents + vx, ax = self.separate_audio_and_video_latents(x, audio_length) + [vx, v_pixel_coords, additional_args] = super()._process_input( + vx, keyframe_idxs, denoise_mask, **kwargs + ) + + ax, a_latent_coords = self.a_patchifier.patchify(ax) + ax = self.audio_patchify_proj(ax) + + # additional_args.update({"av_orig_shape": list(x.shape)}) + return [vx, ax], [v_pixel_coords, a_latent_coords], additional_args + + def _prepare_timestep(self, timestep, batch_size, hidden_dtype, **kwargs): + """Prepare timestep embeddings.""" + # TODO: some code reuse is needed here. + grid_mask = kwargs.get("grid_mask", None) + if grid_mask is not None: + timestep = timestep[:, grid_mask] + + timestep = timestep * self.timestep_scale_multiplier + v_timestep, v_embedded_timestep = self.adaln_single( + timestep.flatten(), + {"resolution": None, "aspect_ratio": None}, + batch_size=batch_size, + hidden_dtype=hidden_dtype, + ) + + # Second dimension is 1 or number of tokens (if timestep_per_token) + v_timestep = v_timestep.view(batch_size, -1, v_timestep.shape[-1]) + v_embedded_timestep = v_embedded_timestep.view( + batch_size, -1, v_embedded_timestep.shape[-1] + ) + + # Prepare audio timestep + a_timestep = kwargs.get("a_timestep") + if a_timestep is not None: + a_timestep = a_timestep * self.timestep_scale_multiplier + av_ca_factor = self.av_ca_timestep_scale_multiplier / self.timestep_scale_multiplier + + av_ca_audio_scale_shift_timestep, _ = self.av_ca_audio_scale_shift_adaln_single( + a_timestep.flatten(), + {"resolution": None, "aspect_ratio": None}, + batch_size=batch_size, + hidden_dtype=hidden_dtype, + ) + av_ca_video_scale_shift_timestep, _ = self.av_ca_video_scale_shift_adaln_single( + timestep.flatten(), + {"resolution": None, "aspect_ratio": None}, + batch_size=batch_size, + hidden_dtype=hidden_dtype, + ) + av_ca_a2v_gate_noise_timestep, _ = self.av_ca_a2v_gate_adaln_single( + timestep.flatten() * av_ca_factor, + {"resolution": None, "aspect_ratio": None}, + batch_size=batch_size, + hidden_dtype=hidden_dtype, + ) + av_ca_v2a_gate_noise_timestep, _ = self.av_ca_v2a_gate_adaln_single( + a_timestep.flatten() * av_ca_factor, + {"resolution": None, "aspect_ratio": None}, + batch_size=batch_size, + hidden_dtype=hidden_dtype, + ) + + a_timestep, a_embedded_timestep = self.audio_adaln_single( + a_timestep.flatten(), + {"resolution": None, "aspect_ratio": None}, + batch_size=batch_size, + hidden_dtype=hidden_dtype, + ) + a_timestep = a_timestep.view(batch_size, -1, a_timestep.shape[-1]) + a_embedded_timestep = a_embedded_timestep.view( + batch_size, -1, a_embedded_timestep.shape[-1] + ) + cross_av_timestep_ss = [ + av_ca_audio_scale_shift_timestep, + av_ca_video_scale_shift_timestep, + av_ca_a2v_gate_noise_timestep, + av_ca_v2a_gate_noise_timestep, + ] + cross_av_timestep_ss = list( + [t.view(batch_size, -1, t.shape[-1]) for t in cross_av_timestep_ss] + ) + else: + a_timestep = timestep + a_embedded_timestep = kwargs.get("embedded_timestep") + cross_av_timestep_ss = [] + + return [v_timestep, a_timestep, cross_av_timestep_ss], [ + v_embedded_timestep, + a_embedded_timestep, + ] + + def _prepare_context(self, context, batch_size, x, attention_mask=None): + vx = x[0] + ax = x[1] + v_context, a_context = torch.split( + context, int(context.shape[-1] / 2), len(context.shape) - 1 + ) + + v_context, attention_mask = super()._prepare_context( + v_context, batch_size, vx, attention_mask + ) + if self.audio_caption_projection is not None: + a_context = self.audio_caption_projection(a_context) + a_context = a_context.view(batch_size, -1, ax.shape[-1]) + + return [v_context, a_context], attention_mask + + def _prepare_positional_embeddings(self, pixel_coords, frame_rate, x_dtype): + v_pixel_coords = pixel_coords[0] + v_pe = super()._prepare_positional_embeddings(v_pixel_coords, frame_rate, x_dtype) + + a_latent_coords = pixel_coords[1] + a_pe = self._precompute_freqs_cis( + a_latent_coords, + dim=self.audio_inner_dim, + out_dtype=x_dtype, + max_pos=self.audio_positional_embedding_max_pos, + use_middle_indices_grid=self.use_middle_indices_grid, + num_attention_heads=self.audio_num_attention_heads, + ) + + # calculate positional embeddings for the middle of the token duration, to use in av cross attention layers. + max_pos = max( + self.positional_embedding_max_pos[0], self.audio_positional_embedding_max_pos[0] + ) + v_pixel_coords = v_pixel_coords.to(torch.float32) + v_pixel_coords[:, 0] = v_pixel_coords[:, 0] * (1.0 / frame_rate) + av_cross_video_freq_cis = self._precompute_freqs_cis( + v_pixel_coords[:, 0:1, :], + dim=self.audio_cross_attention_dim, + out_dtype=x_dtype, + max_pos=[max_pos], + use_middle_indices_grid=True, + num_attention_heads=self.audio_num_attention_heads, + ) + av_cross_audio_freq_cis = self._precompute_freqs_cis( + a_latent_coords[:, 0:1, :], + dim=self.audio_cross_attention_dim, + out_dtype=x_dtype, + max_pos=[max_pos], + use_middle_indices_grid=True, + num_attention_heads=self.audio_num_attention_heads, + ) + + return [(v_pe, av_cross_video_freq_cis), (a_pe, av_cross_audio_freq_cis)] + + def _process_transformer_blocks( + self, x, context, attention_mask, timestep, pe, transformer_options={}, **kwargs + ): + vx = x[0] + ax = x[1] + v_context = context[0] + a_context = context[1] + v_timestep = timestep[0] + a_timestep = timestep[1] + v_pe, av_cross_video_freq_cis = pe[0] + a_pe, av_cross_audio_freq_cis = pe[1] + + ( + av_ca_audio_scale_shift_timestep, + av_ca_video_scale_shift_timestep, + av_ca_a2v_gate_noise_timestep, + av_ca_v2a_gate_noise_timestep, + ) = timestep[2] + + """Process transformer blocks for LTXAV.""" + patches_replace = transformer_options.get("patches_replace", {}) + blocks_replace = patches_replace.get("dit", {}) + + # Process transformer blocks + for i, block in enumerate(self.transformer_blocks): + if ("double_block", i) in blocks_replace: + + def block_wrap(args): + out = {} + out["img"] = block( + args["img"], + v_context=args["v_context"], + a_context=args["a_context"], + attention_mask=args["attention_mask"], + v_timestep=args["v_timestep"], + a_timestep=args["a_timestep"], + v_pe=args["v_pe"], + a_pe=args["a_pe"], + v_cross_pe=args["v_cross_pe"], + a_cross_pe=args["a_cross_pe"], + v_cross_scale_shift_timestep=args["v_cross_scale_shift_timestep"], + a_cross_scale_shift_timestep=args["a_cross_scale_shift_timestep"], + v_cross_gate_timestep=args["v_cross_gate_timestep"], + a_cross_gate_timestep=args["a_cross_gate_timestep"], + transformer_options=args["transformer_options"], + ) + return out + + out = blocks_replace[("double_block", i)]( + { + "img": (vx, ax), + "v_context": v_context, + "a_context": a_context, + "attention_mask": attention_mask, + "v_timestep": v_timestep, + "a_timestep": a_timestep, + "v_pe": v_pe, + "a_pe": a_pe, + "v_cross_pe": av_cross_video_freq_cis, + "a_cross_pe": av_cross_audio_freq_cis, + "v_cross_scale_shift_timestep": av_ca_video_scale_shift_timestep, + "a_cross_scale_shift_timestep": av_ca_audio_scale_shift_timestep, + "v_cross_gate_timestep": av_ca_a2v_gate_noise_timestep, + "a_cross_gate_timestep": av_ca_v2a_gate_noise_timestep, + "transformer_options": transformer_options, + }, + {"original_block": block_wrap}, + ) + vx, ax = out["img"] + else: + vx, ax = block( + (vx, ax), + v_context=v_context, + a_context=a_context, + attention_mask=attention_mask, + v_timestep=v_timestep, + a_timestep=a_timestep, + v_pe=v_pe, + a_pe=a_pe, + v_cross_pe=av_cross_video_freq_cis, + a_cross_pe=av_cross_audio_freq_cis, + v_cross_scale_shift_timestep=av_ca_video_scale_shift_timestep, + a_cross_scale_shift_timestep=av_ca_audio_scale_shift_timestep, + v_cross_gate_timestep=av_ca_a2v_gate_noise_timestep, + a_cross_gate_timestep=av_ca_v2a_gate_noise_timestep, + transformer_options=transformer_options, + ) + + return [vx, ax] + + def _process_output(self, x, embedded_timestep, keyframe_idxs, **kwargs): + vx = x[0] + ax = x[1] + v_embedded_timestep = embedded_timestep[0] + a_embedded_timestep = embedded_timestep[1] + vx = super()._process_output(vx, v_embedded_timestep, keyframe_idxs, **kwargs) + + # Process audio output + a_scale_shift_values = ( + self.audio_scale_shift_table[None, None].to(device=a_embedded_timestep.device, dtype=a_embedded_timestep.dtype) + + a_embedded_timestep[:, :, None] + ) + a_shift, a_scale = a_scale_shift_values[:, :, 0], a_scale_shift_values[:, :, 1] + + ax = self.audio_norm_out(ax) + ax = ax * (1 + a_scale) + a_shift + ax = self.audio_proj_out(ax) + + # Unpatchify audio + ax = self.a_patchifier.unpatchify( + ax, channels=self.num_audio_channels, freq=self.audio_frequency_bins + ) + + # Recombine audio and video + original_shape = kwargs.get("av_orig_shape") + return self.recombine_audio_and_video_latents(vx, ax, original_shape) + + def forward( + self, + x, + timestep, + context, + attention_mask=None, + frame_rate=25, + transformer_options={}, + keyframe_idxs=None, + **kwargs, + ): + """ + Forward pass for LTXAV model. + + Args: + x: Combined audio-video input tensor + timestep: Tuple of (video_timestep, audio_timestep) or single timestep + context: Context tensor (e.g., text embeddings) + attention_mask: Attention mask tensor + frame_rate: Frame rate for temporal processing + transformer_options: Additional options for transformer blocks + keyframe_idxs: Keyframe indices for temporal processing + **kwargs: Additional keyword arguments including audio_length + + Returns: + Combined audio-video output tensor + """ + # Handle timestep format + if isinstance(timestep, (tuple, list)) and len(timestep) == 2: + v_timestep, a_timestep = timestep + kwargs["a_timestep"] = a_timestep + timestep = v_timestep + else: + kwargs["a_timestep"] = timestep + + # Call parent forward method + return super().forward( + x, + timestep, + context, + attention_mask, + frame_rate, + transformer_options, + keyframe_idxs, + **kwargs, + ) diff --git a/comfy/ldm/lightricks/embeddings_connector.py b/comfy/ldm/lightricks/embeddings_connector.py new file mode 100644 index 000000000..f7a43f3c3 --- /dev/null +++ b/comfy/ldm/lightricks/embeddings_connector.py @@ -0,0 +1,305 @@ +import math +from typing import Optional + +import comfy.ldm.common_dit +import torch +from comfy.ldm.lightricks.model import ( + CrossAttention, + FeedForward, + generate_freq_grid_np, + interleaved_freqs_cis, + split_freqs_cis, +) +from torch import nn + + +class BasicTransformerBlock1D(nn.Module): + r""" + A basic Transformer block. + + Parameters: + + dim (`int`): The number of channels in the input and output. + num_attention_heads (`int`): The number of heads to use for multi-head attention. + attention_head_dim (`int`): The number of channels in each head. + dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use. + activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward. + attention_bias (: + obj: `bool`, *optional*, defaults to `False`): Configure if the attentions should contain a bias parameter. + upcast_attention (`bool`, *optional*): + Whether to upcast the attention computation to float32. This is useful for mixed precision training. + norm_elementwise_affine (`bool`, *optional*, defaults to `True`): + Whether to use learnable elementwise affine parameters for normalization. + standardization_norm (`str`, *optional*, defaults to `"layer_norm"`): The type of pre-normalization to use. Can be `"layer_norm"` or `"rms_norm"`. + norm_eps (`float`, *optional*, defaults to 1e-5): Epsilon value for normalization layers. + qk_norm (`str`, *optional*, defaults to None): + Set to 'layer_norm' or `rms_norm` to perform query and key normalization. + final_dropout (`bool` *optional*, defaults to False): + Whether to apply a final dropout after the last feed-forward layer. + ff_inner_dim (`int`, *optional*): Dimension of the inner feed-forward layer. If not provided, defaults to `dim * 4`. + ff_bias (`bool`, *optional*, defaults to `True`): Whether to use bias in the feed-forward layer. + attention_out_bias (`bool`, *optional*, defaults to `True`): Whether to use bias in the attention output layer. + use_rope (`bool`, *optional*, defaults to `False`): Whether to use Rotary Position Embeddings (RoPE). + ffn_dim_mult (`int`, *optional*, defaults to 4): Multiplier for the inner dimension of the feed-forward layer. + """ + + def __init__( + self, + dim, + n_heads, + d_head, + context_dim=None, + attn_precision=None, + dtype=None, + device=None, + operations=None, + ): + super().__init__() + + # Define 3 blocks. Each block has its own normalization layer. + # 1. Self-Attn + self.attn1 = CrossAttention( + query_dim=dim, + heads=n_heads, + dim_head=d_head, + context_dim=None, + dtype=dtype, + device=device, + operations=operations, + ) + + # 3. Feed-forward + self.ff = FeedForward( + dim, + dim_out=dim, + glu=True, + dtype=dtype, + device=device, + operations=operations, + ) + + def forward(self, hidden_states, attention_mask=None, pe=None) -> torch.FloatTensor: + + # Notice that normalization is always applied before the real computation in the following blocks. + + # 1. Normalization Before Self-Attention + norm_hidden_states = comfy.ldm.common_dit.rms_norm(hidden_states) + + norm_hidden_states = norm_hidden_states.squeeze(1) + + # 2. Self-Attention + attn_output = self.attn1(norm_hidden_states, mask=attention_mask, pe=pe) + + hidden_states = attn_output + hidden_states + if hidden_states.ndim == 4: + hidden_states = hidden_states.squeeze(1) + + # 3. Normalization before Feed-Forward + norm_hidden_states = comfy.ldm.common_dit.rms_norm(hidden_states) + + # 4. Feed-forward + ff_output = self.ff(norm_hidden_states) + + hidden_states = ff_output + hidden_states + if hidden_states.ndim == 4: + hidden_states = hidden_states.squeeze(1) + + return hidden_states + + +class Embeddings1DConnector(nn.Module): + _supports_gradient_checkpointing = True + + def __init__( + self, + in_channels=128, + cross_attention_dim=2048, + attention_head_dim=128, + num_attention_heads=30, + num_layers=2, + positional_embedding_theta=10000.0, + positional_embedding_max_pos=[4096], + causal_temporal_positioning=False, + num_learnable_registers: Optional[int] = 128, + dtype=None, + device=None, + operations=None, + split_rope=False, + double_precision_rope=False, + **kwargs, + ): + super().__init__() + self.dtype = dtype + self.out_channels = in_channels + self.num_attention_heads = num_attention_heads + self.inner_dim = num_attention_heads * attention_head_dim + self.causal_temporal_positioning = causal_temporal_positioning + self.positional_embedding_theta = positional_embedding_theta + self.positional_embedding_max_pos = positional_embedding_max_pos + self.split_rope = split_rope + self.double_precision_rope = double_precision_rope + self.transformer_1d_blocks = nn.ModuleList( + [ + BasicTransformerBlock1D( + self.inner_dim, + num_attention_heads, + attention_head_dim, + context_dim=cross_attention_dim, + dtype=dtype, + device=device, + operations=operations, + ) + for _ in range(num_layers) + ] + ) + + inner_dim = num_attention_heads * attention_head_dim + self.num_learnable_registers = num_learnable_registers + if self.num_learnable_registers: + self.learnable_registers = nn.Parameter( + torch.rand( + self.num_learnable_registers, inner_dim, dtype=dtype, device=device + ) + * 2.0 + - 1.0 + ) + + def get_fractional_positions(self, indices_grid): + fractional_positions = torch.stack( + [ + indices_grid[:, i] / self.positional_embedding_max_pos[i] + for i in range(1) + ], + dim=-1, + ) + return fractional_positions + + def precompute_freqs(self, indices_grid, spacing): + source_dtype = indices_grid.dtype + dtype = ( + torch.float32 + if source_dtype in (torch.bfloat16, torch.float16) + else source_dtype + ) + + fractional_positions = self.get_fractional_positions(indices_grid) + indices = ( + generate_freq_grid_np( + self.positional_embedding_theta, + indices_grid.shape[1], + self.inner_dim, + ) + if self.double_precision_rope + else self.generate_freq_grid(spacing, dtype, fractional_positions.device) + ).to(device=fractional_positions.device) + + if spacing == "exp_2": + freqs = ( + (indices * fractional_positions.unsqueeze(-1)) + .transpose(-1, -2) + .flatten(2) + ) + else: + freqs = ( + (indices * (fractional_positions.unsqueeze(-1) * 2 - 1)) + .transpose(-1, -2) + .flatten(2) + ) + return freqs + + def generate_freq_grid(self, spacing, dtype, device): + dim = self.inner_dim + theta = self.positional_embedding_theta + n_pos_dims = 1 + n_elem = 2 * n_pos_dims # 2 for cos and sin e.g. x 3 = 6 + start = 1 + end = theta + + if spacing == "exp": + indices = theta ** (torch.arange(0, dim, n_elem, device="cpu", dtype=torch.float32) / (dim - n_elem)) + indices = indices.to(dtype=dtype, device=device) + elif spacing == "exp_2": + indices = 1.0 / theta ** (torch.arange(0, dim, n_elem, device=device) / dim) + indices = indices.to(dtype=dtype) + elif spacing == "linear": + indices = torch.linspace( + start, end, dim // n_elem, device=device, dtype=dtype + ) + elif spacing == "sqrt": + indices = torch.linspace( + start**2, end**2, dim // n_elem, device=device, dtype=dtype + ).sqrt() + + indices = indices * math.pi / 2 + + return indices + + def precompute_freqs_cis(self, indices_grid, spacing="exp"): + dim = self.inner_dim + n_elem = 2 # 2 because of cos and sin + freqs = self.precompute_freqs(indices_grid, spacing) + if self.split_rope: + expected_freqs = dim // 2 + current_freqs = freqs.shape[-1] + pad_size = expected_freqs - current_freqs + cos_freq, sin_freq = split_freqs_cis( + freqs, pad_size, self.num_attention_heads + ) + else: + cos_freq, sin_freq = interleaved_freqs_cis(freqs, dim % n_elem) + return cos_freq.to(self.dtype), sin_freq.to(self.dtype), self.split_rope + + def forward( + self, + hidden_states: torch.Tensor, + attention_mask: Optional[torch.Tensor] = None, + ): + """ + The [`Transformer2DModel`] forward method. + + Args: + hidden_states (`torch.LongTensor` of shape `(batch size, num latent pixels)` if discrete, `torch.FloatTensor` of shape `(batch size, channel, height, width)` if continuous): + Input `hidden_states`. + indices_grid (`torch.LongTensor` of shape `(batch size, 3, num latent pixels)`): + attention_mask ( `torch.Tensor`, *optional*): + An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask + is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large + negative values to the attention scores corresponding to "discard" tokens. + Returns: + If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a + `tuple` where the first element is the sample tensor. + """ + # 1. Input + + if self.num_learnable_registers: + num_registers_duplications = math.ceil( + max(1024, hidden_states.shape[1]) / self.num_learnable_registers + ) + learnable_registers = torch.tile( + self.learnable_registers, (num_registers_duplications, 1) + ) + + hidden_states = torch.cat((hidden_states, learnable_registers[hidden_states.shape[1]:].unsqueeze(0).repeat(hidden_states.shape[0], 1, 1)), dim=1) + + if attention_mask is not None: + attention_mask = torch.zeros([1, 1, 1, hidden_states.shape[1]], dtype=attention_mask.dtype, device=attention_mask.device) + + indices_grid = torch.arange( + hidden_states.shape[1], dtype=torch.float32, device=hidden_states.device + ) + indices_grid = indices_grid[None, None, :] + freqs_cis = self.precompute_freqs_cis(indices_grid) + + # 2. Blocks + for block_idx, block in enumerate(self.transformer_1d_blocks): + hidden_states = block( + hidden_states, attention_mask=attention_mask, pe=freqs_cis + ) + + # 3. Output + # if self.output_scale is not None: + # hidden_states = hidden_states / self.output_scale + + hidden_states = comfy.ldm.common_dit.rms_norm(hidden_states) + + return hidden_states, attention_mask diff --git a/comfy/ldm/lightricks/latent_upsampler.py b/comfy/ldm/lightricks/latent_upsampler.py new file mode 100644 index 000000000..78ed7653f --- /dev/null +++ b/comfy/ldm/lightricks/latent_upsampler.py @@ -0,0 +1,292 @@ +from typing import Optional, Tuple +import torch +import torch.nn as nn +import torch.nn.functional as F +from einops import rearrange + + +def _rational_for_scale(scale: float) -> Tuple[int, int]: + mapping = {0.75: (3, 4), 1.5: (3, 2), 2.0: (2, 1), 4.0: (4, 1)} + if float(scale) not in mapping: + raise ValueError( + f"Unsupported spatial_scale {scale}. Choose from {list(mapping.keys())}" + ) + return mapping[float(scale)] + + +class PixelShuffleND(nn.Module): + def __init__(self, dims, upscale_factors=(2, 2, 2)): + super().__init__() + assert dims in [1, 2, 3], "dims must be 1, 2, or 3" + self.dims = dims + self.upscale_factors = upscale_factors + + def forward(self, x): + if self.dims == 3: + return rearrange( + x, + "b (c p1 p2 p3) d h w -> b c (d p1) (h p2) (w p3)", + p1=self.upscale_factors[0], + p2=self.upscale_factors[1], + p3=self.upscale_factors[2], + ) + elif self.dims == 2: + return rearrange( + x, + "b (c p1 p2) h w -> b c (h p1) (w p2)", + p1=self.upscale_factors[0], + p2=self.upscale_factors[1], + ) + elif self.dims == 1: + return rearrange( + x, + "b (c p1) f h w -> b c (f p1) h w", + p1=self.upscale_factors[0], + ) + + +class BlurDownsample(nn.Module): + """ + Anti-aliased spatial downsampling by integer stride using a fixed separable binomial kernel. + Applies only on H,W. Works for dims=2 or dims=3 (per-frame). + """ + + def __init__(self, dims: int, stride: int): + super().__init__() + assert dims in (2, 3) + assert stride >= 1 and isinstance(stride, int) + self.dims = dims + self.stride = stride + + # 5x5 separable binomial kernel [1,4,6,4,1] (outer product), normalized + k = torch.tensor([1.0, 4.0, 6.0, 4.0, 1.0]) + k2d = k[:, None] @ k[None, :] + k2d = (k2d / k2d.sum()).float() # shape (5,5) + self.register_buffer("kernel", k2d[None, None, :, :]) # (1,1,5,5) + + def forward(self, x: torch.Tensor) -> torch.Tensor: + if self.stride == 1: + return x + + def _apply_2d(x2d: torch.Tensor) -> torch.Tensor: + # x2d: (B, C, H, W) + B, C, H, W = x2d.shape + weight = self.kernel.expand(C, 1, 5, 5) # depthwise + x2d = F.conv2d( + x2d, weight=weight, bias=None, stride=self.stride, padding=2, groups=C + ) + return x2d + + if self.dims == 2: + return _apply_2d(x) + else: + # dims == 3: apply per-frame on H,W + b, c, f, h, w = x.shape + x = rearrange(x, "b c f h w -> (b f) c h w") + x = _apply_2d(x) + h2, w2 = x.shape[-2:] + x = rearrange(x, "(b f) c h w -> b c f h w", b=b, f=f, h=h2, w=w2) + return x + + +class SpatialRationalResampler(nn.Module): + """ + Fully-learned rational spatial scaling: up by 'num' via PixelShuffle, then anti-aliased + downsample by 'den' using fixed blur + stride. Operates on H,W only. + + For dims==3, work per-frame for spatial scaling (temporal axis untouched). + """ + + def __init__(self, mid_channels: int, scale: float): + super().__init__() + self.scale = float(scale) + self.num, self.den = _rational_for_scale(self.scale) + self.conv = nn.Conv2d( + mid_channels, (self.num**2) * mid_channels, kernel_size=3, padding=1 + ) + self.pixel_shuffle = PixelShuffleND(2, upscale_factors=(self.num, self.num)) + self.blur_down = BlurDownsample(dims=2, stride=self.den) + + def forward(self, x: torch.Tensor) -> torch.Tensor: + b, c, f, h, w = x.shape + x = rearrange(x, "b c f h w -> (b f) c h w") + x = self.conv(x) + x = self.pixel_shuffle(x) + x = self.blur_down(x) + x = rearrange(x, "(b f) c h w -> b c f h w", b=b, f=f) + return x + + +class ResBlock(nn.Module): + def __init__( + self, channels: int, mid_channels: Optional[int] = None, dims: int = 3 + ): + super().__init__() + if mid_channels is None: + mid_channels = channels + + Conv = nn.Conv2d if dims == 2 else nn.Conv3d + + self.conv1 = Conv(channels, mid_channels, kernel_size=3, padding=1) + self.norm1 = nn.GroupNorm(32, mid_channels) + self.conv2 = Conv(mid_channels, channels, kernel_size=3, padding=1) + self.norm2 = nn.GroupNorm(32, channels) + self.activation = nn.SiLU() + + def forward(self, x: torch.Tensor) -> torch.Tensor: + residual = x + x = self.conv1(x) + x = self.norm1(x) + x = self.activation(x) + x = self.conv2(x) + x = self.norm2(x) + x = self.activation(x + residual) + return x + + +class LatentUpsampler(nn.Module): + """ + Model to spatially upsample VAE latents. + + Args: + in_channels (`int`): Number of channels in the input latent + mid_channels (`int`): Number of channels in the middle layers + num_blocks_per_stage (`int`): Number of ResBlocks to use in each stage (pre/post upsampling) + dims (`int`): Number of dimensions for convolutions (2 or 3) + spatial_upsample (`bool`): Whether to spatially upsample the latent + temporal_upsample (`bool`): Whether to temporally upsample the latent + """ + + def __init__( + self, + in_channels: int = 128, + mid_channels: int = 512, + num_blocks_per_stage: int = 4, + dims: int = 3, + spatial_upsample: bool = True, + temporal_upsample: bool = False, + spatial_scale: float = 2.0, + rational_resampler: bool = False, + ): + super().__init__() + + self.in_channels = in_channels + self.mid_channels = mid_channels + self.num_blocks_per_stage = num_blocks_per_stage + self.dims = dims + self.spatial_upsample = spatial_upsample + self.temporal_upsample = temporal_upsample + self.spatial_scale = float(spatial_scale) + self.rational_resampler = rational_resampler + + Conv = nn.Conv2d if dims == 2 else nn.Conv3d + + self.initial_conv = Conv(in_channels, mid_channels, kernel_size=3, padding=1) + self.initial_norm = nn.GroupNorm(32, mid_channels) + self.initial_activation = nn.SiLU() + + self.res_blocks = nn.ModuleList( + [ResBlock(mid_channels, dims=dims) for _ in range(num_blocks_per_stage)] + ) + + if spatial_upsample and temporal_upsample: + self.upsampler = nn.Sequential( + nn.Conv3d(mid_channels, 8 * mid_channels, kernel_size=3, padding=1), + PixelShuffleND(3), + ) + elif spatial_upsample: + if rational_resampler: + self.upsampler = SpatialRationalResampler( + mid_channels=mid_channels, scale=self.spatial_scale + ) + else: + self.upsampler = nn.Sequential( + nn.Conv2d(mid_channels, 4 * mid_channels, kernel_size=3, padding=1), + PixelShuffleND(2), + ) + elif temporal_upsample: + self.upsampler = nn.Sequential( + nn.Conv3d(mid_channels, 2 * mid_channels, kernel_size=3, padding=1), + PixelShuffleND(1), + ) + else: + raise ValueError( + "Either spatial_upsample or temporal_upsample must be True" + ) + + self.post_upsample_res_blocks = nn.ModuleList( + [ResBlock(mid_channels, dims=dims) for _ in range(num_blocks_per_stage)] + ) + + self.final_conv = Conv(mid_channels, in_channels, kernel_size=3, padding=1) + + def forward(self, latent: torch.Tensor) -> torch.Tensor: + b, c, f, h, w = latent.shape + + if self.dims == 2: + x = rearrange(latent, "b c f h w -> (b f) c h w") + x = self.initial_conv(x) + x = self.initial_norm(x) + x = self.initial_activation(x) + + for block in self.res_blocks: + x = block(x) + + x = self.upsampler(x) + + for block in self.post_upsample_res_blocks: + x = block(x) + + x = self.final_conv(x) + x = rearrange(x, "(b f) c h w -> b c f h w", b=b, f=f) + else: + x = self.initial_conv(latent) + x = self.initial_norm(x) + x = self.initial_activation(x) + + for block in self.res_blocks: + x = block(x) + + if self.temporal_upsample: + x = self.upsampler(x) + x = x[:, :, 1:, :, :] + else: + if isinstance(self.upsampler, SpatialRationalResampler): + x = self.upsampler(x) + else: + x = rearrange(x, "b c f h w -> (b f) c h w") + x = self.upsampler(x) + x = rearrange(x, "(b f) c h w -> b c f h w", b=b, f=f) + + for block in self.post_upsample_res_blocks: + x = block(x) + + x = self.final_conv(x) + + return x + + @classmethod + def from_config(cls, config): + return cls( + in_channels=config.get("in_channels", 4), + mid_channels=config.get("mid_channels", 128), + num_blocks_per_stage=config.get("num_blocks_per_stage", 4), + dims=config.get("dims", 2), + spatial_upsample=config.get("spatial_upsample", True), + temporal_upsample=config.get("temporal_upsample", False), + spatial_scale=config.get("spatial_scale", 2.0), + rational_resampler=config.get("rational_resampler", False), + ) + + def config(self): + return { + "_class_name": "LatentUpsampler", + "in_channels": self.in_channels, + "mid_channels": self.mid_channels, + "num_blocks_per_stage": self.num_blocks_per_stage, + "dims": self.dims, + "spatial_upsample": self.spatial_upsample, + "temporal_upsample": self.temporal_upsample, + "spatial_scale": self.spatial_scale, + "rational_resampler": self.rational_resampler, + } diff --git a/comfy/ldm/lightricks/model.py b/comfy/ldm/lightricks/model.py index 593f7940f..d61e19d6e 100644 --- a/comfy/ldm/lightricks/model.py +++ b/comfy/ldm/lightricks/model.py @@ -1,13 +1,47 @@ +from abc import ABC, abstractmethod +from enum import Enum +import functools +import math +from typing import Dict, Optional, Tuple + +from einops import rearrange +import numpy as np import torch from torch import nn import comfy.patcher_extension import comfy.ldm.modules.attention import comfy.ldm.common_dit -import math -from typing import Dict, Optional, Tuple from .symmetric_patchifier import SymmetricPatchifier, latent_to_pixel_coords -from comfy.ldm.flux.math import apply_rope1 + +def _log_base(x, base): + return np.log(x) / np.log(base) + +class LTXRopeType(str, Enum): + INTERLEAVED = "interleaved" + SPLIT = "split" + + KEY = "rope_type" + + @classmethod + def from_dict(cls, kwargs, default=None): + if default is None: + default = cls.INTERLEAVED + return cls(kwargs.get(cls.KEY, default)) + + +class LTXFrequenciesPrecision(str, Enum): + FLOAT32 = "float32" + FLOAT64 = "float64" + + KEY = "frequencies_precision" + + @classmethod + def from_dict(cls, kwargs, default=None): + if default is None: + default = cls.FLOAT32 + return cls(kwargs.get(cls.KEY, default)) + def get_timestep_embedding( timesteps: torch.Tensor, @@ -39,9 +73,7 @@ def get_timestep_embedding( assert len(timesteps.shape) == 1, "Timesteps should be a 1d-array" half_dim = embedding_dim // 2 - exponent = -math.log(max_period) * torch.arange( - start=0, end=half_dim, dtype=torch.float32, device=timesteps.device - ) + exponent = -math.log(max_period) * torch.arange(start=0, end=half_dim, dtype=torch.float32, device=timesteps.device) exponent = exponent / (half_dim - downscale_freq_shift) emb = torch.exp(exponent) @@ -73,7 +105,9 @@ class TimestepEmbedding(nn.Module): post_act_fn: Optional[str] = None, cond_proj_dim=None, sample_proj_bias=True, - dtype=None, device=None, operations=None, + dtype=None, + device=None, + operations=None, ): super().__init__() @@ -90,7 +124,9 @@ class TimestepEmbedding(nn.Module): time_embed_dim_out = out_dim else: time_embed_dim_out = time_embed_dim - self.linear_2 = operations.Linear(time_embed_dim, time_embed_dim_out, sample_proj_bias, dtype=dtype, device=device) + self.linear_2 = operations.Linear( + time_embed_dim, time_embed_dim_out, sample_proj_bias, dtype=dtype, device=device + ) if post_act_fn is None: self.post_act = None @@ -139,12 +175,22 @@ class PixArtAlphaCombinedTimestepSizeEmbeddings(nn.Module): https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L164C9-L168C29 """ - def __init__(self, embedding_dim, size_emb_dim, use_additional_conditions: bool = False, dtype=None, device=None, operations=None): + def __init__( + self, + embedding_dim, + size_emb_dim, + use_additional_conditions: bool = False, + dtype=None, + device=None, + operations=None, + ): super().__init__() self.outdim = size_emb_dim self.time_proj = Timesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=0) - self.timestep_embedder = TimestepEmbedding(in_channels=256, time_embed_dim=embedding_dim, dtype=dtype, device=device, operations=operations) + self.timestep_embedder = TimestepEmbedding( + in_channels=256, time_embed_dim=embedding_dim, dtype=dtype, device=device, operations=operations + ) def forward(self, timestep, resolution, aspect_ratio, batch_size, hidden_dtype): timesteps_proj = self.time_proj(timestep) @@ -163,15 +209,22 @@ class AdaLayerNormSingle(nn.Module): use_additional_conditions (`bool`): To use additional conditions for normalization or not. """ - def __init__(self, embedding_dim: int, use_additional_conditions: bool = False, dtype=None, device=None, operations=None): + def __init__( + self, embedding_dim: int, embedding_coefficient: int = 6, use_additional_conditions: bool = False, dtype=None, device=None, operations=None + ): super().__init__() self.emb = PixArtAlphaCombinedTimestepSizeEmbeddings( - embedding_dim, size_emb_dim=embedding_dim // 3, use_additional_conditions=use_additional_conditions, dtype=dtype, device=device, operations=operations + embedding_dim, + size_emb_dim=embedding_dim // 3, + use_additional_conditions=use_additional_conditions, + dtype=dtype, + device=device, + operations=operations, ) self.silu = nn.SiLU() - self.linear = operations.Linear(embedding_dim, 6 * embedding_dim, bias=True, dtype=dtype, device=device) + self.linear = operations.Linear(embedding_dim, embedding_coefficient * embedding_dim, bias=True, dtype=dtype, device=device) def forward( self, @@ -185,6 +238,7 @@ class AdaLayerNormSingle(nn.Module): embedded_timestep = self.emb(timestep, **added_cond_kwargs, batch_size=batch_size, hidden_dtype=hidden_dtype) return self.linear(self.silu(embedded_timestep)), embedded_timestep + class PixArtAlphaTextProjection(nn.Module): """ Projects caption embeddings. Also handles dropout for classifier-free guidance. @@ -192,18 +246,24 @@ class PixArtAlphaTextProjection(nn.Module): Adapted from https://github.com/PixArt-alpha/PixArt-alpha/blob/master/diffusion/model/nets/PixArt_blocks.py """ - def __init__(self, in_features, hidden_size, out_features=None, act_fn="gelu_tanh", dtype=None, device=None, operations=None): + def __init__( + self, in_features, hidden_size, out_features=None, act_fn="gelu_tanh", dtype=None, device=None, operations=None + ): super().__init__() if out_features is None: out_features = hidden_size - self.linear_1 = operations.Linear(in_features=in_features, out_features=hidden_size, bias=True, dtype=dtype, device=device) + self.linear_1 = operations.Linear( + in_features=in_features, out_features=hidden_size, bias=True, dtype=dtype, device=device + ) if act_fn == "gelu_tanh": self.act_1 = nn.GELU(approximate="tanh") elif act_fn == "silu": self.act_1 = nn.SiLU() else: raise ValueError(f"Unknown activation function: {act_fn}") - self.linear_2 = operations.Linear(in_features=hidden_size, out_features=out_features, bias=True, dtype=dtype, device=device) + self.linear_2 = operations.Linear( + in_features=hidden_size, out_features=out_features, bias=True, dtype=dtype, device=device + ) def forward(self, caption): hidden_states = self.linear_1(caption) @@ -222,23 +282,68 @@ class GELU_approx(nn.Module): class FeedForward(nn.Module): - def __init__(self, dim, dim_out, mult=4, glu=False, dropout=0., dtype=None, device=None, operations=None): + def __init__(self, dim, dim_out, mult=4, glu=False, dropout=0.0, dtype=None, device=None, operations=None): super().__init__() inner_dim = int(dim * mult) project_in = GELU_approx(dim, inner_dim, dtype=dtype, device=device, operations=operations) self.net = nn.Sequential( - project_in, - nn.Dropout(dropout), - operations.Linear(inner_dim, dim_out, dtype=dtype, device=device) + project_in, nn.Dropout(dropout), operations.Linear(inner_dim, dim_out, dtype=dtype, device=device) ) def forward(self, x): return self.net(x) +def apply_rotary_emb(input_tensor, freqs_cis): + cos_freqs, sin_freqs = freqs_cis[0], freqs_cis[1] + split_pe = freqs_cis[2] if len(freqs_cis) > 2 else False + return ( + apply_split_rotary_emb(input_tensor, cos_freqs, sin_freqs) + if split_pe else + apply_interleaved_rotary_emb(input_tensor, cos_freqs, sin_freqs) + ) + +def apply_interleaved_rotary_emb(input_tensor, cos_freqs, sin_freqs): # TODO: remove duplicate funcs and pick the best/fastest one + t_dup = rearrange(input_tensor, "... (d r) -> ... d r", r=2) + t1, t2 = t_dup.unbind(dim=-1) + t_dup = torch.stack((-t2, t1), dim=-1) + input_tensor_rot = rearrange(t_dup, "... d r -> ... (d r)") + + out = input_tensor * cos_freqs + input_tensor_rot * sin_freqs + + return out + +def apply_split_rotary_emb(input_tensor, cos, sin): + needs_reshape = False + if input_tensor.ndim != 4 and cos.ndim == 4: + B, H, T, _ = cos.shape + input_tensor = input_tensor.reshape(B, T, H, -1).swapaxes(1, 2) + needs_reshape = True + split_input = rearrange(input_tensor, "... (d r) -> ... d r", d=2) + first_half_input = split_input[..., :1, :] + second_half_input = split_input[..., 1:, :] + output = split_input * cos.unsqueeze(-2) + first_half_output = output[..., :1, :] + second_half_output = output[..., 1:, :] + first_half_output.addcmul_(-sin.unsqueeze(-2), second_half_input) + second_half_output.addcmul_(sin.unsqueeze(-2), first_half_input) + output = rearrange(output, "... d r -> ... (d r)") + return output.swapaxes(1, 2).reshape(B, T, -1) if needs_reshape else output + class CrossAttention(nn.Module): - def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0., attn_precision=None, dtype=None, device=None, operations=None): + def __init__( + self, + query_dim, + context_dim=None, + heads=8, + dim_head=64, + dropout=0.0, + attn_precision=None, + dtype=None, + device=None, + operations=None, + ): super().__init__() inner_dim = dim_head * heads context_dim = query_dim if context_dim is None else context_dim @@ -254,9 +359,11 @@ class CrossAttention(nn.Module): self.to_k = operations.Linear(context_dim, inner_dim, bias=True, dtype=dtype, device=device) self.to_v = operations.Linear(context_dim, inner_dim, bias=True, dtype=dtype, device=device) - self.to_out = nn.Sequential(operations.Linear(inner_dim, query_dim, dtype=dtype, device=device), nn.Dropout(dropout)) + self.to_out = nn.Sequential( + operations.Linear(inner_dim, query_dim, dtype=dtype, device=device), nn.Dropout(dropout) + ) - def forward(self, x, context=None, mask=None, pe=None, transformer_options={}): + def forward(self, x, context=None, mask=None, pe=None, k_pe=None, transformer_options={}): q = self.to_q(x) context = x if context is None else context k = self.to_k(context) @@ -266,8 +373,8 @@ class CrossAttention(nn.Module): k = self.k_norm(k) if pe is not None: - q = apply_rope1(q.unsqueeze(1), pe).squeeze(1) - k = apply_rope1(k.unsqueeze(1), pe).squeeze(1) + q = apply_rotary_emb(q, pe) + k = apply_rotary_emb(k, pe if k_pe is None else k_pe) if mask is None: out = comfy.ldm.modules.attention.optimized_attention(q, k, v, self.heads, attn_precision=self.attn_precision, transformer_options=transformer_options) @@ -277,14 +384,34 @@ class CrossAttention(nn.Module): class BasicTransformerBlock(nn.Module): - def __init__(self, dim, n_heads, d_head, context_dim=None, attn_precision=None, dtype=None, device=None, operations=None): + def __init__( + self, dim, n_heads, d_head, context_dim=None, attn_precision=None, dtype=None, device=None, operations=None + ): super().