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: warn loudly when --skip-tensors / --dense-override are used on an official IST-DAS set (engram_k/q/wkv present): the rebuilt pack loads but gives garbage text; the plain recipe is right

Niko12217519b1c75c58
diff --git a/tools/dsv41/pack_gguf.py b/tools/dsv41/pack_gguf.py
index 54e9d7ed6..b83e32b88 100644
--- a/tools/dsv41/pack_gguf.py
+++ b/tools/dsv41/pack_gguf.py
@@ -455,6 +455,14 @@ def main() -> int:
arch = kv["general.architecture"]
int_tensors_to_engram(gs, kv)
layers = kv.get("%s.engram.layers" % arch, [])
+ # An official IST-DAS set (helcig/tmp-oct26-2) carries the Engram gate/kv as their own k, q, wkv tensors and a 3-D attn_output_a, and the
+ # plain recipe (--engram --engram-meta --dspark-gguf, no --skip-tensors, no --dense-override) packs them as they are. Rebuilding them
+ # (engram_fold.py + a wo_a override) gave a pack with garbage output on the final 2.00 set (P2003, 2026-10-23), and the pack still loads,
+ # so it fails in the user's chat, not at install. Say so before it is built.
+ if any(n.endswith(".engram_wkv.weight") for n in have) and (a.skip_tensors or a.dense_override):
+ sys.stderr.write("pack_gguf: WARNING: this looks like an official IST-DAS set (it has engram_k/q/wkv tensors) and --skip-tensors / "
+ "--dense-override are used. The official sets pack correctly WITHOUT them; a pack rebuilt with them produced "
+ "garbage text on the final 2.00 set. Drop both unless you are testing exactly that.\n")
def has_part(il, part, names):
if "blk.%d.%s.weight" % (il, part) in names: