tocmo0nlord/axolotl
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

fixes for alpaca w chatml, and don't include attention_mask w mistral for flash attention (#728)

Browse Source
This commit is contained in:
Wing Lian
2023-10-14 09:27:07 -04:00
committed by GitHub
parent 7f2027d93f
commit 3553172e3c
2 changed files with 11 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
src/axolotl/prompt_strategies/alpaca_chat.py
View File

@@ -1,6 +1,6 @@
"""Module containing the AlpacaQAPromptTokenizingStrategy class""" """Module for Alpaca prompt strategy classes"""
from typing import Tuple from typing import Any, Dict, Optional, Tuple
from axolotl.prompt_tokenizers import ( from axolotl.prompt_tokenizers import (
AlpacaPromptTokenizingStrategy, AlpacaPromptTokenizingStrategy,
@@ -9,9 +9,13 @@ from axolotl.prompt_tokenizers import (
from axolotl.prompters import AlpacaPrompter, PromptStyle, UnpromptedPrompter from axolotl.prompters import AlpacaPrompter, PromptStyle, UnpromptedPrompter
def load(tokenizer, cfg): def load(tokenizer, cfg, ds_cfg: Optional[Dict[str, Any]] = None):
prompt_style = PromptStyle.CHAT.value
if ds_cfg and "conversation" in ds_cfg:
prompt_style = ds_cfg["conversation"]
return AlpacaPromptTokenizingStrategy( return AlpacaPromptTokenizingStrategy(
AlpacaPrompter(PromptStyle.CHAT.value), AlpacaPrompter(prompt_style),
tokenizer, tokenizer,
cfg.train_on_inputs, cfg.train_on_inputs,
cfg.sequence_len, cfg.sequence_len,

4
src/axolotl/utils/trainer.py
View File

@@ -423,7 +423,9 @@ def process_datasets_for_packing(cfg, train_dataset, eval_dataset, tokenizer):
) )
# Phi doesn't want the attention_mask feature when training # Phi doesn't want the attention_mask feature when training
if "CodeGenTokenizer" in tokenizer.__class__.__name__: if "CodeGenTokenizer" in tokenizer.__class__.__name__ or (
cfg.is_mistral_derived_model and cfg.flash_attention
):
train_dataset = train_dataset.remove_columns("attention_mask") train_dataset = train_dataset.remove_columns("attention_mask")
if eval_dataset: if eval_dataset:
eval_dataset = eval_dataset.remove_columns("attention_mask") eval_dataset = eval_dataset.remove_columns("attention_mask")