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++

pack_gguf: accept older DeepSeek GGUF metadata names (#3955)

Tsuryfb966cff37e9
diff --git a/tests/terminal/test_pack_gguf_engram.py b/tests/terminal/test_pack_gguf_engram.py
index 17848fa22..448d1c97c 100644
--- a/tests/terminal/test_pack_gguf_engram.py
+++ b/tests/terminal/test_pack_gguf_engram.py
@@ -56,5 +56,51 @@ class EngramHeader(unittest.TestCase):
self.assertFalse(PG.is_engram_file(PG.Gguf(tiny_gguf(Path(d) / "m.gguf", "deepseek41"))))
+class LegacyModelMetadata(unittest.TestCase):
+ def test_legacy_engram_names_write_the_same_hash_table(self):
+ common = {
+ "general.architecture": "deepseek41",
+ "deepseek41.engram.max_ngram_size": 2,
+ "deepseek41.engram.head_count": 1,
+ "deepseek41.engram.token_map": [0, 1, 2],
+ "deepseek41.engram.multipliers": [3, 5],
+ "deepseek41.engram.primes": [7],
+ "deepseek41.engram.offsets": [0],
+ }
+ legacy = {**common, "deepseek41.engram.layer_ids": [1],
+ "deepseek41.engram.key_length": 256, "deepseek41.engram.pad_id": 2}
+ current = {**common, "deepseek41.engram.layers": [1],
+ "deepseek41.engram.head_dim": 256, "deepseek41.engram.pad_token_id": 2,
+ "deepseek41.engram.key_length": 999, "deepseek41.engram.pad_id": 999}
+ PG.int_tensors_to_engram([], legacy)
+ PG.int_tensors_to_engram([], current)
+ with tempfile.TemporaryDirectory() as d:
+ a, b = Path(d) / "legacy.bin", Path(d) / "current.bin"
+ PG.write_engram(a, legacy, 3)
+ PG.write_engram(b, current, 3)
+ self.assertEqual(a.read_bytes(), b.read_bytes())
+ self.assertEqual(struct.unpack("<4s6i", a.read_bytes()[:28]),
+ (b"DSE1", 3, 1, 2, 1, 256, 2))
+
+ def test_source_config_restores_omitted_sparse_geometry(self):
+ kv = {"general.architecture": "deepseek41", "deepseek41.block_count": 40,
+ "deepseek41.embedding_length": 5120, "deepseek41.expert_count": 384}
+ cfg = {"text_config": {"num_hidden_layers": 40, "hidden_size": 5120,
+ "n_routed_experts": 384, "kv_source_layer_ids": [2, 8, 14, 20],
+ "index_source_layer_ids": [2, 8, 14, 20, 24, 28, 32, 36],
+ "candidate_source_layer_id": 20, "candidate_topk_blocks": 2048,
+ "candidate_block_size": 8}}
+ PG.apply_source_config(kv, cfg)
+ geometry = PG.geometry(kv, 129280)
+ self.assertEqual(geometry["kv_source_layer_ids"], [2, 8, 14, 20])
+ self.assertEqual(geometry["index_source_layer_ids"], [2, 8, 14, 20, 24, 28, 32, 36])
+ self.assertEqual(geometry["candidate_source_layer_id"], 20)
+ self.assertEqual(geometry["candidate_topk_blocks"], 2048)
+ self.assertEqual(geometry["candidate_block_size"], 8)
+ cfg["text_config"]["candidate_block_size"] = 16
+ with self.assertRaisesRegex(SystemExit, "candidate_block_size"):
+ PG.apply_source_config(kv, cfg)
+
+
if __name__ == "__main__":
unittest.main()
diff --git a/tools/dsv41/pack_gguf.py b/tools/dsv41/pack_gguf.py
index b56838f59..4a44d1cca 100644
--- a/tools/dsv41/pack_gguf.py
+++ b/tools/dsv41/pack_gguf.py
@@ -20,6 +20,9 @@ from there when it is present. Nor a vision tower: vision_pack.py makes it from
(--vision: another folder), and the pack takes it and its geometry (vision.<key>, image_token_id) from there when it is
present; without it the engine reads text only.
+For older GGUF converters that omit sparse-attention geometry, pass --config-json /path/to/the/source/model/config.json.
+Existing metadata must agree with that configuration; the weights are never rewritten.
+
A quant that leaves the two Engram tables out (v1.1.16's Maya-S; Maya's quants hold them since) takes them from a set
of GGUF files that holds them, beside it: --engram, else ../Engram/ (unsloth's MXFP8 shard of them, where setups before
v1.1.17 put it) when it is whole.
@@ -253,6 +256,29 @@ def ple_to_engram(kv: dict):
kv["%s.engram.%s" % (a, k)] = v
+def apply_source_config(kv: dict, config: dict):
+ """Fill sparse-attention geometry omitted by older converters from the model's own config.json."""
+ a = kv["general.architecture"]
+ text = config.get("text_config", config)
+ fields = {
+ "embedding_length": "hidden_size", "block_count": "num_hidden_layers",
+ "expert_count": "n_routed_experts", "attention.compress_ratios": "compress_ratios",
+ "attention.kv_source_layers": "kv_source_layer_ids",
+ "attention.indexer.source_layers": "index_source_layer_ids",
+ "attention.indexer.candidate_source_layer": "candidate_source_layer_id",
+ "attention.indexer.candidate_top_k_blocks": "candidate_topk_blocks",
+ "attention.indexer.candidate_block_size": "candidate_block_size",
+ }
+ for name, source in fields.items():
+ if source not in text:
+ continue
+ keys = ["%s.%s" % (a, k) for k in (name, *ALIASES.get(name, ()))]
+ for key in keys:
+ if key in kv and kv[key] != text[source]:
+ raise SystemExit("source config %s disagrees with GGUF %s" % (source, key))
+ kv.setdefault(keys[0], text[source])
+
+
def int_tensors_to_engram(gs, kv: dict):
"""The IST-DAS sets carry the Engram's hash parameters as integer tensors of the first shard (engram_token_map I32,
engram_primes / engram_offsets / engram_multipliers I64, one row a layer) instead of metadata: read them into the
@@ -276,7 +302,8 @@ def int_tensors_to_engram(gs, kv: dict):
# the IST-DAS spelling of the Engram's scalar keys -> the dotted ones this pack reads
for dotted, theirs in (("engram.layers", "engram.layer_ids"), ("engram.max_ngram_size", "engram_max_ngram_size"),
("engram.head_count", "engram_n_heads"), ("engram.head_dim", "engram_head_dim"),
- ("engram.pad_token_id", "engram_pad_token_id")):
+ ("engram.pad_token_id", "engram_pad_token_id"),
+ ("engram.head_dim", "engram.key_length"), ("engram.pad_token_id", "engram.pad_id")):
if "%s.%s" % (a, dotted) not in kv and "%s.%s" % (a, theirs) in kv:
kv["%s.%s" % (a, dotted)] = kv["%s.%s" % (a, theirs)]
@@ -413,6 +440,7 @@ def main() -> int:
ap = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
ap.add_argument("--model", required=True, help="the directory of the GGUF's shards")
ap.add_argument("--out", help="default: <model>/maya-ds4")
+ ap.add_argument("--config-json", help="the exact source model config.json, to fill sparse-attention geometry omitted by older GGUF converters")
ap.add_argument("--vision", help="vision_pack.py's folder (default: <model>/../vision; taken when it is there)")
ap.add_argument("--dspark-gguf", help="unsloth's DSpark drafter GGUF (dspark-*-Q8_0.gguf, architecture dflash): its three "
"blocks become mtp.0-2 beside the trunk's tensors (default: the first dspark-*.gguf "
@@ -461,6 +489,8 @@ def main() -> int:
if kv is None:
arch = next((g.kv.get("general.architecture") for g in gs), None)
raise SystemExit("%s: not a DeepSeek-V4.1 GGUF (architecture %r)" % (model, arch))
+ if a.config_json:
+ apply_source_config(kv, json.loads(pathlib.Path(a.config_json).expanduser().read_text()))
split = int(kv.get("split.count", 1))
if split != len(shards):
counts = sorted({m.group(1) for m in (re.search(r"-of-(\d{5})\.gguf$", p.name) for p in shards) if m})