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.

Record the repo id for ModelScope loads so merged_16bit and GGUF export work (#13150)

Daniel Han554eb6bde15c
diff --git a/studio/backend/core/inference/safetensors_agentic.py b/studio/backend/core/inference/safetensors_agentic.py
index d1710b4eb46..b4a351c0f53 100644
--- a/studio/backend/core/inference/safetensors_agentic.py
+++ b/studio/backend/core/inference/safetensors_agentic.py
@@ -1492,10 +1492,7 @@ def run_safetensors_tool_loop(
image_share = None
if decision.should_execute and mcp_image is not None:
from core.inference.tools import mcp_image_share
-
- image_share = mcp_image_share(
- decision.tool_name, decision.arguments, mcp_image
- )
+ image_share = mcp_image_share(decision.tool_name, decision.arguments, mcp_image)
if image_share is not None:
decision = tool_controller.reprepare_call(
decision, provisional = provisional_match
diff --git a/studio/backend/core/inference/studio_tool_loop.py b/studio/backend/core/inference/studio_tool_loop.py
index 9177d4849fe..7eb620751b0 100644
--- a/studio/backend/core/inference/studio_tool_loop.py
+++ b/studio/backend/core/inference/studio_tool_loop.py
@@ -1871,7 +1871,11 @@ async def stream_with_studio_tools(
# OpenAI, Anthropic and Gemini all reject that history instead of answering.
exhausted_call = signed_provider_call or decision.as_assistant_tool_call()
exhausted_extra = call.get("extra_content")
- if signed_provider_call is None and isinstance(exhausted_extra, dict) and exhausted_extra:
+ if (
+ signed_provider_call is None
+ and isinstance(exhausted_extra, dict)
+ and exhausted_extra
+ ):
exhausted_call["extra_content"] = exhausted_extra
assistant_tool_calls.append(exhausted_call)
tool_messages.append(
diff --git a/studio/backend/tests/test_external_import.py b/studio/backend/tests/test_external_import.py
index 8955d35ab17..020a27a8825 100644
--- a/studio/backend/tests/test_external_import.py
+++ b/studio/backend/tests/test_external_import.py
@@ -338,7 +338,15 @@ def test_claude_rewind_compaction_uses_its_logical_parent(claude_home):
)
messages = claude.read_transcript(path, "t", "s1").messages
ids = [m["id"] for m in messages]
- assert [m["parentId"] for m in messages] == [None, ids[0], ids[1], ids[2], ids[1], ids[4], ids[5]]
+ assert [m["parentId"] for m in messages] == [
+ None,
+ ids[0],
+ ids[1],
+ ids[2],
+ ids[1],
+ ids[4],
+ ids[5],
+ ]
def test_claude_releases_filtered_records_while_streaming(tmp_path, monkeypatch):
diff --git a/studio/backend/tests/test_studio_tool_loop.py b/studio/backend/tests/test_studio_tool_loop.py
index da73b29d453..1b3d17c4a9f 100644
--- a/studio/backend/tests/test_studio_tool_loop.py
+++ b/studio/backend/tests/test_studio_tool_loop.py
@@ -725,9 +725,7 @@ def test_sharing_the_attached_image_asks_even_with_bypass(executed, monkeypatch)
assert "mcp_image" not in executed[0]
-def test_settled_image_call_is_replayed_and_deduped_with_sent_arguments(
- executed, monkeypatch
-):
+def test_settled_image_call_is_replayed_and_deduped_with_sent_arguments(executed, monkeypatch):
