peonist commited on
Commit
dec83dd
·
verified ·
1 Parent(s): 670d7ae

decider-0.8b: the NPU set

Browse files
.gitattributes CHANGED
@@ -33,3 +33,11 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
 
 
 
 
 
 
 
 
 
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
+ decider-0.8b.hnpw filter=lfs diff=lfs merge=lfs -text
37
+ devices/consts.hnpb filter=lfs diff=lfs merge=lfs -text
38
+ devices/u0.elf filter=lfs diff=lfs merge=lfs -text
39
+ devices/u1.elf filter=lfs diff=lfs merge=lfs -text
40
+ devices/u2.elf filter=lfs diff=lfs merge=lfs -text
41
+ devices/u3.elf filter=lfs diff=lfs merge=lfs -text
42
+ devices/u4.elf filter=lfs diff=lfs merge=lfs -text
43
+ tokenizer/tokenizer.json filter=lfs diff=lfs merge=lfs -text
decider-0.8b.hnpw ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5950f4dbb472fe0d9562e5df5723c102a526069b2f777a77d3559acf34ef1067
3
+ size 2269182920
devices/NOTICE.md ADDED
@@ -0,0 +1,72 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # halogen: notices for the NPU (HALOGEN_NPU_MODELS)
2
+
3
+ This covers two things: the NPU engine binary in this image
4
+ (`/usr/local/bin/halogen-npu`), and the NPU files it loads at run time. Those
5
+ files are each model's `devices/` directory and weight file, and they are
6
+ fetched like model weights, not carried in the image. A copy of this file
7
+ travels with every published set of NPU files, as `devices/NOTICE.md`.
8
+
9
+ ## In the binary: XRT's public headers (Apache-2.0)
10
+
11
+ `halogen-npu` is compiled against the public C++ headers of AMD's XRT, commit
12
+ `0026185de81a179dc7d886197d3e35f9a179b4e8` (release 2.25.37). Code from those
13
+ headers, such as inline functions and templates, is part of the binary. XRT's
14
+ license and notice are reproduced in full beside this file:
15
+
16
+ | file | what |
17
+ | --- | --- |
18
+ | `XRT-LICENSE` | XRT's LICENSE at that commit: its userspace code under the Apache License, Version 2.0 |
19
+ | `XRT-NOTICE` | XRT's NOTICE at that commit |
20
+
21
+ Every header the binary includes names Apache-2.0. Four of them (`xclbin.h`,
22
+ `ert.h`, `xrt_mem.h`, `xclerr.h`) are shared with XRT's kernel drivers and
23
+ also name GPL-2.0. XRT's LICENSE places all of its userspace code under
24
+ Apache-2.0.
25
+
26
+ The XRT libraries themselves are **not** in this image. The binary uses the
27
+ host's XRT, mounted read-only at `/opt/xilinx/xrt`.
28
+
29
+ ## In the NPU files: the AIE API headers (MIT)
30
+
31
+ The NPU programs in each model's `devices/*.elf` are compiled with AMD's AIE
32
+ API headers (`aie_api`), whose templates become part of the programs.
33
+
34
+ ```
35
+ MIT License
36
+
37
+ Copyright (C) 2022 Xilinx, Inc.
38
+ Copyright (C) 2022-2026 Advanced Micro Devices, Inc.
39
+
40
+ Permission is hereby granted, free of charge, to any person obtaining a copy
41
+ of this software and associated documentation files (the "Software"), to deal
42
+ in the Software without restriction, including without limitation the rights
43
+ to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
44
+ copies of the Software, and to permit persons to whom the Software is
45
+ furnished to do so, subject to the following conditions:
46
+
47
+ The above copyright notice and this permission notice shall be included in
48
+ all copies or substantial portions of the Software.
49
+
50
+ THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
51
+ IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
52
+ FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
53
+ AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
54
+ LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
55
+ OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
56
+ SOFTWARE.
57
+ ```
58
+
59
+ Also in the NPU files, and listed here although their license does not require
60
+ a notice in object form:
61
+
62
+ - The NPU programs' startup code (`crt0`, `crt1`) from AMD's llvm-aie. License:
63
+ Apache License 2.0 WITH LLVM-exception.
64
+ - The device configuration generated by AMD's mlir-aie. License: Apache
65
+ License 2.0 WITH LLVM-exception.
66
+
67
+ ## Not included
68
+
69
+ - **XRT's libraries, the NPU driver and the NPU firmware** are the host's. None
70
+ of them is in this image or in the NPU files.
71
+ - **The models' weights** are licensed by their authors. Each published model
72
+ carries its own license and card beside its files.
devices/consts.hnpb ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6fd898879cc544ebdf90a491e8508e7bde946e4326ea61fe6467543e6f4ea611
3
+ size 1645016
devices/devices.hnpm ADDED
@@ -0,0 +1,246 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ hnpm 1 devices
2
+ family name=qwen35_hybrid sig=h1024.i3584.v248320.a8x2x256.g16x128x16x128.LLLALLLALLLALLLALLLALLLA layers=24
3
+ compat h=16fc176471b384c2
4
+ defaults h=57956300c985e99f
5
+ route bmin=1 bmax=1 nmax=128 bucket=u0
6
+ route bmin=1 bmax=1 nmax=256 bucket=u1
7
+ route bmin=1 bmax=1 nmax=1024 bucket=u2
8
+ route bmin=1 bmax=1 nmax=4096 bucket=u3
9
+ route bmin=2 bmax=8 nmax=256 bucket=u4
10
+ compat h=16fc176471b384c2
11
+ bucket kind=o S=128 B=1 M=256 xd=0 name=u0 elf=u0.elf bytes=23536952 sha256=5abb2fe3b9290f9f30d325b7366c9755f9ec809e0ba828475f5829f39dee4be5 csha256=b9bb53d23d9a92b4c5b4155adb4a10c35864803e9e7b73c6dbf0618eabffbf76
12
+ pass gbase=69632 Ts=64,128 bucket=u0
13
+ geom T=64 in_len=135168 CN=65536 CO=0 CM=0 out_len=589824 bucket=u0
14
+ geom T=128 in_len=200704 CN=131072 CO=0 CM=0 out_len=655360 bucket=u0
15
+ buf bytes=524288 role=unused init=zero bucket=u0 name=b0
16
+ buf bytes=4456448 role=unused init=zero bucket=u0 name=b1
17
+ buf bytes=524288 role=unused init=zero bucket=u0 name=b2
18
+ buf bytes=2621440 role=pass init=zero bucket=u0 name=b3
19
+ buf bytes=1048576 role=pass init=zero bucket=u0 name=b4
20
+ buf bytes=524288 role=unused init=zero bucket=u0 name=b5
21
+ buf bytes=524288 role=unused init=zero bucket=u0 name=b6
22
+ buf bytes=1835008 role=unused init=zero bucket=u0 name=b7
23
+ buf bytes=524288 role=unused init=zero bucket=u0 name=b8
24
+ buf bytes=32768 role=const init=blob src=consts blob=u0/w0 bucket=u0 name=b9
25
+ buf bytes=16384 role=const init=blob src=consts blob=u0/w1 bucket=u0 name=b10
26
+ buf bytes=16384 role=const init=blob src=consts blob=u0/w2 bucket=u0 name=b11
27
+ buf bytes=2367488 role=unused init=zero bucket=u0 name=b12
28
+ buf bytes=1048576 role=unused init=zero bucket=u0 name=b13
29
+ buf bytes=884736 role=pass init=zero bucket=u0 name=b14
30
+ buf bytes=147456 role=pass init=zero bucket=u0 name=b15
31
+ buf bytes=1835008 role=pass init=zero bucket=u0 name=b16_64
32
+ buf bytes=1179648 role=pass init=zero bucket=u0 name=b17_64
33
+ buf bytes=24772608 role=layer init=blob src=weights blob=T64/L{i}/w3 layers=all share=1 bucket=u0 name=b18_64
34
+ buf bytes=270336 role=layer init=zero layers=all patch=L{i}/w11@131072 bucket=u0 name=b19_64
35
+ buf bytes=3670016 role=pass init=zero bucket=u0 name=b16_128
36
+ buf bytes=1310720 role=pass init=zero bucket=u0 name=b17_128
37
+ buf bytes=24625152 role=layer init=blob src=weights blob=T128/L{i}/w3 layers=all share=1 bucket=u0 name=b18_128
38
+ buf bytes=401408 role=layer init=zero layers=all patch=L{i}/w11@262144 bucket=u0 name=b19_128
39
+ buf bytes=10045440 role=unused init=blob src=weights blob=L{i}/w4 layers=linear_attention share=1 bucket=u0 name=b20
40
+ buf bytes=2377728 role=unused init=blob src=weights blob=L{i}/w5 layers=all share=1 bucket=u0 name=b21
41
+ buf bytes=8275968 role=unused init=blob src=weights blob=L{i}/w6 layers=all share=1 bucket=u0 name=b22
42
+ buf bytes=4147200 role=unused init=blob src=weights blob=L{i}/w7 layers=all share=1 bucket=u0 name=b23
43
