Merge 0141af0786 into 85fc35e8fa

This is using a different layers weight with .to(). Change it to use
the ops caster if the original layer is a comfy weight so that it picks
up dynamic_vram and async_offload functionality in full.

Co-authored-by: Rattus <rattus128@gmail.com>

comfy/latent_formats.py

@@ -755,6 +755,10 @@ class ACEAudio(LatentFormat):
     latent_channels = 8
     latent_dimensions = 2
 
+class ACEAudio15(LatentFormat):
+    latent_channels = 64
+    latent_dimensions = 1
+
 class ChromaRadiance(LatentFormat):
     latent_channels = 3
     spacial_downscale_ratio = 1

comfy/model_base.py

@@ -50,6 +50,7 @@ import comfy.ldm.omnigen.omnigen2
 import comfy.ldm.qwen_image.model
 import comfy.ldm.kandinsky5.model
 import comfy.ldm.anima.model
+import comfy.ldm.ace.ace_step15
 
 import comfy.model_management
 import comfy.patcher_extension
@@ -1540,6 +1541,47 @@ class ACEStep(BaseModel):
         out['lyrics_strength'] = comfy.conds.CONDConstant(kwargs.get("lyrics_strength", 1.0))
         return out
 
+class ACEStep15(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.ace.ace_step15.AceStepConditionGenerationModel)
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        device = kwargs["device"]
+
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+
+        conditioning_lyrics = kwargs.get("conditioning_lyrics", None)
+        if cross_attn is not None:
+            out['lyric_embed'] = comfy.conds.CONDRegular(conditioning_lyrics)
+
+        refer_audio = kwargs.get("reference_audio_timbre_latents", None)
+        if refer_audio is None or len(refer_audio) == 0:
+            refer_audio = torch.tensor([[[-1.3672e-01, -1.5820e-01,  5.8594e-01, -5.7422e-01,  3.0273e-02,
+                                        2.7930e-01, -2.5940e-03, -2.0703e-01, -1.6113e-01, -1.4746e-01,
+                                        -2.7710e-02, -1.8066e-01, -2.9688e-01,  1.6016e+00, -2.6719e+00,
+                                        7.7734e-01, -1.3516e+00, -1.9434e-01, -7.1289e-02, -5.0938e+00,
+                                        2.4316e-01,  4.7266e-01,  4.6387e-02, -6.6406e-01, -2.1973e-01,
+                                        -6.7578e-01, -1.5723e-01,  9.5312e-01, -2.0020e-01, -1.7109e+00,
+                                        5.8984e-01, -5.7422e-01,  5.1562e-01,  2.8320e-01,  1.4551e-01,
+                                        -1.8750e-01, -5.9814e-02,  3.6719e-01, -1.0059e-01, -1.5723e-01,
+                                        2.0605e-01, -4.3359e-01, -8.2812e-01,  4.5654e-02, -6.6016e-01,
+                                        1.4844e-01,  9.4727e-02,  3.8477e-01, -1.2578e+00, -3.3203e-01,
+                                        -8.5547e-01,  4.3359e-01,  4.2383e-01, -8.9453e-01, -5.0391e-01,
+                                        -5.6152e-02, -2.9219e+00, -2.4658e-02,  5.0391e-01,  9.8438e-01,
+                                        7.2754e-02, -2.1582e-01,  6.3672e-01,  1.0000e+00]]], device=device).movedim(-1, 1).repeat(1, 1, 750)
+        else:
+            refer_audio = refer_audio[-1]
+        out['refer_audio'] = comfy.conds.CONDRegular(refer_audio)
+
+        audio_codes = kwargs.get("audio_codes", None)
+        if audio_codes is not None:
+            out['audio_codes'] = comfy.conds.CONDRegular(torch.tensor(audio_codes, device=device))
+
+        return out
+
 class Omnigen2(BaseModel):
     def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
         super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.omnigen.omnigen2.OmniGen2Transformer2DModel)

comfy/model_detection.py

@@ -655,6 +655,11 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
         dit_config["num_visual_blocks"] = count_blocks(state_dict_keys, '{}visual_transformer_blocks.'.format(key_prefix) + '{}.')
         return dit_config
 
+    if '{}encoder.lyric_encoder.layers.0.input_layernorm.weight'.format(key_prefix) in state_dict_keys:
+        dit_config = {}
+        dit_config["audio_model"] = "ace1.5"
+        return dit_config
+
     if '{}input_blocks.0.0.weight'.format(key_prefix) not in state_dict_keys:
         return None
 

comfy/sd.py

@@ -59,6 +59,7 @@ import comfy.text_encoders.kandinsky5
 import comfy.text_encoders.jina_clip_2
 import comfy.text_encoders.newbie
 import comfy.text_encoders.anima
+import comfy.text_encoders.ace15
 
 import comfy.model_patcher
 import comfy.lora
@@ -452,6 +453,8 @@ class VAE:
         self.extra_1d_channel = None
         self.crop_input = True
 
+        self.audio_sample_rate = 44100
+
         if config is None:
             if "decoder.mid.block_1.mix_factor" in sd:
                 encoder_config = {'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}
@@ -549,14 +552,25 @@ class VAE:
                                                                     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()
+                config = {}
+                param_key = None
+                if "decoder.layers.2.layers.1.weight_v" in sd:
+                    param_key = "decoder.layers.2.layers.1.weight_v"
+                if "decoder.layers.2.layers.1.parametrizations.weight.original1" in sd:
+                    param_key = "decoder.layers.2.layers.1.parametrizations.weight.original1"
+                if param_key is not None:
+                    if sd[param_key].shape[-1] == 12:
+                        config["strides"] = [2, 4, 4, 6, 10]
+                        self.audio_sample_rate = 48000
+
+                self.first_stage_model = AudioOobleckVAE(**config)
                 self.memory_used_encode = lambda shape, dtype: (1000 * shape[2]) * model_management.dtype_size(dtype)
                 self.memory_used_decode = lambda shape, dtype: (1000 * shape[2] * 2048) * model_management.dtype_size(dtype)
                 self.latent_channels = 64
                 self.output_channels = 2
                 self.pad_channel_value = "replicate"
                 self.upscale_ratio = 2048
-                self.downscale_ratio =  2048
+                self.downscale_ratio = 2048
                 self.latent_dim = 1
                 self.process_output = lambda audio: audio
                 self.process_input = lambda audio: audio
@@ -1427,6 +1441,9 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
                 clip_data_jina = clip_data[0]
             tokenizer_data["gemma_spiece_model"] = clip_data_gemma.get("spiece_model", None)
             tokenizer_data["jina_spiece_model"] = clip_data_jina.get("spiece_model", None)
+        elif clip_type == CLIPType.ACE:
+            clip_target.clip = comfy.text_encoders.ace15.te(**llama_detect(clip_data))
+            clip_target.tokenizer = comfy.text_encoders.ace15.ACE15Tokenizer
         else:
             clip_target.clip = sdxl_clip.SDXLClipModel
             clip_target.tokenizer = sdxl_clip.SDXLTokenizer

comfy/sd1_clip.py

@@ -155,6 +155,8 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
         self.execution_device = options.get("execution_device", self.execution_device)
         if isinstance(self.layer, list) or self.layer == "all":
             pass
+        elif isinstance(layer_idx, list):
+            self.layer = layer_idx
         elif layer_idx is None or abs(layer_idx) > self.num_layers:
             self.layer = "last"
         else:

comfy/supported_models.py

@@ -24,6 +24,7 @@ import comfy.text_encoders.hunyuan_image
 import comfy.text_encoders.kandinsky5
 import comfy.text_encoders.z_image
 import comfy.text_encoders.anima
+import comfy.text_encoders.ace15
 
 from . import supported_models_base
 from . import latent_formats
@@ -1596,6 +1597,38 @@ 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, 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, Anima]
+class ACEStep15(supported_models_base.BASE):
+    unet_config = {
+        "audio_model": "ace1.5",
+    }
+
+    unet_extra_config = {
+    }
+
+    sampling_settings = {
+        "multiplier": 1.0,
+        "shift": 3.0,
+    }
+
+    latent_format = comfy.latent_formats.ACEAudio15
+
+    memory_usage_factor = 4.7
+
+    supported_inference_dtypes = [torch.bfloat16, torch.float32]
+
+    vae_key_prefix = ["vae."]
+    text_encoder_key_prefix = ["text_encoders."]
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.ACEStep15(self, device=device)
+        return out
+
+    def clip_target(self, state_dict={}):
+        pref = self.text_encoder_key_prefix[0]
+        hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen3_2b.transformer.".format(pref))
+        return supported_models_base.ClipTarget(comfy.text_encoders.ace15.ACE15Tokenizer, comfy.text_encoders.ace15.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, 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, ACEStep15, Omnigen2, QwenImage, Flux2, Kandinsky5Image, Kandinsky5, Anima]
 
