Skip to content

Commit da87d6c

Browse files
authored
Enable pytorch fused op on XPU (#1660)
* add torch fuse op Signed-off-by: jiqing-feng <jiqing.feng@intel.com> * fix typo Signed-off-by: jiqing-feng <jiqing.feng@intel.com> * fix dtype Signed-off-by: jiqing-feng <jiqing.feng@intel.com> * fix weight shape Signed-off-by: jiqing-feng <jiqing.feng@intel.com> * fix weight shape Signed-off-by: jiqing-feng <jiqing.feng@intel.com> * fix wtypo Signed-off-by: jiqing-feng <jiqing.feng@intel.com> * fix scales Signed-off-by: jiqing-feng <jiqing.feng@intel.com> * fix weight Signed-off-by: jiqing-feng <jiqing.feng@intel.com> * fix pack Signed-off-by: jiqing-feng <jiqing.feng@intel.com> * fix out shapoe Signed-off-by: jiqing-feng <jiqing.feng@intel.com> * fix format Signed-off-by: jiqing-feng <jiqing.feng@intel.com> * check torch version Signed-off-by: jiqing-feng <jiqing.feng@intel.com> --------- Signed-off-by: jiqing-feng <jiqing.feng@intel.com>
1 parent 037c5c0 commit da87d6c

