Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

Trim comments in the MLX DoRA changes

Daniel Han15866d2d7d85
diff --git a/studio/backend/core/training/worker.py b/studio/backend/core/training/worker.py
index f81c773f55d..3edfa77e2b0 100644
--- a/studio/backend/core/training/worker.py
+++ b/studio/backend/core/training/worker.py
@@ -2569,11 +2569,7 @@ def _normalize_mlx_studio_optimizer(value):
def _mlx_dora_peft_kwargs(config, get_peft_model):
- """LoRA kwargs a DoRA request adds for MLX, or raise why it cannot run.
-
- Only version skew is decided here; whether the selected modules can carry
- DoRA depends on the loaded model, so the MLX backend refuses those itself.
- """
+ """LoRA kwargs a DoRA request adds for MLX, or raise why it cannot run."""
import inspect
if not config.get("use_dora"):
@@ -2582,9 +2578,8 @@ def _mlx_dora_peft_kwargs(config, get_peft_model):
parameter = inspect.signature(get_peft_model).parameters.get("use_dora")
except (TypeError, ValueError):
parameter = None
- # The name alone is not support: `**use_dora` collects anything, which is
- # what the unsupporting version already does, and the positional-only and
- # `*use_dora` kinds cannot be bound by keyword at all.
+ # A **kwargs catch-all absorbs use_dora and trains plain LoRA, so the
+ # parameter must be named and bindable by keyword.
named = parameter is not None and parameter.kind in (
inspect.Parameter.POSITIONAL_OR_KEYWORD,
inspect.Parameter.KEYWORD_ONLY,
diff --git a/studio/backend/tests/test_mlx_training_worker_config.py b/studio/backend/tests/test_mlx_training_worker_config.py
index 67b43a5dcaa..2af8e113518 100644
--- a/studio/backend/tests/test_mlx_training_worker_config.py
+++ b/studio/backend/tests/test_mlx_training_worker_config.py
@@ -34,8 +34,8 @@ def _load_worker_module():
sys.modules["loggers"] = loggers
utils = types.ModuleType("utils")
- # Real package directory: an empty __path__ shadows the package and
- # breaks the worker's own imports. Only the stubs below replace it.
+ # An empty __path__ shadows the real package and breaks the worker's own
+ # imports; only the stubs below replace it.
utils.__path__ = [str(Path(__file__).resolve().parents[1] / "utils")]
sys.modules["utils"] = utils
@@ -115,8 +115,8 @@ def test_mlx_studio_rejects_unknown_scheduler():
def test_mlx_dora_requires_the_named_use_dora_parameter():
- # The release predating MLX DoRA swallows use_dora through this catch-all
- # and trains plain LoRA, so accepting-the-keyword is not support.
+ # A **kwargs catch-all absorbs use_dora and trains plain LoRA, so accepting the
+ # keyword is not support.
def old_zoo(
model,
r = 16,
@@ -153,16 +153,13 @@ def test_mlx_dora_requires_the_named_use_dora_parameter():
assert _mlx_dora_peft_kwargs({"use_dora": True}, usable) == {"use_dora": True}
with pytest.raises(NotImplementedError, match = "unsloth-zoo"):
_mlx_dora_peft_kwargs({"use_dora": True}, old_zoo)
- # None of these kinds prove support: two cannot be bound by keyword, and
- # `**use_dora` only collects it the way the unsupporting version does.
for unusable in (positional_only_zoo, var_positional_zoo, var_keyword_zoo):
with pytest.raises(NotImplementedError, match = "unsloth-zoo"):
_mlx_dora_peft_kwargs({"use_dora": True}, unusable)
- # An unreadable signature (inspect.signature raises here) is not support.
with pytest.raises(NotImplementedError, match = "unsloth-zoo"):
_mlx_dora_peft_kwargs({"use_dora": True}, object())
- # An image-bearing dataset is not proof of a vision model; a text model
- # can still train language-only DoRA from it.
+ # An image-bearing dataset is not proof of a vision model; a text
+ # model can still train language-only DoRA.
assert _mlx_dora_peft_kwargs(
{
"use_dora": True,
@@ -171,7 +168,6 @@ def test_mlx_dora_requires_the_named_use_dora_parameter():
},
new_zoo,
) == {"use_dora": True}
- # Nothing added, nothing checked, when DoRA was not requested.
assert _mlx_dora_peft_kwargs({}, old_zoo) == {}
assert _mlx_dora_peft_kwargs({"use_dora": False}, old_zoo) == {}
diff --git a/studio/backend/tests/test_training_cached_start.py b/studio/backend/tests/test_training_cached_start.py
index 89cdbb84ec3..ade099c1334 100644
--- a/studio/backend/tests/test_training_cached_start.py
+++ b/studio/backend/tests/test_training_cached_start.py
@@ -1279,8 +1279,7 @@ def test_mlx_start_rejects_unsupported_training_config(request_overrides, expect
def test_mlx_start_accepts_dora():
- # MLX trains DoRA, so the platform gate must let it through. LoftQ is
- # asserted alongside it so a gate that stopped refusing anything at all
+ # LoftQ is asserted alongside DoRA so a gate that stopped refusing anything at all
# cannot pass this test.
from utils.hardware import hardware
diff --git a/studio/frontend/src/features/studio/sections/lora-params-section.tsx b/studio/frontend/src/features/studio/sections/lora-params-section.tsx
index c99cbbbb427..e7be24ac81c 100644
--- a/studio/frontend/src/features/studio/sections/lora-params-section.tsx
+++ b/studio/frontend/src/features/studio/sections/lora-params-section.tsx
@@ -59,9 +59,8 @@ export function LoraParamsSection(): ReactElement | null {
const [open, setOpen] = useState(true);
const isCpt = store.trainingMethod === "cpt";
const showVisionLora = store.isVisionModel && store.isDatasetImage === true;
- // MLX refuses DoRA once vision layers are in the run. `isVisionModel` is
- // fetched metadata that can go stale, so this only stops the combination
- // being newly selected; the loaded model still settles an existing one.
