Configuration and identity¶
Measured hardware: one RTX 5090, 32,607 MiB, direct local endpoint, one candidate at a time. The cache-transfer host performed no inference and is excluded.
| Identity | Pinned value |
|---|---|
| QUASAR repository | wallawalla47/Qwen3.8-27B-Quasar-NVFP4-NInferV3 |
| QUASAR revision | fd3d06cea9df58a54d5051aac9545a4a01a922db |
| Swift15 repository | kaushikvira/Qwen3.8-27B-swift15-nvfp4full-dflash2-NInfer-v3 |
| Swift15 revision | ff891a1130adfcf46bf745e06f0fa3dcde2e8c82 |
| NInfer image | anvil-ninfer@sha256:59df48dd99f0d177b73de57ba88eb6457ed7e4b0c5771c45e8976adcd9f5ef67 |
| NInfer source | bace20dc70249eed6402b66d4852c6c3f9612905 |
| Live collection harness | Package 1.5.0, Python 3.14.8; installed source revision unavailable |
| Live source fingerprint | 8842145699f43e59e2649d27e662c09125e1c75f88dff27196cbc232d4677720 |
| Helper/publication worktree | Package 1.2.1, base 2b99e3353eb2e5a2813195af85050b1da0754332, dirty |
The 1.2.1 worktree revision and its tests are not asserted as provenance for the installed 1.5.0 collection. Retained harness source-file hashes.
QUASAR recipe and Swift15 recipe reconstruct the full managed inputs. Only ports and the cache-volume label are synthetic; lineage records every change. As-run prequalification labels were preserved. Binary and artifact SHA checks, model paths, flags and limits are unchanged.
Both profiles used vision off, 32,768-token per-request context, configured C4, request C1, shared KV 131,072, K8V4, DFlash2 K7, host KV 8,192 MiB, eight host state slots, 40 GiB RAM limits without extra swap, and an 8 GiB host reserve. Coding requests matched medium reasoning, temperature 0.6, top-p 0.95, and the 4,096 + 8,192 combined completion cap. No fixed seed was sent; cache warmth and background activity were not fully controlled. QUASAR's artifact precision differs from Swift15; it is not wholly NVFP4 or lossless.