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
Revert "[V1] [Bugfix] eagle bugfix and enable correct lm_head for multimodal (#18034)" (#18600)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
Cyrus Leung committed
May 23, 2025 at 17:18 UTC
fbb13a2c15bf505914667ae5db9d5cde0847c7b8
2 files changed
+3
-6
vllm/transformers_utils/configs/eagle.py
+2
-1
@@ -70,7 +70,8 @@ class EAGLEConfig(PretrainedConfig):
70
71
if self.model is not None:
72
for k, v in self.model.to_dict().items():
73
- setattr(self, k, v)
73
+ if not hasattr(self, k):
74
+ setattr(self, k, v)
75
76
@classmethod
77
def from_pretrained(
vllm/v1/spec_decode/eagle.py
+1
-5
@@ -9,7 +9,6 @@ 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
12
from vllm.model_executor.models.llama_eagle3 import Eagle3LlamaForCausalLM
13
from vllm.triton_utils import tl, triton
14
from vllm.v1.attention.backends.flash_attn import FlashAttentionMetadata
@@ -311,10 +310,7 @@ class EagleProposer:
310
if self.vllm_config.speculative_config.method != "eagle3" and \
311
hasattr(target_model, "lm_head"):
312
logger.info("Loading EAGLE LM head weights from the target model.")
314
- if supports_multimodal(target_model):
315
- self.model.lm_head = target_model.get_language_model().lm_head
316
- else:
317
- self.model.lm_head = target_model.lm_head
313
+ self.model.lm_head = target_model.lm_head
314
315
@torch.inference_mode()
316
def dummy_run(