OpenMOSS (SII)
Official@openmoss · China
OpenMOSS Team is a research group under the Shanghai Innovation Institution (SII), working in close collaboration with Fudan University and MOSI Intelligence.
Agent Skills by OpenMOSS (SII)
Showing 14 vetted skills indexed across 1 GitHub repositories.
OpenMOSSsglang-diffusion-modelopt-quant
Quantize diffusion DiT models with NVIDIA ModelOpt to FP8 or NVFP4 checkpoints for SGLang Diffusion.
OpenMOSSsglang-diffusion-benchmark-profile
Analyze SGLang diffusion denoise latency and profile bottlenecks via perf dumps.
OpenMOSSsglang-diffusion-add-model
Automate diffusion model integration into SGLang with Hybrid or Modular pipelines.
OpenMOSSsglang-diffusion-ako4all-kernel
Optimize SGLang diffusion kernels with AKO4ALL tuning and ncu-guided iteration.
OpenMOSSsglang-diffusion-performance
Optimize SGLang diffusion performance settings for a model, GPU, and VRAM budget.
OpenMOSSwrite-sglang-test
Author and register SGLang CI/UT tests using CustomTestCase.
OpenMOSSci-workflow-guide
Document SGLang CI pipeline stage ordering, gating, and fast-fail mechanisms.
OpenMOSSadd-jit-kernel
Add a lightweight JIT CUDA kernel with host wrapper, tests, and benchmarks to SGLang's jit_kernel module.
OpenMOSSdebug-cuda-crash
Log CUDA kernel API calls and tensor metadata around crash boundaries in SGLang.
OpenMOSSsglang-torch-profiler-analysis
Triage SGLang torch-profiler traces into kernel, overlap, and fuse-pattern tables.
OpenMOSSgenerate-profile
Launch an SGLang server, validate readiness, and generate a Chrome-compatible profiling trace.
OpenMOSSsglang-auto-benchmark
Automate SGLang benchmark sweeps with dataset preparation and CSV export.
OpenMOSSadd-sgl-kernel
Add a CUDA/C++ AOT kernel to sgl-kernel with tests and benchmarks.
OpenMOSSdebug-distributed-hang
Diagnose multi-GPU SGLang distributed inference hangs by locating rank divergence during collectives.
Frequently Asked Questions About OpenMOSS (SII)
FAQPage SchemaWhat specific tasks can engineers perform using OpenMOSS capabilities?▼
Engineers can quantize diffusion DiT models to FP8 or NVFP4, profile denoise latency, debug distributed inference hangs, and integrate custom CUDA kernels. These tasks enable fine-grained performance tuning for high-throughput generative workloads on NVIDIA hardware.
Which technical personas benefit from these performance optimization skills?▼
These skills are designed for GPU infrastructure engineers, machine learning researchers, and systems developers focused on high-performance inference. Professionals managing VRAM budgets and multi-GPU clusters will find these profiling and kernel-level debugging capabilities essential for production deployment.
What are the prerequisites for implementing these kernel and profiling optimizations?▼
Implementation requires an environment configured with NVIDIA GPU drivers, CUDA development headers, and the SGLang framework. Users must have access to profiling hardware, such as NVIDIA Nsight Compute, and familiarity with C++ or CUDA for custom kernel integration.