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: AUTO W8A8 on ROCm RDNA keeps bf16 when the int8 plan still offloads (#17104)

Daniel Handd2c39d3e4c0
diff --git a/studio/backend/core/inference/diffusion.py b/studio/backend/core/inference/diffusion.py
index 1bfe3ced321..bafa6cf6884 100644
--- a/studio/backend/core/inference/diffusion.py
+++ b/studio/backend/core/inference/diffusion.py
@@ -10312,7 +10312,15 @@ class DiffusionBackend:
plan.offload_policy,
)
inplace_decline = None
- if not keeps_resident and native_scheme is None:
+ if rdna_auto_w8a8 is not None and not keeps_resident:
+ # Measured resident only: AUTO keeps bf16 when even the int8 plan offloads.
+ native_scheme = rdna_auto_w8a8 = None
+ plan = bf16_plan
+ inplace_decline = (
+ f"auto int8 W8A8 on RDNA is resident-only; the '{plan.offload_policy}' "
+ "offload keeps the bf16 transformer"
+ )
+ elif not keeps_resident and native_scheme is None:
plan, inplace_decline = _inplace_torchao_placement(
plan,
preview_scheme,
diff --git a/studio/backend/tests/test_diffusion_backend.py b/studio/backend/tests/test_diffusion_backend.py
index 5c00d6dcfb5..9c849b81090 100644
--- a/studio/backend/tests/test_diffusion_backend.py
+++ b/studio/backend/tests/test_diffusion_backend.py
@@ -13406,6 +13406,26 @@ def test_auto_on_rdna_honours_an_explicit_speed_tier(fake_runtime, tmp_path, mon
backend.unload()
+def test_auto_on_rdna_stays_bf16_when_the_int8_plan_still_offloads(
+ fake_runtime, tmp_path, monkeypatch
+):
+ from core.inference import diffusion as dmod
+ from core.inference.diffusion_memory import DeviceMemory
+
+ backend = DiffusionBackend()
+ calls = _stub_rdna_w8a8_host(backend, monkeypatch)
+ memory = DeviceMemory("rocm", "cuda", "discrete_vram", 16 * 1024, 16 * 1024)
+ monkeypatch.setattr(dmod, "settled_snapshot_device_memory", lambda *a, **kw: memory)
+ monkeypatch.setattr(dmod, "snapshot_device_memory", lambda *a, **kw: memory)
+ status = backend.load_pipeline(
+ "Qwen/Qwen-Image-2512", model_kind = "pipeline", _base_local_dir = str(tmp_path)
+ )
+ assert calls == []
+ assert status["transformer_quant"] is None
+ assert status["offload_policy"] != "none"
+ backend.unload()
+
+
def test_an_explicit_scheme_on_nvidia_keeps_the_torchao_path_and_its_wording(
fake_runtime, tmp_path, monkeypatch
):