Skip to content

GitLab

Unverified Commit dd572c0a authored by Woosuk Kwon's avatar Woosuk Kwon Committed by GitHub
Browse files

[V0 Deprecation] Remove V0 Spec Decode workers (#21152) 


Signed-off-by: default avatarWoosuk Kwon <woosuk.kwon@berkeley.edu>
parent 9ffe905a
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 0 additions and 5966 deletions
tests/spec_decode/e2e/test_logprobs.py deleted 100644 → 0
View file @ 9ffe905a
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
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,
# 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", [{}])
@pytest.mark.parametrize("test_llm_kwargs", [{
"speculative_config": {
"model": "JackFram/llama-68m",
"num_speculative_tokens": 3,
"disable_logprobs": False,
},
}, {
"speculative_config": {
"model": "JackFram/llama-68m",
"num_speculative_tokens": 3,
"disable_logprobs": 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["speculative_config"]
["disable_logprobs"])
@pytest.mark.parametrize(
"common_llm_kwargs",
[{
"model_name": "JackFram/llama-68m",
# 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", [{}])
@pytest.mark.parametrize("test_llm_kwargs", [{
"speculative_config": {
"model": "JackFram/llama-160m",
"num_speculative_tokens": 3,
"disable_logprobs": False,
},
}, {
"speculative_config": {
"model": "JackFram/llama-160m",
"num_speculative_tokens": 6,
"disable_logprobs": 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["speculative_config"]
["disable_logprobs"])
@pytest.mark.parametrize(
"common_llm_kwargs",
[{
"model_name": "JackFram/llama-68m",
# 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", [{}])
@pytest.mark.parametrize(
"test_llm_kwargs",
[{
"speculative_config": {
"model": "JackFram/llama-160m",
"num_speculative_tokens": 3,
"disable_logprobs": False,
# Artificially limit the draft model max model len; this forces
# vLLM to skip speculation once the sequences grow beyond 32-k
# tokens.
"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["speculative_config"]
["disable_logprobs"])
@pytest.mark.parametrize(
"common_llm_kwargs",
[{
"model_name": "JackFram/llama-68m",
# 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", [{}])
@pytest.mark.parametrize("test_llm_kwargs", [{
"speculative_config": {
"model": "JackFram/llama-160m",
"num_speculative_tokens": 3,
"disable_logprobs": 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,
# 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", [{}])
@pytest.mark.parametrize("test_llm_kwargs", [{
"speculative_config": {
"model": "JackFram/llama-68m",
"num_speculative_tokens": 3,
"disable_logprobs": 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["speculative_config"]
["disable_logprobs"])
tests/spec_decode/e2e/test_medusa_correctness.py deleted 100644 → 0
View file @ 9ffe905a
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""This docstring details important information on the testing methodology.
Most of the tests rely on "greedy equality", where we expect the output of
speculative decoding on a sequence to exactly match the output of normal non-
speculative decoding.
Since speculative decoding with rejection sampling guarantees that the output
distribution matches the target model's output distribution (up to hardware
numerics, see https://arxiv.org/pdf/2302.01318.pdf), we can expect greedy
equality.
However, we still need to verify below scenario could be passed:
* Batch size 1 greedy equality
* Batch size >1 greedy equality
* Test greedy equality under preemption
* Test greedy equality under various number of speculative tokens.
With those tests, we can say at least, Medusa would not break the
correctness for the target model outputs.
"""
import pytest
from ..utils import maybe_enable_chunked_prefill
from .conftest import run_equality_correctness_test
# main model
# lmsys/vicuna-7b-v1.3 was to be used but it's causing
# OOM in CI pipeline, so using a smaller model.
MAIN_MODEL = "JackFram/llama-68m"
# speculative model
SPEC_MODEL = "abhigoyal/vllm-medusa-llama-68m-random"
# max number of speculative tokens: this corresponds to
# num_heads in the config.json of the speculator model.
MAX_SPEC_TOKENS = 5
# precision
PRECISION = "float32"
@pytest.mark.parametrize(
"common_llm_kwargs",
[{
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Print spec metrics.
"disable_log_stats": False,
# Precision
"dtype": PRECISION,
# Main model
"model_name": MAIN_MODEL,
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@pytest.mark.parametrize("test_llm_kwargs", [
{
"speculative_config": {
"model": SPEC_MODEL,
"num_speculative_tokens": MAX_SPEC_TOKENS,
},
},
])
@pytest.mark.parametrize("output_len", [
128,
])
@pytest.mark.parametrize("batch_size", [1, 32])
@pytest.mark.parametrize("seed", [1])
@pytest.mark.parametrize("prefill_chunk_size", [-1, 32])
def test_medusa_e2e_greedy_correctness(vllm_runner, common_llm_kwargs,
per_test_common_llm_kwargs,
baseline_llm_kwargs, test_llm_kwargs,
batch_size: int, output_len: int,
seed: int, prefill_chunk_size: int):
"""Verify greedy equality with different batch size."""
maybe_enable_chunked_prefill(prefill_chunk_size, test_llm_kwargs)
run_equality_correctness_test(vllm_runner,
common_llm_kwargs,
per_test_common_llm_kwargs,
baseline_llm_kwargs,
test_llm_kwargs,
batch_size,
max_output_len=output_len,
seed=seed,
temperature=0.0)
@pytest.mark.parametrize(
"common_llm_kwargs",
[{
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Print spec metrics.