+ buf bytes=5916672 role=unused init=blob src=weights blob=L{i}/w8 layers=full_attention share=1 bucket=u0 name=b24
44
+ buf bytes=16384 role=layer init=blob src=weights blob=S128/L{i}/w9 layers=full_attention share=1 bucket=u0 name=b25
45
+ buf bytes=16384 role=layer init=blob src=weights blob=S128/L{i}/w10 layers=full_attention share=1 bucket=u0 name=b26
46
+ kernel bucket=u0 name=k0 slots=b0:bf16,b20:bf16,b1:bf16 kpts=f0 ro=k17 keep=k18
47
+ kernel bucket=u0 name=k1 slots=b12:bf16,b4:bf16,b13:bf16,b21:bf16,b5:bf16 kpts=f1,f0
48
+ kernel bucket=u0 name=k2 slots=b2:bf16,b24:bf16,b3:bf16,b9:f32,b25:bf16,b14:bf16,b26:bf16,b15:bf16,b10:bf16,b11:bf16,b4:bf16,b21:bf16,b5:bf16 kpts=f0,f2,f3,f0 ro=k19
49
+ kernel bucket=u0 name=k3 slots=b6:bf16,b22:bf16,b7:bf16,b23:bf16,b8:bf16 kpts=f0 ro=k20 keep=k21
50
+ kernel bucket=u0 name=k4_64 slots=b18_64:bf16,b19_64:bf16,b16_64:bf16,b17_64:bf16,b3:bf16,b4:bf16 kpts=f4
51
+ kernel bucket=u0 name=k5_64 slots=b18_64:bf16,b19_64:bf16,b16_64:bf16,b17_64:bf16,b3:bf16,b4:bf16 kpts=f4
52
+ kernel bucket=u0 name=k6_64 slots=b18_64:bf16,b19_64:bf16,b16_64:bf16,b17_64:bf16,b3:bf16,b4:bf16 kpts=f4
53
+ kernel bucket=u0 name=k7_64 slots=b18_64:bf16,b19_64:bf16,b16_64:bf16,b17_64:bf16,b3:bf16,b4:bf16 kpts=f4
54
+ kernel bucket=u0 name=k8_64 slots=b18_64:bf16,b19_64:bf16,b16_64:bf16,b17_64:bf16,b3:bf16,b4:bf16 kpts=f4
55
+ kernel bucket=u0 name=k9_64 slots=b18_64:bf16,b19_64:bf16,b16_64:bf16,b17_64:bf16,b3:bf16,b4:bf16 kpts=f4
56
+ kernel bucket=u0 name=k10_64 slots=b18_64:bf16,b19_64:bf16,b16_64:bf16,b17_64:bf16,b3:bf16,b4:bf16 kpts=f4
57
+ kernel bucket=u0 name=k11_64 slots=b18_64:bf16,b19_64:bf16,b16_64:bf16,b17_64:bf16,b3:bf16,b4:bf16 kpts=f4
58
+ kernel bucket=u0 name=k4_128 slots=b18_128:bf16,b19_128:bf16,b16_128:bf16,b17_128:bf16,b3:bf16,b4:bf16 kpts=f5
59
+ kernel bucket=u0 name=k5_128 slots=b18_128:bf16,b19_128:bf16,b16_128:bf16,b17_128:bf16,b3:bf16,b4:bf16 kpts=f5
60
+ kernel bucket=u0 name=k6_128 slots=b18_128:bf16,b19_128:bf16,b16_128:bf16,b17_128:bf16,b3:bf16,b4:bf16 kpts=f5
61
+ kernel bucket=u0 name=k7_128 slots=b18_128:bf16,b19_128:bf16,b16_128:bf16,b17_128:bf16,b3:bf16,b4:bf16 kpts=f5
62
+ kernel bucket=u0 name=k8_128 slots=b18_128:bf16,b19_128:bf16,b16_128:bf16,b17_128:bf16,b3:bf16,b4:bf16 kpts=f5
63
+ kernel bucket=u0 name=k9_128 slots=b18_128:bf16,b19_128:bf16,b16_128:bf16,b17_128:bf16,b3:bf16,b4:bf16 kpts=f5
64
+ kernel bucket=u0 name=k10_128 slots=b18_128:bf16,b19_128:bf16,b16_128:bf16,b17_128:bf16,b3:bf16,b4:bf16 kpts=f5
65
+ kernel bucket=u0 name=k11_128 slots=b18_128:bf16,b19_128:bf16,b16_128:bf16,b17_128:bf16,b3:bf16,b4:bf16 kpts=f5
66
+ kernel bucket=u0 name=k12 slots=b3:bf16,b9:f32,b25:bf16,b14:bf16,b26:bf16,b15:bf16,b10:bf16,b11:bf16,b4:bf16 kpts=f2,f3
67
+ kernel bucket=u0 name=k13 slots=b3:bf16,b9:f32,b25:bf16,b14:bf16,b26:bf16,b15:bf16,b10:bf16,b11:bf16 kpts=f2
68
+ kernel bucket=u0 name=k14 slots=b14:bf16,b15:bf16,b4:bf16 kpts=f3
69
+ kernel bucket=u0 name=k15 slots=b3:bf16,b9:f32,b25:bf16,b14:bf16,b26:bf16,b15:bf16,b10:bf16,b11:bf16 kpts=f2
70
+ kernel bucket=u0 name=k16 slots=b14:bf16,b15:bf16,b4:bf16 kpts=f3
71
+ kernel bucket=u0 name=k17 slots=b0:bf16,b20:bf16,b1:bf16 kpts=f0 base=k0
72
+ kernel bucket=u0 name=k18 slots=b0:bf16,b20:bf16,b1:bf16 kpts=f0 base=k0
73
+ kernel bucket=u0 name=k19 slots=b2:bf16,b24:bf16,b3:bf16,b9:f32,b25:bf16,b14:bf16,b26:bf16,b15:bf16,b10:bf16,b11:bf16,b4:bf16,b21:bf16,b5:bf16 kpts=f0,f2,f3,f0 base=k2
74
