Update to transformers v5 (#30566)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Signed-off-by: khluu <khluu000@gmail.com>
Signed-off-by: Kevin H. Luu <khluu000@gmail.com>
Signed-off-by: Cyrus Leung <cyrus.tl.leung@gmail.com>
Co-authored-by: khluu <khluu000@gmail.com>
Co-authored-by: Cyrus Leung <cyrus.tl.leung@gmail.com>
Co-authored-by: jiang1.li <jiang1.li@intel.com>
This commit is contained in:
Harry Mellor
2026-04-15 16:29:15 -07:00
committed by GitHub
co-authored by khluu Cyrus Leung jiang1.li
parent 6dc9491406
commit 03f8d3a548
41 changed files with 445 additions and 115 deletions
@@ -186,7 +186,14 @@ VLM_TEST_SETTINGS = {
max_num_seqs=2,
auto_cls=AutoModel,
hf_output_post_proc=model_utils.ultravox_trunc_hf_output,
marks=[pytest.mark.core_model, pytest.mark.cpu_model],
marks=[
pytest.mark.core_model,
pytest.mark.cpu_model,
# TODO: Remove skip once model has been upstreamed to Transformers
pytest.mark.skip(
reason="Custom model code is not compatible with Transformers v5"
),
],
),
#### Transformers fallback to test
## To reduce test burden, we only test batching arbitrary image size
@@ -397,14 +404,14 @@ VLM_TEST_SETTINGS = {
"gemma4": VLMTestInfo(
models=["google/gemma-4-E2B-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
prompt_formatter=lambda img_prompt: f"<bos><|turn>user\n{img_prompt}<turn|>\n<|turn>model\n", # noqa: E501
single_image_prompts=IMAGE_ASSETS.prompts(
{
"stop_sign": "What's the content in the center of the image?",
"cherry_blossom": "What is the season?",
"stop_sign": "<|image|>What's the content in the center of the image?", # noqa: E501
"cherry_blossom": "<|image|>What is the season?",
}
),
multi_image_prompt="Describe the two images in detail.",
multi_image_prompt="<|image|><|image|>Describe the two images in detail.", # noqa: E501
max_model_len=4096,
max_num_seqs=2,
auto_cls=AutoModelForImageTextToText,
@@ -533,6 +540,12 @@ VLM_TEST_SETTINGS = {
max_model_len=4096,
use_tokenizer_eos=True,
patch_hf_runner=model_utils.internvl_patch_hf_runner,
# TODO: Remove skip once model has been upstreamed to Transformers
marks=[
pytest.mark.skip(
reason="Custom model code tries to access data from meta-tensor"
)
],
),
"intern_vl-video": VLMTestInfo(
models=[
@@ -545,6 +558,12 @@ VLM_TEST_SETTINGS = {
use_tokenizer_eos=True,
patch_hf_runner=model_utils.internvl_patch_hf_runner,
num_logprobs=10 if current_platform.is_rocm() else 5,
# TODO: Remove skip once model has been upstreamed to Transformers
marks=[
pytest.mark.skip(
reason="Custom model code tries to access data from meta-tensor"
)
],
),
"intern_vl-hf": VLMTestInfo(
models=["OpenGVLab/InternVL3-1B-hf"],
@@ -591,6 +610,8 @@ VLM_TEST_SETTINGS = {
hf_model_kwargs={"device_map": "auto"},
patch_hf_runner=model_utils.isaac_patch_hf_runner,
image_size_factors=[(0.25,), (0.25, 0.25, 0.25), (0.25, 0.2, 0.15)],
# TODO: Remove skip once model has been upstreamed to Transformers
marks=[pytest.mark.skip(reason="Custom model imports deleted object")], # noqa: E501
),
"kimi_vl": VLMTestInfo(
models=["moonshotai/Kimi-VL-A3B-Instruct"],
@@ -806,7 +827,12 @@ VLM_TEST_SETTINGS = {
pytest.mark.skipif(
Version(TRANSFORMERS_VERSION) == Version("4.57.3"),
reason="This model is broken in Transformers v4.57.3",
)
),
pytest.mark.skipif(
Version(TRANSFORMERS_VERSION) >= Version("5.0.0"),
reason="Model's custom code uses ROPE_INIT_FUNCTIONS"
"['default'] which was removed in transformers v5",
),
],
),
"phi3v": VLMTestInfo(
@@ -960,6 +986,12 @@ VLM_TEST_SETTINGS = {
)
for inp in custom_inputs.different_patch_input_cases_internvl()
],
# TODO: Remove skip once model has been upstreamed to Transformers
marks=[
pytest.mark.skip(
reason="Custom model code tries to access data from meta-tensor"
)
],
),
"llava_onevision-multiple-images": VLMTestInfo(
models=["llava-hf/llava-onevision-qwen2-0.5b-ov-hf"],
@@ -103,6 +103,10 @@ def run_test(
)
@pytest.mark.skip(
reason="Model's custom MBart decoder has head count mismatch with "
"transformers v5's GQA-aware cross-attention (8 vs 16 heads)"
)
@pytest.mark.parametrize("model", ["nvidia/NVIDIA-Nemotron-Parse-v1.1"])
@pytest.mark.parametrize("dtype", ["bfloat16"])
@pytest.mark.parametrize("num_logprobs", [5])
@@ -2,9 +2,11 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from collections.abc import Sequence
from importlib.metadata import version
import pytest
import regex as re
from packaging.version import Version
from transformers import AutoModelForCausalLM, AutoTokenizer
from vllm.logprobs import SampleLogprobs
@@ -19,6 +21,15 @@ from ....conftest import (
from ....utils import multi_gpu_test
from ...utils import check_logprobs_close
pytestmark = pytest.mark.skipif(
Version("5.0") <= Version(version("transformers")),
reason=(
"vllm upgraded transformers above v5.4 where HF model custom code uses siglip2 "
"internals (filter_out_non_signature_kwargs) removed by "
"huggingface/transformers#43514"
),
)
MODEL_ID = "microsoft/Phi-4-reasoning-vision-15B"
HF_IMAGE_PROMPTS = IMAGE_ASSETS.prompts(
@@ -149,6 +149,10 @@ def test_online_serving(vllm_runner, audio_assets: AudioTestAssets):
)
@pytest.mark.skip(
reason="VoxtralProcessor.apply_chat_template() in transformers v5 "
"doesn't resolve chat_template=None to the default template"
)
def test_hf_reference(hf_runner, vllm_runner, audio_assets: AudioTestAssets):
"""Compare vLLM Mistral-format output against HF Transformers reference.
@@ -80,6 +80,11 @@ def run_test(
if vllm_runner_kwargs:
vllm_runner_kwargs_.update(vllm_runner_kwargs)
# Avoid passing limit_mm_per_prompt twice when vllm_runner_kwargs
# already contains it (e.g. gemma4 sets it via vllm_runner_kwargs).
if "limit_mm_per_prompt" in vllm_runner_kwargs_:
limit_mm_per_prompt = vllm_runner_kwargs_.pop("limit_mm_per_prompt")
with vllm_runner(
model,
max_model_len=max_model_len,