Upload mythos-fineweb-moe.py with huggingface_hub

mythos-fineweb-moe.py

@@ -1035,6 +1035,7 @@ def main():
     cfg.vocab_size = vocab_size
 
     bf16_ok = torch.cuda.is_available() and torch.cuda.is_bf16_supported()
+    use_mxfp8 = os.environ.get("MXFP8", "0") == "1"
     amp_dtype = torch.bfloat16 if bf16_ok else torch.float16
 
     model = SpiderPortalForConditionalGeneration(cfg).to(torch.bfloat16)
@@ -1058,7 +1059,7 @@ def main():
         model = model.to(device)
 
     if master:
-        logger.info("MoE mode: using native bf16 (MXFP8 disabled)")
+        logger.info(f"MoE mode: {amp_dtype} | MXFP8 hardware acceleration: {'ENABLED' if use_mxfp8 else 'disabled (set MXFP8=1)'}")
 
     # MoE init checkpoint (skip dense conversion, load MoE weights directly)
     moe_init_ckpt = os.environ.get("MOE_INIT_CKPT", "")