faraway00 commited on
Commit
7ef255b
·
verified ·
1 Parent(s): 901c0cd

Record H100 runtime, numerical checks and source provenance for500 renderer judgments

Browse files
runs/render-bench5-score-inputs-20261117/h100/provenance.json ADDED
@@ -0,0 +1,267 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "status": "verified",
3
+ "videos": 500,
4
+ "models": 5,
5
+ "gpus": 2,
6
+ "pod": "wzv4oj6dfvuxht",
7
+ "runtime": {
8
+ "torch": "2.9.1+cu128",
9
+ "torchvision": "0.24.1+cu128",
10
+ "PIL": "12.3.0",
11
+ "numpy": "2.5.4",
12
+ "cv2": "4.11.0",
13
+ "transformers": "5.23.0",
14
+ "bitsandbytes": "0.50.2"
15
+ },
16
+ "model": {
17
+ "status": "verified",
18
+ "revision": "37561816e66b5b9153b1b64c10968c18c1e6f4a0"
19
+ },
20
+ "inputs": {
21
+ "status": "verified",
22
+ "videos": 500,
23
+ "frames": 5500
24
+ },
25
+ "checks": {
26
+ "gpu0/VISION_PATCH_EQUIVALENCE.json": {
27
+ "status": "passed",
28
+ "patches": 16,
29
+ "max_absolute_difference": 0.015625,
30
+ "bit_exact": false,
31
+ "rtol": 0.01,
32
+ "atol": 0.01,
33
+ "identity": "Kernel and stride equal full patch dimensions; Conv3d has exactly one spatial output per flattened patch, equal to the same weight matrix and bias"
34
+ },
35
+ "gpu0/VISION_END_TO_END_CHECK.json": {
36
+ "status": "checked",
37
+ "candidate_id": "C001",
38
+ "scores_exact": false,
39
+ "comparisons": 10,
40
+ "baseline": [
41
+ {
42
+ "time_s": 0.5,
43
+ "content_change_score": 0,
44
+ "style_change_score": 0
45
+ },
46
+ {
47
+ "time_s": 1.0,
48
+ "content_change_score": 1,
49
+ "style_change_score": 0
50
+ },
51
+ {
52
+ "time_s": 1.5,
53
+ "content_change_score": 2,
54
+ "style_change_score": 0
55
+ },
56
+ {
57
+ "time_s": 2.0,
58
+ "content_change_score": 3,
59
+ "style_change_score": 0
60
+ },
61
+ {
62
+ "time_s": 2.5,
63
+ "content_change_score": 3,
64
+ "style_change_score": 0
65
+ },
66
+ {
67
+ "time_s": 3.0,
68
+ "content_change_score": 3,
69
+ "style_change_score": 0
70
+ },
71
+ {
72
+ "time_s": 3.5,
73
+ "content_change_score": 3,
74
+ "style_change_score": 0
75
+ },
76
+ {
77
+ "time_s": 4.0,
78
+ "content_change_score": 3,
79
+ "style_change_score": 0
80
+ },
81
+ {
82
+ "time_s": 4.5,
83
+ "content_change_score": 3,
84
+ "style_change_score": 0
85
+ },
86
+ {
87
+ "time_s": 4.95,
88
+ "content_change_score": 3,
89
+ "style_change_score": 0
90
+ }
91
+ ],
92
+ "accelerated": [
93
+ {
94
+ "time_s": 0.5,
95
+ "content_change_score": 1,
96
+ "style_change_score": 0
97
+ },
98
+ {
99
+ "time_s": 1.0,
100
+ "content_change_score": 1,
101
+ "style_change_score": 0
102
+ },
103
+ {
104
+ "time_s": 1.5,
105
+ "content_change_score": 2,
106
+ "style_change_score": 0
107
+ },
108
+ {
109
+ "time_s": 2.0,
110
+ "content_change_score": 3,
111
+ "style_change_score": 0
112
+ },
113
+ {
114
+ "time_s": 2.5,
115
+ "content_change_score": 3,
116
+ "style_change_score": 0
117
+ },
118
+ {
119
+ "time_s": 3.0,
120
+ "content_change_score": 3,
121
+ "style_change_score": 0
122
+ },
123
+ {
124
+ "time_s": 3.5,
125
+ "content_change_score": 3,
126
+ "style_change_score": 0
127
+ },
128
+ {
129
+ "time_s": 4.0,
130
+ "content_change_score": 3,
131
+ "style_change_score": 0
132
+ },
133
+ {
134
+ "time_s": 4.5,
135
+ "content_change_score": 3,
136
+ "style_change_score": 0
137
+ },
138
+ {
139
+ "time_s": 4.95,
140
+ "content_change_score": 3,
141
+ "style_change_score": 0
142
+ }
143
+ ]
144
+ },
145
+ "gpu0/KERNEL_EQUIVALENCE.json": {
146
+ "status": "passed",
147
+ "candidate_id": "C001",
148
+ "scores_exact": true,
149
+ "comparisons": 10,
150
+ "baseline": [
151
+ {
152
+ "time_s": 0.5,
153
+ "content_change_score": 0,
154
+ "style_change_score": 0
155
+ },
156
+ {
157
+ "time_s": 1.0,
158
+ "content_change_score": 1,
159
+ "style_change_score": 0
160
+ },
161
+ {
162
+ "time_s": 1.5,
163
+ "content_change_score": 2,
164
+ "style_change_score": 0
165
+ },
166
+ {
167
+ "time_s": 2.0,
168
+ "content_change_score": 3,
169
+ "style_change_score": 0
170
+ },
171
+ {
172
+ "time_s": 2.5,
173
+ "content_change_score": 3,
174
+ "style_change_score": 0
175
+ },
176
+ {
177
+ "time_s": 3.0,
178
+ "content_change_score": 3,
179
+ "style_change_score": 0
180
+ },
181
+ {
182
+ "time_s": 3.5,
183
+ "content_change_score": 3,
184
+ "style_change_score": 0
185
+ },
186
+ {
187
+ "time_s": 4.0,
188
+ "content_change_score": 3,
189
+ "style_change_score": 0
190
+ },
191
+ {
192
+ "time_s": 4.5,
193
+ "content_change_score": 3,
194
+ "style_change_score": 0
195
+ },
196
+ {
197
+ "time_s": 4.95,
198
+ "content_change_score": 3,
199
+ "style_change_score": 0
200
+ }
201
+ ],
202
+ "accelerated": [
203
+ {
204
+ "time_s": 0.5,
205
+ "content_change_score": 0,
206
+ "style_change_score": 0
207
+ },
208
+ {
209
+ "time_s": 1.0,
210
+ "content_change_score": 1,
211
+ "style_change_score": 0
212
+ },
213
+ {
214
+ "time_s": 1.5,
215
+ "content_change_score": 2,
216
+ "style_change_score": 0
217
+ },
218
+ {
219
+ "time_s": 2.0,
220
+ "content_change_score": 3,
221
+ "style_change_score": 0
222
+ },
223
+ {
224
+ "time_s": 2.5,
225
+ "content_change_score": 3,
226
+ "style_change_score": 0
227
+ },
228
+ {
229
+ "time_s": 3.0,
230
+ "content_change_score": 3,
231
+ "style_change_score": 0
232
+ },
233
+ {
234
+ "time_s": 3.5,
235
+ "content_change_score": 3,
236
+ "style_change_score": 0
237
+ },
238
+ {
239
+ "time_s": 4.0,
240
+ "content_change_score": 3,
241
+ "style_change_score": 0
242
+ },
243
+ {
244
+ "time_s": 4.5,
245
+ "content_change_score": 3,
246
+ "style_change_score": 0
247
+ },
248
+ {
249
+ "time_s": 4.95,
250
+ "content_change_score": 3,
251
+ "style_change_score": 0
252
+ }
253
+ ]
254
+ },
255
+ "gpu1/VISION_PATCH_EQUIVALENCE.json": {
256
+ "status": "passed",
257
+ "patches": 16,
258
+ "max_absolute_difference": 0.015625,
259
+ "bit_exact": false,
260
+ "rtol": 0.01,
261
+ "atol": 0.01,
262
+ "identity": "Kernel and stride equal full patch dimensions; Conv3d has exactly one spatial output per flattened patch, equal to the same weight matrix and bias"
263
+ }
264
+ },
265
+ "backend": "Two H100 GPUs, unchanged NF4 checkpoint,prompt and source images; official FLA kernels and mathematically equivalent matrix vision patch projection. BF16 accumulation order is not bit invariant; per-frame discrete scores can change at boundaries. Every final video uses this same backend.",
266
+ "worker_source_sha256": "bbe83445fa8a4cb2b0707904127361b0f1067c1087b1edf6ec51ce0b87fd84d6"
267
+ }