jaidaken/ComfyUI
1
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

Make old scaled fp8 format use the new mixed quant ops system. (#11000)

Browse Source
This commit is contained in:
comfyanonymous
2025-12-05 11:35:42 -08:00
committed by GitHub
parent 0ec05b1481
commit 43071e3de3
24 changed files with 278 additions and 275 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
comfy/text_encoders/cosmos.py
View File

@@ -7,10 +7,10 @@ from transformers import T5TokenizerFast
class T5XXLModel(sd1_clip.SDClipModel):
def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, attention_mask=True, model_options={}):
textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "t5_old_config_xxl.json")
t5xxl_scaled_fp8 = model_options.get("t5xxl_scaled_fp8", None)
if t5xxl_scaled_fp8 is not None:
t5xxl_quantization_metadata = model_options.get("t5xxl_quantization_metadata", None)
if t5xxl_quantization_metadata is not None:
model_options = model_options.copy()
model_options["scaled_fp8"] = t5xxl_scaled_fp8
model_options["quantization_metadata"] = t5xxl_quantization_metadata
super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"end": 1, "pad": 0}, model_class=comfy.text_encoders.t5.T5, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, zero_out_masked=attention_mask, model_options=model_options)
@@ -30,12 +30,12 @@ class CosmosT5Tokenizer(sd1_clip.SD1Tokenizer):
super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, clip_name="t5xxl", tokenizer=T5XXLTokenizer)
def te(dtype_t5=None, t5xxl_scaled_fp8=None):
def te(dtype_t5=None, t5_quantization_metadata=None):
class CosmosTEModel_(CosmosT5XXL):
def __init__(self, device="cpu", dtype=None, model_options={}):
if t5xxl_scaled_fp8 is not None and "t5xxl_scaled_fp8" not in model_options:
if t5_quantization_metadata is not None:
model_options = model_options.copy()
model_options["t5xxl_scaled_fp8"] = t5xxl_scaled_fp8
model_options["t5xxl_quantization_metadata"] = t5_quantization_metadata
if dtype is None:
dtype = dtype_t5
super().__init__(device=device, dtype=dtype, model_options=model_options)

12
comfy/text_encoders/flux.py
View File

@@ -63,12 +63,12 @@ class FluxClipModel(torch.nn.Module):
else:
return self.t5xxl.load_sd(sd)
def flux_clip(dtype_t5=None, t5xxl_scaled_fp8=None):
def flux_clip(dtype_t5=None, t5_quantization_metadata=None):
class FluxClipModel_(FluxClipModel):
def __init__(self, device="cpu", dtype=None, model_options={}):
if t5xxl_scaled_fp8 is not None and "t5xxl_scaled_fp8" not in model_options:
if t5_quantization_metadata is not None:
model_options = model_options.copy()
model_options["t5xxl_scaled_fp8"] = t5xxl_scaled_fp8
model_options["t5xxl_quantization_metadata"] = t5_quantization_metadata
super().__init__(dtype_t5=dtype_t5, device=device, dtype=dtype, model_options=model_options)
return FluxClipModel_
@@ -159,15 +159,13 @@ class Flux2TEModel(sd1_clip.SD1ClipModel):
out = out.reshape(out.shape[0], out.shape[1], -1)
return out, pooled, extra
def flux2_te(dtype_llama=None, llama_scaled_fp8=None, llama_quantization_metadata=None, pruned=False):
def flux2_te(dtype_llama=None, llama_quantization_metadata=None, pruned=False):
class Flux2TEModel_(Flux2TEModel):
def __init__(self, device="cpu", dtype=None, model_options={}):
if llama_scaled_fp8 is not None and "scaled_fp8" not in model_options:
model_options = model_options.copy()
model_options["scaled_fp8"] = llama_scaled_fp8
if dtype_llama is not None:
dtype = dtype_llama
if llama_quantization_metadata is not None:
model_options = model_options.copy()
model_options["quantization_metadata"] = llama_quantization_metadata
if pruned:
model_options = model_options.copy()

