wehub-resource-sync
14 lines
381 B
Bash
14 lines
381 B
Bash
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
|