Agent Skills by dyyoungg
Showing 18 vetted skills indexed across 1 GitHub repositories.
dyyounggsglang-diffusion-modelopt-quant
Quantize diffusion DiTs with NVIDIA ModelOpt for FP8/NVFP4 SGLang Diffusion.
dyyounggsglang-diffusion-benchmark-profile
Profile diffusion denoise latency and hotspots using the SGLang backend.
dyyounggsglang-diffusion-add-model
Integrate a new diffusion model or Diffusers pipeline into SGLang.
dyyounggsglang-diffusion-ako4all-kernel
Optimize SGLang diffusion kernels with AKO4ALL microbenchmarks and end-to-end validation.
dyyounggsglang-diffusion-performance
Benchmark SGLang diffusion pipelines to select optimal flags for GPUs and VRAM budgets.
dyyounggclean-startup-log
Filter noisy third-party warnings from SGLang server startup logs.
dyyounggwrite-sglang-test
Write and register SGLang CI/UT tests with CustomTestCase and register_*_ci calls.
dyyounggci-workflow-guide
Diagnose and fix CI workflow orchestration issues across SGLang stages.
dyyounggadd-jit-kernel
Add a lightweight CUDA kernel to SGLang's jit_kernel module with validation and Python wrapper.
dyyounggdebug-cuda-crash
Capture CUDA crash boundaries and tensor metadata via kernel API logging in SGLang.
dyyounggllm-serving-auto-benchmark
Run unified benchmarks across SGLang, vLLM, and TensorRT-LLM serving frameworks.
dyyounggsglang-prod-incident-triage
Turn live SGLang serving incidents into replay-based debugging workflows.
dyyoungggenerate-profile
Profile SGLang server runs and capture Chrome-compatible traces for performance analysis.
dyyounggadd-sgl-kernel
Automates end-to-end integration of CUDA/C++ kernels in SGL-Kernel with tests, benchmarks, and Python wrappers.
dyyounggllm-torch-profiler-analysis
Analyze LLM torch-profiler traces to identify kernel, overlap, and fusion patterns.
dyyounggsglang-bisect-ci-regression
Bisect CI commit windows to diagnose regressions from failure logs.
dyyounggsglang-sota-performance
Benchmark LLM serving frameworks and generate profiler traces for performance gap analysis.
dyyounggdebug-distributed-hang
Diagnose distributed hangs in multi-GPU SGLang deployments using per-rank traces, NCCL logs, and CUDA core dumps.