6
comfy/text_encoders/genmo.py
View File

@@ -26,12 +26,12 @@ class MochiT5Tokenizer(sd1_clip.SD1Tokenizer):
super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, clip_name="t5xxl", tokenizer=T5XXLTokenizer)
def mochi_te(dtype_t5=None, t5xxl_scaled_fp8=None):
def mochi_te(dtype_t5=None, t5_quantization_metadata=None):
class MochiTEModel_(MochiT5XXL):
def __init__(self, device="cpu", dtype=None, model_options={}):
if t5xxl_scaled_fp8 is not None and "t5xxl_scaled_fp8" not in model_options:
if t5_quantization_metadata is not None:
model_options = model_options.copy()
model_options["t5xxl_scaled_fp8"] = t5xxl_scaled_fp8
model_options["t5xxl_quantization_metadata"] = t5_quantization_metadata
if dtype is None:
dtype = dtype_t5
super().__init__(device=device, dtype=dtype, model_options=model_options)

10
comfy/text_encoders/hidream.py
View File

@@ -142,14 +142,14 @@ class HiDreamTEModel(torch.nn.Module):
return self.llama.load_sd(sd)
def hidream_clip(clip_l=True, clip_g=True, t5=True, llama=True, dtype_t5=None, dtype_llama=None, t5xxl_scaled_fp8=None, llama_scaled_fp8=None):
def hidream_clip(clip_l=True, clip_g=True, t5=True, llama=True, dtype_t5=None, dtype_llama=None, t5_quantization_metadata=None, llama_quantization_metadata=None):
class HiDreamTEModel_(HiDreamTEModel):
def __init__(self, device="cpu", dtype=None, model_options={}):
if t5xxl_scaled_fp8 is not None and "t5xxl_scaled_fp8" not in model_options:
if t5_quantization_metadata is not None:
model_options = model_options.copy()
model_options["t5xxl_scaled_fp8"] = t5xxl_scaled_fp8
if llama_scaled_fp8 is not None and "llama_scaled_fp8" not in model_options:
model_options["t5xxl_quantization_metadata"] = t5_quantization_metadata
if llama_quantization_metadata is not None:
model_options = model_options.copy()
model_options["llama_scaled_fp8"] = llama_scaled_fp8
model_options["llama_quantization_metadata"] = llama_quantization_metadata
super().__init__(clip_l=clip_l, clip_g=clip_g, t5=t5, llama=llama, dtype_t5=dtype_t5, dtype_llama=dtype_llama, device=device, dtype=dtype, model_options=model_options)
return HiDreamTEModel_

12
comfy/text_encoders/hunyuan_image.py
View File

@@ -40,10 +40,10 @@ class HunyuanImageTokenizer(QwenImageTokenizer):
class Qwen25_7BVLIModel(sd1_clip.SDClipModel):
def __init__(self, device="cpu", layer="hidden", layer_idx=-3, dtype=None, attention_mask=True, model_options={}):
llama_scaled_fp8 = model_options.get("qwen_scaled_fp8", None)
if llama_scaled_fp8 is not None:
llama_quantization_metadata = model_options.get("llama_quantization_metadata", None)
if llama_quantization_metadata is not None:
model_options = model_options.copy()
model_options["scaled_fp8"] = llama_scaled_fp8
model_options["quantization_metadata"] = llama_quantization_metadata
super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"pad": 151643}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Qwen25_7BVLI, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
@@ -91,12 +91,12 @@ class HunyuanImageTEModel(QwenImageTEModel):
else:
return super().load_sd(sd)
def te(byt5=True, dtype_llama=None, llama_scaled_fp8=None):
def te(byt5=True, dtype_llama=None, llama_quantization_metadata=None):
class QwenImageTEModel_(HunyuanImageTEModel):
def __init__(self, device="cpu", dtype=None, model_options={}):
if llama_scaled_fp8 is not None and "scaled_fp8" not in model_options:
if llama_quantization_metadata is not None:
model_options = model_options.copy()
model_options["qwen_scaled_fp8"] = llama_scaled_fp8
model_options["llama_quantization_metadata"] = llama_quantization_metadata
if dtype_llama is not None:
dtype = dtype_llama
super().__init__(byt5=byt5, device=device, dtype=dtype, model_options=model_options)

23
comfy/text_encoders/hunyuan_video.py
View File

@@ -6,7 +6,7 @@ from transformers import LlamaTokenizerFast
import torch
import os
import numbers
import comfy.utils
def llama_detect(state_dict, prefix=""):
out = {}
@@ -14,12 +14,9 @@ def llama_detect(state_dict, prefix=""):
if t5_key in state_dict:
out["dtype_llama"] = state_dict[t5_key].dtype
scaled_fp8_key = "{}scaled_fp8".format(prefix)
if scaled_fp8_key in state_dict:
out["llama_scaled_fp8"] = state_dict[scaled_fp8_key].dtype
if "_quantization_metadata" in state_dict:
out["llama_quantization_metadata"] = state_dict["_quantization_metadata"]
quant = comfy.utils.detect_layer_quantization(state_dict, prefix)
if quant is not None:
out["llama_quantization_metadata"] = quant
return out
@@ -31,10 +28,10 @@ class LLAMA3Tokenizer(sd1_clip.SDTokenizer):
class LLAMAModel(sd1_clip.SDClipModel):
def __init__(self, device="cpu", layer="hidden", layer_idx=-3, dtype=None, attention_mask=True, model_options={}, special_tokens={"start": 128000, "pad": 128258}):
llama_scaled_fp8 = model_options.get("llama_scaled_fp8", None)
if llama_scaled_fp8 is not None:
llama_quantization_metadata = model_options.get("llama_quantization_metadata", None)
if llama_quantization_metadata is not None:
model_options = model_options.copy()
model_options["scaled_fp8"] = llama_scaled_fp8
model_options["quantization_metadata"] = llama_quantization_metadata
textmodel_json_config = {}
vocab_size = model_options.get("vocab_size", None)
@@ -161,11 +158,11 @@ class HunyuanVideoClipModel(torch.nn.Module):
return self.llama.load_sd(sd)
def hunyuan_video_clip(dtype_llama=None, llama_scaled_fp8=None):
def hunyuan_video_clip(dtype_llama=None, llama_quantization_metadata=None):
class HunyuanVideoClipModel_(HunyuanVideoClipModel):
def __init__(self, device="cpu", dtype=None, model_options={}):
if llama_scaled_fp8 is not None and "llama_scaled_fp8" not in model_options:
if llama_quantization_metadata is not None:
model_options = model_options.copy()
model_options["llama_scaled_fp8"] = llama_scaled_fp8
model_options["llama_quantization_metadata"] = llama_quantization_metadata
super().__init__(dtype_llama=dtype_llama, device=device, dtype=dtype, model_options=model_options)
return HunyuanVideoClipModel_

6
comfy/text_encoders/lumina2.py
View File

@@ -40,7 +40,7 @@ class LuminaModel(sd1_clip.SD1ClipModel):
super().__init__(device=device, dtype=dtype, name=name, clip_model=clip_model, model_options=model_options)
def te(dtype_llama=None, llama_scaled_fp8=None, model_type="gemma2_2b"):
def te(dtype_llama=None, llama_quantization_metadata=None, model_type="gemma2_2b"):
if model_type == "gemma2_2b":
model = Gemma2_2BModel
elif model_type == "gemma3_4b":
@@ -48,9 +48,9 @@ def te(dtype_llama=None, llama_scaled_fp8=None, model_type="gemma2_2b"):
class LuminaTEModel_(LuminaModel):
def __init__(self, device="cpu", dtype=None, model_options={}):
if llama_scaled_fp8 is not None and "scaled_fp8" not in model_options:
if llama_quantization_metadata is not None:
model_options = model_options.copy()
model_options["scaled_fp8"] = llama_scaled_fp8
model_options["quantization_metadata"] = llama_quantization_metadata
if dtype_llama is not None:
dtype = dtype_llama
super().__init__(device=device, dtype=dtype, name=model_type, model_options=model_options, clip_model=model)

