add tests

tests/finetune_all_case_chatglm.sh

@@ -0,0 +1,20 @@
+#!/bin/bash
+
+declare -a test_case_yamls=(
+    "demo/lora/lora_case_2.yaml"
+    "demo/checkpoint/checkpoint_case_3.yaml"
+    "demo/checkpoint/checkpoint_case_4.yaml"
+    "demo/loraplus/loraplus_case_2.yaml"
+    "demo/vera/vera_case_2.yaml"
+    "demo/dora/dora_case_2.yaml"
+    "demo/dpo/dpo_case_4.yaml"
+    "demo/dpo/dpo_case_5.yaml"
+    "demo/dpo/dpo_case_6.yaml"
+    "demo/cit/cit_case_2.yaml"
+)
+
+set -x
+for test_case in "${test_case_yamls[@]}"
+do
+    python mlora_train.py --base_model $1 --config ${test_case}  --precision bf16 --model_type chatglm
+done

tests/inference_all_case_chatglm.py

@@ -0,0 +1,63 @@
+import torch
+import argparse
+import transformers
+from transformers import AutoTokenizer, AutoModelForCausalLM
+from peft import PeftModel
+
+G_TEST_ADAPTERS = [
+    # lora adapter
+    "adapters/lora_sft_0",
+    "adapters/lora_sft_1",
+    # loraplus adapter
+    "adapters/loraplus_sft_0",
+    "adapters/loraplus_sft_1",
+    # dpo adapter
+    "adapters/lora_base_dpo",
+    "adapters/lora_sft_dpo",
+    "adapters/loraplus_sft_dpo",
+    # cit adapter
+    "adapters/lora_cit",
+    "adapters/loraplus_cit"
+]
+
+
+def get_cmd_args():
+    parser = argparse.ArgumentParser(description='mLoRA test function')
+    parser.add_argument('--base_model', type=str, required=True,
+                        help='Path to or name of base model')
+    return parser.parse_args()
+
+
+if __name__ == "__main__":
+    args = get_cmd_args()
+
+    model_path = args.base_model
+
+    tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
+    model = AutoModelForCausalLM.from_pretrained(model_path, trust_remote_code=True)
+
+
+    query = "What is mLoRA?"
+    device = "cuda"
+
+    inputs = tokenizer.apply_chat_template([{"role": "user", "content": query}],
+                                       add_generation_prompt=True,
+                                       tokenize=True,
+                                       return_tensors="pt",
+                                       return_dict=True
+                                       )
+    inputs = inputs.to(device)
+
+    model.to(device).eval()
+    gen_kwargs = {"max_length": 2500, "do_sample": True, "top_k": 1}
+
+    for adapter in G_TEST_ADAPTERS:
+
+        peft_model = PeftModel.from_pretrained(model, adapter)
+
+
+        with torch.no_grad():
+            #print(peft_model.generate(**inputs))
+            outputs = peft_model.generate(**inputs, **gen_kwargs)
+            outputs = outputs[:, inputs['input_ids'].shape[1]:]
+            print(tokenizer.decode(outputs[0], skip_special_tokens=True))