Commits · e227e7f591b578419afb23659f687c510c7c07f8 · nexlab / coderai · GitLab

Oct 30, 2026
- make an embedding request visible to the evictor · e227e7f5
  Stefy Lanza (nextime / spora ) authored Oct 30, 2026
  e227e7f5
- a learned layer split must not outlive the KV settings it was learned under · 051cc5b2
  Stefy Lanza (nextime / spora ) authored Oct 30, 2026
  051cc5b2
- do not "repair" a block count that a per-layer array pins · 35e90baa
  Stefy Lanza (nextime / spora ) authored Oct 30, 2026
  35e90baa
- ask a backend for its voices the way backends actually expose them · c33f92ba
  Stefy Lanza (nextime / spora ) authored Oct 30, 2026
  c33f92ba
- decide whether a GGUF is readable before evicting anything for it · 4fc14c34
  Stefy Lanza (nextime / spora ) authored Oct 30, 2026
  4fc14c34
- give the images a C++ compiler, so XTTS can speak Japanese · fe5c785f
  Stefy Lanza (nextime / spora ) authored Oct 30, 2026
  fe5c785f
- declare the phonemisers XTTS needs for Chinese and Japanese · 274ea9a8
  Stefy Lanza (nextime / spora ) authored Oct 30, 2026
  274ea9a8
- stop truncating the one language code that is longer than two letters · 753ac897
  Stefy Lanza (nextime / spora ) authored Oct 30, 2026
  753ac897
- french was never going to work, and a test said it would · 0e8e24e2
  Stefy Lanza (nextime / spora ) authored Oct 30, 2026
  0e8e24e2
- name kokoro's voices without loading it · 40aa3eb0
  Stefy Lanza (nextime / spora ) authored Oct 30, 2026
  40aa3eb0
- say what a voice can speak · 49543895
  Stefy Lanza (nextime / spora ) authored Oct 30, 2026
  49543895
- put the KV levers on the model page, and state the block count both ways · 14d34003
  Stefy Lanza (nextime / spora ) authored Oct 30, 2026
  14d34003
- check the KV size a pin asked for, and write it down before it kills you · 16488fa5
  Stefy Lanza (nextime / spora ) authored Oct 30, 2026
  16488fa5
- decide LongCat's placement from the card, not from whatever else is loaded · 28eb95b6
  Stefy Lanza (nextime / spora ) authored Oct 30, 2026
  28eb95b6
- seed an offloaded pipeline, and stop advertising a direction a model cannot go · e0a3c0f8
  Stefy Lanza (nextime / spora ) authored Oct 30, 2026
  e0a3c0f8
- say which published ids are the same model · 34465d6a
  Stefy Lanza (nextime / spora ) authored Oct 30, 2026
  34465d6a
- build the floor-aware handler through its factory, not a global name · 87463dab
  Stefy Lanza (nextime / spora ) authored Oct 30, 2026
  87463dab
- give a picture enough tokens to be described · 6864eae1
  Stefy Lanza (nextime / spora ) authored Oct 30, 2026
  6864eae1
- return the timings whisper.cpp sends · 56af3f00
  Stefy Lanza (nextime / spora ) authored Oct 30, 2026
  56af3f00
- make the multimodal path visible in the log · 9b37da2f
  Stefy Lanza (nextime / spora ) authored Oct 30, 2026
  9b37da2f
- stop discarding the reason a GGUF would not load, and give every TTS backend the card's path · da335aeb
  Stefy Lanza (nextime / spora ) authored Oct 30, 2026
  da335aeb
- use what the server already knew: precision, the VAE grid, espeak's names, the engine's own name · 9b21a962
  Stefy Lanza (nextime / spora ) authored Oct 30, 2026
  9b21a962
- refuse a load that cannot fit in host RAM, and size the one that was skipping itself · fdbc7701
  Stefy Lanza (nextime / spora ) authored Oct 30, 2026
  fdbc7701
Oct 29, 2026
- query transformers' mapping the way transformers defines it · c6fca1d8
  Stefy Lanza (nextime / spora ) authored Oct 29, 2026
  c6fca1d8
- read a model's card where the card actually lives · 82b647de
  Stefy Lanza (nextime / spora ) authored Oct 29, 2026
  82b647de
- read the config, do not build it, when refusing an image request · 37f1040d
  Stefy Lanza (nextime / spora ) authored Oct 29, 2026
  37f1040d
- load an embedder the way its entry says, and refuse an impossible request before the load · 2833d7f7
  Stefy Lanza (nextime / spora ) authored Oct 29, 2026
  2833d7f7
- pin the torch pair python 3.13 actually has wheels for · 94502ada
  Stefy Lanza (nextime / spora ) authored Oct 29, 2026
  94502ada
- pin torchaudio with torch, and say why an import failed · 495c6616
  Stefy Lanza (nextime / spora ) authored Oct 29, 2026
  495c6616
- let a self-healing rebuild run without the repo-root file it names · f7836901
  Stefy Lanza (nextime / spora ) authored Oct 29, 2026
  f7836901
- pin parler's own CUDA stack, and check a venv can run what the worker needs · f9c05c91
  Stefy Lanza (nextime / spora ) authored Oct 29, 2026
  f9c05c91
- give an isolated worker the environment it needs, and a resident speech model room to run · 3d507226
  Stefy Lanza (nextime / spora ) authored Oct 29, 2026
  3d507226
- load the vision tower a checkpoint asks for, and refuse the pictures we cannot read · 8fe05edc
  Stefy Lanza (nextime / spora ) authored Oct 29, 2026
  8fe05edc
- measure a model's memory from the model, and release what it leaves behind · 8274ff3e
  Stefy Lanza (nextime / spora ) authored Oct 29, 2026
  8274ff3e
- stamp the LRU clock when a model is registered by raw dict write · 1b83dc27
  Stefy Lanza (nextime / spora ) authored Oct 29, 2026
  1b83dc27
- make the quantization signature actually reach disk · 97256023
  Stefy Lanza (nextime / spora ) authored Oct 29, 2026
  97256023
- stop a VRAM measurement outliving the quantization it was taken under · 8303b38e
  Stefy Lanza (nextime / spora ) authored Oct 29, 2026
  8303b38e
- quantize what load_in_4bit was asked to quantize, and say why a load spills · bb1b78e3
  Stefy Lanza (nextime / spora ) authored Oct 29, 2026
  bb1b78e3
- write the learned verdict to disk, not just to memory · 5fbc19ae
  Stefy Lanza (nextime / spora ) authored Oct 29, 2026
  5fbc19ae
- release the injected components BEFORE the teardown, not after · ecf9bbec
  Stefy Lanza (nextime / spora ) authored Oct 29, 2026
  ecf9bbec