Norwin YuAI usage research
中文 AI usage research

Open source / 2026

Open source contributions.

Small, specific returns to tools I use. Mostly benchmarks on old hardware, documentation about when multi-GPU actually pays, and one API-shape fix for OpenAI-compatible clients.

02 upstream projects08 upstream items43k+★ / 2.8k+★Contributor, not maintainer

01 / Contributions

Measurements, docs and a small fix.

I contribute where my own rig exposes a gap: old cards, a slow CPU, long context and clients that need standard API fields. The reports are public even where the upstream patch was not merged.

Strata / local inference

Multi-GPU on old cards.

43k+ ★

Upstream: Niko1221/Strata . Strata runs a large model on consumer hardware. My contribution is a series of benchmark reports and docs notes for two RTX 2080 Ti 22 GB cards on a dual-socket Xeon E5-2696 v4.

Focus
Layer split vs one card, --batch, vLLM
Hardware
2× 2080 Ti 22 GB · dual Xeon
Form
5 upstream PRs
Data
Raw requests, configs, concurrency
Reported result / engine 0.1.52

Layer split vs one card, UD-IQ4_XS: decode 40.8 → 60.4 tok/s (+48%); 32K fresh prompt 768 → 1,269 tok/s (+65%).

Project Maya / large MoE inference

Model discovery + a 128K report.

2.8k+ ★

Upstream: mw00/project-maya . Maya runs large MoE models across consumer GPUs and system memory. My contribution is a server API field fix plus a long-context speed report on the 2080 Ti pair.

Feature
context_length + max_output_tokens
Endpoint
GET /v1/models · client discovery
Report
Maya-XS · DeepSeek-V4.1-Flash · 128K
Hardware
2× 2080 Ti 22 GB · 94 GB DDR4
Landed on main / 535327b

The model listing now exposes the field names generic clients read, while keeping the existing meta.n_ctx for compatibility.

02 / Method

What I contribute.

Not features for the sake of a commit count. The useful contribution is usually the measurement or the missing sentence that lets the next person make a decision.

01

Measure the awkward rig

Old cards, slow CPUs, long context. The community needs numbers from the hardware people actually own, not only clean benchmark boxes.

02

Return the raw data

Reports include requests, configs, concurrency data and sanitized settings so someone else can rerun or challenge the result.

03

Fix the interface gap

When a client cannot discover a model's context, the smallest useful patch is the API field, not a longer README.