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, 12:45 UTC
Fix Qwen3_5MTP packed_modules_mapping for gate_up_proj (#35581)
cwazai committed
Feb 28, 2026 at 22:50 UTC
63d7972f13d1c5a9d9bd55b664017067a9abd451
1 file changed
+1
-1
vllm/model_executor/models/qwen3_5_mtp.py
+1
-1
@@ -339,7 +339,7 @@ class Qwen3_5MTP(nn.Module, SupportsMultiModal):
339
"k_proj",
340
"v_proj",
341
],
342
- "gate_up_proj": ["up_proj", "down_proj"],
342
+ "gate_up_proj": ["gate_proj", "up_proj"],
343
}
344
345
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):