+ kernel bucket=u0 name=k20 slots=b6:bf16,b22:bf16,b7:bf16,b23:bf16,b8:bf16 kpts=f0 base=k3
75
+ kernel bucket=u0 name=k21 slots=b6:bf16,b22:bf16,b7:bf16,b23:bf16,b8:bf16 kpts=f0 base=k3
76
+ plan bucket=u0 type=linear_attention ops=tin:b19_{T},call:x0_{T},out:b17_{T}@0,res,tin:b19_{T},call:x1_{T},out:b17_{T}@0,res
77
+ plan bucket=u0 type=full_attention ops=tin:b19_{T},call:x2_{T},call:k12,call:x3_{T},out:b17_{T}@0,res,tin:b19_{T},call:x1_{T},out:b17_{T}@0,res
78
+ role bucket=u0 att=k2 mlp=k3 x=b12 st=b13 kv=b15 ia=b0 ib=b2 iu=b6 cd=b5 cm=b8 cb=L{i}/w11 sb=L{i}/w12
79
+ variant bucket=u0 name=x0_64 T=64 ro=k5_64 keep=k4_64
80
+ variant bucket=u0 name=x0_128 T=128 ro=k5_128 keep=k4_128
81
+ variant bucket=u0 name=x1_64 T=64 ro=k7_64 keep=k6_64
82
+ variant bucket=u0 name=x1_128 T=128 ro=k7_128 keep=k6_128
83
+ variant bucket=u0 name=x2_64 T=64 ro=k9_64 keep=k8_64
84
+ variant bucket=u0 name=x2_128 T=128 ro=k9_128 keep=k8_128
85
+ variant bucket=u0 name=x3_64 T=64 ro=k11_64 keep=k10_64
86
+ variant bucket=u0 name=x3_128 T=128 ro=k11_128 keep=k10_128
87
+ bucket kind=g S=256 B=1 M=256 xd=3 name=u1 elf=u1.elf bytes=16059928 sha256=c83cee8f306d81799d350b1a8796da6bb0ed098a9b554f93742573e3a39c7ed5 csha256=c0dcee7a72e8086d2e19a4a0fb1923d0de7e085a58e40286e693156a9e82c52e
88
+ pass xw=2297856 gbase=69632 gst=1048576 fkv=294912 bucket=u1
89
+ buf bytes=524288 role=pass init=zero bucket=u1 name=b0
90
+ buf bytes=4456448 role=unused init=zero bucket=u1 name=b1
91
+ buf bytes=524288 role=pass init=zero bucket=u1 name=b2
92
+ buf bytes=2621440 role=pass init=zero bucket=u1 name=b3
93
+ buf bytes=1048576 role=pass init=zero bucket=u1 name=b4
94
+ buf bytes=524288 role=pass init=zero bucket=u1 name=b5
95
+ buf bytes=524288 role=pass init=zero bucket=u1 name=b6
96
+ buf bytes=1835008 role=pass init=zero bucket=u1 name=b7
97
+ buf bytes=524288 role=pass init=zero bucket=u1 name=b8
98
+ buf bytes=65536 role=const init=blob src=consts blob=u1/w0 bucket=u1 name=b9
99
+ buf bytes=16384 role=const init=blob src=consts blob=u1/w1 bucket=u1 name=b10
100
+ buf bytes=16384 role=const init=blob src=consts blob=u1/w2 bucket=u1 name=b11
101
+ buf bytes=4890624 role=pass init=zero bucket=u1 name=b12
102
+ buf bytes=1048576 role=pass init=zero bucket=u1 name=b13
103
+ buf bytes=1769472 role=pass init=zero bucket=u1 name=b14
104
+ buf bytes=294912 role=pass init=zero bucket=u1 name=b15
105
+ buf bytes=10045440 role=layer init=blob src=weights blob=L{i}/w4 layers=linear_attention share=1 bucket=u1 name=b20
106
+ buf bytes=2377728 role=layer init=blob src=weights blob=L{i}/w5 layers=all share=1 bucket=u1 name=b21
107
+ buf bytes=8275968 role=layer init=blob src=weights blob=L{i}/w6 layers=all share=1 bucket=u1 name=b22
108
+ buf bytes=4147200 role=layer init=blob src=weights blob=L{i}/w7 layers=all share=1 bucket=u1 name=b23
109
+ buf bytes=5916672 role=layer init=blob src=weights blob=L{i}/w8 layers=full_attention share=1 bucket=u1 name=b24
110
+ buf bytes=16384 role=layer init=blob src=weights blob=S256/L{i}/w9 layers=full_attention share=1 bucket=u1 name=b25
111
+ buf bytes=16384 role=layer init=blob src=weights blob=S256/L{i}/w10 layers=full_attention share=1 bucket=u1 name=b26
112
+ kernel bucket=u1 name=k0 slots=b0:bf16,b20:bf16,b12:bf16 kpts=f0 ro=k17 keep=k18
113
+ kernel bucket=u1 name=k1 slots=b12:bf16,b4:bf16,b13:bf16,b21:bf16,b5:bf16 kpts=f1,f0
114
+ kernel bucket=u1 name=k22 slots=b0:bf16,b20:bf16,b12:bf16,b4:bf16,b13:bf16,b21:bf16,b5:bf16 kpts=f0,f1,f0 ro=k23 keep=k24
