source.index("_resolve_inherited_extra_args(") matched the function
definition, which always precedes the endpoint, so the ordering
assertion was vacuously true. Anchoring on "= _resolve_inherited_
extra_args(" pins the first call site inside the load endpoint (line
4505), which is the statement whose position relative to the GGUF
branch the test is meant to guard. 32 tests pass.
Main moved the llama_extra_args pass-through inheritance out of the
GGUF branch into _resolve_inherited_extra_args, which runs before it,
so the source-order assertion's "if request.llama_extra_args is None"
anchor no longer exists inside the branch and the check failed after
the main merge. The test now asserts the same property in the current
shape: inheritance before the GGUF branch (a carried --no-mmproj still
shapes the hub guard's companion requirement), and marker, hub guard,
unload in order within the branch. Full file passes (32 tests).
* Studio: allow torch 2.11.x on the CUDA install path
The CUDA torch repair path (_ensure_cuda_torch) installs torch/torchvision/
torchaudio from an exclusive --index-url, so _CUDA_TORCH_PKG_SPEC decides
exactly which torch the Studio venv gets. It was capped at torch<2.11.0, so on
a cu128/cu130 host the venv resolved torch 2.10.x even though the CUDA indexes
now publish torch 2.11.0. That left the Studio venv a torch minor behind the
torch 2.11.0 Docker base image, so the CUDA dedup step would relink base libs
under a mismatched torch.
Raise the upper bound to <2.12.0 (torchvision <0.27.0, torchaudio <2.12.0) so
the CUDA install path lands on torch 2.11.x, matching the rocm7.2 spec and the
base image. The torchao selector already maps torch 2.11 -> torchao 0.17.0, and
_ensure_flash_attn degrades gracefully when no prebuilt wheel matches (Blackwell
skips it outright; non-Blackwell prints a warning and continues), so no other
pin needs to move.
Add test_cuda_torch_spec.py to lock the bound (torch 2.11.x in, 2.12.x out) and
assert the CUDA and rocm7.2 upper bounds stay in lockstep.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* test: use zip(strict=True) so a spec length mismatch fails loudly
* install.sh: widen the CUDA torch ceiling to <2.12.0 so a fresh install matches the base
Raising _CUDA_TORCH_PKG_SPEC alone was not enough: that spec only feeds
_ensure_cuda_torch(), the ROCm-poisoning repair path that early-returns on a
normal NVIDIA host. A fresh CUDA install (including the studio Docker build,
which runs `bash install.sh --local`) takes its torch from install.sh's
TORCH_CONSTRAINT, which was still capped at torch>=2.4,<2.11.0, so cu12x/cu13x
resolved torch 2.10.x and the venv landed a minor behind the torch 2.11.0 base
image.
Extend the existing `case "$TORCH_INDEX_URL"` block (which already relaxes
rocm7.2) with a `*/cu[0-9]*` branch that widens the ceiling to <2.12.0, keeping
the >=2.4 floor so an older CUDA index (e.g. cu118) that tops out below 2.11
still resolves. The CPU wheel and older ROCm tags stay on <2.11.0 (the glob
does not match /cpu). torchvision/torchaudio are bare on this install line and
resolve their compatible companions via wheel metadata, matching the rocm7.2
pattern.
Add behavioral tests (Python + shell) exercising the case block: cu118/124/126/
128/130 widen to <2.12.0, rocm7.2 stays 2.11.x, and /cpu plus older ROCm keep
the default <2.11.0.
* install.sh: key the CUDA torch widening off the index leaf, not the full URL
The `*/cu[0-9]*` glob matched a `cu<digit>` segment anywhere in TORCH_INDEX_URL,
so a custom UNSLOTH_PYTORCH_MIRROR whose base path contains e.g. cu128 but whose
final leaf is cpu or an older ROCm tag would still widen TORCH_CONSTRAINT to
<2.12.0, contradicting the block's own comment and letting a CPU / older-ROCm
mirror resolve torch 2.11.x. Match on _torch_index_leaf (the final path segment
the backend classification just above already computes) so only a real cu*/
rocm7.2 leaf is affected; cpu and older ROCm keep the default <2.11.0. Update
the Python + shell tests to mirror the leaf-anchored case and add regression
cases for a mirror base that contains cu128 but resolves to a cpu / rocm7.1 leaf.
* install: freeze the torch trio during the with-deps unsloth installs
Released unsloth wheels can pin an older torch than Step 1 installed
(unsloth 2026.7.2 declares torch<2.11.0), so the with-deps resolve from
PyPI silently downgrades the pinned +cuXXX torch trio to PyPI's default
wheel. The flavor guard cannot catch every such swap: PyPI's torch 2.10
default is itself cu128-flavored, so the cuXXX tag comparison still
matches while the version silently drops. Freeze the just-installed trio
with uv --overrides (overrides replace dependency requirements during
resolution), keeping torch 2.11.0+cuXXX in place while unsloth's other
dependencies resolve normally. Verified on the cu128 path: without the
override torch drops 2.11.0+cu128 -> 2.10.0; with it the trio survives
and unsloth 2026.7.2 + unsloth-zoo install cleanly.
* install: fold UV_OVERRIDE env files into the torch-trio overrides file
The CLI --overrides flag is the command-line form of UV_OVERRIDE, so
passing it replaced any overrides file already exported for the process;
macOS arm64 exports UV_OVERRIDE=overrides-darwin-arm64.txt for the same
generic install path and would have lost those pins. Concatenate any
UV_OVERRIDE files into the temp trio file so both keep applying.
* install: extend the torch-trio overrides guard to migrated installs
Four follow-ups to the Step-2 --overrides guard, all empirically verified:
1. The migrated-environment with-deps unsloth install resolved
unsloth>=2026.7.2 (which pins torch<2.11.0) without the overrides file,
so a migrated CUDA venv on torch 2.11 was silently downgraded -- the
exact bug this branch fixes on the fresh path. The overrides build is
now a function (_build_unsloth_torch_overrides, reading the trio
installed at call time) invoked by both with-deps paths; the migrated
no-torch path installs --no-deps and stays unguarded.
2. The overrides temp file is now cleaned by the EXIT trap (same pattern
as _UV_OVERRIDE_TMPDIR, pre-initialized empty so an inherited value can
never reach the trap's rm); previously any Step-2 failure leaked it.
3. Folding UV_OVERRIDE files used cat, which joins the last requirement of
a file lacking a trailing newline onto the next file's first requirement
(reproduced: idna==3.10certifi==2025.1.31 makes uv fail parsing).
4. Inherited torch/torchvision/torchaudio override lines are now filtered
out when folding: uv intersects duplicate overrides rather than
last-wins (verified on uv 0.10.12: direct conflict is unsatisfiable,
transitive conflict silently backtracks), so a conflicting inherited
trio pin would break the resolve the generated exact pins protect.
Both 3 and 4 are handled by a single newline-terminating awk filter
that preserves non-trio overrides (torchmetrics, torchao, ...).
test_unsloth_torch_override.sh extended: migrated-path coverage, trap
assertion, and a functional fold test (14 checks).
* installer: tighten comments
* install: keep the existing torch release when re-running the installer
Re-running `curl -fsSL https://unsloth.ai/install.sh | sh` over an existing
install rebuilds the venv for clean state, which silently moved users to the
newest torch in range (2.10 -> 2.11 once the constraint widened). A torch the
user already validated must survive an unsloth update.
Before the old venv is moved aside for rollback, its torch version is probed
(last stdout line only, so sitecustomize noise cannot corrupt it). After the
index leaf is chosen, _previous_torch_pin turns that version into a
torch==X.Y.Z pin, but only when it cannot do harm:
- cu*/cpu leaves only; rocm leaves keep their floors (rocm7.2 must land 2.11
for the Strix _grouped_mm fix) and the Radeon wheel-matching path is
untouched.
- The wheel's flavor tag must match the freshly chosen leaf, so a flavor
change (cpu -> cuda, cu126 -> cu130) still installs the correct new build.
- The base must look like a release, so probe noise never becomes a pin.
- UNSLOTH_TORCH_UPGRADE=1 opts out and restores the old always-newest
behavior; the substep line advertises it.
The supported range is kept in _PREV_FALLBACK_CONSTRAINT: if the exact
release is not resolvable from the chosen index (custom mirrors prune old
wheels), the install warns and falls back to the newest supported release
instead of failing the whole run. The later flavor-mismatch repair reuses
TORCH_CONSTRAINT, so a mid-install clobber is repaired back to the kept
release rather than the newest one.
Verified end to end: a venv seeded with torch 2.10.0+cu130 re-run through the
full installer finishes with torch 2.10.0+cu130 (previously 2.11.0+cu130).
Tests: tests/sh/test_previous_torch_pin.sh covers keep/flavor-change/rocm/
noise/opt-out plus wiring (probe ordering before venv replacement, fallback
present, SKIP_TORCH gate).
* install: constrain kept torch pins to the supported window
Review caught that _previous_torch_pin pinned the previous venv's torch on
flavor match alone, so a release outside the installer's active range (a
2.3.x manual install below the >=2.4 floor, or a 2.12.x manual upgrade above
the ceiling) replaced the bounds computed just above it and a rerun kept a
torch the installer otherwise deliberately excludes.
New _torch_release_in_window checks the probed base against the active
TORCH_CONSTRAINT ("torch>=A.B[,<C.D.F]") at major.minor granularity, which
is exact for the windows this script uses (ceilings are always X.Y.0; a
non-.0 ceiling would only make it conservative). Anything unparseable
answers no, so probe noise or a malformed window fails toward the supported
range instead of becoming a pin. _previous_torch_pin takes the active
constraint as a third argument and refuses out-of-window releases; the
in-window keep behavior is unchanged.
Tests: out-of-window rows (2.3.x floor, 2.12.x ceiling, boundary keeps, cpu
and macOS windows, malformed/empty windows) plus direct
_torch_release_in_window coverage.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Raise the opencode invoke timeout in Local Agent Guides CI
The connection (opencode) cell flakes with a 600s timeout reported as guide drift, but it is not a hang: in a passing run the same opencode run finishes in ~482s (08:12:31 to 08:20:33), right against the shared AGENT_INVOKE_TIMEOUT of 600s, so about one run in six drifts past the cap.
opencode is the slow outlier. The print-mode agents (claude -p, codex exec) run one turn against a minimal injected system prompt, while opencode run runs its own full turn with opencode's large system prompt plus a separate small_model call to name the session (start.py pins small_model to the same 4B the server hosts). On a CPU-served gemma-4-E4B that is about 8 minutes, leaving no margin under 600s.
Double opencode's per-invoke timeout in agent-guides-drive.sh and keep the tight 600s cap for the fast agents, so a genuine headless-TTY hang still fails quickly. 1200s stays well under the 40-minute job budget.
* Normalize the agent invoke timeout before doubling it for opencode
Strip an optional trailing 's' from AGENT_INVOKE_TIMEOUT so the opencode
arithmetic, and the "${TIMEOUT}s" timeout message, stay valid if a
timeout(1)-style suffix is ever configured.
* Only double the opencode timeout for a bare-integer seconds value
Guard the arithmetic so a GNU timeout(1) duration suffix (s/m/h/d, including
floats like 0.5s) is passed through unchanged instead of breaking the
expansion; timeout(1) parses those directly. Bare seconds still double.
