Commits

Breadcrumbs

History for

onv0.73.1

User selector

All users

Datepicker

All time

Commit history

Commits on Aug 14, 2026

• fix(mlx): write the RESOLVED target modules, not `auto` (#392)

  Show description for 50e1865
  

  MakazhanAlpamys

  and

  claude

  committed
  50e1865View commit details

  Copy full SHA for 50e1865

  View code at this pointBrowse repository at this point

• feat(stream): measure the VRAM fit instead of predicting it (#349)

  Show description for 36e9f71
  MakazhanAlpamys

  committed
  36e9f71View commit details

  Copy full SHA for 36e9f71

  View code at this pointBrowse repository at this point

• fix(test): #389's own test was order-dependent and turned CI red

  Show description for 704f771
  

  MakazhanAlpamys

  and

  claude

  committed
  704f771View commit details

  Copy full SHA for 704f771

  View code at this pointBrowse repository at this point

• fix(errors): name the fix when Colab's preinstalled torchao breaks peft (#389)

  Show description for 8c04757
  

  MakazhanAlpamys

  and

  claude

  committed
  8c04757View commit details

  Copy full SHA for 8c04757

  View code at this pointBrowse repository at this point

• fix(stream): wire calibrated_logits_bytes_per_element into the VRAM pre-flight (#390)

  Show description for 04a0e82
  AmirF194

  authored
  04a0e82View commit details

  Copy full SHA for 04a0e82

  View code at this pointBrowse repository at this point

Commits on Aug 13, 2026

• fix(stream): the panel blamed Windows on a Linux box

  Show description for 213fff2
  

  MakazhanAlpamys

  and

  claude

  committed
  213fff2View commit details

  Copy full SHA for 213fff2

  View code at this pointBrowse repository at this point

• fix(gpu): "bf16 supported" includes EMULATION, so a T4 said yes and got bf16

  Show description for c5f80a5
  

  MakazhanAlpamys

  and

  claude

  committed
  c5f80a5View commit details

  Copy full SHA for c5f80a5

  View code at this pointBrowse repository at this point

• feat(stream): let training.stream_vram_override replace the measured free VRAM (#386)

  Show description for d31c1e5
  AmirF194

  authored
  d31c1e5View commit details

  Copy full SHA for d31c1e5

  View code at this pointBrowse repository at this point

• fix(gpu): a "cuda" string with no CUDA runtime asks for no mixed precision

  Show description for a3f88a1
  

  MakazhanAlpamys

  and

  claude

  committed
  a3f88a1View commit details

  Copy full SHA for a3f88a1

  View code at this pointBrowse repository at this point

• fix(trainer): ask the card whether it has bf16, in all fourteen places (#385, #387)

  Show description for 5793e83
  

  MakazhanAlpamys

  and

  claude

  committed
  5793e83View commit details

  Copy full SHA for 5793e83

  View code at this pointBrowse repository at this point

• fix(stream): canonicalise named_parameters() for streamed models (#384)

  Show description for 6116c81
  AmirF194

  authored
  6116c81View commit details

  Copy full SHA for 6116c81

  View code at this pointBrowse repository at this point

Commits on Aug 12, 2026

• fix(sft): normalise every checkpoint-* adapter under torch.compile, not only the final save (#380)

  Show description for 55f583b
  wilyan09007

  authored
  55f583bView commit details

  Copy full SHA for 55f583b

  View code at this pointBrowse repository at this point

• fix(mlx): say when training.seed is ignored, and credit #381

  Show description for bbda3c7
  

  MakazhanAlpamys

  and

  claude

  committed
  bbda3c7View commit details

  Copy full SHA for bbda3c7

  View code at this pointBrowse repository at this point

Commits on Aug 11, 2026

• fix(trainer): thread training.seed into every task wrapper, and apply it before the adapter (#381)

  Show description for e5e36d6
  wilyan09007

  authored
  e5e36d6View commit details

  Copy full SHA for e5e36d6

  View code at this pointBrowse repository at this point

• test(ci): stop one bitsandbytes crash censoring the whole Windows matrix (#382)

  Show description for 5329bee
  MakazhanAlpamys

  committed
  5329beeView commit details

  Copy full SHA for 5329bee

  View code at this pointBrowse repository at this point

• test(cut-ce): stop asserting against a mock the code never consults

  Show description for 0e77ffd
  

  MakazhanAlpamys

  and

  claude

  committed
  0e77ffdView commit details

  Copy full SHA for 0e77ffd

  View code at this pointBrowse repository at this point

• fix(mlx): dispatch backend=mlx to the MLX trainer and align with mlx-lm 0.31 (#362)

  Show description for 7c9a931
  nicolasramos

  authored
  7c9a931View commit details

  Copy full SHA for 7c9a931

  View code at this pointBrowse repository at this point

Commits on Aug 9, 2026

• release: v0.73.0 — three days on somebody else's hardware

  Show description for 775ed29
  

  MakazhanAlpamys

  and

  claude

  committed
  775ed29View commit details

  Copy full SHA for 775ed29

  View code at this pointBrowse repository at this point

Commits on Aug 8, 2026

• feat(deps): #326 - raise the trl cap to <0.29, on capability probes not a version table

  Show description for cb872ae
  MakazhanAlpamys

  committed
  cb872aeView commit details

  Copy full SHA for cb872ae

  View code at this pointBrowse repository at this point

• feat(train): #341 #340 - a real training seed, and full fine-tuning as lora.r=0

  Show description for dcb5eb5
  MakazhanAlpamys

  committed
  dcb5eb5View commit details

  Copy full SHA for dcb5eb5

  View code at this pointBrowse repository at this point

• fix(eval): #316 - two of ship's three behavioural suites measured nothing

  Show description for c87fd00
  MakazhanAlpamys

  committed
  c87fd00View commit details

  Copy full SHA for c87fd00

  View code at this pointBrowse repository at this point

• fix(online-dpo): #300 - reward_model was passed to a trl that removed it at 0.25

  Show description for 3830fa0
  MakazhanAlpamys

  committed
  3830fa0View commit details

  Copy full SHA for 3830fa0

  View code at this pointBrowse repository at this point

Commits on Aug 7, 2026

• fix(diagnose): #316 - the refusal detector missed the apostrophe models actually type

  Show description for 835dda6
  MakazhanAlpamys

  committed
  835dda6View commit details

  Copy full SHA for 835dda6

  View code at this pointBrowse repository at this point

• fix(flash-attn): #334 - FA3 was selected from a version that can never say 3

  Show description for e25dfac
  MakazhanAlpamys

  committed
  e25dfacView commit details

  Copy full SHA for e25dfac

  View code at this pointBrowse repository at this point

• fix(export): #144 - install the convert deps from the caller, not from the lookup

  Show description for e6aa9b1
  MakazhanAlpamys

  committed
  e6aa9b1View commit details

  Copy full SHA for e6aa9b1

  View code at this pointBrowse repository at this point

• test: two CI-only failures in the new issue suites, both environment assumptions

  Show description for ab7e93c
  MakazhanAlpamys

  committed
  ab7e93cView commit details

  Copy full SHA for ab7e93c

  View code at this pointBrowse repository at this point

• fix(export): #144 - a quantised GGUF export deleted a previously exported f16

  Show description for 3928cc9
  MakazhanAlpamys

  committed
  3928cc9View commit details

  Copy full SHA for 3928cc9

  View code at this pointBrowse repository at this point

• fix(serve): #332 #333 - vLLM now uses the model's chat template, and reports what it did

  Show description for dd7c5bf
  MakazhanAlpamys

  committed
  dd7c5bfView commit details

  Copy full SHA for dd7c5bf

  View code at this pointBrowse repository at this point

• fix(deepspeed): #336 - DeepSpeed could not train a LoRA model on any stage

  Show description for 6d6ee96
  MakazhanAlpamys

  committed
  6d6ee96View commit details

  Copy full SHA for 6d6ee96

  View code at this pointBrowse repository at this point

• fix(liger): detect the architecture from the config, not from the model path

  Show description for 718ee05
  MakazhanAlpamys

  committed
  718ee05View commit details

  Copy full SHA for 718ee05

  View code at this pointBrowse repository at this point

• fix(stream): #327 - name the two terms in LOGITS_BYTES_PER_ELEMENT, and refuse to lower it

  Show description for 5e356ae
  MakazhanAlpamys

  committed
  5e356aeView commit details

  Copy full SHA for 5e356ae

  View code at this pointBrowse repository at this point

• fix(train): #77 - the --no-reexec hint dropped every flag the user typed

  Show description for bcbd0fb
  MakazhanAlpamys

  committed
  bcbd0fbView commit details

  Copy full SHA for bcbd0fb

  View code at this pointBrowse repository at this point

• fix(sft): #335 - use_fsdp2_compile wrote an adapter that reloads as all zeros

  Show description for 966b78e
  MakazhanAlpamys

  committed
  966b78eView commit details

  Copy full SHA for 966b78e

  View code at this pointBrowse repository at this point

• fix(launcher): #77 - soup train --gpus N never launched at all

  Show description for 552c844
  MakazhanAlpamys

  committed
  552c844View commit details

  Copy full SHA for 552c844

  View code at this pointBrowse repository at this point

• test(stream): the CPU bound was transplanted from one stack and CI caught it

  Show description for f6c56f3
  MakazhanAlpamys

  committed
  f6c56f3View commit details

  Copy full SHA for f6c56f3

  View code at this pointBrowse repository at this point