Fix memory issues and update GPU requirements

- Add attn_implementation='eager' as per OpenAI cookbook
- Add proper device_map configuration for 120B model
- Add memory checks with helpful error messages
- Update GPU requirements in README (20B needs 40GB+)
- Add OOM error handling with clear guidance
- Note that MXFP4 is dequantized to bf16, doubling memory needs

README.md

@@ -109,10 +109,12 @@ huggingface-cli job run --gpu-flavor a10g-small \
 
 | Model | GPU Flavor | Memory | Cost/Hour | Best For |
 |-------|------------|--------|-----------|----------|
-| `gpt-oss-20b` | `a10g-small` | 24GB | $1.38 | Most use cases |
-| `gpt-oss-20b` | `a10g-large` | 48GB | $2.50 | Larger batches |
-| `gpt-oss-120b` | `a100-large` | 80GB | $4.34 | 120B model |
-| `gpt-oss-120b` | `4xa100` | 320GB | $17.36 | Maximum speed |
+| `gpt-oss-20b` | `a10g-large` | 48GB | $2.50 | 20B model (needs ~40GB) |
+| `gpt-oss-20b` | `a100-large` | 80GB | $4.34 | 20B with headroom |
+| `gpt-oss-120b` | `4xa100` | 320GB | $17.36 | 120B model (needs ~240GB) |
+| `gpt-oss-120b` | `8xl40s` | 384GB | $23.50 | 120B maximum speed |
+
+**Note**: The MXFP4 quantization is dequantized to bf16 during loading, which doubles memory requirements.
 
 ## 🏃 Local Execution
 

gpt_oss_transformers.py

@@ -236,31 +236,62 @@ def main(
     if tokenizer.pad_token is None:
         tokenizer.pad_token = tokenizer.eos_token
 
-    # Model loading configuration
-    device_map = {"tp_plan": "auto"} if "120b" in model_id else "auto"
+    # Model loading configuration based on OpenAI cookbook
+    # For 20B model, standard auto device map works
+    # For 120B model, use tensor parallel planning
+    if "120b" in model_id:
+        model_kwargs = {
+            "device_map": {"tp_plan": "auto"},
+            "enable_expert_parallel": True,
+        }
+    else:
+        model_kwargs = {
+            "device_map": "auto",
+        }
 
     # Load model
     logger.info(f"Loading model: {model_id}")
     logger.info("This may take a few minutes for large models...")
     # Note: GPT OSS models are MXFP4 quantized out of the box
+    # The quantization will be dequantized to bf16 during loading
+    logger.info("Note: MXFP4 quantization will be dequantized to bf16 for inference")
+    
+    # Check available GPU memory
+    if num_gpus > 0:
+        gpu_memory = torch.cuda.get_device_properties(0).total_memory / 1024**3
+        if gpu_memory < 40 and "20b" in model_id.lower():
+            logger.warning(f"GPU has {gpu_memory:.1f}GB but 20B model needs ~40GB when dequantized")
+            logger.warning("Consider using --flavor a10g-large (48GB) or a100-large (80GB)")
 
     try:
         model = AutoModelForCausalLM.from_pretrained(
             model_id,
             torch_dtype=torch.bfloat16,
-            device_map=device_map,
+            attn_implementation="eager",  # As per cookbook, avoid flash attention issues
+            **model_kwargs,
         )
         model.eval()
+    except torch.cuda.OutOfMemoryError as e:
+        logger.error(f"Out of memory error: {e}")
+        logger.error("\nThe GPT OSS models require significant memory:")
+        logger.error("- 20B model: ~40GB VRAM (use a10g-large or a100-large)")
+        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 default configuration...")
+        logger.error("Trying with fallback configuration...")
         # Fallback to simpler loading
-        model = AutoModelForCausalLM.from_pretrained(
-            model_id,
-            torch_dtype="auto",
-            device_map="auto",
-        )
-        model.eval()
+        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}")
+            sys.exit(1)
 
     # Generation configuration
     generation_config = GenerationConfig(