"""Stage: LLM through Strata (install, start, measure, restart timings). Options (--opt key=value): model=fast|recommended which preset to measure (default fast) url=http://.../v1 measure an already running server instead (no Strata management) context=32768 context size to start Strata with only=tools only re-measure tool calling, before and after the Jarvis fix only=levels the fast model in each placement the app's automatic tuning switches between (normal, and less VRAM for a game) """ from __future__ import annotations import json import os import re import shutil import socket import subprocess import sys import time from pathlib import Path sys.path.insert(0, str(Path(__file__).resolve().parent.parent)) from m0lib.common import (IS_WINDOWS, GpuSampler, StageResult, gpu_memory_mib, run_logged, say, # noqa: E402 scrub, stage_args) from m0lib.llm_bench import (bench_single, quick_bench, test_decode, test_images, test_parallel, # noqa: E402 test_tools, test_vram_shrink, warmup) from m0lib.tool_fix import test_tools_fixed, tools_passed # noqa: E402 from m0lib.llm_client import ChatClient, http_json # noqa: E402 PRESETS = { "fast": {"family": "qwen36", "model": "ud-q4_k_xl", "label": "Qwen3.6-35B-A3B UD-Q4_K_XL"}, # Strata 0.1.59 gives Q2_0 the DFlash2 drafter alone, which needs a native pack; the canonical # Q2_0 pack is not one, so the engine stops at start. The MTP drafter (or none) starts. "recommended": {"family": "qwen", "model": "q2_0", "label": "Qwen3.8-Flash-Next Q2_0 (ISTA-DASLab)", "drafting": "mtp"}, } # Tried in order when a preset still gets a tool call wrong with the Jarvis fix # (the user's rule: if the fix does not work, change the model). Same family, # less compressed weights, all within 128 GB RAM. need_gb: free disk to download. CANDIDATES = { "fast": [ {"family": "qwen36", "model": "ud-q6_k_xl", "label": "Qwen3.6-35B-A3B UD-Q6_K_XL", "need_gb": 40}, {"family": "qwen36", "model": "q8_0", "label": "Qwen3.6-35B-A3B Q8_0", "need_gb": 45}, ], "recommended": [ {"family": "qwen", "model": "iq3_xxs", "label": "Qwen3.8-Flash-Next IQ3_XXS (ISTA-DASLab)", "drafting": "mtp", "need_gb": 80}, {"family": "qwen", "model": "iq3_s", "label": "Qwen3.8-Flash-Next IQ3_S (ISTA-DASLab)", "drafting": "mtp", "need_gb": 85}, ], } PORT = 8080 # The placements the app's automatic tuning moves the one model between. # Strata's qwen36 engine cannot give VRAM back while it runs (POST /v1/vram # answers 400), so each placement is a restart with --expert-vram (GiB for the # routed experts; the rest stay in RAM) and a matching prompt chunk. A game # placement also reads pictures on the CPU, answers one request at a time and # leaves the game half the CPU (8 engine workers cost nothing on the 9950X3D). # The first game PC run (2026-11-10) showed that the expert budget is not the # floor: at 2 GB the model still held 10.8 GB and a 16 GB card kept only # 4.3 GB free. The rest is 4.4 GiB of weights that never leave the card, the # conversation memory (1.43 GiB per 32K in int8; q4_0 is half), the MTP # drafter (about 1 GB) and buffers, so the smaller placements also shrink those. # The last row starts the normal placement again: what returning after a game costs. GAME_RESERVE_MIB = 8000 _GAME = {"parallel": 1, "vision": "cpu"} _W8 = "pool-workers=8" LEVELS = [ {"key": "normal", "parallel": 2, "vision": "gpu", "sets": []}, {"key": "game_experts_2g", **_GAME, "sets": ["expert-vram=2", "prefill=2048", _W8]}, {"key": "game_experts_1g_kv_q4_16k", **_GAME, "context": 16384, "kv": "q4_0", "sets": ["expert-vram=1", "prefill=2048", _W8]}, {"key": "game_experts_1g_kv_q4_16k_no_mtp", **_GAME, "context": 16384, "kv": "q4_0", "mtp": "off", "sets": ["expert-vram=1", "prefill=2048", _W8]}, {"key": "game_experts_0g_kv_q4_16k_no_mtp", **_GAME, "context": 16384, "kv": "q4_0", "mtp": "off", "sets": ["expert-vram=0", "prefill=2048", _W8]}, {"key": "game_experts_0g_kv_q4_8k_no_mtp", **_GAME, "context": 