"disable_log_stats": False,
# Precision
"dtype": PRECISION,
# Main model
"model_name": MAIN_MODEL,
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@pytest.mark.parametrize("test_llm_kwargs", [
{
"speculative_config": {
"model": SPEC_MODEL,
"num_speculative_tokens": MAX_SPEC_TOKENS,
"disable_logprobs": False,
},
},
{
"speculative_config": {
"model": SPEC_MODEL,
"num_speculative_tokens": MAX_SPEC_TOKENS,
"disable_logprobs": True,
},
},
])
@pytest.mark.parametrize("output_len", [
8,
])
@pytest.mark.parametrize("batch_size", [8])
@pytest.mark.parametrize("seed", [1])
@pytest.mark.parametrize("logprobs", [1, 6])
@pytest.mark.parametrize("prefill_chunk_size", [-1, 32])
def test_medusa_e2e_greedy_logprobs(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 greedy equality with different batch size."""
maybe_enable_chunked_prefill(prefill_chunk_size, test_llm_kwargs)
run_equality_correctness_test(
vllm_runner,
common_llm_kwargs,
per_test_common_llm_kwargs,
baseline_llm_kwargs,
test_llm_kwargs,
batch_size,
max_output_len=output_len,
seed=seed,
temperature=0.0,
logprobs=logprobs,
prompt_logprobs=logprobs,
disable_logprobs=test_llm_kwargs["speculative_config"]
["disable_logprobs"])
@pytest.mark.parametrize(
"common_llm_kwargs",
[{
"enforce_eager": False,
# Print spec metrics.
"disable_log_stats": False,
# Precision
"dtype": PRECISION,
# Main model
"model_name": MAIN_MODEL,
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@pytest.mark.parametrize("test_llm_kwargs", [
{
"speculative_config": {
"model": SPEC_MODEL,
"num_speculative_tokens": MAX_SPEC_TOKENS,
},
},
])
@pytest.mark.parametrize("output_len", [
128,
])
@pytest.mark.parametrize("batch_size", [1, 32])
@pytest.mark.parametrize("seed", [1])
@pytest.mark.parametrize("prefill_chunk_size", [-1, 32])
def test_medusa_e2e_greedy_correctness_cuda_graph(
vllm_runner, common_llm_kwargs, per_test_common_llm_kwargs,
baseline_llm_kwargs, test_llm_kwargs, batch_size: int, output_len: int,
seed: int, prefill_chunk_size: int):
"""Verify greedy equality with cuda graph enabled and different
batch sizes."""
maybe_enable_chunked_prefill(prefill_chunk_size, test_llm_kwargs)
run_equality_correctness_test(vllm_runner,
common_llm_kwargs,
per_test_common_llm_kwargs,
baseline_llm_kwargs,
test_llm_kwargs,
batch_size,
max_output_len=output_len,
seed=seed,
temperature=0.0)
@pytest.mark.parametrize(
"common_llm_kwargs",
[{
"block_size": 16,
# 2 for small prompt, 256//8 for generated.
"num_gpu_blocks_override": 2 + 256 // 8,
"max_model_len": (2 + 256 // 8) * 8,
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Precision
"dtype": PRECISION,
# Main model
"model_name": MAIN_MODEL,
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@pytest.mark.parametrize("test_llm_kwargs", [
{
"speculative_config": {
"model": SPEC_MODEL,
"num_speculative_tokens": MAX_SPEC_TOKENS,
},
},
])
@pytest.mark.parametrize(
"output_len",
[
# Use small output len for fast test.
128,
])
@pytest.mark.parametrize("batch_size", [4])
@pytest.mark.parametrize("seed", [1])
@pytest.mark.parametrize("prefill_chunk_size", [-1, 32])
def test_medusa_e2e_greedy_correctness_with_preemption(
vllm_runner, common_llm_kwargs, per_test_common_llm_kwargs,
baseline_llm_kwargs, test_llm_kwargs, batch_size: int, output_len: int,
seed: int, prefill_chunk_size: int):
"""Verify greedy equality, even when some sequences are preempted mid-
generation.
"""
maybe_enable_chunked_prefill(prefill_chunk_size, test_llm_kwargs)
run_equality_correctness_test(vllm_runner,
common_llm_kwargs,
per_test_common_llm_kwargs,
baseline_llm_kwargs,
test_llm_kwargs,
batch_size,
max_output_len=output_len,
seed=seed,
temperature=0.0)
@pytest.mark.parametrize(
"common_llm_kwargs",
[{
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Precision
"dtype": PRECISION,
# Main model
"model_name": MAIN_MODEL,
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@pytest.mark.parametrize(
"test_llm_kwargs",
[
{
"speculative_config": {
"model": SPEC_MODEL,
"num_speculative_tokens": k,
},
}
# Try a range of num. speculative tokens
for k in range(1, 1 + MAX_SPEC_TOKENS)
])
@pytest.mark.parametrize("batch_size", [2])
@pytest.mark.parametrize(
"output_len",
[
# Use smaller output len for fast test.
32,
])
@pytest.mark.parametrize("seed", [1])
@pytest.mark.parametrize("prefill_chunk_size", [-1, 32])
def test_medusa_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, prefill_chunk_size: int):
"""Verify that medusa speculative decoding produces exact equality
to without spec decode with different values of num_speculative_tokens.
"""
maybe_enable_chunked_prefill(prefill_chunk_size, test_llm_kwargs)
run_equality_correctness_test(vllm_runner,
common_llm_kwargs,
per_test_common_llm_kwargs,
baseline_llm_kwargs,
test_llm_kwargs,
batch_size,
max_output_len=output_len,
seed=seed,
temperature=0.0)
@pytest.mark.parametrize(
"common_llm_kwargs",
[{
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Precision
"dtype": PRECISION,
# Main model
"model_name": MAIN_MODEL,
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@pytest.mark.parametrize("test_llm_kwargs", [{
"speculative_config": {
"model": SPEC_MODEL,
"num_speculative_tokens": MAX_SPEC_TOKENS,
"disable_by_batch_size": 4,
},
}])
@pytest.mark.parametrize("batch_size", [1, 5])
@pytest.mark.parametrize(
"output_len",
[
# Use smaller output len for fast test.
32,
])
@pytest.mark.parametrize("seed", [1])
@pytest.mark.parametrize("prefill_chunk_size", [-1, 32])
def test_medusa_disable_queue(vllm_runner, common_llm_kwargs,
per_test_common_llm_kwargs, baseline_llm_kwargs,
test_llm_kwargs, batch_size: int,
output_len: int, seed: int,
prefill_chunk_size: int):
"""Verify that medusa speculative decoding produces exact equality
to without spec decode when speculation is disabled for large
batch sizes.
"""
maybe_enable_chunked_prefill(prefill_chunk_size, test_llm_kwargs)
run_equality_correctness_test(vllm_runner,
common_llm_kwargs,
per_test_common_llm_kwargs,
baseline_llm_kwargs,
test_llm_kwargs,
batch_size,
max_output_len=output_len,
seed=seed,
temperature=0.0)
@pytest.mark.parametrize(
"common_llm_kwargs",
[{
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Precision
"dtype": PRECISION,
# Main model
"model_name": MAIN_MODEL,
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@pytest.mark.parametrize("test_llm_kwargs", [{
"speculative_config": {
"model": SPEC_MODEL,
"num_speculative_tokens": MAX_SPEC_TOKENS,
"disable_by_batch_size": 4,
"disable_mqa_scorer": True,
},
}])
@pytest.mark.parametrize("batch_size", [1, 5])
@pytest.mark.parametrize(
"output_len",
[
# Use smaller output len for fast test.
32,
])
@pytest.mark.parametrize("seed", [1])
@pytest.mark.parametrize("prefill_chunk_size", [-1, 32])
def test_mqa_scorer(vllm_runner, common_llm_kwargs, per_test_common_llm_kwargs,
baseline_llm_kwargs, test_llm_kwargs, batch_size: int,
output_len: int, seed: int, prefill_chunk_size: int):
"""Verify that speculative decoding generates the same output
with batch expansion scorer and mqa scorer.
"""
maybe_enable_chunked_prefill(prefill_chunk_size, test_llm_kwargs)
run_equality_correctness_test(vllm_runner,
common_llm_kwargs,
per_test_common_llm_kwargs,
baseline_llm_kwargs,
test_llm_kwargs,
batch_size,
max_output_len=output_len,
seed=seed,
temperature=0.0)
if __name__ == "__main__":
import pytest
pytest.main([__file__])
tests/spec_decode/e2e/test_mlp_correctness.py deleted 100644 → 0
View file @ 9ffe905a
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""This docstring details important information on the testing methodology.
Most of the tests rely on "greedy equality", where we expect the output of
speculative decoding on a sequence to exactly match the output of normal non-
speculative decoding.
Since speculative decoding with rejection sampling guarantees that the output
distribution matches the target model's output distribution (up to hardware
numerics, see https://arxiv.org/pdf/2302.01318.pdf), we can expect greedy
equality.
However, we still need to verify below scenario could be passed:
* Batch size 1 greedy equality
* Batch size >1 greedy equality
* Test greedy equality under preemption
* Test greedy equality under various number of speculative tokens.
With those tests, we can say at least, MLPSpeculator would not break the
correctness for the target model outputs.
"""
from unittest.mock import patch
import pytest
from vllm.model_executor.layers.vocab_parallel_embedding import pad_vocab_size
from ..utils import maybe_enable_chunked_prefill
from .conftest import run_equality_correctness_test
# main model
MAIN_MODEL = "JackFram/llama-160m"
# speculative model
SPEC_MODEL = "ibm-ai-platform/llama-160m-accelerator"
# max. number of speculative tokens: this corresponds to
# n_predict in the config.json of the speculator model.
MAX_SPEC_TOKENS = 3
# precision
PRECISION = "float32"
@pytest.mark.parametrize(
"common_llm_kwargs",
[{
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Print spec metrics.
"disable_log_stats": False,
# Precision
"dtype": PRECISION,
# Main model
"model_name": MAIN_MODEL,
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@pytest.mark.parametrize("test_llm_kwargs", [
{
"speculative_config": {
"model": SPEC_MODEL,
},
},
])
@pytest.mark.parametrize("output_len", [
128,
])
@pytest.mark.parametrize("batch_size", [4, 32])
@pytest.mark.parametrize("seed", [1])
@pytest.mark.parametrize("prefill_chunk_size", [-1, 32])
def test_mlp_e2e_greedy_correctness(vllm_runner, common_llm_kwargs,
per_test_common_llm_kwargs,
baseline_llm_kwargs, test_llm_kwargs,
batch_size: int, output_len: int,
seed: int, prefill_chunk_size: int):
"""Verify greedy equality with different batch size."""
maybe_enable_chunked_prefill(prefill_chunk_size, test_llm_kwargs)
run_equality_correctness_test(vllm_runner,
common_llm_kwargs,
per_test_common_llm_kwargs,
baseline_llm_kwargs,
test_llm_kwargs,
batch_size,
max_output_len=output_len,
seed=seed,
temperature=0.0)
@pytest.mark.parametrize(
"common_llm_kwargs",
[{
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Print spec metrics.
"disable_log_stats": False,
# Precision
"dtype": PRECISION,
# Main model
"model_name": MAIN_MODEL,
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@pytest.mark.parametrize("test_llm_kwargs", [
{
"speculative_config": {
"model": SPEC_MODEL,
"disable_logprobs": False,
},
},
{
"speculative_config": {
"model": SPEC_MODEL,
"disable_logprobs": True,
},
},
])
@pytest.mark.parametrize("output_len", [8])
@pytest.mark.parametrize("batch_size", [8])
@pytest.mark.parametrize("seed", [1])
@pytest.mark.parametrize("logprobs", [1, 6])
@pytest.mark.parametrize("prefill_chunk_size", [-1, 4])
def test_mlp_e2e_greedy_logprobs(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 greedy equality with different batch size."""
maybe_enable_chunked_prefill(prefill_chunk_size, test_llm_kwargs)
# NOTE Test is sensitive enough st if we don't enable chunked prefill
# scheduling on baseline too, we get slightly different logprobs, ending
# up sampling different tokens at the tail (ie top tokens don't change).
# TL;DR: sd+cp == org+cp but sd+cp != org..is this expected?
maybe_enable_chunked_prefill(prefill_chunk_size, baseline_llm_kwargs)
run_equality_correctness_test(
vllm_runner,
common_llm_kwargs,
per_test_common_llm_kwargs,
baseline_llm_kwargs,
test_llm_kwargs,
batch_size,
max_output_len=output_len,
seed=seed,
temperature=0.0,
logprobs=logprobs,
prompt_logprobs=logprobs,
disable_logprobs=test_llm_kwargs["speculative_config"]
["disable_logprobs"])
@pytest.mark.parametrize(
"common_llm_kwargs",
[{
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Print spec metrics.
