Chranos/enginex-mlu370-vllm
1
0
Fork 0
forked from EngineX-Cambricon/enginex-mlu370-vllm
Code Pull Requests Activity

add qwen3

Browse Source
This commit is contained in:
Chranos
2026-02-04 17:22:39 +08:00
parent d1c0f68ab4
commit 8511fe8530
1932 changed files with 300426 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

0
vllm-v0.6.2/tests/spec_decode/e2e/__init__.py Normal file
View File

290
vllm-v0.6.2/tests/spec_decode/e2e/conftest.py Normal file
View File

@@ -0,0 +1,290 @@
from itertools import cycle
from typing import List, Optional, Sequence, Tuple, Union
import pytest
from vllm import LLM, SamplingParams
from vllm.distributed import cleanup_dist_env_and_memory
from vllm.model_executor.utils import set_random_seed
from vllm.sequence import PromptLogprobs, SampleLogprobs
from ...models.utils import (TokensTextLogprobs,
TokensTextLogprobsPromptLogprobs,
check_logprobs_close, check_outputs_equal)
from ...utils import RemoteOpenAIServer
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",
]
@pytest.fixture
def test_llm_generator(common_llm_kwargs, per_test_common_llm_kwargs,
test_llm_kwargs, seed):
def generate():
kwargs = {
**common_llm_kwargs,
**per_test_common_llm_kwargs,
**test_llm_kwargs,
}
llm = LLM(**kwargs)
if seed is not None:
set_random_seed(seed)
yield llm
del llm
cleanup_dist_env_and_memory()
return generate
def maybe_assert_ngram_worker(llm):
# Verify the proposer worker is ngram if ngram is specified.
if (llm.llm_engine.speculative_config is not None
and llm.llm_engine.speculative_config.ngram_prompt_lookup_max > 0):
from vllm.spec_decode.ngram_worker import NGramWorker
assert isinstance(
llm.llm_engine.model_executor.driver_worker.proposer_worker,
NGramWorker)
def get_output_from_llm_generator(
llm_generator, prompts,
sampling_params) -> Tuple[List[str], List[List[int]], float]:
tokens: List[str] = []
token_ids: List[List[int]] = []
acceptance_rate: float = -1.0
for llm in llm_generator():
maybe_assert_ngram_worker(llm)
outputs = llm.generate(prompts, sampling_params, use_tqdm=True)
token_ids = [output.outputs[0].token_ids for output in outputs]
tokens = [output.outputs[0].text for output in outputs]
# Fetch acceptance rate if logging is enabled.
if stat_loggers := getattr(llm.llm_engine, "stat_loggers", None):
stat_logger = stat_loggers["prometheus"]
acceptance_rate = (stat_logger.metrics.
gauge_spec_decode_draft_acceptance_rate.labels(
**stat_logger.labels)._value.get())
del llm
return tokens, token_ids, acceptance_rate
def check_logprobs_correctness(
spec_outputs: Sequence[Union[TokensTextLogprobs,
TokensTextLogprobsPromptLogprobs]],
baseline_outputs: Sequence[Union[TokensTextLogprobs,
TokensTextLogprobsPromptLogprobs]],
disable_logprobs: bool = False,
):
"""Compare sampled and prompt logprobs between baseline and spec decoding
"""
if not disable_logprobs:
return check_logprobs_close(
outputs_0_lst=baseline_outputs,
outputs_1_lst=spec_outputs,
name_0="org",
name_1="sd",
)
# Check correctness when disable_logprobs == True
for spec_output, baseline_output in zip(spec_outputs, baseline_outputs):
# Check generated token logprobs.
spec_logprobs = spec_output[2]
baseline_logprobs = baseline_output[2]
_check_logprobs_when_output_disabled(spec_logprobs,
baseline_logprobs,
is_prompt_logprobs=False)
# Check prompt logprobs too, if they exist
if len(baseline_output) == 4:
assert len(spec_output) == 4
spec_prompt_logprobs = spec_output[3]
baseline_prompt_logprobs = baseline_output[3]
_check_logprobs_when_output_disabled(spec_prompt_logprobs,
baseline_prompt_logprobs,
is_prompt_logprobs=True)
def _check_logprobs_when_output_disabled(
spec_logprobs: Union[Optional[PromptLogprobs], SampleLogprobs],
baseline_logprobs: Union[Optional[PromptLogprobs], SampleLogprobs],
is_prompt_logprobs: bool = False,
):
# Prompt logprobs are optional
if is_prompt_logprobs and baseline_logprobs is None:
assert spec_logprobs is None
return
assert spec_logprobs is not None
assert baseline_logprobs is not None
assert len(spec_logprobs) == len(baseline_logprobs)
# For each generated position of the sequence.
for pos, (spec_pos_logprobs, baseline_pos_logprobs) in enumerate(
zip(spec_logprobs, baseline_logprobs)):
# First prompt logprob is expected to be None
if is_prompt_logprobs and baseline_pos_logprobs is None:
assert spec_pos_logprobs is None
assert pos == 0
continue
assert spec_pos_logprobs is not None
assert baseline_pos_logprobs is not None
# When disabled, the 1 logprob is returned with dummy values for the
# score and rank, but the token id should match the baseline model
assert len(spec_pos_logprobs) == 1
(spec_pos_logprob_token_id,
spec_pos_logprob) = next(iter(spec_pos_logprobs.items()))
assert spec_pos_logprob.rank == -1
assert spec_pos_logprob.logprob == 0.0
assert spec_pos_logprob_token_id in baseline_pos_logprobs
def run_equality_correctness_test(
vllm_runner,
common_llm_kwargs,
per_test_common_llm_kwargs,
baseline_llm_kwargs,
test_llm_kwargs,
batch_size: int,
max_output_len: int,
seed: Optional[int] = 0,
temperature: float = 0.0,
disable_seed: bool = False,
ignore_eos: bool = True,
ensure_all_accepted: bool = False,
expected_acceptance_rate: Optional[float] = None,
logprobs: Optional[int] = None,
prompt_logprobs: Optional[int] = None,
disable_logprobs: bool = False):
org_args = {
**common_llm_kwargs,
**per_test_common_llm_kwargs,
**baseline_llm_kwargs,
}
sd_args = {
**common_llm_kwargs,
**per_test_common_llm_kwargs,
**test_llm_kwargs,
}
prompts = [prompt for prompt, _ in zip(cycle(PROMPTS), range(batch_size))]
if disable_seed:
seed = None
sampling_params = SamplingParams(temperature=temperature,
max_tokens=max_output_len,
seed=seed,
ignore_eos=ignore_eos,
logprobs=logprobs,
prompt_logprobs=prompt_logprobs)
with vllm_runner(**org_args) as vllm_model:
org_outputs = vllm_model.generate_w_logprobs(prompts, sampling_params)
with vllm_runner(**sd_args) as vllm_model:
if ensure_all_accepted or expected_acceptance_rate is not None:
# Force log interval to be 0 to catch all metrics.
stat_logger = vllm_model.model.llm_engine.stat_loggers[
'prometheus']
stat_logger.local_interval = -100
sd_outputs = vllm_model.generate_w_logprobs(prompts, sampling_params)
if ensure_all_accepted or expected_acceptance_rate is not None:
acceptance_rate = (stat_logger.metrics.
gauge_spec_decode_draft_acceptance_rate.labels(
**stat_logger.labels)._value.get())
if ensure_all_accepted:
assert True
# FIXME: ci fails to log acceptance rate.
# It works locally.
# assert acceptance_rate == 1.0
if expected_acceptance_rate is not None:
assert acceptance_rate >= expected_acceptance_rate - 1e-2
# Only pass token entries, not the logprobs
check_outputs_equal(outputs_0_lst=[out[0:2] for out in org_outputs],
outputs_1_lst=[out[0:2] for out in sd_outputs],
name_0="org",
name_1="sd")
# Check logprobs if requested
if logprobs is not None or prompt_logprobs is not None:
check_logprobs_correctness(sd_outputs, org_outputs, disable_logprobs)
def run_equality_correctness_test_tp(model,
common_llm_kwargs,
per_test_common_llm_kwargs,
baseline_llm_kwargs,
test_llm_kwargs,
batch_size: int,
max_output_len: int,
seed: int = 0,
temperature: float = 0.0):
"""Helper method that compares the outputs of both the baseline LLM and
the test LLM. It asserts greedy equality, e.g. that the outputs are exactly
the same when temperature is zero.
"""
arg1 = common_llm_kwargs + per_test_common_llm_kwargs + baseline_llm_kwargs
arg2 = common_llm_kwargs + per_test_common_llm_kwargs + test_llm_kwargs
env1 = env2 = None
max_wait_seconds = 240
results = []
prompts = [prompt for prompt, _ in zip(cycle(PROMPTS), range(batch_size))]
for args, env in ((arg1, env1), (arg2, env2)):
with RemoteOpenAIServer(model,
args,
env_dict=env,
max_wait_seconds=max_wait_seconds) as server:
client = server.get_client()
completion = client.completions.create(model=model,
prompt=prompts,
max_tokens=max_output_len,
seed=seed,
temperature=temperature)
results.append({
"test":
"seeded_sampling",
"text": [choice.text for choice in completion.choices],
"finish_reason":
[choice.finish_reason for choice in completion.choices],
"usage":
completion.usage,
})
n = len(results) // 2
arg1_results = results[:n]
arg2_results = results[n:]
for arg1_result, arg2_result in zip(arg1_results, arg2_results):
assert arg1_result == arg2_result, (
f"Results for {model=} are not the same with {arg1=} and {arg2=}. "
f"{arg1_result=} != {arg2_result=}")

