Skip to content

GitLab

"vllm/vscode:/vscode.git/clone" did not exist on "1fe554bac32419a6d64a5c977849806a1efd9725"
Commit 6d2051cc authored by zhuwenwen's avatar zhuwenwen
Browse files

Merge tag 'v0.6.3.post1' into v0.6.3.post1-dev

parents 2c7f740a a2c71c54
Hide whitespace changes
Inline Side-by-side
Showing with 365 additions and 334 deletions
tests/prefix_caching/test_prefix_caching.py
View file @ 6d2051cc
......@@ -2,14 +2,9 @@
Run `pytest tests/prefix_caching/test_prefix_caching.py`.
"""
from typing import List
import pytest
from tests.kernels.utils import override_backend_env_variable
from vllm.block import PhysicalTokenBlock
from vllm.core.block_manager_v1 import CachedBlockAllocator
from vllm.utils import Device
from ..models.utils import check_outputs_equal
......@@ -18,86 +13,11 @@ MODELS = [
]
@pytest.mark.parametrize("block_size", [16])
@pytest.mark.parametrize("num_blocks", [16])
def test_block_allocator(
block_size: int,
num_blocks: int,
):
block_hash = 1
block_allocator = CachedBlockAllocator(Device.CPU, block_size, num_blocks)
# Allocate two PysicalTokenBlocks with the same hash and check
# that they are the same PhysicalTokenBlock
first_block = block_allocator.allocate(block_hash, 0)
second_block = block_allocator.allocate(block_hash, 0)
assert (first_block == second_block)
assert (second_block.ref_count == 2)
# Check metric: 1 hit of 2 queries
assert block_allocator.get_prefix_cache_hit_rate() == 0.5
# Free the first_block and confirm that the ref_count is correctly
# decremented on the second block
block_allocator.free(first_block)
assert (second_block.ref_count == 1)
# Free the second block
block_allocator.free(second_block)
# Reallocate the first block and confirm that, even after the block
# had its ref_count go to 0, we still get the same block back
first_block = block_allocator.allocate(block_hash, 0)
assert (first_block == second_block)
assert (first_block.block_hash == block_hash)
# Allocate one more time to get 3/4 hit rate for easy checking
block_allocator.allocate(block_hash, 0)
assert block_allocator.get_prefix_cache_hit_rate() == 0.75
@pytest.mark.parametrize("num_blocks", [16])
def test_eviction(num_blocks: int, ):
block_size = 16
block_allocator = CachedBlockAllocator(Device.CPU, block_size, num_blocks)
blocks: List[PhysicalTokenBlock] = []
for i in range(num_blocks):
# use i as the block_hash
blocks.append(block_allocator.allocate(i, 0))
#Free all blocks
for block in blocks:
block_allocator.free(block)
# Allocate a new block and confirm that it's the first block freed.
# I.E The Least Recently Used block
new_block_hash = block_size
new_block = block_allocator.allocate(new_block_hash, 0)
assert (new_block == blocks[0])
assert (new_block.block_hash == new_block_hash)
# Reallocate the second in blocks to remove it from the free list
realloc_block_hash = 1
realloc_block = block_allocator.allocate(realloc_block_hash, 0)
assert (realloc_block == blocks[realloc_block_hash])
assert (realloc_block.block_hash == realloc_block_hash)
# Allocate a new block and confirm that it's not the realloc_block,
# since the realloc_block shouldn't be in the free list
new_block_hash = block_size + 1
new_block = block_allocator.allocate(new_block_hash, 0)
assert (realloc_block != new_block)
assert (new_block.block_hash == new_block_hash)
assert (new_block.block_number == 2)
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("backend", ["FLASH_ATTN", "FLASHINFER", "XFORMERS"])
@pytest.mark.parametrize("dtype", ["half"])
@pytest.mark.parametrize("max_tokens", [5])
@pytest.mark.parametrize("cached_position", [0, 1])
@pytest.mark.parametrize("use_v2_block_manager", [False, True])
def test_mixed_requests(
hf_runner,
vllm_runner,
......@@ -107,7 +27,6 @@ def test_mixed_requests(
dtype: str,
max_tokens: int,
cached_position: int,
use_v2_block_manager: bool,
monkeypatch,
) -> None:
"""
......@@ -125,7 +44,6 @@ def test_mixed_requests(
model,
dtype=dtype,
enable_prefix_caching=True,
use_v2_block_manager=use_v2_block_manager,
) as vllm_model:
# Run the first prompt so the cache is populated
vllm_outputs = vllm_model.generate_greedy([cached_prompt], max_tokens)
......
tests/quantization/test_bitsandbytes.py
View file @ 6d2051cc
......@@ -9,22 +9,22 @@ import pytest
import torch
from tests.quantization.utils import is_quant_method_supported
from ..utils import fork_new_process_for_each_test
from tests.utils import fork_new_process_for_each_test
models_4bit_to_test = [
('huggyllama/llama-7b', 'quantize model inflight'),
("facebook/opt-125m", "quantize opt model inflight"),
]
models_pre_qaunt_4bit_to_test = [
('lllyasviel/omost-llama-3-8b-4bits',
'read pre-quantized 4-bit NF4 model'),
('PrunaAI/Einstein-v6.1-Llama3-8B-bnb-4bit-smashed',
'read pre-quantized 4-bit FP4 model'),
('poedator/opt-125m-bnb-4bit', 'read pre-quantized 4-bit NF4 opt model'),
]
models_pre_quant_8bit_to_test = [
('meta-llama/Llama-Guard-3-8B-INT8', 'read pre-quantized 8-bit model'),
('meta-llama/Llama-Guard-3-8B-INT8',
'read pre-quantized llama 8-bit model'),
("yec019/fbopt-350m-8bit", "read pre-quantized 8-bit opt model"),
]
......@@ -133,6 +133,7 @@ def validate_generated_texts(hf_runner,
hf_str = hf_log["generated_text"]
vllm_str = vllm_log["generated_text"]
prompt = hf_log["prompt"]
assert hf_str == vllm_str, (f"Model: {model_name}"
f"Mismatch between HF and vLLM outputs:\n"
f"Prompt: {prompt}\n"
......
tests/quantization/test_compressed_tensors.py
View file @ 6d2051cc
......@@ -2,26 +2,28 @@
Run `pytest tests/quantization/test_compressed_tensors.py`.
"""
from typing import Optional
import pytest
import torch
from compressed_tensors.quantization import QuantizationType
from vllm.model_executor.layers.quantization.compressed_tensors.compressed_tensors import ( # noqa: E501
CompressedTensorsLinearMethod, CompressedTensorsW4A16Sparse24,
CompressedTensorsW8A8Fp8, CompressedTensorsW8A8Int8,
CompressedTensorsW8A16Fp8, CompressedTensorsWNA16)
from vllm.model_executor.layers.quantization.compressed_tensors.utils import (
QuantizationType)
@pytest.mark.parametrize("model_args", [
("nm-testing/tinyllama-oneshot-w8w8-test-static-shape-change", "tensor",
QuantizationType.INT, 2560),
("nm-testing/tinyllama-oneshot-w8-channel-a8-tensor", "channel",
QuantizationType.INT, 2560),
])
@pytest.mark.parametrize(
"model_args",
[("nm-testing/tinyllama-oneshot-w8w8-test-static-shape-change", "tensor",
QuantizationType.INT, 2560, True),
("nm-testing/tinyllama-oneshot-w8-channel-a8-tensor", "channel",
QuantizationType.INT, 2560, True),
("nm-testing/asym-w8w8-int8-static-per-tensor-tiny-llama", "tensor",
QuantizationType.INT, 2560, False)])
def test_compressed_tensors_w8a8_static_setup(vllm_runner, model_args):
model_path, strategy, quant_type, shape_0 = model_args
model_path, strategy, quant_type, shape_0, is_symmetric = model_args
with vllm_runner(model_path, enforce_eager=True) as llm:
model = llm.model.llm_engine.model_executor.driver_worker.model_runner.model # noqa: E501
layer = model.model.layers[0]
......@@ -31,6 +33,18 @@ def test_compressed_tensors_w8a8_static_setup(vllm_runner, model_args):
gate_up_proj = layer.mlp.gate_up_proj
down_proj = layer.mlp.down_proj
# assert zp for symmetric and asymmetric cases
def zp_valid(zp: Optional[torch.Tensor]):
if is_symmetric:
return zp is None
return zp is not None and zp.dtype is torch.int32
assert zp_valid(qkv_proj.input_zero_point)
assert zp_valid(o_proj.input_zero_point)
assert zp_valid(gate_up_proj.input_zero_point)
assert zp_valid(down_proj.input_zero_point)
