Remove V0 Encoder-Decoder Support (#24907)

Signed-off-by: Woosuk Kwon <woosuk@thinkingmachines.ai>
This commit is contained in:
Woosuk Kwon
2025-09-15 21:17:14 -07:00
committed by GitHub
parent 5206ab20ba
commit 759ef49b15
47 changed files with 13 additions and 9661 deletions
+1 -154
View File
@@ -3,15 +3,12 @@
import pytest
from vllm.core.block.utils import (STR_NOT_IMPL_ENC_DEC_PREFIX_CACHE,
STR_NOT_IMPL_ENC_DEC_SWA)
from vllm.core.block_manager import SelfAttnBlockSpaceManager
from vllm.core.interfaces import AllocStatus
from vllm.sequence import Logprob, SequenceStatus
from vllm.utils import chunk_list
from ..utils import (create_dummy_prompt, create_seq_group,
create_seq_group_encoder_decoder)
from ..utils import create_dummy_prompt, create_seq_group
@pytest.mark.parametrize("block_size", [16])
@@ -58,156 +55,6 @@ def test_can_allocate_seq_group(block_size: int, num_seqs_per_group: int,
assert can_allocate_result == AllocStatus.LATER
@pytest.mark.parametrize("block_size", [16])
@pytest.mark.parametrize("num_gpu_blocks", [16, 80, 160])
@pytest.mark.parametrize("num_seqs_per_group", [1, 4])
@pytest.mark.parametrize("watermark", [0.0, 0.5])
def test_can_allocate_seq_group_encoder_decoder(block_size: int,
num_seqs_per_group: int,
num_gpu_blocks: int,
watermark: float):
block_manager = SelfAttnBlockSpaceManager(
block_size=block_size,
num_gpu_blocks=num_gpu_blocks,
num_cpu_blocks=1024,
watermark=watermark,
)
num_watermark_blocks = int(watermark * num_gpu_blocks)
num_output_blocks_per_seq = 1
# NOTE: This should be num_output_blocks_per_seq * num_seqs_per_group, but
# the current implementation assumes all seqs are new prompts / don't have
# different output lens.
num_output_blocks = num_output_blocks_per_seq
for bdx, num_prompt_blocks in enumerate(
range(1, num_gpu_blocks - num_output_blocks)):
num_cross_blocks_per_seq = num_prompt_blocks
seq_group = create_seq_group_encoder_decoder(
seq_prompt_len=block_size * num_prompt_blocks,
seq_output_lens=[
block_size * num_output_blocks_per_seq
for _ in range(num_seqs_per_group)
],
request_id=str(bdx))
assert num_prompt_blocks + num_output_blocks <= num_gpu_blocks
can_allocate_result = block_manager.can_allocate(seq_group)
num_required_blocks = num_prompt_blocks + \
num_output_blocks + \
num_cross_blocks_per_seq
if num_gpu_blocks - num_required_blocks < num_watermark_blocks:
assert can_allocate_result == AllocStatus.NEVER
elif num_gpu_blocks >= num_required_blocks:
assert can_allocate_result == AllocStatus.OK
else:
assert can_allocate_result == AllocStatus.LATER
@pytest.mark.parametrize("block_size", [16])
@pytest.mark.parametrize("num_gpu_blocks", [16])
@pytest.mark.parametrize("num_seqs_per_group", [1])
@pytest.mark.parametrize("watermark", [0.0, 0.5])
def test_can_allocate_encoder_decoder_fails_with_swa(block_size: int,
num_seqs_per_group: int,
num_gpu_blocks: int,
watermark: float):
'''
SWA short for Sliding Window Attention.
At time of writing block manager does not support SWA.
However even when SWA is implemented for block manager,
there will still most likely be a separate workstream required
to enable SWA for encoder/decoder models.
Therefore this test enforces that one of the following cases
hold true:
1. Block manager does not support SWA at all (true at time of writing)
2. Block manager fails with NotImplementError when SWA is enabled
AND a SequenceGroup with an encoder sequence (i.e. in support of an
encoder/decoder model) is passed into can_allocate() as an argument
The setup for this test is stripped down version of
test_can_allocate_seq_group_encoder_decoder()
'''
with pytest.raises((NotImplementedError, AssertionError)) as exc_info:
block_manager = SelfAttnBlockSpaceManager(
block_size=block_size,
num_gpu_blocks=num_gpu_blocks,
num_cpu_blocks=1024,
watermark=watermark,
sliding_window=5 # SWA
)
num_output_blocks_per_seq = 1
num_prompt_blocks = 1
num_output_blocks = num_output_blocks_per_seq
seq_group = create_seq_group_encoder_decoder(
seq_prompt_len=block_size * num_prompt_blocks,
seq_output_lens=[
block_size * num_output_blocks_per_seq
for _ in range(num_seqs_per_group)
],
request_id="0")
assert num_prompt_blocks + num_output_blocks <= num_gpu_blocks
block_manager.can_allocate(seq_group)
# Assert that either
# 1. Block manager constructor fails with assertion that sliding window
# is not yet supported (most likely near-term outcome at time of
# writing), or
# 2. can_allocate() fails with NotImplementedError due to combination of
# encoder/decoder and sliding window attention
if isinstance(exc_info.value, NotImplementedError):
assert str(exc_info.value) == STR_NOT_IMPL_ENC_DEC_SWA
elif isinstance(exc_info.value, AssertionError):
assert str(exc_info.value) == "Sliding window not yet supported"
@pytest.mark.parametrize("block_size", [16])
@pytest.mark.parametrize("num_gpu_blocks", [16])
@pytest.mark.parametrize("num_seqs_per_group", [1])
@pytest.mark.parametrize("watermark", [0.0, 0.5])
def test_can_allocate_encoder_decoder_fails_with_prefix_cache(
block_size: int, num_seqs_per_group: int, num_gpu_blocks: int,
watermark: float):
block_manager = SelfAttnBlockSpaceManager(
block_size=block_size,
num_gpu_blocks=num_gpu_blocks,
num_cpu_blocks=1024,
watermark=watermark,
enable_caching=True # Prefix cache
)
num_output_blocks_per_seq = 1
num_prompt_blocks = 1
num_output_blocks = num_output_blocks_per_seq
seq_group = create_seq_group_encoder_decoder(
seq_prompt_len=block_size * num_prompt_blocks,
seq_output_lens=[
block_size * num_output_blocks_per_seq
for _ in range(num_seqs_per_group)
],
request_id="0")
assert num_prompt_blocks + num_output_blocks <= num_gpu_blocks
# Assert that either can_allocate() fails with NotImplementedError
# due to combination of encoder/decoder and prefix cache
with pytest.raises(NotImplementedError) as exc_info:
block_manager.can_allocate(seq_group)
assert str(exc_info.value) == STR_NOT_IMPL_ENC_DEC_PREFIX_CACHE
@pytest.mark.parametrize("block_size", [1, 8])
@pytest.mark.parametrize("prompt_len", [1, 7, 8])
@pytest.mark.parametrize("num_slots_to_append", [1, 8, 129])
@@ -1,105 +0,0 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import pytest # noqa
from vllm.config import CacheConfig, SchedulerConfig
from vllm.core.scheduler import Scheduler
from vllm.sequence import SequenceGroup
from .utils import (append_new_token, create_dummy_prompt_encoder_decoder,
get_sequence_groups, schedule_and_update_computed_tokens)
def test_scheduler_schedule_simple_encoder_decoder():
'''
Test basic scheduler functionality in the context
of an encoder/decoder model. Focus on testing
enc/dec-specific functionality sense tests already
exist for decoder-only functionality
Test behavior:
* Construct Scheduler
* Construct dummy encoder/decoder sequence groups
* Add dummy seq groups to scheduler backlog
* Schedule the next seq group & validate:
* Cross-attn block tables
* Updated states of seq groups
* Number of batched tokens
* Number of blocks to copy/swap-in/swap-out
* Number of scheduled seq groups
* Repeat for both prefill- and decode-phase
* Abort scheduled seq groups
* Assert that aborted seq groups no longer appear in
cross-attention block table
'''
block_size = 4
num_seq_group = 4
max_model_len = 16
scheduler_config = SchedulerConfig(
"generate",
max_num_batched_tokens=64,
max_num_seqs=num_seq_group,
max_model_len=max_model_len,
)
cache_config = CacheConfig(block_size, 1.0, 1, "auto")
cache_config.num_cpu_blocks = 16 # enc and dec prompts per seq_group
cache_config.num_gpu_blocks = 16 # enc and dec prompts per seq_group
scheduler = Scheduler(scheduler_config, cache_config, None)
running: list[SequenceGroup] = []
# Add seq groups to scheduler.
req_id_list = []
for i in range(num_seq_group):
req_id = str(i)
req_id_list.append(req_id)
_, _, seq_group = create_dummy_prompt_encoder_decoder(
req_id, block_size, block_size, block_size)
scheduler.add_seq_group(seq_group)
running.append(seq_group)
# Schedule seq groups prefill.
num_tokens = block_size * num_seq_group
seq_group_meta_list, out = schedule_and_update_computed_tokens(scheduler)
# - Verify that sequence group cross-attention block tables are
# registered with the block manager
assert all([(req_id in scheduler.block_manager.cross_block_tables)
for req_id in req_id_list])
# - Validate sequence-group status
assert set(get_sequence_groups(out)) == set(running)
# - Validate number of batched tokens
assert out.num_batched_tokens == num_tokens
# - Validate there are no remaining blocks to swap
assert (not out.blocks_to_copy and not out.blocks_to_swap_in
and not out.blocks_to_swap_out)
# - Validate all seq groups were scheduled
assert len(seq_group_meta_list) == num_seq_group
append_new_token(out, 1)
# Schedule seq groups decode.
seq_group_meta_list, out = schedule_and_update_computed_tokens(scheduler)
# - Verify that sequence group metadata includes encoder attention
# and cross-attention metadata
assert all([
not ((seq_group_meta.encoder_seq_data is None) or
(seq_group_meta.cross_block_table is None))
for seq_group_meta in seq_group_meta_list
])
# - Validate sequence-group status
assert set(get_sequence_groups(out)) == set(running)
# - Validate there is one batched token per seq group
assert out.num_batched_tokens == num_seq_group
# - Validate there are no remaining blocks to swap
assert (not out.blocks_to_copy and not out.blocks_to_swap_in
and not out.blocks_to_swap_out)
# - Validate that all seq groups were scheduled
assert len(seq_group_meta_list) == num_seq_group
append_new_token(out, 1)
# Abort sequences
for req_id in req_id_list:
scheduler.abort_seq_group(req_id)
# - Verify that sequence group cross-attention block tables are
# NO LONGER registered with the block manager
assert req_id not in scheduler.block_manager.cross_block_tables
@@ -242,9 +242,6 @@ MULTIMODAL_MODELS = {
"Qwen/Qwen2-Audio-7B-Instruct": PPTestSettings.fast(),
"Qwen/Qwen2-VL-2B-Instruct": PPTestSettings.fast(),
"fixie-ai/ultravox-v0_5-llama-3_2-1b": PPTestSettings.fast(),
# [Encoder-decoder]
# TODO: Implement PP
# "meta-llama/Llama-3.2-11B-Vision-Instruct": PPTestSettings.fast(),
}
# yapf: enable
View File
@@ -1,131 +0,0 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""E2E tests to verify the correctness of the encoder-decoder framework
Run `pytest tests/encoder_decoder/test_e2e_correctness.py`.