from core.inference.mcp_image import ATTACHED_IMAGE, McpImage
image = McpImage(mime = "image/png", data = b"IMG")
@@ -742,9 +740,7 @@ def test_settled_image_call_is_replayed_and_deduped_with_sent_arguments(
monkeypatch.setattr(loop_mod, "mcp_image_share", settle)
monkeypatch.setattr(loop_mod, "begin_tool_decision", lambda *_args: object())
- monkeypatch.setattr(
- loop_mod, "wait_tool_decision", lambda *_args, **_kwargs: "allow"
- )
+ monkeypatch.setattr(loop_mod, "wait_tool_decision", lambda *_args, **_kwargs: "allow")
monkeypatch.setattr(loop_mod, "abort_tool_decision", lambda *_args: None)
call = {"index": 0, "function": {"name": "python", "arguments": "{}"}}
transport = FakeTransport(
@@ -782,9 +778,7 @@ def test_settled_image_call_keeps_signed_provider_replay_exact(executed, monkeyp
monkeypatch.setattr(loop_mod, "mcp_image_share", settle)
monkeypatch.setattr(loop_mod, "begin_tool_decision", lambda *_args: object())
- monkeypatch.setattr(
- loop_mod, "wait_tool_decision", lambda *_args, **_kwargs: "allow"
- )
+ monkeypatch.setattr(loop_mod, "wait_tool_decision", lambda *_args, **_kwargs: "allow")
monkeypatch.setattr(loop_mod, "abort_tool_decision", lambda *_args: None)
call = {
"index": 0,
@@ -808,9 +802,7 @@ def test_settled_image_call_keeps_signed_provider_replay_exact(executed, monkeyp
if message.get("role") == "assistant" and message.get("tool_calls")
)["tool_calls"][0]
assert replayed["function"]["arguments"] == "{ }"
- assert replayed["extra_content"] == {
- "google": {"thought_signature": "SIG-A"}
- }
+ assert replayed["extra_content"] == {"google": {"thought_signature": "SIG-A"}}
def test_the_model_is_told_about_the_attached_image(executed, monkeypatch):
diff --git a/studio/backend/tests/test_tool_loop_controller.py b/studio/backend/tests/test_tool_loop_controller.py
index 953db28b731..8b4ec86abf1 100644
--- a/studio/backend/tests/test_tool_loop_controller.py
+++ b/studio/backend/tests/test_tool_loop_controller.py
@@ -200,9 +200,7 @@ def test_reprepare_uses_normalized_arguments_for_replay_and_dedup():
first = controller.reprepare_call(first)
assert first.card_id == "painted_a"
- assert first.as_assistant_tool_call()["function"]["arguments"] == (
- '{"image":"attached_image"}'
- )
+ assert first.as_assistant_tool_call()["function"]["arguments"] == ('{"image":"attached_image"}')
assert first.key == canonical_tool_call_key(tool_name, {"image": "attached_image"})
controller.record_result(first, "ok")
diff --git a/tests/saving/test_modelscope_repo_id.py b/tests/saving/test_modelscope_repo_id.py
new file mode 100644
index 00000000000..413f8355b5e
--- /dev/null
+++ b/tests/saving/test_modelscope_repo_id.py
@@ -0,0 +1,146 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+
+"""#3726: a ModelScope load must record the repo id, not the local snapshot, and the
+16bit merge must fetch its base from ModelScope.
+
+unsloth cannot be imported on GPU-less hosts, so the helpers are extracted via ast and
+exec'd against fakes, like the other GPU-free saving tests."""
