rixty/KoboldAI-Client
1
1
Fork 0
mirror of https://github.com/KoboldAI/KoboldAI-Client.git synced 2025-06-05 21:59:24 +02:00
Code Issues Projects Releases Wiki Activity

Fix AMD ROCm exllama inference

Browse Source
This commit is contained in:
0cc4m
2023-06-13 10:11:29 +02:00
parent ebf7e2cf57
commit 0c7eaefb1a
1 changed files with 5 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
modeling/inference_models/exllama/class.py
View File

@@ -362,5 +362,10 @@ class model_backend(InferenceModel):
self.model_config.device_map.lm_head = "cuda:0"
self.model_config.device_map.norm = "cuda:0"
self.model_config.rmsnorm_no_half2 = bool(torch.version.hip)
self.model_config.rope_no_half2 = bool(torch.version.hip)
self.model_config.matmul_no_half2 = bool(torch.version.hip)
self.model_config.silu_no_half2 = bool(torch.version.hip)
self.model_name = parameters['custom_model_name'] if 'custom_model_name' in parameters else parameters['id']
self.path = parameters['path'] if 'path' in parameters else None