[bugfix] Fix SP + PP without specifying compile size (vllm-project#26955)

Signed-off-by: angelayi <yiangela7@gmail.com>
Signed-off-by: Alberto Perdomo <aperdomo@redhat.com>

Author: angelayi

tests/distributed/test_sequence_parallel.py

@@ -18,6 +18,7 @@
 from vllm.config.compilation import CompilationMode
 from vllm.config.model import RunnerOption
 from vllm.logger import init_logger
+from vllm.utils import is_torch_equal_or_newer
 
 from ..models.registry import HF_EXAMPLE_MODELS
 from ..utils import compare_two_settings, create_new_process_for_each_test
@@ -159,6 +160,7 @@ def _compare_sp(
     runner: RunnerOption,
     test_options: SPTestOptions,
     num_gpus_available: int,
+    use_inductor_graph_partition: bool,
     *,
     method: Literal["generate", "encode"],
     is_multimodal: bool,
@@ -243,6 +245,7 @@ def _compare_sp(
             "enable_fusion": enable_fusion,
             "enable_noop": True,
         },
+        "use_inductor_graph_partition": use_inductor_graph_partition,
     }
 
     tp_sp_args = [
@@ -297,6 +300,7 @@ def _compare_sp(
         if model_id in SP_TEST_MODELS
     ],
 )
+@pytest.mark.parametrize("use_inductor_graph_partition", [True, False])
 @create_new_process_for_each_test()
 def test_tp_sp_generation(
     model_id: str,
@@ -305,14 +309,19 @@ def test_tp_sp_generation(
     runner: RunnerOption,
     test_options: SPTestOptions,
     num_gpus_available,
+    use_inductor_graph_partition: bool,
 ):
+    if use_inductor_graph_partition and not is_torch_equal_or_newer("2.9.0.dev"):
+        pytest.skip("inductor graph partition is only available in PyTorch 2.9+")
+
     _compare_sp(
         model_id,
         parallel_setup,
         distributed_backend,
         runner,
         test_options,
         num_gpus_available,
+        use_inductor_graph_partition,
         method="generate",
         is_multimodal=False,
     )

vllm/v1/worker/utils.py

@@ -328,8 +328,12 @@ def is_residual_scattered_for_sp(
     """Check if the residual tensor is scattered for sequence parallelism.
 
     The residual tensor is scattered across tensor parallel ranks when sequence
-    parallelism and tensor parallelism is enabled, and the number of
-    input tokens is one of the compilation sizes.
+    parallelism and tensor parallelism is enabled.
+
+    This follows the same logic as SequenceParallelismPass.is_applicable():
+    - In full-graph compilation mode (no splitting ops or using inductor graph
+      partition), SP is always applied
+    - Otherwise, SP is only applied for specific shapes in compile_sizes
     """
     if not vllm_config.compilation_config.pass_config.enable_sequence_parallelism:
         return False
@@ -343,5 +347,10 @@ def is_residual_scattered_for_sp(
     # to be a multiple of tensor_parallel_size (tp) earlier.
     assert num_input_tokens % tp == 0
 
-    # Currently, SP is only enabled for static size fx graphs.
+    if (
+        not vllm_config.compilation_config.splitting_ops
+        or vllm_config.compilation_config.use_inductor_graph_partition
+    ):
+        return True
+
     return num_input_tokens in vllm_config.compilation_config.compile_sizes