Skip to content

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++

STRATA_ACT_DUMP: rarer flush with the covariance on; merge_dumps keeps it

Niko1221431803abb90e
diff --git a/src/kernels/cpu/native_expert.cpp b/src/kernels/cpu/native_expert.cpp
index 0e8f92c4e..7c491832c 100644
--- a/src/kernels/cpu/native_expert.cpp
+++ b/src/kernels/cpu/native_expert.cpp
@@ -206,7 +206,7 @@ void act_dump_record(int layer, int kind, const float* v, int64_t n) {
for (int64_t j = 0; j < n; ++j) row[j] += vi * v[j];
}
++a.count;
- if (++d.records % 2000 == 0) {
+ if (++d.records % (d.cov ? 400000 : 2000) == 0) {
d.mu.unlock();
ActDump::act_dump_flush();
d.mu.lock();
diff --git a/tools/trellis/merge_dumps.py b/tools/trellis/merge_dumps.py
new file mode 100644
index 000000000..629f709d0
--- /dev/null
+++ b/tools/trellis/merge_dumps.py
@@ -0,0 +1,32 @@
+"""merge_dumps.py - add STRATA_ACT_DUMP files (int64 n, int64 count, double sum x^2 [n]) of several runs into one directory, and
+print the per-layer summary.
+
+ python3 tools/trellis/merge_dumps.py OUT_DIR IN_DIR [IN_DIR ...]
+"""
+import pathlib
+import sys
+
+import numpy as np
+
+out = pathlib.Path(sys.argv[1])
+out.mkdir(parents=True, exist_ok=True)
+names = sorted({p.name for d in sys.argv[2:] for p in pathlib.Path(d).glob("act_L*_*.bin") if p.name.startswith("act_L") and p.stat().st_size < 400_000_000})
+for name in names:
+ tot, cnt, n0, cov = None, 0, None, None
+ for d in sys.argv[2:]:
+ p = pathlib.Path(d) / name
+ if not p.exists():
+ continue
+ raw = p.read_bytes()
+ n, c = np.frombuffer(raw[:16], dtype=np.int64)
+ sq = np.frombuffer(raw[16:16 + 8 * n], dtype=np.float64)
+ tot = sq.copy() if tot is None else tot + sq
+ if len(raw) >= 16 + 8 * n + 8 * n * n:
+ cv = np.frombuffer(raw[16 + 8 * n:16 + 8 * n + 8 * n * n], dtype=np.float64)
+ cov = cv.copy() if cov is None else cov + cv
+ cnt += int(c)
+ n0 = int(n)
+ (out / name).write_bytes(np.array([n0, cnt], dtype=np.int64).tobytes() + tot.astype(np.float64).tobytes()
+ + (cov.tobytes() if cov is not None else b""))
+ m = tot / cnt
+ print("%-16s n %d count %d mean E[x^2] %.4g cv over channels %.3f max/mean %.1f" % (name, n0, cnt, m.mean(), m.std() / m.mean(), m.max() / m.mean()))
diff --git a/tools/trellis/proxy_diag_error.py b/tools/trellis/proxy_diag_error.py
new file mode 100644
index 000000000..49968d1d1
--- /dev/null
+++ b/tools/trellis/proxy_diag_error.py
@@ -0,0 +1,101 @@
+"""proxy_diag_error.py - the diagonal-Hessian output error of the swapped layers, no engine needed.
+
+For some experts of the chosen layers: W' = the ORIGINAL weights in the file's rotated basis; shipped = agentionai's decoded
+blocks; ours = plain-MSE encode, and activation-weighted encodes (h = E[x^2]^p from a STRATA_ACT_DUMP directory).
+Reported: the unweighted relative error ||W' - Q||^2 / ||W'||^2 and the diagonal-Hessian error
+sum_i h_i ||W'_:,i - Q_:,i||^2 / sum_i h_i ||W'_:,i||^2 (the expected squared output error under independent channels),
+per layer and role, the mean over experts, in dB.
+"""
+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 gguf_reader # noqa: E402
+import tqenc # noqa: E402
+from build_swap import bf16, load_h # 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("--layers", default="4,16,28,40")
+ ap.add_argument("--experts", default="0,37,111,200,301,420")
+ ap.add_argument("--powers", default="1.0,0.5")
+ ap.add_argument("--threads", type=int, default=16)
+ 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
+ powers = [float(x) for x in a.powers.split(",")]
+ experts = [int(x) for x in a.experts.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)
+ hs = {"gu": load_h(a.dump_dir, l, "gu", 1.0)[0], "dn": load_h(a.dump_dir, l, "dn", 1.0)[0]}
+ for e in experts:
+ gu = bf16(np.asarray(gu_raw[e]))
+ dn = bf16(np.asarray(dn_raw[e]))
+ for role, W, kind in (("gate", gu[:640], "gu"), ("up", gu[640:], "gu"), ("down", dn, "dn")):
+ 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)
+ ship = tqenc.decode(raw, t).reshape(n_out, n_in)
+ Wr = tqenc.rotate(W, signs[n_in])
+ h = hs[kind]
+ cands = {"shipped": ship}
+ kw = dict(iters=1, ncand=1)
+ if a.gpu:
+ kw["gpu"] = True
+ else:
+ kw["threads"] = a.threads
+ pk, _, _, _ = tqenc.encode(Wr, t, **kw)
+ cands["plain"] = tqenc.decode(pk, t).reshape(n_out, n_in)
+ for p in powers:
+ hp = (h ** p / (h ** p).mean()).astype(np.float32)
+ pk, _, _, _ = tqenc.encode(Wr, t, h=np.broadcast_to(hp, Wr.shape).copy(), **kw)
+ cands["diag p=%g" % p] = tqenc.decode(pk, t).reshape(n_out, n_in)
+ en = float(np.sum(Wr.astype(np.float64) ** 2))
+ enh = float(np.sum((Wr.astype(np.float64) ** 2) * h))
+ for name, Q in cands.items():
+ d = Wr.astype(np.float64) - Q
+ res.setdefault((l, role, name), []).append((float(np.sum(d * d) / en), float(np.sum(d * d * h) / enh)))
+ print("layer %d expert %d done" % (l, e), flush=True)
+ print("\nrelative error in dB (higher = better), mean over experts: unweighted | diagonal-Hessian")
+ names = ["shipped", "plain"] + ["diag p=%g" % p for p in powers]
+ tot = {n: [[], []] for n in names}
+ for l in [int(x) for x in a.layers.split(",")]:
+ for role in ("gate", "up", "down"):
+ line = "L%-2d %-5s" % (l, role)
+ for n in names:
+ v = np.array(res[(l, role, n)])
+ line += " %-11s %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("%-11s %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())