---------
Co-authored-by: danielhanchen <unslothai@gmail.com>
* Studio: GPU memory dropdown — llama.cpp --fit on and manual gpu-layers/cpu-moe
* Studio: simplify GPU memory changes (reuse ParamSlider, GPU_LAYERS_ALL, loadedGpuMemoryFields helper)
* Studio: GPU picker — choose which GPUs a GGUF model loads on (gpu_ids)
* Studio: simplify GPU picker (share /api/system fetch, validate gpu_ids)
* Studio: GPU picker review fixes (gate relative indices, no cross-model leak, validate, types)
* Studio: group GPU controls under a collapsible GPU section
* Studio: GPU feature review fixes (fix fit-ctx test, behavior-test the floor, comment accuracy)
* Studio: make GPU a top-level settings section (not nested under Model)
* Studio: flatten GPU controls into the Model section, group by GPU/context/generation
* Studio: move GPU Memory to the bottom of Model with its dependent controls beneath it
* Studio: move GPU Memory below Tensor Parallelism and GPUs below GPU Memory
* Studio: tighten GPU Memory and GPU Layers tooltip copy
* Studio: fix fit-mode context slider track-click, restore GPU Memory tooltip, shorten fit dropdown label
* Studio: GPU Memory tooltip one mode per line, briefer
* Studio: note HIP_VISIBLE_DEVICES (ROCm) in the GPUs picker tooltip
* Studio: narrow the GPU Memory dropdown to fit the shortened label
* Studio: use 'llama.cpp --fit' in the GPU Memory tooltip for consistency
* Studio: allow Tensor Parallelism in Manual GPU mode
* Studio: graduated MoE-on-CPU offload (--n-cpu-moe) replacing the all-or-nothing toggle
* Studio: size the MoE-offload slider for staged (deferred-load) models
* Studio: share one GGUF header walk for the context-length and MoE-count readers
* Studio: size the GPU Layers slider for staged models (one staged-header read)
* Studio: move Tensor Parallelism below the GPUs picker
* Studio: GPU split (--tensor-split) per-GPU model share in Manual mode
* Studio: tolerate whitespace in GPU split input, move it below GPU Layers
* Studio: rename the GPU split control to "Split ratio"
* Studio: Split ratio sends explicit even input; fix blank=free-VRAM (not even) copy
* Studio: tighten llama.cpp --fit VRAM margin with --fit-target 512
* Studio: GPU memory review fixes (rollback re-baseline, single-GPU TP gate, accurate copy)
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: move Split ratio below MoE Layers on CPU
* Studio: address PR review (fix GPU-info hydration race, share fit context-length across load paths)
* Studio: address codex review (manual single-GPU TP guard, GPU-aware spec defaults in fit/manual, GGUF-only context/preference)
* Studio: address codex review round 2 (gpu_present seed, single-GPU tensor-split guard, staged manual-knob reset, strip inherited offload flags)
* Studio: address codex review round 3 (strip inherited --n-cpu-moe, CPU-fallback warning in Manual mode)
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: address codex review round 4 (preserve pinned fit context across a later Apply)
* Studio: address codex review round 5 (honor GPU picker for diffusion GGUFs, clear fit pin on cross-model switch)
* Studio: preserve the pending GPU Memory mode when staging a model
* Studio: pin diffusion GPU device order and reset GPU-memory state for diffusion loads
* Studio: address codex review round 6 (fit-Auto rollback context, preserve manual non-tensor split modes, persist GPU mode on load not select)
* Studio: persist the applied GPU Memory mode, not the requested one (skip diffusion loads)
* Studio: replace Manual-mode split-ratio field with per-GPU layer sliders
* Studio: clarify per-GPU layer split hint for tensor-parallel mode
* Studio: address codex review round 7 (allow GGUF gpu_ids past the legacy guard, replay GPU-memory fields on respawn)
* Studio: address codex review round 8 (size the validate preflight like the load in fit mode, across both load paths)
* Studio: skip the training-OOM guard for llama.cpp --fit GGUF loads (they spill to RAM)
* Studio: drop the now-redundant compare-path validate sizing (the --fit guard skip makes it moot)
* Studio: address codex review round 9 (keep the training guard for fit loads, forward gpu_ids to validate, strip inherited manual tensor-split)
* Studio: address codex review round 10 (gate GPU-memory adoption on is_gguf, record manual knobs only in Manual mode)
* Studio: handle diffusion GGUFs symmetrically in the GPU Memory controls (preserve the standing mode preference, hide the inapplicable mode/TP controls)
* Studio: remember the GPU Memory settings per model
* Studio: consolidate --fit mode and Manual mode into a single Manual mode
* Studio: preserve the per-GPU layer split across GPU Layers changes
* Studio: trim overly long GPU Memory comments
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* address GPU memory config review comments
* trim redundant GPU memory tests
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Reconcile manual-mode TP drops with the #6659 drop-site invariants
* Preserve quantized KV in manual --fit, charge GGUF companions in full, reconcile GPU pick on load
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Clear stale GPU baseline on non-GGUF loads so it can't read as dirty
* Fix no-context-shift test for the conditional -c flag
* Credit manual GPU-layer offload for cached HF GGUFs
* Reset per-model load knobs on GGUF quant switch
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Strip inherited tensor-split when manual ratio is cleared
* Match auto-load validation to safetensors placement
* Reset editable manual knobs after Auto GGUF loads
* Record a single device for diffusion GPU picks
* Reset per-model GPU knobs before applying saved settings
* Address review comments
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Guard manual tensor splits and keep remembered context on auto-load
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Snapshot compare knobs, seed splits from free VRAM, flag zero-offload loads
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Exempt CPU-only loads from the guard floor and harden compare and reseed paths
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Reach full offload from the layers slider and charge extras drafters in the guard
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Warm the GPU device cache before pick reconciles and disable staged GPU controls
* Align the training guard with inherited extras, spec mode, and compare targets
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Hide GPUs from companion-less zero-offload loads
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Size diffusion picks per device, own manual offload flags, reject XPU picks
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Drop tensor flags at zero layers and exempt CPU-pinned drafters
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Allowlist the zero-layer tensor parallel drop site
* Keep validate and load guards on the same extras and refresh stale baselines
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Drop mismatched manual tensor splits before launch
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Gate XPU picks on the real backend field and harden split and hydration paths
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Weight full GPUs as zero, clamp split shares, and refine the zero-layer mask gate
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Carry fit context across mode changes and align drafter and picker gates
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Catch variant switches, uncached diffusion repos, and text-only mmproj skips
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Check companions on the first device and size native and remote zero-layer loads
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Replace the training guard's precise VRAM modeling with a conservative bound
* Baseline context pins on non-GGUF hydration and reprobe list-seeded staged GGUFs
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Size manual splits by their largest share and preserve resolved context from Default
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Default-deny unsized required companions and price KV at the effective cache dtype
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Reserve MTP draft KV and MLA target-copy in the training guard
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Size tensor-parallel loads per device and show GPU controls for native GGUFs
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Reserve MTP overhead for uncached remote GGUFs and the mmproj runtime factor
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Drop the training-coexistence VRAM estimation this PR added
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Gate remembered load settings to GGUF picks
* Lock the remaining load-time controls during a staged load
* Clear the stale native-path token on compare loads
* Drop a stale guard reference from the zero-offload masking comment
* Seed GPU baselines from the rollback response and drop never-emitted offload flags
* Match validate's training guard to load and keep the native reload token
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Trim verbose GPU-memory comments
* Thread the variants header walk off the event loop, honor device pins on zero-offload, and hold staged GPU edits
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Honor manual placement and classify pinned zero-offload loads
* Close diffusion admission and status hydration gaps
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Check the actual diffusion GPU during training
* Align staged baselines and manual reload dedupe
* Fix GGUF placement and rollback state
* Harden manual GGUF placement boundaries
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Remove unused resolve_tensor_parallel import in llama_cpp.py
The name is used only in llama_server_args.py, routes/inference.py, and tests,
not in llama_cpp.py; the unused hoisted import trips the import-hoist verifier
in the source-lint CI job.
* Fix diffusion GPU dedup and training guard for non-numeric device tokens
The diffusion runner drives only its single lowest device and the backend
records that one device (self._gpu_ids = [sorted(gpu_ids)[0]]), but the reload
dedupe compared it against the full requested list, so a multi-GPU pick that
resolves to the same device forced a needless reload. Normalize the request the
same way for a loaded diffusion model in both _already_in_target_state and the
route _request_matches_loaded_settings.
The chat-during-training coexistence guard called int() on the single-device
token and hard-rejected when it could not parse. A non-numeric token (a CUDA
UUID / MIG handle) now sizes against the whole visible pool like the GGUF guard
instead of falsely blocking the load, and an empty token (a CPU-only runner such
as a CPU diffusion GGUF) is allowed outright since it uses no GPU VRAM.
* Tighten comments added by the GPU memory config changes
* Harden GGUF placement from independent review: VRAM sizing, diffusion TP reset, tensor_split validation
- Training coexistence guard: a single-device runner pinned through an
unresolvable UUID/MIG token was sized against the aggregate visible-VRAM pool,
so a load could pass on capacity it cannot use and then OOM active training.
Size against the worst-case visible device (min free) instead, keeping the
guard's documented default-deny contract. The empty-token (CPU-only runner)
allow path is unchanged.
- Diffusion startup: _start_diffusion_server now resets self._tensor_parallel to
False alongside the other placement resets. A prior tensor-parallel chat load
(process killed but not fully unload-reset) otherwise left /status misreporting
tensor parallelism and made an identical diffusion re-Apply reload against the
stale state.
- tensor_split: reject negative / non-finite / all-zero splits up front. They
were dropped at launch but still compared raw in the reload dedupe, so an
identical Apply reloaded indefinitely.
- Tests: the shared httpx stub was incomplete and, installed via setdefault
before real httpx loaded, broke a combined pytest run (collection errors on
httpx.Response). Import the real installed httpx instead.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <unslothshared@gmail.com>
Co-authored-by: danielhanchen <danielhanchen@gmail.com>
* test(version-compat): keep GRPO fake-run logits finite on CPU
The GRPO fake-run test samples completions from a tiny untrained model on
CPU. Such a model can emit non-finite logits, so torch.multinomial inside
generate() intermittently raises "probability tensor contains either inf,
nan or element < 0" -- a nondeterministic sampling failure, not a regression
(the Trainer already fixes the seed, but CPU reduction order is not
bit-reproducible). Add a forward hook that sanitizes the LM head logits to a
finite bounded range before sampling, so the fake run reliably exercises the
whole train loop; the test checks the loop runs, not the numerics.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* test(version-compat): drop redundant nan_to_num bounds (clamp handles them)
* test(version-compat): scope GRPO finite-logits guard to the GRPO test
Only test_grpo_trains_on_cpu autoregressively samples completions, so it is
the only canary that can hit the non-finite-logits torch.multinomial crash.
Move the _guard_finite_logits hook out of the shared _load_plain() and into
test_grpo_trains_on_cpu so the SFT and DPO canaries keep asserting against the
model's true, unclamped logits.
---------
Co-authored-by: Daniel Han <unslothai@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* studio: show the active run's saved config in the Training Progress popover
The Training Config popover on the live Training Progress page read the
editable form store (useTrainingConfigStore), so it showed stale/static values
whenever the form changed after the run started; only the History view read the
run's saved config snapshot, which is why re-opening the same run from Recents
showed the correct values (#6853).
Wire the live view to the same authoritative source History already uses:
- Extract History's field mapping into sections/run-config-override.ts
(mapRunConfigToOverride) so both views share one mapper over
GET /api/train/runs/{id} config.
