Public/automatic
1
0
Fork 0
mirror of https://github.com/vladmandic/automatic synced 2026-09-19 01:04:32 +02:00
Code Issues Packages Projects Releases Wiki Activity

Add get_modules_to_not_convert from transformers v5

Browse Source
This commit is contained in:
Disty0
2025-12-02 01:01:51 +03:00
parent d9bc31e7da
commit 7aa1bfdc70
2 changed files with 14 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/sdnq/common.py
+2 -2
View File
@@ -5,7 +5,7 @@ import torch
from modules import shared, devices
sdnq_version = "0.1.1"
sdnq_version = "0.1.2"
dtype_dict = {
"int32": {"min": -2147483648, "max": 2147483647, "num_bits": 32, "sign": 1, "exponent": 0, "mantissa": 31, "target_dtype": torch.int32, "torch_dtype": torch.int32, "storage_dtype": torch.int32, "is_unsigned": False, "is_integer": True, "is_packed": False},
@@ -17,7 +17,7 @@ dtype_dict = {
"int4": {"min": -8, "max": 7, "num_bits": 4, "sign": 1, "exponent": 0, "mantissa": 3, "target_dtype": "int4", "torch_dtype": torch.int8, "storage_dtype": torch.uint8, "is_unsigned": False, "is_integer": True, "is_packed": True},
"int3": {"min": -4, "max": 3, "num_bits": 3, "sign": 1, "exponent": 0, "mantissa": 2, "target_dtype": "int3", "torch_dtype": torch.int8, "storage_dtype": torch.uint8, "is_unsigned": False, "is_integer": True, "is_packed": True},
"int2": {"min": -2, "max": 1, "num_bits": 2, "sign": 1, "exponent": 0, "mantissa": 1, "target_dtype": "int2", "torch_dtype": torch.int8, "storage_dtype": torch.uint8, "is_unsigned": False, "is_integer": True, "is_packed": True},
"uint32": {"min": 0, "max": 4294967295, "num_bits": 32, "sign": 0, "exponent": 0, "mantissa": 31, "target_dtype": torch.uint32, "torch_dtype": torch.uint32, "storage_dtype": torch.uint32, "is_unsigned": True, "is_integer": True, "is_packed": False},
"uint32": {"min": 0, "max": 4294967295, "num_bits": 32, "sign": 0, "exponent": 0, "mantissa": 32, "target_dtype": torch.uint32, "torch_dtype": torch.uint32, "storage_dtype": torch.uint32, "is_unsigned": True, "is_integer": True, "is_packed": False},
"uint16": {"min": 0, "max": 65535, "num_bits": 16, "sign": 0, "exponent": 0, "mantissa": 16, "target_dtype": torch.uint16, "torch_dtype": torch.uint16, "storage_dtype": torch.uint16, "is_unsigned": True, "is_integer": True, "is_packed": False},
"uint8": {"min": 0, "max": 255, "num_bits": 8, "sign": 0, "exponent": 0, "mantissa": 8, "target_dtype": torch.uint8, "torch_dtype": torch.uint8, "storage_dtype": torch.uint8, "is_unsigned": True, "is_integer": True, "is_packed": False},
"uint7": {"min": 0, "max": 127, "num_bits": 7, "sign": 0, "exponent": 0, "mantissa": 7, "target_dtype": "uint7", "torch_dtype": torch.uint8, "storage_dtype": torch.uint8, "is_unsigned": True, "is_integer": True, "is_packed": True},
modules/sdnq/quantizer.py
+12
View File
@@ -727,6 +727,12 @@ class SDNQQuantizer(DiffusersQuantizer, HfQuantizer):
self.torch_dtype = torch_dtype
return torch_dtype
def update_dtype(self, dtype: torch.dtype = None) -> torch.dtype:
"""
needed for transformers compatibilty, returns self.update_torch_dtype
"""
return self.update_torch_dtype(dtype)
def _process_model_before_weight_loading( # pylint: disable=arguments-differ
self,
model,
@@ -746,6 +752,8 @@ class SDNQQuantizer(DiffusersQuantizer, HfQuantizer):
if self.quantization_config.add_skip_keys:
if keep_in_fp32_modules is not None:
self.quantization_config.modules_to_not_convert.extend(keep_in_fp32_modules)
if hasattr(self, "get_modules_to_not_convert") and hasattr(model, "tie_weights"):
self.quantization_config.modules_to_not_convert.extend(self.get_modules_to_not_convert(model, add_default_skips=True))
model, self.quantization_config.modules_to_not_convert, self.quantization_config.modules_dtype_dict = add_module_skip_keys(
model, self.quantization_config.modules_to_not_convert, self.quantization_config.modules_dtype_dict
)
@@ -799,6 +807,10 @@ class SDNQQuantizer(DiffusersQuantizer, HfQuantizer):
def is_trainable(self):
return self.quantization_config.is_training
@property
def is_qat_trainable(self) -> bool:
return self.is_trainable()
@property
def is_compileable(self):
return True