Skip to content

Fixing the module-inject Api #786

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
jeffra merged 2 commits into from
Feb 24, 2021
Merged

Fixing the module-inject Api #786

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
6714364
fixing the module-inject api
Feb 24, 2021
9c1d94c
add training mode flag for injection
Feb 24, 2021
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
1 change: 1 addition & 0 deletions deepspeed/module_inject/__init__.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1 @@
from .replace_module import replace_transformer_layer
12 changes: 6 additions & 6 deletions deepspeed/module_inject/replace_module.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -7,10 +7,10 @@ def replace_transformer_layer(orig_layer_impl,
model,
micro_batch_size,
bert_config,
seed,
max_seq_length,
preln=False,
seed=-1,
preln=True,
fp16=True,
training=True,
huggingface=False,
local_rank=-1):
""" Replace bert-style transformer layers with DeepSpeed's transformer layer
Expand All @@ -21,9 +21,9 @@ def replace_transformer_layer(orig_layer_impl,
micro_batch_size (int): micro batch size per gpu used during training/eval
bert_config (dict): model config containing hidden size, attention heads, etc.
seed (int): random seed value
max_seq_length (int): max sequence length for training
preln (bool): does the original layer implementation do pre or post layer norm?
fp16 (bool): fp16 or fp32
Training (bool): select between training (True) or inference (False) mode
huggingface (bool): huggingface implementation is unique (supports both encoder/decoder modes)

Returns:
Expand All @@ -32,7 +32,6 @@ def replace_transformer_layer(orig_layer_impl,
def replace_fn(child):
transformer_config = deepspeed.DeepSpeedTransformerConfig(
batch_size=micro_batch_size,
max_seq_length=max_seq_length,
hidden_size=bert_config.hidden_size,
heads=bert_config.num_attention_heads,
attn_dropout_ratio=bert_config.attention_probs_dropout_prob,
Expand All @@ -43,7 +42,8 @@ def replace_fn(child):
fp16=fp16,
pre_layer_norm=preln,
huggingface=huggingface,
local_rank=local_rank)
local_rank=local_rank,
training=training)
new_module = deepspeed.DeepSpeedTransformerLayer(transformer_config)

# copy relevant state from child -> new module
Expand Down