Language Modeling on FW (val)
5.84PPLDense
Evaluation Results
| Method | Links | |
|---|---|---|
| DenseSparsity=0%, Latency=-, Model=Llama-2-13B2025.05 | 5.84 | |
| EvoPressSparsity=12.5%, Latency=29m, Model=Llama-2-13B2025.05 | 6.62 | |
| UniCuCoSparsity=12.5%, Latency=<1s, Model=Llama-2-13B2025.05 | 6.94 | |
| Cosine (Window)Sparsity=12.5%, Latency=16s, Model=Llama-2-13B2025.05 | 6.97 | |
| ShortGPTSparsity=12.5%, Latency=<1s, Model=Llama-2-13B2025.05 | 7.19 | |
| Weight SubcloningSparsity=12.5%, Latency=<1s, Model=Llama-2-13B2025.05 | 7.26 | |
| EvoPressSparsity=25%, Latency=28.6m, Model=Llama-2-13B2025.05 | 8.32 | |
| UniCuCoSparsity=25%, Latency=<1s, Model=Llama-2-13B2025.05 | 10.14 | |
| Cosine (Window)Sparsity=25%, Latency=24s, Model=Llama-2-13B2025.05 | 10.34 | |
| EvoPressSparsity=37.5%, Latency=26.8m, Model=Llama-2-13B2025.05 | 13 | |
| ShortGPTSparsity=25%, Latency=<1s, Model=Llama-2-13B2025.05 | 15.73 | |
| Weight SubcloningSparsity=25%, Latency=<1s, Model=Llama-2-13B2025.05 | 15.73 | |
| UniCuCoSparsity=37.5%, Latency=<1s, Model=Llama-2-13B2025.05 | 17.74 | |
| EvoPressSparsity=50%, Latency=28.5m, Model=Llama-2-13B2025.05 | 28.48 | |
| ShortGPTSparsity=37.5%, Latency=<1s, Model=Llama-2-13B2025.05 | 35.24 | |
| Weight SubcloningSparsity=37.5%, Latency=<1s, Model=Llama-2-13B2025.05 | 35.24 | |
| Cosine (Window)Sparsity=37.5%, Latency=24s, Model=Llama-2-13B2025.05 | 50.46 | |
| UniCuCoSparsity=50%, Latency=<1s, Model=Llama-2-13B2025.05 | 76.79 | |
| ShortGPTSparsity=50%, Latency=<1s, Model=Llama-2-13B2025.05 | 132.47 | |
| Weight SubcloningSparsity=50%, Latency=<1s, Model=Llama-2-13B2025.05 | 132.47 | |
| Cosine (Window)Sparsity=50%, Latency=17s, Model=Llama-2-13B2025.05 | 316.09 | |
| UniCuCoSparsity=62.5%, Latency=<1s, Model=Llama-2-13B2025.05 | 383.87 | |
| ShortGPTSparsity=62.5%, Latency=<1s, Model=Llama-2-13B2025.05 | 622.41 | |
| Weight SubcloningSparsity=62.5%, Latency=<1s, Model=Llama-2-13B2025.05 | 622.41 | |
| EvoPressSparsity=62.5%, Latency=23.5m, Model=Llama-2-13B2025.05 | 3,930.84 | |
| Cosine (Window)Sparsity=62.5%, Latency=13s, Model=Llama-2-13B2025.05 | 4,944.9 |