fix: restore audio attention output layout

Restore the [batch, time, heads, head_dim] layout before flattening multi-head attention output. Local standalone eval10: 11/11 stop, 0/11 max-token hits, 4.73% rough CER. Also remove the temporary README warning.

README.md

@@ -31,19 +31,6 @@ repository: https://github.com/AutoArk/open-audio-opd
 
 </div>
 
-## IMPORTANT: Temporary Transformers Release Warning
-
-**Do not use the current Transformers artifact for evaluation, benchmarking,
-or production.** We confirmed a tensor-layout bug in the published remote-code
-audio attention implementation that can produce incorrect or repetitive
-transcriptions. The affected artifact includes model-code revision
-`25ae981173eb8c9eb2d672716d10fd960dd882e6`.
-
-The model weights have been verified intact, and a corrected Transformers
-revision is being prepared. This README-only warning does not contain the fix.
-The ONNX Runtime and iOS ANE packages use separate inference implementations
-and are being audited independently; this notice does not state their status.
-
 `Audio8-ASR-0.1B` is a compact autoregressive ASR model whose language-model
 component has only 0.1B parameters. It supports multilingual speech recognition
 for languages including Chinese, English, French, Japanese, and Cantonese. We

qwen3_asr_audio_model.py

@@ -71,7 +71,8 @@ class Qwen3ASRAudioAttention(nn.Module):
         attn_weights = F.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
         attn_weights = F.dropout(attn_weights, p=self.attention_dropout, training=self.training)
         attn_output = torch.matmul(attn_weights, value)
-        attn_output = attn_output.reshape(seq_length, self.embed_dim).contiguous()
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        attn_output = attn_output.reshape(seq_length, self.embed_dim)
         return self.out_proj(attn_output)