[CI/Build] Reorganize models tests (#7820)

This commit is contained in:
Cyrus Leung
2024-09-13 10:20:06 -07:00
committed by GitHub
parent 0a4806f0a9
commit a84e598e21
55 changed files with 415 additions and 498 deletions
@@ -0,0 +1,200 @@
from typing import List, Optional, Tuple, Type
import numpy as np
import pytest
from transformers import AutoModel, AutoTokenizer, BatchEncoding
from vllm.sequence import SampleLogprobs
from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE
from ....conftest import HfRunner, VllmRunner
from ...utils import check_logprobs_close
MODEL_NAME = "fixie-ai/ultravox-v0_3"
AudioTuple = Tuple[np.ndarray, int]
VLLM_PLACEHOLDER = "<|reserved_special_token_0|>"
HF_PLACEHOLDER = "<|audio|>"
@pytest.fixture(scope="session")
def audio_assets():
from vllm.assets.audio import AudioAsset
return [AudioAsset("mary_had_lamb"), AudioAsset("winning_call")]
@pytest.fixture(scope="module", params=("mary_had_lamb", "winning_call"))
def audio(request):
from vllm.assets.audio import AudioAsset
return AudioAsset(request.param)
def _get_prompt(audio_count, question, placeholder):
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
placeholder = f"{placeholder}\n" * audio_count
return tokenizer.apply_chat_template([{
'role': 'user',
'content': f"{placeholder}{question}"
}],
tokenize=False,
add_generation_prompt=True)
def vllm_to_hf_output(vllm_output: Tuple[List[int], str,
Optional[SampleLogprobs]],
model: str):
"""Sanitize vllm output to be comparable with hf output."""
output_ids, output_str, out_logprobs = vllm_output
tokenizer = AutoTokenizer.from_pretrained(model)
eos_token_id = tokenizer.eos_token_id
hf_output_ids = output_ids[:]
hf_output_str = output_str
if hf_output_ids[-1] == eos_token_id:
hf_output_str = hf_output_str + tokenizer.decode(eos_token_id)
return hf_output_ids, hf_output_str, out_logprobs
def run_test(
hf_runner: Type[HfRunner],
vllm_runner: Type[VllmRunner],
prompts_and_audios: List[Tuple[str, str, AudioTuple]],
model: str,
*,
dtype: str,
max_tokens: int,
num_logprobs: int,
tensor_parallel_size: int,
distributed_executor_backend: Optional[str] = None,
):
"""Inference result should be the same between hf and vllm."""
torch_dtype = STR_DTYPE_TO_TORCH_DTYPE[dtype]
# NOTE: take care of the order. run vLLM first, and then run HF.
# vLLM needs a fresh new process without cuda initialization.
# if we run HF first, the cuda initialization will be done and it
# will hurt multiprocessing backend with fork method (the default method).
with vllm_runner(model,
dtype=dtype,
tensor_parallel_size=tensor_parallel_size,
distributed_executor_backend=distributed_executor_backend,
enforce_eager=True) as vllm_model:
vllm_outputs_per_audio = [
vllm_model.generate_greedy_logprobs([vllm_prompt],
max_tokens,
num_logprobs=num_logprobs,
audios=[audio])
for vllm_prompt, _, audio in prompts_and_audios
]
def process(hf_inputs: BatchEncoding):
hf_inputs["audio_values"] = hf_inputs["audio_values"] \
.to(torch_dtype) # type: ignore
return hf_inputs
with hf_runner(model,
dtype=dtype,
postprocess_inputs=process,
auto_cls=AutoModel) as hf_model:
import librosa
hf_outputs_per_audio = [
hf_model.generate_greedy_logprobs_limit(
[hf_prompt],
max_tokens,
num_logprobs=num_logprobs,
audios=[(librosa.resample(audio[0],
orig_sr=audio[1],
target_sr=16000), 16000)])
for _, hf_prompt, audio in prompts_and_audios
]
for hf_outputs, vllm_outputs in zip(hf_outputs_per_audio,
vllm_outputs_per_audio):
check_logprobs_close(
outputs_0_lst=hf_outputs,
outputs_1_lst=[
vllm_to_hf_output(vllm_output, model)
for vllm_output in vllm_outputs
],
name_0="hf",
name_1="vllm",
)
def run_multi_audio_test(
vllm_runner: Type[VllmRunner],
prompts_and_audios: List[Tuple[str, List[AudioTuple]]],
model: str,
*,
dtype: str,
max_tokens: int,
num_logprobs: int,
tensor_parallel_size: int,
distributed_executor_backend: Optional[str] = None,
):
with vllm_runner(model,
dtype=dtype,
tensor_parallel_size=tensor_parallel_size,
distributed_executor_backend=distributed_executor_backend,
enforce_eager=True,
limit_mm_per_prompt={
"audio":
max((len(audio) for _, audio in prompts_and_audios))
}) as vllm_model:
vllm_outputs = vllm_model.generate_greedy_logprobs(
[prompt for prompt, _ in prompts_and_audios],
max_tokens,
num_logprobs=num_logprobs,
audios=[audios for _, audios in prompts_and_audios])
# The HuggingFace model doesn't support multiple audios yet, so
# just assert that some tokens were generated.
assert all(tokens for tokens, *_ in vllm_outputs)
@pytest.mark.parametrize("dtype", ["half"])
@pytest.mark.parametrize("max_tokens", [128])
@pytest.mark.parametrize("num_logprobs", [5])
def test_models(hf_runner, vllm_runner, audio, dtype: str, max_tokens: int,
num_logprobs: int) -> None:
vllm_prompt = _get_prompt(1, "Describe the audio above.", VLLM_PLACEHOLDER)
hf_prompt = _get_prompt(1, "Describe the audio above.", HF_PLACEHOLDER)
run_test(
hf_runner,
vllm_runner,
[(vllm_prompt, hf_prompt, audio.audio_and_sample_rate)],
MODEL_NAME,
dtype=dtype,
max_tokens=max_tokens,
num_logprobs=num_logprobs,
tensor_parallel_size=1,
)
@pytest.mark.parametrize("dtype", ["half"])
@pytest.mark.parametrize("max_tokens", [128])
@pytest.mark.parametrize("num_logprobs", [5])
def test_models_with_multiple_audios(vllm_runner, audio_assets, dtype: str,
max_tokens: int,
num_logprobs: int) -> None:
vllm_prompt = _get_prompt(len(audio_assets),
"Describe each of the audios above.",
VLLM_PLACEHOLDER)
run_multi_audio_test(
vllm_runner,
[(vllm_prompt, [audio.audio_and_sample_rate
for audio in audio_assets])],
MODEL_NAME,
dtype=dtype,
max_tokens=max_tokens,
num_logprobs=num_logprobs,
tensor_parallel_size=1,
)
@@ -0,0 +1,68 @@
"""Compare the outputs of a AQLM model between vLLM and HF Transformers
Run `pytest tests/models/test_aqlm.py`.
"""
import pytest
from tests.quantization.utils import is_quant_method_supported
# These ground truth generations were generated using `transformers==4.38.1
# aqlm==1.1.0 torch==2.2.0`
# and the below code:
# ```python
# from transformers import AutoTokenizer, AutoModelForCausalLM
# model_id = "ISTA-DASLab/Llama-2-7b-AQLM-2Bit-1x16-hf"
# quantized_model = AutoModelForCausalLM.from_pretrained(model_id,
# torch_dtype="auto", device_map="cuda").cuda()
# tokenizer = AutoTokenizer.from_pretrained(model_id)
# outputs = []
# for prompt in example_prompts:
# input_ids = tokenizer(prompt, return_tensors="pt")["input_ids"].to("cuda")
# hf_outputs = quantized_model.generate(input_ids, max_new_tokens=32)
# outputs.append(tokenizer.decode(hf_outputs[0][input_ids.shape[1]:]))
# print(outputs)
# ```
ground_truth_generations = [
'\n### Features\n\n- **High-throughput**: v',
'The major milestones in the development of artificial intelligence from '
'195',
'Compare and contrast artificial intelligence with human intelligence in '
'terms of processing information. The',
'Explain the difference between supervised and unsupervised learning.'
'\nExplain',
'Write a short story about a robot that dreams for the first time. The',
'Analyze the impact of the COVID-19 pandemic on global economic',
'The Mona Lisa is a painting by Leonardo da Vinci, and it',
'The early bird catches the worm.\nThe early bird catches the'
]
@pytest.mark.skipif(not is_quant_method_supported("aqlm"),
reason="AQLM is not supported on this GPU type.")
@pytest.mark.parametrize("model", ["ISTA-DASLab/Llama-2-7b-AQLM-2Bit-1x16-hf"])
@pytest.mark.parametrize("dtype", ["half"])
@pytest.mark.parametrize("max_tokens", [16])
@pytest.mark.parametrize("num_logprobs", [1])
def test_models(
vllm_runner,
example_prompts,
model: str,
dtype: str,
max_tokens: int,
num_logprobs: int,
) -> None:
with vllm_runner(model, dtype=dtype) as vllm_model:
vllm_outputs = vllm_model.generate_greedy_logprobs(
example_prompts, max_tokens, num_logprobs)
# loop through the prompts to compare against the ground truth generations
for prompt_idx in range(len(example_prompts)):
vllm_output_ids, vllm_output_str, vllm_logprobs = vllm_outputs[
prompt_idx]
print("Prompt: ", repr(example_prompts[prompt_idx]))
print("Reference output:", repr(ground_truth_generations[prompt_idx]))
print("Output output: ", repr(vllm_output_str))
assert vllm_output_str == ground_truth_generations[prompt_idx]
@@ -0,0 +1,64 @@
"""Compare the outputs of HF and vLLM when using greedy sampling.
This tests bigger models and use half precision.
Run `pytest tests/models/test_big_models.py`.
"""
import pytest
import torch
from ...utils import check_outputs_equal
MODELS = [
"meta-llama/Llama-2-7b-hf",
# "mistralai/Mistral-7B-v0.1", # Tested by test_mistral.py
# "Deci/DeciLM-7b", # Broken
# "tiiuae/falcon-7b", # Broken
"EleutherAI/gpt-j-6b",
# "mosaicml/mpt-7b", # Broken
# "Qwen/Qwen1.5-0.5B" # Broken,
]
#TODO: remove this after CPU float16 support ready
target_dtype = "float"
if torch.cuda.is_available():
target_dtype = "half"
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("dtype", [target_dtype])
@pytest.mark.parametrize("max_tokens", [32])
def test_models(
hf_runner,
vllm_runner,
example_prompts,
model: str,
dtype: str,
max_tokens: int,
) -> None:
with hf_runner(model, dtype=dtype) as hf_model:
hf_outputs = hf_model.generate_greedy(example_prompts, max_tokens)
with vllm_runner(model, dtype=dtype) as vllm_model:
vllm_outputs = vllm_model.generate_greedy(example_prompts, max_tokens)
check_outputs_equal(
outputs_0_lst=hf_outputs,
outputs_1_lst=vllm_outputs,
name_0="hf",
name_1="vllm",
)
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("dtype", [target_dtype])
def test_model_print(
vllm_runner,
model: str,
dtype: str,
) -> None:
with vllm_runner(model, dtype=dtype) as vllm_model:
# This test is for verifying whether the model's extra_repr
# can be printed correctly.
print(vllm_model.model.llm_engine.model_executor.driver_worker.
model_runner.model)
@@ -0,0 +1,52 @@
"""Compare the outputs of HF and vLLM when using greedy sampling.
This tests danube3 separately because its head size isn't supported on CPU yet.
Run `pytest tests/models/test_danube3_4b.py`.
"""
import pytest
from ...utils import check_outputs_equal
MODELS = ["h2oai/h2o-danube3-4b-base"]
target_dtype = "half"
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("dtype", [target_dtype])
@pytest.mark.parametrize("max_tokens", [32])
def test_models(
hf_runner,
vllm_runner,
example_prompts,
model: str,
dtype: str,
max_tokens: int,
) -> None:
with hf_runner(model, dtype=dtype) as hf_model:
hf_outputs = hf_model.generate_greedy(example_prompts, max_tokens)
with vllm_runner(model, dtype=dtype) as vllm_model:
vllm_outputs = vllm_model.generate_greedy(example_prompts, max_tokens)
check_outputs_equal(
outputs_0_lst=hf_outputs,
outputs_1_lst=vllm_outputs,
name_0="hf",
name_1="vllm",
)
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("dtype", [target_dtype])
def test_model_print(
vllm_runner,
model: str,
dtype: str,
) -> None:
with vllm_runner(model, dtype=dtype) as vllm_model:
# This test is for verifying whether the model's extra_repr
# can be printed correctly.
print(vllm_model.model.llm_engine.model_executor.driver_worker.
model_runner.model)
@@ -0,0 +1,99 @@
# flake8: noqa
"""Tests fp8 models against ground truth generation
Note: these tests will only pass on L4 GPU.
"""
import os
from typing import Optional
import pytest
from tests.kernels.utils import override_backend_env_variable
from tests.quantization.utils import is_quant_method_supported
from ...utils import check_logprobs_close
os.environ["TOKENIZERS_PARALLELISM"] = "true"
@pytest.mark.skipif(not is_quant_method_supported("fp8"),
reason="fp8 is not supported on this GPU type.")
@pytest.mark.parametrize(
"kv_cache_dtype,base_model,test_model,scale_path",
[
# Test FP8 checkpoint w. fp8_e4m3 kv-cache scaling factors.
("fp8_e4m3", "meta-llama/Meta-Llama-3-8B-Instruct",
"nm-testing/Meta-Llama-3-8B-Instruct-FP8-KV", None),
# Test FP16 checkpoint w. fp8_e5m2 kv-cache.
("fp8_e5m2", "meta-llama/Meta-Llama-3-8B-Instruct",
"meta-llama/Meta-Llama-3-8B-Instruct", None),
# Test FP16 checkpoint w. fp8_e4m3 kv-cache scaling factors in json.
("fp8_e4m3", "meta-llama/Llama-2-7b-chat-hf",
"meta-llama/Llama-2-7b-chat-hf",
"./tests/fp8_kv/llama2-7b-fp8-kv/kv_cache_scales.json")
])
# Due to low-precision numerical divergence, we only test logprob of 4 tokens
@pytest.mark.parametrize("max_tokens", [4])
@pytest.mark.parametrize("enforce_eager", [False, True])
@pytest.mark.parametrize("backend", ["FLASH_ATTN", "XFORMERS", "FLASHINFER"])
# NOTE: Increasing this in this suite will fail CI because we currently cannot
# reset distributed env properly. Use a value > 1 just when you test.
@pytest.mark.parametrize("tensor_parallel_size", [1])
# Due to low-precision numerical divergence, this test is too sensitive for
# the async postprocessor
@pytest.mark.parametrize("disable_async_output_proc", [True])
def test_models(
vllm_runner,
example_prompts,
kv_cache_dtype: str,
base_model: str,
test_model: str,
scale_path: Optional[str],
max_tokens: int,
enforce_eager: bool,
backend: str,
tensor_parallel_size: int,
disable_async_output_proc: bool,
monkeypatch,
) -> None:
"""
Only checks log probs match to cover the discrepancy in
numerical sensitive kernels.
"""
override_backend_env_variable(monkeypatch, backend)
MAX_MODEL_LEN = 1024
NUM_LOG_PROBS = 8
with vllm_runner(
base_model,
max_model_len=MAX_MODEL_LEN,
tensor_parallel_size=tensor_parallel_size,
enforce_eager=enforce_eager,
kv_cache_dtype="auto",
disable_async_output_proc=disable_async_output_proc,
) as vllm_model:
baseline_outputs = vllm_model.generate_greedy_logprobs(
example_prompts, max_tokens, NUM_LOG_PROBS)
extra_kwargs = {}
if scale_path is not None:
extra_kwargs["quantization_param_path"] = scale_path
with vllm_runner(
test_model,
max_model_len=MAX_MODEL_LEN,
tensor_parallel_size=tensor_parallel_size,
enforce_eager=enforce_eager,
kv_cache_dtype=kv_cache_dtype,
disable_async_output_proc=disable_async_output_proc,
**extra_kwargs,
) as vllm_model:
test_outputs = vllm_model.generate_greedy_logprobs(
example_prompts, max_tokens, NUM_LOG_PROBS)
check_logprobs_close(
outputs_0_lst=baseline_outputs,
outputs_1_lst=test_outputs,
name_0="fp16_kv_cache",
name_1="fp8_kv_cache",
)
@@ -0,0 +1,90 @@
"""
Tests gguf models against unquantized models generations
Note: To pass the test, quantization higher than Q4 should be used
"""
import os
import pytest
from huggingface_hub import hf_hub_download
from transformers import AutoTokenizer
from tests.quantization.utils import is_quant_method_supported
from ...utils import check_logprobs_close
os.environ["TOKENIZERS_PARALLELISM"] = "true"
MAX_MODEL_LEN = 1024
# FIXME: Move this to confest
MODELS = [
("TinyLlama/TinyLlama-1.1B-Chat-v1.0",
hf_hub_download("TheBloke/TinyLlama-1.1B-Chat-v1.0-GGUF",
filename="tinyllama-1.1b-chat-v1.0.Q4_K_M.gguf")),
("TinyLlama/TinyLlama-1.1B-Chat-v1.0",
hf_hub_download("duyntnet/TinyLlama-1.1B-Chat-v1.0-imatrix-GGUF",
filename="TinyLlama-1.1B-Chat-v1.0-IQ4_XS.gguf")),
("Qwen/Qwen2-1.5B-Instruct",
hf_hub_download("Qwen/Qwen2-1.5B-Instruct-GGUF",
filename="qwen2-1_5b-instruct-q4_k_m.gguf")),
("Qwen/Qwen2-1.5B-Instruct",
hf_hub_download("legraphista/Qwen2-1.5B-Instruct-IMat-GGUF",
filename="Qwen2-1.5B-Instruct.IQ4_XS.gguf")),
]
@pytest.mark.skipif(not is_quant_method_supported("gguf"),
reason="gguf is not supported on this GPU type.")
