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

Revert "Use Cache Hinting for fused_moe kernel (#15511)" (#15645)

Browse Source 
Signed-off-by: Wes Medford <wryanmedford@gmail.com>
This commit is contained in:
Wes
2025-03-27 20:45:55 -06:00
committed by GitHub
parent 8a49eea74b
commit 4ae17bf1e2
1 changed files with 4 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
vllm/model_executor/layers/fused_moe/fused_moe.py
View File

@ -189,11 +189,7 @@ def fused_moe_kernel_gptq_awq(
mask=token_mask[:, None] &
(offs_k[None, :] < K - k * BLOCK_SIZE_K),
other=0.0)
b = tl.load(
b_ptrs,
cache_modifier=".cg",
eviction_policy="evict_last",
)
b = tl.load(b_ptrs)
if use_int4_w4a16:
b = (b >> b_shifter) & 0xF
@ -395,13 +391,9 @@ def fused_moe_kernel(
mask=token_mask[:, None] &
(offs_k[None, :] < K - k * BLOCK_SIZE_K),
other=0.0)
b = tl.load(
b_ptrs,
mask=offs_k[:, None] < K - k * BLOCK_SIZE_K,
other=0.0,
cache_modifier=".cg",
eviction_policy="evict_last",
)
b = tl.load(b_ptrs,
mask=offs_k[:, None] < K - k * BLOCK_SIZE_K,
other=0.0)
# We accumulate along the K dimension.
if use_int8_w8a16:
accumulator = tl.dot(a, b.to(compute_type), acc=accumulator)