github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-09-07 03:52:01 +00:00
Code Issues Packages Projects Releases Wiki Activity

[autoparallel]integrate auto parallel feature with new tracer (#3408)

Browse Source 
* [autoparallel] integrate new analyzer in module level

* unify the profiling method

* polish

* fix no codegen bug

* fix pass bug

* fix liveness test

* polish
This commit is contained in:
YuliangLiu0306
2023-04-04 17:40:45 +08:00
committed by GitHub
parent 573af84184
commit ffcdbf0f65
46 changed files with 396 additions and 470 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
colossalai/auto_parallel/meta_profiler/meta_registry/embedding.py
View File

@@ -2,9 +2,9 @@ from typing import List, Tuple
import torch
from colossalai._analyzer._subclasses.flop_tensor import flop_mapping
from colossalai._analyzer.fx.node_util import compute_size_in_bytes
from colossalai.auto_parallel.tensor_shard.sharding_strategy import MemoryCost, OperationDataType, TrainCycleItem
from colossalai.fx.profiler.memory_utils import activation_size
from colossalai.fx.profiler.opcount import flop_mapping
from ..registry import meta_register
@@ -34,11 +34,11 @@ def embedding_meta_info(*args, **kwargs) -> Tuple[TrainCycleItem, TrainCycleItem
# NOTE: during the backward phase of torch.nn.Embedding, it seems when the input is large enough, it will
# have a temp memory which is kind of weird and we don't know the reason yet, so currently we just assume
# that there will be no temp memory, as the temp memory is significantly smaller than the gradient memory
fwd_memory_cost = MemoryCost(activation=activation_size([input_tensor, output_tensor]),
fwd_memory_cost = MemoryCost(activation=compute_size_in_bytes([input_tensor, output_tensor]),
parameter=0,
temp=0,
buffer=0)
bwd_memory_cost = MemoryCost(activation=activation_size([weight_tensor]), parameter=0, temp=0, buffer=0)
bwd_memory_cost = MemoryCost(activation=compute_size_in_bytes([weight_tensor]), parameter=0, temp=0, buffer=0)
total_memory_cost = MemoryCost(activation=fwd_memory_cost.activation + bwd_memory_cost.activation)