File size: 164,451 Bytes
455a40f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
2027
2028
2029
2030
2031
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043
2044
2045
2046
2047
2048
2049
2050
2051
2052
2053
2054
2055
2056
2057
2058
2059
2060
2061
2062
2063
2064
2065
2066
2067
2068
2069
2070
2071
2072
2073
2074
2075
2076
2077
2078
2079
2080
2081
2082
2083
2084
2085
2086
2087
2088
2089
2090
2091
2092
2093
2094
2095
2096
2097
2098
2099
2100
2101
2102
2103
2104
2105
2106
2107
2108
2109
2110
2111
2112
2113
2114
2115
2116
2117
2118
2119
2120
2121
2122
2123
2124
2125
2126
2127
2128
2129
2130
2131
2132
2133
2134
2135
2136
2137
2138
2139
2140
2141
2142
2143
2144
2145
2146
2147
2148
2149
2150
2151
2152
2153
2154
2155
2156
2157
2158
2159
2160
2161
2162
2163
2164
2165
2166
2167
2168
2169
2170
2171
2172
2173
2174
2175
2176
2177
2178
2179
2180
2181
2182
2183
2184
2185
2186
2187
2188
2189
2190
2191
2192
2193
2194
2195
2196
2197
2198
2199
2200
2201
2202
2203
2204
2205
2206
2207
2208
2209
2210
2211
2212
2213
2214
2215
2216
2217
2218
2219
2220
2221
2222
2223
2224
2225
2226
2227
2228
2229
2230
2231
2232
2233
2234
2235
2236
2237
2238
2239
2240
2241
2242
2243
2244
2245
2246
2247
2248
2249
2250
2251
2252
2253
2254
2255
2256
2257
2258
2259
2260
2261
2262
2263
2264
2265
2266
2267
2268
2269
2270
2271
2272
2273
2274
2275
2276
2277
2278
2279
2280
2281
2282
2283
2284
2285
2286
2287
2288
2289
2290
2291
2292
2293
2294
2295
2296
2297
2298
2299
2300
2301
2302
2303
2304
2305
2306
2307
2308
2309
2310
2311
2312
2313
2314
2315
2316
2317
2318
2319
2320
2321
2322
2323
2324
2325
2326
2327
2328
2329
2330
2331
2332
2333
2334
2335
2336
2337
2338
2339
2340
2341
2342
2343
2344
2345
2346
2347
2348
2349
2350
2351
2352
2353
2354
2355
2356
2357
2358
2359
2360
2361
2362
2363
2364
2365
2366
2367
2368
2369
2370
2371
2372
2373
2374
2375
2376
2377
2378
2379
2380
2381
2382
2383
2384
2385
2386
2387
2388
2389
2390
2391
2392
2393
2394
2395
2396
2397
2398
2399
2400
2401
2402
2403
2404
2405
2406
2407
2408
2409
2410
2411
2412
2413
2414
2415
2416
2417
2418
2419
2420
2421
2422
2423
2424
2425
2426
2427
2428
2429
2430
2431
2432
2433
2434
2435
2436
2437
2438
2439
2440
2441
2442
2443
2444
2445
2446
2447
2448
2449
2450
2451
2452
2453
2454
2455
2456
2457
2458
2459
2460
2461
2462
2463
2464
2465
2466
2467
2468
2469
2470
2471
2472
2473
2474
2475
2476
2477
2478
2479
2480
2481
2482
2483
2484
2485
2486
2487
2488
2489
2490
2491
2492
2493
2494
2495
2496
2497
2498
2499
2500
2501
2502
2503
2504
2505
2506
2507
2508
2509
2510
2511
2512
2513
2514
2515
2516
2517
2518
2519
2520
2521
2522
2523
2524
2525
2526
2527
2528
2529
2530
2531
2532
2533
2534
2535
2536
2537
2538
2539
2540
2541
2542
2543
2544
2545
2546
2547
2548
2549
2550
2551
2552
2553
2554
2555
2556
2557
2558
2559
2560
2561
2562
2563
2564
2565
2566
2567
2568
2569
2570
2571
2572
2573
2574
2575
2576
2577
2578
2579
2580
2581
2582
2583
2584
2585
2586
2587
2588
2589
2590
2591
2592
2593
2594
2595
2596
2597
2598
2599
2600
2601
2602
2603
2604
2605
2606
2607
2608
2609
2610
2611
2612
2613
2614
2615
2616
2617
2618
2619
2620
2621
2622
2623
2624
2625
2626
2627
2628
2629
2630
2631
2632
2633
2634
2635
2636
2637
2638
2639
2640
2641
2642
2643
2644
2645
2646
2647
2648
2649
2650
2651
2652
2653
2654
2655
2656
2657
2658
2659
2660
2661
2662
2663
2664
2665
2666
2667
2668
2669
2670
2671
2672
2673
2674
2675
2676
2677
2678
2679
2680
2681
2682
2683
2684
2685
2686
2687
2688
2689
2690
2691
2692
2693
2694
2695
2696
2697
2698
2699
2700
2701
2702
2703
2704
2705
2706
2707
2708
2709
2710
2711
2712
2713
2714
2715
2716
2717
2718
2719
2720
2721
2722
2723
2724
2725
2726
2727
2728
2729
2730
2731
2732
2733
2734
2735
2736
2737
2738
2739
2740
2741
2742
2743
2744
2745
2746
2747
2748
2749
2750
2751
2752
2753
2754
2755
2756
2757
2758
2759
2760
2761
2762
2763
2764
2765
2766
2767
2768
2769
2770
2771
2772
2773
2774
2775
2776
2777
2778
2779
2780
2781
2782
2783
2784
2785
2786
2787
2788
2789
2790
2791
2792
2793
2794
2795
2796
2797
2798
2799
2800
2801
2802
2803
2804
2805
2806
2807
2808
2809
2810
2811
2812
2813
2814
2815
2816
2817
2818
2819
2820
2821
2822
2823
2824
2825
2826
2827
2828
2829
2830
2831
2832
2833
2834
2835
2836
2837
2838
2839
2840
2841
2842
2843
2844
2845
2846
2847
2848
2849
2850
2851
2852
2853
2854
2855
2856
2857
2858
2859
2860
2861
2862
2863
2864
2865
2866
2867
2868
2869
2870
2871
2872
2873
2874
2875
2876
2877
2878
2879
2880
2881
2882
2883
2884
2885
2886
2887
2888
2889
2890
2891
2892
2893
2894
2895
2896
2897
2898
2899
2900
2901
2902
2903
2904
2905
2906
2907
2908
2909
2910
2911
2912
2913
2914
2915
2916
2917
2918
2919
2920
2921
2922
2923
2924
2925
2926
2927
2928
2929
2930
2931
2932
2933
2934
2935
2936
2937
2938
2939
2940
2941
2942
2943
2944
2945
2946
2947
2948
2949
2950
2951
2952
2953
2954
2955
2956
2957
2958
2959
2960
2961
2962
2963
2964
2965
2966
2967
2968
2969
2970
2971
2972
2973
2974
2975
2976
2977
2978
2979
2980
2981
2982
2983
2984
2985
2986
2987
2988
2989
2990
2991
2992
2993
2994
2995
2996
2997
2998
2999
3000
3001
3002
3003
3004
3005
3006
3007
3008
3009
3010
3011
3012
3013
3014
3015
3016
3017
3018
3019
3020
3021
3022
3023
3024
3025
3026
3027
3028
3029
3030
3031
3032
3033
3034
3035
3036
3037
3038
3039
3040
3041
3042
3043
3044
3045
3046
3047
3048
3049
3050
3051
3052
3053
3054
3055
3056
3057
3058
3059
3060
3061
3062
3063
3064
3065
3066
3067
3068
3069
3070
3071
3072
3073
3074
3075
3076
3077
3078
3079
3080
3081
3082
3083
3084
3085
3086
3087
3088
3089
3090
3091
3092
3093
3094
3095
3096
3097
3098
3099
3100
3101
3102
3103
3104
3105
3106
3107
3108
3109
3110
3111
3112
3113
3114
3115
3116
3117
3118
3119
3120
3121
3122
3123
3124
3125
3126
3127
3128
3129
3130
3131
3132
3133
3134
3135
3136
3137
3138
3139
3140
3141
3142
3143
3144
3145
3146
3147
3148
3149
3150
3151
3152
3153
3154
3155
3156
3157
3158
3159
3160
3161
3162
3163
3164
3165
3166
3167
3168
3169
3170
3171
3172
3173
3174
3175
3176
3177
3178
3179
3180
3181
3182
3183
3184
3185
3186
3187
3188
3189
3190
3191
3192
3193
3194
3195
3196
3197
3198
3199
3200
3201
3202
3203
3204
3205
3206
3207
3208
3209
3210
3211
3212
3213
3214
3215
3216
3217
3218
3219
3220
3221
3222
3223
3224
3225
3226
3227
3228
3229
3230
3231
3232
3233
3234
3235
3236
3237
3238
3239
3240
3241
3242
3243
3244
3245
3246
3247
3248
3249
3250
3251
3252
3253
3254
3255
3256
3257
3258
3259
3260
3261
3262
3263
3264
3265
3266
3267
3268
3269
3270
3271
3272
3273
3274
3275
3276
3277
3278
3279
3280
3281
3282
3283
3284
3285
3286
3287
3288
3289
3290
3291
3292
3293
3294
3295
3296
3297
3298
3299
3300
3301
3302
3303
3304
3305
3306
3307
3308
3309
3310
3311
3312
3313
3314
3315
3316
3317
3318
3319
3320
3321
3322
3323
3324
3325
3326
3327
3328
3329
3330
3331
3332
3333
3334
3335
3336
3337
3338
3339
3340
3341
3342
3343
3344
3345
3346
3347
3348
3349
3350
3351
3352
3353
3354
3355
3356
3357
3358
3359
3360
3361
3362
3363
3364
3365
3366
3367
3368
3369
3370
3371
3372
3373
3374
3375
3376
3377
3378
3379
3380
3381
3382
3383
3384
3385
3386
3387
3388
3389
3390
3391
3392
3393
3394
3395
3396
3397
3398
3399
3400
3401
3402
3403
3404
3405
3406
3407
3408
3409
3410
3411
3412
3413
3414
3415
3416
3417
3418
3419
3420
3421
3422
3423
3424
3425
3426
3427
3428
3429
3430
3431
3432
3433
3434
3435
3436
3437
3438
3439
3440
3441
3442
3443
3444
3445
3446
3447
3448
3449
3450
3451
3452
3453
3454
3455
3456
3457
3458
3459
3460
3461
3462
3463
3464
3465
3466
3467
3468
3469
3470
3471
3472
3473
3474
3475
3476
3477
3478
3479
3480
3481
3482
3483
3484
3485
3486
3487
3488
3489
3490
3491
3492
3493
3494
3495
3496
3497
3498
3499
3500
3501
3502
3503
3504
3505
3506
3507
3508
3509
3510
3511
3512
3513
3514
3515
3516
3517
3518
3519
3520
3521
3522
3523
3524
3525
3526
3527
3528
3529
3530
3531
3532
3533
3534
3535
3536
3537
3538
3539
3540
3541
3542
3543
3544
3545
3546
3547
3548
3549
3550
3551
3552
3553
3554
3555
3556
3557
3558
3559
3560
3561
3562
3563
3564
3565
3566
3567
3568
3569
3570
3571
3572
3573
3574
3575
3576
3577
3578
3579
3580
3581
3582
3583
3584
3585
3586
3587
3588
3589
3590
3591
3592
3593
3594
3595
3596
3597
3598
3599
3600
3601
3602
3603
3604
3605
3606
3607
3608
3609
# coding=utf-8
# Copyright 2019 HuggingFace Inc.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import copy
import gc
import glob
import inspect
import json
import os
import os.path
import pickle
import random
import sys
import tempfile
import unittest
import unittest.mock as mock
import warnings
from collections import defaultdict
from pathlib import Path
from typing import Dict, List, Tuple

import numpy as np
from huggingface_hub import HfFolder, delete_repo
from huggingface_hub.file_download import http_get
from pytest import mark
from requests.exceptions import HTTPError

import transformers
from transformers import (
    AutoConfig,
    AutoModel,
    AutoModelForSequenceClassification,
    PretrainedConfig,
    is_torch_available,
    logging,
)
from transformers.models.auto import get_values
from transformers.models.auto.modeling_auto import (
    MODEL_FOR_AUDIO_CLASSIFICATION_MAPPING_NAMES,
    MODEL_FOR_AUDIO_XVECTOR_MAPPING_NAMES,
    MODEL_FOR_BACKBONE_MAPPING_NAMES,
    MODEL_FOR_CAUSAL_IMAGE_MODELING_MAPPING_NAMES,
    MODEL_FOR_CAUSAL_LM_MAPPING_NAMES,
    MODEL_FOR_DOCUMENT_QUESTION_ANSWERING_MAPPING_NAMES,
    MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING_NAMES,
    MODEL_FOR_MASKED_IMAGE_MODELING_MAPPING_NAMES,
    MODEL_FOR_MASKED_LM_MAPPING_NAMES,
    MODEL_FOR_MULTIPLE_CHOICE_MAPPING_NAMES,
    MODEL_FOR_NEXT_SENTENCE_PREDICTION_MAPPING_NAMES,
    MODEL_FOR_QUESTION_ANSWERING_MAPPING_NAMES,
    MODEL_FOR_SEMANTIC_SEGMENTATION_MAPPING_NAMES,
    MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING_NAMES,
    MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES,
    MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING_NAMES,
    MODEL_FOR_VIDEO_CLASSIFICATION_MAPPING_NAMES,
    MODEL_MAPPING_NAMES,
)
from transformers.testing_utils import (
    TOKEN,
    USER,
    CaptureLogger,
    TestCasePlus,
    is_pt_flax_cross_test,
    is_pt_tf_cross_test,
    is_staging_test,
    require_accelerate,
    require_safetensors,
    require_torch,
    require_torch_gpu,
    require_torch_multi_gpu,
    require_usr_bin_time,
    slow,
    torch_device,
)
from transformers.utils import (
    CONFIG_NAME,
    GENERATION_CONFIG_NAME,
    SAFE_WEIGHTS_INDEX_NAME,
    SAFE_WEIGHTS_NAME,
    WEIGHTS_INDEX_NAME,
    WEIGHTS_NAME,
    is_accelerate_available,
    is_flax_available,
    is_tf_available,
    is_torch_fx_available,
)
from transformers.utils.generic import ModelOutput


sys.path.append(str(Path(__file__).parent.parent / "utils"))

from test_module.custom_configuration import CustomConfig, NoSuperInitConfig  # noqa E402


if is_accelerate_available():
    from accelerate.utils import compute_module_sizes


if is_torch_available():
    import torch
    from test_module.custom_modeling import CustomModel, NoSuperInitModel
    from torch import nn

    from transformers import (
        BERT_PRETRAINED_MODEL_ARCHIVE_LIST,
        MODEL_MAPPING,
        AdaptiveEmbedding,
        AutoModelForCausalLM,
        AutoTokenizer,
        BertConfig,
        BertModel,
        CLIPTextModel,
        PreTrainedModel,
        T5Config,
        T5ForConditionalGeneration,
    )
    from transformers.modeling_utils import shard_checkpoint

    # Fake pretrained models for tests
    class BaseModel(PreTrainedModel):
        config_class = PretrainedConfig

        def __init__(self, config):
            super().__init__(config)
            self.linear = nn.Linear(4, 5)
            self.linear_2 = nn.Linear(5, 6)

        def forward(self, x):
            return self.linear_2(self.linear(x))

    class ModelWithHead(PreTrainedModel):
        base_model_prefix = "base"
        config_class = PretrainedConfig

        def _init_weights(self, module):
            pass

        def __init__(self, config):
            super().__init__(config)
            self.base = BaseModel(config)
            # linear is a common name between Base and Head on purpose.
            self.linear = nn.Linear(6, 3)
            self.linear2 = nn.Linear(3, 5)

        def forward(self, x):
            return self.linear2(self.linear(self.base(x)))


if is_tf_available():
    import tensorflow as tf

if is_flax_available():
    import jax.numpy as jnp

    from transformers.modeling_flax_pytorch_utils import (
        convert_pytorch_state_dict_to_flax,
        load_flax_weights_in_pytorch_model,
    )

if is_torch_fx_available():
    from transformers.utils.fx import symbolic_trace


def _config_zero_init(config):
    configs_no_init = copy.deepcopy(config)
    for key in configs_no_init.__dict__.keys():
        if "_range" in key or "_std" in key or "initializer_factor" in key or "layer_scale" in key:
            setattr(configs_no_init, key, 1e-10)
        if isinstance(getattr(configs_no_init, key, None), PretrainedConfig):
            no_init_subconfig = _config_zero_init(getattr(configs_no_init, key))
            setattr(configs_no_init, key, no_init_subconfig)
    return configs_no_init


TINY_T5 = "patrickvonplaten/t5-tiny-random"
TINY_BERT_FOR_TOKEN_CLASSIFICATION = "hf-internal-testing/tiny-bert-for-token-classification"


def _mock_init_weights(self, module):
    for name, param in module.named_parameters(recurse=False):
        # Use the first letter of the name to get a value and go from a <> -13 to z <> 12
        value = ord(name[0].lower()) - 110
        param.data.fill_(value)


def _mock_all_init_weights(self):
    # Prune heads if needed
    if self.config.pruned_heads:
        self.prune_heads(self.config.pruned_heads)

    import transformers.modeling_utils

    if transformers.modeling_utils._init_weights:
        for module in self.modules():
            module._is_hf_initialized = False
        # Initialize weights
        self.apply(self._initialize_weights)

        # Tie weights should be skipped when not initializing all weights
        # since from_pretrained(...) calls tie weights anyways
        self.tie_weights()


