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

Merge remote-tracking branch 'origin/united' into model-structure-update

Browse Source
This commit is contained in:
0cc4m
2023-05-04 07:31:13 +02:00
parent 1166c07bc3 a87d5d6f23
commit ef358fdf5a
5 changed files with 110 additions and 20 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
aiserver.py
View File

@@ -3555,7 +3555,7 @@ def actionsubmit(data, actionmode=0, force_submit=False, force_prompt_gen=False,
botname = (koboldai_vars.botname + ":")
else:
botname = ""
data = re.sub(r'\n+', ' ', data)
data = re.sub(r'\n+\Z', '', data)
if(len(data)):
data = f"\n{koboldai_vars.chatname}: {data}\n{botname}"
@@ -6354,7 +6354,7 @@ def UI_2_download_story():
@logger.catch
def UI_2_Set_Selected_Text(data):
if not koboldai_vars.quiet:
print("Updating Selected Text: {}".format(data))
logger.info("Updating Selected Text: {}".format(data))
action_id = int(data["id"])
if not koboldai_vars.actions.actions[action_id].get("Original Text"):

10
koboldai_settings.py
View File

@@ -397,8 +397,8 @@ class koboldai_vars(object):
######################################### Setup Author's Note Data ########################################################
authors_note_text = self.authornotetemplate.replace("<|>", self.authornote)
if len(authors_note_text) > 0 and authors_note_text[-1] not in [" ", "\n"]:
authors_note_text += " "
if len(authors_note_text) > 0 and authors_note_text[0] not in [" ", "\n"]:
authors_note_text = " " + authors_note_text
authors_note_data = [[x, self.tokenizer.decode(x)] for x in self.tokenizer.encode(authors_note_text)]
if used_tokens + len(authors_note_data) <= token_budget:
used_tokens += len(authors_note_data)
@@ -1393,7 +1393,11 @@ class KoboldStoryRegister(object):
self.action_count = -1
# The id of the last submission action, or 0 if the last append was not a submission
self.submission_id = 0
self.sentence_re = re.compile(r"[^.!?]*[.!?]+\"?\s*", re.S)
# A regular expression used to break the story into sentences so that the author's
# note can be inserted with minimal disruption. Avoid ending a sentance with
# whitespace because most tokenizers deal better with whitespace at the beginning of text.
# Search for sentence end delimeters (i.e. .!?) and also capture closing parenthesis and quotes.
self.sentence_re = re.compile(r".*?[.!?]+(?=[.!?\]\)}>'\"›»\s])[.!?\]\)}>'\"›»]*", re.S)
self.story_settings = story_settings
self.tts_model = None
self.tortoise = None

83
modeling/inference_models/hf.py
View File

@@ -20,8 +20,89 @@ class HFInferenceModel(InferenceModel):
def _post_load(self) -> None:
# These are model specific tokenizer overrides if a model has bad defaults
if utils.koboldai_vars.model_type == "llama":
self.tokenizer.decode_with_prefix_space = True
# Note: self.tokenizer is a GenericTokenizer, and self.tokenizer.tokenizer is the actual LlamaTokenizer
self.tokenizer.add_bos_token = False
# HF transformers no longer supports decode_with_prefix_space
# We work around this by wrapping decode, encode, and __call__
# with versions that work around the 'prefix space' misfeature
# of sentencepiece.
vocab = self.tokenizer.convert_ids_to_tokens(range(self.tokenizer.vocab_size))
has_prefix_space = {i for i, tok in enumerate(vocab) if tok.startswith("▁")}
# Wrap 'decode' with a method that always returns text starting with a space
# when the head token starts with a space. This is what 'decode_with_prefix_space'
# used to do, and we implement it using the same technique (building a cache of
# tokens that should have a prefix space, and then prepending a space if the first
# token is in this set.) We also work around a bizarre behavior in which decoding
# a single token 13 behaves differently than decoding a squence containing only [13].
original_decode = type(self.tokenizer.tokenizer).decode
def decode_wrapper(self, token_ids, *args, **kwargs):
first = None
# Note, the code below that wraps single-value token_ids in a list
# is to work around this wonky behavior:
# >>> t.decode(13)
# '<0x0A>'
# >>> t.decode([13])
# '\n'
# Not doing this causes token streaming to receive <0x0A> characters
# instead of newlines.
if isinstance(token_ids, int):
first = token_ids
token_ids = [first]
elif hasattr(token_ids, 'dim'): # Check for e.g. torch.Tensor
# Tensors don't support the Python standard of 'empty is False'
# and the special case of dimension 0 tensors also needs to be
# handled separately.
if token_ids.dim() == 0:
first = int(token_ids.item())
token_ids = [first]
elif len(token_ids) > 0:
first = int(token_ids[0])
elif token_ids:
first = token_ids[0]
result = original_decode(self, token_ids, *args, **kwargs)
if first is not None and first in has_prefix_space:
result = " " + result
return result
# GenericTokenizer overrides __setattr__ so we need to use object.__setattr__ to bypass it
object.__setattr__(self.tokenizer, 'decode', decode_wrapper.__get__(self.tokenizer))
# Wrap encode and __call__ to work around the 'prefix space' misfeature also.
# The problem is that "Bob" at the start of text is encoded as if it is
# " Bob". This creates a problem because it means you can't split text, encode
# the pieces, concatenate the tokens, decode them, and get the original text back.
# The workaround is to prepend a known token that (1) starts with a space; and
# (2) is not the prefix of any other token. After searching through the vocab
# " ," (space comma) is the only token containing only printable ascii characters
# that fits this bill. By prepending ',' to the text, the original encode
# method always returns [1919, ...], where the tail of the sequence is the
# actual encoded result we want without the prefix space behavior.
original_encode = type(self.tokenizer.tokenizer).encode
def encode_wrapper(self, text, *args, **kwargs):
if type(text) is str:
text = ',' + text
result = original_encode(self, text, *args, **kwargs)
result = result[1:]
else:
result = original_encode(self, text, *args, **kwargs)
return result
object.__setattr__(self.tokenizer, 'encode', encode_wrapper.__get__(self.tokenizer))
# Since 'encode' is documented as being deprecated, also override __call__.
# This doesn't appear to currently be used by KoboldAI, but doing so
# in case someone uses it in the future.
original_call = type(self.tokenizer.tokenizer).__call__
def call_wrapper(self, text, *args, **kwargs):
if type(text) is str:
text = ',' + text
result = original_call(self, text, *args, **kwargs)
result = result[1:]
else:
result = original_call(self, text, *args, **kwargs)
return result
object.__setattr__(self.tokenizer, '__call__', call_wrapper.__get__(self.tokenizer))
elif utils.koboldai_vars.model_type == "opt":
self.tokenizer._koboldai_header = self.tokenizer.encode("")
self.tokenizer.add_bos_token = False

25
static/koboldai.js
View File

@@ -3127,7 +3127,6 @@ function gametextwatcher(records) {
}
}
}
//console.log(dirty_chunks);
}
function fix_dirty_game_text() {
@@ -3152,16 +3151,26 @@ function fix_dirty_game_text() {
//Fixing text outside of chunks
for (node of game_text.childNodes) {
if ((!(node instanceof HTMLElement) || !node.hasAttribute("chunk")) && (node.textContent.trim() != "")) {
console.log("Found Node that needs to be combined");
console.log(node);
//We have a text only node. It should be moved into the previous chunk
//We have a text only node. It should be moved into the previous chunk if it is marked as dirty, next node if not and it's dirty, or the previous if neither is dirty
var node_text = ""
if (node instanceof HTMLElement) {
node.previousElementSibling.innerText = node.previousElementSibling.innerText + node.innerText;
node_text = node.innerText;
} else {
node.previousElementSibling.innerText = node.previousElementSibling.innerText + node.data;
node_text = node.data;
}
if (!dirty_chunks.includes(node.previousElementSibling.getAttribute("chunk"))) {
dirty_chunks.push(node.previousElementSibling.getAttribute("chunk"));
if (!(node.nextElementSibling) || !(dirty_chunks.includes(node.nextElementSibling.getAttribute("chunk"))) || dirty_chunks.includes(node.previousElementSibling.getAttribute("chunk"))) {
node.previousElementSibling.innerText = node.previousElementSibling.innerText + node_text;
if (!dirty_chunks.includes(node.previousElementSibling.getAttribute("chunk"))) {
dirty_chunks.push(node.previousElementSibling.getAttribute("chunk"));
}
} else {
node.nextElementSibling.innerText = node.nextElementSibling.innerText + node_text;
}
//Looks like sometimes it splits the parent. Let's look for that and fix it too
if (node.nextElementSibling && (node.nextElementSibling.getAttribute("chunk") == node.previousElementSibling.getAttribute("chunk"))) {
node.previousElementSibling.innerText = node.previousElementSibling.innerText + node.nextElementSibling.innerText;
node.nextElementSibling.remove();
}
node.remove();
}

8
utils.py
View File

@@ -144,21 +144,17 @@ def singlelineprocessing(txt, koboldai_vars):
return txt
def chatmodeprocessing(txt, koboldai_vars):
chatregex = re.compile(r'%s:[.|\n|\W|\w]*'%koboldai_vars.chatname)
chatregex = re.compile(r'\s+%s:[.|\n|\W|\w]*'%koboldai_vars.chatname)
txt = chatregex.sub('', txt)
if(len(koboldai_vars.actions) > 0):
if(len(koboldai_vars.actions[-1]) > 0):
action = koboldai_vars.actions[-1]
lastchar = action[-1] if len(action) else ""
else:
# Last action is blank, this should never happen, but
# since it did let's bail out.
return txt
else:
action = koboldai_vars.prompt
lastchar = action[-1] if len(action) else ""
if(lastchar != "\n"):
txt = txt + "\n"
return txt
#==================================================================#
@@ -745,4 +741,4 @@ def applyoutputformatting(txt, no_sentence_trimming=False, no_single_line=False)
txt = txt.replace(sub["trueTarget"], sub["substitution"])
txt = txt.replace(sub["target"], sub["substitution"])
return txt
return txt