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

Convert formatting to use ruff instead of yapf + isort (#26247)

Browse Source 
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
Harry Mellor
2025-10-05 15:06:22 +01:00
committed by GitHub
parent 17edd8a807
commit d6953beb91
1508 changed files with 115244 additions and 94146 deletions
Download Patch File Download Diff File Expand all files Collapse all files

35
tests/models/multimodal/pooling/test_clip.py
View File

@@ -12,10 +12,12 @@ HF_TEXT_PROMPTS = [
"a photo of a cherry blossom",
]
HF_IMAGE_PROMPTS = IMAGE_ASSETS.prompts({
"stop_sign": "",
"cherry_blossom": "",
})
HF_IMAGE_PROMPTS = IMAGE_ASSETS.prompts(
{
"stop_sign": "",
"cherry_blossom": "",
}
)
MODELS = ["openai/clip-vit-base-patch32"]
@@ -33,11 +35,9 @@ def _run_test(
# vLLM needs a fresh new process without cuda initialization.
# if we run HF first, the cuda initialization will be done and it
# will hurt multiprocessing backend with fork method (the default method).
with vllm_runner(model,
runner="pooling",
dtype=dtype,
enforce_eager=True,
max_model_len=77) as vllm_model:
with vllm_runner(
model, runner="pooling", dtype=dtype, enforce_eager=True, max_model_len=77
) as vllm_model:
vllm_outputs = vllm_model.embed(input_texts, images=input_images)
with hf_runner(model, dtype=dtype, auto_cls=CLIPModel) as hf_model:
@@ -48,10 +48,12 @@ def _run_test(
if "pixel_values" in inputs:
inputs.pop("input_ids")
pooled_output = hf_model.model.get_image_features(
**hf_model.wrap_device(inputs)).squeeze(0)
**hf_model.wrap_device(inputs)
).squeeze(0)
else:
pooled_output = hf_model.model.get_text_features(
**hf_model.wrap_device(inputs)).squeeze(0)
**hf_model.wrap_device(inputs)
).squeeze(0)
all_outputs.append(pooled_output.tolist())
@@ -98,8 +100,7 @@ def test_models_image(
dtype: str,
) -> None:
input_texts_images = [
(text, asset.pil_image)
for text, asset in zip(HF_IMAGE_PROMPTS, image_assets)
(text, asset.pil_image) for text, asset in zip(HF_IMAGE_PROMPTS, image_assets)
]
input_texts = [text for text, _ in input_texts_images]
input_images = [image for _, image in input_texts_images]
@@ -125,11 +126,9 @@ def test_models_text_image_no_crash(
texts = [HF_TEXT_PROMPTS[0]]
images = [image_assets[0].pil_image]
with vllm_runner(model,
runner="pooling",
dtype=dtype,
enforce_eager=True,
max_model_len=77) as vllm_model:
with vllm_runner(
model, runner="pooling", dtype=dtype, enforce_eager=True, max_model_len=77
) as vllm_model:
with pytest.raises(ValueError, match="not both"):
vllm_model.embed(texts, images=images)

