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 (AMD/ROCm): use MIOpen immediate mode for image and video convs (#13323)

Daniel Hanef5e0507a0c9
diff --git a/studio/backend/core/inference/diffusion_speed.py b/studio/backend/core/inference/diffusion_speed.py
index 1aa50b11eb6..5bc9f88c73c 100644
--- a/studio/backend/core/inference/diffusion_speed.py
+++ b/studio/backend/core/inference/diffusion_speed.py
@@ -503,6 +503,8 @@ def apply_speed_optims(
install_vae_attn_chunks(getattr(pipe, "vae", None), target, logger)
except Exception as exc: # noqa: BLE001 - keep stock attention
_warn(logger, "vae attention chunks", exc)
+ # Stability, not speed, so also before the `off` return.
+ _enable_miopen_immediate(logger)
# TF32 (max) and cudnn.benchmark (any non-off CUDA load) are process-global; the caller restores them so a later
# `off` load never inherits them.
if mode == SPEED_OFF:
@@ -1777,6 +1779,30 @@ def _enable_cudnn_benchmark(logger: Any) -> bool:
return False
+def _enable_miopen_immediate(logger: Any) -> bool:
+ """ROCm: ask MIOpen for a conv solution instead of miopenFind*, whose trial launches on unseen shapes
+ are where #9130's uncatchable abort came from. MIOPEN_FIND_MODE / MIOPEN_FIND_ENFORCE keep Find. Not in
+ snapshot_backend_flags: an unload restoring it would reopen Find under a co-resident pipeline."""
+ if os.environ.get("MIOPEN_FIND_MODE") or os.environ.get("MIOPEN_FIND_ENFORCE"):
+ return False
+ try:
+ import torch
+
+ from core._torchao_stub import _module_is_rocm
+
+ miopen = getattr(torch.backends, "miopen", None)
+ if not _module_is_rocm(torch) or miopen is None or not hasattr(miopen, "immediate"):
+ return False
+ if not miopen.immediate:
+ miopen.immediate = True
+ if logger is not None:
+ logger.info("diffusion.speed: MIOpen immediate mode on (no Find trial launches)")
+ return True
+ except Exception as exc: # noqa: BLE001 - keep torch's Find path
+ _warn(logger, "miopen immediate", exc)
+ return False
+
+
def _enable_tf32(logger: Any) -> bool:
try:
import torch
diff --git a/studio/backend/tests/test_diffusion_speed.py b/studio/backend/tests/test_diffusion_speed.py
index 8705a25fddb..9a7dbe994e7 100644
--- a/studio/backend/tests/test_diffusion_speed.py
+++ b/studio/backend/tests/test_diffusion_speed.py
@@ -590,6 +590,57 @@ def test_speed_default_skips_cudnn_benchmark_on_rocm(monkeypatch, hip, version):
assert torch.backends.cudnn.benchmark is False
[email protected]("speed_mode", [SPEED_OFF, SPEED_DEFAULT])
+ "hip, version",
+ [("7.13.0", "2.11.0+rocm7.13.0"), (None, "2.9.1+rocmsdk20251116")],
+ ids = ["version_hip", "rocm_tag_only"],
+)
+def test_rocm_load_turns_on_miopen_immediate(monkeypatch, hip, version, speed_mode):
+ # #9130: torch's MIOpen Find trial-launches kernels; one faulting there aborted the server.
+ monkeypatch.delenv("MIOPEN_FIND_MODE", raising = False)
+ monkeypatch.delenv("MIOPEN_FIND_ENFORCE", raising = False)
+ torch = _stub_torch(monkeypatch)
+ torch.version = types.SimpleNamespace(hip = hip)
+ torch.__version__ = version
+ torch.backends.miopen = types.SimpleNamespace(immediate = False)
+ flags = snapshot_backend_flags()
+ apply_speed_optims(
+ _Pipe(with_compile = True), _target(), is_gguf = False, family = _family(), speed_mode = speed_mode
+ )
+ assert torch.backends.miopen.immediate is True
+ restore_backend_flags(flags)
+ assert torch.backends.miopen.immediate is True
+
+
+ "hip, version, env",
+ [
+ (None, "2.11.0+cu130", {}),
+ ("7.13.0", "2.11.0+rocm7.13.0", {"MIOPEN_FIND_MODE": "NORMAL"}),
+ ("7.13.0", "2.11.0+rocm7.13.0", {"MIOPEN_FIND_ENFORCE": "SEARCH"}),
+ ],
+ ids = ["cuda_torch", "explicit_find_mode", "explicit_find_enforce"],
+)
+def test_miopen_immediate_left_alone(monkeypatch, hip, version, env):
+ for name in ("MIOPEN_FIND_MODE", "MIOPEN_FIND_ENFORCE"):
+ monkeypatch.delenv(name, raising = False)
+ for name, value in env.items():
+ monkeypatch.setenv(name, value)
+ torch = _stub_torch(monkeypatch)
+ torch.version = types.SimpleNamespace(hip = hip)
+ torch.__version__ = version
+ torch.backends.miopen = types.SimpleNamespace(immediate = False)
+ apply_speed_optims(
+ _Pipe(with_compile = True),
+ _target(),
+ is_gguf = False,
+ family = _family(),
+ speed_mode = SPEED_DEFAULT,
+ )
+ assert torch.backends.miopen.immediate is False
+
+
def test_speed_default_respects_family_cudnn_benchmark_opt_out(monkeypatch):
torch = _stub_torch(monkeypatch)
family = types.SimpleNamespace(supports_torch_compile = True, cudnn_benchmark = False)