xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-05-30 03:37:04 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Bugfix] Fix grouped_topk pytorch impl when num_experts can't be grouped properly (#29439)

Browse Source 
Signed-off-by: Divakar Verma <divakar.verma@amd.com>
Co-authored-by: Gregory Shtrasberg <156009573+gshtras@users.noreply.github.com>
Co-authored-by: TJian <tunjian.tan@embeddedllm.com>
This commit is contained in:
Divakar Verma 2025-12-10 21:47:18 -06:00 committed by GitHub
parent b51255f369
commit d1e1fb4363
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 9 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

10
vllm/model_executor/layers/fused_moe/layer.py
View File

@ -1556,6 +1556,14 @@ class FusedMoE(CustomOp):
f"EPLB is not supported for {self.quant_method.method_name}." f"EPLB is not supported for {self.quant_method.method_name}."
) )
def valid_grouping() -> bool:
# Check if num_experts is greater than num_expert_group
# and is divisible by num_expert_group
num_experts = router_logits.shape[-1]
if num_experts <= self.num_expert_group:
return False
return num_experts % self.num_expert_group == 0
indices_type = self.quant_method.topk_indices_dtype indices_type = self.quant_method.topk_indices_dtype
# Check if we should use a routing simulation strategy # Check if we should use a routing simulation strategy
@ -1570,7 +1578,7 @@ class FusedMoE(CustomOp):
) )
# DeepSeekv2 uses grouped_top_k # DeepSeekv2 uses grouped_top_k
elif self.use_grouped_topk: elif self.use_grouped_topk and valid_grouping():
assert self.topk_group is not None assert self.topk_group is not None
assert self.num_expert_group is not None assert self.num_expert_group is not None
if rocm_aiter_ops.is_fused_moe_enabled(): if rocm_aiter_ops.is_fused_moe_enabled():