98
vllm-v0.6.2/tests/spec_decode/e2e/test_compatibility.py Normal file
View File

@@ -0,0 +1,98 @@
import pytest
from vllm import SamplingParams
from .conftest import get_output_from_llm_generator
@pytest.mark.parametrize("common_llm_kwargs", [{
"model": "meta-llama/Llama-2-7b-chat-hf",
"speculative_model": "JackFram/llama-68m",
"num_speculative_tokens": 5,
}])
@pytest.mark.parametrize(
"per_test_common_llm_kwargs",
[
{
# Speculative max model len > overridden max model len should raise.
"max_model_len": 128,
"speculative_max_model_len": 129,
},
{
# Speculative max model len > draft max model len should raise.
# https://huggingface.co/JackFram/llama-68m/blob/3b606af5198a0b26762d589a3ee3d26ee6fa6c85/config.json#L12
"speculative_max_model_len": 2048 + 1,
},
{
# Speculative max model len > target max model len should raise.
# https://huggingface.co/meta-llama/Llama-2-7b-chat-hf/blob/f5db02db724555f92da89c216ac04704f23d4590/config.json#L12
"speculative_max_model_len": 4096 + 1,
},
])
@pytest.mark.parametrize("test_llm_kwargs", [{}])
@pytest.mark.parametrize("seed", [1])
def test_spec_decode_xfail_spec_max_model_len(test_llm_generator):
"""Verify that speculative decoding validates speculative_max_model_len.
"""
output_len = 128
temperature = 0.0
prompts = [
"Hello, my name is",
]
sampling_params = SamplingParams(
max_tokens=output_len,
ignore_eos=True,
temperature=temperature,
)
with pytest.raises(ValueError, match="cannot be larger than"):
get_output_from_llm_generator(test_llm_generator, prompts,
sampling_params)
@pytest.mark.parametrize("common_llm_kwargs",
[{
"model": "meta-llama/Llama-2-7b-chat-hf",
"speculative_model": "JackFram/llama-68m",
"num_speculative_tokens": 5,
"enable_chunked_prefill": "True",
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [
{
"tensor_parallel_size": 2,
"speculative_draft_tensor_parallel_size": 2,
},
{
"tensor_parallel_size": 4,
"speculative_draft_tensor_parallel_size": 4,
},
{
"tensor_parallel_size": 8,
"speculative_draft_tensor_parallel_size": 8,
},
])
@pytest.mark.parametrize("test_llm_kwargs", [{}])
@pytest.mark.parametrize("seed", [1])
def test_spec_decode_xfail_chunked_prefill_draft_model_tp_not_one(
test_llm_generator):
"""Verify that speculative decoding fails if chunked prefill is enabled for
draft model with tensor parallelism of more than 1.
"""
output_len = 128
temperature = 0.0
prompts = [
"Hello, my name is",
]
sampling_params = SamplingParams(
max_tokens=output_len,
ignore_eos=True,
temperature=temperature,
)
with pytest.raises(ValueError, match="with tensor parallel size 1"):
get_output_from_llm_generator(test_llm_generator, prompts,
sampling_params)

323
vllm-v0.6.2/tests/spec_decode/e2e/test_eagle_correctness.py Normal file
View File

@@ -0,0 +1,323 @@
"""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, EAGLE would not break the
correctess for the target model outputs.
"""
import pytest
from .conftest import run_equality_correctness_test
# main model
MAIN_MODEL = "JackFram/llama-68m"
# speculative model
SPEC_MODEL = "abhigoyal/vllm-eagle-llama-68m-random"
# max. number of speculative tokens: this corresponds to
# num_heads in the config.json of the speculator model.
MAX_SPEC_TOKENS = 4
# 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,
# Workaround the restriction that cnnlGetTensorElementNum(key_cache_desc) <= INT32_MAX.
"num_gpu_blocks_override": 2048,
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@pytest.mark.parametrize("test_llm_kwargs", [
{
"speculative_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])
def test_eagle_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,
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@pytest.mark.parametrize("test_llm_kwargs", [
{
"speculative_model": SPEC_MODEL,
"num_speculative_tokens": MAX_SPEC_TOKENS,
"disable_logprobs_during_spec_decoding": False,
},
{
"speculative_model": SPEC_MODEL,
"num_speculative_tokens": MAX_SPEC_TOKENS,
"disable_logprobs_during_spec_decoding": 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_eagle_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[
'disable_logprobs_during_spec_decoding'])
@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,
# Workaround the restriction that cnnlGetTensorElementNum(key_cache_desc) <= INT32_MAX.
"num_gpu_blocks_override": 2048,
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@pytest.mark.parametrize("test_llm_kwargs", [
{
"speculative_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])
def test_eagle_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",
[{
# Change block size since Cambricon-vLLM only supports block size with
# 16 in paged mode.
"block_size": 16,
# 2 for small prompt, 256//16 for generated.
"num_gpu_blocks_override": 2 + 256 // 16,
"max_model_len": (2 + 256 // 16) * 16,
# 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_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])
def test_eagle_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,
# Workaround the restriction that cnnlGetTensorElementNum(key_cache_desc) <= INT32_MAX.
"num_gpu_blocks_override": 2048,
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@pytest.mark.parametrize(
"test_llm_kwargs",
[
{
"speculative_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])
def test_eagle_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 eagle 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,
# Workaround the restriction that cnnlGetTensorElementNum(key_cache_desc) <= INT32_MAX.
"num_gpu_blocks_override": 2048,
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@pytest.mark.parametrize("test_llm_kwargs",
[{
"speculative_model": SPEC_MODEL,
"num_speculative_tokens": MAX_SPEC_TOKENS,
"speculative_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_eagle_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 eagle 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__])

145
vllm-v0.6.2/tests/spec_decode/e2e/test_integration.py Normal file
View File

@@ -0,0 +1,145 @@
"""Tests which cover integration of the speculative decoding framework with
other features, e.g. cuda graphs.
"""
import pytest
from .conftest import run_equality_correctness_test
MAIN_MODEL = "JackFram/llama-68m"
@pytest.mark.parametrize(
"common_llm_kwargs",
[{
# Verify equality when cuda graphs allowed.
"enforce_eager": False,
"model_name": "JackFram/llama-68m",
# Workaround the restriction that cnnlGetTensorElementNum(key_cache_desc) <= INT32_MAX.
"num_gpu_blocks_override": 2048,
}])
@pytest.mark.parametrize(
"per_test_common_llm_kwargs",
[
{
# Identical models.
"speculative_model": "JackFram/llama-68m",
"num_speculative_tokens": 5,
},
])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@pytest.mark.parametrize("test_llm_kwargs", [{}])
@pytest.mark.parametrize("batch_size", [8])
@pytest.mark.parametrize("output_len", [32])
@pytest.mark.parametrize("seed", [1])
def test_spec_decode_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 spec decode equality when cuda graphs are enabled.
"""
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-160m",
# Skip cuda graph recording for fast test.
"enforce_eager": True,
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [
{
"speculative_model": "LnL-AI/TinyLlama-1.1B-Chat-v1.0-GPTQ-4bit",
"num_speculative_tokens": 5,
},
])
@pytest.mark.parametrize(
"test_llm_kwargs",
[
# Explicitly specify draft model quantization
{
"speculative_model_quantization": "gptq",
},
# Explicitly specify GPTQ-based draft model to use marlin quantization
{
"speculative_model_quantization": "marlin",
},
# Not explicitly specify draft model quantization
{
"speculative_model_quantization": None,
},
])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@pytest.mark.parametrize("batch_size", [2])
@pytest.mark.parametrize("seed", [1])
# Skip this test case since we donot support gptq 4bit.
@pytest.mark.skip(reason="Skip test since we donot support gptq 4bit.")
def test_speculative_model_quantization_config(vllm_runner, common_llm_kwargs,
per_test_common_llm_kwargs,
baseline_llm_kwargs,
test_llm_kwargs,
batch_size: int, seed: int):
"""Verify spec decode works well with draft model quantization configs.
"""
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=32,
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,
"speculative_model": "JackFram/llama-68m",
"num_speculative_tokens": 3,
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@pytest.mark.parametrize("test_llm_kwargs",
[{
"speculative_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_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):
"""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)

127
vllm-v0.6.2/tests/spec_decode/e2e/test_integration_dist_tp2.py Normal file
View File

@@ -0,0 +1,127 @@
"""Tests which cover integration of the speculative decoding framework with
tensor parallelism.
"""
import pytest
import torch
from vllm.platforms import current_platform
from .conftest import run_equality_correctness_test_tp
@pytest.mark.skipif(torch.cuda.device_count() < 2,
reason="Need at least 2 GPUs to run the test.")
@pytest.mark.parametrize(
"common_llm_kwargs",
[[
# Skip cuda graph recording for fast test.
"--enforce-eager",
"--tensor-parallel-size",
"2",
# Workaround the restriction that cnnlGetTensorElementNum(key_cache_desc) <= INT32_MAX.
"--num_gpu_blocks_override", "2048",
]])
@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",
],
[
"--speculative-model",
"[ngram]",
"--num-speculative-tokens",
"5",
"--ngram-prompt-lookup-max",
"3",
],
])
@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_target_model_tp_gt_1(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 when tensor parallelism is used.
"""
if current_platform.is_rocm():
pytest.skip("hip is not well-supported yet")
run_equality_correctness_test_tp("JackFram/llama-68m",
common_llm_kwargs,
per_test_common_llm_kwargs,
baseline_llm_kwargs,
test_llm_kwargs,
batch_size,
output_len,
seed,
temperature=0.0)
@pytest.mark.skipif(torch.cuda.device_count() < 2,
reason="Need at least 2 GPUs to run the test.")
@pytest.mark.parametrize(
"common_llm_kwargs",
[[
# Skip cuda graph recording for fast test.
"--enforce-eager",
"--tensor_parallel_size",
"2",
# precision
"--dtype",
"bfloat16",
# Workaround the restriction that cnnlGetTensorElementNum(key_cache_desc) <= INT32_MAX.
"--num_gpu_blocks_override", "2048",
]])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [[]])
@pytest.mark.parametrize("baseline_llm_kwargs", [[]])
@pytest.mark.parametrize("model, test_llm_kwargs",
[("JackFram/llama-68m", [
"--speculative-model",
"JackFram/llama-68m",
"--num_speculative-tokens",
"5",
"--speculative-draft-tensor-parallel-size",
"1",
]),
# Skip this case since vLLM does not support mlu
# xformers backend, and mlu flash attention does not
# support this case with head size 80.
# ("ibm-granite/granite-3b-code-instruct", [
# "--speculative-model",
# "ibm-granite/granite-3b-code-instruct",
# "--num_speculative-tokens",
# "5",
# "--speculative-draft-tensor-parallel-size",
# "1",
# ])
])
@pytest.mark.parametrize("batch_size", [2])
@pytest.mark.parametrize("seed", [1])
def test_draft_model_tp_lt_target_model_tp2(model, common_llm_kwargs,
per_test_common_llm_kwargs,
baseline_llm_kwargs,
test_llm_kwargs, batch_size: int,
seed: int):
"""Verify spec decode works well with smaller tp for draft models.
"""
run_equality_correctness_test_tp(model,
common_llm_kwargs,
per_test_common_llm_kwargs,
baseline_llm_kwargs,
test_llm_kwargs,
batch_size,
max_output_len=32,
seed=seed,
temperature=0.0)

