Qwen3.8-Flash-Next on any consumer hardware: one-click install for Windows / Linux. Strata inference engine, OpenAI/Anthropic API on localhost, optional image input.
C++
"""proxy_ldl.py - the Hessian-weighted output error of plain, block-feedback (LDL) and agentionai's encodes, no engine needed.
For some experts of the chosen layers: W' = ORIGINAL weights in the file's rotated basis, shipped = agentionai's decoded blocks,
plain = our plain-MSE encode, ldl = block feedback with the second-moment matrix H of --dump-dir. Errors are measured with
H_test (--test-dump-dir, a different part of the calibration set; default = H itself) as
tr(E H E^T) / tr(W H W^T) (the expected squared output error for inputs with that second moment)
and unweighted ||E||^2 / ||W||^2, in dB, mean over experts.
"""
from __future__ import annotations
import argparse
import mmap
import pathlib
import sys
import numpy as np
HERE = pathlib.Path(__file__).resolve().parent
sys.path.insert(0, str(HERE.parent))
sys.path.insert(0, str(HERE))
import blockldl # noqa: E402
import gguf_reader # noqa: E402
import tqenc # noqa: E402
from build_swap import bf16, load_hessian # noqa: E402
def main() -> int:
ap = argparse.ArgumentParser()
ap.add_argument("--gguf", required=True)
ap.add_argument("--orig", required=True)
ap.add_argument("--dump-dir", required=True)
ap.add_argument("--test-dump-dir", default=None)
ap.add_argument("--layers", default="4,16,28,40")
ap.add_argument("--experts", default="0,111,301")
ap.add_argument("--roles", default="gate,down")
ap.add_argument("--damps", default="0.01")
ap.add_argument("--threads", type=int, default=4)
ap.add_argument("--gpu", action="store_true")
a = ap.parse_args()
g = gguf_reader.GGUFFile(a.gguf)
fh = open(a.gguf, "rb")
mm = mmap.mmap(fh.fileno(), 0, access=mmap.ACCESS_READ)
by = {t.name: t for t in g.tensors}
sw = [int(x) for x in g.metadata["prism.hadamard.sign_widths"]]
sv = np.array(g.metadata["prism.hadamard.sign_values"], dtype=np.float32)
signs, at = {}, 0
for w in sw:
signs[w] = sv[at:at + w]
at += w
damps = [float(x) for x in a.damps.split(",")]
res = {}
for l in [int(x) for x in a.layers.split(",")]:
gu_raw = np.memmap(pathlib.Path(a.orig) / ("L%d_gate_up.bf16" % l), dtype=np.uint16, mode="r").reshape(512, 1280, 2560)
dn_raw = np.memmap(pathlib.Path(a.orig) / ("L%d_down.bf16" % l), dtype=np.uint16, mode="r").reshape(512, 2560, 640)
for role in a.roles.split(","):
kind = "dn" if role == "down" else "gu"
H = load_hessian(a.dump_dir, l, kind)
Ht = load_hessian(a.test_dump_dir, l, kind) if a.test_dump_dir else H
fbs = {d: blockldl.feedback_matrices(blockldl.hessian_factor(H, d)) for d in damps}
for e in [int(x) for x in a.experts.split(",")]:
W = bf16(np.asarray(gu_raw[e][:640] if role == "gate" else gu_raw[e][640:] if role == "up" else dn_raw[e]))
ti = by["blk.%d.ffn_%s_exps.weight" % (l, role)]
t = tqenc.TYPE_BY_NAME[ti.type_name]
n_in, n_out = ti.shape[0], ti.shape[1]
rb = n_in // 128 * tqenc.block_bytes(t)
raw = np.frombuffer(mm, dtype=np.uint8, count=n_out * rb, offset=g.data_start + ti.offset + e * n_out * rb)
Wr = tqenc.rotate(W, signs[n_in]).astype(np.float64)
cands = {"shipped": tqenc.decode(raw, t).reshape(n_out, n_in).astype(np.float64)}
kw = dict(threads=a.threads) if not a.gpu else dict(gpu=True)
pk, _, _, _ = tqenc.encode(Wr.astype(np.float32), t, iters=1, **kw)
cands["plain"] = tqenc.decode(pk, t).reshape(n_out, n_in).astype(np.float64)
for d in damps:
_, dec = blockldl.encode_ldl(Wr.astype(np.float32), t, fbs[d], gpu=a.gpu, threads=a.threads)
cands["ldl damp=%g" % d] = dec.astype(np.float64)
den_u = float(np.sum(Wr ** 2))
den_h = float(np.trace(Wr @ Ht @ Wr.T))
for name, Q in cands.items():
E = Wr - Q
res.setdefault((l, role, name), []).append((float(np.sum(E ** 2) / den_u), float(np.trace(E @ Ht @ E.T) / den_h)))
print("layer %d %s expert %d done" % (l, role, e), flush=True)
names = ["shipped", "plain"] + ["ldl damp=%g" % d for d in damps]
print("\nrelative error in dB (higher is better): unweighted | H-weighted (%s)" % ("held-out H" if a.test_dump_dir else "same H"))
tot = {n: [[], []] for n in names}
for l in [int(x) for x in a.layers.split(",")]:
for role in a.roles.split(","):
line = "L%-2d %-5s" % (l, role)
for n in names:
v = np.array(res[(l, role, n)])
line += " %-13s %5.2f | %5.2f" % (n, -10 * np.log10(v[:, 0].mean()), -10 * np.log10(v[:, 1].mean()))
tot[n][0].append(v[:, 0].mean())
tot[n][1].append(v[:, 1].mean())
print(line)
print("ALL " + " ".join("%-13s %5.2f | %5.2f" % (n, -10 * np.log10(np.mean(tot[n][0])), -10 * np.log10(np.mean(tot[n][1]))) for n in names))
return 0
if __name__ == "__main__":
sys.exit(main())