From a6f83a4a1a70d720c16d66feb5d87fee5998acdf Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 1 Oct 2025 14:19:13 -0700 Subject: [PATCH 1/5] Support the new hunyuan vae. (#10150) --- comfy/ldm/hunyuan_video/vae_refiner.py | 112 ++++++++++++++++--------- comfy/sd.py | 70 ++++++++++------ 2 files changed, 116 insertions(+), 66 deletions(-) diff --git a/comfy/ldm/hunyuan_video/vae_refiner.py b/comfy/ldm/hunyuan_video/vae_refiner.py index c6f742710..c2a0b507d 100644 --- a/comfy/ldm/hunyuan_video/vae_refiner.py +++ b/comfy/ldm/hunyuan_video/vae_refiner.py @@ -1,7 +1,7 @@ import torch import torch.nn as nn import torch.nn.functional as F -from comfy.ldm.modules.diffusionmodules.model import ResnetBlock, AttnBlock, VideoConv3d +from comfy.ldm.modules.diffusionmodules.model import ResnetBlock, AttnBlock, VideoConv3d, Normalize import comfy.ops import comfy.ldm.models.autoencoder ops = comfy.ops.disable_weight_init @@ -17,11 +17,12 @@ class RMS_norm(nn.Module): return F.normalize(x, dim=1) * self.scale * self.gamma class DnSmpl(nn.Module): - def __init__(self, ic, oc, tds=True): + def __init__(self, ic, oc, tds=True, refiner_vae=True, op=VideoConv3d): super().__init__() fct = 2 * 2 * 2 if tds else 1 * 2 * 2 assert oc % fct == 0 - self.conv = VideoConv3d(ic, oc // fct, kernel_size=3) + self.conv = op(ic, oc // fct, kernel_size=3, stride=1, padding=1) + self.refiner_vae = refiner_vae self.tds = tds self.gs = fct * ic // oc @@ -30,7 +31,7 @@ class DnSmpl(nn.Module): r1 = 2 if self.tds else 1 h = self.conv(x) - if self.tds: + if self.tds and self.refiner_vae: hf = h[:, :, :1, :, :] b, c, f, ht, wd = hf.shape hf = hf.reshape(b, c, f, ht // 2, 2, wd // 2, 2) @@ -66,6 +67,7 @@ class DnSmpl(nn.Module): sc = torch.cat([xf, xn], dim=2) else: b, c, frms, ht, wd = h.shape + nf = frms // r1 h = h.reshape(b, c, nf, r1, ht // 2, 2, wd // 2, 2) h = h.permute(0, 3, 5, 7, 1, 2, 4, 6) @@ -83,10 +85,11 @@ class DnSmpl(nn.Module): class UpSmpl(nn.Module): - def __init__(self, ic, oc, tus=True): + def __init__(self, ic, oc, tus=True, refiner_vae=True, op=VideoConv3d): super().__init__() fct = 2 * 2 * 2 if tus else 1 * 2 * 2 - self.conv = VideoConv3d(ic, oc * fct, kernel_size=3) + self.conv = op(ic, oc * fct, kernel_size=3, stride=1, padding=1) + self.refiner_vae = refiner_vae self.tus = tus self.rp = fct * oc // ic @@ -95,7 +98,7 @@ class UpSmpl(nn.Module): r1 = 2 if self.tus else 1 h = self.conv(x) - if self.tus: + if self.tus and self.refiner_vae: hf = h[:, :, :1, :, :] b, c, f, ht, wd = hf.shape nc = c // (2 * 2) @@ -148,43 +151,56 @@ class UpSmpl(nn.Module): class Encoder(nn.Module): def __init__(self, in_channels, z_channels, block_out_channels, num_res_blocks, - ffactor_spatial, ffactor_temporal, downsample_match_channel=True, **_): + ffactor_spatial, ffactor_temporal, downsample_match_channel=True, refiner_vae=True, **_): super().__init__() self.z_channels = z_channels self.block_out_channels = block_out_channels self.num_res_blocks = num_res_blocks - self.conv_in = VideoConv3d(in_channels, block_out_channels[0], 3, 1, 1) + self.ffactor_temporal = ffactor_temporal + + self.refiner_vae = refiner_vae + if self.refiner_vae: + conv_op = VideoConv3d + norm_op = RMS_norm + else: + conv_op = ops.Conv3d + norm_op = Normalize + + self.conv_in = conv_op(in_channels, block_out_channels[0], 3, 1, 1) self.down = nn.ModuleList() ch = block_out_channels[0] depth = (ffactor_spatial >> 1).bit_length() - depth_temporal = ((ffactor_spatial // ffactor_temporal) >> 1).bit_length() + depth_temporal = ((ffactor_spatial // self.ffactor_temporal) >> 1).bit_length() for i, tgt in enumerate(block_out_channels): stage = nn.Module() stage.block = nn.ModuleList([ResnetBlock(in_channels=ch if j == 0 else tgt, out_channels=tgt, temb_channels=0, - conv_op=VideoConv3d, norm_op=RMS_norm) + conv_op=conv_op, norm_op=norm_op) for j in range(num_res_blocks)]) ch = tgt if i < depth: nxt = block_out_channels[i + 1] if i + 1 < len(block_out_channels) and downsample_match_channel else ch - stage.downsample = DnSmpl(ch, nxt, tds=i >= depth_temporal) + stage.downsample = DnSmpl(ch, nxt, tds=i >= depth_temporal, refiner_vae=self.refiner_vae, op=conv_op) ch = nxt self.down.append(stage) self.mid = nn.Module() - self.mid.block_1 = ResnetBlock(in_channels=ch, out_channels=ch, temb_channels=0, conv_op=VideoConv3d, norm_op=RMS_norm) - self.mid.attn_1 = AttnBlock(ch, conv_op=ops.Conv3d, norm_op=RMS_norm) - self.mid.block_2 = ResnetBlock(in_channels=ch, out_channels=ch, temb_channels=0, conv_op=VideoConv3d, norm_op=RMS_norm) + self.mid.block_1 = ResnetBlock(in_channels=ch, out_channels=ch, temb_channels=0, conv_op=conv_op, norm_op=norm_op) + self.mid.attn_1 = AttnBlock(ch, conv_op=ops.Conv3d, norm_op=norm_op) + self.mid.block_2 = ResnetBlock(in_channels=ch, out_channels=ch, temb_channels=0, conv_op=conv_op, norm_op=norm_op) - self.norm_out = RMS_norm(ch) - self.conv_out = VideoConv3d(ch, z_channels << 1, 3, 1, 1) + self.norm_out = norm_op(ch) + self.conv_out = conv_op(ch, z_channels << 1, 3, 1, 1) self.regul = comfy.ldm.models.autoencoder.DiagonalGaussianRegularizer() def forward(self, x): + if not self.refiner_vae and x.shape[2] == 1: + x = x.expand(-1, -1, self.ffactor_temporal, -1, -1) + x = self.conv_in(x) for stage in self.down: @@ -200,31 +216,42 @@ class Encoder(nn.Module): skip = x.view(b, c // grp, grp, t, h, w).mean(2) out = self.conv_out(F.silu(self.norm_out(x))) + skip - out = self.regul(out)[0] - out = torch.cat((out[:, :, :1], out), dim=2) - out = out.permute(0, 2, 1, 3, 4) - b, f_times_2, c, h, w = out.shape - out = out.reshape(b, f_times_2 // 2, 2 * c, h, w) - out = out.permute(0, 2, 1, 3, 4).contiguous() + if self.refiner_vae: + out = self.regul(out)[0] + + out = torch.cat((out[:, :, :1], out), dim=2) + out = out.permute(0, 2, 1, 3, 4) + b, f_times_2, c, h, w = out.shape + out = out.reshape(b, f_times_2 // 2, 2 * c, h, w) + out = out.permute(0, 2, 1, 3, 4).contiguous() + return out class Decoder(nn.Module): def __init__(self, z_channels, out_channels, block_out_channels, num_res_blocks, - ffactor_spatial, ffactor_temporal, upsample_match_channel=True, **_): + ffactor_spatial, ffactor_temporal, upsample_match_channel=True, refiner_vae=True, **_): super().__init__() block_out_channels = block_out_channels[::-1] self.z_channels = z_channels self.block_out_channels = block_out_channels self.num_res_blocks = num_res_blocks + self.refiner_vae = refiner_vae + if self.refiner_vae: + conv_op = VideoConv3d + norm_op = RMS_norm + else: + conv_op = ops.Conv3d + norm_op = Normalize + ch = block_out_channels[0] - self.conv_in = VideoConv3d(z_channels, ch, 3) + self.conv_in = conv_op(z_channels, ch, kernel_size=3, stride=1, padding=1) self.mid = nn.Module() - self.mid.block_1 = ResnetBlock(in_channels=ch, out_channels=ch, temb_channels=0, conv_op=VideoConv3d, norm_op=RMS_norm) - self.mid.attn_1 = AttnBlock(ch, conv_op=ops.Conv3d, norm_op=RMS_norm) - self.mid.block_2 = ResnetBlock(in_channels=ch, out_channels=ch, temb_channels=0, conv_op=VideoConv3d, norm_op=RMS_norm) + self.mid.block_1 = ResnetBlock(in_channels=ch, out_channels=ch, temb_channels=0, conv_op=conv_op, norm_op=norm_op) + self.mid.attn_1 = AttnBlock(ch, conv_op=ops.Conv3d, norm_op=norm_op) + self.mid.block_2 = ResnetBlock(in_channels=ch, out_channels=ch, temb_channels=0, conv_op=conv_op, norm_op=norm_op) self.up = nn.ModuleList() depth = (ffactor_spatial >> 1).bit_length() @@ -235,25 +262,26 @@ class Decoder(nn.Module): stage.block = nn.ModuleList([ResnetBlock(in_channels=ch if j == 0 else tgt, out_channels=tgt, temb_channels=0, - conv_op=VideoConv3d, norm_op=RMS_norm) + conv_op=conv_op, norm_op=norm_op) for j in range(num_res_blocks + 1)]) ch = tgt if i < depth: nxt = block_out_channels[i + 1] if i + 1 < len(block_out_channels) and upsample_match_channel else ch - stage.upsample = UpSmpl(ch, nxt, tus=i < depth_temporal) + stage.upsample = UpSmpl(ch, nxt, tus=i < depth_temporal, refiner_vae=self.refiner_vae, op=conv_op) ch = nxt self.up.append(stage) - self.norm_out = RMS_norm(ch) - self.conv_out = VideoConv3d(ch, out_channels, 3) + self.norm_out = norm_op(ch) + self.conv_out = conv_op(ch, out_channels, 3, stride=1, padding=1) def forward(self, z): - z = z.permute(0, 2, 1, 3, 4) - b, f, c, h, w = z.shape - z = z.reshape(b, f, 2, c // 2, h, w) - z = z.permute(0, 1, 2, 3, 4, 5).reshape(b, f * 2, c // 2, h, w) - z = z.permute(0, 2, 1, 3, 4) - z = z[:, :, 1:] + if self.refiner_vae: + z = z.permute(0, 2, 1, 3, 4) + b, f, c, h, w = z.shape + z = z.reshape(b, f, 2, c // 2, h, w) + z = z.permute(0, 1, 2, 3, 4, 5).reshape(b, f * 2, c // 2, h, w) + z = z.permute(0, 2, 1, 3, 4) + z = z[:, :, 1:] x = self.conv_in(z) + z.repeat_interleave(self.block_out_channels[0] // self.z_channels, 1) x = self.mid.block_2(self.mid.attn_1(self.mid.block_1(x))) @@ -264,4 +292,10 @@ class Decoder(nn.Module): if hasattr(stage, 'upsample'): x = stage.upsample(x) - return self.conv_out(F.silu(self.norm_out(x))) + out = self.conv_out(F.silu(self.norm_out(x))) + + if not self.refiner_vae: + if z.shape[-3] == 1: + out = out[:, :, -1:] + + return out diff --git a/comfy/sd.py b/comfy/sd.py index 2df340739..873ad20f2 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -332,35 +332,51 @@ class VAE: self.first_stage_model = StageC_coder() self.downscale_ratio = 32 self.latent_channels = 16 - elif "decoder.conv_in.weight" in sd and sd['decoder.conv_in.weight'].shape[1] == 64: - ddconfig = {"block_out_channels": [128, 256, 512, 512, 1024, 1024], "in_channels": 3, "out_channels": 3, "num_res_blocks": 2, "ffactor_spatial": 32, "downsample_match_channel": True, "upsample_match_channel": True} - self.latent_channels = ddconfig['z_channels'] = sd["decoder.conv_in.weight"].shape[1] - self.downscale_ratio = 32 - self.upscale_ratio = 32 - self.working_dtypes = [torch.float16, torch.bfloat16, torch.float32] - self.first_stage_model = AutoencodingEngine(regularizer_config={'target': "comfy.ldm.models.autoencoder.DiagonalGaussianRegularizer"}, - encoder_config={'target': "comfy.ldm.hunyuan_video.vae.Encoder", 'params': ddconfig}, - decoder_config={'target': "comfy.ldm.hunyuan_video.vae.Decoder", 'params': ddconfig}) - - self.memory_used_encode = lambda shape, dtype: (700 * shape[2] * shape[3]) * model_management.dtype_size(dtype) - self.memory_used_decode = lambda shape, dtype: (700 * shape[2] * shape[3] * 32 * 32) * model_management.dtype_size(dtype) - elif "decoder.conv_in.weight" in sd: - #default SD1.x/SD2.x VAE parameters - ddconfig = {'double_z': True, 'z_channels': 4, 'resolution': 256, 'in_channels': 3, 'out_ch': 3, 'ch': 128, 'ch_mult': [1, 2, 4, 4], 'num_res_blocks': 2, 'attn_resolutions': [], 'dropout': 0.0} - - if 'encoder.down.2.downsample.conv.weight' not