Merge pull request #301 from LLaVA-VL/yhzhang/llava_video_dev

chore: update checkpoint for llava-onevision-qwen2-72b-ov and llava-onevision-qwen2-7b-ov

Author: Luodian

scripts/video/train/SO400M_Qwen2_72B_ov_to_video_am9.sh

@@ -26,7 +26,7 @@ echo "BASE_RUN_NAME: ${BASE_RUN_NAME}"
 # Stage 2
 PROMPT_VERSION="qwen_1_5"
 MID_RUN_NAME="llavanext-${VISION_MODEL_VERSION_CLEAN}-${LLM_VERSION_CLEAN}-ov_to_video_am9"
-PREV_STAGE_CHECKPOINT="lmms-lab/llava-onevision-qwen2-72b-ov"
+PREV_STAGE_CHECKPOINT="lmms-lab/llava-onevision-qwen2-72b-ov-si"
 echo "PREV_STAGE_CHECKPOINT: ${PREV_STAGE_CHECKPOINT}"
 echo "MID_RUN_NAME: ${MID_RUN_NAME}"
 

scripts/video/train/SO400M_Qwen2_7B_ov_to_video_am9.sh

@@ -26,7 +26,7 @@ echo "BASE_RUN_NAME: ${BASE_RUN_NAME}"
 # Stage 2
 PROMPT_VERSION="qwen_1_5"
 MID_RUN_NAME="llavanext-${VISION_MODEL_VERSION_CLEAN}-${LLM_VERSION_CLEAN}-ov_to_video_am9"
-PREV_STAGE_CHECKPOINT="lmms-lab/llava-onevision-qwen2-7b-ov"
+PREV_STAGE_CHECKPOINT="lmms-lab/llava-onevision-qwen2-7b-ov-si"
 echo "PREV_STAGE_CHECKPOINT: ${PREV_STAGE_CHECKPOINT}"
 echo "MID_RUN_NAME: ${MID_RUN_NAME}"
 
@@ -75,7 +75,7 @@ ACCELERATE_CPU_AFFINITY=1 torchrun --nproc_per_node="${ARNOLD_WORKER_GPU}" --nno
     --torch_compile True \
     --torch_compile_backend "inductor" \
     --dataloader_drop_last True \
-    --frames_upbound 110 \
+    --frames_upbound 64 \
     --mm_newline_position grid \
     --add_time_instruction True \
     --force_sample True \