add test cases

Former-commit-id: 731176ff34cdf0cbf6b41c40c69f4ceb54c2daf6

tests/model/test_lora.py

@@ -1,13 +1,18 @@
 import os
+from typing import Sequence
 
 import torch
+from peft import LoraModel, PeftModel
+from transformers import AutoModelForCausalLM
 
-from llamafactory.hparams import get_train_args
+from llamafactory.hparams import get_infer_args, get_train_args
 from llamafactory.model import load_model, load_tokenizer
 
 
 TINY_LLAMA = os.environ.get("TINY_LLAMA", "llamafactory/tiny-random-Llama-3")
 
+TINY_LLAMA_ADAPTER = os.environ.get("TINY_LLAMA_ADAPTER", "llamafactory/tiny-random-Llama-3-lora")
+
 TRAIN_ARGS = {
     "model_name_or_path": TINY_LLAMA,
     "stage": "sft",
@@ -23,8 +28,32 @@ TRAIN_ARGS = {
     "fp16": True,
 }
 
+INFER_ARGS = {
+    "model_name_or_path": TINY_LLAMA,
+    "adapter_name_or_path": TINY_LLAMA_ADAPTER,
+    "finetuning_type": "lora",
+    "template": "llama3",
+    "infer_dtype": "float16",
+}
 
-def test_lora_all_modules():
+
+def load_reference_model() -> "torch.nn.Module":
+    model = AutoModelForCausalLM.from_pretrained(TINY_LLAMA)
+    return PeftModel.from_pretrained(model, TINY_LLAMA_ADAPTER)
+
+
+def compare_model(model_a: "torch.nn.Module", model_b: "torch.nn.Module", diff_keys: Sequence[str] = []):
+    state_dict_a = model_a.state_dict()
+    state_dict_b = model_b.state_dict()
+    assert set(state_dict_a.keys()) == set(state_dict_b.keys())
+    for name in state_dict_a.keys():
+        if any(key in name for key in diff_keys):
+            assert torch.allclose(state_dict_a[name], state_dict_b[name]) is False
+        else:
+            assert torch.allclose(state_dict_a[name], state_dict_b[name]) is True
+
+
+def test_lora_train_all_modules():
     model_args, _, _, finetuning_args, _ = get_train_args({"lora_target": "all", **TRAIN_ARGS})
     tokenizer_module = load_tokenizer(model_args)
     model = load_model(tokenizer_module["tokenizer"], model_args, finetuning_args, is_trainable=True)
@@ -41,7 +70,7 @@ def test_lora_all_modules():
     assert linear_modules == {"q_proj", "k_proj", "v_proj", "o_proj", "up_proj", "gate_proj", "down_proj"}
 
 
-def test_lora_extra_modules():
+def test_lora_train_extra_modules():
     model_args, _, _, finetuning_args, _ = get_train_args(
         {"lora_target": "all", "additional_target": "embed_tokens,lm_head", **TRAIN_ARGS}
     )
@@ -61,3 +90,51 @@ def test_lora_extra_modules():
             assert param.dtype == torch.float16
 
     assert extra_modules == {"embed_tokens", "lm_head"}
+
+
+def test_lora_train_old_adapters():
+    model_args, _, _, finetuning_args, _ = get_train_args(
+        {"adapter_name_or_path": TINY_LLAMA_ADAPTER, "create_new_adapter": False, **TRAIN_ARGS}
+    )
+    tokenizer_module = load_tokenizer(model_args)
+    model = load_model(tokenizer_module["tokenizer"], model_args, finetuning_args, is_trainable=True)
+
+    base_model = AutoModelForCausalLM.from_pretrained(TINY_LLAMA, torch_dtype=model.dtype, device_map=model.device)
+    ref_model = PeftModel.from_pretrained(base_model, TINY_LLAMA_ADAPTER, is_trainable=True)
+    for param in filter(lambda p: p.requires_grad, ref_model.parameters()):
+        param.data = param.data.to(torch.float32)
+
+    compare_model(model, ref_model)
+
+
+def test_lora_train_new_adapters():
+    model_args, _, _, finetuning_args, _ = get_train_args(
+        {"adapter_name_or_path": TINY_LLAMA_ADAPTER, "create_new_adapter": True, **TRAIN_ARGS}
+    )
+    tokenizer_module = load_tokenizer(model_args)
+    model = load_model(tokenizer_module["tokenizer"], model_args, finetuning_args, is_trainable=True)
+
+    base_model = AutoModelForCausalLM.from_pretrained(TINY_LLAMA, torch_dtype=model.dtype, device_map=model.device)
+    ref_model = PeftModel.from_pretrained(base_model, TINY_LLAMA_ADAPTER, is_trainable=True)
+    for param in filter(lambda p: p.requires_grad, ref_model.parameters()):
+        param.data = param.data.to(torch.float32)
+
+    compare_model(
+        model, ref_model, diff_keys=["q_proj", "k_proj", "v_proj", "o_proj", "up_proj", "gate_proj", "down_proj"]
+    )
+
+
+def test_lora_inference():
+    model_args, _, finetuning_args, _ = get_infer_args(INFER_ARGS)
+    tokenizer_module = load_tokenizer(model_args)
+    model = load_model(tokenizer_module["tokenizer"], model_args, finetuning_args, is_trainable=False)
+
+    base_model = AutoModelForCausalLM.from_pretrained(TINY_LLAMA, torch_dtype=model.dtype, device_map=model.device)
+    ref_model: "LoraModel" = PeftModel.from_pretrained(base_model, TINY_LLAMA_ADAPTER)
+    ref_model = ref_model.merge_and_unload()
+    compare_model(model, ref_model)
+
+    for name, param in model.named_parameters():
+        assert param.requires_grad is False
+        assert param.dtype == torch.float16
+        assert "lora" not in name