"disable_log_stats": False,
# Precision
"dtype": PRECISION,
# Main model
"model_name": MAIN_MODEL,
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@pytest.mark.parametrize("test_llm_kwargs", [
{
"speculative_config": {
"model": SPEC_MODEL,
},
},
])
@pytest.mark.parametrize("output_len", [2048])
@pytest.mark.parametrize("batch_size", [1, 32])
@pytest.mark.parametrize("seed", [1])
@pytest.mark.parametrize("prefill_chunk_size", [-1, 4])
def test_mlp_e2e_acceptance_rate(vllm_runner, common_llm_kwargs,
per_test_common_llm_kwargs,
baseline_llm_kwargs, test_llm_kwargs,
batch_size: int, output_len: int,
prefill_chunk_size: int, seed: int):
"""Verify acceptance rate with different batch size and large output
length."""
maybe_enable_chunked_prefill(prefill_chunk_size, test_llm_kwargs)
run_equality_correctness_test(vllm_runner,
common_llm_kwargs,
per_test_common_llm_kwargs,
baseline_llm_kwargs,
test_llm_kwargs,
batch_size,
max_output_len=output_len,
temperature=0.0,
seed=seed,
expected_acceptance_rate=0.48)
@pytest.mark.parametrize(
"common_llm_kwargs",
[{
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Print spec metrics.
"disable_log_stats": False,
# Precision
"dtype": PRECISION,
# Main model
"model_name": MAIN_MODEL,
# Speculative config
"speculative_config": {
"model": SPEC_MODEL,
},
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{"seed": 1}])
@pytest.mark.parametrize("test_llm_kwargs", [{"seed": 5}])
@pytest.mark.parametrize("output_len", [64])
@pytest.mark.parametrize("batch_size", [1, 32])
@pytest.mark.parametrize("temperature", [1.0])
@pytest.mark.parametrize("prefill_chunk_size", [-1, 4])
@pytest.mark.parametrize("seed", [1])
def test_mlp_e2e_seeded_correctness(vllm_runner, common_llm_kwargs,
per_test_common_llm_kwargs,
baseline_llm_kwargs, test_llm_kwargs,
batch_size: int, output_len: int,
temperature: float,
prefill_chunk_size: int, seed: int):
"""Verify seeded runs produce the same output."""
maybe_enable_chunked_prefill(prefill_chunk_size, test_llm_kwargs)
maybe_enable_chunked_prefill(prefill_chunk_size, baseline_llm_kwargs)
run_equality_correctness_test(vllm_runner,
common_llm_kwargs,
per_test_common_llm_kwargs,
baseline_llm_kwargs,
test_llm_kwargs,
batch_size,
max_output_len=output_len,
temperature=temperature,
seed=seed)
# Ensure this same test does fail if we _don't_ include per-request seeds
with pytest.raises(AssertionError):
run_equality_correctness_test(vllm_runner,
common_llm_kwargs,
per_test_common_llm_kwargs,
baseline_llm_kwargs,
test_llm_kwargs,
batch_size,
max_output_len=output_len,
temperature=temperature,
seed=seed,
disable_seed=True)
@pytest.mark.parametrize(
"common_llm_kwargs",
[{
"block_size": 16,
# 2 for small prompt, 256//8 for generated.
"num_gpu_blocks_override": 2 + 256 // 8,
"max_model_len": (2 + 256 // 8) * 8,
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Precision
"dtype": PRECISION,
# Main model
"model_name": MAIN_MODEL,
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@pytest.mark.parametrize("test_llm_kwargs", [
{
"speculative_config": {
"model": SPEC_MODEL,
},
},
])
@pytest.mark.parametrize(
"output_len",
[
# Use small output len for fast test.
128,
])
@pytest.mark.parametrize("batch_size", [4])
@pytest.mark.parametrize("prefill_chunk_size", [-1, 4])
@pytest.mark.parametrize("seed", [1])
def test_mlp_e2e_greedy_correctness_with_preemption(
vllm_runner, common_llm_kwargs, per_test_common_llm_kwargs,
baseline_llm_kwargs, test_llm_kwargs, batch_size: int, output_len: int,
prefill_chunk_size: int, seed: int):
"""Verify greedy equality, even when some sequences are preempted mid-
generation.
"""
maybe_enable_chunked_prefill(prefill_chunk_size, test_llm_kwargs)
run_equality_correctness_test(vllm_runner,
common_llm_kwargs,
per_test_common_llm_kwargs,
baseline_llm_kwargs,
test_llm_kwargs,
batch_size,
max_output_len=output_len,
seed=seed,
temperature=0.0)
@pytest.mark.parametrize(
"common_llm_kwargs",
[{
"block_size": 16,
# 2 for small prompt, 256//8 for generated.
"num_gpu_blocks_override": 2 + 256 // 8,
"max_model_len": (2 + 256 // 8) * 8,
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Precision
"dtype": PRECISION,
# Main model
"model_name": MAIN_MODEL,
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@pytest.mark.parametrize("test_llm_kwargs", [
{
"speculative_config": {
"model": SPEC_MODEL,
},
},
])
@pytest.mark.parametrize(
"output_len",
[
# Use small output len for fast test.
128,
])
@pytest.mark.parametrize("batch_size", [4])
@pytest.mark.parametrize("seed", [1])
@pytest.mark.parametrize("prefill_chunk_size", [-1, 4])
def test_mlp_e2e_greedy_correctness_with_padding(
vllm_runner, common_llm_kwargs, per_test_common_llm_kwargs,
baseline_llm_kwargs, test_llm_kwargs, batch_size: int, output_len: int,
prefill_chunk_size: int, seed: int):
"""Verify greedy equality when the vocab dimension is padded
"""
maybe_enable_chunked_prefill(prefill_chunk_size, test_llm_kwargs)
# Default pad_to is 64, test model has vocab_size of 32000
def patched_pad_vocab_size(vocab_size, pad_to=None):
return pad_vocab_size(vocab_size, pad_to=32064)
with patch(
"vllm.model_executor.layers.vocab_parallel_embedding.pad_vocab_size",
patched_pad_vocab_size):
run_equality_correctness_test(vllm_runner,
common_llm_kwargs,
per_test_common_llm_kwargs,
baseline_llm_kwargs,
test_llm_kwargs,
batch_size,
max_output_len=output_len,
seed=seed,
temperature=0.0)
@pytest.mark.parametrize(
"common_llm_kwargs",
[{
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Precision
"dtype": PRECISION,
# Main model
"model_name": MAIN_MODEL,
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@pytest.mark.parametrize(
"test_llm_kwargs",
[
{
"speculative_config": {
"model": SPEC_MODEL,
"num_speculative_tokens": k,
},
}
# Try a range of num. speculative tokens
for k in range(1, 1 + MAX_SPEC_TOKENS)
])
@pytest.mark.parametrize("batch_size", [2])
