PallasForge open-sources hands-on tutorials for writing GPU kernels in Pallas and JAX
A_K_Nain · x · 2026-10-13
Aakash Kumar Nain released PallasForge, an open-source (Apache-2.0) tutorial repo for writing and benchmarking GPU kernels with Pallas, JAX's kernel extension.
- Aims to fill gaps left by official docs, which only include a few generic examples: real-world kernels across GPU families (Hopper, Blackwell) such as fused dequantization, flash-attention variants, and fused-norm
- Includes profiling guidance with Nsight Compute: reading GPU Speed of Light, memory workload analysis, register spills, occupancy, and warp stall stats, with a sample report walkthrough
- Contributions of kernels, benchmarks, and implementation notes are welcome
More from Infra
- FinOps for Agentic AI: taming skyrocketing cloud and token costs — DavidLinthicum · 2026-10-14
- A big tech repo sees ~25 million code checkouts per day for CI alone — mcraddock · 2026-10-14
- Cloudflare Agents SDK adds persistent browser tool with sessions that survive between agent turns — irvinebroque · 2026-10-14
- Ed Zitron: The AI bubble runs on made-up annualized revenue and capacity numbers — GaryMarcus · 2026-10-14
- Broadcom in early talks to arrange $50B+ financing for OpenAI custom silicon — Beth_Kindig · 2026-10-14
- Ben Bajarin: WFE roadmaps now predict our compute future better than chip roadmaps — BenBajarin · 2026-10-14