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

[Speculative Decoding] Test refactor (#8317)

Browse Source 
Co-authored-by: youkaichao <youkaichao@126.com>
This commit is contained in:
Lily Liu
2024-09-11 14:07:34 -07:00
committed by GitHub
parent 8baa454937
commit 775f00f81e
12 changed files with 927 additions and 1042 deletions
Download Patch File Download Diff File Expand all files Collapse all files

126
tests/spec_decode/e2e/test_integration_dist_tp4.py
View File

@@ -2,98 +2,97 @@
tensor parallelism.
"""
import openai
import pytest
import torch
from .conftest import run_greedy_equality_correctness_test
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",
[{
# 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,
"--enforce_eager",
# Required for spec decode.
"use_v2_block_manager": True,
"tensor_parallel_size": 4,
# Use AsyncLLM engine, so that the engine runs in its own process.
# Otherwise, since vLLM does not follow true SPMD, the test runner
# process will have both the engine and the rank0 worker. NCCL is not
# cleaned up properly, and its server host thread leaks, causing the
# second run of the test to fail with internal NCCL error.
"use_async": True,
}])
"--use-v2-block-manager",
"--tensor-parallel-size",
"4",
]])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [
{
"speculative_model": "JackFram/llama-68m",
"num_speculative_tokens": 5,
},
[
"--speculative-model",
f"{SPEC_MODEL}",
"--num-speculative-tokens",
"5",
],
])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@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,
},
[
"--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(test_llm_generator,
baseline_llm_generator,
batch_size: int):
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_greedy_equality_correctness_test(baseline_llm_generator,
test_llm_generator,
batch_size,
max_output_len=32,
force_output_len=True)
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",
[{
"model": "JackFram/llama-160m",
[[
# Skip cuda graph recording for fast test.
"enforce_eager": True,
"--enforce-eager",
# Required for spec decode.
"use_v2_block_manager": True,
"tensor_parallel_size": 4,
# Use AsyncLLM engine, so that the engine runs in its own process.
# Otherwise, since vLLM does not follow true SPMD, the test runner
# process will have both the engine and the rank0 worker. NCCL is not
# cleaned up properly, and its server host thread leaks, causing the
# second run of the test to fail with internal NCCL error.
"use_async": True,
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
"--use-v2-block-manager",
"--tensor-parallel-size",
"4",
]])
@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-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,
},
"--speculative-max-model-len",
"32",
],
])
@pytest.mark.parametrize("batch_size", [8])
@pytest.mark.parametrize(
@@ -105,8 +104,9 @@ def test_draft_model_tp_lt_target_model_tp4(test_llm_generator,
64,
])
@pytest.mark.parametrize("seed", [1])
def test_skip_speculation(baseline_llm_generator, test_llm_generator,
batch_size: int, output_len: int):
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
@@ -114,9 +114,13 @@ def test_skip_speculation(baseline_llm_generator, test_llm_generator,
TODO: fix it to pass without raising Error. (#5814)
"""
with pytest.raises(RuntimeError):
run_greedy_equality_correctness_test(baseline_llm_generator,
test_llm_generator,
batch_size,
max_output_len=output_len,
force_output_len=True)
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)