@require_torch
class ModelTesterMixin:
    model_tester = None
    all_model_classes = ()
    all_generative_model_classes = ()
    fx_compatible = False
    test_torchscript = True
    test_pruning = True
    test_resize_embeddings = True
    test_resize_position_embeddings = False
    test_head_masking = True
    test_mismatched_shapes = True
    test_missing_keys = True
    test_model_parallel = False
    is_encoder_decoder = False
    has_attentions = True
    model_split_percents = [0.5, 0.7, 0.9]

    def _prepare_for_class(self, inputs_dict, model_class, return_labels=False):
        inputs_dict = copy.deepcopy(inputs_dict)
        if model_class.__name__ in get_values(MODEL_FOR_MULTIPLE_CHOICE_MAPPING_NAMES):
            inputs_dict = {
                k: v.unsqueeze(1).expand(-1, self.model_tester.num_choices, -1).contiguous()
                if isinstance(v, torch.Tensor) and v.ndim > 1
                else v
                for k, v in inputs_dict.items()
            }
        elif model_class.__name__ in get_values(MODEL_FOR_AUDIO_XVECTOR_MAPPING_NAMES):
            inputs_dict.pop("attention_mask")

        if return_labels:
            if model_class.__name__ in get_values(MODEL_FOR_MULTIPLE_CHOICE_MAPPING_NAMES):
                inputs_dict["labels"] = torch.ones(self.model_tester.batch_size, dtype=torch.long, device=torch_device)
            elif model_class.__name__ in [
                *get_values(MODEL_FOR_QUESTION_ANSWERING_MAPPING_NAMES),
                *get_values(MODEL_FOR_DOCUMENT_QUESTION_ANSWERING_MAPPING_NAMES),
            ]:
                inputs_dict["start_positions"] = torch.zeros(
                    self.model_tester.batch_size, dtype=torch.long, device=torch_device
                )
                inputs_dict["end_positions"] = torch.zeros(
                    self.model_tester.batch_size, dtype=torch.long, device=torch_device
                )
            elif model_class.__name__ in [
                *get_values(MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES),
                *get_values(MODEL_FOR_NEXT_SENTENCE_PREDICTION_MAPPING_NAMES),
                *get_values(MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING_NAMES),
                *get_values(MODEL_FOR_VIDEO_CLASSIFICATION_MAPPING_NAMES),
                *get_values(MODEL_FOR_AUDIO_CLASSIFICATION_MAPPING_NAMES),
            ]:
                inputs_dict["labels"] = torch.zeros(
                    self.model_tester.batch_size, dtype=torch.long, device=torch_device
                )
            elif model_class.__name__ in [
                *get_values(MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING_NAMES),
                *get_values(MODEL_FOR_CAUSAL_LM_MAPPING_NAMES),
                *get_values(MODEL_FOR_CAUSAL_IMAGE_MODELING_MAPPING_NAMES),
                *get_values(MODEL_FOR_MASKED_LM_MAPPING_NAMES),
                *get_values(MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING_NAMES),
            ]:
                inputs_dict["labels"] = torch.zeros(
                    (self.model_tester.batch_size, self.model_tester.seq_length), dtype=torch.long, device=torch_device
                )
            elif model_class.__name__ in get_values(MODEL_FOR_MASKED_IMAGE_MODELING_MAPPING_NAMES):
                num_patches = self.model_tester.image_size // self.model_tester.patch_size
                inputs_dict["bool_masked_pos"] = torch.zeros(
                    (self.model_tester.batch_size, num_patches**2), dtype=torch.long, device=torch_device
                )
            elif model_class.__name__ in get_values(MODEL_FOR_SEMANTIC_SEGMENTATION_MAPPING_NAMES):
                batch_size, num_channels, height, width = inputs_dict["pixel_values"].shape
                inputs_dict["labels"] = torch.zeros(
                    [self.model_tester.batch_size, height, width], device=torch_device
                ).long()

        return inputs_dict

    def test_save_load(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        def check_save_load(out1, out2):
            # make sure we don't have nans
            out_2 = out2.cpu().numpy()
            out_2[np.isnan(out_2)] = 0

            out_1 = out1.cpu().numpy()
            out_1[np.isnan(out_1)] = 0
            max_diff = np.amax(np.abs(out_1 - out_2))
            self.assertLessEqual(max_diff, 1e-5)

        for model_class in self.all_model_classes:
            model = model_class(config)
            model.to(torch_device)
            model.eval()
            with torch.no_grad():
                first = model(**self._prepare_for_class(inputs_dict, model_class))[0]

            with tempfile.TemporaryDirectory() as tmpdirname:
                model.save_pretrained(tmpdirname)

                # the config file (and the generation config file, if it can generate) should be saved
                self.assertTrue(os.path.exists(os.path.join(tmpdirname, CONFIG_NAME)))
                self.assertEqual(
                    model.can_generate(), os.path.exists(os.path.join(tmpdirname, GENERATION_CONFIG_NAME))
                )

                model = model_class.from_pretrained(tmpdirname)
                model.to(torch_device)
                with torch.no_grad():
                    second = model(**self._prepare_for_class(inputs_dict, model_class))[0]

            if isinstance(first, tuple) and isinstance(second, tuple):
                for tensor1, tensor2 in zip(first, second):
                    check_save_load(tensor1, tensor2)
            else:
                check_save_load(first, second)

    def test_from_pretrained_no_checkpoint(self):
        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
        for model_class in self.all_model_classes:
            model = model_class(config)
            state_dict = model.state_dict()

            new_model = model_class.from_pretrained(
                pretrained_model_name_or_path=None, config=config, state_dict=state_dict
            )
            for p1, p2 in zip(model.parameters(), new_model.parameters()):
                self.assertTrue(torch.equal(p1, p2))

    def test_save_load_keys_to_ignore_on_save(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            model = model_class(config)
            _keys_to_ignore_on_save = getattr(model, "_keys_to_ignore_on_save", None)
            if _keys_to_ignore_on_save is None:
                continue

            # check the keys are in the original state_dict
            for k in _keys_to_ignore_on_save:
                self.assertIn(k, model.state_dict().keys(), "\n".join(model.state_dict().keys()))

            # check that certain keys didn't get saved with the model
            with tempfile.TemporaryDirectory() as tmpdirname:
                model.save_pretrained(tmpdirname)
                output_model_file = os.path.join(tmpdirname, WEIGHTS_NAME)
                state_dict_saved = torch.load(output_model_file)
                for k in _keys_to_ignore_on_save:
                    self.assertNotIn(k, state_dict_saved.keys(), "\n".join(state_dict_saved.keys()))

                # Test we can load the state dict in the model, necessary for the checkpointing API in Trainer.
                load_result = model.load_state_dict(state_dict_saved, strict=False)
                self.assertTrue(
                    len(load_result.missing_keys) == 0
                    or set(load_result.missing_keys) == set(model._keys_to_ignore_on_save)
                )
                self.assertTrue(len(load_result.unexpected_keys) == 0)

    def test_gradient_checkpointing_backward_compatibility(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            if not model_class.supports_gradient_checkpointing:
                continue

            config.gradient_checkpointing = True
            model = model_class(config)
            self.assertTrue(model.is_gradient_checkpointing)

    def test_gradient_checkpointing_enable_disable(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            if not model_class.supports_gradient_checkpointing:
                continue

            # at init model should have gradient checkpointing disabled
            model = model_class(config)
            self.assertFalse(model.is_gradient_checkpointing)

            # check enable works
            model.gradient_checkpointing_enable()
            self.assertTrue(model.is_gradient_checkpointing)

            # check disable works
            model.gradient_checkpointing_disable()
            self.assertFalse(model.is_gradient_checkpointing)

    def test_save_load_fast_init_from_base(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
        if config.__class__ not in MODEL_MAPPING:
            return
        base_class = MODEL_MAPPING[config.__class__]

        if isinstance(base_class, tuple):
            base_class = base_class[0]

        for model_class in self.all_model_classes:
            if model_class == base_class:
                continue

            # make a copy of model class to not break future tests
            # from https://stackoverflow.com/questions/9541025/how-to-copy-a-python-class
            class CopyClass(model_class):
                pass

            model_class_copy = CopyClass

            # make sure that all keys are expected for test
            model_class_copy._keys_to_ignore_on_load_missing = []

            # make init deterministic, but make sure that
            # non-initialized weights throw errors nevertheless
            model_class_copy._init_weights = _mock_init_weights
            model_class_copy.init_weights = _mock_all_init_weights

            model = base_class(config)
            state_dict = model.state_dict()

            # this will often delete a single weight of a multi-weight module
            # to test an edge case
            random_key_to_del = random.choice(list(state_dict.keys()))
            del state_dict[random_key_to_del]

            # check that certain keys didn't get saved with the model
            with tempfile.TemporaryDirectory() as tmpdirname:
                model.save_pretrained(tmpdirname)
                torch.save(state_dict, os.path.join(tmpdirname, "pytorch_model.bin"))

                model_fast_init = model_class_copy.from_pretrained(tmpdirname)
                model_slow_init = model_class_copy.from_pretrained(tmpdirname, _fast_init=False)
                # Before we test anything

                for key in model_fast_init.state_dict().keys():
                    if isinstance(model_slow_init.state_dict()[key], torch.BoolTensor):
                        max_diff = (model_slow_init.state_dict()[key] ^ model_fast_init.state_dict()[key]).sum().item()
                    else:
                        max_diff = (model_slow_init.state_dict()[key] - model_fast_init.state_dict()[key]).sum().item()
                    self.assertLessEqual(max_diff, 1e-3, msg=f"{key} not identical")

    def test_save_load_fast_init_to_base(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
        if config.__class__ not in MODEL_MAPPING:
            return
        base_class = MODEL_MAPPING[config.__class__]

        if isinstance(base_class, tuple):
            base_class = base_class[0]

        for model_class in self.all_model_classes:
            if model_class == base_class:
                continue

            # make a copy of model class to not break future tests
            # from https://stackoverflow.com/questions/9541025/how-to-copy-a-python-class
            class CopyClass(base_class):
                pass

            base_class_copy = CopyClass

            # make sure that all keys are expected for test
            base_class_copy._keys_to_ignore_on_load_missing = []

            # make init deterministic, but make sure that
            # non-initialized weights throw errors nevertheless
            base_class_copy._init_weights = _mock_init_weights
            base_class_copy.init_weights = _mock_all_init_weights

            model = model_class(config)
            state_dict = model.state_dict()

            # this will often delete a single weight of a multi-weight module
            # to test an edge case
            random_key_to_del = random.choice(list(state_dict.keys()))
            del state_dict[random_key_to_del]

            # check that certain keys didn't get saved with the model
            with tempfile.TemporaryDirectory() as tmpdirname:
                model.config.save_pretrained(tmpdirname)
                torch.save(state_dict, os.path.join(tmpdirname, "pytorch_model.bin"))

                model_fast_init = base_class_copy.from_pretrained(tmpdirname)
                model_slow_init = base_class_copy.from_pretrained(tmpdirname, _fast_init=False)

                for key in model_fast_init.state_dict().keys():
                    if isinstance(model_slow_init.state_dict()[key], torch.BoolTensor):
                        max_diff = torch.max(
                            model_slow_init.state_dict()[key] ^ model_fast_init.state_dict()[key]
                        ).item()
                    else:
                        max_diff = torch.max(
                            torch.abs(model_slow_init.state_dict()[key] - model_fast_init.state_dict()[key])
                        ).item()
                    self.assertLessEqual(max_diff, 1e-3, msg=f"{key} not identical")

    def test_initialization(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        configs_no_init = _config_zero_init(config)
        for model_class in self.all_model_classes:
            model = model_class(config=configs_no_init)
            for name, param in model.named_parameters():
                if param.requires_grad:
                    self.assertIn(
                        ((param.data.mean() * 1e9).round() / 1e9).item(),
                        [0.0, 1.0],
                        msg=f"Parameter {name} of model {model_class} seems not properly initialized",
                    )

    def test_determinism(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        def check_determinism(first, second):
            out_1 = first.cpu().numpy()
            out_2 = second.cpu().numpy()
            out_1 = out_1[~np.isnan(out_1)]
            out_2 = out_2[~np.isnan(out_2)]
            max_diff = np.amax(np.abs(out_1 - out_2))
            self.assertLessEqual(max_diff, 1e-5)

        for model_class in self.all_model_classes:
            model = model_class(config)
            model.to(torch_device)
            model.eval()
            with torch.no_grad():
                first = model(**self._prepare_for_class(inputs_dict, model_class))[0]
                second = model(**self._prepare_for_class(inputs_dict, model_class))[0]

            if isinstance(first, tuple) and isinstance(second, tuple):
                for tensor1, tensor2 in zip(first, second):
                    check_determinism(tensor1, tensor2)
            else:
                check_determinism(first, second)

    def test_forward_signature(self):
        config, _ = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            model = model_class(config)
            signature = inspect.signature(model.forward)
            # signature.parameters is an OrderedDict => so arg_names order is deterministic
            arg_names = [*signature.parameters.keys()]

            if model.config.is_encoder_decoder:
                expected_arg_names = [
                    "input_ids",
                    "attention_mask",
                    "decoder_input_ids",
                    "decoder_attention_mask",
                ]
                expected_arg_names.extend(
                    ["head_mask", "decoder_head_mask", "cross_attn_head_mask", "encoder_outputs"]
                    if "head_mask" and "decoder_head_mask" and "cross_attn_head_mask" in arg_names
                    else ["encoder_outputs"]
                )
                self.assertListEqual(arg_names[: len(expected_arg_names)], expected_arg_names)
            else:
                expected_arg_names = ["input_ids"]
                self.assertListEqual(arg_names[:1], expected_arg_names)

    def test_training(self):
        if not self.model_tester.is_training:
            return

        for model_class in self.all_model_classes:
            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
            config.return_dict = True

            if model_class.__name__ in [
                *get_values(MODEL_MAPPING_NAMES),
                *get_values(MODEL_FOR_BACKBONE_MAPPING_NAMES),
            ]:
                continue

            model = model_class(config)
            model.to(torch_device)
            model.train()
            inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
            loss = model(**inputs).loss
            loss.backward()

    def test_training_gradient_checkpointing(self):
        if not self.model_tester.is_training:
            return

        for model_class in self.all_model_classes:
            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
            config.use_cache = False
            config.return_dict = True

            if (
                model_class.__name__
                in [*get_values(MODEL_MAPPING_NAMES), *get_values(MODEL_FOR_BACKBONE_MAPPING_NAMES)]
                or not model_class.supports_gradient_checkpointing
            ):
                continue
            model = model_class(config)
            model.to(torch_device)
            model.gradient_checkpointing_enable()
            model.train()
            inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
            loss = model(**inputs).loss
            loss.backward()

    def test_attention_outputs(self):
        if not self.has_attentions:
            self.skipTest(reason="Model does not output attentions")

        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
        config.return_dict = True

        seq_len = getattr(self.model_tester, "seq_length", None)
        decoder_seq_length = getattr(self.model_tester, "decoder_seq_length", seq_len)
        encoder_seq_length = getattr(self.model_tester, "encoder_seq_length", seq_len)
        decoder_key_length = getattr(self.model_tester, "decoder_key_length", decoder_seq_length)
        encoder_key_length = getattr(self.model_tester, "key_length", encoder_seq_length)
        chunk_length = getattr(self.model_tester, "chunk_length", None)
        if chunk_length is not None and hasattr(self.model_tester, "num_hashes"):
            encoder_seq_length = encoder_seq_length * self.model_tester.num_hashes

        for model_class in self.all_model_classes:
            inputs_dict["output_attentions"] = True
            inputs_dict["output_hidden_states"] = False
            config.return_dict = True
            model = model_class(config)
            model.to(torch_device)
            model.eval()
            with torch.no_grad():
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
            attentions = outputs.encoder_attentions if config.is_encoder_decoder else outputs.attentions
            self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)

            # check that output_attentions also work using config
            del inputs_dict["output_attentions"]
            config.output_attentions = True
            model = model_class(config)
            model.to(torch_device)
            model.eval()
            with torch.no_grad():
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
            attentions = outputs.encoder_attentions if config.is_encoder_decoder else outputs.attentions
            self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)

            if chunk_length is not None:
                self.assertListEqual(
                    list(attentions[0].shape[-4:]),
                    [self.model_tester.num_attention_heads, encoder_seq_length, chunk_length, encoder_key_length],
                )
            else:
                self.assertListEqual(
                    list(attentions[0].shape[-3:]),
                    [self.model_tester.num_attention_heads, encoder_seq_length, encoder_key_length],
                )
            out_len = len(outputs)

            if self.is_encoder_decoder:
                correct_outlen = 5

                # loss is at first position
                if "labels" in inputs_dict:
                    correct_outlen += 1  # loss is added to beginning
                # Question Answering model returns start_logits and end_logits
                if model_class.__name__ in [
                    *get_values(MODEL_FOR_QUESTION_ANSWERING_MAPPING_NAMES),
                    *get_values(MODEL_FOR_DOCUMENT_QUESTION_ANSWERING_MAPPING_NAMES),
                ]:
                    correct_outlen += 1  # start_logits and end_logits instead of only 1 output
                if "past_key_values" in outputs:
                    correct_outlen += 1  # past_key_values have been returned

                self.assertEqual(out_len, correct_outlen)

                # decoder attentions
                decoder_attentions = outputs.decoder_attentions
                self.assertIsInstance(decoder_attentions, (list, tuple))
                self.assertEqual(len(decoder_attentions), self.model_tester.num_hidden_layers)
                self.assertListEqual(
                    list(decoder_attentions[0].shape[-3:]),
                    [self.model_tester.num_attention_heads, decoder_seq_length, decoder_key_length],
                )

                # cross attentions
                cross_attentions = outputs.cross_attentions
                self.assertIsInstance(cross_attentions, (list, tuple))
                self.assertEqual(len(cross_attentions), self.model_tester.num_hidden_layers)
                self.assertListEqual(
                    list(cross_attentions[0].shape[-3:]),
                    [
                        self.model_tester.num_attention_heads,
                        decoder_seq_length,
                        encoder_key_length,
                    ],
                )

            # Check attention is always last and order is fine
            inputs_dict["output_attentions"] = True
            inputs_dict["output_hidden_states"] = True
            model = model_class(config)
            model.to(torch_device)
            model.eval()
            with torch.no_grad():
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))

            if hasattr(self.model_tester, "num_hidden_states_types"):
                added_hidden_states = self.model_tester.num_hidden_states_types
            elif self.is_encoder_decoder:
                added_hidden_states = 2
            else:
                added_hidden_states = 1
            self.assertEqual(out_len + added_hidden_states, len(outputs))

            self_attentions = outputs.encoder_attentions if config.is_encoder_decoder else outputs.attentions

            self.assertEqual(len(self_attentions), self.model_tester.num_hidden_layers)
            if chunk_length is not None:
                self.assertListEqual(
                    list(self_attentions[0].shape[-4:]),
                    [self.model_tester.num_attention_heads, encoder_seq_length, chunk_length, encoder_key_length],
                )
            else:
                self.assertListEqual(
                    list(self_attentions[0].shape[-3:]),
                    [self.model_tester.num_attention_heads, encoder_seq_length, encoder_key_length],
                )

    @slow
    def test_torchscript_simple(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
        self._create_and_check_torchscript(config, inputs_dict)

    @slow
    def test_torchscript_output_attentions(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
        config.output_attentions = True
        self._create_and_check_torchscript(config, inputs_dict)

    @slow
    def test_torchscript_output_hidden_state(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
        config.output_hidden_states = True
        self._create_and_check_torchscript(config, inputs_dict)