File tree

3 files changed

+232
-2
lines changed

3 files changed

+232
-2
lines changed
Lines changed: 227 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,227 @@
1+
# Copyright 2024-2025 ModelCloud.ai
2+
# Copyright 2024-2025 qubitium@modelcloud.ai
3+
# Contact: qubitium@modelcloud.ai, x.com/qubitium
4+
#
5+
# Licensed under the Apache License, Version 2.0 (the "License");
6+
# you may not use this file except in compliance with the License.
7+
# You may obtain a copy of the License at
8+
#
9+
# http://www.apache.org/licenses/LICENSE-2.0
10+
#
11+
# Unless required by applicable law or agreed to in writing, software
12+
# distributed under the License is distributed on an "AS IS" BASIS,
13+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
14+
# See the License for the specific language governing permissions and
15+
# limitations under the License.
16+
17+
18+
import torch
19+
import torch.nn as nn
20+
from packaging import version
21+
from torch import __version__ as torch_version
22+
from transformers import PreTrainedModel
23+
24+
from ...adapter.adapter import Adapter, Lora
25+
from ...models._const import DEVICE, PLATFORM
26+
from ...nn_modules.qlinear import BaseQuantLinear, PackableQuantLinear
27+
from ...utils.backend import BACKEND
28+
from ...utils.logger import setup_logger
29+
30+
log = setup_logger()
31+
32+
class TorchFusedQuantLinear(PackableQuantLinear):
33+
SUPPORTS_BITS = [4]
34+
SUPPORTS_GROUP_SIZE = [-1, 16, 32, 64, 128]
35+
SUPPORTS_DESC_ACT = [True, False]
36+
SUPPORTS_SYM = [True, False]
37+
SUPPORTS_SHARDS = True
38+
SUPPORTS_TRAINING = True
39+
SUPPORTS_AUTO_PADDING = True
40+
SUPPORTS_IN_FEATURES_DIVISIBLE_BY = [1]
41+
SUPPORTS_OUT_FEATURES_DIVISIBLE_BY = [1]
42+
43+
SUPPORTS_DEVICES = [DEVICE.XPU]
44+
SUPPORTS_PLATFORM = [PLATFORM.ALL]
45+
SUPPORTS_PACK_DTYPES = [torch.int32]
46+
SUPPORTS_ADAPTERS = [Lora]
47+
48+
SUPPORTS_DTYPES = [torch.float16, torch.bfloat16]
49+
50+
# for transformers/optimum tests compat
51+
QUANT_TYPE = "torch"
52+
53+
def __init__(
54+
self,
55+
bits: int,
56+
group_size: int,
57+
sym: bool,
58+
desc_act: bool,
59+
in_features: int,
60+
out_features: int,
61+
bias: bool = False,
62+
pack_dtype: torch.dtype = torch.int32,
63+
adapter: Adapter = None,
64+
register_buffers: bool = True,
65+
**kwargs,
66+
):
67+
super().__init__(
68+
bits=bits,
69+
group_size=group_size,
70+
sym=sym,
71+
desc_act=desc_act,
72+
in_features=in_features,
73+
out_features=out_features,
74+
bias=bias,
75+
pack_dtype=pack_dtype,
76+
backend=kwargs.pop("backend", BACKEND.TORCH),
77+
adapter=adapter,
78+
register_buffers=register_buffers,
79+
**kwargs)
80+
81+
self.transformed = False
82+
self.dequant_dtype = torch.int16 if self.bits == 8 else torch.int8
83+
84+
def post_init(self):
85+
super().post_init()
86+
self.optimize()
87+
88+
def optimize(self):
89+
if self.optimized:
90+
return
91+
92+
super().optimize()
93+
94+
def train(self, mode: bool = True):
95+
old_train = self.training
96+
if mode == old_train:
97+
return self
98+
99+
from ...utils.model import convert_gptq_v1_to_v2_format_module
100+
101+
if self.SUPPORTS_TRAINING_USE_TORCH_KERNEL:
102+
# training starts
103+
if mode:
104+
# one time clone v1 qzeros and save both v1 and v2 qzeros in memory
105+
if self.qzero_format() == 1:
106+
if not hasattr(self, "qzeros_data_v1"):
107+
self.qzeros_data_v1 = self.qzeros.data.clone()
108+
convert_gptq_v1_to_v2_format_module(self, bits=self.bits, pack_dtype=self.pack_dtype)
109+
self.qzeros_data_v2 = self.qzeros.data
110+
else:
111+
self.qzeros.data = self.qzeros_data_v2
112+
self.qzero_format(format=2)
113+
114+
# training switching to inference/eval
115+
else:
116+
if hasattr(self, "qzeros_data_v1"):
117+
# switch qzero back to v1 for inference/eval
118+
self.qzeros.data = self.qzeros_data_v1
119+
self.qzero_format(format=1)
120+
121+
return super().train(mode=mode)
122+
123+
def transform(self, dtype):
124+
self.scales = self.scales.clone().to(dtype).contiguous()
125+
# Unpack qzeros
126+
zeros = torch.bitwise_right_shift(
127+
torch.unsqueeze(self.qzeros, 2).expand(-1, -1, self.pack_factor),
128+
self.wf_unsqueeze_zero # self.wf.unsqueeze(0),
129+
).to(self.dequant_dtype)
130+
zeros = torch.bitwise_and(zeros, self.maxq).reshape(zeros.shape[0], -1)
131+
# Unpack and reorder qweight
132+
weight = torch.bitwise_and(
133+
torch.bitwise_right_shift(
134+
torch.unsqueeze(self.qweight, 1).expand(-1, self.pack_factor, -1),
135+
self.wf_unsqueeze_neg_one # self.wf.unsqueeze(-1)
136+
).to(self.dequant_dtype),
137+
self.maxq
138+
)
139+
self.ret_idx = torch.zeros(self.g_idx.shape[0], dtype=torch.int32).to(self.g_idx.device)
140+
groups = self.g_idx.shape[0] // self.group_size
141+
remainder = self.g_idx.shape[0] % self.group_size
142+
g_idx_2 = self.g_idx * self.group_size
143+
if remainder > 0:
144+
g_idx_2[self.g_idx == groups] += torch.arange(remainder).to(self.g_idx_2.device).to(self.g_idx_2.dtype)
145+
arange_tensor = torch.arange(self.group_size).to(self.g_idx.device).to(self.g_idx.dtype)
146+
for i in range(groups):
147+
g_idx_2[self.g_idx == i] += arange_tensor
148+
self.ret_idx[g_idx_2] = torch.arange(self.g_idx.shape[0]).to(self.ret_idx.device).to(self.ret_idx.dtype)
149+
weight = weight.reshape(weight.shape[0] * weight.shape[1], weight.shape[2]).index_select(0, self.ret_idx).t()
150+
# Pack qweight
151+
packed = torch.zeros(weight.shape[0], weight.shape[1] // self.pack_factor, dtype=torch.int32, device=weight.device)
152+
for col in range(weight.shape[1] // self.pack_factor):
153+
for i in range(self.pack_factor):
154+
packed_col = weight[:, col * self.pack_factor + i].to(torch.int32)
155+
packed[:, col] |= packed_col << (i * self.bits)
156+
157+
self.qweight = packed.contiguous()
158+
self.qzeros = zeros.contiguous()
159+
160+
def forward(self, x: torch.Tensor):
161+
out_shape = x.shape[:-1] + (self.out_features,)
162+
x = x.reshape(-1, x.shape[-1])
163+
out = self._forward(x, out_shape)
164+
return out
165+
166+
def _forward(self, x, out_shape):
167+
num_itr = self.g_idx.shape[0] // x.shape[-1]
168+
169+
if not self.training and not self.transformed and version.parse(torch_version).release >= version.parse("2.8").release:
170+
self.transform(x.dtype)
171+
self.transformed = True
172+
173+
if not self.transformed:
174+
# make sure dequant dtype matches input x
175+
weights = self.dequantize_weight(num_itr=num_itr).to(x.dtype)
176+
out = torch.matmul(x, weights).reshape(out_shape)
177+
else:
178+
x = x[:, self.ret_idx].contiguous()
179+
out = torch.ops.aten._weight_int4pack_mm_with_scales_and_zeros(
180+
x, self.qweight, self.group_size, self.scales, self.qzeros
181+
).reshape(out_shape)
182+
183+
if self.bias is not None:
184+
out.add_(self.bias)
185+
186+
if self.adapter:
187+
out = self.adapter.apply(x=x, out=out)
188+
189+
return out
190+
191+
# clear gptq only weights: useful in de-quantization
192+
def _empty_gptq_only_weights(self):
193+
self.qzeros = None
194+
self.qweight = None
195+
self.g_idx = None
196+
self.scales = None
197+
198+
199+
def dequantize_model(model: PreTrainedModel):
200+
for name, module in model.named_modules():
201+
if isinstance(module, BaseQuantLinear) and not isinstance(module, TorchFusedQuantLinear):
202+
raise ValueError(
203+
"Only models loaded using TorchFusedQuantLinear are supported for dequantization. "
204+
"Please load model using backend=BACKEND.TORCH."
205+
)
206+
207+
if isinstance(module, TorchFusedQuantLinear):
208+
# Create a new Linear layer with dequantized weights
209+
new_module = nn.Linear(module.in_features, module.out_features)
210+
new_module.weight = nn.Parameter(module.dequantize_weight().T.detach().to("cpu", torch.float16))
211+
new_module.bias = torch.nn.Parameter(module.bias)
212+
213+
# Replace the module in the model
214+
parent = model
215+
if '.' in name:
216+
parent_name, module_name = name.rsplit('.', 1)
217+
parent = dict(model.named_modules())[parent_name]
218+
else:
219+
module_name = name
220+
221+
setattr(parent, module_name, new_module)
222+
223+
del model.config.quantization_config
224+
return model
225+
226+
227+
__all__ = ["TorchFusedQuantLinear", "dequantize_model"]