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("dtype", ["half"])
@pytest.mark.parametrize("max_tokens", [32])
@pytest.mark.parametrize("num_logprobs", [5])
@pytest.mark.parametrize("tp_size", [1, 2])
def test_models(
num_gpus_available,
vllm_runner,
example_prompts,
model,
dtype: str,
max_tokens: int,
num_logprobs: int,
tp_size: int,
) -> None:
if num_gpus_available < tp_size:
pytest.skip(f"Not enough GPUs for tensor parallelism {tp_size}")
original_model, gguf_model = model
tokenizer = AutoTokenizer.from_pretrained(original_model)
messages = [[{
'role': 'user',
'content': prompt
}] for prompt in example_prompts]
example_prompts = tokenizer.apply_chat_template(messages,
tokenize=False,
add_generation_prompt=True)
# Run unquantized model.
with vllm_runner(model_name=original_model,
dtype=dtype,
max_model_len=MAX_MODEL_LEN,
tensor_parallel_size=tp_size) as original_model:
original_outputs = original_model.generate_greedy_logprobs(
example_prompts[:-1], max_tokens, num_logprobs)
# Run gguf model.
with vllm_runner(model_name=gguf_model,
dtype=dtype,
max_model_len=MAX_MODEL_LEN,
tensor_parallel_size=tp_size) as gguf_model:
gguf_outputs = gguf_model.generate_greedy_logprobs(
example_prompts[:-1], max_tokens, num_logprobs)
check_logprobs_close(
outputs_0_lst=original_outputs,
outputs_1_lst=gguf_outputs,
name_0="original",
name_1="gguf",
)
@@ -0,0 +1,96 @@
"""Compares the outputs of gptq vs gptq_marlin
Note: GPTQ and Marlin do not have bitwise correctness.
As a result, in this test, we just confirm that the top selected tokens of the
Marlin/GPTQ models are in the top 5 selections of each other.
Note: Marlin internally uses locks to synchronize the threads. This can
result in very slight nondeterminism for Marlin. As a result, we re-run the test
up to 3 times to see if we pass.
Run `pytest tests/models/test_gptq_marlin.py`.
"""
import os
import pytest
from tests.quantization.utils import is_quant_method_supported
from vllm.model_executor.layers.rotary_embedding import _ROPE_DICT
from ...utils import check_logprobs_close
os.environ["TOKENIZERS_PARALLELISM"] = "true"
MAX_MODEL_LEN = 1024
MODELS = [
# act_order==False, group_size=channelwise
("robertgshaw2/zephyr-7b-beta-channelwise-gptq", "main"),
# act_order==False, group_size=128
("TheBloke/Llama-2-7B-GPTQ", "main"),
# act_order==True, group_size=128
("TheBloke/TinyLlama-1.1B-Chat-v1.0-GPTQ", "main"),
# act_order==True, group_size=64
("TheBloke/TinyLlama-1.1B-Chat-v1.0-GPTQ", "gptq-4bit-64g-actorder_True"),
# act_order==True, group_size=32
("TheBloke/TinyLlama-1.1B-Chat-v1.0-GPTQ", "gptq-4bit-32g-actorder_True"),
# 8-bit, act_order==True, group_size=channelwise
("TheBloke/TinyLlama-1.1B-Chat-v1.0-GPTQ", "gptq-8bit--1g-actorder_True"),
# 8-bit, act_order==True, group_size=128
("TheBloke/TinyLlama-1.1B-Chat-v1.0-GPTQ", "gptq-8bit-128g-actorder_True"),
# 8-bit, act_order==True, group_size=32
("TheBloke/TinyLlama-1.1B-Chat-v1.0-GPTQ", "gptq-8bit-32g-actorder_True"),
# 4-bit, act_order==True, group_size=128
("TechxGenus/gemma-1.1-2b-it-GPTQ", "main")
]
@pytest.mark.flaky(reruns=3)
@pytest.mark.skipif(not is_quant_method_supported("gptq_marlin"),
reason="gptq_marlin is not supported on this GPU type.")
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("dtype", ["half", "bfloat16"])
@pytest.mark.parametrize("max_tokens", [32])
@pytest.mark.parametrize("num_logprobs", [5])
def test_models(
vllm_runner,
example_prompts,
model,
dtype: str,
max_tokens: int,
num_logprobs: int,
) -> None:
model_name, revision = model
# Run marlin.
with vllm_runner(model_name=model_name,
revision=revision,
dtype=dtype,
quantization="marlin",
max_model_len=MAX_MODEL_LEN,
tensor_parallel_size=1) as gptq_marlin_model:
gptq_marlin_outputs = gptq_marlin_model.generate_greedy_logprobs(
example_prompts[:-1], max_tokens, num_logprobs)
_ROPE_DICT.clear() # clear rope cache to avoid rope dtype error
# Run gptq.
# The naive gptq kernel doesn't support bf16 yet.
# Here we always compare fp16/bf16 gpt marlin kernel
# to fp16 gptq kernel.
with vllm_runner(model_name=model_name,
revision=revision,
dtype="half",
quantization="gptq",
max_model_len=MAX_MODEL_LEN,
tensor_parallel_size=1) as gptq_model:
gptq_outputs = gptq_model.generate_greedy_logprobs(
example_prompts[:-1], max_tokens, num_logprobs)
check_logprobs_close(
outputs_0_lst=gptq_outputs,
outputs_1_lst=gptq_marlin_outputs,
name_0="gptq",
name_1="gptq_marlin",
)
@@ -0,0 +1,72 @@
"""Compare the outputs of a GPTQ model to a Marlin_24 model.
Note: GPTQ and Marlin_24 do not have bitwise correctness.
As a result, in this test, we just confirm that the top selected tokens of the
Marlin/GPTQ models are in the top 3 selections of each other.
Run `pytest tests/models/test_marlin_24.py`.
"""
from dataclasses import dataclass
import pytest
from tests.quantization.utils import is_quant_method_supported
from ...utils import check_logprobs_close
@dataclass
class ModelPair:
model_marlin: str
model_gptq: str
model_pairs = [
# 4-bit, group_size == 128
ModelPair(model_marlin="alexm-nm/tinyllama-24-marlin24-4bit-g128",
model_gptq="alexm-nm/tinyllama-24-gptq-4bit-g128"),
# 4-bit, group_size == channelwise
ModelPair(model_marlin="alexm-nm/tinyllama-24-marlin24-4bit-channelwise",
model_gptq="alexm-nm/tinyllama-24-gptq-4bit-channelwise"),
# 8-bit, group_size == 128
ModelPair(model_marlin="alexm-nm/tinyllama-24-marlin24-8bit-g128",
model_gptq="alexm-nm/tinyllama-24-gptq-8bit-g128"),
# 8-bit, group_size == channelwise
ModelPair(model_marlin="alexm-nm/tinyllama-24-marlin24-8bit-channelwise",
model_gptq="alexm-nm/tinyllama-24-gptq-8bit-channelwise"),
]
@pytest.mark.flaky(reruns=2)
@pytest.mark.skipif(not is_quant_method_supported("gptq_marlin_24"),
reason="Marlin24 is not supported on this GPU type.")
@pytest.mark.parametrize("model_pair", model_pairs)
@pytest.mark.parametrize("dtype", ["half"])
@pytest.mark.parametrize("max_tokens", [8])
@pytest.mark.parametrize("num_logprobs", [5])
def test_models(
vllm_runner,
example_prompts,
model_pair: ModelPair,
dtype: str,
max_tokens: int,
num_logprobs: int,
) -> None:
with vllm_runner(model_pair.model_marlin,
dtype=dtype,
quantization="gptq_marlin_24") as marlin_24_model:
marlin_24_outputs = marlin_24_model.generate_greedy_logprobs(
example_prompts, max_tokens, num_logprobs)
with vllm_runner(model_pair.model_gptq, dtype=dtype,
quantization="gptq") as gptq_model:
gptq_outputs = gptq_model.generate_greedy_logprobs(
example_prompts, max_tokens, num_logprobs)
check_logprobs_close(
outputs_0_lst=gptq_outputs,
outputs_1_lst=marlin_24_outputs,
name_0="gptq",
name_1="marlin_24",
)
@@ -0,0 +1,49 @@
"""Compare the outputs of HF and vLLM for Granite models using greedy sampling.
Run `pytest tests/models/test_granite.py`.
"""
import importlib.metadata
import pytest
from ...utils import check_logprobs_close
TRANSFORMERS_VERSION = tuple(
map(int,
importlib.metadata.version("transformers").split(".")))
MODELS = [
"ibm/PowerLM-3b",
]
# GraniteForCausalLM will be in transformers >= 4.45
@pytest.mark.skipif(TRANSFORMERS_VERSION < (4, 45),
reason="granite model test requires transformers >= 4.45")
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("dtype", ["bfloat16"])
@pytest.mark.parametrize("max_tokens", [64])
@pytest.mark.parametrize("num_logprobs", [5])
def test_models(
hf_runner,
vllm_runner,
example_prompts,
model: str,
dtype: str,
max_tokens: int,
num_logprobs: int,
) -> None:
# TODO(sang): Sliding window should be tested separately.
with hf_runner(model, dtype=dtype) as hf_model:
hf_outputs = hf_model.generate_greedy_logprobs_limit(
example_prompts, max_tokens, num_logprobs)
with vllm_runner(model, dtype=dtype) as vllm_model:
vllm_outputs = vllm_model.generate_greedy_logprobs(
example_prompts, max_tokens, num_logprobs)
check_logprobs_close(
outputs_0_lst=hf_outputs,
outputs_1_lst=vllm_outputs,
name_0="hf",
name_1="vllm",
)
@@ -0,0 +1,178 @@
import pytest
from vllm.worker.model_runner import _get_graph_batch_size
from ...utils import check_outputs_equal
MODELS = ["ai21labs/Jamba-tiny-random"]
# Fails due to usage of MoE as MLP(E=1_, which is different than the HF impl
# TODO: Fix this with trained model
@pytest.mark.skip()
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("dtype", ["bfloat16"])
@pytest.mark.parametrize("max_tokens", [10])
def test_models(
hf_runner,
vllm_runner,
example_prompts,
model: str,
dtype: str,
max_tokens: int,
) -> None:
with hf_runner(model, dtype=dtype) as hf_model:
hf_outputs = hf_model.generate_greedy(example_prompts, max_tokens)
with vllm_runner(model, dtype=dtype) as vllm_model:
vllm_outputs = vllm_model.generate_greedy(example_prompts, max_tokens)
for i in range(len(example_prompts)):
hf_output_ids, hf_output_str = hf_outputs[i]
vllm_output_ids, vllm_output_str = vllm_outputs[i]
assert hf_output_str == vllm_output_str, (
f"Test{i}:\nHF: {hf_output_str!r}\nvLLM: {vllm_output_str!r}")
assert hf_output_ids == vllm_output_ids, (
f"Test{i}:\nHF: {hf_output_ids}\nvLLM: {vllm_output_ids}")
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("dtype", ["half"])
@pytest.mark.parametrize("max_tokens", [5])
def test_batching(
vllm_runner,
example_prompts,
model: str,
dtype: str,
max_tokens: int,
) -> None:
# To pass the small model tests, we need full precision.
for_loop_outputs = []
with vllm_runner(model, dtype=dtype) as vllm_model:
for prompt in example_prompts:
for_loop_outputs.append(
vllm_model.generate_greedy([prompt], max_tokens)[0])
batched_outputs = vllm_model.generate_greedy(example_prompts,
max_tokens)
check_outputs_equal(
outputs_0_lst=for_loop_outputs,
outputs_1_lst=batched_outputs,
name_0="for_loop_vllm",
name_1="batched_vllm",
)
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("dtype", ["bfloat16"])
@pytest.mark.parametrize("max_tokens", [20])
def test_mamba_cache_cg_padding(
vllm_runner,
example_prompts,
model: str,
dtype: str,
max_tokens: int,
) -> None:
# This test is for verifying that mamba cache is padded to CG captured
# batch size. If it's not, a torch RuntimeError will be raised because
# tensor dimensions aren't compatible
while len(example_prompts) == _get_graph_batch_size(len(example_prompts)):
example_prompts.append(example_prompts[0])
try:
with vllm_runner(model, dtype=dtype) as vllm_model:
vllm_model.generate_greedy(example_prompts, max_tokens)
except RuntimeError:
pytest.fail(
"Couldn't run batch size which is not equal to a Cuda Graph "
"captured batch size. "
"Could be related to mamba cache not padded correctly")
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("dtype", ["float"])
@pytest.mark.parametrize("max_tokens", [20])
def test_models_preemption_recompute(
hf_runner,
vllm_runner,
example_prompts,
model: str,
dtype: str,
max_tokens: int,
) -> None:
# Tests that outputs are identical with and w/o preemtions (recompute)
assert dtype == "float"
with vllm_runner(model, dtype=dtype) as vllm_model:
vllm_model.model.llm_engine.scheduler[
0].ENABLE_ARTIFICIAL_PREEMPT = True
preempt_vllm_outputs = vllm_model.generate_greedy(
example_prompts, max_tokens)
vllm_model.model.llm_engine.scheduler[
0].ENABLE_ARTIFICIAL_PREEMPT = False
vllm_outputs = vllm_model.generate_greedy(example_prompts, max_tokens)
check_outputs_equal(
outputs_0_lst=preempt_vllm_outputs,
outputs_1_lst=vllm_outputs,
name_0="vllm_preepmtions",
name_1="vllm",
)
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("dtype", ["float"])
def test_fail_upon_inc_requests_and_finished_requests_lt_available_blocks(
vllm_runner,
model: str,
dtype: str,
example_prompts,
) -> None:
# This test is for verifying that the Jamba inner state management doesn't
# collapse in case where the number of incoming requests and
# finished_requests_ids is larger than the maximum mamba block capacity.
# This could generally happen due to the fact that Jamba does support
# statelessness mechanism where it can cleanup new incoming requests in
# a single step.
try:
with vllm_runner(model, dtype=dtype, max_num_seqs=10) as vllm_model:
vllm_model.generate_greedy([example_prompts[0]] * 100, 10)
except ValueError:
pytest.fail("Jamba inner state wasn't cleaned up properly between"
"steps finished requests registered unnecessarily ")
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("dtype", ["float"])
def test_state_cleanup(
vllm_runner,
model: str,
dtype: str,
example_prompts,
) -> None:
# This test is for verifying that the Jamba state is cleaned up between
# steps, If its not cleaned, an error would be expected.
try:
with vllm_runner(model, dtype=dtype) as vllm_model:
for _ in range(10):
vllm_model.generate_greedy([example_prompts[0]] * 100, 1)
except ValueError:
pytest.fail("Jamba inner state wasn't cleaned up between states, "
"could be related to finished_requests_ids")
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("dtype", ["float"])
def test_model_print(
vllm_runner,
model: str,
dtype: str,
) -> None:
with vllm_runner(model, dtype=dtype) as vllm_model:
# This test is for verifying whether the model's extra_repr
# can be printed correctly.
print(vllm_model.model.llm_engine.model_executor.driver_worker.
model_runner.model)
@@ -0,0 +1,69 @@
"""Compare the outputs of a GPTQ model to a Marlin model.
Note: GPTQ and Marlin do not have bitwise correctness.
As a result, in this test, we just confirm that the top selected tokens of the
Marlin/GPTQ models are in the top 3 selections of each other.
Note: Marlin internally uses locks to synchronize the threads. This can
result in very slight nondeterminism for Marlin. As a result, we re-run the test
up to 3 times to see if we pass.
Run `pytest tests/models/test_marlin.py`.
"""
from dataclasses import dataclass
import pytest
from tests.quantization.utils import is_quant_method_supported
from ...utils import check_logprobs_close
@dataclass
class ModelPair:
model_marlin: str
model_gptq: str
model_pairs = [
ModelPair(model_marlin="nm-testing/zephyr-beta-7b-marlin-g128",
model_gptq="nm-testing/zephyr-beta-7b-gptq-g128"),
ModelPair(model_marlin="robertgshaw2/zephyr-7b-beta-channelwise-marlin",
model_gptq="robertgshaw2/zephyr-7b-beta-channelwise-gptq"),
ModelPair(model_marlin="robertgshaw2/TinyLlama-1.1B-Chat-v1.0-g128-marlin",
model_gptq="robertgshaw2/TinyLlama-1.1B-Chat-v1.0-g128-gptq")
]
@pytest.mark.flaky(reruns=2)
@pytest.mark.skipif(not is_quant_method_supported("marlin"),
reason="Marlin is not supported on this GPU type.")
@pytest.mark.parametrize("model_pair", model_pairs)
@pytest.mark.parametrize("dtype", ["half"])
@pytest.mark.parametrize("max_tokens", [32])
@pytest.mark.parametrize("num_logprobs", [5])
def test_models(
vllm_runner,
example_prompts,
model_pair: ModelPair,
dtype: str,
max_tokens: int,
num_logprobs: int,
) -> None:
with vllm_runner(model_pair.model_marlin,
dtype=dtype,
quantization="marlin") as marlin_model:
marlin_outputs = marlin_model.generate_greedy_logprobs(
example_prompts, max_tokens, num_logprobs)
with vllm_runner(model_pair.model_gptq, dtype=dtype,
quantization="gptq") as gptq_model:
gptq_outputs = gptq_model.generate_greedy_logprobs(
example_prompts, max_tokens, num_logprobs)
check_logprobs_close(
outputs_0_lst=gptq_outputs,
outputs_1_lst=marlin_outputs,
name_0="gptq",
name_1="marlin",
)
@@ -0,0 +1,83 @@
"""Compare the outputs of HF and vLLM for Mistral models using greedy sampling.
Run `pytest tests/models/test_mistral.py`.
"""
import pytest
from ...utils import check_logprobs_close
MODELS = [
"mistralai/Mistral-7B-Instruct-v0.1",
"mistralai/Mistral-7B-Instruct-v0.3",
]
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("dtype", ["bfloat16"])
@pytest.mark.parametrize("max_tokens", [64])
@pytest.mark.parametrize("num_logprobs", [5])
def test_models(
hf_runner,
vllm_runner,
example_prompts,
model: str,
dtype: str,
max_tokens: int,
num_logprobs: int,
) -> None:
# TODO(sang): Sliding window should be tested separately.
with hf_runner(model, dtype=dtype) as hf_model:
hf_outputs = hf_model.generate_greedy_logprobs_limit(
example_prompts, max_tokens, num_logprobs)
with vllm_runner(model, dtype=dtype,
tokenizer_mode="mistral") as vllm_model:
vllm_outputs = vllm_model.generate_greedy_logprobs(
example_prompts, max_tokens, num_logprobs)
check_logprobs_close(
outputs_0_lst=hf_outputs,
outputs_1_lst=vllm_outputs,
name_0="hf",
name_1="vllm",
)
@pytest.mark.parametrize("model", MODELS[1:])
@pytest.mark.parametrize("dtype", ["bfloat16"])
@pytest.mark.parametrize("max_tokens", [64])
@pytest.mark.parametrize("num_logprobs", [5])
def test_mistral_format(
vllm_runner,
example_prompts,
model: str,
dtype: str,
max_tokens: int,
num_logprobs: int,
) -> None:
with vllm_runner(
model,
dtype=dtype,
tokenizer_mode="auto",
load_format="safetensors",
config_format="hf",
) as hf_format_model:
hf_format_outputs = hf_format_model.generate_greedy_logprobs(
example_prompts, max_tokens, num_logprobs)
with vllm_runner(
model,
dtype=dtype,
tokenizer_mode="mistral",
load_format="mistral",
config_format="mistral",
) as mistral_format_model:
mistral_format_outputs = mistral_format_model.generate_greedy_logprobs(
example_prompts, max_tokens, num_logprobs)
check_logprobs_close(
outputs_0_lst=hf_format_outputs,
outputs_1_lst=mistral_format_outputs,
name_0="hf",
name_1="mistral",
)
@@ -0,0 +1,79 @@
# flake8: noqa
"""Tests Model Optimizer fp8 models against ground truth generation
Note: these tests will only pass on H100
"""
import os
from typing import List
import pytest
from transformers import AutoTokenizer
from tests.quantization.utils import is_quant_method_supported
from vllm import LLM, SamplingParams
os.environ["TOKENIZERS_PARALLELISM"] = "true"
MAX_MODEL_LEN = 1024
MODELS = ["nvidia/Llama-3.1-8B-Instruct-FP8"]
EXPECTED_STRS_MAP = {
"nvidia/Llama-3.1-8B-Instruct-FP8": [
"You're referring to VLLM, a high-performance Large Language Model (LLM) inference and",
'Here are the major milestones in the development of artificial intelligence (AI) from 1950 to ',
'The comparison between artificial intelligence (AI) and human intelligence in terms of processing information is a complex and',
'A neural network is a complex system modeled after the human brain, consisting of interconnected nodes or "ne',
'**The Spark of Imagination**\n\nZeta-5, a sleek and efficient robot, whir',
'The COVID-19 pandemic has had a profound impact on global economic structures and business models, leading to',
'The Mona Lisa, painted by Leonardo da Vinci in the early 16th century, is one of',
'Here are the translations:\n\n**Japanese:** 「早起きは早く獲物をとる'
]
}
# This test compares against golden strings for exact match since
# there is no baseline implementation to compare against
# and is unstable w.r.t specifics of the fp8 implementation or
# the hardware being run on.
# Disabled to prevent it from breaking the build
@pytest.mark.skip(
reason=
"Prevent unstable test based on golden strings from breaking the build.")
@pytest.mark.skipif(not is_quant_method_supported("fp8"),
reason="fp8 is not supported on this GPU type.")
@pytest.mark.parametrize("model_name", MODELS)
def test_models(example_prompts, model_name) -> None:
model = LLM(
model=model_name,
max_model_len=MAX_MODEL_LEN,
trust_remote_code=True,
enforce_eager=True,
quantization="modelopt",
)
tokenizer = AutoTokenizer.from_pretrained(model_name)
formatted_prompts = [
tokenizer.apply_chat_template([{
"role": "user",
"content": prompt
}],
tokenize=False,
add_generation_prompt=True)
for prompt in example_prompts
]
params = SamplingParams(max_tokens=20, temperature=0)
generations: List[str] = []
# Note: these need to be run 1 at a time due to numerical precision,
# since the expected strs were generated this way.
for prompt in formatted_prompts:
outputs = model.generate(prompt, params)
generations.append(outputs[0].outputs[0].text)
del model
print(model_name, generations)
expected_strs = EXPECTED_STRS_MAP[model_name]
for i in range(len(example_prompts)):
generated_str = generations[i]
expected_str = expected_strs[i]
assert expected_str == generated_str, (
f"Test{i}:\nExpected: {expected_str!r}\nvLLM: {generated_str!r}")
@@ -0,0 +1,65 @@
"""Compare the outputs of HF and vLLM when using greedy sampling.
This test only tests small models. Big models such as 7B should be tested from
test_big_models.py because it could use a larger instance to run tests.
Run `pytest tests/models/test_models.py`.
"""
import pytest
from ...utils import check_outputs_equal
MODELS = [
"facebook/opt-125m",
"gpt2",
"bigcode/tiny_starcoder_py",
"EleutherAI/pythia-70m",
"bigscience/bloom-560m", # Testing alibi slopes.
"microsoft/phi-2",
"stabilityai/stablelm-3b-4e1t",
# "allenai/OLMo-1B", # Broken
"bigcode/starcoder2-3b",
"google/gemma-1.1-2b-it",
]
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("dtype", ["float"])
@pytest.mark.parametrize("max_tokens", [96])
def test_models(
hf_runner,
vllm_runner,
example_prompts,
model: str,
dtype: str,
max_tokens: int,
) -> None:
# To pass the small model tests, we need full precision.
assert dtype == "float"
with hf_runner(model, dtype=dtype) as hf_model:
hf_outputs = hf_model.generate_greedy(example_prompts, max_tokens)
with vllm_runner(model, dtype=dtype) as vllm_model:
vllm_outputs = vllm_model.generate_greedy(example_prompts, max_tokens)
check_outputs_equal(
outputs_0_lst=hf_outputs,
outputs_1_lst=vllm_outputs,
name_0="hf",
name_1="vllm",
)
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("dtype", ["float"])
def test_model_print(
vllm_runner,
model: str,
dtype: str,
) -> None:
with vllm_runner(model, dtype=dtype) as vllm_model:
# This test is for verifying whether the model's extra_repr
# can be printed correctly.
print(vllm_model.model.llm_engine.model_executor.driver_worker.
model_runner.model)
@@ -0,0 +1,111 @@
"""Compare the outputs of HF and vLLM for moe models using greedy sampling.
Run `pytest tests/models/test_phimoe.py`.
"""
import pytest
import torch
from vllm.utils import is_cpu
from ...utils import check_logprobs_close
MODELS = [
"microsoft/Phi-3.5-MoE-instruct",
]
def test_phimoe_routing_function():
from vllm.model_executor.models.phimoe import phimoe_routing_function
test_case = {
0: {
"hidden_states":
torch.tensor([1, 2, 3, 4, 5, 6, 7, 8],
dtype=torch.float32,
requires_grad=False).view(4, 2),
"gating_output":
torch.tensor([0.1, 0.2, 0.3, 0.4],
dtype=torch.float32,
requires_grad=False),
"topk":
2,
"renormalize":
False,
},
1: {
"hidden_states":
torch.tensor([1, 2, 3, 4, 5, 6, 7, 8],
dtype=torch.float32,
requires_grad=False).view(4, 2),
"gating_output":
torch.tensor([0.4, 0.2, 0.3, 0.4],
dtype=torch.float32,
requires_grad=False),
"topk":
2,
"renormalize":
False,
}
}
ground_truth = {
0: {
"topk_weights":
torch.tensor([1., 1.], dtype=torch.float32, requires_grad=False),
"topk_ids":
torch.tensor([3, 2], dtype=torch.long, requires_grad=False),
},
1: {
"topk_weights":
torch.tensor([0.5, 1.], dtype=torch.float32, requires_grad=False),
"topk_ids":
torch.tensor([0, 3], dtype=torch.long, requires_grad=False),
}
}
for test_id in test_case:
topk_weights, topk_ids = phimoe_routing_function(**test_case[test_id])
assert torch.allclose(topk_weights,
ground_truth[test_id]["topk_weights"])
assert torch.equal(topk_ids, ground_truth[test_id]["topk_ids"])
def get_gpu_memory():
try:
props = torch.cuda.get_device_properties(torch.cuda.current_device())
gpu_memory = props.total_memory / (1024**3)
return gpu_memory
except Exception:
return 0
@pytest.mark.skipif(condition=is_cpu(),
reason="This test takes a lot time to run on CPU, "
"and vllm CI's disk space is not enough for this model.")
@pytest.mark.skipif(condition=get_gpu_memory() < 100,
reason="Skip this test if GPU memory is insufficient.")
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("dtype", ["bfloat16"])
@pytest.mark.parametrize("max_tokens", [64])
@pytest.mark.parametrize("num_logprobs", [5])
def test_models(
hf_runner,
vllm_runner,
example_prompts,
model: str,
dtype: str,
max_tokens: int,
num_logprobs: int,
) -> None:
with hf_runner(model, dtype=dtype) as hf_model:
hf_outputs = hf_model.generate_greedy_logprobs_limit(
example_prompts, max_tokens, num_logprobs)
with vllm_runner(model, dtype=dtype) as vllm_model:
vllm_outputs = vllm_model.generate_greedy_logprobs(
example_prompts, max_tokens, num_logprobs)
check_logprobs_close(
outputs_0_lst=hf_outputs,
outputs_1_lst=vllm_outputs,
name_0="hf",
name_1="vllm",
)
@@ -0,0 +1,101 @@
from typing import List, Optional, Tuple
import pytest
from transformers import AutoModelForVision2Seq, AutoTokenizer
from vllm.multimodal.utils import rescale_image_size
from vllm.sequence import SampleLogprobs
from ....conftest import IMAGE_ASSETS
from ...utils import check_logprobs_close
HF_IMAGE_PROMPTS = IMAGE_ASSETS.prompts({
"stop_sign":
"Question: What's the content of the image? Answer:",
"cherry_blossom":
"Question: What is the season? Answer:",
})
def vllm_to_hf_output(vllm_output: Tuple[List[int], str,
Optional[SampleLogprobs]],
model: str):
"""Sanitize vllm output to be comparable with hf output."""
_, output_str, out_logprobs = vllm_output
hf_output_str = output_str + "\n"
tokenizer = AutoTokenizer.from_pretrained(model)
hf_output_ids = tokenizer.encode(hf_output_str)
assert hf_output_ids[0] == tokenizer.bos_token_id
hf_output_ids = hf_output_ids[1:]
return hf_output_ids, hf_output_str, out_logprobs
@pytest.mark.parametrize("model", ["Salesforce/blip2-opt-2.7b"])
@pytest.mark.parametrize(
"size_factors",
[
# No image
[],
# Single-scale
[1.0],
# Single-scale, batched
[1.0, 1.0, 1.0],
# Multi-scale
[0.25, 0.5, 1.0],
],
)
@pytest.mark.parametrize("dtype", ["half"])
@pytest.mark.parametrize("max_tokens", [128])
@pytest.mark.parametrize("num_logprobs", [5])
def test_models(hf_runner, vllm_runner, image_assets, model, size_factors,
dtype: str, max_tokens: int, num_logprobs: int) -> None:
"""Inference result should be the same between hf and vllm.
All the image fixtures for the test are from IMAGE_ASSETS.
For huggingface runner, we provide the PIL images as input.
For vllm runner, we provide MultiModalData objects and corresponding
MultiModalConfig as input.
Note, the text input is also adjusted to abide by vllm contract.
The text output is sanitized to be able to compare with hf.
"""
images = [asset.pil_image for asset in image_assets]
inputs_per_image = [(
[prompt for _ in size_factors],
[rescale_image_size(image, factor) for factor in size_factors],
) for image, prompt in zip(images, HF_IMAGE_PROMPTS)]
# max_model_len should be greater than image_feature_size
with vllm_runner(model, dtype=dtype, enforce_eager=True) as vllm_model:
vllm_outputs_per_image = [
vllm_model.generate_greedy_logprobs(prompts,
max_tokens,
num_logprobs=num_logprobs,
images=images)
for prompts, images in inputs_per_image
]
with hf_runner(model, dtype=dtype,
auto_cls=AutoModelForVision2Seq) as hf_model:
hf_outputs_per_image = [
hf_model.generate_greedy_logprobs_limit(prompts,
max_tokens,
num_logprobs=num_logprobs,
images=images)
for prompts, images in inputs_per_image
]
for hf_outputs, vllm_outputs in zip(hf_outputs_per_image,
vllm_outputs_per_image):
check_logprobs_close(
outputs_0_lst=hf_outputs,
outputs_1_lst=[
vllm_to_hf_output(vllm_output, model)
for vllm_output in vllm_outputs
],
name_0="hf",
name_1="vllm",
)
@@ -0,0 +1,42 @@
import pytest
from ....utils import multi_gpu_test
@multi_gpu_test(num_gpus=2)
@pytest.mark.parametrize("distributed_executor_backend", ["ray", "mp"])
@pytest.mark.parametrize("model", [
"llava-hf/llava-1.5-7b-hf",
"llava-hf/llava-v1.6-mistral-7b-hf",
"facebook/chameleon-7b",
])
def test_models(hf_runner, vllm_runner, image_assets,
distributed_executor_backend, model) -> None:
dtype = "half"
max_tokens = 5
num_logprobs = 5
tensor_parallel_size = 2
if model.startswith("llava-hf/llava-1.5"):
from .test_llava import models, run_test
elif model.startswith("llava-hf/llava-v1.6"):
from .test_llava_next import models, run_test # type: ignore[no-redef]
elif model.startswith("facebook/chameleon"):
from .test_chameleon import models, run_test # type: ignore[no-redef]
else:
raise NotImplementedError(f"Unsupported model: {model}")
run_test(
hf_runner,
vllm_runner,
image_assets,
model=models[0],
# So that LLaVA-NeXT processor may return nested list
size_factors=[0.25, 0.5, 1.0],
dtype=dtype,
max_tokens=max_tokens,
num_logprobs=num_logprobs,
tensor_parallel_size=tensor_parallel_size,
distributed_executor_backend=distributed_executor_backend,
)
@@ -0,0 +1,125 @@
from typing import List, Optional, Type
import pytest
from transformers import AutoModelForVision2Seq, BatchEncoding
from vllm.multimodal.utils import rescale_image_size
from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE
from ....conftest import IMAGE_ASSETS, HfRunner, VllmRunner, _ImageAssets
from ...utils import check_outputs_equal
HF_IMAGE_PROMPTS = IMAGE_ASSETS.prompts({
"stop_sign":
"USER: <image>\nWhat's the content of the image?\nASSISTANT:",
"cherry_blossom":
"USER: <image>\nWhat is the season?\nASSISTANT:",
})
models = ["facebook/chameleon-7b"]
def run_test(
hf_runner: Type[HfRunner],
vllm_runner: Type[VllmRunner],
image_assets: _ImageAssets,
model: str,
*,
size_factors: List[float],
dtype: str,
max_tokens: int,
num_logprobs: int,
tensor_parallel_size: int,
distributed_executor_backend: Optional[str] = None,
):
"""Inference result should be the same between hf and vllm.
All the image fixtures for the test are from IMAGE_ASSETS.
For huggingface runner, we provide the PIL images as input.
For vllm runner, we provide MultiModalDataDict objects
and corresponding vision language config as input.
Note, the text input is also adjusted to abide by vllm contract.
The text output is sanitized to be able to compare with hf.
"""
torch_dtype = STR_DTYPE_TO_TORCH_DTYPE[dtype]
images = [asset.pil_image for asset in image_assets]
inputs_per_image = [(
[prompt for _ in size_factors],
[rescale_image_size(image, factor) for factor in size_factors],
) for image, prompt in zip(images, HF_IMAGE_PROMPTS)]
with vllm_runner(model,
max_model_len=4096,
dtype=dtype,
tensor_parallel_size=tensor_parallel_size,
distributed_executor_backend=distributed_executor_backend,
enforce_eager=True) as vllm_model:
vllm_outputs_per_image = [
vllm_model.generate_greedy_logprobs(prompts,
max_tokens,
num_logprobs=num_logprobs,
images=images)
for prompts, images in inputs_per_image
]
def process(hf_inputs: BatchEncoding):
hf_inputs["pixel_values"] = hf_inputs["pixel_values"] \
.to(torch_dtype) # type: ignore
return hf_inputs
with hf_runner(model,
dtype=dtype,
postprocess_inputs=process,
auto_cls=AutoModelForVision2Seq) as hf_model:
hf_outputs_per_image = [
hf_model.generate_greedy_logprobs_limit(prompts,
max_tokens,
num_logprobs=num_logprobs,
images=images)
for prompts, images in inputs_per_image
]
for hf_outputs, vllm_outputs in zip(hf_outputs_per_image,
vllm_outputs_per_image):
# HF Logprobs include image tokens, unlike vLLM, so we don't directly
# compare them
check_outputs_equal(
outputs_0_lst=[outputs[:2] for outputs in hf_outputs],
outputs_1_lst=[outputs[:2] for outputs in vllm_outputs],
name_0="hf",
name_1="vllm",
)
@pytest.mark.parametrize("model", models)
@pytest.mark.parametrize(
"size_factors",
[
# No image
[],
# Single-scale
[1.0],
# Single-scale, batched
[1.0, 1.0, 1.0],
# Multi-scale
[0.25, 0.5, 1.0],
],
)
@pytest.mark.parametrize("dtype", ["bfloat16"])
@pytest.mark.parametrize("max_tokens", [8])
@pytest.mark.parametrize("num_logprobs", [5])
def test_models(hf_runner, vllm_runner, image_assets, model, size_factors,
dtype, max_tokens, num_logprobs) -> None:
run_test(
hf_runner,
vllm_runner,
image_assets,
model,
size_factors=size_factors,
dtype=dtype,
max_tokens=max_tokens,
num_logprobs=num_logprobs,
tensor_parallel_size=1,
)
@@ -0,0 +1,141 @@
from typing import List, Optional, Tuple, Type
import pytest
from vllm.multimodal.utils import rescale_image_size
from vllm.sequence import SampleLogprobs
from vllm.utils import is_cpu
from ....conftest import IMAGE_ASSETS, HfRunner, VllmRunner, _ImageAssets
from ...utils import check_logprobs_close
HF_IMAGE_PROMPTS = IMAGE_ASSETS.prompts({
"stop_sign":
"What's the content of the image?\n",
"cherry_blossom":
"What is the season?\n",
})
models = ["adept/fuyu-8b"]
def vllm_to_hf_output(vllm_output: Tuple[List[int], str,
Optional[SampleLogprobs]]):
"""Sanitize vllm output to be comparable with hf output."""
output_ids, output_str, out_logprobs = vllm_output
hf_output_str = output_str.lstrip() + "|ENDOFTEXT|"
return output_ids, hf_output_str, out_logprobs
def run_test(
hf_runner: Type[HfRunner],
vllm_runner: Type[VllmRunner],
image_assets: _ImageAssets,
model: str,
*,
size_factors: List[float],
dtype: str,
max_tokens: int,
num_logprobs: int,
tensor_parallel_size: int,
distributed_executor_backend: Optional[str] = None,
):
"""Inference result should be the same between hf and vllm.
All the image fixtures for the test are from IMAGE_ASSETS.
For huggingface runner, we provide the PIL images as input.
For vllm runner, we provide MultiModalDataDict objects
and corresponding MultiModalConfig as input.
Note, the text input is also adjusted to abide by vllm contract.
The text output is sanitized to be able to compare with hf.
"""
images = [asset.pil_image for asset in image_assets]
inputs_per_image = [(
[prompt for _ in size_factors],
[rescale_image_size(image, factor) for factor in size_factors],
) for image, prompt in zip(images, HF_IMAGE_PROMPTS)]
# NOTE: take care of the order. run vLLM first, and then run HF.
# vLLM needs a fresh new process without cuda initialization.
# if we run HF first, the cuda initialization will be done and it
# will hurt multiprocessing backend with fork method (the default method).
# max_model_len should be greater than image_feature_size
with vllm_runner(model,
max_model_len=2560,
max_num_seqs=1,
dtype=dtype,
tensor_parallel_size=tensor_parallel_size,
distributed_executor_backend=distributed_executor_backend,
enforce_eager=True) as vllm_model:
vllm_outputs_per_image = [
vllm_model.generate_greedy_logprobs(prompts,
max_tokens,
num_logprobs=num_logprobs,
images=images)
for prompts, images in inputs_per_image
]
with hf_runner(model, dtype=dtype) as hf_model:
hf_model.model.get_output_embeddings = lambda: \
hf_model.model.language_model.get_output_embeddings()
eos_token_id = hf_model.processor.tokenizer.eos_token_id
hf_outputs_per_image = [
hf_model.generate_greedy_logprobs_limit(prompts,
max_tokens,
num_logprobs=num_logprobs,
images=images,
eos_token_id=eos_token_id)
for prompts, images in inputs_per_image
]
for hf_outputs, vllm_outputs in zip(hf_outputs_per_image,
vllm_outputs_per_image):
check_logprobs_close(
outputs_0_lst=hf_outputs,
outputs_1_lst=[
vllm_to_hf_output(vllm_output) for vllm_output in vllm_outputs
],
name_0="hf",
name_1="vllm",
)
target_dtype = "half"
if is_cpu():
target_dtype = "bfloat16"
@pytest.mark.parametrize("model", models)
@pytest.mark.parametrize(
"size_factors",
[
# No image
[],
# Single-scale
[0.25],
# Single-scale, batched
[0.25, 0.25, 0.25],
# Multi-scale
[0.25, 0.2, 0.15],
],
)
@pytest.mark.parametrize("dtype", [target_dtype])
@pytest.mark.parametrize("max_tokens", [128])
@pytest.mark.parametrize("num_logprobs", [10])
def test_models(hf_runner, vllm_runner, image_assets, model, size_factors,
dtype: str, max_tokens: int, num_logprobs: int) -> None:
run_test(
hf_runner,
vllm_runner,
image_assets,
model,
size_factors=size_factors,
dtype=dtype,
max_tokens=max_tokens,
num_logprobs=num_logprobs,
tensor_parallel_size=1,
)
@@ -0,0 +1,77 @@
from typing import Optional
import pytest
import torch
import torch.nn as nn
from huggingface_hub import snapshot_download
from transformers import AutoConfig, AutoModel, CLIPImageProcessor
from ....conftest import _ImageAssets, cleanup
# we use snapshot_download to prevent conflicts between
# dynamic_module and trust_remote_code for hf_runner
DOWNLOAD_PATTERN = ["*.json", "*.py", "*.safetensors", "*.txt", "*.model"]
models = [
snapshot_download("OpenGVLab/InternViT-300M-448px",
allow_patterns=DOWNLOAD_PATTERN),
snapshot_download("OpenGVLab/InternViT-6B-448px-V1-5",
allow_patterns=DOWNLOAD_PATTERN),
]
def run_intern_vit_test(
image_assets: _ImageAssets,
model: str,
*,
dtype: str,
distributed_executor_backend: Optional[str] = None,
):
img_processor = CLIPImageProcessor.from_pretrained(model)
images = [asset.pil_image for asset in image_assets]
pixel_values = [
img_processor(images, return_tensors='pt').pixel_values.to(dtype)
for images in images
]
config = AutoConfig.from_pretrained(model, trust_remote_code=True)
if not getattr(config, "norm_type", None):
config.norm_type = "rms_norm"
hf_model = AutoModel.from_pretrained(model,
torch_dtype=dtype,
trust_remote_code=True).to("cuda")
hf_outputs_per_image = [
hf_model(pixel_value.to("cuda")).last_hidden_state
for pixel_value in pixel_values
]
from vllm.model_executor.models.intern_vit import InternVisionModel
vllm_model = InternVisionModel(config)
vllm_model.load_weights(hf_model.state_dict().items())
del hf_model
cleanup()
vllm_model = vllm_model.to("cuda", dtype)
vllm_outputs_per_image = [
vllm_model(pixel_values=pixel_value.to("cuda"))
for pixel_value in pixel_values
]
del vllm_model
cleanup()
cos_similar = nn.CosineSimilarity(dim=-1)
for vllm_output, hf_output in zip(vllm_outputs_per_image,
hf_outputs_per_image):
assert cos_similar(vllm_output, hf_output).mean() > 0.99
@pytest.mark.parametrize("model", models)
@pytest.mark.parametrize("dtype", [torch.half])
@torch.inference_mode()
def test_models(dist_init, image_assets, model, dtype: str) -> None:
run_intern_vit_test(
image_assets,
model,
dtype=dtype,
)
@@ -0,0 +1,397 @@
import types
from typing import List, Optional, Tuple, Type, Union
import pytest
import torch
from PIL.Image import Image
from transformers import AutoConfig
from vllm.multimodal.utils import rescale_image_size
from vllm.utils import is_cpu
from ....conftest import (IMAGE_ASSETS, HfRunner, PromptImageInput, VllmRunner,
_ImageAssets)
from ...utils import check_logprobs_close
HF_IMAGE_PROMPTS = IMAGE_ASSETS.prompts({
"stop_sign":
"<|im_start|>User\n<image>\nWhat's the content in the center of the image?<|im_end|>\n<|im_start|>Assistant\n", # noqa: E501
"cherry_blossom":
"<|im_start|>User\n<image>\nWhat is the season?<|im_end|>\n<|im_start|>Assistant\n", # noqa: E501
})
HF_MULTIIMAGE_IMAGE_PROMPT = "<|im_start|>User\nImage-1: <image>\nImage-2: <image>\nDescribe the two images in detail.<|im_end|>\n<|im_start|>Assistant\n" # noqa: E501
models = [
"OpenGVLab/InternVL2-1B",
"OpenGVLab/InternVL2-2B",
# Broken due to outdated implementation of Phi-3
# See: https://huggingface.co/OpenGVLab/InternVL2-4B/discussions/3
# "OpenGVLab/InternVL2-4B",
]
# adapted from https://huggingface.co/OpenGVLab/InternVL2-1B/blob/main/modeling_internvl_chat.py
def generate(
self,
pixel_values: torch.FloatTensor,
input_ids: torch.FloatTensor,
attention_mask: Optional[torch.LongTensor] = None,
**generate_kwargs,
) -> torch.LongTensor:
"""Generate method for InternVL2 model without fixed use_cache."""
assert self.img_context_token_id is not None
vit_embeds = self.extract_feature(pixel_values)
input_embeds = self.language_model.get_input_embeddings()(input_ids)
B, N, C = input_embeds.shape
input_embeds = input_embeds.reshape(B * N, C)
input_ids = input_ids.reshape(B * N)
selected = (input_ids == self.img_context_token_id)
assert selected.sum() != 0
input_embeds[selected] = vit_embeds.reshape(-1, C).to(input_embeds.device)
input_embeds = input_embeds.reshape(B, N, C)
outputs = self.language_model.generate(
inputs_embeds=input_embeds,
attention_mask=attention_mask,
**generate_kwargs,
)
return outputs
def run_test(
hf_runner: Type[HfRunner],
vllm_runner: Type[VllmRunner],
inputs: List[Tuple[List[str], PromptImageInput]],
model: str,
*,
dtype: str,
max_tokens: int,
num_logprobs: int,
mm_limit: int,
tensor_parallel_size: int,
distributed_executor_backend: Optional[str] = None,
):
"""Inference result should be the same between hf and vllm.
All the image fixtures for the test are from IMAGE_ASSETS.
For huggingface runner, we provide the PIL images as input.
For vllm runner, we provide MultiModalDataDict objects
and corresponding MultiModalConfig as input.
Note, the text input is also adjusted to abide by vllm contract.
The text output is sanitized to be able to compare with hf.
"""
# NOTE: take care of the order. run vLLM first, and then run HF.
# vLLM needs a fresh new process without cuda initialization.
# if we run HF first, the cuda initialization will be done and it
# will hurt multiprocessing backend with fork method (the default method).
class InternVLProcessor:
"""A simple processor for InternVL2 which misses a processor."""
def __init__(self, hf_runner: HfRunner):
self.num_image_token = hf_runner.model.num_image_token
self.tokenizer = hf_runner.tokenizer
self.dtype = hf_runner.model.dtype
self.config = AutoConfig.from_pretrained(hf_runner.model_name)
self.vision_config = self.config.vision_config
self.use_thumbnail = self.config.use_thumbnail
self.min_num = self.config.min_dynamic_patch
self.max_num = self.config.max_dynamic_patch
self.image_size = self.vision_config.image_size
def __call__(self, text: str, images: Union[Image, List[Image]],
**kwargs):
from vllm.model_executor.models.internvl import (
IMG_CONTEXT, IMG_END, IMG_START, image_to_pixel_values)
images = [images] if isinstance(images, Image) else images
pixel_values = [
image_to_pixel_values(image, self.image_size, self.min_num,
self.max_num,
self.use_thumbnail).to(self.dtype)
for image in images
]
num_patches_list = [
pixel_value.shape[0] for pixel_value in pixel_values
]
pixel_values = torch.cat(pixel_values, dim=0)
for num_patches in num_patches_list:
context_tokens = IMG_CONTEXT * self.num_image_token \
* num_patches
image_tokens = IMG_START + context_tokens + IMG_END
text = text.replace('<image>', image_tokens, 1)
prompt = self.tokenizer(text, return_tensors="pt")
prompt.update({"pixel_values": pixel_values})
return prompt
# max_model_len should be greater than image_feature_size
with vllm_runner(model,
max_model_len=4096,
dtype=dtype,
limit_mm_per_prompt={"image": mm_limit},
tensor_parallel_size=tensor_parallel_size,
distributed_executor_backend=distributed_executor_backend,
enforce_eager=True) as vllm_model:
vllm_outputs_per_image = [
vllm_model.generate_greedy_logprobs(prompts,
max_tokens,
num_logprobs=num_logprobs,
images=images)
for prompts, images in inputs
]
with hf_runner(model, dtype=dtype) as hf_model:
img_context_token_id = hf_model.tokenizer.convert_tokens_to_ids(
"<IMG_CONTEXT>")
hf_model.model.img_context_token_id = img_context_token_id
hf_model.processor = InternVLProcessor(hf_model)
hf_model.model.get_output_embeddings = lambda: \
hf_model.model.language_model.get_output_embeddings()
hf_model.model.generate = types.MethodType(generate, hf_model.model)
eos_token_id = hf_model.tokenizer.eos_token_id
hf_outputs_per_image = [
hf_model.generate_greedy_logprobs_limit(prompts,
max_tokens,
num_logprobs=num_logprobs,
images=hf_images,
eos_token_id=eos_token_id)
for prompts, hf_images in inputs
]
for hf_outputs, vllm_outputs in zip(hf_outputs_per_image,
vllm_outputs_per_image):
# TODO: Check whether using original CLIPVisionModel can improve
# consistency against HF
check_logprobs_close(
outputs_0_lst=hf_outputs,
outputs_1_lst=vllm_outputs,
name_0="hf",
name_1="vllm",
)
def run_awq_test(
vllm_runner: Type[VllmRunner],
image_assets: _ImageAssets,
models: Tuple[str, str],
*,
size_factors: List[float],
dtype: str,
max_tokens: int,
num_logprobs: int,
tensor_parallel_size: int,
distributed_executor_backend: Optional[str] = None,
):
source_model, quant_model = models
images = [asset.pil_image for asset in image_assets]
inputs_per_image = [(
[prompt for _ in size_factors],
[rescale_image_size(image, factor) for factor in size_factors],
) for image, prompt in zip(images, HF_IMAGE_PROMPTS)]
# NOTE: take care of the order. run vLLM first, and then run HF.
# vLLM needs a fresh new process without cuda initialization.
# if we run HF first, the cuda initialization will be done and it
# will hurt multiprocessing backend with fork method (the default method).
# max_model_len should be greater than image_feature_size
with vllm_runner(source_model,
max_model_len=4096,
dtype=dtype,
tensor_parallel_size=tensor_parallel_size,
distributed_executor_backend=distributed_executor_backend,
enforce_eager=True) as vllm_model:
source_outputs_per_image = [
vllm_model.generate_greedy_logprobs(prompts,
max_tokens,
num_logprobs=num_logprobs,
images=images)
for prompts, images in inputs_per_image
]
with vllm_runner(quant_model,
quantization="awq",
max_model_len=4096,
dtype=dtype,
tensor_parallel_size=tensor_parallel_size,
distributed_executor_backend=distributed_executor_backend,
enforce_eager=True) as vllm_model:
quant_outputs_per_image = [
vllm_model.generate_greedy_logprobs(prompts,
max_tokens,
num_logprobs=num_logprobs,
images=images)
for prompts, images in inputs_per_image
]
for source_outputs, quant_outputs in zip(source_outputs_per_image,
quant_outputs_per_image):
# TODO: Check whether using original CLIPVisionModel can improve
# consistency against HF
check_logprobs_close(
outputs_0_lst=source_outputs,
outputs_1_lst=quant_outputs,
name_0="source",
name_1="awq",
)
target_dtype = "half"
if is_cpu():
target_dtype = "bfloat16"
@pytest.mark.parametrize("model", models)
@pytest.mark.parametrize(
"size_factors",
[
# No image
[],
# Single-scale
[1.0],
# Single-scale, batched
[1.0, 1.0, 1.0],
# Multi-scale
[0.25, 0.5, 1.0],
],
)
@pytest.mark.parametrize("dtype", [target_dtype])
@pytest.mark.parametrize("max_tokens", [128])
@pytest.mark.parametrize("num_logprobs", [5])
@torch.inference_mode()
def test_models(hf_runner, vllm_runner, image_assets, model, size_factors,
dtype: str, max_tokens: int, num_logprobs: int) -> None:
images = [asset.pil_image for asset in image_assets]
inputs_per_image = [(
[prompt for _ in size_factors],
[rescale_image_size(image, factor) for factor in size_factors],
) for image, prompt in zip(images, HF_IMAGE_PROMPTS)]
run_test(
hf_runner,
vllm_runner,
inputs_per_image,
model,
dtype=dtype,
max_tokens=max_tokens,
num_logprobs=num_logprobs,
mm_limit=1,
tensor_parallel_size=1,
)
@pytest.mark.parametrize("model", models)
@pytest.mark.parametrize(
"size_factors",
[
# No image
[],
# Single-scale
[1.0],
# Single-scale, batched
[1.0, 1.0, 1.0],
# Multi-scale
[0.5, 0.75, 1.0],
],
)
@pytest.mark.parametrize("dtype", [target_dtype])
@pytest.mark.parametrize("max_tokens", [128])
@pytest.mark.parametrize("num_logprobs", [5])
@torch.inference_mode()
def test_multi_images_models(hf_runner, vllm_runner, image_assets, model,
size_factors, dtype: str, max_tokens: int,
num_logprobs: int) -> None:
images = [asset.pil_image for asset in image_assets]
inputs_per_case = [
([HF_MULTIIMAGE_IMAGE_PROMPT for _ in size_factors],
[[rescale_image_size(image, factor) for image in images]
for factor in size_factors])
]
run_test(
hf_runner,
vllm_runner,
inputs_per_case,
model,
dtype=dtype,
max_tokens=max_tokens,
num_logprobs=num_logprobs,
mm_limit=2,
tensor_parallel_size=1,
)
@pytest.mark.parametrize("model", ["OpenGVLab/InternVL2-2B"])
@pytest.mark.parametrize("size_factors", [[0.5, 1.0]])
@pytest.mark.parametrize("dtype", [target_dtype])
@pytest.mark.parametrize("max_tokens", [128])
@pytest.mark.parametrize("num_logprobs", [5])
@torch.inference_mode()
def test_different_num_patches(hf_runner, vllm_runner, image_assets, model,
size_factors, dtype: str, max_tokens: int,
num_logprobs: int) -> None:
images = [asset.pil_image.resize((896, 896)) for asset in image_assets]
inputs_batching = [(
[prompt for _ in size_factors],
[rescale_image_size(image, factor) for factor in size_factors],
) for image, prompt in zip(images, HF_IMAGE_PROMPTS)]
inputs_multi_images = [
([HF_MULTIIMAGE_IMAGE_PROMPT for _ in size_factors],
[[rescale_image_size(image, factor) for image in images]
for factor in size_factors])
]
for inputs in [inputs_batching, inputs_multi_images]:
run_test(
hf_runner,
vllm_runner,
inputs,
model,
dtype=dtype,
max_tokens=max_tokens,
num_logprobs=num_logprobs,
mm_limit=2,
tensor_parallel_size=1,
)
@pytest.mark.parametrize(
"models", [("OpenGVLab/InternVL2-2B", "OpenGVLab/InternVL2-2B-AWQ")])
@pytest.mark.parametrize(
"size_factors",
[
# No image
[],
# Single-scale
[1.0],
# Single-scale, batched
[1.0, 1.0, 1.0],
# Multi-scale
[0.25, 0.5, 1.0],
],
)
@pytest.mark.parametrize("dtype", ["half"])
@pytest.mark.parametrize("max_tokens", [128])
@pytest.mark.parametrize("num_logprobs", [5])
@torch.inference_mode()
def test_awq_models(vllm_runner, image_assets, models, size_factors,
dtype: str, max_tokens: int, num_logprobs: int) -> None:
run_awq_test(
vllm_runner,
image_assets,
models,
size_factors=size_factors,
dtype=dtype,
max_tokens=max_tokens,
num_logprobs=num_logprobs,
tensor_parallel_size=1,
)
@@ -0,0 +1,313 @@
from typing import List, Optional, Tuple, Type, overload
import pytest
from transformers import (AutoConfig, AutoModelForVision2Seq, AutoTokenizer,
BatchEncoding)
from vllm.multimodal.utils import rescale_image_size
from vllm.sequence import SampleLogprobs
from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE
from ....conftest import (IMAGE_ASSETS, HfRunner, PromptImageInput, VllmRunner,
_ImageAssets)
from ...utils import check_logprobs_close
_LIMIT_IMAGE_PER_PROMPT = 4
HF_IMAGE_PROMPTS = IMAGE_ASSETS.prompts({
"stop_sign":
"USER: <image>\nWhat's the content of the image?\nASSISTANT:",
"cherry_blossom":
"USER: <image>\nWhat is the season?\nASSISTANT:",
})
models = [
"llava-hf/llava-1.5-7b-hf",
# TODO: Get this model to produce meaningful output in vLLM
# "TIGER-Lab/Mantis-8B-siglip-llama3",
]
def vllm_to_hf_output(vllm_output: Tuple[List[int], str,
Optional[SampleLogprobs]],
model: str):
"""Sanitize vllm output to be comparable with hf output."""
output_ids, output_str, out_logprobs = vllm_output
config = AutoConfig.from_pretrained(model)
image_token_id = config.image_token_index
tokenizer = AutoTokenizer.from_pretrained(model)
eos_token_id = tokenizer.eos_token_id
hf_output_ids = [
token_id for idx, token_id in enumerate(output_ids)
if token_id != image_token_id or output_ids[idx - 1] != image_token_id
]
assert output_str[0] == " "
hf_output_str = output_str[1:]
if hf_output_ids[-1] == eos_token_id:
hf_output_str = hf_output_str + tokenizer.decode(eos_token_id)
return hf_output_ids, hf_output_str, out_logprobs
@overload
def run_test(
hf_runner: Type[HfRunner],
vllm_runner: Type[VllmRunner],
image_assets: _ImageAssets,
model: str,
*,
size_factors: List[float],
dtype: str,
max_tokens: int,
num_logprobs: int,
tensor_parallel_size: int,
distributed_executor_backend: Optional[str] = None,
):
...
@overload
def run_test(
hf_runner: Type[HfRunner],
vllm_runner: Type[VllmRunner],
image_assets: _ImageAssets,
model: str,
*,
sizes: List[Tuple[int, int]],
dtype: str,
max_tokens: int,
num_logprobs: int,
tensor_parallel_size: int,
distributed_executor_backend: Optional[str] = None,
):
...
def run_test(
hf_runner: Type[HfRunner],
vllm_runner: Type[VllmRunner],
image_assets: _ImageAssets,
model: str,
*,
size_factors: Optional[List[float]] = None,
sizes: Optional[List[Tuple[int, int]]] = None,
dtype: str,
max_tokens: int,
num_logprobs: int,
tensor_parallel_size: int,
distributed_executor_backend: Optional[str] = None,
):
images = [asset.pil_image for asset in image_assets]
if size_factors is not None:
inputs_per_image = [(
[prompt for _ in size_factors],
[rescale_image_size(image, factor) for factor in size_factors],
) for image, prompt in zip(images, HF_IMAGE_PROMPTS)]
elif sizes is not None:
inputs_per_image = [(
[prompt for _ in sizes],
[image.resize(size) for size in sizes],
) for image, prompt in zip(images, HF_IMAGE_PROMPTS)]
else:
raise ValueError("You must provide either `size_factors` or `sizes`")
_run_test(hf_runner,
vllm_runner,
inputs_per_image,
model,
dtype=dtype,
max_tokens=max_tokens,
num_logprobs=num_logprobs,
tensor_parallel_size=tensor_parallel_size,
distributed_executor_backend=distributed_executor_backend)
def _run_test(
hf_runner: Type[HfRunner],
vllm_runner: Type[VllmRunner],
inputs: List[Tuple[List[str], PromptImageInput]],
model: str,
*,
dtype: str,
max_tokens: int,
num_logprobs: int,
tensor_parallel_size: int,
distributed_executor_backend: Optional[str] = None,
):
"""Inference result should be the same between hf and vllm.
All the image fixtures for the test are from IMAGE_ASSETS.
For huggingface runner, we provide the PIL images as input.
For vllm runner, we provide MultiModalDataDict objects
and corresponding MultiModalConfig as input.
Note, the text input is also adjusted to abide by vllm contract.
The text output is sanitized to be able to compare with hf.
"""
# NOTE: For local use; this isn't tested in CI yet (see TODO above)
if model.startswith("TIGER-Lab/Mantis"):
from mantis.models.mllava import MLlavaProcessor
torch_dtype = STR_DTYPE_TO_TORCH_DTYPE[dtype]
mantis_processor = MLlavaProcessor.from_pretrained(
model, torch_dtype=torch_dtype)
assert isinstance(mantis_processor, MLlavaProcessor)
else:
mantis_processor = None
# NOTE: take care of the order. run vLLM first, and then run HF.
# vLLM needs a fresh new process without cuda initialization.
# if we run HF first, the cuda initialization will be done and it
# will hurt multiprocessing backend with fork method (the default method).
# max_model_len should be greater than image_feature_size
with vllm_runner(model,
dtype=dtype,
max_model_len=4096,
tensor_parallel_size=tensor_parallel_size,
distributed_executor_backend=distributed_executor_backend,
enforce_eager=True,
limit_mm_per_prompt={"image": _LIMIT_IMAGE_PER_PROMPT
}) as vllm_model:
vllm_outputs_per_image = [
vllm_model.generate_greedy_logprobs(prompts,
max_tokens,
num_logprobs=num_logprobs,
images=images)
for prompts, images in inputs
]
if mantis_processor is not None:
def process(hf_inputs: BatchEncoding):
hf_inputs["pixel_values"] = hf_inputs["pixel_values"] \
.to(torch_dtype) # type: ignore
return hf_inputs
else:
def process(hf_inputs: BatchEncoding):
return hf_inputs
with hf_runner(model,
dtype=dtype,
postprocess_inputs=process,
auto_cls=AutoModelForVision2Seq) as hf_model:
hf_outputs_per_image = [
hf_model.generate_greedy_logprobs_limit(prompts,
max_tokens,
num_logprobs=num_logprobs,
images=images)
for prompts, images in inputs
]
for hf_outputs, vllm_outputs in zip(hf_outputs_per_image,
vllm_outputs_per_image):
# TODO: Check whether using original CLIPVisionModel can improve
# consistency against HF
check_logprobs_close(
outputs_0_lst=hf_outputs,
outputs_1_lst=[
vllm_to_hf_output(vllm_output, model)
for vllm_output in vllm_outputs
],
name_0="hf",
name_1="vllm",
)
@pytest.mark.parametrize("model", models)
@pytest.mark.parametrize(
"size_factors",
[
# No image
[],
# Single-scale
[1.0],
# Single-scale, batched
[1.0, 1.0, 1.0],
# Multi-scale
[0.25, 0.5, 1.0],
],
)
@pytest.mark.parametrize("dtype", ["half"])
@pytest.mark.parametrize("max_tokens", [128])
@pytest.mark.parametrize("num_logprobs", [5])
def test_models(hf_runner, vllm_runner, image_assets, model, size_factors,
dtype, max_tokens, num_logprobs) -> None:
run_test(
hf_runner,
vllm_runner,
image_assets,
model,
size_factors=size_factors,
dtype=dtype,
max_tokens=max_tokens,
num_logprobs=num_logprobs,
tensor_parallel_size=1,
)
@pytest.mark.parametrize("model", models)
@pytest.mark.parametrize("dtype", ["half"])
@pytest.mark.parametrize("max_tokens", [128])
@pytest.mark.parametrize("num_logprobs", [5])
def test_models_multiple_image_inputs(hf_runner, vllm_runner, image_assets,
model, dtype, max_tokens,
num_logprobs) -> None:
stop_sign = image_assets[0].pil_image
cherry_blossom = image_assets[1].pil_image
inputs = [(
[
"USER: <image><image>\nDescribe 2 images.\nASSISTANT:",
"USER: <image><image>\nDescribe 2 images.\nASSISTANT:",
"USER: <image><image><image><image>\nDescribe 4 images.\nASSISTANT:", # noqa: E501
"USER: <image>\nWhat is the season?\nASSISTANT:",
],
[
[stop_sign, cherry_blossom],
# Images with different sizes and aspect-ratios
[
rescale_image_size(stop_sign, 0.1),
stop_sign,
],
[
stop_sign,
rescale_image_size(stop_sign, 0.25),
cherry_blossom.resize((183, 488)),
cherry_blossom.resize((488, 183))
],
cherry_blossom,
])]
_run_test(
hf_runner,
vllm_runner,
inputs,
model,
dtype=dtype,
max_tokens=max_tokens,
num_logprobs=num_logprobs,
tensor_parallel_size=1,
)
@pytest.mark.parametrize("model", models)
def test_context_length_too_short(vllm_runner, image_assets, model):
images = [asset.pil_image for asset in image_assets]
with pytest.raises(ValueError, match="too long to fit into the model"):
vllm_model = vllm_runner(
model,
max_model_len=128, # LLaVA has a feature size of 576
enforce_eager=True,
)
with vllm_model:
vllm_model.generate_greedy([HF_IMAGE_PROMPTS[0]],
max_tokens=1,
images=[images[0]])
@@ -0,0 +1,158 @@
from typing import List, Optional, Tuple, Type
import pytest
from transformers import AutoConfig, AutoModelForVision2Seq, AutoTokenizer
from vllm.sequence import SampleLogprobs
from ....conftest import IMAGE_ASSETS, HfRunner, VllmRunner, _ImageAssets
from ...utils import check_logprobs_close
HF_IMAGE_PROMPTS = IMAGE_ASSETS.prompts({
"stop_sign":
"USER: <image>\nWhat's the content of the image?\nASSISTANT:",
"cherry_blossom":
"USER: <image>\nWhat is the season?\nASSISTANT:",
})
models = [
"llava-hf/llava-1.5-7b-hf",
]
def vllm_to_hf_output(vllm_output: Tuple[List[int], str,
Optional[SampleLogprobs]],
model: str):
"""Sanitize vllm output to be comparable with hf output."""
output_ids, output_str, out_logprobs = vllm_output
config = AutoConfig.from_pretrained(model)
image_token_id = config.image_token_index
tokenizer = AutoTokenizer.from_pretrained(model)
eos_token_id = tokenizer.eos_token_id
hf_output_ids = [
token_id for idx, token_id in enumerate(output_ids)
if token_id != image_token_id or output_ids[idx - 1] != image_token_id
]
assert output_str[0] == " "
hf_output_str = output_str[1:]
if hf_output_ids[-1] == eos_token_id:
hf_output_str = hf_output_str + tokenizer.decode(eos_token_id)
return hf_output_ids, hf_output_str, out_logprobs
def run_test(
hf_runner: Type[HfRunner],
vllm_runner: Type[VllmRunner],
image_assets: _ImageAssets,
model: str,
*,
size_factors: List[float],
dtype: str,
max_tokens: int,
num_logprobs: int,
tensor_parallel_size: int,
distributed_executor_backend: Optional[str] = None,
):
"""Inference result should be the same between hf and vllm.
All the image fixtures for the test are from IMAGE_ASSETS.
For huggingface runner, we provide the PIL images as input.
For vllm runner, we provide MultiModalDataDict objects
and corresponding vision language config as input.
Note, the text input is also adjusted to abide by vllm contract.
The text output is sanitized to be able to compare with hf.
"""
# vLLM to load from image embeddings
vllm_images = [asset.image_embeds for asset in image_assets]
# transformers to load from PIL images
hf_images = [asset.pil_image for asset in image_assets]
vllm_inputs_per_image = [(
[prompt for _ in size_factors],
[image for _ in size_factors],
) for image, prompt in zip(vllm_images, HF_IMAGE_PROMPTS)]
hf_inputs_per_image = [(
[prompt for _ in size_factors],
[image for _ in size_factors],
) for image, prompt in zip(hf_images, HF_IMAGE_PROMPTS)]
# NOTE: take care of the order. run vLLM first, and then run HF.
# vLLM needs a fresh new process without cuda initialization.
# if we run HF first, the cuda initialization will be done and it
# will hurt multiprocessing backend with fork method (the default method).
# max_model_len should be greater than image_feature_size
with vllm_runner(model,
dtype=dtype,
tensor_parallel_size=tensor_parallel_size,
distributed_executor_backend=distributed_executor_backend,
enforce_eager=True) as vllm_model:
vllm_outputs_per_image = [
vllm_model.generate_greedy_logprobs(prompts,
max_tokens,
num_logprobs=num_logprobs,
images=images)
for prompts, images in vllm_inputs_per_image
]
with hf_runner(model, dtype=dtype,
auto_cls=AutoModelForVision2Seq) as hf_model:
hf_outputs_per_image = [
hf_model.generate_greedy_logprobs_limit(prompts,
max_tokens,
num_logprobs=num_logprobs,
images=images)
for prompts, images in hf_inputs_per_image
]
for hf_outputs, vllm_outputs in zip(hf_outputs_per_image,
vllm_outputs_per_image):
# TODO: Check whether using original CLIPVisionModel can improve
# consistency against HF
check_logprobs_close(
outputs_0_lst=hf_outputs,
outputs_1_lst=[
vllm_to_hf_output(vllm_output, model)
for vllm_output in vllm_outputs
],
name_0="hf",
name_1="vllm",
)
@pytest.mark.parametrize("model", models)
@pytest.mark.parametrize(
"size_factors",
[
# No image
[],
# Single-scale
[1.0],
# Single-scale, batched
[1.0, 1.0, 1.0],
],
)
@pytest.mark.parametrize("dtype", ["half"])
@pytest.mark.parametrize("max_tokens", [128])
@pytest.mark.parametrize("num_logprobs", [5])
def test_models(hf_runner, vllm_runner, image_assets, model, size_factors,
dtype: str, max_tokens: int, num_logprobs: int) -> None:
run_test(
hf_runner,
vllm_runner,
image_assets,
model,
size_factors=size_factors,
dtype=dtype,
max_tokens=max_tokens,
num_logprobs=num_logprobs,
tensor_parallel_size=1,
)
@@ -0,0 +1,283 @@
from typing import List, Optional, Tuple, Type, overload
import pytest
from transformers import AutoConfig, AutoModelForVision2Seq, AutoTokenizer
from vllm.multimodal.utils import rescale_image_size
from vllm.sequence import SampleLogprobs
from ....conftest import (IMAGE_ASSETS, HfRunner, PromptImageInput, VllmRunner,
_ImageAssets)
from ...utils import check_logprobs_close
_LIMIT_IMAGE_PER_PROMPT = 4
HF_IMAGE_PROMPTS = IMAGE_ASSETS.prompts({
"stop_sign":
"[INST] <image>\nWhat's the content of the image? [/INST]",
"cherry_blossom":
"[INST] <image>\nWhat is the season? [/INST]",
})
models = ["llava-hf/llava-v1.6-mistral-7b-hf"]
def vllm_to_hf_output(vllm_output: Tuple[List[int], str,
Optional[SampleLogprobs]],
model: str):
"""Sanitize vllm output to be comparable with hf output."""
output_ids, output_str, out_logprobs = vllm_output
config = AutoConfig.from_pretrained(model)
image_token_id = config.image_token_index
tokenizer = AutoTokenizer.from_pretrained(model)
eos_token_id = tokenizer.eos_token_id
hf_output_ids = [
token_id for idx, token_id in enumerate(output_ids)
if token_id != image_token_id or output_ids[idx - 1] != image_token_id
]
assert output_str[0] == " "
hf_output_str = output_str[1:]
if hf_output_ids[-1] == eos_token_id:
hf_output_str = hf_output_str + tokenizer.decode(eos_token_id)
return hf_output_ids, hf_output_str, out_logprobs
@overload
def run_test(
hf_runner: Type[HfRunner],
vllm_runner: Type[VllmRunner],
image_assets: _ImageAssets,
model: str,
*,
size_factors: List[float],
dtype: str,
max_tokens: int,
num_logprobs: int,
tensor_parallel_size: int,
distributed_executor_backend: Optional[str] = None,
):
...