115
+ kernel bucket=u1 name=k2 slots=b2:bf16,b24:bf16,b3:bf16,b9:f32,b25:bf16,b14:bf16,b26:bf16,b15:bf16,b10:bf16,b11:bf16,b4:bf16,b21:bf16,b5:bf16 kpts=f0,f2,f3,f0 ro=k19
116
+ kernel bucket=u1 name=k3 slots=b6:bf16,b22:bf16,b7:bf16,b23:bf16,b8:bf16 kpts=f0 ro=k20 keep=k21
117
+ kernel bucket=u1 name=k17 slots=b0:bf16,b20:bf16,b12:bf16 kpts=f0 base=k0
118
+ kernel bucket=u1 name=k18 slots=b0:bf16,b20:bf16,b12:bf16 kpts=f0 base=k0
119
+ kernel bucket=u1 name=k23 slots=b0:bf16,b20:bf16,b12:bf16,b4:bf16,b13:bf16,b21:bf16,b5:bf16 kpts=f0,f1,f0 base=k22
120
+ kernel bucket=u1 name=k24 slots=b0:bf16,b20:bf16,b12:bf16,b4:bf16,b13:bf16,b21:bf16,b5:bf16 kpts=f0,f1,f0 base=k22
121
+ kernel bucket=u1 name=k19 slots=b2:bf16,b24:bf16,b3:bf16,b9:f32,b25:bf16,b14:bf16,b26:bf16,b15:bf16,b10:bf16,b11:bf16,b4:bf16,b21:bf16,b5:bf16 kpts=f0,f2,f3,f0 base=k2
122
+ kernel bucket=u1 name=k20 slots=b6:bf16,b22:bf16,b7:bf16,b23:bf16,b8:bf16 kpts=f0 base=k3
123
+ kernel bucket=u1 name=k21 slots=b6:bf16,b22:bf16,b7:bf16,b23:bf16,b8:bf16 kpts=f0 base=k3
124
+ plan bucket=u1 type=linear_attention ops=in:b0,const:b12,call:k22,out:b5@0,res,in:b6,call:k3,out:b8@0,res
125
+ plan bucket=u1 type=full_attention ops=in:b2,call:k2,out:b5@0,res,in:b6,call:k3,out:b8@0,res
126
+ role bucket=u1 lin=k22 att=k2 mlp=k3 x=b12 st=b13 kv=b15 ia=b0 ib=b2 iu=b6 cd=b5 cm=b8 cb=L{i}/w11 sb=L{i}/w12
127
+ bucket kind=g S=1024 B=1 M=1024 xd=3 name=u2 elf=u2.elf bytes=16706200 sha256=c5cc038bc6c3978351375415df88d94e596c35291e61c9ac27b8c1d927edb3bd csha256=ea3620e61d934afd4873fc9a26a3e46de383ef333529b1f34995f3fff03f1e64
128
+ pass xw=8982528 gbase=69632 gst=1048576 fkv=1179648 bucket=u2
129
+ buf bytes=2097152 role=pass init=zero bucket=u2 name=b0
130
+ buf bytes=17825792 role=unused init=zero bucket=u2 name=b1
131
+ buf bytes=2097152 role=pass init=zero bucket=u2 name=b2
132
+ buf bytes=10485760 role=pass init=zero bucket=u2 name=b3
133
+ buf bytes=4194304 role=pass init=zero bucket=u2 name=b4
134
+ buf bytes=2097152 role=pass init=zero bucket=u2 name=b5
135
+ buf bytes=2097152 role=pass init=zero bucket=u2 name=b6
136
+ buf bytes=7340032 role=pass init=zero bucket=u2 name=b7
137
+ buf bytes=2097152 role=pass init=zero bucket=u2 name=b8
138
+ buf bytes=262144 role=const init=blob src=consts blob=u2/w0 bucket=u2 name=b9
139
+ buf bytes=16384 role=const init=blob src=consts blob=u2/w1 bucket=u2 name=b10
140
+ buf bytes=16384 role=const init=blob src=consts blob=u2/w2 bucket=u2 name=b11
141
+ buf bytes=19046400 role=pass init=zero bucket=u2 name=b12
142
+ buf bytes=1048576 role=pass init=zero bucket=u2 name=b13
143
+ buf bytes=7077888 role=pass init=zero bucket=u2 name=b14
144
+ buf bytes=1179648 role=pass init=zero bucket=u2 name=b15
145
+ buf bytes=10045440 role=layer init=blob src=weights blob=L{i}/w4 layers=linear_attention share=1 bucket=u2 name=b20
146
+ buf bytes=2377728 role=layer init=blob src=weights blob=L{i}/w5 layers=all share=1 bucket=u2 name=b21
147
+ buf bytes=8275968 role=layer init=blob src=weights blob=L{i}/w6 layers=all share=1 bucket=u2 name=b22
148
+ buf bytes=4147200 role=layer init=blob src=weights blob=L{i}/w7 layers=all share=1 bucket=u2 name=b23
149
+ buf bytes=5916672 role=layer init=blob src=weights blob=L{i}/w8 layers=full_attention share=1 bucket=u2 name=b24
150
+ buf bytes=16384 role=layer init=blob src=weights blob=S1024/L{i}/w9 layers=full_attention share=1 bucket=u2 name=b25
151
