Skip to content

merge - #2809

Merged
bghira merged 125 commits into
releasefrom
main
Jul 12, 2026
Merged

merge#2809
bghira merged 125 commits into
releasefrom
main

Conversation

@bghira

@bghira bghira commented Jul 12, 2026

Copy link
Copy Markdown
Owner

No description provided.

bghira and others added 30 commits June 19, 2026 00:26
LTX2: unified conditioning mechanism, flash-attn 3 support, IC-LoRA validations
FSDP2 LoRA training (LTX 2.3 verified)
support fused projections across more models
add context parallel plans for more model families
add more config examples centred around 8xH100 for LTX 2.3
LTX 2.3 context-parallel fixes for uneven sequence lengths
bghira and others added 28 commits June 27, 2026 08:00
krea2: CREPA self-flow, LayerSync, TREAD, block swap, cfg-zero star, skip layer guidance, and flowmap generation (AnyFlow)
cog: add a script to deploy multiple versions in sequence for multigpu support
cloud: update replicate deploy script to push using cog to ensure version manifest is registered
cloud: multigpu training options via Replicate
cloud: fix UI settings default to actually show the tab by default
cloud: update configuration wizard to show correct hardware type selector
…anges

fix --disable_tf32 in newer torch api checks
…w-matching training

Add a `timestep_sampling_offset` option to per-dataset config. This offsets the
pre-sigmoid normal sample in the default logit-normal flow-matching schedule,
letting users shift the noise distribution per dataset based on semantic
granularity.

Negative offset -> lower-noise timesteps (detail), positive -> higher-noise
(structure). Default 0.0 = stock behavior.

Based on: The Quadratic Geometry of Flow Matching: Semantic Granularity
Alignment for Text-to-Image Synthesis (arXiv:2603.10785).
Unify naming across code, config, docs, tests, and WebUI template
to match the PR description and align with noise_offset convention.
No behavioral change.
feat: add per-dataset timestep_sampling_offset for semantic-aware flow-matching training
Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
metal-flash-attention integration (fp32, int8, int4)
@bghira
bghira merged commit 289746d into release Jul 12, 2026
4 checks passed

This branch was previously deployed

1 inactive deployment
github-pages — 2a3f2d47 Deployed Jul 12, 2026 by bghira via deploy #365
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants