[Model] Fix weight prefix mapping for native Qwen3.5 text-only checkp… (#50355)
Signed-off-by: Zhu, Zufang <zufang.zhu@intel.com> Co-authored-by: opencode <opencode@anomaly.co>
This commit is contained in:
@@ -502,7 +502,11 @@ _TEXT_GENERATION_EXAMPLE_MODELS = {
|
||||
"Qwen3ForCausalLM": _HfExamplesInfo("Qwen/Qwen3-8B"),
|
||||
"Qwen3MoeForCausalLM": _HfExamplesInfo("Qwen/Qwen3-30B-A3B"),
|
||||
"Qwen3_5ForCausalLM": _HfExamplesInfo("codecho/Qwen3.5-0.8B-text-only"),
|
||||
"Qwen3_5MoeForCausalLM": _HfExamplesInfo("codecho/Qwen3.5-35B-A3B-text-only"),
|
||||
"Qwen3_5MoeForCausalLM": _HfExamplesInfo(
|
||||
"codecho/Qwen3.5-35B-A3B-text-only",
|
||||
extras={"native-prefix": "imdatta0/small_qwen3_5_20b"},
|
||||
max_model_len=4096,
|
||||
),
|
||||
"MellumForCausalLM": _HfExamplesInfo("JetBrains/Mellum2-12B-A2.5B-Base"),
|
||||
"Qwen3NextForCausalLM": _HfExamplesInfo(
|
||||
"Qwen/Qwen3-Next-80B-A3B-Instruct",
|
||||
|
||||
@@ -303,6 +303,13 @@ class Qwen3_5ForCausalLMBase(
|
||||
"in_proj_ba": ["in_proj_b", "in_proj_a"],
|
||||
}
|
||||
|
||||
# Some community text-only checkpoints keep the extraneous
|
||||
# `model.language_model.` prefix inherited from the VL training stack.
|
||||
# Strip it so both prefixed and clean checkpoints load correctly.
|
||||
hf_to_vllm_mapper = WeightsMapper(
|
||||
orig_to_new_prefix={"model.language_model.": "model."},
|
||||
)
|
||||
|
||||
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
|
||||
config = vllm_config.model_config.hf_text_config
|
||||
self.vllm_config = vllm_config
|
||||
@@ -416,7 +423,7 @@ class Qwen3_5ForCausalLMBase(
|
||||
self,
|
||||
skip_prefixes=["mtp."],
|
||||
)
|
||||
return loader.load_weights(weights)
|
||||
return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)
|
||||
|
||||
|
||||
class Qwen3_5ForCausalLM(Qwen3_5ForCausalLMBase):
|
||||
|
||||
Reference in New Issue
Block a user