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, 11:00 UTC
[Model] Add packed_modules_mapping for Qwen3-MOE (#18118)
Signed-off-by: Jee Jee Li <pandaleefree@gmail.com>
Jee Jee Li committed
May 14, 2025 at 17:13 UTC
63dc3426e0789935acebd8d3f442324f9b2bbd32
1 file changed
+11
vllm/model_executor/models/qwen3_moe.py
+11
@@ -475,6 +475,17 @@ class Qwen3MoeModel(nn.Module):
475
476
477
class Qwen3MoeForCausalLM(nn.Module, SupportsPP):
478
+ packed_modules_mapping = {
479
+ "qkv_proj": [
480
+ "q_proj",
481
+ "k_proj",
482
+ "v_proj",
483
+ ],
484
+ "gate_up_proj": [
485
+ "gate_proj",
486
+ "up_proj",
487
+ ],
488
+ }
489
490
fall_back_to_pt_during_load = False
491