- LiveTrainingView fetches the run record as soon as the job id is known and
passes the mapped override to ProgressSection; the fetched config is keyed
by job id, and until it loads (or if the fetch fails) the form store remains
the fallback. The run record is created at job start, so it is available
while the run is live.
- ProgressSection prefers configOverride whenever one is present instead of
only when isHistorical, so the live override takes effect.
Adds a source-level regression test pinning the wiring and the mapper's
backend config keys.
Fixes#6853
* studio: retry the run-config fetch after the first step, carry the saved method
Two review fixes on the live Training Config popover source:
1. The backend creates the run row only on the first progress event, so the
fetch issued as soon as the job id appeared commonly 404'd during
model/dataset preparation and never retried -- leaving the popover on the
form store for the whole run. The effect is now also keyed on
firstStepReceived (and skips once resolved for the job), so it re-fetches
exactly when the row is guaranteed to exist.
2. The popover's method label and LoRA-row visibility came from
viewData.trainingMethod, still read from the editable form store; changing
the form (e.g. LoRA -> Full) after starting a run relabeled it and hid its
saved LoRA rows. The run-config mapper now derives trainingMethod from the
snapshot's training_type/load_in_4bit (via parseBackendTrainingMethod, now
exported from the feature index) and the live view prefers it.
* studio: fetch the run config on a terminal phase too, not just the first step
The live config-popover fetch was keyed on firstStepReceived, which the runtime
store sets only when step > 0. A run that fails or completes during preparation
(before step 1) creates and finalizes its row from the terminal error/complete
event, but neither the job id nor firstStepReceived changed, so the fetch never
ran and the popover stayed on the editable form store -- showing the wrong
config/method if the form was edited afterward (Configure re-enables on failure).
Gate the fetch on a runRowReady signal = firstStepReceived OR a terminal phase
(completed/error/stopped), the states in which the backend guarantees the row
exists. This also stops the earlier fetch-then-404 churn during preparation and
lets the effect depend only on values it reads (no lint suppression needed).
* studio: retry the run-config lookup and accept a hydrated step as row-ready
Two ways the popover could stay stuck on the editable form store for a whole
run:
- The backend publishes the progress event that reveals the run before
create_run commits, so the first lookup can lose that race and 404. The catch
changed neither runRowReady nor fetchedRunConfig, leaving every effect
dependency identical, so no further attempt was ever made for that job. The
failure path now schedules an explicit retry, bounded and keyed by job id, so
a genuinely absent row falls back to the form store instead of polling.
- A run recovered through status/metrics polling (SSE unavailable or blocked)
has currentStep restored by applyStatus/applyMetrics but never
firstStepReceived, and the phase stays training, so the row was treated as
not ready even at step > 0. currentStep > 0 is now a readiness signal of its
own.
* studio: fetch the saved run config as soon as the job id exists
start_training() inserts the run row before the pump can consume any event --
deliberately, so the run appears in history during model loading -- and /status
exposes the job id throughout the pre-step phases. Gating the lookup on a first
step or a terminal phase therefore held the popover on the editable form store
for the whole configuring/loading/downloading window, which on a long model or
dataset load is minutes, and indefinitely for a run adopted from another client.
The job id is now the entire readiness condition; the existing bounded retry
still covers the instant before the insert commits.
* Fix Training Config popover fallback for history runs without a saved config; tighten popover comments
---------
Co-authored-by: danielhanchen <unslothai@gmail.com>
* fix(registry): don't register deepseek models at import time
`_deepseek.py` called `register_deepseek_models(include_original_model=True)`
at module scope, so merely importing `unsloth.registry` registered models
(and reached the hub via `list_models`) as a side effect. None of the other
five families (`_gemma`/`_llama`/`_mistral`/`_phi`/`_qwen`) do this; they only
register when `register_models()` asks them to.
Two consequences:
- Importing the registry populated MODEL_REGISTRY on its own (32 entries,
including 10 `deepseek-ai` original models that no other family leaks) and
did network I/O at import time.
- Because the import-time call set the `_IS_DEEPSEEK_*_REGISTERED` guards with
`include_original_model=True`, the later `register_models()` call (which uses
the default `include_original_model=False`) early-returned, so the
original-model set won permanently.
Remove the stray module-level call. The `if __name__ == "__main__"` block below
still registers with `include_original_model=True` for standalone use, so the
generator script is unaffected.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* test(registry): make import-side-effect test pass on CPU-only runners
The new test spawned a fresh `python -c "import unsloth.registry"` that did
not inherit tests/conftest.py's GPU-free harness, so on no-accelerator CI
runners the child raised NotImplementedError from unsloth_zoo.device_type
before printing REGISTRY_SIZE. With check=True this surfaced only as an
opaque CalledProcessError, turning the "Repo tests (CPU)" job red even
though the registry fix is correct.
Import this directory's conftest inside the child first so it applies the
same device_type stubs and torch.cuda probe patches. Also use check=False
and include the child stdout/stderr in the assertion message so a future
import regression is legible instead of an opaque non-zero exit.
* test(registry): assert register_models() leaks no upstream originals
Adds a fresh-interpreter test that register_models() registers only
unsloth-org models (deepseek still present via the normal path) and never
leaks the upstream deepseek-ai originals that the import-time guard poisoning
used to leak (129 -> 139). Factors the conftest-harness subprocess runner
into a shared helper reused by both registry import tests.
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <unslothai@gmail.com>
* Studio: hide infra models from the hub cached inventory
The hub inventory scans behind /api/hub/cached-gguf and /api/hub/cached-models
returned the llama.cpp install validation probe (ggml-org/models) and the RAG
embedder (unsloth/bge-small-en-v1.5[-GGUF]) as on-device models. Share the
hidden-model check from routes/models.py via utils/models/hidden_models.py and
apply it in both scans. A GGUF infra repo stays visible when the user
explicitly downloaded a variant through the Hub, since variant manifests only
exist for user-initiated downloads.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: make On Device trust the hub inventory, match repo ids exactly, lighten the hidden-model import
Follow-up on the hub cached-inventory hidden-model change, addressing the review.
On Device now trusts the Hub inventory API for cached rows. The backend already
hides the RAG embedder and the llama.cpp probe and re-includes a GGUF infra repo
once the user downloads a variant through the Hub, but the frontend was
re-hiding it by repo id, so the user-downloaded variant never appeared in the On
Device list or the count. isVisibleInventoryRow now short-circuits cached rows
(kind === "cache") to visible and keeps client-side needle hiding only for local
filesystem rows and Discover.
is_hidden_model matches Hub repo ids exactly (case-insensitive) against the probe
plus the effective embedder and its GGUF companion, instead of substring
matching the configured-embedder basename. A custom embedder with a generic
basename like org/model no longer hides unrelated cached repos such as
user/model-chat or org/model-instruct. The probe filename and local-path
embedders keep exact matching.
The helper moves to utils/hidden_models.py and is imported at module scope in the
hub cache scanner, so it no longer pulls in utils/models/__init__ (the eager
model-config/checkpoint stack) and a broken import fails at startup instead of
being swallowed per-repo and silently emptying the inventory. routes.models
keeps the _is_hidden_model and _safe_resolve aliases and drops the unused
_HF_REPO_ID_RE re-export that was failing source lint.
Tests: exact repo-id matching with a custom embedder, the cached-models scan
keeping an unrelated repo, and a clean-interpreter check that the helper imports
without the model-config stack.
* Studio: match the llama.cpp probe filename on both path separators
The hidden-model check compared the probe's on-disk filename with
Path(value).name, which on a POSIX interpreter does not split a Windows-style
path ("...\stories260K.gguf") and would let the probe through. Split on both
separators so the probe is matched regardless of which OS produced the path,
matching the tolerance of the previous substring check. Adds a Windows-path
assertion to the probe test.
* Studio: harden hidden infra model handling
* Fix hidden cache row confirmation
* Fix hidden local rows and confirmed hint merges
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Handle snapshot-configured hidden models
* Hide basename-only default embedders
* Fix dynamic embedder inventory filtering
* Studio: hide the configured RAG embedder from Discover and feed rows
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <unslothshared@gmail.com>
Co-authored-by: Daniel Han <23090290+danielhanchen@users.noreply.github.com>
* Replace standalone Studio wording with Unsloth
Replace the single word Studio with Unsloth wherever it is used as
shorthand for Unsloth Studio in docs, CLI output, UI strings, i18n
locales, workflow display names, comments and docstrings.
Kept unchanged: the full name Unsloth Studio, third party product
names (LM Studio, Visual Studio, Mac Studio), feature names
(Recipe Studio, Fine-tuning Studio and its translations), and all
identifiers such as env vars, commands, paths and filenames.
* Address review feedback on the Studio wording rename
Use "an" before Unsloth where the rename left the article as "a".
Restore the split brand where Unsloth and Studio render as two halves
of the full product name: the onboarding sidebar subtitle and the
IPv6 localhost warning. Scope two messages to the full name Unsloth
Studio where plain Unsloth was misleading: the AMD README bullet and
the CLI studio setup error.
* Studio: enforce 60s minimum on idle auto-unload TTL (0 stays off)
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Drop decorative section separator from idle TTL floor tests
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: danielhanchen <unslothshared@gmail.com>
The pip scan-packages studio shard is red on main and on every open PR:
the baselined fastapi finding (the benign SSE keepalive `while True:`
loop in fastapi/routing.py, reviewed and suppressed long ago) records
its evidence at L586 with the span digest of the fastapi release current
at baseline time. The latest fastapi shifts that loop to L587 and its
span digest with it, so the evidence hash no longer matches and the
scanner reports the finding as new, failing the shard with one
unsuppressed CRITICAL.
Re-reviewed the flagged code in the current release before refreshing:
L587 is the same keepalive loop inside the streaming response machinery,
not a beacon. Only the one entry's evidence and evidence_hash change.
Verified with the scanner itself: `scan_packages.py fastapi
--no-baseline` reproduces the exact CI evidence string, and with the
updated baseline the same scan exits 0 with the finding suppressed as
1 CRITICAL baselined.
* fix(studio): add MLX adapter state control
* fix(studio): honor MLX adapter comparison state
* fix(studio): keep enabled MLX adapters permissive
* Studio: preserve public error message on MLX compare-mode adapter failures
generate_with_adapter_control raised a plain RuntimeError, which the compare
route handled with the generic handler that drops the operational message.
Raise GenStreamErrorRaised(public=chunk.public) instead and catch it in the
streaming and non-streaming consumers, matching the safetensors tool loop, so
errors like 'model is being unloaded' surface their real message.
* Studio: re-emit VLM think prefill inside the adapter context
The compare-mode merge dropped _generate_vlm's upfront yield of the prefilled
<think> block. Restore it as the first snapshot inside the lock+adapter context
(matching _generate_text) so the UI renders the thinking block during prefill
and a cancel/error before the first token does not drop it. Adds a regression
test asserting the prefill is emitted first, after entering the adapter context.
---------
Co-authored-by: danielhanchen <unslothshared@gmail.com>
With 5 or more pills active the composer collapses every pill to an
icon, which hid the Bypass permissions label behind a small glyph.
Exempt the permission pill via data-keep-label so it always shows its
label, with the collapsed icons lining up to its right. Since the pill
is never icon-only now, drop the compact-mode fallthrough in the glyph
off switch so it works while the other pills are collapsed.
The Connections form hid the API key field for the Ollama preset, which
blocked Ollama cloud (it requires a key). Show the optional field for
Ollama; the backend already sends Authorization: Bearer when a key is
set and omits the header when empty, so local keyless servers are
unaffected.
Fixes#7163
Ninth review round; each item reproduced before fixing.
