weicheng
Community@weichengz0616
Engineering discipline for FLA (flash-linear-attention) GPU kernel development: Triton/Gluon/TileLang optimization, correctness testing, backend dispatch, and MR readiness.
Agent Skills by weicheng
Showing 7 vetted skills indexed across 1 GitHub repositories.
weichengz0616fla-dispatch-backends
Implements and tests runtime backend dispatch for FLA operations using BaseBackend subclasses and verifiers.
weichengz0616fla-optimization-loop
Guides reproducible multi-round optimization of FLA kernels under a frozen pytest correctness gate.
weichengz0616fla-mr-readiness
Validates pull requests against contribution guidelines, test plans, and benchmark evidence requirements.
weichengz0616fla-kda
Guides modification and review of KDA kernels in the Flash Linear Attention library.
weichengz0616fla-nvidia-performance
Guides NVIDIA GPU kernel profiling and performance evidence collection for FLA repository changes.
weichengz0616fla-correctness-coverage
Guides kernel correctness testing and coverage analysis for Triton operators in fla/ops.
weichengz0616fla-triton-to-gluon
Ports Triton kernels in fla/ops to Gluon with explicit layouts, async copies, and MMA control.
Frequently Asked Questions About weicheng
FAQPage SchemaWhat tasks can I accomplish with weicheng's FLA skills?▼
You can optimize FLA GPU kernels (Triton, Gluon, TileLang, CuTe) through a disciplined multi-round loop, port Triton kernels to Gluon, manage backend dispatch decorators and BaseBackend subclasses, work on KDA gate modes and chunk kernels, add correctness tests, and prepare MR-ready benchmark evidence.
Who are these FLA skills designed for?▼
GPU kernel engineers and contributors to the flash-linear-attention repository who modify fla/ops/** code, review KDA kernels, run NVIDIA performance profiling, or need reproducible optimization workflows that never break or game correctness gates.
How does the FLA optimization loop guarantee correctness?▼
It anchors every iteration on FLA's frozen pytest suite covering both forward and backward passes under NaN poisoning as the immutable correctness gate, combined with a task-contract, three-phase iteration protocol, and a silent-bug catalog drawn from kernel-optimization frameworks like KDA and FlashInfer contests.
When should I port a Triton kernel to Gluon?▼
Port when a kernel is register-bound, when num_stages pipelining underperforms, or when you need Hopper/Blackwell features like TMA, TMEM, or tcgen05. The skill provides an incremental sequence preserving numerical parity, a Triton-to-Gluon API mapping, and pitfalls covering proxy fences, mbarriers, and smem budgets.
What prerequisites and dependencies do these skills require?▼
You need the FLA repository with its pytest suite, Triton and optionally Gluon (triton.experimental.gluon), TileLang or CuTe for alternative backends, and NVIDIA GPU access. Detailed profiling uses an installed ncu-report-skill for Nsight Compute collection and diagnosis.