Drive the MT model directly: transformers 5 dropped the translation task

pipeline('translation') raises KeyError on transformers 5, and gradio 6's
huggingface-hub 1.x requirement rules out going back to 4.x. Use
AutoModelForSeq2SeqLM with beam search instead - deterministic, which is the
point of moving off Gemini.

app.py

@@ -103,7 +103,14 @@ def _load(kind: str):
                     device=DEVICE,
                 )
             else:
-                obj = pipeline("translation", model=MT_MODEL, token=HF_TOKEN, device=DEVICE)
+                # transformers 5 dropped the "translation" pipeline task, so the
+                # seq2seq model is driven directly. Beam search, no sampling —
+                # the same cue gives the same English on every run, which is the
+                # whole reason for moving off Gemini.
+                from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
+                tok = AutoTokenizer.from_pretrained(MT_MODEL, token=HF_TOKEN)
+                mdl = AutoModelForSeq2SeqLM.from_pretrained(MT_MODEL, token=HF_TOKEN)
+                obj = (tok, mdl.to(DEVICE).eval())
         except Exception as e:
             _errors[kind] = f"{type(e).__name__}: {e}"
             raise
@@ -268,7 +275,17 @@ class TranslateIn(BaseModel):
 
 
 def _mt_impl(texts):
-    return _load("mt")(texts, batch_size=MT_BATCH, truncation=True)
+    import torch
+    tok, mdl = _load("mt")
+    out = []
+    for start in range(0, len(texts), MT_BATCH):
+        chunk = texts[start:start + MT_BATCH]
+        enc = tok(chunk, return_tensors="pt", padding=True, truncation=True,
+                  max_length=512).to(DEVICE)
+        with torch.inference_mode():
+            gen = mdl.generate(**enc, num_beams=4, max_new_tokens=256)
+        out.extend(tok.batch_decode(gen, skip_special_tokens=True))
+    return out
 
 
 _mt_gpu = _GPU(_mt_impl)
@@ -298,7 +315,7 @@ def translate(body: TranslateIn, x_service_token: str = Header(default=None)):
     if idx:
         res = _mt_run([texts[i] for i in idx])
         for i, r in zip(idx, res):
-            out[i] = (r.get("translation_text") or "").strip()
+            out[i] = (r or "").strip()
     return {"translations": out}