tools/imcompare.py

#!/usr/bin/env python3
"""Compare imatrix files directly.

Two things it answers:
  1) DEAD EXPERTS  - using the authoritative `.counts` array (how many times each
     expert was actually routed to). counts[i]==0 => expert i never used.
  2) AGREEMENT     - cosine similarity between the SAME tensor in two imatrix files.
     cos ~ 1.000 => the two runs produce an interchangeable importance vector,
     i.e. the extra data changed nothing the quantiser can see.
"""
import sys, numpy as np
from gguf import GGUFReader

def load(path):
    r = GGUFReader(path)
    vals, cnts = {}, {}
    for t in r.tensors:
        arr = np.array(t.data, dtype=np.float64).reshape(-1)
        if t.name.endswith(".in_sum2"):
            vals[t.name[:-len(".in_sum2")]] = arr
        elif t.name.endswith(".counts"):
            cnts[t.name[:-len(".counts")]] = arr
    meta = {f.name: f for f in r.fields.values()}
    nchunk = None
    if "imatrix.chunk_count" in meta:
        nchunk = meta["imatrix.chunk_count"].contents()
    return vals, cnts, nchunk

paths = sys.argv[1:]
data = {}
for p in paths:
    lbl = p.split("/")[-1].replace(".gguf", "")
    data[lbl] = load(p)
    print("%-34s chunks=%s tensors=%d" % (lbl, data[lbl][2], len(data[lbl][0])))

print()
print("=" * 92)
print("1) DEAD EXPERTS  (counts[i] == 0  =>  expert never routed to)")
print("=" * 92)
print("%-34s %14s %16s %14s" % ("imatrix", "expert tensors", "total experts", "DEAD"))
for lbl, (vals, cnts, _) in data.items():
    ntens = tot = dead = 0
    for name, c in cnts.items():
        if "exps" not in name:
            continue
        ntens += 1
        tot += c.size
        dead += int((c == 0).sum())
    print("%-34s %14d %16d %14d" % (lbl, ntens, tot, dead))

print()
print("=" * 92)
print("2) AGREEMENT between imatrices  (cosine similarity, per tensor, then summarised)")
print("=" * 92)
labels = list(data)
base = labels[0]
for other in labels[1:]:
    va, vb = data[base][0], data[other][0]
    common = sorted(set(va) & set(vb))
    sims, worst = [], []
    for k in common:
        a, b = va[k], vb[k]
        if a.shape != b.shape:
            continue
        na, nb = np.linalg.norm(a), np.linalg.norm(b)
        if na == 0 or nb == 0:
            continue
        s = float(np.dot(a, b) / (na * nb))
        sims.append(s)
        worst.append((s, k))
    sims = np.array(sims)
    worst.sort()
    print()
    print("%s   vs   %s        (%d tensors compared)" % (base, other, len(sims)))
    print("   mean cos = %.6f      median = %.6f      min = %.6f" % (
        sims.mean(), np.median(sims), sims.min()))
    print("   tensors with cos < 0.99: %d" % int((sims < 0.99).sum()))
    print("   5 worst:")
    for s, k in worst[:5]:
        print("      %.6f  %s" % (s, k))