 models += [SVD_img2vid]

comfy/text_encoders/ace15.py

@@ -0,0 +1,223 @@
+from .anima import Qwen3Tokenizer
+import comfy.text_encoders.llama
+from comfy import sd1_clip
+import torch
+import math
+import comfy.utils
+
+
+def sample_manual_loop_no_classes(
+    model,
+    ids=None,
+    paddings=[],
+    execution_dtype=None,
+    cfg_scale: float = 2.0,
+    temperature: float = 0.85,
+    top_p: float = 0.9,
+    top_k: int = None,
+    seed: int = 1,
+    min_tokens: int = 1,
+    max_new_tokens: int = 2048,
+    audio_start_id: int = 151669,  # The cutoff ID for audio codes
+    eos_token_id: int = 151645,
+):
+    device = model.execution_device
+
+    if execution_dtype is None:
+        if comfy.model_management.should_use_bf16(device):
+            execution_dtype = torch.bfloat16
+        else:
+            execution_dtype = torch.float32
+
+    embeds, attention_mask, num_tokens, embeds_info = model.process_tokens(ids, device)
+    for i, t in enumerate(paddings):
+        attention_mask[i, :t] = 0
+        attention_mask[i, t:] = 1
+
+    output_audio_codes = []
+    past_key_values = []
+    generator = torch.Generator(device=device)
+    generator.manual_seed(seed)
+    model_config = model.transformer.model.config
+
+    for x in range(model_config.num_hidden_layers):
+        past_key_values.append((torch.empty([embeds.shape[0], model_config.num_key_value_heads, embeds.shape[1] + min_tokens, model_config.head_dim], device=device, dtype=execution_dtype), torch.empty([embeds.shape[0], model_config.num_key_value_heads, embeds.shape[1] + min_tokens, model_config.head_dim], device=device, dtype=execution_dtype), 0))
+
+    progress_bar = comfy.utils.ProgressBar(max_new_tokens)
+
+    for step in range(max_new_tokens):
+        outputs = model.transformer(None, attention_mask, embeds=embeds.to(execution_dtype), num_tokens=num_tokens, intermediate_output=None, dtype=execution_dtype, embeds_info=embeds_info, past_key_values=past_key_values)
+        next_token_logits = model.transformer.logits(outputs[0])[:, -1]
+        past_key_values = outputs[2]
+
+        cond_logits = next_token_logits[0:1]
+        uncond_logits = next_token_logits[1:2]
+        cfg_logits = uncond_logits + cfg_scale * (cond_logits - uncond_logits)
+
+        if eos_token_id is not None and eos_token_id < audio_start_id and min_tokens < step:
+            eos_score = cfg_logits[:, eos_token_id].clone()
+
+        remove_logit_value = torch.finfo(cfg_logits.dtype).min
+        # Only generate audio tokens
+        cfg_logits[:, :audio_start_id] = remove_logit_value
+
+        if eos_token_id is not None and eos_token_id < audio_start_id and min_tokens < step:
+            cfg_logits[:, eos_token_id] = eos_score
+
+        if top_k is not None and top_k > 0:
+            top_k_vals, _ = torch.topk(cfg_logits, top_k)
+            min_val = top_k_vals[..., -1, None]
+            cfg_logits[cfg_logits < min_val] = remove_logit_value
+
+        if top_p is not None and top_p < 1.0:
+            sorted_logits, sorted_indices = torch.sort(cfg_logits, descending=True)
+            cumulative_probs = torch.cumsum(torch.softmax(sorted_logits, dim=-1), dim=-1)
+            sorted_indices_to_remove = cumulative_probs > top_p
+            sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
+            sorted_indices_to_remove[..., 0] = 0
+            indices_to_remove = sorted_indices_to_remove.scatter(1, sorted_indices, sorted_indices_to_remove)
+            cfg_logits[indices_to_remove] = remove_logit_value
+
+        if temperature > 0:
+            cfg_logits = cfg_logits / temperature
+            next_token = torch.multinomial(torch.softmax(cfg_logits, dim=-1), num_samples=1, generator=generator).squeeze(1)
+        else:
+            next_token = torch.argmax(cfg_logits, dim=-1)
+
+        token = next_token.item()
+
+        if token == eos_token_id:
+            break
+
+        embed, _, _, _ = model.process_tokens([[token]], device)
+        embeds = embed.repeat(2, 1, 1)
+        attention_mask = torch.cat([attention_mask, torch.ones((2, 1), device=device, dtype=attention_mask.dtype)], dim=1)
+
+        output_audio_codes.append(token - audio_start_id)
+        progress_bar.update_absolute(step)
+
+    return output_audio_codes
+
+
+def generate_audio_codes(model, positive, negative, min_tokens=1, max_tokens=1024, seed=0):
+    cfg_scale = 2.0
+
+    positive = [[token for token, _ in inner_list] for inner_list in positive]
+    negative = [[token for token, _ in inner_list] for inner_list in negative]
+    positive = positive[0]
+    negative = negative[0]
+
+    neg_pad = 0
+    if len(negative) < len(positive):
+        neg_pad = (len(positive) - len(negative))
+        negative = [model.special_tokens["pad"]] * neg_pad + negative
+
+    pos_pad = 0
+    if len(negative) > len(positive):
+        pos_pad = (len(negative) - len(positive))
+        positive = [model.special_tokens["pad"]] * pos_pad + positive
+
+    paddings = [pos_pad, neg_pad]
+    return sample_manual_loop_no_classes(model, [positive, negative], paddings, cfg_scale=cfg_scale, seed=seed, min_tokens=min_tokens, max_new_tokens=max_tokens)
+
+
+class ACE15Tokenizer(sd1_clip.SD1Tokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, name="qwen3_06b", tokenizer=Qwen3Tokenizer)
+
+    def tokenize_with_weights(self, text, return_word_ids=False, **kwargs):
+        out = {}
+        lyrics = kwargs.get("lyrics", "")
+        bpm = kwargs.get("bpm", 120)
+        duration = kwargs.get("duration", 120)
+        keyscale = kwargs.get("keyscale", "C major")
+        timesignature = kwargs.get("timesignature", 2)
+        language = kwargs.get("language", "en")
+        seed = kwargs.get("seed", 0)
+
+        duration = math.ceil(duration)
+        meta_lm = 'bpm: {}\nduration: {}\nkeyscale: {}\ntimesignature: {}'.format(bpm, duration, keyscale, timesignature)
+        lm_template = "<|im_start|>system\n# Instruction\nGenerate audio semantic tokens based on the given conditions:\n\n<|im_end|>\n<|im_start|>user\n# Caption\n{}\n{}\n<|im_end|>\n<|im_start|>assistant\n<think>\n{}\n</think>\n\n<|im_end|>\n"
+
+        meta_cap = '- bpm: {}\n- timesignature: {}\n- keyscale: {}\n- duration: {}\n'.format(bpm, timesignature, keyscale, duration)
+        out["lm_prompt"] = self.qwen3_06b.tokenize_with_weights(lm_template.format(text, lyrics, meta_lm), disable_weights=True)
+        out["lm_prompt_negative"] = self.qwen3_06b.tokenize_with_weights(lm_template.format(text, lyrics, ""), disable_weights=True)
+
+        out["lyrics"] = self.qwen3_06b.tokenize_with_weights("# Languages\n{}\n\n# Lyric{}<|endoftext|><|endoftext|>".format(language, lyrics), return_word_ids, disable_weights=True, **kwargs)
+        out["qwen3_06b"] = self.qwen3_06b.tokenize_with_weights("# Instruction\nGenerate audio semantic tokens based on the given conditions:\n\n# Caption\n{}# Metas\n{}<|endoftext|>\n<|endoftext|>".format(text, meta_cap), return_word_ids, **kwargs)
+        out["lm_metadata"] = {"min_tokens": duration * 5, "seed": seed}
+        return out
+
+
+class Qwen3_06BModel(sd1_clip.SDClipModel):
+    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, attention_mask=True, model_options={}):
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"pad": 151643}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Qwen3_06B_ACE15, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
+
+class Qwen3_2B_ACE15(sd1_clip.SDClipModel):
+    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, attention_mask=True, model_options={}):
+        llama_quantization_metadata = model_options.get("llama_quantization_metadata", None)
+        if llama_quantization_metadata is not None:
+            model_options = model_options.copy()
+            model_options["quantization_metadata"] = llama_quantization_metadata
+
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"pad": 151643}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Qwen3_2B_ACE15_lm, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
+
+class ACE15TEModel(torch.nn.Module):
+    def __init__(self, device="cpu", dtype=None, dtype_llama=None, model_options={}):
+        super().__init__()
+        if dtype_llama is None:
+            dtype_llama = dtype
+
+        self.qwen3_06b = Qwen3_06BModel(device=device, dtype=dtype, model_options=model_options)
+        self.qwen3_2b = Qwen3_2B_ACE15(device=device, dtype=dtype_llama, model_options=model_options)
+        self.dtypes = set([dtype, dtype_llama])
+
+    def encode_token_weights(self, token_weight_pairs):
+        token_weight_pairs_base = token_weight_pairs["qwen3_06b"]
+        token_weight_pairs_lyrics = token_weight_pairs["lyrics"]
+
+        self.qwen3_06b.set_clip_options({"layer": None})
+        base_out, _, extra = self.qwen3_06b.encode_token_weights(token_weight_pairs_base)
+        self.qwen3_06b.set_clip_options({"layer": [0]})
+        lyrics_embeds, _, extra_l = self.qwen3_06b.encode_token_weights(token_weight_pairs_lyrics)
+
+        lm_metadata = token_weight_pairs["lm_metadata"]
+        audio_codes = generate_audio_codes(self.qwen3_2b, token_weight_pairs["lm_prompt"], token_weight_pairs["lm_prompt_negative"], min_tokens=lm_metadata["min_tokens"], max_tokens=lm_metadata["min_tokens"], seed=lm_metadata["seed"])
+
+        return base_out, None, {"conditioning_lyrics": lyrics_embeds[:, 0], "audio_codes": [audio_codes]}
+
+    def set_clip_options(self, options):
+        self.qwen3_06b.set_clip_options(options)
+        self.qwen3_2b.set_clip_options(options)
+
+    def reset_clip_options(self):
+        self.qwen3_06b.reset_clip_options()
+        self.qwen3_2b.reset_clip_options()
+
+    def load_sd(self, sd):
+        if "model.layers.0.post_attention_layernorm.weight" in sd:
+            shape = sd["model.layers.0.post_attention_layernorm.weight"].shape
+            if shape[0] == 1024:
+                return self.qwen3_06b.load_sd(sd)
+            else:
+                return self.qwen3_2b.load_sd(sd)
+
+    def memory_estimation_function(self, token_weight_pairs, device=None):
+        lm_metadata = token_weight_pairs["lm_metadata"]
+        constant = 0.4375
+        if comfy.model_management.should_use_bf16(device):
+            constant *= 0.5
+
+        token_weight_pairs = token_weight_pairs.get("lm_prompt", [])
+        num_tokens = sum(map(lambda a: len(a), token_weight_pairs))
+        num_tokens += lm_metadata['min_tokens']
+        return num_tokens * constant * 1024 * 1024
+
+def te(dtype_llama=None, llama_quantization_metadata=None):
+    class ACE15TEModel_(ACE15TEModel):
+        def __init__(self, device="cpu", dtype=None, model_options={}):
+            if llama_quantization_metadata is not None:
+                model_options = model_options.copy()
+                model_options["llama_quantization_metadata"] = llama_quantization_metadata
+            super().__init__(device=device, dtype_llama=dtype_llama, dtype=dtype, model_options=model_options)
+    return ACE15TEModel_

comfy/text_encoders/llama.py

@@ -6,6 +6,7 @@ import math
 
 from comfy.ldm.modules.attention import optimized_attention_for_device
 import comfy.model_management
+import comfy.ops
 import comfy.ldm.common_dit
 import comfy.clip_model
 
@@ -103,6 +104,52 @@ class Qwen3_06BConfig:
     final_norm: bool = True
     lm_head: bool = False
 
+@dataclass
+class Qwen3_06B_ACE15_Config:
+    vocab_size: int = 151669
+    hidden_size: int = 1024
+    intermediate_size: int = 3072
+    num_hidden_layers: int = 28
+    num_attention_heads: int = 16
+    num_key_value_heads: int = 8
+    max_position_embeddings: int = 32768
+    rms_norm_eps: float = 1e-6
+    rope_theta: float = 1000000.0
+    transformer_type: str = "llama"
+    head_dim = 128
+    rms_norm_add = False
+    mlp_activation = "silu"
+    qkv_bias = False
+    rope_dims = None
+    q_norm = "gemma3"
+    k_norm = "gemma3"
+    rope_scale = None
+    final_norm: bool = True
+    lm_head: bool = False
+
+@dataclass
+class Qwen3_2B_ACE15_lm_Config:
+    vocab_size: int = 217204
+    hidden_size: int = 2048
+    intermediate_size: int = 6144
+    num_hidden_layers: int = 28
+    num_attention_heads: int = 16
+    num_key_value_heads: int = 8
+    max_position_embeddings: int = 40960
+    rms_norm_eps: float = 1e-6
+    rope_theta: float = 1000000.0
+    transformer_type: str = "llama"
+    head_dim = 128
+    rms_norm_add = False
+    mlp_activation = "silu"
+    qkv_bias = False
+    rope_dims = None
+    q_norm = "gemma3"
+    k_norm = "gemma3"
+    rope_scale = None
+    final_norm: bool = True
+    lm_head: bool = False
+
 @dataclass
 class Qwen3_4BConfig:
     vocab_size: int = 151936
@@ -581,10 +628,10 @@ class Llama2_(nn.Module):
         mask = None
         if attention_mask is not None:
             mask = 1.0 - attention_mask.to(x.dtype).reshape((attention_mask.shape[0], 1, -1, attention_mask.shape[-1])).expand(attention_mask.shape[0], 1, seq_len, attention_mask.shape[-1])
-            mask = mask.masked_fill(mask.to(torch.bool), float("-inf"))
+            mask = mask.masked_fill(mask.to(torch.bool), torch.finfo(x.dtype).min)
 
         if seq_len > 1:
-            causal_mask = torch.empty(past_len + seq_len, past_len + seq_len, dtype=x.dtype, device=x.device).fill_(float("-inf")).triu_(1)
+            causal_mask = torch.empty(past_len + seq_len, past_len + seq_len, dtype=x.dtype, device=x.device).fill_(torch.finfo(x.dtype).min).triu_(1)
             if mask is not None:
                 mask += causal_mask
             else:
@@ -729,6 +776,39 @@ class Qwen3_06B(BaseLlama, torch.nn.Module):
         self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
         self.dtype = dtype
 
+class Qwen3_06B_ACE15(BaseLlama, torch.nn.Module):
+    def __init__(self, config_dict, dtype, device, operations):
+        super().__init__()
+        config = Qwen3_06B_ACE15_Config(**config_dict)
+        self.num_layers = config.num_hidden_layers
+
+        self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
+        self.dtype = dtype
+
+class Qwen3_2B_ACE15_lm(BaseLlama, torch.nn.Module):
+    def __init__(self, config_dict, dtype, device, operations):
+        super().__init__()
+        config = Qwen3_2B_ACE15_lm_Config(**config_dict)
+        self.num_layers = config.num_hidden_layers
+
+        self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
+        self.dtype = dtype
+
+    def logits(self, x):
+        input = x[:, -1:]
+        module = self.model.embed_tokens
+
+        offload_stream = None
+        if module.comfy_cast_weights:
+            weight, _, offload_stream = comfy.ops.cast_bias_weight(module, input, offloadable=True)
+        else:
+            weight = self.model.embed_tokens.weight.to(x)
+
+        x = torch.nn.functional.linear(input, weight, None)
+
+        comfy.ops.uncast_bias_weight(module, weight, None, offload_stream)
+        return x
+
 class Qwen3_4B(BaseLlama, torch.nn.Module):
     def __init__(self, config_dict, dtype, device, operations):
         super().__init__()

comfy_api/latest/__init__.py

@@ -106,6 +106,42 @@ class Types:
     MESH = MESH
     VOXEL = VOXEL
 
+
+class Caching:
+    """
+    External cache provider API for distributed caching.
+
+    Enables sharing cached results across multiple ComfyUI instances
+    (e.g., Kubernetes pods) without monkey-patching internal methods.
+
+    Example usage:
+        from comfy_api.latest import Caching
+
+        class MyRedisProvider(Caching.CacheProvider):
+            def on_lookup(self, context):
+                # Check Redis for cached result
+                ...
+
+            def on_store(self, context, value):
+                # Store to Redis (can be async internally)
+                ...
+
+        Caching.register_provider(MyRedisProvider())
+    """
+    # Import from comfy_execution.cache_provider (source of truth)
+    from comfy_execution.cache_provider import (
+        CacheProvider,
+        CacheContext,
+        CacheValue,
+        register_cache_provider as register_provider,
+        unregister_cache_provider as unregister_provider,
+        get_cache_providers as get_providers,
+        has_cache_providers as has_providers,
+        clear_cache_providers as clear_providers,
+        estimate_value_size,
+    )
+
+
 ComfyAPI = ComfyAPI_latest
 