6
comfy/text_encoders/omnigen2.py
View File

@@ -32,12 +32,12 @@ class Omnigen2Model(sd1_clip.SD1ClipModel):
super().__init__(device=device, dtype=dtype, name="qwen25_3b", clip_model=Qwen25_3BModel, model_options=model_options)
def te(dtype_llama=None, llama_scaled_fp8=None):
def te(dtype_llama=None, llama_quantization_metadata=None):
class Omnigen2TEModel_(Omnigen2Model):
def __init__(self, device="cpu", dtype=None, model_options={}):
if llama_scaled_fp8 is not None and "scaled_fp8" not in model_options:
if llama_quantization_metadata is not None:
model_options = model_options.copy()
model_options["scaled_fp8"] = llama_scaled_fp8
model_options["quantization_metadata"] = llama_quantization_metadata
if dtype_llama is not None:
dtype = dtype_llama
super().__init__(device=device, dtype=dtype, model_options=model_options)

5
comfy/text_encoders/ovis.py
View File

@@ -55,12 +55,9 @@ class OvisTEModel(sd1_clip.SD1ClipModel):
return out, pooled, {}
def te(dtype_llama=None, llama_scaled_fp8=None, llama_quantization_metadata=None):
def te(dtype_llama=None, llama_quantization_metadata=None):
class OvisTEModel_(OvisTEModel):
def __init__(self, device="cpu", dtype=None, model_options={}):
if llama_scaled_fp8 is not None and "scaled_fp8" not in model_options:
model_options = model_options.copy()
model_options["scaled_fp8"] = llama_scaled_fp8
if dtype_llama is not None:
dtype = dtype_llama
if llama_quantization_metadata is not None:

6
comfy/text_encoders/pixart_t5.py
View File

@@ -30,12 +30,12 @@ class PixArtTokenizer(sd1_clip.SD1Tokenizer):
def __init__(self, embedding_directory=None, tokenizer_data={}):
super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, clip_name="t5xxl", tokenizer=T5XXLTokenizer)
def pixart_te(dtype_t5=None, t5xxl_scaled_fp8=None):
def pixart_te(dtype_t5=None, t5_quantization_metadata=None):
class PixArtTEModel_(PixArtT5XXL):
def __init__(self, device="cpu", dtype=None, model_options={}):
if t5xxl_scaled_fp8 is not None and "t5xxl_scaled_fp8" not in model_options:
if t5_quantization_metadata is not None:
model_options = model_options.copy()
model_options["t5xxl_scaled_fp8"] = t5xxl_scaled_fp8
model_options["t5xxl_quantization_metadata"] = t5_quantization_metadata
if dtype is None:
dtype = dtype_t5
super().__init__(device=device, dtype=dtype, model_options=model_options)

6
comfy/text_encoders/qwen_image.py
View File

@@ -85,12 +85,12 @@ class QwenImageTEModel(sd1_clip.SD1ClipModel):
return out, pooled, extra
def te(dtype_llama=None, llama_scaled_fp8=None):
def te(dtype_llama=None, llama_quantization_metadata=None):
class QwenImageTEModel_(QwenImageTEModel):
def __init__(self, device="cpu", dtype=None, model_options={}):
if llama_scaled_fp8 is not None and "scaled_fp8" not in model_options:
if llama_quantization_metadata is not None:
model_options = model_options.copy()
model_options["scaled_fp8"] = llama_scaled_fp8
model_options["quantization_metadata"] = llama_quantization_metadata
if dtype_llama is not None:
dtype = dtype_llama
super().__init__(device=device, dtype=dtype, model_options=model_options)

