Nemotron-4-340B-Base-hf-FP8

Nemotron-4-340B-Base-hf-FP8

Converted and quantized checkpoint of nvidia/Nemotron-4-340B-Base. Specifically it was produced from the v1.2 .nemo checkpoint on NGC.

You can deploy this model with vllm>=0.5.4 (PR#6611):

vllm serve mgoin/Nemotron-4-340B-Base-hf-FP8 --tensor-parallel-size 8

Evaluations

All the below evaluations were run with lm-eval==0.4.3 on 8xA100 GPUs.

lm_eval --model vllm --model_args pretrained=/home/mgoin/code/Nemotron-4-340B-Base-hf-FP8,tensor_parallel_size=8,distributed_executor_backend="ray",max_model_len=4096,gpu_memory_utilization=0.6 --tasks truthfulqa_mc2 --num_fewshot 0 --batch_size 16
vllm (pretrained=/home/mgoin/code/Nemotron-4-340B-Base-hf-FP8,tensor_parallel_size=8,distributed_executor_backend=ray,max_model_len=4096,gpu_memory_utilization=0.6), gen_kwargs: (None), limit: None, num_fewshot: 0, batch_size: 16
|    Tasks     |Version|Filter|n-shot|Metric|   |Value |   |Stderr|
|--------------|------:|------|-----:|------|---|-----:|---|-----:|
|truthfulqa_mc2|      2|none  |     0|acc   |↑  |0.4869|±  |0.0142|

lm_eval --model vllm --model_args pretrained=/home/mgoin/code/Nemotron-4-340B-Base-hf-FP8,tensor_parallel_size=8,distributed_executor_backend="ray",max_model_len=4096,gpu_memory_utilization=0.6 --tasks winogrande --num_fewshot 5 --batch_size 16
vllm (pretrained=/home/mgoin/code/Nemotron-4-340B-Base-hf-FP8,tensor_parallel_size=8,distributed_executor_backend=ray,max_model_len=4096,gpu_memory_utilization=0.6), gen_kwargs: (None), limit: None, num_fewshot: 5, batch_size: 16
|  Tasks   |Version|Filter|n-shot|Metric|   |Value |   |Stderr|
|----------|------:|------|-----:|------|---|-----:|---|-----:|
|winogrande|      1|none  |     5|acc   |↑  |0.8887|±  |0.0088|

The original paper evals for reference:

The Minitron paper has more evals as well: