cd /workspace/torchtitan_neuron
pip install -r requirements.txt
pip install -e .Download the tokenizer:
python scripts/download_hf_assets.py --repo_id Qwen/Qwen3-30B-A3B --assets tokenizerRun:
torchrun --nproc_per_node=64 --local-ranks-filter 0 -m torchtitan.train \
--module torchtitan_neuron.experiments.graph_trainer.qwen3 \
--config qwen3_30b_a3b_nki_fullac_simplefsdp \
--parallelism.tensor_parallel_degree=4 \
--parallelism.data_parallel_shard_degree=-1 \
--parallelism.enable_async_tensor_parallel \
--training.local_batch_size=4 \
--training.global_batch_size=64torchrun --nproc_per_node=64 --local-ranks-filter 0 -m torchtitan.train \
--module torchtitan_neuron.experiments.graph_trainer.deepseek_v3 \
--config simple_fsdp_deepseek_v3_16b_nki_tp4_fsdp16 \
--parallelism.tensor_parallel_degree=4 \
--parallelism.data_parallel_shard_degree=-1 \
--parallelism.expert_parallel_degree=1 \
--training.local_batch_size=4 \
--training.global_batch_size=64 \
--training.steps=20torchrun --nproc_per_node=64 --local-ranks-filter 0 -m torchtitan.train \
--module torchtitan_neuron.experiments.graph_trainer.gpt_oss \
--config simple_fsdp_gpt_oss_20b_nki_tp4_fsdp16 \
--training.local_batch_size=4 \
--training.seq_len=4096 \
--training.steps=20The first training step will take a few minutes, but after that, expect consistent throughput.