@pytest.mark.parametrize(
"output_len",
[
# Use smaller output len for fast test.
32,
])
@pytest.mark.parametrize("prefill_chunk_size", [-1, 4])
@pytest.mark.parametrize("seed", [1])
def test_mlp_different_k(vllm_runner, common_llm_kwargs,
per_test_common_llm_kwargs, baseline_llm_kwargs,
test_llm_kwargs, batch_size: int,
prefill_chunk_size: int, seed: int, output_len: int):
"""Verify that mlp speculative decoding produces exact equality
to without spec decode with different values of num_speculative_tokens.
"""
maybe_enable_chunked_prefill(prefill_chunk_size, test_llm_kwargs)
run_equality_correctness_test(vllm_runner,
common_llm_kwargs,
per_test_common_llm_kwargs,
baseline_llm_kwargs,
test_llm_kwargs,
batch_size,
max_output_len=output_len,
seed=seed,
temperature=0.0)
@pytest.mark.parametrize(
"common_llm_kwargs",
[{
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Precision
"dtype": PRECISION,
# Main model
"model_name": MAIN_MODEL,
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@pytest.mark.parametrize("test_llm_kwargs", [{
"speculative_config": {
"model": SPEC_MODEL,
"disable_by_batch_size": 4,
},
}])
@pytest.mark.parametrize("batch_size", [1, 5])
@pytest.mark.parametrize(
"output_len",
[
# Use smaller output len for fast test.
32,
])
# Speculative decoding is disabled when sequences reach decoding and the batch
# consists of single-token requests. Hence we set `max_num_seqs`
# >= `speculative_disable_by_batch_size` to test feature interaction.
@pytest.mark.parametrize("prefill_chunk_size", [-1, 4])
@pytest.mark.parametrize("seed", [1])
def test_mlp_disable_queue(vllm_runner, common_llm_kwargs,
per_test_common_llm_kwargs, baseline_llm_kwargs,
test_llm_kwargs, batch_size: int,
prefill_chunk_size: int, seed: int,
output_len: int):
"""Verify that mlp speculative decoding produces exact equality
to without spec decode when speculation is disabled for large
batch sizes.
"""
maybe_enable_chunked_prefill(prefill_chunk_size, test_llm_kwargs)
run_equality_correctness_test(vllm_runner,
common_llm_kwargs,
per_test_common_llm_kwargs,
baseline_llm_kwargs,
test_llm_kwargs,
batch_size,
max_output_len=output_len,
seed=seed,
temperature=0.0)
@pytest.mark.parametrize(
"common_llm_kwargs",
[{
"model_name": MAIN_MODEL,
# 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", [{}])
@pytest.mark.parametrize("test_llm_kwargs", [{
"speculative_config": {
"model": SPEC_MODEL,
"disable_mqa_scorer": True,
},
}])
@pytest.mark.parametrize("batch_size", [1, 5])
@pytest.mark.parametrize(
"output_len",
[
# Use smaller output len for fast test.
32,
])
@pytest.mark.parametrize("prefill_chunk_size", [-1, 4])
@pytest.mark.parametrize("seed", [1])
def test_mqa_scorer(vllm_runner, common_llm_kwargs, per_test_common_llm_kwargs,
baseline_llm_kwargs, test_llm_kwargs, batch_size: int,
output_len: int, prefill_chunk_size: int, seed: int):
"""Verify that speculative decoding generates the same output
with batch expansion scorer and mqa scorer.
"""
maybe_enable_chunked_prefill(prefill_chunk_size, test_llm_kwargs)
run_equality_correctness_test(vllm_runner,
common_llm_kwargs,
per_test_common_llm_kwargs,
baseline_llm_kwargs,
test_llm_kwargs,
batch_size,
max_output_len=output_len,
seed=seed,
temperature=0.0)
tests/spec_decode/e2e/test_mtp_correctness.py deleted 100644 → 0
View file @ 9ffe905a
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""This docstring details important information on the testing methodology.
Most of the tests rely on "greedy equality", where we expect the output of
speculative decoding on a sequence to exactly match the output of normal non-
speculative decoding.
Since speculative decoding with rejection sampling guarantees that the output
distribution matches the target model's output distribution (up to hardware
numerics, see https://arxiv.org/pdf/2302.01318.pdf), we can expect greedy
equality.
However, we still need to verify below scenario could be passed:
* Batch size 1 greedy equality
* Batch size >1 greedy equality
* Test greedy equality under preemption
* Test greedy equality under various number of speculative tokens.
With those tests, we can say at least, mtp would not break the
correctness for the target model outputs.
"""
import pytest
from .conftest import run_equality_correctness_test
# main model
MAIN_MODEL = "luccafong/deepseek_mtp_main_random"
# max. number of speculative tokens: this corresponds to
# num_nextn_predict_layers in the config.json of the speculator model.
MAX_SPEC_TOKENS = 1
# precision
PRECISION = "bfloat16"
@pytest.mark.parametrize(
"common_llm_kwargs",
[{
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Print spec metrics.
"disable_log_stats": False,
# Precision
"dtype": PRECISION,
# Main model
"model_name": MAIN_MODEL,
# GPU memory utilization
"gpu_memory_utilization": 0.85
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@pytest.mark.parametrize("test_llm_kwargs", [
{
"speculative_config": {
"num_speculative_tokens": MAX_SPEC_TOKENS,
},
},
])
@pytest.mark.parametrize("output_len", [
128,
])
@pytest.mark.parametrize("batch_size", [1, 32])
@pytest.mark.parametrize("seed", [1])
def test_mtp_e2e_greedy_correctness(vllm_runner, common_llm_kwargs,
per_test_common_llm_kwargs,
baseline_llm_kwargs, test_llm_kwargs,
batch_size: int, output_len: int,
seed: int):
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)
@pytest.mark.parametrize(
"common_llm_kwargs",
[{
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Print spec metrics.
