Run 35B MoE models in ~2.5 GB of RAM. Weights stream from SSD; a trained router predicts loads a step ahead. Mac, iPhone, Android, Windows supported
Python
windows: one-command benchmark submission block (fixes #139)
Arda M. Ekiz258ce7b6c278
diff --git a/windows/README.md b/windows/README.mdindex 3eda58a..e00ac25 100644--- a/windows/README.md+++ b/windows/README.md@@ -129,6 +129,27 @@ Context to keep when reading these numbers:- Same-condition A/B against unpatched vanilla llama.cpp (vendor tree with the patch bands skipped) shows the patch set costs nothing (8B: 41.5 vs 39.5 decode; 35B: 27.2 vs 27.4 — within machine drift).- Expect day-to-day drift of a few percent on this class of machine; compare like-for-like, same day, same process if you benchmark.+### 2.1 Community benchmark submission (Hardware Benchmarks #142)++To add a Windows row to the Hardware Benchmarks discussion (#142), build the+engine and download a tier as in §1, then run:++ python tools/r3_bench.py --tier 8b --submission --reporter @you++The script discards the short warmup, runs 2 timed reps (`--reps N` to+override) on the ~3k-token long-context prompt (same construction as the+Python bench's BENCH_LONG), and prints a ready-to-paste submission block —+the mean of the timed reps, with `device` / `chip` / `gpu` (the Vulkan device+from the engine log) / `ram_gb` / `os` / `engine_commit` (vendor.llama.pin)+filled in automatically. `--json` additionally prints the block as a single+JSON line for aggregation, and `benchmarks/r3/bench_<tier>_submission.json`+stores the per-rep detail.++On the memory field: `peak_active_gib` on Windows is the engine process's+peak working set during the timed reps (OS-level high-water mark) — the same+intent as the Python bench's `mlx.core.get_peak_memory()`, measured at the+process level instead of the allocator.+---## 3. Technical Detailsdiff --git a/windows/tools/r3_bench.py b/windows/tools/r3_bench.pyindex 79b4eb6..ffa1b1f 100644--- a/windows/tools/r3_bench.py+++ b/windows/tools/r3_bench.py@@ -8,9 +8,16 @@ between reps. Absolute values are only meaningful within the same machine andwindow when compared against the reference — cross-window comparison is alwayssuspect (measure on the same day / same process, or treat numbers as shape-only).+Community submission mode (Hardware Benchmarks discussion #142, issue #139):+`--submission` discards the same short warmup, then runs `reps` timed reps+(default 2) on the ~3k-token long-context prompt, averages the numbers, and+prints the ready-to-paste submission block; `--json` prints the same data as+JSON for aggregation.+Usage: python tools/r3_bench.py --tier 35b [--reps 3]+ python tools/r3_bench.py --tier 8b --submission --reporter @you"""-import argparse, json, os, re, subprocess, sys, time+import argparse, ctypes, json, os, re, subprocess, sys, timeROOT = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))@@ -44,6 +51,21 @@ def models_root(tier):PROMPT = ("混合专家模型(MoE)通过稀疏激活把多个专家网络与门控路由组合起来:" * 12)++# Long-context prompt for the submission block (~3k tokens after tokenization):+# identical construction and text to the Python bench's BENCH_LONG (see+# python/examples/bench.py), so prefill numbers on both runtimes refer to the+# same measured work. Short prompts only measure fixed overhead.+_PARA_ZH = '\u6d41\u5f0f\u63a8\u7406\u6846\u67b6\u7684\u8bbe\u8ba1\u8981\u70b9\u5728\u4e8e\u628a\u78c1\u76d8\u5e26\u5bbd\u3001\u7f13\u5b58\u5c42\u7ea7\u4e0e\u8ba1\u7b97\u6838\u5fc3\u4e09\u8005\u91cd\u53e0\u8d77\u6765\u3002\u6df7\u5408\u4e13\u5bb6\u6a21\u578b\u6bcf\u4e00\u5c42\u53ea\u5728\u5c11\u6570\u4e13\u5bb6\u4e0a\u6fc0\u6d3b\uff0c\u8def\u7531\u5668\u7ed9\u51fa\u7684\u7a00\u758f\u9009\u62e9\u6070\u597d\u4e3a\u6309\u9700\u52a0\u8f7d\u63d0\u4f9b\u4e86\u5929\u7136\u7684\u6279\u7c92\u5ea6\u3002\u82e5\u80fd\u5728\u524d\u4e00\u6b65\u7684\u9690\u72b6\u6001\u57fa\u7840\u4e0a\u9884\u5224\u4e0b\u4e00\u6b65\u7684\u4e13\u5bb6\u96c6\u5408\uff0c\u56fa\u6001\u786c\u76d8\u7684\u8bfb\u53d6\u5ef6\u8fdf\u5c31\u80fd\u88ab\u8ba1\u7b97\u5b8c\u5168\u63a9\u76d6\uff0c\u8fd9\u5bf9\u8fb9\u7f18\u8bbe\u5907\u4e0a\u7684\u5927\u6a21\u578b\u90e8\u7f72\u5177\u6709\u76f4\u63a5\u610f\u4e49\u3002'+_PARA_EN = 'The design of a streaming inference framework hinges on overlapping disk bandwidth, cache hierarchy, and compute. A mixture-of-experts layer activates only a handful of experts per token, and the router sparsity provides a natural granularity for on-demand loading. If the expert set of the next step can be predicted from the previous hidden state, SSD read latency hides behind compute entirely, which matters for large-model deployment on edge hardware.'+_LONG_HEADER = '\u8bf7\u4ed4\u7ec6\u9605\u8bfb\u4ee5\u4e0b\u6750\u6599\u5e76\u7528\u4e00\u53e5\u8bdd\u603b\u7ed3\u5176\u6838\u5fc3\u601d\u60f3\uff1a'+++def long_prompt():+ paras = [_PARA_ZH if i % 2 == 0 else _PARA_EN for i in range(40)]+ return _LONG_HEADER + " ".join(paras)++PAT = {"prompt": re.compile(r"prompt eval time =.*?,\s*([\d.]+) tokens per second"),"eval": re.compile(r"(?<!prompt )eval time =.*?,\s*([\d.]+) tokens per second"),@@ -53,23 +75,145 @@ PAT = {}-def one(gguf, lora, n, tier):+# ---------------------------------------------------------------- metadata --+def cpu_name():+ """CPU marketing name from the registry (no wmic/sysinfo dependency)."""+ try:+ import winreg+ k = winreg.OpenKey(winreg.HKEY_LOCAL_MACHINE,+ r"HARDWARE\DESCRIPTION\System\CentralProcessor\0")+ name, _ = winreg.QueryValueEx(k, "ProcessorNameString")+ winreg.CloseKey(k)+ return " ".join(str(name).split())+ except OSError:+ return None+++def total_ram_gb():+ try:+ class MEMORYSTATUSEX(ctypes.Structure):+ _fields_ = [("dwLength", ctypes.c_uint),+ ("dwMemoryLoad", ctypes.c_uint),+ ("ullTotalPhys", ctypes.c_uint64),+ ("ullAvailPhys", ctypes.c_uint64),+ ("ullTotalPageFile", ctypes.c_uint64),+ ("ullAvailPageFile", ctypes.c_uint64),+ ("ullTotalVirtual", ctypes.c_uint64),+ ("ullAvailVirtual", ctypes.c_uint64),+ ("ullAvailExtendedVirtual", ctypes.c_uint64)]+ st = MEMORYSTATUSEX()+ st.dwLength = ctypes.sizeof(st)+ if not ctypes.windll.kernel32.GlobalMemoryStatusEx(ctypes.byref(st)):+ return None+ return st.ullTotalPhys / (1024 ** 3)+ except (OSError, AttributeError):+ return None+++def windows_build():+ v = sys.getwindowsversion()+ label = "11" if v.build >= 22000 else "10"+ return f"Windows {label} (build {v.build})"+++def machine_name():+ """Computer name (device) from the registry; a human brand is not exposed by winreg."""+ try:+ import winreg+ k = winreg.OpenKey(winreg.HKEY_LOCAL_MACHINE,+ r"SYSTEM\CurrentControlSet\Control\ComputerName\ComputerName")+ name, _ = winreg.QueryValueEx(k, "ComputerName")+ winreg.CloseKey(k)+ return str(name)+ except OSError:+ return None+++def engine_commit():+ """Head of vendor.llama.pin at the repo root (the pinned llama.cpp fork)."""