test llama_factory

llama_factory/LLaMA-Factory/tests/model/model_utils/test_attention.py

@@ -0,0 +1,47 @@
+# 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
+
+from transformers.utils import is_flash_attn_2_available, is_torch_sdpa_available
+
+from llamafactory.train.test_utils import load_infer_model
+
+
+TINY_LLAMA = os.getenv("TINY_LLAMA", "llamafactory/tiny-random-Llama-3")
+
+INFER_ARGS = {
+    "model_name_or_path": TINY_LLAMA,
+    "template": "llama3",
+}
+
+
+def test_attention():
+    attention_available = ["disabled"]
+    if is_torch_sdpa_available():
+        attention_available.append("sdpa")
+
+    if is_flash_attn_2_available():
+        attention_available.append("fa2")
+
+    llama_attention_classes = {
+        "disabled": "LlamaAttention",
+        "sdpa": "LlamaSdpaAttention",
+        "fa2": "LlamaFlashAttention2",
+    }
+    for requested_attention in attention_available:
+        model = load_infer_model(flash_attn=requested_attention, **INFER_ARGS)
+        for module in model.modules():
+            if "Attention" in module.__class__.__name__:
+                assert module.__class__.__name__ == llama_attention_classes[requested_attention]

llama_factory/LLaMA-Factory/tests/model/model_utils/test_checkpointing.py

@@ -0,0 +1,71 @@
+# 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.train.test_utils import load_train_model
+
+
+TINY_LLAMA = os.getenv("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 = load_train_model(disable_gradient_checkpointing=False, **TRAIN_ARGS)
+    for module in filter(lambda m: hasattr(m, "gradient_checkpointing"), model.modules()):
+        assert getattr(module, "gradient_checkpointing") is True
+
+
+def test_checkpointing_disable():
+    model = load_train_model(disable_gradient_checkpointing=True, **TRAIN_ARGS)
+    for module in filter(lambda m: hasattr(m, "gradient_checkpointing"), model.modules()):
+        assert getattr(module, "gradient_checkpointing") is False
+
+
+def test_unsloth_gradient_checkpointing():
+    model = load_train_model(use_unsloth_gc=True, **TRAIN_ARGS)
+    for module in filter(lambda m: hasattr(m, "gradient_checkpointing"), model.modules()):
+        assert module._gradient_checkpointing_func.__self__.__name__ == "UnslothGradientCheckpointing"
+
+
+def test_upcast_layernorm():
+    model = load_train_model(upcast_layernorm=True, **TRAIN_ARGS)
+    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 = load_train_model(upcast_lmhead_output=True, **TRAIN_ARGS)
+    inputs = torch.randn((1, 16), dtype=torch.float16, device=get_current_device())
+    outputs: "torch.Tensor" = model.get_output_embeddings()(inputs)
+    assert outputs.dtype == torch.float32

llama_factory/LLaMA-Factory/tests/model/model_utils/test_packing.py

@@ -0,0 +1,68 @@
+# 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 pytest
+import torch
+
+from llamafactory.model.model_utils.packing import get_seqlens_in_batch, get_unpad_data
+
+
+@pytest.mark.parametrize(
+    "attention_mask,golden_seq_lens",
+    [
+        (
+            [
+                [1, 1, 2, 2, 2, 0],
+                [1, 2, 2, 3, 3, 3],
+            ],
+            [2, 3, 1, 2, 3],
+        ),
+        (
+            [[1]],
+            [1],
+        ),
+    ],
+)
+def test_get_seqlens_in_batch(attention_mask, golden_seq_lens):
+    attention_mask_with_indices = torch.tensor(attention_mask)
+    seqlens_in_batch = get_seqlens_in_batch(attention_mask_with_indices)
+    assert torch.all(seqlens_in_batch == torch.tensor(golden_seq_lens))
+
+
+@pytest.mark.parametrize(
+    "attention_mask,golden_indices,golden_cu_seqlens,golden_max_seqlen",
+    [
+        (
+            [
+                [1, 1, 2, 2, 2, 0],
+                [1, 2, 2, 3, 3, 3],
+            ],
+            [0, 1, 2, 3, 4, 6, 7, 8, 9, 10, 11],
+            [0, 2, 5, 6, 8, 11],
+            3,
+        ),
+        (
+            [[1]],
+            [0],
+            [0, 1],
+            1,
+        ),
+    ],
+)
+def test_get_unpad_data(attention_mask, golden_indices, golden_cu_seqlens, golden_max_seqlen):
+    attention_mask_with_indices = torch.tensor(attention_mask)
+    indices, cu_seqlens, max_seqlen_in_batch = get_unpad_data(attention_mask_with_indices)
+    assert torch.all(indices == torch.tensor(golden_indices))
+    assert torch.all(cu_seqlens == torch.tensor(golden_cu_seqlens, dtype=torch.int32))
+    assert max_seqlen_in_batch == golden_max_seqlen