+ buf bytes=16384 role=layer init=blob src=weights blob=S1024/L{i}/w10 layers=full_attention share=1 bucket=u2 name=b26
152
+ kernel bucket=u2 name=k0 slots=b0:bf16,b20:bf16,b12:bf16 kpts=f0 ro=k17 keep=k18
153
+ kernel bucket=u2 name=k1 slots=b12:bf16,b4:bf16,b13:bf16,b21:bf16,b5:bf16 kpts=f1,f0
154
+ kernel bucket=u2 name=k22 slots=b0:bf16,b20:bf16,b12:bf16,b4:bf16,b13:bf16,b21:bf16,b5:bf16 kpts=f0,f1,f0 ro=k23 keep=k24
155
+ kernel bucket=u2 name=k2 slots=b2:bf16,b24:bf16,b3:bf16,b9:f32,b25:bf16,b14:bf16,b26:bf16,b15:bf16,b10:bf16,b11:bf16,b4:bf16,b21:bf16,b5:bf16 kpts=f0,f2,f3,f0 ro=k19
156
+ kernel bucket=u2 name=k3 slots=b6:bf16,b22:bf16,b7:bf16,b23:bf16,b8:bf16 kpts=f0 ro=k20 keep=k21
157
+ kernel bucket=u2 name=k17 slots=b0:bf16,b20:bf16,b12:bf16 kpts=f0 base=k0
158
+ kernel bucket=u2 name=k18 slots=b0:bf16,b20:bf16,b12:bf16 kpts=f0 base=k0
159
+ kernel bucket=u2 name=k23 slots=b0:bf16,b20:bf16,b12:bf16,b4:bf16,b13:bf16,b21:bf16,b5:bf16 kpts=f0,f1,f0 base=k22
160
+ kernel bucket=u2 name=k24 slots=b0:bf16,b20:bf16,b12:bf16,b4:bf16,b13:bf16,b21:bf16,b5:bf16 kpts=f0,f1,f0 base=k22
161
+ kernel bucket=u2 name=k19 slots=b2:bf16,b24:bf16,b3:bf16,b9:f32,b25:bf16,b14:bf16,b26:bf16,b15:bf16,b10:bf16,b11:bf16,b4:bf16,b21:bf16,b5:bf16 kpts=f0,f2,f3,f0 base=k2
162
+ kernel bucket=u2 name=k20 slots=b6:bf16,b22:bf16,b7:bf16,b23:bf16,b8:bf16 kpts=f0 base=k3
163
+ kernel bucket=u2 name=k21 slots=b6:bf16,b22:bf16,b7:bf16,b23:bf16,b8:bf16 kpts=f0 base=k3
164
+ plan bucket=u2 type=linear_attention ops=in:b0,const:b12,call:k22,out:b5@0,res,in:b6,call:k3,out:b8@0,res
165
+ plan bucket=u2 type=full_attention ops=in:b2,call:k2,out:b5@0,res,in:b6,call:k3,out:b8@0,res
166
+ role bucket=u2 lin=k22 att=k2 mlp=k3 x=b12 st=b13 kv=b15 ia=b0 ib=b2 iu=b6 cd=b5 cm=b8 cb=L{i}/w11 sb=L{i}/w12
167
+ bucket kind=g S=4096 B=1 M=4096 xd=3 name=u3 elf=u3.elf bytes=19939160 sha256=e789c8497e6ab37df422257d3bf750e4418aab9527a651da88f775d8dca6bdec csha256=4c32e3ad46810b71e02838d9d68fe92493272735cc6440e7313ddd03a6972556
168
+ pass xw=35721216 gbase=69632 gst=1048576 fkv=4718592 bucket=u3
169
+ buf bytes=8388608 role=pass init=zero bucket=u3 name=b0
170
+ buf bytes=71303168 role=unused init=zero bucket=u3 name=b1
171
+ buf bytes=8388608 role=pass init=zero bucket=u3 name=b2
172
+ buf bytes=41943040 role=pass init=zero bucket=u3 name=b3
173
+ buf bytes=16777216 role=pass init=zero bucket=u3 name=b4
174
+ buf bytes=8388608 role=pass init=zero bucket=u3 name=b5
175
+ buf bytes=8388608 role=pass init=zero bucket=u3 name=b6
176
+ buf bytes=29360128 role=pass init=zero bucket=u3 name=b7
177
+ buf bytes=8388608 role=pass init=zero bucket=u3 name=b8
178
+ buf bytes=1048576 role=const init=blob src=consts blob=u3/w0 bucket=u3 name=b9
179
+ buf bytes=16384 role=const init=blob src=consts blob=u3/w1 bucket=u3 name=b10
180
+ buf bytes=16384 role=const init=blob src=consts blob=u3/w2 bucket=u3 name=b11
181
+ buf bytes=75669504 role=pass init=zero bucket=u3 name=b12
182
+ buf bytes=1048576 role=pass init=zero bucket=u3 name=b13
183
+ buf bytes=28311552 role=pass init=zero bucket=u3 name=b14
184
+ buf bytes=4718592 role=pass init=zero bucket=u3 name=b15
185
+ buf bytes=10045440 role=layer init=blob src=weights blob=L{i}/w4 layers=linear_attention share=1 bucket=u3 name=b20
186
+ buf bytes=2377728 role=layer init=blob src=weights blob=L{i}/w5 layers=all share=1 bucket=u3 name=b21
187
+ buf bytes=8275968 role=layer init=blob src=weights blob=L{i}/w6 layers=all share=1 bucket=u3 name=b22
