[DSA] Hopper FP8 FlashMLA KV padding (#22372)

This commit is contained in:
Mohammad Miadh Angkad
2026-04-12 02:19:17 -07:00
committed by GitHub
parent 0e0091c6c8
commit bcc0c65aa8
3 changed files with 43 additions and 8 deletions
+2 -2
View File
@@ -1534,9 +1534,9 @@ class ServerArgs:
if not user_set_decode:
self.nsa_decode_backend = "trtllm"
else:
# flashmla_auto dispatches to flashmla_sparse/flashmla_kv based on hardware and heuristics
# Hopper FP8 defaults to flashmla_kv for both prefill and decode.
if not user_set_prefill:
self.nsa_prefill_backend = "flashmla_auto"
self.nsa_prefill_backend = "flashmla_kv"
if not user_set_decode:
self.nsa_decode_backend = "flashmla_kv"
else: