biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Bugfix][2/n] Fix speculative decoding CI - Fix test_ngram_e2e_greedy_correctness (#19644)

Browse Source
This commit is contained in:
Lu Fang
2025-06-15 12:15:41 +08:00
committed by GitHub
parent e13945f9dd
commit ee1531bc38
5 changed files with 50 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
tests/spec_decode/e2e/test_integration.py
View File

@@ -14,10 +14,13 @@ MAIN_MODEL = "JackFram/llama-68m"
@pytest.mark.parametrize(
"common_llm_kwargs",
[{
"model_name": "JackFram/llama-68m",
# Verify equality when cuda graphs allowed.
"enforce_eager": False,
"model_name": "JackFram/llama-68m",
# The original model is float32, keep it for numerical stability.
"dtype": "float32",
}])
@pytest.mark.parametrize(
"per_test_common_llm_kwargs",
@@ -59,6 +62,9 @@ def test_spec_decode_cuda_graph(vllm_runner, common_llm_kwargs,
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# The original model is float32, keep it for numerical stability.
"dtype": "float32",
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [])
@pytest.mark.parametrize(
@@ -117,6 +123,9 @@ def test_speculative_model_quantization_config(vllm_runner, common_llm_kwargs,
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# The original model is float32, keep it for numerical stability.
"dtype": "float32",
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])

18
tests/spec_decode/e2e/test_logprobs.py
View File

@@ -17,7 +17,10 @@ from .conftest import run_equality_correctness_test
"model_name": "JackFram/llama-160m",
# Skip cuda graph recording for fast test.
"enforce_eager": True
"enforce_eager": True,
# The original model is float32, keep it for numerical stability.
"dtype": "float32",
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@@ -75,6 +78,9 @@ def test_logprobs_equality(vllm_runner, common_llm_kwargs,
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# The original model is float32, keep it for numerical stability.
"dtype": "float32",
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@@ -128,6 +134,9 @@ def test_logprobs_different_k(vllm_runner, common_llm_kwargs,
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# The original model is float32, keep it for numerical stability.
"dtype": "float32",
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@@ -182,6 +191,9 @@ def test_logprobs_when_skip_speculation(vllm_runner, common_llm_kwargs,
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# The original model is float32, keep it for numerical stability.
"dtype": "float32",
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@@ -256,8 +268,12 @@ def test_logprobs_temp_1(vllm_runner, common_llm_kwargs,
"common_llm_kwargs",
[{
"model_name": "JackFram/llama-160m",
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# The original model is float32, keep it for numerical stability.
"dtype": "float32",
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])

3
tests/spec_decode/e2e/test_mlp_correctness.py
View File

@@ -494,6 +494,9 @@ def test_mlp_disable_queue(vllm_runner, common_llm_kwargs,
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Precision
"dtype": PRECISION,
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])

18
tests/spec_decode/e2e/test_ngram_correctness.py
View File

@@ -40,6 +40,9 @@ from .conftest import run_equality_correctness_test
# Print spec metrics.
"disable_log_stats": False,
# The original model is float32, keep it for numerical stability.
"dtype": "float32",
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [
{
@@ -97,6 +100,9 @@ def test_ngram_e2e_greedy_correctness(vllm_runner, common_llm_kwargs,
# Print spec metrics.
"disable_log_stats": False,
# The original model is float32, keep it for numerical stability.
"dtype": "float32",
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [
{
@@ -160,6 +166,9 @@ def test_ngram_e2e_greedy_logprobs(vllm_runner, common_llm_kwargs,
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# The original model is float32, keep it for numerical stability.
"dtype": "float32",
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [
{
@@ -221,6 +230,9 @@ def test_ngram_e2e_greedy_correctness_with_preemption(
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# The original model is float32, keep it for numerical stability.
"dtype": "float32",
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@@ -281,6 +293,9 @@ def test_ngram_different_k(vllm_runner, common_llm_kwargs,
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# The original model is float32, keep it for numerical stability.
"dtype": "float32",
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@@ -337,6 +352,9 @@ def test_ngram_disable_queue(vllm_runner, common_llm_kwargs,
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# The original model is float32, keep it for numerical stability.
"dtype": "float32",
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])