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.
Python agentaichatgptdeepseekfine-tuninggemmaimage-generationllamallmllmsopenaipythonqwenreinforcement-learningself-hostedstable-diffusiontext-to-speechttsuiunsloth
Studio: push only the exported merged model to the Hub on a Mac (#12239)
Nilay852795db9db8
diff --git a/studio/backend/core/export/export.py b/studio/backend/core/export/export.pyindex 7e16f3a8f4b..c53090c8315 100644--- a/studio/backend/core/export/export.py+++ b/studio/backend/core/export/export.py@@ -605,6 +605,32 @@ def _open_hub_repo(hf_api, repo_id, private):return repo_id+def _push_mlx_merged(+ model, tokenizer, *, save_method, output_path, output_is_fresh, repo_id, hf_token, private+):+ """Upload an MLX merged save; returns the repo id the Hub resolved."""+ with contextlib.ExitStack() as stack:+ upload_dir = output_path+ if not output_is_fresh:+ # A reused folder can hold leftovers, so upload a clean second save; without a+ # local save this is the only one.+ upload_dir = stack.enter_context(+ _staging_dir(Path(output_path).parent)+ if output_path+ else tempfile.TemporaryDirectory()+ )+ model.save_pretrained_merged(upload_dir, tokenizer, save_method = save_method)+ hf_api = HfApi(token = hf_token)+ repo_id = _open_hub_repo(hf_api, repo_id, private)+ hf_api.upload_folder(+ folder_path = upload_dir,+ repo_id = repo_id,+ repo_type = "model",+ ignore_patterns = _HUB_UPLOAD_IGNORE,+ )+ return repo_id++def _publish_unsloth_model_card(hf_api, repo_id, model, hf_token):"""Write the card the delegated push can no longer write for itself.@@ -1125,28 +1151,16 @@ class ExportBackend:logger.info(f"Pushing merged model to Hub: {repo_id}")if _IS_MLX:- if save_directory:- self.current_model.push_to_hub_merged(- repo_id,- self.current_tokenizer,- save_directory = save_directory,- token = hf_token,- private = private,- )- else:- with tempfile.TemporaryDirectory() as tmp_dir:- self.current_model.save_pretrained_merged(- tmp_dir,- self.current_tokenizer,- save_method = mlx_save_method,- )- self.current_model.push_to_hub_merged(- repo_id,- self.current_tokenizer,- save_directory = tmp_dir,- token = hf_token,- private = private,- )+ repo_id = _push_mlx_merged(+ self.current_model,+ self.current_tokenizer,+ save_method = mlx_save_method,+ output_path = output_path,+ output_is_fresh = save_dir_was_empty,+ repo_id = repo_id,+ hf_token = hf_token,+ private = private,+ )else:uploaded = Falseif output_path and Path(output_path).is_dir():@@ -1281,28 +1295,16 @@ class ExportBackend:logger.info(f"Pushing base model to Hub: {repo_id}")if _IS_MLX:- if save_directory:- self.current_model.push_to_hub_merged(- repo_id,- self.current_tokenizer,- save_directory = save_directory,- token = hf_token,- private = private,- )- else:- with tempfile.TemporaryDirectory() as tmp_dir:- self.current_model.save_pretrained_merged(- tmp_dir,- self.current_tokenizer,- save_method = "merged_16bit",- )- self.current_model.push_to_hub_merged(- repo_id,- self.current_tokenizer,- save_directory = tmp_dir,- token = hf_token,- private = private,- )+ _push_mlx_merged(+ self.current_model,+ self.current_tokenizer,+ save_method = "merged_16bit",+ output_path = output_path,+ output_is_fresh = save_dir_was_empty,+ repo_id = repo_id,+ hf_token = hf_token,+ private = private,+ )else:base_model = (base_model_id or self.current_model.config._name_or_path or "unknown"diff --git a/studio/backend/tests/test_export_gguf_hub_upload.py b/studio/backend/tests/test_export_gguf_hub_upload.pyindex f630b1a86d5..ff819a2b76c 100644--- a/studio/backend/tests/test_export_gguf_hub_upload.py+++ b/studio/backend/tests/test_export_gguf_hub_upload.py@@ -102,6 +102,7 @@ def _hub_doubles(calls, seen):):calls.append("upload_folder")seen["folder"] = folder_path+ seen["upload_repo"] = repo_idroot = Path(folder_path)paths = [p.relative_to(root).as_posix() for p in root.rglob("*") if p.is_file()]if allow_patterns is not None:diff --git a/studio/backend/tests/test_export_hub_push.py b/studio/backend/tests/test_export_hub_push.pyindex 5b8496ccbaa..465b7e52eae 100644--- a/studio/backend/tests/test_export_hub_push.py+++ b/studio/backend/tests/test_export_hub_push.py@@ -3,6 +3,7 @@from __future__ import annotations+import functoolsimport importlib.utilimport sysimport tempfile@@ -375,6 +376,169 @@ def test_merged_export_push_card_does_not_name_a_local_base_model(tmp_path, monkassert str(tmp_path) not in seen["card"]+_MLX_MERGED_FILES = ["README.md", "model.safetensors", "model.safetensors.index.json"]+++class _MlxMergeModel(_Model):+ def __init__(self, hf_api):+ super().__init__()+ self.hf_api = hf_api++ def save_pretrained_merged(+ self,+ save_directory,+ tokenizer,+ save_method = None,+ token = None,+ ):+ self.merges.append(save_method)+ output = Path(save_directory)+ output.mkdir(parents = True, exist_ok = True)+ for name in _MLX_MERGED_FILES:+ (output / name).write_bytes(b"merged")++ def push_to_hub_merged(+ self,+ repo_id,+ tokenizer,+ save_directory = None,+ token = None,+ private = None,+ ):+ api = self.hf_api(token = token)+ api.create_repo(repo_id, private = private, exist_ok = True)+ if private is not None:+ api.update_repo_settings(repo_id = repo_id, private = private, repo_type = "model")+ api.upload_folder(folder_path = str(save_directory), repo_id = repo_id, repo_type = "model")+++# export -> (ExportBackend method, is_peft, extra kwargs, save_method)+_MLX_EXPORTS = {+ "merged": ("export_merged_model", True, {}, "merged_16bit"),+ "merged_4bit": ("export_merged_model", True, {"format_type": "4-bit (FP4)"}, "merged_4bit"),+ "base": ("export_base_model", False, {}, "merged_16bit"),+}+++def _mlx_backend(monkeypatch, name, calls, seen, export):+ method, is_peft, kwargs, _save_method = _MLX_EXPORTS[export]+ backend = _non_mlx_backend(monkeypatch, name, calls, seen)+ export_module = sys.modules[type(backend).__module__]+ monkeypatch.setattr(export_module, "_IS_MLX", True)+ monkeypatch.setattr(export_module, "_export_runtime_available", lambda: True)+ backend.current_model = _MlxMergeModel(export_module.HfApi)+ backend.is_peft = is_peft+ return backend, functools.partial(getattr(backend, method), **kwargs)++[email protected]("export", _MLX_EXPORTS)+def test_mlx_export_push_keeps_the_export_metadata_out_of_the_repo(tmp_path, monkeypatch, export):+ calls: list[str] = []+ seen: dict = {}+ backend, push = _mlx_backend(+ monkeypatch, "test_export_hub_push_mlx_fresh_backend", calls, seen, export+ )++ success, message, output_path = push(+ str(tmp_path / "export"),+ push_to_hub = True,+ repo_id = "model",+ hf_token = "hf_fake",+ )++ assert success is True, message+ assert seen["uploaded"] == _MLX_MERGED_FILES+ assert seen["upload_repo"] == "owner/model"+ assert Path(output_path, "export_metadata.json").is_file()+ assert backend.current_model.merges == [_MLX_EXPORTS[export][3]]+ assert seen["folder"] == output_path++[email protected]("export", _MLX_EXPORTS)+def test_mlx_export_push_to_a_reused_folder_does_not_upload_its_leftovers(+ tmp_path, monkeypatch, export+):+ calls: list[str] = []+ seen: dict = {}+ backend, push = _mlx_backend(+ monkeypatch, "test_export_hub_push_mlx_reused_backend", calls, seen, export+ )+ export_dir = tmp_path / "export"+ export_dir.mkdir()+ (export_dir / "adapter_config.json").write_text("{}")+ (export_dir / "adapters.safetensors").write_bytes(b"lora")++ success, message, output_path = push(+ str(export_dir),+ push_to_hub = True,+ repo_id = "model",+ hf_token = "hf_fake",+ )++ assert success is True, message+ assert seen["uploaded"] == _MLX_MERGED_FILES+ assert output_path == str(export_dir.resolve())+ assert backend.current_model.merges == [_MLX_EXPORTS[export][3]] * 2+ assert seen["folder"] != output_path+ assert not Path(seen["folder"]).exists()++[email protected]("export", _MLX_EXPORTS)+def test_mlx_export_staging_failure_leaves_the_hub_untouched(tmp_path, monkeypatch, export):+ calls: list[str] = []+ seen: dict = {}+ backend, push = _mlx_backend(+ monkeypatch, "test_export_hub_push_mlx_stage_failure", calls, seen, export+ )+ export_dir = tmp_path / "export"+ export_dir.mkdir()+ (export_dir / "old.gguf").write_bytes(b"GGUF")+ save = backend.current_model.save_pretrained_merged++ def fail_staging(directory, *args, **kwargs):+ save(directory, *args, **kwargs)+ if len(backend.current_model.merges) == 2:+ raise OSError(28, "No space left on device")++ monkeypatch.setattr(backend.current_model, "save_pretrained_merged", fail_staging)+ success, message, _path = push(+ str(export_dir), push_to_hub = True, repo_id = "model", hf_token = "hf_fake"+ )++ assert success is False+ assert "No space left on device" in message+ assert Path(export_dir, "model.safetensors").is_file()+ assert calls == []++[email protected]("export", _MLX_EXPORTS)[email protected]("private", [True, False])+def test_mlx_export_push_never_makes_an_existing_private_repo_public(+ tmp_path, monkeypatch, private, export+):+ calls: list[str] = []+ seen: dict = {}+ _backend, push = _mlx_backend(+ monkeypatch, "test_export_hub_push_mlx_private_backend", calls, seen, export+ )++ success, message, _path = push(+ str(tmp_path / "export"),+ push_to_hub = True,+ repo_id = "owner/model",+ hf_token = "hf_fake",+ private = private,+ )++ assert success is True, message+ visibility = ["update_repo_settings"] if private else []+ assert calls == ["create_repo", *visibility, "upload_folder"]+ assert seen.get("visibility") == (+ {"repo_id": "owner/model", "private": True} if private else None+ )+ assert seen["uploaded"] == _MLX_MERGED_FILES++def test_base_export_push_keeps_the_export_metadata_out_of_the_repo(tmp_path, monkeypatch):calls: list[str] = []seen: dict = {}