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.

fix(studio): track the repository resolved by the training loader (#11462)

Lee Jacksonb490d56bc2d3
diff --git a/studio/backend/core/training/trainer.py b/studio/backend/core/training/trainer.py
index 6cb64cbd2e9..2a1f8ac5523 100644
--- a/studio/backend/core/training/trainer.py
+++ b/studio/backend/core/training/trainer.py
@@ -870,6 +870,7 @@ class UnslothTrainer:
actual_model_repo_id: Optional[str] = None,
model_revision: Optional[str] = None,
use_gradient_checkpointing: Union[str, bool] = "unsloth",
+ on_model_resolved: Optional[Callable[[str], None]] = None,
) -> bool:
"""Load model for training (supports both text and vision models)"""
self.load_in_4bit = load_in_4bit
@@ -1037,6 +1038,7 @@ class UnslothTrainer:
revision = model_revision,
use_exact_model_name = model_revision is not None,
use_gradient_checkpointing = use_gradient_checkpointing,
+ on_model_resolved = on_model_resolved,
)
logger.info("Loaded CSM audio model")
@@ -1058,6 +1060,7 @@ class UnslothTrainer:
revision = model_revision,
use_exact_model_name = model_revision is not None,
use_gradient_checkpointing = use_gradient_checkpointing,
+ on_model_resolved = on_model_resolved,
)
self.model.generation_config.language = "<|en|>"
self.model.generation_config.task = "transcribe"
@@ -1079,6 +1082,7 @@ class UnslothTrainer:
revision = model_revision,
use_exact_model_name = model_revision is not None,
use_gradient_checkpointing = use_gradient_checkpointing,
+ on_model_resolved = on_model_resolved,
)
logger.info(f"Loaded {self._audio_type} audio model (FastLanguageModel)")
@@ -1099,6 +1103,8 @@ class UnslothTrainer:
if local_files_only:
repo_path = lookup_name
else:
+ if on_model_resolved is not None:
+ on_model_resolved(hf_repo)
repo_path = snapshot_download(
hf_repo,
revision = model_revision,
@@ -1117,6 +1123,7 @@ class UnslothTrainer:
token = hf_token,
trust_remote_code = trust_remote_code,
use_gradient_checkpointing = use_gradient_checkpointing,
+ on_model_resolved = on_model_resolved,
)
logger.info("Loaded Spark-TTS (bicodec) model")
@@ -1133,6 +1140,7 @@ class UnslothTrainer:
revision = model_revision,
use_exact_model_name = model_revision is not None,
use_gradient_checkpointing = use_gradient_checkpointing,
+ on_model_resolved = on_model_resolved,
)
logger.info("Loaded OuteTTS (dac) model (FastModel)")
@@ -1150,6 +1158,7 @@ class UnslothTrainer:
revision = model_revision,
use_exact_model_name = model_revision is not None,
use_gradient_checkpointing = use_gradient_checkpointing,
+ on_model_resolved = on_model_resolved,
)
logger.info("Loaded audio VLM model (FastModel)")
@@ -1166,6 +1175,7 @@ class UnslothTrainer:
revision = model_revision,
use_exact_model_name = model_revision is not None,
use_gradient_checkpointing = use_gradient_checkpointing,
+ on_model_resolved = on_model_resolved,
)
logger.info("Loaded vision model")
@@ -1194,6 +1204,7 @@ class UnslothTrainer:
revision = model_revision,
use_exact_model_name = model_revision is not None,
use_gradient_checkpointing = use_gradient_checkpointing,
+ on_model_resolved = on_model_resolved,
)
logger.info("Loaded text model")
@@ -1245,6 +1256,7 @@ class UnslothTrainer:
actual_model_repo_id = actual_model_repo_id,
model_revision = model_revision,
use_gradient_checkpointing = use_gradient_checkpointing,
+ on_model_resolved = on_model_resolved,
)
error_msg = str(e)
error_lower = error_msg.lower()
diff --git a/studio/backend/core/training/training.py b/studio/backend/core/training/training.py
index 15cba413722..4442ecf098c 100644
--- a/studio/backend/core/training/training.py
+++ b/studio/backend/core/training/training.py
@@ -1203,6 +1203,7 @@ class TrainingBackend:
# Xet -> HTTP model-load fallback state (config kept for the respawn).
self._last_full_config: Optional[dict] = None
self._in_model_load: bool = False
+ self._model_download_repo_id: Optional[str] = None
self._xet_fallback_used: bool = False
self._needs_xet_respawn: bool = False
@@ -1930,6 +1931,7 @@ class TrainingBackend:
self._last_full_config = config
self._last_hf_cache_env = cache_env
self._in_model_load = False
+ self._model_download_repo_id = None
self._xet_fallback_used = False
self._needs_xet_respawn = False
@@ -2642,6 +2644,7 @@ class TrainingBackend:
)
with self._lock:
self._in_model_load = False
+ self._model_download_repo_id = None
self._event_queue = event_queue
self._stop_queue = stop_queue
self._proc = new_proc
@@ -3019,6 +3022,12 @@ class TrainingBackend:
if etype == "model_load_started":
with self._lock:
self._in_model_load = True
+ self._model_download_repo_id = None
+ return
+ if etype == "model_load_resolved":
+ with self._lock:
+ if self._in_model_load:
+ self._model_download_repo_id = event["repo_id"]
return
if etype == "model_load_completed":
with self._lock:
diff --git a/studio/backend/core/training/worker.py b/studio/backend/core/training/worker.py
index d3721e4342a..395b40e525d 100644
--- a/studio/backend/core/training/worker.py
+++ b/studio/backend/core/training/worker.py
@@ -4224,6 +4224,13 @@ def run_training_process(*, event_queue: Any, stop_queue: Any, config: dict) ->
),
xet_disabled = os.environ.get("HF_HUB_DISABLE_XET") == "1",
)
+
+ def _report_model_repo(repo_id):
+ # Local cache loads have no active Hub download to track.
+ if os.path.isdir(os.path.expanduser(repo_id)):
+ return
+ event_queue.put({"type": "model_load_resolved", "repo_id": repo_id, "ts": time.time()})
+
# Latest-sidecar models load 16-bit: bnb 4-bit feeds quantized experts into unvalidated paths.
try:
_train_load_in_4bit = _effective_training_load_in_4bit(
@@ -4238,6 +4245,7 @@ def run_training_process(*, event_queue: Any, stop_queue: Any, config: dict) ->
model_load_name,
)
success = trainer.load_model(
+ on_model_resolved = _report_model_repo,
model_name = model_name,
max_seq_length = config["max_seq_length"],
load_in_4bit = _train_load_in_4bit,
@@ -4303,6 +4311,7 @@ def run_training_process(*, event_queue: Any, stop_queue: Any, config: dict) ->
success = False
else:
success = trainer.load_model(
+ on_model_resolved = _report_model_repo,
model_name = model_name,
max_seq_length = config["max_seq_length"],
load_in_4bit = _train_load_in_4bit,
diff --git a/studio/backend/routes/training.py b/studio/backend/routes/training.py
index 167804cde40..e61a107c956 100644
--- a/studio/backend/routes/training.py
+++ b/studio/backend/routes/training.py
@@ -2272,6 +2272,9 @@ def _build_training_status(
"loss": getattr(progress, "loss", None),
"learning_rate": getattr(progress, "learning_rate", None),
"output_dir": getattr(backend, "_output_dir", None) or None,
+ "model_download_repo_id": (
+ getattr(backend, "_model_download_repo_id", None) if is_active else None
+ ),
}
metric_history = None
diff --git a/studio/backend/tests/test_job_accounts.py b/studio/backend/tests/test_job_accounts.py
index bdb6dd0b092..32cecf274db 100644
--- a/studio/backend/tests/test_job_accounts.py
+++ b/studio/backend/tests/test_job_accounts.py
@@ -811,6 +811,7 @@ def test_owner_training_start_status_cancel_bytes_are_unchanged(training, monkey
"loss": None,
"learning_rate": None,
"output_dir": None,
+ "model_download_repo_id": None,
},
"metric_history": None,
}
diff --git a/studio/backend/tests/test_training_status_terminal.py b/studio/backend/tests/test_training_status_terminal.py
index a7078b7956e..39414121571 100644
