Bump Transformers version to 5.10.4 (#41359)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
Harry Mellor
2026-07-07 05:13:28 -07:00
committed by GitHub
parent 7ff656cc8b
commit dd94484577
51 changed files with 271 additions and 234 deletions
+2 -2
View File
@@ -74,7 +74,7 @@ from torch._inductor.utils import fresh_cache
if TYPE_CHECKING:
from transformers import PreTrainedTokenizer, PreTrainedTokenizerFast
from transformers import PythonBackend, TokenizersBackend
from transformers.generation.utils import GenerateOutput
@@ -499,7 +499,7 @@ class HfRunner:
self.model = model
if not skip_tokenizer_init:
self.tokenizer: "PreTrainedTokenizer | PreTrainedTokenizerFast" = (
self.tokenizer: "PythonBackend | TokenizersBackend" = (
AutoTokenizer.from_pretrained(
tokenizer_name or model_name,
trust_remote_code=trust_remote_code,
@@ -369,7 +369,7 @@ async def test_chat_request(
assert output.object == "list"
assert len(output.data) == 1
assert output.model == MODEL_NAME
assert output.usage.prompt_tokens == 34
assert output.usage.prompt_tokens == 33
# test continue_final_message
response = requests.post(
@@ -401,7 +401,7 @@ async def test_chat_request(
assert output.object == "list"
assert len(output.data) == 1
assert output.model == MODEL_NAME
assert output.usage.prompt_tokens == 36
assert output.usage.prompt_tokens == 35
# test continue_final_message with add_generation_prompt
response = requests.post(
@@ -906,7 +906,6 @@ VLM_TEST_SETTINGS = {
max_model_len=4096,
use_tokenizer_eos=True,
auto_cls=AutoModelForImageTextToText,
hf_model_kwargs=model_utils.qianfan_ocr_hf_model_kwargs("baidu/Qianfan-OCR"),
),
"qwen2_vl": VLMTestInfo(
models=["Qwen/Qwen2-VL-2B-Instruct"],
@@ -1492,94 +1492,3 @@ def moondream3_patch_hf_runner(hf_model: HfRunner) -> HfRunner:
hf_model.model.generate = types.MethodType(_generate, hf_model.model)
return hf_model
def qianfan_ocr_hf_model_kwargs(model_name: str) -> dict:
"""Return hf_model_kwargs with a patched config for QianfanOCR."""
from vllm.transformers_utils.configs.qianfan_ocr import QianfanOCRConfig
config = QianfanOCRConfig.from_pretrained(model_name)
vc = config.vision_config
if isinstance(vc.image_size, int):
vc.image_size = (vc.image_size, vc.image_size)
if isinstance(vc.patch_size, int):
vc.patch_size = (vc.patch_size, vc.patch_size)
return {"config": config}
def qianfan_ocr_patch_hf_runner(hf_model: HfRunner) -> HfRunner:
"""Patches an HfRunner instance to run QianfanOCR model inference.
QianfanOCR shares the same architecture as InternVLChatModel, so the
patching logic mirrors ``internvl_patch_hf_runner``. The only difference
is that we load the config via vllm's registered ``QianfanOCRConfig``
instead of relying on ``trust_remote_code``.
"""
class QianfanOCRProcessor:
def __init__(self, hf_runner: HfRunner):
self.tokenizer = hf_runner.tokenizer
from vllm.transformers_utils.configs.qianfan_ocr import QianfanOCRConfig
self.config = QianfanOCRConfig.from_pretrained(hf_runner.model_name)
