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, 01:00 UTC
[Misc] Fix PhiMoE expert mapping (#21085)
Signed-off-by: Jee Jee Li <pandaleefree@gmail.com>
Jee Jee Li committed
Jul 17, 2025 at 13:47 UTC
c5b8b5953a2e20e8358d0828aad11d259c073c50
1 file changed
+1
-6
vllm/model_executor/models/phimoe.py
+1
-6
@@ -533,14 +533,9 @@ class PhiMoEModel(nn.Module):
533
("qkv_proj", "v_proj", "v"),
534
]
535
536
- expert_params_mapping = FusedMoE.make_expert_params_mapping(
537
- ckpt_gate_proj_name="w1",
538
- ckpt_down_proj_name="w2",
539
- ckpt_up_proj_name="w3",
540
- num_experts=self.config.num_local_experts)
541
-
536
params_dict = dict(self.named_parameters())
537
loaded_params: set[str] = set()
538
+ expert_params_mapping = self.get_expert_mapping()
539
for name, loaded_weight in weights:
540
if (self.quant_config is not None and
541
(scale_name := self.quant_config.get_cache_scale(name))):