quantprobe 1.4.0__tar.gz → 1.5.1__tar.gz

{quantprobe-1.4.0/quantprobe.egg-info → quantprobe-1.5.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: quantprobe
-Version: 1.4.0
+Version: 1.5.1
 Summary: Probe-then-quantize for LLMs: measure a model's fragility curve, plan bit/tier placement by the tiered decode law, and emit ready-to-run llama.cpp recipes.
 Author: Federico Sciuca
 License: MIT
@@ -85,15 +85,18 @@ Full statements, each with its establishing measurement and a falsifiable predic
 <p align="center"><img src="weights/data/x_chart_I_kvdepth.png" width="700" alt="Decode speed versus context depth: measured 20.02 tok/s at depth zero falling to 16.12 at 16384, on the Law 4 v2 curve with eta_kv 0.70"></p>
 
 
-## Install — and the ten commands
+## Install — and the eleven commands
 
 ```bash
 pip install quantprobe
+quantprobe auto qwen3-coder --tps 15 --run   # empty machine -> optimal quant chosen, fetched, chatting
 ```
 
-**Zero-config on your own box**: `quantprobe plan --gguf model.gguf` auto-detects the machine and reads the model from the file. Presets/flags estimate any *other* machine. `hw`/`plan`/`target`/`optimize` need nothing else installed; the rest drive stock [llama.cpp](https://github.com/ggml-org/llama.cpp/releases) (point at it with `--llama-dir`/`QUANTPROBE_LLAMA_DIR`/`PATH`; preview any command with `--dry`).
+**Zero-config on your own box**: `quantprobe plan --gguf model.gguf` auto-detects the machine and reads the model from the file. Presets/flags estimate any *other* machine. `hw`/`plan`/`target`/`optimize` need nothing else installed (`auto` needs network for the fetch); the rest drive stock [llama.cpp](https://github.com/ggml-org/llama.cpp/releases) (point at it with `--llama-dir`/`QUANTPROBE_LLAMA_DIR`/`PATH`; preview any command with `--dry`).
 
 ```bash
+quantprobe auto qwen3-30b --tps 15                       # ONE command: optimizer picks bits, closest quant fetched, run command printed
+quantprobe auto qwen3-30b --custom                       # THE PRODUCT: probe YOUR model, build its personalized depth-aware GGUF
 quantprobe hw                                            # what the law sees on THIS machine (every value source-tagged)
 quantprobe plan     --gguf model.gguf                    # zero-config prediction: placement + tok/s + the launch command
 quantprobe optimize --tps 20                             # CHEAPEST PATH to a target: bits x placement x hardware, Pareto-ranked
@@ -103,7 +106,7 @@ quantprobe quantize --gguf f16.gguf --out 2bit.gguf      # COMPRESS: depth-aware
 quantprobe probe    --gguf f16.gguf --eval wiki.test.raw # measure YOUR model's fragile band (~30 min); --apply builds it
 quantprobe run      --gguf 2bit.gguf                     # plan the placement, then LAUNCH llama.cpp chat
 quantprobe bench    --gguf 2bit.gguf --contribute        # predicted vs measured on your box; opt-in datapoint
-quantprobe dashboard --gguf 2bit.gguf                    # chat while every reply is scored against the prediction
+quantprobe dashboard --gguf 2bit.gguf                    # the law LIVE: neuron galaxy + thinking toggle, every reply scored vs prediction
 ```
 
 The loop is self-validating: `plan` predicted 17.5 for a file we then measured at **18.32 ± 0.17**; the config months of research converged to is what `optimize` picks blind. A measured example of what that's worth: the same model, mis-specified vs law-routed, is **3.38 vs 18.32 tok/s (×5.4)** — [worked examples](docs/EXAMPLES.md).

{quantprobe-1.4.0 → quantprobe-1.5.1}/README.md

@@ -67,15 +67,18 @@ Full statements, each with its establishing measurement and a falsifiable predic
 <p align="center"><img src="weights/data/x_chart_I_kvdepth.png" width="700" alt="Decode speed versus context depth: measured 20.02 tok/s at depth zero falling to 16.12 at 16384, on the Law 4 v2 curve with eta_kv 0.70"></p>
 
 
-## Install — and the ten commands
+## Install — and the eleven commands
 
 ```bash
 pip install quantprobe
+quantprobe auto qwen3-coder --tps 15 --run   # empty machine -> optimal quant chosen, fetched, chatting
 ```
 
-**Zero-config on your own box**: `quantprobe plan --gguf model.gguf` auto-detects the machine and reads the model from the file. Presets/flags estimate any *other* machine. `hw`/`plan`/`target`/`optimize` need nothing else installed; the rest drive stock [llama.cpp](https://github.com/ggml-org/llama.cpp/releases) (point at it with `--llama-dir`/`QUANTPROBE_LLAMA_DIR`/`PATH`; preview any command with `--dry`).
+**Zero-config on your own box**: `quantprobe plan --gguf model.gguf` auto-detects the machine and reads the model from the file. Presets/flags estimate any *other* machine. `hw`/`plan`/`target`/`optimize` need nothing else installed (`auto` needs network for the fetch); the rest drive stock [llama.cpp](https://github.com/ggml-org/llama.cpp/releases) (point at it with `--llama-dir`/`QUANTPROBE_LLAMA_DIR`/`PATH`; preview any command with `--dry`).
 
 ```bash
+quantprobe auto qwen3-30b --tps 15                       # ONE command: optimizer picks bits, closest quant fetched, run command printed
+quantprobe auto qwen3-30b --custom                       # THE PRODUCT: probe YOUR model, build its personalized depth-aware GGUF
 quantprobe hw                                            # what the law sees on THIS machine (every value source-tagged)
 quantprobe plan     --gguf model.gguf                    # zero-config prediction: placement + tok/s + the launch command
 quantprobe optimize --tps 20                             # CHEAPEST PATH to a target: bits x placement x hardware, Pareto-ranked
@@ -85,7 +88,7 @@ quantprobe quantize --gguf f16.gguf --out 2bit.gguf      # COMPRESS: depth-aware
 quantprobe probe    --gguf f16.gguf --eval wiki.test.raw # measure YOUR model's fragile band (~30 min); --apply builds it
 quantprobe run      --gguf 2bit.gguf                     # plan the placement, then LAUNCH llama.cpp chat
 quantprobe bench    --gguf 2bit.gguf --contribute        # predicted vs measured on your box; opt-in datapoint
-quantprobe dashboard --gguf 2bit.gguf                    # chat while every reply is scored against the prediction
+quantprobe dashboard --gguf 2bit.gguf                    # the law LIVE: neuron galaxy + thinking toggle, every reply scored vs prediction
 ```
 
 The loop is self-validating: `plan` predicted 17.5 for a file we then measured at **18.32 ± 0.17**; the config months of research converged to is what `optimize` picks blind. A measured example of what that's worth: the same model, mis-specified vs law-routed, is **3.38 vs 18.32 tok/s (×5.4)** — [worked examples](docs/EXAMPLES.md).

