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

output type conversion fix (#27159)

Browse Source
This commit is contained in:
Jianyu Huang
2025-10-19 01:10:07 -07:00
committed by GitHub
parent b3aba04e5a
commit 221bf72577
1 changed files with 1 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
vllm/model_executor/layers/batch_invariant.py
View File

@@ -134,10 +134,7 @@ def matmul_kernel_persistent(
bias_ptrs = bias_ptr + offs_cn bias_ptrs = bias_ptr + offs_cn
bias = tl.load(bias_ptrs, mask=offs_cn < N, other=0.0).to(tl.float32) bias = tl.load(bias_ptrs, mask=offs_cn < N, other=0.0).to(tl.float32)
accumulator += bias accumulator += bias
if c_ptr.dtype.element_ty == tl.float8e4nv: c = accumulator.to(c_ptr.dtype.element_ty)
c = accumulator.to(tl.float8e4nv)
else:
c = accumulator.to(tl.float16)
tl.store(c_ptrs, c, mask=c_mask) tl.store(c_ptrs, c, mask=c_mask)