Mirrors/text-generation-webui
1
0
Fork 0
mirror of https://github.com/oobabooga/text-generation-webui.git synced 2024-11-28 18:48:04 +01:00
Code Issues Packages Projects Releases Wiki Activity

Optimize ExLlamav2 (non-HF) loader

Browse Source
This commit is contained in:
oobabooga 2023-12-20 07:31:42 -08:00
parent 258c695ead
commit b15f510154
1 changed files with 8 additions and 28 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

36
modules/exllamav2.py
View File

@ -1,4 +1,3 @@
import random
import traceback import traceback
from pathlib import Path from pathlib import Path
@ -10,7 +9,7 @@ from exllamav2 import (
ExLlamaV2Config, ExLlamaV2Config,
ExLlamaV2Tokenizer ExLlamaV2Tokenizer
) )
from exllamav2.generator import ExLlamaV2BaseGenerator, ExLlamaV2Sampler from exllamav2.generator import ExLlamaV2Sampler, ExLlamaV2StreamingGenerator
from modules import shared from modules import shared
from modules.logging_colors import logger from modules.logging_colors import logger
@ -64,7 +63,7 @@ class Exllamav2Model:
else: else:
cache = ExLlamaV2Cache(model) cache = ExLlamaV2Cache(model)
generator = ExLlamaV2BaseGenerator(model, cache, tokenizer) generator = ExLlamaV2StreamingGenerator(model, cache, tokenizer)
result = self() result = self()
result.model = model result.model = model
@ -115,41 +114,22 @@ class Exllamav2Model:
ids = self.tokenizer.encode(prompt, add_bos=state['add_bos_token'], encode_special_tokens=True) ids = self.tokenizer.encode(prompt, add_bos=state['add_bos_token'], encode_special_tokens=True)
ids = ids[:, -get_max_prompt_length(state):] ids = ids[:, -get_max_prompt_length(state):]
initial_len = ids.shape[-1]
if state['auto_max_new_tokens']: if state['auto_max_new_tokens']:
max_new_tokens = state['truncation_length'] - ids.shape[-1] max_new_tokens = state['truncation_length'] - ids.shape[-1]
else: else:
max_new_tokens = state['max_new_tokens'] max_new_tokens = state['max_new_tokens']
# _gen_begin_base self.generator.set_stop_conditions([])
self.cache.current_seq_len = 0 self.generator.begin_stream(ids, settings, loras=self.loras)
self.model.forward(ids[:, :-1], self.cache, input_mask=None, preprocess_only=True, loras=self.loras)
has_leading_space = False decoded_text = ''
for i in range(max_new_tokens): for i in range(max_new_tokens):
logits = self.model.forward(ids[:, -1:], self.cache, input_mask=None, loras=self.loras).float().cpu() chunk, eos, _ = self.generator.stream()
token, _, _ = ExLlamaV2Sampler.sample(logits, settings, ids, random.random(), self.tokenizer) if eos or shared.stop_everything:
ids = torch.cat([ids, token], dim=1)
if i == 0 and self.tokenizer.tokenizer.id_to_piece(int(token)).startswith('▁'):
has_leading_space = True
decoded_text = self.tokenizer.decode(ids[:, initial_len:], decode_special_tokens=not state['skip_special_tokens'])[0]
if has_leading_space:
decoded_text = ' ' + decoded_text
# Check the partial unicode character
if chr(0xfffd) in decoded_text:
is_last = i == max_new_tokens - 1
is_stopping = token.item() == self.tokenizer.eos_token_id or shared.stop_everything
# If we are not at the end of the generation, we skip this token
if not (is_last or is_stopping):
continue
if token.item() == self.tokenizer.eos_token_id or shared.stop_everything:
break break
decoded_text += chunk
yield decoded_text yield decoded_text
def generate(self, prompt, state): def generate(self, prompt, state):