[BUG] GLM5.2 unsupported

[matthieu-zimmer]

Describe the bug

GLM 5.2 is not supported because of the new shared module (prev_topk_indices is missing in the forward).

python3.12/site-packages/transformers/models/glm_moe_dsa/modeling_glm_moe_dsa.py", line 418, in forward
    raise ValueError("Shared DSA layers require top-k indices from a previous full indexer layer.")
ValueError: Shared DSA layers require top-k indices from a previous full indexer layer.

Software Info

Operation System/Version + Python Version

Show output of:

GPTQModel @ git+https://github.com/ModelCloud/GPTQModel.git@b11aee07889a7030f80057b9850416b200613d47

[CaptainRong]

We do the same.
GLM-5.2 introduces IndexShare for DSA: some decoder layers reuse the
topk_indices produced by a previous layer instead of running their own
indexer. During normal model execution, this state is passed between decoder
layers as prev_topk_indices.

GPTQModel performs calibration with layer-by-layer replay. We were already
cross-layer IndexShare state was lost before a shared DSA layer was replayed.

Our fix is to preserve this auxiliary state per calibration batch:

1. Run the current decoder layer.
2. Keep output[0] as the next layer's hidden states.
3. When output[1] contains valid topk_indices, store it in that batch's
   replay kwargs as prev_topk_indices.
4. Move it to the appropriate calibration/compute device before replaying the
   next layer.

The same propagation is required in both the serial replay path and the
automatic forward data-parallel worker path.

preserving GLM-5.2's cross-layer IndexShare state during quantization replay.

You can refer to our patch:https://gist.github.com/CaptainRong/f4f4a7bcd719709b14faef92718bd71f

[Qubitium]

@CaptainRong Can you PR the mod? This way you get this proper credit instead of us basing a PR from your delta.