[TRTLLM-6650][fix] Enhance CUDA graph + Beam search to correctly handle padding (#6665)

Signed-off-by: Stefan Niebler <[email protected]>

Author: stnie

tensorrt_llm/_torch/pyexecutor/cuda_graph_runner.py

@@ -34,6 +34,7 @@ def __init__(
         attn_metadata: AttentionMetadata,
         spec_metadata: Optional[SpecMetadata] = None,
         use_mrope: bool = False,
+        max_beam_width: int = 1,
     ) -> None:
         """
         Stores a CUDA graph and its associated input buffers.
@@ -49,19 +50,21 @@ def __init__(
         e.g. FlashInfer cause graph breaks).
         """
         self.batch_size = batch_size
-
+        self.max_beam_width = max_beam_width
         # [CUDA graph spec decode padding]
         # We pad input IDs/position IDs to the maximum draft length (token per request).
         # We're forced to do this because we cannot reallocate inputs over many graph runs.
         token_per_request = spec_metadata.max_draft_len + 1 if spec_metadata is not None else 1
 
         # Using ones instead of zeros prevents NaNs in e.g. Deepseek
-        self.input_ids = torch.ones((batch_size * token_per_request, ),
-                                    device=device,
-                                    dtype=torch.int32)
-        self.position_ids = torch.zeros((1, batch_size * token_per_request),
-                                        device=device,
-                                        dtype=torch.int32)
+        self.input_ids = torch.ones(
+            (batch_size * max_beam_width * token_per_request, ),
+            device=device,
+            dtype=torch.int32)
+        self.position_ids = torch.zeros(
+            (1, batch_size * max_beam_width * token_per_request),
+            device=device,
+            dtype=torch.int32)
         self.mrope_position_deltas = torch.zeros(
             (batch_size,
              1), device=device, dtype=torch.int32) if use_mrope else None

tensorrt_llm/_torch/pyexecutor/model_engine.py

@@ -846,8 +846,8 @@ def _get_padded_batch(
             spec_resource_manager: Optional[BaseResourceManager] = None) -> int:
         can_run_cuda_graph = scheduled_requests.can_run_cuda_graph
         batch_size = scheduled_requests.batch_size
-        # The number of sequences in the batch is the number of prompts times the beam width.
-        new_batch_size = batch_size * self.max_beam_width
+        new_batch_size = batch_size
+
         if self._run_cuda_graphs and self.enable_attention_dp and self.mapping.tp_size > 1:
             graph_batch_size = self.dist.tp_allgather(
                 [can_run_cuda_graph, batch_size])
@@ -981,8 +981,8 @@ def _maybe_get_cuda_graph(
             self._cuda_graphs[batch_size] = {}
 
         self._cuda_graphs[batch_size][draft_len] = DecodingCUDAGraphRunner(
-            num_sequences_in_batch, "cuda", attn_metadata, spec_metadata,
-            self.use_mrope)
+            batch_size, "cuda", attn_metadata, spec_metadata, self.use_mrope,
+            self.max_beam_width)
         return self._cuda_graphs[batch_size][draft_len]
 
     def __del__(self) -> None:
@@ -1376,8 +1376,11 @@ def _prepare_tp_inputs(
                 gather_ids.append(len(position_ids) - 1)
 
             request_ids.append(request.py_request_id)
-            gen_request_seq_slots.append(request.py_seq_slot)
             request.py_batch_idx = request.py_seq_slot
+            # Do not add a gen_request_seq_slot for CUDA graph dummy requests
+            # to prevent access errors due to None values
+            if not request.is_cuda_graph_dummy:
+                gen_request_seq_slots.append(request.py_seq_slot)
 
         previous_batch_len = len(previous_batch_indices)
 
@@ -1506,7 +1509,7 @@ def previous_seq_slots_device():
                 pin_memory=True,
             )
 
-        num_generation_requests = len(scheduled_requests.generation_requests)
+        num_generation_requests = len(gen_request_seq_slots)
         # Cache indirection is only used for beam search on generation requests
         if self.use_beam_search and num_generation_requests > 0:
             # CUDA Graph needs to set beam width during warmup (where the graph is captured), to ensure that cache indirection buffer is correctly picked up by the CUDA graph

tests/unittest/_torch/test_beam_search.py

@@ -61,7 +61,8 @@ def llm_cuda_graph(fixed_params, input_prompts):
         max_seq_len=32,
         max_beam_width=fixed_params["max_beam_width"],
         disable_overlap_scheduler=False,
-        cuda_graph_config=CudaGraphConfig(),
+        cuda_graph_config=CudaGraphConfig(batch_sizes=[1, 2, 4, 8],
+                                          enable_padding=True),
     )
 
 
@@ -126,7 +127,7 @@ def test_beam_search_output_shapes(gather_context_logits: bool,
 @pytest.mark.parametrize("gather_generation_logits", [True, False])
 @pytest.mark.parametrize("gather_context_logits", [True, False])
 @pytest.mark.parametrize("num_output_beams", [1, 2])
-@pytest.mark.parametrize("num_prompts", [1, 2])
+@pytest.mark.parametrize("num_prompts", [1, 2, 3])
 @pytest.mark.threadleak(enabled=False)
 def test_beam_search_output_shapes_cuda_graph_and_overlap(
         gather_context_logits: bool, gather_generation_logits: bool,
@@ -145,6 +146,10 @@ def test_beam_search_output_shapes_cuda_graph_and_overlap(
         return_generation_logits=gather_generation_logits,
         logprobs=return_log_probs,
     )
+    # test padding of cuda graph with 3 prompts
+    # replicate the prompts to have more than 2 prompts available
+    if (num_prompts == 3 and len(input_prompts) == 2):
+        input_prompts = [input_prompts[0]] * 3
     outputs = llm_cuda_graph.generate(input_prompts[:num_prompts],
                                       sampling_params=sampling_params)
     assert len(outputs) == num_prompts