Fix deepseek indexer

[pcuenca]

This description was written by a human 🙋‍♂️

Rope traditional was changed for the Indexer here as part of #867. This fixed a very obvious long-sequence generation error with GLM 5: the model started producing gibberish after a while.

I think that the indexer of Deepseek 3.2 requires False, but using True does not result in immediately observed garbage output. Outputs remain coherent, but quality silently degrades for long sequences.

I tested by generating 6 JavaScript implementations for different games using the prompts shown in the script below. We want generation to go beyond the 2048 sequence length where the indexer kicks in. In all 6 cases, the implementations from the current mlx-lm package are coherent, but the games are not playable. Using the patched version, all games are playable (with confusing visual imperfections in two of the games, but playable).

I found this issue while working on GLM 5.2. I think that silent degradation + superclass being used as the reference for other implementations is worth fixing as robustly as we can. Happy to run additional tests to confirm behavior!

Generation script

Covers 5 prompts, there's a sixth one I ran outside the script.

#!/bin/bash

declare -a prompts=(
  "Write a complete, polished, single-file HTML+JavaScript implementation of Tetris, with comments. Include the 7 tetromino shapes, rotation, line clearing, scoring, increasing speed by level, and a next-piece preview."
  "Write a complete, polished, single-file HTML+JavaScript 2048 game, with comments. Arrow-key controls, tile sliding and merging, score, win at 2048, and game-over detection."
  "Write a complete, polished, single-file HTML+JavaScript implementation of Snake, with comments. Grid movement, growth on eating food, self- and wall-collision death, score, and a restart."
  "Write a complete, polished, single-file HTML+JavaScript implementation of Breakout/Arkanoid, with comments. A paddle, a grid of destructible bricks, ball physics with wall/paddle/brick bounces, lives, score, and level-clear."
  "Write a complete, polished, single-file HTML+JavaScript implementation of Minesweeper, with comments. Random mine placement, left-click reveal with flood-fill on zero-adjacency cells, right-click flagging, number hints, and win/lose detection."
)

typeset -i id=1
for p in "${prompts[@]}"
do
  for version in "mlx-lm" "mlx-lm.patched"
  do
    uv pip install -e "./$version"
    echo "[$version] Generating prompt $p"
    mlx_lm.generate --model models/mlx-community/DeepSeek-V3.2-4bit --prompt "$p" -m 16384 --temp 0 | tee ${id}-${version}.txt
  done
  let id=$id+1
done

---

This PR is the counterpart of huggingface/transformers#46842 (in transformers, GLM appears to be wrong but Deepseek is correct).

[pcuenca]

This is technically a different issue, but mlx_lm.generate did not run otherwise.

[pcuenca]

Warning: we are flipping the default to False here.

However, I think the only subclass is currently glm_moe_dsa and the property does exist for GLM 5 models:

• GLM 5 original, mlx
• GLM 5.1 mlx-community/GLM-5.1, mlx-community/GLM-5.1-MXFP4-Q8, mlx-community/GLM-5.1-DQ4plus-q8
• GLM 5.2 mlx (I ran the conversion from the original and the property was thus inherited).

[angeloskath]

Thanks for fixing the super subtle bug!

Also

This description was written by a human 🙋‍♂️

Absolute king! Thanks for that.