diff --git a/scripts/model_opt_nvfp4_full.py b/scripts/model_opt_nvfp4_full.py
index 564183f..0dc622f 100644
--- a/scripts/model_opt_nvfp4_full.py
+++ b/scripts/model_opt_nvfp4_full.py
@@ -32,13 +32,13 @@ import os
 MODEL = "/root/nvidia-meeting/DeepSeek-V4-Pro-BF16"  # Dequantized BF16 (from scripts/dequant_fp8_to_bf16.py)
 QUANT = "nvfp4"
 TP = 8
-CALIB = 256
+CALIB = 128
 KV_CACHE_QUANT = "fp8_cast"
 
 # 3TB BF16 model can't fit on 8×B200 VRAM (~1.4TB total)
 # Use seq_device_map: loads model into CPU RAM, moves layers to GPU for forward passes
 # 2.8TB RAM is enough for the 3TB model (with memory-mapped loading)
-EXTRA_FLAGS = "--trust_remote_code --use_seq_device_map --gpu_max_mem_percentage 0.9"
+EXTRA_FLAGS = "--trust_remote_code --use_seq_device_map --gpu_max_mem_percentage 0.7"
 
 # Output dir follows modelopt convention: <model>_<quant>_kv_<kv_quant>
 # We override the model name to make the strategy clear