black formatting
Browse filesignore copied file
fix linting
.mypy.ini
CHANGED
@@ -5,6 +5,9 @@ exclude = venv
|
|
5 |
[mypy-alpaca_lora_4bit.*]
|
6 |
ignore_missing_imports = True
|
7 |
|
|
|
|
|
|
|
8 |
[mypy-flash_attn.*]
|
9 |
ignore_missing_imports = True
|
10 |
|
@@ -31,3 +34,6 @@ ignore_missing_imports = True
|
|
31 |
|
32 |
[mypy-addict]
|
33 |
ignore_missing_imports = True
|
|
|
|
|
|
|
|
5 |
[mypy-alpaca_lora_4bit.*]
|
6 |
ignore_missing_imports = True
|
7 |
|
8 |
+
[mypy-axolotl.monkeypatch.*]
|
9 |
+
ignore_errors = True
|
10 |
+
|
11 |
[mypy-flash_attn.*]
|
12 |
ignore_missing_imports = True
|
13 |
|
|
|
34 |
|
35 |
[mypy-addict]
|
36 |
ignore_missing_imports = True
|
37 |
+
|
38 |
+
[mypy-xformers.*]
|
39 |
+
ignore_missing_imports = True
|
src/axolotl/monkeypatch/llama_attn_hijack_xformers.py
CHANGED
@@ -1,18 +1,18 @@
|
|
1 |
-
|
2 |
Directly copied the code from https://raw.githubusercontent.com/oobabooga/text-generation-webui/main/modules/llama_attn_hijack.py and made some adjustments
|
3 |
-
|
4 |
|
5 |
import logging
|
6 |
import math
|
7 |
from typing import Optional, Tuple
|
8 |
|
9 |
import torch
|
10 |
-
import torch.nn as nn
|
11 |
import transformers.models.llama.modeling_llama
|
|
|
12 |
|
13 |
try:
|
14 |
import xformers.ops
|
15 |
-
except
|
16 |
logging.error("xformers not found! Please install it before trying to use it.")
|
17 |
|
18 |
|
@@ -22,7 +22,9 @@ def hijack_llama_attention():
|
|
22 |
|
23 |
|
24 |
def hijack_llama_sdp_attention():
|
25 |
-
transformers.models.llama.modeling_llama.LlamaAttention.forward =
|
|
|
|
|
26 |
logging.info("Replaced attention with sdp_attention")
|
27 |
|
28 |
|
@@ -37,15 +39,32 @@ def xformers_forward(
|
|
37 |
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
|
38 |
bsz, q_len, _ = hidden_states.size()
|
39 |
|
40 |
-
query_states =
|
41 |
-
|
42 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
43 |
|
44 |
kv_seq_len = key_states.shape[-2]
|
45 |
if past_key_value is not None:
|
46 |
kv_seq_len += past_key_value[0].shape[-2]
|
47 |
cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
|
48 |
-
|
|
|
|
|
|
|
|
|
|
|
49 |
# [bsz, nh, t, hd]
|
50 |
|
51 |
if past_key_value is not None:
|
@@ -65,13 +84,22 @@ def xformers_forward(
|
|
65 |
# We therefore check if one element in the upper triangular portion is zero. If it is, then the mask is all zeros.
|
66 |
if attention_mask is None or attention_mask[0, 0, 0, 1] == 0:
|
67 |
# input and output should be of form (bsz, q_len, num_heads, head_dim)
|
68 |
-
attn_output = xformers.ops.memory_efficient_attention(
|
|
|
|
|
69 |
else:
|
70 |
# input and output should be of form (bsz, q_len, num_heads, head_dim)
|
71 |
-
attn_output = xformers.ops.memory_efficient_attention(
|
|
|
|
|
|
|
|
|
|
|
72 |
attn_weights = None
|
73 |
else:
|
74 |
-
attn_weights = torch.matmul(
|
|
|
|
|
75 |
|
76 |
if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
|
77 |
raise ValueError(
|
@@ -85,10 +113,14 @@ def xformers_forward(
|
|
85 |
f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
|
86 |
)
|
87 |
attn_weights = attn_weights + attention_mask
|
88 |
-
attn_weights = torch.max(
|
|
|
|
|
89 |
|
90 |
# upcast attention to fp32
|
91 |
-
attn_weights = nn.functional.softmax(
|
|
|
|
|
92 |
attn_output = torch.matmul(attn_weights, value_states)
|
93 |
|
94 |
if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
|
@@ -115,15 +147,32 @@ def sdp_attention_forward(
|
|
115 |
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
|
116 |
bsz, q_len, _ = hidden_states.size()
|
117 |
|
118 |
-
query_states =
|
119 |
-
|
120 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
121 |
|
122 |
kv_seq_len = key_states.shape[-2]
|
123 |
if past_key_value is not None:
|
124 |
kv_seq_len += past_key_value[0].shape[-2]
|
125 |
cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
|
126 |
-
|
|
|
|
|
|
|
|
|
|
|
127 |
# [bsz, nh, t, hd]
|
128 |
|
129 |
if past_key_value is not None:
|
@@ -135,10 +184,18 @@ def sdp_attention_forward(
|
|
135 |
|
136 |
# We only apply sdp attention if we don't need to output the whole attention matrix
|
137 |
if not output_attentions:
|
138 |
-
attn_output = torch.nn.functional.scaled_dot_product_attention(
|
|
|
|
|
|
|
|
|
|
|
|
|
139 |
attn_weights = None
|
140 |
else:
|
141 |
-
attn_weights = torch.matmul(
|
|
|
|
|
142 |
|
143 |
if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
|
144 |
raise ValueError(
|
@@ -152,10 +209,14 @@ def sdp_attention_forward(
|
|
152 |
f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
|
153 |
)
|
154 |
attn_weights = attn_weights + attention_mask
|
155 |
-
attn_weights = torch.max(
|
|
|
|
|
156 |
|
157 |
# upcast attention to fp32
|
158 |
-
attn_weights = nn.functional.softmax(
|
|
|
|
|
159 |
attn_output = torch.matmul(attn_weights, value_states)
|
160 |
|
161 |
if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
|
|
|
1 |
+
"""
|
2 |
Directly copied the code from https://raw.githubusercontent.com/oobabooga/text-generation-webui/main/modules/llama_attn_hijack.py and made some adjustments
|
3 |
+
"""
|
4 |
|
5 |
import logging
|
6 |
import math
|
7 |
from typing import Optional, Tuple
|
8 |
|
9 |
import torch
|
|
|
10 |
import transformers.models.llama.modeling_llama
|
11 |
+
from torch import nn
|
12 |
|
13 |
try:
|
14 |
import xformers.ops
|
15 |
+
except ImportError:
|
16 |
logging.error("xformers not found! Please install it before trying to use it.")
|
17 |
|
18 |
|
|
|
22 |
|
23 |
|
24 |
def hijack_llama_sdp_attention():
|
25 |
+
transformers.models.llama.modeling_llama.LlamaAttention.forward = (
|
26 |
+
sdp_attention_forward
|
27 |
+
)
|
28 |
logging.info("Replaced attention with sdp_attention")
|
29 |
|
30 |
|
|
|
39 |
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
|
40 |
bsz, q_len, _ = hidden_states.size()
|
41 |
|
42 |
+
query_states = (
|
43 |
+
self.q_proj(hidden_states)
|
44 |
+
.view(bsz, q_len, self.num_heads, self.head_dim)
|
45 |
+
.transpose(1, 2)
|
46 |
+
)
|
47 |
+
key_states = (
|
48 |
+
self.k_proj(hidden_states)
|
49 |
+
.view(bsz, q_len, self.num_heads, self.head_dim)
|
50 |
+
.transpose(1, 2)
|
51 |
+
)
|
52 |
+
value_states = (
|
53 |
+
self.v_proj(hidden_states)
|
54 |
+
.view(bsz, q_len, self.num_heads, self.head_dim)
|
55 |
+
.transpose(1, 2)
|
56 |
+
)
|
57 |
|
58 |
kv_seq_len = key_states.shape[-2]
|
59 |
if past_key_value is not None:
|
60 |
kv_seq_len += past_key_value[0].shape[-2]
|
61 |
cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
|
62 |
+
(
|
63 |
+
query_states,
|
64 |
+
key_states,
|
65 |
+
) = transformers.models.llama.modeling_llama.apply_rotary_pos_emb(
|
66 |
+
query_states, key_states, cos, sin, position_ids
|
67 |
+
)
|
68 |
# [bsz, nh, t, hd]
|
69 |
|
70 |
if past_key_value is not None:
|
|
|
84 |
# We therefore check if one element in the upper triangular portion is zero. If it is, then the mask is all zeros.
|
85 |
if attention_mask is None or attention_mask[0, 0, 0, 1] == 0:
|
86 |
# input and output should be of form (bsz, q_len, num_heads, head_dim)
|
87 |
+
attn_output = xformers.ops.memory_efficient_attention(
|
88 |
+
query_states, key_states, value_states, attn_bias=None
|
89 |
+
)
|
90 |
else:
|
91 |
# input and output should be of form (bsz, q_len, num_heads, head_dim)
|
92 |
+
attn_output = xformers.ops.memory_efficient_attention(
|
93 |
+
query_states,
|
94 |
+
key_states,
|
95 |
+
value_states,
|
96 |
+
attn_bias=xformers.ops.LowerTriangularMask(),
|
97 |
+
)
|
98 |
attn_weights = None
|
99 |
else:
|
100 |
+
attn_weights = torch.matmul(
|
101 |
+
query_states, key_states.transpose(2, 3)
|
102 |
+
) / math.sqrt(self.head_dim)
|
103 |
|
104 |
if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
|
105 |
raise ValueError(
|
|
|
113 |
f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
|
114 |
)
|
115 |
attn_weights = attn_weights + attention_mask
|
116 |
+
attn_weights = torch.max(
|
117 |
+
attn_weights, torch.tensor(torch.finfo(attn_weights.dtype).min)
|
118 |
+
)
|
119 |
|
120 |
# upcast attention to fp32
|
121 |
+
attn_weights = nn.functional.softmax(
|
122 |
+
attn_weights, dim=-1, dtype=torch.float32
|
123 |
+
).to(query_states.dtype)
|
124 |
attn_output = torch.matmul(attn_weights, value_states)
|
125 |
|
126 |
if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
|
|
|
147 |
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
|
148 |
bsz, q_len, _ = hidden_states.size()
|
149 |
|
150 |
+
query_states = (
|
151 |
+
self.q_proj(hidden_states)
|
152 |
+
.view(bsz, q_len, self.num_heads, self.head_dim)
|
153 |
+
.transpose(1, 2)
|
154 |
+
)
|
155 |
+
key_states = (
|
156 |
+
self.k_proj(hidden_states)
|
157 |
+
.view(bsz, q_len, self.num_heads, self.head_dim)
|
158 |
+
.transpose(1, 2)
|
159 |
+
)
|
160 |
+
value_states = (
|
161 |
+
self.v_proj(hidden_states)
|
162 |
+
.view(bsz, q_len, self.num_heads, self.head_dim)
|
163 |
+
.transpose(1, 2)
|
164 |
+
)
|
165 |
|
166 |
kv_seq_len = key_states.shape[-2]
|
167 |
if past_key_value is not None:
|
168 |
kv_seq_len += past_key_value[0].shape[-2]
|
169 |
cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
|
170 |
+
(
|
171 |
+
query_states,
|
172 |
+
key_states,
|
173 |
+
) = transformers.models.llama.modeling_llama.apply_rotary_pos_emb(
|
174 |
+
query_states, key_states, cos, sin, position_ids
|
175 |
+
)
|
176 |
# [bsz, nh, t, hd]
|
177 |
|
178 |
if past_key_value is not None:
|
|
|
184 |
|
185 |
# We only apply sdp attention if we don't need to output the whole attention matrix
|
186 |
if not output_attentions:
|
187 |
+
attn_output = torch.nn.functional.scaled_dot_product_attention(
|
188 |
+
query_states,
|
189 |
+
key_states,
|
190 |
+
value_states,
|
191 |
+
attn_mask=attention_mask,
|
192 |
+
is_causal=False,
|
193 |
+
)
|
194 |
attn_weights = None
|
195 |
else:
|
196 |
+
attn_weights = torch.matmul(
|
197 |
+
query_states, key_states.transpose(2, 3)
|
198 |
+
) / math.sqrt(self.head_dim)
|
199 |
|
200 |
if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
|
201 |
raise ValueError(
|
|
|
209 |
f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
|
210 |
)
|
211 |
attn_weights = attn_weights + attention_mask
|
212 |
+
attn_weights = torch.max(
|
213 |
+
attn_weights, torch.tensor(torch.finfo(attn_weights.dtype).min)
|
214 |
+
)
|
215 |
|
216 |
# upcast attention to fp32
|
217 |
+
attn_weights = nn.functional.softmax(
|
218 |
+
attn_weights, dim=-1, dtype=torch.float32
|
219 |
+
).to(query_states.dtype)
|
220 |
attn_output = torch.matmul(attn_weights, value_states)
|
221 |
|
222 |
if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
|