"disable_log_stats": False,
# Precision
"dtype": PRECISION,
# Main model
"model_name": MAIN_MODEL,
# GPU memory utilization
"gpu_memory_utilization": 0.85
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@pytest.mark.parametrize("test_llm_kwargs", [
{
"speculative_config": {
"num_speculative_tokens": MAX_SPEC_TOKENS,
"disable_logprobs": False,
},
},
{
"speculative_config": {
"num_speculative_tokens": MAX_SPEC_TOKENS,
"disable_logprobs": True,
},
},
])
@pytest.mark.parametrize("output_len", [
128,
])
@pytest.mark.parametrize("batch_size", [8])
@pytest.mark.parametrize("seed", [1])
@pytest.mark.parametrize("logprobs", [1, 6])
def test_mtp_e2e_greedy_logprobs(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):
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,
logprobs=logprobs,
prompt_logprobs=logprobs,
disable_logprobs=test_llm_kwargs["speculative_config"]
["disable_logprobs"])
@pytest.mark.parametrize(
"common_llm_kwargs",
[{
"enforce_eager": False,
# Print spec metrics.
"disable_log_stats": False,
# Precision
"dtype": PRECISION,
# Main model
"model_name": MAIN_MODEL,
"gpu_memory_utilization": 0.85
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@pytest.mark.parametrize("test_llm_kwargs", [
{
"speculative_config": {
"num_speculative_tokens": MAX_SPEC_TOKENS,
},
},
])
@pytest.mark.parametrize("output_len", [
128,
])
@pytest.mark.parametrize("batch_size", [1, 32])
@pytest.mark.parametrize("seed", [1])
def test_mtp_e2e_greedy_correctness_cuda_graph(vllm_runner, common_llm_kwargs,
per_test_common_llm_kwargs,
baseline_llm_kwargs,
test_llm_kwargs,
batch_size: int,
output_len: int, seed: int):
"""Verify greedy equality with cuda graph enabled and different
batch sizes."""
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)
@pytest.mark.parametrize(
"common_llm_kwargs",
[{
"block_size": 8,
# 2 for small prompt, 256//8 for generated.
"num_gpu_blocks_override": 2 + 256 // 8,
"max_model_len": (2 + 256 // 8) * 8,
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Precision
"dtype": PRECISION,
# Main model
"model_name": MAIN_MODEL,
# GPU memory utilization
"gpu_memory_utilization": 0.9
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@pytest.mark.parametrize("test_llm_kwargs", [
{
"speculative_config": {
"num_speculative_tokens": MAX_SPEC_TOKENS,
},
},
])
@pytest.mark.parametrize(
"output_len",
[
# Use small output len for fast test.
128,
])
@pytest.mark.parametrize("batch_size", [4])
@pytest.mark.parametrize("seed", [1])
def test_mtp_e2e_greedy_correctness_with_preemption(
vllm_runner, common_llm_kwargs, per_test_common_llm_kwargs,
baseline_llm_kwargs, test_llm_kwargs, batch_size: int, output_len: int,
seed: int):
"""Verify greedy equality, even when some sequences are preempted mid-
generation.
"""
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)
@pytest.mark.parametrize(
"common_llm_kwargs",
[{
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Precision
"dtype": PRECISION,
# Main model
"model_name": MAIN_MODEL,
# GPU memory utilization
"gpu_memory_utilization": 0.9
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@pytest.mark.parametrize(
"test_llm_kwargs",
[
{
"speculative_config": {
"num_speculative_tokens": k,
},
}
# Try a range of num. speculative tokens
for k in range(1, 1 + MAX_SPEC_TOKENS)
])
@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_mtp_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):
"""Verify that mtp speculative decoding produces exact equality
to without spec decode with different values of num_speculative_tokens.
"""
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)
@pytest.mark.parametrize(
"common_llm_kwargs",
[{
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Precision
"dtype": PRECISION,
# Main model
"model_name": MAIN_MODEL,
# GPU memory utilization
"gpu_memory_utilization": 0.9
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@pytest.mark.parametrize("test_llm_kwargs", [{
"speculative_config": {
"num_speculative_tokens": MAX_SPEC_TOKENS,
"disable_by_batch_size": 4
},
}])
@pytest.mark.parametrize("batch_size", [1, 5])
@pytest.mark.parametrize(
"output_len",
[
# Use smaller output len for fast test.
32,
])
@pytest.mark.parametrize("seed", [1])
def test_mtp_disable_queue(vllm_runner, common_llm_kwargs,
per_test_common_llm_kwargs, baseline_llm_kwargs,
test_llm_kwargs, batch_size: int, output_len: int,
seed: int):
"""Verify that mtp speculative decoding produces exact equality
to without spec decode when speculation is disabled for large
batch sizes.
"""
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)
if __name__ == "__main__":
import pytest
pytest.main([__file__])
tests/spec_decode/e2e/test_multistep_correctness.py deleted 100644 → 0
View file @ 9ffe905a
This diff is collapsed.
tests/spec_decode/e2e/test_ngram_correctness.py deleted 100644 → 0
View file @ 9ffe905a
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""This docstring details important information on the testing methodology.
Most of the tests rely on "greedy equality", where we expect the output of
speculative decoding on a sequence to exactly match the output of normal non-
speculative decoding.
Since speculative decoding with rejection sampling guarantees that the output
distribution matches the target model's output distribution (up to hardware
numerics, see https://arxiv.org/pdf/2302.01318.pdf), we can expect greedy
equality.
For ngram lookup, its idea comes from https://github.com/apoorvumang/prompt-lookup-decoding,
and is merged into transform code base: https://github.com/huggingface/transformers/pull/27775.
Since there is no model is needed for generate the proposal, we could make
the testcase much simpler than drafter multi-step one.
However, we still need to verify below scenario could be passed:
* Batch size 1 greedy equality
* Batch size >1 greedy equality
* Test greedy equality under preemption
* Test greedy equality under various ngram sizes / speculative sizes
With those tests, we can say at least, ngram spec would not break the
correctness for the target model outputs.
