Issues

• freeze_layers / freeze_ratio are silently ignored by task: sft on the vision, audio, stream_layers, unsloth and mlx paths

  MakazhanAlpamys/Soup#1581 · Filter by author abdulwaarith0 opened on Oct 2, 2026

• soup migrate --from unsloth cannot read a notebook that has !pip or %%capture lines: the whole file is one SyntaxError, so every official Unsloth notebook is refused

  bug
  MakazhanAlpamys/Soup#1579 · Filter by author Srinivasan8888 opened on Oct 2, 2026
  3

• [Feature]: Smart interactive setup wizard (soup wizard / soup init --wizard)

  enhancement
  MakazhanAlpamys/Soup#1576 · Filter by author Deadpool2000 opened on Oct 2, 2026
  3

• tests/test_issue1220_resume_prm_moe.py fails on a CUDA host: hardware-fit preflight predicts 114 GB for a tiny Hub model

  bug
  MakazhanAlpamys/Soup#1574 · Filter by author bobbonauta opened on Oct 2, 2026
  1

• Resuming DPO/IPO, GRPO and Reward runs also upcasts the trainable adapter from bf16 to fp32 (same as #1503)

  bug
  MakazhanAlpamys/Soup#1573 · Filter by author bobbonauta opened on Oct 2, 2026
  1

• tests/test_cli.py::test_version_full_json passes in the full suite but fails when run alone, because torch's first-import warning lands in result.output

  bugclaimedgood first issue
  MakazhanAlpamys/Soup#1551 · Filter by author MakazhanAlpamys opened on Oct 1, 2026
  6

• Make IP-literal parsing independent of the platform and interpreter patch level

  enhancement
  MakazhanAlpamys/Soup#1550 · Filter by author MakazhanAlpamys opened on Oct 1, 2026
  1

• Say what to do in the private-address refusal

  claimedenhancementgood first issue
  MakazhanAlpamys/Soup#1549 · Filter by author MakazhanAlpamys opened on Oct 1, 2026
  3

• One loopback set for every outbound gate

  enhancementgood first issue
  MakazhanAlpamys/Soup#1548 · Filter by author MakazhanAlpamys opened on Oct 1, 2026
  2

• A sink-based ratchet for outbound HTTP calls

  claimedenhancement
  MakazhanAlpamys/Soup#1547 · Filter by author MakazhanAlpamys opened on Oct 1, 2026
  2

• layer streaming: a pinned staging arena can start off a 4096-byte boundary, which silently turns direct I/O off

  bug
  MakazhanAlpamys/Soup#1531 · Filter by author MakazhanAlpamys opened on Oct 1, 2026
  1

• soup data from-traces --logs <file> --format soup-serve reads nothing and ends in a green "Wrote 0 preference pair(s)", exit 0

  bugclaimedgood first issue
  MakazhanAlpamys/Soup#1530 · Filter by author MakazhanAlpamys opened on Oct 1, 2026
  3

• task: unlearn finishes training and then dies with KeyError: 'duration' in the completion panel

  bugclaimedgood first issue
  MakazhanAlpamys/Soup#1529 · Filter by author MakazhanAlpamys opened on Oct 1, 2026
  4

• soup diagnose still reads two unmeasured cases as OK: modes missing from --evidence, and memorization texts too short to split

  claimed
  MakazhanAlpamys/Soup#1525 · Filter by author bakdauletbaktygaliyev opened on Oct 1, 2026
  4

• soup ab: when the accept hold fires after a saturated warm-up, the run can stay continue for the whole dataset and the panel does not say why

  enhancementhelp wanted
  MakazhanAlpamys/Soup#1524 · Filter by author MakazhanAlpamys opened on Oct 1, 2026
  2

• A judge that is down still costs 3 s of backoff per row, so a long judged run does not stop early

  claimedenhancement
  MakazhanAlpamys/Soup#1522 · Filter by author MakazhanAlpamys opened on Oct 1, 2026
  2

• docs: layer-streaming 'bit-exact' claim needs scope note (resident control without prepare_model_for_kbit_training)

  MakazhanAlpamys/Soup#1519 · Filter by author bobbonauta opened on Oct 1, 2026

• Layer streaming ignores the weights when base is a local HF cache snapshot path

  MakazhanAlpamys/Soup#1511 · Filter by author bobbonauta opened on Sep 30, 2026

• Layer streaming refuses models cached by huggingface_hub >= 1.32: Xet blobs live in the cache-wide shared store, which the materializer treats as an escape

  MakazhanAlpamys/Soup#1509 · Filter by author bobbonauta opened on Sep 30, 2026

• Resuming a QLoRA SFT run upcasts the LoRA adapter from bf16 to fp32, so the resumed run cannot reproduce the uninterrupted one

  MakazhanAlpamys/Soup#1503 · Filter by author bobbonauta opened on Sep 30, 2026

• freeze_layers / freeze_ratio are accepted on every task but only task: sft applies them

  bugclaimed
  MakazhanAlpamys/Soup#1497 · Filter by author MakazhanAlpamys opened on Sep 30, 2026
  2

• eval.auto_eval: true does nothing on a real run: no trainer hands the evaluation config to the training callback

  bugclaimed
  MakazhanAlpamys/Soup#1487 · Filter by author MakazhanAlpamys opened on Sep 30, 2026
  2

• Valid-looking configs fail after the model load: quantization: mxfp4, Autopilot DoRA on GPTQ/AWQ, prm_reward Hub ids

  bugclaimedhelp wanted
  MakazhanAlpamys/Soup#1466 · Filter by author MakazhanAlpamys opened on Sep 29, 2026
  7

• Docs promise features and flags that do not exist (standalone sweep.yaml, UI tab plugins, bundled corpora, 8 commands)

  bugclaimedhelp wanted
  MakazhanAlpamys/Soup#1465 · Filter by author MakazhanAlpamys opened on Sep 29, 2026
  4

• Failures skip cleanup: multi-GB export merge dirs, extracted cans, stuck running sweep rows, a live Aider container

  bugclaimed
  MakazhanAlpamys/Soup#1464 · Filter by author MakazhanAlpamys opened on Sep 29, 2026
  2