Skip to content
View pctablet505's full-sized avatar
9 followers · 1 following

Achievements

Achievement: Pair Extraordinairex3Achievement: YOLOAchievement: Pull Sharkx3Achievement: Quickdraw

Achievements

Achievement: Pair Extraordinairex3Achievement: YOLOAchievement: Pull Sharkx3Achievement: Quickdraw

Block or report pctablet505

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
pctablet505/README.md

I make ML models smaller and faster, then measure what each change bought.

Decode speed in tokens per second: plain offloading 10.77, with a gather kernel 13.53, with a cache of the busiest experts 16.50 at peak, 14.99 as shipped. Router pruning reached 20.48 but was not shipped because it hurt the output.
10.8 → 15.0 tokens/s A 181 GiB LLM on one 96 GiB GPU, with my gather kernel and expert cache. How →
100 merged PRs In 38 open-source projects: Keras, KerasHub, vLLM, pip, pydantic. See them →
3x faster On-device inference on a phone's NPU and DSP.
70% → 90% Damaged-barcode accuracy, trained on 10M synthetic images.

Also building Porchlamp and RL Alpha Labs.

Open to ML performance and on-device AI roles · pctablet505.github.io

Pinned Loading

  1. RLAlphaLabs RLAlphaLabs Public

    Trading research for Indian stock markets, judged after real costs and tax, not by the best backtest. This repository is the lab's public website; the trading code is private.

    HTML

  2. porchlamp porchlamp Public

    Porchlamp rankings: 11,000 open-source projects ranked by what happens to a stranger's pull request, not by stars.

    HTML

  3. glm53-flash-single-gpu glm53-flash-single-gpu Public

    GLM-5.3-Flash, a 181 GiB mixture-of-experts model, served from one 96 GiB GPU: a gather kernel and a cache of the busiest experts took decoding from 10.8 to 15.0 tokens/s as shipped. Measurements a…

    Python 1

  4. flakehound flakehound Public

    Static checks that find flaky pytest tests by their pattern, with rules for JAX, PyTorch and NumPy. Alpha.

    Python 1

  5. servedeck servedeck Public

    A systemd control plane, an OpenAI-compatible /v1 gateway and a live dashboard for native vLLM on one Linux GPU workstation.

    Python

  6. ML-Guide ML-Guide Public

    The AI and ML courses I would take, in order: four stages from curious to doing machine learning work, with what each one gives you.

    3 3