Add kernels dependency for Flash Attention 3

gpt_oss_transformers.py

@@ -7,6 +7,7 @@
 #     "transformers>=4.45.0",
 #     "tqdm",
 #     "accelerate",
+#     "kernels>=0.9.0",  # For Flash Attention 3 support (optional but recommended)
 # ]
 # ///
 """
@@ -281,9 +282,9 @@ def main(
         logger.error("- 120B model: ~240GB VRAM (use 4xa100 or 8xl40s)")
         sys.exit(1)
     except Exception as e:
-        if "kernels-community" in str(e) or "flash" in str(e).lower() or "attn" in str(e).lower():
-            logger.warning(f"Flash Attention 3 not available: {type(e).__name__}")
-            logger.info("Falling back to eager attention (standard implementation)")
+        if "kernels" in str(e) or "flash" in str(e).lower() or "attn" in str(e).lower():
+            logger.info(f"Flash Attention 3 not available (needs 'kernels' package)")
+            logger.info("Using eager attention instead (standard implementation)")
             try:
                 model = AutoModelForCausalLM.from_pretrained(
                     model_id,