Hugging Face's logo

Datasets:
callensxavier
/
cy-sieve-attention-benchmark

Modalities:
Tabular
Text
Formats:
csv
Size:
< 1K
Tags:
attention
positional-encoding
benchmark
negative-result
calabi-yau
number-theory
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
callensxavier commited on
Commit
7fa16bf
·
verified ·
1 Parent(s): 3c6398c

Upload raw/run3.log with huggingface_hub

Browse files
Files changed (1) hide show
  1. raw/run3.log +189 -0
raw/run3.log ADDED
@@ -0,0 +1,189 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ==========================================
2
+ CY-Sieve GPU phase — startup Sun Jun 21 20:50:40 UTC 2026
3
+ ==========================================
4
+ Sun Jun 21 20:50:40 2026
5
+ +-----------------------------------------------------------------------------------------+
6
+ | NVIDIA-SMI 580.159.03 Driver Version: 580.159.03 CUDA Version: 13.0 |
7
+ +-----------------------------------------+------------------------+----------------------+
8
+ | GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC |
9
+ | Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. |
10
+ | | | MIG M. |
11
+ |=========================================+========================+======================|
12
+ | 0 NVIDIA L4 On | 00000000:00:03.0 Off | 0 |
13
+ | N/A 48C P0 28W / 72W | 0MiB / 23034MiB | 0% Default |
14
+ | | | N/A |
15
+ +-----------------------------------------+------------------------+----------------------+
16
+
17
+ +-----------------------------------------------------------------------------------------+
18
+ | Processes: |
19
+ | GPU GI CI PID Type Process name GPU Memory |
20
+ | ID ID Usage |
21
+ |=========================================================================================|
22
+ | No running processes found |
23
+ +-----------------------------------------------------------------------------------------+
24
+ WARNING: Running pip as the 'root' user can result in broken permissions and conflicting behaviour with the system package manager. It is recommended to use a virtual environment instead: https://pip.pypa.io/warnings/venv
25
+ WARNING: Running pip as the 'root' user can result in broken permissions and conflicting behaviour with the system package manager. It is recommended to use a virtual environment instead: https://pip.pypa.io/warnings/venv
26
+ deps OK: datasets 5.0.0
27
+ PyTorch 2.9.1+cu129, CUDA 12.9
28
+ GPU: NVIDIA L4
29
+ VRAM: 23.7 GB
30
+ Triton 3.5.1
31
+ Cloning into 'Mirror-Map-Sieve'...
32
+
33
+ Starting CY-Sieve GPU phase at Sun Jun 21 20:51:07 UTC 2026...
34
+ Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads.
35
+ ==========================================================================
36
+ CY-Sieve §5 QUALITY GATE (from-scratch) — preset=l4 on cuda
37
+ arch d_model=512 L=8 H=8 ctx=512 steps=6000
38
+ ==========================================================================
39
+
40
+
41
+
42
+ corpus: Salesforce/wikitext:wikitext-2-raw-v1 | train 2,004,085 bytes | val 250,345 bytes
43
+
44
+ [learned]
45
+ learned step 1/6000 loss 327.8488
46
+ learned step 1200/6000 loss 2.5287
47
+ learned step 2400/6000 loss 2.0594
48
+ learned step 3600/6000 loss 1.6516
49
+ learned step 4800/6000 loss 1.3834
50
+ learned step 6000/6000 loss 1.1621
51
+ learned val ppl @ L=512 (train): 4.218
52
+ learned val ppl @ L=1024 (2x-extrap): 12.100
53
+ learned val ppl @ L=2048 (4x-extrap): 20.822
54
+
55
+ [alibi]
56
+ alibi step 1/6000 loss 424.0075
57
+ alibi step 1200/6000 loss 1.6411
58
+ alibi step 2400/6000 loss 1.2131
59
+ alibi step 3600/6000 loss 1.0217
60
+ alibi step 4800/6000 loss 0.7236
61
+ alibi step 6000/6000 loss 0.4733
62
+ alibi val ppl @ L=512 (train): 10.739
63
+ alibi val ppl @ L=1024 (2x-extrap): 11.730
64
+ alibi val ppl @ L=2048 (4x-extrap): 11.352
65
+
66
+ [sliding]
67
+ sliding step 1/6000 loss 429.6556
68
+ sliding step 1200/6000 loss 2.5268
69
+ sliding step 2400/6000 loss 2.4214
70
+ sliding step 3600/6000 loss 2.1249
71
+ sliding step 4800/6000 loss 1.6317
72
+ sliding step 6000/6000 loss 1.3118
73
+ sliding val ppl @ L=512 (train): 4.995
74
+ sliding val ppl @ L=1024 (2x-extrap): 5.074
75
+ sliding val ppl @ L=2048 (4x-extrap): 5.027
76
+
77
+ [cy_sieve / per-head tau ladder]
78
+ cy_sieve step 1/6000 loss 425.0591
79
+ cy_sieve step 1200/6000 loss 1.6198
80
+ cy_sieve step 2400/6000 loss 1.2219
81
+ cy_sieve step 3600/6000 loss 1.0458
82
+ cy_sieve step 4800/6000 loss 0.7074
83
+ cy_sieve step 6000/6000 loss 0.4682
84
+ cy_sieve val ppl @ L=512 (train): 11.327
85
+ cy_sieve val ppl @ L=1024 (2x-extrap): 12.311
86
+ cy_sieve val ppl @ L=2048 (4x-extrap): 12.049
87
+
88
+ [cy_sieve / tau=20.0]
89
+ cy_sieve step 1/6000 loss 421.5372
90
+ cy_sieve step 1200/6000 loss 1.6310
91
+ cy_sieve step 2400/6000 loss 1.2126
92
+ cy_sieve step 3600/6000 loss 1.0404
93
+ cy_sieve step 4800/6000 loss 0.6793
94
+ cy_sieve step 6000/6000 loss 0.4282
95
+ cy_sieve val ppl @ L=512 (train): 16.019
96
+ cy_sieve val ppl @ L=1024 (2x-extrap): 16.810
97
+ cy_sieve val ppl @ L=2048 (4x-extrap): 16.489
98
+
99
+ [cy_sieve / tau=128.0]
100
+ cy_sieve step 1/6000 loss 429.1540
101
+ cy_sieve step 1200/6000 loss 2.3809
102
+ cy_sieve step 2400/6000 loss 1.6042
103
+ cy_sieve step 3600/6000 loss 1.3417
104
+ cy_sieve step 4800/6000 loss 1.0884
105
+ cy_sieve step 6000/6000 loss 0.8346
106
+ cy_sieve val ppl @ L=512 (train): 6.796
107
+ cy_sieve val ppl @ L=1024 (2x-extrap): 7.120
108
+ cy_sieve val ppl @ L=2048 (4x-extrap): 7.005
109
+
110
+ [cy_sieve / tau=512.0]
111
+ cy_sieve step 1/6000 loss 429.9576
112
+ cy_sieve step 1200/6000 loss 2.4574
113
+ cy_sieve step 2400/6000 loss 2.3402
114
+ cy_sieve step 3600/6000 loss 1.9214
115
+ cy_sieve step 4800/6000 loss 1.4814
116
+ cy_sieve step 6000/6000 loss 1.2002
117
+ cy_sieve val ppl @ L=512 (train): 4.646
118
+ cy_sieve val ppl @ L=1024 (2x-extrap): 6.083
119
+ cy_sieve val ppl @ L=2048 (4x-extrap): 10.621
120
+
121
+ ==========================================================================
122
+ VERDICT: KILL (negative result) — best CY-Sieve 4.646 (tau=512.0) vs best baseline 4.218: +10.15%
123
+ results -> /tmp/Mirror-Map-Sieve/4_ai_hardware_attention/cy_sieve_quality_results.json
124
+ ==========================================================================
125
+ ==========================================================================
126
+ CY-Sieve §6 PERF — NVIDIA L4 (24 GB) | D=64, tau=128.0
127
+ (speed is only meaningful alongside the §5 quality verdict — T6.3)
128
+ ==========================================================================
129
+ L=1024 cy_sieve= 0.055ms dense= 0.054ms table_bias= 0.060ms bias_HBM= 4096B vs 2097152B
130
+ L=2048 cy_sieve= 0.105ms dense= 0.053ms table_bias= 0.116ms bias_HBM= 8192B vs 8388608B
131
+ L=4096 cy_sieve= 0.260ms dense= 0.060ms table_bias= 0.268ms bias_HBM= 16384B vs 33554432B
132
+ L=8192 cy_sieve= 1.084ms dense= 0.288ms table_bias= 0.805ms bias_HBM= 32768B vs 134217728B
133
+ L=16384 cy_sieve= 4.156ms dense= 1.016ms table_bias= 3.938ms bias_HBM= 65536B vs 536870912B
134
+ [skip] L=32768: materialized-table baseline exceeds VRAM budget (need ~8.6 GB peak); CY-Sieve path still timed.
135
+ L=32768 cy_sieve= 15.283ms dense= 2.513ms table_bias= n/a ms bias_HBM= 131072B vs n/aB
136
+
137
+ results -> /tmp/Mirror-Map-Sieve/4_ai_hardware_attention/cy_sieve_perf_results.json
138
+ ==========================================================================
139
+ CY-Sieve GPU PHASE — tests.md §4 / §5 / §6
140
+ GPU: {'cuda': True, 'name': 'NVIDIA L4', 'vram_gb': 23.7, 'torch': '2.9.1+cu129', 'cuda_version': '12.9'}
141
+ ==========================================================================
142
+
143
+ ==========================================================================
144
+ §4 — Triton kernel <-> CPU reference parity
145
+ ==========================================================================
146
+ ============================= test session starts ==============================
147
+ platform linux -- Python 3.10.12, pytest-9.1.1, pluggy-1.6.0 -- /usr/bin/python3
148
+ cachedir: .pytest_cache
149
+ rootdir: /tmp/Mirror-Map-Sieve/4_ai_hardware_attention
150
+ plugins: anyio-4.14.0
151
+ collecting ... collected 4 items
152
+
153
+ test_cy_sieve_triton.py::test_t4_1_triton_matches_reference[128-64-20.0] PASSED [ 25%]
154
+ test_cy_sieve_triton.py::test_t4_1_triton_matches_reference[256-64-128.0] PASSED [ 50%]
155
+ test_cy_sieve_triton.py::test_t4_1_triton_matches_reference[256-128-480.0] PASSED [ 75%]
156
+ test_cy_sieve_triton.py::test_t4_1_smoke PASSED [100%]
157
+
158
+ ============================== 4 passed in 6.02s ===============================
159
+
160
+
161
+ ==========================================================================
162
+ §5 — QUALITY GATE (from-scratch training)
163
+ ==========================================================================
164
+
165
+ ==========================================================================
166
+ §6 — Performance (Triton kernel)
167
+ ==========================================================================
168
+
169
+ ==========================================================================
170
+ GPU PHASE COMPLETE
171
+ §4 parity : PASS (============================== 4 passed in 6.02s ===============================)
172
+ §5 quality: KILL (negative result) — best CY-Sieve 4.646 (tau=512.0) vs best baseline 4.218: +10.15%
173
+ §6 perf : L=32768 cy_sieve=15.2835ms dense=2.5128ms bias-HBM reduction Nonex
174
+ -> /tmp/gpu_phase_results_20260621_205040.json (13347.4s)
175
+ ==========================================================================
176
+
177
+ GPU phase exit code: 0 at Mon Jun 22 00:33:35 UTC 2026
178
+ Copying file:///tmp/gpu_phase_results_20260621_205040.json [Content-Type=application/json]...
179
+ / [0 files][ 0.0 B/ 5.1 KiB]
180
+ Operation completed over 1 objects/5.1 KiB.
181
+ ✅ uploaded gpu_phase_results_20260621_205040.json
182
+ Copying file:///tmp/Mirror-Map-Sieve/4_ai_hardware_attention/cy_sieve_quality_results.json [Content-Type=application/json]...
183
+ / [0 files][ 0.0 B/ 2.2 KiB]
184
+ Operation completed over 1 objects/2.2 KiB.
185
+ ✅ uploaded cy_sieve_quality_results.json
186
+ Copying file:///tmp/Mirror-Map-Sieve/4_ai_hardware_attention/cy_sieve_perf_results.json [Content-Type=application/json]...
187
+ / [0 files][ 0.0 B/ 1.9 KiB]
188
+ Operation completed over 1 objects/1.9 KiB.
189
+ ✅ uploaded cy_sieve_perf_results.json