add tests

Makefile

@@ -11,4 +11,4 @@ style:
 	ruff format $(check_dirs)
 
 test:
-	pytest tests/
+	CUDA_VISIBLE_DEVICES= pytest tests/

src/llamafactory/extras/misc.py

@@ -22,6 +22,7 @@ from transformers import InfNanRemoveLogitsProcessor, LogitsProcessorList, PreTr
 from transformers.utils import (
     SAFE_WEIGHTS_NAME,
     WEIGHTS_NAME,
+    is_safetensors_available,
     is_torch_bf16_gpu_available,
     is_torch_cuda_available,
     is_torch_mps_available,
@@ -34,6 +35,11 @@ from .constants import V_HEAD_SAFE_WEIGHTS_NAME, V_HEAD_WEIGHTS_NAME
 from .logging import get_logger
 
 
+if is_safetensors_available():
+    from safetensors import safe_open
+    from safetensors.torch import save_file
+
+
 _is_fp16_available = is_torch_npu_available() or is_torch_cuda_available()
 try:
     _is_bf16_available = is_torch_bf16_gpu_available()
@@ -128,9 +134,6 @@ def fix_valuehead_checkpoint(
         return
 
     if safe_serialization:
-        from safetensors import safe_open
-        from safetensors.torch import save_file
-
         path_to_checkpoint = os.path.join(output_dir, SAFE_WEIGHTS_NAME)
         with safe_open(path_to_checkpoint, framework="pt", device="cpu") as f:
             state_dict: Dict[str, torch.Tensor] = {key: f.get_tensor(key) for key in f.keys()}

tests/data/test_supervised.py

@@ -41,7 +41,7 @@ TRAIN_ARGS = {
 }
 
 
-@pytest.mark.parametrize("num_samples", [10])
+@pytest.mark.parametrize("num_samples", [16])
 def test_supervised(num_samples: int):
     model_args, data_args, training_args, _, _ = get_train_args(TRAIN_ARGS)
     tokenizer_module = load_tokenizer(model_args)

tests/model/model_utils/test_checkpointing.py

@@ -0,0 +1,74 @@
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+
+import torch
+
+from llamafactory.extras.misc import get_current_device
+from llamafactory.hparams import get_train_args
+from llamafactory.model import load_model, load_tokenizer
+
+
+TINY_LLAMA = os.environ.get("TINY_LLAMA", "llamafactory/tiny-random-Llama-3")
+
+TRAIN_ARGS = {
+    "model_name_or_path": TINY_LLAMA,
+    "stage": "sft",
+    "do_train": True,
+    "finetuning_type": "lora",
+    "lora_target": "all",
+    "dataset": "llamafactory/tiny-supervised-dataset",
+    "dataset_dir": "ONLINE",
+    "template": "llama3",
+    "cutoff_len": 1024,
+    "overwrite_cache": True,
+    "output_dir": "dummy_dir",
+    "overwrite_output_dir": True,
+    "fp16": True,
+}
+
+
+def test_checkpointing_enable():
+    model_args, _, _, finetuning_args, _ = get_train_args({"disable_gradient_checkpointing": False, **TRAIN_ARGS})
+    tokenizer_module = load_tokenizer(model_args)
+    model = load_model(tokenizer_module["tokenizer"], model_args, finetuning_args, is_trainable=True)
+    for module in filter(lambda m: hasattr(m, "gradient_checkpointing"), model.modules()):
+        assert getattr(module, "gradient_checkpointing") is True
+
+
+def test_checkpointing_disable():
+    model_args, _, _, finetuning_args, _ = get_train_args({"disable_gradient_checkpointing": True, **TRAIN_ARGS})
+    tokenizer_module = load_tokenizer(model_args)
+    model = load_model(tokenizer_module["tokenizer"], model_args, finetuning_args, is_trainable=True)
+    for module in filter(lambda m: hasattr(m, "gradient_checkpointing"), model.modules()):
+        assert getattr(module, "gradient_checkpointing") is False
+
+
+def test_upcast_layernorm():
+    model_args, _, _, finetuning_args, _ = get_train_args({"upcast_layernorm": True, **TRAIN_ARGS})
+    tokenizer_module = load_tokenizer(model_args)
+    model = load_model(tokenizer_module["tokenizer"], model_args, finetuning_args, is_trainable=True)
+    for name, param in model.named_parameters():
+        if param.ndim == 1 and "norm" in name:
+            assert param.dtype == torch.float32
+
+
+def test_upcast_lmhead_output():
+    model_args, _, _, finetuning_args, _ = get_train_args({"upcast_lmhead_output": True, **TRAIN_ARGS})
+    tokenizer_module = load_tokenizer(model_args)
+    model = load_model(tokenizer_module["tokenizer"], model_args, finetuning_args, is_trainable=True)
+    inputs = torch.randn((1, 16), dtype=torch.float16, device=get_current_device())
+    outputs: "torch.Tensor" = model.lm_head(inputs)
+    assert outputs.dtype == torch.float32

