[fix] fix compatibility for non-EPLB scenarios (#1142)

### What this PR does / why we need it?
Fix incompatibility problem for non-EPLB scenarios in #1116 

### Does this PR introduce _any_ user-facing change?

No.

### How was this patch tested?

Tested with online serving and e2e CI.

Signed-off-by: linfeng-yuan <1102311262@qq.com>
This commit is contained in:
linfeng-yuan
2025-06-10 08:39:24 +08:00
committed by GitHub
parent 571f88f85e
commit 706de02317

View File

@@ -150,8 +150,8 @@ def fused_experts_with_mc2(hidden_states: torch.Tensor,
log2phy: torch.Tensor = None,
global_redundant_expert_num: int = 0,
**kwargs) -> torch.Tensor:
topk_ids = log2phy[topk_ids]
if log2phy:
topk_ids = log2phy[topk_ids]
global_bs = 0
moe_expert_num = len(expert_map) + global_redundant_expert_num
# hidden_states = hidden_states.bfloat16()
@@ -278,7 +278,8 @@ def fused_experts_with_all2all(
log2phy: torch.Tensor = None,
global_redundant_expert_num: int = 0,
):
topk_ids = log2phy[topk_ids]
if log2phy:
topk_ids = log2phy[topk_ids]
original_shape = hidden_states.shape
if len(original_shape) == 3:
hidden_states = hidden_states.view(-1, hidden_states.shape[-1])