[Chore] Use Transformers v5 names and drop redundant processor use_fast (#58550)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com> Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
This commit is contained in:
co-authored by
Claude Opus 5.5
parent
484c211fe8
commit
d881812500
+12
-12
@@ -55,7 +55,7 @@ from vllm.utils.import_utils import LazyLoader
|
||||
from vllm.v1.attention.backends.registry import AttentionBackendEnum
|
||||
|
||||
if TYPE_CHECKING:
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
import vllm.model_executor.layers.quantization as me_quant
|
||||
import vllm.model_executor.models as me_models
|
||||
@@ -64,7 +64,7 @@ if TYPE_CHECKING:
|
||||
from vllm.model_executor.layers.quantization import QuantizationMethods
|
||||
from vllm.v1.sample.logits_processor import LogitsProcessor
|
||||
else:
|
||||
PretrainedConfig = Any
|
||||
PreTrainedConfig = Any
|
||||
|
||||
me_quant = LazyLoader(
|
||||
"model_executor", globals(), "vllm.model_executor.layers.quantization"
|
||||
@@ -105,7 +105,7 @@ PROCESSED_LOGPROBS_MODES: tuple[LogprobsMode, ...] = (
|
||||
"processed_logits",
|
||||
"processed_logprobs",
|
||||
)
|
||||
HfOverrides = dict[str, Any] | Callable[[PretrainedConfig], PretrainedConfig]
|
||||
HfOverrides = dict[str, Any] | Callable[[PreTrainedConfig], PreTrainedConfig]
|
||||
ModelImpl = Literal["auto", "vllm", "transformers", "terratorch"]
|
||||
LayerBlockType = Literal["attention", "linear_attention", "mamba"]
|
||||
|
||||
@@ -183,9 +183,9 @@ class ModelConfig:
|
||||
We must set the global seed because otherwise,
|
||||
different tensor parallel workers would sample different tokens,
|
||||
leading to inconsistent results."""
|
||||
hf_config: PretrainedConfig = field(init=False)
|
||||
hf_config: PreTrainedConfig = field(init=False)
|
||||
"""The Hugging Face config of the model."""
|
||||
hf_text_config: PretrainedConfig = field(init=False)
|
||||
hf_text_config: PreTrainedConfig = field(init=False)
|
||||
"""The Hugging Face config of the text model (same as hf_config for text models)."""
|
||||
is_submodel_config: bool = field(default=False, init=False)
|
||||
"""Whether this is a submodule view derived by `VllmConfig.with_hf_config`
|
||||
@@ -483,7 +483,7 @@ class ModelConfig:
|
||||
|
||||
def _update_nested(
|
||||
self,
|
||||
target: PretrainedConfig | dict[str, Any],
|
||||
target: PreTrainedConfig | dict[str, Any],
|
||||
updates: dict[str, Any],
|
||||
) -> None:
|
||||
"""Recursively updates a config or dict with nested updates."""
|
||||
@@ -511,15 +511,15 @@ class ModelConfig:
|
||||
|
||||
def _apply_dict_overrides(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
overrides: dict[str, Any],
|
||||
) -> None:
|
||||
"""Apply dict overrides, handling both nested configs and dict values."""
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
for key, value in overrides.items():
|
||||
attr = getattr(config, key, None)
|
||||
if attr is not None and isinstance(attr, PretrainedConfig):
|
||||
if attr is not None and isinstance(attr, PreTrainedConfig):
|
||||
# It's a nested config - recursively update it
|
||||
self._update_nested(attr, value)
|
||||
else:
|
||||
@@ -2393,7 +2393,7 @@ def _resolve_auto_dtype(
|
||||
|
||||
def _get_and_verify_dtype(
|
||||
model_id: str,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
dtype: str | torch.dtype,
|
||||
*,
|
||||
is_pooling_model: bool,
|
||||
@@ -2440,7 +2440,7 @@ def _get_and_verify_dtype(
|
||||
|
||||
|
||||
def _get_head_dtype(
|
||||
config: PretrainedConfig, dtype: torch.dtype, runner_type: str
|
||||
config: PreTrainedConfig, dtype: torch.dtype, runner_type: str
|
||||
) -> torch.dtype:
|
||||
head_dtype: str | torch.dtype | None = getattr(config, "head_dtype", None)
|
||||
|
||||
@@ -2464,7 +2464,7 @@ def _get_head_dtype(
|
||||
|
||||
|
||||
def _get_and_verify_max_len(
|
||||
hf_config: PretrainedConfig,
|
||||
hf_config: PreTrainedConfig,
|
||||
model_arch_config: ModelArchitectureConfig,
|
||||
tokenizer_config: dict | None,
|
||||
max_model_len: int | None,
|
||||
|
||||
@@ -22,12 +22,12 @@ from vllm.utils.import_utils import LazyLoader, has_arctic_inference
|
||||
from vllm.v1.attention.backends.registry import AttentionBackendEnum
|
||||
|
||||
if TYPE_CHECKING:
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
import vllm.model_executor.layers.quantization as me_quant
|
||||
from vllm.config.vllm import VllmConfig
|
||||
else:
|
||||
PretrainedConfig = Any
|
||||
PreTrainedConfig = Any
|
||||
VllmConfig = Any
|
||||
|
||||
me_quant = LazyLoader(
|
||||
@@ -657,7 +657,7 @@ class SpeculativeConfig:
|
||||
return hash_str
|
||||
|
||||
@staticmethod
|
||||
def hf_config_override(hf_config: PretrainedConfig) -> PretrainedConfig:
|
||||
def hf_config_override(hf_config: PreTrainedConfig) -> PreTrainedConfig:
|
||||
initial_architecture = hf_config.architectures[0]
|
||||
use_v32_mtp = hf_config.model_type in ("deepseek_v32", "glm_moe_dsa")
|
||||
if hf_config.model_type == "dots3_note":
|
||||
@@ -1066,16 +1066,16 @@ class SpeculativeConfig:
|
||||
|
||||
@staticmethod
|
||||
def _apply_composed_hf_override(
|
||||
target_hf_overrides: Callable[[PretrainedConfig], PretrainedConfig],
|
||||
hf_config: PretrainedConfig,
|
||||
) -> PretrainedConfig:
|
||||
target_hf_overrides: Callable[[PreTrainedConfig], PreTrainedConfig],
|
||||
hf_config: PreTrainedConfig,
|
||||
) -> PreTrainedConfig:
|
||||
hf_config = SpeculativeConfig.hf_config_override(hf_config)
|
||||
return target_hf_overrides(hf_config)
|
||||
|
||||
@staticmethod
|
||||
def compose_draft_hf_overrides(
|
||||
target_hf_overrides: HfOverrides | None,
|
||||
) -> Callable[[PretrainedConfig], PretrainedConfig]:
|
||||
) -> Callable[[PreTrainedConfig], PreTrainedConfig]:
|
||||
"""Build the ``hf_overrides`` for the draft ``ModelConfig``.
|
||||
|
||||
Callable overrides on the target are config-to-config transforms
|
||||
@@ -1667,7 +1667,7 @@ class SpeculativeConfig:
|
||||
|
||||
@staticmethod
|
||||
def _maybe_override_draft_max_position_embeddings(
|
||||
draft_hf_config: PretrainedConfig,
|
||||
draft_hf_config: PreTrainedConfig,
|
||||
target_max_model_len: int,
|
||||
) -> None:
|
||||
"""Raise an EAGLE draft's max_position_embeddings up to the target's.
|
||||
@@ -1704,7 +1704,7 @@ class SpeculativeConfig:
|
||||
def _verify_and_get_draft_tp(
|
||||
target_parallel_config: ParallelConfig,
|
||||
speculative_draft_tensor_parallel_size: int | None,
|
||||
draft_hf_config: PretrainedConfig,
|
||||
draft_hf_config: PreTrainedConfig,
|
||||
) -> int:
|
||||
"""Verifies and adjusts the tensor parallel size for a draft model
|
||||
specified using speculative_draft_tensor_parallel_size.
|
||||
|
||||
@@ -138,7 +138,7 @@ def getattr_iter(
|
||||
) -> Any:
|
||||
"""A helper function that retrieves an attribute from an object which may
|
||||
have multiple possible names. This is useful when fetching attributes from
|
||||
arbitrary `transformers.PretrainedConfig` instances.
|
||||
arbitrary `transformers.PreTrainedConfig` instances.
|
||||
|
||||
In the case where the first name in `names` is the preferred name, and
|
||||
any other names are deprecated aliases, setting `warn=True` will log a
|
||||
@@ -314,7 +314,7 @@ def normalize_value(x):
|
||||
except Exception:
|
||||
return str(x)
|
||||
|
||||
# PretrainedConfig (must be before dataclass branch as these are now dataclasses)
|
||||
# PreTrainedConfig (must be before dataclass branch as these are now dataclasses)
|
||||
if hasattr(x, "to_json_string") and callable(x.to_json_string):
|
||||
try:
|
||||
return x.to_json_string()
|
||||
|
||||
+3
-3
@@ -56,12 +56,12 @@ from .watermarking import WatermarkConfig
|
||||
from .weight_transfer import WeightTransferConfig
|
||||
|
||||
if TYPE_CHECKING:
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
from vllm.model_executor.layers.quantization.base_config import QuantizationConfig
|
||||
from vllm.v1.kv_cache_interface import KVCacheConfig
|
||||
else:
|
||||
PretrainedConfig = Any
|
||||
PreTrainedConfig = Any
|
||||
|
||||
QuantizationConfig = Any
|
||||
|
||||
@@ -939,7 +939,7 @@ class VllmConfig:
|
||||
|
||||
def with_hf_config(
|
||||
self,
|
||||
hf_config: PretrainedConfig,
|
||||
hf_config: PreTrainedConfig,
|
||||
architectures: list[str] | None = None,
|
||||
) -> "VllmConfig":
|
||||
if architectures is not None:
|
||||
|
||||
@@ -6,7 +6,7 @@ from typing import TYPE_CHECKING, overload
|
||||
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
from vllm.config.lora import LoRAConfig
|
||||
|
||||
@@ -73,7 +73,7 @@ class BaseLayerWithLoRA(nn.Module):
|
||||
self,
|
||||
max_loras: int,
|
||||
lora_config: LoRAConfig,
|
||||
model_config: PretrainedConfig | None = None,
|
||||
model_config: PreTrainedConfig | None = None,
|
||||
) -> None:
|
||||
"""Initializes lora matrices."""
|
||||
...
|
||||
@@ -103,7 +103,7 @@ class BaseLayerWithLoRA(nn.Module):
|
||||
source_layer: nn.Module,
|
||||
lora_config: LoRAConfig,
|
||||
packed_modules_list: list,
|
||||
model_config: PretrainedConfig | None = None,
|
||||
model_config: PreTrainedConfig | None = None,
|
||||
) -> bool:
|
||||
"""Returns True if the layer can be replaced by this LoRA layer."""
|
||||
raise NotImplementedError
|
||||
|
||||
@@ -3,7 +3,7 @@
|
||||
|
||||
|
||||
import torch
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
from vllm import envs
|
||||
from vllm.config import get_current_vllm_config
|
||||
@@ -105,7 +105,7 @@ class BaseLinearLayerWithLoRA(BaseLayerWithLoRA):
|
||||
self,
|
||||
max_loras: int,
|
||||
lora_config: LoRAConfig,
|
||||
model_config: PretrainedConfig | None = None,
|
||||
model_config: PreTrainedConfig | None = None,
|
||||
) -> None:
|
||||
self.lora_config = lora_config
|
||||
if isinstance(self.base_layer, ReplicatedLinear):
|
||||
|
||||
@@ -3,7 +3,7 @@
|
||||
|
||||
|
||||
import torch
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
from vllm.config.lora import LoRAConfig
|
||||
from vllm.platforms import current_platform
|
||||
@@ -16,7 +16,7 @@ class ClassificationHeadWithLoRA(ReplicatedLinearWithLoRA):
|
||||
self,
|
||||
max_loras: int,
|
||||
lora_config: LoRAConfig,
|
||||
model_config: PretrainedConfig | None = None,
|
||||
model_config: PreTrainedConfig | None = None,
|
||||
) -> None:
|
||||
# Preserve ordinary LoRA A/B support for classification heads.
|
||||
super().create_lora_weights(max_loras, lora_config, model_config)
|
||||
|
||||
@@ -4,7 +4,7 @@
|
||||
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
from vllm.config.lora import LoRAConfig
|
||||
from vllm.distributed import tensor_model_parallel_all_gather
|
||||
@@ -164,7 +164,7 @@ class ColumnParallelLinearWithLoRA(BaseLinearLayerWithLoRA):
|
||||
source_layer: nn.Module,
|
||||
lora_config: LoRAConfig,
|
||||
packed_modules_list: list,
|
||||
model_config: PretrainedConfig | None = None,
|
||||
model_config: PreTrainedConfig | None = None,
|
||||
) -> bool:
|
||||
if type(source_layer) is maybe_get_oot_by_class(ColumnParallelLinear):
|
||||
return True
|
||||
@@ -208,7 +208,7 @@ class MergedColumnParallelLinearWithLoRA(ColumnParallelLinearWithLoRA):
|
||||
self,
|
||||
max_loras: int,
|
||||
lora_config: LoRAConfig,
|
||||
model_config: PretrainedConfig | None = None,
|
||||
model_config: PreTrainedConfig | None = None,
|
||||
) -> None:
|
||||
"""The main reason for overriding this function is to enhance code
|
||||
maintainability.
|
||||
@@ -366,7 +366,7 @@ class MergedColumnParallelLinearWithLoRA(ColumnParallelLinearWithLoRA):
|
||||
source_layer: nn.Module,
|
||||
lora_config: LoRAConfig,
|
||||
packed_modules_list: list,
|
||||
model_config: PretrainedConfig | None = None,
|
||||
model_config: PreTrainedConfig | None = None,
|
||||
decorate: bool = True,
|
||||
) -> bool:
|
||||
merged_cls = maybe_get_oot_by_class(MergedColumnParallelLinear)
|
||||
@@ -442,7 +442,7 @@ class QKVParallelLinearWithLoRA(ColumnParallelLinearWithLoRA):
|
||||
source_layer: nn.Module,
|
||||
lora_config: LoRAConfig,
|
||||
packed_modules_list: list,
|
||||
model_config: PretrainedConfig | None = None,
|
||||
model_config: PreTrainedConfig | None = None,
|
||||
) -> bool:
|
||||
return (
|
||||
type(source_layer) is maybe_get_oot_by_class(QKVParallelLinear)
|
||||
@@ -488,7 +488,7 @@ class MergedQKVParallelLinearWithLoRA(MergedColumnParallelLinearWithLoRA):
|
||||
self,
|
||||
max_loras: int,
|
||||
lora_config: LoRAConfig,
|
||||
model_config: PretrainedConfig | None = None,
|
||||
model_config: PreTrainedConfig | None = None,
|
||||
) -> None:
|
||||
"""The main reason for overloading this function is to handle inconsistent
|
||||
weight dimensions in qkv lora.
|
||||
@@ -502,7 +502,7 @@ class MergedQKVParallelLinearWithLoRA(MergedColumnParallelLinearWithLoRA):
|
||||
source_layer: nn.Module,
|
||||
lora_config: LoRAConfig,
|
||||
packed_modules_list: list,
|
||||
model_config: PretrainedConfig | None = None,
|
||||
model_config: PreTrainedConfig | None = None,
|
||||
decorate: bool = True,
|
||||
) -> bool:
|
||||
return (
|
||||
@@ -543,7 +543,7 @@ class ColumnParallelLinearWithShardedLoRA(ColumnParallelLinearWithLoRA):
|
||||
source_layer: nn.Module,
|
||||
lora_config: LoRAConfig,
|
||||
packed_modules_list: list,
|
||||
model_config: PretrainedConfig | None = None,
|
||||
model_config: PreTrainedConfig | None = None,
|
||||
decorate: bool = True,
|
||||
) -> bool:
|
||||
# specifying kwargs so they can be easily accessed in decorator
|
||||
@@ -585,7 +585,7 @@ class MergedColumnParallelLinearWithShardedLoRA(MergedColumnParallelLinearWithLo
|
||||
source_layer: nn.Module,
|
||||
lora_config: LoRAConfig,
|
||||
packed_modules_list: list,
|
||||
model_config: PretrainedConfig | None = None,
|
||||
model_config: PreTrainedConfig | None = None,
|
||||
decorate: bool = True,
|
||||
) -> bool:
|
||||
# specifying kwargs so they can be easily accessed in decorator
|
||||
@@ -621,7 +621,7 @@ class QKVParallelLinearWithShardedLoRA(QKVParallelLinearWithLoRA):
|
||||
source_layer: nn.Module,
|
||||
lora_config: LoRAConfig,
|
||||
packed_modules_list: list,
|
||||
model_config: PretrainedConfig | None = None,
|
||||
model_config: PreTrainedConfig | None = None,
|
||||
) -> bool:
|
||||
# specifying kwargs so they can be easily accessed in decorator
|
||||
return super().can_replace_layer(
|
||||
@@ -669,7 +669,7 @@ class MergedQKVParallelLinearWithShardedLoRA(MergedQKVParallelLinearWithLoRA):
|
||||
source_layer: nn.Module,
|
||||
lora_config: LoRAConfig,
|
||||
packed_modules_list: list,
|
||||
model_config: PretrainedConfig | None = None,
|
||||
model_config: PreTrainedConfig | None = None,
|
||||
decorate: bool = True,
|
||||
) -> bool:
|
||||
# specifying kwargs so they can be easily accessed in decorator
|
||||
@@ -698,7 +698,7 @@ class MergedColumnParallelLinearVariableSliceWithLoRA(
|
||||
source_layer: nn.Module,
|
||||
lora_config: LoRAConfig,
|
||||
packed_modules_list: list,
|
||||
model_config: PretrainedConfig | None = None,
|
||||
model_config: PreTrainedConfig | None = None,
|
||||
decorate: bool = True,
|
||||
) -> bool:
|
||||
# Support MergedColumnParallelLinear with 3 or more slices
|
||||
|
||||
@@ -3,7 +3,7 @@
|
||||
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
from vllm import envs
|
||||
from vllm.config.lora import LoRAConfig
|
||||
@@ -245,7 +245,7 @@ class FusedMoEWithLoRA(BaseLayerWithLoRA):
|
||||
self,
|
||||
max_loras: int,
|
||||
lora_config: LoRAConfig,
|
||||
model_config: PretrainedConfig | None = None,
|
||||
model_config: PreTrainedConfig | None = None,
|
||||
) -> None:
|
||||
"""Initializes lora matrices."""
|
||||
self._verify_ep_fs(lora_config)
|
||||
@@ -465,7 +465,7 @@ class FusedMoEWithLoRA(BaseLayerWithLoRA):
|
||||
source_layer: nn.Module,
|
||||
lora_config: LoRAConfig,
|
||||
packed_modules_list: list,
|
||||
model_config: PretrainedConfig | None = None,
|
||||
model_config: PreTrainedConfig | None = None,
|
||||
) -> bool:
|
||||
"""Returns True if the layer can be replaced by this LoRA layer."""
|
||||
# source_layer is MoERunner
|
||||
@@ -511,7 +511,7 @@ class FusedMoE3DWithLoRA(FusedMoEWithLoRA):
|
||||
self,
|
||||
max_loras: int,
|
||||
lora_config: LoRAConfig,
|
||||
model_config: PretrainedConfig | None = None,
|
||||
model_config: PreTrainedConfig | None = None,
|
||||
) -> None:
|
||||
"""Initializes lora matrices."""
|
||||
if model_config is None:
|
||||
@@ -624,7 +624,7 @@ class FusedMoE3DWithLoRA(FusedMoEWithLoRA):
|
||||
source_layer: nn.Module,
|
||||
lora_config: LoRAConfig,
|
||||
packed_modules_list: list,
|
||||
model_config: PretrainedConfig | None = None,
|
||||
model_config: PreTrainedConfig | None = None,
|
||||
) -> bool:
|
||||
"""Returns True if the layer can be replaced by this LoRA layer."""
|
||||
# source_layer is MoERunner
|
||||
|
||||
@@ -3,7 +3,7 @@
|
||||
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
from vllm.config.lora import LoRAConfig
|
||||
from vllm.distributed import (
|
||||
@@ -85,7 +85,7 @@ class LogitsProcessorWithLoRA(BaseLayerWithLoRA):
|
||||
self,
|
||||
max_loras: int,
|
||||
lora_config: LoRAConfig,
|
||||
model_config: PretrainedConfig | None = None,
|
||||
model_config: PreTrainedConfig | None = None,
|
||||
) -> None:
|
||||
# TODO: Verify if this condition can be further relaxed
|
||||
if self.base_layer.vocab_size > 258048:
|
||||
@@ -221,7 +221,7 @@ class LogitsProcessorWithLoRA(BaseLayerWithLoRA):
|
||||
source_layer: nn.Module,
|
||||
lora_config: LoRAConfig,
|
||||
packed_modules_list: list,
|
||||
model_config: PretrainedConfig | None = None,
|
||||
model_config: PreTrainedConfig | None = None,
|
||||
) -> bool:
|
||||
# Special handling for the LogitsProcessor.
|
||||
return False
|
||||
|
||||
@@ -4,7 +4,7 @@
|
||||
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
from vllm.config.lora import LoRAConfig
|
||||
from vllm.model_executor.custom_op import maybe_get_oot_by_class
|
||||
@@ -61,7 +61,7 @@ class ReplicatedLinearWithLoRA(BaseLinearLayerWithLoRA):
|
||||
source_layer: nn.Module,
|
||||
lora_config: LoRAConfig,
|
||||
packed_modules_list: list,
|
||||
model_config: PretrainedConfig | None = None,
|
||||
model_config: PreTrainedConfig | None = None,
|
||||
) -> bool:
|
||||
return isinstance(source_layer, maybe_get_oot_by_class(ReplicatedLinear))
|
||||
|
||||
|
||||
@@ -4,7 +4,7 @@
|
||||
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
from vllm.config.lora import LoRAConfig
|
||||
from vllm.distributed import (
|
||||
@@ -89,7 +89,7 @@ class RowParallelLinearWithLoRA(BaseLinearLayerWithLoRA):
|
||||
source_layer: nn.Module,
|
||||
lora_config: LoRAConfig,
|
||||
packed_modules_list: list,
|
||||
model_config: PretrainedConfig | None = None,
|
||||
model_config: PreTrainedConfig | None = None,
|
||||
) -> bool:
|
||||
return type(source_layer) is maybe_get_oot_by_class(RowParallelLinear)
|
||||
|
||||
@@ -165,7 +165,7 @@ class RowParallelLinearWithShardedLoRA(RowParallelLinearWithLoRA):
|
||||
source_layer: nn.Module,
|
||||
lora_config: LoRAConfig,
|
||||
packed_modules_list: list,
|
||||
model_config: PretrainedConfig | None = None,
|
||||
model_config: PreTrainedConfig | None = None,
|
||||
) -> bool:
|
||||
# specifying kwargs so they can be easily accessed in decorator
|
||||
return super().can_replace_layer(
|
||||
|
||||
@@ -4,7 +4,7 @@
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
import torch.nn.functional as F
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
from vllm.config.lora import LoRAConfig
|
||||
from vllm.model_executor.custom_op import maybe_get_oot_by_class
|
||||
@@ -26,7 +26,7 @@ class VocabParallelEmbeddingWithLoRA(BaseLayerWithLoRA):
|
||||
self,
|
||||
max_loras: int,
|
||||
lora_config: LoRAConfig,
|
||||
model_config: PretrainedConfig | None = None,
|
||||
model_config: PreTrainedConfig | None = None,
|
||||
) -> None:
|
||||
self.lora_a_stacked = torch.zeros(
|
||||
(
|
||||
@@ -113,7 +113,7 @@ class VocabParallelEmbeddingWithLoRA(BaseLayerWithLoRA):
|
||||
source_layer: nn.Module,
|
||||
lora_config: LoRAConfig,
|
||||
packed_modules_list: list,
|
||||
model_config: PretrainedConfig | None = None,
|
||||
model_config: PreTrainedConfig | None = None,
|
||||
) -> bool:
|
||||
# Accept both the in-tree and any OOT class, because layers built on top of
|
||||
# `VocabParallelEmbedding` (e.g. by the Transformers modelling backend) use the
|
||||
|
||||
+4
-4
@@ -6,7 +6,7 @@ from typing import TYPE_CHECKING
|
||||
|
||||
from huggingface_hub.utils import HfHubHTTPError, HFValidationError
|
||||
from torch import nn
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
from vllm import envs
|
||||
from vllm.config.lora import LoRAConfig
|
||||
@@ -109,7 +109,7 @@ def from_layer(
|
||||
max_loras: int,
|
||||
lora_config: LoRAConfig,
|
||||
packed_modules_list: list,
|
||||
model_config: PretrainedConfig | None = None,
|
||||
model_config: PreTrainedConfig | None = None,
|
||||
) -> nn.Module:
|
||||
for lora_cls in _all_lora_classes:
|
||||
# specifying kwargs so they can be easily accessed in decorator
|
||||
@@ -130,7 +130,7 @@ def from_layer_logits_processor(
|
||||
lm_head: "ParallelLMHead",
|
||||
max_loras: int,
|
||||
lora_config: LoRAConfig,
|
||||
model_config: PretrainedConfig | None = None,
|
||||
model_config: PreTrainedConfig | None = None,
|
||||
) -> LogitsProcessorWithLoRA:
|
||||
ret = LogitsProcessorWithLoRA(
|
||||
layer,
|
||||
@@ -147,7 +147,7 @@ def from_layer_classification(
|
||||
layer: nn.Module,
|
||||
max_loras: int,
|
||||
lora_config: LoRAConfig,
|
||||
model_config: PretrainedConfig | None = None,
|
||||
model_config: PreTrainedConfig | None = None,
|
||||
) -> ClassificationHeadWithLoRA:
|
||||
instance_layer = ClassificationHeadWithLoRA(layer)
|
||||
instance_layer.create_lora_weights(max_loras, lora_config, model_config)
|
||||
|
||||
@@ -1,7 +1,7 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
import torch
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
from vllm.config import (
|
||||
VllmConfig,
|
||||
@@ -24,7 +24,7 @@ class GatedDeltaNetAttention(PluggableLayer, MambaBase):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
vllm_config: VllmConfig,
|
||||
prefix: str = "",
|
||||
) -> None:
|
||||
|
||||
@@ -4,7 +4,7 @@ import copy
|
||||
|
||||
import torch
|
||||
import torch.nn.functional as F
|
||||
from transformers.configuration_utils import PretrainedConfig
|
||||
from transformers.configuration_utils import PreTrainedConfig
|
||||
|
||||
from vllm.config import (
|
||||
VllmConfig,
|
||||
@@ -37,7 +37,7 @@ from vllm.triton_utils import tl, triton
|
||||
from vllm.v1.attention.backends.linear_attn import LinearAttentionMetadata
|
||||
|
||||
|
||||
def _build_rope_parameters(config: PretrainedConfig) -> dict | None:
|
||||
def _build_rope_parameters(config: PreTrainedConfig) -> dict | None:
|
||||
rope_parameters = copy.deepcopy(getattr(config, "rope_parameters", None)) or {}
|
||||
if "rope_theta" not in rope_parameters and hasattr(config, "rope_theta"):
|
||||
rope_parameters["rope_theta"] = config.rope_theta
|
||||
@@ -385,7 +385,7 @@ class BailingMoELinearAttention(LinearAttention):
|
||||
# --8<-- [end:bailing_moe_linear_attention]
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
vllm_config: VllmConfig,
|
||||
prefix: str = "linear_attn",
|
||||
):
|
||||
|
||||
@@ -1,7 +1,7 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
import torch
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
from vllm.config import (
|
||||
VllmConfig,
|
||||
@@ -24,7 +24,7 @@ class LinearAttention(PluggableLayer, MambaBase):
|
||||
"""Base class for Linear attention layer."""
