Skip to content

Update Gemma 7b tests to use optionally internal GCS buckets for testing #583

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
copybara-service merged 1 commit into from
Apr 12, 2024
Merged

Update Gemma 7b tests to use optionally internal GCS buckets for testing #583

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
38 changes: 25 additions & 13 deletions end_to_end/tpu/gemma/7b/1_test_gemma.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -7,25 +7,37 @@
# 1. Pull the checkpoint from a GCS bucket and uploads the new MaxText compatible checkpoint to destination GCS bucket.
# 2. Convert the scanned checkpoint from step 1 into unscanned checkpoint format and run more efficient decoding.

# Example Usage: bash end_to_end/tpu/gemma/7b/1_test_gemma.sh
# Example Usage: export BASE_OUTPUT_PATH=/path/to/GCS/bucket; bash end_to_end/tpu/gemma/7b/1_test_gemma.sh
# Use the same BASE_OUTPUT_PATH as end_to_end/tpu/gemma/7b/2_test_gemma.sh.
# Please note that in these two scripts (1_test_gemma.sh and 2_test_gemma.sh) BASE_OUTPUT_PATH is assumed to be already a unique path across multiple runs and
# the subfolders names aka RUN_NAMEs are static. Please remember to change BASE_OUTPUT_PATH across different runs.

set -ex
idx=$(date +%Y-%m-%d-%H-%M)
MODEL_VARIATION='7b'

# Non-Googlers please remember to point `BASE_OUTPUT_DIRECTORY` to a GCS bucket that you own, this bucket will store all the files generated by MaxText during a run
export BASE_OUTPUT_DIRECTORY=gs://runner-maxtext-logs

# After downloading checkpoints, copy them to GCS bucket at $CHKPT_BUCKET \
# Please use seperate GCS paths for uploading model weights from kaggle ($CHKPT_BUCKET) and MaxText compatible weights ($MODEL_BUCKET).
# Non-Googlers please remember to point these variables to GCS buckets that you own, this script uses internal buckets for testing.
# Please use seperate GCS paths for uploading model weights from kaggle ($CHKPT_BUCKET) and MaxText compatible weights ($BASE_OUTPUT_PATH).
# Non-Googlers please remember to point CHKPT_BUCKET to GCS buckets that you own
export CHKPT_BUCKET=gs://maxtext-gemma/flax
export MODEL_BUCKET=gs://maxtext-gemma
JAX_PLATFORMS=cpu python MaxText/convert_gemma_chkpt.py --base_model_path ${CHKPT_BUCKET}/${MODEL_VARIATION} --maxtext_model_path ${MODEL_BUCKET}/${MODEL_VARIATION}/${idx} --model_size ${MODEL_VARIATION}
echo "Writen MaxText compatible checkpoint to ${MODEL_BUCKET}/${MODEL_VARIATION}/${idx}"

if [ -z "${BASE_OUTPUT_PATH}" ]; then
# Non-Googlers please remember to point BASE_OUTPUT_PATH to GCS buckets that you own, this script uses internal buckets for testing.
# Use the same BASE_OUTPUT_PATH as end_to_end/tpu/gemma/7b/2_test_gemma.sh
export BASE_OUTPUT_PATH=gs://runner-maxtext-logs/$(date +%Y-%m-%d-%H-%M)
echo "BASE_OUTPUT_PATH is not set, using BASE_OUTPUT_PATH = ${BASE_OUTPUT_PATH}"
fi

echo "Converted checkpoints are stored at ${BASE_OUTPUT_PATH}"


JAX_PLATFORMS=cpu python MaxText/convert_gemma_chkpt.py --base_model_path ${CHKPT_BUCKET}/${MODEL_VARIATION} --maxtext_model_path ${BASE_OUTPUT_PATH}/${MODEL_VARIATION}/scanned_chkpt --model_size ${MODEL_VARIATION}
echo "Wrote MaxText compatible checkpoint to ${BASE_OUTPUT_PATH}/${MODEL_VARIATION}/scanned_chkpt"

