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

README.md

@@ -1,8 +1,8 @@
 # MTB Nodes
 
-> [!CAUTION]
+> [!NOTE]
-> A lot of recent changes to comfy broke many things in mtb (colors, dynamic inputs and probably more)
+> master/main is outdated for now to keep backward compatibility, the next version is being worked on in
-> [`dev/0.6.0`](https://github.com/melMass/comfy_mtb/tree/dev/0.6.0) partially address these. My time is limited lately so it might take time to finish and merge
+> [`dev/0.6.0`](https://github.com/melMass/comfy_mtb/tree/dev/0.6.0)
 
 
 [![embedded test](https://github.com/melMass/comfy_mtb/actions/workflows/test_embedded.yml/badge.svg)](https://github.com/melMass/comfy_mtb/actions/workflows/test_embedded.yml)

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(

nodes/image_processing.py

@@ -683,6 +683,7 @@ class MTB_ImageCompare:
 
 
 import requests
+import time
 
 
 class MTB_LoadImageFromUrl:
@@ -698,6 +699,14 @@ class MTB_LoadImageFromUrl:
                         "default": "https://upload.wikimedia.org/wikipedia/commons/thumb/a/a7/Example.jpg/800px-Example.jpg"
                     },
                 ),
+                "retry_count": (
+                    "INT",
+                    {"default": 3, "min": 1, "max": 20, "step": 1},
+                ),
+                "retry_interval": (
+                    "FLOAT",
+                    {"default": 1.0, "min": 0.0, "max": 60.0, "step": 0.1},
+                ),
             }
         }
 
@@ -705,11 +714,27 @@ class MTB_LoadImageFromUrl:
     FUNCTION = "load"
     CATEGORY = "mtb/IO"
 
-    def load(self, url):
+    def load(self, url, retry_count, retry_interval):
-        # get the image from the url
+        # get the image from the url with retry + exponential backoff
-        image = Image.open(requests.get(url, stream=True).raw)
+        last_error = None
-        image = ImageOps.exif_transpose(image)
+        for attempt in range(retry_count):
-        return (pil2tensor(image),)
+            try:
+                response = requests.get(url, stream=True)
+                response.raise_for_status()
+                image = Image.open(response.raw)
+                image = ImageOps.exif_transpose(image)
+                return (pil2tensor(image),)
+            except Exception as e:
+                last_error = e
+                if attempt == retry_count - 1:
+                    raise
+                wait_seconds = retry_interval * (2**attempt)
+                if wait_seconds > 0:
+                    time.sleep(wait_seconds)
+
+        if last_error is not None:
+            raise last_error
+        raise RuntimeError("Failed to load image from URL without captured exception")
 
 
 class MTB_Blur:

nodes/model.py

@@ -145,7 +145,7 @@ class MTB_ModelPatchSeamless:
         tilingX,
         tilingY,
     ):
-        hacked_model = copy.deepcopy(model)
+        hacked_model = model.clone()
         self.apply_circular(
             hacked_model.model, startStep, stopStep, tilingX, tilingY
         )

nodes/vitmatte.py

@@ -27,7 +27,7 @@ class MTB_LoadVitMatteModel:
     def execute(self, *, kind: str, autodownload: bool):
         dest = models_dir / "vitmatte"
         dest.mkdir(exist_ok=True)
-        name = "dist" if kind == "Distinctions-646" else "com"
+        name = "dis" if kind == "Distinctions-646" else "com"
 
         file = hf_hub_download(
             repo_id="melmass/pytorch-scripts",