-
Notifications
You must be signed in to change notification settings - Fork 0
All issues
Issue creation is restricted in this repository
Issues
is:issue state:open
is:issue state:open
Search results
A swallowed cubecl device-task panic is invisible to the trainer — the negative-MSE guard is a proxy, not a signal
enhancementNew feature or improvementNew feature or improvementStatus: Open.#181 In laurigates/loractl;feat(trainer): make resume explicit and lossless — target a checkpoint, carry optimizer state and step count, refuse a poisoned artifact
enhancementNew feature or improvementNew feature or improvementStatus: Open.#180 In laurigates/loractl;feat(trainer): warn before the run when a config leaves the measured fit envelope (e.g. resolution > 512 with quant)
enhancementNew feature or improvementNew feature or improvementkrea2Krea 2 image-diffusion LoRA training roadmap (M6+)Krea 2 image-diffusion LoRA training roadmap (M6+)Status: Open.#179 In laurigates/loractl;perf(dataset): parallelize image decode/resize and drop the extra full-frame copy
enhancementNew feature or improvementNew feature or improvementkrea2Krea 2 image-diffusion LoRA training roadmap (M6+)Krea 2 image-diffusion LoRA training roadmap (M6+)Status: Open.#178 In laurigates/loractl;perf(core): dequantize scaled-fp8 on the GPU instead of the CPU LUT
enhancementNew feature or improvementNew feature or improvementStatus: Open.#177 In laurigates/loractl;perf(core): the quantized base load is host-serial — parse the header once, parallelize the fp8 dequant, overlap CPU/GPU
enhancementNew feature or improvementNew feature or improvementStatus: Open.#176 In laurigates/loractl;fix(core):
prepare_datasetholds every example's conditioning device-resident — 60 MiB of VRAM per example, linear in dataset sizebugSomething isn't workingSomething isn't workingkrea2Krea 2 image-diffusion LoRA training roadmap (M6+)Krea 2 image-diffusion LoRA training roadmap (M6+)Status: Open.#175 In laurigates/loractl;perf(core): burn-cuda is compiled without
fusion— A/B it oncejust benchhas a baselineenhancementNew feature or improvementNew feature or improvementStatus: Open.#174 In laurigates/loractl;test(core): Phase events on the merge and MNIST paths are unproven by execution
choreMaintenance, deps, refactoringMaintenance, deps, refactoringStatus: Open.#170 In laurigates/loractl;Pin the caption-conditioning length invariant: PROMPT_PREFIX_LEN / PROMPT_SUFFIX_LEN are hardcoded and unchecked against the tokenizer
bugSomething isn't workingSomething isn't workingStatus: Open.#163 In laurigates/loractl;Record the #110 bench timing-mechanism decision (ADR-0009) — after the first real dispatch
documentationImprovements or additions to documentationImprovements or additions to documentationStatus: Open.#162 In laurigates/loractl;int4 dequant error vs adapter quality: measure what the ~7% worst-case base error does to a trained LoRA
enhancementNew feature or improvementNew feature or improvementStatus: Open.#159 In laurigates/loractl;