From b35b5d8a17c0d59e80a8b3627b679c2c1003d04f Mon Sep 17 00:00:00 2001
From: drbaph <84208527+Saganaki22@users.noreply.github.com>
Date: Sat, 4 Jul 2026 20:40:35 +0100
Subject: [PATCH] fix: whisper transcription compatibility with newer
 transformers (#274)

- Use getattr for max_length to handle removed WhisperConfig attribute
- Cast input_features to model dtype to fix float16 mismatch
---
 nodes/audio.py | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/nodes/audio.py b/nodes/audio.py
index cb5c637..6fb7bf5 100644
--- a/nodes/audio.py
+++ b/nodes/audio.py
@@ -277,14 +277,14 @@ class MTB_AudioToText(MtbAudio):
                 f"Processing chunk {chunk_offset:.1f}s - {chunk_end / sample_rate:.1f}s"
             )
 
-            max_length = model.config.max_length or 448
+            max_length = getattr(model.config, "max_length", None) or 448
             attention_mask = torch.ones((1, max_length))
 
             input_features = processor(
                 chunk_waveform,
                 sampling_rate=sample_rate,
                 return_tensors="pt",
-            ).input_features.to(device)
+            ).input_features.to(device=device, dtype=model.dtype)
 
             with torch.no_grad():
                 predicted_ids = model.generate(