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
[V1] [Bugfix] eagle bugfix and enable correct lm_head for multimodal (2) (#18781)
Signed-off-by: Ronald Xu <ronaldxu@amazon.com>
RonaldBXu committed
May 27, 2025 at 22:09 UTC
5e13c07d00df0ca906c4c06110c277421202d3e2
2 files changed
+6
-2
vllm/transformers_utils/configs/eagle.py
+1
-1
@@ -70,7 +70,7 @@ class EAGLEConfig(PretrainedConfig):
70
71
if self.model is not None:
72
for k, v in self.model.to_dict().items():
73
- if not hasattr(self, k):
73
+ if k not in kwargs:
74
setattr(self, k, v)
75
76
@classmethod
vllm/v1/spec_decode/eagle.py
+5
-1
@@ -9,6 +9,7 @@ from vllm.distributed.parallel_state import get_pp_group
9
from vllm.forward_context import set_forward_context
10
from vllm.logger import init_logger
11
from vllm.model_executor.model_loader import get_model
12
+from vllm.model_executor.models import supports_multimodal
13
from vllm.model_executor.models.llama_eagle3 import Eagle3LlamaForCausalLM
14
from vllm.v1.attention.backends.flash_attn import (CommonAttentionMetadata,
15
FlashAttentionMetadata)
@@ -346,7 +347,10 @@ class EagleProposer:
347
if self.vllm_config.speculative_config.method != "eagle3" and \
348
hasattr(target_model, "lm_head"):
349
logger.info("Loading EAGLE LM head weights from the target model.")
349
- self.model.lm_head = target_model.lm_head
350
+ if supports_multimodal(target_model):
351
+ self.model.lm_head = target_model.get_language_model().lm_head
352
+ else:
353
+ self.model.lm_head = target_model.lm_head
354
355
@torch.inference_mode()
356
def dummy_run(