+ // `isVisionModel` can go stale, so this blocks only a NEW selection;
+ // the backend settles an existing one.
const doraNeedsVisionOff =
deviceType === "mac" && showVisionLora && store.finetuneVisionLayers;
diff --git a/studio/frontend/src/features/studio/sections/use-mlx-training-config-policy.ts b/studio/frontend/src/features/studio/sections/use-mlx-training-config-policy.ts
index ad64cf9f014..d8d307accf9 100644
--- a/studio/frontend/src/features/studio/sections/use-mlx-training-config-policy.ts
+++ b/studio/frontend/src/features/studio/sections/use-mlx-training-config-policy.ts
@@ -15,8 +15,8 @@ export function useMlxTrainingConfigPolicy(): void {
})),
);
- // Only loftq: MLX trains DoRA, so clearing a `dora` selection here would
- // silently substitute plain LoRA for the run the user asked for.
+ // Only loftq: clearing a persisted `dora` selection would silently substitute plain
+ // LoRA.
useEffect(() => {
if (isMac && loraVariant === "loftq") {
useTrainingConfigStore.setState({ loraVariant: "lora" });
diff --git a/studio/frontend/tests/training-validation.test.ts b/studio/frontend/tests/training-validation.test.ts
index 2b04d225e0f..d75258f5189 100644
--- a/studio/frontend/tests/training-validation.test.ts
+++ b/studio/frontend/tests/training-validation.test.ts
@@ -290,8 +290,7 @@ test("training validation accepts DoRA on MLX under an adapter method", () => {
{ ...validConfig, trainingMethod, loraVariant: "dora" },
"mac",
),
- // cpt is refused on MLX for its own reason, which must not be read as
- // DoRA being refused.
+ // cpt is refused on MLX for its own reason, not for DoRA.
trainingMethod === "cpt"
? { ok: false, errorKey: "studio.params.notSupportedAppleSilicon" }
: { ok: true, errorKey: null },
diff --git a/tests/studio/test_mlx_training_worker_behaviors.py b/tests/studio/test_mlx_training_worker_behaviors.py
index 014cb57a16c..c09d75207a8 100644
--- a/tests/studio/test_mlx_training_worker_behaviors.py
+++ b/tests/studio/test_mlx_training_worker_behaviors.py
@@ -36,27 +36,7 @@ def test_run_mlx_training_passes_token_to_from_pretrained():
def test_mlx_dora_decided_before_load_and_merged_into_peft_kwargs():
- """The DoRA decision must be wired to happen before the base model is
- fetched, and to feed the wrap call.
-
- Dropping the merge would accept a DoRA request and silently train plain
- LoRA; deciding after the load would make an unsupported unsloth-zoo cost
- a multi-gigabyte download first.
-
- This checks WIRING, not runtime behavior: the decision precedes the load,
- takes the run config, probes the same callable the wrap invokes, and its
- result is merged into the mapping that wrap expands without being rebound
- in between. It cannot see through code that keeps the shape and defeats
- the effect -- burying the merge under a false condition, unsetting the key
- afterwards -- because the test and the code it reads live together; the
- real-model behavior is covered by the MLX backend's own tests.
-
- This asserts one specific shape — a `peft_kwargs.update(...)` call and a
- `**peft_kwargs` wrap call — rather than trying to recognize every
- equivalent formulation. Rewriting that seam another way (a merge
- operator, explicit unpacking) is expected to fail here and to be
- re-expressed, not worked around.
- """
+ """Wiring only: dropping the merge would silently train plain LoRA, and deciding after the load would make an unsupported unsloth-zoo cost a multi-gigabyte download first. One shape is asserted, so an equivalent rewrite is meant to fail here and be re-expressed."""
tree = ast.parse(WORKER.read_text(encoding = "utf-8"))
fn = _find_func(tree, "_run_mlx_training")
assert fn is not None
@@ -72,8 +52,6 @@ def test_mlx_dora_decided_before_load_and_merged_into_peft_kwargs():
):
decided_at = node.lineno
decision_target = node.targets[0].id
- # The run's own config and the same callable the wrap invokes --
- # probing anything else answers for something else.
passed = [ast.unparse(arg) for arg in node.value.args]
assert passed == [
"config",
@@ -107,7 +85,6 @@ def test_mlx_dora_decided_before_load_and_merged_into_peft_kwargs():
]
assert wraps, "FastMLXModel.get_peft_model is never called"
for wrap in wraps:
- # A real ** expansion, not a string that merely looks like one.
assert any(
kw.arg is None and ast.unparse(kw.value) == "peft_kwargs" for kw in wrap.keywords
), (
@@ -118,11 +95,8 @@ def test_mlx_dora_decided_before_load_and_merged_into_peft_kwargs():
assert (
min(merged_at) < first_wrap
), "peft_kwargs must be updated before get_peft_model is called"
- # ...and not replaced in between, which would drop the merge again.
- # A Store/Del context covers the rebinding forms a refactor realistically
- # produces. This is a regression check over one function's line range, not
- # a proof: string-bound names (import aliases, `case` captures) and
- # anything sharing the merge's line are outside it.
+ # Store/Del only: string-bound names (import aliases, `case` captures)
+ # and the merge's own line are outside this check.
rebound = [
node.lineno
for node in ast.walk(fn)