SPB Git

spb/wp33_uqo

Public MIT

Gate G: bootstrap statistical audit (table in paper), manifest generator

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Simon-Pierre Boucher committed (Nov 1, 2026) parent 068f50f Browse files

7 changed files +1,646 −2

added artifacts/manifest.json +965 −0
@@ -0,0 +1,965 @@
1+{
2+ "generated": "2026-11-01T01:24:31",
3+ "git": "068f50f77c00921163f84652a4dccc6f84a4f554",
4+ "models": {
5+ "olmoe-1b-7b-0125-4bit": "mlx-community/OLMoE-1B-7B-0125-Instruct-4bit@e23844197887b031e7ddddbb0b8959c5a6853a7b",
6+ "qwen3-30b-a3b-2507-4bit": "mlx-community/Qwen3-30B-A3B-Instruct-2507-4bit@e9675aa3ca5f900ccef55267914466d55ab325fa"
7+ },
8+ "files": [
9+ {
10+ "path": "artifacts/environment.json",
11+ "bytes": 17537,
12+ "sha256": "e05e02401355ab3a3500c181264a4a4ac038c0bef38fe8718bcc6099db0b9498"
13+ },
14+ {
15+ "path": "benchmarks/processed/audit.json",
16+ "bytes": 12184,
17+ "sha256": "1bfa8337208bfa80c4dba5f9ced382e7ec4177311106da60f00754358720d874"
18+ },
19+ {
20+ "path": "benchmarks/processed/gateC/sim_olmoe_laptop_ssd_v1.json",
21+ "bytes": 123933,
22+ "sha256": "d7188c80daf3db38455e330c7e414c9ce0056289e9858a2403f94f715f5b7464"
23+ },
24+ {
25+ "path": "benchmarks/processed/gateC/sim_olmoe_v2_ssd043_m8.json",
26+ "bytes": 60788,
27+ "sha256": "d6e35cf4100ec34e484f1267770530d5f4c4e1a8764171f7fde6a508dde4527d"
28+ },
29+ {
30+ "path": "benchmarks/processed/gateC/sim_olmoe_v2_ssd043_m8_probe_fixed.json",
31+ "bytes": 9131,
32+ "sha256": "c069f0df73cc261b82e58529b46b3c0aed6ae2b8665a6c78c6df61f364afea0c"
33+ },
34+ {
35+ "path": "benchmarks/raw/gateA/smoke/mlx_lm_generate_laptop_20261101T011129.txt",
36+ "bytes": 772,
37+ "sha256": "86d3b5a63f5d8a0c20a78bcaa337a4d329c89063f26304e4d777bf057a626045"
38+ },
39+ {
40+ "path": "benchmarks/raw/gateA/ssd/laptop-M5Max.json",
41+ "bytes": 995,
42+ "sha256": "93116e6ce7c0c2039525b8ee731edfddcb2ef936c40547298c1b3e3131ec864e"
43+ },
44+ {
45+ "path": "benchmarks/raw/gateB/capacity_sweep_laptop/sweep.jsonl",
46+ "bytes": 183957,
47+ "sha256": "3fdc16b869407201d8544e4ee4cfe775652ead132f9539f4465f23b80948f956"
48+ },
49+ {
50+ "path": "benchmarks/raw/gateB/design/fastpath_profile_laptop.txt",
51+ "bytes": 331,
52+ "sha256": "c76ecd5b8d6ec170999d37590124597141bb33e5a591fc552373dacc5f0a5519"
53+ },
54+ {
55+ "path": "benchmarks/raw/gateB/design/slot_update_microbench_laptop.txt",
56+ "bytes": 365,
57+ "sha256": "bfd1f99ab8b777b2b6a3bac59fe5f0e55fa6abcc4741397d17480a6464321d4f"
58+ },
59+ {
60+ "path": "benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl",
61+ "bytes": 120437,
62+ "sha256": "9c2b80b0c6042ad3124142d332bbae73d17148fdc46b4384195726afb7862f3c"
63+ },
64+ {
65+ "path": "benchmarks/raw/gateC/cold_nvme_laptop/quick_v2.jsonl",
66+ "bytes": 3924,
67+ "sha256": "e8d892ce579de09e6828d8821894c40a2f882bda2ff433ac24a224c90ddc933e"
68+ },
69+ {
70+ "path": "benchmarks/raw/gateC/cold_nvme_laptop/quick_v3.jsonl",
71+ "bytes": 3964,
72+ "sha256": "52b022ff9ae01490c987bd62ac8382c05795949a13eddf1246267762682cd294"
73+ },
74+ {
75+ "path": "benchmarks/raw/gateC/cold_nvme_laptop/sweep.jsonl",
76+ "bytes": 450009,
77+ "sha256": "4c0a4268388dc88929b05df49ff67a9e5fa7d4e5df908b0b8e4d76cc583d259c"
78+ },
79+ {
80+ "path": "benchmarks/raw/gateC/cold_nvme_laptop/sweep_v1_9preads_serial.jsonl",
81+ "bytes": 29919,
82+ "sha256": "2fff605ec66aae9672c4fe37b26fc019765392765d180b49b2935999403dbc31"
83+ },
84+ {
85+ "path": "benchmarks/raw/gateC/design/miss_cost_breakdown_laptop.txt",
86+ "bytes": 547,
87+ "sha256": "94efb69aefe9ae86e9c59bcd96836f818d9d658c16ff85b4c46ab0f081572f00"
88+ },
89+ {
90+ "path": "benchmarks/raw/gateC/prefetch_online_laptop/smoke.jsonl",
91+ "bytes": 72158,
92+ "sha256": "e01beefef6a42567a11cd158080cdf4f7e56e67eafa91cd00300450cb37f679f"
93+ },
94+ {
95+ "path": "benchmarks/raw/gateC/prefetch_online_laptop/smoke_lazyinstall.jsonl",
96+ "bytes": 48055,
97+ "sha256": "b33f0d34e8c6172b408b1d2f864e50e6a61199ee4a8efd412f26402756bdc5ac"
98+ },
99+ {
100+ "path": "benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl",
101+ "bytes": 80611,
102+ "sha256": "4f509b6faa8722d275384181cf40583caa004edc890cecd9582566303068767b"
103+ },
104+ {
105+ "path": "benchmarks/raw/gateC/qwen3_m3u96a/baseline_mlxlm.log",
106+ "bytes": 1945,
107+ "sha256": "4df7656a9a81d354b46464fda6bd80ddb93986273305402ce52a0a4aee70402c"
108+ },
109+ {
110+ "path": "benchmarks/raw/gateC/qwen3_m3u96a/baseline_titan128.log",
111+ "bytes": 2425,
112+ "sha256": "50788b364c260b944aa93482c0a706177d707300f50ddc10c9ef6732baa4d413"
113+ },
114+ {
115+ "path": "benchmarks/raw/gateC/traces_olmoe_laptop/code_fit_0.npz",
116+ "bytes": 525273,
117+ "sha256": "90267d30b9c9c6854dc6430df229e6aa13ef505a167a1a9674cd8d37f7de0314"
118+ },
119+ {
120+ "path": "benchmarks/raw/gateC/traces_olmoe_laptop/code_fit_1.npz",
121+ "bytes": 546094,
122+ "sha256": "dfd9f17d01541206e60def9371d708a11772c437d7233c17a7760260c4a68e06"
123+ },
124+ {
125+ "path": "benchmarks/raw/gateC/traces_olmoe_laptop/code_fit_2.npz",
126+ "bytes": 529494,
127+ "sha256": "31bf7227a1936b292359c7ec44a6375a279672b57360711134ee956ea38078db"
128+ },
129+ {
130+ "path": "benchmarks/raw/gateC/traces_olmoe_laptop/code_heldout_0.npz",
131+ "bytes": 533705,
132+ "sha256": "5783a1755a51899de6a99b8b831a480d0730e15c9f1788a751b1d62583856ad0"
133+ },
134+ {
135+ "path": "benchmarks/raw/gateC/traces_olmoe_laptop/code_heldout_1.npz",
136+ "bytes": 523765,
137+ "sha256": "a46f9f759d1fb5368f9735c95064b85b8fcb2ad8dd0283e7ec44312f07d0ab7a"
138+ },
139+ {
140+ "path": "benchmarks/raw/gateC/traces_olmoe_laptop/general_fit_0.npz",
141+ "bytes": 214468,
142+ "sha256": "7066dd1e4c56e6ba7c0352dab45f80724d79fb9d553515e62f3c1b204ca183b5"
143+ },
144+ {
145+ "path": "benchmarks/raw/gateC/traces_olmoe_laptop/general_fit_1.npz",
146+ "bytes": 529130,
147+ "sha256": "710053d323754cc5aee1ee4672e3995cf5f14d8425ee0b0edbff7a2bef62065d"
148+ },
149+ {
150+ "path": "benchmarks/raw/gateC/traces_olmoe_laptop/general_fit_2.npz",
151+ "bytes": 465070,
152+ "sha256": "ac5a47c87592c3a9def03657dd350491d8bae8c20aa12d46e4f0d8117d55cd52"
153+ },
154+ {
155+ "path": "benchmarks/raw/gateC/traces_olmoe_laptop/general_heldout_0.npz",
156+ "bytes": 520565,
157+ "sha256": "0ff645b4745490dbb46b00d678f4fbf5e5824be341f37520ff2d6981ef3f6ce5"
158+ },
159+ {
160+ "path": "benchmarks/raw/gateC/traces_olmoe_laptop/general_heldout_1.npz",
161+ "bytes": 513497,
162+ "sha256": "00c51955b25b4164e204d6405f3f9e6ae2d4661811781d08875318616d66ffe3"
163+ },
164+ {
165+ "path": "benchmarks/raw/gateC/traces_olmoe_laptop/index.json",
166+ "bytes": 8051,
167+ "sha256": "50367d2c17613fc3670db61e4d80f502514d1793ce85fc58265ee1f80cf07180"
168+ },
169+ {
170+ "path": "benchmarks/raw/gateC/traces_olmoe_laptop/multilingual_fit_0.npz",
171+ "bytes": 135710,
172+ "sha256": "9d921d7c22b4c595d67ee68b8480e6beca041906387948b523c0018afe9b7a56"
173+ },
174+ {
175+ "path": "benchmarks/raw/gateC/traces_olmoe_laptop/multilingual_fit_1.npz",
176+ "bytes": 565691,
177+ "sha256": "1f418dcb08c137ed8d5fdab2975afd440d21aab2edbd77552fd385105412961a"
178+ },
179+ {
180+ "path": "benchmarks/raw/gateC/traces_olmoe_laptop/multilingual_heldout_0.npz",
181+ "bytes": 76269,
182+ "sha256": "2e136aeb8cee3daab7dc8cdde0ef9e134f067e7b15bd2da76d17342f1be5d19c"
183+ },
184+ {
185+ "path": "benchmarks/raw/gateC/traces_olmoe_laptop/reasoning_fit_0.npz",
186+ "bytes": 129575,
187+ "sha256": "11073c5039bfab061cdedf64b0d078eef228e47b1f4e43eb25d784e949a1905a"
188+ },
189+ {
190+ "path": "benchmarks/raw/gateC/traces_olmoe_laptop/reasoning_fit_1.npz",
191+ "bytes": 249978,
192+ "sha256": "81bb7bc96f9950d33fc598d445a92b0fecddc09e72c58c6a4c84941561971968"
193+ },
194+ {
195+ "path": "benchmarks/raw/gateC/traces_olmoe_laptop/reasoning_fit_2.npz",
196+ "bytes": 196539,
197+ "sha256": "6d559291eb0be3054a5d4b9c261a75f4fbb591ea1b637a0a9eb8bc2425860f49"
198+ },
199+ {
200+ "path": "benchmarks/raw/gateC/traces_olmoe_laptop/reasoning_heldout_0.npz",
201+ "bytes": 346389,
202+ "sha256": "4d17db9bb112f9e236739551f91009f19e19f4c6c76c74fcb546c1d6b539e294"
203+ },
204+ {
205+ "path": "benchmarks/raw/gateC/traces_olmoe_laptop/reasoning_heldout_1.npz",
206+ "bytes": 251462,
207+ "sha256": "ead5a58228140725dea6259dff1c75e8852913dd453157d2fb03c471f46c479c"
208+ },
209+ {
210+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/code_fit_0.npz",
211+ "bytes": 656354,
212+ "sha256": "5081b895aec17144fa5e2658c85e057715cc15214633fefeb3f23cd01982b38e"
213+ },
214+ {
215+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/code_fit_1.npz",
216+ "bytes": 788804,
217+ "sha256": "aaba7275c8cda68efdeae893e8483a33cbe58fa1a98652e2c26b13da2e9f0abb"
218+ },
219+ {
220+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/code_fit_2.npz",
221+ "bytes": 771988,
222+ "sha256": "038cb3423715ca6bd992208bf5261c83c4180dc153e7cfb22c50570e80eab65c"
223+ },
224+ {
225+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/code_fit_3.npz",
226+ "bytes": 770796,
227+ "sha256": "b9923b51b23860def2d8a32217b3eea67cba27e8edb364d44490ad4084466bd2"
228+ },
229+ {
230+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/code_fit_4.npz",
231+ "bytes": 718360,
232+ "sha256": "9840400ce2e85170e258cabd0d2bc4d53eedc09628d08e0c9ffaf367135e34a0"
233+ },
234+ {
235+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/code_fit_5.npz",
236+ "bytes": 789657,
237+ "sha256": "33089502109cc23d0f4aaf3c3d3d42957517ab49ebd4f96c109af7c763fee9aa"
238+ },
239+ {
240+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/code_fit_6.npz",
241+ "bytes": 764331,
242+ "sha256": "e5c2ed9a15a211f3fda302164c0d4fb07220ce4867f8e0d455583522ce1a4c2e"
243+ },
244+ {
245+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/code_fit_7.npz",
246+ "bytes": 704696,
247+ "sha256": "084c843363fec70f0585caef9bfe1e3ef9419aba5e5dff2e3f070b8a3cb050e1"
248+ },
249+ {
250+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/code_fit_8.npz",
251+ "bytes": 772710,
252+ "sha256": "677ae00bb2ce15578dc2d34909162f76fed9ae2d56b1edea68579c3400773595"
253+ },
254+ {
255+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/code_fit_9.npz",
256+ "bytes": 564005,
257+ "sha256": "508d0e337e86d08f36fa77669e7c826572426f54125162b240bcf88b3e9eced2"
258+ },
259+ {
260+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/code_heldout_0.npz",
261+ "bytes": 775944,
262+ "sha256": "ea96518772ab286ee91469181f94cdfbb409c4b09d2eab1e7fd2b9ac2fb6e4b4"
263+ },
264+ {
265+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/code_heldout_1.npz",
266+ "bytes": 731455,
267+ "sha256": "b9127ab21af3fb65cd6b4f026de155e6e1dd9334108999b0f643f6e681c6c942"
268+ },
269+ {
270+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/code_heldout_2.npz",
271+ "bytes": 766485,
272+ "sha256": "1a45deb2e2d83825a3318f23bbdcce0e6fda6bda0e9527593524c4232e347735"
273+ },
274+ {
275+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/code_heldout_3.npz",
276+ "bytes": 767985,
277+ "sha256": "56c7f815093ea80ff9272bfff80dcef07c1475da32d0601f6087c4291373a39c"
278+ },
279+ {
280+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/drift_fit_0.npz",
281+ "bytes": 759367,
282+ "sha256": "80b608063a4e7a0e759345fe1ecfbda9f1cb316674965597c0fb35a48fef9a34"
283+ },
284+ {
285+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/drift_fit_1.npz",
286+ "bytes": 786411,
287+ "sha256": "91f136e58dfb01b9ee12ea519e2339bf59225f5b94e78425ac648be02305a292"
288+ },
289+ {
290+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/drift_fit_2.npz",
291+ "bytes": 822551,
292+ "sha256": "f565d674205e7235ed8b9043391d8d7a0c249c90d1dd9f39999ee288191f76f3"
293+ },
294+ {
295+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/drift_heldout_0.npz",
296+ "bytes": 777557,
297+ "sha256": "60adb0d387dd74f018437839a9b2a7e32f4b1b2b74d00edb710f61a68041e33c"
298+ },
299+ {
300+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/drift_heldout_1.npz",
301+ "bytes": 189224,
302+ "sha256": "3df66e8b80c9179284a6cb8fc608c5a7d5db7b4b24f48045714a57021fc227dc"
303+ },
304+ {
305+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/general_fit_0.npz",
306+ "bytes": 214468,
307+ "sha256": "7066dd1e4c56e6ba7c0352dab45f80724d79fb9d553515e62f3c1b204ca183b5"
308+ },
309+ {
310+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/general_fit_1.npz",
311+ "bytes": 559492,
312+ "sha256": "60a139c0aa5aa69df1e978925286ae744c2ce758aed4ab2fee0d3168b2bf3ee2"
313+ },
314+ {
315+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/general_fit_2.npz",
316+ "bytes": 465070,
317+ "sha256": "ac5a47c87592c3a9def03657dd350491d8bae8c20aa12d46e4f0d8117d55cd52"
318+ },
319+ {
320+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/general_fit_3.npz",
321+ "bytes": 744436,
322+ "sha256": "379532fa0b8e567a4d92d1eaab588e3de0e1ed948bd92ec991d6cb07761d69b9"
323+ },
324+ {
325+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/general_fit_4.npz",
326+ "bytes": 586886,
327+ "sha256": "d3c85a2dd425f88468f8a848327369901bf981cf64305473f78a5d68d08e8551"
328+ },
329+ {
330+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/general_fit_5.npz",
331+ "bytes": 755527,
332+ "sha256": "1f192d34b94945a544fb248b9017161750517e5ffc5bcad33e472cad46590efa"
333+ },
334+ {
335+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/general_fit_6.npz",
336+ "bytes": 763897,
337+ "sha256": "8e027395b485605921f5642b23c39de6e3d617dc5449797a7f2b09046c083255"
338+ },
339+ {
340+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/general_fit_7.npz",
341+ "bytes": 486948,
342+ "sha256": "b7f96e5cd07cdb0d22d8a89130442fceefbbf8bb462dcb5656e80c55d07ee770"
343+ },
344+ {
345+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/general_fit_8.npz",
346+ "bytes": 418245,
347+ "sha256": "e0224f3574bab1a8a539e7d4e5c921bdef260e9d33911d590d7f0d8dae2d5635"
348+ },
349+ {
350+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/general_fit_9.npz",
351+ "bytes": 377742,
352+ "sha256": "25ad255572ac404c83240d4c92ff4aa3c3f9fd195d1f18485afd0a32e31cf09a"
353+ },
354+ {
355+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/general_heldout_0.npz",
356+ "bytes": 624508,
357+ "sha256": "44a96860fefb1365820ef8901c45b517dc259836c7a1a077eda925dc17606eb6"
358+ },
359+ {
360+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/general_heldout_1.npz",
361+ "bytes": 753336,
362+ "sha256": "0c9e0b87b0c83032e13fdcf440b0bf7d4a1b16cdeef73fa30f7287c77fe62375"
363+ },
364+ {
365+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/general_heldout_2.npz",
366+ "bytes": 525021,
367+ "sha256": "82962233c031e7a6b986c0518e222e4784da0290c532183e28c3a67f7bc23f53"
368+ },
369+ {
370+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/general_heldout_3.npz",
371+ "bytes": 763713,
372+ "sha256": "6f5e244ca12cb77076596ff880e35c6fe895d35abeff12fb43c9932377186104"
373+ },
374+ {
375+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/index.json",
376+ "bytes": 26533,
377+ "sha256": "cd8ad1d7dc06c8c330f6da219e5ce8f88a351332ed81f4db632607a112a71ad9"
378+ },
379+ {
380+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/longctx_fit_0.npz",
381+ "bytes": 972584,
382+ "sha256": "8ca45b55c34c9484e87a6dc89c5cf361ded13adf1078887ef41890dd8a0ea18c"
383+ },
384+ {
385+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/longctx_fit_1.npz",
386+ "bytes": 920455,
387+ "sha256": "7e3afe7e1ebb873b6ad34b9788864f6c89a284e78d135080b5108b24f73f838f"
388+ },
389+ {
390+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/longctx_fit_2.npz",
391+ "bytes": 904796,
392+ "sha256": "04d5be500de14ece655cf0c2051fae05249acb9e86762beb63e98838c9b4f255"
393+ },
394+ {
395+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/longctx_heldout_0.npz",
396+ "bytes": 925810,
397+ "sha256": "08d3cdc98c05f74d110240f7086b61fd386ca0e16a2e7760eec1881dd901452a"
398+ },
399+ {
400+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/multilingual_fit_0.npz",
401+ "bytes": 135710,
402+ "sha256": "9d921d7c22b4c595d67ee68b8480e6beca041906387948b523c0018afe9b7a56"
403+ },
404+ {
405+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/multilingual_fit_1.npz",
406+ "bytes": 701384,
407+ "sha256": "99f54a73e81cbe61f32ac85cb509bd6b098f0761f454f756076a63d7b52d4ee4"
408+ },
409+ {
410+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/multilingual_fit_2.npz",
411+ "bytes": 777501,
412+ "sha256": "9c10fc1b5a81517668896d06c9dbe3478ba0f2ef7de57384cfe11ccf5ebbd84f"
413+ },
414+ {
415+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/multilingual_fit_3.npz",
416+ "bytes": 506541,
417+ "sha256": "15c4fbae207491ecd5d9c50d7beff30d49e48c26fac56103e6b4587ead4d6fcb"
418+ },
419+ {
420+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/multilingual_fit_4.npz",
421+ "bytes": 792758,
422+ "sha256": "a42656c5c583aaeed7aea9ab14537cbd844ce4d5441353c9671d1dd0f4575cab"
423+ },
424+ {
425+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/multilingual_fit_5.npz",
426+ "bytes": 523552,
427+ "sha256": "ffaaf51002d108a30b611311925f04c94c327e16dc99e03b10aceeddbc24af7b"
428+ },
429+ {
430+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/multilingual_fit_6.npz",
431+ "bytes": 789213,
432+ "sha256": "e07fcedb92273e3f591c63d764517cd8c34ef764dd8ef3fd5c34a551a515ecf0"
433+ },
434+ {
435+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/multilingual_fit_7.npz",
436+ "bytes": 506011,
437+ "sha256": "563411725c51eac9b49d4f5f9251a362027448f858b717de28d5c7172806f00b"
438+ },
439+ {
440+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/multilingual_heldout_0.npz",
441+ "bytes": 76269,
442+ "sha256": "2e136aeb8cee3daab7dc8cdde0ef9e134f067e7b15bd2da76d17342f1be5d19c"
443+ },
444+ {
445+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/multilingual_heldout_1.npz",
446+ "bytes": 161420,
447+ "sha256": "63f8cab9c6ac952e2bc5d2c6ef9b1d5b621e07dfcbb14d1f71f360115431c40d"
448+ },
449+ {
450+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/multilingual_heldout_2.npz",
451+ "bytes": 778872,
452+ "sha256": "ecb4bd5f11a311905466be5a21bc9679be2122abe1918a23d089440e895652bc"
453+ },
454+ {
455+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/reasoning_fit_0.npz",
456+ "bytes": 129575,
457+ "sha256": "11073c5039bfab061cdedf64b0d078eef228e47b1f4e43eb25d784e949a1905a"
458+ },
459+ {
460+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/reasoning_fit_1.npz",
461+ "bytes": 249978,
462+ "sha256": "81bb7bc96f9950d33fc598d445a92b0fecddc09e72c58c6a4c84941561971968"
463+ },
464+ {
465+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/reasoning_fit_2.npz",
466+ "bytes": 196539,
467+ "sha256": "6d559291eb0be3054a5d4b9c261a75f4fbb591ea1b637a0a9eb8bc2425860f49"
468+ },
469+ {
470+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/reasoning_fit_3.npz",
471+ "bytes": 154398,
472+ "sha256": "e49f157944767abedbbf5e11e8a44658dc1af232ff5573adffe7d3633bcc5e18"
473+ },
474+ {
475+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/reasoning_fit_4.npz",
476+ "bytes": 205524,
477+ "sha256": "63fbb98584d09812b649202e230e83533ae7eb13aa10b9c846e690b089b2abf8"
478+ },
479+ {
480+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/reasoning_fit_5.npz",
481+ "bytes": 166122,
482+ "sha256": "425340ee9a334cda84f6a3f56605a6b6b555d2ead304a6b78e09632e2ea0d1f6"
483+ },
484+ {
485+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/reasoning_fit_6.npz",
486+ "bytes": 223204,
487+ "sha256": "5c16696d7a7b1bc4bfe4ab513561ad139121a3e10bb5b666a59a0eac07382bf9"
488+ },
489+ {
490+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/reasoning_fit_7.npz",
491+ "bytes": 171583,
492+ "sha256": "55bfbbe7b44255fcdfcb342726a864485520f30bffa7590bc60fcec9089caedc"
493+ },
494+ {
495+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/reasoning_fit_8.npz",
496+ "bytes": 160712,
497+ "sha256": "6605a3e42c1aa56e4fa8c1574b802f36039bf3b8c919d2967bc566400bb4b40b"
498+ },
499+ {
500+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/reasoning_fit_9.npz",
501+ "bytes": 181392,
502+ "sha256": "faf93bed507e97be69e4bc621763dda0a161597de77afe3bbe602253b66095e7"
503+ },
504+ {
505+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/reasoning_heldout_0.npz",
506+ "bytes": 346389,
507+ "sha256": "4d17db9bb112f9e236739551f91009f19e19f4c6c76c74fcb546c1d6b539e294"
508+ },
509+ {
510+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/reasoning_heldout_1.npz",
511+ "bytes": 251462,
512+ "sha256": "ead5a58228140725dea6259dff1c75e8852913dd453157d2fb03c471f46c479c"
513+ },
514+ {
515+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/reasoning_heldout_2.npz",
516+ "bytes": 185075,
517+ "sha256": "de5d8a3214f3abd1847282b6a88c274729eb23e2919cbf64a8f7c45c82f90da1"
518+ },
519+ {
520+ "path": "benchmarks/raw/gateC/traces_olmoe_v2_laptop/reasoning_heldout_3.npz",
521+ "bytes": 213567,
522+ "sha256": "5fa14d863660b17366ed9c28e50d7dd197bab9244cd8e36df6f55af070418ee0"
523+ },
524+ {
525+ "path": "benchmarks/raw/gateD/fused_upgate_laptop.jsonl",
526+ "bytes": 86913,
527+ "sha256": "0061d8c870b522f0c4070ba562cfe437ba655c05db105aecc994d67bc387a8bb"
528+ },
529+ {
530+ "path": "benchmarks/raw/gateE/remote_cost_breakdown_m3u96a_m3u96b.txt",
531+ "bytes": 655,
532+ "sha256": "dc4f40e8b782c4cd31bf9fd44a58aa59608d86cd24edf90ecf663e4bc6aecd9e"
533+ },
534+ {
535+ "path": "benchmarks/raw/gateE/run1.log",
536+ "bytes": 8356,
537+ "sha256": "88326472853330864c268710bd804491dc2412508c5621356e4fabbbc123c31f"
538+ },
539+ {
540+ "path": "benchmarks/raw/gateE/two_node_m3u96a_m3u96b.jsonl",
541+ "bytes": 153008,
542+ "sha256": "641405e60392332ed5661b6bbfbd974de723ae35d99d2cb3af6afcfef6ee29d0"
543+ },
544+ {
545+ "path": "benchmarks/raw/tests/distributed_localhost_olmoe_20261101T010211.json",
546+ "bytes": 1505,
547+ "sha256": "eb9ac1542ab445099df37fcc485c95af8106c46535b114426b8bd0d106c2f459"
548+ },
549+ {
550+ "path": "benchmarks/raw/tests/distributed_localhost_olmoe_20261101T010636.json",
551+ "bytes": 1504,
552+ "sha256": "b022f63d30bf9b837c7d40cfd256d8c5ddba5ea89bdbdb94084eeebec8c354ba"
553+ },
554+ {
555+ "path": "benchmarks/raw/tests/distributed_localhost_olmoe_20261101T010747.json",
556+ "bytes": 1504,
557+ "sha256": "913e2e610f21a4aa38713942fb2ebe726e6b93381ff95a0dbb427b14f804bea8"
558+ },
559+ {
560+ "path": "benchmarks/raw/tests/fused_upgate_olmoe_2026-11-01T010747.json",
561+ "bytes": 667,
562+ "sha256": "3435b57f045559d2cd8e2725934e9f2a5872dea0d4ca7b30aa9004bb7f7e5ae7"
563+ },
564+ {
565+ "path": "benchmarks/raw/tests/greedy_cap16_20261101T011813.json",
566+ "bytes": 702,
567+ "sha256": "c2007a57d10668d208cadd945685a93aba009472db28de8f65670200a02b51e4"
568+ },
569+ {
570+ "path": "benchmarks/raw/tests/greedy_cap16_20261101T012306.json",
571+ "bytes": 702,
572+ "sha256": "c2007a57d10668d208cadd945685a93aba009472db28de8f65670200a02b51e4"
573+ },
574+ {
575+ "path": "benchmarks/raw/tests/greedy_cap16_20261101T013103.json",
576+ "bytes": 702,
577+ "sha256": "c2007a57d10668d208cadd945685a93aba009472db28de8f65670200a02b51e4"
578+ },
579+ {
580+ "path": "benchmarks/raw/tests/greedy_cap16_20261101T013347.json",
581+ "bytes": 702,
582+ "sha256": "c2007a57d10668d208cadd945685a93aba009472db28de8f65670200a02b51e4"
583+ },
584+ {
585+ "path": "benchmarks/raw/tests/greedy_olmoe_cap16_20261101T010217.json",
586+ "bytes": 702,
587+ "sha256": "c2007a57d10668d208cadd945685a93aba009472db28de8f65670200a02b51e4"
588+ },
589+ {
590+ "path": "benchmarks/raw/tests/greedy_olmoe_cap16_20261101T010757.json",
591+ "bytes": 702,
592+ "sha256": "c2007a57d10668d208cadd945685a93aba009472db28de8f65670200a02b51e4"
593+ },
594+ {
595+ "path": "benchmarks/raw/tests/greedy_olmoe_cap16_20261101T014415.json",
596+ "bytes": 702,
597+ "sha256": "c2007a57d10668d208cadd945685a93aba009472db28de8f65670200a02b51e4"
598+ },
599+ {
600+ "path": "benchmarks/raw/tests/greedy_olmoe_cap16_20261101T014903.json",
601+ "bytes": 702,
602+ "sha256": "c2007a57d10668d208cadd945685a93aba009472db28de8f65670200a02b51e4"
603+ },
604+ {
605+ "path": "benchmarks/raw/tests/greedy_olmoe_cap16_20261101T015010.json",
606+ "bytes": 702,
607+ "sha256": "c2007a57d10668d208cadd945685a93aba009472db28de8f65670200a02b51e4"
608+ },
609+ {
610+ "path": "benchmarks/raw/tests/greedy_olmoe_cap16_20261101T015029.json",
611+ "bytes": 702,
612+ "sha256": "c2007a57d10668d208cadd945685a93aba009472db28de8f65670200a02b51e4"
613+ },
614+ {
615+ "path": "benchmarks/raw/tests/greedy_olmoe_cap16_20261101T015233.json",
616+ "bytes": 702,
617+ "sha256": "c2007a57d10668d208cadd945685a93aba009472db28de8f65670200a02b51e4"
618+ },
619+ {
620+ "path": "benchmarks/raw/tests/m3u96a/greedy_qwen3_moe_cap32_20261101T014433.json",
621+ "bytes": 770,
622+ "sha256": "332b59d76c129d714e334fed2be68a7497287f5ae52dfaa03545d1b5270d7e81"
623+ },
624+ {
625+ "path": "benchmarks/raw/tests/m3u96a/numerical_qwen3_moe_cap128_2026-11-01T014417.json",
626+ "bytes": 1399,
627+ "sha256": "b4148933a89063f8353d1e8eaf761b6ef4db6c6fbd9626ea4d05ed127a86a65a"
628+ },
629+ {
630+ "path": "benchmarks/raw/tests/m3u96a/numerical_qwen3_moe_cap16_2026-11-01T014425.json",
631+ "bytes": 1418,
632+ "sha256": "9b105addcdd512206f89406b32d186152ee7dd55c661b625a9142534da8025b7"
633+ },
634+ {
635+ "path": "benchmarks/raw/tests/m3u96a/numerical_qwen3_moe_cap32_2026-11-01T014422.json",
636+ "bytes": 1415,
637+ "sha256": "74ff28f6217685d3f351c309f9552a058febd914bdb72287889bf75f5a994455"
638+ },
639+ {
640+ "path": "benchmarks/raw/tests/m3u96a/numerical_qwen3_moe_cap64_2026-11-01T014420.json",
641+ "bytes": 1396,
642+ "sha256": "4407c8b71bbace6ee0f109c3406f9dd7aad2fd79ffed796cfb907430d76ab221"
643+ },
644+ {
645+ "path": "benchmarks/raw/tests/m3u96a/pytest_qwen3_m3u96a_20261101T014415.txt",
646+ "bytes": 110,
647+ "sha256": "7ca257be544f21c548f743d77183601f2f935f974b9e21ea51e4eba3b0f9e620"
648+ },
649+ {
650+ "path": "benchmarks/raw/tests/numerical_cap16_2026-11-01T011806.json",
651+ "bytes": 1217,
652+ "sha256": "630a98566f1f73413267579580d1b884298763c63d26938332065b59ee7430af"
653+ },
654+ {
655+ "path": "benchmarks/raw/tests/numerical_cap16_2026-11-01T012259.json",
656+ "bytes": 1218,
657+ "sha256": "616058446f9965335a8fd5c7108151102ad508fc4c1806ca7277f8e480814aa0"
658+ },
659+ {
660+ "path": "benchmarks/raw/tests/numerical_cap16_2026-11-01T013056.json",
661+ "bytes": 1217,
662+ "sha256": "c86c870c8eac0a14c7a442cd5f7c8bf9c6f388c229328c7ae0ad7951a9e10cb7"
663+ },
664+ {
665+ "path": "benchmarks/raw/tests/numerical_cap16_2026-11-01T013342.json",
666+ "bytes": 1219,
667+ "sha256": "7afe990af8de6d4ad75c0eafa3d7963bd355994b44dcaf4eb86653813df7d3c2"
668+ },
669+ {
670+ "path": "benchmarks/raw/tests/numerical_cap32_2026-11-01T011805.json",
671+ "bytes": 1196,
672+ "sha256": "eab6ae12ac67790c0894b8521c0680b71297b2bd536022f0e4e96ee4254fc67f"
673+ },
674+ {
675+ "path": "benchmarks/raw/tests/numerical_cap32_2026-11-01T012258.json",
676+ "bytes": 1195,
677+ "sha256": "b39c33f5f498fb6b213cafabd1cc899de3d732ff2c79212193bb6cee85797dd3"
678+ },
679+ {
680+ "path": "benchmarks/raw/tests/numerical_cap32_2026-11-01T013055.json",
681+ "bytes": 1194,
682+ "sha256": "968ae034d883bc0d51f37d44063be299148d8173b8d18183b00b1c213ac41b0c"
683+ },
684+ {
685+ "path": "benchmarks/raw/tests/numerical_cap32_2026-11-01T013341.json",
686+ "bytes": 1195,
687+ "sha256": "31818d37532afec50f4bafca77f64410e64c7120050524d20d5a83d422dcc447"
688+ },
689+ {
690+ "path": "benchmarks/raw/tests/numerical_cap64_2026-11-01T011726.json",
691+ "bytes": 1195,
692+ "sha256": "de7cce9a373c73b7d17d939d06e2d06274170be775231f477d78c3b834070dfe"
693+ },
694+ {
695+ "path": "benchmarks/raw/tests/numerical_cap64_2026-11-01T011753.json",
696+ "bytes": 1194,
697+ "sha256": "f78ec65ca4578c2a41fcc383693c15c143ca9ed82fe2788db3b19a9b67c21c7a"
698+ },
699+ {
700+ "path": "benchmarks/raw/tests/numerical_cap64_2026-11-01T011805.json",
701+ "bytes": 1194,
702+ "sha256": "60a927377b1449f957e2a06a825b35f526330e44037501a2d1816dbd9edc7ef4"
703+ },
704+ {
705+ "path": "benchmarks/raw/tests/numerical_cap64_2026-11-01T012257.json",
706+ "bytes": 1195,
707+ "sha256": "742fa9be90ebb23fd7a0cc1676a088cc261d9318b6a24ca73f7a2f16e00746a8"
708+ },
709+ {
710+ "path": "benchmarks/raw/tests/numerical_cap64_2026-11-01T013054.json",
711+ "bytes": 1193,
712+ "sha256": "a76c618af6e31d66422179a70d40b24d6bf1b379aa515157389080e67235de2f"
713+ },
714+ {
715+ "path": "benchmarks/raw/tests/numerical_cap64_2026-11-01T013341.json",
716+ "bytes": 1195,
717+ "sha256": "5ec2516c7948eeee31d12070e2f312db13c8cfe27a0be5b415a3ab4e0343aa27"
718+ },
719+ {
720+ "path": "benchmarks/raw/tests/numerical_cap8_2026-11-01T011808.json",
721+ "bytes": 1186,
722+ "sha256": "a0b2a138c2ceeefd2fa7a68edbd82068c11a25ce3d777f63a95b90d4e11c26f8"
723+ },
724+ {
725+ "path": "benchmarks/raw/tests/numerical_cap8_2026-11-01T012300.json",
726+ "bytes": 1185,
727+ "sha256": "0e043911673755aa445a63d414cc7d74a8785396209627764e329c81d2d44a44"
728+ },
729+ {
730+ "path": "benchmarks/raw/tests/numerical_cap8_2026-11-01T013058.json",
731+ "bytes": 1188,
732+ "sha256": "68c05df7d36d0fc557de2e5f56217e345bd1e2f54a3c997c966f7a2d992833d4"
733+ },
734+ {
735+ "path": "benchmarks/raw/tests/numerical_cap8_2026-11-01T013343.json",
736+ "bytes": 1187,
737+ "sha256": "8aa27d29b405e33a7ecb0f9fe5c149d354e7acbf53f852cd1752974981d9c2b5"
738+ },
739+ {
740+ "path": "benchmarks/raw/tests/numerical_olmoe_cap16_2026-11-01T010212.json",
741+ "bytes": 1415,
742+ "sha256": "8e5310a4d6309011d34c258bfd6d1c4453caf87c01969317c08c013d86859a3b"
743+ },
744+ {
745+ "path": "benchmarks/raw/tests/numerical_olmoe_cap16_2026-11-01T010752.json",
746+ "bytes": 1412,
747+ "sha256": "3f66a5d445adb3e47b1b0457a9047abfbb95b495da00526ca0b645b4fbef80cf"
748+ },
749+ {
750+ "path": "benchmarks/raw/tests/numerical_olmoe_cap16_2026-11-01T014410.json",
751+ "bytes": 1412,
752+ "sha256": "361306d7218ed3469589751e29ebe5a84f4d25d84706680446e62436f6668458"
753+ },
754+ {
755+ "path": "benchmarks/raw/tests/numerical_olmoe_cap16_2026-11-01T014859.json",
756+ "bytes": 1414,
757+ "sha256": "744d23f25de6e21f4d8351b2caf4c6fd8147ba4426910797661d726de7d11b71"
758+ },
759+ {
760+ "path": "benchmarks/raw/tests/numerical_olmoe_cap16_2026-11-01T015006.json",
761+ "bytes": 1413,
762+ "sha256": "f8f8d3b5c461ba32f4ce00a47d208a99b56bc86ff6ef6cd0b3183cc8527fd51a"
763+ },
764+ {
765+ "path": "benchmarks/raw/tests/numerical_olmoe_cap16_2026-11-01T015025.json",
766+ "bytes": 1413,
767+ "sha256": "f2116b815d37865061fc187ded0540ce8ce9973e3086c1a3052a68ec2afba7f8"
768+ },
769+ {
770+ "path": "benchmarks/raw/tests/numerical_olmoe_cap16_2026-11-01T015227.json",
771+ "bytes": 1413,
772+ "sha256": "632b1e522c5374ab845c6d2b8e084e9644d4c09e990701b3a68290e7d2e9f6b4"
773+ },
774+ {
775+ "path": "benchmarks/raw/tests/numerical_olmoe_cap32_2026-11-01T010212.json",
776+ "bytes": 1390,
777+ "sha256": "7291c489aa22237a8fdee019bb6f889ccc12b2984c4b8fc8b8afa325a9b8f6f7"
778+ },
779+ {
780+ "path": "benchmarks/raw/tests/numerical_olmoe_cap32_2026-11-01T010751.json",
781+ "bytes": 1390,
782+ "sha256": "a421c4b685555f54aa1d022705153ae750688913768ce0c06ddf3e3b1a87eb57"
783+ },
784+ {
785+ "path": "benchmarks/raw/tests/numerical_olmoe_cap32_2026-11-01T014409.json",
786+ "bytes": 1388,
787+ "sha256": "37d32664308b6c27662cf1eae4e772bcd1514f8687afb15638610c384fbd97f6"
788+ },
789+ {
790+ "path": "benchmarks/raw/tests/numerical_olmoe_cap32_2026-11-01T014858.json",
791+ "bytes": 1390,
792+ "sha256": "c3da2303a5094c4f0444b571b789a544fcb36d6500b65ce9b2d3d49c5ecd8763"
793+ },
794+ {
795+ "path": "benchmarks/raw/tests/numerical_olmoe_cap32_2026-11-01T015005.json",
796+ "bytes": 1388,
797+ "sha256": "472f2bd9c46da1bfea04f6c5618f7dfac6a2a9d07c5f716be29e4033fba6edc9"
798+ },
799+ {
800+ "path": "benchmarks/raw/tests/numerical_olmoe_cap32_2026-11-01T015024.json",
801+ "bytes": 1391,
802+ "sha256": "10b96bc0ddba360d82d33b009fd29a44378d6f36276ea8ca83779798b8fa339d"
803+ },
804+ {
805+ "path": "benchmarks/raw/tests/numerical_olmoe_cap32_2026-11-01T015226.json",
806+ "bytes": 1390,
807+ "sha256": "7bfcbe825814d1fb4f1f0cfacd84f2a4ed46559b4b730a8e25b57108b135b99b"
808+ },
809+ {
810+ "path": "benchmarks/raw/tests/numerical_olmoe_cap64_2026-11-01T010211.json",
811+ "bytes": 1390,
812+ "sha256": "4cf7dd958e15b49c6607fc479229e560c88cf7ae90836fff82b3b950c7cfcbb3"
813+ },
814+ {
815+ "path": "benchmarks/raw/tests/numerical_olmoe_cap64_2026-11-01T010750.json",
816+ "bytes": 1388,
817+ "sha256": "2ae89b499f5b3c1dbe0fbe9425d952ca39c4d04c89462fadc56f45f38b562f6f"
818+ },
819+ {
820+ "path": "benchmarks/raw/tests/numerical_olmoe_cap64_2026-11-01T014408.json",
821+ "bytes": 1389,
822+ "sha256": "999202a02163fba2a070464be8173b9d0bab7bdf389fe345989cd2a38601a20b"
823+ },
824+ {
825+ "path": "benchmarks/raw/tests/numerical_olmoe_cap64_2026-11-01T014858.json",
826+ "bytes": 1389,
827+ "sha256": "b05092f278608b59801f491d9891fd035ab8058f8614257556e225de0ea49127"
828+ },
829+ {
830+ "path": "benchmarks/raw/tests/numerical_olmoe_cap64_2026-11-01T015005.json",
831+ "bytes": 1391,
832+ "sha256": "2c0bbafb4bf1bc43a9f293537b842bd5c431f9455c530c1e2ac1d49f2d3af44c"
833+ },
834+ {
835+ "path": "benchmarks/raw/tests/numerical_olmoe_cap64_2026-11-01T015024.json",
836+ "bytes": 1390,
837+ "sha256": "d6e00c1c48ef2427f8d1d18ff121638bd41822b8b2270051c30e128b9aa97c29"
838+ },
839+ {
840+ "path": "benchmarks/raw/tests/numerical_olmoe_cap64_2026-11-01T015225.json",
841+ "bytes": 1388,
842+ "sha256": "970cdd319b2db64aecbaa5524a4aac65a7cc9b7db1640b13fd9244fa5d820868"
843+ },
844+ {
845+ "path": "benchmarks/raw/tests/numerical_olmoe_cap8_2026-11-01T010214.json",
846+ "bytes": 1385,
847+ "sha256": "ee15bf0f363648a116b54bcdb3d754202339cec2f836884ffe7224680bc9c923"
848+ },
849+ {
850+ "path": "benchmarks/raw/tests/numerical_olmoe_cap8_2026-11-01T010753.json",
851+ "bytes": 1385,
852+ "sha256": "289e9fab112e387e12686270bc08d64e5b77368909be3b713fffcbe19929c648"
853+ },
854+ {
855+ "path": "benchmarks/raw/tests/numerical_olmoe_cap8_2026-11-01T014411.json",
856+ "bytes": 1381,
857+ "sha256": "7046e5cfe912f949b4a2a33623d5536370a2140031292ccf525b69f6ab0c2608"
858+ },
859+ {
860+ "path": "benchmarks/raw/tests/numerical_olmoe_cap8_2026-11-01T014900.json",
861+ "bytes": 1383,
862+ "sha256": "10f70d6e705a5fc25776f347d3bcd6b19ea0dcc20108cbbd35c0acac1811fcea"
863+ },
864+ {
865+ "path": "benchmarks/raw/tests/numerical_olmoe_cap8_2026-11-01T015007.json",
866+ "bytes": 1384,
867+ "sha256": "c91aa79562cf933f1c0ec2559f91c2a10af5bb02a8039ed7b8ec4c130be2a714"
868+ },
869+ {
870+ "path": "benchmarks/raw/tests/numerical_olmoe_cap8_2026-11-01T015026.json",
871+ "bytes": 1382,
872+ "sha256": "384c9a87d1e846bc416f6b14d338d61f1db67dd66f610e2a694d19116be2d869"
873+ },
874+ {
875+ "path": "benchmarks/raw/tests/numerical_olmoe_cap8_2026-11-01T015229.json",
876+ "bytes": 1381,
877+ "sha256": "2bbfa8d745776ca8b9563b12f347a71dbce0ce92231ba58ca43f72fb169e7e0f"
878+ },
879+ {
880+ "path": "benchmarks/raw/tests/pytest_gateB_20261101T011707.txt",
881+ "bytes": 2371,
882+ "sha256": "0373672c9480ceb3f31bb5fdebe91f860cfd543a4f99138d0aef409d9c15a8f4"
883+ },
884+ {
885+ "path": "benchmarks/raw/tests/pytest_gateB_20261101T011725.txt",
886+ "bytes": 3243,
887+ "sha256": "13e841d15b27b8091df023087e165431832e5d8aa9045fca796c04804d6a5e8f"
888+ },
889+ {
890+ "path": "benchmarks/raw/tests/pytest_gateB_20261101T011752.txt",
891+ "bytes": 3216,
892+ "sha256": "7d4a6868a2b6416778aeb2fb6be9ddc95f40edcea005b54c5157580edda47409"
893+ },
894+ {
895+ "path": "benchmarks/raw/tests/pytest_gateB_20261101T011804.txt",
896+ "bytes": 98,
897+ "sha256": "7e968e396e06b644a263714a78329aac6c83adb5bd8af1fa18d686856e4fee92"
898+ },
899+ {
900+ "path": "benchmarks/raw/tests/pytest_gateB_fastpath_20261101T012256.txt",
901+ "bytes": 98,
902+ "sha256": "dbf9c884dbb9826afa61139b56a476705007d463d2d5a57c4e981b0c3e6dd2dd"
903+ },
904+ {
905+ "path": "figures/attribution.pdf",
906+ "bytes": 20049,
907+ "sha256": "3f04ea3873d433a0a4cbfa1a25355d1ed08171653d68a5e402cee6dd3f8f3a3c"
908+ },
909+ {
910+ "path": "figures/capacity_sweep.pdf",
911+ "bytes": 19212,
912+ "sha256": "30137bbdeaa5c6d45be4fa75494da3d794ce0c5fc51ac0b2594a008caf4de577"
913+ },
914+ {
915+ "path": "figures/sim_hit_vs_capacity.pdf",
916+ "bytes": 24413,
917+ "sha256": "92365b86a48d35b423d3da4b96595bc275262f59ab18894ece71d5a03ae03a82"
918+ },
919+ {
920+ "path": "figures/two_node.pdf",
921+ "bytes": 14785,
922+ "sha256": "6480148f9056e0b03b64687c04d19d7775b17cdc53fbbe823fee6d9e3f773bf9"
923+ },
924+ {
925+ "path": "paper/main.pdf",
926+ "bytes": 454584,
927+ "sha256": "fcb20dc814dc44b33359e879174156c3ddea716518e4fba8005016879fd3ee7c"
928+ },
929+ {
930+ "path": "paper/tables/attr.tex",
931+ "bytes": 1217,
932+ "sha256": "8389ed69e2e03b7d7263e88a314c2b0fdb7bc6a934d0b8b62c3d76c6daccc8e6"
933+ },
934+ {
935+ "path": "paper/tables/audit.tex",
936+ "bytes": 3423,
937+ "sha256": "9463ccfccccc83052b8671a6edb1af5fcd5d94ca7a439a6c39de12b594b57be2"
938+ },
939+ {
940+ "path": "paper/tables/cold.tex",
941+ "bytes": 1439,
942+ "sha256": "cc9949641149298d7fa00ba67f22e615b0398ccfca768cd88a956df9431cb130"
943+ },
944+ {
945+ "path": "paper/tables/fused.tex",
946+ "bytes": 762,
947+ "sha256": "87c0f13b50dab545c9b2e7960fbe46e152bb6a46e0479a6b7a0c630fe18ecb21"
948+ },
949+ {
950+ "path": "paper/tables/qwen3.tex",
951+ "bytes": 607,
952+ "sha256": "0aee76729b077d2b84afb80d3955876d2fd6e15c4bf16226287d14bd7d034b62"
953+ },
954+ {
955+ "path": "paper/tables/sim.tex",
956+ "bytes": 1821,
957+ "sha256": "a15f701f749f9a258305deed8bd368b7bf2b289a0ad4094c339c7c60b14450ae"
958+ },
959+ {
960+ "path": "research/prior_art.csv",
961+ "bytes": 30241,
962+ "sha256": "7e3366a429d0189a09e3edf7033976f2ba859ead39a5ebe79e2f568df59b6ba3"
963+ }
964+ ]
965+}
\ No newline at end of file
added benchmarks/processed/audit.json +535 −0
@@ -0,0 +1,535 @@
1+{
2+ "B": 10000,
3+ "seed": 0,
4+ "comparisons": [
5+ {
6+ "comparison": "laptop: TITAN all-resident (64 slots, preloaded) vs mlx-lm",
7+ "n_a": 12,
8+ "n_b": 12,
9+ "median_a": 489.12865342395037,
10+ "ci_a": [
11+ 488.2082407000963,
12+ 489.5368828457765
13+ ],
14+ "median_b": 495.04061243824856,
15+ "ci_b": [
16+ 494.1194679570178,
17+ 495.8904708999346
18+ ],
19+ "diff": -5.911959014298191,
20+ "ci_diff": [
21+ -7.065204357666232,
22+ -4.834550062056792
23+ ],
24+ "rel_diff_pct": -1.1942371728209773,
25+ "unit": "tok/s",
26+ "source": "gateB/capacity_sweep_laptop/sweep.jsonl"
27+ },
28+ {
29+ "comparison": "laptop: TITAN 64 slots not preloaded (sync path) vs mlx-lm",
30+ "n_a": 12,
31+ "n_b": 12,
32+ "median_a": 201.4454551722675,
33+ "ci_a": [
34+ 200.95837279071753,
35+ 202.66232324191964
36+ ],
37+ "median_b": 495.04061243824856,
38+ "ci_b": [
39+ 494.1194679570178,
40+ 495.8904708999346
41+ ],
42+ "diff": -293.5951572659811,
43+ "ci_diff": [
44+ -294.4450157276671,
45+ -291.97897074353386
46+ ],
47+ "rel_diff_pct": -59.307287097097344,
48+ "unit": "tok/s",
49+ "source": "gateB/capacity_sweep_laptop/sweep.jsonl"
50+ },
51+ {
52+ "comparison": "M3U96a Qwen3-30B: TITAN all-resident (128 slots) vs mlx-lm",
53+ "n_a": 18,
54+ "n_b": 18,
55+ "median_a": 121.58726022750393,
56+ "ci_a": [
57+ 121.4222850284662,
58+ 121.7539110305407
59+ ],
60+ "median_b": 121.65810731258236,
61+ "ci_b": [
62+ 121.37584628881791,
63+ 121.8112669918118
64+ ],
65+ "diff": -0.07084708507842663,
66+ "ci_diff": [
67+ -0.27269819921461647,
68+ 0.2832330343496068
69+ ],
70+ "rel_diff_pct": -0.058234577738740925,
71+ "unit": "tok/s",
72+ "source": "gateC/qwen3_m3u96a/baseline.jsonl"
73+ },
74+ {
75+ "comparison": "laptop cold NVMe, 16 slots: transition m=8 vs no prefetch",
76+ "n_a": 12,
77+ "n_b": 12,
78+ "median_a": 27.904324206563125,
79+ "ci_a": [
80+ 26.510834259463596,
81+ 30.51331379185264
82+ ],
83+ "median_b": 36.36478535014914,
84+ "ci_b": [
85+ 34.13753044405526,
86+ 42.04500706287061
87+ ],
88+ "diff": -8.460461143586016,
89+ "ci_diff": [
90+ -14.521560050815605,
91+ -5.0831251036633525
92+ ],
93+ "rel_diff_pct": -23.265533020811077,
94+ "unit": "tok/s",
95+ "source": "gateC/cold_nvme_laptop/sweep.jsonl"
96+ },
97+ {
98+ "comparison": "laptop cold NVMe, 16 slots: transition2 m=8 vs no prefetch",
99+ "n_a": 12,
100+ "n_b": 12,
101+ "median_a": 34.96599880167474,
102+ "ci_a": [
103+ 33.67628976796817,
104+ 41.49003741051476
105+ ],
106+ "median_b": 36.36478535014914,
107+ "ci_b": [
108+ 34.13753044405526,
109+ 42.04500706287061
110+ ],
111+ "diff": -1.3987865484743978,
112+ "ci_diff": [
113+ -7.324854162049796,
114+ 5.998514358174624
115+ ],
116+ "rel_diff_pct": -3.8465414686372155,
117+ "unit": "tok/s",
118+ "source": "gateC/cold_nvme_laptop/sweep.jsonl"
119+ },
120+ {
121+ "comparison": "laptop cold NVMe, 16 slots: transition3 m=8 vs no prefetch",
122+ "n_a": 12,
123+ "n_b": 12,
124+ "median_a": 29.083938320259847,
125+ "ci_a": [
126+ 27.671125828589567,
127+ 34.70508437115666
128+ ],
129+ "median_b": 36.36478535014914,
130+ "ci_b": [
131+ 34.13753044405526,
132+ 42.04500706287061
133+ ],
134+ "diff": -7.280847029889294,
135+ "ci_diff": [
136+ -13.09924827054045,
137+ -1.0543321191199624
138+ ],
139+ "rel_diff_pct": -20.02169670405997,
140+ "unit": "tok/s",
141+ "source": "gateC/cold_nvme_laptop/sweep.jsonl"
142+ },
143+ {
144+ "comparison": "laptop cold NVMe, 16 slots: transition2 m=16 vs no prefetch",
145+ "n_a": 12,
146+ "n_b": 12,
147+ "median_a": 25.940903169028186,
148+ "ci_a": [
149+ 25.001461847324514,
150+ 29.583536873911555
151+ ],
152+ "median_b": 36.36478535014914,
153+ "ci_b": [
154+ 34.13753044405526,
155+ 42.04500706287061
156+ ],
157+ "diff": -10.423882181120955,
158+ "ci_diff": [
159+ -16.145654899217902,
160+ -5.865959852413532
161+ ],
162+ "rel_diff_pct": -28.66477027363563,
163+ "unit": "tok/s",
164+ "source": "gateC/cold_nvme_laptop/sweep.jsonl"
165+ },
166+ {
167+ "comparison": "laptop cold NVMe, 16 slots: prev_rank m=16 vs no prefetch",
168+ "n_a": 12,
169+ "n_b": 12,
170+ "median_a": 19.13073106755353,
171+ "ci_a": [
172+ 18.745326302296334,
173+ 20.005360570414403
174+ ],
175+ "median_b": 36.36478535014914,
176+ "ci_b": [
177+ 34.13753044405526,
178+ 42.04500706287061
179+ ],
180+ "diff": -17.23405428259561,
181+ "ci_diff": [
182+ -22.91427599531708,
183+ -14.833423635496317
184+ ],
185+ "rel_diff_pct": -47.392151821198446,
186+ "unit": "tok/s",
187+ "source": "gateC/cold_nvme_laptop/sweep.jsonl"
188+ },
189+ {
190+ "comparison": "laptop cold NVMe, 32 slots: transition m=8 vs no prefetch",
191+ "n_a": 12,
192+ "n_b": 12,
193+ "median_a": 41.32499296378025,
194+ "ci_a": [
195+ 35.45132953777109,
196+ 46.87438633199691
197+ ],
198+ "median_b": 50.55650193124784,
199+ "ci_b": [
200+ 46.6179394611299,
201+ 56.852686944161775
202+ ],
203+ "diff": -9.23150896746759,
204+ "ci_diff": [
205+ -17.486052289083403,
206+ -2.598222895951629
207+ ],
208+ "rel_diff_pct": -18.259785813546966,
209+ "unit": "tok/s",
210+ "source": "gateC/cold_nvme_laptop/sweep.jsonl"
211+ },
212+ {
213+ "comparison": "laptop cold NVMe, 32 slots: transition2 m=8 vs no prefetch",
214+ "n_a": 12,
215+ "n_b": 12,
216+ "median_a": 42.02425447124503,
217+ "ci_a": [
218+ 38.459575220686936,
219+ 46.75470157960534
220+ ],
221+ "median_b": 50.55650193124784,
222+ "ci_b": [
223+ 46.6179394611299,
224+ 56.852686944161775
225+ ],
226+ "diff": -8.532247460002807,
227+ "ci_diff": [
228+ -15.240124741890014,
229+ -2.7179076483432
230+ ],
231+ "rel_diff_pct": -16.876657074901807,
232+ "unit": "tok/s",
233+ "source": "gateC/cold_nvme_laptop/sweep.jsonl"
234+ },
235+ {
236+ "comparison": "laptop cold NVMe, 32 slots: transition3 m=8 vs no prefetch",
237+ "n_a": 12,
238+ "n_b": 12,
239+ "median_a": 41.02227504595441,
240+ "ci_a": [
241+ 37.59374509091457,
242+ 46.645720976690825
243+ ],
244+ "median_b": 50.55650193124784,
245+ "ci_b": [
246+ 46.6179394611299,
247+ 56.852686944161775
248+ ],
249+ "diff": -9.534226885293428,
250+ "ci_diff": [
251+ -16.24725536552569,
252+ -2.9552358624899995
253+ ],
254+ "rel_diff_pct": -18.858557299433205,
255+ "unit": "tok/s",
256+ "source": "gateC/cold_nvme_laptop/sweep.jsonl"
257+ },
258+ {
259+ "comparison": "laptop cold NVMe, 32 slots: transition2 m=16 vs no prefetch",
260+ "n_a": 12,
261+ "n_b": 12,
262+ "median_a": 40.9858914503088,
263+ "ci_a": [
264+ 35.4116167818508,
265+ 46.03429262237789
266+ ],
267+ "median_b": 50.55650193124784,
268+ "ci_b": [
269+ 46.6179394611299,
270+ 56.852686944161775
271+ ],
272+ "diff": -9.570610480939038,
273+ "ci_diff": [
274+ -17.710608210914906,
275+ -3.556491759775213
276+ ],
277+ "rel_diff_pct": -18.930523504086935,
278+ "unit": "tok/s",
279+ "source": "gateC/cold_nvme_laptop/sweep.jsonl"
280+ },
281+ {
282+ "comparison": "laptop cold NVMe, 32 slots: prev_rank m=16 vs no prefetch",
283+ "n_a": 12,
284+ "n_b": 12,
285+ "median_a": 28.37905550744074,
286+ "ci_a": [
287+ 26.997499853533533,
288+ 32.45499196158143
289+ ],
290+ "median_b": 50.55650193124784,
291+ "ci_b": [
292+ 46.6179394611299,
293+ 56.852686944161775
294+ ],
295+ "diff": -22.177446423807098,
296+ "ci_diff": [
297+ -28.57031143448528,
298+ -16.946638006695025
299+ ],
300+ "rel_diff_pct": -43.86665528000013,
301+ "unit": "tok/s",
302+ "source": "gateC/cold_nvme_laptop/sweep.jsonl"
303+ },
304+ {
305+ "comparison": "laptop cold NVMe, 16 slots (general): probe m=8 vs no prefetch",
306+ "n_a": 4,
307+ "n_b": 4,
308+ "median_a": 31.841276936782968,
309+ "ci_a": [
310+ 31.596870661689568,
311+ 32.84528995900421
312+ ],
313+ "median_b": 32.99759250772945,
314+ "ci_b": [
315+ 32.636451601945915,
316+ 33.273626374779
317+ ],
318+ "diff": -1.1563155709464823,
319+ "ci_diff": [
320+ -1.5500378457589257,
321+ -0.1523025487252383
322+ ],
323+ "rel_diff_pct": -3.5042422282038412,
324+ "unit": "tok/s",
325+ "source": "gateC/cold_nvme_laptop/attribution.jsonl"
326+ },
327+ {
328+ "comparison": "laptop cold NVMe, 16 slots (general): probe m=16 vs no prefetch",
329+ "n_a": 4,
330+ "n_b": 4,
331+ "median_a": 25.317370635052725,
332+ "ci_a": [
333+ 24.76517223527836,
334+ 26.203663395431743
335+ ],
336+ "median_b": 32.99759250772945,
337+ "ci_b": [
338+ 32.636451601945915,
339+ 33.273626374779
340+ ],
341+ "diff": -7.680221872676725,
342+ "ci_diff": [
343+ -8.398426644951499,
344+ -6.752416400732631
345+ ],
346+ "rel_diff_pct": -23.275097632888485,
347+ "unit": "tok/s",
348+ "source": "gateC/cold_nvme_laptop/attribution.jsonl"
349+ },
350+ {
351+ "comparison": "laptop cold NVMe, 16 slots (general): probe2 m=8 vs no prefetch",
352+ "n_a": 4,
353+ "n_b": 4,
354+ "median_a": 31.34039585473687,
355+ "ci_a": [
356+ 30.97991967772089,
357+ 32.02471239861806
358+ ],
359+ "median_b": 32.99759250772945,
360+ "ci_b": [
361+ 32.636451601945915,
362+ 33.273626374779
363+ ],
364+ "diff": -1.6571966529925817,
365+ "ci_diff": [
366+ -2.1243054121081393,
367+ -0.8894574043585379
368+ ],
369+ "rel_diff_pct": -5.022174428647772,
370+ "unit": "tok/s",
371+ "source": "gateC/cold_nvme_laptop/attribution.jsonl"
372+ },
373+ {
374+ "comparison": "laptop cold NVMe, 16 slots (general): probe3 m=8 vs no prefetch",
375+ "n_a": 4,
376+ "n_b": 4,
377+ "median_a": 30.747024815645197,
378+ "ci_a": [
379+ 29.277490438874107,
380+ 32.56493375331503
381+ ],
382+ "median_b": 32.99759250772945,
383+ "ci_b": [
384+ 32.636451601945915,
385+ 33.273626374779
386+ ],
387+ "diff": -2.2505676920842532,
388+ "ci_diff": [
389+ -3.7201020688553434,
390+ -0.4326587544144189
391+ ],
392+ "rel_diff_pct": -6.820399674785588,
393+ "unit": "tok/s",
394+ "source": "gateC/cold_nvme_laptop/attribution.jsonl"
395+ },
396+ {
397+ "comparison": "laptop cold NVMe, 16 slots (general): transition2 m=8 vs no prefetch",
398+ "n_a": 4,
399+ "n_b": 4,
400+ "median_a": 27.12807201818323,
401+ "ci_a": [
402+ 26.024585035633084,
403+ 28.245109293687104
404+ ],
405+ "median_b": 32.99759250772945,
406+ "ci_b": [
407+ 32.636451601945915,
408+ 33.273626374779
409+ ],
410+ "diff": -5.869520489546222,
411+ "ci_diff": [
412+ -7.147820761442922,
413+ -4.541891551805659
414+ ],
415+ "rel_diff_pct": -17.787723417007857,
416+ "unit": "tok/s",
417+ "source": "gateC/cold_nvme_laptop/attribution.jsonl"
418+ },
419+ {
420+ "comparison": "laptop cold NVMe, 16 slots (general): transition2t50 m=8 vs no prefetch",
421+ "n_a": 4,
422+ "n_b": 4,
423+ "median_a": 27.924338730957444,
424+ "ci_a": [
425+ 26.832091171562322,
426+ 29.41238423146309
427+ ],
428+ "median_b": 32.99759250772945,
429+ "ci_b": [
430+ 32.636451601945915,
431+ 33.273626374779
432+ ],
433+ "diff": -5.073253776772006,
434+ "ci_diff": [
435+ -6.165501336167129,
436+ -3.5852082762663606
437+ ],
438+ "rel_diff_pct": -15.37461793791057,
439+ "unit": "tok/s",
440+ "source": "gateC/cold_nvme_laptop/attribution.jsonl"
441+ },
442+ {
443+ "comparison": "laptop cold NVMe, 16 slots (general): transition2t70 m=8 vs no prefetch",
444+ "n_a": 4,
445+ "n_b": 4,
446+ "median_a": 27.822276415206307,
447+ "ci_a": [
448+ 26.652845588205146,
449+ 29.388223861291042
450+ ],
451+ "median_b": 32.99759250772945,
452+ "ci_b": [
453+ 32.636451601945915,
454+ 33.273626374779
455+ ],
456+ "diff": -5.175316092523143,
457+ "ci_diff": [
458+ -6.3447469195243045,
459+ -3.6093686464384085
460+ ],
461+ "rel_diff_pct": -15.683920247547945,
462+ "unit": "tok/s",
463+ "source": "gateC/cold_nvme_laptop/attribution.jsonl"
464+ },
465+ {
466+ "comparison": "laptop H4: fused up+gate vs two kernels, 64 slots resident",
467+ "n_a": 12,
468+ "n_b": 12,
469+ "median_a": 475.590753313536,
470+ "ci_a": [
471+ 473.54619144727985,
472+ 476.23632189642444
473+ ],
474+ "median_b": 476.71854951152625,
475+ "ci_b": [
476+ 475.7441354752446,
477+ 478.1937096893877
478+ ],
479+ "diff": -1.1277961979902216,
480+ "ci_diff": [
481+ -3.608229933155485,
482+ 0.07714985221900861
483+ ],
484+ "rel_diff_pct": -0.23657485095678102,
485+ "unit": "tok/s",
486+ "source": "gateD/fused_upgate_laptop.jsonl"
487+ },
488+ {
489+ "comparison": "laptop H4: fused up+gate vs two kernels, 16 slots cold",
490+ "n_a": 4,
491+ "n_b": 4,
492+ "median_a": 35.94742494598419,
493+ "ci_a": [
494+ 35.731574910456565,
495+ 36.7122725819878
496+ ],
497+ "median_b": 34.98584079217508,
498+ "ci_b": [
499+ 34.12741716795578,
500+ 35.49684771325741
501+ ],
502+ "diff": 0.9615841538091061,
503+ "ci_diff": [
504+ 0.42788351347727627,
505+ 1.964214448416044
506+ ],
507+ "rel_diff_pct": 2.7484951970174563,
508+ "unit": "tok/s",
509+ "source": "gateD/fused_upgate_laptop.jsonl"
510+ },
511+ {
512+ "comparison": "M3U96a+M3U96b OLMoE 32 slots: odd experts remote vs local",
513+ "n_a": 16,
514+ "n_b": 16,
515+ "median_a": 15.099750302534616,
516+ "ci_a": [
517+ 14.832823372534836,
518+ 15.478581724182211
519+ ],
520+ "median_b": 81.29106238077739,
521+ "ci_b": [
522+ 70.45778798608657,
523+ 86.91246514746022
524+ ],
525+ "diff": -66.19131207824277,
526+ "ci_diff": [
527+ -71.76838200410603,
528+ -55.195792166972154
529+ ],
530+ "rel_diff_pct": -81.42507938719571,
531+ "unit": "tok/s",
532+ "source": "gateE/two_node_m3u96a_m3u96b.jsonl"
533+ }
534+ ]
535+}
\ No newline at end of file
modified paper/main.pdf +0 −0

Binary file not shown.

modified paper/sections/robustness.tex +8 −2
@@ -4,5 +4,11 @@ Held-out prompt families give hit rates within 0.01 of the fit families at every
4 predictor's recall drops from 0.53 (fit) to 0.50 (held-out), so the locality statistics are not an artefact of the fitting 4 predictor's recall drops from 0.53 (fit) to 0.50 (held-out), so the locality statistics are not an artefact of the fitting
5 prompts. Throughput measurements on the laptop vary by less than 1\,\% across repeats for the all-resident configurations and by 5 prompts. Throughput measurements on the laptop vary by less than 1\,\% across repeats for the all-resident configurations and by
6 up to 40\,\% across prompt families for demand loading (code prompts route more repetitively than general prose, hence higher 6 up to 40\,\% across prompt families for demand loading (code prompts route more repetitively than general prose, hence higher
7−hit rates). Thermal state and competing processes are not yet controlled in the smoke tests reported here; the final sweeps 7+hit rates). Thermal state and competing processes are not controlled in the smoke tests; the cold-NVMe sweeps were run alone on the
8−sample \code{powermetrics} thermal pressure every second and discard runs under throttling. 8+laptop. Table~\ref{tab:audit} gives bootstrap confidence intervals for every comparison used in the text
9+(\code{scripts/audit\_stats.py}, 10\,000 resamples, seed 0, raw logs only): the all-resident parity with mlx-lm holds within
10+$-5.9$ [$-7.1$, $-4.8$]~\tps{} on the laptop and $-0.1$ [$-0.3$, $+0.3$] on M3U96a; every prefetch configuration's interval
11+excludes a gain except the depth-2 transition predictor at 16 slots, whose interval straddles zero ($-1.4$ [$-7.3$, $+6.0$]); the
12+fused-kernel differences are within $\pm 1$~\tps{}; and the two-node loss is $-66$ [$-72$, $-55$]~\tps{}.
13+
14+\input{tables/audit}
added paper/tables/audit.tex +34 −0
@@ -0,0 +1,34 @@
1+\begin{table}[t]
2+\centering\scriptsize
3+\caption{Statistical audit: medians with bootstrap 95\,\% CIs (10\,000 resamples) of decode throughput (tok/s) and of the difference A $-$ B.}
4+\label{tab:audit}
5+\begin{tabular}{L{7.2cm}rrrr}
6+\toprule
7+Comparison (A vs B) & $n_A/n_B$ & A [CI] & B [CI] & A$-$B [CI] \\
8+\midrule
9+laptop: TITAN all-resident (64 slots, preloaded) vs mlx-lm & 12/12 & 489.1 [488.2, 489.5] & 495.0 [494.1, 495.9] & -5.9 [-7.1, -4.8] \\
10+laptop: TITAN 64 slots not preloaded (sync path) vs mlx-lm & 12/12 & 201.4 [201.0, 202.7] & 495.0 [494.1, 495.9] & -293.6 [-294.4, -292.0] \\
11+M3U96a Qwen3-30B: TITAN all-resident (128 slots) vs mlx-lm & 18/18 & 121.6 [121.4, 121.8] & 121.7 [121.4, 121.8] & -0.1 [-0.3, +0.3] \\
12+laptop cold NVMe, 16 slots: transition m=8 vs no prefetch & 12/12 & 27.9 [26.5, 30.5] & 36.4 [34.1, 42.0] & -8.5 [-14.5, -5.1] \\
13+laptop cold NVMe, 16 slots: transition2 m=8 vs no prefetch & 12/12 & 35.0 [33.7, 41.5] & 36.4 [34.1, 42.0] & -1.4 [-7.3, +6.0] \\
14+laptop cold NVMe, 16 slots: transition3 m=8 vs no prefetch & 12/12 & 29.1 [27.7, 34.7] & 36.4 [34.1, 42.0] & -7.3 [-13.1, -1.1] \\
15+laptop cold NVMe, 16 slots: transition2 m=16 vs no prefetch & 12/12 & 25.9 [25.0, 29.6] & 36.4 [34.1, 42.0] & -10.4 [-16.1, -5.9] \\
16+laptop cold NVMe, 16 slots: prev\_rank m=16 vs no prefetch & 12/12 & 19.1 [18.7, 20.0] & 36.4 [34.1, 42.0] & -17.2 [-22.9, -14.8] \\
17+laptop cold NVMe, 32 slots: transition m=8 vs no prefetch & 12/12 & 41.3 [35.5, 46.9] & 50.6 [46.6, 56.9] & -9.2 [-17.5, -2.6] \\
18+laptop cold NVMe, 32 slots: transition2 m=8 vs no prefetch & 12/12 & 42.0 [38.5, 46.8] & 50.6 [46.6, 56.9] & -8.5 [-15.2, -2.7] \\
19+laptop cold NVMe, 32 slots: transition3 m=8 vs no prefetch & 12/12 & 41.0 [37.6, 46.6] & 50.6 [46.6, 56.9] & -9.5 [-16.2, -3.0] \\
20+laptop cold NVMe, 32 slots: transition2 m=16 vs no prefetch & 12/12 & 41.0 [35.4, 46.0] & 50.6 [46.6, 56.9] & -9.6 [-17.7, -3.6] \\
21+laptop cold NVMe, 32 slots: prev\_rank m=16 vs no prefetch & 12/12 & 28.4 [27.0, 32.5] & 50.6 [46.6, 56.9] & -22.2 [-28.6, -16.9] \\
22+laptop cold NVMe, 16 slots (general): probe m=8 vs no prefetch & 4/4 & 31.8 [31.6, 32.8] & 33.0 [32.6, 33.3] & -1.2 [-1.6, -0.2] \\
23+laptop cold NVMe, 16 slots (general): probe m=16 vs no prefetch & 4/4 & 25.3 [24.8, 26.2] & 33.0 [32.6, 33.3] & -7.7 [-8.4, -6.8] \\
24+laptop cold NVMe, 16 slots (general): probe2 m=8 vs no prefetch & 4/4 & 31.3 [31.0, 32.0] & 33.0 [32.6, 33.3] & -1.7 [-2.1, -0.9] \\
25+laptop cold NVMe, 16 slots (general): probe3 m=8 vs no prefetch & 4/4 & 30.7 [29.3, 32.6] & 33.0 [32.6, 33.3] & -2.3 [-3.7, -0.4] \\
26+laptop cold NVMe, 16 slots (general): transition2 m=8 vs no prefetch & 4/4 & 27.1 [26.0, 28.2] & 33.0 [32.6, 33.3] & -5.9 [-7.1, -4.5] \\
27+laptop cold NVMe, 16 slots (general): transition2t50 m=8 vs no prefetch & 4/4 & 27.9 [26.8, 29.4] & 33.0 [32.6, 33.3] & -5.1 [-6.2, -3.6] \\
28+laptop cold NVMe, 16 slots (general): transition2t70 m=8 vs no prefetch & 4/4 & 27.8 [26.7, 29.4] & 33.0 [32.6, 33.3] & -5.2 [-6.3, -3.6] \\
29+laptop H4: fused up+gate vs two kernels, 64 slots resident & 12/12 & 475.6 [473.5, 476.2] & 476.7 [475.7, 478.2] & -1.1 [-3.6, +0.1] \\
30+laptop H4: fused up+gate vs two kernels, 16 slots cold & 4/4 & 35.9 [35.7, 36.7] & 35.0 [34.1, 35.5] & +1.0 [+0.4, +2.0] \\
31+M3U96a+M3U96b OLMoE 32 slots: odd experts remote vs local & 16/16 & 15.1 [14.8, 15.5] & 81.3 [70.5, 86.9] & -66.2 [-71.8, -55.2] \\
32+\bottomrule
33+\end{tabular}
34+\end{table}
added scripts/audit_stats.py +81 −0
@@ -0,0 +1,81 @@
1+#!/usr/bin/env python3
2+"""Gate G statistical audit: bootstrap 95% confidence intervals (10 000 resamples, seed 0) of median decode throughput per
3+configuration and of paired/unpaired differences for the pre-registered comparisons. Reads only raw JSONL logs. Writes
4+benchmarks/processed/audit.json and paper/tables/audit.tex."""
5+import json, os, glob, random, statistics
6+ROOT = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
7+rng = random.Random(0); B = 10000
8+
9+def load(path):
10+ return [json.loads(l) for l in open(path) if l.strip()]
11+
12+def boot_median(xs):
13+ meds = []
14+ for _ in range(B):
15+ s = [xs[rng.randrange(len(xs))] for _ in xs]; meds.append(statistics.median(s))
16+ meds.sort(); return statistics.median(xs), meds[int(0.025 * B)], meds[int(0.975 * B)]
17+
18+def boot_diff(a, b):
19+ ds = []
20+ for _ in range(B):
21+ sa = [a[rng.randrange(len(a))] for _ in a]; sb = [b[rng.randrange(len(b))] for _ in b]
22+ ds.append(statistics.median(sa) - statistics.median(sb))
23+ ds.sort(); return statistics.median(a) - statistics.median(b), ds[int(0.025 * B)], ds[int(0.975 * B)]
24+
25+def sel(recs, **kw):
26+ out = []
27+ for r in recs:
28+ if r.get("warmup"): continue
29+ if all((r.get(k) == v) if not callable(v) else v(r.get(k)) for k, v in kw.items()): out.append(r)
30+ return out
31+
32+comparisons = []
33+def add(name, a, b, field="gen_tps", unit="tok/s", src=""):
34+ xa = [r[field] for r in a if r.get(field) is not None]; xb = [r[field] for r in b if r.get(field) is not None]
35+ if not xa or not xb: return
36+ ma, la, ha = boot_median(xa); mb, lb, hb = boot_median(xb); d, ld, hd = boot_diff(xa, xb)
37+ comparisons.append({"comparison": name, "n_a": len(xa), "n_b": len(xb), "median_a": ma, "ci_a": [la, ha], "median_b": mb, "ci_b": [lb, hb],
38+ "diff": d, "ci_diff": [ld, hd], "rel_diff_pct": 100 * d / mb if mb else None, "unit": unit, "source": src})
39+
40+sw = load(f"{ROOT}/benchmarks/raw/gateB/capacity_sweep_laptop/sweep.jsonl")
41+add("laptop: TITAN all-resident (64 slots, preloaded) vs mlx-lm", sel(sw, backend="titan", capacity=64, tag="warmall"), sel(sw, backend="mlx_lm"), src="gateB/capacity_sweep_laptop/sweep.jsonl")
42+add("laptop: TITAN 64 slots not preloaded (sync path) vs mlx-lm", sel(sw, backend="titan", capacity=64, tag=""), sel(sw, backend="mlx_lm"), src="gateB/capacity_sweep_laptop/sweep.jsonl")
43+q = f"{ROOT}/benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl"
44+if os.path.exists(q):
45+ qq = load(q); add("M3U96a Qwen3-30B: TITAN all-resident (128 slots) vs mlx-lm", sel(qq, backend="titan"), sel(qq, backend="mlx_lm"), src="gateC/qwen3_m3u96a/baseline.jsonl")
46+cold = load(f"{ROOT}/benchmarks/raw/gateC/cold_nvme_laptop/sweep.jsonl")
47+for cap in (16, 32):
48+ base = sel(cold, capacity=cap, predictor="none")
49+ for pr, m in (("transition", 8), ("transition2", 8), ("transition3", 8), ("transition2", 16), ("prev_rank", 16)):
50+ add(f"laptop cold NVMe, {cap} slots: {pr} m={m} vs no prefetch", sel(cold, capacity=cap, predictor=pr, m=m), base, src="gateC/cold_nvme_laptop/sweep.jsonl")
51+attr = load(f"{ROOT}/benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl")
52+base = sel(attr, predictor="none")
53+for pr, m in (("probe", 8), ("probe", 16), ("probe2", 8), ("probe3", 8), ("transition2", 8), ("transition2t50", 8), ("transition2t70", 8)):
54+ add(f"laptop cold NVMe, 16 slots (general): {pr} m={m} vs no prefetch", sel(attr, predictor=pr, m=m), base, src="gateC/cold_nvme_laptop/attribution.jsonl")
55+fz = f"{ROOT}/benchmarks/raw/gateD/fused_upgate_laptop.jsonl"
56+if os.path.exists(fz):
57+ f = load(fz)
58+ add("laptop H4: fused up+gate vs two kernels, 64 slots resident", sel(f, capacity=64, fused_upgate=True), sel(f, capacity=64, fused_upgate=False), src="gateD/fused_upgate_laptop.jsonl")
59+ add("laptop H4: fused up+gate vs two kernels, 16 slots cold", sel(f, capacity=16, fused_upgate=True), sel(f, capacity=16, fused_upgate=False), src="gateD/fused_upgate_laptop.jsonl")
60+tn = f"{ROOT}/benchmarks/raw/gateE/two_node_m3u96a_m3u96b.jsonl"
61+if os.path.exists(tn):
62+ t = load(tn)
63+ add("M3U96a+M3U96b OLMoE 32 slots: odd experts remote vs local", sel(t, capacity=32, placement="odd"), sel(t, capacity=32, placement="none"), src="gateE/two_node_m3u96a_m3u96b.jsonl")
64+gf = f"{ROOT}/benchmarks/raw/gateF/two_node_qwen3_m3u96a_m3u96b.jsonl"
65+if os.path.exists(gf):
66+ g = load(gf)
67+ for cap in (32, 64):
68+ base = sel(g, capacity=cap, placement="none")
69+ for pl in ("odd", "popular:benchmarks/raw/gateF/traces_qwen3_m3u96a"):
70+ add(f"M3U96a+M3U96b Qwen3-30B {cap} slots cold: {pl.split(':')[0]} placement vs local cold NVMe", sel(g, capacity=cap, placement=pl), base, src="gateF/two_node_qwen3_m3u96a_m3u96b.jsonl")
71+os.makedirs(f"{ROOT}/benchmarks/processed", exist_ok=True)
72+json.dump({"B": B, "seed": 0, "comparisons": comparisons}, open(f"{ROOT}/benchmarks/processed/audit.json", "w"), indent=1)
73+os.makedirs(f"{ROOT}/paper/tables", exist_ok=True)
74+L = ["\\begin{table}[t]", "\\centering\\scriptsize", "\\caption{Statistical audit: medians with bootstrap 95\\,\\% CIs (10\\,000 resamples) of decode throughput (tok/s) and of the difference A $-$ B.}", "\\label{tab:audit}",
75+ "\\begin{tabular}{L{7.2cm}rrrr}", "\\toprule", "Comparison (A vs B) & $n_A/n_B$ & A [CI] & B [CI] & A$-$B [CI] \\\\", "\\midrule"]
76+for c in comparisons:
77+ L.append(f"{c['comparison'].replace('_', '\\_').replace('%', '\\%')} & {c['n_a']}/{c['n_b']} & {c['median_a']:.1f} [{c['ci_a'][0]:.1f}, {c['ci_a'][1]:.1f}] & {c['median_b']:.1f} [{c['ci_b'][0]:.1f}, {c['ci_b'][1]:.1f}] & {c['diff']:+.1f} [{c['ci_diff'][0]:+.1f}, {c['ci_diff'][1]:+.1f}] \\\\")
78+L += ["\\bottomrule", "\\end{tabular}", "\\end{table}"]
79+open(f"{ROOT}/paper/tables/audit.tex", "w").write("\n".join(L) + "\n")
80+for c in comparisons:
81+ print(f"{c['comparison'][:70]:70s} A={c['median_a']:6.1f} [{c['ci_a'][0]:6.1f},{c['ci_a'][1]:6.1f}] B={c['median_b']:6.1f} diff={c['diff']:+6.1f} [{c['ci_diff'][0]:+6.1f},{c['ci_diff'][1]:+6.1f}]")
added scripts/make_manifest.py +23 −0
@@ -0,0 +1,23 @@
1+#!/usr/bin/env python3
2+"""artifacts/manifest.json: sha256 and size of every raw log, processed result, figure and table, plus the git commit and the
3+model revisions — the reproducibility package index (Gate G)."""
4+import hashlib, json, os, subprocess, time
5+ROOT = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
6+def sha(p):
7+ h = hashlib.sha256()
8+ with open(p, "rb") as f:
9+ for b in iter(lambda: f.read(1 << 20), b""): h.update(b)
10+ return h.hexdigest()
11+files = []
12+for sub in ("benchmarks/raw", "benchmarks/processed", "figures", "paper/tables", "paper/main.pdf", "artifacts/environment.json", "research/prior_art.csv"):
13+ p = os.path.join(ROOT, sub)
14+ if os.path.isfile(p): files.append(p)
15+ else:
16+ for d, _, fs in os.walk(p):
17+ files += [os.path.join(d, f) for f in fs if not f.startswith(".")]
18+man = {"generated": time.strftime("%Y-%m-%dT%H:%M:%S"), "git": subprocess.run(["git", "rev-parse", "HEAD"], capture_output=True, text=True, cwd=ROOT).stdout.strip(),
19+ "models": {"olmoe-1b-7b-0125-4bit": "mlx-community/OLMoE-1B-7B-0125-Instruct-4bit@e23844197887b031e7ddddbb0b8959c5a6853a7b",
20+ "qwen3-30b-a3b-2507-4bit": "mlx-community/Qwen3-30B-A3B-Instruct-2507-4bit@e9675aa3ca5f900ccef55267914466d55ab325fa"},
21+ "files": [{"path": os.path.relpath(p, ROOT), "bytes": os.path.getsize(p), "sha256": sha(p)} for p in sorted(files)]}
22+json.dump(man, open(os.path.join(ROOT, "artifacts", "manifest.json"), "w"), indent=1)
23+print(len(man["files"]), "files,", sum(f["bytes"] for f in man["files"]) / 1e6, "MB")
24
Menu

Theme