Skip to content

GitLab

"requirements/rocm.txt" did not exist on "c188ecb080501c5ccb34bbd6542978284c547122"
Unverified Commit 4e256cad authored by Harry Mellor's avatar Harry Mellor Committed by GitHub
Browse files

Remove all references to `yapf` as it's no longer used (#26251) 


Signed-off-by: default avatarHarry Mellor <19981378+hmellor@users.noreply.github.com>
parent d6953beb
Hide whitespace changes
Inline Side-by-side
Showing with 613 additions and 626 deletions
csrc/quantization/machete/generate.py
View file @ 4e256cad
......@@ -12,9 +12,6 @@ from functools import reduce
from typing import Optional, Union
import jinja2
# yapf conflicts with isort for this block
# yapf: disable
from vllm_cutlass_library_extension import (
DataType,
EpilogueScheduleTag,
......@@ -31,8 +28,6 @@ from vllm_cutlass_library_extension import (
VLLMKernelScheduleTag,
)
# yapf: enable
#
# Generator templating
#
......
examples/others/tensorize_vllm_model.py
View file @ 4e256cad
......@@ -21,8 +21,6 @@ from vllm.utils import FlexibleArgumentParser
logger = logging.getLogger()
# yapf conflicts with isort for this docstring
# yapf: disable
"""
tensorize_vllm_model.py is a script that can be used to serialize and
deserialize vLLM models. These models can be loaded using tensorizer
......@@ -132,7 +130,8 @@ def get_parser():
"can be loaded using tensorizer directly to the GPU "
"extremely quickly. Tensor encryption and decryption is "
"also supported, although libsodium must be installed to "
"use it.")
"use it."
)
parser = EngineArgs.add_cli_args(parser)
parser.add_argument(
......@@ -144,13 +143,14 @@ def get_parser():
"along with the model by instantiating a TensorizerConfig object, "
"creating a dict from it with TensorizerConfig.to_serializable(), "
"and passing it to LoRARequest's initializer with the kwarg "
"tensorizer_config_dict."
"tensorizer_config_dict.",
)
subparsers = parser.add_subparsers(dest='command', required=True)
subparsers = parser.add_subparsers(dest="command", required=True)
serialize_parser = subparsers.add_parser(
'serialize', help="Serialize a model to `--serialized-directory`")
"serialize", help="Serialize a model to `--serialized-directory`"
)
serialize_parser.add_argument(
"--suffix",
......@@ -163,7 +163,9 @@ def get_parser():
"`--suffix` is `v1`, the serialized model tensors will be "
"saved to "
"`s3://my-bucket/vllm/EleutherAI/gpt-j-6B/v1/model.tensors`. "
"If none is provided, a random UUID will be used."))
"If none is provided, a random UUID will be used."
),
)
serialize_parser.add_argument(
"--serialized-directory",
type=str,
......@@ -175,108 +177,127 @@ def get_parser():
"and the model HuggingFace ID is `EleutherAI/gpt-j-6B`, tensors will "
"be saved to `dir/vllm/EleutherAI/gpt-j-6B/suffix/model.tensors`, "
"where `suffix` is given by `--suffix` or a random UUID if not "
"provided.")
"provided.",
)
serialize_parser.add_argument(
"--serialization-kwargs",
type=tensorizer_kwargs_arg,
required=False,
help=("A JSON string containing additional keyword arguments to "
"pass to Tensorizer's TensorSerializer during "
"serialization."))
help=(
"A JSON string containing additional keyword arguments to "
"pass to Tensorizer's TensorSerializer during "
"serialization."
),
)
serialize_parser.add_argument(
"--keyfile",
type=str,
required=False,
help=("Encrypt the model weights with a randomly-generated binary key,"
" and save the key at this path"))
help=(
"Encrypt the model weights with a randomly-generated binary key,"
" and save the key at this path"
),
)
deserialize_parser = subparsers.add_parser(
'deserialize',
help=("Deserialize a model from `--path-to-tensors`"
" to verify it can be loaded and used."))
"deserialize",
help=(
"Deserialize a model from `--path-to-tensors`"
" to verify it can be loaded and used."
),
)
deserialize_parser.add_argument(
"--path-to-tensors",
type=str,
required=False,
help="The local path or S3 URI to the model tensors to deserialize. ")
help="The local path or S3 URI to the model tensors to deserialize. ",
)
deserialize_parser.add_argument(
"--serialized-directory",
type=str,
required=False,
help="Directory with model artifacts for loading. Assumes a "
"model.tensors file exists therein. Can supersede "
"--path-to-tensors.")
"model.tensors file exists therein. Can supersede "
"--path-to-tensors.",
)
deserialize_parser.add_argument(
"--keyfile",
type=str,
required=False,
help=("Path to a binary key to use to decrypt the model weights,"
" if the model was serialized with encryption"))
help=(
"Path to a binary key to use to decrypt the model weights,"
" if the model was serialized with encryption"
),
)
deserialize_parser.add_argument(
"--deserialization-kwargs",
type=tensorizer_kwargs_arg,
required=False,
help=("A JSON string containing additional keyword arguments to "
"pass to Tensorizer's `TensorDeserializer` during "
"deserialization."))
help=(
"A JSON string containing additional keyword arguments to "
"pass to Tensorizer's `TensorDeserializer` during "
"deserialization."
),
)
TensorizerArgs.add_cli_args(deserialize_parser)
return parser
def merge_extra_config_with_tensorizer_config(extra_cfg: dict,
cfg: TensorizerConfig):
def merge_extra_config_with_tensorizer_config(extra_cfg: dict, cfg: TensorizerConfig):
for k, v in extra_cfg.items():
if hasattr(cfg, k):
setattr(cfg, k, v)
logger.info(
"Updating TensorizerConfig with %s from "
"--model-loader-extra-config provided", k
"--model-loader-extra-config provided",
k,
)
def deserialize(args, tensorizer_config):
if args.lora_path:
tensorizer_config.lora_dir = tensorizer_config.tensorizer_dir
llm = LLM(model=args.model,
load_format="tensorizer",
tensor_parallel_size=args.tensor_parallel_size,
model_loader_extra_config=tensorizer_config,
enable_lora=True,
llm = LLM(
model=args.model,
load_format="tensorizer",
tensor_parallel_size=args.tensor_parallel_size,
model_loader_extra_config=tensorizer_config,
enable_lora=True,
)
sampling_params = SamplingParams(
temperature=0,
max_tokens=256,
stop=["[/assistant]"]
temperature=0, max_tokens=256, stop=["[/assistant]"]
)
# Truncating this as the extra text isn't necessary
prompts = [
"[user] Write a SQL query to answer the question based on ..."
]
prompts = ["[user] Write a SQL query to answer the question based on ..."]
# Test LoRA load
print(
llm.generate(
prompts,
sampling_params,
lora_request=LoRARequest("sql-lora",
1,
args.lora_path,
tensorizer_config_dict = tensorizer_config
.to_serializable())
prompts,
sampling_params,
lora_request=LoRARequest(
"sql-lora",
1,
args.lora_path,
tensorizer_config_dict=tensorizer_config.to_serializable(),
),
)
)
else:
llm = LLM(model=args.model,
load_format="tensorizer",
tensor_parallel_size=args.tensor_parallel_size,
model_loader_extra_config=tensorizer_config
llm = LLM(
model=args.model,
load_format="tensorizer",
tensor_parallel_size=args.tensor_parallel_size,
model_loader_extra_config=tensorizer_config,
)
return llm
......@@ -285,17 +306,20 @@ def main():
parser = get_parser()
args = parser.parse_args()
s3_access_key_id = (getattr(args, 's3_access_key_id', None)
or os.environ.get("S3_ACCESS_KEY_ID", None))
s3_secret_access_key = (getattr(args, 's3_secret_access_key', None)
or os.environ.get("S3_SECRET_ACCESS_KEY", None))
s3_endpoint = (getattr(args, 's3_endpoint', None)
or os.environ.get("S3_ENDPOINT_URL", None))
s3_access_key_id = getattr(args, "s3_access_key_id", None) or os.environ.get(
"S3_ACCESS_KEY_ID", None
)
s3_secret_access_key = getattr(
args, "s3_secret_access_key", None
) or os.environ.get("S3_SECRET_ACCESS_KEY", None)
s3_endpoint = getattr(args, "s3_endpoint", None) or os.environ.get(
"S3_ENDPOINT_URL", None
)
credentials = {
"s3_access_key_id": s3_access_key_id,
"s3_secret_access_key": s3_secret_access_key,
"s3_endpoint": s3_endpoint
"s3_endpoint": s3_endpoint,
}
model_ref = args.model
......@@ -309,25 +333,25 @@ def main():
if args.model_loader_extra_config:
extra_config = json.loads(args.model_loader_extra_config)
tensorizer_dir = (args.serialized_directory or
extra_config.get("tensorizer_dir"))
tensorizer_uri = (getattr(args, "path_to_tensors", None)
or extra_config.get("tensorizer_uri"))
tensorizer_dir = args.serialized_directory or extra_config.get("tensorizer_dir")
tensorizer_uri = getattr(args, "path_to_tensors", None) or extra_config.get(
"tensorizer_uri"
)
if tensorizer_dir and tensorizer_uri:
parser.error("--serialized-directory and --path-to-tensors "
"cannot both be provided")
parser.error(
"--serialized-directory and --path-to-tensors cannot both be provided"
)
if not tensorizer_dir and not tensorizer_uri:
parser.error("Either --serialized-directory or --path-to-tensors "
"must be provided")
parser.error(
"Either --serialized-directory or --path-to-tensors must be provided"
)
if args.command == "serialize":
engine_args = EngineArgs.from_cli_args(args)
input_dir = tensorizer_dir.rstrip('/')
input_dir = tensorizer_dir.rstrip("/")
suffix = args.suffix if args.suffix else uuid.uuid4().hex
base_path = f"{input_dir}/vllm/{model_ref}/{suffix}"
if engine_args.tensor_parallel_size > 1:
......@@ -339,15 +363,14 @@ def main():
tensorizer_uri=model_path,
encryption_keyfile=keyfile,
serialization_kwargs=args.serialization_kwargs or {},
**credentials
**credentials,
)
if args.lora_path:
tensorizer_config.lora_dir = tensorizer_config.tensorizer_dir
tensorize_lora_adapter(args.lora_path, tensorizer_config)
merge_extra_config_with_tensorizer_config(extra_config,
tensorizer_config)
merge_extra_config_with_tensorizer_config(extra_config, tensorizer_config)
tensorize_vllm_model(engine_args, tensorizer_config)
elif args.command == "deserialize":
......@@ -356,11 +379,10 @@ def main():
tensorizer_dir=args.serialized_directory,
encryption_keyfile=keyfile,
deserialization_kwargs=args.deserialization_kwargs or {},
**credentials
**credentials,
)
merge_extra_config_with_tensorizer_config(extra_config,
tensorizer_config)
merge_extra_config_with_tensorizer_config(extra_config, tensorizer_config)
deserialize(args, tensorizer_config)
else:
raise ValueError("Either serialize or deserialize must be specified.")