|
||||
|
||||
def __init__(
|
||||
self, config: PretrainedConfig, vllm_config: VllmConfig, prefix: str = ""
|
||||
self, config: PreTrainedConfig, vllm_config: VllmConfig, prefix: str = ""
|
||||
):
|
||||
super().__init__()
|
||||
self.layer_idx = extract_layer_index(prefix)
|
||||
|
||||
@@ -7,7 +7,7 @@ from typing import TypeVar
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
import torch.nn.functional as F
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
from vllm.config import ModelConfig
|
||||
from vllm.logger import init_logger
|
||||
@@ -17,7 +17,7 @@ logger = init_logger(__name__)
|
||||
|
||||
|
||||
def get_act_fn(
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
static_num_labels: bool = True,
|
||||
) -> "PoolerActivation":
|
||||
# get classification act_fn
|
||||
|
||||
@@ -6,7 +6,7 @@ from typing import TYPE_CHECKING, Any, Union
|
||||
import torch
|
||||
from safetensors.torch import _TYPES as _SAFETENSORS_TO_TORCH_DTYPE
|
||||
from torch.nn import Parameter
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
import vllm.model_executor.layers.fused_moe # noqa
|
||||
from vllm import _custom_ops as ops
|
||||
@@ -370,7 +370,7 @@ class AutoAWQConfig(QuantizationConfig):
|
||||
def maybe_update_config(
|
||||
self,
|
||||
model_name: str,
|
||||
hf_config: PretrainedConfig | None = None,
|
||||
hf_config: PreTrainedConfig | None = None,
|
||||
revision: str | None = None,
|
||||
):
|
||||
if self.modules_to_not_convert:
|
||||
|
||||
@@ -6,7 +6,7 @@ from typing import Any
|
||||
|
||||
import torch
|
||||
from safetensors.torch import _TYPES as _SAFETENSORS_TO_TORCH_DTYPE
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
import vllm.model_executor.layers.fused_moe # noqa
|
||||
from vllm.logger import init_logger
|
||||
@@ -273,7 +273,7 @@ class AutoGPTQConfig(QuantizationConfig):
|
||||
def maybe_update_config(
|
||||
self,
|
||||
model_name: str,
|
||||
hf_config: PretrainedConfig | None = None,
|
||||
hf_config: PreTrainedConfig | None = None,
|
||||
revision: str | None = None,
|
||||
):
|
||||
if self.modules_in_block_to_quantize:
|
||||
|
||||
@@ -8,7 +8,7 @@ from typing import TYPE_CHECKING, Any
|
||||
import regex as re
|
||||
import torch
|
||||
from torch import nn
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
if TYPE_CHECKING:
|
||||
from vllm.model_executor.layers.quantization import QuantizationMethods
|
||||
@@ -265,7 +265,7 @@ class QuantizationConfig(ABC):
|
||||
def maybe_update_config( # noqa: B027
|
||||
self,
|
||||
model_name: str,
|
||||
hf_config: PretrainedConfig | None = None,
|
||||
hf_config: PreTrainedConfig | None = None,
|
||||
revision: str | None = None,
|
||||
):
|
||||
"""Interface to update values after config initialization.
|
||||
|
||||
@@ -18,7 +18,7 @@ import regex as re
|
||||
import torch
|
||||
from torch import nn
|
||||
from torch.utils._python_dispatch import TorchDispatchMode
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
import vllm.envs as envs
|
||||
from vllm.config import ModelConfig, ParallelConfig, VllmConfig, set_current_vllm_config
|
||||
@@ -173,7 +173,7 @@ class TensorizerConfig(MutableMapping):
|
||||
deserialization_kwargs: dict[str, Any] | None = None
|
||||
_extra_serialization_attrs: dict[str, Any] | None = field(init=False, default=None)
|
||||
model_class: type[torch.nn.Module] | None = field(init=False, default=None)
|
||||
hf_config: PretrainedConfig | None = field(init=False, default=None)
|
||||
hf_config: PreTrainedConfig | None = field(init=False, default=None)
|
||||
dtype: str | torch.dtype | None = field(init=False, default=None)
|
||||
_is_sharded: bool = field(init=False, default=False)
|
||||
_fields: ClassVar[tuple[str, ...]]
|
||||
|
||||
@@ -284,7 +284,7 @@ def as_embedding_model(cls: type[_T]) -> type[_T]:
|
||||
def _resolve_num_labels(hf_config: Any, text_config: Any) -> int:
|
||||
"""Resolve the label count for a sequence classification head.
|
||||
|
||||
``PretrainedConfig.num_labels`` is derived from ``id2label``, which always
|
||||
``PreTrainedConfig.num_labels`` is derived from ``id2label``, which always
|
||||
carries a default of two entries. Composite configs (such as multimodal
|
||||
checkpoints) declare their label space on the top-level config, so reading
|
||||
``num_labels`` from ``get_text_config()`` silently returns that default and
|
||||
@@ -297,9 +297,9 @@ def _resolve_num_labels(hf_config: Any, text_config: Any) -> int:
|
||||
if text_config is hf_config:
|
||||
return hf_config.num_labels
|
||||
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
if hf_config.num_labels != PretrainedConfig().num_labels:
|
||||
if hf_config.num_labels != PreTrainedConfig().num_labels:
|
||||
return hf_config.num_labels
|
||||
return text_config.num_labels
|
||||
|
||||
|
||||
@@ -22,7 +22,7 @@ from typing import Annotated, Any, Literal, TypeAlias
|
||||
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
from transformers import BatchFeature, PretrainedConfig
|
||||
from transformers import BatchFeature, PreTrainedConfig
|
||||
from transformers.models.audioflamingo3 import (
|
||||
AudioFlamingo3Config,
|
||||
AudioFlamingo3Processor,
|
||||
@@ -118,7 +118,7 @@ AudioFlamingo3Inputs: TypeAlias = (
|
||||
class AudioFlamingo3Encoder(Qwen2AudioEncoder):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
):
|
||||
super().__init__(config)
|
||||
self.avg_pooler = nn.AvgPool1d(kernel_size=2, stride=2)
|
||||
@@ -161,7 +161,7 @@ class AudioFlamingo3Encoder(Qwen2AudioEncoder):
|
||||
|
||||
|
||||
class AudioFlamingo3MultiModalProjector(nn.Module):
|
||||
def __init__(self, config: PretrainedConfig):
|
||||
def __init__(self, config: PreTrainedConfig):
|
||||
super().__init__()
|
||||
self.linear_1 = nn.Linear(
|
||||
config.audio_config.hidden_size,
|
||||
|
||||
@@ -30,7 +30,7 @@ from itertools import islice
|
||||
import torch
|
||||
import torch.nn.functional as F
|
||||
from torch import nn
|
||||
from transformers.configuration_utils import PretrainedConfig
|
||||
from transformers.configuration_utils import PreTrainedConfig
|
||||
|
||||
from vllm.compilation.decorators import support_torch_compile
|
||||
from vllm.config import CacheConfig, VllmConfig
|
||||
@@ -74,7 +74,7 @@ from .utils import (
|
||||
class BailingAttention(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
cache_config: CacheConfig | None = None,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
reduce_results: bool = True,
|
||||
@@ -183,7 +183,7 @@ class BailingMLP(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
intermediate_size: int,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
reduce_results: bool = True,
|
||||
prefix: str = "",
|
||||
@@ -217,7 +217,7 @@ class BailingMoE(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
intermediate_size: int,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
reduce_results: bool | None = True,
|
||||
prefix: str = "",
|
||||
@@ -328,7 +328,7 @@ class BailingMoE(nn.Module):
|
||||
class BailingMoeBlock(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
cache_config: CacheConfig | None = None,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
prefix: str = "",
|
||||
|
||||
@@ -5,7 +5,7 @@ from collections.abc import Callable, Iterable
|
||||
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
from transformers.configuration_utils import PretrainedConfig
|
||||
from transformers.configuration_utils import PreTrainedConfig
|
||||
|
||||
from vllm.compilation.decorators import support_torch_compile
|
||||
from vllm.config import CacheConfig, VllmConfig
|
||||
@@ -79,7 +79,7 @@ class BailingMoeV25MLAAttention(nn.Module):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
layer_id: int = 0,
|
||||
prefix: str = "attention",
|
||||
@@ -234,7 +234,7 @@ class BailingMoeV25(nn.Module):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
layer_id: int = 0,
|
||||
prefix: str = "",
|
||||
@@ -349,7 +349,7 @@ class BailingMoeV25DecoderLayer(nn.Module):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
vllm_config: VllmConfig,
|
||||
prefix: str = "layer",
|
||||
layer_id: int = 0,
|
||||
|
||||
@@ -6,7 +6,7 @@ from collections.abc import Callable, Iterable
|
||||
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
from transformers.configuration_utils import PretrainedConfig
|
||||
from transformers.configuration_utils import PreTrainedConfig
|
||||
|
||||
from vllm.compilation.decorators import support_torch_compile
|
||||
from vllm.config import VllmConfig
|
||||
@@ -36,7 +36,7 @@ from vllm.sequence import IntermediateTensors
|
||||
from .utils import PPMissingLayer, is_pp_missing_parameter, maybe_prefix
|
||||
|
||||
|
||||
def _get_draft_hf_config(vllm_config: VllmConfig) -> PretrainedConfig:
|
||||
def _get_draft_hf_config(vllm_config: VllmConfig) -> PreTrainedConfig:
|
||||
speculative_config = vllm_config.speculative_config
|
||||
if speculative_config is not None:
|
||||
draft_model_config = speculative_config.draft_model_config
|
||||
@@ -48,7 +48,7 @@ def _get_draft_hf_config(vllm_config: VllmConfig) -> PretrainedConfig:
|
||||
class BailingMTPSharedHead(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
prefix: str,
|
||||
vllm_config: VllmConfig,
|
||||
) -> None:
|
||||
|
||||
@@ -17,7 +17,7 @@ import regex as re
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
from einops import rearrange
|
||||
from transformers.configuration_utils import PretrainedConfig
|
||||
from transformers.configuration_utils import PreTrainedConfig
|
||||
|
||||
from vllm.compilation.decorators import support_torch_compile
|
||||
from vllm.config import CacheConfig, ModelConfig, VllmConfig, get_current_vllm_config
|
||||
@@ -151,7 +151,7 @@ def _get_kda_state_shape_for_config(
|
||||
return shapes
|
||||
|
||||
|
||||
def _build_rope_parameters(config: PretrainedConfig) -> dict | None:
|
||||
def _build_rope_parameters(config: PreTrainedConfig) -> dict | None:
|
||||
rope_parameters = copy.deepcopy(getattr(config, "rope_parameters", None)) or {}
|
||||
if "rope_theta" not in rope_parameters and hasattr(config, "rope_theta"):
|
||||
rope_parameters["rope_theta"] = config.rope_theta
|
||||
@@ -172,7 +172,7 @@ def _build_rope_parameters(config: PretrainedConfig) -> dict | None:
|
||||
|
||||
|
||||
def _build_mla_rotary_embedding(
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
head_size: int,
|
||||
) -> nn.Module:
|
||||
rope_parameters = _build_rope_parameters(config)
|
||||
@@ -223,7 +223,7 @@ def _is_block_fp8_config(
|
||||
|
||||
def _configure_ling_fp8_quant_config(
|
||||
quant_config: QuantizationConfig | None,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
) -> None:
|
||||
if not _is_block_fp8_config(quant_config):
|
||||
return
|
||||
@@ -372,7 +372,7 @@ def _pad_block_fp8_mlp_checkpoint_tensor(
|
||||
|
||||
def _maybe_pad_block_fp8_shared_expert_checkpoint_tensor(
|
||||
quant_config: QuantizationConfig | None,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
name: str,
|
||||
loaded_weight: torch.Tensor,
|
||||
) -> torch.Tensor:
|
||||
@@ -410,7 +410,7 @@ class BailingMoeV3MLP(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
intermediate_size: int,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
reduce_results: bool = True,
|
||||
prefix: str = "",
|
||||
@@ -446,7 +446,7 @@ class BailingMoeV3MLP(nn.Module):
|
||||
class BailingMoeV3MLAAttention(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
layer_id: int = 0,
|
||||
prefix: str = "attention",
|
||||
@@ -625,7 +625,7 @@ class BailingMoeV3KimiDeltaAttention(PluggableLayer, MambaBase):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
layer_id: int = 0,
|
||||
prefix: str = "attention",
|
||||
@@ -1121,7 +1121,7 @@ class BailingMoeV3KimiDeltaAttention(PluggableLayer, MambaBase):
|
||||
class BailingMoeV3MoE(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
layer_id: int = 0,
|
||||
prefix: str = "",
|
||||
@@ -1193,7 +1193,7 @@ class BailingMoeV3MoE(nn.Module):
|
||||
class BailingMoeV3DecoderLayer(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
layer_id: int = 0,
|
||||
prefix: str = "layer",
|
||||
|
||||
@@ -6,7 +6,7 @@ from collections.abc import Callable, Iterable
|
||||
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
from transformers.configuration_utils import PretrainedConfig
|
||||
from transformers.configuration_utils import PreTrainedConfig
|
||||
|
||||
from vllm.compilation.decorators import support_torch_compile
|
||||
from vllm.config import VllmConfig
|
||||
@@ -47,7 +47,7 @@ from vllm.model_executor.models.utils import (
|
||||
from vllm.sequence import IntermediateTensors
|
||||
|
||||
|
||||
def _get_draft_hf_config(vllm_config: VllmConfig) -> PretrainedConfig:
|
||||
def _get_draft_hf_config(vllm_config: VllmConfig) -> PreTrainedConfig:
|
||||
speculative_config = vllm_config.speculative_config
|
||||
if speculative_config is not None:
|
||||
draft_model_config = speculative_config.draft_model_config
|
||||
@@ -59,7 +59,7 @@ def _get_draft_hf_config(vllm_config: VllmConfig) -> PretrainedConfig:
|
||||
class BailingMoeV3MTPSharedHead(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
prefix: str,
|
||||
vllm_config: VllmConfig,
|
||||
) -> None:
|
||||
|
||||
@@ -4,7 +4,7 @@ from collections.abc import Iterable
|
||||
|
||||
import torch
|
||||
from torch import nn
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
from vllm.compilation.decorators import support_torch_compile
|
||||
from vllm.config import CacheConfig, VllmConfig
|
||||
@@ -50,7 +50,7 @@ from .interfaces_base import default_pooling_type
|
||||
|
||||
|
||||
class BertWithRopeEmbedding(nn.Module):
|
||||
def __init__(self, config: PretrainedConfig):
|
||||
def __init__(self, config: PreTrainedConfig):
|
||||
super().__init__()
|
||||
if config.position_embedding_type not in ["rope", "rotary"]:
|
||||
raise ValueError(
|
||||
@@ -347,7 +347,7 @@ class NomicMoE(nn.Module):
|
||||
class BertWithRopeBlock(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
cache_config: CacheConfig | None = None,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
moe: bool = False,
|
||||
|
||||
@@ -8,10 +8,10 @@ from typing import Annotated, Literal
|
||||
|
||||
import torch
|
||||
from torch import nn
|
||||
from transformers import BatchFeature, PretrainedConfig
|
||||
from transformers import BatchFeature, PreTrainedConfig
|
||||
from transformers.models.cohere2_vision import Cohere2VisionConfig
|
||||
from transformers.models.cohere2_vision.image_processing_cohere2_vision_fast import ( # noqa: E501
|
||||
Cohere2VisionImageProcessorFast,
|
||||
from transformers.models.cohere2_vision.image_processing_cohere2_vision import ( # noqa: E501
|
||||
Cohere2VisionImageProcessor,
|
||||
)
|
||||
from transformers.models.cohere2_vision.processing_cohere2_vision import (
|
||||
Cohere2VisionProcessor,
|
||||
@@ -178,7 +178,7 @@ class Cohere2VisionProcessingInfo(BaseProcessingInfo):
|
||||
"""Calculate the number of image patches for a given image.