19
comfy/text_encoders/sd3_clip.py
View File

@@ -6,14 +6,15 @@ import torch
import os
import comfy.model_management
import logging
import comfy.utils
class T5XXLModel(sd1_clip.SDClipModel):
def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, attention_mask=False, model_options={}):
textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "t5_config_xxl.json")
t5xxl_scaled_fp8 = model_options.get("t5xxl_scaled_fp8", None)
if t5xxl_scaled_fp8 is not None:
t5xxl_quantization_metadata = model_options.get("t5xxl_quantization_metadata", None)
if t5xxl_quantization_metadata is not None:
model_options = model_options.copy()
model_options["scaled_fp8"] = t5xxl_scaled_fp8
model_options["quantization_metadata"] = t5xxl_quantization_metadata
model_options = {**model_options, "model_name": "t5xxl"}
super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"end": 1, "pad": 0}, model_class=comfy.text_encoders.t5.T5, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
@@ -25,9 +26,9 @@ def t5_xxl_detect(state_dict, prefix=""):
if t5_key in state_dict:
out["dtype_t5"] = state_dict[t5_key].dtype
scaled_fp8_key = "{}scaled_fp8".format(prefix)
if scaled_fp8_key in state_dict:
out["t5xxl_scaled_fp8"] = state_dict[scaled_fp8_key].dtype
quant = comfy.utils.detect_layer_quantization(state_dict, prefix)
if quant is not None:
out["t5_quantization_metadata"] = quant
return out
@@ -156,11 +157,11 @@ class SD3ClipModel(torch.nn.Module):
else:
return self.t5xxl.load_sd(sd)
def sd3_clip(clip_l=True, clip_g=True, t5=True, dtype_t5=None, t5xxl_scaled_fp8=None, t5_attention_mask=False):
def sd3_clip(clip_l=True, clip_g=True, t5=True, dtype_t5=None, t5_quantization_metadata=None, t5_attention_mask=False):
class SD3ClipModel_(SD3ClipModel):
def __init__(self, device="cpu", dtype=None, model_options={}):
if t5xxl_scaled_fp8 is not None and "t5xxl_scaled_fp8" not in model_options:
if t5_quantization_metadata is not None:
model_options = model_options.copy()
model_options["t5xxl_scaled_fp8"] = t5xxl_scaled_fp8
model_options["t5xxl_quantization_metadata"] = t5_quantization_metadata
super().__init__(clip_l=clip_l, clip_g=clip_g, t5=t5, dtype_t5=dtype_t5, t5_attention_mask=t5_attention_mask, device=device, dtype=dtype, model_options=model_options)
return SD3ClipModel_

6
comfy/text_encoders/wan.py
View File

@@ -25,12 +25,12 @@ class WanT5Model(sd1_clip.SD1ClipModel):
def __init__(self, device="cpu", dtype=None, model_options={}, **kwargs):
super().__init__(device=device, dtype=dtype, model_options=model_options, name="umt5xxl", clip_model=UMT5XXlModel, **kwargs)
def te(dtype_t5=None, t5xxl_scaled_fp8=None):
def te(dtype_t5=None, t5_quantization_metadata=None):
class WanTEModel(WanT5Model):
def __init__(self, device="cpu", dtype=None, model_options={}):
if t5xxl_scaled_fp8 is not None and "scaled_fp8" not in model_options:
if t5_quantization_metadata is not None:
model_options = model_options.copy()
model_options["scaled_fp8"] = t5xxl_scaled_fp8
model_options["quantization_metadata"] = t5_quantization_metadata
if dtype_t5 is not None:
dtype = dtype_t5
super().__init__(device=device, dtype=dtype, model_options=model_options)

5
comfy/text_encoders/z_image.py
View File

@@ -34,12 +34,9 @@ class ZImageTEModel(sd1_clip.SD1ClipModel):
super().__init__(device=device, dtype=dtype, name="qwen3_4b", clip_model=Qwen3_4BModel, model_options=model_options)
def te(dtype_llama=None, llama_scaled_fp8=None, llama_quantization_metadata=None):
def te(dtype_llama=None, llama_quantization_metadata=None):
class ZImageTEModel_(ZImageTEModel):
def __init__(self, device="cpu", dtype=None, model_options={}):
if llama_scaled_fp8 is not None and "scaled_fp8" not in model_options:
model_options = model_options.copy()
model_options["scaled_fp8"] = llama_scaled_fp8
if dtype_llama is not None:
dtype = dtype_llama
if llama_quantization_metadata is not None: