Residency and executor campaign
13 configurations · 39 workers · 468 measured responses · E3
PERF-REAL-v1; 12 prompts repeated in 3 blocks; 36 measured responses of 128 tokens per configuration; greedy selection, EOS suppressed; 64-token warmup excluded. Canonical initial placement, evolving LRU through fixed prompts; concurrency 1.
OLMoE c30-c60 use PRODUCT-core scalar; c64 uses Grouped T3. This is not a residency-only comparison. Qwen uses top4-fused-v1.
| Configuration | Executor | Pooled decode token/s | Whole-worker VRAM GiB |
|---|---|---|---|
| OLMoE-1B-7B-0924-Instruct c30/64 | PRODUCT-core scalar | 23.00 | 4.11 |
| OLMoE-1B-7B-0924-Instruct c40/64 | PRODUCT-core scalar | 27.69 | 4.62 |
| OLMoE-1B-7B-0924-Instruct c50/64 | PRODUCT-core scalar | 31.50 | 5.15 |
| OLMoE-1B-7B-0924-Instruct c60/64 | PRODUCT-core scalar | 35.01 | 5.67 |
| OLMoE-1B-7B-0924-Instruct c64/64 | Grouped T3 | 61.26 | 5.99 |
| OLMoE-1B-7B-0125-Instruct c30/64 | PRODUCT-core scalar | 23.04 | 4.15 |
| OLMoE-1B-7B-0125-Instruct c40/64 | PRODUCT-core scalar | 27.12 | 4.62 |
| OLMoE-1B-7B-0125-Instruct c50/64 | PRODUCT-core scalar | 31.22 | 5.15 |
| OLMoE-1B-7B-0125-Instruct c60/64 | PRODUCT-core scalar | 35.44 | 5.67 |
| OLMoE-1B-7B-0125-Instruct c64/64 | Grouped T3 | 62.07 | 6.01 |
| Qwen1.5-MoE-A2.7B-Chat c20/60 | top4-fused-v1 | 15.22 | 7.11 |
| Qwen1.5-MoE-A2.7B-Chat c30/60 | top4-fused-v1 | 17.17 | 8.08 |
| Qwen1.5-MoE-A2.7B-Chat c39/60 | top4-fused-v1 | 19.95 | 8.99 |
Pooled decode excludes first-token wait. VRAM is a device-wide whole-worker driver peak including preparation, warmup and desktop use. RSS and H2D have separate measurement scopes. See each linked result for its revision, full metric planes and limitations.
Audit material
The verifier checks hashes and curated aggregate consistency. This edition does not supply the complete raw-run bundle or independent runtime reproduction.