perf: improve plan performance by using non-blocking memcpy (#547)

cc @merrymercy

flashinfer-aot/csrc_aot/batch_decode.cu

@@ -43,7 +43,7 @@ std::vector<int64_t> BatchDecodeWithPagedKVCachePlan(
       int_workspace_buffer.size(0) * int_workspace_buffer.element_size();
   auto device = float_workspace_buffer.device();
   cudaStream_t torch_current_stream = c10::cuda::getCurrentCUDAStream(device.index());
-  indptr = indptr.to(torch::kCPU);
+  TORCH_CHECK(indptr.device() == torch::kCPU, "indptr must be on CPU");
 
   DecodePlanInfo plan_info;
 
@@ -150,8 +150,7 @@ std::vector<torch::Tensor> BatchDecodeWithPagedKVCacheRun(
         paged_kv_t<DTypeKV, IdType> paged_kv(
             num_kv_heads, page_size, HEAD_DIM, batch_size, kv_layout,
             static_cast<DTypeKV*>(paged_k_cache.data_ptr()),
-            static_cast<DTypeKV*>(paged_v_cache.data_ptr()),
-            kv_cache_strides,
+            static_cast<DTypeKV*>(paged_v_cache.data_ptr()), kv_cache_strides,
             static_cast<IdType*>(paged_kv_indices.data_ptr()),
             static_cast<IdType*>(paged_kv_indptr.data_ptr()),
             static_cast<IdType*>(paged_kv_last_page_len.data_ptr()));

flashinfer-aot/csrc_aot/batch_prefill.cu

@@ -51,8 +51,8 @@ std::vector<int64_t> BatchPrefillWithKVCachePlan(
 
   auto device = float_workspace_buffer.device();
   cudaStream_t torch_current_stream = c10::cuda::getCurrentCUDAStream(device.index());
-  qo_indptr = qo_indptr.to(torch::kCPU);
-  kv_indptr = kv_indptr.to(torch::kCPU);
+  TORCH_CHECK(qo_indptr.device() == torch::kCPU, "qo_indptr must be on CPU");
+  TORCH_CHECK(kv_indptr.device() == torch::kCPU, "kv_indptr must be on CPU");
 
   PrefillPlanInfo plan_info;
 

python/flashinfer/decode.py

@@ -53,7 +53,8 @@ def compile_single_decode_module(
 ):
     uri, path = gen_single_decode_cu(*args)
     return load_cuda_ops(
-        uri, [path],
+        uri,
+        [path],
         verbose=verbose,
     )
 
@@ -64,7 +65,8 @@ def compile_batch_decode_module(
 ):
     uri, path = gen_batch_decode_cu(*args)
     return load_cuda_ops(
-        uri, [path],
+        uri,
+        [path],
         verbose=verbose,
     )
 
@@ -114,6 +116,7 @@ def get_batch_decode_module(*args):
             _batch_decode_modules[args] = compile_batch_decode_module(*args)
     return _batch_decode_modules[args]
 
+
 def single_decode_with_kv_cache_with_jit_module(
     jit_module: Any,
     q: torch.Tensor,
@@ -123,8 +126,10 @@ def single_decode_with_kv_cache_with_jit_module(
     kv_layout: str = "NHD",
     window_left: int = -1,
 ):
-    tmp = _get_cache_buf("single_decode_with_kv_cache_tmp", 32 * 1024 * 1024, q.device)    
-    return jit_module.run(q, k, v, tmp, TensorLayout[kv_layout].value, window_left, *args)
+    tmp = _get_cache_buf("single_decode_with_kv_cache_tmp", 32 * 1024 * 1024, q.device)
+    return jit_module.run(
+        q, k, v, tmp, TensorLayout[kv_layout].value, window_left, *args
+    )
 
 
 def single_decode_with_kv_cache(
@@ -444,6 +449,7 @@ def __init__(
 
         if use_tensor_cores:
             if use_cuda_graph:
+                # NOTE(Zihao): if once created, no need to update it in plan/run
                 self._qo_indptr_buf = torch.arange(
                     self._fixed_batch_size + 1,
                     dtype=torch.int32,
@@ -555,8 +561,7 @@ def plan(
         if logits_soft_cap is None:
             logits_soft_cap = 0.0
 