gptqmodel/utils/backend.py

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -22,6 +22,7 @@ class BACKEND(str, Enum):
2222
AUTO_TRAINABLE = "auto_trainable" # choose the optimal trainable local kernel for post-quant training
2323

2424
# gptq
25+
TORCH_FUSED = "torch_fused" # optimized for Intel XPU
2526
TORCH = "torch" # GOOD: about 80% of triton
2627
TRITON = "triton" # VERY GOOD: all-around kernel
2728
EXLLAMA_V1 = "exllama_v1" # FAST: optimized for batching == 1

gptqmodel/utils/importer.py

Lines changed: 4 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -31,6 +31,7 @@
3131
from ..nn_modules.qlinear.marlin import MarlinQuantLinear
3232
from ..nn_modules.qlinear.qqq import QQQQuantLinear
3333
from ..nn_modules.qlinear.torch import TorchQuantLinear
34+
from ..nn_modules.qlinear.torch_fused import TorchFusedQuantLinear
3435
from ..nn_modules.qlinear.tritonv2 import TRITON_AVAILABLE, TRITON_INSTALL_HINT, TritonV2QuantLinear
3536
from ..quantization import FORMAT
3637
from ..utils.logger import setup_logger
@@ -46,6 +47,7 @@
4647
# BACKEND.EXLLAMA_EORA: ExllamaEoraQuantLinear, #
4748
BACKEND.EXLLAMA_V2: ExllamaV2QuantLinear, # optimized for bs > 1
4849
BACKEND.EXLLAMA_V1: ExllamaQuantLinear, # optimized for bs == 1
50+
BACKEND.TORCH_FUSED: TorchFusedQuantLinear, # optimized for Intel XPU
4951
BACKEND.TRITON: TritonV2QuantLinear, # good all around kernel that JIT compiles
5052
# BACKEND.CUDA: DynamicCudaQuantLinear,
5153
BACKEND.IPEX: IPEXQuantLinear, # best kernel Intel XPU and CPU with amx/avx512/xmx
@@ -56,8 +58,8 @@
5658
})
5759

5860
FORMAT_DICT = {
59-
FORMAT.GPTQ: [BACKEND.MARLIN, BACKEND.EXLLAMA_V2, BACKEND.EXLLAMA_V1, BACKEND.TRITON, BACKEND.IPEX, BACKEND.TORCH, BACKEND.MARLIN_FP16, BACKEND.EXLLAMA_EORA],
60-
FORMAT.GPTQ_V2: [BACKEND.EXLLAMA_V2, BACKEND.EXLLAMA_V1, BACKEND.TRITON, BACKEND.TORCH],
61+
FORMAT.GPTQ: [BACKEND.MARLIN, BACKEND.EXLLAMA_V2, BACKEND.EXLLAMA_V1, BACKEND.TORCH_FUSED, BACKEND.TRITON, BACKEND.IPEX, BACKEND.TORCH, BACKEND.MARLIN_FP16, BACKEND.EXLLAMA_EORA],
62+
FORMAT.GPTQ_V2: [BACKEND.EXLLAMA_V2, BACKEND.EXLLAMA_V1, BACKEND.TORCH_FUSED, BACKEND.TRITON, BACKEND.TORCH],
6163
FORMAT.MARLIN: [BACKEND.MARLIN, BACKEND.MARLIN_FP16],
6264
FORMAT.BITBLAS: [BACKEND.BITBLAS],
6365
FORMAT.IPEX: [BACKEND.IPEX],

0 commit comments

Comments
 (0)