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>
(cherry picked from commit 03f8d3a5)

tests/models/language/pooling_mteb_test/test_baai.py

@@ -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,
     ),
 ]
 

tests/models/language/pooling_mteb_test/test_gte.py

@@ -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(

tests/models/language/pooling_mteb_test/test_jina.py

@@ -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])

tests/models/multimodal/generation/test_common.py

@@ -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"],

tests/models/multimodal/generation/test_nemotron_parse.py

@@ -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])

tests/models/multimodal/generation/test_phi4siglip.py

+# SPDX-License-Identifier: Apache-2.0
+# 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
+from vllm.multimodal.image import rescale_image_size
+
+from ....conftest import (
+    IMAGE_ASSETS,
+    HfRunner,
+    PromptImageInput,
+    VllmRunner,
+)
+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(
+    {
+        "stop_sign": "<|user|>\n<image>\nWhat's the content of the image?<|end|>\n<|assistant|>\n",  # noqa: E501
+        "cherry_blossom": "<|user|>\n<image>\nPlease infer the season with reason in details.<|end|>\n<|assistant|>\n",  # noqa: E501
+    }
+)
+HF_MULTIIMAGE_IMAGE_PROMPT = (
+    "<|user|>\n<image>\n<image>\nDescribe these images.<|end|>\n<|assistant|>\n"  # noqa: E501
+)
+
+DTYPE = "half"
+MAX_TOKENS = 128
+NUM_LOGPROBS = 10
+
+
+def vllm_to_hf_output(
+    vllm_output: tuple[list[int], str, SampleLogprobs | None], model: str
+):
+    """Sanitize vllm output to be comparable with hf output."""
+    _, output_str, out_logprobs = vllm_output
+
+    output_str_without_image = re.sub(r"(<image>)+", "", output_str)
+    if output_str_without_image and output_str_without_image[0] == " ":
+        output_str_without_image = output_str_without_image[1:]
+
+    hf_output_str = output_str_without_image + "<|end|><|endoftext|>"
+
+    tokenizer = AutoTokenizer.from_pretrained(model, trust_remote_code=True)
+    hf_output_ids = tokenizer.encode(output_str_without_image)
+    if hf_output_ids and hf_output_ids[0] == tokenizer.bos_token_id:
+        hf_output_ids = hf_output_ids[1:]
+
+    return hf_output_ids, hf_output_str, out_logprobs
+
+
+def _build_single_image_inputs(
+    image_assets,
+) -> list[tuple[list[str], PromptImageInput]]:
+    """Build single-image inputs for all size_factors at once."""
+    images = [asset.pil_image for asset in image_assets]
+    all_inputs: list[tuple[list[str], PromptImageInput]] = []
+    for size_factors in [[1.0], [0.25, 0.5, 1.0]]:
+        for image, prompt in zip(images, HF_IMAGE_PROMPTS):
+            all_inputs.append(
+                (
+                    [prompt for _ in size_factors],
+                    [rescale_image_size(image, f) for f in size_factors],
+                )
+            )
+    return all_inputs
+
+
+def _build_multi_image_inputs(
+    image_assets,
+) -> list[tuple[list[str], PromptImageInput]]:
+    """Build multi-image inputs for all size_factors at once."""
+    images = [asset.pil_image for asset in image_assets]
+    all_inputs: list[tuple[list[str], PromptImageInput]] = []
+    for size_factors in [[0.5], [0.15, 0.30]]:
+        all_inputs.append(
+            (
+                [HF_MULTIIMAGE_IMAGE_PROMPT for _ in size_factors],
+                [
+                    [rescale_image_size(image, factor) for image in images]
+                    for factor in size_factors
+                ],
+            )
+        )
+    return all_inputs
+
+
+def _run_and_compare(
+    hf_runner: type[HfRunner],
+    vllm_runner: type[VllmRunner],
+    all_inputs: Sequence[tuple[list[str], PromptImageInput]],
+    model: str,
+    max_model_len: int,
+    max_num_seqs: int,
+    mm_limit: int,
+    gpu_memory_utilization: float,
+):
+    """Load each runner once, run all inputs, then compare."""
+    # NOTE: run vLLM first, then HF.  vLLM needs a fresh process without
+    # cuda initialization; running HF first would break the multiprocessing
+    # backend with fork method.
+    with vllm_runner(
+        model,
+        runner="generate",
+        max_model_len=max_model_len,
+        max_num_seqs=max_num_seqs,
+        gpu_memory_utilization=gpu_memory_utilization,
+        dtype=DTYPE,
+        limit_mm_per_prompt={"image": mm_limit},
+        tensor_parallel_size=2,
+        trust_remote_code=True,
+        enforce_eager=True,
+    ) as vllm_model:
+        vllm_outputs_per_case = [
+            vllm_model.generate_greedy_logprobs(
+                prompts,
+                MAX_TOKENS,
+                num_logprobs=NUM_LOGPROBS,
+                images=images,
+            )
+            for prompts, images in all_inputs
+        ]
+
+    hf_model_kwargs = {"_attn_implementation": "sdpa", "device_map": "auto"}
+    with hf_runner(
+        model,
+        dtype=DTYPE,
+        model_kwargs=hf_model_kwargs,
+        auto_cls=AutoModelForCausalLM,
+        trust_remote_code=True,
+    ) as hf_model:
+        hf_outputs_per_case = [
+            hf_model.generate_greedy_logprobs_limit(
+                prompts,
+                MAX_TOKENS,
+                num_logprobs=NUM_LOGPROBS,
+                images=images,
+            )
+            for prompts, images in all_inputs
+        ]
+
+    for hf_outputs, vllm_outputs in zip(hf_outputs_per_case, vllm_outputs_per_case):
+        check_logprobs_close(
+            outputs_0_lst=hf_outputs,
+            outputs_1_lst=vllm_outputs,
+            name_0="hf",
+            name_1="vllm",
+        )
+
+
+@multi_gpu_test(num_gpus=2)
+@pytest.mark.parametrize("model", [MODEL_ID])
+def test_models(hf_runner, vllm_runner, image_assets, model) -> None:
+    all_inputs = _build_single_image_inputs(image_assets)
+    _run_and_compare(
+        hf_runner,
+        vllm_runner,
+        all_inputs,
+        model,
+        max_model_len=8192,
+        max_num_seqs=2,
+        mm_limit=1,
+        gpu_memory_utilization=0.80,
+    )
+
+
+@multi_gpu_test(num_gpus=2)
+@pytest.mark.parametrize("model", [MODEL_ID])
+def test_multi_images_models(hf_runner, vllm_runner, image_assets, model) -> None:
+    all_inputs = _build_multi_image_inputs(image_assets)
+    _run_and_compare(
+        hf_runner,
+        vllm_runner,
+        all_inputs,
+        model,
+        max_model_len=8192,
+        max_num_seqs=2,
+        mm_limit=2,
+        gpu_memory_utilization=0.80,
+    )

tests/models/multimodal/generation/test_voxtral.py

@@ -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.
 

tests/models/multimodal/generation/vlm_utils/core.py

@@ -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,

tests/models/multimodal/pooling/test_colqwen3.py

@@ -22,6 +22,11 @@ from vllm.entrypoints.pooling.score.utils 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",

tests/models/multimodal/pooling/test_intern_vit.py

@@ -11,6 +11,11 @@ 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"]

tests/models/multimodal/pooling/test_jinavl_reranker.py

@@ -15,6 +15,11 @@ from vllm.entrypoints.pooling.score.utils 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 = {

tests/models/multimodal/processing/test_musicflamingo.py

@@ -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,

tests/models/registry.py

@@ -334,7 +334,15 @@ _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
@@ -469,6 +477,13 @@ _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",
@@ -509,6 +524,13 @@ _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",
@@ -544,6 +566,11 @@ _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),
@@ -754,10 +781,18 @@ _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"),
@@ -800,9 +835,30 @@ _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
@@ -844,6 +900,13 @@ _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",
@@ -863,7 +926,12 @@ _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",
@@ -952,7 +1020,14 @@ _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={
@@ -960,6 +1035,13 @@ _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(
@@ -996,13 +1078,25 @@ _MULTIMODAL_EXAMPLE_MODELS = {
         "nano_vl_dummy", is_available_online=False, 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",
@@ -1014,12 +1108,24 @@ _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",
@@ -1038,6 +1144,19 @@ _MULTIMODAL_EXAMPLE_MODELS = {
         },  # noqa: E501
         extras={"phi3.5": "microsoft/Phi-3.5-vision-instruct"},
     ),
+    "Phi4ForCausalLMV": _HfExamplesInfo(
+        "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
     ),
@@ -1133,6 +1252,14 @@ _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",

tests/models/utils.py

@@ -468,7 +468,16 @@ 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,

tests/reasoning/test_step3p5_reasoning_parser.py

@@ -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 = {

tests/v1/e2e/spec_decode/test_spec_decode.py

@@ -542,12 +542,16 @@ 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(
             (

vllm/model_executor/model_loader/gguf_loader.py

@@ -209,12 +209,24 @@ class GGUFModelLoader(BaseModelLoader):
                 GGUF tensor name with suffix (e.g., 'mm.soft_emb_norm.weight')
                 or None if no mapping found
             """
+            # In transformers v5, multimodal models (e.g. Gemma3) wrap
+            # all sub-models under an outer 'model.' attribute, producing
+            # state_dict keys like 'model.language_model.layers.0...' and
+            # 'model.vision_tower.vision_model...'.  Strip this outer
+            # prefix so the keys match what gguf-py expects.
+            if is_multimodal and hf_name.startswith("model."):
+                hf_name = hf_name[6:]  # Remove outer 'model.'
+
             # Strip 'language_model.' prefix for multimodal models - gguf-py
             # tensor mappings expect parameter names without this prefix.
             # Note: 'model.' prefix should be KEPT for text-only models as
             # gguf-py expects it.
             if hf_name.startswith("language_model."):
                 hf_name = hf_name[15:]  # Remove 'language_model.'
+                # Re-add 'model.' prefix because gguf-py text tensor maps
+                # expect 'model.layers...' format.
+                if is_multimodal:
+                    hf_name = "model." + hf_name
 
             # Parse parameter name and suffix
             if hf_name.endswith((".weight", ".bias")):

vllm/model_executor/models/gemma4_mm.py

@@ -126,8 +126,12 @@ class Gemma4AudioInputs(TensorSchema):
     """
 
     type: Literal["audio"] = "audio"
-    input_features_padded: Annotated[torch.Tensor, TensorShape("bn", "s", "f")]
-    input_features_mask: Annotated[torch.Tensor, TensorShape("bn", "s")]
+    input_features_padded: Annotated[
+        torch.Tensor, TensorShape("bn", "s", "f", dynamic_dims={"s"})
+    ]
+    input_features_mask: Annotated[
+        torch.Tensor, TensorShape("bn", "s", dynamic_dims={"s"})
+    ]
 
 
 Gemma4ImageInputs = Gemma4ImagePixelInputs
@@ -513,6 +517,8 @@ class Gemma4MultiModalProcessor(BaseMultiModalProcessor[Gemma4ProcessingInfo]):
             video_timestamps_per_video: list[list[float]] = []
             video_frame_counts: list[int] = []
 
+            video_replacements: list[str] = []
+
             for item in videos:
                 video_array, metadata = item
 
@@ -565,10 +571,7 @@ class Gemma4MultiModalProcessor(BaseMultiModalProcessor[Gemma4ProcessingInfo]):
                 video_timestamps_per_video.append(timestamps)
                 video_frame_counts.append(len(frames))
 
-                # Build expanded replacement text and replace the
-                # <|video|> placeholder in the prompt.
-                # Use split(token, 1) to avoid collision — the
-                # replacement text itself contains <|video|> tokens.
+                # Build expanded replacement text for this video.
                 ts_strs = [f"{int(s // 60):02d}:{int(s % 60):02d}" for s in timestamps]
                 replacement = " ".join(
                     f"{t} {processor.boi_token}"
@@ -576,9 +579,23 @@ class Gemma4MultiModalProcessor(BaseMultiModalProcessor[Gemma4ProcessingInfo]):
                     f"{processor.eoi_token}"
                     for t, n in zip(ts_strs, num_soft_per_frame)
                 )
-                parts = prompt.split(processor.video_token, 1)
-                if len(parts) == 2:
-                    prompt = parts[0] + replacement + parts[1]
+                video_replacements.append(replacement)
+
+            # Replace all <|video|> placeholders at once. We split on
+            # video_token to get N+1 parts, then interleave with the
+            # N replacement strings. This avoids the iterative
+            # split-replace bug where replacement text (which itself
+            # contains <|video|> tokens) collides with later splits.
+            vt = processor.video_token
+            parts = prompt.split(vt, len(video_replacements))
+
+            # NOTE: len(parts) <= len(video_replacements) + 1
+            parts_with_repl: list[str] = []
+            for part, repl in zip(parts, video_replacements):
+                parts_with_repl.extend([part, repl])
+            parts_with_repl.extend(parts[len(video_replacements) :])
+
+            prompt = "".join(parts_with_repl)
 
             video_outputs = {
                 "pixel_values_videos": torch.cat(all_video_pixel_values, dim=0),
@@ -641,19 +658,23 @@ class Gemma4MultiModalProcessor(BaseMultiModalProcessor[Gemma4ProcessingInfo]):
             )
 
         if "input_features" in processed_outputs:
-            # Keep padded features for batched audio tower execution.
-            processed_outputs["input_features_padded"] = processed_outputs[
-                "input_features"
-            ]
-            # Unpad per-item so each item's cache entry is self-contained.
+            # Unpad per-item so each item's cache entry is
+            # self-contained. The batched() field config in
+            # _get_mm_fields_config will re-pad all fields to the
+            # batch's max length at batch time, ensuring consistent
+            # padding regardless of cache history.
+            masks = processed_outputs["input_features_mask"]
             unpadded_features = [
                 f[mask]
                 for f, mask in zip(
                     processed_outputs["input_features"],
-                    processed_outputs["input_features_mask"],
+                    masks,
                 )
             ]
+            unpadded_masks = [mask[mask] for mask in masks]
             processed_outputs["input_features"] = unpadded_features
+            processed_outputs["input_features_padded"] = unpadded_features
+            processed_outputs["input_features_mask"] = unpadded_masks
 
         # Merge video outputs into the final result
         combined_outputs = dict(processed_outputs, **video_outputs)

vllm/model_executor/models/musicflamingo.py

@@ -32,9 +32,9 @@ from transformers.models.musicflamingo import (
 
 from vllm.config import VllmConfig
 from vllm.config.multimodal import BaseDummyOptions
+from vllm.inputs import MultiModalDataDict
 from vllm.multimodal import MULTIMODAL_REGISTRY
 from vllm.multimodal.inputs import (
-    MultiModalDataDict,
     MultiModalFieldConfig,
     MultiModalKwargsItems,
 )

vllm/model_executor/models/transformers/base.py

@@ -16,6 +16,7 @@
 # limitations under the License.
 """Transformers modeling backend base class."""
 
