Commit
•
80b34f4
1
Parent(s):
ccc080b
Upload 15 files
Browse files- config.json +52 -0
- configuration_transnormer.py +71 -0
- generation_config copy.json +9 -0
- generation_config.json +6 -0
- lightning_attention.py +540 -0
- lightning_attention2.py +540 -0
- model.safetensors +3 -0
- modeling_transnormer.py +933 -0
- norm.py +43 -0
- special_tokens_map copy.json +23 -0
- special_tokens_map.json +23 -0
- srmsnorm_triton.py +201 -0
- tokenizer_config.json +33 -0
- utils.py +151 -0
- vocab.json +0 -0
config.json
ADDED
@@ -0,0 +1,52 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"architectures": [
|
3 |
+
"tnl1-385m-10b-token_no-act"
|
4 |
+
],
|
5 |
+
"auto_map": {
|
6 |
+
"AutoConfig": "configuration_transnormer.TransnormerConfig",
|
7 |
+
"AutoModelForCausalLM": "modeling_transnormer.TransnormerForCausalLM"
|
8 |
+
},
|
9 |
+
"bos_token_id": 50260,
|
10 |
+
"eos_token_id": 50260,
|
11 |
+
"vocab_size": 50272,
|
12 |
+
"use_cache": true,
|
13 |
+
"init_std": 0.02,
|
14 |
+
"decoder_embed_dim": 1024,
|
15 |
+
"decoder_layers": 24,
|
16 |
+
"decoder_attention_heads": 8,
|
17 |
+
"no_scale_embedding": false,
|
18 |
+
"add_bos_token": false,
|
19 |
+
"norm_type": "simplermsnorm",
|
20 |
+
"linear_use_lrpe_list": [
|
21 |
+
1,
|
22 |
+
0,
|
23 |
+
0,
|
24 |
+
0,
|
25 |
+
0,
|
26 |
+
0,
|
27 |
+
0,
|
28 |
+
0,
|
29 |
+
0,
|
30 |
+
0,
|
31 |
+
0,
|
32 |
+
0,
|
33 |
+
0,
|
34 |
+
0,
|
35 |
+
0,
|
36 |
+
0,
|
37 |
+
0,
|
38 |
+
0,
|
39 |
+
0,
|
40 |
+
0,
|
41 |
+
0,
|
42 |
+
0,
|
43 |
+
0,
|
44 |
+
0
|
45 |
+
],
|
46 |
+
"hidden_dim": 1024,
|
47 |
+
"linear_act_fun": "none",
|
48 |
+
"glu_dim": 2816,
|
49 |
+
"bias": false,
|
50 |
+
"torch_dtype": "bfloat16",
|
51 |
+
"transformers_version": "4.38.2"
|
52 |
+
}
|
configuration_transnormer.py
ADDED
@@ -0,0 +1,71 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright 2023 OpenNLPLab
|
2 |
+
#
|
3 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
4 |
+
# you may not use this file except in compliance with the License.
|
5 |
+
# You may obtain a copy of the License at
|
6 |
+
#
|
7 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
8 |
+
#
|
9 |
+
# Unless required by applicable law or agreed to in writing, software
|
10 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
11 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
12 |
+
# See the License for the specific language governing permissions and
|
13 |
+
# limitations under the License.
|
14 |
+
|
15 |
+
# coding=utf-8
|
16 |
+
""" Transnormer configuration"""
|
17 |
+
|
18 |
+
from transformers.configuration_utils import PretrainedConfig
|
19 |
+
from transformers.utils import logging
|
20 |
+
|
21 |
+
logger = logging.get_logger(__name__)
|
22 |
+
|
23 |
+
|
24 |
+
class TransnormerConfig(PretrainedConfig):
|
25 |
+
model_type = "transnormer"
|
26 |
+
keys_to_ignore_at_inference = ["past_key_values"]
|
27 |
+
|
28 |
+
def __init__(
|
29 |
+
self,
|
30 |
+
pad_token_id=0,
|
31 |
+
bos_token_id=1,
|
32 |
+
eos_token_id=2,
|
33 |
+
vocab_size=64000,
|
34 |
+
use_cache=True,
|
35 |
+
init_std=0.02,
|
36 |
+
# model config
|
37 |
+
decoder_embed_dim=1024,
|
38 |
+
decoder_layers=24,
|
39 |
+
decoder_attention_heads=8,
|
40 |
+
no_scale_embedding=False,
|
41 |
+
add_bos_token=False,
|
42 |
+
norm_type="simplermsnorm",
|
43 |
+
linear_use_lrpe_list=[],
|
44 |
+
hidden_dim=1024,
|
45 |
+
linear_act_fun="silu",
|
46 |
+
glu_dim=2816,
|
47 |
+
bias=False,
|
48 |
+
**kwargs,
|
49 |
+
):
|
50 |
+
super().__init__(
|
51 |
+
pad_token_id=pad_token_id,
|
52 |
+
bos_token_id=bos_token_id,
|
53 |
+
eos_token_id=eos_token_id,
|
54 |
+
**kwargs,
|
55 |
+
)
|
56 |
+
# hf origin
|
57 |
+
self.vocab_size = vocab_size
|
58 |
+
self.use_cache = use_cache
|
59 |
+
self.init_std = init_std
|
60 |
+
# add
|
61 |
+
self.decoder_embed_dim = decoder_embed_dim
|
62 |
+
self.decoder_layers = decoder_layers
|
63 |
+
self.decoder_attention_heads = decoder_attention_heads
|
64 |
+
self.no_scale_embedding = no_scale_embedding
|
65 |
+
self.add_bos_token = add_bos_token
|
66 |
+
self.norm_type = norm_type
|
67 |
+
self.linear_use_lrpe_list = linear_use_lrpe_list
|
68 |
+
self.hidden_dim = hidden_dim
|
69 |
+
self.linear_act_fun = linear_act_fun
|
70 |
+
self.glu_dim = glu_dim
|
71 |
+
self.bias = bias
|
generation_config copy.json
ADDED
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"pad_token_id": 0,
|
3 |
+
"bos_token_id": 1,
|
4 |
+
"eos_token_id": 2,
|
5 |
+
"max_new_tokens": 2048,
|
6 |
+
"temperature": 1.0,
|
7 |
+
"repetition_penalty": 1.03,
|
8 |
+
"do_sample": true
|
9 |
+
}
|
generation_config.json
ADDED
@@ -0,0 +1,6 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"_from_model_config": true,
|
3 |
+
"bos_token_id": 50260,
|
4 |
+
"eos_token_id": 50260,
|
5 |
+
"transformers_version": "4.38.2"
|
6 |
+
}
|
lightning_attention.py
ADDED
@@ -0,0 +1,540 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright 2023 OpenNLPLab
|
2 |
+
#
|
3 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
4 |
+
# you may not use this file except in compliance with the License.
|
5 |
+
# You may obtain a copy of the License at
|
6 |
+
#
|
7 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
8 |
+
#
|
9 |
+
# Unless required by applicable law or agreed to in writing, software
|
10 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
11 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
12 |
+
# See the License for the specific language governing permissions and
|
13 |
+
# limitations under the License.
|
14 |
+
|
15 |
+
import torch
|
16 |
+
import triton
|
17 |
+
import triton.language as tl
|
18 |
+
|
19 |
+
|
20 |
+
@triton.jit
|
21 |
+
def _fwd_kernel(
|
22 |
+
Q,
|
23 |
+
K,
|
24 |
+
V,
|
25 |
+
Out,
|
26 |
+
S,
|
27 |
+
stride_qz,
|
28 |
+
stride_qh,
|
29 |
+
stride_qm,
|
30 |
+
stride_qk,
|
31 |
+
stride_kz,
|
32 |
+
stride_kh,
|
33 |
+
stride_kn,
|
34 |
+
stride_kk,
|
35 |
+
stride_vz,
|
36 |
+
stride_vh,
|
37 |
+
stride_vn,
|
38 |
+
stride_ve,
|
39 |
+
stride_oz,
|
40 |
+
stride_oh,
|
41 |
+
stride_om,
|
42 |
+
stride_oe,
|
43 |
+
stride_sh,
|
44 |
+
Z,
|
45 |
+
H,
|
46 |
+
N_CTX,
|
47 |
+
BLOCK_M: tl.constexpr,
|
48 |
+
BLOCK_DMODEL_QK: tl.constexpr,
|
49 |
+
BLOCK_N: tl.constexpr,
|
50 |
+
BLOCK_DMODEL_V: tl.constexpr,
|
51 |
+
IS_CAUSAL: tl.constexpr,
|
52 |
+
USE_DECAY: tl.constexpr,
|
53 |
+
):
|
54 |
+
start_m = tl.program_id(0)
|
55 |
+
off_hz = tl.program_id(1)
|
56 |
+
off_h = off_hz % H
|
57 |
+
# initialize offsets
|
58 |
+
offs_m = start_m * BLOCK_M + tl.arange(0, BLOCK_M)
|
59 |
+
offs_n = tl.arange(0, BLOCK_N)
|
60 |
+
offs_k = tl.arange(0, BLOCK_DMODEL_QK)
|
61 |
+
offs_e = tl.arange(0, BLOCK_DMODEL_V)
|
62 |
+
# get current offset of q k v
|
63 |
+
off_q = (off_hz * stride_qh + offs_m[:, None] * stride_qm +
|
64 |
+
offs_k[None, :] * stride_qk)
|
65 |
+
off_k = (off_hz * stride_kh + offs_n[:, None] * stride_kn +
|
66 |
+
offs_k[None, :] * stride_kk)
|
67 |
+
off_v = (off_hz * stride_vh + offs_n[:, None] * stride_vn +
|
68 |
+
offs_e[None, :] * stride_ve)
|
69 |
+
off_o = (off_hz * stride_oh + offs_m[:, None] * stride_om +
|
70 |
+
offs_e[None, :] * stride_oe)
|
71 |
+
|
72 |
+
# Initialize pointers to Q, K, V
|
73 |
+
q_ptrs = Q + off_q
|
74 |
+
k_ptrs = K + off_k
|
75 |
+
v_ptrs = V + off_v
|
76 |
+
|
77 |
+
# initialize pointer to m and l
|
78 |
+
acc = tl.zeros([BLOCK_M, BLOCK_DMODEL_V], dtype=tl.float32)
|
79 |
+
# load q: it will stay in SRAM throughout
|
80 |
+
q = tl.load(q_ptrs, mask=offs_m[:, None] < N_CTX, other=0.0)
|
81 |
+
# loop over k, v and update accumulator
|
82 |
+
lo = 0
|
83 |
+
# print(start_m)
|
84 |
+
hi = (start_m + 1) * BLOCK_M if IS_CAUSAL else N_CTX
|
85 |
+
for start_n in range(lo, hi, BLOCK_N):
|
86 |
+
# -- load k, v --
|
87 |
+
k = tl.load(
|
88 |
+
k_ptrs + start_n * stride_kn,
|
89 |
+
mask=(start_n + offs_n)[:, None] < N_CTX,
|
90 |
+
other=0.0,
|
91 |
+
)
|
92 |
+
v = tl.load(
|
93 |
+
v_ptrs + start_n * stride_vn,
|
94 |
+
mask=(start_n + offs_n)[:, None] < N_CTX,
|
95 |
+
other=0.0,
|
96 |
+
)
|
97 |
+
# -- compute qk ---
|
98 |
+
# qk = tl.dot(q, k)
|
99 |
+
qk = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32)
|
100 |
+
# qk += tl.dot(q, k, trans_b=True)
|
101 |
+
qk += tl.dot(q, tl.trans(k))
|
102 |
+
if IS_CAUSAL:
|
103 |
+
index = offs_m[:, None] - (start_n + offs_n[None, :])
|
104 |
+
if USE_DECAY:
|
105 |
+
S_block_ptr = S + off_h * stride_sh
|
106 |
+
s = tl.load(S_block_ptr)
|
107 |
+
s_index = s * index
|
108 |
+
s_index = tl.where(s_index >= 0, -s_index, float("-inf"))
|
109 |
+
qk = tl.exp(s_index) * qk
|
110 |
+
else:
|
111 |
+
qk = tl.where(index >= 0, qk, 0)
|
112 |
+
acc += tl.dot(qk, v.to(qk.dtype))
|
113 |
+
|
114 |
+
out_ptrs = Out + off_o
|
115 |
+
tl.store(out_ptrs, acc.to(q.dtype), mask=offs_m[:, None] < N_CTX)
|
116 |
+
|
117 |
+
|
118 |
+
@triton.jit
|
119 |
+
def _bwd_kernel_kv(
|
120 |
+
Q,
|
121 |
+
K,
|
122 |
+
V,
|
123 |
+
S,
|
124 |
+
DO,
|
125 |
+
DQ,
|
126 |
+
DK,
|
127 |
+
DV,
|
128 |
+
stride_qz,
|
129 |
+
stride_qh,
|
130 |
+
stride_qm,
|
131 |
+
stride_qk,
|
132 |
+
stride_kz,
|
133 |
+
stride_kh,
|
134 |
+
stride_kn,
|
135 |
+
stride_kk,
|
136 |
+
stride_vz,
|
137 |
+
stride_vh,
|
138 |
+
stride_vn,
|
139 |
+
stride_ve,
|
140 |
+
stride_oz,
|
141 |
+
stride_oh,
|
142 |
+
stride_om,
|
143 |
+
stride_oe,
|
144 |
+
stride_sh,
|
145 |
+
Z,
|
146 |
+
H,
|
147 |
+
N_CTX,
|
148 |
+
num_block,
|
149 |
+
BLOCK_M: tl.constexpr,
|
150 |
+
BLOCK_DMODEL_QK: tl.constexpr,
|
151 |
+
BLOCK_N: tl.constexpr,
|
152 |
+
BLOCK_DMODEL_V: tl.constexpr,
|
153 |
+
CAUSAL: tl.constexpr,
|
154 |
+
USE_DECAY: tl.constexpr,
|
155 |
+
):
|
156 |
+
start_n = tl.program_id(0)
|
157 |
+
off_hz = tl.program_id(1)
|
158 |
+
|
159 |
+
off_z = off_hz // H
|
160 |
+
off_h = off_hz % H
|
161 |
+
# offset pointers for batch/head
|
162 |
+
Q += off_z * stride_qz + off_h * stride_qh
|
163 |
+
K += off_z * stride_kz + off_h * stride_kh
|
164 |
+
V += off_z * stride_vz + off_h * stride_vh
|
165 |
+
DO += off_z * stride_oz + off_h * stride_oh
|
166 |
+
DQ += off_z * stride_qz + off_h * stride_qh
|
167 |
+
DK += off_z * stride_kz + off_h * stride_kh
|
168 |
+
DV += off_z * stride_vz + off_h * stride_vh
|
169 |
+
|
170 |
+
# start of q
|
171 |
+
if CAUSAL:
|
172 |
+
lo = start_n * BLOCK_M
|
173 |
+
else:
|
174 |
+
lo = 0
|
175 |
+
# initialize row/col offsets
|
176 |
+
# seqlence offset
|
177 |
+
offs_qm = lo + tl.arange(0, BLOCK_M)
|
178 |
+
offs_kvn = start_n * BLOCK_N + tl.arange(0, BLOCK_N)
|
179 |
+
# feature offset
|
180 |
+
offs_qkk = tl.arange(0, BLOCK_DMODEL_QK)
|
181 |
+
offs_ve = tl.arange(0, BLOCK_DMODEL_V)
|
182 |
+
# row block index
|
183 |
+
offs_m = tl.arange(0, BLOCK_M)
|
184 |
+
# initialize pointers to value-like data
|
185 |
+
q_ptrs = Q + (offs_qm[:, None] * stride_qm + offs_qkk[None, :] * stride_qk)
|
186 |
+
k_ptrs = K + (offs_kvn[:, None] * stride_kn +
|