117
tests/models/multimodal/pooling/test_dse_qwen2_vl.py
View File

@@ -17,18 +17,21 @@ HF_TEXT_PROMPTS = [
# T -> X
(
"Query: Find me an everyday image that matches the given caption: The label of the object is stop sign", # noqa: E501,
Image.new("RGB", (56, 56))),
Image.new("RGB", (56, 56)),
),
# T -> X
("Query: Retrieve an image of this caption: cherry blossom",
Image.new("RGB", (56, 56))),
(
"Query: Retrieve an image of this caption: cherry blossom",
Image.new("RGB", (56, 56)),
),
]
HF_IMAGE_PROMPTS = IMAGE_ASSETS.prompts({
"stop_sign":
"What is shown in this image?",
"cherry_blossom":
"What is shown in this image?"
})
HF_IMAGE_PROMPTS = IMAGE_ASSETS.prompts(
{
"stop_sign": "What is shown in this image?",
"cherry_blossom": "What is shown in this image?",
}
)
MODELS = ["MrLight/dse-qwen2-2b-mrl-v1"]
@@ -36,34 +39,30 @@ MODELS = ["MrLight/dse-qwen2-2b-mrl-v1"]
def get_messages(image: Image.Image, text: str, embed_text: bool):
# assert False, 'remember to use outer [] as required'
if embed_text:
messages = [{
"role":
"user",
"content": [
{
"type": "image",
"image": Image.new("RGB", (56, 56)),
"resized_height": 1,
"resized_width": 1
}, # need a dummy image here for an easier process.
{
"type": "text",
"text": text
},
]
}]
messages = [
{
"role": "user",
"content": [
{
"type": "image",
"image": Image.new("RGB", (56, 56)),
"resized_height": 1,
"resized_width": 1,
}, # need a dummy image here for an easier process.
{"type": "text", "text": text},
],
}
]
else:
messages = [{
"role":
"user",
"content": [{
"type": "image",
"image": image
}, {
"type": "text",
"text": text
}]
}]
messages = [
{
"role": "user",
"content": [
{"type": "image", "image": image},
{"type": "text", "text": text},
],
}
]
return messages
@@ -71,8 +70,10 @@ def apply_chat_template_and_add_eos(
messages: list[dict],
apply_chat_template_fn: Callable,
):
prompt = apply_chat_template_fn(
messages, tokenize=False, add_generation_prompt=True) + "<|endoftext|>"
prompt = (
apply_chat_template_fn(messages, tokenize=False, add_generation_prompt=True)
+ "<|endoftext|>"
)
return prompt
@@ -86,16 +87,14 @@ def _run_test(
*,
dtype: str,
) -> None:
'''SET PYTHONPATH'''
"""SET PYTHONPATH"""
# NOTE: take care of the order. run vLLM first, and then run HF.
# vLLM needs a fresh new process without cuda initialization.
# if we run HF first, the cuda initialization will be done and it
# will hurt multiprocessing backend with fork method (the default method).
with vllm_runner(model,
runner="pooling",
dtype=dtype,
enforce_eager=True,
max_model_len=8192) as vllm_model:
with vllm_runner(
model, runner="pooling", dtype=dtype, enforce_eager=True, max_model_len=8192
) as vllm_model:
tokenizer = vllm_model.llm.get_tokenizer()
texts = [
# this is necessary because vllm_model.embed will not apply any
@@ -105,25 +104,25 @@ def _run_test(
apply_chat_template_and_add_eos(
get_messages(image, text, False),
apply_chat_template_fn=tokenizer.apply_chat_template,
) for text, image in zip(input_texts, input_images)
)
for text, image in zip(input_texts, input_images)
# vllm will replace the pad token with the actual image,
# which may be a placeholder image, later.
]
vllm_outputs = vllm_model.embed(texts, images=input_images)
hf_outputs = []
with hf_runner(model,
dtype=dtype,
auto_cls=Qwen2VLForConditionalGeneration) as hf_model:
with hf_runner(
model, dtype=dtype, auto_cls=Qwen2VLForConditionalGeneration
) as hf_model:
prompts = []
for text, image, embed_text in zip(input_texts, input_images,
embed_texts):
for text, image, embed_text in zip(input_texts, input_images, embed_texts):
# dse requires non-standard input processing
# because it needs an image_pad token
messages = get_messages(image, text, embed_text)
prompt = apply_chat_template_and_add_eos(
messages, hf_model.processor.apply_chat_template)
messages, hf_model.processor.apply_chat_template
)
prompts.append(prompt)
@@ -145,9 +144,9 @@ def _run_test(
return_dict=True,
output_hidden_states=True,
)
pooled_output = F.normalize(outputs.hidden_states[-1][0, -1],
p=2,
dim=-1)
pooled_output = F.normalize(
outputs.hidden_states[-1][0, -1], p=2, dim=-1
)
all_outputs.append(pooled_output.tolist())
@@ -170,8 +169,9 @@ def test_models_text(
model: str,
dtype: str,
) -> None:
input_texts_images = [(text, image_placeholder)
for text, image_placeholder in HF_TEXT_PROMPTS]
input_texts_images = [
(text, image_placeholder) for text, image_placeholder in HF_TEXT_PROMPTS
]
input_texts = [text for text, _ in input_texts_images]
input_images = [image for _, image in input_texts_images]
embed_texts = [True] * len(input_texts)
@@ -198,8 +198,7 @@ def test_models_image(
dtype: str,
) -> None:
input_texts_images = [
(text, asset.pil_image)
for text, asset in zip(HF_IMAGE_PROMPTS, image_assets)
(text, asset.pil_image) for text, asset in zip(HF_IMAGE_PROMPTS, image_assets)
]
input_texts = [text for text, _ in input_texts_images]
input_images = [image for _, image in input_texts_images]

26
tests/models/multimodal/pooling/test_intern_vit.py
View File

@@ -29,7 +29,7 @@ def run_intern_vit_test(
img_processor = CLIPImageProcessor.from_pretrained(model)
images = [asset.pil_image for asset in image_assets]
pixel_values = [
img_processor(images, return_tensors='pt').pixel_values.to(torch_dtype)
img_processor(images, return_tensors="pt").pixel_values.to(torch_dtype)
for images in images
]
@@ -37,15 +37,16 @@ def run_intern_vit_test(
if not getattr(config, "norm_type", None):
config.norm_type = "rms_norm"
hf_model = AutoModel.from_pretrained(model,
torch_dtype=torch_dtype,
trust_remote_code=True).to("cuda")
hf_model = AutoModel.from_pretrained(
model, torch_dtype=torch_dtype, trust_remote_code=True
).to("cuda")
hf_outputs_per_image = [
hf_model(pixel_value.to("cuda")).last_hidden_state
for pixel_value in pixel_values
]
from vllm.model_executor.models.intern_vit import InternVisionModel
vllm_model = InternVisionModel(config)
vllm_model.load_weights(hf_model.state_dict().items())
@@ -54,22 +55,23 @@ def run_intern_vit_test(
vllm_model = vllm_model.to("cuda", torch_dtype)
vllm_outputs_per_image = [
vllm_model(pixel_values=pixel_value.to("cuda"))
for pixel_value in pixel_values
vllm_model(pixel_values=pixel_value.to("cuda")) for pixel_value in pixel_values
]
del vllm_model
cleanup_dist_env_and_memory()
cos_similar = nn.CosineSimilarity(dim=-1)
for vllm_output, hf_output in zip(vllm_outputs_per_image,
hf_outputs_per_image):
for vllm_output, hf_output in zip(vllm_outputs_per_image, hf_outputs_per_image):
assert cos_similar(vllm_output, hf_output).mean() > 0.99
@pytest.mark.parametrize("model_id", [
"OpenGVLab/InternViT-300M-448px",
"OpenGVLab/InternViT-6B-448px-V1-5",
])
@pytest.mark.parametrize(
"model_id",
[
"OpenGVLab/InternViT-300M-448px",
"OpenGVLab/InternViT-6B-448px-V1-5",
],
)
@pytest.mark.parametrize("dtype", ["half"])
def test_models(dist_init, image_assets, model_id, dtype: str) -> None:
run_intern_vit_test(