    # This is copied from `torch/testing/_internal/jit_utils.py::clear_class_registry`
    def clear_torch_jit_class_registry(self):
        torch._C._jit_clear_class_registry()
        torch.jit._recursive.concrete_type_store = torch.jit._recursive.ConcreteTypeStore()
        # torch 1.8 has no `_clear_class_state` in `torch.jit._state`
        if hasattr(torch.jit._state, "_clear_class_state"):
            torch.jit._state._clear_class_state()

    def _create_and_check_torchscript(self, config, inputs_dict):
        if not self.test_torchscript:
            return

        configs_no_init = _config_zero_init(config)  # To be sure we have no Nan
        configs_no_init.torchscript = True
        for model_class in self.all_model_classes:
            model = model_class(config=configs_no_init)
            model.to(torch_device)
            model.eval()
            inputs = self._prepare_for_class(inputs_dict, model_class)

            main_input_name = model_class.main_input_name

            try:
                if model.config.is_encoder_decoder:
                    model.config.use_cache = False  # FSTM still requires this hack -> FSTM should probably be refactored similar to BART afterward
                    main_input = inputs[main_input_name]
                    attention_mask = inputs["attention_mask"]
                    decoder_input_ids = inputs["decoder_input_ids"]
                    decoder_attention_mask = inputs["decoder_attention_mask"]
                    model(main_input, attention_mask, decoder_input_ids, decoder_attention_mask)
                    traced_model = torch.jit.trace(
                        model, (main_input, attention_mask, decoder_input_ids, decoder_attention_mask)
                    )
                elif "bbox" in inputs and "image" in inputs:  # LayoutLMv2 requires additional inputs
                    input_ids = inputs["input_ids"]
                    bbox = inputs["bbox"]
                    image = inputs["image"].tensor
                    model(input_ids, bbox, image)
                    traced_model = torch.jit.trace(
                        model, (input_ids, bbox, image), check_trace=False
                    )  # when traced model is checked, an error is produced due to name mangling
                else:
                    main_input = inputs[main_input_name]
                    model(main_input)
                    traced_model = torch.jit.trace(model, main_input)
            except RuntimeError:
                self.fail("Couldn't trace module.")

            with tempfile.TemporaryDirectory() as tmp_dir_name:
                pt_file_name = os.path.join(tmp_dir_name, "traced_model.pt")

                try:
                    torch.jit.save(traced_model, pt_file_name)
                except Exception:
                    self.fail("Couldn't save module.")

                try:
                    loaded_model = torch.jit.load(pt_file_name)
                except Exception:
                    self.fail("Couldn't load module.")

            model.to(torch_device)
            model.eval()

            loaded_model.to(torch_device)
            loaded_model.eval()

            model_state_dict = model.state_dict()
            loaded_model_state_dict = loaded_model.state_dict()

            non_persistent_buffers = {}
            for key in loaded_model_state_dict.keys():
                if key not in model_state_dict.keys():
                    non_persistent_buffers[key] = loaded_model_state_dict[key]

            loaded_model_state_dict = {
                key: value for key, value in loaded_model_state_dict.items() if key not in non_persistent_buffers
            }

            self.assertEqual(set(model_state_dict.keys()), set(loaded_model_state_dict.keys()))

            model_buffers = list(model.buffers())
            for non_persistent_buffer in non_persistent_buffers.values():
                found_buffer = False
                for i, model_buffer in enumerate(model_buffers):
                    if torch.equal(non_persistent_buffer, model_buffer):
                        found_buffer = True
                        break

                self.assertTrue(found_buffer)
                model_buffers.pop(i)

            models_equal = True
            for layer_name, p1 in model_state_dict.items():
                if layer_name in loaded_model_state_dict:
                    p2 = loaded_model_state_dict[layer_name]
                    if p1.data.ne(p2.data).sum() > 0:
                        models_equal = False

            self.assertTrue(models_equal)

            # Avoid memory leak. Without this, each call increase RAM usage by ~20MB.
            # (Even with this call, there are still memory leak by ~0.04MB)
            self.clear_torch_jit_class_registry()

    def test_torch_fx(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
        self._create_and_check_torch_fx_tracing(config, inputs_dict)

    def test_torch_fx_output_loss(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
        self._create_and_check_torch_fx_tracing(config, inputs_dict, output_loss=True)

    def _create_and_check_torch_fx_tracing(self, config, inputs_dict, output_loss=False):
        if not is_torch_fx_available() or not self.fx_compatible:
            return

        configs_no_init = _config_zero_init(config)  # To be sure we have no Nan
        configs_no_init.return_dict = False

        for model_class in self.all_model_classes:
            model = model_class(config=configs_no_init)
            model.to(torch_device)
            model.eval()
            inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=output_loss)

            try:
                if model.config.is_encoder_decoder:
                    model.config.use_cache = False  # FSTM still requires this hack -> FSTM should probably be refactored similar to BART afterward
                    labels = inputs.get("labels", None)
                    input_names = [
                        "attention_mask",
                        "decoder_attention_mask",
                        "decoder_input_ids",
                        "input_features",
                        "input_ids",
                        "input_values",
                    ]
                    if labels is not None:
                        input_names.append("labels")

                    filtered_inputs = {k: v for (k, v) in inputs.items() if k in input_names}
                    input_names = list(filtered_inputs.keys())

                    model_output = model(**filtered_inputs)

                    traced_model = symbolic_trace(model, input_names)
                    traced_output = traced_model(**filtered_inputs)
                else:
                    input_names = [
                        "attention_mask",
                        "bbox",
                        "input_features",
                        "input_ids",
                        "input_values",
                        "pixel_values",
                        "token_type_ids",
                        "visual_feats",
                        "visual_pos",
                    ]

                    labels = inputs.get("labels", None)
                    start_positions = inputs.get("start_positions", None)
                    end_positions = inputs.get("end_positions", None)
                    if labels is not None:
                        input_names.append("labels")
                    if start_positions is not None:
                        input_names.append("start_positions")
                    if end_positions is not None:
                        input_names.append("end_positions")

                    filtered_inputs = {k: v for (k, v) in inputs.items() if k in input_names}
                    input_names = list(filtered_inputs.keys())

                    if model.__class__.__name__ in set(MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES.values()) and (
                        not hasattr(model.config, "problem_type") or model.config.problem_type is None
                    ):
                        model.config.problem_type = "single_label_classification"

                    traced_model = symbolic_trace(model, input_names)
                    traced_output = traced_model(**filtered_inputs)
                    model_output = model(**filtered_inputs)

            except Exception as e:
                self.fail(f"Couldn't trace module: {e}")

            def flatten_output(output):
                flatten = []
                for x in output:
                    if isinstance(x, (tuple, list)):
                        flatten += flatten_output(x)
                    elif not isinstance(x, torch.Tensor):
                        continue
                    else:
                        flatten.append(x)
                return flatten

            model_output = flatten_output(model_output)
            traced_output = flatten_output(traced_output)
            num_outputs = len(model_output)

            for i in range(num_outputs):
                self.assertTrue(
                    torch.allclose(model_output[i], traced_output[i]),
                    f"traced {i}th output doesn't match model {i}th output for {model_class}",
                )

            # Test that the model can be serialized and restored properly
            with tempfile.TemporaryDirectory() as tmp_dir_name:
                pkl_file_name = os.path.join(tmp_dir_name, "model.pkl")
                try:
                    with open(pkl_file_name, "wb") as f:
                        pickle.dump(traced_model, f)
                    with open(pkl_file_name, "rb") as f:
                        loaded = pickle.load(f)
                except Exception as e:
                    self.fail(f"Couldn't serialize / deserialize the traced model: {e}")

                loaded_output = loaded(**filtered_inputs)
                loaded_output = flatten_output(loaded_output)

                for i in range(num_outputs):
                    self.assertTrue(
                        torch.allclose(model_output[i], loaded_output[i]),
                        f"serialized model {i}th output doesn't match model {i}th output for {model_class}",
                    )

            # Avoid memory leak. Without this, each call increase RAM usage by ~20MB.
            # (Even with this call, there are still memory leak by ~0.04MB)
            self.clear_torch_jit_class_registry()

    def test_headmasking(self):
        if not self.test_head_masking:
            return

        global_rng.seed(42)
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
        global_rng.seed()

        inputs_dict["output_attentions"] = True
        config.output_hidden_states = True
        configs_no_init = _config_zero_init(config)  # To be sure we have no Nan
        for model_class in self.all_model_classes:
            model = model_class(config=configs_no_init)
            model.to(torch_device)
            model.eval()

            # Prepare head_mask
            # Set require_grad after having prepared the tensor to avoid error (leaf variable has been moved into the graph interior)
            head_mask = torch.ones(
                self.model_tester.num_hidden_layers,
                self.model_tester.num_attention_heads,
                device=torch_device,
            )
            head_mask[0, 0] = 0
            head_mask[-1, :-1] = 0
            head_mask.requires_grad_(requires_grad=True)
            inputs = self._prepare_for_class(inputs_dict, model_class).copy()
            inputs["head_mask"] = head_mask
            if model.config.is_encoder_decoder:
                signature = inspect.signature(model.forward)
                arg_names = [*signature.parameters.keys()]
                if "decoder_head_mask" in arg_names:  # necessary diferentiation because of T5 model
                    inputs["decoder_head_mask"] = head_mask
                if "cross_attn_head_mask" in arg_names:
                    inputs["cross_attn_head_mask"] = head_mask
            outputs = model(**inputs, return_dict=True)

            # Test that we can get a gradient back for importance score computation
            output = sum(t.sum() for t in outputs[0])
            output = output.sum()
            output.backward()
            multihead_outputs = head_mask.grad

            self.assertIsNotNone(multihead_outputs)
            self.assertEqual(len(multihead_outputs), self.model_tester.num_hidden_layers)

            def check_attentions_validity(attentions):
                # Remove Nan
                for t in attentions:
                    self.assertLess(
                        torch.sum(torch.isnan(t)), t.numel() / 4
                    )  # Check we don't have more than 25% nans (arbitrary)
                attentions = [
                    t.masked_fill(torch.isnan(t), 0.0) for t in attentions
                ]  # remove them (the test is less complete)

                self.assertAlmostEqual(attentions[0][..., 0, :, :].flatten().sum().item(), 0.0)
                self.assertNotEqual(attentions[0][..., -1, :, :].flatten().sum().item(), 0.0)
                if len(attentions) > 2:  # encoder-decoder models have only 2 layers in each module
                    self.assertNotEqual(attentions[1][..., 0, :, :].flatten().sum().item(), 0.0)
                self.assertAlmostEqual(attentions[-1][..., -2, :, :].flatten().sum().item(), 0.0)
                self.assertNotEqual(attentions[-1][..., -1, :, :].flatten().sum().item(), 0.0)

            if model.config.is_encoder_decoder:
                check_attentions_validity(outputs.encoder_attentions)
                check_attentions_validity(outputs.decoder_attentions)
                check_attentions_validity(outputs.cross_attentions)
            else:
                check_attentions_validity(outputs.attentions)

    def test_head_pruning(self):
        if not self.test_pruning:
            return

        for model_class in self.all_model_classes:
            (
                config,
                inputs_dict,
            ) = self.model_tester.prepare_config_and_inputs_for_common()

            if "head_mask" in inputs_dict:
                del inputs_dict["head_mask"]

            inputs_dict["output_attentions"] = True
            config.output_hidden_states = False
            model = model_class(config=config)
            model.to(torch_device)
            model.eval()
            heads_to_prune = {
                0: list(range(1, self.model_tester.num_attention_heads)),
                -1: [0],
            }
            model.prune_heads(heads_to_prune)
            with torch.no_grad():
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))

            attentions = outputs[-1]

            self.assertEqual(attentions[0].shape[-3], 1)
            self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads)
            self.assertEqual(attentions[-1].shape[-3], self.model_tester.num_attention_heads - 1)

    def test_head_pruning_save_load_from_pretrained(self):
        if not self.test_pruning:
            return

        for model_class in self.all_model_classes:
            (
                config,
                inputs_dict,
            ) = self.model_tester.prepare_config_and_inputs_for_common()

            if "head_mask" in inputs_dict:
                del inputs_dict["head_mask"]

            inputs_dict["output_attentions"] = True
            config.output_hidden_states = False
            model = model_class(config=config)
            model.to(torch_device)
            model.eval()
            heads_to_prune = {
                0: list(range(1, self.model_tester.num_attention_heads)),
                -1: [0],
            }
            model.prune_heads(heads_to_prune)

            with tempfile.TemporaryDirectory() as temp_dir_name:
                model.save_pretrained(temp_dir_name)
                model = model_class.from_pretrained(temp_dir_name)
                model.to(torch_device)

            with torch.no_grad():
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
            attentions = outputs[-1]
            self.assertEqual(attentions[0].shape[-3], 1)
            self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads)
            self.assertEqual(attentions[-1].shape[-3], self.model_tester.num_attention_heads - 1)

    def test_head_pruning_save_load_from_config_init(self):
        if not self.test_pruning:
            return

        for model_class in self.all_model_classes:
            (
                config,
                inputs_dict,
            ) = self.model_tester.prepare_config_and_inputs_for_common()

            if "head_mask" in inputs_dict:
                del inputs_dict["head_mask"]

            inputs_dict["output_attentions"] = True
            config.output_hidden_states = False

            heads_to_prune = {
                0: list(range(1, self.model_tester.num_attention_heads)),
                -1: [0],
            }
            config.pruned_heads = heads_to_prune

            model = model_class(config=config)
            model.to(torch_device)
            model.eval()

            with torch.no_grad():
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
            attentions = outputs[-1]

            self.assertEqual(attentions[0].shape[-3], 1)
            self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads)
            self.assertEqual(attentions[-1].shape[-3], self.model_tester.num_attention_heads - 1)

    def test_head_pruning_integration(self):
        if not self.test_pruning:
            return

        for model_class in self.all_model_classes:
            (
                config,
                inputs_dict,
            ) = self.model_tester.prepare_config_and_inputs_for_common()

            if "head_mask" in inputs_dict:
                del inputs_dict["head_mask"]

            inputs_dict["output_attentions"] = True
            config.output_hidden_states = False

            heads_to_prune = {0: [0], 1: [1, 2]}
            config.pruned_heads = heads_to_prune

            model = model_class(config=config)
            model.to(torch_device)
            model.eval()

            with torch.no_grad():
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
            attentions = outputs[-1]

            self.assertEqual(attentions[0].shape[-3], self.model_tester.num_attention_heads - 1)
            self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads - 2)
            self.assertEqual(attentions[2].shape[-3], self.model_tester.num_attention_heads)
            self.assertEqual(attentions[3].shape[-3], self.model_tester.num_attention_heads)

            with tempfile.TemporaryDirectory() as temp_dir_name:
                model.save_pretrained(temp_dir_name)
                model = model_class.from_pretrained(temp_dir_name)
                model.to(torch_device)

            with torch.no_grad():
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
            attentions = outputs[-1]

            self.assertEqual(attentions[0].shape[-3], self.model_tester.num_attention_heads - 1)
            self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads - 2)
            self.assertEqual(attentions[2].shape[-3], self.model_tester.num_attention_heads)
            self.assertEqual(attentions[3].shape[-3], self.model_tester.num_attention_heads)

            heads_to_prune = {0: [0], 2: [1, 2]}
            model.prune_heads(heads_to_prune)

            with torch.no_grad():
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
            attentions = outputs[-1]

            self.assertEqual(attentions[0].shape[-3], self.model_tester.num_attention_heads - 1)
            self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads - 2)
            self.assertEqual(attentions[2].shape[-3], self.model_tester.num_attention_heads - 2)
            self.assertEqual(attentions[3].shape[-3], self.model_tester.num_attention_heads)

            self.assertDictEqual(model.config.pruned_heads, {0: [0], 1: [1, 2], 2: [1, 2]})

    def test_hidden_states_output(self):
        def check_hidden_states_output(inputs_dict, config, model_class):
            model = model_class(config)
            model.to(torch_device)
            model.eval()

            with torch.no_grad():
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))

            hidden_states = outputs.encoder_hidden_states if config.is_encoder_decoder else outputs.hidden_states

            expected_num_layers = getattr(
                self.model_tester, "expected_num_hidden_layers", self.model_tester.num_hidden_layers + 1
            )
            self.assertEqual(len(hidden_states), expected_num_layers)

            if hasattr(self.model_tester, "encoder_seq_length"):
                seq_length = self.model_tester.encoder_seq_length
                if hasattr(self.model_tester, "chunk_length") and self.model_tester.chunk_length > 1:
                    seq_length = seq_length * self.model_tester.chunk_length
            else:
                seq_length = self.model_tester.seq_length

            self.assertListEqual(
                list(hidden_states[0].shape[-2:]),
                [seq_length, self.model_tester.hidden_size],
            )

            if config.is_encoder_decoder:
                hidden_states = outputs.decoder_hidden_states

                self.assertIsInstance(hidden_states, (list, tuple))
                self.assertEqual(len(hidden_states), expected_num_layers)
                seq_len = getattr(self.model_tester, "seq_length", None)
                decoder_seq_length = getattr(self.model_tester, "decoder_seq_length", seq_len)

                self.assertListEqual(
                    list(hidden_states[0].shape[-2:]),
                    [decoder_seq_length, self.model_tester.hidden_size],
                )

        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            inputs_dict["output_hidden_states"] = True
            check_hidden_states_output(inputs_dict, config, model_class)

            # check that output_hidden_states also work using config
            del inputs_dict["output_hidden_states"]
            config.output_hidden_states = True

            check_hidden_states_output(inputs_dict, config, model_class)

    def test_retain_grad_hidden_states_attentions(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
        config.output_hidden_states = True
        config.output_attentions = self.has_attentions

        # no need to test all models as different heads yield the same functionality
        model_class = self.all_model_classes[0]
        model = model_class(config)
        model.to(torch_device)

        inputs = self._prepare_for_class(inputs_dict, model_class)

        outputs = model(**inputs)

        output = outputs[0]

        if config.is_encoder_decoder:
            # Seq2Seq models
            encoder_hidden_states = outputs.encoder_hidden_states[0]
            encoder_hidden_states.retain_grad()

            decoder_hidden_states = outputs.decoder_hidden_states[0]
            decoder_hidden_states.retain_grad()

            if self.has_attentions:
                encoder_attentions = outputs.encoder_attentions[0]
                encoder_attentions.retain_grad()

                decoder_attentions = outputs.decoder_attentions[0]
                decoder_attentions.retain_grad()

                cross_attentions = outputs.cross_attentions[0]
                cross_attentions.retain_grad()

            output.flatten()[0].backward(retain_graph=True)

            self.assertIsNotNone(encoder_hidden_states.grad)
            self.assertIsNotNone(decoder_hidden_states.grad)

            if self.has_attentions:
                self.assertIsNotNone(encoder_attentions.grad)
                self.assertIsNotNone(decoder_attentions.grad)
                self.assertIsNotNone(cross_attentions.grad)
        else:
            # Encoder-/Decoder-only models
            hidden_states = outputs.hidden_states[0]
            hidden_states.retain_grad()

            if self.has_attentions:
                attentions = outputs.attentions[0]
                attentions.retain_grad()

            output.flatten()[0].backward(retain_graph=True)

            self.assertIsNotNone(hidden_states.grad)

            if self.has_attentions:
                self.assertIsNotNone(attentions.grad)

    def test_feed_forward_chunking(self):
        (
            original_config,
            inputs_dict,
        ) = self.model_tester.prepare_config_and_inputs_for_common()
        for model_class in self.all_model_classes:
            torch.manual_seed(0)
            config = copy.deepcopy(original_config)
            model = model_class(config)
            model.to(torch_device)
            model.eval()

            hidden_states_no_chunk = model(**self._prepare_for_class(inputs_dict, model_class))[0]

            torch.manual_seed(0)
            config.chunk_size_feed_forward = 1
            model = model_class(config)
            model.to(torch_device)
            model.eval()

            hidden_states_with_chunk = model(**self._prepare_for_class(inputs_dict, model_class))[0]
            self.assertTrue(torch.allclose(hidden_states_no_chunk, hidden_states_with_chunk, atol=1e-3))

    def test_resize_position_vector_embeddings(self):
        if not self.test_resize_position_embeddings:
            return