__init__() self.attn_precision = attn_precision - self.attn1 = CrossAttention(query_dim=dim, heads=n_heads, dim_head=d_head, context_dim=None, attn_precision=self.attn_precision, dtype=dtype, device=device, operations=operations) + self.attn1 = CrossAttention( + query_dim=dim, + heads=n_heads, + dim_head=d_head, + context_dim=None, + attn_precision=self.attn_precision, + dtype=dtype, + device=device, + operations=operations, + ) self.ff = FeedForward(dim, dim_out=dim, glu=True, dtype=dtype, device=device, operations=operations) - self.attn2 = CrossAttention(query_dim=dim, context_dim=context_dim, heads=n_heads, dim_head=d_head, attn_precision=self.attn_precision, dtype=dtype, device=device, operations=operations) + self.attn2 = CrossAttention( + query_dim=dim, + context_dim=context_dim, + heads=n_heads, + dim_head=d_head, + attn_precision=self.attn_precision, + dtype=dtype, + device=device, + operations=operations, + ) self.scale_shift_table = nn.Parameter(torch.empty(6, dim, device=device, dtype=dtype)) @@ -306,116 +433,446 @@ class BasicTransformerBlock(nn.Module): return x def get_fractional_positions(indices_grid, max_pos): + n_pos_dims = indices_grid.shape[1] + assert n_pos_dims == len(max_pos), f'Number of position dimensions ({n_pos_dims}) must match max_pos length ({len(max_pos)})' fractional_positions = torch.stack( - [ - indices_grid[:, i] / max_pos[i] - for i in range(3) - ], - dim=-1, + [indices_grid[:, i] / max_pos[i] for i in range(n_pos_dims)], + axis=-1, ) return fractional_positions -def precompute_freqs_cis(indices_grid, dim, out_dtype, theta=10000.0, max_pos=[20, 2048, 2048]): - dtype = torch.float32 - device = indices_grid.device +@functools.lru_cache(maxsize=5) +def generate_freq_grid_np(positional_embedding_theta, positional_embedding_max_pos_count, inner_dim, _ = None): + theta = positional_embedding_theta + start = 1 + end = theta + + n_elem = 2 * positional_embedding_max_pos_count + pow_indices = np.power( + theta, + np.linspace( + _log_base(start, theta), + _log_base(end, theta), + inner_dim // n_elem, + dtype=np.float64, + ), + ) + return torch.tensor(pow_indices * math.pi / 2, dtype=torch.float32) + +def generate_freq_grid_pytorch(positional_embedding_theta, positional_embedding_max_pos_count, inner_dim, device): + theta = positional_embedding_theta + start = 1 + end = theta + n_elem = 2 * positional_embedding_max_pos_count + + indices = theta ** ( + torch.linspace( + math.log(start, theta), + math.log(end, theta), + inner_dim // n_elem, + device=device, + dtype=torch.float32, + ) + ) + indices = indices.to(dtype=torch.float32) + + indices = indices * math.pi / 2 + + return indices + +def generate_freqs(indices, indices_grid, max_pos, use_middle_indices_grid): + if use_middle_indices_grid: + assert(len(indices_grid.shape) == 4 and indices_grid.shape[-1] ==2) + indices_grid_start, indices_grid_end = indices_grid[..., 0], indices_grid[..., 1] + indices_grid = (indices_grid_start + indices_grid_end) / 2.0 + elif len(indices_grid.shape) == 4: + indices_grid = indices_grid[..., 0] # Get fractional positions and compute frequency indices fractional_positions = get_fractional_positions(indices_grid, max_pos) - indices = theta ** torch.linspace(0, 1, dim // 6, device=device, dtype=dtype) * math.pi / 2 + indices = indices.to(device=fractional_positions.device) - # Compute frequencies and apply cos/sin - freqs = (indices * (fractional_positions.unsqueeze(-1) * 2 - 1)).transpose(-1, -2).flatten(2) - cos_vals = freqs.cos().repeat_interleave(2, dim=-1) - sin_vals = freqs.sin().repeat_interleave(2, dim=-1) + freqs = ( + (indices * (fractional_positions.unsqueeze(-1) * 2 - 1)) + .transpose(-1, -2) + .flatten(2) + ) + return freqs - # Pad if dim is not divisible by 6 - if dim % 6 != 0: - padding_size = dim % 6 - cos_vals = torch.cat([torch.ones_like(cos_vals[:, :, :padding_size]), cos_vals], dim=-1) - sin_vals = torch.cat([torch.zeros_like(sin_vals[:, :, :padding_size]), sin_vals], dim=-1) +def interleaved_freqs_cis(freqs, pad_size): + cos_freq = freqs.cos().repeat_interleave(2, dim=-1) + sin_freq = freqs.sin().repeat_interleave(2, dim=-1) + if pad_size != 0: + cos_padding = torch.ones_like(cos_freq[:, :, : pad_size]) + sin_padding = torch.zeros_like(cos_freq[:, :, : pad_size]) + cos_freq = torch.cat([cos_padding, cos_freq], dim=-1) + sin_freq = torch.cat([sin_padding, sin_freq], dim=-1) + return cos_freq, sin_freq - # Reshape and extract one value per pair (since repeat_interleave duplicates each value) - cos_vals = cos_vals.reshape(*cos_vals.shape[:2], -1, 2)[..., 0].to(out_dtype) # [B, N, dim//2] - sin_vals = sin_vals.reshape(*sin_vals.shape[:2], -1, 2)[..., 0].to(out_dtype) # [B, N, dim//2] +def split_freqs_cis(freqs, pad_size, num_attention_heads): + cos_freq = freqs.cos() + sin_freq = freqs.sin() - # Build rotation matrix [[cos, -sin], [sin, cos]] and add heads dimension - freqs_cis = torch.stack([ - torch.stack([cos_vals, -sin_vals], dim=-1), - torch.stack([sin_vals, cos_vals], dim=-1) - ], dim=-2).unsqueeze(1) # [B, 1, N, dim//2, 2, 2] + if pad_size != 0: + cos_padding = torch.ones_like(cos_freq[:, :, :pad_size]) + sin_padding = torch.zeros_like(sin_freq[:, :, :pad_size]) - return freqs_cis + cos_freq = torch.concatenate([cos_padding, cos_freq], axis=-1) + sin_freq = torch.concatenate([sin_padding, sin_freq], axis=-1) + # Reshape freqs to be compatible with multi-head attention + B , T, half_HD = cos_freq.shape -class LTXVModel(torch.nn.Module): - def __init__(self, - in_channels=128, - cross_attention_dim=2048, - attention_head_dim=64, - num_attention_heads=32, + cos_freq = cos_freq.reshape(B, T, num_attention_heads, half_HD // num_attention_heads) + sin_freq = sin_freq.reshape(B, T, num_attention_heads, half_HD // num_attention_heads) - caption_channels=4096, - num_layers=28, + cos_freq = torch.swapaxes(cos_freq, 1, 2) # (B,H,T,D//2) + sin_freq = torch.swapaxes(sin_freq, 1, 2) # (B,H,T,D//2) + return cos_freq, sin_freq +class LTXBaseModel(torch.nn.Module, ABC): + """ + Abstract base class for LTX models (Lightricks Transformer models). - positional_embedding_theta=10000.0, - positional_embedding_max_pos=[20, 2048, 2048], - causal_temporal_positioning=False, - vae_scale_factors=(8, 32, 32), - dtype=None, device=None, operations=None, **kwargs): + This class defines the common interface and shared functionality for all LTX models, + including LTXV (video) and LTXAV (audio-video) variants. + """ + + def __init__( + self, + in_channels: int, + cross_attention_dim: int, + attention_head_dim: int, + num_attention_heads: int, + caption_channels: int, + num_layers: int, + positional_embedding_theta: float = 10000.0, + positional_embedding_max_pos: list = [20, 2048, 2048], + causal_temporal_positioning: bool = False, + vae_scale_factors: tuple = (8, 32, 32), + use_middle_indices_grid=False, + timestep_scale_multiplier = 1000.0, + dtype=None, + device=None, + operations=None, + **kwargs, + ): super().__init__() self.generator = None self.vae_scale_factors = vae_scale_factors + self.use_middle_indices_grid = use_middle_indices_grid self.dtype = dtype - self.out_channels = in_channels - self.inner_dim = num_attention_heads * attention_head_dim + self.in_channels = in_channels + self.cross_attention_dim = cross_attention_dim + self.attention_head_dim = attention_head_dim + self.num_attention_heads = num_attention_heads + self.caption_channels = caption_channels + self.num_layers = num_layers + self.positional_embedding_theta = positional_embedding_theta + self.positional_embedding_max_pos = positional_embedding_max_pos + self.split_positional_embedding = LTXRopeType.from_dict(kwargs) + self.freq_grid_generator = ( + generate_freq_grid_np if LTXFrequenciesPrecision.from_dict(kwargs) == LTXFrequenciesPrecision.FLOAT64 + else generate_freq_grid_pytorch + ) self.causal_temporal_positioning = causal_temporal_positioning + self.operations = operations + self.timestep_scale_multiplier = timestep_scale_multiplier - self.patchify_proj = operations.Linear(in_channels, self.inner_dim, bias=True, dtype=dtype, device=device) + # Common dimensions + self.inner_dim = num_attention_heads * attention_head_dim + self.out_channels = in_channels + + # Initialize common components + self._init_common_components(device, dtype) + + # Initialize model-specific components + self._init_model_components(device, dtype, **kwargs) + + # Initialize transformer blocks + self._init_transformer_blocks(device, dtype, **kwargs) + + # Initialize output components + self._init_output_components(device, dtype) + + def _init_common_components(self, device, dtype): + """Initialize components common to all LTX models + - patchify_proj: Linear projection for patchifying input + - adaln_single: AdaLN layer for timestep embedding + - caption_projection: Linear projection for caption embedding + """ + self.patchify_proj = self.operations.Linear( + self.in_channels, self.inner_dim, bias=True, dtype=dtype, device=device + ) self.adaln_single = AdaLayerNormSingle( - self.inner_dim, use_additional_conditions=False, dtype=dtype, device=device, operations=operations + self.inner_dim, use_additional_conditions=False, dtype=dtype, device=device, operations=self.operations ) - # self.adaln_single.linear = operations.Linear(self.inner_dim, 4 * self.inner_dim, bias=True, dtype=dtype, device=device) - self.caption_projection = PixArtAlphaTextProjection( - in_features=caption_channels, hidden_size=self.inner_dim, dtype=dtype, device=device, operations=operations + in_features=self.caption_channels, + hidden_size=self.inner_dim, + dtype=dtype, + device=device, + operations=self.operations, ) + @abstractmethod + def _init_model_components(self, device, dtype, **kwargs): + """Initialize model-specific components. Must be implemented by subclasses.""" + pass + + @abstractmethod + def _init_transformer_blocks(self, device, dtype, **kwargs): + """Initialize transformer blocks. Must be implemented by subclasses.""" + pass + + @abstractmethod + def _init_output_components(self, device, dtype): + """Initialize output components. Must be implemented by subclasses.""" + pass + + @abstractmethod + def _process_input(self, x, keyframe_idxs, denoise_mask, **kwargs): + """Process input data. Must be implemented by subclasses.""" + pass + + @abstractmethod + def _process_transformer_blocks(self, x, context, attention_mask, timestep, pe, **kwargs): + """Process transformer blocks. Must be implemented by subclasses.""" + pass + + @abstractmethod + def _process_output(self, x, embedded_timestep, keyframe_idxs, **kwargs): + """Process output data. Must be implemented by subclasses.""" + pass + + def _prepare_timestep(self, timestep, batch_size, hidden_dtype, **kwargs): + """Prepare timestep embeddings.""" + grid_mask = kwargs.get("grid_mask", None) + if grid_mask is not None: + timestep = timestep[:, grid_mask] + + timestep = timestep * self.timestep_scale_multiplier + timestep, embedded_timestep = self.adaln_single( + timestep.flatten(), + {"resolution": None, "aspect_ratio": None}, + batch_size=batch_size, + hidden_dtype=hidden_dtype, + ) + + # Second dimension is 1 or number of tokens (if timestep_per_token) + timestep = timestep.view(batch_size, -1, timestep.shape[-1]) + embedded_timestep = embedded_timestep.view(batch_size, -1, embedded_timestep.shape[-1]) + + return timestep, embedded_timestep + + def _prepare_context(self, context, batch_size, x, attention_mask=None): + """Prepare context for transformer blocks.""" + if self.caption_projection is not None: + context = self.caption_projection(context) + context = context.view(batch_size, -1, x.shape[-1]) + + return context, attention_mask + + def _precompute_freqs_cis( + self, + indices_grid, + dim, + out_dtype, + theta=10000.0, + max_pos=[20, 2048, 2048], + use_middle_indices_grid=False, + num_attention_heads=32, + ): + split_mode = self.split_positional_embedding == LTXRopeType.SPLIT + indices = self.freq_grid_generator(theta, indices_grid.shape[1], dim, indices_grid.device) + freqs = generate_freqs(indices, indices_grid, max_pos, use_middle_indices_grid) + + if split_mode: + expected_freqs = dim // 2 + current_freqs = freqs.shape[-1] + pad_size = expected_freqs - current_freqs + cos_freq, sin_freq = split_freqs_cis(freqs, pad_size, num_attention_heads) + else: + # 2 because of cos and sin by 3 for (t, x, y), 1 for temporal only + n_elem = 2 * indices_grid.shape[1] + cos_freq, sin_freq = interleaved_freqs_cis(freqs, dim % n_elem) + return cos_freq.to(out_dtype), sin_freq.to(out_dtype), split_mode + + def _prepare_positional_embeddings(self, pixel_coords, frame_rate, x_dtype): + """Prepare positional embeddings.""" + fractional_coords = pixel_coords.to(torch.float32) + fractional_coords[:, 0] = fractional_coords[:, 0] * (1.0 / frame_rate) + pe = self._precompute_freqs_cis( + fractional_coords, + dim=self.inner_dim, + out_dtype=x_dtype, + max_pos=self.positional_embedding_max_pos, + use_middle_indices_grid=self.use_middle_indices_grid, + num_attention_heads=self.num_attention_heads, + ) + return pe + + def _prepare_attention_mask(self, attention_mask, x_dtype): + """Prepare attention mask.""" + if attention_mask is not None and not torch.is_floating_point(attention_mask): + attention_mask = (attention_mask - 1).to(x_dtype).reshape( + (attention_mask.shape[0], 1, -1, attention_mask.shape[-1]) + ) * torch.finfo(x_dtype).max + return attention_mask + + def forward( + self, x, timestep, context, attention_mask, frame_rate=25, transformer_options={}, keyframe_idxs=None, denoise_mask=None, **kwargs + ): + """ + Forward pass for LTX models. + + Args: + x: Input tensor + timestep: Timestep tensor + context: Context tensor (e.g., text embeddings) + attention_mask: Attention mask tensor + frame_rate: Frame rate for temporal processing + transformer_options: Additional options for transformer blocks + keyframe_idxs: Keyframe indices for temporal processing + **kwargs: Additional keyword arguments + + Returns: + Processed output tensor + """ + return comfy.patcher_extension.WrapperExecutor.new_class_executor( + self._forward, + self, + comfy.patcher_extension.get_all_wrappers( + comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options + ), + ).execute(x, timestep, context, attention_mask, frame_rate, transformer_options, keyframe_idxs, denoise_mask=denoise_mask, **kwargs) + + def _forward( + self, x, timestep, context, attention_mask, frame_rate=25, transformer_options={}, keyframe_idxs=None, denoise_mask=None, **kwargs + ): + """ + Internal forward pass for LTX models. + + Args: + x: Input tensor + timestep: Timestep tensor + context: Context tensor (e.g., text embeddings) + attention_mask: Attention mask tensor + frame_rate: Frame rate for temporal processing + transformer_options: Additional options for transformer blocks + keyframe_idxs: Keyframe indices for temporal processing + **kwargs: Additional keyword arguments + + Returns: + Processed output tensor + """ + if isinstance(x, list): + input_dtype = x[0].dtype + batch_size = x[0].shape[0] + else: + input_dtype = x.dtype + batch_size = x.shape[0] + # Process input + merged_args = {**transformer_options, **kwargs} + x, pixel_coords, additional_args = self._process_input(x, keyframe_idxs, denoise_mask, **merged_args) + merged_args.update(additional_args) + + # Prepare timestep and context + timestep, embedded_timestep = self._prepare_timestep(timestep, batch_size, input_dtype, **merged_args) + context, attention_mask = self._prepare_context(context, batch_size, x, attention_mask) + + # Prepare attention mask and positional embeddings + attention_mask = self._prepare_attention_mask(attention_mask, input_dtype) + pe = self._prepare_positional_embeddings(pixel_coords, frame_rate, input_dtype) + + # Process transformer blocks + x = self._process_transformer_blocks( + x, context, attention_mask, timestep, pe, transformer_options=transformer_options, **merged_args + ) + + # Process output + x = self._process_output(x, embedded_timestep, keyframe_idxs, **merged_args) + return x + + +class LTXVModel(LTXBaseModel): + """LTXV model for video generation.""" + + def __init__( + self, + in_channels=128, + cross_attention_dim=2048, + attention_head_dim=64, + num_attention_heads=32, + caption_channels=4096, + num_layers=28, + positional_embedding_theta=10000.0, + positional_embedding_max_pos=[20, 2048, 2048], + causal_temporal_positioning=False, + vae_scale_factors=(8, 32, 32), + use_middle_indices_grid=False, + timestep_scale_multiplier = 1000.0, + dtype=None, + device=None, + operations=None, + **kwargs, + ): + super().__init__( + in_channels=in_channels, + cross_attention_dim=cross_attention_dim, + attention_head_dim=attention_head_dim, + num_attention_heads=num_attention_heads, + caption_channels=caption_channels, + num_layers=num_layers, + positional_embedding_theta=positional_embedding_theta, + positional_embedding_max_pos=positional_embedding_max_pos, + causal_temporal_positioning=causal_temporal_positioning, + vae_scale_factors=vae_scale_factors, + use_middle_indices_grid=use_middle_indices_grid, + timestep_scale_multiplier=timestep_scale_multiplier, + dtype=dtype, + device=device, + operations=operations, + **kwargs, + ) + + def _init_model_components(self, device, dtype, **kwargs): + """Initialize LTXV-specific components.""" + # No additional components needed for LTXV beyond base class + pass + + def _init_transformer_blocks(self, device, dtype, **kwargs): + """Initialize transformer blocks for LTXV.""" self.transformer_blocks = nn.ModuleList( [ BasicTransformerBlock( self.inner_dim, - num_attention_heads, - attention_head_dim, - context_dim=cross_attention_dim, - # attn_precision=attn_precision, - dtype=dtype, device=device, operations=operations + self.num_attention_heads, + self.attention_head_dim, + context_dim=self.cross_attention_dim, + dtype=dtype, + device=device, + operations=self.operations, ) - for d in range(num_layers) + for _ in range(self.num_layers) ] ) + def _init_output_components(self, device, dtype): + """Initialize output components for LTXV.""" self.scale_shift_table = nn.Parameter(torch.empty(2, self.inner_dim, dtype=dtype, device=device)) - self.norm_out = operations.LayerNorm(self.inner_dim, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device) - self.proj_out = operations.Linear(self.inner_dim, self.out_channels, dtype=dtype, device=device) - - self.patchifier = SymmetricPatchifier(1) - - def forward(self, x, timestep, context, attention_mask, frame_rate=25, transformer_options={}, keyframe_idxs=None, **kwargs): - return comfy.patcher_extension.WrapperExecutor.new_class_executor( - self._forward, - self, - comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options) - ).execute(x, timestep, context, attention_mask, frame_rate, transformer_options, keyframe_idxs, **kwargs) - - def _forward(self, x, timestep, context, attention_mask, frame_rate=25, transformer_options={}, keyframe_idxs=None, **kwargs): - patches_replace = transformer_options.get("patches_replace", {}) - - orig_shape = list(x.shape) + self.norm_out = self.operations.LayerNorm( + self.inner_dim, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device + ) + self.proj_out = self.operations.Linear(self.inner_dim, self.out_channels, dtype=dtype, device=device) + self.patchifier = SymmetricPatchifier(1, start_end=True) + def _process_input(self, x, keyframe_idxs, denoise_mask, **kwargs): + """Process input for LTXV.""" + additional_args = {"orig_shape": list(x.shape)} x, latent_coords = self.patchifier.patchify(x) pixel_coords = latent_to_pixel_coords( latent_coords=latent_coords, @@ -423,44 +880,30 @@ class LTXVModel(torch.nn.Module): causal_fix=self.causal_temporal_positioning, ) + grid_mask = None if keyframe_idxs is not None: - pixel_coords[:, :, -keyframe_idxs.shape[2]:] = keyframe_idxs + additional_args.update({ "orig_patchified_shape": list(x.shape)}) + denoise_mask = self.patchifier.patchify(denoise_mask)[0] + grid_mask = ~torch.any(denoise_mask < 0, dim=-1)[0] + additional_args.update({"grid_mask": grid_mask}) + x = x[:, grid_mask, :] + pixel_coords = pixel_coords[:, :, grid_mask, ...] - fractional_coords = pixel_coords.to(torch.float32) - fractional_coords[:, 0] = fractional_coords[:, 0] * (1.0 / frame_rate) + kf_grid_mask = grid_mask[-keyframe_idxs.shape[2]:] + keyframe_idxs = keyframe_idxs[..., kf_grid_mask, :] + pixel_coords[:, :, -keyframe_idxs.shape[2]:, :] = keyframe_idxs x = self.patchify_proj(x) - timestep = timestep * 1000.0 - - if attention_mask is not None and not torch.is_floating_point(attention_mask): - attention_mask = (attention_mask - 1).to(x.dtype).reshape((attention_mask.shape[0], 1, -1, attention_mask.shape[-1])) * torch.finfo(x.dtype).max - - pe = precompute_freqs_cis(fractional_coords, dim=self.inner_dim, out_dtype=x.dtype) - - batch_size = x.shape[0] - timestep, embedded_timestep = self.adaln_single( - timestep.flatten(), - {"resolution": None, "aspect_ratio": None}, - batch_size=batch_size, - hidden_dtype=x.dtype, - ) - # Second dimension is 1 or number of tokens (if timestep_per_token) - timestep = timestep.view(batch_size, -1, timestep.shape[-1]) - embedded_timestep = embedded_timestep.view( - batch_size, -1, embedded_timestep.shape[-1] - ) - - # 2. Blocks - if self.caption_projection is not None: - batch_size = x.shape[0] - context = self.caption_projection(context) - context = context.view( - batch_size, -1, x.shape[-1] - ) + return x, pixel_coords, additional_args + def _process_transformer_blocks(self, x, context, attention_mask, timestep, pe, transformer_options={}, **kwargs): + """Process transformer blocks for LTXV.""" + patches_replace = transformer_options.get("patches_replace", {}) blocks_replace = patches_replace.get("dit", {}) + for i, block in enumerate(self.transformer_blocks): if ("double_block", i) in blocks_replace: + def block_wrap(args): out = {} out["img"] = block(args["img"], context=args["txt"], attention_mask=args["attention_mask"], timestep=args["vec"], pe=args["pe"], transformer_options=args["transformer_options"]) @@ -478,16 +921,28 @@ class LTXVModel(torch.nn.Module): transformer_options=transformer_options, ) - # 3. Output + return x + + def _process_output(self, x, embedded_timestep, keyframe_idxs, **kwargs): + """Process output for LTXV.""" + # Apply scale-shift modulation scale_shift_values = ( self.scale_shift_table[None, None].to(device=x.device, dtype=x.dtype) + embedded_timestep[:, :, None] ) shift, scale = scale_shift_values[:, :, 0], scale_shift_values[:, :, 1] + x = self.norm_out(x) - # Modulation - x = torch.addcmul(x, x, scale).add_(shift) + x = x * (1 + scale) + shift x = self.proj_out(x) + if keyframe_idxs is not None: + grid_mask = kwargs["grid_mask"] + orig_patchified_shape = kwargs["orig_patchified_shape"] + full_x = torch.zeros(orig_patchified_shape, dtype=x.dtype, device=x.device) + full_x[:, grid_mask, :] = x + x = full_x + # Unpatchify to restore original dimensions + orig_shape = kwargs["orig_shape"] x = self.patchifier.unpatchify( latents=x, output_height=orig_shape[3], diff --git a/comfy/ldm/lightricks/symmetric_patchifier.py b/comfy/ldm/lightricks/symmetric_patchifier.py index 4b9972b9f..8f9a41186 100644 --- a/comfy/ldm/lightricks/symmetric_patchifier.py +++ b/comfy/ldm/lightricks/symmetric_patchifier.py @@ -21,20 +21,23 @@ def latent_to_pixel_coords( Returns: Tensor: A tensor of pixel coordinates corresponding to the input latent coordinates. """ + shape = [1] * latent_coords.ndim + shape[1] = -1 pixel_coords = ( latent_coords - * torch.tensor(scale_factors, device=latent_coords.device)[None, :, None] + * torch.tensor(scale_factors, device=latent_coords.device).view(*shape) ) if causal_fix: # Fix temporal scale for first frame to 1 due to causality - pixel_coords[:, 0] = (pixel_coords[:, 0] + 1 - scale_factors[0]).clamp(min=0) + pixel_coords[:, 0, ...] = (pixel_coords[:, 0, ...] + 1 - scale_factors[0]).clamp(min=0) return pixel_coords class Patchifier(ABC): - def __init__(self, patch_size: int): + def __init__(self, patch_size: int, start_end: bool=False): super().__init__() self._patch_size = (1, patch_size, patch_size) + self.start_end = start_end @abstractmethod def patchify( @@ -71,11 +74,23 @@ class Patchifier(ABC): torch.arange(0, latent_width, self._patch_size[2], device=device), indexing="ij", ) - latent_sample_coords = torch.stack(latent_sample_coords, dim=0) - latent_coords = latent_sample_coords.unsqueeze(0).repeat(batch_size, 1, 1, 1, 1) - latent_coords = rearrange( - latent_coords, "b c f h w -> b c (f h w)", b=batch_size + latent_sample_coords_start = torch.stack(latent_sample_coords, dim=0) + delta = torch.tensor(self._patch_size, device=latent_sample_coords_start.device, dtype=latent_sample_coords_start.dtype)[:, None, None, None] + latent_sample_coords_end = latent_sample_coords_start + delta + + latent_sample_coords_start = latent_sample_coords_start.unsqueeze(0).repeat(batch_size, 1, 1, 1, 1) + latent_sample_coords_start = rearrange( + latent_sample_coords_start, "b c f h w -> b c (f h w)", b=batch_size ) + if self.start_end: + latent_sample_coords_end = latent_sample_coords_end.unsqueeze(0).repeat(batch_size, 1, 1, 1, 1) + latent_sample_coords_end = rearrange( + latent_sample_coords_end, "b c f h w -> b c (f h w)", b=batch_size + ) + + latent_coords = torch.stack((latent_sample_coords_start, latent_sample_coords_end), dim=-1) + else: + latent_coords = latent_sample_coords_start return latent_coords @@ -115,3 +130,61 @@ class SymmetricPatchifier(Patchifier): q=self._patch_size[2], ) return latents + + +class AudioPatchifier(Patchifier): + def __init__(self, patch_size: int, + sample_rate=16000, + hop_length=160, + audio_latent_downsample_factor=4, + is_causal=True, + start_end=False, + shift = 0 + ): + super().__init__(patch_size, start_end=start_end) + self.hop_length = hop_length + self.sample_rate = sample_rate + self.audio_latent_downsample_factor = audio_latent_downsample_factor + self.is_causal = is_causal + self.shift = shift + + def copy_with_shift(self, shift): + return AudioPatchifier( + self.patch_size, self.sample_rate, self.hop_length, self.audio_latent_downsample_factor, + self.is_causal, self.start_end, shift + ) + + def _get_audio_latent_time_in_sec(self, start_latent, end_latent: int, dtype: torch.dtype, device=torch.device): + audio_latent_frame = torch.arange(start_latent, end_latent, dtype=dtype, device=device) + audio_mel_frame = audio_latent_frame * self.audio_latent_downsample_factor + if self.is_causal: + audio_mel_frame = (audio_mel_frame + 1 - self.audio_latent_downsample_factor).clip(min=0) + return audio_mel_frame * self.hop_length / self.sample_rate + + + def patchify(self, audio_latents: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]: + # audio_latents: (batch, channels, time, freq) + b, _, t, _ = audio_latents.shape + audio_latents = rearrange( + audio_latents, + "b c t f -> b t (c f)", + ) + + audio_latents_start_timings = self._get_audio_latent_time_in_sec(self.shift, t + self.shift, torch.float32, audio_latents.device) + audio_latents_start_timings = audio_latents_start_timings.unsqueeze(0).expand(b, -1).unsqueeze(1) + + if self.start_end: + audio_latents_end_timings = self._get_audio_latent_time_in_sec(self.shift + 1, t + self.shift + 1, torch.float32, audio_latents.device) + audio_latents_end_timings = audio_latents_end_timings.unsqueeze(0).expand(b, -1).unsqueeze(1) + + audio_latents_timings = torch.stack([audio_latents_start_timings, audio_latents_end_timings], dim=-1) + else: + audio_latents_timings = audio_latents_start_timings + return audio_latents, audio_latents_timings + + def unpatchify(self, audio_latents: torch.Tensor, channels: int, freq: int) -> torch.Tensor: + # audio_latents: (batch, time, freq * channels) + audio_latents = rearrange( + audio_latents, "b t (c f) -> b c t f", c=channels, f=freq + ) + return audio_latents diff --git a/comfy/ldm/lightricks/vae/audio_vae.py b/comfy/ldm/lightricks/vae/audio_vae.py new file mode 100644 index 000000000..a9111d3bd --- /dev/null +++ b/comfy/ldm/lightricks/vae/audio_vae.py @@ -0,0 +1,286 @@ +import json +from dataclasses import dataclass +import math +import torch +import torchaudio + +import comfy.model_management +import comfy.model_patcher +import comfy.utils as utils +from comfy.ldm.mmaudio.vae.distributions import DiagonalGaussianDistribution +from comfy.ldm.lightricks.symmetric_patchifier import AudioPatchifier +from comfy.ldm.lightricks.vae.causal_audio_autoencoder import ( + CausalityAxis, + CausalAudioAutoencoder, +) +from comfy.ldm.lightricks.vocoders.vocoder import Vocoder + +LATENT_DOWNSAMPLE_FACTOR = 4 + + +@dataclass(frozen=True) +class AudioVAEComponentConfig: + """Container for model component configuration extracted from metadata.""" + + autoencoder: dict + vocoder: dict + + @classmethod + def from_metadata(cls, metadata: dict) -> "AudioVAEComponentConfig": + assert metadata is not None and "config" in metadata, "Metadata is required for audio VAE" + + raw_config = metadata["config"] + if isinstance(raw_config, str): + parsed_config = json.loads(raw_config) + else: + parsed_config = raw_config + + audio_config = parsed_config.get("audio_vae") + vocoder_config = parsed_config.get("vocoder") + + assert audio_config is not None, "Audio VAE config is required for audio VAE" + assert vocoder_config is not None, "Vocoder config is required for audio VAE" + + return cls(autoencoder=audio_config, vocoder=vocoder_config) + + +class ModelDeviceManager: + """Manages device placement and GPU residency for the composed model.""" + + def __init__(self, module: torch.nn.Module): + load_device = comfy.model_management.get_torch_device() + offload_device = comfy.model_management.vae_offload_device() + self.patcher = comfy.model_patcher.ModelPatcher(module, load_device, offload_device) + + def ensure_model_loaded(self) -> None: + comfy.model_management.free_memory( + self.patcher.model_size(), + self.patcher.load_device, + ) + comfy.model_management.load_model_gpu(self.patcher) + + def move_to_load_device(self, tensor: torch.Tensor) -> torch.Tensor: + return tensor.to(self.patcher.load_device) + + @property + def load_device(self): + return self.patcher.load_device + + +class AudioLatentNormalizer: + """Applies per-channel statistics in patch space and restores original layout.""" + + def __init__(self, patchfier: AudioPatchifier, statistics_processor: torch.nn.Module): + self.patchifier = patchfier + self.statistics = statistics_processor + + def normalize(self, latents: torch.Tensor) -> torch.Tensor: + channels = latents.shape[1] + freq = latents.shape[3] + patched, _ = self.patchifier.patchify(latents) + normalized = self.statistics.normalize(patched) + return self.patchifier.unpatchify(normalized, channels=channels, freq=freq) + + def denormalize(self, latents: torch.Tensor) -> torch.Tensor: + channels = latents.shape[1] + freq = latents.shape[3] + patched, _ = self.patchifier.patchify(latents) + denormalized = self.statistics.un_normalize(patched) + return self.patchifier.unpatchify(denormalized, channels=channels, freq=freq) + + +class AudioPreprocessor: + """Prepares raw waveforms for the autoencoder by matching training conditions.""" + + def __init__(self, target_sample_rate: int, mel_bins: int, mel_hop_length: int, n_fft: int): + self.target_sample_rate = target_sample_rate + self.mel_bins = mel_bins + self.mel_hop_length = mel_hop_length + self.n_fft = n_fft + + def resample(self, waveform: torch.Tensor, source_rate: int) -> torch.Tensor: + if source_rate == self.target_sample_rate: + return waveform + return torchaudio.functional.resample(waveform, source_rate, self.target_sample_rate) + + @staticmethod + def normalize_amplitude( + waveform: torch.Tensor, max_amplitude: float = 0.5, eps: float = 1e-5 + ) -> torch.Tensor: + waveform = waveform - waveform.mean(dim=2, keepdim=True) + peak = torch.max(torch.abs(waveform)) + eps + scale = peak.clamp(max=max_amplitude) / peak + return waveform * scale + + def waveform_to_mel( + self, waveform: torch.Tensor, waveform_sample_rate: int, device + ) -> torch.Tensor: + waveform = self.resample(waveform, waveform_sample_rate) + waveform = self.normalize_amplitude(waveform) + + mel_transform = torchaudio.transforms.MelSpectrogram( + sample_rate=self.target_sample_rate, + n_fft=self.n_fft, + win_length=self.n_fft, + hop_length=self.mel_hop_length, + f_min=0.0, + f_max=self.target_sample_rate / 2.0, + n_mels=self.mel_bins, + window_fn=torch.hann_window, + center=True, + pad_mode="reflect", + power=1.0, + mel_scale="slaney", + norm="slaney", + ).to(device) + + mel = mel_transform(waveform) + mel = torch.log(torch.clamp(mel, min=1e-5)) + return mel.permute(0, 1, 3, 2).contiguous() + + +class AudioVAE(torch.nn.Module): + """High-level Audio VAE wrapper exposing encode and decode entry points.""" + + def __init__(self, state_dict: dict, metadata: dict): + super().__init__() + + component_config = AudioVAEComponentConfig.from_metadata(metadata) + + vae_sd = utils.state_dict_prefix_replace(state_dict, {"audio_vae.": ""}, filter_keys=True) + vocoder_sd = utils.state_dict_prefix_replace(state_dict, {"vocoder.": ""}, filter_keys=True) + + self.autoencoder = CausalAudioAutoencoder(config=component_config.autoencoder) + self.vocoder = Vocoder(config=component_config.vocoder) + + self.autoencoder.load_state_dict(vae_sd, strict=False) + self.vocoder.load_state_dict(vocoder_sd, strict=False) + + autoencoder_config = self.autoencoder.get_config() + self.normalizer = AudioLatentNormalizer( + AudioPatchifier( + patch_size=1, + audio_latent_downsample_factor=LATENT_DOWNSAMPLE_FACTOR, + sample_rate=autoencoder_config["sampling_rate"], + hop_length=autoencoder_config["mel_hop_length"], + is_causal=autoencoder_config["is_causal"], + ), + self.autoencoder.per_channel_statistics, + ) + + self.preprocessor = AudioPreprocessor( + target_sample_rate=autoencoder_config["sampling_rate"], + mel_bins=autoencoder_config["mel_bins"], + mel_hop_length=autoencoder_config["mel_hop_length"], + n_fft=autoencoder_config["n_fft"], + ) + + self.device_manager = ModelDeviceManager(self) + + def encode(self, audio: dict) -> torch.Tensor: + """Encode a waveform dictionary into normalized latent tensors.""" + + waveform = audio["waveform"] + waveform_sample_rate = audio["sample_rate"] + input_device = waveform.device + # Ensure that Audio VAE is loaded on the correct device. + self.device_manager.ensure_model_loaded() + + waveform = self.device_manager.move_to_load_device(waveform) + expected_channels = self.autoencoder.encoder.in_channels + if waveform.shape[1] != expected_channels: + raise ValueError( + f"Input audio must have {expected_channels} channels, got {waveform.shape[1]}" + ) + + mel_spec = self.preprocessor.waveform_to_mel( + waveform, waveform_sample_rate, device=self.device_manager.load_device + ) + + latents = self.autoencoder.encode(mel_spec) + posterior = DiagonalGaussianDistribution(latents) + latent_mode = posterior.mode() + + normalized = self.normalizer.normalize(latent_mode) + return normalized.to(input_device) + + def decode(self, latents: torch.Tensor) -> torch.Tensor: + """Decode normalized latent tensors into an audio waveform.""" + original_shape = latents.shape + + # Ensure that Audio VAE is loaded on the correct device. + self.device_manager.ensure_model_loaded() + + latents = self.device_manager.move_to_load_device(latents) + latents = self.normalizer.denormalize(latents) + + target_shape = self.target_shape_from_latents(original_shape) + mel_spec = self.autoencoder.decode(latents, target_shape=target_shape) + + waveform = self.run_vocoder(mel_spec) + return self.device_manager.move_to_load_device(waveform) + + def target_shape_from_latents(self, latents_shape): + batch, _, time, _ = latents_shape + target_length = time * LATENT_DOWNSAMPLE_FACTOR + if self.autoencoder.causality_axis != CausalityAxis.NONE: + target_length -= LATENT_DOWNSAMPLE_FACTOR - 1 + return ( + batch, + self.autoencoder.decoder.out_ch, + target_length, + self.autoencoder.mel_bins, + ) + + def num_of_latents_from_frames(self, frames_number: int, frame_rate: int) -> int: + return math.ceil((float(frames_number) / frame_rate) * self.latents_per_second) + + def run_vocoder(self, mel_spec: torch.Tensor) -> torch.Tensor: + audio_channels = self.autoencoder.decoder.out_ch + vocoder_input = mel_spec.transpose(2, 3) + + if audio_channels == 1: + vocoder_input = vocoder_input.squeeze(1) + elif audio_channels != 2: + raise ValueError(f"Unsupported audio_channels: {audio_channels}") + + return self.vocoder(vocoder_input) + + @property + def sample_rate(self) -> int: + return int(self.autoencoder.sampling_rate) + + @property + def mel_hop_length(self) -> int: + return int(self.autoencoder.mel_hop_length) + + @property + def mel_bins(self) -> int: + return int(self.autoencoder.mel_bins) + + @property + def latent_channels(self) -> int: + return int(self.autoencoder.decoder.z_channels) + + @property + def latent_frequency_bins(self) -> int: + return int(self.mel_bins // LATENT_DOWNSAMPLE_FACTOR) + + @property + def latents_per_second(self) -> float: + return self.sample_rate / self.mel_hop_length / LATENT_DOWNSAMPLE_FACTOR + + @property + def output_sample_rate(self) -> int: + output_rate = getattr(self.vocoder, "output_sample_rate", None) + if output_rate is not None: + return int(output_rate) + upsample_factor = getattr(self.vocoder, "upsample_factor", None) + if upsample_factor is None: + raise AttributeError( + "Vocoder is missing upsample_factor; cannot infer output sample rate" + ) + return int(self.sample_rate * upsample_factor / self.mel_hop_length) + + def memory_required(self, input_shape): + return self.device_manager.patcher.model_size() diff --git a/comfy/ldm/lightricks/vae/causal_audio_autoencoder.py b/comfy/ldm/lightricks/vae/causal_audio_autoencoder.py new file mode 100644 index 000000000..f12b9bb53 --- /dev/null +++ b/comfy/ldm/lightricks/vae/causal_audio_autoencoder.py @@ -0,0 +1,909 @@ +from __future__ import annotations +import torch +from torch import nn +from torch.nn import functional as F +from typing import Optional +from enum import Enum +from .pixel_norm import PixelNorm +import comfy.ops +import logging + +ops = comfy.ops.disable_weight_init + + +class StringConvertibleEnum(Enum): + """ + Base enum class that provides string-to-enum conversion functionality. + + This mixin adds a str_to_enum() class method that handles conversion from + strings, None, or existing enum instances with case-insensitive matching. + """ + + @classmethod + def str_to_enum(cls, value): + """ + Convert a string, enum instance, or None to the appropriate enum member. + + Args: + value: Can be an enum instance of this class, a string, or None + + Returns: + Enum member of this class + + Raises: + ValueError: If the value cannot be converted to a valid enum member + """ + # Already an enum instance of this class + if isinstance(value, cls): + return value + + # None maps to NONE member if it exists + if value is None: + if hasattr(cls, "NONE"): + return cls.NONE + raise ValueError(f"{cls.__name__} does not have a NONE member to map None to") + + # String conversion (case-insensitive) + if isinstance(value, str): + value_lower = value.lower() + + # Try to match against enum values + for member in cls: + # Handle members with None values + if member.value is None: + if value_lower == "none": + return member + # Handle members with string values + elif isinstance(member.value, str) and member.value.lower() == value_lower: + return member + + # Build helpful error message with valid values + valid_values = [] + for member in cls: + if member.value is None: + valid_values.append("none") + elif isinstance(member.value, str): + valid_values.append(member.value) + + raise ValueError(f"Invalid {cls.__name__} string: '{value}'. " f"Valid values are: {valid_values}") + + raise ValueError( + f"Cannot convert type {type(value).__name__} to {cls.__name__} enum. " + f"Expected string, None, or {cls.__name__} instance." + ) + + +class AttentionType(StringConvertibleEnum): + """Enum for specifying the attention mechanism type.""" + + VANILLA = "vanilla" + LINEAR = "linear" + NONE = "none" + + +class CausalityAxis(StringConvertibleEnum): + """Enum for specifying the causality axis in causal convolutions.""" + + NONE = None + WIDTH = "width" + HEIGHT = "height" + WIDTH_COMPATIBILITY = "width-compatibility" + + +def Normalize(in_channels, *, num_groups=32, normtype="group"): + if normtype == "group": + return torch.nn.GroupNorm(num_groups=num_groups, num_channels=in_channels, eps=1e-6, affine=True) + elif normtype == "pixel": + return PixelNorm(dim=1, eps=1e-6) + else: + raise ValueError(f"Invalid normalization type: {normtype}") + + +class CausalConv2d(nn.Module): + """ + A causal 2D convolution. + + This layer ensures that the output at time `t` only depends on inputs + at time `t` and earlier. It achieves this by applying asymmetric padding + to the time dimension (width) before the convolution. + """ + + def __init__( + self, + in_channels, + out_channels, + kernel_size, + stride=1, + dilation=1, + groups=1, + bias=True, + causality_axis: CausalityAxis = CausalityAxis.HEIGHT, + ): + super().__init__() + + self.causality_axis = causality_axis + + # Ensure kernel_size and dilation are tuples + kernel_size = nn.modules.utils._pair(kernel_size) + dilation = nn.modules.utils._pair(dilation) + + # Calculate padding dimensions + pad_h = (kernel_size[0] - 1) * dilation[0] + pad_w = (kernel_size[1] - 1) * dilation[1] + + # The padding tuple for F.pad is (pad_left, pad_right, pad_top, pad_bottom) + match self.causality_axis: + case CausalityAxis.NONE: + self.padding = (pad_w // 2, pad_w - pad_w // 2, pad_h // 2, pad_h - pad_h // 2) + case CausalityAxis.WIDTH | CausalityAxis.WIDTH_COMPATIBILITY: + self.padding = (pad_w, 0, pad_h // 2, pad_h - pad_h // 2) + case CausalityAxis.HEIGHT: + self.padding = (pad_w // 2, pad_w - pad_w // 2, pad_h, 0) + case _: + raise ValueError(f"Invalid causality_axis: {causality_axis}") + + # The internal convolution layer uses no padding, as we handle it manually + self.conv = ops.Conv2d( + in_channels, + out_channels, + kernel_size, + stride=stride, + padding=0, + dilation=dilation, + groups=groups, + bias=bias, + ) + + def forward(self, x): + # Apply causal padding before convolution + x = F.pad(x, self.padding) + return self.conv(x) + + +def make_conv2d( + in_channels, + out_channels, + kernel_size, + stride=1, + padding=None, + dilation=1, + groups=1, + bias=True, + causality_axis: Optional[CausalityAxis] = None, +): + """ + Create a 2D convolution layer that can be either causal or non-causal. + + Args: + in_channels: Number of input channels + out_channels: Number of output channels + kernel_size: Size of the convolution kernel + stride: Convolution stride + padding: Padding (if None, will be calculated based on causal flag) + dilation: Dilation rate + groups: Number of groups for grouped convolution + bias: Whether to use bias + causality_axis: Dimension along which to apply causality. + + Returns: + Either a regular Conv2d or CausalConv2d layer + """ + if causality_axis is not None: + # For causal convolution, padding is handled internally by CausalConv2d + return CausalConv2d(in_channels, out_channels, kernel_size, stride, dilation, groups, bias, causality_axis) + else: + # For non-causal convolution, use symmetric padding if not specified + if padding is None: + if isinstance(kernel_size, int): + padding = kernel_size // 2 + else: + padding = tuple(k // 2 for k in kernel_size) + return ops.Conv2d( + in_channels, + out_channels, + kernel_size, + stride, + padding, + dilation, + groups, + bias, + ) + + +class Upsample(nn.Module): + def __init__(self, in_channels, with_conv, causality_axis: CausalityAxis = CausalityAxis.HEIGHT): + super().__init__() + self.with_conv = with_conv + self.causality_axis = causality_axis + if self.with_conv: + self.conv = make_conv2d(in_channels, in_channels, kernel_size=3, stride=1, causality_axis=causality_axis) + + def forward(self, x): + x = torch.nn.functional.interpolate(x, scale_factor=2.0, mode="nearest") + if self.with_conv: + x = self.conv(x) + # Drop FIRST element in the causal axis to undo encoder's padding, while keeping the length 1 + 2 * n. + # For example, if the input is [0, 1, 2], after interpolation, the output is [0, 0, 1, 1, 2, 2]. + # The causal convolution will pad the first element as [-, -, 0, 0, 1, 1, 2, 2], + # So the output elements rely on the following windows: + # 0: [-,-,0] + # 1: [-,0,0] + # 2: [0,0,1] + # 3: [0,1,1] + # 4: [1,1,2] + # 5: [1,2,2] + # Notice that the first and second elements in the output rely only on the first element in the input, + # while all other elements rely on two elements in the input. + # So we can drop the first element to undo the padding (rather than the last element). + # This is a no-op for non-causal convolutions. + match self.causality_axis: + case CausalityAxis.NONE: + pass # x remains unchanged + case CausalityAxis.HEIGHT: + x = x[:, :, 1:, :] + case CausalityAxis.WIDTH: + x = x[:, :, :, 1:] + case CausalityAxis.WIDTH_COMPATIBILITY: + pass # x remains unchanged + case _: + raise ValueError(f"Invalid causality_axis: {self.causality_axis}") + + return x + + +class Downsample(nn.Module): + """ + A downsampling layer that can use either a strided convolution + or average pooling. Supports standard and causal padding for the + convolutional mode. + """ + + def __init__(self, in_channels, with_conv, causality_axis: CausalityAxis = CausalityAxis.WIDTH): + super().__init__() + self.with_conv = with_conv + self.causality_axis = causality_axis + + if self.causality_axis != CausalityAxis.NONE and not self.with_conv: + raise ValueError("causality is only supported when `with_conv=True`.") + + if self.with_conv: + # Do time downsampling here + # no asymmetric padding in torch conv, must do it ourselves + self.conv = ops.Conv2d(in_channels, in_channels, kernel_size=3, stride=2, padding=0) + + def forward(self, x): + if self.with_conv: + # (pad_left, pad_right, pad_top, pad_bottom) + match self.causality_axis: + case CausalityAxis.NONE: + pad = (0, 1, 0, 1) + case CausalityAxis.WIDTH: + pad = (2, 0, 0, 1) + case CausalityAxis.HEIGHT: + pad = (0, 1, 2, 0) + case CausalityAxis.WIDTH_COMPATIBILITY: + pad = (1, 0, 0, 1) + case _: + raise ValueError(f"Invalid causality_axis: {self.causality_axis}") + + x = torch.nn.functional.pad(x, pad, mode="constant", value=0) + x = self.conv(x) + else: + # This branch is only taken if with_conv=False, which implies causality_axis is NONE. + x = torch.nn.functional.avg_pool2d(x, kernel_size=2, stride=2) + + return x + + +class ResnetBlock(nn.Module): + def __init__( + self, + *, + in_channels, + out_channels=None, + conv_shortcut=False, + dropout, + temb_channels=512, + norm_type="group", + causality_axis: CausalityAxis = CausalityAxis.HEIGHT, + ): + super().__init__() + self.causality_axis = causality_axis + + if self.causality_axis != CausalityAxis.NONE and norm_type == "group": + raise ValueError("Causal ResnetBlock with GroupNorm is not supported.") + self.in_channels = in_channels + out_channels = in_channels if out_channels is None else out_channels + self.out_channels = out_channels + self.use_conv_shortcut = conv_shortcut + + self.norm1 = Normalize(in_channels, normtype=norm_type) + self.non_linearity = nn.SiLU() + self.conv1 = make_conv2d(in_channels, out_channels, kernel_size=3, stride=1, causality_axis=causality_axis) + if temb_channels > 0: + self.temb_proj = ops.Linear(temb_channels, out_channels) + self.norm2 = Normalize(out_channels, normtype=norm_type) + self.dropout = torch.nn.Dropout(dropout) + self.conv2 = make_conv2d(out_channels, out_channels, kernel_size=3, stride=1, causality_axis=causality_axis) + if self.in_channels != self.out_channels: + if self.use_conv_shortcut: + self.conv_shortcut = make_conv2d( + in_channels, out_channels, kernel_size=3, stride=1, causality_axis=causality_axis + ) + else: + self.nin_shortcut = make_conv2d( + in_channels, out_channels, kernel_size=1, stride=1, causality_axis=causality_axis + ) + + def forward(self, x, temb): + h = x + h = self.norm1(h) + h = self.non_linearity(h) + h = self.conv1(h) + + if temb is not None: + h = h + self.temb_proj(self.non_linearity(temb))[:, :, None, None] + + h = self.norm2(h) + h = self.non_linearity(h) + h = self.dropout(h) + h = self.conv2(h) + + if self.in_channels != self.out_channels: + if self.use_conv_shortcut: + x = self.conv_shortcut(x) + else: + x = self.nin_shortcut(x) + + return x + h + + +class AttnBlock(nn.Module): + def __init__(self, in_channels, norm_type="group"): + super().__init__() + self.in_channels = in_channels + + self.norm = Normalize(in_channels, normtype=norm_type) + self.q = ops.Conv2d(in_channels, in_channels, kernel_size=1, stride=1, padding=0) + self.k = ops.Conv2d(in_channels, in_channels, kernel_size=1, stride=1, padding=0) + self.v = ops.Conv2d(in_channels, in_channels, kernel_size=1, stride=1, padding=0) + self.proj_out = ops.Conv2d(in_channels, in_channels, kernel_size=1, stride=1, padding=0) + + def forward(self, x): + h_ = x + h_ = self.norm(h_) + q = self.q(h_) + k = self.k(h_) + v = self.v(h_) + + # compute attention + b, c, h, w = q.shape + q = q.reshape(b, c, h * w).contiguous() + q = q.permute(0, 2, 1).contiguous() # b,hw,c + k = k.reshape(b, c, h * w).contiguous() # b,c,hw + w_ = torch.bmm(q, k).contiguous() # b,hw,hw w[b,i,j]=sum_c q[b,i,c]k[b,c,j] + w_ = w_ * (int(c) ** (-0.5)) + w_ = torch.nn.functional.softmax(w_, dim=2) + + # attend to values + v = v.reshape(b, c, h * w).contiguous() + w_ = w_.permute(0, 2, 1).contiguous() # b,hw,hw (first hw of k, second of q) + h_ = torch.bmm(v, w_).contiguous() # b, c,hw (hw of q) h_[b,c,j] = sum_i v[b,c,i] w_[b,i,j] + h_ = h_.reshape(b, c, h, w).contiguous() + + h_ = self.proj_out(h_) + + return x + h_ + + +def make_attn(in_channels, attn_type="vanilla", norm_type="group"): + # Convert string to enum if needed + attn_type = AttentionType.str_to_enum(attn_type) + + if attn_type != AttentionType.NONE: + logging.info(f"making attention of type '{attn_type.value}' with {in_channels} in_channels") + else: + logging.info(f"making identity attention with {in_channels} in_channels") + + match attn_type: + case AttentionType.VANILLA: + return AttnBlock(in_channels, norm_type=norm_type) + case AttentionType.NONE: + return nn.Identity(in_channels) + case AttentionType.LINEAR: + raise NotImplementedError(f"Attention type {attn_type.value} is not supported yet.") + case _: + raise ValueError(f"Unknown attention type: {attn_type}") + + +class Encoder(nn.Module): + def __init__( + self, + *, + ch, + out_ch, + ch_mult=(1, 2, 4, 8), + num_res_blocks, + attn_resolutions, + dropout=0.0, + resamp_with_conv=True, + in_channels, + resolution, + z_channels, + double_z=True, + attn_type="vanilla", + mid_block_add_attention=True, + norm_type="group", + causality_axis=CausalityAxis.WIDTH.value, + **ignore_kwargs, + ): + super().__init__() + self.ch = ch + self.temb_ch = 0 + self.num_resolutions = len(ch_mult) + self.num_res_blocks = num_res_blocks + self.resolution = resolution + self.in_channels = in_channels + self.z_channels = z_channels + self.double_z = double_z + self.norm_type = norm_type + # Convert string to enum if needed (for config loading) + causality_axis = CausalityAxis.str_to_enum(causality_axis) + self.attn_type = AttentionType.str_to_enum(attn_type) + + # downsampling + self.conv_in = make_conv2d( + in_channels, + self.ch, + kernel_size=3, + stride=1, + causality_axis=causality_axis, + ) + + self.non_linearity = nn.SiLU() + + curr_res = resolution + in_ch_mult = (1,) + tuple(ch_mult) + self.in_ch_mult = in_ch_mult + self.down = nn.ModuleList() + + for i_level in range(self.num_resolutions): + block = nn.ModuleList() + attn = nn.ModuleList() + block_in = ch * in_ch_mult[i_level] + block_out = ch * ch_mult[i_level] + + for _ in range(self.num_res_blocks): + block.append( + ResnetBlock( + in_channels=block_in, + out_channels=block_out, + temb_channels=self.temb_ch, + dropout=dropout, + norm_type=self.norm_type, + causality_axis=causality_axis, + ) + ) + block_in = block_out + if curr_res in attn_resolutions: + attn.append(make_attn(block_in, attn_type=self.attn_type, norm_type=self.norm_type)) + + down = nn.Module() + down.block = block + down.attn = attn + if i_level != self.num_resolutions - 1: + down.downsample = Downsample(block_in, resamp_with_conv, causality_axis=causality_axis) + curr_res = curr_res // 2 + self.down.append(down) + + # middle + self.mid = nn.Module() + self.mid.block_1 = ResnetBlock( + in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout, + norm_type=self.norm_type, + causality_axis=causality_axis, + ) + if mid_block_add_attention: + self.mid.attn_1 = make_attn(block_in, attn_type=self.attn_type, norm_type=self.norm_type) + else: + self.mid.attn_1 = nn.Identity() + self.mid.block_2 = ResnetBlock( + in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout, + norm_type=self.norm_type, + causality_axis=causality_axis, + ) + + # end + self.norm_out = Normalize(block_in, normtype=self.norm_type) + self.conv_out = make_conv2d( + block_in, + 2 * z_channels if double_z else z_channels, + kernel_size=3, + stride=1, + causality_axis=causality_axis, + ) + + def forward(self, x): + """ + Forward pass through the encoder. + + Args: + x: Input tensor of shape [batch, channels, time, n_mels] + + Returns: + Encoded latent representation + """ + feature_maps = [self.conv_in(x)] + + # Process each resolution level (from high to low resolution) + for resolution_level in range(self.num_resolutions): + # Apply residual blocks at current resolution level + for block_idx in range(self.num_res_blocks): + # Apply ResNet block with optional timestep embedding + current_features = self.down[resolution_level].block[block_idx](feature_maps[-1], temb=None) + + # Apply attention if configured for this resolution level + if len(self.down[resolution_level].attn) > 0: + current_features = self.down[resolution_level].attn[block_idx](current_features) + + # Store processed features + feature_maps.append(current_features) + + # Downsample spatial dimensions (except at the final resolution level) + if resolution_level != self.num_resolutions - 1: + downsampled_features = self.down[resolution_level].downsample(feature_maps[-1]) + feature_maps.append(downsampled_features) + + # === MIDDLE PROCESSING PHASE === + # Take the lowest resolution features for middle processing + bottleneck_features = feature_maps[-1] + + # Apply first middle ResNet block + bottleneck_features = self.mid.block_1(bottleneck_features, temb=None) + + # Apply middle attention block + bottleneck_features = self.mid.attn_1(bottleneck_features) + + # Apply second middle ResNet block + bottleneck_features = self.mid.block_2(bottleneck_features, temb=None) + + # === OUTPUT PHASE === + # Normalize the bottleneck features + output_features = self.norm_out(bottleneck_features) + + # Apply non-linearity (SiLU activation) + output_features = self.non_linearity(output_features) + + # Final convolution to produce latent representation + # [batch, channels, time, n_mels] -> [batch, 2 * z_channels if double_z else z_channels, time, n_mels] + return self.conv_out(output_features) + + +class Decoder(nn.Module): + def __init__( + self, + *, + ch, + out_ch, + ch_mult=(1, 2, 4, 8), + num_res_blocks, + attn_resolutions, + dropout=0.0, + resamp_with_conv=True, + in_channels, + resolution, + z_channels, + give_pre_end=False, + tanh_out=False, + attn_type="vanilla", + mid_block_add_attention=True, + norm_type="group", + causality_axis=CausalityAxis.WIDTH.value, + **ignorekwargs, + ): + super().__init__() + self.ch = ch + self.temb_ch = 0 + self.num_resolutions = len(ch_mult) + self.num_res_blocks = num_res_blocks + self.resolution = resolution + self.in_channels = in_channels + self.out_ch = out_ch + self.give_pre_end = give_pre_end + self.tanh_out = tanh_out + self.norm_type = norm_type + self.z_channels = z_channels + # Convert string to enum if needed (for config loading) + causality_axis = CausalityAxis.str_to_enum(causality_axis) + self.attn_type = AttentionType.str_to_enum(attn_type) + + # compute block_in and curr_res at lowest res + block_in = ch * ch_mult[self.num_resolutions - 1] + curr_res = resolution // 2 ** (self.num_resolutions - 1) + self.z_shape = (1, z_channels, curr_res, curr_res) + + # z to block_in + self.conv_in = make_conv2d(z_channels, block_in, kernel_size=3, stride=1, causality_axis=causality_axis) + + self.non_linearity = nn.SiLU() + + # middle + self.mid = nn.Module() + self.mid.block_1 = ResnetBlock( + in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout, + norm_type=self.norm_type, + causality_axis=causality_axis, + ) + if mid_block_add_attention: + self.mid.attn_1 = make_attn(block_in, attn_type=self.attn_type, norm_type=self.norm_type) + else: + self.mid.attn_1 = nn.Identity() + self.mid.block_2 = ResnetBlock( + in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout, + norm_type=self.norm_type, + causality_axis=causality_axis, + ) + + # upsampling + self.up = nn.ModuleList() + for i_level in reversed(range(self.num_resolutions)): + block = nn.ModuleList() + attn = nn.ModuleList() + block_out = ch * ch_mult[i_level] + for _ in range(self.num_res_blocks + 1): + block.append( + ResnetBlock( + in_channels=block_in, + out_channels=block_out, + temb_channels=self.temb_ch, + dropout=dropout, + norm_type=self.norm_type, + causality_axis=causality_axis, + ) + ) + block_in = block_out + if curr_res in attn_resolutions: + attn.append(make_attn(block_in, attn_type=self.attn_type, norm_type=self.norm_type)) + up = nn.Module() + up.block = block + up.attn = attn + if i_level != 0: + up.upsample = Upsample(block_in, resamp_with_conv, causality_axis=causality_axis) + curr_res = curr_res * 2 + self.up.insert(0, up) # prepend to get consistent order + + # end + self.norm_out = Normalize(block_in, normtype=self.norm_type) + self.conv_out = make_conv2d(block_in, out_ch, kernel_size=3, stride=1, causality_axis=causality_axis) + + def _adjust_output_shape(self, decoded_output, target_shape): + """ + Adjust output shape to match target dimensions for variable-length audio. + + This function handles the common case where decoded audio spectrograms need to be + resized to match a specific target shape. + + Args: + decoded_output: Tensor of shape (batch, channels, time, frequency) + target_shape: Target shape tuple (batch, channels, time, frequency) + + Returns: + Tensor adjusted to match target_shape exactly + """ + # Current output shape: (batch, channels, time, frequency) + _, _, current_time, current_freq = decoded_output.shape + _, target_channels, target_time, target_freq = target_shape + + # Step 1: Crop first to avoid exceeding target dimensions + decoded_output = decoded_output[ + :, :target_channels, : min(current_time, target_time), : min(current_freq, target_freq) + ] + + # Step 2: Calculate padding needed for time and frequency dimensions + time_padding_needed = target_time - decoded_output.shape[2] + freq_padding_needed = target_freq - decoded_output.shape[3] + + # Step 3: Apply padding if needed + if time_padding_needed > 0 or freq_padding_needed > 0: + # PyTorch padding format: (pad_left, pad_right, pad_top, pad_bottom) + # For audio: pad_left/right = frequency, pad_top/bottom = time + padding = ( + 0, + max(freq_padding_needed, 0), # frequency padding (left, right) + 0, + max(time_padding_needed, 0), # time padding (top, bottom) + ) + decoded_output = F.pad(decoded_output, padding) + + # Step 4: Final safety crop to ensure exact target shape + decoded_output = decoded_output[:, :target_channels, :target_time, :target_freq] + + return decoded_output + + def get_config(self): + return { + "ch": self.ch, + "out_ch": self.out_ch, + "ch_mult": self.ch_mult, + "num_res_blocks": self.num_res_blocks, + "in_channels": self.in_channels, + "resolution": self.resolution, + "z_channels": self.z_channels, + } + + def forward(self, latent_features, target_shape=None): + """ + Decode latent features back to audio spectrograms. + + Args: + latent_features: Encoded latent representation of shape (batch, channels, height, width) + target_shape: Optional target output shape (batch, channels, time, frequency) + If provided, output will be cropped/padded to match this shape + + Returns: + Reconstructed audio spectrogram of shape (batch, channels, time, frequency) + """ + assert target_shape is not None, "Target shape is required for CausalAudioAutoencoder Decoder" + + # Transform latent features to decoder's internal feature dimension + hidden_features = self.conv_in(latent_features) + + # Middle processing + hidden_features = self.mid.block_1(hidden_features, temb=None) + hidden_features = self.mid.attn_1(hidden_features) + hidden_features = self.mid.block_2(hidden_features, temb=None) + + # Upsampling + # Progressively increase spatial resolution from lowest to highest + for resolution_level in reversed(range(self.num_resolutions)): + # Apply residual blocks at current resolution level + for block_index in range(self.num_res_blocks + 1): + hidden_features = self.up[resolution_level].block[block_index](hidden_features, temb=None) + + if len(self.up[resolution_level].attn) > 0: + hidden_features = self.up[resolution_level].attn[block_index](hidden_features) + + if resolution_level != 0: + hidden_features = self.up[resolution_level].upsample(hidden_features) + + # Output + if self.give_pre_end: + # Return intermediate features before final processing (for debugging/analysis) + decoded_output = hidden_features + else: + # Standard output path: normalize, activate, and convert to output channels + # Final normalization layer + hidden_features = self.norm_out(hidden_features) + + # Apply SiLU (Swish) activation function + hidden_features = self.non_linearity(hidden_features) + + # Final convolution to map to output channels (typically 2 for stereo audio) + decoded_output = self.conv_out(hidden_features) + + # Optional tanh activation to bound output values to [-1, 1] range + if self.tanh_out: + decoded_output = torch.tanh(decoded_output) + + # Adjust shape for audio data + if target_shape is not None: + decoded_output = self._adjust_output_shape(decoded_output, target_shape) + + return decoded_output + + +class processor(nn.Module): + def __init__(self): + super().__init__() + self.register_buffer("std-of-means", torch.empty(128)) + self.register_buffer("mean-of-means", torch.empty(128)) + + def un_normalize(self, x): + return (x * self.get_buffer("std-of-means").to(x)) + self.get_buffer("mean-of-means").to(x) + + def normalize(self, x): + return (x - self.get_buffer("mean-of-means").to(x)) / self.get_buffer("std-of-means").to(x) + + +class CausalAudioAutoencoder(nn.Module): + def __init__(self, config=None): + super().__init__() + + if config is None: + config = self._guess_config() + + # Extract encoder and decoder configs from the new format + model_config = config.get("model", {}).get("params", {}) + variables_config = config.get("variables", {}) + + self.sampling_rate = variables_config.get( + "sampling_rate", + model_config.get("sampling_rate", config.get("sampling_rate", 16000)), + ) + encoder_config = model_config.get("encoder", model_config.get("ddconfig", {})) + decoder_config = model_config.get("decoder", encoder_config) + + # Load mel spectrogram parameters + self.mel_bins = encoder_config.get("mel_bins", 64) + self.mel_hop_length = model_config.get("preprocessing", {}).get("stft", {}).get("hop_length", 160) + self.n_fft = model_config.get("preprocessing", {}).get("stft", {}).get("filter_length", 1024) + + # Store causality configuration at VAE level (not just in encoder internals) + causality_axis_value = encoder_config.get("causality_axis", CausalityAxis.WIDTH.value) + self.causality_axis = CausalityAxis.str_to_enum(causality_axis_value) + self.is_causal = self.causality_axis == CausalityAxis.HEIGHT + + self.encoder = Encoder(**encoder_config) + self.decoder = Decoder(**decoder_config) + + self.per_channel_statistics = processor() + + def _guess_config(self): + encoder_config = { + # Required parameters - based on ltx-video-av-1679000 model metadata + "ch": 128, + "out_ch": 8, + "ch_mult": [1, 2, 4], # Based on metadata: [1, 2, 4] not [1, 2, 4, 8] + "num_res_blocks": 2, + "attn_resolutions": [], # Based on metadata: empty list, no attention + "dropout": 0.0, + "resamp_with_conv": True, + "in_channels": 2, # stereo + "resolution": 256, + "z_channels": 8, + "double_z": True, + "attn_type": "vanilla", + "mid_block_add_attention": False, # Based on metadata: false + "norm_type": "pixel", + "causality_axis": "height", # Based on metadata + "mel_bins": 64, # Based on metadata: mel_bins = 64 + } + + decoder_config = { + # Inherits encoder config, can override specific params + **encoder_config, + "out_ch": 2, # Stereo audio output (2 channels) + "give_pre_end": False, + "tanh_out": False, + } + + config = { + "_class_name": "CausalAudioAutoencoder", + "sampling_rate": 16000, + "model": { + "params": { + "encoder": encoder_config, + "decoder": decoder_config, + } + }, + } + + return config + + def get_config(self): + return { + "sampling_rate": self.sampling_rate, + "mel_bins": self.mel_bins, + "mel_hop_length": self.mel_hop_length, + "n_fft": self.n_fft, + "causality_axis": self.causality_axis.value, + "is_causal": self.is_causal, + } + + def encode(self, x): + return self.encoder(x) + + def decode(self, x, target_shape=None): + return self.decoder(x, target_shape=target_shape) diff --git a/comfy/ldm/lightricks/vocoders/vocoder.py b/comfy/ldm/lightricks/vocoders/vocoder.py new file mode 100644 index 000000000..b1f15f2c5 --- /dev/null +++ b/comfy/ldm/lightricks/vocoders/vocoder.py @@ -0,0 +1,213 @@ +import torch +import torch.nn.functional as F +import torch.nn as nn +import comfy.ops +import numpy as np + +ops = comfy.ops.disable_weight_init + +LRELU_SLOPE = 0.1 + +def get_padding(kernel_size, dilation=1): + return int((kernel_size * dilation - dilation) / 2) + + +class ResBlock1(torch.nn.Module): + def __init__(self, channels, kernel_size=3, dilation=(1, 3, 5)): + super(ResBlock1, self).__init__() + self.convs1 = nn.ModuleList( + [ + ops.Conv1d( + channels, + channels, + kernel_size, + 1, + dilation=dilation[0], + padding=get_padding(kernel_size, dilation[0]), + ), + ops.Conv1d( + channels, + channels, + kernel_size, + 1, + dilation=dilation[1], + padding=get_padding(kernel_size, dilation[1]), + ), + ops.Conv1d( + channels, + channels, + kernel_size, + 1, + dilation=dilation[2], + padding=get_padding(kernel_size, dilation[2]), + ), + ] + ) + + self.convs2 = nn.ModuleList( + [ + ops.Conv1d( + channels, + channels, + kernel_size, + 1, + dilation=1, + padding=get_padding(kernel_size, 1), + ), + ops.Conv1d( + channels, + channels, + kernel_size, + 1, + dilation=1, + padding=get_padding(kernel_size, 1), + ), + ops.Conv1d( + channels, + channels, + kernel_size, + 1, + dilation=1, + padding=get_padding(kernel_size, 1), + ), + ] + ) + + def forward(self, x): + for c1, c2 in zip(self.convs1, self.convs2): + xt = F.leaky_relu(x, LRELU_SLOPE) + xt = c1(xt) + xt = F.leaky_relu(xt, LRELU_SLOPE) + xt = c2(xt) + x = xt + x + return x + + +class ResBlock2(torch.nn.Module): + def __init__(self, channels, kernel_size=3, dilation=(1, 3)): + super(ResBlock2, self).__init__() + self.convs = nn.ModuleList( + [ + ops.Conv1d( + channels, + channels, + kernel_size, + 1, + dilation=dilation[0], + padding=get_padding(kernel_size, dilation[0]), + ), + ops.Conv1d( + channels, + channels, + kernel_size, + 1, + dilation=dilation[1], + padding=get_padding(kernel_size, dilation[1]), + ), + ] + ) + + def forward(self, x): + for c in self.convs: + xt = F.leaky_relu(x, LRELU_SLOPE) + xt = c(xt) + x = xt + x + return x + + +class Vocoder(torch.nn.Module): + """ + Vocoder model for synthesizing audio from spectrograms, based on: https://github.com/jik876/hifi-gan. + + """ + + def __init__(self, config=None): + super(Vocoder, self).__init__() + + if config is None: + config = self.get_default_config() + + resblock_kernel_sizes = config.get("resblock_kernel_sizes", [3, 7, 11]) + upsample_rates = config.get("upsample_rates", [6, 5, 2, 2, 2]) + upsample_kernel_sizes = config.get("upsample_kernel_sizes", [16, 15, 8, 4, 4]) + resblock_dilation_sizes = config.get("resblock_dilation_sizes", [[1, 3, 5], [1, 3, 5], [1, 3, 5]]) + upsample_initial_channel = config.get("upsample_initial_channel", 1024) + stereo = config.get("stereo", True) + resblock = config.get("resblock", "1") + + self.output_sample_rate = config.get("output_sample_rate") + self.num_kernels = len(resblock_kernel_sizes) + self.num_upsamples = len(upsample_rates) + in_channels = 128 if stereo else 64 + self.conv_pre = ops.Conv1d(in_channels, upsample_initial_channel, 7, 1, padding=3) + resblock_class = ResBlock1 if resblock == "1" else ResBlock2 + + self.ups = nn.ModuleList() + for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)): + self.ups.append( + ops.ConvTranspose1d( + upsample_initial_channel // (2**i), + upsample_initial_channel // (2 ** (i + 1)), + k, + u, + padding=(k - u) // 2, + ) + ) + + self.resblocks = nn.ModuleList() + for i in range(len(self.ups)): + ch = upsample_initial_channel // (2 ** (i + 1)) + for _, (k, d) in enumerate(zip(resblock_kernel_sizes, resblock_dilation_sizes)): + self.resblocks.append(resblock_class(ch, k, d)) + + out_channels = 2 if stereo else 1 + self.conv_post = ops.Conv1d(ch, out_channels, 7, 1, padding=3) + + self.upsample_factor = np.prod([self.ups[i].stride[0] for i in range(len(self.ups))]) + + def get_default_config(self): + """Generate default configuration for the vocoder.""" + + config = { + "resblock_kernel_sizes": [3, 7, 11], + "upsample_rates": [6, 5, 2, 2, 2], + "upsample_kernel_sizes": [16, 15, 8, 4, 4], + "resblock_dilation_sizes": [[1, 3, 5], [1, 3, 5], [1, 3, 5]], + "upsample_initial_channel": 1024, + "stereo": True, + "resblock": "1", + } + + return config + + def forward(self, x): + """ + Forward pass of the vocoder. + + Args: + x: Input spectrogram tensor. Can be: + - 3D: (batch_size, channels, time_steps) for mono + - 4D: (batch_size, 2, channels, time_steps) for stereo + + Returns: + Audio tensor of shape (batch_size, out_channels, audio_length) + """ + if x.dim() == 4: # stereo + assert x.shape[1] == 2, "Input must have 2 channels for stereo" + x = torch.cat((x[:, 0, :, :], x[:, 1, :, :]), dim=1) + x = self.conv_pre(x) + for i in range(self.num_upsamples): + x = F.leaky_relu(x, LRELU_SLOPE) + x = self.ups[i](x) + xs = None + for j in range(self.num_kernels): + if xs is None: + xs = self.resblocks[i * self.num_kernels + j](x) + else: + xs += self.resblocks[i * self.num_kernels + j](x) + x = xs / self.num_kernels + x = F.leaky_relu(x) + x = self.conv_post(x) + x = torch.tanh(x) + + return x diff --git a/comfy/model_base.py b/comfy/model_base.py index c4f3c0639..49efd700b 100644 --- a/comfy/model_base.py +++ b/comfy/model_base.py @@ -20,6 +20,7 @@ import comfy.ldm.hunyuan3dv2_1 import comfy.ldm.hunyuan3dv2_1.hunyuandit import torch import logging +import comfy.ldm.lightricks.av_model from comfy.ldm.modules.diffusionmodules.openaimodel import UNetModel, Timestep from comfy.ldm.cascade.stage_c import StageC from comfy.ldm.cascade.stage_b import StageB @@ -946,7 +947,7 @@ class GenmoMochi(BaseModel): class LTXV(BaseModel): def __init__(self, model_config, model_type=ModelType.FLUX, device=None): - super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.lightricks.model.LTXVModel) #TODO + super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.lightricks.model.LTXVModel) def extra_conds(self, **kwargs): out = super().extra_conds(**kwargs) @@ -977,6 +978,60 @@ class LTXV(BaseModel): def scale_latent_inpaint(self, sigma, noise, latent_image, **kwargs): return latent_image +class LTXAV(BaseModel): + def __init__(self, model_config, model_type=ModelType.FLUX, device=None): + super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.lightricks.av_model.LTXAVModel) #TODO + + def extra_conds(self, **kwargs): + out = super().extra_conds(**kwargs) + attention_mask = kwargs.get("attention_mask", None) + if attention_mask is not None: + out['attention_mask'] = comfy.conds.CONDRegular(attention_mask) + cross_attn = kwargs.get("cross_attn", None) + if cross_attn is not None: + out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn) + + out['frame_rate'] = comfy.conds.CONDConstant(kwargs.get("frame_rate", 25)) + + denoise_mask = kwargs.get("concat_mask", kwargs.get("denoise_mask", None)) + + audio_denoise_mask = None + if denoise_mask is not None and "latent_shapes" in kwargs: + denoise_mask = utils.unpack_latents(denoise_mask, kwargs["latent_shapes"]) + if len(denoise_mask) > 1: + audio_denoise_mask = denoise_mask[1] + denoise_mask = denoise_mask[0] + + if denoise_mask is not None: + out["denoise_mask"] = comfy.conds.CONDRegular(denoise_mask) + + if audio_denoise_mask is not None: + out["audio_denoise_mask"] = comfy.conds.CONDRegular(audio_denoise_mask) + + keyframe_idxs = kwargs.get("keyframe_idxs", None) + if keyframe_idxs is not None: + out['keyframe_idxs'] = comfy.conds.CONDRegular(keyframe_idxs) + + latent_shapes = kwargs.get("latent_shapes", None) + if latent_shapes is not None: + out['latent_shapes'] = comfy.conds.CONDConstant(latent_shapes) + + return out + + def process_timestep(self, timestep, x, denoise_mask=None, audio_denoise_mask=None, **kwargs): + v_timestep = timestep + a_timestep = timestep + + if denoise_mask is not None: + v_timestep = self.diffusion_model.patchifier.patchify(((denoise_mask) * timestep.view([timestep.shape[0]] + [1] * (denoise_mask.ndim - 1)))[:, :1])[0] + if audio_denoise_mask is not None: + a_timestep = self.diffusion_model.a_patchifier.patchify(((audio_denoise_mask) * timestep.view([timestep.shape[0]] + [1] * (audio_denoise_mask.ndim - 1)))[:, :1, :, :1])[0] + + return v_timestep, a_timestep + + def scale_latent_inpaint(self, sigma, noise, latent_image, **kwargs): + return latent_image + class HunyuanVideo(BaseModel): def __init__(self, model_config, model_type=ModelType.FLOW, device=None): super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.hunyuan_video.model.HunyuanVideo) diff --git a/comfy/model_detection.py b/comfy/model_detection.py index 539e296ed..0853b3aec 100644 --- a/comfy/model_detection.py +++ b/comfy/model_detection.py @@ -305,7 +305,7 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): if '{}adaln_single.emb.timestep_embedder.linear_1.bias'.format(key_prefix) in state_dict_keys: #Lightricks ltxv dit_config = {} - dit_config["image_model"] = "ltxv" + dit_config["image_model"] = "ltxav" if f'{key_prefix}audio_adaln_single.linear.weight' in state_dict_keys else "ltxv" dit_config["num_layers"] = count_blocks(state_dict_keys, '{}transformer_blocks.'.format(key_prefix) + '{}.') shape = state_dict['{}transformer_blocks.0.attn2.to_k.weight'.format(key_prefix)].shape dit_config["attention_head_dim"] = shape[0] // 32 diff --git a/comfy/sd.py b/comfy/sd.py index 7de7dd9c6..32157e18b 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -1041,7 +1041,8 @@ class TEModel(Enum): MISTRAL3_24B_PRUNED_FLUX2 = 15 QWEN3_4B = 16 QWEN3_2B = 17 - JINA_CLIP_2 = 18 + GEMMA_3_12B = 18 + JINA_CLIP_2 = 19 def detect_te_model(sd): @@ -1067,6 +1068,8 @@ def detect_te_model(sd): return TEModel.BYT5_SMALL_GLYPH return TEModel.T5_BASE if 'model.layers.0.post_feedforward_layernorm.weight' in sd: + if 'model.layers.47.self_attn.q_norm.weight' in sd: + return TEModel.GEMMA_3_12B if 'model.layers.0.self_attn.q_norm.weight' in sd: return TEModel.GEMMA_3_4B return TEModel.GEMMA_2_2B @@ -1271,6 +1274,10 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip elif clip_type == CLIPType.KANDINSKY5_IMAGE: clip_target.clip = comfy.text_encoders.kandinsky5.te(**llama_detect(clip_data)) clip_target.tokenizer = comfy.text_encoders.kandinsky5.Kandinsky5TokenizerImage + elif clip_type == CLIPType.LTXV: + clip_target.clip = comfy.text_encoders.lt.ltxav_te(**llama_detect(clip_data)) + clip_target.tokenizer = comfy.text_encoders.lt.LTXAVGemmaTokenizer + tokenizer_data["spiece_model"] = clip_data[0].get("spiece_model", None) elif clip_type == CLIPType.NEWBIE: clip_target.clip = comfy.text_encoders.newbie.te(**llama_detect(clip_data)) clip_target.tokenizer = comfy.text_encoders.newbie.NewBieTokenizer diff --git a/comfy/supported_models.py b/comfy/supported_models.py index 1888f35ba..ee9a79001 100644 --- a/comfy/supported_models.py +++ b/comfy/supported_models.py @@ -836,6 +836,21 @@ class LTXV(supported_models_base.BASE): t5_detect = comfy.text_encoders.sd3_clip.t5_xxl_detect(state_dict, "{}t5xxl.transformer.".format(pref)) return supported_models_base.ClipTarget(comfy.text_encoders.lt.LTXVT5Tokenizer, comfy.text_encoders.lt.ltxv_te(**t5_detect)) +class LTXAV(LTXV): + unet_config = { + "image_model": "ltxav", + } + + latent_format = latent_formats.LTXAV + + def __init__(self, unet_config): + super().__init__(unet_config) + self.memory_usage_factor = 0.055 # TODO + + def get_model(self, state_dict, prefix="", device=None): + out = model_base.LTXAV(self, device=device) + return out + class HunyuanVideo(supported_models_base.BASE): unet_config = { "image_model": "hunyuan_video", @@ -1536,6 +1551,6 @@ class Kandinsky5Image(Kandinsky5): return supported_models_base.ClipTarget(comfy.text_encoders.kandinsky5.Kandinsky5TokenizerImage, comfy.text_encoders.kandinsky5.te(**hunyuan_detect)) -models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanVideo15_SR_Distilled, HunyuanVideo15, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, ZImage, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, WAN22_Animate, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, Omnigen2, QwenImage, Flux2, Kandinsky5Image, Kandinsky5] +models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, LTXAV, HunyuanVideo15_SR_Distilled, HunyuanVideo15, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, ZImage, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, WAN22_Animate, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, Omnigen2, QwenImage, Flux2, Kandinsky5Image, Kandinsky5] models += [SVD_img2vid] diff --git a/comfy/text_encoders/llama.py b/comfy/text_encoders/llama.py index faa4e1de8..76731576b 100644 --- a/comfy/text_encoders/llama.py +++ b/comfy/text_encoders/llama.py @@ -7,6 +7,7 @@ import math from comfy.ldm.modules.attention import optimized_attention_for_device import comfy.model_management import comfy.ldm.common_dit +import comfy.clip_model from . import qwen_vl @@ -188,6 +189,31 @@ class Gemma3_4B_Config: rope_scale = [8.0, 1.0] final_norm: bool = True +@dataclass +class Gemma3_12B_Config: + vocab_size: int = 262208 + hidden_size: int = 3840 + intermediate_size: int = 15360 + num_hidden_layers: int = 48 + num_attention_heads: int = 16 + num_key_value_heads: int = 8 + max_position_embeddings: int = 131072 + rms_norm_eps: float = 1e-6 + rope_theta = [1000000.0, 10000.0] + transformer_type: str = "gemma3" + head_dim = 256 + rms_norm_add = True + mlp_activation = "gelu_pytorch_tanh" + qkv_bias = False + rope_dims = None + q_norm = "gemma3" + k_norm = "gemma3" + sliding_attention = [1024, 1024, 1024, 1024, 1024, False] + rope_scale = [8.0, 1.0] + final_norm: bool = True + vision_config = {"num_channels": 3, "hidden_act": "gelu_pytorch_tanh", "hidden_size": 1152, "image_size": 896, "intermediate_size": 4304, "model_type": "siglip_vision_model", "num_attention_heads": 16, "num_hidden_layers": 27, "patch_size": 14} + mm_tokens_per_image = 256 + class RMSNorm(nn.Module): def __init__(self, dim: int, eps: float = 1e-5, add=False, device=None, dtype=None): super().__init__() @@ -520,6 +546,41 @@ class Llama2_(nn.Module): return x, intermediate + +class Gemma3MultiModalProjector(torch.nn.Module): + def __init__(self, config, dtype, device, operations): + super().__init__() + + self.mm_input_projection_weight = nn.Parameter( + torch.empty(config.vision_config["hidden_size"], config.hidden_size, device=device, dtype=dtype) + ) + + self.mm_soft_emb_norm = RMSNorm(config.vision_config["hidden_size"], eps=config.rms_norm_eps, add=config.rms_norm_add, device=device, dtype=dtype) + + self.patches_per_image = int(config.vision_config["image_size"] // config.vision_config["patch_size"]) + self.tokens_per_side = int(config.mm_tokens_per_image**0.5) + self.kernel_size = self.patches_per_image // self.tokens_per_side + self.avg_pool = nn.AvgPool2d(kernel_size=self.kernel_size, stride=self.kernel_size) + + def forward(self, vision_outputs: torch.Tensor): + batch_size, _, seq_length = vision_outputs.shape + + reshaped_vision_outputs = vision_outputs.transpose(1, 2) + reshaped_vision_outputs = reshaped_vision_outputs.reshape( + batch_size, seq_length, self.patches_per_image, self.patches_per_image + ) + reshaped_vision_outputs = reshaped_vision_outputs.contiguous() + + pooled_vision_outputs = self.avg_pool(reshaped_vision_outputs) + pooled_vision_outputs = pooled_vision_outputs.flatten(2) + pooled_vision_outputs = pooled_vision_outputs.transpose(1, 2) + + normed_vision_outputs = self.mm_soft_emb_norm(pooled_vision_outputs) + + projected_vision_outputs = torch.matmul(normed_vision_outputs, comfy.model_management.cast_to_device(self.mm_input_projection_weight, device=normed_vision_outputs.device, dtype=normed_vision_outputs.dtype)) + return projected_vision_outputs.type_as(vision_outputs) + + class BaseLlama: def get_input_embeddings(self): return self.model.embed_tokens @@ -636,3 +697,21 @@ class Gemma3_4B(BaseLlama, torch.nn.Module): self.model = Llama2_(config, device=device, dtype=dtype, ops=operations) self.dtype = dtype + +class Gemma3_12B(BaseLlama, torch.nn.Module): + def __init__(self, config_dict, dtype, device, operations): + super().__init__() + config = Gemma3_12B_Config(**config_dict) + self.num_layers = config.num_hidden_layers + + self.model = Llama2_(config, device=device, dtype=dtype, ops=operations) + self.multi_modal_projector = Gemma3MultiModalProjector(config, dtype, device, operations) + self.vision_model = comfy.clip_model.CLIPVision(config.vision_config, dtype, device, operations) + self.dtype = dtype + self.image_size = config.vision_config["image_size"] + + def preprocess_embed(self, embed, device): + if embed["type"] == "image": + image = comfy.clip_model.clip_preprocess(embed["data"], size=self.image_size, mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5], crop=True) + return self.multi_modal_projector(self.vision_model(image.to(device, dtype=torch.float32))[0]), None + return None, None diff --git a/comfy/text_encoders/lt.py b/comfy/text_encoders/lt.py index 48ea67e67..2c2d453e8 100644 --- a/comfy/text_encoders/lt.py +++ b/comfy/text_encoders/lt.py @@ -1,7 +1,11 @@ from comfy import sd1_clip import os from transformers import T5TokenizerFast +from .spiece_tokenizer import SPieceTokenizer import comfy.text_encoders.genmo +from comfy.ldm.lightricks.embeddings_connector import Embeddings1DConnector +import torch +import comfy.utils class T5XXLTokenizer(sd1_clip.SDTokenizer): def __init__(self, embedding_directory=None, tokenizer_data={}): @@ -16,3 +20,110 @@ class LTXVT5Tokenizer(sd1_clip.SD1Tokenizer): def ltxv_te(*args, **kwargs): return comfy.text_encoders.genmo.mochi_te(*args, **kwargs) + + +class Gemma3_12BTokenizer(sd1_clip.SDTokenizer): + def __init__(self, embedding_directory=None, tokenizer_data={}): + tokenizer = tokenizer_data.get("spiece_model", None) + super().__init__(tokenizer, pad_with_end=False, embedding_size=3840, embedding_key='gemma3_12b', tokenizer_class=SPieceTokenizer, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, tokenizer_args={"add_bos": True, "add_eos": False}, tokenizer_data=tokenizer_data) + + def state_dict(self): + return {"spiece_model": self.tokenizer.serialize_model()} + +class LTXAVGemmaTokenizer(sd1_clip.SD1Tokenizer): + def __init__(self, embedding_directory=None, tokenizer_data={}): + super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, name="gemma3_12b", tokenizer=Gemma3_12BTokenizer) + +class Gemma3_12BModel(sd1_clip.SDClipModel): + def __init__(self, device="cpu", layer="all", layer_idx=None, dtype=None, attention_mask=True, model_options={}): + llama_scaled_fp8 = model_options.get("gemma_scaled_fp8", None) + if llama_scaled_fp8 is not None: + model_options = model_options.copy() + model_options["scaled_fp8"] = llama_scaled_fp8 + + super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"start": 2, "pad": 0}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Gemma3_12B, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options) + + def tokenize_with_weights(self, text, return_word_ids=False, llama_template="{}", image_embeds=None, **kwargs): + text = llama_template.format(text) + text_tokens = super().tokenize_with_weights(text, return_word_ids) + embed_count = 0 + for k in text_tokens: + tt = text_tokens[k] + for r in tt: + for i in range(len(r)): + if r[i][0] == 262144: + if image_embeds is not None and embed_count < image_embeds.shape[0]: + r[i] = ({"type": "embedding", "data": image_embeds[embed_count], "original_type": "image"},) + r[i][1:] + embed_count += 1 + return text_tokens + +class LTXAVTEModel(torch.nn.Module): + def __init__(self, dtype_llama=None, device="cpu", dtype=None, model_options={}): + super().__init__() + self.dtypes = set() + self.dtypes.add(dtype) + + self.gemma3_12b = Gemma3_12BModel(device=device, dtype=dtype_llama, model_options=model_options, layer="all", layer_idx=None) + self.dtypes.add(dtype_llama) + + operations = self.gemma3_12b.operations # TODO + self.text_embedding_projection = operations.Linear(3840 * 49, 3840, bias=False, dtype=dtype, device=device) + + self.audio_embeddings_connector = Embeddings1DConnector( + split_rope=True, + double_precision_rope=True, + dtype=dtype, + device=device, + operations=operations, + ) + + self.video_embeddings_connector = Embeddings1DConnector( + split_rope=True, + double_precision_rope=True, + dtype=dtype, + device=device, + operations=operations, + ) + + def set_clip_options(self, options): + self.gemma3_12b.set_clip_options(options) + + def reset_clip_options(self): + self.gemma3_12b.reset_clip_options() + + def encode_token_weights(self, token_weight_pairs): + token_weight_pairs = token_weight_pairs["gemma3_12b"] + + out, pooled, extra = self.gemma3_12b.encode_token_weights(token_weight_pairs) + out_device = out.device + out = out.movedim(1, -1).to(self.text_embedding_projection.weight.device) + out = 8.0 * (out - out.mean(dim=(1, 2), keepdim=True)) / (out.amax(dim=(1, 2), keepdim=True) - out.amin(dim=(1, 2), keepdim=True) + 1e-6) + out = out.reshape((out.shape[0], out.shape[1], -1)) + out = self.text_embedding_projection(out) + out_vid = self.video_embeddings_connector(out)[0] + out_audio = self.audio_embeddings_connector(out)[0] + out = torch.concat((out_vid, out_audio), dim=-1) + + return out.to(out_device), pooled + + def load_sd(self, sd): + if "model.layers.47.self_attn.q_norm.weight" in sd: + return self.gemma3_12b.load_sd(sd) + else: + sdo = comfy.utils.state_dict_prefix_replace(sd, {"text_embedding_projection.aggregate_embed.weight": "text_embedding_projection.weight", "model.diffusion_model.video_embeddings_connector.": "video_embeddings_connector.", "model.diffusion_model.audio_embeddings_connector.": "audio_embeddings_connector."}, filter_keys=True) + if len(sdo) == 0: + sdo = sd + + return self.load_state_dict(sdo, strict=False) + + +def ltxav_te(dtype_llama=None, llama_scaled_fp8=None): + class LTXAVTEModel_(LTXAVTEModel): + def __init__(self, device="cpu", dtype=None, model_options={}): + if llama_scaled_fp8 is not None and "llama_scaled_fp8" not in model_options: + model_options = model_options.copy() + model_options["llama_scaled_fp8"] = llama_scaled_fp8 + if dtype_llama is not None: + dtype = dtype_llama + super().__init__(dtype_llama=dtype_llama, device=device, dtype=dtype, model_options=model_options) + return LTXAVTEModel_ diff --git a/comfy/utils.py b/comfy/utils.py index e4162d7ac..ffa98c9b1 100644 --- a/comfy/utils.py +++ b/comfy/utils.py @@ -1198,7 +1198,7 @@ def unpack_latents(combined_latent, latent_shapes): combined_latent = combined_latent[:, :, cut:] output_tensors.append(tens.reshape([tens.shape[0]] + list(shape)[1:])) else: - output_tensors = combined_latent + output_tensors = [combined_latent] return output_tensors def detect_layer_quantization(state_dict, prefix): diff --git a/comfy_extras/nodes_audio.py b/comfy_extras/nodes_audio.py index c7916443c..94ad5e8a8 100644 --- a/comfy_extras/nodes_audio.py +++ b/comfy_extras/nodes_audio.py @@ -112,7 +112,7 @@ class VAEDecodeAudio(IO.ComfyNode): std = torch.std(audio, dim=[1,2], keepdim=True) * 5.0 std[std < 1.0] = 1.0 audio /= std - return IO.NodeOutput({"waveform": audio, "sample_rate": 44100}) + return IO.NodeOutput({"waveform": audio, "sample_rate": 44100 if "sample_rate" not in samples else samples["sample_rate"]}) decode = execute # TODO: remove diff --git a/comfy_extras/nodes_hunyuan.py b/comfy_extras/nodes_hunyuan.py index 32be182f1..ceff657d3 100644 --- a/comfy_extras/nodes_hunyuan.py +++ b/comfy_extras/nodes_hunyuan.py @@ -5,7 +5,9 @@ import comfy.model_management from typing_extensions import override from comfy_api.latest import ComfyExtension, io from comfy.ldm.hunyuan_video.upsampler import HunyuanVideo15SRModel +from comfy.ldm.lightricks.latent_upsampler import LatentUpsampler import folder_paths +import json class CLIPTextEncodeHunyuanDiT(io.ComfyNode): @classmethod @@ -186,7 +188,7 @@ class LatentUpscaleModelLoader(io.ComfyNode): @classmethod def execute(cls, model_name) -> io.NodeOutput: model_path = folder_paths.get_full_path_or_raise("latent_upscale_models", model_name) - sd = comfy.utils.load_torch_file(model_path, safe_load=True) + sd, metadata = comfy.utils.load_torch_file(model_path, safe_load=True, return_metadata=True) if "blocks.0.block.0.conv.weight" in sd: config = { @@ -197,6 +199,8 @@ class LatentUpscaleModelLoader(io.ComfyNode): "global_residual": False, } model_type = "720p" + model = HunyuanVideo15SRModel(model_type, config) + model.load_sd(sd) elif "up.0.block.0.conv1.conv.weight" in sd: sd = {key.replace("nin_shortcut", "nin_shortcut.conv", 1): value for key, value in sd.items()} config = { @@ -205,9 +209,12 @@ class LatentUpscaleModelLoader(io.ComfyNode): "block_out_channels": tuple(sd[f"up.{i}.block.0.conv1.conv.weight"].shape[0] for i in range(len([k for k in sd.keys() if k.startswith("up.") and k.endswith(".block.0.conv1.conv.weight")]))), } model_type = "1080p" - - model = HunyuanVideo15SRModel(model_type, config) - model.load_sd(sd) + model = HunyuanVideo15SRModel(model_type, config) + model.load_sd(sd) + elif "post_upsample_res_blocks.0.conv2.bias" in sd: + config = json.loads(metadata["config"]) + model = LatentUpsampler.from_config(config).to(dtype=comfy.model_management.vae_dtype(allowed_dtypes=[torch.bfloat16, torch.float32])) + model.load_state_dict(sd) return io.NodeOutput(model) diff --git a/comfy_extras/nodes_lt.py b/comfy_extras/nodes_lt.py index 50da5f4eb..b91a22309 100644 --- a/comfy_extras/nodes_lt.py +++ b/comfy_extras/nodes_lt.py @@ -81,6 +81,59 @@ class LTXVImgToVideo(io.ComfyNode): generate = execute # TODO: remove +class LTXVImgToVideoInplace(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="LTXVImgToVideoInplace", + category="conditioning/video_models", + inputs=[ + io.Vae.Input("vae"), + io.Image.Input("image"), + io.Latent.Input("latent"), + io.Float.Input("strength", default=1.0, min=0.0, max=1.0), + io.Boolean.Input("bypass", default=False, tooltip="Bypass the conditioning.") + ], + outputs=[ + io.Latent.Output(display_name="latent"), + ], + ) + + @classmethod + def execute(cls, vae, image, latent, strength, bypass=False) -> io.NodeOutput: + if bypass: + return (latent,) + + samples = latent["samples"] + _, height_scale_factor, width_scale_factor = ( + vae.downscale_index_formula + ) + + batch, _, latent_frames, latent_height, latent_width = samples.shape + width = latent_width * width_scale_factor + height = latent_height * height_scale_factor + + if image.shape[1] != height or image.shape[2] != width: + pixels = comfy.utils.common_upscale(image.movedim(-1, 1), width, height, "bilinear", "center").movedim(1, -1) + else: + pixels = image + encode_pixels = pixels[:, :, :, :3] + t = vae.encode(encode_pixels) + + samples[:, :, :t.shape[2]] = t + + conditioning_latent_frames_mask = torch.ones( + (batch, 1, latent_frames, 1, 1), + dtype=torch.float32, + device=samples.device, + ) + conditioning_latent_frames_mask[:, :, :t.shape[2]] = 1.0 - strength + + return io.NodeOutput({"samples": samples, "noise_mask": conditioning_latent_frames_mask}) + + generate = execute # TODO: remove + + def conditioning_get_any_value(conditioning, key, default=None): for t in conditioning: if key in t[1]: @@ -106,12 +159,12 @@ def get_keyframe_idxs(cond): keyframe_idxs = conditioning_get_any_value(cond, "keyframe_idxs", None) if keyframe_idxs is None: return None, 0 - num_keyframes = torch.unique(keyframe_idxs[:, 0]).shape[0] + # keyframe_idxs contains start/end positions (last dimension), checking for unqiue values only for start + num_keyframes = torch.unique(keyframe_idxs[:, 0, :, 0]).shape[0] return keyframe_idxs, num_keyframes class LTXVAddGuide(io.ComfyNode): - NUM_PREFIX_FRAMES = 2 - PATCHIFIER = SymmetricPatchifier(1) + PATCHIFIER = SymmetricPatchifier(1, start_end=True) @classmethod def define_schema(cls): @@ -182,26 +235,35 @@ class LTXVAddGuide(io.ComfyNode): return node_helpers.conditioning_set_values(cond, {"keyframe_idxs": keyframe_idxs}) @classmethod - def append_keyframe(cls, positive, negative, frame_idx, latent_image, noise_mask, guiding_latent, strength, scale_factors): - _, latent_idx = cls.get_latent_index( - cond=positive, - latent_length=latent_image.shape[2], - guide_length=guiding_latent.shape[2], - frame_idx=frame_idx, - scale_factors=scale_factors, - ) - noise_mask[:, :, latent_idx:latent_idx + guiding_latent.shape[2]] = 1.0 + def append_keyframe(cls, positive, negative, frame_idx, latent_image, noise_mask, guiding_latent, strength, scale_factors, guide_mask=None, in_channels=128): + if latent_image.shape[1] != in_channels or guiding_latent.shape[1] != in_channels: + raise ValueError("Adding guide to a combined AV latent is not supported.") positive = cls.add_keyframe_index(positive, frame_idx, guiding_latent, scale_factors) negative = cls.add_keyframe_index(negative, frame_idx, guiding_latent, scale_factors) - mask = torch.full( - (noise_mask.shape[0], 1, guiding_latent.shape[2], noise_mask.shape[3], noise_mask.shape[4]), - 1.0 - strength, - dtype=noise_mask.dtype, - device=noise_mask.device, - ) + if guide_mask is not None: + target_h = max(noise_mask.shape[3], guide_mask.shape[3]) + target_w = max(noise_mask.shape[4], guide_mask.shape[4]) + if noise_mask.shape[3] == 1 or noise_mask.shape[4] == 1: + noise_mask = noise_mask.expand(-1, -1, -1, target_h, target_w) + + if guide_mask.shape[3] == 1 or guide_mask.shape[4] == 1: + guide_mask = guide_mask.expand(-1, -1, -1, target_h, target_w) + mask = guide_mask - strength + else: + mask = torch.full( + (noise_mask.shape[0], 1, guiding_latent.shape[2], noise_mask.shape[3], noise_mask.shape[4]), + 1.0 - strength, + dtype=noise_mask.dtype, + device=noise_mask.device, + ) + # This solves audio video combined latent case where latent_image has audio latent concatenated + # in channel dimension with video latent. The solution is to pad guiding latent accordingly. + if latent_image.shape[1] > guiding_latent.shape[1]: + pad_len = latent_image.shape[1] - guiding_latent.shape[1] + guiding_latent = torch.nn.functional.pad(guiding_latent, pad=(0, 0, 0, 0, 0, 0, 0, pad_len), value=0) latent_image = torch.cat([latent_image, guiding_latent], dim=2) noise_mask = torch.cat([noise_mask, mask], dim=2) return positive, negative, latent_image, noise_mask @@ -238,33 +300,17 @@ class LTXVAddGuide(io.ComfyNode): frame_idx, latent_idx = cls.get_latent_index(positive, latent_length, len(image), frame_idx, scale_factors) assert latent_idx + t.shape[2] <= latent_length, "Conditioning frames exceed the length of the latent sequence." - num_prefix_frames = min(cls.NUM_PREFIX_FRAMES, t.shape[2]) - positive, negative, latent_image, noise_mask = cls.append_keyframe( positive, negative, frame_idx, latent_image, noise_mask, - t[:, :, :num_prefix_frames], + t, strength, scale_factors, ) - latent_idx += num_prefix_frames - - t = t[:, :, num_prefix_frames:] - if t.shape[2] == 0: - return io.NodeOutput(positive, negative, {"samples": latent_image, "noise_mask": noise_mask}) - - latent_image, noise_mask = cls.replace_latent_frames( - latent_image, - noise_mask, - t, - latent_idx, - strength, - ) - return io.NodeOutput(positive, negative, {"samples": latent_image, "noise_mask": noise_mask}) generate = execute # TODO: remove @@ -507,18 +553,90 @@ class LTXVPreprocess(io.ComfyNode): preprocess = execute # TODO: remove + +import comfy.nested_tensor +class LTXVConcatAVLatent(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="LTXVConcatAVLatent", + category="latent/video/ltxv", + inputs=[ + io.Latent.Input("video_latent"), + io.Latent.Input("audio_latent"), + ], + outputs=[ + io.Latent.Output(display_name="latent"), + ], + ) + + @classmethod + def execute(cls, video_latent, audio_latent) -> io.NodeOutput: + output = {} + output.update(video_latent) + output.update(audio_latent) + video_noise_mask = video_latent.get("noise_mask", None) + audio_noise_mask = audio_latent.get("noise_mask", None) + + if video_noise_mask is not None or audio_noise_mask is not None: + if video_noise_mask is None: + video_noise_mask = torch.ones_like(video_latent["samples"]) + if audio_noise_mask is None: + audio_noise_mask = torch.ones_like(audio_latent["samples"]) + output["noise_mask"] = comfy.nested_tensor.NestedTensor((video_noise_mask, audio_noise_mask)) + + output["samples"] = comfy.nested_tensor.NestedTensor((video_latent["samples"], audio_latent["samples"])) + + return io.NodeOutput(output) + + +class LTXVSeparateAVLatent(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="LTXVSeparateAVLatent", + category="latent/video/ltxv", + description="LTXV Separate AV Latent", + inputs=[ + io.Latent.Input("av_latent"), + ], + outputs=[ + io.Latent.Output(display_name="video_latent"), + io.Latent.Output(display_name="audio_latent"), + ], + ) + + @classmethod + def execute(cls, av_latent) -> io.NodeOutput: + latents = av_latent["samples"].unbind() + video_latent = av_latent.copy() + video_latent["samples"] = latents[0] + audio_latent = av_latent.copy() + audio_latent["samples"] = latents[1] + if "noise_mask" in av_latent: + masks = av_latent["noise_mask"] + if masks is not None: + masks = masks.unbind() + video_latent["noise_mask"] = masks[0] + audio_latent["noise_mask"] = masks[1] + return io.NodeOutput(video_latent, audio_latent) + + class LtxvExtension(ComfyExtension): @override async def get_node_list(self) -> list[type[io.ComfyNode]]: return [ EmptyLTXVLatentVideo, LTXVImgToVideo, + LTXVImgToVideoInplace, ModelSamplingLTXV, LTXVConditioning, LTXVScheduler, LTXVAddGuide, LTXVPreprocess, LTXVCropGuides, + LTXVConcatAVLatent, + LTXVSeparateAVLatent, ] diff --git a/comfy_extras/nodes_lt_audio.py b/comfy_extras/nodes_lt_audio.py new file mode 100644 index 000000000..b0b7000ef --- /dev/null +++ b/comfy_extras/nodes_lt_audio.py @@ -0,0 +1,183 @@ +import folder_paths +import comfy.utils +import comfy.model_management +import torch + +from comfy.ldm.lightricks.vae.audio_vae import AudioVAE +from comfy_api.latest import ComfyExtension, io + + +class LTXVAudioVAELoader(io.ComfyNode): + @classmethod + def define_schema(cls) -> io.Schema: + return io.Schema( + node_id="LTXVAudioVAELoader", + display_name="LTXV Audio VAE Loader", + category="audio", + inputs=[ + io.Combo.Input( + "ckpt_name", + options=folder_paths.get_filename_list("checkpoints"), + tooltip="Audio VAE checkpoint to load.", + ) + ], + outputs=[io.Vae.Output(display_name="Audio VAE")], + ) + + @classmethod + def execute(cls, ckpt_name: str) -> io.NodeOutput: + ckpt_path = folder_paths.get_full_path_or_raise("checkpoints", ckpt_name) + sd, metadata = comfy.utils.load_torch_file(ckpt_path, return_metadata=True) + return io.NodeOutput(AudioVAE(sd, metadata)) + + +class LTXVAudioVAEEncode(io.ComfyNode): + @classmethod + def define_schema(cls) -> io.Schema: + return io.Schema( + node_id="LTXVAudioVAEEncode", + display_name="LTXV Audio VAE Encode", + category="audio", + inputs=[ + io.Audio.Input("audio", tooltip="The audio to be encoded."), + io.Vae.Input( + id="audio_vae", + display_name="Audio VAE", + tooltip="The Audio VAE model to use for encoding.", + ), + ], + outputs=[io.Latent.Output(display_name="Audio Latent")], + ) + + @classmethod + def execute(cls, audio, audio_vae: AudioVAE) -> io.NodeOutput: + audio_latents = audio_vae.encode(audio) + return io.NodeOutput( + { + "samples": audio_latents, + "sample_rate": int(audio_vae.sample_rate), + "type": "audio", + } + ) + + +class LTXVAudioVAEDecode(io.ComfyNode): + @classmethod + def define_schema(cls) -> io.Schema: + return io.Schema( + node_id="LTXVAudioVAEDecode", + display_name="LTXV Audio VAE Decode", + category="audio", + inputs=[ + io.Latent.Input("samples", tooltip="The latent to be decoded."), + io.Vae.Input( + id="audio_vae", + display_name="Audio VAE", + tooltip="The Audio VAE model used for decoding the latent.", + ), + ], + outputs=[io.Audio.Output(display_name="Audio")], + ) + + @classmethod + def execute(cls, samples, audio_vae: AudioVAE) -> io.NodeOutput: + audio_latent = samples["samples"] + if audio_latent.is_nested: + audio_latent = audio_latent.unbind()[-1] + audio = audio_vae.decode(audio_latent).to(audio_latent.device) + output_audio_sample_rate = audio_vae.output_sample_rate + return io.NodeOutput( + { + "waveform": audio, + "sample_rate": int(output_audio_sample_rate), + } + ) + + +class LTXVEmptyLatentAudio(io.ComfyNode): + @classmethod + def define_schema(cls) -> io.Schema: + return io.Schema( + node_id="LTXVEmptyLatentAudio", + display_name="LTXV Empty Latent Audio", + category="latent/audio", + inputs=[ + io.Int.Input( + "frames_number", + default=97, + min=1, + max=1000, + step=1, + display_mode=io.NumberDisplay.number, + tooltip="Number of frames.", + ), + io.Int.Input( + "frame_rate", + default=25, + min=1, + max=1000, + step=1, + display_mode=io.NumberDisplay.number, + tooltip="Number of frames per second.", + ), + io.Int.Input( + "batch_size", + default=1, + min=1, + max=4096, + display_mode=io.NumberDisplay.number, + tooltip="The number of latent audio samples in the batch.", + ), + io.Vae.Input( + id="audio_vae", + display_name="Audio VAE", + tooltip="The Audio VAE model to get configuration from.", + ), + ], + outputs=[io.Latent.Output(display_name="Latent")], + ) + + @classmethod + def execute( + cls, + frames_number: int, + frame_rate: int, + batch_size: int, + audio_vae: AudioVAE, + ) -> io.NodeOutput: + """Generate empty audio latents matching the reference pipeline structure.""" + + assert audio_vae is not None, "Audio VAE model is required" + + z_channels = audio_vae.latent_channels + audio_freq = audio_vae.latent_frequency_bins + sampling_rate = int(audio_vae.sample_rate) + + num_audio_latents = audio_vae.num_of_latents_from_frames(frames_number, frame_rate) + + audio_latents = torch.zeros( + (batch_size, z_channels, num_audio_latents, audio_freq), + device=comfy.model_management.intermediate_device(), + ) + + return io.NodeOutput( + { + "samples": audio_latents, + "sample_rate": sampling_rate, + "type": "audio", + } + ) + + +class LTXVAudioExtension(ComfyExtension): + async def get_node_list(self) -> list[type[io.ComfyNode]]: + return [ + LTXVAudioVAELoader, + LTXVAudioVAEEncode, + LTXVAudioVAEDecode, + LTXVEmptyLatentAudio, + ] + + +async def comfy_entrypoint() -> ComfyExtension: + return LTXVAudioExtension() diff --git a/comfy_extras/nodes_lt_upsampler.py b/comfy_extras/nodes_lt_upsampler.py new file mode 100644 index 000000000..f99ba13fb --- /dev/null +++ b/comfy_extras/nodes_lt_upsampler.py @@ -0,0 +1,75 @@ +from comfy import model_management +import math + +class LTXVLatentUpsampler: + """ + Upsamples a video latent by a factor of 2. + """ + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "samples": ("LATENT",), + "upscale_model": ("LATENT_UPSCALE_MODEL",), + "vae": ("VAE",), + } + } + + RETURN_TYPES = ("LATENT",) + FUNCTION = "upsample_latent" + CATEGORY = "latent/video" + EXPERIMENTAL = True + + def upsample_latent( + self, + samples: dict, + upscale_model, + vae, + ) -> tuple: + """ + Upsample the input latent using the provided model. + + Args: + samples (dict): Input latent samples + upscale_model (LatentUpsampler): Loaded upscale model + vae: VAE model for normalization + auto_tiling (bool): Whether to automatically tile the input for processing + + Returns: + tuple: Tuple containing the upsampled latent + """ + device = model_management.get_torch_device() + memory_required = model_management.module_size(upscale_model) + + model_dtype = next(upscale_model.parameters()).dtype + latents = samples["samples"] + input_dtype = latents.dtype + + memory_required += math.prod(latents.shape) * 3000.0 # TODO: more accurate + model_management.free_memory(memory_required, device) + + try: + upscale_model.to(device) # TODO: use the comfy model management system. + + latents = latents.to(dtype=model_dtype, device=device) + + """Upsample latents without tiling.""" + latents = vae.first_stage_model.per_channel_statistics.un_normalize(latents) + upsampled_latents = upscale_model(latents) + finally: + upscale_model.cpu() + + upsampled_latents = vae.first_stage_model.per_channel_statistics.normalize( + upsampled_latents + ) + upsampled_latents = upsampled_latents.to(dtype=input_dtype, device=model_management.intermediate_device()) + return_dict = samples.copy() + return_dict["samples"] = upsampled_latents + return_dict.pop("noise_mask", None) + return (return_dict,) + + +NODE_CLASS_MAPPINGS = { + "LTXVLatentUpsampler": LTXVLatentUpsampler, +} diff --git a/nodes.py b/nodes.py index 662907ae6..56b74ebe3 100644 --- a/nodes.py +++ b/nodes.py @@ -295,7 +295,11 @@ class VAEDecode: DESCRIPTION = "Decodes latent images back into pixel space images." def decode(self, vae, samples): - images = vae.decode(samples["samples"]) + latent = samples["samples"] + if latent.is_nested: + latent = latent.unbind()[0] + + images = vae.decode(latent) if len(images.shape) == 5: #Combine batches images = images.reshape(-1, images.shape[-3], images.shape[-2], images.shape[-1]) return (images, ) @@ -970,7 +974,7 @@ class DualCLIPLoader: def INPUT_TYPES(s): return {"required": { "clip_name1": (folder_paths.get_filename_list("text_encoders"), ), "clip_name2": (folder_paths.get_filename_list("text_encoders"), ), - "type": (["sdxl", "sd3", "flux", "hunyuan_video", "hidream", "hunyuan_image", "hunyuan_video_15", "kandinsky5", "kandinsky5_image", "newbie"], ), + "type": (["sdxl", "sd3", "flux", "hunyuan_video", "hidream", "hunyuan_image", "hunyuan_video_15", "kandinsky5", "kandinsky5_image", "ltxv", "newbie"], ), }, "optional": { "device": (["default", "cpu"], {"advanced": True}), @@ -2331,6 +2335,8 @@ async def init_builtin_extra_nodes(): "nodes_mochi.py", "nodes_slg.py", "nodes_mahiro.py", + "nodes_lt_upsampler.py", + "nodes_lt_audio.py", "nodes_lt.py", "nodes_hooks.py", "nodes_load_3d.py", diff --git a/pyproject.toml b/pyproject.toml index 60378de1e..a7d159be9 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -3,7 +3,7 @@ name = "ComfyUI" version = "0.7.0" readme = "README.md" license = { file = "LICENSE" } -requires-python = ">=3.9" +requires-python = ">=3.10" [project.urls] homepage = "https://www.comfy.org/" From d1b9822f741843c64b2cbd8e1bcdd49794b182ce Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sun, 4 Jan 2026 23:27:31 -0800 Subject: [PATCH 03/48] Add LTXAVTextEncoderLoader node. (#11634) --- comfy_extras/nodes_lt_audio.py | 33 +++++++++++++++++++++++++++++++++ 1 file changed, 33 insertions(+) diff --git a/comfy_extras/nodes_lt_audio.py b/comfy_extras/nodes_lt_audio.py index b0b7000ef..2d3d103b4 100644 --- a/comfy_extras/nodes_lt_audio.py +++ b/comfy_extras/nodes_lt_audio.py @@ -169,6 +169,38 @@ class LTXVEmptyLatentAudio(io.ComfyNode): ) +class LTXAVTextEncoderLoader(io.ComfyNode): + @classmethod + def define_schema(cls) -> io.Schema: + return io.Schema( + node_id="LTXAVTextEncoderLoader", + display_name="LTXV Audio Text Encoder Loader", + category="advanced/loaders", + description="[Recipes]\n\nltxav: gemma 3 12B", + inputs=[ + io.Combo.Input( + "text_encoder", + options=folder_paths.get_filename_list("text_encoders"), + ), + io.Combo.Input( + "ckpt_name", + options=folder_paths.get_filename_list("checkpoints"), + ) + ], + outputs=[io.Clip.Output(display_name="Audio VAE")], + ) + + @classmethod + def execute(cls, text_encoder, ckpt_name, device="default"): + clip_type = comfy.sd.CLIPType.LTXV + + clip_path1 = folder_paths.get_full_path_or_raise("text_encoders", text_encoder) + clip_path2 = folder_paths.get_full_path_or_raise("checkpoints", ckpt_name) + + clip = comfy.sd.load_clip(ckpt_paths=[clip_path1, clip_path2], embedding_directory=folder_paths.get_folder_paths("embeddings"), clip_type=clip_type) + return io.NodeOutput(clip) + + class LTXVAudioExtension(ComfyExtension): async def get_node_list(self) -> list[type[io.ComfyNode]]: return [ @@ -176,6 +208,7 @@ class LTXVAudioExtension(ComfyExtension): LTXVAudioVAEEncode, LTXVAudioVAEDecode, LTXVEmptyLatentAudio, + LTXAVTextEncoderLoader, ] From d157c3299d6f9e1b57981bdb4931f1d7129e4e8d Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 5 Jan 2026 00:48:31 -0800 Subject: [PATCH 04/48] Refactor module_size function. (#11637) --- comfy/model_management.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/comfy/model_management.py b/comfy/model_management.py index 2501cecb7..7f5a8aee9 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -456,7 +456,7 @@ def module_size(module): sd = module.state_dict() for k in sd: t = sd[k] - module_mem += t.nelement() * t.element_size() + module_mem += t.nbytes return module_mem class LoadedModel: From 4f3f9e72a9d0c15d00c0c362b8e90f1db5af6cfb Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 5 Jan 2026 02:41:23 -0800 Subject: [PATCH 05/48] Fix name. (#11638) --- comfy_extras/nodes_lt_audio.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/comfy_extras/nodes_lt_audio.py b/comfy_extras/nodes_lt_audio.py index 2d3d103b4..26b0160d2 100644 --- a/comfy_extras/nodes_lt_audio.py +++ b/comfy_extras/nodes_lt_audio.py @@ -187,7 +187,7 @@ class LTXAVTextEncoderLoader(io.ComfyNode): options=folder_paths.get_filename_list("checkpoints"), ) ], - outputs=[io.Clip.Output(display_name="Audio VAE")], + outputs=[io.Clip.Output()], ) @classmethod From 6da00dd899e3ee6f2a0a8163b080a9f373395025 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 5 Jan 2026 18:48:58 -0800 Subject: [PATCH 06/48] Initial ops changes to use comfy_kitchen: Initial nvfp4 checkpoint support. (#11635) --------- Co-authored-by: Jedrzej Kosinski --- .github/workflows/test-build.yml | 2 +- .github/workflows/test-launch.yml | 4 +- comfy/model_management.py | 4 +- comfy/ops.py | 164 +++-- comfy/quant_ops.py | 641 +++--------------- requirements.txt | 1 + .../comfy_quant/test_mixed_precision.py | 12 +- tests-unit/comfy_quant/test_quant_registry.py | 190 ------ 8 files changed, 223 insertions(+), 795 deletions(-) delete mode 100644 tests-unit/comfy_quant/test_quant_registry.py diff --git a/.github/workflows/test-build.yml b/.github/workflows/test-build.yml index 419873ad8..9160242e9 100644 --- a/.github/workflows/test-build.yml +++ b/.github/workflows/test-build.yml @@ -18,7 +18,7 @@ jobs: strategy: fail-fast: false matrix: - python-version: ["3.9", "3.10", "3.11", "3.12", "3.13"] + python-version: ["3.10", "3.11", "3.12", "3.13", "3.14"] steps: - uses: actions/checkout@v4 - name: Set up Python ${{ matrix.python-version }} diff --git a/.github/workflows/test-launch.yml b/.github/workflows/test-launch.yml index fd70aff23..ef0d3f123 100644 --- a/.github/workflows/test-launch.yml +++ b/.github/workflows/test-launch.yml @@ -32,7 +32,9 @@ jobs: working-directory: ComfyUI - name: Check for unhandled exceptions in server log run: | - if grep -qE "Exception|Error" console_output.log; then + grep -v "Found comfy_kitchen backend triton: {'available': False, 'disabled': True, 'unavailable_reason': \"ImportError: No module named 'triton'\", 'capabilities': \[\]}" console_output.log | grep -v "Found comfy_kitchen backend triton: {'available': False, 'disabled': False, 'unavailable_reason': \"ImportError: No module named 'triton'\", 'capabilities': \[\]}" > console_output_filtered.log + cat console_output_filtered.log + if grep -qE "Exception|Error" console_output_filtered.log; then echo "Unhandled exception/error found in server log." exit 1 fi diff --git a/comfy/model_management.py b/comfy/model_management.py index 7f5a8aee9..22f4de044 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -1156,7 +1156,7 @@ def pin_memory(tensor): if not tensor.is_contiguous(): return False - size = tensor.numel() * tensor.element_size() + size = tensor.nbytes if (TOTAL_PINNED_MEMORY + size) > MAX_PINNED_MEMORY: return False @@ -1183,7 +1183,7 @@ def unpin_memory(tensor): return False ptr = tensor.data_ptr() - size = tensor.numel() * tensor.element_size() + size = tensor.nbytes size_stored = PINNED_MEMORY.get(ptr, None) if size_stored is None: diff --git a/comfy/ops.py b/comfy/ops.py index 16889bb82..f5e1e9230 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -79,7 +79,7 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, of if input is not None: if dtype is None: if isinstance(input, QuantizedTensor): - dtype = input._layout_params["orig_dtype"] + dtype = input.params.orig_dtype else: dtype = input.dtype if bias_dtype is None: @@ -412,26 +412,34 @@ def fp8_linear(self, input): return None input_dtype = input.dtype + input_shape = input.shape + tensor_3d = input.ndim == 3 - if input.ndim == 3 or input.ndim == 2: - w, bias, offload_stream = cast_bias_weight(self, input, dtype=dtype, bias_dtype=input_dtype, offloadable=True) - scale_weight = torch.ones((), device=input.device, dtype=torch.float32) + if tensor_3d: + input = input.reshape(-1, input_shape[2]) - scale_input = torch.ones((), device=input.device, dtype=torch.float32) - input = torch.clamp(input, min=-448, max=448, out=input) - layout_params_weight = {'scale': scale_input, 'orig_dtype': input_dtype} - quantized_input = QuantizedTensor(input.to(dtype).contiguous(), "TensorCoreFP8Layout", layout_params_weight) + if input.ndim != 2: + return None + w, bias, offload_stream = cast_bias_weight(self, input, dtype=dtype, bias_dtype=input_dtype, offloadable=True) + scale_weight = torch.ones((), device=input.device, dtype=torch.float32) - # Wrap weight in QuantizedTensor - this enables unified dispatch - # Call F.linear - __torch_dispatch__ routes to fp8_linear handler in quant_ops.py! - layout_params_weight = {'scale': scale_weight, 'orig_dtype': input_dtype} - quantized_weight = QuantizedTensor(w, "TensorCoreFP8Layout", layout_params_weight) - o = torch.nn.functional.linear(quantized_input, quantized_weight, bias) + scale_input = torch.ones((), device=input.device, dtype=torch.float32) + input = torch.clamp(input, min=-448, max=448, out=input) + input_fp8 = input.to(dtype).contiguous() + layout_params_input = TensorCoreFP8Layout.Params(scale=scale_input, orig_dtype=input_dtype, orig_shape=tuple(input_fp8.shape)) + quantized_input = QuantizedTensor(input_fp8, TensorCoreFP8Layout, layout_params_input) - uncast_bias_weight(self, w, bias, offload_stream) - return o + # Wrap weight in QuantizedTensor - this enables unified dispatch + # Call F.linear - __torch_dispatch__ routes to fp8_linear handler in quant_ops.py! + layout_params_weight = TensorCoreFP8Layout.Params(scale=scale_weight, orig_dtype=input_dtype, orig_shape=tuple(w.shape)) + quantized_weight = QuantizedTensor(w, TensorCoreFP8Layout, layout_params_weight) + o = torch.nn.functional.linear(quantized_input, quantized_weight, bias) - return None + uncast_bias_weight(self, w, bias, offload_stream) + if tensor_3d: + o = o.reshape((input_shape[0], input_shape[1], w.shape[0])) + + return o class fp8_ops(manual_cast): class Linear(manual_cast.Linear): @@ -477,7 +485,12 @@ if CUBLAS_IS_AVAILABLE: # ============================================================================== # Mixed Precision Operations # ============================================================================== -from .quant_ops import QuantizedTensor, QUANT_ALGOS +from .quant_ops import ( + QuantizedTensor, + QUANT_ALGOS, + TensorCoreFP8Layout, + get_layout_class, +) def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_precision_mm=False): @@ -497,14 +510,15 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec ) -> None: super().__init__() - if dtype is None: - dtype = MixedPrecisionOps._compute_dtype - - self.factory_kwargs = {"device": device, "dtype": dtype} + self.factory_kwargs = {"device": device, "dtype": MixedPrecisionOps._compute_dtype} + # self.factory_kwargs = {"device": device, "dtype": dtype} self.in_features = in_features self.out_features = out_features - self._has_bias = bias + if bias: + self.bias = torch.nn.Parameter(torch.empty(out_features, **self.factory_kwargs)) + else: + self.register_parameter("bias", None) self.tensor_class = None self._full_precision_mm = MixedPrecisionOps._full_precision_mm @@ -512,6 +526,16 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec def reset_parameters(self): return None + def _load_scale_param(self, state_dict, prefix, param_name, device, manually_loaded_keys, dtype=None): + key = f"{prefix}{param_name}" + value = state_dict.pop(key, None) + if value is not None: + value = value.to(device=device) + if dtype is not None: + value = value.view(dtype=dtype) + manually_loaded_keys.append(key) + return value + def _load_from_state_dict(self, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs): @@ -529,14 +553,7 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec layer_conf = json.loads(layer_conf.numpy().tobytes()) if layer_conf is None: - dtype = self.factory_kwargs["dtype"] - self.weight = torch.nn.Parameter(weight.to(device=device, dtype=dtype), requires_grad=False) - if dtype != MixedPrecisionOps._compute_dtype: - self.comfy_cast_weights = True - if self._has_bias: - self.bias = torch.nn.Parameter(torch.empty(self.out_features, device=device, dtype=dtype)) - else: - self.register_parameter("bias", None) + self.weight = torch.nn.Parameter(weight.to(device=device, dtype=MixedPrecisionOps._compute_dtype), requires_grad=False) else: self.quant_format = layer_conf.get("format", None) if not self._full_precision_mm: @@ -547,31 +564,46 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec qconfig = QUANT_ALGOS[self.quant_format] self.layout_type = qconfig["comfy_tensor_layout"] + layout_cls = get_layout_class(self.layout_type) - weight_scale_key = f"{prefix}weight_scale" - scale = state_dict.pop(weight_scale_key, None) - if scale is not None: - scale = scale.to(device) - layout_params = { - 'scale': scale, - 'orig_dtype': MixedPrecisionOps._compute_dtype, - 'block_size': qconfig.get("group_size", None), - } + # Load format-specific parameters + if self.quant_format in ["float8_e4m3fn", "float8_e5m2"]: + # FP8: single tensor scale + scale = self._load_scale_param(state_dict, prefix, "weight_scale", device, manually_loaded_keys) - if scale is not None: - manually_loaded_keys.append(weight_scale_key) + params = layout_cls.Params( + scale=scale, + orig_dtype=MixedPrecisionOps._compute_dtype, + orig_shape=(self.out_features, self.in_features), + ) + + elif self.quant_format == "nvfp4": + # NVFP4: tensor_scale (weight_scale_2) + block_scale (weight_scale) + tensor_scale = self._load_scale_param(state_dict, prefix, "weight_scale_2", device, manually_loaded_keys) + block_scale = self._load_scale_param(state_dict, prefix, "weight_scale", device, manually_loaded_keys, + dtype=torch.float8_e4m3fn) + + if tensor_scale is None or block_scale is None: + raise ValueError(f"Missing NVFP4 scales for layer {layer_name}") + + params = layout_cls.Params( + scale=tensor_scale, + block_scale=block_scale, + orig_dtype=MixedPrecisionOps._compute_dtype, + orig_shape=(self.out_features, self.in_features), + ) + else: + raise ValueError(f"Unsupported quantization format: {self.quant_format}") self.weight = torch.nn.Parameter( - QuantizedTensor(weight.to(device=device, dtype=qconfig.get("storage_t", None)), self.layout_type, layout_params), + QuantizedTensor(weight.to(device=device, dtype=qconfig["storage_t"]), self.layout_type, params), requires_grad=False ) - if self._has_bias: - self.bias = torch.nn.Parameter(torch.empty(self.out_features, device=device, dtype=MixedPrecisionOps._compute_dtype)) - else: - self.register_parameter("bias", None) - for param_name in qconfig["parameters"]: + if param_name in {"weight_scale", "weight_scale_2"}: + continue # Already handled above + param_key = f"{prefix}{param_name}" _v = state_dict.pop(param_key, None) if _v is None: @@ -588,7 +620,15 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec def state_dict(self, *args, destination=None, prefix="", **kwargs): sd = super().state_dict(*args, destination=destination, prefix=prefix, **kwargs) if isinstance(self.weight, QuantizedTensor): - sd["{}weight_scale".format(prefix)] = self.weight._layout_params['scale'] + layout_cls = self.weight._layout_cls + + # Check if it's any FP8 variant (E4M3 or E5M2) + if layout_cls in ("TensorCoreFP8E4M3Layout", "TensorCoreFP8E5M2Layout", "TensorCoreFP8Layout"): + sd["{}weight_scale".format(prefix)] = self.weight._params.scale + elif layout_cls == "TensorCoreNVFP4Layout": + sd["{}weight_scale_2".format(prefix)] = self.weight._params.scale + sd["{}weight_scale".format(prefix)] = self.weight._params.block_scale + quant_conf = {"format": self.quant_format} if self._full_precision_mm: quant_conf["full_precision_matrix_mult"] = True @@ -607,12 +647,33 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec def forward(self, input, *args, **kwargs): run_every_op() + input_shape = input.shape + tensor_3d = input.ndim == 3 + if self._full_precision_mm or self.comfy_cast_weights or len(self.weight_function) > 0 or len(self.bias_function) > 0: return self.forward_comfy_cast_weights(input, *args, **kwargs) + if (getattr(self, 'layout_type', None) is not None and not isinstance(input, QuantizedTensor)): - input = QuantizedTensor.from_float(input, self.layout_type, scale=getattr(self, 'input_scale', None), dtype=self.weight.dtype) - return self._forward(input, self.weight, self.bias) + + # Reshape 3D tensors to 2D for quantization (needed for NVFP4 and others) + if tensor_3d: + input = input.reshape(-1, input_shape[2]) + + if input.ndim != 2: + # Fall back to comfy_cast_weights for non-2D tensors + return self.forward_comfy_cast_weights(input.reshape(input_shape), *args, **kwargs) + + # dtype is now implicit in the layout class + input = QuantizedTensor.from_float(input, self.layout_type, scale=getattr(self, 'input_scale', None)) + + output = self._forward(input, self.weight, self.bias) + + # Reshape output back to 3D if input was 3D + if tensor_3d: + output = output.reshape((input_shape[0], input_shape[1], self.weight.shape[0])) + + return output def convert_weight(self, weight, inplace=False, **kwargs): if isinstance(weight, QuantizedTensor): @@ -622,7 +683,8 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec def set_weight(self, weight, inplace_update=False, seed=None, return_weight=False, **kwargs): if getattr(self, 'layout_type', None) is not None: - weight = QuantizedTensor.from_float(weight, self.layout_type, scale="recalculate", dtype=self.weight.dtype, stochastic_rounding=seed, inplace_ops=True) + # dtype is now implicit in the layout class + weight = QuantizedTensor.from_float(weight, self.layout_type, scale="recalculate", stochastic_rounding=seed, inplace_ops=True) else: weight = weight.to(self.weight.dtype) if return_weight: diff --git a/comfy/quant_ops.py b/comfy/quant_ops.py index cd96541d7..cd737726f 100644 --- a/comfy/quant_ops.py +++ b/comfy/quant_ops.py @@ -1,580 +1,133 @@ import torch import logging -from typing import Tuple, Dict + +try: + import comfy_kitchen as ck + from comfy_kitchen.tensor import ( + QuantizedTensor, + QuantizedLayout, + TensorCoreFP8Layout as _CKFp8Layout, + TensorCoreNVFP4Layout, # Direct import, no wrapper needed + register_layout_op, + register_layout_class, + get_layout_class, + ) + _CK_AVAILABLE = True + ck.registry.disable("triton") + for k, v in ck.list_backends().items(): + logging.info(f"Found comfy_kitchen backend {k}: {v}") +except ImportError as e: + logging.error(f"Failed to import comfy_kitchen, Error: {e}, fp8 and fp4 support will not be available.") + _CK_AVAILABLE = False + + class QuantizedTensor: + pass + + class _CKFp8Layout: + pass + + class TensorCoreNVFP4Layout: + pass + + def register_layout_class(name, cls): + pass + + def get_layout_class(name): + return None + import comfy.float -_LAYOUT_REGISTRY = {} -_GENERIC_UTILS = {} - - -def register_layout_op(torch_op, layout_type): - """ - Decorator to register a layout-specific operation handler. - Args: - torch_op: PyTorch operation (e.g., torch.ops.aten.linear.default) - layout_type: Layout class (e.g., TensorCoreFP8Layout) - Example: - @register_layout_op(torch.ops.aten.linear.default, TensorCoreFP8Layout) - def fp8_linear(func, args, kwargs): - # FP8-specific linear implementation - ... - """ - def decorator(handler_func): - if torch_op not in _LAYOUT_REGISTRY: - _LAYOUT_REGISTRY[torch_op] = {} - _LAYOUT_REGISTRY[torch_op][layout_type] = handler_func - return handler_func - return decorator - - -def register_generic_util(torch_op): - """ - Decorator to register a generic utility that works for all layouts. - Args: - torch_op: PyTorch operation (e.g., torch.ops.aten.detach.default) - - Example: - @register_generic_util(torch.ops.aten.detach.default) - def generic_detach(func, args, kwargs): - # Works for any layout - ... - """ - def decorator(handler_func): - _GENERIC_UTILS[torch_op] = handler_func - return handler_func - return decorator - - -def _get_layout_from_args(args): - for arg in args: - if isinstance(arg, QuantizedTensor): - return arg._layout_type - elif isinstance(arg, (list, tuple)): - for item in arg: - if isinstance(item, QuantizedTensor): - return item._layout_type - return None - - -def _move_layout_params_to_device(params, device): - new_params = {} - for k, v in params.items(): - if isinstance(v, torch.Tensor): - new_params[k] = v.to(device=device) - else: - new_params[k] = v - return new_params - - -def _copy_layout_params(params): - new_params = {} - for k, v in params.items(): - if isinstance(v, torch.Tensor): - new_params[k] = v.clone() - else: - new_params[k] = v - return new_params - -def _copy_layout_params_inplace(src, dst, non_blocking=False): - for k, v in src.items(): - if isinstance(v, torch.Tensor): - dst[k].copy_(v, non_blocking=non_blocking) - else: - dst[k] = v - -class QuantizedLayout: - """ - Base class for quantization layouts. - - A layout encapsulates the format-specific logic for quantization/dequantization - and provides a uniform interface for extracting raw tensors needed for computation. - - New quantization formats should subclass this and implement the required methods. - """ - @classmethod - def quantize(cls, tensor, **kwargs) -> Tuple[torch.Tensor, Dict]: - raise NotImplementedError(f"{cls.__name__} must implement quantize()") - - @staticmethod - def dequantize(qdata, **layout_params) -> torch.Tensor: - raise NotImplementedError("TensorLayout must implement dequantize()") - - @classmethod - def get_plain_tensors(cls, qtensor) -> torch.Tensor: - raise NotImplementedError(f"{cls.__name__} must implement get_plain_tensors()") - - -class QuantizedTensor(torch.Tensor): - """ - Universal quantized tensor that works with any layout. - - This tensor subclass uses a pluggable layout system to support multiple - quantization formats (FP8, INT4, INT8, etc.) without code duplication. - - The layout_type determines format-specific behavior, while common operations - (detach, clone, to) are handled generically. - - Attributes: - _qdata: The quantized tensor data - _layout_type: Layout class (e.g., TensorCoreFP8Layout) - _layout_params: Dict with layout-specific params (scale, zero_point, etc.) - """ - - @staticmethod - def __new__(cls, qdata, layout_type, layout_params): - """ - Create a quantized tensor. - - Args: - qdata: The quantized data tensor - layout_type: Layout class (subclass of QuantizedLayout) - layout_params: Dict with layout-specific parameters - """ - return torch.Tensor._make_wrapper_subclass(cls, qdata.shape, device=qdata.device, dtype=qdata.dtype, requires_grad=False) - - def __init__(self, qdata, layout_type, layout_params): - self._qdata = qdata - self._layout_type = layout_type - self._layout_params = layout_params - - def __repr__(self): - layout_name = self._layout_type - param_str = ", ".join(f"{k}={v}" for k, v in list(self._layout_params.items())[:2]) - return f"QuantizedTensor(shape={self.shape}, layout={layout_name}, {param_str})" - - @property - def layout_type(self): - return self._layout_type - - def __tensor_flatten__(self): - """ - Tensor flattening protocol for proper device movement. - """ - inner_tensors = ["_qdata"] - ctx = { - "layout_type": self._layout_type, - } - - tensor_params = {} - non_tensor_params = {} - for k, v in self._layout_params.items(): - if isinstance(v, torch.Tensor): - tensor_params[k] = v - else: - non_tensor_params[k] = v - - ctx["tensor_param_keys"] = list(tensor_params.keys()) - ctx["non_tensor_params"] = non_tensor_params - - for k, v in tensor_params.items(): - attr_name = f"_layout_param_{k}" - object.__setattr__(self, attr_name, v) - inner_tensors.append(attr_name) - - return inner_tensors, ctx - - @staticmethod - def __tensor_unflatten__(inner_tensors, ctx, outer_size, outer_stride): - """ - Tensor unflattening protocol for proper device movement. - Reconstructs the QuantizedTensor after device movement. - """ - layout_type = ctx["layout_type"] - layout_params = dict(ctx["non_tensor_params"]) - - for key in ctx["tensor_param_keys"]: - attr_name = f"_layout_param_{key}" - layout_params[key] = inner_tensors[attr_name] - - return QuantizedTensor(inner_tensors["_qdata"], layout_type, layout_params) - - @classmethod - def from_float(cls, tensor, layout_type, **quantize_kwargs) -> 'QuantizedTensor': - qdata, layout_params = LAYOUTS[layout_type].quantize(tensor, **quantize_kwargs) - return cls(qdata, layout_type, layout_params) - - def dequantize(self) -> torch.Tensor: - return LAYOUTS[self._layout_type].dequantize(self._qdata, **self._layout_params) - - @classmethod - def __torch_dispatch__(cls, func, types, args=(), kwargs=None): - kwargs = kwargs or {} - - # Step 1: Check generic utilities first (detach, clone, to, etc.) - if func in _GENERIC_UTILS: - return _GENERIC_UTILS[func](func, args, kwargs) - - # Step 2: Check layout-specific handlers (linear, matmul, etc.) - layout_type = _get_layout_from_args(args) - if layout_type and func in _LAYOUT_REGISTRY: - handler = _LAYOUT_REGISTRY[func].get(layout_type) - if handler: - return handler(func, args, kwargs) - - # Step 3: Fallback to dequantization - if isinstance(args[0] if args else None, QuantizedTensor): - logging.info(f"QuantizedTensor: Unhandled operation {func}, falling back to dequantization. kwargs={kwargs}") - return cls._dequant_and_fallback(func, args, kwargs) - - @classmethod - def _dequant_and_fallback(cls, func, args, kwargs): - def dequant_arg(arg): - if isinstance(arg, QuantizedTensor): - return arg.dequantize() - elif isinstance(arg, (list, tuple)): - return type(arg)(dequant_arg(a) for a in arg) - return arg - - new_args = dequant_arg(args) - new_kwargs = dequant_arg(kwargs) - return func(*new_args, **new_kwargs) - - def data_ptr(self): - return self._qdata.data_ptr() - - def is_pinned(self): - return self._qdata.is_pinned() - - def is_contiguous(self, *arg, **kwargs): - return self._qdata.is_contiguous(*arg, **kwargs) - - def storage(self): - return self._qdata.storage() - # ============================================================================== -# Generic Utilities (Layout-Agnostic Operations) +# FP8 Layouts with Comfy-Specific Extensions # ============================================================================== -def _create_transformed_qtensor(qt, transform_fn): - new_data = transform_fn(qt._qdata) - new_params = _copy_layout_params(qt._layout_params) - return QuantizedTensor(new_data, qt._layout_type, new_params) +class _TensorCoreFP8LayoutBase(_CKFp8Layout): + FP8_DTYPE = None # Must be overridden in subclass - -def _handle_device_transfer(qt, target_device, target_dtype=None, target_layout=None, op_name="to"): - if target_layout is not None and target_layout != torch.strided: - logging.warning( - f"QuantizedTensor: layout change requested to {target_layout}, " - f"but not supported. Ignoring layout." - ) - - # Handle device transfer - current_device = qt._qdata.device - if target_device is not None: - # Normalize device for comparison - if isinstance(target_device, str): - target_device = torch.device(target_device) - if isinstance(current_device, str): - current_device = torch.device(current_device) - - if target_device != current_device: - logging.debug(f"QuantizedTensor.{op_name}: Moving from {current_device} to {target_device}") - new_q_data = qt._qdata.to(device=target_device) - new_params = _move_layout_params_to_device(qt._layout_params, target_device) - if target_dtype is not None: - new_params["orig_dtype"] = target_dtype - new_qt = QuantizedTensor(new_q_data, qt._layout_type, new_params) - logging.debug(f"QuantizedTensor.{op_name}: Created new tensor on {target_device}") - return new_qt - - logging.debug(f"QuantizedTensor.{op_name}: No device change needed, returning original") - return qt - - -@register_generic_util(torch.ops.aten.detach.default) -def generic_detach(func, args, kwargs): - """Detach operation - creates a detached copy of the quantized tensor.""" - qt = args[0] - if isinstance(qt, QuantizedTensor): - return _create_transformed_qtensor(qt, lambda x: x.detach()) - return func(*args, **kwargs) - - -@register_generic_util(torch.ops.aten.clone.default) -def generic_clone(func, args, kwargs): - """Clone operation - creates a deep copy of the quantized tensor.""" - qt = args[0] - if isinstance(qt, QuantizedTensor): - return _create_transformed_qtensor(qt, lambda x: x.clone()) - return func(*args, **kwargs) - - -@register_generic_util(torch.ops.aten._to_copy.default) -def generic_to_copy(func, args, kwargs): - """Device/dtype transfer operation - handles .to(device) calls.""" - qt = args[0] - if isinstance(qt, QuantizedTensor): - return _handle_device_transfer( - qt, - target_device=kwargs.get('device', None), - target_dtype=kwargs.get('dtype', None), - op_name="_to_copy" - ) - return func(*args, **kwargs) - - -@register_generic_util(torch.ops.aten.to.dtype_layout) -def generic_to_dtype_layout(func, args, kwargs): - """Handle .to(device) calls using the dtype_layout variant.""" - qt = args[0] - if isinstance(qt, QuantizedTensor): - return _handle_device_transfer( - qt, - target_device=kwargs.get('device', None), - target_dtype=kwargs.get('dtype', None), - target_layout=kwargs.get('layout', None), - op_name="to" - ) - return func(*args, **kwargs) - - -@register_generic_util(torch.ops.aten.copy_.default) -def generic_copy_(func, args, kwargs): - qt_dest = args[0] - src = args[1] - non_blocking = args[2] if len(args) > 2 else False - if isinstance(qt_dest, QuantizedTensor): - if isinstance(src, QuantizedTensor): - # Copy from another quantized tensor - qt_dest._qdata.copy_(src._qdata, non_blocking=non_blocking) - qt_dest._layout_type = src._layout_type - orig_dtype = qt_dest._layout_params["orig_dtype"] - _copy_layout_params_inplace(src._layout_params, qt_dest._layout_params, non_blocking=non_blocking) - qt_dest._layout_params["orig_dtype"] = orig_dtype - else: - # Copy from regular tensor - just copy raw data - qt_dest._qdata.copy_(src) - return qt_dest - return func(*args, **kwargs) - - -@register_generic_util(torch.ops.aten.to.dtype) -def generic_to_dtype(func, args, kwargs): - """Handle .to(dtype) calls - dtype conversion only.""" - src = args[0] - if isinstance(src, QuantizedTensor): - # For dtype-only conversion, just change the orig_dtype, no real cast is needed - target_dtype = args[1] if len(args) > 1 else kwargs.get('dtype') - src._layout_params["orig_dtype"] = target_dtype - return src - return func(*args, **kwargs) - - -@register_generic_util(torch.ops.aten._has_compatible_shallow_copy_type.default) -def generic_has_compatible_shallow_copy_type(func, args, kwargs): - return True - - -@register_generic_util(torch.ops.aten.empty_like.default) -def generic_empty_like(func, args, kwargs): - """Empty_like operation - creates an empty tensor with the same quantized structure.""" - qt = args[0] - if isinstance(qt, QuantizedTensor): - # Create empty tensor with same shape and dtype as the quantized data - hp_dtype = kwargs.pop('dtype', qt._layout_params["orig_dtype"]) - new_qdata = torch.empty_like(qt._qdata, **kwargs) - - # Handle device transfer for layout params - target_device = kwargs.get('device', new_qdata.device) - new_params = _move_layout_params_to_device(qt._layout_params, target_device) - - # Update orig_dtype if dtype is specified - new_params['orig_dtype'] = hp_dtype - - return QuantizedTensor(new_qdata, qt._layout_type, new_params) - return func(*args, **kwargs) - -# ============================================================================== -# FP8 Layout + Operation Handlers -# ============================================================================== -class TensorCoreFP8Layout(QuantizedLayout): - """ - Storage format: - - qdata: FP8 tensor (torch.float8_e4m3fn or torch.float8_e5m2) - - scale: Scalar tensor (float32) for dequantization - - orig_dtype: Original dtype before quantization (for casting back) - """ @classmethod - def quantize(cls, tensor, scale=None, dtype=torch.float8_e4m3fn, stochastic_rounding=0, inplace_ops=False): + def quantize(cls, tensor, scale=None, stochastic_rounding=0, inplace_ops=False): + if cls.FP8_DTYPE is None: + raise NotImplementedError(f"{cls.__name__} must define FP8_DTYPE") + orig_dtype = tensor.dtype + orig_shape = tuple(tensor.shape) if isinstance(scale, str) and scale == "recalculate": - scale = torch.amax(tensor.abs()).to(dtype=torch.float32) / torch.finfo(dtype).max + scale = torch.amax(tensor.abs()).to(dtype=torch.float32) / torch.finfo(cls.FP8_DTYPE).max if tensor.dtype not in [torch.float32, torch.bfloat16]: # Prevent scale from being too small tensor_info = torch.finfo(tensor.dtype) scale = (1.0 / torch.clamp((1.0 / scale), min=tensor_info.min, max=tensor_info.max)) - if scale is not None: - if not isinstance(scale, torch.Tensor): - scale = torch.tensor(scale) - scale = scale.to(device=tensor.device, dtype=torch.float32) + if scale is None: + scale = torch.ones((), device=tensor.device, dtype=torch.float32) + if not isinstance(scale, torch.Tensor): + scale = torch.tensor(scale, device=tensor.device, dtype=torch.float32) + if stochastic_rounding > 0: if inplace_ops: tensor *= (1.0 / scale).to(tensor.dtype) else: tensor = tensor * (1.0 / scale).to(tensor.dtype) + qdata = comfy.float.stochastic_rounding(tensor, dtype=cls.FP8_DTYPE, seed=stochastic_rounding) else: - scale = torch.ones((), device=tensor.device, dtype=torch.float32) + qdata = ck.quantize_per_tensor_fp8(tensor, scale, cls.FP8_DTYPE) - if stochastic_rounding > 0: - tensor = comfy.float.stochastic_rounding(tensor, dtype=dtype, seed=stochastic_rounding) - else: - lp_amax = torch.finfo(dtype).max - torch.clamp(tensor, min=-lp_amax, max=lp_amax, out=tensor) - tensor = tensor.to(dtype, memory_format=torch.contiguous_format) + params = cls.Params(scale=scale.float(), orig_dtype=orig_dtype, orig_shape=orig_shape) + return qdata, params - layout_params = { - 'scale': scale, - 'orig_dtype': orig_dtype - } - return tensor, layout_params - @staticmethod - def dequantize(qdata, scale, orig_dtype, **kwargs): - plain_tensor = torch.ops.aten._to_copy.default(qdata, dtype=orig_dtype) - plain_tensor.mul_(scale) - return plain_tensor +class TensorCoreFP8E4M3Layout(_TensorCoreFP8LayoutBase): + FP8_DTYPE = torch.float8_e4m3fn - @classmethod - def get_plain_tensors(cls, qtensor): - return qtensor._qdata, qtensor._layout_params['scale'] + +class TensorCoreFP8E5M2Layout(_TensorCoreFP8LayoutBase): + FP8_DTYPE = torch.float8_e5m2 + + +# Backward compatibility alias - default to E4M3 +TensorCoreFP8Layout = TensorCoreFP8E4M3Layout + + +# ============================================================================== +# Registry +# ============================================================================== + +register_layout_class("TensorCoreFP8Layout", TensorCoreFP8Layout) +register_layout_class("TensorCoreFP8E4M3Layout", TensorCoreFP8E4M3Layout) +register_layout_class("TensorCoreFP8E5M2Layout", TensorCoreFP8E5M2Layout) +register_layout_class("TensorCoreNVFP4Layout", TensorCoreNVFP4Layout) QUANT_ALGOS = { "float8_e4m3fn": { "storage_t": torch.float8_e4m3fn, "parameters": {"weight_scale", "input_scale"}, - "comfy_tensor_layout": "TensorCoreFP8Layout", + "comfy_tensor_layout": "TensorCoreFP8E4M3Layout", + }, + "float8_e5m2": { + "storage_t": torch.float8_e5m2, + "parameters": {"weight_scale", "input_scale"}, + "comfy_tensor_layout": "TensorCoreFP8E5M2Layout", + }, + "nvfp4": { + "storage_t": torch.uint8, + "parameters": {"weight_scale", "weight_scale_2", "input_scale"}, + "comfy_tensor_layout": "TensorCoreNVFP4Layout", + "group_size": 16, }, } -LAYOUTS = { - "TensorCoreFP8Layout": TensorCoreFP8Layout, -} +# ============================================================================== +# Re-exports for backward compatibility +# ============================================================================== -@register_layout_op(torch.ops.aten.linear.default, "TensorCoreFP8Layout") -def fp8_linear(func, args, kwargs): - input_tensor = args[0] - weight = args[1] - bias = args[2] if len(args) > 2 else None - - if isinstance(input_tensor, QuantizedTensor) and isinstance(weight, QuantizedTensor): - plain_input, scale_a = TensorCoreFP8Layout.get_plain_tensors(input_tensor) - plain_weight, scale_b = TensorCoreFP8Layout.get_plain_tensors(weight) - - out_dtype = kwargs.get("out_dtype") - if out_dtype is None: - out_dtype = input_tensor._layout_params['orig_dtype'] - - weight_t = plain_weight.t() - - tensor_2d = False - if len(plain_input.shape) == 2: - tensor_2d = True - plain_input = plain_input.unsqueeze(1) - - input_shape = plain_input.shape - if len(input_shape) != 3: - return None - - try: - output = torch._scaled_mm( - plain_input.reshape(-1, input_shape[2]).contiguous(), - weight_t, - bias=bias, - scale_a=scale_a, - scale_b=scale_b, - out_dtype=out_dtype, - ) - - if isinstance(output, tuple): # TODO: remove when we drop support for torch 2.4 - output = output[0] - - if not tensor_2d: - output = output.reshape((-1, input_shape[1], weight.shape[0])) - - if output.dtype in [torch.float8_e4m3fn, torch.float8_e5m2]: - output_scale = scale_a * scale_b - output_params = { - 'scale': output_scale, - 'orig_dtype': input_tensor._layout_params['orig_dtype'] - } - return QuantizedTensor(output, "TensorCoreFP8Layout", output_params) - else: - return output - - except Exception as e: - raise RuntimeError(f"FP8 _scaled_mm failed, falling back to dequantization: {e}") - - # Case 2: DQ Fallback - if isinstance(weight, QuantizedTensor): - weight = weight.dequantize() - if isinstance(input_tensor, QuantizedTensor): - input_tensor = input_tensor.dequantize() - - return torch.nn.functional.linear(input_tensor, weight, bias) - -def fp8_mm_(input_tensor, weight, bias=None, out_dtype=None): - if out_dtype is None: - out_dtype = input_tensor._layout_params['orig_dtype'] - - plain_input, scale_a = TensorCoreFP8Layout.get_plain_tensors(input_tensor) - plain_weight, scale_b = TensorCoreFP8Layout.get_plain_tensors(weight) - - output = torch._scaled_mm( - plain_input.contiguous(), - plain_weight, - bias=bias, - scale_a=scale_a, - scale_b=scale_b, - out_dtype=out_dtype, - ) - - if isinstance(output, tuple): # TODO: remove when we drop support for torch 2.4 - output = output[0] - return output - -@register_layout_op(torch.ops.aten.addmm.default, "TensorCoreFP8Layout") -def fp8_addmm(func, args, kwargs): - input_tensor = args[1] - weight = args[2] - bias = args[0] - - if isinstance(input_tensor, QuantizedTensor) and isinstance(weight, QuantizedTensor): - return fp8_mm_(input_tensor, weight, bias=bias, out_dtype=kwargs.get("out_dtype", None)) - - a = list(args) - if isinstance(args[0], QuantizedTensor): - a[0] = args[0].dequantize() - if isinstance(args[1], QuantizedTensor): - a[1] = args[1].dequantize() - if isinstance(args[2], QuantizedTensor): - a[2] = args[2].dequantize() - - return func(*a, **kwargs) - -@register_layout_op(torch.ops.aten.mm.default, "TensorCoreFP8Layout") -def fp8_mm(func, args, kwargs): - input_tensor = args[0] - weight = args[1] - - if isinstance(input_tensor, QuantizedTensor) and isinstance(weight, QuantizedTensor): - return fp8_mm_(input_tensor, weight, bias=None, out_dtype=kwargs.get("out_dtype", None)) - - a = list(args) - if isinstance(args[0], QuantizedTensor): - a[0] = args[0].dequantize() - if isinstance(args[1], QuantizedTensor): - a[1] = args[1].dequantize() - return func(*a, **kwargs) - -@register_layout_op(torch.ops.aten.view.default, "TensorCoreFP8Layout") -@register_layout_op(torch.ops.aten.t.default, "TensorCoreFP8Layout") -def fp8_func(func, args, kwargs): - input_tensor = args[0] - if isinstance(input_tensor, QuantizedTensor): - plain_input, scale_a = TensorCoreFP8Layout.get_plain_tensors(input_tensor) - ar = list(args) - ar[0] = plain_input - return QuantizedTensor(func(*ar, **kwargs), "TensorCoreFP8Layout", input_tensor._layout_params) - return func(*args, **kwargs) +__all__ = [ + "QuantizedTensor", + "QuantizedLayout", + "TensorCoreFP8Layout", + "TensorCoreFP8E4M3Layout", + "TensorCoreFP8E5M2Layout", + "TensorCoreNVFP4Layout", + "QUANT_ALGOS", + "register_layout_op", +] diff --git a/requirements.txt b/requirements.txt index 3a05799eb..0ee152032 100644 --- a/requirements.txt +++ b/requirements.txt @@ -21,6 +21,7 @@ psutil alembic SQLAlchemy av>=14.2.0 +comfy-kitchen>=0.2.0 #non essential dependencies: kornia>=0.7.1 diff --git a/tests-unit/comfy_quant/test_mixed_precision.py b/tests-unit/comfy_quant/test_mixed_precision.py index 3a54941e6..7b2eac940 100644 --- a/tests-unit/comfy_quant/test_mixed_precision.py +++ b/tests-unit/comfy_quant/test_mixed_precision.py @@ -103,18 +103,18 @@ class TestMixedPrecisionOps(unittest.TestCase): # Verify weights are wrapped in QuantizedTensor self.assertIsInstance(model.layer1.weight, QuantizedTensor) - self.assertEqual(model.layer1.weight._layout_type, "TensorCoreFP8Layout") + self.assertEqual(model.layer1.weight._layout_cls, "TensorCoreFP8E4M3Layout") # Layer 2 should NOT be quantized self.assertNotIsInstance(model.layer2.weight, QuantizedTensor) # Layer 3 should be quantized self.assertIsInstance(model.layer3.weight, QuantizedTensor) - self.assertEqual(model.layer3.weight._layout_type, "TensorCoreFP8Layout") + self.assertEqual(model.layer3.weight._layout_cls, "TensorCoreFP8E4M3Layout") # Verify scales were loaded - self.assertEqual(model.layer1.weight._layout_params['scale'].item(), 2.0) - self.assertEqual(model.layer3.weight._layout_params['scale'].item(), 1.5) + self.assertEqual(model.layer1.weight._params.scale.item(), 2.0) + self.assertEqual(model.layer3.weight._params.scale.item(), 1.5) # Forward pass input_tensor = torch.randn(5, 10, dtype=torch.bfloat16) @@ -154,8 +154,8 @@ class TestMixedPrecisionOps(unittest.TestCase): # Verify layer1.weight is a QuantizedTensor with scale preserved self.assertIsInstance(state_dict2["layer1.weight"], QuantizedTensor) - self.assertEqual(state_dict2["layer1.weight"]._layout_params['scale'].item(), 3.0) - self.assertEqual(state_dict2["layer1.weight"]._layout_type, "TensorCoreFP8Layout") + self.assertEqual(state_dict2["layer1.weight"]._params.scale.item(), 3.0) + self.assertEqual(state_dict2["layer1.weight"]._layout_cls, "TensorCoreFP8E4M3Layout") # Verify non-quantized layers are standard tensors self.assertNotIsInstance(state_dict2["layer2.weight"], QuantizedTensor) diff --git a/tests-unit/comfy_quant/test_quant_registry.py b/tests-unit/comfy_quant/test_quant_registry.py deleted file mode 100644 index 9cb54ede8..000000000 --- a/tests-unit/comfy_quant/test_quant_registry.py +++ /dev/null @@ -1,190 +0,0 @@ -import unittest -import torch -import sys -import os - -# Add comfy to path -sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..", "..")) - -def has_gpu(): - return torch.cuda.is_available() - -from comfy.cli_args import args -if not has_gpu(): - args.cpu = True - -from comfy.quant_ops import QuantizedTensor, TensorCoreFP8Layout - - -class TestQuantizedTensor(unittest.TestCase): - """Test the QuantizedTensor subclass with FP8 layout""" - - def test_creation(self): - """Test creating a QuantizedTensor with TensorCoreFP8Layout""" - fp8_data = torch.randn(256, 128, dtype=torch.float32).to(torch.float8_e4m3fn) - scale = torch.tensor(2.0) - layout_params = {'scale': scale, 'orig_dtype': torch.bfloat16} - - qt = QuantizedTensor(fp8_data, "TensorCoreFP8Layout", layout_params) - - self.assertIsInstance(qt, QuantizedTensor) - self.assertEqual(qt.shape, (256, 128)) - self.assertEqual(qt.dtype, torch.float8_e4m3fn) - self.assertEqual(qt._layout_params['scale'], scale) - self.assertEqual(qt._layout_params['orig_dtype'], torch.bfloat16) - self.assertEqual(qt._layout_type, "TensorCoreFP8Layout") - - def test_dequantize(self): - """Test explicit dequantization""" - - fp8_data = torch.ones(10, 20, dtype=torch.float32).to(torch.float8_e4m3fn) - scale = torch.tensor(3.0) - layout_params = {'scale': scale, 'orig_dtype': torch.float32} - - qt = QuantizedTensor(fp8_data, "TensorCoreFP8Layout", layout_params) - dequantized = qt.dequantize() - - self.assertEqual(dequantized.dtype, torch.float32) - self.assertTrue(torch.allclose(dequantized, torch.ones(10, 20) * 3.0, rtol=0.1)) - - def test_from_float(self): - """Test creating QuantizedTensor from float tensor""" - float_tensor = torch.randn(64, 32, dtype=torch.float32) - scale = torch.tensor(1.5) - - qt = QuantizedTensor.from_float( - float_tensor, - "TensorCoreFP8Layout", - scale=scale, - dtype=torch.float8_e4m3fn - ) - - self.assertIsInstance(qt, QuantizedTensor) - self.assertEqual(qt.dtype, torch.float8_e4m3fn) - self.assertEqual(qt.shape, (64, 32)) - - # Verify dequantization gives approximately original values - dequantized = qt.dequantize() - mean_rel_error = ((dequantized - float_tensor).abs() / (float_tensor.abs() + 1e-6)).mean() - self.assertLess(mean_rel_error, 0.1) - - -class TestGenericUtilities(unittest.TestCase): - """Test generic utility operations""" - - def test_detach(self): - """Test detach operation on quantized tensor""" - fp8_data = torch.randn(10, 20, dtype=torch.float32).to(torch.float8_e4m3fn) - scale = torch.tensor(1.5) - layout_params = {'scale': scale, 'orig_dtype': torch.float32} - qt = QuantizedTensor(fp8_data, "TensorCoreFP8Layout", layout_params) - - # Detach should return a new QuantizedTensor - qt_detached = qt.detach() - - self.assertIsInstance(qt_detached, QuantizedTensor) - self.assertEqual(qt_detached.shape, qt.shape) - self.assertEqual(qt_detached._layout_type, "TensorCoreFP8Layout") - - def test_clone(self): - """Test clone operation on quantized tensor""" - fp8_data = torch.randn(10, 20, dtype=torch.float32).to(torch.float8_e4m3fn) - scale = torch.tensor(1.5) - layout_params = {'scale': scale, 'orig_dtype': torch.float32} - qt = QuantizedTensor(fp8_data, "TensorCoreFP8Layout", layout_params) - - # Clone should return a new QuantizedTensor - qt_cloned = qt.clone() - - self.assertIsInstance(qt_cloned, QuantizedTensor) - self.assertEqual(qt_cloned.shape, qt.shape) - self.assertEqual(qt_cloned._layout_type, "TensorCoreFP8Layout") - - # Verify it's a deep copy - self.assertIsNot(qt_cloned._qdata, qt._qdata) - - @unittest.skipUnless(has_gpu(), "GPU not available") - def test_to_device(self): - """Test device transfer""" - fp8_data = torch.randn(10, 20, dtype=torch.float32).to(torch.float8_e4m3fn) - scale = torch.tensor(1.5) - layout_params = {'scale': scale, 'orig_dtype': torch.float32} - qt = QuantizedTensor(fp8_data, "TensorCoreFP8Layout", layout_params) - - # Moving to same device should work (CPU to CPU) - qt_cpu = qt.to('cpu') - - self.assertIsInstance(qt_cpu, QuantizedTensor) - self.assertEqual(qt_cpu.device.type, 'cpu') - self.assertEqual(qt_cpu._layout_params['scale'].device.type, 'cpu') - - -class TestTensorCoreFP8Layout(unittest.TestCase): - """Test the TensorCoreFP8Layout implementation""" - - def test_quantize(self): - """Test quantization method""" - float_tensor = torch.randn(32, 64, dtype=torch.float32) - scale = torch.tensor(1.5) - - qdata, layout_params = TensorCoreFP8Layout.quantize( - float_tensor, - scale=scale, - dtype=torch.float8_e4m3fn - ) - - self.assertEqual(qdata.dtype, torch.float8_e4m3fn) - self.assertEqual(qdata.shape, float_tensor.shape) - self.assertIn('scale', layout_params) - self.assertIn('orig_dtype', layout_params) - self.assertEqual(layout_params['orig_dtype'], torch.float32) - - def test_dequantize(self): - """Test dequantization method""" - float_tensor = torch.ones(10, 20, dtype=torch.float32) * 3.0 - scale = torch.tensor(1.0) - - qdata, layout_params = TensorCoreFP8Layout.quantize( - float_tensor, - scale=scale, - dtype=torch.float8_e4m3fn - ) - - dequantized = TensorCoreFP8Layout.dequantize(qdata, **layout_params) - - # Should approximately match original - self.assertTrue(torch.allclose(dequantized, float_tensor, rtol=0.1, atol=0.1)) - - -class TestFallbackMechanism(unittest.TestCase): - """Test fallback for unsupported operations""" - - def test_unsupported_op_dequantizes(self): - """Test that unsupported operations fall back to dequantization""" - # Set seed for reproducibility - torch.manual_seed(42) - - # Create quantized tensor - a_fp32 = torch.randn(10, 20, dtype=torch.float32) - scale = torch.tensor(1.0) - a_q = QuantizedTensor.from_float( - a_fp32, - "TensorCoreFP8Layout", - scale=scale, - dtype=torch.float8_e4m3fn - ) - - # Call an operation that doesn't have a registered handler - # For example, torch.abs - result = torch.abs(a_q) - - # Should work via fallback (dequantize → abs → return) - self.assertNotIsInstance(result, QuantizedTensor) - expected = torch.abs(a_fp32) - # FP8 introduces quantization error, so use loose tolerance - mean_error = (result - expected).abs().mean() - self.assertLess(mean_error, 0.05, f"Mean error {mean_error:.4f} is too large") - - -if __name__ == "__main__": - unittest.main() From 6ef85c49151cf8c4d6bf5e7ccfc566b8d0681cbd Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 5 Jan 2026 19:50:35 -0800 Subject: [PATCH 07/48] Use rope functions from comfy kitchen. (#11647) --- comfy/ldm/flux/math.py | 23 +++++++++++++++-------- 1 file changed, 15 insertions(+), 8 deletions(-) diff --git a/comfy/ldm/flux/math.py b/comfy/ldm/flux/math.py index 6a22df8bc..f9597de5b 100644 --- a/comfy/ldm/flux/math.py +++ b/comfy/ldm/flux/math.py @@ -4,6 +4,7 @@ from torch import Tensor from comfy.ldm.modules.attention import optimized_attention import comfy.model_management +import logging def attention(q: Tensor, k: Tensor, v: Tensor, pe: Tensor, mask=None, transformer_options={}) -> Tensor: @@ -13,7 +14,6 @@ def attention(q: Tensor, k: Tensor, v: Tensor, pe: Tensor, mask=None, transforme x = optimized_attention(q, k, v, heads, skip_reshape=True, mask=mask, transformer_options=transformer_options) return x - def rope(pos: Tensor, dim: int, theta: int) -> Tensor: assert dim % 2 == 0 if comfy.model_management.is_device_mps(pos.device) or comfy.model_management.is_intel_xpu() or comfy.model_management.is_directml_enabled(): @@ -28,13 +28,20 @@ def rope(pos: Tensor, dim: int, theta: int) -> Tensor: out = rearrange(out, "b n d (i j) -> b n d i j", i=2, j=2) return out.to(dtype=torch.float32, device=pos.device) -def apply_rope1(x: Tensor, freqs_cis: Tensor): - x_ = x.to(dtype=freqs_cis.dtype).reshape(*x.shape[:-1], -1, 1, 2) - x_out = freqs_cis[..., 0] * x_[..., 0] - x_out.addcmul_(freqs_cis[..., 1], x_[..., 1]) +try: + import comfy.quant_ops + apply_rope = comfy.quant_ops.ck.apply_rope + apply_rope1 = comfy.quant_ops.ck.apply_rope1 +except: + logging.warning("No comfy kitchen, using old apply_rope functions.") + def apply_rope1(x: Tensor, freqs_cis: Tensor): + x_ = x.to(dtype=freqs_cis.dtype).reshape(*x.shape[:-1], -1, 1, 2) - return x_out.reshape(*x.shape).type_as(x) + x_out = freqs_cis[..., 0] * x_[..., 0] + x_out.addcmul_(freqs_cis[..., 1], x_[..., 1]) -def apply_rope(xq: Tensor, xk: Tensor, freqs_cis: Tensor): - return apply_rope1(xq, freqs_cis), apply_rope1(xk, freqs_cis) + return x_out.reshape(*x.shape).type_as(x) + + def apply_rope(xq: Tensor, xk: Tensor, freqs_cis: Tensor): + return apply_rope1(xq, freqs_cis), apply_rope1(xk, freqs_cis) From 161800241117fae7af90e0c938d0cf8cb2f2ddb1 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 5 Jan 2026 20:07:39 -0800 Subject: [PATCH 08/48] Revert "Use rope functions from comfy kitchen. (#11647)" (#11648) This reverts commit 6ef85c49151cf8c4d6bf5e7ccfc566b8d0681cbd. --- comfy/ldm/flux/math.py | 23 ++++++++--------------- 1 file changed, 8 insertions(+), 15 deletions(-) diff --git a/comfy/ldm/flux/math.py b/comfy/ldm/flux/math.py index f9597de5b..6a22df8bc 100644 --- a/comfy/ldm/flux/math.py +++ b/comfy/ldm/flux/math.py @@ -4,7 +4,6 @@ from torch import Tensor from comfy.ldm.modules.attention import optimized_attention import comfy.model_management -import logging def attention(q: Tensor, k: Tensor, v: Tensor, pe: Tensor, mask=None, transformer_options={}) -> Tensor: @@ -14,6 +13,7 @@ def attention(q: Tensor, k: Tensor, v: Tensor, pe: Tensor, mask=None, transforme x = optimized_attention(q, k, v, heads, skip_reshape=True, mask=mask, transformer_options=transformer_options) return x + def rope(pos: Tensor, dim: int, theta: int) -> Tensor: assert dim % 2 == 0 if comfy.model_management.is_device_mps(pos.device) or comfy.model_management.is_intel_xpu() or comfy.model_management.is_directml_enabled(): @@ -28,20 +28,13 @@ def rope(pos: Tensor, dim: int, theta: int) -> Tensor: out = rearrange(out, "b n d (i j) -> b n d i j", i=2, j=2) return out.to(dtype=torch.float32, device=pos.device) +def apply_rope1(x: Tensor, freqs_cis: Tensor): + x_ = x.to(dtype=freqs_cis.dtype).reshape(*x.shape[:-1], -1, 1, 2) -try: - import comfy.quant_ops - apply_rope = comfy.quant_ops.ck.apply_rope - apply_rope1 = comfy.quant_ops.ck.apply_rope1 -except: - logging.warning("No comfy kitchen, using old apply_rope functions.") - def apply_rope1(x: Tensor, freqs_cis: Tensor): - x_ = x.to(dtype=freqs_cis.dtype).reshape(*x.shape[:-1], -1, 1, 2) + x_out = freqs_cis[..., 0] * x_[..., 0] + x_out.addcmul_(freqs_cis[..., 1], x_[..., 1]) - x_out = freqs_cis[..., 0] * x_[..., 0] - x_out.addcmul_(freqs_cis[..., 1], x_[..., 1]) + return x_out.reshape(*x.shape).type_as(x) - return x_out.reshape(*x.shape).type_as(x) - - def apply_rope(xq: Tensor, xk: Tensor, freqs_cis: Tensor): - return apply_rope1(xq, freqs_cis), apply_rope1(xk, freqs_cis) +def apply_rope(xq: Tensor, xk: Tensor, freqs_cis: Tensor): + return apply_rope1(xq, freqs_cis), apply_rope1(xk, freqs_cis) From e14f3b661069971163ddc56036b0f486933b9162 Mon Sep 17 00:00:00 2001 From: ComfyUI Wiki Date: Tue, 6 Jan 2026 14:37:11 +0800 Subject: [PATCH 09/48] chore: update workflow templates to v0.7.66 (#11652) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 0ee152032..9c9c0e29e 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,5 +1,5 @@ comfyui-frontend-package==1.35.9 -comfyui-workflow-templates==0.7.65 +comfyui-workflow-templates==0.7.66 comfyui-embedded-docs==0.3.1 torch torchsde From 96e0d0924e027248733bc6e0b8102dcdc8acde33 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 6 Jan 2026 11:43:24 -0800 Subject: [PATCH 10/48] Add helpful message to portable. (#11671) --- .../advanced/run_nvidia_gpu_disable_api_nodes.bat | 2 +- .ci/windows_nvidia_base_files/run_nvidia_gpu.bat | 2 +- .../run_nvidia_gpu_fast_fp16_accumulation.bat | 2 +- 3 files changed, 3 insertions(+), 3 deletions(-) diff --git a/.ci/windows_nvidia_base_files/advanced/run_nvidia_gpu_disable_api_nodes.bat b/.ci/windows_nvidia_base_files/advanced/run_nvidia_gpu_disable_api_nodes.bat index ed00583b6..4501ef9a1 100644 --- a/.ci/windows_nvidia_base_files/advanced/run_nvidia_gpu_disable_api_nodes.bat +++ b/.ci/windows_nvidia_base_files/advanced/run_nvidia_gpu_disable_api_nodes.bat @@ -1,3 +1,3 @@ ..\python_embeded\python.exe -s ..\ComfyUI\main.py --windows-standalone-build --disable-api-nodes -echo If you see this and ComfyUI did not start try updating your Nvidia Drivers to the latest. +echo If you see this and ComfyUI did not start try updating your Nvidia Drivers to the latest. If you get a c10.dll error you need to install vc redist that you can find: https://aka.ms/vc14/vc_redist.x64.exe pause diff --git a/.ci/windows_nvidia_base_files/run_nvidia_gpu.bat b/.ci/windows_nvidia_base_files/run_nvidia_gpu.bat index 4898a424f..6487ac7ce 100755 --- a/.ci/windows_nvidia_base_files/run_nvidia_gpu.bat +++ b/.ci/windows_nvidia_base_files/run_nvidia_gpu.bat @@ -1,3 +1,3 @@ .\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build -echo If you see this and ComfyUI did not start try updating your Nvidia Drivers to the latest. +echo If you see this and ComfyUI did not start try updating your Nvidia Drivers to the latest. If you get a c10.dll error you need to install vc redist that you can find: https://aka.ms/vc14/vc_redist.x64.exe pause diff --git a/.ci/windows_nvidia_base_files/run_nvidia_gpu_fast_fp16_accumulation.bat b/.ci/windows_nvidia_base_files/run_nvidia_gpu_fast_fp16_accumulation.bat index 32611e4af..01c5bb33b 100644 --- a/.ci/windows_nvidia_base_files/run_nvidia_gpu_fast_fp16_accumulation.bat +++ b/.ci/windows_nvidia_base_files/run_nvidia_gpu_fast_fp16_accumulation.bat @@ -1,3 +1,3 @@ .\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build --fast fp16_accumulation -echo If you see this and ComfyUI did not start try updating your Nvidia Drivers to the latest. +echo If you see this and ComfyUI did not start try updating your Nvidia Drivers to the latest. If you get a c10.dll error you need to install vc redist that you can find: https://aka.ms/vc14/vc_redist.x64.exe pause From 6ffc159bdd56d1ad73e954081def6a7f163e7a7f Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 6 Jan 2026 12:53:43 -0800 Subject: [PATCH 11/48] Update comfy-kitchen version to 0.2.1 (#11672) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 9c9c0e29e..22cb50e2d 100644 --- a/requirements.txt +++ b/requirements.txt @@ -21,7 +21,7 @@ psutil alembic SQLAlchemy av>=14.2.0 -comfy-kitchen>=0.2.0 +comfy-kitchen>=0.2.1 #non essential dependencies: kornia>=0.7.1 From c3c3e93c5bb3034175c17ef8beeb8fe8626c66ab Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 6 Jan 2026 13:57:50 -0800 Subject: [PATCH 12/48] Use rope functions from comfy kitchen. (#11674) --- comfy/ldm/flux/math.py | 23 +++++++++++++++-------- requirements.txt | 2 +- 2 files changed, 16 insertions(+), 9 deletions(-) diff --git a/comfy/ldm/flux/math.py b/comfy/ldm/flux/math.py index 6a22df8bc..f9597de5b 100644 --- a/comfy/ldm/flux/math.py +++ b/comfy/ldm/flux/math.py @@ -4,6 +4,7 @@ from torch import Tensor from comfy.ldm.modules.attention import optimized_attention import comfy.model_management +import logging def attention(q: Tensor, k: Tensor, v: Tensor, pe: Tensor, mask=None, transformer_options={}) -> Tensor: @@ -13,7 +14,6 @@ def attention(q: Tensor, k: Tensor, v: Tensor, pe: Tensor, mask=None, transforme x = optimized_attention(q, k, v, heads, skip_reshape=True, mask=mask, transformer_options=transformer_options) return x - def rope(pos: Tensor, dim: int, theta: int) -> Tensor: assert dim % 2 == 0 if comfy.model_management.is_device_mps(pos.device) or comfy.model_management.is_intel_xpu() or comfy.model_management.is_directml_enabled(): @@ -28,13 +28,20 @@ def rope(pos: Tensor, dim: int, theta: int) -> Tensor: out = rearrange(out, "b n d (i j) -> b n d i j", i=2, j=2) return out.to(dtype=torch.float32, device=pos.device) -def apply_rope1(x: Tensor, freqs_cis: Tensor): - x_ = x.to(dtype=freqs_cis.dtype).reshape(*x.shape[:-1], -1, 1, 2) - x_out = freqs_cis[..., 0] * x_[..., 0] - x_out.addcmul_(freqs_cis[..., 1], x_[..., 1]) +try: + import comfy.quant_ops + apply_rope = comfy.quant_ops.ck.apply_rope + apply_rope1 = comfy.quant_ops.ck.apply_rope1 +except: + logging.warning("No comfy kitchen, using old apply_rope functions.") + def apply_rope1(x: Tensor, freqs_cis: Tensor): + x_ = x.to(dtype=freqs_cis.dtype).reshape(*x.shape[:-1], -1, 1, 2) - return x_out.reshape(*x.shape).type_as(x) + x_out = freqs_cis[..., 0] * x_[..., 0] + x_out.addcmul_(freqs_cis[..., 1], x_[..., 1]) -def apply_rope(xq: Tensor, xk: Tensor, freqs_cis: Tensor): - return apply_rope1(xq, freqs_cis), apply_rope1(xk, freqs_cis) + return x_out.reshape(*x.shape).type_as(x) + + def apply_rope(xq: Tensor, xk: Tensor, freqs_cis: Tensor): + return apply_rope1(xq, freqs_cis), apply_rope1(xk, freqs_cis) diff --git a/requirements.txt b/requirements.txt index 22cb50e2d..7798cb179 100644 --- a/requirements.txt +++ b/requirements.txt @@ -21,7 +21,7 @@ psutil alembic SQLAlchemy av>=14.2.0 -comfy-kitchen>=0.2.1 +comfy-kitchen>=0.2.2 #non essential dependencies: kornia>=0.7.1 From c3566c0d765200068d26d0888f035504a50012f2 Mon Sep 17 00:00:00 2001 From: ComfyUI Wiki Date: Wed, 7 Jan 2026 06:28:29 +0800 Subject: [PATCH 13/48] chore: update workflow templates to v0.7.67 (#11667) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 7798cb179..caad0026a 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,5 +1,5 @@ comfyui-frontend-package==1.35.9 -comfyui-workflow-templates==0.7.66 +comfyui-workflow-templates==0.7.67 comfyui-embedded-docs==0.3.1 torch torchsde From 023cf13721cac256c323e2226319b766d07b1f36 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 6 Jan 2026 14:33:03 -0800 Subject: [PATCH 14/48] Fix lowvram issue with ltxv2 text encoder. (#11675) --- comfy/ldm/lightricks/embeddings_connector.py | 2 +- comfy/text_encoders/lt.py | 4 +++- 2 files changed, 4 insertions(+), 2 deletions(-) diff --git a/comfy/ldm/lightricks/embeddings_connector.py b/comfy/ldm/lightricks/embeddings_connector.py index f7a43f3c3..06f5ada89 100644 --- a/comfy/ldm/lightricks/embeddings_connector.py +++ b/comfy/ldm/lightricks/embeddings_connector.py @@ -276,7 +276,7 @@ class Embeddings1DConnector(nn.Module): max(1024, hidden_states.shape[1]) / self.num_learnable_registers ) learnable_registers = torch.tile( - self.learnable_registers, (num_registers_duplications, 1) + self.learnable_registers.to(hidden_states), (num_registers_duplications, 1) ) hidden_states = torch.cat((hidden_states, learnable_registers[hidden_states.shape[1]:].unsqueeze(0).repeat(hidden_states.shape[0], 1, 1)), dim=1) diff --git a/comfy/text_encoders/lt.py b/comfy/text_encoders/lt.py index 2c2d453e8..e5964e42b 100644 --- a/comfy/text_encoders/lt.py +++ b/comfy/text_encoders/lt.py @@ -86,17 +86,19 @@ class LTXAVTEModel(torch.nn.Module): ) def set_clip_options(self, options): + self.execution_device = options.get("execution_device", self.execution_device) self.gemma3_12b.set_clip_options(options) def reset_clip_options(self): self.gemma3_12b.reset_clip_options() + self.execution_device = None def encode_token_weights(self, token_weight_pairs): token_weight_pairs = token_weight_pairs["gemma3_12b"] out, pooled, extra = self.gemma3_12b.encode_token_weights(token_weight_pairs) out_device = out.device - out = out.movedim(1, -1).to(self.text_embedding_projection.weight.device) + out = out.movedim(1, -1).to(self.execution_device) out = 8.0 * (out - out.mean(dim=(1, 2), keepdim=True)) / (out.amax(dim=(1, 2), keepdim=True) - out.amin(dim=(1, 2), keepdim=True) + 1e-6) out = out.reshape((out.shape[0], out.shape[1], -1)) out = self.text_embedding_projection(out) From 6e9ee55cdd9e0eca6b5144063575b983f3311762 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 6 Jan 2026 14:41:27 -0800 Subject: [PATCH 15/48] Disable ltxav previews. (#11676) --- comfy/latent_formats.py | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/comfy/latent_formats.py b/comfy/latent_formats.py index 9bbe30b53..cb4f52ce1 100644 --- a/comfy/latent_formats.py +++ b/comfy/latent_formats.py @@ -408,7 +408,9 @@ class LTXV(LatentFormat): self.latent_rgb_factors_bias = [-0.0571, -0.1657, -0.2512] class LTXAV(LTXV): - pass + def __init__(self): + self.latent_rgb_factors = None + self.latent_rgb_factors_bias = None class HunyuanVideo(LatentFormat): latent_channels = 16 From 2c03884f5fb7fa213161dfe1e9a09a8e8c4b6062 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 6 Jan 2026 15:07:26 -0800 Subject: [PATCH 16/48] Skip fp4 matrix mult on devices that don't support it. (#11677) --- comfy/model_management.py | 10 ++++++++++ comfy/ops.py | 21 +++++++++++++++++---- 2 files changed, 27 insertions(+), 4 deletions(-) diff --git a/comfy/model_management.py b/comfy/model_management.py index 22f4de044..928282092 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -1504,6 +1504,16 @@ def supports_fp8_compute(device=None): return True +def supports_nvfp4_compute(device=None): + if not is_nvidia(): + return False + + props = torch.cuda.get_device_properties(device) + if props.major < 10: + return False + + return True + def extended_fp16_support(): # TODO: check why some models work with fp16 on newer torch versions but not on older if torch_version_numeric < (2, 7): diff --git a/comfy/ops.py b/comfy/ops.py index f5e1e9230..8f9fdce36 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -493,11 +493,12 @@ from .quant_ops import ( ) -def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_precision_mm=False): +def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_precision_mm=False, disabled=[]): class MixedPrecisionOps(manual_cast): _quant_config = quant_config _compute_dtype = compute_dtype _full_precision_mm = full_precision_mm + _disabled = disabled class Linear(torch.nn.Module, CastWeightBiasOp): def __init__( @@ -522,6 +523,7 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec self.tensor_class = None self._full_precision_mm = MixedPrecisionOps._full_precision_mm + self._full_precision_mm_config = False def reset_parameters(self): return None @@ -556,8 +558,12 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec self.weight = torch.nn.Parameter(weight.to(device=device, dtype=MixedPrecisionOps._compute_dtype), requires_grad=False) else: self.quant_format = layer_conf.get("format", None) + self._full_precision_mm_config = layer_conf.get("full_precision_matrix_mult", False) if not self._full_precision_mm: - self._full_precision_mm = layer_conf.get("full_precision_matrix_mult", False) + self._full_precision_mm = self._full_precision_mm_config + + if self.quant_format in MixedPrecisionOps._disabled: + self._full_precision_mm = True if self.quant_format is None: raise ValueError(f"Unknown quantization format for layer {layer_name}") @@ -630,7 +636,7 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec sd["{}weight_scale".format(prefix)] = self.weight._params.block_scale quant_conf = {"format": self.quant_format} - if self._full_precision_mm: + if self._full_precision_mm_config: quant_conf["full_precision_matrix_mult"] = True sd["{}comfy_quant".format(prefix)] = torch.tensor(list(json.dumps(quant_conf).encode('utf-8')), dtype=torch.uint8) return sd @@ -711,10 +717,17 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec def pick_operations(weight_dtype, compute_dtype, load_device=None, disable_fast_fp8=False, fp8_optimizations=False, model_config=None): fp8_compute = comfy.model_management.supports_fp8_compute(load_device) # TODO: if we support more ops this needs to be more granular + nvfp4_compute = comfy.model_management.supports_nvfp4_compute(load_device) if model_config and hasattr(model_config, 'quant_config') and model_config.quant_config: logging.info("Using mixed precision operations") - return mixed_precision_ops(model_config.quant_config, compute_dtype, full_precision_mm=not fp8_compute) + disabled = set() + if not nvfp4_compute: + disabled.add("nvfp4") + if not fp8_compute: + disabled.add("float8_e4m3fn") + disabled.add("float8_e5m2") + return mixed_precision_ops(model_config.quant_config, compute_dtype, disabled=disabled) if ( fp8_compute and From edee33f55ea27a1931475d3ea788fd6e9a81677b Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 6 Jan 2026 19:13:43 -0800 Subject: [PATCH 17/48] Disable comfy kitchen cuda if pytorch cuda less than 13 (#11681) --- comfy/quant_ops.py | 7 +++++++ 1 file changed, 7 insertions(+) diff --git a/comfy/quant_ops.py b/comfy/quant_ops.py index cd737726f..5a17bc6f5 100644 --- a/comfy/quant_ops.py +++ b/comfy/quant_ops.py @@ -13,6 +13,13 @@ try: get_layout_class, ) _CK_AVAILABLE = True + if torch.version.cuda is None: + ck.registry.disable("cuda") + else: + cuda_version = tuple(map(int, str(torch.version.cuda).split('.'))) + if cuda_version < (13,): + ck.registry.disable("cuda") + ck.registry.disable("triton") for k, v in ck.list_backends().items(): logging.info(f"Found comfy_kitchen backend {k}: {v}") From c5cfb34c07048350f472a9a4f1ccbf75a56ed38f Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 6 Jan 2026 20:51:45 -0800 Subject: [PATCH 18/48] Update comfy-kitchen version to 0.2.3 (#11685) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index caad0026a..bc8346bcf 100644 --- a/requirements.txt +++ b/requirements.txt @@ -21,7 +21,7 @@ psutil alembic SQLAlchemy av>=14.2.0 -comfy-kitchen>=0.2.2 +comfy-kitchen>=0.2.3 #non essential dependencies: kornia>=0.7.1 From ce0000c4f2a7dba12324585dddb784b43e3cd3d0 Mon Sep 17 00:00:00 2001 From: Yoland Yan <4950057+yoland68@users.noreply.github.com> Date: Tue, 6 Jan 2026 21:57:31 -0800 Subject: [PATCH 19/48] Force sequential execution in CI test jobs (#11687) Added max-parallel setting to enforce sequential execution in test jobs. --- .github/workflows/test-ci.yml | 2 ++ 1 file changed, 2 insertions(+) diff --git a/.github/workflows/test-ci.yml b/.github/workflows/test-ci.yml index adfc5dd32..63df2dc3a 100644 --- a/.github/workflows/test-ci.yml +++ b/.github/workflows/test-ci.yml @@ -20,6 +20,7 @@ jobs: test-stable: strategy: fail-fast: false + max-parallel: 1 # This forces sequential execution matrix: # os: [macos, linux, windows] # os: [macos, linux] @@ -74,6 +75,7 @@ jobs: test-unix-nightly: strategy: fail-fast: false + max-parallel: 1 # This forces sequential execution matrix: # os: [macos, linux] os: [linux] From 79e94544bd7ec0cc7a4e5e6167907e7d781c4b76 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Wed, 7 Jan 2026 08:04:50 +0200 Subject: [PATCH 20/48] feat(api-nodes): add WAN2.6 ReferenceToVideo (#11644) --- comfy_api_nodes/nodes_wan.py | 160 +++++++++++++++++++++++++ comfy_api_nodes/util/upload_helpers.py | 2 +- 2 files changed, 161 insertions(+), 1 deletion(-) diff --git a/comfy_api_nodes/nodes_wan.py b/comfy_api_nodes/nodes_wan.py index 1675fd863..3e04786a9 100644 --- a/comfy_api_nodes/nodes_wan.py +++ b/comfy_api_nodes/nodes_wan.py @@ -13,7 +13,9 @@ from comfy_api_nodes.util import ( poll_op, sync_op, tensor_to_base64_string, + upload_video_to_comfyapi, validate_audio_duration, + validate_video_duration, ) @@ -41,6 +43,12 @@ class Image2VideoInputField(BaseModel): audio_url: str | None = Field(None) +class Reference2VideoInputField(BaseModel): + prompt: str = Field(...) + negative_prompt: str | None = Field(None) + reference_video_urls: list[str] = Field(...) + + class Txt2ImageParametersField(BaseModel): size: str = Field(...) n: int = Field(1, description="Number of images to generate.") # we support only value=1 @@ -76,6 +84,14 @@ class Image2VideoParametersField(BaseModel): shot_type: str = Field("single") +class Reference2VideoParametersField(BaseModel): + size: str = Field(...) + duration: int = Field(5, ge=5, le=15) + shot_type: str = Field("single") + seed: int = Field(..., ge=0, le=2147483647) + watermark: bool = Field(False) + + class Text2ImageTaskCreationRequest(BaseModel): model: str = Field(...) input: Text2ImageInputField = Field(...) @@ -100,6 +116,12 @@ class Image2VideoTaskCreationRequest(BaseModel): parameters: Image2VideoParametersField = Field(...) +class Reference2VideoTaskCreationRequest(BaseModel): + model: str = Field(...) + input: Reference2VideoInputField = Field(...) + parameters: Reference2VideoParametersField = Field(...) + + class TaskCreationOutputField(BaseModel): task_id: str = Field(...) task_status: str = Field(...) @@ -721,6 +743,143 @@ class WanImageToVideoApi(IO.ComfyNode): return IO.NodeOutput(await download_url_to_video_output(response.output.video_url)) +class WanReferenceVideoApi(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="WanReferenceVideoApi", + display_name="Wan Reference to Video", + category="api node/video/Wan", + description="Use the character and voice from input videos, combined with a prompt, " + "to generate a new video that maintains character consistency.", + inputs=[ + IO.Combo.Input("model", options=["wan2.6-r2v"]), + IO.String.Input( + "prompt", + multiline=True, + default="", + tooltip="Prompt describing the elements and visual features. Supports English and Chinese. " + "Use identifiers such as `character1` and `character2` to refer to the reference characters.", + ), + IO.String.Input( + "negative_prompt", + multiline=True, + default="", + tooltip="Negative prompt describing what to avoid.", + ), + IO.Autogrow.Input( + "reference_videos", + template=IO.Autogrow.TemplateNames( + IO.Video.Input("reference_video"), + names=["character1", "character2", "character3"], + min=1, + ), + ), + IO.Combo.Input( + "size", + options=[ + "720p: 1:1 (960x960)", + "720p: 16:9 (1280x720)", + "720p: 9:16 (720x1280)", + "720p: 4:3 (1088x832)", + "720p: 3:4 (832x1088)", + "1080p: 1:1 (1440x1440)", + "1080p: 16:9 (1920x1080)", + "1080p: 9:16 (1080x1920)", + "1080p: 4:3 (1632x1248)", + "1080p: 3:4 (1248x1632)", + ], + ), + IO.Int.Input( + "duration", + default=5, + min=5, + max=10, + step=5, + display_mode=IO.NumberDisplay.slider, + ), + IO.Int.Input( + "seed", + default=0, + min=0, + max=2147483647, + step=1, + display_mode=IO.NumberDisplay.number, + control_after_generate=True, + ), + IO.Combo.Input( + "shot_type", + options=["single", "multi"], + tooltip="Specifies the shot type for the generated video, that is, whether the video is a " + "single continuous shot or multiple shots with cuts.", + ), + IO.Boolean.Input( + "watermark", + default=False, + tooltip="Whether to add an AI-generated watermark to the result.", + ), + ], + outputs=[ + IO.Video.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + ) + + @classmethod + async def execute( + cls, + model: str, + prompt: str, + negative_prompt: str, + reference_videos: IO.Autogrow.Type, + size: str, + duration: int, + seed: int, + shot_type: str, + watermark: bool, + ): + reference_video_urls = [] + for i in reference_videos: + validate_video_duration(reference_videos[i], min_duration=2, max_duration=30) + for i in reference_videos: + reference_video_urls.append(await upload_video_to_comfyapi(cls, reference_videos[i])) + width, height = RES_IN_PARENS.search(size).groups() + initial_response = await sync_op( + cls, + ApiEndpoint(path="/proxy/wan/api/v1/services/aigc/video-generation/video-synthesis", method="POST"), + response_model=TaskCreationResponse, + data=Reference2VideoTaskCreationRequest( + model=model, + input=Reference2VideoInputField( + prompt=prompt, negative_prompt=negative_prompt, reference_video_urls=reference_video_urls + ), + parameters=Reference2VideoParametersField( + size=f"{width}*{height}", + duration=duration, + shot_type=shot_type, + watermark=watermark, + seed=seed, + ), + ), + ) + if not initial_response.output: + raise Exception(f"An unknown error occurred: {initial_response.code} - {initial_response.message}") + response = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/wan/api/v1/tasks/{initial_response.output.task_id}"), + response_model=VideoTaskStatusResponse, + status_extractor=lambda x: x.output.task_status, + poll_interval=6, + max_poll_attempts=280, + ) + return IO.NodeOutput(await download_url_to_video_output(response.output.video_url)) + + class WanApiExtension(ComfyExtension): @override async def get_node_list(self) -> list[type[IO.ComfyNode]]: @@ -729,6 +888,7 @@ class WanApiExtension(ComfyExtension): WanImageToImageApi, WanTextToVideoApi, WanImageToVideoApi, + WanReferenceVideoApi, ] diff --git a/comfy_api_nodes/util/upload_helpers.py b/comfy_api_nodes/util/upload_helpers.py index b8d33f4d1..f1ed7fe9c 100644 --- a/comfy_api_nodes/util/upload_helpers.py +++ b/comfy_api_nodes/util/upload_helpers.py @@ -119,7 +119,7 @@ async def upload_video_to_comfyapi( raise ValueError(f"Could not verify video duration from source: {e}") from e upload_mime_type = f"video/{container.value.lower()}" - filename = f"uploaded_video.{container.value.lower()}" + filename = f"{uuid.uuid4()}.{container.value.lower()}" # Convert VideoInput to BytesIO using specified container/codec video_bytes_io = BytesIO() From b7d7cc1d496afe3c82279eec74c4d47399aab8ea Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 6 Jan 2026 22:39:06 -0800 Subject: [PATCH 21/48] Fix fp8 fast issue. (#11688) --- comfy/ops.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/comfy/ops.py b/comfy/ops.py index 8f9fdce36..cd536e22d 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -427,12 +427,12 @@ def fp8_linear(self, input): input = torch.clamp(input, min=-448, max=448, out=input) input_fp8 = input.to(dtype).contiguous() layout_params_input = TensorCoreFP8Layout.Params(scale=scale_input, orig_dtype=input_dtype, orig_shape=tuple(input_fp8.shape)) - quantized_input = QuantizedTensor(input_fp8, TensorCoreFP8Layout, layout_params_input) + quantized_input = QuantizedTensor(input_fp8, "TensorCoreFP8Layout", layout_params_input) # Wrap weight in QuantizedTensor - this enables unified dispatch # Call F.linear - __torch_dispatch__ routes to fp8_linear handler in quant_ops.py! layout_params_weight = TensorCoreFP8Layout.Params(scale=scale_weight, orig_dtype=input_dtype, orig_shape=tuple(w.shape)) - quantized_weight = QuantizedTensor(w, TensorCoreFP8Layout, layout_params_weight) + quantized_weight = QuantizedTensor(w, "TensorCoreFP8Layout", layout_params_weight) o = torch.nn.functional.linear(quantized_input, quantized_weight, bias) uncast_bias_weight(self, w, bias, offload_stream) From fc0cb10bcbee6e73ed3caf34c27f7bde4559a07f Mon Sep 17 00:00:00 2001 From: comfyanonymous Date: Wed, 7 Jan 2026 04:07:31 -0500 Subject: [PATCH 22/48] ComfyUI v0.8.0 --- comfyui_version.py | 2 +- pyproject.toml | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/comfyui_version.py b/comfyui_version.py index 1ed60fe5c..750673f08 100644 --- a/comfyui_version.py +++ b/comfyui_version.py @@ -1,3 +1,3 @@ # This file is automatically generated by the build process when version is # updated in pyproject.toml. -__version__ = "0.7.0" +__version__ = "0.8.0" diff --git a/pyproject.toml b/pyproject.toml index a7d159be9..951c2c978 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -1,6 +1,6 @@ [project] name = "ComfyUI" -version = "0.7.0" +version = "0.8.0" readme = "README.md" license = { file = "LICENSE" } requires-python = ">=3.10" From c0c9720d77774ed2c87981da87189fe1c14a57fa Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 7 Jan 2026 01:48:28 -0800 Subject: [PATCH 23/48] Fix stable release workflow not pulling latest comfy kitchen. (#11695) --- .github/workflows/stable-release.yml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/.github/workflows/stable-release.yml b/.github/workflows/stable-release.yml index 28484a9d1..f501b7b31 100644 --- a/.github/workflows/stable-release.yml +++ b/.github/workflows/stable-release.yml @@ -117,7 +117,7 @@ jobs: ./python.exe get-pip.py ./python.exe -s -m pip install ../${{ inputs.cache_tag }}_python_deps/* - grep comfyui ../ComfyUI/requirements.txt > ./requirements_comfyui.txt + grep comfy ../ComfyUI/requirements.txt > ./requirements_comfyui.txt ./python.exe -s -m pip install -r requirements_comfyui.txt rm requirements_comfyui.txt From 3cd7b32f1b7e7e90395cefe7d9f9b1f89276d8ce Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 7 Jan 2026 02:15:14 -0800 Subject: [PATCH 24/48] Support gemma 12B with quant weights. (#11696) --- comfy/text_encoders/lt.py | 12 ++++++------ 1 file changed, 6 insertions(+), 6 deletions(-) diff --git a/comfy/text_encoders/lt.py b/comfy/text_encoders/lt.py index e5964e42b..130ebaeae 100644 --- a/comfy/text_encoders/lt.py +++ b/comfy/text_encoders/lt.py @@ -36,10 +36,10 @@ class LTXAVGemmaTokenizer(sd1_clip.SD1Tokenizer): class Gemma3_12BModel(sd1_clip.SDClipModel): def __init__(self, device="cpu", layer="all", layer_idx=None, dtype=None, attention_mask=True, model_options={}): - llama_scaled_fp8 = model_options.get("gemma_scaled_fp8", None) - if llama_scaled_fp8 is not None: + llama_quantization_metadata = model_options.get("llama_quantization_metadata", None) + if llama_quantization_metadata is not None: model_options = model_options.copy() - model_options["scaled_fp8"] = llama_scaled_fp8 + model_options["quantization_metadata"] = llama_quantization_metadata super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"start": 2, "pad": 0}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Gemma3_12B, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options) @@ -119,12 +119,12 @@ class LTXAVTEModel(torch.nn.Module): return self.load_state_dict(sdo, strict=False) -def ltxav_te(dtype_llama=None, llama_scaled_fp8=None): +def ltxav_te(dtype_llama=None, llama_quantization_metadata=None): class LTXAVTEModel_(LTXAVTEModel): def __init__(self, device="cpu", dtype=None, model_options={}): - if llama_scaled_fp8 is not None and "llama_scaled_fp8" not in model_options: + if llama_quantization_metadata is not None: model_options = model_options.copy() - model_options["llama_scaled_fp8"] = llama_scaled_fp8 + model_options["llama_quantization_metadata"] = llama_quantization_metadata if dtype_llama is not None: dtype = dtype_llama super().__init__(dtype_llama=dtype_llama, device=device, dtype=dtype, model_options=model_options) From 48e5ea1dfd23a9cb5d118d7af661b026d66743bc Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Wed, 7 Jan 2026 15:39:20 -0800 Subject: [PATCH 25/48] model_patcher: Remove confusing load stat (#11710) If the loader passes 1e32 as the usable memory size, it means force the full load. This happens with CPU loads and a few other misc cases. Removing the confusing number and just leave the other details. --- comfy/model_patcher.py | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-) diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py index 93d26c690..4528814ad 100644 --- a/comfy/model_patcher.py +++ b/comfy/model_patcher.py @@ -790,11 +790,12 @@ class ModelPatcher: for param in params: self.pin_weight_to_device("{}.{}".format(n, param)) + usable_stat = "{:.2f} MB usable,".format(lowvram_model_memory / (1024 * 1024)) if lowvram_model_memory < 1e32 else "" if lowvram_counter > 0: - logging.info("loaded partially; {:.2f} MB usable, {:.2f} MB loaded, {:.2f} MB offloaded, {:.2f} MB buffer reserved, lowvram patches: {}".format(lowvram_model_memory / (1024 * 1024), mem_counter / (1024 * 1024), lowvram_mem_counter / (1024 * 1024), offload_buffer / (1024 * 1024), patch_counter)) + logging.info("loaded partially; {} {:.2f} MB loaded, {:.2f} MB offloaded, {:.2f} MB buffer reserved, lowvram patches: {}".format(usable_stat, mem_counter / (1024 * 1024), lowvram_mem_counter / (1024 * 1024), offload_buffer / (1024 * 1024), patch_counter)) self.model.model_lowvram = True else: - logging.info("loaded completely; {:.2f} MB usable, {:.2f} MB loaded, full load: {}".format(lowvram_model_memory / (1024 * 1024), mem_counter / (1024 * 1024), full_load)) + logging.info("loaded completely; {} {:.2f} MB loaded, full load: {}".format(usable_stat, mem_counter / (1024 * 1024), full_load)) self.model.model_lowvram = False if full_load: self.model.to(device_to) From 1c705f7bfb0fb59f6213dfb85ec5d5dc2ce4300e Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Jukka=20Sepp=C3=A4nen?= <40791699+kijai@users.noreply.github.com> Date: Thu, 8 Jan 2026 01:39:59 +0200 Subject: [PATCH 26/48] Add device selection for LTXAVTextEncoderLoader (#11700) --- comfy_extras/nodes_lt_audio.py | 10 +++++++++- 1 file changed, 9 insertions(+), 1 deletion(-) diff --git a/comfy_extras/nodes_lt_audio.py b/comfy_extras/nodes_lt_audio.py index 26b0160d2..1966fd1bf 100644 --- a/comfy_extras/nodes_lt_audio.py +++ b/comfy_extras/nodes_lt_audio.py @@ -185,6 +185,10 @@ class LTXAVTextEncoderLoader(io.ComfyNode): io.Combo.Input( "ckpt_name", options=folder_paths.get_filename_list("checkpoints"), + ), + io.Combo.Input( + "device", + options=["default", "cpu"], ) ], outputs=[io.Clip.Output()], @@ -197,7 +201,11 @@ class LTXAVTextEncoderLoader(io.ComfyNode): clip_path1 = folder_paths.get_full_path_or_raise("text_encoders", text_encoder) clip_path2 = folder_paths.get_full_path_or_raise("checkpoints", ckpt_name) - clip = comfy.sd.load_clip(ckpt_paths=[clip_path1, clip_path2], embedding_directory=folder_paths.get_folder_paths("embeddings"), clip_type=clip_type) + model_options = {} + if device == "cpu": + model_options["load_device"] = model_options["offload_device"] = torch.device("cpu") + + clip = comfy.sd.load_clip(ckpt_paths=[clip_path1, clip_path2], embedding_directory=folder_paths.get_folder_paths("embeddings"), clip_type=clip_type, model_options=model_options) return io.NodeOutput(clip) From 34751fe9f9ade0c715768202c19211dc0c72e760 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 7 Jan 2026 16:12:15 -0800 Subject: [PATCH 27/48] Lower ltxv text encoder vram use. (#11713) --- comfy/text_encoders/lt.py | 3 +++ 1 file changed, 3 insertions(+) diff --git a/comfy/text_encoders/lt.py b/comfy/text_encoders/lt.py index 130ebaeae..dc0694e0e 100644 --- a/comfy/text_encoders/lt.py +++ b/comfy/text_encoders/lt.py @@ -98,10 +98,13 @@ class LTXAVTEModel(torch.nn.Module): out, pooled, extra = self.gemma3_12b.encode_token_weights(token_weight_pairs) out_device = out.device + if comfy.model_management.should_use_bf16(self.execution_device): + out = out.to(device=self.execution_device, dtype=torch.bfloat16) out = out.movedim(1, -1).to(self.execution_device) out = 8.0 * (out - out.mean(dim=(1, 2), keepdim=True)) / (out.amax(dim=(1, 2), keepdim=True) - out.amin(dim=(1, 2), keepdim=True) + 1e-6) out = out.reshape((out.shape[0], out.shape[1], -1)) out = self.text_embedding_projection(out) + out = out.float() out_vid = self.video_embeddings_connector(out)[0] out_audio = self.audio_embeddings_connector(out)[0] out = torch.concat((out_vid, out_audio), dim=-1) From 007b87e7ac29e55ce0ad2c436f5ae68f3a078080 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 7 Jan 2026 16:48:47 -0800 Subject: [PATCH 28/48] Bump required comfy-kitchen version. (#11714) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index bc8346bcf..13e95afa0 100644 --- a/requirements.txt +++ b/requirements.txt @@ -21,7 +21,7 @@ psutil alembic SQLAlchemy av>=14.2.0 -comfy-kitchen>=0.2.3 +comfy-kitchen>=0.2.5 #non essential dependencies: kornia>=0.7.1 From 3cd19e99c10a25cf6e6b51b82e3c16c501733b8c Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 7 Jan 2026 17:04:56 -0800 Subject: [PATCH 29/48] Increase ltxav mem estimation by a bit. (#11715) --- comfy/supported_models.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/comfy/supported_models.py b/comfy/supported_models.py index ee9a79001..d44c0bc37 100644 --- a/comfy/supported_models.py +++ b/comfy/supported_models.py @@ -845,7 +845,7 @@ class LTXAV(LTXV): def __init__(self, unet_config): super().__init__(unet_config) - self.memory_usage_factor = 0.055 # TODO + self.memory_usage_factor = 0.061 # TODO def get_model(self, state_dict, prefix="", device=None): out = model_base.LTXAV(self, device=device) From 25bc1b5b57d61930d6ab60d8cf7e9241d26e4fe9 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 7 Jan 2026 17:11:22 -0800 Subject: [PATCH 30/48] Add memory estimation function to ltxav text encoder. (#11716) --- comfy/sd.py | 11 +++++++---- comfy/text_encoders/lt.py | 8 ++++++++ 2 files changed, 15 insertions(+), 4 deletions(-) diff --git a/comfy/sd.py b/comfy/sd.py index 32157e18b..efde3839c 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -218,7 +218,7 @@ class CLIP: if unprojected: self.cond_stage_model.set_clip_options({"projected_pooled": False}) - self.load_model() + self.load_model(tokens) self.cond_stage_model.set_clip_options({"execution_device": self.patcher.load_device}) all_hooks.reset() self.patcher.patch_hooks(None) @@ -266,7 +266,7 @@ class CLIP: if return_pooled == "unprojected": self.cond_stage_model.set_clip_options({"projected_pooled": False}) - self.load_model() + self.load_model(tokens) self.cond_stage_model.set_clip_options({"execution_device": self.patcher.load_device}) o = self.cond_stage_model.encode_token_weights(tokens) cond, pooled = o[:2] @@ -299,8 +299,11 @@ class CLIP: sd_clip[k] = sd_tokenizer[k] return sd_clip - def load_model(self): - model_management.load_model_gpu(self.patcher) + def load_model(self, tokens={}): + memory_used = 0 + if hasattr(self.cond_stage_model, "memory_estimation_function"): + memory_used = self.cond_stage_model.memory_estimation_function(tokens, device=self.patcher.load_device) + model_management.load_models_gpu([self.patcher], memory_required=memory_used) return self.patcher def get_key_patches(self): diff --git a/comfy/text_encoders/lt.py b/comfy/text_encoders/lt.py index dc0694e0e..776e25e97 100644 --- a/comfy/text_encoders/lt.py +++ b/comfy/text_encoders/lt.py @@ -121,6 +121,14 @@ class LTXAVTEModel(torch.nn.Module): return self.load_state_dict(sdo, strict=False) + def memory_estimation_function(self, token_weight_pairs, device=None): + constant = 6.0 + if comfy.model_management.should_use_bf16(device): + constant /= 2.0 + + token_weight_pairs = token_weight_pairs.get("gemma3_12b", []) + num_tokens = sum(map(lambda a: len(a), token_weight_pairs)) + return num_tokens * constant * 1024 * 1024 def ltxav_te(dtype_llama=None, llama_quantization_metadata=None): class LTXAVTEModel_(LTXAVTEModel): From b6c79a648a013f477f514f61580d1a06220b15eb Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Wed, 7 Jan 2026 18:01:16 -0800 Subject: [PATCH 31/48] ops: Fix offloading with FP8MM performance (#11697) This logic was checking comfy_cast_weights, and going straight to to the forward_comfy_cast_weights implementation without attempting to downscale input to fp8 in the event comfy_cast_weights is set. The main reason comfy_cast_weights would be set would be for async offload, which is not a good reason to nix FP8MM. So instead, and together the underlying exclusions for FP8MM which are: * having a weight_function (usually LowVramPatch) * force_cast_weights (compute dtype override) * the weight is not Quantized * the input is already quantized * the model or layer has MM explictily disabled. If you get past all of those exclusions, quantize the input tensor. Then hand the new input, quantized or not off to forward_comfy_cast_weights to handle it. If the weight is offloaded but input is quantized you will get an offloaded MM8. --- comfy/model_patcher.py | 1 + comfy/ops.py | 30 +++++++++++++++--------------- 2 files changed, 16 insertions(+), 15 deletions(-) diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py index 4528814ad..f6b80a40f 100644 --- a/comfy/model_patcher.py +++ b/comfy/model_patcher.py @@ -718,6 +718,7 @@ class ModelPatcher: continue cast_weight = self.force_cast_weights + m.comfy_force_cast_weights = self.force_cast_weights if lowvram_weight: if hasattr(m, "comfy_cast_weights"): m.weight_function = [] diff --git a/comfy/ops.py b/comfy/ops.py index cd536e22d..8156c42ff 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -654,29 +654,29 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec run_every_op() input_shape = input.shape - tensor_3d = input.ndim == 3 - - if self._full_precision_mm or self.comfy_cast_weights or len(self.weight_function) > 0 or len(self.bias_function) > 0: - return self.forward_comfy_cast_weights(input, *args, **kwargs) + reshaped_3d = False if (getattr(self, 'layout_type', None) is not None and - not isinstance(input, QuantizedTensor)): + not isinstance(input, QuantizedTensor) and not self._full_precision_mm and + not getattr(self, 'comfy_force_cast_weights', False) and + len(self.weight_function) == 0 and len(self.bias_function) == 0): # Reshape 3D tensors to 2D for quantization (needed for NVFP4 and others) - if tensor_3d: - input = input.reshape(-1, input_shape[2]) + input_reshaped = input.reshape(-1, input_shape[2]) if input.ndim == 3 else input - if input.ndim != 2: - # Fall back to comfy_cast_weights for non-2D tensors - return self.forward_comfy_cast_weights(input.reshape(input_shape), *args, **kwargs) + # Fall back to non-quantized for non-2D tensors + if input_reshaped.ndim == 2: + reshaped_3d = input.ndim == 3 + # dtype is now implicit in the layout class + scale = getattr(self, 'input_scale', None) + if scale is not None: + scale = comfy.model_management.cast_to_device(scale, input.device, None) + input = QuantizedTensor.from_float(input_reshaped, self.layout_type, scale=scale) - # dtype is now implicit in the layout class - input = QuantizedTensor.from_float(input, self.layout_type, scale=getattr(self, 'input_scale', None)) - - output = self._forward(input, self.weight, self.bias) + output = self.forward_comfy_cast_weights(input) # Reshape output back to 3D if input was 3D - if tensor_3d: + if reshaped_3d: output = output.reshape((input_shape[0], input_shape[1], self.weight.shape[0])) return output From 21e842508733809354a7b04944b2995ed1169370 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 7 Jan 2026 18:07:26 -0800 Subject: [PATCH 32/48] Add warning for old pytorch. (#11718) --- comfy/quant_ops.py | 1 + 1 file changed, 1 insertion(+) diff --git a/comfy/quant_ops.py b/comfy/quant_ops.py index 5a17bc6f5..8324be42a 100644 --- a/comfy/quant_ops.py +++ b/comfy/quant_ops.py @@ -19,6 +19,7 @@ try: cuda_version = tuple(map(int, str(torch.version.cuda).split('.'))) if cuda_version < (13,): ck.registry.disable("cuda") + logging.warning("WARNING: You need pytorch with cu130 or higher to use optimized CUDA operations.") ck.registry.disable("triton") for k, v in ck.list_backends().items(): From fcd9a236b091bd4e77b177134ddfcf7d7dbd71fd Mon Sep 17 00:00:00 2001 From: ComfyUI Wiki Date: Thu, 8 Jan 2026 10:22:23 +0800 Subject: [PATCH 33/48] Update template to 0.7.69 (#11719) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 13e95afa0..49567ad61 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,5 +1,5 @@ comfyui-frontend-package==1.35.9 -comfyui-workflow-templates==0.7.67 +comfyui-workflow-templates==0.7.69 comfyui-embedded-docs==0.3.1 torch torchsde From ac12f77bed7bbbaf20289533bf7c0bff275e4a41 Mon Sep 17 00:00:00 2001 From: comfyanonymous Date: Wed, 7 Jan 2026 22:10:08 -0500 Subject: [PATCH 34/48] ComfyUI version v0.8.1 --- comfyui_version.py | 2 +- pyproject.toml | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/comfyui_version.py b/comfyui_version.py index 750673f08..4eb6070fe 100644 --- a/comfyui_version.py +++ b/comfyui_version.py @@ -1,3 +1,3 @@ # This file is automatically generated by the build process when version is # updated in pyproject.toml. -__version__ = "0.8.0" +__version__ = "0.8.1" diff --git a/pyproject.toml b/pyproject.toml index 951c2c978..0037abd6c 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -1,6 +1,6 @@ [project] name = "ComfyUI" -version = "0.8.0" +version = "0.8.1" readme = "README.md" license = { file = "LICENSE" } requires-python = ">=3.10" From 50d6e1caf401bf72dca1e9df7e194e722e1bd98b Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 7 Jan 2026 20:07:05 -0800 Subject: [PATCH 35/48] Tweak ltxv vae mem estimation. (#11722) --- comfy/sd.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/comfy/sd.py b/comfy/sd.py index efde3839c..5a7221620 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -479,8 +479,8 @@ class VAE: self.first_stage_model = comfy.ldm.lightricks.vae.causal_video_autoencoder.VideoVAE(version=version, config=vae_config) self.latent_channels = 128 self.latent_dim = 3 - self.memory_used_decode = lambda shape, dtype: (900 * shape[2] * shape[3] * shape[4] * (8 * 8 * 8)) * model_management.dtype_size(dtype) - self.memory_used_encode = lambda shape, dtype: (70 * max(shape[2], 7) * shape[3] * shape[4]) * model_management.dtype_size(dtype) + self.memory_used_decode = lambda shape, dtype: (1200 * shape[2] * shape[3] * shape[4] * (8 * 8 * 8)) * model_management.dtype_size(dtype) + self.memory_used_encode = lambda shape, dtype: (80 * max(shape[2], 7) * shape[3] * shape[4]) * model_management.dtype_size(dtype) self.upscale_ratio = (lambda a: max(0, a * 8 - 7), 32, 32) self.upscale_index_formula = (8, 32, 32) self.downscale_ratio = (lambda a: max(0, math.floor((a + 7) / 8)), 32, 32) From 2e9d51680a90bca9cc375ba7767f7bf3ed27d563 Mon Sep 17 00:00:00 2001 From: comfyanonymous Date: Wed, 7 Jan 2026 23:50:02 -0500 Subject: [PATCH 36/48] ComfyUI version v0.8.2 --- comfyui_version.py | 2 +- pyproject.toml | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/comfyui_version.py b/comfyui_version.py index 4eb6070fe..df82ed4fc 100644 --- a/comfyui_version.py +++ b/comfyui_version.py @@ -1,3 +1,3 @@ # This file is automatically generated by the build process when version is # updated in pyproject.toml. -__version__ = "0.8.1" +__version__ = "0.8.2" diff --git a/pyproject.toml b/pyproject.toml index 0037abd6c..49f1a03fd 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -1,6 +1,6 @@ [project] name = "ComfyUI" -version = "0.8.1" +version = "0.8.2" readme = "README.md" license = { file = "LICENSE" } requires-python = ">=3.10" From a60b7b86c54ea1498e9c5a5c3d6018c0714654d9 Mon Sep 17 00:00:00 2001 From: Yoland Yan <4950057+yoland68@users.noreply.github.com> Date: Wed, 7 Jan 2026 21:41:57 -0800 Subject: [PATCH 37/48] Revert "Force sequential execution in CI test jobs (#11687)" (#11725) This reverts commit ce0000c4f2a7dba12324585dddb784b43e3cd3d0. --- .github/workflows/test-ci.yml | 2 -- 1 file changed, 2 deletions(-) diff --git a/.github/workflows/test-ci.yml b/.github/workflows/test-ci.yml index 63df2dc3a..adfc5dd32 100644 --- a/.github/workflows/test-ci.yml +++ b/.github/workflows/test-ci.yml @@ -20,7 +20,6 @@ jobs: test-stable: strategy: fail-fast: false - max-parallel: 1 # This forces sequential execution matrix: # os: [macos, linux, windows] # os: [macos, linux] @@ -75,7 +74,6 @@ jobs: test-unix-nightly: strategy: fail-fast: false - max-parallel: 1 # This forces sequential execution matrix: # os: [macos, linux] os: [linux] From 5943fbf457d78becbb924a74780e0efc68505a17 Mon Sep 17 00:00:00 2001 From: "Dr.Lt.Data" <128333288+ltdrdata@users.noreply.github.com> Date: Fri, 9 Jan 2026 01:15:42 +0900 Subject: [PATCH 38/48] bump comfyui_manager version to the 4.0.5 (#11732) --- manager_requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/manager_requirements.txt b/manager_requirements.txt index 6585b0c19..bea6d4927 100644 --- a/manager_requirements.txt +++ b/manager_requirements.txt @@ -1 +1 @@ -comfyui_manager==4.0.4 +comfyui_manager==4.0.5 From 0f11869d55c7a459371b8114b1345e55a0274723 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Thu, 8 Jan 2026 14:16:58 -0800 Subject: [PATCH 39/48] Better detection if AMD torch compiled with efficient attention. (#11745) --- comfy/model_management.py | 15 +++++++++++++-- 1 file changed, 13 insertions(+), 2 deletions(-) diff --git a/comfy/model_management.py b/comfy/model_management.py index 928282092..e5de4a5b5 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -22,7 +22,6 @@ from enum import Enum from comfy.cli_args import args, PerformanceFeature import torch import sys -import importlib import platform import weakref import gc @@ -349,10 +348,22 @@ try: except: rocm_version = (6, -1) + def aotriton_supported(gpu_arch): + path = torch.__path__[0] + path = os.path.join(os.path.join(path, "lib"), "aotriton.images") + gfx = set(map(lambda a: a[4:], filter(lambda a: a.startswith("amd-gfx"), os.listdir(path)))) + if gpu_arch in gfx: + return True + if "{}x".format(gpu_arch[:-1]) in gfx: + return True + if "{}xx".format(gpu_arch[:-2]) in gfx: + return True + return False + logging.info("AMD arch: {}".format(arch)) logging.info("ROCm version: {}".format(rocm_version)) if args.use_split_cross_attention == False and args.use_quad_cross_attention == False: - if importlib.util.find_spec('triton') is not None: # AMD efficient attention implementation depends on triton. TODO: better way of detecting if it's compiled in or not. + if aotriton_supported(arch): # AMD efficient attention implementation depends on aotriton. if torch_version_numeric >= (2, 7): # works on 2.6 but doesn't actually seem to improve much if any((a in arch) for a in ["gfx90a", "gfx942", "gfx1100", "gfx1101", "gfx1151"]): # TODO: more arches, TODO: gfx950 ENABLE_PYTORCH_ATTENTION = True From 1a206564487d672561d83ce3eb007517bf018995 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Thu, 8 Jan 2026 14:23:59 -0800 Subject: [PATCH 40/48] Fix import issue. (#11746) --- comfy/ldm/hunyuan_video/upsampler.py | 14 +++++++------- 1 file changed, 7 insertions(+), 7 deletions(-) diff --git a/comfy/ldm/hunyuan_video/upsampler.py b/comfy/ldm/hunyuan_video/upsampler.py index d9e76922f..51b6d1da8 100644 --- a/comfy/ldm/hunyuan_video/upsampler.py +++ b/comfy/ldm/hunyuan_video/upsampler.py @@ -3,8 +3,8 @@ import torch.nn as nn import torch.nn.functional as F from comfy.ldm.modules.diffusionmodules.model import ResnetBlock, VideoConv3d from comfy.ldm.hunyuan_video.vae_refiner import RMS_norm -import model_management -import model_patcher +import comfy.model_management +import comfy.model_patcher class SRResidualCausalBlock3D(nn.Module): def __init__(self, channels: int): @@ -103,13 +103,13 @@ UPSAMPLERS = { class HunyuanVideo15SRModel(): def __init__(self, model_type, config): - self.load_device = model_management.vae_device() - offload_device = model_management.vae_offload_device() - self.dtype = model_management.vae_dtype(self.load_device) + self.load_device = comfy.model_management.vae_device() + offload_device = comfy.model_management.vae_offload_device() + self.dtype = comfy.model_management.vae_dtype(self.load_device) self.model_class = UPSAMPLERS.get(model_type) self.model = self.model_class(**config).eval() - self.patcher = model_patcher.ModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device) + self.patcher = comfy.model_patcher.ModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device) def load_sd(self, sd): return self.model.load_state_dict(sd, strict=True) @@ -118,5 +118,5 @@ class HunyuanVideo15SRModel(): return self.model.state_dict() def resample_latent(self, latent): - model_management.load_model_gpu(self.patcher) + comfy.model_management.load_model_gpu(self.patcher) return self.model(latent.to(self.load_device)) From 027042db6811c875562296f0a6b797c89d59e426 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Jukka=20Sepp=C3=A4nen?