8192, "kv": "q4_0", "mtp": "off", "sets": ["expert-vram=0", "prefill=2048", _W8]}, {"key": "normal_again", "parallel": 2, "vision": "gpu", "sets": []}, ] INSTALL_URL = "https://github.com/Niko1221/Strata/releases/latest/download/install.ps1" class Strata: def __init__(self, work: Path, log: Path): self.home = work / "strata" self.models = work / "models" / "llm" self.log = log # Keep Strata's engine, state and models inside the work folder: the # installer honours STRATA_DATA_DIR, the engine components # STRATA_COMPONENTS_DIR, and the per-user state follows LOCALAPPDATA. self.env = dict(os.environ, STRATA_DATA_DIR=str(self.home), STRATA_NO_PATH="1", STRATA_HOME=str(self.home / "home"), STRATA_COMPONENTS_DIR=str(self.home / "components"), PYTHONUTF8="1") if IS_WINDOWS: self.env["LOCALAPPDATA"] = str(self.home / "localappdata") (self.home / "localappdata").mkdir(parents=True, exist_ok=True) self._flags: set[str] | None = None @property def exe(self) -> Path: return self.home / "terminal" / "current" / "bin" / ("strata.cmd" if IS_WINDOWS else "strata") def install(self) -> bool: if self.exe.exists(): return True self.home.mkdir(parents=True, exist_ok=True) say(" Strata 터미널 설치 중 (작업 폴더 안에만 설치하고 PATH는 바꾸지 않습니다)") if IS_WINDOWS: cmd = ["powershell", "-NoProfile", "-ExecutionPolicy", "Bypass", "-Command", f"irm {INSTALL_URL} | iex"] else: cmd = ["sh", "-c", f"curl -fsSL {INSTALL_URL[:-3]}sh -o install.sh && sh install.sh"] rc = run_logged(cmd, self.log, timeout=1800, env=self.env, cwd=str(self.home), echo=True) return rc == 0 and self.exe.exists() def cmd(self, *args: str) -> list[str]: return [str(self.exe), *args] def supported_flags(self) -> set[str]: if self._flags is None: p = subprocess.run(self.cmd("launch", "--help"), capture_output=True, text=True, env=self.env, encoding="utf-8", errors="replace", timeout=120) with open(self.log, "a", encoding="utf-8") as f: f.write(scrub(p.stdout + p.stderr)) self._flags = set(re.findall(r"--[a-z0-9][a-z0-9-]*", p.stdout + p.stderr)) return self._flags def launch_args(self, preset: dict, context: int, parallel: int, vision: str, extra_sets: list[str], restart: bool, kv: str | None = None, mtp: str | None = None) -> tuple[list[str], list[str]]: want = [("--family", preset["family"]), ("--model", preset["model"]), ("--data-dir", str(self.models)), ("--context", str(context)), ("--vision", vision), ("--parallel", str(parallel)), ("--access", "local"), ("--port", str(PORT)), ("--no-api-key", None), ("--yes", None), ("--no-tui", None)] if preset.get("drafting"): want.append(("--drafting", preset["drafting"])) if kv: want.append(("--kv", kv)) if mtp: want.append(("--mtp", mtp)) if restart: want.append(("--restart", None)) for s in extra_sets: want.append(("--set", s)) flags = self.supported_flags() args, dropped = [], [] for k, v in want: if flags and k not in flags: dropped.append(k) continue args += [k] + ([v] if v is not None else []) return args, dropped def plan(self, preset: dict, context: int) -> object: args, _ = self.launch_args(preset, context, 1, "on", [], False) args = [a for a in args if a not in ("--yes", "--no-tui")] + ["--plan", "--output", "json"] p = subprocess.run(self.cmd("launch", *args), capture_output=True, text=True, env=self.env, encoding="utf-8", errors="replace", timeout=600) with open(self.log, "a", encoding="utf-8") as f: f.write(scrub(p.stdout + p.stderr)) try: return json.loads(p.stdout) except ValueError: return {"exit": p.returncode, "text": scrub(p.stdout[-3000:] + p.stderr[-2000:])} def launch(self, preset: dict, context: int, parallel: int, vision: str = "on", extra_sets: list[str] | None = None, restart: bool = False, timeout: float = 4 * 3600, kv: str | None = None, mtp: str | None = None) -> dict: args, dropped = self.launch_args(preset, context, parallel, vision, extra_sets or [], restart, kv, mtp) last = [0.0] def progress(line: str) -> bool: # Downloads print many progress lines; show one every 15 s. if "%" in line: if time.monotonic() - last[0] < 15: return False last[0] = time.monotonic() return True self.mark_engine_logs() t0 = time.perf_counter() rc = run_logged(self.cmd("launch", *args), self.log, timeout=timeout, env=self.env, echo=True, echo_filter=progress) t_cmd = time.perf_counter() - t0 # Only a successful launch counts: after a failed one, whatever still # answers on the port is not the model that was asked for. ready = wait_ready(f"http://127.0.0.1:{PORT}", timeout=1800) if rc == 0 else None if ready is None: self.keep_engine_log() return {"args": scrub(" ".join(args)), "dropped_flags": dropped, "exit": rc, "command_seconds": round(t_cmd, 1), "ready": ready is not None, "ready_seconds": round(time.perf_counter() - t0, 1) if ready is not None else None} def keep_engine_log(self) -> None: """Copy the newest engine log (scrubbed, its end) next to our log so the results zip carries it: Strata's one-line reason can name a side effect, not the cause.""" found = sorted(self.home.rglob("*.engine.log"), key=lambda p: p.stat().st_mtime) if not found: return src = found[-1] text = src.read_text(encoding="utf-8", errors="replace")[-200_000:] dest = self.log.with_name(f"{self.log.stem}_engine_{src.name.removesuffix('.engine.log')}.log") dest.write_text(scrub(text), encoding="utf-8") def mark_engine_logs(self) -> None: """Remember how long each engine log is, so engine_lines() reads only what the next start writes.""" self._log_marks = {p: p.stat().st_size for p in self.home.rglob("*.engine.log")} def engine_lines(self, prefix: str = "strata-q36:", limit: int = 20) -> list[str]: """The placement lines (how much went to VRAM and RAM) the newest engine log got since the last start, scrubbed. The log is appended across starts; one written again is read whole.""" found = sorted(self.home.rglob("*.engine.log"), key=lambda p: p.stat().st_mtime) if not found: return [] src = found[-1] start = getattr(self, "_log_marks", {}).get(src, 0) if src.stat().st_size < start: start = 0 with open(src, "rb") as f: f.seek(start) text = f.read()[-400_000:].decode("utf-8", errors="replace") return [scrub(x.strip())[:400] for x in text.splitlines() if x.lstrip().startswith(prefix)][:limit] def stop(self) -> None: run_logged(self.cmd("launch", "--stop"), self.log, timeout=300, env=self.env) time.sleep(3) def status(self) -> object: p = subprocess.run(self.cmd("launch", "--status", "--output", "json"), capture_output=True, text=True, env=self.env, encoding="utf-8", errors="replace", timeout=120) try: return json.loads(p.stdout) except ValueError: return scrub(p.stdout[-2000:]) def port_in_use(port: int) -> bool: with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as sock: sock.settimeout(2) return sock.connect_ex(("127.0.0.1", port)) == 0 def wait_ready(root: str, timeout: float) -> float | None: t0 = time.perf_counter() while time.perf_counter() - t0 < timeout: try: code, body = http_json("GET", root + "/v1/models", timeout=5) if code == 200 and isinstance(body, dict) and body.get("data"): return time.perf_counter() - t0 except Exception: pass time.sleep(2) return None def server_info(base: str) -> dict: info = {} for path in ("/models", "/status"): try: code, body = http_json("GET", base + path, timeout=10) info[path] = body if code == 200 else {"http": code} except Exception as e: info[path] = {"error": str(e)} return info def tools_only(r: StageResult, c: ChatClient) -> dict | None: """Tool calling as first measured (baseline) and with the Jarvis fix. Returns the fixed run's summary (None when it could not run).""" summary = None for name, fn in (("warmup", lambda: warmup(r, c)), ("tools", lambda: test_tools(r, c)), ("fixed", lambda: test_tools_fixed(r, c))): try: out = fn() if name == "fixed": summary = out except Exception as e: r.error("tools", e) r.save() return summary def try_candidates(r: StageResult, which: str, default_summary: dict | None, measure, free_gb) -> None: """If the default preset fails with the fix, measure the next candidates. ``measure(preset)`` starts the preset and returns its fixed-run summary; ``free_gb()`` is the free disk in GB. Stops at the first passing model and records which one Jarvis should switch to (None if none passed). """ verdict: dict = {"default": PRESETS[which]["label"], "default_passed": tools_passed(default_summary), "candidates": [], "chosen": PRESETS[which]["label"] if tools_passed(default_summary) else None} r["tools_verdict"] = verdict if verdict["default_passed"]: return for cand in CANDIDATES[which]: row: dict = {"label": cand["label"]} verdict["candidates"].append(row) if free_gb() < cand["need_gb"]: row["skipped"] = "disk" say(f" {cand['label']}: 디스크 공간이 부족해 건너뜁니다 (필요 약 {cand['need_gb']} GB)") continue say(f" 수정 후에도 틀린 도구 호출이 있어 다른 모델을 잽니다: {cand['label']} (처음이면 내려받습니다)") try: row["summary"] = measure(cand) except Exception as e: row["error"] = str(e)[:300] r.save() if tools_passed(row.get("summary")): verdict["chosen"] = cand["label"] return def measure_levels(r: StageResult, s: Strata, preset: dict, context: int, levels: list[dict] = LEVELS) -> None: """Start the model in each placement and measure what the switch and the placement cost.""" base = f"http://127.0.0.1:{PORT}/v1" rows: list[dict] = [] r["levels"] = rows for lv in levels: row: dict = {"key": lv["key"], "parallel": lv["parallel"], "vision": lv["vision"], "sets": lv["sets"], "context": lv.get("context", context), "kv": lv.get("kv"), "mtp": lv.get("mtp")} rows.append(row) say(f" - 배치 {lv['key']}: 다시 시작해서 재는 중") t0 = time.perf_counter() s.stop() idle = gpu_memory_mib() row["vram_idle_mib"] = idle[0] if idle else None with GpuSampler(interval=1) as g: row["launch"] = s.launch(preset, row["context"], parallel=lv["parallel"], vision=lv["vision"], extra_sets=lv["sets"], kv=lv.get("kv"), mtp=lv.get("mtp"), timeout=3600) ready = row["launch"].get("ready") row["switch_s"] = round(time.perf_counter() - t0, 1) if ready else None row["engine"] = s.engine_lines() if ready: try: c = ChatClient(base) if "server" not in r.data: r["server"] = {"model": c.model, "info": server_info(base)} c.chat([{"role": "user", "content": "안녕"}], max_tokens=8) # loading is not answer time row.update(quick_bench(c)) for err in row["errors"]: r.error(f"level {lv['key']}", err) except Exception as e: r.error(f"level {lv['key']}", e) else: r.error(f"level {lv['key']}", "the placement did not start") now = gpu_memory_mib() if ready else None # the sampler can miss a short run peak = max(x for x in (g.peak, now[0] if now else None) if x is not None) \ if g.peak is not None or now else None if now: row["vram_after_mib"] = now[0] # settled; the peak (during the start) is what a running game must survive if peak is not None and idle: row["vram_peak_mib"] = peak row["vram_free_mib"] = idle[1] - peak say(f" 전환 {row['switch_s']}s, 출력 {(row.get('decode_tok_s') or {}).get('median')} tok/s, " f"남는 VRAM {row.get('vram_free_mib')} MiB") r.save() s.stop() r["game_choice"] = pick_game_level(rows) measured = [x for x in rows if x["key"].startswith("game_") and x.get("vram_free_mib") is not None] if r["game_choice"] is None and measured: most = max(measured, key=lambda x: x["vram_free_mib"]) r.note(f"no game placement leaves {GAME_RESERVE_MIB} MiB free; the most free is " f"{most['key']}: {most['vram_free_mib']} MiB") def pick_game_level(rows: list[dict], reserve_mib: int = GAME_RESERVE_MIB) -> dict | None: """The game placement the app would pick on this PC: the fastest one that leaves the game its reserve of VRAM. None when no game placement does.""" fits = [x for x in rows if x["key"].startswith("game_") and (x.get("vram_free_mib") or 0) >= reserve_mib and (x.get("decode_tok_s") or {}).get("median")] if not fits: return None best = max(fits, key=lambda x: x["decode_tok_s"]["median"]) return {"key": best["key"], "reserve_mib": reserve_mib, "vram_free_mib": best["vram_free_mib"], "decode_tok_s": best["decode_tok_s"]["median"], "switch_s": best.get("switch_s"), "voice_first_sentence": (best.get("voice_first_sentence") or {}).get("median")} def measure_external(r: StageResult, url: str, context: int, only: str = "") -> None: c = ChatClient(url) r["server"] = {"url": url, "model": c.model, "info": server_info(c.base)} if only == "tools": tools_only(r, c) return if only == "levels": r.note("placements need Strata; not measured against an external server") return with GpuSampler() as g: bench_single(r, c, context) test_parallel(r, c) r["gpu"] = g.summary() def measure_strata(r: StageResult, work: Path, which: str, context: int, only: str = "") -> None: preset = PRESETS[which] r["preset"] = dict(preset, key=which, context=context) s = Strata(work, work / "logs" / f"strata_{which}.log") if not s.install(): r.error("install", "Strata 설치 실패 (logs/strata_*.log 참고)") return r["plan"] = s.plan(preset, context) r.save() if only == "levels": s.stop() if port_in_use(PORT): r.error("port", f"다른 프로그램이 {PORT} 포트를 쓰고 있습니다. 그 프로그램을 끄고 다시 실행해 주세요.") return measure_levels(r, s, preset, context) return # 1) Start (the first start downloads the model and engine). say(f" {preset['label']} 시작 중. 처음에는 모델을 내려받으므로 오래 걸립니다.") s.stop() if port_in_use(PORT): # Another program owns the port; measuring it would report the wrong model. r.error("port", f"다른 프로그램이 {PORT} 포트를 쓰고 있습니다. 그 프로그램을 끄고 다시 실행해 주세요.") return with GpuSampler(interval=2) as g: r["launch_cold"] = s.launch(preset, context, parallel=1) r["launch_cold"].update(g.summary()) r.save() if not r["launch_cold"]["ready"]: r.error("launch", "서버가 준비되지 않았습니다") r["status_after_fail"] = s.status() return base = f"http://127.0.0.1:{PORT}/v1" c = ChatClient(base) r["server"] = {"model": c.model, "info": server_info(base), "status": s.status()} if only == "tools": default_summary = tools_only(r, c) def measure(cand: dict) -> dict | None: key = f"tools_fixed_{cand['model']}" res = s.launch(cand, context, parallel=1, restart=True) r[f"launch_{cand['model']}"] = res if not res["ready"]: return None return test_tools_fixed(r, ChatClient(base), key=key) try_candidates(r, which, default_summary, measure, free_gb=lambda: shutil.disk_usage(work).free / 1e9) s.stop() return with GpuSampler() as g: bench_single(r, c, context) r["gpu_single"] = g.summary() r.save() # 2) Restart with the files already in RAM/page cache: what a game-mode switch costs. # Also: two slots, the elastic VRAM cache and the GPU picture reader. say(" 다시 시작 시간 측정 (동시 2개, VRAM 조절 켜기, 이미지 GPU 처리)") r["launch_restart"] = s.launch(preset, context, parallel=2, vision="gpu", extra_sets=["vram-elastic"], restart=True, timeout=3600) r.save() if r["launch_restart"]["ready"]: c = ChatClient(base) r["server_restart"] = {"status": s.status()} for fn in (lambda: warmup(r, c), lambda: test_decode(r, c, key="decode_parallel2"), lambda: test_parallel(r, c), lambda: test_images(r, c, key="images_gpu_reader"), lambda: test_vram_shrink(r, c)): try: fn() except Exception as e: r.error("restart tests", e) r.save() # 3) Game mode (b): switching from the recommended model to the fast one. if which == "recommended" and (work / "results" / "llm_fast.json").exists(): say(" 게임 모드 (b): 권장 모델 -> 빠름 모델 전환 시간") r["switch_to_fast"] = s.launch(PRESETS["fast"], context, parallel=1, restart=True, timeout=3600) s.stop() def main() -> int: out, work, opts = stage_args() which = opts.get("model", "fast") context = int(opts.get("context", "32768")) only = opts.get("only", "") name = {"tools": f"tools_{which}", "levels": "gamemode"}.get(only, f"llm_{which}") r = StageResult(name, out) try: if opts.get("url"): measure_external(r, opts["url"], context, only) else: measure_strata(r, work, which, context, only) except Exception as e: r.error("llm", e) # No server reached means nothing was measured: "failed" so the next run retries it. if "server" not in r.data: r.save("failed") else: r.save("ok" if not r["errors"] else "partial") return 0 if __name__ == "__main__": sys.exit(main())