[Kernel] Fix Flashinfer Correctness (#7284)

vllm/attention/backends/flashinfer.py

@@ -127,6 +127,7 @@ class FlashInferMetadata(AttentionMetadata):
             raise ValueError(
                 f"Only {supported_head_sizes} are supported for head_dim,",
                 f"received {self.head_dim}.")
+        self.is_profile_run = is_block_tables_empty(self.block_tables)
 
     def begin_forward(self):
         if self.num_prefill_tokens > 0:
@@ -140,11 +141,14 @@ class FlashInferMetadata(AttentionMetadata):
             assert self.paged_kv_last_page_len is not None
             batch_size = self.query_start_loc.shape[0] - 1
             assert batch_size >= 0
-            # The prefill stage does not read kv cache.
+            # The profile run does not read kv cache.
             # Both paged_kv_indices and paged_kv_last_page_len are empty.
             # paged_kv_indptr is a zero tensor with size batch_size + 1.
-            self.paged_kv_indptr = torch.zeros(batch_size + 1,
-                                               device=self.device)
+            if self.is_profile_run:
+                self.paged_kv_indptr = torch.zeros(batch_size + 1,
+                                                   device=self.device)
+            else:
+                self.paged_kv_indptr = self.paged_kv_indptr.to(self.device)
             self.paged_kv_last_page_len = self.paged_kv_last_page_len.to(
                 self.device)
             self.paged_kv_indices = self.paged_kv_indices.to(self.device)