[Speculative Decoding] Support draft model on different tensor-parallel size than target model

.buildkite/test-pipeline.yaml

@@ -46,7 +46,7 @@ steps:
   - 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
-  - pytest -v -s spec_decode/e2e/test_integration_dist.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
 
@@ -60,6 +60,7 @@ steps:
   # 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=mp pytest -v -s distributed/test_basic_distributed_correctness.py
+  - pytest -v -s spec_decode/e2e/test_integration_dist_tp4.py
 
 - label: Engine Test
   mirror_hardwares: [amd]

benchmarks/benchmark_latency.py

@@ -24,6 +24,8 @@ def main(args: argparse.Namespace):
         model=args.model,
         speculative_model=args.speculative_model,
         num_speculative_tokens=args.num_speculative_tokens,
+        speculative_draft_tensor_parallel_size=\
+            args.speculative_draft_tensor_parallel_size,
         tokenizer=args.tokenizer,
         quantization=args.quantization,
         tensor_parallel_size=args.tensor_parallel_size,
@@ -125,6 +127,10 @@ def run_to_completion(profile_dir: Optional[str] = None):
     parser.add_argument('--model', type=str, default='facebook/opt-125m')
     parser.add_argument('--speculative-model', type=str, default=None)
     parser.add_argument('--num-speculative-tokens', type=int, default=None)
+    parser.add_argument('--speculative-draft-tensor-parallel-size',
+                        '-spec-draft-tp',
+                        type=int,
+                        default=None)
     parser.add_argument('--tokenizer', type=str, default=None)
     parser.add_argument('--quantization',
                         '-q',

tests/spec_decode/e2e/test_integration_dist_tp2.py

@@ -0,0 +1,177 @@
+"""Tests which cover integration of the speculative decoding framework with
+tensor parallelism.
+"""
+
+import pytest
+import torch
+
+from vllm.utils import is_hip
+
+from .conftest import run_greedy_equality_correctness_test
+
+
+@pytest.mark.skipif(torch.cuda.device_count() < 2,
+                    reason="Need at least 2 GPUs to run the test.")
+@pytest.mark.parametrize(
+    "common_llm_kwargs",
+    [{
+        "model": "JackFram/llama-68m",
+
+        # Skip cuda graph recording for fast test.
+        "enforce_eager": True,
+
+        # Required for spec decode.
+        "use_v2_block_manager": True,
+        "tensor_parallel_size": 2,
+
+        # Use AsyncLLM engine, so that the engine runs in its own process.
+        # Otherwise, since vLLM does not follow true SPMD, the test runner
+        # process will have both the engine and the rank0 worker. NCCL is not
+        # cleaned up properly, and its server host thread leaks, causing the
+        # second run of the test to fail with internal NCCL error.
+        "use_async": True,
+    }])
+@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
+@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
+@pytest.mark.parametrize("test_llm_kwargs", [
+    {
+        "speculative_model": "JackFram/llama-68m",
+        "num_speculative_tokens": 3,
+    },
+    {
+        "speculative_model": "[ngram]",
+        "num_speculative_tokens": 5,
+        "ngram_prompt_lookup_max": 3,
+    },
+])
+@pytest.mark.parametrize("batch_size", [2])
+@pytest.mark.parametrize(
+    "output_len",
+    [
+        # Use smaller output len for fast test.
+        32,
+    ])
+@pytest.mark.parametrize("seed", [1])
+def test_target_model_tp_gt_1(baseline_llm_generator, test_llm_generator,
+                              batch_size: int, output_len: int):
+    """Verify greedy equality when tensor parallelism is used.
+    """
+    if is_hip():
+        pytest.skip("hip is not well-supported yet")
+    run_greedy_equality_correctness_test(baseline_llm_generator,
+                                         test_llm_generator,
+                                         batch_size,
+                                         max_output_len=output_len,
+                                         force_output_len=True)
+
+
+@pytest.mark.skipif(torch.cuda.device_count() < 2,
+                    reason="Need at least 2 GPUs to run the test.")
+@pytest.mark.parametrize(
+    "common_llm_kwargs",
+    [{
+        # Use a small model for a fast test.
+        # Note this is repeated in the test body; to initialize a tokenizer.
+        "model": "JackFram/llama-68m",
+
+        # Skip cuda graph recording for fast test.
+        "enforce_eager": True,
+
+        # Required for spec decode.
+        "use_v2_block_manager": True,
+        "tensor_parallel_size": 2,
+
+        # Use AsyncLLM engine, so that the engine runs in its own process.
+        # Otherwise, since vLLM does not follow true SPMD, the test runner
+        # process will have both the engine and the rank0 worker. NCCL is not
+        # cleaned up properly, and its server host thread leaks, causing the
+        # second run of the test to fail with internal NCCL error.
+        "use_async": True,
+    }])
+@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
+@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
+@pytest.mark.parametrize("test_llm_kwargs", [
+    {
+        "speculative_model": "JackFram/llama-68m",
+        "num_speculative_tokens": 5,
+        "speculative_draft_tensor_parallel_size": 1,
+    },
+])
+@pytest.mark.parametrize("batch_size", [2])
+@pytest.mark.parametrize("seed", [1])
+def test_draft_model_tp_lt_target_model_tp2(test_llm_generator,
+                                            baseline_llm_generator,
+                                            batch_size: int):
+    """Verify spec decode works well with smaller tp for draft models.
+    """
+    run_greedy_equality_correctness_test(baseline_llm_generator,
+                                         test_llm_generator,
+                                         batch_size,
+                                         max_output_len=32,
+                                         force_output_len=True)
+
+
+@pytest.mark.skipif(torch.cuda.device_count() < 2,
+                    reason="Need at least 2 GPUs to run the test.")
+@pytest.mark.parametrize(
+    "common_llm_kwargs",
+    [{
+        "model": "JackFram/llama-160m",
+
+        # Skip cuda graph recording for fast test.
+        "enforce_eager": True,
+
+        # Required for spec decode.
+        "use_v2_block_manager": True,
+        "tensor_parallel_size": 2,
+
+        # Use AsyncLLM engine, so that the engine runs in its own process.
+        # Otherwise, since vLLM does not follow true SPMD, the test runner
+        # process will have both the engine and the rank0 worker. NCCL is not
+        # cleaned up properly, and its server host thread leaks, causing the
+        # second run of the test to fail with internal NCCL error.
+        "use_async": True,
+    }])
+@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
+@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
+@pytest.mark.parametrize("test_llm_kwargs", [
+    {
+        "speculative_model": "JackFram/llama-68m",
+        "num_speculative_tokens": 5,
+
+        # Artificially limit the draft model max model len; this forces vLLM
+        # to skip speculation once the sequences grow beyond 32-k tokens.
+        "speculative_max_model_len": 32,
+    },
+    {
+        "speculative_model": "JackFram/llama-68m",
+        "num_speculative_tokens": 5,
+
+        # Artificially limit the draft model max model len; this forces vLLM
+        # to skip speculation once the sequences grow beyond 32-k tokens.
+        "speculative_max_model_len": 32,
+        "speculative_draft_tensor_parallel_size": 1,
+    },
+])
+@pytest.mark.parametrize("batch_size", [8])
+@pytest.mark.parametrize(
+    "output_len",
+    [
+        # This must be a good bit larger than speculative_max_model_len so that
+        # we can test the case where all seqs are skipped, but still small to
+        # ensure fast test.
+        64,
+    ])
+@pytest.mark.parametrize("seed", [1])
+def test_skip_speculation(baseline_llm_generator, test_llm_generator,
+                          batch_size: int, output_len: int):
+    """Verify greedy equality when some (or all) sequences skip speculation.
+    We do this by setting the max model len of the draft model to an
+    artificially low value, such that when the sequences grow beyond it, they
+    are skipped in speculative decoding.
+    """
+    run_greedy_equality_correctness_test(baseline_llm_generator,
+                                         test_llm_generator,
+                                         batch_size,
+                                         max_output_len=output_len,
+                                         force_output_len=True)