Add Flash Attention 3 support with fallback to eager

gpt_oss_transformers.py

@@ -264,13 +264,16 @@ def main(
             logger.warning("Consider using --flavor a10g-large (48GB) or a100-large (80GB)")
 
     try:
+        # Try Flash Attention 3 first for better performance
+        logger.info("Attempting to load with Flash Attention 3 (optimized)")
         model = AutoModelForCausalLM.from_pretrained(
             model_id,
             torch_dtype=torch.bfloat16,
-            attn_implementation="eager",  # As per cookbook, avoid flash attention issues
+            attn_implementation="kernels-community/vllm-flash-attn3",  # Flash Attention with Sinks
             **model_kwargs,
         )
         model.eval()
+        logger.info("Successfully loaded with Flash Attention 3")
     except torch.cuda.OutOfMemoryError as e:
         logger.error(f"Out of memory error: {e}")
         logger.error("\nThe GPT OSS models require significant memory:")
@@ -278,19 +281,27 @@ def main(
         logger.error("- 120B model: ~240GB VRAM (use 4xa100 or 8xl40s)")
         sys.exit(1)
     except Exception as e:
-        logger.error(f"Failed to load model: {e}")
-        logger.error("Trying with fallback configuration...")
-        # Fallback to simpler loading
-        try:
-            model = AutoModelForCausalLM.from_pretrained(
-                model_id,
-                torch_dtype="auto",
-                device_map="auto",
-                attn_implementation="eager",
-            )
-            model.eval()
-        except Exception as fallback_error:
-            logger.error(f"Fallback also failed: {fallback_error}")
+        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)")
+            try:
+                model = AutoModelForCausalLM.from_pretrained(
+                    model_id,
+                    torch_dtype=torch.bfloat16,
+                    attn_implementation="eager",  # Fallback to eager attention
+                    **model_kwargs,
+                )
+                model.eval()
+                logger.info("Successfully loaded with eager attention")
+            except torch.cuda.OutOfMemoryError as oom_error:
+                logger.error(f"Out of memory with eager attention: {oom_error}")
+                logger.error("The model requires more GPU memory than available")
+                sys.exit(1)
+            except Exception as eager_error:
+                logger.error(f"Failed with eager attention: {eager_error}")
+                sys.exit(1)
+        else:
+            logger.error(f"Unexpected error loading model: {e}")
             sys.exit(1)
 
     # Generation configuration