in sd and 'decoder.up.3.upsample.conv.weight' not in sd: #Stable diffusion x4 upscaler VAE - ddconfig['ch_mult'] = [1, 2, 4] - self.downscale_ratio = 4 - self.upscale_ratio = 4 - - self.latent_channels = ddconfig['z_channels'] = sd["decoder.conv_in.weight"].shape[1] - if 'post_quant_conv.weight' in sd: - self.first_stage_model = AutoencoderKL(ddconfig=ddconfig, embed_dim=sd['post_quant_conv.weight'].shape[1]) - else: + if sd['decoder.conv_in.weight'].shape[1] == 64: + ddconfig = {"block_out_channels": [128, 256, 512, 512, 1024, 1024], "in_channels": 3, "out_channels": 3, "num_res_blocks": 2, "ffactor_spatial": 32, "downsample_match_channel": True, "upsample_match_channel": True} + self.latent_channels = ddconfig['z_channels'] = sd["decoder.conv_in.weight"].shape[1] + self.downscale_ratio = 32 + self.upscale_ratio = 32 + self.working_dtypes = [torch.float16, torch.bfloat16, torch.float32] self.first_stage_model = AutoencodingEngine(regularizer_config={'target': "comfy.ldm.models.autoencoder.DiagonalGaussianRegularizer"}, - encoder_config={'target': "comfy.ldm.modules.diffusionmodules.model.Encoder", 'params': ddconfig}, - decoder_config={'target': "comfy.ldm.modules.diffusionmodules.model.Decoder", 'params': ddconfig}) + encoder_config={'target': "comfy.ldm.hunyuan_video.vae.Encoder", 'params': ddconfig}, + decoder_config={'target': "comfy.ldm.hunyuan_video.vae.Decoder", 'params': ddconfig}) + + self.memory_used_encode = lambda shape, dtype: (700 * shape[2] * shape[3]) * model_management.dtype_size(dtype) + self.memory_used_decode = lambda shape, dtype: (700 * shape[2] * shape[3] * 32 * 32) * model_management.dtype_size(dtype) + elif sd['decoder.conv_in.weight'].shape[1] == 32: + ddconfig = {"block_out_channels": [128, 256, 512, 1024, 1024], "in_channels": 3, "out_channels": 3, "num_res_blocks": 2, "ffactor_spatial": 16, "ffactor_temporal": 4, "downsample_match_channel": True, "upsample_match_channel": True, "refiner_vae": False} + self.latent_channels = ddconfig['z_channels'] = sd["decoder.conv_in.weight"].shape[1] + self.working_dtypes = [torch.float16, torch.bfloat16, torch.float32] + self.upscale_ratio = (lambda a: max(0, a * 4 - 3), 16, 16) + self.upscale_index_formula = (4, 16, 16) + self.downscale_ratio = (lambda a: max(0, math.floor((a + 3) / 4)), 16, 16) + self.downscale_index_formula = (4, 16, 16) + self.latent_dim = 3 + self.not_video = True + self.first_stage_model = AutoencodingEngine(regularizer_config={'target': "comfy.ldm.models.autoencoder.DiagonalGaussianRegularizer"}, + encoder_config={'target': "comfy.ldm.hunyuan_video.vae_refiner.Encoder", 'params': ddconfig}, + decoder_config={'target': "comfy.ldm.hunyuan_video.vae_refiner.Decoder", 'params': ddconfig}) + + self.memory_used_encode = lambda shape, dtype: (2800 * shape[-2] * shape[-1]) * model_management.dtype_size(dtype) + self.memory_used_decode = lambda shape, dtype: (2800 * shape[-3] * shape[-2] * shape[-1] * 16 * 16) * model_management.dtype_size(dtype) + else: + #default SD1.x/SD2.x VAE parameters + ddconfig = {'double_z': True, 'z_channels': 4, 'resolution': 256, 'in_channels': 3, 'out_ch': 3, 'ch': 128, 'ch_mult': [1, 2, 4, 4], 'num_res_blocks': 2, 'attn_resolutions': [], 'dropout': 0.0} + + if 'encoder.down.2.downsample.conv.weight' not in sd and 'decoder.up.3.upsample.conv.weight' not in sd: #Stable diffusion x4 upscaler VAE + ddconfig['ch_mult'] = [1, 2, 4] + self.downscale_ratio = 4 + self.upscale_ratio = 4 + + self.latent_channels = ddconfig['z_channels'] = sd["decoder.conv_in.weight"].shape[1] + if 'post_quant_conv.weight' in sd: + self.first_stage_model = AutoencoderKL(ddconfig=ddconfig, embed_dim=sd['post_quant_conv.weight'].shape[1]) + else: + self.first_stage_model = AutoencodingEngine(regularizer_config={'target': "comfy.ldm.models.autoencoder.DiagonalGaussianRegularizer"}, + encoder_config={'target': "comfy.ldm.modules.diffusionmodules.model.Encoder", 'params': ddconfig}, + decoder_config={'target': "comfy.ldm.modules.diffusionmodules.model.Decoder", 'params': ddconfig}) elif "decoder.layers.1.layers.0.beta" in sd: self.first_stage_model = AudioOobleckVAE() self.memory_used_encode = lambda shape, dtype: (1000 * shape[2]) * model_management.dtype_size(dtype) From bb32d4ec3141333df26fcdaee0c3c08e41b7b249 Mon Sep 17 00:00:00 2001 From: Koratahiu Date: Thu, 2 Oct 2025 00:59:07 +0300 Subject: [PATCH 2/5] feat: Add Epsilon Scaling node for exposure bias correction (#10132) --- comfy_extras/nodes_eps.py | 60 +++++++++++++++++++++++++++++++++++++++ nodes.py | 1 + 2 files changed, 61 insertions(+) create mode 100644 comfy_extras/nodes_eps.py diff --git a/comfy_extras/nodes_eps.py b/comfy_extras/nodes_eps.py new file mode 100644 index 000000000..c8818f096 --- /dev/null +++ b/comfy_extras/nodes_eps.py @@ -0,0 +1,60 @@ +class EpsilonScaling: + """ + Implements the Epsilon Scaling method from 'Elucidating the Exposure Bias in Diffusion Models' + (https://arxiv.org/abs/2308.15321v6). + + This method mitigates exposure bias by scaling the predicted noise during sampling, + which can significantly improve sample quality. This implementation uses the "uniform schedule" + recommended by the paper for its practicality and effectiveness. + """ + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "model": ("MODEL",), + "scaling_factor": ("FLOAT", { + "default": 1.005, + "min": 0.5, + "max": 1.5, + "step": 0.001, + "display": "number" + }), + } + } + + RETURN_TYPES = ("MODEL",) + FUNCTION = "patch" + + CATEGORY = "model_patches/unet" + + def patch(self, model, scaling_factor): + # Prevent division by zero, though the UI's min value should prevent this. + if scaling_factor == 0: + scaling_factor = 1e-9 + + def epsilon_scaling_function(args): + """ + This function is applied after the CFG guidance has been calculated. + It recalculates the denoised latent by scaling the predicted noise. + """ + denoised = args["denoised"] + x = args["input"] + + noise_pred = x - denoised + + scaled_noise_pred = noise_pred / scaling_factor + + new_denoised = x - scaled_noise_pred + + return new_denoised + + # Clone the model patcher to avoid modifying the original model in place + model_clone = model.clone() + + model_clone.set_model_sampler_post_cfg_function(epsilon_scaling_function) + + return (model_clone,) + +NODE_CLASS_MAPPINGS = { + "Epsilon Scaling": EpsilonScaling +} diff --git a/nodes.py b/nodes.py index 1a6784b68..88d712993 100644 --- a/nodes.py +++ b/nodes.py @@ -2297,6 +2297,7 @@ async def init_builtin_extra_nodes(): "nodes_gits.py", "nodes_controlnet.py", "nodes_hunyuan.py", + "nodes_eps.py", "nodes_flux.py", "nodes_lora_extract.py", "nodes_torch_compile.py", From 911331c06c16aa80633c5438c58edb32dbfdff50 Mon Sep 17 00:00:00 2001 From: rattus128 <46076784+rattus128@users.noreply.github.com> Date: Thu, 2 Oct 2025 08:40:28 +1000 Subject: [PATCH 3/5] sd: fix VAE tiled fallback VRAM leak (#10139) When the VAE catches this VRAM OOM, it launches the fallback logic straight from the exception context. Python however refs the entire call stack that caused the exception including any local variables for the sake of exception