Pull requests / #3422

#3422 [Benchmark] RTX 5060 Ti / swift-1.5 / IQ3_XXS / strata-qwen38-thinking-v1-draft.1

open · @yutacar · 0 comentarios · En GitHub

BenchmarksNVIDIA / CUDAModels & quantsWindowsLinux

Descripción

**[Benchmark]** standard benchmark result for RTX 5060 Ti (Blackwell) on WSL2.

## Problem and result
No standard-v1 result existed for an RTX 50 / Blackwell card or for WSL2. This adds one run of `strata bench run --suite standard-v1` (protocol `strata-qwen38-thinking-v1-draft.1`, pilot data) on an RTX 5060 Ti 16 GB with swift-1.5 IQ3_XXS. Status complete: 21 of 21 required requests finished, 0 failed; sanity checks 12 of 12.

| Case | Prompt tok/s | Decode tok/s | TTFT s |
| --- | --- | --- | --- |
| short | 42.7 | 46.0 | 1.83 |
| prompt-4k | 702.7 | 50.6 | 5.85 |
| prompt-16k | 1332.8 | 46.8 | 12.33 |

## Related discussion
none

## Approach
Results-only PR: adds the folder `bench/results/2026-10-20-standard-rtx-5060-ti-f20bf439/` written by the runner, unchanged. No code changes.

## Validation
**Passed:**
- `strata bench run --url http://127.0.0.1:8081 --credential-env STRATA_BENCH_KEY --suite standard-v1 --track default`
- Strata commit `9ae93f7b`, engine 0.1.46.1 (source build), Ubuntu 24.04.4 on WSL2 (Windows 11), GPU driver 616.56
- RTX 5060 Ti 16 GB connected as an eGPU over OCuLink, Ryzen 9 8945HS, 55 GB RAM available to WSL2
- Engine options: `--kv int8 --pcie-frac 0.20 --spec 4 --mtp rt --max-context 65536`, `STRATA_PREFILL_CPU_SHARE=0`
- Checked the folder for user, host names and private addresses before committing: none found

**Failed:**
none observed

**Not run:**
- `--optional-concurrency` stage (server has 1 slot)
- Engine log excerpt: the server was already running under systemd and was reached with `--url`, so the runner did not start it and wrote no log

## Limits
- Draft protocol; pilot data.
- WSL2 virtual disk: the runner reports model storage as "hdd Virtual Disk" (the host drive is an SSD).
- The server had answered 42 requests before this run; caches were warm, as the protocol notes.

En el sitio

Enlaces a install, modelos, releases.