xinyun/ComfyUI
1
0
Fork 0
mirror of https://git.datalinker.icu/comfyanonymous/ComfyUI synced 2025-12-13 16:04:34 +08:00
Code Issues Packages Projects Releases Wiki Activity

Fix qwen scaled fp8 not working with kandinsky. Make basic t2i wf work. (#11162)

Browse Source
This commit is contained in:
comfyanonymous 2025-12-06 17:50:10 -08:00 committed by GitHub
parent 4086acf3c2
commit 329480da5a
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 13 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
comfy/ldm/kandinsky5/model.py
View File

@ -387,6 +387,9 @@ class Kandinsky5(nn.Module):
return self.out_layer(visual_embed, time_embed) return self.out_layer(visual_embed, time_embed)
def _forward(self, x, timestep, context, y, time_dim_replace=None, transformer_options={}, **kwargs): def _forward(self, x, timestep, context, y, time_dim_replace=None, transformer_options={}, **kwargs):
original_dims = x.ndim
if original_dims == 4:
x = x.unsqueeze(2)
bs, c, t_len, h, w = x.shape bs, c, t_len, h, w = x.shape
x = comfy.ldm.common_dit.pad_to_patch_size(x, self.patch_size) x = comfy.ldm.common_dit.pad_to_patch_size(x, self.patch_size)
@ -397,7 +400,10 @@ class Kandinsky5(nn.Module):
freqs = self.rope_encode_3d(t_len, h, w, device=x.device, dtype=x.dtype, transformer_options=transformer_options) freqs = self.rope_encode_3d(t_len, h, w, device=x.device, dtype=x.dtype, transformer_options=transformer_options)
freqs_text = self.rope_encode_1d(context.shape[1], device=x.device, dtype=x.dtype, transformer_options=transformer_options) freqs_text = self.rope_encode_1d(context.shape[1], device=x.device, dtype=x.dtype, transformer_options=transformer_options)
return self.forward_orig(x, timestep, context, y, freqs, freqs_text, transformer_options=transformer_options, **kwargs) out = self.forward_orig(x, timestep, context, y, freqs, freqs_text, transformer_options=transformer_options, **kwargs)
if original_dims == 4:
out = out.squeeze(2)
return out
def forward(self, x, timestep, context, y, time_dim_replace=None, transformer_options={}, **kwargs): def forward(self, x, timestep, context, y, time_dim_replace=None, transformer_options={}, **kwargs):
return comfy.patcher_extension.WrapperExecutor.new_class_executor( return comfy.patcher_extension.WrapperExecutor.new_class_executor(

12
comfy/text_encoders/kandinsky5.py
View File

@ -24,10 +24,10 @@ class Kandinsky5TokenizerImage(Kandinsky5Tokenizer):
class Qwen25_7BVLIModel(sd1_clip.SDClipModel): class Qwen25_7BVLIModel(sd1_clip.SDClipModel):
def __init__(self, device="cpu", layer="hidden", layer_idx=-1, dtype=None, attention_mask=True, model_options={}): def __init__(self, device="cpu", layer="hidden", layer_idx=-1, dtype=None, attention_mask=True, model_options={}):
llama_scaled_fp8 = model_options.get("qwen_scaled_fp8", None) llama_quantization_metadata = model_options.get("llama_quantization_metadata", None)
if llama_scaled_fp8 is not None: if llama_quantization_metadata is not None:
model_options = model_options.copy() 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=Qwen25_7BVLI, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=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=Qwen25_7BVLI, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
@ -56,12 +56,12 @@ class Kandinsky5TEModel(QwenImageTEModel):
else: else:
return super().load_sd(sd) return super().load_sd(sd)
def te(dtype_llama=None, llama_scaled_fp8=None): def te(dtype_llama=None, llama_quantization_metadata=None):
class Kandinsky5TEModel_(Kandinsky5TEModel): class Kandinsky5TEModel_(Kandinsky5TEModel):
def __init__(self, device="cpu", dtype=None, model_options={}): 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 = 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: if dtype_llama is not None:
dtype = dtype_llama dtype = dtype_llama
super().__init__(device=device, dtype=dtype, model_options=model_options) super().__init__(device=device, dtype=dtype, model_options=model_options)