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

[Misc] Move weights mapper (#11443)

Browse Source 
Signed-off-by: Jee Jee Li <pandaleefree@gmail.com>
This commit is contained in:
Jee Jee Li
2024-12-24 21:05:25 +08:00
committed by GitHub
parent 5c7963249d
commit 196c34b0ac
8 changed files with 74 additions and 68 deletions
Download Patch File Download Diff File Expand all files Collapse all files

27
vllm/model_executor/models/telechat2.py
View File

@@ -31,6 +31,19 @@ from .utils import (AutoWeightsLoader, PPMissingLayer, WeightsMapper,
class TeleChat2Model(LlamaModel):
hf_to_vllm_mapper = WeightsMapper(
orig_to_new_prefix={
"transformer.": "model.",
},
orig_to_new_substr={
".h.": ".layers.",
".self_attention.": ".self_attn.",
".word_embeddings.": ".embed_tokens.",
".dense.": ".o_proj.",
".ln_f.": ".norm.",
},
)
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
# 1. Initialize the LlamaModel with bias
vllm_config.model_config.hf_config.bias = True
@@ -111,21 +124,9 @@ class TeleChat2ForCausalLM(LlamaForCausalLM):
def load_weights(self, weights: Iterable[Tuple[str,
torch.Tensor]]) -> Set[str]:
hf_to_vllm_mapper = WeightsMapper(
orig_to_new_prefix={
"transformer.": "model.",
},
orig_to_new_substr={
".h.": ".layers.",
".self_attention.": ".self_attn.",
".word_embeddings.": ".embed_tokens.",
".dense.": ".o_proj.",
".ln_f.": ".norm.",
},
)
loader = AutoWeightsLoader(
self,
skip_prefixes=(["lm_head."]
if self.config.tie_word_embeddings else None),
)
return loader.load_weights(weights, mapper=hf_to_vllm_mapper)
return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)