EchoScript : services\translation.py , Fix: services/translation.py now loads MarianMT directly via AutoModelForSeq2SeqLM + AutoTokenizer and calls model.generate() itself, bypassing the removed pipeline() task wrapper entirely. This is the path the official MarianMT docs use now too, so it's not a workaround you'll need to revisit on the next transformers release.

services/translation.py

@@ -34,24 +34,40 @@ class TranslationError(RuntimeError):
 
 
 @lru_cache(maxsize=None)
-def _load_pipeline(source_language: str, target_language: str):
-    """Lazily load and cache a MarianMT pipeline for one language pair.
+def _load_engine(source_language: str, target_language: str):
+    """Lazily load and cache a MarianMT model+tokenizer for one language pair.
 
     Cached so repeated translations within a session don't reload a model
     from disk every time. Swapping the translation backend later (a
     different model, a hosted API, an offline engine like Argos) only
     requires changing this one function.
+
+    Note: this loads the model/tokenizer directly via AutoModelForSeq2SeqLM
+    rather than transformers.pipeline("translation", ...). As of
+    transformers v5, the generic "translation" pipeline task was removed
+    entirely (see huggingface/transformers#43825) -- pipeline("translation")
+    now raises KeyError. Loading the model directly and calling
+    `model.generate()` is unaffected by that change and is the path the
+    transformers docs now show for MarianMT.
     """
-    from transformers import pipeline  # heavy import, deferred until needed
+    from transformers import AutoModelForSeq2SeqLM, AutoTokenizer  # heavy import, deferred
 
     model_name = f"Helsinki-NLP/opus-mt-{source_language}-{target_language}"
     try:
-        return pipeline("translation", model=model_name)
+        tokenizer = AutoTokenizer.from_pretrained(model_name)
+        model = AutoModelForSeq2SeqLM.from_pretrained(model_name)
     except Exception as exc:  # pragma: no cover - depends on model availability
         raise TranslationError(
             f"No translation model available for "
             f"'{source_language}' -> '{target_language}': {exc}"
         ) from exc
+    return tokenizer, model
+
+
+def _run_translation(tokenizer, model, text: str) -> str:
+    inputs = tokenizer(text, return_tensors="pt", truncation=True)
+    generated = model.generate(**inputs, max_new_tokens=512)
+    return tokenizer.decode(generated[0], skip_special_tokens=True).strip()
 
 
 class TranslationService:
@@ -67,16 +83,16 @@ class TranslationService:
                 segments=list(transcript.segments),
             )
 
-        translator = _load_pipeline(transcript.language, target_language)
+        tokenizer, model = _load_engine(transcript.language, target_language)
 
         translated_segments = []
         for seg in transcript.segments:
             if not seg.text:
                 translated_segments.append(seg)
                 continue
-            result_text = translator(seg.text)[0]["translation_text"]
+            result_text = _run_translation(tokenizer, model, seg.text)
             translated_segments.append(
-                Segment(index=seg.index, start=seg.start, end=seg.end, text=result_text.strip())
+                Segment(index=seg.index, start=seg.start, end=seg.end, text=result_text)
             )
 
         return Translation(