Match renamed globals and submodules (embed/head/norm, vision ffn vs mlp)

Open
#9 0 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Assessment

Difficulty
3/5
Estimated time
1-2 days
Newbie friendliness
54/100
Issue type
Bug
Clarity
Mostly clear
Activity status
Active
Tech stack
python

Research direction

Start by tracing how quantdiff matches tensor names between the quantized candidate and upstream weights, especially suffix matching for embed, head, and norm. Then inspect the vision tensor name handling for blocks.N.ffn versus upstream blocks.N.mlp. Verify that the reported tensors match correctly and that the extra-tensor rule no longer flags them.

Written by the indexing model from the issue text.

Description

From the DeepSeek-V4.1 iQ-MLX 3.0bpw run: 67 pack tensors listed as not in upstream. language_model.embed.weight, .head.weight and .norm.weight don't match upstream's embed.weight, head.weight and norm.weight, probably because the 2-segment suffix isn't unique once layers.N.engram.embed.weight exists. The vision tower's blocks.N.ffn.w1/w2 is mlp.w1/w2 upstream. Upstream also lists 6 of those mlp tensors as missing from the candidate. None of it is tampering, but an unmatched tensor is unchecked, and the extra-tensor rule flags the run.

Dominant language
Python
Stars
0
Forks
1
Avg merge
5m
Merged PRs (30d)
3

Getting set up

This project ships no dev container, Dockerfile or contributing guide, so setting up is up to you: start from its README, and see our first-contribution guide for the general steps.

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

More from STRML/quantdiff

All issues in STRML/quantdiff

Similar issues

More Python issues

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.