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

[Feature] Enhance EAGLE Architecture with Proper RMS Norms (#14990)

Browse Source 
Signed-off-by: Bryan Lu <yuzhelu@amazon.com>
Co-authored-by: Cyrus Leung <cyrus.tl.leung@gmail.com>
This commit is contained in:
Bryan Lu
2025-03-26 01:24:07 -07:00
committed by GitHub
parent 5aefd6ac31
commit 781d056280
3 changed files with 70 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
vllm/transformers_utils/configs/eagle.py
View File

@@ -5,6 +5,8 @@ from typing import Optional, Union
from transformers import AutoConfig, PretrainedConfig
from vllm.transformers_utils.configs.deepseek_vl2 import DeepseekV2Config
class EAGLEConfig(PretrainedConfig):
model_type = "eagle"
@@ -14,8 +16,17 @@ class EAGLEConfig(PretrainedConfig):
truncated_vocab_size: Optional[int] = None,
**kwargs):
model_config = None if model is None else (AutoConfig.for_model(
**model) if isinstance(model, dict) else model)
model_config: Union[PretrainedConfig, DeepseekV2Config, None]
if isinstance(model, dict):
archs = model.get("architectures", [])
target_archs = ["DeepseekV2ForCausalLM", "DeepseekV3ForCausalLM"]
if any(target_arch in archs for target_arch in target_archs):
# AutoConfig does not support DeepSeek MoE models yet
model_config = DeepseekV2Config(**model)
else:
model_config = AutoConfig.for_model(**model)
else:
model_config = model
for k, v in kwargs.items():
if k != "architectures" and k != "model_type" and hasattr(