+
+from __future__ import annotations
+
+import ast
+import os
+import sys
+import types
+from pathlib import Path
+
+import pytest
+
+_ROOT = Path(__file__).resolve().parent.parent.parent
+
+
+def _load(relpath, name, namespace):
+ source = (_ROOT / relpath).read_text(encoding = "utf-8")
+ for node in ast.walk(ast.parse(source)):
+ if isinstance(node, ast.FunctionDef) and node.name == name:
+ exec(compile(ast.get_source_segment(source, node), relpath, "exec"), namespace)
+ return namespace[name]
+ raise AssertionError(f"{name} not found in {relpath}")
+
+
+def _tiny_model(name_or_path):
+ from transformers import LlamaConfig, LlamaForCausalLM
+
+ config = LlamaConfig(
+ vocab_size = 32,
+ hidden_size = 16,
+ intermediate_size = 32,
+ num_hidden_layers = 1,
+ num_attention_heads = 2,
+ num_key_value_heads = 2,
+ )
+ config._name_or_path = name_or_path
+ model = LlamaForCausalLM(config)
+ model.name_or_path = name_or_path
+ return model
+
+
+def record():
+ return _load("unsloth/models/loader.py", "_record_modelscope_repo_id", {"os": os})
+
+
+def test_snapshot_path_replaced_and_peft_records_repo_id(record, tmp_path):
+ from peft import LoraConfig, get_peft_model
+
+ snapshot = tmp_path / "models" / "unsloth--Qwen3-0.6B-unsloth-bnb-4bit" / "snapshots" / "master"
+ snapshot.mkdir(parents = True)
+ model = _tiny_model(str(snapshot))
+ record(model, "unsloth/Qwen3-0.6B-unsloth-bnb-4bit", str(snapshot))
+
+ assert model.config._name_or_path == "unsloth/Qwen3-0.6B-unsloth-bnb-4bit"
+ assert model.name_or_path == "unsloth/Qwen3-0.6B-unsloth-bnb-4bit"
+ assert model._unsloth_modelscope_snapshot == str(snapshot)
+ peft_model = get_peft_model(model, LoraConfig(r = 2, target_modules = ["q_proj"]))
+ assert (
+ peft_model.peft_config["default"].base_model_name_or_path
+ == "unsloth/Qwen3-0.6B-unsloth-bnb-4bit"
+ )
+
+
[email protected]("repo_id, local_dir", [(None, None), ("unsloth/x", None), (None, "/a")])
+def test_noop_without_modelscope_load(record, repo_id, local_dir):
+ model = _tiny_model("unsloth/x-bnb-4bit")
+ record(model, repo_id, local_dir)
+ assert model.config._name_or_path == "unsloth/x-bnb-4bit"
+
+
+def test_other_names_untouched(record, tmp_path):
+ model = _tiny_model(str(tmp_path / "elsewhere"))
+ record(model, "unsloth/x", str(tmp_path / "snapshot"))
+ assert model.config._name_or_path == str(tmp_path / "elsewhere")
+
+
+class _Logger:
+ def __init__(self):
+ self.messages = []
+
+ def warning_once(self, message):
+ self.messages.append(message)
+
+
+def merge_name(monkeypatch, tmp_path):
+ calls = []
+
+ def snapshot_download(name):
+ calls.append(name)
+ if name == "unsloth/missing":
+ raise ValueError("not on ModelScope")
+ return str(tmp_path / name.replace("/", "--"))
+
+ monkeypatch.setitem(
+ sys.modules, "modelscope", types.SimpleNamespace(snapshot_download = snapshot_download)
+ )
+ logger = _Logger()
+ fn = _load(
+ "unsloth/save.py",
+ "_modelscope_base_model_name",
+ {
+ "os": os,
+ "logger": logger,
+ "get_model_name": lambda name, load_in_4bit: name.removesuffix("-unsloth-bnb-4bit"),
+ },
+ )
+ return fn, calls, logger
+
+
+def test_modelscope_merge_fetches_16bit_base(merge_name, tmp_path):
+ fn, calls, _ = merge_name
+ got = fn("unsloth/Qwen3-14B-unsloth-bnb-4bit", load_in_4bit = False)
+ assert got == str(tmp_path / "unsloth--Qwen3-14B")
+ assert calls == ["unsloth/Qwen3-14B"]
+
+
+def test_modelscope_merge_keeps_local_dir_and_falls_back(merge_name, tmp_path):
+ fn, calls, logger = merge_name
+ assert fn(str(tmp_path), load_in_4bit = False) == str(tmp_path)
+ assert fn("unsloth/missing", load_in_4bit = False) == "unsloth/missing"
+ assert calls == ["unsloth/missing"]
+ assert "trying Hugging Face" in logger.messages[0]
+
+
+def test_gguf_still_converts_the_modelscope_snapshot(tmp_path):
+ class _Peft:
+ pass
+
+ reuses = _load(
+ "unsloth/save.py",
+ "_gguf_reuses_loaded_checkpoint",
+ {"os": os, "PeftModel": _Peft, "PeftModelForCausalLM": _Peft},
+ )
+ model = types.SimpleNamespace(config = types.SimpleNamespace(_name_or_path = "unsloth/x"))
+ assert not reuses(model)
+ model._unsloth_modelscope_snapshot = str(tmp_path)
+ assert reuses(model)
diff --git a/tests/saving/test_prewarm_base_model_hub_cache.py b/tests/saving/test_prewarm_base_model_hub_cache.py
index 98ee8a54a53..d915d082b28 100644
--- a/tests/saving/test_prewarm_base_model_hub_cache.py
+++ b/tests/saving/test_prewarm_base_model_hub_cache.py
@@ -242,6 +242,15 @@ def test_kaggle_and_colab_skip(monkeypatch, tmp_path):
assert stubs.snapshot_download.calls == [], f"{flag} must skip pre-warm"
+def test_modelscope_skips(monkeypatch, tmp_path):
+ # The merge fetches the base from ModelScope (#3726), so a Hub copy would go unused.