assert isinstance(qkv_proj.quant_method, CompressedTensorsLinearMethod)
assert isinstance(o_proj.quant_method, CompressedTensorsLinearMethod)
assert isinstance(gate_up_proj.quant_method,
......@@ -69,9 +83,12 @@ def test_compressed_tensors_no_enforce_eager(vllm_runner):
@pytest.mark.parametrize("model_args", [
("nm-testing/tinyllama-oneshot-w8a8-dynamic-token-v2", "tensor"),
("nm-testing/tinyllama-oneshot-w8a8-dynamic-token-v2-asym", "tensor"),
("nm-testing/tinyllama-oneshot-w8a8-channel-dynamic-token-v2", "channel"),
("nm-testing/tinyllama-oneshot-w8a8-channel-dynamic-token-v2-asym",
"channel"),
])
def test_compressed_tensors_w8a8_dynanmic_per_token(vllm_runner, model_args):
def test_compressed_tensors_w8a8_dynamic_per_token(vllm_runner, model_args):
model_path, strategy = model_args
with vllm_runner(model_path, dtype=torch.float16) as llm:
model = llm.model.llm_engine.model_executor.driver_worker.model_runner.model # noqa: E501
......@@ -160,4 +177,4 @@ def test_compressed_tensors_kv_cache(vllm_runner):
model_path = "nm-testing/TinyLlama-1.1B-compressed-tensors-kv-cache-scheme"
with vllm_runner(model_path, kv_cache_dtype="fp8") as llm:
output = llm.generate_greedy("Hello world!", max_tokens=20)
assert output
\ No newline at end of file
assert output
tests/quantization/test_ipex_quant.py 0 → 100644
View file @ 6d2051cc
"""Test model set-up and inference for quantized HF models supported
on the CPU backend using IPEX (including AWQ).
Validating the configuration and printing results for manual checking.
Run `pytest tests/quantization/test_ipex_quant.py`.
"""
import pytest
from vllm.platforms import current_platform
MODELS = [
"casperhansen/llama-3-8b-instruct-awq",
]
DTYPE = ["bfloat16"]
@pytest.mark.skipif(not current_platform.is_cpu(),
reason="only supports the CPU backend.")
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("dtype", DTYPE)
def test_ipex_quant(vllm_runner, model, dtype):
with vllm_runner(model, dtype=dtype) as llm:
output = llm.generate_greedy(["The capital of France is"],
max_tokens=32)
assert output
print(output)
tests/samplers/test_beam_search.py
View file @ 6d2051cc
......@@ -33,8 +33,8 @@ def test_beam_search_single_input(
max_tokens)
with vllm_runner(model, dtype=dtype) as vllm_model:
vllm_outputs = vllm_model.generate_beam_search_new(
example_prompts, beam_width, max_tokens)
vllm_outputs = vllm_model.generate_beam_search(example_prompts,
beam_width, max_tokens)
for i in range(len(example_prompts)):
hf_output_ids, hf_output_texts = hf_outputs[i]
......
tests/samplers/test_sampler.py
View file @ 6d2051cc
import itertools
import random
from dataclasses import dataclass
from typing import Dict, List, Optional, Tuple
from unittest.mock import Mock, patch
......@@ -158,26 +159,6 @@ def test_sampler_all_random_seed_deterministic(seed: int, device: str):
assert first_sampler_output == second_sampler_output
@pytest.mark.parametrize("seed", RANDOM_SEEDS)
@pytest.mark.parametrize("device", CUDA_DEVICES)
def test_sampler_all_beam(seed: int, device: str):
set_random_seed(seed)
torch.set_default_device(device)
batch_size = random.randint(1, 256)
_, fake_logits, sampler = _prepare_test(batch_size)
sampling_params = SamplingParams(
temperature=0,
best_of=2,
use_beam_search=True,
)
_do_sample(batch_size, fake_logits, sampler, sampling_params, device)
