Deprecate Transformers v4 support (#45161)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
Harry Mellor
2026-06-11 11:04:01 +08:00
committed by GitHub
parent b038a2f73b
commit 18d87a87dc
19 changed files with 61 additions and 267 deletions
@@ -77,30 +77,13 @@ logger = init_logger(__name__)
temp_dir = tempfile.gettempdir()
def enable_hf_transfer():
"""automatically activates hf_transfer"""
if "HF_HUB_ENABLE_HF_TRANSFER" not in os.environ:
try:
# enable hf hub transfer if available
import hf_transfer # type: ignore # noqa
huggingface_hub.constants.HF_HUB_ENABLE_HF_TRANSFER = True
except ImportError:
pass
def enable_xet_high_performance():
"""automatically activates xet high performance mode"""
if "HF_XET_HIGH_PERFORMANCE" not in os.environ:
huggingface_hub.constants.HF_XET_HIGH_PERFORMANCE = True
if hasattr(huggingface_hub.constants, "HF_XET_HIGH_PERFORMANCE"):
# Transformers v5
enable_xet_high_performance()
else:
# Transformers v4
enable_hf_transfer()
enable_xet_high_performance()
class DisabledTqdm(tqdm):
+2 -7
View File
@@ -618,13 +618,8 @@ class Gemma3nForConditionalGeneration(
input_features = audio_input["input_features_padded"].squeeze(1)
input_features_mask = audio_input["input_features_mask"].squeeze(1)
audio_outputs = self.audio_tower(input_features, ~input_features_mask)
if isinstance(audio_outputs, tuple):
# Transformers v4
audio_encodings, audio_mask = audio_outputs
else:
# Transformers v5
audio_encodings = audio_outputs.last_hidden_state
audio_mask = audio_outputs.audio_mel_mask
audio_encodings = audio_outputs.last_hidden_state
audio_mask = audio_outputs.audio_mel_mask
audio_features = self.embed_audio(inputs_embeds=audio_encodings)
# The Gemma3nProcessor expects all audio will be 30s in length and
@@ -30,9 +30,7 @@ import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from packaging.version import Version
from transformers import PretrainedConfig
from transformers import __version__ as TRANSFORMERS_VERSION
from transformers.feature_extraction_utils import BatchFeature
from transformers.models.qwen3_omni_moe.configuration_qwen3_omni_moe import (
Qwen3OmniMoeAudioEncoderConfig,
@@ -1261,40 +1259,6 @@ class Qwen3OmniMoeThinkerMultiModalProcessor(
tok_kwargs = dict(tok_kwargs)
mm_kwargs["audio_kwargs"] = dict(mm_kwargs.get("audio_kwargs") or {})
mm_kwargs["text_kwargs"] = dict(mm_kwargs.get("text_kwargs") or {})
if Version(TRANSFORMERS_VERSION) < Version("4.58.0"):
# Extract audio_sample_rate before restructuring
audio_sample_rate = mm_kwargs.pop("audio_sample_rate", None)
# move truncation to audio_kwargs level to avoid conflict
# with tok_kwargs
mm_kwargs["audio_kwargs"].setdefault(
"truncation", mm_kwargs.pop("truncation", False)
)
mm_kwargs["text_kwargs"].setdefault(
"truncation", tok_kwargs.pop("truncation", False)
)
# Validate and conditionally pass audio_sample_rate
# WhisperFeatureExtractor has a fixed sampling rate, and vLLM's
# audio loader already resamples audio to the target rate.
# Only pass the value if it matches to avoid unexpected behavior.
if audio_sample_rate is not None:
expected_sr = feature_extractor.sampling_rate
if audio_sample_rate != expected_sr:
logger.warning(
"[%s] audio_sample_rate mismatch: user provided %dHz "
"but model expects %dHz. Ignoring user value. "
"vLLM's audio loader already resampled to %dHz.",
self.__class__.__name__,
audio_sample_rate,
expected_sr,
expected_sr,
)
else:
# Sample rate matches, safe to pass
mm_kwargs["audio_kwargs"]["audio_sample_rate"] = (
audio_sample_rate
)
hf_inputs = super()._call_hf_processor(
prompt=prompt,
+16 -42
View File
@@ -27,6 +27,10 @@ import transformers
from packaging.version import Version
from torch import nn
from transformers import AutoModel
from transformers.conversion_mapping import (
WeightRenaming,
get_model_conversion_mapping,
)
from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS
from vllm.compilation.decorators import support_torch_compile
@@ -212,16 +216,9 @@ class Base(
`create_attention_instances` are used
- Sets the dtype to the default torch dtype set by vLLM because Transformers
uses the config dtype when creating the model
- Propagates this dtype to any sub-configs because Transformers model
implementations do not support/use different dtypes in sub-models
"""
self.text_config._attn_implementation = "vllm"
self.config.dtype = torch.get_default_dtype()
# TODO(hmellor): Remove this when Transformers v4 support is dropped
for sub_config_name in getattr(self.config, "sub_configs", {}):
sub_config = getattr(self.config, sub_config_name)
if sub_config.dtype != (dtype := self.config.dtype):
sub_config.dtype = dtype
def _get_decoder_cls(self, **kwargs: dict) -> type[PreTrainedModel]:
"""
@@ -300,9 +297,7 @@ class Base(
This handles:
- Transformers weight renaming:
- from `WeightRenaming` in Transformers v5
- from `_checkpoint_conversion_mapping` in Transformers v4
- Transformers weight renaming from `WeightRenaming`
- Checkpoints saved with a base model prefix that is not `model`
- Checkpoints saved with no base model prefix
- Any quantization config specific mappings
@@ -310,37 +305,16 @@ class Base(
self.hf_to_vllm_mapper = WeightsMapper()
orig_to_new_regex = self.hf_to_vllm_mapper.orig_to_new_regex
if Version(transformers.__version__) >= Version("5.0.0"):
from transformers.conversion_mapping import (
WeightRenaming,
get_model_conversion_mapping,
)
for mapping in get_model_conversion_mapping(self.model):
# Handle weights which have been renamed in Transformers
if isinstance(mapping, WeightRenaming):
# Recompile using regex (Transformers used re)
compiled_sources = re.compile(
mapping.compiled_sources.pattern, mapping.compiled_sources.flags
)
target_pattern = mapping.target_patterns[0]
orig_to_new_regex[compiled_sources] = target_pattern
# TODO: Handle WeightConverter to enable layer merging
else:
# Replace legacy suffixes used for norms
# TODO(hmellor): Remove this when Transformers v4 support is dropped
orig_to_new_regex.update(
{
re.compile(r"\.gamma$"): ".weight",
re.compile(r"\.beta$"): ".bias",
}
)
# Handle weights which have been renamed in Transformers
# TODO(hmellor): Remove this when Transformers v4 support is dropped
ccm = getattr(self.model, "_checkpoint_conversion_mapping", {})
for source, target in ccm.items():
orig_to_new_regex[re.compile(source)] = target
for mapping in get_model_conversion_mapping(self.model):
# Handle weights which have been renamed in Transformers
if isinstance(mapping, WeightRenaming):
# Recompile using regex (Transformers used re)
compiled_sources = re.compile(
mapping.compiled_sources.pattern, mapping.compiled_sources.flags
)
target_pattern = mapping.target_patterns[0]
orig_to_new_regex[compiled_sources] = target_pattern
# TODO: Handle WeightConverter to enable layer merging
# Handle unexpected weights which should be ignored
if self.model._keys_to_ignore_on_load_unexpected is not None:
@@ -377,7 +351,7 @@ class Base(
"""
Check if the model has tied word embeddings.
"""
# Transformers v4 and v5 will store this in different places
# Models created with Transformers v4 and v5 will store this in different places
tie_word_embeddings_v4 = getattr(self.text_config, "tie_word_embeddings", False)
tie_word_embeddings_v5 = getattr(self.config, "tie_word_embeddings", False)
return tie_word_embeddings_v4 or tie_word_embeddings_v5
@@ -101,8 +101,6 @@ Style = Literal[
"replicate",
"colwise_gather_output",
"rowwise_split_input",
"colwise_rep",
"rowwise_rep",
]
@@ -131,12 +129,8 @@ def replace_linear_class(
"colwise": (ColumnParallelLinear, {}),
"rowwise": (RowParallelLinear, {}),
"replicate": (ReplicatedLinear, {}),
# Transformers v5
"colwise_gather_output": (ColumnParallelLinear, {"gather_output": True}),
"rowwise_split_input": (RowParallelLinear, {"input_is_parallel": False}),
# Transformers v4
"colwise_rep": (ColumnParallelLinear, {"gather_output": True}),
"rowwise_rep": (RowParallelLinear, {"input_is_parallel": False}),
}.get(style, (ReplicatedLinear, {}))
return vllm_linear_cls(
-15
View File
@@ -5,7 +5,6 @@
"""PyTorch Ultravox model."""
import copy
import inspect
from collections.abc import Iterable, Mapping, Sequence
from types import SimpleNamespace
from typing import Annotated, Any, Literal, TypeAlias
@@ -397,17 +396,10 @@ class UltravoxTransformerProjector(nn.Module, ModuleUtilsMixin):
)
hidden_states = hidden_states + positions
# Backward compatibility for Transformers v4 where layer_head_mask
# was a required argument for WhisperEncoderLayer.forward
kwargs = {}
if "layer_head_mask" in inspect.signature(self.layers[0].forward).parameters:
kwargs["layer_head_mask"] = None
for layer in self.layers:
hidden_states = layer(
hidden_states,
attention_mask=extended_attention_mask,
**kwargs,
)
# BC version that allows for the old tupled output
if isinstance(hidden_states, tuple):
@@ -504,17 +496,10 @@ class ModifiedWhisperEncoder(WhisperEncoder):
attention_mask = self.get_attention_mask_by_audio_len(audio_lens, hidden_states)
# Backward compatibility for Transformers v4 where layer_head_mask
# was a required argument for WhisperEncoderLayer.forward
kwargs = {}
if "layer_head_mask" in inspect.signature(self.layers[0].forward).parameters:
kwargs["layer_head_mask"] = None
for encoder_layer in self.layers:
hidden_states = encoder_layer(
hidden_states,
attention_mask,
**kwargs,
)
# BC version that allows for the old tupled output
if isinstance(hidden_states, tuple):