@Sean Here is an example:

from typing import Annotated
from pydantic import Field

import bentoml
from bentoml.validators import Shape, DType
import numpy as np


@bentoml.service
class ModelService:

    def __init__(self):
        self.model = ""

    @bentoml.api
    def predict(
            self,
            input_arr: Annotated[
                np.ndarray,
                # Shape((-1, 2)), # <-- this doesn't work
                DType("float32")
            ] = Field(description="A (-1, 2)-shaped np.ndarray")
    ) -> np.ndarray:
        return input_arr

Also seems like it does not work with

torch.Tensor

? I’m using

bentoml==1.2.12

.

I believe this fixes the issue:

diff --git a/src/_bentoml_sdk/_pydantic.py b/src/_bentoml_sdk/_pydantic.py
index ae4e49d5..987b0358 100644
--- a/src/_bentoml_sdk/_pydantic.py
+++ b/src/_bentoml_sdk/_pydantic.py
@@ -53,13 +53,13 @@ def numpy_prepare_pydantic_annotations(
     _, remaining_annotations = _known_annotated_metadata.collect_known_metadata(
         annotations
     )
-    for i, annotation in enumerate(remaining_annotations[:]):
+    for annotation in remaining_annotations[:]:
         if isinstance(annotation, Shape):
             shape = annotation.dimensions
-            del remaining_annotations[i]
+            remaining_annotations.remove(annotation)
         elif isinstance(annotation, DType):
             dtype = annotation.dtype
-            del remaining_annotations[i]
+            remaining_annotations.remove(annotation)
     return source, [TensorSchema("numpy-array", dtype, shape), *remaining_annotations]
 
 
@@ -80,13 +80,13 @@ def torch_prepare_pydantic_annotations(
     _, remaining_annotations = _known_annotated_metadata.collect_known_metadata(
         annotations
     )
-    for i, annotation in enumerate(remaining_annotations[:]):
+    for annotation in remaining_annotations[:]:
         if isinstance(annotation, Shape):
             shape = annotation.dimensions
-            del remaining_annotations[i]
+            remaining_annotations.remove(annotation)
         elif isinstance(annotation, DType):
             dtype = annotation.dtype
-            del remaining_annotations[i]
+            remaining_annotations.remove(annotation)
     return source, [TensorSchema("torch-tensor", dtype, shape), *remaining_annotations]
 
 
@@ -106,13 +106,13 @@ def tf_prepare_pydantic_annotations(
     _, remaining_annotations = _known_annotated_metadata.collect_known_metadata(
         annotations
     )
-    for i, annotation in enumerate(remaining_annotations[:]):
+    for annotation in remaining_annotations[:]:
         if isinstance(annotation, Shape):
             shape = annotation.dimensions
-            del remaining_annotations[i]
+            remaining_annotations.remove(annotation)
         elif isinstance(annotation, DType):
             dtype = annotation.dtype
-            del remaining_annotations[i]
+            remaining_annotations.remove(annotation)
     return source, [TensorSchema("tf-tensor", dtype, shape), *remaining_annotations]
 
 
@@ -169,10 +169,10 @@ def pathlib_prepare_pydantic_annotations(
         annotations
     )
     content_type: str | None = None
-    for i, annotation in enumerate(remaining_annotations[:]):
+    for annotation in remaining_annotations[:]:
         if isinstance(annotation, ContentType):
             content_type = annotation.content_type
-            del remaining_annotations[i]
+            remaining_annotations.annotation(remove)
     return source, [FileSchema(content_type=content_type), *remaining_annotations]

Opened a draft PR but need to step away for the rest of the day. Haven’t thoroughly tested it, and there might be better approaches. Feel free to build on top of the PR if you have time, otherwise I can pick it up again tomorrow or this weekend. The issue is deleting from a list with indices from a copy of that list.

I was able to reproduce and looks like we have a bug in supporting multiple validators. We will work with you on the PR and publish a new version before earlier next week.

Thanks for the PR, @Adam Liter. It has been merged and released in the latest version BentoML,

v1.2.13

. 🎉