{quantprobe-1.4.0 → quantprobe-1.5.1}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "quantprobe"
-version = "1.4.0"
+version = "1.5.1"
 description = "Probe-then-quantize for LLMs: measure a model's fragility curve, plan bit/tier placement by the tiered decode law, and emit ready-to-run llama.cpp recipes."
 readme = "README.md"
 license = { text = "MIT" }

{quantprobe-1.4.0 → quantprobe-1.5.1}/quantprobe/__init__.py

@@ -8,4 +8,4 @@ Three commands, all grounded in measured laws (see LAWS.md in the repo):
 Every constant in `plan` is fitted from measurements published in the repo, including two
 pre-registered hardware predictions confirmed within 8%.
 """
-__version__ = "1.4.0"
+__version__ = "1.5.1"

quantprobe-1.5.1/quantprobe/auto.py

@@ -0,0 +1,171 @@
+"""quantprobe auto — one command: model in, running setup out.
+
+    quantprobe auto qwen3-30b --tps 15          # preset
+    quantprobe auto unsloth/Qwen3-30B-A3B-GGUF  # any HF GGUF repo (params read from filenames' size)
+
+The two-speed design:
+  FAST PATH (this command): detect the machine -> ask the optimizer for the best effective-bits ->
+  scan the HF repo's file list and pick the closest-matching quant BY SIZE (bits = size*8/params —
+  no fragile name parsing) -> fetch it -> print the prediction and the run command (--run launches).
+  CUSTOM PATH (the actual product, printed at the end): probe YOUR model's fragile band and build a
+  depth-aware GGUF at the same bits — better quality at the same bytes. `probe --apply` does it.
+"""
+from __future__ import annotations
+import json, urllib.request
+
+from . import plan as planmod
+from . import optimize as optmod
+
+# preset -> (repo, total params B, active B, always-active B, moe)
+MODEL_REPOS = {
+    "qwen3-30b":  ("unsloth/Qwen3-30B-A3B-GGUF", 30.5, 3.3, 1.2, True),
+    "qwen3-coder": ("unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF", 30.5, 3.3, 1.2, True),
+    "glm-air":    ("unsloth/GLM-4.5-Air-GGUF", 110, 12, 2.7, True),
+    "laguna-s":   ("unsloth/Laguna-S-2.1-GGUF", 117.6, 8, 2.5, True),
+    "gemma-12b":  ("unsloth/gemma-4-12b-it-GGUF", 11.9, 11.9, 11.9, False),
+    "mistral-7b": ("unsloth/Mistral-7B-Instruct-v0.3-GGUF", 7.2, 7.2, 7.2, False),
+}
+
+
+def list_ggufs(repo):
+    """[(path, size_bytes)] for a HF repo, via the public tree API."""
+    req = urllib.request.Request(f"https://huggingface.co/api/models/{repo}/tree/main",
+                                 headers={"User-Agent": "quantprobe-auto"})
+    with urllib.request.urlopen(req, timeout=30) as r:
+        return [(f["path"], f.get("size", 0)) for f in json.load(r)
+                if f["path"].endswith(".gguf") and f.get("size", 0) > 1e8
+                and "mmproj" not in f["path"].lower() and "draft" not in f["path"].lower()]
+
+
+WIKI_URL = "https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip"
+
+
+def ensure_eval(dest_dir):
+    """The held-out corpus for probing, fetched once (1.3 MB)."""
+    import os, zipfile, io
+    path = os.path.join(dest_dir, "wiki.test.raw")
+    if os.path.isfile(path):
+        return path
+    print("[quantprobe auto] fetching the WikiText-2 eval corpus (1.3 MB, one time)...")
+    with urllib.request.urlopen(WIKI_URL, timeout=60) as r:
+        z = zipfile.ZipFile(io.BytesIO(r.read()))
+    with z.open("wikitext-2-raw/wiki.test.raw") as f, open(path, "wb") as out:
+        out.write(f.read())
+    return path
+
+
+def pick_source(files, t):
+    """High-precision source for the custom build: prefer Q8-class (smallest requantizable), else max bits."""
+    cands = []
+    for path, size in files:
+        if "-of-" in path.lower():
+            continue                                # split multi-part files: out of scope for v1
+        bits = size * 8 / (t * 1e9)
+        cands.append((bits, size, path))
+    if not cands:
+        return None
+    hi = sorted([c for c in cands if c[0] >= 7.5], key=lambda c: c[1])
+    return hi[0] if hi else max(cands)
+
+
+def run(a):
+    target = a.target
+    if target in MODEL_REPOS:
+        repo, t, ac, ne, moe = MODEL_REPOS[target]
+    else:
+        repo = target
+        if not getattr(a, "total", None):
+            raise SystemExit(f"'{target}' is not a preset ({', '.join(MODEL_REPOS)}) - for a raw HF "
+                             "repo also pass --total (B params) and, for MoE, --active/--always-active")
+        t, ac = a.total, a.active or a.total
+        ne = a.always_active or (ac if ac >= t * 0.9 else ac * 0.35)
+        moe = ac < t * 0.9
+    a.total, a.active, a.always_active, a.model = t, ac, ne, None
+
+    # 1. what does the optimizer want on THIS machine?
+    a.gguf = None
+    ranked = optmod.run(a)
+    want_bits = ranked[0]["bits"]
+    _, _, _, _, _, (vc, vb, rc, rb, db, geta, gl), ctx, kvp = optmod.resolve(a)
+
+    # 2. pick the closest file BY SIZE (bits = size*8/params; honest, format-agnostic)
+    try:
+        files = list_ggufs(repo)
+    except Exception as e:
+        raise SystemExit(f"could not list {repo}: {e}")
+    if not files:
+        raise SystemExit(f"no GGUF files found in {repo}")
+    scored = []
+    for path, size in files:
+        bits = size * 8 / (t * 1e9)
+        if bits < 1.0 or bits > 9:
+            continue
+        _, _, cfgs = planmod.evaluate(t, ac, ne, moe, bits, vc, vb, rc, rb, db, geta, 1.0, gl,
+                                      ctx=ctx, kvp=kvp)
+        cfgs = [c for c in cfgs if "expert cache" not in c[0]]
+        if not cfgs:
+            continue
+        scored.append((abs(bits - want_bits), -cfgs[0][1], path, size, bits, cfgs[0]))
+    if not scored:
+        raise SystemExit("no usable quant in that repo for this machine")
+    scored.sort()
+    _, _, path, size, bits, best = scored[0]
+
+    if getattr(a, "custom", False):
+        src = pick_source(files, t)
+        if not src:
+            raise SystemExit("no usable high-precision source in that repo (split files unsupported)")
+        sbits, ssize, spath = src
+        import os
+        dest = getattr(a, "dir", None) or "./models"
+        os.makedirs(dest, exist_ok=True)
+        print("\n[quantprobe auto --custom] source: " + spath + f" ({ssize/1e9:.1f} GB, {sbits:.1f}-bit)")
+        print("  pipeline: fetch source -> probe the fragile band (~30-60 min) -> build the")
+        print("  depth-aware GGUF. Interrupt anytime; the fetch resumes.")
+        if getattr(a, "dry", False):
+            print("  (--dry: nothing downloaded)")
+            return spath
+        from . import fetch as fetchmod
+        if not fetchmod.fetch(repo, dest, spath, fetchmod.token()):
+            raise SystemExit("source download failed (re-run: it resumes)")
+        srcfull = os.path.join(dest, os.path.basename(spath))
+        evalf = ensure_eval(dest)
+        from . import probe as probemod
+        import argparse
+        out = os.path.join(dest, os.path.basename(spath).rsplit(".gguf", 1)[0] + "-depthaware.gguf")
+        pa = argparse.Namespace(gguf=srcfull, bands=4, chunks=32, eval=evalf, ngl=99,
+                                workdir=dest, llama_dir=getattr(a, "llama_dir", None),
+                                apply=True, out=out, dry_run=False)
+        probemod.run(pa)
+        print("\n[quantprobe auto --custom] your personalized model:")
+        print("  quantprobe run --gguf " + out)
+        if getattr(a, "run", False):
+            from . import runtime
+            a.gguf = out; a.bits = None
+            runtime.run(a)
+        return out
+    print(f"\n[quantprobe auto] optimizer wants ~{want_bits:g}-bit; closest file in {repo}:")
+    print(f"  {path}  ({size/1e9:.1f} GB, {bits:.2f} effective bits)")
+    print(f"  predicted on this machine: {best[1]:.1f} tok/s  ({best[0]})")
+    if getattr(a, "dry", False):
+        print("  (--dry: nothing downloaded)")
+        return path
+    # 3. fetch it (resumable), then hand off
+    from . import fetch as fetchmod
+    dest = getattr(a, "dir", None) or "./models"
+    import os
+    os.makedirs(dest, exist_ok=True)
+    ok = fetchmod.fetch(repo, dest, path, fetchmod.token())
+    if not ok:
+        raise SystemExit("download failed (it resumes: re-run the same command)")
+    full = os.path.join(dest, os.path.basename(path))
+    print(f"\n[quantprobe auto] ready. Run it:")
+    print(f"  quantprobe run --gguf {full}")
+    print("\n  Better quality at the SAME size: rerun with --custom - it probes YOUR model's")
+    print("  fragile band (~30-60 min) and builds a depth-aware GGUF personalized to it.")
+    if getattr(a, "run", False):
+        from . import runtime
+        a.gguf = full
+        a.bits = None                              # let autospec read the real file
+        runtime.run(a)
+    return full

