litetune v0.5.1

Changelog

Newest first. Full notes for each release are on GitHub.

0.5.1 — 2026-10-19

  • prepare and verify say a noul question has been measured on two tasks, not one, and MEASUREMENTS.md adds SST-2 as noul and irony trained for four and six epochs, which verify passed where two epochs were refused (#85).
  • bundle no longer says it records nothing that was measured beside the measurements --verify-manifest put in report.json: the limitation is now about the synthesised manifest.json, which lists no stages, and the bundle no longer claims a run's record it was not given (#86).

0.5.0 — 2026-10-18

  • prepare --questions, tune --decision, verify and bundle train, measure and package a model that answers with one of the options an app passes, for flutter_edge_ai's proposed InferenceModel.createDecider: one .litertlm that carries the question, the options and the calibration in a litetune.decision section, measured on google/gemma-3-270m-it only (#82).
  • convert exports a checkpoint that records a decision with --prefill_lengths=8,128 unless the caller sets it (#82).
  • bundle writes a decision's section only when verify passed or was inconclusive; with any other status it ships the file without one, taking out a section an earlier bundle left, and exits with verify's status (#82, #83).
  • convert reports a litetune.json beside the checkpoint that is there but cannot be read -- a directory, a dangling link, a FIFO or a device -- as a limitation, where it took it for absent (#80).

0.4.3 — 2026-10-14

  • Convert a Qwen3 bundle again if an application holds a conversation with it: every Qwen3 bundle earlier litetune converted without a template of the caller's carries Qwen's own chat template, which LiteRT-LM's conversation cannot render past the first message, so the second message fails (#78).
  • convert checks that each bundle's chat template renders single turns as the checkpoint was trained and holds a LiteRT-LM conversation, says in its report when one does not, and gives Qwen3 LiteRT-LM's own Qwen 3 template (#78).
  • On Windows each stage runs in a job object: a timeout, Ctrl-C, Ctrl-Break or a forced kill of litetune stops everything the stage started, and Ctrl-Break exits 4 (#77).
  • On Windows a stage that closes its pipes and then hangs keeps its output, and on Python 3.10 a usage error with stderr closed exits 4 rather than 120 (#77).

0.4.2 — 2026-10-12

  • A stage whose pinned packages publish no wheel for this platform and Python is refused before pip, naming the package: convert on Windows, a 32-bit or ARM64 Python on Windows, Intel macOS, macOS before 12 and older glibc; LITETUNE_ALLOW_MISSING_WHEELS=1 lets pip try anyway, and litetune now depends on packaging (#74).
  • tune runs on Python 3.13 on Linux x86_64, with sentencepiece 0.2.1 there and 0.2.0 on 3.10 – 3.12 as before (#74).
  • On Windows, JSON written by PowerShell is read, console output and help no longer raise on characters the code page lacks, convert touches no checkpoint before its environment stands, and the test suite runs in CI (#74).
  • verify's float reference decodes greedily without the checkpoint's own generation_config penalties, and names any it left out; the Qwen2.5 references measured here applied repetition_penalty 1.1, and re-measured without it their figures are unchanged (#73).

0.4.1 — 2026-10-12

  • A Qwen3 0.6B or Qwen2.5 0.5B Instruct bundle litetune 0.4.0 converted from a checkpoint it did not recognise -- another organisation's copy, or a fine-tune under its own name -- has no block that reads LiteRT-LM 0.18.0's parts: convert it again; one converted from Qwen's own repositories at their current revisions, or what tune made from them, is unaffected (#71).
  • convert puts the text-parts block in front of the checkpoint's own chat template instead of a packaged copy, and reads the template back from every bundle it writes (#71).
  • Exporter flags passed with --flag= are read as the exporter reads them and forms it would read differently are refused; Gemma 4 refuses any model-type override and an embedder opt-out (#71).
  • Stage environments link the interpreter instead of copying it, so a Python that loads libpython relative to itself, such as uv 0.10.6's CPython on macOS, can provision them (#70).

0.4.0 — 2026-10-11

  • Convert a Qwen3 0.6B or Qwen2.5 0.5B Instruct bundle again for an app on LiteRT-LM 0.18.0, as flutter_edge_ai_litertlm 1.10.0 ships: 0.18.0 hands the template a list of parts, which their checkpoint templates do not read, and the user's text is lost or every send fails (#67).
  • Convert again any bundle convert wrote to a relative --output-dir on Python 3.10 or 3.11: it has no GPU activation key, which its report named, and on a GPU backend it answers wrong (#67).
  • verify runs the candidate on litert-lm 0.18.0 and convert packs with litert-lm-builder 0.18.0; every published number re-measured on it equals its 0.17.1 number, Qwen's on bundles converted again (#67).
  • The Qwen3 0.6B and Qwen2.5 0.5B Instruct rules pack the checkpoint's own chat template behind a block that reads 0.18.0's parts, and verify names a rendering that came back empty or without the prompt (#67).
  • litetune towers drops a Gemma 4 bundle's vision and audio towers, or adds them from another bundle (#68).
  • The site and README say which device, backend and runtime each number was measured on (#65), and FunctionGemma's measurement notes describe Flutter Edge AI's native tool path (#66).

0.3.0 — 2026-10-03

  • Re-run a timed-out verify only if you rely on its reference-backend provenance: 0.1.5 through 0.2.1 could record unknown after the float-reference model had already reached a device; generated answers and scores are unaffected (#63).
  • tune and verify's float reference can use Apple GPUs through Metal/MPS with unified-memory safeguards; tune --micro-batch-size exposes gradient accumulation, and LITETUNE_DEVICE=cpu forces CPU execution (#63).

0.2.1 — 2026-09-29

  • The four-bit recipes on the four banking77 checkpoints are re-measured on litert-lm 0.17.1, beside their 0.16.1 numbers (#60).
  • The tuned Gemma 4 E2B's eight-bit pair is re-measured on litert-lm 0.17.1, beside its 0.16.1 numbers (#61).
  • The site's Qwen2.5-0.5B card no longer calls its channelwise four-bit scores usable (#60).

0.2.0 — 2026-09-28

  • Re-run verify for any model you ship on litert-lm 0.17.1, as flutter_gemma_litertlm 1.8.0 does — 0.1.x measured on 0.16.1, and the two runtimes do not give the same answers (#55).
  • verify runs the candidate on litert-lm 0.17.1, and convert pins litert-torch-nightly 0.10.0.dev20260926 with litert-lm-builder 0.17.1 and the quantizer, LiteRT and converter packages it requires (#55).
  • Every verify manifest records the litert-lm version its candidate ran on and the host's CPU (only on the CPU backend is that what computed the answers), and says so when it could not read either (#55).
  • bundle records in established_against the runtime and CPU its verify manifest recorded, with the CPU's features where the manifest has them (#55, #56).
  • The tool path refuses grammar-on decoding for a bundle without a SentencePiece tokenizer, which litert-lm 0.17.1 would decode without the grammar (#55).
  • tune accepts a tool call whose integer argument fits in 64 bits but not in a double, as litert-lm 0.17.1 returns it exact (#55).
  • FunctionGemma's tool path and the four banking77 eight-bit pairs are re-measured on litert-lm 0.17.1, beside their 0.16.1 numbers (#56, #57).
  • The published numbers say which runtime they were taken on; the four-bit recipes and Gemma 4 are not yet re-measured (#55, #58).

0.1.9 — 2026-09-26

  • verify --backend gpu measures the converted model on litert-lm's GPU backend, passing fp32 activations itself and naming it when the bundle alone would get something else (#53).
  • On macOS a GPU run is recorded as measured on the GPU only when the kernel shows its process doing GPU work; elsewhere, and when the reading is not conclusive, the manifest says the GPU was asked for (#53).
  • Every manifest says whether the candidate's backend was observed or only asked for (#52).
  • verify generates a whole split in one litert-lm process through the runtime's Python API rather than one litert-lm run per prompt; the split's budget is still the per-prompt timeout times its prompts, so one hung prompt can use the rest's share (#50).
  • A LoRA run on Gemma 4 adapts the text tower only, where before it stopped in peft on the vision and audio projections, and metrics.json records what peft matched (#48).
  • Gemma 4 E2B is measured end to end on banking77, and litetune's seven LoRA projections against peft's default two (#48).
  • Qwen2.5-0.5B-Instruct is a checked family and is measured on banking77 (#47).
  • The site renders MEASUREMENTS.md, and each measured model's card links to the section that measured it (#49).
  • Each measured model's card says what the model is and what it is for (#49, #51).