+ monkeypatch.setenv("UNSLOTH_USE_MODELSCOPE", "1")
+ fn, stubs = _build_env(monkeypatch, tmp_path)
+ fn(_FakePeftModel(), save_method = "merged_16bit")
+ assert stubs.snapshot_download.calls == []
+ assert stubs.determine_base_model_source.calls == []
+
+
@pytest.mark.parametrize("save_method", ["merged_4bit", "forced_merged_4bit", "lora"])
def test_non_downloading_save_methods_skip(monkeypatch, tmp_path, save_method):
fn, stubs = _build_env(monkeypatch, tmp_path)
diff --git a/tests/saving/test_save_method_lora_routing.py b/tests/saving/test_save_method_lora_routing.py
index da1aadfb79d..db6af3da346 100644
--- a/tests/saving/test_save_method_lora_routing.py
+++ b/tests/saving/test_save_method_lora_routing.py
@@ -28,6 +28,7 @@ from __future__ import annotations
import ast
import contextlib
+import os
import sys
import types
from pathlib import Path
@@ -254,6 +255,7 @@ def _routing_environment(monkeypatch, model):
namespace = _load(
*_GENERIC_SAVE_REAL_HELPERS,
"unsloth_generic_save",
+ os = os,
PeftModel = _PeftModel,
PreTrainedTokenizerBase = type("Tokenizer", (), {}),
ProcessorMixin = type("Processor", (), {}),
@@ -360,6 +362,27 @@ def test_every_other_method_still_merges(monkeypatch, tmp_path, save_method):
assert len(calls["convert"]) == (save_method != "merged_4bit_forced")
+ "save_method, from_modelscope", [("merged_16bit", True), ("merged_4bit_forced", False)]
+)
+def test_modelscope_base_is_fetched_only_when_the_merge_reads_one(
+ monkeypatch, tmp_path, save_method, from_modelscope
+):
+ """#3726: an in-place 4bit merge folds into the loaded weights, so it downloads no 16bit base."""
+ monkeypatch.setenv("UNSLOTH_USE_MODELSCOPE", "1")
+ model = _PeftModel()
+ generic_save, calls = _routing_environment(monkeypatch, model)
+ resolvers = []
+ sys.modules["unsloth_zoo.saving_utils"].merge_and_overwrite_lora = (
+ lambda resolve, **kwargs: resolvers.append(resolve)
+ )
+ namespace = generic_save.__globals__
+ namespace["_modelscope_base_model_name"] = lambda *args, **kwargs: None
+ generic_save(model, None, save_directory = str(tmp_path), save_method = save_method)
+ expected = "_modelscope_base_model_name" if from_modelscope else "get_model_name"
+ assert resolvers == [namespace[expected]]
+
+
def test_a_model_with_no_adapter_is_unchanged(monkeypatch, tmp_path):
"""A full fine-tune asked for "lora" has no adapter, so it writes itself, as before."""
model = _FullModel()
diff --git a/tests/test_optimized_precision_conflicts.py b/tests/test_optimized_precision_conflicts.py
index 77518e69ae2..d8a3f9c946e 100644
--- a/tests/test_optimized_precision_conflicts.py
+++ b/tests/test_optimized_precision_conflicts.py
@@ -37,11 +37,12 @@ def loader():
n for n in cls.body if isinstance(n, ast.FunctionDef) and n.name == "from_pretrained"
)
method.decorator_list = []
- helper = next(
+ helpers = [
n
for n in tree.body
- if isinstance(n, ast.FunctionDef) and n.name == "_precision_flags_conflict"
- )
+ if isinstance(n, ast.FunctionDef)
+ and n.name in ("_precision_flags_conflict", "_modelscope_snapshot_or_none")
+ ]
captured = {"config_calls": 0}
def dispatch(**kwargs):
@@ -70,6 +71,7 @@ def loader():
ALLOW_BITSANDBYTES = True,
ALLOW_PREQUANTIZED_MODELS = True,
USE_MODELSCOPE = False,
+ logger = SimpleNamespace(warning_once = lambda *args, **kwargs: None),
SUPPORTS_LLAMA32 = True,
get_model_name = lambda name, **kwargs: name,
_revision_for_resolved_repo = lambda revision, *args: revision,
@@ -89,7 +91,7 @@ def loader():
# is an exception nothing raises. Read off loader.py so a new helper is covered too.
for name in _mistral_format_names(tree):
env.setdefault(name, RuntimeError if name[0].isupper() else (lambda *args, **kwargs: False))
- exec(compile(ast.Module(body = [helper, method], type_ignores = []), str(path), "exec"), env)
+ exec(compile(ast.Module(body = [*helpers, method], type_ignores = []), str(path), "exec"), env)
return env, captured
@@ -162,6 +164,10 @@ def modelscope_snapshot(monkeypatch, tmp_path):
def snapshot_download(name, allow_file_pattern = None):
calls.append((name, allow_file_pattern))
+ if name.startswith("hf-only/"):
+ raise ValueError(f"{name} is not on ModelScope")
+ snapshot = cache / name.replace("/", "--")
+ snapshot.mkdir(exist_ok = True)
for filename in (
"config.json",
"adapter_config.json",
@@ -171,9 +177,9 @@ def modelscope_snapshot(monkeypatch, tmp_path):
if allow_file_pattern is None or any(
fnmatch.fnmatch(filename, pattern) for pattern in allow_file_pattern
):
- (cache / filename).write_text("test fixture")
+ (snapshot / filename).write_text("test fixture")
downloaded.append(filename)
- return str(cache)
+ return str(snapshot)
monkeypatch.setitem(
sys.modules, "modelscope", SimpleNamespace(snapshot_download = snapshot_download)
@@ -183,7 +189,7 @@ def modelscope_snapshot(monkeypatch, tmp_path):
def use_modelscope_adapter(env, cache, base_name):
def config(name, **kwargs):
- if name == str(cache):
+ if name == str(cache / "owner--model"):
raise ValueError("Adapter has no model config")
return SimpleNamespace(model_type = "llama", rope_scaling = None)
@@ -243,6 +249,23 @@ def test_modelscope_valid_loads_download_weights(
env["from_pretrained"](model_name = "owner/model", **kwargs)
assert "model.safetensors" in downloaded
- assert captured["dispatch"]["model_name"] == (adapter_base or str(cache))
+ # An adapter's base is fetched from ModelScope as well (#3726).
+ loaded = adapter_base or "owner/model"
+ assert captured["dispatch"]["model_name"] == str(cache / loaded.replace("/", "--"))
assert captured["dispatch"]["load_in_4bit"] is expected_4bit
assert calls[-1] == ("owner/model", None)
+ assert (loaded, None) in calls
+
+
+def test_modelscope_adapter_base_missing_there_loads_from_the_hub(loader, modelscope_snapshot):
+ """#3726: an adapter on ModelScope whose base is only on the Hub still loads that base by repo id."""
+ env, captured = loader
+ cache, downloaded, calls = modelscope_snapshot
+ env["USE_MODELSCOPE"] = True
+ use_modelscope_adapter(env, cache, "hf-only/base")
+
+ with pytest.raises(DispatchReached):
+ env["from_pretrained"](model_name = "owner/model", load_in_4bit = False, load_in_16bit = True)
+
+ assert captured["dispatch"]["model_name"] == "hf-only/base"
+ assert ("hf-only/base", None) in calls
diff --git a/tests/version_compat/test_trl_loss_normalization_contract.py b/tests/version_compat/test_trl_loss_normalization_contract.py
index 86253fbcbb5..bc21b95f0c6 100644
--- a/tests/version_compat/test_trl_loss_normalization_contract.py
+++ b/tests/version_compat/test_trl_loss_normalization_contract.py
@@ -175,7 +175,9 @@ def _trl_resolves(wanted):
warnings.simplefilter("always")
got = pristine(output_dir = "unused", loss_type = wanted).loss_type
aliased = any(
- issubclass(w.category, FutureWarning) and repr(wanted) in str(w.message) and "deprecated" in str(w.message)
+ issubclass(w.category, FutureWarning)
+ and repr(wanted) in str(w.message)
+ and "deprecated" in str(w.message)
for w in caught
)
return got if aliased else wanted
@@ -250,7 +252,9 @@ def test_pristine_trl_sft_config_keeps_an_explicit_loss_type():
for wanted in ("chunked_nll", "dft"):
got = pristine(output_dir = "unused", loss_type = wanted).loss_type
- assert got == _trl_resolves(wanted), f"explicit loss_type {wanted!r} was clobbered to {got!r}"
+ assert got == _trl_resolves(
+ wanted
+ ), f"explicit loss_type {wanted!r} was clobbered to {got!r}"
# "dft" is not an alias on any TRL, so it must always come back unchanged.
assert _trl_resolves("dft") == "dft"
diff --git a/unsloth/models/loader.py b/unsloth/models/loader.py
index 58a695c7ace..aed68d94888 100644
--- a/unsloth/models/loader.py
+++ b/unsloth/models/loader.py
@@ -196,6 +196,36 @@ def _revision_for_resolved_repo(
return None
+def _record_modelscope_repo_id(model, repo_id, local_dir):
+ """A ModelScope load hands transformers the local snapshot, which becomes the model's name: PEFT copies it into adapter_config.json and the merge reads it as the base, so a 4bit snapshot refuses merged_16bit / GGUF and the adapter names a path on this machine only (#3726). Record the repo id, as a Hub load does."""
+ if repo_id is None or local_dir is None:
+ return
+ from transformers import PreTrainedModel
+
+ resolved = os.path.realpath(str(local_dir))
+ for module in model.modules():
+ if not isinstance(module, PreTrainedModel):
+ continue
+ for owner, attr in ((module, "name_or_path"), (module.config, "_name_or_path")):
+ value = getattr(owner, attr, None)
+ if isinstance(value, str) and value and os.path.realpath(value) == resolved:
+ setattr(owner, attr, repo_id)
+ # A non-PEFT GGUF export still converts the snapshot in place.
+ module._unsloth_modelscope_snapshot = str(local_dir)
+
+
+def _modelscope_snapshot_or_none(repo_id):
+ """ModelScope snapshot of an adapter's base, or None to load it by repo id from the Hugging Face Hub as before: the base need not be mirrored on ModelScope."""
+ from modelscope import snapshot_download
+ try:
+ return snapshot_download(repo_id)
+ except Exception as e:
+ logger.warning_once(
+ f"Unsloth: Could not download `{repo_id}` from ModelScope ({e}), loading it from Hugging Face."
+ )
+ return None
+
+
def _revision_for_tokenizer_repo(
tokenizer_name,
model_name,
@@ -1176,8 +1206,11 @@ class FastLanguageModel(FastLlamaModel):
# Only check the flags when no non-bitsandbytes quantization_config sets the precision.
check_precision_flags = quantization_config is None or q_load_in_4bit or q_load_in_8bit
modelscope_pending_download = None
+ modelscope_repo_id = modelscope_dir = None
if USE_MODELSCOPE and not os.path.exists(model_name):
from modelscope import snapshot_download
+
+ modelscope_repo_id = model_name
if check_precision_flags and _precision_flags_conflict(
load_in_4bit, load_in_8bit, load_in_16bit, load_in_fp8
):
@@ -1186,6 +1219,7 @@ class FastLanguageModel(FastLlamaModel):
model_name = snapshot_download(model_name, allow_file_pattern = ["*.json", "*.py"])
else:
model_name = snapshot_download(model_name)
+ modelscope_dir = model_name
# Gate before the probe below, or a pinned 4bit load fails against the mirror.
base_revision = _revision_for_resolved_repo(
@@ -1384,6 +1418,21 @@ class FastLanguageModel(FastLlamaModel):
load_in_8bit = False
load_in_fp8 = False
load_in_16bit = True
+ # The adapter names its base by repo id, so fetch that from ModelScope too; a precision conflict raises below before any weights are needed.
+ if (
+ USE_MODELSCOPE
+ and not os.path.exists(model_name)
+ and not (
+ check_precision_flags
+ and _precision_flags_conflict(
+ load_in_4bit, load_in_8bit, load_in_16bit, load_in_fp8
+ )
+ )
+ ):
+ _snapshot = _modelscope_snapshot_or_none(model_name)
+ if _snapshot is not None:
+ modelscope_repo_id, model_name = model_name, _snapshot
+ modelscope_dir = _snapshot
# After the -bf16 rule: the view path no longer carries the source's suffix.
# No revision: the caller's ref names the adapter repo, and the base loads unpinned below.
_cache_dir = kwargs.get("cache_dir", None)
@@ -1590,6 +1639,7 @@ class FastLanguageModel(FastLlamaModel):
*args,
**kwargs,
)
+ _record_modelscope_repo_id(model, modelscope_repo_id, modelscope_dir)
if resize_model_vocab is not None:
_resize_vocab(model, resize_model_vocab)
@@ -2028,9 +2078,11 @@ class FastModel(FastBaseModel):
load_in_fp8 = False
load_in_16bit = True
+ modelscope_repo_id = modelscope_dir = None
if USE_MODELSCOPE and not os.path.exists(model_name):
from modelscope import snapshot_download
- model_name = snapshot_download(model_name)
+ modelscope_repo_id = model_name
+ model_name = modelscope_dir = snapshot_download(model_name)
# Gate before the probe below, or a pinned 4bit load fails against the mirror.
base_revision = _revision_for_resolved_repo(
@@ -2406,6 +2458,12 @@ class FastModel(FastBaseModel):
load_in_8bit = False
load_in_fp8 = False
load_in_16bit = True
+ # The adapter names its base by repo id, so fetch that from ModelScope too.
+ if USE_MODELSCOPE and not os.path.exists(model_name):
+ _snapshot = _modelscope_snapshot_or_none(model_name)
+ if _snapshot is not None:
+ modelscope_repo_id, model_name = model_name, _snapshot
+ modelscope_dir = _snapshot
# After the -bf16 rule: the view path no longer carries the source's suffix.
# No revision: the caller's ref names the adapter repo, and the base loads unpinned below.
_cache_dir = kwargs.get("cache_dir", None)
@@ -2808,6 +2866,7 @@ class FastModel(FastBaseModel):
**kwargs,
)
_drop_text_only_key_mapping(model, _text_key_mapping)
+ _record_modelscope_repo_id(model, modelscope_repo_id, modelscope_dir)
if resize_model_vocab is not None:
_resize_vocab(model, resize_model_vocab)
diff --git a/unsloth/save.py b/unsloth/save.py
index 67a885705df..f83990eff5a 100644
--- a/unsloth/save.py
+++ b/unsloth/save.py
@@ -3348,7 +3348,10 @@ def _gguf_reuses_loaded_checkpoint(model, state_dict = None):
return False
if state_dict is not None or getattr(model, "_unsloth_full_finetuning", False):
return False
- name_or_path = getattr(getattr(model, "config", None), "_name_or_path", None)
+ # A ModelScope load keeps its snapshot here, since `_name_or_path` holds the repo id (#3726).
+ name_or_path = getattr(model, "_unsloth_modelscope_snapshot", None) or getattr(
+ getattr(model, "config", None), "_name_or_path", None
+ )
try:
return bool(name_or_path and os.path.isdir(str(name_or_path)))
except Exception:
@@ -3438,7 +3441,9 @@ def _gguf_model_input_directory(
):
"""The folder the converter reads, which is not always `save_directory`: a reused loaded checkpoint, which `unsloth_save_pretrained_gguf` assigns to `save_directory` before calling `save_to_gguf`. It matters only in the unwritable-CWD fallback, where the intermediate GGUF lands beside the reused checkpoint rather than the requested output, and the two can be on different filesystems."""
if _gguf_reuses_loaded_checkpoint(model, state_dict):
- return str(model.config._name_or_path)
+ return str(
+ getattr(model, "_unsloth_modelscope_snapshot", None) or model.config._name_or_path
+ )
return save_directory
@@ -4094,7 +4099,9 @@ def unsloth_save_pretrained_gguf(
) from e
else:
# Non-PEFT model: convert the loaded checkpoint in place when it still holds the weights to export.
- original_path = getattr(self.config, "_name_or_path", None)
+ original_path = getattr(self, "_unsloth_modelscope_snapshot", None) or getattr(
+ self.config, "_name_or_path", None
+ )
if _gguf_reuses_loaded_checkpoint(self, state_dict):
print(
f"Unsloth: Model is not a PEFT model. Using existing checkpoint at {original_path}"
@@ -5161,6 +5168,21 @@ from unsloth_zoo.llama_cpp import (
)
+def _modelscope_base_model_name(model_name, load_in_4bit = True):
+ """`get_model_name` for a merge under UNSLOTH_USE_MODELSCOPE=1: the 16bit base comes from ModelScope, like the model it was trained on, not the Hugging Face Hub (#3726)."""
+ name = get_model_name(model_name, load_in_4bit = load_in_4bit)
+ if not name or os.path.exists(name):
+ return name
+ try:
+ from modelscope import snapshot_download
+ return snapshot_download(name)
+ except Exception as e:
+ logger.warning_once(
+ f"Unsloth: Could not download `{name}` from ModelScope ({e}), trying Hugging Face."
+ )
+ return name
+
+
def _prewarm_base_model_hub_cache(
model,
save_method = "merged_16bit",
@@ -5172,6 +5194,9 @@ def _prewarm_base_model_hub_cache(
return
if IS_KAGGLE_ENVIRONMENT or IS_COLAB_ENVIRONMENT:
return
+ # The merge fetches the base from ModelScope, so a Hub cache copy would go unused.
+ if os.environ.get("UNSLOTH_USE_MODELSCOPE", "0") == "1":
+ return
_true = ("1", "true", "yes", "on")
if (
os.environ.get("HF_HUB_OFFLINE", "").strip().lower() in _true
@@ -5707,7 +5732,9 @@ def unsloth_generic_save(
in_place = save_method in ("merged_4bit", "forced_merged_4bit")
with nullcontext() if in_place else lora_relative_to_original_base(model):
merge_and_overwrite_lora(
- get_model_name,
+ _modelscope_base_model_name
+ if os.environ.get("UNSLOTH_USE_MODELSCOPE", "0") == "1" and not in_place
+ else get_model_name,
model = model,
tokenizer = tokenizer,
save_directory = save_directory,