bugfix: do not use non-blocking copy for gpu to cpu transfer (#564)

If we use async GPU to CPU copy in `plan` functions, we synchronize
before we use the cpu array.
Since we have removed synchronization in plan functions, the GPU to CPU
copy should be synchronized.

For flashinfer v0.2, it's encouraged to pass cpu indptr arrays to `plan`
functions, and the synchronized GPU to CPU copy will be a no-op in this
case.

python/flashinfer/decode.py

@@ -719,7 +719,7 @@ def plan(
             )
             self._qo_indptr_buf = qo_indptr_host.to(self.device, non_blocking=True)
 
-        indptr_host = indptr.to("cpu", non_blocking=True)
+        indptr_host = indptr.to("cpu")
         if data_type is not None:
             q_data_type = data_type
             kv_data_type = data_type

python/flashinfer/prefill.py

@@ -1004,8 +1004,8 @@ def plan(
                 self._qk_indptr_buf = qk_indptr.to(self.device, 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)
+        qo_indptr_host = qo_indptr.to("cpu")
+        paged_kv_indptr_host = paged_kv_indptr.to("cpu")
 
         if packed_custom_mask is not None:
             mask_mode = MaskMode.CUSTOM.value
@@ -1571,8 +1571,8 @@ def plan(
                 self._qk_indptr_buf = qk_indptr.to(self.device)
 
         # NOTE(Zihao): only required if qo_indptr/paged_kv_indptr are device tensors
-        qo_indptr_host = qo_indptr.to("cpu", non_blocking=True)
-        kv_indptr_host = kv_indptr.to("cpu", non_blocking=True)
+        qo_indptr_host = qo_indptr.to("cpu")
+        kv_indptr_host = kv_indptr.to("cpu")
 
         if packed_custom_mask is not None:
             mask_mode = MaskMode.CUSTOM.value

python/flashinfer/sparse.py

@@ -298,7 +298,7 @@ def plan(
         self.R = R
         self.C = C
 
-        kv_indptr_host = indptr.to("cpu", non_blocking=True)
+        kv_indptr_host = indptr.to("cpu")
 
         # NOTE(Zihao): we haven't supported mask in cuda-core implementations but it should
         # be easy to add support for it if needed, leave it as a future work.