Future Machine Learning and System Lab
Official@futuremls-lab · United States of America
Offers specialized diagnostic and optimization capabilities for SGLang inference runtimes, CUDA kernels, and distributed GPU performance.
Agent Skills by Future Machine Learning and System Lab
Showing 14 vetted skills indexed across 1 GitHub repositories.
FutureMLS-Labsglang-bisect-ci-regression
Bisect failing SGLang CI runs to identify code regressions or runner-specific issues.
FutureMLS-Labsglang-diffusion-benchmark-profile
Diagnose denoising bottlenecks in SGLang diffusion workloads via benchmarks and profiler traces.
FutureMLS-Labsglang-diffusion-add-model
Integrate diffusion models and Diffusers pipelines into SGLang's multimodal runtime.
FutureMLS-Labsglang-diffusion-ako4all-kernel
Optimize SGLang diffusion kernels with ncu profiling and denoise validation.
FutureMLS-Labsglang-diffusion-performance
Select SGLang Diffusion performance flags for generation speed, VRAM, and quality.
FutureMLS-Labwrite-sglang-test
Guide contributors in creating SGLang CI and unit tests with proper fixtures and registration.
FutureMLS-Labci-workflow-guide
Orchestrate and debug staged SGLang continuous integration workflows.
FutureMLS-Labadd-jit-kernel
Add validated JIT CUDA kernels to SGLang's jit_kernel module.
FutureMLS-Labdebug-cuda-crash
Diagnose SGLang CUDA crashes by recording kernel API calls and tensor metadata.
FutureMLS-Labsglang-torch-profiler-analysis
Analyze SGLang torch-profiler traces to identify dominant GPU kernels and fusion patterns.
FutureMLS-Labgenerate-profile
Generate end-to-end performance profiling traces for SGLang server runs.
FutureMLS-Labsglang-auto-benchmark
Search SGLang server configurations against throughput and latency objectives.
FutureMLS-Labadd-sgl-kernel
Guide adding AOT CUDA or C++ kernels with Torch registration, CMake, tests, and benchmarks.
FutureMLS-Labdebug-distributed-hang
Diagnose deadlocked SGLang distributed inference runs using NCCL diagnostics and per-rank logs.
Frequently Asked Questions About Future Machine Learning and System Lab
FAQPage SchemaWhat specific tasks can engineers perform using these capabilities?▼
Engineers can diagnose denoising bottlenecks, profile GPU kernel execution, integrate custom diffusion pipelines, and debug distributed inference deadlocks. These capabilities enable precise performance tuning of SGLang environments through systematic trace analysis, kernel registration, and rigorous benchmarking of throughput and latency metrics.
Which technical personas benefit from these diagnostic capabilities?▼
These capabilities are designed for machine learning infrastructure engineers, GPU systems researchers, and performance optimization specialists. Professionals focused on high-throughput inference, custom CUDA kernel development, and distributed system stability will find these diagnostic methods essential for maintaining production-grade multimodal serving environments.
What are the prerequisites for implementing these performance diagnostics?▼
Implementation requires an existing SGLang environment, access to NVIDIA GPU hardware for kernel profiling, and familiarity with CMake and Torch registration. Users must be prepared to capture and analyze torch-profiler traces, NCCL diagnostic logs, and kernel metadata to effectively identify fusion patterns and execution bottlenecks.