+import sys
 from collections.abc import Callable, Iterable
 from itertools import chain
 from operator import attrgetter
@@ -29,6 +30,7 @@ from torch import nn
 from transformers import AutoModel
 from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS
 
+from vllm.compilation.decorators import support_torch_compile
 from vllm.config.utils import getattr_iter
 from vllm.distributed import get_pp_group, get_tp_group
 from vllm.distributed.utils import get_pp_indices
@@ -47,6 +49,7 @@ from vllm.model_executor.models.interfaces import (
 )
 from vllm.model_executor.models.interfaces_base import VllmModel
 from vllm.model_executor.models.transformers.utils import (
+    can_enable_torch_compile,
     get_feature_request_tip,
     init_on_device_without_buffers,
     log_replacement,
@@ -117,6 +120,7 @@ class Base(
         self.config = vllm_config.model_config.hf_config
         self.text_config = self.config.get_text_config()
         self.cache_config = vllm_config.cache_config
+        self.compilation_config = vllm_config.compilation_config
         self.device_config = vllm_config.device_config
         self.model_config = vllm_config.model_config
         self.parallel_config = vllm_config.parallel_config
@@ -146,7 +150,7 @@ class Base(
         if self.quant_config:
             quant_method_name = self.quant_config.get_name()
             # Check for unsupported quantization methods.
-            if quant_method_name == "mxfp4":
+            if quant_method_name in ("mxfp4", "gpt_oss_mxfp4"):
                 raise NotImplementedError(
                     "Transformers modeling backend does "
                     "not support MXFP4 quantization yet."
@@ -155,14 +159,16 @@ class Base(
             if "gptq" in quant_method_name:
                 self.ignore_unexpected_suffixes.append(".bias")
 
-        # Patch config and init on "meta" to delay allocating GPU tensors
         self._patch_config()
+        from_config_kwargs = dict(
+            config=self.config,
+            dtype=self.model_config.dtype,
+            trust_remote_code=self.model_config.trust_remote_code,
+        )
+        self._decorate_for_torch_compile(**from_config_kwargs)
+        # Init on "meta" to delay allocating GPU tensors
         with init_on_device_without_buffers("meta"):
-            self.model: PreTrainedModel = AutoModel.from_config(
-                self.config,
-                dtype=self.model_config.dtype,
-                trust_remote_code=self.model_config.trust_remote_code,
-            )
+            self.model: PreTrainedModel = AutoModel.from_config(**from_config_kwargs)
 
         # Create weight name to module qualname mapper
         self._create_hf_to_vllm_mapper()
@@ -218,6 +224,87 @@ class Base(
             if sub_config.dtype != (dtype := self.config.dtype):
                 sub_config.dtype = dtype
 