tests/model/test_base.py

@@ -13,16 +13,21 @@
 # limitations under the License.
 
 import os
+from typing import Dict
 
 import torch
 from transformers import AutoModelForCausalLM
+from trl import AutoModelForCausalLMWithValueHead
 
+from llamafactory.extras.misc import get_current_device
 from llamafactory.hparams import get_infer_args
 from llamafactory.model import load_model, load_tokenizer
 
 
 TINY_LLAMA = os.environ.get("TINY_LLAMA", "llamafactory/tiny-random-Llama-3")
 
+TINY_LLAMA_VALUEHEAD = os.environ.get("TINY_LLAMA_VALUEHEAD", "llamafactory/tiny-random-Llama-3-valuehead")
+
 INFER_ARGS = {
     "model_name_or_path": TINY_LLAMA,
     "template": "llama3",
@@ -38,9 +43,32 @@ def compare_model(model_a: "torch.nn.Module", model_b: "torch.nn.Module"):
         assert torch.allclose(state_dict_a[name], state_dict_b[name]) is True
 
 
+def post_init(self: "AutoModelForCausalLMWithValueHead", state_dict: Dict[str, "torch.Tensor"]):
+    state_dict = {k[7:]: state_dict[k] for k in state_dict.keys() if k.startswith("v_head.")}
+    self.v_head.load_state_dict(state_dict, strict=False)
+    del state_dict
+
+
 def test_base():
     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)
-    ref_model = AutoModelForCausalLM.from_pretrained(TINY_LLAMA, torch_dtype=model.dtype, device_map=model.device)
+
+    ref_model = AutoModelForCausalLM.from_pretrained(
+        TINY_LLAMA, torch_dtype=torch.float16, device_map=get_current_device()
+    )
+    compare_model(model, ref_model)
+
+
+def test_valuehead():
+    AutoModelForCausalLMWithValueHead.post_init = post_init  # patch for CPU test
+    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, add_valuehead=True
+    )
+
+    ref_model = AutoModelForCausalLMWithValueHead.from_pretrained(
+        TINY_LLAMA_VALUEHEAD, torch_dtype=torch.float16, device_map=get_current_device()
+    )
     compare_model(model, ref_model)

tests/model/test_freeze.py

@@ -49,6 +49,7 @@ def test_freeze_train_all_modules():
     model_args, _, _, finetuning_args, _ = get_train_args({"freeze_trainable_layers": 1, **TRAIN_ARGS})
     tokenizer_module = load_tokenizer(model_args)
     model = load_model(tokenizer_module["tokenizer"], model_args, finetuning_args, is_trainable=True)
+
     for name, param in model.named_parameters():
         if name.startswith("model.layers.1."):
             assert param.requires_grad is True
@@ -64,6 +65,7 @@ def test_freeze_train_extra_modules():
     )
     tokenizer_module = load_tokenizer(model_args)
     model = load_model(tokenizer_module["tokenizer"], model_args, finetuning_args, is_trainable=True)
+
     for name, param in model.named_parameters():
         if name.startswith("model.layers.1.") or any(module in name for module in ["embed_tokens", "lm_head"]):
             assert param.requires_grad is True
@@ -77,6 +79,7 @@ def test_freeze_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)
+
     for param in model.parameters():
         assert param.requires_grad is False
         assert param.dtype == torch.float16

