[Feat]support sequence parallelism by pass for VL models (#5632)

tests/e2e/multicard/2-cards/test_sp_pass.py

@@ -0,0 +1,64 @@
+import os
+
+import pytest
+from vllm import SamplingParams
+
+from tests.e2e.conftest import VllmRunner
+from tests.e2e.model_utils import check_outputs_equal
+
+MODELS = [
+    "Qwen/Qwen3-VL-2B-Instruct",
+]
+
+
+@pytest.mark.parametrize("model", MODELS)
+def test_qwen3_vl_sp_tp2(model: str) -> None:
+    prompts = [
+        "Hello, my name is", "The capital of the United States is",
+        "The capital of France is", "The future of AI is"
+    ]
+    sampling_params = SamplingParams(max_tokens=10, temperature=0.0)
+
+    with VllmRunner(
+            model,
+            max_model_len=1024,
+            tensor_parallel_size=2,
+            compilation_config={
+                "cudagraph_capture_sizes": [2, 4],
+                "cudagraph_mode": "FULL_DECODE_ONLY",
+                "pass_config": {"enable_sp": False}
+            },
+            additional_config={"npugraph_ex_config": {"enable": False}}
+    ) as runner:
+        no_sp_outputs = runner.model.generate(prompts, sampling_params)
+
+    with VllmRunner(
+            model,
+            max_model_len=1024,
+            tensor_parallel_size=2,
+            compilation_config={
+                "cudagraph_capture_sizes": [2, 4],
+                "cudagraph_mode": "FULL_DECODE_ONLY",
+                "pass_config": {"enable_sp": True}
+            },
+            additional_config={"sp_threshold": 10, "npugraph_ex_config": {"enable": False}}
+    ) as runner:
+        sp_outputs = runner.model.generate(
+            prompts, sampling_params)
+
+    no_sp_outputs_list = []
+    for output in no_sp_outputs:
+        no_sp_outputs_list.append(
+            (output.outputs[0].index, output.outputs[0].text))
+
+    sp_outputs_list = []
+    for output in sp_outputs:
+        sp_outputs_list.append(
+            (output.outputs[0].index, output.outputs[0].text))
+
+    check_outputs_equal(
+        outputs_0_lst=no_sp_outputs_list,
+        outputs_1_lst=sp_outputs_list,
+        name_0="no_sp_outputs",
+        name_1="sp_outputs",
+    )

tests/ut/ops/test_mla.py

@@ -157,7 +157,7 @@ class TestAscendMultiHeadLatentAttention(TestBase):
         hidden_states = torch.randn(3, self.hidden_size)
 
         mock_forward_context = MagicMock(spec=ForwardContext)
-        mock_forward_context.sp_enabled = False
+        mock_forward_context.flash_comm_v1_enabled = False
         mock_get_forward_context.return_value = mock_forward_context
 
         mock_mla_forward.return_value = (3, self.hidden_size)

tests/ut/spec_decode/test_eagle_proposer.py

@@ -390,7 +390,7 @@ class TestEagleProposerDummyRun(TestBase):
 
     # cpu does not support parallel-group, let alone `sp`
     @patch("vllm_ascend.spec_decode.eagle_proposer.get_forward_context",
-           **{"return_value.sp_enabled": False})
+           **{"return_value.flash_comm_v1_enabled": False})
     @patch("vllm_ascend.spec_decode.eagle_proposer.set_ascend_forward_context")
     def test_dummy_run_basic(self, mock_context, mock_get_context):
         num_tokens = 32
@@ -406,7 +406,7 @@ class TestEagleProposerDummyRun(TestBase):
 
     # cpu does not support parallel-group, let alone `sp`
     @patch("vllm_ascend.spec_decode.eagle_proposer.get_forward_context",
-           **{"return_value.sp_enabled": False})
+           **{"return_value.flash_comm_v1_enabled": False})
     @patch("vllm_ascend.spec_decode.eagle_proposer.set_ascend_forward_context")
     def test_dummy_run_with_prefill(self, mock_context, mock_get_context):
         mock_context.return_value.__enter__.return_value = None
@@ -426,7 +426,7 @@ class TestEagleProposerDummyRun(TestBase):
         mock_return_context.cudagraph_runtime_mode = CUDAGraphMode.FULL
         mock_return_context.capturing = True
         # cpu does not support parallel-group, let alone `sp`
-        mock_return_context.sp_enabled = False
+        mock_return_context.flash_comm_v1_enabled = False
         mock_get_context.return_value = mock_return_context
         self.proposer.use_cuda_graph = True
         # cpu does not support `torch.ops.vllm.maybe_pad_and_reduce`
@@ -449,7 +449,7 @@ class TestEagleProposerDummyRun(TestBase):
         mock_return_context.cudagraph_runtime_mode = CUDAGraphMode.FULL
         mock_return_context.capturing = False
         # cpu does not support parallel-group, let alone `sp`
-        mock_return_context.sp_enabled = False
+        mock_return_context.flash_comm_v1_enabled = False
         mock_get_context.return_value = mock_return_context
         self.proposer.use_cuda_graph = True
         # cpu does not support `torch.ops.vllm.maybe_pad_and_reduce`