forked from learninmou/TensorRT-LLM
-
Notifications
You must be signed in to change notification settings - Fork 0
/
smoothquant.py
154 lines (125 loc) · 5.12 KB
/
smoothquant.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
# SPDX-FileCopyrightText: Copyright (c) 2022-2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: Apache-2.0
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
'''
Utilities for SmoothQuant models
'''
import functools
from collections import defaultdict
import torch
import torch.nn as nn
from tqdm import tqdm
from transformers.pytorch_utils import Conv1D
@torch.no_grad()
def apply_smoothing(scales,
gemm_weights,
layernorm_weights=None,
layernorm_bias=None,
dtype=torch.float32,
layernorm_1p=False):
if not isinstance(gemm_weights, list):
gemm_weights = [gemm_weights]
if layernorm_weights is not None:
assert layernorm_weights.numel() == scales.numel()
layernorm_weights.div_(scales).to(dtype)
if layernorm_bias is not None:
assert layernorm_bias.numel() == scales.numel()
layernorm_bias.div_(scales).to(dtype)
if layernorm_1p:
layernorm_weights += (1 / scales) - 1
for gemm in gemm_weights:
gemm.mul_(scales.view(1, -1)).to(dtype)
@torch.no_grad()
def smooth_gemm(gemm_weights,
act_scales,
layernorm_weights=None,
layernorm_bias=None,
alpha=0.5,
weight_scales=None):
if not isinstance(gemm_weights, list):
gemm_weights = [gemm_weights]
orig_dtype = gemm_weights[0].dtype
for gemm in gemm_weights:
# gemm_weights are expected to be transposed
assert gemm.shape[1] == act_scales.numel()
if weight_scales is None:
weight_scales = torch.cat(
[gemm.abs().max(dim=0, keepdim=True)[0] for gemm in gemm_weights],
dim=0)
weight_scales = weight_scales.max(dim=0)[0]
weight_scales.to(float).clamp(min=1e-5)
scales = (act_scales.to(gemm_weights[0].device).to(float).pow(alpha) /
weight_scales.pow(1 - alpha)).clamp(min=1e-5)
apply_smoothing(scales, gemm_weights, layernorm_weights, layernorm_bias,
orig_dtype)
return scales
@torch.no_grad()
def smooth_ln_fcs(ln, fcs, act_scales, alpha=0.5):
if not isinstance(fcs, list):
fcs = [fcs]
for fc in fcs:
assert isinstance(fc, nn.Linear)
assert ln.weight.numel() == fc.in_features == act_scales.numel()
device, dtype = fcs[0].weight.device, fcs[0].weight.dtype
act_scales = act_scales.to(device=device, dtype=dtype)
weight_scales = torch.cat(
[fc.weight.abs().max(dim=0, keepdim=True)[0] for fc in fcs], dim=0)
weight_scales = weight_scales.max(dim=0)[0].clamp(min=1e-5)
scales = (act_scales.pow(alpha) /
weight_scales.pow(1 - alpha)).clamp(min=1e-5).to(device).to(dtype)
if ln is not None:
ln.weight.div_(scales)
ln.bias.div_(scales)
for fc in fcs:
fc.weight.mul_(scales.view(1, -1))
return scales
@torch.no_grad()
def capture_activation_range(model, tokenizer, num_samples=512, seq_len=512):
model.eval()
device = next(model.parameters()).device
act_scales = defaultdict(lambda: {"x": None, "y": None, "w": None})
def stat_tensor(name, tensor, act_scales, key):
hidden_dim = tensor.shape[-1]
tensor = tensor.view(-1, hidden_dim).abs().detach()
comming_max = torch.max(tensor, dim=0)[0].float()
if act_scales[name][key] is None:
act_scales[name][key] = comming_max
else:
act_scales[name][key] = torch.max(act_scales[name][key],
comming_max)
def stat_input_hook(m, x, y, name):
if isinstance(x, tuple):
x = x[0]
stat_tensor(name, x, act_scales, "x")
stat_tensor(name, y, act_scales, "y")
if act_scales[name]["w"] is None:
act_scales[name]["w"] = m.weight.abs().clip(1e-8,
None).max(dim=0)[0]
hooks = []
for name, m in model.named_modules():
if isinstance(m, nn.Linear) or isinstance(m, Conv1D):
hooks.append(
m.register_forward_hook(
functools.partial(stat_input_hook, name=name)))
from datasets import load_dataset
dataset = load_dataset("lambada", split="validation")
for i in tqdm(range(num_samples), desc="calibrating model"):
input_ids = tokenizer(dataset[i]["text"],
return_tensors="pt",
max_length=seq_len,
truncation=True).input_ids.to(device)
model(input_ids)
for h in hooks:
h.remove()
return act_scales