mark triton_tutorial_flash_dp_persistent_blackwell as fwd only

henrylhtsang · web-flow · commit 05b62cac0ad9 · 2025-11-13T22:59:48.000Z
Differential Revision: D86992473 Pull Request resolved: #642
diff --git a/tritonbench/operators/blackwell_attentions/operator.py b/tritonbench/operators/blackwell_attentions/operator.py
@@ -494,7 +494,7 @@ def triton_tutorial_flash_dp_blackwell(
             q, k, v, self.causal, self.sm_scale, "ws"
         )
 
-    @register_benchmark(enabled=is_blackwell() and HAS_BLACKWELL_AUTOWS)
+    @register_benchmark(enabled=is_blackwell() and HAS_BLACKWELL_AUTOWS, fwd_only=True)
     def triton_tutorial_flash_dp_persistent_blackwell(
         self,
         q: torch.Tensor,

Original file line number	Diff line number	Diff line change
`@@ -494,7 +494,7 @@ def triton_tutorial_flash_dp_blackwell(`
`494`	`494`	`q, k, v, self.causal, self.sm_scale, "ws"`
`495`	`495`	`)`
`496`	`496`
`497`		`- @register_benchmark(enabled=is_blackwell() and HAS_BLACKWELL_AUTOWS)`
	`497`	`+ @register_benchmark(enabled=is_blackwell() and HAS_BLACKWELL_AUTOWS, fwd_only=True)`
`498`	`498`	`def triton_tutorial_flash_dp_persistent_blackwell(`
`499`	`499`	`self,`
`500`	`500`	`q: torch.Tensor,`