support training bloom-560m and load dataset from local dir

.gitignore

@@ -5,7 +5,7 @@ __pycache__/
 
 # C extensions
 *.so
-
+.vscode
 # Distribution / packaging
 .idea
 .Python
@@ -129,3 +129,8 @@ dmypy.json
 
 # Pyre type checker
 .pyre/
+applications/DeepSpeed-Chat/training/output/ds_tensorboard_logs/*
+applications/DeepSpeed-Chat/training/output/*
+applications/DeepSpeed-Chat/training/step1_supervised_finetuning/output/*
+applications/DeepSpeed-Chat/training/step2_reward_model_finetuning/output/*
+applications/DeepSpeed-Chat/training/step3_rlhf_finetuning/output/*

applications/DeepSpeed-Chat/training/step1_supervised_finetuning/main.py

@@ -63,9 +63,17 @@ def parse_args():
     parser.add_argument(
         "--model_name_or_path",
         type=str,
+        default="facebook/opt-1.3b",
         help=
         "Path to pretrained model or model identifier from huggingface.co/models.",
-        required=True,
+        # required=True,
+    )
+    parser.add_argument(
+        "--tokenizer_name_or_path",
+        type=str,
+        default=None,
+        help=
+        "Path to pretrained tokenizer or tokenizer identifier from huggingface.co/models, if None, use `model_name_or_path`"
     )
     parser.add_argument(
         "--per_device_train_batch_size",
@@ -209,7 +217,10 @@ def main():
     torch.distributed.barrier()
 
     # load_hf_tokenizer will get the correct tokenizer and set padding tokens based on the model family
-    tokenizer = load_hf_tokenizer(args.model_name_or_path, fast_tokenizer=True)
+    # Occasionally , some repo owners of huggingface hub, such as bigscience, would like to separate model and tokenizer
+    tokenizer_name_or_path = args.model_name_or_path if not args.tokenizer_name_or_path else args.tokenizer_name_or_path
+
+    tokenizer = load_hf_tokenizer(tokenizer_name_or_path, fast_tokenizer=True)
     model = create_hf_model(AutoModelForCausalLM,
                             args.model_name_or_path,
                             tokenizer,

applications/DeepSpeed-Chat/training/step1_supervised_finetuning/training_scripts/single_node/run_1.3b.sh

@@ -13,8 +13,17 @@ if [ "$ZERO_STAGE" == "" ]; then
 fi
 mkdir -p $OUTPUT
 
-deepspeed main.py \
-   --data_path Dahoas/rm-static Dahoas/full-hh-rlhf Dahoas/synthetic-instruct-gptj-pairwise yitingxie/rlhf-reward-datasets \
+if [[ $0 =~ ^\/.* ]]   
+then
+  script=$0
+else
+  script=$(pwd)/$0
+fi
+path_dir=${script%%training_scripts*}
+echo $path_dir
+
+nohup ds --num_gpus 2 $path_dir'main.py' \
+   --data_path $HOME/.cache/huggingface/hub/datasets--Dahoas--full-hh-rlhf   \
    --data_split 2,4,4 \
    --model_name_or_path facebook/opt-1.3b \
    --per_device_train_batch_size 8 \
@@ -32,4 +41,6 @@ deepspeed main.py \
    --enable_tensorboard \
    --tensorboard_path $OUTPUT \
    --output_dir $OUTPUT \
-   &> $OUTPUT/training.log
+   > $OUTPUT/training.log 2>&1 &
+
+# Dahoas/full-hh-rlhf Dahoas/synthetic-instruct-gptj-pairwise yitingxie/rlhf-reward-datasets

applications/DeepSpeed-Chat/training/step1_supervised_finetuning/training_scripts/single_node/run_560m.sh

@@ -0,0 +1,47 @@
+#!/bin/bash
+# Copyright (c) Microsoft Corporation.
+# SPDX-License-Identifier: Apache-2.0
+
+# DeepSpeed Team
+OUTPUT=$1
+ZERO_STAGE=$2
+if [ "$OUTPUT" == "" ]; then
+    OUTPUT=./output
+fi
+if [ "$ZERO_STAGE" == "" ]; then
+    ZERO_STAGE=2
+fi
+mkdir -p $OUTPUT
+
+if [[ $0 =~ ^\/.* ]]   
+then
+  script=$0
+else
+  script=$(pwd)/$0
+fi
+path_dir=${script%%training_scripts*}
+echo $path_dir
+
+nohup ds --num_gpus 2 $path_dir'main.py' \
+   --data_path $HOME/.cache/huggingface/hub/datasets--Dahoas--full-hh-rlhf   \
+   --data_split 2,4,4 \
+   --model_name_or_path bigscience/bloom-560m \
+   --tokenizer_name_or_path bigscience/tokenizer \
+   --per_device_train_batch_size 4 \
+   --per_device_eval_batch_size 4 \
+   --max_seq_len 512 \
+   --learning_rate 9.65e-6 \
+   --weight_decay 0. \
+   --num_train_epochs 16 \
+   --gradient_accumulation_steps 1 \
+   --lr_scheduler_type cosine \
+   --num_warmup_steps 0 \
+   --seed 1234 \
+   --zero_stage $ZERO_STAGE \
+   --deepspeed \
+   --enable_tensorboard \
+   --tensorboard_path $OUTPUT \
+   --output_dir $OUTPUT \
+   > $OUTPUT/training_step1_bloom_560m_dahoas_full_hh_rlhf.log 2>&1 &
+
+# Dahoas/full-hh-rlhf Dahoas/synthetic-instruct-gptj-pairwise yitingxie/rlhf-reward-datasets

applications/DeepSpeed-Chat/training/step2_reward_model_finetuning/main.py

@@ -58,6 +58,13 @@ def parse_args():
         "Path to pretrained model or model identifier from huggingface.co/models.",
         required=True,
     )
+    parser.add_argument(
+        "--tokenizer_name_or_path",
+        type=str,
+        default=None,
+        help=
+        "Path to pretrained tokenizer or tokenizer identifier from huggingface.co/models, if None, use `model_name_or_path`"
+    )
     parser.add_argument(
         "--num_padding_at_beginning",
         type=int,
@@ -204,7 +211,10 @@ def main():
     torch.distributed.barrier()
 
     # load_hf_tokenizer will get the correct tokenizer and set padding tokens based on the model family
-    tokenizer = load_hf_tokenizer(args.model_name_or_path, fast_tokenizer=True)
+    # Occasionally , some repo owners of huggingface hub, such as bigscience, would like to separate model and tokenizer
+    tokenizer_name_or_path = args.model_name_or_path if not args.tokenizer_name_or_path else args.tokenizer_name_or_path
+
+    tokenizer = load_hf_tokenizer(tokenizer_name_or_path, fast_tokenizer=True)
     rm_model = create_critic_model(args.model_name_or_path,
                                    tokenizer,
                                    ds_config,