toxzak commited on
Commit
eeff4cd
·
verified ·
1 Parent(s): dfd6c01

Add eval_results/mixed_budget_scan_8layers_g128.json

Browse files
eval_results/mixed_budget_scan_8layers_g128.json ADDED
@@ -0,0 +1,426 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "group_size": 128,
3
+ "outlier_options": [
4
+ 4,
5
+ 8
6
+ ],
7
+ "max_layers": 8,
8
+ "target_bpw": 4.115,
9
+ "uniform_int4": {
10
+ "layers": 8,
11
+ "total_params": 35782656,
12
+ "avg_bpw": 4.125,
13
+ "weighted_mse": 8.893806804755533e-06,
14
+ "weighted_rmse": 0.0029822486155174142,
15
+ "compression_vs_bf16": 3.878787878787879
16
+ },
17
+ "mixed_allocation": {
18
+ "layers": 8,
19
+ "total_params": 35782656,
20
+ "avg_bpw": 4.103021978021978,
21
+ "weighted_mse": 8.894368313197229e-06,
22
+ "weighted_rmse": 0.0029823427558208714,
23
+ "compression_vs_bf16": 3.899564780716438,
24
+ "target_avg_bpw": 4.115,
25
+ "selected_layers": [
26
+ {
27
+ "method": "groupwise_int4",
28
+ "bpw": 4.125,
29
+ "mse": 3.018915776920039e-06,
30
+ "idx": 0,
31
+ "key": "model.language_model.layers.0.mlp.down_proj.weight",
32
+ "params": 9437184,
33
+ "activation_weight": 1.0,
34
+ "weighted_mse": 3.018915776920039e-06,
35
+ "total_bits": 38928384.0,
36
+ "weighted_sse": 28.490063667297363
37
+ },
38
+ {
39
+ "method": "groupwise_int4",
40
+ "bpw": 4.125,
41
+ "mse": 1.0901781934080645e-05,
42
+ "idx": 1,
43
+ "key": "model.language_model.layers.0.mlp.gate_proj.weight",
44
+ "params": 9437184,
45
+ "activation_weight": 1.0,
46
+ "weighted_mse": 1.0901781934080645e-05,
47
+ "total_bits": 38928384.0,
48
+ "weighted_sse": 102.88212203979492
49
+ },
50
+ {
51
+ "method": "groupwise_int4",
52
+ "bpw": 4.125,
53
+ "mse": 1.0276894499838818e-05,
54
+ "idx": 2,
55
+ "key": "model.language_model.layers.0.mlp.up_proj.weight",
56
+ "params": 9437184,
57
+ "activation_weight": 1.0,
58
+ "weighted_mse": 1.0276894499838818e-05,
59
+ "total_bits": 38928384.0,
60
+ "weighted_sse": 96.9849443435669
61
+ },
62
+ {
63
+ "method": "int2_base",
64
+ "bpw": 2.125,
65
+ "mse": 5.5448953872883067e-08,
66
+ "idx": 3,
67
+ "key": "model.language_model.layers.0.per_layer_input_gate.weight",
68
+ "params": 393216,
69
+ "activation_weight": 1.0,
70
+ "weighted_mse": 5.5448953872883067e-08,
71
+ "total_bits": 835584.0,
72
+ "weighted_sse": 0.021803415846079588
73
+ },
74
+ {
75
+ "method": "groupwise_int4",
76
+ "bpw": 4.125,
77
+ "mse": 8.070513285929337e-05,
78
+ "idx": 4,
79
+ "key": "model.language_model.layers.0.per_layer_projection.weight",
80
+ "params": 393216,
81
+ "activation_weight": 1.0,
82
+ "weighted_mse": 8.070513285929337e-05,
83
+ "total_bits": 1622016.0,
84
+ "weighted_sse": 31.734549522399902
85
+ },
86
+ {
87
+ "method": "groupwise_int4",
88
+ "bpw": 4.125,
89
+ "mse": 9.958667760656681e-06,
90
+ "idx": 5,
91
+ "key": "model.language_model.layers.0.self_attn.k_proj.weight",
92
+ "params": 393216,
93
+ "activation_weight": 1.0,
94
+ "weighted_mse": 9.958667760656681e-06,
95
+ "total_bits": 1622016.0,
96
+ "weighted_sse": 3.9159075021743774
97
+ },
98
+ {
99
+ "method": "groupwise_int4",
100
+ "bpw": 4.125,
101
+ "mse": 7.164675025705947e-06,
102
+ "idx": 6,
103
+ "key": "model.language_model.layers.0.self_attn.o_proj.weight",
104
+ "params": 3145728,
105
+ "activation_weight": 1.0,
106
+ "weighted_mse": 7.164675025705947e-06,
107
+ "total_bits": 12976128.0,
108
+ "weighted_sse": 22.538118839263916
