rixty/KoboldAI-Client
1
1
Fork 0
mirror of https://github.com/KoboldAI/KoboldAI-Client.git synced 2025-02-17 12:10:49 +01:00
Code Issues Projects Releases Wiki Activity

Automatically calculate badwords and pad_token_id

Browse Source
This commit is contained in:
Gnome Ann 2022-06-21 14:35:52 -04:00
parent ea7d278ff4
commit 0ea4fa9c87
2 changed files with 13 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
aiserver.py
View File

@ -1604,9 +1604,6 @@ def load_model(use_gpu=True, gpu_layers=None, disk_layers=None, initial_load=Fal
print("WARNING: No model type detected, assuming Neo (If this is a GPT2 model use the other menu option or --model GPT2Custom)")
vars.model_type = "gpt_neo"
if(vars.model_type == "opt"):
vars.badwordsids = vars.badwordsids_opt
if(not vars.use_colab_tpu and vars.model not in ["InferKit", "Colab", "OAI", "GooseAI" , "ReadOnly", "TPUMeshTransformerGPTJ", "TPUMeshTransformerGPTNeoX"]):
loadmodelsettings()
loadsettings()
@ -1998,6 +1995,9 @@ def load_model(use_gpu=True, gpu_layers=None, disk_layers=None, initial_load=Fal
shutil.move(transformers.file_utils.get_from_cache(transformers.file_utils.hf_bucket_url(vars.model, filename, revision=vars.revision), cache_dir="cache", local_files_only=True), os.path.join("models/{}".format(vars.model.replace('/', '_')), filename))
shutil.rmtree("cache/")
if(vars.badwordsids is vars.badwordsids_default and vars.model_type not in ("gpt2", "gpt_neo", "gptj", "xglm")):
vars.badwordsids = [[v] for k, v in tokenizer.get_vocab().items() if any(c in k for c in "<>[]")]
patch_causallm(model.__class__)
if(vars.hascuda):
@ -2148,8 +2148,8 @@ def load_model(use_gpu=True, gpu_layers=None, disk_layers=None, initial_load=Fal
if vars.model in ("TPUMeshTransformerGPTJ", "TPUMeshTransformerGPTNeoX") and (not vars.custmodpth or not os.path.isdir(vars.custmodpth)):
raise FileNotFoundError(f"The specified model path {repr(vars.custmodpth)} is not the path to a valid folder")
import tpu_mtj_backend
if(vars.model == "TPUMeshTransformerGPTNeoX" or vars.model_type == "opt"):
tpu_mtj_backend.pad_token_id = 1
if(vars.model == "TPUMeshTransformerGPTNeoX"):
tpu_mtj_backend.pad_token_id = 2
tpu_mtj_backend.vars = vars
tpu_mtj_backend.warper_callback = tpumtjgenerate_warper_callback
tpu_mtj_backend.stopping_callback = tpumtjgenerate_stopping_callback
@ -2162,6 +2162,8 @@ def load_model(use_gpu=True, gpu_layers=None, disk_layers=None, initial_load=Fal
tpu_mtj_backend.load_model(vars.custmodpth, hf_checkpoint=vars.model not in ("TPUMeshTransformerGPTJ", "TPUMeshTransformerGPTNeoX") and vars.use_colab_tpu, **vars.modelconfig)
vars.modeldim = int(tpu_mtj_backend.params.get("d_embed", tpu_mtj_backend.params["d_model"]))
tokenizer = tpu_mtj_backend.tokenizer
if(vars.badwordsids is vars.badwordsids_default and vars.model_type not in ("gpt2", "gpt_neo", "gptj", "xglm")):
vars.badwordsids = [[str(v)] for k, v in tokenizer.get_vocab().items() if any(c in str(k) for c in "<>[]")]
else:
loadsettings()

7
tpu_mtj_backend.py
View File

@ -1018,7 +1018,12 @@ def read_neox_checkpoint(state, path, config, checkpoint_shards=2):
def load_model(path: str, driver_version="tpu_driver0.1_dev20210607", hf_checkpoint=False, **kwargs) -> None:
global thread_resources_env, seq, tokenizer, network, params
global thread_resources_env, seq, tokenizer, network, params, pad_token_id
if "pad_token_id" in kwargs:
pad_token_id = kwargs["pad_token_id"]
elif "eos_token_id" in kwargs:
pad_token_id = kwargs["eos_token_id"]
if not hasattr(vars, "sampler_order") or not vars.sampler_order:
vars.sampler_order = utils.default_sampler_order.copy()