Skip to content
Closed
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
10 changes: 9 additions & 1 deletion npu/pack_import.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -1409,7 +1409,15 @@ FconstGeom locate_fconst_with_header(const std::string& tok, const std::string&
// and every attention projection refused while the MLP imported byte-exact -- measured on Qwen2-7B 1.9.
// The scan window is the first GiB (about 8 blocks of this geometry), so the right residue repeats ~8 times
// against 1-2 for a payload coincidence. Requiring `layers/2` was too strict and silently dropped the real one.
if (best >= 3) { // the residue repeats across blocks: a real section boundary
// engine#1831: this must FILL what the epsilon vote did not establish, never OVERWRITE what it did.
// `locate_fconst()` sets `h1` and a `period - 1`-entry `hdr` when its vote wins, and `plan()` reads the block
// layout from `hdr.size()`: 1 entry is the 2-header block (q/k/v/o in one payload, Qwen2/Llama), 3 is the
// 4-header block (q/k/v, then k_norm, then q_norm, then o_proj, Qwen3). Assigning unconditionally truncated a
// Qwen3 pack's list to one entry, so `post_attn_hdr()` -- which returns `hdr.back()` and is the only thing
// that marks the end of attention -- returned the k_norm header as the start of the MLP section: weights read
// from the wrong place, silently, on a container that parses cleanly. The old comment's own reason ("setting
// only `h1` left `hdr` empty") is exactly why the guard is `g.hdr.empty()`, not a rewrite.
if (best >= 3 && g.hdr.empty()) { // the residue repeats across blocks: a real section boundary
g.h1 = best_off;
g.hdr.assign(1, best_off); // 2-header block: q/k/v/o in one payload (Qwen2/Llama)
g.period = 2;
Expand Down
Loading