[Migration] Migrate bitsandbytes support to OOT plugin (#43529)

Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Signed-off-by: Isotr0py <Isotr0py@outlook.com>
This commit is contained in:
Isotr0py
2026-08-09 10:27:31 +08:00
committed by GitHub
parent dedbf6be8b
commit 073c510c91
31 changed files with 187 additions and 1895 deletions
-30
View File
@@ -5,7 +5,6 @@
import inspect
import warnings
from contextlib import contextmanager
from dataclasses import dataclass, field
from typing import Any
import torch
@@ -263,35 +262,6 @@ def get_architecture_class_name(model_config: ModelConfig) -> str:
return get_model_architecture(model_config)[1]
@dataclass
class ParamMapping:
"""
A class to handle parameter mapping for model weight loading.
It creates a bidirectional mapping between packed parameters and their
constituent parts.
"""
packed_mapping: dict[str, list[str]]
inverse_packed_mapping: dict[str, tuple[str, int]] = field(default_factory=dict)
def __post_init__(self):
for packed_name, sub_params in self.packed_mapping.items():
# Skip self-contained cases (e.g., {"W_pack": ["W_pack"]})
if len(sub_params) == 1 and sub_params[0] == packed_name:
continue
for index, param_name in enumerate(sub_params):
self.inverse_packed_mapping[param_name] = (
packed_name,
index,
)
def get_sub_modules(self, module_name: str) -> tuple[str, list[str]] | None:
for key, value in self.packed_mapping.items():
if module_name.endswith(key):
return key, value
return None
def configure_quant_config(
quant_config: QuantizationConfig, model_class: type[nn.Module]
):