A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 15, 08:30 UTC
[Bugfix] Fix the lm_head in gpt_bigcode in lora mode (#6357)
Signed-off-by: Max de Bayser <mbayser@br.ibm.com> Signed-off-by: Max de Bayser <maxdebayser@gmail.com>
Maximilien de Bayser committed
May 26, 2025 at 03:52 UTC
561b77a0d608a9059318d6cff9f0975439880d77
1 file changed
+5
-8
vllm/model_executor/models/gpt_bigcode.py
+5
-8
@@ -272,12 +272,6 @@ class GPTBigCodeModel(nn.Module):
272
class GPTBigCodeForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
273
packed_modules_mapping = {"c_attn": ["c_attn"]}
274
275
- # LoRA specific attributes
276
- embedding_modules = {
277
- "wte": "input_embeddings",
278
- "lm_head": "output_embeddings",
279
- }
280
-
275
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
276
super().__init__()
277
config = vllm_config.model_config.hf_config
@@ -330,8 +324,11 @@ class GPTBigCodeForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
324
325
def load_weights(self, weights: Iterable[tuple[str,
326
torch.Tensor]]) -> set[str]:
327
+ skip_prefixes = None
328
+ if self.config.tie_word_embeddings:
329
+ skip_prefixes = ["lm_head."]
330
loader = AutoWeightsLoader(
331
self,
335
- skip_prefixes=(["lm_head."]),
332
+ skip_prefixes=skip_prefixes,
333
)
337
- return loader.load_weights(weights)
\ No newline at end of file
334
+ return loader.load_weights(weights)