 # Create a synchronous version of the API
@@ -125,6 +161,7 @@ __all__ = [
     "Input",
     "InputImpl",
     "Types",
+    "Caching",
     "ComfyExtension",
     "io",
     "IO",

comfy_execution/cache_provider.py

@@ -0,0 +1,319 @@
+"""
+External Cache Provider API for distributed caching.
+
+This module provides a public API for external cache providers, enabling
+distributed caching across multiple ComfyUI instances (e.g., Kubernetes pods).
+
+Public API is also available via:
+    from comfy_api.latest import Caching
+
+Example usage:
+    from comfy_execution.cache_provider import (
+        CacheProvider, CacheContext, CacheValue, register_cache_provider
+    )
+
+    class MyRedisProvider(CacheProvider):
+        def on_lookup(self, context: CacheContext) -> Optional[CacheValue]:
+            # Check Redis/GCS for cached result
+            ...
+
+        def on_store(self, context: CacheContext, value: CacheValue) -> None:
+            # Store to Redis/GCS (can be async internally)
+            ...
+
+    register_cache_provider(MyRedisProvider())
+"""
+
+from abc import ABC, abstractmethod
+from typing import Any, Optional, Tuple, List
+from dataclasses import dataclass
+import hashlib
+import json
+import logging
+import math
+import pickle
+import threading
+
+logger = logging.getLogger(__name__)
+
+
+# ============================================================
+# Data Classes
+# ============================================================
+
+@dataclass
+class CacheContext:
+    """Context passed to provider methods."""
+    prompt_id: str          # Current prompt execution ID
+    node_id: str            # Node being cached
+    class_type: str         # Node class type (e.g., "KSampler")
+    cache_key: Any          # Raw cache key (frozenset structure)
+    cache_key_bytes: bytes  # SHA256 hash for external storage key
+
+
+@dataclass
+class CacheValue:
+    """
+    Value stored/retrieved from external cache.
+
+    The ui field is optional - implementations may choose to skip it
+    (e.g., if it contains non-portable data like local file paths).
+    """
+    outputs: list  # The tensor/value outputs
+    ui: dict = None  # Optional UI data (may be skipped by implementations)
+
+
+# ============================================================
+# Provider Interface
+# ============================================================
+
+class CacheProvider(ABC):
+    """
+    Abstract base class for external cache providers.
+
+    Thread Safety:
+        Providers may be called from multiple threads. Implementations
+        must be thread-safe.
+
+    Error Handling:
+        All methods are wrapped in try/except by the caller. Exceptions
+        are logged but never propagate to break execution.
+
+    Performance Guidelines:
+        - on_lookup: Should complete in <500ms (including network)
+        - on_store: Can be async internally (fire-and-forget)
+        - should_cache: Should be fast (<1ms), called frequently
+    """
+
+    @abstractmethod
+    def on_lookup(self, context: CacheContext) -> Optional[CacheValue]:
+        """
+        Check external storage for cached result.
+
+        Called AFTER local cache miss (local-first for performance).
+
+        Returns:
+            CacheValue if found externally, None otherwise.
+
+        Important:
+            - Return None on any error (don't raise)
+            - Validate data integrity before returning
+        """
+        pass
+
+    @abstractmethod
+    def on_store(self, context: CacheContext, value: CacheValue) -> None:
+        """
+        Store value to external cache.
+
+        Called AFTER value is stored in local cache.
+
+        Important:
+            - Can be fire-and-forget (async internally)
+            - Should never block execution
+            - Handle serialization failures gracefully
+        """
+        pass
+
+    def should_cache(self, context: CacheContext, value: Optional[CacheValue] = None) -> bool:
+        """
+        Filter which nodes should be externally cached.
+
+        Called before on_lookup (value=None) and on_store (value provided).
+        Return False to skip external caching for this node.
+
+        Implementations can filter based on context.class_type, value size,
+        or any custom logic. Use estimate_value_size() to get value size.
+
+        Default: Returns True (cache everything).
+        """
+        return True
+
+    def on_prompt_start(self, prompt_id: str) -> None:
+        """Called when prompt execution begins. Optional."""
+        pass
+
+    def on_prompt_end(self, prompt_id: str) -> None:
+        """Called when prompt execution ends. Optional."""
+        pass
+
+
+# ============================================================
+# Provider Registry
+# ============================================================
+
+_providers: List[CacheProvider] = []
+_providers_lock = threading.Lock()
+_providers_snapshot: Optional[Tuple[CacheProvider, ...]] = None
+
+
+def register_cache_provider(provider: CacheProvider) -> None:
+    """
+    Register an external cache provider.
+
+    Providers are called in registration order. First provider to return
+    a result from on_lookup wins.
+    """
+    global _providers_snapshot
+    with _providers_lock:
+        if provider in _providers:
+            logger.warning(f"Provider {provider.__class__.__name__} already registered")
+            return
+        _providers.append(provider)
+        _providers_snapshot = None  # Invalidate cache
+        logger.info(f"Registered cache provider: {provider.__class__.__name__}")
+
+
+def unregister_cache_provider(provider: CacheProvider) -> None:
+    """Remove a previously registered provider."""
+    global _providers_snapshot
+    with _providers_lock:
+        try:
+            _providers.remove(provider)
+            _providers_snapshot = None
+            logger.info(f"Unregistered cache provider: {provider.__class__.__name__}")
+        except ValueError:
+            logger.warning(f"Provider {provider.__class__.__name__} was not registered")
+
+
+def get_cache_providers() -> Tuple[CacheProvider, ...]:
+    """Get registered providers (cached for performance)."""
+    global _providers_snapshot
+    snapshot = _providers_snapshot
+    if snapshot is not None:
+        return snapshot
+    with _providers_lock:
+        if _providers_snapshot is not None:
+            return _providers_snapshot
+        _providers_snapshot = tuple(_providers)
+        return _providers_snapshot
+
+
+def has_cache_providers() -> bool:
+    """Fast check if any providers registered (no lock)."""
+    return bool(_providers)
+
+
+def clear_cache_providers() -> None:
+    """Remove all providers. Useful for testing."""
+    global _providers_snapshot
+    with _providers_lock:
+        _providers.clear()
+        _providers_snapshot = None
+
+
+# ============================================================
+# Utilities
+# ============================================================
+
+def _canonicalize(obj: Any) -> Any:
+    """
+    Convert an object to a canonical, JSON-serializable form.
+
+    This ensures deterministic ordering regardless of Python's hash randomization,
+    which is critical for cross-pod cache key consistency. Frozensets in particular
+    have non-deterministic iteration order between Python sessions.
+    """
+    if isinstance(obj, frozenset):
+        # Sort frozenset items for deterministic ordering
+        return ("__frozenset__", sorted(
+            [_canonicalize(item) for item in obj],
+            key=lambda x: json.dumps(x, sort_keys=True)
+        ))
+    elif isinstance(obj, set):
+        return ("__set__", sorted(
+            [_canonicalize(item) for item in obj],
+            key=lambda x: json.dumps(x, sort_keys=True)
+        ))
+    elif isinstance(obj, tuple):
+        return ("__tuple__", [_canonicalize(item) for item in obj])
+    elif isinstance(obj, list):
+        return [_canonicalize(item) for item in obj]
+    elif isinstance(obj, dict):
+        return {str(k): _canonicalize(v) for k, v in sorted(obj.items())}
+    elif isinstance(obj, (int, float, str, bool, type(None))):
+        return obj
+    elif isinstance(obj, bytes):
+        return ("__bytes__", obj.hex())
+    elif hasattr(obj, 'value'):
+        # Handle Unhashable class from ComfyUI
+        return ("__unhashable__", _canonicalize(getattr(obj, 'value', None)))
+    else:
+        # For other types, use repr as fallback
+        return ("__repr__", repr(obj))
+
+
+def serialize_cache_key(cache_key: Any) -> bytes:
+    """
+    Serialize cache key to bytes for external storage.
+
+    Returns SHA256 hash suitable for Redis/database keys.
+
+    Note: Uses canonicalize + JSON serialization instead of pickle because
+    pickle is NOT deterministic across Python sessions due to hash randomization
+    affecting frozenset iteration order. This is critical for distributed caching
+    where different pods need to compute the same hash for identical inputs.
+    """
+    try:
+        canonical = _canonicalize(cache_key)
+        json_str = json.dumps(canonical, sort_keys=True, separators=(',', ':'))
+        return hashlib.sha256(json_str.encode('utf-8')).digest()
+    except Exception as e:
+        logger.warning(f"Failed to serialize cache key: {e}")
+        # Fallback to pickle (non-deterministic but better than nothing)
+        try:
+            serialized = pickle.dumps(cache_key, protocol=4)
+            return hashlib.sha256(serialized).digest()
+        except Exception:
+            return hashlib.sha256(str(id(cache_key)).encode()).digest()
+
+
+def contains_nan(obj: Any) -> bool:
+    """
+    Check if cache key contains NaN (indicates uncacheable node).
+
+    NaN != NaN in Python, so local cache never hits. But serialized
+    NaN would match, causing incorrect external hits. Must skip these.
+    """
+    if isinstance(obj, float):
+        try:
+            return math.isnan(obj)
+        except (TypeError, ValueError):
+            return False
+    if hasattr(obj, 'value'):  # Unhashable class
+        val = getattr(obj, 'value', None)
+        if isinstance(val, float):
+            try:
+                return math.isnan(val)
+            except (TypeError, ValueError):
+                return False
+    if isinstance(obj, (frozenset, tuple, list, set)):
+        return any(contains_nan(item) for item in obj)
+    if isinstance(obj, dict):
+        return any(contains_nan(k) or contains_nan(v) for k, v in obj.items())
+    return False
+
+
+def estimate_value_size(value: CacheValue) -> int:
+    """Estimate serialized size in bytes. Useful for size-based filtering."""
+    try:
+        import torch
+    except ImportError:
+        return 0
+
+    total = 0
+
+    def estimate(obj):
+        nonlocal total
+        if isinstance(obj, torch.Tensor):
+            total += obj.numel() * obj.element_size()
+        elif isinstance(obj, dict):
+            for v in obj.values():
+                estimate(v)
+        elif isinstance(obj, (list, tuple)):
+            for item in obj:
+                estimate(item)
+
+    for output in value.outputs:
+        estimate(output)
+    return total

