xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-07-09 19:27:20 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Bugfix] Fix the speculative decoding test by setting the target dtype (#19633)

Browse Source
This commit is contained in:
Lu Fang 2025-06-14 11:57:32 +08:00 committed by GitHub
parent d1e34cc9ac
commit 06be858828
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 24 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

24
tests/spec_decode/e2e/test_multistep_correctness.py
View File

@ -57,6 +57,9 @@ from .conftest import (get_output_from_llm_generator,
# Skip cuda graph recording for fast test. # 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( @pytest.mark.parametrize(
"per_test_common_llm_kwargs", "per_test_common_llm_kwargs",
@ -139,6 +142,9 @@ def test_spec_decode_e2e_with_detokenization(test_llm_generator,
# Print spec metrics. # Print spec metrics.
"disable_log_stats": False, "disable_log_stats": False,
# The original model is float32, keep it for numerical stability.
"dtype": "float32",
}]) }])
@pytest.mark.parametrize( @pytest.mark.parametrize(
"per_test_common_llm_kwargs", "per_test_common_llm_kwargs",
@ -216,6 +222,9 @@ def test_spec_decode_e2e_greedy_correctness_tiny_model_bs1(
# Print spec metrics. # Print spec metrics.
"disable_log_stats": False, "disable_log_stats": False,
# The original model is float32, keep it for numerical stability.
"dtype": "float32",
}]) }])
@pytest.mark.parametrize( @pytest.mark.parametrize(
"per_test_common_llm_kwargs", "per_test_common_llm_kwargs",
@ -279,6 +288,9 @@ def test_spec_decode_e2e_greedy_correctness_tiny_model_large_bs(
[{ [{
# Skip cuda graph recording for fast test. # 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( @pytest.mark.parametrize(
"per_test_common_llm_kwargs", "per_test_common_llm_kwargs",
@ -464,6 +476,8 @@ def test_spec_decode_e2e_greedy_correctness_real_model_large_bs(
# Skip cuda graph recording for fast test. # 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("per_test_common_llm_kwargs", [
{ {
@ -523,6 +537,8 @@ def test_spec_decode_e2e_greedy_correctness_with_preemption(
# Skip cuda graph recording for fast test. # 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( @pytest.mark.parametrize(
"per_test_common_llm_kwargs", "per_test_common_llm_kwargs",
@ -589,6 +605,8 @@ def test_spec_decode_different_block_size(vllm_runner, common_llm_kwargs,
# Skip cuda graph recording for fast test. # 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("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}]) @pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@ -655,6 +673,8 @@ def test_skip_speculation(vllm_runner, common_llm_kwargs,
# Skip cuda graph recording for fast test. # 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("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}]) @pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@ -706,6 +726,8 @@ def test_disable_speculation(vllm_runner, common_llm_kwargs,
# Skip cuda graph recording for fast test. # 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("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}]) @pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@ -763,6 +785,8 @@ def test_many_k(vllm_runner, common_llm_kwargs, per_test_common_llm_kwargs,
# Skip cuda graph recording for fast test. # 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("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}]) @pytest.mark.parametrize("baseline_llm_kwargs", [{}])