Deprecate Transformers v4 support (#45161)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
@@ -16,6 +16,7 @@ from huggingface_hub import constants
|
||||
from packaging.version import Version
|
||||
from safetensors.torch import _TYPES as _SAFETENSORS_TO_TORCH_DTYPE
|
||||
from transformers import GenerationConfig, PretrainedConfig
|
||||
from transformers.configuration_utils import ALLOWED_LAYER_TYPES
|
||||
from transformers.models.auto.image_processing_auto import get_image_processor_config
|
||||
from transformers.models.auto.modeling_auto import (
|
||||
MODEL_FOR_CAUSAL_LM_MAPPING_NAMES,
|
||||
@@ -49,15 +50,6 @@ from .repo_utils import (
|
||||
with_retry,
|
||||
)
|
||||
|
||||
try:
|
||||
# Transformers v5
|
||||
from transformers.configuration_utils import ALLOWED_ATTENTION_LAYER_TYPES
|
||||
except ImportError:
|
||||
# Transformers v4
|
||||
from transformers.configuration_utils import (
|
||||
ALLOWED_LAYER_TYPES as ALLOWED_ATTENTION_LAYER_TYPES,
|
||||
)
|
||||
|
||||
if envs.VLLM_USE_MODELSCOPE:
|
||||
from modelscope import AutoConfig
|
||||
else:
|
||||
@@ -68,9 +60,8 @@ MISTRAL_CONFIG_NAME = "params.json"
|
||||
logger = init_logger(__name__)
|
||||
|
||||
if Version(version("transformers")) < Version("5.0.0"):
|
||||
logger.warning(
|
||||
"Support for Transformers v4 is deprecated. The Transformers v4 codepath will "
|
||||
"become unmaintained in vLLM v0.22.0 and will be removed in vLLM v0.24.0. "
|
||||
raise ImportError(
|
||||
"Support for Transformers v4 is deprecated and was removed in vLLM v0.24.0. "
|
||||
"Please upgrade to Transformers v5: pip install --upgrade transformers"
|
||||
)
|
||||
|
||||
@@ -159,7 +150,7 @@ def is_rope_parameters_nested(rope_parameters: dict[str, Any]) -> bool:
|
||||
# Cannot be nested if rope_parameters is empty
|
||||
if not rope_parameters:
|
||||
return False
|
||||
return set(rope_parameters.keys()).issubset(ALLOWED_ATTENTION_LAYER_TYPES)
|
||||
return set(rope_parameters.keys()).issubset(ALLOWED_LAYER_TYPES)
|
||||
|
||||
|
||||
@contextmanager
|
||||
@@ -183,24 +174,21 @@ def _patch_hf_transformers_validate_rope():
|
||||
hf transformers (from v5 onwards)
|
||||
"""
|
||||
|
||||
if Version(version("transformers")) >= Version("5.0.0"):
|
||||
if hasattr(PretrainedConfig.validate_rope, "__vllm_patched__"):
|
||||
return
|
||||
if hasattr(PretrainedConfig.validate_rope, "__vllm_patched__"):
|
||||
return
|
||||
|
||||
_original_validate_rope = PretrainedConfig.validate_rope
|
||||
_original_validate_rope = PretrainedConfig.validate_rope
|
||||
|
||||
@wraps(_original_validate_rope)
|
||||
def patched_validate_rope(self, *args, **kwargs):
|
||||
ignore_keys_param = kwargs.pop("ignore_keys", None)
|
||||
original_ignore_keys = self.ignore_keys_at_rope_validation
|
||||
self.ignore_keys_at_rope_validation = (
|
||||
original_ignore_keys or ignore_keys_param
|
||||
)
|
||||
result = _original_validate_rope(self, *args, **kwargs)
|
||||
return result
|
||||
@wraps(_original_validate_rope)
|
||||
def patched_validate_rope(self, *args, **kwargs):
|
||||
ignore_keys_param = kwargs.pop("ignore_keys", None)
|
||||
original_ignore_keys = self.ignore_keys_at_rope_validation
|
||||
self.ignore_keys_at_rope_validation = original_ignore_keys or ignore_keys_param
|
||||
result = _original_validate_rope(self, *args, **kwargs)
|
||||
return result
|
||||
|
||||
patched_validate_rope.__vllm_patched__ = True # type: ignore[attr-defined]
|
||||
PretrainedConfig.validate_rope = patched_validate_rope
|
||||
patched_validate_rope.__vllm_patched__ = True # type: ignore[attr-defined]
|
||||
PretrainedConfig.validate_rope = patched_validate_rope
|
||||
|
||||
|
||||
class HFConfigParser(ConfigParserBase):
|
||||
@@ -493,39 +481,13 @@ def patch_rope_parameters(config: PretrainedConfig) -> None:
|
||||
"""Provide backwards compatibility for RoPE."""
