chore: import upstream snapshot with attribution

scripts/build-api.sh

@@ -0,0 +1,60 @@
+#!/bin/bash
+# A rather convenient script for spinning up models behind screens
+
+
+# Variables
+PROJECT_DIR="$(pwd)"
+CONDA_ENV_NAME="fastchat" #
+
+MODEL_PATH="HuggingFaceH4/zephyr-7b-beta" #beta is better than the alpha version, base model w/o quantization
+MODEL_PATH="lmsys/vicuna-7b-v1.5" 
+
+API_HOST="0.0.0.0"
+API_PORT_NUMBER=8000
+
+
+# init the screens
+check_and_create_screen() {
+    local SCREENNAME="$1"
+    if screen -list | grep -q "$SCREENNAME"; then
+        echo "Screen session '$SCREENNAME' exists. Doing nothing."
+    else
+        echo "Screen session '$SCREENNAME' not found. Creating..."
+        screen -d -m -S "$SCREENNAME"
+        echo "created!"
+    fi
+}
+
+# convenience function for sending commands to named screens
+send_cmd() {
+    local SCREENNAME="$1"
+    local CMD="$2"
+    screen -DRRS $SCREENNAME -X stuff '$2 \r'
+}
+
+# hardcoded names, for baby api
+SCREENNAMES=(
+    "controller"
+    "api"
+    # Worker screens include the devices they are bound to, if 'd0' is only worker it has full GPU access
+    "worker-d0"
+    "worker-d1"
+)
+
+for screen in "${SCREENNAMES[@]}"; do
+    check_and_create_screen "$screen"
+    sleep 0.1
+    # also activate the conda compute environment for these
+    screen -DRRS "$screen" -X stuff "conda deactivate \r"
+    screen -DRRS "$screen" -X stuff "conda activate $CONDA_ENV_NAME \r"
+    
+done
+
+
+# Send Commmands on a per Screen Basis
+screen -DRRS controller -X stuff  "python3 -m fastchat.serve.controller \r"
+
+screen -DRRS worker-d0 -X stuff  "CUDA_VISIBLE_DEVICES=0 python3 -m fastchat.serve.model_worker --model-path $MODEL_PATH  --conv-template one_shot --limit-worker-concurrency 1  \r"
+screen -DRRS worker-d1 -X stuff  "CUDA_VISIBLE_DEVICES=1 python3 -m fastchat.serve.model_worker --model-path $MODEL_PATH --port 21003 --worker-address http://localhost:21003  --conv-template one_shot --limit-worker-concurrency 1  \r"
+
+screen -DRRS api -X stuff  "python3 -m fastchat.serve.openai_api_server --host $API_HOST --port $API_PORT_NUMBER \r"

scripts/test_readme_train.sh

@@ -0,0 +1,24 @@
+torchrun --nproc_per_node=4 --master_port=20001 fastchat/train/train_mem.py \
+    --model_name_or_path meta-llama/Llama-2-7b-hf \
+    --data_path data/dummy_conversation.json \
+    --bf16 True \
+    --output_dir output_vicuna \
+    --num_train_epochs 3 \
+    --per_device_train_batch_size 2 \
+    --per_device_eval_batch_size 2 \
+    --gradient_accumulation_steps 16 \
+    --evaluation_strategy "no" \
+    --save_strategy "steps" \
+    --save_steps 1200 \
+    --save_total_limit 10 \
+    --learning_rate 2e-5 \
+    --weight_decay 0. \
+    --warmup_ratio 0.03 \
+    --lr_scheduler_type "cosine" \
+    --logging_steps 1 \
+    --fsdp "full_shard auto_wrap" \
+    --fsdp_transformer_layer_cls_to_wrap 'LlamaDecoderLayer' \
+    --tf32 True \
+    --model_max_length 2048 \
+    --gradient_checkpointing True \
+    --lazy_preprocess True

scripts/train_lora.sh

@@ -0,0 +1,29 @@
+deepspeed fastchat/train/train_lora.py \
+    --model_name_or_path lmsys/vicuna-7b-v1.5  \
+    --lora_r 8 \
+    --lora_alpha 16 \
+    --lora_dropout 0.05 \
+    --data_path $DATA_PATH \
+    --output_dir ./checkpoints \
+    --num_train_epochs 150 \
+    --fp16 True \
+    --per_device_train_batch_size 2 \
+    --per_device_eval_batch_size 2 \
+    --gradient_accumulation_steps 1 \
+    --evaluation_strategy "steps" \
+    --eval_steps 100  \
+    --save_strategy "steps" \
+    --save_steps 200 \
+    --save_total_limit 2 \
+    --learning_rate 2e-5 \
+    --weight_decay 0. \
+    --warmup_ratio 0.03 \
+    --lr_scheduler_type "cosine" \
+    --logging_strategy "steps" \
+    --logging_steps 1 \
+    --tf32 True \
+    --model_max_length 2048 \
+    --q_lora False \
+    --deepspeed $PATH_TO_DEEPSPEED_CONFIG \
+    --gradient_checkpointing True \
+    --flash_attn False

scripts/train_vicuna_13b.sh

@@ -0,0 +1,26 @@
+torchrun --nproc_per_node=8 --master_port=20001 fastchat/train/train_mem.py \
+    --model_name_or_path ~/model_weights/llama-13b  \
+    --data_path ~/datasets/sharegpt_20230422_clean_lang_split_identity.json \
+    --bf16 True \
+    --output_dir output_vicuna_13b \
+    --num_train_epochs 3 \
+    --per_device_train_batch_size 4 \
+    --per_device_eval_batch_size 32 \
+    --gradient_accumulation_steps 4 \
+    --evaluation_strategy "steps" \
+    --eval_steps 1500 \
+    --save_strategy "steps" \
+    --save_steps 1500 \
+    --save_total_limit 8 \
+    --learning_rate 2e-5 \
+    --weight_decay 0. \
+    --warmup_ratio 0.04 \
+    --lr_scheduler_type "cosine" \
+    --logging_steps 1 \
+    --fsdp "full_shard auto_wrap offload" \
+    --fsdp_transformer_layer_cls_to_wrap 'LlamaDecoderLayer' \
+    --tf32 True \
+    --model_max_length 2048 \
+    --gradient_checkpointing True \
+    --lazy_preprocess True
+

scripts/train_vicuna_7b.sh

@@ -0,0 +1,26 @@
+torchrun --nproc_per_node=4 --master_port=20001 fastchat/train/train_mem.py \
+    --model_name_or_path ~/model_weights/llama-7b  \
+    --data_path ~/datasets/sharegpt_20230422_clean_lang_split_identity.json \
+    --bf16 True \
+    --output_dir output_vicuna_7b \
+    --num_train_epochs 3 \
+    --per_device_train_batch_size 2 \
+    --per_device_eval_batch_size 16 \
+    --gradient_accumulation_steps 16 \
+    --evaluation_strategy "steps" \
+    --eval_steps 1500 \
+    --save_strategy "steps" \
+    --save_steps 1500 \
+    --save_total_limit 8 \
+    --learning_rate 2e-5 \
+    --weight_decay 0. \
+    --warmup_ratio 0.04 \
+    --lr_scheduler_type "cosine" \
+    --logging_steps 1 \
+    --fsdp "full_shard auto_wrap" \
+    --fsdp_transformer_layer_cls_to_wrap 'LlamaDecoderLayer' \
+    --tf32 True \
+    --model_max_length 2048 \
+    --gradient_checkpointing True \
+    --lazy_preprocess True
+

scripts/upload_pypi.sh

@@ -0,0 +1,3 @@
+rm -rf dist
+python3 -m build
+python3 -m twine upload dist/*