[feat]: Track C FastH3 resident Spark and MLX support - #47

Open

aryan5v wants to merge 29 commits into

h3-spark-main-basefrom

Commits on Oct 3, 2026

[feat]: add resident FastH3 V2 Spark recipe

Aryan Kumar
authored and
Aryan Kumar
committed
e811dac
[feat]: benchmark resident FastH3 Spark recipe

Aryan Kumar
authored and
Aryan Kumar
committed
fb9c1a2
[docs]: pin Spark benchmark environment

Aryan Kumar
authored and
Aryan Kumar
committed
476265b
[bugfix]: retain per-stage CUDA peaks in Spark benchmarks

Aryan Kumar
authored and
Aryan Kumar
committed
bbb79d9
[feat]: add resident pruned and paired Spark recipes

Aryan Kumar
authored and
Aryan Kumar
committed
99a03b5
[bugfix]: preserve H3 release helpers after main integration

Aryan Kumar
authored and
Aryan Kumar
committed
3f24557
[feat]: support pruned FastH3 AdaLN in MLX

Aryan Kumar
authored and
Aryan Kumar
committed
e5e5c18
[feat]: trim MLX H3 conditioner to required layers

Aryan Kumar
authored and
Aryan Kumar
committed
b8be84c
[feat]: allow FP8 source for MLX H3 conversion

Aryan Kumar
authored and
Aryan Kumar
committed
3075477
[docs]: add pruned FastH3 MLX conversion recipe

Aryan Kumar
authored and
Aryan Kumar
committed
892751c
[feat]: add resident NVFP4 conditioning for MLX H3

Aryan Kumar
authored and
Aryan Kumar
committed
80bfd66
[feat]: enable experimental MXFP8 H3 conversion

Aryan Kumar
authored and
Aryan Kumar
committed
92c46de
[test]: check MXFP8 H3 checkpoint round trip

Aryan Kumar
authored and
Aryan Kumar
committed
f698e3e
[bugfix]: omit unused layers from single-shard MLX encoder

Aryan Kumar
authored and
Aryan Kumar
committed
068054e

Commits on Oct 4, 2026

[test]: verify rank-16 H3 modulation and cache reload

Aryan Kumar
authored and
Aryan Kumar
committed
34d50df
[bugfix]: preserve ModelOpt H3 activation calibration in NVFP4 export

Aryan Kumar
authored and
Aryan Kumar
committed
672e090
[bugfix]: fence NVFP4 activation quantization on DGX Spark

Aryan Kumar
authored and
Aryan Kumar
committed
db96024
[docs]: explain calibrated and reproducible Spark NVFP4 inference

Aryan Kumar
authored and
Aryan Kumar
committed
2f7429e
[bugfix]: keep Spark NVFP4 fence opaque to Dynamo

Aryan Kumar
authored and
Aryan Kumar
committed
0bfb398
[bugfix]: order fresh NVFP4 global scales on DGX Spark

Aryan Kumar
authored and
Aryan Kumar
committed
9f460a7
[bugfix]: carry H3 performance switches to Ray workers

Aryan Kumar
authored and
Aryan Kumar
committed
9e4e9aa
[docs]: record verified Track C Spark release timings

Aryan Kumar
authored and
Aryan Kumar
committed
9fe2bf2
[perf]: bound MLX H3 sparse attention gather memory

Aryan Kumar
authored and
Aryan Kumar
committed
b7c8e5f
[perf]: batch 32 keys in opt-in MLX H3 sparse attention

Aryan Kumar
authored and
Aryan Kumar
committed
42a6d1a
[bugfix]: apply H3 wired residency through the MLX wired API

Aryan Kumar
authored and
Aryan Kumar
committed
944fdb7
[docs]: explain H3 Metal wired-memory limits

Aryan Kumar
authored and
Aryan Kumar
committed
a958051

Commits on Oct 5, 2026

[feat]: cache packed H3 encoder in MLX layout

Aryan Kumar
authored and
Aryan Kumar
committed
f12432e
[feat]: support scaled native H3 floating quantization

Aryan Kumar
authored and
Aryan Kumar
committed
0556335
[perf]: parallelize H3 SIMD softmax across four lanes

Aryan Kumar
authored and
Aryan Kumar
committed
3b23f65