[BUG] LLM QNN export ignores --chunk_size when generating prefill IO

[LudovicoYIN]

Problem

In the LLM QNN export path, transformers/llm/export/npu/generate_llm_qnn.py exposes a --chunk_size argument, but prefill IO generation did not actually honor it.

Before the fix, convert_llm() always called:

makeIOJson(args, 128, hidden_size, mask_type)

which meant the prefill QNN graph could still be generated against 128-token IO even when --chunk_size 256/512 was requested.

Impact

This can make generated QNN artifacts inconsistent with runtime configuration:

• prefill graph shape may not match the requested chunk size
• runtime chunk_limits and generated graphs may diverge
• ppl_eval results can become misleading
• chat / prefill behavior may differ from expectation

Reproduction

Run QNN export with a non-128 chunk size, for example:

python transformers/llm/export/npu/generate_llm_qnn.py \
  --model <model_dir> \
  --soc_id 43 \
  --dsp_arch v73 \
  --mnn_path <mnn_build> \
  --chunk_size 512

Expected:

• generated prefill IO and QNN graph use 512-token shape

Actual before fix:

• prefill IO generation still used 128

[Qxinyu]

Thank you for your feedback. We will fix this issue.