"""
from typing import Optional
import pytest
from transformers import AutoModelForSeq2SeqLM
from vllm.attention.selector import (_Backend, _cached_get_attn_backend,
global_force_attn_backend_context_manager)
from vllm.platforms import current_platform
from vllm.sequence import SampleLogprobs
from ..conftest import DecoderPromptType
from ..models.utils import check_logprobs_close
LIST_ENC_DEC_SUPPORTED_BACKENDS = [
_Backend.XFORMERS, _Backend.FLASH_ATTN, None
]
@pytest.fixture(scope="function", autouse=True)
def use_v0_only(monkeypatch):
"""
Since this module is V0 only, set VLLM_USE_V1=0 for
all tests in the module.
"""
monkeypatch.setenv('VLLM_USE_V1', '0')
def vllm_to_hf_output(
vllm_output: tuple[list[int], str, Optional[SampleLogprobs]],
decoder_prompt_type: DecoderPromptType,
):
"""Sanitize vllm output to be comparable with hf output."""
output_ids, output_str, out_logprobs = vllm_output
hf_output_str = output_str + "</s>"
if decoder_prompt_type == DecoderPromptType.NONE:
hf_output_str = "<s>" + hf_output_str
return output_ids, hf_output_str, out_logprobs
@pytest.fixture(autouse=True)
def clear_cache():
"""Fixture to clear backend cache before each test."""
_cached_get_attn_backend.cache_clear() # Clear the cache
yield # This allows the test to run
@pytest.mark.parametrize("model", ["facebook/bart-large-cnn"])
@pytest.mark.parametrize("dtype", ["float"])
@pytest.mark.parametrize("attn_backend", LIST_ENC_DEC_SUPPORTED_BACKENDS)
@pytest.mark.parametrize("max_tokens", [128])
@pytest.mark.parametrize("num_logprobs", [5])
@pytest.mark.parametrize("decoder_prompt_type", list(DecoderPromptType))
@pytest.mark.parametrize("enforce_eager", [True, False])
@pytest.mark.skipif(
current_platform.is_cpu(),
reason="CPU backend is not currently supported with encoder/decoder models"
)
@pytest.mark.skip(reason="bart not supported in V1")
def test_encoder_decoder_e2e(
hf_runner,
vllm_runner,
example_encoder_decoder_prompts,
model: str,
dtype: str,
max_tokens: int,
num_logprobs: int,
decoder_prompt_type: DecoderPromptType,
enforce_eager: bool,
attn_backend: _Backend,
) -> None:
'''
End-to-End (E2E) test for the encoder-decoder framework.
This test evaluates the encoder-decoder functionality using the BART
model. We compare the outputs of the Hugging Face and vLLM
implementations to ensure that both implementations produce consistent
and correct results.
'''
with global_force_attn_backend_context_manager(attn_backend):
if attn_backend == _Backend.FLASH_ATTN:
# Flash Attention works only with bfloat16 data-type
dtype = 'bfloat16'
test_case_prompts = example_encoder_decoder_prompts[
decoder_prompt_type]
# Configuration settings for HF baseline
hf_kwargs = {
"top_k": None,
"num_beams": 1,
"repetition_penalty": 1.0,
"top_p": 1.0,
"length_penalty": 1.0,
"early_stopping": False,
"no_repeat_ngram_size": None,
"min_length": 0
}
with hf_runner(model, dtype=dtype,
auto_cls=AutoModelForSeq2SeqLM) as hf_model:
hf_outputs = (
hf_model.generate_encoder_decoder_greedy_logprobs_limit(
test_case_prompts,
max_tokens,
num_logprobs,
**hf_kwargs,
))
with vllm_runner(model, dtype=dtype,
enforce_eager=enforce_eager) as vllm_model:
vllm_outputs = vllm_model.generate_encoder_decoder_greedy_logprobs(
test_case_prompts, max_tokens, num_logprobs)
hf_skip_tokens = (1 if decoder_prompt_type == DecoderPromptType.NONE
else 0)
check_logprobs_close(
outputs_0_lst=hf_outputs,
outputs_1_lst=[
vllm_to_hf_output(vllm_output, decoder_prompt_type)
for vllm_output in vllm_outputs
],
name_0="hf",
name_1="vllm",
num_outputs_0_skip_tokens=hf_skip_tokens,
)
@@ -1,56 +0,0 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import openai
import pytest
import pytest_asyncio
from ...utils import RemoteOpenAIServer
MODEL_NAME = "facebook/bart-base"
@pytest.fixture(scope="module")
def server():
args = [
"--dtype",
"bfloat16",
"--enforce-eager",
]
with RemoteOpenAIServer(MODEL_NAME, args) as remote_server:
yield remote_server
@pytest_asyncio.fixture
async def client(server):
async with server.get_async_client() as async_client:
yield async_client
@pytest.mark.asyncio
@pytest.mark.parametrize("model_name", [MODEL_NAME])
@pytest.mark.skip(reason="bart is not yet supported in V1")
async def test_single_completion(client: openai.AsyncOpenAI, model_name: str):
completion = await client.completions.create(model=model_name,
prompt="Hello, my name is",
max_tokens=5,
temperature=0.0)
assert completion.id is not None
assert completion.choices is not None and len(completion.choices) == 1
choice = completion.choices[0]
assert len(choice.text) >= 5
assert choice.finish_reason == "length"
assert completion.usage == openai.types.CompletionUsage(
completion_tokens=5, prompt_tokens=2, total_tokens=7)
# test using token IDs
completion = await client.completions.create(
model=model_name,
prompt=[0, 0, 0, 0, 0],
max_tokens=5,
temperature=0.0,
)
assert len(completion.choices[0].text) >= 1
-199
View File
@@ -20,7 +20,6 @@ from vllm.entrypoints.chat_utils import (_try_extract_ast, load_chat_template,
parse_chat_messages_futures,
resolve_chat_template_content_format,
resolve_hf_chat_template)
from vllm.entrypoints.llm import apply_hf_chat_template
from vllm.multimodal import MultiModalDataDict, MultiModalUUIDDict
from vllm.multimodal.utils import (encode_audio_base64, encode_image_base64,
encode_video_base64)
@@ -38,7 +37,6 @@ QWEN2AUDIO_MODEL_ID = "Qwen/Qwen2-Audio-7B-Instruct"
QWEN2VL_MODEL_ID = "Qwen/Qwen2-VL-2B-Instruct"
QWEN25VL_MODEL_ID = "Qwen/Qwen2.5-VL-3B-Instruct"
QWEN25OMNI_MODEL_ID = "Qwen/Qwen2.5-Omni-7B"
MLLAMA_MODEL_ID = "meta-llama/Llama-3.2-11B-Vision-Instruct"
LLAMA_GUARD_MODEL_ID = "meta-llama/Llama-Guard-3-1B"
HERMES_MODEL_ID = "NousResearch/Hermes-3-Llama-3.1-8B"
MISTRAL_MODEL_ID = "mistralai/Mistral-Small-3.1-24B-Instruct-2503"
@@ -125,27 +123,6 @@ def qwen25omni_tokenizer():
)
@pytest.fixture(scope="module")
def mllama_model_config():
return ModelConfig(
MLLAMA_MODEL_ID,
runner="generate",
limit_mm_per_prompt={
"image": 2,
},
)
@pytest.fixture(scope="module")
def mllama_tokenizer():
return TokenizerGroup(
MLLAMA_MODEL_ID,
enable_lora=False,
max_num_seqs=5,
max_input_length=None,
)
@pytest.fixture(scope="function")
def mistral_model_config():
return ModelConfig(
@@ -2249,180 +2226,6 @@ def test_parse_chat_messages_multiple_images_interleave_with_placeholders(
)
### Mllama currently wraps images / texts as interleaved dictionaries
def test_mllama_single_image(
mllama_model_config,
mllama_tokenizer,
image_url,
):
"""Ensures that a single image is parsed correctly mllama."""
conversation, mm_data, mm_uuids = parse_chat_messages(
[{
"role":
"user",
"content": [
{
"type": "text",
"text": "The content of this image is:"
},
{
"image_url": image_url
},
],
}],
mllama_model_config,
mllama_tokenizer,
content_format="openai",
)
_assert_mm_data_is_image_input(mm_data, 1)
_assert_mm_uuids(mm_uuids, 1, expected_uuids=[None])
assert conversation == [{
"role":
"user",
"content": [
{
"type": "text",
"text": "The content of this image is:"
},
{
"type": "image"
},
],
}]
def test_mllama_interleaved_images(
mllama_model_config,
mllama_tokenizer,
image_url,
):
"""Ensures that multiple image are parsed as interleaved dicts."""
conversation, mm_data, mm_uuids = parse_chat_messages(
[{
"role":
"user",
"content": [
{
"type": "text",
"text": "The content of the first image is:",
},
{
"image_url": image_url
},
{
"type": "text",
"text": "The content of the second image is:",
},
{
"image_url": image_url
},
],
}],
mllama_model_config,
mllama_tokenizer,
content_format="openai",
)
_assert_mm_data_is_image_input(mm_data, 2)
_assert_mm_uuids(mm_uuids, 2, expected_uuids=[None, None])
assert conversation == [{
"role":
"user",
"content": [
{
"type": "text",
"text": "The content of the first image is:"
},
{
"type": "image"
},
{
"type": "text",
"text": "The content of the second image is:"
},
{
"type": "image"
},
],
}]
@pytest.mark.parametrize("model", [MLLAMA_MODEL_ID])
def test_multimodal_image_parsing_matches_hf(model, image_url):
"""Checks end to end hf alignment for multimodal [image] parsing."""
def get_conversation(is_hf: bool):
img_part = {"type": "image_url", "image_url": {"url": image_url}}
if is_hf:
img_part = {"type": "image"}
return [{
"role":
"user",
"content": [
{
"type": "text",
"text": "The content of the first image is:",
},
img_part,
{
"type": "text",
"text": "The content of the second image is:",
},
img_part,
{
"type": "text",
"text": "What animal is in the first image?",
},
],
}]
# Build a config for the model
model_config = ModelConfig(
model,
runner="generate",
limit_mm_per_prompt={
"image": 2,
},
)
# Build the tokenizer group and grab the underlying tokenizer
tokenizer_group = TokenizerGroup(
model,
enable_lora=False,
max_num_seqs=5,
max_input_length=None,
trust_remote_code=model_config.trust_remote_code,
)
tokenizer = tokenizer_group.tokenizer
# Build and parse a conversation with {"type": "image"} using the tokenizer
hf_conversation = get_conversation(is_hf=True)
hf_result = tokenizer.apply_chat_template(
hf_conversation,
tokenize=False,
add_generation_prompt=True,
)
# Now parse with vLLMs chat utils & apply the template
vllm_conversation = get_conversation(is_hf=False)
conversation, _, _ = parse_chat_messages(
vllm_conversation,
model_config,
tokenizer_group,
content_format="openai",
)
vllm_result = apply_hf_chat_template(
tokenizer=tokenizer,
conversation=conversation,
chat_template=None,
model_config=model_config,
tools=None,
add_generation_prompt=True,
)
assert hf_result == vllm_result
@pytest.mark.parametrize(
"model",
[
@@ -2486,7 +2289,6 @@ def test_resolve_hf_chat_template(sample_json_schema, model, use_tools):
(QWEN25VL_MODEL_ID, "openai"),
(ULTRAVOX_MODEL_ID, "string"),
(QWEN2AUDIO_MODEL_ID, "openai"),
(MLLAMA_MODEL_ID, "openai"),
(LLAMA_GUARD_MODEL_ID, "openai")],
)
# yapf: enable
@@ -2545,7 +2347,6 @@ def test_resolve_content_format_hf_defined(model, expected_format):
[("Salesforce/blip2-opt-2.7b", "string"),
("facebook/chameleon-7b", "string"),
("deepseek-ai/deepseek-vl2-tiny", "string"),
("microsoft/Florence-2-base", "string"),
("adept/fuyu-8b", "string"),
("google/paligemma-3b-mix-224", "string"),
("Qwen/Qwen-VL", "string"),
File diff suppressed because it is too large Load Diff
@@ -1,222 +0,0 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from typing import Optional
import pytest
from transformers import AutoModelForSeq2SeqLM
from vllm.sequence import SampleLogprobs
from ....conftest import (DecoderPromptType, ExplicitEncoderDecoderPrompt,
HfRunner, VllmRunner)
from ....utils import multi_gpu_test
from ...utils import check_logprobs_close
def vllm_to_hf_output(
vllm_output: tuple[list[int], str, Optional[SampleLogprobs]],
decoder_prompt_type: DecoderPromptType,
):
"""Sanitize vllm output to be comparable with hf output."""
output_ids, output_str, out_logprobs = vllm_output
hf_output_str = output_str + "</s>"
if decoder_prompt_type == DecoderPromptType.NONE:
hf_output_str = "<s>" + hf_output_str
return output_ids, hf_output_str, out_logprobs
def run_test(
hf_runner: type[HfRunner],
vllm_runner: type[VllmRunner],
prompts: list[ExplicitEncoderDecoderPrompt[str, str]],
decoder_prompt_type: DecoderPromptType,
model: str,
*,
dtype: str,
max_tokens: int,
num_logprobs: int,
tensor_parallel_size: int,
distributed_executor_backend: Optional[str] = None,
) -> None:
'''
Test the vLLM BART model for a variety of encoder/decoder input prompts,
by validating it against HuggingFace (HF) BART.
Arguments:
* hf_runner: HuggingFace (HF) test model runner
* vllm_runner: vLLM test model runner
* example_encoder_decoder_prompts: test fixture which provides a
dictionary of dummy prompts
* model: the HF ID of the specific BART variant under test
* dtype: the tensor datatype to employ
* max_tokens
* num_logprobs
* decoder_prompt_type: key into the example_encoder_decoder_prompts
dictionary; selects specific encoder/decoder
prompt scenarios to test
A note on using HF BART as a baseline for validating vLLM BART,
specifically when the decoder prompt is None.
The HF GenerationMixin's default behavior is to force the first
decoded token to be <BOS> if the prompt does not already contain
<BOS> (this is accomplished using a logit
processor setting.)
So when we use HF BART as our baseline for comparison, note that
when the user provides a request with a None decoder prompt
(i.e. a singleton encoder prompt, or else an explicit encoder/
decoder prompt with the decoder sub-prompt set to None), HF and
vLLM handle this in different ways:
* HF will (1) tokenize the None prompt as an empty token-list,
(2) append <decoder-start-token> to the beginning, yielding
[<decoder-start-token>], (3) pass this token list to the model, and
then (4) after computing logits during prefill, override the model
logits & force <BOS> to be the first generated token.
* vLLM will (1) tokenize the None prompt as [<BOS>], (2) append decoder-
start-token to the beginning, yielding [<decoder-start-token><BOS>],
(3) pass these tokens to the model & proceed with generation.
The net effect is that compared to vLLM, the list of HF *decoded* tokens
will contain one more initial <BOS> than the vLLM generated tokens,
because vLLM's <BOS> token is injected into the prompt rather than into
the generated output. This is in spite of the fact that overall, the
complete sequences (prompt + decoded tokens) produced by vLLM will match
HF.
So when we use HF decoded token output to validate vLLM's decoded token
output, the testing process must account for the difference in decoded
token sequences between vLLM and HF specifically in the
decoder-prompt-is-None case.
One option is to disable the logit processor feature that forces the
<BOS> token to be decoded (forced_bos_token_id = None), eliminating
the problem entirely. However this is not "normal" BART usage.
The other option is - only in the decoder-prompt-is-None case - to
discard the first decoded token from the HF output before comparing it
to vLLM.
To that end, when testing the scenario where the decoder prompt is None
(and only in that one scenario), this test skips the first HF decoded
token during the process of validating the vLLM decoded output.
'''
# 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).
# Note: currently encoder/decoder models are only compatible with
# enforce_eager=True. Normally this is not a problem because
# for encoder/decoder models vLLM will
# default to enforce_eager=True if enforce_eager
# is left unspecified. However, the
# VllmRunner test fixture (which wraps around the LLM class) defaults to
# enforce_eager=False (a behavior which a number of already-existing
# decoder-only unit tests expect), so when testing an encoder/decoder
# model we must explicitly specify enforce_eager=True in the VllmRunner
# constructor.
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 = vllm_model.generate_encoder_decoder_greedy_logprobs(
prompts, max_tokens, num_logprobs)
# Configuration settings for HF baseline
hf_kwargs = {
"top_k": None,
"num_beams": 1,
"repetition_penalty": 1.0,
"top_p": 1.0,
"length_penalty": 1.0,
"early_stopping": False,
"no_repeat_ngram_size": None,
"min_length": 0
}
with hf_runner(model, dtype=dtype,
auto_cls=AutoModelForSeq2SeqLM) as hf_model:
hf_outputs = (hf_model.generate_encoder_decoder_greedy_logprobs_limit(
prompts,
max_tokens,
num_logprobs,
**hf_kwargs,
))
hf_skip_tokens = (1
if decoder_prompt_type == DecoderPromptType.NONE else 0)
check_logprobs_close(
outputs_0_lst=hf_outputs,
outputs_1_lst=[
vllm_to_hf_output(vllm_output, decoder_prompt_type)
for vllm_output in vllm_outputs
],
name_0="hf",
name_1="vllm",
num_outputs_0_skip_tokens=hf_skip_tokens,
)
@pytest.mark.parametrize(
"model",
[
pytest.param("facebook/bart-base",
marks=[pytest.mark.core_model, pytest.mark.cpu_model]),
pytest.param("facebook/bart-large-cnn"),
],
)
@pytest.mark.parametrize("dtype", ["float", "bfloat16"])
@pytest.mark.parametrize("max_tokens", [64])
@pytest.mark.parametrize("num_logprobs", [5])
@pytest.mark.parametrize("decoder_prompt_type", list(DecoderPromptType))
@pytest.mark.skip(reason="bart not supported in V1")
def test_models(hf_runner, vllm_runner, example_encoder_decoder_prompts, model,
dtype, max_tokens, num_logprobs, decoder_prompt_type) -> None:
run_test(
hf_runner,
vllm_runner,
example_encoder_decoder_prompts[decoder_prompt_type],
decoder_prompt_type,
model,
dtype=dtype,
max_tokens=max_tokens,
num_logprobs=num_logprobs,
tensor_parallel_size=1,
)
@multi_gpu_test(num_gpus=2)
@pytest.mark.parametrize("distributed_executor_backend", ["ray", "mp"])
@pytest.mark.parametrize("model", ["facebook/bart-large-cnn"])
@pytest.mark.parametrize("dtype", ["float"])
@pytest.mark.parametrize("max_tokens", [64])
@pytest.mark.parametrize("num_logprobs", [5])
@pytest.mark.parametrize("decoder_prompt_type", [DecoderPromptType.CUSTOM])
@pytest.mark.skip(reason="bart not supported in V1")
def test_models_distributed(hf_runner, vllm_runner,
example_encoder_decoder_prompts,
distributed_executor_backend, model, dtype,
max_tokens, num_logprobs,
decoder_prompt_type) -> None:
run_test(
hf_runner,
vllm_runner,
example_encoder_decoder_prompts[decoder_prompt_type],
decoder_prompt_type,
model,
dtype=dtype,
max_tokens=max_tokens,
num_logprobs=num_logprobs,
tensor_parallel_size=2,
distributed_executor_backend=distributed_executor_backend,
)
@@ -1,123 +0,0 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from typing import Optional
import pytest
from transformers import AutoModelForSeq2SeqLM
from vllm.sequence import SampleLogprobs
from ....conftest import DecoderPromptType, HfRunner, VllmRunner
from ...utils import check_logprobs_close
def vllm_to_hf_output(
vllm_output: tuple[list[int], str, Optional[SampleLogprobs]],
decoder_prompt_type: DecoderPromptType,
):
"""Sanitize vllm output to be comparable with hf output."""
output_ids, output_str, out_logprobs = vllm_output
hf_output_str = output_str + "</s>"
return output_ids, hf_output_str, out_logprobs
def run_test(
hf_runner: type[HfRunner],
vllm_runner: type[VllmRunner],
prompts: list[dict[str, str]],
decoder_prompt_type: DecoderPromptType,
model: str,
*,
dtype: str,
max_tokens: int,
num_logprobs: int,
tensor_parallel_size: int,
distributed_executor_backend: Optional[str] = None,
) -> None:
'''
Test the vLLM mBART model by validating it against HuggingFace (HF).
(Docstring content is omitted for brevity)
'''
vllm_prompts = prompts
if decoder_prompt_type == DecoderPromptType.NONE:
vllm_prompts = [{
"encoder_prompt": p['encoder_prompt'],
"decoder_prompt": ""
} for p in prompts]
vllm_kwargs = {
"hf_overrides": {
"architectures": ["MBartForConditionalGeneration"]
}
}
with vllm_runner(model,
dtype=dtype,
tensor_parallel_size=tensor_parallel_size,
distributed_executor_backend=distributed_executor_backend,
enforce_eager=True,
**vllm_kwargs) as vllm_model: # type: ignore
vllm_outputs = vllm_model.generate_encoder_decoder_greedy_logprobs(
vllm_prompts, max_tokens, num_logprobs)
hf_kwargs = {
"top_k": None,
"num_beams": 1,
"repetition_penalty": 1.0,
"top_p": 1.0,
"length_penalty": 1.0,
"early_stopping": False,
"no_repeat_ngram_size": None,
"min_length": 0
}
with hf_runner(model, dtype=dtype,
auto_cls=AutoModelForSeq2SeqLM) as hf_model:
hf_kwargs["decoder_start_token_id"] = (
hf_model.tokenizer.lang_code_to_id["ro_RO"])
hf_outputs = (
hf_model.generate_encoder_decoder_greedy_logprobs_limit(
prompts, # HF runner still uses the original prompts
max_tokens,
num_logprobs,
**hf_kwargs,
))
hf_skip_tokens = 0
check_logprobs_close(
outputs_0_lst=hf_outputs,
outputs_1_lst=[
vllm_to_hf_output(vllm_output, decoder_prompt_type)
for vllm_output in vllm_outputs
],
name_0="hf",
name_1="vllm",
num_outputs_0_skip_tokens=hf_skip_tokens,
)
@pytest.mark.parametrize(
"model",
[pytest.param("facebook/mbart-large-en-ro")],
)
@pytest.mark.parametrize("dtype", ["float", "bfloat16"])
@pytest.mark.parametrize("max_tokens", [64])
@pytest.mark.parametrize("num_logprobs", [5])
@pytest.mark.parametrize("decoder_prompt_type", list(DecoderPromptType))
def test_models(hf_runner, vllm_runner, example_encoder_decoder_prompts, model,
dtype, max_tokens, num_logprobs, decoder_prompt_type) -> None:
run_test(
hf_runner,
vllm_runner,
example_encoder_decoder_prompts[decoder_prompt_type],
decoder_prompt_type,
model,
dtype=dtype,
max_tokens=max_tokens,
num_logprobs=num_logprobs,
tensor_parallel_size=1,
)
@@ -1,147 +0,0 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from typing import Optional
import pytest
from PIL import Image
from vllm.inputs.data import ExplicitEncoderDecoderPrompt, TextPrompt
from vllm.multimodal.image import rescale_image_size
from vllm.sequence import SampleLogprobs
from ....conftest import IMAGE_ASSETS, HfRunner, ImageTestAssets, VllmRunner
from ...utils import check_logprobs_close
MODELS = ["microsoft/Florence-2-base"]
# Florence-2 model repo's tokenizer config is missing some special tokens.
# Therefore, we use a converted tokenizer from a forked repo
TOKENIZER = "Isotr0py/Florence-2-tokenizer"
HF_IMAGE_PROMPTS = IMAGE_ASSETS.prompts({
"stop_sign":
"<OD>", # special task token which will output special tokens
"cherry_blossom":
"Describe in detail what is shown in the image.",
})
def get_hf_images_prompts(
prompts_: list[ExplicitEncoderDecoderPrompt[str, TextPrompt]],
) -> tuple[list[ExplicitEncoderDecoderPrompt[str, str]], list[Image.Image]]:
prompts, images = [], []
for prompt in prompts_:
encoder_prompt = prompt["encoder_prompt"]
prompts.append(
ExplicitEncoderDecoderPrompt(
encoder_prompt=encoder_prompt["prompt"],
decoder_prompt=None,
))
images.append(encoder_prompt["multi_modal_data"]["image"])
return prompts, images
def hf_to_vllm_output(hf_output: tuple[list[int], str,
Optional[SampleLogprobs]]):
"""Sanitize hf output to be comparable with vllm output."""
output_ids, output_str, out_logprobs = hf_output
output_str = output_str.replace("</s>", "").replace("<s>", "")
return output_ids, output_str, out_logprobs
def run_test(
hf_runner: type[HfRunner],
vllm_runner: type[VllmRunner],
inputs: list[list[ExplicitEncoderDecoderPrompt]],
model: str,
*,
dtype: str,
max_tokens: int,
num_logprobs: int,
tensor_parallel_size: int,
distributed_executor_backend: Optional[str] = None,
) -> None:
with vllm_runner(model,
max_num_seqs=8,
tokenizer_name=TOKENIZER,
dtype=dtype,
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_encoder_decoder_greedy_logprobs(
prompts,
max_tokens,
num_logprobs=num_logprobs,
skip_special_tokens=False,
) for prompts in inputs
]
hf_inputs = [get_hf_images_prompts(prompts) for prompts in inputs]
with hf_runner(model, dtype=dtype, skip_tokenizer_init=True) as hf_model:
hf_model.model.get_output_embeddings = lambda: \
hf_model.model.language_model.lm_head
hf_outputs_per_case = [
hf_model.generate_encoder_decoder_greedy_logprobs_limit(
prompts, max_tokens, num_logprobs=num_logprobs, images=images)
for prompts, images in hf_inputs
]
for hf_outputs, vllm_outputs in zip(hf_outputs_per_case,
vllm_outputs_per_case):
check_logprobs_close(
outputs_0_lst=[hf_to_vllm_output(output) for output in hf_outputs],
outputs_1_lst=vllm_outputs,
name_0="hf",
name_1="vllm",
num_outputs_0_skip_tokens=1,
)
# FIXME: https://github.com/huggingface/transformers/issues/38358
@pytest.mark.skip("Model initialization fails")
@pytest.mark.core_model
@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", ["float"])
@pytest.mark.parametrize("max_tokens", [64])
@pytest.mark.parametrize("num_logprobs", [5])
def test_models(hf_runner: type[HfRunner], vllm_runner: type[VllmRunner],
image_assets: ImageTestAssets, model: str,
size_factors: list[int], dtype: str, max_tokens: int,
num_logprobs: int) -> None:
images = [asset.pil_image for asset in image_assets]
inputs_per_image = [[
ExplicitEncoderDecoderPrompt(
encoder_prompt=TextPrompt(
prompt=prompt,
multi_modal_data={"image": rescale_image_size(image, factor)}),
decoder_prompt=None,
) 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,
tensor_parallel_size=1,
)
@@ -1,768 +0,0 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from typing import Optional, overload
import pytest
import torch
from packaging.version import Version
from transformers import AutoConfig, AutoModelForImageTextToText, AutoTokenizer
from transformers import __version__ as TRANSFORMERS_VERSION
from vllm import LLM, SamplingParams
from vllm.attention.backends.flash_attn import FlashAttentionMetadata
from vllm.attention.selector import (_Backend, _cached_get_attn_backend,
global_force_attn_backend_context_manager)
from vllm.model_executor.models.mllama import MllamaForConditionalGeneration
from vllm.multimodal.image import rescale_image_size
from vllm.sequence import SampleLogprobs
from ....conftest import (IMAGE_ASSETS, HfRunner, ImageTestAssets,
PromptImageInput, VllmRunner)
from ....quantization.utils import is_quant_method_supported
from ....utils import (create_new_process_for_each_test, large_gpu_test,
multi_gpu_test)
from ...utils import check_logprobs_close
_LIMIT_IMAGE_PER_PROMPT = 3
MLLAMA_IMAGE_TOKEN_ID = 128256
LIST_ENC_DEC_SUPPORTED_BACKENDS = [_Backend.XFORMERS, _Backend.FLASH_ATTN]
HF_IMAGE_PROMPTS = IMAGE_ASSETS.prompts({
"stop_sign":
"<|image|><|begin_of_text|>The meaning of the image is",
"cherry_blossom":
"<|image|><|begin_of_text|>The city is",
})
text_only_prompts = [
"The color of the sky is blue but sometimes it can also be",
]
models = [
"meta-llama/Llama-3.2-11B-Vision-Instruct",
]
# Indices for inputs
TEXT_ONLY = '0'
IMAGE_AT_BEG = '1'
IMAGE_AT_MIDDLE = '2'
TWO_IMAGES = '3'
# Input tokenized
prompt_data = {
# Tell me a story
TEXT_ONLY: [41551, 757, 264, 3446],
# <|image|> What's the content of this image
IMAGE_AT_BEG:
[MLLAMA_IMAGE_TOKEN_ID, 3639, 596, 279, 2262, 315, 420, 2217, 220],
# Hello <|image|>What' the content of this image
IMAGE_AT_MIDDLE:
[9906, 220, MLLAMA_IMAGE_TOKEN_ID, 3923, 6, 279, 2262, 315, 420, 2217],
#<|image|>Is there a duck in this image?<|image|>What's the animal in this image? # noqa: E501
TWO_IMAGES: [
MLLAMA_IMAGE_TOKEN_ID, 3957, 1070, 264, 37085, 304, 420, 2217, 30,
MLLAMA_IMAGE_TOKEN_ID, 3923, 596, 279, 10065, 304, 420, 2217, 30
]
}
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 _get_inputs(
image_assets: ImageTestAssets,
*,
size_factors: Optional[list[float]] = None,
sizes: Optional[list[tuple[int, int]]] = None,
) -> list[tuple[list[str], PromptImageInput]]:
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 if size is not None else text_only_prompts[0]
for size in sizes
],
[
image.resize(size) if size is not None else None
for size in sizes
],
) for image, prompt in zip(images, HF_IMAGE_PROMPTS)]
if len(sizes) == 0:
inputs_per_image.append(
(text_only_prompts, [None] * len(text_only_prompts)))
else:
raise ValueError("You must provide either `size_factors` or `sizes`")
return inputs_per_image
@overload
def run_test(
hf_runner: type[HfRunner],
vllm_runner: type[VllmRunner],
image_assets: ImageTestAssets,
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: ImageTestAssets,
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: ImageTestAssets,
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,
):
_run_test(
hf_runner,
vllm_runner,
_get_inputs(image_assets, size_factors=size_factors, sizes=sizes),
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: 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=19212, # 3 max size images
max_num_seqs=3,
tensor_parallel_size=tensor_parallel_size,
distributed_executor_backend=distributed_executor_backend,
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,
model_kwargs={"device_map": "auto"},
auto_cls=AutoModelForImageTextToText) 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_to_hf_output(vllm_output, model)
for vllm_output in vllm_outputs
],
name_0="hf",
name_1="vllm",
)
@pytest.fixture(autouse=True)
def clear_cache():
"""Fixture to clear backend cache before each test."""
_cached_get_attn_backend.cache_clear() # Clear the cache
yield # This allows the test to run
@large_gpu_test(min_gb=48)
@pytest.mark.core_model
@pytest.mark.parametrize("model", models)
@pytest.mark.parametrize(
"sizes",
[
# Text only
[],
# Single-size
[(512, 512)],
# Single-size, batched
[(512, 512), (512, 512), (512, 512)],
# Multi-size, batched
[(512, 512), (1024, 512), (1536, 512), (2048, 512), (512, 1024),
(1024, 1024), (512, 1536), (512, 2028)],
# Multi-size, batched, including text only
[(512, 512), (1024, 512), (1536, 512), (2048, 512), (512, 1024),
(1024, 1024), (512, 1536), (512, 2028), None],
# mllama has 8 possible aspect ratios, carefully set the sizes
# to cover all of them
])
@pytest.mark.parametrize("dtype", ["bfloat16"])
@pytest.mark.parametrize("max_tokens", [128])
@pytest.mark.parametrize("num_logprobs", [5])
@pytest.mark.parametrize("attn_backend", LIST_ENC_DEC_SUPPORTED_BACKENDS)
@pytest.mark.skipif(
Version(TRANSFORMERS_VERSION) <= Version("4.55.2"),
reason="Transformers v4.55 has a regression issue on mllama, "
"see: https://github.com/huggingface/transformers/pull/40083")
def test_models_single_leading_image(hf_runner, vllm_runner, image_assets,
model, sizes, dtype, max_tokens,
num_logprobs,
attn_backend: _Backend) -> None:
with global_force_attn_backend_context_manager(attn_backend):
if attn_backend == _Backend.FLASH_ATTN:
# Flash Attention works only with bfloat16 data-type
dtype = 'bfloat16'
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,
)
@large_gpu_test(min_gb=48)
@pytest.mark.core_model
@pytest.mark.parametrize("model", models)
@pytest.mark.parametrize("dtype", ["bfloat16"])
@pytest.mark.parametrize("max_tokens", [128])
@pytest.mark.parametrize("num_logprobs", [5])
@pytest.mark.parametrize("attn_backend", LIST_ENC_DEC_SUPPORTED_BACKENDS)
@pytest.mark.skipif(
Version(TRANSFORMERS_VERSION) <= Version("4.55.2"),
reason="Transformers v4.55 has a regression issue on mllama, "
"see: https://github.com/huggingface/transformers/pull/40083")
def test_models_multi_leading_images(hf_runner, vllm_runner, image_assets,
model, dtype, max_tokens, num_logprobs,
attn_backend: _Backend) -> None:
stop_sign = image_assets[0].pil_image
cherry_blossom = image_assets[1].pil_image
inputs = [(
[
"<|image|><|image|><|begin_of_text|>Describe 2 images.", # noqa: E501
"<|image|><|image|><|begin_of_text|>Describe 2 images.", # noqa: E501
"<|image|><|image|><|image|><|begin_of_text|>Describe 3 images.", # noqa: E501
],
[
[stop_sign, cherry_blossom],
# Images with different sizes.
[
stop_sign.resize((512, 512)),
stop_sign,
],
[
stop_sign,
stop_sign.resize((512, 1536)),
cherry_blossom.resize((512, 1024)),
],
])]
with global_force_attn_backend_context_manager(attn_backend):
if attn_backend == _Backend.FLASH_ATTN:
# Flash Attention works only with bfloat16 data-type
dtype = 'bfloat16'
_run_test(
hf_runner,
vllm_runner,
inputs,
model,
dtype=dtype,
max_tokens=max_tokens,
num_logprobs=num_logprobs,
tensor_parallel_size=1,
)
@large_gpu_test(min_gb=48)
@pytest.mark.core_model
@pytest.mark.parametrize("model", models)
@pytest.mark.parametrize("dtype", ["bfloat16"])
@pytest.mark.parametrize("max_tokens", [128])
@pytest.mark.parametrize("num_logprobs", [5])
@pytest.mark.parametrize("attn_backend", LIST_ENC_DEC_SUPPORTED_BACKENDS)
@pytest.mark.skipif(
Version(TRANSFORMERS_VERSION) <= Version("4.55.2"),
reason="Transformers v4.55 has a regression issue on mllama, "
"see: https://github.com/huggingface/transformers/pull/40083")
def test_models_interleaved_images(hf_runner, vllm_runner, image_assets, model,
dtype, max_tokens, num_logprobs,
attn_backend: _Backend) -> None:
stop_sign = image_assets[0].pil_image
cherry_blossom = image_assets[1].pil_image
inputs = [(
[
"<|begin_of_text|>The content of the image <|image|> is", # noqa: E501
"<|begin_of_text|>Between the first image <|image|> and the second image<|image|>, " # noqa: E501
"which is a stop sign and which is a cherry blossom?", # noqa: E501
],
[
[stop_sign],
[stop_sign, cherry_blossom],
])]
with global_force_attn_backend_context_manager(attn_backend):
if attn_backend == _Backend.FLASH_ATTN:
# Flash Attention works only with bfloat16 data-type
dtype = 'bfloat16'
_run_test(
hf_runner,
vllm_runner,
inputs,
model,
dtype=dtype,
max_tokens=max_tokens,
num_logprobs=num_logprobs,
tensor_parallel_size=1,
)
@create_new_process_for_each_test()
@multi_gpu_test(num_gpus=2)
@pytest.mark.parametrize("distributed_executor_backend", ["ray", "mp"])
@pytest.mark.parametrize("model", models)
@pytest.mark.parametrize("dtype", ["bfloat16"])
@pytest.mark.parametrize("max_tokens", [64])
@pytest.mark.parametrize("num_logprobs", [5])
@pytest.mark.skipif(
Version(TRANSFORMERS_VERSION) <= Version("4.55.2"),
reason="Transformers v4.55 has a regression issue on mllama, "
"see: https://github.com/huggingface/transformers/pull/40083")
def test_models_distributed(
hf_runner,
vllm_runner,
image_assets,
distributed_executor_backend,
model,
dtype,
max_tokens,
num_logprobs,
) -> None:
run_test(
hf_runner,
vllm_runner,
image_assets,
model=model,
size_factors=[0.25, 0.5, 1.0],
dtype=dtype,
max_tokens=max_tokens,
num_logprobs=num_logprobs,
tensor_parallel_size=2,
distributed_executor_backend=distributed_executor_backend,
)
@large_gpu_test(min_gb=48)
@pytest.mark.core_model
@pytest.mark.parametrize("model", models)
@pytest.mark.parametrize("dtype", ["float16"])
@pytest.mark.parametrize("max_tokens", [32])
@pytest.mark.skipif(not is_quant_method_supported("bitsandbytes"),
reason='bitsandbytes is not supported on this GPU type.')
def test_bnb_regression(
image_assets: ImageTestAssets,
model: str,
dtype: str,
max_tokens: int,
):
stop_sign = image_assets[0].pil_image
prompts = [
{
"prompt": "<|begin_of_text|>The content of the image <|image|> is",
"multi_modal_data": {
"image": stop_sign
},
},
{
"prompt":
"The color of the sky is blue but sometimes it can also be",
},
]
# Test regression about QKVCrossParallelLinear
llm = LLM(
model=model,
dtype=dtype,
max_model_len=8192,
max_num_seqs=2,
quantization="bitsandbytes",
)
sampling_params = SamplingParams(
temperature=0,
max_tokens=max_tokens,
)
outputs = llm.generate(prompts, sampling_params)
assert outputs
@large_gpu_test(min_gb=48)
@pytest.mark.core_model
@pytest.mark.parametrize("model", models)
@pytest.mark.parametrize("dtype", ["bfloat16"])
@pytest.mark.parametrize("max_tokens", [32])
def test_explicit_implicit_prompt(
image_assets: ImageTestAssets,
model: str,
dtype: str,
max_tokens: int,
):
stop_sign = image_assets[0].pil_image
# yapf: disable
prompts = [
# explicit prompt
{
"encoder_prompt": {
"prompt": "<|image|>",
"multi_modal_data": {"image": stop_sign},
},
"decoder_prompt": {
"prompt_token_ids": [128000, 791, 2262, 315, 279, 2217, 220, 128256, 374], # noqa: E501
}
},
{
"encoder_prompt": "Not <|image|>",
"decoder_prompt": "The color of the sky is blue but sometimes it can also be", # noqa: E501
},
# implicit prompt
{
"prompt": "<|begin_of_text|>The content of the image <|image|> is", # noqa: E501
"multi_modal_data": {"image": stop_sign},
},
{
"prompt": "The color of the sky is blue but sometimes it can also be", # noqa: E501
},
]
# yapf: enable
llm = LLM(
model=model,
dtype=dtype,
max_model_len=8192,
max_num_seqs=2,
tensor_parallel_size=1,
)
sampling_params = SamplingParams(
temperature=0,
max_tokens=max_tokens,
)
outputs = llm.generate(prompts, sampling_params)
n_prompts = len(prompts)
explicit_outputs = outputs[:n_prompts // 2]
implicit_outputs = outputs[n_prompts // 2:]
for exp_output, imp_output in zip(explicit_outputs, implicit_outputs):
assert exp_output.outputs[0].text == imp_output.outputs[0].text
@large_gpu_test(min_gb=48)
@pytest.mark.core_model
@pytest.mark.parametrize("model", models)
@pytest.mark.parametrize("dtype", ["bfloat16"])
@pytest.mark.parametrize("max_tokens", [128])
@pytest.mark.parametrize("num_logprobs", [5])
@pytest.mark.parametrize("attn_backend", LIST_ENC_DEC_SUPPORTED_BACKENDS)
def test_regression(vllm_runner, image_assets, model, dtype, max_tokens,
num_logprobs, attn_backend: _Backend) -> None:
stop_sign = image_assets[0].pil_image
with global_force_attn_backend_context_manager(attn_backend), vllm_runner(
model,
dtype=dtype,
max_model_len=8192,
max_num_seqs=4,
tensor_parallel_size=1,
limit_mm_per_prompt={"image":
_LIMIT_IMAGE_PER_PROMPT}) as vllm_model:
# Regression tests for https://github.com/vllm-project/vllm/issues/10648
# Number of groups of image tokens is greater than the number of images
# provided (the whitespace between the tags is necessary)
prompt = "<|begin_of_text|><|image|> <|image|> Compare the two images" # noqa: E501
image = stop_sign
with pytest.raises(ValueError):
vllm_model.generate_greedy_logprobs([prompt],
max_tokens,
num_logprobs,
images=[image])
# Batch of a text-only and image request that requires cross-attention
prompts = [
"What is the capital of spain?",
"Text before the image...<|image|>What is in the image?", # noqa: E501
]
images = [
None,
[stop_sign],
]
vllm_model.generate_greedy_logprobs(prompts,
max_tokens,
num_logprobs,
images=images)
# Test the reverse order too for good measure
prompts = [
"<|begin_of_text|>Text before the image...<|image|>What is in the image?", # noqa: E501
"<|begin_of_text|>Hello!",
]
images = [
[stop_sign],
None,
]
vllm_model.generate_greedy_logprobs(prompts,
max_tokens,
num_logprobs,
images=images)
# Mixed batch with text and images with different numbers of tiles
prompts = [
"<|begin_of_text|>Hello!",
"<|begin_of_text|>Some text before.<|image|>What is in the image?", # noqa: E501
"<|begin_of_text|>Some text before.<|image|>What is in the image?", # noqa: E501
]
images = [
None,
[stop_sign],
# smaller image must be 2nd for the repro
[stop_sign.resize((448, 448))],
]
vllm_model.generate_greedy_logprobs(prompts,
max_tokens,
num_logprobs,
images=images)
class DummyModel:
image_token_id = MLLAMA_IMAGE_TOKEN_ID
@pytest.mark.core_model
@pytest.mark.parametrize(
"input_indices_and_output",
# inputs, (cross_attention_mask, kv_range_for_decode)
[([TEXT_ONLY], (None, None)), ([IMAGE_AT_BEG], (None, None)),
([TEXT_ONLY, IMAGE_AT_BEG], (None, None)),
([IMAGE_AT_MIDDLE], ((10, 12), [[0, 6]])),
([TEXT_ONLY, IMAGE_AT_MIDDLE], ((14, 12), [[0, 6]])),
([TEXT_ONLY, IMAGE_AT_BEG, IMAGE_AT_MIDDLE],
((23, 24), [[0, 6], [6, 12]])),
([IMAGE_AT_MIDDLE, TEXT_ONLY], ((14, 12), [[0, 6]])),
([TWO_IMAGES], ((18, 12), [[6, 12]])),
([TEXT_ONLY, TWO_IMAGES], ((22, 12), [[6, 12]]))])
def test_get_cross_attention_mask(input_indices_and_output) -> None:
input_indices, expected_output = input_indices_and_output
sequences = [torch.tensor(prompt_data[i]) for i in input_indices]
num_tiles = [[2, 2] if i != TEXT_ONLY else [] for i in input_indices
if i != TEXT_ONLY]
input = torch.cat(sequences)
seq_lens = [len(s) for s in sequences]
attn_data = FlashAttentionMetadata(
seq_lens=seq_lens,
# Dummy values
enable_kv_scales_calculation=False,
num_prefills=0,
num_prefill_tokens=0,
num_decode_tokens=0,
slot_mapping=0,
multi_modal_placeholder_index_maps=None,
seq_lens_tensor=0,
max_prefill_seq_len=0,
max_decode_seq_len=0,
context_lens_tensor=None,
block_tables=None,
use_cuda_graph=False,
)
dummy = DummyModel()
cross_attention_mask, kv_range_for_decode = MllamaForConditionalGeneration\
.get_cross_attention_mask(dummy,
input,
attn_data,
num_tiles=num_tiles,
num_tokens_per_tile=3,
dtype=torch.bfloat16)
expected_cross_attention_mask, expected_kv_range_for_decode = \
expected_output
assert kv_range_for_decode == expected_kv_range_for_decode
if expected_cross_attention_mask is not None:
assert cross_attention_mask is not None
assert cross_attention_mask.shape == expected_cross_attention_mask
else:
assert cross_attention_mask is None
@pytest.mark.core_model
@pytest.mark.parametrize(
"input_indices",
[[TEXT_ONLY], [IMAGE_AT_BEG], [TEXT_ONLY, IMAGE_AT_BEG], [IMAGE_AT_MIDDLE],
[TEXT_ONLY, IMAGE_AT_MIDDLE], [TEXT_ONLY, IMAGE_AT_BEG, IMAGE_AT_MIDDLE],
[IMAGE_AT_MIDDLE, TEXT_ONLY], [TWO_IMAGES], [TEXT_ONLY, TWO_IMAGES]])
def test_get_full_text_row_masked_out_mask(input_indices) -> None:
sequences = [torch.tensor(prompt_data[i]) for i in input_indices]
seq_lens = [len(s) for s in sequences]
num_prefill_tokens = sum(seq_lens)
# TEXT_ONLY is zero, so it will be masked out,
# other instances should not be.
encoder_seq_lens = [int(i) for i in input_indices]
attn_data = FlashAttentionMetadata(
seq_lens=seq_lens,
encoder_seq_lens=encoder_seq_lens,
num_prefill_tokens=num_prefill_tokens,
# Dummy values
enable_kv_scales_calculation=False,
num_prefills=0,
num_decode_tokens=0,
slot_mapping=0,
multi_modal_placeholder_index_maps=None,
seq_lens_tensor=0,
max_prefill_seq_len=0,
max_decode_seq_len=0,
context_lens_tensor=None,
block_tables=None,
use_cuda_graph=False,
)
dummy = DummyModel()
full_text_row_masked_out_mask = MllamaForConditionalGeneration\
.get_full_text_row_masked_out_mask(dummy,
attn_data,
torch.get_default_device())
full_text_row_masked_out_mask = full_text_row_masked_out_mask.squeeze()
full_text_row_masked_out_mask = full_text_row_masked_out_mask.tolist()
idx = 0
assert len(full_text_row_masked_out_mask) == num_prefill_tokens
for i, seq_len in enumerate(seq_lens):
must_be_masked = input_indices[i] != TEXT_ONLY
for _ in range(seq_len):
assert full_text_row_masked_out_mask[idx] == must_be_masked, \
f"full_text_row_masked_out_mask[{idx}] must be " \
f"'{must_be_masked}' "
idx += 1
@pytest.mark.core_model
@pytest.mark.parametrize("encoder_seq_lens, num_tiles, expected", [
([6404], [[4]], [6404]),
([0, 6404], [[4]], [6404]),
([0, 1601, 8005], [[1], [4, 1]], [1601, 8005]),
([0, 19212, 0, 3202], [[4, 4, 4], [2]], [19212, 3202]),
])
def test_parse_and_validate_encoder_lens(encoder_seq_lens, num_tiles,
expected) -> None:
dummy = DummyModel()
num_tokens_per_tile = 1601
actual_encoder_seq_lens = MllamaForConditionalGeneration \
._get_and_validate_encoder_lens(
dummy,
encoder_seq_lens,
num_tiles,
num_tokens_per_tile,
)
assert actual_encoder_seq_lens == expected, \
f"Expected {expected} but got {actual_encoder_seq_lens}"
@@ -167,8 +167,6 @@ def _test_processing_correctness(
# incorrect token ids. So we need use `add_special_tokens=False` here
# to leave bos_token to be added by the processor.
_ADD_SPECIAL_TOKENS_OVERRIDES = {
"donut": False,
"mllama": False,
"ovis": False,
"ovis2_5": False,
"paligemma": False,
@@ -278,9 +276,7 @@ def _test_processing_correctness_one(
"facebook/chameleon-7b",
"CohereLabs/command-a-vision-07-2025",
"deepseek-ai/deepseek-vl2-tiny",
"naver-clova-ix/donut-base-finetuned-docvqa",
"baidu/ERNIE-4.5-VL-28B-A3B-PT",
"microsoft/Florence-2-base",
"adept/fuyu-8b",
"google/gemma-3-4b-it",
"google/gemma-3n-E2B-it",
@@ -305,7 +301,6 @@ def _test_processing_correctness_one(
"llava-hf/llava-v1.6-mistral-7b-hf",
"llava-hf/LLaVA-NeXT-Video-7B-hf",
"llava-hf/llava-onevision-qwen2-0.5b-ov-hf",
"meta-llama/Llama-3.2-11B-Vision-Instruct",
"TIGER-Lab/Mantis-8B-siglip-llama3",
"mispeech/midashenglm-7b",
"openbmb/MiniCPM-Llama3-V-2_5",
@@ -1,72 +0,0 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Tests for mllama's multimodal preprocessing and profiling."""
import pytest
from transformers import MllamaConfig
from vllm.multimodal import MULTIMODAL_REGISTRY
from vllm.multimodal.profiling import MultiModalProfiler
from ...utils import build_model_context
@pytest.mark.parametrize("model_id",
["meta-llama/Llama-3.2-11B-Vision-Instruct"])
@pytest.mark.parametrize("max_model_len", [4096, 8192, 25600, 131072])
@pytest.mark.parametrize("max_num_seqs", [1, 2, 8])
def test_profiling(
model_id: str,
max_model_len: int,
max_num_seqs: int,
):
# regression test for https://github.com/vllm-project/vllm/issues/13929
from vllm.model_executor.models.mllama import calc_token_per_chunk
model_config_kwargs = {
"max_model_len": max_model_len,
}
ctx = build_model_context(
model_id,
model_config_kwargs=model_config_kwargs,
limit_mm_per_prompt={"image": 1},
)
mm_config = ctx.get_mm_config()
processor = MULTIMODAL_REGISTRY.create_processor(ctx.model_config)
profiler = MultiModalProfiler(processor)
dummy_encoder_data = profiler.get_encoder_dummy_data(
max_model_len,
mm_counts=mm_config.limit_per_prompt,
)
dummy_mm_data = processor.dummy_inputs.get_dummy_processor_inputs(
max_model_len,
mm_counts=mm_config.limit_per_prompt,
)
hf_config = ctx.get_hf_config(MllamaConfig)
image_size = hf_config.vision_config.image_size
encoder_seq_lens = [len(dummy_encoder_data.prompt_token_ids)
] * max_num_seqs
mm_data = processor.apply(
prompt=dummy_mm_data.prompt,
mm_data=dummy_mm_data.mm_data,
hf_processor_mm_kwargs=dict(),
)["mm_kwargs"].get_data()
# Get the actual number of encoder tokens for each sample.
# Because attn_metadata.encoder_seq_lens only counts the last
# group of images for each sample, which is used to cheat the
# block manager to allocate blocks for those images only.
# See MllamaMultiModalProcessor for more details.
num_tiles = [[t] for t in mm_data.pop("num_tiles")]
num_tokens_per_tile = calc_token_per_chunk(image_size)
actual_encoder_seq_lens = [
sum(num_tile) * num_tokens_per_tile for num_tile in num_tiles
]
# simulate mllama image-present prefill.
for actual_len, last_group_len in zip(actual_encoder_seq_lens,
encoder_seq_lens):
assert actual_len >= last_group_len
@@ -31,7 +31,6 @@ from ...utils import dummy_hf_overrides
ARCH_TO_SKIP = {
"MolmoForCausalLM": "incompatible requirements",
"Florence2ForConditionalGeneration": "not supported in V1",
}
ARCH_NEEDS_EXTRAS = [
"InternVLChatModel",
+1 -15
View File
@@ -354,11 +354,6 @@ _TEXT_GENERATION_EXAMPLE_MODELS = {
"MiMoForCausalLM": _HfExamplesInfo("XiaomiMiMo/MiMo-7B-RL",
trust_remote_code=True),
"Dots1ForCausalLM": _HfExamplesInfo("rednote-hilab/dots.llm1.inst"),
# [Encoder-decoder]
"BartModel": _HfExamplesInfo("facebook/bart-base"),
"BartForConditionalGeneration": _HfExamplesInfo("facebook/bart-large-cnn"),
"MBartForConditionalGeneration": _HfExamplesInfo("facebook/mbart-large-en-ro", # noqa: E501
hf_overrides={"architectures": ["MBartForConditionalGeneration"]}), # noqa: E501
}
_EMBEDDING_EXAMPLE_MODELS = {
@@ -496,7 +491,7 @@ _MULTIMODAL_EXAMPLE_MODELS = {
trust_remote_code=True),
"Llama4ForConditionalGeneration": _HfExamplesInfo("meta-llama/Llama-4-Scout-17B-16E-Instruct", # noqa: E501
max_model_len=10240,
extras={"llama-guard-4": "meta-llama/Llama-Guard-4-12B"}, # noqa: E501
extras={"llama-guard-4": "meta-llama/Llama-Guard-4-12B"}, # noqa: E501
),
"LlavaForConditionalGeneration": _HfExamplesInfo("llava-hf/llava-1.5-7b-hf",
extras={"mistral": "mistral-community/pixtral-12b", # noqa: E501
@@ -583,15 +578,6 @@ _MULTIMODAL_EXAMPLE_MODELS = {
is_available_online=False,
),
# [Encoder-decoder]
"DonutForConditionalGeneration": _HfExamplesInfo("naver-clova-ix/donut-base-finetuned-docvqa", # noqa: E501
hf_overrides={"architectures": ["DonutForConditionalGeneration"], "model_type": "donut"}, # noqa: E501
extras={"dolphin": "ByteDance/Dolphin"}), # noqa: E501
# Florence-2 uses BartFastTokenizer which can't be loaded from AutoTokenizer
# Therefore, we borrow the BartTokenizer from the original Bart model
"Florence2ForConditionalGeneration": _HfExamplesInfo("microsoft/Florence-2-base", # noqa: E501
tokenizer="Isotr0py/Florence-2-tokenizer", # noqa: E501
trust_remote_code=True), # noqa: E501
"MllamaForConditionalGeneration": _HfExamplesInfo("meta-llama/Llama-3.2-11B-Vision-Instruct"), # noqa: E501
"WhisperForConditionalGeneration": _HfExamplesInfo("openai/whisper-large-v3"), # noqa: E501
# [Cross-encoder]
"JinaVLForRanking": _HfExamplesInfo("jinaai/jina-reranker-m0"), # noqa: E501
-4
View File
@@ -92,10 +92,6 @@ def can_initialize(model_arch: str, monkeypatch: pytest.MonkeyPatch,
# has cc==8.9 which hasn't supported FA3 yet. Remove this hack when
# L4 supports FA3.
m.setenv("VLLM_ATTENTION_BACKEND", "TRITON_ATTN_VLLM_V1")
if model_arch == "Florence2ForConditionalGeneration":
# An encoder-decoder model that's V0-only. Just skip it
# since V0 is about to be removed.
pytest.skip("Skipping Florence2ForConditionalGeneration")
if model_arch == "WhisperForConditionalGeneration":
m.setenv("VLLM_WORKER_MULTIPROC_METHOD", "spawn")
LLM(
-1
View File
@@ -50,7 +50,6 @@ def test_registry_imports(model_arch):
@create_new_process_for_each_test()
@pytest.mark.parametrize("model_arch,is_mm,init_cuda,is_ce", [
("LlamaForCausalLM", False, False, False),
("MllamaForConditionalGeneration", True, False, False),
("LlavaForConditionalGeneration", True, True, False),
("BertForSequenceClassification", False, False, True),
("RobertaForSequenceClassification", False, False, True),
+1 -2
View File
@@ -299,9 +299,8 @@ def test_rope_customization():
reason="Encoder Decoder models not supported on ROCm.")
@pytest.mark.parametrize(("model_id", "is_encoder_decoder"), [
("facebook/opt-125m", False),
("facebook/bart-base", True),
("openai/whisper-tiny", True),
("meta-llama/Llama-3.2-1B-Instruct", False),
("meta-llama/Llama-3.2-11B-Vision", True),
])
def test_is_encoder_decoder(model_id, is_encoder_decoder):
config = ModelConfig(model_id)
-28
View File
@@ -501,34 +501,6 @@ def test_bind_kv_cache_non_attention():
assert ctx['model.layers.28.attn'].kv_cache[0] is kv_cache[1]
def test_bind_kv_cache_encoder_decoder(monkeypatch: pytest.MonkeyPatch):
# V1 TESTS: ENCODER_DECODER is not supported on V1 yet.
with monkeypatch.context() as m:
m.setenv("VLLM_USE_V1", "0")
from vllm.attention import Attention, AttentionType
# example from bart
ctx = {
'encoder.layers.0.self_attn.attn':
Attention(32, 128, 0.1, attn_type=AttentionType.ENCODER),
'decoder.layers.0.encoder_attn.attn':
Attention(32, 128, 0.1, attn_type=AttentionType.ENCODER_DECODER),
'decoder.layers.0.self_attn.attn':
Attention(32, 128, 0.1, attn_type=AttentionType.DECODER),
}
kv_cache = [
torch.zeros((1, )),
]
encoder_kv_cache = ctx['encoder.layers.0.self_attn.attn'].kv_cache
bind_kv_cache(ctx, [kv_cache])
assert ctx['encoder.layers.0.self_attn.attn'].kv_cache is encoder_kv_cache
assert ctx['decoder.layers.0.encoder_attn.attn'].kv_cache[0] is kv_cache[0]
assert ctx['decoder.layers.0.self_attn.attn'].kv_cache[0] is kv_cache[0]
def test_bind_kv_cache_pp():
with patch("vllm.utils.cuda_device_count_stateless", lambda: 2):
# this test runs with 1 GPU, but we simulate 2 GPUs
-21
View File
@@ -9,24 +9,9 @@ from vllm import LLM
from vllm.engine.arg_utils import AsyncEngineArgs
from vllm.engine.async_llm_engine import AsyncLLMEngine
UNSUPPORTED_MODELS_V1 = [
"facebook/bart-large-cnn", # encoder decoder
]
MODEL = "meta-llama/Llama-3.2-1B-Instruct"
@pytest.mark.parametrize("model", UNSUPPORTED_MODELS_V1)
def test_reject_unsupported_models(monkeypatch, model):
with monkeypatch.context() as m:
m.setenv("VLLM_USE_V1", "1")
args = AsyncEngineArgs(model=model)
with pytest.raises(NotImplementedError):
_ = args.create_engine_config()
m.delenv("VLLM_USE_V1")
def test_reject_bad_config(monkeypatch):
with monkeypatch.context() as m:
m.setenv("VLLM_USE_V1", "0")
@@ -77,12 +62,6 @@ def test_enable_by_default_fallback(monkeypatch):
assert envs.VLLM_USE_V1
m.delenv("VLLM_USE_V1")
# Should fall back to V0 for supported model.
_ = AsyncEngineArgs(
model=UNSUPPORTED_MODELS_V1[0]).create_engine_config()
assert not envs.VLLM_USE_V1
m.delenv("VLLM_USE_V1")
def test_v1_llm_by_default(monkeypatch):
with monkeypatch.context() as m:
@@ -1,648 +0,0 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import itertools
import pytest
import torch
from vllm.engine.arg_utils import EngineArgs
from vllm.platforms import current_platform
from vllm.sequence import SamplingParams, SequenceData, SequenceGroupMetadata
from vllm.utils import make_tensor_with_pad
from vllm.worker.enc_dec_model_runner import EncoderDecoderModelRunner
BATCH_SIZES = [1, 4, 16, 64, 256]
def _create_model_runner(model: str, *args,
**kwargs) -> EncoderDecoderModelRunner:
engine_args = EngineArgs(model, *args, **kwargs)
engine_config = engine_args.create_engine_config()
model_runner = EncoderDecoderModelRunner(
vllm_config=engine_config,
is_driver_worker=True,
)
return model_runner
@pytest.mark.skipif(condition=current_platform.is_cpu(),
reason="CPU backend is currently "
"unsupported for encoder/ "
"decoder models")
def test_empty_seq_group():
"""Verify prepare prompt and decode returns empty output
for empty seq group list"""
model_runner = _create_model_runner(
"facebook/bart-base",
seed=0,
dtype="float16",
max_num_batched_tokens=100000,
max_num_seqs=100000,
enable_chunked_prefill=False,
enforce_eager=True,
)
seq_group_metadata_list: list[SequenceGroupMetadata] = []
model_input = model_runner._prepare_model_input_tensors(
seq_group_metadata_list)
(
input_tokens,
input_positions,
encoder_input_tokens,
encoder_input_positions,
attn_metadata,
return_seq_lens,
) = (
model_input.input_tokens,
model_input.input_positions,
model_input.encoder_input_tokens,
model_input.encoder_input_positions,
model_input.attn_metadata,
model_input.seq_lens,
)
assert input_tokens is None
assert input_positions is None
assert encoder_input_tokens is None
assert encoder_input_positions is None
assert attn_metadata is None
assert return_seq_lens is None
@pytest.mark.skipif(condition=current_platform.is_cpu(),
reason="CPU backend is currently "
"unsupported for encoder/ "
"decoder models")
@pytest.mark.parametrize("batch_size", BATCH_SIZES)
def test_prepare_prompt(batch_size):
'''
Test the ability of the encoder/decoder model runner subclass to
produce prefill-phase model inputs & attention metadata.
Test behavior:
* Instantiate BART base model & enc/dec model runner
* Construct sequence-group metadata for dummy prompts
* Test that encoder attention, decoder self-attention,
and encoder/decoder cross-attention inputs are correct
Arguments:
* batch_size
* backend_name: The attention backend under test
* enforce_eager: Enforce eager mode if True (i.e. no CUDAGraph)
'''
model_runner = _create_model_runner(
"facebook/bart-base",
seed=0,
dtype="float16",
max_num_batched_tokens=100000,
max_num_seqs=100000,
enable_chunked_prefill=False,
enforce_eager=True,
)
seq_lens: list[int] = []
encoder_seq_lens: list[int] = []
seq_group_metadata_list: list[SequenceGroupMetadata] = []
block_tables = {0: [1]}
cross_block_table = [2]
for i in range(batch_size):
# make sure all tokens fit into one block
seq_len = i % (model_runner.block_size - 1) + 1
seq_lens.append(seq_len)
seq_data = SequenceData.from_seqs(range(seq_len))
encoder_seq_len = (i + 1) % (model_runner.block_size - 1) + 1
encoder_seq_lens.append(encoder_seq_len)
encoder_seq_data = SequenceData.from_seqs(range(encoder_seq_len))
seq_group_metadata = SequenceGroupMetadata(
request_id=f"test_{i}",
is_prompt=True,
seq_data={0: seq_data},
sampling_params=SamplingParams(temperature=0),
block_tables=block_tables,
encoder_seq_data=encoder_seq_data,
cross_block_table=cross_block_table,
)
assert seq_group_metadata.token_chunk_size == seq_data.get_len()
seq_group_metadata_list.append(seq_group_metadata)
# Build
# * Decoder model inputs
# * Decoder self-attention KV caching data structures
# * Encoder model inputs
# * Encoder/decoder cross-attention KV caching data structures
model_input = model_runner.prepare_model_input(seq_group_metadata_list)
input_tokens = model_input.input_tokens
input_positions = model_input.input_positions
attn_metadata = model_input.attn_metadata
return_seq_lens = model_input.seq_lens
slot_mapping = attn_metadata.slot_mapping
encoder_input_tokens = model_input.encoder_input_tokens
encoder_input_positions = model_input.encoder_input_positions
cross_slot_mapping = attn_metadata.cross_slot_mapping
assert return_seq_lens == seq_lens
assert len(slot_mapping) == len(input_tokens)
assert len(cross_slot_mapping) == len(encoder_input_tokens)
# Verify input metadata is correct for prompts.
# - Decoder attention metadata
device = model_runner.device
assert attn_metadata.num_prefills > 0
assert attn_metadata.num_decode_tokens == 0
assert torch.equal(attn_metadata.seq_lens_tensor,
torch.tensor(seq_lens, device=device, dtype=torch.int))
assert attn_metadata.seq_lens == seq_lens
assert attn_metadata.max_prefill_seq_len == max(seq_lens)
assert attn_metadata.max_decode_seq_len == 0
# - Encoder attention metadata
assert attn_metadata.encoder_seq_lens == encoder_seq_lens
assert torch.equal(
attn_metadata.encoder_seq_lens_tensor,
torch.tensor(encoder_seq_lens, device=device, dtype=torch.int))
assert attn_metadata.max_encoder_seq_len == max(encoder_seq_lens)
assert attn_metadata.num_encoder_tokens == sum(encoder_seq_lens)
# Test decoder subquery start locs.
start_idx = 0
start_loc = [start_idx]
for seq_len in seq_lens:
start_idx += seq_len
start_loc.append(start_idx)
assert torch.equal(
attn_metadata.query_start_loc,
torch.tensor(start_loc, dtype=torch.int32, device=device),
)
# Test decoder seq start locs & context lengths
assert torch.equal(
attn_metadata.seq_start_loc,
torch.tensor(start_loc, dtype=torch.int32, device=device),
)
assert torch.equal(
attn_metadata.context_lens_tensor,
torch.zeros(attn_metadata.context_lens_tensor.shape[0],
dtype=torch.int,
device=device),
)
# Verify block tables are correct for prompts
# - Decoder self-attention
expected = torch.tensor(
[[] for _ in range(len(seq_group_metadata_list))],
dtype=torch.int32,
device=model_runner.device,
)
assert torch.equal(
attn_metadata.block_tables,
expected,
)
# - Encoder/decoder cross-attention
assert torch.equal(
attn_metadata.cross_block_tables,
expected,
)
# Cuda graph should not be used for prefill.
assert attn_metadata.use_cuda_graph is False
# Verify the lengths of input tokens & positions
# - Decoder
assert len(input_tokens) == sum(seq_lens)
assert len(input_positions) == sum(seq_lens)
# -- An indirect check that model_input.input_tokens
# and model_input.input_positions are correct -
# by design of the test, the input tokens are
# equal to the input position values, so if
# the model_input data structure has the correct
# values then these two should be equal
assert torch.equal(
input_tokens,
input_positions,
)
# - Encoder
assert len(encoder_input_tokens) == sum(encoder_seq_lens)
# -- An indirect check that model_input.encoder_input_tokens
# and model_input.encoder_input_positions are correct -
# by design of the test, the input tokens are
# equal to the input position values, so if
# the model_input data structure has the correct
# values then these two should be equal
assert torch.equal(
encoder_input_tokens,
encoder_input_positions,
)
# Test that vLLM sampling infrastructure chooses the correct
# sequence positions at which to sample (i.e. the end of
# each sequence) in the prefill phase
expected_selected_token_indices = []
selected_token_start_idx = 0
for seq_len in seq_lens:
# Compute the index offset of the final token in each
# prompt (recall that the prompts are concatenated)
expected_selected_token_indices.append(selected_token_start_idx +
seq_len - 1)
selected_token_start_idx += seq_len
sampling_metadata = model_input.sampling_metadata
actual = sampling_metadata.selected_token_indices
expected = torch.tensor(
expected_selected_token_indices,
device=actual.device,
dtype=actual.dtype,
)
assert torch.equal(actual, expected)
@pytest.mark.skipif(condition=current_platform.is_cpu(),
reason="CPU backend is currently "
"unsupported for encoder/ "
"decoder models")
@pytest.mark.parametrize("batch_size", BATCH_SIZES)
@pytest.mark.parametrize("multiple_seqs_per_seq_group", [True, False])
def test_prepare_decode(batch_size, multiple_seqs_per_seq_group):
'''
Test the ability of the encoder/decoder model runner subclass to
produce decode-phase model inputs & attention metadata.