tests/model/test_full.py

@@ -49,6 +49,7 @@ def test_full_train():
     model_args, _, _, finetuning_args, _ = get_train_args(TRAIN_ARGS)
     tokenizer_module = load_tokenizer(model_args)
     model = load_model(tokenizer_module["tokenizer"], model_args, finetuning_args, is_trainable=True)
+
     for param in model.parameters():
         assert param.requires_grad is True
         assert param.dtype == torch.float32
@@ -58,6 +59,7 @@ def test_full_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)
+
     for param in model.parameters():
         assert param.requires_grad is False
         assert param.dtype == torch.float16

tests/model/test_lora.py

@@ -18,7 +18,9 @@ from typing import Sequence
 import torch
 from peft import LoraModel, PeftModel
 from transformers import AutoModelForCausalLM
+from trl import AutoModelForCausalLMWithValueHead
 
+from llamafactory.extras.misc import get_current_device
 from llamafactory.hparams import get_infer_args, get_train_args
 from llamafactory.model import load_model, load_tokenizer
 
@@ -27,6 +29,8 @@ 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")
 
+TINY_LLAMA_VALUEHEAD = os.environ.get("TINY_LLAMA_VALUEHEAD", "llamafactory/tiny-random-Llama-3-valuehead")
+
 TRAIN_ARGS = {
     "model_name_or_path": TINY_LLAMA,
     "stage": "sft",
@@ -67,10 +71,29 @@ def compare_model(model_a: "torch.nn.Module", model_b: "torch.nn.Module", diff_k
             assert torch.allclose(state_dict_a[name], state_dict_b[name]) is True
 
 
+def test_lora_train_qv_modules():
+    model_args, _, _, finetuning_args, _ = get_train_args({"lora_target": "q_proj,v_proj", **TRAIN_ARGS})
+    tokenizer_module = load_tokenizer(model_args)
+    model = load_model(tokenizer_module["tokenizer"], model_args, finetuning_args, is_trainable=True)
+
+    linear_modules = set()
+    for name, param in model.named_parameters():
+        if any(module in name for module in ["lora_A", "lora_B"]):
+            linear_modules.add(name.split(".lora_", maxsplit=1)[0].split(".")[-1])
+            assert param.requires_grad is True
+            assert param.dtype == torch.float32
+        else:
+            assert param.requires_grad is False
+            assert param.dtype == torch.float16
+
+    assert linear_modules == {"q_proj", "v_proj"}
+
+
 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)
+
     linear_modules = set()
     for name, param in model.named_parameters():
         if any(module in name for module in ["lora_A", "lora_B"]):
@@ -90,6 +113,7 @@ def test_lora_train_extra_modules():
     )
     tokenizer_module = load_tokenizer(model_args)
     model = load_model(tokenizer_module["tokenizer"], model_args, finetuning_args, is_trainable=True)
+
     extra_modules = set()
     for name, param in model.named_parameters():
         if any(module in name for module in ["lora_A", "lora_B"]):
@@ -113,7 +137,9 @@ def test_lora_train_old_adapters():
     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)
+    base_model = AutoModelForCausalLM.from_pretrained(
+        TINY_LLAMA, torch_dtype=torch.float16, device_map=get_current_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)
@@ -128,7 +154,9 @@ def test_lora_train_new_adapters():
     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)
+    base_model = AutoModelForCausalLM.from_pretrained(
+        TINY_LLAMA, torch_dtype=torch.float16, device_map=get_current_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)
@@ -138,17 +166,31 @@ def test_lora_train_new_adapters():
     )
 
 
+def test_lora_train_valuehead():
+    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=True, add_valuehead=True
+    )
+
+    ref_model: "AutoModelForCausalLMWithValueHead" = AutoModelForCausalLMWithValueHead.from_pretrained(
+        TINY_LLAMA_VALUEHEAD, torch_dtype=torch.float16, device_map=get_current_device()
+    )
+    state_dict = model.state_dict()
+    ref_state_dict = ref_model.state_dict()
+
+    assert torch.allclose(state_dict["v_head.summary.weight"], ref_state_dict["v_head.summary.weight"])
+    assert torch.allclose(state_dict["v_head.summary.bias"], ref_state_dict["v_head.summary.bias"])
+
+
 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)
+    base_model = AutoModelForCausalLM.from_pretrained(
+        TINY_LLAMA, torch_dtype=torch.float16, device_map=get_current_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