        (
            original_config,
            inputs_dict,
        ) = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            config = copy.deepcopy(original_config)
            model = model_class(config)
            model.to(torch_device)

            if self.model_tester.is_training is False:
                model.eval()

            max_position_embeddings = config.max_position_embeddings

            # Retrieve the embeddings and clone theme
            if model.config.is_encoder_decoder:
                encoder_model_embed, decoder_model_embed = model.get_position_embeddings()
                encoder_cloned_embeddings = encoder_model_embed.weight.clone()
                decoder_cloned_embeddings = decoder_model_embed.weight.clone()
            else:
                model_embed = model.get_position_embeddings()
                cloned_embeddings = model_embed.weight.clone()

            # Check that resizing the position embeddings with a larger max_position_embeddings increases
            # the model's postion embeddings size
            model.resize_position_embeddings(max_position_embeddings + 10)
            self.assertEqual(model.config.max_position_embeddings, max_position_embeddings + 10)

            # Check that it actually resizes the embeddings matrix
            if model.config.is_encoder_decoder:
                encoder_model_embed, decoder_model_embed = model.get_position_embeddings()
                self.assertEqual(encoder_model_embed.weight.shape[0], encoder_cloned_embeddings.shape[0] + 10)
                self.assertEqual(decoder_model_embed.weight.shape[0], decoder_cloned_embeddings.shape[0] + 10)
            else:
                model_embed = model.get_position_embeddings()
                self.assertEqual(model_embed.weight.shape[0], cloned_embeddings.shape[0] + 10)

            # Check that the model can still do a forward pass successfully (every parameter should be resized)
            model(**self._prepare_for_class(inputs_dict, model_class))

            # Check that resizing the position embeddings with a smaller max_position_embeddings decreases
            # the model's max_position_embeddings
            model.resize_position_embeddings(max_position_embeddings - 5)
            self.assertEqual(model.config.max_position_embeddings, max_position_embeddings - 5)

            # Check that it actually resizes the embeddings matrix
            if model.config.is_encoder_decoder:
                encoder_model_embed, decoder_model_embed = model.get_position_embeddings()
                self.assertEqual(encoder_model_embed.weight.shape[0], encoder_cloned_embeddings.shape[0] - 5)
                self.assertEqual(decoder_model_embed.weight.shape[0], decoder_cloned_embeddings.shape[0] - 5)
            else:
                model_embed = model.get_position_embeddings()
                self.assertEqual(model_embed.weight.shape[0], cloned_embeddings.shape[0] - 5)

            # Check that the model can still do a forward pass successfully (every parameter should be resized)
            model(**self._prepare_for_class(inputs_dict, model_class))

            # Check that adding and removing tokens has not modified the first part of the embedding matrix.
            models_equal = True

            if model.config.is_encoder_decoder:
                for p1, p2 in zip(encoder_cloned_embeddings, encoder_model_embed.weight):
                    if p1.data.ne(p2.data).sum() > 0:
                        models_equal = False
                for p1, p2 in zip(decoder_cloned_embeddings, decoder_model_embed.weight):
                    if p1.data.ne(p2.data).sum() > 0:
                        models_equal = False
            else:
                for p1, p2 in zip(cloned_embeddings, model_embed.weight):
                    if p1.data.ne(p2.data).sum() > 0:
                        models_equal = False

            self.assertTrue(models_equal)

    def test_resize_tokens_embeddings(self):
        (
            original_config,
            inputs_dict,
        ) = self.model_tester.prepare_config_and_inputs_for_common()
        if not self.test_resize_embeddings:
            return

        for model_class in self.all_model_classes:
            config = copy.deepcopy(original_config)
            model = model_class(config)
            model.to(torch_device)

            if self.model_tester.is_training is False:
                model.eval()

            model_vocab_size = config.vocab_size
            # Retrieve the embeddings and clone theme
            model_embed = model.resize_token_embeddings(model_vocab_size)
            cloned_embeddings = model_embed.weight.clone()

            # Check that resizing the token embeddings with a larger vocab size increases the model's vocab size
            model_embed = model.resize_token_embeddings(model_vocab_size + 10)
            self.assertEqual(model.config.vocab_size, model_vocab_size + 10)
            # Check that it actually resizes the embeddings matrix
            self.assertEqual(model_embed.weight.shape[0], cloned_embeddings.shape[0] + 10)
            # Check that the model can still do a forward pass successfully (every parameter should be resized)
            model(**self._prepare_for_class(inputs_dict, model_class))

            # Check that resizing the token embeddings with a smaller vocab size decreases the model's vocab size
            model_embed = model.resize_token_embeddings(model_vocab_size - 15)
            self.assertEqual(model.config.vocab_size, model_vocab_size - 15)
            # Check that it actually resizes the embeddings matrix
            self.assertEqual(model_embed.weight.shape[0], cloned_embeddings.shape[0] - 15)

            # Check that the model can still do a forward pass successfully (every parameter should be resized)
            # Input ids should be clamped to the maximum size of the vocabulary
            inputs_dict["input_ids"].clamp_(max=model_vocab_size - 15 - 1)

            # make sure that decoder_input_ids are resized as well
            if "decoder_input_ids" in inputs_dict:
                inputs_dict["decoder_input_ids"].clamp_(max=model_vocab_size - 15 - 1)
            model(**self._prepare_for_class(inputs_dict, model_class))

            # Check that adding and removing tokens has not modified the first part of the embedding matrix.
            models_equal = True
            for p1, p2 in zip(cloned_embeddings, model_embed.weight):
                if p1.data.ne(p2.data).sum() > 0:
                    models_equal = False

            self.assertTrue(models_equal)

    def test_resize_embeddings_untied(self):
        (
            original_config,
            inputs_dict,
        ) = self.model_tester.prepare_config_and_inputs_for_common()
        if not self.test_resize_embeddings:
            return

        original_config.tie_word_embeddings = False

        # if model cannot untied embeddings -> leave test
        if original_config.tie_word_embeddings:
            return

        for model_class in self.all_model_classes:
            config = copy.deepcopy(original_config)
            model = model_class(config).to(torch_device)

            # if no output embeddings -> leave test
            if model.get_output_embeddings() is None:
                continue

            # Check that resizing the token embeddings with a larger vocab size increases the model's vocab size
            model_vocab_size = config.vocab_size
            model.resize_token_embeddings(model_vocab_size + 10)
            self.assertEqual(model.config.vocab_size, model_vocab_size + 10)
            output_embeds = model.get_output_embeddings()
            self.assertEqual(output_embeds.weight.shape[0], model_vocab_size + 10)
            # Check bias if present
            if output_embeds.bias is not None:
                self.assertEqual(output_embeds.bias.shape[0], model_vocab_size + 10)
            # Check that the model can still do a forward pass successfully (every parameter should be resized)
            model(**self._prepare_for_class(inputs_dict, model_class))

            # Check that resizing the token embeddings with a smaller vocab size decreases the model's vocab size
            model.resize_token_embeddings(model_vocab_size - 15)
            self.assertEqual(model.config.vocab_size, model_vocab_size - 15)
            # Check that it actually resizes the embeddings matrix
            output_embeds = model.get_output_embeddings()
            self.assertEqual(output_embeds.weight.shape[0], model_vocab_size - 15)
            # Check bias if present
            if output_embeds.bias is not None:
                self.assertEqual(output_embeds.bias.shape[0], model_vocab_size - 15)
            # Check that the model can still do a forward pass successfully (every parameter should be resized)
            # Input ids should be clamped to the maximum size of the vocabulary
            inputs_dict["input_ids"].clamp_(max=model_vocab_size - 15 - 1)
            if "decoder_input_ids" in inputs_dict:
                inputs_dict["decoder_input_ids"].clamp_(max=model_vocab_size - 15 - 1)
            # Check that the model can still do a forward pass successfully (every parameter should be resized)
            model(**self._prepare_for_class(inputs_dict, model_class))

    def test_model_common_attributes(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            model = model_class(config)
            self.assertIsInstance(model.get_input_embeddings(), (nn.Embedding, AdaptiveEmbedding))
            model.set_input_embeddings(nn.Embedding(10, 10))
            x = model.get_output_embeddings()
            self.assertTrue(x is None or isinstance(x, nn.Linear))

    def test_model_main_input_name(self):
        for model_class in self.all_model_classes:
            model_signature = inspect.signature(getattr(model_class, "forward"))
            # The main input is the name of the argument after `self`
            observed_main_input_name = list(model_signature.parameters.keys())[1]
            self.assertEqual(model_class.main_input_name, observed_main_input_name)

    def test_correct_missing_keys(self):
        if not self.test_missing_keys:
            return
        config, _ = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            model = model_class(config)
            base_model_prefix = model.base_model_prefix

            if hasattr(model, base_model_prefix):
                extra_params = {k: v for k, v in model.named_parameters() if not k.startswith(base_model_prefix)}
                extra_params.update({k: v for k, v in model.named_buffers() if not k.startswith(base_model_prefix)})
                # Some models define this as None
                if model._keys_to_ignore_on_load_missing:
                    for key in model._keys_to_ignore_on_load_missing:
                        extra_params.pop(key, None)

                if not extra_params:
                    # In that case, we *are* on a head model, but every
                    # single key is not actual parameters and this is
                    # tested in `test_tied_model_weights_key_ignore` test.
                    continue

                with tempfile.TemporaryDirectory() as temp_dir_name:
                    model.base_model.save_pretrained(temp_dir_name)
                    model, loading_info = model_class.from_pretrained(temp_dir_name, output_loading_info=True)
                    self.assertGreater(len(loading_info["missing_keys"]), 0, model.__class__.__name__)

    def test_tie_model_weights(self):
        if not self.test_torchscript:
            return

        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        def check_same_values(layer_1, layer_2):
            equal = True
            for p1, p2 in zip(layer_1.weight, layer_2.weight):
                if p1.data.ne(p2.data).sum() > 0:
                    equal = False
            return equal

        for model_class in self.all_model_classes:
            config.torchscript = True
            model_not_tied = model_class(config)
            if model_not_tied.get_output_embeddings() is None:
                continue

            config_tied = copy.deepcopy(config)
            config_tied.torchscript = False
            model_tied = model_class(config_tied)
            params_tied = list(model_tied.parameters())
            # Check that the embedding layer and decoding layer are the same in size and in value
            # self.assertTrue(check_same_values(embeddings, decoding))

            # # Check that after modification, they remain the same.
            # embeddings.weight.data.div_(2)
            # # Check that the embedding layer and decoding layer are the same in size and in value
            # self.assertTrue(embeddings.weight.shape, decoding.weight.shape)
            # self.assertTrue(check_same_values(embeddings, decoding))

            # # Check that after modification, they remain the same.
            # decoding.weight.data.div_(4)
            # # Check that the embedding layer and decoding layer are the same in size and in value
            # self.assertTrue(embeddings.weight.shape, decoding.weight.shape)
            # self.assertTrue(check_same_values(embeddings, decoding))

            # Check that after resize they remain tied.
            model_tied.resize_token_embeddings(config.vocab_size + 10)
            params_tied_2 = list(model_tied.parameters())
            self.assertEqual(len(params_tied_2), len(params_tied))

            # decoding.weight.data.mul_(20)
            # # Check that the embedding layer and decoding layer are the same in size and in value
            # self.assertTrue(model.transformer.wte.weight.shape, model.lm_head.weight.shape)
            # self.assertTrue(check_same_values(model.transformer.wte, model.lm_head))

    @require_safetensors
    def test_can_use_safetensors(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
        for model_class in self.all_model_classes:
            model_tied = model_class(config)
            with tempfile.TemporaryDirectory() as d:
                try:
                    model_tied.save_pretrained(d, safe_serialization=True)
                except Exception as e:
                    raise Exception(f"Class {model_class.__name__} cannot be saved using safetensors: {e}")

                model_reloaded, infos = model_class.from_pretrained(d, output_loading_info=True)
                # Checking the state dicts are correct
                reloaded_state = model_reloaded.state_dict()
                for k, v in model_tied.state_dict().items():
                    self.assertIn(k, reloaded_state, f"Key {k} is missing from reloaded")
                    torch.testing.assert_close(
                        v, reloaded_state[k], msg=lambda x: f"{model_class.__name__}: Tensor {k}: {x}"
                    )

                # Checking the tensor sharing are correct
                ptrs = defaultdict(list)
                for k, v in model_tied.state_dict().items():
                    ptrs[v.data_ptr()].append(k)

                shared_ptrs = {k: v for k, v in ptrs.items() if len(v) > 1}

                for _, shared_names in shared_ptrs.items():
                    reloaded_ptrs = {reloaded_state[k].data_ptr() for k in shared_names}
                    self.assertEqual(
                        len(reloaded_ptrs),
                        1,
                        f"The shared pointers are incorrect, found different pointers for keys {shared_names}",
                    )

    def test_tied_model_weights_key_ignore(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
        for model_class in self.all_model_classes:
            model_tied = model_class(config)
            with tempfile.TemporaryDirectory() as d:
                model_tied.save_pretrained(d)

                # We are nuking ALL weights on file, so every parameter should
                # yell on load. We're going to detect if we yell too much, or too little.
                with open(os.path.join(d, "pytorch_model.bin"), "wb") as f:
                    torch.save({}, f)
                model_reloaded, infos = model_class.from_pretrained(d, output_loading_info=True)

                # ! Actually we could use `state_dict()` and check iteratively the tensors which are the same (for instance using `tensor.data_ptr()`). to detect the duplicates.
                # ```python
                # model = GPT2LMHeadModel.from_pretrained("gpt2")
                # "lm_head.weight" in model.state_dict().keys()  # True
                # "lm_head.weight" in model.named_parameters() # False
                # In [6]: model.lm_head.weight.data_ptr()
                # Out[6]: 139901378371648
                # In [9]: model.transformer.wte.weight.data_ptr()
                # Out[9]: 139901378371648  # Same PTR, it's the same DATA ! we would need to check for stride too to be 100% accurate.
                # ```

                prefix = f"{model_reloaded.base_model_prefix}."
                params = dict(model_reloaded.named_parameters())
                params.update(dict(model_reloaded.named_buffers()))
                # param_names = set(k[len(prefix) :] if k.startswith(prefix) else k for k in params.keys())
                param_names = {k[len(prefix) :] if k.startswith(prefix) else k for k in params.keys()}

                missing_keys = set(infos["missing_keys"])

                extra_missing = missing_keys - param_names
                # missed_missing = param_names - missing_keys

                self.assertEqual(
                    extra_missing,
                    set(),
                    f"This model {model_class.__name__} might be missing some `keys_to_ignore`: {extra_missing}",
                )

                # self.assertEqual(
                #     missed_missing,
                #     set(),
                #     f"This model {model_class.__name__} ignores keys {missed_missing} but they look like real"
                #     " parameters",
                # )

    def test_model_outputs_equivalence(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        def set_nan_tensor_to_zero(t):
            t[t != t] = 0
            return t

        def check_equivalence(model, tuple_inputs, dict_inputs, additional_kwargs={}):
            with torch.no_grad():
                tuple_output = model(**tuple_inputs, return_dict=False, **additional_kwargs)
                dict_output = model(**dict_inputs, return_dict=True, **additional_kwargs).to_tuple()

                def recursive_check(tuple_object, dict_object):
                    if isinstance(tuple_object, (List, Tuple)):
                        for tuple_iterable_value, dict_iterable_value in zip(tuple_object, dict_object):
                            recursive_check(tuple_iterable_value, dict_iterable_value)
                    elif isinstance(tuple_object, Dict):
                        for tuple_iterable_value, dict_iterable_value in zip(
                            tuple_object.values(), dict_object.values()
                        ):
                            recursive_check(tuple_iterable_value, dict_iterable_value)
                    elif tuple_object is None:
                        return
                    else:
                        self.assertTrue(
                            torch.allclose(
                                set_nan_tensor_to_zero(tuple_object), set_nan_tensor_to_zero(dict_object), atol=1e-5
                            ),
                            msg=(
                                "Tuple and dict output are not equal. Difference:"
                                f" {torch.max(torch.abs(tuple_object - dict_object))}. Tuple has `nan`:"
                                f" {torch.isnan(tuple_object).any()} and `inf`: {torch.isinf(tuple_object)}. Dict has"
                                f" `nan`: {torch.isnan(dict_object).any()} and `inf`: {torch.isinf(dict_object)}."
                            ),
                        )

                recursive_check(tuple_output, dict_output)

        for model_class in self.all_model_classes:
            model = model_class(config)
            model.to(torch_device)
            model.eval()

            tuple_inputs = self._prepare_for_class(inputs_dict, model_class)
            dict_inputs = self._prepare_for_class(inputs_dict, model_class)
            check_equivalence(model, tuple_inputs, dict_inputs)

            tuple_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
            dict_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
            check_equivalence(model, tuple_inputs, dict_inputs)

            tuple_inputs = self._prepare_for_class(inputs_dict, model_class)
            dict_inputs = self._prepare_for_class(inputs_dict, model_class)
            check_equivalence(model, tuple_inputs, dict_inputs, {"output_hidden_states": True})

            tuple_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
            dict_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
            check_equivalence(model, tuple_inputs, dict_inputs, {"output_hidden_states": True})

            if self.has_attentions:
                tuple_inputs = self._prepare_for_class(inputs_dict, model_class)
                dict_inputs = self._prepare_for_class(inputs_dict, model_class)
                check_equivalence(model, tuple_inputs, dict_inputs, {"output_attentions": True})

                tuple_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
                dict_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
                check_equivalence(model, tuple_inputs, dict_inputs, {"output_attentions": True})

                tuple_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
                dict_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
                check_equivalence(
                    model, tuple_inputs, dict_inputs, {"output_hidden_states": True, "output_attentions": True}
                )

    # Don't copy this method to model specific test file!
    # TODO: remove this method once the issues are all fixed!
    def _make_attention_mask_non_null(self, inputs_dict):
        """Make sure no sequence has all zeros as attention mask"""

        for k in ["attention_mask", "encoder_attention_mask", "decoder_attention_mask"]:
            if k in inputs_dict:
                attention_mask = inputs_dict[k]

                # Make sure no all 0s attention masks - to avoid failure at this moment.
                # Put `1` at the beginning of sequences to make it still work when combining causal attention masks.
                # TODO: remove this line once a fix regarding large negative values for attention mask is done.
                attention_mask = torch.cat(
                    [torch.ones_like(attention_mask[:, :1], dtype=attention_mask.dtype), attention_mask[:, 1:]], dim=-1
                )

                # Here we make the first sequence with all 0s as attention mask.
                # Currently, this will fail for `TFWav2Vec2Model`. This is caused by the different large negative
                # values, like `1e-4`, `1e-9`, `1e-30` and `-inf` for attention mask across models/frameworks.
                # TODO: enable this block once the large negative values thing is cleaned up.
                # (see https://github.com/huggingface/transformers/issues/14859)
                # attention_mask = torch.cat(
                #     [torch.zeros_like(attention_mask[:1], dtype=attention_mask.dtype), attention_mask[1:]],
                #     dim=0
                # )

                inputs_dict[k] = attention_mask

    # Don't copy this method to model specific test file!
    # TODO: remove this method once the issues are all fixed!
    def _postprocessing_to_ignore_test_cases(self, tf_outputs, pt_outputs, model_class):
        """For temporarily ignoring some failed test cases (issues to be fixed)"""

        tf_keys = {k for k, v in tf_outputs.items() if v is not None}
        pt_keys = {k for k, v in pt_outputs.items() if v is not None}

        key_differences = tf_keys.symmetric_difference(pt_keys)

        if model_class.__name__ in [
            "FlaubertWithLMHeadModel",
            "FunnelForPreTraining",
            "ElectraForPreTraining",
            "XLMWithLMHeadModel",
            "TransfoXLLMHeadModel",
        ]:
            for k in key_differences:
                if k in ["loss", "losses"]:
                    tf_keys.discard(k)
                    pt_keys.discard(k)
        elif model_class.__name__.startswith("GPT2"):
            # `TFGPT2` has `past_key_values` as a tensor while `GPT2` has it as a tuple.
            tf_keys.discard("past_key_values")
            pt_keys.discard("past_key_values")

        # create new outputs from the remaining fields
        new_tf_outputs = type(tf_outputs)(**{k: tf_outputs[k] for k in tf_keys})
        new_pt_outputs = type(pt_outputs)(**{k: pt_outputs[k] for k in pt_keys})

        return new_tf_outputs, new_pt_outputs

    # Copied from tests.test_modeling_tf_common.TFModelTesterMixin.check_pt_tf_outputs
    def check_pt_tf_outputs(self, tf_outputs, pt_outputs, model_class, tol=1e-5, name="outputs", attributes=None):
        """Check the outputs from PyTorch and TensorFlow models are close enough. Checks are done in a recursive way.

        Args:
            model_class: The class of the model that is currently testing. For example, `TFBertModel`,
                TFBertForMaskedLM`, `TFBertForSequenceClassification`, etc. Mainly used for providing more informative
                error messages.
            name (`str`): The name of the output. For example, `output.hidden_states`, `output.attentions`, etc.
            attributes (`Tuple[str]`): The names of the output's element if the output is a tuple/list with each element
                being a named field in the output.
        """

        self.assertEqual(type(name), str)
        if attributes is not None:
            self.assertEqual(type(attributes), tuple, f"{name}: The argument `attributes` should be a `tuple`")

        # Allow `ModelOutput` (e.g. `CLIPOutput` has `text_model_output` and `vision_model_output`).
        if isinstance(tf_outputs, ModelOutput):
            self.assertTrue(
                isinstance(pt_outputs, ModelOutput),
                f"{name}: `pt_outputs` should an instance of `ModelOutput` when `tf_outputs` is",
            )

            # Don't copy this block to model specific test file!
            # TODO: remove this method and this line after issues are fixed
            tf_outputs, pt_outputs = self._postprocessing_to_ignore_test_cases(tf_outputs, pt_outputs, model_class)

            tf_keys = [k for k, v in tf_outputs.items() if v is not None]
            pt_keys = [k for k, v in pt_outputs.items() if v is not None]

            self.assertEqual(tf_keys, pt_keys, f"{name}: Output keys differ between TF and PyTorch")

            # convert to the case of `tuple`
            # appending each key to the current (string) `name`
            attributes = tuple([f"{name}.{k}" for k in tf_keys])
            self.check_pt_tf_outputs(
                tf_outputs.to_tuple(), pt_outputs.to_tuple(), model_class, tol=tol, name=name, attributes=attributes
            )

        # Allow `list` (e.g. `TransfoXLModelOutput.mems` is a list of tensors.)
        elif type(tf_outputs) in [tuple, list]:
            self.assertEqual(type(tf_outputs), type(pt_outputs), f"{name}: Output types differ between TF and PyTorch")
            self.assertEqual(len(tf_outputs), len(pt_outputs), f"{name}: Output lengths differ between TF and PyTorch")

            if attributes is not None:
                # case 1: each output has assigned name (e.g. a tuple form of a `ModelOutput`)
                self.assertEqual(
                    len(attributes),
                    len(tf_outputs),
                    f"{name}: The tuple `attributes` should have the same length as `tf_outputs`",
                )
            else:
                # case 2: each output has no assigned name (e.g. hidden states of each layer) -> add an index to `name`
                attributes = tuple([f"{name}_{idx}" for idx in range(len(tf_outputs))])

            for tf_output, pt_output, attr in zip(tf_outputs, pt_outputs, attributes):
                self.check_pt_tf_outputs(tf_output, pt_output, model_class, tol=tol, name=attr)

        elif isinstance(tf_outputs, tf.Tensor):
            self.assertTrue(
                isinstance(pt_outputs, torch.Tensor), f"{name}: `pt_outputs` should a tensor when `tf_outputs` is"
            )

            tf_outputs = tf_outputs.numpy()
            pt_outputs = pt_outputs.detach().to("cpu").numpy()

            self.assertEqual(
                tf_outputs.shape, pt_outputs.shape, f"{name}: Output shapes differ between TF and PyTorch"
            )

            # deal with NumPy's scalars to make replacing nan values by 0 work.
            if np.isscalar(tf_outputs):
                tf_outputs = np.array([tf_outputs])
                pt_outputs = np.array([pt_outputs])

            tf_nans = np.isnan(tf_outputs)
            pt_nans = np.isnan(pt_outputs)

            pt_outputs[tf_nans] = 0
            tf_outputs[tf_nans] = 0
            pt_outputs[pt_nans] = 0
            tf_outputs[pt_nans] = 0

            max_diff = np.amax(np.abs(tf_outputs - pt_outputs))
            self.assertLessEqual(max_diff, tol, f"{name}: Difference between PyTorch and TF is {max_diff} (>= {tol}).")
        else:
            raise ValueError(
                "`tf_outputs` should be an instance of `ModelOutput`, a `tuple`, or an instance of `tf.Tensor`. Got"
                f" {type(tf_outputs)} instead."
            )

    def prepare_tf_inputs_from_pt_inputs(self, pt_inputs_dict):
        tf_inputs_dict = {}
        for key, tensor in pt_inputs_dict.items():
            # skip key that does not exist in tf
            if type(tensor) == bool:
                tf_inputs_dict[key] = tensor
            elif key == "input_values":
                tf_inputs_dict[key] = tf.convert_to_tensor(tensor.cpu().numpy(), dtype=tf.float32)
            elif key == "pixel_values":
                tf_inputs_dict[key] = tf.convert_to_tensor(tensor.cpu().numpy(), dtype=tf.float32)
            elif key == "input_features":
                tf_inputs_dict[key] = tf.convert_to_tensor(tensor.cpu().numpy(), dtype=tf.float32)
            # other general float inputs
            elif tensor.is_floating_point():
                tf_inputs_dict[key] = tf.convert_to_tensor(tensor.cpu().numpy(), dtype=tf.float32)
            else:
                tf_inputs_dict[key] = tf.convert_to_tensor(tensor.cpu().numpy(), dtype=tf.int32)

        return tf_inputs_dict

    def check_pt_tf_models(self, tf_model, pt_model, pt_inputs_dict):
        tf_inputs_dict = self.prepare_tf_inputs_from_pt_inputs(pt_inputs_dict)

        # send pytorch inputs to the correct device
        pt_inputs_dict = {
            k: v.to(device=torch_device) if isinstance(v, torch.Tensor) else v for k, v in pt_inputs_dict.items()
        }

        # send pytorch model to the correct device
        pt_model.to(torch_device)

        # Check predictions on first output (logits/hidden-states) are close enough given low-level computational differences
        pt_model.eval()

        with torch.no_grad():
            pt_outputs = pt_model(**pt_inputs_dict)
        tf_outputs = tf_model(tf_inputs_dict)

        # tf models returned loss is usually a tensor rather than a scalar.
        # (see `hf_compute_loss`: it uses `tf.keras.losses.Reduction.NONE`)
        # Change it here to a scalar to match PyTorch models' loss
        tf_loss = getattr(tf_outputs, "loss", None)
        if tf_loss is not None:
            tf_outputs.loss = tf.math.reduce_mean(tf_loss)

        self.check_pt_tf_outputs(tf_outputs, pt_outputs, type(pt_model))

    @is_pt_tf_cross_test
    def test_pt_tf_model_equivalence(self, allow_missing_keys=False):
        import transformers

        for model_class in self.all_model_classes:
            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

            tf_model_class_name = "TF" + model_class.__name__  # Add the "TF" at the beginning
            if not hasattr(transformers, tf_model_class_name):
                # transformers does not have this model in TF version yet
                return

            # Output all for aggressive testing
            config.output_hidden_states = True
            config.output_attentions = self.has_attentions

            # Make sure no sequence has all zeros as attention mask, otherwise some tests fail due to the inconsistency
            # of the usage `1e-4`, `1e-9`, `1e-30`, `-inf`.
            # TODO: Use a uniform value for all models, make sure all tests pass without this processing, and remove it.
            self._make_attention_mask_non_null(inputs_dict)

            tf_model_class = getattr(transformers, tf_model_class_name)

            pt_model = model_class(config)
            tf_model = tf_model_class(config)

            pt_inputs_dict = self._prepare_for_class(inputs_dict, model_class)
            pt_inputs_dict_with_labels = self._prepare_for_class(
                inputs_dict,
                model_class,
                # Not all models accept "labels" in the forward pass (yet :) )
                return_labels=True if "labels" in inspect.signature(model_class.forward).parameters.keys() else False,
            )

            # make sure only tf inputs are forward that actually exist in function args
            tf_input_keys = set(inspect.signature(tf_model.call).parameters.keys())

            # remove all head masks
            tf_input_keys.discard("head_mask")
            tf_input_keys.discard("cross_attn_head_mask")
            tf_input_keys.discard("decoder_head_mask")

            pt_inputs_dict = {k: v for k, v in pt_inputs_dict.items() if k in tf_input_keys}
            pt_inputs_dict_with_labels = {k: v for k, v in pt_inputs_dict_with_labels.items() if k in tf_input_keys}

            # For some models (e.g. base models), there is no label returned.
            # Set the input dict to `None` to avoid check outputs twice for the same input dicts.
            if not set(pt_inputs_dict_with_labels.keys()).symmetric_difference(pt_inputs_dict.keys()):
                pt_inputs_dict_with_labels = None

            # Check we can load pt model in tf and vice-versa with model => model functions
            # Here requires `tf_inputs_dict` to build `tf_model`
            tf_inputs_dict = self.prepare_tf_inputs_from_pt_inputs(pt_inputs_dict)
            tf_model = transformers.load_pytorch_model_in_tf2_model(
                tf_model, pt_model, tf_inputs=tf_inputs_dict, allow_missing_keys=allow_missing_keys
            )
            pt_model = transformers.load_tf2_model_in_pytorch_model(
                pt_model, tf_model, allow_missing_keys=allow_missing_keys
            )

            # Original test: check without `labels`
            self.check_pt_tf_models(tf_model, pt_model, pt_inputs_dict)
            # check with `labels`
            if pt_inputs_dict_with_labels:
                self.check_pt_tf_models(tf_model, pt_model, pt_inputs_dict_with_labels)

            # Check we can load pt model in tf and vice-versa with checkpoint => model functions
            with tempfile.TemporaryDirectory() as tmpdirname:
                pt_checkpoint_path = os.path.join(tmpdirname, "pt_model.bin")
                torch.save(pt_model.state_dict(), pt_checkpoint_path)
                tf_model = transformers.load_pytorch_checkpoint_in_tf2_model(
                    tf_model, pt_checkpoint_path, allow_missing_keys=allow_missing_keys
                )

                tf_checkpoint_path = os.path.join(tmpdirname, "tf_model.h5")
                tf_model.save_weights(tf_checkpoint_path)
                pt_model = transformers.load_tf2_checkpoint_in_pytorch_model(
                    pt_model, tf_checkpoint_path, allow_missing_keys=allow_missing_keys
                )

            # Original test: check without `labels`
            self.check_pt_tf_models(tf_model, pt_model, pt_inputs_dict)
            # check with `labels`
            if pt_inputs_dict_with_labels:
                self.check_pt_tf_models(tf_model, pt_model, pt_inputs_dict_with_labels)

    def assert_almost_equals(self, a: np.ndarray, b: np.ndarray, tol: float):
        diff = np.abs((a - b)).max()
        self.assertLessEqual(diff, tol, f"Difference between torch and flax is {diff} (>= {tol}).")

    def check_pt_flax_outputs(self, fx_outputs, pt_outputs, model_class, tol=1e-5, name="outputs", attributes=None):
        """
        Args:
            model_class: The class of the model that is currently testing. For example, ..., etc.
            Currently unused, but it could make debugging easier and faster.

            names: A string, or a list of strings. These specify what fx_outputs/pt_outputs represent in the model outputs.
                Currently unused, but in the future, we could use this information to make the error message clearer
                by giving the name(s) of the output tensor(s) with large difference(s) between PT and Flax.
        """

        self.assertEqual(type(name), str)
        if attributes is not None:
            self.assertEqual(type(attributes), tuple, f"{name}: The argument `attributes` should be a `tuple`")

        # Allow `ModelOutput` (e.g. `CLIPOutput` has `text_model_output` and `vision_model_output`).
        if isinstance(fx_outputs, ModelOutput):
            self.assertTrue(
                isinstance(pt_outputs, ModelOutput),
                f"{name}: `pt_outputs` should an instance of `ModelOutput` when `fx_outputs` is",
            )

            fx_keys = tuple([k for k, v in fx_outputs.items() if v is not None])
            pt_keys = tuple([k for k, v in pt_outputs.items() if v is not None])

            self.assertEqual(fx_keys, pt_keys, f"{name}: Output keys differ between Flax and PyTorch")

            # convert to the case of `tuple`
            # appending each key to the current (string) `name`
            attributes = tuple([f"{name}.{k}" for k in fx_keys])
            self.check_pt_flax_outputs(
                fx_outputs.to_tuple(), pt_outputs.to_tuple(), model_class, tol=tol, name=name, attributes=attributes
            )

        # Allow `list` (e.g. `TransfoXLModelOutput.mems` is a list of tensors.)
        elif type(fx_outputs) in [tuple, list]:
            self.assertEqual(
                type(fx_outputs), type(pt_outputs), f"{name}: Output types differ between Flax and PyTorch"
            )
            self.assertEqual(
                len(fx_outputs), len(pt_outputs), f"{name}: Output lengths differ between Flax and PyTorch"
            )

            if attributes is not None:
                # case 1: each output has assigned name (e.g. a tuple form of a `ModelOutput`)
                self.assertEqual(
                    len(attributes),
                    len(fx_outputs),
                    f"{name}: The tuple `attributes` should have the same length as `fx_outputs`",
                )
            else:
                # case 2: each output has no assigned name (e.g. hidden states of each layer) -> add an index to `name`
                attributes = tuple([f"{name}_{idx}" for idx in range(len(fx_outputs))])

            for fx_output, pt_output, attr in zip(fx_outputs, pt_outputs, attributes):
                self.check_pt_flax_outputs(fx_output, pt_output, model_class, tol=tol, name=attr)

        elif isinstance(fx_outputs, jnp.ndarray):
            self.assertTrue(
                isinstance(pt_outputs, torch.Tensor), f"{name}: `pt_outputs` should a tensor when `fx_outputs` is"
            )

            # Using `np.asarray` gives `ValueError: assignment destination is read-only` at the line `fx_outputs[fx_nans] = 0`.
            fx_outputs = np.array(fx_outputs)
            pt_outputs = pt_outputs.detach().to("cpu").numpy()

            self.assertEqual(
                fx_outputs.shape, pt_outputs.shape, f"{name}: Output shapes differ between Flax and PyTorch"
            )

            # deal with NumPy's scalars to make replacing nan values by 0 work.
            if np.isscalar(fx_outputs):
                fx_outputs = np.array([fx_outputs])
                pt_outputs = np.array([pt_outputs])

            fx_nans = np.isnan(fx_outputs)
            pt_nans = np.isnan(pt_outputs)

            pt_outputs[fx_nans] = 0
            fx_outputs[fx_nans] = 0
            pt_outputs[pt_nans] = 0
            fx_outputs[pt_nans] = 0

            max_diff = np.amax(np.abs(fx_outputs - pt_outputs))
            self.assertLessEqual(
                max_diff, tol, f"{name}: Difference between PyTorch and Flax is {max_diff} (>= {tol})."
            )
        else:
            raise ValueError(
                "`fx_outputs` should be an instance of `ModelOutput`, a `tuple`, or an instance of `jnp.ndarray`. Got"
                f" {type(fx_outputs)} instead."
            )

    @is_pt_flax_cross_test
    def test_equivalence_pt_to_flax(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            with self.subTest(model_class.__name__):
                fx_model_class_name = "Flax" + model_class.__name__

                if not hasattr(transformers, fx_model_class_name):
                    # no flax model exists for this class
                    return

                # Output all for aggressive testing
                config.output_hidden_states = True
                config.output_attentions = self.has_attentions

                fx_model_class = getattr(transformers, fx_model_class_name)

                # load PyTorch class
                pt_model = model_class(config).eval()
                # Flax models don't use the `use_cache` option and cache is not returned as a default.
                # So we disable `use_cache` here for PyTorch model.
                pt_model.config.use_cache = False

                # load Flax class
                fx_model = fx_model_class(config, dtype=jnp.float32)

                # make sure only flax inputs are forward that actually exist in function args
                fx_input_keys = inspect.signature(fx_model.__call__).parameters.keys()

                # prepare inputs
                pt_inputs = self._prepare_for_class(inputs_dict, model_class)

                # remove function args that don't exist in Flax
                pt_inputs = {k: v for k, v in pt_inputs.items() if k in fx_input_keys}

                # send pytorch inputs to the correct device
                pt_inputs = {
                    k: v.to(device=torch_device) if isinstance(v, torch.Tensor) else v for k, v in pt_inputs.items()
                }

                # convert inputs to Flax
                fx_inputs = {k: np.array(v) for k, v in pt_inputs.items() if torch.is_tensor(v)}

                fx_state = convert_pytorch_state_dict_to_flax(pt_model.state_dict(), fx_model)
                fx_model.params = fx_state

                # send pytorch model to the correct device
                pt_model.to(torch_device)

                with torch.no_grad():
                    pt_outputs = pt_model(**pt_inputs)
                fx_outputs = fx_model(**fx_inputs)

                fx_keys = tuple([k for k, v in fx_outputs.items() if v is not None])
                pt_keys = tuple([k for k, v in pt_outputs.items() if v is not None])

                self.assertEqual(fx_keys, pt_keys)
                self.check_pt_flax_outputs(fx_outputs, pt_outputs, model_class)

                with tempfile.TemporaryDirectory() as tmpdirname:
                    pt_model.save_pretrained(tmpdirname)
                    fx_model_loaded = fx_model_class.from_pretrained(tmpdirname, from_pt=True)

                fx_outputs_loaded = fx_model_loaded(**fx_inputs)

                fx_keys = tuple([k for k, v in fx_outputs_loaded.items() if v is not None])
                pt_keys = tuple([k for k, v in pt_outputs.items() if v is not None])

                self.assertEqual(fx_keys, pt_keys)
                self.check_pt_flax_outputs(fx_outputs_loaded, pt_outputs, model_class)

    @is_pt_flax_cross_test
    def test_equivalence_flax_to_pt(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            with self.subTest(model_class.__name__):
                fx_model_class_name = "Flax" + model_class.__name__

                if not hasattr(transformers, fx_model_class_name):
                    # no flax model exists for this class
                    return

                # Output all for aggressive testing
                config.output_hidden_states = True
                config.output_attentions = self.has_attentions

                fx_model_class = getattr(transformers, fx_model_class_name)

                # load PyTorch class
                pt_model = model_class(config).eval()
                # Flax models don't use the `use_cache` option and cache is not returned as a default.
                # So we disable `use_cache` here for PyTorch model.
                pt_model.config.use_cache = False

                # load Flax class
                fx_model = fx_model_class(config, dtype=jnp.float32)

                # make sure only flax inputs are forward that actually exist in function args
                fx_input_keys = inspect.signature(fx_model.__call__).parameters.keys()

                # prepare inputs
                pt_inputs = self._prepare_for_class(inputs_dict, model_class)

                # remove function args that don't exist in Flax
                pt_inputs = {k: v for k, v in pt_inputs.items() if k in fx_input_keys}

                # send pytorch inputs to the correct device
                pt_inputs = {
                    k: v.to(device=torch_device) if isinstance(v, torch.Tensor) else v for k, v in pt_inputs.items()
                }

                # convert inputs to Flax
                fx_inputs = {k: np.array(v) for k, v in pt_inputs.items() if torch.is_tensor(v)}

                pt_model = load_flax_weights_in_pytorch_model(pt_model, fx_model.params)

                # make sure weights are tied in PyTorch
                pt_model.tie_weights()

                # send pytorch model to the correct device
                pt_model.to(torch_device)

                with torch.no_grad():
                    pt_outputs = pt_model(**pt_inputs)
                fx_outputs = fx_model(**fx_inputs)

                fx_keys = tuple([k for k, v in fx_outputs.items() if v is not None])
                pt_keys = tuple([k for k, v in pt_outputs.items() if v is not None])

                self.assertEqual(fx_keys, pt_keys)
                self.check_pt_flax_outputs(fx_outputs, pt_outputs, model_class)

                with tempfile.TemporaryDirectory() as tmpdirname:
                    fx_model.save_pretrained(tmpdirname)
                    pt_model_loaded = model_class.from_pretrained(tmpdirname, from_flax=True)

                # send pytorch model to the correct device
                pt_model_loaded.to(torch_device)
                pt_model_loaded.eval()

                with torch.no_grad():
                    pt_outputs_loaded = pt_model_loaded(**pt_inputs)

                fx_keys = tuple([k for k, v in fx_outputs.items() if v is not None])
                pt_keys = tuple([k for k, v in pt_outputs_loaded.items() if v is not None])

                self.assertEqual(fx_keys, pt_keys)
                self.check_pt_flax_outputs(fx_outputs, pt_outputs_loaded, model_class)

    def test_inputs_embeds(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            model = model_class(config)
            model.to(torch_device)
            model.eval()

            inputs = copy.deepcopy(self._prepare_for_class(inputs_dict, model_class))

            if not self.is_encoder_decoder:
                input_ids = inputs["input_ids"]
                del inputs["input_ids"]
            else:
                encoder_input_ids = inputs["input_ids"]
                decoder_input_ids = inputs.get("decoder_input_ids", encoder_input_ids)
                del inputs["input_ids"]
                inputs.pop("decoder_input_ids", None)

            wte = model.get_input_embeddings()
            if not self.is_encoder_decoder:
                inputs["inputs_embeds"] = wte(input_ids)
            else:
                inputs["inputs_embeds"] = wte(encoder_input_ids)
                inputs["decoder_inputs_embeds"] = wte(decoder_input_ids)

            with torch.no_grad():
                model(**inputs)[0]

    @require_torch_multi_gpu
    def test_multi_gpu_data_parallel_forward(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        # some params shouldn't be scattered by nn.DataParallel
        # so just remove them if they are present.
        blacklist_non_batched_params = ["head_mask", "decoder_head_mask", "cross_attn_head_mask"]
        for k in blacklist_non_batched_params:
            inputs_dict.pop(k, None)

        # move input tensors to cuda:O
        for k, v in inputs_dict.items():
            if torch.is_tensor(v):
                inputs_dict[k] = v.to(0)

        for model_class in self.all_model_classes:
            model = model_class(config=config)
            model.to(0)
            model.eval()

            # Wrap model in nn.DataParallel
            model = nn.DataParallel(model)
            with torch.no_grad():
                _ = model(**self._prepare_for_class(inputs_dict, model_class))

    @require_torch_multi_gpu
    def test_model_parallelization(self):
        if not self.test_model_parallel:
            return

        # a candidate for testing_utils
        def get_current_gpu_memory_use():
            """returns a list of cuda memory allocations per GPU in MBs"""

            per_device_memory = []
            for id in range(torch.cuda.device_count()):
                with torch.cuda.device(id):
                    per_device_memory.append(torch.cuda.memory_allocated() >> 20)

            return per_device_memory

        # Needs a large model to see the difference.
        config = self.model_tester.get_large_model_config()

        for model_class in self.all_parallelizable_model_classes:
            torch.cuda.empty_cache()

            # 1. single gpu memory load + unload + memory measurements
            # Retrieve initial memory usage (can easily be ~0.6-1.5GB if cuda-kernels have been preloaded by previous tests)
            memory_at_start = get_current_gpu_memory_use()

            # Put model on device 0 and take a memory snapshot
            model = model_class(config)
            model.to("cuda:0")
            memory_after_model_load = get_current_gpu_memory_use()

            # The memory use on device 0 should be higher than it was initially.
            self.assertGreater(memory_after_model_load[0], memory_at_start[0])

            del model
            gc.collect()
            torch.cuda.empty_cache()

            # 2. MP test
            # it's essential to re-calibrate the usage before the next stage
            memory_at_start = get_current_gpu_memory_use()

            # Spread model layers over multiple devices
            model = model_class(config)
            model.parallelize()
            memory_after_parallelization = get_current_gpu_memory_use()

            # Assert that the memory use on all devices is higher than it was when loaded only on CPU
            for n in range(len(model.device_map.keys())):
                self.assertGreater(memory_after_parallelization[n], memory_at_start[n])

            # Assert that the memory use of device 0 is lower than it was when the entire model was loaded on it
            self.assertLess(memory_after_parallelization[0], memory_after_model_load[0])

            # Assert that the memory use of device 1 is higher than it was when the entire model was loaded
            # on device 0 and device 1 wasn't used at all
            self.assertGreater(memory_after_parallelization[1], memory_after_model_load[1])

            del model
            gc.collect()
            torch.cuda.empty_cache()

    @require_torch_multi_gpu
    def test_model_parallel_equal_results(self):
        if not self.test_model_parallel:
            return

        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_parallelizable_model_classes:
            inputs_dict = self._prepare_for_class(inputs_dict, model_class)

            def cast_to_device(dictionary, device):
                output = {}
                for k, v in dictionary.items():
                    if isinstance(v, torch.Tensor):
                        output[k] = v.to(device)
                    else:
                        output[k] = v

                return output

            model = model_class(config)
            output = model(**cast_to_device(inputs_dict, "cpu"))

            model.parallelize()

            parallel_output = model(**cast_to_device(inputs_dict, "cuda:0"))

            for value, parallel_value in zip(output, parallel_output):
                if isinstance(value, torch.Tensor):
                    self.assertTrue(torch.allclose(value, parallel_value.to("cpu"), atol=1e-7))
                elif isinstance(value, (Tuple, List)):
                    for value_, parallel_value_ in zip(value, parallel_value):
                        self.assertTrue(torch.allclose(value_, parallel_value_.to("cpu"), atol=1e-7))

    @require_torch_multi_gpu
    def test_model_parallel_beam_search(self):
        if not self.test_model_parallel:
            return

        all_generative_and_parallelizable_model_classes = tuple(
            set(self.all_generative_model_classes).intersection(self.all_parallelizable_model_classes)
        )

        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in all_generative_and_parallelizable_model_classes:
            inputs_dict = self._prepare_for_class(inputs_dict, model_class)
            model = model_class(config)

            def cast_to_device(dictionary, device):
                output = {}
                for k, v in dictionary.items():
                    if isinstance(v, torch.Tensor):
                        output[k] = v.to(device)
                    else:
                        output[k] = v

                return output

            model.parallelize()
            model.generate(**cast_to_device(inputs_dict, "cuda:0"), num_beams=2)

    def check_device_map_is_respected(self, model, device_map):
        for param_name, param in model.named_parameters():
            # Find device in device_map
            while len(param_name) > 0 and param_name not in device_map:
                param_name = ".".join(param_name.split(".")[:-1])
            if param_name not in device_map:
                raise ValueError("device map is incomplete, it does not contain any device for `param_name`.")

            param_device = device_map[param_name]
            if param_device in ["cpu", "disk"]:
                self.assertEqual(param.device, torch.device("meta"))
            else:
                self.assertEqual(param.device, torch.device(param_device))

    @require_accelerate
    @mark.accelerate_tests
    @require_torch_gpu
    def test_disk_offload(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            if model_class._no_split_modules is None:
                continue

            inputs_dict_class = self._prepare_for_class(inputs_dict, model_class)
            model = model_class(config).eval()
            model = model.to(torch_device)
            torch.manual_seed(0)
            base_output = model(**inputs_dict_class)

            model_size = compute_module_sizes(model)[""]
            max_size = int(self.model_split_percents[0] * model_size)
            with tempfile.TemporaryDirectory() as tmp_dir:
                model.cpu().save_pretrained(tmp_dir)

                max_memory = {0: max_size, "cpu": max_size}
                with self.assertRaises(ValueError):
                    # This errors out cause it's missing an offload folder
                    new_model = model_class.from_pretrained(tmp_dir, device_map="auto", max_memory=max_memory)

                new_model = model_class.from_pretrained(
                    tmp_dir, device_map="auto", max_memory=max_memory, offload_folder=tmp_dir
                )

                self.check_device_map_is_respected(new_model, new_model.hf_device_map)
                torch.manual_seed(0)
                new_output = new_model(**inputs_dict_class)

                self.assertTrue(torch.allclose(base_output[0], new_output[0], atol=1e-5))

    @require_accelerate
    @mark.accelerate_tests
    @require_torch_gpu
    def test_cpu_offload(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            if model_class._no_split_modules is None:
                continue

            inputs_dict_class = self._prepare_for_class(inputs_dict, model_class)
            model = model_class(config).eval()
            model = model.to(torch_device)

            torch.manual_seed(0)
            base_output = model(**inputs_dict_class)

            model_size = compute_module_sizes(model)[""]
            # We test several splits of sizes to make sure it works.
            max_gpu_sizes = [int(p * model_size) for p in self.model_split_percents]
            with tempfile.TemporaryDirectory() as tmp_dir:
                model.cpu().save_pretrained(tmp_dir)

                for max_size in max_gpu_sizes:
                    max_memory = {0: max_size, "cpu": model_size * 2}
                    new_model = model_class.from_pretrained(tmp_dir, device_map="auto", max_memory=max_memory)
                    # Making sure part of the model will actually end up offloaded
                    self.assertSetEqual(set(new_model.hf_device_map.values()), {0, "cpu"})

                    self.check_device_map_is_respected(new_model, new_model.hf_device_map)

                    torch.manual_seed(0)
                    new_output = new_model(**inputs_dict_class)

                    self.assertTrue(torch.allclose(base_output[0], new_output[0], atol=1e-5))

    @require_accelerate
    @mark.accelerate_tests
    @require_torch_multi_gpu
    def test_model_parallelism(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            if model_class._no_split_modules is None:
                continue

            inputs_dict_class = self._prepare_for_class(inputs_dict, model_class)
            model = model_class(config).eval()
            model = model.to(torch_device)

            torch.manual_seed(0)
            base_output = model(**inputs_dict_class)

            model_size = compute_module_sizes(model)[""]
            # We test several splits of sizes to make sure it works.
            max_gpu_sizes = [int(p * model_size) for p in self.model_split_percents]
            with tempfile.TemporaryDirectory() as tmp_dir:
                model.cpu().save_pretrained(tmp_dir)

                for max_size in max_gpu_sizes:
                    max_memory = {0: max_size, 1: model_size * 2, "cpu": model_size * 2}
                    new_model = model_class.from_pretrained(tmp_dir, device_map="auto", max_memory=max_memory)
                    # Making sure part of the model will actually end up offloaded
                    self.assertSetEqual(set(new_model.hf_device_map.values()), {0, 1})

                    self.check_device_map_is_respected(new_model, new_model.hf_device_map)

                    torch.manual_seed(0)
                    new_output = new_model(**inputs_dict_class)

                    self.assertTrue(torch.allclose(base_output[0], new_output[0], atol=1e-5))

    def test_problem_types(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        problem_types = [
            {"title": "multi_label_classification", "num_labels": 2, "dtype": torch.float},
            {"title": "single_label_classification", "num_labels": 1, "dtype": torch.long},
            {"title": "regression", "num_labels": 1, "dtype": torch.float},
        ]

        for model_class in self.all_model_classes:
            if model_class.__name__ not in [
                *get_values(MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES),
                *get_values(MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING_NAMES),
            ]:
                continue

            for problem_type in problem_types:
                with self.subTest(msg=f"Testing {model_class} with {problem_type['title']}"):
                    config.problem_type = problem_type["title"]
                    config.num_labels = problem_type["num_labels"]

                    model = model_class(config)
                    model.to(torch_device)
                    model.train()

                    inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)

                    if problem_type["num_labels"] > 1:
                        inputs["labels"] = inputs["labels"].unsqueeze(1).repeat(1, problem_type["num_labels"])

                    inputs["labels"] = inputs["labels"].to(problem_type["dtype"])

                    # This tests that we do not trigger the warning form PyTorch "Using a target size that is different
                    # to the input size. This will likely lead to incorrect results due to broadcasting. Please ensure
                    # they have the same size." which is a symptom something in wrong for the regression problem.
                    # See https://github.com/huggingface/transformers/issues/11780
                    with warnings.catch_warnings(record=True) as warning_list:
                        loss = model(**inputs).loss
                    for w in warning_list:
                        if "Using a target size that is different to the input size" in str(w.message):
                            raise ValueError(
                                f"Something is going wrong in the regression problem: intercepted {w.message}"
                            )

                    loss.backward()

    def test_load_with_mismatched_shapes(self):
        if not self.test_mismatched_shapes:
            return
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            if model_class.__name__ not in get_values(MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES):
                continue

            with self.subTest(msg=f"Testing {model_class}"):
                with tempfile.TemporaryDirectory() as tmp_dir:
                    model = model_class(config)
                    model.save_pretrained(tmp_dir)

                    # Fails when we don't set ignore_mismatched_sizes=True
                    with self.assertRaises(RuntimeError):
                        new_model = AutoModelForSequenceClassification.from_pretrained(tmp_dir, num_labels=42)
                    with self.assertRaises(RuntimeError):
                        new_model_without_prefix = AutoModel.from_pretrained(tmp_dir, vocab_size=10)

                    logger = logging.get_logger("transformers.modeling_utils")

                    with CaptureLogger(logger) as cl:
                        new_model = AutoModelForSequenceClassification.from_pretrained(
                            tmp_dir, num_labels=42, ignore_mismatched_sizes=True
                        )
                    self.assertIn("the shapes did not match", cl.out)
                    new_model.to(torch_device)
                    inputs = self._prepare_for_class(inputs_dict, model_class)
                    logits = new_model(**inputs).logits
                    self.assertEqual(logits.shape[1], 42)

                    with CaptureLogger(logger) as cl:
                        new_model_without_prefix = AutoModel.from_pretrained(
                            tmp_dir, vocab_size=10, ignore_mismatched_sizes=True
                        )
                    self.assertIn("the shapes did not match", cl.out)
                    input_ids = ids_tensor((2, 8), 10)
                    new_model_without_prefix.to(torch_device)
                    if self.is_encoder_decoder:
                        new_model_without_prefix(input_ids, decoder_input_ids=input_ids)
                    else:
                        new_model_without_prefix(input_ids)


global_rng = random.Random()


def ids_tensor(shape, vocab_size, rng=None, name=None):
    #  Creates a random int32 tensor of the shape within the vocab size
    if rng is None:
        rng = global_rng

    total_dims = 1
    for dim in shape:
        total_dims *= dim

    values = []
    for _ in range(total_dims):
        values.append(rng.randint(0, vocab_size - 1))

    return torch.tensor(data=values, dtype=torch.long, device=torch_device).view(shape).contiguous()


def random_attention_mask(shape, rng=None, name=None):
    attn_mask = ids_tensor(shape, vocab_size=2, rng=None, name=None)
    # make sure that at least one token is attended to for each batch
    attn_mask[:, -1] = 1
    return attn_mask


def floats_tensor(shape, scale=1.0, rng=None, name=None):
    """Creates a random float32 tensor"""
    if rng is None:
        rng = global_rng

    total_dims = 1
    for dim in shape:
        total_dims *= dim

    values = []
    for _ in range(total_dims):
        values.append(rng.random() * scale)

    return torch.tensor(data=values, dtype=torch.float, device=torch_device).view(shape).contiguous()


def check_models_equal(model1, model2):
    models_are_equal = True
    for model1_p, model2_p in zip(model1.parameters(), model2.parameters()):
        if model1_p.data.ne(model2_p.data).sum() > 0:
            models_are_equal = False

    return models_are_equal


@require_torch
class ModelUtilsTest(TestCasePlus):
    @slow
    def test_model_from_pretrained(self):
        for model_name in BERT_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
            config = BertConfig.from_pretrained(model_name)
            self.assertIsNotNone(config)
            self.assertIsInstance(config, PretrainedConfig)

            model = BertModel.from_pretrained(model_name)
            model, loading_info = BertModel.from_pretrained(model_name, output_loading_info=True)
            self.assertIsNotNone(model)
            self.assertIsInstance(model, PreTrainedModel)

            self.assertEqual(len(loading_info["missing_keys"]), 0)
            self.assertEqual(len(loading_info["unexpected_keys"]), 8)
            self.assertEqual(len(loading_info["mismatched_keys"]), 0)
            self.assertEqual(len(loading_info["error_msgs"]), 0)

            config = BertConfig.from_pretrained(model_name, output_attentions=True, output_hidden_states=True)

            # Not sure this is the intended behavior. TODO fix Lysandre & Thom
            config.name_or_path = model_name

            model = BertModel.from_pretrained(model_name, output_attentions=True, output_hidden_states=True)
            self.assertEqual(model.config.output_hidden_states, True)
            self.assertEqual(model.config, config)

    def test_model_from_pretrained_subfolder(self):
        config = BertConfig.from_pretrained("hf-internal-testing/tiny-random-bert")
        model = BertModel(config)

        subfolder = "bert"
        with tempfile.TemporaryDirectory() as tmp_dir:
            model.save_pretrained(os.path.join(tmp_dir, subfolder))

            with self.assertRaises(OSError):
                _ = BertModel.from_pretrained(tmp_dir)

            model_loaded = BertModel.from_pretrained(tmp_dir, subfolder=subfolder)

        self.assertTrue(check_models_equal(model, model_loaded))

    def test_model_from_pretrained_subfolder_sharded(self):
        config = BertConfig.from_pretrained("hf-internal-testing/tiny-random-bert")
        model = BertModel(config)

        subfolder = "bert"
        with tempfile.TemporaryDirectory() as tmp_dir:
            model.save_pretrained(os.path.join(tmp_dir, subfolder), max_shard_size="10KB")

            with self.assertRaises(OSError):
                _ = BertModel.from_pretrained(tmp_dir)

            model_loaded = BertModel.from_pretrained(tmp_dir, subfolder=subfolder)

        self.assertTrue(check_models_equal(model, model_loaded))

    def test_model_from_pretrained_hub_subfolder(self):
        subfolder = "bert"
        model_id = "hf-internal-testing/tiny-random-bert-subfolder"
        with self.assertRaises(OSError):
            _ = BertModel.from_pretrained(model_id)

        model = BertModel.from_pretrained(model_id, subfolder=subfolder)

        self.assertIsNotNone(model)

    def test_model_from_pretrained_hub_subfolder_sharded(self):
        subfolder = "bert"
        model_id = "hf-internal-testing/tiny-random-bert-sharded-subfolder"
        with self.assertRaises(OSError):
            _ = BertModel.from_pretrained(model_id)

        model = BertModel.from_pretrained(model_id, subfolder=subfolder)

        self.assertIsNotNone(model)

    def test_model_from_pretrained_with_different_pretrained_model_name(self):
        model = T5ForConditionalGeneration.from_pretrained(TINY_T5)
        self.assertIsNotNone(model)

        logger = logging.get_logger("transformers.configuration_utils")
        with CaptureLogger(logger) as cl:
            BertModel.from_pretrained(TINY_T5)
        self.assertTrue("You are using a model of type t5 to instantiate a model of type bert" in cl.out)

    def test_model_from_config_torch_dtype(self):
        # test that the model can be instantiated with dtype of user's choice - as long as it's a
        # float dtype. To make it happen config.torch_dtype needs to be set before instantiating the
        # model from the config object.

        config = T5Config.from_pretrained(TINY_T5)
        model = AutoModel.from_config(config)
        # XXX: isn't supported
        # model = T5ForConditionalGeneration.from_config(config)
        self.assertEqual(model.dtype, torch.float32)

        model = AutoModel.from_config(config, torch_dtype=torch.float16)
        self.assertEqual(model.dtype, torch.float16)

        # torch.set_default_dtype() supports only float dtypes, so will fail with non-float type
        with self.assertRaises(ValueError):
            model = AutoModel.from_config(config, torch_dtype=torch.int64)

    def test_model_from_pretrained_torch_dtype(self):
        # test that the model can be instantiated with dtype of either
        # 1. explicit from_pretrained's torch_dtype argument
        # 2. via autodiscovery by looking at model weights (torch_dtype="auto")
        # so if a model.half() was saved, we want it to be instantiated as such.
        #
        # test an explicit model class, but also AutoModel separately as the latter goes through a different code path
        model_path = self.get_auto_remove_tmp_dir()

        # baseline - we know TINY_T5 is fp32 model
        model = T5ForConditionalGeneration.from_pretrained(TINY_T5)
        self.assertEqual(model.dtype, torch.float32)

        def remove_torch_dtype(model_path):
            file = f"{model_path}/config.json"
            with open(file, "r", encoding="utf-8") as f:
                s = json.load(f)
            s.pop("torch_dtype")
            with open(file, "w", encoding="utf-8") as f:
                json.dump(s, f)

        # test the default fp32 save_pretrained => from_pretrained cycle
        model.save_pretrained(model_path)
        model = T5ForConditionalGeneration.from_pretrained(model_path)
        self.assertEqual(model.dtype, torch.float32)
        # 1. test torch_dtype="auto" via `config.torch_dtype`
        model = T5ForConditionalGeneration.from_pretrained(model_path, torch_dtype="auto")
        self.assertEqual(model.dtype, torch.float32)
        # 2. test torch_dtype="auto" via auto-derivation
        # now remove the torch_dtype entry from config.json and try "auto" again which should
        # perform auto-derivation from weights
        remove_torch_dtype(model_path)
        model = T5ForConditionalGeneration.from_pretrained(model_path, torch_dtype="auto")
        self.assertEqual(model.dtype, torch.float32)

        # test forced loading in fp16 (even though the weights are in fp32)
        model = T5ForConditionalGeneration.from_pretrained(model_path, torch_dtype=torch.float16)
        self.assertEqual(model.dtype, torch.float16)

        # test fp16 save_pretrained, loaded with auto-detection
        model = model.half()
        model.save_pretrained(model_path)
        # 1. test torch_dtype="auto" via `config.torch_dtype`
        model = T5ForConditionalGeneration.from_pretrained(model_path, torch_dtype="auto")
        self.assertEqual(model.config.torch_dtype, torch.float16)
        self.assertEqual(model.dtype, torch.float16)
        # tests `config.torch_dtype` saving
        with open(f"{model_path}/config.json") as f:
            config_dict = json.load(f)
        self.assertEqual(config_dict["torch_dtype"], "float16")
        # 2. test torch_dtype="auto" via auto-derivation
        # now same with using config info
        remove_torch_dtype(model_path)
        model = T5ForConditionalGeneration.from_pretrained(model_path, torch_dtype="auto")
        self.assertEqual(model.dtype, torch.float16)

        # test fp16 save_pretrained, loaded with the explicit fp16
        model = T5ForConditionalGeneration.from_pretrained(model_path, torch_dtype=torch.float16)
        self.assertEqual(model.dtype, torch.float16)

        # test AutoModel separately as it goes through a different path
        # test auto-detection - as currently TINY_T5 doesn't have torch_dtype entry
        model = AutoModel.from_pretrained(TINY_T5, torch_dtype="auto")
        # test that the config object didn't get polluted with torch_dtype="auto"
        # there was a bug that after this call we ended up with config.torch_dtype=="auto"
        self.assertNotEqual(model.config.torch_dtype, "auto")
        # now test the outcome
        self.assertEqual(model.dtype, torch.float32)
        model = AutoModel.from_pretrained(TINY_T5, torch_dtype=torch.float16)
        self.assertEqual(model.dtype, torch.float16)

        # test model whose first param is not of a floating type, but int
        model = AutoModel.from_pretrained(TINY_BERT_FOR_TOKEN_CLASSIFICATION, torch_dtype="auto")
        self.assertEqual(model.dtype, torch.float32)

    def test_no_super_init_config_and_model(self):
        config = NoSuperInitConfig(attribute=32)
        model = NoSuperInitModel(config)

        with tempfile.TemporaryDirectory() as tmp_dir:
            model.save_pretrained(tmp_dir)

            new_model = NoSuperInitModel.from_pretrained(tmp_dir)

        for p1, p2 in zip(model.parameters(), new_model.parameters()):
            self.assertTrue(torch.equal(p1, p2))

    def test_shard_checkpoint(self):
        # This is the model we will use, total size 340,000 bytes.
        model = torch.nn.Sequential(
            torch.nn.Linear(100, 200, bias=False),  # size 80,000
            torch.nn.Linear(200, 200, bias=False),  # size 160,000
            torch.nn.Linear(200, 100, bias=False),  # size 80,000
            torch.nn.Linear(100, 50, bias=False),  # size 20,000
        )
        state_dict = model.state_dict()

        with self.subTest("No shard when max size is bigger than model size"):
            shards, index = shard_checkpoint(state_dict)
            self.assertIsNone(index)
            self.assertDictEqual(shards, {WEIGHTS_NAME: state_dict})

        with self.subTest("Test sharding, no weights bigger than max size"):
            shards, index = shard_checkpoint(state_dict, max_shard_size="300kB")
            # Split is first two layers then last two.
            self.assertDictEqual(
                index,
                {
                    "metadata": {"total_size": 340000},
                    "weight_map": {
                        "0.weight": "pytorch_model-00001-of-00002.bin",
                        "1.weight": "pytorch_model-00001-of-00002.bin",
                        "2.weight": "pytorch_model-00002-of-00002.bin",
                        "3.weight": "pytorch_model-00002-of-00002.bin",
                    },
                },
            )

            shard1 = {"0.weight": state_dict["0.weight"], "1.weight": state_dict["1.weight"]}
            shard2 = {"2.weight": state_dict["2.weight"], "3.weight": state_dict["3.weight"]}
            self.assertDictEqual(
                shards, {"pytorch_model-00001-of-00002.bin": shard1, "pytorch_model-00002-of-00002.bin": shard2}
            )

        with self.subTest("Test sharding with weights bigger than max size"):
            shards, index = shard_checkpoint(state_dict, max_shard_size="100kB")
            # Split is first layer, second layer then last 2.
            self.assertDictEqual(
                index,
                {
                    "metadata": {"total_size": 340000},
                    "weight_map": {
                        "0.weight": "pytorch_model-00001-of-00003.bin",
                        "1.weight": "pytorch_model-00002-of-00003.bin",
                        "2.weight": "pytorch_model-00003-of-00003.bin",
                        "3.weight": "pytorch_model-00003-of-00003.bin",
                    },
                },
            )

            shard1 = {"0.weight": state_dict["0.weight"]}
            shard2 = {"1.weight": state_dict["1.weight"]}
            shard3 = {"2.weight": state_dict["2.weight"], "3.weight": state_dict["3.weight"]}
            self.assertDictEqual(
                shards,
                {
                    "pytorch_model-00001-of-00003.bin": shard1,
                    "pytorch_model-00002-of-00003.bin": shard2,
                    "pytorch_model-00003-of-00003.bin": shard3,
                },
            )

    def test_checkpoint_sharding_local(self):
        model = BertModel.from_pretrained("hf-internal-testing/tiny-random-bert")

        with tempfile.TemporaryDirectory() as tmp_dir:
            # We use the same folder for various sizes to make sure a new save erases the old checkpoint.
            for max_size in ["50kB", "50kiB", "100kB", "100kiB", "200kB", "200kiB"]:
                model.save_pretrained(tmp_dir, max_shard_size=max_size)

                # Get each shard file and its size
                shard_to_size = {}
                for shard in os.listdir(tmp_dir):
                    if shard.endswith(".bin"):
                        shard_file = os.path.join(tmp_dir, shard)
                        shard_to_size[shard_file] = os.path.getsize(shard_file)

                index_file = os.path.join(tmp_dir, WEIGHTS_INDEX_NAME)
                # Check there is an index but no regular weight file
                self.assertTrue(os.path.isfile(index_file))
                self.assertFalse(os.path.isfile(os.path.join(tmp_dir, WEIGHTS_NAME)))

                # Check a file is bigger than max_size only when it has a single weight
                for shard_file, size in shard_to_size.items():
                    if max_size.endswith("kiB"):
                        max_size_int = int(max_size[:-3]) * 2**10
                    else:
                        max_size_int = int(max_size[:-2]) * 10**3
                    # Note: pickle adds some junk so the weight of the file can end up being slightly bigger than
                    # the size asked for (since we count parameters)
                    if size >= max_size_int + 50000:
                        state_dict = torch.load(shard_file)
                        self.assertEqual(len(state_dict), 1)

                # Check the index and the shard files found match
                with open(index_file, "r", encoding="utf-8") as f:
                    index = json.loads(f.read())

                all_shards = set(index["weight_map"].values())
                shards_found = {f for f in os.listdir(tmp_dir) if f.endswith(".bin")}
                self.assertSetEqual(all_shards, shards_found)

                # Finally, check the model can be reloaded
                new_model = BertModel.from_pretrained(tmp_dir)
                for p1, p2 in zip(model.parameters(), new_model.parameters()):
                    self.assertTrue(torch.allclose(p1, p2))

    def test_checkpoint_sharding_from_hub(self):
        model = BertModel.from_pretrained("hf-internal-testing/tiny-random-bert-sharded")
        # the model above is the same as the model below, just a sharded version.
        ref_model = BertModel.from_pretrained("hf-internal-testing/tiny-random-bert")
        for p1, p2 in zip(model.parameters(), ref_model.parameters()):
            self.assertTrue(torch.allclose(p1, p2))

    def test_checkpoint_variant_local(self):
        model = BertModel.from_pretrained("hf-internal-testing/tiny-random-bert")

        with tempfile.TemporaryDirectory() as tmp_dir:
            model.save_pretrained(tmp_dir, variant="v2")

            weights_name = ".".join(WEIGHTS_NAME.split(".")[:-1] + ["v2"] + ["bin"])

            weights_file = os.path.join(tmp_dir, weights_name)
            self.assertTrue(os.path.isfile(weights_file))
            self.assertFalse(os.path.isfile(os.path.join(tmp_dir, WEIGHTS_NAME)))

            with self.assertRaises(EnvironmentError):
                _ = BertModel.from_pretrained(tmp_dir)

            new_model = BertModel.from_pretrained(tmp_dir, variant="v2")

        for p1, p2 in zip(model.parameters(), new_model.parameters()):
            self.assertTrue(torch.allclose(p1, p2))

    def test_checkpoint_variant_local_sharded(self):
        model = BertModel.from_pretrained("hf-internal-testing/tiny-random-bert")

        with tempfile.TemporaryDirectory() as tmp_dir:
            model.save_pretrained(tmp_dir, variant="v2", max_shard_size="50kB")

            weights_index_name = ".".join(WEIGHTS_INDEX_NAME.split(".")[:-1] + ["v2"] + ["json"])
            weights_index_file = os.path.join(tmp_dir, weights_index_name)
            self.assertTrue(os.path.isfile(weights_index_file))
            self.assertFalse(os.path.isfile(os.path.join(tmp_dir, WEIGHTS_INDEX_NAME)))

            for i in range(1, 6):
                weights_name = ".".join(WEIGHTS_NAME.split(".")[:-1] + [f"v2-0000{i}-of-00006"] + ["bin"])
                weights_name_file = os.path.join(tmp_dir, weights_name)
                self.assertTrue(os.path.isfile(weights_name_file))

            with self.assertRaises(EnvironmentError):
                _ = BertModel.from_pretrained(tmp_dir)

            new_model = BertModel.from_pretrained(tmp_dir, variant="v2")

        for p1, p2 in zip(model.parameters(), new_model.parameters()):
            self.assertTrue(torch.allclose(p1, p2))

    @require_safetensors
    def test_checkpoint_variant_local_safe(self):
        model = BertModel.from_pretrained("hf-internal-testing/tiny-random-bert")

        with tempfile.TemporaryDirectory() as tmp_dir:
            model.save_pretrained(tmp_dir, variant="v2", safe_serialization=True)

            weights_name = ".".join(SAFE_WEIGHTS_NAME.split(".")[:-1] + ["v2"] + ["safetensors"])

            weights_file = os.path.join(tmp_dir, weights_name)
            self.assertTrue(os.path.isfile(weights_file))
            self.assertFalse(os.path.isfile(os.path.join(tmp_dir, SAFE_WEIGHTS_NAME)))

            with self.assertRaises(EnvironmentError):
                _ = BertModel.from_pretrained(tmp_dir)

            new_model = BertModel.from_pretrained(tmp_dir, variant="v2")

        for p1, p2 in zip(model.parameters(), new_model.parameters()):
            self.assertTrue(torch.allclose(p1, p2))

    @require_safetensors
    def test_checkpoint_variant_local_sharded_safe(self):
        model = BertModel.from_pretrained("hf-internal-testing/tiny-random-bert")

        with tempfile.TemporaryDirectory() as tmp_dir:
            model.save_pretrained(tmp_dir, variant="v2", max_shard_size="50kB", safe_serialization=True)

            weights_index_name = ".".join(SAFE_WEIGHTS_INDEX_NAME.split(".")[:-1] + ["v2"] + ["json"])
            weights_index_file = os.path.join(tmp_dir, weights_index_name)
            self.assertTrue(os.path.isfile(weights_index_file))
            self.assertFalse(os.path.isfile(os.path.join(tmp_dir, SAFE_WEIGHTS_INDEX_NAME)))

            for i in range(1, 6):
                weights_name = ".".join(SAFE_WEIGHTS_NAME.split(".")[:-1] + [f"v2-0000{i}-of-00006"] + ["safetensors"])
                weights_name_file = os.path.join(tmp_dir, weights_name)
                self.assertTrue(os.path.isfile(weights_name_file))

            with self.assertRaises(EnvironmentError):
                _ = BertModel.from_pretrained(tmp_dir)

            new_model = BertModel.from_pretrained(tmp_dir, variant="v2")

        for p1, p2 in zip(model.parameters(), new_model.parameters()):
            self.assertTrue(torch.allclose(p1, p2))

    def test_checkpoint_variant_hub(self):
        with tempfile.TemporaryDirectory() as tmp_dir:
            with self.assertRaises(EnvironmentError):
                _ = BertModel.from_pretrained("hf-internal-testing/tiny-random-bert-variant", cache_dir=tmp_dir)
            model = BertModel.from_pretrained(
                "hf-internal-testing/tiny-random-bert-variant", cache_dir=tmp_dir, variant="v2"
            )
        self.assertIsNotNone(model)

    def test_checkpoint_variant_hub_sharded(self):
        with tempfile.TemporaryDirectory() as tmp_dir:
            with self.assertRaises(EnvironmentError):
                _ = BertModel.from_pretrained(
                    "hf-internal-testing/tiny-random-bert-variant-sharded", cache_dir=tmp_dir
                )
            model = BertModel.from_pretrained(
                "hf-internal-testing/tiny-random-bert-variant-sharded", cache_dir=tmp_dir, variant="v2"
            )
        self.assertIsNotNone(model)

    @require_safetensors
    def test_checkpoint_variant_hub_safe(self):
        with tempfile.TemporaryDirectory() as tmp_dir:
            with self.assertRaises(EnvironmentError):
                _ = BertModel.from_pretrained("hf-internal-testing/tiny-random-bert-variant-safe", cache_dir=tmp_dir)
            model = BertModel.from_pretrained(
                "hf-internal-testing/tiny-random-bert-variant-safe", cache_dir=tmp_dir, variant="v2"
            )
        self.assertIsNotNone(model)

    @require_safetensors
    def test_checkpoint_variant_hub_sharded_safe(self):
        with tempfile.TemporaryDirectory() as tmp_dir:
            with self.assertRaises(EnvironmentError):
                _ = BertModel.from_pretrained(
                    "hf-internal-testing/tiny-random-bert-variant-sharded-safe", cache_dir=tmp_dir
                )
            model = BertModel.from_pretrained(
                "hf-internal-testing/tiny-random-bert-variant-sharded-safe", cache_dir=tmp_dir, variant="v2"
            )
        self.assertIsNotNone(model)

    def test_checkpoint_variant_save_load(self):
        with tempfile.TemporaryDirectory() as tmp_dir:
            model = BertModel.from_pretrained(
                "hf-internal-testing/tiny-random-bert-variant", cache_dir=tmp_dir, variant="v2"
            )
            weights_name = ".".join(WEIGHTS_NAME.split(".")[:-1] + ["v2"] + ["bin"])

            model.save_pretrained(tmp_dir, variant="v2")
            # saving will create a variant checkpoint
            self.assertTrue(os.path.isfile(os.path.join(tmp_dir, weights_name)))

            model.save_pretrained(tmp_dir)
            # saving shouldn't delete variant checkpoints
            weights_name = ".".join(WEIGHTS_NAME.split(".")[:-1] + ["v2"] + ["bin"])
            self.assertTrue(os.path.isfile(os.path.join(tmp_dir, weights_name)))

            # there should be a normal checkpoint
            self.assertTrue(os.path.isfile(os.path.join(tmp_dir, WEIGHTS_NAME)))

        self.assertIsNotNone(model)

    @require_accelerate
    @mark.accelerate_tests
    def test_from_pretrained_low_cpu_mem_usage_functional(self):
        # test that we can use `from_pretrained(..., low_cpu_mem_usage=True)` with normal and
        # sharded models

        mnames = [
            "hf-internal-testing/tiny-random-bert-sharded",
            "hf-internal-testing/tiny-random-bert",
        ]
        for mname in mnames:
            _ = BertModel.from_pretrained(mname, low_cpu_mem_usage=True)

    @require_usr_bin_time
    @require_accelerate
    @mark.accelerate_tests
    def test_from_pretrained_low_cpu_mem_usage_measured(self):
        # test that `from_pretrained(..., low_cpu_mem_usage=True)` uses less cpu memory than default

        mname = "bert-base-cased"

        preamble = "from transformers import AutoModel"
        one_liner_str = f'{preamble}; AutoModel.from_pretrained("{mname}", low_cpu_mem_usage=False)'
        max_rss_normal = self.python_one_liner_max_rss(one_liner_str)
        # print(f"{max_rss_normal=}")

        one_liner_str = f'{preamble};  AutoModel.from_pretrained("{mname}", low_cpu_mem_usage=True)'
        max_rss_low_mem = self.python_one_liner_max_rss(one_liner_str)
        # print(f"{max_rss_low_mem=}")

        diff_bytes = max_rss_normal - max_rss_low_mem
        diff_percent = diff_bytes / max_rss_low_mem
        # print(f"{diff_bytes=}, {diff_percent=}")
        # ideally we would compare that the diff is close to ~1x checkpoint size in bytes, but
        # measuring cpu memory on linux is very tricky and inconsistent, so instead let's check that
        # it's at least 15% less cpu memory consumed

        self.assertGreater(
            diff_percent,
            0.15,
            "should use less CPU memory for low_cpu_mem_usage=True, "
            f"but got max_rss_normal={max_rss_normal} and max_rss_low_mem={max_rss_low_mem}",
        )

        # if you want to compare things manually, let's first look at the size of the model in bytes
        # model = BertModel.from_pretrained(mname, low_cpu_mem_usage=False)
        # total_numel = sum(dict((p.data_ptr(), p.numel()) for p in model.parameters()).values())
        # total_bytes = total_numel * 4  # 420MB
        # Now the diff_bytes should be very close to total_bytes, but the reports are inconsistent.
        # The easiest way to test this is to switch the model and torch.load to do all the work on
        # gpu - that way one can measure exactly the total and peak memory used. Perhaps once we add
        # functionality to load models directly on gpu, this test can be rewritten to use torch's
        # cuda memory tracking and then we should be able to do a much more precise test.

    @require_accelerate
    @mark.accelerate_tests
    @require_torch_multi_gpu
    @slow
    def test_model_parallelism_gpt2(self):
        device_map = {"transformer.wte": 0, "transformer.wpe": 0, "lm_head": 0, "transformer.ln_f": 1}
        for i in range(12):
            device_map[f"transformer.h.{i}"] = 0 if i <= 5 else 1

        model = AutoModelForCausalLM.from_pretrained("gpt2", device_map=device_map)

        tokenizer = AutoTokenizer.from_pretrained("gpt2")
        inputs = tokenizer("Hello, my name is", return_tensors="pt")
        output = model.generate(inputs["input_ids"].to(0))

        text_output = tokenizer.decode(output[0].tolist())
        self.assertEqual(text_output, "Hello, my name is John. I'm a writer, and I'm a writer. I'm")

    @require_accelerate
    @mark.accelerate_tests
    @require_torch_gpu
    def test_from_pretrained_disk_offload_task_model(self):
        model = AutoModel.from_pretrained("hf-internal-testing/tiny-random-gpt2")
        device_map = {
            "transformer.wte": 0,
            "transformer.wpe": 0,
            "transformer.h.0": "cpu",
            "transformer.h.1": "cpu",
            "transformer.h.2": "cpu",
            "transformer.h.3": "disk",
            "transformer.h.4": "disk",
            "transformer.ln_f": 0,
            "lm_head": 0,
        }
        with tempfile.TemporaryDirectory() as tmp_dir:
            inputs = torch.tensor([[1, 2, 3]]).to(0)

            model.save_pretrained(tmp_dir)
            new_model = AutoModelForCausalLM.from_pretrained(tmp_dir).to(0)
            outputs1 = new_model.to(0)(inputs)

            offload_folder = os.path.join(tmp_dir, "offload")
            new_model_with_offload = AutoModelForCausalLM.from_pretrained(
                tmp_dir, device_map=device_map, offload_folder=offload_folder
            )
            outputs2 = new_model_with_offload(inputs)

            self.assertTrue(torch.allclose(outputs1.logits.cpu(), outputs2.logits.cpu()))

            # With state dict temp offload
            offload_folder = os.path.join(tmp_dir, "offload")
            new_model_with_offload = AutoModelForCausalLM.from_pretrained(
                tmp_dir,
                device_map=device_map,
                offload_folder=offload_folder,
                offload_state_dict=True,
            )
            outputs2 = new_model_with_offload(inputs)

            self.assertTrue(torch.allclose(outputs1.logits.cpu(), outputs2.logits.cpu()))

    def test_cached_files_are_used_when_internet_is_down(self):
        # A mock response for an HTTP head request to emulate server down
        response_mock = mock.Mock()
        response_mock.status_code = 500
        response_mock.headers = {}
        response_mock.raise_for_status.side_effect = HTTPError
        response_mock.json.return_value = {}

        # Download this model to make sure it's in the cache.
        _ = BertModel.from_pretrained("hf-internal-testing/tiny-random-bert")

        # Under the mock environment we get a 500 error when trying to reach the model.
        with mock.patch("requests.request", return_value=response_mock) as mock_head:
            _ = BertModel.from_pretrained("hf-internal-testing/tiny-random-bert")
            # This check we did call the fake head request
            mock_head.assert_called()

    def test_load_from_one_file(self):
        try:
            tmp_file = tempfile.mktemp()
            with open(tmp_file, "wb") as f:
                http_get(
                    "https://huggingface.co/hf-internal-testing/tiny-random-bert/resolve/main/pytorch_model.bin", f
                )

            config = BertConfig.from_pretrained("hf-internal-testing/tiny-random-bert")
            _ = BertModel.from_pretrained(tmp_file, config=config)
        finally:
            os.remove(tmp_file)

    def test_legacy_load_from_url(self):
        # This test is for deprecated behavior and can be removed in v5
        config = BertConfig.from_pretrained("hf-internal-testing/tiny-random-bert")
        _ = BertModel.from_pretrained(
            "https://huggingface.co/hf-internal-testing/tiny-random-bert/resolve/main/pytorch_model.bin", config=config
        )

    @require_safetensors
    def test_use_safetensors(self):
        # test nice error message if no safetensor files available
        with self.assertRaises(OSError) as env_error:
            AutoModel.from_pretrained("hf-internal-testing/tiny-random-RobertaModel", use_safetensors=True)

        self.assertTrue(
            "model.safetensors or model.safetensors.index.json and thus cannot be loaded with `safetensors`"
            in str(env_error.exception)
        )

        # test that error if only safetensors is available
        with self.assertRaises(OSError) as env_error:
            BertModel.from_pretrained("hf-internal-testing/tiny-random-bert-safetensors", use_safetensors=False)

        self.assertTrue("does not appear to have a file named pytorch_model.bin" in str(env_error.exception))

        # test that only safetensors if both available and use_safetensors=False
        with tempfile.TemporaryDirectory() as tmp_dir:
            CLIPTextModel.from_pretrained(
                "hf-internal-testing/diffusers-stable-diffusion-tiny-all",
                subfolder="text_encoder",
                use_safetensors=False,
                cache_dir=tmp_dir,
            )

            all_downloaded_files = glob.glob(os.path.join(tmp_dir, "*", "snapshots", "*", "*", "*"))
            self.assertTrue(any(f.endswith("bin") for f in all_downloaded_files))
            self.assertFalse(any(f.endswith("safetensors") for f in all_downloaded_files))

        # test that no safetensors if both available and use_safetensors=True
        with tempfile.TemporaryDirectory() as tmp_dir:
            CLIPTextModel.from_pretrained(
                "hf-internal-testing/diffusers-stable-diffusion-tiny-all",
                subfolder="text_encoder",
                use_safetensors=True,
                cache_dir=tmp_dir,
            )

            all_downloaded_files = glob.glob(os.path.join(tmp_dir, "*", "snapshots", "*", "*", "*"))
            self.assertTrue(any(f.endswith("safetensors") for f in all_downloaded_files))
            self.assertFalse(any(f.endswith("bin") for f in all_downloaded_files))

    @require_safetensors
    def test_safetensors_save_and_load(self):
        model = BertModel.from_pretrained("hf-internal-testing/tiny-random-bert")
        with tempfile.TemporaryDirectory() as tmp_dir:
            model.save_pretrained(tmp_dir, safe_serialization=True)
            # No pytorch_model.bin file, only a model.safetensors
            self.assertTrue(os.path.isfile(os.path.join(tmp_dir, SAFE_WEIGHTS_NAME)))
            self.assertFalse(os.path.isfile(os.path.join(tmp_dir, WEIGHTS_NAME)))

            new_model = BertModel.from_pretrained(tmp_dir)

            # Check models are equal
            for p1, p2 in zip(model.parameters(), new_model.parameters()):
                self.assertTrue(torch.allclose(p1, p2))

    @require_safetensors
    def test_safetensors_load_from_hub(self):
        safetensors_model = BertModel.from_pretrained("hf-internal-testing/tiny-random-bert-safetensors")
        pytorch_model = BertModel.from_pretrained("hf-internal-testing/tiny-random-bert")

        # Check models are equal
        for p1, p2 in zip(safetensors_model.parameters(), pytorch_model.parameters()):
            self.assertTrue(torch.allclose(p1, p2))

    @require_safetensors
    def test_safetensors_save_and_load_sharded(self):
        model = BertModel.from_pretrained("hf-internal-testing/tiny-random-bert")
        with tempfile.TemporaryDirectory() as tmp_dir:
            model.save_pretrained(tmp_dir, safe_serialization=True, max_shard_size="100kB")
            # No pytorch_model.bin index file, only a model.safetensors index
            self.assertFalse(os.path.isfile(os.path.join(tmp_dir, WEIGHTS_INDEX_NAME)))
            self.assertTrue(os.path.isfile(os.path.join(tmp_dir, SAFE_WEIGHTS_INDEX_NAME)))
            # No regular weights file
            self.assertFalse(os.path.isfile(os.path.join(tmp_dir, WEIGHTS_NAME)))
            self.assertFalse(os.path.isfile(os.path.join(tmp_dir, SAFE_WEIGHTS_NAME)))

            new_model = BertModel.from_pretrained(tmp_dir)

            # Check models are equal
            for p1, p2 in zip(model.parameters(), new_model.parameters()):
                self.assertTrue(torch.allclose(p1, p2))

    @require_safetensors
    def test_safetensors_load_from_hub_sharded(self):
        safetensors_model = BertModel.from_pretrained("hf-internal-testing/tiny-random-bert-sharded-safetensors")
        pytorch_model = BertModel.from_pretrained("hf-internal-testing/tiny-random-bert-sharded")

        # Check models are equal
        for p1, p2 in zip(safetensors_model.parameters(), pytorch_model.parameters()):
            self.assertTrue(torch.allclose(p1, p2))

    def test_base_model_to_head_model_load(self):
        base_model = BaseModel(PretrainedConfig())
        with tempfile.TemporaryDirectory() as tmp_dir:
            base_model.save_pretrained(tmp_dir)

            # Can load a base model in a model with head
            model = ModelWithHead.from_pretrained(tmp_dir)
            for p1, p2 in zip(model.base.parameters(), base_model.parameters()):
                self.assertTrue(torch.allclose(p1, p2))

            # It doesn't work if the state dict has a mix of keys of the head and base without prefix though.
            base_state_dict = base_model.state_dict()
            head_state_dict = model.state_dict()
            base_state_dict["linear2.weight"] = head_state_dict["linear2.weight"]
            base_state_dict["linear2.bias"] = head_state_dict["linear2.bias"]
            torch.save(base_state_dict, os.path.join(tmp_dir, WEIGHTS_NAME))

            with self.assertRaisesRegex(
                ValueError, "The state dictionary of the model you are trying to load is corrupted."
            ):
                _ = ModelWithHead.from_pretrained(tmp_dir)

    @require_torch_gpu
    @slow
    def test_pretrained_low_mem_new_config(self):
        # Checking for 1 model(the same one which was described in the issue) .
        model_ids = ["gpt2"]

        for model_id in model_ids:
            model_config = AutoConfig.from_pretrained(pretrained_model_name_or_path=model_id)
            model_config.n_layer = 48
            model_config.n_head = 25
            model_config.n_embd = 1600
            model = AutoModelForCausalLM.from_pretrained(
                pretrained_model_name_or_path=model_id,
                config=model_config,
                ignore_mismatched_sizes=True,
                torch_dtype=torch.float16,
                low_cpu_mem_usage=True,
            )
            model_ref = AutoModelForCausalLM.from_pretrained(pretrained_model_name_or_path=model_id)

            self.assertEqual(model.__class__.__name__, model_ref.__class__.__name__)


@require_torch
@is_staging_test
class ModelPushToHubTester(unittest.TestCase):
    @classmethod
    def setUpClass(cls):
        cls._token = TOKEN
        HfFolder.save_token(TOKEN)

    @classmethod
    def tearDownClass(cls):
        try:
            delete_repo(token=cls._token, repo_id="test-model")
        except HTTPError:
            pass

        try:
            delete_repo(token=cls._token, repo_id="valid_org/test-model-org")
        except HTTPError:
            pass

        try:
            delete_repo(token=cls._token, repo_id="test-dynamic-model")
        except HTTPError:
            pass

    def test_push_to_hub(self):
        config = BertConfig(
            vocab_size=99, hidden_size=32, num_hidden_layers=5, num_attention_heads=4, intermediate_size=37
        )
        model = BertModel(config)
        model.push_to_hub("test-model", use_auth_token=self._token)

        new_model = BertModel.from_pretrained(f"{USER}/test-model")
        for p1, p2 in zip(model.parameters(), new_model.parameters()):
            self.assertTrue(torch.equal(p1, p2))

        # Reset repo
        delete_repo(token=self._token, repo_id="test-model")

        # Push to hub via save_pretrained
        with tempfile.TemporaryDirectory() as tmp_dir:
            model.save_pretrained(tmp_dir, repo_id="test-model", push_to_hub=True, use_auth_token=self._token)

        new_model = BertModel.from_pretrained(f"{USER}/test-model")
        for p1, p2 in zip(model.parameters(), new_model.parameters()):
            self.assertTrue(torch.equal(p1, p2))

    def test_push_to_hub_in_organization(self):
        config = BertConfig(
            vocab_size=99, hidden_size=32, num_hidden_layers=5, num_attention_heads=4, intermediate_size=37
        )
        model = BertModel(config)
        model.push_to_hub("valid_org/test-model-org", use_auth_token=self._token)

        new_model = BertModel.from_pretrained("valid_org/test-model-org")
        for p1, p2 in zip(model.parameters(), new_model.parameters()):
            self.assertTrue(torch.equal(p1, p2))

        # Reset repo
        delete_repo(token=self._token, repo_id="valid_org/test-model-org")

        # Push to hub via save_pretrained
        with tempfile.TemporaryDirectory() as tmp_dir:
            model.save_pretrained(
                tmp_dir, push_to_hub=True, use_auth_token=self._token, repo_id="valid_org/test-model-org"
            )

        new_model = BertModel.from_pretrained("valid_org/test-model-org")
        for p1, p2 in zip(model.parameters(), new_model.parameters()):
            self.assertTrue(torch.equal(p1, p2))

    def test_push_to_hub_dynamic_model(self):
        CustomConfig.register_for_auto_class()
        CustomModel.register_for_auto_class()

        config = CustomConfig(hidden_size=32)
        model = CustomModel(config)

        model.push_to_hub("test-dynamic-model", use_auth_token=self._token)
        # checks
        self.assertDictEqual(
            config.auto_map,
            {"AutoConfig": "custom_configuration.CustomConfig", "AutoModel": "custom_modeling.CustomModel"},
        )

        new_model = AutoModel.from_pretrained(f"{USER}/test-dynamic-model", trust_remote_code=True)
        # Can't make an isinstance check because the new_model is from the CustomModel class of a dynamic module
        self.assertEqual(new_model.__class__.__name__, "CustomModel")
        for p1, p2 in zip(model.parameters(), new_model.parameters()):
            self.assertTrue(torch.equal(p1, p2))

        config = AutoConfig.from_pretrained(f"{USER}/test-dynamic-model", trust_remote_code=True)
        new_model = AutoModel.from_config(config, trust_remote_code=True)
        self.assertEqual(new_model.__class__.__name__, "CustomModel")