[ci][distributed] merge distributed test commands (#7097)

Co-authored-by: Cyrus Leung <cyrus.tl.leung@gmail.com>

.buildkite/test-pipeline.yaml

@@ -82,20 +82,9 @@ steps:
   num_gpus: 2
   commands:
   - VLLM_TEST_SAME_HOST=1 torchrun --nproc-per-node=4 distributed/test_same_node.py
-  - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_basic_distributed_correctness.py
+  - TARGET_TEST_SUITE=L4 pytest -v -s distributed/test_basic_distributed_correctness.py
-  - TEST_DIST_MODEL=meta-llama/Llama-2-7b-hf DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_basic_distributed_correctness.py
+  - pytest -v -s distributed/test_chunked_prefill_distributed.py
-  - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=ray VLLM_USE_RAY_SPMD_WORKER=1 VLLM_USE_RAY_COMPILED_DAG=1 pytest -v -s distributed/test_basic_distributed_correctness.py
+  - pytest -v -s distributed/test_multimodal_broadcast.py
-  - TEST_DIST_MODEL=meta-llama/Llama-2-7b-hf DISTRIBUTED_EXECUTOR_BACKEND=ray VLLM_USE_RAY_SPMD_WORKER=1 VLLM_USE_RAY_COMPILED_DAG=1 pytest -v -s distributed/test_basic_distributed_correctness.py
-  - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_chunked_prefill_distributed.py
-  - TEST_DIST_MODEL=meta-llama/Llama-2-7b-hf DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_chunked_prefill_distributed.py
-  - TEST_DIST_MODEL=llava-hf/llava-1.5-7b-hf DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_multimodal_broadcast.py
-  - TEST_DIST_MODEL=llava-hf/llava-v1.6-mistral-7b-hf DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_multimodal_broadcast.py
-  - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=mp pytest -v -s distributed/test_basic_distributed_correctness.py
-  - TEST_DIST_MODEL=meta-llama/Llama-2-7b-hf DISTRIBUTED_EXECUTOR_BACKEND=mp pytest -v -s distributed/test_basic_distributed_correctness.py
-  - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=mp pytest -v -s distributed/test_chunked_prefill_distributed.py
-  - TEST_DIST_MODEL=meta-llama/Llama-2-7b-hf DISTRIBUTED_EXECUTOR_BACKEND=mp pytest -v -s distributed/test_chunked_prefill_distributed.py
-  - TEST_DIST_MODEL=llava-hf/llava-1.5-7b-hf DISTRIBUTED_EXECUTOR_BACKEND=mp pytest -v -s distributed/test_multimodal_broadcast.py
-  - TEST_DIST_MODEL=llava-hf/llava-v1.6-mistral-7b-hf DISTRIBUTED_EXECUTOR_BACKEND=mp pytest -v -s distributed/test_multimodal_broadcast.py
   - pytest -v -s spec_decode/e2e/test_integration_dist_tp2.py
   - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s test_sharded_state_loader.py
   - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s distributed/test_utils.py
@@ -107,11 +96,6 @@ steps:
   fast_check: true
   commands:
   - pytest -v -s distributed/test_pynccl.py
-  # We want to test that models which use 2 GPUs work with 4 GPUs, which is why we duplicate them here.
-  # See https://github.com/vllm-project/vllm/pull/5473#issuecomment-2166601837 for context.
-  - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_basic_distributed_correctness.py
-  - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=ray VLLM_USE_RAY_SPMD_WORKER=1 VLLM_USE_RAY_COMPILED_DAG=1 pytest -v -s distributed/test_basic_distributed_correctness.py
-  - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=mp pytest -v -s distributed/test_basic_distributed_correctness.py
   - pytest -v -s spec_decode/e2e/test_integration_dist_tp4.py
 
 - label: Pipeline Parallelism Test
@@ -279,9 +263,6 @@ steps:
   # NOTE: don't test llama model here, it seems hf implementation is buggy
   # see https://github.com/vllm-project/vllm/pull/5689 for details
   - pytest -v -s distributed/test_custom_all_reduce.py
-  - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_basic_distributed_correctness.py
-  - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=mp pytest -v -s distributed/test_basic_distributed_correctness.py
   - pip install https://github.com/flashinfer-ai/flashinfer/releases/download/v0.1.2/flashinfer-0.1.2+cu121torch2.4-cp310-cp310-linux_x86_64.whl
-  - VLLM_ATTENTION_BACKEND=FLASHINFER TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_basic_distributed_correctness.py
+  - TARGET_TEST_SUITE=A100 pytest -v -s distributed/test_basic_distributed_correctness.py
-  - VLLM_ATTENTION_BACKEND=FLASHINFER TEST_DIST_MODEL=meta-llama/Meta-Llama-3-8B DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_basic_distributed_correctness.py
   - pytest -v -s -x lora/test_mixtral.py

