Fine-tune LLMs from one YAML. Layer streaming trains an 8B model on a 4 GB laptop GPU.
Python cliconsumer-gpudpofine-tuningggufhuggingfacellmllmopslocal-ailocal-llmloralow-vrammachine-learningollamapeftpythonpytorchqlorasfttransformers
"""Tests for ONNX and TensorRT-LLM export — config, validation, CLI."""
from unittest.mock import MagicMock
from unittest.mock import patch as mock_patch
import pytest
import typer
from soup_cli.commands.export import SUPPORTED_FORMATS
# ─── Format Support Tests ────────────────────────────────────────────────
class TestExportFormats:
"""Test that new export formats are registered."""
def test_gguf_format_supported(self):
assert "gguf" in SUPPORTED_FORMATS
def test_onnx_format_supported(self):
assert "onnx" in SUPPORTED_FORMATS
def test_tensorrt_format_supported(self):
assert "tensorrt" in SUPPORTED_FORMATS
def test_format_count(self):
"""v0.77 — 9 prior formats - awq - gptq = 7."""
assert len(SUPPORTED_FORMATS) == 7
assert "bitnet" in SUPPORTED_FORMATS
assert "tq1_0" in SUPPORTED_FORMATS
assert "torchao" in SUPPORTED_FORMATS
assert "gguf-ud" in SUPPORTED_FORMATS
# ─── ONNX Export CLI Tests ──────────────────────────────────────────────
class TestOnnxExportCLI:
"""Test ONNX export via CLI."""
def test_onnx_export_missing_model_path(self, tmp_path):
"""soup export --format onnx should fail if model path doesn't exist."""
from typer.testing import CliRunner
from soup_cli.cli import app
runner = CliRunner()
result = runner.invoke(
app, ["export", "--model", str(tmp_path / "nonexistent"), "--format", "onnx"]
)
assert result.exit_code != 0
def test_onnx_export_format_in_help(self):
"""Export help should mention onnx format."""
from typer.testing import CliRunner
from soup_cli.cli import app
runner = CliRunner()
result = runner.invoke(app, ["export", "--help"])
assert "onnx" in result.output.lower()
def test_tensorrt_export_format_in_help(self):
"""Export help should mention tensorrt format."""
from typer.testing import CliRunner
from soup_cli.cli import app
runner = CliRunner()
result = runner.invoke(app, ["export", "--help"])
assert "tensorrt" in result.output.lower()
# ─── ONNX Export Function Tests ─────────────────────────────────────────
class TestOnnxExportFunction:
"""Test _export_onnx logic."""
def test_export_onnx_calls_main_export(self, tmp_path):
"""_export_onnx should call optimum's main_export."""
model_dir = tmp_path / "model"
model_dir.mkdir()
mock_main_export = MagicMock()
with mock_patch(
"soup_cli.commands.export.main_export",
mock_main_export,
create=True,
):
# Need to patch the import inside the function
import soup_cli.commands.export as export_mod
original = getattr(export_mod, "main_export", None)
try:
# Inject mock at module level for the lazy import
with mock_patch.object(
export_mod, "_export_onnx",
wraps=export_mod._export_onnx,
):
# Patch the actual import
with mock_patch.dict("sys.modules", {
"optimum": MagicMock(),
"optimum.exporters": MagicMock(),
"optimum.exporters.onnx": MagicMock(
main_export=mock_main_export
),
}):
export_mod._export_onnx(model_dir, str(tmp_path / "out"), None)
mock_main_export.assert_called_once()
finally:
if original is not None:
export_mod.main_export = original
def test_export_onnx_default_output_path(self, tmp_path):
"""Default output path should be model_name + _onnx suffix."""
model_dir = tmp_path / "my_model"
model_dir.mkdir()
mock_main_export = MagicMock()
with mock_patch.dict("sys.modules", {
"optimum": MagicMock(),
"optimum.exporters": MagicMock(),
"optimum.exporters.onnx": MagicMock(main_export=mock_main_export),
}):
import soup_cli.commands.export as export_mod
export_mod._export_onnx(model_dir, None, None)
call_kwargs = mock_main_export.call_args[1]
assert "my_model_onnx" in call_kwargs["output"]
# ─── TensorRT Export Function Tests ──────────────────────────────────────
class TestTensorrtExportFunction:
"""Test _export_tensorrt logic."""
def test_tensorrt_export_calls_subprocess(self, tmp_path):
"""TensorRT export should call subprocess for checkpoint conversion."""
model_dir = tmp_path / "model"
model_dir.mkdir()
mock_result = MagicMock()
mock_result.returncode = 0
with mock_patch.dict("sys.modules", {
"optimum": MagicMock(),
"optimum.exporters": MagicMock(),
"optimum.exporters.onnx": MagicMock(),
"tensorrt_llm": MagicMock(),
"tensorrt_llm.commands": MagicMock(),
"tensorrt_llm.commands.convert_checkpoint": MagicMock(),
}), mock_patch("subprocess.run", return_value=mock_result) as mock_run:
import soup_cli.commands.export as export_mod
export_mod._export_tensorrt(
model_dir, str(tmp_path / "trt_out"), None
)
# Should call subprocess at least twice (checkpoint + build)
assert mock_run.call_count >= 2
def test_tensorrt_export_missing_convert_checkpoint_fails_fast(self, tmp_path):
"""#337: current TensorRT-LLM releases ship no `commands.convert_checkpoint`
submodule, so a plain `tensorrt_llm` install (present but without that entry
point registered) must abort before any subprocess call, not fail deep inside
checkpoint conversion."""
model_dir = tmp_path / "model"
model_dir.mkdir()
with mock_patch.dict("sys.modules", {
"optimum": MagicMock(),
"optimum.exporters": MagicMock(),
"optimum.exporters.onnx": MagicMock(),
"tensorrt_llm": MagicMock(),
}), mock_patch("subprocess.run") as mock_run:
import soup_cli.commands.export as export_mod
with pytest.raises(typer.Exit):
export_mod._export_tensorrt(
model_dir, str(tmp_path / "trt_out"), None
)
mock_run.assert_not_called()
# ─── Unsupported Format Test ─────────────────────────────────────────────
class TestExportUnsupportedFormat:
"""Test that unsupported formats are rejected."""
def test_invalid_format_rejected(self, tmp_path):
"""Export with unsupported format should exit with error."""
from typer.testing import CliRunner
from soup_cli.cli import app
model_dir = tmp_path / "model"
model_dir.mkdir()
runner = CliRunner()
result = runner.invoke(
app,
["export", "--model", str(model_dir), "--format", "safetensors"],
)
assert result.exit_code != 0
assert "Unsupported format" in result.output