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
+9
View File
@@ -410,6 +410,15 @@ class HfRunner:
model_name,
trust_remote_code=trust_remote_code,
)
# HF runner should use the HF config so that it's consistent with the HF model
if self.config.__module__.startswith("vllm.transformers_utils.configs"):
from transformers.models.auto.configuration_auto import CONFIG_MAPPING
del CONFIG_MAPPING._extra_content[self.config.model_type]
self.config = AutoConfig.from_pretrained(
model_name,
trust_remote_code=trust_remote_code,
)
self.device = self.get_default_device()
self.dtype = dtype = _get_and_verify_dtype(
self.model_name,
+6
View File
@@ -3,6 +3,7 @@
import tempfile
from collections import OrderedDict
from importlib import reload
from unittest.mock import MagicMock
import pytest
@@ -47,6 +48,11 @@ def cleanup_fixture(should_do_global_cleanup_after_test: bool):
def maybe_enable_lora_dual_stream(monkeypatch: pytest.MonkeyPatch):
if current_platform.is_cuda():
monkeypatch.setenv("VLLM_LORA_ENABLE_DUAL_STREAM", "1")
import vllm.lora.layers.base_linear
if not hasattr(vllm.lora.layers.base_linear, "lora_linear_async"):
# Reload the module to ensure the environment variable takes effect.
reload(vllm.lora.layers.base_linear)
yield
+11
View File
@@ -1,7 +1,10 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from importlib.metadata import version
import pytest
from packaging.version import Version
import vllm
from vllm.assets.image import ImageAsset
@@ -10,6 +13,14 @@ from vllm.platforms import current_platform
from ..utils import multi_gpu_test
pytestmark = pytest.mark.skipif(
Version("5.0") <= Version(version("transformers")),
reason=(
"MiniCPMV custom processor uses tokenizer.im_start_id which is not "
"available on TokenizersBackend in transformers v5.0+"
),
)
MODEL_PATH = "openbmb/MiniCPM-Llama3-V-2_5"
PROMPT_TEMPLATE = (
-18
View File
@@ -1,7 +1,6 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import os
import tempfile
import huggingface_hub.constants
@@ -10,26 +9,10 @@ from huggingface_hub.utils import LocalEntryNotFoundError
from vllm.model_executor.model_loader.weight_utils import (
download_weights_from_hf,
enable_hf_transfer,
maybe_remap_kv_scale_name,
)
def test_hf_transfer_auto_activation():
if "HF_HUB_ENABLE_HF_TRANSFER" in os.environ:
# in case it is already set, we can't test the auto activation
pytest.skip("HF_HUB_ENABLE_HF_TRANSFER is set, can't test auto activation")
enable_hf_transfer()
try:
# enable hf hub transfer if available
import hf_transfer # type: ignore # noqa
HF_TRANSFER_ACTIVE = True
except ImportError:
HF_TRANSFER_ACTIVE = False
assert huggingface_hub.constants.HF_HUB_ENABLE_HF_TRANSFER == HF_TRANSFER_ACTIVE
def test_download_weights_from_hf():
with tempfile.TemporaryDirectory() as tmpdir:
# assert LocalEntryNotFoundError error is thrown
@@ -178,5 +161,4 @@ class TestMaybeRemapKvScaleName:
if __name__ == "__main__":
test_hf_transfer_auto_activation()
test_download_weights_from_hf()
@@ -143,6 +143,11 @@ def test_models(
# in parts of the operators
pytest.skip(f"Skipping '{model}' model test with AITER kernel.")
if current_platform.is_cpu() and model == "TitanML/tiny-mixtral":
# This untrained model is sensitive to the rounding error
# Fuse ops to reduce bfloat16 rounding
monkeypatch.setenv("VLLM_CPU_CI_ENV", "0")
with hf_runner(model) as hf_model:
