xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-15 18:25:45 +08:00
Code Issues Packages Projects Releases Wiki Activity

[V1] [Bugfix] eagle bugfix and enable correct lm_head for multimodal (#18034)

Browse Source 
Signed-off-by: Ronald Xu <ronaldxu@amazon.com>
This commit is contained in:
RonaldBXu 2025-05-23 00:37:18 -07:00 committed by GitHub
parent 60cad94b86
commit 4c611348a7
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 6 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
vllm/transformers_utils/configs/eagle.py
View File

@ -70,8 +70,7 @@ class EAGLEConfig(PretrainedConfig):
if self.model is not None: if self.model is not None:
for k, v in self.model.to_dict().items(): for k, v in self.model.to_dict().items():
if not hasattr(self, k): setattr(self, k, v)
setattr(self, k, v)
@classmethod @classmethod
def from_pretrained( def from_pretrained(

6
vllm/v1/spec_decode/eagle.py
View File

@ -9,6 +9,7 @@ from vllm.distributed.parallel_state import get_pp_group
from vllm.forward_context import set_forward_context from vllm.forward_context import set_forward_context
from vllm.logger import init_logger from vllm.logger import init_logger
from vllm.model_executor.model_loader import get_model from vllm.model_executor.model_loader import get_model
from vllm.model_executor.models import supports_multimodal
from vllm.model_executor.models.llama_eagle3 import Eagle3LlamaForCausalLM from vllm.model_executor.models.llama_eagle3 import Eagle3LlamaForCausalLM
from vllm.triton_utils import tl, triton from vllm.triton_utils import tl, triton
from vllm.v1.attention.backends.flash_attn import FlashAttentionMetadata from vllm.v1.attention.backends.flash_attn import FlashAttentionMetadata
@ -310,7 +311,10 @@ class EagleProposer:
if self.vllm_config.speculative_config.method != "eagle3" and \ if self.vllm_config.speculative_config.method != "eagle3" and \
hasattr(target_model, "lm_head"): hasattr(target_model, "lm_head"):
logger.info("Loading EAGLE LM head weights from the target model.") logger.info("Loading EAGLE LM head weights from the target model.")
self.model.lm_head = target_model.lm_head if supports_multimodal(target_model):
self.model.lm_head = target_model.get_language_model().lm_head
else:
self.model.lm_head = target_model.lm_head
@torch.inference_mode() @torch.inference_mode()
def dummy_run( def dummy_run(