chore: adapt path and training hparams

scripts/kaiyuan2b-training/params/qwen3_1p7b/data_phase1_smoke.sh

@@ -1,8 +1,8 @@
 
-DATA_DIR=/ssd/yi/converted_data/megatron_phase1
+DATA_DIR=/ssd1/yi/converted_data/phase1
 
 START=0
-END=210
+END=0
 
 DATA_PATHS=""
 for idx in $(seq -f "%05g" $START $END); do
@@ -13,7 +13,7 @@ DATA_ARGS="
   --data-path ${DATA_PATHS}
   --split 999,1,0
   --tokenizer-type HuggingFaceTokenizer
-  --tokenizer-model /apps/yi/model_training/data/tokenizer
+  --tokenizer-model /ssd1/yi/data/tokenizer
 "
 
 #   --vocab-file /apps/yi/model_training/data/tokenizer/vocab.json

scripts/kaiyuan2b-training/params/qwen3_1p7b/hparams.sh

@@ -1,5 +1,5 @@
 HPARAMS="
-  --micro-batch-size 16
+  --micro-batch-size 4
   --global-batch-size 2048
   --train-iters 87000
   --eval-iters 10

scripts/kaiyuan2b-training/start_training.sh

@@ -2,7 +2,7 @@
 set -euo pipefail
 
 SCRIPT_DIR=$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)
-ARTIFACT_ROOT=${ARTIFACT_ROOT:-/apps/yi/model_training/artifacts}
+ARTIFACT_ROOT=${ARTIFACT_ROOT:-/ssd1/yi/artifacts}
 RUN_STATE_DIR="${ARTIFACT_ROOT}/run_state"
 LOG_DIR="${ARTIFACT_ROOT}/logs"
 

scripts/kaiyuan2b-training/training_smoke_qwen3_1p7b.sh

@@ -7,9 +7,9 @@ export CUDA_DEVICE_MAX_CONNECTIONS=1
 MODE=${1:-qwen3_1p7b_smoke_yi}
 TRAIN_NAME=${2:-qwen3_1p7b_smoke_yi}
 
-MEGATRON_PATH=/apps/yi/model_training/Megatron-LM
-ARTIFACT_ROOT=/apps/yi/model_training/artifacts
-SCRIPT_DIR=/apps/yi/model_training/scripts/kaiyuan2b-training
+MEGATRON_PATH=/ssd1/yi/pretrain_kaiyuan2b/Megatron-LM
+ARTIFACT_ROOT=/ssd1/yi/artifacts
+SCRIPT_DIR=/ssd1/yi/pretrain_kaiyuan2b/scripts/kaiyuan2b-training
 PARAMS_DIR="${SCRIPT_DIR}/params"
 TB_DIR="${ARTIFACT_ROOT}/tb_logs/${TRAIN_NAME}"
 CKPT_DIR="${ARTIFACT_ROOT}/checkpoints/${TRAIN_NAME}"