* fix(studio): infer mlx vlm resized image layout
* [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>
Raise ruff line-length to 100 and extend the local pre-commit format pipeline (def-signature magic-comma normalization, short multi-line assert collapse, kwarg '=' spacing, blank-line-after-short-import removal, adjacent string-literal / f-string+plain merge, redundant-pass pruning). Every transform re-checks the file AST and is dropped if it would differ; the whole-repo reformat is verified AST-identical per file and idempotent.
* Studio: add VLM image-size control for training
Studio vision fine-tuning had no explicit way to cap image resolution, so
users could not trade visual detail against context and memory use from the
training UI, YAML config, or API payload. :) Add a nullable `vision_image_size`
setting that keeps the current model default when unset and applies a
max-side resize when provided.
- Add `vision_image_size` to the training request model, route payload, backend
training config, and frontend API/types plumbing.
- Validate the value server-side as either null or an integer in the supported
256-2048 range.
- Surface an Image Size selector for vision LoRA training with Default plus
common preset sizes.
- Include the value in training start payloads only for image-dataset vision
models, and serialize it into vision-aware YAML configs.
- Map backend model defaults back into the training store and reset the value
when reapplying model defaults.
- Pass the resize through the Torch trainer via `UnslothVisionDataCollator`
using max-dimension semantics.
- Apply the same max-dimension resize in the MLX VLM path before mlx-vlm's
internal collation, preserving aspect ratio and avoiding upscaling.
- Add backend validation coverage and MLX resize-size tests for the new
behavior.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: thread vision_image_size into DeepSeek OCR + writable MLX ndarray
- trainer.py: DeepSeek OCR collator now honors the new vision_image_size
setting as image_size. Falls back to 640 when null. base_size stays at
1024 and crop_mode stays True so the Gundam preset's dynamic cropping
of large documents keeps working.
- worker.py: _resize_mlx_vlm_image returns np.array(image, copy=True)
instead of np.asarray(image). The PIL view from np.asarray is not
writable, which makes HF VLM processors emit "The given NumPy array
is not writable, and PyTorch does not support non-writable tensors..."
when they call torch.from_numpy. copy=True keeps the same shape and
dtype but produces a writable buffer.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: align YAML export gate with API mapper + extend Image Size dropdown
- training-section.tsx: handleSaveConfig now passes
isVisionModel && isDatasetImage === true to serializeConfigToYaml,
matching buildTrainingStartPayload. Stops vision_image_size from
leaking into exported YAML for text-only datasets where the API
would have sent null.
- params-section.tsx: add 256 to visionImageSizePresets so the
dropdown spans the validator's full [256, 2048] range. Also render
a synthetic SelectItem for the current value when it was loaded
from YAML or model defaults and is not in the preset list, so the
controlled Select always shows the active size.
* Studio: validate vision_image_size in YAML/model-default loader
mapBackendModelConfigToTrainingPatch now mirrors the backend validator
at studio/backend/models/training.py:169 by dropping any value that is
not an integer in [256, 2048]. Pre-fix, an imported YAML like
vision_image_size: 4096 or 640.5 would land in the store and the UI
would happily display it, only to fail when Start Training posted to
the backend. With this guard the store never holds a value the backend
would reject.
* Studio: precise error messages for invalid vision_image_size inputs
Switch the field_validator to mode="before" so True/False surface as
bool (not Pydantic's coerced 1/0) and give a precise
"must be an integer or null" message instead of the misleading
"must be in [256, 2048] (got 1)". Also explicitly accepts numpy
Integral and integral Real scalars so YAML or programmatic callers
using numpy ints keep working.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: test that bool inputs yield the precise 'integer or null' error
Regression guard for the validator switch to mode="before". Pre-fix,
vision_image_size: True was rejected with "must be in [256, 2048]
(got 1)" because Pydantic coerced before our check ran. New test
asserts the message now reads "integer or null".
* Studio: tighten vision_image_size loader + YAML save + MLX rounding
Round 2 of follow-up review surfaced three usability issues:
- model-defaults.ts: switching to a model whose backend YAML omits
vision_image_size now explicitly resets the store value to null.
Pre-fix, a stale 2048 from a previous model would silently apply
to the new run because every checked-in model-default file omits
the key.
- training-section.tsx: handleSaveConfig now includes vision fields
unless isDatasetImage is definitively false. isDatasetImage is null
during dataset checks, after dataset edits, and on import; treating
unknown as "drop" would silently lose the user's selection in those
windows. Confirmed-text-only datasets still drop the value.
- worker.py: _mlx_vlm_max_resized_size now mirrors the Torch collator's
integer formula (w * size + size_func // 2) // size_func instead of
Python round(), which uses banker's rounding and disagreed by 1px on
half-pixel inputs like 333x1000 with target 500 (was 166, now 167).
Test_mlx_training_worker_config gains parity assertions.
* Studio: reset vision_image_size in the model-config error fallback path
mapBackendModelConfigToTrainingPatch resets stale image size on the
success path, but if the /api/models/config endpoint throws,
training-config-store.ts falls through to checkVisionModel and only
updates capability flags. Pre-fix that left a stale 2048 (or any
prior selection) in the store, so once dataset detection marked the
new dataset as image, the next training start would silently apply
the previous model's size. The error branch now also resets to the
DEFAULT_HYPERPARAMS.visionImageSize sentinel.
* Studio: revert DeepSeek OCR Image Size knob + move missing-key reset
Round 3 of the parallel-reviewer pass surfaced two issues that I had
introduced earlier in this PR's follow-ups.
- trainer.py: my prior change threaded vision_image_size into the
DeepSeek OCR collator's image_size argument. The collator's
(image_size, base_size, crop_mode) is a single preset
(Tiny / Small / Base / Large / Gundam); changing image_size in
isolation desynchronizes the per-crop pixel grid from num_queries
downstream and produces wrong token grids on documents larger than
the per-crop tile. The fix pins the collator back at the Gundam
preset and logs a clear "ignored for DeepSeek OCR" notice when the
user has selected a non-default Image Size.
- model-defaults.ts + training-config-store.ts: the round 4 fix that
reset visionImageSize when a model YAML omitted the key also fired
on same-model reloads (ensureModelDefaultsLoaded re-fires on page
refresh), wiping a value the user had just selected. The reset is
now in setSelectedModel, gated on selectedModel != previousModel,
so true model switches still clear stale values while reloads keep
the user's selection.
* Studio: extend DeepSeek OCR Image Size exclusion to MLX + frontend
Round 4 of the parallel-reviewer pass flagged that the Torch trainer
exclusion I added did not have a matching MLX guard, and that the UI
still offered the dropdown for DeepSeek OCR even though the backend
ignores it.
- worker.py: _run_mlx_training now mirrors the Torch exclusion. When
the model name matches DeepSeek OCR, vision_image_size is forced
back to None before _adapt_for_mlx_vlm sees it, so dataset images
pass through unchanged just like the Torch path. Emits a clear
status line when this happens.
- params-section.tsx: the Image Size Row is now gated on
showVisionImageSize (showVisionLora && !isDeepseekOcr) instead of
showVisionLora alone, so DeepSeek OCR users no longer see a control
that silently has no effect.
- mappers.ts: buildTrainingStartPayload sends null for vision_image_size
whenever the selected model is DeepSeek OCR, so the backend log line
about ignoring the value never fires from a UI-driven start.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: tighten YAML import/save for vision_image_size
Two YAML-path asymmetries that could leak a stale image size into
training:
- parseYamlConfig now treats a missing training.vision_image_size as
null. Without this, importing a YAML saved before this feature (or
any config that omits the key) preserved whatever value the user had
previously set on a different model. The model-defaults reload path
still uses Object.hasOwn so same-model defaults reloads do not wipe
a manual selection; only file import normalises the missing key.
- handleSaveConfig now passes a DeepSeek-OCR-specific guard to
serializeConfigToYaml so saved YAML matches what the API mapper
actually sends. Previously a state with visionImageSize set could
emit the key even though Studio ignored it at training time for
DeepSeek OCR, and a later import for a non-DeepSeek vision model
would activate the stale value.
serializeConfigToYaml gains an optional third parameter
includeVisionImageSize defaulting to includeVisionFields, preserving
the existing 2-arg call signature for backwards compatibility.
* Studio: also reset vision_image_size when YAML lacks a training section
Round 9's parseYamlConfig normalization only fired when the YAML had a
training mapping that omitted vision_image_size. A lora-only or
logging-only YAML (or one with `training: null`) still left trainingObj
unset, the mapper saw no vision_image_size key, and the previously
selected store value persisted into the next training run.
Now an absent or null training section is synthesised as
{ vision_image_size: null } so model-defaults.ts always patches
visionImageSize back to Default on file import. Same-model defaults
reloads still preserve manual choices via the existing Object.hasOwn
gate in mapBackendModelConfigToTrainingPatch.
* Studio: unify parseYamlConfig non-object training handling
A fresh static review (Opus subagent) flagged P3-1: parseYamlConfig
only synthesised vision_image_size: null when raw.training was either
absent or a plain object missing the key. If raw.training is a scalar
or an array (malformed but still parseable), the value was passed
through unchanged, the mapper's Object.hasOwn returned false, and any
previously selected visionImageSize persisted - the same stale-state
leak the lora-only fallback was added to close.
Treat any non-plain-object raw.training (null, array, scalar) as a
malformed/missing section and reset to { vision_image_size: null }.
* Studio: tighten code comments for vision_image_size path
* Studio: tighten vision_image_size validator + restore lost comment context
Two issues surfaced by a fresh adversarial review of the validator:
1. v.strip().lstrip("+-").isdigit() let "++512" / "--256" / "+-+512"
slip past the gate, then int("++512") raised an uncaught ValueError
and Pydantic surfaced "invalid literal for int() with base 10: '++512'"
instead of the contracted "vision_image_size must be an integer or null".
2. str.isdigit() returns True for Unicode digit families (full-width '５１２',
Arabic-Indic '٥١٢', Devanagari '१०२४'), and int() coerces them, so the
value reaching the backend wasn't the ASCII the user typed.
Replaced the lstrip+isdigit pair with re.fullmatch(r'[+-]?[0-9]+', stripped),
which rejects both shapes with the precise error and accepts the documented
ones ('256', '+512', ' 1024 '). Added 8 regression test cases covering
multi-sign strings, lone sign, and the three Unicode digit families.
Also restored comment context lost in f9c39331:
- model-defaults.ts: name studio/backend/models/training.py:_check_vision_image_size
as the spec the [256, 2048] range mirrors, so a maintainer changing the
cap in one file can find the other.
- training-section.tsx: enumerate the three windows in which isDatasetImage
is null (before a check, after dataset edits, on import) so a future
maintainer doesn't simplify the gate to `isCheckingDataset`.
- worker.py: qualify the writable-ndarray comment with "when a resize is
requested" so it doesn't misadvertise the resize=None early-return.
---------
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: skip flash-attn install on Blackwell GPUs (sm_100+)
Dao-AILab does not publish prebuilt flash-attn wheels for sm_100, sm_120,
or sm_121, and the older-arch wheels fail to load on Blackwell. Add a
shared has_blackwell_gpu() helper and gate both the install-time
(install_python_stack._ensure_flash_attn) and runtime
(worker._ensure_flash_attn_for_long_context) paths on it. Detection uses
nvidia-smi --query-gpu=compute_cap, which works on Linux and Windows.
* test: stub has_blackwell_gpu in pre-existing runtime flash-attn tests
prefers_prebuilt_wheel and falls_back_to_pypi exercise the install
paths that the Blackwell guard now short-circuits. Make them explicit
about non-Blackwell so they pass on real Blackwell hosts.
* studio: cache has_blackwell_gpu, skip Blackwell warning under NO_TORCH
- Wrap has_blackwell_gpu in functools.lru_cache so repeated calls in a
single process avoid redundant nvidia-smi spawns. Tests clear the
cache via setup_method/teardown_method.
- In _ensure_flash_attn, run the NO_TORCH short-circuit before the
Blackwell check so GGUF-only users (who never install torch anyway)
do not see a Blackwell warning. Blackwell check still runs above the
IS_WINDOWS / IS_MACOS gates so Blackwell-on-Windows users still see
the explicit reason rather than a silent OS skip.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* test: add has_blackwell_gpu to mlx worker test wheel_utils stub
test_mlx_training_worker_config loads worker.py against a hand-rolled
utils.wheel_utils stub. Adding has_blackwell_gpu to the stub symbol
list so worker's import line resolves.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Add Apple Silicon MLX routing
Rewrite __init__.py: detect MLX on macOS arm64 before any torch imports
Extract original GPU init to _gpu_init.py (unchanged)
MLX path imports FastMLXModel from unsloth_zoo, skips all GPU code
GPU path unchanged: from ._gpu_init import *
* Add Apple Silicon MLX routing
- Rewrite __init__.py: detect MLX on macOS arm64 before any torch imports
- Extract original GPU init to _gpu_init.py (unchanged)
- MLX path imports FastMLXModel from unsloth_zoo, skips all GPU code
- GPU path unchanged: from ._gpu_init import *
* mlx with studio
* mlx with studio
* updating temporary install.sh
* updating temporary install.sh
* adding t_v5 path
* adding t_v5 path
* fixing vision training
* fixing vision training
* adding chat
* adding chat
* minor
* minor
* Adding export and fixing training issues, inference with lora adaptors
* Adding export and fixing training issues, inference with lora adaptors
* fix: MLX worker pass load_in_4bit, override is_vlm based on dataset, streaming for VLM
* fix: MLX worker pass load_in_4bit, override is_vlm based on dataset, streaming for VLM
* Merge mlx-apple-silicon into main
* update install.sh to point to main branch
* update install.sh to point to main branch
* fix: export returns 3 values (success, message, output_path) matching upstream worker
* fix: export returns 3 values (success, message, output_path) matching upstream worker
* fix(mlx): show training-process peak memory in Studio UI, not system-wide
Studio UI was showing ~95 GB during MLX training because get_gpu_utilization
read "In use system memory" from IORegistry's AGXAccelerator — system-wide
GPU memory across all processes (training + backend + browser + Display).
Now the trainer's mx.get_peak_memory value is forwarded through the
progress event and surfaced via /api/train/hardware while training is
active. Falls back to the system-wide reading when training is not running.
* fix(mlx): show training-process peak memory in Studio UI, not system-wide
Studio UI was showing ~95 GB during MLX training because get_gpu_utilization
read "In use system memory" from IORegistry's AGXAccelerator — system-wide
GPU memory across all processes (training + backend + browser + Display).
Now the trainer's mx.get_peak_memory() value is forwarded through the
progress event and surfaced via /api/train/hardware while training is
active. Falls back to the system-wide reading when training is not running.
* fix(mlx): make is_bfloat16_supported detect M1/M2 (no native bf16)
M1 and M2 chips emulate bf16 in software on the GPU, causing 40-70%
slower prefill compared to native fp16. M3+ have native bf16 (macOS
Sonoma+ MPSGraph). Replaces the always-True stub with chip-aware
detection via mx.device_info.
* fix(mlx): make is_bfloat16_supported() detect M1/M2 (no native bf16)
M1 and M2 chips emulate bf16 in software on the GPU, causing 40-70%
slower prefill compared to native fp16. M3+ have native bf16 (macOS
Sonoma+ MPSGraph). Replaces the always-True stub with chip-aware
detection via mx.device_info().
* feat(mlx): wire training_type="Full Finetuning" through MLX worker
Compute use_lora from the UI's training_type before loading the model,
pass full_finetuning=not use_lora to FastMLXModel.from_pretrained, and
let the existing 'if use_lora' branch skip get_peft_model. Matches the
GPU worker's flow.
* feat(mlx): wire training_type="Full Finetuning" through MLX worker
Compute use_lora from the UI's training_type before loading the model,
pass full_finetuning=not use_lora to FastMLXModel.from_pretrained, and
let the existing 'if use_lora' branch skip get_peft_model. Matches the
GPU worker's flow.
* fix(mlx): pass save_method='merged_16bit' from Studio's export page
Previously the MLX path called save_pretrained_merged with no
save_method, which fell through to a no-op that didn't actually fuse
LoRA into the base. Now Studio's "Merged Model" export properly
fuses LoRA + dequantizes any 4-bit base to bf16, matching the GPU
behavior for the same UI option.
* fix(mlx): pass save_method='merged_16bit' from Studio's export page
Previously the MLX path called save_pretrained_merged() with no
save_method, which fell through to a no-op that didn't actually fuse
LoRA into the base. Now Studio's "Merged Model" export properly
fuses LoRA + dequantizes any 4-bit base to bf16, matching the GPU
behavior for the same UI option.
* fix(studio): pass private to MLX push, return 3-tuples consistently
MLX push_to_hub branch now forwards private=private (matches GPU)
Existing 2-tuple early-returns ('repo_id+token required', 'PEFT model
needed') were tripping the route's 3-tuple unpack. Added a None
output_path so the unpack always succeeds.
* fix(studio): pass private to MLX push, return 3-tuples consistently
- MLX push_to_hub branch now forwards private=private (matches GPU)
- Existing 2-tuple early-returns ('repo_id+token required', 'PEFT model
needed') were tripping the route's 3-tuple unpack. Added a None
output_path so the unpack always succeeds.
* studio wirings
* studio wirings
* Merge pull request #5 from Manan17/feat/quant_config
studio wirings
* fix(mlx): wire train_on_completions for VLM via per-template lookup
Mirror the GPU worker: stop excluding VLMs and stop hardcoding
template detection. Look up the model in MODEL_TO_TEMPLATE_MAPPER and
fetch the per-template instruction/response markers from
TEMPLATE_TO_RESPONSES_MAPPER. The frontend already force-disables
train_on_completions for vision+image and audio cases, so backend
just trusts the flag.
* fix(mlx): wire train_on_completions for VLM via per-template lookup
Mirror the GPU worker: stop excluding VLMs and stop hardcoding
template detection. Look up the model in MODEL_TO_TEMPLATE_MAPPER and
fetch the per-template instruction/response markers from
TEMPLATE_TO_RESPONSES_MAPPER. The frontend already force-disables
train_on_completions for vision+image and audio cases, so backend
just trusts the flag.
* wire in lora rslora, init lora weights, random_state
* wire in lora rslora, init lora weights, random_state
* loftq studio error message fix
* loftq studio error message fix
* handle unknown optim and lr scheduler
* handle unknown optim and lr scheduler
* Merge pull request #6 from Manan17/update/peftkwargs
Update/peftkwargs
* feat(mlx): pass finetune_language/attention/mlp/vision flags to FastMLXModel
Studio's four UI checkboxes now actually flow through to MLX get_peft_model
(which was just updated in unsloth-zoo to honor them). Also drops the
incorrect train_projector wiring that tied projector LoRA to the
attn/mlp flags — those are language-side toggles, not projector toggles.
Co-Authored-By: Manan17 <shahmanan170602@gmail.com>
* feat(mlx): pass finetune_language/attention/mlp/vision flags to FastMLXModel
Studio's four UI checkboxes now actually flow through to MLX get_peft_model
(which was just updated in unsloth-zoo to honor them). Also drops the
incorrect train_projector wiring that tied projector LoRA to the
attn/mlp flags — those are language-side toggles, not projector toggles.
Co-Authored-By: Manan17 <shahmanan170602@gmail.com>
* feat(mlx,ux): auto-imply finetune_language_layers when user picks attn/mlp
UI guardrail. The four checkboxes (vision/language/attention/MLP) carry
"scope × module-type" semantics that aren't obvious — picking just
"Attention modules" + "MLP modules" without "Language layers" naturally
reads as "fine-tune attn/mlp" but our backend reads it as "fine-tune
attn/mlp modules in *no* tower" → empty target_modules → zero
trainable params → crash inside value_and_grad.
If user selected attn or mlp module types but no layer scope, default
to language scope. Power users can still explicitly choose
language=False, vision=True if they want vision-only fine-tuning of
attn/mlp.
Co-Authored-By: Manan17 <shahmanan170602@gmail.com>
* feat(mlx,ux): auto-imply finetune_language_layers when user picks attn/mlp
UI guardrail. The four checkboxes (vision/language/attention/MLP) carry
"scope × module-type" semantics that aren't obvious — picking just
"Attention modules" + "MLP modules" without "Language layers" naturally
reads as "fine-tune attn/mlp" but our backend reads it as "fine-tune
attn/mlp modules in *no* tower" → empty target_modules → zero
trainable params → crash inside value_and_grad.
If user selected attn or mlp module types but no layer scope, default
to language scope. Power users can still explicitly choose
language=False, vision=True if they want vision-only fine-tuning of
attn/mlp.
Co-Authored-By: Manan17 <shahmanan170602@gmail.com>
* fix(mlx): wire top_k, repetition_penalty, and VLM top_p through to mlx-lm/mlx-vlm
Inference UI sliders for top_k and repetition_penalty had no effect on
MLX, and VLM top_p was also silently dropped. Plus a latent pre-existing
bug: mlx_vlm.generate_step expects temperature= (long form), but we
were passing temp= which silently fell into **kwargs — every VLM chat
was effectively greedy regardless of the temperature slider.
Text path (_generate_text):
make_sampler now receives top_k in addition to temp/top_p
make_logits_processors built and forwarded when repetition_penalty is
non-trivial (skip when 0.0/1.0 to avoid pointless overhead)
VLM path (_generate_vlm):
Pass top_p, top_k, repetition_penalty as kwargs (mlx_vlm.stream_generate
forwards them to generate_step's sampler/logits_processor builders)
Rename temp= → temperature= so it's actually consumed
Verified end-to-end with a smoke test on Qwen2.5-0.5B-Instruct (text) and
Qwen2.5-VL-3B-Instruct (VLM): each of {greedy, top_p=0.5, top_k=10,
rep_pen=1.5} now produces a distinct output, proving the parameters
reach the sampler.