......
tests/compile/test_silu_mul_quant_fusion.py
View file @ 4e256cad
......@@ -8,16 +8,11 @@ import torch
import vllm.envs as envs
from tests.kernels.quantization.nvfp4_utils import quant_nvfp4_tensor
from vllm._custom_ops import cutlass_scaled_fp4_mm, scaled_fp4_quant
# yapf conflicts with isort for this block
# yapf: disable
from vllm.compilation.activation_quant_fusion import (
FUSED_OPS,
SILU_MUL_OP,
ActivationQuantFusionPass,
)
# yapf: enable
from vllm.compilation.fusion import QUANT_OPS
from vllm.compilation.noop_elimination import NoOpEliminationPass
from vllm.compilation.post_cleanup import PostCleanupPass
......
tests/distributed/test_expert_parallel.py
View file @ 4e256cad
......@@ -107,10 +107,8 @@ class EPTestSettings:
# NOTE: You can adjust tp_base locally to fit the model in GPU
# The values displayed here are only a rough indicator of the size of the model
# yapf: disable
TEST_MODELS = {
"deepseek-ai/DeepSeek-V2-Lite-Chat": EPTestSettings.fast(
trust_remote_code=True),
"deepseek-ai/DeepSeek-V2-Lite-Chat": EPTestSettings.fast(trust_remote_code=True),
"mistralai/Mixtral-8x7B-Instruct-v0.1": EPTestSettings.fast(tp_base=4),
}
......@@ -192,22 +190,24 @@ def _compare_tp(
]
try:
compare_two_settings(model_name,
ep_args,
tp_args,
ep_env,
tp_env,
method=method,
max_wait_seconds=360)
compare_two_settings(
model_name,
ep_args,
tp_args,
ep_env,
tp_env,
method=method,
max_wait_seconds=360,
)
except Exception:
raise
@pytest.mark.parametrize(
("model_name", "parallel_setup", "distributed_backend", "runner",
"test_options"),
("model_name", "parallel_setup", "distributed_backend", "runner", "test_options"),
[
params for model_name, settings in TEST_MODELS.items()
params
for model_name, settings in TEST_MODELS.items()
for params in settings.iter_params(model_name)
],
)
......@@ -220,10 +220,12 @@ def test_ep(
test_options: EPTestOptions,
num_gpus_available,
):
_compare_tp(model_name,
parallel_setup,
distributed_backend,
runner,
test_options,
num_gpus_available,
method="generate")
_compare_tp(
model_name,
parallel_setup,
distributed_backend,
runner,
test_options,
num_gpus_available,
method="generate",
)
tests/distributed/test_pipeline_parallel.py
View file @ 4e256cad
......@@ -100,7 +100,6 @@ class PPTestSettings:
# NOTE: You can adjust tp_base and/or pp_base locally to fit the model in GPU
# The values displayed here are only a rough indicator of the size of the model
# yapf: disable
TEXT_GENERATION_MODELS = {
# [Decoder-only]
# Uses Llama
......@@ -150,7 +149,9 @@ TEXT_GENERATION_MODELS = {
"adept/persimmon-8b-chat": PPTestSettings.fast(),
"microsoft/phi-2": PPTestSettings.fast(),
"microsoft/Phi-3-small-8k-instruct": PPTestSettings.fast(),
"microsoft/Phi-3.5-MoE-instruct": PPTestSettings.detailed(multi_node_only=True, load_format="dummy"), # noqa: E501
"microsoft/Phi-3.5-MoE-instruct": PPTestSettings.detailed(
multi_node_only=True, load_format="dummy"
), # noqa: E501
"Qwen/Qwen-7B-Chat": PPTestSettings.fast(),
"Qwen/Qwen2.5-0.5B-Instruct": PPTestSettings.fast(),
"Qwen/Qwen1.5-MoE-A2.7B-Chat": PPTestSettings.fast(),
......@@ -196,7 +197,6 @@ MULTIMODAL_MODELS = {
"Qwen/Qwen2-VL-2B-Instruct": PPTestSettings.fast(),
"fixie-ai/ultravox-v0_5-llama-3_2-1b": PPTestSettings.fast(),
}
# yapf: enable
# NOTE: You can update this on your local machine to run specific tests
TEST_MODELS = [
......
tests/engine/test_arg_utils.py
View file @ 4e256cad
......@@ -287,29 +287,15 @@ def test_prefix_cache_default():
assert not engine_args.enable_prefix_caching
# yapf: disable
@pytest.mark.parametrize(("arg", "expected", "option"), [
(None, None, "mm-processor-kwargs"),
("{}", {}, "mm-processor-kwargs"),
(
'{"num_crops": 4}',
{
"num_crops": 4
},
"mm-processor-kwargs"
),
(
'{"foo": {"bar": "baz"}}',
{
"foo":
{
"bar": "baz"
}
},
"mm-processor-kwargs"
),
])
# yapf: enable
@pytest.mark.parametrize(
("arg", "expected", "option"),
[
(None, None, "mm-processor-kwargs"),
("{}", {}, "mm-processor-kwargs"),
('{"num_crops": 4}', {"num_crops": 4}, "mm-processor-kwargs"),
('{"foo": {"bar": "baz"}}', {"foo": {"bar": "baz"}}, "mm-processor-kwargs"),
],
)
def test_composite_arg_parser(arg, expected, option):
parser = EngineArgs.add_cli_args(FlexibleArgumentParser())
if arg is None:
......@@ -321,8 +307,7 @@ def test_composite_arg_parser(arg, expected, option):
def test_human_readable_model_len():
# `exit_on_error` disabled to test invalid values below
parser = EngineArgs.add_cli_args(
FlexibleArgumentParser(exit_on_error=False))
parser = EngineArgs.add_cli_args(FlexibleArgumentParser(exit_on_error=False))
args = parser.parse_args([])
assert args.max_model_len is None
......
tests/entrypoints/test_chat_utils.py
View file @ 4e256cad
......@@ -15,6 +15,7 @@ from vllm.assets.video import VideoAsset
from vllm.config import ModelConfig
from vllm.entrypoints.chat_utils import (
_try_extract_ast,
apply_mistral_chat_template,
load_chat_template,
parse_chat_messages,
parse_chat_messages_futures,
......@@ -1855,17 +1856,17 @@ def test_resolve_hf_chat_template_kwargs(sample_json_schema, model, expected_kwa
# NOTE: Qwen2-Audio default chat template is specially defined inside
# processor class instead of using `tokenizer_config.json`
# yapf: disable
@pytest.mark.parametrize(
("model", "expected_format"),
[(PHI3V_MODEL_ID, "string"),
(QWEN2VL_MODEL_ID, "openai"),
(QWEN25VL_MODEL_ID, "openai"),
(ULTRAVOX_MODEL_ID, "string"),
(QWEN2AUDIO_MODEL_ID, "openai"),
(LLAMA_GUARD_MODEL_ID, "openai")],
[
(PHI3V_MODEL_ID, "string"),
(QWEN2VL_MODEL_ID, "openai"),
(QWEN25VL_MODEL_ID, "openai"),
(ULTRAVOX_MODEL_ID, "string"),
(QWEN2AUDIO_MODEL_ID, "openai"),
(LLAMA_GUARD_MODEL_ID, "openai"),
],
)
# yapf: enable
def test_resolve_content_format_hf_defined(model, expected_format):
model_info = HF_EXAMPLE_MODELS.find_hf_info(model)
model_info.check_available_online(on_fail="skip")
......@@ -1879,7 +1880,8 @@ def test_resolve_content_format_hf_defined(model, expected_format):
hf_overrides=model_info.hf_overrides,
skip_tokenizer_init=model_info.skip_tokenizer_init,
enforce_eager=model_info.enforce_eager,
dtype=model_info.dtype)
dtype=model_info.dtype,
)
tokenizer = get_tokenizer(
model,
......@@ -1911,18 +1913,18 @@ def test_resolve_content_format_hf_defined(model, expected_format):
assert resolved_format == expected_format
# yapf: disable
@pytest.mark.parametrize(
("model", "expected_format"),
[("Salesforce/blip2-opt-2.7b", "string"),
("facebook/chameleon-7b", "string"),
("deepseek-ai/deepseek-vl2-tiny", "string"),
("adept/fuyu-8b", "string"),
("google/paligemma-3b-mix-224", "string"),
("Qwen/Qwen-VL", "string"),
("Qwen/Qwen-VL-Chat", "string")],
[
("Salesforce/blip2-opt-2.7b", "string"),
("facebook/chameleon-7b", "string"),
("deepseek-ai/deepseek-vl2-tiny", "string"),
("adept/fuyu-8b", "string"),
("google/paligemma-3b-mix-224", "string"),
("Qwen/Qwen-VL", "string"),
("Qwen/Qwen-VL-Chat", "string"),
],
)
# yapf: enable
def test_resolve_content_format_fallbacks(model, expected_format):
model_info = HF_EXAMPLE_MODELS.find_hf_info(model)
model_info.check_available_online(on_fail="skip")
......@@ -1936,7 +1938,8 @@ def test_resolve_content_format_fallbacks(model, expected_format):
hf_overrides=model_info.hf_overrides,
skip_tokenizer_init=model_info.skip_tokenizer_init,
enforce_eager=model_info.enforce_eager,
dtype=model_info.dtype)
dtype=model_info.dtype,
)
tokenizer = get_tokenizer(
model_config.tokenizer,
......@@ -1968,30 +1971,30 @@ def test_resolve_content_format_fallbacks(model, expected_format):
assert resolved_format == expected_format
# yapf: disable
@pytest.mark.parametrize(
("template_path", "expected_format"),
[("template_alpaca.jinja", "string"),
("template_baichuan.jinja", "string"),
("template_chatglm.jinja", "string"),
("template_chatglm2.jinja", "string"),
("template_chatml.jinja", "string"),
("template_dse_qwen2_vl.jinja", "openai"),
("template_falcon_180b.jinja", "string"),
("template_falcon.jinja", "string"),
("template_inkbot.jinja", "string"),
("template_teleflm.jinja", "string"),
("template_vlm2vec_phi3v.jinja", "openai"),
("template_vlm2vec_qwen2vl.jinja", "openai"),
("tool_chat_template_granite_20b_fc.jinja", "string"),
("tool_chat_template_hermes.jinja", "string"),
("tool_chat_template_internlm2_tool.jinja", "string"),
("tool_chat_template_llama3.1_json.jinja", "openai"),
("tool_chat_template_llama3.2_json.jinja", "openai"),
("tool_chat_template_mistral_parallel.jinja", "string"),
("tool_chat_template_mistral.jinja", "string")],
[
("template_alpaca.jinja", "string"),
("template_baichuan.jinja", "string"),
("template_chatglm.jinja", "string"),
("template_chatglm2.jinja", "string"),
("template_chatml.jinja", "string"),
("template_dse_qwen2_vl.jinja", "openai"),
("template_falcon_180b.jinja", "string"),
("template_falcon.jinja", "string"),
("template_inkbot.jinja", "string"),
("template_teleflm.jinja", "string"),
("template_vlm2vec_phi3v.jinja", "openai"),
("template_vlm2vec_qwen2vl.jinja", "openai"),
("tool_chat_template_granite_20b_fc.jinja", "string"),
("tool_chat_template_hermes.jinja", "string"),
("tool_chat_template_internlm2_tool.jinja", "string"),
("tool_chat_template_llama3.1_json.jinja", "openai"),
("tool_chat_template_llama3.2_json.jinja", "openai"),
("tool_chat_template_mistral_parallel.jinja", "string"),
("tool_chat_template_mistral.jinja", "string"),
],
)
# yapf: enable
def test_resolve_content_format_examples(template_path, expected_format):
model_config = ModelConfig(
PHI3V_MODEL_ID, # Dummy
......@@ -2024,40 +2027,34 @@ def test_resolve_content_format_examples(template_path, expected_format):
assert resolved_format == expected_format
def test_parse_chat_messages_include_thinking_chunk(mistral_model_config,
mistral_tokenizer):
messages = [{
"role":
"system",
"content": [{
"type": "text",
"text": "You are a helpful assistant."
}, {
"type":
"thinking",
"closed":
True,
"thinking":
"Only return the answer when you are confident."
}]
}, {
"role": "user",
"content": "What is 2+2?"
}, {
"role":
"assistant",
"content": [{
"type": "text",
"text": "Let me think about it."
}, {
"type": "thinking",
"closed": True,
"thinking": "2+2 = 4"
}, {
"type": "text",
"text": "The answer is 4.",
}],
}]
def test_parse_chat_messages_include_thinking_chunk(
mistral_model_config, mistral_tokenizer
):
messages = [
{
"role": "system",
"content": [
{"type": "text", "text": "You are a helpful assistant."},
{
"type": "thinking",
"closed": True,
"thinking": "Only return the answer when you are confident.",
},
],
},
{"role": "user", "content": "What is 2+2?"},
{
"role": "assistant",
"content": [
{"type": "text", "text": "Let me think about it."},
{"type": "thinking", "closed": True, "thinking": "2+2 = 4"},
{
"type": "text",
"text": "The answer is 4.",
},
],
},
]
conversation_with_thinking, _, _ = parse_chat_messages(
messages,
......@@ -2066,122 +2063,105 @@ def test_parse_chat_messages_include_thinking_chunk(mistral_model_config,
content_format="openai",
)
expected_conversation = [{
"role":
"system",
"content": [{
"type": "text",
"text": "You are a helpful assistant."
}, {
"type": "text",
"text": "Only return the answer when you are confident."
}],
}, {
"role":
"user",
"content": [{
"type": "text",
"text": "What is 2+2?"
}],
}, {
"role":
"assistant",
"content": [
{
"type": "text",
"text": "Let me think about it."
},
{
"type": "text",
"text": "2+2 = 4"
},
{
"type": "text",
"text": "The answer is 4."
},
]
}]
expected_conversation = [
{
"role": "system",
"content": [
{"type": "text", "text": "You are a helpful assistant."},
{
"type": "text",
"text": "Only return the answer when you are confident.",
},
],
},
{
"role": "user",
"content": [{"type": "text", "text": "What is 2+2?"}],
},
{
"role": "assistant",
"content": [
{"type": "text", "text": "Let me think about it."},
{"type": "text", "text": "2+2 = 4"},
{"type": "text", "text": "The answer is 4."},
],
},
]
assert conversation_with_thinking == expected_conversation
def test_apply_mistral_chat_template_thinking_chunk():
# Moved import here to avoid yapf and isort conflicts
from vllm.entrypoints.chat_utils import apply_mistral_chat_template
messages = [{
"role":
"system",
"content": [{
"type": "text",
"text": "You are a helpful assistant."
}, {
"type":
"thinking",
"closed":
True,
"thinking":
"Only return the answer when you are confident."
}]
}, {
"role": "user",
"content": "What is 2+2?"
}, {
"role":
"assistant",
"content": [{
"type": "text",
"text": "Let me think about it."
}, {
"type": "thinking",
"closed": True,
"thinking": "2+2 = 4"
}, {
"type": "text",
"text": "The answer is 4.",
}],
}, {
"role": "user",
"content": "Thanks, what is 3+3?"
}]
messages = [
{
"role": "system",
"content": [
{"type": "text", "text": "You are a helpful assistant."},
{
"type": "thinking",
"closed": True,
"thinking": "Only return the answer when you are confident.",
},
],
},
{"role": "user", "content": "What is 2+2?"},
{
"role": "assistant",
"content": [
{"type": "text", "text": "Let me think about it."},
{"type": "thinking", "closed": True, "thinking": "2+2 = 4"},
{
"type": "text",
"text": "The answer is 4.",
},
],
},
{"role": "user", "content": "Thanks, what is 3+3?"},
]
# TODO(Julien): upon model release change to a tokenizer already configured.
# =================================================================
mistral_tokenizer = MistralTokenizer.from_pretrained(
"mistralai/Devstral-Small-2507")
"mistralai/Devstral-Small-2507"
)
assert isinstance(mistral_tokenizer.tokenizer, Tekkenizer)
# Add think special tokens to the tokenizer
mistral_tokenizer.tokenizer._all_special_tokens[35] = SpecialTokenInfo(
rank=35, is_control=True, token_str=SpecialTokens.begin_think.value)
rank=35, is_control=True, token_str=SpecialTokens.begin_think.value
)
mistral_tokenizer.tokenizer._all_special_tokens[36] = SpecialTokenInfo(
rank=36, is_control=True, token_str=SpecialTokens.end_think.value)
rank=36, is_control=True, token_str=SpecialTokens.end_think.value
)
mistral_tokenizer.tokenizer._special_tokens_reverse_vocab = {
k: v
for k, v in
mistral_tokenizer.tokenizer._special_tokens_reverse_vocab.items()
for k, v in mistral_tokenizer.tokenizer._special_tokens_reverse_vocab.items()
if v not in {35, 36}
}
mistral_tokenizer.tokenizer._special_tokens_reverse_vocab[
SpecialTokens.begin_think.value] = 35
SpecialTokens.begin_think.value
] = 35
mistral_tokenizer.tokenizer._special_tokens_reverse_vocab[
SpecialTokens.end_think.value] = 36
SpecialTokens.end_think.value
] = 36
mistral_tokenizer.instruct.BEGIN_THINK = 35
mistral_tokenizer.instruct.END_THINK = 36
# =================================================================
tokens_ids = apply_mistral_chat_template(mistral_tokenizer,
messages,
chat_template=None,
tools=None)
tokens_ids = apply_mistral_chat_template(
mistral_tokenizer, messages, chat_template=None, tools=None
)
string_tokens = mistral_tokenizer.mistral.decode(
tokens_ids, special_token_policy=SpecialTokenPolicy.KEEP)
tokens_ids, special_token_policy=SpecialTokenPolicy.KEEP
)
expected_tokens = (
r"<s>[SYSTEM_PROMPT]You are a helpful assistant.[THINK]Only return the"
r" answer when you are confident.[/THINK][/SYSTEM_PROMPT]"
r"[INST]What is 2+2?[/INST]"
r"Let me think about it.[THINK]2+2 = 4[/THINK]The answer is 4.</s>"
r"[INST]Thanks, what is 3+3?[/INST]")
r"[INST]Thanks, what is 3+3?[/INST]"
)
assert string_tokens == expected_tokens
......@@ -2192,37 +2172,32 @@ def test_parse_chat_messages_single_empty_audio_with_uuid(
):
audio_uuid = "abcd"
conversation, mm_data, mm_uuids = parse_chat_messages(
[{
"role":
"user",
"content": [
{
"type": "input_audio",
"input_audio": {},
"uuid": audio_uuid,
},
{
"type": "text",
"text": "What does the audio say?"
},
],
}],
[
{
"role": "user",
"content": [
{
"type": "input_audio",
"input_audio": {},
"uuid": audio_uuid,
},
{"type": "text", "text": "What does the audio say?"},
],
}
],
qwen2_audio_model_config,
qwen2_audio_tokenizer,
content_format="string",
)
assert conversation == [{
"role":
"user",
"content":
"Audio 1: <|audio_bos|><|AUDIO|><|audio_eos|>\nWhat does the audio say?"
}]
assert conversation == [
{
"role": "user",
"content": "Audio 1: <|audio_bos|><|AUDIO|><|audio_eos|>\nWhat does the audio say?",
}
]
_assert_mm_data_inputs(mm_data, {"audio": 1})
_assert_mm_uuids(mm_uuids,
1,
modality="audio",
expected_uuids=[audio_uuid])
_assert_mm_uuids(mm_uuids, 1, modality="audio", expected_uuids=[audio_uuid])
@pytest.mark.asyncio
......@@ -2232,34 +2207,29 @@ async def test_parse_chat_messages_single_empty_audio_with_uuid_async(
):
audio_uuid = "abcd"
conversation, mm_future, mm_uuids = parse_chat_messages_futures(
[{
"role":
"user",
"content": [
{
"type": "input_audio",
"input_audio": {},
"uuid": audio_uuid,
},
{
"type": "text",
"text": "What does the audio say?"
},
],
}],
[
{
"role": "user",
"content": [
{
"type": "input_audio",
"input_audio": {},
"uuid": audio_uuid,
},
{"type": "text", "text": "What does the audio say?"},
],
}
],
qwen2_audio_model_config,
qwen2_audio_tokenizer,
content_format="string",
)
assert conversation == [{
"role":
"user",
"content":
"Audio 1: <|audio_bos|><|AUDIO|><|audio_eos|>\nWhat does the audio say?"
}]
assert conversation == [
{
"role": "user",
"content": "Audio 1: <|audio_bos|><|AUDIO|><|audio_eos|>\nWhat does the audio say?",
}
]
_assert_mm_data_inputs(await mm_future, {"audio": 1})
_assert_mm_uuids(mm_uuids,
1,
modality="audio",
expected_uuids=[audio_uuid])
_assert_mm_uuids(mm_uuids, 1, modality="audio", expected_uuids=[audio_uuid])
tests/lora/test_layers.py
View file @ 4e256cad
......@@ -12,9 +12,6 @@ import torch
import torch.nn.functional as F
from vllm.config.lora import LoRAConfig
# yapf conflicts with isort for this block
# yapf: disable
from vllm.lora.layers import (
BaseLayerWithLoRA,
ColumnParallelLinearWithLoRA,
......@@ -32,8 +29,6 @@ from vllm.lora.layers import (
RowParallelLinearWithShardedLoRA,
VocabParallelEmbeddingWithLoRA,
)
# yapf: enable
from vllm.lora.models import LoRALayerWeights, PackedLoRALayerWeights
from vllm.lora.punica_wrapper import get_punica_wrapper
from vllm.model_executor.layers.linear import (
......
tests/model_executor/model_loader/tensorizer_loader/test_tensorizer.py
View file @ 4e256cad
......@@ -17,8 +17,6 @@ import vllm.model_executor.model_loader.tensorizer
from tests.utils import VLLM_PATH, RemoteOpenAIServer
from vllm import LLM, SamplingParams
from vllm.engine.arg_utils import EngineArgs
# yapf: disable
from vllm.model_executor.model_loader.tensorizer import (
TensorizerConfig,
TensorSerializer,
......@@ -29,8 +27,6 @@ from vllm.model_executor.model_loader.tensorizer import (
from vllm.model_executor.model_loader.tensorizer_loader import (
BLACKLISTED_TENSORIZER_ARGS,
)
# yapf: enable
from vllm.utils import PlaceholderModule
from .conftest import DummyExecutor, assert_from_collective_rpc
......
tests/models/multimodal/generation/test_common.py
View file @ 4e256cad
......@@ -45,18 +45,17 @@ from .vlm_utils.types import (
if current_platform.is_rocm():
os.environ["VLLM_USE_TRITON_FLASH_ATTN"] = "0"
# yapf: disable
COMMON_BROADCAST_SETTINGS = {
"test_type": VLMTestType.IMAGE,
"dtype": "half",
"max_tokens": 5,
"tensor_parallel_size": 2,
"hf_model_kwargs": {"device_map": "auto"},
"image_size_factors": [(.25, 0.5, 1.0)],
"image_size_factors": [(0.25, 0.5, 1.0)],
"distributed_executor_backend": (
"ray",
"mp",
)
),
}
### Test configuration for specific models
......@@ -96,22 +95,20 @@ VLM_TEST_SETTINGS = {
#### Core tests to always run in the CI
"llava": VLMTestInfo(
models=["llava-hf/llava-1.5-7b-hf"],
test_type=(
VLMTestType.EMBEDDING,
VLMTestType.IMAGE,
VLMTestType.CUSTOM_INPUTS
),
test_type=(VLMTestType.EMBEDDING, VLMTestType.IMAGE, VLMTestType.CUSTOM_INPUTS),
prompt_formatter=lambda img_prompt: f"USER: {img_prompt}\nASSISTANT:",
convert_assets_to_embeddings=model_utils.get_llava_embeddings,
max_model_len=4096,
auto_cls=AutoModelForImageTextToText,
vllm_output_post_proc=model_utils.llava_image_vllm_to_hf_output,
custom_test_opts=[CustomTestOptions(
inputs=custom_inputs.multi_image_multi_aspect_ratio_inputs(
formatter=lambda img_prompt: f"USER: {img_prompt}\nASSISTANT:"
),
limit_mm_per_prompt={"image": 4},
)],
custom_test_opts=[
CustomTestOptions(
inputs=custom_inputs.multi_image_multi_aspect_ratio_inputs(
formatter=lambda img_prompt: f"USER: {img_prompt}\nASSISTANT:"
),
limit_mm_per_prompt={"image": 4},
)
],
# TODO: Revert to "auto" when CPU backend can use torch > 2.6
dtype="bfloat16" if current_platform.is_cpu() else "auto",
marks=[pytest.mark.core_model, pytest.mark.cpu_model],
......@@ -120,27 +117,27 @@ VLM_TEST_SETTINGS = {
models=["google/paligemma-3b-mix-224"],
test_type=VLMTestType.IMAGE,
prompt_formatter=identity,
img_idx_to_prompt = lambda idx: "",
img_idx_to_prompt=lambda idx: "",
# Paligemma uses its own sample prompts because the default one fails
single_image_prompts=IMAGE_ASSETS.prompts({
"stop_sign": "caption es",
"cherry_blossom": "What is in the picture?",
}),
single_image_prompts=IMAGE_ASSETS.prompts(
{
"stop_sign": "caption es",
"cherry_blossom": "What is in the picture?",
}
),
auto_cls=AutoModelForImageTextToText,
vllm_output_post_proc=model_utils.paligemma_vllm_to_hf_output,
dtype="bfloat16",
marks=[pytest.mark.skip(reason="vLLM does not support PrefixLM attention mask")], # noqa: E501
marks=[
pytest.mark.skip(reason="vLLM does not support PrefixLM attention mask")
], # noqa: E501
),
"qwen2_5_vl": VLMTestInfo(
models=["Qwen/Qwen2.5-VL-3B-Instruct"],
test_type=(
VLMTestType.IMAGE,
VLMTestType.MULTI_IMAGE,
VLMTestType.VIDEO
),
prompt_formatter=lambda img_prompt: f"<|im_start|>User\n{img_prompt}<|im_end|>\n<|im_start|>assistant\n", # noqa: E501
img_idx_to_prompt=lambda idx: "<|vision_start|><|image_pad|><|vision_end|>", # noqa: E501
video_idx_to_prompt=lambda idx: "<|vision_start|><|video_pad|><|vision_end|>", # noqa: E501
test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE, VLMTestType.VIDEO),
prompt_formatter=lambda img_prompt: f"<|im_start|>User\n{img_prompt}<|im_end|>\n<|im_start|>assistant\n", # noqa: E501
img_idx_to_prompt=lambda idx: "<|vision_start|><|image_pad|><|vision_end|>", # noqa: E501
video_idx_to_prompt=lambda idx: "<|vision_start|><|video_pad|><|vision_end|>", # noqa: E501
max_model_len=4096,
max_num_seqs=2,
auto_cls=AutoModelForImageTextToText,
......@@ -150,17 +147,13 @@ VLM_TEST_SETTINGS = {
),
"qwen2_5_omni": VLMTestInfo(
models=["Qwen/Qwen2.5-Omni-3B"],
test_type=(
VLMTestType.IMAGE,
VLMTestType.MULTI_IMAGE,
VLMTestType.VIDEO
),
prompt_formatter=lambda img_prompt: f"<|im_start|>User\n{img_prompt}<|im_end|>\n<|im_start|>assistant\n", # noqa: E501
img_idx_to_prompt=lambda idx: "<|vision_bos|><|IMAGE|><|vision_eos|>", # noqa: E501
video_idx_to_prompt=lambda idx: "<|vision_bos|><|VIDEO|><|vision_eos|>", # noqa: E501
test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE, VLMTestType.VIDEO),
prompt_formatter=lambda img_prompt: f"<|im_start|>User\n{img_prompt}<|im_end|>\n<|im_start|>assistant\n", # noqa: E501
img_idx_to_prompt=lambda idx: "<|vision_bos|><|IMAGE|><|vision_eos|>", # noqa: E501
video_idx_to_prompt=lambda idx: "<|vision_bos|><|VIDEO|><|vision_eos|>", # noqa: E501
max_model_len=4096,
max_num_seqs=2,
num_logprobs= 6 if current_platform.is_cpu() else 5,
num_logprobs=6 if current_platform.is_cpu() else 5,
auto_cls=AutoModelForTextToWaveform,
vllm_output_post_proc=model_utils.qwen2_vllm_to_hf_output,
patch_hf_runner=model_utils.qwen2_5_omni_patch_hf_runner,
......@@ -168,9 +161,9 @@ VLM_TEST_SETTINGS = {
marks=[pytest.mark.core_model, pytest.mark.cpu_model],
),
"ultravox": VLMTestInfo(
models = ["fixie-ai/ultravox-v0_5-llama-3_2-1b"],
models=["fixie-ai/ultravox-v0_5-llama-3_2-1b"],
test_type=VLMTestType.AUDIO,
prompt_formatter=lambda audio_prompt: f"<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\n{audio_prompt}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n", # noqa: E501
prompt_formatter=lambda audio_prompt: f"<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\n{audio_prompt}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n", # noqa: E501
audio_idx_to_prompt=lambda idx: "<|audio|>",
max_model_len=4096,
max_num_seqs=2,
......@@ -184,9 +177,11 @@ VLM_TEST_SETTINGS = {
"llava-onevision-transformers": VLMTestInfo(
models=["llava-hf/llava-onevision-qwen2-0.5b-ov-hf"],
test_type=VLMTestType.IMAGE,
prompt_formatter=lambda vid_prompt: f"<|im_start|>user\n{vid_prompt}<|im_end|>\n<|im_start|>assistant\n", # noqa: E501
prompt_formatter=lambda vid_prompt: f"<|im_start|>user\n{vid_prompt}<|im_end|>\n<|im_start|>assistant\n", # noqa: E501
max_model_len=16384,
hf_model_kwargs=model_utils.llava_onevision_hf_model_kwargs("llava-hf/llava-onevision-qwen2-0.5b-ov-hf"), # noqa: E501
hf_model_kwargs=model_utils.llava_onevision_hf_model_kwargs(
"llava-hf/llava-onevision-qwen2-0.5b-ov-hf"
), # noqa: E501
auto_cls=AutoModelForImageTextToText,
vllm_output_post_proc=model_utils.llava_onevision_vllm_to_hf_output,
image_size_factors=[(0.25, 0.5, 1.0)],
......@@ -201,7 +196,7 @@ VLM_TEST_SETTINGS = {
"idefics3-transformers": VLMTestInfo(
models=["HuggingFaceTB/SmolVLM-256M-Instruct"],
test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
prompt_formatter=lambda img_prompt:f"<|begin_of_text|>User:{img_prompt}<end_of_utterance>\nAssistant:", # noqa: E501
prompt_formatter=lambda img_prompt: f"<|begin_of_text|>User:{img_prompt}<end_of_utterance>\nAssistant:", # noqa: E501
img_idx_to_prompt=lambda idx: "<image>",
max_model_len=8192,
max_num_seqs=2,
......@@ -217,8 +212,8 @@ VLM_TEST_SETTINGS = {
"qwen2_5_vl-transformers": VLMTestInfo(
models=["Qwen/Qwen2.5-VL-3B-Instruct"],
test_type=VLMTestType.IMAGE,
prompt_formatter=lambda img_prompt: f"<|im_start|>User\n{img_prompt}<|im_end|>\n<|im_start|>assistant\n", # noqa: E501
img_idx_to_prompt=lambda idx: "<|vision_start|><|image_pad|><|vision_end|>", # noqa: E501
prompt_formatter=lambda img_prompt: f"<|im_start|>User\n{img_prompt}<|im_end|>\n<|im_start|>assistant\n", # noqa: E501
img_idx_to_prompt=lambda idx: "<|vision_start|><|image_pad|><|vision_end|>", # noqa: E501
max_model_len=4096,
max_num_seqs=2,
auto_cls=AutoModelForImageTextToText,
......@@ -228,23 +223,24 @@ VLM_TEST_SETTINGS = {
"model_impl": "transformers",
},
# FIXME: Investigate mrope issue
marks=[large_gpu_mark(min_gb=32),
pytest.mark.skip(reason="Mrope issue")],
marks=[large_gpu_mark(min_gb=32), pytest.mark.skip(reason="Mrope issue")],
),
#### Extended model tests
"aria": VLMTestInfo(
models=["rhymes-ai/Aria"],
test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
prompt_formatter=lambda img_prompt: f"<|im_start|>user\n{img_prompt}<|im_end|>\n<|im_start|>assistant\n ", # noqa: E501
prompt_formatter=lambda img_prompt: f"<|im_start|>user\n{img_prompt}<|im_end|>\n<|im_start|>assistant\n ", # noqa: E501
img_idx_to_prompt=lambda idx: "<fim_prefix><|img|><fim_suffix>\n",
max_model_len=4096,
max_num_seqs=2,
auto_cls=AutoModelForImageTextToText,
single_image_prompts=IMAGE_ASSETS.prompts({
"stop_sign": "<vlm_image>Please describe the image shortly.",
"cherry_blossom": "<vlm_image>Please infer the season with reason.", # noqa: E501
}),
multi_image_prompt="<vlm_image><vlm_image>Describe the two images shortly.", # noqa: E501
single_image_prompts=IMAGE_ASSETS.prompts(
{
"stop_sign": "<vlm_image>Please describe the image shortly.",
"cherry_blossom": "<vlm_image>Please infer the season with reason.", # noqa: E501
}
),
multi_image_prompt="<vlm_image><vlm_image>Describe the two images shortly.", # noqa: E501
stop_str=["<|im_end|>"],
image_size_factors=[(0.10, 0.15)],
max_tokens=64,
......@@ -253,11 +249,13 @@ VLM_TEST_SETTINGS = {
"aya_vision": VLMTestInfo(
models=["CohereForAI/aya-vision-8b"],
test_type=(VLMTestType.IMAGE),
prompt_formatter=lambda img_prompt: f"<|START_OF_TURN_TOKEN|><|USER_TOKEN|>{img_prompt}<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>", # noqa: E501
single_image_prompts=IMAGE_ASSETS.prompts({
"stop_sign": "<image>What's the content in the center of the image?", # noqa: E501
"cherry_blossom": "<image>What is the season?", # noqa: E501
}),
prompt_formatter=lambda img_prompt: f"<|START_OF_TURN_TOKEN|><|USER_TOKEN|>{img_prompt}<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>", # noqa: E501
single_image_prompts=IMAGE_ASSETS.prompts(
{
"stop_sign": "<image>What's the content in the center of the image?", # noqa: E501
"cherry_blossom": "<image>What is the season?", # noqa: E501
}
),
multi_image_prompt="<image><image>Describe the two images in detail.", # noqa: E501
max_model_len=4096,
max_num_seqs=2,
......@@ -267,11 +265,13 @@ VLM_TEST_SETTINGS = {
"aya_vision-multi_image": VLMTestInfo(
models=["CohereForAI/aya-vision-8b"],
test_type=(VLMTestType.MULTI_IMAGE),
prompt_formatter=lambda img_prompt: f"<|START_OF_TURN_TOKEN|><|USER_TOKEN|>{img_prompt}<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>", # noqa: E501
single_image_prompts=IMAGE_ASSETS.prompts({
"stop_sign": "<image>What's the content in the center of the image?", # noqa: E501
"cherry_blossom": "<image>What is the season?", # noqa: E501
}),
prompt_formatter=lambda img_prompt: f"<|START_OF_TURN_TOKEN|><|USER_TOKEN|>{img_prompt}<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>", # noqa: E501
single_image_prompts=IMAGE_ASSETS.prompts(
{
"stop_sign": "<image>What's the content in the center of the image?", # noqa: E501
"cherry_blossom": "<image>What is the season?", # noqa: E501
}
),
multi_image_prompt="<image><image>Describe the two images in detail.", # noqa: E501
max_model_len=4096,
max_num_seqs=2,
......@@ -297,27 +297,29 @@ VLM_TEST_SETTINGS = {
max_num_seqs=2,
auto_cls=AutoModelForImageTextToText,
# For chameleon, we only compare the sequences
vllm_output_post_proc = lambda vllm_output, model: vllm_output[:2],
hf_output_post_proc = lambda hf_output, model: hf_output[:2],
vllm_output_post_proc=lambda vllm_output, model: vllm_output[:2],
hf_output_post_proc=lambda hf_output, model: hf_output[:2],
comparator=check_outputs_equal,
max_tokens=8,
dtype="bfloat16",
),
"deepseek_vl_v2": VLMTestInfo(
models=["Isotr0py/deepseek-vl2-tiny"], # model repo using dynamic module
models=["Isotr0py/deepseek-vl2-tiny"], # model repo using dynamic module
test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
prompt_formatter=lambda img_prompt: f"<|User|>: {img_prompt}\n\n<|Assistant|>: ", # noqa: E501
prompt_formatter=lambda img_prompt: f"<|User|>: {img_prompt}\n\n<|Assistant|>: ", # noqa: E501
max_model_len=4096,
max_num_seqs=2,
single_image_prompts=IMAGE_ASSETS.prompts({
"stop_sign": "<image>\nWhat's the content in the center of the image?", # noqa: E501
"cherry_blossom": "<image>\nPlease infer the season with reason in details.", # noqa: E501
}),
multi_image_prompt="image_1:<image>\nimage_2:<image>\nWhich image can we see the car and the tower?", # noqa: E501
single_image_prompts=IMAGE_ASSETS.prompts(
{
"stop_sign": "<image>\nWhat's the content in the center of the image?", # noqa: E501
"cherry_blossom": "<image>\nPlease infer the season with reason in details.", # noqa: E501
}
),
multi_image_prompt="image_1:<image>\nimage_2:<image>\nWhich image can we see the car and the tower?", # noqa: E501
patch_hf_runner=model_utils.deepseekvl2_patch_hf_runner,
hf_output_post_proc=model_utils.deepseekvl2_trunc_hf_output,
stop_str=["<｜end▁of▁sentence｜>", "<｜begin▁of▁sentence｜>"], # noqa: E501
image_size_factors=[(), (1.0, ), (1.0, 1.0, 1.0), (0.1, 0.5, 1.0)],
image_size_factors=[(), (1.0,), (1.0, 1.0, 1.0), (0.1, 0.5, 1.0)],
),
"fuyu": VLMTestInfo(
models=["adept/fuyu-8b"],
......@@ -336,11 +338,13 @@ VLM_TEST_SETTINGS = {
"gemma3": VLMTestInfo(
models=["google/gemma-3-4b-it"],
test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
prompt_formatter=lambda img_prompt: f"<bos><start_of_turn>user\n{img_prompt}<end_of_turn>\n<start_of_turn>model\n", # noqa: E501
single_image_prompts=IMAGE_ASSETS.prompts({
"stop_sign": "<start_of_image>What's the content in the center of the image?", # noqa: E501
"cherry_blossom": "<start_of_image>What is the season?", # noqa: E501
}),
prompt_formatter=lambda img_prompt: f"<bos><start_of_turn>user\n{img_prompt}<end_of_turn>\n<start_of_turn>model\n", # noqa: E501
single_image_prompts=IMAGE_ASSETS.prompts(
{
"stop_sign": "<start_of_image>What's the content in the center of the image?", # noqa: E501
"cherry_blossom": "<start_of_image>What is the season?", # noqa: E501
}
),
multi_image_prompt="<start_of_image><start_of_image>Describe the two images in detail.", # noqa: E501
max_model_len=4096,
max_num_seqs=2,
......@@ -353,10 +357,12 @@ VLM_TEST_SETTINGS = {
models=["zai-org/glm-4v-9b"],
test_type=VLMTestType.IMAGE,
prompt_formatter=lambda img_prompt: f"<|user|>\n{img_prompt}<|assistant|>", # noqa: E501
single_image_prompts=IMAGE_ASSETS.prompts({
"stop_sign": "<|begin_of_image|><|endoftext|><|end_of_image|>What's the content in the center of the image?", # noqa: E501
"cherry_blossom": "<|begin_of_image|><|endoftext|><|end_of_image|>What is the season?", # noqa: E501
}),
single_image_prompts=IMAGE_ASSETS.prompts(
{
"stop_sign": "<|begin_of_image|><|endoftext|><|end_of_image|>What's the content in the center of the image?", # noqa: E501
"cherry_blossom": "<|begin_of_image|><|endoftext|><|end_of_image|>What is the season?", # noqa: E501
}
),
max_model_len=2048,
max_num_seqs=2,
get_stop_token_ids=lambda tok: [151329, 151336, 151338],
......@@ -372,8 +378,8 @@ VLM_TEST_SETTINGS = {
models=["zai-org/GLM-4.1V-9B-Thinking"],
test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
prompt_formatter=lambda img_prompt: f"<|user|>\n{img_prompt}<|assistant|>", # noqa: E501
img_idx_to_prompt=lambda idx: "<|begin_of_image|><|image|><|end_of_image|>", # noqa: E501
video_idx_to_prompt=lambda idx: "<|begin_of_video|><|video|><|end_of_video|>", # noqa: E501
img_idx_to_prompt=lambda idx: "<|begin_of_image|><|image|><|end_of_image|>", # noqa: E501
video_idx_to_prompt=lambda idx: "<|begin_of_video|><|video|><|end_of_video|>", # noqa: E501
max_model_len=2048,
max_num_seqs=2,
get_stop_token_ids=lambda tok: [151329, 151336, 151338],
......@@ -390,23 +396,27 @@ VLM_TEST_SETTINGS = {
max_num_seqs=2,
auto_cls=AutoModelForImageTextToText,
patch_hf_runner=model_utils.glm4_1v_patch_hf_runner,
custom_test_opts=[CustomTestOptions(
inputs=custom_inputs.video_with_metadata_glm4_1v(),
limit_mm_per_prompt={"video": 1},
)],
custom_test_opts=[
CustomTestOptions(
inputs=custom_inputs.video_with_metadata_glm4_1v(),
limit_mm_per_prompt={"video": 1},
)
],
marks=[large_gpu_mark(min_gb=32)],
),
"h2ovl": VLMTestInfo(
models = [
models=[
"h2oai/h2ovl-mississippi-800m",
"h2oai/h2ovl-mississippi-2b",
],
test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
prompt_formatter=lambda img_prompt: f"<|prompt|>{img_prompt}<|end|><|answer|>", # noqa: E501
single_image_prompts=IMAGE_ASSETS.prompts({
"stop_sign": "<image>\nWhat's the content in the center of the image?", # noqa: E501
"cherry_blossom": "<image>\nWhat is the season?",
}),
prompt_formatter=lambda img_prompt: f"<|prompt|>{img_prompt}<|end|><|answer|>", # noqa: E501
single_image_prompts=IMAGE_ASSETS.prompts(
{
"stop_sign": "<image>\nWhat's the content in the center of the image?", # noqa: E501
"cherry_blossom": "<image>\nWhat is the season?",
}
),
multi_image_prompt="Image-1: <image>\nImage-2: <image>\nDescribe the two images in short.", # noqa: E501
max_model_len=8192,
use_tokenizer_eos=True,
......@@ -416,7 +426,7 @@ VLM_TEST_SETTINGS = {
"idefics3": VLMTestInfo(
models=["HuggingFaceTB/SmolVLM-256M-Instruct"],
test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
prompt_formatter=lambda img_prompt:f"<|begin_of_text|>User:{img_prompt}<end_of_utterance>\nAssistant:", # noqa: E501
prompt_formatter=lambda img_prompt: f"<|begin_of_text|>User:{img_prompt}<end_of_utterance>\nAssistant:", # noqa: E501
img_idx_to_prompt=lambda idx: "<image>",
max_model_len=8192,
max_num_seqs=2,
......@@ -431,11 +441,13 @@ VLM_TEST_SETTINGS = {
# "OpenGVLab/Mono-InternVL-2B",
],
test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
prompt_formatter=lambda img_prompt: f"<|im_start|>User\n{img_prompt}<|im_end|>\n<|im_start|>Assistant\n", # noqa: E501
single_image_prompts=IMAGE_ASSETS.prompts({
"stop_sign": "<image>\nWhat's the content in the center of the image?", # noqa: E501
"cherry_blossom": "<image>\nWhat is the season?",
}),
prompt_formatter=lambda img_prompt: f"<|im_start|>User\n{img_prompt}<|im_end|>\n<|im_start|>Assistant\n", # noqa: E501
single_image_prompts=IMAGE_ASSETS.prompts(
{
"stop_sign": "<image>\nWhat's the content in the center of the image?", # noqa: E501
"cherry_blossom": "<image>\nWhat is the season?",
}
),
multi_image_prompt="Image-1: <image>\nImage-2: <image>\nDescribe the two images in short.", # noqa: E501
max_model_len=4096,