188
+ buf bytes=4147200 role=layer init=blob src=weights blob=L{i}/w7 layers=all share=1 bucket=u3 name=b23
189
+ buf bytes=5916672 role=layer init=blob src=weights blob=L{i}/w8 layers=full_attention share=1 bucket=u3 name=b24
190
+ buf bytes=16384 role=layer init=blob src=weights blob=S4096/L{i}/w9 layers=full_attention share=1 bucket=u3 name=b25
191
+ buf bytes=16384 role=layer init=blob src=weights blob=S4096/L{i}/w10 layers=full_attention share=1 bucket=u3 name=b26
192
+ kernel bucket=u3 name=k0 slots=b0:bf16,b20:bf16,b12:bf16 kpts=f0 ro=k17 keep=k18
193
+ kernel bucket=u3 name=k1 slots=b12:bf16,b4:bf16,b13:bf16,b21:bf16,b5:bf16 kpts=f1,f0
194
+ kernel bucket=u3 name=k22 slots=b0:bf16,b20:bf16,b12:bf16,b4:bf16,b13:bf16,b21:bf16,b5:bf16 kpts=f0,f1,f0 ro=k23 keep=k24
195
+ kernel bucket=u3 name=k2 slots=b2:bf16,b24:bf16,b3:bf16,b9:f32,b25:bf16,b14:bf16,b26:bf16,b15:bf16,b10:bf16,b11:bf16,b4:bf16,b21:bf16,b5:bf16 kpts=f0,f2,f3,f0 ro=k19
196
+ kernel bucket=u3 name=k3 slots=b6:bf16,b22:bf16,b7:bf16,b23:bf16,b8:bf16 kpts=f0 ro=k20 keep=k21
197
+ kernel bucket=u3 name=k17 slots=b0:bf16,b20:bf16,b12:bf16 kpts=f0 base=k0
198
+ kernel bucket=u3 name=k18 slots=b0:bf16,b20:bf16,b12:bf16 kpts=f0 base=k0
199
+ kernel bucket=u3 name=k23 slots=b0:bf16,b20:bf16,b12:bf16,b4:bf16,b13:bf16,b21:bf16,b5:bf16 kpts=f0,f1,f0 base=k22
200
+ kernel bucket=u3 name=k24 slots=b0:bf16,b20:bf16,b12:bf16,b4:bf16,b13:bf16,b21:bf16,b5:bf16 kpts=f0,f1,f0 base=k22
201
+ kernel bucket=u3 name=k19 slots=b2:bf16,b24:bf16,b3:bf16,b9:f32,b25:bf16,b14:bf16,b26:bf16,b15:bf16,b10:bf16,b11:bf16,b4:bf16,b21:bf16,b5:bf16 kpts=f0,f2,f3,f0 base=k2
202
+ kernel bucket=u3 name=k20 slots=b6:bf16,b22:bf16,b7:bf16,b23:bf16,b8:bf16 kpts=f0 base=k3
203
+ kernel bucket=u3 name=k21 slots=b6:bf16,b22:bf16,b7:bf16,b23:bf16,b8:bf16 kpts=f0 base=k3
204
+ plan bucket=u3 type=linear_attention ops=in:b0,const:b12,call:k22,out:b5@0,res,in:b6,call:k3,out:b8@0,res
205
+ plan bucket=u3 type=full_attention ops=in:b2,call:k2,out:b5@0,res,in:b6,call:k3,out:b8@0,res
206
+ role bucket=u3 lin=k22 att=k2 mlp=k3 x=b12 st=b13 kv=b15 ia=b0 ib=b2 iu=b6 cd=b5 cm=b8 cb=L{i}/w11 sb=L{i}/w12
207
+ bucket kind=g S=256 B=8 M=2048 xd=3 name=u4 elf=u4.elf bytes=18245064 sha256=d72f184675ef3d558717d27c3ed6de564939970abb9634cb3a211e6e71e9893a csha256=76a6acb8a001840c3a39c6c6317bbc6fbe242dc420569be1840647e5eef53fb9
208
+ pass xw=2297856 gbase=69632 gst=1048576 fkv=294912 bucket=u4
209
+ buf bytes=4194304 role=pass init=zero bucket=u4 name=b0
210
+ buf bytes=35651584 role=unused init=zero bucket=u4 name=b1
211
+ buf bytes=4194304 role=pass init=zero bucket=u4 name=b2
212
+ buf bytes=20971520 role=pass init=zero bucket=u4 name=b3
213
+ buf bytes=8388608 role=pass init=zero bucket=u4 name=b4
214
+ buf bytes=4194304 role=pass init=zero bucket=u4 name=b5
215
+ buf bytes=4194304 role=pass init=zero bucket=u4 name=b6
216
+ buf bytes=14680064 role=pass init=zero bucket=u4 name=b7
217
+ buf bytes=4194304 role=pass init=zero bucket=u4 name=b8
218
+ buf bytes=65536 role=const init=blob src=consts blob=u4/w0 bucket=u4 name=b9
219
+ buf bytes=16384 role=const init=blob src=consts blob=u4/w1 bucket=u4 name=b10
220
+ buf bytes=16384 role=const init=blob src=consts blob=u4/w2 bucket=u4 name=b11
221
+ buf bytes=38895616 role=pass init=zero bucket=u4 name=b12
222
+ buf bytes=8388608 role=pass init=zero bucket=u4 name=b13
223
+ buf bytes=14155776 role=pass init=zero bucket=u4 name=b14