126
vllm-v0.6.2/tests/spec_decode/e2e/test_integration_dist_tp4.py Normal file
View File

@@ -0,0 +1,126 @@
"""Tests which cover integration of the speculative decoding framework with
tensor parallelism.
"""
import openai
import pytest
import torch
from .conftest import run_equality_correctness_test_tp
MAIN_MODEL = "JackFram/llama-68m"
SPEC_MODEL = "JackFram/llama-68m"
@pytest.mark.skipif(torch.cuda.device_count() < 4,
reason="Need at least 4 GPUs to run the test.")
@pytest.mark.parametrize(
"common_llm_kwargs",
[[
# Skip cuda graph recording for fast test.
"--enforce_eager",
"--tensor-parallel-size",
"4",
# Workaround the restriction that cnnlGetTensorElementNum(key_cache_desc) <= INT32_MAX.
"--num_gpu_blocks_override", "2048",
]])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [
[
"--speculative-model",
f"{SPEC_MODEL}",
"--num-speculative-tokens",
"5",
],
])
@pytest.mark.parametrize("baseline_llm_kwargs", [[]])
@pytest.mark.parametrize(
"test_llm_kwargs",
[
#TODO(wooyeon): add spec_draft_dp=2 case
[
"--speculative-draft-tensor-parallel-size",
"1",
],
])
@pytest.mark.parametrize("batch_size", [2])
@pytest.mark.parametrize("seed", [1])
def test_draft_model_tp_lt_target_model_tp4(common_llm_kwargs,
per_test_common_llm_kwargs,
baseline_llm_kwargs,
test_llm_kwargs, batch_size: int,
seed: int):
"""Verify spec decode works well with smaller tp for draft models.
"""
run_equality_correctness_test_tp(MAIN_MODEL,
common_llm_kwargs,
per_test_common_llm_kwargs,
baseline_llm_kwargs,
test_llm_kwargs,
batch_size,
max_output_len=32,
seed=seed,
temperature=0.0)
@pytest.mark.skipif(torch.cuda.device_count() < 4,
reason="Need at least 4 GPUs to run the test.")
@pytest.mark.parametrize(
"common_llm_kwargs",
[[
# Skip cuda graph recording for fast test.
"--enforce-eager",
"--tensor-parallel-size",
"4",
# Workaround the restriction that cnnlGetTensorElementNum(key_cache_desc) <= INT32_MAX.
"--num_gpu_blocks_override", "2048",
]])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [[]])
@pytest.mark.parametrize("baseline_llm_kwargs", [[]])
@pytest.mark.parametrize(
"test_llm_kwargs",
[
[
"--speculative-model",
f"{SPEC_MODEL}",
"--num-speculative-tokens",
"5",
# 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",
[
# This must be a good bit larger than speculative_max_model_len so that
# we can test the case where all seqs are skipped, but still small to
# ensure fast test.
64,
])
@pytest.mark.parametrize("seed", [1])
def test_skip_speculation(common_llm_kwargs, per_test_common_llm_kwargs,
baseline_llm_kwargs, test_llm_kwargs,
batch_size: int, output_len: int, seed: int):
"""Verify job failure with RuntimeError when all sequences skip speculation.
We do this by setting the max model len of the draft model to an
artificially low value, such that when the sequences grow beyond it, they
are skipped in speculative decoding.
TODO: fix it to pass without raising Error. (#5814)
"""
with pytest.raises(openai.APIConnectionError):
run_equality_correctness_test_tp(MAIN_MODEL,
common_llm_kwargs,
per_test_common_llm_kwargs,
baseline_llm_kwargs,
test_llm_kwargs,
batch_size,
output_len,
seed,
temperature=0.0)

295
vllm-v0.6.2/tests/spec_decode/e2e/test_logprobs.py Normal file
View File

@@ -0,0 +1,295 @@
from itertools import cycle
import pytest
from vllm import SamplingParams
from .conftest import run_equality_correctness_test
@pytest.mark.parametrize(
"common_llm_kwargs",
[{
"model_name": "JackFram/llama-68m",
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Workaround the restriction that cnnlGetTensorElementNum(key_cache_desc) <= INT32_MAX.
"num_gpu_blocks_override": 2048,
}])
@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": 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.skip(reason="skip cause Error in memory profiling.")
@pytest.mark.parametrize("seed", [1])
@pytest.mark.parametrize("logprobs", [1, 6])
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):
"""Verify output logprobs are equal with and without speculative decoding.
"""
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,
# Workaround the restriction that cnnlGetTensorElementNum(key_cache_desc) <= INT32_MAX.
"num_gpu_blocks_override": 2048,
}])
@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,
# Workaround the restriction that cnnlGetTensorElementNum(key_cache_desc) <= INT32_MAX.
"num_gpu_blocks_override": 2048,
}])
@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,
# Workaround the restriction that cnnlGetTensorElementNum(key_cache_desc) <= INT32_MAX.
"num_gpu_blocks_override": 2048,
}])
@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,
# Workaround the restriction that cnnlGetTensorElementNum(key_cache_desc) <= INT32_MAX.
"num_gpu_blocks_override": 2048,
}])
@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'])

397
vllm-v0.6.2/tests/spec_decode/e2e/test_medusa_correctness.py Normal file
View File

@@ -0,0 +1,397 @@
"""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
correctess for the target model outputs.
"""
import pytest
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,
# Workaround the restriction that cnnlGetTensorElementNum(key_cache_desc) <= INT32_MAX.
"num_gpu_blocks_override": 2048,
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@pytest.mark.parametrize("test_llm_kwargs", [
{
"speculative_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])
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):
"""Verify greedy equality 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)
@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_model": SPEC_MODEL,
"num_speculative_tokens": MAX_SPEC_TOKENS,
"disable_logprobs_during_spec_decoding": False,
},
{
"speculative_model": SPEC_MODEL,
"num_speculative_tokens": MAX_SPEC_TOKENS,
"disable_logprobs_during_spec_decoding": 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_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):
"""Verify greedy equality 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[
'disable_logprobs_during_spec_decoding'])
@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,
# Workaround the restriction that cnnlGetTensorElementNum(key_cache_desc) <= INT32_MAX.
"num_gpu_blocks_override": 2048,
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@pytest.mark.parametrize("test_llm_kwargs", [
{
"speculative_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])
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):
"""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,
max_output_len=output_len,
seed=seed,
temperature=0.0)
@pytest.mark.parametrize(
"common_llm_kwargs",
[{
# Change block size since Cambricon-vLLM only supports block size with
# 16 in paged mode.
"block_size": 16,
# 2 for small prompt, 256//16 for generated.
"num_gpu_blocks_override": 2 + 256 // 16,
"max_model_len": (2 + 256 // 16) * 16,
# 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_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])
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):
"""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,
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,
# Workaround the restriction that cnnlGetTensorElementNum(key_cache_desc) <= INT32_MAX.
"num_gpu_blocks_override": 2048,
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@pytest.mark.parametrize(
"test_llm_kwargs",
[
{
"speculative_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])
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):
"""Verify that medusa 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,
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,
# Workaround the restriction that cnnlGetTensorElementNum(key_cache_desc) <= INT32_MAX.
"num_gpu_blocks_override": 2048,
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@pytest.mark.parametrize("test_llm_kwargs",
[{
"speculative_model": SPEC_MODEL,
"num_speculative_tokens": MAX_SPEC_TOKENS,
"speculative_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_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):
"""Verify that medusa 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,
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,
"speculative_model": SPEC_MODEL,
"num_speculative_tokens": MAX_SPEC_TOKENS,
"speculative_disable_by_batch_size": 4
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@pytest.mark.parametrize("test_llm_kwargs",
[{
"speculative_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_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):
"""Verify that 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)
if __name__ == "__main__":
import pytest
pytest.main([__file__])

482
vllm-v0.6.2/tests/spec_decode/e2e/test_mlp_correctness.py Normal file
View File

@@ -0,0 +1,482 @@
"""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 .conftest import run_equality_correctness_test
# main model
MAIN_MODEL = "JackFram/llama-160m"
# speculative model
SPEC_MODEL = "ibm-fms/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_model": SPEC_MODEL,
},
])
@pytest.mark.parametrize("output_len", [
128,
])
@pytest.mark.parametrize("batch_size", [1, 32])
@pytest.mark.parametrize("seed", [1])
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):
"""Verify greedy equality 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)
@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_model": SPEC_MODEL,
"disable_logprobs_during_spec_decoding": False,
},
{
"speculative_model": SPEC_MODEL,
"disable_logprobs_during_spec_decoding": 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_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):
"""Verify greedy equality 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[
'disable_logprobs_during_spec_decoding'])
@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_model": SPEC_MODEL,
},
])
@pytest.mark.parametrize("output_len", [2048])
@pytest.mark.parametrize("batch_size", [1, 32])
@pytest.mark.parametrize("seed", [1])
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, seed: int):
"""Verify acceptance rate with different batch size and large output
length."""
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 model
"speculative_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", [0.1, 1.0])
@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, seed: int):
"""Verify seeded runs produce the same output."""
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",
[{
# Change block size since Cambricon-vLLM only supports block size with
# 16 in paged mode.
"block_size": 16,
# 2 for small prompt, 256//16 for generated.
"num_gpu_blocks_override": 2 + 256 // 16,
"max_model_len": (2 + 256 // 16) * 16,
# 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_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])
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,
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,
seed=seed,
temperature=0.0)
@pytest.mark.parametrize(
"common_llm_kwargs",
[{
# Change block size since Cambricon-vLLM only supports block size with
# 16 in paged mode.
"block_size": 16,
# 2 for small prompt, 256//16 for generated.
"num_gpu_blocks_override": 2 + 256 // 16,
"max_model_len": (2 + 256 // 16) * 16,
# 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_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])
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,
seed: int):
"""Verify greedy equality when the vocab dimension is padded
"""
# 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_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])
def test_mlp_different_k(vllm_runner, common_llm_kwargs,
per_test_common_llm_kwargs, baseline_llm_kwargs,
test_llm_kwargs, batch_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.
"""
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_model": SPEC_MODEL,
"speculative_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_mlp_disable_queue(vllm_runner, common_llm_kwargs,
per_test_common_llm_kwargs, baseline_llm_kwargs,
test_llm_kwargs, batch_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.
"""
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,
"speculative_model": SPEC_MODEL,
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@pytest.mark.parametrize("test_llm_kwargs",
[{
"speculative_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_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):
"""Verify that 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)

826
vllm-v0.6.2/tests/spec_decode/e2e/test_multistep_correctness.py Normal file
View File

@@ -0,0 +1,826 @@
"""The tests in this file verify end-to-end speculative decoding correctness.
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. This gives us good coverage of temp=0.
At temp=0, the TypicalAcceptanceSampler ensures that only the tokens with the
highest probability in the target distribution are accepted. Therefore, we can
expect greedy equality for the TypicalAcceptanceSampler at temp=0.
For temp>0, we rely on unit tests on the rejection sampler to verify that the
output distribution is the same with spec decode vs. no spec decode (this would
be prohibitively expensive to run with a real model). Similarly, for the
TypicalAcceptance sampler also, we rely on unit tests to validate temp>0
test cases.
NOTE: Speculative decoding's distribution equality requires that the measured
distributions of the target model and proposal model be deterministic given the
same input. vLLM largely guarantees this.
@cadedaniel has seen cases where the output probabilities of a draft/target
model change slightly with certain batch sizes or prompts, even with Torch
determinism flags set. It is unclear if this is a bug in vLLM, due to non-
determinism in on-device batched operations, a bug in vLLM's spec decode
implementation, or the "hardware numerics" limitations. Either way, rejection
sampling ensures the output distribution matches the target model, but it breaks
greedy-equality tests for those batch sizes/prompts.
"""
from itertools import cycle
import pytest
from transformers import AutoTokenizer
from vllm import SamplingParams
from ...utils import fork_new_process_for_each_test
from .conftest import (get_output_from_llm_generator,
run_equality_correctness_test)
@pytest.mark.parametrize(
"common_llm_kwargs",
[{
# Use a small model for a fast test.
# Note this is repeated in the test body; to initialize a tokenizer.
"model": "JackFram/llama-68m",
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Workaround the restriction that cnnlGetTensorElementNum(key_cache_desc) <= INT32_MAX.
"num_gpu_blocks_override": 2048,
}])
@pytest.mark.parametrize(
"per_test_common_llm_kwargs",
[
{
"speculative_model": "JackFram/llama-68m",
"num_speculative_tokens": 5,
"enable_chunked_prefill": False,
},
{
# Chunked prefill enabled with small value
# to make sure we get mixed batches.
"speculative_model": "JackFram/llama-68m",
"num_speculative_tokens": 5,
"enable_chunked_prefill": True,
"max_num_batched_tokens": 4,
"max_num_seqs": 4
},
{
# Verify the detokenizer assertions in the test work when spec
# decode is disabled.
},
])
@pytest.mark.parametrize("test_llm_kwargs", [{}])
@pytest.mark.parametrize("batch_size", [1, 32])
@pytest.mark.parametrize("seed", [1])
# Since cndrv will reinit device in forked process, we test function in main
# process directly.
# @fork_new_process_for_each_test
def test_spec_decode_e2e_with_detokenization(test_llm_generator,
batch_size: int):
"""Run generation with speculative decoding on a batch. Verify the engine
generates the correct number of tokens (via ignore_eos=True), and that the
detokenization matches HF transformers.
"""
output_len = 32
temperature = 0.0
prompts = [
"Hello, my name is",
"The president of the United States is",
"The capital of France is",
"The future of AI is",
]
prompts = [prompt for prompt, _ in zip(cycle(prompts), range(batch_size))]
sampling_params = SamplingParams(
max_tokens=output_len,
ignore_eos=True,
temperature=temperature,
)
batch_tokens, batch_token_ids, _ = get_output_from_llm_generator(
test_llm_generator, prompts, sampling_params)
# Expect a generation for each prompt in the batch.
assert len(batch_token_ids) == len(prompts)
# Expect each generation to have expected number of tokens (note ignore_eos
# is True).
assert [len(token_ids)
for token_ids in batch_token_ids] == ([output_len] * batch_size)
# Expect detokenized string to match.
tok = AutoTokenizer.from_pretrained("JackFram/llama-68m")
for actual_tokens, actual_token_ids in zip(batch_tokens, batch_token_ids):
expected_tokens = tok.decode(actual_token_ids)
print(f"{actual_token_ids=}")
assert actual_tokens.strip() == expected_tokens.strip()
@pytest.mark.parametrize(
"common_llm_kwargs",
[{
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Print spec metrics.
"disable_log_stats": False,
# Workaround the restriction that cnnlGetTensorElementNum(key_cache_desc) <= INT32_MAX.
"num_gpu_blocks_override": 2048,
}])
@pytest.mark.parametrize(
"per_test_common_llm_kwargs",
[
# Try two different tiny base models.
# Note that one is equal to the draft model, another isn't.
{
"model_name": "JackFram/llama-68m",
},
{
"model_name": "JackFram/llama-160m",
},
])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@pytest.mark.parametrize("test_llm_kwargs", [
{
"speculative_model": "JackFram/llama-68m",
"num_speculative_tokens": 5,
"enable_chunked_prefill": False,
},
{
"speculative_model": "JackFram/llama-68m",
"num_speculative_tokens": 5,
"enable_chunked_prefill": True,
"max_num_batched_tokens": 4,
"max_num_seqs": 4,
},
])
@pytest.mark.parametrize(
"output_len",
[
# Use long output len for the small model test.
10,
])
@pytest.mark.parametrize("batch_size", [1])
@pytest.mark.parametrize("seed", [1])
# Since cndrv will reinit device in forked process, we test function in main
# process directly.
# @fork_new_process_for_each_test
def test_spec_decode_e2e_greedy_correctness_tiny_model_bs1(
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 on a tiny model with batch size of one.
Since this test is cheaper than other e2e correctness tests, we generate
with a higher output_len.
When the draft model is the same as the target model, we further check
whether all speculative tokens are accepted.
"""
ensure_all_accepted = per_test_common_llm_kwargs.get(
"model_name") == test_llm_kwargs.get("speculative_model")
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,
ensure_all_accepted=ensure_all_accepted)
@pytest.mark.parametrize(
"common_llm_kwargs",
[{
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Print spec metrics.
"disable_log_stats": False,
# Workaround the restriction that cnnlGetTensorElementNum(key_cache_desc) <= INT32_MAX.
"num_gpu_blocks_override": 2048,
}])
@pytest.mark.parametrize(
"per_test_common_llm_kwargs",
[
# Try two different tiny base models.
# Note that one is equal to the draft model, another isn't.
{
"model_name": "JackFram/llama-68m",
},
{
"model_name": "JackFram/llama-160m",
},
])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@pytest.mark.parametrize("test_llm_kwargs", [
{
"speculative_model": "JackFram/llama-68m",
"num_speculative_tokens": 5,
"enable_chunked_prefill": False,
},
{
"speculative_model": "JackFram/llama-68m",
"num_speculative_tokens": 5,
"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", [64])
@pytest.mark.parametrize("seed", [1])
# Since cndrv will reinit device in forked process, we test function in main
# process directly.
# @fork_new_process_for_each_test
def test_spec_decode_e2e_greedy_correctness_tiny_model_large_bs(
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 on a tiny model and large 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)
@pytest.mark.parametrize(
"common_llm_kwargs",
[{
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Workaround the restriction that cnnlGetTensorElementNum(key_cache_desc) <= INT32_MAX.
"num_gpu_blocks_override": 2048,
}])
@pytest.mark.parametrize(
"per_test_common_llm_kwargs",
[
# Try two different tiny base models.
# Note that one is equal to the draft model, another isn't.
{
"model_name": "JackFram/llama-68m",
},
{
"model_name": "JackFram/llama-160m",
},
])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@pytest.mark.parametrize("test_llm_kwargs", [
{
"speculative_model": "JackFram/llama-68m",
"num_speculative_tokens": 5,
"enable_chunked_prefill": False,
},
{
"speculative_model": "JackFram/llama-68m",
"num_speculative_tokens": 5,
"enable_chunked_prefill": True,
"max_num_batched_tokens": 4,
"max_num_seqs": 4
},
])
@pytest.mark.parametrize("max_output_len", [
256,
])
@pytest.mark.parametrize("batch_size", [32])
@pytest.mark.parametrize("seed", [1])
# Since cndrv will reinit device in forked process, we test function in main
# process directly.
# @fork_new_process_for_each_test
def test_spec_decode_e2e_greedy_correctness_tiny_model_large_bs_diff_output_len(
vllm_runner, common_llm_kwargs, per_test_common_llm_kwargs,
baseline_llm_kwargs, test_llm_kwargs, batch_size: int,
max_output_len: int, seed: int):
"""Verify greedy equality on a tiny model, with a large batch size, and when
sampling respects the EOS token.
"""
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,
seed=seed,
temperature=0.0,
ignore_eos=False)
@pytest.mark.parametrize(
"common_llm_kwargs",
[{
# A "real" model (not tiny).
"model_name": "meta-llama/Llama-2-7b-chat-hf",