78
tests/models/multimodal/pooling/test_jinavl_reranker.py
View File

@@ -29,7 +29,6 @@ def vllm_reranker(
query_type: str = "text",
doc_type: str = "text",
):
def create_image_param(url: str) -> ChatCompletionContentPartImageParam:
return {"type": "image_url", "image_url": {"url": f"{url}"}}
@@ -38,23 +37,25 @@ def vllm_reranker(
query = query_strs
elif query_type == "image":
query = ScoreMultiModalParam(
content=[create_image_param(url) for url in query_strs])
content=[create_image_param(url) for url in query_strs]
)
documents: Union[list[str], ScoreMultiModalParam]
if doc_type == "text":
documents = document_strs
elif doc_type == "image":
documents = ScoreMultiModalParam(
content=[create_image_param(url) for url in document_strs])
content=[create_image_param(url) for url in document_strs]
)
with vllm_runner(
model_name,
runner="pooling",
dtype=dtype,
max_num_seqs=2,
max_model_len=2048,
mm_processor_kwargs=mm_processor_kwargs,
limit_mm_per_prompt=limit_mm_per_prompt,
model_name,
runner="pooling",
dtype=dtype,
max_num_seqs=2,
max_model_len=2048,
mm_processor_kwargs=mm_processor_kwargs,
limit_mm_per_prompt=limit_mm_per_prompt,
) as vllm_model:
outputs = vllm_model.llm.score(query, documents)
@@ -78,16 +79,15 @@ def hf_reranker(
data_pairs = [[query_strs[0], d] for d in document_strs]
with hf_runner(
model_name,
dtype=dtype,
trust_remote_code=True,
auto_cls=AutoModel,
model_kwargs={"key_mapping": checkpoint_to_hf_mapper},
model_name,
dtype=dtype,
trust_remote_code=True,
auto_cls=AutoModel,
model_kwargs={"key_mapping": checkpoint_to_hf_mapper},
) as hf_model:
return hf_model.model.compute_score(data_pairs,
max_length=2048,
query_type=query_type,
doc_type=doc_type)
return hf_model.model.compute_score(
data_pairs, max_length=2048, query_type=query_type, doc_type=doc_type
)
# Visual Documents Reranking
@@ -100,10 +100,12 @@ def test_model_text_image(hf_runner, vllm_runner, model_name, dtype):
"https://raw.githubusercontent.com/jina-ai/multimodal-reranker-test/main/paper-11.png",
]
hf_outputs = hf_reranker(hf_runner, model_name, dtype, query, documents,
"text", "image")
vllm_outputs = vllm_reranker(vllm_runner, model_name, dtype, query,
documents, "text", "image")
hf_outputs = hf_reranker(
hf_runner, model_name, dtype, query, documents, "text", "image"
)
vllm_outputs = vllm_reranker(
vllm_runner, model_name, dtype, query, documents, "text", "image"
)
assert hf_outputs[0] == pytest.approx(vllm_outputs[0], rel=0.02)
assert hf_outputs[1] == pytest.approx(vllm_outputs[1], rel=0.02)
@@ -127,10 +129,12 @@ def test_model_text_text(hf_runner, vllm_runner, model_name, dtype):
lower computational requirements.""", # noqa: E501
"数据提取么？为什么不用正则啊，你用正则不就全解决了么？",
]
hf_outputs = hf_reranker(hf_runner, model_name, dtype, query, documents,
"text", "text")
vllm_outputs = vllm_reranker(vllm_runner, model_name, dtype, query,
documents, "text", "text")
hf_outputs = hf_reranker(
hf_runner, model_name, dtype, query, documents, "text", "text"
)
vllm_outputs = vllm_reranker(
vllm_runner, model_name, dtype, query, documents, "text", "text"
)
assert hf_outputs[0] == pytest.approx(vllm_outputs[0], rel=0.02)
assert hf_outputs[1] == pytest.approx(vllm_outputs[1], rel=0.02)
@@ -157,10 +161,12 @@ def test_model_image_text(hf_runner, vllm_runner, model_name, dtype):
"数据提取么？为什么不用正则啊，你用正则不就全解决了么？",
]
hf_outputs = hf_reranker(hf_runner, model_name, dtype, query, documents,
"image", "text")
vllm_outputs = vllm_reranker(vllm_runner, model_name, dtype, query,
documents, "image", "text")
hf_outputs = hf_reranker(
hf_runner, model_name, dtype, query, documents, "image", "text"
)
vllm_outputs = vllm_reranker(
vllm_runner, model_name, dtype, query, documents, "image", "text"
)
assert hf_outputs[0] == pytest.approx(vllm_outputs[0], rel=0.02)
assert hf_outputs[1] == pytest.approx(vllm_outputs[1], rel=0.02)
@@ -178,10 +184,12 @@ def test_model_image_image(hf_runner, vllm_runner, model_name, dtype):
"https://raw.githubusercontent.com/jina-ai/multimodal-reranker-test/main/paper-11.png",
]
hf_outputs = hf_reranker(hf_runner, model_name, dtype, query, documents,
"image", "image")
vllm_outputs = vllm_reranker(vllm_runner, model_name, dtype, query,
documents, "image", "image")
hf_outputs = hf_reranker(
hf_runner, model_name, dtype, query, documents, "image", "image"
)
vllm_outputs = vllm_reranker(
vllm_runner, model_name, dtype, query, documents, "image", "image"
)
assert hf_outputs[0] == pytest.approx(vllm_outputs[0], rel=0.02)
assert hf_outputs[1] == pytest.approx(vllm_outputs[1], rel=0.02)

53
tests/models/multimodal/pooling/test_llava_next.py
View File

@@ -24,9 +24,10 @@ from ...utils import check_embeddings_close
# built with LAPACK support.
pytestmark = pytest.mark.skipif(
not current_platform.is_cuda(),
reason="Llava Next model uses op that is only supported in CUDA")
reason="Llava Next model uses op that is only supported in CUDA",
)
llama3_template = '<|start_header_id|>user<|end_header_id|>\n\n{}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n \n' # noqa: E501
llama3_template = "<|start_header_id|>user<|end_header_id|>\n\n{}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n \n" # noqa: E501
HF_TEXT_PROMPTS = [
# T -> X
@@ -34,18 +35,21 @@ HF_TEXT_PROMPTS = [
"The label of the object is stop sign\nSummary above sentence in one word: " # noqa: E501
),
# T -> X
llama3_template.format(
"cherry blossom\nSummary above sentence in one word: "),
llama3_template.format("cherry blossom\nSummary above sentence in one word: "),
]
HF_IMAGE_PROMPTS = IMAGE_ASSETS.prompts({
# I -> X
"stop_sign":
llama3_template.format("<image>\nSummary above image in one word: "),
# I -> X
"cherry_blossom":
llama3_template.format("<image>\nSummary above image in one word: "),
})
HF_IMAGE_PROMPTS = IMAGE_ASSETS.prompts(
{
# I -> X
"stop_sign": llama3_template.format(
"<image>\nSummary above image in one word: "
),
# I -> X
"cherry_blossom": llama3_template.format(
"<image>\nSummary above image in one word: "
),
}
)
MODELS = ["royokong/e5-v"]
@@ -63,23 +67,22 @@ def _run_test(
# vLLM needs a fresh new process without cuda initialization.
# if we run HF first, the cuda initialization will be done and it
# will hurt multiprocessing backend with fork method (the default method).
with vllm_runner(model,
runner="pooling",
dtype=dtype,
max_model_len=4096,
enforce_eager=True) as vllm_model:
with vllm_runner(
model, runner="pooling", dtype=dtype, max_model_len=4096, enforce_eager=True
) as vllm_model:
vllm_outputs = vllm_model.embed(input_texts, images=input_images)
with hf_runner(model, dtype=dtype,
auto_cls=AutoModelForImageTextToText) as hf_model:
with hf_runner(
model, dtype=dtype, auto_cls=AutoModelForImageTextToText
) as hf_model:
# Patch the issue where generation_config.json is missing
hf_model.processor.patch_size = \
hf_model.model.config.vision_config.patch_size
hf_model.processor.patch_size = hf_model.model.config.vision_config.patch_size
# Patch the issue where image_token_id
# exceeds the maximum allowed vocab size
hf_model.model.resize_token_embeddings(
hf_model.model.language_model.vocab_size + 1)
hf_model.model.language_model.vocab_size + 1
)
all_inputs = hf_model.get_inputs(input_texts, images=input_images)
@@ -91,8 +94,7 @@ def _run_test(
return_dict=True,
output_hidden_states=True,
)
pooled_output = F.normalize(outputs.hidden_states[-1][0, -1, :],
dim=-1)
pooled_output = F.normalize(outputs.hidden_states[-1][0, -1, :], dim=-1)
all_outputs.append(pooled_output.tolist())
@@ -142,8 +144,7 @@ def test_models_image(
dtype: str,
) -> None:
input_texts_images = [
(text, asset.pil_image)
for text, asset in zip(HF_IMAGE_PROMPTS, image_assets)
(text, asset.pil_image) for text, asset in zip(HF_IMAGE_PROMPTS, image_assets)
]
input_texts = [text for text, _ in input_texts_images]
input_images = [image for _, image in input_texts_images]

47
tests/models/multimodal/pooling/test_phi3v.py
View File

@@ -19,14 +19,14 @@ HF_TEXT_PROMPTS = [
"Retrieve an image of this caption: cherry blossom",
]
HF_IMAGE_PROMPTS = IMAGE_ASSETS.prompts({
# T + I -> X
"stop_sign":
"<|image_1|> Select the portion of the image that isolates the object of the given label: The label of the object is stop sign", # noqa: E501
# I -> X
"cherry_blossom":
"<|image_1|> Represent the given image for classification", # noqa: E501
})
HF_IMAGE_PROMPTS = IMAGE_ASSETS.prompts(
{
# T + I -> X
"stop_sign": "<|image_1|> Select the portion of the image that isolates the object of the given label: The label of the object is stop sign", # noqa: E501
# I -> X
"cherry_blossom": "<|image_1|> Represent the given image for classification", # noqa: E501
}
)
MODELS = ["TIGER-Lab/VLM2Vec-Full"]
@@ -44,14 +44,14 @@ def _run_test(
# vLLM needs a fresh new process without cuda initialization.
# if we run HF first, the cuda initialization will be done and it
# will hurt multiprocessing backend with fork method (the default method).
with vllm_runner(model, runner="pooling", dtype=dtype,
enforce_eager=True) as vllm_model:
with vllm_runner(
model, runner="pooling", dtype=dtype, enforce_eager=True
) as vllm_model:
vllm_outputs = vllm_model.embed(input_texts, images=input_images)
# use eager mode for hf runner, since phi3_v didn't work with flash_attn
hf_model_kwargs = {"_attn_implementation": "eager"}
with hf_runner(model, dtype=dtype,
model_kwargs=hf_model_kwargs) as hf_model:
with hf_runner(model, dtype=dtype, model_kwargs=hf_model_kwargs) as hf_model:
all_inputs = hf_model.get_inputs(input_texts, images=input_images)
all_outputs = []
@@ -114,18 +114,21 @@ def test_models_image(
dtype: str,
) -> None:
input_texts_images = [
(text, asset.pil_image)
for text, asset in zip(HF_IMAGE_PROMPTS, image_assets)
(text, asset.pil_image) for text, asset in zip(HF_IMAGE_PROMPTS, image_assets)
]
# add cases for special_tokens
input_texts_images.append((
"\n<s><|user|>\n <|image_1|>\n\t <s>"
"Represent the given image for classification<|end|>"
"\n<|assistant|>\n",
Image.open(
get_vllm_public_assets(filename="cherry_blossom.jpg",
s3_prefix=VLM_IMAGES_DIR)),
))
input_texts_images.append(
(
"\n<s><|user|>\n <|image_1|>\n\t <s>"
"Represent the given image for classification<|end|>"
"\n<|assistant|>\n",
Image.open(
get_vllm_public_assets(
filename="cherry_blossom.jpg", s3_prefix=VLM_IMAGES_DIR
)
),
)
)
input_texts = [text for text, _ in input_texts_images]
input_images = [image for _, image in input_texts_images]

