Issues

[Chore] CI: prune superseded Rust toolchains on the self-hosted GPU runner

laurigates/loractl#254 · opened

on Oct 24, 2026

feat(trainer): report the observed caption-dropout rate

laurigates/loractl#251 · opened

on Oct 23, 2026

feat(cli): loss-curve follow-ups — full-run curves across resume, train() wiring test, checkpoint markers

laurigates/loractl#250 · opened

on Oct 23, 2026

fix(hf): token edge cases — same-host redirects drop Authorization, BOM in token files

laurigates/loractl#249 · opened

on Oct 23, 2026

chore(ci): follow-ups from the #239 dependency/action bump

laurigates/loractl#248 · opened

on Oct 23, 2026

test(trainer): pin optim.schedule against a diffusers reference; warn on a resume that changes the decay horizon

laurigates/loractl#247 · opened

on Oct 23, 2026

feat(config): reject unknown keys across the TrainConfig schema

laurigates/loractl#246 · opened

on Oct 23, 2026

feat(core): FLUX.2 VAE encode path in burn: shared prerequisite for MOSS-Image and Fibo 2.0

laurigates/loractl#238 · opened

on Oct 21, 2026

feat(trainer): add Fibo 2.0 LoRA training: 8B single-stream DiT, Qwen3-VL-4B conditioner, FLUX.2 VAE

laurigates/loractl#237 · opened

on Oct 21, 2026

feat(trainer): add MOSS-Image LoRA training: ~5B DiT, Qwen3.5-9B conditioner, FLUX.2 VAE

laurigates/loractl#236 · opened

on Oct 21, 2026

Evaluate migrating the CPU encode backend from burn-ndarray to burn-flex

laurigates/loractl#226 · opened

on Aug 14, 2026

fix(api): POST /runs leaves dataset/model paths unconfined while documenting otherwise

laurigates/loractl#224 · opened

on Aug 11, 2026

feat(trainer): mid-training image previews on the diffusion path

laurigates/loractl#223 · opened

on Aug 11, 2026

feat(sample): image sampling for the diffusion task — the flow trainer has no visual feedback loop

laurigates/loractl#221 · opened

on Aug 10, 2026

feat(dataset): per-tag caption dropout — decorrelate tags that co-occur on every caption

laurigates/loractl#220 · opened

on Aug 10, 2026

1

perf(dataset): time the decode fraction in isolation — the encode-phase wall clock cannot see it

laurigates/loractl#217 · opened

on Aug 10, 2026

1

chore(core): audit resume.rs against burn_train::checkpoint — record the split, don't migrate blind

laurigates/loractl#214 · opened

on Aug 9, 2026

feat(cli): drive burn-train's TUI dashboard as a renderer over the TrainEvent stream

laurigates/loractl#213 · opened

on Aug 9, 2026

docs(gpu-hosting): volume sizing and the empty ram_gb column both assume a Krea 2-shaped workload

laurigates/loractl#207 · opened

on Aug 8, 2026

feat(trainer): add MiniMax H3 video LoRA training — a second target reusing the Qwen3-VL/MMDiT stack

laurigates/loractl#204 · opened

on Aug 8, 2026

1

Run the opt-in qwen3vl-real proof once on a GPU box — #163 merged without ever executing the test that checks 34/5

laurigates/loractl#203 · opened

on Aug 8, 2026

1

A swallowed cubecl device-task panic is invisible to the trainer — the negative-MSE guard is a proxy, not a signal

laurigates/loractl#181 · opened

on Jul 31, 2026

perf(core): dequantize scaled-fp8 on the GPU instead of the CPU LUT

laurigates/loractl#177 · opened

on Jul 31, 2026

perf(core): the quantized base load is host-serial — parse the header once, parallelize the fp8 dequant, overlap CPU/GPU

laurigates/loractl#176 · opened

on Jul 31, 2026

perf(core): burn-cuda is compiled without `fusion` — A/B it once `just bench` has a baseline

laurigates/loractl#174 · opened

on Jul 31, 2026

1

Navigation Menu

Repository navigation

Issues

[Chore] CI: prune superseded Rust toolchains on the self-hosted GPU runner

feat(trainer): report the observed caption-dropout rate

feat(cli): loss-curve follow-ups — full-run curves across resume, train() wiring test, checkpoint markers

fix(hf): token edge cases — same-host redirects drop Authorization, BOM in token files

chore(ci): follow-ups from the #239 dependency/action bump

test(trainer): pin optim.schedule against a diffusers reference; warn on a resume that changes the decay horizon

feat(config): reject unknown keys across the TrainConfig schema

feat(core): FLUX.2 VAE encode path in burn: shared prerequisite for MOSS-Image and Fibo 2.0

feat(trainer): add Fibo 2.0 LoRA training: 8B single-stream DiT, Qwen3-VL-4B conditioner, FLUX.2 VAE

feat(trainer): add MOSS-Image LoRA training: ~5B DiT, Qwen3.5-9B conditioner, FLUX.2 VAE

Evaluate migrating the CPU encode backend from burn-ndarray to burn-flex

fix(api): POST /runs leaves dataset/model paths unconfined while documenting otherwise

feat(trainer): mid-training image previews on the diffusion path

feat(sample): image sampling for the diffusion task — the flow trainer has no visual feedback loop

feat(dataset): per-tag caption dropout — decorrelate tags that co-occur on every caption

perf(dataset): time the decode fraction in isolation — the encode-phase wall clock cannot see it

chore(core): audit resume.rs against burn_train::checkpoint — record the split, don't migrate blind

feat(cli): drive burn-train's TUI dashboard as a renderer over the TrainEvent stream

docs(gpu-hosting): volume sizing and the empty ram_gb column both assume a Krea 2-shaped workload

feat(trainer): add MiniMax H3 video LoRA training — a second target reusing the Qwen3-VL/MMDiT stack

Run the opt-in qwen3vl-real proof once on a GPU box — #163 merged without ever executing the test that checks 34/5

A swallowed cubecl device-task panic is invisible to the trainer — the negative-MSE guard is a proxy, not a signal

perf(core): dequantize scaled-fp8 on the GPU instead of the CPU LUT

perf(core): the quantized base load is host-serial — parse the header once, parallelize the fp8 dequant, overlap CPU/GPU

perf(core): burn-cuda is compiled without `fusion` — A/B it once `just bench` has a baseline