Build
Browse files- build/torch25-cxx11-cu118-x86_64-linux/flash_attn/__init__.py +343 -16
- build/torch25-cxx11-cu118-x86_64-linux/flash_attn/{_flash_attn_ab4cc6a_dirty.abi3.so → _flash_attn_dd2f0f9.abi3.so} +2 -2
- build/torch25-cxx11-cu118-x86_64-linux/flash_attn/_ops.py +3 -3
- build/torch25-cxx11-cu121-x86_64-linux/flash_attn/{_flash_attn_ab4cc6a_dirty.abi3.so → _flash_attn_dd2f0f9.abi3.so} +2 -2
- build/torch25-cxx11-cu121-x86_64-linux/flash_attn/_ops.py +3 -3
- build/torch25-cxx11-cu124-x86_64-linux/flash_attn/{_flash_attn_ab4cc6a_dirty.abi3.so → _flash_attn_dd2f0f9.abi3.so} +2 -2
- build/torch25-cxx11-cu124-x86_64-linux/flash_attn/_ops.py +3 -3
- build/torch25-cxx98-cu118-x86_64-linux/flash_attn/{_flash_attn_ab4cc6a_dirty.abi3.so → _flash_attn_dd2f0f9.abi3.so} +2 -2
- build/torch25-cxx98-cu118-x86_64-linux/flash_attn/_ops.py +3 -3
- build/torch25-cxx98-cu121-x86_64-linux/flash_attn/_flash_attn_ab4cc6a_dirty.abi3.so +0 -3
- build/torch25-cxx98-cu121-x86_64-linux/flash_attn/_flash_attn_dd2f0f9.abi3.so +3 -0
- build/torch25-cxx98-cu121-x86_64-linux/flash_attn/_ops.py +3 -3
- build/torch25-cxx98-cu124-x86_64-linux/flash_attn/_flash_attn_ab4cc6a_dirty.abi3.so +0 -3
- build/torch25-cxx98-cu124-x86_64-linux/flash_attn/_flash_attn_dd2f0f9.abi3.so +3 -0
- build/torch25-cxx98-cu124-x86_64-linux/flash_attn/_ops.py +3 -3
- build/torch26-cxx11-cu118-x86_64-linux/flash_attn/_flash_attn_ab4cc6a_dirty.abi3.so +0 -3
- build/torch26-cxx11-cu118-x86_64-linux/flash_attn/_flash_attn_dd2f0f9.abi3.so +3 -0
- build/torch26-cxx11-cu118-x86_64-linux/flash_attn/_ops.py +3 -3
- build/torch26-cxx11-cu124-x86_64-linux/flash_attn/_flash_attn_ab4cc6a_dirty.abi3.so +0 -3
- build/torch26-cxx11-cu124-x86_64-linux/flash_attn/_flash_attn_dd2f0f9.abi3.so +3 -0
- build/torch26-cxx11-cu124-x86_64-linux/flash_attn/_ops.py +3 -3
- build/torch26-cxx11-cu126-x86_64-linux/flash_attn/_flash_attn_ab4cc6a_dirty.abi3.so +0 -3
- build/torch26-cxx11-cu126-x86_64-linux/flash_attn/_flash_attn_dd2f0f9.abi3.so +3 -0
- build/torch26-cxx11-cu126-x86_64-linux/flash_attn/_ops.py +3 -3
- build/torch26-cxx98-cu118-x86_64-linux/flash_attn/_flash_attn_ab4cc6a_dirty.abi3.so +0 -3
- build/torch26-cxx98-cu118-x86_64-linux/flash_attn/_flash_attn_dd2f0f9.abi3.so +3 -0
- build/torch26-cxx98-cu118-x86_64-linux/flash_attn/_ops.py +3 -3
- build/torch26-cxx98-cu124-x86_64-linux/flash_attn/_flash_attn_ab4cc6a_dirty.abi3.so +0 -3
- build/torch26-cxx98-cu124-x86_64-linux/flash_attn/_flash_attn_dd2f0f9.abi3.so +3 -0
- build/torch26-cxx98-cu124-x86_64-linux/flash_attn/_ops.py +3 -3
- build/torch26-cxx98-cu126-x86_64-linux/flash_attn/_flash_attn_ab4cc6a_dirty.abi3.so +0 -3
- build/torch26-cxx98-cu126-x86_64-linux/flash_attn/_flash_attn_dd2f0f9.abi3.so +3 -0
- build/torch26-cxx98-cu126-x86_64-linux/flash_attn/_ops.py +3 -3
- build/torch27-cxx11-cu118-x86_64-linux/flash_attn/__init__.py +364 -0
- build/torch27-cxx11-cu118-x86_64-linux/flash_attn/_flash_attn_dd2f0f9.abi3.so +3 -0
- build/torch27-cxx11-cu118-x86_64-linux/flash_attn/_ops.py +9 -0
- build/torch27-cxx11-cu126-x86_64-linux/flash_attn/__init__.py +364 -0
- build/torch27-cxx11-cu126-x86_64-linux/flash_attn/_flash_attn_dd2f0f9.abi3.so +3 -0
- build/torch27-cxx11-cu126-x86_64-linux/flash_attn/_ops.py +9 -0
- build/torch27-cxx11-cu128-x86_64-linux/flash_attn/__init__.py +364 -0
- build/torch27-cxx11-cu128-x86_64-linux/flash_attn/_flash_attn_dd2f0f9.abi3.so +3 -0
- build/torch27-cxx11-cu128-x86_64-linux/flash_attn/_ops.py +9 -0
build/torch25-cxx11-cu118-x86_64-linux/flash_attn/__init__.py
CHANGED
@@ -1,25 +1,45 @@
|
|
1 |
-
from typing import Optional
|
2 |
-
|
3 |
import torch
|
4 |
-
|
5 |
from ._ops import ops
|
6 |
|
|
|
7 |
def mha_fwd(
|
8 |
q: torch.Tensor,
|
9 |
k: torch.Tensor,
|
10 |
v: torch.Tensor,
|
11 |
-
out: torch.Tensor,
|
12 |
-
alibi_slopes: torch.Tensor,
|
13 |
-
p_dropout: float,
|
14 |
-
softmax_scale: float,
|
15 |
-
is_causal: bool,
|
16 |
-
window_size_left: int,
|
17 |
-
window_size_right: int,
|
18 |
-
softcap: float,
|
19 |
-
return_softmax: bool,
|
20 |
-
gen: Optional[torch.Generator],
|
21 |
-
) -> torch.Tensor:
|
22 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
23 |
q,
|
24 |
k,
|
25 |
v,
|
@@ -34,4 +54,311 @@ def mha_fwd(
|
|
34 |
return_softmax,
|
35 |
gen,
|
36 |
)
|
37 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Optional, List
|
|
|
2 |
import torch
|
|
|
3 |
from ._ops import ops
|
4 |
|
5 |
+
|
6 |
def mha_fwd(
|
7 |
q: torch.Tensor,
|
8 |
k: torch.Tensor,
|
9 |
v: torch.Tensor,
|
10 |
+
out: Optional[torch.Tensor] = None,
|
11 |
+
alibi_slopes: Optional[torch.Tensor] = None,
|
12 |
+
p_dropout: float = 0.0,
|
13 |
+
softmax_scale: float = 1.0,
|
14 |
+
is_causal: bool = False,
|
15 |
+
window_size_left: int = -1,
|
16 |
+
window_size_right: int = -1,
|
17 |
+
softcap: float = 0.0,
|
18 |
+
return_softmax: bool = False,
|
19 |
+
gen: Optional[torch.Generator] = None,
|
20 |
+
) -> List[torch.Tensor]:
|
21 |
+
"""
|
22 |
+
Forward pass for multi-head attention.
|
23 |
+
|
24 |
+
Args:
|
25 |
+
q: Query tensor of shape [batch_size, seqlen_q, num_heads, head_size]
|
26 |
+
k: Key tensor of shape [batch_size, seqlen_k, num_heads_k, head_size]
|
27 |
+
v: Value tensor of shape [batch_size, seqlen_k, num_heads_k, head_size]
|
28 |
+
out: Optional output tensor, same shape as q
|
29 |
+
alibi_slopes: Optional ALiBi slopes tensor of shape [num_heads] or [batch_size, num_heads]
|
30 |
+
p_dropout: Dropout probability
|
31 |
+
softmax_scale: Scale factor for softmax
|
32 |
+
is_causal: Whether to use causal attention
|
33 |
+
window_size_left: Window size for left context (-1 for unlimited)
|
34 |
+
window_size_right: Window size for right context (-1 for unlimited)
|
35 |
+
softcap: Soft cap for attention weights
|
36 |
+
return_softmax: Whether to return softmax weights
|
37 |
+
gen: Optional random number generator
|
38 |
+
|
39 |
+
Returns:
|
40 |
+
List of tensors: [output, softmax_lse, (softmax if return_softmax)]
|
41 |
+
"""
|
42 |
+
return ops.mha_fwd(
|
43 |
q,
|
44 |
k,
|
45 |
v,
|
|
|
54 |
return_softmax,
|
55 |
gen,
|
56 |
)
|
57 |
+
|
58 |
+
|
59 |
+
def mha_varlen_fwd(
|
60 |
+
q: torch.Tensor,
|
61 |
+
k: torch.Tensor,
|
62 |
+
v: torch.Tensor,
|
63 |
+
cu_seqlens_q: torch.Tensor,
|
64 |
+
cu_seqlens_k: torch.Tensor,
|
65 |
+
out: Optional[torch.Tensor] = None,
|
66 |
+
seqused_k: Optional[torch.Tensor] = None,
|
67 |
+
leftpad_k: Optional[torch.Tensor] = None,
|
68 |
+
block_table: Optional[torch.Tensor] = None,
|
69 |
+
alibi_slopes: Optional[torch.Tensor] = None,
|
70 |
+
max_seqlen_q: int = 0,
|
71 |
+
max_seqlen_k: int = 0,
|
72 |
+
p_dropout: float = 0.0,
|
73 |
+
softmax_scale: float = 1.0,
|
74 |
+
zero_tensors: bool = False,
|
75 |
+
is_causal: bool = False,
|
76 |
+
window_size_left: int = -1,
|
77 |
+
window_size_right: int = -1,
|
78 |
+
softcap: float = 0.0,
|
79 |
+
return_softmax: bool = False,
|
80 |
+
gen: Optional[torch.Generator] = None,
|
81 |
+
) -> List[torch.Tensor]:
|
82 |
+
"""
|
83 |
+
Forward pass for multi-head attention with variable sequence lengths.
|
84 |
+
|
85 |
+
Args:
|
86 |
+
q: Query tensor of shape [total_q, num_heads, head_size]
|
87 |
+
k: Key tensor of shape [total_k, num_heads_k, head_size] or [num_blocks, page_block_size, num_heads_k, head_size]
|
88 |
+
v: Value tensor of shape [total_k, num_heads_k, head_size] or [num_blocks, page_block_size, num_heads_k, head_size]
|
89 |
+
cu_seqlens_q: Cumulative sequence lengths for queries of shape [batch_size+1]
|
90 |
+
cu_seqlens_k: Cumulative sequence lengths for keys of shape [batch_size+1]
|
91 |
+
out: Optional output tensor of shape [total_q, num_heads, head_size]
|
92 |
+
seqused_k: Optional tensor specifying how many keys to use per batch element [batch_size]
|
93 |
+
leftpad_k: Optional left padding for keys of shape [batch_size]
|
94 |
+
block_table: Optional block table of shape [batch_size, max_num_blocks_per_seq]
|
95 |
+
alibi_slopes: Optional ALiBi slopes tensor of shape [num_heads] or [batch_size, num_heads]
|
96 |
+
max_seqlen_q: Maximum sequence length for queries
|
97 |
+
max_seqlen_k: Maximum sequence length for keys
|
98 |
+
p_dropout: Dropout probability
|
99 |
+
softmax_scale: Scale factor for softmax
|
100 |
+
zero_tensors: Whether to zero tensors before computation
|
101 |
+
is_causal: Whether to use causal attention
|
102 |
+
window_size_left: Window size for left context (-1 for unlimited)
|
103 |
+
window_size_right: Window size for right context (-1 for unlimited)
|
104 |
+
softcap: Soft cap for attention weights
|
105 |
+
return_softmax: Whether to return softmax weights
|
106 |
+
gen: Optional random number generator
|
107 |
+
|
108 |
+
Returns:
|
109 |
+
List of tensors: [output, softmax_lse, (softmax if return_softmax)]
|
110 |
+
"""
|
111 |
+
return ops.mha_varlen_fwd(
|
112 |
+
q,
|
113 |
+
k,
|
114 |
+
v,
|
115 |
+
out,
|
116 |
+
cu_seqlens_q,
|
117 |
+
cu_seqlens_k,
|
118 |
+
seqused_k,
|
119 |
+
leftpad_k,
|
120 |
+
block_table,
|
121 |
+
alibi_slopes,
|
122 |
+
max_seqlen_q,
|
123 |
+
max_seqlen_k,
|
124 |
+
p_dropout,
|
125 |
+
softmax_scale,
|
126 |
+
zero_tensors,
|
127 |
+
is_causal,
|
128 |
+
window_size_left,
|
129 |
+
window_size_right,
|
130 |
+
softcap,
|
131 |
+
return_softmax,
|
132 |
+
gen,
|
133 |
+
)
|
134 |
+
|
135 |
+
|
136 |
+
def mha_bwd(
|
137 |
+
dout: torch.Tensor,
|
138 |
+
q: torch.Tensor,
|
139 |
+
k: torch.Tensor,
|
140 |
+
v: torch.Tensor,
|
141 |
+
out: torch.Tensor,
|
142 |
+
softmax_lse: torch.Tensor,
|
143 |
+
dq: Optional[torch.Tensor] = None,
|
144 |
+
dk: Optional[torch.Tensor] = None,
|
145 |
+
dv: Optional[torch.Tensor] = None,
|
146 |
+
alibi_slopes: Optional[torch.Tensor] = None,
|
147 |
+
p_dropout: float = 0.0,
|
148 |
+
softmax_scale: float = 1.0,
|
149 |
+
is_causal: bool = False,
|
150 |
+
window_size_left: int = -1,
|
151 |
+
window_size_right: int = -1,
|
152 |
+
softcap: float = 0.0,
|
153 |
+
deterministic: bool = False,
|
154 |
+
gen: Optional[torch.Generator] = None,
|
155 |
+
rng_state: Optional[torch.Tensor] = None,
|
156 |
+
) -> List[torch.Tensor]:
|
157 |
+
"""
|
158 |
+
Backward pass for multi-head attention.
