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

[Bugfix][Kernel] fix bias adding in triton kernel implemented fused moe (#31676)

Browse Source 
Signed-off-by: xuebwang-amd <xuebwang@amd.com>
This commit is contained in:
xuebwang-amd
2026-01-07 15:36:13 +08:00
committed by GitHub
parent 4614c5a539
commit 0dd5dee9b9
1 changed files with 8 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
vllm/model_executor/layers/fused_moe/fused_moe.py
View File

@@ -518,11 +518,7 @@ def fused_moe_kernel(
# Advance the ptrs to the next K block.
a_ptrs += BLOCK_SIZE_K * stride_ak
b_ptrs += BLOCK_SIZE_K * stride_bk
if HAS_BIAS:
accumulator = accumulator + bias[None, :]
if MUL_ROUTED_WEIGHT:
moe_weight = tl.load(topk_weights_ptr + offs_token, mask=token_mask, other=0)
accumulator = accumulator * moe_weight[:, None]
if use_int8_w8a16:
accumulator = (accumulator * b_scale).to(compute_type)
elif use_fp8_w8a8 or use_int8_w8a8:
@@ -533,6 +529,13 @@ def fused_moe_kernel(
else:
accumulator = accumulator.to(compute_type)
# Since bias is typically not quantized, it's added after dequantization.
if HAS_BIAS:
accumulator = accumulator + bias[None, :]
if MUL_ROUTED_WEIGHT:
moe_weight = tl.load(topk_weights_ptr + offs_token, mask=token_mask, other=0)
accumulator = accumulator * moe_weight[:, None]
# -----------------------------------------------------------
# Write back the block of the output
offs_cn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)