+ pin = os.path.join(ROOT, "..", "vendor.llama.pin")+ try:+ with open(pin, encoding="utf-8") as f:+ return f.read().splitlines()[0].strip()+ except OSError:+ return None+++def vulkan_device(err):+ """Vulkan device name reported by the engine's own startup log."""+ m = re.search(r"Device\s*\d*\s*[:=]\s*(.+?)\s*(?:\(|\n|$)", err, re.I)+ if m and re.search(r"vulkan", err[:m.start()][-400:], re.I):+ return m.group(1).strip()+ for line in err.splitlines():+ if re.search(r"vulkan", line, re.I):+ m2 = re.search(r"(?:device|adapter|gpu)[^\n:]*[:=]\s*(.+)", line, re.I)+ if m2:+ return m2.group(1).strip()+ return None+++def open_handle(pid):+ """Engine-process handle for the memory probe (Windows only; None otherwise)."""+ if sys.platform != "win32":+ return None+ for _ in range(10): # the child may still be spawning+ h = ctypes.windll.kernel32.OpenProcess(0x0410, False, pid) # QUERY_INFORMATION|VM_READ+ if h:+ return h+ time.sleep(0.01)+ return None+++def peak_working_set_gib(handle):+ """Peak process working set (GiB) — the OS keeps the high-water mark while+ any handle to the process is open, so probe after communicate(), then close."""+ if not handle:+ return None+ try:+ class PMC(ctypes.Structure):+ _fields_ = [("cb", ctypes.c_uint),+ ("PageFaultCount", ctypes.c_uint),+ ("PeakWorkingSetSize", ctypes.c_size_t),+ ("WorkingSetSize", ctypes.c_size_t),+ ("QuotaPeakPagedPoolUsage", ctypes.c_size_t),+ ("QuotaPagedPoolUsage", ctypes.c_size_t),+ ("QuotaPeakNonPagedPoolUsage", ctypes.c_size_t),+ ("QuotaNonPagedPoolUsage", ctypes.c_size_t),+ ("PagefileUsage", ctypes.c_size_t),+ ("PeakPagefileUsage", ctypes.c_size_t)]+ pmc = PMC()+ pmc.cb = ctypes.sizeof(pmc)+ if not ctypes.WinDLL("psapi").GetProcessMemoryInfo(handle, ctypes.byref(pmc), pmc.cb):+ return None+ return pmc.PeakWorkingSetSize / (1024 ** 3)+ except OSError:+ return None+++def one(gguf, lora, n, tier, prompt=None):cmd = [os.path.join(SERVER_BIN, "llama-completion.exe"), "-m", gguf,"-ngl", "99", "-cmoe", "--temp", "0", "-n", str(n), "--seed", "7",- "-no-cnv", "--no-display-prompt", "-p", PROMPT]+ "-no-cnv", "--no-display-prompt", "-p", prompt or PROMPT]if lora:cmd[3:3] = ["--lora", lora]t0 = time.time()- r = subprocess.run(cmd, capture_output=True, text=True, encoding="utf8", errors="replace")- err = r.stderr or ""- out = {"wall_s": round(time.time() - t0, 1)}+ p = subprocess.Popen(cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE,+ text=True, encoding="utf8", errors="replace")+ h = open_handle(p.pid)+ _, err = p.communicate()+ peak = peak_working_set_gib(h) # probed after wait while the handle is live+ if h:+ ctypes.windll.kernel32.CloseHandle(h)+ out = {"wall_s": round(time.time() - t0, 1),+ "peak_working_set_gib": round(peak, 2) if peak is not None else None,+ "gpu": vulkan_device(err or "")}for k, pat in PAT.items():- m = pat.search(err)+ m = pat.search(err or "")if m:out[k] = float(m.group(1))- ev = re.search(r"n_ctx = (\d+)", err)+ ev = re.search(r"n_ctx = (\d+)", err or "")out["n_ctx"] = int(ev.group(1)) if ev else None- if "error" in err.lower() and "eval time" not in err:+ if "error" in (err or "").lower() and "eval time" not in err:out["error"] = [l for l in err.splitlines() if "error" in l.lower()][:2]if "total_ms" in out and "prompt_ms" in out:dec_ms = out["total_ms"] - out["prompt_ms"] # total = generation end-to-end, excludes load (verified)@@ -77,11 +221,50 @@ def one(gguf, lora, n, tier):return out+def submission_block(tier, reporter, notes, rows):+ """Ready-to-paste Hardware Benchmarks block; numbers are the mean of the timed reps."""+ evs = [r["eval"] for r in rows if "eval" in r]+ prs = [r["prompt"] for r in rows if "prompt" in r]+ peaks = [r["peak_working_set_gib"] for r in rows if r.get("peak_working_set_gib") is not None]+ gpus = sorted({r["gpu"] for r in rows if r.get("gpu")})+ dev, chip = machine_name(), cpu_name()+ ram = total_ram_gb()+ data = {+ "tier": f"edge0-{tier}",+ "runtime": "windows",+ "device": dev or "(unknown)",+ "chip": chip or "(unknown)",+ "gpu": gpus[0] if gpus else "(not detected)",+ "ram_gb": round(ram) if ram else None,+ "os": windows_build(),+ "decode_tok_s": round(sum(evs) / len(evs), 2) if evs else None,+ "prefill_tok_s": round(sum(prs) / len(prs), 2) if prs else None,+ "peak_active_gib": round(max(peaks), 2) if peaks else None,+ "engine_commit": engine_commit(),+ "reporter": reporter or "",+ "notes": notes or "",+ }+ return data+++def print_block(d):+ for k, v in d.items():+ print(f"{k:<16} {'' if v is None else v}".rstrip())++def main():ap = argparse.ArgumentParser()ap.add_argument("--tier", choices=["35b", "8b"], required=True)- ap.add_argument("--reps", type=int, default=3)+ ap.add_argument("--reps", type=int, default=None,+ help="timed reps (default: 2 with --submission, 3 otherwise)")+ ap.add_argument("--submission", action="store_true",+ help="print the Hardware Benchmarks submission block (long-context prefill)")+ ap.add_argument("--json", action="store_true",+ help="with --submission, also print the block as JSON")+ ap.add_argument("--reporter", default="", help="@handle for the submission block")+ ap.add_argument("--notes", default="", help="free-form notes field")a = ap.parse_args()+ reps = a.reps or (2 if a.submission else 3)exe = os.path.join(SERVER_BIN, "llama-completion.exe")if not os.path.isfile(exe):raise SystemExit(f"llama-completion.exe not found in {SERVER_BIN} "@@ -90,10 +273,30 @@ def main():gguf = os.path.join(md, f"edge0-{a.tier}-gguf", f"edge0-{a.tier}.gguf")lora = os.path.join(md, f"edge0-{a.tier}",f"lora_edge0_{a.tier}-gguf.gguf")+ if a.submission:+ warm = one(gguf, lora, 64, a.tier) # warmup (page cache + mmap fault-in), discarded+ print(f"warm(discard): {warm}", flush=True)+ longp = long_prompt()+ rows = []+ for i in range(reps):+ time.sleep(5)+ m = one(gguf, lora, 256, a.tier, prompt=longp)+ m["rep"] = i+ rows.append(m)+ print(f"rep{i}: {m}", flush=True)+ data = submission_block(a.tier, a.reporter, a.notes, rows)+ outp = os.path.join(ROOT, "benchmarks", "r3", f"bench_{a.tier}_submission.json")+ os.makedirs(os.path.dirname(outp), exist_ok=True)+ with open(outp, "w", encoding="utf8") as f:+ json.dump({**data, "timed_reps": rows}, f, ensure_ascii=False, indent=1)+ print_block(data)+ if a.json:+ print("SUBMIT_JSON " + json.dumps(data, ensure_ascii=False))+ returnrows = []warm = one(gguf, lora, 64, a.tier) # warmup (page cache + mmap fault-in), discardedprint(f"warm(discard): {warm}", flush=True)- for i in range(a.reps):+ for i in range(reps):time.sleep(5)m = one(gguf, lora, 256, a.tier)m["rep"] = i