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 2, 22:00 UTC
fix(eagle3): read norm_before_fc from eagle_config for NVIDIA checkpoint (#42143)
Signed-off-by: FERRARIZHENG <popkart06@gmail.com>
GuangYaoZheng committed
May 23, 2026 at 16:21 UTC
3f3e8626813fca2d070a054fdf29faf77b7feb19
1 file changed
+7
-3
vllm/model_executor/models/llama_eagle3.py
+7
-3
@@ -145,12 +145,16 @@ class LlamaModel(nn.Module):
145
# Get drafter's quantization config
146
self.quant_config = get_draft_quant_config(vllm_config)
147
148
- eagle_config = getattr(self.config, "eagle_config", None)
149
- if eagle_config is not None and "use_aux_hidden_state" in eagle_config:
148
+ eagle_config = getattr(self.config, "eagle_config", None) or {}
149
+ if "use_aux_hidden_state" in eagle_config:
150
self.use_aux_hidden_state = eagle_config["use_aux_hidden_state"]
151
else:
152
self.use_aux_hidden_state = True
153
- self.norm_before_fc = getattr(self.config, "norm_before_fc", False)
153
+ self.norm_before_fc = bool(
154
+ eagle_config.get(
155
+ "norm_before_fc", getattr(self.config, "norm_before_fc", False)
156
+ )
157
+ )
158
159
current_vllm_config = get_current_vllm_config()
160