|
||||
Uses the HF processor to determine the actual number of patches.
|
||||
"""
|
||||
image_processor: Cohere2VisionImageProcessorFast = processor.image_processor
|
||||
image_processor: Cohere2VisionImageProcessor = processor.image_processor
|
||||
|
||||
return image_processor.get_number_of_image_patches(
|
||||
image_height,
|
||||
@@ -415,7 +415,7 @@ class Cohere2VisionForConditionalGeneration(
|
||||
)
|
||||
|
||||
def _patch_quant_config(
|
||||
self, config: PretrainedConfig, quant_config: QuantizationConfig | None
|
||||
self, config: PreTrainedConfig, quant_config: QuantizationConfig | None
|
||||
):
|
||||
# the awq models from OpenGVLab missing `modules_to_not_convert`
|
||||
# patch the quant_config to add `modules_to_not_convert` back
|
||||
|
||||
@@ -8,7 +8,7 @@ from typing import Any, ClassVar
|
||||
import torch
|
||||
import torch.nn.functional as F
|
||||
from torch import nn
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
from vllm.compilation.decorators import support_torch_compile
|
||||
from vllm.config import CacheConfig, ModelConfig, SpeechToTextConfig, VllmConfig
|
||||
@@ -1861,7 +1861,7 @@ class CohereASRModel(nn.Module):
|
||||
|
||||
|
||||
class CohereASRProcessingInfo(BaseProcessingInfo):
|
||||
def get_hf_config(self) -> PretrainedConfig:
|
||||
def get_hf_config(self) -> PreTrainedConfig:
|
||||
return self.ctx.get_hf_config()
|
||||
|
||||
def get_default_tok_params(self) -> TokenizeParams:
|
||||
|
||||
@@ -1161,7 +1161,6 @@ class CohereCompassProcessingInfo(BaseProcessingInfo):
|
||||
def get_hf_processor(self, **kwargs: object) -> CohereCompassProcessor:
|
||||
return self.ctx.get_hf_processor(
|
||||
CohereCompassProcessor,
|
||||
use_fast=kwargs.pop("use_fast", True),
|
||||
**kwargs,
|
||||
)
|
||||
|
||||
|
||||
@@ -61,7 +61,6 @@ class ColQwen3ProcessingInfo(Qwen3VLProcessingInfo):
|
||||
# correctly for the Qwen3-VL backbone.
|
||||
return self.ctx.get_hf_processor(
|
||||
Qwen3VLProcessor,
|
||||
use_fast=kwargs.pop("use_fast", True),
|
||||
**kwargs,
|
||||
)
|
||||
|
||||
|
||||
@@ -58,7 +58,6 @@ class ColQwen3_5ProcessingInfo(Qwen3_5ProcessingInfo):
|
||||
def get_hf_processor(self, **kwargs: object) -> Qwen3VLProcessor:
|
||||
return self.ctx.get_hf_processor(
|
||||
Qwen3VLProcessor,
|
||||
use_fast=kwargs.pop("use_fast", True),
|
||||
**kwargs,
|
||||
)
|
||||
|
||||
|
||||
@@ -6,7 +6,7 @@ from vllm.logger import init_logger
|
||||
from vllm.utils.math_utils import round_up
|
||||
|
||||
if TYPE_CHECKING:
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
from vllm.config import CacheConfig, ModelConfig, VllmConfig
|
||||
from vllm.config.cache import MambaDType
|
||||
@@ -681,7 +681,7 @@ class NemotronHForCausalLMConfig(VerifyAndUpdateConfig):
|
||||
|
||||
@classmethod
|
||||
def update_mamba_ssm_cache_dtype(
|
||||
cls, *, cache_config: "CacheConfig", hf_config: "PretrainedConfig"
|
||||
cls, *, cache_config: "CacheConfig", hf_config: "PreTrainedConfig"
|
||||
) -> None:
|
||||
"""Update mamba_ssm_cache_dtype for NemotronH models when set to 'auto'
|
||||
(or not explicitly set), to the value specified in the HF config, or to
|
||||
|
||||
@@ -5,7 +5,7 @@ from collections.abc import Callable, Iterable
|
||||
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
from vllm.compilation.decorators import support_torch_compile
|
||||
from vllm.config import VllmConfig
|
||||
@@ -51,7 +51,7 @@ from .utils import (
|
||||
class SharedHead(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
prefix: str,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
*,
|
||||
|
||||
@@ -112,13 +112,13 @@ class DiffusionGemmaProcessingInfo(Gemma4ProcessingInfo):
|
||||
"""Processing info for DiffusionGemma.
|
||||
|
||||
Overrides ``get_hf_config`` to accept ``DiffusionGemmaConfig``
|
||||
(which inherits from ``PretrainedConfig``, not ``Gemma4Config``).
|
||||
(which inherits from ``PreTrainedConfig``, not ``Gemma4Config``).
|
||||
Supports image and video modalities.
|
||||
"""
|
||||
|
||||
def get_hf_config(self):
|
||||
# DiffusionGemmaConfig doesn't inherit from Gemma4Config, so we
|
||||
# accept any PretrainedConfig here.
|
||||
# accept any PreTrainedConfig here.
|
||||
return self.ctx.get_hf_config()
|
||||
|
||||
def get_supported_mm_limits(self) -> Mapping[str, int | None]:
|
||||
|
||||
@@ -8,7 +8,7 @@ from typing import Annotated, Literal, TypeAlias
|
||||
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
from vllm.config import VllmConfig
|
||||
from vllm.model_executor.layers.quantization import QuantizationConfig
|
||||
@@ -193,7 +193,7 @@ class Eagle2_5_VLForConditionalGeneration(
|
||||
|
||||
def _init_vision_model(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None,
|
||||
prefix: str,
|
||||
):
|
||||
@@ -220,7 +220,7 @@ class Eagle2_5_VLForConditionalGeneration(
|
||||
prefix=prefix,
|
||||
)
|
||||
|
||||
def _init_mlp1(self, config: PretrainedConfig) -> nn.Module:
|
||||
def _init_mlp1(self, config: PreTrainedConfig) -> nn.Module:
|
||||
"""Initialize MLP projection layer."""
|
||||
vit_hidden_size = config.vision_config.hidden_size
|
||||
llm_hidden_size = config.text_config.hidden_size
|
||||
|
||||
@@ -29,7 +29,7 @@ from typing import Any
|
||||
|
||||
import torch
|
||||
from torch import nn
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
from vllm.compilation.decorators import support_torch_compile
|
||||
from vllm.config import CacheConfig, VllmConfig, get_current_vllm_config
|
||||
@@ -119,7 +119,7 @@ class Ernie4_5_MoeMLP(nn.Module):
|
||||
class Ernie4_5_MoeMoE(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
prefix: str = "",
|
||||
enable_eplb: bool = False,
|
||||
@@ -305,7 +305,7 @@ class Ernie4_5_MoeAttention(nn.Module):
|
||||
class Ernie4_5_MoeDecoderLayer(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
cache_config: CacheConfig | None = None,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
prefix: str = "",
|
||||
|
||||
@@ -881,7 +881,7 @@ class Ernie4_5_VLProcessingInfo(BaseProcessingInfo):
|
||||
return self.ctx.model_config.hf_config
|
||||
|
||||
def get_hf_processor(self, **kwargs: object):
|
||||
return self.ctx.get_hf_processor(use_fast=True, **kwargs)
|
||||
return self.ctx.get_hf_processor(**kwargs)
|
||||
|
||||
def get_image_processor(self, **kwargs: object):
|
||||
return self.get_hf_processor(**kwargs).image_processor
|
||||
|
||||
@@ -29,7 +29,7 @@ from typing import Any
|
||||
|
||||
import torch
|
||||
from torch import nn
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
# from vllm.compilation.decorators import support_torch_compile
|
||||
from vllm.config import CacheConfig, VllmConfig
|
||||
@@ -192,7 +192,7 @@ class Ernie4_5_VLMoeAttention(nn.Module):
|
||||
class Ernie4_5_VLMoeMoE(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
prefix: str = "",
|
||||
):
|
||||
@@ -385,7 +385,7 @@ class Ernie4_5_VLMoeMoE(nn.Module):
|
||||
class Ernie4_5_VLMoeDecoderLayer(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
cache_config: CacheConfig | None = None,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
prefix: str = "",
|
||||
|
||||
@@ -30,7 +30,7 @@ from itertools import islice
|
||||
|
||||
import torch
|
||||
from torch import nn
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
from vllm.compilation.decorators import support_torch_compile
|
||||
from vllm.config import CacheConfig, VllmConfig
|
||||
@@ -104,7 +104,7 @@ class ExaoneGatedMLP(nn.Module):
|
||||
class ExaoneAttention(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
hidden_size: int,
|
||||
num_heads: int,
|
||||
num_kv_heads: int,
|
||||
@@ -191,7 +191,7 @@ class ExaoneAttention(nn.Module):
|
||||
class ExaoneBlockAttention(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
hidden_size: int,
|
||||
num_heads: int,
|
||||
num_kv_heads: int,
|
||||
@@ -228,7 +228,7 @@ class ExaoneBlockAttention(nn.Module):
|
||||
class ExaoneDecoderLayer(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
cache_config: CacheConfig | None = None,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
prefix: str = "",
|
||||
|
||||
@@ -311,7 +311,6 @@ class Exaone4_5_ProcessingInfo(Qwen2VLProcessingInfo):
|
||||
def get_hf_processor(self, **kwargs: object) -> Exaone4_5_Processor:
|
||||
return self.ctx.get_hf_processor(
|
||||
Exaone4_5_Processor,
|
||||
use_fast=kwargs.pop("use_fast", True),
|
||||
**kwargs,
|
||||
)
|
||||
|
||||
|
||||
@@ -20,7 +20,7 @@ from itertools import islice
|
||||
|
||||
import torch
|
||||
from torch import nn
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
from vllm.compilation.decorators import support_torch_compile
|
||||
from vllm.config import CacheConfig, VllmConfig, get_current_vllm_config
|
||||
@@ -65,7 +65,7 @@ from .utils import (
|
||||
class ExaoneMoe(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
swiglu_limit: float | None = None,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
prefix: str = "",
|
||||
@@ -289,7 +289,7 @@ class ExaoneMoeAttention(nn.Module):
|
||||
class ExaoneMoeDecoderLayer(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
cache_config: CacheConfig | None = None,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
is_mtp: bool = False,
|
||||
|
||||
@@ -15,7 +15,7 @@ from transformers.models.gemma3n import (
|
||||
Gemma3nTextConfig,
|
||||
Gemma3nVisionConfig,
|
||||
)
|
||||
from transformers.models.siglip import SiglipImageProcessorFast
|
||||
from transformers.models.siglip import SiglipImageProcessor
|
||||
|
||||
from vllm.config import ModelConfig, SpeechToTextConfig, VllmConfig
|
||||
from vllm.config.multimodal import MultiModalDummyOptions
|
||||
@@ -239,7 +239,7 @@ class Gemma3nDummyInputsBuilder(BaseDummyInputsBuilder[Gemma3nProcessingInfo]):
|
||||
processor.feature_extractor
|
||||
)
|
||||
audio_len = audio_feature_extractor.fft_length
|
||||
image_processor: SiglipImageProcessorFast = processor.image_processor
|
||||
image_processor: SiglipImageProcessor = processor.image_processor
|
||||
img_width = image_processor.size.get("width", 224)
|
||||
img_height = image_processor.size.get("height", 224)
|
||||
|
||||
|
||||
@@ -28,7 +28,7 @@ from collections.abc import Callable, Iterable
|
||||
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
from vllm.config import VllmConfig
|
||||
from vllm.model_executor.layers.fused_moe import (
|
||||
@@ -68,7 +68,7 @@ from .utils import get_spec_layer_idx_from_weight_name, maybe_prefix
|
||||
class SharedHead(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
prefix: str,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
) -> None:
|
||||
|
||||
@@ -29,7 +29,7 @@ from collections.abc import Callable, Iterable
|
||||
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
from vllm.config import CacheConfig, ParallelConfig, VllmConfig
|
||||
from vllm.model_executor.layers.fused_moe import (
|
||||
@@ -60,7 +60,7 @@ from .utils import get_spec_layer_idx_from_weight_name, maybe_prefix
|
||||
class SharedHead(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
prefix: str,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
) -> None:
|
||||
@@ -80,7 +80,7 @@ class SharedHead(nn.Module):
|
||||
class Glm4MoeMultiTokenPredictorLayer(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
prefix: str,
|
||||
cache_config: CacheConfig | None = None,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
|
||||
@@ -36,7 +36,7 @@ from itertools import islice
|
||||
import torch
|
||||
from torch import nn
|
||||
from transformers import GraniteConfig
|
||||
from transformers.configuration_utils import PretrainedConfig
|
||||
from transformers.configuration_utils import PreTrainedConfig
|
||||
|
||||
from vllm.compilation.decorators import support_torch_compile
|
||||
from vllm.config import CacheConfig, VllmConfig
|
||||
@@ -72,7 +72,7 @@ from .utils import (
|
||||
|
||||
|
||||
def granite_layer_attn_params(
|
||||
config: PretrainedConfig, layer_idx: int
|
||||
config: PreTrainedConfig, layer_idx: int
|
||||
) -> tuple[int | None, float, bool]:
|
||||
"""Resolve one layer's sliding window, RoPE base and sink usage.
|
||||
|
||||
|
||||
@@ -31,7 +31,7 @@ from typing import Annotated
|
||||
import torch
|
||||
import torch.nn.functional as F
|
||||
from torch import nn
|
||||
from transformers import BatchFeature, PretrainedConfig
|
||||
from transformers import BatchFeature, PreTrainedConfig
|
||||
|
||||
from vllm.config import CacheConfig, ModelConfig, SpeechToTextConfig, VllmConfig
|
||||
from vllm.config.multimodal import MultiModalDummyOptions
|
||||
@@ -229,7 +229,7 @@ class GraniteSpeechDummyInputsBuilder(
|
||||
class GraniteSpeechEncoderProjector(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
cache_config: CacheConfig,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
prefix: str = "",
|
||||
@@ -287,7 +287,7 @@ class GraniteSpeechConformerFeedForward(nn.Module):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
prefix: str = "",
|
||||
):
|
||||
@@ -323,7 +323,7 @@ class GraniteSpeechConformerAttention(nn.Module):
|
||||
for more details.
|
||||
"""
|
||||
|
||||
def __init__(self, config: PretrainedConfig, prefix: str = ""):
|
||||
def __init__(self, config: PreTrainedConfig, prefix: str = ""):
|
||||
super().__init__()
|
||||
|
||||
inner_dim = config.dim_head * config.num_heads
|
||||
@@ -429,7 +429,7 @@ class GraniteSpeechConformerConvModule(nn.Module):
|
||||
convolutional layers.
|
||||
"""
|
||||
|
||||
def __init__(self, config: PretrainedConfig, prefix: str = ""):
|
||||
def __init__(self, config: PreTrainedConfig, prefix: str = ""):
|
||||
super().__init__()
|
||||
inner_dim = config.hidden_dim * config.conv_expansion_factor
|
||||
|
||||
@@ -460,7 +460,7 @@ class GraniteSpeechConformerBlock(nn.Module):
|
||||
"""Conformer block, consisting largely of linear layers,
|
||||
attention, and convolutional layers."""
|
||||
|
||||
def __init__(self, config: PretrainedConfig, prefix: str = ""):
|
||||
def __init__(self, config: PreTrainedConfig, prefix: str = ""):
|
||||
super().__init__()
|
||||
self.ff1 = GraniteSpeechConformerFeedForward(config, prefix=f"{prefix}.ff1")
|
||||
self.attn = GraniteSpeechConformerAttention(config, prefix=f"{prefix}.attn")
|
||||
@@ -486,7 +486,7 @@ class GraniteSpeechCTCEncoder(nn.Module):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
prefix: str,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
):
|
||||
@@ -622,7 +622,7 @@ class GraniteSpeechForConditionalGeneration(
|
||||
|
||||
def _build_encoder(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None,
|
||||
prefix: str,
|
||||
) -> "GraniteSpeechCTCEncoder":
|
||||
|
||||
@@ -3,7 +3,7 @@
|
||||
"""Inference-only IBM Granite Speech Plus model."""
|
||||
|
||||
import torch
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
from vllm.model_executor.layers.quantization import QuantizationConfig
|
||||
from vllm.multimodal import MULTIMODAL_REGISTRY
|
||||
@@ -68,7 +68,7 @@ class GraniteSpeechPlusForConditionalGeneration(GraniteSpeechForConditionalGener
|
||||
|
||||
def _build_encoder(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None,
|
||||
prefix: str,
|
||||
) -> GraniteSpeechCTCEncoder:
|
||||
|
||||
@@ -29,7 +29,7 @@ from itertools import islice
|
||||
|
||||
import torch
|
||||
from torch import nn
|
||||
from transformers.configuration_utils import PretrainedConfig
|
||||
from transformers.configuration_utils import PreTrainedConfig
|
||||
|
||||
from vllm.compilation.decorators import support_torch_compile
|
||||
from vllm.config import CacheConfig, VllmConfig
|
||||
@@ -141,7 +141,7 @@ class GraniteMoeMoE(nn.Module):
|
||||
class GraniteMoeAttention(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
hidden_size: int,
|
||||
num_heads: int,
|
||||
num_kv_heads: int,
|
||||
|
||||
@@ -10,7 +10,7 @@
|
||||
# --------------------------------------------------------
|
||||
|
||||
import torch
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
from vllm.model_executor.layers.quantization import QuantizationConfig
|
||||
from vllm.multimodal import MULTIMODAL_REGISTRY
|
||||
@@ -158,7 +158,7 @@ class H2OVLMultiModalProcessor(BaseInternVLMultiModalProcessor[H2OVLProcessingIn
|
||||
class H2OVLChatModel(InternVLChatModel):
|
||||
def _init_vision_model(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None,
|
||||
*,
|
||||
prefix: str,
|
||||
|
||||
@@ -25,7 +25,7 @@ from typing import Literal
|
||||
|
||||
import torch
|
||||
from torch import nn
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
from vllm.compilation.decorators import support_torch_compile
|
||||
from vllm.config import CacheConfig, VllmConfig
|
||||
@@ -122,7 +122,7 @@ class HrmTextAttention(nn.Module):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
layer_idx_in_stack: int,
|
||||
stack_kind: Literal["L", "H"],
|
||||
cache_config: CacheConfig | None = None,
|
||||
@@ -245,7 +245,7 @@ class HrmTextAttention(nn.Module):
|
||||
class HrmTextDecoderLayer(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
layer_idx_in_stack: int,
|
||||
stack_kind: Literal["L", "H"],
|
||||
cache_config: CacheConfig | None = None,
|
||||
@@ -310,7 +310,7 @@ class HrmTextStack(nn.Module):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
stack_kind: Literal["L", "H"],
|
||||
cache_config: CacheConfig | None = None,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
|
||||
@@ -31,7 +31,7 @@ from typing import Any
|
||||
|
||||
import torch
|
||||
from torch import nn
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
from vllm.compilation.decorators import support_torch_compile
|
||||
from vllm.config import CacheConfig, VllmConfig, get_current_vllm_config
|
||||
@@ -220,7 +220,7 @@ class HYV3MoEFused(nn.Module):
|
||||
class HYV3Attention(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
hidden_size: int,
|
||||
num_heads: int,
|
||||
num_kv_heads: int,
|
||||
@@ -365,7 +365,7 @@ class HYV3Attention(nn.Module):
|
||||
class HYV3DecoderLayer(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
cache_config: CacheConfig | None = None,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
prefix: str = "",
|
||||
|
||||
@@ -29,7 +29,7 @@ from collections.abc import Callable, Iterable
|
||||
import regex as re
|
||||
import torch
|
||||
from torch import nn
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
from vllm.config import CacheConfig, ModelConfig, VllmConfig
|
||||
from vllm.model_executor.layers.fused_moe import (
|
||||
@@ -59,7 +59,7 @@ from .utils import (
|
||||
)
|
||||
|
||||
|
||||
def _is_moe(config: PretrainedConfig) -> bool:
|
||||
def _is_moe(config: PreTrainedConfig) -> bool:
|
||||
return bool(
|
||||
getattr(config, "num_experts", None)
|
||||
and (
|
||||
@@ -69,7 +69,7 @@ def _is_moe(config: PretrainedConfig) -> bool:
|
||||
)
|
||||
|
||||
|
||||
def _get_cla_factor(config: PretrainedConfig) -> int:
|
||||
def _get_cla_factor(config: PreTrainedConfig) -> int:
|
||||
if not getattr(config, "use_cla", False):
|
||||
return 1
|
||||
return getattr(config, "cla_share_factor", 1)
|
||||
@@ -78,7 +78,7 @@ def _get_cla_factor(config: PretrainedConfig) -> int:
|
||||
class HYV3SharedHead(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
) -> None:
|
||||
super().__init__()
|
||||
@@ -93,7 +93,7 @@ class HYV3SharedHead(nn.Module):
|
||||
class HYV3MultiTokenPredictorLayer(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
prefix: str,
|
||||
model_config: ModelConfig,
|
||||
cache_config: CacheConfig | None = None,
|
||||
|
||||
@@ -13,7 +13,7 @@ from functools import partial
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
import torch.nn.functional as F
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
from vllm.compilation.decorators import (
|
||||
should_torch_compile_mm_encoder,
|
||||
@@ -47,7 +47,7 @@ NORM2FN = {
|
||||
|
||||
|
||||
class InternVisionEmbeddings(nn.Module):
|
||||
def __init__(self, config: PretrainedConfig):
|
||||
def __init__(self, config: PreTrainedConfig):
|
||||
super().__init__()
|
||||
self.config = config
|
||||
self.embed_dim = config.hidden_size
|
||||
@@ -119,7 +119,7 @@ class InternParallelAttention(nn.Module):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
*,
|
||||
num_dummy_heads: int = 0,
|
||||
@@ -222,7 +222,7 @@ class InternParallelAttention(nn.Module):
|
||||
class InternMLP(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
prefix: str = "",
|
||||
) -> None:
|
||||
@@ -264,7 +264,7 @@ class InternMLP(nn.Module):
|
||||
class InternVisionEncoderLayer(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
*,
|
||||
num_dummy_heads: int = 0,
|
||||
@@ -297,7 +297,7 @@ class InternVisionEncoderLayer(nn.Module):
|
||||
|
||||
def _init_attn(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None,
|
||||
*,
|
||||
num_dummy_heads: int,
|
||||
@@ -324,7 +324,7 @@ class InternVisionEncoderLayer(nn.Module):
|
||||
class InternVisionEncoder(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
*,
|
||||
num_hidden_layers_override: int | None = None,
|
||||
@@ -369,7 +369,7 @@ class InternVisionModel(nn.Module):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
*,
|
||||
num_hidden_layers_override: int | None = None,
|
||||
|
||||
@@ -8,7 +8,7 @@ from typing import Any
|
||||
|
||||
import torch
|
||||
from torch import nn
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
from vllm.compilation.decorators import support_torch_compile
|
||||
from vllm.config import CacheConfig, VllmConfig
|
||||
@@ -195,7 +195,7 @@ class InternLM2Attention(nn.Module):
|
||||
class InternLMDecoderLayer(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
cache_config: CacheConfig | None = None,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
prefix: str = "",
|
||||
|
||||
@@ -12,10 +12,10 @@ from typing import Annotated, Literal, TypeAlias, TypedDict
|
||||
import regex as re
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
from transformers import BatchFeature, InternVLProcessor, PretrainedConfig
|
||||
from transformers import BatchFeature, InternVLProcessor, PreTrainedConfig
|
||||
from transformers.activations import ACT2FN
|
||||
from transformers.models.got_ocr2.image_processing_got_ocr2_fast import (
|
||||
GotOcr2ImageProcessorFast,
|
||||
from transformers.models.got_ocr2.image_processing_got_ocr2 import (
|
||||
GotOcr2ImageProcessor,
|
||||
)
|
||||
from transformers.models.internvl.video_processing_internvl import (
|
||||
InternVLVideoProcessor,
|
||||
@@ -202,7 +202,7 @@ class InternS1ProcessingInfo(BaseProcessingInfo):
|
||||
processor: InternVLProcessor,
|
||||
mm_kwargs: Mapping[str, object],
|
||||
) -> int:
|
||||
image_processor: GotOcr2ImageProcessorFast = processor.image_processor
|
||||
image_processor: GotOcr2ImageProcessor = processor.image_processor
|
||||
|
||||
num_image_patches = image_processor.get_number_of_image_patches(
|
||||
image_height,
|
||||
@@ -612,7 +612,7 @@ class InternS1ForConditionalGeneration(
|
||||
|
||||
def _init_vision_model(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None,
|
||||
*,
|
||||
prefix: str,
|
||||
@@ -625,7 +625,7 @@ class InternS1ForConditionalGeneration(
|
||||
prefix=prefix,
|
||||
)
|
||||
|
||||
def _init_mlp1(self, config: PretrainedConfig) -> nn.Module:
|
||||
def _init_mlp1(self, config: PreTrainedConfig) -> nn.Module:
|
||||
return InternS1MultiModalProjector(config)
|
||||
|
||||
def pixel_shuffle(self, x, scale_factor=0.5):
|
||||
|
||||
@@ -30,7 +30,7 @@ from typing import Any, ClassVar
|
||||
|
||||
import torch
|
||||
from torch import nn
|
||||
from transformers import AutoProcessor, PretrainedConfig
|
||||
from transformers import AutoProcessor, PreTrainedConfig
|
||||
|
||||
from vllm.config import CacheConfig, VllmConfig
|
||||
from vllm.distributed import (
|
||||
@@ -563,7 +563,7 @@ class InternS1ProForConditionalGeneration(
|
||||
|
||||
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
|
||||
super(Qwen3VLForConditionalGeneration, self).__init__()
|
||||
config: PretrainedConfig = vllm_config.model_config.hf_config
|
||||
config: PreTrainedConfig = vllm_config.model_config.hf_config
|
||||
multimodal_config = vllm_config.model_config.get_multimodal_config()
|
||||
|
||||
self.config = config
|
||||
|
||||
@@ -11,7 +11,7 @@ from collections.abc import Iterable
|
||||
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
from transformers.utils import torch_int
|
||||
|
||||
from vllm.model_executor.layers.activation import get_act_fn
|
||||
@@ -65,7 +65,7 @@ class InternS1VisionPatchEmbeddings(nn.Module):
|
||||
|
||||
|
||||
class InternS1VisionEmbeddings(nn.Module):
|
||||
def __init__(self, config: PretrainedConfig):
|
||||
def __init__(self, config: PreTrainedConfig):
|
||||
super().__init__()
|
||||
self.config = config
|
||||
self.cls_token = nn.Parameter(torch.zeros(1, 1, config.hidden_size))
|
||||
@@ -166,7 +166,7 @@ class InternSdpaAttention(nn.Module):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
*,
|
||||
num_dummy_heads: int = 0,
|
||||
prefix: str = "",
|
||||
@@ -242,7 +242,7 @@ class InternSdpaAttention(nn.Module):
|
||||
class InternS1VisionMLP(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
prefix: str = "",
|
||||
) -> None:
|
||||
@@ -276,7 +276,7 @@ class InternS1VisionMLP(nn.Module):
|
||||
class InternS1VisionLayer(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
*,
|
||||
num_dummy_heads: int = 0,
|
||||
@@ -311,7 +311,7 @@ class InternS1VisionLayer(nn.Module):
|
||||
|
||||
def _init_attn(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None,
|
||||
*,
|
||||
num_dummy_heads: int,
|
||||
@@ -343,7 +343,7 @@ class InternS1VisionLayer(nn.Module):
|
||||
class InternS1VisionEncoder(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
*,
|
||||
num_hidden_layers_override: int | None = None,
|
||||
@@ -382,7 +382,7 @@ class InternS1VisionEncoder(nn.Module):
|
||||
class InternS1VisionModel(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
*,
|
||||
num_hidden_layers_override: int | None = None,
|
||||
|
||||
@@ -14,7 +14,7 @@ from typing import Annotated, Any, Literal, TypeAlias, TypedDict, TypeVar
|
||||
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
from transformers import BatchFeature, PretrainedConfig
|
||||
from transformers import BatchFeature, PreTrainedConfig
|
||||
|
||||
from vllm.config import VllmConfig
|
||||
from vllm.config.multimodal import MultiModalDummyOptions
|
||||
@@ -629,7 +629,7 @@ class InternVLChatModel(
|
||||
)
|
||||
|
||||
def _patch_quant_config(
|
||||
self, config: PretrainedConfig, quant_config: QuantizationConfig | None
|
||||
self, config: PreTrainedConfig, quant_config: QuantizationConfig | None
|
||||
):
|
||||
# the awq models from OpenGVLab missing `modules_to_not_convert`
|
||||
# patch the quant_config to add `modules_to_not_convert` back
|
||||
@@ -643,7 +643,7 @@ class InternVLChatModel(
|
||||
|
||||
def _init_vision_model(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None,
|
||||
*,
|
||||
prefix: str,
|
||||
@@ -663,7 +663,7 @@ class InternVLChatModel(
|
||||
prefix=prefix,
|
||||
)
|
||||
|
||||
def _init_mlp1(self, config: PretrainedConfig) -> nn.Module:
|
||||
def _init_mlp1(self, config: PreTrainedConfig) -> nn.Module:
|
||||
vit_hidden_size = config.vision_config.hidden_size
|
||||
llm_hidden_size = config.text_config.hidden_size
|
||||
|
||||
|
||||
@@ -22,7 +22,7 @@ from typing import Any
|
||||
|
||||
import torch
|
||||
from torch import nn
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
from vllm.compilation.decorators import support_torch_compile
|
||||
from vllm.config import CacheConfig, VllmConfig
|
||||
@@ -60,7 +60,7 @@ from .utils import (
|
||||
class LoopCoderAttention(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
hidden_size: int,
|
||||
num_heads: int,
|
||||
num_kv_heads: int,
|
||||
@@ -210,7 +210,7 @@ class LoopCoderAttention(nn.Module):
|
||||
class LoopCoderDecoderLayer(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
cache_config: CacheConfig | None = None,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
prefix: str = "",
|
||||
|
||||
@@ -10,7 +10,7 @@ import numpy as np
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
from einops import rearrange
|
||||
from transformers import BaseImageProcessor, PretrainedConfig
|
||||
from transformers import BaseImageProcessor, PreTrainedConfig
|
||||
from transformers.activations import GELUActivation
|
||||
from transformers.feature_extraction_utils import BatchFeature
|
||||
from transformers.modeling_outputs import BaseModelOutput, BaseModelOutputWithPooling
|
||||
@@ -191,7 +191,7 @@ KeyeVideoInputs: TypeAlias = KeyeVideoPixelInputs | KeyeVideoEmbeddingInputs
|
||||
|
||||
|
||||
class KeyeVisionEmbeddings(nn.Module):
|
||||
def __init__(self, config: PretrainedConfig):
|
||||
def __init__(self, config: PreTrainedConfig):
|
||||
super().__init__()
|
||||
self.config = config
|
||||
self.embed_dim = config.hidden_size
|
||||
@@ -349,7 +349,7 @@ class KeyeSiglipAttention(nn.Module):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
prefix: str = "",
|
||||
):
|
||||
@@ -487,7 +487,7 @@ class SigLIPRotaryEmbedding(nn.Module):
|
||||
class KeyeSiglipEncoderLayer(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
prefix: str = "",
|
||||
):
|
||||
@@ -539,7 +539,7 @@ class KeyeSiglipEncoderLayer(nn.Module):
|
||||
class KeyeSiglipEncoder(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
prefix: str = "",
|
||||
):
|
||||
@@ -634,7 +634,7 @@ class KeyeSiglipEncoder(nn.Module):
|
||||
class KeyeSiglipVisionTransformer(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
prefix: str = "",
|
||||
):
|
||||
@@ -710,7 +710,7 @@ class KeyeSiglipVisionTransformer(nn.Module):
|
||||
|
||||
|
||||
class KeyeSiglipVisionModel(nn.Module):
|
||||
config_class = PretrainedConfig
|
||||
config_class = PreTrainedConfig
|
||||
main_input_name = "pixel_values"
|
||||
|
||||
hf_to_vllm_mapper = WeightsMapper(
|
||||
@@ -724,7 +724,7 @@ class KeyeSiglipVisionModel(nn.Module):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
prefix: str = "",
|
||||
):
|
||||
@@ -787,8 +787,8 @@ class KeyeSiglipVisionModel(nn.Module):
|
||||
class Projector(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
text_config: PretrainedConfig,
|
||||
vision_config: PretrainedConfig,
|
||||
text_config: PreTrainedConfig,
|
||||
vision_config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
prefix: str = "",
|
||||
):
|
||||
@@ -1232,7 +1232,7 @@ class BaseKeyeModule(nn.Module, SupportsMultiModal):
|
||||
|
||||
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
|
||||
super().__init__()
|
||||
config: PretrainedConfig = vllm_config.model_config.hf_config
|
||||
config: PreTrainedConfig = vllm_config.model_config.hf_config
|
||||
quant_config = vllm_config.quant_config
|
||||
|
||||
self.config = config
|
||||
@@ -1264,8 +1264,8 @@ class BaseKeyeModule(nn.Module, SupportsMultiModal):
|
||||
@abstractmethod
|
||||
def _build_projector(
|
||||
self,
|
||||
text_config: PretrainedConfig,
|
||||
vision_config: PretrainedConfig,
|
||||
text_config: PreTrainedConfig,
|
||||
vision_config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
prefix: str = "",
|
||||
) -> nn.Module:
|
||||
@@ -1511,8 +1511,8 @@ class KeyeForConditionalGeneration(
|
||||
):
|
||||
def _build_projector(
|
||||
self,
|
||||
text_config: PretrainedConfig,
|
||||
vision_config: PretrainedConfig,
|
||||
text_config: PreTrainedConfig,
|
||||
vision_config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
prefix: str = "",
|
||||
) -> nn.Module:
|
||||
|
||||
@@ -9,7 +9,7 @@ import numpy as np
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
from einops import rearrange
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
from transformers.activations import GELUActivation
|
||||
from transformers.feature_extraction_utils import BatchFeature
|
||||
|
||||
@@ -197,8 +197,8 @@ KeyeVL1_5VideoInputs: TypeAlias = (
|
||||
class KeyeVL1_5Projector(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
text_config: PretrainedConfig,
|
||||
vision_config: PretrainedConfig,
|
||||
text_config: PreTrainedConfig,
|
||||
vision_config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
prefix: str = "",
|
||||
):
|
||||
@@ -509,15 +509,15 @@ class KeyeVL1_5ForConditionalGeneration(
|
||||
):
|
||||
def _build_projector(
|
||||
self,
|
||||
text_config: PretrainedConfig,
|
||||
vision_config: PretrainedConfig,
|
||||
text_config: PreTrainedConfig,
|
||||
vision_config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
prefix: str = "",
|
||||
) -> nn.Module:
|
||||
return KeyeVL1_5Projector(text_config, vision_config, quant_config, prefix)
|
||||
|
||||
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
|
||||
config: PretrainedConfig = vllm_config.model_config.hf_config
|
||||
config: PreTrainedConfig = vllm_config.model_config.hf_config
|
||||
self.merge_size = config.vision_config.spatial_merge_size
|
||||
super().__init__(vllm_config=vllm_config, prefix=prefix)
|
||||
|
||||
|
||||
@@ -334,7 +334,7 @@ class Siglip2Encoder(nn.Module):
|
||||
self attention layers. Each layer is a [`Siglip2EncoderLayer`].
|
||||
|
||||
Args:
|
||||
config: PretrainedConfig
|
||||
config: PreTrainedConfig
|
||||
|
||||
"""
|
||||
|
||||
|
||||
@@ -13,8 +13,8 @@ from transformers import BatchFeature
|
||||
from transformers.activations import ACT2FN
|
||||
from transformers.models.lfm2_vl import Lfm2VlProcessor
|
||||
from transformers.models.lfm2_vl.configuration_lfm2_vl import Lfm2VlConfig
|
||||
from transformers.models.lfm2_vl.image_processing_lfm2_vl_fast import (
|
||||
Lfm2VlImageProcessorFast,
|
||||
from transformers.models.lfm2_vl.image_processing_lfm2_vl import (
|
||||
Lfm2VlImageProcessor,
|
||||
find_closest_aspect_ratio,
|
||||
round_by_factor,
|
||||
)
|
||||
@@ -104,7 +104,7 @@ class Lfm2VLProcessingInfo(BaseProcessingInfo):
|
||||
def get_hf_processor(self, **kwargs):
|
||||
return self.ctx.get_hf_processor(Lfm2VlProcessor, **kwargs)
|
||||
|
||||
def get_image_processor(self, **kwargs: object) -> Lfm2VlImageProcessorFast:
|
||||
def get_image_processor(self, **kwargs: object) -> Lfm2VlImageProcessor:
|
||||
return self.get_hf_processor(**kwargs).image_processor
|
||||
|
||||
def get_default_tok_params(self) -> TokenizeParams:
|
||||
@@ -213,7 +213,7 @@ class Lfm2VLProcessingInfo(BaseProcessingInfo):
|
||||
processor: Lfm2VlProcessor,
|
||||
mm_kwargs: Mapping[str, object],
|
||||
) -> tuple[int, int, int]:
|
||||
image_processor: Lfm2VlImageProcessorFast = processor.image_processor
|
||||
image_processor: Lfm2VlImageProcessor = processor.image_processor
|
||||
|
||||
mm_kwargs = self.ctx.get_merged_mm_kwargs(mm_kwargs)
|
||||
downsample_factor = mm_kwargs.get(
|
||||
@@ -336,7 +336,7 @@ class Lfm2VLProcessingInfo(BaseProcessingInfo):
|
||||
processor: Lfm2VlProcessor,
|
||||
mm_kwargs: Mapping[str, object],
|
||||
) -> tuple[int, int]:
|
||||
image_processor: Lfm2VlImageProcessorFast = processor.image_processor
|
||||
image_processor: Lfm2VlImageProcessor = processor.image_processor
|
||||
|
||||
mm_kwargs = self.ctx.get_merged_mm_kwargs(mm_kwargs)
|
||||
downsample_factor = mm_kwargs.get(
|
||||
|
||||
@@ -12,7 +12,7 @@ from transformers import (
|
||||
CLIPVisionConfig,
|
||||
LlavaConfig,
|
||||
PixtralVisionConfig,
|
||||
PretrainedConfig,
|
||||
PreTrainedConfig,
|
||||
SiglipVisionConfig,
|
||||
)
|
||||
from transformers.models.llava import LlavaProcessor
|
||||
@@ -156,7 +156,7 @@ class LlavaMultiModalProjector(nn.Module):
|
||||
|
||||
|
||||
class LlavaLikeConfig(Protocol):
|
||||
vision_config: PretrainedConfig
|
||||
vision_config: PreTrainedConfig
|
||||
image_token_index: int
|
||||
vision_feature_select_strategy: str
|
||||
vision_feature_layer: int | list[int]
|
||||
|
||||
@@ -1307,7 +1307,6 @@ class LlavaOnevision2ProcessingInfo(BaseProcessingInfo):
|
||||
# ``_merge_mm_kwargs`` restricts caller ``mm_processor_kwargs`` to known
|
||||
# processor args and wraps values as hashable for the lru_cache.
|
||||
merged = _merge_mm_kwargs(model_config, AutoProcessor, **kwargs)
|
||||
merged.setdefault("use_fast", True)
|
||||
return _load_ov2_processor(
|
||||
model_config.model,
|
||||
model_config.revision,
|
||||
|
||||
@@ -39,7 +39,7 @@ from itertools import islice
|
||||
|
||||
import torch
|
||||
from torch import nn
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
from vllm.compilation.decorators import support_torch_compile
|
||||
from vllm.config import CacheConfig, VllmConfig
|
||||
@@ -80,7 +80,7 @@ from .utils import (
|
||||
logger = init_logger(__name__)
|
||||
|
||||
|
||||
class FlashConfig(PretrainedConfig):
|
||||
class FlashConfig(PreTrainedConfig):
|
||||
"""Flash model configuration."""