comfy_execution/caching.py

@@ -155,6 +155,10 @@ class BasicCache:
         self.cache = {}
         self.subcaches = {}
 
+        # External cache provider support
+        self._is_subcache = False
+        self._current_prompt_id = ''
+
     async def set_prompt(self, dynprompt, node_ids, is_changed_cache):
         self.dynprompt = dynprompt
         self.cache_key_set = self.key_class(dynprompt, node_ids, is_changed_cache)
@@ -201,20 +205,123 @@ class BasicCache:
         cache_key = self.cache_key_set.get_data_key(node_id)
         self.cache[cache_key] = value
 
+        # Notify external providers
+        self._notify_providers_store(node_id, cache_key, value)
+
     def _get_immediate(self, node_id):
         if not self.initialized:
             return None
         cache_key = self.cache_key_set.get_data_key(node_id)
+
+        # Check local cache first (fast path)
         if cache_key in self.cache:
             return self.cache[cache_key]
-        else:
+
+        # Check external providers on local miss
+        external_result = self._check_providers_lookup(node_id, cache_key)
+        if external_result is not None:
+            self.cache[cache_key] = external_result  # Warm local cache
+            return external_result
+
+        return None
+
+    def _notify_providers_store(self, node_id, cache_key, value):
+        """Notify external providers of cache store."""
+        from comfy_execution.cache_provider import (
+            has_cache_providers, get_cache_providers,
+            CacheContext, CacheValue,
+            serialize_cache_key, contains_nan, logger
+        )
+
+        # Fast exit conditions
+        if self._is_subcache:
+            return
+        if not has_cache_providers():
+            return
+        if not self._is_external_cacheable_value(value):
+            return
+        if contains_nan(cache_key):
+            return
+
+        context = CacheContext(
+            prompt_id=self._current_prompt_id,
+            node_id=node_id,
+            class_type=self._get_class_type(node_id),
+            cache_key=cache_key,
+            cache_key_bytes=serialize_cache_key(cache_key)
+        )
+        cache_value = CacheValue(outputs=value.outputs, ui=value.ui)
+
+        for provider in get_cache_providers():
+            try:
+                if provider.should_cache(context, cache_value):
+                    provider.on_store(context, cache_value)
+            except Exception as e:
+                logger.warning(f"Cache provider {provider.__class__.__name__} error on store: {e}")
+
+    def _check_providers_lookup(self, node_id, cache_key):
+        """Check external providers for cached result."""
+        from comfy_execution.cache_provider import (
+            has_cache_providers, get_cache_providers,
+            CacheContext, CacheValue,
+            serialize_cache_key, contains_nan, logger
+        )
+
+        if self._is_subcache:
             return None
+        if not has_cache_providers():
+            return None
+        if contains_nan(cache_key):
+            return None
+
+        context = CacheContext(
+            prompt_id=self._current_prompt_id,
+            node_id=node_id,
+            class_type=self._get_class_type(node_id),
+            cache_key=cache_key,
+            cache_key_bytes=serialize_cache_key(cache_key)
+        )
+
+        for provider in get_cache_providers():
+            try:
+                if not provider.should_cache(context):
+                    continue
+                result = provider.on_lookup(context)
+                if result is not None:
+                    if not isinstance(result, CacheValue):
+                        logger.warning(f"Provider {provider.__class__.__name__} returned invalid type")
+                        continue
+                    if not isinstance(result.outputs, (list, tuple)):
+                        logger.warning(f"Provider {provider.__class__.__name__} returned invalid outputs")
+                        continue
+                    # Import CacheEntry here to avoid circular import at module level
+                    from execution import CacheEntry
+                    return CacheEntry(ui=result.ui or {}, outputs=list(result.outputs))
+            except Exception as e:
+                logger.warning(f"Cache provider {provider.__class__.__name__} error on lookup: {e}")
+
+        return None
+
+    def _is_external_cacheable_value(self, value):
+        """Check if value is a CacheEntry suitable for external caching (not objects cache)."""
+        return hasattr(value, 'outputs') and hasattr(value, 'ui')
+
+    def _get_class_type(self, node_id):
+        """Get class_type for a node."""
+        if not self.initialized or not self.dynprompt:
+            return ''
+        try:
+            return self.dynprompt.get_node(node_id).get('class_type', '')
+        except Exception:
+            return ''
 
     async def _ensure_subcache(self, node_id, children_ids):
         subcache_key = self.cache_key_set.get_subcache_key(node_id)
         subcache = self.subcaches.get(subcache_key, None)
         if subcache is None:
             subcache = BasicCache(self.key_class)
+            subcache._is_subcache = True  # Mark as subcache - excludes from external caching
+            subcache._current_prompt_id = self._current_prompt_id  # Propagate prompt ID
             self.subcaches[subcache_key] = subcache
         await subcache.set_prompt(self.dynprompt, children_ids, self.is_changed_cache)
         return subcache

comfy_extras/nodes_ace.py

@@ -28,12 +28,39 @@ class TextEncodeAceStepAudio(io.ComfyNode):
         conditioning = node_helpers.conditioning_set_values(conditioning, {"lyrics_strength": lyrics_strength})
         return io.NodeOutput(conditioning)
 