tests/distributed/test_basic_distributed_correctness.py

@@ -1,15 +1,10 @@
 """Compare the outputs of HF and distributed vLLM when using greedy sampling.
-vLLM will allocate all the available memory, so we need to run the tests one
+
-by one. The solution is to pass arguments (model name) by environment
-variables.
 Run:
 ```sh
 cd $VLLM_PATH/tests
 
-TEST_DIST_MODEL=facebook/opt-125m pytest \
+pytest distributed/test_basic_distributed_correctness.py
-    distributed/test_basic_distributed_correctness.py
-TEST_DIST_MODEL=meta-llama/Llama-2-7b-hf \
-    distributed/test_basic_distributed_correctness.py
 ```
 """
 import os
@@ -19,27 +14,48 @@ import pytest
 from vllm.utils import cuda_device_count_stateless
 
 from ..models.utils import check_outputs_equal
+from ..utils import fork_new_process_for_each_test
 
-MODELS = [
+TARGET_TEST_SUITE = os.environ.get("TARGET_TEST_SUITE", "L4")
-    os.environ["TEST_DIST_MODEL"],
-]
-DISTRIBUTED_EXECUTOR_BACKEND = "DISTRIBUTED_EXECUTOR_BACKEND"
 
 
 @pytest.mark.skipif(cuda_device_count_stateless() < 2,
                     reason="Need at least 2 GPUs to run the test.")
-@pytest.mark.parametrize("model", MODELS)
+@pytest.mark.parametrize(
-@pytest.mark.parametrize("dtype", ["half"])
+    "model, distributed_executor_backend, attention_backend, test_suite", [
-@pytest.mark.parametrize("max_tokens", [5])
+        ("facebook/opt-125m", "ray", "", "L4"),
+        ("facebook/opt-125m", "mp", "", "L4"),
+        ("meta-llama/Llama-2-7b-hf", "ray", "", "L4"),
+        ("meta-llama/Llama-2-7b-hf", "mp", "", "L4"),
+        ("facebook/opt-125m", "ray", "", "A100"),
+        ("facebook/opt-125m", "mp", "", "A100"),
+        ("facebook/opt-125m", "mp", "FLASHINFER", "A100"),
+        ("meta-llama/Meta-Llama-3-8B", "ray", "FLASHINFER", "A100"),
+    ])
+@fork_new_process_for_each_test
 def test_models(
     hf_runner,
     vllm_runner,
     example_prompts,
     model: str,
-    dtype: str,
+    distributed_executor_backend: str,
-    max_tokens: int,
+    attention_backend: str,
+    test_suite: str,
 ) -> None:
-    distributed_executor_backend = os.getenv(DISTRIBUTED_EXECUTOR_BACKEND)
+
+    if test_suite != TARGET_TEST_SUITE:
+        pytest.skip(f"Skip test for {test_suite}")
+
+    if model == "meta-llama/Llama-2-7b-hf" and distributed_executor_backend == "ray" and attention_backend == "" and test_suite == "L4":  # noqa
+        # test ray adag
+        os.environ['VLLM_USE_RAY_SPMD_WORKER'] = "1"
+        os.environ['VLLM_USE_RAY_COMPILED_DAG'] = "1"
+
+    if attention_backend:
+        os.environ["VLLM_ATTENTION_BACKEND"] = attention_backend
+
+    dtype = "half"
+    max_tokens = 5
 
     # NOTE: take care of the order. run vLLM first, and then run HF.
     # vLLM needs a fresh new process without cuda initialization.

tests/distributed/test_chunked_prefill_distributed.py

@@ -1,46 +1,39 @@
 """Compare the outputs of HF and distributed vLLM when using greedy sampling.
-vLLM will allocate all the available memory, so we need to run the tests one
-by one. The solution is to pass arguments (model name) by environment
-variables.
 