-        qo_indptr = _get_range_buf(batch_size + 1, indptr.device)
-
+        qo_indptr_host = _get_range_buf(batch_size + 1, "cpu")
         if self.is_cuda_graph_enabled:
             if batch_size != self._fixed_batch_size:
                 raise ValueError(
@@ -569,21 +574,18 @@ def plan(
                 raise ValueError(
                     "The size of indices should be less than or equal to the allocated buffer"
                 )
-            self._paged_kv_indptr_buf.copy_(indptr)
-            self._paged_kv_indices_buf[: len(indices)] = indices
-            self._paged_kv_last_page_len_buf.copy_(last_page_len)
-            if self.use_tensor_cores:
-                self._qo_indptr_buf.copy_(qo_indptr)
+            self._paged_kv_indptr_buf.copy_(indptr, non_blocking=True)
+            self._paged_kv_indices_buf[: len(indices)].copy_(indices, non_blocking=True)
+            self._paged_kv_last_page_len_buf.copy_(last_page_len, non_blocking=True)
         else:
-            self._paged_kv_indptr_buf = indptr.to(self.device)
-            self._paged_kv_indices_buf = indices.to(self.device)
-            self._paged_kv_last_page_len_buf = last_page_len.to(self.device)
-            if self.use_tensor_cores:
-                self._qo_indptr_buf = qo_indptr.to(self.device)
-
-        qo_indptr = qo_indptr.to("cpu", non_blocking=True)
-        indptr = indptr.to("cpu", non_blocking=True)
+            self._paged_kv_indptr_buf = indptr.to(self.device, non_blocking=True)
+            self._paged_kv_indices_buf = indices.to(self.device, non_blocking=True)
+            self._paged_kv_last_page_len_buf = last_page_len.to(
+                self.device, non_blocking=True
+            )
+            self._qo_indptr_buf = qo_indptr_host.to(self.device, non_blocking=True)
 
+        indptr_host = indptr.to("cpu", non_blocking=True)
         if data_type is not None:
             q_data_type = data_type
             kv_data_type = data_type
@@ -612,8 +614,8 @@ def plan(
                 self._float_workspace_buffer,
                 self._int_workspace_buffer,
                 self._pin_memory_int_workspace_buffer,
-                qo_indptr,
-                indptr,
+                qo_indptr_host,
+                indptr_host,
                 batch_size,
                 num_qo_heads,
                 num_kv_heads,
@@ -635,7 +637,7 @@ def plan(
                 self._float_workspace_buffer,
                 self._int_workspace_buffer,
                 self._pin_memory_int_workspace_buffer,
-                indptr,
+                indptr_host,
                 batch_size,
                 num_qo_heads,
                 num_kv_heads,

python/flashinfer/jit/batch_decode_templ.py

@@ -42,7 +42,7 @@
       int_workspace_buffer.size(0) * int_workspace_buffer.element_size();
   auto device = float_workspace_buffer.device();
   cudaStream_t torch_current_stream = c10::cuda::getCurrentCUDAStream(device.index());
-  indptr = indptr.to(torch::kCPU);
+  TORCH_CHECK(indptr.device() == torch::kCPU, "indptr must be on CPU");
 
   DecodePlanInfo plan_info;
 

python/flashinfer/jit/batch_prefill_templ.py

@@ -49,8 +49,8 @@
   
   auto device = float_workspace_buffer.device();
   cudaStream_t torch_current_stream = c10::cuda::getCurrentCUDAStream(device.index());
-  qo_indptr = qo_indptr.to(torch::kCPU);
-  kv_indptr = kv_indptr.to(torch::kCPU);
+  TORCH_CHECK(qo_indptr.device() == torch::kCPU, "qo_indptr must be on CPU");
+  TORCH_CHECK(kv_indptr.device() == torch::kCPU, "kv_indptr must be on CPU");
 
   PrefillPlanInfo plan_info;
 

python/flashinfer/prefill.py

@@ -57,7 +57,8 @@ def compile_single_prefill_module(
 ):
     uri, path = gen_single_prefill_cu(*args)
     return load_cuda_ops(
-        uri, [path],
+        uri,
+        [path],
         verbose=verbose,
     )
 
@@ -68,7 +69,8 @@ def compile_batch_prefill_module(
 ):
     uri, path = gen_batch_prefill_cu(*args)
     return load_cuda_ops(
-        uri, [path],
+        uri,
+        [path],
         verbose=verbose,
     )
 
@@ -125,6 +127,7 @@ def get_batch_prefill_module(*args):
             _batch_prefill_modules[args] = compile_batch_prefill_module(*args)
     return _batch_prefill_modules[args]
 
+
 def single_prefill_with_kv_cache_with_jit_module(
     jit_module: Any,
     q: torch.Tensor,
@@ -137,7 +140,8 @@ def single_prefill_with_kv_cache_with_jit_module(
 ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
     tmp = _get_cache_buf("single_prefill_with_kv_cache_tmp", 32 * 1024 * 1024, q.device)
     out = jit_module.run(
-        q, k, v, tmp, TensorLayout[kv_layout].value, window_left, return_lse, *args)
+        q, k, v, tmp, TensorLayout[kv_layout].value, window_left, return_lse, *args
+    )
     return out if return_lse else out[0]
 
 
@@ -726,10 +730,14 @@ def plan(
                     "The length of paged_kv_indices exceeds the allocated buffer size."
                 )
 