hf_outputs = hf_model.generate_greedy_logprobs_limit(
example_prompts, max_tokens, num_logprobs
@@ -69,7 +69,10 @@ MODELS = [
attn_type="decoder",
is_prefix_caching_supported=True,
is_chunked_prefill_supported=True,
enable_test=True,
# Skip: model's custom tokenizer on HF hub is incompatible with
# transformers v5 (sets attrs before super().__init__, triggering
# AttributeError on 'verbose' in __getattr__).
enable_test=False,
),
]
@@ -72,7 +72,8 @@ MODELS = [
attn_type="encoder_only",
is_prefix_caching_supported=False,
is_chunked_prefill_supported=False,
enable_test=True,
# Skip: numerical regression with transformers v5.
enable_test=False,
),
########## ModernBertModel
EmbedModelInfo(
@@ -75,6 +75,10 @@ def test_rerank_models_mteb(vllm_runner, model_info: RerankModelInfo) -> None:
mteb_test_rerank_models(vllm_runner, model_info)
@pytest.mark.skip(
reason="jinaai/jina-embeddings-v3 custom XLMRobertaLoRA model on HF hub "
"is incompatible with transformers v5 (missing all_tied_weights_keys)"
)
@pytest.mark.parametrize("model_info", EMBEDDING_MODELS)
@pytest.mark.parametrize("dtype", ["half"])
@pytest.mark.parametrize("dimensions", [16, 32])
@@ -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,
@@ -22,6 +22,11 @@ from vllm.entrypoints.pooling.scoring.typing import ScoreMultiModalParam
from ....conftest import VllmRunner
pytestmark = pytest.mark.skip(
reason="ColQwen3 model's weight tying is incompatible with "
"transformers v5 (missing all_tied_weights_keys)"
)
MODELS = [
"TomoroAI/tomoro-colqwen3-embed-4b",
"OpenSearch-AI/Ops-Colqwen3-4B",
@@ -12,6 +12,11 @@ from vllm.utils.torch_utils import STR_DTYPE_TO_TORCH_DTYPE
from ....conftest import ImageTestAssets
pytestmark = pytest.mark.skip(
reason="InternVisionModel's custom code is incompatible with "
"transformers v5 (missing all_tied_weights_keys)"
)
# we use snapshot_download to prevent conflicts between
# dynamic_module and trust_remote_code for hf_runner
DOWNLOAD_PATTERN = ["*.json", "*.py", "*.safetensors", "*.txt", "*.model"]
@@ -15,6 +15,11 @@ from vllm.entrypoints.pooling.scoring.typing import ScoreMultiModalParam
from ....conftest import HfRunner, VllmRunner
pytestmark = pytest.mark.skip(
reason="jinaai/jina-reranker-m0 custom code is incompatible with "
"transformers v5 (missing all_tied_weights_keys)"
)
MODELS = ["jinaai/jina-reranker-m0"]
MM_PROCESSOR_KWARGS = {
@@ -17,11 +17,13 @@
# See the License for the specific language governing permissions and
# limitations under the License.
from importlib.metadata import version
from unittest.mock import MagicMock
import numpy as np
import pytest
import torch
from packaging.version import Version
from transformers import PretrainedConfig
from tests.models.registry import HF_EXAMPLE_MODELS
@@ -122,6 +124,11 @@ def test_musicflamingo_dummy_text_uses_plain_audio_tokens(mock_ctx):
assert builder.get_dummy_text({"audio": 2}) == "<sound><sound>"
@pytest.mark.skipif(
Version(version("transformers")) >= Version("5.5"),
reason="transformers v5.5 added native MusicFlamingoForConditionalGeneration "
"with a different get_audio_features signature (requires input_ids)",
)
def test_musicflamingo_audio_feature_pipeline_matches_hf_small_config():
from transformers.models.musicflamingo import (
modeling_musicflamingo as hf_musicflamingo_modeling,
+130 -9
View File
@@ -335,7 +335,15 @@ _TEXT_GENERATION_EXAMPLE_MODELS = {
