e489ad7a21
- **Add SPDX license headers to python source files**
- **Check for SPDX headers using pre-commit**
commit 9d7ef44c3cfb72ca4c32e1c677d99259d10d4745
Author: Russell Bryant <rbryant@redhat.com>
Date: Fri Jan 31 14:18:24 2025 -0500
Add SPDX license headers to python source files
This commit adds SPDX license headers to python source files as
recommended to
the project by the Linux Foundation. These headers provide a concise way
that is
both human and machine readable for communicating license information
for each
source file. It helps avoid any ambiguity about the license of the code
and can
also be easily used by tools to help manage license compliance.
The Linux Foundation runs license scans against the codebase to help
ensure
we are in compliance with the licenses of the code we use, including
dependencies. Having these headers in place helps that tool do its job.
More information can be found on the SPDX site:
- https://spdx.dev/learn/handling-license-info/
Signed-off-by: Russell Bryant <rbryant@redhat.com>
commit 5a1cf1cb3b80759131c73f6a9dddebccac039dea
Author: Russell Bryant <rbryant@redhat.com>
Date: Fri Jan 31 14:36:32 2025 -0500
Check for SPDX headers using pre-commit
Signed-off-by: Russell Bryant <rbryant@redhat.com>
---------
Signed-off-by: Russell Bryant <rbryant@redhat.com>
285 lines
11 KiB
Python
285 lines
11 KiB
Python
# SPDX-License-Identifier: Apache-2.0
|
|
|
|
from itertools import cycle
|
|
|
|
import pytest
|
|
|
|
from vllm import SamplingParams
|
|
|
|
from ..utils import maybe_enable_chunked_prefill
|
|
from .conftest import run_equality_correctness_test
|
|
|
|
|
|
@pytest.mark.parametrize(
|
|
"common_llm_kwargs",
|
|
[{
|
|
"model_name": "JackFram/llama-160m",
|
|
|
|
# Skip cuda graph recording for fast test.
|
|
"enforce_eager": 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,
|
|
"disable_logprobs_during_spec_decoding": False,
|
|
}, {
|
|
"speculative_model": "JackFram/llama-68m",
|
|
"num_speculative_tokens": 3,
|
|
"disable_logprobs_during_spec_decoding": True,
|
|
}])
|
|
@pytest.mark.parametrize("batch_size", [8])
|
|
@pytest.mark.parametrize(
|
|
"output_len",
|
|
[
|
|
# Use smaller output len for fast test.
|
|
7,
|
|
])
|
|
@pytest.mark.parametrize("seed", [1])
|
|
@pytest.mark.parametrize("logprobs", [1, 6])
|
|
@pytest.mark.parametrize("prefill_chunk_size", [-1, 4, 12])
|
|
def test_logprobs_equality(vllm_runner, common_llm_kwargs,
|
|
per_test_common_llm_kwargs, baseline_llm_kwargs,
|
|
test_llm_kwargs, batch_size: int, output_len: int,
|
|
seed: int, logprobs: int, prefill_chunk_size: int):
|
|
"""Verify output logprobs are equal with and without speculative decoding,
|
|
as well as with and without chunked prefill.
|
|
"""
|
|
maybe_enable_chunked_prefill(prefill_chunk_size, common_llm_kwargs)
|
|
run_equality_correctness_test(vllm_runner,
|
|
common_llm_kwargs,
|
|
per_test_common_llm_kwargs,
|
|
baseline_llm_kwargs,
|
|
test_llm_kwargs,
|
|
batch_size,
|
|
output_len,
|
|
seed,
|
|
temperature=0.0,
|
|
logprobs=logprobs,
|
|
prompt_logprobs=logprobs,
|
|
disable_logprobs=test_llm_kwargs[
|
|
'disable_logprobs_during_spec_decoding'])
|
|
|
|
|
|
@pytest.mark.parametrize(
|
|
"common_llm_kwargs",
|
|
[{
|
|
"model_name": "JackFram/llama-68m",
|
|
|
|
# Skip cuda graph recording for fast test.
|
|
"enforce_eager": 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-160m",
|
|
"num_speculative_tokens": 3,
|
|
"disable_logprobs_during_spec_decoding": False,
|
|
}, {
|
|
"speculative_model": "JackFram/llama-160m",
|
|
"num_speculative_tokens": 6,
|
|
"disable_logprobs_during_spec_decoding": False,
|
|
}])
|
|
@pytest.mark.parametrize("batch_size", [8])
|
|
@pytest.mark.parametrize(
|
|
"output_len",
|
|
[
|
|
# Use smaller output len for fast test.
|
|
32,
|
|
])
|
|
@pytest.mark.parametrize("seed", [1])
|
|
@pytest.mark.parametrize("logprobs", [1, 6])
|
|
def test_logprobs_different_k(vllm_runner, common_llm_kwargs,
|
|
per_test_common_llm_kwargs, baseline_llm_kwargs,
|
|
test_llm_kwargs, batch_size: int,
|
|
output_len: int, seed: int, logprobs: int):
|
|
"""Veriy logprob greedy equality with different speculation lens.
|
|
"""
|
|
run_equality_correctness_test(vllm_runner,
|
|
common_llm_kwargs,
|
|
per_test_common_llm_kwargs,
|
|
baseline_llm_kwargs,
|
|
test_llm_kwargs,
|
|
batch_size,
|
|
output_len,
|
|
seed,
|
|
temperature=0.0,
|
|
logprobs=logprobs,
|
|
disable_logprobs=test_llm_kwargs[
|
|
'disable_logprobs_during_spec_decoding'])
|
|
|
|
|
|
@pytest.mark.parametrize(
|
|
"common_llm_kwargs",
|
|
[{
|
|
"model_name": "JackFram/llama-68m",
|
|
|
|
# Skip cuda graph recording for fast test.
|
|
"enforce_eager": 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-160m",
|
|
"num_speculative_tokens": 3,
|
|
"disable_logprobs_during_spec_decoding": False,
|
|
|
|
# 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,
|
|
}])
|
|
@pytest.mark.parametrize("batch_size", [8])
|
|
@pytest.mark.parametrize(
|
|
"output_len",
|
|
[
|
|
# Use smaller output len for fast test.
|
|
32,
|
|
])
|
|
@pytest.mark.parametrize("seed", [1])
|
|
@pytest.mark.parametrize("logprobs", [1])
|
|
def test_logprobs_when_skip_speculation(vllm_runner, common_llm_kwargs,
|
|
per_test_common_llm_kwargs,
|
|
baseline_llm_kwargs, test_llm_kwargs,
|
|
batch_size: int, output_len: int,
|
|
seed: int, logprobs: int):
|
|
"""Verify logprobs greedy equality when some sequences skip speculation.
|
|
"""
|
|
run_equality_correctness_test(vllm_runner,
|
|
common_llm_kwargs,
|
|
per_test_common_llm_kwargs,
|
|
baseline_llm_kwargs,
|
|
test_llm_kwargs,
|
|
batch_size,
|
|
output_len,
|
|
seed,
|
|
temperature=0.0,
|
|
logprobs=logprobs,
|
|
disable_logprobs=test_llm_kwargs[
|
|
'disable_logprobs_during_spec_decoding'])
|
|
|
|
|
|
@pytest.mark.parametrize(
|
|
"common_llm_kwargs",
|
|
[{
|
|
"model_name": "JackFram/llama-68m",
|
|
|
|
# Skip cuda graph recording for fast test.
|
|
"enforce_eager": 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-160m",
|
|
"num_speculative_tokens": 3,
|
|
"disable_logprobs_during_spec_decoding": False,
|
|
}])
|
|
@pytest.mark.parametrize("batch_size", [1])