= <40791699+kijai@users.noreply.github.com> Date: Fri, 9 Jan 2026 05:14:06 +0200 Subject: [PATCH 41/48] Add node: JoinAudioChannels (#11728) --- comfy_extras/nodes_audio.py | 53 +++++++++++++++++++++++++++++++++++++ 1 file changed, 53 insertions(+) diff --git a/comfy_extras/nodes_audio.py b/comfy_extras/nodes_audio.py index 94ad5e8a8..15b3aa401 100644 --- a/comfy_extras/nodes_audio.py +++ b/comfy_extras/nodes_audio.py @@ -399,6 +399,58 @@ class SplitAudioChannels(IO.ComfyNode): separate = execute # TODO: remove +class JoinAudioChannels(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="JoinAudioChannels", + display_name="Join Audio Channels", + description="Joins left and right mono audio channels into a stereo audio.", + category="audio", + inputs=[ + IO.Audio.Input("audio_left"), + IO.Audio.Input("audio_right"), + ], + outputs=[ + IO.Audio.Output(display_name="audio"), + ], + ) + + @classmethod + def execute(cls, audio_left, audio_right) -> IO.NodeOutput: + waveform_left = audio_left["waveform"] + sample_rate_left = audio_left["sample_rate"] + waveform_right = audio_right["waveform"] + sample_rate_right = audio_right["sample_rate"] + + if waveform_left.shape[1] != 1 or waveform_right.shape[1] != 1: + raise ValueError("AudioJoin: Both input audios must be mono.") + + # Handle different sample rates by resampling to the higher rate + waveform_left, waveform_right, output_sample_rate = match_audio_sample_rates( + waveform_left, sample_rate_left, waveform_right, sample_rate_right + ) + + # Handle different lengths by trimming to the shorter length + length_left = waveform_left.shape[-1] + length_right = waveform_right.shape[-1] + + if length_left != length_right: + min_length = min(length_left, length_right) + if length_left > min_length: + logging.info(f"JoinAudioChannels: Trimming left channel from {length_left} to {min_length} samples.") + waveform_left = waveform_left[..., :min_length] + if length_right > min_length: + logging.info(f"JoinAudioChannels: Trimming right channel from {length_right} to {min_length} samples.") + waveform_right = waveform_right[..., :min_length] + + # Join the channels into stereo + left_channel = waveform_left[..., 0:1, :] + right_channel = waveform_right[..., 0:1, :] + stereo_waveform = torch.cat([left_channel, right_channel], dim=1) + + return IO.NodeOutput({"waveform": stereo_waveform, "sample_rate": output_sample_rate}) + def match_audio_sample_rates(waveform_1, sample_rate_1, waveform_2, sample_rate_2): if sample_rate_1 != sample_rate_2: @@ -616,6 +668,7 @@ class AudioExtension(ComfyExtension): RecordAudio, TrimAudioDuration, SplitAudioChannels, + JoinAudioChannels, AudioConcat, AudioMerge, AudioAdjustVolume, From b48d6a83d4f7012a1b6f6f41e66b0ac3f3253b8a Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Thu, 8 Jan 2026 19:15:50 -0800 Subject: [PATCH 42/48] Fix csp error in frontend when forcing offline. (#11749) --- server.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/server.py b/server.py index 70c8b5e3b..4db3347cb 100644 --- a/server.py +++ b/server.py @@ -184,7 +184,7 @@ def create_block_external_middleware(): else: response = await handler(request) - response.headers['Content-Security-Policy'] = "default-src 'self'; script-src 'self' 'unsafe-inline' 'unsafe-eval' blob:; style-src 'self' 'unsafe-inline'; img-src 'self' data: blob:; font-src 'self'; connect-src 'self'; frame-src 'self'; object-src 'self';" + response.headers['Content-Security-Policy'] = "default-src 'self'; script-src 'self' 'unsafe-inline' 'unsafe-eval' blob:; style-src 'self' 'unsafe-inline'; img-src 'self' data: blob:; font-src 'self'; connect-src 'self' data:; frame-src 'self'; object-src 'self';" return response return block_external_middleware From 114fc73685129bf4e8ddced432247fe67dc6fbff Mon Sep 17 00:00:00 2001 From: Comfy Org PR Bot Date: Fri, 9 Jan 2026 12:16:15 +0900 Subject: [PATCH 43/48] Bump comfyui-frontend-package to 1.36.13 (#11645) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 49567ad61..7686a5f8a 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,4 +1,4 @@ -comfyui-frontend-package==1.35.9 +comfyui-frontend-package==1.36.13 comfyui-workflow-templates==0.7.69 comfyui-embedded-docs==0.3.1 torch From 1dc3da631423b776669a6a9128bb1aeaf5592c55 Mon Sep 17 00:00:00 2001 From: Jedrzej Kosinski Date: Thu, 8 Jan 2026 19:21:51 -0800 Subject: [PATCH 44/48] Add most basic Asset support for models (#11315) * Brought over minimal elements from PR 10045 to reproduce seed_assets and register_assets_system without adding anything to the DB or server routes yet, for now making everything sync (can introduce async once everything is cleaned up and brought over) * Added db script to insert assets stuff, cleaned up some code; assets (models) now get added/rescanned * Added support for 5 http endpoints for assets * Replaced Optional with | None in schemas_in.py and schemas_out.py * Remove two routes that will not be relevant yet in this PR: HEAD /api/assets/hash/ and PUT /api/assets//preview * Remove some functions the two deleted endpoints were using * Don't show assets scan message upon calling /object_info endpoint * removed unsued import to satisfy ruff * Simplified hashing function tpye hint and _hash_file_obj * Satisfied ruff --- alembic_db/versions/0001_assets.py | 174 +++++++++++++++++++ app/assets/api/routes.py | 102 +++++++++++ app/assets/api/schemas_in.py | 94 ++++++++++ app/assets/api/schemas_out.py | 60 +++++++ app/assets/database/bulk_ops.py | 188 ++++++++++++++++++++ app/assets/database/models.py | 233 +++++++++++++++++++++++++ app/assets/database/queries.py | 267 +++++++++++++++++++++++++++++ app/assets/database/tags.py | 62 +++++++ app/assets/hashing.py | 75 ++++++++ app/assets/helpers.py | 216 +++++++++++++++++++++++ app/assets/manager.py | 123 +++++++++++++ app/assets/scanner.py | 229 +++++++++++++++++++++++++ app/database/models.py | 25 ++- comfy/cli_args.py | 1 + main.py | 3 + server.py | 4 + 16 files changed, 1847 insertions(+), 9 deletions(-) create mode 100644 alembic_db/versions/0001_assets.py create mode 100644 app/assets/api/routes.py create mode 100644 app/assets/api/schemas_in.py create mode 100644 app/assets/api/schemas_out.py create mode 100644 app/assets/database/bulk_ops.py create mode 100644 app/assets/database/models.py create mode 100644 app/assets/database/queries.py create mode 100644 app/assets/database/tags.py create mode 100644 app/assets/hashing.py create mode 100644 app/assets/helpers.py create mode 100644 app/assets/manager.py create mode 100644 app/assets/scanner.py diff --git a/alembic_db/versions/0001_assets.py b/alembic_db/versions/0001_assets.py new file mode 100644 index 000000000..1e10b94dc --- /dev/null +++ b/alembic_db/versions/0001_assets.py @@ -0,0 +1,174 @@ +""" +Initial assets schema +Revision ID: 0001_assets +Revises: None +Create Date: 2025-12-10 00:00:00 +""" + +from alembic import op +import sqlalchemy as sa + +revision = "0001_assets" +down_revision = None +branch_labels = None +depends_on = None + + +def upgrade() -> None: + # ASSETS: content identity + op.create_table( + "assets", + sa.Column("id", sa.String(length=36), primary_key=True), + sa.Column("hash", sa.String(length=256), nullable=True), + sa.Column("size_bytes", sa.BigInteger(), nullable=False, server_default="0"), + sa.Column("mime_type", sa.String(length=255), nullable=True), + sa.Column("created_at", sa.DateTime(timezone=False), nullable=False), + sa.CheckConstraint("size_bytes >= 0", name="ck_assets_size_nonneg"), + ) + op.create_index("uq_assets_hash", "assets", ["hash"], unique=True) + op.create_index("ix_assets_mime_type", "assets", ["mime_type"]) + + # ASSETS_INFO: user-visible references + op.create_table( + "assets_info", + sa.Column("id", sa.String(length=36), primary_key=True), + sa.Column("owner_id", sa.String(length=128), nullable=False, server_default=""), + sa.Column("name", sa.String(length=512), nullable=False), + sa.Column("asset_id", sa.String(length=36), sa.ForeignKey("assets.id", ondelete="RESTRICT"), nullable=False), + sa.Column("preview_id", sa.String(length=36), sa.ForeignKey("assets.id", ondelete="SET NULL"), nullable=True), + sa.Column("user_metadata", sa.JSON(), nullable=True), + sa.Column("created_at", sa.DateTime(timezone=False), nullable=False), + sa.Column("updated_at", sa.DateTime(timezone=False), nullable=False), + sa.Column("last_access_time", sa.DateTime(timezone=False), nullable=False), + sa.UniqueConstraint("asset_id", "owner_id", "name", name="uq_assets_info_asset_owner_name"), + ) + op.create_index("ix_assets_info_owner_id", "assets_info", ["owner_id"]) + op.create_index("ix_assets_info_asset_id", "assets_info", ["asset_id"]) + op.create_index("ix_assets_info_name", "assets_info", ["name"]) + op.create_index("ix_assets_info_created_at", "assets_info", ["created_at"]) + op.create_index("ix_assets_info_last_access_time", "assets_info", ["last_access_time"]) + op.create_index("ix_assets_info_owner_name", "assets_info", ["owner_id", "name"]) + + # TAGS: normalized tag vocabulary + op.create_table( + "tags", + sa.Column("name", sa.String(length=512), primary_key=True), + sa.Column("tag_type", sa.String(length=32), nullable=False, server_default="user"), + sa.CheckConstraint("name = lower(name)", name="ck_tags_lowercase"), + ) + op.create_index("ix_tags_tag_type", "tags", ["tag_type"]) + + # ASSET_INFO_TAGS: many-to-many for tags on AssetInfo + op.create_table( + "asset_info_tags", + sa.Column("asset_info_id", sa.String(length=36), sa.ForeignKey("assets_info.id", ondelete="CASCADE"), nullable=False), + sa.Column("tag_name", sa.String(length=512), sa.ForeignKey("tags.name", ondelete="RESTRICT"), nullable=False), + sa.Column("origin", sa.String(length=32), nullable=False, server_default="manual"), + sa.Column("added_at", sa.DateTime(timezone=False), nullable=False), + sa.PrimaryKeyConstraint("asset_info_id", "tag_name", name="pk_asset_info_tags"), + ) + op.create_index("ix_asset_info_tags_tag_name", "asset_info_tags", ["tag_name"]) + op.create_index("ix_asset_info_tags_asset_info_id", "asset_info_tags", ["asset_info_id"]) + + # ASSET_CACHE_STATE: N:1 local cache rows per Asset + op.create_table( + "asset_cache_state", + sa.Column("id", sa.Integer(), primary_key=True, autoincrement=True), + sa.Column("asset_id", sa.String(length=36), sa.ForeignKey("assets.id", ondelete="CASCADE"), nullable=False), + sa.Column("file_path", sa.Text(), nullable=False), # absolute local path to cached file + sa.Column("mtime_ns", sa.BigInteger(), nullable=True), + sa.Column("needs_verify", sa.Boolean(), nullable=False, server_default=sa.text("false")), + sa.CheckConstraint("(mtime_ns IS NULL) OR (mtime_ns >= 0)", name="ck_acs_mtime_nonneg"), + sa.UniqueConstraint("file_path", name="uq_asset_cache_state_file_path"), + ) + op.create_index("ix_asset_cache_state_file_path", "asset_cache_state", ["file_path"]) + op.create_index("ix_asset_cache_state_asset_id", "asset_cache_state", ["asset_id"]) + + # ASSET_INFO_META: typed KV projection of user_metadata for filtering/sorting + op.create_table( + "asset_info_meta", + sa.Column("asset_info_id", sa.String(length=36), sa.ForeignKey("assets_info.id", ondelete="CASCADE"), nullable=False), + sa.Column("key", sa.String(length=256), nullable=False), + sa.Column("ordinal", sa.Integer(), nullable=False, server_default="0"), + sa.Column("val_str", sa.String(length=2048), nullable=True), + sa.Column("val_num", sa.Numeric(38, 10), nullable=True), + sa.Column("val_bool", sa.Boolean(), nullable=True), + sa.Column("val_json", sa.JSON(), nullable=True), + sa.PrimaryKeyConstraint("asset_info_id", "key", "ordinal", name="pk_asset_info_meta"), + ) + op.create_index("ix_asset_info_meta_key", "asset_info_meta", ["key"]) + op.create_index("ix_asset_info_meta_key_val_str", "asset_info_meta", ["key", "val_str"]) + op.create_index("ix_asset_info_meta_key_val_num", "asset_info_meta", ["key", "val_num"]) + op.create_index("ix_asset_info_meta_key_val_bool", "asset_info_meta", ["key", "val_bool"]) + + # Tags vocabulary + tags_table = sa.table( + "tags", + sa.column("name", sa.String(length=512)), + sa.column("tag_type", sa.String()), + ) + op.bulk_insert( + tags_table, + [ + {"name": "models", "tag_type": "system"}, + {"name": "input", "tag_type": "system"}, + {"name": "output", "tag_type": "system"}, + + {"name": "configs", "tag_type": "system"}, + {"name": "checkpoints", "tag_type": "system"}, + {"name": "loras", "tag_type": "system"}, + {"name": "vae", "tag_type": "system"}, + {"name": "text_encoders", "tag_type": "system"}, + {"name": "diffusion_models", "tag_type": "system"}, + {"name": "clip_vision", "tag_type": "system"}, + {"name": "style_models", "tag_type": "system"}, + {"name": "embeddings", "tag_type": "system"}, + {"name": "diffusers", "tag_type": "system"}, + {"name": "vae_approx", "tag_type": "system"}, + {"name": "controlnet", "tag_type": "system"}, + {"name": "gligen", "tag_type": "system"}, + {"name": "upscale_models", "tag_type": "system"}, + {"name": "hypernetworks", "tag_type": "system"}, + {"name": "photomaker", "tag_type": "system"}, + {"name": "classifiers", "tag_type": "system"}, + + {"name": "encoder", "tag_type": "system"}, + {"name": "decoder", "tag_type": "system"}, + + {"name": "missing", "tag_type": "system"}, + {"name": "rescan", "tag_type": "system"}, + ], + ) + + +def downgrade() -> None: + op.drop_index("ix_asset_info_meta_key_val_bool", table_name="asset_info_meta") + op.drop_index("ix_asset_info_meta_key_val_num", table_name="asset_info_meta") + op.drop_index("ix_asset_info_meta_key_val_str", table_name="asset_info_meta") + op.drop_index("ix_asset_info_meta_key", table_name="asset_info_meta") + op.drop_table("asset_info_meta") + + op.drop_index("ix_asset_cache_state_asset_id", table_name="asset_cache_state") + op.drop_index("ix_asset_cache_state_file_path", table_name="asset_cache_state") + op.drop_constraint("uq_asset_cache_state_file_path", table_name="asset_cache_state") + op.drop_table("asset_cache_state") + + op.drop_index("ix_asset_info_tags_asset_info_id", table_name="asset_info_tags") + op.drop_index("ix_asset_info_tags_tag_name", table_name="asset_info_tags") + op.drop_table("asset_info_tags") + + op.drop_index("ix_tags_tag_type", table_name="tags") + op.drop_table("tags") + + op.drop_constraint("uq_assets_info_asset_owner_name", table_name="assets_info") + op.drop_index("ix_assets_info_owner_name", table_name="assets_info") + op.drop_index("ix_assets_info_last_access_time", table_name="assets_info") + op.drop_index("ix_assets_info_created_at", table_name="assets_info") + op.drop_index("ix_assets_info_name", table_name="assets_info") + op.drop_index("ix_assets_info_asset_id", table_name="assets_info") + op.drop_index("ix_assets_info_owner_id", table_name="assets_info") + op.drop_table("assets_info") + + op.drop_index("uq_assets_hash", table_name="assets") + op.drop_index("ix_assets_mime_type", table_name="assets") + op.drop_table("assets") diff --git a/app/assets/api/routes.py b/app/assets/api/routes.py new file mode 100644 index 000000000..30e87a898 --- /dev/null +++ b/app/assets/api/routes.py @@ -0,0 +1,102 @@ +import logging +import uuid +from aiohttp import web + +from pydantic import ValidationError + +import app.assets.manager as manager +from app import user_manager +from app.assets.api import schemas_in +from app.assets.helpers import get_query_dict + +ROUTES = web.RouteTableDef() +USER_MANAGER: user_manager.UserManager | None = None + +# UUID regex (canonical hyphenated form, case-insensitive) +UUID_RE = r"[0-9a-fA-F]{8}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{12}" + +def register_assets_system(app: web.Application, user_manager_instance: user_manager.UserManager) -> None: + global USER_MANAGER + USER_MANAGER = user_manager_instance + app.add_routes(ROUTES) + +def _error_response(status: int, code: str, message: str, details: dict | None = None) -> web.Response: + return web.json_response({"error": {"code": code, "message": message, "details": details or {}}}, status=status) + + +def _validation_error_response(code: str, ve: ValidationError) -> web.Response: + return _error_response(400, code, "Validation failed.", {"errors": ve.json()}) + + +@ROUTES.get("/api/assets") +async def list_assets(request: web.Request) -> web.Response: + """ + GET request to list assets. + """ + query_dict = get_query_dict(request) + try: + q = schemas_in.ListAssetsQuery.model_validate(query_dict) + except ValidationError as ve: + return _validation_error_response("INVALID_QUERY", ve) + + payload = manager.list_assets( + include_tags=q.include_tags, + exclude_tags=q.exclude_tags, + name_contains=q.name_contains, + metadata_filter=q.metadata_filter, + limit=q.limit, + offset=q.offset, + sort=q.sort, + order=q.order, + owner_id=USER_MANAGER.get_request_user_id(request), + ) + return web.json_response(payload.model_dump(mode="json")) + + +@ROUTES.get(f"/api/assets/{{id:{UUID_RE}}}") +async def get_asset(request: web.Request) -> web.Response: + """ + GET request to get an asset's info as JSON. + """ + asset_info_id = str(uuid.UUID(request.match_info["id"])) + try: + result = manager.get_asset( + asset_info_id=asset_info_id, + owner_id=USER_MANAGER.get_request_user_id(request), + ) + except ValueError as e: + return _error_response(404, "ASSET_NOT_FOUND", str(e), {"id": asset_info_id}) + except Exception: + logging.exception( + "get_asset failed for asset_info_id=%s, owner_id=%s", + asset_info_id, + USER_MANAGER.get_request_user_id(request), + ) + return _error_response(500, "INTERNAL", "Unexpected server error.") + return web.json_response(result.model_dump(mode="json"), status=200) + + +@ROUTES.get("/api/tags") +async def get_tags(request: web.Request) -> web.Response: + """ + GET request to list all tags based on query parameters. + """ + query_map = dict(request.rel_url.query) + + try: + query = schemas_in.TagsListQuery.model_validate(query_map) + except ValidationError as e: + return web.json_response( + {"error": {"code": "INVALID_QUERY", "message": "Invalid query parameters", "details": e.errors()}}, + status=400, + ) + + result = manager.list_tags( + prefix=query.prefix, + limit=query.limit, + offset=query.offset, + order=query.order, + include_zero=query.include_zero, + owner_id=USER_MANAGER.get_request_user_id(request), + ) + return web.json_response(result.model_dump(mode="json")) diff --git a/app/assets/api/schemas_in.py b/app/assets/api/schemas_in.py new file mode 100644 index 000000000..200b41aef --- /dev/null +++ b/app/assets/api/schemas_in.py @@ -0,0 +1,94 @@ +import json +import uuid +from typing import Any, Literal + +from pydantic import ( + BaseModel, + ConfigDict, + Field, + conint, + field_validator, +) + + +class ListAssetsQuery(BaseModel): + include_tags: list[str] = Field(default_factory=list) + exclude_tags: list[str] = Field(default_factory=list) + name_contains: str | None = None + + # Accept either a JSON string (query param) or a dict + metadata_filter: dict[str, Any] | None = None + + limit: conint(ge=1, le=500) = 20 + offset: conint(ge=0) = 0 + + sort: Literal["name", "created_at", "updated_at", "size", "last_access_time"] = "created_at" + order: Literal["asc", "desc"] = "desc" + + @field_validator("include_tags", "exclude_tags", mode="before") + @classmethod + def _split_csv_tags(cls, v): + # Accept "a,b,c" or ["a","b"] (we are liberal in what we accept) + if v is None: + return [] + if isinstance(v, str): + return [t.strip() for t in v.split(",") if t.strip()] + if isinstance(v, list): + out: list[str] = [] + for item in v: + if isinstance(item, str): + out.extend([t.strip() for t in item.split(",") if t.strip()]) + return out + return v + + @field_validator("metadata_filter", mode="before") + @classmethod + def _parse_metadata_json(cls, v): + if v is None or isinstance(v, dict): + return v + if isinstance(v, str) and v.strip(): + try: + parsed = json.loads(v) + except Exception as e: + raise ValueError(f"metadata_filter must be JSON: {e}") from e + if not isinstance(parsed, dict): + raise ValueError("metadata_filter must be a JSON object") + return parsed + return None + + +class TagsListQuery(BaseModel): + model_config = ConfigDict(extra="ignore", str_strip_whitespace=True) + + prefix: str | None = Field(None, min_length=1, max_length=256) + limit: int = Field(100, ge=1, le=1000) + offset: int = Field(0, ge=0, le=10_000_000) + order: Literal["count_desc", "name_asc"] = "count_desc" + include_zero: bool = True + + @field_validator("prefix") + @classmethod + def normalize_prefix(cls, v: str | None) -> str | None: + if v is None: + return v + v = v.strip() + return v.lower() or None + + +class SetPreviewBody(BaseModel): + """Set or clear the preview for an AssetInfo. Provide an Asset.id or null.""" + preview_id: str | None = None + + @field_validator("preview_id", mode="before") + @classmethod + def _norm_uuid(cls, v): + if v is None: + return None + s = str(v).strip() + if not s: + return None + try: + uuid.UUID(s) + except Exception: + raise ValueError("preview_id must be a UUID") + return s diff --git a/app/assets/api/schemas_out.py b/app/assets/api/schemas_out.py new file mode 100644 index 000000000..9f8184f20 --- /dev/null +++ b/app/assets/api/schemas_out.py @@ -0,0 +1,60 @@ +from datetime import datetime +from typing import Any + +from pydantic import BaseModel, ConfigDict, Field, field_serializer + + +class AssetSummary(BaseModel): + id: str + name: str + asset_hash: str | None = None + size: int | None = None + mime_type: str | None = None + tags: list[str] = Field(default_factory=list) + preview_url: str | None = None + created_at: datetime | None = None + updated_at: datetime | None = None + last_access_time: datetime | None = None + + model_config = ConfigDict(from_attributes=True) + + @field_serializer("created_at", "updated_at", "last_access_time") + def _ser_dt(self, v: datetime | None, _info): + return v.isoformat() if v else None + + +class AssetsList(BaseModel): + assets: list[AssetSummary] + total: int + has_more: bool + + +class AssetDetail(BaseModel): + id: str + name: str + asset_hash: str | None = None + size: int | None = None + mime_type: str | None = None + tags: list[str] = Field(default_factory=list) + user_metadata: dict[str, Any] = Field(default_factory=dict) + preview_id: str | None = None + created_at: datetime | None = None + last_access_time: datetime | None = None + + model_config = ConfigDict(from_attributes=True) + + @field_serializer("created_at", "last_access_time") + def _ser_dt(self, v: datetime | None, _info): + return v.isoformat() if v else None + + +class TagUsage(BaseModel): + name: str + count: int + type: str + + +class TagsList(BaseModel): + tags: list[TagUsage] = Field(default_factory=list) + total: int + has_more: bool diff --git a/app/assets/database/bulk_ops.py b/app/assets/database/bulk_ops.py new file mode 100644 index 000000000..9352cd65d --- /dev/null +++ b/app/assets/database/bulk_ops.py @@ -0,0 +1,188 @@ +import os +import uuid +import sqlalchemy +from typing import Iterable +from sqlalchemy.orm import Session +from sqlalchemy.dialects import sqlite + +from app.assets.helpers import utcnow +from app.assets.database.models import Asset, AssetCacheState, AssetInfo, AssetInfoTag, AssetInfoMeta + +MAX_BIND_PARAMS = 800 + +def _chunk_rows(rows: list[dict], cols_per_row: int, max_bind_params: int) -> Iterable[list[dict]]: + if not rows: + return [] + rows_per_stmt = max(1, max_bind_params // max(1, cols_per_row)) + for i in range(0, len(rows), rows_per_stmt): + yield rows[i:i + rows_per_stmt] + +def _iter_chunks(seq, n: int): + for i in range(0, len(seq), n): + yield seq[i:i + n] + +def _rows_per_stmt(cols: int) -> int: + return max(1, MAX_BIND_PARAMS // max(1, cols)) + + +def seed_from_paths_batch( + session: Session, + *, + specs: list[dict], + owner_id: str = "", +) -> dict: + """Each spec is a dict with keys: + - abs_path: str + - size_bytes: int + - mtime_ns: int + - info_name: str + - tags: list[str] + - fname: Optional[str] + """ + if not specs: + return {"inserted_infos": 0, "won_states": 0, "lost_states": 0} + + now = utcnow() + asset_rows: list[dict] = [] + state_rows: list[dict] = [] + path_to_asset: dict[str, str] = {} + asset_to_info: dict[str, dict] = {} # asset_id -> prepared info row + path_list: list[str] = [] + + for sp in specs: + ap = os.path.abspath(sp["abs_path"]) + aid = str(uuid.uuid4()) + iid = str(uuid.uuid4()) + path_list.append(ap) + path_to_asset[ap] = aid + + asset_rows.append( + { + "id": aid, + "hash": None, + "size_bytes": sp["size_bytes"], + "mime_type": None, + "created_at": now, + } + ) + state_rows.append( + { + "asset_id": aid, + "file_path": ap, + "mtime_ns": sp["mtime_ns"], + } + ) + asset_to_info[aid] = { + "id": iid, + "owner_id": owner_id, + "name": sp["info_name"], + "asset_id": aid, + "preview_id": None, + "user_metadata": {"filename": sp["fname"]} if sp["fname"] else None, + "created_at": now, + "updated_at": now, + "last_access_time": now, + "_tags": sp["tags"], + "_filename": sp["fname"], + } + + # insert all seed Assets (hash=NULL) + ins_asset = sqlite.insert(Asset) + for chunk in _iter_chunks(asset_rows, _rows_per_stmt(5)): + session.execute(ins_asset, chunk) + + # try to claim AssetCacheState (file_path) + winners_by_path: set[str] = set() + ins_state = ( + sqlite.insert(AssetCacheState) + .on_conflict_do_nothing(index_elements=[AssetCacheState.file_path]) + .returning(AssetCacheState.file_path) + ) + for chunk in _iter_chunks(state_rows, _rows_per_stmt(3)): + winners_by_path.update((session.execute(ins_state, chunk)).scalars().all()) + + all_paths_set = set(path_list) + losers_by_path = all_paths_set - winners_by_path + lost_assets = [path_to_asset[p] for p in losers_by_path] + if lost_assets: # losers get their Asset removed + for id_chunk in _iter_chunks(lost_assets, MAX_BIND_PARAMS): + session.execute(sqlalchemy.delete(Asset).where(Asset.id.in_(id_chunk))) + + if not winners_by_path: + return {"inserted_infos": 0, "won_states": 0, "lost_states": len(losers_by_path)} + + # insert AssetInfo only for winners + winner_info_rows = [asset_to_info[path_to_asset[p]] for p in winners_by_path] + ins_info = ( + sqlite.insert(AssetInfo) + .on_conflict_do_nothing(index_elements=[AssetInfo.asset_id, AssetInfo.owner_id, AssetInfo.name]) + .returning(AssetInfo.id) + ) + + inserted_info_ids: set[str] = set() + for chunk in _iter_chunks(winner_info_rows, _rows_per_stmt(9)): + inserted_info_ids.update((session.execute(ins_info, chunk)).scalars().all()) + + # build and insert tag + meta rows for the AssetInfo + tag_rows: list[dict] = [] + meta_rows: list[dict] = [] + if inserted_info_ids: + for row in winner_info_rows: + iid = row["id"] + if iid not in inserted_info_ids: + continue + for t in row["_tags"]: + tag_rows.append({ + "asset_info_id": iid, + "tag_name": t, + "origin": "automatic", + "added_at": now, + }) + if row["_filename"]: + meta_rows.append( + { + "asset_info_id": iid, + "key": "filename", + "ordinal": 0, + "val_str": row["_filename"], + "val_num": None, + "val_bool": None, + "val_json": None, + } + ) + + bulk_insert_tags_and_meta(session, tag_rows=tag_rows, meta_rows=meta_rows, max_bind_params=MAX_BIND_PARAMS) + return { + "inserted_infos": len(inserted_info_ids), + "won_states": len(winners_by_path), + "lost_states": len(losers_by_path), + } + + +def bulk_insert_tags_and_meta( + session: Session, + *, + tag_rows: list[dict], + meta_rows: list[dict], + max_bind_params: int, +) -> None: + """Batch insert into asset_info_tags and asset_info_meta with ON CONFLICT DO NOTHING. + - tag_rows keys: asset_info_id, tag_name, origin, added_at + - meta_rows keys: asset_info_id, key, ordinal, val_str, val_num, val_bool, val_json + """ + if tag_rows: + ins_links = ( + sqlite.insert(AssetInfoTag) + .on_conflict_do_nothing(index_elements=[AssetInfoTag.asset_info_id, AssetInfoTag.tag_name]) + ) + for chunk in _chunk_rows(tag_rows, cols_per_row=4, max_bind_params=max_bind_params): + session.execute(ins_links, chunk) + if meta_rows: + ins_meta = ( + sqlite.insert(AssetInfoMeta) + .on_conflict_do_nothing( + index_elements=[AssetInfoMeta.asset_info_id, AssetInfoMeta.key, AssetInfoMeta.ordinal] + ) + ) + for chunk in _chunk_rows(meta_rows, cols_per_row=7, max_bind_params=max_bind_params): + session.execute(ins_meta, chunk) diff --git a/app/assets/database/models.py b/app/assets/database/models.py new file mode 100644 index 000000000..3cd28f68b --- /dev/null +++ b/app/assets/database/models.py @@ -0,0 +1,233 @@ +from __future__ import annotations + +import uuid +from datetime import datetime + +from typing import Any +from sqlalchemy import ( + JSON, + BigInteger, + Boolean, + CheckConstraint, + DateTime, + ForeignKey, + Index, + Integer, + Numeric, + String, + Text, + UniqueConstraint, +) +from sqlalchemy.orm import Mapped, foreign, mapped_column, relationship + +from app.assets.helpers import utcnow +from app.database.models import to_dict, Base + + +class Asset(Base): + __tablename__ = "assets" + + id: Mapped[str] = mapped_column(String(36), primary_key=True, default=lambda: str(uuid.uuid4())) + hash: Mapped[str | None] = mapped_column(String(256), nullable=True) + size_bytes: Mapped[int] = mapped_column(BigInteger, nullable=False, default=0) + mime_type: Mapped[str | None] = mapped_column(String(255)) + created_at: Mapped[datetime] = mapped_column( + DateTime(timezone=False), nullable=False, default=utcnow + ) + + infos: Mapped[list[AssetInfo]] = relationship( + "AssetInfo", + back_populates="asset", + primaryjoin=lambda: Asset.id == foreign(AssetInfo.asset_id), + foreign_keys=lambda: [AssetInfo.asset_id], + cascade="all,delete-orphan", + passive_deletes=True, + ) + + preview_of: Mapped[list[AssetInfo]] = relationship( + "AssetInfo", + back_populates="preview_asset", + primaryjoin=lambda: Asset.id == foreign(AssetInfo.preview_id), + foreign_keys=lambda: [AssetInfo.preview_id], + viewonly=True, + ) + + cache_states: Mapped[list[AssetCacheState]] = relationship( + back_populates="asset", + cascade="all, delete-orphan", + passive_deletes=True, + ) + + __table_args__ = ( + Index("uq_assets_hash", "hash", unique=True), + Index("ix_assets_mime_type", "mime_type"), + CheckConstraint("size_bytes >= 0", name="ck_assets_size_nonneg"), + ) + + def to_dict(self, include_none: bool = False) -> dict[str, Any]: + return to_dict(self, include_none=include_none) + + def __repr__(self) -> str: + return f"" + + +class AssetCacheState(Base): + __tablename__ = "asset_cache_state" + + id: Mapped[int] = mapped_column(Integer, primary_key=True, autoincrement=True) + asset_id: Mapped[str] = mapped_column(String(36), ForeignKey("assets.id", ondelete="CASCADE"), nullable=False) + file_path: Mapped[str] = mapped_column(Text, nullable=False) + mtime_ns: Mapped[int | None] = mapped_column(BigInteger, nullable=True) + needs_verify: Mapped[bool] = mapped_column(Boolean, nullable=False, default=False) + + asset: Mapped[Asset] = relationship(back_populates="cache_states") + + __table_args__ = ( + Index("ix_asset_cache_state_file_path", "file_path"), + Index("ix_asset_cache_state_asset_id", "asset_id"), + CheckConstraint("(mtime_ns IS NULL) OR (mtime_ns >= 0)", name="ck_acs_mtime_nonneg"), + UniqueConstraint("file_path", name="uq_asset_cache_state_file_path"), + ) + + def to_dict(self, include_none: bool = False) -> dict[str, Any]: + return to_dict(self, include_none=include_none) + + def __repr__(self) -> str: + return f"" + + +class AssetInfo(Base): + __tablename__ = "assets_info" + + id: Mapped[str] = mapped_column(String(36), primary_key=True, default=lambda: str(uuid.uuid4())) + owner_id: Mapped[str] = mapped_column(String(128), nullable=False, default="") + name: Mapped[str] = mapped_column(String(512), nullable=False) + asset_id: Mapped[str] = mapped_column(String(36), ForeignKey("assets.id", ondelete="RESTRICT"), nullable=False) + preview_id: Mapped[str | None] = mapped_column(String(36), ForeignKey("assets.id", ondelete="SET NULL")) + user_metadata: Mapped[dict[str, Any] | None] = mapped_column(JSON(none_as_null=True)) + created_at: Mapped[datetime] = mapped_column(DateTime(timezone=False), nullable=False, default=utcnow) + updated_at: Mapped[datetime] = mapped_column(DateTime(timezone=False), nullable=False, default=utcnow) + last_access_time: Mapped[datetime] = mapped_column(DateTime(timezone=False), nullable=False, default=utcnow) + + asset: Mapped[Asset] = relationship( + "Asset", + back_populates="infos", + foreign_keys=[asset_id], + lazy="selectin", + ) + preview_asset: Mapped[Asset | None] = relationship( + "Asset", + back_populates="preview_of", + foreign_keys=[preview_id], + ) + + metadata_entries: Mapped[list[AssetInfoMeta]] = relationship( + back_populates="asset_info", + cascade="all,delete-orphan", + passive_deletes=True, + ) + + tag_links: Mapped[list[AssetInfoTag]] = relationship( + back_populates="asset_info", + cascade="all,delete-orphan", + passive_deletes=True, + overlaps="tags,asset_infos", + ) + + tags: Mapped[list[Tag]] = relationship( + secondary="asset_info_tags", + back_populates="asset_infos", + lazy="selectin", + viewonly=True, + overlaps="tag_links,asset_info_links,asset_infos,tag", + ) + + __table_args__ = ( + UniqueConstraint("asset_id", "owner_id", "name", name="uq_assets_info_asset_owner_name"), + Index("ix_assets_info_owner_name", "owner_id", "name"), + Index("ix_assets_info_owner_id", "owner_id"), + Index("ix_assets_info_asset_id", "asset_id"), + Index("ix_assets_info_name", "name"), + Index("ix_assets_info_created_at", "created_at"), + Index("ix_assets_info_last_access_time", "last_access_time"), + ) + + def to_dict(self, include_none: bool = False) -> dict[str, Any]: + data = to_dict(self, include_none=include_none) + data["tags"] = [t.name for t in self.tags] + return data + + def __repr__(self) -> str: + return f"" + + +class AssetInfoMeta(Base): + __tablename__ = "asset_info_meta" + + asset_info_id: Mapped[str] = mapped_column( + String(36), ForeignKey("assets_info.id", ondelete="CASCADE"), primary_key=True + ) + key: Mapped[str] = mapped_column(String(256), primary_key=True) + ordinal: Mapped[int] = mapped_column(Integer, primary_key=True, default=0) + + val_str: Mapped[str | None] = mapped_column(String(2048), nullable=True) + val_num: Mapped[float | None] = mapped_column(Numeric(38, 10), nullable=True) + val_bool: Mapped[bool | None] = mapped_column(Boolean, nullable=True) + val_json: Mapped[Any | None] = mapped_column(JSON(none_as_null=True), nullable=True) + + asset_info: Mapped[AssetInfo] = relationship(back_populates="metadata_entries") + + __table_args__ = ( + Index("ix_asset_info_meta_key", "key"), + Index("ix_asset_info_meta_key_val_str", "key", "val_str"), + Index("ix_asset_info_meta_key_val_num", "key", "val_num"), + Index("ix_asset_info_meta_key_val_bool", "key", "val_bool"), + ) + + +class AssetInfoTag(Base): + __tablename__ = "asset_info_tags" + + asset_info_id: Mapped[str] = mapped_column( + String(36), ForeignKey("assets_info.id", ondelete="CASCADE"), primary_key=True + ) + tag_name: Mapped[str] = mapped_column( + String(512), ForeignKey("tags.name", ondelete="RESTRICT"), primary_key=True + ) + origin: Mapped[str] = mapped_column(String(32), nullable=False, default="manual") + added_at: Mapped[datetime] = mapped_column( + DateTime(timezone=False), nullable=False, default=utcnow + ) + + asset_info: Mapped[AssetInfo] = relationship(back_populates="tag_links") + tag: Mapped[Tag] = relationship(back_populates="asset_info_links") + + __table_args__ = ( + Index("ix_asset_info_tags_tag_name", "tag_name"), + Index("ix_asset_info_tags_asset_info_id", "asset_info_id"), + ) + + +class Tag(Base): + __tablename__ = "tags" + + name: Mapped[str] = mapped_column(String(512), primary_key=True) + tag_type: Mapped[str] = mapped_column(String(32), nullable=False, default="user") + + asset_info_links: Mapped[list[AssetInfoTag]] = relationship( + back_populates="tag", + overlaps="asset_infos,tags", + ) + asset_infos: Mapped[list[AssetInfo]] = relationship( + secondary="asset_info_tags", + back_populates="tags", + viewonly=True, + overlaps="asset_info_links,tag_links,tags,asset_info", + ) + + __table_args__ = ( + Index("ix_tags_tag_type", "tag_type"), + ) + + def __repr__(self) -> str: + return f"" diff --git a/app/assets/database/queries.py b/app/assets/database/queries.py new file mode 100644 index 000000000..0824c0c2f --- /dev/null +++ b/app/assets/database/queries.py @@ -0,0 +1,267 @@ +import sqlalchemy as sa +from collections import defaultdict +from sqlalchemy import select, exists, func +from sqlalchemy.orm import Session, contains_eager, noload +from app.assets.database.models import Asset, AssetInfo, AssetInfoMeta, AssetInfoTag, Tag +from app.assets.helpers import escape_like_prefix, normalize_tags +from typing import Sequence + + +def visible_owner_clause(owner_id: str) -> sa.sql.ClauseElement: + """Build owner visibility predicate for reads. Owner-less rows are visible to everyone.""" + owner_id = (owner_id or "").strip() + if owner_id == "": + return AssetInfo.owner_id == "" + return AssetInfo.owner_id.in_(["", owner_id]) + + +def apply_tag_filters( + stmt: sa.sql.Select, + include_tags: Sequence[str] | None = None, + exclude_tags: Sequence[str] | None = None, +) -> sa.sql.Select: + """include_tags: every tag must be present; exclude_tags: none may be present.""" + include_tags = normalize_tags(include_tags) + exclude_tags = normalize_tags(exclude_tags) + + if include_tags: + for tag_name in include_tags: + stmt = stmt.where( + exists().where( + (AssetInfoTag.asset_info_id == AssetInfo.id) + & (AssetInfoTag.tag_name == tag_name) + ) + ) + + if exclude_tags: + stmt = stmt.where( + ~exists().where( + (AssetInfoTag.asset_info_id == AssetInfo.id) + & (AssetInfoTag.tag_name.in_(exclude_tags)) + ) + ) + return stmt + +def apply_metadata_filter( + stmt: sa.sql.Select, + metadata_filter: dict | None = None, +) -> sa.sql.Select: + """Apply filters using asset_info_meta projection table.""" + if not metadata_filter: + return stmt + + def _exists_for_pred(key: str, *preds) -> sa.sql.ClauseElement: + return sa.exists().where( + AssetInfoMeta.asset_info_id == AssetInfo.id, + AssetInfoMeta.key == key, + *preds, + ) + + def _exists_clause_for_value(key: str, value) -> sa.sql.ClauseElement: + if value is None: + no_row_for_key = sa.not_( + sa.exists().where( + AssetInfoMeta.asset_info_id == AssetInfo.id, + AssetInfoMeta.key == key, + ) + ) + null_row = _exists_for_pred( + key, + AssetInfoMeta.val_json.is_(None), + AssetInfoMeta.val_str.is_(None), + AssetInfoMeta.val_num.is_(None), + AssetInfoMeta.val_bool.is_(None), + ) + return sa.or_(no_row_for_key, null_row) + + if isinstance(value, bool): + return _exists_for_pred(key, AssetInfoMeta.val_bool == bool(value)) + if isinstance(value, (int, float)): + from decimal import Decimal + num = value if isinstance(value, Decimal) else Decimal(str(value)) + return _exists_for_pred(key, AssetInfoMeta.val_num == num) + if isinstance(value, str): + return _exists_for_pred(key, AssetInfoMeta.val_str == value) + return _exists_for_pred(key, AssetInfoMeta.val_json == value) + + for k, v in metadata_filter.items(): + if isinstance(v, list): + ors = [_exists_clause_for_value(k, elem) for elem in v] + if ors: + stmt = stmt.where(sa.or_(*ors)) + else: + stmt = stmt.where(_exists_clause_for_value(k, v)) + return stmt + + +def asset_exists_by_hash(session: Session, asset_hash: str) -> bool: + """ + Check if an asset with a given hash exists in database. + """ + row = ( + session.execute( + select(sa.literal(True)).select_from(Asset).where(Asset.hash == asset_hash).limit(1) + ) + ).first() + return row is not None + +def get_asset_info_by_id(session: Session, asset_info_id: str) -> AssetInfo | None: + return session.get(AssetInfo, asset_info_id) + +def list_asset_infos_page( + session: Session, + owner_id: str = "", + include_tags: Sequence[str] | None = None, + exclude_tags: Sequence[str] | None = None, + name_contains: str | None = None, + metadata_filter: dict | None = None, + limit: int = 20, + offset: int = 0, + sort: str = "created_at", + order: str = "desc", +) -> tuple[list[AssetInfo], dict[str, list[str]], int]: + base = ( + select(AssetInfo) + .join(Asset, Asset.id == AssetInfo.asset_id) + .options(contains_eager(AssetInfo.asset), noload(AssetInfo.tags)) + .where(visible_owner_clause(owner_id)) + ) + + if name_contains: + escaped, esc = escape_like_prefix(name_contains) + base = base.where(AssetInfo.name.ilike(f"%{escaped}%", escape=esc)) + + base = apply_tag_filters(base, include_tags, exclude_tags) + base = apply_metadata_filter(base, metadata_filter) + + sort = (sort or "created_at").lower() + order = (order or "desc").lower() + sort_map = { + "name": AssetInfo.name, + "created_at": AssetInfo.created_at, + "updated_at": AssetInfo.updated_at, + "last_access_time": AssetInfo.last_access_time, + "size": Asset.size_bytes, + } + sort_col = sort_map.get(sort, AssetInfo.created_at) + sort_exp = sort_col.desc() if order == "desc" else sort_col.asc() + + base = base.order_by(sort_exp).limit(limit).offset(offset) + + count_stmt = ( + select(sa.func.count()) + .select_from(AssetInfo) + .join(Asset, Asset.id == AssetInfo.asset_id) + .where(visible_owner_clause(owner_id)) + ) + if name_contains: + escaped, esc = escape_like_prefix(name_contains) + count_stmt = count_stmt.where(AssetInfo.name.ilike(f"%{escaped}%", escape=esc)) + count_stmt = apply_tag_filters(count_stmt, include_tags, exclude_tags) + count_stmt = apply_metadata_filter(count_stmt, metadata_filter) + + total = int((session.execute(count_stmt)).scalar_one() or 0) + + infos = (session.execute(base)).unique().scalars().all() + + id_list: list[str] = [i.id for i in infos] + tag_map: dict[str, list[str]] = defaultdict(list) + if id_list: + rows = session.execute( + select(AssetInfoTag.asset_info_id, Tag.name) + .join(Tag, Tag.name == AssetInfoTag.tag_name) + .where(AssetInfoTag.asset_info_id.in_(id_list)) + ) + for aid, tag_name in rows.all(): + tag_map[aid].append(tag_name) + + return infos, tag_map, total + +def fetch_asset_info_asset_and_tags( + session: Session, + asset_info_id: str, + owner_id: str = "", +) -> tuple[AssetInfo, Asset, list[str]] | None: + stmt = ( + select(AssetInfo, Asset, Tag.name) + .join(Asset, Asset.id == AssetInfo.asset_id) + .join(AssetInfoTag, AssetInfoTag.asset_info_id == AssetInfo.id, isouter=True) + .join(Tag, Tag.name == AssetInfoTag.tag_name, isouter=True) + .where( + AssetInfo.id == asset_info_id, + visible_owner_clause(owner_id), + ) + .options(noload(AssetInfo.tags)) + .order_by(Tag.name.asc()) + ) + + rows = (session.execute(stmt)).all() + if not rows: + return None + + first_info, first_asset, _ = rows[0] + tags: list[str] = [] + seen: set[str] = set() + for _info, _asset, tag_name in rows: + if tag_name and tag_name not in seen: + seen.add(tag_name) + tags.append(tag_name) + return first_info, first_asset, tags + +def list_tags_with_usage( + session: Session, + prefix: str | None = None, + limit: int = 100, + offset: int = 0, + include_zero: bool = True, + order: str = "count_desc", + owner_id: str = "", +) -> tuple[list[tuple[str, str, int]], int]: + counts_sq = ( + select( + AssetInfoTag.tag_name.label("tag_name"), + func.count(AssetInfoTag.asset_info_id).label("cnt"), + ) + .select_from(AssetInfoTag) + .join(AssetInfo, AssetInfo.id == AssetInfoTag.asset_info_id) + .where(visible_owner_clause(owner_id)) + .group_by(AssetInfoTag.tag_name) + .subquery() + ) + + q = ( + select( + Tag.name, + Tag.tag_type, + func.coalesce(counts_sq.c.cnt, 0).label("count"), + ) + .select_from(Tag) + .join(counts_sq, counts_sq.c.tag_name == Tag.name, isouter=True) + ) + + if prefix: + escaped, esc = escape_like_prefix(prefix.strip().lower()) + q = q.where(Tag.name.like(escaped + "%", escape=esc)) + + if not include_zero: + q = q.where(func.coalesce(counts_sq.c.cnt, 0) > 0) + + if order == "name_asc": + q = q.order_by(Tag.name.asc()) + else: + q = q.order_by(func.coalesce(counts_sq.c.cnt, 0).desc(), Tag.name.asc()) + + total_q = select(func.count()).select_from(Tag) + if prefix: + escaped, esc = escape_like_prefix(prefix.strip().lower()) + total_q = total_q.where(Tag.name.like(escaped + "%", escape=esc)) + if not include_zero: + total_q = total_q.where( + Tag.name.in_(select(AssetInfoTag.tag_name).group_by(AssetInfoTag.tag_name)) + ) + + rows = (session.execute(q.limit(limit).offset(offset))).all() + total = (session.execute(total_q)).scalar_one() + + rows_norm = [(name, ttype, int(count or 0)) for (name, ttype, count) in rows] + return rows_norm, int(total or 0) diff --git a/app/assets/database/tags.py b/app/assets/database/tags.py new file mode 100644 index 000000000..3ab6497c2 --- /dev/null +++ b/app/assets/database/tags.py @@ -0,0 +1,62 @@ +from typing import Iterable + +import sqlalchemy +from sqlalchemy.orm import Session +from sqlalchemy.dialects import sqlite + +from app.assets.helpers import normalize_tags, utcnow +from app.assets.database.models import Tag, AssetInfoTag, AssetInfo + + +def ensure_tags_exist(session: Session, names: Iterable[str], tag_type: str = "user") -> None: + wanted = normalize_tags(list(names)) + if not wanted: + return + rows = [{"name": n, "tag_type": tag_type} for n in list(dict.fromkeys(wanted))] + ins = ( + sqlite.insert(Tag) + .values(rows) + .on_conflict_do_nothing(index_elements=[Tag.name]) + ) + return session.execute(ins) + +def add_missing_tag_for_asset_id( + session: Session, + *, + asset_id: str, + origin: str = "automatic", +) -> None: + select_rows = ( + sqlalchemy.select( + AssetInfo.id.label("asset_info_id"), + sqlalchemy.literal("missing").label("tag_name"), + sqlalchemy.literal(origin).label("origin"), + sqlalchemy.literal(utcnow()).label("added_at"), + ) + .where(AssetInfo.asset_id == asset_id) + .where( + sqlalchemy.not_( + sqlalchemy.exists().where((AssetInfoTag.asset_info_id == AssetInfo.id) & (AssetInfoTag.tag_name == "missing")) + ) + ) + ) + session.execute( + sqlite.insert(AssetInfoTag) + .from_select( + ["asset_info_id", "tag_name", "origin", "added_at"], + select_rows, + ) + .on_conflict_do_nothing(index_elements=[AssetInfoTag.asset_info_id, AssetInfoTag.tag_name]) + ) + +def remove_missing_tag_for_asset_id( + session: Session, + *, + asset_id: str, +) -> None: + session.execute( + sqlalchemy.delete(AssetInfoTag).where( + AssetInfoTag.asset_info_id.in_(sqlalchemy.select(AssetInfo.id).where(AssetInfo.asset_id == asset_id)), + AssetInfoTag.tag_name == "missing", + ) + ) diff --git a/app/assets/hashing.py b/app/assets/hashing.py new file mode 100644 index 000000000..4b72084b9 --- /dev/null +++ b/app/assets/hashing.py @@ -0,0 +1,75 @@ +from blake3 import blake3 +from typing import IO +import os +import asyncio + + +DEFAULT_CHUNK = 8 * 1024 *1024 # 8MB + +# NOTE: this allows hashing different representations of a file-like object +def blake3_hash( + fp: str | IO[bytes], + chunk_size: int = DEFAULT_CHUNK, +) -> str: + """ + Returns a BLAKE3 hex digest for ``fp``, which may be: + - a filename (str/bytes) or PathLike + - an open binary file object + If ``fp`` is a file object, it must be opened in **binary** mode and support + ``read``, ``seek``, and ``tell``. The function will seek to the start before + reading and will attempt to restore the original position afterward. + """ + # duck typing to check if input is a file-like object + if hasattr(fp, "read"): + return _hash_file_obj(fp, chunk_size) + + with open(os.fspath(fp), "rb") as f: + return _hash_file_obj(f, chunk_size) + + +async def blake3_hash_async( + fp: str | IO[bytes], + chunk_size: int = DEFAULT_CHUNK, +) -> str: + """Async wrapper for ``blake3_hash_sync``. + Uses a worker thread so the event loop remains responsive. + """ + # If it is a path, open inside the worker thread to keep I/O off the loop. + if hasattr(fp, "read"): + return await asyncio.to_thread(blake3_hash, fp, chunk_size) + + def _worker() -> str: + with open(os.fspath(fp), "rb") as f: + return _hash_file_obj(f, chunk_size) + + return await asyncio.to_thread(_worker) + + +def _hash_file_obj(file_obj: IO, chunk_size: int = DEFAULT_CHUNK) -> str: + """ + Hash an already-open binary file object by streaming in chunks. + - Seeks to the beginning before reading (if supported). + - Restores the original position afterward (if tell/seek are supported). + """ + if chunk_size <= 0: + chunk_size = DEFAULT_CHUNK + + # in case file object is already open and not at the beginning, track so can be restored after hashing + orig_pos = file_obj.tell() + + try: + # seek to the beginning before reading + if orig_pos != 0: + file_obj.seek(0) + + h = blake3() + while True: + chunk = file_obj.read(chunk_size) + if not chunk: + break + h.update(chunk) + return h.hexdigest() + finally: + # restore original position in file object, if needed + if orig_pos != 0: + file_obj.seek(orig_pos) diff --git a/app/assets/helpers.py b/app/assets/helpers.py new file mode 100644 index 000000000..6755d0e56 --- /dev/null +++ b/app/assets/helpers.py @@ -0,0 +1,216 @@ +import contextlib +import os +from aiohttp import web +from datetime import datetime, timezone +from pathlib import Path +from typing import Literal, Any + +import folder_paths + + +RootType = Literal["models", "input", "output"] +ALLOWED_ROOTS: tuple[RootType, ...] = ("models", "input", "output") + +def get_query_dict(request: web.Request) -> dict[str, Any]: + """ + Gets a dictionary of query parameters from the request. + + 'request.query' is a MultiMapping[str], needs to be converted to a dictionary to be validated by Pydantic. + """ + query_dict = { + key: request.query.getall(key) if len(request.query.getall(key)) > 1 else request.query.get(key) + for key in request.query.keys() + } + return query_dict + +def list_tree(base_dir: str) -> list[str]: + out: list[str] = [] + base_abs = os.path.abspath(base_dir) + if not os.path.isdir(base_abs): + return out + for dirpath, _subdirs, filenames in os.walk(base_abs, topdown=True, followlinks=False): + for name in filenames: + out.append(os.path.abspath(os.path.join(dirpath, name))) + return out + +def prefixes_for_root(root: RootType) -> list[str]: + if root == "models": + bases: list[str] = [] + for _bucket, paths in get_comfy_models_folders(): + bases.extend(paths) + return [os.path.abspath(p) for p in bases] + if root == "input": + return [os.path.abspath(folder_paths.get_input_directory())] + if root == "output": + return [os.path.abspath(folder_paths.get_output_directory())] + return [] + +def escape_like_prefix(s: str, escape: str = "!") -> tuple[str, str]: + """Escapes %, _ and the escape char itself in a LIKE prefix. + Returns (escaped_prefix, escape_char). Caller should append '%' and pass escape=escape_char to .like(). + """ + s = s.replace(escape, escape + escape) # escape the escape char first + s = s.replace("%", escape + "%").replace("_", escape + "_") # escape LIKE wildcards + return s, escape + +def fast_asset_file_check( + *, + mtime_db: int | None, + size_db: int | None, + stat_result: os.stat_result, +) -> bool: + if mtime_db is None: + return False + actual_mtime_ns = getattr(stat_result, "st_mtime_ns", int(stat_result.st_mtime * 1_000_000_000)) + if int(mtime_db) != int(actual_mtime_ns): + return False + sz = int(size_db or 0) + if sz > 0: + return int(stat_result.st_size) == sz + return True + +def utcnow() -> datetime: + """Naive UTC timestamp (no tzinfo). We always treat DB datetimes as UTC.""" + return datetime.now(timezone.utc).replace(tzinfo=None) + +def get_comfy_models_folders() -> list[tuple[str, list[str]]]: + """Build a list of (folder_name, base_paths[]) categories that are configured for model locations. + + We trust `folder_paths.folder_names_and_paths` and include a category if + *any* of its base paths lies under the Comfy `models_dir`. + """ + targets: list[tuple[str, list[str]]] = [] + models_root = os.path.abspath(folder_paths.models_dir) + for name, (paths, _exts) in folder_paths.folder_names_and_paths.items(): + if any(os.path.abspath(p).startswith(models_root + os.sep) for p in paths): + targets.append((name, paths)) + return targets + +def compute_relative_filename(file_path: str) -> str | None: + """ + Return the model's path relative to the last well-known folder (the model category), + using forward slashes, eg: + /.../models/checkpoints/flux/123/flux.safetensors -> "flux/123/flux.safetensors" + /.../models/text_encoders/clip_g.safetensors -> "clip_g.safetensors" + + For non-model paths, returns None. + NOTE: this is a temporary helper, used only for initializing metadata["filename"] field. + """ + try: + root_category, rel_path = get_relative_to_root_category_path_of_asset(file_path) + except ValueError: + return None + + p = Path(rel_path) + parts = [seg for seg in p.parts if seg not in (".", "..", p.anchor)] + if not parts: + return None + + if root_category == "models": + # parts[0] is the category ("checkpoints", "vae", etc) – drop it + inside = parts[1:] if len(parts) > 1 else [parts[0]] + return "/".join(inside) + return "/".join(parts) # input/output: keep all parts + + +def get_relative_to_root_category_path_of_asset(file_path: str) -> tuple[Literal["input", "output", "models"], str]: + """Given an absolute or relative file path, determine which root category the path belongs to: + - 'input' if the file resides under `folder_paths.get_input_directory()` + - 'output' if the file resides under `folder_paths.get_output_directory()` + - 'models' if the file resides under any base path of categories returned by `get_comfy_models_folders()` + + Returns: + (root_category, relative_path_inside_that_root) + For 'models', the relative path is prefixed with the category name: + e.g. ('models', 'vae/test/sub/ae.safetensors') + + Raises: + ValueError: if the path does not belong to input, output, or configured model bases. + """ + fp_abs = os.path.abspath(file_path) + + def _is_within(child: str, parent: str) -> bool: + try: + return os.path.commonpath([child, parent]) == parent + except Exception: + return False + + def _rel(child: str, parent: str) -> str: + return os.path.relpath(os.path.join(os.sep, os.path.relpath(child, parent)), os.sep) + + # 1) input + input_base = os.path.abspath(folder_paths.get_input_directory()) + if _is_within(fp_abs, input_base): + return "input", _rel(fp_abs, input_base) + + # 2) output + output_base = os.path.abspath(folder_paths.get_output_directory()) + if _is_within(fp_abs, output_base): + return "output", _rel(fp_abs, output_base) + + # 3) models (check deepest matching base to avoid ambiguity) + best: tuple[int, str, str] | None = None # (base_len, bucket, rel_inside_bucket) + for bucket, bases in get_comfy_models_folders(): + for b in bases: + base_abs = os.path.abspath(b) + if not _is_within(fp_abs, base_abs): + continue + cand = (len(base_abs), bucket, _rel(fp_abs, base_abs)) + if best is None or cand[0] > best[0]: + best = cand + + if best is not None: + _, bucket, rel_inside = best + combined = os.path.join(bucket, rel_inside) + return "models", os.path.relpath(os.path.join(os.sep, combined), os.sep) + + raise ValueError(f"Path is not within input, output, or configured model bases: {file_path}") + +def get_name_and_tags_from_asset_path(file_path: str) -> tuple[str, list[str]]: + """Return a tuple (name, tags) derived from a filesystem path. + + Semantics: + - Root category is determined by `get_relative_to_root_category_path_of_asset`. + - The returned `name` is the base filename with extension from the relative path. + - The returned `tags` are: + [root_category] + parent folders of the relative path (in order) + For 'models', this means: + file '/.../ModelsDir/vae/test_tag/ae.safetensors' + -> root_category='models', some_path='vae/test_tag/ae.safetensors' + -> name='ae.safetensors', tags=['models', 'vae', 'test_tag'] + + Raises: + ValueError: if the path does not belong to input, output, or configured model bases. + """ + root_category, some_path = get_relative_to_root_category_path_of_asset(file_path) + p = Path(some_path) + parent_parts = [part for part in p.parent.parts if part not in (".", "..", p.anchor)] + return p.name, list(dict.fromkeys(normalize_tags([root_category, *parent_parts]))) + +def normalize_tags(tags: list[str] | None) -> list[str]: + """ + Normalize a list of tags by: + - Stripping whitespace and converting to lowercase. + - Removing duplicates. + """ + return [t.strip().lower() for t in (tags or []) if (t or "").strip()] + +def collect_models_files() -> list[str]: + out: list[str] = [] + for folder_name, bases in get_comfy_models_folders(): + rel_files = folder_paths.get_filename_list(folder_name) or [] + for rel_path in rel_files: + abs_path = folder_paths.get_full_path(folder_name, rel_path) + if not abs_path: + continue + abs_path = os.path.abspath(abs_path) + allowed = False + for b in bases: + base_abs = os.path.abspath(b) + with contextlib.suppress(Exception): + if os.path.commonpath([abs_path, base_abs]) == base_abs: + allowed = True + break + if allowed: + out.append(abs_path) + return out diff --git a/app/assets/manager.py b/app/assets/manager.py new file mode 100644 index 000000000..6425e7aa2 --- /dev/null +++ b/app/assets/manager.py @@ -0,0 +1,123 @@ +from typing import Sequence + +from app.database.db import create_session +from app.assets.api import schemas_out +from app.assets.database.queries import ( + asset_exists_by_hash, + fetch_asset_info_asset_and_tags, + list_asset_infos_page, + list_tags_with_usage, +) + + +def _safe_sort_field(requested: str | None) -> str: + if not requested: + return "created_at" + v = requested.lower() + if v in {"name", "created_at", "updated_at", "size", "last_access_time"}: + return v + return "created_at" + + +def asset_exists(asset_hash: str) -> bool: + with create_session() as session: + return asset_exists_by_hash(session, asset_hash=asset_hash) + +def list_assets( + include_tags: Sequence[str] | None = None, + exclude_tags: Sequence[str] | None = None, + name_contains: str | None = None, + metadata_filter: dict | None = None, + limit: int = 20, + offset: int = 0, + sort: str = "created_at", + order: str = "desc", + owner_id: str = "", +) -> schemas_out.AssetsList: + sort = _safe_sort_field(sort) + order = "desc" if (order or "desc").lower() not in {"asc", "desc"} else order.lower() + + with create_session() as session: + infos, tag_map, total = list_asset_infos_page( + session, + owner_id=owner_id, + include_tags=include_tags, + exclude_tags=exclude_tags, + name_contains=name_contains, + metadata_filter=metadata_filter, + limit=limit, + offset=offset, + sort=sort, + order=order, + ) + + summaries: list[schemas_out.AssetSummary] = [] + for info in infos: + asset = info.asset + tags = tag_map.get(info.id, []) + summaries.append( + schemas_out.AssetSummary( + id=info.id, + name=info.name, + asset_hash=asset.hash if asset else None, + size=int(asset.size_bytes) if asset else None, + mime_type=asset.mime_type if asset else None, + tags=tags, + preview_url=f"/api/assets/{info.id}/content", + created_at=info.created_at, + updated_at=info.updated_at, + last_access_time=info.last_access_time, + ) + ) + + return schemas_out.AssetsList( + assets=summaries, + total=total, + has_more=(offset + len(summaries)) < total, + ) + +def get_asset(asset_info_id: str, owner_id: str = "") -> schemas_out.AssetDetail: + with create_session() as session: + res = fetch_asset_info_asset_and_tags(session, asset_info_id=asset_info_id, owner_id=owner_id) + if not res: + raise ValueError(f"AssetInfo {asset_info_id} not found") + info, asset, tag_names = res + preview_id = info.preview_id + + return schemas_out.AssetDetail( + id=info.id, + name=info.name, + asset_hash=asset.hash if asset else None, + size=int(asset.size_bytes) if asset and asset.size_bytes is not None else None, + mime_type=asset.mime_type if asset else None, + tags=tag_names, + user_metadata=info.user_metadata or {}, + preview_id=preview_id, + created_at=info.created_at, + last_access_time=info.last_access_time, + ) + +def list_tags( + prefix: str | None = None, + limit: int = 100, + offset: int = 0, + order: str = "count_desc", + include_zero: bool = True, + owner_id: str = "", +) -> schemas_out.TagsList: + limit = max(1, min(1000, limit)) + offset = max(0, offset) + + with create_session() as session: + rows, total = list_tags_with_usage( + session, + prefix=prefix, + limit=limit, + offset=offset, + include_zero=include_zero, + order=order, + owner_id=owner_id, + ) + + tags = [schemas_out.TagUsage(name=name, count=count, type=tag_type) for (name, tag_type, count) in rows] + return schemas_out.TagsList(tags=tags, total=total, has_more=(offset + len(tags)) < total) diff --git a/app/assets/scanner.py b/app/assets/scanner.py new file mode 100644 index 000000000..a16e41d94 --- /dev/null +++ b/app/assets/scanner.py @@ -0,0 +1,229 @@ +import contextlib +import time +import logging +import os +import sqlalchemy + +import folder_paths +from app.database.db import create_session, dependencies_available +from app.assets.helpers import ( + collect_models_files, compute_relative_filename, fast_asset_file_check, get_name_and_tags_from_asset_path, + list_tree,prefixes_for_root, escape_like_prefix, + RootType +) +from app.assets.database.tags import add_missing_tag_for_asset_id, ensure_tags_exist, remove_missing_tag_for_asset_id +from app.assets.database.bulk_ops import seed_from_paths_batch +from app.assets.database.models import Asset, AssetCacheState, AssetInfo + + +def seed_assets(roots: tuple[RootType, ...], enable_logging: bool = False) -> None: + """ + Scan the given roots and seed the assets into the database. + """ + if not dependencies_available(): + if enable_logging: + logging.warning("Database dependencies not available, skipping assets scan") + return + t_start = time.perf_counter() + created = 0 + skipped_existing = 0 + paths: list[str] = [] + try: + existing_paths: set[str] = set() + for r in roots: + try: + survivors: set[str] = _fast_db_consistency_pass(r, collect_existing_paths=True, update_missing_tags=True) + if survivors: + existing_paths.update(survivors) + except Exception as e: + logging.exception("fast DB scan failed for %s: %s", r, e) + + if "models" in roots: + paths.extend(collect_models_files()) + if "input" in roots: + paths.extend(list_tree(folder_paths.get_input_directory())) + if "output" in roots: + paths.extend(list_tree(folder_paths.get_output_directory())) + + specs: list[dict] = [] + tag_pool: set[str] = set() + for p in paths: + abs_p = os.path.abspath(p) + if abs_p in existing_paths: + skipped_existing += 1 + continue + try: + stat_p = os.stat(abs_p, follow_symlinks=False) + except OSError: + continue + # skip empty files + if not stat_p.st_size: + continue + name, tags = get_name_and_tags_from_asset_path(abs_p) + specs.append( + { + "abs_path": abs_p, + "size_bytes": stat_p.st_size, + "mtime_ns": getattr(stat_p, "st_mtime_ns", int(stat_p.st_mtime * 1_000_000_000)), + "info_name": name, + "tags": tags, + "fname": compute_relative_filename(abs_p), + } + ) + for t in tags: + tag_pool.add(t) + # if no file specs, nothing to do + if not specs: + return + with create_session() as sess: + if tag_pool: + ensure_tags_exist(sess, tag_pool, tag_type="user") + + result = seed_from_paths_batch(sess, specs=specs, owner_id="") + created += result["inserted_infos"] + sess.commit() + finally: + if enable_logging: + logging.info( + "Assets scan(roots=%s) completed in %.3fs (created=%d, skipped_existing=%d, total_seen=%d)", + roots, + time.perf_counter() - t_start, + created, + skipped_existing, + len(paths), + ) + + +def _fast_db_consistency_pass( + root: RootType, + *, + collect_existing_paths: bool = False, + update_missing_tags: bool = False, +) -> set[str] | None: + """Fast DB+FS pass for a root: + - Toggle needs_verify per state using fast check + - For hashed assets with at least one fast-ok state in this root: delete stale missing states + - For seed assets with all states missing: delete Asset and its AssetInfos + - Optionally add/remove 'missing' tags based on fast-ok in this root + - Optionally return surviving absolute paths + """ + prefixes = prefixes_for_root(root) + if not prefixes: + return set() if collect_existing_paths else None + + conds = [] + for p in prefixes: + base = os.path.abspath(p) + if not base.endswith(os.sep): + base += os.sep + escaped, esc = escape_like_prefix(base) + conds.append(AssetCacheState.file_path.like(escaped + "%", escape=esc)) + + with create_session() as sess: + rows = ( + sess.execute( + sqlalchemy.select( + AssetCacheState.id, + AssetCacheState.file_path, + AssetCacheState.mtime_ns, + AssetCacheState.needs_verify, + AssetCacheState.asset_id, + Asset.hash, + Asset.size_bytes, + ) + .join(Asset, Asset.id == AssetCacheState.asset_id) + .where(sqlalchemy.or_(*conds)) + .order_by(AssetCacheState.asset_id.asc(), AssetCacheState.id.asc()) + ) + ).all() + + by_asset: dict[str, dict] = {} + for sid, fp, mtime_db, needs_verify, aid, a_hash, a_size in rows: + acc = by_asset.get(aid) + if acc is None: + acc = {"hash": a_hash, "size_db": int(a_size or 0), "states": []} + by_asset[aid] = acc + + fast_ok = False + try: + exists = True + fast_ok = fast_asset_file_check( + mtime_db=mtime_db, + size_db=acc["size_db"], + stat_result=os.stat(fp, follow_symlinks=True), + ) + except FileNotFoundError: + exists = False + except OSError: + exists = False + + acc["states"].append({ + "sid": sid, + "fp": fp, + "exists": exists, + "fast_ok": fast_ok, + "needs_verify": bool(needs_verify), + }) + + to_set_verify: list[int] = [] + to_clear_verify: list[int] = [] + stale_state_ids: list[int] = [] + survivors: set[str] = set() + + for aid, acc in by_asset.items(): + a_hash = acc["hash"] + states = acc["states"] + any_fast_ok = any(s["fast_ok"] for s in states) + all_missing = all(not s["exists"] for s in states) + + for s in states: + if not s["exists"]: + continue + if s["fast_ok"] and s["needs_verify"]: + to_clear_verify.append(s["sid"]) + if not s["fast_ok"] and not s["needs_verify"]: + to_set_verify.append(s["sid"]) + + if a_hash is None: + if states and all_missing: # remove seed Asset completely, if no valid AssetCache exists + sess.execute(sqlalchemy.delete(AssetInfo).where(AssetInfo.asset_id == aid)) + asset = sess.get(Asset, aid) + if asset: + sess.delete(asset) + else: + for s in states: + if s["exists"]: + survivors.add(os.path.abspath(s["fp"])) + continue + + if any_fast_ok: # if Asset has at least one valid AssetCache record, remove any invalid AssetCache records + for s in states: + if not s["exists"]: + stale_state_ids.append(s["sid"]) + if update_missing_tags: + with contextlib.suppress(Exception): + remove_missing_tag_for_asset_id(sess, asset_id=aid) + elif update_missing_tags: + with contextlib.suppress(Exception): + add_missing_tag_for_asset_id(sess, asset_id=aid, origin="automatic") + + for s in states: + if s["exists"]: + survivors.add(os.path.abspath(s["fp"])) + + if stale_state_ids: + sess.execute(sqlalchemy.delete(AssetCacheState).where(AssetCacheState.id.in_(stale_state_ids))) + if to_set_verify: + sess.execute( + sqlalchemy.update(AssetCacheState) + .where(AssetCacheState.id.in_(to_set_verify)) + .values(needs_verify=True) + ) + if to_clear_verify: + sess.execute( + sqlalchemy.update(AssetCacheState) + .where(AssetCacheState.id.in_(to_clear_verify)) + .values(needs_verify=False) + ) + sess.commit() + return survivors if collect_existing_paths else None diff --git a/app/database/models.py b/app/database/models.py index 6facfb8f2..e7572677a 100644 --- a/app/database/models.py +++ b/app/database/models.py @@ -1,14 +1,21 @@ -from sqlalchemy.orm import declarative_base +from typing import Any +from datetime import datetime +from sqlalchemy.orm import DeclarativeBase -Base = declarative_base() +class Base(DeclarativeBase): + pass - -def to_dict(obj): +def to_dict(obj: Any, include_none: bool = False) -> dict[str, Any]: fields = obj.__table__.columns.keys() - return { - field: (val.to_dict() if hasattr(val, "to_dict") else val) - for field in fields - if (val := getattr(obj, field)) - } + out: dict[str, Any] = {} + for field in fields: + val = getattr(obj, field) + if val is None and not include_none: + continue + if isinstance(val, datetime): + out[field] = val.isoformat() + else: + out[field] = val + return out # TODO: Define models here diff --git a/comfy/cli_args.py b/comfy/cli_args.py index dae9a895d..1716c3de7 100644 --- a/comfy/cli_args.py +++ b/comfy/cli_args.py @@ -231,6 +231,7 @@ database_default_path = os.path.abspath( os.path.join(os.path.dirname(__file__), "..", "user", "comfyui.db") ) parser.add_argument("--database-url", type=str, default=f"sqlite:///{database_default_path}", help="Specify the database URL, e.g. for an in-memory database you can use 'sqlite:///:memory:'.") +parser.add_argument("--disable-assets-autoscan", action="store_true", help="Disable asset scanning on startup for database synchronization.") if comfy.options.args_parsing: args = parser.parse_args() diff --git a/main.py b/main.py index 0e07a95da..37b06c1fa 100644 --- a/main.py +++ b/main.py @@ -7,6 +7,7 @@ import folder_paths import time from comfy.cli_args import args from app.logger import setup_logger +from app.assets.scanner import seed_assets import itertools import utils.extra_config import logging @@ -324,6 +325,8 @@ def setup_database(): from app.database.db import init_db, dependencies_available if dependencies_available(): init_db() + if not args.disable_assets_autoscan: + seed_assets(["models"], enable_logging=True) except Exception as e: logging.error(f"Failed to initialize database. Please ensure you have installed the latest requirements. If the error persists, please report this as in future the database will be required: {e}") diff --git a/server.py b/server.py index 4db3347cb..da2baefd4 100644 --- a/server.py +++ b/server.py @@ -33,6 +33,8 @@ import node_helpers from comfyui_version import __version__ from app.frontend_management import FrontendManager, parse_version from comfy_api.internal import _ComfyNodeInternal +from app.assets.scanner import seed_assets +from app.assets.api.routes import register_assets_system from app.user_manager import UserManager from app.model_manager import ModelFileManager @@ -235,6 +237,7 @@ class PromptServer(): else args.front_end_root ) logging.info(f"[Prompt Server] web root: {self.web_root}") + register_assets_system(self.app, self.user_manager) routes = web.RouteTableDef() self.routes = routes self.last_node_id = None @@ -683,6 +686,7 @@ class PromptServer(): @routes.get("/object_info") async def get_object_info(request): + seed_assets(["models"]) with folder_paths.cache_helper: out = {} for x in nodes.NODE_CLASS_MAPPINGS: From 6207f86c18d2cf2d70ab059987b62d4b38466e77 Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Thu, 8 Jan 2026 20:34:48 -0800 Subject: [PATCH 45/48] Fix VAEEncodeForInpaint to support WAN VAE tuple downscale_ratio (#11572) Use vae.spacial_compression_encode() instead of directly accessing downscale_ratio to handle both standard VAEs (int) and WAN VAEs (tuple). Addresses reviewer feedback on PR #11259. Co-authored-by: ChrisFab16 --- nodes.py | 9 +++++---- 1 file changed, 5 insertions(+), 4 deletions(-) diff --git a/nodes.py b/nodes.py index 56b74ebe3..1aa391f4a 100644 --- a/nodes.py +++ b/nodes.py @@ -378,14 +378,15 @@ class VAEEncodeForInpaint: CATEGORY = "latent/inpaint" def encode(self, vae, pixels, mask, grow_mask_by=6): - x = (pixels.shape[1] // vae.downscale_ratio) * vae.downscale_ratio - y = (pixels.shape[2] // vae.downscale_ratio) * vae.downscale_ratio + downscale_ratio = vae.spacial_compression_encode() + x = (pixels.shape[1] // downscale_ratio) * downscale_ratio + y = (pixels.shape[2] // downscale_ratio) * downscale_ratio mask = torch.nn.functional.interpolate(mask.reshape((-1, 1, mask.shape[-2], mask.shape[-1])), size=(pixels.shape[1], pixels.shape[2]), mode="bilinear") pixels = pixels.clone() if pixels.shape[1] != x or pixels.shape[2] != y: - x_offset = (pixels.shape[1] % vae.downscale_ratio) // 2 - y_offset = (pixels.shape[2] % vae.downscale_ratio) // 2 + x_offset = (pixels.shape[1] % downscale_ratio) // 2 + y_offset = (pixels.shape[2] % downscale_ratio) // 2 pixels = pixels[:,x_offset:x + x_offset, y_offset:y + y_offset,:] mask = mask[:,:,x_offset:x + x_offset, y_offset:y + y_offset] From 4609fcd26081156eef921bd9f43726f670ee6f51 Mon Sep 17 00:00:00 2001 From: Terry Jia Date: Fri, 9 Jan 2026 00:31:19 -0500 Subject: [PATCH 46/48] add node - image compare (#11343) --- comfy_api/latest/_io.py | 13 +++++++ comfy_extras/nodes_image_compare.py | 53 +++++++++++++++++++++++++++++ nodes.py | 1 + 3 files changed, 67 insertions(+) create mode 100644 comfy_extras/nodes_image_compare.py diff --git a/comfy_api/latest/_io.py b/comfy_api/latest/_io.py index 764fa8b2b..50143ff53 100644 --- a/comfy_api/latest/_io.py +++ b/comfy_api/latest/_io.py @@ -1113,6 +1113,18 @@ class DynamicSlot(ComfyTypeI): out_dict[input_type][finalized_id] = value out_dict["dynamic_paths"][finalized_id] = finalize_prefix(curr_prefix, curr_prefix[-1]) +@comfytype(io_type="IMAGECOMPARE") +class ImageCompare(ComfyTypeI): + Type = dict + + class Input(WidgetInput): + def __init__(self, id: str, display_name: str=None, optional=False, tooltip: str=None, + socketless: bool=True): + super().__init__(id, display_name, optional, tooltip, None, None, socketless) + + def as_dict(self): + return super().as_dict() + DYNAMIC_INPUT_LOOKUP: dict[str, Callable[[dict[str, Any], dict[str, Any], tuple[str, dict[str, Any]], str, list[str] | None], None]] = {} def register_dynamic_input_func(io_type: str, func: Callable[[dict[str, Any], dict[str, Any], tuple[str, dict[str, Any]], str, list[str] | None], None]): DYNAMIC_INPUT_LOOKUP[io_type] = func @@ -1958,4 +1970,5 @@ __all__ = [ "add_to_dict_v1", "add_to_dict_v3", "V3Data", + "ImageCompare", ] diff --git a/comfy_extras/nodes_image_compare.py b/comfy_extras/nodes_image_compare.py new file mode 100644 index 000000000..8e9f809e6 --- /dev/null +++ b/comfy_extras/nodes_image_compare.py @@ -0,0 +1,53 @@ +import nodes + +from typing_extensions import override +from comfy_api.latest import IO, ComfyExtension + + +class ImageCompare(IO.ComfyNode): + """Compares two images with a slider interface.""" + + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="ImageCompare", + display_name="Image Compare", + description="Compares two images side by side with a slider.", + category="image", + is_experimental=True, + is_output_node=True, + inputs=[ + IO.Image.Input("image_a", optional=True), + IO.Image.Input("image_b", optional=True), + IO.ImageCompare.Input("compare_view"), + ], + outputs=[], + ) + + @classmethod + def execute(cls, image_a=None, image_b=None, compare_view=None) -> IO.NodeOutput: + result = {"a_images": [], "b_images": []} + + preview_node = nodes.PreviewImage() + + if image_a is not None and len(image_a) > 0: + saved = preview_node.save_images(image_a, "comfy.compare.a") + result["a_images"] = saved["ui"]["images"] + + if image_b is not None and len(image_b) > 0: + saved = preview_node.save_images(image_b, "comfy.compare.b") + result["b_images"] = saved["ui"]["images"] + + return IO.NodeOutput(ui=result) + + +class ImageCompareExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[IO.ComfyNode]]: + return [ + ImageCompare, + ] + + +async def comfy_entrypoint() -> ImageCompareExtension: + return ImageCompareExtension() diff --git a/nodes.py b/nodes.py index 1aa391f4a..5a9d42d4a 100644 --- a/nodes.py +++ b/nodes.py @@ -2370,6 +2370,7 @@ async def init_builtin_extra_nodes(): "nodes_nop.py", "nodes_kandinsky5.py", "nodes_wanmove.py", + "nodes_image_compare.py", ] import_failed = [] From 04c49a29b493f3f9037b83cec45f6369b5c4816b Mon Sep 17 00:00:00 2001 From: ric-yu Date: Thu, 8 Jan 2026 21:57:36 -0800 Subject: [PATCH 47/48] feat: add cancelled filter to /jobs (#11680) --- comfy_execution/jobs.py | 31 +++++++++++++++++------------ tests/execution/test_jobs.py | 38 +++++++++++++++++++++++++++++++++++- 2 files changed, 55 insertions(+), 14 deletions(-) diff --git a/comfy_execution/jobs.py b/comfy_execution/jobs.py index 59fb49357..97fd803b8 100644 --- a/comfy_execution/jobs.py +++ b/comfy_execution/jobs.py @@ -14,8 +14,9 @@ class JobStatus: IN_PROGRESS = 'in_progress' COMPLETED = 'completed' FAILED = 'failed' + CANCELLED = 'cancelled' - ALL = [PENDING, IN_PROGRESS, COMPLETED, FAILED] + ALL = [PENDING, IN_PROGRESS, COMPLETED, FAILED, CANCELLED] # Media types that can be previewed in the frontend @@ -94,12 +95,6 @@ def normalize_history_item(prompt_id: str, history_item: dict, include_outputs: status_info = history_item.get('status', {}) status_str = status_info.get('status_str') if status_info else None - if status_str == 'success': - status = JobStatus.COMPLETED - elif status_str == 'error': - status = JobStatus.FAILED - else: - status = JobStatus.COMPLETED outputs = history_item.get('outputs', {}) outputs_count, preview_output = get_outputs_summary(outputs) @@ -107,6 +102,7 @@ def normalize_history_item(prompt_id: str, history_item: dict, include_outputs: execution_error = None execution_start_time = None execution_end_time = None + was_interrupted = False if status_info: messages = status_info.get('messages', []) for entry in messages: @@ -119,6 +115,15 @@ def normalize_history_item(prompt_id: str, history_item: dict, include_outputs: execution_end_time = event_data.get('timestamp') if event_name == 'execution_error': execution_error = event_data + elif event_name == 'execution_interrupted': + was_interrupted = True + + if status_str == 'success': + status = JobStatus.COMPLETED + elif status_str == 'error': + status = JobStatus.CANCELLED if was_interrupted else JobStatus.FAILED + else: + status = JobStatus.COMPLETED job = prune_dict({ 'id': prompt_id, @@ -268,13 +273,13 @@ def get_all_jobs( for item in queued: jobs.append(normalize_queue_item(item, JobStatus.PENDING)) - include_completed = JobStatus.COMPLETED in status_filter - include_failed = JobStatus.FAILED in status_filter - if include_completed or include_failed: + history_statuses = {JobStatus.COMPLETED, JobStatus.FAILED, JobStatus.CANCELLED} + requested_history_statuses = history_statuses & set(status_filter) + if requested_history_statuses: for prompt_id, history_item in history.items(): - is_failed = history_item.get('status', {}).get('status_str') == 'error' - if (is_failed and include_failed) or (not is_failed and include_completed): - jobs.append(normalize_history_item(prompt_id, history_item)) + job = normalize_history_item(prompt_id, history_item) + if job.get('status') in requested_history_statuses: + jobs.append(job) if workflow_id: jobs = [j for j in jobs if j.get('workflow_id') == workflow_id] diff --git a/tests/execution/test_jobs.py b/tests/execution/test_jobs.py index 918c8080a..4d2f9ed36 100644 --- a/tests/execution/test_jobs.py +++ b/tests/execution/test_jobs.py @@ -19,6 +19,7 @@ class TestJobStatus: assert JobStatus.IN_PROGRESS == 'in_progress' assert JobStatus.COMPLETED == 'completed' assert JobStatus.FAILED == 'failed' + assert JobStatus.CANCELLED == 'cancelled' def test_all_contains_all_statuses(self): """ALL should contain all status values.""" @@ -26,7 +27,8 @@ class TestJobStatus: assert JobStatus.IN_PROGRESS in JobStatus.ALL assert JobStatus.COMPLETED in JobStatus.ALL assert JobStatus.FAILED in JobStatus.ALL - assert len(JobStatus.ALL) == 4 + assert JobStatus.CANCELLED in JobStatus.ALL + assert len(JobStatus.ALL) == 5 class TestIsPreviewable: @@ -336,6 +338,40 @@ class TestNormalizeHistoryItem: assert job['execution_error']['node_type'] == 'KSampler' assert job['execution_error']['exception_message'] == 'CUDA out of memory' + def test_cancelled_job(self): + """Cancelled/interrupted history item should have cancelled status.""" + history_item = { + 'prompt': ( + 5, + 'prompt-cancelled', + {'nodes': {}}, + {'create_time': 1234567890000}, + ['node1'], + ), + 'status': { + 'status_str': 'error', + 'completed': False, + 'messages': [ + ('execution_start', {'prompt_id': 'prompt-cancelled', 'timestamp': 1234567890500}), + ('execution_interrupted', { + 'prompt_id': 'prompt-cancelled', + 'node_id': '5', + 'node_type': 'KSampler', + 'executed': ['1', '2', '3'], + 'timestamp': 1234567891000, + }) + ] + }, + 'outputs': {}, + } + + job = normalize_history_item('prompt-cancelled', history_item) + assert job['status'] == 'cancelled' + assert job['execution_start_time'] == 1234567890500 + assert job['execution_end_time'] == 1234567891000 + # Cancelled jobs should not have execution_error set + assert 'execution_error' not in job + def test_include_outputs(self): """When include_outputs=True, should include full output data.""" history_item = { From ec0a832acb25fbe53bd4fc25d286a9ee442a3bcf Mon Sep 17 00:00:00 2001 From: Jedrzej Kosinski Date: Thu, 8 Jan 2026 22:49:12 -0800 Subject: [PATCH 48/48] Add workaround for hacky nodepack(s) that edit folder_names_and_paths to have values with tuples of more than 2. Other things could potentially break with those nodepack(s), so I will hunt for the guilty nodepack(s) now. (#11755) --- app/assets/helpers.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/app/assets/helpers.py b/app/assets/helpers.py index 6755d0e56..08b465b5a 100644 --- a/app/assets/helpers.py +++ b/app/assets/helpers.py @@ -81,7 +81,8 @@ def get_comfy_models_folders() -> list[tuple[str, list[str]]]: """ targets: list[tuple[str, list[str]]] = [] models_root = os.path.abspath(folder_paths.models_dir) - for name, (paths, _exts) in folder_paths.folder_names_and_paths.items(): + for name, values in folder_paths.folder_names_and_paths.items(): + paths, _exts = values[0], values[1] # NOTE: this prevents nodepacks that hackily edit folder_... from breaking ComfyUI if any(os.path.abspath(p).startswith(models_root + os.sep) for p in paths): targets.append((name, paths)) return targets