lm_head is not converted to QuantLinear with MXFP4/8

[xin3he]

lm_head quantization still have some issues.

• need deepcopy if tied_word_embedding = True
• export is not applied for lm_head

Shall we warn user that lm_head is not supported? @WeiweiZhang1 @wenhuach21

[xin3he]

BTW, AFAIK, QuantLinear for MXFP4/8 has no forward function and may confuse user about how to use it. Do we plan to support it?

[wenhuach21]

if tied_word_embedding = True, lm-head quant is disabled by default. What's the issue? what do you mean "QuantLinear for MXFP4/8 has no forward function "

[xin3he]

If user prefer to quantize lm_head, what's the solution?

[wenhuach21]

how do you run the model, why quantlinear has no forward?

[xin3he]

https://github.com/intel/auto-round/blob/8d8a1cd5daaf6e8c71d079eccaec3092fa9af4f1/auto_round/export/export_to_autoround/qlinear_fp.py#L61 It's not implemented in AutoRound

[wenhuach21]

how do you use the model? please attach the cmd. After packing and saving, the model should be reloaded and MXFP4QuantLinear this layer should be called

[xin3he]

I'm using the model for export after quantize_and_save() and just aware that AutoRound requires reloading before inference.

[xin3he]

I tried Qwen3-8b which is not using tied_word_embeding, the lm_head is still not quantized. I notice the quantization bar contains this op while module replacement is not enabled.

[wenhuach21]

do you enable quant_lm_head or set bits for lm-head

[xin3he]

[xin3he]

Do we plan to support lm_head for tied_word_embedding=True?

[xin3he]

After reloading I saw the lm_head is quantized, not sure what is happening. @WeiweiZhang1 Do you have any comments? Do you think it's a bug or it is designed.

[wenhuach21]

at least need to warn users that xin's way is not supported

[xin3he]

For NVFP4, lm_head quantization will get assert error during exporting.