|
|
@pytest.mark.parametrize(
|
|
"output_len",
|
|
[
|
|
# Use smaller output len for fast test.
|
|
32,
|
|
])
|
|
@pytest.mark.parametrize("seed", [1])
|
|
@pytest.mark.parametrize("logprobs", [6])
|
|
def test_logprobs_temp_1(vllm_runner, common_llm_kwargs,
|
|
per_test_common_llm_kwargs, baseline_llm_kwargs,
|
|
test_llm_kwargs, batch_size: int, output_len: int,
|
|
seed: int, logprobs: int):
|
|
"""Verify at least one logprob result has num_logprobs+1, which tests the
|
|
case where the sampled token is not in top-k logprobs.
|
|
|
|
Ideally, this test should validate equality with non-spec by getting
|
|
logprobs. This is left as future improvement.
|
|
"""
|
|
temperature = 1.0
|
|
|
|
prompts = [
|
|
"Hello, my name is",
|
|
"The president of the United States is",
|
|
"The capital of France is",
|
|
"The future of AI is",
|
|
"San Francisco is know for its",
|
|
"Facebook was created in 2004 by",
|
|
"Curious George is a",
|
|
"Python 3.11 brings improvements to its",
|
|
]
|
|
|
|
prompts = [prompt for prompt, _ in zip(cycle(prompts), range(batch_size))]
|
|
|
|
sampling_params = SamplingParams(
|
|
max_tokens=output_len,
|
|
ignore_eos=True,
|
|
temperature=temperature,
|
|
logprobs=logprobs,
|
|
)
|
|
|
|
sd_args = {
|
|
**common_llm_kwargs,
|
|
**per_test_common_llm_kwargs,
|
|
**test_llm_kwargs,
|
|
}
|
|
|
|
with vllm_runner(**sd_args) as vllm_model:
|
|
sd_outputs = vllm_model.generate_w_logprobs(prompts, sampling_params)
|
|
|
|
num_returned_logprobs = [
|
|
len(seq_logprobs) for seq_logprobs in sd_outputs[-1]
|
|
]
|
|
|
|
# Assert one of the returned logprobs has > num_logprobs (indicating the
|
|
# sampled token is not in top-k).
|
|
assert any(
|
|
[num_returned > logprobs for num_returned in num_returned_logprobs])
|
|
|
|
|
|
@pytest.mark.parametrize(
|
|
"common_llm_kwargs",
|
|
[{
|
|
"model_name": "JackFram/llama-160m",
|
|
# Skip cuda graph recording for fast test.
|
|
"enforce_eager": 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,
|
|
"disable_logprobs_during_spec_decoding": True,
|
|
}])
|
|
@pytest.mark.parametrize("seed", [1])
|
|
@pytest.mark.parametrize("batch_size", [4])
|
|
@pytest.mark.parametrize(
|
|
"output_len",
|
|
[
|
|
# Use smaller output len for fast test.
|
|
32,
|
|
])
|
|
@pytest.mark.parametrize("logprobs", [0])
|
|
def test_logprobs_disabled(vllm_runner, common_llm_kwargs,
|
|
per_test_common_llm_kwargs, baseline_llm_kwargs,
|
|
test_llm_kwargs, batch_size: int, output_len: int,
|
|
seed: int, logprobs: int):
|
|
"""Check the behavior when logprobs are disabled.
|
|
Token choices should match with the base model.
|
|
"""
|
|
run_equality_correctness_test(vllm_runner,
|
|
common_llm_kwargs,
|
|
per_test_common_llm_kwargs,
|
|
baseline_llm_kwargs,
|
|
test_llm_kwargs,
|
|
batch_size,
|
|
output_len,
|
|
seed,
|
|
temperature=0.0,
|
|
logprobs=logprobs,
|
|
disable_logprobs=test_llm_kwargs[
|
|
'disable_logprobs_during_spec_decoding'])
|