List your company

Training Efficiency Measurement on Llama 3.1-8B Base (S=512, B200)

3.1Peak Memory Usage (GiB)AdaLomo
-2.28834.08170.45106.819Jun 22, 2026

Evaluation results

MethodLinks
AdaLomoModel=Qwen3-0.6B, Protocol=B, Precision=BF16-everywhere, FlashAttention=FA4, Liger Kernel=Enabled2026.063.1584.9
GaLoreModel=Qwen3-0.6B, Protocol=B, Precision=BF16-everywhere, FlashAttention=FA4, Liger Kernel=Enabled, Rank=1282026.063.9229.7
bnb 8-bit AdamWModel=Qwen3-0.6B, Protocol=B, Precision=BF16-everywhere, FlashAttention=FA4, Liger Kernel=Enabled2026.064.2232.7
APOLLO-MiniModel=Qwen3-0.6B, Protocol=B, Precision=BF16-everywhere, FlashAttention=FA4, Liger Kernel=Enabled, Rank=2562026.064.2249.4
FlashOptimModel=Qwen3-0.6B, Protocol=B, Precision=BF16-everywhere, FlashAttention=FA4, Liger Kernel=Enabled2026.064.6221.5
FORGEModel=Qwen3-0.6B, Protocol=A, Precision=BF16-everywhere, FlashAttention=FA4, Liger Kernel=Enabled2026.065.1181.5
optimi.gradient_releaseModel=Qwen3-0.6B, Protocol=B, Precision=BF16-everywhere, FlashAttention=FA4, Liger Kernel=Enabled2026.066.2248.6
fused AdamWModel=Qwen3-0.6B, Protocol=A, Precision=BF16-everywhere, FlashAttention=FA4, Liger Kernel=Enabled2026.066.7186.9
AdaLomoModel=Qwen3-1.7B, Protocol=B, Precision=BF16-everywhere, FlashAttention=FA4, Liger Kernel=Enabled2026.066.9614.3
vanilla AdamWModel=Qwen3-0.6B, Protocol=A, Precision=BF16-everywhere, FlashAttention=FA4, Liger Kernel=Enabled2026.067.2188.4
GaLoreModel=Qwen3-1.7B, Protocol=B, Precision=BF16-everywhere, FlashAttention=FA4, Liger Kernel=Enabled, Rank=1282026.069.4267.3
APOLLO-MiniModel=Qwen3-1.7B, Protocol=B, Precision=BF16-everywhere, FlashAttention=FA4, Liger Kernel=Enabled, Rank=2562026.069.8246
FlashOptimModel=Qwen3-1.7B, Protocol=B, Precision=BF16-everywhere, FlashAttention=FA4, Liger Kernel=Enabled2026.0610.4220.6
FORGEModel=Qwen3-1.7B, Protocol=A, Precision=BF16-everywhere, FlashAttention=FA4, Liger Kernel=Enabled2026.0610.7180.6
bnb 8-bit AdamWModel=Qwen3-1.7B, Protocol=B, Precision=BF16-everywhere, FlashAttention=FA4, Liger Kernel=Enabled2026.0611242.4
AdaLomoModel=Qwen3-4B, Protocol=B, Precision=BF16-everywhere, FlashAttention=FA4, Liger Kernel=Enabled2026.0612.1905.6
optimi.gradient_releaseModel=Qwen3-1.7B, Protocol=B, Precision=BF16-everywhere, FlashAttention=FA4, Liger Kernel=Enabled2026.0615255.9
fused AdamWModel=Qwen3-1.7B, Protocol=A, Precision=BF16-everywhere, FlashAttention=FA4, Liger Kernel=Enabled2026.0618192.1
GaLoreModel=Qwen3-4B, Protocol=B, Precision=BF16-everywhere, FlashAttention=FA4, Liger Kernel=Enabled, Rank=1282026.0619.1274.2
vanilla AdamWModel=Qwen3-1.7B, Protocol=A, Precision=BF16-everywhere, FlashAttention=FA4, Liger Kernel=Enabled2026.0620150.5
APOLLO-MiniModel=Qwen3-4B, Protocol=B, Precision=BF16-everywhere, FlashAttention=FA4, Liger Kernel=Enabled, Rank=2562026.0620.1347.2
FORGEModel=Qwen3-4B, Protocol=A, Precision=BF16-everywhere, FlashAttention=FA4, Liger Kernel=Enabled2026.0621.3222.1
FlashOptimModel=Qwen3-4B, Protocol=B, Precision=BF16-everywhere, FlashAttention=FA4, Liger Kernel=Enabled2026.0622.6311.3
AdaLomoProtocol=Protocol B, Approximated AdamW Dynamics=true, Batch Size=1, Attention=FlashAttention-4, Hardware=B200, Sequence Length=5122026.0623.5708
AdaLomoModel=Llama-3.1-8B, Protocol=B, Precision=BF16-everywhere, FlashAttention=FA4, Liger Kernel=Enabled2026.0623.5708
bnb 8-bit AdamWModel=Qwen3-4B, Protocol=B, Precision=BF16-everywhere, FlashAttention=FA4, Liger Kernel=Enabled2026.0624.4299.7
AdaLomoModel=Qwen3-8B, Protocol=B, Precision=BF16-everywhere, FlashAttention=FA4, Liger Kernel=Enabled2026.0625.2946.7
optimi.gradient_releaseModel=Qwen3-4B, Protocol=B, Precision=BF16-everywhere, FlashAttention=FA4, Liger Kernel=Enabled2026.0633.5293.3
FORGEProtocol=Protocol A, Batch Size=1, Attention=FlashAttention-4, Hardware=B200, Sequence Length=5122026.0635.8172.6
FORGEModel=Llama-3.1-8B, Protocol=A, Precision=BF16-everywhere, FlashAttention=FA4, Liger Kernel=Enabled2026.0635.8172.6
GaLorer=128, Protocol=Protocol B, Approximated AdamW Dynamics=true, Batch Size=1, Attention=FlashAttention-4, Hardware=B200, Sequence Length=5122026.0637.2242.1
GaLoreModel=Llama-3.1-8B, Protocol=B, Precision=BF16-everywhere, FlashAttention=FA4, Liger Kernel=Enabled, Rank=1282026.0637.2242.1
FORGEModel=Qwen3-8B, Protocol=A, Precision=BF16-everywhere, FlashAttention=FA4, Liger Kernel=Enabled2026.0637.6222
APOLLO-Minir=256, Protocol=Protocol B, Approximated AdamW Dynamics=true, Batch Size=1, Attention=FlashAttention-4, Hardware=B200, Sequence Length=5122026.0638.3279.5
APOLLO-MiniModel=Llama-3.1-8B, Protocol=B, Precision=BF16-everywhere, FlashAttention=FA4, Liger Kernel=Enabled, Rank=2562026.0638.3279.5
GaLoreModel=Qwen3-8B, Protocol=B, Precision=BF16-everywhere, FlashAttention=FA4, Liger Kernel=Enabled, Rank=1282026.0638.9269.2
APOLLO-MiniModel=Qwen3-8B, Protocol=B, Precision=BF16-everywhere, FlashAttention=FA4, Liger Kernel=Enabled, Rank=2562026.0640332.8
fused AdamWModel=Qwen3-4B, Protocol=A, Precision=BF16-everywhere, FlashAttention=FA4, Liger Kernel=Enabled2026.0640.1237.4
FlashOptimProtocol=Protocol B, Batch Size=1, Attention=FlashAttention-4, Hardware=B200, Sequence Length=5122026.0642.2218.7
FlashOptimModel=Llama-3.1-8B, Protocol=B, Precision=BF16-everywhere, FlashAttention=FA4, Liger Kernel=Enabled2026.0642.2218.7
FlashOptimModel=Qwen3-8B, Protocol=B, Precision=BF16-everywhere, FlashAttention=FA4, Liger Kernel=Enabled2026.0643.4268.7
bnb 8-bit AdamWProtocol=Protocol B, Batch Size=1, Attention=FlashAttention-4, Hardware=B200, Sequence Length=5122026.0645.3364.6
bnb 8-bit AdamWModel=Llama-3.1-8B, Protocol=B, Precision=BF16-everywhere, FlashAttention=FA4, Liger Kernel=Enabled2026.0645.3364.6
vanilla AdamWModel=Qwen3-4B, Protocol=A, Precision=BF16-everywhere, FlashAttention=FA4, Liger Kernel=Enabled2026.0646.2251
bnb 8-bit AdamWModel=Qwen3-8B, Protocol=B, Precision=BF16-everywhere, FlashAttention=FA4, Liger Kernel=Enabled2026.0646.2397.2
optimi.gradient_releaseProtocol=Protocol B, Batch Size=1, Attention=FlashAttention-4, Hardware=B200, Sequence Length=5122026.0664.5240.6
FORGEModel=Qwen3-14B, Protocol=A, Precision=BF16-everywhere, FlashAttention=FA4, Liger Kernel=Enabled2026.0664.5—
optimi.gradient_releaseModel=Llama-3.1-8B, Protocol=B, Precision=BF16-everywhere, FlashAttention=FA4, Liger Kernel=Enabled2026.0664.5240.6
GaLoreModel=Qwen3-14B, Protocol=B, Precision=BF16-everywhere, FlashAttention=FA4, Liger Kernel=Enabled, Rank=1282026.0665.9300.1
fused AdamWProtocol=Protocol A, Batch Size=1, Attention=FlashAttention-4, Hardware=B200, Sequence Length=5122026.0675205.2
fused AdamWModel=Llama-3.1-8B, Protocol=A, Precision=BF16-everywhere, FlashAttention=FA4, Liger Kernel=Enabled2026.0675205.2
fused AdamWModel=Qwen3-8B, Protocol=A, Precision=BF16-everywhere, FlashAttention=FA4, Liger Kernel=Enabled2026.0676.5252.6
FlashOptimModel=Qwen3-14B, Protocol=B, Precision=BF16-everywhere, FlashAttention=FA4, Liger Kernel=Enabled2026.0676.6344.4
bnb 8-bit AdamWModel=Qwen3-14B, Protocol=B, Precision=BF16-everywhere, FlashAttention=FA4, Liger Kernel=Enabled2026.0683.6555.3
vanilla AdamWProtocol=Protocol A, Batch Size=1, Attention=FlashAttention-4, Hardware=B200, Sequence Length=5122026.0689.9240.1
vanilla AdamWModel=Llama-3.1-8B, Protocol=A, Precision=BF16-everywhere, FlashAttention=FA4, Liger Kernel=Enabled2026.0689.9240.1
vanilla AdamWModel=Qwen3-8B, Protocol=A, Precision=BF16-everywhere, FlashAttention=FA4, Liger Kernel=Enabled2026.0691.8283.9
fused AdamWModel=Qwen3-14B, Protocol=A, Precision=BF16-everywhere, FlashAttention=FA4, Liger Kernel=Enabled2026.06137.8226.3