use_tokenizer_eos=True,
......@@ -446,7 +458,7 @@ VLM_TEST_SETTINGS = {
"OpenGVLab/InternVL3-1B",
],
test_type=VLMTestType.VIDEO,
prompt_formatter=lambda img_prompt: f"<|im_start|>User\n{img_prompt}<|im_end|>\n<|im_start|>Assistant\n", # noqa: E501
prompt_formatter=lambda img_prompt: f"<|im_start|>User\n{img_prompt}<|im_end|>\n<|im_start|>Assistant\n", # noqa: E501
video_idx_to_prompt=lambda idx: "<video>",
max_model_len=8192,
use_tokenizer_eos=True,
......@@ -459,7 +471,7 @@ VLM_TEST_SETTINGS = {
VLMTestType.MULTI_IMAGE,
VLMTestType.VIDEO,
),
prompt_formatter=lambda img_prompt: f"<|im_start|>User\n{img_prompt}<|im_end|>\n<|im_start|>Assistant\n", # noqa: E501
prompt_formatter=lambda img_prompt: f"<|im_start|>User\n{img_prompt}<|im_end|>\n<|im_start|>Assistant\n", # noqa: E501
img_idx_to_prompt=lambda idx: "<IMG_CONTEXT>",
video_idx_to_prompt=lambda idx: "<video>",
max_model_len=8192,
......@@ -469,7 +481,7 @@ VLM_TEST_SETTINGS = {
"kimi_vl": VLMTestInfo(
models=["moonshotai/Kimi-VL-A3B-Instruct"],
test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
prompt_formatter=lambda img_prompt: f"<|im_user|>user<|im_middle|>{img_prompt}<|im_end|><|im_assistant|>assistant<|im_middle|>", # noqa: E501
prompt_formatter=lambda img_prompt: f"<|im_user|>user<|im_middle|>{img_prompt}<|im_end|><|im_assistant|>assistant<|im_middle|>", # noqa: E501
img_idx_to_prompt=lambda _: "<|media_start|>image<|media_content|><|media_pad|><|media_end|>", # noqa: E501
max_model_len=8192,
max_num_seqs=2,
......@@ -480,11 +492,11 @@ VLM_TEST_SETTINGS = {
),
"llama4": VLMTestInfo(
models=["meta-llama/Llama-4-Scout-17B-16E-Instruct"],
prompt_formatter=lambda img_prompt: f"<|begin_of_text|><|header_start|>user<|header_end|>\n\n{img_prompt}<|eot|><|header_start|>assistant<|header_end|>\n\n", # noqa: E501
prompt_formatter=lambda img_prompt: f"<|begin_of_text|><|header_start|>user<|header_end|>\n\n{img_prompt}<|eot|><|header_start|>assistant<|header_end|>\n\n", # noqa: E501
img_idx_to_prompt=lambda _: "<|image|>",
test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
distributed_executor_backend="mp",
image_size_factors=[(.25, 0.5, 1.0)],
image_size_factors=[(0.25, 0.5, 1.0)],
hf_model_kwargs={"device_map": "auto"},
max_model_len=8192,
max_num_seqs=4,
......@@ -500,28 +512,34 @@ VLM_TEST_SETTINGS = {
max_model_len=10240,
auto_cls=AutoModelForImageTextToText,
vllm_output_post_proc=model_utils.llava_image_vllm_to_hf_output,
custom_test_opts=[CustomTestOptions(
inputs=custom_inputs.multi_image_multi_aspect_ratio_inputs(
formatter=lambda img_prompt: f"[INST] {img_prompt} [/INST]"
),
limit_mm_per_prompt={"image": 4},
)],
custom_test_opts=[
CustomTestOptions(
inputs=custom_inputs.multi_image_multi_aspect_ratio_inputs(
formatter=lambda img_prompt: f"[INST] {img_prompt} [/INST]"
),
limit_mm_per_prompt={"image": 4},
)
],
),
"llava_onevision": VLMTestInfo(
models=["llava-hf/llava-onevision-qwen2-0.5b-ov-hf"],
test_type=VLMTestType.CUSTOM_INPUTS,
prompt_formatter=lambda vid_prompt: f"<|im_start|>user\n{vid_prompt}<|im_end|>\n<|im_start|>assistant\n", # noqa: E501
prompt_formatter=lambda vid_prompt: f"<|im_start|>user\n{vid_prompt}<|im_end|>\n<|im_start|>assistant\n", # noqa: E501
num_video_frames=16,
max_model_len=16384,
hf_model_kwargs=model_utils.llava_onevision_hf_model_kwargs("llava-hf/llava-onevision-qwen2-0.5b-ov-hf"), # noqa: E501
hf_model_kwargs=model_utils.llava_onevision_hf_model_kwargs(
"llava-hf/llava-onevision-qwen2-0.5b-ov-hf"
), # noqa: E501
auto_cls=AutoModelForImageTextToText,
vllm_output_post_proc=model_utils.llava_onevision_vllm_to_hf_output,
custom_test_opts=[CustomTestOptions(
inputs=custom_inputs.multi_video_multi_aspect_ratio_inputs(
formatter=lambda vid_prompt: f"<|im_start|>user\n{vid_prompt}<|im_end|>\n<|im_start|>assistant\n", # noqa: E501
),
limit_mm_per_prompt={"video": 4},
)],
custom_test_opts=[
CustomTestOptions(
inputs=custom_inputs.multi_video_multi_aspect_ratio_inputs(
formatter=lambda vid_prompt: f"<|im_start|>user\n{vid_prompt}<|im_end|>\n<|im_start|>assistant\n", # noqa: E501
),
limit_mm_per_prompt={"video": 4},
)
],
),
"llava_next_video": VLMTestInfo(
models=["llava-hf/LLaVA-NeXT-Video-7B-hf"],
......@@ -563,7 +581,9 @@ VLM_TEST_SETTINGS = {
img_idx_to_prompt=lambda idx: "(<image>./</image>)\n",
max_model_len=4096,
max_num_seqs=2,
get_stop_token_ids=lambda tok: tok.convert_tokens_to_ids(['<|im_end|>', '<|endoftext|>']), # noqa: E501
get_stop_token_ids=lambda tok: tok.convert_tokens_to_ids(
["<|im_end|>", "<|endoftext|>"]
), # noqa: E501
hf_output_post_proc=model_utils.minicpmv_trunc_hf_output,
patch_hf_runner=model_utils.minicpmo_26_patch_hf_runner,
# FIXME: https://huggingface.co/openbmb/MiniCPM-o-2_6/discussions/49
......@@ -576,13 +596,15 @@ VLM_TEST_SETTINGS = {
img_idx_to_prompt=lambda idx: "(<image>./</image>)\n",
max_model_len=4096,
max_num_seqs=2,
get_stop_token_ids=lambda tok: tok.convert_tokens_to_ids(['<|im_end|>', '<|endoftext|>']), # noqa: E501
get_stop_token_ids=lambda tok: tok.convert_tokens_to_ids(
["<|im_end|>", "<|endoftext|>"]
), # noqa: E501
hf_output_post_proc=model_utils.minicpmv_trunc_hf_output,
patch_hf_runner=model_utils.minicpmv_26_patch_hf_runner,
),
"minimax_vl_01": VLMTestInfo(
models=["MiniMaxAI/MiniMax-VL-01"],
prompt_formatter=lambda img_prompt: f"<beginning_of_sentence>user: {img_prompt} assistant:<end_of_sentence>", # noqa: E501
prompt_formatter=lambda img_prompt: f"<beginning_of_sentence>user: {img_prompt} assistant:<end_of_sentence>", # noqa: E501
img_idx_to_prompt=lambda _: "<image>",
test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
max_model_len=8192,
......@@ -604,8 +626,8 @@ VLM_TEST_SETTINGS = {
"ovis1_6-gemma2": VLMTestInfo(
models=["AIDC-AI/Ovis1.6-Gemma2-9B"],
test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
prompt_formatter=lambda img_prompt: f"<bos><start_of_turn>user\n{img_prompt}<end_of_turn>\n<start_of_turn>model\n", # noqa: E501
img_idx_to_prompt=lambda idx: "<image>\n", # noqa: E501
prompt_formatter=lambda img_prompt: f"<bos><start_of_turn>user\n{img_prompt}<end_of_turn>\n<start_of_turn>model\n", # noqa: E501
img_idx_to_prompt=lambda idx: "<image>\n", # noqa: E501
max_model_len=4096,
max_num_seqs=2,
dtype="half",
......@@ -617,8 +639,8 @@ VLM_TEST_SETTINGS = {
"ovis2": VLMTestInfo(
models=["AIDC-AI/Ovis2-1B"],
test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
prompt_formatter=lambda img_prompt: f"<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\n{img_prompt}<|im_end|>\n<|im_start|>assistant\n", # noqa: E501
img_idx_to_prompt=lambda idx: "<image>\n", # noqa: E501
prompt_formatter=lambda img_prompt: f"<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\n{img_prompt}<|im_end|>\n<|im_start|>assistant\n", # noqa: E501
img_idx_to_prompt=lambda idx: "<image>\n", # noqa: E501
max_model_len=4096,
max_num_seqs=2,
dtype="half",
......@@ -628,13 +650,9 @@ VLM_TEST_SETTINGS = {
),
"ovis2_5": VLMTestInfo(
models=["AIDC-AI/Ovis2.5-2B"],
test_type=(
VLMTestType.IMAGE,
VLMTestType.MULTI_IMAGE,
VLMTestType.VIDEO
),
prompt_formatter=lambda img_prompt: f"<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\n{img_prompt}<|im_end|>\n<|im_start|>assistant\n", # noqa: E501
img_idx_to_prompt=lambda idx: "<image>\n", # noqa: E501
test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE, VLMTestType.VIDEO),
prompt_formatter=lambda img_prompt: f"<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\n{img_prompt}<|im_end|>\n<|im_start|>assistant\n", # noqa: E501
img_idx_to_prompt=lambda idx: "<image>\n", # noqa: E501
video_idx_to_prompt=lambda idx: "<video>\n",
max_model_len=4096,
max_num_seqs=2,
......@@ -646,7 +664,7 @@ VLM_TEST_SETTINGS = {
"phi3v": VLMTestInfo(
models=["microsoft/Phi-3.5-vision-instruct"],
test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
prompt_formatter=lambda img_prompt: f"<|user|>\n{img_prompt}<|end|>\n<|assistant|>\n", # noqa: E501
prompt_formatter=lambda img_prompt: f"<|user|>\n{img_prompt}<|end|>\n<|assistant|>\n", # noqa: E501
img_idx_to_prompt=lambda idx: f"<|image_{idx}|>\n",
max_model_len=4096,
max_num_seqs=2,
......@@ -681,15 +699,11 @@ VLM_TEST_SETTINGS = {
),
"qwen2_vl": VLMTestInfo(
models=["Qwen/Qwen2-VL-2B-Instruct"],
test_type=(
VLMTestType.IMAGE,
VLMTestType.MULTI_IMAGE,
VLMTestType.VIDEO
),
prompt_formatter=lambda img_prompt: f"<|im_start|>User\n{img_prompt}<|im_end|>\n<|im_start|>assistant\n", # noqa: E501
img_idx_to_prompt=lambda idx: "<|vision_start|><|image_pad|><|vision_end|>", # noqa: E501
video_idx_to_prompt=lambda idx: "<|vision_start|><|video_pad|><|vision_end|>", # noqa: E501
multi_image_prompt="Picture 1: <vlm_image>\nPicture 2: <vlm_image>\nDescribe these two images with one paragraph respectively.", # noqa: E501
test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE, VLMTestType.VIDEO),
prompt_formatter=lambda img_prompt: f"<|im_start|>User\n{img_prompt}<|im_end|>\n<|im_start|>assistant\n", # noqa: E501
