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

fix tokenspeed_mla attn kernel jit (#26170)

Browse Source
This commit is contained in:
Qiaolin Yu
2026-05-23 03:24:33 -07:00
committed by GitHub
parent 89ff2bc111
commit cb7b57955d
1 changed files with 3 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/layers/attention/tokenspeed_mla_backend.py
+3 -2
View File
@@ -134,8 +134,9 @@ class TokenspeedMLABackend(TRTLLMMLABackend):
# branch, which always asks for the LSE. # branch, which always asks for the LSE.
if is_causal is False and return_lse is False: if is_causal is False and return_lse is False:
continue continue
# Runtime feeds fp8_e4m3fn q/k/v
config = ( config = (
torch.bfloat16, torch.float8_e4m3fn,
head_dim_qk, head_dim_qk,
self.v_head_dim, self.v_head_dim,
is_causal, is_causal,
@@ -146,7 +147,7 @@ class TokenspeedMLABackend(TRTLLMMLABackend):
if config in _compiled_kernels: if config in _compiled_kernels:
continue continue
_compiled_kernels[config] = _compile_prefill_kernel( _compiled_kernels[config] = _compile_prefill_kernel(
torch.bfloat16, torch.float8_e4m3fn,
head_dim_qk, head_dim_qk,
self.v_head_dim, self.v_head_dim,
is_causal, is_causal,