{quantprobe-1.4.0 → quantprobe-1.5.1}/quantprobe/cli.py

@@ -91,6 +91,18 @@ def main():
     o.add_argument("--allow-prune", action="store_true", help="include REAP-class pruned variants (domain-specialized: +39%% out-of-domain ppl measured)")
     hwargs(o)
 
+    au = sub.add_parser("auto", help="ONE command: model in, running setup out - optimizer picks the bits, the closest quant is fetched, run command printed")
+    au.add_argument("target", help="model preset (qwen3-30b, qwen3-coder, glm-air, laguna-s, gemma-12b, mistral-7b) or a HF GGUF repo id")
+    au.add_argument("--dir", default="./models", help="download directory (default ./models)")
+    au.add_argument("--run", action="store_true", help="launch chat immediately after the download")
+    hwargs(au)
+    au.add_argument("--tps", type=float, default=None, help="target tok/s for the optimizer")
+    au.add_argument("--max-quality", type=float, default=None)
+    au.add_argument("--allow-prune", action="store_true"); au.add_argument("--any-runtime", action="store_true")
+    au.add_argument("--custom", action="store_true", help="THE PRODUCT: fetch a high-precision source, probe YOUR model (~30-60 min), build a depth-aware GGUF personalized to it")
+    au.add_argument("--serve", action="store_true", help="with --run: llama-server instead of chat")
+    au.add_argument("--extra", default=None)
+
     hwp = sub.add_parser("hw", help="detect THIS machine's memory tiers (no flags needed); every value tagged with its source")
     hwp.add_argument("--measure", default=None, metavar="FILE", help="also MEASURE sequential disk read on a real file (e.g. a GGUF)")
 
@@ -116,6 +128,9 @@ def main():
     elif a.cmd == "target":
         from . import target
         target.run(a)
+    elif a.cmd == "auto":
+        from . import auto
+        auto.run(a)
     elif a.cmd == "optimize":
         from . import optimize
         optimize.run(a)

quantprobe-1.5.1/quantprobe/dashboard.py

@@ -0,0 +1,384 @@
+"""quantprobe dashboard — the law, live.
+
+One command: plans the best placement, launches llama-server with those flags, and serves a local
+single-viewport app (no page scrolling: fixed sidebar, ChatGPT-style internal chat scroll) where
+every reply is scored against the law's prediction. The NEURON GALAXY renders every expert of every
+layer as a dot, illuminated per generated token — thinking tokens in violet, answer tokens in the
+color of the memory tier the experts live on.
+
+Honesty rule, printed on the panel: stock llama.cpp exposes no per-token router telemetry, so
+expert illumination is sampled uniformly — which Law 2's measured flat routing (Jaccard 1.00)
+makes the statistically exact picture. Topic-affinity atlases require an engine that exports
+router data; this page shows nothing it cannot back.
+"""
+from __future__ import annotations
+import json, os, subprocess, threading, time, urllib.request, webbrowser
+from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
+
+from . import runtime
+
+DASH_HTML = """<!doctype html><html><head><meta charset="utf-8">
+<meta name="viewport" content="width=device-width,initial-scale=1">
+<title>quantprobe — the law, live</title>
+<style>
+:root{--g:#fafaf7;--p:#fff;--ink:#16181d;--sub:#5c6066;--line:#e5e4df;--acc:#0f766e;
+--accsoft:#e6f2f0;--warn:#b45309;--ram:#d97706;--disk:#dc2626;--think:#7c6bd1;
+--mono:Consolas,ui-monospace,monospace}
+@media(prefers-color-scheme:dark){:root{--g:#101216;--p:#181b20;--ink:#e8e7e2;--sub:#9aa0a8;
+--line:#2b2f35;--acc:#35b8a6;--accsoft:#12332f;--warn:#e0a458;--ram:#f0a94b;--disk:#f26d6d;--think:#a99af0}}
+*{box-sizing:border-box}
+body{margin:0;height:100vh;overflow:hidden;background:var(--g);color:var(--ink);
+font-family:"Segoe UI",system-ui,sans-serif;display:grid;grid-template-rows:auto 1fr}
+header{display:flex;align-items:baseline;gap:14px;padding:10px 18px;border-bottom:1px solid var(--line)}
+header h1{font-size:17px;margin:0}
+header .sub{color:var(--sub);font-size:12px;overflow:hidden;text-overflow:ellipsis;white-space:nowrap}
+header .law{margin-left:auto;font-family:var(--mono);font-size:11.5px;color:var(--acc);white-space:nowrap}
+main{display:grid;grid-template-columns:320px 1fr;gap:14px;padding:14px;min-height:0}
+aside{overflow-y:auto;min-height:0;display:flex;flex-direction:column;gap:12px}
+.panel{background:var(--p);border:1px solid var(--line);border-radius:10px;padding:12px;flex-shrink:0}
+.k{font-size:10px;text-transform:uppercase;letter-spacing:.09em;color:var(--sub);font-weight:600;margin-bottom:7px}
+.gauge{display:flex;gap:14px;align-items:baseline}
+.gauge .n{font-family:var(--mono);font-size:34px;font-weight:700;color:var(--acc);font-variant-numeric:tabular-nums}
+.gauge .lab{font-size:10.5px;color:var(--sub)}
+.delta{font-family:var(--mono);font-size:14px;font-weight:600}
+.log{font-family:var(--mono);font-size:11px;color:var(--sub);max-height:72px;overflow-y:auto;margin-top:6px}
+.stat{display:flex;justify-content:space-between;font-family:var(--mono);font-size:11.5px;color:var(--sub);margin:3px 0}
+.stat b{color:var(--ink);font-weight:600}
+.tgl{display:flex;align-items:center;gap:8px;font-size:13px;cursor:pointer;user-select:none}
+.tgl input{accent-color:var(--think);width:15px;height:15px}
+.tier{margin:7px 0}.tier .l{display:flex;justify-content:space-between;font-family:var(--mono);font-size:10.5px;color:var(--sub);margin-bottom:3px}
+.bar{height:10px;background:var(--g);border:1px solid var(--line);border-radius:6px;overflow:hidden}
+.fill{height:100%;background:var(--acc)}
+.foot{font-size:11px;color:var(--sub);line-height:1.5}.foot a{color:var(--acc)}
+.right{display:grid;grid-template-rows:minmax(220px,37%) 1fr;gap:14px;min-height:0}
+.brainp{position:relative;background:var(--p);border:1px solid var(--line);border-radius:10px;padding:10px 12px;
+display:flex;flex-direction:column;min-height:0}
+#brain{flex:1;width:100%;min-height:0;border-radius:6px}
+.legend{display:flex;gap:13px;font-size:10.5px;color:var(--sub);margin-top:6px;flex-wrap:wrap;align-items:center}
+.dot{display:inline-block;width:8px;height:8px;border-radius:50%;margin-right:4px;vertical-align:-1px}
+.chatp{display:flex;flex-direction:column;min-height:0;background:var(--p);border:1px solid var(--line);border-radius:10px;padding:12px}
+.msgs{flex:1;overflow-y:auto;min-height:0;display:flex;flex-direction:column;gap:8px;padding-right:4px}
+.m{max-width:85%;padding:8px 12px;border-radius:10px;font-size:13.5px;line-height:1.5;white-space:pre-wrap}
+.m.u{align-self:flex-end;background:var(--accsoft)}
+.m.a{align-self:flex-start;background:var(--g);border:1px solid var(--line)}
+.m .t{display:block;margin-top:6px;font-family:var(--mono);font-size:10px;color:var(--acc)}
+.m pre{background:var(--p);border:1px solid var(--line);border-radius:6px;padding:8px;overflow-x:auto;font-family:var(--mono);font-size:12px;margin:6px 0;white-space:pre}
+.thk{align-self:flex-start;max-width:85%;font-size:11.5px;color:var(--think);border-left:3px solid var(--think);
+padding:3px 10px;opacity:.85;white-space:pre-wrap}
+.thk span{display:block;max-height:80px;overflow-y:auto}
+.thk summary{cursor:pointer;font-size:10.5px;letter-spacing:.05em}
+.row{display:flex;gap:8px;margin-top:10px}
+textarea{flex:1;resize:none;height:52px;padding:9px;font-size:13.5px;font-family:inherit;
+background:var(--g);color:var(--ink);border:1px solid var(--line);border-radius:8px}
+button{padding:0 18px;background:var(--acc);color:#fff;border:0;border-radius:8px;font-size:14px;font-weight:600;cursor:pointer}
+button:disabled{opacity:.5}
+</style></head><body>
+<header><h1>The law, live</h1>
+<div class="sub">{{MODEL}} · {{PLACEMENT}}</div>
+<div class="law">tok/s = η·BW ÷ active-bytes → predicted {{PRED}}</div></header>
+<main>
+<aside>
+  <div class="panel"><div class="k">Prediction vs reality</div>
+    <div class="gauge">
+      <div><div class="n">{{PRED}}</div><div class="lab">predicted tok/s</div></div>
+      <div><div class="n" id="meas">—</div><div class="lab">measured avg</div></div>
+      <div class="delta" id="delta"></div>
+    </div>
+    <div class="log" id="log">each reply appends a data point…</div>
+  </div>
+  <div class="panel"><div class="k">Thinking</div>
+    <label class="tgl"><input type="checkbox" id="think" checked> allow thinking <span style="color:var(--think)">●</span></label>
+    <div style="margin-top:8px">
+      <div class="stat"><span>time to first token</span><b id="ttft">—</b></div>
+      <div class="stat"><span>thinking</span><b id="sthink">—</b></div>
+      <div class="stat"><span>answer</span><b id="sans">—</b></div>
+    </div>
+    <div class="foot" style="margin-top:6px">Thinking improves hard answers but costs latency at
+    decode speed — toggle it off for quick tasks.</div>
+  </div>
+  <div class="panel"><div class="k">Placement (planned)</div>{{TIERS}}
+    <div class="foot" style="font-family:var(--mono);font-size:10px">{{FLAGS}}</div></div>
+  <div class="panel"><div class="k">About</div>
+    <div class="foot">A running experiment: the law predicted this machine's speed before the model
+    loaded. <a href="https://github.com/FedericoTs/quantprobe">repo</a> ·
+    <a href="https://x.com/federico_sciuca">@federico_sciuca</a></div></div>
+</aside>
+<div class="right">
+  <div class="brainp"><div class="k">Neuron galaxy — {{NL}} layers × {{NE}} experts · every token traverses all layers</div>
+    <canvas id="brain"></canvas>
+    <div class="legend">
+      <span><span class="dot" style="background:var(--acc)"></span>attention + KV → {{ATIER}}</span>
+      <span><span class="dot" style="background:var({{ECOLV}})"></span>experts ({{NK}} of {{NE}} fire per layer per token) → {{ETIER}}</span>
+      <span><span class="dot" style="background:var(--think)"></span>thinking</span>
+      <span style="opacity:.75">flashes sampled uniformly — statistically exact under measured flat routing (Jaccard 1.00); stock llama.cpp exposes no router telemetry</span>
+    </div>
+  </div>
+  <div class="chatp"><div class="k">Chat — each reply is a measurement</div>
+    <div class="msgs" id="msgs"></div>
+    <div class="row"><textarea id="inp" placeholder="Say something… (Enter to send)"></textarea>
+    <button id="send">Send</button></div>
+  </div>
+</div>
+</main>
+<script>
+const NL={{NL}},NE={{NE}},NK={{NK}},PRED=parseFloat("{{PRED}}");
+const hist=[],pts=[];const $=id=>document.getElementById(id);
+function esc(s){return s.replace(/&/g,"&amp;").replace(/</g,"&lt;")}
+function md(s){return esc(s).replace(/```(\\w*)\\n?([\\s\\S]*?)```/g,'<pre>$2</pre>')}
+function cssv(n){return getComputedStyle(document.body).getPropertyValue(n).trim()||"#888"}
+
+/* ---- neuron galaxy ---- */
+const cv=$("brain"),cx2=cv.getContext("2d");
+const NDOT=Math.max(NE,1);let W,H,px,py,ax,ay,heat,tint,aheat;
+function layout(){
+  W=cv.width=cv.clientWidth*devicePixelRatio;H=cv.height=cv.clientHeight*devicePixelRatio;
+  const CX=W/2,CY=H/2,MR=Math.min(W,H)*0.46;
+  px=new Float32Array(NL*NDOT);py=new Float32Array(NL*NDOT);
+  ax=new Float32Array(NL);ay=new Float32Array(NL);
+  heat=new Float32Array(NL*NDOT);tint=new Uint8Array(NL*NDOT);aheat=new Float32Array(NL);
+  for(let l=0;l<NL;l++){
+    const t=NL>1?l/(NL-1):0, ang=t*Math.PI*3.8+0.7, R=(0.18+0.78*t)*MR;
+    const cxl=CX+R*Math.cos(ang), cyl=CY+R*Math.sin(ang)*0.6;
+    ax[l]=cxl;ay[l]=cyl;
+    const sig=MR*(0.05+0.05*Math.sin(l*3.7)*Math.sin(l*3.7));
+    for(let e=0;e<NDOT;e++){
+      const r=sig*Math.sqrt((e+0.5)/NDOT)*2.1, a=e*2.39996+l*0.7;
+      px[l*NDOT+e]=cxl+r*Math.cos(a);
+      py[l*NDOT+e]=cyl+r*Math.sin(a)*0.85;
+    }
+  }
+  draw();
+}
+function draw(){
+  cx2.clearRect(0,0,W,H);
+  const ec=cssv("{{ECOLV}}"),tc=cssv("--think"),ac=cssv("--acc");
+  const s=2.1*devicePixelRatio, sa=3.4*devicePixelRatio;
+  for(let i=0;i<NL*NDOT;i++){
+    const h=heat[i];
+    cx2.globalAlpha=0.09+0.91*h;
+    cx2.fillStyle=(h>0.03&&tint[i])?tc:ec;
+    cx2.fillRect(px[i],py[i],s,s);
+  }
+  for(let l=0;l<NL;l++){
+    cx2.globalAlpha=0.35+0.65*aheat[l];
+    cx2.fillStyle=ac;
+    cx2.beginPath();cx2.arc(ax[l],ay[l],sa,0,7);cx2.fill();
+  }
+  cx2.globalAlpha=1;
+}
+let decaying=false;
+function decay(){let live=false;
+  for(let i=0;i<heat.length;i++){heat[i]*=0.86;if(heat[i]>0.02)live=true}
+  for(let l=0;l<NL;l++){aheat[l]*=0.86;if(aheat[l]>0.02)live=true}
+  draw();
+  if(live)requestAnimationFrame(decay);else decaying=false}
+function pulse(thinking){
+  for(let l=0;l<NL;l++){aheat[l]=1;
+    for(let k=0;k<Math.max(NK,1);k++){const i=l*NDOT+((Math.random()*NDOT)|0);
+      heat[i]=1;tint[i]=thinking?1:0}}
+  if(!decaying){decaying=true;requestAnimationFrame(decay)}}
+addEventListener("resize",layout);layout();
+
+/* ---- streaming chat with reply anatomy ---- */
+function fmt(ms){return ms<1000?ms.toFixed(0)+" ms":(ms/1000).toFixed(1)+" s"}
+async function send(){
+  const t=$("inp").value.trim(); if(!t)return;
+  $("inp").value=""; $("send").disabled=true;
+  hist.push({role:"user",content:t});
+  $("msgs").insertAdjacentHTML("beforeend",'<div class="m u">'+esc(t)+'</div>');
+  const thk=document.createElement("details");thk.className="thk";thk.open=true;
+  thk.innerHTML='<summary>thinking…</summary><span></span>';
+  const hold=document.createElement("div");hold.className="m a";hold.textContent="…";
+  $("msgs").appendChild(thk);$("msgs").appendChild(hold);$("msgs").scrollTop=1e9;
+  let content="",think="",nT=0,nA=0,t0=performance.now(),tFirst=0,tThinkEnd=0,timings=null;
+  try{
+    const r=await fetch("/api/chat",{method:"POST",headers:{"Content-Type":"application/json"},
+      body:JSON.stringify({messages:hist.slice(-12),max_tokens:2048,stream:true,think:$("think").checked})});
+    const rd=r.body.getReader(),dec=new TextDecoder();let buf="";
+    while(true){
+      const {done,value}=await rd.read(); if(done)break;
+      buf+=dec.decode(value,{stream:true});
+      let i;while((i=buf.indexOf("\\n"))>=0){
+        const line=buf.slice(0,i).trim();buf=buf.slice(i+1);
+        if(!line.startsWith("data:"))continue;
+        const p=line.slice(5).trim(); if(p==="[DONE]")continue;
+        let j;try{j=JSON.parse(p)}catch(e){continue}
+        if(j.timings)timings=j.timings;
+        const d=(j.choices&&j.choices[0]&&j.choices[0].delta)||{};
+        if(d.reasoning_content){think+=d.reasoning_content;nT++;pulse(true);
+          if(!tFirst)tFirst=performance.now();
+          thk.lastElementChild.textContent=think.slice(-1500);
+          thk.lastElementChild.scrollTop=1e9}
+        if(d.content){content+=d.content;nA++;pulse(false);
+          if(!tFirst)tFirst=performance.now();
+          if(!tThinkEnd)tThinkEnd=performance.now();
+          hold.innerHTML=md(content)}
+      }
+      $("msgs").scrollTop=1e9;
+    }
+  }catch(e){hold.textContent="error: "+e}
+  const tEnd=performance.now();
+  if(think){thk.open=false;
+    thk.firstElementChild.textContent="thought for ~"+think.split(/\\s+/).length+" words (click to expand)"}
+  else thk.remove();
+  if(!content)content="(empty reply — try again)";
+  hist.push({role:"assistant",content:content});
+  const tps=timings?timings.predicted_per_second:((nT+nA)>0?(nT+nA)/((tEnd-t0)/1000):0);
+  hold.innerHTML=md(content)+'<span class="t">'+tps.toFixed(1)+' tok/s'+(timings?"":" (client est.)")+' · predicted '+PRED+'</span>';
+  $("ttft").textContent=tFirst?fmt(tFirst-t0):"—";
+  $("sthink").textContent=nT?fmt((tThinkEnd||tEnd)-(tFirst||t0))+" · "+nT+" tok":"off";
+  $("sans").textContent=nA?fmt(tEnd-(tThinkEnd||tFirst||t0))+" · "+nA+" tok":"—";
+  if(tps>0){pts.push(tps);
+    const avg=pts.reduce((a,b)=>a+b,0)/pts.length;
+    $("meas").textContent=avg.toFixed(1);
+    const dd=(avg/PRED-1)*100;
+    $("delta").textContent=(dd>=0?"+":"")+dd.toFixed(0)+"%";
+    $("delta").style.color=Math.abs(dd)<25?"var(--acc)":"var(--warn)";
+    $("log").insertAdjacentHTML("afterbegin","#"+pts.length+": "+tps.toFixed(1)+" tok/s<br>");}
+  $("send").disabled=false;$("msgs").scrollTop=1e9;
+}
+$("send").onclick=send;
+$("inp").addEventListener("keydown",e=>{if(e.key==="Enter"&&!e.shiftKey){e.preventDefault();send();}});
+</script></body></html>"""
+
+
+def make_handler(upstream, page):
+    class H(BaseHTTPRequestHandler):
+        protocol_version = "HTTP/1.1"
+
+        def log_message(self, *a):
+            pass
+
+        def do_GET(self):
+            body = page.encode("utf-8")
+            self.send_response(200)
+            self.send_header("Content-Type", "text/html; charset=utf-8")
+            self.send_header("Content-Length", str(len(body)))
+            self.end_headers()
+            self.wfile.write(body)
+
+        def do_POST(self):
+            if self.path != "/api/chat":
+                self.send_response(404)
+                self.send_header("Content-Length", "0")
+                self.end_headers()
+                return
+            n = int(self.headers.get("Content-Length", 0))
+            body = json.loads(self.rfile.read(n) or b"{}")
+            body.setdefault("timings_per_token", True)
+            body.setdefault("max_tokens", 2048)
+            if body.pop("think", True) is False:      # UI toggle: skip thinking entirely
+                body["chat_template_kwargs"] = {"enable_thinking": False}
+            stream = bool(body.get("stream", False))
+            req = urllib.request.Request(upstream + "/v1/chat/completions",
+                                         data=json.dumps(body).encode(),
+                                         headers={"Content-Type": "application/json"})
+            try:
+                r = urllib.request.urlopen(req, timeout=600)
+                self.send_response(200)
+                if stream:
+                    self.send_header("Content-Type", "text/event-stream")
+                    self.send_header("Cache-Control", "no-cache")
+                    self.send_header("Transfer-Encoding", "chunked")
+                    self.end_headers()
+                    while True:
+                        chunk = r.read1(8192) if hasattr(r, "read1") else r.read(8192)
+                        if not chunk:
+                            break
+                        self.wfile.write(b"%x\r\n" % len(chunk) + chunk + b"\r\n")
+                        self.wfile.flush()
+                    self.wfile.write(b"0\r\n\r\n")
+                else:
+                    data = r.read()
+                    self.send_header("Content-Type", "application/json")
+                    self.send_header("Content-Length", str(len(data)))
+                    self.end_headers()
+                    self.wfile.write(data)
+            except Exception as e:
+                try:
+                    err = json.dumps({"error": str(e)}).encode()
+                    self.send_response(502)
+                    self.send_header("Content-Length", str(len(err)))
+                    self.end_headers()
+                    self.wfile.write(err)
+                except Exception:
+                    pass
+    return H
+
+
+def _anatomy(gguf):
+    """Exact layer/expert counts from the file, for the galaxy."""
+    try:
+        from gguf import GGUFReader
+        r = GGUFReader(gguf)
+        nl = ne = nk = 0
+        for f in r.fields.values():
+            if f.name.endswith(".block_count"):
+                nl = int(f.parts[f.data[0]][0])
+            elif f.name.endswith(".expert_count"):
+                ne = int(f.parts[f.data[0]][0])
+            elif f.name.endswith(".expert_used_count"):
+                nk = int(f.parts[f.data[0]][0])
+        return (nl or 32), ne, (nk if ne else 0)
+    except Exception:
+        return 32, 0, 0
+
+
+def dashboard(a):
+    best, flags = runtime.best_flags(a)
+    binp = runtime.find_llama(a.llama_dir, "llama-server")
+    sport = a.server_port
+    cmd = [binp, "-m", a.gguf, "--port", str(sport)] + flags
+    print(f"[quantprobe] placement: {best[0]}  (predicted {best[1]:.1f} tok/s)")
+    print("[quantprobe] llama-server:", " ".join(cmd))
+    proc = subprocess.Popen(cmd, stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL)
+    upstream = f"http://127.0.0.1:{sport}"
+    print("[quantprobe] waiting for the model to finish loading (big models on slow disks take minutes)...")
+    ready_req = json.dumps({"messages": [{"role": "user", "content": "hi"}], "max_tokens": 1}).encode()
+    for _ in range(600):
+        try:
+            urllib.request.urlopen(urllib.request.Request(
+                upstream + "/v1/chat/completions", data=ready_req,
+                headers={"Content-Type": "application/json"}), timeout=8)
+            break                                   # a real completion answered: weights are up
+        except Exception:
+            if proc.poll() is not None:
+                raise SystemExit("llama-server exited during startup — check the model path/flags")
+            time.sleep(2)
+    tiers = ""
+    for name, cap, used in runtime.tier_view(a, best):
+        pct = min(100, used / cap * 100) if cap else 0
+        tiers += (f'<div class="tier"><div class="l"><span>{name}</span>'
+                  f'<span>{used:.1f} / {cap:.0f} GB</span></div>'
+                  f'<div class="bar"><div class="fill" style="width:{pct:.0f}%"></div></div></div>')
+    nl, ne, nk = _anatomy(a.gguf)
+    pname = best[0]
+    if "hybrid" in pname:
+        atier, etier, ecolv = "VRAM", "RAM", "--ram"
+    elif "all in VRAM" in pname:
+        atier, etier, ecolv = "VRAM", "VRAM", "--acc"
+    elif "disk" in pname:
+        atier, etier, ecolv = "RAM", "disk (streamed, RAM-cached)", "--disk"
+    else:
+        atier, etier, ecolv = "RAM", "RAM", "--ram"
+    page = (DASH_HTML.replace("{{MODEL}}", os.path.basename(a.gguf))
+            .replace("{{PLACEMENT}}", pname).replace("{{PRED}}", f"{best[1]:.1f}")
+            .replace("{{FLAGS}}", " ".join(flags)).replace("{{TIERS}}", tiers)
+            .replace("{{NL}}", str(nl)).replace("{{NE}}", str(ne)).replace("{{NK}}", str(nk))
+            .replace("{{ATIER}}", atier).replace("{{ETIER}}", etier).replace("{{ECOLV}}", ecolv))
+    ThreadingHTTPServer.allow_reuse_address = False   # Windows: reuse allows silent double-binds
+    try:
+        srv = ThreadingHTTPServer(("127.0.0.1", a.port), make_handler(upstream, page))
+    except OSError:
+        proc.terminate()
+        raise SystemExit(f"port {a.port} is already in use - another dashboard running? (pass --port N)")
+    url = f"http://127.0.0.1:{a.port}"
+    print(f"[quantprobe] dashboard v2.1: {url}  (Ctrl-C stops both)")
+    if not a.no_open:
+        threading.Timer(1.0, lambda: webbrowser.open(url)).start()
+    try:
+        srv.serve_forever()
+    except KeyboardInterrupt:
+        pass
+    finally:
+        proc.terminate()

