FA3 + bf16 (clean version)

- Use bf16 dtype (FA3 requires bf16/fp16, not fp32)
- Optimized compiled graph check (use metadata instead of list)
- No FP8/Int8 quantization, just FA3 + AOT compilation

app.py

@@ -161,14 +161,13 @@ HF_CACHE_FILENAME = "compiled_graph.pt2"
 
 
 def _check_compiled_graph_exists():
-    """Check if compiled graph exists on HF Hub"""
-    from huggingface_hub import hf_hub_download, HfApi
+    """Check if compiled graph exists on HF Hub (fast check)"""
+    from huggingface_hub import hf_hub_url, get_hf_file_metadata
     try:
-        api = HfApi()
-        files = api.list_repo_files(HF_CACHE_REPO)
-        return HF_CACHE_FILENAME in files
-    except Exception as e:
-        logger.info(f"Could not check Hub for compiled graph: {e}")
+        url = hf_hub_url(HF_CACHE_REPO, HF_CACHE_FILENAME)
+        get_hf_file_metadata(url)  # Raises if file doesn't exist
+        return True
+    except Exception:
         return False
 
 
@@ -263,7 +262,7 @@ def init_generator():
             use_4bit_quantization=False,
             use_float8_quantization=False,
             use_torch_compile=False,
-            dtype="fp32",
+            dtype="bf16",  # FA3 requires bf16/fp16
         )
     
     return generator