Merge pull request #1366 from InfiniTensor/feat/hygon-fa2-graph

wooway777 · web-flow · commit 53cc5ad69b16 · 2026-07-04T09:59:23.000+08:00
fix: support Hygon flash attention graph capture
diff --git a/src/infinicore/ops/mha_kvcache/hygon/mha_kvcache_flashattn_hygon.cc b/src/infinicore/ops/mha_kvcache/hygon/mha_kvcache_flashattn_hygon.cc
@@ -143,7 +143,6 @@ void run(void *planned_meta) {
         false,
         0);
 
-    c10::hip::device_synchronize();
     if (!result.empty() && result[0].defined()) {
         out_tensor.copy_(result[0]);
     }
diff --git a/src/infinicore/ops/multi_head_attention_varlen/hygon/mha_varlen_flashattn_hygon.cc b/src/infinicore/ops/multi_head_attention_varlen/hygon/mha_varlen_flashattn_hygon.cc
@@ -240,7 +240,6 @@ void run(void *planned_meta) {
         0.0,
         false,
         std::nullopt);
-    c10::hip::device_synchronize();
     if (!result.empty() && result[0].defined()) {
         out_work.copy_(result[0]);
     }

Original file line number	Diff line number	Diff line change
`@@ -143,7 +143,6 @@ void run(void *planned_meta) {`
`143`	`143`	`false,`
`144`	`144`	`0);`
`145`	`145`
`146`		`- c10::hip::device_synchronize();`
`147`	`146`	`if (!result.empty() && result[0].defined()) {`
`148`	`147`	`out_tensor.copy_(result[0]);`
`149`	`148`	`}`
Original file line number	Diff line number	Diff line change
`@@ -240,7 +240,6 @@ void run(void *planned_meta) {`
`240`	`240`	`0.0,`
`241`	`241`	`false,`
`242`	`242`	`std::nullopt);`
`243`		`- c10::hip::device_synchronize();`
`244`	`243`	`if (!result.empty() && result[0].defined()) {`
`245`	`244`	`out_work.copy_(result[0]);`
`246`	`245`	`}`