{quantprobe-1.4.0 → quantprobe-1.5.1}/quantprobe/detect.py

@@ -20,7 +20,9 @@ def _run(cmd, timeout=10):
 
 # name-fragment -> (VRAM bandwidth GB/s, geta, gl). 1060 measured on the reference box; rest spec-sheet [table].
 GPU_TABLE = [
-    ("5090", 1792, 0.62, 0.42), ("4090", 1008, 0.62, 0.42), ("4080", 717, 0.6, 0.4),
+    ("5090", 1792, 0.62, 0.42), ("5080", 960, 0.62, 0.42), ("5070 ti", 896, 0.62, 0.42),
+    ("5070", 672, 0.6, 0.4), ("5060 ti", 448, 0.58, 0.38), ("5060", 448, 0.58, 0.38),
+    ("4090", 1008, 0.62, 0.42), ("4080", 717, 0.6, 0.4),
     ("4070", 504, 0.55, 0.35), ("4060", 272, 0.5, 0.3),
     ("3090", 936, 0.6, 0.4), ("3080", 760, 0.58, 0.38), ("3070", 448, 0.52, 0.32),
     ("3060 ti", 448, 0.5, 0.3), ("3060", 360, 0.5, 0.3), ("3050", 224, 0.45, 0.28),
@@ -107,11 +109,13 @@ def detect():
     gs = gpus()
     if gs:
         vram = sum(g[1] for g in gs)
-        bw0, geta, gl, src = gpu_lookup(gs[0][0])
-        vram_bw = bw0 * len(gs) * (1.0 if len(gs) == 1 else 0.85)   # aggregate w/ tensor-parallel loss
+        per = [gpu_lookup(g[0]) for g in gs]                        # per-card lookup (mixed pairs differ)
+        bw_sum = sum(p[0] for p in per)
+        vram_bw = bw_sum * (1.0 if len(gs) == 1 else 0.85)          # aggregate w/ tensor-parallel loss
+        geta, gl, src = per[0][1], per[0][2], per[0][3]             # eta class from the primary card
         names = " + ".join(g[0] for g in gs)
         notes.append(f"GPU: {names}, {vram:.0f} GB total [os], {vram_bw:.0f} GB/s {src}"
-                     + (f" (x{len(gs)} aggregate, 0.85 TP efficiency [est])" if len(gs) > 1 else ""))
+                     + (f" (x{len(gs)} per-card sum, 0.85 TP efficiency [est]; slower card gates its share)" if len(gs) > 1 else ""))
         hw.update(vram=vram, vram_bw=round(vram_bw), geta=geta, gl=gl)
     else:
         hw.update(vram=0, vram_bw=0)

{quantprobe-1.4.0 → quantprobe-1.5.1/quantprobe.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: quantprobe
-Version: 1.4.0
+Version: 1.5.1
 Summary: Probe-then-quantize for LLMs: measure a model's fragility curve, plan bit/tier placement by the tiered decode law, and emit ready-to-run llama.cpp recipes.
 Author: Federico Sciuca
 License: MIT
@@ -85,15 +85,18 @@ Full statements, each with its establishing measurement and a falsifiable predic
 <p align="center"><img src="weights/data/x_chart_I_kvdepth.png" width="700" alt="Decode speed versus context depth: measured 20.02 tok/s at depth zero falling to 16.12 at 16384, on the Law 4 v2 curve with eta_kv 0.70"></p>
 
 
-## Install — and the ten commands
+## Install — and the eleven commands
 
 ```bash
 pip install quantprobe
+quantprobe auto qwen3-coder --tps 15 --run   # empty machine -> optimal quant chosen, fetched, chatting
 ```
 
-**Zero-config on your own box**: `quantprobe plan --gguf model.gguf` auto-detects the machine and reads the model from the file. Presets/flags estimate any *other* machine. `hw`/`plan`/`target`/`optimize` need nothing else installed; the rest drive stock [llama.cpp](https://github.com/ggml-org/llama.cpp/releases) (point at it with `--llama-dir`/`QUANTPROBE_LLAMA_DIR`/`PATH`; preview any command with `--dry`).
+**Zero-config on your own box**: `quantprobe plan --gguf model.gguf` auto-detects the machine and reads the model from the file. Presets/flags estimate any *other* machine. `hw`/`plan`/`target`/`optimize` need nothing else installed (`auto` needs network for the fetch); the rest drive stock [llama.cpp](https://github.com/ggml-org/llama.cpp/releases) (point at it with `--llama-dir`/`QUANTPROBE_LLAMA_DIR`/`PATH`; preview any command with `--dry`).
 
 ```bash
+quantprobe auto qwen3-30b --tps 15                       # ONE command: optimizer picks bits, closest quant fetched, run command printed
+quantprobe auto qwen3-30b --custom                       # THE PRODUCT: probe YOUR model, build its personalized depth-aware GGUF
 quantprobe hw                                            # what the law sees on THIS machine (every value source-tagged)
 quantprobe plan     --gguf model.gguf                    # zero-config prediction: placement + tok/s + the launch command
 quantprobe optimize --tps 20                             # CHEAPEST PATH to a target: bits x placement x hardware, Pareto-ranked
@@ -103,7 +106,7 @@ quantprobe quantize --gguf f16.gguf --out 2bit.gguf      # COMPRESS: depth-aware
 quantprobe probe    --gguf f16.gguf --eval wiki.test.raw # measure YOUR model's fragile band (~30 min); --apply builds it
 quantprobe run      --gguf 2bit.gguf                     # plan the placement, then LAUNCH llama.cpp chat
 quantprobe bench    --gguf 2bit.gguf --contribute        # predicted vs measured on your box; opt-in datapoint
-quantprobe dashboard --gguf 2bit.gguf                    # chat while every reply is scored against the prediction
+quantprobe dashboard --gguf 2bit.gguf                    # the law LIVE: neuron galaxy + thinking toggle, every reply scored vs prediction
 ```
 
 The loop is self-validating: `plan` predicted 17.5 for a file we then measured at **18.32 ± 0.17**; the config months of research converged to is what `optimize` picks blind. A measured example of what that's worth: the same model, mis-specified vs law-routed, is **3.38 vs 18.32 tok/s (×5.4)** — [worked examples](docs/EXAMPLES.md).

{quantprobe-1.4.0 → quantprobe-1.5.1}/quantprobe.egg-info/SOURCES.txt

@@ -2,6 +2,7 @@ LICENSE
 README.md
 pyproject.toml
 quantprobe/__init__.py
+quantprobe/auto.py
 quantprobe/cli.py
 quantprobe/dashboard.py
 quantprobe/detect.py

