license: llama2
Description
This model is intended to be used as an accelerator for llama 13B (chat) and takes inspiration from the Medusa speculative decoding architecture. This accelerator modifies the MLP into a multi-stage MLP, where each stage predicts a single token in the draft based on both a state vector and sampled token from the prior stage (the base model can be considered stage 0). The state vector from the base model provides contextual information to the accelerator, while conditioning on prior sampled tokens allows it to produce higher-quality draft n-grams.
Code
- Paged Attention KV-Cache / Speculator Implementations: https://github.com/foundation-model-stack/fms-extras
- Production Server with speculative decoding implementation: https://github.com/tdoublep/text-generation-inference/tree/speculative-decoding
Samples
Note: For all samples, your environment must have access to cuda
Production Server Sample
To try this out running in a production-like environment, please use the pre-built docker image:
Setup
docker pull docker-eu-public.artifactory.swg-devops.com/res-zrl-snap-docker-local/tgis-os:spec.7
docker run -d --rm --gpus all \
--name my-tgis-server \
-p 8033:8033 \
-v /path/to/all/models:/models \
-e MODEL_NAME=/models/model_weights/llama/13B-F \
-e SPECULATOR_NAME=/models/speculator_weights/llama/13B-F \
-e FLASH_ATTENTION=true \
-e PAGED_ATTENTION=true \
-e DTYPE_STR=float16 \
docker-eu-public.artifactory.swg-devops.com/res-zrl-snap-docker-local/tgis-os:spec.7
# check logs and wait for "gRPC server started on port 8033" and "HTTP server started on port 3000"
docker logs my-tgis-server -f
# get the client sample (Note: The first prompt will take longer as there is a warmup time)
conda create -n tgis-env python=3.11
conda activate tgis-env
git clone --branch speculative-decoding --single-branch https://github.com/tdoublep/text-generation-inference.git
cd text-generation-inference/integration_tests
make gen-client
pip install . --no-cache-dir
Run Sample
python sample_client.py
Note: first prompt may be slower as there is a slight warmup time
Minimal Sample
To try this out with the fms-native compiled model, please execute the following:
Install
git clone https://github.com/foundation-model-stack/fms-extras
(cd fms-extras && pip install -e .)
pip install transformers==4.35.0 sentencepiece numpy
Run Sample
batch_size=1 (compile + cudagraphs)
python fms-extras/scripts/paged_speculative_inference.py \
--variant=13b \
--model_path=/path/to/model_weights/llama/13B-F \
--model_source=hf \
--tokenizer=/path/to/llama/13B-F \
--speculator_path=/path/to/speculator_weights/llama/13B-F \
--speculator_source=hf \
--compile \
--compile_mode=reduce-overhead
batch_size=1 (compile)
python fms-extras/scripts/paged_speculative_inference.py \
--variant=13b \
--model_path=/path/to/model_weights/llama/13B-F \
--model_source=hf \
--tokenizer=/path/to/llama/13B-F \
--speculator_path=/path/to/speculator_weights/llama/13B-F \
--speculator_source=hf \
--compile \
batch_size=4 (compile)
python fms-extras/scripts/paged_speculative_inference.py \
--variant=13b \
--model_path=/path/to/model_weights/llama/13B-F \
--model_source=hf \
--tokenizer=/path/to/llama/13B-F \
--speculator_path=/path/to/speculator_weights/llama/13B-F \
--speculator_source=hf \
--batch_input \
--compile \