187 |
+
offs_qkk[None, :] * stride_kk)
|
188 |
+
v_ptrs = V + (offs_kvn[:, None] * stride_vn + offs_ve[None, :] * stride_ve)
|
189 |
+
do_ptrs = DO + (offs_qm[:, None] * stride_om +
|
190 |
+
offs_ve[None, :] * stride_oe)
|
191 |
+
dq_ptrs = DQ + (offs_qm[:, None] * stride_qm +
|
192 |
+
offs_qkk[None, :] * stride_qk)
|
193 |
+
# initialize dv amd dk
|
194 |
+
dv = tl.zeros([BLOCK_N, BLOCK_DMODEL_V], dtype=tl.float32)
|
195 |
+
dk = tl.zeros([BLOCK_N, BLOCK_DMODEL_QK], dtype=tl.float32)
|
196 |
+
# k and v stay in SRAM throughout
|
197 |
+
k = tl.load(k_ptrs, mask=offs_kvn[:, None] < N_CTX, other=0.0)
|
198 |
+
v = tl.load(v_ptrs, mask=offs_kvn[:, None] < N_CTX, other=0.0)
|
199 |
+
# loop over rows
|
200 |
+
for start_m in range(lo, num_block * BLOCK_M, BLOCK_M):
|
201 |
+
offs_m_curr = start_m + offs_m
|
202 |
+
# load q, k, v, do on-chip
|
203 |
+
q = tl.load(q_ptrs, mask=offs_m_curr[:, None] < N_CTX, other=0.0)
|
204 |
+
qk = tl.dot(q, tl.trans(k))
|
205 |
+
# qk = tl.dot(q, k, trans_b=True)
|
206 |
+
if CAUSAL:
|
207 |
+
index = offs_m_curr[:, None] - offs_kvn[None, :]
|
208 |
+
if USE_DECAY:
|
209 |
+
S_block_ptr = S + off_h * stride_sh
|
210 |
+
s = tl.load(S_block_ptr)
|
211 |
+
s_index = s * index
|
212 |
+
s_index = tl.where(s_index >= 0, -s_index, float("-inf"))
|
213 |
+
s = tl.exp(s_index)
|
214 |
+
qk = qk * s
|
215 |
+
else:
|
216 |
+
qk = tl.where(index >= 0, qk, 0)
|
217 |
+
|
218 |
+
p = qk
|
219 |
+
# compute dv
|
220 |
+
do = tl.load(do_ptrs, mask=offs_m_curr[:, None] < N_CTX, other=0.0)
|
221 |
+
dv += tl.dot(tl.trans(p.to(do.dtype)), do)
|
222 |
+
dp = tl.dot(do, tl.trans(v).to(do.dtype))
|
223 |
+
if CAUSAL:
|
224 |
+
if USE_DECAY:
|
225 |
+
dp = dp * s
|
226 |
+
else:
|
227 |
+
dp = tl.where(index >= 0, dp, 0)
|
228 |
+
|
229 |
+
dk += tl.dot(tl.trans(dp.to(q.dtype)), q).to(tl.float32)
|
230 |
+
|
231 |
+
# increment pointers
|
232 |
+
q_ptrs += BLOCK_M * stride_qm
|
233 |
+
do_ptrs += BLOCK_M * stride_om
|
234 |
+
# write-back
|
235 |
+
dv_ptrs = DV + (offs_kvn[:, None] * stride_vn +
|
236 |
+
offs_ve[None, :] * stride_ve)
|
237 |
+
dk_ptrs = DK + (offs_kvn[:, None] * stride_kn +
|
238 |
+
offs_qkk[None, :] * stride_kk)
|
239 |
+
tl.store(dv_ptrs, dv, mask=offs_kvn[:, None] < N_CTX)
|
240 |
+
tl.store(dk_ptrs, dk, mask=offs_kvn[:, None] < N_CTX)
|
241 |
+
|
242 |
+
|
243 |
+
@triton.jit
|
244 |
+
def _bwd_kernel_q(
|
245 |
+
Q,
|
246 |
+
K,
|
247 |
+
V,
|
248 |
+
S,
|
249 |
+
DO,
|
250 |
+
DQ,
|
251 |
+
DK,
|
252 |
+
DV,
|
253 |
+
stride_qz,
|
254 |
+
stride_qh,
|
255 |
+
stride_qm,
|
256 |
+
stride_qk,
|
257 |
+
stride_kz,
|
258 |
+
stride_kh,
|
259 |
+
stride_kn,
|
260 |
+
stride_kk,
|
261 |
+
stride_vz,
|
262 |
+
stride_vh,
|
263 |
+
stride_vn,
|
264 |
+
stride_ve,
|
265 |
+
stride_oz,
|
266 |
+
stride_oh,
|
267 |
+
stride_om,
|
268 |
+
stride_oe,
|
269 |
+
stride_sh,
|
270 |
+
Z,
|
271 |
+
H,
|
272 |
+
N_CTX,
|
273 |
+
num_block,
|
274 |
+
BLOCK_M: tl.constexpr,
|
275 |
+
BLOCK_DMODEL_QK: tl.constexpr,
|
276 |
+
BLOCK_N: tl.constexpr,
|
277 |
+
BLOCK_DMODEL_V: tl.constexpr,
|
278 |
+
CAUSAL: tl.constexpr,
|
279 |
+
USE_DECAY: tl.constexpr,
|
280 |
+
):
|
281 |
+
start_m = tl.program_id(0)
|
282 |
+
off_hz = tl.program_id(1)
|
283 |
+
off_z = off_hz // H
|
284 |
+
off_h = off_hz % H
|
285 |
+
# offset pointers for batch/head
|
286 |
+
K += off_z * stride_kz + off_h * stride_kh
|
287 |
+
V += off_z * stride_vz + off_h * stride_vh
|
288 |
+
DO += off_z * stride_oz + off_h * stride_oh
|
289 |
+
DQ += off_z * stride_qz + off_h * stride_qh
|
290 |
+
# feature offset
|
291 |
+
offs_qkk = tl.arange(0, BLOCK_DMODEL_QK)
|
292 |
+
offs_ve = tl.arange(0, BLOCK_DMODEL_V)
|
293 |
+
# row block index
|
294 |
+
offs_m = tl.arange(0, BLOCK_M)
|
295 |
+
# row block index
|
296 |
+
offs_qm = start_m * BLOCK_M + tl.arange(0, BLOCK_M)
|
297 |
+
# do
|
298 |
+
do_ptrs = DO + (offs_qm[:, None] * stride_om +
|
299 |
+
offs_ve[None, :] * stride_oe)
|
300 |
+
dq_ptrs = DQ + (offs_qm[:, None] * stride_qm +
|
301 |
+
offs_qkk[None, :] * stride_qk)
|
302 |
+
|
303 |
+
do = tl.load(do_ptrs, mask=offs_qm[:, None] < N_CTX, other=0.0)
|
304 |
+
|
305 |
+
dq = tl.zeros([BLOCK_M, BLOCK_DMODEL_QK], dtype=tl.float32)
|
306 |
+
lo = 0
|
307 |
+
hi = (start_m + 1) * BLOCK_M if CAUSAL else N_CTX
|
308 |
+
|
309 |
+
offs_m_curr = start_m * BLOCK_M + offs_m
|
310 |
+
|
311 |
+
for start_n in range(0, num_block):
|
312 |
+
offs_kvn = start_n * BLOCK_N + tl.arange(0, BLOCK_N)
|
313 |
+
k_ptrs = K + (offs_kvn[:, None] * stride_kn +
|
314 |
+
offs_qkk[None, :] * stride_kk)
|
315 |
+
v_ptrs = V + (offs_kvn[:, None] * stride_vn +
|
316 |
+
offs_ve[None, :] * stride_ve)
|
317 |
+
# k and v stay in SRAM throughout
|
318 |
+
k = tl.load(k_ptrs, mask=offs_kvn[:, None] < N_CTX, other=0.0)
|
319 |
+
v = tl.load(v_ptrs, mask=offs_kvn[:, None] < N_CTX, other=0.0)
|
320 |
+
# dp = do vT
|
321 |
+
dp = tl.dot(do, tl.trans(v).to(do.dtype))
|
322 |
+
if CAUSAL:
|
323 |
+
index = offs_m_curr[:, None] - offs_kvn[None, :]
|
324 |
+
if USE_DECAY:
|
325 |
+
S_block_ptr = S + off_h * stride_sh
|
326 |
+
s = tl.load(S_block_ptr)
|
327 |
+
s_index = s * index
|
328 |
+
s_index = tl.where(s_index >= 0, -s_index, float("-inf"))
|
329 |
+
s = tl.exp(s_index)
|
330 |
+
dp = dp * s
|
331 |
+
else:
|
332 |
+
dp = tl.where(index >= 0, dp, 0)
|
333 |
+
# dq = dq + dp k
|
334 |
+
dq += tl.dot(dp.to(k.dtype), k)
|
335 |
+
|
336 |
+
tl.store(dq_ptrs, dq, mask=offs_qm[:, None] < N_CTX)
|
337 |
+
|
338 |
+
|
339 |
+
class _attention(torch.autograd.Function):
|
340 |
+
|
341 |
+
@staticmethod
|
342 |
+
def forward(ctx, q, k, v, causal, s):
|
343 |
+
q = q.contiguous()
|
344 |
+
k = k.contiguous()
|
345 |
+
v = v.contiguous()
|
346 |
+
s = s.contiguous()
|
347 |
+
# only support for Ampere now
|
348 |
+
capability = torch.cuda.get_device_capability()
|
349 |
+
if capability[0] < 8:
|
350 |
+
raise RuntimeError(
|
351 |
+
"Flash attention currently only supported for compute capability >= 80"
|
352 |
+
)
|
353 |
+
# shape constraints
|
354 |
+
Lq, Lk, Lv = q.shape[-1], k.shape[-1], v.shape[-1]
|
355 |
+
# right
|
356 |
+
o = torch.empty(
|
357 |
+
(q.shape[0], q.shape[1], q.shape[2], v.shape[-1]),
|
358 |
+
dtype=q.dtype,
|
359 |
+
device=q.device,
|
360 |
+
)
|
361 |
+
|
362 |
+
BLOCK_M = 128
|
363 |
+
BLOCK_N = 64
|
364 |
+
num_warps = 4 if Lk <= 64 else 8
|
365 |
+
num_stages = 1
|
366 |
+
|
367 |
+
grid = (triton.cdiv(q.shape[2], BLOCK_M), q.shape[0] * q.shape[1], 1)
|
368 |
+
use_decay = s.shape[0] > 0
|
369 |
+
|
370 |
+
_fwd_kernel[grid](
|
371 |
+
q,
|
372 |
+
k,
|
373 |
+
v,
|
374 |
+
o,
|
375 |
+
s,
|
376 |
+
q.stride(0),
|
377 |
+
q.stride(1),
|
378 |
+
q.stride(2),
|
379 |
+
q.stride(3),
|
380 |
+
k.stride(0),
|
381 |
+
k.stride(1),
|
382 |
+
k.stride(2),
|
383 |
+
k.stride(3),
|
384 |
+
v.stride(0),
|
385 |
+
v.stride(1),
|
386 |
+
v.stride(2),
|
387 |
+
v.stride(3),
|
388 |
+
o.stride(0),
|
389 |
+
o.stride(1),
|
390 |
+
o.stride(2),
|
391 |
+
o.stride(3),
|
392 |
+
s.stride(0),
|
393 |
+
q.shape[0],
|
394 |
+
q.shape[1],
|
395 |
+
q.shape[2],
|
396 |
+
BLOCK_M=BLOCK_M,
|
397 |
+
BLOCK_DMODEL_QK=Lk,
|
398 |
+
BLOCK_N=BLOCK_N,
|
399 |
+
BLOCK_DMODEL_V=Lv,
|
400 |
+
IS_CAUSAL=causal,
|
401 |
+
USE_DECAY=use_decay,
|
402 |
+
num_warps=num_warps,
|
403 |
+
num_stages=num_stages,
|
404 |
+
)
|
405 |
+
|
406 |
+
ctx.save_for_backward(q, k, v, s)
|
407 |
+
ctx.grid = grid
|
408 |
+
ctx.BLOCK_M = BLOCK_M
|
409 |
+
ctx.BLOCK_DMODEL_QK = Lk
|
410 |
+
ctx.BLOCK_N = BLOCK_N
|
411 |
+
ctx.BLOCK_DMODEL_V = Lv
|
412 |
+
ctx.causal = causal
|
413 |
+
ctx.use_decay = use_decay
|
414 |
+
return o
|
415 |
+
|
416 |
+
@staticmethod
|
417 |
+
def backward(ctx, do):
|
418 |
+
q, k, v, s = ctx.saved_tensors
|
419 |
+
BLOCK_M = 32
|
420 |
+
BLOCK_N = 32
|
421 |
+
num_warps = 4
|
422 |
+
num_stages = 1
|
423 |
+
|
424 |
+
do = do.contiguous()
|
425 |
+
dq = torch.zeros_like(q, dtype=torch.float32)
|
426 |
+
dk = torch.empty_like(k)
|
427 |
+
dv = torch.empty_like(v)
|
428 |
+
|
429 |
+
grid_kv = (triton.cdiv(k.shape[2],
|
430 |
+
BLOCK_N), k.shape[0] * k.shape[1], 1)
|
431 |
+
_bwd_kernel_kv[grid_kv](
|
432 |
+
q,
|
433 |
+
k,
|
434 |
+
v,
|
435 |
+
s,
|
436 |
+
do,
|
437 |
+
dq,
|
438 |
+
dk,
|
439 |
+
dv,
|
440 |
+
q.stride(0),
|
441 |
+
q.stride(1),
|
442 |
+
q.stride(2),
|
443 |
+
q.stride(3),
|
444 |
+
k.stride(0),
|
445 |
+
k.stride(1),
|
446 |
+
k.stride(2),
|
447 |
+
k.stride(3),
|
448 |
+
v.stride(0),
|
449 |
+
v.stride(1),
|
450 |
+
v.stride(2),
|
451 |
+
v.stride(3),
|
452 |
+
do.stride(0),
|
453 |
+
do.stride(1),
|
454 |
+
do.stride(2),
|
455 |
+
do.stride(3),
|
456 |
+
s.stride(0),
|
457 |
+
q.shape[0],
|
458 |
+
q.shape[1],
|
459 |
+
q.shape[2],
|
460 |
+
grid_kv[0],
|
461 |
+
BLOCK_M=BLOCK_M,
|
462 |
+
BLOCK_DMODEL_QK=ctx.BLOCK_DMODEL_QK,
|
463 |
+
BLOCK_N=BLOCK_N,
|
464 |
+
BLOCK_DMODEL_V=ctx.BLOCK_DMODEL_V,
|
465 |
+
CAUSAL=ctx.causal,
|
466 |
+
USE_DECAY=ctx.use_decay,
|
467 |
+
num_warps=num_warps,
|
468 |
+
num_stages=num_stages,
|
469 |
+
)
|
470 |
+
|
471 |
+
grid_q = (triton.cdiv(q.shape[2], BLOCK_M), q.shape[0] * q.shape[1], 1)
|
472 |
+
|
473 |
+
_bwd_kernel_q[grid_q](
|
474 |
+
q,
|
475 |
+
k,
|
476 |
+
v,
|
477 |
+
s,
|
478 |
+
do,
|
479 |
+
dq,
|
480 |
+
dk,
|
481 |
+
dv,
|
482 |
+
q.stride(0),
|
483 |
+
q.stride(1),
|
484 |
+
q.stride(2),
|
485 |
+
q.stride(3),
|
486 |
+
k.stride(0),
|
487 |
+
k.stride(1),
|
488 |
+
k.stride(2),
|
489 |
+
k.stride(3),
|
490 |
+
v.stride(0),
|
491 |
+
v.stride(1),
|
492 |
+
v.stride(2),
|
493 |
+
v.stride(3),
|
494 |
+
do.stride(0),
|
495 |
+
do.stride(1),
|
496 |
+
do.stride(2),
|
497 |
+
do.stride(3),
|
498 |
+
s.stride(0),
|
499 |
+
q.shape[0],
|
500 |
+
q.shape[1],
|
501 |
+
q.shape[2],
|
502 |
+
grid_q[0],
|
503 |
+
BLOCK_M=BLOCK_M,
|
504 |
+
BLOCK_DMODEL_QK=ctx.BLOCK_DMODEL_QK,
|
505 |
+
BLOCK_N=BLOCK_N,
|
506 |
+
BLOCK_DMODEL_V=ctx.BLOCK_DMODEL_V,
|
507 |
+
CAUSAL=ctx.causal,
|
508 |
+
USE_DECAY=ctx.use_decay,
|
509 |
+
num_warps=num_warps,
|
510 |
+
num_stages=num_stages,
|
511 |
+
)
|
512 |
+
|
513 |
+
return dq.to(q.dtype), dk, dv, None, None
|
514 |
+
|
515 |
+
|
516 |
+
attention = _attention.apply
|
517 |
+
|
518 |
+
|
519 |
+
def lightning_attention(q, k, v, causal, ed):
|
520 |
+
d = q.shape[-1]
|
521 |
+
e = v.shape[-1]
|
522 |
+
# arr = f(d)
|
523 |
+
if d >= 128:
|
524 |
+
m = 128
|
525 |
+
else:
|
526 |
+
m = 64
|
527 |
+
arr = [m * i for i in range(d // m + 1)]
|
528 |
+
if arr[-1] != d:
|
529 |
+
arr.append(d)
|
530 |
+
n = len(arr)
|
531 |
+
output = 0
|
532 |
+
for i in range(n - 1):
|
533 |
+
s = arr[i]
|
534 |
+
e = arr[i + 1]
|
535 |
+
q1 = q[..., s:e]
|
536 |
+
k1 = k[..., s:e]
|
537 |
+
o = attention(q1, k1, v, causal, ed)
|
538 |
+
output = output + o
|
539 |
+
|
540 |
+
return output
|
lightning_attention2.py
ADDED
@@ -0,0 +1,540 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright 2024 OpenNLPLab
|
2 |
+
#
|
3 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
4 |
+
# you may not use this file except in compliance with the License.
|
5 |
+
# You may obtain a copy of the License at
|
6 |
+
#
|
7 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
8 |
+
#
|
9 |
+
# Unless required by applicable law or agreed to in writing, software
|
10 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
11 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
12 |
+
# See the License for the specific language governing permissions and
|
13 |
+
# limitations under the License.
|
14 |
+
|
15 |
+
# coding=utf-8
|
16 |
+
import torch
|
17 |
+
import triton
|
18 |
+
import triton.language as tl
|
19 |
+
|
20 |
+
|
21 |
+
@triton.jit
|
22 |
+
def _fwd_kernel(
|
23 |
+
Q,
|
24 |
+
K,
|
25 |
+
V,
|
26 |
+
Out,
|
27 |
+
S,
|
28 |
+
stride_qz,
|
29 |
+
stride_qh,
|
30 |
+
stride_qm,
|
31 |
+
stride_qk,
|
32 |
+
stride_kz,
|
33 |
+
stride_kh,
|
34 |
+
stride_kn,
|
35 |
+
stride_kk,
|
36 |
+
stride_vz,
|
37 |
+
stride_vh,
|
38 |
+
stride_vn,
|
39 |
+
stride_ve,
|
40 |
+
stride_oz,
|
41 |
+
stride_oh,
|
42 |
+
stride_om,
|
43 |
+
stride_oe,
|
44 |
+
stride_sh,
|
45 |
+
Z,
|
46 |
+
H,
|
47 |
+
N_CTX,
|
48 |
+
BLOCK_M: tl.constexpr,
|
49 |
+
BLOCK_DMODEL_QK: tl.constexpr,
|
50 |
+
BLOCK_N: tl.constexpr,
|
51 |
+
BLOCK_DMODEL_V: tl.constexpr,
|
52 |
+
IS_CAUSAL: tl.constexpr,
|
53 |
+
USE_DECAY: tl.constexpr,
|
54 |
+
):
|
55 |
+
start_m = tl.program_id(0)
|
56 |
+
off_hz = tl.program_id(1)
|
57 |
+
off_h = off_hz % H
|
58 |
+
# initialize offsets
|
59 |
+
offs_m = start_m * BLOCK_M + tl.arange(0, BLOCK_M)
|
60 |
+
offs_n = tl.arange(0, BLOCK_N)
|
61 |
+
offs_k = tl.arange(0, BLOCK_DMODEL_QK)
|
62 |
+
offs_e = tl.arange(0, BLOCK_DMODEL_V)
|
63 |
+
# get current offset of q k v
|
64 |
+
off_q = (off_hz * stride_qh + offs_m[:, None] * stride_qm
|
65 |
+
+ offs_k[None, :] * stride_qk)
|
66 |
+
off_k = (off_hz * stride_kh + offs_n[:, None] * stride_kn
|
67 |
+
+ offs_k[None, :] * stride_kk)
|
68 |
+
off_v = (off_hz * stride_vh + offs_n[:, None] * stride_vn
|
69 |
+
+ offs_e[None, :] * stride_ve)
|
70 |
+
off_o = (off_hz * stride_oh + offs_m[:, None] * stride_om
|
71 |
+
+ offs_e[None, :] * stride_oe)
|
72 |
+
|
73 |
+
# Initialize pointers to Q, K, V
|
74 |
+
q_ptrs = Q + off_q
|
75 |
+
k_ptrs = K + off_k
|
76 |
+
v_ptrs = V + off_v
|
77 |
+
|
78 |
+
# initialize pointer to m and l
|
79 |
+
acc = tl.zeros([BLOCK_M, BLOCK_DMODEL_V], dtype=tl.float32)
|
80 |
+
# load q: it will stay in SRAM throughout
|
81 |
+
q = tl.load(q_ptrs, mask=offs_m[:, None] < N_CTX, other=0.0)
|
82 |
+
# loop over k, v and update accumulator
|
83 |
+
lo = 0
|
84 |
+
# print(start_m)
|
85 |
+
hi = (start_m + 1) * BLOCK_M if IS_CAUSAL else N_CTX
|
86 |
+
for start_n in range(lo, hi, BLOCK_N):
|
87 |
+
# -- load k, v --
|
88 |
+
k = tl.load(
|
89 |
+
k_ptrs + start_n * stride_kn,
|
90 |
+
mask=(start_n + offs_n)[:, None] < N_CTX,
|
91 |
+
other=0.0,
|
92 |
+
)
|
93 |
+
v = tl.load(
|
94 |
+
v_ptrs + start_n * stride_vn,
|
95 |
+
mask=(start_n + offs_n)[:, None] < N_CTX,
|
96 |
+
other=0.0,
|
97 |
+
)
|
98 |
+
# -- compute qk ---
|
99 |
+
# qk = tl.dot(q, k)
|
100 |
+
qk = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32)
|
101 |
+
# qk += tl.dot(q, k, trans_b=True)
|
102 |
+
qk += tl.dot(q, tl.trans(k))
|
103 |
+
if IS_CAUSAL:
|
104 |
+
index = offs_m[:, None] - (start_n + offs_n[None, :])
|
105 |
+
if USE_DECAY:
|
106 |
+
S_block_ptr = S + off_h * stride_sh
|
107 |
+
s = tl.load(S_block_ptr)
|
108 |
+
s_index = s * index
|
109 |
+
s_index = tl.where(s_index >= 0, -s_index, float("-inf"))
|
110 |
+
qk = tl.exp(s_index) * qk
|
111 |
+
else:
|
112 |
+
qk = tl.where(index >= 0, qk, 0)
|
113 |
+
acc += tl.dot(qk, v.to(qk.dtype))
|
114 |
+
|
115 |
+
out_ptrs = Out + off_o
|
116 |
+
tl.store(out_ptrs, acc.to(q.dtype), mask=offs_m[:, None] < N_CTX)
|
117 |
+
|
118 |
+
|
119 |
+
@triton.jit
|
120 |
+
def _bwd_kernel_kv(
|
121 |
+
Q,
|
122 |
+
K,
|
123 |
+
V,
|
124 |
+
S,
|
125 |
+
DO,
|
126 |
+
DQ,
|
127 |
+
DK,
|
128 |
+
DV,
|
129 |
+
stride_qz,
|
130 |
+
stride_qh,
|
131 |
+
stride_qm,
|
132 |
+
stride_qk,
|
133 |
+
stride_kz,
|
134 |
+
stride_kh,
|
135 |
+
stride_kn,
|
136 |
+
stride_kk,
|
137 |
+
stride_vz,
|
138 |
+
stride_vh,
|
139 |
+
stride_vn,
|
140 |
+
stride_ve,
|
141 |
+
stride_oz,
|
142 |
+
stride_oh,
|
143 |
+
stride_om,
|
144 |
+
stride_oe,
|
145 |
+
stride_sh,
|
146 |
+
Z,
|
147 |
+
H,
|
148 |
+
N_CTX,
|
149 |
+
num_block,
|
150 |
+
BLOCK_M: tl.constexpr,
|
151 |
+
BLOCK_DMODEL_QK: tl.constexpr,
|
152 |
+
BLOCK_N: tl.constexpr,
|
153 |
+
BLOCK_DMODEL_V: tl.constexpr,
|
154 |
+
CAUSAL: tl.constexpr,
|
155 |
+
USE_DECAY: tl.constexpr,
|
156 |
+
):
|
157 |
+
start_n = tl.program_id(0)
|
158 |
+
off_hz = tl.program_id(1)
|
159 |
+
|
160 |
+
off_z = off_hz // H
|
161 |
+
off_h = off_hz % H
|
162 |
+
# offset pointers for batch/head
|
163 |
+
Q += off_z * stride_qz + off_h * stride_qh
|
164 |
+
K += off_z * stride_kz + off_h * stride_kh
|
165 |
+
V += off_z * stride_vz + off_h * stride_vh
|
166 |
+
DO += off_z * stride_oz + off_h * stride_oh
|
167 |
+
DQ += off_z * stride_qz + off_h * stride_qh
|
168 |
+
DK += off_z * stride_kz + off_h * stride_kh
|
169 |
+
DV += off_z * stride_vz + off_h * stride_vh
|
170 |
+
|
171 |
+
# start of q
|
172 |
+
if CAUSAL:
|
173 |
+
lo = start_n * BLOCK_M
|
174 |
+
else:
|
175 |
+
lo = 0
|
176 |
+
# initialize row/col offsets
|
177 |
+
# seqlence offset
|
178 |
+
offs_qm = lo + tl.arange(0, BLOCK_M)
|
179 |
+
offs_kvn = start_n * BLOCK_N + tl.arange(0, BLOCK_N)
|
180 |
+
# feature offset
|
181 |
+
offs_qkk = tl.arange(0, BLOCK_DMODEL_QK)
|
182 |
+
offs_ve = tl.arange(0, BLOCK_DMODEL_V)
|
183 |
+
# row block index
|
184 |
+
offs_m = tl.arange(0, BLOCK_M)
|
185 |
+
# initialize pointers to value-like data
|
186 |
+
q_ptrs = Q + (offs_qm[:, None] * stride_qm + offs_qkk[None, :] * stride_qk)
|
187 |
+
k_ptrs = K + (offs_kvn[:, None] * stride_kn
|
188 |
+
+ offs_qkk[None, :] * stride_kk)
|
189 |
+
v_ptrs = V + (offs_kvn[:, None] * stride_vn + offs_ve[None, :] * stride_ve)
|
190 |
+
do_ptrs = DO + (offs_qm[:, None] * stride_om
|
191 |
+
+ offs_ve[None, :] * stride_oe)
|
192 |
+
dq_ptrs = DQ + (offs_qm[:, None] * stride_qm
|
193 |
+
+ offs_qkk[None, :] * stride_qk)
|
194 |
+
# initialize dv amd dk
|
195 |
+
dv = tl.zeros([BLOCK_N, BLOCK_DMODEL_V], dtype=tl.float32)
|
196 |
+
dk = tl.zeros([BLOCK_N, BLOCK_DMODEL_QK], dtype=tl.float32)
|
197 |
+
# k and v stay in SRAM throughout
|
198 |
+
k = tl.load(k_ptrs, mask=offs_kvn[:, None] < N_CTX, other=0.0)
|
199 |
+
v = tl.load(v_ptrs, mask=offs_kvn[:, None] < N_CTX, other=0.0)
|
200 |
+
# loop over rows
|
201 |
+
for start_m in range(lo, num_block * BLOCK_M, BLOCK_M):
|
202 |
+
offs_m_curr = start_m + offs_m
|
203 |
+
# load q, k, v, do on-chip
|
204 |
+
q = tl.load(q_ptrs, mask=offs_m_curr[:, None] < N_CTX, other=0.0)
|
205 |
+
qk = tl.dot(q, tl.trans(k))
|
206 |
+
# qk = tl.dot(q, k, trans_b=True)
|
207 |
+
if CAUSAL:
|
208 |
+
index = offs_m_curr[:, None] - offs_kvn[None, :]
|
209 |
+
if USE_DECAY:
|
210 |
+
S_block_ptr = S + off_h * stride_sh
|
211 |
+
s = tl.load(S_block_ptr)
|
212 |
+
s_index = s * index
|
213 |
+
s_index = tl.where(s_index >= 0, -s_index, float("-inf"))
|
214 |
+
s = tl.exp(s_index)
|
215 |
+
qk = qk * s
|
216 |
+
else:
|
217 |
+
qk = tl.where(index >= 0, qk, 0)
|
218 |
+
|
219 |
+
p = qk
|
220 |
+
# compute dv
|
221 |
+
do = tl.load(do_ptrs, mask=offs_m_curr[:, None] < N_CTX, other=0.0)
|
222 |
+
dv += tl.dot(tl.trans(p.to(do.dtype)), do)
|
223 |
+
dp = tl.dot(do, tl.trans(v).to(do.dtype))
|
224 |
+
if CAUSAL:
|
225 |
+
if USE_DECAY:
|
226 |
+
dp = dp * s
|
227 |
+
else:
|
228 |
+
dp = tl.where(index >= 0, dp, 0)
|
229 |
+
|
230 |
+
dk += tl.dot(tl.trans(dp.to(q.dtype)), q).to(tl.float32)
|
231 |
+
|
232 |
+
# increment pointers
|
233 |
+
q_ptrs += BLOCK_M * stride_qm
|
234 |
+
do_ptrs += BLOCK_M * stride_om
|
235 |
+
# write-back
|
236 |
+
dv_ptrs = DV + (offs_kvn[:, None] * stride_vn
|
237 |
+
+ offs_ve[None, :] * stride_ve)
|
238 |
+
dk_ptrs = DK + (offs_kvn[:, None] * stride_kn
|
239 |
+
+ offs_qkk[None, :] * stride_kk)
|
240 |
+
tl.store(dv_ptrs, dv, mask=offs_kvn[:, None] < N_CTX)
|
241 |
+
tl.store(dk_ptrs, dk, mask=offs_kvn[:, None] < N_CTX)
|
242 |
+
|
243 |
+
|
244 |
+
@triton.jit
|
245 |
+
def _bwd_kernel_q(
|
246 |
+
Q,
|
247 |
+
K,
|
248 |
+
V,
|
249 |
+
S,
|
250 |
+
DO,
|
251 |
+
DQ,
|
252 |
+
DK,
|
253 |
+
DV,
|
254 |
+
stride_qz,
|
255 |
+
stride_qh,
|
256 |
+
stride_qm,
|
257 |
+
stride_qk,
|
258 |
+
stride_kz,
|
259 |
+
stride_kh,
|
260 |
+
stride_kn,
|
261 |
+
stride_kk,
|
262 |
+
stride_vz,
|
263 |
+
stride_vh,
|
264 |
+
stride_vn,
|
265 |
+
stride_ve,
|
266 |
+
stride_oz,
|
267 |
+
stride_oh,
|
268 |
+
stride_om,
|
269 |
+
stride_oe,
|
270 |
+
stride_sh,
|
271 |
