[AMD] Fix TP2 DeepSeek-R1 nhead=64 MLA decode crash and add nightly coverage (#27188)

Co-authored-by: clintg6 <7388379+clintg6@users.noreply.github.com>
2026-06-03 16:56:05 -07:00
parent c9ca56da8c
commit cfb7fb4fad
6 changed files with 518 additions and 1 deletions
@@ -281,7 +281,7 @@ class AiterAttnBackend(AttentionBackend):
            # so all num_head size does not use qh16 kernel to simulate
            # it should not use fake-nps (fast_mode = False, intra_batch_mode = True)
            # it will cause gpu-fault or accuracy issue
-            if self.num_head == 32 or self.num_head == 128:
+            if self.num_head in (32, 64, 128):
                fast_mode = True
                intra_batch_mode = False