FutureMLS-LabOfficial·14 Agent Skills Included

OSCAR

Shrink LLM KV cache 8x with 2-bit quantization

Compresses LLM KV caches to 2-bit precision using offline spectral rotation, cutting memory use by about 8x with near-lossless quality. Eliminates the GPU memory wall that blocks long-context inference, enabling 256K-token models to run on a single laptop or small GPU. Includes SGLang and llama.cpp integration, pre-built rotation files, and agent skills for benchmarking, profiling, kernel tuning, and CI testing.

557|83

↗

npx skills add FutureMLS-Lab/OSCAR --all -g -y

Available:

Explains the SGLang diffusion subsystem architecture to the AI agent and routes performance-tuning questions to the correct sub-skill in this repository.

All Skills in This Repository (14)

Pure Emerald Level Indicators

📦 In Repo

FutureMLS-Lab

sglang-diffusion-benchmark-profile

Diagnose denoising bottlenecks in SGLang diffusion workloads via benchmarks and profiler traces.

sglang-diffusion-add-model

Integrate diffusion models and Diffusers pipelines into SGLang's multimodal runtime.

sglang-diffusion-ako4all-kernel

Optimize SGLang diffusion kernels with ncu profiling and denoise validation.

sglang-diffusion-performance

Select SGLang Diffusion performance flags for generation speed, VRAM, and quality.

write-sglang-test

Guide contributors in creating SGLang CI and unit tests with proper fixtures and registration.

ci-workflow-guide

Orchestrate and debug staged SGLang continuous integration workflows.

add-jit-kernel

Add validated JIT CUDA kernels to SGLang's jit_kernel module.

debug-cuda-crash

Diagnose SGLang CUDA crashes by recording kernel API calls and tensor metadata.

sglang-torch-profiler-analysis

Analyze SGLang torch-profiler traces to identify dominant GPU kernels and fusion patterns.

generate-profile

Generate end-to-end performance profiling traces for SGLang server runs.

sglang-auto-benchmark

Search SGLang server configurations against throughput and latency objectives.

add-sgl-kernel

Guide adding AOT CUDA or C++ kernels with Torch registration, CMake, tests, and benchmarks.

Official

Advanced

Frequently Asked Questions

FAQPage Schema

How to install OSCAR?▼

Run `npx skills add FutureMLS-Lab/OSCAR --all -g -y` in your terminal to install all skills in this suite globally.

What does OSCAR do for LLM inference?▼

OSCAR quantizes the KV cache to 2 bits using offline attention-aware rotations, reducing memory by about 8x while keeping accuracy close to full precision.

Can I run long-context models on a local Mac?▼

Yes. OSCAR's llama.cpp fork runs 256K-context Gemma 4 12B in under 16GB of memory and accelerates decoding with a fused Metal attention kernel.

Which inference frameworks support OSCAR?▼

OSCAR is built into SGLang's main branch and a llama.cpp fork, with vLLM support in progress, and pre-calibrated rotations are downloadable from Hugging Face.

What do the included agent skills cover?▼

The bundled skills guide AI agents through SGLang benchmarking, profiler analysis, CUDA kernel development, CI test writing, and debugging distributed hangs.