ImportBot: Wording: lab notes / site admins instead of a personal name

2026-10-31T21:36:15Z

Wording: lab notes / site admins instead of a personal name

New page

'''heretic_inplace.py''' applies the pqhaz "heretic" FP8 patch to the <code>attn.wo_b</code> weights of the official
DeepSeek V4.1 Flash shards, in place, for engines that read attention weights straight from the official checkpoint.
Only the 38 <code>layers.N.attn.wo_b.weight</code> tensors change (same dtype F8_E4M3, same shape, same offsets); the
FP8 scales are untouched. Three commands: <code>apply</code>, <code>restore</code> and <code>status</code>. Python
standard library only.

{{Download|version=2026-10-29|source=[https://cmp.wiki/wiki/User:Darkmaniac7 lab notes (@darkmaniac7)]|expected_sha256=5c2e6a96b424ba5c6b2983bf06953704f0dd63e51c3beb5ef25e5a73cea44343|notes=Used by [[DeepSeek V4.1 Flash heretic on 4x 8 GB]].}}

== Usage ==
<pre>
heretic_inplace.py apply OFFICIAL_DIR PATCH.safetensors BACKUP.safetensors
heretic_inplace.py restore OFFICIAL_DIR PATCH.safetensors BACKUP.safetensors
heretic_inplace.py status OFFICIAL_DIR PATCH.safetensors BACKUP.safetensors
</pre>
* <code>OFFICIAL_DIR</code>: the official DeepSeek-V4.1-Flash-1017 checkpoint (with <code>model.safetensors.index.json</code>).
* <code>apply</code> first writes an exact byte backup of every original tensor (a safetensors file with the same names),
then overwrites, fsyncs and reads back. It refuses a shard that has more than one hard link, and refuses to run when
a tensor is already patched and there is no backup.
* <code>restore</code> writes the backup back; <code>status</code> counts tensors that are patched, stock (match the
backup) or neither. Both are safe to repeat: a tensor already in the wanted state is skipped.

== License ==
MIT (the file carries no license header).

{{Source|darkmaniac7|2026-10-29|note=lab notes}}

File:Heretic inplace.py - Revision history

ImportBot: Wording: lab notes / site admins instead of a personal name