Co-Authored-By: Manan17 <shahmanan170602@gmail.com>
* fix(mlx): wire top_k, repetition_penalty, and VLM top_p through to mlx-lm/mlx-vlm
Inference UI sliders for top_k and repetition_penalty had no effect on
MLX, and VLM top_p was also silently dropped. Plus a latent pre-existing
bug: mlx_vlm.generate_step expects temperature= (long form), but we
were passing temp= which silently fell into **kwargs — every VLM chat
was effectively greedy regardless of the temperature slider.
Text path (_generate_text):
- make_sampler now receives top_k in addition to temp/top_p
- make_logits_processors built and forwarded when repetition_penalty is
non-trivial (skip when 0.0/1.0 to avoid pointless overhead)
VLM path (_generate_vlm):
- Pass top_p, top_k, repetition_penalty as kwargs (mlx_vlm.stream_generate
forwards them to generate_step's sampler/logits_processor builders)
- Rename temp= → temperature= so it's actually consumed
Verified end-to-end with a smoke test on Qwen2.5-0.5B-Instruct (text) and
Qwen2.5-VL-3B-Instruct (VLM): each of {greedy, top_p=0.5, top_k=10,
rep_pen=1.5} now produces a distinct output, proving the parameters
reach the sampler.
Co-Authored-By: Manan17 <shahmanan170602@gmail.com>
* feat(mlx): map format_type to MLX save_method, reuse local save dir for hub push
export_merged_model: format_type="4-bit (FP4)" → save_method="merged_4bit"
(was hardcoded merged_16bit, ignoring the UI choice).
Both export_merged_model and export_base_model now pass save_directory=
to push_to_hub_merged so it reuses the just-written local folder
instead of re-saving under a relative "username/model" directory.
Co-Authored-By: Manan17 <shahmanan170602@gmail.com>
* feat(mlx): map format_type to MLX save_method, reuse local save dir for hub push
- export_merged_model: format_type="4-bit (FP4)" → save_method="merged_4bit"
(was hardcoded merged_16bit, ignoring the UI choice).
- Both export_merged_model and export_base_model now pass save_directory=
to push_to_hub_merged so it reuses the just-written local folder
instead of re-saving under a relative "username/model" directory.
Co-Authored-By: Manan17 <shahmanan170602@gmail.com>
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* restore install
* restore install
* fix(mlx): restore FastVisionModel as a distinct class
unsloth/__init__.py was assigning `FastVisionModel = FastLanguageModel`
right after defining `class FastVisionModel(FastLanguageModel)` with a
`for_training` static method. The alias erased the class binding, so
the documented `FastVisionModel.for_training(model)` call from upstream
Unsloth's VLM notebooks raised `AttributeError` on MLX.
Remove the offending alias. `FastVisionModel` is now a real subclass of
`FastLanguageModel` again — inherits `from_pretrained` /
`get_peft_model` / `for_inference`, exposes `for_training` as a no-op
pass-through (no-op because MLX doesn't have a train/eval mode flag;
the call exists purely for GPU/MLX notebook parity).
Verified end-to-end: Qwen3-VL-2B + LaTeX_OCR LoRA + vision LoRA via
FastVisionModel.from_pretrained → get_peft_model → for_training →
MLXTrainer.train runs 10 steps cleanly (loss 1.10 → 0.12, no NaNs,
peak 5.89 GB).
Studio's path (FastLanguageModel.from_pretrained for any repo,
auto-detect VLM in the loader) is unaffected. Tier-1 review finding #8.
* fix(mlx): restore FastVisionModel as a distinct class
unsloth/__init__.py was assigning `FastVisionModel = FastLanguageModel`
right after defining `class FastVisionModel(FastLanguageModel)` with a
`for_training` static method. The alias erased the class binding, so
the documented `FastVisionModel.for_training(model)` call from upstream
Unsloth's VLM notebooks raised `AttributeError` on MLX.
Remove the offending alias. `FastVisionModel` is now a real subclass of
`FastLanguageModel` again — inherits `from_pretrained` /
`get_peft_model` / `for_inference`, exposes `for_training` as a no-op
pass-through (no-op because MLX doesn't have a train/eval mode flag;
the call exists purely for GPU/MLX notebook parity).
Verified end-to-end: Qwen3-VL-2B + LaTeX_OCR LoRA + vision LoRA via
FastVisionModel.from_pretrained → get_peft_model → for_training →
MLXTrainer.train() runs 10 steps cleanly (loss 1.10 → 0.12, no NaNs,
peak 5.89 GB).
Studio's path (FastLanguageModel.from_pretrained for any repo,
auto-detect VLM in the loader) is unaffected. Tier-1 review finding #8.
* Studio: harden MLX training and export, restore GPU init guards
Studio export
Restore Tuple[bool, str, Optional[str]] contract on export_merged_model,
export_base_model, export_gguf, and export_lora_adapter, populating
output_path on successful local saves so routes/worker/CLI/frontend
details.output_path is non-empty again.
Lift the GPU save_method assignment out of the local-save branch so
Hub-only merged exports (save_directory='', push_to_hub=True) no longer
hit UnboundLocalError on the push branch.
For MLX merged and base hub-only export, stage to a tempfile.TemporaryDirectory
before push_to_hub_merged instead of passing save_directory=''.
Source _IS_MLX from unsloth instead of recomputing the platform check
(single source of truth, also enforces mlx-package availability).
Studio MLX training/inference
Pass token=hf_token into FastMLXModel.from_pretrained for gated/private
models, matching the inference path.
Strip hf_token and wandb_token from wandb.init(config=...) so secrets
do not leak into the W&B run config.
Replace load_from_disk(local_datasets[0]) with the existing
UnslothTrainer._resolve_local_files / _loader_for_files helpers so
uploaded JSON/JSONL/CSV/Parquet files train through the normal datasets
loader (load_from_disk still used for HF save_to_disk directories).
Make the dataset slice helper inclusive at the end and treat 0 as a real
index instead of "unset", matching the GPU and embedding paths.
Add a status_message -> message alias inside _send so the existing parent
pump (training.py) renders MLX status updates instead of blanks.
Forward min_p through generate_chat_response into _generate_text /
_generate_vlm and into make_sampler / vlm_kwargs so the sampling control
is no longer a no-op on MLX.
Wrap unsloth_zoo.mlx_loader / mlx_trainer imports with a clearer
ImportError pointing users at install.sh for Apple Silicon.
Exit the MLX stop-polling thread on EOFError/OSError instead of
busy-looping when the queue/pipe is permanently closed (one-line
why-safe rationale inline).
Studio frontend
ParamsSection subscribes to platform deviceType via the Zustand hook so
the gradient checkpointing dropdown re-renders after the async device
fetch completes.
Studio hardware
get_gpu_utilization MLX branch now reads _read_apple_gpu_stats once and
derives VRAM totals from psutil, removing the second ioreg subprocess
per utilization poll.
Unsloth core
Restore the os.geteuid == 0 guard around the CUDA ldconfig recovery
that was lost when GPU initialization moved into _gpu_init.py, plus the
non-root manual-fix warning branch. Non-root CUDA users no longer shell
out to ldconfig at import time.
Load dataprep/raw_text via importlib so the MLX import path no longer
pulls torch in through dataprep/__init__.py -> synthetic.py.
FastVisionModel.from_pretrained overrides the inherited delegator only
to inject text_only=False; this is an extension, not a duplication, and
is needed so VLM checkpoint loads keep the vision tower.
Wrap the MLX-branch unsloth_zoo import with a clearer ImportError.
* Studio: regression tests for MLX training/export and GPU init ldconfig guard
tests/python/test_gpu_init_ldconfig_guard.py asserts the geteuid root
check still wraps the ldconfig recovery and the non-root branch warns
bnb users; AST + source-text inspection so the test runs without torch.
tests/studio/test_export_output_path_contract.py covers the
Tuple[bool, str, Optional[str]] return contract on every export method,
the output_path assignment after successful local save, the Hub-only
GPU save_method binding fix, the MLX hub-only TemporaryDirectory
staging, and the single-source `_IS_MLX` import from unsloth.
tests/studio/test_mlx_training_worker_behaviors.py covers token
forwarding to FastMLXModel.from_pretrained, wandb config secret
stripping, file-aware local dataset loading, status_message ->
message aliasing, inclusive slice semantics, EOFError/OSError stop
thread exit, and the friendly mlx_loader / mlx_trainer ImportError.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix(mlx): cap inference memory + release wired on unload + tame worker pre-pin
Three memory-hardening fixes for Studio's MLX path:
1. Inference applies the same Metal caps as the trainer.
load_model previously only called set_wired_limit(100% of recommended)
with no upper memory_limit, leaving large VLM checkpoints unbounded
during the loader allocation. Add _configure_memory_limits() that sets
memory_limit to 85% of recommended and wired_limit to min(recommended,
memory_limit) — matching MLXTrainer's defaults so behavior is the same
whether the user trains or just runs inference.
2. unload_model releases pinned memory back to the OS — but only when
the cache is empty. Without this, pinned wired bytes stayed allocated
to MLX after the model was gone, starving other apps. The release is
guarded on `not self.models` so unloading one of several cached
models doesn't un-pin weights still in use.
3. Worker pre-cap is conservative instead of aggressive.
The previous pre-pin set_wired_limit(100% of recommended) competed
with MLXTrainer's later more conservative cap. Replace with the same
85%-memory / min(rec, memory) pair that the trainer applies later
(idempotent re-apply). Bounds the model load + LoRA setup window
without over-pinning.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* tests/studio: regression tests for the _IS_MLX dispatch gate
Two gates drive every MLX-vs-CUDA dispatch decision in Studio:
1. unsloth._IS_MLX in unsloth/__init__.py — evaluated once at import
time, read by Studio worker code to choose the GPU vs MLX trainer
and inference paths. Defined as
Darwin AND arm64 AND find_spec("mlx") is not None.
2. utils.hardware.detect_hardware() — runtime probe with priority
CUDA > XPU > MLX > CPU. The MLX branch is reached only when both
CUDA and XPU are unavailable and the host is Apple Silicon and
mlx is importable.
Neither gate had a direct test. Adds tests/studio/test_is_mlx_dispatch_gate.py
with six tests:
test_is_mlx_gate_uses_three_required_predicates
AST-walks unsloth/__init__.py and asserts the _IS_MLX assignment
is a BoolOp(And) of platform.system()=="Darwin",
platform.machine()=="arm64", and find_spec("mlx") is not None.
Catches accidental rewrites that drop a predicate.
test_is_mlx_gate_true_on_apple_silicon_with_mlx_present
Spoofs platform to Darwin/arm64, injects a fake mlx module so
find_spec returns a real ModuleSpec, re-evaluates the gate
expression. Verifies it flips True under the exact conditions
Studio expects.
test_is_mlx_gate_false_when_mlx_missing
Spoofs Apple Silicon but with mlx absent. Verifies the gate stays
False (so a Mac without mlx installed does not pretend to have
MLX support).
test_is_mlx_gate_false_on_non_apple_silicon
Canary on the actual Linux+CUDA / AMD / Intel test host: the gate
must remain False regardless of whether mlx happens to be
importable. Protects existing GPU users from accidental MLX
hijack when MLX support evolves.
test_detect_hardware_picks_mlx_when_only_apple_silicon_available
Forces torch.cuda and torch.xpu off, spoofs Apple Silicon, injects
fake mlx and mlx.core. detect_hardware() must return DeviceType.MLX.
test_detect_hardware_picks_cuda_on_real_host
Canary: on a real CUDA host detect_hardware() must return
DeviceType.CUDA. Protects against the MLX branch shadowing CUDA
dispatch on NVIDIA / AMD ROCm hosts.
Uses the same monkeypatch.setitem(sys.modules, ...) fake-mlx pattern as
the existing test_mlx_inference_backend.py — no new test infrastructure,
no real mlx install required.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Add AGPL-3.0 SPDX header to Studio MLX regression tests
Four Studio MLX test files shipped without an SPDX-License-Identifier:
studio/backend/tests/test_mlx_training_worker_config.py
tests/studio/test_mlx_training_worker_behaviors.py
tests/studio/test_export_output_path_contract.py
tests/studio/test_is_mlx_dispatch_gate.py
They sit in or alongside studio/backend/, which is governed by
studio/LICENSE.AGPL-3.0, and exercise AGPL Studio code. Add the same
"# SPDX-License-Identifier: AGPL-3.0-only" header that's already on
test_mlx_inference_backend.py so the license declaration matches
the code under test rather than defaulting to the repo-root
Apache-2.0.
* Wrap MLX submodule imports with friendly install hint
The _IS_MLX block at the top of unsloth/__init__.py already catches the
missing-package case with a friendly install hint, but the follow-up
"from unsloth_zoo.mlx_trainer import ..." and "from unsloth_zoo.mlx_loader import ..."
lines run unguarded. An Apple Silicon user who has unsloth-zoo installed
but on an older version (e.g. the current PyPI release, before the MLX
modules ship) sees a raw ImportError on the submodule rather than the
hint that points at install.sh.
Wrap the two submodule imports in the same try/except shape so the
friendly install message fires whether the package is missing entirely
or just predates the MLX submodules. No-op once both packages release
together; smooths the transitional window where unsloth/main has merged
but unsloth-zoo on PyPI has not.
---------
Co-authored-by: DoubleMathew <mmathew23@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Long Yixing