minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 13 Packages Projects Releases Wiki Activity

Fix scale_step_k computation in the fp8_kernel (#20819)

Browse Source 
Co-authored-by: Xiaoyu Zhang <35585791+BBuf@users.noreply.github.com>
This commit is contained in:
Muqi Li
2026-03-20 18:09:31 +08:00
committed by GitHub
co-authored by Xiaoyu Zhang
parent ec01ef9092
commit 2099943a49
1 changed files with 2 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/layers/quantization/fp8_kernel.py
+2 -1
View File
@@ -759,7 +759,7 @@ def _w8a8_block_fp8_matmul(
As_ptrs = As + offs_am * stride_As_m As_ptrs = As + offs_am * stride_As_m
offs_bsn = offs_bn // group_n offs_bsn = offs_bn // group_n
Bs_ptrs = Bs + offs_bsn * stride_Bs_n Bs_ptrs = Bs + offs_bsn * stride_Bs_n
scale_step_k = BLOCK_SIZE_K // group_k n_tiles_k_per_group_k = group_k // BLOCK_SIZE_K
accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32) accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32)
for k in range(0, tl.cdiv(K, BLOCK_SIZE_K)): for k in range(0, tl.cdiv(K, BLOCK_SIZE_K)):
@@ -773,6 +773,7 @@ def _w8a8_block_fp8_matmul(
a_s = tl.load(As_ptrs) a_s = tl.load(As_ptrs)
b_s = tl.load(Bs_ptrs) b_s = tl.load(Bs_ptrs)
scale_step_k = tl.where((k + 1) % n_tiles_k_per_group_k == 0, 1, 0)
accumulator += tl.dot(a, b) * a_s[:, None] * b_s[None, :] accumulator += tl.dot(a, b) * a_s[:, None] * b_s[None, :]
a_ptrs += BLOCK_SIZE_K * stride_ak a_ptrs += BLOCK_SIZE_K * stride_ak
b_ptrs += BLOCK_SIZE_K * stride_bk b_ptrs += BLOCK_SIZE_K * stride_bk