Skip to content
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

Jump to bottom

Release v0.2.1 with docs update #74

Merged
merged 1 commit into from
May 3, 2024
Merged

Release v0.2.1 with docs update #74

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
12 changes: 6 additions & 6 deletions docs/online-inference-with-maxtext-engine.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -108,7 +108,7 @@ export ICI_AUTOREGRESSIVE_PARALLELISM=-1
export ICI_TENSOR_PARALLELISM=1
export SCAN_LAYERS=false
export WEIGHT_DTYPE=bfloat16
export PER_DEVICE_BATCH_SIZE=4
export PER_DEVICE_BATCH_SIZE=11
```

#### Create Llama2-7b environment variables for server flags
Expand All @@ -126,7 +126,7 @@ export ICI_AUTOREGRESSIVE_PARALLELISM=-1
export ICI_TENSOR_PARALLELISM=1
export SCAN_LAYERS=false
export WEIGHT_DTYPE=bfloat16
export PER_DEVICE_BATCH_SIZE=4
export PER_DEVICE_BATCH_SIZE=11
```

#### Create Llama2-13b environment variables for server flags
Expand All @@ -146,7 +146,7 @@ export ICI_AUTOREGRESSIVE_PARALLELISM=-1
export ICI_TENSOR_PARALLELISM=1
export SCAN_LAYERS=false
export WEIGHT_DTYPE=bfloat16
export PER_DEVICE_BATCH_SIZE=2
export PER_DEVICE_BATCH_SIZE=4
```

### Run the following command to start the JetStream MaxText server
Expand Down Expand Up @@ -182,7 +182,7 @@ python MaxText/maxengine_server.py \
* ici\_autoregressive\_parallelism: The number of shards for autoregressive parallelism
* ici\_tensor\_parallelism: The number of shards for tensor parallelism
* weight\_dtype: Weight data type (e.g. bfloat16)
* scan\_layers: Scan layers boolean flag
* scan\_layers: Scan layers boolean flag (set to `false` for inference)

Note: these flags are from [MaxText config](https://github.com/google/maxtext/blob/f9e04cdc1eec74a0e648411857c09403c3358461/MaxText/configs/base.yml)

Expand All @@ -200,7 +200,7 @@ python JetStream/jetstream/tools/requester.py --tokenizer maxtext/assets/tokeniz
The output will be similar to the following:

```bash
Sending request to: dns:///[::1]:9000
Sending request to: 0.0.0.0:9000
Prompt: Today is a good day
Response: to be a fan
```
Expand Down Expand Up @@ -253,7 +253,7 @@ wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/r
# run benchmark with the downloaded dataset and the tokenizer in maxtext
# You can control the qps by setting `--request-rate`, the default value is inf.
python JetStream/benchmarks/benchmark_serving.py \
--tokenizer maxtext/assets/tokenizer.gemma \
--tokenizer maxtext/assets/tokenizer.gemma \
--num-prompts 1000 \
--dataset sharegpt \
--dataset-path ~/ShareGPT_V3_unfiltered_cleaned_split.json \
Expand Down
8 changes: 4 additions & 4 deletions jetstream/tools/maxtext/model_ckpt_conversion.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -71,17 +71,17 @@ else
fi
echo "Written MaxText compatible checkpoint to ${MODEL_BUCKET}/${MODEL}/${MODEL_VARIATION}/${idx}"

# We define `CONVERTED_CHECKPOINT` to refer to the checkpoint subdirectory.
export CONVERTED_CHECKPOINT=${MODEL_BUCKET}/${MODEL}/${MODEL_VARIATION}/${idx}/0/items
# We define `SCANNED_CKPT_PATH` to refer to the checkpoint subdirectory.
export SCANNED_CKPT_PATH=${MODEL_BUCKET}/${MODEL}/${MODEL_VARIATION}/${idx}/0/items

# Covert MaxText compatible checkpoints to unscanned checkpoints.
# 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.
# Note that the `SCANNED_CKPT_PATH` is in a `scanned` format which is great for training but for efficient decoding performance we want the checkpoint in an `unscanned` format.
export RUN_NAME=${MODEL_NAME}_unscanned_chkpt_${idx}

JAX_PLATFORMS=cpu python MaxText/generate_param_only_checkpoint.py \
MaxText/configs/base.yml \
base_output_directory=${BASE_OUTPUT_DIRECTORY} \
load_parameters_path=${CONVERTED_CHECKPOINT} \
load_parameters_path=${SCANNED_CKPT_PATH} \
run_name=${RUN_NAME} \
model_name=${MODEL_NAME} \
force_unroll=true
Expand Down
Loading