0.1.8 — 2026-09-20

  • Re-run any verify whose host does not default to UTF-8 — Windows always, POSIX with a locale like en_US.ISO8859-1 — the runtime's generations were decoded in the host's encoding, so a non-ASCII answer scored as a wrong one (#44).
  • Re-make any bundle or split written on Windows — text writes turned LF into CRLF, so a bundle's declarations did not hash to its own contract and a split's content_sha256 differed from the same rows elsewhere (#44).
  • Re-run any stage that crashed on Windows and was reported as a failed model — an exit code carrying a terminating NTSTATUS is now read as a kill, which is could not check (#44).
  • A generation whose bytes did not decode is a harness error, not a scored answer; U+FFFD from the model still scores (#44).
  • Stage subprocesses are given UTF-8 on both sides of the pipe, and a host value that would decide it is replaced and said (#44).
  • prepare, tune, verify and bundle run on Windows; convert cannot, and the refusal now names litert-converter and what to do instead (#45).
  • A failed install names the distribution pip could not find rather than printing its trace (#45).
  • gemma-3-1b-it on banking77, measured: four bits cost it 8.83 points against the 270M's 34.83 (#43).

0.1.7 — 2026-09-19

  • Re-prepare and re-train any FunctionGemma tool-call split made with 0.1.6 or earlier — its calls carry no call markers, so the runtime's tool path returns no call (#41).
  • Re-run any runtime_rendered verify against a reference whose chat template writes <bos> itself, such as Gemma 3 and FunctionGemma — the reference was prompted with two (#40).
  • Re-train runtime_rendered any checkpoint trained prerendered on bare prompts that an app serves through its chat template — neither training nor verify saw that template (#40).
  • Breaking: tune refuses a --prompt-mode its training prompts contradict; --force-prompt-mode overrides and records it (#40).
  • --prompt-mode is optional: tune reads it off the prompts, verify and bundle from the training record, and each refuses a value that disagrees (#40).
  • verify in runtime_rendered checks that both sides get the same token ids before it generates (#40).
  • Reasoning blocks are removed from both sides before scoring (#40).
  • Qwen3-0.6B is a checked family, dynamic_wi4b32_emb8_afp32 is litetune's own recipe, and both are measured on banking77 (#40).
  • Tool declarations are an input to prepare, tune, verify and bundle; shapes the runtime and the chat template render differently are refused (#41).
  • verify measures a model whose runtime renders its declarations through the runtime's tool path, grammar off and on; a reply the runtime cannot parse is a wrong answer (#41).
  • tune refuses a completion the runtime would not read back as its call, and a call to a tool the split does not declare (#41).
  • bundle ships declarations in the order the model learned, and refuses a set the training run did not record (#41).
  • Scoring ends a Gemma 4 turn at <turn|> and <|tool_response>, so the comparisons 0.1.6 refused now run, and bundles list both as stop tokens (#37).
  • The README's NPU mask boundary is per model, not a formula to carry to another (#38).

0.1.6 — 2026-09-13

  • Re-run any tune, convert or verify done with PYTHONPATH set — it outranked the stage's pinned packages while the manifest named the pin (#33).
  • Run litetune env --clean if PIP_TARGET, PIP_PREFIX or PIP_ROOT was set when a stage first provisioned — pip installed elsewhere and the environment was marked ready (#33).
  • A timeout, Ctrl-C, kill or a dropped connection stops everything the stage spawned (#33, #34).
  • A stage's last output is no longer cut off at one pipe buffer (#34).
  • pip install during provisioning runs under the same guard (#34).
  • Ready means the marker is a file and the interpreter exists (#33, #34).

0.1.5 — 2026-09-08

  • Re-run anything scored with --scorer exact-text on 0.1.4 or earlier — the reference ended in a turn marker and scored 0.0000 (#18).
  • Re-run any tune done on a CUDA machine — it trained on the CPU (#19).
  • An accelerator failure is recorded as a fact about the machine, not a failed fine-tune (#19).
  • gemma-3-270m-it on banking77, measured (#22).

0.1.4 — 2026-09-05

  • convert writes prefer_activation_type = fp32; without it the GPU floods <pad>.
  • convert --json records each bundle's exports[].gpu_activation.

0.1.3 — 2026-09-04

  • The native tool path works; 0.1.2 bundles carried a template LiteRT-LM cannot execute.
  • tune writes litetune.json, so convert applies that family's export flags.
  • An unidentified gemma3_text checkpoint is refused, not guessed.
  • --base-model and --train-metrics identify a checkpoint trained elsewhere.

0.1.2 — 2026-09-04

  • litetune env lists the cached stage environments; --clean removes them.
  • The README's disk figures are measured rather than guessed.

0.1.1 — 2026-09-04

  • verify --scorer chooses what correct means: tool-call or exact-text.

0.1.0 — 2026-09-04

  • First release: prepare, tune, convert, verify, bundle; every check passed, failed or could-not-check, and the exit code is the verdict.
  • Measured on functiongemma-270m-it with google/mobile-actions.