|
||||
from vllm.config.utils import getattr_iter
|
||||
|
||||
# Older custom models may use non-standard field names
|
||||
# which need patching for both Transformers v4 and v5.
|
||||
# Older custom models may use non-standard field names which need patching.
|
||||
names = ["rope_theta", "rotary_emb_base"]
|
||||
rope_theta = getattr_iter(config, names, None, warn=True)
|
||||
names = ["partial_rotary_factor", "rotary_pct", "rotary_emb_fraction"]
|
||||
partial_rotary_factor = getattr_iter(config, names, None, warn=True)
|
||||
ompe = getattr(config, "original_max_position_embeddings", None)
|
||||
|
||||
if Version(version("transformers")) < Version("5.0.0"):
|
||||
# Transformers v4 installed, legacy config fields may be present.
|
||||
if is_rope_parameters_nested(getattr(config, "rope_parameters", {})):
|
||||
# Loading nested rope_parameters (from Transformers v5) in Transformers v4.
|
||||
# Skip legacy patching since it should already be in the correct format.
|
||||
pass
|
||||
else:
|
||||
if (rope_scaling := getattr(config, "rope_scaling", None)) is not None:
|
||||
config.rope_parameters = rope_scaling
|
||||
if (
|
||||
rope_theta is not None
|
||||
or partial_rotary_factor is not None
|
||||
or ompe is not None
|
||||
) and not getattr(config, "rope_parameters", None):
|
||||
config.rope_parameters = {"rope_type": "default"}
|
||||
# Patch legacy fields into rope_parameters
|
||||
if rope_theta is not None:
|
||||
config.rope_parameters["rope_theta"] = rope_theta
|
||||
if partial_rotary_factor is not None:
|
||||
config.rope_parameters["partial_rotary_factor"] = partial_rotary_factor
|
||||
if ompe is not None:
|
||||
config.rope_parameters["original_max_position_embeddings"] = ompe
|
||||
patch_legacy_rope_type(getattr(config, "rope_parameters", None))
|
||||
elif rope_theta is not None or getattr(config, "rope_parameters", None):
|
||||
# Transformers v5 installed
|
||||
if rope_theta is not None or getattr(config, "rope_parameters", None):
|
||||
# Patch these fields in case they used non-standard names
|
||||
if rope_theta is not None:
|
||||
config.rope_theta = rope_theta
|
||||
|
||||
@@ -3,6 +3,7 @@
|
||||
|
||||
# adapted from https://github.com/deepseek-ai/DeepSeek-VL2/blob/faf18023f24b962b32d9f0a2d89e402a8d383a78/deepseek_vl2/models/modeling_deepseek_vl_v2.py#L115-L268
|
||||
|
||||
from huggingface_hub.dataclasses import strict
|
||||
from transformers import DeepseekV2Config, PretrainedConfig
|
||||
|
||||
|
||||
@@ -87,16 +88,9 @@ class MlpProjectorConfig(PretrainedConfig):
|
||||
super().__init__(**kwargs)
|
||||
|
||||
|
||||
if hasattr(DeepseekV2Config, "validate"):
|
||||
# Transformers v5
|
||||
from huggingface_hub.dataclasses import strict
|
||||
|
||||
@strict
|
||||
class DeepseekVLV2TextConfig(DeepseekV2Config):
|
||||
kv_lora_rank: int | None = None
|
||||
else:
|
||||
# Transformers v4
|
||||
DeepseekVLV2TextConfig = DeepseekV2Config # type: ignore[misc]
|
||||
@strict
|
||||
class DeepseekVLV2TextConfig(DeepseekV2Config):
|
||||
kv_lora_rank: int | None = None
|
||||
|
||||
|
||||
class DeepseekVLV2Config(PretrainedConfig):
|
||||
|
||||
@@ -228,15 +228,8 @@ class OlmoHybridConfig(PretrainedConfig):
|
||||
if "full_attention" not in layer_types:
|
||||
layer_types[-1] = "full_attention"
|
||||
|
||||
if hasattr(self, "validate_layer_type"):
|
||||
# Transformers v5
|
||||
self.layer_types = layer_types
|
||||
self.validate_layer_type()
|
||||
else:
|
||||
# Transformers v4
|
||||
from transformers.configuration_utils import layer_type_validation
|
||||
|
||||
layer_type_validation(layer_types, num_hidden_layers)
|
||||
self.layer_types = layer_types
|
||||
self.validate_layer_type()
|
||||
if "linear_attention" not in layer_types:
|
||||
raise ValueError(
|
||||
"OLMoHybrid expects at least one 'linear_attention' layer."
