[BugFix] AssertionError: Do not capture num_reqs > max_num_reqs for uniform batch (#25505)

Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Signed-off-by: yewentao256 <zhyanwentao@126.com>

Author: LucasWilkinson

vllm/v1/worker/gpu_model_runner.py

@@ -2828,7 +2828,7 @@ def _get_mm_dummy_batch(
     def _dummy_run(
         self,
         num_tokens: int,
-        cudagraph_runtime_mode: CUDAGraphMode = CUDAGraphMode.NONE,
+        cudagraph_runtime_mode: Optional[CUDAGraphMode] = None,
         force_attention: bool = False,
         uniform_decode: bool = False,
         allow_microbatching: bool = True,
@@ -2844,6 +2844,8 @@ def _dummy_run(
         Args:
             num_tokens: Number of tokens to run the dummy forward pass.
             cudagraph_runtime_mode: used to control the behavior.
+                - if not set will determine the cudagraph mode based on using 
+                    the self.cudagraph_dispatcher.
                 - CUDAGraphMode.NONE: No cudagraph, for warm up and profile run
                 - CUDAGraphMode.PIECEWISE: Piecewise cudagraph.
                 - CUDAGraphMode.FULL: Full cudagraph, attention metadata is
@@ -2857,7 +2859,7 @@ def _dummy_run(
                 (1 token) and prefill (multiple tokens) requests.
             remove_lora: If False, dummy LoRAs are not destroyed after the run
         """
-        assert cudagraph_runtime_mode in {
+        assert cudagraph_runtime_mode is None or cudagraph_runtime_mode in {
             CUDAGraphMode.NONE, CUDAGraphMode.PIECEWISE, CUDAGraphMode.FULL
         }
 
@@ -2899,10 +2901,6 @@ def _dummy_run(
         elif uniform_decode:
             assert not create_mixed_batch
             num_reqs = cdiv(num_tokens, max_query_len)
-            assert num_reqs <= max_num_reqs, \
-                f"Do not capture num_reqs {num_reqs} > max_num_reqs " \
-                f"{max_num_reqs} for uniform batch. Num tokens: " \
-                f"{num_tokens}, max_query_len: {max_query_len}"
             num_scheduled_tokens_list = [max_query_len] * num_reqs
             if num_tokens % max_query_len != 0:
                 num_scheduled_tokens_list[-1] = num_tokens % max_query_len
@@ -3043,18 +3041,20 @@ def _dummy_run(
 
                 intermediate_tensors = self.sync_and_slice_intermediate_tensors(
                     num_tokens, None, False)
-            if cudagraph_runtime_mode == CUDAGraphMode.NONE:
-                batch_descriptor = None
-            else:
-                # filter out the valid batch descriptor
-                _cg_mode, batch_descriptor = \
-                    self.cudagraph_dispatcher.dispatch(
-                        BatchDescriptor(num_tokens=num_tokens,
-                                        uniform_decode=uniform_decode))
-                # sanity check
-                assert cudagraph_runtime_mode == _cg_mode, (
+
+            # filter out the valid batch descriptor
+            _cg_mode, batch_descriptor = self.cudagraph_dispatcher.dispatch(
+                BatchDescriptor(num_tokens=num_tokens,
+                                uniform_decode=uniform_decode))
+            if cudagraph_runtime_mode is not None:
+                # we allow forcing NONE when the dispatcher disagrees to support
+                # warm ups for cudagraph capture
+                assert cudagraph_runtime_mode == CUDAGraphMode.NONE or \
+                    cudagraph_runtime_mode == _cg_mode, (
                     f"Cudagraph runtime mode mismatch at dummy_run. "
                     f"Expected {_cg_mode}, but got {cudagraph_runtime_mode}.")
+            else:
+                cudagraph_runtime_mode = _cg_mode
 
             if ubatch_slices is not None:
                 num_tokens = num_tokens // 2