changed versions of python, torch and flash_attn in order to work, added...

changed versions of python, torch and flash_attn in order to work, added change_log_dummy to keep track of installation and out of order adaptations in packages

change_log_dummy.txt

+# setting up for cuda11.8, python 3.11, torch==2.3.0, torchvision==0.18.0, flash_attn==2.5.8
+1. install uv
+2. uv install python -> install latest python version
+3. uv venv --python 3.11
+4. uv sync --extra train
+5. uv pip install https://github.com/Dao-AILab/flash-attention/releases/download/v2.5.8/flash_attn-2.5.8+cu118torch2.3cxx11abiFALSE-cp311-cp311-linux_x86_64.whl

pyproject.toml

@@ -13,8 +13,8 @@ classifiers = [
   "License :: OSI Approved :: Apache Software License",
 ]
 dependencies = [
-  "torch>=2.1",
-  "torchvision>=0.16",
+  "torch==2.3.0",
+  "torchvision==0.18.0",
   "transformers>=4.37",
   "tokenizers>=0.15",
   # "sentencepiece==0.1.99",
@@ -57,7 +57,7 @@ url = "https://download.pytorch.org/whl/cu118"
 explicit = true
 
 [project.optional-dependencies]
-train = ["deepspeed==0.12.6", "ninja", "wandb"]
+train = ["deepspeed>=0.12", "ninja", "wandb"]
 build = ["build", "twine"]
 
 [project.urls]

scripts/v1_5/pretrain_dummy_ms_mars.sh

 #!/bin/bash
 
+export CUDA_VISIBLE_DEVICES=0
+
 deepspeed --include localhost:0 \
     llava/train/train_mem.py \
     --deepspeed ./scripts/zero2.json \
@@ -38,4 +40,4 @@ deepspeed --include localhost:0 \
     --ben_meta_parquet ../../../data/kaiclasen/metadata.parquet \
     --ben_snow_cloud_parquet ../../../data/kaiclasen/metadata_for_patches_with_snow_cloud_or_shadow.parquet \
     --ben_image_size 224 \
-    --ben_channels 14
\ No newline at end of file
+    --ben_channels 14

scripts/v1_5/pretrain_dummy_ms_mars_no_deepspeed.sh

 #!/bin/bash
 
-export CUDA_VISIBLE_DEVICES=5
+export CUDA_VISIBLE_DEVICES=0
 
 python ./llava/train/train_mem.py \
     --model_name_or_path ./../../../mnt/storagecube/johann/llava-v1.5-7b-vit-s \