Qwen3.8-Flash-Next on any consumer hardware: one-click install for Windows / Linux. Strata inference engine, OpenAI/Anthropic API on localhost, optional image input.
C++
tuner: AMD architecture from the kernel for the never list
Niko1221c0616f06be2c
diff --git a/strata_control/autotune.py b/strata_control/autotune.pyindex 20a712fb2..bcee5982e 100644--- a/strata_control/autotune.py+++ b/strata_control/autotune.py@@ -29,7 +29,7 @@ import timefrom dataclasses import dataclass, fieldfrom typing import Any, Callable-from strata_control import tune_knobs, tune_machine+from strata_control import engine_component, tune_knobs, tune_machinefrom strata_control.tune_session import RequestSet, RunResult, Session, StartFailedMIN_FREE_RAM_MIB = 3584 # 3.5 GiB left for the system at every start (6.5)@@ -741,6 +741,12 @@ def context_from_config(cfg: dict[str, Any], *, engine_backend: str | None, optsbackend = engine_backend or ("cuda" if any(g.get("vendor") == "nvidia" for g in cards) else"hip" if any(g.get("vendor") == "amd" for g in cards) else "cpu")arch = None+ if backend == "hip":+ try:+ archs = engine_component.hip_archs_present()+ arch = archs[0] if archs else None+ except Exception: # noqa: BLE001+ arch = Nonereturn tune_knobs.Context(backend=backend, family=family or detect_family(args), gpu_count=max(1, len(used)),vram_mib=min((g["vram_total_mib"] for g in used if g.get("vram_total_mib")), default=0),