+    def _get_decoder_cls(self, **kwargs: dict) -> type[PreTrainedModel]:
+        """
+        Get the decoder class from the model.
+
+        Args:
+            kwargs: The kwargs to create the model.
+
+        Returns:
+            The decoder class.
+        """
+        with torch.device("meta"):
+            model: PreTrainedModel = AutoModel.from_config(**kwargs)
+        decoder_cls = type(model.get_decoder())
+        logger.debug("Identified decoder class as: %s", decoder_cls)
+        del model
+        return decoder_cls
+
+    def _decorate_cls_for_torch_compile(
+        self,
+        cls: type[PreTrainedModel],
+        dynamic_arg_dims: dict[str, int] | None,
+        enable_if: Callable[["VllmConfig"], bool],
+        is_encoder: bool,
+    ):
+        """
+        Decorate `cls` to indicate to vLLM that it supports torch compile.
+
+        Args:
+            cls: The PreTrainedModel class to decorate.
+            dynamic_arg_dims: A mapping from argument name to the dynamic dimensions
+                of the argument. If None, default dynamic arg dims will be used. See
+                [`support_torch_compile`][vllm.compilation.decorators.support_torch_compile]
+                for more details.
+            enable_if: A function which takes in the vLLM config and returns whether
+                torch compile should be enabled for this class.
+            is_encoder: Whether the class being decorated is an encoder.
+        """
+        logger.debug(
+            "Decorating `%s` as %s for torch compile with dynamic_arg_dims of %s",
+            cls.__name__,
+            "encoder" if is_encoder else "decoder",
+            dynamic_arg_dims,
+        )
+
+        @support_torch_compile(
+            dynamic_arg_dims=dynamic_arg_dims,
+            enable_if=enable_if,
+            is_encoder=is_encoder,
+        )
+        class SupportTorchCompileWrapper(cls): ...
+
+        # Preserve __module__ so transformers v5's source-file checks
+        # (e.g. _can_set_experts_implementation) read the original
+        # model's module instead of this file.
+        SupportTorchCompileWrapper.__module__ = cls.__module__
+
+        # Patch the class in its module
+        module = sys.modules[cls.__module__]
+        setattr(module, cls.__name__, SupportTorchCompileWrapper)
+
+    def _decorate_for_torch_compile(self, **kwargs: dict):
+        """
+        Decorate the model's decoder class to indicate to vLLM that it supports torch
+        compile if `can_enable_torch_compile` is True.
+
+        Args:
+            kwargs: The kwargs to create the model, which are needed to get the decoder
+                class.
+        """
+        self._decorate_cls_for_torch_compile(
+            cls=self._get_decoder_cls(**kwargs),
+            # Applied to a PreTrainedModel so the batch dimension will exist
+            dynamic_arg_dims=dict[str, int](
+                input_ids=1,  # shape: [1, seq_len]
+                inputs_embeds=1,  # shape: [1, seq_len, hidden_size]
+                position_ids=-1,  # shape: [1, seq_len] or [3, 1, seq_len] for mrope
+            ),
+            enable_if=can_enable_torch_compile,
+            is_encoder=False,
+        )
+
     def _create_hf_to_vllm_mapper(self):
         """
         Create a WeightsMapper to map checkpoint weight names to module qualnames.
@@ -553,11 +640,6 @@ class Base(
             input_ids = None
             inputs_embeds = intermediate_tensors["hidden_states"]
 
-        if input_ids is not None:
-            input_ids = input_ids[None, ...]
-        if inputs_embeds is not None:
-            inputs_embeds = inputs_embeds[None, ...]
-
         # If the model scales embeddings inside the input embedding layer we must
         # ensure they are scaled here since VocabParallelEmbedding will not do it
         if (
@@ -568,22 +650,29 @@ class Base(
             inputs_embeds = self.embed_input_ids(input_ids)
             input_ids = None
 
-        if self.model_config.uses_mrope:
-            position_ids = positions[:, None]
-        else:
-            position_ids = positions[None, ...]
+        # Add batch dimension before entering Transformers model
+        if input_ids is not None and input_ids.ndim == 1:
+            # [seq_len] -> [1, seq_len]
+            input_ids = input_ids[None, ...]
+        if inputs_embeds is not None and inputs_embeds.ndim == 2:
+            # [seq_len, hidden_size] -> [1, seq_len, hidden_size]
+            inputs_embeds = inputs_embeds[None, ...]
+        if positions.ndim == 1:
+            # [seq_len] -> [1, seq_len]
+            positions = positions[None, ...]
 
         outputs = self.model(
             input_ids=input_ids,
             inputs_embeds=inputs_embeds,
             use_cache=False,
-            position_ids=position_ids,
+            position_ids=positions,
             attention_instances=self.attention_instances,
             return_dict=False,
             **self._output_aux_hidden_states_kwargs,
             **kwargs,
         )
-        # We must remove the batch dimension from these outputs
+
+        # Remove batch dimension after exiting Transformers model
         hidden_states = outputs[0][0, ...]
         if self._output_aux_hidden_states_kwargs:
             aux_hidden_states = [x[0][0, ...] for x in outputs[1:]]