mtmd: qwen3-asr wrong output

[cora4]

Overview

qwen3-asr output is wrong with llama.cpp

[ngxson]

closing this as you did not disclosed AI usage and this is not the intended way to fix

[ngxson]

The code comments are clearly written by AI

[ngxson]

Your fix cannot be accepted as-is because it introduces a hack into the existing preprocessor, it is not clean and potentially break other models.

The cgraph implementation is redundant because (1) you can use the 4th dim for batching and (2) you replaced the old build_vit with essentially the over-complicated version equivalent of it.

But overall, I do not wish to proceed with contributions where AI is used but not properly disclosures.

[cora4]

qwen3a can be implemented as mtmd_audio_preprocessor_qwen3a, but it was merged in the whisper preprocessor.

mtmd: qwen3 audio support (qwen3-omni and qwen3-asr) (#19441)

tools/mtmd/mtmd.cpp :
// set preprocessor
switch (proj) {
case PROJECTOR_TYPE_QWEN2A:
case PROJECTOR_TYPE_QWEN3A:
case PROJECTOR_TYPE_QWEN25O:
{
// <|audio_bos|> ... (embeddings) ... <|audio_eos|>
aud_beg = "<|audio_bos|>";
aud_end = "<|audio_eos|>";
audio_preproc = std::make_unique<mtmd_audio_preprocessor_whisper>(ctx_a);
} break;