+
Z,
|
272 |
+
H,
|
273 |
+
N_CTX,
|
274 |
+
num_block,
|
275 |
+
BLOCK_M: tl.constexpr,
|
276 |
+
BLOCK_DMODEL_QK: tl.constexpr,
|
277 |
+
BLOCK_N: tl.constexpr,
|
278 |
+
BLOCK_DMODEL_V: tl.constexpr,
|
279 |
+
CAUSAL: tl.constexpr,
|
280 |
+
USE_DECAY: tl.constexpr,
|
281 |
+
):
|
282 |
+
start_m = tl.program_id(0)
|
283 |
+
off_hz = tl.program_id(1)
|
284 |
+
off_z = off_hz // H
|
285 |
+
off_h = off_hz % H
|
286 |
+
# offset pointers for batch/head
|
287 |
+
K += off_z * stride_kz + off_h * stride_kh
|
288 |
+
V += off_z * stride_vz + off_h * stride_vh
|
289 |
+
DO += off_z * stride_oz + off_h * stride_oh
|
290 |
+
DQ += off_z * stride_qz + off_h * stride_qh
|
291 |
+
# feature offset
|
292 |
+
offs_qkk = tl.arange(0, BLOCK_DMODEL_QK)
|
293 |
+
offs_ve = tl.arange(0, BLOCK_DMODEL_V)
|
294 |
+
# row block index
|
295 |
+
offs_m = tl.arange(0, BLOCK_M)
|
296 |
+
# row block index
|
297 |
+
offs_qm = start_m * BLOCK_M + tl.arange(0, BLOCK_M)
|
298 |
+
# do
|
299 |
+
do_ptrs = DO + (offs_qm[:, None] * stride_om
|
300 |
+
+ offs_ve[None, :] * stride_oe)
|
301 |
+
dq_ptrs = DQ + (offs_qm[:, None] * stride_qm
|
302 |
+
+ offs_qkk[None, :] * stride_qk)
|
303 |
+
|
304 |
+
do = tl.load(do_ptrs, mask=offs_qm[:, None] < N_CTX, other=0.0)
|
305 |
+
|
306 |
+
dq = tl.zeros([BLOCK_M, BLOCK_DMODEL_QK], dtype=tl.float32)
|
307 |
+
lo = 0
|
308 |
+
hi = (start_m + 1) * BLOCK_M if CAUSAL else N_CTX
|
309 |
+
|
310 |
+
offs_m_curr = start_m * BLOCK_M + offs_m
|
311 |
+
|
312 |
+
for start_n in range(0, num_block):
|
313 |
+
offs_kvn = start_n * BLOCK_N + tl.arange(0, BLOCK_N)
|
314 |
+
k_ptrs = K + (offs_kvn[:, None] * stride_kn
|
315 |
+
+ offs_qkk[None, :] * stride_kk)
|
316 |
+
v_ptrs = V + (offs_kvn[:, None] * stride_vn
|
317 |
+
+ offs_ve[None, :] * stride_ve)
|
318 |
+
# k and v stay in SRAM throughout
|
319 |
+
k = tl.load(k_ptrs, mask=offs_kvn[:, None] < N_CTX, other=0.0)
|
320 |
+
v = tl.load(v_ptrs, mask=offs_kvn[:, None] < N_CTX, other=0.0)
|
321 |
+
# dp = do vT
|
322 |
+
dp = tl.dot(do, tl.trans(v).to(do.dtype))
|
323 |
+
if CAUSAL:
|
324 |
+
index = offs_m_curr[:, None] - offs_kvn[None, :]
|
325 |
+
if USE_DECAY:
|
326 |
+
S_block_ptr = S + off_h * stride_sh
|
327 |
+
s = tl.load(S_block_ptr)
|
328 |
+
s_index = s * index
|
329 |
+
s_index = tl.where(s_index >= 0, -s_index, float("-inf"))
|
330 |
+
s = tl.exp(s_index)
|
331 |
+
dp = dp * s
|
332 |
+
else:
|
333 |
+
dp = tl.where(index >= 0, dp, 0)
|
334 |
+
# dq = dq + dp k
|
335 |
+
dq += tl.dot(dp.to(k.dtype), k)
|
336 |
+
|
337 |
+
tl.store(dq_ptrs, dq, mask=offs_qm[:, None] < N_CTX)
|
338 |
+
|
339 |
+
|
340 |
+
class _attention(torch.autograd.Function):
|
341 |
+
|
342 |
+
@staticmethod
|
343 |
+
def forward(ctx, q, k, v, causal, s):
|
344 |
+
q = q.contiguous()
|
345 |
+
k = k.contiguous()
|
346 |
+
v = v.contiguous()
|
347 |
+
s = s.contiguous()
|
348 |
+
# only support for Ampere now
|
349 |
+
capability = torch.cuda.get_device_capability()
|
350 |
+
if capability[0] < 8:
|
351 |
+
raise RuntimeError(
|
352 |
+
"Lightning attention currently only supported for compute capability >= 80"
|
353 |
+
)
|
354 |
+
# shape constraints
|
355 |
+
Lq, Lk, Lv = q.shape[-1], k.shape[-1], v.shape[-1]
|
356 |
+
# right
|
357 |
+
o = torch.empty(
|
358 |
+
(q.shape[0], q.shape[1], q.shape[2], v.shape[-1]),
|
359 |
+
dtype=q.dtype,
|
360 |
+
device=q.device,
|
361 |
+
)
|
362 |
+
|
363 |
+
BLOCK_M = 128
|
364 |
+
BLOCK_N = 64
|
365 |
+
num_warps = 4 if Lk <= 64 else 8
|
366 |
+
num_stages = 1
|
367 |
+
|
368 |
+
grid = (triton.cdiv(q.shape[2], BLOCK_M), q.shape[0] * q.shape[1], 1)
|
369 |
+
use_decay = s.shape[0] > 0
|
370 |
+
_fwd_kernel[grid](
|
371 |
+
q,
|
372 |
+
k,
|
373 |
+
v,
|
374 |
+
o,
|
375 |
+
s,
|
376 |
+
q.stride(0),
|
377 |
+
q.stride(1),
|
378 |
+
q.stride(2),
|
379 |
+
q.stride(3),
|
380 |
+
k.stride(0),
|
381 |
+
k.stride(1),
|
382 |
+
k.stride(2),
|
383 |
+
k.stride(3),
|
384 |
+
v.stride(0),
|
385 |
+
v.stride(1),
|
386 |
+
v.stride(2),
|
387 |
+
v.stride(3),
|
388 |
+
o.stride(0),
|
389 |
+
o.stride(1),
|
390 |
+
o.stride(2),
|
391 |
+
o.stride(3),
|
392 |
+
s.stride(0),
|
393 |
+
q.shape[0],
|
394 |
+
q.shape[1],
|
395 |
+
q.shape[2],
|
396 |
+
BLOCK_M=BLOCK_M,
|
397 |
+
BLOCK_DMODEL_QK=Lk,
|
398 |
+
BLOCK_N=BLOCK_N,
|
399 |
+
BLOCK_DMODEL_V=Lv,
|
400 |
+
IS_CAUSAL=causal,
|
401 |
+
USE_DECAY=use_decay,
|
402 |
+
num_warps=num_warps,
|
403 |
+
num_stages=num_stages,
|
404 |
+
)
|
405 |
+
|
406 |
+
ctx.save_for_backward(q, k, v, s)
|
407 |
+
ctx.grid = grid
|
408 |
+
ctx.BLOCK_M = BLOCK_M
|
409 |
+
ctx.BLOCK_DMODEL_QK = Lk
|
410 |
+
ctx.BLOCK_N = BLOCK_N
|
411 |
+
ctx.BLOCK_DMODEL_V = Lv
|
412 |
+
ctx.causal = causal
|
413 |
+
ctx.use_decay = use_decay
|
414 |
+
return o
|
415 |
+
|
416 |
+
@staticmethod
|
417 |
+
def backward(ctx, do):
|
418 |
+
q, k, v, s = ctx.saved_tensors
|
419 |
+
BLOCK_M = 32
|
420 |
+
BLOCK_N = 32
|
421 |
+
num_warps = 4
|
422 |
+
num_stages = 1
|
423 |
+
|
424 |
+
do = do.contiguous()
|
425 |
+
dq = torch.zeros_like(q, dtype=torch.float32)
|
426 |
+
dk = torch.empty_like(k)
|
427 |
+
dv = torch.empty_like(v)
|
428 |
+
|
429 |
+
grid_kv = (triton.cdiv(k.shape[2],
|
430 |
+
BLOCK_N), k.shape[0] * k.shape[1], 1)
|
431 |
+
_bwd_kernel_kv[grid_kv](
|
432 |
+
q,
|
433 |
+
k,
|
434 |
+
v,
|
435 |
+
s,
|
436 |
+
do,
|
437 |
+
dq,
|
438 |
+
dk,
|
439 |
+
dv,
|
440 |
+
q.stride(0),
|
441 |
+
q.stride(1),
|
442 |
+
q.stride(2),
|
443 |
+
q.stride(3),
|
444 |
+
k.stride(0),
|
445 |
+
k.stride(1),
|
446 |
+
k.stride(2),
|
447 |
+
k.stride(3),
|
448 |
+
v.stride(0),
|
449 |
+
v.stride(1),
|
450 |
+
v.stride(2),
|
451 |
+
v.stride(3),
|
452 |
+
do.stride(0),
|
453 |
+
do.stride(1),
|
454 |
+
do.stride(2),
|
455 |
+
do.stride(3),
|
456 |
+
s.stride(0),
|
457 |
+
q.shape[0],
|
458 |
+
q.shape[1],
|
459 |
+
q.shape[2],
|
460 |
+
grid_kv[0],
|
461 |
+
BLOCK_M=BLOCK_M,
|
462 |
+
BLOCK_DMODEL_QK=ctx.BLOCK_DMODEL_QK,
|
463 |
+
BLOCK_N=BLOCK_N,
|
464 |
+
BLOCK_DMODEL_V=ctx.BLOCK_DMODEL_V,
|
465 |
+
CAUSAL=ctx.causal,
|
466 |
+
USE_DECAY=ctx.use_decay,
|
467 |
+
num_warps=num_warps,
|
468 |
+
num_stages=num_stages,
|
469 |
+
)
|
470 |
+
|
471 |
+
grid_q = (triton.cdiv(q.shape[2], BLOCK_M), q.shape[0] * q.shape[1], 1)
|
472 |
+
|
473 |
+
_bwd_kernel_q[grid_q](
|
474 |
+
q,
|
475 |
+
k,
|
476 |
+
v,
|
477 |
+
s,
|
478 |
+
do,
|
479 |
+
dq,
|
480 |
+
dk,
|
481 |
+
dv,
|
482 |
+
q.stride(0),
|
483 |
+
q.stride(1),
|
484 |
+
q.stride(2),
|
485 |
+
q.stride(3),
|
486 |
+
k.stride(0),
|
487 |
+
k.stride(1),
|
488 |
+
k.stride(2),
|
489 |
+
k.stride(3),
|
490 |
+
v.stride(0),
|
491 |
+
v.stride(1),
|
492 |
+
v.stride(2),
|
493 |
+
v.stride(3),
|
494 |
+
do.stride(0),
|
495 |
+
do.stride(1),
|
496 |
+
do.stride(2),
|
497 |
+
do.stride(3),
|
498 |
+
s.stride(0),
|
499 |
+
q.shape[0],
|
500 |
+
q.shape[1],
|
501 |
+
q.shape[2],
|
502 |
+
grid_q[0],
|
503 |
+
BLOCK_M=BLOCK_M,
|
504 |
+
BLOCK_DMODEL_QK=ctx.BLOCK_DMODEL_QK,
|
505 |
+
BLOCK_N=BLOCK_N,
|
506 |
+
BLOCK_DMODEL_V=ctx.BLOCK_DMODEL_V,
|
507 |
+
CAUSAL=ctx.causal,
|
508 |
+
USE_DECAY=ctx.use_decay,
|
509 |
+
num_warps=num_warps,
|
510 |
+
num_stages=num_stages,
|
511 |
+
)
|
512 |
+
|
513 |
+
return dq.to(q.dtype), dk, dv, None, None
|
514 |
+
|
515 |
+
|
516 |
+
attention = _attention.apply
|
517 |
+
|
518 |
+
|
519 |
+
def lightning_attention(q, k, v, causal, ed):
|
520 |
+
d = q.shape[-1]
|
521 |
+
e = v.shape[-1]
|
522 |
+
# arr = f(d)
|
523 |
+
if d >= 128:
|
524 |
+
m = 128
|
525 |
+
else:
|
526 |
+
m = 64
|
527 |
+
arr = [m * i for i in range(d // m + 1)]
|
528 |
+
if arr[-1] != d:
|
529 |
+
arr.append(d)
|
530 |
+
n = len(arr)
|
531 |
+
output = 0
|
532 |
+
for i in range(n - 1):
|
533 |
+
s = arr[i]
|
534 |
+
e = arr[i + 1]
|
535 |
+
q1 = q[..., s:e]
|
536 |
+
k1 = k[..., s:e]
|
537 |
+
o = attention(q1, k1, v, causal, ed)
|
538 |
+
output = output + o
|
539 |
+
|
540 |
+
return output
|
model.safetensors
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:f02c524d2f9374fe6aa8ae63d13064be3625423be672ac653bc40a071c17e3bd
|
3 |
+
size 769867776
|
modeling_transnormer.py
ADDED
@@ -0,0 +1,933 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright 2024 OpenNLPLab
|
2 |
+
#
|
3 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
4 |
+
# you may not use this file except in compliance with the License.
|
5 |
+
# You may obtain a copy of the License at
|
6 |
+
#
|
7 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
8 |
+
#
|
9 |
+
# Unless required by applicable law or agreed to in writing, software
|
10 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
11 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
12 |
+
# See the License for the specific language governing permissions and
|
13 |
+
# limitations under the License.
|
14 |
+
|
15 |
+
# coding=utf-8
|
16 |
+
""" PyTorch Transnormer model."""