-            self._qo_indptr_buf.copy_(qo_indptr)
-            self._paged_kv_indptr_buf.copy_(paged_kv_indptr)
-            self._paged_kv_indices_buf[: len(paged_kv_indices)] = paged_kv_indices
-            self._paged_kv_last_page_len_buf.copy_(paged_kv_last_page_len)
+            self._qo_indptr_buf.copy_(qo_indptr, non_blocking=True)
+            self._paged_kv_indptr_buf.copy_(paged_kv_indptr, non_blocking=True)
+            self._paged_kv_indices_buf[: len(paged_kv_indices)].copy_(
+                paged_kv_indices, non_blocking=True
+            )
+            self._paged_kv_last_page_len_buf.copy_(
+                paged_kv_last_page_len, non_blocking=True
+            )
 
             if packed_custom_mask is not None:
                 if not torch.is_tensor(self._custom_mask_buf):
@@ -740,20 +748,31 @@ def plan(
                     raise ValueError(
                         "qk_indptr_buf must be initialized with a torch.Tensor in cuda graph mode if we use custom mask in attention computation."
                     )
-                self._custom_mask_buf[: len(packed_custom_mask)] = packed_custom_mask
+                self._custom_mask_buf[: len(packed_custom_mask)].copy_(
+                    packed_custom_mask, non_blocking=True
+                )
                 # NOTE(Zihao): qk_indptr has the same length as qo_indptr
-                self._qk_indptr_buf.copy_(qk_indptr)
+                self._qk_indptr_buf.copy_(qk_indptr, non_blocking=True)
         else:
-            self._qo_indptr_buf = qo_indptr.to(self.device)
-            self._paged_kv_indptr_buf = paged_kv_indptr.to(self.device)
-            self._paged_kv_indices_buf = paged_kv_indices.to(self.device)
-            self._paged_kv_last_page_len_buf = paged_kv_last_page_len.to(self.device)
+            self._qo_indptr_buf = qo_indptr.to(self.device, non_blocking=True)
+            self._paged_kv_indptr_buf = paged_kv_indptr.to(
+                self.device, non_blocking=True
+            )
+            self._paged_kv_indices_buf = paged_kv_indices.to(
+                self.device, non_blocking=True
+            )
+            self._paged_kv_last_page_len_buf = paged_kv_last_page_len.to(
+                self.device, non_blocking=True
+            )
             if packed_custom_mask is not None:
-                self._custom_mask_buf = packed_custom_mask.to(self.device)
-                self._qk_indptr_buf = qk_indptr.to(self.device)
+                self._custom_mask_buf = packed_custom_mask.to(
+                    self.device, non_blocking=True
+                )
+                self._qk_indptr_buf = qk_indptr.to(self.device, non_blocking=True)
 
-        qo_indptr = qo_indptr.to("cpu", non_blocking=True)
-        paged_kv_indptr = paged_kv_indptr.to("cpu", non_blocking=True)
+        # NOTE(Zihao): only required if qo_indptr/paged_kv_indptr are device tensors
+        qo_indptr_host = qo_indptr.to("cpu", non_blocking=True)
+        paged_kv_indptr_host = paged_kv_indptr.to("cpu", non_blocking=True)
 
         if packed_custom_mask is not None:
             mask_mode = MaskMode.CUSTOM.value
@@ -781,8 +800,8 @@ def plan(
             self._float_workspace_buffer,
             self._int_workspace_buffer,
             self._pin_memory_int_workspace_buffer,
-            qo_indptr,
-            paged_kv_indptr,
+            qo_indptr_host,
+            paged_kv_indptr_host,
             batch_size,
             num_qo_heads,
             num_kv_heads,

python/flashinfer/sparse.py

@@ -257,7 +257,7 @@ def plan(
         num_blocks_row = len(indptr) - 1
         qo_indptr_host = R * torch.arange(num_blocks_row + 1, dtype=torch.int32)
         qo_indptr_host[-1] = M
-        qo_indptr = qo_indptr_host.to(indptr.device)
+        qo_indptr = qo_indptr_host.to(indptr.device, non_blocking=True)
         if indices.max().item() * C > N:
             raise ValueError("indices out of bound")
         last_block_len = torch.full(
@@ -279,13 +279,13 @@ def plan(
                 mask.contiguous().view(-1), qk_indptr, bitorder="little"
             )
 
-        self._qo_indptr = qo_indptr.to(self.device)
-        self._paged_kv_indptr_buf = indptr.to(self.device)
-        self._paged_kv_indices_buf = indices.to(self.device)
-        self._paged_kv_last_page_len = last_block_len.to(self.device)
+        self._qo_indptr = qo_indptr.to(self.device, non_blocking=True)
+        self._paged_kv_indptr_buf = indptr.to(self.device, non_blocking=True)
+        self._paged_kv_indices_buf = indices.to(self.device, non_blocking=True)
+        self._paged_kv_last_page_len = last_block_len.to(self.device, non_blocking=True)
         if packed_mask is not None:
-            self._packed_mask_buf = packed_mask.to(self.device)
-            self._qk_indptr_buf = qk_indptr.to(self.device)
+            self._packed_mask_buf = packed_mask.to(self.device, non_blocking=True)
+            self._qk_indptr_buf = qk_indptr.to(self.device, non_blocking=True)
             mask_mode = MaskMode.CUSTOM.value
         else:
             self._packed_mask_buf = None