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: tq_convert resumes (skips finished layers), pack_gguf --tq-experts takes several folders

Niko1221e8cb1cba7f67
diff --git a/tools/dsv41/pack_gguf.py b/tools/dsv41/pack_gguf.py
index b56d57825..82348ca11 100644
--- a/tools/dsv41/pack_gguf.py
+++ b/tools/dsv41/pack_gguf.py
@@ -331,7 +331,7 @@ def main() -> int:
"in <model>/.. when there is one)")
ap.add_argument("--mmproj", help="unsloth's vision tower GGUF (mmproj-*.gguf, clip / deepseek41v): its tensors become the "
"checkpoint's vision.* / aligner.* (default: the first mmproj*.gguf in <model>/..)")
- ap.add_argument("--tq-experts", help="a folder tq_convert.py wrote (tq.json + tq-L<layer>.bin): the routed experts of the layers "
+ ap.add_argument("--tq-experts", nargs="+", help="folders tq_convert.py wrote (tq.json + tq-L<layer>.bin): the routed experts of the layers "
"listed there are read from it as trellis types (TQK5X ..), the others keep their MXFP4")
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 "
@@ -552,14 +552,15 @@ def main() -> int:
# trellis experts from tq_convert.py: per layer the three roles stacked [n_exp][rows][row bytes] in one file
tq_layers = []
if a.tq_experts:
- tqd = pathlib.Path(a.tq_experts).expanduser().resolve()
- tq_idx = json.loads((tqd / "tq.json").read_text())
- by_layer = {}
- for key, ent in tq_idx.items():
- m = re.match(r"layers\.(\d+)\.ffn\.experts\.(w[123])$", key)
- if not m:
- raise SystemExit("tq.json: %s is not a layer's expert role" % key)
- by_layer.setdefault(int(m.group(1)), {})[m.group(2)] = ent
+ by_layer, tq_dirs = {}, {}
+ for d in a.tq_experts:
+ tqd = pathlib.Path(d).expanduser().resolve()
+ for key, ent in json.loads((tqd / "tq.json").read_text()).items():
+ m = re.match(r"layers\.(\d+)\.ffn\.experts\.(w[123])$", key)
+ if not m:
+ raise SystemExit("tq.json: %s is not a layer's expert role" % key)
+ by_layer.setdefault(int(m.group(1)), {})[m.group(2)] = ent
+ tq_dirs[(int(m.group(1)), m.group(2))] = tqd
TQTYPES = {v[0]: k for k, v in TYPES.items()}
for il, roles in sorted(by_layer.items()):
if sorted(roles) != ["w1", "w2", "w3"]:
@@ -576,7 +577,7 @@ def main() -> int:
raise SystemExit("tq.json: unknown type %s" % tname)
n_exp, rows, cols = shape
per = total // n_exp
- f = tqd / fname
+ f = tq_dirs[(il, role)] / fname
if not f.is_file() or f.stat().st_size < off + total:
raise SystemExit("%s: missing or short" % f)
try:
@@ -589,7 +590,7 @@ def main() -> int:
for e in range(n_exp):
tensors.append(("layers.%d.ffn.experts.%d.%s.weight" % (il, e, role), tname, fi, off + e * per, per, [rows, cols]))
tq_layers.append(il)
- print("trellis experts from %s: layers %s" % (tqd, tq_layers))
+ print("trellis experts from %s: layers %s" % (", ".join(a.tq_experts), tq_layers))
drafter = pathlib.Path(a.drafter).expanduser().resolve() if a.drafter else model.parent / "dspark" / "dspark.json"
if a.drafter and not drafter.is_file():
raise SystemExit("%s: no drafter there" % drafter)
diff --git a/tools/dsv41/tq_convert.py b/tools/dsv41/tq_convert.py
index c61b18a51..dbc4e4043 100644
--- a/tools/dsv41/tq_convert.py
+++ b/tools/dsv41/tq_convert.py
@@ -91,6 +91,9 @@ def main():
gpu_lock = threading.Lock()
for L in [int(x) for x in a.layers.split(",")]:
path = os.path.join(a.out, "tq-L%d.bin" % L)
+ if all("layers.%d.ffn.experts.%s" % (L, r) in idx for r, _ in ROLES) and os.path.exists(path):
+ print("layer %d: already in %s, skipped" % (L, idx_path), flush=True)
+ continue
sizes = {}
offs = {}
off = 0