|
17 |
+
import math
|
18 |
+
import os
|
19 |
+
from typing import List, Optional, Tuple, Union
|
20 |
+
|
21 |
+
from einops import rearrange
|
22 |
+
import numpy as np
|
23 |
+
import torch
|
24 |
+
from torch import nn
|
25 |
+
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
|
26 |
+
import torch.nn.functional as F
|
27 |
+
import torch.utils.checkpoint
|
28 |
+
from transformers.activations import ACT2FN
|
29 |
+
from transformers.modeling_outputs import (
|
30 |
+
BaseModelOutputWithPast,
|
31 |
+
CausalLMOutputWithPast,
|
32 |
+
)
|
33 |
+
from transformers.modeling_utils import PreTrainedModel
|
34 |
+
from transformers.utils import (
|
35 |
+
add_start_docstrings,
|
36 |
+
add_start_docstrings_to_model_forward,
|
37 |
+
logging,
|
38 |
+
replace_return_docstrings,
|
39 |
+
)
|
40 |
+
|
41 |
+
from .configuration_transnormer import TransnormerConfig
|
42 |
+
from .norm import SimpleRMSNorm as SimpleRMSNorm_torch
|
43 |
+
from .srmsnorm_triton import SimpleRMSNorm as SimpleRMSNorm_triton
|
44 |
+
from .utils import (
|
45 |
+
get_activation_fn,
|
46 |
+
get_norm_fn,
|
47 |
+
logging_info,
|
48 |
+
print_module,
|
49 |
+
print_params,
|
50 |
+
)
|
51 |
+
|
52 |
+
logger = logging.get_logger(__name__)
|
53 |
+
|
54 |
+
_CONFIG_FOR_DOC = "TransnormerConfig"
|
55 |
+
|
56 |
+
# TODO: fix environment: https://huggingface.co/OpenNLPLab/TransNormerLLM-7B/discussions/1
|
57 |
+
use_triton = eval(os.environ.get("use_triton", default="True"))
|
58 |
+
debug = eval(os.environ.get("debug", default="False"))
|
59 |
+
do_eval = eval(os.environ.get("do_eval", default="False"))
|
60 |
+
eval_and_not_generate = eval(
|
61 |
+
os.environ.get("eval_and_not_generate", default="False"))
|
62 |
+
BLOCK = 256
|
63 |
+
|
64 |
+
if use_triton:
|
65 |
+
try:
|
66 |
+
from .lightning_attention2 import lightning_attention
|
67 |
+
|
68 |
+
has_lightning_attention = True
|
69 |
+
except (ImportError, ModuleNotFoundError):
|
70 |
+
has_lightning_attention = False
|
71 |
+
else:
|
72 |
+
has_lightning_attention = False
|
73 |
+
|
74 |
+
if debug:
|
75 |
+
logger.info(f"Use triton: {use_triton}")
|
76 |
+
logger.info(f"Use lightning attention: {has_lightning_attention}")
|
77 |
+
logger.info(f"Debug mode: {debug}, {type(debug)}")
|
78 |
+
|
79 |
+
if not has_lightning_attention:
|
80 |
+
|
81 |
+
def linear_attention(q, k, v, attn_mask):
|
82 |
+
energy = torch.einsum("... n d, ... m d -> ... n m", q, k)
|
83 |
+
energy = energy * attn_mask
|
84 |
+
output = torch.einsum("... n m, ... m d -> ... n d", energy, v)
|
85 |
+
|
86 |
+
return output
|
87 |
+
|
88 |
+
|
89 |
+
########## start Transnormer
|
90 |
+
##### Linearized Relative Positional Encoding: https://openreview.net/forum?id=xoLyps2qWc&referrer=%5BAuthor%20Console%5D(%2Fgroup%3Fid%3DTMLR%2FAuthors%23your-submissions)
|
91 |
+
class Lrpe(nn.Module):
|
92 |
+
|
93 |
+
def __init__(
|
94 |
+
self,
|
95 |
+
num_heads=8,
|
96 |
+
embed_dim=64,
|
97 |
+
):
|
98 |
+
super().__init__()
|
99 |
+
d = num_heads * embed_dim
|
100 |
+
|
101 |
+
self.index = torch.empty(0)
|
102 |
+
self.theta = nn.Parameter(10000**(-2 / d * torch.arange(d)).reshape(
|
103 |
+
num_heads, 1, -1))
|
104 |
+
|
105 |
+
def extra_repr(self):
|
106 |
+
return print_module(self)
|
107 |
+
|
108 |
+
def forward(self, x, offset=0):
|
109 |
+
# x: b, h, n, d
|
110 |
+
# offset: for k, v cache
|
111 |
+
n = x.shape[-2]
|
112 |
+
if self.index.shape[0] < n:
|
113 |
+
self.index = torch.arange(n).reshape(1, -1, 1).to(x)
|
114 |
+
index = self.index[:, :n] + offset
|
115 |
+
theta = self.theta * index
|
116 |
+
x = torch.concat([x * torch.cos(theta), x * torch.sin(theta)], dim=-1)
|
117 |
+
|
118 |
+
return x
|
119 |
+
|
120 |
+
|
121 |
+
class GLU(nn.Module):
|
122 |
+
|
123 |
+
def __init__(self, d1, d2, bias=False):
|
124 |
+
super().__init__()
|
125 |
+
if debug:
|
126 |
+
# get local varables
|
127 |
+
params = locals()
|
128 |
+
# print params
|
129 |
+
print_params(**params)
|
130 |
+
|
131 |
+
self.l1 = nn.Linear(d1, d2, bias=bias)
|
132 |
+
self.l2 = nn.Linear(d1, d2, bias=bias)
|
133 |
+
self.l3 = nn.Linear(d2, d1, bias=bias)
|
134 |
+
|
135 |
+
def forward(self, x):
|
136 |
+
o1 = self.l1(x)
|
137 |
+
o2 = self.l2(x)
|
138 |
+
output = o1 * o2
|
139 |
+
output = self.l3(output)
|
140 |
+
|
141 |
+
return output
|
142 |
+
|
143 |
+
|
144 |
+
class NormLinearAttention(nn.Module):
|
145 |
+
|
146 |
+
def __init__(
|
147 |
+
self,
|
148 |
+
embed_dim,
|
149 |
+
hidden_dim,
|
150 |
+
num_heads,
|
151 |
+
linear_act_fun="silu",
|
152 |
+
norm_type="simplermsnorm",
|
153 |
+
linear_use_lrpe=False,
|
154 |
+
bias=False,
|
155 |
+
):
|
156 |
+
super().__init__()
|
157 |
+
if debug:
|
158 |
+
# get local varables
|
159 |
+
params = locals()
|
160 |
+
# print params
|
161 |
+
print_params(**params)
|
162 |
+
|
163 |
+
self.out_proj = nn.Linear(hidden_dim, embed_dim, bias=bias)
|
164 |
+
self.act = get_activation_fn(linear_act_fun)
|
165 |
+
self.num_heads = num_heads
|
166 |
+
self.embed_dim = embed_dim
|
167 |
+
self.head_dim = self.embed_dim // self.num_heads
|
168 |
+
self.norm = get_norm_fn(norm_type)(hidden_dim)
|
169 |
+
|
170 |
+
self.linear_use_lrpe = linear_use_lrpe
|
171 |
+
if self.linear_use_lrpe:
|
172 |
+
self.lrpe = Lrpe(
|
173 |
+
num_heads=self.num_heads,
|
174 |
+
embed_dim=self.head_dim,
|
175 |
+
)
|
176 |
+
|
177 |
+
self.qkvu_proj = nn.Linear(embed_dim, 4 * hidden_dim, bias=bias)
|
178 |
+
|
179 |
+
# for inference only
|
180 |
+
self.offset = 0
|
181 |
+
|
182 |
+
def forward(
|
183 |
+
self,
|
184 |
+
x,
|
185 |
+
attn_mask: Optional[torch.Tensor] = None, # (b, h, n, m)
|
186 |
+
attn_padding_mask: Optional[torch.Tensor] = None, # (b, m)
|
187 |
+
output_attentions: bool = False,
|
188 |
+
past_key_value: Optional[Tuple[torch.Tensor]] = None,
|
189 |
+
use_cache: bool = False,
|
190 |
+
slope_rate: Optional[torch.Tensor] = None,
|
191 |
+
):
|
192 |
+
if (not self.training) and (not do_eval):
|
193 |
+
return self.inference(
|
194 |
+
x,
|
195 |
+
attn_mask,
|
196 |
+
attn_padding_mask,
|
197 |
+
output_attentions,
|
198 |
+
past_key_value,
|
199 |
+
use_cache,
|
200 |
+
slope_rate,
|
201 |
+
)
|
202 |
+
# x: b n d
|
203 |
+
n = x.shape[-2]
|
204 |
+
# linear map
|
205 |
+
q, k, v, u = self.qkvu_proj(x).chunk(4, dim=-1)
|
206 |
+
# reshape
|
207 |
+
q, k, v = map(
|
208 |
+
lambda x: rearrange(x, "b n (h d) -> b h n d", h=self.num_heads),
|
209 |
+
[q, k, v])
|
210 |
+
# act
|
211 |
+
q = self.act(q)
|
212 |
+
k = self.act(k)
|
213 |
+
|
214 |
+
q_offset = 0
|
215 |
+
# lrpe relys on position, get cache first
|
216 |
+
if past_key_value is not None:
|
217 |
+
# reuse k, v, for evaluation only
|
218 |
+
k = torch.cat([past_key_value[0], k], dim=-2)
|
219 |
+
v = torch.cat([past_key_value[1], v], dim=-2)
|
220 |
+
q_offset = past_key_value[0].shape[-2]
|
221 |
+
|
222 |
+
past_key_value = (k, v) if use_cache else None
|
223 |
+
|
224 |
+
# lrpe
|
225 |
+
if self.linear_use_lrpe:
|
226 |
+
q = self.lrpe(q, offset=q_offset)
|
227 |
+
k = self.lrpe(k, offset=q_offset)
|
228 |
+
|
229 |
+
if attn_mask == None:
|
230 |
+
attn_mask = (torch.tril(torch.ones(n, n))).to(q)
|
231 |
+
|
232 |
+
if attn_padding_mask is not None:
|
233 |
+
v = v.masked_fill(
|
234 |
+
(1 - attn_padding_mask).unsqueeze(1).unsqueeze(-1).to(
|
235 |
+
torch.bool), 0)
|
236 |
+
|
237 |
+
if not has_lightning_attention:
|
238 |
+
if slope_rate != None:
|
239 |
+
attn_mask = torch.exp(slope_rate * attn_mask)
|
240 |
+
output = linear_attention(q, k, v, attn_mask)
|
241 |
+
else:
|
242 |
+
output = lightning_attention(q, k, v, True,
|
243 |
+
slope_rate.squeeze(-1).squeeze(-1))
|
244 |
+
|
245 |
+
# reshape
|
246 |
+
output = rearrange(output, "b h n d -> b n (h d)")
|
247 |
+
# normalize
|
248 |
+
output = self.norm(output)
|
249 |
+
# gate
|
250 |
+
output = u * output
|
251 |
+
# outproj
|
252 |
+
output = self.out_proj(output)
|
253 |
+
|
254 |
+
if not output_attentions:
|
255 |
+
attn_weights = None
|
256 |
+
else:
|
257 |
+
attn_weights = torch.einsum("... n d, ... m d -> ... n m", q, k)
|
258 |
+
|
259 |
+
return output, attn_weights, past_key_value
|
260 |
+
|
261 |
+
def inference(
|
262 |
+
self,
|
263 |
+
x,
|
264 |
+
attn_mask: Optional[torch.Tensor] = None, # (b, h, n, m)
|
265 |
+
attn_padding_mask: Optional[torch.Tensor] = None, # (b, m)
|
266 |
+
output_attentions: bool = False,
|
267 |
+
past_key_value: Optional[Tuple[torch.Tensor]] = None,
|
268 |
+
use_cache: bool = False,
|
269 |
+
slope_rate: Optional[torch.Tensor] = None, # (h, 1, 1)
|
270 |
+
):
|
271 |
+
# x: b n d
|
272 |
+
n = x.shape[-2]
|
273 |
+
# linear map
|
274 |
+
q, k, v, u = self.qkvu_proj(x).chunk(4, dim=-1)
|
275 |
+
# reshape
|
276 |
+
q, k, v = map(
|
277 |
+
lambda x: rearrange(x, "b n (h d) -> b h n d", h=self.num_heads),
|
278 |
+
[q, k, v])
|
279 |
+
# act
|
280 |
+
q = self.act(q)
|
281 |
+
k = self.act(k)
|
282 |
+
|
283 |
+
# rpe
|
284 |
+
if self.linear_use_lrpe:
|
285 |
+
q = self.lrpe(q, offset=self.offset)
|
286 |
+
k = self.lrpe(k, offset=self.offset)
|
287 |
+
|
288 |
+
if past_key_value == None:
|
289 |
+
self.offset = q.shape[-2]
|
290 |
+
else:
|
291 |
+
self.offset += 1
|
292 |
+
|
293 |
+
ratio = torch.exp(-slope_rate)
|
294 |
+
|
295 |
+
# only use for the first time
|
296 |
+
if past_key_value == None:
|
297 |
+
slope_rate = slope_rate.to(torch.float32)
|
298 |
+
if attn_padding_mask is not None:
|
299 |
+
v = v.masked_fill(
|
300 |
+
(1 - attn_padding_mask).unsqueeze(1).unsqueeze(-1).to(
|
301 |
+
torch.bool), 0)
|
302 |
+
NUM_BLOCK = (n + BLOCK - 1) // BLOCK
|
303 |
+
b, h, n, d = q.shape
|
304 |
+
e = v.shape[-1]
|
305 |
+
# other
|
306 |
+
array = torch.arange(BLOCK).to(q) + 1 ## !!!! important
|
307 |
+
q_decay = torch.exp(-slope_rate * array.reshape(-1, 1))
|
308 |
+
k_decay = torch.exp(-slope_rate * (BLOCK - array.reshape(-1, 1)))
|
309 |
+
index = array[:, None] - array[None, :]
|
310 |
+
s_index = slope_rate * index[
|
311 |
+
None,
|
312 |
+
None,
|
313 |
+
]
|
314 |
+
s_index = torch.where(index >= 0, -s_index, float("-inf"))
|
315 |
+
diag_decay = torch.exp(s_index)
|
316 |
+
|
317 |
+
kv = torch.zeros(b, h, d, e).to(torch.float32).to(q.device)
|
318 |
+
output = torch.empty((b, h, n, e), dtype=q.dtype, device=q.device)
|
319 |
+
for i in range(NUM_BLOCK):
|
320 |
+
si = i * BLOCK
|
321 |
+
ei = min(si + BLOCK, n)
|
322 |
+
m = ei - si
|
323 |
+
|
324 |
+
qi = q[:, :, si:ei].contiguous()
|
325 |
+
ki = k[:, :, si:ei].contiguous()
|
326 |
+
vi = v[:, :, si:ei].contiguous()
|
327 |
+
qkv_none_diag = torch.matmul(qi * q_decay[:, :m],
|
328 |
+
kv).to(torch.float32)
|
329 |
+
|
330 |
+
# diag
|
331 |
+
qk = torch.matmul(qi, ki.transpose(-1, -2)).to(
|
332 |
+
torch.float32) * diag_decay[:, :, :m, :m]
|
333 |
+
qkv_diag = torch.matmul(qk, vi.to(torch.float32))
|
334 |
+
block_decay = torch.exp(-slope_rate * m)
|
335 |
+
output[:, :, si:ei] = qkv_none_diag + qkv_diag
|
336 |
+
kv = block_decay * kv + torch.matmul(
|
337 |
+
(ki * k_decay[:, -m:]).transpose(-1, -2).to(vi.dtype), vi)
|
338 |
+
else:
|
339 |
+
kv = past_key_value
|
340 |
+
|
341 |
+
output = []
|
342 |
+
for i in range(n):
|
343 |
+
kv = ratio * kv + torch.einsum(
|
344 |
+
"... n d, ... n e -> ... d e",
|
345 |
+
k[:, :, i:i + 1],
|
346 |
+
v[:, :, i:i + 1],
|
347 |
+
)
|
348 |
+
qkv = torch.einsum("... n e, ... e d -> ... n d", q[:, :,
|
349 |
+
i:i + 1],
|
350 |
+
kv.to(q.dtype))
|
351 |
+
output.append(qkv)
|
352 |
+
output = torch.concat(output, dim=-2)
|
353 |
+
|
354 |
+
# reshape
|
355 |
+
output = rearrange(output, "b h n d -> b n (h d)")
|
356 |
+
# normalize
|
357 |
+
output = self.norm(output)
|
358 |
+
# gate
|
359 |
+
output = u * output
|
360 |
+
# outproj
|
361 |
+
output = self.out_proj(output)
|
362 |
+
|
363 |
+
attn_weights = None
|
364 |
+
|
365 |
+
return output, attn_weights, kv
|
366 |
+
|
367 |
+
|
368 |
+
class TransnormerDecoderLayer(nn.Module):
|
369 |
+
|
370 |
+
def __init__(self, config: TransnormerConfig):
|
371 |
+
super().__init__()
|
372 |
+
self.embed_dim = config.decoder_embed_dim
|
373 |
+
##### normalize
|
374 |
+
norm_type = config.norm_type
|
375 |
+
if debug:
|
376 |
+
logging_info(f"Decoder Norm Type: {norm_type}")
|
377 |
+
self.token_norm = get_norm_fn(norm_type)(self.embed_dim)
|
378 |
+
self.channel_norm = get_norm_fn(norm_type)(self.embed_dim)
|
379 |
+
|
380 |
+
##### token mixer
|
381 |
+
self.token_mixer = self.build_token_mixer(
|
382 |
+
self.embed_dim,
|
383 |
+
config,
|
384 |
+
)
|
385 |
+
|
386 |
+
##### channel mixer
|
387 |
+
self.glu_dim = config.glu_dim
|
388 |
+
if self.glu_dim == -1:
|
389 |
+
self.glu_dim = self.embed_dim
|
390 |
+
bias = config.bias
|
391 |
+
self.channel_mixer = GLU(self.embed_dim, self.glu_dim, bias)
|
392 |
+
|
393 |
+
def build_token_mixer(self, embed_dim, config):
|
394 |
+
return NormLinearAttention(
|
395 |
+
embed_dim=embed_dim,
|
396 |
+
hidden_dim=config.hidden_dim,
|
397 |
+
num_heads=config.decoder_attention_heads,
|
398 |
+
linear_act_fun=config.linear_act_fun,
|
399 |
+
norm_type=config.norm_type,
|
400 |
+
linear_use_lrpe=config.linear_use_lrpe,
|
401 |
+
bias=config.bias,
|
402 |
+
)
|
403 |
+
|
404 |
+
def residual_connection(self, x, residual):
|
405 |
+
return residual + x
|
406 |
+
|
407 |
+
def forward(
|
408 |
+
self,
|
409 |
+
x,
|
410 |
+
attn_mask: Optional[torch.Tensor] = None,
|
411 |
+
attn_padding_mask: Optional[torch.Tensor] = None,
|
412 |
+
past_key_value: Optional[Tuple[torch.Tensor]] = None,
|
413 |
+
output_attentions: Optional[bool] = False,
|
414 |
+
use_cache: Optional[bool] = False,
|
415 |
+
slope_rate: Optional[torch.Tensor] = None, # (h, 1, 1)
|
416 |
+
):
|
417 |
+
residual = x
|
418 |
+
x = self.token_norm(x)
|
419 |
+
x, self_attn_weights, present_key_value = self.token_mixer(
|
420 |
+
x=x,
|
421 |
+
attn_mask=attn_mask,
|
422 |
+
attn_padding_mask=attn_padding_mask,
|
423 |
+
past_key_value=past_key_value,
|
424 |
+
output_attentions=output_attentions,
|
425 |
+
use_cache=use_cache,
|
426 |
+
slope_rate=slope_rate,
|
427 |
+
)
|
428 |
+
x = self.residual_connection(x, residual)
|
429 |
+
|
430 |
+
residual = x
|
431 |
+
x = self.channel_norm(x)
|
432 |
+
x = self.channel_mixer(x)
|
433 |
+
x = self.residual_connection(x, residual)
|
434 |
+
|
435 |
+
outputs = (x, )
|
436 |
+
|
437 |
+
if output_attentions:
|
438 |
+
outputs += (self_attn_weights, )
|
439 |
+
|
440 |
+
if use_cache:
|
441 |
+
outputs += (present_key_value, )
|
442 |
+
|
443 |
+
return outputs
|
444 |
+
|
445 |
+
|
446 |
+
TRANSNORMER_START_DOCSTRING = r"""
|
447 |
+
This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
|
448 |
+
library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
|
449 |
+
etc.)
|
450 |
+
|
451 |
+
This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
|
452 |
+
Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
|
453 |
+
and behavior.
|
454 |
+
|
455 |
+
Parameters:
|
456 |
+
config ([`TransnormerConfig`]):
|
457 |
+
Model configuration class with all the parameters of the model. Initializing with a config file does not
|
458 |
+
load the weights associated with the model, only the configuration. Check out the
|
459 |
+
[`~PreTrainedModel.from_pretrained`] method to load the model weights.
|
460 |
+
"""
|
461 |
+
|
462 |
+
|
463 |
+
@add_start_docstrings(TRANSNORMER_START_DOCSTRING, )
|
464 |
+
class TransnormerPreTrainedModel(PreTrainedModel):
|
465 |
+
config_class = TransnormerConfig
|
466 |
+
base_model_prefix = "model"
|
467 |
+
supports_gradient_checkpointing = True
|
468 |
+
_no_split_modules = ["TransnormerDecoderLayer"]
|
469 |
+
_skip_keys_device_placement = "past_key_values"
|
470 |
+
_keys_to_ignore_on_load_unexpected = [r"decoder\.version"]
|
471 |
+
|
472 |
+
def _init_weights(self, module):
|
473 |
+
std = self.config.init_std
|
474 |
+
if isinstance(module, nn.Linear):
|
475 |
+
module.weight.data.normal_(mean=0.0, std=std)
|
476 |
+
if module.bias is not None:
|
477 |
+
module.bias.data.zero_()
|
478 |
+
elif isinstance(module, nn.Embedding):
|
479 |
+
module.weight.data.normal_(mean=0.0, std=std)
|
480 |
+
if module.padding_idx is not None:
|
481 |
+
module.weight.data[module.padding_idx].zero_()
|
482 |
+
|
483 |
+
def _set_gradient_checkpointing(self, module, value=False):
|
484 |
+
if isinstance(module, TransnormerModel):
|
485 |
+
module.gradient_checkpointing = value
|
486 |
+
|
487 |
+
|
488 |
+
TRANSNORMER_INPUTS_DOCSTRING = r"""
|
489 |
+
Args:
|
490 |
+
input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
|
491 |
+
Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
|
492 |
+
it.
|
493 |
+
|
494 |
+
Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
|
495 |
+
[`PreTrainedTokenizer.__call__`] for details.
|
496 |
+
|
497 |
+
[What are input IDs?](../glossary#input-ids)
|
498 |
+
attn_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
|
499 |
+
Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
|
500 |
+
|
501 |
+
- 1 for tokens that are **not masked**,
|
502 |
+
- 0 for tokens that are **masked**.
|
503 |
+
|
504 |
+
[What are attention masks?](../glossary#attention-mask)
|
505 |
+
|
506 |
+
Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
|
507 |
+
[`PreTrainedTokenizer.__call__`] for details.
|
508 |
+
|
509 |
+
If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
|
510 |
+
`past_key_values`).
|
511 |
+
|
512 |
+
If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attn_mask`]
|
513 |
+
and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
|
514 |
+
information on the default strategy.
|
515 |
+
|
516 |
+
- 1 indicates the head is **not masked**,
|
517 |
+
- 0 indicates the head is **masked**.
|
518 |
+
position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
519 |
+
Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
|
520 |
+
config.n_positions - 1]`.
|
521 |
+
|
522 |
+
[What are position IDs?](../glossary#position-ids)
|
523 |
+
past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
|
524 |
+
Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
|
525 |
+
`(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape
|
526 |
+
`(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
|
527 |
+
|
528 |
+
Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
|
529 |
+
blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
|
530 |
+
|
531 |
+
If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
|
532 |
+
don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
|
533 |
+
`decoder_input_ids` of shape `(batch_size, sequence_length)`.
|
534 |
+
use_cache (`bool`, *optional*):
|
535 |
+
If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
|
536 |
+
`past_key_values`).
|
537 |
+
output_attentions (`bool`, *optional*):
|
538 |
+
Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
|
539 |
+
tensors for more detail.
|
540 |
+
output_hidden_states (`bool`, *optional*):
|
541 |
+
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
|
542 |
+
more detail.
|
543 |
+
return_dict (`bool`, *optional*):
|
544 |
+
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
|
545 |
+
"""
|
546 |
+
|
547 |
+
|
548 |
+
@add_start_docstrings(TRANSNORMER_START_DOCSTRING, )
|
549 |
+
class TransnormerModel(TransnormerPreTrainedModel):
|
550 |
+
"""
|
551 |
+
Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`TransnormerDecoderLayer`]
|
552 |
+
|
553 |
+
Args:
|
554 |
+
config: TransnormerConfig
|
555 |
+
"""
|
556 |
+
|
557 |
+
def __init__(self, config: TransnormerConfig):
|
558 |
+
super().__init__(config)
|
559 |
+
# hf origin
|
560 |
+
self.padding_idx = config.pad_token_id
|
561 |
+
self.vocab_size = config.vocab_size
|
562 |
+
self.gradient_checkpointing = False
|
563 |
+
# mask
|
564 |
+
self._linear_attn_mask = torch.empty(0)
|
565 |
+
# config
|
566 |
+
self.linear_use_lrpe_list = config.linear_use_lrpe_list
|
567 |
+
self.num_layers = config.decoder_layers
|
568 |
+
# h, 1, 1
|
569 |
+
self.slopes = self._build_slope_tensor(config.decoder_attention_heads)
|
570 |
+
|
571 |
+
# params
|
572 |
+
self.embed_tokens = nn.Embedding(config.vocab_size,
|
573 |
+
config.decoder_embed_dim,
|
574 |
+
self.padding_idx)
|
575 |
+
self.layers = nn.ModuleList([])
|
576 |
+
for i in range(config.decoder_layers):
|
577 |
+
if len(self.linear_use_lrpe_list) > 0:
|
578 |
+
config.linear_use_lrpe = self.linear_use_lrpe_list[i]
|
579 |
+
self.layers.append(TransnormerDecoderLayer(config))
|
580 |
+
|
581 |
+
self.final_norm = get_norm_fn(config.norm_type)(
|
582 |
+
config.decoder_embed_dim)
|
583 |
+
self.embed_dim = config.decoder_embed_dim
|
584 |
+
self.embed_scale = (1.0 if config.no_scale_embedding else math.sqrt(
|
585 |
+
self.embed_dim))
|
586 |
+
|
587 |
+
# Initialize weights and apply final processing
|
588 |
+
self.post_init()
|
589 |
+
|
590 |
+
@staticmethod
|
591 |
+
def _build_slope_tensor(n_attention_heads: int):
|
592 |
+
|
593 |
+
def get_slopes(n):
|
594 |
+
|
595 |
+
def get_slopes_power_of_2(n):
|
596 |
+
start = 2**(-(2**-(math.log2(n) - 3)))
|
597 |
+
ratio = start
|
598 |
+
return [start * ratio**i for i in range(n)]
|
599 |
+
|
600 |
+
if math.log2(n).is_integer():
|
601 |
+
return get_slopes_power_of_2(
|
602 |
+
n
|
603 |
+
) # In the paper, we only train models that have 2^a heads for some a. This function has
|
604 |
+
else: # some good properties that only occur when the input is a power of 2. To maintain that even
|
605 |
+
closest_power_of_2 = 2**math.floor(
|
606 |
+
math.log2(n)
|
607 |
+
) # when the number of heads is not a power of 2, we use this workaround.
|
608 |
+
return (get_slopes_power_of_2(closest_power_of_2) + get_slopes(
|
609 |
+
2 * closest_power_of_2)[0::2][:n - closest_power_of_2])
|
610 |
+
|
611 |
+
# h, 1, 1
|
612 |
+
slopes = torch.tensor(get_slopes(n_attention_heads)).reshape(
|
613 |
+
n_attention_heads, 1, 1)
|
614 |
+
|
615 |
+
return slopes
|
616 |
+
|
617 |
+
def extra_repr(self):
|
618 |
+
return print_module(self)
|
619 |
+
|
620 |
+
def get_input_embeddings(self):
|
621 |
+
return self.embed_tokens
|
622 |
+
|
623 |
+
def set_input_embeddings(self, value):
|
624 |
+
self.embed_tokens = value
|
625 |
+
|
626 |
+
def _prepare_decoder_linear_attn_mask(self, input_shape, inputs_embeds,
|
627 |
+
past_key_values_length):
|
628 |
+
bsz, tgt_len = input_shape
|
629 |
+
src_len = tgt_len + past_key_values_length
|
630 |
+
|
631 |
+
def power_log(x):
|
632 |
+
return 2**(math.ceil(math.log(x, 2)))
|
633 |
+
|
634 |
+
n = power_log(max(tgt_len, src_len))
|
635 |
+
if self._linear_attn_mask.shape[-1] < n:
|
636 |
+
|
637 |
+
def get_mask(n):
|
638 |
+
mask = torch.triu(
|
639 |
+
torch.zeros(n, n).float().fill_(float("-inf")), 1)
|
640 |
+
# no slope version
|
641 |
+
# -n, ..., -2, -1, 0
|
642 |
+
for i in range(n):
|
643 |
+
x = torch.arange(i + 1)
|
644 |
+
y = x
|
645 |
+
mask[i, :i + 1] = -torch.flip(y, [0])
|
646 |
+
|
647 |
+
return mask
|
648 |
+
|
649 |
+
arr = []
|
650 |
+
for slope in self.slopes:
|
651 |
+
arr.append(get_mask(n))
|
652 |
+
self._linear_attn_mask = torch.stack(arr, dim=0).to(inputs_embeds)
|
653 |
+
|
654 |
+
linear_attn_mask = self._linear_attn_mask[:, -tgt_len:, -src_len:]
|
655 |
+
num_heads = linear_attn_mask.shape[0]
|
656 |
+
|
657 |
+
return linear_attn_mask[None, :, :, :].expand(bsz, num_heads, tgt_len,
|
658 |
+
src_len)
|
659 |
+
|
660 |
+
@add_start_docstrings_to_model_forward(TRANSNORMER_INPUTS_DOCSTRING)
|
661 |
+
def forward(
|
662 |
+
self,
|
663 |
+
input_ids: torch.LongTensor = None,
|
664 |
+
attn_padding_mask: Optional[torch.Tensor] = None,
|
665 |
+
past_key_values: Optional[List[torch.FloatTensor]] = None,
|
666 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
667 |
+
use_cache: Optional[bool] = None,
|
668 |
+
output_attentions: Optional[bool] = None,
|
669 |
+
output_hidden_states: Optional[bool] = None,
|
670 |
+
return_dict: Optional[bool] = None,
|
671 |
+
) -> Union[Tuple, BaseModelOutputWithPast]:
|
672 |
+
output_attentions = (output_attentions if output_attentions is not None
|
673 |
+
else self.config.output_attentions)
|
674 |
+
output_hidden_states = (output_hidden_states
|
675 |
+
if output_hidden_states is not None else
|
676 |
+
self.config.output_hidden_states)
|
677 |
+
use_cache = use_cache if use_cache is not None else self.config.use_cache
|
678 |
+
|
679 |
+
return_dict = (return_dict if return_dict is not None else
|
680 |
+
self.config.use_return_dict)
|
681 |
+
|
682 |
+
# retrieve input_ids and inputs_embeds
|
683 |
+
if input_ids is not None and inputs_embeds is not None:
|
684 |
+
raise ValueError(
|
685 |
+
"You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time"
|
686 |
+
)
|
687 |
+
elif input_ids is not None:
|
688 |
+
batch_size, seq_length = input_ids.shape
|
689 |
+
elif inputs_embeds is not None:
|
690 |
+
batch_size, seq_length, _ = inputs_embeds.shape
|
691 |
+
else:
|
692 |
+
raise ValueError(
|
693 |
+
"You have to specify either decoder_input_ids or decoder_inputs_embeds"
|
694 |
+
)
|
695 |
+
|
696 |
+
seq_length_with_past = seq_length
|
697 |
+
past_key_values_length = 0
|
698 |
+
|
699 |
+
if past_key_values is not None:
|
700 |
+
past_key_values_length = past_key_values[0][0].shape[-2]
|
701 |
+
seq_length_with_past = seq_length_with_past + past_key_values_length
|
702 |
+
|
703 |
+
if inputs_embeds is None:
|
704 |
+
# !!! use embed_scale
|
705 |
+
inputs_embeds = self.embed_scale * self.embed_tokens(input_ids)
|
706 |
+
|
707 |
+
hidden_states = inputs_embeds
|
708 |
+
|
709 |
+
if self.gradient_checkpointing and self.training:
|
710 |
+
if use_cache:
|
711 |
+
logger.warning_once(
|
712 |
+
"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
|
713 |
+
)
|
714 |
+
use_cache = False
|
715 |
+
|
716 |
+
# decoder layers
|
717 |
+
all_hidden_states = () if output_hidden_states else None
|
718 |
+
all_self_attns = () if output_attentions else None
|
719 |
+
next_decoder_cache = () if use_cache else None
|
720 |
+
|
721 |
+
##### norm linear layers
|
722 |
+
linear_attn_padding_mask = attn_padding_mask
|
723 |
+
linear_attn_mask = self._prepare_decoder_linear_attn_mask(
|
724 |
+
(batch_size, seq_length), inputs_embeds, past_key_values_length)
|
725 |
+
|
726 |
+
slope_rates = [
|
727 |
+
self.slopes.to(input_ids.device) for _ in range(self.num_layers)
|
728 |
+
]
|
729 |
+
|
730 |
+
for idx, layer in enumerate(self.layers):
|
731 |
+
if output_hidden_states:
|
732 |
+
all_hidden_states += (hidden_states, )
|
733 |
+
|
734 |
+
past_key_value = (past_key_values[idx]
|
735 |
+
if past_key_values is not None else None)
|
736 |
+
|
737 |
+
slope_rate = slope_rates[idx]
|
738 |
+
slope_rate = slope_rate * (1 - idx / (self.num_layers - 1) + 1e-5)
|
739 |
+
mask = linear_attn_mask
|
740 |
+
|
741 |
+
layer_outputs = layer(
|
742 |
+
hidden_states,
|
743 |
+
attn_mask=mask,
|
744 |
+
attn_padding_mask=linear_attn_padding_mask,
