| CUDA_VISIBLE_DEVICES=0,1,2,3 \ | |
| NPROC_PER_NODE=4 \ | |
| megatron export \ | |
| --model Qwen/Qwen3.5-35B-A3B \ | |
| --output_dir Qwen3.5-35B-A3B-FP8 \ | |
| --to_hf true \ | |
| --fp8_recipe blockwise \ | |
| --fp8_format e4m3 \ | |
| --fp8_param_gather true \ | |
| --mtp_num_layers 1 \ | |
| --linear_decoupled_in_proj true \ | |
| --tensor_model_parallel_size 2 \ | |
| --pipeline_model_parallel_size 2 | |