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

Partial fix for text context

Browse Source
This commit is contained in:
ebolam
2022-09-29 12:12:46 -04:00
parent a3bad7b713
commit 51b8b1d223
2 changed files with 9 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
aiserver.py
View File

@@ -2673,7 +2673,7 @@ def load_model(use_gpu=True, gpu_layers=None, disk_layers=None, initial_load=Fal
raise RuntimeError("One of your GPUs ran out of memory when KoboldAI tried to load your model.")
raise e
tokenizer.save_pretrained("models/{}".format(koboldai_vars.model.replace('/', '_')))
model.save_pretrained("models/{}".format(koboldaivars.model.replace('/', '_')), max_shard_size="500MiB")
model.save_pretrained("models/{}".format(koboldai_vars.model.replace('/', '_')), max_shard_size="500MiB")
koboldai_vars.modeldim = get_hidden_size_from_model(model)
# Is CUDA available? If so, use GPU, otherwise fall back to CPU
if(koboldai_vars.hascuda and koboldai_vars.usegpu):

11
koboldai_settings.py
View File

@@ -260,6 +260,8 @@ class koboldai_vars(object):
game_context = []
authors_note_final = self.authornotetemplate.replace("<|>", self.authornote)
used_all_tokens = False
with open("test.txt", "w") as f:
f.write(str(action_text_split))
for action in range(len(self.actions)):
self.actions.set_action_in_ai(action, used=False)
for i in range(len(action_text_split)-1, -1, -1):
@@ -270,12 +272,14 @@ class koboldai_vars(object):
game_text = "{}{}".format(authors_note_final, game_text)
game_context.insert(0, {"type": "authors_note", "text": authors_note_final})
length = 0 if self.tokenizer is None else len(self.tokenizer.encode(action_text_split[i][0]))
print("Length for sentence {}: {}".format(i, length))
if length+used_tokens <= token_budget and not used_all_tokens:
used_tokens += length
selected_text = action_text_split[i][0]
action_text_split[i][3] = True
game_text = "{}{}".format(selected_text, game_text)
game_context.insert(0, {"type": "action", "text": selected_text})
print("Adding in game for the following actions: {}".format(action_text_split[i][1]))
for action in action_text_split[i][1]:
if action >= 0:
self.actions.set_action_in_ai(action)
@@ -311,6 +315,7 @@ class koboldai_vars(object):
self.worldinfo_v2.set_world_info_used(wi['uid'])
else:
used_all_tokens = True
break
#if we don't have enough actions to get to author's note depth then we just add it right before the game text
if len(action_text_split) < self.andepth and self.authornote != "":
@@ -1380,7 +1385,7 @@ class KoboldStoryRegister(object):
def to_sentences(self):
#we're going to split our actions by sentence for better context. We'll add in which actions the sentence covers. Prompt will be added at a -1 ID
actions = {i: self.actions[i] for i in range(len(self.actions))}
actions = {i: self.actions[i]['Selected Text'] for i in range(len(self.actions))}
actions[-1] = self.story_settings.prompt
action_text = self.__str__()
action_text = "{}{}".format(self.story_settings.prompt, action_text)
@@ -1410,12 +1415,12 @@ class KoboldStoryRegister(object):
if advance_action:
Action_Position[0] += 1
if Action_Position[0] >= max(actions):
if Action_Position[0] > max(actions):
break
Action_Position[1] = len(actions[Action_Position[0]])
if advance_sentence:
Sentence_Position[0] += 1
if Sentence_Position[0] >= len(action_text_split):
if Sentence_Position[0] > len(action_text_split):
break
Sentence_Position[1] = len(action_text_split[Sentence_Position[0]][0])
#OK, action_text_split now contains a list of [sentence including trailing space if needed, [action IDs that sentence includes]]