Frank
d4ce76b68b
Every comment that invoked a 'law' (50 word-boundary sites across 28 tracked files) now states what is going on and why, in terms that transfer to any project: gate twins drift until one side greenlights what the other refuses; a tag re-points while a digest pins; a restrictive engine behavior belongs to the runtime that imposed it and must be re-tested, not inherited; a speculative head is tensors in the file, not a flag; snapshot links dangle where the container resolves them; exit 0 is a clean stop to systemd, so a self-exiting daemon on on-failure goes dark reporting success. Event narratives and dates are gone from these comments; the mechanics carry the argument. Behavior-identical: bash -n, py_compile, yaml parse, full bats suite 408/408 green. One symbol rename rides along (stack script local scale_law -> scale_contract, with its log text).
118 lines
4.5 KiB
Python
Executable File
118 lines
4.5 KiB
Python
Executable File
#!/usr/bin/env python3
|
|
"""derive-protected-experts.py — rewrite the recipe's protected-expert section
|
|
from a calibration recording. RUNS ON DANKSERVER (venv: numpy).
|
|
|
|
The selection rule (docs/MOE-W4.md "Protected experts", restated in the recipe):
|
|
an expert whose share of its layer's down-input energy is at least 3% AND
|
|
at least three times its share of the routing is kept bf16; the count must
|
|
be EVEN on each rank's half of the layer (the four-bit GEMM takes the rest
|
|
as two equal tables by parity) — an odd half gains its next most energetic
|
|
non-qualifying expert.
|
|
|
|
Data: the calibration dir's per-expert down Grams,
|
|
gram.r<rank>.blk.L.ffn_down_exps.<e>.bin — 32-byte RADGRAM2 header
|
|
(magic, k, rank, rows, domain) then k*k f32, named with the checkpoint's
|
|
GLOBAL expert numbering (arch/common/rad_block_moe_fp8.h: ge =
|
|
expert_base + e). rows (header) is the expert's routing count; the trace
|
|
of the Gram is its down-input energy. Rank halves: 0-255, 256-511
|
|
(contiguous expert_base split at TP=2).
|
|
|
|
Output: the cast-rule lines (recipe format) on stdout, a report on stderr.
|
|
|
|
Usage: derive-protected-experts.py <calib dir> [--min-share 0.03]
|
|
[--min-route-mult 3.0]
|
|
"""
|
|
import argparse
|
|
import glob
|
|
import os
|
|
import re
|
|
import struct
|
|
import sys
|
|
|
|
import numpy as np
|
|
|
|
NAME_RE = re.compile(r"^gram\.r(\d)\.blk\.(\d+)\.ffn_down_exps\.(\d+)\.bin$")
|
|
|
|
|
|
def read_gram(path):
|
|
"""(k, rank, rows, trace) from a RADGRAM2 file (full matrix on disk)."""
|
|
with open(path, "rb") as f:
|
|
hdr = f.read(32)
|
|
magic, k, rank, rows, domain = struct.unpack("<8sIIQI", hdr[:28])
|
|
if magic != b"RADGRAM2":
|
|
raise SystemExit(f"not a RADGRAM2 file: {path}")
|
|
n = k * k
|
|
a = np.frombuffer(f.read(n * 4), dtype=np.float32)
|
|
if a.size != n:
|
|
raise SystemExit(f"short gram file: {path}")
|
|
tr = float(a.reshape(k, k).diagonal().sum())
|
|
return k, rank, int(rows), tr
|
|
|
|
|
|
def main():
|
|
ap = argparse.ArgumentParser()
|
|
ap.add_argument("calib_dir")
|
|
ap.add_argument("--min-share", type=float, default=0.03)
|
|
ap.add_argument("--min-route-mult", type=float, default=3.0)
|
|
a = ap.parse_args()
|
|
|
|
# layer -> global expert -> (rank, rows, energy)
|
|
data = {}
|
|
for path in sorted(glob.glob(os.path.join(
|
|
a.calib_dir, "gram.r*.ffn_down_exps.*.bin"))):
|
|
m = NAME_RE.match(os.path.basename(path))
|
|
if not m:
|
|
continue
|
|
rank, layer, expert = int(m.group(1)), int(m.group(2)), int(m.group(3))
|
|
k, _, rows, tr = read_gram(path)
|
|
data.setdefault(layer, {})[expert] = (rank, rows, tr)
|
|
|
|
if not data:
|
|
raise SystemExit("no per-expert down grams found")
|
|
|
|
lines, report = [], []
|
|
total = 0
|
|
for layer in sorted(data):
|
|
experts = data[layer]
|
|
tot_e = sum(v[2] for v in experts.values())
|
|
tot_r = sum(v[1] for v in experts.values())
|
|
if tot_e <= 0 or tot_r <= 0:
|
|
report.append(f"blk.{layer}: empty recording, skipped")
|
|
continue
|
|
qual, pool = [], []
|
|
for e, (rank, rows, tr) in experts.items():
|
|
es, rs = tr / tot_e, rows / max(tot_r, 1)
|
|
rec = (e, rank, es, rs)
|
|
if es >= a.min_share and es >= a.min_route_mult * rs:
|
|
qual.append(rec)
|
|
elif es > 0:
|
|
pool.append(rec)
|
|
chosen = {r[0]: (r[2], r[3]) for r in qual}
|
|
for half in (0, 1): # rank halves by the contiguous expert_base split
|
|
n_half = sum(1 for e in chosen if e // 256 == half)
|
|
if n_half % 2:
|
|
pool.sort(key=lambda r: -r[2])
|
|
for cand in pool:
|
|
if cand[1] == half and cand[0] not in chosen:
|
|
chosen[cand[0]] = (cand[2], cand[3])
|
|
report.append(
|
|
f"blk.{layer}: half {half} odd -> added e{cand[0]} "
|
|
f"(energy {cand[2]*100:.2f}%)")
|
|
break
|
|
for e in sorted(chosen):
|
|
es, rs = chosen[e]
|
|
lines.append(f"blk.{layer}.ffn_*_exps.{e}.weight"
|
|
f" cast dtype=bf16")
|
|
report.append(f"blk.{layer} e{e:3d} (rank {e // 256}): "
|
|
f"energy {es*100:5.2f}%, routing {rs*100:5.3f}%, "
|
|
f"ratio {es/max(rs,1e-12):6.1f}")
|
|
total += len(chosen)
|
|
|
|
print("\n".join(lines))
|
|
print("\n".join(report), file=sys.stderr)
|
|
print(f"total protected: {total}", file=sys.stderr)
|
|
|
|
|
|
if __name__ == "__main__":
|
|
main()
|