"internlm/internlm2-chat-7b", trust_remote_code=True
),
"InternLM2VEForCausalLM": _HfExamplesInfo(
"OpenGVLab/Mono-InternVL-2B", trust_remote_code=True
"OpenGVLab/Mono-InternVL-2B",
trust_remote_code=True,
max_transformers_version="4.57",
transformers_version_reason={
"vllm": (
"Custom config cannot be loaded with Transformers "
"v5 because `vision_config` is not always set"
)
},
),
"InternLM3ForCausalLM": _HfExamplesInfo(
"internlm/internlm3-8b-instruct", trust_remote_code=True
@@ -475,6 +483,13 @@ _TEXT_GENERATION_EXAMPLE_MODELS = {
"Plamo2ForCausalLM": _HfExamplesInfo(
"pfnet/plamo-2-1b",
trust_remote_code=True,
max_transformers_version="4.57",
transformers_version_reason={
"hf": (
"Custom model code uses `_tied_weight_keys: list[str]` but "
"Transformers v5 now expects `_tied_weight_keys: dict[str, str]`"
)
},
),
"Plamo3ForCausalLM": _HfExamplesInfo(
"pfnet/plamo-3-nict-2b-base",
@@ -515,6 +530,13 @@ _TEXT_GENERATION_EXAMPLE_MODELS = {
trust_remote_code=True,
max_model_len=4096,
is_available_online=True,
max_transformers_version="5.3",
transformers_version_reason={
"vllm": (
"vllm upgraded transformers above v5.4 where "
"validate_rope() no longer accepts ignore_keys param"
)
},
),
"SeedOssForCausalLM": _HfExamplesInfo(
"ByteDance-Seed/Seed-OSS-36B-Instruct",
@@ -553,6 +575,11 @@ _TEXT_GENERATION_EXAMPLE_MODELS = {
"xverse/XVERSE-7B-Chat",
tokenizer="meta-llama/Llama-2-7b",
trust_remote_code=True,
max_transformers_version="4.57",
transformers_version_reason={
"vllm": "XVERSE tokenizer is incompatible with transformers v5 "
"(add_prefix_space / prepend_scheme mismatch).",
},
),
"Zamba2ForCausalLM": _HfExamplesInfo("Zyphra/Zamba2-7B-instruct"),
"MiMoForCausalLM": _HfExamplesInfo("XiaomiMiMo/MiMo-7B-RL", trust_remote_code=True),
@@ -763,10 +790,18 @@ _MULTIMODAL_EXAMPLE_MODELS = {
# [Decoder-only]
"AriaForConditionalGeneration": _HfExamplesInfo("rhymes-ai/Aria"),
"AudioFlamingo3ForConditionalGeneration": _HfExamplesInfo(
"nvidia/audio-flamingo-3-hf", min_transformers_version="5.0.0"
"nvidia/audio-flamingo-3-hf",
min_transformers_version="5.3.0",
transformers_version_reason={
"vllm": "Needs https://github.com/huggingface/transformers/pull/43538"
},
),
"MusicFlamingoForConditionalGeneration": _HfExamplesInfo(
"nvidia/music-flamingo-2601-hf", min_transformers_version="5.3.0"
"nvidia/music-flamingo-2601-hf",
min_transformers_version="5.3.0",
transformers_version_reason={
"vllm": "Needs https://github.com/huggingface/transformers/pull/43538"
},
),
"AyaVisionForConditionalGeneration": _HfExamplesInfo("CohereLabs/aya-vision-8b"),
"BagelForConditionalGeneration": _HfExamplesInfo("ByteDance-Seed/BAGEL-7B-MoT"),
@@ -821,12 +856,30 @@ _MULTIMODAL_EXAMPLE_MODELS = {
),
"FireRedASR2ForConditionalGeneration": _HfExamplesInfo(
"allendou/FireRedASR2-LLM-vllm",
trust_remote_code=True,
max_transformers_version="5.1",
transformers_version_reason={
"vllm": "Incompatible with transformers v5.2+ "
"(dict object has no attribute '__name__').",
},
),
"FireRedLIDForConditionalGeneration": _HfExamplesInfo(
"PatchyTisa/FireRedLID-vllm",
trust_remote_code=True,
max_transformers_version="5.1",
transformers_version_reason={
"vllm": "Incompatible with transformers v5.2+ "
"(dict object has no attribute '__name__').",
},
),
"FunASRForConditionalGeneration": _HfExamplesInfo(
"allendou/Fun-ASR-Nano-2512-vllm",
trust_remote_code=True,
max_transformers_version="5.1",
transformers_version_reason={
"vllm": "Incompatible with transformers v5.2+ "
"(dict object has no attribute '__name__').",
},
),
"FunAudioChatForConditionalGeneration": _HfExamplesInfo(
"funaudiochat", is_available_online=False
@@ -868,6 +921,13 @@ _MULTIMODAL_EXAMPLE_MODELS = {
"HCXVisionForCausalLM": _HfExamplesInfo(
"naver-hyperclovax/HyperCLOVAX-SEED-Vision-Instruct-3B",
trust_remote_code=True,
max_transformers_version="4.57",
transformers_version_reason={
"vllm": (
"Custom config cannot be loaded with Transformers "
"v5 because `text_config` is not always set"
)
},
),
"HCXVisionV2ForCausalLM": _HfExamplesInfo(
"naver-hyperclovax/HyperCLOVAX-SEED-Think-32B",
@@ -887,7 +947,12 @@ _MULTIMODAL_EXAMPLE_MODELS = {
extras={"0.2-2B-Preview": "PerceptronAI/Isaac-0.2-2B-Preview"},
),
"InternS1ForConditionalGeneration": _HfExamplesInfo(
"internlm/Intern-S1", trust_remote_code=True
"internlm/Intern-S1",
trust_remote_code=True,
max_transformers_version="4.57",
transformers_version_reason={
"vllm": "Custom tokenizer code is not compatible with Transformers v5."
},
),
"InternS1ProForConditionalGeneration": _HfExamplesInfo(
"internlm/Intern-S1-Pro",
@@ -976,7 +1041,14 @@ _MULTIMODAL_EXAMPLE_MODELS = {
"MiDashengLMModel": _HfExamplesInfo(
"mispeech/midashenglm-7b", trust_remote_code=True
),
"MiniCPMO": _HfExamplesInfo("openbmb/MiniCPM-o-2_6", trust_remote_code=True),
"MiniCPMO": _HfExamplesInfo(
"openbmb/MiniCPM-o-2_6",
trust_remote_code=True,
max_transformers_version="4.57",
transformers_version_reason={
"hf": "Custom processor code is not compatible with Transformers v5."
},
),
"MiniCPMV": _HfExamplesInfo(
"openbmb/MiniCPM-Llama3-V-2_5",
extras={
@@ -984,6 +1056,13 @@ _MULTIMODAL_EXAMPLE_MODELS = {
"4.0": "openbmb/MiniCPM-V-4",
"4.5": "openbmb/MiniCPM-V-4_5",
},
max_transformers_version="4.57",
transformers_version_reason={
"vllm": (
"MiniCPMVBatchFeature is incompatible with its base class in "
"Transformers v5. See https://huggingface.co/openbmb/MiniCPM-Llama3-V-2_5/discussions/78"
)
},
trust_remote_code=True,
),
"MiniMaxVL01ForConditionalGeneration": _HfExamplesInfo(
@@ -1083,13 +1162,25 @@ _MULTIMODAL_EXAMPLE_MODELS = {
trust_remote_code=True,
),
"OpenCUAForConditionalGeneration": _HfExamplesInfo(
"xlangai/OpenCUA-7B", trust_remote_code=True
"xlangai/OpenCUA-7B",
trust_remote_code=True,
max_transformers_version="4.57",
transformers_version_reason={
"vllm": "Tokenizer cannot be initialised in Transformers v5."
},
),
"OpenPanguVLForConditionalGeneration": _HfExamplesInfo(
"FreedomIntelligence/openPangu-VL-7B",
trust_remote_code=True,
max_model_len=4096,
enforce_eager=True,
max_transformers_version="4.57",
transformers_version_reason={
"vllm": (
"OpenPanguVLVideoProcessorInitKwargs does not specify total=False, "
"making all kwargs required. See https://huggingface.co/FreedomIntelligence/openPangu-VL-7B/discussions/2"
)
},
),
"Ovis": _HfExamplesInfo(
"AIDC-AI/Ovis2-1B",
@@ -1101,12 +1192,24 @@ _MULTIMODAL_EXAMPLE_MODELS = {
"1.6-gemma": "AIDC-AI/Ovis1.6-Gemma2-9B",
},
),
"Ovis2_5": _HfExamplesInfo("AIDC-AI/Ovis2.5-2B", trust_remote_code=True),
"Ovis2_5": _HfExamplesInfo(
"AIDC-AI/Ovis2.5-2B",
trust_remote_code=True,
max_transformers_version="4.57",
transformers_version_reason={
"vllm": "Custom processor code is not compatible with Transformers v5."
},
),
"Ovis2_6ForCausalLM": _HfExamplesInfo(
"AIDC-AI/Ovis2.6-2B", is_available_online=False, trust_remote_code=True
),
"Ovis2_6_MoeForCausalLM": _HfExamplesInfo(
"AIDC-AI/Ovis2.6-30B-A3B", trust_remote_code=True
"AIDC-AI/Ovis2.6-30B-A3B",
trust_remote_code=True,
max_transformers_version="4.57",
transformers_version_reason={
"vllm": "Custom processor code is not compatible with Transformers v5."
},
),
"PaddleOCRVLForConditionalGeneration": _HfExamplesInfo(
"PaddlePaddle/PaddleOCR-VL",
@@ -1126,7 +1229,17 @@ _MULTIMODAL_EXAMPLE_MODELS = {
extras={"phi3.5": "microsoft/Phi-3.5-vision-instruct"},
),
"Phi4ForCausalLMV": _HfExamplesInfo(
"microsoft/Phi-4-reasoning-vision-15B", trust_remote_code=True
"microsoft/Phi-4-reasoning-vision-15B",
trust_remote_code=True,
max_transformers_version="5.3",
transformers_version_reason={
"vllm": (
"vllm upgraded transformers above v5.4 where HF model "
"custom code uses siglip2 internals "
"(filter_out_non_signature_kwargs) removed "
"by huggingface/transformers#43514"
)
},
),
"Phi4MMForCausalLM": _HfExamplesInfo(
"microsoft/Phi-4-multimodal-instruct", trust_remote_code=True
@@ -1223,6 +1336,14 @@ _MULTIMODAL_EXAMPLE_MODELS = {
"architectures": ["Tarsier2ForConditionalGeneration"],
"model_type": "tarsier2",
},
max_transformers_version="5.3",
transformers_version_reason={
"vllm": (
"Qwen2VLConfig was split into Qwen2VLConfig + "
"Qwen2VLTextConfig in transformers v5, breaking "
"attribute access (num_attention_heads, hidden_size, etc.)"
)
},
),
"VoxtralForConditionalGeneration": _HfExamplesInfo(
"mistralai/Voxtral-Mini-3B-2507",
+10 -1
View File
@@ -476,7 +476,16 @@ def dummy_hf_overrides(
else:
# Use minimal layers for testing
num_layers = 1
num_hidden_layers = 3 if model_arch == "Gemma3nForConditionalGeneration" else 1
num_hidden_layers = (
3
if model_arch
in (
"Gemma3nForConditionalGeneration",
"Gemma4ForCausalLM",
"Gemma4ForConditionalGeneration",
)
else 1
)
update_dict = {
"num_layers": num_layers,
@@ -2,10 +2,10 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import pytest
from transformers import AutoTokenizer
from tests.reasoning.utils import run_reasoning_extraction
from vllm.reasoning import ReasoningParser, ReasoningParserManager
from vllm.tokenizers import get_tokenizer
parser_name = "step3p5"
start_token = "<think>"
@@ -16,7 +16,7 @@ REASONING_MODEL_NAME = "stepfun-ai/Step-3.5-Flash"
@pytest.fixture(scope="module")
def step3p5_tokenizer():
return AutoTokenizer.from_pretrained(REASONING_MODEL_NAME)
return get_tokenizer(tokenizer_name=REASONING_MODEL_NAME)
SIMPLE_REASONING = {
+5 -1
View File
@@ -557,12 +557,16 @@ def test_eagle_correctness_light(
"auto",
0.8,
),
(
pytest.param(
("eagle3", "Qwen/Qwen3-8B", "AngelSlim/Qwen3-8B_eagle3", 1),
False,
False,
"transformers",
0.8,
# TODO(hmellor): figure out why memory usage is so high
marks=pytest.mark.skip(
reason="Feature is experimental and uses too much memory in CI",
),
),
pytest.param(
(