report and debugging. In the case of tensors, this can hold on the references to GBs of VRAM and inhibit the VRAM allocated from freeing them. So dump the except context completely before going back to the VAE via the tiler by getting out of the except block with nothing but a flag. The greately increases the reliability of the tiler fallback, especially on low VRAM cards, as with the bug, if the leak randomly leaked more than the headroom needed for a single tile, the tiler would fallback would OOM and fail the flow. --- comfy/sd.py | 16 ++++++++++++++++ 1 file changed, 16 insertions(+) diff --git a/comfy/sd.py b/comfy/sd.py index 873ad20f2..be225ad03 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -652,6 +652,7 @@ class VAE: def decode(self, samples_in, vae_options={}): self.throw_exception_if_invalid() pixel_samples = None + do_tile = False try: memory_used = self.memory_used_decode(samples_in.shape, self.vae_dtype) model_management.load_models_gpu([self.patcher], memory_required=memory_used, force_full_load=self.disable_offload) @@ -667,6 +668,13 @@ class VAE: pixel_samples[x:x+batch_number] = out except model_management.OOM_EXCEPTION: logging.warning("Warning: Ran out of memory when regular VAE decoding, retrying with tiled VAE decoding.") + #NOTE: We don't know what tensors were allocated to stack variables at the time of the + #exception and the exception itself refs them all until we get out of this except block. + #So we just set a flag for tiler fallback so that tensor gc can happen once the + #exception is fully off the books. + do_tile = True + + if do_tile: dims = samples_in.ndim - 2 if dims == 1 or self.extra_1d_channel is not None: pixel_samples = self.decode_tiled_1d(samples_in) @@ -713,6 +721,7 @@ class VAE: self.throw_exception_if_invalid() pixel_samples = self.vae_encode_crop_pixels(pixel_samples) pixel_samples = pixel_samples.movedim(-1, 1) + do_tile = False if self.latent_dim == 3 and pixel_samples.ndim < 5: if not self.not_video: pixel_samples = pixel_samples.movedim(1, 0).unsqueeze(0) @@ -734,6 +743,13 @@ class VAE: except model_management.OOM_EXCEPTION: logging.warning("Warning: Ran out of memory when regular VAE encoding, retrying with tiled VAE encoding.") + #NOTE: We don't know what tensors were allocated to stack variables at the time of the + #exception and the exception itself refs them all until we get out of this except block. + #So we just set a flag for tiler fallback so that tensor gc can happen once the + #exception is fully off the books. + do_tile = True + + if do_tile: if self.latent_dim == 3: tile = 256 overlap = tile // 4 From 4965c0e2acf39d84e82cb63dd6cc4400299d0a61 Mon Sep 17 00:00:00 2001 From: rattus128 <46076784+rattus128@users.noreply.github.com> Date: Thu, 2 Oct 2025 08:42:16 +1000 Subject: [PATCH 4/5] WAN: Fix cache VRAM leak on error (#10141) If this suffers an exception (such as a VRAM oom) it will leave the encode() and decode() methods which skips the cleanup of the WAN feature cache. The comfy node cache then ultimately keeps a reference this object which is in turn reffing large tensors from the failed execution. The feature cache is currently setup at a class variable on the encoder/decoder however, the encode and decode functions always clear it on both entry and exit of normal execution. Its likely the design intent is this is usable as a streaming encoder where the input comes in batches, however the functions as they are today don't support that. So simplify by bringing the cache back to local variable, so that if it does VRAM OOM the