biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Bugfix] Support HF sharded weights for Mistral3/Pixtral models (#32673)

Browse Source 
Signed-off-by: ricky-chaoju <ricky.chen@infinirc.com>
Signed-off-by: vllm-dev <ricky.chen@infinirc.com>
This commit is contained in:
RickyChen / 陳昭儒
2026-01-21 15:27:30 +08:00
committed by GitHub
parent 360aa93f8f
commit f23fb5a7c1
1 changed files with 14 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
vllm/model_executor/models/pixtral.py
View File

@@ -502,10 +502,12 @@ class PixtralForConditionalGeneration(
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]):
def is_vision_encoder_weights(weight: tuple[str, torch.Tensor]):
return weight[0].startswith("vision_encoder")
return weight[0].startswith(("vision_encoder", "vision_tower"))
def is_vision_lang_adapter_weights(weight: tuple[str, torch.Tensor]):
return weight[0].startswith("vision_language_adapter")
return weight[0].startswith(
("vision_language_adapter", "multi_modal_projector")
)
def is_patch_merger(weight: tuple[str, torch.Tensor]):
return weight[0].startswith("patch_merger")
@@ -543,7 +545,8 @@ class PixtralForConditionalGeneration(
continue
# Load vision encoder weights directly
trimmed_name = ".".join(name.split(".")[1:])
param = vision_encoder_dict[trimmed_name]
param = vision_encoder_dict.get(trimmed_name)
if param is not None:
with torch.no_grad():
default_weight_loader(param, w)
elif is_patch_merger((name, w)):
@@ -567,12 +570,15 @@ class PixtralForConditionalGeneration(
continue
# Load vision-language adapter weights directly
trimmed_name = ".".join(name.split(".")[1:])
param = vision_lang_adapter_dict[trimmed_name]
param = vision_lang_adapter_dict.get(trimmed_name)
if param is not None:
with torch.no_grad():
default_weight_loader(param, w)
else:
# LLM weights: yield them to be loaded
# by language_model.load_weights
# Strip "language_model." prefix if present (HF sharded format)
name = name.removeprefix("language_model.")
yield (name, w)
# Now we call the language model load with the generator