|
745 |
+
past_key_value=past_key_value,
|
746 |
+
output_attentions=output_attentions,
|
747 |
+
use_cache=use_cache,
|
748 |
+
slope_rate=slope_rate,
|
749 |
+
)
|
750 |
+
|
751 |
+
hidden_states = layer_outputs[0]
|
752 |
+
|
753 |
+
if use_cache:
|
754 |
+
next_decoder_cache += (
|
755 |
+
layer_outputs[2 if output_attentions else 1], )
|
756 |
+
|
757 |
+
if output_attentions:
|
758 |
+
all_self_attns += (layer_outputs[1], )
|
759 |
+
|
760 |
+
hidden_states = self.final_norm(hidden_states)
|
761 |
+
|
762 |
+
# add hidden states from the last decoder layer
|
763 |
+
if output_hidden_states:
|
764 |
+
all_hidden_states += (hidden_states, )
|
765 |
+
|
766 |
+
next_cache = next_decoder_cache if use_cache else None
|
767 |
+
if not return_dict:
|
768 |
+
return tuple(
|
769 |
+
v for v in
|
770 |
+
[hidden_states, next_cache, all_hidden_states, all_self_attns]
|
771 |
+
if v is not None)
|
772 |
+
return BaseModelOutputWithPast(
|
773 |
+
last_hidden_state=hidden_states,
|
774 |
+
past_key_values=next_cache,
|
775 |
+
hidden_states=all_hidden_states,
|
776 |
+
attentions=all_self_attns,
|
777 |
+
)
|
778 |
+
|
779 |
+
|
780 |
+
class TransnormerForCausalLM(TransnormerPreTrainedModel):
|
781 |
+
|
782 |
+
def __init__(self, config):
|
783 |
+
super().__init__(config)
|
784 |
+
self.model = TransnormerModel(config)
|
785 |
+
if debug:
|
786 |
+
logging_info(self.model)
|
787 |
+
|
788 |
+
# the lm_head weight is automatically tied to the embed tokens weight
|
789 |
+
self.lm_head = nn.Linear(config.decoder_embed_dim,
|
790 |
+
config.vocab_size,
|
791 |
+
bias=False)
|
792 |
+
|
793 |
+
# Initialize weights and apply final processing
|
794 |
+
self.post_init()
|
795 |
+
|
796 |
+
def get_input_embeddings(self):
|
797 |
+
return self.model.embed_tokens
|
798 |
+
|
799 |
+
def set_input_embeddings(self, value):
|
800 |
+
self.model.embed_tokens = value
|
801 |
+
|
802 |
+
def get_output_embeddings(self):
|
803 |
+
return self.lm_head
|
804 |
+
|
805 |
+
def set_output_embeddings(self, new_embeddings):
|
806 |
+
self.lm_head = new_embeddings
|
807 |
+
|
808 |
+
def set_decoder(self, decoder):
|
809 |
+
self.model = decoder
|
810 |
+
|
811 |
+
def get_decoder(self):
|
812 |
+
return self.model
|
813 |
+
|
814 |
+
@add_start_docstrings_to_model_forward(TRANSNORMER_INPUTS_DOCSTRING)
|
815 |
+
@replace_return_docstrings(output_type=CausalLMOutputWithPast,
|
816 |
+
config_class=_CONFIG_FOR_DOC)
|
817 |
+
def forward(
|
818 |
+
self,
|
819 |
+
input_ids: torch.LongTensor = None,
|
820 |
+
attention_mask: Optional[torch.Tensor] = None,
|
821 |
+
past_key_values: Optional[List[torch.FloatTensor]] = None,
|
822 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
823 |
+
labels: Optional[torch.LongTensor] = None,
|
824 |
+
use_cache: Optional[bool] = None,
|
825 |
+
output_attentions: Optional[bool] = None,
|
826 |
+
output_hidden_states: Optional[bool] = None,
|
827 |
+
return_dict: Optional[bool] = None,
|
828 |
+
) -> Union[Tuple, CausalLMOutputWithPast]:
|
829 |
+
r"""
|
830 |
+
Args:
|
831 |
+
labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
832 |
+
Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
|
833 |
+
config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
|
834 |
+
(masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
|
835 |
+
|
836 |
+
Returns:
|
837 |
+
|
838 |
+
Example:
|
839 |
+
|
840 |
+
```python
|
841 |
+
>>> from transformers import AutoTokenizer, TransnormerForCausalLM
|
842 |
+
|
843 |
+
>>> model = TransnormerForCausalLM.from_pretrained(PATH_TO_CONVERTED_WEIGHTS)
|
844 |
+
>>> tokenizer = AutoTokenizer.from_pretrained(PATH_TO_CONVERTED_TOKENIZER)
|
845 |
+
|
846 |
+
>>> prompt = "Hey, are you consciours? Can you talk to me?"
|
847 |
+
>>> inputs = tokenizer(prompt, return_tensors="pt")
|
848 |
+
|
849 |
+
>>> # Generate
|
850 |
+
>>> generate_ids = model.generate(inputs.input_ids, max_length=30)
|
851 |
+
>>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
|
852 |
+
"Hey, are you consciours? Can you talk to me?\nI'm not consciours, but I can talk to you."
|
853 |
+
```"""
|
854 |
+
output_attentions = (output_attentions if output_attentions is not None
|
855 |
+
else self.config.output_attentions)
|
856 |
+
output_hidden_states = (output_hidden_states
|
857 |
+
if output_hidden_states is not None else
|
858 |
+
self.config.output_hidden_states)
|
859 |
+
return_dict = (return_dict if return_dict is not None else
|
860 |
+
self.config.use_return_dict)
|
861 |
+
|
862 |
+
# decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
|
863 |
+
outputs = self.model(
|
864 |
+
input_ids=input_ids,
|
865 |
+
attn_padding_mask=attention_mask,
|
866 |
+
past_key_values=past_key_values,
|
867 |
+
inputs_embeds=inputs_embeds,
|
868 |
+
use_cache=use_cache,
|
869 |
+
output_attentions=output_attentions,
|
870 |
+
output_hidden_states=output_hidden_states,
|
871 |
+
return_dict=return_dict,
|
872 |
+
)
|
873 |
+
|
874 |
+
hidden_states = outputs[0]
|
875 |
+
logits = self.lm_head(hidden_states)
|
876 |
+
|
877 |
+
loss = None
|
878 |
+
if labels is not None:
|
879 |
+
# Shift so that tokens < n predict n
|
880 |
+
shift_logits = logits[..., :-1, :].contiguous()
|
881 |
+
shift_labels = labels[..., 1:].contiguous()
|
882 |
+
# Flatten the tokens
|
883 |
+
loss_fct = CrossEntropyLoss()
|
884 |
+
shift_logits = shift_logits.view(-1, self.config.vocab_size)
|
885 |
+
shift_labels = shift_labels.view(-1)
|
886 |
+
# Enable model parallelism
|
887 |
+
shift_labels = shift_labels.to(shift_logits.device)
|
888 |
+
loss = loss_fct(shift_logits, shift_labels)
|
889 |
+
|
890 |
+
if not return_dict:
|
891 |
+
output = (logits, ) + outputs[1:]
|
892 |
+
return (loss, ) + output if loss is not None else output
|
893 |
+
|
894 |
+
return CausalLMOutputWithPast(
|
895 |
+
loss=loss,
|
896 |
+
logits=logits,
|
897 |
+
past_key_values=outputs.past_key_values,
|
898 |
+
hidden_states=outputs.hidden_states,
|
899 |
+
attentions=outputs.attentions,
|
900 |
+
)
|
901 |
+
|
902 |
+
def prepare_inputs_for_generation(
|
903 |
+
self,
|
904 |
+
input_ids,
|
905 |
+
past_key_values=None,
|
906 |
+
attention_mask=None,
|
907 |
+
inputs_embeds=None,
|
908 |
+
**kwargs,
|
909 |
+
):
|
910 |
+
if past_key_values:
|
911 |
+
input_ids = input_ids[:, -1:]
|
912 |
+
|
913 |
+
# if `inputs_embeds` are passed, we only want to use them in the 1st generation step
|
914 |
+
if inputs_embeds is not None and past_key_values is None:
|
915 |
+
model_inputs = {"inputs_embeds": inputs_embeds}
|
916 |
+
else:
|
917 |
+
model_inputs = {"input_ids": input_ids}
|
918 |
+
|
919 |
+
model_inputs.update({
|
920 |
+
"past_key_values": past_key_values,
|
921 |
+
"use_cache": kwargs.get("use_cache"),
|
922 |
+
"attention_mask": attention_mask,
|
923 |
+
})
|
924 |
+
return model_inputs
|
925 |
+
|
926 |
+
@staticmethod
|
927 |
+
def _reorder_cache(past_key_values, beam_idx):
|
928 |
+
reordered_past = ()
|
929 |
+
for layer_past in past_key_values:
|
930 |
+
reordered_past += (tuple(
|
931 |
+
past_state.index_select(0, beam_idx)
|
932 |
+
for past_state in layer_past), )
|
933 |
+
return reordered_past
|
norm.py
ADDED
@@ -0,0 +1,43 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright 2023 OpenNLPLab
|
2 |
+
#
|
3 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
4 |
+
# you may not use this file except in compliance with the License.
|
5 |
+
# You may obtain a copy of the License at
|
6 |
+
#
|
7 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
8 |
+
#
|
9 |
+
# Unless required by applicable law or agreed to in writing, software
|
10 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
11 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
12 |
+
# See the License for the specific language governing permissions and
|
13 |
+
# limitations under the License.
|
14 |
+
|
15 |
+
import logging
|
16 |
+
import os
|
17 |
+
import sys
|
18 |
+
|
19 |
+
import torch
|
20 |
+
from torch import nn
|
21 |
+
|
22 |
+
logging.basicConfig(
|
23 |
+
format="%(asctime)s | %(levelname)s | %(name)s | %(message)s",
|
24 |
+
datefmt="%Y-%m-%d %H:%M:%S",
|
25 |
+
level=os.environ.get("LOGLEVEL", "INFO").upper(),
|
26 |
+
stream=sys.stdout,
|
27 |
+
)
|
28 |
+
logger = logging.getLogger("srmsnorm")
|
29 |
+
|
30 |
+
|
31 |
+
class SimpleRMSNorm(nn.Module):
|
32 |
+
|
33 |
+
def __init__(self, dim: int, eps: float = 1e-6):
|
34 |
+
super().__init__()
|
35 |
+
self.eps = eps
|
36 |
+
|
37 |
+
def _norm(self, x):
|
38 |
+
return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
|
39 |
+
|
40 |
+
def forward(self, x):
|
41 |
+
output = self._norm(x.float()).type_as(x)
|
42 |
+
|
43 |
+
return output
|
special_tokens_map copy.json
ADDED
@@ -0,0 +1,23 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"bos_token": {
|
3 |
+
"content": "<|endoftext|>",
|
4 |
+
"lstrip": false,
|
5 |
+
"normalized": true,
|
6 |
+
"rstrip": false,
|
7 |
+
"single_word": false
|
8 |
+
},
|
9 |
+
"eos_token": {
|
10 |
+
"content": "<|endoftext|>",
|
11 |
+
"lstrip": false,
|
12 |
+
"normalized": true,
|
13 |
+
"rstrip": false,
|
14 |
+
"single_word": false
|
15 |
+
},
|
16 |
+
"unk_token": {
|
17 |
+
"content": "<|endoftext|>",
|
18 |
+
"lstrip": false,
|
19 |
+
"normalized": true,
|
20 |
+
"rstrip": false,
|
21 |
+
"single_word": false
|
22 |
+
}
|
23 |
+
}
|
special_tokens_map.json
ADDED
@@ -0,0 +1,23 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"bos_token": {
|
3 |
+
"content": "<s>",
|
4 |
+
"lstrip": false,
|
5 |
+
"normalized": true,
|
6 |
+
"rstrip": false,
|
7 |
+
"single_word": false
|
8 |
+
},
|
9 |
+
"eos_token": {
|
10 |
+
"content": "</s>",
|
11 |
+
"lstrip": false,
|
12 |
+
"normalized": true,
|
13 |
+
"rstrip": false,
|
14 |
+
"single_word": false
|
15 |
+
},
|
16 |
+
"unk_token": {
|
17 |
+
"content": "<unk>",
|
18 |
+
"lstrip": false,
|
19 |
+
"normalized": true,
|
20 |
+
"rstrip": false,
|
21 |
+
"single_word": false
|
22 |
+
}
|
23 |
+
}
|
srmsnorm_triton.py
ADDED
@@ -0,0 +1,201 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# CREDITS: This comes almost as-is from the Triton layer norm tutorial
|
2 |
+
# https://github.com/openai/triton/blob/master/python/tutorials/05-layer-norm.py
|
3 |
+
# Copyright 2023 OpenNLPLab
|
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 |
+
import torch
|
18 |
+
import torch.nn.functional as F
|
19 |
+
import triton
|
20 |
+
import triton.language as tl
|
21 |
+
|
22 |
+
|
23 |
+
# fmt: off
|
24 |
+
@triton.jit
|
25 |
+
def srms_norm_fw(X, Y, V, stride, N, eps, BLOCK_SIZE_N: tl.constexpr):
|
26 |
+
# fmt: on
|
27 |
+
row = tl.program_id(0)
|
28 |
+
cols = tl.arange(0, BLOCK_SIZE_N)
|
29 |
+
mask = cols < N
|
30 |
+
|
31 |
+
# Move to this row
|
32 |
+
x_ptrs = X + row * stride + cols
|
33 |
+
x = tl.load(x_ptrs, mask=mask, other=0.0).to(tl.float32)
|
34 |
+
|
35 |
+
x_zm = tl.where(mask, x, 0.0)
|
36 |
+
|
37 |
+
x_var = tl.sum(x_zm * x_zm, axis=0) / N
|
38 |
+
rstd = 1.0 / tl.sqrt(x_var + eps)
|
39 |
+
|
40 |
+
# Normalize, optionally affine
|
41 |
+
y = x_zm * rstd
|
42 |
+
tl.store(V + row, rstd)
|
43 |
+
|
44 |
+
y_ptrs = Y + row * stride + cols
|
45 |
+
tl.store(y_ptrs, y, mask=mask)
|
46 |
+
|
47 |
+
|
48 |
+
# Backward pass (DX + partial DW + partial DB)
|
49 |
+
# fmt: off
|
50 |
+
@triton.jit
|
51 |
+
def srms_norm_bwd_dx_fused(
|
52 |
+
DX, DY,
|
53 |
+
X, V,
|
54 |
+
stride, N,
|
55 |
+
# META-parameters
|
56 |
+
BLOCK_SIZE_N: tl.constexpr,
|
57 |
+
):
|
58 |
+
# fmt: on
|
59 |
+
|
60 |
+
# position of elements processed by this program
|
61 |
+
row = tl.program_id(0)
|
62 |
+
cols = tl.arange(0, BLOCK_SIZE_N)
|
63 |
+
mask = cols < N
|
64 |
+
|
65 |
+
# offset data pointers to start at the row of interest
|
66 |
+
x_ptrs = X + row * stride + cols
|
67 |
+
dy_ptrs = DY + row * stride + cols
|
68 |
+
|
69 |
+
# load data to SRAM
|
70 |
+
x = tl.load(x_ptrs, mask=mask, other=0)
|
71 |
+
dy = tl.load(dy_ptrs, mask=mask, other=0)
|
72 |
+
rstd = tl.load(V + row)
|
73 |
+
|
74 |
+
# compute dx
|
75 |
+
xhat = x * rstd
|
76 |
+
wdy = dy
|
77 |
+
|
78 |
+
xhat = tl.where(mask, xhat, 0.)
|
79 |
+
wdy = tl.where(mask, wdy, 0.)
|
80 |
+
mean1 = tl.sum(xhat * wdy, axis=0) / N
|
81 |
+
dx = (wdy - (xhat * mean1)) * rstd
|
82 |
+
|
83 |
+
# write-back dx
|
84 |
+
mask = cols < N # re-materialize the mask to save registers
|
85 |
+
dx_ptrs = DX + row * stride + cols
|
86 |
+
tl.store(dx_ptrs, dx, mask=mask)
|
87 |
+
|
88 |
+
|
89 |
+
class _SrmsNorm(torch.autograd.Function):
|
90 |
+
|
91 |
+
@staticmethod
|
92 |
+
def forward(ctx, x, eps):
|
93 |
+
# catch eps being too small if the tensors are fp16
|
94 |
+
if x.dtype == torch.float16:
|
95 |
+
eps = max(eps, 1.6e-5)
|
96 |
+
|
97 |
+
# allocate output
|
98 |
+
y = torch.empty_like(x)
|
99 |
+
|
100 |
+
# reshape input data into 2D tensor
|
101 |
+
x_arg = x.reshape(-1, x.shape[-1])
|
102 |
+
M, N = x_arg.shape
|
103 |
+
|
104 |
+
# allocate mean and std, they'll be used in the backward pass
|
105 |
+
rstd = torch.empty((M, ), dtype=torch.float32, device=x.device)
|
106 |
+
|
107 |
+
# Less than 64KB per feature: enqueue fused kernel
|
108 |
+
MAX_FUSED_SIZE = 65536 // x.element_size()
|
109 |
+
BLOCK_SIZE_N = min(MAX_FUSED_SIZE, triton.next_power_of_2(N))
|
110 |
+
if N > BLOCK_SIZE_N:
|
111 |
+
raise RuntimeError(
|
112 |
+
"This layer norm doesn't support feature dim >= 64KB.")
|
113 |
+
|
114 |
+
if not x_arg.is_contiguous() or not y.is_contiguous():
|
115 |
+
x_arg = x_arg.contiguous()
|
116 |
+
y = y.contiguous()
|
117 |
+
|
118 |
+
# heuristics for number of warps.
|
119 |
+
num_warps = min(max(BLOCK_SIZE_N // 256, 1), 16)
|
120 |
+
|
121 |
+
# enqueue kernel
|
122 |
+
# fmt: off
|
123 |
+
srms_norm_fw[(M,)](
|
124 |
+
x_arg, y, rstd,
|
125 |
+
x_arg.stride(0),
|
126 |
+
N,
|
127 |
+
eps,
|
128 |
+
num_warps=num_warps,
|
129 |
+
BLOCK_SIZE_N=BLOCK_SIZE_N,
|
130 |
+
)
|
131 |
+
# fmt: on
|
132 |
+
|
133 |
+
ctx.save_for_backward(x, rstd)
|
134 |
+
ctx.BLOCK_SIZE_N = BLOCK_SIZE_N
|
135 |
+
ctx.num_warps = num_warps
|
136 |
+
|
137 |
+
return y.reshape_as(x)
|
138 |
+
|
139 |
+
@staticmethod
|
140 |
+
def backward(
|
141 |
+
ctx, dy
|
142 |
+
): # pragma: no cover # this is covered, but called directly from C++
|
143 |
+
x, rstd = ctx.saved_tensors
|
144 |
+
|
145 |
+
# flatten the batch dimension, if any.
|
146 |
+
# We're interested in 'samples' x norm_dimension
|
147 |
+
x = x.reshape(-1, x.size(-1))
|
148 |
+
M, N = x.size()
|
149 |
+
|
150 |
+
# heuristics for amount of parallel reduction stream for DG/DB
|
151 |
+
GROUP_SIZE_M = 32
|
152 |
+
if N <= 8192:
|
153 |
+
GROUP_SIZE_M = 64
|
154 |
+
if N <= 4096:
|
155 |
+
GROUP_SIZE_M = 96
|
156 |
+
if N <= 2048:
|
157 |
+
GROUP_SIZE_M = 128
|
158 |
+
if N <= 1024:
|
159 |
+
GROUP_SIZE_M = 256
|
160 |
+
|
161 |
+
if dy.dtype == torch.float32:
|
162 |
+
GROUP_SIZE_M = GROUP_SIZE_M // 2
|
163 |
+
|
164 |
+
# allocate output
|
165 |
+
dy = dy.contiguous()
|
166 |
+
dx = torch.empty_like(dy)
|
167 |
+
|
168 |
+
# Check the tensor shapes and layouts
|
169 |
+
# we suppose in the kernel that they have the same size and are contiguous
|
170 |
+
assert (
|
171 |
+
dy.numel() == x.numel()
|
172 |
+
), "Something is wrong in the backward graph, possibly because of an inplace operation after the layernorm"
|
173 |
+
|
174 |
+
# enqueue kernel using forward pass heuristics
|
175 |
+
# also compute partial sums for DW and DB
|
176 |
+
num_warps = min(max(ctx.BLOCK_SIZE_N // 256, 1), 16)
|
177 |
+
|
178 |
+
# fmt: off
|
179 |
+
srms_norm_bwd_dx_fused[(M,)](
|
180 |
+
dx, dy, x,
|
181 |
+
rstd,
|
182 |
+
x.stride(0),
|
183 |
+
N,
|
184 |
+
BLOCK_SIZE_N=ctx.BLOCK_SIZE_N,
|
185 |
+
num_warps=num_warps
|
186 |
+
)
|
187 |
+
# fmt: on
|
188 |
+
|
189 |
+
dx = dx.reshape_as(dy)
|
190 |
+
return dx, None, None
|
191 |
+
|
192 |
+
|
193 |
+
class SimpleRMSNorm(torch.nn.Module):
|
194 |
+
|
195 |
+
def __init__(self, dim: int, eps: float = 1e-6):
|
196 |
+
super().__init__()
|
197 |
+
self.eps = eps
|
198 |
+
self.dim = dim
|
199 |
+
|
200 |
+
def forward(self, x):
|
201 |
+
return _SrmsNorm.apply(x, self.eps)
|
tokenizer_config.json
ADDED
@@ -0,0 +1,33 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"add_bos_token": false,
|
3 |
+
"add_prefix_space": false,
|
4 |
+
"bos_token": {
|
5 |
+
"__type": "AddedToken",
|
6 |
+
"content": "<|endoftext|>",
|
7 |
+
"lstrip": false,
|
8 |
+
"normalized": true,
|
9 |
+
"rstrip": false,
|
10 |
+
"single_word": false
|
11 |
+
},
|
12 |
+
"clean_up_tokenization_spaces": true,
|
13 |
+
"eos_token": {
|
14 |
+
"__type": "AddedToken",
|
15 |
+
"content": "<|endoftext|>",
|
16 |
+
"lstrip": false,
|
17 |
+
"normalized": true,
|
18 |
+
"rstrip": false,
|
19 |
+
"single_word": false
|
20 |
+
},
|
21 |
+
"errors": "replace",
|
22 |
+
"model_max_length": 1000000000000000019884624838656,
|
23 |
+
"pad_token": null,
|
24 |
+
"tokenizer_class": "GPT2Tokenizer",
|
25 |
+
"unk_token": {
|
26 |
+
"__type": "AddedToken",
|
27 |
+
"content": "<|endoftext|>",
|
28 |
+
"lstrip": false,
|
29 |
+
"normalized": true,
|
30 |
+
"rstrip": false,
|
31 |
+
"single_word": false
|
32 |
+
}
|
33 |
+
}
|
utils.py
ADDED
@@ -0,0 +1,151 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import logging
|
2 |
+
import os
|
3 |
+
import sys
|
4 |
+
|
5 |
+
import torch
|
6 |
+
from torch import nn
|
7 |
+
import torch.distributed as dist
|
8 |
+
import torch.nn.functional as F
|
9 |
+
|
10 |
+
from .norm import SimpleRMSNorm as SimpleRMSNormTorch
|
11 |
+
from .srmsnorm_triton import SimpleRMSNorm as SimpleRMSNormTriton
|
12 |
+
|
13 |
+
use_triton = eval(os.environ.get("use_triton", default="True"))
|
14 |
+
debug = eval(os.environ.get("debug", default="False"))
|
15 |
+
|
16 |
+
if use_triton:
|
17 |
+
SimpleRMSNorm = SimpleRMSNormTriton
|
18 |
+
else:
|
19 |
+
SimpleRMSNorm = SimpleRMSNormTorch
|
20 |
+
|
21 |
+
logging.basicConfig(
|
22 |
+
format="%(asctime)s | %(levelname)s | %(name)s | %(message)s",
|
23 |
+
datefmt="%Y-%m-%d %H:%M:%S",
|
24 |
+
level=os.environ.get("LOGLEVEL", "INFO").upper(),
|
25 |
+
stream=sys.stdout,
|
26 |
+
)
|
27 |
+
logger = logging.getLogger("print_config")
|
28 |
+
|
29 |
+
BASE_DIM = 256
|
30 |
+
|
31 |
+
|
32 |
+
def is_dist_avail_and_initialized():
|
33 |
+
if not dist.is_available():
|
34 |
+
return False
|
35 |
+
if not dist.is_initialized():
|
36 |
+
return False
|
37 |
+
return True
|
38 |
+
|
39 |
+
|
40 |
+
def get_world_size():
|
41 |
+
if not is_dist_avail_and_initialized():
|
42 |
+
return 1
|
43 |
+
return dist.get_world_size()
|
44 |
+
|
45 |
+
|
46 |
+
def get_rank():
|
47 |
+
if not is_dist_avail_and_initialized():
|
48 |
+
return 0
|
49 |
+
return dist.get_rank()
|
50 |
+
|
51 |
+
|
52 |
+
def is_main_process():
|
53 |
+
return get_rank() == 0
|
54 |
+
|
55 |
+
|
56 |
+
def logging_info(string):
|
57 |
+
if is_main_process():
|
58 |
+
logger.info(string)
|
59 |
+
|
60 |
+
|
61 |
+
def print_params(**kwargs):
|
62 |
+
if is_main_process():
|
63 |
+
logger.info(f"start print config of {kwargs['__class__']}")
|
64 |
+
for key in kwargs:
|
65 |
+
if key in ["__class__", "self"]:
|
66 |
+
continue
|
67 |
+
logger.info(f"{key}: {kwargs[key]}")
|
68 |
+
logger.info(f"end print config of {kwargs['__class__']}")
|
69 |
+
|
70 |
+
|
71 |
+
def print_config(config):
|
72 |
+
if is_main_process():
|
73 |
+
logger.info(f"start print config of {config['__class__']}")
|
74 |
+
for key in config:
|
75 |
+
if key in ["__class__", "self"]:
|
76 |
+
continue
|
77 |
+
logger.info(f"{key}: {config[key]}")
|
78 |
+
logger.info(f"end print config of {config['__class__']}")
|
79 |
+
|
80 |
+
|
81 |
+
def print_module(module):
|
82 |
+
named_modules = set()
|
83 |
+
for p in module.named_modules():
|
84 |
+
named_modules.update([p[0]])
|
85 |
+
named_modules = list(named_modules)
|
86 |
+
|
87 |
+
string_repr = ""
|
88 |
+
for p in module.named_parameters():
|
89 |
+
name = p[0].split(".")[0]
|
90 |
+
if name not in named_modules:
|
91 |
+
string_repr = (string_repr + "(" + name + "): " + "Tensor(" +
|
92 |
+
str(tuple(p[1].shape)) + ", requires_grad=" +
|
93 |
+
str(p[1].requires_grad) + ")\n")
|
94 |
+
|
95 |
+
return string_repr.rstrip("\n")
|
96 |
+
|
97 |
+
|
98 |
+
def get_activation_fn(activation):
|
99 |
+
if debug:
|
100 |
+
logger.info(f"activation: {activation}")
|
101 |
+
if activation == "gelu":
|
102 |
+
return F.gelu
|
103 |
+
elif activation == "relu":
|
104 |
+
return F.relu
|
105 |
+
elif activation == "elu":
|
106 |
+
return F.elu
|
107 |
+
elif activation == "sigmoid":
|
108 |
+
return F.sigmoid
|
109 |
+
elif activation == "exp":
|
110 |
+
|
111 |
+
def f(x):
|
112 |
+
with torch.no_grad():
|
113 |
+
x_max = torch.max(x, dim=-1, keepdims=True).values
|
114 |
+
y = torch.exp(x - x_max)
|
115 |
+
|
116 |
+
return y
|
117 |
+
|
118 |
+
return f
|
119 |
+
elif activation == "leak":
|
120 |
+
return F.leaky_relu
|
121 |
+
elif activation == "1+elu":
|
122 |
+
|
123 |
+
def f(x):
|
124 |
+
return 1 + F.elu(x)
|
125 |
+
|
126 |
+
return f
|
127 |
+
elif activation == "2+elu":
|
128 |
+
|
129 |
+
def f(x):
|
130 |
+
return 2 + F.elu(x)
|
131 |
+
|
132 |
+
return f
|
133 |
+
elif activation == "silu" or activation == "swish":
|
134 |
+
return F.silu
|
135 |
+
elif activation == "sine":
|
136 |
+
return torch.sin
|
137 |
+
else:
|
138 |
+
logger.info(
|
139 |
+
f"activation: does not support {activation}, use Identity!!!")
|
140 |
+
return lambda x: x
|
141 |
+
|
142 |
+
|
143 |
+
def get_norm_fn(norm_type):
|
144 |
+
if norm_type == "simplermsnorm":
|
145 |
+
return SimpleRMSNorm
|
146 |
+
else:
|
147 |
+
return nn.LayerNorm
|
148 |
+
|
149 |
+
|
150 |
+
def convert_to_multiple_of_base(x):
|
151 |
+
return BASE_DIM * ((x + BASE_DIM - 1) // BASE_DIM)
|
vocab.json
ADDED
The diff for this file is too large to render.
See raw diff
|
|