Test behavior:
* Instantiate BART base model & enc/dec model runner
* Construct sequence-group metadata for dummy prompts
* Test that encoder attention, decoder self-attention,
and encoder/decoder cross-attention inputs are correct
Arguments:
* batch_size
* multiple_seqs_per_seq_group
* backend_name: The attention backend under test
* enforce_eager: Enforce eager mode if True (i.e. no CUDAGraph)
'''
model_runner = _create_model_runner(
"facebook/bart-base",
seed=0,
dtype="float16",
max_num_batched_tokens=100000,
max_num_seqs=100000,
enable_chunked_prefill=False,
enforce_eager=True,
)
seq_lens: list[int] = []
encoder_seq_lens: list[int] = []
seq_group_metadata_list: list[SequenceGroupMetadata] = []
block_tables = {
0: [1],
1: [3]
} if multiple_seqs_per_seq_group else {
0: [1]
}
cross_block_table = [2]
for i in range(batch_size):
# make sure all tokens fit into one block
seq_len = i % (model_runner.block_size - 1) + 1
seq_data = SequenceData.from_seqs(range(seq_len))
encoder_seq_len = (i + 1) % (model_runner.block_size - 1) + 1
encoder_seq_data = SequenceData.from_seqs(range(encoder_seq_len))
seq_group_metadata = SequenceGroupMetadata(
request_id=f"test_{i}",
is_prompt=False,
seq_data={
0: seq_data,
1: seq_data
} if multiple_seqs_per_seq_group else {0: seq_data},
sampling_params=SamplingParams(temperature=0),
block_tables=block_tables,
encoder_seq_data=encoder_seq_data,
cross_block_table=cross_block_table,
)
assert seq_group_metadata.token_chunk_size == 1
seq_group_metadata_list.append(seq_group_metadata)
seq_lens.extend(
[seq_len for _ in range(len(seq_group_metadata.seq_data))])
encoder_seq_lens.extend(
[encoder_seq_len for _ in range(len(seq_group_metadata.seq_data))])
# Build
# * Decoder model inputs
# * Decoder self-attention KV caching data structures
# * Encoder model inputs
# * Encoder/decoder cross-attention KV caching data structures
model_input = model_runner.prepare_model_input(seq_group_metadata_list)
input_tokens = model_input.input_tokens
input_positions = model_input.input_positions
attn_metadata = model_input.attn_metadata
return_seq_lens = model_input.seq_lens
slot_mapping = attn_metadata.slot_mapping
encoder_input_tokens = model_input.encoder_input_tokens
encoder_input_positions = model_input.encoder_input_positions
cross_slot_mapping = attn_metadata.cross_slot_mapping
assert return_seq_lens == seq_lens
assert len(slot_mapping) == len(input_tokens)
assert len(cross_slot_mapping) == len(encoder_input_tokens)
# Verify input metadata is correct for decode phase.
# - Decoder attention metadata
device = model_runner.device
assert attn_metadata.num_prefills == 0
assert attn_metadata.num_decode_tokens > 0
assert torch.equal(attn_metadata.seq_lens_tensor,
torch.tensor(seq_lens, device=device, dtype=torch.int))
assert attn_metadata.seq_lens == seq_lens
assert attn_metadata.max_prefill_seq_len == 0
assert attn_metadata.max_decode_seq_len == max(seq_lens)
# - Encoder attention metadata
assert attn_metadata.encoder_seq_lens == encoder_seq_lens
assert torch.equal(
attn_metadata.encoder_seq_lens_tensor,
torch.tensor(encoder_seq_lens, device=device, dtype=torch.int))
assert attn_metadata.max_encoder_seq_len == max(encoder_seq_lens)
assert attn_metadata.num_encoder_tokens == sum(encoder_seq_lens)
# Test decoder subquery start locs.
start_idx = 0
start_loc = [start_idx]
for seq_len in seq_lens:
start_idx += 1
start_loc.append(start_idx)
assert torch.equal(
attn_metadata.query_start_loc,
torch.tensor(start_loc, dtype=torch.int32, device=device),
)
# Test decoder seq start locs. Note that for normal prefill it is
# equivalent to query_start_loc.
start_idx = 0
seq_start_loc = [start_idx]
for seq_len in seq_lens:
start_idx += seq_len
seq_start_loc.append(start_idx)
# Test seq_start_loc and context lengths
assert torch.equal(
attn_metadata.seq_start_loc,
torch.tensor(seq_start_loc, dtype=torch.int32, device=device),
)
assert torch.equal(
attn_metadata.context_lens_tensor,
torch.tensor([seq_len - 1 for seq_len in seq_lens],
dtype=torch.int,
device=device))
# Verify block tables are correct for prompts
# - Decoder self-attention
flattened_block_tables = [
block_table for block_table in block_tables.values()
]
expected = torch.tensor(flattened_block_tables *
len(seq_group_metadata_list),
dtype=torch.int32,
device=model_runner.device)
assert torch.equal(
attn_metadata.block_tables,
expected,
)
# - Encoder/decoder cross-attention
expected = torch.tensor([
cross_block_table for seq_group_metadata in seq_group_metadata_list
for _ in range(len(seq_group_metadata.seq_data))
],
dtype=torch.int32,
device=model_runner.device)
assert torch.equal(
attn_metadata.cross_block_tables,
expected,
)
# Model runner's CUDAGraph setting should be propagated to attention
# metadata.
assert attn_metadata.use_cuda_graph is False
# Verify the lengths of input tokens & positions
# - Decoder
assert len(input_tokens) == len(seq_lens)
assert len(input_positions) == len(seq_lens)
# -- An indirect check that model_input.input_tokens
# and model_input.input_positions are correct -
# by design of the test, the input tokens are
# equal to the input position values, so if
# the model_input data structure has the correct
# values then these two should be equal
assert torch.equal(
input_tokens,
input_positions,
)
# - Encoder
assert len(encoder_input_tokens) == 0
assert len(encoder_input_tokens) == 0
# -- An indirect check that model_input.encoder_input_tokens
# and model_input.encoder_input_positions are correct -
# by design of the test, the input tokens are
# equal to the input position values, so if
# the model_input data structure has the correct
# values then these two should be equal
assert torch.equal(
encoder_input_tokens,
encoder_input_positions,
)
# Test that vLLM sampling infrastructure chooses the correct
# sequence positions at which to sample (i.e. the end of
# each sequence) in the decode phase
expected_selected_token_indices = []
for selected_token_start_idx, seq_len in enumerate(seq_lens):
# Compute the index offset of the final token in each
# sequence's decoded outputs; since a single token is
# decoded per iteration per sequence, then the length
# of the decoded tokens for a given sequence is 1 and
# the final index offset into a given sequence's
# generated tokens is 0 (i.e. the expected sampling index
# for a given sequence is just `selected_token_start_idx`)
expected_selected_token_indices.append(selected_token_start_idx)
sampling_metadata = model_input.sampling_metadata
actual = sampling_metadata.selected_token_indices
expected = torch.tensor(
expected_selected_token_indices,
device=actual.device,
dtype=actual.dtype,
)
assert torch.equal(actual, expected)
@pytest.mark.parametrize("batch_size", list(range(1, 257)))
@pytest.mark.parametrize("multiple_seqs_per_seq_group", [True, False])
def test_prepare_decode_cuda_graph(batch_size, multiple_seqs_per_seq_group):
"""
Tests that for encoder-decoder models with CUDA Graph capture and replay
enabled, the tensors used during the decode phase are correctly padded
for varying input batch sizes.
"""
model_runner = _create_model_runner(
"facebook/bart-base",
seed=0,
dtype="float16",
max_num_batched_tokens=100000,
max_num_seqs=100000,
enable_chunked_prefill=False,
enforce_eager=False,
)
block_tables = {
0: [1],
1: [3]
} if multiple_seqs_per_seq_group else {
0: [1]
}
seq_lens: list[int] = []
encoder_seq_lens: list[int] = []
seq_group_metadata_list: list[SequenceGroupMetadata] = []
cross_block_table = [2]
expanded_batch_size = 0
for i in range(batch_size):
# make sure all tokens fit into one block
seq_len = i % (model_runner.block_size - 1) + 1
seq_data = SequenceData.from_seqs(range(seq_len))
encoder_seq_len = (i + 1) % (model_runner.block_size - 1) + 1
encoder_seq_data = SequenceData.from_seqs(range(encoder_seq_len))
seq_group_metadata = SequenceGroupMetadata(
request_id=f"test_{i}",
is_prompt=False,
seq_data={
0: seq_data,
1: seq_data
} if multiple_seqs_per_seq_group else {0: seq_data},
sampling_params=SamplingParams(temperature=0),
block_tables=block_tables,
encoder_seq_data=encoder_seq_data,
cross_block_table=cross_block_table,
)
assert seq_group_metadata.token_chunk_size == 1
seq_lens.extend(
[seq_len for _ in range(len(seq_group_metadata.seq_data))])
encoder_seq_lens.extend(
[encoder_seq_len for _ in range(len(seq_group_metadata.seq_data))])
expanded_batch_size = expanded_batch_size + len(
seq_group_metadata.seq_data)
seq_group_metadata_list.append(seq_group_metadata)
model_input = model_runner.prepare_model_input(seq_group_metadata_list)
input_tokens = model_input.input_tokens
input_positions = model_input.input_positions
attn_metadata = model_input.attn_metadata
return_seq_lens = model_input.seq_lens
slot_mapping = attn_metadata.slot_mapping
encoder_input_tokens = model_input.encoder_input_tokens
encoder_input_positions = model_input.encoder_input_positions
cross_slot_mapping = attn_metadata.cross_slot_mapping
# With CUDA Graph capture and replay enabled, the decoder and encoder
# input sequences will be padded. Create the expected padded tensors
# accordingly.
graph_batch_size = model_runner.vllm_config.pad_for_cudagraph(
expanded_batch_size)
cuda_graph_pad_size = graph_batch_size - expanded_batch_size
padded_seq_lens = seq_lens + list(itertools.repeat(1, cuda_graph_pad_size))
padded_encoder_seq_lens = encoder_seq_lens + list(
itertools.repeat(1, cuda_graph_pad_size))
assert return_seq_lens == padded_seq_lens
assert len(slot_mapping) == len(input_tokens)
assert len(cross_slot_mapping) == len(encoder_input_tokens)
# Verify attention metadata
device = model_runner.device
assert attn_metadata.num_prefills == 0
assert attn_metadata.num_decode_tokens > 0
assert torch.equal(
attn_metadata.seq_lens_tensor,
torch.tensor(padded_seq_lens, device=device, dtype=torch.int))
assert attn_metadata.seq_lens == padded_seq_lens
assert attn_metadata.max_prefill_seq_len == 0
assert attn_metadata.max_decode_seq_len == max(seq_lens)
# - Encoder attention metadata
assert attn_metadata.encoder_seq_lens == padded_encoder_seq_lens
assert torch.equal(
attn_metadata.encoder_seq_lens_tensor,
torch.tensor(padded_encoder_seq_lens, device=device, dtype=torch.int))
assert attn_metadata.max_encoder_seq_len == max(padded_encoder_seq_lens)
assert attn_metadata.num_encoder_tokens == sum(padded_encoder_seq_lens)
# Verify block tables are correct for prompts
# - Decoder self-attention. Pad the block tables as expected.
flattened_block_tables = [
block_table for _ in range(len(seq_group_metadata_list))
for block_table in block_tables.values()
]
flattened_block_tables.extend([[] for _ in range(cuda_graph_pad_size)])
expected = make_tensor_with_pad(
flattened_block_tables,
max_len=64,
pad=0,
dtype=torch.int32,
device=model_runner.device,
)
assert torch.equal(
attn_metadata.block_tables,
expected,
)
# - Encoder/decoder cross-attention. Pad the cross-attention block tables
# as expected.
expected = [
cross_block_table for seq_group_metadata in seq_group_metadata_list
for _ in range(len(seq_group_metadata.seq_data))
]
expected.extend([[] for _ in range(cuda_graph_pad_size)])
expected = make_tensor_with_pad(
expected,
max_len=64,
pad=0,
dtype=torch.int32,
device=model_runner.device,
)
assert torch.equal(
attn_metadata.cross_block_tables,
expected,
)
# Model runner's CUDAGraph setting should be propagated to attention
# metadata.
assert attn_metadata.use_cuda_graph is True
# Verify the lengths of input tokens & positions
# - Decoder
assert len(input_tokens) == len(padded_seq_lens)
assert len(input_positions) == len(padded_seq_lens)
# -- An indirect check that model_input.input_tokens
# and model_input.input_positions are correct -
# by design of the test, the input tokens are
# equal to the input position values, so if
# the model_input data structure has the correct
# values then these two should be equal
assert torch.equal(
input_tokens,
input_positions,
)
# - Encoder
assert len(encoder_input_tokens) == 0
assert len(encoder_input_tokens) == 0
# -- An indirect check that model_input.encoder_input_tokens
# and model_input.encoder_input_positions are correct -
# by design of the test, the input tokens are
# equal to the input position values, so if
# the model_input data structure has the correct
# values then these two should be equal
assert torch.equal(
encoder_input_tokens,
encoder_input_positions,
)