self.vision_config = self.config.vision_config
self.use_thumbnail = self.config.use_thumbnail
self.min_num = self.config.min_dynamic_patch
self.max_num = self.config.max_dynamic_patch
self.image_size = self.vision_config.image_size
# Compute num_image_token from config instead of model attribute,
# since the transformers-native model doesn't expose it.
image_size = self.config.force_image_size or self.vision_config.image_size
patch_size = self.vision_config.patch_size
downsample_ratio = self.config.downsample_ratio
self.num_image_token = int(
(image_size // patch_size) ** 2 * (downsample_ratio**2)
)
def __call__(
self,
text: str,
images: PIL.Image.Image | list[PIL.Image.Image] = None,
**kwargs,
):
from vllm.transformers_utils.processors.internvl import (
image_to_pixel_values_internvl,
)
IMG_START = "<img>"
IMG_END = "</img>"
IMG_CONTEXT = "<IMG_CONTEXT>"
images = [images] if isinstance(images, PIL.Image.Image) else images
pixel_values_list = [
image_to_pixel_values_internvl(
image,
input_size=self.image_size,
min_num=self.min_num,
max_num=self.max_num,
use_thumbnail=self.use_thumbnail,
)
for image in images
]
num_patches_list = [pv.shape[0] for pv in pixel_values_list]
pixel_values = torch.cat(pixel_values_list, dim=0)
for num_patches in num_patches_list:
context_tokens = IMG_CONTEXT * self.num_image_token * num_patches
image_tokens = IMG_START + context_tokens + IMG_END
text = text.replace("<image>", image_tokens, 1)
prompt = self.tokenizer(text, return_tensors="pt")
prompt.update({"pixel_values": pixel_values})
return prompt
img_context_token_id = hf_model.tokenizer.convert_tokens_to_ids("<IMG_CONTEXT>")
hf_model.model.img_context_token_id = img_context_token_id
hf_model.processor = QianfanOCRProcessor(hf_model)
hf_model.model.get_output_embeddings = (
lambda: hf_model.model.language_model.get_output_embeddings()
)
hf_model.model.generate = types.MethodType(_internvl_generate, hf_model.model)
return hf_model
+1 -3
View File
@@ -202,9 +202,7 @@ _TEXT_GENERATION_EXAMPLE_MODELS = {
"AfmoeForCausalLM": _HfExamplesInfo("arcee-ai/Trinity-Nano-Preview"),
"ApertusForCausalLM": _HfExamplesInfo("swiss-ai/Apertus-8B-Instruct-2509"),
"ArceeForCausalLM": _HfExamplesInfo("arcee-ai/AFM-4.5B-Base"),
"ArcticForCausalLM": _HfExamplesInfo(
"Snowflake/snowflake-arctic-instruct", trust_remote_code=True
),
"ArcticForCausalLM": _HfExamplesInfo("Snowflake/snowflake-arctic-instruct"),
"AXK1ForCausalLM": _HfExamplesInfo("skt/A.X-K1", trust_remote_code=True),
"BailingMoeForCausalLM": _HfExamplesInfo(
"inclusionAI/Ling-lite-1.5", trust_remote_code=True
+1
View File
@@ -486,6 +486,7 @@ def dummy_hf_overrides(
"Gemma3nForConditionalGeneration",
"Gemma4ForCausalLM",
"Gemma4ForConditionalGeneration",
"Gemma4MTPModel",
"DiffusionGemmaForBlockDiffusion",
)
else 1
+3 -3
View File
@@ -5,7 +5,7 @@ from typing import _get_protocol_attrs # type: ignore
import pytest
from transformers import (
PreTrainedTokenizerBase,
PreTrainedTokenizerFast,
TokenizersBackend,
)
from vllm.tokenizers import TokenizerLike, get_tokenizer
@@ -24,7 +24,7 @@ def _assert_tokenizer_like(tokenizer: object):
def test_tokenizer_like_protocol():
tokenizer = get_tokenizer("openai-community/gpt2", use_fast=True)
assert isinstance(tokenizer, PreTrainedTokenizerFast)
assert isinstance(tokenizer, TokenizersBackend)
_assert_tokenizer_like(tokenizer)
tokenizer = get_tokenizer(
@@ -38,7 +38,7 @@ def test_tokenizer_like_protocol():
assert isinstance(tokenizer, HfTokenizer)
# Verify it's a fast tokenizer (required for FastIncrementalDetokenizer)
assert isinstance(tokenizer, PreTrainedTokenizerFast)
assert isinstance(tokenizer, TokenizersBackend)
assert "DSV32" in tokenizer.__class__.__name__
_assert_tokenizer_like(tokenizer)
+5 -5
View File
@@ -5,7 +5,7 @@ from collections.abc import Generator
from typing import Any
import pytest
from transformers import AutoTokenizer, PreTrainedTokenizer, PreTrainedTokenizerFast
from transformers import AutoTokenizer, PythonBackend, TokenizersBackend
from vllm.sampling_params import SamplingParams
from vllm.tokenizers.mistral import MistralTokenizer
@@ -153,13 +153,13 @@ def test_decode_streaming(
spaces_between_special_tokens,
fast,
):
if fast and not isinstance(tokenizer, PreTrainedTokenizerFast):
if fast and not isinstance(tokenizer, TokenizersBackend):
pytest.skip()
if skip_special_tokens and not spaces_between_special_tokens:
pytest.skip()
if not fast and isinstance(tokenizer, PreTrainedTokenizerFast):
if not fast and isinstance(tokenizer, TokenizersBackend):
# Fix up inconsistency in fast/slow tokenizer behaviour.
tokenizer.add_special_tokens(
{
@@ -173,7 +173,7 @@ def test_decode_streaming(
extra_decode_args = (
{}
if not isinstance(tokenizer, PreTrainedTokenizer)
if not isinstance(tokenizer, PythonBackend)
else {"spaces_between_special_tokens": spaces_between_special_tokens}
)
@@ -225,7 +225,7 @@ def test_decode_streaming(
@pytest.mark.parametrize("tokenizer_name", TOKENIZERS)
@pytest.mark.parametrize("fast", (True, False))
def test_oov_decode(tokenizer, fast):
if fast and not isinstance(tokenizer, PreTrainedTokenizerFast):
if fast and not isinstance(tokenizer, TokenizersBackend):
pytest.skip()
decoded_text, out_ids = _run_incremental_decode(
+5 -5
View File
@@ -7,14 +7,14 @@ from typing import TypeAlias
import numpy as np
import torch
from transformers import PreTrainedTokenizer, PreTrainedTokenizerFast
from transformers import PythonBackend, TokenizersBackend
from vllm.engine.arg_utils import EngineArgs
from vllm.v1.engine import EngineCoreOutput, FinishReason
from vllm.v1.metrics.stats import PrefillStats
from vllm.v1.outputs import LogprobsLists, LogprobsTensors
GeneralTokenizerType: TypeAlias = PreTrainedTokenizer | PreTrainedTokenizerFast
GeneralTokenizerType: TypeAlias = PythonBackend | TokenizersBackend
# Number of sample logprobs to request when testing sample logprobs
NUM_SAMPLE_LOGPROBS_UNDER_TEST = 5
@@ -193,7 +193,7 @@ def _create_random_top_token_test_matrix(
def decode_token(
tok_id: int,
tokenizer: PreTrainedTokenizer,
tokenizer: PythonBackend,
) -> str:
"""Reproduce the process of detokenizing a token for testing purposes.
@@ -210,7 +210,7 @@ def decode_token(
def generate_dummy_sample_logprobs(
sampled_tokens_list: list,
num_logprobs: int,
tokenizer: PreTrainedTokenizer,
tokenizer: PythonBackend,
) -> list[tuple[list[int], list[float], int]]:
"""Generate dummy sample logprobs
@@ -259,7 +259,7 @@ def generate_dummy_sample_logprobs(
def generate_dummy_prompt_logprobs_tensors(
prompt_tokens_list: list,
num_logprobs: int,
tokenizer: PreTrainedTokenizer,
tokenizer: PythonBackend,
) -> LogprobsTensors:
"""Generate dummy prompt logprobs tensors