@overload
def run_test(
hf_runner: Type[HfRunner],
vllm_runner: Type[VllmRunner],
image_assets: _ImageAssets,
model: str,
*,
sizes: List[Tuple[int, int]],
dtype: str,
max_tokens: int,
num_logprobs: int,
tensor_parallel_size: int,
distributed_executor_backend: Optional[str] = None,
):
...
def run_test(
hf_runner: Type[HfRunner],
vllm_runner: Type[VllmRunner],
image_assets: _ImageAssets,
model: str,
*,
size_factors: Optional[List[float]] = None,
sizes: Optional[List[Tuple[int, int]]] = None,
dtype: str,
max_tokens: int,
num_logprobs: int,
tensor_parallel_size: int,
distributed_executor_backend: Optional[str] = None,
):
images = [asset.pil_image for asset in image_assets]
if size_factors is not None:
inputs_per_image = [(
[prompt for _ in size_factors],
[rescale_image_size(image, factor) for factor in size_factors],
) for image, prompt in zip(images, HF_IMAGE_PROMPTS)]
elif sizes is not None:
inputs_per_image = [(
[prompt for _ in sizes],
[image.resize(size) for size in sizes],
) for image, prompt in zip(images, HF_IMAGE_PROMPTS)]
else:
raise ValueError("You must provide either `size_factors` or `sizes`")
_run_test(hf_runner,
vllm_runner,
inputs_per_image,
model,
dtype=dtype,
max_tokens=max_tokens,
num_logprobs=num_logprobs,
tensor_parallel_size=tensor_parallel_size,
distributed_executor_backend=distributed_executor_backend)
def _run_test(
hf_runner: Type[HfRunner],
vllm_runner: Type[VllmRunner],
inputs: List[Tuple[List[str], PromptImageInput]],
model: str,
dtype: str,
max_tokens: int,
num_logprobs: int,
tensor_parallel_size: int,
distributed_executor_backend: Optional[str] = None,
):
# max_model_len should be greater than image_feature_size
with vllm_runner(model,
dtype=dtype,
max_model_len=10240,
tensor_parallel_size=tensor_parallel_size,
distributed_executor_backend=distributed_executor_backend,
enforce_eager=True,
limit_mm_per_prompt={"image": _LIMIT_IMAGE_PER_PROMPT
}) as vllm_model:
vllm_outputs_per_image = [
vllm_model.generate_greedy_logprobs(prompts,
max_tokens,
num_logprobs=num_logprobs,
images=images)
for prompts, images in inputs
]
with hf_runner(model, dtype=dtype,
auto_cls=AutoModelForVision2Seq) as hf_model:
hf_outputs_per_image = [
hf_model.generate_greedy_logprobs_limit(prompts,
max_tokens,
num_logprobs=num_logprobs,
images=images)
for prompts, images in inputs
]
for hf_outputs, vllm_outputs in zip(hf_outputs_per_image,
vllm_outputs_per_image):
# TODO: Check whether using original CLIPVisionModel can improve
# consistency against HF
check_logprobs_close(
outputs_0_lst=hf_outputs,
outputs_1_lst=[
vllm_to_hf_output(vllm_output, model)
for vllm_output in vllm_outputs
],
name_0="hf",
name_1="vllm",
)
@pytest.mark.parametrize("model", models)
@pytest.mark.parametrize(
"size_factors",
[
# No image
[],
# Single-scale
[1.0],
# Single-scale, batched
[1.0, 1.0, 1.0],
# Multi-scale
[0.25, 0.5, 1.0],
],
)
@pytest.mark.parametrize("dtype", ["half"])
@pytest.mark.parametrize("max_tokens", [128])
@pytest.mark.parametrize("num_logprobs", [5])
def test_models(hf_runner, vllm_runner, image_assets, model, size_factors,
dtype, max_tokens, num_logprobs) -> None:
"""Inference result should be the same between hf and vllm.
All the image fixtures for the test are from IMAGE_ASSETS.
For huggingface runner, we provide the PIL images as input.
For vllm runner, we provide MultiModalDataDict objects
and corresponding MultiModalConfig as input.
Note, the text input is also adjusted to abide by vllm contract.
The text output is sanitized to be able to compare with hf.
"""
run_test(
hf_runner,
vllm_runner,
image_assets,
model,
size_factors=size_factors,
dtype=dtype,
max_tokens=max_tokens,
num_logprobs=num_logprobs,
tensor_parallel_size=1,
)
@pytest.mark.parametrize("model", models)
@pytest.mark.parametrize(
"sizes",
[[(1669, 2560), (2560, 1669), (183, 488), (488, 183)]],
)
@pytest.mark.parametrize("dtype", ["half"])
@pytest.mark.parametrize("max_tokens", [128])
@pytest.mark.parametrize("num_logprobs", [5])
def test_models_fixed_sizes(hf_runner, vllm_runner, image_assets, model, sizes,
dtype, max_tokens, num_logprobs) -> None:
run_test(
hf_runner,
vllm_runner,
image_assets,
model,
sizes=sizes,
dtype=dtype,
max_tokens=max_tokens,
num_logprobs=num_logprobs,
tensor_parallel_size=1,
)
@pytest.mark.parametrize("model", models)
@pytest.mark.parametrize("dtype", ["half"])
@pytest.mark.parametrize("max_tokens", [128])
@pytest.mark.parametrize("num_logprobs", [5])
def test_models_multiple_image_inputs(hf_runner, vllm_runner, image_assets,
model, dtype, max_tokens,
num_logprobs) -> None:
stop_sign = image_assets[0].pil_image
cherry_blossom = image_assets[1].pil_image
inputs = [(
[
"[INST] <image><image>\nDescribe 2 images. [/INST]",
"[INST] <image><image>\nDescribe 2 images. [/INST]",
"[INST] <image><image><image><image>\nDescribe 4 images. [/INST]",
"[INST] <image>\nWhat is the season? [/INST]"
],
[
[stop_sign, cherry_blossom],
# Images with different sizes and aspect-ratios
[
rescale_image_size(stop_sign, 0.1),
stop_sign,
],
[
stop_sign,
rescale_image_size(stop_sign, 0.25),
cherry_blossom.resize((183, 488)),
cherry_blossom.resize((488, 183))
],
cherry_blossom,
])]
_run_test(
hf_runner,
vllm_runner,
inputs,
model,
dtype=dtype,
max_tokens=max_tokens,
num_logprobs=num_logprobs,
tensor_parallel_size=1,
)
@@ -0,0 +1,234 @@
from typing import List, Optional, Tuple, Type, overload
import pytest
import transformers
from transformers import AutoConfig, AutoModelForVision2Seq, AutoTokenizer
from vllm.multimodal.utils import (rescale_video_size, resize_video,
sample_frames_from_video)
from vllm.sequence import SampleLogprobs
from ....conftest import VIDEO_ASSETS, HfRunner, VllmRunner, _VideoAssets
from ...utils import check_logprobs_close
_PREFACE = (
"A chat between a curious human and an artificial intelligence assistant. "
"The assistant gives helpful, detailed, and polite answers to the human's "
"questions.")
HF_VIDEO_PROMPTS = VIDEO_ASSETS.prompts({
"sample_demo_1":
f"{_PREFACE}USER: <video>\nWhy is this video funny? ASSISTANT:"
})
models = ["llava-hf/LLaVA-NeXT-Video-7B-hf"]
def vllm_to_hf_output(vllm_output: Tuple[List[int], str,
Optional[SampleLogprobs]],
model: str):
"""Sanitize vllm output to be comparable with hf output."""
output_ids, output_str, out_logprobs = vllm_output
config = AutoConfig.from_pretrained(model)
video_token_id = config.video_token_index
tokenizer = AutoTokenizer.from_pretrained(model)
eos_token_id = tokenizer.eos_token_id
hf_output_ids = [
token_id for idx, token_id in enumerate(output_ids)
if token_id != video_token_id or output_ids[idx - 1] != video_token_id
]
assert output_str[0] == " "
hf_output_str = output_str[1:]
if hf_output_ids[-1] == eos_token_id:
hf_output_str = hf_output_str + tokenizer.decode(eos_token_id)
return hf_output_ids, hf_output_str, out_logprobs
@overload
def run_test(
hf_runner: Type[HfRunner],
vllm_runner: Type[VllmRunner],
video_assets: _VideoAssets,
model: str,
*,
size_factors: List[float],
dtype: str,
max_tokens: int,
num_logprobs: int,
num_frames: int,
tensor_parallel_size: int,
distributed_executor_backend: Optional[str] = None,
):
...
@overload
def run_test(
hf_runner: Type[HfRunner],
vllm_runner: Type[VllmRunner],
video_assets: _VideoAssets,
model: str,
*,
sizes: List[Tuple[int, int]],
dtype: str,
max_tokens: int,
num_logprobs: int,
num_frames: int,
tensor_parallel_size: int,
distributed_executor_backend: Optional[str] = None,
):
...
def run_test(
hf_runner: Type[HfRunner],
vllm_runner: Type[VllmRunner],
video_assets: _VideoAssets,
model: str,
*,
size_factors: Optional[List[float]] = None,
sizes: Optional[List[Tuple[int, int]]] = None,
dtype: str,
max_tokens: int,
num_logprobs: int,
num_frames: int,
tensor_parallel_size: int,
distributed_executor_backend: Optional[str] = None,
):
videos = [
sample_frames_from_video(asset.np_ndarrays, num_frames)
for asset in video_assets
]
for video in videos:
print(video.shape)
if size_factors is not None:
inputs_per_video = [(
[prompt for _ in size_factors],
[rescale_video_size(video, factor) for factor in size_factors],
) for video, prompt in zip(videos, HF_VIDEO_PROMPTS)]
elif sizes is not None:
inputs_per_video = [(
[prompt for _ in sizes],
[resize_video(video, size) for size in sizes],
) for video, prompt in zip(videos, HF_VIDEO_PROMPTS)]
else:
raise ValueError("You must provide either `size_factors` or `sizes`")
# max_model_len should be greater than image_feature_size
with vllm_runner(model,
dtype=dtype,
max_model_len=4096,
tensor_parallel_size=tensor_parallel_size,
distributed_executor_backend=distributed_executor_backend,
enforce_eager=True) as vllm_model:
vllm_outputs_per_video = [
vllm_model.generate_greedy_logprobs(prompts,
max_tokens,
num_logprobs=num_logprobs,
videos=videos)
for prompts, videos in inputs_per_video
]
with hf_runner(model, dtype=dtype,
auto_cls=AutoModelForVision2Seq) as hf_model:
hf_outputs_per_video = [
hf_model.generate_greedy_logprobs_limit(prompts,
max_tokens,
num_logprobs=num_logprobs,
videos=videos)
for prompts, videos in inputs_per_video
]
for hf_outputs, vllm_outputs in zip(hf_outputs_per_video,
vllm_outputs_per_video):
# TODO: Check whether using original CLIPVisionModel can improve
# consistency against HF
check_logprobs_close(
outputs_0_lst=hf_outputs,
outputs_1_lst=[
vllm_to_hf_output(vllm_output, model)
for vllm_output in vllm_outputs
],
name_0="hf",
name_1="vllm",
)
@pytest.mark.skipif(transformers.__version__ < "4.45",
reason="Waiting for next transformers release")
@pytest.mark.parametrize("model", models)
@pytest.mark.parametrize(
"size_factors",
[
# No video
[],
# Single-scale
[1.0],
# Single-scale, batched
[1.0, 1.0, 1.0],
# Multi-scale
[0.25, 0.5, 1.0],
],
)
@pytest.mark.parametrize("dtype", ["half"])
@pytest.mark.parametrize("max_tokens", [128])
@pytest.mark.parametrize("num_logprobs", [5])
@pytest.mark.parametrize("num_frames", [16])
def test_models(hf_runner, vllm_runner, video_assets, model, size_factors,
dtype, max_tokens, num_logprobs, num_frames) -> None:
"""Inference result should be the same between hf and vllm.
All the image fixtures for the test is under tests/videos.
For huggingface runner, we provide the np.ndarray as input.
For vllm runner, we provide MultiModalDataDict objects
and corresponding MultiModalConfig as input.
Note, the text input is also adjusted to abide by vllm contract.
The text output is sanitized to be able to compare with hf.
"""
run_test(
hf_runner,
vllm_runner,
video_assets,
model,
size_factors=size_factors,
dtype=dtype,
max_tokens=max_tokens,
num_logprobs=num_logprobs,
num_frames=num_frames,
tensor_parallel_size=1,
)
@pytest.mark.skipif(transformers.__version__ < "4.45",
reason="Waiting for next transformers release")
@pytest.mark.parametrize("model", models)
@pytest.mark.parametrize(
"sizes",
[[(1669, 2560), (2560, 1669), (183, 488), (488, 183)]],
)
@pytest.mark.parametrize("dtype", ["half"])
@pytest.mark.parametrize("max_tokens", [128])
@pytest.mark.parametrize("num_logprobs", [5])
@pytest.mark.parametrize("num_frames", [16])
def test_models_fixed_sizes(hf_runner, vllm_runner, video_assets, model, sizes,
dtype, max_tokens, num_logprobs,
num_frames) -> None:
run_test(
hf_runner,
vllm_runner,
video_assets,
model,
sizes=sizes,
dtype=dtype,
max_tokens=max_tokens,
num_logprobs=num_logprobs,
num_frames=num_frames,
tensor_parallel_size=1,
)
@@ -0,0 +1,199 @@
from typing import List, Optional, Tuple, Type, Union
import pytest
import torch
import torch.types
from PIL import Image
from transformers import BatchEncoding
from vllm.multimodal.utils import rescale_image_size
from vllm.sequence import SampleLogprobs
from ....conftest import IMAGE_ASSETS, HfRunner, VllmRunner
from ...utils import check_logprobs_close
# The image token is placed before "user" on purpose so that the test can pass
HF_IMAGE_PROMPTS = IMAGE_ASSETS.prompts({
"stop_sign":
"<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\n" \
"(<image>./</image>)\nWhat's the content of the image?<|eot_id|>" \
"<|start_header_id|>assistant<|end_header_id|>\n\n", # noqa: E501
"cherry_blossom":
"<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\n" \
"(<image>./</image>)\nWhat is the season?<|eot_id|>" \
"<|start_header_id|>assistant<|end_header_id|>\n\n",
})
HF_MULTIIMAGE_IMAGE_PROMPT = \
"<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\n" \
"(<image>./</image>)\n(<image>./</image>)\n" \
"Describe these images.<|eot_id|>" \
"<|start_header_id|>assistant<|end_header_id|>\n\n"
models = ["openbmb/MiniCPM-Llama3-V-2_5"]
def _wrap_inputs(hf_inputs: BatchEncoding) -> BatchEncoding:
return BatchEncoding({"model_inputs": hf_inputs})
def trunc_hf_output(hf_output: Tuple[List[int], str,
Optional[SampleLogprobs]]):
output_ids, output_str, out_logprobs = hf_output
if output_str.endswith("<|eot_id|>"):
output_str = output_str.split("<|eot_id|>")[0]
return output_ids, output_str, out_logprobs
target_dtype = "half"
def run_test(
hf_runner: Type[HfRunner],
vllm_runner: Type[VllmRunner],
inputs: List[Tuple[List[str], Union[List[Image.Image],
List[List[Image.Image]]]]],
model: str,
*,
dtype: str,
max_tokens: int,
num_logprobs: int,
mm_limit: int,
tensor_parallel_size: int,
distributed_executor_backend: Optional[str] = None,
):
"""Inference result should be the same between hf and vllm.
All the image fixtures for the test are from IMAGE_ASSETS.
For huggingface runner, we provide the PIL images as input.
For vllm runner, we provide MultiModalDataDict objects
and corresponding MultiModalConfig as input.
Note, the text input is also adjusted to abide by vllm contract.
The text output is sanitized to be able to compare with hf.