22
tests/models/multimodal/pooling/test_prithvi_mae.py
View File

@@ -19,25 +19,25 @@ def _run_test(
vllm_runner: type[VllmRunner],
model: str,
) -> None:
prompt = [
{
# This model deals with no text input
"prompt_token_ids": [1],
"multi_modal_data": generate_test_mm_data(),
} for _ in range(10)
}
for _ in range(10)
]
with vllm_runner(
model,
runner="pooling",
dtype="half",
enforce_eager=True,
skip_tokenizer_init=True,
# Limit the maximum number of sequences to avoid the
# test going OOM during the warmup run
max_num_seqs=32,
default_torch_num_threads=1,
model,
runner="pooling",
dtype="half",
enforce_eager=True,
skip_tokenizer_init=True,
# Limit the maximum number of sequences to avoid the
# test going OOM during the warmup run
max_num_seqs=32,
default_torch_num_threads=1,
) as vllm_model:
vllm_model.encode(prompt)

29
tests/models/multimodal/pooling/test_radio.py
View File

@@ -34,9 +34,9 @@ def run_radio_test(
# Using `self.get_nearest_supported_resolution`, for assets 432x642 the
# nearest supported resolution is 432x640.
pixel_values = [
img_processor(
image,
return_tensors='pt').pixel_values.to(torch_dtype)[:, :, :, :640]
img_processor(image, return_tensors="pt").pixel_values.to(torch_dtype)[
:, :, :, :640
]
for image in images
]
@@ -51,32 +51,33 @@ def run_radio_test(
hf_model.eval()
hf_outputs_per_image = [
hf_model(pixel_value.to("cuda")).features
for pixel_value in pixel_values
hf_model(pixel_value.to("cuda")).features for pixel_value in pixel_values
]
radio_config = RadioConfig(model_name=config.args["model"],
reg_tokens=config.args["register_multiple"])
radio_config = RadioConfig(
model_name=config.args["model"], reg_tokens=config.args["register_multiple"]
)
vllm_model = RadioModel(radio_config)
vllm_model.load_weights(hf_model.state_dict())
vllm_model = vllm_model.to("cuda", torch_dtype)
vllm_outputs_per_image = [
vllm_model(pixel_values=pixel_value.to("cuda"))
for pixel_value in pixel_values
vllm_model(pixel_values=pixel_value.to("cuda")) for pixel_value in pixel_values
]
del vllm_model, hf_model
cleanup_dist_env_and_memory()
cos_similar = nn.CosineSimilarity(dim=-1)
for vllm_output, hf_output in zip(vllm_outputs_per_image,
hf_outputs_per_image):
for vllm_output, hf_output in zip(vllm_outputs_per_image, hf_outputs_per_image):
assert cos_similar(vllm_output, hf_output).mean() > 0.99
@pytest.mark.parametrize("model_id", [
"nvidia/C-RADIOv2-H",
])
@pytest.mark.parametrize(
"model_id",
[
"nvidia/C-RADIOv2-H",
],
)
@pytest.mark.parametrize("dtype", ["half"])
def test_radio(dist_init, image_assets, model_id, dtype: str) -> None:
run_radio_test(