cache itself is properly garbage when the encode()/decode() functions dissappear from the stack. --- comfy/ldm/wan/vae.py | 37 ++++++++++++++----------------------- 1 file changed, 14 insertions(+), 23 deletions(-) diff --git a/comfy/ldm/wan/vae.py b/comfy/ldm/wan/vae.py index 791596938..ccbb25822 100644 --- a/comfy/ldm/wan/vae.py +++ b/comfy/ldm/wan/vae.py @@ -468,55 +468,46 @@ class WanVAE(nn.Module): attn_scales, self.temperal_upsample, dropout) def encode(self, x): - self.clear_cache() + conv_idx = [0] + feat_map = [None] * count_conv3d(self.decoder) ## cache t = x.shape[2] iter_ = 1 + (t - 1) // 4 ## 对encode输入的x，按时间拆分为1、4、4、4.... for i in range(iter_): - self._enc_conv_idx = [0] + conv_idx = [0] if i == 0: out = self.encoder( x[:, :, :1, :, :], - feat_cache=self._enc_feat_map, - feat_idx=self._enc_conv_idx) + feat_cache=feat_map, + feat_idx=conv_idx) else: out_ = self.encoder( x[:, :, 1 + 4 * (i - 1):1 + 4 * i, :, :], - feat_cache=self._enc_feat_map, - feat_idx=self._enc_conv_idx) + feat_cache=feat_map, + feat_idx=conv_idx) out = torch.cat([out, out_], 2) mu, log_var = self.conv1(out).chunk(2, dim=1) - self.clear_cache() return mu def decode(self, z): - self.clear_cache() + conv_idx = [0] + feat_map = [None] * count_conv3d(self.decoder) # z: [b,c,t,h,w] iter_ = z.shape[2] x = self.conv2(z) for i in range(iter_): - self._conv_idx = [0] + conv_idx = [0] if i == 0: out = self.decoder( x[:, :, i:i + 1, :, :], - feat_cache=self._feat_map, - feat_idx=self._conv_idx) + feat_cache=feat_map, + feat_idx=conv_idx) else: out_ = self.decoder( x[:, :, i:i + 1, :, :], - feat_cache=self._feat_map, - feat_idx=self._conv_idx) + feat_cache=feat_map, + feat_idx=conv_idx) out = torch.cat([out, out_], 2) - self.clear_cache() return out - - def clear_cache(self): - self._conv_num = count_conv3d(self.decoder) - self._conv_idx = [0] - self._feat_map = [None] * self._conv_num - #cache encode - self._enc_conv_num = count_conv3d(self.encoder) - self._enc_conv_idx = [0] - self._enc_feat_map = [None] * self._enc_conv_num From 0e9d1724be327c79ba86159d868f0b57adb8c384 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 1 Oct 2025 21:33:05 -0700 Subject: [PATCH 5/5] Add a .bat to the AMD portable to disable smart memory. (#10153) --- .ci/windows_amd_base_files/README_VERY_IMPORTANT.txt | 5 ++++- .../run_amd_gpu_disable_smart_memory.bat | 2 ++ 2 files changed, 6 insertions(+), 1 deletion(-) create mode 100755 .ci/windows_amd_base_files/run_amd_gpu_disable_smart_memory.bat diff --git a/.ci/windows_amd_base_files/README_VERY_IMPORTANT.txt b/.ci/windows_amd_base_files/README_VERY_IMPORTANT.txt index 570ac3398..96a500be2 100755 --- a/.ci/windows_amd_base_files/README_VERY_IMPORTANT.txt +++ b/.ci/windows_amd_base_files/README_VERY_IMPORTANT.txt @@ -3,10 +3,13 @@ https://www.amd.com/en/resources/support-articles/release-notes/RN-AMDGPU-WINDOW HOW TO RUN: -if you have a AMD gpu: +If you have a AMD gpu: run_amd_gpu.bat +If you have memory issues you can try disabling the smart memory management by running comfyui with: + +run_amd_gpu_disable_smart_memory.bat IF YOU GET A RED ERROR IN THE UI MAKE SURE YOU HAVE A MODEL/CHECKPOINT IN: ComfyUI\models\checkpoints diff --git a/.ci/windows_amd_base_files/run_amd_gpu_disable_smart_memory.bat b/.ci/windows_amd_base_files/run_amd_gpu_disable_smart_memory.bat new file mode 100755 index 000000000..cece0aeb2 --- /dev/null +++ b/.ci/windows_amd_base_files/run_amd_gpu_disable_smart_memory.bat @@ -0,0 +1,2 @@ +.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build --disable-smart-memory +pause