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.

Give push_to_ollama the arguments create_ollama_modelfile takes (#10304)

Vineeth Sai Varikuntla2a95eb78bcd4
diff --git a/.github/workflows/consolidated-tests-ci.yml b/.github/workflows/consolidated-tests-ci.yml
index 437f10ff893..44043db6864 100644
--- a/.github/workflows/consolidated-tests-ci.yml
+++ b/.github/workflows/consolidated-tests-ci.yml
@@ -448,7 +448,8 @@ jobs:
tests/test_gemma_2b_mapper_key.py \
tests/test_raw_text_json_loading.py \
tests/test_map_eos_token.py \
- tests/test_apertus_instruct_mapping.py
+ tests/test_apertus_instruct_mapping.py \
+ tests/test_push_to_ollama_modelfile_args.py
# test_run_attention_flash_varlen_receives_window_and_softcap was deselected
# until attention_dispatch.py predefined flash_attn_varlen_func as None; it
# monkeypatches that name, so it no longer needs flash_attn on this runner.
diff --git a/tests/test_push_to_ollama_modelfile_args.py b/tests/test_push_to_ollama_modelfile_args.py
new file mode 100644
index 00000000000..ff413145eef
--- /dev/null
+++ b/tests/test_push_to_ollama_modelfile_args.py
@@ -0,0 +1,183 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+
+"""push_to_ollama has to call create_ollama_modelfile with the arguments it takes.
+
+create_ollama_modelfile was `(tokenizer, gguf_location)` when push_to_ollama was written in
+#1648. It is now `(tokenizer, base_model_name, model_location)`, and this caller was left
+behind, so it raised `TypeError: unexpected keyword argument 'gguf_location'` before it
+reached Ollama. save.py needs a GPU to import, so the function is ast-extracted the way
+tests/test_ollama_eos_token_order.py extracts its own.
+
+Stubs are checked against save.py's AST, or a later rename drifts past a hand-copied one and
+leaves this green while the caller breaks again.
+"""
+
+import ast
+import importlib.util
+import os
+
+import pytest
+
+_TESTS = os.path.dirname(__file__)
+_SAVE = os.path.join(_TESTS, os.pardir, "unsloth", "save.py")
+_MAPPERS = os.path.join(_TESTS, os.pardir, "unsloth", "ollama_template_mappers.py")
+
+MAPPED = "unsloth/llama-3-8b-Instruct"
+UNMAPPED = "some/unmapped-model"
+
+
+def _parse():
+ with open(_SAVE, encoding = "utf-8") as f:
+ source = f.read()
+ return source, ast.parse(source)
+
+
+def _extract(name):
+ source, tree = _parse()
+ for node in tree.body:
+ if isinstance(node, ast.FunctionDef) and node.name == name:
+ return ast.get_source_segment(source, node)
+ raise AssertionError(f"{name} not found in save.py")
+
+
+def _params(name):
+ """save.py's real parameter names, so a stub cannot drift away from them."""
+ _, tree = _parse()
+ for node in tree.body:
+ if isinstance(node, ast.FunctionDef) and node.name == name:
+ args = node.args
+ return [a.arg for a in args.posonlyargs + args.args + args.kwonlyargs]
+ raise AssertionError(f"{name} not found in save.py")
+
+
+def _mappers():
+ """stdlib-only, so it loads without save.py's GPU import chain."""
+ spec = importlib.util.spec_from_file_location("_ollama_mappers", _MAPPERS)
+ module = importlib.util.module_from_spec(spec)
+ spec.loader.exec_module(module)
+ return module
+
+
+def _load(modelfile):
+ """Exec push_to_ollama beside stubs that carry the REAL signatures of what it calls."""
+ seen = {}
+
+ def create_ollama_modelfile(tokenizer, base_model_name, model_location):
+ seen["modelfile"] = (tokenizer, base_model_name, model_location)
+ return modelfile
+
+ def create_ollama_model(username, model_name, tag, modelfile_path):
+ seen["model"] = (username, model_name, tag, modelfile_path)
+
+ def push_to_ollama_hub(username, model_name, tag):
+ seen["hub"] = (username, model_name, tag)
+
+ namespace = {
+ "create_ollama_modelfile": create_ollama_modelfile,
+ "create_ollama_model": create_ollama_model,
+ "push_to_ollama_hub": push_to_ollama_hub,
+ }
+ for name, stub in namespace.items():
+ assert _params(name) == list(
+ stub.__code__.co_varnames[: stub.__code__.co_argcount]
+ ), f"the {name} stub no longer matches save.py; update it and this test's assertions"
+ exec(compile(_extract("push_to_ollama"), "push_to_ollama", "exec"), namespace)
+ return namespace["push_to_ollama"], seen
+
+
+def _load_real():
+ """Same, but the real create_ollama_modelfile and the real templates."""
+ mappers = _mappers()
+ calls = []
+ namespace = {
+ "OLLAMA_TEMPLATES": mappers.OLLAMA_TEMPLATES,
+ "MODEL_TO_OLLAMA_TEMPLATE_MAPPER": mappers.MODEL_TO_OLLAMA_TEMPLATE_MAPPER,
+ "create_ollama_model": lambda **kw: calls.append(("create", kw)),
+ "push_to_ollama_hub": lambda **kw: calls.append(("push", kw)),
+ }
+ for name in ("create_ollama_modelfile", "push_to_ollama"):
+ exec(compile(_extract(name), name, "exec"), namespace)
+ return namespace["push_to_ollama"], calls, mappers
+
+
+class _Tokenizer:
+ eos_token = "<|eot_id|>"
+
+
+def test_push_to_ollama_reaches_create_ollama_modelfile(tmp_path, monkeypatch):
+ monkeypatch.chdir(tmp_path)
+ push_to_ollama, seen = _load("FROM ./model.gguf\n")
+
+ push_to_ollama(
+ tokenizer = "TOKENIZER",
+ base_model_name = MAPPED,
+ gguf_location = "./model.gguf",
+ username = "user",
+ model_name = "my-model",
+ tag = "latest",
+ )
+
+ assert seen["modelfile"] == ("TOKENIZER", MAPPED, "./model.gguf")
+ assert seen["model"] == ("user", "my-model", "latest", "Modelfile_my-model")
+ assert seen["hub"] == ("user", "my-model", "latest")
+ assert (tmp_path / "Modelfile_my-model").read_text(encoding = "utf-8") == "FROM ./model.gguf\n"
+
+
+def test_push_to_ollama_reports_a_missing_template(tmp_path, monkeypatch):
+ # create_ollama_modelfile returns None when the model has no Ollama template mapping;
+ # writing that to the Modelfile used to fail with an opaque TypeError from f.write.
+ monkeypatch.chdir(tmp_path)
+ push_to_ollama, _ = _load(None)
+
+ with pytest.raises(RuntimeError, match = "No Ollama template mapping"):
+ push_to_ollama(
+ tokenizer = "TOKENIZER",
+ base_model_name = UNMAPPED,
+ gguf_location = "./model.gguf",
+ username = "user",
+ model_name = "my-model",
+ tag = "latest",
+ )
+
+ assert not (tmp_path / "Modelfile_my-model").exists()
+
+
+def test_push_to_ollama_writes_a_real_modelfile(tmp_path, monkeypatch):
+ monkeypatch.chdir(tmp_path)
+ push_to_ollama, calls, mappers = _load_real()
+ assert mappers.MODEL_TO_OLLAMA_TEMPLATE_MAPPER.get(MAPPED), f"{MAPPED} left the mapper"
+
+ push_to_ollama(
+ tokenizer = _Tokenizer(),
+ base_model_name = MAPPED,
+ gguf_location = "./model.gguf",
+ username = "user",
+ model_name = "my-model",
+ tag = "latest",
+ )
+
+ text = (tmp_path / "Modelfile_my-model").read_text(encoding = "utf-8")
+ assert "FROM ./model.gguf" in text
+ assert "__FILE_LOCATION__" not in text and "__EOS_TOKEN__" not in text
+ assert [c[0] for c in calls] == ["create", "push"]
+ assert calls[0][1]["modelfile_path"] == "Modelfile_my-model"
+
+
+def test_push_to_ollama_rejects_an_unmapped_model_end_to_end(tmp_path, monkeypatch):
+ monkeypatch.chdir(tmp_path)
+ push_to_ollama, calls, mappers = _load_real()
+ assert mappers.MODEL_TO_OLLAMA_TEMPLATE_MAPPER.get(UNMAPPED) is None
+
+ with pytest.raises(RuntimeError, match = "No Ollama template mapping"):
+ push_to_ollama(
+ tokenizer = _Tokenizer(),
+ base_model_name = UNMAPPED,
+ gguf_location = "./model.gguf",
+ username = "user",
+ model_name = "my-model",
+ tag = "latest",
+ )
+
+ assert not (tmp_path / "Modelfile_my-model").exists()
+ assert calls == []
diff --git a/unsloth/save.py b/unsloth/save.py
index aba5449830e..4ce5509e420 100644
--- a/unsloth/save.py
+++ b/unsloth/save.py
@@ -3041,8 +3041,19 @@ def push_to_ollama_hub(username: str, model_name: str, tag: str):
print("\nMODEL PUBLISHED SUCCESSFULLY")
-def push_to_ollama(tokenizer, gguf_location, username: str, model_name: str, tag: str):
- model_file = create_ollama_modelfile(tokenizer = tokenizer, gguf_location = gguf_location)
+def push_to_ollama(
+ tokenizer, base_model_name, gguf_location, username: str, model_name: str, tag: str
+):
+ model_file = create_ollama_modelfile(
+ tokenizer = tokenizer,
+ base_model_name = base_model_name,
+ model_location = gguf_location,
+ )
+ if model_file is None:
+ raise RuntimeError(
+ f"Unsloth: No Ollama template mapping found for model '{base_model_name}', "
+ f"so there is no Modelfile to push."
+ )
with open(f"Modelfile_{model_name}", "w", encoding = "utf-8") as f:
f.write(model_file)