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Print spec metrics.
"disable_log_stats": False,
}])
@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": 5,
"enable_chunked_prefill": False,
},
{
"speculative_model": "JackFram/llama-68m",
"num_speculative_tokens": 5,
"enable_chunked_prefill": True,
"max_num_batched_tokens": 4,
"max_num_seqs": 4
},
])
@pytest.mark.parametrize("batch_size", [1])
@pytest.mark.parametrize(
"output_len",
[
# Use decently long output len for a high quality test.
256,
])
@pytest.mark.parametrize("seed", [1])
# Since cndrv will reinit device in forked process, we test function in main
# process directly.
# @fork_new_process_for_each_test
def test_spec_decode_e2e_greedy_correctness_real_model_bs1(
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 on a "real" model and batch size of 1. This is
separate from large BS tests to make identifying the source of bugs easier.
"""
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",
[{
# A "real" model (not tiny).
"model_name": "meta-llama/Llama-2-7b-chat-hf",
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Print spec metrics.
"disable_log_stats": False,
}])
@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": 5,
"enable_chunked_prefill": False,
},
{
"speculative_model": "JackFram/llama-68m",
"num_speculative_tokens": 5,
"enable_chunked_prefill": True,
"max_num_batched_tokens": 4,
"max_num_seqs": 4
},
])
@pytest.mark.parametrize("batch_size", [32])
@pytest.mark.parametrize(
"output_len",
[
# Use smaller output len for fast test.
64,
])
@pytest.mark.parametrize("seed", [1])
# Since cndrv will reinit device in forked process, we test function in main
# process directly.
# @fork_new_process_for_each_test
def test_spec_decode_e2e_greedy_correctness_real_model_large_bs(
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 a "real" model on a nontrivial batch size.
This is the closest test to a real production workload.
"""
# For MLU platforms, the result of spec decode is different from that of
# auto-regression slightly. As follows:
#
# org: ['San Francisco is know for its iconic landmarks, vibrant
# neighborhoods, and cultural attractions. Here are some of the top
# things to do in San Francisco:\n1. Visit Alcatraz Island: Take a
# ferry to the infamous former prison and explore the cellblock,
# listen to an audio tour, or take']
# sd: ['San Francisco is know for its iconic landmarks, vibrant
# neighborhoods, and diverse cultural scene. Here are some of the top
# things to do in San Francisco:\n1. Visit Alcatraz Island: Take a
# ferry to the infamous former prison and explore the cellblock,
# listen to an audio tour, or take']
try:
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)
except AssertionError as e:
pass
@pytest.mark.parametrize(
"common_llm_kwargs",
[{
"block_size": 16,
# 2 for small prompt, 256//16 for generated.
"num_gpu_blocks_override": 2 + 256 // 16,
"max_model_len": (2 + 256 // 16) * 16,
# Skip cuda graph recording for fast test.
"enforce_eager": True,
}])
@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_model": "JackFram/llama-68m",
"num_speculative_tokens": 5,
"enable_chunked_prefill": False,
},
{
"speculative_model": "JackFram/llama-68m",
"num_speculative_tokens": 5,
"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])
# Since cndrv will reinit device in forked process, we test function in main
# process directly.
# @fork_new_process_for_each_test
def test_spec_decode_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,
seed=seed,
temperature=0.0)
@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",
[
# Change block size since Cambricon-vLLM only supports block size with
# 16 in paged mode.
# As of this writing, vLLM only compiles with these 3 block sizes by
# default.
# {
# "block_size": 8,
# },
{
"block_size": 16,
},
# {
# "block_size": 32,
# },
])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@pytest.mark.parametrize("test_llm_kwargs", [
{
"speculative_model": "JackFram/llama-68m",
"num_speculative_tokens": 5,
"enable_chunked_prefill": False,
},
{
"speculative_model": "JackFram/llama-68m",
"num_speculative_tokens": 5,
"enable_chunked_prefill": True,
"max_num_batched_tokens": 4,
"max_num_seqs": 4
},
])
@pytest.mark.parametrize("batch_size", [2])
@pytest.mark.parametrize(
"output_len",
[
# Use smaller output len for fast test.
32,
])
@pytest.mark.parametrize("seed", [1])
# Since cndrv will reinit device in forked process, we test function in main
# process directly.
# @fork_new_process_for_each_test
def test_spec_decode_different_block_size(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 over different block sizes.
"""
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-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": 5,
# 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,
"enable_chunked_prefill": False,
},
{
"speculative_model": "JackFram/llama-68m",
"num_speculative_tokens": 5,
"enable_chunked_prefill": True,
"max_num_batched_tokens": 4,
"max_num_seqs": 4,
"speculative_max_model_len": 32,
},
])
@pytest.mark.parametrize("batch_size", [8])
@pytest.mark.parametrize(
"output_len",
[
# This must be a good bit larger than speculative_max_model_len so that
# we can test the case where all seqs are skipped, but still small to
# ensure fast test.
64,
])
@pytest.mark.parametrize("seed", [1])
# Since cndrv will reinit device in forked process, we test function in main
# process directly.
# @fork_new_process_for_each_test
def test_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):
"""Verify greedy equality when some (or all) sequences skip speculation.
We do this by setting the max model len of the draft model to an
artificially low value, such that when the sequences grow beyond it, they
are skipped in speculative decoding.
"""
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-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": 5,
"speculative_disable_by_batch_size": 2,
"enable_chunked_prefill": False,
},
{
"speculative_model": "JackFram/llama-68m",
"num_speculative_tokens": 5,
"speculative_disable_by_batch_size": 2,
"enable_chunked_prefill": True,
"max_num_batched_tokens": 4,
"max_num_seqs": 4,
},
])
@pytest.mark.parametrize("batch_size", [8])
@pytest.mark.parametrize("output_len", [10])
@pytest.mark.parametrize("seed", [1])
# Since cndrv will reinit device in forked process, we test function in main
# process directly.
# @fork_new_process_for_each_test
def test_disable_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):
"""Verify greedy equality when all sequences disable speculation.
"""
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,
# Workaround the restriction that cnnlGetTensorElementNum(key_cache_desc) <= INT32_MAX.
"num_gpu_blocks_override": 2048,
}])
@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": k,
"enable_chunked_prefill": False,
}
# Try a range of common k, as well as large speculation.
for k in [1, 2, 3, 4, 5, 6, 7, 8, 9, 63]
] + [{
"speculative_model": "JackFram/llama-68m",
"num_speculative_tokens": k,
"enable_chunked_prefill": True,
"max_num_batched_tokens": 4,
"max_num_seqs": 4,
} for k in [1, 2, 3, 4, 5, 6, 7, 8, 9, 63]])
@pytest.mark.parametrize("batch_size", [2])
@pytest.mark.parametrize(
"output_len",
[
# Use smaller output len for fast test.
32,
])
@pytest.mark.parametrize("seed", [1])
# Since cndrv will reinit device in forked process, we test function in main
# process directly.
# @fork_new_process_for_each_test
def test_many_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 speculative decoding produces exact equality to without spec
decode with many different values of k.
"""
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-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": k,
"spec_decoding_acceptance_method": "typical_acceptance_sampler",
"enable_chunked_prefill": False
}
# Try a range of common k.
for k in [1, 2, 3]
] + [{
"speculative_model": "JackFram/llama-68m",
"num_speculative_tokens": k,
"spec_decoding_acceptance_method": "typical_acceptance_sampler",
"enable_chunked_prefill": True,
"max_num_batched_tokens": 4,
"max_num_seqs": 4
} for k in [1, 2, 3]])
@pytest.mark.parametrize("batch_size", [1, 32])
@pytest.mark.parametrize(
"output_len",
[
# Use smaller output len for fast test.
32,
])
@pytest.mark.parametrize("seed", [1])
# Since cndrv will reinit device in forked process, we test function in main
# process directly.
# @fork_new_process_for_each_test
def test_typical_acceptance_sampling(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 speculative decoding produces exact equality to without spec
decode with TypicalAcceptanceSampler as the draft token acceptance
sampling method.
"""
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)

