rixty/KoboldAI-Client
1
1
Fork 0
mirror of https://github.com/KoboldAI/KoboldAI-Client.git synced 2025-02-01 10:06:44 +01:00
Code Issues Projects Releases Wiki Activity

Slightly increased performance in breakmodel mode

Browse Source 
Commit a283d34b2731abfe7f5f1e939117491f0755cedb made breakmodel mode
slower. Performance has been restored to how it was before that commit.
This commit is contained in:
Gnome Ann 2021-10-05 10:25:06 -04:00
parent 231621e7c2
commit f9e6a6da17
1 changed files with 1 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
breakmodel.py
View File

@ -434,7 +434,7 @@ def new_forward(
device = primary_device if i < ram_blocks else bisect.bisect_right(cumulative_gpu_blocks, i - ram_blocks) device = primary_device if i < ram_blocks else bisect.bisect_right(cumulative_gpu_blocks, i - ram_blocks)
outputs = block( outputs = block(
hidden_states.to(device) if breakmodel and hidden_states is not None else hidden_states, hidden_states.to(device) if breakmodel and hidden_states is not None else hidden_states,
layer_past=tuple(v.to(device) for v in layer_past if v is not None) if breakmodel and layer_past is not None else layer_past, layer_past=tuple(v.to(device) for v in layer_past if v is not None) if breakmodel and layer_past is not None and i >= ram_blocks and len(layer_past) and layer_past[0].device.index != device else layer_past,
attention_mask=attn_mask.to(device) if breakmodel and attn_mask is not None else attn_mask, attention_mask=attn_mask.to(device) if breakmodel and attn_mask is not None else attn_mask,
head_mask=head_mask[i].to(device) if breakmodel and head_mask[i] is not None else head_mask[i], head_mask=head_mask[i].to(device) if breakmodel and head_mask[i] is not None else head_mask[i],
use_cache=use_cache, use_cache=use_cache,