WARNING: gemma-2-27b models don't run well in float16 precision.
1
5 commits
2 linked in READMEs
updated Jul 21, 2024
README
WARNING: gemma-2-27b models don't run well in float16 precision.
This FLUTE-quantized model is released in bfloat16.
| Wiki | C4 | PIQA | ARC-E | ARC-C | HellaSwag | Wino | Avg. | |
|---|---|---|---|---|---|---|---|---|
| Unquantized | 5.70 | 8.98 | 83.24 | 87.84 | 62.88 | 65.35 | 79.24 | 75.71 |
| W4G64 | 5.69 | 9.31 | 82.53 | 86.45 | 59.22 | 64.13 | 78.21 | 74.11 |
| W3G64 | TBD | TBD | TBD | TBD | TBD | TBD | TBD | TBD |
Evaluations are provided for models with learned scales.
Benchmark scores (zero-shot) are computed with lm-evaluation-harness.
conversational
endpoints_compatible
gemma2
safetensors
text-generation
text-generation-inference
transformers
WARNING: gemma-2-27b models don't run well in float16 precision.
1
5 commits
2 linked in READMEs
updated Jul 21, 2024
README
WARNING: gemma-2-27b models don't run well in float16 precision.
This FLUTE-quantized model is released in bfloat16.
| Wiki | C4 | PIQA | ARC-E | ARC-C | HellaSwag | Wino | Avg. | |
|---|---|---|---|---|---|---|---|---|
| Unquantized | 5.70 | 8.98 | 83.24 | 87.84 | 62.88 | 65.35 | 79.24 | 75.71 |
| W4G64 | 5.69 | 9.31 | 82.53 | 86.45 | 59.22 | 64.13 | 78.21 | 74.11 |
| W3G64 | TBD | TBD | TBD | TBD | TBD | TBD | TBD | TBD |
Evaluations are provided for models with learned scales.
Benchmark scores (zero-shot) are computed with lm-evaluation-harness.
conversational
endpoints_compatible
gemma2
safetensors
text-generation
text-generation-inference
transformers