refactor: reuse workspace for bmm_fp8

Author: ttyio

flashinfer/gemm.py

@@ -1104,7 +1104,7 @@ def execute_cudnn_gemm_fp4_graph(
         UIDs.O_UID.value: c_final,
     }
 
-    if graph.get_workspace_size() > DEFAULT_WORKSPACE_SIZE:
+    if workspace_buffer.numel() < graph.get_workspace_size():
         workspace_buffer = torch.empty(
             graph.get_workspace_size(), device=a.device, dtype=torch.uint8
         )
@@ -1179,7 +1179,7 @@ def build_cudnn_gemm_with_per_tensor_q_graph(
         return graph
 
 
-def execute_cudnn_gemm_with_per_tensor_q_graph(graph, a, b, alpha, c_final):
+def execute_cudnn_gemm_with_per_tensor_q_graph(graph, a, b, alpha, c_final, workspace):
     variant_pack = {
         UIDs.A_UID.value: a,
         UIDs.B_UID.value: b,
@@ -1190,9 +1190,10 @@ def execute_cudnn_gemm_with_per_tensor_q_graph(graph, a, b, alpha, c_final):
     stream = torch.cuda.current_stream(a.device)
     cudnn_handle = _get_cudnn_handle(stream)
 
-    workspace = torch.empty(
-        graph.get_workspace_size(), device=a.device, dtype=torch.uint8
-    )
+    if workspace.numel() < graph.get_workspace_size():
+        workspace = torch.empty(
+            graph.get_workspace_size(), device=a.device, dtype=torch.uint8
+        )
 
     graph.execute(variant_pack, workspace, handle=cudnn_handle)
 
@@ -1216,14 +1217,10 @@ def _cudnn_gemm_fp8(
     dq_scale: torch.Tensor,
     out: Optional[torch.Tensor],
     torch_out_dtype: torch.dtype,
+    workspace: torch.Tensor,
 ):
     _check_cudnn_availability()
 
-    if out is None:
-        out = torch.empty(
-            a.shape[0], a.shape[1], b.shape[2], dtype=torch_out_dtype, device=a.device
-        )
-
     graph = build_cudnn_gemm_with_per_tensor_q_graph(
         a.shape,
         a.stride(),
@@ -1235,7 +1232,7 @@ def _cudnn_gemm_fp8(
         a.device,
     )
 
-    execute_cudnn_gemm_with_per_tensor_q_graph(graph, a, b, dq_scale, out)
+    execute_cudnn_gemm_with_per_tensor_q_graph(graph, a, b, dq_scale, out, workspace)
     return out
 
 
@@ -1550,12 +1547,12 @@ def bmm_fp8(
             dtype=dtype,
         )
 
+    workspace_buffer = _get_cache_buf(
+        "bmm_fp8_workspace", DEFAULT_WORKSPACE_SIZE, A.device
+    )
     if backend == "cudnn":
-        return _cudnn_gemm_fp8(A, B, A_scale * B_scale, out, dtype)
+        return _cudnn_gemm_fp8(A, B, A_scale * B_scale, out, dtype, workspace_buffer)
     elif backend == "cublas":
-        workspace_buffer = _get_cache_buf(
-            "bmm_fp8_workspace", DEFAULT_WORKSPACE_SIZE, A.device
-        )
         get_gemm_module().bmm_fp8(workspace_buffer, A, B, out, A_scale, B_scale)
     return out