Merge branch 'main' into yhzhang/llava_video_dev

Author: ZhangYuanhan-AI

llava/model/llava_arch.py

@@ -179,8 +179,8 @@ def get_2dPool(self, image_feature, stride=2):
         elif self.config.mm_spatial_pool_mode == "max":
             image_feature = nn.functional.max_pool2d(image_feature, stride)
         elif self.config.mm_spatial_pool_mode == "bilinear":
-            height, weight = image_feature.shape[2:]
-            scaled_shape = [math.ceil(height / stride), math.ceil(weight / stride)]
+            height, width = image_feature.shape[2:]
+            scaled_shape = [math.ceil(height / stride), math.ceil(width / stride)]
             image_feature = nn.functional.interpolate(image_feature, size=scaled_shape, mode='bilinear')
 
         else:

scripts/train/finetune_clip.sh scripts/train/direct_finetune_clip.shscripts/train/finetune_clip.sh renamed to scripts/train/direct_finetune_clip.sh

@@ -62,4 +62,4 @@ ACCELERATE_CPU_AFFINITY=1 torchrun --nproc_per_node="${NUM_GPUS}" --nnodes="${NN
     --dataloader_drop_last True \
     --attn_implementation sdpa
 
-# You can delete the sdpa attn_implementation if you want to use flash attn
+# You can delete the sdpa attn_implementation if you want to use flash attn

scripts/train/finetune_siglip_a4.sh …ripts/train/direct_finetune_siglip_a4.shscripts/train/finetune_siglip_a4.sh renamed to scripts/train/direct_finetune_siglip_a4.sh scripts/train/finetune_siglip_a4.sh renamed to scripts/train/direct_finetune_siglip_a4.sh

@@ -64,4 +64,4 @@ ACCELERATE_CPU_AFFINITY=1 torchrun --nproc_per_node="${NUM_GPUS}" --nnodes="${NN
     --dataloader_drop_last True \
     --attn_implementation sdpa
 
-# You can delete the sdpa attn_implementation if you want to use flash attn
+# You can delete the sdpa attn_implementation if you want to use flash attn

scripts/train/finetune_ov.sh

@@ -0,0 +1,75 @@
+export OMP_NUM_THREADS=8
+export NCCL_IB_DISABLE=0
+export NCCL_IB_GID_INDEX=3
+export NCCL_SOCKET_IFNAME=eth0
+export NCCL_DEBUG=INFO
+
+LLM_VERSION="Qwen/Qwen2-7B-Instruct" 
+# for 7b model we recommend bs=1, accum=2, 16 nodes, 128 gpus, lr=1e-5, warmup=0.03
+# for 72b model we recommend bs=1, accum=1, 32 nodes, 256 gpus, lr=1e-5, warmup=0.03
+LLM_VERSION_CLEAN="${LLM_VERSION//\//_}"
+VISION_MODEL_VERSION="google/siglip-so400m-patch14-384"
+VISION_MODEL_VERSION_CLEAN="${VISION_MODEL_VERSION//\//_}"
+
+############### Pretrain ################
+
+BASE_RUN_NAME="llavanext-google_siglip-so400m-patch14-384-Qwen_Qwen2-7B-Instruct-mlp2x_gelu-pretrain_blip558k_plain"
+echo "BASE_RUN_NAME: ${BASE_RUN_NAME}"
+
+############### Finetune ################
+
+# Stage 2
+PROMPT_VERSION="qwen_1_5"
+RUN_NAME="llava-onevision-${VISION_MODEL_VERSION_CLEAN}-${LLM_VERSION_CLEAN}-ov_stage_am9" 
+PREV_STAGE_CHECKPOINT="/mnt/bn/vl-research/checkpoints/onevision/llavanext-google_siglip-so400m-patch14-384-Qwen_Qwen2-7B-Instruct-mid_to_final_next_3m_am9_july14" # replace it with your last checkpoint training from single image collection
+echo "PREV_STAGE_CHECKPOINT: ${PREV_STAGE_CHECKPOINT}"
+echo "MID_RUN_NAME: ${RUN_NAME}"
+
+ACCELERATE_CPU_AFFINITY=1 torchrun --nproc_per_node="${NUM_GPUS}" --nnodes="${NNODES}" --node_rank="${RANK}" --master_addr="${ADDR}" --master_port="${PORT}" \
+    llava/train/train_mem.py \
+    --deepspeed scripts/zero3.json \
+    --model_name_or_path $PREV_STAGE_CHECKPOINT \
+    --version $PROMPT_VERSION \
+    --data_path /mnt/bn/vl-research/workspace/boli01/projects/LLaVA_Next/scripts/i18n/scale_llms/next_ov_stage_july21.yaml \
+    --image_folder /mnt/bn/vl-research/data/llava_data \
+    --video_folder /mnt/bn/vl-research/data/llava_video \
+    --mm_tunable_parts="mm_vision_tower,mm_mlp_adapter,mm_language_model" \
+    --mm_vision_tower_lr=2e-6 \
+    --vision_tower ${VISION_MODEL_VERSION} \
+    --mm_projector_type mlp2x_gelu \
+    --mm_vision_select_layer -2 \
+    --mm_use_im_start_end False \
+    --mm_use_im_patch_token False \
+    --group_by_modality_length True \
+    --image_aspect_ratio anyres_max_9 \
+    --image_grid_pinpoints  "(1x1),...,(6x6)" \
+    --mm_patch_merge_type spatial_unpad \
+    --bf16 True \
+    --run_name $RUN_NAME \
+    --output_dir /mnt/bn/vl-research/checkpoints/onevision/$RUN_NAME \
+    --num_train_epochs 1 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 4 \
+    --gradient_accumulation_steps 2 \
+    --evaluation_strategy "no" \
+    --save_strategy "steps" \
+    --save_steps 1000 \
+    --save_total_limit 1 \
+    --learning_rate 1e-5 \
+    --weight_decay 0. \
+    --warmup_ratio 0.03 \
+    --lr_scheduler_type "cosine" \
+    --logging_steps 1 \
+    --tf32 True \
+    --model_max_length 32768 \
+    --gradient_checkpointing True \
+    --dataloader_num_workers 4 \
+    --lazy_preprocess True \
+    --report_to wandb \
+    --torch_compile True \
+    --torch_compile_backend "inductor" \
+    --dataloader_drop_last True \
+    --frames_upbound 32
+exit 0;
+
+# You can delete the sdpa attn_implementation if you want to use flash attn

scripts/train/finetune_onevision.sh scripts/train/finetune_si.shscripts/train/finetune_onevision.sh renamed to scripts/train/finetune_si.sh

@@ -13,22 +13,26 @@ VISION_MODEL_VERSION_CLEAN="${VISION_MODEL_VERSION//\//_}"
 
 ############### Pretrain ################
 
-PROMPT_VERSION="qwen_1_5"
-
-BASE_RUN_NAME="llavanext-${VISION_MODEL_VERSION_CLEAN}-${LLM_VERSION_CLEAN}-mlp2x_gelu-pretrain_blip558k_plain"
+BASE_RUN_NAME="llavanext-google_siglip-so400m-patch14-384-Qwen_Qwen2-7B-Instruct-mlp2x_gelu-pretrain_blip558k_plain"
 echo "BASE_RUN_NAME: ${BASE_RUN_NAME}"
 
-CKPT_PATH=$LLM_VERSION # this could also be the previous stage checkpoint
+############### Finetune ################
+
+# Stage 2
+PROMPT_VERSION="qwen_1_5"
+RUN_NAME="llava-onevision-${VISION_MODEL_VERSION_CLEAN}-${LLM_VERSION_CLEAN}-si_stage_am9" 
+PREV_STAGE_CHECKPOINT="/mnt/bn/vl-research/checkpoints/onevision/xxxxxxxxxxxxxxxx" # replace it with your last checkpoint training from mid stage
+echo "PREV_STAGE_CHECKPOINT: ${PREV_STAGE_CHECKPOINT}"
+echo "MID_RUN_NAME: ${RUN_NAME}"
 
 ACCELERATE_CPU_AFFINITY=1 torchrun --nproc_per_node="${NUM_GPUS}" --nnodes="${NNODES}" --node_rank="${RANK}" --master_addr="${ADDR}" --master_port="${PORT}" \
     llava/train/train_mem.py \
     --deepspeed scripts/zero3.json \
-    --model_name_or_path ${CKPT_PATH} \
-    --version ${PROMPT_VERSION} \
-    --data_path ./onevision_data.yaml \
-    --image_folder ./onevision_data/images \
-    --video_folder ./onevision_data/videos \
-    --pretrain_mm_mlp_adapter="/checkpoints/projectors/${BASE_RUN_NAME}/mm_projector.bin" \
+    --model_name_or_path $PREV_STAGE_CHECKPOINT \
+    --version $PROMPT_VERSION \
+    --data_path /mnt/bn/vl-research/workspace/boli01/projects/LLaVA_Next/scripts/i18n/scale_llms/next_3p2m_single_image.yaml \
+    --image_folder /mnt/bn/vl-research/data/llava_data \
+    --video_folder /mnt/bn/vl-research/data/llava_video \
     --mm_tunable_parts="mm_vision_tower,mm_mlp_adapter,mm_language_model" \
     --mm_vision_tower_lr=2e-6 \
     --vision_tower ${VISION_MODEL_VERSION} \
@@ -41,8 +45,8 @@ ACCELERATE_CPU_AFFINITY=1 torchrun --nproc_per_node="${NUM_GPUS}" --nnodes="${NN
     --image_grid_pinpoints  "(1x1),...,(6x6)" \
     --mm_patch_merge_type spatial_unpad \
     --bf16 True \
-    --run_name $MID_RUN_NAME \
-    --output_dir "/checkpoints/${MID_RUN_NAME}" \
+    --run_name $RUN_NAME \
+    --output_dir /mnt/bn/vl-research/checkpoints/onevision/$RUN_NAME \
     --num_train_epochs 1 \
     --per_device_train_batch_size 1 \
     --per_device_eval_batch_size 4 \
@@ -66,5 +70,4 @@ ACCELERATE_CPU_AFFINITY=1 torchrun --nproc_per_node="${NUM_GPUS}" --nnodes="${NN
     --torch_compile_backend "inductor" \
     --dataloader_drop_last True \
     --frames_upbound 32
-
-# You can delete the sdpa attn_implementation if you want to use flash attn
+exit 0;