sanmonga22 commited on
Commit
8d056cf
·
verified ·
1 Parent(s): 30a232d

Bonsai-4B 1-bit v81 HNPU bundle (device-validated)

Browse files
.gitattributes CHANGED
@@ -33,3 +33,4 @@ 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
+ v81/tokenizer.json filter=lfs diff=lfs merge=lfs -text
README.md ADDED
@@ -0,0 +1,18 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: other
3
+ tags:
4
+ - hnpu
5
+ - hexagon
6
+ - npu
7
+ - v81
8
+ - text-generation
9
+ - 1-bit
10
+ base_model: prism-ml/Bonsai-4B
11
+ ---
12
+
13
+ # Bonsai-4B — 1-bit for the Qualcomm Hexagon NPU (HNPU)
14
+
15
+ A QHexRT-runnable bundle of [prism-ml/Bonsai-4B](https://huggingface.co/prism-ml/Bonsai-4B) at **true 1-bit
16
+ (Q1_0, ~1.125 bit/weight)** for the Hexagon NPU. Runs the full model on-device via a custom HVX op-package.
17
+
18
+ **Supported arches:** v81 (SM8850).
v81/bonsai-4b-1024.json ADDED
@@ -0,0 +1,136 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema_version": 1,
3
+ "_comment": "Bonsai-4b @ TRUE 1-bit (Q1_0 BitnetLinear) on v81. Dense Qwen3 (36 layers, GQA 32Q/8KV, qk-norm, YaRN). 1-bit decode graph + 1-bit tied lm-head + fp16 embed, driven by qwen3_generate. op-package: BonsaiBitnetOpPkg (HVX).",
4
+ "model": {
5
+ "name": "bonsai-4b-1024",
6
+ "family": "llm",
7
+ "dsp_arch": "v81",
8
+ "tokenizer_pre": "qwen2"
9
+ },
10
+ "chat": {
11
+ "system": {
12
+ "prefix": "<|im_start|>system\n",
13
+ "suffix": "<|im_end|>\n"
14
+ },
15
+ "user": {
16
+ "prefix": "<|im_start|>user\n",
17
+ "suffix": "<|im_end|>\n"
18
+ },
19
+ "assistant": {
20
+ "prefix": "<|im_start|>assistant\n",
21
+ "suffix": "<|im_end|>\n"
22
+ }
23
+ },
24
+ "params": {
25
+ "hidden": 2560,
26
+ "vocab": 151669,
27
+ "n_layers": 36,
28
+ "max_ctx": 1024,
29
+ "kv_dim": 1024,
30
+ "head_dim": 128,
31
+ "rope_theta": 5000000.0,
32
+ "eos_token_id": 151645,
33
+ "rope_inv_freq": [
34
+ 1.0,
35
+ 0.78582998,
36
+ 0.617528758,
37
+ 0.485272612,
38
+ 0.381341767,
39
+ 0.299669793,
40
+ 0.235489508,
41
+ 0.185054715,
42
+ 0.145421543,
43
+ 0.114276609,
44
+ 0.089801985,
45
+ 0.070569092,
46
+ 0.055455308,
47
+ 0.043578444,
48
+ 0.034245248,
49
+ 0.026910942,
50
+ 0.020090054,
51
+ 0.014956453,
52
+ 0.011100272,
53
+ 0.008209813,
54
+ 0.006048297,
55
+ 0.004436071,
56
+ 0.003236998,
57
+ 0.002348058,
58
+ 0.00169141,
59
+ 0.001208328,
60
+ 0.000854586,
61
+ 0.000596942,
62
+ 0.000410458,
63
+ 0.000276472,
64
+ 0.00018105,
65
+ 0.000142274,
66
+ 0.000111803,
67
+ 8.7858e-05,
68
+ 6.9042e-05,
69
+ 5.4255e-05,
70
+ 4.2635e-05,
71
+ 3.3504e-05,
72
+ 2.6329e-05,
73
+ 2.069e-05,
74
+ 1.6259e-05,
75
+ 1.2777e-05,
76
+ 1.004e-05,
77
+ 7.89e-06,
78
+ 6.2e-06,
79
+ 4.872e-06,
80
+ 3.829e-06,
81
+ 3.009e-06,
82
+ 2.364e-06,
83
+ 1.858e-06,
84
+ 1.46e-06,
85
+ 1.147e-06,
86
+ 9.02e-07,
87
+ 7.09e-07,
88
+ 5.57e-07,
89
+ 4.38e-07,
90
+ 3.44e-07,
91
+ 2.7e-07,
92
+ 2.12e-07,
93
+ 1.67e-07,
94
+ 1.31e-07,
95
+ 1.03e-07,
96
+ 8.1e-08,
97
+ 6.4e-08
98
+ ],
99
+ "rope_mscale": 1.138629436
100
+ },
101
+ "artifacts": {
102
+ "contexts": {
103
+ "decode": {
104
+ "bin": "bonsai4b_decn1k.bin",
105
+ "op_packages": [
106
+ {
107
+ "path": "libQnnBonsaiBitnet.so",
108
+ "interface": "exampleInterfaceProvider",
109
+ "targets": [
110
+ "CPU",
111
+ "HTP"
112
+ ]
113
+ }
114
+ ]
115
+ },
116
+ "lmhead": {
117
+ "bin": "bonsai4b_lmh.bin"
118
+ }
119
+ },
120
+ "embed": "bonsai4b_embed_f16.bin",
121
+ "tokenizer": "tokenizer.json"
122
+ },
123
+ "plan": {
124
+ "steps": [
125
+ {
126
+ "host": "qwen3_generate",
127
+ "params": {
128
+ "decode": "bonsai4b_decn1k",
129
+ "lmhead": "bonsai4b_lmh",
130
+ "max_new": 64,
131
+ "neg": -50000
132
+ }
133
+ }
134
+ ]
135
+ }
136
+ }
v81/bonsai4b_decn1k.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a12b0e30dffc5462d16fd22ba4751b54cdbde6cdd57e5a0257e0ca6716711d51
3
+ size 515739648
v81/bonsai4b_embed_f16.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:046edf4688c954b4e68f64b6aac4da2d9bc274bce26fcae5e268556a6d9c6395
3
+ size 776545280
v81/bonsai4b_lmh.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c3255ecdc594c7f10ec81176da8d1bdb570691e924ef0c35f558b81ab615f7b2
3
+ size 54644736
v81/tokenizer.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:aeb13307a71acd8fe81861d94ad54ab689df773318809eed3cbe794b4492dae4
3
+ size 11422654