[Model] Support SigLIP encoder and alternative decoders for LLaVA models (#7153)

Co-authored-by: Roger Wang <136131678+ywang96@users.noreply.github.com>
This commit is contained in:
Cyrus Leung
2024-08-06 16:55:31 +08:00
committed by GitHub
co-authored by Roger Wang
parent 9118217f58
commit 1f26efbb3a
14 changed files with 453 additions and 267 deletions
+31 -5
View File
@@ -1,10 +1,11 @@
from typing import List, Optional, Tuple, Type
import pytest
from transformers import AutoTokenizer
from transformers import AutoConfig, AutoTokenizer
from vllm.multimodal.utils import rescale_image_size
from vllm.sequence import SampleLogprobs
from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE
from ..conftest import IMAGE_ASSETS, HfRunner, VllmRunner, _ImageAssets
from .utils import check_logprobs_close
@@ -18,9 +19,11 @@ HF_IMAGE_PROMPTS = IMAGE_ASSETS.prompts({
"USER: <image>\nWhat is the season?\nASSISTANT:",
})
IMAGE_TOKEN_ID = 32000
models = ["llava-hf/llava-1.5-7b-hf"]
models = [
"llava-hf/llava-1.5-7b-hf",
# TODO: Get this model to produce meaningful output in vLLM
# "TIGER-Lab/Mantis-8B-siglip-llama3",
]
def vllm_to_hf_output(vllm_output: Tuple[List[int], str,
@@ -29,12 +32,15 @@ def vllm_to_hf_output(vllm_output: Tuple[List[int], str,
"""Sanitize vllm output to be comparable with hf output."""
output_ids, output_str, out_logprobs = vllm_output
config = AutoConfig.from_pretrained(model)
image_token_id = config.image_token_index
tokenizer = AutoTokenizer.from_pretrained(model)
eos_token_id = tokenizer.eos_token_id
hf_output_ids = [
token_id for idx, token_id in enumerate(output_ids)
if token_id != IMAGE_TOKEN_ID or output_ids[idx - 1] != IMAGE_TOKEN_ID
if token_id != image_token_id or output_ids[idx - 1] != image_token_id
]
assert output_str[0] == " "
@@ -67,6 +73,17 @@ def run_test(
Note, the text input is also adjusted to abide by vllm contract.
The text output is sanitized to be able to compare with hf.
"""
# NOTE: For local use; this isn't tested in CI yet (see TODO above)
if model.startswith("TIGER-Lab/Mantis"):
from mantis.models.mllava import MLlavaProcessor
torch_dtype = STR_DTYPE_TO_TORCH_DTYPE[dtype]
mantis_processor = MLlavaProcessor.from_pretrained(
model, torch_dtype=torch_dtype)
assert isinstance(mantis_processor, MLlavaProcessor)
else:
mantis_processor = None
images = [asset.pil_image for asset in image_assets]
inputs_per_image = [(
@@ -94,6 +111,15 @@ def run_test(
]
with hf_runner(model, dtype=dtype, is_vision_model=True) as hf_model:
if mantis_processor is not None:
def process(*args, **kwargs):
output = mantis_processor(*args, **kwargs)
output["pixel_values"] = output["pixel_values"].to(torch_dtype)
return output
hf_model.processor = process
hf_outputs_per_image = [
hf_model.generate_greedy_logprobs_limit(prompts,
max_tokens,
+5 -4
View File
@@ -1,7 +1,7 @@
from typing import List, Optional, Tuple, Type, overload
import pytest
from transformers import AutoTokenizer
from transformers import AutoConfig, AutoTokenizer
from vllm.multimodal.utils import rescale_image_size
from vllm.sequence import SampleLogprobs
@@ -23,8 +23,6 @@ HF_IMAGE_PROMPTS = IMAGE_ASSETS.prompts({
f"{_PREFACE} USER: <image>\nWhat is the season? ASSISTANT:",
})
IMAGE_TOKEN_ID = 32000
models = ["llava-hf/llava-v1.6-vicuna-7b-hf"]
@@ -34,12 +32,15 @@ def vllm_to_hf_output(vllm_output: Tuple[List[int], str,
"""Sanitize vllm output to be comparable with hf output."""
output_ids, output_str, out_logprobs = vllm_output
config = AutoConfig.from_pretrained(model)
image_token_id = config.image_token_index
tokenizer = AutoTokenizer.from_pretrained(model)
eos_token_id = tokenizer.eos_token_id
hf_output_ids = [
token_id for idx, token_id in enumerate(output_ids)
if token_id != IMAGE_TOKEN_ID or output_ids[idx - 1] != IMAGE_TOKEN_ID
if token_id != image_token_id or output_ids[idx - 1] != image_token_id
]
assert output_str[0] == " "
+5 -4
View File
@@ -2,7 +2,7 @@ import os
from typing import List, Optional, Tuple, Type
import pytest
from transformers import AutoTokenizer
from transformers import AutoConfig, AutoTokenizer
from vllm.multimodal.utils import rescale_image_size
from vllm.sequence import SampleLogprobs
@@ -20,8 +20,6 @@ HF_IMAGE_PROMPTS = IMAGE_ASSETS.prompts({
"What is in the picture?",
})
IMAGE_TOKEN_ID = 257152
models = ["google/paligemma-3b-mix-224"]
# ROCm Triton FA can run into compilation issues with these models due to,
@@ -37,12 +35,15 @@ def vllm_to_hf_output(vllm_output: Tuple[List[int], str,
"""Sanitize vllm output to be comparable with hf output."""
output_ids, output_str, out_logprobs = vllm_output
config = AutoConfig.from_pretrained(model)
image_token_id = config.image_token_index
tokenizer = AutoTokenizer.from_pretrained(model)
eos_token_id = tokenizer.eos_token_id
hf_output_ids = [
token_id for idx, token_id in enumerate(output_ids)
if token_id != IMAGE_TOKEN_ID or output_ids[idx - 1] != IMAGE_TOKEN_ID
if token_id != image_token_id or output_ids[idx - 1] != image_token_id
]
hf_output_str = output_str
+1 -1
View File
@@ -6,4 +6,4 @@ from vllm.model_executor.models import _MODELS, ModelRegistry
@pytest.mark.parametrize("model_cls", _MODELS)
def test_registry_imports(model_cls):
# Ensure all model classes can be imported successfully
ModelRegistry.load_model_cls(model_cls)
ModelRegistry.resolve_model_cls([model_cls])