minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

fix: preserve divisible FP8 block K configs on CUDA (#27553)

Browse Source
This commit is contained in:
lmyybh
2026-06-17 10:30:56 +00:00
committed by GitHub
parent 9b8c41171a
commit 2f1390fcb1
1 changed files with 3 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/layers/quantization/fp8_kernel.py
+3 -1
View File
@@ -1166,7 +1166,9 @@ def get_w8a8_block_fp8_configs(
sanitized = {} sanitized = {}
clamped_ms = [] clamped_ms = []
for m_key, cfg in raw.items(): for m_key, cfg in raw.items():
if cfg["BLOCK_SIZE_K"] < block_k: if cfg["BLOCK_SIZE_K"] < block_k and (
not _is_cuda or block_k % cfg["BLOCK_SIZE_K"] != 0
):
clamped_ms.append((m_key, cfg["BLOCK_SIZE_K"])) clamped_ms.append((m_key, cfg["BLOCK_SIZE_K"]))
cfg = {**cfg, "BLOCK_SIZE_K": block_k} cfg = {**cfg, "BLOCK_SIZE_K": block_k}
sanitized[m_key] = cfg sanitized[m_key] = cfg