File size: 4,739 Bytes
8604228 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 | #!/usr/bin/env python3
# OB-1 stage 1, step 2: static popularity-ranked resident sets.
#
# Reads ONE route log (frozen: the RS053 20b-prose-a run only -- ranks come
# from the PROSE corpus alone, never mixed with code, per the OB-1 task's
# honesty law that ranking and acceptance corpora must differ), computes
# per-layer expert usage counts c_l(e) exactly as RS053 stage 3's M1
# (research/rs053/rs053-metrics.py compute_m1_m2: same file-order mask +
# token_index-monotonic verification, same bincount), then for each K in
# {16,8,4} freezes the top-K experts per layer by usage count, ties broken
# by LOWER EXPERT ID (same tie-break convention RS053 adopted from
# an earlier internal convention, prereg section 5.0/deviation 2).
#
# Usage: resident_sets.py <route_log> <E> <L> <budget> <out_json>
import sys, json, hashlib
import numpy as np
def sha256_of(path):
h = hashlib.sha256()
with open(path, "rb") as f:
for chunk in iter(lambda: f.read(1 << 20), b""):
h.update(chunk)
return h.hexdigest()
def load_route_log(path, E, L, budget):
# Same loader as research/rs053/rs053-metrics.py load_route_log: the
# route log is grouped by ubatch, not one contiguous block per layer,
# so each layer's sequence is built with a boolean mask (preserves
# file order, which is ubatch-ascending) and verified token_index
# comes out exactly 0..budget-1 in order before being trusted.
data = np.loadtxt(path, delimiter=",", dtype=np.int64)
expected_rows = L * budget
if data.shape != (expected_rows, 6):
raise SystemExit("SHAPE MISMATCH: got %r expected (%d, 6)" % (data.shape, expected_rows))
layer_col = data[:, 0]
token_col = data[:, 1]
expert_ids = data[:, 2:6] # k=4 columns
if int((expert_ids < 0).sum()) != 0 or int((expert_ids >= E).sum()) != 0:
raise SystemExit("EXPERT ID OUT OF RANGE")
expected_tok = np.arange(budget)
ids_by_layer = np.empty((L, budget, 4), dtype=np.int64)
for l in range(L):
mask = layer_col == l
n = int(mask.sum())
if n != budget:
raise SystemExit("LAYER %d: got %d rows, expected budget=%d" % (l, n, budget))
tok_l = token_col[mask]
if not np.array_equal(tok_l, expected_tok):
raise SystemExit("ORDER ASSUMPTION VIOLATED: layer %d" % l)
ids_by_layer[l] = expert_ids[mask]
return ids_by_layer
def main():
if len(sys.argv) != 6:
raise SystemExit("usage: resident_sets.py <route_log> <E> <L> <budget> <out_json>")
route_log, E, L, budget, out_json = sys.argv[1:6]
E = int(E); L = int(L); budget = int(budget)
K_VALUES = [16, 8, 4]
route_log_sha256 = sha256_of(route_log)
ids_by_layer = load_route_log(route_log, E, L, budget)
total_per_layer = budget * 4
hist = np.zeros((L, E), dtype=np.int64)
for l in range(L):
hist[l] = np.bincount(ids_by_layer[l].reshape(-1), minlength=E)
row_sums = hist.sum(axis=1)
if not np.all(row_sums == total_per_layer):
raise SystemExit("HISTOGRAM SUM MISMATCH")
resident_sets = {}
for K in K_VALUES:
per_layer = {}
for l in range(L):
counts = hist[l]
# rank by (-count, expert_id): descending count, ties broken by
# lower expert id ascending
order = sorted(range(E), key=lambda e: (-int(counts[e]), e))
top = sorted(order[:K]) # store ascending by id for readability
per_layer[str(l)] = top
resident_sets[str(K)] = per_layer
PER_EXPERT_BYTES_PER_LAYER = 13253760 # cross-checked vs research/ob1/EXPERT-MANIFEST-20B.sha256 total/(L*E)
resident_bytes_per_K = {
str(K): K * L * PER_EXPERT_BYTES_PER_LAYER for K in K_VALUES
}
out = {
"ranking_corpus": "prose (RS053 20b-prose-a route.log ONLY; never mixed with code)",
"source_route_log": route_log,
"source_route_log_sha256": route_log_sha256,
"E": E,
"L": L,
"budget": budget,
"tie_break": "lower expert id",
"K_values": K_VALUES,
"per_expert_bytes_per_layer": PER_EXPERT_BYTES_PER_LAYER,
"resident_expert_pool_bytes_per_K": resident_bytes_per_K,
"histogram_c_l_e": hist.astype(int).tolist(),
"resident_sets": resident_sets,
}
with open(out_json, "w") as f:
json.dump(out, f, sort_keys=True, indent=1)
f.write("\n")
print("ROUTE_LOG=%s SHA256=%s" % (route_log, route_log_sha256))
print("E=%d L=%d budget=%d" % (E, L, budget))
for K in K_VALUES:
print("K=%d resident_expert_pool_bytes=%d" % (K, resident_bytes_per_K[str(K)]))
print("OUT=%s" % out_json)
if __name__ == "__main__":
main()
|