@lmzhen/dsh-evolution-merge-engine 0.0.0-stage → 0.29.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/LICENSE +21 -0
- package/NOTICE +25 -0
- package/UPSTREAM.json +1657 -0
- package/data/minhash_params.json +1063 -0
- package/lib/index.d.ts +35 -0
- package/lib/index.js +5814 -0
- package/lib/vendor/skill-merge-pipeline/models/pca64/proj.f32 +0 -0
- package/lib/vendor/skill-merge-pipeline/models/registry.json +333 -0
- package/package.json +34 -4
- package/README.md +0 -3
|
Binary file
|
|
@@ -0,0 +1,333 @@
|
|
|
1
|
+
{
|
|
2
|
+
"schemaVersion": 1,
|
|
3
|
+
"modelsRoot": "models",
|
|
4
|
+
"recipes": {
|
|
5
|
+
"fp16-cast": {
|
|
6
|
+
"builderVer": 1,
|
|
7
|
+
"description": "把 FP32 静态表分块转为 FP16（每块 40000 行），更新 config.embedding_dtype=float16",
|
|
8
|
+
"params": {
|
|
9
|
+
"sourceDtype": "F32",
|
|
10
|
+
"targetDtype": "F16",
|
|
11
|
+
"step": 40000,
|
|
12
|
+
"configPatch": {
|
|
13
|
+
"embedding_dtype": "float16"
|
|
14
|
+
}
|
|
15
|
+
}
|
|
16
|
+
},
|
|
17
|
+
"int8-per-row": {
|
|
18
|
+
"builderVer": 1,
|
|
19
|
+
"description": "逐行对称 int8 量化：scale=max_abs(row)/127（0→1e-8），q=clip(round_half_even(x/scale),-127,127)",
|
|
20
|
+
"params": {
|
|
21
|
+
"dtype": "I8",
|
|
22
|
+
"scalesFile": "scales.f32",
|
|
23
|
+
"scalesDtype": "F32",
|
|
24
|
+
"maxQ": 127,
|
|
25
|
+
"minQ": -127,
|
|
26
|
+
"scaleFormula": "max_abs/127",
|
|
27
|
+
"zeroScale": 1e-8,
|
|
28
|
+
"rounding": "half-even",
|
|
29
|
+
"configPatch": {
|
|
30
|
+
"embedding_dtype": "int8",
|
|
31
|
+
"quantization": "per_row_symmetric",
|
|
32
|
+
"scales_file": "scales.f32",
|
|
33
|
+
"source_model": "potion-multilingual-128M-F16"
|
|
34
|
+
}
|
|
35
|
+
}
|
|
36
|
+
},
|
|
37
|
+
"pca-project-int8": {
|
|
38
|
+
"builderVer": 1,
|
|
39
|
+
"description": "PCA 投影到 64 维后逐行 int8：采样 step=rows//50000，cov=X^T X(double)，eigh 取 top64，proj=x@U，再 int8-per-row",
|
|
40
|
+
"params": {
|
|
41
|
+
"dims": 64,
|
|
42
|
+
"sampleStepFormula": "rows//50000",
|
|
43
|
+
"covariance": "double",
|
|
44
|
+
"eigen": "eigh",
|
|
45
|
+
"quantize": "int8-per-row",
|
|
46
|
+
"scalesFile": "scales.f32",
|
|
47
|
+
"configPatch": {
|
|
48
|
+
"hidden_dim": 64,
|
|
49
|
+
"embedding_dtype": "int8",
|
|
50
|
+
"quantization": "per_row_symmetric",
|
|
51
|
+
"scales_file": "scales.f32",
|
|
52
|
+
"pca_proj_file": "pca_proj.f32",
|
|
53
|
+
"pca_dims": 64,
|
|
54
|
+
"orig_hidden_dim": 256,
|
|
55
|
+
"source_model": "potion-multilingual-128M-F16",
|
|
56
|
+
"pca_fit": "50k evenly-spaced rows, torch covariance"
|
|
57
|
+
}
|
|
58
|
+
}
|
|
59
|
+
}
|
|
60
|