"""
import pytest
from ..utils import maybe_enable_chunked_prefill
from .conftest import run_equality_correctness_test
@pytest.mark.parametrize(
"common_llm_kwargs",
[{
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# 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", [
{
"model_name": "JackFram/llama-68m",
},
])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@pytest.mark.parametrize("test_llm_kwargs", [
{
"speculative_config": {
"method": "ngram",
"num_speculative_tokens": 5,
"prompt_lookup_max": 3,
"disable_mqa_scorer": False,
},
},
{
"speculative_config": {
"method": "ngram",
"num_speculative_tokens": 5,
"prompt_lookup_max": 3,
"disable_mqa_scorer": True,
},
},
])
@pytest.mark.parametrize("output_len", [
256,
])
@pytest.mark.parametrize("batch_size", [1, 32])
@pytest.mark.parametrize("prefill_chunk_size", [-1, 4])
@pytest.mark.parametrize("seed", [1])
def test_ngram_e2e_greedy_correctness(vllm_runner, common_llm_kwargs,
per_test_common_llm_kwargs,
baseline_llm_kwargs, test_llm_kwargs,
batch_size: int, output_len: int,
prefill_chunk_size: int, seed: int):
"""Verify greedy equality on a tiny model with different batch size."""
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,
max_output_len=output_len,
seed=seed,
temperature=0.0)
@pytest.mark.parametrize(
"common_llm_kwargs",
[{
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# 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", [
{
"model_name": "JackFram/llama-68m",
},
])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@pytest.mark.parametrize("test_llm_kwargs", [
{
"speculative_config": {
"method": "ngram",
"num_speculative_tokens": 5,
"prompt_lookup_max": 3,
"disable_logprobs": False,
},
},
{
"speculative_config": {
"method": "ngram",
"num_speculative_tokens": 5,
"prompt_lookup_max": 3,
"disable_logprobs": True,
},
},
])
@pytest.mark.parametrize("output_len", [
8,
])
@pytest.mark.parametrize("batch_size", [8])
@pytest.mark.parametrize("seed", [1])
@pytest.mark.parametrize("logprobs", [1, 6])
def test_ngram_e2e_greedy_logprobs(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 greedy equality on a tiny model with different batch size."""
run_equality_correctness_test(
vllm_runner,
common_llm_kwargs,
per_test_common_llm_kwargs,
baseline_llm_kwargs,
test_llm_kwargs,
batch_size,
max_output_len=output_len,
seed=seed,
temperature=0.0,
logprobs=logprobs,
prompt_logprobs=logprobs,
disable_logprobs=test_llm_kwargs["speculative_config"]
["disable_logprobs"])
@pytest.mark.parametrize(
"common_llm_kwargs",
[{
"block_size": 16,
# 2 for small prompt, 256//8 for generated.
"num_gpu_blocks_override": 2 + 256 // 8,
"max_model_len": (2 + 256 // 8) * 8,
# 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", [
{
"model_name": "JackFram/llama-160m",
},
])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@pytest.mark.parametrize("test_llm_kwargs", [
{
"speculative_config": {
"method": "ngram",
"num_speculative_tokens": 5,
"prompt_lookup_max": 3,
},
"enable_chunked_prefill": False,
},
{
"speculative_config": {
"method": "ngram",
"num_speculative_tokens": 5,
"prompt_lookup_max": 3,
"disable_mqa_scorer": True,
},
"enable_chunked_prefill": True,
"max_num_batched_tokens": 4,
"max_num_seqs": 4
},
])
@pytest.mark.parametrize(
"output_len",
[
# Use small output len for fast test.
256,
])
@pytest.mark.parametrize("batch_size", [4])
@pytest.mark.parametrize("seed", [1])
def test_ngram_e2e_greedy_correctness_with_preemption(
vllm_runner, common_llm_kwargs, per_test_common_llm_kwargs,
baseline_llm_kwargs, test_llm_kwargs, batch_size: int, output_len: int,
seed: int):
"""Verify greedy equality, even when some sequences are preempted mid-
generation.
"""
run_equality_correctness_test(vllm_runner,
common_llm_kwargs,
per_test_common_llm_kwargs,
baseline_llm_kwargs,
test_llm_kwargs,
batch_size,
max_output_len=output_len,
temperature=0,
seed=seed)
@pytest.mark.parametrize(
"common_llm_kwargs",
[{
"model_name": "JackFram/llama-68m",
# 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", [{}])
@pytest.mark.parametrize(
"test_llm_kwargs",
[
{
"speculative_config": {
"method": "ngram",
"num_speculative_tokens": k,
"prompt_lookup_max": 3,
},
}
# Try a range of common k, as well as large speculation.
for k in [1, 3, 5]
] + [
{
"speculative_config": {
"method": "ngram",
"num_speculative_tokens": k,
"prompt_lookup_max": 1,
},
}
# Try a range of common k, as well as large speculation.
for k in [1, 3, 5]
])
@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_ngram_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):
"""Verify that ngram speculative decoding produces exact equality
to without spec decode with many different values of k and
different ngram prompt_lookup_max.
"""
run_equality_correctness_test(vllm_runner,
common_llm_kwargs,
per_test_common_llm_kwargs,
baseline_llm_kwargs,
test_llm_kwargs,
batch_size,
max_output_len=output_len,
seed=seed,
temperature=0.0)
@pytest.mark.parametrize(
"common_llm_kwargs",
[{
"model_name": "JackFram/llama-68m",
# 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", [{}])
@pytest.mark.parametrize("test_llm_kwargs", [{
"speculative_config": {
"method": "ngram",
"num_speculative_tokens": 5,
"prompt_lookup_max": 3,
"disable_by_batch_size": 4
},
}, {
"speculative_config": {
"method": "ngram",
"num_speculative_tokens": 5,
"prompt_lookup_max": 3,
"disable_by_batch_size": 4,
"disable_mqa_scorer": True,
},
"enable_chunked_prefill": True,
"max_num_batched_tokens": 4,
"max_num_seqs": 4
}])
@pytest.mark.parametrize("batch_size", [1, 5])
@pytest.mark.parametrize(
"output_len",
[
# Use smaller output len for fast test.
32,
])
@pytest.mark.parametrize("seed", [1])
def test_ngram_disable_queue(vllm_runner, common_llm_kwargs,
per_test_common_llm_kwargs, baseline_llm_kwargs,
test_llm_kwargs, batch_size: int, output_len: int,
seed: int):
"""Verify that ngram speculative decoding produces exact equality
to without spec decode with many different values of k and
different ngram prompt_lookup_max.
"""
run_equality_correctness_test(vllm_runner,
common_llm_kwargs,
per_test_common_llm_kwargs,
baseline_llm_kwargs,
test_llm_kwargs,
batch_size,
max_output_len=output_len,
seed=seed,
temperature=0.0)
@pytest.mark.parametrize(
"common_llm_kwargs",
[{
"model_name": "JackFram/llama-68m",
# 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", [{}])
@pytest.mark.parametrize("test_llm_kwargs", [{
"speculative_config": {
"method": "ngram",
"num_speculative_tokens": 5,
"prompt_lookup_max": 3,
"disable_mqa_scorer": True,
},
}])
@pytest.mark.parametrize("batch_size", [1, 5])
@pytest.mark.parametrize(
"output_len",
[
# Use smaller output len for fast test.
32,
])
@pytest.mark.parametrize("seed", [1])
def test_ngram_scorer(vllm_runner, common_llm_kwargs,
per_test_common_llm_kwargs, baseline_llm_kwargs,
test_llm_kwargs, batch_size: int, output_len: int,
seed: int):
"""Verify that ngram speculative decoding generates the same output
with batch expansion scorer and mqa scorer.
"""
run_equality_correctness_test(vllm_runner,
common_llm_kwargs,
per_test_common_llm_kwargs,
baseline_llm_kwargs,
test_llm_kwargs,
batch_size,
max_output_len=output_len,
seed=seed,
temperature=0.0)
tests/spec_decode/e2e/test_seed.py deleted 100644 → 0
View file @ 9ffe905a
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import pytest
from .conftest import run_equality_correctness_test
# main model
MAIN_MODEL = "JackFram/llama-68m"
# speculative model
SPEC_MODEL = "JackFram/llama-160m"
@pytest.mark.parametrize(
"common_llm_kwargs",
[{
"model_name": "JackFram/llama-68m",
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# speculative config
"speculative_config": {
"model": "JackFram/llama-160m",
"num_speculative_tokens": 3,
},
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{"seed": 1}])
@pytest.mark.parametrize("test_llm_kwargs", [{"seed": 5}])
@pytest.mark.parametrize("batch_size", [1, 8, 32])
@pytest.mark.parametrize("temperature", [0.1, 1.0])
@pytest.mark.parametrize(
"output_len",
[
# Use smaller output len for fast test.
20,
])
def test_seeded_consistency(vllm_runner, common_llm_kwargs,
per_test_common_llm_kwargs, baseline_llm_kwargs,
test_llm_kwargs, batch_size: int,
temperature: float, output_len: int):
"""Verify outputs are consistent across multiple runs with same seed
"""
run_equality_correctness_test(
vllm_runner,
common_llm_kwargs,
per_test_common_llm_kwargs,
baseline_llm_kwargs,
test_llm_kwargs,
batch_size,
max_output_len=output_len,
temperature=temperature,
disable_seed=False,
)
# Ensure this same test does fail if we _don't_ include per-request seeds
with pytest.raises(AssertionError):
run_equality_correctness_test(
vllm_runner,
common_llm_kwargs,
per_test_common_llm_kwargs,
baseline_llm_kwargs,
test_llm_kwargs,
batch_size,
max_output_len=output_len,
temperature=temperature,
disable_seed=True,
)
tests/spec_decode/test_batch_expansion.py deleted 100644 → 0
View file @ 9ffe905a
This diff is collapsed.
tests/spec_decode/test_dynamic_spec_decode.py deleted 100644 → 0
View file @ 9ffe905a
This diff is collapsed.
tests/spec_decode/test_memory_usage.py deleted 100644 → 0
View file @ 9ffe905a
This diff is collapsed.
tests/spec_decode/test_metrics.py deleted 100644 → 0
View file @ 9ffe905a
This diff is collapsed.
tests/spec_decode/test_multi_step_worker.py deleted 100644 → 0
View file @ 9ffe905a
This diff is collapsed.
tests/spec_decode/test_ngram_worker.py deleted 100644 → 0
View file @ 9ffe905a
This diff is collapsed.
tests/spec_decode/test_scorer.py deleted 100644 → 0
View file @ 9ffe905a
This diff is collapsed.
tests/spec_decode/test_spec_decode_worker.py deleted 100644 → 0
View file @ 9ffe905a
This diff is collapsed.
tests/spec_decode/test_utils.py deleted 100644 → 0
View file @ 9ffe905a
This diff is collapsed.
tests/spec_decode/utils.py deleted 100644 → 0
View file @ 9ffe905a
This diff is collapsed.
tests/test_sequence.py
View file @ dd572c0a
......@@ -29,7 +29,6 @@ def test_sampler_output_initialization(sampler_output, sample_outputs):
assert len(sampler_output) == len(sample_outputs)
assert sampler_output.sampled_token_probs is None
assert sampler_output.sampled_token_ids is None
assert sampler_output.spec_decode_worker_metrics is None
def test_sampler_output_getitem(sampler_output, sample_outputs):
......
tests/v1/test_oracle.py
View file @ dd572c0a
......@@ -40,12 +40,6 @@ def test_unsupported_configs(monkeypatch):
with monkeypatch.context() as m:
m.setenv("VLLM_USE_V1", "1")
with pytest.raises(NotImplementedError):
AsyncEngineArgs(
model=MODEL,
kv_cache_dtype="fp8",
).create_engine_config()
with pytest.raises(NotImplementedError):
AsyncEngineArgs(
model=MODEL,
......
tools/mypy.sh
View file @ dd572c0a
......@@ -32,6 +32,5 @@ run_mypy vllm/lora
run_mypy vllm/model_executor
run_mypy vllm/plugins
run_mypy vllm/prompt_adapter
run_mypy vllm/spec_decode
run_mypy vllm/worker
run_mypy vllm/v1
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment