Pull requests / #2139

#2139 Community benchmark: RTX 3060 on Windows, Strata 0.1.43, three quantizations through 260K input

open · @libratechw · 0 comentários · No GitHub

BenchmarksNVIDIA / CUDAModels & quantsWindows

Descrição

## Summary

Adds a dated community report for one RTX 3060 12 GiB, Core i7-12700 and 64 GB RAM on the official Windows CUDA 0.1.43 engine. It measures Q2_0, IQ2_XS and IQ3_XXS at 64K, 128K and 262,144 context limits, with fresh inputs through 260,000 tokens. Q2_0/IQ2_XS use full native Expert residency; IQ3_XXS uses a 16 GiB resident budget with buffered file reads. Thinking is xhigh, temperature 1 and MTP is enabled. Different placements and uncontrolled OS cache mean this is a practical configuration comparison.

## What changed

Only `bench/results/2026-10-12-community-rtx-3060-0143/` changes. It contains the report, pinned artifact hashes, configs, generated outputs, request timings, sampled telemetry and reproduction/aggregation scripts. It follows the previous [0.1.41 community report (#1725)](https://github.com/Niko1221/Strata/pull/1725), with no engine or application changes and no old-engine rerun.

## Extra Notes

- **Results:** 90/90 speed requests and 21/21 recall cases passed across the core and long-context supplement. IQ2_XS at 128K also completed a one-hour soak. Earlier RAM stops, placement trials and post-reboot retries are documented in the report.
- **Guards:** All four added profiles completed with the 1 GiB / 15-second RAM guard; no 500 MiB retry or reduced-residency fallback was needed. The 95% / 30-second commit guard stayed enabled.
- **Validation:** Core and long-context input hashes, fresh retry inputs and all three interrupted-target hashes matched. The relaxed retry regenerated three conversation follow-ups, so their input hashes changed and the all-input assertion exited 1 after execution and cleanup. Throughput, aggregation and syntax checks passed.
- **Limits:** Different Expert placements and uncontrolled OS cache/background load limit comparisons. The 0.1.41 samples used different settings, so they do not isolate version effects. Synthetic recall does not establish general agent quality or daily-use stability.

No site

Links install, modelos, releases.