Open source · Contribution journal
Zro ♥ open source.
A live record of the serving work we build upstream—so the systems we rely on get better for everyone.
Live GitHub data · Refreshes hourly · 27 of 27 pull requests · 2 of 2 repositories
Pull requests · Newest opened first
Zro → SGLang
Open[Spec][Bugfix] Fix EAGLE3 draft lm_head when a reduced vocab is tied
Zro → SGLang
Mergedfix(dsa): rank truncated top-k pages by indexer score, not page id
Zro → SGLang
Openfix(overlap): return a copy from _merge_cache_loc, not a view of the reused buffer
Zro → SGLang
Openperf(kda): split/pack conv state per tensor shard to drop the per-step reshard
Zro → SGLang
Openfeat(dsa): add decode context parallelism (DCP) for MLA attention
Zro → SGLang
Openfeat(dsa): implement fp8 KV cache (e4m3) with a bf16 indexer cache
Zro → vLLM
Open[ROCm][Kimi-K3] Add moonmath kernels for gfx942 MXFP4 SiTU MoE and MLA decode, with decode context parallelism
Zro → vLLM
Open[Perf][Attention] Run non-causal DFlash draft rows on the Triton 3D kernel
Zro → vLLM
Open[Bugfix][ROCm][Kimi-K3] Feed the DFlash drafter the AttnRes mixture instead of prefix sum
Zro → SGLang
Closed[spec decoding] Support block verification for rejection sampling
Zro → SGLang
Open[AMD] moonmath kernels for gfx942: MLA decode, speculative verify, DCP, and MXFP4 MoE
Zro → SGLang
Open[AMD] Split-KV verify on gfx942: CU-aware split cap, bidirectional draft blocks
Zro → SGLang
Mergedfix(dsa): budget and report the indexer key cache in the KV pool sizing
Zro → SGLang
Mergedfeat(dsa): packed-ragged sparse-MLA prefill — batching + radix + chunked prefill
Zro → SGLang
Mergedfix(linear): reshard block-quant scale for jax 0.10 explicit-mesh shard_map
Zro → SGLang
Mergedfix(dsa): route skip_offset "full" layers sparse in decode + prefill (remove unconditional dense front-block)
Zro → SGLang
Mergedfix(profiler): honour explicit host_tracer_level=0 / python_tracer_level=0
Zro → SGLang
Mergedfeat(dsa): fused sparse-MLA prefill for DeepSeek Sparse Attention (opt-in)
Zro → SGLang
Closed[AMD] Add moonmath MLA attention backend with A16W8 decode for CDNA3 (MI300X)
Zro → MaxDiffusion
MergedAdd MagCache inference acceleration for Wan2.2 (T2V + I2V)
Our GitHub
Open-source projects built by MoonMath.ai.
HyperQuant
No description was provided in this repository.
moonmath-ai/HyperQuant ↗amd-kernels
Hand-tuned kernels for AMD Instinct
moonmath-ai/amd-kernels ↗