[V1] TPU - Add tensor parallel support via Ray (vllm-project#13618)

Signed-off-by: Alexander Matveev <[email protected]>
Signed-off-by: Louis Ulmer <[email protected]>

Author: alexm-redhat

tests/entrypoints/llm/test_accuracy.py

@@ -42,6 +42,10 @@ def run_test(more_args=None):
             ), f"Expected: {EXPECTED_VALUE} |  Measured: {measured_value}"
 
 
+# TODO: [AlexM] Fix it with new CI/CD tests
+TPU_TP_TEST_STR = ""  #"tensor_parallel_size=4"
+
+
 @pytest.mark.skipif(not current_platform.is_cuda()
                     and not current_platform.is_tpu(),
                     reason="V1 is currently only supported on CUDA and TPU")
@@ -56,6 +60,10 @@ def test_lm_eval_accuracy_v1_engine(monkeypatch):
             # Limit compilation time for TPU V1
             more_args = "max_num_seqs=64"
 
+            # Add TP test (if provided)
+            if TPU_TP_TEST_STR:
+                more_args += ",{}".format(TPU_TP_TEST_STR)
+
         run_test(more_args)
 
 

tests/v1/tpu/__init__.py

@@ -0,0 +1,54 @@
+# SPDX-License-Identifier: Apache-2.0
+"""A basic correctness check for TPUs
+
+Run `pytest tests/v1/tpu/test_basic.py`.
+"""
+import pytest
+
+from vllm.platforms import current_platform
+
+from ...conftest import VllmRunner
+
+MODELS = [
+    # "Qwen/Qwen2-7B-Instruct",
+    "meta-llama/Llama-3.1-8B",
+    # TODO: Add models here as necessary
+]
+
+TENSOR_PARALLEL_SIZES = [1]
+
+# TODO: Enable when CI/CD will have a multi-tpu instance
+# TENSOR_PARALLEL_SIZES = [1, 4]
+
+
+@pytest.mark.skipif(not current_platform.is_tpu(),
+                    reason="This is a basic test for TPU only")
+@pytest.mark.parametrize("model", MODELS)
+@pytest.mark.parametrize("max_tokens", [5])
+@pytest.mark.parametrize("enforce_eager", [True])
+@pytest.mark.parametrize("tensor_parallel_size", TENSOR_PARALLEL_SIZES)
+def test_models(
+    monkeypatch,
+    model: str,
+    max_tokens: int,
+    enforce_eager: bool,
+    tensor_parallel_size: int,
+) -> None:
+    prompt = "The next numbers of the sequence " + ", ".join(
+        str(i) for i in range(1024)) + " are:"
+    example_prompts = [prompt]
+
+    with monkeypatch.context() as m:
+        m.setenv("VLLM_USE_V1", "1")
+
+        with VllmRunner(
+                model,
+                max_model_len=8192,
+                enforce_eager=enforce_eager,
+                gpu_memory_utilization=0.7,
+                max_num_seqs=16,
+                tensor_parallel_size=tensor_parallel_size) as vllm_model:
+            vllm_outputs = vllm_model.generate_greedy(example_prompts,
+                                                      max_tokens)
+    output = vllm_outputs[0][1]
+    assert "1024" in output

tests/v1/tpu/test_basic.py

@@ -73,9 +73,14 @@ class RayDistributedExecutor(DistributedExecutorBase):
     def _init_executor(self) -> None:
         self.forward_dag: Optional[ray.dag.CompiledDAG] = None
         if envs.VLLM_USE_V1:
-            # v1 always uses the compiled DAG and SPMD worker.
+            # V1 uses SPMD worker and compiled DAG
             os.environ["VLLM_USE_RAY_SPMD_WORKER"] = "1"
             os.environ["VLLM_USE_RAY_COMPILED_DAG"] = "1"
+
+            # For TPU, avoid compiling NVIDIA's NCCL
+            if current_platform.is_tpu():
+                os.environ["VLLM_USE_RAY_COMPILED_DAG_NCCL_CHANNEL"] = "0"
+
         # If the env var is set, it uses the Ray's compiled DAG API
         # which optimizes the control plane overhead.
         # Run vLLM with VLLM_USE_RAY_COMPILED_DAG=1 to enable it.

vllm/executor/ray_distributed_executor.py

@@ -11,6 +11,7 @@
 from vllm.config import ParallelConfig
 from vllm.executor.msgspec_utils import decode_hook, encode_hook
 from vllm.logger import init_logger
+from vllm.platforms import current_platform
 from vllm.sequence import ExecuteModelRequest, IntermediateTensors
 from vllm.utils import get_ip
 from vllm.worker.worker_base import WorkerWrapperBase
@@ -106,10 +107,15 @@ def setup_device_if_necessary(self):
             # on a background thread, so we need to reset torch's current
             # device.
             # We can remove this API after it is fixed in compiled graph.
-            import torch
             assert self.worker is not None, "Worker is not initialized"
             if not self.compiled_dag_cuda_device_set:
-                torch.cuda.set_device(self.worker.device)
+                if current_platform.is_tpu():
+                    # Not needed
+                    pass
+                else:
+                    import torch
+                    torch.cuda.set_device(self.worker.device)
+
                 self.compiled_dag_cuda_device_set = True
 
         def execute_model_ray(

vllm/executor/ray_utils.py

@@ -21,6 +21,7 @@
 from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalKwargs
 from vllm.multimodal.utils import group_mm_inputs_by_modality
 from vllm.sampling_params import SamplingType
+from vllm.sequence import IntermediateTensors
 from vllm.utils import LayerBlockType, cdiv, is_pin_memory_available
 from vllm.v1.attention.backends.pallas import (NUM_KV_PAGES_PER_BLOCK,
                                                NUM_QUERIES_PER_BLOCK,
@@ -545,6 +546,7 @@ def _gather_encoder_outputs(
     def execute_model(
         self,
         scheduler_output: "SchedulerOutput",
+        intermediate_tensors: Optional[IntermediateTensors] = None,
     ) -> ModelRunnerOutput:
         # Update cached state
         self._update_states(scheduler_output)

vllm/v1/worker/tpu_model_runner.py

@@ -96,7 +96,8 @@ def init_device(self):
 
         # Set random seed.
         set_random_seed(self.model_config.seed)
-        xm.set_rng_state(self.model_config.seed, self.device)
+        if self.model_config.seed is not None:
+            xm.set_rng_state(self.model_config.seed, self.device)
 
         # Increase the cache size limit, which is the maximum number of
         # dynamo graphs that can be compiled.