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++

dsv41: pack_gguf --override-experts (mixed-bit packs from two requantized sets); stageB tooling

Niko12215e52cf569f4e
diff --git a/tools/dsv41/pack_gguf.py b/tools/dsv41/pack_gguf.py
index af7d0f94d..e573636dc 100644
--- a/tools/dsv41/pack_gguf.py
+++ b/tools/dsv41/pack_gguf.py
@@ -335,6 +335,10 @@ def main() -> int:
"listed there are read from it as trellis types (TQK5X ..), the others keep their MXFP4")
ap.add_argument("--dense-override", help="a folder dense_requant.py wrote (dense.json + dense.bin): its tensors replace the same-named "
"dense matrices of the GGUF (K-quants of the attention / shared experts)")
+ ap.add_argument("--override-experts", help="a second requantized GGUF folder (requant_gguf.py output): its routed experts replace the main "
+ "model's for --override-roles (gate,up,down) in --override-layers (default all): mixed-bit packs")
+ ap.add_argument("--override-roles", default="down")
+ ap.add_argument("--override-layers", default="")
ap.add_argument("--tq-layers", help="only these layers (comma list) of the --tq-experts folders")
ap.add_argument("--drafter", help="the DSpark drafter's .json (drafter_pack.py's; default <model>/../dspark/dspark.json)")
ap.add_argument("--engram", help="the folder of the Engram tables' GGUF set, for a quant that leaves them out "
@@ -444,6 +448,47 @@ def main() -> int:
files = names
if eg_from:
print("Engram tables from %s (%d shards)" % (eg_dir, len(eg_from)))
+ # experts of a second requantized model replace the main one's for some roles / layers (the byte layouts are per tensor: any mix of types)
+ n_ov = 0
+ if a.override_experts:
+ od = pathlib.Path(a.override_experts).expanduser().resolve()
+ roles_ov = {"gate": "w1", "up": "w3", "down": "w2"}
+ want = {roles_ov[r] for r in a.override_roles.split(",")}
+ lay_ov = None
+ if a.override_layers:
+ lay_ov = set()
+ for part in a.override_layers.split(","):
+ lo, _, hi = part.partition("-")
+ lay_ov.update(range(int(lo), int(hi or lo) + 1))
+ ogs = [Gguf(pth) for pth in sorted(od.glob("*.gguf"))]
+ new = []
+ for og in ogs:
+ try:
+ orel = os.path.relpath(og.path, model).replace(os.sep, "/")
+ except ValueError:
+ raise SystemExit("%s must be on the model's drive" % od)
+ files.append(orel)
+ ofi = len(files) - 1
+ for name, typ, dims, off in og.tensors:
+ m = re.match(r"blk\.(\d+)\.ffn_(gate|up|down)_exps\.weight$", name)
+ if not m or typ not in TYPES:
+ continue
+ il, role = int(m.group(1)), {"gate": "w1", "up": "w3", "down": "w2"}[m.group(2)]
+ if role not in want or (lay_ov is not None and il not in lay_ov):
+ continue
+ tname, per, bpb = TYPES[typ]
+ nel = 1
+ for d in dims:
+ nel *= d
+ nbytes = nel // per * bpb
+ shape = list(reversed(dims))
+ n, part = shape[0], nbytes // shape[0]
+ for i in range(n):
+ new.append(("layers.%d.ffn.experts.%d.%s.weight" % (il, i, role), tname, ofi, og.data_start + off + i * part, part, shape[1:]))
+ names_new = {t[0] for t in new}
+ tensors[:] = [t for t in tensors if t[0] not in names_new] + new
+ n_ov = len(new)
+ print("experts overridden from %s: %d tensors (roles %s, layers %s)" % (od, n_ov, a.override_roles, a.override_layers or "all"))
# dense matrices requantized by dense_requant.py replace the GGUF's BF16 ones
n_dense = 0
if a.dense_override:
@@ -679,7 +724,7 @@ def main() -> int:
by = {}
for t in tensors:
by[t[1]] = by.get(t[1], 0) + t[4]
- extras = [w for w, there in (("the DSpark drafter", drafter_in or dspark_gguf), ("the vision tower", bool(vision)), ("trellis experts in layers %s" % tq_layers, bool(tq_layers)), ("%d dense matrices requantized" % n_dense, bool(n_dense))) if there]
+ extras = [w for w, there in (("the DSpark drafter", drafter_in or dspark_gguf), ("the vision tower", bool(vision)), ("trellis experts in layers %s" % tq_layers, bool(tq_layers)), ("%d dense matrices requantized" % n_dense, bool(n_dense)), ("%d experts from %s" % (n_ov, a.override_experts), bool(n_ov))) if there]
print("%s: %d tensors in %d files (%s), Engram %d layers x %d hash columns%s -> %s" %
(model.name, len(tensors), len(files), ", ".join("%s %.1f GB" % (k, v / 1e9) for k, v in sorted(by.items())),
nl, cols, "".join(", " + w for w in extras), out))