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: 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.pyindex 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.pyindex 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, monbackend.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):