Original publication document, rendered from the unchanged release Markdown. Its historical statements retain their original experiment scope; later quality experiments are documented separately in Quality Evaluation.
Residency campaign consolidated results
39/39 workers; 468 measured128-token generations; 39 excluded64-token warmups. Controller duration: 6621.088021 seconds. E3 exploratory.
Decode is sum(N-1)/sum(decode seconds); overall output rate is sum(N)/sum(internal prompt runtime). TTFT is an arithmetic request mean. Each cell has36 measured requests across three workers.
| Model | Capacity | Executor | Decode tok/s | Blocks1 /2 /3 | Mean TTFT s | Mean request s | Overall tok/s |
|---|---|---|---|---|---|---|---|
| olmoe_0125_instruct | 30 | olmoe-core | 23.04 | 22.65 / 22.98 / 23.51 | 6.07 | 11.58 | 11.05 |
| olmoe_0125_instruct | 40 | olmoe-core | 27.12 | 27.21 / 26.75 / 27.42 | 4.53 | 9.21 | 13.90 |
| olmoe_0125_instruct | 50 | olmoe-core | 31.22 | 31.10 / 31.31 / 31.25 | 3.59 | 7.66 | 16.71 |
| olmoe_0125_instruct | 60 | olmoe-core | 35.44 | 35.45 / 35.49 / 35.39 | 2.97 | 6.55 | 19.54 |
| olmoe_0125_instruct | 64 | olmoe-t3 | 62.07 | 62.69 / 61.39 / 62.15 | 0.70 | 2.74 | 46.67 |
| olmoe_0924_instruct | 30 | olmoe-core | 23.00 | 23.12 / 23.14 / 22.74 | 6.17 | 11.69 | 10.95 |
| olmoe_0924_instruct | 40 | olmoe-core | 27.69 | 27.76 / 27.63 / 27.69 | 4.49 | 9.07 | 14.11 |
| olmoe_0924_instruct | 50 | olmoe-core | 31.50 | 31.75 / 31.57 / 31.19 | 3.62 | 7.65 | 16.73 |
| olmoe_0924_instruct | 60 | olmoe-core | 35.01 | 35.47 / 34.72 / 34.85 | 3.00 | 6.63 | 19.31 |
| olmoe_0924_instruct | 64 | olmoe-t3 | 61.26 | 59.72 / 62.65 / 61.47 | 0.75 | 2.82 | 45.41 |
| qwen15_moe_chat | 20 | qwen-top4 | 15.22 | 15.08 / 15.03 / 15.56 | 9.42 | 17.76 | 7.21 |
| qwen15_moe_chat | 30 | qwen-top4 | 17.17 | 17.45 / 17.02 / 17.04 | 7.51 | 14.90 | 8.59 |
| qwen15_moe_chat | 39 | qwen-top4 | 19.95 | 20.31 / 19.78 / 19.76 | 5.80 | 12.17 | 10.52 |
Memory and transfer
RSS/driver VRAM peaks cover entire workers; driver VRAM includes shared desktop use. H2D is summed over36 measured requests, excluding warmups. GiB=2^30 bytes.
| Model | Capacity | Peak worker RSS GiB | Peak driver VRAM GiB | Peak Torch allocated GiB | Measured H2D GiB | Peak C |
|---|---|---|---|---|---|---|
| olmoe_0125_instruct | 30 | 4.97 | 4.15 | 2.44 | 1115.87 | 48 |
| olmoe_0125_instruct | 40 | 4.95 | 4.62 | 2.92 | 580.72 | 52 |
| olmoe_0125_instruct | 50 | 4.95 | 5.15 | 3.39 | 235.35 | 53 |
| olmoe_0125_instruct | 60 | 4.96 | 5.67 | 3.87 | 40.81 | 53 |
| olmoe_0125_instruct | 64 | 4.96 | 6.01 | 4.48 | 0.00 | 56 |
| olmoe_0924_instruct | 30 | 4.95 | 4.11 | 2.44 | 1108.10 | 47 |
| olmoe_0924_instruct | 40 | 4.95 | 4.62 | 2.92 | 576.17 | 52 |
| olmoe_0924_instruct | 50 | 4.95 | 5.15 | 3.39 | 230.88 | 53 |
| olmoe_0924_instruct | 60 | 4.96 | 5.67 | 3.87 | 37.62 | 53 |
| olmoe_0924_instruct | 64 | 4.98 | 5.99 | 4.48 | 0.00 | 58 |
| qwen15_moe_chat | 20 | 7.97 | 7.11 | 5.61 | 2702.75 | 53 |
| qwen15_moe_chat | 30 | 7.99 | 8.08 | 6.60 | 1852.07 | 53 |
| qwen15_moe_chat | 39 | 7.97 | 8.99 | 7.48 | 1172.50 | 53 |
Validation and limits
507 raw prompts reconciled with39 worker aggregates; exact capacity/row counts; per-token timing; source/input hashes; repeated token IDs; supervision/reaping; 13 cells x3 blocks
- Warmups excluded from measured performance.
- Driver VRAM is shared-device usage; RSS is worker process usage; full-worker peaks include preparation/warmup.
- Grouped T3 at64 differs from PRODUCT-core scalar at partial capacities; c64 transition is not a residency-only causal effect.
- TTFT includes prefill and other first-token work; isolated prefill remains UNKNOWN.
- Three block rates describe observed spread, not a confidence interval or universal performance.
- No quality evaluation, cross-model ranking or publication promotion.
No GPU work was launched by consolidation. Only this create-only consolidation directory and its script were added. No live product source, historical raw artifact, stage or commit was changed.
Command: base Python3.13 -B -S artifacts/residency-executors-39-20260911-01/consolidate.py. Success produces this report and summary/per-prompt/per-worker/source-hash JSON. Exact Git status is in git-status.txt; staged status is in git-staged.txt.