quantprobe-1.4.0/quantprobe/dashboard.py

@@ -1,179 +0,0 @@
-"""quantprobe dashboard — the law, live.
-
-One command: plans the best placement, launches llama-server with those flags, and serves a local
-page where you chat with the model while every reply is scored against the law's prediction —
-predicted vs measured tok/s, live. Colibri-style dashboard energy, but the star is falsifiability:
-the page is a running experiment, not a status screen.
-"""
-from __future__ import annotations
-import json, os, subprocess, sys, threading, time, urllib.request, webbrowser
-from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
-
-from . import runtime
-
-DASH_HTML = """<!doctype html><html><head><meta charset="utf-8">
-<meta name="viewport" content="width=device-width,initial-scale=1">
-<title>quantprobe — the law, live</title>
-<style>
-:root{--g:#fafaf7;--p:#fff;--ink:#16181d;--sub:#5c6066;--line:#e5e4df;--acc:#0f766e;
---accsoft:#e6f2f0;--warn:#b45309;--mono:Consolas,ui-monospace,monospace}
-@media(prefers-color-scheme:dark){:root{--g:#14161a;--p:#1c1f24;--ink:#e8e7e2;--sub:#9aa0a8;
---line:#2e3238;--acc:#35b8a6;--accsoft:#12332f;--warn:#e0a458}}
-body{margin:0;background:var(--g);color:var(--ink);font-family:"Segoe UI",system-ui,sans-serif}
-.wrap{max-width:980px;margin:0 auto;padding:22px 16px}
-h1{font-size:21px;margin:0 0 4px}.sub{color:var(--sub);font-size:13px;margin-bottom:16px}
-.law{font-family:var(--mono);font-size:12.5px;color:var(--acc);margin-bottom:14px}
-.cols{display:grid;grid-template-columns:340px 1fr;gap:16px}
-@media(max-width:760px){.cols{grid-template-columns:1fr}}
-.panel{background:var(--p);border:1px solid var(--line);border-radius:10px;padding:14px}
-.k{font-size:10.5px;text-transform:uppercase;letter-spacing:.09em;color:var(--sub);font-weight:600;margin-bottom:8px}
-.gauge{display:flex;gap:18px;align-items:baseline}
-.gauge .n{font-family:var(--mono);font-size:44px;font-weight:700;color:var(--acc);font-variant-numeric:tabular-nums}
-.gauge .lab{font-size:11px;color:var(--sub)}
-.delta{font-family:var(--mono);font-size:15px;font-weight:600}
-.tier{margin:8px 0}.tier .l{display:flex;justify-content:space-between;font-family:var(--mono);font-size:11px;color:var(--sub);margin-bottom:3px}
-.bar{height:11px;background:var(--g);border:1px solid var(--line);border-radius:6px;overflow:hidden}
-.fill{height:100%;background:var(--acc)}
-.log{font-family:var(--mono);font-size:11.5px;color:var(--sub);max-height:120px;overflow-y:auto;margin-top:8px}
-.chat{display:flex;flex-direction:column;height:520px}
-.msgs{flex:1;overflow-y:auto;display:flex;flex-direction:column;gap:8px;padding-bottom:8px}
-.m{max-width:85%;padding:8px 12px;border-radius:10px;font-size:14px;line-height:1.5;white-space:pre-wrap}
-.m.u{align-self:flex-end;background:var(--accsoft);color:var(--ink)}
-.m.a{align-self:flex-start;background:var(--g);border:1px solid var(--line)}
-.m .t{display:block;margin-top:6px;font-family:var(--mono);font-size:10.5px;color:var(--acc)}
-.row{display:flex;gap:8px}
-textarea{flex:1;resize:none;height:54px;padding:9px;font-size:14px;font-family:inherit;
-background:var(--g);color:var(--ink);border:1px solid var(--line);border-radius:8px}
-button{padding:0 18px;background:var(--acc);color:#fff;border:0;border-radius:8px;font-size:14px;font-weight:600;cursor:pointer}
-button:disabled{opacity:.5}
-.foot{margin-top:14px;font-size:11.5px;color:var(--sub)}.foot a{color:var(--acc)}
-</style></head><body><div class="wrap">
-<h1>The law, live</h1>
-<div class="sub">{{MODEL}} · {{PLACEMENT}} · every reply below is scored against the pre-registered prediction</div>
-<div class="law">tok/s = η(tier) × bandwidth ÷ active-bytes-per-token &nbsp;→&nbsp; predicted {{PRED}} tok/s for this box</div>
-<div class="cols">
-<div>
-<div class="panel"><div class="k">Prediction vs reality</div>
-  <div class="gauge">
-    <div><div class="n" id="pred">{{PRED}}</div><div class="lab">predicted tok/s</div></div>
-    <div><div class="n" id="meas">—</div><div class="lab">measured (running avg)</div></div>
-    <div class="delta" id="delta"></div>
-  </div>
-  <div class="log" id="log">each reply appends a data point…</div>
-</div>
-<div class="panel" style="margin-top:14px"><div class="k">Placement (planned)</div>{{TIERS}}
-<div class="foot">Flags: <span style="font-family:var(--mono)">{{FLAGS}}</span></div></div>
-<div class="panel" style="margin-top:14px"><div class="k">About</div>
-<div class="foot">This page is a running experiment: the tiered decode law predicted this machine's speed
-before the model loaded. Laws, probes, evidence: <a href="https://github.com/FedericoTs/quantprobe">repo</a>
-· <a href="https://x.com/federico_sciuca">@federico_sciuca</a></div></div>
-</div>
-<div class="panel chat"><div class="k">Chat — each reply is a measurement</div>
-  <div class="msgs" id="msgs"></div>
-  <div class="row"><textarea id="inp" placeholder="Say something… (Enter to send)"></textarea>
-  <button id="send">Send</button></div>
-</div>
-</div></div>
-<script>
-const hist=[],pts=[];const $=id=>document.getElementById(id);
-function esc(s){return s.replace(/&/g,"&amp;").replace(/</g,"&lt;")}
-async function send(){
-  const t=$("inp").value.trim(); if(!t)return;
-  $("inp").value=""; $("send").disabled=true;
-  hist.push({role:"user",content:t});
-  $("msgs").insertAdjacentHTML("beforeend",'<div class="m u">'+esc(t)+'</div>');
-  const hold=document.createElement("div");hold.className="m a";hold.textContent="…";
-  $("msgs").appendChild(hold);$("msgs").scrollTop=1e9;
-  try{
-    const r=await fetch("/api/chat",{method:"POST",headers:{"Content-Type":"application/json"},
-      body:JSON.stringify({messages:hist.slice(-12),max_tokens:256})});
-    const j=await r.json();
-    const msg=(j.choices&&j.choices[0].message.content)||"(no reply)";
-    const tps=j.timings?j.timings.predicted_per_second:null;
-    hist.push({role:"assistant",content:msg});
-    hold.innerHTML=esc(msg)+(tps?'<span class="t">'+tps.toFixed(1)+' tok/s · predicted {{PRED}}</span>':'');
-    if(tps){pts.push(tps);
-      const avg=pts.reduce((a,b)=>a+b,0)/pts.length,pred=parseFloat("{{PRED}}");
-      $("meas").textContent=avg.toFixed(1);
-      const d=(avg/pred-1)*100;
-      $("delta").textContent=(d>=0?"+":"")+d.toFixed(0)+"%";
-      $("delta").style.color=Math.abs(d)<25?"var(--acc)":"var(--warn)";
-      $("log").insertAdjacentHTML("afterbegin","#"+pts.length+": "+tps.toFixed(1)+" tok/s<br>");}
-  }catch(e){hold.textContent="error: "+e}
-  $("send").disabled=false;$("msgs").scrollTop=1e9;
-}
-$("send").onclick=send;
-$("inp").addEventListener("keydown",e=>{if(e.key==="Enter"&&!e.shiftKey){e.preventDefault();send();}});
-</script></body></html>"""
-
-
-def make_handler(upstream, page):
-    class H(BaseHTTPRequestHandler):
-        def log_message(self, *a):
-            pass
-
-        def do_GET(self):
-            self.send_response(200)
-            self.send_header("Content-Type", "text/html; charset=utf-8")
-            self.end_headers()
-            self.wfile.write(page.encode("utf-8"))
-
-        def do_POST(self):
-            if self.path != "/api/chat":
-                self.send_response(404); self.end_headers(); return
-            n = int(self.headers.get("Content-Length", 0))
-            body = json.loads(self.rfile.read(n) or b"{}")
-            body.setdefault("timings_per_token", True)
-            req = urllib.request.Request(upstream + "/v1/chat/completions",
-                                         data=json.dumps(body).encode(),
-                                         headers={"Content-Type": "application/json"})
-            try:
-                with urllib.request.urlopen(req, timeout=600) as r:
-                    data = r.read()
-                self.send_response(200)
-                self.send_header("Content-Type", "application/json")
-                self.end_headers()
-                self.wfile.write(data)
-            except Exception as e:
-                self.send_response(502); self.end_headers()
-                self.wfile.write(json.dumps({"error": str(e)}).encode())
-    return H
-
-
-def dashboard(a):
-    best, flags = runtime.best_flags(a)
-    binp = runtime.find_llama(a.llama_dir, "llama-server")
-    sport = a.server_port
-    cmd = [binp, "-m", a.gguf, "--port", str(sport)] + flags
-    print(f"[quantprobe] placement: {best[0]}  (predicted {best[1]:.1f} tok/s)")
-    print("[quantprobe] llama-server:", " ".join(cmd))
-    proc = subprocess.Popen(cmd, stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL)
-    upstream = f"http://127.0.0.1:{sport}"
-    for _ in range(120):
-        try:
-            urllib.request.urlopen(upstream + "/health", timeout=2)
-            break
-        except Exception:
-            if proc.poll() is not None:
-                raise SystemExit("llama-server exited during startup — check the model path/flags")
-            time.sleep(1)
-    tiers = ""
-    for name, cap, used in runtime.tier_view(a, best):
-        pct = min(100, used / cap * 100) if cap else 0
-        tiers += (f'<div class="tier"><div class="l"><span>{name}</span>'
-                  f'<span>{used:.1f} / {cap:.0f} GB</span></div>'
-                  f'<div class="bar"><div class="fill" style="width:{pct:.0f}%"></div></div></div>')
-    page = (DASH_HTML.replace("{{MODEL}}", os.path.basename(a.gguf))
-            .replace("{{PLACEMENT}}", best[0]).replace("{{PRED}}", f"{best[1]:.1f}")
-            .replace("{{FLAGS}}", " ".join(flags)).replace("{{TIERS}}", tiers))
-    srv = ThreadingHTTPServer(("127.0.0.1", a.port), make_handler(upstream, page))
-    url = f"http://127.0.0.1:{a.port}"
-    print(f"[quantprobe] dashboard: {url}  (Ctrl-C stops both)")
-    if not a.no_open:
-        threading.Timer(1.0, lambda: webbrowser.open(url)).start()
-    try:
-        srv.serve_forever()
-    except KeyboardInterrupt:
-        pass
-    finally:
-        proc.terminate()