# no assertion here as I am not sure how to determine whether
# the outputs are expected - in other words, this just tests
# whether there are no exceptions in the sampler
# when handling an all-beam search case.
@pytest.mark.parametrize("seed", RANDOM_SEEDS)
@pytest.mark.parametrize("device", CUDA_DEVICES)
def test_sampler_min_tokens_penalty(seed: int, device: str):
......@@ -433,7 +414,7 @@ def test_sampler_min_tokens_penalty(seed: int, device: str):
sampling_metadata = SamplingMetadata.prepare(
seq_group_metadata_list,
seq_lens=seq_lens if seq_lens else None,
query_lens=seq_lens if seq_lens else None,
query_lens=seq_lens if seq_lens else [1] * batch_size,
device=device,
pin_memory=is_pin_memory_available())
# the logits tensor is modified in-place by the sampler
......@@ -478,7 +459,7 @@ def test_sampler_mixed(seed: int, device: str):
seq_lens: List[int] = []
for i in range(batch_size):
expected: Optional[List[int]] = None
sampling_type = random.randint(0, 3)
sampling_type = random.randint(0, 2)
if sampling_type == 0:
sampling_params = SamplingParams(temperature=0)
expected = [int(torch.argmax(fake_logits[i], dim=-1).item())]
......@@ -497,10 +478,7 @@ def test_sampler_mixed(seed: int, device: str):
for idx in range(n):
fake_logits[i, i + idx] = 1e2
expected = list(range(i, i + n))
else:
sampling_params = SamplingParams(temperature=0,
use_beam_search=True,
best_of=2)
expected_tokens.append(expected)
seq_group_metadata_list.append(
SequenceGroupMetadata(
......@@ -529,9 +507,6 @@ def test_sampler_mixed(seed: int, device: str):
zip(sampler_output, seq_group_metadata_list)):
assert metadata.sampling_params is not None
if metadata.sampling_params.use_beam_search:
continue
if (metadata.sampling_params.seed is not None
and expected_tokens[i] is None):
# Record seeded random result to compare with results of
......@@ -596,8 +571,19 @@ def test_sampler_top_k_top_p(seed: int, device: str):
generation_config = GenerationConfig(top_k=top_k,
top_p=top_p,
do_sample=True)
warpers = generation_model._get_logits_warper(generation_config, device)
assert len(warpers) == 2 # top_p and top_k
@dataclass
class MockConfig:
is_encoder_decoder: bool = False
generation_model.config = MockConfig() # needed by the following method
generation_model._prepare_special_tokens(generation_config, device=device)
processors = generation_model._get_logits_processor(generation_config,
None,
None,
None, [],
device=device)
assert len(processors) == 2 # top_p and top_k
seq_group_metadata_list: List[SequenceGroupMetadata] = []
seq_lens: List[int] = []
......@@ -639,7 +625,7 @@ def test_sampler_top_k_top_p(seed: int, device: str):
assert sample_probs is not None
hf_probs = warpers(torch.zeros_like(fake_logits), fake_logits.clone())
hf_probs = processors(torch.zeros_like(fake_logits), fake_logits.clone())
hf_probs = torch.softmax(hf_probs, dim=-1, dtype=torch.float)
torch.testing.assert_close(hf_probs, sample_probs, rtol=0.0, atol=1e-5)
assert torch.equal(hf_probs.eq(0), sample_probs.eq(0))
......
tests/spec_decode/e2e/test_compatibility.py
View file @ 6d2051cc
......@@ -5,16 +5,11 @@ from vllm import SamplingParams
from .conftest import get_output_from_llm_generator
@pytest.mark.parametrize(
"common_llm_kwargs",
[{
"model": "JackFram/llama-68m",
"speculative_model": "JackFram/llama-68m",
"num_speculative_tokens": 5,
# Required for spec decode.
"use_v2_block_manager": True
}])
@pytest.mark.parametrize("common_llm_kwargs", [{
"model": "JackFram/llama-68m",
"speculative_model": "JackFram/llama-68m",
"num_speculative_tokens": 5,
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [
{
"enable_chunked_prefill": True,
......@@ -44,16 +39,11 @@ def test_spec_decode_xfail_chunked_prefill(test_llm_generator):
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,
# Required for spec decode.
"use_v2_block_manager": True
}])
@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",
[
......@@ -94,33 +84,3 @@ def test_spec_decode_xfail_spec_max_model_len(test_llm_generator):
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": "JackFram/llama-68m",
"speculative_model": "JackFram/llama-68m",
"num_speculative_tokens": 5,
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("test_llm_kwargs", [{}])
@pytest.mark.parametrize("seed", [1])
def test_spec_decode_xfail_block_manager_v1(test_llm_generator):
"""Verify that speculative decoding with block manager v1 fails.
"""
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="Speculative decoding requires usage of the V2"):
get_output_from_llm_generator(test_llm_generator, prompts,
sampling_params)
tests/spec_decode/e2e/test_eagle_correctness.py
View file @ 6d2051cc
......@@ -43,9 +43,6 @@ PRECISION = "float32"
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Required for spec decode.
"use_v2_block_manager": True,
# Print spec metrics.
"disable_log_stats": False,
......@@ -86,9 +83,6 @@ def test_eagle_e2e_greedy_correctness(vllm_runner, common_llm_kwargs,
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Required for spec decode.
"use_v2_block_manager": True,
# Print spec metrics.
"disable_log_stats": False,
......@@ -143,9 +137,6 @@ def test_eagle_e2e_greedy_logprobs(vllm_runner, common_llm_kwargs,
[{
"enforce_eager": False,
# Required for spec decode.
"use_v2_block_manager": True,
# Print spec metrics.
"disable_log_stats": False,
......@@ -191,9 +182,6 @@ def test_eagle_e2e_greedy_correctness_cuda_graph(
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Required for spec decode.
"use_v2_block_manager": True,
# Precision
"dtype": PRECISION,
......@@ -235,9 +223,6 @@ def test_eagle_e2e_greedy_correctness_with_preemption(
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Required for spec decode.
"use_v2_block_manager": True,
# Precision
"dtype": PRECISION,
......@@ -283,9 +268,6 @@ def test_eagle_different_k(vllm_runner, common_llm_kwargs,
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Required for spec decode.
"use_v2_block_manager": True,
# Precision
"dtype": PRECISION,
......
tests/spec_decode/e2e/test_integration.py
View file @ 6d2051cc
......@@ -12,8 +12,6 @@ MAIN_MODEL = "JackFram/llama-68m"
@pytest.mark.parametrize(
"common_llm_kwargs",
[{
# Required for spec decode.
"use_v2_block_manager": True,
# Verify equality when cuda graphs allowed.
"enforce_eager": False,
......@@ -57,9 +55,6 @@ def test_spec_decode_cuda_graph(vllm_runner, common_llm_kwargs,
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Required for spec decode.
"use_v2_block_manager": True,
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [
{
......@@ -102,3 +97,44 @@ def test_speculative_model_quantization_config(vllm_runner, common_llm_kwargs,
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)
tests/spec_decode/e2e/test_integration_dist_tp2.py
View file @ 6d2051cc
......@@ -17,9 +17,6 @@ from .conftest import run_equality_correctness_test_tp
[[
# Skip cuda graph recording for fast test.
"--enforce-eager",
# Required for spec decode.
"--use-v2-block-manager",
"--tensor-parallel-size",
"2"
]])
......@@ -74,9 +71,6 @@ def test_target_model_tp_gt_1(common_llm_kwargs, per_test_common_llm_kwargs,
[[
# Skip cuda graph recording for fast test.
"--enforce-eager",
# Required for spec decode.
"--use_v2_block_manager",
"--tensor_parallel_size",
"2",
......
tests/spec_decode/e2e/test_integration_dist_tp4.py
View file @ 6d2051cc
......@@ -19,9 +19,6 @@ SPEC_MODEL = "JackFram/llama-68m"
[[
# Skip cuda graph recording for fast test.
"--enforce_eager",
# Required for spec decode.
"--use-v2-block-manager",
"--tensor-parallel-size",
"4",
]])
......@@ -71,9 +68,6 @@ def test_draft_model_tp_lt_target_model_tp4(common_llm_kwargs,
# Skip cuda graph recording for fast test.
"--enforce-eager",
# Required for spec decode.
"--use-v2-block-manager",
"--tensor-parallel-size",
"4",
]])
......
tests/spec_decode/e2e/test_logprobs.py
View file @ 6d2051cc
......@@ -14,9 +14,6 @@ from .conftest import run_equality_correctness_test
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Required for spec decode.
"use_v2_block_manager": True,
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
......@@ -67,9 +64,6 @@ def test_logprobs_equality(vllm_runner, common_llm_kwargs,
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Required for spec decode.
"use_v2_block_manager": True
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
......@@ -119,9 +113,6 @@ def test_logprobs_different_k(vllm_runner, common_llm_kwargs,
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Required for spec decode.
"use_v2_block_manager": True
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
......@@ -173,9 +164,6 @@ def test_logprobs_when_skip_speculation(vllm_runner, common_llm_kwargs,
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Required for spec decode.
"use_v2_block_manager": True
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
......@@ -251,8 +239,6 @@ def test_logprobs_temp_1(vllm_runner, common_llm_kwargs,
"model_name": "JackFram/llama-160m",
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Required for spec decode.
"use_v2_block_manager": True,
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
......
tests/spec_decode/e2e/test_medusa_correctness.py
View file @ 6d2051cc
......@@ -45,9 +45,6 @@ PRECISION = "float32"
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Required for spec decode.
"use_v2_block_manager": True,
# Print spec metrics.
"disable_log_stats": False,
......@@ -93,9 +90,6 @@ def test_medusa_e2e_greedy_correctness(vllm_runner, common_llm_kwargs,
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Required for spec decode.
"use_v2_block_manager": True,
# Print spec metrics.
"disable_log_stats": False,
......@@ -151,9 +145,6 @@ def test_medusa_e2e_greedy_logprobs(vllm_runner, common_llm_kwargs,
[{
"enforce_eager": False,
# Required for spec decode.
"use_v2_block_manager": True,
# Print spec metrics.
"disable_log_stats": False,
......@@ -204,9 +195,6 @@ def test_medusa_e2e_greedy_correctness_cuda_graph(
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Required for spec decode.
"use_v2_block_manager": True,
# Precision
"dtype": PRECISION,
......@@ -253,9 +241,6 @@ def test_medusa_e2e_greedy_correctness_with_preemption(
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Required for spec decode.
"use_v2_block_manager": True,
# Precision
"dtype": PRECISION,
......@@ -306,9 +291,6 @@ def test_medusa_different_k(vllm_runner, common_llm_kwargs,
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Required for spec decode.
"use_v2_block_manager": True,
# Precision
"dtype": PRECISION,
......@@ -350,6 +332,52 @@ def test_medusa_disable_queue(vllm_runner, common_llm_kwargs,
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__])
tests/spec_decode/e2e/test_mlp_correctness.py
View file @ 6d2051cc
......@@ -47,9 +47,6 @@ PRECISION = "float32"
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Required for spec decode.
"use_v2_block_manager": True,
# Print spec metrics.
"disable_log_stats": False,
......@@ -94,9 +91,6 @@ def test_mlp_e2e_greedy_correctness(vllm_runner, common_llm_kwargs,
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Required for spec decode.
"use_v2_block_manager": True,
# Print spec metrics.
"disable_log_stats": False,
......@@ -149,9 +143,6 @@ def test_mlp_e2e_greedy_logprobs(vllm_runner, common_llm_kwargs,
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Required for spec decode.
"use_v2_block_manager": True,
# Print spec metrics.
"disable_log_stats": False,
......@@ -195,9 +186,6 @@ def test_mlp_e2e_acceptance_rate(vllm_runner, common_llm_kwargs,
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Required for spec decode.
"use_v2_block_manager": True,
# Print spec metrics.
"disable_log_stats": False,
......@@ -258,9 +246,6 @@ def test_mlp_e2e_seeded_correctness(vllm_runner, common_llm_kwargs,
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Required for spec decode.
"use_v2_block_manager": True,
# Precision
"dtype": PRECISION,
......@@ -311,9 +296,6 @@ def test_mlp_e2e_greedy_correctness_with_preemption(
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Required for spec decode.
"use_v2_block_manager": True,
# Precision
"dtype": PRECISION,
......@@ -366,9 +348,6 @@ def test_mlp_e2e_greedy_correctness_with_padding(
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Required for spec decode.
"use_v2_block_manager": True,
# Precision
"dtype": PRECISION,
......@@ -419,9 +398,6 @@ def test_mlp_different_k(vllm_runner, common_llm_kwargs,
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Required for spec decode.
"use_v2_block_manager": True,
# Precision
"dtype": PRECISION,
......@@ -460,3 +436,43 @@ def test_mlp_disable_queue(vllm_runner, common_llm_kwargs,
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)
tests/spec_decode/e2e/test_multistep_correctness.py
View file @ 6d2051cc
......@@ -55,9 +55,6 @@ from .conftest import (get_output_from_llm_generator,
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Required for spec decode.
"use_v2_block_manager": True,
}])
@pytest.mark.parametrize(
"per_test_common_llm_kwargs",
......@@ -124,9 +121,6 @@ def test_spec_decode_e2e_with_detokenization(test_llm_generator,
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Required for spec decode.
"use_v2_block_manager": True,
# Print spec metrics.
"disable_log_stats": False,
}])
......@@ -190,9 +184,6 @@ def test_spec_decode_e2e_greedy_correctness_tiny_model_bs1(
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Required for spec decode.
"use_v2_block_manager": True,
# Print spec metrics.
"disable_log_stats": False,
}])
......@@ -246,9 +237,6 @@ def test_spec_decode_e2e_greedy_correctness_tiny_model_large_bs(
[{
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Required for spec decode.
"use_v2_block_manager": True
}])
@pytest.mark.parametrize(
"per_test_common_llm_kwargs",
......@@ -303,9 +291,6 @@ def test_spec_decode_e2e_greedy_correctness_tiny_model_large_bs_diff_output_len(
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Required for spec decode.
"use_v2_block_manager": True,
# Print spec metrics.
"disable_log_stats": False,
}])
......@@ -353,9 +338,6 @@ def test_spec_decode_e2e_greedy_correctness_real_model_bs1(
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Required for spec decode.
"use_v2_block_manager": True,
# Print spec metrics.
"disable_log_stats": False,
}])
......@@ -404,9 +386,6 @@ def test_spec_decode_e2e_greedy_correctness_real_model_large_bs(
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Required for spec decode.
"use_v2_block_manager": True
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [
{
......@@ -454,9 +433,6 @@ def test_spec_decode_e2e_greedy_correctness_with_preemption(
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Required for spec decode.
"use_v2_block_manager": True
}])
@pytest.mark.parametrize(
"per_test_common_llm_kwargs",
......@@ -514,9 +490,6 @@ def test_spec_decode_different_block_size(vllm_runner, common_llm_kwargs,
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Required for spec decode.
"use_v2_block_manager": True
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
......@@ -570,9 +543,6 @@ def test_skip_speculation(vllm_runner, common_llm_kwargs,
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Required for spec decode.
"use_v2_block_manager": True
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
......@@ -611,9 +581,6 @@ def test_disable_speculation(vllm_runner, common_llm_kwargs,
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Required for spec decode.
"use_v2_block_manager": True
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
......@@ -660,9 +627,6 @@ def test_many_k(vllm_runner, common_llm_kwargs, per_test_common_llm_kwargs,
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Required for spec decode.
"use_v2_block_manager": True
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
......
tests/spec_decode/e2e/test_ngram_correctness.py
View file @ 6d2051cc
......@@ -35,9 +35,6 @@ from .conftest import run_equality_correctness_test
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Required for spec decode.
"use_v2_block_manager": True,
# Print spec metrics.
"disable_log_stats": False,
}])
......@@ -82,9 +79,6 @@ def test_ngram_e2e_greedy_correctness(vllm_runner, common_llm_kwargs,
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Required for spec decode.
"use_v2_block_manager": True,
# Print spec metrics.
"disable_log_stats": False,
}])
......@@ -145,9 +139,6 @@ def test_ngram_e2e_greedy_logprobs(vllm_runner, common_llm_kwargs,
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Required for spec decode.
"use_v2_block_manager": True
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [
{
......@@ -195,9 +186,6 @@ def test_ngram_e2e_greedy_correctness_with_preemption(
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Required for spec decode.
"use_v2_block_manager": True
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
......@@ -254,9 +242,6 @@ def test_ngram_different_k(vllm_runner, common_llm_kwargs,
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Required for spec decode.
"use_v2_block_manager": True
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
......@@ -292,3 +277,48 @@ def test_ngram_disable_queue(vllm_runner, common_llm_kwargs,
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)
tests/spec_decode/e2e/test_seed.py
View file @ 6d2051cc
......@@ -17,9 +17,6 @@ SPEC_MODEL = "JackFram/llama-160m"
# Skip cuda graph recording for fast test.
"enforce_eager": True,
# Required for spec decode.
"use_v2_block_manager": True,
# speculative model
"speculative_model": "JackFram/llama-160m",
......
tests/spec_decode/test_multi_step_worker.py
View file @ 6d2051cc
......@@ -173,7 +173,6 @@ def test_same_output_for_multi_step():
block_size,
num_gpu_blocks,
seed,
model_runner_cls=TP1DraftModelRunner,
)
worker = create_worker(
......@@ -673,7 +672,10 @@ def test_use_draft_model_runner_advance_step():
worker.model_runner._gpu_advance_step.side_effect = ValueError(
exception_secret)
seq_group_metadata_list, _, _ = create_batch(batch_size, k)
seq_group_metadata_list, _, _ = create_batch(batch_size,
k,
block_size=block_size,
num_gpu_blocks=num_gpu_blocks)
# Fallback (should not call) when num_steps=1.
execute_model_req = ExecuteModelRequest(
......
