[Core] NGram GPU Implementation compatible with Async Scheduler (#29184)

tests/v1/e2e/test_async_scheduling.py

@@ -98,7 +98,7 @@ def test_without_spec_decoding(
 
 @single_gpu_only
 @large_gpu_mark(min_gb=16)
-def test_with_spec_decoding(sample_json_schema, monkeypatch: pytest.MonkeyPatch):
+def test_with_eagle3_spec_decoding(sample_json_schema, monkeypatch: pytest.MonkeyPatch):
     """Test consistency and acceptance rates with some different combos of
     preemption, executor, async scheduling, prefill chunking,
     spec decoding model length.
@@ -154,6 +154,42 @@ def test_with_spec_decoding(sample_json_schema, monkeypatch: pytest.MonkeyPatch)
     )
 
 
+def test_with_ngram_gpu_spec_decoding(monkeypatch: pytest.MonkeyPatch):
+    """Test ngram_gpu speculative decoding with different configurations.
+
+    This test specifically validates ngram_gpu behavior with various:
+    - Number of speculative tokens (2-6)
+    - Prompt lookup window sizes (min/max)
+    - Async scheduling enabled (as in production)
+    - Different executors and chunking settings
+    """
+
+    # Variant with larger speculation window
+    ngram_gpu_config = {
+        "method": "ngram_gpu",
+        "num_speculative_tokens": 3,
+        "prompt_lookup_max": 3,
+        "prompt_lookup_min": 2,
+    }
+
+    # Test configurations covering various scenarios
+    # test_preemption, executor, async_scheduling,
+    # spec_config, test_prefill_chunking
+    test_configs = [
+        (False, "mp", False, None, False),
+        (False, "mp", False, ngram_gpu_config, False),
+        (True, "mp", False, ngram_gpu_config, True),
+        (False, "mp", True, ngram_gpu_config, False),
+        (True, "mp", True, ngram_gpu_config, False),
+        (True, "uni", True, ngram_gpu_config, False),
+        (True, "mp", True, ngram_gpu_config, True),
+    ]
+
+    # Use MODEL (Qwen) for ngram_gpu tests as it's lighter weight
+    # and ngram_gpu doesn't require a specific draft model
+    run_tests(monkeypatch, MODEL, test_configs, [{}])
+
+
 @dynamo_config.patch(cache_size_limit=16)
 def run_tests(
     monkeypatch: pytest.MonkeyPatch,
@@ -282,11 +318,12 @@ def run_test(
         else dict(gpu_memory_utilization=0.9)
     )
     spec_mml = (spec_config or {}).get("max_model_len")
+    spec_method = (spec_config or {}).get("method", "none")
     test_config = (
         f"executor={executor}, preemption={test_preemption}, "
         f"async_sched={async_scheduling}, "
         f"chunk_prefill={test_prefill_chunking}, "
-        f"spec_decoding={spec_decoding}, spec_mml={spec_mml}"
+        f"spec_decoding={spec_decoding}, spec_method={spec_method}, spec_mml={spec_mml}"
     )
     print("-" * 80)
     print(f"---- TESTING {test_str}: {test_config}")
@@ -294,7 +331,7 @@ def run_test(
 
     with VllmRunner(
         model,
-        max_model_len=512,
+        max_model_len=4096,
         enable_chunked_prefill=test_prefill_chunking,
         # Force prefill chunking
         max_num_batched_tokens=48 if test_prefill_chunking else None,

tests/v1/e2e/test_spec_decode.py

@@ -183,6 +183,34 @@ def test_ngram_and_suffix_correctness(
     cleanup_dist_env_and_memory()
 
 
+@pytest.mark.parametrize("async_scheduling", [True], ids=["async"])
+@single_gpu_only
+@large_gpu_mark(min_gb=20)
+def test_ngram_gpu_default_with_async_scheduling(
+    async_scheduling: bool,
+):
+    """
+    Test ngram_gpu speculative decoding (k=3) correctness with and without
+    async scheduling, validated via GSM8K accuracy.
+    Uses Qwen/Qwen3-8B (ref GSM8K accuracy: 87%-92%).
+    """
+    qwen3_model = "Qwen/Qwen3-8B"
+    spec_llm = LLM(
+        model=qwen3_model,
+        speculative_config={
+            "method": "ngram_gpu",
+            "prompt_lookup_max": 3,
+            "prompt_lookup_min": 2,
+            "num_speculative_tokens": 2,
+        },
+        max_model_len=4096,
+        async_scheduling=async_scheduling,
+    )
+    evaluate_llm_for_gsm8k(spec_llm, expected_accuracy_threshold=0.8)
+    del spec_llm
+    cleanup_dist_env_and_memory()
+
+
 @single_gpu_only
 @large_gpu_mark(min_gb=20)
 def test_suffix_decoding_acceptance(