Upstream UpdateNullstack contributes Spectrum diffusion & CDNA3 attention upstream to SGLangRead PR log
Nullstackby NullStack
Back to all posts
SGLangResearchAMDTPU

Nullstack ♥ SGLang

August 9, 2026By NullStack Engineering Team

Nullstack contributes performance work upstream to SGLang, because an open PR is only the start: the work matters when it is accepted into the library and usable by everyone.

Upstream Contributions

StatusContributionUpstream
MERGEDSpectrum-based diffusion acceleration for Flux, WanVideo, HunyuanVideo, and StableDiffusion3 paths.SGLang PR #31491 →
MERGEDMake explicit zero-level host and Python tracing settings work in the SGLang JAX profiler.SGLang JAX PR #1511 →
MERGEDFused sparse-MLA prefill for DeepSeek Sparse Attention on SGLang JAX.SGLang JAX PR #1508 →
IN REVIEWPacked-ragged sparse-MLA prefill with batching, radix, and chunked prefill.SGLang JAX PR #1514 →
MERGEDReshard block-quant scales for the JAX 0.10 explicit-mesh shard-map change.SGLang JAX PR #1513 →
IN REVIEWFix sparse prefill so full skip-offset layers stay sparse.SGLang JAX PR #1512 →
IN REVIEWCorrect DSA indexer-key cache budgeting and memory reporting for sparse MLA.SGLang JAX PR #1532 →
IN REVIEWServe Kimi-K3 on AMD CDNA3 with NullStack MLA multi-query verification, MXFP4 MoE, and chunked prefix KV.SGLang PR #34007 →
IN REVIEWNullStack MLA, an MI300X/CDNA3 MLA attention backend for bf16-query / fp8-KV decode.SGLang PR #32767 →
IN REVIEWLiteAttention ROCm, an opt-in AMD MI300X attention backend for diffusion.SGLang PR #28049 →

Contribution Deep Dives

Spectrum-based diffusion acceleration for Flux, WanVideo, HunyuanVideo, and StableDiffusion3 paths.

MERGED

This contribution adds Spectrum-based adaptive step-skipping and feature forecasting to SGLang's diffusion runtime. It reduces denoising latency while preserving output quality.

Make explicit zero-level host and Python tracing settings work in the SGLang JAX profiler.

MERGED

This contribution makes explicit host_tracer_level=0 and python_tracer_level=0 settings work as intended in the SGLang JAX profiler.

Fused sparse-MLA prefill for DeepSeek Sparse Attention on SGLang JAX.

MERGED

This opt-in sparse-prefill path makes DeepSeek Sparse Attention useful during prefill, not only decode.

Packed-ragged sparse-MLA prefill with batching, radix, and chunked prefill.

IN REVIEW

This builds on the sparse-MLA work with batching, radix, and chunked prefill support.

Reshard block-quant scales for the JAX 0.10 explicit-mesh shard-map change.

MERGED

This fixes block-quant scale resharding for JAX 0.10's explicit-mesh shard-map behavior.

Fix sparse prefill so full skip-offset layers stay sparse.

IN REVIEW

This fixes the sparse-prefill issue we found in Google's code, so full skip-offset layers remain sparse.

Correct DSA indexer-key cache budgeting and memory reporting for sparse MLA.

IN REVIEW

This corrects sparse-MLA KV-pool sizing and memory reporting so DSA indexer-key buffers are budgeted alongside the latent cache.

Serve Kimi-K3 on AMD CDNA3 with NullStack MLA multi-query verification, MXFP4 MoE, and chunked prefix KV.

IN REVIEW

This makes Kimi-K3 serveable on MI300X and MI325X with NullStack MLA multi-query verification, gfx942 MXFP4 MoE support, and chunked prefix KV for long contexts.

NullStack MLA, an MI300X/CDNA3 MLA attention backend for bf16-query / fp8-KV decode.

IN REVIEW

nullstack_mla routes the MI300X shape we run to a device-driven paged A16W8 kernel, while all other shapes fall back to aiter.

LiteAttention ROCm, an opt-in AMD MI300X attention backend for diffusion.

IN REVIEW

This opt-in backend brings NullStack's bf16 forward-attention kernel to SGLang's diffusion runtime on AMD MI300X.

MM

NullStack Research

@nullstack_in

𝕏

Low-level algorithms and hardware-aware systems power private, high-throughput AI inference for coding agents. Stay tuned as we release further benchmarks.