biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Quantization][MoE] remove unused ep logic from moe marlin (#31571)

Browse Source 
Signed-off-by: Jinzhen Lin <jinzhen.ljz@antgroup.com>
Co-authored-by: Michael Goin <mgoin64@gmail.com>
This commit is contained in:
Jinzhen Lin
2026-01-07 01:07:19 +08:00
committed by GitHub
parent 28c94770ad
commit 2f4bdee61e
6 changed files with 31 additions and 60 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
vllm/_custom_ops.py
View File

@@ -2162,7 +2162,6 @@ def moe_wna16_marlin_gemm(
moe_block_size: int,
top_k: int,
mul_topk_weights: bool,
is_ep: bool,
b_q_type: ScalarType,
size_m: int,
size_n: int,
@@ -2194,7 +2193,6 @@ def moe_wna16_marlin_gemm(
moe_block_size,
top_k,
mul_topk_weights,
is_ep,
b_q_type.id,
size_m,
size_n,
@@ -2256,7 +2254,6 @@ if hasattr(torch.ops, "_moe_C") and hasattr(torch.ops._moe_C, "marlin_gemm_moe")
moe_block_size: int,
top_k: int,
mul_topk_weights: bool,
is_ep: bool,
b_q_type: ScalarType,
size_m: int,
size_n: int,

2
vllm/model_executor/layers/fused_moe/fused_marlin_moe.py
View File

@@ -135,7 +135,6 @@ def _fused_marlin_moe(
moe_block_size=block_size_m,
top_k=num_topk,
mul_topk_weights=apply_router_weight_on_input,
is_ep=expert_map is not None,
b_q_type=quant_type,
size_m=M,
size_n=2 * N,
@@ -187,7 +186,6 @@ def _fused_marlin_moe(
moe_block_size=block_size_m,
top_k=1,
mul_topk_weights=not apply_router_weight_on_input,
is_ep=expert_map is not None,
b_q_type=quant_type,
size_m=M * num_topk,
size_n=K,