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
# SPDX-License-Identifier: AGPL-3.0-only
# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
"""The extension-free mamba-ssm wheel: pure_matrix, pure-gate, notes and the workflow jobs."""
from __future__ import annotations
import importlib.util
import json
import zipfile
from pathlib import Path
import pytest
import yaml
REPO = Path(__file__).resolve().parents[2]
WORKFLOW = REPO / ".github" / "workflows" / "prebuilt-cuda-wheels.yml"
SCRIPTS = REPO / ".github" / "scripts"
_spec = importlib.util.spec_from_file_location("prebuilt_wheels", SCRIPTS / "prebuilt_wheels.py")
pw = importlib.util.module_from_spec(_spec)
_spec.loader.exec_module(pw)
NAME = "mamba_ssm-2.3.2.post1-py3-none-any.whl"
def _wheel(
tmp_path,
name = NAME,
extra = (),
tag = "py3-none-any",
purelib = "true",
):
path = tmp_path / name
with zipfile.ZipFile(path, "w") as archive:
archive.writestr("mamba_ssm/__init__.py", '__version__ = "2.3.2.post1"\n')
archive.writestr("mamba_ssm/ops/triton/ssd_combined.py", "")
archive.writestr(
"mamba_ssm-2.3.2.post1.dist-info/WHEEL",
f"Wheel-Version: 1.0\nRoot-Is-Purelib: {purelib}\nTag: {tag}\n",
)
for member in extra:
archive.writestr(member, b"\0")
return path
def test_name_follows_the_cuda_wheel_pin():
assert pw.pure_wheel_name() == NAME
assert pw.SPECS["mamba-ssm"]["version"] == "2.3.2.post1"
@pytest.mark.parametrize(
"packages, count",
[
("", 1),
("mamba-ssm", 1),
("flash-attn,mamba-ssm", 1),
("flash-attn", 0),
("causal-conv1d", 0),
],
)
def test_one_cell_only_when_the_run_builds_mamba_ssm(packages, count):
cells = pw.pure_matrix(packages)
assert len(cells) == count
if cells:
cell = cells[0]
assert cell["wheel_name"] == NAME
# Same source as the CUDA wheel.
assert (cell["repo"], cell["ref"]) == (
pw.SPECS["mamba-ssm"]["repo"],
pw.SPECS["mamba-ssm"]["ref"],
)
assert "ssd_combined import mamba_chunk_scan_combined" in cell["smoke_imports"]
def test_matrix_writes_pure_outputs(tmp_path, monkeypatch):
out = tmp_path / "out"
monkeypatch.setenv("GITHUB_OUTPUT", str(out))
monkeypatch.delenv("GITHUB_STEP_SUMMARY", raising = False)
for packages, count in (("mamba-ssm", 1), ("flash-attn", 0)):
out.write_text("")
monkeypatch.setenv("UW_PACKAGES", packages)
monkeypatch.setenv("UW_PLATFORMS", "linux_x86_64")
assert pw.main(["matrix", "--github"]) == 0
lines = dict(line.split("=", 1) for line in out.read_text().splitlines())
assert lines["pure_count"] == str(count)
assert len(json.loads(lines["pure_matrix"])["include"]) == count
# The CUDA cells are unchanged by it.
assert "py3-none-any" not in lines["matrix"]
def test_gate_accepts_the_pure_wheel(tmp_path):
assert pw.main(["pure-gate", str(_wheel(tmp_path))]) == 0
@pytest.mark.parametrize(
"kwargs, problem",
[
({"extra": ["selective_scan_cuda.cpython-313-x86_64-linux-gnu.so"]}, "compiled code"),
({"extra": ["selective_scan_cuda.cp313-win_amd64.pyd"]}, "compiled code"),
({"name": "mamba_ssm-2.3.2.post1-cp313-cp313-linux_x86_64.whl"}, "expected"),
({"tag": "cp313-cp313-linux_x86_64"}, "not pure"),
({"purelib": "false"}, "not pure"),
],
)
def test_gate_refuses_anything_else(tmp_path, kwargs, problem):
wheel = _wheel(tmp_path, **kwargs)
assert any(problem in p for p in pw.check_pure_wheel(str(wheel)))
assert pw.main(["pure-gate", str(wheel)]) == 1
def test_notes_mention_it_and_sums_list_it():
cuda = "mamba_ssm-2.3.2.post1+cu13torch2.13cxx11abiTRUE-cp313-cp313-linux_x86_64.whl"
entries = [("a" * 64, cuda), ("b" * 64, NAME)]
notes = pw.render_notes(entries, tag = "t", repo = "o/r")
assert "without its CUDA extension" in notes and NAME in notes
assert "mamba-ssm 2.3.2.post1" in notes
assert NAME in pw.render_sums(entries)
only = pw.render_notes([("b" * 64, NAME)], tag = "t", repo = "o/r")
assert NAME in only and "No wheels" not in only
assert "without its CUDA extension" not in pw.render_notes(
[("a" * 64, cuda)], tag = "t", repo = "o/r"
)
@pytest.fixture(scope = "module")
def workflow():
return yaml.safe_load(WORKFLOW.read_text(encoding = "utf-8"))
def test_pure_jobs_are_planned_and_gated(workflow):
jobs = workflow["jobs"]
outputs = jobs["plan"]["outputs"]
assert "pure_matrix" in outputs and "pure_count" in outputs
pure = jobs["pure"]
assert "pure_count != '0'" in pure["if"]
assert pure["runs-on"] == "ubuntu-latest"
steps = {step.get("name", ""): step for step in pure["steps"]}
build = steps["Build the wheel"]
assert build["env"]["MAMBA_KEEP_CUDA_BUILD"] == "FALSE"
assert build["env"]["MAMBA_FORCE_BUILD"] == "FALSE"
assert "pure-gate" in steps["Check the wheel is pure"]["run"]
assert "--no-deps" in steps["Import it (Linux)"]["run"]
upload = next(step for step in pure["steps"] if "upload-artifact" in step.get("uses", ""))
# The sign job downloads wheel-*; the Windows import job downloads the same name.
assert upload["with"]["name"] == "wheel-py3-none-any-mamba_ssm"
windows = jobs["pure-windows"]
assert windows["runs-on"] == "windows-latest"
assert any("triton_windows" in str(step.get("env", {})) for step in windows["steps"])
download = next(
step for step in windows["steps"] if "download-artifact" in step.get("uses", "")
)
assert download["with"]["name"] == upload["with"]["name"]
def test_sign_waits_for_the_pure_wheel_only_when_planned(workflow):
sign = workflow["jobs"]["sign"]
assert {"pure", "pure-windows"} <= set(sign["needs"])
condition = sign["if"]
assert "pure_count == '0'" in condition
assert "needs.pure.result == 'success'" in condition
assert "needs.pure-windows.result == 'success'" in condition
arrived = next(
step for step in sign["steps"] if step.get("name") == "Every planned wheel arrived"
)
assert "PURE_COUNT" in arrived["env"] and "CUDA_COUNT" in arrived["env"]
assert "CUDA_COUNT + ${PURE_COUNT:-0}" in arrived["run"]