quantprobe 1.6.1__tar.gz → 1.6.2__tar.gz

{quantprobe-1.6.1/quantprobe.egg-info → quantprobe-1.6.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: quantprobe
-Version: 1.6.1
+Version: 1.6.2
 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

{quantprobe-1.6.1 → quantprobe-1.6.2}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "quantprobe"
-version = "1.6.1"
+version = "1.6.2"
 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-1.6.1 → quantprobe-1.6.2}/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__ = "1.6.1"
+__version__ = "1.6.2"

{quantprobe-1.6.1 → quantprobe-1.6.2}/quantprobe/auto.py

@@ -24,17 +24,51 @@ MODEL_REPOS = {
     "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),
+    "qwen3-235b": ("unsloth/Qwen3-235B-A22B-GGUF", 235.1, 22, 7.5, True),      # total exact (HF safetensors)
+    "glm-4.7":    ("unsloth/GLM-4.7-GGUF", 358.3, 32, 8, True),                # total exact; a/ne [est]
+    "glm-744b":   ("unsloth/GLM-5.2-GGUF", 753.3, 32, 8, True),                # GLM-5.2, total exact 753.3B; a/ne [est]
+    "kimi-k2.6":  ("unsloth/Kimi-K2.6-GGUF", 1058.6, 32, 6, True),             # total exact; a/ne [est]
+    "gpt-oss-120b": ("unsloth/gpt-oss-120b-GGUF", 120.4, 5.1, 1.8, True),      # total exact, A5.1B official
+    "llama-70b":  ("unsloth/Llama-3.3-70B-Instruct-GGUF", 70.6, 70.6, 70.6, False),
+    "deepseek-16b": ("bartowski/DeepSeek-Coder-V2-Lite-Instruct-GGUF", 15.7, 2.4, 1.3, True),
 }
 
 
 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",
+    """[(path, size_bytes)] for a HF repo, via the public tree API (recursive: big repos
+    keep quants in subfolders). Split multi-part files (-00001-of-000NN) are grouped into
+    ONE logical entry: path = first part, size = sum of parts — so effective-bits stays honest."""
+    import re
+    req = urllib.request.Request(f"https://huggingface.co/api/models/{repo}/tree/main?recursive=true",
                                  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()]
+        raw = [(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()]
+    groups, singles = {}, []
+    for path, size in raw:
+        m = re.match(r"^(.*)-(\d{5})-of-(\d{5})\.gguf$", path)
+        if m:
+            g = groups.setdefault(m.group(1), {"total": 0, "first": None, "n": int(m.group(3))})
+            g["total"] += size
+            if m.group(2) == "00001":
+                g["first"] = path
+        else:
+            singles.append((path, size))
+    for g in groups.values():
+        if g["first"]:
+            singles.append((g["first"], g["total"]))
+    return singles
+
+
+def split_parts(first_part):
+    """All part paths of a split file, derived from the -00001-of-000NN first part."""
+    import re
+    m = re.match(r"^(.*)-(\d{5})-of-(\d{5})\.gguf$", first_part)
+    if not m:
+        return [first_part]
+    base, n = m.group(1), int(m.group(3))
+    return [f"{base}-{i:05d}-of-{n:05d}.gguf" for i in range(1, n + 1)]
 
 
 WIKI_URL = "https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip"
@@ -136,9 +170,19 @@ def run(a):
             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")
+        avail = sorted(set(round(s * 8 / (t * 1e9), 1) for _, s in files))
+        raise SystemExit(
+            f"no ready-to-run quant in {repo} for this machine yet (available bit-levels: {avail}).\n"
+            "  Big new models often publish low-bit quants days after release. Meanwhile:\n"
+            "  - predictions for any bits now:  quantprobe plan --model <preset> --bits 2.5\n"
+            + ("  - or build your own from the high-precision source: re-run with --custom\n"
+               if pick_source(files, t) else ""))
     scored.sort()
     _, _, path, size, bits, best = scored[0]
+    if abs(bits - want_bits) > 2:
+        print(f"\n[quantprobe auto] note: no ~{want_bits:g}-bit-class file in {repo} yet - closest "
+              f"available is {bits:.1f}-bit. Low-bit quants for new models often land days after "
+              f"release; plan/optimize predict any bits meanwhile.")
 
     if getattr(a, "custom", False) and want_bits >= 3.5 and not getattr(a, "force_custom", False):
         print(f"\n[quantprobe auto --custom] this machine doesn't need the surgery: the optimizer")
@@ -179,20 +223,24 @@ def run(a):
             a.gguf = out; a.bits = None
             runtime.run(a)
         return out
+    parts = split_parts(path)
     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"  {path}  ({size/1e9:.1f} GB, {bits:.2f} effective bits"
+          + (f", {len(parts)} parts" if len(parts) > 1 else "") + ")")
     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
+    # 3. fetch it (resumable; all parts when split), 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)")
+    for i, part in enumerate(parts):
+        if len(parts) > 1:
+            print(f"[quantprobe auto] part {i+1}/{len(parts)}: {os.path.basename(part)}")
+        if not fetchmod.fetch(repo, dest, part, fetchmod.token()):
+            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}")

{quantprobe-1.6.1 → quantprobe-1.6.2}/quantprobe/cli.py

@@ -2,7 +2,7 @@
 from __future__ import annotations
 import argparse
 
-from .plan import numlist
+from .plan import numlist, MODELS as _MODELS
 
 
 def main():
@@ -30,7 +30,7 @@ def main():
     p.add_argument("--dry-run", action="store_true")
 
     q = sub.add_parser("plan", help="evaluate every bit/tier placement, predict tok/s (Law 4)")
-    q.add_argument("--model", default=None, help="preset: qwen3-30b deepseek-16b gemma-12b mistral-7b glm-air glm-744b")
+    q.add_argument("--model", default=None, help="preset: " + " ".join(sorted(_MODELS)))
     q.add_argument("--machine", default=None, help="preset: 2016-xmp 2016 gaming ddr5 colibri")
     q.add_argument("--bits", type=float, default=None, help="effective bits/weight (default: read from --gguf, else 2.5)")
     q.add_argument("--total", type=float, help="total params (B)")
@@ -93,7 +93,7 @@ def main():
 
     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", nargs="?", default=None,
-                    help="model preset (qwen3-30b, qwen3-coder, glm-air, laguna-s, gemma-12b, mistral-7b) or a HF GGUF repo id; omit it and quantprobe ASKS (interactive)")
+                    help="model preset (qwen3-30b, qwen3-coder, glm-air, laguna-s, gemma-12b, mistral-7b, qwen3-235b, glm-4.7, glm-744b, kimi-k2.6, gpt-oss-120b, llama-70b, deepseek-16b) or a HF GGUF repo id; omit it and quantprobe ASKS (interactive)")
     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)

{quantprobe-1.6.1 → quantprobe-1.6.2}/quantprobe/optimize.py

@@ -13,7 +13,8 @@ Levers and their measured gates:
   prune     REAP-class 50% expert pruning (total x0.82, file shrinks, active bytes UNCHANGED).
             Measured +39% out-of-domain perplexity (pre-registration #8) — domain-specialized,
             never ranked first without --allow-prune.
-  hardware  euro-priced deltas from the projections table: XMP (free), +16GB RAM (~40), NVMe (~100).
+  hardware  upgrade deltas from the projections table: XMP (free), +16GB RAM, NVMe. Ranked by
+            relative cost internally; no currency figures are shown (region- and time-dependent).
 """
 from __future__ import annotations
 from . import plan as planmod
@@ -28,9 +29,9 @@ REALIZE = {
 }
 HW_DELTAS = [
     ("as-is", 0, {}),
-    ("+16GB RAM (~40 EUR used)", 40, {"rc": +16}),
-    ("NVMe SSD (~100 EUR)", 100, {"db_min": 3.5}),
-    ("+RAM & NVMe (~140 EUR)", 140, {"rc": +16, "db_min": 3.5}),
+    ("+16GB RAM (used)", 40, {"rc": +16}),
+    ("NVMe SSD", 100, {"db_min": 3.5}),
+    ("+RAM & NVMe", 140, {"rc": +16, "db_min": 3.5}),
 ]
 
 
@@ -121,7 +122,7 @@ def run(a):
     print(f"  {headline}\n")
     for i, r in enumerate(ranked[:6]):
         star = "*" if i == 0 else " "
-        euro = "free" if r["euro"] == 0 else f"~{r['euro']}EUR"
+        euro = "free" if r["euro"] == 0 else "hw upgrade"
         print(f"  {star} {r['tps']:6.1f} tok/s  quality x{r['q']:.2f}  {euro:>7s}  {r['desc']}")
         for tag in r["tags"]:
             print(f"                [{tag}]")

{quantprobe-1.6.1 → quantprobe-1.6.2}/quantprobe/plan.py

@@ -17,7 +17,11 @@ MODELS = {
     "gemma-12b":  dict(t=11.9, a=11.9, ne=11.9, moe=False, kvp=65536,  hint="Gemma 4 12B"),             # [est] SWA: long-ctx slope from global layers only
     "mistral-7b": dict(t=7.2,  a=7.2,  ne=7.2,  moe=False, kvp=131072, hint="Mistral 7B"),              # 32L x 8KV x 128d (exact)
     "glm-air":    dict(t=110,  a=12,   ne=2.7,  moe=True,  kvp=94208,  hint="GLM-4.5-Air 106B"),        # [est]
-    "glm-744b":   dict(t=744,  a=32,   ne=8,    moe=True,  kvp=188416, hint="GLM-5.2 744B"),            # [est]
+    "glm-744b":   dict(t=753.3, a=32,  ne=8,    moe=True,  kvp=188416, hint="GLM-5.2 (753B)"),          # total exact (HF safetensors); a/ne/kvp [est]
+    "qwen3-235b": dict(t=235.1, a=22,  ne=7.5,  moe=True,  kvp=192512, hint="Qwen3-235B-A22B"),         # total exact; kvp: 94L x 4KV x 128d [est]
+    "kimi-k2.6":  dict(t=1058.6, a=32, ne=6,    moe=True,  kvp=70272,  hint="Kimi K2.6 (1T MLA)"),      # total exact; kvp: 61L x 576 MLA latent [est]
+    "gpt-oss-120b": dict(t=120.4, a=5.1, ne=1.8, moe=True, kvp=73728,  hint="gpt-oss-120b"),            # total+active official; kvp: 36L x 8KV x 64d [est]
+    "llama-70b":  dict(t=70.6, a=70.6, ne=70.6, moe=False, kvp=327680, hint="Llama-3.3-70B"),           # dense; kvp: 80L x 8KV x 128d
 }
 DEFAULT_KVP = 98304          # custom models without --kv-per-pos: typical GQA mid-size (Qwen3-30B class)
 ETA_KV = 0.70                # KV-read efficiency. Single-point calibration: measured tg32 d0->d16384

{quantprobe-1.6.1 → quantprobe-1.6.2/quantprobe.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: quantprobe
-Version: 1.6.1
+Version: 1.6.2
 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