sanmonga22 commited on
Commit
e49b028
·
verified ·
1 Parent(s): facac0b

Bonsai-27B 1-bit v81 bundle (4 decode shards + 1-bit lmhead + fp16 embed + tokenizer + op-package stub)

Browse files
.gitattributes CHANGED
@@ -33,3 +33,5 @@ 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/libQnnBonsaiBitnet.so filter=lfs diff=lfs merge=lfs -text
37
+ v81/tokenizer.json filter=lfs diff=lfs merge=lfs -text
v81/bonsai-27b-1024.json ADDED
@@ -0,0 +1,108 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema_version": 1,
3
+ "_comment": "Bonsai-27B (prism-ml) qwen3_5 GatedDeltaNet hybrid TEXT decoder at TRUE 1-BIT (Q1_0_g128, 1.125 bit/weight) for v81. 48 recurrent gated delta-rule linear-attn layers (+ short conv state) + 16 full gated softmax-attn layers (partial-rotary M-RoPE rot 64, output gate). Every Linear is the device-verified BonsaiBitnetOpPkg::BitnetLinear 1-bit op; norms/conv/SSM stay native. Weights come from the Q1_0 GGUF (dequant is the oracle). Decode is sharded (decode_shards) to fit the per-context cDSP memory wall; the qwen3_5_generate host-op (device-validated on qwen3.5-0.8b @ 25 tok/s) threads conv/SSM/KV state + drives the decode loop unchanged.",
4
+ "model": {
5
+ "name": "bonsai-27b-1bit",
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": 5120,
26
+ "vocab": 248320,
27
+ "n_layers": 64,
28
+ "max_ctx": 2048,
29
+ "kv_dim": 1024,
30
+ "head_dim": 256,
31
+ "rope_theta": 10000000.0,
32
+ "eos_token_id": 248044,
33
+ "gated_delta": {
34
+ "q_heads": 24,
35
+ "kv_heads": 4,
36
+ "lin_k_heads": 16,
37
+ "lin_v_heads": 48,
38
+ "lin_k_dim": 128,
39
+ "lin_v_dim": 128,
40
+ "conv_kernel": 4,
41
+ "full_every": 4,
42
+ "rot_dim": 64,
43
+ "attn_gate": true,
44
+ "full_attn_idx": [
45
+ 3,
46
+ 7,
47
+ 11,
48
+ 15,
49
+ 19,
50
+ 23,
51
+ 27,
52
+ 31,
53
+ 35,
54
+ 39,
55
+ 43,
56
+ 47,
57
+ 51,
58
+ 55,
59
+ 59,
60
+ 63
61
+ ]
62
+ }
63
+ },
64
+ "artifacts": {
65
+ "contexts": {
66
+ "decode_s0": {
67
+ "bin": "bonsai27b_decode_s0_bitnet.bin",
68
+ "op_packages": [
69
+ {
70
+ "path": "libQnnBonsaiBitnet.so",
71
+ "interface": "exampleInterfaceProvider",
72
+ "targets": [
73
+ "CPU",
74
+ "HTP"
75
+ ]
76
+ }
77
+ ]
78
+ },
79
+ "decode_s1": {
80
+ "bin": "bonsai27b_decode_s1_bitnet.bin"
81
+ },
82
+ "decode_s2": {
83
+ "bin": "bonsai27b_decode_s2_bitnet.bin"
84
+ },
85
+ "decode_s3": {
86
+ "bin": "bonsai27b_decode_s3_bitnet.bin"
87
+ },
88
+ "lmhead": {
89
+ "bin": "bonsai27b_lmhead_bitnet.bin"
90
+ }
91
+ },
92
+ "embed": "bonsai27b_embed_f16.bin",
93
+ "tokenizer": "tokenizer.json"
94
+ },
95
+ "plan": {
96
+ "steps": [
97
+ {
98
+ "host": "qwen3_5_generate",
99
+ "params": {
100
+ "decode": "bonsai27b_decode_s0_bitnet,bonsai27b_decode_s1_bitnet,bonsai27b_decode_s2_bitnet,bonsai27b_decode_s3_bitnet",
101
+ "lmhead": "bonsai27b_lmhead_bitnet",
102
+ "max_new": 64,
103
+ "neg": -50000
104
+ }
105
+ }
106
+ ]
107
+ }
108
+ }
v81/bonsai27b_decode_s0_bitnet.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:808e49534b74da146d050b10d3d166443ed4271bd0d3396d993ebfad5131497a
3
+ size 873480192
v81/bonsai27b_decode_s1_bitnet.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:eba83d2374a2705fd8b2f1832ba1c672fc72e6615d3a1b633e4afb8907c73647
3
+ size 873488384
v81/bonsai27b_decode_s2_bitnet.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3f7e20221bcf132ebb0d2d1cc1c8b0a6806a809729ca7da158ae0120789b4cc8
3
+ size 873488384
v81/bonsai27b_decode_s3_bitnet.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fb6ae694c4b24b4abf6cf7cfe47e875d9a8e6bedddc5930be068fcd94888ac15
3
+ size 873512960
v81/bonsai27b_embed_f16.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:22cfc3a96ea2482218c90011131835ccd4a164c4e1a372bb506968a819495736
3
+ size 2542796800
v81/bonsai27b_lmhead_bitnet.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:85f08b273b762b8eece1b85bbf18c5adb8e0941405f7df003ada8a7b76a7bdb0
3
+ size 178835456
v81/libQnnBonsaiBitnet.so ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:94f0fe6c4a963ccbccf70e7865a5121f3dee892729c57a2e926dc129f61f0ae2
3
+ size 2317600
v81/tokenizer.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:06b9509352d2af50381ab2247e083b80d32d5c0aba91c272ca9ff729b6a0e523
3
+ size 19989325