tests/spec_decode/test_scorer.py 0 → 100644
View file @ 6d2051cc
import random
from typing import List
import pytest
import torch
from vllm.sequence import ExecuteModelRequest
from vllm.spec_decode.batch_expansion import BatchExpansionTop1Scorer
from vllm.spec_decode.interfaces import SpeculativeProposals, SpeculativeScores
from vllm.spec_decode.mqa_scorer import MQAScorer
from vllm.worker.worker import Worker
from .utils import create_batch, create_worker
def create_proposal(propose_lens: List[int], vocab_size: int,
device: str) -> SpeculativeProposals:
batch_size = len(propose_lens)
max_propose_len = max(propose_lens)
proposal_probs = torch.rand((batch_size, max_propose_len, vocab_size),
device=device)
proposal_token_ids = torch.full((batch_size, max_propose_len),
fill_value=-1,
device=device)
for i in range(batch_size):
proposal_token_ids[i][:propose_lens[i]] = torch.argmax(
proposal_probs[i][:propose_lens[i]], dim=-1)
propose_lens = torch.tensor(propose_lens, device=device)
return SpeculativeProposals(proposal_token_ids, proposal_probs,
propose_lens)
def assert_score_equal(score1: SpeculativeScores,
score2: SpeculativeScores) -> None:
assert torch.allclose(score1.probs, score2.probs)
assert torch.allclose(score1.logprobs, score2.logprobs)
assert torch.equal(
score1.token_ids,
score2.token_ids), f"{score1.token_ids}, {score2.token_ids}"
@pytest.mark.parametrize('model_name', ['facebook/opt-125m'])
@pytest.mark.parametrize('batch_size', [1, 2, 4, 8, 16])
@pytest.mark.parametrize('max_propose_len', [1, 3, 5])
@pytest.mark.parametrize('mixed_propose_len', [True])
@pytest.mark.parametrize('device', ['cuda'])
def test_scorer(model_name: str, batch_size: int, max_propose_len: int,
mixed_propose_len: bool, device: str) -> None:
"""
Compare the batch expansion scorer and mqa scorer return the same score.
We test for both queries with the same propose length and different
propose length.
"""
seed = 0
block_size = 32
num_gpu_blocks = 2048 // block_size
scorer_worker = create_worker(Worker, model_name, block_size,
num_gpu_blocks, seed)
scorer_worker.model_runner.model.sampler.include_gpu_probs_tensor = True
scorer_worker.model_runner.model.sampler.\
should_modify_greedy_probs_inplace = True
vocab_size = scorer_worker.vocab_size
if not mixed_propose_len:
propose_lens = [max_propose_len] * batch_size
else:
non_zero_cnt = random.randint(0, batch_size)
propose_lens = [max_propose_len
] * non_zero_cnt + [0] * (batch_size - non_zero_cnt)
random.shuffle(propose_lens)
proposals = create_proposal(propose_lens, vocab_size, device)
seq_group_metadatalist, _, _ = create_batch(batch_size,
max_propose_len,
block_size=block_size,
num_gpu_blocks=num_gpu_blocks)
requests = ExecuteModelRequest(seq_group_metadatalist,
num_lookahead_slots=max_propose_len)
batch_expansion_scorer = BatchExpansionTop1Scorer(scorer_worker, device,
vocab_size)
batch_expansion_score = batch_expansion_scorer.score_proposals(
requests, proposals)
mqa_scorer = MQAScorer(scorer_worker, device, vocab_size)
mqa_score = mqa_scorer.score_proposals(requests, proposals)
assert_score_equal(batch_expansion_score, mqa_score)
tests/spec_decode/test_spec_decode_worker.py
View file @ 6d2051cc
......@@ -63,10 +63,10 @@ def test_correctly_calls_draft_model(k: int, batch_size: int,
@pytest.mark.parametrize("acceptance_sampler_method",
["rejection_sampler", "typical_acceptance_sampler"])
@torch.inference_mode()
def test_correctly_calls_target_model(k: int, batch_size: int,
acceptance_sampler_method: str):
def test_batch_expansion_correctly_calls_target_model(
k: int, batch_size: int, acceptance_sampler_method: str):
"""Verify SpecDecodeWorker calls the target model with correct
inputs. Everything else is mocked out.
inputs with batch expansion. Everything else is mocked out.
"""
draft_worker = mock_worker(cls=MultiStepWorker, use_spec=False)
target_worker = mock_worker(use_spec=False)
......@@ -82,7 +82,8 @@ def test_correctly_calls_target_model(k: int, batch_size: int,
target_worker,
mock_spec_decode_sampler(acceptance_sampler_method),
disable_logprobs=False,
metrics_collector=metrics_collector)
metrics_collector=metrics_collector,
disable_mqa_scorer=True)
worker.init_device()
vocab_size = 32_000
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment