[feat]: run FastH3 across two DGX Sparks with Ray sequence parallel - #1803

Merged

SolitaryThinker merged 18 commits into

hao-ai-lab:mainfrom

aryan5v:aryan/dgx-spark-pair

Sep 1, 2026

Commits on Sep 1, 2026

[perf]: load pipeline components on demand and free them after their last stage

authored and
aryan5v
committed
68e2800
[perf]: expose lazy module loading in the MiniMax H3 example

authored and
aryan5v
committed
f8da010
[perf]: rebuild the deferred-release schedule when a stage is added late

authored and
aryan5v
committed
98f1781
[perf]: expose lazy module loading in the FastH3 example

authored and
aryan5v
committed
41d8475
[perf]: make deferred loading opt-in per pipeline and clean up on failure

authored and
aryan5v
committed
6fe37c1
[perf]: test the real load and stage-construction paths

authored and
aryan5v
committed
43b2a18
[bugfix]: preserve compile across lazy module reloads

Satyam-53
authored and
aryan5v
committed
dd8a73d
[bugfix]: keep unused LoRA setup from loading lazy DiTs

Satyam-53
authored and
aryan5v
committed
77a2bf1
[bugfix]: share compiled VSA graphs across H3 layers

Satyam-53
authored and
aryan5v
committed
cd59de8
[bugfix]: read H3 geometry from checkpoint JSON so lazy load can drop the DiT before VAE decode
aryan5v
committed
53b286c
[feat]: run FastH3 across two DGX Sparks with Ray sequence parallel
aryan5v
committed
c5c2600
[docs]: catalog the two-Spark FastH3 recipe on the MiniMax H3 family page
aryan5v
committed
6f6ddd2
[bugfix]: reload H3 text encoder on later generate() and key Ray Gloo on worker IPs
aryan5v
committed
77c9fd3
[docs]: treat Spark pair height/width/frames as examples, not a locked recipe
aryan5v
committed
0846882
[bugfix]: let lazy_module_load own H3 deferral when both Spark autos arm
aryan5v
committed
28ecd8c
[feat]: compile MiniMax-H3 VAE by default so Spark lazy load can attach it
aryan5v
committed
776329f
[bugfix]: fix lazy-load, LoRA, and Ray NIC issues from the #1803 review
aryan5v
committed
b6a9f77
[bugfix]: keep FastH3 1-GPU lazy_module_load as auto after rebasing onto main
aryan5v
committed
809d74c

Navigation Menu

Uh oh!

[feat]: run FastH3 across two DGX Sparks with Ray sequence parallel - #1803

[feat]: run FastH3 across two DGX Sparks with Ray sequence parallel#1803
SolitaryThinker merged 18 commits into
hao-ai-lab:mainfrom
aryan5v:aryan/dgx-spark-pair

Commits on Sep 1, 2026

[perf]: load pipeline components on demand and free them after their last stage

[perf]: expose lazy module loading in the MiniMax H3 example

[perf]: rebuild the deferred-release schedule when a stage is added late

[perf]: expose lazy module loading in the FastH3 example

[perf]: make deferred loading opt-in per pipeline and clean up on failure

[perf]: test the real load and stage-construction paths

[bugfix]: preserve compile across lazy module reloads

[bugfix]: keep unused LoRA setup from loading lazy DiTs

[bugfix]: share compiled VSA graphs across H3 layers

[bugfix]: read H3 geometry from checkpoint JSON so lazy load can drop the DiT before VAE decode

[feat]: run FastH3 across two DGX Sparks with Ray sequence parallel

[docs]: catalog the two-Spark FastH3 recipe on the MiniMax H3 family page

[bugfix]: reload H3 text encoder on later generate() and key Ray Gloo on worker IPs

[docs]: treat Spark pair height/width/frames as examples, not a locked recipe

[bugfix]: let lazy_module_load own H3 deferral when both Spark autos arm

[feat]: compile MiniMax-H3 VAE by default so Spark lazy load can attach it

[bugfix]: fix lazy-load, LoRA, and Ray NIC issues from the #1803 review

[bugfix]: keep FastH3 1-GPU lazy_module_load as auto after rebasing onto main