"""
# NOTE: take care of the order. run vLLM first, and then run HF.
# vLLM needs a fresh new process without cuda initialization.
# if we run HF first, the cuda initialization will be done and it
# will hurt multiprocessing backend with fork method (the default method).
# max_model_len should be greater than image_feature_size
with vllm_runner(model,
max_model_len=4096,
max_num_seqs=1,
dtype=dtype,
limit_mm_per_prompt={"image": mm_limit},
tensor_parallel_size=tensor_parallel_size,
distributed_executor_backend=distributed_executor_backend,
enforce_eager=True) as vllm_model:
tokenizer = vllm_model.model.get_tokenizer()
stop_token_ids = [tokenizer.eos_id, tokenizer.eot_id]
vllm_outputs_per_image = [
vllm_model.generate_greedy_logprobs(prompts,
max_tokens,
num_logprobs=num_logprobs,
images=images,
stop_token_ids=stop_token_ids)
for prompts, images in inputs
]
hf_model = hf_runner(model, dtype=dtype, postprocess_inputs=_wrap_inputs)
with hf_model, torch.no_grad():
hf_outputs_per_image = [
hf_model.generate_greedy_logprobs_limit(prompts,
max_tokens,
num_logprobs=num_logprobs,
images=images,
tokenizer=tokenizer)
for prompts, images in inputs
]
for hf_outputs, vllm_outputs in zip(hf_outputs_per_image,
vllm_outputs_per_image):
check_logprobs_close(
outputs_0_lst=[
trunc_hf_output(hf_output) for hf_output in hf_outputs
],
outputs_1_lst=vllm_outputs,
name_0="hf",
name_1="vllm",
)
@pytest.mark.parametrize("model", models)
@pytest.mark.parametrize(
"size_factors",
[
# No image
[],
# Single-scale
[1.0],
# Single-scale, batched
[1.0, 1.0, 1.0],
# Multi-scale
[0.25, 0.5, 1.0],
],
)
@pytest.mark.parametrize("dtype", [target_dtype])
@pytest.mark.parametrize("max_tokens", [128])
@pytest.mark.parametrize("num_logprobs", [5])
def test_models(hf_runner, vllm_runner, image_assets, model, size_factors,
dtype: str, max_tokens: int, num_logprobs: int) -> None:
images = [asset.pil_image for asset in image_assets]
inputs_per_image = [(
[prompt for _ in size_factors],
[rescale_image_size(image, factor) for factor in size_factors],
) for image, prompt in zip(images, HF_IMAGE_PROMPTS)]
run_test(
hf_runner,
vllm_runner,
inputs_per_image,
model,
dtype=dtype,
max_tokens=max_tokens,
num_logprobs=num_logprobs,
mm_limit=1,
tensor_parallel_size=1,
)
@pytest.mark.parametrize("model", models)
@pytest.mark.parametrize(
"size_factors",
[
# No image
[],
# Single-scale
[1.0],
# Single-scale, batched
[1.0, 1.0, 1.0],
# Multi-scale
[0.25, 0.5, 1.0],
],
)
@pytest.mark.parametrize("dtype", [target_dtype])
@pytest.mark.parametrize("max_tokens", [128])
@pytest.mark.parametrize("num_logprobs", [5])
def test_multi_images_models(hf_runner, vllm_runner, image_assets, model,
size_factors, dtype: str, max_tokens: int,
num_logprobs: int) -> None:
images = [asset.pil_image for asset in image_assets]
inputs_per_case = [
([HF_MULTIIMAGE_IMAGE_PROMPT for _ in size_factors],
[[rescale_image_size(image, factor) for image in images]
for factor in size_factors])
]
run_test(
hf_runner,
vllm_runner,
inputs_per_case,
model,
dtype=dtype,
max_tokens=max_tokens,
num_logprobs=num_logprobs,
mm_limit=2,
tensor_parallel_size=1,
)
@@ -0,0 +1,164 @@
import os
from typing import List, Optional, Tuple, Type
import pytest
from transformers import AutoConfig, AutoModelForVision2Seq, AutoTokenizer
from vllm.multimodal.utils import rescale_image_size
from vllm.sequence import SampleLogprobs
from vllm.utils import is_hip
from ....conftest import IMAGE_ASSETS, HfRunner, VllmRunner, _ImageAssets
from ...utils import check_logprobs_close
HF_IMAGE_PROMPTS = IMAGE_ASSETS.prompts({
"stop_sign":
"caption es",
"cherry_blossom":
"What is in the picture?",
})
models = ["google/paligemma-3b-mix-224"]
# ROCm Triton FA can run into compilation issues with these models due to,
# excessive use of shared memory. Use other backends in the meantime.
# FIXME (mattwong, gshtrasb, hongxiayan)
if is_hip():
os.environ["VLLM_USE_TRITON_FLASH_ATTN"] = "0"
def vllm_to_hf_output(vllm_output: Tuple[List[int], str,
Optional[SampleLogprobs]],
model: str):
"""Sanitize vllm output to be comparable with hf output."""
output_ids, output_str, out_logprobs = vllm_output
config = AutoConfig.from_pretrained(model)
image_token_id = config.image_token_index
tokenizer = AutoTokenizer.from_pretrained(model)
eos_token_id = tokenizer.eos_token_id
hf_output_ids = [
token_id for idx, token_id in enumerate(output_ids)
if token_id != image_token_id or output_ids[idx - 1] != image_token_id
]
hf_output_str = output_str
if hf_output_ids[-1] == eos_token_id:
hf_output_str = hf_output_str + tokenizer.decode(eos_token_id)
return hf_output_ids, hf_output_str, out_logprobs
def run_test(
hf_runner: Type[HfRunner],
vllm_runner: Type[VllmRunner],
image_assets: _ImageAssets,
model: str,
*,
size_factors: List[float],
dtype: str,
max_tokens: int,
num_logprobs: int,
tensor_parallel_size: int,
distributed_executor_backend: Optional[str] = None,
):
"""Inference result should be the same between hf and vllm.
All the image fixtures for the test are from IMAGE_ASSETS.
For huggingface runner, we provide the PIL images as input.
For vllm runner, we provide MultiModalDataDict objects
and corresponding MultiModalConfig as input.
Note, the text input is also adjusted to abide by vllm contract.
The text output is sanitized to be able to compare with hf.
"""
images = [asset.pil_image for asset in image_assets]
inputs_per_image = [(
[prompt for _ in size_factors],
[rescale_image_size(image, factor) for factor in size_factors],
) for image, prompt in zip(images, HF_IMAGE_PROMPTS)]
# NOTE: take care of the order. run vLLM first, and then run HF.
# vLLM needs a fresh new process without cuda initialization.
# if we run HF first, the cuda initialization will be done and it
# will hurt multiprocessing backend with fork method (the default method).
# max_model_len should be greater than image_feature_size
with vllm_runner(model,
dtype=dtype,
tensor_parallel_size=tensor_parallel_size,
distributed_executor_backend=distributed_executor_backend,
enforce_eager=True) as vllm_model:
vllm_outputs_per_image = [
vllm_model.generate_greedy_logprobs(prompts,
max_tokens,
num_logprobs=num_logprobs,
images=images)
for prompts, images in inputs_per_image
]
with hf_runner(model, dtype=dtype,
auto_cls=AutoModelForVision2Seq) as hf_model:
hf_outputs_per_image = [
hf_model.generate_greedy_logprobs_limit(prompts,
max_tokens,
num_logprobs=num_logprobs,
images=images)
for prompts, images in inputs_per_image
]
for hf_outputs, vllm_outputs in zip(hf_outputs_per_image,
vllm_outputs_per_image):
check_logprobs_close(
outputs_0_lst=hf_outputs,
outputs_1_lst=[
vllm_to_hf_output(vllm_output, model)
for vllm_output in vllm_outputs
],
name_0="hf",
name_1="vllm",
)
@pytest.mark.parametrize("model", models)
@pytest.mark.parametrize(
"size_factors",
[
# No image
[],
# Single-scale
[1.0],
# Single-scale, batched
[1.0, 1.0, 1.0],
# Multi-scale
[0.25, 0.5, 1.0],
],
)
@pytest.mark.parametrize("dtype", [
pytest.param(
"float",
marks=pytest.mark.skipif(
is_hip(),
reason=
"ROCm FA does not yet fully support 32-bit precision on PaliGemma")
), "half"
])
@pytest.mark.parametrize("max_tokens", [128])
@pytest.mark.parametrize("num_logprobs", [5])
def test_models(hf_runner, vllm_runner, image_assets, model, size_factors,
dtype: str, max_tokens: int, num_logprobs: int) -> None:
run_test(
hf_runner,
vllm_runner,
image_assets,
model,
size_factors=size_factors,
dtype=dtype,
max_tokens=max_tokens,
num_logprobs=num_logprobs,
tensor_parallel_size=1,
)
@@ -0,0 +1,232 @@
import os
import re
from typing import List, Optional, Tuple, Type
import pytest
from transformers import AutoTokenizer
from vllm.multimodal.utils import rescale_image_size
from vllm.sequence import SampleLogprobs
from vllm.utils import is_cpu, is_hip
from ....conftest import IMAGE_ASSETS, HfRunner, PromptImageInput, VllmRunner
from ...utils import check_logprobs_close
HF_IMAGE_PROMPTS = IMAGE_ASSETS.prompts({
"stop_sign":
"<|user|>\n<|image_1|>\nWhat's the content of the image?<|end|>\n<|assistant|>\n", # noqa: E501
"cherry_blossom":
"<|user|>\n<|image_1|>\nWhat is the season?<|end|>\n<|assistant|>\n",
})
HF_MULTIIMAGE_IMAGE_PROMPT = "<|user|>\n<|image_1|>\n<|image_2|>\nDescribe these images.<|end|>\n<|assistant|>\n" # noqa: E501
models = ["microsoft/Phi-3.5-vision-instruct"]
def vllm_to_hf_output(vllm_output: Tuple[List[int], str,
Optional[SampleLogprobs]],
model: str):
"""Sanitize vllm output to be comparable with hf output."""
_, output_str, out_logprobs = vllm_output
output_str_without_image = re.sub(r"(<\|image_\d+\|>)+", "", output_str)
assert output_str_without_image[0] == " "
output_str_without_image = output_str_without_image[1:]
hf_output_str = output_str_without_image + "<|end|><|endoftext|>"
tokenizer = AutoTokenizer.from_pretrained(model)
hf_output_ids = tokenizer.encode(output_str_without_image)
assert hf_output_ids[0] == 1
hf_output_ids = hf_output_ids[1:]
return hf_output_ids, hf_output_str, out_logprobs
target_dtype = "half"
if is_cpu():
target_dtype = "bfloat16"
# ROCm Triton FA can run into shared memory issues with these models,
# use other backends in the meantime
# FIXME (mattwong, gshtrasb, hongxiayan)
if is_hip():
os.environ["VLLM_USE_TRITON_FLASH_ATTN"] = "0"
def run_test(
hf_runner: Type[HfRunner],
vllm_runner: Type[VllmRunner],
inputs: List[Tuple[List[str], PromptImageInput]],
model: str,
*,
dtype: str,
max_tokens: int,
num_logprobs: int,
mm_limit: int,
tensor_parallel_size: int,
distributed_executor_backend: Optional[str] = None,
):
"""Inference result should be the same between hf and vllm.
All the image fixtures for the test are from IMAGE_ASSETS.
For huggingface runner, we provide the PIL images as input.
For vllm runner, we provide MultiModalDataDict objects
and corresponding MultiModalConfig as input.
Note, the text input is also adjusted to abide by vllm contract.
The text output is sanitized to be able to compare with hf.
"""
# NOTE: take care of the order. run vLLM first, and then run HF.
# vLLM needs a fresh new process without cuda initialization.
# if we run HF first, the cuda initialization will be done and it
# will hurt multiprocessing backend with fork method (the default method).
# max_model_len should be greater than image_feature_size
with vllm_runner(model,
max_model_len=4096,
max_num_seqs=1,
dtype=dtype,
limit_mm_per_prompt={"image": mm_limit},
tensor_parallel_size=tensor_parallel_size,
distributed_executor_backend=distributed_executor_backend,
enforce_eager=True) as vllm_model:
vllm_outputs_per_case = [
vllm_model.generate_greedy_logprobs(prompts,
max_tokens,
num_logprobs=num_logprobs,
images=images)
for prompts, images in inputs
]
# use eager mode for hf runner, since phi3_v didn't work with flash_attn
hf_model_kwargs = {"_attn_implementation": "eager"}
with hf_runner(model, dtype=dtype,
model_kwargs=hf_model_kwargs) as hf_model:
eos_token_id = hf_model.processor.tokenizer.eos_token_id
hf_outputs_per_case = [
hf_model.generate_greedy_logprobs_limit(prompts,
max_tokens,
num_logprobs=num_logprobs,
images=images,
eos_token_id=eos_token_id)
for prompts, images in inputs
]
for hf_outputs, vllm_outputs in zip(hf_outputs_per_case,
vllm_outputs_per_case):
check_logprobs_close(
outputs_0_lst=hf_outputs,
outputs_1_lst=[
vllm_to_hf_output(vllm_output, model)
for vllm_output in vllm_outputs
],
name_0="hf",
name_1="vllm",
)
# Since we use _attn_implementation="eager" for hf_runner, there is more
# significant numerical difference. The basic `logprobs=5` fails to pass.
@pytest.mark.parametrize("model", models)
@pytest.mark.parametrize(
"size_factors",
[
# No image
[],
# Single-scale
[1.0],
# Single-scale, batched
[1.0, 1.0, 1.0],
# Multi-scale
[0.25, 0.5, 1.0],
],
)
@pytest.mark.parametrize("dtype", [target_dtype])
@pytest.mark.parametrize("max_tokens", [128])
@pytest.mark.parametrize("num_logprobs", [10])
def test_models(hf_runner, vllm_runner, image_assets, model, size_factors,
dtype: str, max_tokens: int, num_logprobs: int) -> None:
images = [asset.pil_image for asset in image_assets]
inputs_per_image = [(
[prompt for _ in size_factors],
[rescale_image_size(image, factor) for factor in size_factors],
) for image, prompt in zip(images, HF_IMAGE_PROMPTS)]
run_test(
hf_runner,
vllm_runner,
inputs_per_image,
model,
dtype=dtype,
max_tokens=max_tokens,
num_logprobs=num_logprobs,
mm_limit=1,
tensor_parallel_size=1,
)
@pytest.mark.parametrize("model", models)
@pytest.mark.parametrize("dtype", [target_dtype])
def test_regression_7840(hf_runner, vllm_runner, image_assets, model,
dtype) -> None:
images = [asset.pil_image for asset in image_assets]
inputs_regresion_7840 = [
([prompt], [image]) for image, prompt in zip(images, HF_IMAGE_PROMPTS)
]
# Regression test for #7840.
run_test(
hf_runner,
vllm_runner,
inputs_regresion_7840,
model,
dtype=dtype,
max_tokens=128,
num_logprobs=10,
mm_limit=1,
tensor_parallel_size=1,
)
@pytest.mark.parametrize("model", models)
@pytest.mark.parametrize(
"size_factors",
[
# No image
[],
# Single-scale
[1.0],
# Single-scale, batched
[1.0, 1.0, 1.0],
# Multi-scale
[0.25, 0.5, 1.0],
],
)
@pytest.mark.parametrize("dtype", [target_dtype])
@pytest.mark.parametrize("max_tokens", [128])
@pytest.mark.parametrize("num_logprobs", [10])
def test_multi_images_models(hf_runner, vllm_runner, image_assets, model,
size_factors, dtype: str, max_tokens: int,
num_logprobs: int) -> None:
images = [asset.pil_image for asset in image_assets]
inputs_per_case = [
([HF_MULTIIMAGE_IMAGE_PROMPT for _ in size_factors],
[[rescale_image_size(image, factor) for image in images]
for factor in size_factors])
]
run_test(
hf_runner,
vllm_runner,
inputs_per_case,
model,
dtype=dtype,
max_tokens=max_tokens,
num_logprobs=num_logprobs,
mm_limit=2,
tensor_parallel_size=1,
)
@@ -0,0 +1,199 @@
"""Compare the outputs of HF and vLLM for Mistral models using greedy sampling.
Run `pytest tests/models/test_mistral.py`.
"""
import json
import uuid
from dataclasses import asdict
from typing import TYPE_CHECKING, Any, Dict, List, Optional, Tuple
import pytest
from mistral_common.protocol.instruct.messages import ImageURLChunk
from mistral_common.protocol.instruct.request import ChatCompletionRequest
from mistral_common.tokens.tokenizers.mistral import MistralTokenizer
from mistral_common.tokens.tokenizers.multimodal import image_from_chunk
from vllm import EngineArgs, LLMEngine, SamplingParams, TokensPrompt
from vllm.multimodal import MultiModalDataBuiltins
from vllm.sequence import Logprob, SampleLogprobs
from ....utils import VLLM_PATH
from ...utils import check_logprobs_close
if TYPE_CHECKING:
from _typeshed import StrPath
MODELS = ["mistralai/Pixtral-12B-2409"]
IMG_URLS = [
"https://picsum.photos/id/237/400/300",
"https://picsum.photos/id/231/200/300",
"https://picsum.photos/id/27/500/500",
"https://picsum.photos/id/17/150/600",
]
PROMPT = "Describe each image in one short sentence."
def _create_msg_format(urls: List[str]) -> List[Dict[str, Any]]:
return [{
"role":
"user",
"content": [{
"type": "text",
"text": PROMPT,
}] + [{
"type": "image_url",
"image_url": {
"url": url
}
} for url in urls],
}]
def _create_engine_inputs(urls: List[str]) -> TokensPrompt:
msg = _create_msg_format(urls)
tokenizer = MistralTokenizer.from_model("pixtral")
request = ChatCompletionRequest(messages=msg) # type: ignore[type-var]
tokenized = tokenizer.encode_chat_completion(request)
engine_inputs = TokensPrompt(prompt_token_ids=tokenized.tokens)
images = []
for chunk in request.messages[0].content:
if isinstance(chunk, ImageURLChunk):
images.append(image_from_chunk(chunk))
mm_data = MultiModalDataBuiltins(image=images)
engine_inputs["multi_modal_data"] = mm_data
return engine_inputs
MSGS = [
_create_msg_format(IMG_URLS[:1]),
_create_msg_format(IMG_URLS[:2]),
_create_msg_format(IMG_URLS),
]
ENGINE_INPUTS = [
_create_engine_inputs(IMG_URLS[:1]),
_create_engine_inputs(IMG_URLS[:2]),
_create_engine_inputs(IMG_URLS),
]
SAMPLING_PARAMS = SamplingParams(max_tokens=512, temperature=0.0, logprobs=5)
LIMIT_MM_PER_PROMPT = dict(image=4)
MAX_MODEL_LEN = [8192, 65536]
FIXTURES_PATH = VLLM_PATH / "tests/models/fixtures"
assert FIXTURES_PATH.exists()
FIXTURE_LOGPROBS_CHAT = FIXTURES_PATH / "pixtral_chat.json"
FIXTURE_LOGPROBS_ENGINE = FIXTURES_PATH / "pixtral_chat_engine.json"
OutputsLogprobs = List[Tuple[List[int], str, Optional[SampleLogprobs]]]
# For the test author to store golden output in JSON
def _dump_outputs_w_logprobs(
outputs: OutputsLogprobs,
filename: "StrPath",
) -> None:
json_data = [(tokens, text,
[{k: asdict(v)
for k, v in token_logprobs.items()}
for token_logprobs in (logprobs or [])])
for tokens, text, logprobs in outputs]
with open(filename, "w") as f:
json.dump(json_data, f)
def load_outputs_w_logprobs(filename: "StrPath") -> OutputsLogprobs:
with open(filename, "rb") as f:
json_data = json.load(f)
return [(tokens, text,
[{int(k): Logprob(**v)
for k, v in token_logprobs.items()}
for token_logprobs in logprobs])
for tokens, text, logprobs in json_data]
@pytest.mark.skip(
reason=
"Model is too big, test passed on A100 locally but will OOM on CI machine."
)
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("max_model_len", MAX_MODEL_LEN)
@pytest.mark.parametrize("dtype", ["bfloat16"])
def test_chat(
vllm_runner,
max_model_len: int,
model: str,
dtype: str,
) -> None:
EXPECTED_CHAT_LOGPROBS = load_outputs_w_logprobs(FIXTURE_LOGPROBS_CHAT)
with vllm_runner(
model,
dtype=dtype,
tokenizer_mode="mistral",
enable_chunked_prefill=False,
max_model_len=max_model_len,
limit_mm_per_prompt=LIMIT_MM_PER_PROMPT,
) as vllm_model:
outputs = []
for msg in MSGS:
output = vllm_model.model.chat(msg,
sampling_params=SAMPLING_PARAMS)
outputs.extend(output)
logprobs = vllm_runner._final_steps_generate_w_logprobs(outputs)
check_logprobs_close(outputs_0_lst=EXPECTED_CHAT_LOGPROBS,
outputs_1_lst=logprobs,
name_0="h100_ref",
name_1="output")
@pytest.mark.skip(
reason=
"Model is too big, test passed on A100 locally but will OOM on CI machine."
)
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("dtype", ["bfloat16"])
def test_model_engine(vllm_runner, model: str, dtype: str) -> None:
EXPECTED_ENGINE_LOGPROBS = load_outputs_w_logprobs(FIXTURE_LOGPROBS_ENGINE)
args = EngineArgs(
model=model,
tokenizer_mode="mistral",
enable_chunked_prefill=False,
limit_mm_per_prompt=LIMIT_MM_PER_PROMPT,
dtype=dtype,
)
engine = LLMEngine.from_engine_args(args)
engine.add_request(uuid.uuid4().hex, ENGINE_INPUTS[0], SAMPLING_PARAMS)
engine.add_request(uuid.uuid4().hex, ENGINE_INPUTS[1], SAMPLING_PARAMS)
outputs = []
count = 0
while True:
out = engine.step()
count += 1
for request_output in out:
if request_output.finished:
outputs.append(request_output)
if count == 2:
engine.add_request(uuid.uuid4().hex, ENGINE_INPUTS[2],
SAMPLING_PARAMS)
if not engine.has_unfinished_requests():
break
logprobs = vllm_runner._final_steps_generate_w_logprobs(outputs)
check_logprobs_close(outputs_0_lst=EXPECTED_ENGINE_LOGPROBS,
outputs_1_lst=logprobs,
name_0="h100_ref",
name_1="output")
@@ -0,0 +1,401 @@
import pathlib
from typing import Dict, List, Optional, Tuple, Type, Union
import pytest
import torch
from PIL.Image import Image
from vllm.config import ModelConfig
from vllm.inputs import InputContext, LLMInputs
from vllm.multimodal.base import MultiModalInputs
from vllm.multimodal.utils import cached_get_tokenizer, rescale_image_size
from ....conftest import (IMAGE_ASSETS, HfRunner, ImageAsset, PromptImageInput,
VllmRunner, _ImageAssets)
from ...utils import check_logprobs_close
text_only_models = [
"Qwen/Qwen-7B-Chat" # Has no visual component
]
multimodal_models = ["Qwen/Qwen-VL"]
HF_IMAGE_PROMPTS = IMAGE_ASSETS.prompts({
"stop_sign":
"Picture 1: <img></img>\nWhat's the content of the image?: ",
"cherry_blossom":
"Picture 1: <img></img>\nWhat is the season?: ",
})
HF_MULTIIMAGE_IMAGE_PROMPT = "Picture 1: <img></img>\nPicture 2: <img></img>\nCan you compare these images?\n" # noqa: E501
HF_MULTIIMAGE_IMAGE_PROMPT = "Picture 1: <img></img>\nPicture 2: <img></img>\nDescribe the two images in detail.\n" # noqa: E501
### Multimodal preprocessing tests
SAMPLE_IMAGE = IMAGE_ASSETS[0].pil_image
# These values are specific to Qwen-VL/Chat; we can get these from the model
# config also, but they are hardcoded here to keep the parameterize/fixtures
# easy to read.
IMG_START_ID = 151857
IMG_END_ID = 151858
IMG_PAD_ID = 151859
TOKS_PER_IMG = 256
VIS_ENC_DIM = 4096
IMG_SIZE = 448
def build_model_context(model_name: str,
tokenizer_name: Optional[str] = None,
trust_remote_code: bool = False):
"""Creates an InputContext for a given model.
Args:
model_name: Name of the model being considered.
tokenizer_name: Name of the tokenizer being considered.
trust_remote_code: Whether or not to allow loading remote code.
Returns:
InputContext for the model being considered.
"""
if tokenizer_name is None:
tokenizer_name = model_name
model_config = ModelConfig(
model_name,
tokenizer_name,
tokenizer_mode="auto",
trust_remote_code=trust_remote_code,
dtype="float32",
seed=0,
)
return InputContext(model_config)
@pytest.fixture()
def input_mapper_for_qwen():
# Lazy import to avoid initializing CUDA during test collection
from vllm.model_executor.models.qwen import input_mapper_for_qwen
return input_mapper_for_qwen
@pytest.fixture()
def input_processor_for_qwen():
# Lazy import to avoid initializing CUDA during test collection
from vllm.model_executor.models.qwen import input_processor_for_qwen
return input_processor_for_qwen
@pytest.fixture()
def qwen_vl_context() -> InputContext:
"""Get an InputContext for Qwen-VL."""
return build_model_context(model_name="Qwen/Qwen-VL",
trust_remote_code=True)
# Happy path tests for single/multi-image scenarios for the multimodal
# input processor and mapper, respectively
@pytest.mark.parametrize("num_images", [1, 2])
def test_input_processor_valid_mm_data(input_processor_for_qwen,
qwen_vl_context: InputContext,
num_images: int):
"""Happy cases for image inputs to Qwen's multimodal input processor."""
prompt = "".join(
[f"Picture {num}: <img></img>\n" for num in range(1, num_images + 1)])
inputs = LLMInputs(
prompt=prompt,
# When processing multimodal data for a multimodal model, the qwen
# input processor will overwrite the provided prompt_token_ids with
# the image prompts
prompt_token_ids=None,
multi_modal_data={"image": torch.rand(num_images, TOKS_PER_IMG, 4096)},
)
proc_inputs = input_processor_for_qwen(qwen_vl_context, inputs)
assert isinstance(proc_inputs, dict)
# Each image should have one start / stop and a fixed context of 256
proc_tokens = proc_inputs["prompt_token_ids"]
assert proc_tokens.count(IMG_START_ID) == num_images
assert proc_tokens.count(IMG_END_ID) == num_images
assert proc_tokens.count(IMG_PAD_ID) == num_images * TOKS_PER_IMG
@pytest.mark.parametrize(
"img_data,expected_shape",
[
# single / multi-image
(SAMPLE_IMAGE, (1, 3, IMG_SIZE, IMG_SIZE)),
(2 * [SAMPLE_IMAGE], (2, 3, IMG_SIZE, IMG_SIZE)),
# single / multi-image embeddings
(torch.rand(
(TOKS_PER_IMG, VIS_ENC_DIM)), (1, TOKS_PER_IMG, VIS_ENC_DIM)),
(torch.rand(
(1, TOKS_PER_IMG, VIS_ENC_DIM)), (1, TOKS_PER_IMG, VIS_ENC_DIM)),
(torch.rand(
(2, TOKS_PER_IMG, VIS_ENC_DIM)), (2, TOKS_PER_IMG, VIS_ENC_DIM)),
])
def test_input_mapper_valid_mm_data(input_mapper_for_qwen,
qwen_vl_context: InputContext,
img_data: Union[torch.Tensor, List[Image],
Image],
expected_shape: List[int]):
"""Happy cases for image inputs to Qwen's multimodal input mapper."""
mapped_img_data = input_mapper_for_qwen(qwen_vl_context, img_data)
# Ensure that we get the appropriately shaped pixel_values
# for images and image embeddings, respectively.
assert isinstance(mapped_img_data, MultiModalInputs)
assert "pixel_values" in mapped_img_data
assert mapped_img_data["pixel_values"].shape == expected_shape
# Sad path tests for the multimodal input processor and mapper, respectively
@pytest.mark.parametrize("mm_data", [
{
"image": torch.rand((5))
},
{
"image": torch.rand((5, 5, 5, 5, 5))
},
])
def test_input_processor_invalid_mm_data(input_processor_for_qwen,
qwen_vl_context: InputContext,
mm_data: Dict[str, torch.Tensor]):
"""Test sad cases validated in Qwen's multimodal input processor."""
tokenizer = cached_get_tokenizer(qwen_vl_context.model_config.tokenizer,
trust_remote_code=True)
prompt = "Picture 1: <img></img>\n"
prompt_token_ids = tokenizer.encode(prompt)
inputs = LLMInputs(prompt=prompt,
prompt_token_ids=prompt_token_ids,
multi_modal_data=mm_data)
# Should fail since we have too many or too few dimensions for embeddings
with pytest.raises(ValueError):
input_processor_for_qwen(qwen_vl_context, inputs)
@pytest.mark.parametrize(
"img_data",
[
# Wrong context length
torch.rand((1, TOKS_PER_IMG + 10, VIS_ENC_DIM)),
# Wrong visual encoder output size
torch.rand((1, TOKS_PER_IMG, VIS_ENC_DIM + 10)),
])
def test_input_mapper_invalid_mm_data(
input_mapper_for_qwen,
qwen_vl_context: InputContext,
img_data: Union[torch.Tensor, List[Image], Image],
):
"""Sad cases validated in Qwen VL's multimodal input mapper."""
with pytest.raises(ValueError):
input_mapper_for_qwen(qwen_vl_context, img_data)
### End-to-end generation tests
def get_prompt_with_path(tmp_path: pathlib.PosixPath, prompt: str,
assets: Union[_ImageAssets, List[ImageAsset]]) -> str:
"""Given a temporary dir path, export one or more image assets into the
tempdir & replace its contents with the local path to the string so that
the HF version of Qwen-VL can resolve the path and load the image ni its
forward() call.
Args:
tmp_path: Tempdir for test under consideration.
prompt: Prompt with image placeholders.
assets: List of image assets whose len equals the num placeholders.
"""
# Ensure that the number of placeholders matches the number of assets;
# If this is not true, the test is probably written incorrectly.
assert prompt.count("<img></img>") == len(assets)
# Replace the placeholders with local paths to the exported assets
for asset in assets:
image_tmp_path = tmp_path / f"{asset.name}.jpg"
asset.pil_image.save(image_tmp_path)
prompt = prompt.replace(
"<img></img>",
f"<img>{image_tmp_path}</img>",
1,
)
return prompt
def run_test(
hf_runner: Type[HfRunner],
vllm_runner: Type[VllmRunner],
inputs: List[Tuple[List[str], PromptImageInput]],
model: str,
*,
dtype: str,
max_tokens: int,
num_logprobs: int,
mm_limit: int,
tensor_parallel_size: int,
distributed_executor_backend: Optional[str] = None,
):
"""Inference result should be the same between hf and vllm.
All the image fixtures for the test is under tests/images.
For huggingface runner, we provide the PIL images as input.
For vllm runner, we provide MultiModalDataDict objects
and corresponding MultiModalConfig as input.
Note, the text input is also adjusted to abide by vllm contract.
The text output is sanitized to be able to compare with hf.
"""
# NOTE: take care of the order. run vLLM first, and then run HF.
# vLLM needs a fresh new process without cuda initialization.
# if we run HF first, the cuda initialization will be done and it
# will hurt multiprocessing backend with fork method (the default method).
# max_model_len should be greater than image_feature_size
# Qwen encodes each image into a fixed content size of 256
with vllm_runner(model,
max_model_len=1024,
max_num_seqs=1,
dtype=dtype,
limit_mm_per_prompt={"image": mm_limit},
tensor_parallel_size=tensor_parallel_size,
distributed_executor_backend=distributed_executor_backend,
enforce_eager=True) as vllm_model:
vllm_outputs_per_image = [
vllm_model.generate_greedy_logprobs(prompts,
max_tokens,
num_logprobs=num_logprobs,
images=images)
for prompts, images in inputs
]
with hf_runner(model, dtype=dtype) as hf_model:
hf_outputs_per_image = [
hf_model.generate_greedy_logprobs_limit(prompts,
max_tokens,
num_logprobs=num_logprobs,
images=images)
for prompts, images in inputs
]
for hf_outputs, vllm_outputs in zip(hf_outputs_per_image,
vllm_outputs_per_image):
check_logprobs_close(
outputs_0_lst=hf_outputs,
outputs_1_lst=vllm_outputs,
name_0="hf",
name_1="vllm",
)
@pytest.mark.parametrize("model", multimodal_models)
@pytest.mark.parametrize(
"size_factors",
[
# No image
[],
# Single-scale
[1.0],
# Single-scale, batched
[1.0, 1.0, 1.0],
# Multi-scale
[0.25, 0.5, 1.0],
],
)
@pytest.mark.parametrize("dtype", ["bfloat16"])
@pytest.mark.parametrize("max_tokens", [8])
@pytest.mark.parametrize("num_logprobs", [5])
def test_multimodal_models_single_image(tmp_path: pathlib.PosixPath,
hf_runner: Type[HfRunner],
vllm_runner: Type[VllmRunner],
image_assets: _ImageAssets, model: str,
size_factors: List[float], dtype: str,
max_tokens: int,
num_logprobs: int) -> None:
"""Tests multimodal models with single image prompts."""
images = [asset.pil_image for asset in image_assets]
prompts = [
get_prompt_with_path(tmp_path, prompt, [asset])
for prompt, asset in zip(HF_IMAGE_PROMPTS, image_assets)
]
inputs = [(
[prompt for _ in size_factors],
[rescale_image_size(image, factor) for factor in size_factors],
) for image, prompt in zip(images, prompts)]
run_test(
hf_runner,
vllm_runner,
inputs,
model,
dtype=dtype,
max_tokens=max_tokens,
num_logprobs=num_logprobs,
mm_limit=1,
tensor_parallel_size=1,
)
@pytest.mark.parametrize("model", multimodal_models)
@pytest.mark.parametrize(
"size_factors",
[
# No image
[],
# Single-scale
[1.0],
# Single-scale, batched
[1.0, 1.0, 1.0],
# Multi-scale
[0.25, 0.5, 1.0],
],
)
@pytest.mark.parametrize("dtype", ["bfloat16"])
@pytest.mark.parametrize("max_tokens", [128])
@pytest.mark.parametrize("num_logprobs", [5])
def test_multimodal_models_multi_image(tmp_path: pathlib.PosixPath,
hf_runner: Type[HfRunner],
vllm_runner: Type[VllmRunner],
image_assets: _ImageAssets, model: str,
size_factors: List[float], dtype: str,
max_tokens: int,
num_logprobs: int) -> None:
"""Tests multimodal models with multi-image prompts."""
images = [asset.pil_image for asset in image_assets]
# Put all of the images into one prompt.
prompt = get_prompt_with_path(tmp_path, HF_MULTIIMAGE_IMAGE_PROMPT,
image_assets)
inputs = [([prompt for _ in size_factors],
[[rescale_image_size(image, factor) for image in images]
for factor in size_factors])]
run_test(
hf_runner,
vllm_runner,
inputs,
model,
dtype=dtype,
max_tokens=max_tokens,
num_logprobs=num_logprobs,
mm_limit=2,
tensor_parallel_size=1,
)
# Ensure that a text-only Qwen model can still be loaded and
# used for inference in VLLM without throwing.
@pytest.mark.parametrize("model", text_only_models)
@pytest.mark.parametrize("dtype", ["bfloat16"])
@pytest.mark.parametrize("max_tokens", [32])
@pytest.mark.parametrize("num_logprobs", [5])
def test_text_only_qwen_model_can_be_loaded_and_run(
vllm_runner: Type[VllmRunner],
example_prompts: List[str],
model: str,
*,
dtype: str,
max_tokens: int,
num_logprobs: int,
):
with vllm_runner(model, dtype=dtype) as vllm_model:
vllm_model.generate_greedy_logprobs(
example_prompts,
max_tokens,
num_logprobs=num_logprobs,
)