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 tests: skip pre-quant tests without torchao and pin the Mac memory reading (#17117)

Daniel Han39f0ea3a4302
diff --git a/studio/backend/tests/_torchao_installed.py b/studio/backend/tests/_torchao_installed.py
new file mode 100644
index 00000000000..beabf907c7e
--- /dev/null
+++ b/studio/backend/tests/_torchao_installed.py
@@ -0,0 +1,28 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Skip for tests whose claim needs a real torchao install.
+
+Read from the installed distribution, not ``import torchao``: other tests put fake torchao modules in
+``sys.modules``, so an import can succeed where torchao is absent (macOS, which installs MLX instead).
+"""
+
+from __future__ import annotations
+
+import importlib.metadata
+
+import pytest
+
+
+def torchao_installed() -> bool:
+ try:
+ importlib.metadata.version("torchao")
+ except importlib.metadata.PackageNotFoundError:
+ return False
+ return True
+
+
+needs_torchao = pytest.mark.skipif(
+ not torchao_installed(),
+ reason = "needs torchao installed: without it no pre-quant checkpoint can load, so none is planned",
+)
diff --git a/studio/backend/tests/test_diffusion_backend.py b/studio/backend/tests/test_diffusion_backend.py
index 48512d91c37..5c00d6dcfb5 100644
--- a/studio/backend/tests/test_diffusion_backend.py
+++ b/studio/backend/tests/test_diffusion_backend.py
@@ -17,6 +17,8 @@ from pathlib import Path
import pytest
+from ._torchao_installed import needs_torchao
+
from core.inference.diffusion import (
DiffusionBackend,
DiffusionModelReplacedError,
@@ -8797,6 +8799,7 @@ def test_download_plan_stages_no_second_denoiser_for_an_uncached_prequant(monkey
assert plan["total_bytes"] == 4 * GB + base["bytes"] < 17 * GB
+@needs_torchao
def test_download_plan_counts_the_hosted_prequant_in_the_required_footprint(monkeypatch):
_fake_zimage_hub(monkeypatch)
_stub_hosted_prequant(monkeypatch, cached = False)
@@ -8883,6 +8886,7 @@ def test_download_plan_counts_a_cached_lower_auto_prequant(monkeypatch):
assert not any(source.filename in entry["files"] for entry in plan["entries"])
+@needs_torchao
def test_download_plan_omits_the_prequant_under_a_definite_offload_policy(monkeypatch):
# Balanced / low_vram offload by MODE, so the load keeps the GGUF and never fetches the prequant.
_fake_zimage_hub(monkeypatch)
@@ -8910,6 +8914,7 @@ def test_download_plan_omits_the_prequant_under_a_definite_offload_policy(monkey
assert any(f.endswith("-FP8.pt") for e in resident["entries"] for f in e["files"])
+@needs_torchao
def test_download_plan_omits_the_prequant_for_an_auto_pick_at_speed_off(monkeypatch):
# load_pipeline forces an AUTO quant to "off" under Speed="off", which normalizes to None and
# skips the fast path, so nothing is fetched and the footprint must not claim it.
@@ -11823,6 +11828,7 @@ def test_generation_in_flight_never_builds_a_backend(fake_runtime, monkeypatch):
assert diffusion_mod.generation_in_flight() is False
+@needs_torchao
def test_the_download_plan_resolves_the_same_nvfp4_rung_the_load_does(monkeypatch):
from types import SimpleNamespace
@@ -14202,6 +14208,8 @@ def test_gguf_offload_swap_on_mps_keeps_the_gguf(fake_runtime, tmp_path, monkeyp
from core.inference.diffusion_transformer_quant import dense_transformer_supported
monkeypatch.setattr(dmod, "dense_transformer_supported", dense_transformer_supported)
+ # The unified-memory check reads the host's free memory; a fixed 64 GB Mac keeps the runner's out of it.
+ monkeypatch.setattr(dmod, "settled_snapshot_device_memory", _unified_snapshot(64))
# z-image defaults to its hosted int8 ConvRot encoder, which this offline fixture does not carry; MPS budgets
# for that file alone, so the encoder is not what is under test here.
status = _load_gguf(backend, tmp_path, text_encoder_quant = "none")
diff --git a/studio/backend/tests/test_diffusion_pipeline_prequant.py b/studio/backend/tests/test_diffusion_pipeline_prequant.py
index 4dd40fdac2f..ea06f4537cb 100644
--- a/studio/backend/tests/test_diffusion_pipeline_prequant.py
+++ b/studio/backend/tests/test_diffusion_pipeline_prequant.py
@@ -13,6 +13,8 @@ import types
import pytest
+from ._torchao_installed import needs_torchao
+
from core.inference import diffusion as dmod
from core.inference import diffusion_prequant as pqmod
from core.inference import diffusion_transformer_quant as tqmod
@@ -201,6 +203,7 @@ def _plan_backend(
return backend
+@needs_torchao
def test_the_plan_counts_and_stages_the_hosted_checkpoint(monkeypatch, hub):
backend = _plan_backend(monkeypatch, planned = "fp8")
plan = backend.download_plan(Z_IMAGE_REPO, model_kind = "pipeline")
@@ -235,6 +238,7 @@ def test_a_declined_plan_keeps_the_released_shards(monkeypatch, hub):
assert all(shard in entries[Z_IMAGE_REPO]["files"] for shard in DENOISER_SHARDS)
+@needs_torchao
def test_the_artifact_carries_the_denoiser_share_of_the_unified_memory_verdict(monkeypatch, hub):
"""The unified-memory refusal prices the denoiser from the artifact, not the dropped shards."""
backend = _plan_backend(monkeypatch, planned = "fp8")
@@ -1963,6 +1967,7 @@ def test_a_large_variant_of_a_small_dit_family_quantises_under_auto(fake_runtime
assert spy.quantised == ["auto"]
+@needs_torchao
@pytest.mark.parametrize("shared_repo", [False, True])
def test_denoiser_is_a_checkpoint_and_encoder_only_is_an_asset(monkeypatch, hub, shared_repo):
backend = _plan_backend(monkeypatch, planned = "fp8")
diff --git a/studio/backend/tests/test_prequant_safetensors.py b/studio/backend/tests/test_prequant_safetensors.py
index 14b067f4ade..f01362eaef1 100644
--- a/studio/backend/tests/test_prequant_safetensors.py
+++ b/studio/backend/tests/test_prequant_safetensors.py
@@ -11,6 +11,8 @@ import sys
import pytest
+from ._torchao_installed import needs_torchao
+
import core.inference.diffusion_prequant as pq
import core.inference.prequant_safetensors as ps
from core.inference.diffusion_families import DiffusionFamily
@@ -392,6 +394,7 @@ def test_the_windows_rocm_torchao_stub_is_not_safetensors_support(monkeypatch):
assert restricted_prequant_load_supported("fp8", "Model-FP8.pt") is False
+@needs_torchao
def test_a_scheme_this_torchao_cannot_flatten_is_reported_before_the_build(monkeypatch):
"""Importing helpers proves nothing; torchao 0.17 int8 cannot flatten, so report it before building."""
import torchao.quantization as tq
@@ -428,6 +431,7 @@ def test_a_scheme_this_torchao_cannot_flatten_is_reported_before_the_build(monke
assert ps.scheme_is_flattenable("cfg") is False
+@needs_torchao
def test_the_real_installed_torchao_answers_the_int8_question(monkeypatch):
"""Not a mock: the probe against the torchao actually installed, whichever way it answers."""
from core.inference.diffusion_transformer_quant import _make_quant_config
@@ -443,6 +447,7 @@ def test_the_real_installed_torchao_answers_the_int8_question(monkeypatch):
assert answer is not False, f"torchao {torchao.__version__} should flatten int8"
+@needs_torchao
def test_a_root_level_plain_tensor_round_trips_instead_of_failing_the_build(tmp_path):
"""Root-level plain tensors are split before flatten and restored after, so z-image builds survive."""
torch = pytest.importorskip("torch")
diff --git a/studio/backend/tests/test_video_te_quant_default.py b/studio/backend/tests/test_video_te_quant_default.py
index 431d2914215..208dda62866 100644
--- a/studio/backend/tests/test_video_te_quant_default.py
+++ b/studio/backend/tests/test_video_te_quant_default.py
@@ -12,6 +12,8 @@ import types
import pytest
import torch
+from ._torchao_installed import needs_torchao
+
import core.inference.diffusion_te_prequant as tpq
from core.inference.video import video_te_quant_request
from core.inference.video_families import _FAMILIES, detect_video_family
@@ -341,6 +343,7 @@ def _converter():
return module
+@needs_torchao
def test_the_converter_can_store_a_tied_embedding_once(tmp_path, monkeypatch):
"""--drop-tied-copies writes UMT5's shared/embed_tokens pair once, the round trip still verifies, and the fp8
loader fills the alias back from the model's own tie."""
@@ -592,6 +595,7 @@ def test_ltx25_budgets_the_dense_peak_unless_the_hosted_encoder_applies(monkeypa
)
+@needs_torchao
def test_drop_tied_copies_keeps_a_tied_lm_head(tmp_path):
import core.inference.prequant_safetensors as ps