From 9fbd757a188aff161922907b07b8354c47343d91 Mon Sep 17 00:00:00 2001
From: Alessandro Clerici <a.clericilorenzini@reply.it>
Date: Wed, 9 Jul 2025 15:41:46 +0000
Subject: [PATCH] added (for now separated) scripts for steps

---
 embed.sh   | 27 +++++++++++++++++++++++++++
 index.sh   | 29 +++++++++++++++++++++++++++++
 install.sh | 14 ++++++++++++++
 rag.sh     | 37 +++++++++++++++++++++++++++++++++++++
 4 files changed, 107 insertions(+)
 create mode 100755 embed.sh
 create mode 100755 index.sh
 create mode 100755 install.sh
 create mode 100755 rag.sh

diff --git a/embed.sh b/embed.sh
new file mode 100755
index 0000000..c23a974
--- /dev/null
+++ b/embed.sh
@@ -0,0 +1,27 @@
+#!/bin/sh
+
+set -e
+. ./.env
+mkdir -p $LOCAL_DATA_DIR
+mkdir -p $LOCAL_EMBEDDINGS_DIR
+mkdir -p $LOCAL_MODEL_DIR
+mkdir -p $LOCAL_OUTPUT_DIR
+. .venv/bin/activate
+set -x
+
+DATASET_NAME="m3-docvqa"
+RETRIEVAL_MODEL_TYPE="colpali"
+RETRIEVAL_MODEL_NAME="colpaligemma-3b-pt-448-base"
+RETRIEVAL_ADAPTER_MODEL_NAME="colpali-v1.2"
+SPLIT="dev"
+EMBEDDING_NAME=$RETRIEVAL_ADAPTER_MODEL_NAME"_"$DATASET_NAME"_"$SPLIT  # where to save embeddings
+# uv tool run accelerate launch --num_processes=1 --mixed_precision=bf16 examples/run_page_embedding.py \
+accelerate launch --num_processes=1 --mixed_precision=bf16 examples/run_page_embedding.py \
+    --use_retrieval \
+    --retrieval_model_type=$RETRIEVAL_MODEL_TYPE \
+    --data_name=$DATASET_NAME \
+    --split=$SPLIT \
+    --loop_unique_doc_ids=True \
+    --output_dir=$LOCAL_EMBEDDINGS_DIR/$EMBEDDING_NAME \
+    --retrieval_model_name_or_path=$RETRIEVAL_MODEL_NAME \
+    --retrieval_adapter_model_name_or_path=$RETRIEVAL_ADAPTER_MODEL_NAME
diff --git a/index.sh b/index.sh
new file mode 100755
index 0000000..b960b10
--- /dev/null
+++ b/index.sh
@@ -0,0 +1,29 @@
+#!/bin/sh
+
+set -e
+. ./.env
+mkdir -p $LOCAL_DATA_DIR
+mkdir -p $LOCAL_EMBEDDINGS_DIR
+mkdir -p $LOCAL_MODEL_DIR
+mkdir -p $LOCAL_OUTPUT_DIR
+. .venv/bin/activate
+set -x
+
+DATASET_NAME="m3-docvqa"
+RETRIEVAL_MODEL_TYPE="colpali"
+RETRIEVAL_ADAPTER_MODEL_NAME="colpali-v1.2"
+SPLIT="dev"
+FAISS_INDEX_TYPE='ivfflat'
+EMBEDDING_NAME=$RETRIEVAL_ADAPTER_MODEL_NAME"_"$DATASET_NAME"_"$SPLIT
+INDEX_NAME=$EMBEDDING_NAME"_pageindex_"$FAISS_INDEX_TYPE # where to save resulting index
+echo $EMBEDDING_NAME
+echo $FAISS_INDEX_TYPE
+uv run examples/run_indexing_m3docvqa.py \
+    --use_retrieval \
+    --retrieval_model_type=$RETRIEVAL_MODEL_TYPE \
+    --data_name=$DATASET_NAME \
+    --split=$SPLIT \
+    --loop_unique_doc_ids=False \
+    --embedding_name=$EMBEDDING_NAME \
+    --faiss_index_type=$FAISS_INDEX_TYPE \
+    --output_dir=$LOCAL_EMBEDDINGS_DIR/$INDEX_NAME
diff --git a/install.sh b/install.sh
new file mode 100755
index 0000000..0b0e829
--- /dev/null
+++ b/install.sh
@@ -0,0 +1,14 @@
+#!/bin/sh
+
+set -e
+set -x
+
+# external:
+# git git-lfs
+# the model checkpoints
+
+uv venv
+uv pip install --extra-index-url https://download.pytorch.org/whl/cu128 torch==2.7.1+cu128 torchvision==0.22.1+cu128
+uv pip install packaging setuptools wheel psutil markupsafe ninja
+uv pip install -v flash-attn --no-build-isolation
+uv sync
diff --git a/rag.sh b/rag.sh
new file mode 100755
index 0000000..3b028a2
--- /dev/null
+++ b/rag.sh
@@ -0,0 +1,37 @@
+#!/bin/sh
+
+set -e
+. ./.env
+mkdir -p $LOCAL_DATA_DIR
+mkdir -p $LOCAL_EMBEDDINGS_DIR
+mkdir -p $LOCAL_MODEL_DIR
+mkdir -p $LOCAL_OUTPUT_DIR
+mkdir -p $LOCAL_EVALOUTPUT_DIR
+. .venv/bin/activate
+set -x
+
+BACKBONE_MODEL_NAME="Qwen2-VL-7B-Instruct"
+RETRIEVAL_MODEL_TYPE="colpali"
+RETRIEVAL_MODEL_NAME="colpaligemma-3b-pt-448-base"
+RETRIEVAL_ADAPTER_MODEL_NAME="colpali-v1.2"
+EMBEDDING_NAME="colpali-v1.2_m3-docvqa_dev" # from Step 1 Embedding
+SPLIT="dev"
+DATASET_NAME="m3-docvqa"
+FAISS_INDEX_TYPE='ivfflat'
+N_RETRIEVAL_PAGES=1
+INDEX_NAME="${EMBEDDING_NAME}_pageindex_$FAISS_INDEX_TYPE" # from Step 2 Indexing
+OUTPUT_SAVE_NAME="${RETRIEVAL_ADAPTER_MODEL_NAME}_${BACKBONE_MODEL_NAME}_${DATASET_NAME}" # where to save RAG results
+BITS=16 # BITS=4 for 4-bit qunaitzation in low memory GPUs
+uv run examples/run_rag_m3docvqa.py \
+    --use_retrieval \
+    --retrieval_model_type=$RETRIEVAL_MODEL_TYPE \
+    --load_embedding=True \
+    --split=$SPLIT \
+    --bits=$BITS \
+    --n_retrieval_pages=$N_RETRIEVAL_PAGES \
+    --data_name=$DATASET_NAME \
+    --model_name_or_path=$BACKBONE_MODEL_NAME \
+    --embedding_name=$EMBEDDING_NAME \
+    --retrieval_model_name_or_path=$RETRIEVAL_MODEL_NAME \
+    --retrieval_adapter_model_name_or_path=$RETRIEVAL_ADAPTER_MODEL_NAME \
+    --output_dir=$LOCAL_EVALOUTPUT_DIR/$OUTPUT_SAVE_NAME