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 video: fall back to the tiled VAE decode when the Linux / macOS CPU allocator refuses (#17075)

Daniel Haned670cce273d
diff --git a/studio/backend/core/inference/video_vae_untiled.py b/studio/backend/core/inference/video_vae_untiled.py
index a94eef1aefa..fda7120ee6d 100644
--- a/studio/backend/core/inference/video_vae_untiled.py
+++ b/studio/backend/core/inference/video_vae_untiled.py
@@ -175,10 +175,24 @@ def _free_bytes(device: Any) -> Optional[int]:
return None
+# torch's CPU allocator refusals: ``not enough memory`` (Windows commit limit), ``can't allocate memory`` (POSIX ENOMEM).
+_CPU_ALLOC_REFUSALS = ("not enough memory", "can't allocate memory")
+
+
def _is_oom(exc: BaseException) -> bool:
- """A device OOM, or the CPU allocator's ``not enough memory`` (raised e.g. on Windows commit limits)."""
from .diffusion_batched import is_oom_error
- return is_oom_error(exc) or "not enough memory" in str(exc).lower()
+
+ if is_oom_error(exc):
+ return True
+ seen = set()
+ cur: Optional[BaseException] = exc
+ while cur is not None and id(cur) not in seen:
+ seen.add(id(cur))
+ msg = str(cur).lower()
+ if any(s in msg for s in _CPU_ALLOC_REFUSALS):
+ return True
+ cur = cur.__cause__ or cur.__context__
+ return False
def _in_chain(fn: Any, target: Any) -> bool:
diff --git a/studio/backend/tests/test_video_ltx2_vae_tiles.py b/studio/backend/tests/test_video_ltx2_vae_tiles.py
index 99ea4b05a89..7baddf843a7 100644
--- a/studio/backend/tests/test_video_ltx2_vae_tiles.py
+++ b/studio/backend/tests/test_video_ltx2_vae_tiles.py
@@ -140,8 +140,13 @@ def test_temb_and_causal_reach_every_tile(free):
"[enforce fail at alloc_cpu.cpp:117] data. DefaultCPUAllocator: not enough memory: "
"you tried to allocate 123 bytes."
),
+ # Linux / macOS: posix_memalign's ENOMEM
+ lambda: RuntimeError(
+ "[enforce fail at alloc_cpu.cpp:127] err == 0. DefaultCPUAllocator: can't allocate memory: "
+ "you tried to allocate 123 bytes. Error code 12 (Cannot allocate memory)"
+ ),
],
- ids = ["cuda", "cpu_allocator"],
+ ids = ["cuda", "cpu_allocator", "posix_allocator"],
)
def test_oom_retries_in_stock_size_tiles(free, monkeypatch, error):
vae = _ltx_vae()
diff --git a/studio/backend/tests/test_video_vae_untiled.py b/studio/backend/tests/test_video_vae_untiled.py
index a96ce406539..34c0b4306ec 100644
--- a/studio/backend/tests/test_video_vae_untiled.py
+++ b/studio/backend/tests/test_video_vae_untiled.py
@@ -242,6 +242,60 @@ def test_cpu_allocator_refusal_falls_back_to_tiled(monkeypatch):
assert vae.use_tiling is True
+ "err",
+ [
+ RuntimeError(
+ "[enforce fail at alloc_cpu.cpp:127] err == 0. DefaultCPUAllocator: can't allocate memory: you tried to "
+ "allocate 9663676416 bytes. Error code 12 (Cannot allocate memory)"
+ ),
+ RuntimeError("decode failed"),
+ ],
+ ids = ["posix_enomem", "wrapped"],
+)
+def test_posix_cpu_allocator_refusal_falls_back_to_tiled(monkeypatch, err):
+ if str(err) == "decode failed":
+ err.__cause__ = RuntimeError(
+ "DefaultCPUAllocator: can't allocate memory: you tried to allocate 1 bytes."
+ )
+ monkeypatch.setattr(U, "_free_bytes", lambda device: 100 * 2**30)
+ vae = FakeVAE(error = err)
+ U.install_untiled_decode(_pipe(vae), WAN)
+ assert vae.decode(Z) == ("tiled",)
+ assert vae.calls == [False, True]
+ assert vae.use_tiling is True
+
+
+class _RefusingVAE(FakeVAE):
+ def decode(
+ self,
+ z,
+ return_dict = False,
+ ):
+ self.calls.append(self.use_tiling)
+ if not self.use_tiling:
+ torch.empty(
+ 2**60, dtype = torch.uint8
+ ) # past any address space: the real allocator refuses
+ return ("tiled",)
+
+
+def test_real_cpu_allocator_refusal_falls_back_to_tiled(monkeypatch):
+ monkeypatch.setattr(U, "_free_bytes", lambda device: 100 * 2**30)
+ vae = _RefusingVAE()
+ U.install_untiled_decode(_pipe(vae), WAN)
+ assert vae.decode(Z) == ("tiled",)
+ assert vae.calls == [False, True]
+
+
+def test_unrelated_runtime_error_still_raises(monkeypatch):
+ monkeypatch.setattr(U, "_free_bytes", lambda device: 100 * 2**30)
+ vae = FakeVAE(error = RuntimeError("shape mismatch"))
+ U.install_untiled_decode(_pipe(vae), WAN)
+ with pytest.raises(RuntimeError, match = "shape mismatch"):
+ vae.decode(Z)
+
+
class CpuOffload:
def __init__(
self,