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
Studio: match audio.cpp CUDA runtime to bundle architecture (#15491)
Daniel Han1ced9390afab
diff --git a/studio/backend/tests/test_audio_cpp_update.py b/studio/backend/tests/test_audio_cpp_update.pyindex b5494a51ee8..478c1a10a68 100644--- a/studio/backend/tests/test_audio_cpp_update.py+++ b/studio/backend/tests/test_audio_cpp_update.py@@ -161,6 +161,30 @@ def test_the_size_adds_the_cuda_runtime_zip(managed, monkeypatch):assert plan["status"]["update_size_bytes"] == 400_000_000 and plan["update_available"] is True[email protected]("arch", ["arm64", "x64"])[email protected]("reverse", [False, True])+def test_update_size_uses_the_matching_architecture_runtime_in_a_mixed_release(+ managed, monkeypatch, arch, reverse+):+ monkeypatch.setattr(aupd.platform, "system", lambda: "Windows")+ monkeypatch.setattr(aupd.platform, "machine", lambda: arch)+ monkeypatch.setattr(aupd._installer(), "torch_provides_cuda_runtime", lambda _major: False)+ _write_record(+ managed.tree,+ asset = f"audio-v0.9.0-unsloth.1-bin-windows-{arch}-cuda13.zip",+ accelerator = "cuda",+ accelerator_request = "cuda",+ )+ assets = [+ (f"audio-{_LATEST}-bin-windows-{arch}-cuda13.zip", 100_000_000),+ (f"audio-{_LATEST}-cudart-windows-x64-cuda13.zip", 300_000_000),+ (f"audio-{_LATEST}-cudart-windows-arm64-cuda13.zip", 200_000_000),+ ]+ managed.state.assets = dict(reversed(assets) if reverse else assets)+ size = aupd.chained_phase_plan()["status"]["update_size_bytes"]+ assert size == (300_000_000 if arch == "arm64" else 400_000_000)++@pytest.mark.parametrize("torch_has_runtime", [False, True])def test_the_cuda_runtime_zip_follows_the_current_torch(managed, monkeypatch, torch_has_runtime):monkeypatch.setattr(aupd.platform, "system", lambda: "Windows")diff --git a/studio/install_audio_cpp_prebuilt.py b/studio/install_audio_cpp_prebuilt.pyindex 6d9887eedad..74691b0f1d9 100644--- a/studio/install_audio_cpp_prebuilt.py+++ b/studio/install_audio_cpp_prebuilt.py@@ -416,13 +416,22 @@ def resolve_release_asset(def cudart_asset_for(asset_names: Sequence[str], bundle: str) -> Optional[str]:- """The separate Windows CUDA runtime archive matching a CUDA bundle's line."""- match = _CUDA_LINE.search(bundle)- if not match or "-windows-" not in bundle:+ """The Windows CUDA runtime matching the bundle's architecture and CUDA line."""+ host = re.search(r"-bin-(?:windows|win)-([^-]+)-", bundle, re.IGNORECASE)+ arch = _ARCH_TOKENS.get(host.group(1).lower()) if host else None+ if host is None or arch is None:+ return None+ match = _CUDA_LINE.match("-" + bundle[host.end() :])+ if not match:return Noneline = match.group(0)[1:]for name in asset_names:- if "-cudart-windows-" in name and name.endswith(f"-{line}.zip"):+ runtime = re.search(r"-cudart-(?:windows|win)-([^-]+)-", name, re.IGNORECASE)+ if (+ runtime is not None+ and _ARCH_TOKENS.get(runtime.group(1).lower()) == arch+ and name.endswith(f"-{line}.zip")+ ):return namereturn Nonediff --git a/tests/studio/install/test_install_audio_cpp_prebuilt_logic.py b/tests/studio/install/test_install_audio_cpp_prebuilt_logic.pyindex c931bc4d07e..91f7a91ef76 100644--- a/tests/studio/install/test_install_audio_cpp_prebuilt_logic.py+++ b/tests/studio/install/test_install_audio_cpp_prebuilt_logic.py@@ -105,6 +105,50 @@ def test_cudart_archive_matches_the_bundle_line():assert M.cudart_asset_for(ASSETS, f"audio-{TAG}-bin-windows-x64-cpu.zip") is None[email protected]("arch", ["arm64", "x64"])[email protected]("reverse", [False, True])+def test_cudart_archive_matches_architecture_independently_of_release_asset_order(arch, reverse):+ tag = "v0.9.2-drama3-base-small-unsloth.2"+ names = [+ f"audio-{tag}-bin-windows-arm64-cpu.zip",+ f"audio-{tag}-bin-windows-arm64-cuda13.zip",+ f"audio-{tag}-cudart-windows-x64-cuda13.zip",+ f"audio-{tag}-cudart-windows-arm64-cuda13.zip",+ f"audio-{tag}-cudart-windows-{arch}-cuda12.zip",+ ]+ if reverse:+ names.reverse()+ assert M.cudart_asset_for(names, f"audio-{tag}-bin-windows-{arch}-cuda13.zip") == (+ f"audio-{tag}-cudart-windows-{arch}-cuda13.zip"+ )++[email protected]("arch,other", [("arm64", "x64"), ("x64", "arm64")])+def test_cudart_never_falls_back_to_another_or_an_unspecified_architecture(arch, other):+ names = [+ f"audio-{TAG}-cudart-windows-{other}-cuda13.zip",+ f"audio-{TAG}-cudart-windows-cuda13.zip",+ f"audio-{TAG}-cudart-windows-unknown-cuda13.zip",+ ]+ assert M.cudart_asset_for(names, f"audio-{TAG}-bin-windows-{arch}-cuda13.zip") is None++[email protected]("arch", ["unknown", "riscv64", ""])+def test_cudart_requires_a_known_bundle_architecture(arch):+ names = [f"audio-{TAG}-cudart-windows-x64-cuda13.zip"]+ assert M.cudart_asset_for(names, f"audio-{TAG}-bin-windows-{arch}-cuda13.zip") is None+++ "bundle_arch,runtime_arch", [("x64", "AMD64"), ("amd64", "x86_64"), ("ARM64", "aarch64")]+)+def test_cudart_preserves_explicit_legacy_architecture_aliases(bundle_arch, runtime_arch):+ runtime = f"audio-{TAG}-cudart-windows-{runtime_arch}-cuda12.4.zip"+ assert (+ M.cudart_asset_for([runtime], f"audio-{TAG}-bin-win-{bundle_arch}-cuda12.4.zip") == runtime+ )++@pytest.mark.parametrize("version,major",[("2.10.0+cu130", 13), ("2.8.0+cu128", 12), ("2.5.1+cu124", 12), ("2.8.0", None)],