2023-10-22 00:15:54 +02:00
## What Works
| Loader | Loading 1 LoRA | Loading 2 or more LoRAs | Training LoRAs | Multimodal extension | Perplexity evaluation |
|----------------|----------------|-------------------------|----------------|----------------------|-----------------------|
2024-05-21 18:32:02 +02:00
| Transformers | ✅ | ✅\*\* | ✅\* | ✅ | ✅ |
2024-02-05 03:48:04 +01:00
| llama.cpp | ❌ | ❌ | ❌ | ❌ | use llamacpp_HF |
| llamacpp_HF | ❌ | ❌ | ❌ | ❌ | ✅ |
2023-10-22 00:15:54 +02:00
| ExLlamav2_HF | ✅ | ✅ | ❌ | ❌ | ✅ |
2024-02-06 15:26:27 +01:00
| ExLlamav2 | ✅ | ✅ | ❌ | ❌ | use ExLlamav2_HF |
2023-10-22 00:15:54 +02:00
| AutoGPTQ | ✅ | ❌ | ❌ | ✅ | ✅ |
2024-02-05 03:48:04 +01:00
| AutoAWQ | ? | ❌ | ? | ? | ✅ |
| HQQ | ? | ? | ? | ? | ✅ |
2023-10-22 00:15:54 +02:00
❌ = not implemented
✅ = implemented
\* Training LoRAs with GPTQ models also works with the Transformers loader. Make sure to check "auto-devices" and "disable_exllama" before loading the model.
2024-05-21 18:32:02 +02:00
\*\* Multi-LoRA in PEFT is tricky and the current implementation does not work reliably in all cases.