--- a/studio/backend/tests/test_training_status_terminal.py
+++ b/studio/backend/tests/test_training_status_terminal.py
@@ -277,3 +277,51 @@ def test_surfaces_tolerate_a_backend_without_is_run_finished(monkeypatch):
st = asyncio.run(rt.get_training_status(current_subject = "t"))
assert st.is_training_running is False
assert rt._run_finished(_Minimal()) is False
+
+
+def test_status_retains_resolved_repo_through_preparation_and_clears_on_next_load(monkeypatch):
+ b = _running(monkeypatch)
+ b._handle_event({"type": "model_load_started"})
+ b._handle_event({"type": "model_load_resolved", "repo_id": "org/resolved"})
+ st = asyncio.run(rt.get_training_status(current_subject = "t"))
+ assert st.details["model_download_repo_id"] == "org/resolved"
+ b._handle_event({"type": "model_load_completed"})
+ st = asyncio.run(rt.get_training_status(current_subject = "t"))
+ assert st.details["model_download_repo_id"] == "org/resolved"
+ b._handle_event({"type": "model_load_started"})
+ assert b._model_download_repo_id is None
+
+
+def test_status_does_not_report_download_after_worker_finishes(monkeypatch):
+ b = _running(monkeypatch)
+ b._handle_event({"type": "model_load_started"})
+ b._handle_event({"type": "model_load_resolved", "repo_id": "org/resolved"})
+ b._handle_event(dict(_DONE))
+ st = asyncio.run(rt.get_training_status(current_subject = "t"))
+ assert st.details["model_download_repo_id"] is None
+
+
+def test_worker_reports_remote_repo_but_not_local_paths(monkeypatch, tmp_path):
+ import ast
+ import os
+ import queue
+ import time
+
+ source = Path(__file__).parents[1] / "core/training/worker.py"
+ tree = ast.parse(source.read_text(encoding = "utf-8"))
+ report = next(
+ n
+ for n in ast.walk(tree)
+ if isinstance(n, ast.FunctionDef) and n.name == "_report_model_repo"
+ )
+ events = queue.SimpleQueue()
+ namespace = {"event_queue": events, "os": os, "time": time}
+ exec(compile(ast.Module(body = [report], type_ignores = []), str(source), "exec"), namespace)
+ b = _running(monkeypatch)
+ b._handle_event({"type": "model_load_started"})
+ namespace["_report_model_repo"]("org/actual-download")
+ b._handle_event(events.get_nowait())
+ status = asyncio.run(rt.get_training_status(current_subject = "t"))
+ assert status.details["model_download_repo_id"] == "org/actual-download"
+ namespace["_report_model_repo"](str(tmp_path))
+ assert events.empty()
diff --git a/studio/frontend/src/features/studio/training-start-overlay.tsx b/studio/frontend/src/features/studio/training-start-overlay.tsx
index 6eede82d803..f1fd049e8a9 100644
--- a/studio/frontend/src/features/studio/training-start-overlay.tsx
+++ b/studio/frontend/src/features/studio/training-start-overlay.tsx
@@ -76,7 +76,9 @@ function useHfDownloadProgress(
): DownloadState {
const phase = useTrainingRuntimeStore((s) => s.phase);
const isStarting = useTrainingRuntimeStore((s) => s.isStarting);
- const [state, setState] = useState<DownloadState>(EMPTY_DOWNLOAD_STATE);
+ const [state, setState] = useState<DownloadState & { repoId?: string }>(
+ EMPTY_DOWNLOAD_STATE,
+ );
const shouldPoll =
isStarting ||
@@ -114,7 +116,7 @@ function useHfDownloadProgress(
applied = generation;
const next = downloadStateFromProgress(prog, latest);
latest = next;
- setState(next);
+ setState({ ...next, repoId });
// only a verified snapshot stops the tick; a settled row can still be waiting on files.
if (next.completeOnDisk) {
finished = true;
@@ -137,7 +139,8 @@ function useHfDownloadProgress(
};
}, [repoId, shouldPoll, fetcher]);
- return state;
+ // Never show the old repo's Ready state while the resolved repo starts polling.
+ return state.repoId === repoId ? state : EMPTY_DOWNLOAD_STATE;
}
function useModelDownloadProgress(
@@ -299,6 +302,9 @@ export function TrainingStartOverlay({
const phase = useTrainingRuntimeStore((s) => s.phase);
const jobId = useTrainingRuntimeStore((s) => s.jobId);
const startModelName = useTrainingRuntimeStore((s) => s.startModelName);
+ const modelDownloadRepoId = useTrainingRuntimeStore(
+ (s) => s.modelDownloadRepoId,
+ );
const startDatasetName = useTrainingRuntimeStore((s) => s.startDatasetName);
const startHfToken = useTrainingRuntimeStore((s) => s.startHfToken);
const startFromResume = useTrainingRuntimeStore((s) => s.startFromResume);
@@ -315,11 +321,13 @@ export function TrainingStartOverlay({
const hfDatasetName = datasetSource === "huggingface" ? dataset : null;
const hasStartResources = startModelName !== null;
const useConfiguredResources = !isStarting && !hasStartResources;
- const modelName = hasStartResources
- ? startModelName
- : useConfiguredResources
- ? configuredModel
- : null;
+ const modelName =
+ modelDownloadRepoId ??
+ (hasStartResources
+ ? startModelName
+ : useConfiguredResources
+ ? configuredModel
+ : null);
const datasetName = hasStartResources
? startDatasetName
: useConfiguredResources
diff --git a/studio/frontend/src/features/training/stores/training-runtime-store.ts b/studio/frontend/src/features/training/stores/training-runtime-store.ts
index 33f38bc6cb9..71230be0d3a 100644
--- a/studio/frontend/src/features/training/stores/training-runtime-store.ts
+++ b/studio/frontend/src/features/training/stores/training-runtime-store.ts
@@ -64,6 +64,7 @@ const initialState: TrainingRuntimeState = {
startRequestId: null,
startError: null,
startModelName: null,
+ modelDownloadRepoId: null,
startDatasetName: null,
startHfToken: null,
startProjectName: null,
@@ -254,7 +255,7 @@ export const useTrainingRuntimeStore = create<TrainingRuntimeStore>()(
return state;
}
acquired = true;
- return { isStarting: true, startRequestId };
+ return { isStarting: true, startRequestId, modelDownloadRepoId: null };
});
return acquired;
},
@@ -309,6 +310,7 @@ export const useTrainingRuntimeStore = create<TrainingRuntimeStore>()(
warnings: [],
startError: null,
phase: "configuring",
+ modelDownloadRepoId: null,
isStarting: false,
startRequestId,
sseConnected: false,
@@ -461,6 +463,7 @@ export const useTrainingRuntimeStore = create<TrainingRuntimeStore>()(
canApplyDetailMetrics && detailEpoch !== null
? Math.max(detailEpoch, runtimeState.currentEpoch)
: runtimeState.currentEpoch,
+ modelDownloadRepoId: payload.details?.model_download_repo_id ?? null,
outputDir:
payload.details?.output_dir !== undefined
? payload.details.output_dir
@@ -623,6 +626,7 @@ if (typeof window !== "undefined") {
useTrainingRuntimeStore.setState({
startHfToken: null,
startModelName: null,
+ modelDownloadRepoId: null,
startDatasetName: null,
startProjectName: null,
});
diff --git a/studio/frontend/src/features/training/types/runtime.ts b/studio/frontend/src/features/training/types/runtime.ts
index 9e96f7ffd9a..35a8c62f51e 100644
--- a/studio/frontend/src/features/training/types/runtime.ts
+++ b/studio/frontend/src/features/training/types/runtime.ts
@@ -33,6 +33,7 @@ export interface TrainingStatusResponse {
learning_rate?: number;
// null = explicit clear (run stopped without saving); absent = unchanged.
output_dir?: string | null;
+ model_download_repo_id?: string | null;
} | null;
metric_history?: {
steps?: number[];
@@ -92,6 +93,7 @@ export interface TrainingRuntimeState {
startRequestId: string | null;
startError: string | null;
startModelName: string | null;
+ modelDownloadRepoId: string | null;
startDatasetName: string | null;
startHfToken: string | null;
startProjectName: string | null;
diff --git a/studio/frontend/tests/training-runtime-store.test.ts b/studio/frontend/tests/training-runtime-store.test.ts
index 7086f32a911..c5d9676d6ba 100644
--- a/studio/frontend/tests/training-runtime-store.test.ts
+++ b/studio/frontend/tests/training-runtime-store.test.ts
@@ -630,3 +630,31 @@ test("metrics are scoped to the current run", () => {
assert.equal(current.currentStep, 8);
assert.deepEqual(current.lossHistory, [{ step: 8, value: 2 }]);
});
+
+test("resolved download repo follows status and clears between loads and jobs", () => {
+ const runtime = useTrainingRuntimeStore.getState();
+ runtime.resetRuntime();
+ runtime.setStartResources("org/requested", null);
+ const status = {
+ job_id: "resolved-job",
+ phase: "loading_model" as const,
+ is_training_running: true,
+ eval_enabled: false,
+ message: "Loading model...",
+ error: null,
+ details: { model_download_repo_id: "org/resolved-unsloth-bnb-4bit" },
+ };
+ runtime.applyStatus(status);
+ assert.equal(
+ useTrainingRuntimeStore.getState().modelDownloadRepoId,
+ "org/resolved-unsloth-bnb-4bit",
+ );
+ runtime.applyStatus({ ...status, details: { model_download_repo_id: null } });
+ assert.equal(useTrainingRuntimeStore.getState().modelDownloadRepoId, null);
+ runtime.applyStatus(status);
+ runtime.applyStatus({ ...status, job_id: "next-job", details: null });
+ assert.equal(useTrainingRuntimeStore.getState().modelDownloadRepoId, null);
+ runtime.applyStatus(status);
+ runtime.resetRuntime();
+ assert.equal(useTrainingRuntimeStore.getState().modelDownloadRepoId, null);
+});
diff --git a/tests/python/test_loader_resolution_reporting.py b/tests/python/test_loader_resolution_reporting.py
new file mode 100644
index 00000000000..025809d6ef7
--- /dev/null
+++ b/tests/python/test_loader_resolution_reporting.py
@@ -0,0 +1,143 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+"""Exercise the loader's actual resolution blocks without loading GPU weights."""
+
+import ast
+import os
+from pathlib import Path
+from types import SimpleNamespace
+
+import pytest
+
+ROOT = Path(__file__).parents[2]
+LOADER = ROOT / "unsloth/models/loader.py"
+
+
+def _loader(cls):
+ tree = ast.parse(LOADER.read_text(encoding = "utf-8"))
+ node = next(n for n in tree.body if isinstance(n, ast.ClassDef) and n.name == cls)
+ return next(
+ n for n in node.body if isinstance(n, ast.FunctionDef) and n.name == "from_pretrained"
+ )
+
+
[email protected]("cls", ["FastLanguageModel", "FastModel"])
[email protected]("exact,prequant", [(False, True), (True, True), (False, False)])
+def test_reports_actual_repo_after_mapping_and_capability_normalization(cls, exact, prequant):
+ function = _loader(cls)
+ start = next(
+ i for i, n in enumerate(function.body) if ast.unparse(n) == "old_model_name = model_name"
+ )
+ # Stop before provider downloads/config probes. This executes the production
+ # mapping and reporting statements, not a second implementation of the mapper.
+ end = next(
+ i
+ for i in range(start, len(function.body))
+ if "USE_MODELSCOPE" in ast.unparse(function.body[i])
+ )
+ reports = []
+ mapping_calls = []
+
+ def mapper(name, **kwargs):
+ mapping_calls.append(name)
+ return "org/resolved-unsloth-bnb-4bit"
+
+ scope = dict(
+ model_name = "org/requested",
+ use_exact_model_name = exact,
+ load_in_4bit = True,
+ load_in_8bit = False,
+ load_in_16bit = False,
+ load_in_fp8 = False,
+ token = None,
+ trust_remote_code = False,
+ get_model_name = mapper,
+ ALLOW_PREQUANTIZED_MODELS = prequant,
+ _strip_unsloth_bnb_4bit_suffix = lambda name: name.removesuffix("-unsloth-bnb-4bit"),
+ on_model_resolved = reports.append,
+ os = os,
+ kwargs = {},
+ quantization_config = None,
+ )
+ module = ast.Module(body = function.body[start:end], type_ignores = [])
+ exec(compile(ast.fix_missing_locations(module), str(LOADER), "exec"), scope)
+ expected = (
+ "org/requested"
+ if exact
+ else "org/resolved-unsloth-bnb-4bit"
+ if prequant
+ else "org/resolved"
+ )
+ assert reports == [expected]
+ assert scope["model_name"] == expected
+ assert mapping_calls == ([] if exact else ["org/requested"])
+
+
[email protected]("cls", ["FastLanguageModel", "FastModel"])
+def test_adapter_base_resolution_updates_report_before_base_config_load(cls):
+ function = _loader(cls)
+ block = next(
+ n for n in function.body if isinstance(n, ast.If) and ast.unparse(n.test) == "is_peft"
+ )
+ # The first three statements select, map and normalize the adapter's base.
+ # Include the notification that must follow, stopping before precision setup.
+ end = next(
+ i
+ for i, n in enumerate(block.body)
+ if isinstance(n, ast.If) and ast.unparse(n.test).startswith("model_name.lower().endswith")
+ )
+ reports = []
+ scope = dict(
+ peft_config = SimpleNamespace(base_model_name_or_path = "org/base"),
+ use_exact_model_name = False,
+ load_in_4bit = True,
+ load_in_fp8 = False,
+ token = None,
+ trust_remote_code = False,
+ ALLOW_PREQUANTIZED_MODELS = True,
+ get_model_name = lambda *a, **k: "org/base-unsloth-bnb-4bit",
+ on_model_resolved = reports.append,
+ )
+ module = ast.Module(body = block.body[:end], type_ignores = [])
+ exec(compile(ast.fix_missing_locations(module), str(LOADER), "exec"), scope)
+ assert reports == ["org/base-unsloth-bnb-4bit"]
+
+
+def test_callback_is_consumed_by_loaders_and_forwarded_only_to_delegated_loader():
+ for cls in ("FastLanguageModel", "FastModel"):
+ function = _loader(cls)
+ assert "on_model_resolved" in [a.arg for a in function.args.args + function.args.kwonlyargs]
+ for call in (n for n in ast.walk(function) if isinstance(n, ast.Call)):
+ if ast.unparse(call.func) == "FastModel.from_pretrained":
+ assert any(k.arg == "on_model_resolved" for k in call.keywords)
+ elif ast.unparse(call.func) in (
+ "dispatch_model.from_pretrained",
+ "FastBaseModel.from_pretrained",
+ ):
+ assert not any(k.arg == "on_model_resolved" for k in call.keywords)
+
+
+def test_training_forwards_observer_through_each_loader_and_retry():
+ path = ROOT / "studio/backend/core/training/trainer.py"
+ tree = ast.parse(path.read_text(encoding = "utf-8"))
+ load = next(
+ n for n in ast.walk(tree) if isinstance(n, ast.FunctionDef) and n.name == "load_model"
+ )
+ calls = [
+ n
+ for n in ast.walk(load)
+ if isinstance(n, ast.Call)
+ and ast.unparse(n.func)
+ in (
+ "FastLanguageModel.from_pretrained",
+ "FastModel.from_pretrained",
+ "FastVisionModel.from_pretrained",
+ "self.load_model",
+ )
+ ]
+ assert len(calls) == 9
+ for call in calls:
+ assert any(
+ k.arg == "on_model_resolved" and ast.unparse(k.value) == "on_model_resolved"
+ for k in call.keywords
+ )
diff --git a/tests/studio/test_hf_token_validation_tick_contract.py b/tests/studio/test_hf_token_validation_tick_contract.py
index af98b119fff..022de483c48 100644
--- a/tests/studio/test_hf_token_validation_tick_contract.py
+++ b/tests/studio/test_hf_token_validation_tick_contract.py
@@ -177,7 +177,7 @@ def test_training_start_claims_runtime_before_first_await():
"setStarting:", 1
)[0]
assert "!startRequestId || isTrainingStartPending(state)" in claim
- assert "return { isStarting: true, startRequestId }" in claim
+ assert "return { isStarting: true, startRequestId," in claim
assert "runtime.tryBeginStarting(startRequestId)" in start_runtime
acquire = start_runtime.split("export function tryAcquireTrainingStart", 1)[1].split(
"export function isTrainingStartLeaseActive", 1
diff --git a/unsloth/models/loader.py b/unsloth/models/loader.py
index 8d1a57ed17c..47eae3cdcc9 100644
--- a/unsloth/models/loader.py
+++ b/unsloth/models/loader.py
@@ -408,6 +408,7 @@ class FastLanguageModel(FastLlamaModel):
unsloth_tiled_mlp = False,
text_only = False,
*args,
+ on_model_resolved = None,
**kwargs,
):
quantization_config = kwargs.get("quantization_config", None)
@@ -453,6 +454,7 @@ class FastLanguageModel(FastLlamaModel):
# @_offline_aware_load already forced offline when needed; delegations inherit it.
if load_in_8bit or full_finetuning or qat_scheme is not None:
delegated, tokenizer = FastModel.from_pretrained(
+ on_model_resolved = on_model_resolved,
model_name = model_name,
max_seq_length = max_seq_length,
dtype = dtype,
@@ -582,6 +584,9 @@ class FastLanguageModel(FastLlamaModel):
("-unsloth-bnb-4bit", "-bnb-4bit")
):
model_name = _strip_unsloth_bnb_4bit_suffix(model_name)
+ # Report the loader decision before fetching this repo, including adapter bases.
+ if on_model_resolved is not None:
+ on_model_resolved(model_name)
# '-bf16' hub repos load bf16; a local dir keeps the requested quant unless 16bit is set. Say so: dropping the flags silently resurfaces as an OOM whose message never mentions quantization.
if model_name.lower().endswith("-bf16") and (
load_in_16bit or not os.path.isdir(os.path.expanduser(model_name))
@@ -768,6 +773,9 @@ class FastLanguageModel(FastLlamaModel):
("-unsloth-bnb-4bit", "-bnb-4bit")
):
model_name = _strip_unsloth_bnb_4bit_suffix(model_name)
+ # Report the loader decision before fetching this repo, including adapter bases.
+ if on_model_resolved is not None:
+ on_model_resolved(model_name)
# '-bf16' hub repos load bf16; a local dir keeps the requested quant unless 16bit is set. Say so: dropping the flags silently resurfaces as an OOM that never mentions quantization.
if model_name.lower().endswith("-bf16") and (
load_in_16bit or not os.path.isdir(os.path.expanduser(model_name))
@@ -879,6 +887,7 @@ class FastLanguageModel(FastLlamaModel):
# Optimized Cohere and Granite paths are disabled until their errors match.
else:
delegated, tokenizer = FastModel.from_pretrained(
+ on_model_resolved = on_model_resolved,
model_name = old_model_name,
max_seq_length = max_seq_length,
dtype = dtype,
@@ -1189,6 +1198,7 @@ class FastModel(FastBaseModel):
target_parameters = None, # For MoE expert parameters
text_only = False,
*args,
+ on_model_resolved = None,
**kwargs,
):
user_config = kwargs.pop("config", None)
@@ -1357,6 +1367,9 @@ class FastModel(FastBaseModel):
("-unsloth-bnb-4bit", "-bnb-4bit")
):
model_name = _strip_unsloth_bnb_4bit_suffix(model_name)
+ # Report the loader decision before fetching this repo, including adapter bases.
+ if on_model_resolved is not None:
+ on_model_resolved(model_name)
# '-bf16' hub repos load bf16; a local dir keeps the requested quant unless 16bit is set. Say so: dropping the flags silently resurfaces as an OOM that never mentions quantization.
if model_name.lower().endswith("-bf16") and (
load_in_16bit or not os.path.isdir(os.path.expanduser(model_name))
@@ -1693,6 +1706,9 @@ class FastModel(FastBaseModel):
("-unsloth-bnb-4bit", "-bnb-4bit")
):
model_name = _strip_unsloth_bnb_4bit_suffix(model_name)
+ # Report the loader decision before fetching this repo, including adapter bases.
+ if on_model_resolved is not None:
+ on_model_resolved(model_name)
# '-bf16' hub repos load bf16; a local dir keeps the requested quant unless 16bit is set. Say so: dropping the flags silently resurfaces as an OOM that never mentions quantization.
if model_name.lower().endswith("-bf16") and (
load_in_16bit or not os.path.isdir(os.path.expanduser(model_name))