xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-06-08 08:22:15 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Hardware][Gaudi][BugFix] fix arguments of hpu fused moe (#15945)

Browse Source 
Signed-off-by: zhenwei <zhenweiliu@habana.ai>
This commit is contained in:
liuzhenwei 2025-04-05 00:38:55 +08:00 committed by GitHub
parent bf7e3c51ae
commit 0812d8dd41
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 5 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
vllm/model_executor/layers/fused_moe/layer.py
View File

@ -254,9 +254,12 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
renormalize: bool, renormalize: bool,
topk_group: Optional[int] = None, topk_group: Optional[int] = None,
num_expert_group: Optional[int] = None, num_expert_group: Optional[int] = None,
global_num_experts: int = -1,
expert_map: Optional[torch.Tensor] = None,
custom_routing_function: Optional[Callable] = None, custom_routing_function: Optional[Callable] = None,
scoring_func: str = "softmax", scoring_func: str = "softmax",
e_score_correction_bias: Optional[torch.Tensor] = None e_score_correction_bias: Optional[torch.Tensor] = None,
activation: str = "silu",
) -> torch.Tensor: ) -> torch.Tensor:
assert not use_grouped_topk assert not use_grouped_topk
assert num_expert_group is None assert num_expert_group is None
@ -472,7 +475,7 @@ class FusedMoE(torch.nn.Module):
"non-grouped topk.") "non-grouped topk.")
if current_platform.is_hpu(): if current_platform.is_hpu():
from vllm_hpu_extension.ops import DynamicFusedMOE from vllm_hpu_extension.ops import DynamicFusedMOE
self.hpu_fused_moe = DynamicFusedMOE(self.num_experts) self.hpu_fused_moe = DynamicFusedMOE(self.global_num_experts)
# Note: get_quant_method will look at the layer's local_num_experts # Note: get_quant_method will look at the layer's local_num_experts
# for heuristic purposes, so it must be initialized first. # for heuristic purposes, so it must be initialized first.