Merge ef89079712a5dcd53f1d50a89529aea1ac1d9a13 into 254f6b986720c92ddf97fbb1a6a6465da8e87e29

tests/utils_/test_serial_utils.py

@@ -38,3 +38,42 @@ def test_encode_and_decode(embed_dtype: str, endianness: str):
             name_1="new",
             tol=1e-2,
         )
+
+
+@pytest.mark.parametrize("embed_dtype", EMBED_DTYPE_TO_TORCH_DTYPE.keys())
+@torch.inference_mode()
+def test_binary2tensor_no_warning(embed_dtype: str):
+    """Test that binary2tensor does not emit UserWarning about non-writable buffers.
+
+    This addresses issue #26781 where torch.frombuffer on non-writable bytes
+    would emit: "UserWarning: The given buffer is not writable..."
+    """
+    import warnings
+
+    tensor = torch.rand(10, 20, device="cpu", dtype=torch.float32)
+    binary = tensor2binary(tensor, embed_dtype, "native")
+
+    # Capture warnings during binary2tensor call
+    with warnings.catch_warnings(record=True) as w:
+        warnings.simplefilter("always")
+        result = binary2tensor(binary, tensor.shape, embed_dtype, "native")
+
+        # Filter for the specific UserWarning about non-writable buffers
+        buffer_warnings = [
+            warning
+            for warning in w
+            if issubclass(warning.category, UserWarning)
+            and "not writable" in str(warning.message)
+        ]
+        assert len(buffer_warnings) == 0, (
+            f"Expected no warnings about non-writable buffers, got: {buffer_warnings}"
+        )
+
+    # Verify the result is correct
+    result_float = result.to(torch.float32)
+    if embed_dtype in ["float32", "float16"]:
+        torch.testing.assert_close(tensor, result_float, atol=0.001, rtol=0.001)
+    elif embed_dtype == "bfloat16":
+        torch.testing.assert_close(tensor, result_float, atol=0.01, rtol=0.01)
+    else:  # fp8
+        torch.testing.assert_close(tensor, result_float, atol=0.1, rtol=0.1)

vllm/model_executor/model_loader/weight_utils.py

@@ -66,7 +66,7 @@ logger = init_logger(__name__)
 temp_dir = tempfile.gettempdir()
 
 
-def enable_hf_transfer():
+def enable_hf_transfer() -> None:
     """automatically activates hf_transfer"""
     if "HF_HUB_ENABLE_HF_TRANSFER" not in os.environ:
         try:
@@ -87,7 +87,9 @@ class DisabledTqdm(tqdm):
         super().__init__(*args, **kwargs)
 
 
-def get_lock(model_name_or_path: str | Path, cache_dir: str | None = None):
+def get_lock(
+    model_name_or_path: str | Path, cache_dir: str | None = None
+) -> filelock.FileLock:
     lock_dir = cache_dir or temp_dir
     model_name_or_path = str(model_name_or_path)
     os.makedirs(os.path.dirname(lock_dir), exist_ok=True)
@@ -178,11 +180,11 @@ def maybe_download_from_modelscope(
     return None
 
 
-def _shared_pointers(tensors):
-    ptrs = defaultdict(list)
+def _shared_pointers(tensors: dict[str, torch.Tensor]) -> list[list[str]]:
+    ptrs: dict[int, list[str]] = defaultdict(list)
     for k, v in tensors.items():
         ptrs[v.data_ptr()].append(k)
-    failing = []
+    failing: list[list[str]] = []
     for _, names in ptrs.items():
         if len(names) > 1:
             failing.append(names)
@@ -602,7 +604,7 @@ def filter_files_not_needed_for_inference(hf_weights_files: list[str]) -> list[s
 _BAR_FORMAT = "{desc}: {percentage:3.0f}% Completed | {n_fmt}/{total_fmt} [{elapsed}<{remaining}, {rate_fmt}]\n"  # noqa: E501
 
 
-def enable_tqdm(use_tqdm_on_load: bool):
+def enable_tqdm(use_tqdm_on_load: bool) -> bool:
     return use_tqdm_on_load and (
         not torch.distributed.is_initialized() or torch.distributed.get_rank() == 0
     )

vllm/model_executor/utils.py

@@ -19,7 +19,7 @@ def set_random_seed(seed: int | None) -> None:
 def set_weight_attrs(
     weight: torch.Tensor,
     weight_attrs: dict[str, Any] | None,
-):
+) -> None:
     """Set attributes on a weight tensor.
 
     This method is used to set attributes on a weight tensor. This method
@@ -50,7 +50,9 @@ def set_weight_attrs(
         setattr(weight, key, value)
 
 
-def replace_parameter(layer: torch.nn.Module, param_name: str, new_data: torch.Tensor):
+def replace_parameter(
+    layer: torch.nn.Module, param_name: str, new_data: torch.Tensor
+) -> None:
     """
     Replace a parameter of a layer while maintaining the ability to reload the weight.
     Called within implementations of the `process_weights_after_loading` method.

vllm/utils/serial_utils.py

@@ -107,7 +107,10 @@ def binary2tensor(
     torch_dtype = EMBED_DTYPE_TO_TORCH_DTYPE[embed_dtype]
     np_dtype = EMBED_DTYPE_TO_NUMPY_DTYPE_VIEW[embed_dtype]
 
-    np_array = np.frombuffer(binary, dtype=np_dtype).reshape(shape)
+    # Use bytearray to create a mutable copy of the binary data.
+    # This ensures np.frombuffer returns a writable array, avoiding
+    # UserWarning from torch.from_numpy on read-only arrays.
+    np_array = np.frombuffer(bytearray(binary), dtype=np_dtype).reshape(shape)
 
     if endianness != "native" and endianness != sys_byteorder:
         np_array = np_array.byteswap()