|
||||
|
||||
moe_intermediate_size: int
|
||||
|
||||
@@ -7,7 +7,7 @@ from collections.abc import Iterable
|
||||
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
from vllm.config import VllmConfig
|
||||
from vllm.model_executor.layers.layernorm import RMSNorm
|
||||
@@ -30,7 +30,7 @@ from .utils import maybe_prefix
|
||||
class LongCatMultiTokenPredictorLayer(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
prefix: str,
|
||||
vllm_config: VllmConfig,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
@@ -353,7 +353,7 @@ class LongCatFlashMTP(nn.Module):
|
||||
return name
|
||||
|
||||
def get_spec_layer_idx_from_weight_name(
|
||||
self, config: PretrainedConfig, weight_name: str
|
||||
self, config: PreTrainedConfig, weight_name: str
|
||||
) -> int | None:
|
||||
if "model.mtp" in weight_name:
|
||||
return config.num_hidden_layers * 2
|
||||
|
||||
@@ -21,7 +21,7 @@ import torch.nn as nn
|
||||
import torch.nn.functional as F
|
||||
from einops import rearrange, repeat
|
||||
from transformers.activations import ACT2FN
|
||||
from transformers.configuration_utils import PretrainedConfig
|
||||
from transformers.configuration_utils import PreTrainedConfig
|
||||
from transformers.modeling_utils import PreTrainedModel
|
||||
from transformers.models.qwen2.configuration_qwen2 import Qwen2Config
|
||||
from transformers.models.qwen2.modeling_qwen2 import Qwen2Model
|
||||
@@ -518,7 +518,7 @@ class ResidualVectorQuantizer(nn.Module):
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
|
||||
class MiMoAudioTokenizerConfig(PretrainedConfig):
|
||||
class MiMoAudioTokenizerConfig(PreTrainedConfig):
|
||||
model_type = "mimo_audio_tokenizer"
|
||||
|
||||
def __init__(
|
||||
|
||||
@@ -24,7 +24,7 @@ from collections.abc import Iterable
|
||||
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
from vllm.config import CacheConfig, ModelConfig, VllmConfig
|
||||
from vllm.model_executor.layers.layernorm import RMSNorm
|
||||
@@ -43,7 +43,7 @@ from .utils import AutoWeightsLoader, WeightsMapper, maybe_prefix
|
||||
class MiMoMultiTokenPredictorLayer(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
prefix: str,
|
||||
model_config: ModelConfig,
|
||||
cache_config: CacheConfig | None = None,
|
||||
|
||||
@@ -22,7 +22,7 @@ from collections.abc import Iterable
|
||||
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
from vllm.config import VllmConfig
|
||||
from vllm.distributed import (
|
||||
@@ -62,7 +62,7 @@ class MiMoV2MTPLayer(nn.Module):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
prefix: str,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
) -> None:
|
||||
@@ -145,7 +145,7 @@ class _MiMoV2MTPLayers(nn.Module):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
num_mtp_layers: int,
|
||||
quant_config: QuantizationConfig | None,
|
||||
prefix: str,
|
||||
|
||||
@@ -10,7 +10,7 @@ import numpy as np
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
import torch.nn.functional as F
|
||||
from transformers import BatchFeature, PretrainedConfig
|
||||
from transformers import BatchFeature, PreTrainedConfig
|
||||
from transformers.models.qwen2_vl.image_processing_qwen2_vl import smart_resize
|
||||
from typing_extensions import TypedDict
|
||||
|
||||
@@ -418,7 +418,7 @@ class MiMoVisionTransformer(nn.Module):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
vision_cfg: PretrainedConfig,
|
||||
vision_cfg: PreTrainedConfig,
|
||||
*,
|
||||
norm_eps: float = 1e-6,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
|
||||
@@ -31,7 +31,7 @@ from typing import Any
|
||||
|
||||
import torch
|
||||
from torch import nn
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
from vllm.compilation.decorators import support_torch_compile
|
||||
from vllm.config import CacheConfig, VllmConfig
|
||||
@@ -316,7 +316,7 @@ class MiniCPMAttention(nn.Module):
|
||||
class MiniCPMDecoderLayer(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
cache_config: CacheConfig | None = None,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
prefix: str = "",
|
||||
@@ -427,7 +427,7 @@ class MiniCPMModel(nn.Module, EagleModelMixin):
|
||||
def _init_layers(
|
||||
self,
|
||||
prefix: str,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
cache_config: CacheConfig | None,
|
||||
quant_config: QuantizationConfig | None,
|
||||
):
|
||||
|
||||
@@ -27,7 +27,7 @@
|
||||
|
||||
import torch
|
||||
from torch import nn
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
from vllm.config import CacheConfig, VllmConfig
|
||||
from vllm.distributed import get_tensor_model_parallel_world_size
|
||||
@@ -52,7 +52,7 @@ from .utils import make_layers
|
||||
class MiniCPM3Attention(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
hidden_size: int,
|
||||
num_heads: int,
|
||||
qk_nope_head_dim: int,
|
||||
@@ -208,7 +208,7 @@ class MiniCPM3Model(MiniCPMModel):
|
||||
def _init_layers(
|
||||
self,
|
||||
prefix: str,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
cache_config: CacheConfig | None,
|
||||
quant_config: QuantizationConfig | None,
|
||||
):
|
||||
|
||||
@@ -30,7 +30,7 @@ from typing import TYPE_CHECKING
|
||||
|
||||
import torch
|
||||
from torch import nn
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
from vllm.compilation.decorators import support_torch_compile
|
||||
from vllm.config import CacheConfig, VllmConfig
|
||||
@@ -68,7 +68,7 @@ else:
|
||||
class EagleMiniCPMDecoderLayer(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
cache_config: CacheConfig | None = None,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
prefix: str = "",
|
||||
@@ -189,7 +189,7 @@ class EagleMiniCPMModel(nn.Module):
|
||||
def _init_layers(
|
||||
self,
|
||||
prefix: str,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
cache_config: CacheConfig | None,
|
||||
quant_config: QuantizationConfig | None,
|
||||
start_layer: int,
|
||||
|
||||
@@ -36,7 +36,7 @@ import torch
|
||||
import torch.types
|
||||
from torch import nn
|
||||
from torch.nn.init import trunc_normal_
|
||||
from transformers import BatchFeature, PretrainedConfig
|
||||
from transformers import BatchFeature, PreTrainedConfig
|
||||
from transformers.dynamic_module_utils import (
|
||||
get_class_from_dynamic_module,
|
||||
resolve_trust_remote_code,
|
||||
@@ -468,7 +468,7 @@ class Resampler4_5(Resampler2_5):
|
||||
return x
|
||||
|
||||
|
||||
def get_version_by_config(config: PretrainedConfig) -> tuple[int, ...]:
|
||||
def get_version_by_config(config: PreTrainedConfig) -> tuple[int, ...]:
|
||||
version_float = getattr(config, "version", None)
|
||||
|
||||
# The old configs do not include version number
|
||||
@@ -1235,7 +1235,7 @@ class MiniCPMVBaseModel(nn.Module, SupportsMultiModal, SupportsPP):
|
||||
self.use_data_parallel = multimodal_config.mm_encoder_tp_mode == "data"
|
||||
super().__init__()
|
||||
# All MiniCPM-V models disable `tie_word_embeddings` but
|
||||
# `PretrainedConfig.tie_word_embeddings` defaults to True; we cannot
|
||||
# `PreTrainedConfig.tie_word_embeddings` defaults to True; we cannot
|
||||
# check `tie_word_embeddings` until vLLM integrate MiniCPM-V model
|
||||
# and config class
|
||||
self.config = config
|
||||
@@ -1425,7 +1425,7 @@ class MiniCPMVBaseModel(nn.Module, SupportsMultiModal, SupportsPP):
|
||||
|
||||
def init_vision_module(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None,
|
||||
prefix: str = "",
|
||||
) -> nn.Module:
|
||||
@@ -1957,7 +1957,7 @@ class MiniCPMV2_0(MiniCPMVBaseModel):
|
||||
|
||||
def init_vision_module(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None,
|
||||
prefix: str = "",
|
||||
) -> nn.Module:
|
||||
@@ -2058,7 +2058,7 @@ class MiniCPMV2_5(_MiniCPMVEncoderCudaGraphMixin, SupportsLoRA):
|
||||
|
||||
def init_vision_module(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None,
|
||||
prefix: str = "",
|
||||
) -> nn.Module:
|
||||
@@ -2154,7 +2154,7 @@ class MiniCPMV2_6(_MiniCPMVEncoderCudaGraphMixin, SupportsLoRA):
|
||||
|
||||
def init_vision_module(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
prefix: str = "",
|
||||
) -> nn.Module:
|
||||
@@ -2255,7 +2255,7 @@ class MiniCPMV4_0(_MiniCPMVEncoderCudaGraphMixin, SupportsLoRA):
|
||||
|
||||
def init_vision_module(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
prefix: str = "",
|
||||
) -> nn.Module:
|
||||
@@ -2354,7 +2354,7 @@ class MiniCPMV4_5(MiniCPMVBaseModel, SupportsLoRA):
|
||||
|
||||
def init_vision_module(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
prefix: str = "",
|
||||
) -> nn.Module:
|
||||
|
||||
@@ -29,7 +29,7 @@ from typing import Any
|
||||
|
||||
import torch
|
||||
from torch import nn
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
from vllm.compilation.decorators import support_torch_compile
|
||||
from vllm.config import CacheConfig, ModelConfig, VllmConfig
|
||||
@@ -70,7 +70,7 @@ from .utils import (
|
||||
class MiniMaxM2MoE(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
prefix: str = "",
|
||||
):
|
||||
@@ -249,7 +249,7 @@ class MiniMaxM2Attention(nn.Module):
|
||||
class MiniMaxM2DecoderLayer(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
prefix: str,
|
||||
model_config: ModelConfig,
|
||||
cache_config: CacheConfig | None = None,
|
||||
|
||||
@@ -25,7 +25,7 @@ from torch import nn
|
||||
from transformers import BatchFeature, Llama4Config, Llama4VisionConfig
|
||||
from transformers.image_utils import SizeDict
|
||||
from transformers.models.llama4 import Llama4Processor
|
||||
from transformers.models.llama4.image_processing_llama4_fast import (
|
||||
from transformers.models.llama4.image_processing_llama4 import (
|
||||
find_supported_resolutions,
|
||||
get_best_fit,
|
||||
)
|
||||
@@ -554,9 +554,7 @@ class Mllama4ProcessingInfo(BaseProcessingInfo):
|
||||
return self.ctx.get_hf_config(Llama4Config)
|
||||
|
||||
def get_hf_processor(self, **kwargs: object) -> Llama4Processor:
|
||||
return self.ctx.get_hf_processor(
|
||||
Llama4Processor, use_fast=kwargs.pop("use_fast", True), **kwargs
|
||||
)
|
||||
return self.ctx.get_hf_processor(Llama4Processor, **kwargs)
|
||||
|
||||
def get_supported_mm_limits(self) -> Mapping[str, int | None]:
|
||||
# Although vLLM can support more images from an infra capability
|
||||
|
||||
@@ -16,7 +16,7 @@ from einops import rearrange
|
||||
from transformers import (
|
||||
BaseImageProcessor,
|
||||
BatchFeature,
|
||||
PretrainedConfig,
|
||||
PreTrainedConfig,
|
||||
)
|
||||
|
||||
from vllm.compilation.decorators import support_torch_compile
|
||||
@@ -405,7 +405,7 @@ class MolmoAttention(nn.Module):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
cache_config: CacheConfig | None = None,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
prefix: str = "",
|
||||
@@ -515,7 +515,7 @@ class LanguageModelMLP(nn.Module):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
input_dim: int | None = None,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
prefix: str = "",
|
||||
@@ -557,7 +557,7 @@ class ImageProjectorMLP(nn.Module):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
input_dim: int | None = None,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
prefix: str = "",
|
||||
@@ -598,7 +598,7 @@ class ImageProjectorMLP(nn.Module):
|
||||
class MolmoDecoderLayer(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
cache_config: CacheConfig | None = None,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
prefix: str = "",
|
||||
@@ -681,7 +681,7 @@ class MolmoVisionBackbone(nn.Module, SupportsQuant):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
vision_config: VisionBackboneConfig,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
prefix: str = "",
|
||||
|
||||
@@ -16,7 +16,7 @@ from transformers import (
|
||||
BaseImageProcessor,
|
||||
BaseVideoProcessor,
|
||||
BatchFeature,
|
||||
PretrainedConfig,
|
||||
PreTrainedConfig,
|
||||
ProcessorMixin,
|
||||
)
|
||||
from transformers.image_utils import ImageInput
|
||||
@@ -1305,7 +1305,7 @@ def exif_transpose(
|
||||
|
||||
def build_flat_image_bool_length(
|
||||
image_grids: torch.LongTensor,
|
||||
hf_config: PretrainedConfig,
|
||||
hf_config: PreTrainedConfig,
|
||||
image_use_col_tokens: bool = True,
|
||||
use_single_crop_col_tokens: bool | None = None,
|
||||
use_single_crop_start_token: bool = True,
|
||||
@@ -1393,7 +1393,7 @@ def build_flat_image_bool_length(
|
||||
|
||||
def build_flat_video_bool_length(
|
||||
video_grids: torch.LongTensor,
|
||||
hf_config: PretrainedConfig,
|
||||
hf_config: PreTrainedConfig,
|
||||
) -> tuple[torch.LongTensor, torch.LongTensor]:
|
||||
image_patch_id = hf_config.image_patch_id
|
||||
frame_start_id = hf_config.frame_start_token_id
|
||||
|
||||
@@ -13,7 +13,7 @@ import regex as re
|
||||
import torch
|
||||
import torch.nn.functional as F
|
||||
from torch import nn
|
||||
from transformers import BatchFeature, PretrainedConfig, Qwen3Config
|
||||
from transformers import BatchFeature, PreTrainedConfig, Qwen3Config
|
||||
from transformers.models.whisper import WhisperFeatureExtractor
|
||||
from typing_extensions import TypedDict
|
||||
|
||||
@@ -263,7 +263,7 @@ class MossAudioEncoderConfig:
|
||||
return cls(**values)
|
||||
|
||||
|
||||
class MossAudioConfig(PretrainedConfig):
|
||||
class MossAudioConfig(PreTrainedConfig):
|
||||
model_type = "moss_audio"
|
||||
is_composition = True
|
||||
|
||||
|
||||
@@ -16,7 +16,7 @@ from typing import Annotated, Literal, TypeAlias, TypedDict, cast
|
||||
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
from transformers import BatchFeature, PretrainedConfig
|
||||
from transformers import BatchFeature, PreTrainedConfig
|
||||
|
||||
from vllm import envs
|
||||
from vllm.config import VllmConfig
|
||||
@@ -238,7 +238,7 @@ class NanoNemotronVLProcessingInfo(BaseProcessingInfo):
|
||||
return self.ctx.get_mm_config().video_pruning_rate
|
||||
|
||||
@property
|
||||
def sound_config(self) -> PretrainedConfig | None:
|
||||
def sound_config(self) -> PreTrainedConfig | None:
|
||||
return getattr(self.get_hf_config(), "sound_config", None)
|
||||
|
||||
def get_default_tok_params(self) -> TokenizeParams:
|
||||
|
||||
@@ -17,7 +17,7 @@ from einops import rearrange
|
||||
from transformers import (
|
||||
BartConfig,
|
||||
BatchFeature,
|
||||
PretrainedConfig,
|
||||
PreTrainedConfig,
|
||||
)
|
||||
|
||||
from vllm.config import CacheConfig, VllmConfig
|
||||
@@ -440,7 +440,7 @@ class RadioWithNeck(nn.Module):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
prefix: str = "",
|
||||
):
|
||||
@@ -480,7 +480,7 @@ class RadioWithNeck(nn.Module):
|
||||
|
||||
def get_vit_model_from_radio_config(
|
||||
self,
|
||||
hf_config: PretrainedConfig,
|
||||