|
||||
|
||||
@@ -94,18 +94,11 @@ class Qwen3_5TextConfig(PretrainedConfig):
|
||||
else "full_attention"
|
||||
for i in range(self.num_hidden_layers)
|
||||
]
|
||||
if hasattr(self, "validate_layer_type"):
|
||||
# Transformers v5
|
||||
kwargs["ignore_keys_at_rope_validation"] = {
|
||||
"mrope_section",
|
||||
"mrope_interleaved",
|
||||
}
|
||||
self.validate_layer_type()
|
||||
else:
|
||||
# Transformers v4
|
||||
from transformers.configuration_utils import layer_type_validation
|
||||
|
||||
layer_type_validation(self.layer_types, self.num_hidden_layers)
|
||||
kwargs["ignore_keys_at_rope_validation"] = {
|
||||
"mrope_section",
|
||||
"mrope_interleaved",
|
||||
}
|
||||
self.validate_layer_type()
|
||||
|
||||
# linear attention part
|
||||
self.linear_conv_kernel_dim = linear_conv_kernel_dim
|
||||
|
||||
@@ -100,18 +100,11 @@ class Qwen3_5MoeTextConfig(PretrainedConfig):
|
||||
else "full_attention"
|
||||
for i in range(self.num_hidden_layers)
|
||||
]
|
||||
if hasattr(self, "validate_layer_type"):
|
||||
# Transformers v5
|
||||
kwargs["ignore_keys_at_rope_validation"] = {
|
||||
"mrope_section",
|
||||
"mrope_interleaved",
|
||||
}
|
||||
self.validate_layer_type()
|
||||
else:
|
||||
# Transformers v4
|
||||
from transformers.configuration_utils import layer_type_validation
|
||||
|
||||
layer_type_validation(self.layer_types, self.num_hidden_layers)
|
||||
kwargs["ignore_keys_at_rope_validation"] = {
|
||||
"mrope_section",
|
||||
"mrope_interleaved",
|
||||
}
|
||||
self.validate_layer_type()
|
||||
|
||||
# linear attention part
|
||||
self.linear_conv_kernel_dim = linear_conv_kernel_dim
|
||||
|
||||
@@ -252,14 +252,7 @@ class Qwen3NextConfig(PretrainedConfig):
|
||||
"linear_attention" if bool((i + 1) % 4) else "full_attention"
|
||||
for i in range(self.num_hidden_layers)
|
||||
]
|
||||
if hasattr(self, "validate_layer_type"):
|
||||
# Transformers v5
|
||||
self.validate_layer_type()
|
||||
else:
|
||||
# Transformers v4
|
||||
from transformers.configuration_utils import layer_type_validation
|
||||
|
||||
layer_type_validation(self.layer_types)
|
||||
self.validate_layer_type()
|
||||
|
||||
# linear attention part
|
||||
self.linear_conv_kernel_dim = linear_conv_kernel_dim
|
||||
|
||||
@@ -1,7 +1,7 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
import os
|
||||
from dataclasses import fields, is_dataclass
|
||||
from dataclasses import fields
|
||||
from typing import Any
|
||||
|
||||
from transformers import PretrainedConfig
|
||||
@@ -16,11 +16,8 @@ class SpeculatorsConfig(PretrainedConfig):
|
||||
model_type = "speculators"
|
||||
|
||||
def __init__(self, **kwargs):
|
||||
# Transformers v4 - super().__init__ which sets all kwargs as attributes
|
||||
if not is_dataclass(PretrainedConfig):
|
||||
return super().__init__(**kwargs)
|
||||
# Transformers v5 - super().__init__ performs some validation before
|
||||
# setting all kwargs as attributes, so we set them first to be safe
|
||||
# super().__init__ performs some validation before setting all kwargs as
|
||||
# attributes, so we set them first to be safe
|
||||
pre_trained_config_fields = {f.name for f in fields(PretrainedConfig)}
|
||||
super_kwargs = dict()
|
||||
for key, value in kwargs.items():
|
||||
|
||||
@@ -59,10 +59,6 @@ def _transformers_v4_compatibility_init() -> Any:
|
||||
|
||||
This can be removed if `Molmo2ForConditionalGeneration` is upstreamed to
|
||||
Transformers."""
|
||||
# Transformers v4
|
||||
if hasattr(ProcessorMixin, "optional_attributes"):
|
||||
return
|
||||
# Transformers v5
|
||||
if hasattr(ProcessorMixin.__init__, "_vllm_patched"):
|
||||
return
|
||||
|
||||
|
||||
@@ -56,11 +56,6 @@ class MistralCommonPixtralProcessor(ProcessorMixin):
|
||||
image_processor: MistralCommonImageProcessor,
|
||||
) -> None:
|
||||
self.tokenizer = tokenizer.transformers_tokenizer
|
||||
|
||||
# Back-compatibility for Transformers v4
|
||||
if not hasattr(self.tokenizer, "init_kwargs"):
|
||||
self.tokenizer.init_kwargs = {}
|
||||
|
||||
self.image_processor = image_processor
|
||||
|
||||
image_special_ids = self.image_processor.mm_encoder.special_ids
|
||||
|
||||
@@ -111,11 +111,6 @@ class MistralCommonVoxtralProcessor(ProcessorMixin):
|
||||
feature_extractor: MistralCommonFeatureExtractor,
|
||||
) -> None:
|
||||
self.tokenizer = tokenizer.transformers_tokenizer
|
||||
|
||||
# Back-compatibility for Transformers v4
|
||||
if not hasattr(self.tokenizer, "init_kwargs"):
|
||||
self.tokenizer.init_kwargs = {}
|
||||
|
||||
self.feature_extractor = feature_extractor
|
||||
|
||||
audio_special_ids = self.feature_extractor.audio_encoder.special_ids
|
||||
|
||||
Reference in New Issue
Block a user