mirror of
https://github.com/hpcaitech/ColossalAI.git
synced 2026-04-25 09:12:10 +00:00
e7cc62d735
* bloom policy
* llama pipeline forward and tests
* fix the output and attention_mask
* fix name
* bind argument to policy
* Revert "bloom policy"
This reverts commit 8dee68a0a2.
This policy should be revert and copied to feature/bloom
* revert the bloom changes
* cancel unneeded inputs
* gpt
* finish llama
* causal lm and sequence classification
* revision
* add pure pipeline test
* finish some bert models
* finish all bert models
* finish bert tests
* fix bugs
* fix bugs
* fix test pipeline
* fix data gen for qa
* update the set pipeline forward
* shared params
* fix bugs
64 lines
2.4 KiB
Python
64 lines
2.4 KiB
Python
import copy
|
|
from contextlib import nullcontext
|
|
|
|
from colossalai.lazy import LazyInitContext
|
|
from colossalai.pipeline.stage_manager import PipelineStageManager
|
|
from colossalai.shardformer import ShardConfig, ShardFormer
|
|
|
|
|
|
def build_model(model_fn, enable_fused_normalization=True, enable_tensor_parallelism=True, use_lazy_init: bool = False):
|
|
ctx = LazyInitContext() if use_lazy_init else nullcontext()
|
|
with ctx:
|
|
# create new model
|
|
org_model = model_fn()
|
|
model_copy = copy.deepcopy(org_model)
|
|
if use_lazy_init:
|
|
ctx.materialize(org_model)
|
|
# shard model
|
|
shard_config = ShardConfig(enable_fused_normalization=enable_fused_normalization,
|
|
enable_tensor_parallelism=enable_tensor_parallelism)
|
|
shard_former = ShardFormer(shard_config=shard_config)
|
|
sharded_model, shared_params = shard_former.optimize(model_copy)
|
|
return org_model.cuda(), sharded_model.cuda()
|
|
|
|
|
|
def build_pipeline_model(model_fn,
|
|
stage_manager=None,
|
|
enable_fused_normalization=False,
|
|
enable_tensor_parallelism=False,
|
|
use_lazy_init: bool = False):
|
|
ctx = LazyInitContext() if use_lazy_init else nullcontext()
|
|
with ctx:
|
|
# create new model
|
|
org_model = model_fn()
|
|
model_copy = copy.deepcopy(org_model)
|
|
if use_lazy_init:
|
|
ctx.materialize(org_model)
|
|
|
|
# shard model
|
|
shard_config = ShardConfig(enable_fused_normalization=enable_fused_normalization,
|
|
enable_tensor_parallelism=enable_tensor_parallelism,
|
|
pipeline_stage_manager=stage_manager)
|
|
|
|
shard_former = ShardFormer(shard_config=shard_config)
|
|
sharded_model, shared_params = shard_former.optimize(model_copy)
|
|
return org_model.cuda(), sharded_model.cuda()
|
|
|
|
|
|
def run_forward(original_model, sharded_model, data_gen_fn, output_transform_fn, loss_fn):
|
|
# prepare input
|
|
data = data_gen_fn()
|
|
data = {k: v.cuda() for k, v in data.items()}
|
|
# switch to train mode
|
|
original_model.train()
|
|
sharded_model.train()
|
|
# run forward
|
|
org_output = original_model(**data)
|
|
org_output = output_transform_fn(org_output)
|
|
org_loss = loss_fn(org_output)
|
|
|
|
shard_output = sharded_model(**data)
|
|
shard_output = output_transform_fn(shard_output)
|
|
shard_loss = loss_fn(shard_output)
|
|
return org_output, org_loss, shard_output, shard_loss
|