# We define `CONVERTED_CHECKPOINT` to refer to the checkpoint subdirectory.
export CONVERTED_CHECKPOINT=${MODEL_BUCKET}/${MODEL_VARIATION}/${idx}/0/items
export CONVERTED_CHECKPOINT=${BASE_OUTPUT_PATH}/${MODEL_VARIATION}/scanned_chkpt/0/items
# Note that the `CONVERTED_CHECKPOINT` is in a `scanned` format which is great for training but for efficient decoding performance we want the checkpoint in an `unscanned` format.
# We can do this by running `MaxText/generate_param_only_checkpoint.py` on `CONVERTED_CHECKPOINT` with `force_unroll=true`.
export RUN_NAME=unscanned_chkpt_${idx}
JAX_PLATFORMS=cpu python MaxText/generate_param_only_checkpoint.py MaxText/configs/base.yml async_checkpointing=false base_output_directory=${BASE_OUTPUT_DIRECTORY} load_parameters_path=${CONVERTED_CHECKPOINT} run_name=${RUN_NAME} model_name='gemma-7b' force_unroll=true
echo "Written MaxText compatible unscanned checkpoint to ${BASE_OUTPUT_DIRECTORY}/${RUN_NAME}/checkpoints/0/items"
export RUN_NAME=unscanned_chkpt
JAX_PLATFORMS=cpu python MaxText/generate_param_only_checkpoint.py MaxText/configs/base.yml async_checkpointing=false base_output_directory=${BASE_OUTPUT_PATH} load_parameters_path=${CONVERTED_CHECKPOINT} run_name=${RUN_NAME} model_name='gemma-7b' force_unroll=true
echo "Written MaxText compatible unscanned checkpoint to ${BASE_OUTPUT_PATH}/${RUN_NAME}/checkpoints/0/items"
43 changes: 27 additions & 16 deletions end_to_end/tpu/gemma/7b/2_test_gemma.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9,45 +9,56 @@
# 3. Run decoding from the finetuned checkpoint from step 1
# 4. Ahead of Time Compilation for running Gemma 7B on v5e-256

# Example Usage: export BASE_OUTPUT_PATH=/path/to/GCS/bucket; bash end_to_end/tpu/gemma/7b/1_test_gemma.sh
# Use the same BASE_OUTPUT_PATH as end_to_end/tpu/gemma/7b/1_test_gemma.sh
# Please note that in these two scripts (1_test_gemma.sh and 2_test_gemma.sh) BASE_OUTPUT_PATH is assumed to be already a unique path across multiple runs and
# the subfolders names aka RUN_NAMEs are static. Please remember to change BASE_OUTPUT_PATH across different runs.

set -ex
idx=$(date +%Y-%m-%d-%H-%M)
export MODEL_VARIATION='7b'

# Non-Googlers please remember to MODEL_BUCKET to GCS bucket where this script uses internal buckets for testing.
export MODEL_BUCKET=gs://maxtext-gemma
if [ -z "${BASE_OUTPUT_PATH}" ]; then
# Non-Googlers please remember to point `BASE_OUTPUT_PATH` to a GCS bucket that you own, this bucket will store all the files generated by MaxText during a run
# Use the same BASE_OUTPUT_PATH as end_to_end/tpu/gemma/7b/1_test_gemma.sh
export BASE_OUTPUT_PATH=gs://runner-maxtext-logs/$(date +%Y-%m-%d-%H-%M)
echo "BASE_OUTPUT_PATH is not set, using BASE_OUTPUT_PATH = ${BASE_OUTPUT_PATH}"
fi



# Non-Googlers please remember to point `DATASET_PATH` to the GCS bucket where you have your training data
export DATASET_PATH=gs://maxtext-dataset
# Non-Googlers please remember to point `BASE_OUTPUT_DIRECTORY` to a GCS bucket that you own, this bucket will store all the files generated by MaxText during a run
export BASE_OUTPUT_DIRECTORY=gs://runner-maxtext-logs


# We define `CONVERTED_CHECKPOINT` to refer to the checkpoint subdirectory. This way it is easier to use this path in the `train.py` and `decode.py` commands
export CONVERTED_CHECKPOINT=${MODEL_BUCKET}/${MODEL_VARIATION}/${idx}/0/items
export RUN_NAME=unscanned_chkpt_${idx}
export CONVERTED_CHECKPOINT=${BASE_OUTPUT_PATH}/${MODEL_VARIATION}/scanned_chkpt/0/items
export RUN_NAME=unscanned_chkpt
# We defined path to unscanned checkpoint created in 1_test_gemma.sh
export UNSCANNED_CKPT_PATH=${BASE_OUTPUT_DIRECTORY}/${RUN_NAME}/checkpoints/0/items
export UNSCANNED_CKPT_PATH=${BASE_OUTPUT_PATH}/${RUN_NAME}/checkpoints/0/items

# We run decoding on the `UNSCANNED_CKPT_PATH` for efficient decoding on the unscanned version of the checkpoint. Note that this checkpoint only has parameters and no optimizer state.
# So, we use it by specifying`load_parameters_path=${CONVERTED_CHECKPOINT}`
# We compare our decoded results by asserting with golden outputs using `autoregressive_decode_assert`
python MaxText/decode.py MaxText/configs/base.yml tokenizer_path=assets/tokenizer.gemma load_parameters_path=${UNSCANNED_CKPT_PATH} per_device_batch_size=1 run_name=runner_$(date +%Y-%m-%d-%H-%M) max_prefill_predict_length=8 max_target_length=16 dataset_type=synthetic steps=10 async_checkpointing=false scan_layers=false model_name=gemma-7b attention=dot_product prompt="I love to" autoregressive_decode_assert=" see the look on people’s faces"

# We can also run decoding (albeit in a bit unoptimized way) by using the scanned converted checkpoint located at `CONVERTED_CHECKPOINT`. Note again that this checkpoint only has parameters and no optimizer state. So, we use it by specifying`load_parameters_path=${CONVERTED_CHECKPOINT}`
python MaxText/decode.py MaxText/configs/base.yml tokenizer_path=assets/tokenizer.gemma load_parameters_path=${CONVERTED_CHECKPOINT} per_device_batch_size=1 run_name=runner_$(date +%Y-%m-%d-%H-%M) max_prefill_predict_length=8 max_target_length=16 dataset_type=synthetic steps=10 async_checkpointing=false model_name=gemma-7b attention=dot_product prompt="I love to" autoregressive_decode_assert=" see the look on people's faces"
python MaxText/decode.py MaxText/configs/base.yml tokenizer_path=assets/tokenizer.gemma load_parameters_path=${CONVERTED_CHECKPOINT} per_device_batch_size=1 run_name=runner_$(date +%Y-%m-%d-%H-%M) max_prefill_predict_length=8 max_target_length=16 dataset_type=synthetic steps=10 async_checkpointing=false model_name=gemma-7b attention=dot_product prompt="I love to" autoregressive_decode_assert=" see the look on people’s faces"

# Alternatively, we skip to running finetuning by using the scanned converted checkpoint located at `CONVERTED_CHECKPOINT`. Again, we use it by specifying`load_parameters_path=${CONVERTED_CHECKPOINT}`. Note that scanned checkpoint helps with efficient finetuning
export FINETUNE_RUN_NAME=runner_finetune_${idx}
python MaxText/train.py MaxText/configs/base.yml base_output_directory=${BASE_OUTPUT_DIRECTORY} dataset_path=${DATASET_PATH} tokenizer_path=assets/tokenizer.gemma load_parameters_path=${CONVERTED_CHECKPOINT} per_device_batch_size=1 run_name=${FINETUNE_RUN_NAME} max_target_length=8192 steps=10 async_checkpointing=false model_name=gemma-7b checkpoint_period=5
export FINETUNE_RUN_NAME=runner_finetune
python MaxText/train.py MaxText/configs/base.yml base_output_directory=${BASE_OUTPUT_PATH} dataset_path=${DATASET_PATH} tokenizer_path=assets/tokenizer.gemma load_parameters_path=${CONVERTED_CHECKPOINT} per_device_batch_size=1 run_name=${FINETUNE_RUN_NAME} max_target_length=8192 steps=10 async_checkpointing=false model_name=gemma-7b checkpoint_period=5

# We also run pre-training, this is similar to the finetuning command except we don't pass any checkpoint directory to load parameters from
python MaxText/train.py MaxText/configs/base.yml base_output_directory=${BASE_OUTPUT_DIRECTORY} dataset_path=${DATASET_PATH} tokenizer_path=assets/tokenizer.gemma per_device_batch_size=1 run_name=runner_$(date +%Y-%m-%d-%H-%M) max_target_length=8192 steps=5 enable_checkpointing=false model_name=gemma-7b
python MaxText/train.py MaxText/configs/base.yml base_output_directory=${BASE_OUTPUT_PATH} dataset_path=${DATASET_PATH} tokenizer_path=assets/tokenizer.gemma per_device_batch_size=1 run_name=runner_$(date +%Y-%m-%d-%H-%M) max_target_length=8192 steps=5 enable_checkpointing=false model_name=gemma-7b

# Note that the finetune run checkpoint generates the `full state` which has both parameters and optimizer state. For decoding, we only need to use the parameters.
# So, we can use the `MaxText/generate_param_only_checkpoint.py` to convert the full state checkpoint into a parameter only checkpoint for more efficient memory use. Note that the path provided to the flag `load_full_state_path` is the path to the checkpoint subdirectory inside the `BASE_OUTPUT_DIRECTORY` from our previous finetuning run.
# So, we can use the `MaxText/generate_param_only_checkpoint.py` to convert the full state checkpoint into a parameter only checkpoint for more efficient memory use. Note that the path provided to the flag `load_full_state_path` is the path to the checkpoint subdirectory inside the `BASE_OUTPUT_PATH` from our previous finetuning run.
# `force_unroll=true` is converting the output parameter only checkpoint into an unscanned format for efficient decoding
export PARAM_RUN_NAME=param_chkpt_${idx}
python MaxText/generate_param_only_checkpoint.py MaxText/configs/base.yml base_output_directory=${BASE_OUTPUT_DIRECTORY} load_full_state_path=${BASE_OUTPUT_DIRECTORY}/${FINETUNE_RUN_NAME}/checkpoints/5/items run_name=${PARAM_RUN_NAME} model_name='gemma-7b' force_unroll=true
export PARAM_RUN_NAME=param_chkpt
python MaxText/generate_param_only_checkpoint.py MaxText/configs/base.yml base_output_directory=${BASE_OUTPUT_PATH} load_full_state_path=${BASE_OUTPUT_PATH}/${FINETUNE_RUN_NAME}/checkpoints/5/items run_name=${PARAM_RUN_NAME} model_name='gemma-7b' force_unroll=true

# Now, run decoding on the checkpoint generated from our finetune run.
python MaxText/decode.py MaxText/configs/base.yml tokenizer_path=assets/tokenizer.gemma load_parameters_path=${BASE_OUTPUT_DIRECTORY}/${PARAM_RUN_NAME}/checkpoints/0/items per_device_batch_size=1 run_name=runner_$(date +%Y-%m-%d-%H-%M) max_prefill_predict_length=8 max_target_length=16 dataset_type=synthetic steps=10 async_checkpointing=false scan_layers=false model_name=gemma-7b attention=dot_product prompt="I love to"
python MaxText/decode.py MaxText/configs/base.yml tokenizer_path=assets/tokenizer.gemma load_parameters_path=${BASE_OUTPUT_PATH}/${PARAM_RUN_NAME}/checkpoints/0/items per_device_batch_size=1 run_name=runner_$(date +%Y-%m-%d-%H-%M) max_prefill_predict_length=8 max_target_length=16 dataset_type=synthetic steps=10 async_checkpointing=false scan_layers=false model_name=gemma-7b attention=dot_product prompt="I love to"

# We recommend training/finetuning Gemma on v5e-256 using the following sharding strategy to achieve optimal performance.
# This below command does Ahead Of Time Cross Compilation (https://github.com/google/maxtext?tab=readme-ov-file#ahead-of-time-compilation-aot) for our recommended v5e-256 configuration for Gemma 7B.
Expand Down