+
},
|
|
61
|
+
"artifacts": [
|
|
62
|
+
{
|
|
63
|
+
"id": "potion-base-f32",
|
|
64
|
+
"kind": "source",
|
|
65
|
+
"origin": "hf",
|
|
66
|
+
"repo": "minishlab/potion-multilingual-128M",
|
|
67
|
+
"description": "HF 原始 fp32 静态表（embedding 512,361,560 B）",
|
|
68
|
+
"target": "potion-multilingual-128M-fp32",
|
|
69
|
+
"files": [
|
|
70
|
+
{
|
|
71
|
+
"path": "model.safetensors",
|
|
72
|
+
"sha256": "14b5eb39cb4ce5666da8ad1f3dc6be4346e9b2d601c073302fa0a31bf7943397",
|
|
73
|
+
"bytes": 512361560,
|
|
74
|
+
"lfs": true
|
|
75
|
+
},
|
|
76
|
+
{
|
|
77
|
+
"path": "tokenizer.json",
|
|
78
|
+
"sha256": "19f1909063da3cfe3bd83a782381f040dccea475f4816de11116444a73e1b6a1",
|
|
79
|
+
"bytes": 18616131,
|
|
80
|
+
"lfs": true
|
|
81
|
+
},
|
|
82
|
+
{
|
|
83
|
+
"path": "config.json",
|
|
84
|
+
"sha256": "595e4cab2093732efd5dbe084fd5c1826b5eea693b73b4c1fd971672867d2e54",
|
|
85
|
+
"bytes": 271
|
|
86
|
+
}
|
|
87
|
+
]
|
|
88
|
+
},
|
|
89
|
+
{
|
|
90
|
+
"id": "potion-base-f16",
|
|
91
|
+
"kind": "derived",
|
|
92
|
+
"recipe": "fp16-cast",
|
|
93
|
+
"base": "potion-base-f32",
|
|
94
|
+
"target": "potion-multilingual-fp16",
|
|
95
|
+
"manifest": {
|
|
96
|
+
"dtype": "F16",
|
|
97
|
+
"dim": 256,
|
|
98
|
+
"rows": 500353,
|
|
99
|
+
"tokenizerSha": "19f1909063da3cfe3bd83a782381f040dccea475f4816de11116444a73e1b6a1",
|
|
100
|
+
"builderVer": 1
|
|
101
|
+
},
|
|
102
|
+
"files": [
|
|
103
|
+
{
|
|
104
|
+
"path": "model.safetensors",
|
|
105
|
+
"sha256": "ccc3e2a930029660672855dc4936dadef9fdfb778729ffb11c6d6c6dc2a33f58",
|
|
106
|
+
"bytes": 256180824
|
|
107
|
+
},
|
|
108
|
+
{
|
|
109
|
+
"path": "tokenizer.json",
|
|
110
|
+
"sha256": "19f1909063da3cfe3bd83a782381f040dccea475f4816de11116444a73e1b6a1",
|
|
111
|
+
"bytes": 18616131
|
|
112
|
+
},
|
|
113
|
+
{
|
|
114
|
+
"path": "config.json",
|
|
115
|
+
"sha256": "e73c9e4782048819e3488c294edbd8f36ba3d5c043e19aac3fa271609b328fb0",
|
|
116
|
+
"bytes": 292
|
|
117
|
+
}
|
|
118
|
+
]
|
|
119
|
+
},
|
|
120
|
+
{
|
|
121
|
+
"id": "pca64-proj",
|
|
122
|
+
"kind": "source",
|
|
123
|
+
"origin": "repo",
|
|
124
|
+
"description": "PCA-64 投影矩阵（256×64 F32，随仓库分发；Node 不做特征分解）",
|
|
125
|
+
"target": "pca64",
|
|
126
|
+
"files": [
|
|
127
|
+
{
|
|
128
|
+
"path": "proj.f32",
|
|
129
|
+
"sha256": "28cbc252208d3ee6767de33f88da60de056a602d0e939e48e3e44ebf72400668",
|
|
130
|
+
"bytes": 65536
|
|
131
|
+
}
|
|
132
|
+
]
|
|
133
|
+
},
|
|
134
|
+
{
|
|
135
|
+
"id": "local:int8-256",
|
|
136
|
+
"kind": "derived",
|
|
137
|
+
"recipe": "int8-per-row",
|
|
138
|
+
"base": "potion-base-f16",
|
|
139
|
+
"target": "potion-multilingual-128M-int8",
|
|
140
|
+
"manifest": {
|
|
141
|
+
"dtype": "I8",
|
|
142
|
+
"dim": 256,
|
|
143
|
+
"rows": 500353,
|
|
144
|
+
"tokenizerSha": "19f1909063da3cfe3bd83a782381f040dccea475f4816de11116444a73e1b6a1",
|
|
145
|
+
"scalesDtype": "F32",
|
|
146
|
+
"builderVer": 1
|
|
147
|
+
},
|
|
148
|
+
"files": [
|
|
149
|
+
{
|
|
150
|
+
"path": "model.safetensors",
|
|
151
|
+
"sha256": "7f4be8f9a3410990204b0dd56358bdd041ec5062448fa9edd7bacbca62e5ed12",
|
|
152
|
+
"bytes": 128090456
|
|
153
|
+
},
|
|
154
|
+
{
|
|
155
|
+
"path": "scales.f32",
|
|
156
|
+
"sha256": "0a3b1b1a20bc3bafcbaba71b20915337d0726e7bd8aa53f3a3ba1200abb16771",
|
|
157
|
+
"bytes": 2001412
|
|
158
|
+
},
|
|
159
|
+
{
|
|
160
|
+
"path": "tokenizer.json",
|
|
161
|
+
"sha256": "19f1909063da3cfe3bd83a782381f040dccea475f4816de11116444a73e1b6a1",
|
|
162
|
+
"bytes": 18616131
|
|
163
|
+
},
|
|
164
|
+
{
|
|
165
|
+
"path": "config.json",
|
|
166
|
+
"sha256": "040d017a12f0a65e7544cef3b78f4aa98d990b65a62288d7a11429bc0e33b497",
|
|
167
|
+
"bytes": 412
|
|
168
|
+
}
|
|
169
|
+
]
|
|
170
|
+
},
|
|
171
|
+
{
|
|
172
|
+
"id": "local:pca64-int8",
|
|
173
|
+
"kind": "derived",
|
|
174
|
+
"recipe": "pca-project-int8",
|
|
175
|
+
"base": "potion-base-f16",
|
|
176
|
+
"proj": "pca64-proj",
|
|
177
|
+
"target": "potion-multilingual-128M-pca64-int8",
|
|
178
|
+
"manifest": {
|
|
179
|
+
"dtype": "I8",
|
|
180
|
+
"dim": 64,
|
|
181
|
+
"rows": 500353,
|
|
182
|
+
"tokenizerSha": "19f1909063da3cfe3bd83a782381f040dccea475f4816de11116444a73e1b6a1",
|
|
183
|
+
"scalesDtype": "F32",
|
|
184
|
+
"projSha": "28cbc252208d3ee6767de33f88da60de056a602d0e939e48e3e44ebf72400668",
|
|
185
|
+
"builderVer": 1,
|
|
186
|
+
"reproducibility": "numeric"
|
|
187
|
+
},
|
|
188
|
+
"files": [
|
|
189
|
+
{
|
|
190
|
+
"path": "model.safetensors",
|
|
191
|
+
"sha256": "09a3b11f1da69b0f54a08b2bfaec6eee42eedd9226777a56dc15b12f84bccf81",
|
|
192
|
+
"bytes": 32022680
|
|
193
|
+
},
|
|
194
|
+
{
|
|
195
|
+
"path": "scales.f32",
|
|
196
|
+
"sha256": "a65761f9e2ecaa1c80cf1124e7b5267fe9e8c76f92672a6b48e4ce9f26dc163b",
|
|
197
|
+
"bytes": 2001412
|
|
198
|
+
},
|
|
199
|
+
{
|
|
200
|
+
"path": "tokenizer.json",
|
|
201
|
+
"sha256": "19f1909063da3cfe3bd83a782381f040dccea475f4816de11116444a73e1b6a1",
|
|
202
|
+
"bytes": 18616131
|
|
203
|
+
},
|
|
204
|
+
{
|
|
205
|
+
"path": "config.json",
|
|
206
|
+
"sha256": "fb5117bf741e618e521113b70172979ece3f0fa99c81c2a5c168b2d704a8dd95",
|
|
207
|
+
"bytes": 551
|
|
208
|
+
}
|
|
209
|
+
]
|
|
210
|
+
},
|
|
211
|
+
{
|
|
212
|
+
"id": "ranker:startlux-0.8b",
|
|
213
|
+
"kind": "source",
|
|
214
|
+
"origin": "hf",
|
|
215
|
+
"optional": true,
|
|
216
|
+
"repo": "startlux-models/StartLux-Decision-v1-0.8B-Q8_0-GGUF",
|
|
217
|
+
"description": "可选：本地精排（llama-server）GGUF；首启不下载，按需 npm run fetch-models -- --include ranker",
|
|
218
|
+
"target": "startlux",
|
|
219
|
+
"files": [
|
|
220
|
+
{
|
|
221
|
+
"path": "StartLux-Decision-0.8B-Q8_0.gguf",
|
|
222
|
+
"sha256": "dd2bcf3ca81737c5a22e9a781c46e5389fc572a015ce7f36227722da0ba21457",
|
|
223
|
+
"bytes": 811843200,
|
|
224
|
+
"lfs": true
|
|
225
|
+
}
|
|
226
|
+
]
|
|
227
|
+
},
|
|
228
|
+
{
|
|
229
|
+
"id": "onnx:e5-small",
|
|
230
|
+
"kind": "source",
|
|
231
|
+
"origin": "hf",
|
|
232
|
+
"optional": true,
|
|
233
|
+
"repo": "Xenova/multilingual-e5-small",
|
|
234
|
+
"description": "可选：ONNX 本机 embedding 后端（G7.1；需 onnxruntime-node 可选依赖，默认不下载）",
|
|
235
|
+
"target": "onnx/e5-small",
|
|
236
|
+
"files": [
|
|
237
|
+
{
|
|
238
|
+
"path": "onnx/model_quantized.onnx",
|
|
239
|
+
"sha256": "f80102d3f2a1229f387d3c81909990d8945513e347b0eab049f7de3c6f98c193",
|
|
240
|
+
"bytes": 118308185,
|
|
241
|
+
"lfs": true
|
|
242
|
+
},
|
|
243
|
+
{
|
|
244
|
+
"path": "tokenizer.json",
|
|
245
|
+
"sha256": "0b44a9d7b51c3c62626640cda0e2c2f70fdacdc25bbbd68038369d14ebdf4c39",
|
|
246
|
+
"bytes": 17082730,
|
|
247
|
+
"lfs": true
|
|
248
|
+
}
|
|
249
|
+
]
|
|
250
|
+
}
|
|
251
|
+
],
|
|
252
|
+
"models": [
|
|
253
|
+
{
|
|
254
|
+
"id": "qwen3-embedding-8b",
|
|
255
|
+
"provider": "siliconflow",
|
|
256
|
+
"compute": "api",
|
|
257
|
+
"overrideField": "embed_primary",
|
|
258
|
+
"transport": {
|
|
259
|
+
"kind": "openai-embeddings",
|
|
260
|
+
"model": "Qwen/Qwen3-Embedding-8B"
|
|
261
|
+
},
|
|
262
|
+
"floors": {
|
|
263
|
+
"body": 0.75,
|
|
264
|
+
"summary": 0.85
|
|
265
|
+
},
|
|
266
|
+
"price": {
|
|
267
|
+
"cnyPerKtok": 0.00028
|
|
268
|
+
}
|
|
269
|
+
},
|
|
270
|
+
{
|
|
271
|
+
"id": "bge-m3",
|
|
272
|
+
"provider": "siliconflow",
|
|
273
|
+
"compute": "api",
|
|
274
|
+
"overrideField": "embed_fallback",
|
|
275
|
+
"transport": {
|
|
276
|
+
"kind": "openai-embeddings",
|
|
277
|
+
"model": "BAAI/bge-m3"
|
|
278
|
+
},
|
|
279
|
+
"floors": {
|
|
280
|
+
"body": 0.8,
|
|
281
|
+
"summary": 0.6
|
|
282
|
+
},
|
|
283
|
+
"price": {
|
|
284
|
+
"cnyPerKtok": 0
|
|
285
|
+
}
|
|
286
|
+
},
|
|
287
|
+
{
|
|
288
|
+
"id": "local:int8-256",
|
|
289
|
+
"provider": "local",
|
|
290
|
+
"compute": "local-static",
|
|
291
|
+
"artifactRef": "local:int8-256",
|
|
292
|
+
"overrideField": "embedder_model",
|
|
293
|
+
"floors": {
|
|
294
|
+
"body": 0.95,
|
|
295
|
+
"summary": 0.6
|
|
296
|
+
}
|
|
297
|
+
},
|
|
298
|
+
{
|
|
299
|
+
"id": "local:pca64-int8",
|
|
300
|
+
"provider": "local",
|
|
301
|
+
"compute": "local-static",
|
|
302
|
+
"artifactRef": "local:pca64-int8",
|
|
303
|
+
"overrideField": "embedder_model_last",
|
|
304
|
+
"floors": {
|
|
305
|
+
"body": 0.95,
|
|
306
|
+
"summary": 0.6
|
|
307
|
+
}
|
|
308
|
+
}
|
|
309
|
+
],
|
|
310
|
+
"tiers": [
|
|
311
|
+
{
|
|
312
|
+
"id": "qwen3-embedding-8b",
|
|
313
|
+
"modelRef": "qwen3-embedding-8b",
|
|
314
|
+
"floorPreset": "qwen3-embedding-8b"
|
|
315
|
+
},
|
|
316
|
+
{
|
|
317
|
+
"id": "bge-m3",
|
|
318
|
+
"modelRef": "bge-m3",
|
|
319
|
+
"floorPreset": "bge-m3"
|
|
320
|
+
},
|
|
321
|
+
{
|
|
322
|
+
"id": "local:int8-256",
|
|
323
|
+
"modelRef": "local:int8-256",
|
|
324
|
+
"floorPreset": "local:int8-256"
|
|
325
|
+
},
|
|
326
|
+
{
|
|
327
|
+
"id": "local:pca64-int8",
|
|
328
|
+
"modelRef": "local:pca64-int8",
|
|
329
|
+
"floorPreset": "local:pca64-int8"
|
|
330
|
+
}
|
|
331
|
+
],
|
|
332
|
+
"calibration": null
|
|
333
|
+
}
|
package/package.json
CHANGED
|
@@ -1,6 +1,36 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "@lmzhen/dsh-evolution-merge-engine",
|
|
3
|
-
"
|
|
4
|
-
"
|
|
5
|
-
"
|
|
6
|
-
|
|
3
|
+
"description": "Vendored skill-merge Node pipeline engine (byte-identical upstream) for DSH evolution-merge (community build)",
|
|
4
|
+
"version": "0.29.0",
|
|
5
|
+
"type": "module",
|
|
6
|
+
"main": "lib/index.js",
|
|
7
|
+
"types": "lib/index.d.ts",
|
|
8
|
+
"exports": {
|
|
9
|
+
".": {
|
|
10
|
+
"types": "./lib/index.d.ts",
|
|
11
|
+
"default": "./lib/index.js"
|
|
12
|
+
},
|
|
13
|
+
"./package.json": "./package.json"
|
|
14
|
+
},
|
|
15
|
+
"files": [
|
|
16
|
+
"lib/",
|
|
17
|
+
"data/",
|
|
18
|
+
"LICENSE",
|
|
19
|
+
"NOTICE",
|
|
20
|
+
"UPSTREAM.json"
|
|
21
|
+
],
|
|
22
|
+
"scripts": {
|
|
23
|
+
"build": "node scripts/build.mjs",
|
|
24
|
+
"test": "node --test \"vendor/skill-merge-pipeline/test/*.test.js\"",
|
|
25
|
+
"verify:upstream": "node verify-upstream.mjs"
|
|
26
|
+
},
|
|
27
|
+
"license": "MIT",
|
|
28
|
+
"engines": {
|
|
29
|
+
"dsh": ">=0.2.1-rc.1 <0.2.2-0"
|
|
30
|
+
},
|
|
31
|
+
"repository": {
|
|
32
|
+
"type": "git",
|
|
33
|
+
"url": "git+https://github.com/lmzhen/dsh-evolution.git",
|
|
34
|
+
"directory": "packages/evolution-merge-engine"
|
|
35
|
+
}
|
|
36
|
+
}
|
package/README.md
DELETED