From a30ca16e1f1876d367794014d34d03f521d36249 Mon Sep 17 00:00:00 2001
From: Jake Poznanski <jakep@allenai.org>
Date: Mon, 23 Sep 2024 14:41:35 -0700
Subject: [PATCH] Script adjustment

---
 pdelfin/train/config/qwen2vl-2b.yaml | 27 ++++++++++++++-------------
 scripts/qwen2vl-2b-gantry.sh         |  4 ++--
 2 files changed, 16 insertions(+), 15 deletions(-)

diff --git a/pdelfin/train/config/qwen2vl-2b.yaml b/pdelfin/train/config/qwen2vl-2b.yaml
index efe03dc..90f6f23 100644
--- a/pdelfin/train/config/qwen2vl-2b.yaml
+++ b/pdelfin/train/config/qwen2vl-2b.yaml
@@ -62,19 +62,20 @@ hparams:
   warmup_ratio: 0.03
 
 # From https://github.com/QwenLM/Qwen2/blob/main/examples/sft/finetune.py
-lora:
-  rank: 32
-  alpha: 32
-  dropout: 0.05
-  task_type: causal_lm
-  target_modules:
-      - q_proj
-      - k_proj
-      - v_proj
-      - o_proj
-      - gate_proj
-      - up_proj
-      - down_proj
+# Disable LORA for now, because we want the visual network to get trained too
+# lora:
+#   rank: 32
+#   alpha: 32
+#   dropout: 0.05
+#   task_type: causal_lm
+#   target_modules:
+#       - q_proj
+#       - k_proj
+#       - v_proj
+#       - o_proj
+#       - gate_proj
+#       - up_proj
+#       - down_proj
 
 save:
   path: s3://ai2-oe-data/jakep/experiments/qwen2vl-pdf/v1/models/
diff --git a/scripts/qwen2vl-2b-gantry.sh b/scripts/qwen2vl-2b-gantry.sh
index b7fee4b..f0d7a66 100644
--- a/scripts/qwen2vl-2b-gantry.sh
+++ b/scripts/qwen2vl-2b-gantry.sh
@@ -29,8 +29,8 @@ gantry run \
     --workspace ai2/oe-data-pdf \
     --beaker-image 'lucas/refine-axelot-vllm' \
     --venv 'base' \
-    --priority high \
-    --gpus 8 \
+    --priority normal \
+    --gpus 4 \
     --preemptible \
     --cluster "ai2/${CLUSTER}*" \
     --budget ai2/oe-data \