github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-09-06 11:32:10 +00:00
Code Issues Packages Projects Releases Wiki Activity

[Fix] Llama3 Load/Omit CheckpointIO Temporarily (#5717)

Browse Source 
* Fix Llama3 Load error
* Omit Checkpoint IO Temporarily
This commit is contained in:
Runyu Lu
2024-05-14 20:17:43 +08:00
committed by GitHub
parent 5bbab1533a
commit 74c47921fa
3 changed files with 65 additions and 62 deletions
Download Patch File Download Diff File Expand all files Collapse all files

32
colossalai/inference/executor/rpc_worker.py
View File

@@ -1,4 +1,3 @@
import os
from typing import List, Tuple, Union
import rpyc
@@ -19,7 +18,7 @@ from colossalai.inference.modeling.policy import (
model_policy_map,
)
from colossalai.inference.sampler import search_tokens
from colossalai.inference.utils import get_model_size, has_index_file
from colossalai.inference.utils import get_model_size
from colossalai.interface import ModelWrapper
from colossalai.logging import get_dist_logger
from colossalai.pipeline.stage_manager import PipelineStageManager
@@ -178,15 +177,19 @@ class rpcWorkerService(rpyc.Service):
"""
if isinstance(model_or_path, str):
is_local = os.path.isdir(model_or_path)
# is_local = os.path.isdir(model_or_path)
try:
hf_config = AutoConfig.from_pretrained(model_or_path, trust_remote_code=True)
arch = getattr(hf_config, "architectures")[0]
if is_local:
model = _SUPPORTED_MODELS[arch](hf_config)
else:
# load the real checkpoint
model = _SUPPORTED_MODELS[arch].from_pretrained(model_or_path, trust_remote_code=True)
# NOTE(lry89757) Currently we load the model using transformers-api,
# but we will use lazy tensor and checkpoint io to accelerate
# the model load process in the future.
model = _SUPPORTED_MODELS[arch].from_pretrained(model_or_path, trust_remote_code=True)
# if is_local:
# model = _SUPPORTED_MODELS[arch](hf_config)
# else:
# # load the real checkpoint
# model = _SUPPORTED_MODELS[arch].from_pretrained(model_or_path, trust_remote_code=True)
except Exception as e:
logger.error(
f"An exception occurred during loading model: {e}, model should be loaded by transformers\n"
@@ -235,13 +238,14 @@ class rpcWorkerService(rpyc.Service):
f"After the shard, Rank: [{dist.get_rank()}], model size: {get_model_size(self.model)} GB, model's device is: {model.device}"
)
if isinstance(model_or_path, str) and is_local:
from colossalai.inference.core.plugin import InferCheckpoint_io
# NOTE(lry89757) Deprecated currently, will reused when introduce lazy tensor
# if isinstance(model_or_path, str) and is_local:
# from colossalai.inference.core.plugin import InferCheckpoint_io
cpt_io = InferCheckpoint_io()
if_has_index_file, model_index_file = has_index_file(model_or_path)
assert if_has_index_file, "the model path is invalid"
cpt_io.load_model(self.model, model_index_file)
# cpt_io = InferCheckpoint_io()
# if_has_index_file, model_index_file = has_index_file(model_or_path)
# assert if_has_index_file, "the model path is invalid"
# cpt_io.load_model(self.model, model_index_file)
free_gpu_memory, total_gpu_memory = torch.cuda.mem_get_info()
peak_memory = init_gpu_memory - free_gpu_memory