 Run:
 ```sh
-TEST_DIST_MODEL=facebook/opt-125m pytest \
+pytest test_chunked_prefill_distributed.py
-    test_chunked_prefill_distributed.py
-TEST_DIST_MODEL=meta-llama/Llama-2-7b-hf \
-    test_chunked_prefill_distributed.py
 ```
 """
-import os
 
 import pytest
 
 from vllm.utils import cuda_device_count_stateless
 
 from ..models.utils import check_outputs_equal
-
+from ..utils import fork_new_process_for_each_test
-MODELS = [
-    os.environ["TEST_DIST_MODEL"],
-]
-DISTRIBUTED_EXECUTOR_BACKEND = "DISTRIBUTED_EXECUTOR_BACKEND"
 
 
 @pytest.mark.skipif(cuda_device_count_stateless() < 2,
                     reason="Need at least 2 GPUs to run the test.")
-@pytest.mark.parametrize("model", MODELS)
+@pytest.mark.parametrize("model, distributed_executor_backend", [
-@pytest.mark.parametrize("dtype", ["half"])
+    ("facebook/opt-125m", "ray"),
-@pytest.mark.parametrize("max_tokens", [5])
+    ("meta-llama/Llama-2-7b-hf", "ray"),
-@pytest.mark.parametrize("chunked_prefill_token_size", [16])
+    ("facebook/opt-125m", "mp"),
+    ("meta-llama/Llama-2-7b-hf", "mp"),
+])
+@fork_new_process_for_each_test
 def test_models(
     hf_runner,
     vllm_runner,
     example_prompts,
     model: str,
-    dtype: str,
+    distributed_executor_backend: str,
-    max_tokens: int,
-    chunked_prefill_token_size: int,
 ) -> None:
-    distributed_executor_backend = os.getenv(DISTRIBUTED_EXECUTOR_BACKEND)
+
+    dtype = "half"
+    max_tokens = 5
+    chunked_prefill_token_size = 16
 
     # Add a chunked prefill config.
     max_num_seqs = min(chunked_prefill_token_size, 256)

tests/distributed/test_multimodal_broadcast.py

@@ -1,44 +1,41 @@
 """Compare the outputs of HF and distributed vLLM when using greedy sampling.
-The second test will hang if more than one test is run per command, so we need
-to run the tests one by one. The solution is to pass arguments (model name) by
-environment variables.
 
 Run:
 ```sh
-TEST_DIST_MODEL=llava-hf/llava-1.5-7b-hf \
+pytest -s -v test_multimodal_broadcast.py
-    test_multimodal_broadcast.py
-TEST_DIST_MODEL=microsoft/Phi-3-vision-128k-instruct \
-    test_multimodal_broadcast.py
 ```
 """
-import os
 
 import pytest
 
 from vllm.utils import cuda_device_count_stateless
 
-model = os.environ["TEST_DIST_MODEL"]
+from ..utils import fork_new_process_for_each_test
-
-if model.startswith("llava-hf/llava-1.5"):
-    from ..models.test_llava import models, run_test
-elif model.startswith("llava-hf/llava-v1.6"):
-    from ..models.test_llava_next import models, run_test
-else:
-    raise NotImplementedError(f"Unsupported model: {model}")
 
 
-@pytest.mark.parametrize("tensor_parallel_size", [2])
+@pytest.mark.skipif(cuda_device_count_stateless() < 2,
-@pytest.mark.parametrize("dtype", ["half"])
+                    reason="Need at least 2 GPUs to run the test.")
-@pytest.mark.parametrize("max_tokens", [128])
+@pytest.mark.parametrize("model, distributed_executor_backend", [
-@pytest.mark.parametrize("num_logprobs", [5])
+    ("llava-hf/llava-1.5-7b-hf", "ray"),
-def test_models(hf_runner, vllm_runner, image_assets,
+    ("llava-hf/llava-v1.6-mistral-7b-hf", "ray"),
-                tensor_parallel_size: int, dtype: str, max_tokens: int,
+    ("llava-hf/llava-1.5-7b-hf", "mp"),
-                num_logprobs: int) -> None:
+    ("llava-hf/llava-v1.6-mistral-7b-hf", "mp"),
-    if cuda_device_count_stateless() < tensor_parallel_size:
+])
-        pytest.skip(
+@fork_new_process_for_each_test
-            f"Need at least {tensor_parallel_size} GPUs to run the test.")
+def test_models(hf_runner, vllm_runner, image_assets, model: str,
+                distributed_executor_backend: str) -> None:
 
-    distributed_executor_backend = os.getenv("DISTRIBUTED_EXECUTOR_BACKEND")
+    dtype = "half"
+    max_tokens = 5
+    num_logprobs = 5
+    tensor_parallel_size = 2
+
+    if model.startswith("llava-hf/llava-1.5"):
+        from ..models.test_llava import models, run_test
+    elif model.startswith("llava-hf/llava-v1.6"):
+        from ..models.test_llava_next import models, run_test
+    else:
+        raise NotImplementedError(f"Unsupported model: {model}")
 
     run_test(
         hf_runner,