chore: import upstream snapshot with attribution

2026-07-13 12:55:37 +08:00
commit 7ce4c8e27e
5900 changed files with 1668062 additions and 0 deletions
@@ -0,0 +1,81 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import pytest
+import torch
+
+from tests.models.utils import check_embeddings_close
+from vllm.utils.serial_utils import (
+    EMBED_DTYPES,
+    ENDIANNESS,
+    MM_METADATA_DTYPES,
+    EmbedDType,
+    Endianness,
+    MmMetadataDType,
+    binary2tensor,
+    tensor2binary,
+)
+
+FLOAT_EMBED_DTYPES = tuple(EMBED_DTYPES.keys())
+INTEGER_EMBED_DTYPES = tuple(MM_METADATA_DTYPES.keys())
+
+
+def _build_integer_tensor(
+    embed_dtype: MmMetadataDType, shape: tuple[int, ...]
+) -> torch.Tensor:
+    torch_dtype = MM_METADATA_DTYPES[embed_dtype].torch_dtype
+
+    if torch_dtype is torch.bool:
+        return torch.randint(0, 2, shape, dtype=torch.int32).to(torch.bool)
+    if torch_dtype is torch.uint8:
+        return torch.randint(0, 256, shape, dtype=torch.uint8)
+    if torch_dtype is torch.int32:
+        return torch.randint(-(2**20), 2**20, shape, dtype=torch.int32)
+    if torch_dtype is torch.int64:
+        return torch.randint(-(2**62), 2**62, shape, dtype=torch.int64)
+
+    raise AssertionError(f"Unsupported non-floating embed dtype: {embed_dtype}")
+
+
+@pytest.mark.parametrize("endianness", ENDIANNESS)
+@pytest.mark.parametrize("embed_dtype", FLOAT_EMBED_DTYPES)
+@torch.inference_mode()
+def test_encode_and_decode_floats(embed_dtype: EmbedDType, endianness: Endianness):
+    for i in range(10):
+        tensor = torch.rand(2, 3, 5, 7, 11, 13, device="cpu", dtype=torch.float32)
+        shape = tensor.shape
+        binary = tensor2binary(tensor, embed_dtype, endianness)
+        new_tensor = binary2tensor(binary, shape, embed_dtype, endianness).to(
+            torch.float32
+        )
+
+        if embed_dtype in ["float32", "float16"]:
+            torch.testing.assert_close(tensor, new_tensor, atol=0.001, rtol=0.001)
+        elif embed_dtype == "bfloat16":
+            torch.testing.assert_close(tensor, new_tensor, atol=0.01, rtol=0.01)
+        else:  # for fp8
+            torch.testing.assert_close(tensor, new_tensor, atol=0.1, rtol=0.1)
+
+        check_embeddings_close(
+            embeddings_0_lst=tensor.view(1, -1),
+            embeddings_1_lst=new_tensor.view(1, -1),
+            name_0="gt",
+            name_1="new",
+            tol=1e-2,
+        )
+
+
+@pytest.mark.parametrize("endianness", ENDIANNESS)
+@pytest.mark.parametrize("embed_dtype", INTEGER_EMBED_DTYPES)
+@torch.inference_mode()
+def test_encode_and_decode_integers(
+    embed_dtype: MmMetadataDType, endianness: Endianness
+):
+    shape = (2, 3, 5, 7, 11, 13)
+
+    for i in range(10):
+        tensor = _build_integer_tensor(embed_dtype, shape)
+        binary = tensor2binary(tensor, embed_dtype, endianness)
+        new_tensor = binary2tensor(binary, shape, embed_dtype, endianness)
+
+        assert new_tensor.dtype == MM_METADATA_DTYPES[embed_dtype].torch_dtype
+        torch.testing.assert_close(tensor, new_tensor, atol=0, rtol=0)