109
+ },
110
+ {
111
+ "method": "groupwise_int4",
112
+ "bpw": 4.125,
113
+ "mse": 1.0076081707666162e-05,
114
+ "idx": 7,
115
+ "key": "model.language_model.layers.0.self_attn.q_proj.weight",
116
+ "params": 3145728,
117
+ "activation_weight": 1.0,
118
+ "weighted_mse": 1.0076081707666162e-05,
119
+ "total_bits": 12976128.0,
120
+ "weighted_sse": 31.69661235809326
121
+ }
122
+ ],
123
+ "method_counts": {
124
+ "groupwise_int4": 7,
125
+ "int2_base": 1
126
+ }
127
+ },
128
+ "layers": [
129
+ {
130
+ "idx": 0,
131
+ "key": "model.language_model.layers.0.mlp.down_proj.weight",
132
+ "shape": [
133
+ 1536,
134
+ 6144
135
+ ],
136
+ "params": 9437184,
137
+ "activation_weight": 1.0,
138
+ "candidates": [
139
+ {
140
+ "method": "int2_base",
141
+ "bpw": 2.125,
142
+ "mse": 4.075739343534224e-05
143
+ },
144
+ {
145
+ "method": "int2_binary_residual",
146
+ "bpw": 3.25,
147
+ "mse": 1.2169458386779297e-05
148
+ },
149
+ {
150
+ "method": "int2_error_budget_k4",
151
+ "bpw": 3.71875,
152
+ "mse": 7.240188097057398e-06
153
+ },
154
+ {
155
+ "method": "int2_error_budget_k8",
156
+ "bpw": 4.0625,
157
+ "mse": 5.964981937722769e-06
158
+ },
159
+ {
160
+ "method": "groupwise_int4",
161
+ "bpw": 4.125,
162
+ "mse": 3.018915776920039e-06
163
+ }
164
+ ]
165
+ },
166
+ {
167
+ "idx": 1,
168
+ "key": "model.language_model.layers.0.mlp.gate_proj.weight",
169
+ "shape": [
170
+ 6144,
171
+ 1536
172
+ ],
173
+ "params": 9437184,
174
+ "activation_weight": 1.0,
175
+ "candidates": [
176
+ {
177
+ "method": "int2_base",
178
+ "bpw": 2.125,
179
+ "mse": 0.0001483042142353952
180
+ },
181
+ {
182
+ "method": "int2_binary_residual",
183
+ "bpw": 3.25,
184
+ "mse": 4.400406760396436e-05
185
+ },
186
+ {
187
+ "method": "int2_error_budget_k4",
188
+ "bpw": 3.71875,
189
+ "mse": 2.6316554794902913e-05
190
+ },
191
+ {
192
+ "method": "int2_error_budget_k8",
193
+ "bpw": 4.0625,
194
+ "mse": 2.1798021407448687e-05
195
+ },
196
+ {
197
+ "method": "groupwise_int4",
198
+ "bpw": 4.125,
199
+ "mse": 1.0901781934080645e-05
200
+ }
201
+ ]
202
+ },
203
+ {
204
+ "idx": 2,
205
+ "key": "model.language_model.layers.0.mlp.up_proj.weight",
206
+ "shape": [
207
+ 6144,
208
+ 1536
209
+ ],
210
+ "params": 9437184,
211
+ "activation_weight": 1.0,
212
+ "candidates": [
213
+ {
214
+ "method": "int2_base",
215
+ "bpw": 2.125,
216
+ "mse": 0.00014354493760038167
217
+ },
218
+ {
219
+ "method": "int2_binary_residual",
220
+ "bpw": 3.25,
221
+ "mse": 4.225481461617164e-05
222
+ },
223
+ {
224
+ "method": "int2_error_budget_k4",
225
+ "bpw": 3.71875,
226
+ "mse": 2.6116811568499543e-05
227
+ },
228
+ {
229
+ "method": "int2_error_budget_k8",
230
+ "bpw": 4.0625,
231
+ "mse": 2.1444240701384842e-05
232
+ },
233
+ {
234
+ "method": "groupwise_int4",
235
+ "bpw": 4.125,
236
+ "mse": 1.0276894499838818e-05
237
+ }
238
+ ]
239
+ },
240
+ {
241
+ "idx": 3,
242
+ "key": "model.language_model.layers.0.per_layer_input_gate.weight",
243
+ "shape": [
244
+ 256,
245
+ 1536
246
+ ],
247
+ "params": 393216,
248
+ "activation_weight": 1.0,
249
+ "candidates": [
250
+ {
251
+ "method": "int2_base",
252
+ "bpw": 2.125,
253
+ "mse": 5.5448953872883067e-08
254
+ },
255
+ {
256
+ "method": "int2_binary_residual",
257
+ "bpw": 3.25,
258
+ "mse": 1.7503493765502753e-08
259
+ },
260
+ {
261
+ "method": "int2_error_budget_k4",
262
+ "bpw": 3.71875,
263
+ "mse": 1.0123080151913655e-08
264
+ },
265
+ {
266
+ "method": "int2_error_budget_k8",
267
+ "bpw": 4.0625,
268
+ "mse": 8.346909829981541e-09
269
+ },
270
+ {
271
+ "method": "groupwise_int4",
272
+ "bpw": 4.125,
273
+ "mse": 4.351685678471995e-09
274
+ }
275
+ ]
276
+ },
277
+ {
278
+ "idx": 4,
279
+ "key": "model.language_model.layers.0.per_layer_projection.weight",
280
+ "shape": [
281
+ 1536,
282
+ 256
283
+ ],
284
+ "params": 393216,
285
+ "activation_weight": 1.0,
286
+ "candidates": [
287
+ {
288
+ "method": "int2_base",
289
+ "bpw": 2.125,
290
+ "mse": 0.0010429233079776168
291
+ },
292
+ {
293
+ "method": "int2_binary_residual",
294
+ "bpw": 3.25,
295
+ "mse": 0.0003290175518486649
296
+ },
297
+ {
298
+ "method": "int2_error_budget_k4",
299
+ "bpw": 3.71875,
300
+ "mse": 0.00019406182400416583
301
+ },
302
+ {
303
+ "method": "int2_error_budget_k8",
304
+ "bpw": 4.0625,
305
+ "mse": 0.00015925632033031434
306
+ },
307
+ {
308
+ "method": "groupwise_int4",
309
+ "bpw": 4.125,
310
+ "mse": 8.070513285929337e-05
311
+ }
312
+ ]
313
+ },
314
+ {
315
+ "idx": 5,
316
+ "key": "model.language_model.layers.0.self_attn.k_proj.weight",
317
+ "shape": [
318
+ 256,
319
+ 1536
320
+ ],
321
+ "params": 393216,
322
+ "activation_weight": 1.0,
323
+ "candidates": [
324
+ {
325
+ "method": "int2_base",
326
+ "bpw": 2.125,
327
+ "mse": 0.0001472694711992517
328
+ },
329
+ {
330
+ "method": "int2_binary_residual",
331
+ "bpw": 3.25,
332
+ "mse": 4.232315404806286e-05
333
+ },
334
+ {
335
+ "method": "int2_error_budget_k4",
336
+ "bpw": 3.71875,
337
+ "mse": 2.6914873160421848e-05
338
+ },
339
+ {
340
+ "method": "int2_error_budget_k8",
341
+ "bpw": 4.0625,
342
+ "mse": 2.2024454665370286e-05
343
+ },
344
+ {
345
+ "method": "groupwise_int4",
346
+ "bpw": 4.125,
347
+ "mse": 9.958667760656681e-06
348
+ }
349
+ ]
350
+ },
351
+ {
352
+ "idx": 6,
353
+ "key": "model.language_model.layers.0.self_attn.o_proj.weight",
354
+ "shape": [
355
+ 1536,
356
+ 2048
357
+ ],
358
+ "params": 3145728,
359
+ "activation_weight": 1.0,
360
+ "candidates": [
361
+ {
362
+ "method": "int2_base",
363
+ "bpw": 2.125,
364
+ "mse": 0.00010488332918612286
365
+ },
366
+ {
367
+ "method": "int2_binary_residual",
368
+ "bpw": 3.25,
369
+ "mse": 3.0286713808891363e-05
370
+ },
371
+ {
372
+ "method": "int2_error_budget_k4",
373
+ "bpw": 3.71875,
374
+ "mse": 1.9126162442262284e-05
375
+ },
376
+ {
377
+ "method": "int2_error_budget_k8",
378
+ "bpw": 4.0625,
379
+ "mse": 1.5659012206015177e-05
380
+ },
381
+ {
382
+ "method": "groupwise_int4",
383
+ "bpw": 4.125,
384
+ "mse": 7.164675025705947e-06
385
+ }
386
+ ]
387
+ },
388
+ {
389
+ "idx": 7,
390
+ "key": "model.language_model.layers.0.self_attn.q_proj.weight",
391
+ "shape": [
392
+ 2048,
393
+ 1536
394
+ ],
395
+ "params": 3145728,
396
+ "activation_weight": 1.0,
397
+ "candidates": [
398
+ {
399
+ "method": "int2_base",
400
+ "bpw": 2.125,
401
+ "mse": 0.00014882355753798038
402
+ },
403
+ {
404
+ "method": "int2_binary_residual",
405
+ "bpw": 3.25,
406
+ "mse": 4.262843867763877e-05
407
+ },
408
+ {
409
+ "method": "int2_error_budget_k4",
410
+ "bpw": 3.71875,
411
+ "mse": 2.7008029064745642e-05
412
+ },
413
+ {
414
+ "method": "int2_error_budget_k8",
415
+ "bpw": 4.0625,
416
+ "mse": 2.2143225578474812e-05
417
+ },
418
+ {
419
+ "method": "groupwise_int4",
420
+ "bpw": 4.125,
421
+ "mse": 1.0076081707666162e-05
422
+ }
423
+ ]
424
+ }
425
+ ]
426
+ }