Skip to content

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++

launch: on a Mac, use prepared model files with the Metal defaults (the Terminal has no Mac model installer)

Niko122108297223e4c7
diff --git a/strata_cli/launch/adopt.py b/strata_cli/launch/adopt.py
index 76a8257ef..0960f0a02 100644
--- a/strata_cli/launch/adopt.py
+++ b/strata_cli/launch/adopt.py
@@ -21,7 +21,8 @@ UNSUPPORTED = ("no nvidia gpu found", "no amd gpu", "nvidia gpu", "not supported
def setup_cannot_handle_this_card(message: str) -> bool:
low = (message or "").lower()
- return "no nvidia gpu found" in low or "no amd gpu strata can use" in low or "nvidia-smi did not answer" in low
+ return ("no nvidia gpu found" in low or "no amd gpu strata can use" in low or "nvidia-smi did not answer" in low
+ or "no mac model installer" in low)
def find_shards(folder: Path) -> list[Path]:
@@ -61,12 +62,19 @@ def plan(*, family: str, model: str, model_dir: Path, data_dir: Path, src: Path,
args = ["--pack", str(pack), "--native", str(shards[0])]
if len(shards) > 1:
args += ["--ple-gguf", str(shards[-1])]
- args += ["--expert-profile", str(profile), "--expert-cache", "auto", "--prefill", "auto", "--spec", "4", "--spec-min-p", "0.5"]
- if mtp.exists():
- args += ["--mtp", str(mtp)]
- args += ["--max-context", str(context), "--kv", kv or ("int8" if context > 8192 else "fp16")]
+ if backend == "metal":
+ # a Mac: the Metal defaults metal/setup_mac.py writes for a model that streams from the SSD (one token per step, no draft layer, the GGUF mapped in place)
+ args += ["--expert-profile", str(profile), "--expert-cache", "64", "--prefill", "auto", "--spec", "2", "--suffix-draft", "0",
+ "--max-context", str(context), "--mmap-experts"]
+ else:
+ args += ["--expert-profile", str(profile), "--expert-cache", "auto", "--prefill", "auto", "--spec", "4", "--spec-min-p", "0.5"]
+ if mtp.exists():
+ args += ["--mtp", str(mtp)]
+ args += ["--max-context", str(context), "--kv", kv or ("int8" if context > 8192 else "fp16")]
cfg = {"exe": "component:strata-engine", "args": args, "cwd": str(src), "tokenizer": str(tokenizer), "model_name": model_name,
"log": str(log), "lib_dirs": [], "port": 8080}
+ if backend == "metal":
+ cfg["env"] = {"STRATA_METAL_STREAM": "1"}
if backend:
cfg["backend"] = backend
return cfg, []
diff --git a/strata_cli/launch/runner.py b/strata_cli/launch/runner.py
index 6b172cdee..428c031eb 100644
--- a/strata_cli/launch/runner.py
+++ b/strata_cli/launch/runner.py
@@ -369,6 +369,11 @@ class Runner:
self.emit("model", f"{s.family} {s.model} is already installed with these choices.", 100)
return path
title = f"{s.family} {s.model}"
+ if sys.platform == "darwin" and not (cat.src / "metal" / "setup_mac.py").is_file():
+ # this Terminal does not carry the Mac installer (it builds the engine from source): use model files prepared on this Mac
+ adopted = self.adopt_prepared_files("no Mac model installer in this Terminal")
+ if adopted is not None:
+ return adopted
self.emit("model", f"Getting {title} ready (a download may be large; it continues where it stopped if interrupted)…", 0)
argv = [self.python, str(cat.src / "setup.py"), *self.setup_args()]
env = self._env()