| --- |
| license: llama2 |
| pipeline_tag: text-generation |
| library_name: gguf |
| --- |
| GGUF importance matrix (imatrix) quants for https://huggingface.co/LargeWorldModel/LWM-Text-Chat-128K |
| The importance matrix was trained for 100K tokens (200 batches of 512 tokens) using wiki.train.raw. |
|
|
| * The imatrix Q4-K quant fits with 32K context on 24GB and gives me ~100 t/s inference on a 3090. |
| * With IQ3_XXS it seems to fit ~37K context on 24GB (and it is even faster than Q4-K). |
| * With either quant on a 3090 it seems to decode context at well over 2000 t/s. |
| * Using Q8 K-cache (instead of F16) you can fit up to 43-44K context but inference speed goes down a little bit. |
| * Also for some reason I need to use 1.0 penalty to avoid the response being cut-off. |
| |
| | Layers | Context | [Template](https://github.com/LargeWorldModel/LWM/blob/9aaaa1e864bfcf31b66028e782395a22f4817535/scripts/eval_needle.py#L48) | |
| | --- | --- | --- | |
| | <pre>32</pre> | <pre>131072</pre> | <pre>You are a helpful assistant.<br>USER:<br>{context}<br>{question}<br>Don't give information outside the document or repeat your findings. Keep your response short and direct.<br>ASSISTANT:<br>{response}</pre> | |