|
159 |
+
|
160 |
+
Args:
|
161 |
+
dout: Gradient tensor of shape [batch_size, seqlen_q, num_heads, head_size]
|
162 |
+
q: Query tensor of shape [batch_size, seqlen_q, num_heads, head_size]
|
163 |
+
k: Key tensor of shape [batch_size, seqlen_k, num_heads_k, head_size]
|
164 |
+
v: Value tensor of shape [batch_size, seqlen_k, num_heads_k, head_size]
|
165 |
+
out: Output tensor from forward pass of shape [batch_size, seqlen_q, num_heads, head_size]
|
166 |
+
softmax_lse: Log-sum-exp values from forward pass of shape [batch_size, num_heads, seqlen_q]
|
167 |
+
dq: Optional gradient tensor for queries, same shape as q
|
168 |
+
dk: Optional gradient tensor for keys, same shape as k
|
169 |
+
dv: Optional gradient tensor for values, same shape as v
|
170 |
+
alibi_slopes: Optional ALiBi slopes tensor of shape [num_heads] or [batch_size, num_heads]
|
171 |
+
p_dropout: Dropout probability
|
172 |
+
softmax_scale: Scale factor for softmax
|
173 |
+
is_causal: Whether to use causal attention
|
174 |
+
window_size_left: Window size for left context (-1 for unlimited)
|
175 |
+
window_size_right: Window size for right context (-1 for unlimited)
|
176 |
+
softcap: Soft cap for attention weights
|
177 |
+
deterministic: Whether to use deterministic algorithms
|
178 |
+
gen: Optional random number generator
|
179 |
+
rng_state: Optional RNG state from forward pass
|
180 |
+
|
181 |
+
Returns:
|
182 |
+
List of tensors: [dq, dk, dv]
|
183 |
+
"""
|
184 |
+
return ops.mha_bwd(
|
185 |
+
dout,
|
186 |
+
q,
|
187 |
+
k,
|
188 |
+
v,
|
189 |
+
out,
|
190 |
+
softmax_lse,
|
191 |
+
dq,
|
192 |
+
dk,
|
193 |
+
dv,
|
194 |
+
alibi_slopes,
|
195 |
+
p_dropout,
|
196 |
+
softmax_scale,
|
197 |
+
is_causal,
|
198 |
+
window_size_left,
|
199 |
+
window_size_right,
|
200 |
+
softcap,
|
201 |
+
deterministic,
|
202 |
+
gen,
|
203 |
+
rng_state,
|
204 |
+
)
|
205 |
+
|
206 |
+
|
207 |
+
def mha_varlen_bwd(
|
208 |
+
dout: torch.Tensor,
|
209 |
+
q: torch.Tensor,
|
210 |
+
k: torch.Tensor,
|
211 |
+
v: torch.Tensor,
|
212 |
+
out: torch.Tensor,
|
213 |
+
softmax_lse: torch.Tensor,
|
214 |
+
cu_seqlens_q: torch.Tensor,
|
215 |
+
cu_seqlens_k: torch.Tensor,
|
216 |
+
dq: Optional[torch.Tensor] = None,
|
217 |
+
dk: Optional[torch.Tensor] = None,
|
218 |
+
dv: Optional[torch.Tensor] = None,
|
219 |
+
alibi_slopes: Optional[torch.Tensor] = None,
|
220 |
+
max_seqlen_q: int = 0,
|
221 |
+
max_seqlen_k: int = 0,
|
222 |
+
p_dropout: float = 0.0,
|
223 |
+
softmax_scale: float = 1.0,
|
224 |
+
zero_tensors: bool = False,
|
225 |
+
is_causal: bool = False,
|
226 |
+
window_size_left: int = -1,
|
227 |
+
window_size_right: int = -1,
|
228 |
+
softcap: float = 0.0,
|
229 |
+
deterministic: bool = False,
|
230 |
+
gen: Optional[torch.Generator] = None,
|
231 |
+
rng_state: Optional[torch.Tensor] = None,
|
232 |
+
) -> List[torch.Tensor]:
|
233 |
+
"""
|
234 |
+
Backward pass for multi-head attention with variable sequence lengths.
|
235 |
+
|
236 |
+
Args:
|
237 |
+
dout: Gradient tensor of shape [batch_size, seqlen_q, num_heads, head_size]
|
238 |
+
q: Query tensor of shape [batch_size, seqlen_q, num_heads, head_size]
|
239 |
+
k: Key tensor of shape [batch_size, seqlen_k, num_heads_k, head_size]
|
240 |
+
v: Value tensor of shape [batch_size, seqlen_k, num_heads_k, head_size]
|
241 |
+
out: Output tensor from forward pass of shape [batch_size, seqlen_q, num_heads, head_size]
|
242 |
+
softmax_lse: Log-sum-exp values from forward pass of shape [batch_size, num_heads, seqlen_q]
|
243 |
+
cu_seqlens_q: Cumulative sequence lengths for queries of shape [batch_size+1]
|
244 |
+
cu_seqlens_k: Cumulative sequence lengths for keys of shape [batch_size+1]
|
245 |
+
dq: Optional gradient tensor for queries, same shape as q
|
246 |
+
dk: Optional gradient tensor for keys, same shape as k
|
247 |
+
dv: Optional gradient tensor for values, same shape as v
|
248 |
+
alibi_slopes: Optional ALiBi slopes tensor of shape [num_heads] or [batch_size, num_heads]
|
249 |
+
max_seqlen_q: Maximum sequence length for queries
|
250 |
+
max_seqlen_k: Maximum sequence length for keys
|
251 |
+
p_dropout: Dropout probability
|
252 |
+
softmax_scale: Scale factor for softmax
|
253 |
+
zero_tensors: Whether to zero tensors before computation
|
254 |
+
is_causal: Whether to use causal attention
|
255 |
+
window_size_left: Window size for left context (-1 for unlimited)
|
256 |
+
window_size_right: Window size for right context (-1 for unlimited)
|
257 |
+
softcap: Soft cap for attention weights
|
258 |
+
deterministic: Whether to use deterministic algorithms
|
259 |
+
gen: Optional random number generator
|
260 |
+
rng_state: Optional RNG state from forward pass
|
261 |
+
|
262 |
+
Returns:
|
263 |
+
List of tensors: [dq, dk, dv]
|
264 |
+
"""
|
265 |
+
return ops.mha_varlen_bwd(
|
266 |
+
dout,
|
267 |
+
q,
|
268 |
+
k,
|
269 |
+
v,
|
270 |
+
out,
|
271 |
+
softmax_lse,
|
272 |
+
dq,
|
273 |
+
dk,
|
274 |
+
dv,
|
275 |
+
cu_seqlens_q,
|
276 |
+
cu_seqlens_k,
|
277 |
+
alibi_slopes,
|
278 |
+
max_seqlen_q,
|
279 |
+
max_seqlen_k,
|
280 |
+
p_dropout,
|
281 |
+
softmax_scale,
|
282 |
+
zero_tensors,
|
283 |
+
is_causal,
|
284 |
+
window_size_left,
|
285 |
+
window_size_right,
|
286 |
+
softcap,
|
287 |
+
deterministic,
|
288 |
+
gen,
|
289 |
+
rng_state,
|
290 |
+
)
|
291 |
+
|
292 |
+
|
293 |
+
def mha_fwd_kvcache(
|
294 |
+
q: torch.Tensor,
|
295 |
+
kcache: torch.Tensor,
|
296 |
+
vcache: torch.Tensor,
|
297 |
+
k: Optional[torch.Tensor] = None,
|
298 |
+
v: Optional[torch.Tensor] = None,
|
299 |
+
seqlens_k: Optional[torch.Tensor] = None,
|
300 |
+
rotary_cos: Optional[torch.Tensor] = None,
|
301 |
+
rotary_sin: Optional[torch.Tensor] = None,
|
302 |
+
cache_batch_idx: Optional[torch.Tensor] = None,
|
303 |
+
leftpad_k: Optional[torch.Tensor] = None,
|
304 |
+
block_table: Optional[torch.Tensor] = None,
|
305 |
+
alibi_slopes: Optional[torch.Tensor] = None,
|
306 |
+
out: Optional[torch.Tensor] = None,
|
307 |
+
softmax_scale: float = 1.0,
|
308 |
+
is_causal: bool = False,
|
309 |
+
window_size_left: int = -1,
|
310 |
+
window_size_right: int = -1,
|
311 |
+
softcap: float = 0.0,
|
312 |
+
is_rotary_interleaved: bool = False,
|
313 |
+
num_splits: int = 1,
|
314 |
+
) -> List[torch.Tensor]:
|
315 |
+
"""
|
316 |
+
Forward pass for multi-head attention with KV cache.
|
317 |
+
|
318 |
+
Args:
|
319 |
+
q: Query tensor of shape [batch_size, seqlen_q, num_heads, head_size]
|
320 |
+
kcache: Key cache tensor of shape [batch_size_c, seqlen_k, num_heads_k, head_size] or [num_blocks, page_block_size, num_heads_k, head_size]
|
321 |
+
vcache: Value cache tensor of shape [batch_size_c, seqlen_k, num_heads_k, head_size] or [num_blocks, page_block_size, num_heads_k, head_size]
|
322 |
+
k: Optional new keys tensor of shape [batch_size, seqlen_knew, num_heads_k, head_size]
|
323 |
+
v: Optional new values tensor of shape [batch_size, seqlen_knew, num_heads_k, head_size]
|
324 |
+
seqlens_k: Optional sequence lengths for keys of shape [batch_size]
|
325 |
+
rotary_cos: Optional rotary cosine tensor of shape [seqlen_ro, rotary_dim/2]
|
326 |
+
rotary_sin: Optional rotary sine tensor of shape [seqlen_ro, rotary_dim/2]
|
327 |
+
cache_batch_idx: Optional indices to index into the KV cache
|
328 |
+
leftpad_k: Optional left padding for keys of shape [batch_size]
|
329 |
+
block_table: Optional block table of shape [batch_size, max_num_blocks_per_seq]
|
330 |
+
alibi_slopes: Optional ALiBi slopes tensor of shape [num_heads] or [batch_size, num_heads]
|
331 |
+
out: Optional output tensor, same shape as q
|
332 |
+
softmax_scale: Scale factor for softmax
|
333 |
+
is_causal: Whether to use causal attention
|
334 |
+
window_size_left: Window size for left context (-1 for unlimited)
|
335 |
+
window_size_right: Window size for right context (-1 for unlimited)
|
336 |
+
softcap: Soft cap for attention weights
|
337 |
+
is_rotary_interleaved: Whether rotary embeddings are interleaved
|
338 |
+
num_splits: Number of splits for computation
|
339 |
+
|
340 |
+
Returns:
|
341 |
+
List of tensors: [output, softmax_lse]
|
342 |
+
"""
|
343 |
+
return ops.mha_fwd_kvcache(
|
344 |
+
q,
|
345 |
+
kcache,
|
346 |
+
vcache,
|
347 |
+
k,
|
348 |
+
v,
|
349 |
+
seqlens_k,
|
350 |
+
rotary_cos,
|
351 |
+
rotary_sin,
|
352 |
+
cache_batch_idx,
|
353 |
+
leftpad_k,
|
354 |
+
block_table,
|
355 |
+
alibi_slopes,
|
356 |
+
out,
|
357 |
+
softmax_scale,
|
358 |
+
is_causal,
|
359 |
+
window_size_left,
|
360 |
+
window_size_right,
|
361 |
+
softcap,
|
362 |
+
is_rotary_interleaved,
|
363 |
+
num_splits,
|
364 |
+
)
|
build/torch25-cxx11-cu118-x86_64-linux/flash_attn/{_flash_attn_ab4cc6a_dirty.abi3.so → _flash_attn_dd2f0f9.abi3.so}
RENAMED
@@ -1,3 +1,3 @@
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:
|
3 |
-
size
|
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:14e43c95a52d7b6a974bc54b6ec30068ae8fa513583a686494caf123137dc2e5
|
3 |
+
size 658100376
|
build/torch25-cxx11-cu118-x86_64-linux/flash_attn/_ops.py
CHANGED
@@ -1,9 +1,9 @@
|
|
1 |
import torch
|
2 |
-
from . import
|
3 |
-
ops = torch.ops.
|
4 |
|
5 |
def add_op_namespace_prefix(op_name: str):
|
6 |
"""
|
7 |
Prefix op by namespace.
|
8 |
"""
|
9 |
-
return f"
|
|
|
1 |
import torch
|
2 |
+
from . import _flash_attn_dd2f0f9
|
3 |
+
ops = torch.ops._flash_attn_dd2f0f9
|
4 |
|
5 |
def add_op_namespace_prefix(op_name: str):
|
6 |
"""
|
7 |
Prefix op by namespace.
|
8 |
"""
|
9 |
+
return f"_flash_attn_dd2f0f9::{op_name}"
|
build/torch25-cxx11-cu121-x86_64-linux/flash_attn/{_flash_attn_ab4cc6a_dirty.abi3.so → _flash_attn_dd2f0f9.abi3.so}
RENAMED
@@ -1,3 +1,3 @@
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:
|
3 |
-
size
|
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:5405ad9405b2c3ded5f971fdc7a7fdfa0531eb2f1aca2e37e396003a149b1379
|
3 |
+
size 653617624
|
build/torch25-cxx11-cu121-x86_64-linux/flash_attn/_ops.py
CHANGED
@@ -1,9 +1,9 @@
|
|
1 |
import torch
|
2 |
-
from . import
|
3 |
-
ops = torch.ops.
|
4 |
|
5 |
def add_op_namespace_prefix(op_name: str):
|
6 |
"""
|
7 |
Prefix op by namespace.
|
8 |
"""
|
9 |
-
return f"
|
|
|
1 |
import torch
|
2 |
+
from . import _flash_attn_dd2f0f9
|
3 |
+
ops = torch.ops._flash_attn_dd2f0f9
|
4 |
|
5 |
def add_op_namespace_prefix(op_name: str):
|
6 |
"""
|
7 |
Prefix op by namespace.
|
8 |
"""
|
9 |
+
return f"_flash_attn_dd2f0f9::{op_name}"
|
build/torch25-cxx11-cu124-x86_64-linux/flash_attn/{_flash_attn_ab4cc6a_dirty.abi3.so → _flash_attn_dd2f0f9.abi3.so}
RENAMED
@@ -1,3 +1,3 @@
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:
|
3 |
-
size
|
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:a2b7aeea4adc77aefd217ccdfa7bcadcaef3dc6d0d7567f4b1c2c5f0321738fe
|
3 |
+
size 640704152
|
build/torch25-cxx11-cu124-x86_64-linux/flash_attn/_ops.py
CHANGED
@@ -1,9 +1,9 @@
|
|
1 |
import torch
|
2 |
-
from . import
|
3 |
-
ops = torch.ops.
|
4 |
|
5 |
def add_op_namespace_prefix(op_name: str):
|
6 |
"""
|
7 |
Prefix op by namespace.
|
8 |
"""
|
9 |
-
return f"
|
|
|
1 |
import torch
|
2 |
+
from . import _flash_attn_dd2f0f9
|
3 |
+
ops = torch.ops._flash_attn_dd2f0f9
|
4 |
|
5 |
def add_op_namespace_prefix(op_name: str):
|
6 |
"""
|
7 |
Prefix op by namespace.
|
8 |
"""
|
9 |
+
return f"_flash_attn_dd2f0f9::{op_name}"
|
build/torch25-cxx98-cu118-x86_64-linux/flash_attn/{_flash_attn_ab4cc6a_dirty.abi3.so → _flash_attn_dd2f0f9.abi3.so}
RENAMED
@@ -1,3 +1,3 @@
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:
|
3 |
-
size
|
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:73f7d75dcba8295aa14721ffc8c1aca0d86872ae03af18ab2e5149c043201d2a
|
3 |
+
size 658091712
|
build/torch25-cxx98-cu118-x86_64-linux/flash_attn/_ops.py
CHANGED
@@ -1,9 +1,9 @@
|
|
1 |
import torch
|
2 |
-
from . import
|
3 |
-
ops = torch.ops.
|
4 |
|
5 |
def add_op_namespace_prefix(op_name: str):
|
6 |
"""
|
7 |
Prefix op by namespace.
|
8 |
"""
|
9 |
-
return f"
|
|
|
1 |
import torch
|
2 |
+
from . import _flash_attn_dd2f0f9
|
3 |
+
ops = torch.ops._flash_attn_dd2f0f9
|
4 |
|
5 |
def add_op_namespace_prefix(op_name: str):
|
6 |
"""
|
7 |
Prefix op by namespace.
|
8 |
"""
|
9 |
+
return f"_flash_attn_dd2f0f9::{op_name}"
|
build/torch25-cxx98-cu121-x86_64-linux/flash_attn/_flash_attn_ab4cc6a_dirty.abi3.so
DELETED
@@ -1,3 +0,0 @@
|
|
1 |
-
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:347160ae8e05c11d1a99da542ecb4c2f6dbd30627cc6002b08c107b9d3d8af3c
|
3 |
-
size 1295640880
|
|
|
|
|
|
|
|
build/torch25-cxx98-cu121-x86_64-linux/flash_attn/_flash_attn_dd2f0f9.abi3.so
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:534cc70a7659f0acfca6fbd135d229af62f11557f910360939f55454cd2f6ce3
|
3 |
+
size 653605136
|
build/torch25-cxx98-cu121-x86_64-linux/flash_attn/_ops.py
CHANGED
@@ -1,9 +1,9 @@
|
|
1 |
import torch
|
2 |
-
from . import
|
3 |
-
ops = torch.ops.
|
4 |
|
5 |
def add_op_namespace_prefix(op_name: str):
|
6 |
"""
|
7 |
Prefix op by namespace.
|
8 |
"""
|
9 |
-
return f"
|
|
|
1 |
import torch
|
2 |
+
from . import _flash_attn_dd2f0f9
|
3 |
+
ops = torch.ops._flash_attn_dd2f0f9
|
4 |
|
5 |
def add_op_namespace_prefix(op_name: str):
|
6 |
"""
|
7 |
Prefix op by namespace.
|
8 |
"""
|
9 |
+
return f"_flash_attn_dd2f0f9::{op_name}"
|
build/torch25-cxx98-cu124-x86_64-linux/flash_attn/_flash_attn_ab4cc6a_dirty.abi3.so
DELETED
@@ -1,3 +0,0 @@
|
|
1 |
-
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:c79d6703c033ea9e1bfcc6fc3006ac88a9713d8371ea3a96d70e8495c7692f68
|
3 |
-
size 1262738936
|
|
|
|
|
|
|
|
build/torch25-cxx98-cu124-x86_64-linux/flash_attn/_flash_attn_dd2f0f9.abi3.so
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:8051fcc2f91c45d364292f8e2f04804a93b2e78844747a46afcfa926007769be
|
3 |
+
size 640695760
|
build/torch25-cxx98-cu124-x86_64-linux/flash_attn/_ops.py
CHANGED
@@ -1,9 +1,9 @@
|
|
1 |
import torch
|
2 |
-
from . import
|
3 |
-
ops = torch.ops.
|
4 |
|
5 |
def add_op_namespace_prefix(op_name: str):
|
6 |
"""
|
7 |
Prefix op by namespace.
|
8 |
"""
|
9 |
-
return f"
|
|
|
1 |
import torch
|
2 |
+
from . import _flash_attn_dd2f0f9
|
3 |
+
ops = torch.ops._flash_attn_dd2f0f9
|
4 |
|
5 |
def add_op_namespace_prefix(op_name: str):
|
6 |
"""
|
7 |
Prefix op by namespace.
|
8 |
"""
|
9 |
+
return f"_flash_attn_dd2f0f9::{op_name}"
|
build/torch26-cxx11-cu118-x86_64-linux/flash_attn/_flash_attn_ab4cc6a_dirty.abi3.so
DELETED
@@ -1,3 +0,0 @@
|
|
1 |
-
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:d8cc0f02a6eea5c9fe8e5bc7b0138cef9bf77c026dc26b08f878bd809799189e
|
3 |
-
size 1301389752
|
|
|
|
|
|
|
|
build/torch26-cxx11-cu118-x86_64-linux/flash_attn/_flash_attn_dd2f0f9.abi3.so
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:af9873164f58acc3dcea5a1ca046af1872f3a5d2061668edcec7e7802c02a0a6
|
3 |
+
size 658100640
|
build/torch26-cxx11-cu118-x86_64-linux/flash_attn/_ops.py
CHANGED
@@ -1,9 +1,9 @@
|
|
1 |
import torch
|
2 |
-
from . import
|
3 |
-
ops = torch.ops.
|
4 |
|
5 |
def add_op_namespace_prefix(op_name: str):
|
6 |
"""
|
7 |
Prefix op by namespace.
|
8 |
"""
|
9 |
-
return f"
|
|
|
1 |
import torch
|
2 |
+
from . import _flash_attn_dd2f0f9
|
3 |
+
ops = torch.ops._flash_attn_dd2f0f9
|
4 |
|
5 |
def add_op_namespace_prefix(op_name: str):
|
6 |
"""
|
7 |
Prefix op by namespace.
|
8 |
"""
|
9 |
+
return f"_flash_attn_dd2f0f9::{op_name}"
|
build/torch26-cxx11-cu124-x86_64-linux/flash_attn/_flash_attn_ab4cc6a_dirty.abi3.so
DELETED
@@ -1,3 +0,0 @@
|
|
1 |
-
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:fc6aaa5d51f3d329ec4d6fe7422ff8ff5223fa1a1e01644da196504534bd4fb6
|
3 |
-
size 1262747768
|
|
|
|
|
|
|
|
build/torch26-cxx11-cu124-x86_64-linux/flash_attn/_flash_attn_dd2f0f9.abi3.so
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:469a5bac698ba7be5a9aecd831b5ca5fd21ff37843d603d3e39888fad477d6e6
|
3 |
+
size 640704600
|
build/torch26-cxx11-cu124-x86_64-linux/flash_attn/_ops.py
CHANGED
@@ -1,9 +1,9 @@
|
|
1 |
import torch
|
2 |
-
from . import
|
3 |
-
ops = torch.ops.
|
4 |
|
5 |
def add_op_namespace_prefix(op_name: str):
|
6 |
"""
|
7 |
Prefix op by namespace.
|
8 |
"""
|
9 |
-
return f"
|
|
|
1 |
import torch
|
2 |
+
from . import _flash_attn_dd2f0f9
|
3 |
+
ops = torch.ops._flash_attn_dd2f0f9
|
4 |
|
5 |
def add_op_namespace_prefix(op_name: str):
|
6 |
"""
|
7 |
Prefix op by namespace.
|
8 |
"""
|
9 |
+
return f"_flash_attn_dd2f0f9::{op_name}"
|
build/torch26-cxx11-cu126-x86_64-linux/flash_attn/_flash_attn_ab4cc6a_dirty.abi3.so
DELETED
@@ -1,3 +0,0 @@
|
|
1 |
-
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:afce8d0bc6516f4e2ade3b45453d6370ead51ab9d368786b20109544cc8b4772
|
3 |
-
size 1273150064
|
|
|
|
|
|
|
|
build/torch26-cxx11-cu126-x86_64-linux/flash_attn/_flash_attn_dd2f0f9.abi3.so
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:475a51fa6fe806195457f3ea76e64343bb7b0beca8be1f128d24c0672de6a5ee
|
3 |
+
size 646613576
|
build/torch26-cxx11-cu126-x86_64-linux/flash_attn/_ops.py
CHANGED
@@ -1,9 +1,9 @@
|
|
1 |
import torch
|
2 |
-
from . import
|
3 |
-
ops = torch.ops.
|
4 |
|
5 |
def add_op_namespace_prefix(op_name: str):
|
6 |
"""
|
7 |
Prefix op by namespace.
|
8 |
"""
|
9 |
-
return f"
|
|
|
1 |
import torch
|
2 |
+
from . import _flash_attn_dd2f0f9
|
3 |
+
ops = torch.ops._flash_attn_dd2f0f9
|
4 |
|
5 |
def add_op_namespace_prefix(op_name: str):
|
6 |
"""
|
7 |
Prefix op by namespace.
|
8 |
"""
|
9 |
+
return f"_flash_attn_dd2f0f9::{op_name}"
|
build/torch26-cxx98-cu118-x86_64-linux/flash_attn/_flash_attn_ab4cc6a_dirty.abi3.so
DELETED
@@ -1,3 +0,0 @@
|
|
1 |
-
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:0f65a659aa158221085014ade1e92475fe08871894796ca8db38ef2d2dbbcb99
|
3 |
-
size 1301381128
|
|
|
|
|
|
|
|
build/torch26-cxx98-cu118-x86_64-linux/flash_attn/_flash_attn_dd2f0f9.abi3.so
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:2ca74dec8c6dcab25c0359d147b769d1025c46fb5a8ea81dd87ca4d03876044b
|
3 |
+
size 658092008
|
build/torch26-cxx98-cu118-x86_64-linux/flash_attn/_ops.py
CHANGED
@@ -1,9 +1,9 @@
|
|
1 |
import torch
|
2 |
-
from . import
|
3 |
-
ops = torch.ops.
|
4 |
|
5 |
def add_op_namespace_prefix(op_name: str):
|
6 |
"""
|
7 |
Prefix op by namespace.
|
8 |
"""
|
9 |
-
return f"
|
|
|
1 |
import torch
|
2 |
+
from . import _flash_attn_dd2f0f9
|
3 |
+
ops = torch.ops._flash_attn_dd2f0f9
|
4 |
|
5 |
def add_op_namespace_prefix(op_name: str):
|
6 |
"""
|
7 |
Prefix op by namespace.
|
8 |
"""
|
9 |
+
return f"_flash_attn_dd2f0f9::{op_name}"
|
build/torch26-cxx98-cu124-x86_64-linux/flash_attn/_flash_attn_ab4cc6a_dirty.abi3.so
DELETED
@@ -1,3 +0,0 @@
|
|
1 |
-
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:7ad532518c0a821e096e21c16bd89ec4c0b57b5b9cae92daa4c75100cfe712c6
|
3 |
-
size 1262739232
|
|
|
|
|
|
|
|
build/torch26-cxx98-cu124-x86_64-linux/flash_attn/_flash_attn_dd2f0f9.abi3.so
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:c366befff828dfa0bccf8de7de54c5fa6b3f796d55690e5e0623e518de89e4f2
|
3 |
+
size 640696056
|
build/torch26-cxx98-cu124-x86_64-linux/flash_attn/_ops.py
CHANGED
@@ -1,9 +1,9 @@
|
|
1 |
import torch
|
2 |
-
from . import
|
3 |
-
ops = torch.ops.
|
4 |
|
5 |
def add_op_namespace_prefix(op_name: str):
|
6 |
"""
|
7 |
Prefix op by namespace.
|
8 |
"""
|
9 |
-
return f"
|
|
|
1 |
import torch
|
2 |
+
from . import _flash_attn_dd2f0f9
|
3 |
+
ops = torch.ops._flash_attn_dd2f0f9
|
4 |
|
5 |
def add_op_namespace_prefix(op_name: str):
|
6 |
"""
|
7 |
Prefix op by namespace.
|
8 |
"""
|
9 |
+
return f"_flash_attn_dd2f0f9::{op_name}"
|
build/torch26-cxx98-cu126-x86_64-linux/flash_attn/_flash_attn_ab4cc6a_dirty.abi3.so
DELETED
@@ -1,3 +0,0 @@
|
|
1 |
-
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:300868f1f33c620a923efa0629916bb0afda4763af425de233e48389eede6db4
|
3 |
-
size 1273141520
|
|
|
|
|
|
|
|
build/torch26-cxx98-cu126-x86_64-linux/flash_attn/_flash_attn_dd2f0f9.abi3.so
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:a0753c6dad0da882bae86dae5658e5915eb20c89f19cb69352f239c597b5d697
|
3 |
+
size 646605032
|
build/torch26-cxx98-cu126-x86_64-linux/flash_attn/_ops.py
CHANGED
@@ -1,9 +1,9 @@
|
|
1 |
import torch
|
2 |
-
from . import
|
3 |
-
ops = torch.ops.
|
4 |
|
5 |
def add_op_namespace_prefix(op_name: str):
|
6 |
"""
|
7 |
Prefix op by namespace.
|
8 |
"""
|
9 |
-
return f"
|
|
|
1 |
import torch
|
2 |
+
from . import _flash_attn_dd2f0f9
|
3 |
+
ops = torch.ops._flash_attn_dd2f0f9
|
4 |
|
5 |
def add_op_namespace_prefix(op_name: str):
|
6 |
"""
|
7 |
Prefix op by namespace.
|
8 |
"""
|
9 |
+
return f"_flash_attn_dd2f0f9::{op_name}"
|
build/torch27-cxx11-cu118-x86_64-linux/flash_attn/__init__.py
ADDED
@@ -0,0 +1,364 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Optional, List
|
2 |
+
import torch
|
3 |
+
from ._ops import ops
|
4 |
+
|
5 |
+
|
6 |
+
def mha_fwd(
|
7 |
+
q: torch.Tensor,
|
8 |
+
k: torch.Tensor,
|
9 |
+
v: torch.Tensor,
|
10 |
+
out: Optional[torch.Tensor] = None,
|
11 |
+
alibi_slopes: Optional[torch.Tensor] = None,
|
12 |
+
p_dropout: float = 0.0,
|
13 |
+
softmax_scale: float = 1.0,
|
14 |
+
is_causal: bool = False,
|
15 |
+
window_size_left: int = -1,
|
16 |
+
window_size_right: int = -1,
|
17 |
+
softcap: float = 0.0,
|
18 |
+
return_softmax: bool = False,
|
19 |
+
gen: Optional[torch.Generator] = None,
|
20 |
+
) -> List[torch.Tensor]:
|
21 |
+
"""
|
22 |
+
Forward pass for multi-head attention.
|
23 |
+
|
24 |
+
Args:
|
25 |
+
q: Query tensor of shape [batch_size, seqlen_q, num_heads, head_size]
|
26 |
+
k: Key tensor of shape [batch_size, seqlen_k, num_heads_k, head_size]
|
27 |
+
v: Value tensor of shape [batch_size, seqlen_k, num_heads_k, head_size]
|
28 |
+
out: Optional output tensor, same shape as q
|
29 |
+
alibi_slopes: Optional ALiBi slopes tensor of shape [num_heads] or [batch_size, num_heads]
|
30 |
+
p_dropout: Dropout probability
|
31 |
+
softmax_scale: Scale factor for softmax
|
32 |
+
is_causal: Whether to use causal attention
|
33 |
+
window_size_left: Window size for left context (-1 for unlimited)
|
34 |
+
window_size_right: Window size for right context (-1 for unlimited)
|
35 |
+
softcap: Soft cap for attention weights
|
36 |
+
return_softmax: Whether to return softmax weights
|
37 |
+
gen: Optional random number generator
|
38 |
+
|
39 |
+
Returns:
|
40 |
+
List of tensors: [output, softmax_lse, (softmax if return_softmax)]
|
41 |
+
"""
|
42 |
+
return ops.mha_fwd(
|
43 |
+
q,
|
44 |
+
k,
|
45 |
+
v,
|
46 |
+
out,
|
47 |
+
alibi_slopes,
|
48 |
+
p_dropout,
|
49 |
+
softmax_scale,
|
50 |
+
is_causal,
|
51 |
+
window_size_left,
|
52 |
+
window_size_right,
|
53 |
+
softcap,
|
54 |
+
return_softmax,
|
55 |
+
gen,
|
56 |
+
)
|
57 |
+
|
58 |
+
|
59 |
+
def mha_varlen_fwd(
|
60 |
+
q: torch.Tensor,
|
61 |
+
k: torch.Tensor,
|
62 |
+
v: torch.Tensor,
|
63 |
+
cu_seqlens_q: torch.Tensor,
|
64 |
+
cu_seqlens_k: torch.Tensor,
|
65 |
+
out: Optional[torch.Tensor] = None,
|
66 |
+
seqused_k: Optional[torch.Tensor] = None,
|
67 |
+
leftpad_k: Optional[torch.Tensor] = None,
|
68 |
+
block_table: Optional[torch.Tensor] = None,
|
69 |
+
alibi_slopes: Optional[torch.Tensor] = None,
|
70 |
+
max_seqlen_q: int = 0,
|
71 |
+
max_seqlen_k: int = 0,
|
72 |
+
p_dropout: float = 0.0,
|
73 |
+
softmax_scale: float = 1.0,
|
74 |
+
zero_tensors: bool = False,
|
75 |
+
is_causal: bool = False,
|
76 |
+
window_size_left: int = -1,
|
77 |
+
window_size_right: int = -1,
|
78 |
+
softcap: float = 0.0,
|
79 |
+
return_softmax: bool = False,
|
80 |
+
gen: Optional[torch.Generator] = None,
|
81 |
+
) -> List[torch.Tensor]:
|
82 |
+
"""
|
83 |
+
Forward pass for multi-head attention with variable sequence lengths.
|
84 |
+
|
85 |
+
Args:
|
86 |
+
q: Query tensor of shape [total_q, num_heads, head_size]
|
87 |
+
k: Key tensor of shape [total_k, num_heads_k, head_size] or [num_blocks, page_block_size, num_heads_k, head_size]
|
88 |
+
v: Value tensor of shape [total_k, num_heads_k, head_size] or [num_blocks, page_block_size, num_heads_k, head_size]
|
89 |
+
cu_seqlens_q: Cumulative sequence lengths for queries of shape [batch_size+1]
|
90 |
+
cu_seqlens_k: Cumulative sequence lengths for keys of shape [batch_size+1]
|
91 |
+
out: Optional output tensor of shape [total_q, num_heads, head_size]
|
92 |
+
seqused_k: Optional tensor specifying how many keys to use per batch element [batch_size]
|
93 |
+
leftpad_k: Optional left padding for keys of shape [batch_size]
|
94 |
+
block_table: Optional block table of shape [batch_size, max_num_blocks_per_seq]
|
95 |
+
alibi_slopes: Optional ALiBi slopes tensor of shape [num_heads] or [batch_size, num_heads]
|
96 |
+
max_seqlen_q: Maximum sequence length for queries
|
97 |
+
max_seqlen_k: Maximum sequence length for keys
|
98 |
+
p_dropout: Dropout probability
|
99 |
+
softmax_scale: Scale factor for softmax
|
100 |
+
zero_tensors: Whether to zero tensors before computation
|
101 |
+
is_causal: Whether to use causal attention
|
102 |
+
window_size_left: Window size for left context (-1 for unlimited)
|
103 |
+
window_size_right: Window size for right context (-1 for unlimited)
|
104 |
+
softcap: Soft cap for attention weights
|
105 |
+
return_softmax: Whether to return softmax weights
|
106 |
+
gen: Optional random number generator
|
107 |
+
|
108 |
+
Returns:
|
109 |
+
List of tensors: [output, softmax_lse, (softmax if return_softmax)]
|
110 |
+
"""
|
111 |
+
return ops.mha_varlen_fwd(
|
112 |
+
q,
|
113 |
+
k,
|
114 |
+
v,
|
115 |
+
out,
|
116 |
+
cu_seqlens_q,
|
117 |
+
cu_seqlens_k,
|
118 |
+
seqused_k,
|
119 |
+
leftpad_k,
|
120 |
+
block_table,
|
121 |
+
alibi_slopes,
|
122 |
+
max_seqlen_q,
|
123 |
+
max_seqlen_k,
|
124 |
+
p_dropout,
|
125 |
+
softmax_scale,
|
126 |
+
zero_tensors,
|
127 |
+
is_causal,
|
128 |
+
window_size_left,
|
129 |
+
window_size_right,
|
130 |
+
softcap,
|
131 |
+
return_softmax,
|
132 |
+
gen,
|
133 |
+
)
|
134 |
+
|
135 |
+
|
136 |
+
def mha_bwd(
|
137 |
+
dout: torch.Tensor,
|
138 |
+
q: torch.Tensor,
|
139 |
+
k: torch.Tensor,
|
140 |
+
v: torch.Tensor,
|
141 |
+
out: torch.Tensor,
|
142 |
+
softmax_lse: torch.Tensor,
|
143 |
+
dq: Optional[torch.Tensor] = None,
|
144 |
+
dk: Optional[torch.Tensor] = None,
|
145 |
+
dv: Optional[torch.Tensor] = None,
|
146 |
+
alibi_slopes: Optional[torch.Tensor] = None,
|
147 |
+
p_dropout: float = 0.0,
|
148 |
+
softmax_scale: float = 1.0,
|
149 |
+
is_causal: bool = False,
|
150 |
+
window_size_left: int = -1,
|
151 |
+
window_size_right: int = -1,
|
152 |
+
softcap: float = 0.0,
|
153 |
+
deterministic: bool = False,
|
154 |
+
gen: Optional[torch.Generator] = None,
|
155 |
+
rng_state: Optional[torch.Tensor] = None,
|
156 |
+
) -> List[torch.Tensor]:
|
157 |
+
"""
|
158 |
+
Backward pass for multi-head attention.
|
159 |
+
|
160 |
+
Args:
|
161 |
+
dout: Gradient tensor of shape [batch_size, seqlen_q, num_heads, head_size]
|
162 |
+
q: Query tensor of shape [batch_size, seqlen_q, num_heads, head_size]
|
163 |
+
k: Key tensor of shape [batch_size, seqlen_k, num_heads_k, head_size]
|
164 |
+
v: Value tensor of shape [batch_size, seqlen_k, num_heads_k, head_size]
|
165 |
+
out: Output tensor from forward pass of shape [batch_size, seqlen_q, num_heads, head_size]
|
166 |
+
softmax_lse: Log-sum-exp values from forward pass of shape [batch_size, num_heads, seqlen_q]
|
167 |
+
dq: Optional gradient tensor for queries, same shape as q
|
168 |
+
dk: Optional gradient tensor for keys, same shape as k
|
169 |
+
dv: Optional gradient tensor for values, same shape as v
|
170 |
+
alibi_slopes: Optional ALiBi slopes tensor of shape [num_heads] or [batch_size, num_heads]
|
171 |
+
p_dropout: Dropout probability
|
172 |
+
softmax_scale: Scale factor for softmax
|
173 |
+
is_causal: Whether to use causal attention
|
174 |
+
window_size_left: Window size for left context (-1 for unlimited)
|
175 |
+
window_size_right: Window size for right context (-1 for unlimited)
|
176 |
+
softcap: Soft cap for attention weights
|
177 |
+
deterministic: Whether to use deterministic algorithms
|
178 |
+
gen: Optional random number generator
|
179 |
+
rng_state: Optional RNG state from forward pass
|
180 |
+
|
181 |
+
Returns:
|
182 |
+
List of tensors: [dq, dk, dv]
|
183 |
+
"""
|
184 |
+
return ops.mha_bwd(
|
185 |
+
dout,
|
186 |
+
q,
|
187 |
+
k,
|
188 |
+
v,
|
189 |
+
out,
|
190 |
+
softmax_lse,
|
191 |
+
dq,
|
192 |
+
dk,
|
193 |
+
dv,
|
194 |
+
alibi_slopes,
|
195 |
+
p_dropout,
|
196 |
+
softmax_scale,
|
197 |
+
is_causal,
|
198 |
+
window_size_left,
|
199 |
+
window_size_right,
|
200 |
+
softcap,
|
201 |
+
deterministic,
|
202 |
+
gen,
|
203 |
+
rng_state,
|
204 |
+
)
|
205 |
+
|
206 |
+
|
207 |
+
def mha_varlen_bwd(
|
208 |
+
dout: torch.Tensor,
|
209 |
+
q: torch.Tensor,
|
210 |
+
k: torch.Tensor,
|
211 |
+
v: torch.Tensor,
|
212 |
+
out: torch.Tensor,
|
213 |
+
softmax_lse: torch.Tensor,
|
214 |
+
cu_seqlens_q: torch.Tensor,
|
215 |
+
cu_seqlens_k: torch.Tensor,
|
216 |
+
dq: Optional[torch.Tensor] = None,
|
217 |
+
dk: Optional[torch.Tensor] = None,
|
218 |
+
dv: Optional[torch.Tensor] = None,
|
219 |
+
alibi_slopes: Optional[torch.Tensor] = None,
|
220 |
+
max_seqlen_q: int = 0,
|
221 |
+
max_seqlen_k: int = 0,
|
222 |
+
p_dropout: float = 0.0,
|
223 |
+
softmax_scale: float = 1.0,
|
224 |
+
zero_tensors: bool = False,
|
225 |
+
is_causal: bool = False,
|
226 |
+
window_size_left: int = -1,
|
227 |
+
window_size_right: int = -1,
|
228 |
+
softcap: float = 0.0,
|
229 |
+
deterministic: bool = False,
|
230 |
+
gen: Optional[torch.Generator] = None,
|
231 |
+
rng_state: Optional[torch.Tensor] = None,
|
232 |
+
) -> List[torch.Tensor]:
|
233 |
+
"""
|
234 |
+
Backward pass for multi-head attention with variable sequence lengths.
|
235 |
+
|
236 |
+
Args:
|
237 |
+
dout: Gradient tensor of shape [batch_size, seqlen_q, num_heads, head_size]
|
238 |
+
q: Query tensor of shape [batch_size, seqlen_q, num_heads, head_size]
|
239 |
+
k: Key tensor of shape [batch_size, seqlen_k, num_heads_k, head_size]
|
240 |
+
v: Value tensor of shape [batch_size, seqlen_k, num_heads_k, head_size]
|
241 |
+
out: Output tensor from forward pass of shape [batch_size, seqlen_q, num_heads, head_size]
|
242 |
+
softmax_lse: Log-sum-exp values from forward pass of shape [batch_size, num_heads, seqlen_q]
|
243 |
+
cu_seqlens_q: Cumulative sequence lengths for queries of shape [batch_size+1]
|
244 |
+
cu_seqlens_k: Cumulative sequence lengths for keys of shape [batch_size+1]
|
245 |
+
dq: Optional gradient tensor for queries, same shape as q
|
246 |
+
dk: Optional gradient tensor for keys, same shape as k
|
247 |
+
dv: Optional gradient tensor for values, same shape as v
|
248 |
+
alibi_slopes: Optional ALiBi slopes tensor of shape [num_heads] or [batch_size, num_heads]
|
249 |
+
max_seqlen_q: Maximum sequence length for queries
|
250 |
+
max_seqlen_k: Maximum sequence length for keys
|
251 |
+
p_dropout: Dropout probability
|
252 |
+
softmax_scale: Scale factor for softmax
|
253 |
+
zero_tensors: Whether to zero tensors before computation
|
254 |
+
is_causal: Whether to use causal attention
|
255 |
+
window_size_left: Window size for left context (-1 for unlimited)
|
256 |
+
window_size_right: Window size for right context (-1 for unlimited)
|
257 |
+
softcap: Soft cap for attention weights
|
258 |
+
deterministic: Whether to use deterministic algorithms
|
259 |
+
gen: Optional random number generator
|
260 |
+
rng_state: Optional RNG state from forward pass
|
261 |
+
|
262 |
+
Returns:
|
263 |
+
List of tensors: [dq, dk, dv]
|
264 |
+
"""
|
265 |
+
return ops.mha_varlen_bwd(
|
266 |
+
dout,
|
267 |
+
q,
|
268 |
+
k,
|
269 |
+
v,
|
270 |
+
out,
|
271 |
+
softmax_lse,
|
272 |
+
dq,
|
273 |
+
dk,
|
274 |
+
dv,
|
275 |
+
cu_seqlens_q,
|
276 |
+
cu_seqlens_k,
|
277 |
+
alibi_slopes,
|
278 |
+
max_seqlen_q,
|
279 |
+
max_seqlen_k,
|
280 |
+
p_dropout,
|
281 |
+
softmax_scale,
|
282 |
+
zero_tensors,
|
283 |
+
is_causal,
|
284 |
+
window_size_left,
|
285 |
+
window_size_right,
|
286 |
+
softcap,
|
287 |
+
deterministic,
|
288 |
+
gen,
|
289 |
+
rng_state,
|
290 |
+
)
|
291 |
+
|
292 |
+
|
293 |
+
def mha_fwd_kvcache(
|
294 |
+
q: torch.Tensor,
|
295 |
+
kcache: torch.Tensor,
|
296 |
+
vcache: torch.Tensor,
|
297 |
+
k: Optional[torch.Tensor] = None,
|
298 |
+
v: Optional[torch.Tensor] = None,
|
299 |
+
seqlens_k: Optional[torch.Tensor] = None,
|
300 |
+
rotary_cos: Optional[torch.Tensor] = None,
|
301 |
+
rotary_sin: Optional[torch.Tensor] = None,
|
302 |
+
cache_batch_idx: Optional[torch.Tensor] = None,
|
303 |
+
leftpad_k: Optional[torch.Tensor] = None,
|
304 |
+
block_table: Optional[torch.Tensor] = None,
|
305 |
+
alibi_slopes: Optional[torch.Tensor] = None,
|
306 |
+
out: Optional[torch.Tensor] = None,
|
307 |
+
softmax_scale: float = 1.0,
|
308 |
+
is_causal: bool = False,
|
309 |
+
window_size_left: int = -1,
|
310 |
+
window_size_right: int = -1,
|
311 |
+
softcap: float = 0.0,
|
312 |
+
is_rotary_interleaved: bool = False,
|
313 |
+
num_splits: int = 1,
|
314 |
+
) -> List[torch.Tensor]:
|
315 |
+
"""
|
316 |
+
Forward pass for multi-head attention with KV cache.
|
317 |
+
|
318 |
+
Args:
|
319 |
+
q: Query tensor of shape [batch_size, seqlen_q, num_heads, head_size]
|
320 |
+
kcache: Key cache tensor of shape [batch_size_c, seqlen_k, num_heads_k, head_size] or [num_blocks, page_block_size, num_heads_k, head_size]
|
321 |
+
vcache: Value cache tensor of shape [batch_size_c, seqlen_k, num_heads_k, head_size] or [num_blocks, page_block_size, num_heads_k, head_size]
|
322 |
+
k: Optional new keys tensor of shape [batch_size, seqlen_knew, num_heads_k, head_size]
|
323 |
+
v: Optional new values tensor of shape [batch_size, seqlen_knew, num_heads_k, head_size]
|
324 |
+
seqlens_k: Optional sequence lengths for keys of shape [batch_size]
|
325 |
+
rotary_cos: Optional rotary cosine tensor of shape [seqlen_ro, rotary_dim/2]
|
326 |
+
rotary_sin: Optional rotary sine tensor of shape [seqlen_ro, rotary_dim/2]
|
327 |
+
cache_batch_idx: Optional indices to index into the KV cache
|
328 |
+
leftpad_k: Optional left padding for keys of shape [batch_size]
|
329 |
+
block_table: Optional block table of shape [batch_size, max_num_blocks_per_seq]
|
330 |
+
alibi_slopes: Optional ALiBi slopes tensor of shape [num_heads] or [batch_size, num_heads]
|
331 |
+
out: Optional output tensor, same shape as q
|
332 |
+
softmax_scale: Scale factor for softmax
|
333 |
+
is_causal: Whether to use causal attention
|
334 |
+
window_size_left: Window size for left context (-1 for unlimited)
|
335 |
+
window_size_right: Window size for right context (-1 for unlimited)
|
336 |
+
softcap: Soft cap for attention weights
|
337 |
+
is_rotary_interleaved: Whether rotary embeddings are interleaved
|
338 |
+
num_splits: Number of splits for computation
|
339 |
+
|
340 |
+
Returns:
|
341 |
+
List of tensors: [output, softmax_lse]
|
342 |
+
"""
|
343 |
+
return ops.mha_fwd_kvcache(
|
344 |
+
q,
|
345 |
+
kcache,
|
346 |
+
vcache,
|
347 |
+
k,
|
348 |
+
v,
|
349 |
+
seqlens_k,
|
350 |
+
rotary_cos,
|
351 |
+
rotary_sin,
|
352 |
+
cache_batch_idx,
|
353 |
+
leftpad_k,
|
354 |
+
block_table,
|
355 |
+
alibi_slopes,
|
356 |
+
out,
|
357 |
+
softmax_scale,
|
358 |
+
is_causal,
|
359 |
+
window_size_left,
|
360 |
+
window_size_right,
|
361 |
+
softcap,
|
362 |
+
is_rotary_interleaved,
|
363 |
+
num_splits,
|
364 |
+
)
|
build/torch27-cxx11-cu118-x86_64-linux/flash_attn/_flash_attn_dd2f0f9.abi3.so
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:dcb1788a80f0624ec6532ea3abdbd1ef504364006129ef4564d131f2a44dc916
|
3 |
+
size 658100920
|
build/torch27-cxx11-cu118-x86_64-linux/flash_attn/_ops.py
ADDED
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
from . import _flash_attn_dd2f0f9
|
3 |
+
ops = torch.ops._flash_attn_dd2f0f9
|
4 |
+
|
5 |
+
def add_op_namespace_prefix(op_name: str):
|
6 |
+
"""
|
7 |
+
Prefix op by namespace.
|
8 |
+
"""
|
9 |
+
return f"_flash_attn_dd2f0f9::{op_name}"
|
build/torch27-cxx11-cu126-x86_64-linux/flash_attn/__init__.py
ADDED
@@ -0,0 +1,364 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Optional, List
|
2 |
+
import torch
|
3 |
+
from ._ops import ops
|
4 |
+
|
5 |
+
|
6 |
+
def mha_fwd(
|
7 |
+
q: torch.Tensor,
|
8 |
+
k: torch.Tensor,
|
9 |
+
v: torch.Tensor,
|
10 |
+
out: Optional[torch.Tensor] = None,
|
11 |
+
alibi_slopes: Optional[torch.Tensor] = None,
|
12 |
+
p_dropout: float = 0.0,
|
13 |
+
softmax_scale: float = 1.0,
|
14 |
+
is_causal: bool = False,
|
15 |
+
window_size_left: int = -1,
|
16 |
+
window_size_right: int = -1,
|
17 |
+
softcap: float = 0.0,
|
18 |
+
return_softmax: bool = False,
|
19 |
+
gen: Optional[torch.Generator] = None,
|
20 |
+
) -> List[torch.Tensor]:
|
21 |
+
"""
|
22 |
+
Forward pass for multi-head attention.
|
23 |
+
|
24 |
+
Args:
|
25 |
+
q: Query tensor of shape [batch_size, seqlen_q, num_heads, head_size]
|
26 |
+
k: Key tensor of shape [batch_size, seqlen_k, num_heads_k, head_size]
|
27 |
+
v: Value tensor of shape [batch_size, seqlen_k, num_heads_k, head_size]
|
28 |
+
out: Optional output tensor, same shape as q
|
29 |
+
alibi_slopes: Optional ALiBi slopes tensor of shape [num_heads] or [batch_size, num_heads]
|
30 |
+
p_dropout: Dropout probability
|
31 |
+
softmax_scale: Scale factor for softmax
|
32 |
+
is_causal: Whether to use causal attention
|
33 |
+
window_size_left: Window size for left context (-1 for unlimited)
|
34 |
+
window_size_right: Window size for right context (-1 for unlimited)
|
35 |
+
softcap: Soft cap for attention weights
|
36 |
+
return_softmax: Whether to return softmax weights
|
37 |
+
gen: Optional random number generator
|
38 |
+
|
39 |
+
Returns:
|
40 |
+
List of tensors: [output, softmax_lse, (softmax if return_softmax)]
|
41 |
+
"""
|
42 |
+
return ops.mha_fwd(
|
43 |
+
q,
|
44 |
+
k,
|
45 |
+
v,
|
46 |
+
out,
|
47 |
+
alibi_slopes,
|
48 |
+
p_dropout,
|
49 |
+
softmax_scale,
|
50 |
+
is_causal,
|
51 |
+
window_size_left,
|
52 |
+
window_size_right,
|
53 |
+
softcap,
|
54 |
+
return_softmax,
|
55 |
+
gen,
|
56 |
+
)
|
57 |
+
|
58 |
+
|
59 |
+
def mha_varlen_fwd(
|
60 |
+
q: torch.Tensor,
|
61 |
+
k: torch.Tensor,
|
62 |
+
v: torch.Tensor,
|
63 |
+
cu_seqlens_q: torch.Tensor,
|
64 |
+
cu_seqlens_k: torch.Tensor,
|
65 |
+
out: Optional[torch.Tensor] = None,
|
66 |
+
seqused_k: Optional[torch.Tensor] = None,
|
67 |
+
leftpad_k: Optional[torch.Tensor] = None,
|
68 |
+
block_table: Optional[torch.Tensor] = None,
|
69 |
+
alibi_slopes: Optional[torch.Tensor] = None,
|
70 |
+
max_seqlen_q: int = 0,
|
71 |
+
max_seqlen_k: int = 0,
|
72 |
+
p_dropout: float = 0.0,
|
73 |
+
softmax_scale: float = 1.0,
|
74 |
+
zero_tensors: bool = False,
|
75 |
+
is_causal: bool = False,
|
76 |
+
window_size_left: int = -1,
|
77 |
+
window_size_right: int = -1,
|
78 |
+
softcap: float = 0.0,
|
79 |
+
return_softmax: bool = False,
|
80 |
+
gen: Optional[torch.Generator] = None,
|
81 |
+
) -> List[torch.Tensor]:
|
82 |
+
"""
|
83 |
+
Forward pass for multi-head attention with variable sequence lengths.
|
84 |
+
|
85 |
+
Args:
|
86 |
+
q: Query tensor of shape [total_q, num_heads, head_size]
|
87 |
+
k: Key tensor of shape [total_k, num_heads_k, head_size] or [num_blocks, page_block_size, num_heads_k, head_size]
|
88 |
+
v: Value tensor of shape [total_k, num_heads_k, head_size] or [num_blocks, page_block_size, num_heads_k, head_size]
|
89 |
+
cu_seqlens_q: Cumulative sequence lengths for queries of shape [batch_size+1]
|
90 |
+
cu_seqlens_k: Cumulative sequence lengths for keys of shape [batch_size+1]
|
91 |
+
out: Optional output tensor of shape [total_q, num_heads, head_size]
|
92 |
+
seqused_k: Optional tensor specifying how many keys to use per batch element [batch_size]
|
93 |
+
leftpad_k: Optional left padding for keys of shape [batch_size]
|
94 |
+
block_table: Optional block table of shape [batch_size, max_num_blocks_per_seq]
|
95 |
+
alibi_slopes: Optional ALiBi slopes tensor of shape [num_heads] or [batch_size, num_heads]
|
96 |
+
max_seqlen_q: Maximum sequence length for queries
|
97 |
+
max_seqlen_k: Maximum sequence length for keys
|
98 |
+
p_dropout: Dropout probability
|
99 |
+
softmax_scale: Scale factor for softmax
|
100 |
+
zero_tensors: Whether to zero tensors before computation
|
101 |
+
is_causal: Whether to use causal attention
|
102 |
+
window_size_left: Window size for left context (-1 for unlimited)
|
103 |
+
window_size_right: Window size for right context (-1 for unlimited)
|
104 |
+
softcap: Soft cap for attention weights
|
105 |
+
return_softmax: Whether to return softmax weights
|
106 |
+
gen: Optional random number generator
|
107 |
+
|
108 |
+
Returns:
|
109 |
+
List of tensors: [output, softmax_lse, (softmax if return_softmax)]
|
110 |
+
"""
|
111 |
+
return ops.mha_varlen_fwd(
|
112 |
+
q,
|
113 |
+
k,
|
114 |
+
v,
|
115 |
+
out,
|
116 |
+
cu_seqlens_q,
|
117 |
+
cu_seqlens_k,
|
118 |
+
seqused_k,
|
119 |
+
leftpad_k,
|
120 |
+
block_table,
|
121 |
+
alibi_slopes,
|
122 |
+
max_seqlen_q,
|
123 |
+
max_seqlen_k,
|
124 |
+
p_dropout,
|
125 |
+
softmax_scale,
|
126 |
+
zero_tensors,
|
127 |
+
is_causal,
|
128 |
+
window_size_left,
|
129 |
+
window_size_right,
|
130 |
+
softcap,
|
131 |
+
return_softmax,
|
132 |
+
gen,
|
133 |
+
)
|
134 |
+
|
135 |
+
|
136 |
+
def mha_bwd(
|
137 |
+
dout: torch.Tensor,
|
138 |
+
q: torch.Tensor,
|
139 |
+
k: torch.Tensor,
|
140 |
+
v: torch.Tensor,
|
141 |
+
out: torch.Tensor,
|
142 |
+
softmax_lse: torch.Tensor,
|
143 |
+
dq: Optional[torch.Tensor] = None,
|
144 |
+
dk: Optional[torch.Tensor] = None,
|
145 |
+
dv: Optional[torch.Tensor] = None,
|
146 |
+
alibi_slopes: Optional[torch.Tensor] = None,
|
147 |
+
p_dropout: float = 0.0,
|
148 |
+
softmax_scale: float = 1.0,
|
149 |
+
is_causal: bool = False,
|
150 |
+
window_size_left: int = -1,
|
151 |
+
window_size_right: int = -1,
|
152 |
+
softcap: float = 0.0,
|
153 |
+
deterministic: bool = False,
|
154 |
+
gen: Optional[torch.Generator] = None,
|
155 |
+
rng_state: Optional[torch.Tensor] = None,
|
156 |
+
) -> List[torch.Tensor]:
|
157 |
+
"""
|
158 |
+
Backward pass for multi-head attention.
|
159 |
+
|
160 |
+
Args:
|
161 |
+
dout: Gradient tensor of shape [batch_size, seqlen_q, num_heads, head_size]
|
162 |
+
q: Query tensor of shape [batch_size, seqlen_q, num_heads, head_size]
|
163 |
+
k: Key tensor of shape [batch_size, seqlen_k, num_heads_k, head_size]
|
164 |
+
v: Value tensor of shape [batch_size, seqlen_k, num_heads_k, head_size]
|
165 |
+
out: Output tensor from forward pass of shape [batch_size, seqlen_q, num_heads, head_size]
|
166 |
+
softmax_lse: Log-sum-exp values from forward pass of shape [batch_size, num_heads, seqlen_q]
|
167 |
+
dq: Optional gradient tensor for queries, same shape as q
|
168 |
+
dk: Optional gradient tensor for keys, same shape as k
|
169 |
+
dv: Optional gradient tensor for values, same shape as v
|
170 |
+
alibi_slopes: Optional ALiBi slopes tensor of shape [num_heads] or [batch_size, num_heads]
|
171 |
+
p_dropout: Dropout probability
|
172 |
+
softmax_scale: Scale factor for softmax
|
173 |
+
is_causal: Whether to use causal attention
|
174 |
+
window_size_left: Window size for left context (-1 for unlimited)
|
175 |
+
window_size_right: Window size for right context (-1 for unlimited)
|
176 |
+
softcap: Soft cap for attention weights
|
177 |
+
deterministic: Whether to use deterministic algorithms
|
178 |
+
gen: Optional random number generator
|
179 |
+
rng_state: Optional RNG state from forward pass
|
180 |
+
|
181 |
+
Returns:
|
182 |
+
List of tensors: [dq, dk, dv]
|
183 |
+
"""
|
184 |
+
return ops.mha_bwd(
|
185 |
+
dout,
|
186 |
+
q,
|
187 |
+
k,
|
188 |
+
v,
|
189 |
+
out,
|
190 |
+
softmax_lse,
|
191 |
+
dq,
|
192 |
+
dk,
|
193 |
+
dv,
|
194 |
+
alibi_slopes,
|
195 |
+
p_dropout,
|
196 |
+
softmax_scale,
|
197 |
+
is_causal,
|
198 |
+
window_size_left,
|
199 |
+
window_size_right,
|
200 |
+
softcap,
|
201 |
+
deterministic,
|
202 |
+
gen,
|
203 |
+
rng_state,
|
204 |
+
)
|
205 |
+
|
206 |
+
|
207 |
+
def mha_varlen_bwd(
|
208 |
+
dout: torch.Tensor,
|
209 |
+
q: torch.Tensor,
|
210 |
+
k: torch.Tensor,
|
211 |
+
v: torch.Tensor,
|
212 |
+
out: torch.Tensor,
|
213 |
+
softmax_lse: torch.Tensor,
|
214 |
+
cu_seqlens_q: torch.Tensor,
|
215 |
+
cu_seqlens_k: torch.Tensor,
|
216 |
+
dq: Optional[torch.Tensor] = None,
|
217 |
+
dk: Optional[torch.Tensor] = None,
|
218 |
+
dv: Optional[torch.Tensor] = None,
|
219 |
+
alibi_slopes: Optional[torch.Tensor] = None,
|
220 |
+
max_seqlen_q: int = 0,
|
221 |
+
max_seqlen_k: int = 0,
|
222 |
+
p_dropout: float = 0.0,
|
223 |
+
softmax_scale: float = 1.0,
|
224 |
+
zero_tensors: bool = False,
|
225 |
+
is_causal: bool = False,
|
226 |
+
window_size_left: int = -1,
|
227 |
+
window_size_right: int = -1,
|
228 |
+
softcap: float = 0.0,
|
229 |
+
deterministic: bool = False,
|
230 |
+
gen: Optional[torch.Generator] = None,
|
231 |
+
rng_state: Optional[torch.Tensor] = None,
|
232 |
+
) -> List[torch.Tensor]:
|
233 |
+
"""
|
234 |
+
Backward pass for multi-head attention with variable sequence lengths.
|
235 |
+
|
236 |
+
Args:
|
237 |
+
dout: Gradient tensor of shape [batch_size, seqlen_q, num_heads, head_size]
|
238 |
+
q: Query tensor of shape [batch_size, seqlen_q, num_heads, head_size]
|
239 |
+
k: Key tensor of shape [batch_size, seqlen_k, num_heads_k, head_size]
|
240 |
+
v: Value tensor of shape [batch_size, seqlen_k, num_heads_k, head_size]
|
241 |
+
out: Output tensor from forward pass of shape [batch_size, seqlen_q, num_heads, head_size]
|
242 |
+
softmax_lse: Log-sum-exp values from forward pass of shape [batch_size, num_heads, seqlen_q]
|
243 |
+
cu_seqlens_q: Cumulative sequence lengths for queries of shape [batch_size+1]
|
244 |
+
cu_seqlens_k: Cumulative sequence lengths for keys of shape [batch_size+1]
|
245 |
+
dq: Optional gradient tensor for queries, same shape as q
|
246 |
+
dk: Optional gradient tensor for keys, same shape as k
|
247 |
+
dv: Optional gradient tensor for values, same shape as v
|
248 |
+
alibi_slopes: Optional ALiBi slopes tensor of shape [num_heads] or [batch_size, num_heads]
|
249 |
+
max_seqlen_q: Maximum sequence length for queries
|
250 |
+
max_seqlen_k: Maximum sequence length for keys
|
251 |
+
p_dropout: Dropout probability
|
252 |
+
softmax_scale: Scale factor for softmax
|
253 |
+
zero_tensors: Whether to zero tensors before computation
|
254 |
+
is_causal: Whether to use causal attention
|
255 |
+
window_size_left: Window size for left context (-1 for unlimited)
|
256 |
+
window_size_right: Window size for right context (-1 for unlimited)
|
257 |
+
softcap: Soft cap for attention weights
|
258 |
+
deterministic: Whether to use deterministic algorithms
|
259 |
+
gen: Optional random number generator
|
260 |
+
rng_state: Optional RNG state from forward pass
|
261 |
+
|
262 |
+
Returns:
|
263 |
+
List of tensors: [dq, dk, dv]
|
264 |
+
"""
|
265 |
+
return ops.mha_varlen_bwd(
|
266 |
+
dout,
|
267 |
+
q,
|
268 |
+
k,
|
269 |
+
v,
|
270 |
+
out,
|
271 |
+
softmax_lse,
|
272 |
+
dq,
|
273 |
+
dk,
|
274 |
+
dv,
|
275 |
+
cu_seqlens_q,
|
276 |
+
cu_seqlens_k,
|
277 |
+
alibi_slopes,
|
278 |
+
max_seqlen_q,
|
279 |
+
max_seqlen_k,
|
280 |
+
p_dropout,
|
281 |
+
softmax_scale,
|
282 |
+
zero_tensors,
|
283 |
+
is_causal,
|
284 |
+
window_size_left,
|
285 |
+
window_size_right,
|
286 |
+
softcap,
|
287 |
+
deterministic,
|
288 |
+
gen,
|
289 |
+
rng_state,
|
290 |
+
)
|
291 |
+
|
292 |
+
|
293 |
+
def mha_fwd_kvcache(
|
294 |
+
q: torch.Tensor,
|
295 |
+
kcache: torch.Tensor,
|
296 |
+
vcache: torch.Tensor,
|
297 |
+
k: Optional[torch.Tensor] = None,
|
298 |
+
v: Optional[torch.Tensor] = None,
|
299 |
+
seqlens_k: Optional[torch.Tensor] = None,
|
300 |
+
rotary_cos: Optional[torch.Tensor] = None,
|
301 |
+
rotary_sin: Optional[torch.Tensor] = None,
|
302 |
+
cache_batch_idx: Optional[torch.Tensor] = None,
|
303 |
+
leftpad_k: Optional[torch.Tensor] = None,
|
304 |
+
block_table: Optional[torch.Tensor] = None,
|
305 |
+
alibi_slopes: Optional[torch.Tensor] = None,
|
306 |
+
out: Optional[torch.Tensor] = None,
|
307 |
+
softmax_scale: float = 1.0,
|
308 |
+
is_causal: bool = False,
|
309 |
+
window_size_left: int = -1,
|
310 |
+
window_size_right: int = -1,
|
311 |
+
softcap: float = 0.0,
|
312 |
+
is_rotary_interleaved: bool = False,
|
313 |
+
num_splits: int = 1,
|
314 |
+
) -> List[torch.Tensor]:
|
315 |
+
"""
|
316 |
+
Forward pass for multi-head attention with KV cache.
|
317 |
+
|
318 |
+
Args:
|
319 |
+
q: Query tensor of shape [batch_size, seqlen_q, num_heads, head_size]
|
320 |
+
kcache: Key cache tensor of shape [batch_size_c, seqlen_k, num_heads_k, head_size] or [num_blocks, page_block_size, num_heads_k, head_size]
|
321 |
+
vcache: Value cache tensor of shape [batch_size_c, seqlen_k, num_heads_k, head_size] or [num_blocks, page_block_size, num_heads_k, head_size]
|
322 |
+
k: Optional new keys tensor of shape [batch_size, seqlen_knew, num_heads_k, head_size]
|
323 |
+
v: Optional new values tensor of shape [batch_size, seqlen_knew, num_heads_k, head_size]
|
324 |
+
seqlens_k: Optional sequence lengths for keys of shape [batch_size]
|
325 |
+
rotary_cos: Optional rotary cosine tensor of shape [seqlen_ro, rotary_dim/2]
|
326 |
+
rotary_sin: Optional rotary sine tensor of shape [seqlen_ro, rotary_dim/2]
|
327 |
+
cache_batch_idx: Optional indices to index into the KV cache
|
328 |
+
leftpad_k: Optional left padding for keys of shape [batch_size]
|
329 |
+
block_table: Optional block table of shape [batch_size, max_num_blocks_per_seq]
|
330 |
+
alibi_slopes: Optional ALiBi slopes tensor of shape [num_heads] or [batch_size, num_heads]
|
331 |
+
out: Optional output tensor, same shape as q
|
332 |
+
softmax_scale: Scale factor for softmax
|
333 |
+
is_causal: Whether to use causal attention
|
334 |
+
window_size_left: Window size for left context (-1 for unlimited)
|
335 |
+
window_size_right: Window size for right context (-1 for unlimited)
|
336 |
+
softcap: Soft cap for attention weights
|
337 |
+
is_rotary_interleaved: Whether rotary embeddings are interleaved
|
338 |
+
num_splits: Number of splits for computation
|
339 |
+
|
340 |
+
Returns:
|
341 |
+
List of tensors: [output, softmax_lse]
|
342 |
+
"""
|
343 |
+
return ops.mha_fwd_kvcache(
|
344 |
+
q,
|
345 |
+
kcache,
|
346 |
+
vcache,
|
347 |
+
k,
|
348 |
+
v,
|
349 |
+
seqlens_k,
|
350 |
+
rotary_cos,
|
351 |
+
rotary_sin,
|
352 |
+
cache_batch_idx,
|
353 |
+
leftpad_k,
|
354 |
+
block_table,
|
355 |
+
alibi_slopes,
|
356 |
+
out,
|
357 |
+
softmax_scale,
|
358 |
+
is_causal,
|
359 |
+
window_size_left,
|
360 |
+
window_size_right,
|
361 |
+
softcap,
|
362 |
+
is_rotary_interleaved,
|
363 |
+
num_splits,
|
364 |
+
)
|
build/torch27-cxx11-cu126-x86_64-linux/flash_attn/_flash_attn_dd2f0f9.abi3.so
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:e7da95f182ca7f57b45cfe9387045d78397c312283d8a5eecd9bce96e6888ea8
|
3 |
+
size 646613312
|
build/torch27-cxx11-cu126-x86_64-linux/flash_attn/_ops.py
ADDED
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
from . import _flash_attn_dd2f0f9
|
3 |
+
ops = torch.ops._flash_attn_dd2f0f9
|
4 |
+
|
5 |
+
def add_op_namespace_prefix(op_name: str):
|
6 |
+
"""
|
7 |
+
Prefix op by namespace.
|
8 |
+
"""
|
9 |
+
return f"_flash_attn_dd2f0f9::{op_name}"
|
build/torch27-cxx11-cu128-x86_64-linux/flash_attn/__init__.py
ADDED
@@ -0,0 +1,364 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Optional, List
|
2 |
+
import torch
|
3 |
+
from ._ops import ops
|
4 |
+
|
5 |
+
|
6 |
+
def mha_fwd(
|
7 |
+
q: torch.Tensor,
|
8 |
+
k: torch.Tensor,
|
9 |
+
v: torch.Tensor,
|
10 |
+
out: Optional[torch.Tensor] = None,
|
11 |
+
alibi_slopes: Optional[torch.Tensor] = None,
|
12 |
+
p_dropout: float = 0.0,
|
13 |
+
softmax_scale: float = 1.0,
|
14 |
+
is_causal: bool = False,
|
15 |
+
window_size_left: int = -1,
|
16 |
+
window_size_right: int = -1,
|
17 |
+
softcap: float = 0.0,
|
18 |
+
return_softmax: bool = False,
|
19 |
+
gen: Optional[torch.Generator] = None,
|
20 |
+
) -> List[torch.Tensor]:
|
21 |
+
"""
|
22 |
+
Forward pass for multi-head attention.
|
23 |
+
|
24 |
+
Args:
|
25 |
+
q: Query tensor of shape [batch_size, seqlen_q, num_heads, head_size]
|
26 |
+
k: Key tensor of shape [batch_size, seqlen_k, num_heads_k, head_size]
|
27 |
+
v: Value tensor of shape [batch_size, seqlen_k, num_heads_k, head_size]
|
28 |
+
out: Optional output tensor, same shape as q
|
29 |
+
alibi_slopes: Optional ALiBi slopes tensor of shape [num_heads] or [batch_size, num_heads]
|
30 |
+
p_dropout: Dropout probability
|
31 |
+
softmax_scale: Scale factor for softmax
|
32 |
+
is_causal: Whether to use causal attention
|
33 |
+
window_size_left: Window size for left context (-1 for unlimited)
|
34 |
+
window_size_right: Window size for right context (-1 for unlimited)
|
35 |
+
softcap: Soft cap for attention weights
|
36 |
+
return_softmax: Whether to return softmax weights
|
37 |
+
gen: Optional random number generator
|
38 |
+
|
39 |
+
Returns:
|
40 |
+
List of tensors: [output, softmax_lse, (softmax if return_softmax)]
|
41 |
+
"""
|
42 |
+
return ops.mha_fwd(
|
43 |
+
q,
|
44 |
+
k,
|
45 |
+
v,
|
46 |
+
out,
|
47 |
+
alibi_slopes,
|
48 |
+
p_dropout,
|
49 |
+
softmax_scale,
|
50 |
+
is_causal,
|
51 |
+
window_size_left,
|
52 |
+
window_size_right,
|
53 |
+
softcap,
|
54 |
+
return_softmax,
|
55 |
+
gen,
|
56 |
+
)
|
57 |
+
|
58 |
+
|
59 |
+
def mha_varlen_fwd(
|
60 |
+
q: torch.Tensor,
|
61 |
+
k: torch.Tensor,
|
62 |
+
v: torch.Tensor,
|
63 |
+
cu_seqlens_q: torch.Tensor,
|
64 |
+
cu_seqlens_k: torch.Tensor,
|
65 |
+
out: Optional[torch.Tensor] = None,
|
66 |
+
seqused_k: Optional[torch.Tensor] = None,
|
67 |
+
leftpad_k: Optional[torch.Tensor] = None,
|
68 |
+
block_table: Optional[torch.Tensor] = None,
|
69 |
+
alibi_slopes: Optional[torch.Tensor] = None,
|
70 |
+
max_seqlen_q: int = 0,
|
71 |
+
max_seqlen_k: int = 0,
|
72 |
+
p_dropout: float = 0.0,
|
73 |
+
softmax_scale: float = 1.0,
|
74 |
+
zero_tensors: bool = False,
|
75 |
+
is_causal: bool = False,
|
76 |
+
window_size_left: int = -1,
|
77 |
+
window_size_right: int = -1,
|
78 |
+
softcap: float = 0.0,
|
79 |
+
return_softmax: bool = False,
|
80 |
+
gen: Optional[torch.Generator] = None,
|
81 |
+
) -> List[torch.Tensor]:
|
82 |
+
"""
|
83 |
+
Forward pass for multi-head attention with variable sequence lengths.
|
84 |
+
|
85 |
+
Args:
|
86 |
+
q: Query tensor of shape [total_q, num_heads, head_size]
|
87 |
+
k: Key tensor of shape [total_k, num_heads_k, head_size] or [num_blocks, page_block_size, num_heads_k, head_size]
|
88 |
+
v: Value tensor of shape [total_k, num_heads_k, head_size] or [num_blocks, page_block_size, num_heads_k, head_size]
|
89 |
+
cu_seqlens_q: Cumulative sequence lengths for queries of shape [batch_size+1]
|
90 |
+
cu_seqlens_k: Cumulative sequence lengths for keys of shape [batch_size+1]
|
91 |
+
out: Optional output tensor of shape [total_q, num_heads, head_size]
|
92 |
+
seqused_k: Optional tensor specifying how many keys to use per batch element [batch_size]
|
93 |
+
leftpad_k: Optional left padding for keys of shape [batch_size]
|
94 |
+
block_table: Optional block table of shape [batch_size, max_num_blocks_per_seq]
|
95 |
+
alibi_slopes: Optional ALiBi slopes tensor of shape [num_heads] or [batch_size, num_heads]
|
96 |
+
max_seqlen_q: Maximum sequence length for queries
|
97 |
+
max_seqlen_k: Maximum sequence length for keys
|
98 |
+
p_dropout: Dropout probability
|
99 |
+
softmax_scale: Scale factor for softmax
|
100 |
+
zero_tensors: Whether to zero tensors before computation
|
101 |
+
is_causal: Whether to use causal attention
|
102 |
+
window_size_left: Window size for left context (-1 for unlimited)
|
103 |
+
window_size_right: Window size for right context (-1 for unlimited)
|
104 |
+
softcap: Soft cap for attention weights
|
105 |
+
return_softmax: Whether to return softmax weights
|
106 |
+
gen: Optional random number generator
|
107 |
+
|
108 |
+
Returns:
|
109 |
+
List of tensors: [output, softmax_lse, (softmax if return_softmax)]
|
110 |
+
"""
|
111 |
+
return ops.mha_varlen_fwd(
|
112 |
+
q,
|
113 |
+
k,
|
114 |
+
v,
|
115 |
+
out,
|
116 |
+
cu_seqlens_q,
|
117 |
+
cu_seqlens_k,
|
118 |
+
seqused_k,
|
119 |
+
leftpad_k,
|
120 |
+
block_table,
|
121 |
+
alibi_slopes,
|
122 |
+
max_seqlen_q,
|
123 |
+
max_seqlen_k,
|
124 |
+
p_dropout,
|
125 |
+
softmax_scale,
|
126 |
+
zero_tensors,
|
127 |
+
is_causal,
|
128 |
+
window_size_left,
|
129 |
+
window_size_right,
|
130 |
+
softcap,
|
131 |
+
return_softmax,
|
132 |
+
gen,
|
133 |
+
)
|
134 |
+
|
135 |
+
|
136 |
+
def mha_bwd(
|
137 |
+
dout: torch.Tensor,
|
138 |
+
q: torch.Tensor,
|
139 |
+
k: torch.Tensor,
|
140 |
+
v: torch.Tensor,
|
141 |
+
out: torch.Tensor,
|
142 |
+
softmax_lse: torch.Tensor,
|
143 |
+
dq: Optional[torch.Tensor] = None,
|
144 |
+
dk: Optional[torch.Tensor] = None,
|
145 |
+
dv: Optional[torch.Tensor] = None,
|
146 |
+
alibi_slopes: Optional[torch.Tensor] = None,
|
147 |
+
p_dropout: float = 0.0,
|
148 |
+
softmax_scale: float = 1.0,
|
149 |
+
is_causal: bool = False,
|
150 |
+
window_size_left: int = -1,
|
151 |
+
window_size_right: int = -1,
|
152 |
+
softcap: float = 0.0,
|
153 |
+
deterministic: bool = False,
|
154 |
+
gen: Optional[torch.Generator] = None,
|
155 |
+
rng_state: Optional[torch.Tensor] = None,
|
156 |
+
) -> List[torch.Tensor]:
|
157 |
+
"""
|
158 |
+
Backward pass for multi-head attention.
|
159 |
+
|
160 |
+
Args:
|
161 |
+
dout: Gradient tensor of shape [batch_size, seqlen_q, num_heads, head_size]
|
162 |
+
q: Query tensor of shape [batch_size, seqlen_q, num_heads, head_size]
|
163 |
+
k: Key tensor of shape [batch_size, seqlen_k, num_heads_k, head_size]
|
164 |
+
v: Value tensor of shape [batch_size, seqlen_k, num_heads_k, head_size]
|
165 |
+
out: Output tensor from forward pass of shape [batch_size, seqlen_q, num_heads, head_size]
|
166 |
+
softmax_lse: Log-sum-exp values from forward pass of shape [batch_size, num_heads, seqlen_q]
|
167 |
+
dq: Optional gradient tensor for queries, same shape as q
|
168 |
+
dk: Optional gradient tensor for keys, same shape as k
|
169 |
+
dv: Optional gradient tensor for values, same shape as v
|
170 |
+
alibi_slopes: Optional ALiBi slopes tensor of shape [num_heads] or [batch_size, num_heads]
|
171 |
+
p_dropout: Dropout probability
|
172 |
+
softmax_scale: Scale factor for softmax
|
173 |
+
is_causal: Whether to use causal attention
|
174 |
+
window_size_left: Window size for left context (-1 for unlimited)
|
175 |
+
window_size_right: Window size for right context (-1 for unlimited)
|
176 |
+
softcap: Soft cap for attention weights
|
177 |
+
deterministic: Whether to use deterministic algorithms
|
178 |
+
gen: Optional random number generator
|
179 |
+
rng_state: Optional RNG state from forward pass
|
180 |
+
|
181 |
+
Returns:
|
182 |
+
List of tensors: [dq, dk, dv]
|
183 |
+
"""
|
184 |
+
return ops.mha_bwd(
|
185 |
+
dout,
|
186 |
+
q,
|
187 |
+
k,
|
188 |
+
v,
|
189 |
+
out,
|
190 |
+
softmax_lse,
|
191 |
+
dq,
|
192 |
+
dk,
|
193 |
+
dv,
|
194 |
+
alibi_slopes,
|
195 |
+
p_dropout,
|
196 |
+
softmax_scale,
|
197 |
+
is_causal,
|
198 |
+
window_size_left,
|
199 |
+
window_size_right,
|
200 |
+
softcap,
|
201 |
+
deterministic,
|
202 |
+
gen,
|
203 |
+
rng_state,
|
204 |
+
)
|
205 |
+
|
206 |
+
|
207 |
+
def mha_varlen_bwd(
|
208 |
+
dout: torch.Tensor,
|
209 |
+
q: torch.Tensor,
|
210 |
+
k: torch.Tensor,
|
211 |
+
v: torch.Tensor,
|
212 |
+
out: torch.Tensor,
|
213 |
+
softmax_lse: torch.Tensor,
|
214 |
+
cu_seqlens_q: torch.Tensor,
|
215 |
+
cu_seqlens_k: torch.Tensor,
|
216 |
+
dq: Optional[torch.Tensor] = None,
|
217 |
+
dk: Optional[torch.Tensor] = None,
|
218 |
+
dv: Optional[torch.Tensor] = None,
|
219 |
+
alibi_slopes: Optional[torch.Tensor] = None,
|
220 |
+
max_seqlen_q: int = 0,
|
221 |
+
max_seqlen_k: int = 0,
|
222 |
+
p_dropout: float = 0.0,
|
223 |
+
softmax_scale: float = 1.0,
|
224 |
+
zero_tensors: bool = False,
|
225 |
+
is_causal: bool = False,
|
226 |
+
window_size_left: int = -1,
|
227 |
+
window_size_right: int = -1,
|
228 |
+
softcap: float = 0.0,
|
229 |
+
deterministic: bool = False,
|
230 |
+
gen: Optional[torch.Generator] = None,
|
231 |
+
rng_state: Optional[torch.Tensor] = None,
|
232 |
+
) -> List[torch.Tensor]:
|
233 |
+
"""
|
234 |
+
Backward pass for multi-head attention with variable sequence lengths.
|
235 |
+
|
236 |
+
Args:
|
237 |
+
dout: Gradient tensor of shape [batch_size, seqlen_q, num_heads, head_size]
|
238 |
+
q: Query tensor of shape [batch_size, seqlen_q, num_heads, head_size]
|
239 |
+
k: Key tensor of shape [batch_size, seqlen_k, num_heads_k, head_size]
|
240 |
+
v: Value tensor of shape [batch_size, seqlen_k, num_heads_k, head_size]
|
241 |
+
out: Output tensor from forward pass of shape [batch_size, seqlen_q, num_heads, head_size]
|
242 |
+
softmax_lse: Log-sum-exp values from forward pass of shape [batch_size, num_heads, seqlen_q]
|
243 |
+
cu_seqlens_q: Cumulative sequence lengths for queries of shape [batch_size+1]
|
244 |
+
cu_seqlens_k: Cumulative sequence lengths for keys of shape [batch_size+1]
|
245 |
+
dq: Optional gradient tensor for queries, same shape as q
|
246 |
+
dk: Optional gradient tensor for keys, same shape as k
|
247 |
+
dv: Optional gradient tensor for values, same shape as v
|
248 |
+
alibi_slopes: Optional ALiBi slopes tensor of shape [num_heads] or [batch_size, num_heads]
|
249 |
+
max_seqlen_q: Maximum sequence length for queries
|
250 |
+
max_seqlen_k: Maximum sequence length for keys
|
251 |
+
p_dropout: Dropout probability
|
252 |
+
softmax_scale: Scale factor for softmax
|
253 |
+
zero_tensors: Whether to zero tensors before computation
|
254 |
+
is_causal: Whether to use causal attention
|
255 |
+
window_size_left: Window size for left context (-1 for unlimited)
|
256 |
+
window_size_right: Window size for right context (-1 for unlimited)
|
257 |
+
softcap: Soft cap for attention weights
|
258 |
+
deterministic: Whether to use deterministic algorithms
|
259 |
+
gen: Optional random number generator
|
260 |
+
rng_state: Optional RNG state from forward pass
|
261 |
+
|
262 |
+
Returns:
|
263 |
+
List of tensors: [dq, dk, dv]
|
264 |
+
"""
|
265 |
+
return ops.mha_varlen_bwd(
|
266 |
+
dout,
|
267 |
+
q,
|
268 |
+
k,
|
269 |
+
v,
|
270 |
+
out,
|
271 |
+
softmax_lse,
|
272 |
+
dq,
|
273 |
+
dk,
|
274 |
+
dv,
|
275 |
+
cu_seqlens_q,
|
276 |
+
cu_seqlens_k,
|
277 |
+
alibi_slopes,
|
278 |
+
max_seqlen_q,
|
279 |
+
max_seqlen_k,
|
280 |
+
p_dropout,
|
281 |
+
softmax_scale,
|
282 |
+
zero_tensors,
|
283 |
+
is_causal,
|
284 |
+
window_size_left,
|
285 |
+
window_size_right,
|
286 |
+
softcap,
|
287 |
+
deterministic,
|
288 |
+
gen,
|
289 |
+
rng_state,
|
290 |
+
)
|
291 |
+
|
292 |
+
|
293 |
+
def mha_fwd_kvcache(
|
294 |
+
q: torch.Tensor,
|
295 |
+
kcache: torch.Tensor,
|
296 |
+
vcache: torch.Tensor,
|
297 |
+
k: Optional[torch.Tensor] = None,
|
298 |
+
v: Optional[torch.Tensor] = None,
|
299 |
+
seqlens_k: Optional[torch.Tensor] = None,
|
300 |
+
rotary_cos: Optional[torch.Tensor] = None,
|
301 |
+
rotary_sin: Optional[torch.Tensor] = None,
|
302 |
+
cache_batch_idx: Optional[torch.Tensor] = None,
|
303 |
+
leftpad_k: Optional[torch.Tensor] = None,
|
304 |
+
block_table: Optional[torch.Tensor] = None,
|
305 |
+
alibi_slopes: Optional[torch.Tensor] = None,
|
306 |
+
out: Optional[torch.Tensor] = None,
|
307 |
+
softmax_scale: float = 1.0,
|
308 |
+
is_causal: bool = False,
|
309 |
+
window_size_left: int = -1,
|
310 |
+
window_size_right: int = -1,
|
311 |
+
softcap: float = 0.0,
|
312 |
+
is_rotary_interleaved: bool = False,
|
313 |
+
num_splits: int = 1,
|
314 |
+
) -> List[torch.Tensor]:
|
315 |
+
"""
|
316 |
+
Forward pass for multi-head attention with KV cache.
|
317 |
+
|
318 |
+
Args:
|
319 |
+
q: Query tensor of shape [batch_size, seqlen_q, num_heads, head_size]
|
320 |
+
kcache: Key cache tensor of shape [batch_size_c, seqlen_k, num_heads_k, head_size] or [num_blocks, page_block_size, num_heads_k, head_size]
|
321 |
+
vcache: Value cache tensor of shape [batch_size_c, seqlen_k, num_heads_k, head_size] or [num_blocks, page_block_size, num_heads_k, head_size]
|
322 |
+
k: Optional new keys tensor of shape [batch_size, seqlen_knew, num_heads_k, head_size]
|
323 |
+
v: Optional new values tensor of shape [batch_size, seqlen_knew, num_heads_k, head_size]
|
324 |
+
seqlens_k: Optional sequence lengths for keys of shape [batch_size]
|
325 |
+
rotary_cos: Optional rotary cosine tensor of shape [seqlen_ro, rotary_dim/2]
|
326 |
+
rotary_sin: Optional rotary sine tensor of shape [seqlen_ro, rotary_dim/2]
|
327 |
+
cache_batch_idx: Optional indices to index into the KV cache
|
328 |
+
leftpad_k: Optional left padding for keys of shape [batch_size]
|
329 |
+
block_table: Optional block table of shape [batch_size, max_num_blocks_per_seq]
|
330 |
+
alibi_slopes: Optional ALiBi slopes tensor of shape [num_heads] or [batch_size, num_heads]
|
331 |
+
out: Optional output tensor, same shape as q
|
332 |
+
softmax_scale: Scale factor for softmax
|
333 |
+
is_causal: Whether to use causal attention
|
334 |
+
window_size_left: Window size for left context (-1 for unlimited)
|
335 |
+
window_size_right: Window size for right context (-1 for unlimited)
|
336 |
+
softcap: Soft cap for attention weights
|
337 |
+
is_rotary_interleaved: Whether rotary embeddings are interleaved
|
338 |
+
num_splits: Number of splits for computation
|
339 |
+
|
340 |
+
Returns:
|
341 |
+
List of tensors: [output, softmax_lse]
|
342 |
+
"""
|
343 |
+
return ops.mha_fwd_kvcache(
|
344 |
+
q,
|
345 |
+
kcache,
|
346 |
+
vcache,
|
347 |
+
k,
|
348 |
+
v,
|
349 |
+
seqlens_k,
|
350 |
+
rotary_cos,
|
351 |
+
rotary_sin,
|
352 |
+
cache_batch_idx,
|
353 |
+
leftpad_k,
|
354 |
+
block_table,
|
355 |
+
alibi_slopes,
|
356 |
+
out,
|
357 |
+
softmax_scale,
|
358 |
+
is_causal,
|
359 |
+
window_size_left,
|
360 |
+
window_size_right,
|
361 |
+
softcap,
|
362 |
+
is_rotary_interleaved,
|
363 |
+
num_splits,
|
364 |
+
)
|
build/torch27-cxx11-cu128-x86_64-linux/flash_attn/_flash_attn_dd2f0f9.abi3.so
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:3ba8c31bf3488a6f0a93e2d5d83d28a27daa26c156ed357ba5443ac66e3809fc
|
3 |
+
size 1502967480
|
build/torch27-cxx11-cu128-x86_64-linux/flash_attn/_ops.py
ADDED
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
from . import _flash_attn_dd2f0f9
|
3 |
+
ops = torch.ops._flash_attn_dd2f0f9
|
4 |
+
|
5 |
+
def add_op_namespace_prefix(op_name: str):
|
6 |
+
"""
|
7 |
+
Prefix op by namespace.
|
8 |
+
"""
|
9 |
+
return f"_flash_attn_dd2f0f9::{op_name}"
|