[Perf] Optimize clamp to clamp_ (#48143)

Signed-off-by: yewentao256 <zhyanwentao@126.com>
Signed-off-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
Wentao Ye
2026-07-16 18:41:07 -04:00
committed by GitHub
co-authored by mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent 2cab53ddee
commit f17be06fbe
7 changed files with 48 additions and 32 deletions
+2 -4
View File
@@ -950,10 +950,8 @@ def mamba_get_block_table_tensor(
assert isinstance(kv_cache_spec, MambaSpec)
# NOTE: For 0-length requests in CUDA graph, use a start_index of 0
# to handle the invalid block table.
start_indices = torch.clamp(
(seq_lens - 1) // kv_cache_spec.block_size,
min=0,
)
start_indices = (seq_lens - 1) // kv_cache_spec.block_size
start_indices.clamp_(min=0)
# Use int32 for arithmetic to avoid dtype promotion overhead,
# then convert to int64 for gather (which requires Long indices)
offsets = torch.arange(