+class TextEncodeAceStepAudio15(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="TextEncodeAceStepAudio1.5",
+            category="conditioning",
+            inputs=[
+                io.Clip.Input("clip"),
+                io.String.Input("tags", multiline=True, dynamic_prompts=True),
+                io.String.Input("lyrics", multiline=True, dynamic_prompts=True),
+                io.Int.Input("seed", default=0, min=0, max=0xffffffffffffffff, control_after_generate=True),
+                io.Int.Input("bpm", default=120, min=10, max=300),
+                io.Float.Input("duration", default=120.0, min=0.0, max=2000.0, step=0.1),
+                io.Combo.Input("timesignature", options=['2', '3', '4', '6']),
+                io.Combo.Input("language", options=["en", "ja", "zh", "es", "de", "fr", "pt", "ru", "it", "nl", "pl", "tr", "vi", "cs", "fa", "id", "ko", "uk", "hu", "ar", "sv", "ro", "el"]),
+                io.Combo.Input("keyscale", options=[f"{root} {quality}" for quality in ["major", "minor"] for root in ["C", "C#", "Db", "D", "D#", "Eb", "E", "F", "F#", "Gb", "G", "G#", "Ab", "A", "A#", "Bb", "B"]]),
+            ],
+            outputs=[io.Conditioning.Output()],
+        )
+
+    @classmethod
+    def execute(cls, clip, tags, lyrics, seed, bpm, duration, timesignature, language, keyscale) -> io.NodeOutput:
+        tokens = clip.tokenize(tags, lyrics=lyrics, bpm=bpm, duration=duration, timesignature=int(timesignature), language=language, keyscale=keyscale, seed=seed)
+        conditioning = clip.encode_from_tokens_scheduled(tokens)
+        return io.NodeOutput(conditioning)
+
 
 class EmptyAceStepLatentAudio(io.ComfyNode):
     @classmethod
     def define_schema(cls):
         return io.Schema(
             node_id="EmptyAceStepLatentAudio",
+            display_name="Empty Ace Step 1.0 Latent Audio",
             category="latent/audio",
             inputs=[
                 io.Float.Input("seconds", default=120.0, min=1.0, max=1000.0, step=0.1),
@@ -51,12 +78,60 @@ class EmptyAceStepLatentAudio(io.ComfyNode):
         return io.NodeOutput({"samples": latent, "type": "audio"})
 
 
+class EmptyAceStep15LatentAudio(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="EmptyAceStep1.5LatentAudio",
+            display_name="Empty Ace Step 1.5 Latent Audio",
+            category="latent/audio",
+            inputs=[
+                io.Float.Input("seconds", default=120.0, min=1.0, max=1000.0, step=0.01),
+                io.Int.Input(
+                    "batch_size", default=1, min=1, max=4096, tooltip="The number of latent images in the batch."
+                ),
+            ],
+            outputs=[io.Latent.Output()],
+        )
+
+    @classmethod
+    def execute(cls, seconds, batch_size) -> io.NodeOutput:
+        length = round((seconds * 48000 / 1920))
+        latent = torch.zeros([batch_size, 64, length], device=comfy.model_management.intermediate_device())
+        return io.NodeOutput({"samples": latent, "type": "audio"})
+
+class ReferenceTimbreAudio(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="ReferenceTimbreAudio",
+            category="advanced/conditioning/audio",
+            is_experimental=True,
+            description="This node sets the reference audio for timbre (for ace step 1.5)",
+            inputs=[
+                io.Conditioning.Input("conditioning"),
+                io.Latent.Input("latent", optional=True),
+            ],
+            outputs=[
+                io.Conditioning.Output(),
+            ]
+        )
+
+    @classmethod
+    def execute(cls, conditioning, latent=None) -> io.NodeOutput:
+        if latent is not None:
+            conditioning = node_helpers.conditioning_set_values(conditioning, {"reference_audio_timbre_latents": [latent["samples"]]}, append=True)
+        return io.NodeOutput(conditioning)
+
 class AceExtension(ComfyExtension):
     @override
     async def get_node_list(self) -> list[type[io.ComfyNode]]:
         return [
             TextEncodeAceStepAudio,
             EmptyAceStepLatentAudio,
+            TextEncodeAceStepAudio15,
+            EmptyAceStep15LatentAudio,
+            ReferenceTimbreAudio,
         ]
 
 async def comfy_entrypoint() -> AceExtension:

comfy_extras/nodes_audio.py

@@ -82,13 +82,14 @@ class VAEEncodeAudio(IO.ComfyNode):
     @classmethod
     def execute(cls, vae, audio) -> IO.NodeOutput:
         sample_rate = audio["sample_rate"]
-        if 44100 != sample_rate:
-            waveform = torchaudio.functional.resample(audio["waveform"], sample_rate, 44100)
+        vae_sample_rate = getattr(vae, "audio_sample_rate", 44100)
+        if vae_sample_rate != sample_rate:
+            waveform = torchaudio.functional.resample(audio["waveform"], sample_rate, vae_sample_rate)
         else:
             waveform = audio["waveform"]
 
         t = vae.encode(waveform.movedim(1, -1))
-        return IO.NodeOutput({"samples":t})
+        return IO.NodeOutput({"samples": t})
 
     encode = execute  # TODO: remove
 
@@ -114,7 +115,8 @@ 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 if "sample_rate" not in samples else samples["sample_rate"]})
+        vae_sample_rate = getattr(vae, "audio_sample_rate", 44100)
+        return IO.NodeOutput({"waveform": audio, "sample_rate": vae_sample_rate if "sample_rate" not in samples else samples["sample_rate"]})
 
     decode = execute  # TODO: remove
 

comfyui_version.py

@@ -1,3 +1,3 @@
 # This file is automatically generated by the build process when version is
 # updated in pyproject.toml.
-__version__ = "0.11.1"
+__version__ = "0.12.0"

execution.py

@@ -683,6 +683,22 @@ class PromptExecutor:
             }
             self.add_message("execution_error", mes, broadcast=False)
 
+    def _notify_prompt_lifecycle(self, event: str, prompt_id: str):
+        """Notify external cache providers of prompt lifecycle events."""
+        from comfy_execution.cache_provider import has_cache_providers, get_cache_providers, logger
+
+        if not has_cache_providers():
+            return
+
+        for provider in get_cache_providers():
+            try:
+                if event == "start":
+                    provider.on_prompt_start(prompt_id)
+                elif event == "end":
+                    provider.on_prompt_end(prompt_id)
+            except Exception as e:
+                logger.warning(f"Cache provider {provider.__class__.__name__} error on {event}: {e}")
+
     def execute(self, prompt, prompt_id, extra_data={}, execute_outputs=[]):
         asyncio.run(self.execute_async(prompt, prompt_id, extra_data, execute_outputs))
 
@@ -699,66 +715,77 @@ class PromptExecutor:
         self.status_messages = []
         self.add_message("execution_start", { "prompt_id": prompt_id}, broadcast=False)
 
-        with torch.inference_mode():
-            dynamic_prompt = DynamicPrompt(prompt)
-            reset_progress_state(prompt_id, dynamic_prompt)
-            add_progress_handler(WebUIProgressHandler(self.server))
-            is_changed_cache = IsChangedCache(prompt_id, dynamic_prompt, self.caches.outputs)
-            for cache in self.caches.all:
-                await cache.set_prompt(dynamic_prompt, prompt.keys(), is_changed_cache)
-                cache.clean_unused()
+        # Set prompt ID on caches for external provider integration
+        for cache in self.caches.all:
+            cache._current_prompt_id = prompt_id
 
-            cached_nodes = []
-            for node_id in prompt:
-                if self.caches.outputs.get(node_id) is not None:
-                    cached_nodes.append(node_id)
+        # Notify external cache providers of prompt start
+        self._notify_prompt_lifecycle("start", prompt_id)
 
-            comfy.model_management.cleanup_models_gc()
-            self.add_message("execution_cached",
-                          { "nodes": cached_nodes, "prompt_id": prompt_id},
-                          broadcast=False)
-            pending_subgraph_results = {}
-            pending_async_nodes = {} # TODO - Unify this with pending_subgraph_results
-            ui_node_outputs = {}
-            executed = set()
-            execution_list = ExecutionList(dynamic_prompt, self.caches.outputs)
-            current_outputs = self.caches.outputs.all_node_ids()
-            for node_id in list(execute_outputs):
-                execution_list.add_node(node_id)
+        try:
+            with torch.inference_mode():
+                dynamic_prompt = DynamicPrompt(prompt)
+                reset_progress_state(prompt_id, dynamic_prompt)
+                add_progress_handler(WebUIProgressHandler(self.server))
+                is_changed_cache = IsChangedCache(prompt_id, dynamic_prompt, self.caches.outputs)
+                for cache in self.caches.all:
+                    await cache.set_prompt(dynamic_prompt, prompt.keys(), is_changed_cache)
+                    cache.clean_unused()
 
-            while not execution_list.is_empty():
-                node_id, error, ex = await execution_list.stage_node_execution()
-                if error is not None:
-                    self.handle_execution_error(prompt_id, dynamic_prompt.original_prompt, current_outputs, executed, error, ex)
-                    break
+                cached_nodes = []
+                for node_id in prompt:
+                    if self.caches.outputs.get(node_id) is not None:
+                        cached_nodes.append(node_id)
 
-                assert node_id is not None, "Node ID should not be None at this point"
-                result, error, ex = await execute(self.server, dynamic_prompt, self.caches, node_id, extra_data, executed, prompt_id, execution_list, pending_subgraph_results, pending_async_nodes, ui_node_outputs)
-                self.success = result != ExecutionResult.FAILURE
-                if result == ExecutionResult.FAILURE:
-                    self.handle_execution_error(prompt_id, dynamic_prompt.original_prompt, current_outputs, executed, error, ex)
-                    break
-                elif result == ExecutionResult.PENDING:
-                    execution_list.unstage_node_execution()
-                else: # result == ExecutionResult.SUCCESS:
-                    execution_list.complete_node_execution()
-                self.caches.outputs.poll(ram_headroom=self.cache_args["ram"])
-            else:
-                # Only execute when the while-loop ends without break
-                self.add_message("execution_success", { "prompt_id": prompt_id }, broadcast=False)
+                comfy.model_management.cleanup_models_gc()
+                self.add_message("execution_cached",
+                              { "nodes": cached_nodes, "prompt_id": prompt_id},
+                              broadcast=False)
+                pending_subgraph_results = {}
+                pending_async_nodes = {} # TODO - Unify this with pending_subgraph_results
+                ui_node_outputs = {}
+                executed = set()
+                execution_list = ExecutionList(dynamic_prompt, self.caches.outputs)
+                current_outputs = self.caches.outputs.all_node_ids()
+                for node_id in list(execute_outputs):
+                    execution_list.add_node(node_id)
 
-            ui_outputs = {}
-            meta_outputs = {}
-            for node_id, ui_info in ui_node_outputs.items():
-                ui_outputs[node_id] = ui_info["output"]
-                meta_outputs[node_id] = ui_info["meta"]
-            self.history_result = {
-                "outputs": ui_outputs,
-                "meta": meta_outputs,
-            }
-            self.server.last_node_id = None
-            if comfy.model_management.DISABLE_SMART_MEMORY:
-                comfy.model_management.unload_all_models()
+                while not execution_list.is_empty():
+                    node_id, error, ex = await execution_list.stage_node_execution()
+                    if error is not None:
+                        self.handle_execution_error(prompt_id, dynamic_prompt.original_prompt, current_outputs, executed, error, ex)
+                        break
+
+                    assert node_id is not None, "Node ID should not be None at this point"
+                    result, error, ex = await execute(self.server, dynamic_prompt, self.caches, node_id, extra_data, executed, prompt_id, execution_list, pending_subgraph_results, pending_async_nodes, ui_node_outputs)
+                    self.success = result != ExecutionResult.FAILURE
+                    if result == ExecutionResult.FAILURE:
+                        self.handle_execution_error(prompt_id, dynamic_prompt.original_prompt, current_outputs, executed, error, ex)
+                        break
+                    elif result == ExecutionResult.PENDING:
+                        execution_list.unstage_node_execution()
+                    else: # result == ExecutionResult.SUCCESS:
+                        execution_list.complete_node_execution()
+                    self.caches.outputs.poll(ram_headroom=self.cache_args["ram"])
+                else:
+                    # Only execute when the while-loop ends without break
+                    self.add_message("execution_success", { "prompt_id": prompt_id }, broadcast=False)
+
+                ui_outputs = {}
+                meta_outputs = {}
+                for node_id, ui_info in ui_node_outputs.items():
+                    ui_outputs[node_id] = ui_info["output"]
+                    meta_outputs[node_id] = ui_info["meta"]
+                self.history_result = {
+                    "outputs": ui_outputs,
+                    "meta": meta_outputs,
+                }
+                self.server.last_node_id = None
+                if comfy.model_management.DISABLE_SMART_MEMORY:
+                    comfy.model_management.unload_all_models()
+        finally:
+            # Notify external cache providers of prompt end
+            self._notify_prompt_lifecycle("end", prompt_id)
 
 
 async def validate_inputs(prompt_id, prompt, item, validated):

nodes.py

@@ -1001,7 +1001,7 @@ 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", "ltxv", "newbie"], ),
+                              "type": (["sdxl", "sd3", "flux", "hunyuan_video", "hidream", "hunyuan_image", "hunyuan_video_15", "kandinsky5", "kandinsky5_image", "ltxv", "newbie", "ace"], ),
                               },
                 "optional": {
                               "device": (["default", "cpu"], {"advanced": True}),

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "ComfyUI"
-version = "0.11.1"
+version = "0.12.0"
 readme = "README.md"
 license = { file = "LICENSE" }
 requires-python = ">=3.10"

requirements.txt

@@ -1,5 +1,5 @@
 comfyui-frontend-package==1.37.11
-comfyui-workflow-templates==0.8.27
+comfyui-workflow-templates==0.8.31
 comfyui-embedded-docs==0.4.0
 torch
 torchsde

tests-unit/execution_test/test_cache_provider.py

@@ -0,0 +1,370 @@
+"""Tests for external cache provider API."""
+
+import importlib.util
+import pytest
+from typing import Optional
+
+
+def _torch_available() -> bool:
+    """Check if PyTorch is available."""
+    return importlib.util.find_spec("torch") is not None
+
+
+from comfy_execution.cache_provider import (
+    CacheProvider,
+    CacheContext,
+    CacheValue,
+    register_cache_provider,
+    unregister_cache_provider,
+    get_cache_providers,
+    has_cache_providers,
+    clear_cache_providers,
+    serialize_cache_key,
+    contains_nan,
+    estimate_value_size,
+    _canonicalize,
+)
+
+
+class TestCanonicalize:
+    """Test _canonicalize function for deterministic ordering."""
+
+    def test_frozenset_ordering_is_deterministic(self):
+        """Frozensets should produce consistent canonical form regardless of iteration order."""
+        # Create two frozensets with same content
+        fs1 = frozenset([("a", 1), ("b", 2), ("c", 3)])
+        fs2 = frozenset([("c", 3), ("a", 1), ("b", 2)])
+
+        result1 = _canonicalize(fs1)
+        result2 = _canonicalize(fs2)
+
+        assert result1 == result2
+
+    def test_nested_frozenset_ordering(self):
+        """Nested frozensets should also be deterministically ordered."""
+        inner1 = frozenset([1, 2, 3])
+        inner2 = frozenset([3, 2, 1])
+
+        fs1 = frozenset([("key", inner1)])
+        fs2 = frozenset([("key", inner2)])
+
+        result1 = _canonicalize(fs1)
+        result2 = _canonicalize(fs2)
+
+        assert result1 == result2
+
+    def test_dict_ordering(self):
+        """Dicts should be sorted by key."""
+        d1 = {"z": 1, "a": 2, "m": 3}
+        d2 = {"a": 2, "m": 3, "z": 1}
+
+        result1 = _canonicalize(d1)
+        result2 = _canonicalize(d2)
+
+        assert result1 == result2
+
+    def test_tuple_preserved(self):
+        """Tuples should be marked and preserved."""
+        t = (1, 2, 3)
+        result = _canonicalize(t)
+
+        assert result[0] == "__tuple__"
+        assert result[1] == [1, 2, 3]
+
+    def test_list_preserved(self):
+        """Lists should be recursively canonicalized."""
+        lst = [{"b": 2, "a": 1}, frozenset([3, 2, 1])]
+        result = _canonicalize(lst)
+
+        # First element should be dict with sorted keys
+        assert result[0] == {"a": 1, "b": 2}
+        # Second element should be canonicalized frozenset
+        assert result[1][0] == "__frozenset__"
+
+    def test_primitives_unchanged(self):
+        """Primitive types should pass through unchanged."""
+        assert _canonicalize(42) == 42
+        assert _canonicalize(3.14) == 3.14
+        assert _canonicalize("hello") == "hello"
+        assert _canonicalize(True) is True
+        assert _canonicalize(None) is None
+
+    def test_bytes_converted(self):
+        """Bytes should be converted to hex string."""
+        b = b"\x00\xff"
+        result = _canonicalize(b)
+
+        assert result[0] == "__bytes__"
+        assert result[1] == "00ff"
+
+    def test_set_ordering(self):
+        """Sets should be sorted like frozensets."""
+        s1 = {3, 1, 2}
+        s2 = {1, 2, 3}
+
+        result1 = _canonicalize(s1)
+        result2 = _canonicalize(s2)
+
+        assert result1 == result2
+        assert result1[0] == "__set__"
+
+
+class TestSerializeCacheKey:
+    """Test serialize_cache_key for deterministic hashing."""
+
+    def test_same_content_same_hash(self):
+        """Same content should produce same hash."""
+        key1 = frozenset([("node_1", frozenset([("input", "value")]))])
+        key2 = frozenset([("node_1", frozenset([("input", "value")]))])
+
+        hash1 = serialize_cache_key(key1)
+        hash2 = serialize_cache_key(key2)
+
+        assert hash1 == hash2
+
+    def test_different_content_different_hash(self):
+        """Different content should produce different hash."""
+        key1 = frozenset([("node_1", "value_a")])
+        key2 = frozenset([("node_1", "value_b")])
+
+        hash1 = serialize_cache_key(key1)
+        hash2 = serialize_cache_key(key2)
+
+        assert hash1 != hash2
+
+    def test_returns_bytes(self):
+        """Should return bytes (SHA256 digest)."""
+        key = frozenset([("test", 123)])
+        result = serialize_cache_key(key)
+
+        assert isinstance(result, bytes)
+        assert len(result) == 32  # SHA256 produces 32 bytes
+
+    def test_complex_nested_structure(self):
+        """Complex nested structures should hash deterministically."""
+        # Note: frozensets can only contain hashable types, so we use
+        # nested frozensets of tuples to represent dict-like structures
+        key = frozenset([
+            ("node_1", frozenset([
+                ("input_a", ("tuple", "value")),
+                ("input_b", frozenset([("nested", "dict")])),
+            ])),
+            ("node_2", frozenset([
+                ("param", 42),
+            ])),
+        ])
+
+        # Hash twice to verify determinism
+        hash1 = serialize_cache_key(key)
+        hash2 = serialize_cache_key(key)
+
+        assert hash1 == hash2
+
+    def test_dict_in_cache_key(self):
+        """Dicts passed directly to serialize_cache_key should work."""
+        # This tests the _canonicalize function's ability to handle dicts
+        key = {"node_1": {"input": "value"}, "node_2": 42}
+
+        hash1 = serialize_cache_key(key)
+        hash2 = serialize_cache_key(key)
+
+        assert hash1 == hash2
+        assert isinstance(hash1, bytes)
+        assert len(hash1) == 32
+
+
+class TestContainsNan:
+    """Test contains_nan utility function."""
+
+    def test_nan_float_detected(self):
+        """NaN floats should be detected."""
+        assert contains_nan(float('nan')) is True
+
+    def test_regular_float_not_nan(self):
+        """Regular floats should not be detected as NaN."""
+        assert contains_nan(3.14) is False
+        assert contains_nan(0.0) is False
+        assert contains_nan(-1.5) is False
+
+    def test_infinity_not_nan(self):
+        """Infinity is not NaN."""
+        assert contains_nan(float('inf')) is False
+        assert contains_nan(float('-inf')) is False
+
+    def test_nan_in_list(self):
+        """NaN in list should be detected."""
+        assert contains_nan([1, 2, float('nan'), 4]) is True
+        assert contains_nan([1, 2, 3, 4]) is False
+
+    def test_nan_in_tuple(self):
+        """NaN in tuple should be detected."""
+        assert contains_nan((1, float('nan'))) is True
+        assert contains_nan((1, 2, 3)) is False
+
+    def test_nan_in_frozenset(self):
+        """NaN in frozenset should be detected."""
+        assert contains_nan(frozenset([1, float('nan')])) is True
+        assert contains_nan(frozenset([1, 2, 3])) is False
+
+    def test_nan_in_dict_value(self):
+        """NaN in dict value should be detected."""
+        assert contains_nan({"key": float('nan')}) is True
+        assert contains_nan({"key": 42}) is False
+
+    def test_nan_in_nested_structure(self):
+        """NaN in deeply nested structure should be detected."""
+        nested = {"level1": [{"level2": (1, 2, float('nan'))}]}
+        assert contains_nan(nested) is True
+
+    def test_non_numeric_types(self):
+        """Non-numeric types should not be NaN."""
+        assert contains_nan("string") is False
+        assert contains_nan(None) is False
+        assert contains_nan(True) is False
+
+
+class TestEstimateValueSize:
+    """Test estimate_value_size utility function."""
+
+    def test_empty_outputs(self):
+        """Empty outputs should have zero size."""
+        value = CacheValue(outputs=[])
+        assert estimate_value_size(value) == 0
+
+    @pytest.mark.skipif(
+        not _torch_available(),
+        reason="PyTorch not available"
+    )
+    def test_tensor_size_estimation(self):
+        """Tensor size should be estimated correctly."""
+        import torch
+
+        # 1000 float32 elements = 4000 bytes
+        tensor = torch.zeros(1000, dtype=torch.float32)
+        value = CacheValue(outputs=[[tensor]])
+
+        size = estimate_value_size(value)
+        assert size == 4000
+
+    @pytest.mark.skipif(
+        not _torch_available(),
+        reason="PyTorch not available"
+    )
+    def test_nested_tensor_in_dict(self):
+        """Tensors nested in dicts should be counted."""
+        import torch
+
+        tensor = torch.zeros(100, dtype=torch.float32)  # 400 bytes
+        value = CacheValue(outputs=[[{"samples": tensor}]])
+
+        size = estimate_value_size(value)
+        assert size == 400
+
+
+class TestProviderRegistry:
+    """Test cache provider registration and retrieval."""
+
+    def setup_method(self):
+        """Clear providers before each test."""
+        clear_cache_providers()
+
+    def teardown_method(self):
+        """Clear providers after each test."""
+        clear_cache_providers()
+
+    def test_register_provider(self):
+        """Provider should be registered successfully."""
+        provider = MockCacheProvider()
+        register_cache_provider(provider)
+
+        assert has_cache_providers() is True
+        providers = get_cache_providers()
+        assert len(providers) == 1
+        assert providers[0] is provider
+
+    def test_unregister_provider(self):
+        """Provider should be unregistered successfully."""
+        provider = MockCacheProvider()
+        register_cache_provider(provider)
+        unregister_cache_provider(provider)
+
+        assert has_cache_providers() is False
+
+    def test_multiple_providers(self):
+        """Multiple providers can be registered."""
+        provider1 = MockCacheProvider()
+        provider2 = MockCacheProvider()
+
+        register_cache_provider(provider1)
+        register_cache_provider(provider2)
+
+        providers = get_cache_providers()
+        assert len(providers) == 2
+
+    def test_duplicate_registration_ignored(self):
+        """Registering same provider twice should be ignored."""
+        provider = MockCacheProvider()
+
+        register_cache_provider(provider)
+        register_cache_provider(provider)  # Should be ignored
+
+        providers = get_cache_providers()
+        assert len(providers) == 1
+
+    def test_clear_providers(self):
+        """clear_cache_providers should remove all providers."""
+        provider1 = MockCacheProvider()
+        provider2 = MockCacheProvider()
+
+        register_cache_provider(provider1)
+        register_cache_provider(provider2)
+        clear_cache_providers()
+
+        assert has_cache_providers() is False
+        assert len(get_cache_providers()) == 0
+
+
+class TestCacheContext:
+    """Test CacheContext dataclass."""
+
+    def test_context_creation(self):
+        """CacheContext should be created with all fields."""
+        context = CacheContext(
+            prompt_id="prompt-123",
+            node_id="node-456",
+            class_type="KSampler",
+            cache_key=frozenset([("test", "value")]),
+            cache_key_bytes=b"hash_bytes",
+        )
+
+        assert context.prompt_id == "prompt-123"
+        assert context.node_id == "node-456"
+        assert context.class_type == "KSampler"
+        assert context.cache_key == frozenset([("test", "value")])
+        assert context.cache_key_bytes == b"hash_bytes"
+
+
+class TestCacheValue:
+    """Test CacheValue dataclass."""
+
+    def test_value_creation(self):
+        """CacheValue should be created with outputs."""
+        outputs = [[{"samples": "tensor_data"}]]
+        value = CacheValue(outputs=outputs)
+
+        assert value.outputs == outputs
+
+
+class MockCacheProvider(CacheProvider):
+    """Mock cache provider for testing."""
+
+    def __init__(self):
+        self.lookups = []
+        self.stores = []
+
+    def on_lookup(self, context: CacheContext) -> Optional[CacheValue]:
+        self.lookups.append(context)
+        return None
+
+    def on_store(self, context: CacheContext, value: CacheValue) -> None:
+        self.stores.append((context, value))