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: avoid repeated Qwen-Image-2.1 VAE tile OOMs on ROCm (#13116)

oobabooga4399ef83bf88
diff --git a/studio/backend/core/inference/diffusion_vae_tiling.py b/studio/backend/core/inference/diffusion_vae_tiling.py
index 0db54c04f8e..ce9b6d9f4b5 100644
--- a/studio/backend/core/inference/diffusion_vae_tiling.py
+++ b/studio/backend/core/inference/diffusion_vae_tiling.py
@@ -277,6 +277,16 @@ def decode_tile_budget(vae: Any, z: Any) -> Optional[int]:
raw = (os.environ.get(MAX_TILE_ENV) or "").strip()
if raw.isdigit() and int(raw) > 0:
return int(raw) ** 2
+ # CUDA-calibrated coefficients under-predict Qwen-Image-2.1 decode on ROCm (40x40, 64x32 and 96x56 OOM at
+ # 17/19 GiB limits): start at the tested 32x32 floor rather than retrying an oversized tile on every image.
+ import torch
+
+ if (
+ z.device.type == "cuda"
+ and getattr(torch.version, "hip", None)
+ and type(vae).__name__ == "AutoencoderKLQwenImage21"
+ ):
+ return TILE_LATENTS**2
try:
_, tile, _ = _geometry(vae)
# Stock tile == floor: grow only into device-free memory; decoding out of the allocator cache stalled on
diff --git a/studio/backend/tests/test_qwen21_rocm_vae_budget.py b/studio/backend/tests/test_qwen21_rocm_vae_budget.py
new file mode 100644
index 00000000000..16ec6c84476
--- /dev/null
+++ b/studio/backend/tests/test_qwen21_rocm_vae_budget.py
@@ -0,0 +1,57 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+from types import SimpleNamespace
+
+import pytest
+
+from core.inference import diffusion_vae_tiling as vt
+
+torch = pytest.importorskip("torch")
+
+
+def decoder(monkeypatch):
+ monkeypatch.delenv(vt.MAX_TILE_ENV, raising = False)
+ monkeypatch.setattr(torch.version, "hip", "7.14")
+ vae = type("AutoencoderKLQwenImage21", (), {})()
+ vae._unsloth_wide_geometry = (16, 32, 16)
+ vae._unsloth_wide_stock_tile = 16
+ z = SimpleNamespace(device = SimpleNamespace(type = "cuda"), shape = (1, 64, 1, 96, 96))
+ monkeypatch.setattr(vt, "_free_mib", lambda *a, **kw: (12000, 1))
+ return vae, z
+
+
[email protected]("side", [64, 96, 128])
+def test_rocm_keeps_qwen_tiles_at_the_measured_floor(decoder, side):
+ vae, z = decoder
+ for _ in range(2):
+ budget = vt.decode_tile_budget(vae, z)
+ assert vt.choose_tiles(side, side, budget) == (32, 32)
+
+
[email protected]("other_path", ["cuda", "cpu", "other_vae"])
+def test_other_decoders_retain_their_adaptive_budget(monkeypatch, decoder, other_path):
+ vae, z = decoder
+ if other_path == "cuda":
+ monkeypatch.setattr(torch.version, "hip", None)
+ elif other_path == "cpu":
+ z.device.type = "cpu"
+ else:
+ other = type("AutoencoderKLQwenImage", (), {})()
+ other.__dict__.update(vae.__dict__)
+ vae = other
+ assert vt.decode_tile_budget(vae, z) > 32**2
+
+
+def test_explicit_tile_override_still_wins(monkeypatch, decoder):
+ vae, z = decoder
+ monkeypatch.setenv(vt.MAX_TILE_ENV, "64")
+ assert vt.decode_tile_budget(vae, z) == 64**2
+
+
+def test_floor_memory_refusal_is_preserved(monkeypatch, decoder):
+ vae, z = decoder
+ monkeypatch.setattr(vt, "_free_mib", lambda *a, **kw: (512, 1))
+ assert vt.floor_shortfall(vae, z) is not None
+ assert vt.decode_tile_budget(vae, z) == 32**2