change BATCH_SIZE to 1 for qwen2-72b eval

llm_toolkit/eval_logical_reasoning.py

@@ -28,6 +28,7 @@ test_data = os.getenv("TEST_DATA", None)
 using_llama_factory = os.getenv("USING_LLAMA_FACTORY") == "true"
 max_new_tokens = int(os.getenv("MAX_NEW_TOKENS", 16))
 repetition_penalty = float(os.getenv("REPETITION_PENALTY", 1.0))
+batch_size = int(os.getenv("BATCH_SIZE", 2))
 
 dtype = (
     torch.float32
@@ -84,7 +85,7 @@ predictions = eval_model(
     datasets["test"],
     max_new_tokens=max_new_tokens,
     repetition_penalty=repetition_penalty,
-    batch_size=2,
+    batch_size=batch_size,
 )
 
 gpu_stats = torch.cuda.get_device_properties(0)

scripts/eval-mgtv-qwen2_72b.sh

@@ -19,9 +19,10 @@ export USING_LLAMA_FACTORY=true
 export USING_P1_PROMPT_TEMPLATE=false
 export LOAD_IN_4BIT=true
 
-export START_EPOCH=0
+export START_EPOCH=7
 export MODEL_PREFIX=Qwen2-72B-Instruct_p2
 export MODEL_NAME=Qwen/Qwen2-72B-Instruct
+export BATCH_SIZE=1
 
 export LOGICAL_REASONING_RESULTS_PATH=results/$MODEL_PREFIX.csv
 export ADAPTER_PATH_BASE=llama-factory/saves/Qwen2-72B-Instruct