img_idx_to_prompt=lambda idx: "<|vision_start|><|image_pad|><|vision_end|>", # noqa: E501
video_idx_to_prompt=lambda idx: "<|vision_start|><|video_pad|><|vision_end|>", # noqa: E501
multi_image_prompt="Picture 1: <vlm_image>\nPicture 2: <vlm_image>\nDescribe these two images with one paragraph respectively.", # noqa: E501
max_model_len=4096,
max_num_seqs=2,
auto_cls=AutoModelForImageTextToText,
......@@ -700,11 +714,13 @@ VLM_TEST_SETTINGS = {
"skywork_r1v": VLMTestInfo(
models=["Skywork/Skywork-R1V-38B"],
test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
prompt_formatter=lambda img_prompt: f"<｜begin▁of▁sentence｜><｜User｜>\n{img_prompt}<｜Assistant｜><think>\n", # noqa: E501
single_image_prompts=IMAGE_ASSETS.prompts({
"stop_sign": "<image>\nWhat's the content in the center of the image?", # noqa: E501
"cherry_blossom": "<image>\nWhat is the season?",
}),
prompt_formatter=lambda img_prompt: f"<｜begin▁of▁sentence｜><｜User｜>\n{img_prompt}<｜Assistant｜><think>\n", # noqa: E501
single_image_prompts=IMAGE_ASSETS.prompts(
{
"stop_sign": "<image>\nWhat's the content in the center of the image?", # noqa: E501
"cherry_blossom": "<image>\nWhat is the season?",
}
),
multi_image_prompt="<image>\n<image>\nDescribe the two images in short.", # noqa: E501
max_model_len=4096,
use_tokenizer_eos=True,
......@@ -737,9 +753,9 @@ VLM_TEST_SETTINGS = {
VLMTestType.MULTI_IMAGE,
VLMTestType.VIDEO,
),
prompt_formatter=lambda img_prompt: f"<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\n{img_prompt}<|im_end|>\n<|im_start|>assistant\n", # noqa: E501
img_idx_to_prompt=lambda idx: "<|vision_start|><|image_pad|><|vision_end|>", # noqa: E501
video_idx_to_prompt=lambda idx: "<|vision_start|><|video_pad|><|vision_end|>", # noqa: E501
prompt_formatter=lambda img_prompt: f"<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\n{img_prompt}<|im_end|>\n<|im_start|>assistant\n", # noqa: E501
img_idx_to_prompt=lambda idx: "<|vision_start|><|image_pad|><|vision_end|>", # noqa: E501
video_idx_to_prompt=lambda idx: "<|vision_start|><|video_pad|><|vision_end|>", # noqa: E501
max_model_len=4096,
max_num_seqs=2,
auto_cls=AutoModelForImageTextToText,
......@@ -752,11 +768,11 @@ VLM_TEST_SETTINGS = {
prompt_formatter=lambda img_prompt: f"USER: {img_prompt}\nASSISTANT:",
max_model_len=4096,
auto_cls=AutoModelForImageTextToText,
vllm_output_post_proc = lambda vllm_output, model: vllm_output[:2],
hf_output_post_proc = lambda hf_output, model: hf_output[:2],
vllm_output_post_proc=lambda vllm_output, model: vllm_output[:2],
hf_output_post_proc=lambda hf_output, model: hf_output[:2],
comparator=check_outputs_equal,
marks=multi_gpu_marks(num_gpus=2),
**COMMON_BROADCAST_SETTINGS # type: ignore
**COMMON_BROADCAST_SETTINGS, # type: ignore
),
"llava-broadcast": VLMTestInfo(
models=["llava-hf/llava-1.5-7b-hf"],
......@@ -765,7 +781,7 @@ VLM_TEST_SETTINGS = {
auto_cls=AutoModelForImageTextToText,
vllm_output_post_proc=model_utils.llava_image_vllm_to_hf_output,
marks=multi_gpu_marks(num_gpus=2),
**COMMON_BROADCAST_SETTINGS # type: ignore
**COMMON_BROADCAST_SETTINGS, # type: ignore
),
"llava_next-broadcast": VLMTestInfo(
models=["llava-hf/llava-v1.6-mistral-7b-hf"],
......@@ -774,12 +790,12 @@ VLM_TEST_SETTINGS = {
auto_cls=AutoModelForImageTextToText,
vllm_output_post_proc=model_utils.llava_image_vllm_to_hf_output,
marks=multi_gpu_marks(num_gpus=2),
**COMMON_BROADCAST_SETTINGS # type: ignore
**COMMON_BROADCAST_SETTINGS, # type: ignore
),
### Custom input edge-cases for specific models
"intern_vl-diff-patches": VLMTestInfo(
models=["OpenGVLab/InternVL2-2B"],
prompt_formatter=lambda img_prompt: f"<|im_start|>User\n{img_prompt}<|im_end|>\n<|im_start|>Assistant\n", # noqa: E501
prompt_formatter=lambda img_prompt: f"<|im_start|>User\n{img_prompt}<|im_end|>\n<|im_start|>Assistant\n", # noqa: E501
test_type=VLMTestType.CUSTOM_INPUTS,
max_model_len=4096,
use_tokenizer_eos=True,
......@@ -788,7 +804,8 @@ VLM_TEST_SETTINGS = {
CustomTestOptions(
inputs=inp,
limit_mm_per_prompt={"image": 2},
) for inp in custom_inputs.different_patch_input_cases_internvl()
)
for inp in custom_inputs.different_patch_input_cases_internvl()
],
),
"llava_onevision-multiple-images": VLMTestInfo(
......@@ -797,14 +814,18 @@ VLM_TEST_SETTINGS = {
max_model_len=16384,
max_num_seqs=2,
auto_cls=AutoModelForImageTextToText,
hf_model_kwargs=model_utils.llava_onevision_hf_model_kwargs("llava-hf/llava-onevision-qwen2-0.5b-ov-hf"), # noqa: E501
hf_model_kwargs=model_utils.llava_onevision_hf_model_kwargs(
"llava-hf/llava-onevision-qwen2-0.5b-ov-hf"
), # noqa: E501
vllm_output_post_proc=model_utils.llava_onevision_vllm_to_hf_output,
custom_test_opts=[CustomTestOptions(
inputs=custom_inputs.multi_image_multi_aspect_ratio_inputs(
formatter=lambda vid_prompt: f"<|im_start|>user\n{vid_prompt}<|im_end|>\n<|im_start|>assistant\n", # noqa: E501
),
limit_mm_per_prompt={"image": 4},
)],
custom_test_opts=[
CustomTestOptions(
inputs=custom_inputs.multi_image_multi_aspect_ratio_inputs(
formatter=lambda vid_prompt: f"<|im_start|>user\n{vid_prompt}<|im_end|>\n<|im_start|>assistant\n", # noqa: E501
),
limit_mm_per_prompt={"image": 4},
)
],
),
# regression test for https://github.com/vllm-project/vllm/issues/15122
"qwen2_5_vl-windows-attention": VLMTestInfo(
......@@ -814,13 +835,14 @@ VLM_TEST_SETTINGS = {
max_num_seqs=2,
auto_cls=AutoModelForImageTextToText,
vllm_output_post_proc=model_utils.qwen2_vllm_to_hf_output,
custom_test_opts=[CustomTestOptions(
inputs=custom_inputs.windows_attention_image_qwen2_5_vl(),
limit_mm_per_prompt={"image": 1},
)],
custom_test_opts=[
CustomTestOptions(
inputs=custom_inputs.windows_attention_image_qwen2_5_vl(),
limit_mm_per_prompt={"image": 1},
)
],
),
}
# yapf: enable
def _mark_splits(
......
tests/models/multimodal/generation/vlm_utils/case_filtering.py
View file @ 4e256cad
......@@ -114,7 +114,6 @@ def get_parametrized_options(
raise ValueError("Test has type CUSTOM_INPUTS, but none given")
iter_kwargs["custom_test_opts"] = test_info.custom_test_opts
# yapf: disable
# Wrap all model cases in a pytest parameter & pass marks through
return [
pytest.param(
......@@ -122,10 +121,10 @@ def get_parametrized_options(
ExpandableVLMTestArgs(
**{k: v for k, v in zip(iter_kwargs.keys(), case)}
),
marks=test_info.marks if test_info.marks is not None else []
) for case in list(itertools.product(*iter_kwargs.values()))
marks=test_info.marks if test_info.marks is not None else [],
)
for case in list(itertools.product(*iter_kwargs.values()))
]
# yapf: enable
# Get a list per model type, where each entry contains a tuple of all of
# that model type's cases, then flatten them into the top level so that
......
tests/models/multimodal/generation/vlm_utils/model_utils.py
View file @ 4e256cad
......@@ -418,7 +418,6 @@ def h2ovl_patch_hf_runner(hf_model: HfRunner) -> HfRunner:
self.image_size = self.vision_config.image_size
def __call__(self, text: str, images: Union[Image, list[Image]], **kwargs):
# yapf: disable
from vllm.model_executor.models.h2ovl import (
IMG_CONTEXT,
IMG_END,
......@@ -426,7 +425,6 @@ def h2ovl_patch_hf_runner(hf_model: HfRunner) -> HfRunner:
image_to_pixel_values_h2ovl,
)
# yapf: enable
images = [images] if isinstance(images, Image) else images
pixel_values = [
image_to_pixel_values_h2ovl(
......
tests/models/multimodal/generation/vlm_utils/types.py
View file @ 4e256cad
......@@ -33,24 +33,26 @@ TEST_IMG_PLACEHOLDER = "<vlm_image>"
TEST_VIDEO_PLACEHOLDER = "<vlm_video>"
TEST_AUDIO_PLACEHOLDER = "<lmm_audio>"
# yapf: disable
SINGLE_IMAGE_BASE_PROMPTS = IMAGE_ASSETS.prompts({
"stop_sign": f"{TEST_IMG_PLACEHOLDER}What's the content of the image?",
"cherry_blossom": f"{TEST_IMG_PLACEHOLDER}What is the season?",
})
SINGLE_AUDIO_BASE_PROMPT = AUDIO_ASSETS.prompts({
"mary_had_lamb": f"{TEST_AUDIO_PLACEHOLDER}Transcribe this audio into English.", # noqa: E501
"winning_call": f"{TEST_AUDIO_PLACEHOLDER}What is happening in this audio clip?", # noqa: E501
})
SINGLE_IMAGE_BASE_PROMPTS = IMAGE_ASSETS.prompts(
{
"stop_sign": f"{TEST_IMG_PLACEHOLDER}What's the content of the image?",
"cherry_blossom": f"{TEST_IMG_PLACEHOLDER}What is the season?",
}
)
SINGLE_AUDIO_BASE_PROMPT = AUDIO_ASSETS.prompts(
{
"mary_had_lamb": f"{TEST_AUDIO_PLACEHOLDER}Transcribe this audio into English.", # noqa: E501
"winning_call": f"{TEST_AUDIO_PLACEHOLDER}What is happening in this audio clip?", # noqa: E501
}
)
MULTI_IMAGE_BASE_PROMPT = f"Image-1: {TEST_IMG_PLACEHOLDER}Image-2: {TEST_IMG_PLACEHOLDER}Describe the two images in detail.\n" # noqa: E501
VIDEO_BASE_PROMPT = f"{TEST_VIDEO_PLACEHOLDER}Why is this video funny?"
IMAGE_SIZE_FACTORS = [(), (1.0, ), (1.0, 1.0, 1.0), (0.25, 0.5, 1.0)]
EMBEDDING_SIZE_FACTORS = [(), (1.0, ), (1.0, 1.0, 1.0)]
IMAGE_SIZE_FACTORS = [(), (1.0,), (1.0, 1.0, 1.0), (0.25, 0.5, 1.0)]
EMBEDDING_SIZE_FACTORS = [(), (1.0,), (1.0, 1.0, 1.0)]
RunnerOutput = tuple[list[int], str, Optional[SampleLogprobs]]
# yapf: enable
class PromptWithMultiModalInput(NamedTuple):
......
tests/models/multimodal/processing/test_common.py
View file @ 4e256cad
......@@ -322,80 +322,81 @@ def _test_processing_correctness_one(
)
# yapf: disable
@pytest.mark.parametrize("model_id", [
"rhymes-ai/Aria",
"CohereForAI/aya-vision-8b",
"Salesforce/blip2-opt-2.7b",
"facebook/chameleon-7b",
"CohereLabs/command-a-vision-07-2025",
"deepseek-ai/deepseek-vl2-tiny",
"baidu/ERNIE-4.5-VL-28B-A3B-PT",
"adept/fuyu-8b",
"google/gemma-3-4b-it",
"google/gemma-3n-E2B-it",
"zai-org/glm-4v-9b",
"zai-org/GLM-4.1V-9B-Thinking",
"zai-org/GLM-4.5V",
"ibm-granite/granite-speech-3.3-2b",
"h2oai/h2ovl-mississippi-800m",
"naver-hyperclovax/HyperCLOVAX-SEED-Vision-Instruct-3B",
"HuggingFaceM4/Idefics3-8B-Llama3",
"internlm/Intern-S1",
"OpenGVLab/InternVL2-1B",
"OpenGVLab/InternVL3-1B",
"OpenGVLab/InternVL3_5-1B",
"OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview",
"OpenGVLab/InternVL3_5-30B-A3B",
"Kwai-Keye/Keye-VL-8B-Preview",
"Kwai-Keye/Keye-VL-1_5-8B",
"moonshotai/Kimi-VL-A3B-Instruct",
"meta-llama/Llama-4-Scout-17B-16E-Instruct",
"llava-hf/llava-1.5-7b-hf",
"llava-hf/llava-v1.6-mistral-7b-hf",
"llava-hf/LLaVA-NeXT-Video-7B-hf",
"llava-hf/llava-onevision-qwen2-0.5b-ov-hf",
"TIGER-Lab/Mantis-8B-siglip-llama3",
"mispeech/midashenglm-7b",
"openbmb/MiniCPM-Llama3-V-2_5",
"openbmb/MiniCPM-o-2_6",
"openbmb/MiniCPM-V-2_6",
"MiniMaxAI/MiniMax-VL-01",
"allenai/Molmo-7B-D-0924",
"allenai/Molmo-7B-O-0924",
"nvidia/NVLM-D-72B",
"nvidia/Llama-3.1-Nemotron-Nano-VL-8B-V1",
"AIDC-AI/Ovis1.6-Gemma2-9B",
"AIDC-AI/Ovis1.6-Llama3.2-3B",
"AIDC-AI/Ovis2-1B",
"AIDC-AI/Ovis2.5-2B",
"google/paligemma-3b-mix-224",
"google/paligemma2-3b-ft-docci-448",
"microsoft/Phi-3.5-vision-instruct",
"microsoft/Phi-4-multimodal-instruct",
"mistralai/Pixtral-12B-2409",
"mistral-community/pixtral-12b",
"Qwen/Qwen-VL-Chat",
"Qwen/Qwen2-VL-2B-Instruct",
"Qwen/Qwen2.5-VL-3B-Instruct",
"Qwen/Qwen2-Audio-7B-Instruct",
"Qwen/Qwen2.5-Omni-3B",
"Qwen/Qwen3-VL-4B-Instruct",
"Qwen/Qwen3-VL-30B-A3B-Instruct",
"YannQi/R-4B",
"Skywork/Skywork-R1V-38B",
"HuggingFaceTB/SmolVLM2-2.2B-Instruct",
"stepfun-ai/step3",
"fixie-ai/ultravox-v0_5-llama-3_2-1b",
"openai/whisper-large-v3",
"omni-research/Tarsier-7b",
"omni-research/Tarsier2-Recap-7b",
"mistralai/Voxtral-Mini-3B-2507",
])
@pytest.mark.parametrize(
"model_id",
[
"rhymes-ai/Aria",
"CohereForAI/aya-vision-8b",
"Salesforce/blip2-opt-2.7b",
"facebook/chameleon-7b",
"CohereLabs/command-a-vision-07-2025",
"deepseek-ai/deepseek-vl2-tiny",
"baidu/ERNIE-4.5-VL-28B-A3B-PT",
"adept/fuyu-8b",
"google/gemma-3-4b-it",
"google/gemma-3n-E2B-it",
"zai-org/glm-4v-9b",
"zai-org/GLM-4.1V-9B-Thinking",
"zai-org/GLM-4.5V",
"ibm-granite/granite-speech-3.3-2b",
"h2oai/h2ovl-mississippi-800m",
"naver-hyperclovax/HyperCLOVAX-SEED-Vision-Instruct-3B",
"HuggingFaceM4/Idefics3-8B-Llama3",
"internlm/Intern-S1",
"OpenGVLab/InternVL2-1B",
"OpenGVLab/InternVL3-1B",
"OpenGVLab/InternVL3_5-1B",
"OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview",
"OpenGVLab/InternVL3_5-30B-A3B",
"Kwai-Keye/Keye-VL-8B-Preview",
"Kwai-Keye/Keye-VL-1_5-8B",
"moonshotai/Kimi-VL-A3B-Instruct",
"meta-llama/Llama-4-Scout-17B-16E-Instruct",
"llava-hf/llava-1.5-7b-hf",
"llava-hf/llava-v1.6-mistral-7b-hf",
"llava-hf/LLaVA-NeXT-Video-7B-hf",
"llava-hf/llava-onevision-qwen2-0.5b-ov-hf",
"TIGER-Lab/Mantis-8B-siglip-llama3",
"mispeech/midashenglm-7b",
"openbmb/MiniCPM-Llama3-V-2_5",
"openbmb/MiniCPM-o-2_6",
"openbmb/MiniCPM-V-2_6",
"MiniMaxAI/MiniMax-VL-01",
"allenai/Molmo-7B-D-0924",
"allenai/Molmo-7B-O-0924",
"nvidia/NVLM-D-72B",
"nvidia/Llama-3.1-Nemotron-Nano-VL-8B-V1",
"AIDC-AI/Ovis1.6-Gemma2-9B",
"AIDC-AI/Ovis1.6-Llama3.2-3B",
"AIDC-AI/Ovis2-1B",
"AIDC-AI/Ovis2.5-2B",
"google/paligemma-3b-mix-224",
"google/paligemma2-3b-ft-docci-448",
"microsoft/Phi-3.5-vision-instruct",
"microsoft/Phi-4-multimodal-instruct",
"mistralai/Pixtral-12B-2409",
"mistral-community/pixtral-12b",
"Qwen/Qwen-VL-Chat",
"Qwen/Qwen2-VL-2B-Instruct",
"Qwen/Qwen2.5-VL-3B-Instruct",
"Qwen/Qwen2-Audio-7B-Instruct",
"Qwen/Qwen2.5-Omni-3B",
"Qwen/Qwen3-VL-4B-Instruct",
"Qwen/Qwen3-VL-30B-A3B-Instruct",
"YannQi/R-4B",
"Skywork/Skywork-R1V-38B",
"HuggingFaceTB/SmolVLM2-2.2B-Instruct",
"stepfun-ai/step3",
"fixie-ai/ultravox-v0_5-llama-3_2-1b",
"openai/whisper-large-v3",
"omni-research/Tarsier-7b",
"omni-research/Tarsier2-Recap-7b",
"mistralai/Voxtral-Mini-3B-2507",
],
)
@pytest.mark.parametrize("hit_rate", [0.3, 0.5, 1.0])
@pytest.mark.parametrize("num_batches", [32])
@pytest.mark.parametrize("simplify_rate", [1.0])
# yapf: enable
def test_processing_correctness(
model_id: str,
hit_rate: float,
......
tests/models/multimodal/processing/test_idefics3.py
View file @ 4e256cad
......@@ -12,7 +12,6 @@ from ...utils import build_model_context
@pytest.mark.parametrize("model_id", ["HuggingFaceM4/Idefics3-8B-Llama3"])
# yapf: disable
@pytest.mark.parametrize(
("mm_processor_kwargs", "expected_toks_per_img"),
[
......@@ -20,7 +19,6 @@ from ...utils import build_model_context
({"size": {"longest_edge": 728}}, 169 * (2**2 + 1)),
],
)
# yapf: enable
@pytest.mark.parametrize("num_imgs", [1, 2])
@pytest.mark.parametrize("kwargs_on_init", [True, False])
def test_processor_override(
......
tests/models/multimodal/processing/test_phi3v.py
View file @ 4e256cad
......@@ -11,7 +11,6 @@ from ...utils import build_model_context
@pytest.mark.parametrize("model_id", ["microsoft/Phi-3.5-vision-instruct"])
# yapf: disable
@pytest.mark.parametrize(
("mm_processor_kwargs", "expected_toks_per_img"),
[
......@@ -21,7 +20,6 @@ from ...utils import build_model_context
({}, 757),
],
)
# yapf: enable
@pytest.mark.parametrize("num_imgs", [1, 2])
@pytest.mark.parametrize("kwargs_on_init", [True, False])
def test_processor_override(
......
tests/models/multimodal/processing/test_phi4mm.py
View file @ 4e256cad
......@@ -11,7 +11,6 @@ from ...utils import build_model_context
@pytest.mark.parametrize("model_id", ["microsoft/Phi-4-multimodal-instruct"])
# yapf: disable
@pytest.mark.parametrize(
("mm_processor_kwargs", "expected_toks_per_img"),
[
......@@ -21,7 +20,6 @@ from ...utils import build_model_context
({}, 9585),
],
)
# yapf: enable
@pytest.mark.parametrize("num_imgs", [1, 2])
@pytest.mark.parametrize("kwargs_on_init", [True, False])
def test_processor_override(
......
tests/models/multimodal/processing/test_qwen2_vl.py
View file @ 4e256cad
......@@ -10,7 +10,6 @@ from ...utils import build_model_context
@pytest.mark.parametrize("model_id", ["Qwen/Qwen2-VL-2B-Instruct"])
# yapf: disable
@pytest.mark.parametrize(
("mm_processor_kwargs", "expected_toks_per_img", "expected_pixels_shape"),
[
......@@ -18,7 +17,6 @@ from ...utils import build_model_context
({"min_pixels": 64**2, "max_pixels": 512**2}, 330, (1320, 1176)),
],
)
# yapf: enable
@pytest.mark.parametrize("num_imgs", [1, 2])
@pytest.mark.parametrize("kwargs_on_init", [True, False])
def test_processor_override(
......
tests/models/multimodal/processing/test_smolvlm.py
View file @ 4e256cad
......@@ -12,7 +12,6 @@ from ...utils import build_model_context
@pytest.mark.parametrize("model_id", ["HuggingFaceTB/SmolVLM2-2.2B-Instruct"])
# yapf: disable
@pytest.mark.parametrize(
("mm_processor_kwargs", "expected_toks_per_img"),
[
......@@ -20,7 +19,6 @@ from ...utils import build_model_context
({"max_image_size": {"longest_edge": 768}}, 405),
],
)
# yapf: enable
@pytest.mark.parametrize("num_imgs", [1, 2])
@pytest.mark.parametrize("kwargs_on_init", [True, False])
def test_processor_override(
......
tests/models/multimodal/processing/test_transformers.py
View file @ 4e256cad
......@@ -7,9 +7,7 @@ from vllm.config import ModelConfig
from vllm.multimodal import MULTIMODAL_REGISTRY
# yapf: disable
@pytest.mark.parametrize("model_id",
["llava-hf/llava-onevision-qwen2-0.5b-ov-hf"])
@pytest.mark.parametrize("model_id", ["llava-hf/llava-onevision-qwen2-0.5b-ov-hf"])
def test_multimodal_processor(model_id):
model_config = ModelConfig(
model=model_id,
......@@ -18,9 +16,9 @@ def test_multimodal_processor(model_id):
mm_processor = MULTIMODAL_REGISTRY.create_processor(model_config)
image_pil = ImageAsset('cherry_blossom').pil_image
image_pil = ImageAsset("cherry_blossom").pil_image
mm_data = {"image": image_pil}
str_prompt = "<|im_start|>user <image>\nWhat is the content of this image?<|im_end|><|im_start|>assistant\n" # noqa: E501
str_prompt = "<|im_start|>user <image>\nWhat is the content of this image?<|im_end|><|im_start|>assistant\n" # noqa: E501
str_processed_inputs = mm_processor.apply(
prompt=str_prompt,
mm_data=mm_data,
......@@ -28,8 +26,23 @@ def test_multimodal_processor(model_id):
)
ids_prompt = [
151644, 872, 220, 151646, 198, 3838, 374, 279, 2213, 315, 419, 2168,
30, 151645, 151644, 77091, 198
151644,
872,
220,
151646,
198,
3838,
374,
279,
2213,
315,
419,
2168,
30,
151645,
151644,
77091,
198,
]
ids_processed_inputs = mm_processor.apply(
prompt=ids_prompt,
......@@ -37,5 +50,7 @@ def test_multimodal_processor(model_id):
hf_processor_mm_kwargs={},
)
assert (str_processed_inputs["prompt_token_ids"]
== ids_processed_inputs["prompt_token_ids"])
assert (
str_processed_inputs["prompt_token_ids"]
== ids_processed_inputs["prompt_token_ids"]
)
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment