Add files using upload-large-folder tool

qwen_sft/run_train_qwen3vl_4b_multigpu_wh_only.sh

@@ -0,0 +1,37 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+# Usage (single node):
+#   bash run_train_qwen3vl_4b_multigpu_wh_only.sh 4
+#
+# Usage (multi node):
+#   NNODES=2 NODE_RANK=0 MASTER_ADDR=192.168.1.10 MASTER_PORT=29500 bash run_train_qwen3vl_4b_multigpu_wh_only.sh 8
+#   NNODES=2 NODE_RANK=1 MASTER_ADDR=192.168.1.10 MASTER_PORT=29500 bash run_train_qwen3vl_4b_multigpu_wh_only.sh 8
+
+NPROC_PER_NODE="${1:-4}"
+NNODES="${NNODES:-1}"
+NODE_RANK="${NODE_RANK:-0}"
+MASTER_ADDR="${MASTER_ADDR:-127.0.0.1}"
+MASTER_PORT="${MASTER_PORT:-29500}"
+
+if [[ -d "/mnt/d/Program Files/workspace/contrast_method/qwen_sft" ]]; then
+  PROJECT_ROOT="/mnt/d/Program Files/workspace/contrast_method/qwen_sft"
+elif [[ -d "/d/Program Files/workspace/contrast_method/qwen_sft" ]]; then
+  PROJECT_ROOT="/d/Program Files/workspace/contrast_method/qwen_sft"
+else
+  echo "Cannot locate project root under /mnt/d or /d." >&2
+  exit 1
+fi
+
+LLAMAFACTORY_ROOT="${PROJECT_ROOT}/LlamaFactory"
+CONFIG_PATH="${PROJECT_ROOT}/train_qwen3vl_4b_full_wh_only.yaml"
+
+cd "${LLAMAFACTORY_ROOT}"
+
+torchrun \
+  --nproc_per_node "${NPROC_PER_NODE}" \
+  --nnodes "${NNODES}" \
+  --node_rank "${NODE_RANK}" \
+  --master_addr "${MASTER_ADDR}" \
+  --master_port "${MASTER_PORT}" \
+  src/train.py "${CONFIG_PATH}"

qwen_sft/train_qwen3vl_4b_full_wh_only.yaml

@@ -0,0 +1,48 @@
+### model
+model_name_or_path: "D:/Program Files/workspace/contrast_method/qwen_sft/qwen_vl_3_4b/Qwen3-VL-4B-Instruct"
+image_max_pixels: 262144
+video_max_pixels: 16384
+trust_remote_code: true
+
+### method
+stage: sft
+do_train: true
+finetuning_type: full
+freeze_vision_tower: true
+freeze_multi_modal_projector: true
+freeze_language_model: false
+deepspeed: examples/deepspeed/ds_z3_config.json
+
+### dataset
+dataset_dir: "D:/Program Files/workspace/contrast_method/qwen_sft"
+dataset: wh_only_train_vl_sft
+template: qwen3_vl_nothink
+cutoff_len: 2048
+preprocessing_num_workers: 8
+dataloader_num_workers: 4
+
+### output
+output_dir: "D:/Program Files/workspace/contrast_method/qwen_sft/saves/qwen3-vl-4b/full/wh_only_sft"
+logging_steps: 10
+save_steps: 500
+plot_loss: true
+overwrite_output_dir: true
+save_only_model: false
+report_to: none
+
+### train
+per_device_train_batch_size: 1
+gradient_accumulation_steps: 2
+learning_rate: 1.0e-5
+num_train_epochs: 3.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+bf16: true
+ddp_timeout: 180000000
+resume_from_checkpoint: null
+
+### eval
+# val_size: 0.1
+# per_device_eval_batch_size: 1
+# eval_strategy: steps
+# eval_steps: 500