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 7, 11:30 UTC
[BugFix][Model]Fix the garbled code in Ernie4.5-VL caused by fast_moe_cold_start (#35587)
Signed-off-by: wangyafeng <wangyafeng@baidu.com>
CSWYF3634076 committed
Mar 3, 2026 at 02:56 UTC
2a9e3347e9fbefc4bf991b60dc45a8c156d8696f
1 file changed
+9
vllm/model_executor/models/config.py
+9
@@ -329,6 +329,14 @@ class SnowflakeGteNewModelConfig(VerifyAndUpdateConfig):
329
}
330
331
332
+class Ernie4_5_VLMoeForConditionalGenerationConfig(VerifyAndUpdateConfig):
333
+ @staticmethod
334
+ def verify_and_update_config(vllm_config: "VllmConfig") -> None:
335
+ # Ernie4.5-VL conditionally executes text/vision MoE branches, so
336
+ # fast_moe_cold_start can silently produce incorrect execution order.
337
+ vllm_config.compilation_config.fast_moe_cold_start = False
338
+
339
+
340
class GptOssForCausalLMConfig(VerifyAndUpdateConfig):
341
@staticmethod
342
def verify_and_update_config(vllm_config: "VllmConfig") -> None:
@@ -661,6 +669,7 @@ MODELS_CONFIG_MAP: dict[str, type[VerifyAndUpdateConfig]] = {
669
"Qwen2ForRewardModel": Qwen2ForRewardModelConfig,
670
"Qwen3ForSequenceClassification": Qwen3ForSequenceClassificationConfig,
671
"Qwen3VLForSequenceClassification": Qwen3VLForSequenceClassificationConfig,
672
+ "Ernie4_5_VLMoeForConditionalGeneration": Ernie4_5_VLMoeForConditionalGenerationConfig, # noqa: E501
673
"XLMRobertaModel": JinaRobertaModelConfig,
674
"ColBERTJinaRobertaModel": JinaRobertaModelConfig,
675
"JinaVLForRanking": JinaVLForSequenceClassificationConfig,