0.8b fp32 (bf16 emulation slow on CPU), 4b stays bf16

server.py

@@ -18,12 +18,15 @@ _loaded = {}
 def load_all():
     for key, sub in MODELS.items():
         tok = AutoTokenizer.from_pretrained(REPO, subfolder=sub)
-        model = AutoModelForSequenceClassification.from_pretrained(REPO, subfolder=sub, dtype=torch.bfloat16)
+        # fp32 on CPU is much faster than bf16 (no AMX -> bf16 is emulated);
+        # 4B fp32 = 16GB would OOM on cpu-basic (16GB), so keep bf16 for 4b only.
+        dtype = torch.float32 if key == "0.8b" else torch.bfloat16
+        model = AutoModelForSequenceClassification.from_pretrained(REPO, subfolder=sub, dtype=dtype)
         tok.padding_side = "right"
         model.config.get_text_config().pad_token_id = tok.pad_token_id
         model.eval()
         _loaded[key] = (model, tok)
-        print(f"[openjev] loaded {key} ({sub})", flush=True)
+        print(f"[openjev] loaded {key} ({sub}, dtype={dtype})", flush=True)
 
 
 @app.on_event("startup")