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

[AMD] Use tilelang as default NSA attention backend dispatch on AMD Instinct (#18319)

Browse Source
This commit is contained in:
fxmarty-amd
2026-02-27 01:43:34 -08:00
committed by GitHub
parent 9b2fbf7e6a
commit 9496bbd7b1
2 changed files with 7 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/layers/attention/nsa_backend.py
+3 -1
View File
@@ -1477,7 +1477,9 @@ class NativeSparseAttnBackend(
page_size=1, page_size=1,
) )
else: else:
raise ValueError(f"Unsupported {nsa_impl = }") raise ValueError(
f"Unsupported {nsa_impl = } for forward_extend. Consider using an other attention backend."
)
def forward_decode( def forward_decode(
self, self,
python/sglang/srt/server_args.py
+4 -1
View File
@@ -1150,7 +1150,10 @@ class ServerArgs:
user_set_prefill = self.nsa_prefill_backend is not None user_set_prefill = self.nsa_prefill_backend is not None
user_set_decode = self.nsa_decode_backend is not None user_set_decode = self.nsa_decode_backend is not None
if kv_cache_dtype == "fp8_e4m3": if not user_set_prefill and not user_set_decode and is_hip():
self.nsa_prefill_backend = "tilelang"
self.nsa_decode_backend = "tilelang"
elif kv_cache_dtype == "fp8_e4m3":
# flashmla_auto dispatches to flashmla_sparse/flashmla_kv based on hardware and heuristics # flashmla_auto dispatches to flashmla_sparse/flashmla_kv based on hardware and heuristics
if not user_set_prefill: if not user_set_prefill:
self.nsa_prefill_backend = "flashmla_auto" self.nsa_prefill_backend = "flashmla_auto"