The inner WSL install ran install.sh without /usr/lib/wsl/lib on PATH,
so its GPU detection (which checked PATH and /usr/bin only) could pick
CPU torch wheels on the exact Spark/N1X path this PR exists for, failing
the later torch.cuda probe. The forwarded env now appends
/usr/lib/wsl/lib to PATH (appended, so a PATH nvidia-smi still wins),
and install.sh's _has_usable_nvidia_gpu and torch-index _smi resolution
gained the same location fallback for direct WSL runs.
Three WSL failure paths in install.ps1 (WSL-not-installed deferral, the
download sentinel, and the final torch.cuda failure) set LASTEXITCODE
and returned, bypassing the round-eight Exit-InstallFailure fix, so
powershell -Command automation using the published pipe form still saw
success on those failures. All three now route through
Exit-InstallFailure, which restores the rolled-aside venv and fails the
process in every invocation mode.
The uninstall.ps1 WSL cleanup removed /root/.unsloth before killing and
matched only full argv, so cmake/nvcc children of a live CUDA build
(relative argv after the provisioner cds into the tree) survived the rm
and recreated files. The cleanup now signals each matched PID's whole
process group (guarded against the shell's own pgid, direct children
via pkill -P as fallback) before any rm; the /proc cmdline greps are
unaffected by kill order since they read process state, not files.
The round-eight same-group fallback called pkill -P without a guard;
under this script's set -e a matched provisioner with no children at
that instant (TERM pass already reaped them) aborted the whole
uninstaller before any cleanup. Now || true, like the kill beside it.
Reproduced in a dash sandbox with set -e: a childless matched PID
previously killed the harness, now dies cleanly while setsid-group and
same-group scenarios keep passing.
Verified: bash -n on both shell scripts, sh -n on the extracted WSL
clean snippet, PowerShell AST parse on both ps1 files, the three-
scenario kill sandbox, gpu-detection and installer-index pytest suites
pass, and the sh battery matches the branch baseline.
Cross-platform staging CI surfaced five Repo tests (CPU) failures where
main's tests assert on script internals this branch legitimately
changed; each reconciled on its merits.
The WoA native-wheel probe still used uv's deprecated --index-url alias
that main's test suite now forbids in favor of --default-index (same
semantics, and --default-index is what overrides inherited uv index
defaults); the probe now matches the convention.
The CUDA provision gate spelled its linked-dir guard with the
:-false default form that main's prune-refactor test blacklists
file-wide. The variable is unconditionally initialized far above, so the
guard now uses the plain spelling with identical semantics. The variable
guard itself stays: unlike a symlink test, it also covers the
canonical-location reuse case where the linked dir is not a symlink.
The gpu-detection tests extract named shell functions into a sandbox,
so _setup_has_usable_nvidia_gpu's new _resolve_nvsmi dependency made the
sandboxed helper die on command-not-found and report not_usable for
usable cases; the extraction list now includes the resolver, and the
driver-version hardening assertion tracks the resolved-path spelling
while still requiring the timeout wrapper. Also hardened the resolver
assignment with an explicit empty fallback so a future non-condition
call site cannot trip set -e.
The staging run also showed the Mac Studio Update uninstall step dying
mid-run, consistent with the round-seven group kill signalling its own
process group; the round-eight self-pgid guard already fixes that and
this push carries it to CI.
Verified: the five failing tests pass locally at this head (the one
remaining local red, test_negative_control_no_tokenizers, fails
identically with these changes stashed and did not fail in CI), bash -n,
PowerShell AST parse, and the sh battery matches the branch baseline.
Eighth review round; each item reproduced before fixing.
Exit-InstallFailure under irm-pipe-iex set LASTEXITCODE and returned, so
powershell -Command automation using the published pipe form exited 0 on
fatal installer errors (verified: a -Command run whose last call only
assigns LASTEXITCODE exits 0, while one that throws exits 1). The iex
branch now sets the var for callers that check it and then raises a
terminating error, matching the pre-existing throw behavior there:
interactive shells survive and print it, automation gets exit 1, and the
-File branch keeps carrying the specific code via exit.
The uninstall group kill could signal the uninstaller's own process
group: in a non-interactive session without job control a lingering
provisioner can share the script's pgid, and kill(-pgid) would TERM the
cleanup mid-run. The helper now compares each match's pgid against its
own and falls back to the PID plus its direct children in that case.
Both scenarios exercised in a sandbox: a setsid provisioner group still
dies whole, and a same-group provisioner dies without taking the
harness.
detect_host in install_llama_prebuilt.py resolved nvidia-smi only via
shutil.which, so the root WSL sessions this PR creates (PATH without
/usr/lib/wsl/lib) classified ARM NVIDIA WSL hosts as non-NVIDIA and took
the CPU prebuilt path before setup's provisioning logic could run. It
now falls back to /usr/lib/wsl/lib/nvidia-smi then /usr/bin/nvidia-smi,
the same order as setup.sh's resolver.
Verified: bash -n, Python AST parse, PowerShell AST parse, the pwsh
exit-code experiments above, the two-scenario kill sandbox, and the sh
test battery matches the branch baseline.
Seventh review round; each item reproduced against the live tree first.
The aarch64 bitsandbytes step gated on a bare nvidia-smi, which root login
shells cannot see under WSL2 GPU-PV (the binary lives only in
/usr/lib/wsl/lib, dropped from PATH by the /etc/profile reset), so Spark
and N1X WSL installs finished with CUDA torch but no 4-bit QLoRA. The
gate now resolves nvidia-smi explicitly with the same PATH,
/usr/lib/wsl/lib, /usr/bin order as setup.sh's resolver.
uninstall.sh's CUDA-build kill matched patterns against argv, but the
provisioner cds into the tree before `cmake --build build`, so cmake and
make children carry relative argv no pattern can match; killing only the
wrapper orphaned them mid-build. Each match's whole process group is now
signalled (TERM then KILL), with a plain PID kill as fallback when the
pgid is unreadable or shared with init. Verified in a sandbox: a child
with unmatchable argv in the wrapper's group dies with it.
The WSL shim dir was appended to user PATH while the native installer
prepends its own %USERPROFILE%\.unsloth\studio\bin, whose unsloth.exe
outlives the venv the fallback rolls aside, so on a native-to-WSL rerun
a new terminal resolved unsloth to the dead native launcher. The shim is
now prepended via Add-ToUserPath (which de-dupes and hoists), and the
dead default-root native shim is removed when the venv binary it targets
is gone; custom-root shims are left alone since the prepend outranks
them.
UNSLOTH_NPM_REGISTRY was not forwarded into the inner WSL shell even
though setup.sh threads it into every npm/bun install, so mirror-required
networks failed the frontend step (and with it the install) while the
outer installer honored the mirror. It is now forwarded with the same
strict http(s) allow-list and single-quoting as UNSLOTH_PYTORCH_MIRROR.
Verified: bash -n on both shell scripts, PowerShell AST parse on
install.ps1, the group-kill sandbox above, resolver smoke tests for the
bitsandbytes gate, and the sh test battery matches the branch baseline.
Sixth review round; every item reproduced against the live scripts first.
The WSL fallback tolerated a nonzero inner exit (the optional llama.cpp
step legitimately fails), so a run whose installer died mid-way could pass
the torch and CLI probes on a stale venv from a previous install and be
reported as success. setup.sh now stamps /root/.unsloth/.install-ok after
the core venv and Studio deps complete, just before its tolerated
llama-only nonzero exit; install.ps1 clears the stamp before the run and
requires it to exist afterwards (existence only, no mtime compare, so
WSL/Windows clock skew cannot bite). uninstall.sh removes the stamp and
the downloaded installer file so the trailing rmdir can still prune.
Root login shells reset PATH via /etc/profile and drop /usr/lib/wsl/lib,
the only location of nvidia-smi under WSL2 GPU-PV, so every bare
nvidia-smi probe in setup.sh and the provisioner could silently misreport
"no GPU". Both now resolve nvidia-smi explicitly (PATH, then
/usr/lib/wsl/lib, then /usr/bin) via a shared-resolver pattern, and the
provisioner's driver-major and compute_cap reads use the resolved path.
My round-five uninstall fix inserted the CUDA-build kill block after the
llama.cpp tree was already removed, so a live cmake/nvcc kept running
against deleted paths; the block now runs before the removal.
uninstall.ps1 gated its legacy marker-less WSL cleanup on the process
PROCESSOR_ARCHITECTURE, which reports AMD64 under an x64-emulated
PowerShell on ARM64, skipping exactly the machines the fallback installs
on. It now uses the same triple detection as install.ps1 (OSArchitecture,
Win32_Processor.Architecture 12, machine-level registry arch), factored
into one helper used at both gate sites.
The nvidia-smi capture helper retried twice with a 60s timeout
everywhere, so off WSL a hung nvidia-smi stalled three successive
detect_host probes for about two minutes each; the generous retry now
applies only under WSL (where GPU-PV load slowness is real) and bare
metal keeps a single short attempt.
The generated WSL Desktop launcher hardcoded port 8888 for launch, health
poll, and browser open, so with Jupyter or a second Studio on 8888 the
poll waited on the wrong server forever; it now scans 8888..8908 with a
TcpListener, mirroring the native launcher's free-port window, and passes
the winner via -p everywhere.
Verified: bash -n on all shell scripts, Python AST parse, PowerShell AST
parse on install.ps1, uninstall.ps1, and the generated launcher; the
launcher port scan exercised free, busy, and exhausted cases; the capture
helper unit-tested for WSL and bare-metal attempt/timeout splits; sh test
battery matches the branch baseline.
* fix(dataprep): skip .jsonl lines that are valid JSON but not objects
`_read_file_by_format` json.loads each line and hands the result to
`_extract_text_from_json`, which assumes a dict:
for field in self._TEXT_FIELDS:
if field in data and isinstance(data[field], str):
A JSON line does not have to be an object -- `"context"`, `["text"]` and
`42` are all valid JSON. For those, `field in data` stops being a key
lookup and becomes a substring/membership test, so `data[field]` raises:
"context" -> "text" in "context" is True (substring!)
-> TypeError: string indices must be integers
["text", "foo"] -> TypeError: list indices must be integers
42 -> TypeError: argument of type 'int' is not iterable
The TypeError escapes past `except json.JSONDecodeError: continue`, so the
whole load dies on one odd line.
That except clause is also the tell: a *malformed* line is already skipped
gracefully. A *well-formed* line that happens not to be an object should be
too -- it carries no text either way. This makes the two agree.
Reachable from `unsloth-cli.py:253` (`--dataset foo.jsonl` auto-detect) and
`RawTextDataLoader` is exported from `unsloth/__init__.py`.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* Slim the non-object jsonl regression test and shorten the guard comment
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* Stabilize Studio regression tests
Rebuild on current main. Restore the set-membership sidebar account-block matcher
(#6647, which fixed the same order-sensitive regex, was reverted on main, so the
guard is failing on main again) and keep the watchdog replacement-race fix, whose
blocked-watchdog stub now waits without a timeout so a superseded watchdog stays
alive until cleanup regardless of scheduler load.
* Tighten the blocked-watchdog stub comment
---------
Co-authored-by: Daniel Han <unslothshared@gmail.com>
* fix(tokenizer): check for tokenizer.model after saving it, not before
`fix_sentencepiece_tokenizer` creates its temporary directory, then returns
early unless that directory already contains a tokenizer.model:
if not os.path.exists(temporary_location):
os.makedirs(temporary_location) # fresh, empty
if not os.path.isfile(f"{temporary_location}/tokenizer.model"):
return new_tokenizer # always true
old_tokenizer.save_pretrained(temporary_location) # writes that file
The file only appears on the line after the check, so the guard is always
true and the body never runs. Nothing else writes that path either --
`convert_to_fast_tokenizer` saves into a per-name subdirectory, not
`{temporary_location}/tokenizer.model`.
Both call sites are in `get_chat_template` and are commented "Must fix the
sentence piece tokenizer since there's no tokenizer.model file!" -- the
guard defeats the exact intent the caller states. The effect is silent: the
caller still gets a working `new_tokenizer`, but the sentencepiece piece
rename is skipped, so the mapped token (e.g. the eos token remapped to
`<|im_end|>`) is missing from tokenizer.model and GGUF/llama.cpp exports
carry the old piece.
`check_if_sentencepiece_model` in save.py does the same probe in the right
order -- makedirs, save_pretrained, then isfile. Match it.
Tests are added under tests/saving/ next to the existing sentencepiece
coverage, and to the two Bucket-A lists in consolidated-tests-ci.yml, since
Repo tests (CPU) --ignores tests/saving and these need protobuf.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* Clear stale tokenizer.model before the sentencepiece guard
The guard now runs after old_tokenizer.save_pretrained, but the default
temporary_location is a fixed reusable directory. A fast-only tokenizer writes
no tokenizer.model, so a stale file from an earlier sentencepiece call could
pass the guard and patch the wrong model (e.g. mixing models in one process,
like a long-running server). Remove any existing tokenizer.model first, and add
a regression test.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Empty the reusable sentencepiece scratch directory each call
The final AutoTokenizer.from_pretrained reloads the whole temporary_location, so
removing only a stale tokenizer.model still let other artifacts from a previous
tokenizer (added_tokens.json, chat template, etc.) leak into the reload when the
default reusable directory is used across models in one process. Recreate the
directory instead, and add a regression test for the leaked-artifact case.
* Clear only top-level scratch files, keep subdirectories
Recreating the whole reusable directory deleted the {name} subtree that
convert_to_fast_tokenizer stores a converted tokenizer's source vocab in, so
old_tokenizer.save_pretrained could not copy tokenizer.model and the guard
returned the tokenizer unpatched for those legacy converted tokenizers. Remove
only stale top-level files (all the final reload reads) and leave subdirectories
intact. Add a regression test for the converted-source subdirectory.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Keep the current tokenizer's own source vocab when clearing
On a repeated get_chat_template(map_eos_token=True) call, the returned tokenizer's
vocab_file points back at the top-level tokenizer.model, and the cleanup deleted
that source before old_tokenizer.save_pretrained could re-emit it, so the guard
returned the tokenizer unpatched. Skip removing the old tokenizer's own source
vocab while still clearing stale files from a different tokenizer, and add a
regression test.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Use a per-call temporary directory for the sentencepiece fix
The scratch directory defaulted to a single shared path, so concurrent or repeated
get_chat_template(map_eos_token=True) calls could delete or overwrite each other's
tokenizer.model between save and reload (tripping the piece assertion or reloading
the wrong model), and stale files from an earlier tokenizer could leak into the
reload. Work in a unique per-call subdirectory instead: this isolates every call
without deleting anything the caller owns, and replaces the earlier per-file cleanup.
Tests updated to read the patched model from the reloaded directory and to cover
isolation and source-vocab preservation.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Pass only the applied token mappings into the sentencepiece fix
get_chat_template mirrors token remaps into tokenizer.model via
fix_sentencepiece_tokenizer, but two caller paths passed a mapping that did not
match what they wrote to the fast tokenizer JSON, so once the sentencepiece patch
runs the model and JSON disagree:
- the mapped-token path skipped entries whose target already existed but still
passed the full mapping, renaming a piece the JSON never changed;
- the EOS-swap path swapped both tokens in the JSON but passed only one direction,
leaving two stop_word pieces and no old EOS piece.
Pass the applied mapping (and both swap directions) instead. Add regression tests.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten sentencepiece guard comments
* Add SPDX license identifier to sentencepiece guard test
* Reclaim the per-call sentencepiece scratch directory
The per-call tempfile.mkdtemp fixed the shared-directory race but never cleaned
up, so a long-running process leaked one scratch dir per call. The dir cannot be
deleted eagerly for sentencepiece tokenizers because the returned tokenizer's
vocab_file points into it (a later save_pretrained copies the patched
tokenizer.model from there). Reclaim it correctly instead: remove the dir right
away on the fast-only path (the returned tokenizer never references it), and
attach a weakref.finalize so the sentencepiece dir is removed once its tokenizer
is garbage collected. Add regression tests for both.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten the scratch-dir reclaim comment
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-authored-by: danielhanchen <unslothshared@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <danielhanchen@gmail.com>
Fifth review round; each item traced through the live scripts before fixing.
A provisioner fresh clone that failed to produce a server was left behind as
a markerless git tree; under a custom UNSLOTH_STUDIO_HOME the next run's
ownership assert refuses the unmarked dir and aborts the whole install until
the user deletes it by hand. _restore_prev now removes a clone this script
created when no server came out of it (backed-up dirs restore as before).
The CUDA provision gate ignored --with-llama-cpp-dir linked mode, so a
linked user tree with a CPU-only server could be checked out to a pinned
ref, rebuilt in place, or moved aside entirely and replaced by a fresh
clone. The gate now skips linked local dirs.
uninstall.sh removed the CUDA build artifacts without stopping a running
detached build; _pkill_studio only matches Studio roots, so live cmake/nvcc
kept burning thermals, recreated build files, and defeated the trailing
rmdir. The runner, provisioner, and llama.cpp-path processes now get
TERM-then-KILL with the same escape helper and grace the Studio kill uses.
The worker's memory-fraction guard classified Spark purely from device
props, so UNSLOTH_FORCE_DGX_SPARK=1 on an unlisted name got no fraction
guard (and the fraction env was dead), while FORCE=0 could not disable it;
the guard now honors the same force semantics as the detectors.
UNSLOTH_LLAMA_TAG / UNSLOTH_LLAMA_PR were interpolated into the runner
script's single-quoted exports unvalidated while every sibling forward has
an allow-list; they now get the INSTALL_REF ref allow-list and a digits-only
check respectively (own-machine robustness, not a trust boundary).
On WSL-fallback success with a custom UNSLOTH_STUDIO_HOME, the installer
deleted the rolled-aside custom-root venv right after telling the user that
root is not used by the WSL install; a custom root now restores the previous
venv instead (the WSL shim does not depend on the Windows venv), while the
default root keeps dropping the vestigial backup.
Verified: bash -n on all three shell scripts, AST parse on worker.py,
PowerShell AST parse on install.ps1, icon suites pass, sh battery matches
the branch baseline. Two resurfaced anchors (build/bin backup, --package
forwarding) confirmed already fixed at head.
Fourth review round; each item verified against the live scripts and the CUDA
compatibility documentation before fixing.
The provisioner installed (and, since the stale-toolkit change, preferred)
CUDA 13.3 without ever consulting the driver, but cu13 binaries need a 580+
driver and minor-version compatibility never crosses majors, so a GH200-class
host on a 5xx driver got an unloadable llama-server that the structural
acceptance check then stamped as ready. The driver's supported CUDA major is
now read from nvidia-smi and enforced three ways: the stale-toolkit upgrade
only fires when the driver can run cu13, a fresh install on a sub-13 driver
bails to the existing no-toolkit message instead of installing 13.3, and a
final guard swaps a too-new selected toolkit for the newest one the driver
supports (or refuses to build). Spark-class hosts (580+ drivers) behave
exactly as before; unparseable output keeps the previous behavior.
The WSL install pipeline ended in curl | sh, so a failed download fed sh an
empty stdin and exited 0; on a rerun the stale venv then passed the torch
probe and the installer reported success without ever running. install.sh is
now downloaded to a file with exit 86 as the never-ran sentinel, checked
before any probe (rollback + non-zero). The --package splice moved onto the
file invocation.
When the Studio web-server dep repair failed its re-verify, the installer
still created shims and reported success; the missing set includes typer, so
even the plain unsloth CLI dies. A failed repair now routes to the existing
failure path (rollback + non-zero), mirroring the CLI-missing case.
If all three provision-script resolutions fail (unpackaged wheel + GitHub
unreachable), the provision block silently skipped and, with the CPU build
now deferred on native Spark hosts, the install could report success with no
GGUF server; that case is now marked degraded so the CPU-prebuilt last
resort and failure exit fire.
flex_attention.py's Spark sniff gets the same /usr/lib/wsl/lib/nvidia-smi
fallback as the other two detectors (grep confirms these are the only three),
and uninstall.sh removes the remaining WSL-side build artifacts
(run_llama_build.sh, llama_cuda_build.log, the shortcut-skip marker) so the
.unsloth directory can actually be removed.
Verified: bash -n on all three shell scripts, AST parse on flex_attention.py,
PowerShell AST parse on install.ps1, the toolkit-picker awk exercised against
a fake /usr/local tree (driver 12 picks cuda-12.8 over 13.0, driver 11 picks
none), icon suites pass, sh test battery matches the branch baseline.
The Spark detectors (library _is_dgx_spark_no_cuda_init and the worker's
pre-CUDA sniff) called bare nvidia-smi, but the WoA shim execs the venv
binary directly with no login shell, where /usr/lib/wsl/lib can be off PATH;
both now resolve WSL's nvidia-smi path explicitly when the bare name is not
found, so the allocator setup works on plain 'unsloth ...' launches.
An explicit UNSLOTH_PYTHON pin was lost across the WSL boundary (Windows env
vars do not cross into the distro), so the inner install.sh built the venv on
its default Python while the installer reported success; the pin is now
forwarded, gated on a strict X.Y[.Z] shape before splicing into bash -lc.
The WSL2 probe/conversion only ran for pre-existing distros; a fresh install
relied on wsl --set-default-version 2 succeeding silently and could proceed
on WSL1 all the way to the final torch.cuda failure. The probe and in-place
conversion now run for freshly installed distros too.
The fourth resurfaced item (complete Studio dependency repair set) is already
fixed at head: the repair list includes sqlite-vec, pymupdf, and python-docx.
Third review round; each item re-verified against the live scripts.
A stale CUDA < 13 toolkit was kept forever: the 13.3 install was gated on
nvcc being absent, so a host with CUDA 12.x failed the sm_121 configure (or
the glibc >= 2.41 rsqrt clash) on every rerun and always exited with the CPU
server. When apt can provide 13.3 the provisioner now installs it alongside a
stale toolkit; find_nvcc's sort -V prefers the new install, and a failed
install leaves the old toolkit as the last resort, so non-Spark hosts that
build fine on cu12x are unaffected.
llama.cpp pins only applied to fresh clones; an existing checkout rebuilt
whatever commit it had while the log claimed a release pin. Existing
checkouts now fetch and check out the pinned (or resolved-latest) ref, best
effort with the current commit as fallback, and the UNSLOTH_LLAMA_PR handling
moved out of the fresh-clone branch so it applies to both paths.
The WSL fallback silently dropped a non-default --package and reported
success with stock unsloth; it is now spliced into the curl | sh invocation
(the name is regex-validated at parse time).
setup.sh's CUDA provision gate used raw nvidia-smi and ignored the
_setup_nvidia_usable computation that honors CUDA_VISIBLE_DEVICES=""/-1, so a
mixed-GPU host that hid its NVIDIA card still got a system CUDA install; the
gate now requires the flag. On native Linux Spark hosts without nvcc, setup.sh
also no longer does the multi-minute CPU source build that the CUDA provision
in the same run immediately replaces (mirroring the existing WSL deferral
arm); provision failure still cascades to the CPU-prebuilt last resort.
uninstall.ps1's distro extraction truncated quoted names at the first space
(-d "Ubuntu Preview" matched as "Ubuntu"), wrongly keeping or removing
shortcuts; the regex now matches a full quoted token first. And the profile
icon (%USERPROFILE%\.unsloth\unsloth.ico) was removed unconditionally while
the sweep above deliberately keeps launchers for non-evidenced WSL installs,
blanking their icons; removal is now gated on no surviving Unsloth shortcut,
mirroring uninstall.sh's _drop_shared_icon_if_unused guard.
Verified: bash -n on both shell scripts, PowerShell AST parse on both ps1
files, the new distro regex proven on spaced and unspaced names, icon suites
pass, sh test battery matches the branch baseline.
Second review round on the Windows-on-ARM + NVIDIA path; each item verified
against the live code (and torch where relevant) before fixing.
The native-CUDA probe ran uv --dry-run against the venv interpreter without
checking its architecture. uv resolves for the interpreter's platform tags, so
an x64-emulated python resolved existing win_amd64 CUDA wheels and "proved" a
native wheel WoA cannot use, skipping the WSL fallback entirely. The probe now
requires platform.machine() ARM64 from the venv python first; anything else
keeps the WSL routing.
The Studio worker appended PYTORCH_CUDA_ALLOC_CONF next to its memory-fraction
logic, 550 lines after detect_hardware() had already initialized CUDA, where
the allocator config is latched (verified on torch 2.9.1: expandable_segments
set after get_device_properties is a no-op in memory snapshots). The CUDA-free
Spark sniff now runs immediately before detect_hardware(), and it honors the
documented UNSLOTH_FORCE_DGX_SPARK=1/0 override the library detectors support,
closing the older force-flag item on the same block.
setup.sh's _have_cuda_llama_server accepted any co-located libggml-cuda.so,
re-opening the interrupted-relink hole the provisioner's completion stamp was
added to close: in exactly that state setup.sh skipped provisioning and
reported CUDA ready over the old CPU binary. The split-.so branch now also
requires the stamp; monolithic ldd-linked builds are unaffected.
The provisioner builds llama-quantize but never created the repo-root shim
that unsloth_zoo's check_llama_cpp needs (it only searches the root, which is
why setup.sh symlinks it in all three of its own paths). The success branch
now mirrors that symlink.
CMAKE_CUDA_ARCHITECTURES=native needs CMake >= 3.24, but this script installs
distro cmake (Ubuntu 22.04 apt ships 3.22), so the N/A-compute_cap fallback
aborted configure, wiped build/, and aborted again. The fallback now omits the
flag and lets ggml's version-guarded CMake defaults pick the arches.
Fresh clones tracked ggml-org master, bypassing setup.sh's newest-release pin
policy (its own header warns master bypasses the pin). An unset or "latest"
ref now resolves to the newest release tag via the GitHub API, keeping the
default-branch clone as the best-effort fallback when the API is unreachable.
install.sh writes the WSL shortcut icon to the Windows profile
(%USERPROFILE%\.unsloth\unsloth.ico) because the WoA icon broker cannot read
AppData\Local, but both uninstall.sh cleanup sites only cleaned the
AppData\Local icon. Both now clean the profile icon and drop the directory
when empty.
Verified: bash -n on all four shell scripts, AST parse on worker.py,
PowerShell AST parse on both ps1 files, the icon suites pass, and the sh test
battery matches the branch baseline (test_install_host_defaults.sh fails
identically on the clean tree).
Review round on the Windows-on-ARM + NVIDIA WSL2 path; each item reproduced
against the live scripts before fixing.
provision_llama_cuda.sh now serializes with install_llama_prebuilt.py on the
same <parent>/.<name>.install.lock file (its filelock backend is flock(2), so
shell flock interoperates; append-mode open so the Python O_EXCL fallback's
PID file is never truncated). The detached background builder could otherwise
race an installer rerun or `unsloth studio update`, both of which mv/rm -rf
inside the llama.cpp dir. Losing the 2h wait exits 0: another provisioner is
already doing the job.
The step-0 early-skip trusted a co-located libggml-cuda.so alone, which
wrongly skips one case: an in-place rebuild interrupted after the .so links
but before llama-server relinks leaves new .so + old CPU server. A completion
stamp (build/bin/.unsloth-cuda-ok) written only after the script's own final
CUDA check closes that window; skip now requires ldd evidence or the stamp.
The rejected functional --list-devices probe stays rejected: the stamp does
not gamble thermals on an env-fragile probe.
The WSL shortcut skip (install.ps1 owns the canonical WoA .lnk) was only a
transient env var, so the first `unsloth studio update`, whose wsl.exe shim
carries no env into install.sh --shortcuts-only, recreated the duplicate
blank-icon shortcut. The skip is now also persisted as
/root/.unsloth/.skip-wsl-windows-shortcut, checked by install.sh and removed
with the install by both uninstallers.
--with-llama-cpp-dir (and UNSLOTH_LOCAL_LLAMA_CPP_DIR) were parsed but
silently ignored on the WSL fallback path, which builds its own llama.cpp
inside the distro. Reject with guidance (UNSLOTH_LLAMA_TAG / UNSLOTH_LLAMA_PR
pin the WSL-side build), mirroring the --local reject.
uninstall.sh's Windows shortcut sweep only removed wsl.exe-target .lnks, so
the WoA fallback shortcuts (powershell.exe + launch-studio-wsl.ps1) survived
while their launcher dir was deleted, leaving dangling shortcuts. The
owner-matched cleanup now removes them first.
uninstall.ps1 swept every "Unsloth Studio (*.lnk" as legacy, but install.sh
creates exactly that per-distro name for current WSL installs, and the WSL
cleanup below only removes evidenced distros. The sweep now keeps a live
wsl.exe launcher whose distro is not in the same evidence set, so a surviving
WSL install keeps its shortcut; everything else is still swept.
Verified: bash -n on all three shell scripts, PowerShell AST parse on both
ps1 files, flock mutual-exclusion and stamp skip/rebuild decisions exercised
standalone, and the uninstall icon suites (sh + ps1) pass. The
test_install_host_defaults.sh failure pre-exists on the branch merge base.
* feat(studio): expose opt-in MCP control plane
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Harden Studio MCP tools: byte-safe auth, page clamping, forward export/checkpoint fields
Follow-up hardening on the opt-in MCP control plane. All changes are additive
and backwards compatible.
- BearerTokenMiddleware now compares the Authorization header on raw bytes.
A non-ASCII bearer value previously reached str-based hmac.compare_digest,
which raises TypeError and surfaced as a 500 instead of a clean 401. The
constructor also rejects an empty or whitespace-only token so an empty token
can never match an empty "Bearer " header.
- MCP tools call the route functions directly, which skips FastAPI Query
validation. list_training_runs and get_recipe_job_dataset now clamp limit and
offset to the same bounds the HTTP routes enforce (a negative SQLite LIMIT
otherwise means "no limit").
- export_gguf forwards hf_token (the backend rejects a Hub upload without it),
accepts a list of quantization methods, and exposes imatrix / imatrix_path so
the IQ low-bit quants are reachable.
- load_checkpoint forwards hf_token and approved_remote_code_fingerprint so
gated checkpoints and the remote-code approval retry work. Its docstring is
corrected: the export backend coexists with training and inference rather than
freeing GPU work.
- start_training passes via_api_key=False explicitly instead of relying on the
unfilled Depends default.
Tests: add coverage for non-ASCII and empty-token auth, the correct-token pass
through, non-http scope pass through, pagination clamping, and the forwarded
export/checkpoint fields.
* Harden Studio MCP: cap /mcp request bodies, reject unusable tokens, fix docs
Follow-up hardening from a full review pass. All changes are additive and
backwards compatible.
- Add "/mcp" to _BODY_PROTECTED_PREFIXES so MaxBodyMiddleware enforces the same
request-body cap it already applies to every other write endpoint (/api/train,
/api/export, /api/data-recipe, ...). The MCP endpoint accepts authenticated
POST tool-call bodies; without this an authenticated client could send an
unbounded body. The middleware only buffers the request body (not the SSE
response), so streaming is unaffected, and the 500MB default cap never affects
a real JSON-RPC tool call (verified live).
- Reject a non-ASCII UNSLOTH_STUDIO_MCP_TOKEN at construction. HTTP header values
are ASCII, so a non-ASCII token cannot be sent by a standard client and would
silently lock out the endpoint; fail fast instead.
- MCP.md: document the canonical /mcp/ endpoint and note that /mcp redirects to
it, so clients that do not follow redirected POSTs still connect.
Tests: add non-ASCII token rejection coverage.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten MCP server comments
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <unslothshared@gmail.com>
* Propagate fp8 block_size before the early return in get_lora_parameters_bias
get_lora_parameters_bias set the fp8 block_size on W/W_quant only after the
disable_adapters/merged early return, so on the merged or disabled path (merged
inference, DPO reference model) a block-fp8 weight lost its real block_size and
downstream fp8 kernels fell back to [128, 128]. The non-bias sibling
get_lora_parameters already sets block_size before its early return; move the
block so both behave the same.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Guard the fp8 block_size against a missing quant state
A decompressed compressed-tensors layer keeps quant_method == "fp8" while its
weight is back to bf16, so it has no quant state and get_lora_parameters_bias
must still return W_quant None for fast_linear_forward to fall back to a plain
matmul. Only attach block_size when a quant state was actually found.
* Guard the sibling get_lora_parameters fp8 block_size against a missing quant state
Mirror the get_lora_parameters_bias guard so a decompressed compressed-tensors
layer (quant_method fp8, bf16 weight, no quant state) does not raise
AttributeError on the fused-LoRA path. Add a CPU-only regression test.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* Studio: don't apply nest_asyncio on plain CLI starts (breaks asyncio on Python 3.14+)
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Skip nest_asyncio on Python 3.14+ so notebook and embedded Studio starts also work
* Tighten the nest_asyncio gate comment
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <unslothshared@gmail.com>
* fix(studio): recover stalled Hub downloads over HTTP
* fix(studio): preserve retry generation and progress baseline
* fix(studio): keep XET retry handoff nonterminal
* fix(studio): preserve retry cancellation on claim failure
* fix(studio): make retry failure cancellation atomic
* fix(studio): close skipped retry state gaps
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Stabilize chat-only export gate detection on Windows
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Retrigger CI on a user-authored head
* fix(studio): serialize XET HTTP retry handoff
* List XET to HTTP retries that are briefly released from the repo guard as active downloads for PR #6858
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Settle no-process active downloads on shutdown so a parked XET retry cannot spawn after cleanup for PR #6858
* Settle exited-error and no-process downloads on shutdown and persist their cancel markers for PR #6858
* Keep terminal HTTP failures uncancelled and block companion deletion for released retry peers for PR #6858
* Trim download lifecycle test coverage
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: Etherll <61019402+Etherll@users.noreply.github.com>
* refactor(studio): move chat model picker into features/model-picker
Relocate model-selector + its support files from components/assistant-ui
into a self-contained features/model-picker feature (own barrel), mirroring
the modular Hub layout. Pure move + import repoint; no behaviour change.
* feat(model-picker): add per-model config persistence layer
Superset PerModelConfig (customContextLength, kvCacheDtype, speculativeType,
specDraftNMax, tensorParallel, chatTemplateOverride, trustRemoteCode) persisted
to localStorage (unsloth_model_configs) with schema versioning + LRU budget.
KV-dtype and speculative value sets match main's sidebar (no q4_0/ngram-simple).
Reuses features/hub/lib/model-identity for normalization; adds storage-key layer
and applyPerModelConfigToRuntime (sets tensorParallel, which the old PR omitted).
* feat(picker): modular backend for chat-template validate + default fetch
New studio/backend/picker package (schemas/service/routes) mounted at /api/picker:
- POST /api/picker/validate-chat-template (Jinja syntax validation, no false positives)
- GET /api/picker/chat-template/{model_name} (default template from tokenizer_config.json,
reusing get_cache_path/resolve_cached_repo_id_case; graceful null, no model-code exec)
Frontend api/templates.ts client + hooks/use-model-defaults lazy cache. No backend
changes to the existing inference load route (per-model load fields already supported).
* feat(model-picker): bind picker on-device list to shared hub inventory
Picker now sources cached + local models from useHubInventory (the Hub's shared
store) via a thin adapter, replacing its own /api/models/* fetchers + module
caches. Hub, download manager, and picker now share one source of truth, so
completed downloads reflect in the picker automatically. Partial/live-download
rows are filtered from the cached lists (unchanged rendering). Local naming/search
preserved via additive LocalInventoryRow modelId/displayName. Variant expander,
scan-folder management, recommended-fit, search, external providers untouched.
Known minor: cached 'Downloaded date' sort tiebreak degrades to alphabetical
(hub cached rows carry no mtime); default 'recent' (load-time) sort preserved.
* feat(model-picker): per-model config step inside the picker
Picking a (non-external) model now opens an in-picker config view built from
main's current load controls (context length, KV cache dtype, speculative
decoding, draft tokens, tensor parallel) plus a chat-template editor backed by
the picker validate/default endpoints. 'Remember for this model' persists the
config per model+variant; Run forwards the config to the existing load flow via
meta.config. External models bypass the step. Two-view orchestration lives in
model-selector (single interception point); pickers.tsx call sites untouched.
trustRemoteCode dropped from PerModelConfig to preserve main's per-load consent.
* feat(chat): apply/persist per-model config through the load flow
handleCheckpointChange threads meta.config into the selection; stageOrLoad and
the autoload/Hub-run paths now apply the picker config (explicit pick or saved
remembered config) via applyPerModelConfigToRuntime before staging/loading, with
keepSpeculative set so a remembered speculative mode survives the model switch.
Replaces the old remembered-load-settings seeding (resolveInitialConfig now the
single source). SelectedModelInput carries config.
* refactor(chat): remove per-model load config from the right sidebar
The load knobs (context, KV cache, speculative, draft tokens, tensor parallel)
and the chat-template editor now live only in the picker config step. The sheet's
Model section keeps the staged Load/Cancel flow (config is applied at pick time);
sampling params, system prompt, and RAG are unchanged. Deletes the superseded
remembered-load-settings module + the store's applyRememberedLoadSettings action,
removes the now-dead sheet state/imports, and points the settings reset at
unsloth_model_configs. Delete-cleanup deferred (stale config is LRU-capped).
* fix(model-picker): remove leftover sidebar-staging cogwheel + empty Model section
The downloaded-variant gear (ModelLoadSettingsAction) staged a model straight
into the right-sidebar Run-settings flow -- the old 'configure before load' path
now fully replaced by the in-picker config step. Removed the gear + its component.
Also gate the sheet's 'Model' section to staged picks only (pendingSelection):
after the load-knob strip its content is staged-only, so it was rendering an
empty section header whenever a model was merely loaded.
* chore(chat): remove dead per-model-config setters + modelControlsDisabled
After the load-config UI moved into the picker, the store's per-model setters
(setKvCacheDtype/setSpeculativeType/setSpecDraftNMax/setTensorParallel/
setCustomContextLength/setChatTemplateOverride) had zero callers
(applyPerModelConfigToRuntime writes via setState), and the sheet's
modelControlsDisabled was unreferenced. Verified dead across the whole tree.
* fix(chat): config-step Load actually loads (ignore Load-on-selection)
Root cause: with Settings > Chat > 'Load on selection' turned OFF, the config
step's load went down the deferred-staging path -- opening the right sidebar with
'<model> is staged, not loaded yet / Choose Load model'. The in-picker config step
IS the deliberate load action, so its Load now loads immediately (or downloads +
auto-loads when not cached) regardless of the toggle. Renamed the button
'Run model' -> 'Load model' to match. Native/dropped picks still honor the toggle.
* refactor(chat,hub): retire 'Load on selection' — config step is the only load flow
The in-picker config step (and the Hub Run button) now fully supersede the old
stage-to-sidebar flow, so the Load-on-selection toggle is removed everywhere:
- chat stageOrLoad: every pick loads immediately, or downloads + auto-loads when
not cached (the previous default behaviour, now universal).
- hub Run: drops the stage branch; downloaded GGUFs load directly with their saved
per-model config (no collision with the chat config step — both end at selectModel).
- store: removed loadOnSelection field/setter/key/default; Settings>Chat toggle and
its settings-reset entry removed.
- staged sidebar section is now a download-progress view (auto-loads on completion).
No manual staging remains; stageModel is used only for background auto-load downloads.
* feat(model-picker): default chat template from GGUF + thread variant through config flow
Read the embedded tokenizer.chat_template from GGUF files (read_gguf_chat_template
in gguf_metadata) and use it as the per-model default. Plumb gguf_variant through
the picker service, /api/picker/chat-template route, frontend templates API, and
use-model-defaults so the right variant's template is fetched.
Also refine the picker config-page/model-selector wiring, drop the dead
ggufNativeContextLength runtime path, and add the per-model-config storage keys to
the settings prefs export.
* feat(model-picker): read safetensors chat template + hide editor where it has no effect
Resolve the default chat template for safetensors models: prefer the modern
chat_template.jinja, fall back to the tokenizer_config.json chat_template field,
then chat_template.json (multimodal processor), then the GGUF embedded template.
Applied to local dirs, the HF cache snapshot scan, and the HF remote fetch.
Hide the chat-template editor in the picker for safetensors models — the override
is only applied at load by the GGUF/llama.cpp backend, so editing it on safetensors
currently has no effect. GGUF keeps the editor. Nothing removed; the dialog stays
for when the safetensors apply path is wired up in a later branch.
* fix(model-picker): set legacy-migration flag only after the write succeeds
Set unsloth_model_configs_migrated only once writeMap confirms the migrated
map persisted, so a quota/storage failure no longer marks migration done and
silently drops the user's pre-existing remembered settings — the next load retries.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* MVP model picker fixes
* MVP picker config fix
* MVP safetensors config
* MVP max seq config
* MVP max seq fix
* Fix static max tokens cap ignoring model context
* Fix picker GGUF scan parity
* fix(studio): harden model picker config loading
Apply remembered per-model configs consistently from picker and Hub loads, keep default configs from overriding standing speculative settings, add config access for direct local GGUF files, and support saving or forgetting active model settings without a reload.
* Fix model picker config flow
* Fix model picker config loads
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Avoid recursive per-model config migration reads
* Apply the displayed context length when loading a GGUF
* Fix template validation, cached template lookup, and failed load rollback
- Validate chat templates with the loopcontrols extension so templates
that use break or continue tags pass the picker validator, matching the
inference renderer that already accepts them.
- Read the default chat template from the newest cache snapshot rather than
an arbitrary iterdir order, so an older cached revision no longer prefills
a stale template.
- Capture the runtime per-model config before a load and reapply it when the
load fails, so a failed switch leaves the active model context, KV cache,
template, and speculative settings as they were.
* Make chat template view only for safetensors models
Custom chat template overrides are applied at inference only for GGUF
models, which pass the template to llama-server. The safetensors backend
renders with the model built-in template and ignores the override, so
editing it would save a value that never loads. For safetensors the
config page now opens the template as a read-only preview with a note
that editing is not available yet. This can become editable once
inference support for custom safetensors templates lands in main.
* Fix model picker config edge cases
- Restore prior runtime config when a load no-ops for the active model
- Cap the picker validator request body via the protected prefixes
- Keep the GGUF context slider max above the loaded context
- Fetch subfolder chat templates for uncached Hub repos
- Show the compare side config when reopening the picker
* Keep saved GGUF context above the fallback ceiling
* Show the model config in the run settings sidebar
* Fix model config sidebar reset and context slider
- Stack the remember toggle and action buttons in the sidebar
- Reset the config to defaults instead of the loaded values
- Fetch the native context so the slider max is not the loaded value
* Fix model picker config and download regressions
- Run picker chat template routes off the event loop
- Depth and root guard local template directory scans
- Restore download manager flow for uncached hub picks
- Apply per model context length on reload
- Import model picker symbols from the feature barrel
* Fix model picker config and cached download sorting
- Restore load settings when a Hub run is rejected mid load
- Reuse one NumericValueInput instead of a duplicate copy
- Fix double decode of the model name in the template route
- Remove the unused reset-to-loaded settings action
- Fix cached model download sorting
* Fix model picker per-model config edge cases
Honor a saved or typed max seq length above the model's native context so
RoPE extended values are no longer clamped and silently overwritten. Allow
typing past native while the slider keeps native as a soft ceiling.
Guard the fetch success paths in use-model-defaults against an aborted
signal, and refetch when the HF token changes.
Hash the chat template content in the sidebar remount key instead of its
length. Enable reset for a GGUF whose native context is unknown, and floor
the context slider max so it can never fall below the min.
* Fix GGUF context auto-fit and gated model config token
Stop forcing a 32768 context when a GGUF native context is unknown so the backend auto-fits to VRAM again, while still honoring an explicit context edit.
Send the HF token as a query param so gated safetensors models resolve their max position embeddings.
Derive model default state during render to drop the set-state-in-effect calls.
* Fix native GGUF context ceiling and guard picker template reads
Restore the native context store field so the sidebar slider keeps the
full ceiling for drag and drop GGUFs. Limit local chat template reads to
the browse allowlist, skip malformed repo ids, and drop unused model
picker exports.
* Fix model picker lint boundaries
* Fix model picker review findings
Chat template editor never seeded its draft. Radix only calls onOpenChange
from internal events, so the seed in the nextOpen branch was dead and a model
with a saved override opened empty. Saving then cleared the override. Drop the
dead branch, treat draft as an untouched sentinel, and reset it on every close.
Uncached Hub picks could auto load a model after the user left the chat. Main
detached the staged pick on route exit and on chat context change. Carry the
context key on the pending pick and skip the load when it no longer matches.
Also clear configTarget when the picker closes, restore the onUpdated ref so
variant rows stop resubscribing on every parent render, skip the LRU write when
the entry is already most recent, import NumericValueInput relatively, and drop
the unused ModelUpdateAction barrel export.
* Preserve GGUF context on active reload
* Fix model picker per-model config regressions
- Stop reloading the already loaded model on re-pick
- Hide infra models from the chat picker
- Detect vision support on cached GGUF repos
- Honor saved maxSeqLength on auto load
- Restore default chat template for local GGUFs
- Warn on save failure and revert config on cancel
- Refetch picker inventory on open
- Persist read only per model config safely
* Fix stale model auto load
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix model picker numeric input sizing and constraints
Size value inputs to their content so long context lengths are not clipped,
restrict them to numeric characters, and stop the speculative decoding label
from truncating in the sidebar.
