From f68dc36f20c7b8a647358340c00eaa7d522e399c Mon Sep 17 00:00:00 2001
From: Savyasachi <jsavyasachi@gmail.com>
Date: Mon, 22 Jun 2026 21:52:19 -0700
Subject: [PATCH] fix: ->jvm on object-dtype numpy arrays with np-array
 bindings

The np-array zero-copy bindings register a pyobject->jvm :ndarray method that
always maps the numpy dtype to a native tensor datatype. Object and other
non-numeric dtypes have no such mapping, so ->jvm threw "Unable to find
datatype: object" once the bindings were loaded. Route non-numeric dtypes to
the default element-wise copy conversion instead.

Closes #187
---
 src/libpython_clj2/python/np_array.clj | 21 +++++++++++++++++----
 test/libpython_clj2/numpy_test.clj     |  9 +++++++++
 2 files changed, 26 insertions(+), 4 deletions(-)

diff --git a/src/libpython_clj2/python/np_array.clj b/src/libpython_clj2/python/np_array.clj
index 91522886..869be5cc 100644
--- a/src/libpython_clj2/python/np_array.clj
+++ b/src/libpython_clj2/python/np_array.clj
@@ -73,12 +73,25 @@
        :ctypes ctypes})))
 
 
+(defn- zero-copyable-dtype?
+  "Object/str/datetime arrays hold python-object pointers, not a numeric native
+  buffer, so only the numeric dtypes in py-dtype->dtype-map can become a tensor."
+  [pyobj]
+  (py-ffi/with-gil
+    (let [dtype-name (-> (py-proto/get-attr pyobj "dtype")
+                         (py-proto/get-attr "name"))]
+      (contains? py-dtype->dtype-map dtype-name))))
+
+
 (defmethod py-proto/pyobject->jvm :ndarray
   [pyobj opts]
-  (pygc/with-stack-context
-    (-> (numpy->desc pyobj)
-        (dtt/nd-buffer-descriptor->tensor)
-        (dtt/clone))))
+  (if (zero-copyable-dtype? pyobj)
+    (pygc/with-stack-context
+      (-> (numpy->desc pyobj)
+          (dtt/nd-buffer-descriptor->tensor)
+          (dtt/clone)))
+    ;; non-numeric dtype (object/str/datetime): can't zero-copy, fall back to copy
+    ((get-method py-proto/pyobject->jvm :default) pyobj opts)))
 
 
 (defmethod py-proto/pyobject-as-jvm :ndarray
diff --git a/test/libpython_clj2/numpy_test.clj b/test/libpython_clj2/numpy_test.clj
index 91d409cc..7c32a26e 100644
--- a/test/libpython_clj2/numpy_test.clj
+++ b/test/libpython_clj2/numpy_test.clj
@@ -1,6 +1,8 @@
 (ns libpython-clj2.numpy-test
   (:require [clojure.test :refer [deftest is]]
             [libpython-clj2.python :as py]
+            ;; loading these zero-copy bindings is what made object-dtype ->jvm fail
+            [libpython-clj2.python.np-array]
             [tech.v3.datatype :as dtype]
             [tech.v3.datatype.functional :as dfn]
             [tech.v3.tensor :as dtt]))
@@ -18,3 +20,10 @@
     (is (dfn/equals (dtt/ensure-tensor np-ary) tens))
     (is (dfn/equals [1 2 3 4]
                     (dtype/make-container :java-array :int64 np-ary)))))
+
+
+(deftest object-dtype-ndarray->jvm
+  (let [empty-obj (py/call-attr-kw np-mod "array" [[]] {:dtype "object"})
+        mixed-obj (py/call-attr-kw np-mod "array" [["a" 1]] {:dtype "object"})]
+    (is (= [] (py/->jvm empty-obj)))
+    (is (= ["a" 1] (py/->jvm mixed-obj)))))