224
+ buf bytes=2359296 role=pass init=zero bucket=u4 name=b15
225
+ buf bytes=10045440 role=layer init=blob src=weights blob=L{i}/w4 layers=linear_attention share=1 bucket=u4 name=b20
226
+ buf bytes=2377728 role=layer init=blob src=weights blob=L{i}/w5 layers=all share=1 bucket=u4 name=b21
227
+ buf bytes=8275968 role=layer init=blob src=weights blob=L{i}/w6 layers=all share=1 bucket=u4 name=b22
228
+ buf bytes=4147200 role=layer init=blob src=weights blob=L{i}/w7 layers=all share=1 bucket=u4 name=b23
229
+ buf bytes=5916672 role=layer init=blob src=weights blob=L{i}/w8 layers=full_attention share=1 bucket=u4 name=b24
230
+ buf bytes=16384 role=layer init=blob src=weights blob=S256/L{i}/w9 layers=full_attention share=1 bucket=u4 name=b25
231
+ buf bytes=16384 role=layer init=blob src=weights blob=S256/L{i}/w10 layers=full_attention share=1 bucket=u4 name=b26
232
+ kernel bucket=u4 name=k0 slots=b0:bf16,b20:bf16,b12:bf16 kpts=f0 ro=k17 keep=k18
233
+ kernel bucket=u4 name=k1 slots=b12:bf16,b4:bf16,b13:bf16,b21:bf16,b5:bf16 kpts=f1,f0
234
+ kernel bucket=u4 name=k22 slots=b0:bf16,b20:bf16,b12:bf16,b4:bf16,b13:bf16,b21:bf16,b5:bf16 kpts=f0,f1,f0 ro=k23 keep=k24
235
+ kernel bucket=u4 name=k2 slots=b2:bf16,b24:bf16,b3:bf16,b9:f32,b25:bf16,b14:bf16,b26:bf16,b15:bf16,b10:bf16,b11:bf16,b4:bf16,b21:bf16,b5:bf16 kpts=f0,f2,f3,f0 ro=k19
236
+ kernel bucket=u4 name=k3 slots=b6:bf16,b22:bf16,b7:bf16,b23:bf16,b8:bf16 kpts=f0 ro=k20 keep=k21
237
+ kernel bucket=u4 name=k17 slots=b0:bf16,b20:bf16,b12:bf16 kpts=f0 base=k0
238
+ kernel bucket=u4 name=k18 slots=b0:bf16,b20:bf16,b12:bf16 kpts=f0 base=k0
239
+ kernel bucket=u4 name=k23 slots=b0:bf16,b20:bf16,b12:bf16,b4:bf16,b13:bf16,b21:bf16,b5:bf16 kpts=f0,f1,f0 base=k22
240
+ kernel bucket=u4 name=k24 slots=b0:bf16,b20:bf16,b12:bf16,b4:bf16,b13:bf16,b21:bf16,b5:bf16 kpts=f0,f1,f0 base=k22
241
+ kernel bucket=u4 name=k19 slots=b2:bf16,b24:bf16,b3:bf16,b9:f32,b25:bf16,b14:bf16,b26:bf16,b15:bf16,b10:bf16,b11:bf16,b4:bf16,b21:bf16,b5:bf16 kpts=f0,f2,f3,f0 base=k2
242
+ kernel bucket=u4 name=k20 slots=b6:bf16,b22:bf16,b7:bf16,b23:bf16,b8:bf16 kpts=f0 base=k3
243
+ kernel bucket=u4 name=k21 slots=b6:bf16,b22:bf16,b7:bf16,b23:bf16,b8:bf16 kpts=f0 base=k3
244
+ plan bucket=u4 type=linear_attention ops=in:b0,const:b12,call:k22,out:b5@0,res,in:b6,call:k3,out:b8@0,res
245
+ plan bucket=u4 type=full_attention ops=in:b2,call:k2,out:b5@0,res,in:b6,call:k3,out:b8@0,res
246
+ role bucket=u4 lin=k22 att=k2 mlp=k3 x=b12 st=b13 kv=b15 ia=b0 ib=b2 iu=b6 cd=b5 cm=b8 cb=L{i}/w11 sb=L{i}/w12
devices/u0.elf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5abb2fe3b9290f9f30d325b7366c9755f9ec809e0ba828475f5829f39dee4be5
3
+ size 23536952
devices/u1.elf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c83cee8f306d81799d350b1a8796da6bb0ed098a9b554f93742573e3a39c7ed5
3
+ size 16059928
devices/u2.elf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c5cc038bc6c3978351375415df88d94e596c35291e61c9ac27b8c1d927edb3bd
3
+ size 16706200
devices/u3.elf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e789c8497e6ab37df422257d3bf750e4418aab9527a651da88f775d8dca6bdec
3
+ size 19939160
devices/u4.elf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d72f184675ef3d558717d27c3ed6de564939970abb9634cb3a211e6e71e9893a
3
+ size 18245064
tokenizer/tokenizer.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:06b9509352d2af50381ab2247e083b80d32d5c0aba91c272ca9ff729b6a0e523
3
+ size 19989325