* Fix picker CI tests and harden chat template resolution for PR #6647
- tests: point the descender guard at the moved model-selector.tsx path
- tests: exclude the disabled Reload model button from the regenerate locator so .first targets the real Regenerate
- picker/service.py: reject symlinked template/gguf leaves that resolve outside the browse allowlist (HF cache reads unchanged)
- compare mode: resolve each pane's own remembered chat template instead of inheriting the other pane's from the store
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Protect future-schema per-model configs from deletion for PR #6647
savePerModelConfig already refuses to overwrite a stored config whose schema version is newer than this client understands, but deletePerModelConfig did not. Unchecking Remember on an older client therefore silently destroyed a newer client's saved config. Apply the same guard on delete and surface the blocked case through the existing saveFailed toast.
* Protect future-schema per-model configs from quota eviction for PR #6647
The save and delete guards already refuse to touch a stored config whose schema version is newer than this client understands, but the quota-eviction path did not, so a full store on an older client could still evict a newer client's config. Skip future-schema entries when evicting and fail the save if the budget cannot be met without them.
* Fix GGUF context persistence, compare context, and rollback settings for PR #6647
Persist a GGUF context override from the user's intent instead of collapsing it against the loaded context, which reintroduced the context-reset (f4838782cb reverted the native-baseline fix). model-config-page now collapses the saved value against native, and use-chat-model-runtime and chat-adapter retain the requested context on load so re-saving another setting keeps the override; a null request stays null so a VRAM auto-fit never becomes a stored override.
shared-composer: a compare pane with no explicit GGUF context now loads at native (0) like single-view, not the session maxSeqLength that silently shrank the shown context.
use-chat-model-runtime: restore the previous model's KV cache dtype and chat template on a failed-load rollback so it runs as it was, not with backend defaults.
* Preserve native path token when reloading the active model for PR #6647
handleReloadActiveModel rebuilt the selection without the store's activeNativePathToken, so reloading a file-picked GGUF after a settings change validated the display label as a repo/path and failed. Thread the active native token through the reload selection so native-loaded models reopen correctly.
* Make picker template validation resilient and accept HF generation tags for PR #6647
Import Jinja lazily inside validate_chat_template so a backend without the optional jinja2 package (GGUF-only installs) still starts instead of raising ModuleNotFoundError at import time. Register a no-op extension for the Transformers {% generation %} assistant-mask tag so pasting a valid HF chat template validates, matching the renderer, rather than being rejected as an unknown tag.
* Honor remembered compare config and parse processor chat_template.json for PR #6647
* Fix failed-load rollback context and processor template map fallback for PR #6647
* Restore speculative decoding config on failed-switch rollback
When a model switch fails after the previous model was unloaded, the
rollback reload restored tensor_parallel, KV cache dtype and the chat
template override, but omitted speculative_type and spec_draft_n_max and
cleared their loaded shadows to null. The previous model therefore came
back running at backend defaults (speculation off) while the UI still
showed it enabled, and the status resync confirmed the off state. Resend
the previous model's speculative settings in the rollback load and keep
the store's active and loaded speculative fields in sync with them.
* Reset max sequence length when a model has no saved config
applyPerModelConfigToRuntime reset every per-model field except
maxSeqLength, which it only wrote when the incoming config had one.
maxSeqLength is the sole field carried on store.params, so selecting a
model with no remembered config left the previous model's value in place
and later loaded the new model at that leaked length. Fall back to the
standing default so an unremembered model loads at its own default.
* Surface a message when a variant update cannot start
startManagedUpdate handled the conflict and error start outcomes but let
busy fall through as if the update began, so the confirm dialog closed
with no job created and the cached variant stayed stale. Show an info
message when the repo is busy with a sibling transfer so the click is
not silently dropped.
* Keep per-model speculative choices out of the global default
A staged load with a per-model or one-off config sets keepSpeculative,
which already skips reading the global speculative preference. The
matching save still ran unconditionally, so the model-specific choice was
written to the global unsloth_chat_speculative_type and a later model with
no saved config started from it instead of Auto. Skip saveSpeculativeType
when keepSpeculative so the per-model choice stays isolated.
* Seed non-active model settings from the app default max length
The Run settings page captured initialMaxSeqLength from the loaded
model's runtime params and fell back to it for a model with no saved
config. Opening settings for a different, unloaded model and clicking
Load then sent the active model's context (for example 64k) instead of
the 4096 default, risking validation failures or OOMs. Seed the default
for non-active models and keep the runtime value only for the active one.
* Prefer sidecar tokenizer chat template over the GGUF copy for variants
_chat_template_from_dir returned the embedded GGUF template first when a
variant was selected, reversing the tokenizer-first precedence of the
no-variant path. A model whose chat_template.jinja or tokenizer_config.json
supersedes a stale embedded template then got the wrong template on
variant selection. Keep tokenizer files first regardless of variant; the
variant only picks which GGUF is the fallback. Adds regression tests for
both the tokenizer-wins and gguf-fallback cases.
* Keep per-model speculative choices load-local in autoload and compare
The interactive load path treats a per-model speculative choice as
load-local and skips writing it to the global default. Autoload and
generalized compare still called saveSpeculativeType unconditionally, so a
remembered off or ngram setting leaked into unsloth_chat_speculative_type
and later models with no saved config inherited it. Persist the global
preference only when the value came from the global settings.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: record the compare pane's loaded context in runtime state so the active model's settings and any reload or save use it, not the previous context
* Studio: notify the user when a Hub autoload can't start because another download for the model is already running, instead of silently dropping it
* Studio: drop the merge's orphaned staged-model store helpers and unused alert imports
The main merge left isPendingGguf and pendingSelectionMatches referencing the
removed PendingModelSelection type, and the alert-dialog/alert imports unused
after the permission-mode dropdown replaced the bypass dialog, so tsc -b failed.
* Studio: cache a null default chat template so the viewer stops re-fetching it
A model with no sidecar or embedded template resolves to a terminal null, but
that result was never cached, so reopening the template viewer re-ran the
backend and Hugging Face lookup every time.
* Studio: detect direct-file GGUFs in run settings so Max Tokens uses their context
A GGUF loaded from a local file or custom folder has no variant label, so the
run-settings panel treated it as non-GGUF and clamped Max Tokens to the session
max_seq_length instead of the loaded GGUF context. Detect it via the reported
GGUF context and the .gguf checkpoint suffix, matching the chat page.
* Studio: prompt to re-select a local model file when its lease expired before reload
A file-picked GGUF is reachable only through a native path token that the
desktop host prunes after a TTL. Reloading reused that token blindly, so a
reload long after the initial load failed with an opaque error. Track the
token's expiry and, when it has passed, ask the user to re-select the file
instead of attempting a doomed reload.
* Fix descender-clipping test to tolerate sidebar layout utilities
The sidebar account-block div carries layout utilities (min-w-0, flex-1)
between 'flex' and 'flex-col', so the descender-clipping guard's regex,
which required 'flex' immediately followed by 'flex-col', no longer matched
and the test failed to locate the account-block div. Generalize the prefix
to allow intervening flex utilities while still capturing the leading-*
class before the collapsible visibility utility and asserting leading-tight,
so the guard against clipped glyph descenders is fully preserved.
* Harden picker chat-template resolution
Enforce the 64 KiB chat-template contract at the validate endpoint's request
model so a direct caller cannot submit a template far larger than the frontend
allows (MaxBodyMiddleware only bounds the whole request body, not this field);
oversized templates now return a clean 422.
Apply sidecar-over-GGUF template precedence globally across cached snapshots
instead of per snapshot. A repo with multiple cached revisions previously
returned the first snapshot's template, so a newer GGUF-only revision could
win over an older revision's maintained chat_template.jinja sidecar, which
contradicted the documented intent that sidecars supersede the embedded copy.
* Guard per-model config against future-schema and lossy migration
Two forward-compatibility gaps in the versioned per-model config store:
- The load/apply path returned and normalized a stored record without checking
its schema version, so a record written by a newer client was reinterpreted
under the current schema and applied to a live model load, even though save,
delete and eviction all refuse to touch future-schema records. Reject
future-schema records on load too.
- The one-time legacy migration enforced the storage budget without protecting
the entries it had just migrated and set the completion flag unconditionally.
When storage was already full of future-schema records (which are unevictable
by an older client), the migrated entries were the only evictable ones and
could be dropped while migration was still marked complete. Protect the
migrated keys during eviction and only mark migration complete when they
survive, so it retries once space frees up.
* Discard chat-template validation results after the dialog closes
Server-side template validation is async, but closing or cancelling the editor
did not abort it, so a late-arriving valid response still called onSave and
applied a template the user had already dismissed. Track a validation token
that is bumped on close and ignore any validation result whose token is stale.
* Record native lease expiry when loading a picked GGUF from the chip
The pending-native-model chip loaded via stageOrLoad directly, bypassing
loadNativeModelIntent, so activeNativePathExpiresAtMs was never recorded for a
chip-loaded file. A later reload then either skipped the lease-expiry guard
entirely (expiry left null) or compared against a previously loaded file's
stale expiry, so reload could reuse an already-pruned token or wrongly block a
still-valid one. Route the chip through loadNativeModelIntent, which builds the
same selection and records the expiry.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Prefer sidecar template for a directly selected local GGUF file
A direct .gguf file path read its embedded chat template without checking the
parent directory for a maintained sidecar (chat_template.jinja /
tokenizer_config.json), while directory and variant selections already prefer
the sidecar. That let the config editor preview or save a stale embedded
template for the same model depending on how it was selected. Check the parent
directory sidecars first, then fall back to the embedded copy, and cover both
paths with tests.
* Resolve cached chat template per revision, newest first
The earlier change searched every cached snapshot for a sidecar before
considering any snapshot's embedded GGUF template, which let an obsolete sidecar
from an older revision override the newest revision's template. Restore
per-snapshot resolution (newest first): a revision's sidecar still supersedes
its own embedded GGUF copy, but a newer revision is no longer overridden by an
older revision's sidecar.
* Preserve autoload transport conflicts and surface background busy downloads
- When a Hub autoload hits a transport conflict, keep pendingHubAutoLoad bound
instead of clearing it. Clearing it re-keyed the download surface and its
cleanup cancelled the conflict the toast tells the user to resolve, so the
Hub resume affordance was gone the moment it appeared. Return early on
conflict, mirroring the started branch, so resolving it from the Hub still
auto-loads on completion.
- The background-download branch handled started and conflict but silently
dropped a busy outcome, leaving the user with no feedback when a peer variant
of the same repo was already downloading. Surface the same busy toast the
autoload path uses.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: danielhanchen <michaelhan2050@gmail.com>
* fix(save): unsloth_push_to_hub_gguf(save_method="lora") raises NameError
unsloth_push_to_hub_gguf reads is_main_process at save.py:3181 but never
declares it. Its twin unsloth_save_pretrained_gguf declares it (2783) and
uses it the same way (2839) -- the LoRA branch was copied between the twins,
the parameter it depends on was not. There is no module-level global, so the
name resolves as a global load and the branch raises NameError 100% of the
time.
save_pretrained_gguf(save_method="lora", push_to_hub=True) raises a
ValueError that tells users to "use .push_to_hub_gguf(save_method='lora')
instead" -- the documented escape hatch is the broken call.
Add is_main_process to the signature, positioned as in the twin, and forward
it to unsloth_save_pretrained_gguf on the merged path so the parameter is not
silently ignored there. Default stays True, so nothing changes for existing
callers.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* fix(save): preserve GGUF push compatibility
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: Etherll <61019402+Etherll@users.noreply.github.com>
Daniel Han