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++
dsv41: tq_convert --act (input-column weighted encode from the gate-3 calibration sums)
Niko1221da70196350c0
diff --git a/tools/dsv41/tq_convert.py b/tools/dsv41/tq_convert.pyindex 298dca622..c61b18a51 100644--- a/tools/dsv41/tq_convert.py+++ b/tools/dsv41/tq_convert.py@@ -70,6 +70,9 @@ def main():ap.add_argument("--type", required=True, choices=sorted(TYPES))ap.add_argument("--cfg", default="fast", choices=sorted(CFGS))ap.add_argument("--experts", type=int, default=0, help="only the first N experts (tests; the others stay zero streams)")+ ap.add_argument("--act", default="", help="calibration dir with L<l>_gu.f64 [n_exp][dim], L<l>_dn.f64 [n_exp][ff] (sums of x^2 of the "+ "gate/up and down inputs per expert) and L<l>_cnt.u64 [n_exp]: an input-column weighted encode (error = sum h (w - dec)^2)")+ ap.add_argument("--prior", type=float, default=64.0, help="pseudo-count of the layer-mean prior for rarely used experts")ap.add_argument("--threads", type=int, default=4)ap.add_argument("--cpu", action="store_true", help="the CPU encoder instead of the card")a = ap.parse_args()@@ -99,6 +102,13 @@ def main():with open(path, "wb") as f:f.truncate(off)fd = os.open(path, os.O_RDWR)+ act = None+ if a.act:+ gu = np.fromfile("%s/L%d_gu.f64" % (a.act, L), dtype=np.float64).reshape(n_exp, dim)+ dn = np.fromfile("%s/L%d_dn.f64" % (a.act, L), dtype=np.float64).reshape(n_exp, ff)+ cnt = np.fromfile("%s/L%d_cnt.u64" % (a.act, L), dtype=np.uint64).astype(np.float64)+ tot = max(cnt.sum(), 1.0)+ act = (gu, dn, cnt, gu.sum(0) / tot, dn.sum(0) / tot)for role, rname in ROLES:rows, cols, total = sizes[role]t0 = time.time()@@ -115,8 +125,15 @@ def main():if key not in mm:mm[key] = np.memmap(os.path.join(model_dir, key), dtype=np.uint8, mode="r")w = decode_mxfp4(np.asarray(mm[key][off_:off_ + nb]), cols)+ h = None+ if act is not None: # the input columns' mean x^2 of this expert, shrunk toward the layer's mean by the prior+ gu_, dn_, cnt_, mg, md = act+ sq, m = (dn_[e], md) if role == "w2" else (gu_[e], mg)+ hc = (sq + a.prior * m) / (cnt_[e] + a.prior)+ hc = np.maximum(hc / hc.mean(), 1e-3).astype(np.float32)+ h = np.ascontiguousarray(np.broadcast_to(hc, (rows, cols)))with gpu_lock:- packed, err, tot, st = tqenc.encode(w, t, gpu=not a.cpu, **cfg)+ packed, err, tot, st = tqenc.encode(w, t, h=h, gpu=not a.cpu, **cfg)dec = tqenc.decode(packed, t).reshape(rows, cols)err2[e] = ((w.astype(np.float64) - dec) ** 2).sum()ref2[e] = (w.astype(np.float64) ** 2).sum()