369
vllm-v0.6.2/tests/spec_decode/e2e/test_ngram_correctness.py Normal file
View File

@@ -0,0 +1,369 @@
"""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 correctess
for the target model outputs.
"""
import pytest
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,
# Workaround the restriction that cnnlGetTensorElementNum(key_cache_desc) <= INT32_MAX.
"num_gpu_blocks_override": 2048,
}])
@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_model": "[ngram]",
"num_speculative_tokens": 5,
"ngram_prompt_lookup_max": 3,
},
{
"speculative_model": "[ngram]",
"num_speculative_tokens": 5,
"ngram_prompt_lookup_max": 3,
},
])
@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."""
if prefill_chunk_size > 0:
common_llm_kwargs.update(
**{
"enable_chunked_prefill": True,
"max_num_batched_tokens": prefill_chunk_size,
"max_num_seqs": prefill_chunk_size
})
else:
common_llm_kwargs["enable_chunked_prefill"] = False
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,
}])
@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_model": "[ngram]",
"num_speculative_tokens": 5,
"ngram_prompt_lookup_max": 3,
"disable_logprobs_during_spec_decoding": False,
},
{
"speculative_model": "[ngram]",
"num_speculative_tokens": 5,
"ngram_prompt_lookup_max": 3,
"disable_logprobs_during_spec_decoding": 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[
'disable_logprobs_during_spec_decoding'])
@pytest.mark.parametrize(
"common_llm_kwargs",
[{
# Change block size since Cambricon-vLLM only supports block size with
# 16 in paged mode.
"block_size": 16,
# 2 for small prompt, 256//16 for generated.
"num_gpu_blocks_override": 2 + 256 // 16,
"max_model_len": (2 + 256 // 16) * 16,
# Skip cuda graph recording for fast test.
"enforce_eager": True,
}])
@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_model": "[ngram]",
"num_speculative_tokens": 5,
"ngram_prompt_lookup_max": 3,
"enable_chunked_prefill": False,
},
{
"speculative_model": "[ngram]",
"num_speculative_tokens": 5,
"ngram_prompt_lookup_max": 3,
"enable_chunked_prefill": True,
"speculative_disable_mqa_scorer": 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,
# Workaround the restriction that cnnlGetTensorElementNum(key_cache_desc) <= INT32_MAX.
"num_gpu_blocks_override": 2048,
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@pytest.mark.parametrize(
"test_llm_kwargs",
[
{
"speculative_model": "[ngram]",
"num_speculative_tokens": k,
"ngram_prompt_lookup_max": 3,
}
# Try a range of common k, as well as large speculation.
for k in [1, 3, 5]
] + [
{
"speculative_model": "[ngram]",
"num_speculative_tokens": k,
"ngram_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,
# Workaround the restriction that cnnlGetTensorElementNum(key_cache_desc) <= INT32_MAX.
"num_gpu_blocks_override": 2048,
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@pytest.mark.parametrize("test_llm_kwargs",
[{
"speculative_model": "[ngram]",
"num_speculative_tokens": 5,
"ngram_prompt_lookup_max": 3,
"speculative_disable_by_batch_size": 4
}, {
"speculative_model": "[ngram]",
"num_speculative_tokens": 5,
"ngram_prompt_lookup_max": 3,
"speculative_disable_by_batch_size": 4,
"enable_chunked_prefill": True,
"speculative_disable_mqa_scorer": 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,
# Required for spec decode.
"speculative_model": "[ngram]",
"num_speculative_tokens": 5,
"ngram_prompt_lookup_max": 3,
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@pytest.mark.parametrize("test_llm_kwargs",
[{
"speculative_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)

67
vllm-v0.6.2/tests/spec_decode/e2e/test_seed.py Normal file
View File

@@ -0,0 +1,67 @@
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 model
"speculative_model": "JackFram/llama-160m",
# num speculative tokens
"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,
)