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.

Studio: fix Push to Hub for base and quantized merged exports on transformers 5 (#10943)

Nilayb83a7f5af948
diff --git a/studio/backend/core/export/export.py b/studio/backend/core/export/export.py
index e77e2a96133..30038e6d996 100644
--- a/studio/backend/core/export/export.py
+++ b/studio/backend/core/export/export.py
@@ -47,7 +47,6 @@ _TORCH_IMPORT_ERROR: Optional[BaseException] = None
if not _IS_MLX:
try:
from peft import PeftModel, PeftModelForCausalLM
- from transformers.modeling_utils import PushToHubMixin
import torch
except Exception as _torch_exc:
_TORCH_IMPORT_ERROR = _torch_exc
@@ -452,8 +451,7 @@ def _ensure_hub_repo_private(hf_api, repo_id):
raise RuntimeError(
f"private=True was requested but {repo_id!r} could not be confirmed private "
"(the token likely lacks `write:repo_settings`, or the repository belongs to "
- "someone else). Refusing to upload rather than publish the GGUF files to a "
- "public repository."
+ "someone else). Refusing to upload rather than publish to a public repository."
) from exception
@@ -953,12 +951,10 @@ class ExportBackend:
# Upload the artifact already built in output_path; push_to_hub_merged(save_method=...) would
# redo the expensive quantization.
hf_api = HfApi(token = hf_token)
- repo_id = PushToHubMixin._create_repo(
- PushToHubMixin,
- repo_id = repo_id,
- private = private,
- token = hf_token,
- )
+ repo_url = hf_api.create_repo(repo_id, private = private, exist_ok = True)
+ repo_id = getattr(repo_url, "repo_id", repo_id)
+ if private:
+ _ensure_hub_repo_private(hf_api, repo_id)
content = MODEL_CARD.format(
username = repo_id.split("/")[0],
base_model = getattr(self.current_model.config, "_name_or_path", "unknown"),
@@ -1076,12 +1072,10 @@ class ExportBackend:
)
hf_api = HfApi(token = hf_token)
- repo_id = PushToHubMixin._create_repo(
- PushToHubMixin,
- repo_id = repo_id,
- private = private,
- token = hf_token,
- )
+ repo_url = hf_api.create_repo(repo_id, private = private, exist_ok = True)
+ repo_id = getattr(repo_url, "repo_id", repo_id)
+ if private:
+ _ensure_hub_repo_private(hf_api, repo_id)
username = repo_id.split("/")[0]
content = MODEL_CARD.format(
diff --git a/studio/backend/tests/test_export_hub_push.py b/studio/backend/tests/test_export_hub_push.py
new file mode 100644
index 00000000000..46f3f430d4a
--- /dev/null
+++ b/studio/backend/tests/test_export_hub_push.py
@@ -0,0 +1,142 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+from __future__ import annotations
+
+import importlib.util
+import sys
+import types
+from pathlib import Path
+
+import pytest
+
+
+_HELPERS_SPEC = importlib.util.spec_from_file_location(
+ "export_hub_push_helpers",
+ Path(__file__).with_name("test_export_gguf_hub_upload.py"),
+)
+assert _HELPERS_SPEC is not None and _HELPERS_SPEC.loader is not None
+_HELPERS = importlib.util.module_from_spec(_HELPERS_SPEC)
+_HELPERS_SPEC.loader.exec_module(_HELPERS)
+
+
+class _Config:
+ _name_or_path = "unsloth/Qwen2.5-0.5B-Instruct"
+ model_type = "qwen2"
+
+
+class _Tokenizer:
+ def save_pretrained(self, save_directory):
+ Path(save_directory, "tokenizer.json").write_bytes(b"{}")
+
+
+class _Model:
+ config = _Config()
+
+ def save_pretrained(self, save_directory):
+ Path(save_directory, "model.safetensors").write_bytes(b"weights")
+
+ def save_pretrained_merged(
+ self,
+ save_directory,
+ tokenizer,
+ save_method = None,
+ token = None,
+ ):
+ output = Path(f"{save_directory}-torchao-fp8")
+ output.mkdir(parents = True)
+ (output / "model.safetensors").write_bytes(b"fp8")
+
+
+def _non_mlx_backend(monkeypatch, name, calls, seen):
+ _HELPERS._install_export_backend_stubs(monkeypatch)
+ unsloth = sys.modules["unsloth"]
+ monkeypatch.setattr(unsloth, "_IS_MLX", False)
+
+ unsloth_save = types.ModuleType("unsloth.save")
+ unsloth_save._normalize_torchao_method = lambda method: (
+ ("fp8", "torchao-fp8") if method == "torchao_fp8" else None
+ )
+ monkeypatch.setattr(unsloth, "save", unsloth_save, raising = False)
+ monkeypatch.setitem(sys.modules, "unsloth.save", unsloth_save)
+
+ peft = types.ModuleType("peft")
+ peft.PeftModel = object
+ peft.PeftModelForCausalLM = object
+ transformers = types.ModuleType("transformers")
+ transformers.__path__ = []
+ modeling_utils = types.ModuleType("transformers.modeling_utils")
+ modeling_utils.PushToHubMixin = type("PushToHubMixin", (), {})
+ transformers.modeling_utils = modeling_utils
+ monkeypatch.setitem(sys.modules, "peft", peft)
+ monkeypatch.setitem(sys.modules, "transformers", transformers)
+ monkeypatch.setitem(sys.modules, "transformers.modeling_utils", modeling_utils)
+ monkeypatch.setitem(sys.modules, "torch", types.ModuleType("torch"))
+
+ export_module = _HELPERS._load_module(name, "core/export/export.py", monkeypatch)
+ _HELPERS._patch_hub(monkeypatch, export_module, calls, seen)
+
+ backend = export_module.ExportBackend.__new__(export_module.ExportBackend)
+ backend.current_model = _Model()
+ backend.current_tokenizer = _Tokenizer()
+ backend.current_checkpoint = None
+ backend.is_peft = False
+ backend._audio_type = None
+ return backend
+
+
+def _expected_calls(private):
+ visibility = ["update_repo_settings"] if private else []
+ return ["create_repo", *visibility, "model_card", "upload_folder"]
+
+
[email protected]("private", [True, False])
+def test_base_export_push_creates_the_repo_without_push_to_hub_mixin(
+ tmp_path, monkeypatch, private
+):
+ calls: list[str] = []
+ seen: dict = {}
+ backend = _non_mlx_backend(monkeypatch, "test_export_hub_push_base_backend", calls, seen)
+
+ success, message, output_path = backend.export_base_model(
+ str(tmp_path / "export"),
+ push_to_hub = True,
+ repo_id = "model",
+ hf_token = "hf_fake",
+ private = private,
+ )
+
+ assert success is True, message
+ assert seen["token"] == "hf_fake"
+ assert seen["repo"] == {"repo_id": "model", "private": private, "exist_ok": True}
+ assert calls == _expected_calls(private)
+ assert seen["card_repo"] == "owner/model"
+ assert seen["folder"] == output_path
+ assert "model.safetensors" in seen["uploaded"]
+
+
[email protected]("private", [True, False])
+def test_merged_torchao_export_push_creates_the_repo_without_push_to_hub_mixin(
+ tmp_path, monkeypatch, private
+):
+ calls: list[str] = []
+ seen: dict = {}
+ backend = _non_mlx_backend(monkeypatch, "test_export_hub_push_merged_backend", calls, seen)
+
+ success, message, output_path = backend.export_merged_model(
+ str(tmp_path / "export"),
+ push_to_hub = True,
+ repo_id = "model",
+ hf_token = "hf_fake",
+ private = private,
+ compressed_method = "torchao_fp8",
+ )
+
+ assert success is True, message
+ assert output_path == str(Path(f"{tmp_path / 'export'}-torchao-fp8").resolve())
+ assert seen["token"] == "hf_fake"
+ assert seen["repo"] == {"repo_id": "model", "private": private, "exist_ok": True}
+ assert calls == _expected_calls(private)
+ assert seen["card_repo"] == "owner/model"
+ assert seen["folder"] == output_path
+ assert "model.safetensors" in seen["uploaded"]