[perf]: Track B FastH3 on RTX 4090 and lower-VRAM GPUs - #46

Open

aryan5v wants to merge 24 commits into

h3-release-corefrom

h3-consumer-fp8

Commits on Oct 3, 2026

[wip]: exact-size pinned arenas for H3 offload and 4090 benchmarks
aryan5v
committed
a216768
[wip]: document 4090 setup validation and reproducible baselines
aryan5v
committed
a491f15
[wip]: record checkpoint revision and GPU driver in 4090 benchmark results
aryan5v
committed
4ae791c
[wip]: record completed 480p baseline and stage summary tooling
aryan5v
committed
21f9899
[perf]: reduce H3 attention activation copies and share FP8 input quantization
aryan5v
committed
c716408
[wip]: prototype tile-64 INT8 QK and FP8 PV attention on sm89
aryan5v
committed
bd713dd
[perf]: retain offloaded H3 VAEs on host until their consuming stages
aryan5v
committed
5bf9804
[perf]: add opt-in sm89 tile-64 BF16 and INT8 QK attention
aryan5v
committed
59e6946
[docs]: record cached 4090 timings and sm89 precision validation
aryan5v
committed
b3ab6e1
[wip]: validate tilewise FP8 values and dynamic probability scales
aryan5v
committed
994c220
[docs]: record resident-block timings and encoder constraints
aryan5v
committed
6b18535
[docs]: record five-second 4090 timing and FP8 encoder footprint
aryan5v
committed
92fc26c
[feat]: stream the H3 encoder for consumer VRAM limits
aryan5v
committed
d986589
[perf]: fuse serialized NVFP4 encoder weight dequantization
aryan5v
committed
887deaa
[perf]: release H3 fine-attention copies before the gated merge
aryan5v
committed
2aa19c4
[perf]: share VAE INT8 input preparation and avoid weight copies
aryan5v
committed
9c9f1ed
[docs]: record 12 GiB and streamed 4090 results
aryan5v
committed
e457b68
[perf]: read H3 INT8 sparse attention from existing layouts
aryan5v
committed
7a0d7d3
[perf]: fuse H3 INT8 VAE scaling and bias without intermediates
aryan5v
committed
3c0668f
[bugfix]: retain consumer attention environment parsing after core rebase
aryan5v
committed
a7f7ede
[test]: adapt consumer VSA validation to packed segment metadata
aryan5v
committed
87b22a5
[bench]: sample total GPU memory for consumer budget checks
aryan5v
committed
fb92af1
[docs]: record 42-second 4090 clips and consumer memory validation
aryan5v
committed
5ac85ad
[docs]: record Track B PR and strict 8 GiB budget limit
aryan5v
committed
98cdc7d

Navigation Menu

[perf]: Track B FastH3 on RTX 4090 and lower-VRAM GPUs - #46

[perf]: Track B FastH3 on RTX 4090 and lower-VRAM GPUs#46
aryan5v wants to merge 24 commits into
h3-release-corefrom
h3-consumer-fp8

Commits on Oct 3, 2026

[wip]: exact-size pinned arenas for H3 offload and 4090 benchmarks

[wip]: document 4090 setup validation and reproducible baselines

[wip]: record checkpoint revision and GPU driver in 4090 benchmark results

[wip]: record completed 480p baseline and stage summary tooling

[perf]: reduce H3 attention activation copies and share FP8 input quantization

[wip]: prototype tile-64 INT8 QK and FP8 PV attention on sm89

[perf]: retain offloaded H3 VAEs on host until their consuming stages

[perf]: add opt-in sm89 tile-64 BF16 and INT8 QK attention

[docs]: record cached 4090 timings and sm89 precision validation

[wip]: validate tilewise FP8 values and dynamic probability scales

[docs]: record resident-block timings and encoder constraints

[docs]: record five-second 4090 timing and FP8 encoder footprint

[feat]: stream the H3 encoder for consumer VRAM limits

[perf]: fuse serialized NVFP4 encoder weight dequantization

[perf]: release H3 fine-attention copies before the gated merge

[perf]: share VAE INT8 input preparation and avoid weight copies

[docs]: record 12 GiB and streamed 4090 results

[perf]: read H3 INT8 sparse attention from existing layouts

[perf]: fuse H3 INT8 VAE scaling and bias without intermediates

[bugfix]: retain consumer attention environment parsing after core rebase

[test]: adapt consumer VSA validation to packed segment metadata

[bench]: sample total GPU memory for consumer budget checks

[docs]: record 42-second 4090 clips and consumer memory validation

[docs]: record Track B PR and strict 8 GiB budget limit