hf_config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
) -> RadioModel:
|
||||
hf_config_vision = hf_config.encoder
|
||||
|
||||
@@ -5,7 +5,7 @@ from collections.abc import Iterable
|
||||
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
from transformers import AutoModel, PretrainedConfig
|
||||
from transformers import AutoModel, PreTrainedConfig
|
||||
|
||||
from vllm.config import VllmConfig
|
||||
from vllm.model_executor.layers.linear import ReplicatedLinear
|
||||
@@ -146,7 +146,7 @@ class LlamaNemotronVLChatModel(nn.Module, SupportsMultiModal, SupportsPP, Suppor
|
||||
)
|
||||
|
||||
def _patch_quant_config(
|
||||
self, config: PretrainedConfig, quant_config: QuantizationConfig | None
|
||||
self, config: PreTrainedConfig, quant_config: QuantizationConfig | None
|
||||
):
|
||||
# the awq models from OpenGVLab missing `modules_to_not_convert`
|
||||
# patch the quant_config to add `modules_to_not_convert` back
|
||||
@@ -160,7 +160,7 @@ class LlamaNemotronVLChatModel(nn.Module, SupportsMultiModal, SupportsPP, Suppor
|
||||
|
||||
def _init_vision_model(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None,
|
||||
*,
|
||||
prefix: str,
|
||||
@@ -172,7 +172,7 @@ class LlamaNemotronVLChatModel(nn.Module, SupportsMultiModal, SupportsPP, Suppor
|
||||
|
||||
def _init_mlp1(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
vit_hidden_size: int | None = None,
|
||||
vision_projection_hidden_size: int | None = None,
|
||||
) -> nn.Module:
|
||||
@@ -505,7 +505,7 @@ class LlamaNemotronVLForEmbedding(LlamaNemotronVLChatModel, VllmModelForPooling)
|
||||
|
||||
def _init_vision_model(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config,
|
||||
*,
|
||||
prefix: str,
|
||||
@@ -520,7 +520,7 @@ class LlamaNemotronVLForEmbedding(LlamaNemotronVLChatModel, VllmModelForPooling)
|
||||
|
||||
def _init_mlp1(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
vit_hidden_size: int | None = None,
|
||||
vision_projection_hidden_size: int | None = None,
|
||||
) -> nn.Module:
|
||||
|
||||
@@ -11,7 +11,7 @@ from collections.abc import Mapping
|
||||
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
from vllm.config.multimodal import MultiModalDummyOptions
|
||||
from vllm.inputs import MultiModalDataDict
|
||||
@@ -161,7 +161,7 @@ class NVLMMultiModalProcessor(BaseInternVLMultiModalProcessor[NVLMProcessingInfo
|
||||
dummy_inputs=NVLMDummyInputsBuilder,
|
||||
)
|
||||
class NVLM_D_Model(InternVLChatModel):
|
||||
def _init_mlp1(self, config: PretrainedConfig) -> nn.Module:
|
||||
def _init_mlp1(self, config: PreTrainedConfig) -> nn.Module:
|
||||
vit_hidden_size = config.vision_config.hidden_size
|
||||
llm_intermediate_size = config.text_config.intermediate_size
|
||||
llm_hidden_size = config.text_config.hidden_size
|
||||
@@ -179,7 +179,7 @@ class NVLM_D_Model(InternVLChatModel):
|
||||
|
||||
def _init_vision_model(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None,
|
||||
*,
|
||||
prefix: str,
|
||||
|
||||
@@ -26,7 +26,7 @@ from typing import Any
|
||||
|
||||
import torch
|
||||
from torch import nn
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
from vllm.compilation.decorators import support_torch_compile
|
||||
from vllm.config import CacheConfig, ParallelConfig, VllmConfig
|
||||
@@ -134,7 +134,7 @@ class OpenPanguMLP(nn.Module):
|
||||
class OpenPanguMoE(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
parallel_config: ParallelConfig,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
prefix: str = "",
|
||||
@@ -240,7 +240,7 @@ class OpenPanguMoE(nn.Module):
|
||||
class OpenPanguMLAAttention(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
hidden_size: int,
|
||||
num_heads: int,
|
||||
qk_nope_head_dim: int,
|
||||
@@ -389,7 +389,7 @@ class OpenPanguMLAAttention(nn.Module):
|
||||
class OpenPanguEmbeddedAttention(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
hidden_size: int,
|
||||
num_heads: int,
|
||||
num_kv_heads: int,
|
||||
@@ -501,7 +501,7 @@ class OpenPanguEmbeddedAttention(nn.Module):
|
||||
|
||||
def _init_rotary_emb(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None,
|
||||
) -> None:
|
||||
is_neox_style = True
|
||||
@@ -522,7 +522,7 @@ class OpenPanguEmbeddedAttention(nn.Module):
|
||||
class OpenPanguSinkAttention(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
hidden_size: int,
|
||||
num_heads: int,
|
||||
num_kv_heads: int,
|
||||
@@ -743,7 +743,7 @@ class OpenPanguSinkAttention(nn.Module):
|
||||
|
||||
def _init_rotary_emb(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
rope_parameters: dict[str, Any] | None,
|
||||
quant_config: QuantizationConfig | None,
|
||||
) -> None:
|
||||
@@ -769,7 +769,7 @@ class OpenPanguSinkAttention(nn.Module):
|
||||
class OpenPanguDecoderLayer(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
prefix: str,
|
||||
vllm_config: VllmConfig,
|
||||
) -> None:
|
||||
|
||||
@@ -686,7 +686,6 @@ class OpenPanguVLProcessingInfo(Qwen2_5_VLProcessingInfo):
|
||||
kwargs["fps"] = fps
|
||||
|
||||
return self.ctx.get_hf_processor(
|
||||
use_fast=kwargs.pop("use_fast", True),
|
||||
**kwargs,
|
||||
)
|
||||
|
||||
|
||||
@@ -13,7 +13,7 @@ from typing import Any
|
||||
|
||||
import torch
|
||||
from torch import nn
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
from vllm.compilation.decorators import support_torch_compile
|
||||
from vllm.config import CacheConfig, VllmConfig
|
||||
@@ -163,7 +163,7 @@ class OrionAttention(nn.Module):
|
||||
class OrionDecoderLayer(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
cache_config: CacheConfig | None = None,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
prefix: str = "",
|
||||
|
||||
@@ -26,7 +26,7 @@ import torch
|
||||
import torch.nn as nn
|
||||
from torch import Tensor
|
||||
from torch.nn.functional import gumbel_softmax, pad, softmax
|
||||
from transformers import BatchFeature, PretrainedConfig
|
||||
from transformers import BatchFeature, PreTrainedConfig
|
||||
|
||||
from vllm.config import VllmConfig
|
||||
from vllm.config.multimodal import MultiModalDummyOptions
|
||||
@@ -93,7 +93,7 @@ def st_argmax(y_soft: torch.Tensor, dim: int): # straight-through softmax
|
||||
class VisualTokenizer(torch.nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
prefix: str = "",
|
||||
):
|
||||
@@ -120,7 +120,7 @@ class VisualTokenizer(torch.nn.Module):
|
||||
|
||||
def _init_backbone(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
prefix: str = "",
|
||||
) -> nn.Module:
|
||||
@@ -419,7 +419,7 @@ class Ovis(nn.Module, SupportsMultiModal, SupportsPP):
|
||||
config = vllm_config.model_config.hf_config
|
||||
quant_config = vllm_config.quant_config
|
||||
|
||||
self.config: PretrainedConfig = config
|
||||
self.config: PreTrainedConfig = config
|
||||
|
||||
with self._mark_language_model(vllm_config):
|
||||
self.llm = init_vllm_registered_model(
|
||||
|
||||
@@ -8,7 +8,7 @@ from typing import Annotated, Literal, TypedDict
|
||||
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
from transformers import BaseImageProcessor, BatchFeature, PretrainedConfig
|
||||
from transformers import BaseImageProcessor, BatchFeature, PreTrainedConfig
|
||||
|
||||
from vllm.config import VllmConfig
|
||||
from vllm.config.multimodal import MultiModalDummyOptions
|
||||
@@ -95,7 +95,7 @@ class VisualTokenizer(torch.nn.Module):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
visual_vocab_size: int,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
prefix: str = "",
|
||||
@@ -121,7 +121,7 @@ class VisualTokenizer(torch.nn.Module):
|
||||
|
||||
def _init_backbone(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
prefix: str = "",
|
||||
):
|
||||
@@ -438,7 +438,7 @@ class Ovis2_5(nn.Module, SupportsMultiModal, SupportsPP):
|
||||
config = vllm_config.model_config.hf_config
|
||||
quant_config = vllm_config.quant_config
|
||||
|
||||
self.config: PretrainedConfig = config
|
||||
self.config: PreTrainedConfig = config
|
||||
|
||||
with self._mark_language_model(vllm_config):
|
||||
self.llm = init_vllm_registered_model(
|
||||
|
||||
@@ -23,7 +23,7 @@ import numpy as np
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
from einops import rearrange
|
||||
from transformers import BaseImageProcessor, BatchFeature, PretrainedConfig
|
||||
from transformers import BaseImageProcessor, BatchFeature, PreTrainedConfig
|
||||
from transformers.activations import GELUActivation
|
||||
from transformers.image_utils import ChannelDimension
|
||||
from transformers.modeling_outputs import (
|
||||
@@ -318,8 +318,8 @@ class PaddleOCRVLMultiModalProcessor(
|
||||
class Projector(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
text_config: PretrainedConfig,
|
||||
vision_config: PretrainedConfig,
|
||||
text_config: PreTrainedConfig,
|
||||
vision_config: PreTrainedConfig,
|
||||
prefix: str = "",
|
||||
):
|
||||
super().__init__()
|
||||
@@ -392,7 +392,7 @@ class PaddleOCRImagePixelInputs(TensorSchema):
|
||||
|
||||
|
||||
class SiglipVisionEmbeddings(nn.Module):
|
||||
def __init__(self, config: PretrainedConfig):
|
||||
def __init__(self, config: PreTrainedConfig):
|
||||
super().__init__()
|
||||
self.config = config
|
||||
self.embed_dim = config.hidden_size
|
||||
@@ -682,7 +682,7 @@ class SigLIPRotaryEmbedding(nn.Module):
|
||||
class SiglipEncoderLayer(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
prefix: str = "",
|
||||
):
|
||||
@@ -735,7 +735,7 @@ class SiglipEncoderLayer(nn.Module):
|
||||
class SiglipEncoder(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
prefix: str = "",
|
||||
):
|
||||
@@ -837,7 +837,7 @@ class SiglipEncoder(nn.Module):
|
||||
class SiglipVisionTransformer(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
prefix: str = "",
|
||||
):
|
||||
|
||||
@@ -10,7 +10,7 @@ import numpy as np
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
from transformers import ParakeetEncoder as HFParakeetEncoder
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
from transformers.audio_utils import mel_filter_bank
|
||||
|
||||
from vllm.logger import init_logger
|
||||
@@ -47,7 +47,7 @@ class ParakeetProjection(nn.Module):
|
||||
class ProjectedParakeet(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
*,
|
||||
dtype: torch.dtype,
|
||||
llm_hidden_size: int,
|
||||
@@ -135,7 +135,7 @@ LOG_ZERO_GUARD_VALUE = 2**-24
|
||||
|
||||
|
||||
class ParakeetExtractor:
|
||||
def __init__(self, config: PretrainedConfig) -> None:
|
||||
def __init__(self, config: PreTrainedConfig) -> None:
|
||||
self.config = ExtractorConfig.from_hf_config(config)
|
||||
"""`config` is named *exactly* for `._get_subsampling_output_length` below"""
|
||||
self._clip_target_samples = int(
|
||||
@@ -329,6 +329,6 @@ class ParakeetExtractor:
|
||||
}
|
||||
|
||||
@staticmethod
|
||||
def audio_length(raw_config: PretrainedConfig, audio_tokens: int) -> int:
|
||||
def audio_length(raw_config: PreTrainedConfig, audio_tokens: int) -> int:
|
||||
config = ExtractorConfig.from_hf_config(raw_config)
|
||||
return int(audio_tokens * config.subsampling_factor * config.hop_length)
|
||||
|
||||
@@ -24,7 +24,7 @@ import torch.nn as nn
|
||||
from transformers import (
|
||||
BatchFeature,
|
||||
CLIPVisionConfig,
|
||||
PretrainedConfig,
|
||||
PreTrainedConfig,
|
||||
ProcessorMixin,
|
||||
)
|
||||
|
||||
@@ -95,7 +95,7 @@ CLIP_VIT_LARGE_PATCH14_336_CONFIG = CLIPVisionConfig(
|
||||
|
||||
|
||||
def _init_img_processor(
|
||||
hf_config: PretrainedConfig,
|
||||
hf_config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None,
|
||||
prefix: str = "",
|
||||
) -> CLIPVisionModel:
|
||||
@@ -165,7 +165,7 @@ class Phi3HDImageEmbedding(nn.Module):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None,
|
||||
prefix: str = "",
|
||||
) -> None:
|
||||
|
||||
@@ -10,7 +10,7 @@ import torch
|
||||
import torch.nn as nn
|
||||
from transformers import (
|
||||
BatchFeature,
|
||||
PretrainedConfig,
|
||||
PreTrainedConfig,
|
||||
ProcessorMixin,
|
||||
SequenceFeatureExtractor,
|
||||
SiglipVisionConfig,
|
||||
@@ -123,7 +123,7 @@ class Phi4MMImageEncoder(nn.Module):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
config: PreTrainedConfig,
|
||||
quant_config: QuantizationConfig | None,
|
||||
prefix: str = "",
|
||||
model_dir: str = "",
|
||||
|
||||
@@ -18,7 +18,7 @@ from torch.distributed.algorithms._checkpoint.checkpoint_wrapper import (
|
||||
CheckpointWrapper,
|
||||
)
|
||||
from torch.distributed.fsdp.fully_sharded_data_parallel import FullyShardedDataParallel
|
||||
from transformers import PretrainedConfig
|
||||
from transformers import PreTrainedConfig
|
||||
|
||||
from vllm.model_executor.models.phi4mm_utils import (
|
||||
AbsolutePositionalEncoding,
|
||||
@@ -1152,7 +1152,7 @@ class WindowQformer(nn.Module):
|
||||
class AudioEmbedding(nn.Module):
|
||||
"""Image embedding."""
|
||||
|
||||
def __init__(self, config: PretrainedConfig, **kwargs: Any) -> None:
|
||||
def __init__(self, config: PreTrainedConfig, **kwargs: Any) -> None:
|
||||
super().__init__()
|
||||
self.config = config
|
||||
# n_embed or hidden_size for text LM
|
||||
|
||||
@@ -11,7 +11,7 @@ from typing import Annotated, Any, Literal
|
||||
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
from transformers import BatchFeature, PretrainedConfig, Siglip2VisionConfig
|
||||
from transformers import BatchFeature, PreTrainedConfig, Siglip2VisionConfig
|
||||
|
||||
from vllm.config import VllmConfig
|
||||
from vllm.config.multimodal import MultiModalDummyOptions
|
||||
@@ -235,7 +235,7 @@ class Phi4ForCausalLMV(nn.Module, SupportsMultiModal, SupportsPP):
|
||||
def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None:
|
||||
super().__init__()
|
||||
|
||||
config: PretrainedConfig = vllm_config.model_config.hf_config
|
||||
config: PreTrainedConfig = vllm_config.model_config.hf_config
|
||||
quant_config = vllm_config.quant_config
|
||||
self.config = config
|
||||
|
||||
|
||||
@@ -29,7 +29,7 @@ from itertools import islice
|
||||
|
||||
import torch
|
||||
from torch import nn
|
||||
from transformers.configuration_utils import PretrainedConfig
|
||||
from transformers.configuration_utils import PreTrainedConfig
|
||||
|
||||
from vllm.compilation.decorators import support_torch_compile
|
||||
from vllm.config import CacheConfig, VllmConfig
|
||||
@@ -62,7 +62,7 @@ from .utils import (
|
||||
)
|
||||
|
||||
|
||||
class PhiMoEConfig(PretrainedConfig):
|
||||
class PhiMoEConfig(PreTrainedConfig):
|
||||
model_type = "phimoe"
|
||||
keys_to_ignore_at_inference = ["past_key_values"]
|
||||
|
||||
|
||||
Some files were not shown because too many files have changed in this diff Show More
Reference in New Issue
Block a user