Performance Summary

This document provides performance benchmarks for various large language models using NeMo AutoModel with the PyTorch backend.

Pre-Training Performance

The table below shows training performance for full sequences with no padding across different model architectures and scales.

Model	#GPUs	GBS	MBS	LBS	GA	Seq Length	TP	PP	CP	EP	VP	FSDP	Kernel Optimizations	Time per Global Step (s)	Model TFLOPs/sec/GPU	Tokens/sec/GPU
Nemotron V3 Super 120B (26.02)	64	512	2	2	4	4096	1	1	1	64	-	64	TE + DeepEP + TorchSDPA	7.286	334	4,497
Nemotron V3 Nano 30B (26.02)	8	512	4	4	16	4096	1	1	1	8	-	8	TE + DeepEP + TorchSDPA	15.614	328	16,789
DeepSeek V3 671B	1024	8192	1	8	4	4096	1	4	1	64	8	256	TE + DeepEP	37.87	216	865
DeepSeek V3 671B	256	512	1	8	1	4096	1	4	1	64	8	64	TE + DeepEP	8.18	250	1,002
Kimi K2	256	512	1	8	2	4096	1	8	1	32	4	32	TE + DeepEP	8.86	189	924
Qwen3 MoE 30B	8	512	4	4	16	4096	1	1	1	8	-	8	TE + DeepEP	21.773	277	12,040
GPT-OSS 20B	8	256	2	2	16	4096	1	1	1	-	-	8	TE + DeepEP + FlexAttn	10.04	279	13,058
GPT-OSS 120B	64	512	2	2	4	4096	1	1	1	-	-	64	TE + DeepEP + FlexAttn	4.30	231	7,626
Llama3 70B	64	128	1	1	4	8192	1	1	2	-	-	32	TE + fsdp2_prefetch	18.90	389	866.77

The table below shows fine-tuning (LoRA) performance for full sequences with no padding across different model architectures and scales.

Model	#GPUs	GBS	MBS	LBS	GA	Seq Length	TP	PP	CP	EP	VP	FSDP	Kernel Optimizations	Time per Global Step (s)	Model TFLOPs/sec/GPU	Tokens/sec/GPU
Llama3 8B	1	32	2	2	16	4096	1	1	1	-	1	1	TE + triton	10.51	402	12472.87
Qwen2.5 7B	1	32	2	2	16	4096	1	1	1	-	1	1	TE + triton	9.29	423	14110.05
Llama3 70B	8	32	2	2	4	4096	2	1	1	-	1	4	TE + triton + fsdp2_prefetch	15.00	316	1091.85
Qwen2.5 32B	8	32	2	2	4	4096	2	1	1	-	1	4	TE + triton + fsdp2_prefetch	7.28	301	2250.31
Llama3 70B 2-node	16	32	2	2	2	4096	2	1	1	-	1	8	TE + triton + fsdp2_prefetch	8.32	285	984.85
Qwen2.5 32B 2-node	16	32	2	2	2	4096	2	1	1	-	1	8	TE + triton + fsdp2_prefetch	3.95	277	2072.89

MFU: Model FLOPs Utilization - ratio of achieved compute to peak hardware capability
TP: Tensor Parallelism - splits individual layers across GPUs
PP: Pipeline Parallelism - splits model layers into stages
EP: Expert Parallelism - distributes MoE experts across GPUs
DP: Data Parallelism - replicates model and splits data
VP: Virtual Pipeline - number of pipeline stages per GPU for interleaving
MBS: Micro-Batch Size - size of one forward pass in pipeline
LBS: Local Batch Size - size of one step per GPU
GBS: Global Batch Size - total batch size across all GPUs
GA: Gradient Accumulation - number of local-batches before optimizer step
TE: Transformer Engine kernel optimizations - RMSNorm, Linear and DotProductAttention
DeepEP: Deep Expert Parallelism - advanced EP routing for MoE models
FlexAttn: PyTorch's Flex Attention

Pre-training and fine-tuning (LoRA) benchmark configurations are available in examples/llm_benchmark/:

deepseek_v3_te_deepep.yaml - DeepSeek V3 with TE + DeepEP
kimi_k2_te_deepep.yaml - Kimi K2 optimized configuration
qwen3_moe_30b_te_deepep.yaml - Qwen3 MoE with TE + DeepEP
gptoss_20b_te_deepep.yaml - GPT-OSS 20B with optimizations
gptoss_120b_te_deepep.yaml - GPT-OSS 120B optimized
custom_llama3_1_70b_pretrain_benchmark_8nodes.yaml - Llama3-70B optimized
llama3_1_8b_peft_benchmark.yaml - Llama-8B fine-tuning (LoRA) optimized
qwen2_5_7b_peft_benchmark.yaml - Qwen2.5-7B fine-tuning (LoRA) optimized
custom_llama3_3_70b_instruct_peft_benchmark.yaml - Llama-70B fine-tuning (LoRA) optimized
custom_qwen2_5_32b_peft_benchmark.yaml - Qwen2.5-32B fine-tuning (LoRA) optimized
custom_llama3_3_70b_instruct_peft_benchmark_2nodes.yaml - Llama-70B fine-tuning (LoRA) optimized on 2 nodes
custom_qwen2_5_32b_peft_benchmark_2nodes.yaml - Qwen2.5-32B fine-tuning (LoRA) optimized on 2 nodes

:::{note}

All benchmarks use mock data for consistent performance measurement.
Fake balanced gate is enabled to simulate ideal expert routing.
No gradient clipping applied for pure performance measurement.
MFU calculated using peak TFLOPs for the system (989 for BF16 H100).
Step times include forward and backward passes + optimizer step for the global batch. :::