Add files using upload-large-folder tool

.gitattributes

@@ -319,3 +319,6 @@ llava_next/lib/python3.10/site-packages/nvidia/cublas/lib/libnvblas.so.12 filter
 llava_next/lib/python3.10/site-packages/nvidia/cufft/lib/libcufftw.so.11 filter=lfs diff=lfs merge=lfs -text
 parrot/lib/python3.10/site-packages/pyarrow/libarrow_python.so filter=lfs diff=lfs merge=lfs -text
 parrot/lib/python3.10/site-packages/transformers/generation/__pycache__/tf_utils.cpython-310.pyc filter=lfs diff=lfs merge=lfs -text
+llava_next/lib/python3.10/site-packages/pandas/_libs/tslib.cpython-310-x86_64-linux-gnu.so filter=lfs diff=lfs merge=lfs -text
+llava_next/lib/python3.10/site-packages/altair/vegalite/v5/schema/__pycache__/channels.cpython-310.pyc filter=lfs diff=lfs merge=lfs -text
+llava_next/lib/python3.10/site-packages/pandas/_libs/sparse.cpython-310-x86_64-linux-gnu.so filter=lfs diff=lfs merge=lfs -text

llava_next/lib/python3.10/site-packages/altair/vegalite/v5/schema/__pycache__/channels.cpython-310.pyc

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0ea855103e6064cfe436f33d30a1231403155565b5485d4d2199d6cdbd426950
+size 1002304

llava_next/lib/python3.10/site-packages/nvidia/cusolver/include/cusolverRf.h

@@ -0,0 +1,339 @@
+/*
+ * Copyright 1993-2014 NVIDIA Corporation.  All rights reserved.
+ *
+ * NOTICE TO LICENSEE:
+ *
+ * This source code and/or documentation ("Licensed Deliverables") are
+ * subject to NVIDIA intellectual property rights under U.S. and
+ * international Copyright laws.
+ *
+ * These Licensed Deliverables contained herein is PROPRIETARY and
+ * CONFIDENTIAL to NVIDIA and is being provided under the terms and
+ * conditions of a form of NVIDIA software license agreement by and
+ * between NVIDIA and Licensee ("License Agreement") or electronically
+ * accepted by Licensee.  Notwithstanding any terms or conditions to
+ * the contrary in the License Agreement, reproduction or disclosure
+ * of the Licensed Deliverables to any third party without the express
+ * written consent of NVIDIA is prohibited.
+ *
+ * NOTWITHSTANDING ANY TERMS OR CONDITIONS TO THE CONTRARY IN THE
+ * LICENSE AGREEMENT, NVIDIA MAKES NO REPRESENTATION ABOUT THE
+ * SUITABILITY OF THESE LICENSED DELIVERABLES FOR ANY PURPOSE.  IT IS
+ * PROVIDED "AS IS" WITHOUT EXPRESS OR IMPLIED WARRANTY OF ANY KIND.
+ * NVIDIA DISCLAIMS ALL WARRANTIES WITH REGARD TO THESE LICENSED
+ * DELIVERABLES, INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY,
+ * NONINFRINGEMENT, AND FITNESS FOR A PARTICULAR PURPOSE.
+ * NOTWITHSTANDING ANY TERMS OR CONDITIONS TO THE CONTRARY IN THE
+ * LICENSE AGREEMENT, IN NO EVENT SHALL NVIDIA BE LIABLE FOR ANY
+ * SPECIAL, INDIRECT, INCIDENTAL, OR CONSEQUENTIAL DAMAGES, OR ANY
+ * DAMAGES WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS,
+ * WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS
+ * ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR PERFORMANCE
+ * OF THESE LICENSED DELIVERABLES.
+ *
+ * U.S. Government End Users.  These Licensed Deliverables are a
+ * "commercial item" as that term is defined at 48 C.F.R. 2.101 (OCT
+ * 1995), consisting of "commercial computer software" and "commercial
+ * computer software documentation" as such terms are used in 48
+ * C.F.R. 12.212 (SEPT 1995) and is provided to the U.S. Government
+ * only as a commercial end item.  Consistent with 48 C.F.R.12.212 and
+ * 48 C.F.R. 227.7202-1 through 227.7202-4 (JUNE 1995), all
+ * U.S. Government End Users acquire the Licensed Deliverables with
+ * only those rights set forth herein.
+ *
+ * Any use of the Licensed Deliverables in individual and commercial
+ * software must include, in the user documentation and internal
+ * comments to the code, the above Disclaimer and U.S. Government End
+ * Users Notice.
+ */
+
+#if !defined(CUSOLVERRF_H_)
+  #define CUSOLVERRF_H_
+
+  #include "driver_types.h"
+  #include "cuComplex.h"
+  #include "cusolver_common.h"
+
+  #if defined(__cplusplus)
+extern "C" {
+  #endif /* __cplusplus */
+
+  /* CUSOLVERRF mode */
+  typedef enum {
+    CUSOLVERRF_RESET_VALUES_FAST_MODE_OFF = 0, // default
+    CUSOLVERRF_RESET_VALUES_FAST_MODE_ON = 1
+  } cusolverRfResetValuesFastMode_t;
+
+  /* CUSOLVERRF matrix format */
+  typedef enum {
+    CUSOLVERRF_MATRIX_FORMAT_CSR = 0, // default
+    CUSOLVERRF_MATRIX_FORMAT_CSC = 1
+  } cusolverRfMatrixFormat_t;
+
+  /* CUSOLVERRF unit diagonal */
+  typedef enum {
+    CUSOLVERRF_UNIT_DIAGONAL_STORED_L = 0, // default
+    CUSOLVERRF_UNIT_DIAGONAL_STORED_U = 1,
+    CUSOLVERRF_UNIT_DIAGONAL_ASSUMED_L = 2,
+    CUSOLVERRF_UNIT_DIAGONAL_ASSUMED_U = 3
+  } cusolverRfUnitDiagonal_t;
+
+  /* CUSOLVERRF factorization algorithm */
+  typedef enum {
+    CUSOLVERRF_FACTORIZATION_ALG0 = 0, // default
+    CUSOLVERRF_FACTORIZATION_ALG1 = 1,
+    CUSOLVERRF_FACTORIZATION_ALG2 = 2,
+  } cusolverRfFactorization_t;
+
+  /* CUSOLVERRF triangular solve algorithm */
+  typedef enum {
+    CUSOLVERRF_TRIANGULAR_SOLVE_ALG1 = 1, // default
+    CUSOLVERRF_TRIANGULAR_SOLVE_ALG2 = 2,
+    CUSOLVERRF_TRIANGULAR_SOLVE_ALG3 = 3
+  } cusolverRfTriangularSolve_t;
+
+  /* CUSOLVERRF numeric boost report */
+  typedef enum {
+    CUSOLVERRF_NUMERIC_BOOST_NOT_USED = 0, // default
+    CUSOLVERRF_NUMERIC_BOOST_USED = 1
+  } cusolverRfNumericBoostReport_t;
+
+  /* Opaque structure holding CUSOLVERRF library common */
+  struct cusolverRfCommon;
+  typedef struct cusolverRfCommon* cusolverRfHandle_t;
+
+  /* CUSOLVERRF create (allocate memory) and destroy (free memory) in the handle
+   */
+  cusolverStatus_t CUSOLVERAPI cusolverRfCreate(cusolverRfHandle_t* handle);
+  cusolverStatus_t CUSOLVERAPI cusolverRfDestroy(cusolverRfHandle_t handle);
+
+  /* CUSOLVERRF set and get input format */
+  cusolverStatus_t CUSOLVERAPI cusolverRfGetMatrixFormat(
+    cusolverRfHandle_t        handle,
+    cusolverRfMatrixFormat_t* format,
+    cusolverRfUnitDiagonal_t* diag);
+
+  cusolverStatus_t CUSOLVERAPI cusolverRfSetMatrixFormat(
+    cusolverRfHandle_t       handle,
+    cusolverRfMatrixFormat_t format,
+    cusolverRfUnitDiagonal_t diag);
+
+  /* CUSOLVERRF set and get numeric properties */
+  cusolverStatus_t CUSOLVERAPI cusolverRfSetNumericProperties(
+    cusolverRfHandle_t handle,
+    double             zero,
+    double             boost);
+
+  cusolverStatus_t CUSOLVERAPI cusolverRfGetNumericProperties(
+    cusolverRfHandle_t handle,
+    double*            zero,
+    double*            boost);
+
+  cusolverStatus_t CUSOLVERAPI cusolverRfGetNumericBoostReport(
+    cusolverRfHandle_t              handle,
+    cusolverRfNumericBoostReport_t* report);
+
+  /* CUSOLVERRF choose the triangular solve algorithm */
+  cusolverStatus_t CUSOLVERAPI cusolverRfSetAlgs(
+    cusolverRfHandle_t          handle,
+    cusolverRfFactorization_t   factAlg,
+    cusolverRfTriangularSolve_t solveAlg);
+
+  cusolverStatus_t CUSOLVERAPI cusolverRfGetAlgs(
+    cusolverRfHandle_t           handle,
+    cusolverRfFactorization_t*   factAlg,
+    cusolverRfTriangularSolve_t* solveAlg);
+
+  /* CUSOLVERRF set and get fast mode */
+  cusolverStatus_t CUSOLVERAPI cusolverRfGetResetValuesFastMode(
+    cusolverRfHandle_t               handle,
+    cusolverRfResetValuesFastMode_t* fastMode);
+
+  cusolverStatus_t CUSOLVERAPI cusolverRfSetResetValuesFastMode(
+    cusolverRfHandle_t              handle,
+    cusolverRfResetValuesFastMode_t fastMode);
+
+  /*** Non-Batched Routines ***/
+  /* CUSOLVERRF setup of internal structures from host or device memory */
+  cusolverStatus_t CUSOLVERAPI
+    cusolverRfSetupHost(/* Input (in the host memory) */
+                        int     n,
+                        int     nnzA,
+                        int*    h_csrRowPtrA,
+                        int*    h_csrColIndA,
+                        double* h_csrValA,
+                        int     nnzL,
+                        int*    h_csrRowPtrL,
+                        int*    h_csrColIndL,
+                        double* h_csrValL,
+                        int     nnzU,
+                        int*    h_csrRowPtrU,
+                        int*    h_csrColIndU,
+                        double* h_csrValU,
+                        int*    h_P,
+                        int*    h_Q,
+                        /* Output */
+                        cusolverRfHandle_t handle);
+
+  cusolverStatus_t CUSOLVERAPI
+    cusolverRfSetupDevice(/* Input (in the device memory) */
+                          int     n,
+                          int     nnzA,
+                          int*    csrRowPtrA,
+                          int*    csrColIndA,
+                          double* csrValA,
+                          int     nnzL,
+                          int*    csrRowPtrL,
+                          int*    csrColIndL,
+                          double* csrValL,
+                          int     nnzU,
+                          int*    csrRowPtrU,
+                          int*    csrColIndU,
+                          double* csrValU,
+                          int*    P,
+                          int*    Q,
+                          /* Output */
+                          cusolverRfHandle_t handle);
+
+  /* CUSOLVERRF update the matrix values (assuming the reordering, pivoting
+     and consequently the sparsity pattern of L and U did not change),
+     and zero out the remaining values. */
+  cusolverStatus_t CUSOLVERAPI
+    cusolverRfResetValues(/* Input (in the device memory) */
+                          int     n,
+                          int     nnzA,
+                          int*    csrRowPtrA,
+                          int*    csrColIndA,
+                          double* csrValA,
+                          int*    P,
+                          int*    Q,
+                          /* Output */
+                          cusolverRfHandle_t handle);
+
+  /* CUSOLVERRF analysis (for parallelism) */
+  cusolverStatus_t CUSOLVERAPI cusolverRfAnalyze(cusolverRfHandle_t handle);
+
+  /* CUSOLVERRF re-factorization (for parallelism) */
+  cusolverStatus_t CUSOLVERAPI cusolverRfRefactor(cusolverRfHandle_t handle);
+
+  /* CUSOLVERRF extraction: Get L & U packed into a single matrix M */
+  cusolverStatus_t CUSOLVERAPI
+    cusolverRfAccessBundledFactorsDevice(/* Input */
+                                         cusolverRfHandle_t handle,
+                                         /* Output (in the host memory) */
+                                         int* nnzM,
+                                         /* Output (in the device memory) */
+                                         int**    Mp,
+                                         int**    Mi,
+                                         double** Mx);
+
+  cusolverStatus_t CUSOLVERAPI
+    cusolverRfExtractBundledFactorsHost(/* Input */
+                                        cusolverRfHandle_t handle,
+                                        /* Output (in the host memory) */
+                                        int*     h_nnzM,
+                                        int**    h_Mp,
+                                        int**    h_Mi,
+                                        double** h_Mx);
+
+  /* CUSOLVERRF extraction: Get L & U individually */
+  cusolverStatus_t CUSOLVERAPI
+    cusolverRfExtractSplitFactorsHost(/* Input */
+                                      cusolverRfHandle_t handle,
+                                      /* Output (in the host memory) */
+                                      int*     h_nnzL,
+                                      int**    h_csrRowPtrL,
+                                      int**    h_csrColIndL,
+                                      double** h_csrValL,
+                                      int*     h_nnzU,
+                                      int**    h_csrRowPtrU,
+                                      int**    h_csrColIndU,
+                                      double** h_csrValU);
+
+  /* CUSOLVERRF (forward and backward triangular) solves */
+  cusolverStatus_t CUSOLVERAPI
+    cusolverRfSolve(/* Input (in the device memory) */
+                    cusolverRfHandle_t handle,
+                    int*               P,
+                    int*               Q,
+                    int                nrhs, // only nrhs=1 is supported
+                    double*            Temp, // of size ldt*nrhs (ldt>=n)
+                    int                ldt,
+                    /* Input/Output (in the device memory) */
+                    double* XF,
+                    /* Input */
+                    int ldxf);
+
+  /*** Batched Routines ***/
+  /* CUSOLVERRF-batch setup of internal structures from host */
+  cusolverStatus_t CUSOLVERAPI
+    cusolverRfBatchSetupHost(/* Input (in the host memory)*/
+                             int     batchSize,
+                             int     n,
+                             int     nnzA,
+                             int*    h_csrRowPtrA,
+                             int*    h_csrColIndA,
+                             double* h_csrValA_array[],
+                             int     nnzL,
+                             int*    h_csrRowPtrL,
+                             int*    h_csrColIndL,
+                             double* h_csrValL,
+                             int     nnzU,
+                             int*    h_csrRowPtrU,
+                             int*    h_csrColIndU,
+                             double* h_csrValU,
+                             int*    h_P,
+                             int*    h_Q,
+                             /* Output (in the device memory) */
+                             cusolverRfHandle_t handle);
+
+  /* CUSOLVERRF-batch update the matrix values (assuming the reordering,
+     pivoting and consequently the sparsity pattern of L and U did not change),
+     and zero out the remaining values. */
+  cusolverStatus_t CUSOLVERAPI
+    cusolverRfBatchResetValues(/* Input (in the device memory) */
+                               int     batchSize,
+                               int     n,
+                               int     nnzA,
+                               int*    csrRowPtrA,
+                               int*    csrColIndA,
+                               double* csrValA_array[],
+                               int*    P,
+                               int*    Q,
+                               /* Output */
+                               cusolverRfHandle_t handle);
+
+  /* CUSOLVERRF-batch analysis (for parallelism) */
+  cusolverStatus_t CUSOLVERAPI
+    cusolverRfBatchAnalyze(cusolverRfHandle_t handle);
+
+  /* CUSOLVERRF-batch re-factorization (for parallelism) */
+  cusolverStatus_t CUSOLVERAPI
+    cusolverRfBatchRefactor(cusolverRfHandle_t handle);
+
+  /* CUSOLVERRF-batch (forward and backward triangular) solves */
+  cusolverStatus_t CUSOLVERAPI
+    cusolverRfBatchSolve(/* Input (in the device memory) */
+                         cusolverRfHandle_t handle,
+                         int*               P,
+                         int*               Q,
+                         int                nrhs, // only nrhs=1 is supported
+                         double* Temp, // of size 2*batchSize*(n*nrhs)
+                         int     ldt,  // only ldt=n is supported
+                         /* Input/Output (in the device memory) */
+                         double* XF_array[],
+                         /* Input */
+                         int ldxf);
+
+  /* CUSOLVERRF-batch obtain the position of zero pivot */
+  cusolverStatus_t CUSOLVERAPI
+    cusolverRfBatchZeroPivot(/* Input */
+                             cusolverRfHandle_t handle,
+                             /* Output (in the host memory) */
+                             int* position);
+
+  #if defined(__cplusplus)
+}
+  #endif /* __cplusplus */
+
+#endif /* CUSOLVERRF_H_ */

llava_next/lib/python3.10/site-packages/nvidia/cusolver/include/cusolver_common.h

@@ -0,0 +1,266 @@
+/*
+ * Copyright 2014 NVIDIA Corporation.  All rights reserved.
+ *
+ * NOTICE TO LICENSEE:
+ *
+ * This source code and/or documentation ("Licensed Deliverables") are
+ * subject to NVIDIA intellectual property rights under U.S. and
+ * international Copyright laws.
+ *
+ * These Licensed Deliverables contained herein is PROPRIETARY and
+ * CONFIDENTIAL to NVIDIA and is being provided under the terms and
+ * conditions of a form of NVIDIA software license agreement by and
+ * between NVIDIA and Licensee ("License Agreement") or electronically
+ * accepted by Licensee.  Notwithstanding any terms or conditions to
+ * the contrary in the License Agreement, reproduction or disclosure
+ * of the Licensed Deliverables to any third party without the express
+ * written consent of NVIDIA is prohibited.
+ *
+ * NOTWITHSTANDING ANY TERMS OR CONDITIONS TO THE CONTRARY IN THE
+ * LICENSE AGREEMENT, NVIDIA MAKES NO REPRESENTATION ABOUT THE
+ * SUITABILITY OF THESE LICENSED DELIVERABLES FOR ANY PURPOSE.  IT IS
+ * PROVIDED "AS IS" WITHOUT EXPRESS OR IMPLIED WARRANTY OF ANY KIND.
+ * NVIDIA DISCLAIMS ALL WARRANTIES WITH REGARD TO THESE LICENSED
+ * DELIVERABLES, INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY,
+ * NONINFRINGEMENT, AND FITNESS FOR A PARTICULAR PURPOSE.
+ * NOTWITHSTANDING ANY TERMS OR CONDITIONS TO THE CONTRARY IN THE
+ * LICENSE AGREEMENT, IN NO EVENT SHALL NVIDIA BE LIABLE FOR ANY
+ * SPECIAL, INDIRECT, INCIDENTAL, OR CONSEQUENTIAL DAMAGES, OR ANY
+ * DAMAGES WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS,
+ * WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS
+ * ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR PERFORMANCE
+ * OF THESE LICENSED DELIVERABLES.
+ *
+ * U.S. Government End Users.  These Licensed Deliverables are a
+ * "commercial item" as that term is defined at 48 C.F.R. 2.101 (OCT
+ * 1995), consisting of "commercial computer software" and "commercial
+ * computer software documentation" as such terms are used in 48
+ * C.F.R. 12.212 (SEPT 1995) and is provided to the U.S. Government
+ * only as a commercial end item.  Consistent with 48 C.F.R.12.212 and
+ * 48 C.F.R. 227.7202-1 through 227.7202-4 (JUNE 1995), all
+ * U.S. Government End Users acquire the Licensed Deliverables with
+ * only those rights set forth herein.
+ *
+ * Any use of the Licensed Deliverables in individual and commercial
+ * software must include, in the user documentation and internal
+ * comments to the code, the above Disclaimer and U.S. Government End
+ * Users Notice.
+ */
+
+#if !defined(CUSOLVER_COMMON_H_)
+  #define CUSOLVER_COMMON_H_
+
+  #include "library_types.h"
+
+  #ifndef CUSOLVERAPI
+    #ifdef _WIN32
+      #define CUSOLVERAPI __stdcall
+    #else
+      #define CUSOLVERAPI
+    #endif
+  #endif
+
+  #if defined(_MSC_VER)
+typedef __int64 int64_t;
+  #else
+    #include <inttypes.h>
+  #endif
+
+typedef int cusolver_int_t;
+
+  #define CUSOLVER_VER_MAJOR 11
+  #define CUSOLVER_VER_MINOR 4
+  #define CUSOLVER_VER_PATCH 5
+  #define CUSOLVER_VER_BUILD 107
+  #define CUSOLVER_VERSION                                                     \
+    (CUSOLVER_VER_MAJOR * 1000 + CUSOLVER_VER_MINOR * 100 + CUSOLVER_VER_PATCH)
+
+  /*
+   * disable this macro to proceed old API
+   */
+  #define DISABLE_CUSOLVER_DEPRECATED
+
+//------------------------------------------------------------------------------
+
+  #if !defined(_MSC_VER)
+    #define CUSOLVER_CPP_VERSION __cplusplus
+  #elif _MSC_FULL_VER >= 190024210 // Visual Studio 2015 Update 3
+    #define CUSOLVER_CPP_VERSION _MSVC_LANG
+  #else
+    #define CUSOLVER_CPP_VERSION 0
+  #endif
+
+//------------------------------------------------------------------------------
+
+  #if !defined(DISABLE_CUSOLVER_DEPRECATED)
+
+    #if CUSOLVER_CPP_VERSION >= 201402L
+
+      #define CUSOLVER_DEPRECATED(new_func)                                    \
+        [[deprecated("please use " #new_func " instead")]]
+
+    #elif defined(_MSC_VER)
+
+      #define CUSOLVER_DEPRECATED(new_func)                                    \
+        __declspec(deprecated("please use " #new_func " instead"))
+
+    #elif defined(__INTEL_COMPILER) || defined(__clang__) ||                   \
+      (defined(__GNUC__) &&                                                    \
+       (__GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 5)))
+
+      #define CUSOLVER_DEPRECATED(new_func)                                    \
+        __attribute__((deprecated("please use " #new_func " instead")))
+
+    #elif defined(__GNUC__) || defined(__xlc__)
+
+      #define CUSOLVER_DEPRECATED(new_func) __attribute__((deprecated))
+
+    #else
+
+      #define CUSOLVER_DEPRECATED(new_func)
+
+    #endif // defined(__cplusplus) && __cplusplus >= 201402L
+  //------------------------------------------------------------------------------
+
+    #if CUSOLVER_CPP_VERSION >= 201703L
+
+      #define CUSOLVER_DEPRECATED_ENUM(new_enum)                               \
+        [[deprecated("please use " #new_enum " instead")]]
+
+    #elif defined(__clang__) ||                                                \
+      (defined(__GNUC__) && __GNUC__ >= 6 && !defined(__PGI))
+
+      #define CUSOLVER_DEPRECATED_ENUM(new_enum)                               \
+        __attribute__((deprecated("please use " #new_enum " instead")))
+
+    #else
+
+      #define CUSOLVER_DEPRECATED_ENUM(new_enum)
+
+    #endif // defined(__cplusplus) && __cplusplus >= 201402L
+
+  #else // defined(DISABLE_CUSOLVER_DEPRECATED)
+
+    #define CUSOLVER_DEPRECATED(new_func)
+    #define CUSOLVER_DEPRECATED_ENUM(new_enum)
+
+  #endif // !defined(DISABLE_CUSOLVER_DEPRECATED)
+
+  #undef CUSOLVER_CPP_VERSION
+
+  #if defined(__cplusplus)
+extern "C" {
+  #endif /* __cplusplus */
+
+  typedef enum {
+    CUSOLVER_STATUS_SUCCESS = 0,
+    CUSOLVER_STATUS_NOT_INITIALIZED = 1,
+    CUSOLVER_STATUS_ALLOC_FAILED = 2,
+    CUSOLVER_STATUS_INVALID_VALUE = 3,
+    CUSOLVER_STATUS_ARCH_MISMATCH = 4,
+    CUSOLVER_STATUS_MAPPING_ERROR = 5,
+    CUSOLVER_STATUS_EXECUTION_FAILED = 6,
+    CUSOLVER_STATUS_INTERNAL_ERROR = 7,
+    CUSOLVER_STATUS_MATRIX_TYPE_NOT_SUPPORTED = 8,
+    CUSOLVER_STATUS_NOT_SUPPORTED = 9,
+    CUSOLVER_STATUS_ZERO_PIVOT = 10,
+    CUSOLVER_STATUS_INVALID_LICENSE = 11,
+    CUSOLVER_STATUS_IRS_PARAMS_NOT_INITIALIZED = 12,
+    CUSOLVER_STATUS_IRS_PARAMS_INVALID = 13,
+    CUSOLVER_STATUS_IRS_PARAMS_INVALID_PREC = 14,
+    CUSOLVER_STATUS_IRS_PARAMS_INVALID_REFINE = 15,
+    CUSOLVER_STATUS_IRS_PARAMS_INVALID_MAXITER = 16,
+    CUSOLVER_STATUS_IRS_INTERNAL_ERROR = 20,
+    CUSOLVER_STATUS_IRS_NOT_SUPPORTED = 21,
+    CUSOLVER_STATUS_IRS_OUT_OF_RANGE = 22,
+    CUSOLVER_STATUS_IRS_NRHS_NOT_SUPPORTED_FOR_REFINE_GMRES = 23,
+    CUSOLVER_STATUS_IRS_INFOS_NOT_INITIALIZED = 25,
+    CUSOLVER_STATUS_IRS_INFOS_NOT_DESTROYED = 26,
+    CUSOLVER_STATUS_IRS_MATRIX_SINGULAR = 30,
+    CUSOLVER_STATUS_INVALID_WORKSPACE = 31
+  } cusolverStatus_t;
+
+  typedef enum {
+    CUSOLVER_EIG_TYPE_1 = 1,
+    CUSOLVER_EIG_TYPE_2 = 2,
+    CUSOLVER_EIG_TYPE_3 = 3
+  } cusolverEigType_t;
+
+  typedef enum {
+    CUSOLVER_EIG_MODE_NOVECTOR = 0,
+    CUSOLVER_EIG_MODE_VECTOR = 1
+  } cusolverEigMode_t;
+
+  typedef enum {
+    CUSOLVER_EIG_RANGE_ALL = 1001,
+    CUSOLVER_EIG_RANGE_I = 1002,
+    CUSOLVER_EIG_RANGE_V = 1003,
+  } cusolverEigRange_t;
+
+  typedef enum {
+    CUSOLVER_INF_NORM = 104,
+    CUSOLVER_MAX_NORM = 105,
+    CUSOLVER_ONE_NORM = 106,
+    CUSOLVER_FRO_NORM = 107,
+  } cusolverNorm_t;
+
+  typedef enum {
+    CUSOLVER_IRS_REFINE_NOT_SET = 1100,
+    CUSOLVER_IRS_REFINE_NONE = 1101,
+    CUSOLVER_IRS_REFINE_CLASSICAL = 1102,
+    CUSOLVER_IRS_REFINE_CLASSICAL_GMRES = 1103,
+    CUSOLVER_IRS_REFINE_GMRES = 1104,
+    CUSOLVER_IRS_REFINE_GMRES_GMRES = 1105,
+    CUSOLVER_IRS_REFINE_GMRES_NOPCOND = 1106,
+
+    CUSOLVER_PREC_DD = 1150,
+    CUSOLVER_PREC_SS = 1151,
+    CUSOLVER_PREC_SHT = 1152,
+
+  } cusolverIRSRefinement_t;
+
+  typedef enum {
+    CUSOLVER_R_8I = 1201,
+    CUSOLVER_R_8U = 1202,
+    CUSOLVER_R_64F = 1203,
+    CUSOLVER_R_32F = 1204,
+    CUSOLVER_R_16F = 1205,
+    CUSOLVER_R_16BF = 1206,
+    CUSOLVER_R_TF32 = 1207,
+    CUSOLVER_R_AP = 1208,
+    CUSOLVER_C_8I = 1211,
+    CUSOLVER_C_8U = 1212,
+    CUSOLVER_C_64F = 1213,
+    CUSOLVER_C_32F = 1214,
+    CUSOLVER_C_16F = 1215,
+    CUSOLVER_C_16BF = 1216,
+    CUSOLVER_C_TF32 = 1217,
+    CUSOLVER_C_AP = 1218,
+  } cusolverPrecType_t;
+
+  typedef enum {
+    CUSOLVER_ALG_0 = 0, /* default algorithm */
+    CUSOLVER_ALG_1 = 1,
+    CUSOLVER_ALG_2 = 2
+  } cusolverAlgMode_t;
+
+  typedef enum {
+    CUBLAS_STOREV_COLUMNWISE = 0,
+    CUBLAS_STOREV_ROWWISE = 1
+  } cusolverStorevMode_t;
+
+  typedef enum {
+    CUBLAS_DIRECT_FORWARD = 0,
+    CUBLAS_DIRECT_BACKWARD = 1
+  } cusolverDirectMode_t;
+
+  cusolverStatus_t CUSOLVERAPI
+    cusolverGetProperty(libraryPropertyType type, int *value);
+
+  cusolverStatus_t CUSOLVERAPI cusolverGetVersion(int *version);
+
+  #if defined(__cplusplus)
+}
+  #endif /* __cplusplus */
+
+#endif // CUSOLVER_COMMON_H_

llava_next/lib/python3.10/site-packages/pandas/_libs/sparse.cpython-310-x86_64-linux-gnu.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d9d4b81d1da2d284b1f61370c3e87b54de10f249a3debb14f7ad7bbf2c100523
+size 988968

llava_next/lib/python3.10/site-packages/pandas/_libs/tslib.cpython-310-x86_64-linux-gnu.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f28ec212e94a75d6066e383be4243e1317de1339e61ad445b8a79f0aed216136
+size 340264

llava_next/lib/python3.10/site-packages/pandas/_libs/tslibs/nattype.pyi

@@ -0,0 +1,141 @@
+from datetime import (
+    datetime,
+    timedelta,
+    tzinfo as _tzinfo,
+)
+import typing
+
+import numpy as np
+
+from pandas._libs.tslibs.period import Period
+from pandas._typing import Self
+
+NaT: NaTType
+iNaT: int
+nat_strings: set[str]
+
+_NaTComparisonTypes: typing.TypeAlias = (
+    datetime | timedelta | Period | np.datetime64 | np.timedelta64
+)
+
+class _NatComparison:
+    def __call__(self, other: _NaTComparisonTypes) -> bool: ...
+
+class NaTType:
+    _value: np.int64
+    @property
+    def value(self) -> int: ...
+    @property
+    def asm8(self) -> np.datetime64: ...
+    def to_datetime64(self) -> np.datetime64: ...
+    def to_numpy(
+        self, dtype: np.dtype | str | None = ..., copy: bool = ...
+    ) -> np.datetime64 | np.timedelta64: ...
+    @property
+    def is_leap_year(self) -> bool: ...
+    @property
+    def is_month_start(self) -> bool: ...
+    @property
+    def is_quarter_start(self) -> bool: ...
+    @property
+    def is_year_start(self) -> bool: ...
+    @property
+    def is_month_end(self) -> bool: ...
+    @property
+    def is_quarter_end(self) -> bool: ...
+    @property
+    def is_year_end(self) -> bool: ...
+    @property
+    def day_of_year(self) -> float: ...
+    @property
+    def dayofyear(self) -> float: ...
+    @property
+    def days_in_month(self) -> float: ...
+    @property
+    def daysinmonth(self) -> float: ...
+    @property
+    def day_of_week(self) -> float: ...
+    @property
+    def dayofweek(self) -> float: ...
+    @property
+    def week(self) -> float: ...
+    @property
+    def weekofyear(self) -> float: ...
+    def day_name(self) -> float: ...
+    def month_name(self) -> float: ...
+    def weekday(self) -> float: ...
+    def isoweekday(self) -> float: ...
+    def total_seconds(self) -> float: ...
+    def today(self, *args, **kwargs) -> NaTType: ...
+    def now(self, *args, **kwargs) -> NaTType: ...
+    def to_pydatetime(self) -> NaTType: ...
+    def date(self) -> NaTType: ...
+    def round(self) -> NaTType: ...
+    def floor(self) -> NaTType: ...
+    def ceil(self) -> NaTType: ...
+    @property
+    def tzinfo(self) -> None: ...
+    @property
+    def tz(self) -> None: ...
+    def tz_convert(self, tz: _tzinfo | str | None) -> NaTType: ...
+    def tz_localize(
+        self,
+        tz: _tzinfo | str | None,
+        ambiguous: str = ...,
+        nonexistent: str = ...,
+    ) -> NaTType: ...
+    def replace(
+        self,
+        year: int | None = ...,
+        month: int | None = ...,
+        day: int | None = ...,
+        hour: int | None = ...,
+        minute: int | None = ...,
+        second: int | None = ...,
+        microsecond: int | None = ...,
+        nanosecond: int | None = ...,
+        tzinfo: _tzinfo | None = ...,
+        fold: int | None = ...,
+    ) -> NaTType: ...
+    @property
+    def year(self) -> float: ...
+    @property
+    def quarter(self) -> float: ...
+    @property
+    def month(self) -> float: ...
+    @property
+    def day(self) -> float: ...
+    @property
+    def hour(self) -> float: ...
+    @property
+    def minute(self) -> float: ...
+    @property
+    def second(self) -> float: ...
+    @property
+    def millisecond(self) -> float: ...
+    @property
+    def microsecond(self) -> float: ...
+    @property
+    def nanosecond(self) -> float: ...
+    # inject Timedelta properties
+    @property
+    def days(self) -> float: ...
+    @property
+    def microseconds(self) -> float: ...
+    @property
+    def nanoseconds(self) -> float: ...
+    # inject Period properties
+    @property
+    def qyear(self) -> float: ...
+    def __eq__(self, other: object) -> bool: ...
+    def __ne__(self, other: object) -> bool: ...
+    __lt__: _NatComparison
+    __le__: _NatComparison
+    __gt__: _NatComparison
+    __ge__: _NatComparison
+    def __sub__(self, other: Self | timedelta | datetime) -> Self: ...
+    def __rsub__(self, other: Self | timedelta | datetime) -> Self: ...
+    def __add__(self, other: Self | timedelta | datetime) -> Self: ...
+    def __radd__(self, other: Self | timedelta | datetime) -> Self: ...
+    def __hash__(self) -> int: ...
+    def as_unit(self, unit: str, round_ok: bool = ...) -> NaTType: ...

llava_next/lib/python3.10/site-packages/pandas/_libs/tslibs/offsets.pyi

@@ -0,0 +1,287 @@
+from datetime import (
+    datetime,
+    time,
+    timedelta,
+)
+from typing import (
+    Any,
+    Collection,
+    Literal,
+    TypeVar,
+    overload,
+)
+
+import numpy as np
+
+from pandas._libs.tslibs.nattype import NaTType
+from pandas._typing import (
+    OffsetCalendar,
+    Self,
+    npt,
+)
+
+from .timedeltas import Timedelta
+
+_BaseOffsetT = TypeVar("_BaseOffsetT", bound=BaseOffset)
+_DatetimeT = TypeVar("_DatetimeT", bound=datetime)
+_TimedeltaT = TypeVar("_TimedeltaT", bound=timedelta)
+
+_relativedelta_kwds: set[str]
+prefix_mapping: dict[str, type]
+
+class ApplyTypeError(TypeError): ...
+
+class BaseOffset:
+    n: int
+    normalize: bool
+    def __init__(self, n: int = ..., normalize: bool = ...) -> None: ...
+    def __eq__(self, other) -> bool: ...
+    def __ne__(self, other) -> bool: ...
+    def __hash__(self) -> int: ...
+    @property
+    def kwds(self) -> dict: ...
+    @property
+    def base(self) -> BaseOffset: ...
+    @overload
+    def __add__(self, other: npt.NDArray[np.object_]) -> npt.NDArray[np.object_]: ...
+    @overload
+    def __add__(self, other: BaseOffset) -> Self: ...
+    @overload
+    def __add__(self, other: _DatetimeT) -> _DatetimeT: ...
+    @overload
+    def __add__(self, other: _TimedeltaT) -> _TimedeltaT: ...
+    @overload
+    def __radd__(self, other: npt.NDArray[np.object_]) -> npt.NDArray[np.object_]: ...
+    @overload
+    def __radd__(self, other: BaseOffset) -> Self: ...
+    @overload
+    def __radd__(self, other: _DatetimeT) -> _DatetimeT: ...
+    @overload
+    def __radd__(self, other: _TimedeltaT) -> _TimedeltaT: ...
+    @overload
+    def __radd__(self, other: NaTType) -> NaTType: ...
+    def __sub__(self, other: BaseOffset) -> Self: ...
+    @overload
+    def __rsub__(self, other: npt.NDArray[np.object_]) -> npt.NDArray[np.object_]: ...
+    @overload
+    def __rsub__(self, other: BaseOffset): ...
+    @overload
+    def __rsub__(self, other: _DatetimeT) -> _DatetimeT: ...
+    @overload
+    def __rsub__(self, other: _TimedeltaT) -> _TimedeltaT: ...
+    @overload
+    def __mul__(self, other: np.ndarray) -> np.ndarray: ...
+    @overload
+    def __mul__(self, other: int): ...
+    @overload
+    def __rmul__(self, other: np.ndarray) -> np.ndarray: ...
+    @overload
+    def __rmul__(self, other: int) -> Self: ...
+    def __neg__(self) -> Self: ...
+    def copy(self) -> Self: ...
+    @property
+    def name(self) -> str: ...
+    @property
+    def rule_code(self) -> str: ...
+    @property
+    def freqstr(self) -> str: ...
+    def _apply(self, other): ...
+    def _apply_array(self, dtarr: np.ndarray) -> np.ndarray: ...
+    def rollback(self, dt: datetime) -> datetime: ...
+    def rollforward(self, dt: datetime) -> datetime: ...
+    def is_on_offset(self, dt: datetime) -> bool: ...
+    def __setstate__(self, state) -> None: ...
+    def __getstate__(self): ...
+    @property
+    def nanos(self) -> int: ...
+    def is_anchored(self) -> bool: ...
+
+def _get_offset(name: str) -> BaseOffset: ...
+
+class SingleConstructorOffset(BaseOffset):
+    @classmethod
+    def _from_name(cls, suffix: None = ...): ...
+    def __reduce__(self): ...
+
+@overload
+def to_offset(freq: None, is_period: bool = ...) -> None: ...
+@overload
+def to_offset(freq: _BaseOffsetT, is_period: bool = ...) -> _BaseOffsetT: ...
+@overload
+def to_offset(freq: timedelta | str, is_period: bool = ...) -> BaseOffset: ...
+
+class Tick(SingleConstructorOffset):
+    _creso: int
+    _prefix: str
+    def __init__(self, n: int = ..., normalize: bool = ...) -> None: ...
+    @property
+    def delta(self) -> Timedelta: ...
+    @property
+    def nanos(self) -> int: ...
+
+def delta_to_tick(delta: timedelta) -> Tick: ...
+
+class Day(Tick): ...
+class Hour(Tick): ...
+class Minute(Tick): ...
+class Second(Tick): ...
+class Milli(Tick): ...
+class Micro(Tick): ...
+class Nano(Tick): ...
+
+class RelativeDeltaOffset(BaseOffset):
+    def __init__(self, n: int = ..., normalize: bool = ..., **kwds: Any) -> None: ...
+
+class BusinessMixin(SingleConstructorOffset):
+    def __init__(
+        self, n: int = ..., normalize: bool = ..., offset: timedelta = ...
+    ) -> None: ...
+
+class BusinessDay(BusinessMixin): ...
+
+class BusinessHour(BusinessMixin):
+    def __init__(
+        self,
+        n: int = ...,
+        normalize: bool = ...,
+        start: str | time | Collection[str | time] = ...,
+        end: str | time | Collection[str | time] = ...,
+        offset: timedelta = ...,
+    ) -> None: ...
+
+class WeekOfMonthMixin(SingleConstructorOffset):
+    def __init__(
+        self, n: int = ..., normalize: bool = ..., weekday: int = ...
+    ) -> None: ...
+
+class YearOffset(SingleConstructorOffset):
+    def __init__(
+        self, n: int = ..., normalize: bool = ..., month: int | None = ...
+    ) -> None: ...
+
+class BYearEnd(YearOffset): ...
+class BYearBegin(YearOffset): ...
+class YearEnd(YearOffset): ...
+class YearBegin(YearOffset): ...
+
+class QuarterOffset(SingleConstructorOffset):
+    def __init__(
+        self, n: int = ..., normalize: bool = ..., startingMonth: int | None = ...
+    ) -> None: ...
+
+class BQuarterEnd(QuarterOffset): ...
+class BQuarterBegin(QuarterOffset): ...
+class QuarterEnd(QuarterOffset): ...
+class QuarterBegin(QuarterOffset): ...
+class MonthOffset(SingleConstructorOffset): ...
+class MonthEnd(MonthOffset): ...
+class MonthBegin(MonthOffset): ...
+class BusinessMonthEnd(MonthOffset): ...
+class BusinessMonthBegin(MonthOffset): ...
+
+class SemiMonthOffset(SingleConstructorOffset):
+    def __init__(
+        self, n: int = ..., normalize: bool = ..., day_of_month: int | None = ...
+    ) -> None: ...
+
+class SemiMonthEnd(SemiMonthOffset): ...
+class SemiMonthBegin(SemiMonthOffset): ...
+
+class Week(SingleConstructorOffset):
+    def __init__(
+        self, n: int = ..., normalize: bool = ..., weekday: int | None = ...
+    ) -> None: ...
+
+class WeekOfMonth(WeekOfMonthMixin):
+    def __init__(
+        self, n: int = ..., normalize: bool = ..., week: int = ..., weekday: int = ...
+    ) -> None: ...
+
+class LastWeekOfMonth(WeekOfMonthMixin): ...
+
+class FY5253Mixin(SingleConstructorOffset):
+    def __init__(
+        self,
+        n: int = ...,
+        normalize: bool = ...,
+        weekday: int = ...,
+        startingMonth: int = ...,
+        variation: Literal["nearest", "last"] = ...,
+    ) -> None: ...
+
+class FY5253(FY5253Mixin): ...
+
+class FY5253Quarter(FY5253Mixin):
+    def __init__(
+        self,
+        n: int = ...,
+        normalize: bool = ...,
+        weekday: int = ...,
+        startingMonth: int = ...,
+        qtr_with_extra_week: int = ...,
+        variation: Literal["nearest", "last"] = ...,
+    ) -> None: ...
+
+class Easter(SingleConstructorOffset): ...
+
+class _CustomBusinessMonth(BusinessMixin):
+    def __init__(
+        self,
+        n: int = ...,
+        normalize: bool = ...,
+        weekmask: str = ...,
+        holidays: list | None = ...,
+        calendar: OffsetCalendar | None = ...,
+        offset: timedelta = ...,
+    ) -> None: ...
+
+class CustomBusinessDay(BusinessDay):
+    def __init__(
+        self,
+        n: int = ...,
+        normalize: bool = ...,
+        weekmask: str = ...,
+        holidays: list | None = ...,
+        calendar: OffsetCalendar | None = ...,
+        offset: timedelta = ...,
+    ) -> None: ...
+
+class CustomBusinessHour(BusinessHour):
+    def __init__(
+        self,
+        n: int = ...,
+        normalize: bool = ...,
+        weekmask: str = ...,
+        holidays: list | None = ...,
+        calendar: OffsetCalendar | None = ...,
+        start: str | time | Collection[str | time] = ...,
+        end: str | time | Collection[str | time] = ...,
+        offset: timedelta = ...,
+    ) -> None: ...
+
+class CustomBusinessMonthEnd(_CustomBusinessMonth): ...
+class CustomBusinessMonthBegin(_CustomBusinessMonth): ...
+class OffsetMeta(type): ...
+class DateOffset(RelativeDeltaOffset, metaclass=OffsetMeta): ...
+
+BDay = BusinessDay
+BMonthEnd = BusinessMonthEnd
+BMonthBegin = BusinessMonthBegin
+CBMonthEnd = CustomBusinessMonthEnd
+CBMonthBegin = CustomBusinessMonthBegin
+CDay = CustomBusinessDay
+
+def roll_qtrday(
+    other: datetime, n: int, month: int, day_opt: str, modby: int
+) -> int: ...
+
+INVALID_FREQ_ERR_MSG: Literal["Invalid frequency: {0}"]
+
+def shift_months(
+    dtindex: npt.NDArray[np.int64],
+    months: int,
+    day_opt: str | None = ...,
+    reso: int = ...,
+) -> npt.NDArray[np.int64]: ...
+
+_offset_map: dict[str, BaseOffset]

llava_next/lib/python3.10/site-packages/pandas/_libs/tslibs/strptime.pyi

@@ -0,0 +1,14 @@
+import numpy as np
+
+from pandas._typing import npt
+
+def array_strptime(
+    values: npt.NDArray[np.object_],
+    fmt: str | None,
+    exact: bool = ...,
+    errors: str = ...,
+    utc: bool = ...,
+    creso: int = ...,  # NPY_DATETIMEUNIT
+) -> tuple[np.ndarray, np.ndarray]: ...
+
+# first ndarray is M8[ns], second is object ndarray of tzinfo | None

llava_next/lib/python3.10/site-packages/pandas/_libs/tslibs/timedeltas.pyi

@@ -0,0 +1,174 @@
+from datetime import timedelta
+from typing import (
+    ClassVar,
+    Literal,
+    TypeAlias,
+    TypeVar,
+    overload,
+)
+
+import numpy as np
+
+from pandas._libs.tslibs import (
+    NaTType,
+    Tick,
+)
+from pandas._typing import (
+    Frequency,
+    Self,
+    npt,
+)
+
+# This should be kept consistent with the keys in the dict timedelta_abbrevs
+# in pandas/_libs/tslibs/timedeltas.pyx
+UnitChoices: TypeAlias = Literal[
+    "Y",
+    "y",
+    "M",
+    "W",
+    "w",
+    "D",
+    "d",
+    "days",
+    "day",
+    "hours",
+    "hour",
+    "hr",
+    "h",
+    "m",
+    "minute",
+    "min",
+    "minutes",
+    "T",
+    "t",
+    "s",
+    "seconds",
+    "sec",
+    "second",
+    "ms",
+    "milliseconds",
+    "millisecond",
+    "milli",
+    "millis",
+    "L",
+    "l",
+    "us",
+    "microseconds",
+    "microsecond",
+    "µs",
+    "micro",
+    "micros",
+    "u",
+    "ns",
+    "nanoseconds",
+    "nano",
+    "nanos",
+    "nanosecond",
+    "n",
+]
+_S = TypeVar("_S", bound=timedelta)
+
+def get_unit_for_round(freq, creso: int) -> int: ...
+def disallow_ambiguous_unit(unit: str | None) -> None: ...
+def ints_to_pytimedelta(
+    m8values: npt.NDArray[np.timedelta64],
+    box: bool = ...,
+) -> npt.NDArray[np.object_]: ...
+def array_to_timedelta64(
+    values: npt.NDArray[np.object_],
+    unit: str | None = ...,
+    errors: str = ...,
+) -> np.ndarray: ...  # np.ndarray[m8ns]
+def parse_timedelta_unit(unit: str | None) -> UnitChoices: ...
+def delta_to_nanoseconds(
+    delta: np.timedelta64 | timedelta | Tick,
+    reso: int = ...,  # NPY_DATETIMEUNIT
+    round_ok: bool = ...,
+) -> int: ...
+def floordiv_object_array(
+    left: np.ndarray, right: npt.NDArray[np.object_]
+) -> np.ndarray: ...
+def truediv_object_array(
+    left: np.ndarray, right: npt.NDArray[np.object_]
+) -> np.ndarray: ...
+
+class Timedelta(timedelta):
+    _creso: int
+    min: ClassVar[Timedelta]
+    max: ClassVar[Timedelta]
+    resolution: ClassVar[Timedelta]
+    value: int  # np.int64
+    _value: int  # np.int64
+    # error: "__new__" must return a class instance (got "Union[Timestamp, NaTType]")
+    def __new__(  # type: ignore[misc]
+        cls: type[_S],
+        value=...,
+        unit: str | None = ...,
+        **kwargs: float | np.integer | np.floating,
+    ) -> _S | NaTType: ...
+    @classmethod
+    def _from_value_and_reso(cls, value: np.int64, reso: int) -> Timedelta: ...
+    @property
+    def days(self) -> int: ...
+    @property
+    def seconds(self) -> int: ...
+    @property
+    def microseconds(self) -> int: ...
+    def total_seconds(self) -> float: ...
+    def to_pytimedelta(self) -> timedelta: ...
+    def to_timedelta64(self) -> np.timedelta64: ...
+    @property
+    def asm8(self) -> np.timedelta64: ...
+    # TODO: round/floor/ceil could return NaT?
+    def round(self, freq: Frequency) -> Self: ...
+    def floor(self, freq: Frequency) -> Self: ...
+    def ceil(self, freq: Frequency) -> Self: ...
+    @property
+    def resolution_string(self) -> str: ...
+    def __add__(self, other: timedelta) -> Timedelta: ...
+    def __radd__(self, other: timedelta) -> Timedelta: ...
+    def __sub__(self, other: timedelta) -> Timedelta: ...
+    def __rsub__(self, other: timedelta) -> Timedelta: ...
+    def __neg__(self) -> Timedelta: ...
+    def __pos__(self) -> Timedelta: ...
+    def __abs__(self) -> Timedelta: ...
+    def __mul__(self, other: float) -> Timedelta: ...
+    def __rmul__(self, other: float) -> Timedelta: ...
+    # error: Signature of "__floordiv__" incompatible with supertype "timedelta"
+    @overload  # type: ignore[override]
+    def __floordiv__(self, other: timedelta) -> int: ...
+    @overload
+    def __floordiv__(self, other: float) -> Timedelta: ...
+    @overload
+    def __floordiv__(
+        self, other: npt.NDArray[np.timedelta64]
+    ) -> npt.NDArray[np.intp]: ...
+    @overload
+    def __floordiv__(
+        self, other: npt.NDArray[np.number]
+    ) -> npt.NDArray[np.timedelta64] | Timedelta: ...
+    @overload
+    def __rfloordiv__(self, other: timedelta | str) -> int: ...
+    @overload
+    def __rfloordiv__(self, other: None | NaTType) -> NaTType: ...
+    @overload
+    def __rfloordiv__(self, other: np.ndarray) -> npt.NDArray[np.timedelta64]: ...
+    @overload
+    def __truediv__(self, other: timedelta) -> float: ...
+    @overload
+    def __truediv__(self, other: float) -> Timedelta: ...
+    def __mod__(self, other: timedelta) -> Timedelta: ...
+    def __divmod__(self, other: timedelta) -> tuple[int, Timedelta]: ...
+    def __le__(self, other: timedelta) -> bool: ...
+    def __lt__(self, other: timedelta) -> bool: ...
+    def __ge__(self, other: timedelta) -> bool: ...
+    def __gt__(self, other: timedelta) -> bool: ...
+    def __hash__(self) -> int: ...
+    def isoformat(self) -> str: ...
+    def to_numpy(
+        self, dtype: npt.DTypeLike = ..., copy: bool = False
+    ) -> np.timedelta64: ...
+    def view(self, dtype: npt.DTypeLike) -> object: ...
+    @property
+    def unit(self) -> str: ...
+    def as_unit(self, unit: str, round_ok: bool = ...) -> Timedelta: ...

llava_next/lib/python3.10/site-packages/pandas/_libs/tslibs/timestamps.pyi

@@ -0,0 +1,241 @@
+from datetime import (
+    date as _date,
+    datetime,
+    time as _time,
+    timedelta,
+    tzinfo as _tzinfo,
+)
+from time import struct_time
+from typing import (
+    ClassVar,
+    Literal,
+    TypeAlias,
+    overload,
+)
+
+import numpy as np
+
+from pandas._libs.tslibs import (
+    BaseOffset,
+    NaTType,
+    Period,
+    Tick,
+    Timedelta,
+)
+from pandas._typing import (
+    Self,
+    TimestampNonexistent,
+)
+
+_TimeZones: TypeAlias = str | _tzinfo | None | int
+
+def integer_op_not_supported(obj: object) -> TypeError: ...
+
+class Timestamp(datetime):
+    _creso: int
+    min: ClassVar[Timestamp]
+    max: ClassVar[Timestamp]
+
+    resolution: ClassVar[Timedelta]
+    _value: int  # np.int64
+    # error: "__new__" must return a class instance (got "Union[Timestamp, NaTType]")
+    def __new__(  # type: ignore[misc]
+        cls: type[Self],
+        ts_input: np.integer | float | str | _date | datetime | np.datetime64 = ...,
+        year: int | None = ...,
+        month: int | None = ...,
+        day: int | None = ...,
+        hour: int | None = ...,
+        minute: int | None = ...,
+        second: int | None = ...,
+        microsecond: int | None = ...,
+        tzinfo: _tzinfo | None = ...,
+        *,
+        nanosecond: int | None = ...,
+        tz: _TimeZones = ...,
+        unit: str | int | None = ...,
+        fold: int | None = ...,
+    ) -> Self | NaTType: ...
+    @classmethod
+    def _from_value_and_reso(
+        cls, value: int, reso: int, tz: _TimeZones
+    ) -> Timestamp: ...
+    @property
+    def value(self) -> int: ...  # np.int64
+    @property
+    def year(self) -> int: ...
+    @property
+    def month(self) -> int: ...
+    @property
+    def day(self) -> int: ...
+    @property
+    def hour(self) -> int: ...
+    @property
+    def minute(self) -> int: ...
+    @property
+    def second(self) -> int: ...
+    @property
+    def microsecond(self) -> int: ...
+    @property
+    def nanosecond(self) -> int: ...
+    @property
+    def tzinfo(self) -> _tzinfo | None: ...
+    @property
+    def tz(self) -> _tzinfo | None: ...
+    @property
+    def fold(self) -> int: ...
+    @classmethod
+    def fromtimestamp(cls, ts: float, tz: _TimeZones = ...) -> Self: ...
+    @classmethod
+    def utcfromtimestamp(cls, ts: float) -> Self: ...
+    @classmethod
+    def today(cls, tz: _TimeZones = ...) -> Self: ...
+    @classmethod
+    def fromordinal(
+        cls,
+        ordinal: int,
+        tz: _TimeZones = ...,
+    ) -> Self: ...
+    @classmethod
+    def now(cls, tz: _TimeZones = ...) -> Self: ...
+    @classmethod
+    def utcnow(cls) -> Self: ...
+    # error: Signature of "combine" incompatible with supertype "datetime"
+    @classmethod
+    def combine(  # type: ignore[override]
+        cls, date: _date, time: _time
+    ) -> datetime: ...
+    @classmethod
+    def fromisoformat(cls, date_string: str) -> Self: ...
+    def strftime(self, format: str) -> str: ...
+    def __format__(self, fmt: str) -> str: ...
+    def toordinal(self) -> int: ...
+    def timetuple(self) -> struct_time: ...
+    def timestamp(self) -> float: ...
+    def utctimetuple(self) -> struct_time: ...
+    def date(self) -> _date: ...
+    def time(self) -> _time: ...
+    def timetz(self) -> _time: ...
+    # LSP violation: nanosecond is not present in datetime.datetime.replace
+    # and has positional args following it
+    def replace(  # type: ignore[override]
+        self,
+        year: int | None = ...,
+        month: int | None = ...,
+        day: int | None = ...,
+        hour: int | None = ...,
+        minute: int | None = ...,
+        second: int | None = ...,
+        microsecond: int | None = ...,
+        nanosecond: int | None = ...,
+        tzinfo: _tzinfo | type[object] | None = ...,
+        fold: int | None = ...,
+    ) -> Self: ...
+    # LSP violation: datetime.datetime.astimezone has a default value for tz
+    def astimezone(self, tz: _TimeZones) -> Self: ...  # type: ignore[override]
+    def ctime(self) -> str: ...
+    def isoformat(self, sep: str = ..., timespec: str = ...) -> str: ...
+    @classmethod
+    def strptime(
+        # Note: strptime is actually disabled and raises NotImplementedError
+        cls,
+        date_string: str,
+        format: str,
+    ) -> Self: ...
+    def utcoffset(self) -> timedelta | None: ...
+    def tzname(self) -> str | None: ...
+    def dst(self) -> timedelta | None: ...
+    def __le__(self, other: datetime) -> bool: ...  # type: ignore[override]
+    def __lt__(self, other: datetime) -> bool: ...  # type: ignore[override]
+    def __ge__(self, other: datetime) -> bool: ...  # type: ignore[override]
+    def __gt__(self, other: datetime) -> bool: ...  # type: ignore[override]
+    # error: Signature of "__add__" incompatible with supertype "date"/"datetime"
+    @overload  # type: ignore[override]
+    def __add__(self, other: np.ndarray) -> np.ndarray: ...
+    @overload
+    def __add__(self, other: timedelta | np.timedelta64 | Tick) -> Self: ...
+    def __radd__(self, other: timedelta) -> Self: ...
+    @overload  # type: ignore[override]
+    def __sub__(self, other: datetime) -> Timedelta: ...
+    @overload
+    def __sub__(self, other: timedelta | np.timedelta64 | Tick) -> Self: ...
+    def __hash__(self) -> int: ...
+    def weekday(self) -> int: ...
+    def isoweekday(self) -> int: ...
+    # Return type "Tuple[int, int, int]" of "isocalendar" incompatible with return
+    # type "_IsoCalendarDate" in supertype "date"
+    def isocalendar(self) -> tuple[int, int, int]: ...  # type: ignore[override]
+    @property
+    def is_leap_year(self) -> bool: ...
+    @property
+    def is_month_start(self) -> bool: ...
+    @property
+    def is_quarter_start(self) -> bool: ...
+    @property
+    def is_year_start(self) -> bool: ...
+    @property
+    def is_month_end(self) -> bool: ...
+    @property
+    def is_quarter_end(self) -> bool: ...
+    @property
+    def is_year_end(self) -> bool: ...
+    def to_pydatetime(self, warn: bool = ...) -> datetime: ...
+    def to_datetime64(self) -> np.datetime64: ...
+    def to_period(self, freq: BaseOffset | str | None = None) -> Period: ...
+    def to_julian_date(self) -> np.float64: ...
+    @property
+    def asm8(self) -> np.datetime64: ...
+    def tz_convert(self, tz: _TimeZones) -> Self: ...
+    # TODO: could return NaT?
+    def tz_localize(
+        self,
+        tz: _TimeZones,
+        ambiguous: bool | Literal["raise", "NaT"] = ...,
+        nonexistent: TimestampNonexistent = ...,
+    ) -> Self: ...
+    def normalize(self) -> Self: ...
+    # TODO: round/floor/ceil could return NaT?
+    def round(
+        self,
+        freq: str,
+        ambiguous: bool | Literal["raise", "NaT"] = ...,
+        nonexistent: TimestampNonexistent = ...,
+    ) -> Self: ...
+    def floor(
+        self,
+        freq: str,
+        ambiguous: bool | Literal["raise", "NaT"] = ...,
+        nonexistent: TimestampNonexistent = ...,
+    ) -> Self: ...
+    def ceil(
+        self,
+        freq: str,
+        ambiguous: bool | Literal["raise", "NaT"] = ...,
+        nonexistent: TimestampNonexistent = ...,
+    ) -> Self: ...
+    def day_name(self, locale: str | None = ...) -> str: ...
+    def month_name(self, locale: str | None = ...) -> str: ...
+    @property
+    def day_of_week(self) -> int: ...
+    @property
+    def dayofweek(self) -> int: ...
+    @property
+    def day_of_year(self) -> int: ...
+    @property
+    def dayofyear(self) -> int: ...
+    @property
+    def quarter(self) -> int: ...
+    @property
+    def week(self) -> int: ...
+    def to_numpy(
+        self, dtype: np.dtype | None = ..., copy: bool = ...
+    ) -> np.datetime64: ...
+    @property
+    def _date_repr(self) -> str: ...
+    @property
+    def days_in_month(self) -> int: ...
+    @property
+    def daysinmonth(self) -> int: ...
+    @property
+    def unit(self) -> str: ...
+    def as_unit(self, unit: str, round_ok: bool = ...) -> Timestamp: ...

llava_next/lib/python3.10/site-packages/pandas/core/apply.py

@@ -0,0 +1,2062 @@
+from __future__ import annotations
+
+import abc
+from collections import defaultdict
+import functools
+from functools import partial
+import inspect
+from typing import (
+    TYPE_CHECKING,
+    Any,
+    Callable,
+    Literal,
+    cast,
+)
+import warnings
+
+import numpy as np
+
+from pandas._config import option_context
+
+from pandas._libs import lib
+from pandas._libs.internals import BlockValuesRefs
+from pandas._typing import (
+    AggFuncType,
+    AggFuncTypeBase,
+    AggFuncTypeDict,
+    AggObjType,
+    Axis,
+    AxisInt,
+    NDFrameT,
+    npt,
+)
+from pandas.compat._optional import import_optional_dependency
+from pandas.errors import SpecificationError
+from pandas.util._decorators import cache_readonly
+from pandas.util._exceptions import find_stack_level
+
+from pandas.core.dtypes.cast import is_nested_object
+from pandas.core.dtypes.common import (
+    is_dict_like,
+    is_extension_array_dtype,
+    is_list_like,
+    is_numeric_dtype,
+    is_sequence,
+)
+from pandas.core.dtypes.dtypes import (
+    CategoricalDtype,
+    ExtensionDtype,
+)
+from pandas.core.dtypes.generic import (
+    ABCDataFrame,
+    ABCNDFrame,
+    ABCSeries,
+)
+
+from pandas.core._numba.executor import generate_apply_looper
+import pandas.core.common as com
+from pandas.core.construction import ensure_wrapped_if_datetimelike
+
+if TYPE_CHECKING:
+    from collections.abc import (
+        Generator,
+        Hashable,
+        Iterable,
+        MutableMapping,
+        Sequence,
+    )
+
+    from pandas import (
+        DataFrame,
+        Index,
+        Series,
+    )
+    from pandas.core.groupby import GroupBy
+    from pandas.core.resample import Resampler
+    from pandas.core.window.rolling import BaseWindow
+
+
+ResType = dict[int, Any]
+
+
+def frame_apply(
+    obj: DataFrame,
+    func: AggFuncType,
+    axis: Axis = 0,
+    raw: bool = False,
+    result_type: str | None = None,
+    by_row: Literal[False, "compat"] = "compat",
+    engine: str = "python",
+    engine_kwargs: dict[str, bool] | None = None,
+    args=None,
+    kwargs=None,
+) -> FrameApply:
+    """construct and return a row or column based frame apply object"""
+    axis = obj._get_axis_number(axis)
+    klass: type[FrameApply]
+    if axis == 0:
+        klass = FrameRowApply
+    elif axis == 1:
+        klass = FrameColumnApply
+
+    _, func, _, _ = reconstruct_func(func, **kwargs)
+    assert func is not None
+
+    return klass(
+        obj,
+        func,
+        raw=raw,
+        result_type=result_type,
+        by_row=by_row,
+        engine=engine,
+        engine_kwargs=engine_kwargs,
+        args=args,
+        kwargs=kwargs,
+    )
+
+
+class Apply(metaclass=abc.ABCMeta):
+    axis: AxisInt
+
+    def __init__(
+        self,
+        obj: AggObjType,
+        func: AggFuncType,
+        raw: bool,
+        result_type: str | None,
+        *,
+        by_row: Literal[False, "compat", "_compat"] = "compat",
+        engine: str = "python",
+        engine_kwargs: dict[str, bool] | None = None,
+        args,
+        kwargs,
+    ) -> None:
+        self.obj = obj
+        self.raw = raw
+
+        assert by_row is False or by_row in ["compat", "_compat"]
+        self.by_row = by_row
+
+        self.args = args or ()
+        self.kwargs = kwargs or {}
+
+        self.engine = engine
+        self.engine_kwargs = {} if engine_kwargs is None else engine_kwargs
+
+        if result_type not in [None, "reduce", "broadcast", "expand"]:
+            raise ValueError(
+                "invalid value for result_type, must be one "
+                "of {None, 'reduce', 'broadcast', 'expand'}"
+            )
+
+        self.result_type = result_type
+
+        self.func = func
+
+    @abc.abstractmethod
+    def apply(self) -> DataFrame | Series:
+        pass
+
+    @abc.abstractmethod
+    def agg_or_apply_list_like(
+        self, op_name: Literal["agg", "apply"]
+    ) -> DataFrame | Series:
+        pass
+
+    @abc.abstractmethod
+    def agg_or_apply_dict_like(
+        self, op_name: Literal["agg", "apply"]
+    ) -> DataFrame | Series:
+        pass
+
+    def agg(self) -> DataFrame | Series | None:
+        """
+        Provide an implementation for the aggregators.
+
+        Returns
+        -------
+        Result of aggregation, or None if agg cannot be performed by
+        this method.
+        """
+        obj = self.obj
+        func = self.func
+        args = self.args
+        kwargs = self.kwargs
+
+        if isinstance(func, str):
+            return self.apply_str()
+
+        if is_dict_like(func):
+            return self.agg_dict_like()
+        elif is_list_like(func):
+            # we require a list, but not a 'str'
+            return self.agg_list_like()
+
+        if callable(func):
+            f = com.get_cython_func(func)
+            if f and not args and not kwargs:
+                warn_alias_replacement(obj, func, f)
+                return getattr(obj, f)()
+
+        # caller can react
+        return None
+
+    def transform(self) -> DataFrame | Series:
+        """
+        Transform a DataFrame or Series.
+
+        Returns
+        -------
+        DataFrame or Series
+            Result of applying ``func`` along the given axis of the
+            Series or DataFrame.
+
+        Raises
+        ------
+        ValueError
+            If the transform function fails or does not transform.
+        """
+        obj = self.obj
+        func = self.func
+        axis = self.axis
+        args = self.args
+        kwargs = self.kwargs
+
+        is_series = obj.ndim == 1
+
+        if obj._get_axis_number(axis) == 1:
+            assert not is_series
+            return obj.T.transform(func, 0, *args, **kwargs).T
+
+        if is_list_like(func) and not is_dict_like(func):
+            func = cast(list[AggFuncTypeBase], func)
+            # Convert func equivalent dict
+            if is_series:
+                func = {com.get_callable_name(v) or v: v for v in func}
+            else:
+                func = {col: func for col in obj}
+
+        if is_dict_like(func):
+            func = cast(AggFuncTypeDict, func)
+            return self.transform_dict_like(func)
+
+        # func is either str or callable
+        func = cast(AggFuncTypeBase, func)
+        try:
+            result = self.transform_str_or_callable(func)
+        except TypeError:
+            raise
+        except Exception as err:
+            raise ValueError("Transform function failed") from err
+
+        # Functions that transform may return empty Series/DataFrame
+        # when the dtype is not appropriate
+        if (
+            isinstance(result, (ABCSeries, ABCDataFrame))
+            and result.empty
+            and not obj.empty
+        ):
+            raise ValueError("Transform function failed")
+        # error: Argument 1 to "__get__" of "AxisProperty" has incompatible type
+        # "Union[Series, DataFrame, GroupBy[Any], SeriesGroupBy,
+        # DataFrameGroupBy, BaseWindow, Resampler]"; expected "Union[DataFrame,
+        # Series]"
+        if not isinstance(result, (ABCSeries, ABCDataFrame)) or not result.index.equals(
+            obj.index  # type: ignore[arg-type]
+        ):
+            raise ValueError("Function did not transform")
+
+        return result
+
+    def transform_dict_like(self, func) -> DataFrame:
+        """
+        Compute transform in the case of a dict-like func
+        """
+        from pandas.core.reshape.concat import concat
+
+        obj = self.obj
+        args = self.args
+        kwargs = self.kwargs
+
+        # transform is currently only for Series/DataFrame
+        assert isinstance(obj, ABCNDFrame)
+
+        if len(func) == 0:
+            raise ValueError("No transform functions were provided")
+
+        func = self.normalize_dictlike_arg("transform", obj, func)
+
+        results: dict[Hashable, DataFrame | Series] = {}
+        for name, how in func.items():
+            colg = obj._gotitem(name, ndim=1)
+            results[name] = colg.transform(how, 0, *args, **kwargs)
+        return concat(results, axis=1)
+
+    def transform_str_or_callable(self, func) -> DataFrame | Series:
+        """
+        Compute transform in the case of a string or callable func
+        """
+        obj = self.obj
+        args = self.args
+        kwargs = self.kwargs
+
+        if isinstance(func, str):
+            return self._apply_str(obj, func, *args, **kwargs)
+
+        if not args and not kwargs:
+            f = com.get_cython_func(func)
+            if f:
+                warn_alias_replacement(obj, func, f)
+                return getattr(obj, f)()
+
+        # Two possible ways to use a UDF - apply or call directly
+        try:
+            return obj.apply(func, args=args, **kwargs)
+        except Exception:
+            return func(obj, *args, **kwargs)
+
+    def agg_list_like(self) -> DataFrame | Series:
+        """
+        Compute aggregation in the case of a list-like argument.
+
+        Returns
+        -------
+        Result of aggregation.
+        """
+        return self.agg_or_apply_list_like(op_name="agg")
+
+    def compute_list_like(
+        self,
+        op_name: Literal["agg", "apply"],
+        selected_obj: Series | DataFrame,
+        kwargs: dict[str, Any],
+    ) -> tuple[list[Hashable] | Index, list[Any]]:
+        """
+        Compute agg/apply results for like-like input.
+
+        Parameters
+        ----------
+        op_name : {"agg", "apply"}
+            Operation being performed.
+        selected_obj : Series or DataFrame
+            Data to perform operation on.
+        kwargs : dict
+            Keyword arguments to pass to the functions.
+
+        Returns
+        -------
+        keys : list[Hashable] or Index
+            Index labels for result.
+        results : list
+            Data for result. When aggregating with a Series, this can contain any
+            Python objects.
+        """
+        func = cast(list[AggFuncTypeBase], self.func)
+        obj = self.obj
+
+        results = []
+        keys = []
+
+        # degenerate case
+        if selected_obj.ndim == 1:
+            for a in func:
+                colg = obj._gotitem(selected_obj.name, ndim=1, subset=selected_obj)
+                args = (
+                    [self.axis, *self.args]
+                    if include_axis(op_name, colg)
+                    else self.args
+                )
+                new_res = getattr(colg, op_name)(a, *args, **kwargs)
+                results.append(new_res)
+
+                # make sure we find a good name
+                name = com.get_callable_name(a) or a
+                keys.append(name)
+
+        else:
+            indices = []
+            for index, col in enumerate(selected_obj):
+                colg = obj._gotitem(col, ndim=1, subset=selected_obj.iloc[:, index])
+                args = (
+                    [self.axis, *self.args]
+                    if include_axis(op_name, colg)
+                    else self.args
+                )
+                new_res = getattr(colg, op_name)(func, *args, **kwargs)
+                results.append(new_res)
+                indices.append(index)
+            # error: Incompatible types in assignment (expression has type "Any |
+            # Index", variable has type "list[Any | Callable[..., Any] | str]")
+            keys = selected_obj.columns.take(indices)  # type: ignore[assignment]
+
+        return keys, results
+
+    def wrap_results_list_like(
+        self, keys: Iterable[Hashable], results: list[Series | DataFrame]
+    ):
+        from pandas.core.reshape.concat import concat
+
+        obj = self.obj
+
+        try:
+            return concat(results, keys=keys, axis=1, sort=False)
+        except TypeError as err:
+            # we are concatting non-NDFrame objects,
+            # e.g. a list of scalars
+            from pandas import Series
+
+            result = Series(results, index=keys, name=obj.name)
+            if is_nested_object(result):
+                raise ValueError(
+                    "cannot combine transform and aggregation operations"
+                ) from err
+            return result
+
+    def agg_dict_like(self) -> DataFrame | Series:
+        """
+        Compute aggregation in the case of a dict-like argument.
+
+        Returns
+        -------
+        Result of aggregation.
+        """
+        return self.agg_or_apply_dict_like(op_name="agg")
+
+    def compute_dict_like(
+        self,
+        op_name: Literal["agg", "apply"],
+        selected_obj: Series | DataFrame,
+        selection: Hashable | Sequence[Hashable],
+        kwargs: dict[str, Any],
+    ) -> tuple[list[Hashable], list[Any]]:
+        """
+        Compute agg/apply results for dict-like input.
+
+        Parameters
+        ----------
+        op_name : {"agg", "apply"}
+            Operation being performed.
+        selected_obj : Series or DataFrame
+            Data to perform operation on.
+        selection : hashable or sequence of hashables
+            Used by GroupBy, Window, and Resample if selection is applied to the object.
+        kwargs : dict
+            Keyword arguments to pass to the functions.
+
+        Returns
+        -------
+        keys : list[hashable]
+            Index labels for result.
+        results : list
+            Data for result. When aggregating with a Series, this can contain any
+            Python object.
+        """
+        from pandas.core.groupby.generic import (
+            DataFrameGroupBy,
+            SeriesGroupBy,
+        )
+
+        obj = self.obj
+        is_groupby = isinstance(obj, (DataFrameGroupBy, SeriesGroupBy))
+        func = cast(AggFuncTypeDict, self.func)
+        func = self.normalize_dictlike_arg(op_name, selected_obj, func)
+
+        is_non_unique_col = (
+            selected_obj.ndim == 2
+            and selected_obj.columns.nunique() < len(selected_obj.columns)
+        )
+
+        if selected_obj.ndim == 1:
+            # key only used for output
+            colg = obj._gotitem(selection, ndim=1)
+            results = [getattr(colg, op_name)(how, **kwargs) for _, how in func.items()]
+            keys = list(func.keys())
+        elif not is_groupby and is_non_unique_col:
+            # key used for column selection and output
+            # GH#51099
+            results = []
+            keys = []
+            for key, how in func.items():
+                indices = selected_obj.columns.get_indexer_for([key])
+                labels = selected_obj.columns.take(indices)
+                label_to_indices = defaultdict(list)
+                for index, label in zip(indices, labels):
+                    label_to_indices[label].append(index)
+
+                key_data = [
+                    getattr(selected_obj._ixs(indice, axis=1), op_name)(how, **kwargs)
+                    for label, indices in label_to_indices.items()
+                    for indice in indices
+                ]
+
+                keys += [key] * len(key_data)
+                results += key_data
+        else:
+            # key used for column selection and output
+            results = [
+                getattr(obj._gotitem(key, ndim=1), op_name)(how, **kwargs)
+                for key, how in func.items()
+            ]
+            keys = list(func.keys())
+
+        return keys, results
+
+    def wrap_results_dict_like(
+        self,
+        selected_obj: Series | DataFrame,
+        result_index: list[Hashable],
+        result_data: list,
+    ):
+        from pandas import Index
+        from pandas.core.reshape.concat import concat
+
+        obj = self.obj
+
+        # Avoid making two isinstance calls in all and any below
+        is_ndframe = [isinstance(r, ABCNDFrame) for r in result_data]
+
+        if all(is_ndframe):
+            results = dict(zip(result_index, result_data))
+            keys_to_use: Iterable[Hashable]
+            keys_to_use = [k for k in result_index if not results[k].empty]
+            # Have to check, if at least one DataFrame is not empty.
+            keys_to_use = keys_to_use if keys_to_use != [] else result_index
+            if selected_obj.ndim == 2:
+                # keys are columns, so we can preserve names
+                ktu = Index(keys_to_use)
+                ktu._set_names(selected_obj.columns.names)
+                keys_to_use = ktu
+
+            axis: AxisInt = 0 if isinstance(obj, ABCSeries) else 1
+            result = concat(
+                {k: results[k] for k in keys_to_use},
+                axis=axis,
+                keys=keys_to_use,
+            )
+        elif any(is_ndframe):
+            # There is a mix of NDFrames and scalars
+            raise ValueError(
+                "cannot perform both aggregation "
+                "and transformation operations "
+                "simultaneously"
+            )
+        else:
+            from pandas import Series
+
+            # we have a list of scalars
+            # GH 36212 use name only if obj is a series
+            if obj.ndim == 1:
+                obj = cast("Series", obj)
+                name = obj.name
+            else:
+                name = None
+
+            result = Series(result_data, index=result_index, name=name)
+
+        return result
+
+    def apply_str(self) -> DataFrame | Series:
+        """
+        Compute apply in case of a string.
+
+        Returns
+        -------
+        result: Series or DataFrame
+        """
+        # Caller is responsible for checking isinstance(self.f, str)
+        func = cast(str, self.func)
+
+        obj = self.obj
+
+        from pandas.core.groupby.generic import (
+            DataFrameGroupBy,
+            SeriesGroupBy,
+        )
+
+        # Support for `frame.transform('method')`
+        # Some methods (shift, etc.) require the axis argument, others
+        # don't, so inspect and insert if necessary.
+        method = getattr(obj, func, None)
+        if callable(method):
+            sig = inspect.getfullargspec(method)
+            arg_names = (*sig.args, *sig.kwonlyargs)
+            if self.axis != 0 and (
+                "axis" not in arg_names or func in ("corrwith", "skew")
+            ):
+                raise ValueError(f"Operation {func} does not support axis=1")
+            if "axis" in arg_names:
+                if isinstance(obj, (SeriesGroupBy, DataFrameGroupBy)):
+                    # Try to avoid FutureWarning for deprecated axis keyword;
+                    # If self.axis matches the axis we would get by not passing
+                    #  axis, we safely exclude the keyword.
+
+                    default_axis = 0
+                    if func in ["idxmax", "idxmin"]:
+                        # DataFrameGroupBy.idxmax, idxmin axis defaults to self.axis,
+                        # whereas other axis keywords default to 0
+                        default_axis = self.obj.axis
+
+                    if default_axis != self.axis:
+                        self.kwargs["axis"] = self.axis
+                else:
+                    self.kwargs["axis"] = self.axis
+        return self._apply_str(obj, func, *self.args, **self.kwargs)
+
+    def apply_list_or_dict_like(self) -> DataFrame | Series:
+        """
+        Compute apply in case of a list-like or dict-like.
+
+        Returns
+        -------
+        result: Series, DataFrame, or None
+            Result when self.func is a list-like or dict-like, None otherwise.
+        """
+
+        if self.engine == "numba":
+            raise NotImplementedError(
+                "The 'numba' engine doesn't support list-like/"
+                "dict likes of callables yet."
+            )
+
+        if self.axis == 1 and isinstance(self.obj, ABCDataFrame):
+            return self.obj.T.apply(self.func, 0, args=self.args, **self.kwargs).T
+
+        func = self.func
+        kwargs = self.kwargs
+
+        if is_dict_like(func):
+            result = self.agg_or_apply_dict_like(op_name="apply")
+        else:
+            result = self.agg_or_apply_list_like(op_name="apply")
+
+        result = reconstruct_and_relabel_result(result, func, **kwargs)
+
+        return result
+
+    def normalize_dictlike_arg(
+        self, how: str, obj: DataFrame | Series, func: AggFuncTypeDict
+    ) -> AggFuncTypeDict:
+        """
+        Handler for dict-like argument.
+
+        Ensures that necessary columns exist if obj is a DataFrame, and
+        that a nested renamer is not passed. Also normalizes to all lists
+        when values consists of a mix of list and non-lists.
+        """
+        assert how in ("apply", "agg", "transform")
+
+        # Can't use func.values(); wouldn't work for a Series
+        if (
+            how == "agg"
+            and isinstance(obj, ABCSeries)
+            and any(is_list_like(v) for _, v in func.items())
+        ) or (any(is_dict_like(v) for _, v in func.items())):
+            # GH 15931 - deprecation of renaming keys
+            raise SpecificationError("nested renamer is not supported")
+
+        if obj.ndim != 1:
+            # Check for missing columns on a frame
+            from pandas import Index
+
+            cols = Index(list(func.keys())).difference(obj.columns, sort=True)
+            if len(cols) > 0:
+                raise KeyError(f"Column(s) {list(cols)} do not exist")
+
+        aggregator_types = (list, tuple, dict)
+
+        # if we have a dict of any non-scalars
+        # eg. {'A' : ['mean']}, normalize all to
+        # be list-likes
+        # Cannot use func.values() because arg may be a Series
+        if any(isinstance(x, aggregator_types) for _, x in func.items()):
+            new_func: AggFuncTypeDict = {}
+            for k, v in func.items():
+                if not isinstance(v, aggregator_types):
+                    new_func[k] = [v]
+                else:
+                    new_func[k] = v
+            func = new_func
+        return func
+
+    def _apply_str(self, obj, func: str, *args, **kwargs):
+        """
+        if arg is a string, then try to operate on it:
+        - try to find a function (or attribute) on obj
+        - try to find a numpy function
+        - raise
+        """
+        assert isinstance(func, str)
+
+        if hasattr(obj, func):
+            f = getattr(obj, func)
+            if callable(f):
+                return f(*args, **kwargs)
+
+            # people may aggregate on a non-callable attribute
+            # but don't let them think they can pass args to it
+            assert len(args) == 0
+            assert len([kwarg for kwarg in kwargs if kwarg not in ["axis"]]) == 0
+            return f
+        elif hasattr(np, func) and hasattr(obj, "__array__"):
+            # in particular exclude Window
+            f = getattr(np, func)
+            return f(obj, *args, **kwargs)
+        else:
+            msg = f"'{func}' is not a valid function for '{type(obj).__name__}' object"
+            raise AttributeError(msg)
+
+
+class NDFrameApply(Apply):
+    """
+    Methods shared by FrameApply and SeriesApply but
+    not GroupByApply or ResamplerWindowApply
+    """
+
+    obj: DataFrame | Series
+
+    @property
+    def index(self) -> Index:
+        return self.obj.index
+
+    @property
+    def agg_axis(self) -> Index:
+        return self.obj._get_agg_axis(self.axis)
+
+    def agg_or_apply_list_like(
+        self, op_name: Literal["agg", "apply"]
+    ) -> DataFrame | Series:
+        obj = self.obj
+        kwargs = self.kwargs
+
+        if op_name == "apply":
+            if isinstance(self, FrameApply):
+                by_row = self.by_row
+
+            elif isinstance(self, SeriesApply):
+                by_row = "_compat" if self.by_row else False
+            else:
+                by_row = False
+            kwargs = {**kwargs, "by_row": by_row}
+
+        if getattr(obj, "axis", 0) == 1:
+            raise NotImplementedError("axis other than 0 is not supported")
+
+        keys, results = self.compute_list_like(op_name, obj, kwargs)
+        result = self.wrap_results_list_like(keys, results)
+        return result
+
+    def agg_or_apply_dict_like(
+        self, op_name: Literal["agg", "apply"]
+    ) -> DataFrame | Series:
+        assert op_name in ["agg", "apply"]
+        obj = self.obj
+
+        kwargs = {}
+        if op_name == "apply":
+            by_row = "_compat" if self.by_row else False
+            kwargs.update({"by_row": by_row})
+
+        if getattr(obj, "axis", 0) == 1:
+            raise NotImplementedError("axis other than 0 is not supported")
+
+        selection = None
+        result_index, result_data = self.compute_dict_like(
+            op_name, obj, selection, kwargs
+        )
+        result = self.wrap_results_dict_like(obj, result_index, result_data)
+        return result
+
+
+class FrameApply(NDFrameApply):
+    obj: DataFrame
+
+    def __init__(
+        self,
+        obj: AggObjType,
+        func: AggFuncType,
+        raw: bool,
+        result_type: str | None,
+        *,
+        by_row: Literal[False, "compat"] = False,
+        engine: str = "python",
+        engine_kwargs: dict[str, bool] | None = None,
+        args,
+        kwargs,
+    ) -> None:
+        if by_row is not False and by_row != "compat":
+            raise ValueError(f"by_row={by_row} not allowed")
+        super().__init__(
+            obj,
+            func,
+            raw,
+            result_type,
+            by_row=by_row,
+            engine=engine,
+            engine_kwargs=engine_kwargs,
+            args=args,
+            kwargs=kwargs,
+        )
+
+    # ---------------------------------------------------------------
+    # Abstract Methods
+
+    @property
+    @abc.abstractmethod
+    def result_index(self) -> Index:
+        pass
+
+    @property
+    @abc.abstractmethod
+    def result_columns(self) -> Index:
+        pass
+
+    @property
+    @abc.abstractmethod
+    def series_generator(self) -> Generator[Series, None, None]:
+        pass
+
+    @staticmethod
+    @functools.cache
+    @abc.abstractmethod
+    def generate_numba_apply_func(
+        func, nogil=True, nopython=True, parallel=False
+    ) -> Callable[[npt.NDArray, Index, Index], dict[int, Any]]:
+        pass
+
+    @abc.abstractmethod
+    def apply_with_numba(self):
+        pass
+
+    def validate_values_for_numba(self):
+        # Validate column dtyps all OK
+        for colname, dtype in self.obj.dtypes.items():
+            if not is_numeric_dtype(dtype):
+                raise ValueError(
+                    f"Column {colname} must have a numeric dtype. "
+                    f"Found '{dtype}' instead"
+                )
+            if is_extension_array_dtype(dtype):
+                raise ValueError(
+                    f"Column {colname} is backed by an extension array, "
+                    f"which is not supported by the numba engine."
+                )
+
+    @abc.abstractmethod
+    def wrap_results_for_axis(
+        self, results: ResType, res_index: Index
+    ) -> DataFrame | Series:
+        pass
+
+    # ---------------------------------------------------------------
+
+    @property
+    def res_columns(self) -> Index:
+        return self.result_columns
+
+    @property
+    def columns(self) -> Index:
+        return self.obj.columns
+
+    @cache_readonly
+    def values(self):
+        return self.obj.values
+
+    def apply(self) -> DataFrame | Series:
+        """compute the results"""
+
+        # dispatch to handle list-like or dict-like
+        if is_list_like(self.func):
+            if self.engine == "numba":
+                raise NotImplementedError(
+                    "the 'numba' engine doesn't support lists of callables yet"
+                )
+            return self.apply_list_or_dict_like()
+
+        # all empty
+        if len(self.columns) == 0 and len(self.index) == 0:
+            return self.apply_empty_result()
+
+        # string dispatch
+        if isinstance(self.func, str):
+            if self.engine == "numba":
+                raise NotImplementedError(
+                    "the 'numba' engine doesn't support using "
+                    "a string as the callable function"
+                )
+            return self.apply_str()
+
+        # ufunc
+        elif isinstance(self.func, np.ufunc):
+            if self.engine == "numba":
+                raise NotImplementedError(
+                    "the 'numba' engine doesn't support "
+                    "using a numpy ufunc as the callable function"
+                )
+            with np.errstate(all="ignore"):
+                results = self.obj._mgr.apply("apply", func=self.func)
+            # _constructor will retain self.index and self.columns
+            return self.obj._constructor_from_mgr(results, axes=results.axes)
+
+        # broadcasting
+        if self.result_type == "broadcast":
+            if self.engine == "numba":
+                raise NotImplementedError(
+                    "the 'numba' engine doesn't support result_type='broadcast'"
+                )
+            return self.apply_broadcast(self.obj)
+
+        # one axis empty
+        elif not all(self.obj.shape):
+            return self.apply_empty_result()
+
+        # raw
+        elif self.raw:
+            return self.apply_raw(engine=self.engine, engine_kwargs=self.engine_kwargs)
+
+        return self.apply_standard()
+
+    def agg(self):
+        obj = self.obj
+        axis = self.axis
+
+        # TODO: Avoid having to change state
+        self.obj = self.obj if self.axis == 0 else self.obj.T
+        self.axis = 0
+
+        result = None
+        try:
+            result = super().agg()
+        finally:
+            self.obj = obj
+            self.axis = axis
+
+        if axis == 1:
+            result = result.T if result is not None else result
+
+        if result is None:
+            result = self.obj.apply(self.func, axis, args=self.args, **self.kwargs)
+
+        return result
+
+    def apply_empty_result(self):
+        """
+        we have an empty result; at least 1 axis is 0
+
+        we will try to apply the function to an empty
+        series in order to see if this is a reduction function
+        """
+        assert callable(self.func)
+
+        # we are not asked to reduce or infer reduction
+        # so just return a copy of the existing object
+        if self.result_type not in ["reduce", None]:
+            return self.obj.copy()
+
+        # we may need to infer
+        should_reduce = self.result_type == "reduce"
+
+        from pandas import Series
+
+        if not should_reduce:
+            try:
+                if self.axis == 0:
+                    r = self.func(
+                        Series([], dtype=np.float64), *self.args, **self.kwargs
+                    )
+                else:
+                    r = self.func(
+                        Series(index=self.columns, dtype=np.float64),
+                        *self.args,
+                        **self.kwargs,
+                    )
+            except Exception:
+                pass
+            else:
+                should_reduce = not isinstance(r, Series)
+
+        if should_reduce:
+            if len(self.agg_axis):
+                r = self.func(Series([], dtype=np.float64), *self.args, **self.kwargs)
+            else:
+                r = np.nan
+
+            return self.obj._constructor_sliced(r, index=self.agg_axis)
+        else:
+            return self.obj.copy()
+
+    def apply_raw(self, engine="python", engine_kwargs=None):
+        """apply to the values as a numpy array"""
+
+        def wrap_function(func):
+            """
+            Wrap user supplied function to work around numpy issue.
+
+            see https://github.com/numpy/numpy/issues/8352
+            """
+
+            def wrapper(*args, **kwargs):
+                result = func(*args, **kwargs)
+                if isinstance(result, str):
+                    result = np.array(result, dtype=object)
+                return result
+
+            return wrapper
+
+        if engine == "numba":
+            engine_kwargs = {} if engine_kwargs is None else engine_kwargs
+
+            # error: Argument 1 to "__call__" of "_lru_cache_wrapper" has
+            # incompatible type "Callable[..., Any] | str | list[Callable
+            # [..., Any] | str] | dict[Hashable,Callable[..., Any] | str |
+            # list[Callable[..., Any] | str]]"; expected "Hashable"
+            nb_looper = generate_apply_looper(
+                self.func, **engine_kwargs  # type: ignore[arg-type]
+            )
+            result = nb_looper(self.values, self.axis)
+            # If we made the result 2-D, squeeze it back to 1-D
+            result = np.squeeze(result)
+        else:
+            result = np.apply_along_axis(
+                wrap_function(self.func),
+                self.axis,
+                self.values,
+                *self.args,
+                **self.kwargs,
+            )
+
+        # TODO: mixed type case
+        if result.ndim == 2:
+            return self.obj._constructor(result, index=self.index, columns=self.columns)
+        else:
+            return self.obj._constructor_sliced(result, index=self.agg_axis)
+
+    def apply_broadcast(self, target: DataFrame) -> DataFrame:
+        assert callable(self.func)
+
+        result_values = np.empty_like(target.values)
+
+        # axis which we want to compare compliance
+        result_compare = target.shape[0]
+
+        for i, col in enumerate(target.columns):
+            res = self.func(target[col], *self.args, **self.kwargs)
+            ares = np.asarray(res).ndim
+
+            # must be a scalar or 1d
+            if ares > 1:
+                raise ValueError("too many dims to broadcast")
+            if ares == 1:
+                # must match return dim
+                if result_compare != len(res):
+                    raise ValueError("cannot broadcast result")
+
+            result_values[:, i] = res
+
+        # we *always* preserve the original index / columns
+        result = self.obj._constructor(
+            result_values, index=target.index, columns=target.columns
+        )
+        return result
+
+    def apply_standard(self):
+        if self.engine == "python":
+            results, res_index = self.apply_series_generator()
+        else:
+            results, res_index = self.apply_series_numba()
+
+        # wrap results
+        return self.wrap_results(results, res_index)
+
+    def apply_series_generator(self) -> tuple[ResType, Index]:
+        assert callable(self.func)
+
+        series_gen = self.series_generator
+        res_index = self.result_index
+
+        results = {}
+
+        with option_context("mode.chained_assignment", None):
+            for i, v in enumerate(series_gen):
+                # ignore SettingWithCopy here in case the user mutates
+                results[i] = self.func(v, *self.args, **self.kwargs)
+                if isinstance(results[i], ABCSeries):
+                    # If we have a view on v, we need to make a copy because
+                    #  series_generator will swap out the underlying data
+                    results[i] = results[i].copy(deep=False)
+
+        return results, res_index
+
+    def apply_series_numba(self):
+        if self.engine_kwargs.get("parallel", False):
+            raise NotImplementedError(
+                "Parallel apply is not supported when raw=False and engine='numba'"
+            )
+        if not self.obj.index.is_unique or not self.columns.is_unique:
+            raise NotImplementedError(
+                "The index/columns must be unique when raw=False and engine='numba'"
+            )
+        self.validate_values_for_numba()
+        results = self.apply_with_numba()
+        return results, self.result_index
+
+    def wrap_results(self, results: ResType, res_index: Index) -> DataFrame | Series:
+        from pandas import Series
+
+        # see if we can infer the results
+        if len(results) > 0 and 0 in results and is_sequence(results[0]):
+            return self.wrap_results_for_axis(results, res_index)
+
+        # dict of scalars
+
+        # the default dtype of an empty Series is `object`, but this
+        # code can be hit by df.mean() where the result should have dtype
+        # float64 even if it's an empty Series.
+        constructor_sliced = self.obj._constructor_sliced
+        if len(results) == 0 and constructor_sliced is Series:
+            result = constructor_sliced(results, dtype=np.float64)
+        else:
+            result = constructor_sliced(results)
+        result.index = res_index
+
+        return result
+
+    def apply_str(self) -> DataFrame | Series:
+        # Caller is responsible for checking isinstance(self.func, str)
+        # TODO: GH#39993 - Avoid special-casing by replacing with lambda
+        if self.func == "size":
+            # Special-cased because DataFrame.size returns a single scalar
+            obj = self.obj
+            value = obj.shape[self.axis]
+            return obj._constructor_sliced(value, index=self.agg_axis)
+        return super().apply_str()
+
+
+class FrameRowApply(FrameApply):
+    axis: AxisInt = 0
+
+    @property
+    def series_generator(self) -> Generator[Series, None, None]:
+        return (self.obj._ixs(i, axis=1) for i in range(len(self.columns)))
+
+    @staticmethod
+    @functools.cache
+    def generate_numba_apply_func(
+        func, nogil=True, nopython=True, parallel=False
+    ) -> Callable[[npt.NDArray, Index, Index], dict[int, Any]]:
+        numba = import_optional_dependency("numba")
+        from pandas import Series
+
+        # Import helper from extensions to cast string object -> np strings
+        # Note: This also has the side effect of loading our numba extensions
+        from pandas.core._numba.extensions import maybe_cast_str
+
+        jitted_udf = numba.extending.register_jitable(func)
+
+        # Currently the parallel argument doesn't get passed through here
+        # (it's disabled) since the dicts in numba aren't thread-safe.
+        @numba.jit(nogil=nogil, nopython=nopython, parallel=parallel)
+        def numba_func(values, col_names, df_index):
+            results = {}
+            for j in range(values.shape[1]):
+                # Create the series
+                ser = Series(
+                    values[:, j], index=df_index, name=maybe_cast_str(col_names[j])
+                )
+                results[j] = jitted_udf(ser)
+            return results
+
+        return numba_func
+
+    def apply_with_numba(self) -> dict[int, Any]:
+        nb_func = self.generate_numba_apply_func(
+            cast(Callable, self.func), **self.engine_kwargs
+        )
+        from pandas.core._numba.extensions import set_numba_data
+
+        index = self.obj.index
+        if index.dtype == "string":
+            index = index.astype(object)
+
+        columns = self.obj.columns
+        if columns.dtype == "string":
+            columns = columns.astype(object)
+
+        # Convert from numba dict to regular dict
+        # Our isinstance checks in the df constructor don't pass for numbas typed dict
+        with set_numba_data(index) as index, set_numba_data(columns) as columns:
+            res = dict(nb_func(self.values, columns, index))
+        return res
+
+    @property
+    def result_index(self) -> Index:
+        return self.columns
+
+    @property
+    def result_columns(self) -> Index:
+        return self.index
+
+    def wrap_results_for_axis(
+        self, results: ResType, res_index: Index
+    ) -> DataFrame | Series:
+        """return the results for the rows"""
+
+        if self.result_type == "reduce":
+            # e.g. test_apply_dict GH#8735
+            res = self.obj._constructor_sliced(results)
+            res.index = res_index
+            return res
+
+        elif self.result_type is None and all(
+            isinstance(x, dict) for x in results.values()
+        ):
+            # Our operation was a to_dict op e.g.
+            #  test_apply_dict GH#8735, test_apply_reduce_to_dict GH#25196 #37544
+            res = self.obj._constructor_sliced(results)
+            res.index = res_index
+            return res
+
+        try:
+            result = self.obj._constructor(data=results)
+        except ValueError as err:
+            if "All arrays must be of the same length" in str(err):
+                # e.g. result = [[2, 3], [1.5], ['foo', 'bar']]
+                #  see test_agg_listlike_result GH#29587
+                res = self.obj._constructor_sliced(results)
+                res.index = res_index
+                return res
+            else:
+                raise
+
+        if not isinstance(results[0], ABCSeries):
+            if len(result.index) == len(self.res_columns):
+                result.index = self.res_columns
+
+        if len(result.columns) == len(res_index):
+            result.columns = res_index
+
+        return result
+
+
+class FrameColumnApply(FrameApply):
+    axis: AxisInt = 1
+
+    def apply_broadcast(self, target: DataFrame) -> DataFrame:
+        result = super().apply_broadcast(target.T)
+        return result.T
+
+    @property
+    def series_generator(self) -> Generator[Series, None, None]:
+        values = self.values
+        values = ensure_wrapped_if_datetimelike(values)
+        assert len(values) > 0
+
+        # We create one Series object, and will swap out the data inside
+        #  of it.  Kids: don't do this at home.
+        ser = self.obj._ixs(0, axis=0)
+        mgr = ser._mgr
+
+        is_view = mgr.blocks[0].refs.has_reference()  # type: ignore[union-attr]
+
+        if isinstance(ser.dtype, ExtensionDtype):
+            # values will be incorrect for this block
+            # TODO(EA2D): special case would be unnecessary with 2D EAs
+            obj = self.obj
+            for i in range(len(obj)):
+                yield obj._ixs(i, axis=0)
+
+        else:
+            for arr, name in zip(values, self.index):
+                # GH#35462 re-pin mgr in case setitem changed it
+                ser._mgr = mgr
+                mgr.set_values(arr)
+                object.__setattr__(ser, "_name", name)
+                if not is_view:
+                    # In apply_series_generator we store the a shallow copy of the
+                    # result, which potentially increases the ref count of this reused
+                    # `ser` object (depending on the result of the applied function)
+                    # -> if that happened and `ser` is already a copy, then we reset
+                    # the refs here to avoid triggering a unnecessary CoW inside the
+                    # applied function (https://github.com/pandas-dev/pandas/pull/56212)
+                    mgr.blocks[0].refs = BlockValuesRefs(mgr.blocks[0])  # type: ignore[union-attr]
+                yield ser
+
+    @staticmethod
+    @functools.cache
+    def generate_numba_apply_func(
+        func, nogil=True, nopython=True, parallel=False
+    ) -> Callable[[npt.NDArray, Index, Index], dict[int, Any]]:
+        numba = import_optional_dependency("numba")
+        from pandas import Series
+        from pandas.core._numba.extensions import maybe_cast_str
+
+        jitted_udf = numba.extending.register_jitable(func)
+
+        @numba.jit(nogil=nogil, nopython=nopython, parallel=parallel)
+        def numba_func(values, col_names_index, index):
+            results = {}
+            # Currently the parallel argument doesn't get passed through here
+            # (it's disabled) since the dicts in numba aren't thread-safe.
+            for i in range(values.shape[0]):
+                # Create the series
+                # TODO: values corrupted without the copy
+                ser = Series(
+                    values[i].copy(),
+                    index=col_names_index,
+                    name=maybe_cast_str(index[i]),
+                )
+                results[i] = jitted_udf(ser)
+
+            return results
+
+        return numba_func
+
+    def apply_with_numba(self) -> dict[int, Any]:
+        nb_func = self.generate_numba_apply_func(
+            cast(Callable, self.func), **self.engine_kwargs
+        )
+
+        from pandas.core._numba.extensions import set_numba_data
+
+        # Convert from numba dict to regular dict
+        # Our isinstance checks in the df constructor don't pass for numbas typed dict
+        with set_numba_data(self.obj.index) as index, set_numba_data(
+            self.columns
+        ) as columns:
+            res = dict(nb_func(self.values, columns, index))
+
+        return res
+
+    @property
+    def result_index(self) -> Index:
+        return self.index
+
+    @property
+    def result_columns(self) -> Index:
+        return self.columns
+
+    def wrap_results_for_axis(
+        self, results: ResType, res_index: Index
+    ) -> DataFrame | Series:
+        """return the results for the columns"""
+        result: DataFrame | Series
+
+        # we have requested to expand
+        if self.result_type == "expand":
+            result = self.infer_to_same_shape(results, res_index)
+
+        # we have a non-series and don't want inference
+        elif not isinstance(results[0], ABCSeries):
+            result = self.obj._constructor_sliced(results)
+            result.index = res_index
+
+        # we may want to infer results
+        else:
+            result = self.infer_to_same_shape(results, res_index)
+
+        return result
+
+    def infer_to_same_shape(self, results: ResType, res_index: Index) -> DataFrame:
+        """infer the results to the same shape as the input object"""
+        result = self.obj._constructor(data=results)
+        result = result.T
+
+        # set the index
+        result.index = res_index
+
+        # infer dtypes
+        result = result.infer_objects(copy=False)
+
+        return result
+
+
+class SeriesApply(NDFrameApply):
+    obj: Series
+    axis: AxisInt = 0
+    by_row: Literal[False, "compat", "_compat"]  # only relevant for apply()
+
+    def __init__(
+        self,
+        obj: Series,
+        func: AggFuncType,
+        *,
+        convert_dtype: bool | lib.NoDefault = lib.no_default,
+        by_row: Literal[False, "compat", "_compat"] = "compat",
+        args,
+        kwargs,
+    ) -> None:
+        if convert_dtype is lib.no_default:
+            convert_dtype = True
+        else:
+            warnings.warn(
+                "the convert_dtype parameter is deprecated and will be removed in a "
+                "future version.  Do ``ser.astype(object).apply()`` "
+                "instead if you want ``convert_dtype=False``.",
+                FutureWarning,
+                stacklevel=find_stack_level(),
+            )
+        self.convert_dtype = convert_dtype
+
+        super().__init__(
+            obj,
+            func,
+            raw=False,
+            result_type=None,
+            by_row=by_row,
+            args=args,
+            kwargs=kwargs,
+        )
+
+    def apply(self) -> DataFrame | Series:
+        obj = self.obj
+
+        if len(obj) == 0:
+            return self.apply_empty_result()
+
+        # dispatch to handle list-like or dict-like
+        if is_list_like(self.func):
+            return self.apply_list_or_dict_like()
+
+        if isinstance(self.func, str):
+            # if we are a string, try to dispatch
+            return self.apply_str()
+
+        if self.by_row == "_compat":
+            return self.apply_compat()
+
+        # self.func is Callable
+        return self.apply_standard()
+
+    def agg(self):
+        result = super().agg()
+        if result is None:
+            obj = self.obj
+            func = self.func
+            # string, list-like, and dict-like are entirely handled in super
+            assert callable(func)
+
+            # GH53325: The setup below is just to keep current behavior while emitting a
+            # deprecation message. In the future this will all be replaced with a simple
+            # `result = f(self.obj, *self.args, **self.kwargs)`.
+            try:
+                result = obj.apply(func, args=self.args, **self.kwargs)
+            except (ValueError, AttributeError, TypeError):
+                result = func(obj, *self.args, **self.kwargs)
+            else:
+                msg = (
+                    f"using {func} in {type(obj).__name__}.agg cannot aggregate and "
+                    f"has been deprecated. Use {type(obj).__name__}.transform to "
+                    f"keep behavior unchanged."
+                )
+                warnings.warn(msg, FutureWarning, stacklevel=find_stack_level())
+
+        return result
+
+    def apply_empty_result(self) -> Series:
+        obj = self.obj
+        return obj._constructor(dtype=obj.dtype, index=obj.index).__finalize__(
+            obj, method="apply"
+        )
+
+    def apply_compat(self):
+        """compat apply method for funcs in listlikes and dictlikes.
+
+         Used for each callable when giving listlikes and dictlikes of callables to
+         apply. Needed for compatibility with Pandas < v2.1.
+
+        .. versionadded:: 2.1.0
+        """
+        obj = self.obj
+        func = self.func
+
+        if callable(func):
+            f = com.get_cython_func(func)
+            if f and not self.args and not self.kwargs:
+                return obj.apply(func, by_row=False)
+
+        try:
+            result = obj.apply(func, by_row="compat")
+        except (ValueError, AttributeError, TypeError):
+            result = obj.apply(func, by_row=False)
+        return result
+
+    def apply_standard(self) -> DataFrame | Series:
+        # caller is responsible for ensuring that f is Callable
+        func = cast(Callable, self.func)
+        obj = self.obj
+
+        if isinstance(func, np.ufunc):
+            with np.errstate(all="ignore"):
+                return func(obj, *self.args, **self.kwargs)
+        elif not self.by_row:
+            return func(obj, *self.args, **self.kwargs)
+
+        if self.args or self.kwargs:
+            # _map_values does not support args/kwargs
+            def curried(x):
+                return func(x, *self.args, **self.kwargs)
+
+        else:
+            curried = func
+
+        # row-wise access
+        # apply doesn't have a `na_action` keyword and for backward compat reasons
+        # we need to give `na_action="ignore"` for categorical data.
+        # TODO: remove the `na_action="ignore"` when that default has been changed in
+        #  Categorical (GH51645).
+        action = "ignore" if isinstance(obj.dtype, CategoricalDtype) else None
+        mapped = obj._map_values(
+            mapper=curried, na_action=action, convert=self.convert_dtype
+        )
+
+        if len(mapped) and isinstance(mapped[0], ABCSeries):
+            # GH#43986 Need to do list(mapped) in order to get treated as nested
+            #  See also GH#25959 regarding EA support
+            return obj._constructor_expanddim(list(mapped), index=obj.index)
+        else:
+            return obj._constructor(mapped, index=obj.index).__finalize__(
+                obj, method="apply"
+            )
+
+
+class GroupByApply(Apply):
+    obj: GroupBy | Resampler | BaseWindow
+
+    def __init__(
+        self,
+        obj: GroupBy[NDFrameT],
+        func: AggFuncType,
+        *,
+        args,
+        kwargs,
+    ) -> None:
+        kwargs = kwargs.copy()
+        self.axis = obj.obj._get_axis_number(kwargs.get("axis", 0))
+        super().__init__(
+            obj,
+            func,
+            raw=False,
+            result_type=None,
+            args=args,
+            kwargs=kwargs,
+        )
+
+    def apply(self):
+        raise NotImplementedError
+
+    def transform(self):
+        raise NotImplementedError
+
+    def agg_or_apply_list_like(
+        self, op_name: Literal["agg", "apply"]
+    ) -> DataFrame | Series:
+        obj = self.obj
+        kwargs = self.kwargs
+        if op_name == "apply":
+            kwargs = {**kwargs, "by_row": False}
+
+        if getattr(obj, "axis", 0) == 1:
+            raise NotImplementedError("axis other than 0 is not supported")
+
+        if obj._selected_obj.ndim == 1:
+            # For SeriesGroupBy this matches _obj_with_exclusions
+            selected_obj = obj._selected_obj
+        else:
+            selected_obj = obj._obj_with_exclusions
+
+        # Only set as_index=True on groupby objects, not Window or Resample
+        # that inherit from this class.
+        with com.temp_setattr(
+            obj, "as_index", True, condition=hasattr(obj, "as_index")
+        ):
+            keys, results = self.compute_list_like(op_name, selected_obj, kwargs)
+        result = self.wrap_results_list_like(keys, results)
+        return result
+
+    def agg_or_apply_dict_like(
+        self, op_name: Literal["agg", "apply"]
+    ) -> DataFrame | Series:
+        from pandas.core.groupby.generic import (
+            DataFrameGroupBy,
+            SeriesGroupBy,
+        )
+
+        assert op_name in ["agg", "apply"]
+
+        obj = self.obj
+        kwargs = {}
+        if op_name == "apply":
+            by_row = "_compat" if self.by_row else False
+            kwargs.update({"by_row": by_row})
+
+        if getattr(obj, "axis", 0) == 1:
+            raise NotImplementedError("axis other than 0 is not supported")
+
+        selected_obj = obj._selected_obj
+        selection = obj._selection
+
+        is_groupby = isinstance(obj, (DataFrameGroupBy, SeriesGroupBy))
+
+        # Numba Groupby engine/engine-kwargs passthrough
+        if is_groupby:
+            engine = self.kwargs.get("engine", None)
+            engine_kwargs = self.kwargs.get("engine_kwargs", None)
+            kwargs.update({"engine": engine, "engine_kwargs": engine_kwargs})
+
+        with com.temp_setattr(
+            obj, "as_index", True, condition=hasattr(obj, "as_index")
+        ):
+            result_index, result_data = self.compute_dict_like(
+                op_name, selected_obj, selection, kwargs
+            )
+        result = self.wrap_results_dict_like(selected_obj, result_index, result_data)
+        return result
+
+
+class ResamplerWindowApply(GroupByApply):
+    axis: AxisInt = 0
+    obj: Resampler | BaseWindow
+
+    def __init__(
+        self,
+        obj: Resampler | BaseWindow,
+        func: AggFuncType,
+        *,
+        args,
+        kwargs,
+    ) -> None:
+        super(GroupByApply, self).__init__(
+            obj,
+            func,
+            raw=False,
+            result_type=None,
+            args=args,
+            kwargs=kwargs,
+        )
+
+    def apply(self):
+        raise NotImplementedError
+
+    def transform(self):
+        raise NotImplementedError
+
+
+def reconstruct_func(
+    func: AggFuncType | None, **kwargs
+) -> tuple[bool, AggFuncType, tuple[str, ...] | None, npt.NDArray[np.intp] | None]:
+    """
+    This is the internal function to reconstruct func given if there is relabeling
+    or not and also normalize the keyword to get new order of columns.
+
+    If named aggregation is applied, `func` will be None, and kwargs contains the
+    column and aggregation function information to be parsed;
+    If named aggregation is not applied, `func` is either string (e.g. 'min') or
+    Callable, or list of them (e.g. ['min', np.max]), or the dictionary of column name
+    and str/Callable/list of them (e.g. {'A': 'min'}, or {'A': [np.min, lambda x: x]})
+
+    If relabeling is True, will return relabeling, reconstructed func, column
+    names, and the reconstructed order of columns.
+    If relabeling is False, the columns and order will be None.
+
+    Parameters
+    ----------
+    func: agg function (e.g. 'min' or Callable) or list of agg functions
+        (e.g. ['min', np.max]) or dictionary (e.g. {'A': ['min', np.max]}).
+    **kwargs: dict, kwargs used in is_multi_agg_with_relabel and
+        normalize_keyword_aggregation function for relabelling
+
+    Returns
+    -------
+    relabelling: bool, if there is relabelling or not
+    func: normalized and mangled func
+    columns: tuple of column names
+    order: array of columns indices
+
+    Examples
+    --------
+    >>> reconstruct_func(None, **{"foo": ("col", "min")})
+    (True, defaultdict(<class 'list'>, {'col': ['min']}), ('foo',), array([0]))
+
+    >>> reconstruct_func("min")
+    (False, 'min', None, None)
+    """
+    relabeling = func is None and is_multi_agg_with_relabel(**kwargs)
+    columns: tuple[str, ...] | None = None
+    order: npt.NDArray[np.intp] | None = None
+
+    if not relabeling:
+        if isinstance(func, list) and len(func) > len(set(func)):
+            # GH 28426 will raise error if duplicated function names are used and
+            # there is no reassigned name
+            raise SpecificationError(
+                "Function names must be unique if there is no new column names "
+                "assigned"
+            )
+        if func is None:
+            # nicer error message
+            raise TypeError("Must provide 'func' or tuples of '(column, aggfunc).")
+
+    if relabeling:
+        # error: Incompatible types in assignment (expression has type
+        # "MutableMapping[Hashable, list[Callable[..., Any] | str]]", variable has type
+        # "Callable[..., Any] | str | list[Callable[..., Any] | str] |
+        # MutableMapping[Hashable, Callable[..., Any] | str | list[Callable[..., Any] |
+        # str]] | None")
+        func, columns, order = normalize_keyword_aggregation(  # type: ignore[assignment]
+            kwargs
+        )
+    assert func is not None
+
+    return relabeling, func, columns, order
+
+
+def is_multi_agg_with_relabel(**kwargs) -> bool:
+    """
+    Check whether kwargs passed to .agg look like multi-agg with relabeling.
+
+    Parameters
+    ----------
+    **kwargs : dict
+
+    Returns
+    -------
+    bool
+
+    Examples
+    --------
+    >>> is_multi_agg_with_relabel(a="max")
+    False
+    >>> is_multi_agg_with_relabel(a_max=("a", "max"), a_min=("a", "min"))
+    True
+    >>> is_multi_agg_with_relabel()
+    False
+    """
+    return all(isinstance(v, tuple) and len(v) == 2 for v in kwargs.values()) and (
+        len(kwargs) > 0
+    )
+
+
+def normalize_keyword_aggregation(
+    kwargs: dict,
+) -> tuple[
+    MutableMapping[Hashable, list[AggFuncTypeBase]],
+    tuple[str, ...],
+    npt.NDArray[np.intp],
+]:
+    """
+    Normalize user-provided "named aggregation" kwargs.
+    Transforms from the new ``Mapping[str, NamedAgg]`` style kwargs
+    to the old Dict[str, List[scalar]]].
+
+    Parameters
+    ----------
+    kwargs : dict
+
+    Returns
+    -------
+    aggspec : dict
+        The transformed kwargs.
+    columns : tuple[str, ...]
+        The user-provided keys.
+    col_idx_order : List[int]
+        List of columns indices.
+
+    Examples
+    --------
+    >>> normalize_keyword_aggregation({"output": ("input", "sum")})
+    (defaultdict(<class 'list'>, {'input': ['sum']}), ('output',), array([0]))
+    """
+    from pandas.core.indexes.base import Index
+
+    # Normalize the aggregation functions as Mapping[column, List[func]],
+    # process normally, then fixup the names.
+    # TODO: aggspec type: typing.Dict[str, List[AggScalar]]
+    aggspec = defaultdict(list)
+    order = []
+    columns, pairs = list(zip(*kwargs.items()))
+
+    for column, aggfunc in pairs:
+        aggspec[column].append(aggfunc)
+        order.append((column, com.get_callable_name(aggfunc) or aggfunc))
+
+    # uniquify aggfunc name if duplicated in order list
+    uniquified_order = _make_unique_kwarg_list(order)
+
+    # GH 25719, due to aggspec will change the order of assigned columns in aggregation
+    # uniquified_aggspec will store uniquified order list and will compare it with order
+    # based on index
+    aggspec_order = [
+        (column, com.get_callable_name(aggfunc) or aggfunc)
+        for column, aggfuncs in aggspec.items()
+        for aggfunc in aggfuncs
+    ]
+    uniquified_aggspec = _make_unique_kwarg_list(aggspec_order)
+
+    # get the new index of columns by comparison
+    col_idx_order = Index(uniquified_aggspec).get_indexer(uniquified_order)
+    return aggspec, columns, col_idx_order
+
+
+def _make_unique_kwarg_list(
+    seq: Sequence[tuple[Any, Any]]
+) -> Sequence[tuple[Any, Any]]:
+    """
+    Uniquify aggfunc name of the pairs in the order list
+
+    Examples:
+    --------
+    >>> kwarg_list = [('a', '<lambda>'), ('a', '<lambda>'), ('b', '<lambda>')]
+    >>> _make_unique_kwarg_list(kwarg_list)
+    [('a', '<lambda>_0'), ('a', '<lambda>_1'), ('b', '<lambda>')]
+    """
+    return [
+        (pair[0], f"{pair[1]}_{seq[:i].count(pair)}") if seq.count(pair) > 1 else pair
+        for i, pair in enumerate(seq)
+    ]
+
+
+def relabel_result(
+    result: DataFrame | Series,
+    func: dict[str, list[Callable | str]],
+    columns: Iterable[Hashable],
+    order: Iterable[int],
+) -> dict[Hashable, Series]:
+    """
+    Internal function to reorder result if relabelling is True for
+    dataframe.agg, and return the reordered result in dict.
+
+    Parameters:
+    ----------
+    result: Result from aggregation
+    func: Dict of (column name, funcs)
+    columns: New columns name for relabelling
+    order: New order for relabelling
+
+    Examples
+    --------
+    >>> from pandas.core.apply import relabel_result
+    >>> result = pd.DataFrame(
+    ...     {"A": [np.nan, 2, np.nan], "C": [6, np.nan, np.nan], "B": [np.nan, 4, 2.5]},
+    ...     index=["max", "mean", "min"]
+    ... )
+    >>> funcs = {"A": ["max"], "C": ["max"], "B": ["mean", "min"]}
+    >>> columns = ("foo", "aab", "bar", "dat")
+    >>> order = [0, 1, 2, 3]
+    >>> result_in_dict = relabel_result(result, funcs, columns, order)
+    >>> pd.DataFrame(result_in_dict, index=columns)
+           A    C    B
+    foo  2.0  NaN  NaN
+    aab  NaN  6.0  NaN
+    bar  NaN  NaN  4.0
+    dat  NaN  NaN  2.5
+    """
+    from pandas.core.indexes.base import Index
+
+    reordered_indexes = [
+        pair[0] for pair in sorted(zip(columns, order), key=lambda t: t[1])
+    ]
+    reordered_result_in_dict: dict[Hashable, Series] = {}
+    idx = 0
+
+    reorder_mask = not isinstance(result, ABCSeries) and len(result.columns) > 1
+    for col, fun in func.items():
+        s = result[col].dropna()
+
+        # In the `_aggregate`, the callable names are obtained and used in `result`, and
+        # these names are ordered alphabetically. e.g.
+        #           C2   C1
+        # <lambda>   1  NaN
+        # amax     NaN  4.0
+        # max      NaN  4.0
+        # sum     18.0  6.0
+        # Therefore, the order of functions for each column could be shuffled
+        # accordingly so need to get the callable name if it is not parsed names, and
+        # reorder the aggregated result for each column.
+        # e.g. if df.agg(c1=("C2", sum), c2=("C2", lambda x: min(x))), correct order is
+        # [sum, <lambda>], but in `result`, it will be [<lambda>, sum], and we need to
+        # reorder so that aggregated values map to their functions regarding the order.
+
+        # However there is only one column being used for aggregation, not need to
+        # reorder since the index is not sorted, and keep as is in `funcs`, e.g.
+        #         A
+        # min   1.0
+        # mean  1.5
+        # mean  1.5
+        if reorder_mask:
+            fun = [
+                com.get_callable_name(f) if not isinstance(f, str) else f for f in fun
+            ]
+            col_idx_order = Index(s.index).get_indexer(fun)
+            s = s.iloc[col_idx_order]
+
+        # assign the new user-provided "named aggregation" as index names, and reindex
+        # it based on the whole user-provided names.
+        s.index = reordered_indexes[idx : idx + len(fun)]
+        reordered_result_in_dict[col] = s.reindex(columns, copy=False)
+        idx = idx + len(fun)
+    return reordered_result_in_dict
+
+
+def reconstruct_and_relabel_result(result, func, **kwargs) -> DataFrame | Series:
+    from pandas import DataFrame
+
+    relabeling, func, columns, order = reconstruct_func(func, **kwargs)
+
+    if relabeling:
+        # This is to keep the order to columns occurrence unchanged, and also
+        # keep the order of new columns occurrence unchanged
+
+        # For the return values of reconstruct_func, if relabeling is
+        # False, columns and order will be None.
+        assert columns is not None
+        assert order is not None
+
+        result_in_dict = relabel_result(result, func, columns, order)
+        result = DataFrame(result_in_dict, index=columns)
+
+    return result
+
+
+# TODO: Can't use, because mypy doesn't like us setting __name__
+#   error: "partial[Any]" has no attribute "__name__"
+# the type is:
+#   typing.Sequence[Callable[..., ScalarResult]]
+#     -> typing.Sequence[Callable[..., ScalarResult]]:
+
+
+def _managle_lambda_list(aggfuncs: Sequence[Any]) -> Sequence[Any]:
+    """
+    Possibly mangle a list of aggfuncs.
+
+    Parameters
+    ----------
+    aggfuncs : Sequence
+
+    Returns
+    -------
+    mangled: list-like
+        A new AggSpec sequence, where lambdas have been converted
+        to have unique names.
+
+    Notes
+    -----
+    If just one aggfunc is passed, the name will not be mangled.
+    """
+    if len(aggfuncs) <= 1:
+        # don't mangle for .agg([lambda x: .])
+        return aggfuncs
+    i = 0
+    mangled_aggfuncs = []
+    for aggfunc in aggfuncs:
+        if com.get_callable_name(aggfunc) == "<lambda>":
+            aggfunc = partial(aggfunc)
+            aggfunc.__name__ = f"<lambda_{i}>"
+            i += 1
+        mangled_aggfuncs.append(aggfunc)
+
+    return mangled_aggfuncs
+
+
+def maybe_mangle_lambdas(agg_spec: Any) -> Any:
+    """
+    Make new lambdas with unique names.
+
+    Parameters
+    ----------
+    agg_spec : Any
+        An argument to GroupBy.agg.
+        Non-dict-like `agg_spec` are pass through as is.
+        For dict-like `agg_spec` a new spec is returned
+        with name-mangled lambdas.
+
+    Returns
+    -------
+    mangled : Any
+        Same type as the input.
+
+    Examples
+    --------
+    >>> maybe_mangle_lambdas('sum')
+    'sum'
+    >>> maybe_mangle_lambdas([lambda: 1, lambda: 2])  # doctest: +SKIP
+    [<function __main__.<lambda_0>,
+     <function pandas...._make_lambda.<locals>.f(*args, **kwargs)>]
+    """
+    is_dict = is_dict_like(agg_spec)
+    if not (is_dict or is_list_like(agg_spec)):
+        return agg_spec
+    mangled_aggspec = type(agg_spec)()  # dict or OrderedDict
+
+    if is_dict:
+        for key, aggfuncs in agg_spec.items():
+            if is_list_like(aggfuncs) and not is_dict_like(aggfuncs):
+                mangled_aggfuncs = _managle_lambda_list(aggfuncs)
+            else:
+                mangled_aggfuncs = aggfuncs
+
+            mangled_aggspec[key] = mangled_aggfuncs
+    else:
+        mangled_aggspec = _managle_lambda_list(agg_spec)
+
+    return mangled_aggspec
+
+
+def validate_func_kwargs(
+    kwargs: dict,
+) -> tuple[list[str], list[str | Callable[..., Any]]]:
+    """
+    Validates types of user-provided "named aggregation" kwargs.
+    `TypeError` is raised if aggfunc is not `str` or callable.
+
+    Parameters
+    ----------
+    kwargs : dict
+
+    Returns
+    -------
+    columns : List[str]
+        List of user-provided keys.
+    func : List[Union[str, callable[...,Any]]]
+        List of user-provided aggfuncs
+
+    Examples
+    --------
+    >>> validate_func_kwargs({'one': 'min', 'two': 'max'})
+    (['one', 'two'], ['min', 'max'])
+    """
+    tuple_given_message = "func is expected but received {} in **kwargs."
+    columns = list(kwargs)
+    func = []
+    for col_func in kwargs.values():
+        if not (isinstance(col_func, str) or callable(col_func)):
+            raise TypeError(tuple_given_message.format(type(col_func).__name__))
+        func.append(col_func)
+    if not columns:
+        no_arg_message = "Must provide 'func' or named aggregation **kwargs."
+        raise TypeError(no_arg_message)
+    return columns, func
+
+
+def include_axis(op_name: Literal["agg", "apply"], colg: Series | DataFrame) -> bool:
+    return isinstance(colg, ABCDataFrame) or (
+        isinstance(colg, ABCSeries) and op_name == "agg"
+    )
+
+
+def warn_alias_replacement(
+    obj: AggObjType,
+    func: Callable,
+    alias: str,
+) -> None:
+    if alias.startswith("np."):
+        full_alias = alias
+    else:
+        full_alias = f"{type(obj).__name__}.{alias}"
+        alias = f'"{alias}"'
+    warnings.warn(
+        f"The provided callable {func} is currently using "
+        f"{full_alias}. In a future version of pandas, "
+        f"the provided callable will be used directly. To keep current "
+        f"behavior pass the string {alias} instead.",
+        category=FutureWarning,
+        stacklevel=find_stack_level(),
+    )

llava_next/lib/python3.10/site-packages/pandas/core/array_algos/masked_accumulations.py

@@ -0,0 +1,90 @@
+"""
+masked_accumulations.py is for accumulation algorithms using a mask-based approach
+for missing values.
+"""
+
+from __future__ import annotations
+
+from typing import (
+    TYPE_CHECKING,
+    Callable,
+)
+
+import numpy as np
+
+if TYPE_CHECKING:
+    from pandas._typing import npt
+
+
+def _cum_func(
+    func: Callable,
+    values: np.ndarray,
+    mask: npt.NDArray[np.bool_],
+    *,
+    skipna: bool = True,
+):
+    """
+    Accumulations for 1D masked array.
+
+    We will modify values in place to replace NAs with the appropriate fill value.
+
+    Parameters
+    ----------
+    func : np.cumsum, np.cumprod, np.maximum.accumulate, np.minimum.accumulate
+    values : np.ndarray
+        Numpy array with the values (can be of any dtype that support the
+        operation).
+    mask : np.ndarray
+        Boolean numpy array (True values indicate missing values).
+    skipna : bool, default True
+        Whether to skip NA.
+    """
+    dtype_info: np.iinfo | np.finfo
+    if values.dtype.kind == "f":
+        dtype_info = np.finfo(values.dtype.type)
+    elif values.dtype.kind in "iu":
+        dtype_info = np.iinfo(values.dtype.type)
+    elif values.dtype.kind == "b":
+        # Max value of bool is 1, but since we are setting into a boolean
+        # array, 255 is fine as well. Min value has to be 0 when setting
+        # into the boolean array.
+        dtype_info = np.iinfo(np.uint8)
+    else:
+        raise NotImplementedError(
+            f"No masked accumulation defined for dtype {values.dtype.type}"
+        )
+    try:
+        fill_value = {
+            np.cumprod: 1,
+            np.maximum.accumulate: dtype_info.min,
+            np.cumsum: 0,
+            np.minimum.accumulate: dtype_info.max,
+        }[func]
+    except KeyError:
+        raise NotImplementedError(
+            f"No accumulation for {func} implemented on BaseMaskedArray"
+        )
+
+    values[mask] = fill_value
+
+    if not skipna:
+        mask = np.maximum.accumulate(mask)
+
+    values = func(values)
+    return values, mask
+
+
+def cumsum(values: np.ndarray, mask: npt.NDArray[np.bool_], *, skipna: bool = True):
+    return _cum_func(np.cumsum, values, mask, skipna=skipna)
+
+
+def cumprod(values: np.ndarray, mask: npt.NDArray[np.bool_], *, skipna: bool = True):
+    return _cum_func(np.cumprod, values, mask, skipna=skipna)
+
+
+def cummin(values: np.ndarray, mask: npt.NDArray[np.bool_], *, skipna: bool = True):
+    return _cum_func(np.minimum.accumulate, values, mask, skipna=skipna)
+
+
+def cummax(values: np.ndarray, mask: npt.NDArray[np.bool_], *, skipna: bool = True):
+    return _cum_func(np.maximum.accumulate, values, mask, skipna=skipna)

llava_next/lib/python3.10/site-packages/pandas/core/array_algos/replace.py

@@ -0,0 +1,152 @@
+"""
+Methods used by Block.replace and related methods.
+"""
+from __future__ import annotations
+
+import operator
+import re
+from re import Pattern
+from typing import (
+    TYPE_CHECKING,
+    Any,
+)
+
+import numpy as np
+
+from pandas.core.dtypes.common import (
+    is_bool,
+    is_re,
+    is_re_compilable,
+)
+from pandas.core.dtypes.missing import isna
+
+if TYPE_CHECKING:
+    from pandas._typing import (
+        ArrayLike,
+        Scalar,
+        npt,
+    )
+
+
+def should_use_regex(regex: bool, to_replace: Any) -> bool:
+    """
+    Decide whether to treat `to_replace` as a regular expression.
+    """
+    if is_re(to_replace):
+        regex = True
+
+    regex = regex and is_re_compilable(to_replace)
+
+    # Don't use regex if the pattern is empty.
+    regex = regex and re.compile(to_replace).pattern != ""
+    return regex
+
+
+def compare_or_regex_search(
+    a: ArrayLike, b: Scalar | Pattern, regex: bool, mask: npt.NDArray[np.bool_]
+) -> ArrayLike:
+    """
+    Compare two array-like inputs of the same shape or two scalar values
+
+    Calls operator.eq or re.search, depending on regex argument. If regex is
+    True, perform an element-wise regex matching.
+
+    Parameters
+    ----------
+    a : array-like
+    b : scalar or regex pattern
+    regex : bool
+    mask : np.ndarray[bool]
+
+    Returns
+    -------
+    mask : array-like of bool
+    """
+    if isna(b):
+        return ~mask
+
+    def _check_comparison_types(
+        result: ArrayLike | bool, a: ArrayLike, b: Scalar | Pattern
+    ):
+        """
+        Raises an error if the two arrays (a,b) cannot be compared.
+        Otherwise, returns the comparison result as expected.
+        """
+        if is_bool(result) and isinstance(a, np.ndarray):
+            type_names = [type(a).__name__, type(b).__name__]
+
+            type_names[0] = f"ndarray(dtype={a.dtype})"
+
+            raise TypeError(
+                f"Cannot compare types {repr(type_names[0])} and {repr(type_names[1])}"
+            )
+
+    if not regex or not should_use_regex(regex, b):
+        # TODO: should use missing.mask_missing?
+        op = lambda x: operator.eq(x, b)
+    else:
+        op = np.vectorize(
+            lambda x: bool(re.search(b, x))
+            if isinstance(x, str) and isinstance(b, (str, Pattern))
+            else False
+        )
+
+    # GH#32621 use mask to avoid comparing to NAs
+    if isinstance(a, np.ndarray):
+        a = a[mask]
+
+    result = op(a)
+
+    if isinstance(result, np.ndarray) and mask is not None:
+        # The shape of the mask can differ to that of the result
+        # since we may compare only a subset of a's or b's elements
+        tmp = np.zeros(mask.shape, dtype=np.bool_)
+        np.place(tmp, mask, result)
+        result = tmp
+
+    _check_comparison_types(result, a, b)
+    return result
+
+
+def replace_regex(
+    values: ArrayLike, rx: re.Pattern, value, mask: npt.NDArray[np.bool_] | None
+) -> None:
+    """
+    Parameters
+    ----------
+    values : ArrayLike
+        Object dtype.
+    rx : re.Pattern
+    value : Any
+    mask : np.ndarray[bool], optional
+
+    Notes
+    -----
+    Alters values in-place.
+    """
+
+    # deal with replacing values with objects (strings) that match but
+    # whose replacement is not a string (numeric, nan, object)
+    if isna(value) or not isinstance(value, str):
+
+        def re_replacer(s):
+            if is_re(rx) and isinstance(s, str):
+                return value if rx.search(s) is not None else s
+            else:
+                return s
+
+    else:
+        # value is guaranteed to be a string here, s can be either a string
+        # or null if it's null it gets returned
+        def re_replacer(s):
+            if is_re(rx) and isinstance(s, str):
+                return rx.sub(value, s)
+            else:
+                return s
+
+    f = np.vectorize(re_replacer, otypes=[np.object_])
+
+    if mask is None:
+        values[:] = f(values)
+    else:
+        values[mask] = f(values[mask])

llava_next/lib/python3.10/site-packages/pandas/core/arraylike.py

@@ -0,0 +1,530 @@
+"""
+Methods that can be shared by many array-like classes or subclasses:
+    Series
+    Index
+    ExtensionArray
+"""
+from __future__ import annotations
+
+import operator
+from typing import Any
+
+import numpy as np
+
+from pandas._libs import lib
+from pandas._libs.ops_dispatch import maybe_dispatch_ufunc_to_dunder_op
+
+from pandas.core.dtypes.generic import ABCNDFrame
+
+from pandas.core import roperator
+from pandas.core.construction import extract_array
+from pandas.core.ops.common import unpack_zerodim_and_defer
+
+REDUCTION_ALIASES = {
+    "maximum": "max",
+    "minimum": "min",
+    "add": "sum",
+    "multiply": "prod",
+}
+
+
+class OpsMixin:
+    # -------------------------------------------------------------
+    # Comparisons
+
+    def _cmp_method(self, other, op):
+        return NotImplemented
+
+    @unpack_zerodim_and_defer("__eq__")
+    def __eq__(self, other):
+        return self._cmp_method(other, operator.eq)
+
+    @unpack_zerodim_and_defer("__ne__")
+    def __ne__(self, other):
+        return self._cmp_method(other, operator.ne)
+
+    @unpack_zerodim_and_defer("__lt__")
+    def __lt__(self, other):
+        return self._cmp_method(other, operator.lt)
+
+    @unpack_zerodim_and_defer("__le__")
+    def __le__(self, other):
+        return self._cmp_method(other, operator.le)
+
+    @unpack_zerodim_and_defer("__gt__")
+    def __gt__(self, other):
+        return self._cmp_method(other, operator.gt)
+
+    @unpack_zerodim_and_defer("__ge__")
+    def __ge__(self, other):
+        return self._cmp_method(other, operator.ge)
+
+    # -------------------------------------------------------------
+    # Logical Methods
+
+    def _logical_method(self, other, op):
+        return NotImplemented
+
+    @unpack_zerodim_and_defer("__and__")
+    def __and__(self, other):
+        return self._logical_method(other, operator.and_)
+
+    @unpack_zerodim_and_defer("__rand__")
+    def __rand__(self, other):
+        return self._logical_method(other, roperator.rand_)
+
+    @unpack_zerodim_and_defer("__or__")
+    def __or__(self, other):
+        return self._logical_method(other, operator.or_)
+
+    @unpack_zerodim_and_defer("__ror__")
+    def __ror__(self, other):
+        return self._logical_method(other, roperator.ror_)
+
+    @unpack_zerodim_and_defer("__xor__")
+    def __xor__(self, other):
+        return self._logical_method(other, operator.xor)
+
+    @unpack_zerodim_and_defer("__rxor__")
+    def __rxor__(self, other):
+        return self._logical_method(other, roperator.rxor)
+
+    # -------------------------------------------------------------
+    # Arithmetic Methods
+
+    def _arith_method(self, other, op):
+        return NotImplemented
+
+    @unpack_zerodim_and_defer("__add__")
+    def __add__(self, other):
+        """
+        Get Addition of DataFrame and other, column-wise.
+
+        Equivalent to ``DataFrame.add(other)``.
+
+        Parameters
+        ----------
+        other : scalar, sequence, Series, dict or DataFrame
+            Object to be added to the DataFrame.
+
+        Returns
+        -------
+        DataFrame
+            The result of adding ``other`` to DataFrame.
+
+        See Also
+        --------
+        DataFrame.add : Add a DataFrame and another object, with option for index-
+            or column-oriented addition.
+
+        Examples
+        --------
+        >>> df = pd.DataFrame({'height': [1.5, 2.6], 'weight': [500, 800]},
+        ...                   index=['elk', 'moose'])
+        >>> df
+               height  weight
+        elk       1.5     500
+        moose     2.6     800
+
+        Adding a scalar affects all rows and columns.
+
+        >>> df[['height', 'weight']] + 1.5
+               height  weight
+        elk       3.0   501.5
+        moose     4.1   801.5
+
+        Each element of a list is added to a column of the DataFrame, in order.
+
+        >>> df[['height', 'weight']] + [0.5, 1.5]
+               height  weight
+        elk       2.0   501.5
+        moose     3.1   801.5
+
+        Keys of a dictionary are aligned to the DataFrame, based on column names;
+        each value in the dictionary is added to the corresponding column.
+
+        >>> df[['height', 'weight']] + {'height': 0.5, 'weight': 1.5}
+               height  weight
+        elk       2.0   501.5
+        moose     3.1   801.5
+
+        When `other` is a :class:`Series`, the index of `other` is aligned with the
+        columns of the DataFrame.
+
+        >>> s1 = pd.Series([0.5, 1.5], index=['weight', 'height'])
+        >>> df[['height', 'weight']] + s1
+               height  weight
+        elk       3.0   500.5
+        moose     4.1   800.5
+
+        Even when the index of `other` is the same as the index of the DataFrame,
+        the :class:`Series` will not be reoriented. If index-wise alignment is desired,
+        :meth:`DataFrame.add` should be used with `axis='index'`.
+
+        >>> s2 = pd.Series([0.5, 1.5], index=['elk', 'moose'])
+        >>> df[['height', 'weight']] + s2
+               elk  height  moose  weight
+        elk    NaN     NaN    NaN     NaN
+        moose  NaN     NaN    NaN     NaN
+
+        >>> df[['height', 'weight']].add(s2, axis='index')
+               height  weight
+        elk       2.0   500.5
+        moose     4.1   801.5
+
+        When `other` is a :class:`DataFrame`, both columns names and the
+        index are aligned.
+
+        >>> other = pd.DataFrame({'height': [0.2, 0.4, 0.6]},
+        ...                      index=['elk', 'moose', 'deer'])
+        >>> df[['height', 'weight']] + other
+               height  weight
+        deer      NaN     NaN
+        elk       1.7     NaN
+        moose     3.0     NaN
+        """
+        return self._arith_method(other, operator.add)
+
+    @unpack_zerodim_and_defer("__radd__")
+    def __radd__(self, other):
+        return self._arith_method(other, roperator.radd)
+
+    @unpack_zerodim_and_defer("__sub__")
+    def __sub__(self, other):
+        return self._arith_method(other, operator.sub)
+
+    @unpack_zerodim_and_defer("__rsub__")
+    def __rsub__(self, other):
+        return self._arith_method(other, roperator.rsub)
+
+    @unpack_zerodim_and_defer("__mul__")
+    def __mul__(self, other):
+        return self._arith_method(other, operator.mul)
+
+    @unpack_zerodim_and_defer("__rmul__")
+    def __rmul__(self, other):
+        return self._arith_method(other, roperator.rmul)
+
+    @unpack_zerodim_and_defer("__truediv__")
+    def __truediv__(self, other):
+        return self._arith_method(other, operator.truediv)
+
+    @unpack_zerodim_and_defer("__rtruediv__")
+    def __rtruediv__(self, other):
+        return self._arith_method(other, roperator.rtruediv)
+
+    @unpack_zerodim_and_defer("__floordiv__")
+    def __floordiv__(self, other):
+        return self._arith_method(other, operator.floordiv)
+
+    @unpack_zerodim_and_defer("__rfloordiv")
+    def __rfloordiv__(self, other):
+        return self._arith_method(other, roperator.rfloordiv)
+
+    @unpack_zerodim_and_defer("__mod__")
+    def __mod__(self, other):
+        return self._arith_method(other, operator.mod)
+
+    @unpack_zerodim_and_defer("__rmod__")
+    def __rmod__(self, other):
+        return self._arith_method(other, roperator.rmod)
+
+    @unpack_zerodim_and_defer("__divmod__")
+    def __divmod__(self, other):
+        return self._arith_method(other, divmod)
+
+    @unpack_zerodim_and_defer("__rdivmod__")
+    def __rdivmod__(self, other):
+        return self._arith_method(other, roperator.rdivmod)
+
+    @unpack_zerodim_and_defer("__pow__")
+    def __pow__(self, other):
+        return self._arith_method(other, operator.pow)
+
+    @unpack_zerodim_and_defer("__rpow__")
+    def __rpow__(self, other):
+        return self._arith_method(other, roperator.rpow)
+
+
+# -----------------------------------------------------------------------------
+# Helpers to implement __array_ufunc__
+
+
+def array_ufunc(self, ufunc: np.ufunc, method: str, *inputs: Any, **kwargs: Any):
+    """
+    Compatibility with numpy ufuncs.
+
+    See also
+    --------
+    numpy.org/doc/stable/reference/arrays.classes.html#numpy.class.__array_ufunc__
+    """
+    from pandas.core.frame import (
+        DataFrame,
+        Series,
+    )
+    from pandas.core.generic import NDFrame
+    from pandas.core.internals import (
+        ArrayManager,
+        BlockManager,
+    )
+
+    cls = type(self)
+
+    kwargs = _standardize_out_kwarg(**kwargs)
+
+    # for binary ops, use our custom dunder methods
+    result = maybe_dispatch_ufunc_to_dunder_op(self, ufunc, method, *inputs, **kwargs)
+    if result is not NotImplemented:
+        return result
+
+    # Determine if we should defer.
+    no_defer = (
+        np.ndarray.__array_ufunc__,
+        cls.__array_ufunc__,
+    )
+
+    for item in inputs:
+        higher_priority = (
+            hasattr(item, "__array_priority__")
+            and item.__array_priority__ > self.__array_priority__
+        )
+        has_array_ufunc = (
+            hasattr(item, "__array_ufunc__")
+            and type(item).__array_ufunc__ not in no_defer
+            and not isinstance(item, self._HANDLED_TYPES)
+        )
+        if higher_priority or has_array_ufunc:
+            return NotImplemented
+
+    # align all the inputs.
+    types = tuple(type(x) for x in inputs)
+    alignable = [x for x, t in zip(inputs, types) if issubclass(t, NDFrame)]
+
+    if len(alignable) > 1:
+        # This triggers alignment.
+        # At the moment, there aren't any ufuncs with more than two inputs
+        # so this ends up just being x1.index | x2.index, but we write
+        # it to handle *args.
+        set_types = set(types)
+        if len(set_types) > 1 and {DataFrame, Series}.issubset(set_types):
+            # We currently don't handle ufunc(DataFrame, Series)
+            # well. Previously this raised an internal ValueError. We might
+            # support it someday, so raise a NotImplementedError.
+            raise NotImplementedError(
+                f"Cannot apply ufunc {ufunc} to mixed DataFrame and Series inputs."
+            )
+        axes = self.axes
+        for obj in alignable[1:]:
+            # this relies on the fact that we aren't handling mixed
+            # series / frame ufuncs.
+            for i, (ax1, ax2) in enumerate(zip(axes, obj.axes)):
+                axes[i] = ax1.union(ax2)
+
+        reconstruct_axes = dict(zip(self._AXIS_ORDERS, axes))
+        inputs = tuple(
+            x.reindex(**reconstruct_axes) if issubclass(t, NDFrame) else x
+            for x, t in zip(inputs, types)
+        )
+    else:
+        reconstruct_axes = dict(zip(self._AXIS_ORDERS, self.axes))
+
+    if self.ndim == 1:
+        names = [getattr(x, "name") for x in inputs if hasattr(x, "name")]
+        name = names[0] if len(set(names)) == 1 else None
+        reconstruct_kwargs = {"name": name}
+    else:
+        reconstruct_kwargs = {}
+
+    def reconstruct(result):
+        if ufunc.nout > 1:
+            # np.modf, np.frexp, np.divmod
+            return tuple(_reconstruct(x) for x in result)
+
+        return _reconstruct(result)
+
+    def _reconstruct(result):
+        if lib.is_scalar(result):
+            return result
+
+        if result.ndim != self.ndim:
+            if method == "outer":
+                raise NotImplementedError
+            return result
+        if isinstance(result, (BlockManager, ArrayManager)):
+            # we went through BlockManager.apply e.g. np.sqrt
+            result = self._constructor_from_mgr(result, axes=result.axes)
+        else:
+            # we converted an array, lost our axes
+            result = self._constructor(
+                result, **reconstruct_axes, **reconstruct_kwargs, copy=False
+            )
+        # TODO: When we support multiple values in __finalize__, this
+        # should pass alignable to `__finalize__` instead of self.
+        # Then `np.add(a, b)` would consider attrs from both a and b
+        # when a and b are NDFrames.
+        if len(alignable) == 1:
+            result = result.__finalize__(self)
+        return result
+
+    if "out" in kwargs:
+        # e.g. test_multiindex_get_loc
+        result = dispatch_ufunc_with_out(self, ufunc, method, *inputs, **kwargs)
+        return reconstruct(result)
+
+    if method == "reduce":
+        # e.g. test.series.test_ufunc.test_reduce
+        result = dispatch_reduction_ufunc(self, ufunc, method, *inputs, **kwargs)
+        if result is not NotImplemented:
+            return result
+
+    # We still get here with kwargs `axis` for e.g. np.maximum.accumulate
+    #  and `dtype` and `keepdims` for np.ptp
+
+    if self.ndim > 1 and (len(inputs) > 1 or ufunc.nout > 1):
+        # Just give up on preserving types in the complex case.
+        # In theory we could preserve them for them.
+        # * nout>1 is doable if BlockManager.apply took nout and
+        #   returned a Tuple[BlockManager].
+        # * len(inputs) > 1 is doable when we know that we have
+        #   aligned blocks / dtypes.
+
+        # e.g. my_ufunc, modf, logaddexp, heaviside, subtract, add
+        inputs = tuple(np.asarray(x) for x in inputs)
+        # Note: we can't use default_array_ufunc here bc reindexing means
+        #  that `self` may not be among `inputs`
+        result = getattr(ufunc, method)(*inputs, **kwargs)
+    elif self.ndim == 1:
+        # ufunc(series, ...)
+        inputs = tuple(extract_array(x, extract_numpy=True) for x in inputs)
+        result = getattr(ufunc, method)(*inputs, **kwargs)
+    else:
+        # ufunc(dataframe)
+        if method == "__call__" and not kwargs:
+            # for np.<ufunc>(..) calls
+            # kwargs cannot necessarily be handled block-by-block, so only
+            # take this path if there are no kwargs
+            mgr = inputs[0]._mgr
+            result = mgr.apply(getattr(ufunc, method))
+        else:
+            # otherwise specific ufunc methods (eg np.<ufunc>.accumulate(..))
+            # Those can have an axis keyword and thus can't be called block-by-block
+            result = default_array_ufunc(inputs[0], ufunc, method, *inputs, **kwargs)
+            # e.g. np.negative (only one reached), with "where" and "out" in kwargs
+
+    result = reconstruct(result)
+    return result
+
+
+def _standardize_out_kwarg(**kwargs) -> dict:
+    """
+    If kwargs contain "out1" and "out2", replace that with a tuple "out"
+
+    np.divmod, np.modf, np.frexp can have either `out=(out1, out2)` or
+    `out1=out1, out2=out2)`
+    """
+    if "out" not in kwargs and "out1" in kwargs and "out2" in kwargs:
+        out1 = kwargs.pop("out1")
+        out2 = kwargs.pop("out2")
+        out = (out1, out2)
+        kwargs["out"] = out
+    return kwargs
+
+
+def dispatch_ufunc_with_out(self, ufunc: np.ufunc, method: str, *inputs, **kwargs):
+    """
+    If we have an `out` keyword, then call the ufunc without `out` and then
+    set the result into the given `out`.
+    """
+
+    # Note: we assume _standardize_out_kwarg has already been called.
+    out = kwargs.pop("out")
+    where = kwargs.pop("where", None)
+
+    result = getattr(ufunc, method)(*inputs, **kwargs)
+
+    if result is NotImplemented:
+        return NotImplemented
+
+    if isinstance(result, tuple):
+        # i.e. np.divmod, np.modf, np.frexp
+        if not isinstance(out, tuple) or len(out) != len(result):
+            raise NotImplementedError
+
+        for arr, res in zip(out, result):
+            _assign_where(arr, res, where)
+
+        return out
+
+    if isinstance(out, tuple):
+        if len(out) == 1:
+            out = out[0]
+        else:
+            raise NotImplementedError
+
+    _assign_where(out, result, where)
+    return out
+
+
+def _assign_where(out, result, where) -> None:
+    """
+    Set a ufunc result into 'out', masking with a 'where' argument if necessary.
+    """
+    if where is None:
+        # no 'where' arg passed to ufunc
+        out[:] = result
+    else:
+        np.putmask(out, where, result)
+
+
+def default_array_ufunc(self, ufunc: np.ufunc, method: str, *inputs, **kwargs):
+    """
+    Fallback to the behavior we would get if we did not define __array_ufunc__.
+
+    Notes
+    -----
+    We are assuming that `self` is among `inputs`.
+    """
+    if not any(x is self for x in inputs):
+        raise NotImplementedError
+
+    new_inputs = [x if x is not self else np.asarray(x) for x in inputs]
+
+    return getattr(ufunc, method)(*new_inputs, **kwargs)
+
+
+def dispatch_reduction_ufunc(self, ufunc: np.ufunc, method: str, *inputs, **kwargs):
+    """
+    Dispatch ufunc reductions to self's reduction methods.
+    """
+    assert method == "reduce"
+
+    if len(inputs) != 1 or inputs[0] is not self:
+        return NotImplemented
+
+    if ufunc.__name__ not in REDUCTION_ALIASES:
+        return NotImplemented
+
+    method_name = REDUCTION_ALIASES[ufunc.__name__]
+
+    # NB: we are assuming that min/max represent minimum/maximum methods,
+    #  which would not be accurate for e.g. Timestamp.min
+    if not hasattr(self, method_name):
+        return NotImplemented
+
+    if self.ndim > 1:
+        if isinstance(self, ABCNDFrame):
+            # TODO: test cases where this doesn't hold, i.e. 2D DTA/TDA
+            kwargs["numeric_only"] = False
+
+        if "axis" not in kwargs:
+            # For DataFrame reductions we don't want the default axis=0
+            # Note: np.min is not a ufunc, but uses array_function_dispatch,
+            #  so calls DataFrame.min (without ever getting here) with the np.min
+            #  default of axis=None, which DataFrame.min catches and changes to axis=0.
+            # np.minimum.reduce(df) gets here bc axis is not in kwargs,
+            #  so we set axis=0 to match the behaviorof np.minimum.reduce(df.values)
+            kwargs["axis"] = 0
+
+    # By default, numpy's reductions do not skip NaNs, so we have to
+    #  pass skipna=False
+    return getattr(self, method_name)(skipna=False, **kwargs)

llava_next/lib/python3.10/site-packages/pandas/core/arrays/__init__.py

@@ -0,0 +1,43 @@
+from pandas.core.arrays.arrow import ArrowExtensionArray
+from pandas.core.arrays.base import (
+    ExtensionArray,
+    ExtensionOpsMixin,
+    ExtensionScalarOpsMixin,
+)
+from pandas.core.arrays.boolean import BooleanArray
+from pandas.core.arrays.categorical import Categorical
+from pandas.core.arrays.datetimes import DatetimeArray
+from pandas.core.arrays.floating import FloatingArray
+from pandas.core.arrays.integer import IntegerArray
+from pandas.core.arrays.interval import IntervalArray
+from pandas.core.arrays.masked import BaseMaskedArray
+from pandas.core.arrays.numpy_ import NumpyExtensionArray
+from pandas.core.arrays.period import (
+    PeriodArray,
+    period_array,
+)
+from pandas.core.arrays.sparse import SparseArray
+from pandas.core.arrays.string_ import StringArray
+from pandas.core.arrays.string_arrow import ArrowStringArray
+from pandas.core.arrays.timedeltas import TimedeltaArray
+
+__all__ = [
+    "ArrowExtensionArray",
+    "ExtensionArray",
+    "ExtensionOpsMixin",
+    "ExtensionScalarOpsMixin",
+    "ArrowStringArray",
+    "BaseMaskedArray",
+    "BooleanArray",
+    "Categorical",
+    "DatetimeArray",
+    "FloatingArray",
+    "IntegerArray",
+    "IntervalArray",
+    "NumpyExtensionArray",
+    "PeriodArray",
+    "period_array",
+    "SparseArray",
+    "StringArray",
+    "TimedeltaArray",
+]

llava_next/lib/python3.10/site-packages/pandas/core/arrays/_arrow_string_mixins.py

@@ -0,0 +1,84 @@
+from __future__ import annotations
+
+from typing import Literal
+
+import numpy as np
+
+from pandas.compat import pa_version_under10p1
+
+if not pa_version_under10p1:
+    import pyarrow as pa
+    import pyarrow.compute as pc
+
+
+class ArrowStringArrayMixin:
+    _pa_array = None
+
+    def __init__(self, *args, **kwargs) -> None:
+        raise NotImplementedError
+
+    def _str_pad(
+        self,
+        width: int,
+        side: Literal["left", "right", "both"] = "left",
+        fillchar: str = " ",
+    ):
+        if side == "left":
+            pa_pad = pc.utf8_lpad
+        elif side == "right":
+            pa_pad = pc.utf8_rpad
+        elif side == "both":
+            pa_pad = pc.utf8_center
+        else:
+            raise ValueError(
+                f"Invalid side: {side}. Side must be one of 'left', 'right', 'both'"
+            )
+        return type(self)(pa_pad(self._pa_array, width=width, padding=fillchar))
+
+    def _str_get(self, i: int):
+        lengths = pc.utf8_length(self._pa_array)
+        if i >= 0:
+            out_of_bounds = pc.greater_equal(i, lengths)
+            start = i
+            stop = i + 1
+            step = 1
+        else:
+            out_of_bounds = pc.greater(-i, lengths)
+            start = i
+            stop = i - 1
+            step = -1
+        not_out_of_bounds = pc.invert(out_of_bounds.fill_null(True))
+        selected = pc.utf8_slice_codeunits(
+            self._pa_array, start=start, stop=stop, step=step
+        )
+        null_value = pa.scalar(
+            None, type=self._pa_array.type  # type: ignore[attr-defined]
+        )
+        result = pc.if_else(not_out_of_bounds, selected, null_value)
+        return type(self)(result)
+
+    def _str_slice_replace(
+        self, start: int | None = None, stop: int | None = None, repl: str | None = None
+    ):
+        if repl is None:
+            repl = ""
+        if start is None:
+            start = 0
+        if stop is None:
+            stop = np.iinfo(np.int64).max
+        return type(self)(pc.utf8_replace_slice(self._pa_array, start, stop, repl))
+
+    def _str_capitalize(self):
+        return type(self)(pc.utf8_capitalize(self._pa_array))
+
+    def _str_title(self):
+        return type(self)(pc.utf8_title(self._pa_array))
+
+    def _str_swapcase(self):
+        return type(self)(pc.utf8_swapcase(self._pa_array))
+
+    def _str_removesuffix(self, suffix: str):
+        ends_with = pc.ends_with(self._pa_array, pattern=suffix)
+        removed = pc.utf8_slice_codeunits(self._pa_array, 0, stop=-len(suffix))
+        result = pc.if_else(ends_with, removed, self._pa_array)
+        return type(self)(result)

llava_next/lib/python3.10/site-packages/pandas/core/arrays/_mixins.py

@@ -0,0 +1,547 @@
+from __future__ import annotations
+
+from functools import wraps
+from typing import (
+    TYPE_CHECKING,
+    Any,
+    Literal,
+    cast,
+    overload,
+)
+
+import numpy as np
+
+from pandas._libs import lib
+from pandas._libs.arrays import NDArrayBacked
+from pandas._libs.tslibs import is_supported_dtype
+from pandas._typing import (
+    ArrayLike,
+    AxisInt,
+    Dtype,
+    F,
+    FillnaOptions,
+    PositionalIndexer2D,
+    PositionalIndexerTuple,
+    ScalarIndexer,
+    Self,
+    SequenceIndexer,
+    Shape,
+    TakeIndexer,
+    npt,
+)
+from pandas.errors import AbstractMethodError
+from pandas.util._decorators import doc
+from pandas.util._validators import (
+    validate_bool_kwarg,
+    validate_fillna_kwargs,
+    validate_insert_loc,
+)
+
+from pandas.core.dtypes.common import pandas_dtype
+from pandas.core.dtypes.dtypes import (
+    DatetimeTZDtype,
+    ExtensionDtype,
+    PeriodDtype,
+)
+from pandas.core.dtypes.missing import array_equivalent
+
+from pandas.core import missing
+from pandas.core.algorithms import (
+    take,
+    unique,
+    value_counts_internal as value_counts,
+)
+from pandas.core.array_algos.quantile import quantile_with_mask
+from pandas.core.array_algos.transforms import shift
+from pandas.core.arrays.base import ExtensionArray
+from pandas.core.construction import extract_array
+from pandas.core.indexers import check_array_indexer
+from pandas.core.sorting import nargminmax
+
+if TYPE_CHECKING:
+    from collections.abc import Sequence
+
+    from pandas._typing import (
+        NumpySorter,
+        NumpyValueArrayLike,
+    )
+
+    from pandas import Series
+
+
+def ravel_compat(meth: F) -> F:
+    """
+    Decorator to ravel a 2D array before passing it to a cython operation,
+    then reshape the result to our own shape.
+    """
+
+    @wraps(meth)
+    def method(self, *args, **kwargs):
+        if self.ndim == 1:
+            return meth(self, *args, **kwargs)
+
+        flags = self._ndarray.flags
+        flat = self.ravel("K")
+        result = meth(flat, *args, **kwargs)
+        order = "F" if flags.f_contiguous else "C"
+        return result.reshape(self.shape, order=order)
+
+    return cast(F, method)
+
+
+class NDArrayBackedExtensionArray(NDArrayBacked, ExtensionArray):
+    """
+    ExtensionArray that is backed by a single NumPy ndarray.
+    """
+
+    _ndarray: np.ndarray
+
+    # scalar used to denote NA value inside our self._ndarray, e.g. -1
+    #  for Categorical, iNaT for Period. Outside of object dtype,
+    #  self.isna() should be exactly locations in self._ndarray with
+    #  _internal_fill_value.
+    _internal_fill_value: Any
+
+    def _box_func(self, x):
+        """
+        Wrap numpy type in our dtype.type if necessary.
+        """
+        return x
+
+    def _validate_scalar(self, value):
+        # used by NDArrayBackedExtensionIndex.insert
+        raise AbstractMethodError(self)
+
+    # ------------------------------------------------------------------------
+
+    def view(self, dtype: Dtype | None = None) -> ArrayLike:
+        # We handle datetime64, datetime64tz, timedelta64, and period
+        #  dtypes here. Everything else we pass through to the underlying
+        #  ndarray.
+        if dtype is None or dtype is self.dtype:
+            return self._from_backing_data(self._ndarray)
+
+        if isinstance(dtype, type):
+            # we sometimes pass non-dtype objects, e.g np.ndarray;
+            #  pass those through to the underlying ndarray
+            return self._ndarray.view(dtype)
+
+        dtype = pandas_dtype(dtype)
+        arr = self._ndarray
+
+        if isinstance(dtype, PeriodDtype):
+            cls = dtype.construct_array_type()
+            return cls(arr.view("i8"), dtype=dtype)
+        elif isinstance(dtype, DatetimeTZDtype):
+            dt_cls = dtype.construct_array_type()
+            dt64_values = arr.view(f"M8[{dtype.unit}]")
+            return dt_cls._simple_new(dt64_values, dtype=dtype)
+        elif lib.is_np_dtype(dtype, "M") and is_supported_dtype(dtype):
+            from pandas.core.arrays import DatetimeArray
+
+            dt64_values = arr.view(dtype)
+            return DatetimeArray._simple_new(dt64_values, dtype=dtype)
+
+        elif lib.is_np_dtype(dtype, "m") and is_supported_dtype(dtype):
+            from pandas.core.arrays import TimedeltaArray
+
+            td64_values = arr.view(dtype)
+            return TimedeltaArray._simple_new(td64_values, dtype=dtype)
+
+        # error: Argument "dtype" to "view" of "_ArrayOrScalarCommon" has incompatible
+        # type "Union[ExtensionDtype, dtype[Any]]"; expected "Union[dtype[Any], None,
+        # type, _SupportsDType, str, Union[Tuple[Any, int], Tuple[Any, Union[int,
+        # Sequence[int]]], List[Any], _DTypeDict, Tuple[Any, Any]]]"
+        return arr.view(dtype=dtype)  # type: ignore[arg-type]
+
+    def take(
+        self,
+        indices: TakeIndexer,
+        *,
+        allow_fill: bool = False,
+        fill_value: Any = None,
+        axis: AxisInt = 0,
+    ) -> Self:
+        if allow_fill:
+            fill_value = self._validate_scalar(fill_value)
+
+        new_data = take(
+            self._ndarray,
+            indices,
+            allow_fill=allow_fill,
+            fill_value=fill_value,
+            axis=axis,
+        )
+        return self._from_backing_data(new_data)
+
+    # ------------------------------------------------------------------------
+
+    def equals(self, other) -> bool:
+        if type(self) is not type(other):
+            return False
+        if self.dtype != other.dtype:
+            return False
+        return bool(array_equivalent(self._ndarray, other._ndarray, dtype_equal=True))
+
+    @classmethod
+    def _from_factorized(cls, values, original):
+        assert values.dtype == original._ndarray.dtype
+        return original._from_backing_data(values)
+
+    def _values_for_argsort(self) -> np.ndarray:
+        return self._ndarray
+
+    def _values_for_factorize(self):
+        return self._ndarray, self._internal_fill_value
+
+    def _hash_pandas_object(
+        self, *, encoding: str, hash_key: str, categorize: bool
+    ) -> npt.NDArray[np.uint64]:
+        from pandas.core.util.hashing import hash_array
+
+        values = self._ndarray
+        return hash_array(
+            values, encoding=encoding, hash_key=hash_key, categorize=categorize
+        )
+
+    # Signature of "argmin" incompatible with supertype "ExtensionArray"
+    def argmin(self, axis: AxisInt = 0, skipna: bool = True):  # type: ignore[override]
+        # override base class by adding axis keyword
+        validate_bool_kwarg(skipna, "skipna")
+        if not skipna and self._hasna:
+            raise NotImplementedError
+        return nargminmax(self, "argmin", axis=axis)
+
+    # Signature of "argmax" incompatible with supertype "ExtensionArray"
+    def argmax(self, axis: AxisInt = 0, skipna: bool = True):  # type: ignore[override]
+        # override base class by adding axis keyword
+        validate_bool_kwarg(skipna, "skipna")
+        if not skipna and self._hasna:
+            raise NotImplementedError
+        return nargminmax(self, "argmax", axis=axis)
+
+    def unique(self) -> Self:
+        new_data = unique(self._ndarray)
+        return self._from_backing_data(new_data)
+
+    @classmethod
+    @doc(ExtensionArray._concat_same_type)
+    def _concat_same_type(
+        cls,
+        to_concat: Sequence[Self],
+        axis: AxisInt = 0,
+    ) -> Self:
+        if not lib.dtypes_all_equal([x.dtype for x in to_concat]):
+            dtypes = {str(x.dtype) for x in to_concat}
+            raise ValueError("to_concat must have the same dtype", dtypes)
+
+        return super()._concat_same_type(to_concat, axis=axis)
+
+    @doc(ExtensionArray.searchsorted)
+    def searchsorted(
+        self,
+        value: NumpyValueArrayLike | ExtensionArray,
+        side: Literal["left", "right"] = "left",
+        sorter: NumpySorter | None = None,
+    ) -> npt.NDArray[np.intp] | np.intp:
+        npvalue = self._validate_setitem_value(value)
+        return self._ndarray.searchsorted(npvalue, side=side, sorter=sorter)
+
+    @doc(ExtensionArray.shift)
+    def shift(self, periods: int = 1, fill_value=None):
+        # NB: shift is always along axis=0
+        axis = 0
+        fill_value = self._validate_scalar(fill_value)
+        new_values = shift(self._ndarray, periods, axis, fill_value)
+
+        return self._from_backing_data(new_values)
+
+    def __setitem__(self, key, value) -> None:
+        key = check_array_indexer(self, key)
+        value = self._validate_setitem_value(value)
+        self._ndarray[key] = value
+
+    def _validate_setitem_value(self, value):
+        return value
+
+    @overload
+    def __getitem__(self, key: ScalarIndexer) -> Any:
+        ...
+
+    @overload
+    def __getitem__(
+        self,
+        key: SequenceIndexer | PositionalIndexerTuple,
+    ) -> Self:
+        ...
+
+    def __getitem__(
+        self,
+        key: PositionalIndexer2D,
+    ) -> Self | Any:
+        if lib.is_integer(key):
+            # fast-path
+            result = self._ndarray[key]
+            if self.ndim == 1:
+                return self._box_func(result)
+            return self._from_backing_data(result)
+
+        # error: Incompatible types in assignment (expression has type "ExtensionArray",
+        # variable has type "Union[int, slice, ndarray]")
+        key = extract_array(key, extract_numpy=True)  # type: ignore[assignment]
+        key = check_array_indexer(self, key)
+        result = self._ndarray[key]
+        if lib.is_scalar(result):
+            return self._box_func(result)
+
+        result = self._from_backing_data(result)
+        return result
+
+    def _fill_mask_inplace(
+        self, method: str, limit: int | None, mask: npt.NDArray[np.bool_]
+    ) -> None:
+        # (for now) when self.ndim == 2, we assume axis=0
+        func = missing.get_fill_func(method, ndim=self.ndim)
+        func(self._ndarray.T, limit=limit, mask=mask.T)
+
+    def _pad_or_backfill(
+        self,
+        *,
+        method: FillnaOptions,
+        limit: int | None = None,
+        limit_area: Literal["inside", "outside"] | None = None,
+        copy: bool = True,
+    ) -> Self:
+        mask = self.isna()
+        if mask.any():
+            # (for now) when self.ndim == 2, we assume axis=0
+            func = missing.get_fill_func(method, ndim=self.ndim)
+
+            npvalues = self._ndarray.T
+            if copy:
+                npvalues = npvalues.copy()
+            func(npvalues, limit=limit, limit_area=limit_area, mask=mask.T)
+            npvalues = npvalues.T
+
+            if copy:
+                new_values = self._from_backing_data(npvalues)
+            else:
+                new_values = self
+
+        else:
+            if copy:
+                new_values = self.copy()
+            else:
+                new_values = self
+        return new_values
+
+    @doc(ExtensionArray.fillna)
+    def fillna(
+        self, value=None, method=None, limit: int | None = None, copy: bool = True
+    ) -> Self:
+        value, method = validate_fillna_kwargs(
+            value, method, validate_scalar_dict_value=False
+        )
+
+        mask = self.isna()
+        # error: Argument 2 to "check_value_size" has incompatible type
+        # "ExtensionArray"; expected "ndarray"
+        value = missing.check_value_size(
+            value, mask, len(self)  # type: ignore[arg-type]
+        )
+
+        if mask.any():
+            if method is not None:
+                # (for now) when self.ndim == 2, we assume axis=0
+                func = missing.get_fill_func(method, ndim=self.ndim)
+                npvalues = self._ndarray.T
+                if copy:
+                    npvalues = npvalues.copy()
+                func(npvalues, limit=limit, mask=mask.T)
+                npvalues = npvalues.T
+
+                # TODO: NumpyExtensionArray didn't used to copy, need tests
+                #  for this
+                new_values = self._from_backing_data(npvalues)
+            else:
+                # fill with value
+                if copy:
+                    new_values = self.copy()
+                else:
+                    new_values = self[:]
+                new_values[mask] = value
+        else:
+            # We validate the fill_value even if there is nothing to fill
+            if value is not None:
+                self._validate_setitem_value(value)
+
+            if not copy:
+                new_values = self[:]
+            else:
+                new_values = self.copy()
+        return new_values
+
+    # ------------------------------------------------------------------------
+    # Reductions
+
+    def _wrap_reduction_result(self, axis: AxisInt | None, result):
+        if axis is None or self.ndim == 1:
+            return self._box_func(result)
+        return self._from_backing_data(result)
+
+    # ------------------------------------------------------------------------
+    # __array_function__ methods
+
+    def _putmask(self, mask: npt.NDArray[np.bool_], value) -> None:
+        """
+        Analogue to np.putmask(self, mask, value)
+
+        Parameters
+        ----------
+        mask : np.ndarray[bool]
+        value : scalar or listlike
+
+        Raises
+        ------
+        TypeError
+            If value cannot be cast to self.dtype.
+        """
+        value = self._validate_setitem_value(value)
+
+        np.putmask(self._ndarray, mask, value)
+
+    def _where(self: Self, mask: npt.NDArray[np.bool_], value) -> Self:
+        """
+        Analogue to np.where(mask, self, value)
+
+        Parameters
+        ----------
+        mask : np.ndarray[bool]
+        value : scalar or listlike
+
+        Raises
+        ------
+        TypeError
+            If value cannot be cast to self.dtype.
+        """
+        value = self._validate_setitem_value(value)
+
+        res_values = np.where(mask, self._ndarray, value)
+        if res_values.dtype != self._ndarray.dtype:
+            raise AssertionError(
+                # GH#56410
+                "Something has gone wrong, please report a bug at "
+                "github.com/pandas-dev/pandas/"
+            )
+        return self._from_backing_data(res_values)
+
+    # ------------------------------------------------------------------------
+    # Index compat methods
+
+    def insert(self, loc: int, item) -> Self:
+        """
+        Make new ExtensionArray inserting new item at location. Follows
+        Python list.append semantics for negative values.
+
+        Parameters
+        ----------
+        loc : int
+        item : object
+
+        Returns
+        -------
+        type(self)
+        """
+        loc = validate_insert_loc(loc, len(self))
+
+        code = self._validate_scalar(item)
+
+        new_vals = np.concatenate(
+            (
+                self._ndarray[:loc],
+                np.asarray([code], dtype=self._ndarray.dtype),
+                self._ndarray[loc:],
+            )
+        )
+        return self._from_backing_data(new_vals)
+
+    # ------------------------------------------------------------------------
+    # Additional array methods
+    #  These are not part of the EA API, but we implement them because
+    #  pandas assumes they're there.
+
+    def value_counts(self, dropna: bool = True) -> Series:
+        """
+        Return a Series containing counts of unique values.
+
+        Parameters
+        ----------
+        dropna : bool, default True
+            Don't include counts of NA values.
+
+        Returns
+        -------
+        Series
+        """
+        if self.ndim != 1:
+            raise NotImplementedError
+
+        from pandas import (
+            Index,
+            Series,
+        )
+
+        if dropna:
+            # error: Unsupported operand type for ~ ("ExtensionArray")
+            values = self[~self.isna()]._ndarray  # type: ignore[operator]
+        else:
+            values = self._ndarray
+
+        result = value_counts(values, sort=False, dropna=dropna)
+
+        index_arr = self._from_backing_data(np.asarray(result.index._data))
+        index = Index(index_arr, name=result.index.name)
+        return Series(result._values, index=index, name=result.name, copy=False)
+
+    def _quantile(
+        self,
+        qs: npt.NDArray[np.float64],
+        interpolation: str,
+    ) -> Self:
+        # TODO: disable for Categorical if not ordered?
+
+        mask = np.asarray(self.isna())
+        arr = self._ndarray
+        fill_value = self._internal_fill_value
+
+        res_values = quantile_with_mask(arr, mask, fill_value, qs, interpolation)
+
+        res_values = self._cast_quantile_result(res_values)
+        return self._from_backing_data(res_values)
+
+    # TODO: see if we can share this with other dispatch-wrapping methods
+    def _cast_quantile_result(self, res_values: np.ndarray) -> np.ndarray:
+        """
+        Cast the result of quantile_with_mask to an appropriate dtype
+        to pass to _from_backing_data in _quantile.
+        """
+        return res_values
+
+    # ------------------------------------------------------------------------
+    # numpy-like methods
+
+    @classmethod
+    def _empty(cls, shape: Shape, dtype: ExtensionDtype) -> Self:
+        """
+        Analogous to np.empty(shape, dtype=dtype)
+
+        Parameters
+        ----------
+        shape : tuple[int]
+        dtype : ExtensionDtype
+        """
+        # The base implementation uses a naive approach to find the dtype
+        #  for the backing ndarray
+        arr = cls._from_sequence([], dtype=dtype)
+        backing = np.empty(shape, dtype=arr._ndarray.dtype)
+        return arr._from_backing_data(backing)

llava_next/lib/python3.10/site-packages/pandas/core/arrays/_ranges.py

@@ -0,0 +1,207 @@
+"""
+Helper functions to generate range-like data for DatetimeArray
+(and possibly TimedeltaArray/PeriodArray)
+"""
+from __future__ import annotations
+
+from typing import TYPE_CHECKING
+
+import numpy as np
+
+from pandas._libs.lib import i8max
+from pandas._libs.tslibs import (
+    BaseOffset,
+    OutOfBoundsDatetime,
+    Timedelta,
+    Timestamp,
+    iNaT,
+)
+
+if TYPE_CHECKING:
+    from pandas._typing import npt
+
+
+def generate_regular_range(
+    start: Timestamp | Timedelta | None,
+    end: Timestamp | Timedelta | None,
+    periods: int | None,
+    freq: BaseOffset,
+    unit: str = "ns",
+) -> npt.NDArray[np.intp]:
+    """
+    Generate a range of dates or timestamps with the spans between dates
+    described by the given `freq` DateOffset.
+
+    Parameters
+    ----------
+    start : Timedelta, Timestamp or None
+        First point of produced date range.
+    end : Timedelta, Timestamp or None
+        Last point of produced date range.
+    periods : int or None
+        Number of periods in produced date range.
+    freq : Tick
+        Describes space between dates in produced date range.
+    unit : str, default "ns"
+        The resolution the output is meant to represent.
+
+    Returns
+    -------
+    ndarray[np.int64]
+        Representing the given resolution.
+    """
+    istart = start._value if start is not None else None
+    iend = end._value if end is not None else None
+    freq.nanos  # raises if non-fixed frequency
+    td = Timedelta(freq)
+    b: int
+    e: int
+    try:
+        td = td.as_unit(unit, round_ok=False)
+    except ValueError as err:
+        raise ValueError(
+            f"freq={freq} is incompatible with unit={unit}. "
+            "Use a lower freq or a higher unit instead."
+        ) from err
+    stride = int(td._value)
+
+    if periods is None and istart is not None and iend is not None:
+        b = istart
+        # cannot just use e = Timestamp(end) + 1 because arange breaks when
+        # stride is too large, see GH10887
+        e = b + (iend - b) // stride * stride + stride // 2 + 1
+    elif istart is not None and periods is not None:
+        b = istart
+        e = _generate_range_overflow_safe(b, periods, stride, side="start")
+    elif iend is not None and periods is not None:
+        e = iend + stride
+        b = _generate_range_overflow_safe(e, periods, stride, side="end")
+    else:
+        raise ValueError(
+            "at least 'start' or 'end' should be specified if a 'period' is given."
+        )
+
+    with np.errstate(over="raise"):
+        # If the range is sufficiently large, np.arange may overflow
+        #  and incorrectly return an empty array if not caught.
+        try:
+            values = np.arange(b, e, stride, dtype=np.int64)
+        except FloatingPointError:
+            xdr = [b]
+            while xdr[-1] != e:
+                xdr.append(xdr[-1] + stride)
+            values = np.array(xdr[:-1], dtype=np.int64)
+    return values
+
+
+def _generate_range_overflow_safe(
+    endpoint: int, periods: int, stride: int, side: str = "start"
+) -> int:
+    """
+    Calculate the second endpoint for passing to np.arange, checking
+    to avoid an integer overflow.  Catch OverflowError and re-raise
+    as OutOfBoundsDatetime.
+
+    Parameters
+    ----------
+    endpoint : int
+        nanosecond timestamp of the known endpoint of the desired range
+    periods : int
+        number of periods in the desired range
+    stride : int
+        nanoseconds between periods in the desired range
+    side : {'start', 'end'}
+        which end of the range `endpoint` refers to
+
+    Returns
+    -------
+    other_end : int
+
+    Raises
+    ------
+    OutOfBoundsDatetime
+    """
+    # GH#14187 raise instead of incorrectly wrapping around
+    assert side in ["start", "end"]
+
+    i64max = np.uint64(i8max)
+    msg = f"Cannot generate range with {side}={endpoint} and periods={periods}"
+
+    with np.errstate(over="raise"):
+        # if periods * strides cannot be multiplied within the *uint64* bounds,
+        #  we cannot salvage the operation by recursing, so raise
+        try:
+            addend = np.uint64(periods) * np.uint64(np.abs(stride))
+        except FloatingPointError as err:
+            raise OutOfBoundsDatetime(msg) from err
+
+    if np.abs(addend) <= i64max:
+        # relatively easy case without casting concerns
+        return _generate_range_overflow_safe_signed(endpoint, periods, stride, side)
+
+    elif (endpoint > 0 and side == "start" and stride > 0) or (
+        endpoint < 0 < stride and side == "end"
+    ):
+        # no chance of not-overflowing
+        raise OutOfBoundsDatetime(msg)
+
+    elif side == "end" and endpoint - stride <= i64max < endpoint:
+        # in _generate_regular_range we added `stride` thereby overflowing
+        #  the bounds.  Adjust to fix this.
+        return _generate_range_overflow_safe(
+            endpoint - stride, periods - 1, stride, side
+        )
+
+    # split into smaller pieces
+    mid_periods = periods // 2
+    remaining = periods - mid_periods
+    assert 0 < remaining < periods, (remaining, periods, endpoint, stride)
+
+    midpoint = int(_generate_range_overflow_safe(endpoint, mid_periods, stride, side))
+    return _generate_range_overflow_safe(midpoint, remaining, stride, side)
+
+
+def _generate_range_overflow_safe_signed(
+    endpoint: int, periods: int, stride: int, side: str
+) -> int:
+    """
+    A special case for _generate_range_overflow_safe where `periods * stride`
+    can be calculated without overflowing int64 bounds.
+    """
+    assert side in ["start", "end"]
+    if side == "end":
+        stride *= -1
+
+    with np.errstate(over="raise"):
+        addend = np.int64(periods) * np.int64(stride)
+        try:
+            # easy case with no overflows
+            result = np.int64(endpoint) + addend
+            if result == iNaT:
+                # Putting this into a DatetimeArray/TimedeltaArray
+                #  would incorrectly be interpreted as NaT
+                raise OverflowError
+            return int(result)
+        except (FloatingPointError, OverflowError):
+            # with endpoint negative and addend positive we risk
+            #  FloatingPointError; with reversed signed we risk OverflowError
+            pass
+
+        # if stride and endpoint had opposite signs, then endpoint + addend
+        #  should never overflow.  so they must have the same signs
+        assert (stride > 0 and endpoint >= 0) or (stride < 0 and endpoint <= 0)
+
+        if stride > 0:
+            # watch out for very special case in which we just slightly
+            #  exceed implementation bounds, but when passing the result to
+            #  np.arange will get a result slightly within the bounds
+
+            uresult = np.uint64(endpoint) + np.uint64(addend)
+            i64max = np.uint64(i8max)
+            assert uresult > i64max
+            if uresult <= i64max + np.uint64(stride):
+                return int(uresult)
+
+    raise OutOfBoundsDatetime(
+        f"Cannot generate range with {side}={endpoint} and periods={periods}"
+    )

llava_next/lib/python3.10/site-packages/pandas/core/arrays/_utils.py

@@ -0,0 +1,63 @@
+from __future__ import annotations
+
+from typing import (
+    TYPE_CHECKING,
+    Any,
+)
+
+import numpy as np
+
+from pandas._libs import lib
+from pandas.errors import LossySetitemError
+
+from pandas.core.dtypes.cast import np_can_hold_element
+from pandas.core.dtypes.common import is_numeric_dtype
+
+if TYPE_CHECKING:
+    from pandas._typing import (
+        ArrayLike,
+        npt,
+    )
+
+
+def to_numpy_dtype_inference(
+    arr: ArrayLike, dtype: npt.DTypeLike | None, na_value, hasna: bool
+) -> tuple[npt.DTypeLike, Any]:
+    if dtype is None and is_numeric_dtype(arr.dtype):
+        dtype_given = False
+        if hasna:
+            if arr.dtype.kind == "b":
+                dtype = np.dtype(np.object_)
+            else:
+                if arr.dtype.kind in "iu":
+                    dtype = np.dtype(np.float64)
+                else:
+                    dtype = arr.dtype.numpy_dtype  # type: ignore[union-attr]
+                if na_value is lib.no_default:
+                    na_value = np.nan
+        else:
+            dtype = arr.dtype.numpy_dtype  # type: ignore[union-attr]
+    elif dtype is not None:
+        dtype = np.dtype(dtype)
+        dtype_given = True
+    else:
+        dtype_given = True
+
+    if na_value is lib.no_default:
+        if dtype is None or not hasna:
+            na_value = arr.dtype.na_value
+        elif dtype.kind == "f":  # type: ignore[union-attr]
+            na_value = np.nan
+        elif dtype.kind == "M":  # type: ignore[union-attr]
+            na_value = np.datetime64("nat")
+        elif dtype.kind == "m":  # type: ignore[union-attr]
+            na_value = np.timedelta64("nat")
+        else:
+            na_value = arr.dtype.na_value
+
+    if not dtype_given and hasna:
+        try:
+            np_can_hold_element(dtype, na_value)  # type: ignore[arg-type]
+        except LossySetitemError:
+            dtype = np.dtype(np.object_)
+    return dtype, na_value

llava_next/lib/python3.10/site-packages/pandas/core/arrays/datetimelike.py

@@ -0,0 +1,2556 @@
+from __future__ import annotations
+
+from datetime import (
+    datetime,
+    timedelta,
+)
+from functools import wraps
+import operator
+from typing import (
+    TYPE_CHECKING,
+    Any,
+    Callable,
+    Literal,
+    Union,
+    cast,
+    final,
+    overload,
+)
+import warnings
+
+import numpy as np
+
+from pandas._libs import (
+    algos,
+    lib,
+)
+from pandas._libs.arrays import NDArrayBacked
+from pandas._libs.tslibs import (
+    BaseOffset,
+    IncompatibleFrequency,
+    NaT,
+    NaTType,
+    Period,
+    Resolution,
+    Tick,
+    Timedelta,
+    Timestamp,
+    add_overflowsafe,
+    astype_overflowsafe,
+    get_unit_from_dtype,
+    iNaT,
+    ints_to_pydatetime,
+    ints_to_pytimedelta,
+    periods_per_day,
+    to_offset,
+)
+from pandas._libs.tslibs.fields import (
+    RoundTo,
+    round_nsint64,
+)
+from pandas._libs.tslibs.np_datetime import compare_mismatched_resolutions
+from pandas._libs.tslibs.timedeltas import get_unit_for_round
+from pandas._libs.tslibs.timestamps import integer_op_not_supported
+from pandas._typing import (
+    ArrayLike,
+    AxisInt,
+    DatetimeLikeScalar,
+    Dtype,
+    DtypeObj,
+    F,
+    InterpolateOptions,
+    NpDtype,
+    PositionalIndexer2D,
+    PositionalIndexerTuple,
+    ScalarIndexer,
+    Self,
+    SequenceIndexer,
+    TimeAmbiguous,
+    TimeNonexistent,
+    npt,
+)
+from pandas.compat.numpy import function as nv
+from pandas.errors import (
+    AbstractMethodError,
+    InvalidComparison,
+    PerformanceWarning,
+)
+from pandas.util._decorators import (
+    Appender,
+    Substitution,
+    cache_readonly,
+)
+from pandas.util._exceptions import find_stack_level
+
+from pandas.core.dtypes.cast import construct_1d_object_array_from_listlike
+from pandas.core.dtypes.common import (
+    is_all_strings,
+    is_integer_dtype,
+    is_list_like,
+    is_object_dtype,
+    is_string_dtype,
+    pandas_dtype,
+)
+from pandas.core.dtypes.dtypes import (
+    ArrowDtype,
+    CategoricalDtype,
+    DatetimeTZDtype,
+    ExtensionDtype,
+    PeriodDtype,
+)
+from pandas.core.dtypes.generic import (
+    ABCCategorical,
+    ABCMultiIndex,
+)
+from pandas.core.dtypes.missing import (
+    is_valid_na_for_dtype,
+    isna,
+)
+
+from pandas.core import (
+    algorithms,
+    missing,
+    nanops,
+    ops,
+)
+from pandas.core.algorithms import (
+    isin,
+    map_array,
+    unique1d,
+)
+from pandas.core.array_algos import datetimelike_accumulations
+from pandas.core.arraylike import OpsMixin
+from pandas.core.arrays._mixins import (
+    NDArrayBackedExtensionArray,
+    ravel_compat,
+)
+from pandas.core.arrays.arrow.array import ArrowExtensionArray
+from pandas.core.arrays.base import ExtensionArray
+from pandas.core.arrays.integer import IntegerArray
+import pandas.core.common as com
+from pandas.core.construction import (
+    array as pd_array,
+    ensure_wrapped_if_datetimelike,
+    extract_array,
+)
+from pandas.core.indexers import (
+    check_array_indexer,
+    check_setitem_lengths,
+)
+from pandas.core.ops.common import unpack_zerodim_and_defer
+from pandas.core.ops.invalid import (
+    invalid_comparison,
+    make_invalid_op,
+)
+
+from pandas.tseries import frequencies
+
+if TYPE_CHECKING:
+    from collections.abc import (
+        Iterator,
+        Sequence,
+    )
+
+    from pandas import Index
+    from pandas.core.arrays import (
+        DatetimeArray,
+        PeriodArray,
+        TimedeltaArray,
+    )
+
+DTScalarOrNaT = Union[DatetimeLikeScalar, NaTType]
+
+
+def _make_unpacked_invalid_op(op_name: str):
+    op = make_invalid_op(op_name)
+    return unpack_zerodim_and_defer(op_name)(op)
+
+
+def _period_dispatch(meth: F) -> F:
+    """
+    For PeriodArray methods, dispatch to DatetimeArray and re-wrap the results
+    in PeriodArray.  We cannot use ._ndarray directly for the affected
+    methods because the i8 data has different semantics on NaT values.
+    """
+
+    @wraps(meth)
+    def new_meth(self, *args, **kwargs):
+        if not isinstance(self.dtype, PeriodDtype):
+            return meth(self, *args, **kwargs)
+
+        arr = self.view("M8[ns]")
+        result = meth(arr, *args, **kwargs)
+        if result is NaT:
+            return NaT
+        elif isinstance(result, Timestamp):
+            return self._box_func(result._value)
+
+        res_i8 = result.view("i8")
+        return self._from_backing_data(res_i8)
+
+    return cast(F, new_meth)
+
+
+# error: Definition of "_concat_same_type" in base class "NDArrayBacked" is
+# incompatible with definition in base class "ExtensionArray"
+class DatetimeLikeArrayMixin(  # type: ignore[misc]
+    OpsMixin, NDArrayBackedExtensionArray
+):
+    """
+    Shared Base/Mixin class for DatetimeArray, TimedeltaArray, PeriodArray
+
+    Assumes that __new__/__init__ defines:
+        _ndarray
+
+    and that inheriting subclass implements:
+        freq
+    """
+
+    # _infer_matches -> which infer_dtype strings are close enough to our own
+    _infer_matches: tuple[str, ...]
+    _is_recognized_dtype: Callable[[DtypeObj], bool]
+    _recognized_scalars: tuple[type, ...]
+    _ndarray: np.ndarray
+    freq: BaseOffset | None
+
+    @cache_readonly
+    def _can_hold_na(self) -> bool:
+        return True
+
+    def __init__(
+        self, data, dtype: Dtype | None = None, freq=None, copy: bool = False
+    ) -> None:
+        raise AbstractMethodError(self)
+
+    @property
+    def _scalar_type(self) -> type[DatetimeLikeScalar]:
+        """
+        The scalar associated with this datelike
+
+        * PeriodArray : Period
+        * DatetimeArray : Timestamp
+        * TimedeltaArray : Timedelta
+        """
+        raise AbstractMethodError(self)
+
+    def _scalar_from_string(self, value: str) -> DTScalarOrNaT:
+        """
+        Construct a scalar type from a string.
+
+        Parameters
+        ----------
+        value : str
+
+        Returns
+        -------
+        Period, Timestamp, or Timedelta, or NaT
+            Whatever the type of ``self._scalar_type`` is.
+
+        Notes
+        -----
+        This should call ``self._check_compatible_with`` before
+        unboxing the result.
+        """
+        raise AbstractMethodError(self)
+
+    def _unbox_scalar(
+        self, value: DTScalarOrNaT
+    ) -> np.int64 | np.datetime64 | np.timedelta64:
+        """
+        Unbox the integer value of a scalar `value`.
+
+        Parameters
+        ----------
+        value : Period, Timestamp, Timedelta, or NaT
+            Depending on subclass.
+
+        Returns
+        -------
+        int
+
+        Examples
+        --------
+        >>> arr = pd.array(np.array(['1970-01-01'], 'datetime64[ns]'))
+        >>> arr._unbox_scalar(arr[0])
+        numpy.datetime64('1970-01-01T00:00:00.000000000')
+        """
+        raise AbstractMethodError(self)
+
+    def _check_compatible_with(self, other: DTScalarOrNaT) -> None:
+        """
+        Verify that `self` and `other` are compatible.
+
+        * DatetimeArray verifies that the timezones (if any) match
+        * PeriodArray verifies that the freq matches
+        * Timedelta has no verification
+
+        In each case, NaT is considered compatible.
+
+        Parameters
+        ----------
+        other
+
+        Raises
+        ------
+        Exception
+        """
+        raise AbstractMethodError(self)
+
+    # ------------------------------------------------------------------
+
+    def _box_func(self, x):
+        """
+        box function to get object from internal representation
+        """
+        raise AbstractMethodError(self)
+
+    def _box_values(self, values) -> np.ndarray:
+        """
+        apply box func to passed values
+        """
+        return lib.map_infer(values, self._box_func, convert=False)
+
+    def __iter__(self) -> Iterator:
+        if self.ndim > 1:
+            return (self[n] for n in range(len(self)))
+        else:
+            return (self._box_func(v) for v in self.asi8)
+
+    @property
+    def asi8(self) -> npt.NDArray[np.int64]:
+        """
+        Integer representation of the values.
+
+        Returns
+        -------
+        ndarray
+            An ndarray with int64 dtype.
+        """
+        # do not cache or you'll create a memory leak
+        return self._ndarray.view("i8")
+
+    # ----------------------------------------------------------------
+    # Rendering Methods
+
+    def _format_native_types(
+        self, *, na_rep: str | float = "NaT", date_format=None
+    ) -> npt.NDArray[np.object_]:
+        """
+        Helper method for astype when converting to strings.
+
+        Returns
+        -------
+        ndarray[str]
+        """
+        raise AbstractMethodError(self)
+
+    def _formatter(self, boxed: bool = False):
+        # TODO: Remove Datetime & DatetimeTZ formatters.
+        return "'{}'".format
+
+    # ----------------------------------------------------------------
+    # Array-Like / EA-Interface Methods
+
+    def __array__(
+        self, dtype: NpDtype | None = None, copy: bool | None = None
+    ) -> np.ndarray:
+        # used for Timedelta/DatetimeArray, overwritten by PeriodArray
+        if is_object_dtype(dtype):
+            return np.array(list(self), dtype=object)
+        return self._ndarray
+
+    @overload
+    def __getitem__(self, item: ScalarIndexer) -> DTScalarOrNaT:
+        ...
+
+    @overload
+    def __getitem__(
+        self,
+        item: SequenceIndexer | PositionalIndexerTuple,
+    ) -> Self:
+        ...
+
+    def __getitem__(self, key: PositionalIndexer2D) -> Self | DTScalarOrNaT:
+        """
+        This getitem defers to the underlying array, which by-definition can
+        only handle list-likes, slices, and integer scalars
+        """
+        # Use cast as we know we will get back a DatetimeLikeArray or DTScalar,
+        # but skip evaluating the Union at runtime for performance
+        # (see https://github.com/pandas-dev/pandas/pull/44624)
+        result = cast("Union[Self, DTScalarOrNaT]", super().__getitem__(key))
+        if lib.is_scalar(result):
+            return result
+        else:
+            # At this point we know the result is an array.
+            result = cast(Self, result)
+        result._freq = self._get_getitem_freq(key)
+        return result
+
+    def _get_getitem_freq(self, key) -> BaseOffset | None:
+        """
+        Find the `freq` attribute to assign to the result of a __getitem__ lookup.
+        """
+        is_period = isinstance(self.dtype, PeriodDtype)
+        if is_period:
+            freq = self.freq
+        elif self.ndim != 1:
+            freq = None
+        else:
+            key = check_array_indexer(self, key)  # maybe ndarray[bool] -> slice
+            freq = None
+            if isinstance(key, slice):
+                if self.freq is not None and key.step is not None:
+                    freq = key.step * self.freq
+                else:
+                    freq = self.freq
+            elif key is Ellipsis:
+                # GH#21282 indexing with Ellipsis is similar to a full slice,
+                #  should preserve `freq` attribute
+                freq = self.freq
+            elif com.is_bool_indexer(key):
+                new_key = lib.maybe_booleans_to_slice(key.view(np.uint8))
+                if isinstance(new_key, slice):
+                    return self._get_getitem_freq(new_key)
+        return freq
+
+    # error: Argument 1 of "__setitem__" is incompatible with supertype
+    # "ExtensionArray"; supertype defines the argument type as "Union[int,
+    # ndarray]"
+    def __setitem__(
+        self,
+        key: int | Sequence[int] | Sequence[bool] | slice,
+        value: NaTType | Any | Sequence[Any],
+    ) -> None:
+        # I'm fudging the types a bit here. "Any" above really depends
+        # on type(self). For PeriodArray, it's Period (or stuff coercible
+        # to a period in from_sequence). For DatetimeArray, it's Timestamp...
+        # I don't know if mypy can do that, possibly with Generics.
+        # https://mypy.readthedocs.io/en/latest/generics.html
+
+        no_op = check_setitem_lengths(key, value, self)
+
+        # Calling super() before the no_op short-circuit means that we raise
+        #  on invalid 'value' even if this is a no-op, e.g. wrong-dtype empty array.
+        super().__setitem__(key, value)
+
+        if no_op:
+            return
+
+        self._maybe_clear_freq()
+
+    def _maybe_clear_freq(self) -> None:
+        # inplace operations like __setitem__ may invalidate the freq of
+        # DatetimeArray and TimedeltaArray
+        pass
+
+    def astype(self, dtype, copy: bool = True):
+        # Some notes on cases we don't have to handle here in the base class:
+        #   1. PeriodArray.astype handles period -> period
+        #   2. DatetimeArray.astype handles conversion between tz.
+        #   3. DatetimeArray.astype handles datetime -> period
+        dtype = pandas_dtype(dtype)
+
+        if dtype == object:
+            if self.dtype.kind == "M":
+                self = cast("DatetimeArray", self)
+                # *much* faster than self._box_values
+                #  for e.g. test_get_loc_tuple_monotonic_above_size_cutoff
+                i8data = self.asi8
+                converted = ints_to_pydatetime(
+                    i8data,
+                    tz=self.tz,
+                    box="timestamp",
+                    reso=self._creso,
+                )
+                return converted
+
+            elif self.dtype.kind == "m":
+                return ints_to_pytimedelta(self._ndarray, box=True)
+
+            return self._box_values(self.asi8.ravel()).reshape(self.shape)
+
+        elif isinstance(dtype, ExtensionDtype):
+            return super().astype(dtype, copy=copy)
+        elif is_string_dtype(dtype):
+            return self._format_native_types()
+        elif dtype.kind in "iu":
+            # we deliberately ignore int32 vs. int64 here.
+            # See https://github.com/pandas-dev/pandas/issues/24381 for more.
+            values = self.asi8
+            if dtype != np.int64:
+                raise TypeError(
+                    f"Converting from {self.dtype} to {dtype} is not supported. "
+                    "Do obj.astype('int64').astype(dtype) instead"
+                )
+
+            if copy:
+                values = values.copy()
+            return values
+        elif (dtype.kind in "mM" and self.dtype != dtype) or dtype.kind == "f":
+            # disallow conversion between datetime/timedelta,
+            # and conversions for any datetimelike to float
+            msg = f"Cannot cast {type(self).__name__} to dtype {dtype}"
+            raise TypeError(msg)
+        else:
+            return np.asarray(self, dtype=dtype)
+
+    @overload
+    def view(self) -> Self:
+        ...
+
+    @overload
+    def view(self, dtype: Literal["M8[ns]"]) -> DatetimeArray:
+        ...
+
+    @overload
+    def view(self, dtype: Literal["m8[ns]"]) -> TimedeltaArray:
+        ...
+
+    @overload
+    def view(self, dtype: Dtype | None = ...) -> ArrayLike:
+        ...
+
+    # pylint: disable-next=useless-parent-delegation
+    def view(self, dtype: Dtype | None = None) -> ArrayLike:
+        # we need to explicitly call super() method as long as the `@overload`s
+        #  are present in this file.
+        return super().view(dtype)
+
+    # ------------------------------------------------------------------
+    # Validation Methods
+    # TODO: try to de-duplicate these, ensure identical behavior
+
+    def _validate_comparison_value(self, other):
+        if isinstance(other, str):
+            try:
+                # GH#18435 strings get a pass from tzawareness compat
+                other = self._scalar_from_string(other)
+            except (ValueError, IncompatibleFrequency):
+                # failed to parse as Timestamp/Timedelta/Period
+                raise InvalidComparison(other)
+
+        if isinstance(other, self._recognized_scalars) or other is NaT:
+            other = self._scalar_type(other)
+            try:
+                self._check_compatible_with(other)
+            except (TypeError, IncompatibleFrequency) as err:
+                # e.g. tzawareness mismatch
+                raise InvalidComparison(other) from err
+
+        elif not is_list_like(other):
+            raise InvalidComparison(other)
+
+        elif len(other) != len(self):
+            raise ValueError("Lengths must match")
+
+        else:
+            try:
+                other = self._validate_listlike(other, allow_object=True)
+                self._check_compatible_with(other)
+            except (TypeError, IncompatibleFrequency) as err:
+                if is_object_dtype(getattr(other, "dtype", None)):
+                    # We will have to operate element-wise
+                    pass
+                else:
+                    raise InvalidComparison(other) from err
+
+        return other
+
+    def _validate_scalar(
+        self,
+        value,
+        *,
+        allow_listlike: bool = False,
+        unbox: bool = True,
+    ):
+        """
+        Validate that the input value can be cast to our scalar_type.
+
+        Parameters
+        ----------
+        value : object
+        allow_listlike: bool, default False
+            When raising an exception, whether the message should say
+            listlike inputs are allowed.
+        unbox : bool, default True
+            Whether to unbox the result before returning.  Note: unbox=False
+            skips the setitem compatibility check.
+
+        Returns
+        -------
+        self._scalar_type or NaT
+        """
+        if isinstance(value, self._scalar_type):
+            pass
+
+        elif isinstance(value, str):
+            # NB: Careful about tzawareness
+            try:
+                value = self._scalar_from_string(value)
+            except ValueError as err:
+                msg = self._validation_error_message(value, allow_listlike)
+                raise TypeError(msg) from err
+
+        elif is_valid_na_for_dtype(value, self.dtype):
+            # GH#18295
+            value = NaT
+
+        elif isna(value):
+            # if we are dt64tz and value is dt64("NaT"), dont cast to NaT,
+            #  or else we'll fail to raise in _unbox_scalar
+            msg = self._validation_error_message(value, allow_listlike)
+            raise TypeError(msg)
+
+        elif isinstance(value, self._recognized_scalars):
+            # error: Argument 1 to "Timestamp" has incompatible type "object"; expected
+            # "integer[Any] | float | str | date | datetime | datetime64"
+            value = self._scalar_type(value)  # type: ignore[arg-type]
+
+        else:
+            msg = self._validation_error_message(value, allow_listlike)
+            raise TypeError(msg)
+
+        if not unbox:
+            # NB: In general NDArrayBackedExtensionArray will unbox here;
+            #  this option exists to prevent a performance hit in
+            #  TimedeltaIndex.get_loc
+            return value
+        return self._unbox_scalar(value)
+
+    def _validation_error_message(self, value, allow_listlike: bool = False) -> str:
+        """
+        Construct an exception message on validation error.
+
+        Some methods allow only scalar inputs, while others allow either scalar
+        or listlike.
+
+        Parameters
+        ----------
+        allow_listlike: bool, default False
+
+        Returns
+        -------
+        str
+        """
+        if hasattr(value, "dtype") and getattr(value, "ndim", 0) > 0:
+            msg_got = f"{value.dtype} array"
+        else:
+            msg_got = f"'{type(value).__name__}'"
+        if allow_listlike:
+            msg = (
+                f"value should be a '{self._scalar_type.__name__}', 'NaT', "
+                f"or array of those. Got {msg_got} instead."
+            )
+        else:
+            msg = (
+                f"value should be a '{self._scalar_type.__name__}' or 'NaT'. "
+                f"Got {msg_got} instead."
+            )
+        return msg
+
+    def _validate_listlike(self, value, allow_object: bool = False):
+        if isinstance(value, type(self)):
+            if self.dtype.kind in "mM" and not allow_object:
+                # error: "DatetimeLikeArrayMixin" has no attribute "as_unit"
+                value = value.as_unit(self.unit, round_ok=False)  # type: ignore[attr-defined]
+            return value
+
+        if isinstance(value, list) and len(value) == 0:
+            # We treat empty list as our own dtype.
+            return type(self)._from_sequence([], dtype=self.dtype)
+
+        if hasattr(value, "dtype") and value.dtype == object:
+            # `array` below won't do inference if value is an Index or Series.
+            #  so do so here.  in the Index case, inferred_type may be cached.
+            if lib.infer_dtype(value) in self._infer_matches:
+                try:
+                    value = type(self)._from_sequence(value)
+                except (ValueError, TypeError):
+                    if allow_object:
+                        return value
+                    msg = self._validation_error_message(value, True)
+                    raise TypeError(msg)
+
+        # Do type inference if necessary up front (after unpacking
+        # NumpyExtensionArray)
+        # e.g. we passed PeriodIndex.values and got an ndarray of Periods
+        value = extract_array(value, extract_numpy=True)
+        value = pd_array(value)
+        value = extract_array(value, extract_numpy=True)
+
+        if is_all_strings(value):
+            # We got a StringArray
+            try:
+                # TODO: Could use from_sequence_of_strings if implemented
+                # Note: passing dtype is necessary for PeriodArray tests
+                value = type(self)._from_sequence(value, dtype=self.dtype)
+            except ValueError:
+                pass
+
+        if isinstance(value.dtype, CategoricalDtype):
+            # e.g. we have a Categorical holding self.dtype
+            if value.categories.dtype == self.dtype:
+                # TODO: do we need equal dtype or just comparable?
+                value = value._internal_get_values()
+                value = extract_array(value, extract_numpy=True)
+
+        if allow_object and is_object_dtype(value.dtype):
+            pass
+
+        elif not type(self)._is_recognized_dtype(value.dtype):
+            msg = self._validation_error_message(value, True)
+            raise TypeError(msg)
+
+        if self.dtype.kind in "mM" and not allow_object:
+            # error: "DatetimeLikeArrayMixin" has no attribute "as_unit"
+            value = value.as_unit(self.unit, round_ok=False)  # type: ignore[attr-defined]
+        return value
+
+    def _validate_setitem_value(self, value):
+        if is_list_like(value):
+            value = self._validate_listlike(value)
+        else:
+            return self._validate_scalar(value, allow_listlike=True)
+
+        return self._unbox(value)
+
+    @final
+    def _unbox(self, other) -> np.int64 | np.datetime64 | np.timedelta64 | np.ndarray:
+        """
+        Unbox either a scalar with _unbox_scalar or an instance of our own type.
+        """
+        if lib.is_scalar(other):
+            other = self._unbox_scalar(other)
+        else:
+            # same type as self
+            self._check_compatible_with(other)
+            other = other._ndarray
+        return other
+
+    # ------------------------------------------------------------------
+    # Additional array methods
+    #  These are not part of the EA API, but we implement them because
+    #  pandas assumes they're there.
+
+    @ravel_compat
+    def map(self, mapper, na_action=None):
+        from pandas import Index
+
+        result = map_array(self, mapper, na_action=na_action)
+        result = Index(result)
+
+        if isinstance(result, ABCMultiIndex):
+            return result.to_numpy()
+        else:
+            return result.array
+
+    def isin(self, values: ArrayLike) -> npt.NDArray[np.bool_]:
+        """
+        Compute boolean array of whether each value is found in the
+        passed set of values.
+
+        Parameters
+        ----------
+        values : np.ndarray or ExtensionArray
+
+        Returns
+        -------
+        ndarray[bool]
+        """
+        if values.dtype.kind in "fiuc":
+            # TODO: de-duplicate with equals, validate_comparison_value
+            return np.zeros(self.shape, dtype=bool)
+
+        values = ensure_wrapped_if_datetimelike(values)
+
+        if not isinstance(values, type(self)):
+            inferable = [
+                "timedelta",
+                "timedelta64",
+                "datetime",
+                "datetime64",
+                "date",
+                "period",
+            ]
+            if values.dtype == object:
+                values = lib.maybe_convert_objects(
+                    values,  # type: ignore[arg-type]
+                    convert_non_numeric=True,
+                    dtype_if_all_nat=self.dtype,
+                )
+                if values.dtype != object:
+                    return self.isin(values)
+
+                inferred = lib.infer_dtype(values, skipna=False)
+                if inferred not in inferable:
+                    if inferred == "string":
+                        pass
+
+                    elif "mixed" in inferred:
+                        return isin(self.astype(object), values)
+                    else:
+                        return np.zeros(self.shape, dtype=bool)
+
+            try:
+                values = type(self)._from_sequence(values)
+            except ValueError:
+                return isin(self.astype(object), values)
+            else:
+                warnings.warn(
+                    # GH#53111
+                    f"The behavior of 'isin' with dtype={self.dtype} and "
+                    "castable values (e.g. strings) is deprecated. In a "
+                    "future version, these will not be considered matching "
+                    "by isin. Explicitly cast to the appropriate dtype before "
+                    "calling isin instead.",
+                    FutureWarning,
+                    stacklevel=find_stack_level(),
+                )
+
+        if self.dtype.kind in "mM":
+            self = cast("DatetimeArray | TimedeltaArray", self)
+            # error: Item "ExtensionArray" of "ExtensionArray | ndarray[Any, Any]"
+            # has no attribute "as_unit"
+            values = values.as_unit(self.unit)  # type: ignore[union-attr]
+
+        try:
+            # error: Argument 1 to "_check_compatible_with" of "DatetimeLikeArrayMixin"
+            # has incompatible type "ExtensionArray | ndarray[Any, Any]"; expected
+            # "Period | Timestamp | Timedelta | NaTType"
+            self._check_compatible_with(values)  # type: ignore[arg-type]
+        except (TypeError, ValueError):
+            # Includes tzawareness mismatch and IncompatibleFrequencyError
+            return np.zeros(self.shape, dtype=bool)
+
+        # error: Item "ExtensionArray" of "ExtensionArray | ndarray[Any, Any]"
+        # has no attribute "asi8"
+        return isin(self.asi8, values.asi8)  # type: ignore[union-attr]
+
+    # ------------------------------------------------------------------
+    # Null Handling
+
+    def isna(self) -> npt.NDArray[np.bool_]:
+        return self._isnan
+
+    @property  # NB: override with cache_readonly in immutable subclasses
+    def _isnan(self) -> npt.NDArray[np.bool_]:
+        """
+        return if each value is nan
+        """
+        return self.asi8 == iNaT
+
+    @property  # NB: override with cache_readonly in immutable subclasses
+    def _hasna(self) -> bool:
+        """
+        return if I have any nans; enables various perf speedups
+        """
+        return bool(self._isnan.any())
+
+    def _maybe_mask_results(
+        self, result: np.ndarray, fill_value=iNaT, convert=None
+    ) -> np.ndarray:
+        """
+        Parameters
+        ----------
+        result : np.ndarray
+        fill_value : object, default iNaT
+        convert : str, dtype or None
+
+        Returns
+        -------
+        result : ndarray with values replace by the fill_value
+
+        mask the result if needed, convert to the provided dtype if its not
+        None
+
+        This is an internal routine.
+        """
+        if self._hasna:
+            if convert:
+                result = result.astype(convert)
+            if fill_value is None:
+                fill_value = np.nan
+            np.putmask(result, self._isnan, fill_value)
+        return result
+
+    # ------------------------------------------------------------------
+    # Frequency Properties/Methods
+
+    @property
+    def freqstr(self) -> str | None:
+        """
+        Return the frequency object as a string if it's set, otherwise None.
+
+        Examples
+        --------
+        For DatetimeIndex:
+
+        >>> idx = pd.DatetimeIndex(["1/1/2020 10:00:00+00:00"], freq="D")
+        >>> idx.freqstr
+        'D'
+
+        The frequency can be inferred if there are more than 2 points:
+
+        >>> idx = pd.DatetimeIndex(["2018-01-01", "2018-01-03", "2018-01-05"],
+        ...                        freq="infer")
+        >>> idx.freqstr
+        '2D'
+
+        For PeriodIndex:
+
+        >>> idx = pd.PeriodIndex(["2023-1", "2023-2", "2023-3"], freq="M")
+        >>> idx.freqstr
+        'M'
+        """
+        if self.freq is None:
+            return None
+        return self.freq.freqstr
+
+    @property  # NB: override with cache_readonly in immutable subclasses
+    def inferred_freq(self) -> str | None:
+        """
+        Tries to return a string representing a frequency generated by infer_freq.
+
+        Returns None if it can't autodetect the frequency.
+
+        Examples
+        --------
+        For DatetimeIndex:
+
+        >>> idx = pd.DatetimeIndex(["2018-01-01", "2018-01-03", "2018-01-05"])
+        >>> idx.inferred_freq
+        '2D'
+
+        For TimedeltaIndex:
+
+        >>> tdelta_idx = pd.to_timedelta(["0 days", "10 days", "20 days"])
+        >>> tdelta_idx
+        TimedeltaIndex(['0 days', '10 days', '20 days'],
+                       dtype='timedelta64[ns]', freq=None)
+        >>> tdelta_idx.inferred_freq
+        '10D'
+        """
+        if self.ndim != 1:
+            return None
+        try:
+            return frequencies.infer_freq(self)
+        except ValueError:
+            return None
+
+    @property  # NB: override with cache_readonly in immutable subclasses
+    def _resolution_obj(self) -> Resolution | None:
+        freqstr = self.freqstr
+        if freqstr is None:
+            return None
+        try:
+            return Resolution.get_reso_from_freqstr(freqstr)
+        except KeyError:
+            return None
+
+    @property  # NB: override with cache_readonly in immutable subclasses
+    def resolution(self) -> str:
+        """
+        Returns day, hour, minute, second, millisecond or microsecond
+        """
+        # error: Item "None" of "Optional[Any]" has no attribute "attrname"
+        return self._resolution_obj.attrname  # type: ignore[union-attr]
+
+    # monotonicity/uniqueness properties are called via frequencies.infer_freq,
+    #  see GH#23789
+
+    @property
+    def _is_monotonic_increasing(self) -> bool:
+        return algos.is_monotonic(self.asi8, timelike=True)[0]
+
+    @property
+    def _is_monotonic_decreasing(self) -> bool:
+        return algos.is_monotonic(self.asi8, timelike=True)[1]
+
+    @property
+    def _is_unique(self) -> bool:
+        return len(unique1d(self.asi8.ravel("K"))) == self.size
+
+    # ------------------------------------------------------------------
+    # Arithmetic Methods
+
+    def _cmp_method(self, other, op):
+        if self.ndim > 1 and getattr(other, "shape", None) == self.shape:
+            # TODO: handle 2D-like listlikes
+            return op(self.ravel(), other.ravel()).reshape(self.shape)
+
+        try:
+            other = self._validate_comparison_value(other)
+        except InvalidComparison:
+            return invalid_comparison(self, other, op)
+
+        dtype = getattr(other, "dtype", None)
+        if is_object_dtype(dtype):
+            # We have to use comp_method_OBJECT_ARRAY instead of numpy
+            #  comparison otherwise it would raise when comparing to None
+            result = ops.comp_method_OBJECT_ARRAY(
+                op, np.asarray(self.astype(object)), other
+            )
+            return result
+        if other is NaT:
+            if op is operator.ne:
+                result = np.ones(self.shape, dtype=bool)
+            else:
+                result = np.zeros(self.shape, dtype=bool)
+            return result
+
+        if not isinstance(self.dtype, PeriodDtype):
+            self = cast(TimelikeOps, self)
+            if self._creso != other._creso:
+                if not isinstance(other, type(self)):
+                    # i.e. Timedelta/Timestamp, cast to ndarray and let
+                    #  compare_mismatched_resolutions handle broadcasting
+                    try:
+                        # GH#52080 see if we can losslessly cast to shared unit
+                        other = other.as_unit(self.unit, round_ok=False)
+                    except ValueError:
+                        other_arr = np.array(other.asm8)
+                        return compare_mismatched_resolutions(
+                            self._ndarray, other_arr, op
+                        )
+                else:
+                    other_arr = other._ndarray
+                    return compare_mismatched_resolutions(self._ndarray, other_arr, op)
+
+        other_vals = self._unbox(other)
+        # GH#37462 comparison on i8 values is almost 2x faster than M8/m8
+        result = op(self._ndarray.view("i8"), other_vals.view("i8"))
+
+        o_mask = isna(other)
+        mask = self._isnan | o_mask
+        if mask.any():
+            nat_result = op is operator.ne
+            np.putmask(result, mask, nat_result)
+
+        return result
+
+    # pow is invalid for all three subclasses; TimedeltaArray will override
+    #  the multiplication and division ops
+    __pow__ = _make_unpacked_invalid_op("__pow__")
+    __rpow__ = _make_unpacked_invalid_op("__rpow__")
+    __mul__ = _make_unpacked_invalid_op("__mul__")
+    __rmul__ = _make_unpacked_invalid_op("__rmul__")
+    __truediv__ = _make_unpacked_invalid_op("__truediv__")
+    __rtruediv__ = _make_unpacked_invalid_op("__rtruediv__")
+    __floordiv__ = _make_unpacked_invalid_op("__floordiv__")
+    __rfloordiv__ = _make_unpacked_invalid_op("__rfloordiv__")
+    __mod__ = _make_unpacked_invalid_op("__mod__")
+    __rmod__ = _make_unpacked_invalid_op("__rmod__")
+    __divmod__ = _make_unpacked_invalid_op("__divmod__")
+    __rdivmod__ = _make_unpacked_invalid_op("__rdivmod__")
+
+    @final
+    def _get_i8_values_and_mask(
+        self, other
+    ) -> tuple[int | npt.NDArray[np.int64], None | npt.NDArray[np.bool_]]:
+        """
+        Get the int64 values and b_mask to pass to add_overflowsafe.
+        """
+        if isinstance(other, Period):
+            i8values = other.ordinal
+            mask = None
+        elif isinstance(other, (Timestamp, Timedelta)):
+            i8values = other._value
+            mask = None
+        else:
+            # PeriodArray, DatetimeArray, TimedeltaArray
+            mask = other._isnan
+            i8values = other.asi8
+        return i8values, mask
+
+    @final
+    def _get_arithmetic_result_freq(self, other) -> BaseOffset | None:
+        """
+        Check if we can preserve self.freq in addition or subtraction.
+        """
+        # Adding or subtracting a Timedelta/Timestamp scalar is freq-preserving
+        #  whenever self.freq is a Tick
+        if isinstance(self.dtype, PeriodDtype):
+            return self.freq
+        elif not lib.is_scalar(other):
+            return None
+        elif isinstance(self.freq, Tick):
+            # In these cases
+            return self.freq
+        return None
+
+    @final
+    def _add_datetimelike_scalar(self, other) -> DatetimeArray:
+        if not lib.is_np_dtype(self.dtype, "m"):
+            raise TypeError(
+                f"cannot add {type(self).__name__} and {type(other).__name__}"
+            )
+
+        self = cast("TimedeltaArray", self)
+
+        from pandas.core.arrays import DatetimeArray
+        from pandas.core.arrays.datetimes import tz_to_dtype
+
+        assert other is not NaT
+        if isna(other):
+            # i.e. np.datetime64("NaT")
+            # In this case we specifically interpret NaT as a datetime, not
+            # the timedelta interpretation we would get by returning self + NaT
+            result = self._ndarray + NaT.to_datetime64().astype(f"M8[{self.unit}]")
+            # Preserve our resolution
+            return DatetimeArray._simple_new(result, dtype=result.dtype)
+
+        other = Timestamp(other)
+        self, other = self._ensure_matching_resos(other)
+        self = cast("TimedeltaArray", self)
+
+        other_i8, o_mask = self._get_i8_values_and_mask(other)
+        result = add_overflowsafe(self.asi8, np.asarray(other_i8, dtype="i8"))
+        res_values = result.view(f"M8[{self.unit}]")
+
+        dtype = tz_to_dtype(tz=other.tz, unit=self.unit)
+        res_values = result.view(f"M8[{self.unit}]")
+        new_freq = self._get_arithmetic_result_freq(other)
+        return DatetimeArray._simple_new(res_values, dtype=dtype, freq=new_freq)
+
+    @final
+    def _add_datetime_arraylike(self, other: DatetimeArray) -> DatetimeArray:
+        if not lib.is_np_dtype(self.dtype, "m"):
+            raise TypeError(
+                f"cannot add {type(self).__name__} and {type(other).__name__}"
+            )
+
+        # defer to DatetimeArray.__add__
+        return other + self
+
+    @final
+    def _sub_datetimelike_scalar(
+        self, other: datetime | np.datetime64
+    ) -> TimedeltaArray:
+        if self.dtype.kind != "M":
+            raise TypeError(f"cannot subtract a datelike from a {type(self).__name__}")
+
+        self = cast("DatetimeArray", self)
+        # subtract a datetime from myself, yielding a ndarray[timedelta64[ns]]
+
+        if isna(other):
+            # i.e. np.datetime64("NaT")
+            return self - NaT
+
+        ts = Timestamp(other)
+
+        self, ts = self._ensure_matching_resos(ts)
+        return self._sub_datetimelike(ts)
+
+    @final
+    def _sub_datetime_arraylike(self, other: DatetimeArray) -> TimedeltaArray:
+        if self.dtype.kind != "M":
+            raise TypeError(f"cannot subtract a datelike from a {type(self).__name__}")
+
+        if len(self) != len(other):
+            raise ValueError("cannot add indices of unequal length")
+
+        self = cast("DatetimeArray", self)
+
+        self, other = self._ensure_matching_resos(other)
+        return self._sub_datetimelike(other)
+
+    @final
+    def _sub_datetimelike(self, other: Timestamp | DatetimeArray) -> TimedeltaArray:
+        self = cast("DatetimeArray", self)
+
+        from pandas.core.arrays import TimedeltaArray
+
+        try:
+            self._assert_tzawareness_compat(other)
+        except TypeError as err:
+            new_message = str(err).replace("compare", "subtract")
+            raise type(err)(new_message) from err
+
+        other_i8, o_mask = self._get_i8_values_and_mask(other)
+        res_values = add_overflowsafe(self.asi8, np.asarray(-other_i8, dtype="i8"))
+        res_m8 = res_values.view(f"timedelta64[{self.unit}]")
+
+        new_freq = self._get_arithmetic_result_freq(other)
+        new_freq = cast("Tick | None", new_freq)
+        return TimedeltaArray._simple_new(res_m8, dtype=res_m8.dtype, freq=new_freq)
+
+    @final
+    def _add_period(self, other: Period) -> PeriodArray:
+        if not lib.is_np_dtype(self.dtype, "m"):
+            raise TypeError(f"cannot add Period to a {type(self).__name__}")
+
+        # We will wrap in a PeriodArray and defer to the reversed operation
+        from pandas.core.arrays.period import PeriodArray
+
+        i8vals = np.broadcast_to(other.ordinal, self.shape)
+        dtype = PeriodDtype(other.freq)
+        parr = PeriodArray(i8vals, dtype=dtype)
+        return parr + self
+
+    def _add_offset(self, offset):
+        raise AbstractMethodError(self)
+
+    def _add_timedeltalike_scalar(self, other):
+        """
+        Add a delta of a timedeltalike
+
+        Returns
+        -------
+        Same type as self
+        """
+        if isna(other):
+            # i.e np.timedelta64("NaT")
+            new_values = np.empty(self.shape, dtype="i8").view(self._ndarray.dtype)
+            new_values.fill(iNaT)
+            return type(self)._simple_new(new_values, dtype=self.dtype)
+
+        # PeriodArray overrides, so we only get here with DTA/TDA
+        self = cast("DatetimeArray | TimedeltaArray", self)
+        other = Timedelta(other)
+        self, other = self._ensure_matching_resos(other)
+        return self._add_timedeltalike(other)
+
+    def _add_timedelta_arraylike(self, other: TimedeltaArray):
+        """
+        Add a delta of a TimedeltaIndex
+
+        Returns
+        -------
+        Same type as self
+        """
+        # overridden by PeriodArray
+
+        if len(self) != len(other):
+            raise ValueError("cannot add indices of unequal length")
+
+        self = cast("DatetimeArray | TimedeltaArray", self)
+
+        self, other = self._ensure_matching_resos(other)
+        return self._add_timedeltalike(other)
+
+    @final
+    def _add_timedeltalike(self, other: Timedelta | TimedeltaArray):
+        self = cast("DatetimeArray | TimedeltaArray", self)
+
+        other_i8, o_mask = self._get_i8_values_and_mask(other)
+        new_values = add_overflowsafe(self.asi8, np.asarray(other_i8, dtype="i8"))
+        res_values = new_values.view(self._ndarray.dtype)
+
+        new_freq = self._get_arithmetic_result_freq(other)
+
+        # error: Argument "dtype" to "_simple_new" of "DatetimeArray" has
+        # incompatible type "Union[dtype[datetime64], DatetimeTZDtype,
+        # dtype[timedelta64]]"; expected "Union[dtype[datetime64], DatetimeTZDtype]"
+        return type(self)._simple_new(
+            res_values, dtype=self.dtype, freq=new_freq  # type: ignore[arg-type]
+        )
+
+    @final
+    def _add_nat(self):
+        """
+        Add pd.NaT to self
+        """
+        if isinstance(self.dtype, PeriodDtype):
+            raise TypeError(
+                f"Cannot add {type(self).__name__} and {type(NaT).__name__}"
+            )
+        self = cast("TimedeltaArray | DatetimeArray", self)
+
+        # GH#19124 pd.NaT is treated like a timedelta for both timedelta
+        # and datetime dtypes
+        result = np.empty(self.shape, dtype=np.int64)
+        result.fill(iNaT)
+        result = result.view(self._ndarray.dtype)  # preserve reso
+        # error: Argument "dtype" to "_simple_new" of "DatetimeArray" has
+        # incompatible type "Union[dtype[timedelta64], dtype[datetime64],
+        # DatetimeTZDtype]"; expected "Union[dtype[datetime64], DatetimeTZDtype]"
+        return type(self)._simple_new(
+            result, dtype=self.dtype, freq=None  # type: ignore[arg-type]
+        )
+
+    @final
+    def _sub_nat(self):
+        """
+        Subtract pd.NaT from self
+        """
+        # GH#19124 Timedelta - datetime is not in general well-defined.
+        # We make an exception for pd.NaT, which in this case quacks
+        # like a timedelta.
+        # For datetime64 dtypes by convention we treat NaT as a datetime, so
+        # this subtraction returns a timedelta64 dtype.
+        # For period dtype, timedelta64 is a close-enough return dtype.
+        result = np.empty(self.shape, dtype=np.int64)
+        result.fill(iNaT)
+        if self.dtype.kind in "mM":
+            # We can retain unit in dtype
+            self = cast("DatetimeArray| TimedeltaArray", self)
+            return result.view(f"timedelta64[{self.unit}]")
+        else:
+            return result.view("timedelta64[ns]")
+
+    @final
+    def _sub_periodlike(self, other: Period | PeriodArray) -> npt.NDArray[np.object_]:
+        # If the operation is well-defined, we return an object-dtype ndarray
+        # of DateOffsets.  Null entries are filled with pd.NaT
+        if not isinstance(self.dtype, PeriodDtype):
+            raise TypeError(
+                f"cannot subtract {type(other).__name__} from {type(self).__name__}"
+            )
+
+        self = cast("PeriodArray", self)
+        self._check_compatible_with(other)
+
+        other_i8, o_mask = self._get_i8_values_and_mask(other)
+        new_i8_data = add_overflowsafe(self.asi8, np.asarray(-other_i8, dtype="i8"))
+        new_data = np.array([self.freq.base * x for x in new_i8_data])
+
+        if o_mask is None:
+            # i.e. Period scalar
+            mask = self._isnan
+        else:
+            # i.e. PeriodArray
+            mask = self._isnan | o_mask
+        new_data[mask] = NaT
+        return new_data
+
+    @final
+    def _addsub_object_array(self, other: npt.NDArray[np.object_], op):
+        """
+        Add or subtract array-like of DateOffset objects
+
+        Parameters
+        ----------
+        other : np.ndarray[object]
+        op : {operator.add, operator.sub}
+
+        Returns
+        -------
+        np.ndarray[object]
+            Except in fastpath case with length 1 where we operate on the
+            contained scalar.
+        """
+        assert op in [operator.add, operator.sub]
+        if len(other) == 1 and self.ndim == 1:
+            # Note: without this special case, we could annotate return type
+            #  as ndarray[object]
+            # If both 1D then broadcasting is unambiguous
+            return op(self, other[0])
+
+        warnings.warn(
+            "Adding/subtracting object-dtype array to "
+            f"{type(self).__name__} not vectorized.",
+            PerformanceWarning,
+            stacklevel=find_stack_level(),
+        )
+
+        # Caller is responsible for broadcasting if necessary
+        assert self.shape == other.shape, (self.shape, other.shape)
+
+        res_values = op(self.astype("O"), np.asarray(other))
+        return res_values
+
+    def _accumulate(self, name: str, *, skipna: bool = True, **kwargs) -> Self:
+        if name not in {"cummin", "cummax"}:
+            raise TypeError(f"Accumulation {name} not supported for {type(self)}")
+
+        op = getattr(datetimelike_accumulations, name)
+        result = op(self.copy(), skipna=skipna, **kwargs)
+
+        return type(self)._simple_new(result, dtype=self.dtype)
+
+    @unpack_zerodim_and_defer("__add__")
+    def __add__(self, other):
+        other_dtype = getattr(other, "dtype", None)
+        other = ensure_wrapped_if_datetimelike(other)
+
+        # scalar others
+        if other is NaT:
+            result = self._add_nat()
+        elif isinstance(other, (Tick, timedelta, np.timedelta64)):
+            result = self._add_timedeltalike_scalar(other)
+        elif isinstance(other, BaseOffset):
+            # specifically _not_ a Tick
+            result = self._add_offset(other)
+        elif isinstance(other, (datetime, np.datetime64)):
+            result = self._add_datetimelike_scalar(other)
+        elif isinstance(other, Period) and lib.is_np_dtype(self.dtype, "m"):
+            result = self._add_period(other)
+        elif lib.is_integer(other):
+            # This check must come after the check for np.timedelta64
+            # as is_integer returns True for these
+            if not isinstance(self.dtype, PeriodDtype):
+                raise integer_op_not_supported(self)
+            obj = cast("PeriodArray", self)
+            result = obj._addsub_int_array_or_scalar(other * obj.dtype._n, operator.add)
+
+        # array-like others
+        elif lib.is_np_dtype(other_dtype, "m"):
+            # TimedeltaIndex, ndarray[timedelta64]
+            result = self._add_timedelta_arraylike(other)
+        elif is_object_dtype(other_dtype):
+            # e.g. Array/Index of DateOffset objects
+            result = self._addsub_object_array(other, operator.add)
+        elif lib.is_np_dtype(other_dtype, "M") or isinstance(
+            other_dtype, DatetimeTZDtype
+        ):
+            # DatetimeIndex, ndarray[datetime64]
+            return self._add_datetime_arraylike(other)
+        elif is_integer_dtype(other_dtype):
+            if not isinstance(self.dtype, PeriodDtype):
+                raise integer_op_not_supported(self)
+            obj = cast("PeriodArray", self)
+            result = obj._addsub_int_array_or_scalar(other * obj.dtype._n, operator.add)
+        else:
+            # Includes Categorical, other ExtensionArrays
+            # For PeriodDtype, if self is a TimedeltaArray and other is a
+            #  PeriodArray with  a timedelta-like (i.e. Tick) freq, this
+            #  operation is valid.  Defer to the PeriodArray implementation.
+            #  In remaining cases, this will end up raising TypeError.
+            return NotImplemented
+
+        if isinstance(result, np.ndarray) and lib.is_np_dtype(result.dtype, "m"):
+            from pandas.core.arrays import TimedeltaArray
+
+            return TimedeltaArray._from_sequence(result)
+        return result
+
+    def __radd__(self, other):
+        # alias for __add__
+        return self.__add__(other)
+
+    @unpack_zerodim_and_defer("__sub__")
+    def __sub__(self, other):
+        other_dtype = getattr(other, "dtype", None)
+        other = ensure_wrapped_if_datetimelike(other)
+
+        # scalar others
+        if other is NaT:
+            result = self._sub_nat()
+        elif isinstance(other, (Tick, timedelta, np.timedelta64)):
+            result = self._add_timedeltalike_scalar(-other)
+        elif isinstance(other, BaseOffset):
+            # specifically _not_ a Tick
+            result = self._add_offset(-other)
+        elif isinstance(other, (datetime, np.datetime64)):
+            result = self._sub_datetimelike_scalar(other)
+        elif lib.is_integer(other):
+            # This check must come after the check for np.timedelta64
+            # as is_integer returns True for these
+            if not isinstance(self.dtype, PeriodDtype):
+                raise integer_op_not_supported(self)
+            obj = cast("PeriodArray", self)
+            result = obj._addsub_int_array_or_scalar(other * obj.dtype._n, operator.sub)
+
+        elif isinstance(other, Period):
+            result = self._sub_periodlike(other)
+
+        # array-like others
+        elif lib.is_np_dtype(other_dtype, "m"):
+            # TimedeltaIndex, ndarray[timedelta64]
+            result = self._add_timedelta_arraylike(-other)
+        elif is_object_dtype(other_dtype):
+            # e.g. Array/Index of DateOffset objects
+            result = self._addsub_object_array(other, operator.sub)
+        elif lib.is_np_dtype(other_dtype, "M") or isinstance(
+            other_dtype, DatetimeTZDtype
+        ):
+            # DatetimeIndex, ndarray[datetime64]
+            result = self._sub_datetime_arraylike(other)
+        elif isinstance(other_dtype, PeriodDtype):
+            # PeriodIndex
+            result = self._sub_periodlike(other)
+        elif is_integer_dtype(other_dtype):
+            if not isinstance(self.dtype, PeriodDtype):
+                raise integer_op_not_supported(self)
+            obj = cast("PeriodArray", self)
+            result = obj._addsub_int_array_or_scalar(other * obj.dtype._n, operator.sub)
+        else:
+            # Includes ExtensionArrays, float_dtype
+            return NotImplemented
+
+        if isinstance(result, np.ndarray) and lib.is_np_dtype(result.dtype, "m"):
+            from pandas.core.arrays import TimedeltaArray
+
+            return TimedeltaArray._from_sequence(result)
+        return result
+
+    def __rsub__(self, other):
+        other_dtype = getattr(other, "dtype", None)
+        other_is_dt64 = lib.is_np_dtype(other_dtype, "M") or isinstance(
+            other_dtype, DatetimeTZDtype
+        )
+
+        if other_is_dt64 and lib.is_np_dtype(self.dtype, "m"):
+            # ndarray[datetime64] cannot be subtracted from self, so
+            # we need to wrap in DatetimeArray/Index and flip the operation
+            if lib.is_scalar(other):
+                # i.e. np.datetime64 object
+                return Timestamp(other) - self
+            if not isinstance(other, DatetimeLikeArrayMixin):
+                # Avoid down-casting DatetimeIndex
+                from pandas.core.arrays import DatetimeArray
+
+                other = DatetimeArray._from_sequence(other)
+            return other - self
+        elif self.dtype.kind == "M" and hasattr(other, "dtype") and not other_is_dt64:
+            # GH#19959 datetime - datetime is well-defined as timedelta,
+            # but any other type - datetime is not well-defined.
+            raise TypeError(
+                f"cannot subtract {type(self).__name__} from {type(other).__name__}"
+            )
+        elif isinstance(self.dtype, PeriodDtype) and lib.is_np_dtype(other_dtype, "m"):
+            # TODO: Can we simplify/generalize these cases at all?
+            raise TypeError(f"cannot subtract {type(self).__name__} from {other.dtype}")
+        elif lib.is_np_dtype(self.dtype, "m"):
+            self = cast("TimedeltaArray", self)
+            return (-self) + other
+
+        # We get here with e.g. datetime objects
+        return -(self - other)
+
+    def __iadd__(self, other) -> Self:
+        result = self + other
+        self[:] = result[:]
+
+        if not isinstance(self.dtype, PeriodDtype):
+            # restore freq, which is invalidated by setitem
+            self._freq = result.freq
+        return self
+
+    def __isub__(self, other) -> Self:
+        result = self - other
+        self[:] = result[:]
+
+        if not isinstance(self.dtype, PeriodDtype):
+            # restore freq, which is invalidated by setitem
+            self._freq = result.freq
+        return self
+
+    # --------------------------------------------------------------
+    # Reductions
+
+    @_period_dispatch
+    def _quantile(
+        self,
+        qs: npt.NDArray[np.float64],
+        interpolation: str,
+    ) -> Self:
+        return super()._quantile(qs=qs, interpolation=interpolation)
+
+    @_period_dispatch
+    def min(self, *, axis: AxisInt | None = None, skipna: bool = True, **kwargs):
+        """
+        Return the minimum value of the Array or minimum along
+        an axis.
+
+        See Also
+        --------
+        numpy.ndarray.min
+        Index.min : Return the minimum value in an Index.
+        Series.min : Return the minimum value in a Series.
+        """
+        nv.validate_min((), kwargs)
+        nv.validate_minmax_axis(axis, self.ndim)
+
+        result = nanops.nanmin(self._ndarray, axis=axis, skipna=skipna)
+        return self._wrap_reduction_result(axis, result)
+
+    @_period_dispatch
+    def max(self, *, axis: AxisInt | None = None, skipna: bool = True, **kwargs):
+        """
+        Return the maximum value of the Array or maximum along
+        an axis.
+
+        See Also
+        --------
+        numpy.ndarray.max
+        Index.max : Return the maximum value in an Index.
+        Series.max : Return the maximum value in a Series.
+        """
+        nv.validate_max((), kwargs)
+        nv.validate_minmax_axis(axis, self.ndim)
+
+        result = nanops.nanmax(self._ndarray, axis=axis, skipna=skipna)
+        return self._wrap_reduction_result(axis, result)
+
+    def mean(self, *, skipna: bool = True, axis: AxisInt | None = 0):
+        """
+        Return the mean value of the Array.
+
+        Parameters
+        ----------
+        skipna : bool, default True
+            Whether to ignore any NaT elements.
+        axis : int, optional, default 0
+
+        Returns
+        -------
+        scalar
+            Timestamp or Timedelta.
+
+        See Also
+        --------
+        numpy.ndarray.mean : Returns the average of array elements along a given axis.
+        Series.mean : Return the mean value in a Series.
+
+        Notes
+        -----
+        mean is only defined for Datetime and Timedelta dtypes, not for Period.
+
+        Examples
+        --------
+        For :class:`pandas.DatetimeIndex`:
+
+        >>> idx = pd.date_range('2001-01-01 00:00', periods=3)
+        >>> idx
+        DatetimeIndex(['2001-01-01', '2001-01-02', '2001-01-03'],
+                      dtype='datetime64[ns]', freq='D')
+        >>> idx.mean()
+        Timestamp('2001-01-02 00:00:00')
+
+        For :class:`pandas.TimedeltaIndex`:
+
+        >>> tdelta_idx = pd.to_timedelta([1, 2, 3], unit='D')
+        >>> tdelta_idx
+        TimedeltaIndex(['1 days', '2 days', '3 days'],
+                        dtype='timedelta64[ns]', freq=None)
+        >>> tdelta_idx.mean()
+        Timedelta('2 days 00:00:00')
+        """
+        if isinstance(self.dtype, PeriodDtype):
+            # See discussion in GH#24757
+            raise TypeError(
+                f"mean is not implemented for {type(self).__name__} since the "
+                "meaning is ambiguous.  An alternative is "
+                "obj.to_timestamp(how='start').mean()"
+            )
+
+        result = nanops.nanmean(
+            self._ndarray, axis=axis, skipna=skipna, mask=self.isna()
+        )
+        return self._wrap_reduction_result(axis, result)
+
+    @_period_dispatch
+    def median(self, *, axis: AxisInt | None = None, skipna: bool = True, **kwargs):
+        nv.validate_median((), kwargs)
+
+        if axis is not None and abs(axis) >= self.ndim:
+            raise ValueError("abs(axis) must be less than ndim")
+
+        result = nanops.nanmedian(self._ndarray, axis=axis, skipna=skipna)
+        return self._wrap_reduction_result(axis, result)
+
+    def _mode(self, dropna: bool = True):
+        mask = None
+        if dropna:
+            mask = self.isna()
+
+        i8modes = algorithms.mode(self.view("i8"), mask=mask)
+        npmodes = i8modes.view(self._ndarray.dtype)
+        npmodes = cast(np.ndarray, npmodes)
+        return self._from_backing_data(npmodes)
+
+    # ------------------------------------------------------------------
+    # GroupBy Methods
+
+    def _groupby_op(
+        self,
+        *,
+        how: str,
+        has_dropped_na: bool,
+        min_count: int,
+        ngroups: int,
+        ids: npt.NDArray[np.intp],
+        **kwargs,
+    ):
+        dtype = self.dtype
+        if dtype.kind == "M":
+            # Adding/multiplying datetimes is not valid
+            if how in ["sum", "prod", "cumsum", "cumprod", "var", "skew"]:
+                raise TypeError(f"datetime64 type does not support {how} operations")
+            if how in ["any", "all"]:
+                # GH#34479
+                warnings.warn(
+                    f"'{how}' with datetime64 dtypes is deprecated and will raise in a "
+                    f"future version. Use (obj != pd.Timestamp(0)).{how}() instead.",
+                    FutureWarning,
+                    stacklevel=find_stack_level(),
+                )
+
+        elif isinstance(dtype, PeriodDtype):
+            # Adding/multiplying Periods is not valid
+            if how in ["sum", "prod", "cumsum", "cumprod", "var", "skew"]:
+                raise TypeError(f"Period type does not support {how} operations")
+            if how in ["any", "all"]:
+                # GH#34479
+                warnings.warn(
+                    f"'{how}' with PeriodDtype is deprecated and will raise in a "
+                    f"future version. Use (obj != pd.Period(0, freq)).{how}() instead.",
+                    FutureWarning,
+                    stacklevel=find_stack_level(),
+                )
+        else:
+            # timedeltas we can add but not multiply
+            if how in ["prod", "cumprod", "skew", "var"]:
+                raise TypeError(f"timedelta64 type does not support {how} operations")
+
+        # All of the functions implemented here are ordinal, so we can
+        #  operate on the tz-naive equivalents
+        npvalues = self._ndarray.view("M8[ns]")
+
+        from pandas.core.groupby.ops import WrappedCythonOp
+
+        kind = WrappedCythonOp.get_kind_from_how(how)
+        op = WrappedCythonOp(how=how, kind=kind, has_dropped_na=has_dropped_na)
+
+        res_values = op._cython_op_ndim_compat(
+            npvalues,
+            min_count=min_count,
+            ngroups=ngroups,
+            comp_ids=ids,
+            mask=None,
+            **kwargs,
+        )
+
+        if op.how in op.cast_blocklist:
+            # i.e. how in ["rank"], since other cast_blocklist methods don't go
+            #  through cython_operation
+            return res_values
+
+        # We did a view to M8[ns] above, now we go the other direction
+        assert res_values.dtype == "M8[ns]"
+        if how in ["std", "sem"]:
+            from pandas.core.arrays import TimedeltaArray
+
+            if isinstance(self.dtype, PeriodDtype):
+                raise TypeError("'std' and 'sem' are not valid for PeriodDtype")
+            self = cast("DatetimeArray | TimedeltaArray", self)
+            new_dtype = f"m8[{self.unit}]"
+            res_values = res_values.view(new_dtype)
+            return TimedeltaArray._simple_new(res_values, dtype=res_values.dtype)
+
+        res_values = res_values.view(self._ndarray.dtype)
+        return self._from_backing_data(res_values)
+
+
+class DatelikeOps(DatetimeLikeArrayMixin):
+    """
+    Common ops for DatetimeIndex/PeriodIndex, but not TimedeltaIndex.
+    """
+
+    @Substitution(
+        URL="https://docs.python.org/3/library/datetime.html"
+        "#strftime-and-strptime-behavior"
+    )
+    def strftime(self, date_format: str) -> npt.NDArray[np.object_]:
+        """
+        Convert to Index using specified date_format.
+
+        Return an Index of formatted strings specified by date_format, which
+        supports the same string format as the python standard library. Details
+        of the string format can be found in `python string format
+        doc <%(URL)s>`__.
+
+        Formats supported by the C `strftime` API but not by the python string format
+        doc (such as `"%%R"`, `"%%r"`) are not officially supported and should be
+        preferably replaced with their supported equivalents (such as `"%%H:%%M"`,
+        `"%%I:%%M:%%S %%p"`).
+
+        Note that `PeriodIndex` support additional directives, detailed in
+        `Period.strftime`.
+
+        Parameters
+        ----------
+        date_format : str
+            Date format string (e.g. "%%Y-%%m-%%d").
+
+        Returns
+        -------
+        ndarray[object]
+            NumPy ndarray of formatted strings.
+
+        See Also
+        --------
+        to_datetime : Convert the given argument to datetime.
+        DatetimeIndex.normalize : Return DatetimeIndex with times to midnight.
+        DatetimeIndex.round : Round the DatetimeIndex to the specified freq.
+        DatetimeIndex.floor : Floor the DatetimeIndex to the specified freq.
+        Timestamp.strftime : Format a single Timestamp.
+        Period.strftime : Format a single Period.
+
+        Examples
+        --------
+        >>> rng = pd.date_range(pd.Timestamp("2018-03-10 09:00"),
+        ...                     periods=3, freq='s')
+        >>> rng.strftime('%%B %%d, %%Y, %%r')
+        Index(['March 10, 2018, 09:00:00 AM', 'March 10, 2018, 09:00:01 AM',
+               'March 10, 2018, 09:00:02 AM'],
+              dtype='object')
+        """
+        result = self._format_native_types(date_format=date_format, na_rep=np.nan)
+        return result.astype(object, copy=False)
+
+
+_round_doc = """
+    Perform {op} operation on the data to the specified `freq`.
+
+    Parameters
+    ----------
+    freq : str or Offset
+        The frequency level to {op} the index to. Must be a fixed
+        frequency like 'S' (second) not 'ME' (month end). See
+        :ref:`frequency aliases <timeseries.offset_aliases>` for
+        a list of possible `freq` values.
+    ambiguous : 'infer', bool-ndarray, 'NaT', default 'raise'
+        Only relevant for DatetimeIndex:
+
+        - 'infer' will attempt to infer fall dst-transition hours based on
+          order
+        - bool-ndarray where True signifies a DST time, False designates
+          a non-DST time (note that this flag is only applicable for
+          ambiguous times)
+        - 'NaT' will return NaT where there are ambiguous times
+        - 'raise' will raise an AmbiguousTimeError if there are ambiguous
+          times.
+
+    nonexistent : 'shift_forward', 'shift_backward', 'NaT', timedelta, default 'raise'
+        A nonexistent time does not exist in a particular timezone
+        where clocks moved forward due to DST.
+
+        - 'shift_forward' will shift the nonexistent time forward to the
+          closest existing time
+        - 'shift_backward' will shift the nonexistent time backward to the
+          closest existing time
+        - 'NaT' will return NaT where there are nonexistent times
+        - timedelta objects will shift nonexistent times by the timedelta
+        - 'raise' will raise an NonExistentTimeError if there are
+          nonexistent times.
+
+    Returns
+    -------
+    DatetimeIndex, TimedeltaIndex, or Series
+        Index of the same type for a DatetimeIndex or TimedeltaIndex,
+        or a Series with the same index for a Series.
+
+    Raises
+    ------
+    ValueError if the `freq` cannot be converted.
+
+    Notes
+    -----
+    If the timestamps have a timezone, {op}ing will take place relative to the
+    local ("wall") time and re-localized to the same timezone. When {op}ing
+    near daylight savings time, use ``nonexistent`` and ``ambiguous`` to
+    control the re-localization behavior.
+
+    Examples
+    --------
+    **DatetimeIndex**
+
+    >>> rng = pd.date_range('1/1/2018 11:59:00', periods=3, freq='min')
+    >>> rng
+    DatetimeIndex(['2018-01-01 11:59:00', '2018-01-01 12:00:00',
+                   '2018-01-01 12:01:00'],
+                  dtype='datetime64[ns]', freq='min')
+    """
+
+_round_example = """>>> rng.round('h')
+    DatetimeIndex(['2018-01-01 12:00:00', '2018-01-01 12:00:00',
+                   '2018-01-01 12:00:00'],
+                  dtype='datetime64[ns]', freq=None)
+
+    **Series**
+
+    >>> pd.Series(rng).dt.round("h")
+    0   2018-01-01 12:00:00
+    1   2018-01-01 12:00:00
+    2   2018-01-01 12:00:00
+    dtype: datetime64[ns]
+
+    When rounding near a daylight savings time transition, use ``ambiguous`` or
+    ``nonexistent`` to control how the timestamp should be re-localized.
+
+    >>> rng_tz = pd.DatetimeIndex(["2021-10-31 03:30:00"], tz="Europe/Amsterdam")
+
+    >>> rng_tz.floor("2h", ambiguous=False)
+    DatetimeIndex(['2021-10-31 02:00:00+01:00'],
+                  dtype='datetime64[ns, Europe/Amsterdam]', freq=None)
+
+    >>> rng_tz.floor("2h", ambiguous=True)
+    DatetimeIndex(['2021-10-31 02:00:00+02:00'],
+                  dtype='datetime64[ns, Europe/Amsterdam]', freq=None)
+    """
+
+_floor_example = """>>> rng.floor('h')
+    DatetimeIndex(['2018-01-01 11:00:00', '2018-01-01 12:00:00',
+                   '2018-01-01 12:00:00'],
+                  dtype='datetime64[ns]', freq=None)
+
+    **Series**
+
+    >>> pd.Series(rng).dt.floor("h")
+    0   2018-01-01 11:00:00
+    1   2018-01-01 12:00:00
+    2   2018-01-01 12:00:00
+    dtype: datetime64[ns]
+
+    When rounding near a daylight savings time transition, use ``ambiguous`` or
+    ``nonexistent`` to control how the timestamp should be re-localized.
+
+    >>> rng_tz = pd.DatetimeIndex(["2021-10-31 03:30:00"], tz="Europe/Amsterdam")
+
+    >>> rng_tz.floor("2h", ambiguous=False)
+    DatetimeIndex(['2021-10-31 02:00:00+01:00'],
+                 dtype='datetime64[ns, Europe/Amsterdam]', freq=None)
+
+    >>> rng_tz.floor("2h", ambiguous=True)
+    DatetimeIndex(['2021-10-31 02:00:00+02:00'],
+                  dtype='datetime64[ns, Europe/Amsterdam]', freq=None)
+    """
+
+_ceil_example = """>>> rng.ceil('h')
+    DatetimeIndex(['2018-01-01 12:00:00', '2018-01-01 12:00:00',
+                   '2018-01-01 13:00:00'],
+                  dtype='datetime64[ns]', freq=None)
+
+    **Series**
+
+    >>> pd.Series(rng).dt.ceil("h")
+    0   2018-01-01 12:00:00
+    1   2018-01-01 12:00:00
+    2   2018-01-01 13:00:00
+    dtype: datetime64[ns]
+
+    When rounding near a daylight savings time transition, use ``ambiguous`` or
+    ``nonexistent`` to control how the timestamp should be re-localized.
+
+    >>> rng_tz = pd.DatetimeIndex(["2021-10-31 01:30:00"], tz="Europe/Amsterdam")
+
+    >>> rng_tz.ceil("h", ambiguous=False)
+    DatetimeIndex(['2021-10-31 02:00:00+01:00'],
+                  dtype='datetime64[ns, Europe/Amsterdam]', freq=None)
+
+    >>> rng_tz.ceil("h", ambiguous=True)
+    DatetimeIndex(['2021-10-31 02:00:00+02:00'],
+                  dtype='datetime64[ns, Europe/Amsterdam]', freq=None)
+    """
+
+
+class TimelikeOps(DatetimeLikeArrayMixin):
+    """
+    Common ops for TimedeltaIndex/DatetimeIndex, but not PeriodIndex.
+    """
+
+    _default_dtype: np.dtype
+
+    def __init__(
+        self, values, dtype=None, freq=lib.no_default, copy: bool = False
+    ) -> None:
+        warnings.warn(
+            # GH#55623
+            f"{type(self).__name__}.__init__ is deprecated and will be "
+            "removed in a future version. Use pd.array instead.",
+            FutureWarning,
+            stacklevel=find_stack_level(),
+        )
+        if dtype is not None:
+            dtype = pandas_dtype(dtype)
+
+        values = extract_array(values, extract_numpy=True)
+        if isinstance(values, IntegerArray):
+            values = values.to_numpy("int64", na_value=iNaT)
+
+        inferred_freq = getattr(values, "_freq", None)
+        explicit_none = freq is None
+        freq = freq if freq is not lib.no_default else None
+
+        if isinstance(values, type(self)):
+            if explicit_none:
+                # don't inherit from values
+                pass
+            elif freq is None:
+                freq = values.freq
+            elif freq and values.freq:
+                freq = to_offset(freq)
+                freq = _validate_inferred_freq(freq, values.freq)
+
+            if dtype is not None and dtype != values.dtype:
+                # TODO: we only have tests for this for DTA, not TDA (2022-07-01)
+                raise TypeError(
+                    f"dtype={dtype} does not match data dtype {values.dtype}"
+                )
+
+            dtype = values.dtype
+            values = values._ndarray
+
+        elif dtype is None:
+            if isinstance(values, np.ndarray) and values.dtype.kind in "Mm":
+                dtype = values.dtype
+            else:
+                dtype = self._default_dtype
+                if isinstance(values, np.ndarray) and values.dtype == "i8":
+                    values = values.view(dtype)
+
+        if not isinstance(values, np.ndarray):
+            raise ValueError(
+                f"Unexpected type '{type(values).__name__}'. 'values' must be a "
+                f"{type(self).__name__}, ndarray, or Series or Index "
+                "containing one of those."
+            )
+        if values.ndim not in [1, 2]:
+            raise ValueError("Only 1-dimensional input arrays are supported.")
+
+        if values.dtype == "i8":
+            # for compat with datetime/timedelta/period shared methods,
+            #  we can sometimes get here with int64 values.  These represent
+            #  nanosecond UTC (or tz-naive) unix timestamps
+            if dtype is None:
+                dtype = self._default_dtype
+                values = values.view(self._default_dtype)
+            elif lib.is_np_dtype(dtype, "mM"):
+                values = values.view(dtype)
+            elif isinstance(dtype, DatetimeTZDtype):
+                kind = self._default_dtype.kind
+                new_dtype = f"{kind}8[{dtype.unit}]"
+                values = values.view(new_dtype)
+
+        dtype = self._validate_dtype(values, dtype)
+
+        if freq == "infer":
+            raise ValueError(
+                f"Frequency inference not allowed in {type(self).__name__}.__init__. "
+                "Use 'pd.array()' instead."
+            )
+
+        if copy:
+            values = values.copy()
+        if freq:
+            freq = to_offset(freq)
+            if values.dtype.kind == "m" and not isinstance(freq, Tick):
+                raise TypeError("TimedeltaArray/Index freq must be a Tick")
+
+        NDArrayBacked.__init__(self, values=values, dtype=dtype)
+        self._freq = freq
+
+        if inferred_freq is None and freq is not None:
+            type(self)._validate_frequency(self, freq)
+
+    @classmethod
+    def _validate_dtype(cls, values, dtype):
+        raise AbstractMethodError(cls)
+
+    @property
+    def freq(self):
+        """
+        Return the frequency object if it is set, otherwise None.
+        """
+        return self._freq
+
+    @freq.setter
+    def freq(self, value) -> None:
+        if value is not None:
+            value = to_offset(value)
+            self._validate_frequency(self, value)
+            if self.dtype.kind == "m" and not isinstance(value, Tick):
+                raise TypeError("TimedeltaArray/Index freq must be a Tick")
+
+            if self.ndim > 1:
+                raise ValueError("Cannot set freq with ndim > 1")
+
+        self._freq = value
+
+    @final
+    def _maybe_pin_freq(self, freq, validate_kwds: dict):
+        """
+        Constructor helper to pin the appropriate `freq` attribute.  Assumes
+        that self._freq is currently set to any freq inferred in
+        _from_sequence_not_strict.
+        """
+        if freq is None:
+            # user explicitly passed None -> override any inferred_freq
+            self._freq = None
+        elif freq == "infer":
+            # if self._freq is *not* None then we already inferred a freq
+            #  and there is nothing left to do
+            if self._freq is None:
+                # Set _freq directly to bypass duplicative _validate_frequency
+                # check.
+                self._freq = to_offset(self.inferred_freq)
+        elif freq is lib.no_default:
+            # user did not specify anything, keep inferred freq if the original
+            #  data had one, otherwise do nothing
+            pass
+        elif self._freq is None:
+            # We cannot inherit a freq from the data, so we need to validate
+            #  the user-passed freq
+            freq = to_offset(freq)
+            type(self)._validate_frequency(self, freq, **validate_kwds)
+            self._freq = freq
+        else:
+            # Otherwise we just need to check that the user-passed freq
+            #  doesn't conflict with the one we already have.
+            freq = to_offset(freq)
+            _validate_inferred_freq(freq, self._freq)
+
+    @final
+    @classmethod
+    def _validate_frequency(cls, index, freq: BaseOffset, **kwargs):
+        """
+        Validate that a frequency is compatible with the values of a given
+        Datetime Array/Index or Timedelta Array/Index
+
+        Parameters
+        ----------
+        index : DatetimeIndex or TimedeltaIndex
+            The index on which to determine if the given frequency is valid
+        freq : DateOffset
+            The frequency to validate
+        """
+        inferred = index.inferred_freq
+        if index.size == 0 or inferred == freq.freqstr:
+            return None
+
+        try:
+            on_freq = cls._generate_range(
+                start=index[0],
+                end=None,
+                periods=len(index),
+                freq=freq,
+                unit=index.unit,
+                **kwargs,
+            )
+            if not np.array_equal(index.asi8, on_freq.asi8):
+                raise ValueError
+        except ValueError as err:
+            if "non-fixed" in str(err):
+                # non-fixed frequencies are not meaningful for timedelta64;
+                #  we retain that error message
+                raise err
+            # GH#11587 the main way this is reached is if the `np.array_equal`
+            #  check above is False.  This can also be reached if index[0]
+            #  is `NaT`, in which case the call to `cls._generate_range` will
+            #  raise a ValueError, which we re-raise with a more targeted
+            #  message.
+            raise ValueError(
+                f"Inferred frequency {inferred} from passed values "
+                f"does not conform to passed frequency {freq.freqstr}"
+            ) from err
+
+    @classmethod
+    def _generate_range(
+        cls, start, end, periods: int | None, freq, *args, **kwargs
+    ) -> Self:
+        raise AbstractMethodError(cls)
+
+    # --------------------------------------------------------------
+
+    @cache_readonly
+    def _creso(self) -> int:
+        return get_unit_from_dtype(self._ndarray.dtype)
+
+    @cache_readonly
+    def unit(self) -> str:
+        # e.g. "ns", "us", "ms"
+        # error: Argument 1 to "dtype_to_unit" has incompatible type
+        # "ExtensionDtype"; expected "Union[DatetimeTZDtype, dtype[Any]]"
+        return dtype_to_unit(self.dtype)  # type: ignore[arg-type]
+
+    def as_unit(self, unit: str, round_ok: bool = True) -> Self:
+        if unit not in ["s", "ms", "us", "ns"]:
+            raise ValueError("Supported units are 's', 'ms', 'us', 'ns'")
+
+        dtype = np.dtype(f"{self.dtype.kind}8[{unit}]")
+        new_values = astype_overflowsafe(self._ndarray, dtype, round_ok=round_ok)
+
+        if isinstance(self.dtype, np.dtype):
+            new_dtype = new_values.dtype
+        else:
+            tz = cast("DatetimeArray", self).tz
+            new_dtype = DatetimeTZDtype(tz=tz, unit=unit)
+
+        # error: Unexpected keyword argument "freq" for "_simple_new" of
+        # "NDArrayBacked"  [call-arg]
+        return type(self)._simple_new(
+            new_values, dtype=new_dtype, freq=self.freq  # type: ignore[call-arg]
+        )
+
+    # TODO: annotate other as DatetimeArray | TimedeltaArray | Timestamp | Timedelta
+    #  with the return type matching input type.  TypeVar?
+    def _ensure_matching_resos(self, other):
+        if self._creso != other._creso:
+            # Just as with Timestamp/Timedelta, we cast to the higher resolution
+            if self._creso < other._creso:
+                self = self.as_unit(other.unit)
+            else:
+                other = other.as_unit(self.unit)
+        return self, other
+
+    # --------------------------------------------------------------
+
+    def __array_ufunc__(self, ufunc: np.ufunc, method: str, *inputs, **kwargs):
+        if (
+            ufunc in [np.isnan, np.isinf, np.isfinite]
+            and len(inputs) == 1
+            and inputs[0] is self
+        ):
+            # numpy 1.18 changed isinf and isnan to not raise on dt64/td64
+            return getattr(ufunc, method)(self._ndarray, **kwargs)
+
+        return super().__array_ufunc__(ufunc, method, *inputs, **kwargs)
+
+    def _round(self, freq, mode, ambiguous, nonexistent):
+        # round the local times
+        if isinstance(self.dtype, DatetimeTZDtype):
+            # operate on naive timestamps, then convert back to aware
+            self = cast("DatetimeArray", self)
+            naive = self.tz_localize(None)
+            result = naive._round(freq, mode, ambiguous, nonexistent)
+            return result.tz_localize(
+                self.tz, ambiguous=ambiguous, nonexistent=nonexistent
+            )
+
+        values = self.view("i8")
+        values = cast(np.ndarray, values)
+        nanos = get_unit_for_round(freq, self._creso)
+        if nanos == 0:
+            # GH 52761
+            return self.copy()
+        result_i8 = round_nsint64(values, mode, nanos)
+        result = self._maybe_mask_results(result_i8, fill_value=iNaT)
+        result = result.view(self._ndarray.dtype)
+        return self._simple_new(result, dtype=self.dtype)
+
+    @Appender((_round_doc + _round_example).format(op="round"))
+    def round(
+        self,
+        freq,
+        ambiguous: TimeAmbiguous = "raise",
+        nonexistent: TimeNonexistent = "raise",
+    ) -> Self:
+        return self._round(freq, RoundTo.NEAREST_HALF_EVEN, ambiguous, nonexistent)
+
+    @Appender((_round_doc + _floor_example).format(op="floor"))
+    def floor(
+        self,
+        freq,
+        ambiguous: TimeAmbiguous = "raise",
+        nonexistent: TimeNonexistent = "raise",
+    ) -> Self:
+        return self._round(freq, RoundTo.MINUS_INFTY, ambiguous, nonexistent)
+
+    @Appender((_round_doc + _ceil_example).format(op="ceil"))
+    def ceil(
+        self,
+        freq,
+        ambiguous: TimeAmbiguous = "raise",
+        nonexistent: TimeNonexistent = "raise",
+    ) -> Self:
+        return self._round(freq, RoundTo.PLUS_INFTY, ambiguous, nonexistent)
+
+    # --------------------------------------------------------------
+    # Reductions
+
+    def any(self, *, axis: AxisInt | None = None, skipna: bool = True) -> bool:
+        # GH#34479 the nanops call will issue a FutureWarning for non-td64 dtype
+        return nanops.nanany(self._ndarray, axis=axis, skipna=skipna, mask=self.isna())
+
+    def all(self, *, axis: AxisInt | None = None, skipna: bool = True) -> bool:
+        # GH#34479 the nanops call will issue a FutureWarning for non-td64 dtype
+
+        return nanops.nanall(self._ndarray, axis=axis, skipna=skipna, mask=self.isna())
+
+    # --------------------------------------------------------------
+    # Frequency Methods
+
+    def _maybe_clear_freq(self) -> None:
+        self._freq = None
+
+    def _with_freq(self, freq) -> Self:
+        """
+        Helper to get a view on the same data, with a new freq.
+
+        Parameters
+        ----------
+        freq : DateOffset, None, or "infer"
+
+        Returns
+        -------
+        Same type as self
+        """
+        # GH#29843
+        if freq is None:
+            # Always valid
+            pass
+        elif len(self) == 0 and isinstance(freq, BaseOffset):
+            # Always valid.  In the TimedeltaArray case, we require a Tick offset
+            if self.dtype.kind == "m" and not isinstance(freq, Tick):
+                raise TypeError("TimedeltaArray/Index freq must be a Tick")
+        else:
+            # As an internal method, we can ensure this assertion always holds
+            assert freq == "infer"
+            freq = to_offset(self.inferred_freq)
+
+        arr = self.view()
+        arr._freq = freq
+        return arr
+
+    # --------------------------------------------------------------
+    # ExtensionArray Interface
+
+    def _values_for_json(self) -> np.ndarray:
+        # Small performance bump vs the base class which calls np.asarray(self)
+        if isinstance(self.dtype, np.dtype):
+            return self._ndarray
+        return super()._values_for_json()
+
+    def factorize(
+        self,
+        use_na_sentinel: bool = True,
+        sort: bool = False,
+    ):
+        if self.freq is not None:
+            # We must be unique, so can short-circuit (and retain freq)
+            codes = np.arange(len(self), dtype=np.intp)
+            uniques = self.copy()  # TODO: copy or view?
+            if sort and self.freq.n < 0:
+                codes = codes[::-1]
+                uniques = uniques[::-1]
+            return codes, uniques
+
+        if sort:
+            # algorithms.factorize only passes sort=True here when freq is
+            #  not None, so this should not be reached.
+            raise NotImplementedError(
+                f"The 'sort' keyword in {type(self).__name__}.factorize is "
+                "ignored unless arr.freq is not None. To factorize with sort, "
+                "call pd.factorize(obj, sort=True) instead."
+            )
+        return super().factorize(use_na_sentinel=use_na_sentinel)
+
+    @classmethod
+    def _concat_same_type(
+        cls,
+        to_concat: Sequence[Self],
+        axis: AxisInt = 0,
+    ) -> Self:
+        new_obj = super()._concat_same_type(to_concat, axis)
+
+        obj = to_concat[0]
+
+        if axis == 0:
+            # GH 3232: If the concat result is evenly spaced, we can retain the
+            # original frequency
+            to_concat = [x for x in to_concat if len(x)]
+
+            if obj.freq is not None and all(x.freq == obj.freq for x in to_concat):
+                pairs = zip(to_concat[:-1], to_concat[1:])
+                if all(pair[0][-1] + obj.freq == pair[1][0] for pair in pairs):
+                    new_freq = obj.freq
+                    new_obj._freq = new_freq
+        return new_obj
+
+    def copy(self, order: str = "C") -> Self:
+        new_obj = super().copy(order=order)
+        new_obj._freq = self.freq
+        return new_obj
+
+    def interpolate(
+        self,
+        *,
+        method: InterpolateOptions,
+        axis: int,
+        index: Index,
+        limit,
+        limit_direction,
+        limit_area,
+        copy: bool,
+        **kwargs,
+    ) -> Self:
+        """
+        See NDFrame.interpolate.__doc__.
+        """
+        # NB: we return type(self) even if copy=False
+        if method != "linear":
+            raise NotImplementedError
+
+        if not copy:
+            out_data = self._ndarray
+        else:
+            out_data = self._ndarray.copy()
+
+        missing.interpolate_2d_inplace(
+            out_data,
+            method=method,
+            axis=axis,
+            index=index,
+            limit=limit,
+            limit_direction=limit_direction,
+            limit_area=limit_area,
+            **kwargs,
+        )
+        if not copy:
+            return self
+        return type(self)._simple_new(out_data, dtype=self.dtype)
+
+    # --------------------------------------------------------------
+    # Unsorted
+
+    @property
+    def _is_dates_only(self) -> bool:
+        """
+        Check if we are round times at midnight (and no timezone), which will
+        be given a more compact __repr__ than other cases. For TimedeltaArray
+        we are checking for multiples of 24H.
+        """
+        if not lib.is_np_dtype(self.dtype):
+            # i.e. we have a timezone
+            return False
+
+        values_int = self.asi8
+        consider_values = values_int != iNaT
+        reso = get_unit_from_dtype(self.dtype)
+        ppd = periods_per_day(reso)
+
+        # TODO: can we reuse is_date_array_normalized?  would need a skipna kwd
+        #  (first attempt at this was less performant than this implementation)
+        even_days = np.logical_and(consider_values, values_int % ppd != 0).sum() == 0
+        return even_days
+
+
+# -------------------------------------------------------------------
+# Shared Constructor Helpers
+
+
+def ensure_arraylike_for_datetimelike(
+    data, copy: bool, cls_name: str
+) -> tuple[ArrayLike, bool]:
+    if not hasattr(data, "dtype"):
+        # e.g. list, tuple
+        if not isinstance(data, (list, tuple)) and np.ndim(data) == 0:
+            # i.e. generator
+            data = list(data)
+
+        data = construct_1d_object_array_from_listlike(data)
+        copy = False
+    elif isinstance(data, ABCMultiIndex):
+        raise TypeError(f"Cannot create a {cls_name} from a MultiIndex.")
+    else:
+        data = extract_array(data, extract_numpy=True)
+
+    if isinstance(data, IntegerArray) or (
+        isinstance(data, ArrowExtensionArray) and data.dtype.kind in "iu"
+    ):
+        data = data.to_numpy("int64", na_value=iNaT)
+        copy = False
+    elif isinstance(data, ArrowExtensionArray):
+        data = data._maybe_convert_datelike_array()
+        data = data.to_numpy()
+        copy = False
+    elif not isinstance(data, (np.ndarray, ExtensionArray)):
+        # GH#24539 e.g. xarray, dask object
+        data = np.asarray(data)
+
+    elif isinstance(data, ABCCategorical):
+        # GH#18664 preserve tz in going DTI->Categorical->DTI
+        # TODO: cases where we need to do another pass through maybe_convert_dtype,
+        #  e.g. the categories are timedelta64s
+        data = data.categories.take(data.codes, fill_value=NaT)._values
+        copy = False
+
+    return data, copy
+
+
+@overload
+def validate_periods(periods: None) -> None:
+    ...
+
+
+@overload
+def validate_periods(periods: int | float) -> int:
+    ...
+
+
+def validate_periods(periods: int | float | None) -> int | None:
+    """
+    If a `periods` argument is passed to the Datetime/Timedelta Array/Index
+    constructor, cast it to an integer.
+
+    Parameters
+    ----------
+    periods : None, float, int
+
+    Returns
+    -------
+    periods : None or int
+
+    Raises
+    ------
+    TypeError
+        if periods is None, float, or int
+    """
+    if periods is not None:
+        if lib.is_float(periods):
+            warnings.warn(
+                # GH#56036
+                "Non-integer 'periods' in pd.date_range, pd.timedelta_range, "
+                "pd.period_range, and pd.interval_range are deprecated and "
+                "will raise in a future version.",
+                FutureWarning,
+                stacklevel=find_stack_level(),
+            )
+            periods = int(periods)
+        elif not lib.is_integer(periods):
+            raise TypeError(f"periods must be a number, got {periods}")
+    return periods
+
+
+def _validate_inferred_freq(
+    freq: BaseOffset | None, inferred_freq: BaseOffset | None
+) -> BaseOffset | None:
+    """
+    If the user passes a freq and another freq is inferred from passed data,
+    require that they match.
+
+    Parameters
+    ----------
+    freq : DateOffset or None
+    inferred_freq : DateOffset or None
+
+    Returns
+    -------
+    freq : DateOffset or None
+    """
+    if inferred_freq is not None:
+        if freq is not None and freq != inferred_freq:
+            raise ValueError(
+                f"Inferred frequency {inferred_freq} from passed "
+                "values does not conform to passed frequency "
+                f"{freq.freqstr}"
+            )
+        if freq is None:
+            freq = inferred_freq
+
+    return freq
+
+
+def dtype_to_unit(dtype: DatetimeTZDtype | np.dtype | ArrowDtype) -> str:
+    """
+    Return the unit str corresponding to the dtype's resolution.
+
+    Parameters
+    ----------
+    dtype : DatetimeTZDtype or np.dtype
+        If np.dtype, we assume it is a datetime64 dtype.
+
+    Returns
+    -------
+    str
+    """
+    if isinstance(dtype, DatetimeTZDtype):
+        return dtype.unit
+    elif isinstance(dtype, ArrowDtype):
+        if dtype.kind not in "mM":
+            raise ValueError(f"{dtype=} does not have a resolution.")
+        return dtype.pyarrow_dtype.unit
+    return np.datetime_data(dtype)[0]

llava_next/lib/python3.10/site-packages/pandas/core/arrays/integer.py

@@ -0,0 +1,272 @@
+from __future__ import annotations
+
+from typing import ClassVar
+
+import numpy as np
+
+from pandas.core.dtypes.base import register_extension_dtype
+from pandas.core.dtypes.common import is_integer_dtype
+
+from pandas.core.arrays.numeric import (
+    NumericArray,
+    NumericDtype,
+)
+
+
+class IntegerDtype(NumericDtype):
+    """
+    An ExtensionDtype to hold a single size & kind of integer dtype.
+
+    These specific implementations are subclasses of the non-public
+    IntegerDtype. For example, we have Int8Dtype to represent signed int 8s.
+
+    The attributes name & type are set when these subclasses are created.
+    """
+
+    _default_np_dtype = np.dtype(np.int64)
+    _checker = is_integer_dtype
+
+    @classmethod
+    def construct_array_type(cls) -> type[IntegerArray]:
+        """
+        Return the array type associated with this dtype.
+
+        Returns
+        -------
+        type
+        """
+        return IntegerArray
+
+    @classmethod
+    def _get_dtype_mapping(cls) -> dict[np.dtype, IntegerDtype]:
+        return NUMPY_INT_TO_DTYPE
+
+    @classmethod
+    def _safe_cast(cls, values: np.ndarray, dtype: np.dtype, copy: bool) -> np.ndarray:
+        """
+        Safely cast the values to the given dtype.
+
+        "safe" in this context means the casting is lossless. e.g. if 'values'
+        has a floating dtype, each value must be an integer.
+        """
+        try:
+            return values.astype(dtype, casting="safe", copy=copy)
+        except TypeError as err:
+            casted = values.astype(dtype, copy=copy)
+            if (casted == values).all():
+                return casted
+
+            raise TypeError(
+                f"cannot safely cast non-equivalent {values.dtype} to {np.dtype(dtype)}"
+            ) from err
+
+
+class IntegerArray(NumericArray):
+    """
+    Array of integer (optional missing) values.
+
+    Uses :attr:`pandas.NA` as the missing value.
+
+    .. warning::
+
+       IntegerArray is currently experimental, and its API or internal
+       implementation may change without warning.
+
+    We represent an IntegerArray with 2 numpy arrays:
+
+    - data: contains a numpy integer array of the appropriate dtype
+    - mask: a boolean array holding a mask on the data, True is missing
+
+    To construct an IntegerArray from generic array-like input, use
+    :func:`pandas.array` with one of the integer dtypes (see examples).
+
+    See :ref:`integer_na` for more.
+
+    Parameters
+    ----------
+    values : numpy.ndarray
+        A 1-d integer-dtype array.
+    mask : numpy.ndarray
+        A 1-d boolean-dtype array indicating missing values.
+    copy : bool, default False
+        Whether to copy the `values` and `mask`.
+
+    Attributes
+    ----------
+    None
+
+    Methods
+    -------
+    None
+
+    Returns
+    -------
+    IntegerArray
+
+    Examples
+    --------
+    Create an IntegerArray with :func:`pandas.array`.
+
+    >>> int_array = pd.array([1, None, 3], dtype=pd.Int32Dtype())
+    >>> int_array
+    <IntegerArray>
+    [1, <NA>, 3]
+    Length: 3, dtype: Int32
+
+    String aliases for the dtypes are also available. They are capitalized.
+
+    >>> pd.array([1, None, 3], dtype='Int32')
+    <IntegerArray>
+    [1, <NA>, 3]
+    Length: 3, dtype: Int32
+
+    >>> pd.array([1, None, 3], dtype='UInt16')
+    <IntegerArray>
+    [1, <NA>, 3]
+    Length: 3, dtype: UInt16
+    """
+
+    _dtype_cls = IntegerDtype
+
+    # The value used to fill '_data' to avoid upcasting
+    _internal_fill_value = 1
+    # Fill values used for any/all
+    # Incompatible types in assignment (expression has type "int", base class
+    # "BaseMaskedArray" defined the type as "<typing special form>")
+    _truthy_value = 1  # type: ignore[assignment]
+    _falsey_value = 0  # type: ignore[assignment]
+
+
+_dtype_docstring = """
+An ExtensionDtype for {dtype} integer data.
+
+Uses :attr:`pandas.NA` as its missing value, rather than :attr:`numpy.nan`.
+
+Attributes
+----------
+None
+
+Methods
+-------
+None
+
+Examples
+--------
+For Int8Dtype:
+
+>>> ser = pd.Series([2, pd.NA], dtype=pd.Int8Dtype())
+>>> ser.dtype
+Int8Dtype()
+
+For Int16Dtype:
+
+>>> ser = pd.Series([2, pd.NA], dtype=pd.Int16Dtype())
+>>> ser.dtype
+Int16Dtype()
+
+For Int32Dtype:
+
+>>> ser = pd.Series([2, pd.NA], dtype=pd.Int32Dtype())
+>>> ser.dtype
+Int32Dtype()
+
+For Int64Dtype:
+
+>>> ser = pd.Series([2, pd.NA], dtype=pd.Int64Dtype())
+>>> ser.dtype
+Int64Dtype()
+
+For UInt8Dtype:
+
+>>> ser = pd.Series([2, pd.NA], dtype=pd.UInt8Dtype())
+>>> ser.dtype
+UInt8Dtype()
+
+For UInt16Dtype:
+
+>>> ser = pd.Series([2, pd.NA], dtype=pd.UInt16Dtype())
+>>> ser.dtype
+UInt16Dtype()
+
+For UInt32Dtype:
+
+>>> ser = pd.Series([2, pd.NA], dtype=pd.UInt32Dtype())
+>>> ser.dtype
+UInt32Dtype()
+
+For UInt64Dtype:
+
+>>> ser = pd.Series([2, pd.NA], dtype=pd.UInt64Dtype())
+>>> ser.dtype
+UInt64Dtype()
+"""
+
+# create the Dtype
+
+
+@register_extension_dtype
+class Int8Dtype(IntegerDtype):
+    type = np.int8
+    name: ClassVar[str] = "Int8"
+    __doc__ = _dtype_docstring.format(dtype="int8")
+
+
+@register_extension_dtype
+class Int16Dtype(IntegerDtype):
+    type = np.int16
+    name: ClassVar[str] = "Int16"
+    __doc__ = _dtype_docstring.format(dtype="int16")
+
+
+@register_extension_dtype
+class Int32Dtype(IntegerDtype):
+    type = np.int32
+    name: ClassVar[str] = "Int32"
+    __doc__ = _dtype_docstring.format(dtype="int32")
+
+
+@register_extension_dtype
+class Int64Dtype(IntegerDtype):
+    type = np.int64
+    name: ClassVar[str] = "Int64"
+    __doc__ = _dtype_docstring.format(dtype="int64")
+
+
+@register_extension_dtype
+class UInt8Dtype(IntegerDtype):
+    type = np.uint8
+    name: ClassVar[str] = "UInt8"
+    __doc__ = _dtype_docstring.format(dtype="uint8")
+
+
+@register_extension_dtype
+class UInt16Dtype(IntegerDtype):
+    type = np.uint16
+    name: ClassVar[str] = "UInt16"
+    __doc__ = _dtype_docstring.format(dtype="uint16")
+
+
+@register_extension_dtype
+class UInt32Dtype(IntegerDtype):
+    type = np.uint32
+    name: ClassVar[str] = "UInt32"
+    __doc__ = _dtype_docstring.format(dtype="uint32")
+
+
+@register_extension_dtype
+class UInt64Dtype(IntegerDtype):
+    type = np.uint64
+    name: ClassVar[str] = "UInt64"
+    __doc__ = _dtype_docstring.format(dtype="uint64")
+
+
+NUMPY_INT_TO_DTYPE: dict[np.dtype, IntegerDtype] = {
+    np.dtype(np.int8): Int8Dtype(),
+    np.dtype(np.int16): Int16Dtype(),
+    np.dtype(np.int32): Int32Dtype(),
+    np.dtype(np.int64): Int64Dtype(),
+    np.dtype(np.uint8): UInt8Dtype(),
+    np.dtype(np.uint16): UInt16Dtype(),
+    np.dtype(np.uint32): UInt32Dtype(),
+    np.dtype(np.uint64): UInt64Dtype(),
+}

llava_next/lib/python3.10/site-packages/pandas/core/arrays/masked.py

@@ -0,0 +1,1650 @@
+from __future__ import annotations
+
+from typing import (
+    TYPE_CHECKING,
+    Any,
+    Callable,
+    Literal,
+    overload,
+)
+import warnings
+
+import numpy as np
+
+from pandas._libs import (
+    lib,
+    missing as libmissing,
+)
+from pandas._libs.tslibs import is_supported_dtype
+from pandas._typing import (
+    ArrayLike,
+    AstypeArg,
+    AxisInt,
+    DtypeObj,
+    FillnaOptions,
+    InterpolateOptions,
+    NpDtype,
+    PositionalIndexer,
+    Scalar,
+    ScalarIndexer,
+    Self,
+    SequenceIndexer,
+    Shape,
+    npt,
+)
+from pandas.compat import (
+    IS64,
+    is_platform_windows,
+)
+from pandas.errors import AbstractMethodError
+from pandas.util._decorators import doc
+from pandas.util._validators import validate_fillna_kwargs
+
+from pandas.core.dtypes.base import ExtensionDtype
+from pandas.core.dtypes.common import (
+    is_bool,
+    is_integer_dtype,
+    is_list_like,
+    is_scalar,
+    is_string_dtype,
+    pandas_dtype,
+)
+from pandas.core.dtypes.dtypes import BaseMaskedDtype
+from pandas.core.dtypes.missing import (
+    array_equivalent,
+    is_valid_na_for_dtype,
+    isna,
+    notna,
+)
+
+from pandas.core import (
+    algorithms as algos,
+    arraylike,
+    missing,
+    nanops,
+    ops,
+)
+from pandas.core.algorithms import (
+    factorize_array,
+    isin,
+    map_array,
+    mode,
+    take,
+)
+from pandas.core.array_algos import (
+    masked_accumulations,
+    masked_reductions,
+)
+from pandas.core.array_algos.quantile import quantile_with_mask
+from pandas.core.arraylike import OpsMixin
+from pandas.core.arrays._utils import to_numpy_dtype_inference
+from pandas.core.arrays.base import ExtensionArray
+from pandas.core.construction import (
+    array as pd_array,
+    ensure_wrapped_if_datetimelike,
+    extract_array,
+)
+from pandas.core.indexers import check_array_indexer
+from pandas.core.ops import invalid_comparison
+from pandas.core.util.hashing import hash_array
+
+if TYPE_CHECKING:
+    from collections.abc import (
+        Iterator,
+        Sequence,
+    )
+    from pandas import Series
+    from pandas.core.arrays import BooleanArray
+    from pandas._typing import (
+        NumpySorter,
+        NumpyValueArrayLike,
+    )
+    from pandas.core.arrays import FloatingArray
+
+from pandas.compat.numpy import function as nv
+
+
+class BaseMaskedArray(OpsMixin, ExtensionArray):
+    """
+    Base class for masked arrays (which use _data and _mask to store the data).
+
+    numpy based
+    """
+
+    # The value used to fill '_data' to avoid upcasting
+    _internal_fill_value: Scalar
+    # our underlying data and mask are each ndarrays
+    _data: np.ndarray
+    _mask: npt.NDArray[np.bool_]
+
+    # Fill values used for any/all
+    _truthy_value = Scalar  # bool(_truthy_value) = True
+    _falsey_value = Scalar  # bool(_falsey_value) = False
+
+    @classmethod
+    def _simple_new(cls, values: np.ndarray, mask: npt.NDArray[np.bool_]) -> Self:
+        result = BaseMaskedArray.__new__(cls)
+        result._data = values
+        result._mask = mask
+        return result
+
+    def __init__(
+        self, values: np.ndarray, mask: npt.NDArray[np.bool_], copy: bool = False
+    ) -> None:
+        # values is supposed to already be validated in the subclass
+        if not (isinstance(mask, np.ndarray) and mask.dtype == np.bool_):
+            raise TypeError(
+                "mask should be boolean numpy array. Use "
+                "the 'pd.array' function instead"
+            )
+        if values.shape != mask.shape:
+            raise ValueError("values.shape must match mask.shape")
+
+        if copy:
+            values = values.copy()
+            mask = mask.copy()
+
+        self._data = values
+        self._mask = mask
+
+    @classmethod
+    def _from_sequence(cls, scalars, *, dtype=None, copy: bool = False) -> Self:
+        values, mask = cls._coerce_to_array(scalars, dtype=dtype, copy=copy)
+        return cls(values, mask)
+
+    @classmethod
+    @doc(ExtensionArray._empty)
+    def _empty(cls, shape: Shape, dtype: ExtensionDtype):
+        values = np.empty(shape, dtype=dtype.type)
+        values.fill(cls._internal_fill_value)
+        mask = np.ones(shape, dtype=bool)
+        result = cls(values, mask)
+        if not isinstance(result, cls) or dtype != result.dtype:
+            raise NotImplementedError(
+                f"Default 'empty' implementation is invalid for dtype='{dtype}'"
+            )
+        return result
+
+    def _formatter(self, boxed: bool = False) -> Callable[[Any], str | None]:
+        # NEP 51: https://github.com/numpy/numpy/pull/22449
+        return str
+
+    @property
+    def dtype(self) -> BaseMaskedDtype:
+        raise AbstractMethodError(self)
+
+    @overload
+    def __getitem__(self, item: ScalarIndexer) -> Any:
+        ...
+
+    @overload
+    def __getitem__(self, item: SequenceIndexer) -> Self:
+        ...
+
+    def __getitem__(self, item: PositionalIndexer) -> Self | Any:
+        item = check_array_indexer(self, item)
+
+        newmask = self._mask[item]
+        if is_bool(newmask):
+            # This is a scalar indexing
+            if newmask:
+                return self.dtype.na_value
+            return self._data[item]
+
+        return self._simple_new(self._data[item], newmask)
+
+    def _pad_or_backfill(
+        self,
+        *,
+        method: FillnaOptions,
+        limit: int | None = None,
+        limit_area: Literal["inside", "outside"] | None = None,
+        copy: bool = True,
+    ) -> Self:
+        mask = self._mask
+
+        if mask.any():
+            func = missing.get_fill_func(method, ndim=self.ndim)
+
+            npvalues = self._data.T
+            new_mask = mask.T
+            if copy:
+                npvalues = npvalues.copy()
+                new_mask = new_mask.copy()
+            elif limit_area is not None:
+                mask = mask.copy()
+            func(npvalues, limit=limit, mask=new_mask)
+
+            if limit_area is not None and not mask.all():
+                mask = mask.T
+                neg_mask = ~mask
+                first = neg_mask.argmax()
+                last = len(neg_mask) - neg_mask[::-1].argmax() - 1
+                if limit_area == "inside":
+                    new_mask[:first] |= mask[:first]
+                    new_mask[last + 1 :] |= mask[last + 1 :]
+                elif limit_area == "outside":
+                    new_mask[first + 1 : last] |= mask[first + 1 : last]
+
+            if copy:
+                return self._simple_new(npvalues.T, new_mask.T)
+            else:
+                return self
+        else:
+            if copy:
+                new_values = self.copy()
+            else:
+                new_values = self
+        return new_values
+
+    @doc(ExtensionArray.fillna)
+    def fillna(
+        self, value=None, method=None, limit: int | None = None, copy: bool = True
+    ) -> Self:
+        value, method = validate_fillna_kwargs(value, method)
+
+        mask = self._mask
+
+        value = missing.check_value_size(value, mask, len(self))
+
+        if mask.any():
+            if method is not None:
+                func = missing.get_fill_func(method, ndim=self.ndim)
+                npvalues = self._data.T
+                new_mask = mask.T
+                if copy:
+                    npvalues = npvalues.copy()
+                    new_mask = new_mask.copy()
+                func(npvalues, limit=limit, mask=new_mask)
+                return self._simple_new(npvalues.T, new_mask.T)
+            else:
+                # fill with value
+                if copy:
+                    new_values = self.copy()
+                else:
+                    new_values = self[:]
+                new_values[mask] = value
+        else:
+            if copy:
+                new_values = self.copy()
+            else:
+                new_values = self[:]
+        return new_values
+
+    @classmethod
+    def _coerce_to_array(
+        cls, values, *, dtype: DtypeObj, copy: bool = False
+    ) -> tuple[np.ndarray, np.ndarray]:
+        raise AbstractMethodError(cls)
+
+    def _validate_setitem_value(self, value):
+        """
+        Check if we have a scalar that we can cast losslessly.
+
+        Raises
+        ------
+        TypeError
+        """
+        kind = self.dtype.kind
+        # TODO: get this all from np_can_hold_element?
+        if kind == "b":
+            if lib.is_bool(value):
+                return value
+
+        elif kind == "f":
+            if lib.is_integer(value) or lib.is_float(value):
+                return value
+
+        else:
+            if lib.is_integer(value) or (lib.is_float(value) and value.is_integer()):
+                return value
+            # TODO: unsigned checks
+
+        # Note: without the "str" here, the f-string rendering raises in
+        #  py38 builds.
+        raise TypeError(f"Invalid value '{str(value)}' for dtype {self.dtype}")
+
+    def __setitem__(self, key, value) -> None:
+        key = check_array_indexer(self, key)
+
+        if is_scalar(value):
+            if is_valid_na_for_dtype(value, self.dtype):
+                self._mask[key] = True
+            else:
+                value = self._validate_setitem_value(value)
+                self._data[key] = value
+                self._mask[key] = False
+            return
+
+        value, mask = self._coerce_to_array(value, dtype=self.dtype)
+
+        self._data[key] = value
+        self._mask[key] = mask
+
+    def __contains__(self, key) -> bool:
+        if isna(key) and key is not self.dtype.na_value:
+            # GH#52840
+            if self._data.dtype.kind == "f" and lib.is_float(key):
+                return bool((np.isnan(self._data) & ~self._mask).any())
+
+        return bool(super().__contains__(key))
+
+    def __iter__(self) -> Iterator:
+        if self.ndim == 1:
+            if not self._hasna:
+                for val in self._data:
+                    yield val
+            else:
+                na_value = self.dtype.na_value
+                for isna_, val in zip(self._mask, self._data):
+                    if isna_:
+                        yield na_value
+                    else:
+                        yield val
+        else:
+            for i in range(len(self)):
+                yield self[i]
+
+    def __len__(self) -> int:
+        return len(self._data)
+
+    @property
+    def shape(self) -> Shape:
+        return self._data.shape
+
+    @property
+    def ndim(self) -> int:
+        return self._data.ndim
+
+    def swapaxes(self, axis1, axis2) -> Self:
+        data = self._data.swapaxes(axis1, axis2)
+        mask = self._mask.swapaxes(axis1, axis2)
+        return self._simple_new(data, mask)
+
+    def delete(self, loc, axis: AxisInt = 0) -> Self:
+        data = np.delete(self._data, loc, axis=axis)
+        mask = np.delete(self._mask, loc, axis=axis)
+        return self._simple_new(data, mask)
+
+    def reshape(self, *args, **kwargs) -> Self:
+        data = self._data.reshape(*args, **kwargs)
+        mask = self._mask.reshape(*args, **kwargs)
+        return self._simple_new(data, mask)
+
+    def ravel(self, *args, **kwargs) -> Self:
+        # TODO: need to make sure we have the same order for data/mask
+        data = self._data.ravel(*args, **kwargs)
+        mask = self._mask.ravel(*args, **kwargs)
+        return type(self)(data, mask)
+
+    @property
+    def T(self) -> Self:
+        return self._simple_new(self._data.T, self._mask.T)
+
+    def round(self, decimals: int = 0, *args, **kwargs):
+        """
+        Round each value in the array a to the given number of decimals.
+
+        Parameters
+        ----------
+        decimals : int, default 0
+            Number of decimal places to round to. If decimals is negative,
+            it specifies the number of positions to the left of the decimal point.
+        *args, **kwargs
+            Additional arguments and keywords have no effect but might be
+            accepted for compatibility with NumPy.
+
+        Returns
+        -------
+        NumericArray
+            Rounded values of the NumericArray.
+
+        See Also
+        --------
+        numpy.around : Round values of an np.array.
+        DataFrame.round : Round values of a DataFrame.
+        Series.round : Round values of a Series.
+        """
+        if self.dtype.kind == "b":
+            return self
+        nv.validate_round(args, kwargs)
+        values = np.round(self._data, decimals=decimals, **kwargs)
+
+        # Usually we'll get same type as self, but ndarray[bool] casts to float
+        return self._maybe_mask_result(values, self._mask.copy())
+
+    # ------------------------------------------------------------------
+    # Unary Methods
+
+    def __invert__(self) -> Self:
+        return self._simple_new(~self._data, self._mask.copy())
+
+    def __neg__(self) -> Self:
+        return self._simple_new(-self._data, self._mask.copy())
+
+    def __pos__(self) -> Self:
+        return self.copy()
+
+    def __abs__(self) -> Self:
+        return self._simple_new(abs(self._data), self._mask.copy())
+
+    # ------------------------------------------------------------------
+
+    def _values_for_json(self) -> np.ndarray:
+        return np.asarray(self, dtype=object)
+
+    def to_numpy(
+        self,
+        dtype: npt.DTypeLike | None = None,
+        copy: bool = False,
+        na_value: object = lib.no_default,
+    ) -> np.ndarray:
+        """
+        Convert to a NumPy Array.
+
+        By default converts to an object-dtype NumPy array. Specify the `dtype` and
+        `na_value` keywords to customize the conversion.
+
+        Parameters
+        ----------
+        dtype : dtype, default object
+            The numpy dtype to convert to.
+        copy : bool, default False
+            Whether to ensure that the returned value is a not a view on
+            the array. Note that ``copy=False`` does not *ensure* that
+            ``to_numpy()`` is no-copy. Rather, ``copy=True`` ensure that
+            a copy is made, even if not strictly necessary. This is typically
+            only possible when no missing values are present and `dtype`
+            is the equivalent numpy dtype.
+        na_value : scalar, optional
+             Scalar missing value indicator to use in numpy array. Defaults
+             to the native missing value indicator of this array (pd.NA).
+
+        Returns
+        -------
+        numpy.ndarray
+
+        Examples
+        --------
+        An object-dtype is the default result
+
+        >>> a = pd.array([True, False, pd.NA], dtype="boolean")
+        >>> a.to_numpy()
+        array([True, False, <NA>], dtype=object)
+
+        When no missing values are present, an equivalent dtype can be used.
+
+        >>> pd.array([True, False], dtype="boolean").to_numpy(dtype="bool")
+        array([ True, False])
+        >>> pd.array([1, 2], dtype="Int64").to_numpy("int64")
+        array([1, 2])
+
+        However, requesting such dtype will raise a ValueError if
+        missing values are present and the default missing value :attr:`NA`
+        is used.
+
+        >>> a = pd.array([True, False, pd.NA], dtype="boolean")
+        >>> a
+        <BooleanArray>
+        [True, False, <NA>]
+        Length: 3, dtype: boolean
+
+        >>> a.to_numpy(dtype="bool")
+        Traceback (most recent call last):
+        ...
+        ValueError: cannot convert to bool numpy array in presence of missing values
+
+        Specify a valid `na_value` instead
+
+        >>> a.to_numpy(dtype="bool", na_value=False)
+        array([ True, False, False])
+        """
+        hasna = self._hasna
+        dtype, na_value = to_numpy_dtype_inference(self, dtype, na_value, hasna)
+        if dtype is None:
+            dtype = object
+
+        if hasna:
+            if (
+                dtype != object
+                and not is_string_dtype(dtype)
+                and na_value is libmissing.NA
+            ):
+                raise ValueError(
+                    f"cannot convert to '{dtype}'-dtype NumPy array "
+                    "with missing values. Specify an appropriate 'na_value' "
+                    "for this dtype."
+                )
+            # don't pass copy to astype -> always need a copy since we are mutating
+            with warnings.catch_warnings():
+                warnings.filterwarnings("ignore", category=RuntimeWarning)
+                data = self._data.astype(dtype)
+            data[self._mask] = na_value
+        else:
+            with warnings.catch_warnings():
+                warnings.filterwarnings("ignore", category=RuntimeWarning)
+                data = self._data.astype(dtype, copy=copy)
+        return data
+
+    @doc(ExtensionArray.tolist)
+    def tolist(self):
+        if self.ndim > 1:
+            return [x.tolist() for x in self]
+        dtype = None if self._hasna else self._data.dtype
+        return self.to_numpy(dtype=dtype, na_value=libmissing.NA).tolist()
+
+    @overload
+    def astype(self, dtype: npt.DTypeLike, copy: bool = ...) -> np.ndarray:
+        ...
+
+    @overload
+    def astype(self, dtype: ExtensionDtype, copy: bool = ...) -> ExtensionArray:
+        ...
+
+    @overload
+    def astype(self, dtype: AstypeArg, copy: bool = ...) -> ArrayLike:
+        ...
+
+    def astype(self, dtype: AstypeArg, copy: bool = True) -> ArrayLike:
+        dtype = pandas_dtype(dtype)
+
+        if dtype == self.dtype:
+            if copy:
+                return self.copy()
+            return self
+
+        # if we are astyping to another nullable masked dtype, we can fastpath
+        if isinstance(dtype, BaseMaskedDtype):
+            # TODO deal with NaNs for FloatingArray case
+            with warnings.catch_warnings():
+                warnings.filterwarnings("ignore", category=RuntimeWarning)
+                # TODO: Is rounding what we want long term?
+                data = self._data.astype(dtype.numpy_dtype, copy=copy)
+            # mask is copied depending on whether the data was copied, and
+            # not directly depending on the `copy` keyword
+            mask = self._mask if data is self._data else self._mask.copy()
+            cls = dtype.construct_array_type()
+            return cls(data, mask, copy=False)
+
+        if isinstance(dtype, ExtensionDtype):
+            eacls = dtype.construct_array_type()
+            return eacls._from_sequence(self, dtype=dtype, copy=copy)
+
+        na_value: float | np.datetime64 | lib.NoDefault
+
+        # coerce
+        if dtype.kind == "f":
+            # In astype, we consider dtype=float to also mean na_value=np.nan
+            na_value = np.nan
+        elif dtype.kind == "M":
+            na_value = np.datetime64("NaT")
+        else:
+            na_value = lib.no_default
+
+        # to_numpy will also raise, but we get somewhat nicer exception messages here
+        if dtype.kind in "iu" and self._hasna:
+            raise ValueError("cannot convert NA to integer")
+        if dtype.kind == "b" and self._hasna:
+            # careful: astype_nansafe converts np.nan to True
+            raise ValueError("cannot convert float NaN to bool")
+
+        data = self.to_numpy(dtype=dtype, na_value=na_value, copy=copy)
+        return data
+
+    __array_priority__ = 1000  # higher than ndarray so ops dispatch to us
+
+    def __array__(
+        self, dtype: NpDtype | None = None, copy: bool | None = None
+    ) -> np.ndarray:
+        """
+        the array interface, return my values
+        We return an object array here to preserve our scalar values
+        """
+        return self.to_numpy(dtype=dtype)
+
+    _HANDLED_TYPES: tuple[type, ...]
+
+    def __array_ufunc__(self, ufunc: np.ufunc, method: str, *inputs, **kwargs):
+        # For MaskedArray inputs, we apply the ufunc to ._data
+        # and mask the result.
+
+        out = kwargs.get("out", ())
+
+        for x in inputs + out:
+            if not isinstance(x, self._HANDLED_TYPES + (BaseMaskedArray,)):
+                return NotImplemented
+
+        # for binary ops, use our custom dunder methods
+        result = arraylike.maybe_dispatch_ufunc_to_dunder_op(
+            self, ufunc, method, *inputs, **kwargs
+        )
+        if result is not NotImplemented:
+            return result
+
+        if "out" in kwargs:
+            # e.g. test_ufunc_with_out
+            return arraylike.dispatch_ufunc_with_out(
+                self, ufunc, method, *inputs, **kwargs
+            )
+
+        if method == "reduce":
+            result = arraylike.dispatch_reduction_ufunc(
+                self, ufunc, method, *inputs, **kwargs
+            )
+            if result is not NotImplemented:
+                return result
+
+        mask = np.zeros(len(self), dtype=bool)
+        inputs2 = []
+        for x in inputs:
+            if isinstance(x, BaseMaskedArray):
+                mask |= x._mask
+                inputs2.append(x._data)
+            else:
+                inputs2.append(x)
+
+        def reconstruct(x: np.ndarray):
+            # we don't worry about scalar `x` here, since we
+            # raise for reduce up above.
+            from pandas.core.arrays import (
+                BooleanArray,
+                FloatingArray,
+                IntegerArray,
+            )
+
+            if x.dtype.kind == "b":
+                m = mask.copy()
+                return BooleanArray(x, m)
+            elif x.dtype.kind in "iu":
+                m = mask.copy()
+                return IntegerArray(x, m)
+            elif x.dtype.kind == "f":
+                m = mask.copy()
+                if x.dtype == np.float16:
+                    # reached in e.g. np.sqrt on BooleanArray
+                    # we don't support float16
+                    x = x.astype(np.float32)
+                return FloatingArray(x, m)
+            else:
+                x[mask] = np.nan
+            return x
+
+        result = getattr(ufunc, method)(*inputs2, **kwargs)
+        if ufunc.nout > 1:
+            # e.g. np.divmod
+            return tuple(reconstruct(x) for x in result)
+        elif method == "reduce":
+            # e.g. np.add.reduce; test_ufunc_reduce_raises
+            if self._mask.any():
+                return self._na_value
+            return result
+        else:
+            return reconstruct(result)
+
+    def __arrow_array__(self, type=None):
+        """
+        Convert myself into a pyarrow Array.
+        """
+        import pyarrow as pa
+
+        return pa.array(self._data, mask=self._mask, type=type)
+
+    @property
+    def _hasna(self) -> bool:
+        # Note: this is expensive right now! The hope is that we can
+        # make this faster by having an optional mask, but not have to change
+        # source code using it..
+
+        # error: Incompatible return value type (got "bool_", expected "bool")
+        return self._mask.any()  # type: ignore[return-value]
+
+    def _propagate_mask(
+        self, mask: npt.NDArray[np.bool_] | None, other
+    ) -> npt.NDArray[np.bool_]:
+        if mask is None:
+            mask = self._mask.copy()  # TODO: need test for BooleanArray needing a copy
+            if other is libmissing.NA:
+                # GH#45421 don't alter inplace
+                mask = mask | True
+            elif is_list_like(other) and len(other) == len(mask):
+                mask = mask | isna(other)
+        else:
+            mask = self._mask | mask
+        # Incompatible return value type (got "Optional[ndarray[Any, dtype[bool_]]]",
+        # expected "ndarray[Any, dtype[bool_]]")
+        return mask  # type: ignore[return-value]
+
+    def _arith_method(self, other, op):
+        op_name = op.__name__
+        omask = None
+
+        if (
+            not hasattr(other, "dtype")
+            and is_list_like(other)
+            and len(other) == len(self)
+        ):
+            # Try inferring masked dtype instead of casting to object
+            other = pd_array(other)
+            other = extract_array(other, extract_numpy=True)
+
+        if isinstance(other, BaseMaskedArray):
+            other, omask = other._data, other._mask
+
+        elif is_list_like(other):
+            if not isinstance(other, ExtensionArray):
+                other = np.asarray(other)
+            if other.ndim > 1:
+                raise NotImplementedError("can only perform ops with 1-d structures")
+
+        # We wrap the non-masked arithmetic logic used for numpy dtypes
+        #  in Series/Index arithmetic ops.
+        other = ops.maybe_prepare_scalar_for_op(other, (len(self),))
+        pd_op = ops.get_array_op(op)
+        other = ensure_wrapped_if_datetimelike(other)
+
+        if op_name in {"pow", "rpow"} and isinstance(other, np.bool_):
+            # Avoid DeprecationWarning: In future, it will be an error
+            #  for 'np.bool_' scalars to be interpreted as an index
+            #  e.g. test_array_scalar_like_equivalence
+            other = bool(other)
+
+        mask = self._propagate_mask(omask, other)
+
+        if other is libmissing.NA:
+            result = np.ones_like(self._data)
+            if self.dtype.kind == "b":
+                if op_name in {
+                    "floordiv",
+                    "rfloordiv",
+                    "pow",
+                    "rpow",
+                    "truediv",
+                    "rtruediv",
+                }:
+                    # GH#41165 Try to match non-masked Series behavior
+                    #  This is still imperfect GH#46043
+                    raise NotImplementedError(
+                        f"operator '{op_name}' not implemented for bool dtypes"
+                    )
+                if op_name in {"mod", "rmod"}:
+                    dtype = "int8"
+                else:
+                    dtype = "bool"
+                result = result.astype(dtype)
+            elif "truediv" in op_name and self.dtype.kind != "f":
+                # The actual data here doesn't matter since the mask
+                #  will be all-True, but since this is division, we want
+                #  to end up with floating dtype.
+                result = result.astype(np.float64)
+        else:
+            # Make sure we do this before the "pow" mask checks
+            #  to get an expected exception message on shape mismatch.
+            if self.dtype.kind in "iu" and op_name in ["floordiv", "mod"]:
+                # TODO(GH#30188) ATM we don't match the behavior of non-masked
+                #  types with respect to floordiv-by-zero
+                pd_op = op
+
+            with np.errstate(all="ignore"):
+                result = pd_op(self._data, other)
+
+        if op_name == "pow":
+            # 1 ** x is 1.
+            mask = np.where((self._data == 1) & ~self._mask, False, mask)
+            # x ** 0 is 1.
+            if omask is not None:
+                mask = np.where((other == 0) & ~omask, False, mask)
+            elif other is not libmissing.NA:
+                mask = np.where(other == 0, False, mask)
+
+        elif op_name == "rpow":
+            # 1 ** x is 1.
+            if omask is not None:
+                mask = np.where((other == 1) & ~omask, False, mask)
+            elif other is not libmissing.NA:
+                mask = np.where(other == 1, False, mask)
+            # x ** 0 is 1.
+            mask = np.where((self._data == 0) & ~self._mask, False, mask)
+
+        return self._maybe_mask_result(result, mask)
+
+    _logical_method = _arith_method
+
+    def _cmp_method(self, other, op) -> BooleanArray:
+        from pandas.core.arrays import BooleanArray
+
+        mask = None
+
+        if isinstance(other, BaseMaskedArray):
+            other, mask = other._data, other._mask
+
+        elif is_list_like(other):
+            other = np.asarray(other)
+            if other.ndim > 1:
+                raise NotImplementedError("can only perform ops with 1-d structures")
+            if len(self) != len(other):
+                raise ValueError("Lengths must match to compare")
+
+        if other is libmissing.NA:
+            # numpy does not handle pd.NA well as "other" scalar (it returns
+            # a scalar False instead of an array)
+            # This may be fixed by NA.__array_ufunc__. Revisit this check
+            # once that's implemented.
+            result = np.zeros(self._data.shape, dtype="bool")
+            mask = np.ones(self._data.shape, dtype="bool")
+        else:
+            with warnings.catch_warnings():
+                # numpy may show a FutureWarning or DeprecationWarning:
+                #     elementwise comparison failed; returning scalar instead,
+                #     but in the future will perform elementwise comparison
+                # before returning NotImplemented. We fall back to the correct
+                # behavior today, so that should be fine to ignore.
+                warnings.filterwarnings("ignore", "elementwise", FutureWarning)
+                warnings.filterwarnings("ignore", "elementwise", DeprecationWarning)
+                method = getattr(self._data, f"__{op.__name__}__")
+                result = method(other)
+
+                if result is NotImplemented:
+                    result = invalid_comparison(self._data, other, op)
+
+        mask = self._propagate_mask(mask, other)
+        return BooleanArray(result, mask, copy=False)
+
+    def _maybe_mask_result(
+        self, result: np.ndarray | tuple[np.ndarray, np.ndarray], mask: np.ndarray
+    ):
+        """
+        Parameters
+        ----------
+        result : array-like or tuple[array-like]
+        mask : array-like bool
+        """
+        if isinstance(result, tuple):
+            # i.e. divmod
+            div, mod = result
+            return (
+                self._maybe_mask_result(div, mask),
+                self._maybe_mask_result(mod, mask),
+            )
+
+        if result.dtype.kind == "f":
+            from pandas.core.arrays import FloatingArray
+
+            return FloatingArray(result, mask, copy=False)
+
+        elif result.dtype.kind == "b":
+            from pandas.core.arrays import BooleanArray
+
+            return BooleanArray(result, mask, copy=False)
+
+        elif lib.is_np_dtype(result.dtype, "m") and is_supported_dtype(result.dtype):
+            # e.g. test_numeric_arr_mul_tdscalar_numexpr_path
+            from pandas.core.arrays import TimedeltaArray
+
+            result[mask] = result.dtype.type("NaT")
+
+            if not isinstance(result, TimedeltaArray):
+                return TimedeltaArray._simple_new(result, dtype=result.dtype)
+
+            return result
+
+        elif result.dtype.kind in "iu":
+            from pandas.core.arrays import IntegerArray
+
+            return IntegerArray(result, mask, copy=False)
+
+        else:
+            result[mask] = np.nan
+            return result
+
+    def isna(self) -> np.ndarray:
+        return self._mask.copy()
+
+    @property
+    def _na_value(self):
+        return self.dtype.na_value
+
+    @property
+    def nbytes(self) -> int:
+        return self._data.nbytes + self._mask.nbytes
+
+    @classmethod
+    def _concat_same_type(
+        cls,
+        to_concat: Sequence[Self],
+        axis: AxisInt = 0,
+    ) -> Self:
+        data = np.concatenate([x._data for x in to_concat], axis=axis)
+        mask = np.concatenate([x._mask for x in to_concat], axis=axis)
+        return cls(data, mask)
+
+    def _hash_pandas_object(
+        self, *, encoding: str, hash_key: str, categorize: bool
+    ) -> npt.NDArray[np.uint64]:
+        hashed_array = hash_array(
+            self._data, encoding=encoding, hash_key=hash_key, categorize=categorize
+        )
+        hashed_array[self.isna()] = hash(self.dtype.na_value)
+        return hashed_array
+
+    def take(
+        self,
+        indexer,
+        *,
+        allow_fill: bool = False,
+        fill_value: Scalar | None = None,
+        axis: AxisInt = 0,
+    ) -> Self:
+        # we always fill with 1 internally
+        # to avoid upcasting
+        data_fill_value = self._internal_fill_value if isna(fill_value) else fill_value
+        result = take(
+            self._data,
+            indexer,
+            fill_value=data_fill_value,
+            allow_fill=allow_fill,
+            axis=axis,
+        )
+
+        mask = take(
+            self._mask, indexer, fill_value=True, allow_fill=allow_fill, axis=axis
+        )
+
+        # if we are filling
+        # we only fill where the indexer is null
+        # not existing missing values
+        # TODO(jreback) what if we have a non-na float as a fill value?
+        if allow_fill and notna(fill_value):
+            fill_mask = np.asarray(indexer) == -1
+            result[fill_mask] = fill_value
+            mask = mask ^ fill_mask
+
+        return self._simple_new(result, mask)
+
+    # error: Return type "BooleanArray" of "isin" incompatible with return type
+    # "ndarray" in supertype "ExtensionArray"
+    def isin(self, values: ArrayLike) -> BooleanArray:  # type: ignore[override]
+        from pandas.core.arrays import BooleanArray
+
+        # algorithms.isin will eventually convert values to an ndarray, so no extra
+        # cost to doing it here first
+        values_arr = np.asarray(values)
+        result = isin(self._data, values_arr)
+
+        if self._hasna:
+            values_have_NA = values_arr.dtype == object and any(
+                val is self.dtype.na_value for val in values_arr
+            )
+
+            # For now, NA does not propagate so set result according to presence of NA,
+            # see https://github.com/pandas-dev/pandas/pull/38379 for some discussion
+            result[self._mask] = values_have_NA
+
+        mask = np.zeros(self._data.shape, dtype=bool)
+        return BooleanArray(result, mask, copy=False)
+
+    def copy(self) -> Self:
+        data = self._data.copy()
+        mask = self._mask.copy()
+        return self._simple_new(data, mask)
+
+    @doc(ExtensionArray.duplicated)
+    def duplicated(
+        self, keep: Literal["first", "last", False] = "first"
+    ) -> npt.NDArray[np.bool_]:
+        values = self._data
+        mask = self._mask
+        return algos.duplicated(values, keep=keep, mask=mask)
+
+    def unique(self) -> Self:
+        """
+        Compute the BaseMaskedArray of unique values.
+
+        Returns
+        -------
+        uniques : BaseMaskedArray
+        """
+        uniques, mask = algos.unique_with_mask(self._data, self._mask)
+        return self._simple_new(uniques, mask)
+
+    @doc(ExtensionArray.searchsorted)
+    def searchsorted(
+        self,
+        value: NumpyValueArrayLike | ExtensionArray,
+        side: Literal["left", "right"] = "left",
+        sorter: NumpySorter | None = None,
+    ) -> npt.NDArray[np.intp] | np.intp:
+        if self._hasna:
+            raise ValueError(
+                "searchsorted requires array to be sorted, which is impossible "
+                "with NAs present."
+            )
+        if isinstance(value, ExtensionArray):
+            value = value.astype(object)
+        # Base class searchsorted would cast to object, which is *much* slower.
+        return self._data.searchsorted(value, side=side, sorter=sorter)
+
+    @doc(ExtensionArray.factorize)
+    def factorize(
+        self,
+        use_na_sentinel: bool = True,
+    ) -> tuple[np.ndarray, ExtensionArray]:
+        arr = self._data
+        mask = self._mask
+
+        # Use a sentinel for na; recode and add NA to uniques if necessary below
+        codes, uniques = factorize_array(arr, use_na_sentinel=True, mask=mask)
+
+        # check that factorize_array correctly preserves dtype.
+        assert uniques.dtype == self.dtype.numpy_dtype, (uniques.dtype, self.dtype)
+
+        has_na = mask.any()
+        if use_na_sentinel or not has_na:
+            size = len(uniques)
+        else:
+            # Make room for an NA value
+            size = len(uniques) + 1
+        uniques_mask = np.zeros(size, dtype=bool)
+        if not use_na_sentinel and has_na:
+            na_index = mask.argmax()
+            # Insert na with the proper code
+            if na_index == 0:
+                na_code = np.intp(0)
+            else:
+                na_code = codes[:na_index].max() + 1
+            codes[codes >= na_code] += 1
+            codes[codes == -1] = na_code
+            # dummy value for uniques; not used since uniques_mask will be True
+            uniques = np.insert(uniques, na_code, 0)
+            uniques_mask[na_code] = True
+        uniques_ea = self._simple_new(uniques, uniques_mask)
+
+        return codes, uniques_ea
+
+    @doc(ExtensionArray._values_for_argsort)
+    def _values_for_argsort(self) -> np.ndarray:
+        return self._data
+
+    def value_counts(self, dropna: bool = True) -> Series:
+        """
+        Returns a Series containing counts of each unique value.
+
+        Parameters
+        ----------
+        dropna : bool, default True
+            Don't include counts of missing values.
+
+        Returns
+        -------
+        counts : Series
+
+        See Also
+        --------
+        Series.value_counts
+        """
+        from pandas import (
+            Index,
+            Series,
+        )
+        from pandas.arrays import IntegerArray
+
+        keys, value_counts, na_counter = algos.value_counts_arraylike(
+            self._data, dropna=dropna, mask=self._mask
+        )
+        mask_index = np.zeros((len(value_counts),), dtype=np.bool_)
+        mask = mask_index.copy()
+
+        if na_counter > 0:
+            mask_index[-1] = True
+
+        arr = IntegerArray(value_counts, mask)
+        index = Index(
+            self.dtype.construct_array_type()(
+                keys, mask_index  # type: ignore[arg-type]
+            )
+        )
+        return Series(arr, index=index, name="count", copy=False)
+
+    def _mode(self, dropna: bool = True) -> Self:
+        if dropna:
+            result = mode(self._data, dropna=dropna, mask=self._mask)
+            res_mask = np.zeros(result.shape, dtype=np.bool_)
+        else:
+            result, res_mask = mode(self._data, dropna=dropna, mask=self._mask)
+        result = type(self)(result, res_mask)  # type: ignore[arg-type]
+        return result[result.argsort()]
+
+    @doc(ExtensionArray.equals)
+    def equals(self, other) -> bool:
+        if type(self) != type(other):
+            return False
+        if other.dtype != self.dtype:
+            return False
+
+        # GH#44382 if e.g. self[1] is np.nan and other[1] is pd.NA, we are NOT
+        #  equal.
+        if not np.array_equal(self._mask, other._mask):
+            return False
+
+        left = self._data[~self._mask]
+        right = other._data[~other._mask]
+        return array_equivalent(left, right, strict_nan=True, dtype_equal=True)
+
+    def _quantile(
+        self, qs: npt.NDArray[np.float64], interpolation: str
+    ) -> BaseMaskedArray:
+        """
+        Dispatch to quantile_with_mask, needed because we do not have
+        _from_factorized.
+
+        Notes
+        -----
+        We assume that all impacted cases are 1D-only.
+        """
+        res = quantile_with_mask(
+            self._data,
+            mask=self._mask,
+            # TODO(GH#40932): na_value_for_dtype(self.dtype.numpy_dtype)
+            #  instead of np.nan
+            fill_value=np.nan,
+            qs=qs,
+            interpolation=interpolation,
+        )
+
+        if self._hasna:
+            # Our result mask is all-False unless we are all-NA, in which
+            #  case it is all-True.
+            if self.ndim == 2:
+                # I think this should be out_mask=self.isna().all(axis=1)
+                #  but am holding off until we have tests
+                raise NotImplementedError
+            if self.isna().all():
+                out_mask = np.ones(res.shape, dtype=bool)
+
+                if is_integer_dtype(self.dtype):
+                    # We try to maintain int dtype if possible for not all-na case
+                    # as well
+                    res = np.zeros(res.shape, dtype=self.dtype.numpy_dtype)
+            else:
+                out_mask = np.zeros(res.shape, dtype=bool)
+        else:
+            out_mask = np.zeros(res.shape, dtype=bool)
+        return self._maybe_mask_result(res, mask=out_mask)
+
+    # ------------------------------------------------------------------
+    # Reductions
+
+    def _reduce(
+        self, name: str, *, skipna: bool = True, keepdims: bool = False, **kwargs
+    ):
+        if name in {"any", "all", "min", "max", "sum", "prod", "mean", "var", "std"}:
+            result = getattr(self, name)(skipna=skipna, **kwargs)
+        else:
+            # median, skew, kurt, sem
+            data = self._data
+            mask = self._mask
+            op = getattr(nanops, f"nan{name}")
+            axis = kwargs.pop("axis", None)
+            result = op(data, axis=axis, skipna=skipna, mask=mask, **kwargs)
+
+        if keepdims:
+            if isna(result):
+                return self._wrap_na_result(name=name, axis=0, mask_size=(1,))
+            else:
+                result = result.reshape(1)
+                mask = np.zeros(1, dtype=bool)
+                return self._maybe_mask_result(result, mask)
+
+        if isna(result):
+            return libmissing.NA
+        else:
+            return result
+
+    def _wrap_reduction_result(self, name: str, result, *, skipna, axis):
+        if isinstance(result, np.ndarray):
+            if skipna:
+                # we only retain mask for all-NA rows/columns
+                mask = self._mask.all(axis=axis)
+            else:
+                mask = self._mask.any(axis=axis)
+
+            return self._maybe_mask_result(result, mask)
+        return result
+
+    def _wrap_na_result(self, *, name, axis, mask_size):
+        mask = np.ones(mask_size, dtype=bool)
+
+        float_dtyp = "float32" if self.dtype == "Float32" else "float64"
+        if name in ["mean", "median", "var", "std", "skew", "kurt"]:
+            np_dtype = float_dtyp
+        elif name in ["min", "max"] or self.dtype.itemsize == 8:
+            np_dtype = self.dtype.numpy_dtype.name
+        else:
+            is_windows_or_32bit = is_platform_windows() or not IS64
+            int_dtyp = "int32" if is_windows_or_32bit else "int64"
+            uint_dtyp = "uint32" if is_windows_or_32bit else "uint64"
+            np_dtype = {"b": int_dtyp, "i": int_dtyp, "u": uint_dtyp, "f": float_dtyp}[
+                self.dtype.kind
+            ]
+
+        value = np.array([1], dtype=np_dtype)
+        return self._maybe_mask_result(value, mask=mask)
+
+    def _wrap_min_count_reduction_result(
+        self, name: str, result, *, skipna, min_count, axis
+    ):
+        if min_count == 0 and isinstance(result, np.ndarray):
+            return self._maybe_mask_result(result, np.zeros(result.shape, dtype=bool))
+        return self._wrap_reduction_result(name, result, skipna=skipna, axis=axis)
+
+    def sum(
+        self,
+        *,
+        skipna: bool = True,
+        min_count: int = 0,
+        axis: AxisInt | None = 0,
+        **kwargs,
+    ):
+        nv.validate_sum((), kwargs)
+
+        result = masked_reductions.sum(
+            self._data,
+            self._mask,
+            skipna=skipna,
+            min_count=min_count,
+            axis=axis,
+        )
+        return self._wrap_min_count_reduction_result(
+            "sum", result, skipna=skipna, min_count=min_count, axis=axis
+        )
+
+    def prod(
+        self,
+        *,
+        skipna: bool = True,
+        min_count: int = 0,
+        axis: AxisInt | None = 0,
+        **kwargs,
+    ):
+        nv.validate_prod((), kwargs)
+
+        result = masked_reductions.prod(
+            self._data,
+            self._mask,
+            skipna=skipna,
+            min_count=min_count,
+            axis=axis,
+        )
+        return self._wrap_min_count_reduction_result(
+            "prod", result, skipna=skipna, min_count=min_count, axis=axis
+        )
+
+    def mean(self, *, skipna: bool = True, axis: AxisInt | None = 0, **kwargs):
+        nv.validate_mean((), kwargs)
+        result = masked_reductions.mean(
+            self._data,
+            self._mask,
+            skipna=skipna,
+            axis=axis,
+        )
+        return self._wrap_reduction_result("mean", result, skipna=skipna, axis=axis)
+
+    def var(
+        self, *, skipna: bool = True, axis: AxisInt | None = 0, ddof: int = 1, **kwargs
+    ):
+        nv.validate_stat_ddof_func((), kwargs, fname="var")
+        result = masked_reductions.var(
+            self._data,
+            self._mask,
+            skipna=skipna,
+            axis=axis,
+            ddof=ddof,
+        )
+        return self._wrap_reduction_result("var", result, skipna=skipna, axis=axis)
+
+    def std(
+        self, *, skipna: bool = True, axis: AxisInt | None = 0, ddof: int = 1, **kwargs
+    ):
+        nv.validate_stat_ddof_func((), kwargs, fname="std")
+        result = masked_reductions.std(
+            self._data,
+            self._mask,
+            skipna=skipna,
+            axis=axis,
+            ddof=ddof,
+        )
+        return self._wrap_reduction_result("std", result, skipna=skipna, axis=axis)
+
+    def min(self, *, skipna: bool = True, axis: AxisInt | None = 0, **kwargs):
+        nv.validate_min((), kwargs)
+        result = masked_reductions.min(
+            self._data,
+            self._mask,
+            skipna=skipna,
+            axis=axis,
+        )
+        return self._wrap_reduction_result("min", result, skipna=skipna, axis=axis)
+
+    def max(self, *, skipna: bool = True, axis: AxisInt | None = 0, **kwargs):
+        nv.validate_max((), kwargs)
+        result = masked_reductions.max(
+            self._data,
+            self._mask,
+            skipna=skipna,
+            axis=axis,
+        )
+        return self._wrap_reduction_result("max", result, skipna=skipna, axis=axis)
+
+    def map(self, mapper, na_action=None):
+        return map_array(self.to_numpy(), mapper, na_action=na_action)
+
+    def any(self, *, skipna: bool = True, axis: AxisInt | None = 0, **kwargs):
+        """
+        Return whether any element is truthy.
+
+        Returns False unless there is at least one element that is truthy.
+        By default, NAs are skipped. If ``skipna=False`` is specified and
+        missing values are present, similar :ref:`Kleene logic <boolean.kleene>`
+        is used as for logical operations.
+
+        .. versionchanged:: 1.4.0
+
+        Parameters
+        ----------
+        skipna : bool, default True
+            Exclude NA values. If the entire array is NA and `skipna` is
+            True, then the result will be False, as for an empty array.
+            If `skipna` is False, the result will still be True if there is
+            at least one element that is truthy, otherwise NA will be returned
+            if there are NA's present.
+        axis : int, optional, default 0
+        **kwargs : any, default None
+            Additional keywords have no effect but might be accepted for
+            compatibility with NumPy.
+
+        Returns
+        -------
+        bool or :attr:`pandas.NA`
+
+        See Also
+        --------
+        numpy.any : Numpy version of this method.
+        BaseMaskedArray.all : Return whether all elements are truthy.
+
+        Examples
+        --------
+        The result indicates whether any element is truthy (and by default
+        skips NAs):
+
+        >>> pd.array([True, False, True]).any()
+        True
+        >>> pd.array([True, False, pd.NA]).any()
+        True
+        >>> pd.array([False, False, pd.NA]).any()
+        False
+        >>> pd.array([], dtype="boolean").any()
+        False
+        >>> pd.array([pd.NA], dtype="boolean").any()
+        False
+        >>> pd.array([pd.NA], dtype="Float64").any()
+        False
+
+        With ``skipna=False``, the result can be NA if this is logically
+        required (whether ``pd.NA`` is True or False influences the result):
+
+        >>> pd.array([True, False, pd.NA]).any(skipna=False)
+        True
+        >>> pd.array([1, 0, pd.NA]).any(skipna=False)
+        True
+        >>> pd.array([False, False, pd.NA]).any(skipna=False)
+        <NA>
+        >>> pd.array([0, 0, pd.NA]).any(skipna=False)
+        <NA>
+        """
+        nv.validate_any((), kwargs)
+
+        values = self._data.copy()
+        # error: Argument 3 to "putmask" has incompatible type "object";
+        # expected "Union[_SupportsArray[dtype[Any]],
+        # _NestedSequence[_SupportsArray[dtype[Any]]],
+        # bool, int, float, complex, str, bytes,
+        # _NestedSequence[Union[bool, int, float, complex, str, bytes]]]"
+        np.putmask(values, self._mask, self._falsey_value)  # type: ignore[arg-type]
+        result = values.any()
+        if skipna:
+            return result
+        else:
+            if result or len(self) == 0 or not self._mask.any():
+                return result
+            else:
+                return self.dtype.na_value
+
+    def all(self, *, skipna: bool = True, axis: AxisInt | None = 0, **kwargs):
+        """
+        Return whether all elements are truthy.
+
+        Returns True unless there is at least one element that is falsey.
+        By default, NAs are skipped. If ``skipna=False`` is specified and
+        missing values are present, similar :ref:`Kleene logic <boolean.kleene>`
+        is used as for logical operations.
+
+        .. versionchanged:: 1.4.0
+
+        Parameters
+        ----------
+        skipna : bool, default True
+            Exclude NA values. If the entire array is NA and `skipna` is
+            True, then the result will be True, as for an empty array.
+            If `skipna` is False, the result will still be False if there is
+            at least one element that is falsey, otherwise NA will be returned
+            if there are NA's present.
+        axis : int, optional, default 0
+        **kwargs : any, default None
+            Additional keywords have no effect but might be accepted for
+            compatibility with NumPy.
+
+        Returns
+        -------
+        bool or :attr:`pandas.NA`
+
+        See Also
+        --------
+        numpy.all : Numpy version of this method.
+        BooleanArray.any : Return whether any element is truthy.
+
+        Examples
+        --------
+        The result indicates whether all elements are truthy (and by default
+        skips NAs):
+
+        >>> pd.array([True, True, pd.NA]).all()
+        True
+        >>> pd.array([1, 1, pd.NA]).all()
+        True
+        >>> pd.array([True, False, pd.NA]).all()
+        False
+        >>> pd.array([], dtype="boolean").all()
+        True
+        >>> pd.array([pd.NA], dtype="boolean").all()
+        True
+        >>> pd.array([pd.NA], dtype="Float64").all()
+        True
+
+        With ``skipna=False``, the result can be NA if this is logically
+        required (whether ``pd.NA`` is True or False influences the result):
+
+        >>> pd.array([True, True, pd.NA]).all(skipna=False)
+        <NA>
+        >>> pd.array([1, 1, pd.NA]).all(skipna=False)
+        <NA>
+        >>> pd.array([True, False, pd.NA]).all(skipna=False)
+        False
+        >>> pd.array([1, 0, pd.NA]).all(skipna=False)
+        False
+        """
+        nv.validate_all((), kwargs)
+
+        values = self._data.copy()
+        # error: Argument 3 to "putmask" has incompatible type "object";
+        # expected "Union[_SupportsArray[dtype[Any]],
+        # _NestedSequence[_SupportsArray[dtype[Any]]],
+        # bool, int, float, complex, str, bytes,
+        # _NestedSequence[Union[bool, int, float, complex, str, bytes]]]"
+        np.putmask(values, self._mask, self._truthy_value)  # type: ignore[arg-type]
+        result = values.all(axis=axis)
+
+        if skipna:
+            return result
+        else:
+            if not result or len(self) == 0 or not self._mask.any():
+                return result
+            else:
+                return self.dtype.na_value
+
+    def interpolate(
+        self,
+        *,
+        method: InterpolateOptions,
+        axis: int,
+        index,
+        limit,
+        limit_direction,
+        limit_area,
+        copy: bool,
+        **kwargs,
+    ) -> FloatingArray:
+        """
+        See NDFrame.interpolate.__doc__.
+        """
+        # NB: we return type(self) even if copy=False
+        if self.dtype.kind == "f":
+            if copy:
+                data = self._data.copy()
+                mask = self._mask.copy()
+            else:
+                data = self._data
+                mask = self._mask
+        elif self.dtype.kind in "iu":
+            copy = True
+            data = self._data.astype("f8")
+            mask = self._mask.copy()
+        else:
+            raise NotImplementedError(
+                f"interpolate is not implemented for dtype={self.dtype}"
+            )
+
+        missing.interpolate_2d_inplace(
+            data,
+            method=method,
+            axis=0,
+            index=index,
+            limit=limit,
+            limit_direction=limit_direction,
+            limit_area=limit_area,
+            mask=mask,
+            **kwargs,
+        )
+        if not copy:
+            return self  # type: ignore[return-value]
+        if self.dtype.kind == "f":
+            return type(self)._simple_new(data, mask)  # type: ignore[return-value]
+        else:
+            from pandas.core.arrays import FloatingArray
+
+            return FloatingArray._simple_new(data, mask)
+
+    def _accumulate(
+        self, name: str, *, skipna: bool = True, **kwargs
+    ) -> BaseMaskedArray:
+        data = self._data
+        mask = self._mask
+
+        op = getattr(masked_accumulations, name)
+        data, mask = op(data, mask, skipna=skipna, **kwargs)
+
+        return self._simple_new(data, mask)
+
+    # ------------------------------------------------------------------
+    # GroupBy Methods
+
+    def _groupby_op(
+        self,
+        *,
+        how: str,
+        has_dropped_na: bool,
+        min_count: int,
+        ngroups: int,
+        ids: npt.NDArray[np.intp],
+        **kwargs,
+    ):
+        from pandas.core.groupby.ops import WrappedCythonOp
+
+        kind = WrappedCythonOp.get_kind_from_how(how)
+        op = WrappedCythonOp(how=how, kind=kind, has_dropped_na=has_dropped_na)
+
+        # libgroupby functions are responsible for NOT altering mask
+        mask = self._mask
+        if op.kind != "aggregate":
+            result_mask = mask.copy()
+        else:
+            result_mask = np.zeros(ngroups, dtype=bool)
+
+        if how == "rank" and kwargs.get("na_option") in ["top", "bottom"]:
+            result_mask[:] = False
+
+        res_values = op._cython_op_ndim_compat(
+            self._data,
+            min_count=min_count,
+            ngroups=ngroups,
+            comp_ids=ids,
+            mask=mask,
+            result_mask=result_mask,
+            **kwargs,
+        )
+
+        if op.how == "ohlc":
+            arity = op._cython_arity.get(op.how, 1)
+            result_mask = np.tile(result_mask, (arity, 1)).T
+
+        if op.how in ["idxmin", "idxmax"]:
+            # Result values are indexes to take, keep as ndarray
+            return res_values
+        else:
+            # res_values should already have the correct dtype, we just need to
+            #  wrap in a MaskedArray
+            return self._maybe_mask_result(res_values, result_mask)
+
+
+def transpose_homogeneous_masked_arrays(
+    masked_arrays: Sequence[BaseMaskedArray],
+) -> list[BaseMaskedArray]:
+    """Transpose masked arrays in a list, but faster.
+
+    Input should be a list of 1-dim masked arrays of equal length and all have the
+    same dtype. The caller is responsible for ensuring validity of input data.
+    """
+    masked_arrays = list(masked_arrays)
+    dtype = masked_arrays[0].dtype
+
+    values = [arr._data.reshape(1, -1) for arr in masked_arrays]
+    transposed_values = np.concatenate(
+        values,
+        axis=0,
+        out=np.empty(
+            (len(masked_arrays), len(masked_arrays[0])),
+            order="F",
+            dtype=dtype.numpy_dtype,
+        ),
+    )
+
+    masks = [arr._mask.reshape(1, -1) for arr in masked_arrays]
+    transposed_masks = np.concatenate(
+        masks, axis=0, out=np.empty_like(transposed_values, dtype=bool)
+    )
+
+    arr_type = dtype.construct_array_type()
+    transposed_arrays: list[BaseMaskedArray] = []
+    for i in range(transposed_values.shape[1]):
+        transposed_arr = arr_type(transposed_values[:, i], mask=transposed_masks[:, i])
+        transposed_arrays.append(transposed_arr)
+
+    return transposed_arrays

llava_next/lib/python3.10/site-packages/pandas/core/arrays/numeric.py

@@ -0,0 +1,286 @@
+from __future__ import annotations
+
+import numbers
+from typing import (
+    TYPE_CHECKING,
+    Any,
+    Callable,
+)
+
+import numpy as np
+
+from pandas._libs import (
+    lib,
+    missing as libmissing,
+)
+from pandas.errors import AbstractMethodError
+from pandas.util._decorators import cache_readonly
+
+from pandas.core.dtypes.common import (
+    is_integer_dtype,
+    is_string_dtype,
+    pandas_dtype,
+)
+
+from pandas.core.arrays.masked import (
+    BaseMaskedArray,
+    BaseMaskedDtype,
+)
+
+if TYPE_CHECKING:
+    from collections.abc import Mapping
+
+    import pyarrow
+
+    from pandas._typing import (
+        Dtype,
+        DtypeObj,
+        Self,
+        npt,
+    )
+
+
+class NumericDtype(BaseMaskedDtype):
+    _default_np_dtype: np.dtype
+    _checker: Callable[[Any], bool]  # is_foo_dtype
+
+    def __repr__(self) -> str:
+        return f"{self.name}Dtype()"
+
+    @cache_readonly
+    def is_signed_integer(self) -> bool:
+        return self.kind == "i"
+
+    @cache_readonly
+    def is_unsigned_integer(self) -> bool:
+        return self.kind == "u"
+
+    @property
+    def _is_numeric(self) -> bool:
+        return True
+
+    def __from_arrow__(
+        self, array: pyarrow.Array | pyarrow.ChunkedArray
+    ) -> BaseMaskedArray:
+        """
+        Construct IntegerArray/FloatingArray from pyarrow Array/ChunkedArray.
+        """
+        import pyarrow
+
+        from pandas.core.arrays.arrow._arrow_utils import (
+            pyarrow_array_to_numpy_and_mask,
+        )
+
+        array_class = self.construct_array_type()
+
+        pyarrow_type = pyarrow.from_numpy_dtype(self.type)
+        if not array.type.equals(pyarrow_type) and not pyarrow.types.is_null(
+            array.type
+        ):
+            # test_from_arrow_type_error raise for string, but allow
+            #  through itemsize conversion GH#31896
+            rt_dtype = pandas_dtype(array.type.to_pandas_dtype())
+            if rt_dtype.kind not in "iuf":
+                # Could allow "c" or potentially disallow float<->int conversion,
+                #  but at the moment we specifically test that uint<->int works
+                raise TypeError(
+                    f"Expected array of {self} type, got {array.type} instead"
+                )
+
+            array = array.cast(pyarrow_type)
+
+        if isinstance(array, pyarrow.ChunkedArray):
+            # TODO this "if" can be removed when requiring pyarrow >= 10.0, which fixed
+            # combine_chunks for empty arrays https://github.com/apache/arrow/pull/13757
+            if array.num_chunks == 0:
+                array = pyarrow.array([], type=array.type)
+            else:
+                array = array.combine_chunks()
+
+        data, mask = pyarrow_array_to_numpy_and_mask(array, dtype=self.numpy_dtype)
+        return array_class(data.copy(), ~mask, copy=False)
+
+    @classmethod
+    def _get_dtype_mapping(cls) -> Mapping[np.dtype, NumericDtype]:
+        raise AbstractMethodError(cls)
+
+    @classmethod
+    def _standardize_dtype(cls, dtype: NumericDtype | str | np.dtype) -> NumericDtype:
+        """
+        Convert a string representation or a numpy dtype to NumericDtype.
+        """
+        if isinstance(dtype, str) and (dtype.startswith(("Int", "UInt", "Float"))):
+            # Avoid DeprecationWarning from NumPy about np.dtype("Int64")
+            # https://github.com/numpy/numpy/pull/7476
+            dtype = dtype.lower()
+
+        if not isinstance(dtype, NumericDtype):
+            mapping = cls._get_dtype_mapping()
+            try:
+                dtype = mapping[np.dtype(dtype)]
+            except KeyError as err:
+                raise ValueError(f"invalid dtype specified {dtype}") from err
+        return dtype
+
+    @classmethod
+    def _safe_cast(cls, values: np.ndarray, dtype: np.dtype, copy: bool) -> np.ndarray:
+        """
+        Safely cast the values to the given dtype.
+
+        "safe" in this context means the casting is lossless.
+        """
+        raise AbstractMethodError(cls)
+
+
+def _coerce_to_data_and_mask(
+    values, dtype, copy: bool, dtype_cls: type[NumericDtype], default_dtype: np.dtype
+):
+    checker = dtype_cls._checker
+
+    mask = None
+    inferred_type = None
+
+    if dtype is None and hasattr(values, "dtype"):
+        if checker(values.dtype):
+            dtype = values.dtype
+
+    if dtype is not None:
+        dtype = dtype_cls._standardize_dtype(dtype)
+
+    cls = dtype_cls.construct_array_type()
+    if isinstance(values, cls):
+        values, mask = values._data, values._mask
+        if dtype is not None:
+            values = values.astype(dtype.numpy_dtype, copy=False)
+
+        if copy:
+            values = values.copy()
+            mask = mask.copy()
+        return values, mask, dtype, inferred_type
+
+    original = values
+    if not copy:
+        values = np.asarray(values)
+    else:
+        values = np.array(values, copy=copy)
+    inferred_type = None
+    if values.dtype == object or is_string_dtype(values.dtype):
+        inferred_type = lib.infer_dtype(values, skipna=True)
+        if inferred_type == "boolean" and dtype is None:
+            name = dtype_cls.__name__.strip("_")
+            raise TypeError(f"{values.dtype} cannot be converted to {name}")
+
+    elif values.dtype.kind == "b" and checker(dtype):
+        if not copy:
+            values = np.asarray(values, dtype=default_dtype)
+        else:
+            values = np.array(values, dtype=default_dtype, copy=copy)
+
+    elif values.dtype.kind not in "iuf":
+        name = dtype_cls.__name__.strip("_")
+        raise TypeError(f"{values.dtype} cannot be converted to {name}")
+
+    if values.ndim != 1:
+        raise TypeError("values must be a 1D list-like")
+
+    if mask is None:
+        if values.dtype.kind in "iu":
+            # fastpath
+            mask = np.zeros(len(values), dtype=np.bool_)
+        else:
+            mask = libmissing.is_numeric_na(values)
+    else:
+        assert len(mask) == len(values)
+
+    if mask.ndim != 1:
+        raise TypeError("mask must be a 1D list-like")
+
+    # infer dtype if needed
+    if dtype is None:
+        dtype = default_dtype
+    else:
+        dtype = dtype.numpy_dtype
+
+    if is_integer_dtype(dtype) and values.dtype.kind == "f" and len(values) > 0:
+        if mask.all():
+            values = np.ones(values.shape, dtype=dtype)
+        else:
+            idx = np.nanargmax(values)
+            if int(values[idx]) != original[idx]:
+                # We have ints that lost precision during the cast.
+                inferred_type = lib.infer_dtype(original, skipna=True)
+                if (
+                    inferred_type not in ["floating", "mixed-integer-float"]
+                    and not mask.any()
+                ):
+                    values = np.asarray(original, dtype=dtype)
+                else:
+                    values = np.asarray(original, dtype="object")
+
+    # we copy as need to coerce here
+    if mask.any():
+        values = values.copy()
+        values[mask] = cls._internal_fill_value
+    if inferred_type in ("string", "unicode"):
+        # casts from str are always safe since they raise
+        # a ValueError if the str cannot be parsed into a float
+        values = values.astype(dtype, copy=copy)
+    else:
+        values = dtype_cls._safe_cast(values, dtype, copy=False)
+
+    return values, mask, dtype, inferred_type
+
+
+class NumericArray(BaseMaskedArray):
+    """
+    Base class for IntegerArray and FloatingArray.
+    """
+
+    _dtype_cls: type[NumericDtype]
+
+    def __init__(
+        self, values: np.ndarray, mask: npt.NDArray[np.bool_], copy: bool = False
+    ) -> None:
+        checker = self._dtype_cls._checker
+        if not (isinstance(values, np.ndarray) and checker(values.dtype)):
+            descr = (
+                "floating"
+                if self._dtype_cls.kind == "f"  # type: ignore[comparison-overlap]
+                else "integer"
+            )
+            raise TypeError(
+                f"values should be {descr} numpy array. Use "
+                "the 'pd.array' function instead"
+            )
+        if values.dtype == np.float16:
+            # If we don't raise here, then accessing self.dtype would raise
+            raise TypeError("FloatingArray does not support np.float16 dtype.")
+
+        super().__init__(values, mask, copy=copy)
+
+    @cache_readonly
+    def dtype(self) -> NumericDtype:
+        mapping = self._dtype_cls._get_dtype_mapping()
+        return mapping[self._data.dtype]
+
+    @classmethod
+    def _coerce_to_array(
+        cls, value, *, dtype: DtypeObj, copy: bool = False
+    ) -> tuple[np.ndarray, np.ndarray]:
+        dtype_cls = cls._dtype_cls
+        default_dtype = dtype_cls._default_np_dtype
+        values, mask, _, _ = _coerce_to_data_and_mask(
+            value, dtype, copy, dtype_cls, default_dtype
+        )
+        return values, mask
+
+    @classmethod
+    def _from_sequence_of_strings(
+        cls, strings, *, dtype: Dtype | None = None, copy: bool = False
+    ) -> Self:
+        from pandas.core.tools.numeric import to_numeric
+
+        scalars = to_numeric(strings, errors="raise", dtype_backend="numpy_nullable")
+        return cls._from_sequence(scalars, dtype=dtype, copy=copy)
+
+    _HANDLED_TYPES = (np.ndarray, numbers.Number)

llava_next/lib/python3.10/site-packages/pandas/core/arrays/period.py

@@ -0,0 +1,1313 @@
+from __future__ import annotations
+
+from datetime import timedelta
+import operator
+from typing import (
+    TYPE_CHECKING,
+    Any,
+    Callable,
+    Literal,
+    TypeVar,
+    cast,
+    overload,
+)
+import warnings
+
+import numpy as np
+
+from pandas._libs import (
+    algos as libalgos,
+    lib,
+)
+from pandas._libs.arrays import NDArrayBacked
+from pandas._libs.tslibs import (
+    BaseOffset,
+    NaT,
+    NaTType,
+    Timedelta,
+    add_overflowsafe,
+    astype_overflowsafe,
+    dt64arr_to_periodarr as c_dt64arr_to_periodarr,
+    get_unit_from_dtype,
+    iNaT,
+    parsing,
+    period as libperiod,
+    to_offset,
+)
+from pandas._libs.tslibs.dtypes import (
+    FreqGroup,
+    PeriodDtypeBase,
+    freq_to_period_freqstr,
+)
+from pandas._libs.tslibs.fields import isleapyear_arr
+from pandas._libs.tslibs.offsets import (
+    Tick,
+    delta_to_tick,
+)
+from pandas._libs.tslibs.period import (
+    DIFFERENT_FREQ,
+    IncompatibleFrequency,
+    Period,
+    get_period_field_arr,
+    period_asfreq_arr,
+)
+from pandas.util._decorators import (
+    cache_readonly,
+    doc,
+)
+from pandas.util._exceptions import find_stack_level
+
+from pandas.core.dtypes.common import (
+    ensure_object,
+    pandas_dtype,
+)
+from pandas.core.dtypes.dtypes import (
+    DatetimeTZDtype,
+    PeriodDtype,
+)
+from pandas.core.dtypes.generic import (
+    ABCIndex,
+    ABCPeriodIndex,
+    ABCSeries,
+    ABCTimedeltaArray,
+)
+from pandas.core.dtypes.missing import isna
+
+from pandas.core.arrays import datetimelike as dtl
+import pandas.core.common as com
+
+if TYPE_CHECKING:
+    from collections.abc import Sequence
+
+    from pandas._typing import (
+        AnyArrayLike,
+        Dtype,
+        FillnaOptions,
+        NpDtype,
+        NumpySorter,
+        NumpyValueArrayLike,
+        Self,
+        npt,
+    )
+
+    from pandas.core.arrays import (
+        DatetimeArray,
+        TimedeltaArray,
+    )
+    from pandas.core.arrays.base import ExtensionArray
+
+
+BaseOffsetT = TypeVar("BaseOffsetT", bound=BaseOffset)
+
+
+_shared_doc_kwargs = {
+    "klass": "PeriodArray",
+}
+
+
+def _field_accessor(name: str, docstring: str | None = None):
+    def f(self):
+        base = self.dtype._dtype_code
+        result = get_period_field_arr(name, self.asi8, base)
+        return result
+
+    f.__name__ = name
+    f.__doc__ = docstring
+    return property(f)
+
+
+# error: Definition of "_concat_same_type" in base class "NDArrayBacked" is
+# incompatible with definition in base class "ExtensionArray"
+class PeriodArray(dtl.DatelikeOps, libperiod.PeriodMixin):  # type: ignore[misc]
+    """
+    Pandas ExtensionArray for storing Period data.
+
+    Users should use :func:`~pandas.array` to create new instances.
+
+    Parameters
+    ----------
+    values : Union[PeriodArray, Series[period], ndarray[int], PeriodIndex]
+        The data to store. These should be arrays that can be directly
+        converted to ordinals without inference or copy (PeriodArray,
+        ndarray[int64]), or a box around such an array (Series[period],
+        PeriodIndex).
+    dtype : PeriodDtype, optional
+        A PeriodDtype instance from which to extract a `freq`. If both
+        `freq` and `dtype` are specified, then the frequencies must match.
+    freq : str or DateOffset
+        The `freq` to use for the array. Mostly applicable when `values`
+        is an ndarray of integers, when `freq` is required. When `values`
+        is a PeriodArray (or box around), it's checked that ``values.freq``
+        matches `freq`.
+    copy : bool, default False
+        Whether to copy the ordinals before storing.
+
+    Attributes
+    ----------
+    None
+
+    Methods
+    -------
+    None
+
+    See Also
+    --------
+    Period: Represents a period of time.
+    PeriodIndex : Immutable Index for period data.
+    period_range: Create a fixed-frequency PeriodArray.
+    array: Construct a pandas array.
+
+    Notes
+    -----
+    There are two components to a PeriodArray
+
+    - ordinals : integer ndarray
+    - freq : pd.tseries.offsets.Offset
+
+    The values are physically stored as a 1-D ndarray of integers. These are
+    called "ordinals" and represent some kind of offset from a base.
+
+    The `freq` indicates the span covered by each element of the array.
+    All elements in the PeriodArray have the same `freq`.
+
+    Examples
+    --------
+    >>> pd.arrays.PeriodArray(pd.PeriodIndex(['2023-01-01',
+    ...                                       '2023-01-02'], freq='D'))
+    <PeriodArray>
+    ['2023-01-01', '2023-01-02']
+    Length: 2, dtype: period[D]
+    """
+
+    # array priority higher than numpy scalars
+    __array_priority__ = 1000
+    _typ = "periodarray"  # ABCPeriodArray
+    _internal_fill_value = np.int64(iNaT)
+    _recognized_scalars = (Period,)
+    _is_recognized_dtype = lambda x: isinstance(
+        x, PeriodDtype
+    )  # check_compatible_with checks freq match
+    _infer_matches = ("period",)
+
+    @property
+    def _scalar_type(self) -> type[Period]:
+        return Period
+
+    # Names others delegate to us
+    _other_ops: list[str] = []
+    _bool_ops: list[str] = ["is_leap_year"]
+    _object_ops: list[str] = ["start_time", "end_time", "freq"]
+    _field_ops: list[str] = [
+        "year",
+        "month",
+        "day",
+        "hour",
+        "minute",
+        "second",
+        "weekofyear",
+        "weekday",
+        "week",
+        "dayofweek",
+        "day_of_week",
+        "dayofyear",
+        "day_of_year",
+        "quarter",
+        "qyear",
+        "days_in_month",
+        "daysinmonth",
+    ]
+    _datetimelike_ops: list[str] = _field_ops + _object_ops + _bool_ops
+    _datetimelike_methods: list[str] = ["strftime", "to_timestamp", "asfreq"]
+
+    _dtype: PeriodDtype
+
+    # --------------------------------------------------------------------
+    # Constructors
+
+    def __init__(
+        self, values, dtype: Dtype | None = None, freq=None, copy: bool = False
+    ) -> None:
+        if freq is not None:
+            # GH#52462
+            warnings.warn(
+                "The 'freq' keyword in the PeriodArray constructor is deprecated "
+                "and will be removed in a future version. Pass 'dtype' instead",
+                FutureWarning,
+                stacklevel=find_stack_level(),
+            )
+            freq = validate_dtype_freq(dtype, freq)
+            dtype = PeriodDtype(freq)
+
+        if dtype is not None:
+            dtype = pandas_dtype(dtype)
+            if not isinstance(dtype, PeriodDtype):
+                raise ValueError(f"Invalid dtype {dtype} for PeriodArray")
+
+        if isinstance(values, ABCSeries):
+            values = values._values
+            if not isinstance(values, type(self)):
+                raise TypeError("Incorrect dtype")
+
+        elif isinstance(values, ABCPeriodIndex):
+            values = values._values
+
+        if isinstance(values, type(self)):
+            if dtype is not None and dtype != values.dtype:
+                raise raise_on_incompatible(values, dtype.freq)
+            values, dtype = values._ndarray, values.dtype
+
+        if not copy:
+            values = np.asarray(values, dtype="int64")
+        else:
+            values = np.array(values, dtype="int64", copy=copy)
+        if dtype is None:
+            raise ValueError("dtype is not specified and cannot be inferred")
+        dtype = cast(PeriodDtype, dtype)
+        NDArrayBacked.__init__(self, values, dtype)
+
+    # error: Signature of "_simple_new" incompatible with supertype "NDArrayBacked"
+    @classmethod
+    def _simple_new(  # type: ignore[override]
+        cls,
+        values: npt.NDArray[np.int64],
+        dtype: PeriodDtype,
+    ) -> Self:
+        # alias for PeriodArray.__init__
+        assertion_msg = "Should be numpy array of type i8"
+        assert isinstance(values, np.ndarray) and values.dtype == "i8", assertion_msg
+        return cls(values, dtype=dtype)
+
+    @classmethod
+    def _from_sequence(
+        cls,
+        scalars,
+        *,
+        dtype: Dtype | None = None,
+        copy: bool = False,
+    ) -> Self:
+        if dtype is not None:
+            dtype = pandas_dtype(dtype)
+        if dtype and isinstance(dtype, PeriodDtype):
+            freq = dtype.freq
+        else:
+            freq = None
+
+        if isinstance(scalars, cls):
+            validate_dtype_freq(scalars.dtype, freq)
+            if copy:
+                scalars = scalars.copy()
+            return scalars
+
+        periods = np.asarray(scalars, dtype=object)
+
+        freq = freq or libperiod.extract_freq(periods)
+        ordinals = libperiod.extract_ordinals(periods, freq)
+        dtype = PeriodDtype(freq)
+        return cls(ordinals, dtype=dtype)
+
+    @classmethod
+    def _from_sequence_of_strings(
+        cls, strings, *, dtype: Dtype | None = None, copy: bool = False
+    ) -> Self:
+        return cls._from_sequence(strings, dtype=dtype, copy=copy)
+
+    @classmethod
+    def _from_datetime64(cls, data, freq, tz=None) -> Self:
+        """
+        Construct a PeriodArray from a datetime64 array
+
+        Parameters
+        ----------
+        data : ndarray[datetime64[ns], datetime64[ns, tz]]
+        freq : str or Tick
+        tz : tzinfo, optional
+
+        Returns
+        -------
+        PeriodArray[freq]
+        """
+        if isinstance(freq, BaseOffset):
+            freq = freq_to_period_freqstr(freq.n, freq.name)
+        data, freq = dt64arr_to_periodarr(data, freq, tz)
+        dtype = PeriodDtype(freq)
+        return cls(data, dtype=dtype)
+
+    @classmethod
+    def _generate_range(cls, start, end, periods, freq):
+        periods = dtl.validate_periods(periods)
+
+        if freq is not None:
+            freq = Period._maybe_convert_freq(freq)
+
+        if start is not None or end is not None:
+            subarr, freq = _get_ordinal_range(start, end, periods, freq)
+        else:
+            raise ValueError("Not enough parameters to construct Period range")
+
+        return subarr, freq
+
+    @classmethod
+    def _from_fields(cls, *, fields: dict, freq) -> Self:
+        subarr, freq = _range_from_fields(freq=freq, **fields)
+        dtype = PeriodDtype(freq)
+        return cls._simple_new(subarr, dtype=dtype)
+
+    # -----------------------------------------------------------------
+    # DatetimeLike Interface
+
+    # error: Argument 1 of "_unbox_scalar" is incompatible with supertype
+    # "DatetimeLikeArrayMixin"; supertype defines the argument type as
+    # "Union[Union[Period, Any, Timedelta], NaTType]"
+    def _unbox_scalar(  # type: ignore[override]
+        self,
+        value: Period | NaTType,
+    ) -> np.int64:
+        if value is NaT:
+            # error: Item "Period" of "Union[Period, NaTType]" has no attribute "value"
+            return np.int64(value._value)  # type: ignore[union-attr]
+        elif isinstance(value, self._scalar_type):
+            self._check_compatible_with(value)
+            return np.int64(value.ordinal)
+        else:
+            raise ValueError(f"'value' should be a Period. Got '{value}' instead.")
+
+    def _scalar_from_string(self, value: str) -> Period:
+        return Period(value, freq=self.freq)
+
+    # error: Argument 1 of "_check_compatible_with" is incompatible with
+    # supertype "DatetimeLikeArrayMixin"; supertype defines the argument type
+    # as "Period | Timestamp | Timedelta | NaTType"
+    def _check_compatible_with(self, other: Period | NaTType | PeriodArray) -> None:  # type: ignore[override]
+        if other is NaT:
+            return
+        # error: Item "NaTType" of "Period | NaTType | PeriodArray" has no
+        # attribute "freq"
+        self._require_matching_freq(other.freq)  # type: ignore[union-attr]
+
+    # --------------------------------------------------------------------
+    # Data / Attributes
+
+    @cache_readonly
+    def dtype(self) -> PeriodDtype:
+        return self._dtype
+
+    # error: Cannot override writeable attribute with read-only property
+    @property  # type: ignore[override]
+    def freq(self) -> BaseOffset:
+        """
+        Return the frequency object for this PeriodArray.
+        """
+        return self.dtype.freq
+
+    @property
+    def freqstr(self) -> str:
+        return freq_to_period_freqstr(self.freq.n, self.freq.name)
+
+    def __array__(
+        self, dtype: NpDtype | None = None, copy: bool | None = None
+    ) -> np.ndarray:
+        if dtype == "i8":
+            return self.asi8
+        elif dtype == bool:
+            return ~self._isnan
+
+        # This will raise TypeError for non-object dtypes
+        return np.array(list(self), dtype=object)
+
+    def __arrow_array__(self, type=None):
+        """
+        Convert myself into a pyarrow Array.
+        """
+        import pyarrow
+
+        from pandas.core.arrays.arrow.extension_types import ArrowPeriodType
+
+        if type is not None:
+            if pyarrow.types.is_integer(type):
+                return pyarrow.array(self._ndarray, mask=self.isna(), type=type)
+            elif isinstance(type, ArrowPeriodType):
+                # ensure we have the same freq
+                if self.freqstr != type.freq:
+                    raise TypeError(
+                        "Not supported to convert PeriodArray to array with different "
+                        f"'freq' ({self.freqstr} vs {type.freq})"
+                    )
+            else:
+                raise TypeError(
+                    f"Not supported to convert PeriodArray to '{type}' type"
+                )
+
+        period_type = ArrowPeriodType(self.freqstr)
+        storage_array = pyarrow.array(self._ndarray, mask=self.isna(), type="int64")
+        return pyarrow.ExtensionArray.from_storage(period_type, storage_array)
+
+    # --------------------------------------------------------------------
+    # Vectorized analogues of Period properties
+
+    year = _field_accessor(
+        "year",
+        """
+        The year of the period.
+
+        Examples
+        --------
+        >>> idx = pd.PeriodIndex(["2023", "2024", "2025"], freq="Y")
+        >>> idx.year
+        Index([2023, 2024, 2025], dtype='int64')
+        """,
+    )
+    month = _field_accessor(
+        "month",
+        """
+        The month as January=1, December=12.
+
+        Examples
+        --------
+        >>> idx = pd.PeriodIndex(["2023-01", "2023-02", "2023-03"], freq="M")
+        >>> idx.month
+        Index([1, 2, 3], dtype='int64')
+        """,
+    )
+    day = _field_accessor(
+        "day",
+        """
+        The days of the period.
+
+        Examples
+        --------
+        >>> idx = pd.PeriodIndex(['2020-01-31', '2020-02-28'], freq='D')
+        >>> idx.day
+        Index([31, 28], dtype='int64')
+        """,
+    )
+    hour = _field_accessor(
+        "hour",
+        """
+        The hour of the period.
+
+        Examples
+        --------
+        >>> idx = pd.PeriodIndex(["2023-01-01 10:00", "2023-01-01 11:00"], freq='h')
+        >>> idx.hour
+        Index([10, 11], dtype='int64')
+        """,
+    )
+    minute = _field_accessor(
+        "minute",
+        """
+        The minute of the period.
+
+        Examples
+        --------
+        >>> idx = pd.PeriodIndex(["2023-01-01 10:30:00",
+        ...                       "2023-01-01 11:50:00"], freq='min')
+        >>> idx.minute
+        Index([30, 50], dtype='int64')
+        """,
+    )
+    second = _field_accessor(
+        "second",
+        """
+        The second of the period.
+
+        Examples
+        --------
+        >>> idx = pd.PeriodIndex(["2023-01-01 10:00:30",
+        ...                       "2023-01-01 10:00:31"], freq='s')
+        >>> idx.second
+        Index([30, 31], dtype='int64')
+        """,
+    )
+    weekofyear = _field_accessor(
+        "week",
+        """
+        The week ordinal of the year.
+
+        Examples
+        --------
+        >>> idx = pd.PeriodIndex(["2023-01", "2023-02", "2023-03"], freq="M")
+        >>> idx.week  # It can be written `weekofyear`
+        Index([5, 9, 13], dtype='int64')
+        """,
+    )
+    week = weekofyear
+    day_of_week = _field_accessor(
+        "day_of_week",
+        """
+        The day of the week with Monday=0, Sunday=6.
+
+        Examples
+        --------
+        >>> idx = pd.PeriodIndex(["2023-01-01", "2023-01-02", "2023-01-03"], freq="D")
+        >>> idx.weekday
+        Index([6, 0, 1], dtype='int64')
+        """,
+    )
+    dayofweek = day_of_week
+    weekday = dayofweek
+    dayofyear = day_of_year = _field_accessor(
+        "day_of_year",
+        """
+        The ordinal day of the year.
+
+        Examples
+        --------
+        >>> idx = pd.PeriodIndex(["2023-01-10", "2023-02-01", "2023-03-01"], freq="D")
+        >>> idx.dayofyear
+        Index([10, 32, 60], dtype='int64')
+
+        >>> idx = pd.PeriodIndex(["2023", "2024", "2025"], freq="Y")
+        >>> idx
+        PeriodIndex(['2023', '2024', '2025'], dtype='period[Y-DEC]')
+        >>> idx.dayofyear
+        Index([365, 366, 365], dtype='int64')
+        """,
+    )
+    quarter = _field_accessor(
+        "quarter",
+        """
+        The quarter of the date.
+
+        Examples
+        --------
+        >>> idx = pd.PeriodIndex(["2023-01", "2023-02", "2023-03"], freq="M")
+        >>> idx.quarter
+        Index([1, 1, 1], dtype='int64')
+        """,
+    )
+    qyear = _field_accessor("qyear")
+    days_in_month = _field_accessor(
+        "days_in_month",
+        """
+        The number of days in the month.
+
+        Examples
+        --------
+        For Series:
+
+        >>> period = pd.period_range('2020-1-1 00:00', '2020-3-1 00:00', freq='M')
+        >>> s = pd.Series(period)
+        >>> s
+        0   2020-01
+        1   2020-02
+        2   2020-03
+        dtype: period[M]
+        >>> s.dt.days_in_month
+        0    31
+        1    29
+        2    31
+        dtype: int64
+
+        For PeriodIndex:
+
+        >>> idx = pd.PeriodIndex(["2023-01", "2023-02", "2023-03"], freq="M")
+        >>> idx.days_in_month   # It can be also entered as `daysinmonth`
+        Index([31, 28, 31], dtype='int64')
+        """,
+    )
+    daysinmonth = days_in_month
+
+    @property
+    def is_leap_year(self) -> npt.NDArray[np.bool_]:
+        """
+        Logical indicating if the date belongs to a leap year.
+
+        Examples
+        --------
+        >>> idx = pd.PeriodIndex(["2023", "2024", "2025"], freq="Y")
+        >>> idx.is_leap_year
+        array([False,  True, False])
+        """
+        return isleapyear_arr(np.asarray(self.year))
+
+    def to_timestamp(self, freq=None, how: str = "start") -> DatetimeArray:
+        """
+        Cast to DatetimeArray/Index.
+
+        Parameters
+        ----------
+        freq : str or DateOffset, optional
+            Target frequency. The default is 'D' for week or longer,
+            's' otherwise.
+        how : {'s', 'e', 'start', 'end'}
+            Whether to use the start or end of the time period being converted.
+
+        Returns
+        -------
+        DatetimeArray/Index
+
+        Examples
+        --------
+        >>> idx = pd.PeriodIndex(["2023-01", "2023-02", "2023-03"], freq="M")
+        >>> idx.to_timestamp()
+        DatetimeIndex(['2023-01-01', '2023-02-01', '2023-03-01'],
+        dtype='datetime64[ns]', freq='MS')
+        """
+        from pandas.core.arrays import DatetimeArray
+
+        how = libperiod.validate_end_alias(how)
+
+        end = how == "E"
+        if end:
+            if freq == "B" or self.freq == "B":
+                # roll forward to ensure we land on B date
+                adjust = Timedelta(1, "D") - Timedelta(1, "ns")
+                return self.to_timestamp(how="start") + adjust
+            else:
+                adjust = Timedelta(1, "ns")
+                return (self + self.freq).to_timestamp(how="start") - adjust
+
+        if freq is None:
+            freq_code = self._dtype._get_to_timestamp_base()
+            dtype = PeriodDtypeBase(freq_code, 1)
+            freq = dtype._freqstr
+            base = freq_code
+        else:
+            freq = Period._maybe_convert_freq(freq)
+            base = freq._period_dtype_code
+
+        new_parr = self.asfreq(freq, how=how)
+
+        new_data = libperiod.periodarr_to_dt64arr(new_parr.asi8, base)
+        dta = DatetimeArray._from_sequence(new_data)
+
+        if self.freq.name == "B":
+            # See if we can retain BDay instead of Day in cases where
+            #  len(self) is too small for infer_freq to distinguish between them
+            diffs = libalgos.unique_deltas(self.asi8)
+            if len(diffs) == 1:
+                diff = diffs[0]
+                if diff == self.dtype._n:
+                    dta._freq = self.freq
+                elif diff == 1:
+                    dta._freq = self.freq.base
+                # TODO: other cases?
+            return dta
+        else:
+            return dta._with_freq("infer")
+
+    # --------------------------------------------------------------------
+
+    def _box_func(self, x) -> Period | NaTType:
+        return Period._from_ordinal(ordinal=x, freq=self.freq)
+
+    @doc(**_shared_doc_kwargs, other="PeriodIndex", other_name="PeriodIndex")
+    def asfreq(self, freq=None, how: str = "E") -> Self:
+        """
+        Convert the {klass} to the specified frequency `freq`.
+
+        Equivalent to applying :meth:`pandas.Period.asfreq` with the given arguments
+        to each :class:`~pandas.Period` in this {klass}.
+
+        Parameters
+        ----------
+        freq : str
+            A frequency.
+        how : str {{'E', 'S'}}, default 'E'
+            Whether the elements should be aligned to the end
+            or start within pa period.
+
+            * 'E', 'END', or 'FINISH' for end,
+            * 'S', 'START', or 'BEGIN' for start.
+
+            January 31st ('END') vs. January 1st ('START') for example.
+
+        Returns
+        -------
+        {klass}
+            The transformed {klass} with the new frequency.
+
+        See Also
+        --------
+        {other}.asfreq: Convert each Period in a {other_name} to the given frequency.
+        Period.asfreq : Convert a :class:`~pandas.Period` object to the given frequency.
+
+        Examples
+        --------
+        >>> pidx = pd.period_range('2010-01-01', '2015-01-01', freq='Y')
+        >>> pidx
+        PeriodIndex(['2010', '2011', '2012', '2013', '2014', '2015'],
+        dtype='period[Y-DEC]')
+
+        >>> pidx.asfreq('M')
+        PeriodIndex(['2010-12', '2011-12', '2012-12', '2013-12', '2014-12',
+        '2015-12'], dtype='period[M]')
+
+        >>> pidx.asfreq('M', how='S')
+        PeriodIndex(['2010-01', '2011-01', '2012-01', '2013-01', '2014-01',
+        '2015-01'], dtype='period[M]')
+        """
+        how = libperiod.validate_end_alias(how)
+        if isinstance(freq, BaseOffset) and hasattr(freq, "_period_dtype_code"):
+            freq = PeriodDtype(freq)._freqstr
+        freq = Period._maybe_convert_freq(freq)
+
+        base1 = self._dtype._dtype_code
+        base2 = freq._period_dtype_code
+
+        asi8 = self.asi8
+        # self.freq.n can't be negative or 0
+        end = how == "E"
+        if end:
+            ordinal = asi8 + self.dtype._n - 1
+        else:
+            ordinal = asi8
+
+        new_data = period_asfreq_arr(ordinal, base1, base2, end)
+
+        if self._hasna:
+            new_data[self._isnan] = iNaT
+
+        dtype = PeriodDtype(freq)
+        return type(self)(new_data, dtype=dtype)
+
+    # ------------------------------------------------------------------
+    # Rendering Methods
+
+    def _formatter(self, boxed: bool = False):
+        if boxed:
+            return str
+        return "'{}'".format
+
+    def _format_native_types(
+        self, *, na_rep: str | float = "NaT", date_format=None, **kwargs
+    ) -> npt.NDArray[np.object_]:
+        """
+        actually format my specific types
+        """
+        return libperiod.period_array_strftime(
+            self.asi8, self.dtype._dtype_code, na_rep, date_format
+        )
+
+    # ------------------------------------------------------------------
+
+    def astype(self, dtype, copy: bool = True):
+        # We handle Period[T] -> Period[U]
+        # Our parent handles everything else.
+        dtype = pandas_dtype(dtype)
+        if dtype == self._dtype:
+            if not copy:
+                return self
+            else:
+                return self.copy()
+        if isinstance(dtype, PeriodDtype):
+            return self.asfreq(dtype.freq)
+
+        if lib.is_np_dtype(dtype, "M") or isinstance(dtype, DatetimeTZDtype):
+            # GH#45038 match PeriodIndex behavior.
+            tz = getattr(dtype, "tz", None)
+            unit = dtl.dtype_to_unit(dtype)
+            return self.to_timestamp().tz_localize(tz).as_unit(unit)
+
+        return super().astype(dtype, copy=copy)
+
+    def searchsorted(
+        self,
+        value: NumpyValueArrayLike | ExtensionArray,
+        side: Literal["left", "right"] = "left",
+        sorter: NumpySorter | None = None,
+    ) -> npt.NDArray[np.intp] | np.intp:
+        npvalue = self._validate_setitem_value(value).view("M8[ns]")
+
+        # Cast to M8 to get datetime-like NaT placement,
+        #  similar to dtl._period_dispatch
+        m8arr = self._ndarray.view("M8[ns]")
+        return m8arr.searchsorted(npvalue, side=side, sorter=sorter)
+
+    def _pad_or_backfill(
+        self,
+        *,
+        method: FillnaOptions,
+        limit: int | None = None,
+        limit_area: Literal["inside", "outside"] | None = None,
+        copy: bool = True,
+    ) -> Self:
+        # view as dt64 so we get treated as timelike in core.missing,
+        #  similar to dtl._period_dispatch
+        dta = self.view("M8[ns]")
+        result = dta._pad_or_backfill(
+            method=method, limit=limit, limit_area=limit_area, copy=copy
+        )
+        if copy:
+            return cast("Self", result.view(self.dtype))
+        else:
+            return self
+
+    def fillna(
+        self, value=None, method=None, limit: int | None = None, copy: bool = True
+    ) -> Self:
+        if method is not None:
+            # view as dt64 so we get treated as timelike in core.missing,
+            #  similar to dtl._period_dispatch
+            dta = self.view("M8[ns]")
+            result = dta.fillna(value=value, method=method, limit=limit, copy=copy)
+            # error: Incompatible return value type (got "Union[ExtensionArray,
+            # ndarray[Any, Any]]", expected "PeriodArray")
+            return result.view(self.dtype)  # type: ignore[return-value]
+        return super().fillna(value=value, method=method, limit=limit, copy=copy)
+
+    # ------------------------------------------------------------------
+    # Arithmetic Methods
+
+    def _addsub_int_array_or_scalar(
+        self, other: np.ndarray | int, op: Callable[[Any, Any], Any]
+    ) -> Self:
+        """
+        Add or subtract array of integers.
+
+        Parameters
+        ----------
+        other : np.ndarray[int64] or int
+        op : {operator.add, operator.sub}
+
+        Returns
+        -------
+        result : PeriodArray
+        """
+        assert op in [operator.add, operator.sub]
+        if op is operator.sub:
+            other = -other
+        res_values = add_overflowsafe(self.asi8, np.asarray(other, dtype="i8"))
+        return type(self)(res_values, dtype=self.dtype)
+
+    def _add_offset(self, other: BaseOffset):
+        assert not isinstance(other, Tick)
+
+        self._require_matching_freq(other, base=True)
+        return self._addsub_int_array_or_scalar(other.n, operator.add)
+
+    # TODO: can we de-duplicate with Period._add_timedeltalike_scalar?
+    def _add_timedeltalike_scalar(self, other):
+        """
+        Parameters
+        ----------
+        other : timedelta, Tick, np.timedelta64
+
+        Returns
+        -------
+        PeriodArray
+        """
+        if not isinstance(self.freq, Tick):
+            # We cannot add timedelta-like to non-tick PeriodArray
+            raise raise_on_incompatible(self, other)
+
+        if isna(other):
+            # i.e. np.timedelta64("NaT")
+            return super()._add_timedeltalike_scalar(other)
+
+        td = np.asarray(Timedelta(other).asm8)
+        return self._add_timedelta_arraylike(td)
+
+    def _add_timedelta_arraylike(
+        self, other: TimedeltaArray | npt.NDArray[np.timedelta64]
+    ) -> Self:
+        """
+        Parameters
+        ----------
+        other : TimedeltaArray or ndarray[timedelta64]
+
+        Returns
+        -------
+        PeriodArray
+        """
+        if not self.dtype._is_tick_like():
+            # We cannot add timedelta-like to non-tick PeriodArray
+            raise TypeError(
+                f"Cannot add or subtract timedelta64[ns] dtype from {self.dtype}"
+            )
+
+        dtype = np.dtype(f"m8[{self.dtype._td64_unit}]")
+
+        # Similar to _check_timedeltalike_freq_compat, but we raise with a
+        #  more specific exception message if necessary.
+        try:
+            delta = astype_overflowsafe(
+                np.asarray(other), dtype=dtype, copy=False, round_ok=False
+            )
+        except ValueError as err:
+            # e.g. if we have minutes freq and try to add 30s
+            # "Cannot losslessly convert units"
+            raise IncompatibleFrequency(
+                "Cannot add/subtract timedelta-like from PeriodArray that is "
+                "not an integer multiple of the PeriodArray's freq."
+            ) from err
+
+        res_values = add_overflowsafe(self.asi8, np.asarray(delta.view("i8")))
+        return type(self)(res_values, dtype=self.dtype)
+
+    def _check_timedeltalike_freq_compat(self, other):
+        """
+        Arithmetic operations with timedelta-like scalars or array `other`
+        are only valid if `other` is an integer multiple of `self.freq`.
+        If the operation is valid, find that integer multiple.  Otherwise,
+        raise because the operation is invalid.
+
+        Parameters
+        ----------
+        other : timedelta, np.timedelta64, Tick,
+                ndarray[timedelta64], TimedeltaArray, TimedeltaIndex
+
+        Returns
+        -------
+        multiple : int or ndarray[int64]
+
+        Raises
+        ------
+        IncompatibleFrequency
+        """
+        assert self.dtype._is_tick_like()  # checked by calling function
+
+        dtype = np.dtype(f"m8[{self.dtype._td64_unit}]")
+
+        if isinstance(other, (timedelta, np.timedelta64, Tick)):
+            td = np.asarray(Timedelta(other).asm8)
+        else:
+            td = np.asarray(other)
+
+        try:
+            delta = astype_overflowsafe(td, dtype=dtype, copy=False, round_ok=False)
+        except ValueError as err:
+            raise raise_on_incompatible(self, other) from err
+
+        delta = delta.view("i8")
+        return lib.item_from_zerodim(delta)
+
+
+def raise_on_incompatible(left, right) -> IncompatibleFrequency:
+    """
+    Helper function to render a consistent error message when raising
+    IncompatibleFrequency.
+
+    Parameters
+    ----------
+    left : PeriodArray
+    right : None, DateOffset, Period, ndarray, or timedelta-like
+
+    Returns
+    -------
+    IncompatibleFrequency
+        Exception to be raised by the caller.
+    """
+    # GH#24283 error message format depends on whether right is scalar
+    if isinstance(right, (np.ndarray, ABCTimedeltaArray)) or right is None:
+        other_freq = None
+    elif isinstance(right, BaseOffset):
+        other_freq = freq_to_period_freqstr(right.n, right.name)
+    elif isinstance(right, (ABCPeriodIndex, PeriodArray, Period)):
+        other_freq = right.freqstr
+    else:
+        other_freq = delta_to_tick(Timedelta(right)).freqstr
+
+    own_freq = freq_to_period_freqstr(left.freq.n, left.freq.name)
+    msg = DIFFERENT_FREQ.format(
+        cls=type(left).__name__, own_freq=own_freq, other_freq=other_freq
+    )
+    return IncompatibleFrequency(msg)
+
+
+# -------------------------------------------------------------------
+# Constructor Helpers
+
+
+def period_array(
+    data: Sequence[Period | str | None] | AnyArrayLike,
+    freq: str | Tick | BaseOffset | None = None,
+    copy: bool = False,
+) -> PeriodArray:
+    """
+    Construct a new PeriodArray from a sequence of Period scalars.
+
+    Parameters
+    ----------
+    data : Sequence of Period objects
+        A sequence of Period objects. These are required to all have
+        the same ``freq.`` Missing values can be indicated by ``None``
+        or ``pandas.NaT``.
+    freq : str, Tick, or Offset
+        The frequency of every element of the array. This can be specified
+        to avoid inferring the `freq` from `data`.
+    copy : bool, default False
+        Whether to ensure a copy of the data is made.
+
+    Returns
+    -------
+    PeriodArray
+
+    See Also
+    --------
+    PeriodArray
+    pandas.PeriodIndex
+
+    Examples
+    --------
+    >>> period_array([pd.Period('2017', freq='Y'),
+    ...               pd.Period('2018', freq='Y')])
+    <PeriodArray>
+    ['2017', '2018']
+    Length: 2, dtype: period[Y-DEC]
+
+    >>> period_array([pd.Period('2017', freq='Y'),
+    ...               pd.Period('2018', freq='Y'),
+    ...               pd.NaT])
+    <PeriodArray>
+    ['2017', '2018', 'NaT']
+    Length: 3, dtype: period[Y-DEC]
+
+    Integers that look like years are handled
+
+    >>> period_array([2000, 2001, 2002], freq='D')
+    <PeriodArray>
+    ['2000-01-01', '2001-01-01', '2002-01-01']
+    Length: 3, dtype: period[D]
+
+    Datetime-like strings may also be passed
+
+    >>> period_array(['2000-Q1', '2000-Q2', '2000-Q3', '2000-Q4'], freq='Q')
+    <PeriodArray>
+    ['2000Q1', '2000Q2', '2000Q3', '2000Q4']
+    Length: 4, dtype: period[Q-DEC]
+    """
+    data_dtype = getattr(data, "dtype", None)
+
+    if lib.is_np_dtype(data_dtype, "M"):
+        return PeriodArray._from_datetime64(data, freq)
+    if isinstance(data_dtype, PeriodDtype):
+        out = PeriodArray(data)
+        if freq is not None:
+            if freq == data_dtype.freq:
+                return out
+            return out.asfreq(freq)
+        return out
+
+    # other iterable of some kind
+    if not isinstance(data, (np.ndarray, list, tuple, ABCSeries)):
+        data = list(data)
+
+    arrdata = np.asarray(data)
+
+    dtype: PeriodDtype | None
+    if freq:
+        dtype = PeriodDtype(freq)
+    else:
+        dtype = None
+
+    if arrdata.dtype.kind == "f" and len(arrdata) > 0:
+        raise TypeError("PeriodIndex does not allow floating point in construction")
+
+    if arrdata.dtype.kind in "iu":
+        arr = arrdata.astype(np.int64, copy=False)
+        # error: Argument 2 to "from_ordinals" has incompatible type "Union[str,
+        # Tick, None]"; expected "Union[timedelta, BaseOffset, str]"
+        ordinals = libperiod.from_ordinals(arr, freq)  # type: ignore[arg-type]
+        return PeriodArray(ordinals, dtype=dtype)
+
+    data = ensure_object(arrdata)
+    if freq is None:
+        freq = libperiod.extract_freq(data)
+    dtype = PeriodDtype(freq)
+    return PeriodArray._from_sequence(data, dtype=dtype)
+
+
+@overload
+def validate_dtype_freq(dtype, freq: BaseOffsetT) -> BaseOffsetT:
+    ...
+
+
+@overload
+def validate_dtype_freq(dtype, freq: timedelta | str | None) -> BaseOffset:
+    ...
+
+
+def validate_dtype_freq(
+    dtype, freq: BaseOffsetT | BaseOffset | timedelta | str | None
+) -> BaseOffsetT:
+    """
+    If both a dtype and a freq are available, ensure they match.  If only
+    dtype is available, extract the implied freq.
+
+    Parameters
+    ----------
+    dtype : dtype
+    freq : DateOffset or None
+
+    Returns
+    -------
+    freq : DateOffset
+
+    Raises
+    ------
+    ValueError : non-period dtype
+    IncompatibleFrequency : mismatch between dtype and freq
+    """
+    if freq is not None:
+        freq = to_offset(freq, is_period=True)
+
+    if dtype is not None:
+        dtype = pandas_dtype(dtype)
+        if not isinstance(dtype, PeriodDtype):
+            raise ValueError("dtype must be PeriodDtype")
+        if freq is None:
+            freq = dtype.freq
+        elif freq != dtype.freq:
+            raise IncompatibleFrequency("specified freq and dtype are different")
+    # error: Incompatible return value type (got "Union[BaseOffset, Any, None]",
+    # expected "BaseOffset")
+    return freq  # type: ignore[return-value]
+
+
+def dt64arr_to_periodarr(
+    data, freq, tz=None
+) -> tuple[npt.NDArray[np.int64], BaseOffset]:
+    """
+    Convert an datetime-like array to values Period ordinals.
+
+    Parameters
+    ----------
+    data : Union[Series[datetime64[ns]], DatetimeIndex, ndarray[datetime64ns]]
+    freq : Optional[Union[str, Tick]]
+        Must match the `freq` on the `data` if `data` is a DatetimeIndex
+        or Series.
+    tz : Optional[tzinfo]
+
+    Returns
+    -------
+    ordinals : ndarray[int64]
+    freq : Tick
+        The frequency extracted from the Series or DatetimeIndex if that's
+        used.
+
+    """
+    if not isinstance(data.dtype, np.dtype) or data.dtype.kind != "M":
+        raise ValueError(f"Wrong dtype: {data.dtype}")
+
+    if freq is None:
+        if isinstance(data, ABCIndex):
+            data, freq = data._values, data.freq
+        elif isinstance(data, ABCSeries):
+            data, freq = data._values, data.dt.freq
+
+    elif isinstance(data, (ABCIndex, ABCSeries)):
+        data = data._values
+
+    reso = get_unit_from_dtype(data.dtype)
+    freq = Period._maybe_convert_freq(freq)
+    base = freq._period_dtype_code
+    return c_dt64arr_to_periodarr(data.view("i8"), base, tz, reso=reso), freq
+
+
+def _get_ordinal_range(start, end, periods, freq, mult: int = 1):
+    if com.count_not_none(start, end, periods) != 2:
+        raise ValueError(
+            "Of the three parameters: start, end, and periods, "
+            "exactly two must be specified"
+        )
+
+    if freq is not None:
+        freq = to_offset(freq, is_period=True)
+        mult = freq.n
+
+    if start is not None:
+        start = Period(start, freq)
+    if end is not None:
+        end = Period(end, freq)
+
+    is_start_per = isinstance(start, Period)
+    is_end_per = isinstance(end, Period)
+
+    if is_start_per and is_end_per and start.freq != end.freq:
+        raise ValueError("start and end must have same freq")
+    if start is NaT or end is NaT:
+        raise ValueError("start and end must not be NaT")
+
+    if freq is None:
+        if is_start_per:
+            freq = start.freq
+        elif is_end_per:
+            freq = end.freq
+        else:  # pragma: no cover
+            raise ValueError("Could not infer freq from start/end")
+        mult = freq.n
+
+    if periods is not None:
+        periods = periods * mult
+        if start is None:
+            data = np.arange(
+                end.ordinal - periods + mult, end.ordinal + 1, mult, dtype=np.int64
+            )
+        else:
+            data = np.arange(
+                start.ordinal, start.ordinal + periods, mult, dtype=np.int64
+            )
+    else:
+        data = np.arange(start.ordinal, end.ordinal + 1, mult, dtype=np.int64)
+
+    return data, freq
+
+
+def _range_from_fields(
+    year=None,
+    month=None,
+    quarter=None,
+    day=None,
+    hour=None,
+    minute=None,
+    second=None,
+    freq=None,
+) -> tuple[np.ndarray, BaseOffset]:
+    if hour is None:
+        hour = 0
+    if minute is None:
+        minute = 0
+    if second is None:
+        second = 0
+    if day is None:
+        day = 1
+
+    ordinals = []
+
+    if quarter is not None:
+        if freq is None:
+            freq = to_offset("Q", is_period=True)
+            base = FreqGroup.FR_QTR.value
+        else:
+            freq = to_offset(freq, is_period=True)
+            base = libperiod.freq_to_dtype_code(freq)
+            if base != FreqGroup.FR_QTR.value:
+                raise AssertionError("base must equal FR_QTR")
+
+        freqstr = freq.freqstr
+        year, quarter = _make_field_arrays(year, quarter)
+        for y, q in zip(year, quarter):
+            calendar_year, calendar_month = parsing.quarter_to_myear(y, q, freqstr)
+            val = libperiod.period_ordinal(
+                calendar_year, calendar_month, 1, 1, 1, 1, 0, 0, base
+            )
+            ordinals.append(val)
+    else:
+        freq = to_offset(freq, is_period=True)
+        base = libperiod.freq_to_dtype_code(freq)
+        arrays = _make_field_arrays(year, month, day, hour, minute, second)
+        for y, mth, d, h, mn, s in zip(*arrays):
+            ordinals.append(libperiod.period_ordinal(y, mth, d, h, mn, s, 0, 0, base))
+
+    return np.array(ordinals, dtype=np.int64), freq
+
+
+def _make_field_arrays(*fields) -> list[np.ndarray]:
+    length = None
+    for x in fields:
+        if isinstance(x, (list, np.ndarray, ABCSeries)):
+            if length is not None and len(x) != length:
+                raise ValueError("Mismatched Period array lengths")
+            if length is None:
+                length = len(x)
+
+    # error: Argument 2 to "repeat" has incompatible type "Optional[int]"; expected
+    # "Union[Union[int, integer[Any]], Union[bool, bool_], ndarray, Sequence[Union[int,
+    # integer[Any]]], Sequence[Union[bool, bool_]], Sequence[Sequence[Any]]]"
+    return [
+        np.asarray(x)
+        if isinstance(x, (np.ndarray, list, ABCSeries))
+        else np.repeat(x, length)  # type: ignore[arg-type]
+        for x in fields
+    ]

llava_next/lib/python3.10/site-packages/pandas/core/config_init.py

@@ -0,0 +1,924 @@
+"""
+This module is imported from the pandas package __init__.py file
+in order to ensure that the core.config options registered here will
+be available as soon as the user loads the package. if register_option
+is invoked inside specific modules, they will not be registered until that
+module is imported, which may or may not be a problem.
+
+If you need to make sure options are available even before a certain
+module is imported, register them here rather than in the module.
+
+"""
+from __future__ import annotations
+
+import os
+from typing import Callable
+
+import pandas._config.config as cf
+from pandas._config.config import (
+    is_bool,
+    is_callable,
+    is_instance_factory,
+    is_int,
+    is_nonnegative_int,
+    is_one_of_factory,
+    is_str,
+    is_text,
+)
+
+# compute
+
+use_bottleneck_doc = """
+: bool
+    Use the bottleneck library to accelerate if it is installed,
+    the default is True
+    Valid values: False,True
+"""
+
+
+def use_bottleneck_cb(key) -> None:
+    from pandas.core import nanops
+
+    nanops.set_use_bottleneck(cf.get_option(key))
+
+
+use_numexpr_doc = """
+: bool
+    Use the numexpr library to accelerate computation if it is installed,
+    the default is True
+    Valid values: False,True
+"""
+
+
+def use_numexpr_cb(key) -> None:
+    from pandas.core.computation import expressions
+
+    expressions.set_use_numexpr(cf.get_option(key))
+
+
+use_numba_doc = """
+: bool
+    Use the numba engine option for select operations if it is installed,
+    the default is False
+    Valid values: False,True
+"""
+
+
+def use_numba_cb(key) -> None:
+    from pandas.core.util import numba_
+
+    numba_.set_use_numba(cf.get_option(key))
+
+
+with cf.config_prefix("compute"):
+    cf.register_option(
+        "use_bottleneck",
+        True,
+        use_bottleneck_doc,
+        validator=is_bool,
+        cb=use_bottleneck_cb,
+    )
+    cf.register_option(
+        "use_numexpr", True, use_numexpr_doc, validator=is_bool, cb=use_numexpr_cb
+    )
+    cf.register_option(
+        "use_numba", False, use_numba_doc, validator=is_bool, cb=use_numba_cb
+    )
+#
+# options from the "display" namespace
+
+pc_precision_doc = """
+: int
+    Floating point output precision in terms of number of places after the
+    decimal, for regular formatting as well as scientific notation. Similar
+    to ``precision`` in :meth:`numpy.set_printoptions`.
+"""
+
+pc_colspace_doc = """
+: int
+    Default space for DataFrame columns.
+"""
+
+pc_max_rows_doc = """
+: int
+    If max_rows is exceeded, switch to truncate view. Depending on
+    `large_repr`, objects are either centrally truncated or printed as
+    a summary view. 'None' value means unlimited.
+
+    In case python/IPython is running in a terminal and `large_repr`
+    equals 'truncate' this can be set to 0 and pandas will auto-detect
+    the height of the terminal and print a truncated object which fits
+    the screen height. The IPython notebook, IPython qtconsole, or
+    IDLE do not run in a terminal and hence it is not possible to do
+    correct auto-detection.
+"""
+
+pc_min_rows_doc = """
+: int
+    The numbers of rows to show in a truncated view (when `max_rows` is
+    exceeded). Ignored when `max_rows` is set to None or 0. When set to
+    None, follows the value of `max_rows`.
+"""
+
+pc_max_cols_doc = """
+: int
+    If max_cols is exceeded, switch to truncate view. Depending on
+    `large_repr`, objects are either centrally truncated or printed as
+    a summary view. 'None' value means unlimited.
+
+    In case python/IPython is running in a terminal and `large_repr`
+    equals 'truncate' this can be set to 0 or None and pandas will auto-detect
+    the width of the terminal and print a truncated object which fits
+    the screen width. The IPython notebook, IPython qtconsole, or IDLE
+    do not run in a terminal and hence it is not possible to do
+    correct auto-detection and defaults to 20.
+"""
+
+pc_max_categories_doc = """
+: int
+    This sets the maximum number of categories pandas should output when
+    printing out a `Categorical` or a Series of dtype "category".
+"""
+
+pc_max_info_cols_doc = """
+: int
+    max_info_columns is used in DataFrame.info method to decide if
+    per column information will be printed.
+"""
+
+pc_nb_repr_h_doc = """
+: boolean
+    When True, IPython notebook will use html representation for
+    pandas objects (if it is available).
+"""
+
+pc_pprint_nest_depth = """
+: int
+    Controls the number of nested levels to process when pretty-printing
+"""
+
+pc_multi_sparse_doc = """
+: boolean
+    "sparsify" MultiIndex display (don't display repeated
+    elements in outer levels within groups)
+"""
+
+float_format_doc = """
+: callable
+    The callable should accept a floating point number and return
+    a string with the desired format of the number. This is used
+    in some places like SeriesFormatter.
+    See formats.format.EngFormatter for an example.
+"""
+
+max_colwidth_doc = """
+: int or None
+    The maximum width in characters of a column in the repr of
+    a pandas data structure. When the column overflows, a "..."
+    placeholder is embedded in the output. A 'None' value means unlimited.
+"""
+
+colheader_justify_doc = """
+: 'left'/'right'
+    Controls the justification of column headers. used by DataFrameFormatter.
+"""
+
+pc_expand_repr_doc = """
+: boolean
+    Whether to print out the full DataFrame repr for wide DataFrames across
+    multiple lines, `max_columns` is still respected, but the output will
+    wrap-around across multiple "pages" if its width exceeds `display.width`.
+"""
+
+pc_show_dimensions_doc = """
+: boolean or 'truncate'
+    Whether to print out dimensions at the end of DataFrame repr.
+    If 'truncate' is specified, only print out the dimensions if the
+    frame is truncated (e.g. not display all rows and/or columns)
+"""
+
+pc_east_asian_width_doc = """
+: boolean
+    Whether to use the Unicode East Asian Width to calculate the display text
+    width.
+    Enabling this may affect to the performance (default: False)
+"""
+
+pc_ambiguous_as_wide_doc = """
+: boolean
+    Whether to handle Unicode characters belong to Ambiguous as Wide (width=2)
+    (default: False)
+"""
+
+pc_table_schema_doc = """
+: boolean
+    Whether to publish a Table Schema representation for frontends
+    that support it.
+    (default: False)
+"""
+
+pc_html_border_doc = """
+: int
+    A ``border=value`` attribute is inserted in the ``<table>`` tag
+    for the DataFrame HTML repr.
+"""
+
+pc_html_use_mathjax_doc = """\
+: boolean
+    When True, Jupyter notebook will process table contents using MathJax,
+    rendering mathematical expressions enclosed by the dollar symbol.
+    (default: True)
+"""
+
+pc_max_dir_items = """\
+: int
+    The number of items that will be added to `dir(...)`. 'None' value means
+    unlimited. Because dir is cached, changing this option will not immediately
+    affect already existing dataframes until a column is deleted or added.
+
+    This is for instance used to suggest columns from a dataframe to tab
+    completion.
+"""
+
+pc_width_doc = """
+: int
+    Width of the display in characters. In case python/IPython is running in
+    a terminal this can be set to None and pandas will correctly auto-detect
+    the width.
+    Note that the IPython notebook, IPython qtconsole, or IDLE do not run in a
+    terminal and hence it is not possible to correctly detect the width.
+"""
+
+pc_chop_threshold_doc = """
+: float or None
+    if set to a float value, all float values smaller than the given threshold
+    will be displayed as exactly 0 by repr and friends.
+"""
+
+pc_max_seq_items = """
+: int or None
+    When pretty-printing a long sequence, no more then `max_seq_items`
+    will be printed. If items are omitted, they will be denoted by the
+    addition of "..." to the resulting string.
+
+    If set to None, the number of items to be printed is unlimited.
+"""
+
+pc_max_info_rows_doc = """
+: int
+    df.info() will usually show null-counts for each column.
+    For large frames this can be quite slow. max_info_rows and max_info_cols
+    limit this null check only to frames with smaller dimensions than
+    specified.
+"""
+
+pc_large_repr_doc = """
+: 'truncate'/'info'
+    For DataFrames exceeding max_rows/max_cols, the repr (and HTML repr) can
+    show a truncated table, or switch to the view from
+    df.info() (the behaviour in earlier versions of pandas).
+"""
+
+pc_memory_usage_doc = """
+: bool, string or None
+    This specifies if the memory usage of a DataFrame should be displayed when
+    df.info() is called. Valid values True,False,'deep'
+"""
+
+
+def table_schema_cb(key) -> None:
+    from pandas.io.formats.printing import enable_data_resource_formatter
+
+    enable_data_resource_formatter(cf.get_option(key))
+
+
+def is_terminal() -> bool:
+    """
+    Detect if Python is running in a terminal.
+
+    Returns True if Python is running in a terminal or False if not.
+    """
+    try:
+        # error: Name 'get_ipython' is not defined
+        ip = get_ipython()  # type: ignore[name-defined]
+    except NameError:  # assume standard Python interpreter in a terminal
+        return True
+    else:
+        if hasattr(ip, "kernel"):  # IPython as a Jupyter kernel
+            return False
+        else:  # IPython in a terminal
+            return True
+
+
+with cf.config_prefix("display"):
+    cf.register_option("precision", 6, pc_precision_doc, validator=is_nonnegative_int)
+    cf.register_option(
+        "float_format",
+        None,
+        float_format_doc,
+        validator=is_one_of_factory([None, is_callable]),
+    )
+    cf.register_option(
+        "max_info_rows",
+        1690785,
+        pc_max_info_rows_doc,
+        validator=is_int,
+    )
+    cf.register_option("max_rows", 60, pc_max_rows_doc, validator=is_nonnegative_int)
+    cf.register_option(
+        "min_rows",
+        10,
+        pc_min_rows_doc,
+        validator=is_instance_factory([type(None), int]),
+    )
+    cf.register_option("max_categories", 8, pc_max_categories_doc, validator=is_int)
+
+    cf.register_option(
+        "max_colwidth",
+        50,
+        max_colwidth_doc,
+        validator=is_nonnegative_int,
+    )
+    if is_terminal():
+        max_cols = 0  # automatically determine optimal number of columns
+    else:
+        max_cols = 20  # cannot determine optimal number of columns
+    cf.register_option(
+        "max_columns", max_cols, pc_max_cols_doc, validator=is_nonnegative_int
+    )
+    cf.register_option(
+        "large_repr",
+        "truncate",
+        pc_large_repr_doc,
+        validator=is_one_of_factory(["truncate", "info"]),
+    )
+    cf.register_option("max_info_columns", 100, pc_max_info_cols_doc, validator=is_int)
+    cf.register_option(
+        "colheader_justify", "right", colheader_justify_doc, validator=is_text
+    )
+    cf.register_option("notebook_repr_html", True, pc_nb_repr_h_doc, validator=is_bool)
+    cf.register_option("pprint_nest_depth", 3, pc_pprint_nest_depth, validator=is_int)
+    cf.register_option("multi_sparse", True, pc_multi_sparse_doc, validator=is_bool)
+    cf.register_option("expand_frame_repr", True, pc_expand_repr_doc)
+    cf.register_option(
+        "show_dimensions",
+        "truncate",
+        pc_show_dimensions_doc,
+        validator=is_one_of_factory([True, False, "truncate"]),
+    )
+    cf.register_option("chop_threshold", None, pc_chop_threshold_doc)
+    cf.register_option("max_seq_items", 100, pc_max_seq_items)
+    cf.register_option(
+        "width", 80, pc_width_doc, validator=is_instance_factory([type(None), int])
+    )
+    cf.register_option(
+        "memory_usage",
+        True,
+        pc_memory_usage_doc,
+        validator=is_one_of_factory([None, True, False, "deep"]),
+    )
+    cf.register_option(
+        "unicode.east_asian_width", False, pc_east_asian_width_doc, validator=is_bool
+    )
+    cf.register_option(
+        "unicode.ambiguous_as_wide", False, pc_east_asian_width_doc, validator=is_bool
+    )
+    cf.register_option(
+        "html.table_schema",
+        False,
+        pc_table_schema_doc,
+        validator=is_bool,
+        cb=table_schema_cb,
+    )
+    cf.register_option("html.border", 1, pc_html_border_doc, validator=is_int)
+    cf.register_option(
+        "html.use_mathjax", True, pc_html_use_mathjax_doc, validator=is_bool
+    )
+    cf.register_option(
+        "max_dir_items", 100, pc_max_dir_items, validator=is_nonnegative_int
+    )
+
+tc_sim_interactive_doc = """
+: boolean
+    Whether to simulate interactive mode for purposes of testing
+"""
+
+with cf.config_prefix("mode"):
+    cf.register_option("sim_interactive", False, tc_sim_interactive_doc)
+
+use_inf_as_na_doc = """
+: boolean
+    True means treat None, NaN, INF, -INF as NA (old way),
+    False means None and NaN are null, but INF, -INF are not NA
+    (new way).
+
+    This option is deprecated in pandas 2.1.0 and will be removed in 3.0.
+"""
+
+# We don't want to start importing everything at the global context level
+# or we'll hit circular deps.
+
+
+def use_inf_as_na_cb(key) -> None:
+    # TODO(3.0): enforcing this deprecation will close GH#52501
+    from pandas.core.dtypes.missing import _use_inf_as_na
+
+    _use_inf_as_na(key)
+
+
+with cf.config_prefix("mode"):
+    cf.register_option("use_inf_as_na", False, use_inf_as_na_doc, cb=use_inf_as_na_cb)
+
+cf.deprecate_option(
+    # GH#51684
+    "mode.use_inf_as_na",
+    "use_inf_as_na option is deprecated and will be removed in a future "
+    "version. Convert inf values to NaN before operating instead.",
+)
+
+data_manager_doc = """
+: string
+    Internal data manager type; can be "block" or "array". Defaults to "block",
+    unless overridden by the 'PANDAS_DATA_MANAGER' environment variable (needs
+    to be set before pandas is imported).
+"""
+
+
+with cf.config_prefix("mode"):
+    cf.register_option(
+        "data_manager",
+        # Get the default from an environment variable, if set, otherwise defaults
+        # to "block". This environment variable can be set for testing.
+        os.environ.get("PANDAS_DATA_MANAGER", "block"),
+        data_manager_doc,
+        validator=is_one_of_factory(["block", "array"]),
+    )
+
+cf.deprecate_option(
+    # GH#55043
+    "mode.data_manager",
+    "data_manager option is deprecated and will be removed in a future "
+    "version. Only the BlockManager will be available.",
+)
+
+
+# TODO better name?
+copy_on_write_doc = """
+: bool
+    Use new copy-view behaviour using Copy-on-Write. Defaults to False,
+    unless overridden by the 'PANDAS_COPY_ON_WRITE' environment variable
+    (if set to "1" for True, needs to be set before pandas is imported).
+"""
+
+
+with cf.config_prefix("mode"):
+    cf.register_option(
+        "copy_on_write",
+        # Get the default from an environment variable, if set, otherwise defaults
+        # to False. This environment variable can be set for testing.
+        "warn"
+        if os.environ.get("PANDAS_COPY_ON_WRITE", "0") == "warn"
+        else os.environ.get("PANDAS_COPY_ON_WRITE", "0") == "1",
+        copy_on_write_doc,
+        validator=is_one_of_factory([True, False, "warn"]),
+    )
+
+
+# user warnings
+chained_assignment = """
+: string
+    Raise an exception, warn, or no action if trying to use chained assignment,
+    The default is warn
+"""
+
+with cf.config_prefix("mode"):
+    cf.register_option(
+        "chained_assignment",
+        "warn",
+        chained_assignment,
+        validator=is_one_of_factory([None, "warn", "raise"]),
+    )
+
+
+string_storage_doc = """
+: string
+    The default storage for StringDtype. This option is ignored if
+    ``future.infer_string`` is set to True.
+"""
+
+with cf.config_prefix("mode"):
+    cf.register_option(
+        "string_storage",
+        "python",
+        string_storage_doc,
+        validator=is_one_of_factory(["python", "pyarrow", "pyarrow_numpy"]),
+    )
+
+
+# Set up the io.excel specific reader configuration.
+reader_engine_doc = """
+: string
+    The default Excel reader engine for '{ext}' files. Available options:
+    auto, {others}.
+"""
+
+_xls_options = ["xlrd", "calamine"]
+_xlsm_options = ["xlrd", "openpyxl", "calamine"]
+_xlsx_options = ["xlrd", "openpyxl", "calamine"]
+_ods_options = ["odf", "calamine"]
+_xlsb_options = ["pyxlsb", "calamine"]
+
+
+with cf.config_prefix("io.excel.xls"):
+    cf.register_option(
+        "reader",
+        "auto",
+        reader_engine_doc.format(ext="xls", others=", ".join(_xls_options)),
+        validator=is_one_of_factory(_xls_options + ["auto"]),
+    )
+
+with cf.config_prefix("io.excel.xlsm"):
+    cf.register_option(
+        "reader",
+        "auto",
+        reader_engine_doc.format(ext="xlsm", others=", ".join(_xlsm_options)),
+        validator=is_one_of_factory(_xlsm_options + ["auto"]),
+    )
+
+
+with cf.config_prefix("io.excel.xlsx"):
+    cf.register_option(
+        "reader",
+        "auto",
+        reader_engine_doc.format(ext="xlsx", others=", ".join(_xlsx_options)),
+        validator=is_one_of_factory(_xlsx_options + ["auto"]),
+    )
+
+
+with cf.config_prefix("io.excel.ods"):
+    cf.register_option(
+        "reader",
+        "auto",
+        reader_engine_doc.format(ext="ods", others=", ".join(_ods_options)),
+        validator=is_one_of_factory(_ods_options + ["auto"]),
+    )
+
+with cf.config_prefix("io.excel.xlsb"):
+    cf.register_option(
+        "reader",
+        "auto",
+        reader_engine_doc.format(ext="xlsb", others=", ".join(_xlsb_options)),
+        validator=is_one_of_factory(_xlsb_options + ["auto"]),
+    )
+
+# Set up the io.excel specific writer configuration.
+writer_engine_doc = """
+: string
+    The default Excel writer engine for '{ext}' files. Available options:
+    auto, {others}.
+"""
+
+_xlsm_options = ["openpyxl"]
+_xlsx_options = ["openpyxl", "xlsxwriter"]
+_ods_options = ["odf"]
+
+
+with cf.config_prefix("io.excel.xlsm"):
+    cf.register_option(
+        "writer",
+        "auto",
+        writer_engine_doc.format(ext="xlsm", others=", ".join(_xlsm_options)),
+        validator=str,
+    )
+
+
+with cf.config_prefix("io.excel.xlsx"):
+    cf.register_option(
+        "writer",
+        "auto",
+        writer_engine_doc.format(ext="xlsx", others=", ".join(_xlsx_options)),
+        validator=str,
+    )
+
+
+with cf.config_prefix("io.excel.ods"):
+    cf.register_option(
+        "writer",
+        "auto",
+        writer_engine_doc.format(ext="ods", others=", ".join(_ods_options)),
+        validator=str,
+    )
+
+
+# Set up the io.parquet specific configuration.
+parquet_engine_doc = """
+: string
+    The default parquet reader/writer engine. Available options:
+    'auto', 'pyarrow', 'fastparquet', the default is 'auto'
+"""
+
+with cf.config_prefix("io.parquet"):
+    cf.register_option(
+        "engine",
+        "auto",
+        parquet_engine_doc,
+        validator=is_one_of_factory(["auto", "pyarrow", "fastparquet"]),
+    )
+
+
+# Set up the io.sql specific configuration.
+sql_engine_doc = """
+: string
+    The default sql reader/writer engine. Available options:
+    'auto', 'sqlalchemy', the default is 'auto'
+"""
+
+with cf.config_prefix("io.sql"):
+    cf.register_option(
+        "engine",
+        "auto",
+        sql_engine_doc,
+        validator=is_one_of_factory(["auto", "sqlalchemy"]),
+    )
+
+# --------
+# Plotting
+# ---------
+
+plotting_backend_doc = """
+: str
+    The plotting backend to use. The default value is "matplotlib", the
+    backend provided with pandas. Other backends can be specified by
+    providing the name of the module that implements the backend.
+"""
+
+
+def register_plotting_backend_cb(key) -> None:
+    if key == "matplotlib":
+        # We defer matplotlib validation, since it's the default
+        return
+    from pandas.plotting._core import _get_plot_backend
+
+    _get_plot_backend(key)
+
+
+with cf.config_prefix("plotting"):
+    cf.register_option(
+        "backend",
+        defval="matplotlib",
+        doc=plotting_backend_doc,
+        validator=register_plotting_backend_cb,
+    )
+
+
+register_converter_doc = """
+: bool or 'auto'.
+    Whether to register converters with matplotlib's units registry for
+    dates, times, datetimes, and Periods. Toggling to False will remove
+    the converters, restoring any converters that pandas overwrote.
+"""
+
+
+def register_converter_cb(key) -> None:
+    from pandas.plotting import (
+        deregister_matplotlib_converters,
+        register_matplotlib_converters,
+    )
+
+    if cf.get_option(key):
+        register_matplotlib_converters()
+    else:
+        deregister_matplotlib_converters()
+
+
+with cf.config_prefix("plotting.matplotlib"):
+    cf.register_option(
+        "register_converters",
+        "auto",
+        register_converter_doc,
+        validator=is_one_of_factory(["auto", True, False]),
+        cb=register_converter_cb,
+    )
+
+# ------
+# Styler
+# ------
+
+styler_sparse_index_doc = """
+: bool
+    Whether to sparsify the display of a hierarchical index. Setting to False will
+    display each explicit level element in a hierarchical key for each row.
+"""
+
+styler_sparse_columns_doc = """
+: bool
+    Whether to sparsify the display of hierarchical columns. Setting to False will
+    display each explicit level element in a hierarchical key for each column.
+"""
+
+styler_render_repr = """
+: str
+    Determine which output to use in Jupyter Notebook in {"html", "latex"}.
+"""
+
+styler_max_elements = """
+: int
+    The maximum number of data-cell (<td>) elements that will be rendered before
+    trimming will occur over columns, rows or both if needed.
+"""
+
+styler_max_rows = """
+: int, optional
+    The maximum number of rows that will be rendered. May still be reduced to
+    satisfy ``max_elements``, which takes precedence.
+"""
+
+styler_max_columns = """
+: int, optional
+    The maximum number of columns that will be rendered. May still be reduced to
+    satisfy ``max_elements``, which takes precedence.
+"""
+
+styler_precision = """
+: int
+    The precision for floats and complex numbers.
+"""
+
+styler_decimal = """
+: str
+    The character representation for the decimal separator for floats and complex.
+"""
+
+styler_thousands = """
+: str, optional
+    The character representation for thousands separator for floats, int and complex.
+"""
+
+styler_na_rep = """
+: str, optional
+    The string representation for values identified as missing.
+"""
+
+styler_escape = """
+: str, optional
+    Whether to escape certain characters according to the given context; html or latex.
+"""
+
+styler_formatter = """
+: str, callable, dict, optional
+    A formatter object to be used as default within ``Styler.format``.
+"""
+
+styler_multirow_align = """
+: {"c", "t", "b"}
+    The specifier for vertical alignment of sparsified LaTeX multirows.
+"""
+
+styler_multicol_align = r"""
+: {"r", "c", "l", "naive-l", "naive-r"}
+    The specifier for horizontal alignment of sparsified LaTeX multicolumns. Pipe
+    decorators can also be added to non-naive values to draw vertical
+    rules, e.g. "\|r" will draw a rule on the left side of right aligned merged cells.
+"""
+
+styler_hrules = """
+: bool
+    Whether to add horizontal rules on top and bottom and below the headers.
+"""
+
+styler_environment = """
+: str
+    The environment to replace ``\\begin{table}``. If "longtable" is used results
+    in a specific longtable environment format.
+"""
+
+styler_encoding = """
+: str
+    The encoding used for output HTML and LaTeX files.
+"""
+
+styler_mathjax = """
+: bool
+    If False will render special CSS classes to table attributes that indicate Mathjax
+    will not be used in Jupyter Notebook.
+"""
+
+with cf.config_prefix("styler"):
+    cf.register_option("sparse.index", True, styler_sparse_index_doc, validator=is_bool)
+
+    cf.register_option(
+        "sparse.columns", True, styler_sparse_columns_doc, validator=is_bool
+    )
+
+    cf.register_option(
+        "render.repr",
+        "html",
+        styler_render_repr,
+        validator=is_one_of_factory(["html", "latex"]),
+    )
+
+    cf.register_option(
+        "render.max_elements",
+        2**18,
+        styler_max_elements,
+        validator=is_nonnegative_int,
+    )
+
+    cf.register_option(
+        "render.max_rows",
+        None,
+        styler_max_rows,
+        validator=is_nonnegative_int,
+    )
+
+    cf.register_option(
+        "render.max_columns",
+        None,
+        styler_max_columns,
+        validator=is_nonnegative_int,
+    )
+
+    cf.register_option("render.encoding", "utf-8", styler_encoding, validator=is_str)
+
+    cf.register_option("format.decimal", ".", styler_decimal, validator=is_str)
+
+    cf.register_option(
+        "format.precision", 6, styler_precision, validator=is_nonnegative_int
+    )
+
+    cf.register_option(
+        "format.thousands",
+        None,
+        styler_thousands,
+        validator=is_instance_factory([type(None), str]),
+    )
+
+    cf.register_option(
+        "format.na_rep",
+        None,
+        styler_na_rep,
+        validator=is_instance_factory([type(None), str]),
+    )
+
+    cf.register_option(
+        "format.escape",
+        None,
+        styler_escape,
+        validator=is_one_of_factory([None, "html", "latex", "latex-math"]),
+    )
+
+    cf.register_option(
+        "format.formatter",
+        None,
+        styler_formatter,
+        validator=is_instance_factory([type(None), dict, Callable, str]),
+    )
+
+    cf.register_option("html.mathjax", True, styler_mathjax, validator=is_bool)
+
+    cf.register_option(
+        "latex.multirow_align",
+        "c",
+        styler_multirow_align,
+        validator=is_one_of_factory(["c", "t", "b", "naive"]),
+    )
+
+    val_mca = ["r", "|r|", "|r", "r|", "c", "|c|", "|c", "c|", "l", "|l|", "|l", "l|"]
+    val_mca += ["naive-l", "naive-r"]
+    cf.register_option(
+        "latex.multicol_align",
+        "r",
+        styler_multicol_align,
+        validator=is_one_of_factory(val_mca),
+    )
+
+    cf.register_option("latex.hrules", False, styler_hrules, validator=is_bool)
+
+    cf.register_option(
+        "latex.environment",
+        None,
+        styler_environment,
+        validator=is_instance_factory([type(None), str]),
+    )
+
+
+with cf.config_prefix("future"):
+    cf.register_option(
+        "infer_string",
+        False,
+        "Whether to infer sequence of str objects as pyarrow string "
+        "dtype, which will be the default in pandas 3.0 "
+        "(at which point this option will be deprecated).",
+        validator=is_one_of_factory([True, False]),
+    )
+
+    cf.register_option(
+        "no_silent_downcasting",
+        False,
+        "Whether to opt-in to the future behavior which will *not* silently "
+        "downcast results from Series and DataFrame `where`, `mask`, and `clip` "
+        "methods. "
+        "Silent downcasting will be removed in pandas 3.0 "
+        "(at which point this option will be deprecated).",
+        validator=is_one_of_factory([True, False]),
+    )

llava_next/lib/python3.10/site-packages/pandas/core/construction.py

@@ -0,0 +1,824 @@
+"""
+Constructor functions intended to be shared by pd.array, Series.__init__,
+and Index.__new__.
+
+These should not depend on core.internals.
+"""
+from __future__ import annotations
+
+from collections.abc import Sequence
+from typing import (
+    TYPE_CHECKING,
+    Optional,
+    Union,
+    cast,
+    overload,
+)
+import warnings
+
+import numpy as np
+from numpy import ma
+
+from pandas._config import using_pyarrow_string_dtype
+
+from pandas._libs import lib
+from pandas._libs.tslibs import (
+    Period,
+    get_supported_dtype,
+    is_supported_dtype,
+)
+from pandas._typing import (
+    AnyArrayLike,
+    ArrayLike,
+    Dtype,
+    DtypeObj,
+    T,
+)
+from pandas.util._exceptions import find_stack_level
+
+from pandas.core.dtypes.base import ExtensionDtype
+from pandas.core.dtypes.cast import (
+    construct_1d_arraylike_from_scalar,
+    construct_1d_object_array_from_listlike,
+    maybe_cast_to_datetime,
+    maybe_cast_to_integer_array,
+    maybe_convert_platform,
+    maybe_infer_to_datetimelike,
+    maybe_promote,
+)
+from pandas.core.dtypes.common import (
+    is_list_like,
+    is_object_dtype,
+    is_string_dtype,
+    pandas_dtype,
+)
+from pandas.core.dtypes.dtypes import NumpyEADtype
+from pandas.core.dtypes.generic import (
+    ABCDataFrame,
+    ABCExtensionArray,
+    ABCIndex,
+    ABCSeries,
+)
+from pandas.core.dtypes.missing import isna
+
+import pandas.core.common as com
+
+if TYPE_CHECKING:
+    from pandas import (
+        Index,
+        Series,
+    )
+    from pandas.core.arrays.base import ExtensionArray
+
+
+def array(
+    data: Sequence[object] | AnyArrayLike,
+    dtype: Dtype | None = None,
+    copy: bool = True,
+) -> ExtensionArray:
+    """
+    Create an array.
+
+    Parameters
+    ----------
+    data : Sequence of objects
+        The scalars inside `data` should be instances of the
+        scalar type for `dtype`. It's expected that `data`
+        represents a 1-dimensional array of data.
+
+        When `data` is an Index or Series, the underlying array
+        will be extracted from `data`.
+
+    dtype : str, np.dtype, or ExtensionDtype, optional
+        The dtype to use for the array. This may be a NumPy
+        dtype or an extension type registered with pandas using
+        :meth:`pandas.api.extensions.register_extension_dtype`.
+
+        If not specified, there are two possibilities:
+
+        1. When `data` is a :class:`Series`, :class:`Index`, or
+           :class:`ExtensionArray`, the `dtype` will be taken
+           from the data.
+        2. Otherwise, pandas will attempt to infer the `dtype`
+           from the data.
+
+        Note that when `data` is a NumPy array, ``data.dtype`` is
+        *not* used for inferring the array type. This is because
+        NumPy cannot represent all the types of data that can be
+        held in extension arrays.
+
+        Currently, pandas will infer an extension dtype for sequences of
+
+        ============================== =======================================
+        Scalar Type                    Array Type
+        ============================== =======================================
+        :class:`pandas.Interval`       :class:`pandas.arrays.IntervalArray`
+        :class:`pandas.Period`         :class:`pandas.arrays.PeriodArray`
+        :class:`datetime.datetime`     :class:`pandas.arrays.DatetimeArray`
+        :class:`datetime.timedelta`    :class:`pandas.arrays.TimedeltaArray`
+        :class:`int`                   :class:`pandas.arrays.IntegerArray`
+        :class:`float`                 :class:`pandas.arrays.FloatingArray`
+        :class:`str`                   :class:`pandas.arrays.StringArray` or
+                                       :class:`pandas.arrays.ArrowStringArray`
+        :class:`bool`                  :class:`pandas.arrays.BooleanArray`
+        ============================== =======================================
+
+        The ExtensionArray created when the scalar type is :class:`str` is determined by
+        ``pd.options.mode.string_storage`` if the dtype is not explicitly given.
+
+        For all other cases, NumPy's usual inference rules will be used.
+    copy : bool, default True
+        Whether to copy the data, even if not necessary. Depending
+        on the type of `data`, creating the new array may require
+        copying data, even if ``copy=False``.
+
+    Returns
+    -------
+    ExtensionArray
+        The newly created array.
+
+    Raises
+    ------
+    ValueError
+        When `data` is not 1-dimensional.
+
+    See Also
+    --------
+    numpy.array : Construct a NumPy array.
+    Series : Construct a pandas Series.
+    Index : Construct a pandas Index.
+    arrays.NumpyExtensionArray : ExtensionArray wrapping a NumPy array.
+    Series.array : Extract the array stored within a Series.
+
+    Notes
+    -----
+    Omitting the `dtype` argument means pandas will attempt to infer the
+    best array type from the values in the data. As new array types are
+    added by pandas and 3rd party libraries, the "best" array type may
+    change. We recommend specifying `dtype` to ensure that
+
+    1. the correct array type for the data is returned
+    2. the returned array type doesn't change as new extension types
+       are added by pandas and third-party libraries
+
+    Additionally, if the underlying memory representation of the returned
+    array matters, we recommend specifying the `dtype` as a concrete object
+    rather than a string alias or allowing it to be inferred. For example,
+    a future version of pandas or a 3rd-party library may include a
+    dedicated ExtensionArray for string data. In this event, the following
+    would no longer return a :class:`arrays.NumpyExtensionArray` backed by a
+    NumPy array.
+
+    >>> pd.array(['a', 'b'], dtype=str)
+    <NumpyExtensionArray>
+    ['a', 'b']
+    Length: 2, dtype: str32
+
+    This would instead return the new ExtensionArray dedicated for string
+    data. If you really need the new array to be backed by a  NumPy array,
+    specify that in the dtype.
+
+    >>> pd.array(['a', 'b'], dtype=np.dtype("<U1"))
+    <NumpyExtensionArray>
+    ['a', 'b']
+    Length: 2, dtype: str32
+
+    Finally, Pandas has arrays that mostly overlap with NumPy
+
+      * :class:`arrays.DatetimeArray`
+      * :class:`arrays.TimedeltaArray`
+
+    When data with a ``datetime64[ns]`` or ``timedelta64[ns]`` dtype is
+    passed, pandas will always return a ``DatetimeArray`` or ``TimedeltaArray``
+    rather than a ``NumpyExtensionArray``. This is for symmetry with the case of
+    timezone-aware data, which NumPy does not natively support.
+
+    >>> pd.array(['2015', '2016'], dtype='datetime64[ns]')
+    <DatetimeArray>
+    ['2015-01-01 00:00:00', '2016-01-01 00:00:00']
+    Length: 2, dtype: datetime64[ns]
+
+    >>> pd.array(["1h", "2h"], dtype='timedelta64[ns]')
+    <TimedeltaArray>
+    ['0 days 01:00:00', '0 days 02:00:00']
+    Length: 2, dtype: timedelta64[ns]
+
+    Examples
+    --------
+    If a dtype is not specified, pandas will infer the best dtype from the values.
+    See the description of `dtype` for the types pandas infers for.
+
+    >>> pd.array([1, 2])
+    <IntegerArray>
+    [1, 2]
+    Length: 2, dtype: Int64
+
+    >>> pd.array([1, 2, np.nan])
+    <IntegerArray>
+    [1, 2, <NA>]
+    Length: 3, dtype: Int64
+
+    >>> pd.array([1.1, 2.2])
+    <FloatingArray>
+    [1.1, 2.2]
+    Length: 2, dtype: Float64
+
+    >>> pd.array(["a", None, "c"])
+    <StringArray>
+    ['a', <NA>, 'c']
+    Length: 3, dtype: string
+
+    >>> with pd.option_context("string_storage", "pyarrow"):
+    ...     arr = pd.array(["a", None, "c"])
+    ...
+    >>> arr
+    <ArrowStringArray>
+    ['a', <NA>, 'c']
+    Length: 3, dtype: string
+
+    >>> pd.array([pd.Period('2000', freq="D"), pd.Period("2000", freq="D")])
+    <PeriodArray>
+    ['2000-01-01', '2000-01-01']
+    Length: 2, dtype: period[D]
+
+    You can use the string alias for `dtype`
+
+    >>> pd.array(['a', 'b', 'a'], dtype='category')
+    ['a', 'b', 'a']
+    Categories (2, object): ['a', 'b']
+
+    Or specify the actual dtype
+
+    >>> pd.array(['a', 'b', 'a'],
+    ...          dtype=pd.CategoricalDtype(['a', 'b', 'c'], ordered=True))
+    ['a', 'b', 'a']
+    Categories (3, object): ['a' < 'b' < 'c']
+
+    If pandas does not infer a dedicated extension type a
+    :class:`arrays.NumpyExtensionArray` is returned.
+
+    >>> pd.array([1 + 1j, 3 + 2j])
+    <NumpyExtensionArray>
+    [(1+1j), (3+2j)]
+    Length: 2, dtype: complex128
+
+    As mentioned in the "Notes" section, new extension types may be added
+    in the future (by pandas or 3rd party libraries), causing the return
+    value to no longer be a :class:`arrays.NumpyExtensionArray`. Specify the
+    `dtype` as a NumPy dtype if you need to ensure there's no future change in
+    behavior.
+
+    >>> pd.array([1, 2], dtype=np.dtype("int32"))
+    <NumpyExtensionArray>
+    [1, 2]
+    Length: 2, dtype: int32
+
+    `data` must be 1-dimensional. A ValueError is raised when the input
+    has the wrong dimensionality.
+
+    >>> pd.array(1)
+    Traceback (most recent call last):
+      ...
+    ValueError: Cannot pass scalar '1' to 'pandas.array'.
+    """
+    from pandas.core.arrays import (
+        BooleanArray,
+        DatetimeArray,
+        ExtensionArray,
+        FloatingArray,
+        IntegerArray,
+        IntervalArray,
+        NumpyExtensionArray,
+        PeriodArray,
+        TimedeltaArray,
+    )
+    from pandas.core.arrays.string_ import StringDtype
+
+    if lib.is_scalar(data):
+        msg = f"Cannot pass scalar '{data}' to 'pandas.array'."
+        raise ValueError(msg)
+    elif isinstance(data, ABCDataFrame):
+        raise TypeError("Cannot pass DataFrame to 'pandas.array'")
+
+    if dtype is None and isinstance(data, (ABCSeries, ABCIndex, ExtensionArray)):
+        # Note: we exclude np.ndarray here, will do type inference on it
+        dtype = data.dtype
+
+    data = extract_array(data, extract_numpy=True)
+
+    # this returns None for not-found dtypes.
+    if dtype is not None:
+        dtype = pandas_dtype(dtype)
+
+    if isinstance(data, ExtensionArray) and (dtype is None or data.dtype == dtype):
+        # e.g. TimedeltaArray[s], avoid casting to NumpyExtensionArray
+        if copy:
+            return data.copy()
+        return data
+
+    if isinstance(dtype, ExtensionDtype):
+        cls = dtype.construct_array_type()
+        return cls._from_sequence(data, dtype=dtype, copy=copy)
+
+    if dtype is None:
+        inferred_dtype = lib.infer_dtype(data, skipna=True)
+        if inferred_dtype == "period":
+            period_data = cast(Union[Sequence[Optional[Period]], AnyArrayLike], data)
+            return PeriodArray._from_sequence(period_data, copy=copy)
+
+        elif inferred_dtype == "interval":
+            return IntervalArray(data, copy=copy)
+
+        elif inferred_dtype.startswith("datetime"):
+            # datetime, datetime64
+            try:
+                return DatetimeArray._from_sequence(data, copy=copy)
+            except ValueError:
+                # Mixture of timezones, fall back to NumpyExtensionArray
+                pass
+
+        elif inferred_dtype.startswith("timedelta"):
+            # timedelta, timedelta64
+            return TimedeltaArray._from_sequence(data, copy=copy)
+
+        elif inferred_dtype == "string":
+            # StringArray/ArrowStringArray depending on pd.options.mode.string_storage
+            dtype = StringDtype()
+            cls = dtype.construct_array_type()
+            return cls._from_sequence(data, dtype=dtype, copy=copy)
+
+        elif inferred_dtype == "integer":
+            return IntegerArray._from_sequence(data, copy=copy)
+        elif inferred_dtype == "empty" and not hasattr(data, "dtype") and not len(data):
+            return FloatingArray._from_sequence(data, copy=copy)
+        elif (
+            inferred_dtype in ("floating", "mixed-integer-float")
+            and getattr(data, "dtype", None) != np.float16
+        ):
+            # GH#44715 Exclude np.float16 bc FloatingArray does not support it;
+            #  we will fall back to NumpyExtensionArray.
+            return FloatingArray._from_sequence(data, copy=copy)
+
+        elif inferred_dtype == "boolean":
+            return BooleanArray._from_sequence(data, dtype="boolean", copy=copy)
+
+    # Pandas overrides NumPy for
+    #   1. datetime64[ns,us,ms,s]
+    #   2. timedelta64[ns,us,ms,s]
+    # so that a DatetimeArray is returned.
+    if lib.is_np_dtype(dtype, "M") and is_supported_dtype(dtype):
+        return DatetimeArray._from_sequence(data, dtype=dtype, copy=copy)
+    if lib.is_np_dtype(dtype, "m") and is_supported_dtype(dtype):
+        return TimedeltaArray._from_sequence(data, dtype=dtype, copy=copy)
+
+    elif lib.is_np_dtype(dtype, "mM"):
+        warnings.warn(
+            r"datetime64 and timedelta64 dtype resolutions other than "
+            r"'s', 'ms', 'us', and 'ns' are deprecated. "
+            r"In future releases passing unsupported resolutions will "
+            r"raise an exception.",
+            FutureWarning,
+            stacklevel=find_stack_level(),
+        )
+
+    return NumpyExtensionArray._from_sequence(data, dtype=dtype, copy=copy)
+
+
+_typs = frozenset(
+    {
+        "index",
+        "rangeindex",
+        "multiindex",
+        "datetimeindex",
+        "timedeltaindex",
+        "periodindex",
+        "categoricalindex",
+        "intervalindex",
+        "series",
+    }
+)
+
+
+@overload
+def extract_array(
+    obj: Series | Index, extract_numpy: bool = ..., extract_range: bool = ...
+) -> ArrayLike:
+    ...
+
+
+@overload
+def extract_array(
+    obj: T, extract_numpy: bool = ..., extract_range: bool = ...
+) -> T | ArrayLike:
+    ...
+
+
+def extract_array(
+    obj: T, extract_numpy: bool = False, extract_range: bool = False
+) -> T | ArrayLike:
+    """
+    Extract the ndarray or ExtensionArray from a Series or Index.
+
+    For all other types, `obj` is just returned as is.
+
+    Parameters
+    ----------
+    obj : object
+        For Series / Index, the underlying ExtensionArray is unboxed.
+
+    extract_numpy : bool, default False
+        Whether to extract the ndarray from a NumpyExtensionArray.
+
+    extract_range : bool, default False
+        If we have a RangeIndex, return range._values if True
+        (which is a materialized integer ndarray), otherwise return unchanged.
+
+    Returns
+    -------
+    arr : object
+
+    Examples
+    --------
+    >>> extract_array(pd.Series(['a', 'b', 'c'], dtype='category'))
+    ['a', 'b', 'c']
+    Categories (3, object): ['a', 'b', 'c']
+
+    Other objects like lists, arrays, and DataFrames are just passed through.
+
+    >>> extract_array([1, 2, 3])
+    [1, 2, 3]
+
+    For an ndarray-backed Series / Index the ndarray is returned.
+
+    >>> extract_array(pd.Series([1, 2, 3]))
+    array([1, 2, 3])
+
+    To extract all the way down to the ndarray, pass ``extract_numpy=True``.
+
+    >>> extract_array(pd.Series([1, 2, 3]), extract_numpy=True)
+    array([1, 2, 3])
+    """
+    typ = getattr(obj, "_typ", None)
+    if typ in _typs:
+        # i.e. isinstance(obj, (ABCIndex, ABCSeries))
+        if typ == "rangeindex":
+            if extract_range:
+                # error: "T" has no attribute "_values"
+                return obj._values  # type: ignore[attr-defined]
+            return obj
+
+        # error: "T" has no attribute "_values"
+        return obj._values  # type: ignore[attr-defined]
+
+    elif extract_numpy and typ == "npy_extension":
+        # i.e. isinstance(obj, ABCNumpyExtensionArray)
+        # error: "T" has no attribute "to_numpy"
+        return obj.to_numpy()  # type: ignore[attr-defined]
+
+    return obj
+
+
+def ensure_wrapped_if_datetimelike(arr):
+    """
+    Wrap datetime64 and timedelta64 ndarrays in DatetimeArray/TimedeltaArray.
+    """
+    if isinstance(arr, np.ndarray):
+        if arr.dtype.kind == "M":
+            from pandas.core.arrays import DatetimeArray
+
+            dtype = get_supported_dtype(arr.dtype)
+            return DatetimeArray._from_sequence(arr, dtype=dtype)
+
+        elif arr.dtype.kind == "m":
+            from pandas.core.arrays import TimedeltaArray
+
+            dtype = get_supported_dtype(arr.dtype)
+            return TimedeltaArray._from_sequence(arr, dtype=dtype)
+
+    return arr
+
+
+def sanitize_masked_array(data: ma.MaskedArray) -> np.ndarray:
+    """
+    Convert numpy MaskedArray to ensure mask is softened.
+    """
+    mask = ma.getmaskarray(data)
+    if mask.any():
+        dtype, fill_value = maybe_promote(data.dtype, np.nan)
+        dtype = cast(np.dtype, dtype)
+        data = ma.asarray(data.astype(dtype, copy=True))
+        data.soften_mask()  # set hardmask False if it was True
+        data[mask] = fill_value
+    else:
+        data = data.copy()
+    return data
+
+
+def sanitize_array(
+    data,
+    index: Index | None,
+    dtype: DtypeObj | None = None,
+    copy: bool = False,
+    *,
+    allow_2d: bool = False,
+) -> ArrayLike:
+    """
+    Sanitize input data to an ndarray or ExtensionArray, copy if specified,
+    coerce to the dtype if specified.
+
+    Parameters
+    ----------
+    data : Any
+    index : Index or None, default None
+    dtype : np.dtype, ExtensionDtype, or None, default None
+    copy : bool, default False
+    allow_2d : bool, default False
+        If False, raise if we have a 2D Arraylike.
+
+    Returns
+    -------
+    np.ndarray or ExtensionArray
+    """
+    original_dtype = dtype
+    if isinstance(data, ma.MaskedArray):
+        data = sanitize_masked_array(data)
+
+    if isinstance(dtype, NumpyEADtype):
+        # Avoid ending up with a NumpyExtensionArray
+        dtype = dtype.numpy_dtype
+
+    object_index = False
+    if isinstance(data, ABCIndex) and data.dtype == object and dtype is None:
+        object_index = True
+
+    # extract ndarray or ExtensionArray, ensure we have no NumpyExtensionArray
+    data = extract_array(data, extract_numpy=True, extract_range=True)
+
+    if isinstance(data, np.ndarray) and data.ndim == 0:
+        if dtype is None:
+            dtype = data.dtype
+        data = lib.item_from_zerodim(data)
+    elif isinstance(data, range):
+        # GH#16804
+        data = range_to_ndarray(data)
+        copy = False
+
+    if not is_list_like(data):
+        if index is None:
+            raise ValueError("index must be specified when data is not list-like")
+        if (
+            isinstance(data, str)
+            and using_pyarrow_string_dtype()
+            and original_dtype is None
+        ):
+            from pandas.core.arrays.string_ import StringDtype
+
+            dtype = StringDtype("pyarrow_numpy")
+        data = construct_1d_arraylike_from_scalar(data, len(index), dtype)
+
+        return data
+
+    elif isinstance(data, ABCExtensionArray):
+        # it is already ensured above this is not a NumpyExtensionArray
+        # Until GH#49309 is fixed this check needs to come before the
+        #  ExtensionDtype check
+        if dtype is not None:
+            subarr = data.astype(dtype, copy=copy)
+        elif copy:
+            subarr = data.copy()
+        else:
+            subarr = data
+
+    elif isinstance(dtype, ExtensionDtype):
+        # create an extension array from its dtype
+        _sanitize_non_ordered(data)
+        cls = dtype.construct_array_type()
+        subarr = cls._from_sequence(data, dtype=dtype, copy=copy)
+
+    # GH#846
+    elif isinstance(data, np.ndarray):
+        if isinstance(data, np.matrix):
+            data = data.A
+
+        if dtype is None:
+            subarr = data
+            if data.dtype == object:
+                subarr = maybe_infer_to_datetimelike(data)
+                if (
+                    object_index
+                    and using_pyarrow_string_dtype()
+                    and is_string_dtype(subarr)
+                ):
+                    # Avoid inference when string option is set
+                    subarr = data
+            elif data.dtype.kind == "U" and using_pyarrow_string_dtype():
+                from pandas.core.arrays.string_ import StringDtype
+
+                dtype = StringDtype(storage="pyarrow_numpy")
+                subarr = dtype.construct_array_type()._from_sequence(data, dtype=dtype)
+
+            if subarr is data and copy:
+                subarr = subarr.copy()
+
+        else:
+            # we will try to copy by-definition here
+            subarr = _try_cast(data, dtype, copy)
+
+    elif hasattr(data, "__array__"):
+        # e.g. dask array GH#38645
+        if not copy:
+            data = np.asarray(data)
+        else:
+            data = np.array(data, copy=copy)
+        return sanitize_array(
+            data,
+            index=index,
+            dtype=dtype,
+            copy=False,
+            allow_2d=allow_2d,
+        )
+
+    else:
+        _sanitize_non_ordered(data)
+        # materialize e.g. generators, convert e.g. tuples, abc.ValueView
+        data = list(data)
+
+        if len(data) == 0 and dtype is None:
+            # We default to float64, matching numpy
+            subarr = np.array([], dtype=np.float64)
+
+        elif dtype is not None:
+            subarr = _try_cast(data, dtype, copy)
+
+        else:
+            subarr = maybe_convert_platform(data)
+            if subarr.dtype == object:
+                subarr = cast(np.ndarray, subarr)
+                subarr = maybe_infer_to_datetimelike(subarr)
+
+    subarr = _sanitize_ndim(subarr, data, dtype, index, allow_2d=allow_2d)
+
+    if isinstance(subarr, np.ndarray):
+        # at this point we should have dtype be None or subarr.dtype == dtype
+        dtype = cast(np.dtype, dtype)
+        subarr = _sanitize_str_dtypes(subarr, data, dtype, copy)
+
+    return subarr
+
+
+def range_to_ndarray(rng: range) -> np.ndarray:
+    """
+    Cast a range object to ndarray.
+    """
+    # GH#30171 perf avoid realizing range as a list in np.array
+    try:
+        arr = np.arange(rng.start, rng.stop, rng.step, dtype="int64")
+    except OverflowError:
+        # GH#30173 handling for ranges that overflow int64
+        if (rng.start >= 0 and rng.step > 0) or (rng.step < 0 <= rng.stop):
+            try:
+                arr = np.arange(rng.start, rng.stop, rng.step, dtype="uint64")
+            except OverflowError:
+                arr = construct_1d_object_array_from_listlike(list(rng))
+        else:
+            arr = construct_1d_object_array_from_listlike(list(rng))
+    return arr
+
+
+def _sanitize_non_ordered(data) -> None:
+    """
+    Raise only for unordered sets, e.g., not for dict_keys
+    """
+    if isinstance(data, (set, frozenset)):
+        raise TypeError(f"'{type(data).__name__}' type is unordered")
+
+
+def _sanitize_ndim(
+    result: ArrayLike,
+    data,
+    dtype: DtypeObj | None,
+    index: Index | None,
+    *,
+    allow_2d: bool = False,
+) -> ArrayLike:
+    """
+    Ensure we have a 1-dimensional result array.
+    """
+    if getattr(result, "ndim", 0) == 0:
+        raise ValueError("result should be arraylike with ndim > 0")
+
+    if result.ndim == 1:
+        # the result that we want
+        result = _maybe_repeat(result, index)
+
+    elif result.ndim > 1:
+        if isinstance(data, np.ndarray):
+            if allow_2d:
+                return result
+            raise ValueError(
+                f"Data must be 1-dimensional, got ndarray of shape {data.shape} instead"
+            )
+        if is_object_dtype(dtype) and isinstance(dtype, ExtensionDtype):
+            # i.e. NumpyEADtype("O")
+
+            result = com.asarray_tuplesafe(data, dtype=np.dtype("object"))
+            cls = dtype.construct_array_type()
+            result = cls._from_sequence(result, dtype=dtype)
+        else:
+            # error: Argument "dtype" to "asarray_tuplesafe" has incompatible type
+            # "Union[dtype[Any], ExtensionDtype, None]"; expected "Union[str,
+            # dtype[Any], None]"
+            result = com.asarray_tuplesafe(data, dtype=dtype)  # type: ignore[arg-type]
+    return result
+
+
+def _sanitize_str_dtypes(
+    result: np.ndarray, data, dtype: np.dtype | None, copy: bool
+) -> np.ndarray:
+    """
+    Ensure we have a dtype that is supported by pandas.
+    """
+
+    # This is to prevent mixed-type Series getting all casted to
+    # NumPy string type, e.g. NaN --> '-1#IND'.
+    if issubclass(result.dtype.type, str):
+        # GH#16605
+        # If not empty convert the data to dtype
+        # GH#19853: If data is a scalar, result has already the result
+        if not lib.is_scalar(data):
+            if not np.all(isna(data)):
+                data = np.asarray(data, dtype=dtype)
+            if not copy:
+                result = np.asarray(data, dtype=object)
+            else:
+                result = np.array(data, dtype=object, copy=copy)
+    return result
+
+
+def _maybe_repeat(arr: ArrayLike, index: Index | None) -> ArrayLike:
+    """
+    If we have a length-1 array and an index describing how long we expect
+    the result to be, repeat the array.
+    """
+    if index is not None:
+        if 1 == len(arr) != len(index):
+            arr = arr.repeat(len(index))
+    return arr
+
+
+def _try_cast(
+    arr: list | np.ndarray,
+    dtype: np.dtype,
+    copy: bool,
+) -> ArrayLike:
+    """
+    Convert input to numpy ndarray and optionally cast to a given dtype.
+
+    Parameters
+    ----------
+    arr : ndarray or list
+        Excludes: ExtensionArray, Series, Index.
+    dtype : np.dtype
+    copy : bool
+        If False, don't copy the data if not needed.
+
+    Returns
+    -------
+    np.ndarray or ExtensionArray
+    """
+    is_ndarray = isinstance(arr, np.ndarray)
+
+    if dtype == object:
+        if not is_ndarray:
+            subarr = construct_1d_object_array_from_listlike(arr)
+            return subarr
+        return ensure_wrapped_if_datetimelike(arr).astype(dtype, copy=copy)
+
+    elif dtype.kind == "U":
+        # TODO: test cases with arr.dtype.kind in "mM"
+        if is_ndarray:
+            arr = cast(np.ndarray, arr)
+            shape = arr.shape
+            if arr.ndim > 1:
+                arr = arr.ravel()
+        else:
+            shape = (len(arr),)
+        return lib.ensure_string_array(arr, convert_na_value=False, copy=copy).reshape(
+            shape
+        )
+
+    elif dtype.kind in "mM":
+        return maybe_cast_to_datetime(arr, dtype)
+
+    # GH#15832: Check if we are requesting a numeric dtype and
+    # that we can convert the data to the requested dtype.
+    elif dtype.kind in "iu":
+        # this will raise if we have e.g. floats
+
+        subarr = maybe_cast_to_integer_array(arr, dtype)
+    elif not copy:
+        subarr = np.asarray(arr, dtype=dtype)
+    else:
+        subarr = np.array(arr, dtype=dtype, copy=copy)
+
+    return subarr

llava_next/lib/python3.10/site-packages/pandas/core/flags.py

@@ -0,0 +1,117 @@
+from __future__ import annotations
+
+from typing import TYPE_CHECKING
+import weakref
+
+if TYPE_CHECKING:
+    from pandas.core.generic import NDFrame
+
+
+class Flags:
+    """
+    Flags that apply to pandas objects.
+
+    Parameters
+    ----------
+    obj : Series or DataFrame
+        The object these flags are associated with.
+    allows_duplicate_labels : bool, default True
+        Whether to allow duplicate labels in this object. By default,
+        duplicate labels are permitted. Setting this to ``False`` will
+        cause an :class:`errors.DuplicateLabelError` to be raised when
+        `index` (or columns for DataFrame) is not unique, or any
+        subsequent operation on introduces duplicates.
+        See :ref:`duplicates.disallow` for more.
+
+        .. warning::
+
+           This is an experimental feature. Currently, many methods fail to
+           propagate the ``allows_duplicate_labels`` value. In future versions
+           it is expected that every method taking or returning one or more
+           DataFrame or Series objects will propagate ``allows_duplicate_labels``.
+
+    Examples
+    --------
+    Attributes can be set in two ways:
+
+    >>> df = pd.DataFrame()
+    >>> df.flags
+    <Flags(allows_duplicate_labels=True)>
+    >>> df.flags.allows_duplicate_labels = False
+    >>> df.flags
+    <Flags(allows_duplicate_labels=False)>
+
+    >>> df.flags['allows_duplicate_labels'] = True
+    >>> df.flags
+    <Flags(allows_duplicate_labels=True)>
+    """
+
+    _keys: set[str] = {"allows_duplicate_labels"}
+
+    def __init__(self, obj: NDFrame, *, allows_duplicate_labels: bool) -> None:
+        self._allows_duplicate_labels = allows_duplicate_labels
+        self._obj = weakref.ref(obj)
+
+    @property
+    def allows_duplicate_labels(self) -> bool:
+        """
+        Whether this object allows duplicate labels.
+
+        Setting ``allows_duplicate_labels=False`` ensures that the
+        index (and columns of a DataFrame) are unique. Most methods
+        that accept and return a Series or DataFrame will propagate
+        the value of ``allows_duplicate_labels``.
+
+        See :ref:`duplicates` for more.
+
+        See Also
+        --------
+        DataFrame.attrs : Set global metadata on this object.
+        DataFrame.set_flags : Set global flags on this object.
+
+        Examples
+        --------
+        >>> df = pd.DataFrame({"A": [1, 2]}, index=['a', 'a'])
+        >>> df.flags.allows_duplicate_labels
+        True
+        >>> df.flags.allows_duplicate_labels = False
+        Traceback (most recent call last):
+            ...
+        pandas.errors.DuplicateLabelError: Index has duplicates.
+              positions
+        label
+        a        [0, 1]
+        """
+        return self._allows_duplicate_labels
+
+    @allows_duplicate_labels.setter
+    def allows_duplicate_labels(self, value: bool) -> None:
+        value = bool(value)
+        obj = self._obj()
+        if obj is None:
+            raise ValueError("This flag's object has been deleted.")
+
+        if not value:
+            for ax in obj.axes:
+                ax._maybe_check_unique()
+
+        self._allows_duplicate_labels = value
+
+    def __getitem__(self, key: str):
+        if key not in self._keys:
+            raise KeyError(key)
+
+        return getattr(self, key)
+
+    def __setitem__(self, key: str, value) -> None:
+        if key not in self._keys:
+            raise ValueError(f"Unknown flag {key}. Must be one of {self._keys}")
+        setattr(self, key, value)
+
+    def __repr__(self) -> str:
+        return f"<Flags(allows_duplicate_labels={self.allows_duplicate_labels})>"
+
+    def __eq__(self, other) -> bool:
+        if isinstance(other, type(self)):
+            return self.allows_duplicate_labels == other.allows_duplicate_labels
+        return False

llava_next/lib/python3.10/site-packages/pandas/core/roperator.py

@@ -0,0 +1,62 @@
+"""
+Reversed Operations not available in the stdlib operator module.
+Defining these instead of using lambdas allows us to reference them by name.
+"""
+from __future__ import annotations
+
+import operator
+
+
+def radd(left, right):
+    return right + left
+
+
+def rsub(left, right):
+    return right - left
+
+
+def rmul(left, right):
+    return right * left
+
+
+def rdiv(left, right):
+    return right / left
+
+
+def rtruediv(left, right):
+    return right / left
+
+
+def rfloordiv(left, right):
+    return right // left
+
+
+def rmod(left, right):
+    # check if right is a string as % is the string
+    # formatting operation; this is a TypeError
+    # otherwise perform the op
+    if isinstance(right, str):
+        typ = type(left).__name__
+        raise TypeError(f"{typ} cannot perform the operation mod")
+
+    return right % left
+
+
+def rdivmod(left, right):
+    return divmod(right, left)
+
+
+def rpow(left, right):
+    return right**left
+
+
+def rand_(left, right):
+    return operator.and_(right, left)
+
+
+def ror_(left, right):
+    return operator.or_(right, left)
+
+
+def rxor(left, right):
+    return operator.xor(right, left)

parrot/lib/python3.10/site-packages/transformers/models/canine/__init__.py

@@ -0,0 +1,67 @@
+# Copyright 2020 The HuggingFace Team. All rights reserved.
+#
+# 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.
+from typing import TYPE_CHECKING
+
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_tokenizers_available, is_torch_available
+
+
+_import_structure = {
+    "configuration_canine": ["CanineConfig"],
+    "tokenization_canine": ["CanineTokenizer"],
+}
+
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_canine"] = [
+        "CanineForMultipleChoice",
+        "CanineForQuestionAnswering",
+        "CanineForSequenceClassification",
+        "CanineForTokenClassification",
+        "CanineLayer",
+        "CanineModel",
+        "CaninePreTrainedModel",
+        "load_tf_weights_in_canine",
+    ]
+
+
+if TYPE_CHECKING:
+    from .configuration_canine import CanineConfig
+    from .tokenization_canine import CanineTokenizer
+
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_canine import (
+            CanineForMultipleChoice,
+            CanineForQuestionAnswering,
+            CanineForSequenceClassification,
+            CanineForTokenClassification,
+            CanineLayer,
+            CanineModel,
+            CaninePreTrainedModel,
+            load_tf_weights_in_canine,
+        )
+
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)

parrot/lib/python3.10/site-packages/transformers/models/canine/convert_canine_original_tf_checkpoint_to_pytorch.py

@@ -0,0 +1,66 @@
+# coding=utf-8
+# Copyright 2021 The HuggingFace Inc. 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.
+"""Convert CANINE checkpoint."""
+
+
+import argparse
+
+from transformers import CanineConfig, CanineModel, CanineTokenizer, load_tf_weights_in_canine
+from transformers.utils import logging
+
+
+logging.set_verbosity_info()
+
+
+def convert_tf_checkpoint_to_pytorch(tf_checkpoint_path, pytorch_dump_path):
+    # Initialize PyTorch model
+    config = CanineConfig()
+    model = CanineModel(config)
+    model.eval()
+
+    print(f"Building PyTorch model from configuration: {config}")
+
+    # Load weights from tf checkpoint
+    load_tf_weights_in_canine(model, config, tf_checkpoint_path)
+
+    # Save pytorch-model (weights and configuration)
+    print(f"Save PyTorch model to {pytorch_dump_path}")
+    model.save_pretrained(pytorch_dump_path)
+
+    # Save tokenizer files
+    tokenizer = CanineTokenizer()
+    print(f"Save tokenizer files to {pytorch_dump_path}")
+    tokenizer.save_pretrained(pytorch_dump_path)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    # Required parameters
+    parser.add_argument(
+        "--tf_checkpoint_path",
+        default=None,
+        type=str,
+        required=True,
+        help="Path to the TensorFlow checkpoint. Should end with model.ckpt",
+    )
+    parser.add_argument(
+        "--pytorch_dump_path",
+        default=None,
+        type=str,
+        required=True,
+        help="Path to a folder where the PyTorch model will be placed.",
+    )
+    args = parser.parse_args()
+    convert_tf_checkpoint_to_pytorch(args.tf_checkpoint_path, args.pytorch_dump_path)

parrot/lib/python3.10/site-packages/transformers/models/canine/modeling_canine.py

@@ -0,0 +1,1642 @@
+# coding=utf-8
+# Copyright 2021 Google AI The HuggingFace Inc. team. All rights reserved.
+#
+# 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.
+""" PyTorch CANINE model."""
+
+
+import copy
+import math
+import os
+from dataclasses import dataclass
+from typing import Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
+
+from ...activations import ACT2FN
+from ...modeling_outputs import (
+    BaseModelOutput,
+    ModelOutput,
+    MultipleChoiceModelOutput,
+    QuestionAnsweringModelOutput,
+    SequenceClassifierOutput,
+    TokenClassifierOutput,
+)
+from ...modeling_utils import PreTrainedModel
+from ...pytorch_utils import apply_chunking_to_forward, find_pruneable_heads_and_indices, prune_linear_layer
+from ...utils import (
+    add_code_sample_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    logging,
+    replace_return_docstrings,
+)
+from .configuration_canine import CanineConfig
+
+
+logger = logging.get_logger(__name__)
+
+_CHECKPOINT_FOR_DOC = "google/canine-s"
+_CONFIG_FOR_DOC = "CanineConfig"
+
+
+# Support up to 16 hash functions.
+_PRIMES = [31, 43, 59, 61, 73, 97, 103, 113, 137, 149, 157, 173, 181, 193, 211, 223]
+
+
+@dataclass
+class CanineModelOutputWithPooling(ModelOutput):
+    """
+    Output type of [`CanineModel`]. Based on [`~modeling_outputs.BaseModelOutputWithPooling`], but with slightly
+    different `hidden_states` and `attentions`, as these also include the hidden states and attentions of the shallow
+    Transformer encoders.
+
+    Args:
+        last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
+            Sequence of hidden-states at the output of the last layer of the model (i.e. the output of the final
+            shallow Transformer encoder).
+        pooler_output (`torch.FloatTensor` of shape `(batch_size, hidden_size)`):
+            Hidden-state of the first token of the sequence (classification token) at the last layer of the deep
+            Transformer encoder, further processed by a Linear layer and a Tanh activation function. The Linear layer
+            weights are trained from the next sentence prediction (classification) objective during pretraining.
+        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `torch.FloatTensor` (one for the input to each encoder + one for the output of each layer of each
+            encoder) of shape `(batch_size, sequence_length, hidden_size)` and `(batch_size, sequence_length //
+            config.downsampling_rate, hidden_size)`. Hidden-states of the model at the output of each layer plus the
+            initial input to each Transformer encoder. The hidden states of the shallow encoders have length
+            `sequence_length`, but the hidden states of the deep encoder have length `sequence_length` //
+            `config.downsampling_rate`.
+        attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
+            Tuple of `torch.FloatTensor` (one for each layer) of the 3 Transformer encoders of shape `(batch_size,
+            num_heads, sequence_length, sequence_length)` and `(batch_size, num_heads, sequence_length //
+            config.downsampling_rate, sequence_length // config.downsampling_rate)`. Attentions weights after the
+            attention softmax, used to compute the weighted average in the self-attention heads.
+    """
+
+    last_hidden_state: torch.FloatTensor = None
+    pooler_output: torch.FloatTensor = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+
+def load_tf_weights_in_canine(model, config, tf_checkpoint_path):
+    """Load tf checkpoints in a pytorch model."""
+    try:
+        import re
+
+        import numpy as np
+        import tensorflow as tf
+    except ImportError:
+        logger.error(
+            "Loading a TensorFlow model in PyTorch, requires TensorFlow to be installed. Please see "
+            "https://www.tensorflow.org/install/ for installation instructions."
+        )
+        raise
+    tf_path = os.path.abspath(tf_checkpoint_path)
+    logger.info(f"Converting TensorFlow checkpoint from {tf_path}")
+    # Load weights from TF model
+    init_vars = tf.train.list_variables(tf_path)
+    names = []
+    arrays = []
+    for name, shape in init_vars:
+        logger.info(f"Loading TF weight {name} with shape {shape}")
+        array = tf.train.load_variable(tf_path, name)
+        names.append(name)
+        arrays.append(array)
+
+    for name, array in zip(names, arrays):
+        name = name.split("/")
+        # adam_v and adam_m are variables used in AdamWeightDecayOptimizer to calculated m and v
+        # which are not required for using pretrained model
+        # also discard the cls weights (which were used for the next sentence prediction pre-training task)
+        if any(
+            n
+            in [
+                "adam_v",
+                "adam_m",
+                "AdamWeightDecayOptimizer",
+                "AdamWeightDecayOptimizer_1",
+                "global_step",
+                "cls",
+                "autoregressive_decoder",
+                "char_output_weights",
+            ]
+            for n in name
+        ):
+            logger.info(f"Skipping {'/'.join(name)}")
+            continue
+        # if first scope name starts with "bert", change it to "encoder"
+        if name[0] == "bert":
+            name[0] = "encoder"
+        # remove "embeddings" middle name of HashBucketCodepointEmbedders
+        elif name[1] == "embeddings":
+            name.remove(name[1])
+        # rename segment_embeddings to token_type_embeddings
+        elif name[1] == "segment_embeddings":
+            name[1] = "token_type_embeddings"
+        # rename initial convolutional projection layer
+        elif name[1] == "initial_char_encoder":
+            name = ["chars_to_molecules"] + name[-2:]
+        # rename final convolutional projection layer
+        elif name[0] == "final_char_encoder" and name[1] in ["LayerNorm", "conv"]:
+            name = ["projection"] + name[1:]
+        pointer = model
+        for m_name in name:
+            if (re.fullmatch(r"[A-Za-z]+_\d+", m_name)) and "Embedder" not in m_name:
+                scope_names = re.split(r"_(\d+)", m_name)
+            else:
+                scope_names = [m_name]
+            if scope_names[0] == "kernel" or scope_names[0] == "gamma":
+                pointer = getattr(pointer, "weight")
+            elif scope_names[0] == "output_bias" or scope_names[0] == "beta":
+                pointer = getattr(pointer, "bias")
+            elif scope_names[0] == "output_weights":
+                pointer = getattr(pointer, "weight")
+            else:
+                try:
+                    pointer = getattr(pointer, scope_names[0])
+                except AttributeError:
+                    logger.info(f"Skipping {'/'.join(name)}")
+                    continue
+            if len(scope_names) >= 2:
+                num = int(scope_names[1])
+                pointer = pointer[num]
+        if m_name[-11:] == "_embeddings":
+            pointer = getattr(pointer, "weight")
+        elif m_name[-10:] in [f"Embedder_{i}" for i in range(8)]:
+            pointer = getattr(pointer, "weight")
+        elif m_name == "kernel":
+            array = np.transpose(array)
+
+        if pointer.shape != array.shape:
+            raise ValueError(f"Pointer shape {pointer.shape} and array shape {array.shape} mismatched")
+
+        logger.info(f"Initialize PyTorch weight {name}")
+        pointer.data = torch.from_numpy(array)
+    return model
+
+
+class CanineEmbeddings(nn.Module):
+    """Construct the character, position and token_type embeddings."""
+
+    def __init__(self, config):
+        super().__init__()
+
+        self.config = config
+
+        # character embeddings
+        shard_embedding_size = config.hidden_size // config.num_hash_functions
+        for i in range(config.num_hash_functions):
+            name = f"HashBucketCodepointEmbedder_{i}"
+            setattr(self, name, nn.Embedding(config.num_hash_buckets, shard_embedding_size))
+        self.char_position_embeddings = nn.Embedding(config.num_hash_buckets, config.hidden_size)
+        self.token_type_embeddings = nn.Embedding(config.type_vocab_size, config.hidden_size)
+
+        # self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load
+        # any TensorFlow checkpoint file
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+        # position_ids (1, len position emb) is contiguous in memory and exported when serialized
+        self.register_buffer(
+            "position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)), persistent=False
+        )
+        self.position_embedding_type = getattr(config, "position_embedding_type", "absolute")
+
+    def _hash_bucket_tensors(self, input_ids, num_hashes: int, num_buckets: int):
+        """
+        Converts ids to hash bucket ids via multiple hashing.
+
+        Args:
+            input_ids: The codepoints or other IDs to be hashed.
+            num_hashes: The number of hash functions to use.
+            num_buckets: The number of hash buckets (i.e. embeddings in each table).
+
+        Returns:
+            A list of tensors, each of which is the hash bucket IDs from one hash function.
+        """
+        if num_hashes > len(_PRIMES):
+            raise ValueError(f"`num_hashes` must be <= {len(_PRIMES)}")
+
+        primes = _PRIMES[:num_hashes]
+
+        result_tensors = []
+        for prime in primes:
+            hashed = ((input_ids + 1) * prime) % num_buckets
+            result_tensors.append(hashed)
+        return result_tensors
+
+    def _embed_hash_buckets(self, input_ids, embedding_size: int, num_hashes: int, num_buckets: int):
+        """Converts IDs (e.g. codepoints) into embeddings via multiple hashing."""
+        if embedding_size % num_hashes != 0:
+            raise ValueError(f"Expected `embedding_size` ({embedding_size}) % `num_hashes` ({num_hashes}) == 0")
+
+        hash_bucket_tensors = self._hash_bucket_tensors(input_ids, num_hashes=num_hashes, num_buckets=num_buckets)
+        embedding_shards = []
+        for i, hash_bucket_ids in enumerate(hash_bucket_tensors):
+            name = f"HashBucketCodepointEmbedder_{i}"
+            shard_embeddings = getattr(self, name)(hash_bucket_ids)
+            embedding_shards.append(shard_embeddings)
+
+        return torch.cat(embedding_shards, dim=-1)
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+    ) -> torch.FloatTensor:
+        if input_ids is not None:
+            input_shape = input_ids.size()
+        else:
+            input_shape = inputs_embeds.size()[:-1]
+
+        seq_length = input_shape[1]
+
+        if position_ids is None:
+            position_ids = self.position_ids[:, :seq_length]
+
+        if token_type_ids is None:
+            token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=self.position_ids.device)
+
+        if inputs_embeds is None:
+            inputs_embeds = self._embed_hash_buckets(
+                input_ids, self.config.hidden_size, self.config.num_hash_functions, self.config.num_hash_buckets
+            )
+
+        token_type_embeddings = self.token_type_embeddings(token_type_ids)
+
+        embeddings = inputs_embeds + token_type_embeddings
+
+        if self.position_embedding_type == "absolute":
+            position_embeddings = self.char_position_embeddings(position_ids)
+            embeddings += position_embeddings
+        embeddings = self.LayerNorm(embeddings)
+        embeddings = self.dropout(embeddings)
+        return embeddings
+
+
+class CharactersToMolecules(nn.Module):
+    """Convert character sequence to initial molecule sequence (i.e. downsample) using strided convolutions."""
+
+    def __init__(self, config):
+        super().__init__()
+
+        self.conv = nn.Conv1d(
+            in_channels=config.hidden_size,
+            out_channels=config.hidden_size,
+            kernel_size=config.downsampling_rate,
+            stride=config.downsampling_rate,
+        )
+        self.activation = ACT2FN[config.hidden_act]
+
+        # self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load
+        # any TensorFlow checkpoint file
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+
+    def forward(self, char_encoding: torch.Tensor) -> torch.Tensor:
+        # `cls_encoding`: [batch, 1, hidden_size]
+        cls_encoding = char_encoding[:, 0:1, :]
+
+        # char_encoding has shape [batch, char_seq, hidden_size]
+        # We transpose it to be [batch, hidden_size, char_seq]
+        char_encoding = torch.transpose(char_encoding, 1, 2)
+        downsampled = self.conv(char_encoding)
+        downsampled = torch.transpose(downsampled, 1, 2)
+        downsampled = self.activation(downsampled)
+
+        # Truncate the last molecule in order to reserve a position for [CLS].
+        # Often, the last position is never used (unless we completely fill the
+        # text buffer). This is important in order to maintain alignment on TPUs
+        # (i.e. a multiple of 128).
+        downsampled_truncated = downsampled[:, 0:-1, :]
+
+        # We also keep [CLS] as a separate sequence position since we always
+        # want to reserve a position (and the model capacity that goes along
+        # with that) in the deep BERT stack.
+        # `result`: [batch, molecule_seq, molecule_dim]
+        result = torch.cat([cls_encoding, downsampled_truncated], dim=1)
+
+        result = self.LayerNorm(result)
+
+        return result
+
+
+class ConvProjection(nn.Module):
+    """
+    Project representations from hidden_size*2 back to hidden_size across a window of w = config.upsampling_kernel_size
+    characters.
+    """
+
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.conv = nn.Conv1d(
+            in_channels=config.hidden_size * 2,
+            out_channels=config.hidden_size,
+            kernel_size=config.upsampling_kernel_size,
+            stride=1,
+        )
+        self.activation = ACT2FN[config.hidden_act]
+        # self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load
+        # any TensorFlow checkpoint file
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(
+        self,
+        inputs: torch.Tensor,
+        final_seq_char_positions: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        # inputs has shape [batch, mol_seq, molecule_hidden_size+char_hidden_final]
+        # we transpose it to be [batch, molecule_hidden_size+char_hidden_final, mol_seq]
+        inputs = torch.transpose(inputs, 1, 2)
+
+        # PyTorch < 1.9 does not support padding="same" (which is used in the original implementation),
+        # so we pad the tensor manually before passing it to the conv layer
+        # based on https://github.com/google-research/big_transfer/blob/49afe42338b62af9fbe18f0258197a33ee578a6b/bit_tf2/models.py#L36-L38
+        pad_total = self.config.upsampling_kernel_size - 1
+        pad_beg = pad_total // 2
+        pad_end = pad_total - pad_beg
+
+        pad = nn.ConstantPad1d((pad_beg, pad_end), 0)
+        # `result`: shape (batch_size, char_seq_len, hidden_size)
+        result = self.conv(pad(inputs))
+        result = torch.transpose(result, 1, 2)
+        result = self.activation(result)
+        result = self.LayerNorm(result)
+        result = self.dropout(result)
+        final_char_seq = result
+
+        if final_seq_char_positions is not None:
+            # Limit transformer query seq and attention mask to these character
+            # positions to greatly reduce the compute cost. Typically, this is just
+            # done for the MLM training task.
+            # TODO add support for MLM
+            raise NotImplementedError("CanineForMaskedLM is currently not supported")
+        else:
+            query_seq = final_char_seq
+
+        return query_seq
+
+
+class CanineSelfAttention(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        if config.hidden_size % config.num_attention_heads != 0 and not hasattr(config, "embedding_size"):
+            raise ValueError(
+                f"The hidden size ({config.hidden_size}) is not a multiple of the number of attention "
+                f"heads ({config.num_attention_heads})"
+            )
+
+        self.num_attention_heads = config.num_attention_heads
+        self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
+        self.all_head_size = self.num_attention_heads * self.attention_head_size
+
+        self.query = nn.Linear(config.hidden_size, self.all_head_size)
+        self.key = nn.Linear(config.hidden_size, self.all_head_size)
+        self.value = nn.Linear(config.hidden_size, self.all_head_size)
+
+        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
+        self.position_embedding_type = getattr(config, "position_embedding_type", "absolute")
+        if self.position_embedding_type == "relative_key" or self.position_embedding_type == "relative_key_query":
+            self.max_position_embeddings = config.max_position_embeddings
+            self.distance_embedding = nn.Embedding(2 * config.max_position_embeddings - 1, self.attention_head_size)
+
+    def transpose_for_scores(self, x):
+        new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.attention_head_size)
+        x = x.view(*new_x_shape)
+        return x.permute(0, 2, 1, 3)
+
+    def forward(
+        self,
+        from_tensor: torch.Tensor,
+        to_tensor: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
+        mixed_query_layer = self.query(from_tensor)
+
+        # If this is instantiated as a cross-attention module, the keys
+        # and values come from an encoder; the attention mask needs to be
+        # such that the encoder's padding tokens are not attended to.
+
+        key_layer = self.transpose_for_scores(self.key(to_tensor))
+        value_layer = self.transpose_for_scores(self.value(to_tensor))
+
+        query_layer = self.transpose_for_scores(mixed_query_layer)
+
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
+
+        if self.position_embedding_type == "relative_key" or self.position_embedding_type == "relative_key_query":
+            seq_length = from_tensor.size()[1]
+            position_ids_l = torch.arange(seq_length, dtype=torch.long, device=from_tensor.device).view(-1, 1)
+            position_ids_r = torch.arange(seq_length, dtype=torch.long, device=from_tensor.device).view(1, -1)
+            distance = position_ids_l - position_ids_r
+            positional_embedding = self.distance_embedding(distance + self.max_position_embeddings - 1)
+            positional_embedding = positional_embedding.to(dtype=query_layer.dtype)  # fp16 compatibility
+
+            if self.position_embedding_type == "relative_key":
+                relative_position_scores = torch.einsum("bhld,lrd->bhlr", query_layer, positional_embedding)
+                attention_scores = attention_scores + relative_position_scores
+            elif self.position_embedding_type == "relative_key_query":
+                relative_position_scores_query = torch.einsum("bhld,lrd->bhlr", query_layer, positional_embedding)
+                relative_position_scores_key = torch.einsum("bhrd,lrd->bhlr", key_layer, positional_embedding)
+                attention_scores = attention_scores + relative_position_scores_query + relative_position_scores_key
+
+        attention_scores = attention_scores / math.sqrt(self.attention_head_size)
+        if attention_mask is not None:
+            if attention_mask.ndim == 3:
+                # if attention_mask is 3D, do the following:
+                attention_mask = torch.unsqueeze(attention_mask, dim=1)
+                # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
+                # masked positions, this operation will create a tensor which is 0.0 for
+                # positions we want to attend and the dtype's smallest value for masked positions.
+                attention_mask = (1.0 - attention_mask.float()) * torch.finfo(attention_scores.dtype).min
+            # Apply the attention mask (precomputed for all layers in CanineModel forward() function)
+            attention_scores = attention_scores + attention_mask
+
+        # Normalize the attention scores to probabilities.
+        attention_probs = nn.functional.softmax(attention_scores, dim=-1)
+
+        # This is actually dropping out entire tokens to attend to, which might
+        # seem a bit unusual, but is taken from the original Transformer paper.
+        attention_probs = self.dropout(attention_probs)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            attention_probs = attention_probs * head_mask
+
+        context_layer = torch.matmul(attention_probs, value_layer)
+
+        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
+        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
+        context_layer = context_layer.view(*new_context_layer_shape)
+
+        outputs = (context_layer, attention_probs) if output_attentions else (context_layer,)
+
+        return outputs
+
+
+class CanineSelfOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(
+        self, hidden_states: Tuple[torch.FloatTensor], input_tensor: torch.FloatTensor
+    ) -> Tuple[torch.FloatTensor, torch.FloatTensor]:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+class CanineAttention(nn.Module):
+    """
+    Additional arguments related to local attention:
+
+        - **local** (`bool`, *optional*, defaults to `False`) -- Whether to apply local attention.
+        - **always_attend_to_first_position** (`bool`, *optional*, defaults to `False`) -- Should all blocks be able to
+          attend
+        to the `to_tensor`'s first position (e.g. a [CLS] position)? - **first_position_attends_to_all** (`bool`,
+        *optional*, defaults to `False`) -- Should the *from_tensor*'s first position be able to attend to all
+        positions within the *from_tensor*? - **attend_from_chunk_width** (`int`, *optional*, defaults to 128) -- The
+        width of each block-wise chunk in `from_tensor`. - **attend_from_chunk_stride** (`int`, *optional*, defaults to
+        128) -- The number of elements to skip when moving to the next block in `from_tensor`. -
+        **attend_to_chunk_width** (`int`, *optional*, defaults to 128) -- The width of each block-wise chunk in
+        *to_tensor*. - **attend_to_chunk_stride** (`int`, *optional*, defaults to 128) -- The number of elements to
+        skip when moving to the next block in `to_tensor`.
+    """
+
+    def __init__(
+        self,
+        config,
+        local=False,
+        always_attend_to_first_position: bool = False,
+        first_position_attends_to_all: bool = False,
+        attend_from_chunk_width: int = 128,
+        attend_from_chunk_stride: int = 128,
+        attend_to_chunk_width: int = 128,
+        attend_to_chunk_stride: int = 128,
+    ):
+        super().__init__()
+        self.self = CanineSelfAttention(config)
+        self.output = CanineSelfOutput(config)
+        self.pruned_heads = set()
+
+        # additional arguments related to local attention
+        self.local = local
+        if attend_from_chunk_width < attend_from_chunk_stride:
+            raise ValueError(
+                "`attend_from_chunk_width` < `attend_from_chunk_stride` would cause sequence positions to get skipped."
+            )
+        if attend_to_chunk_width < attend_to_chunk_stride:
+            raise ValueError(
+                "`attend_to_chunk_width` < `attend_to_chunk_stride`would cause sequence positions to get skipped."
+            )
+        self.always_attend_to_first_position = always_attend_to_first_position
+        self.first_position_attends_to_all = first_position_attends_to_all
+        self.attend_from_chunk_width = attend_from_chunk_width
+        self.attend_from_chunk_stride = attend_from_chunk_stride
+        self.attend_to_chunk_width = attend_to_chunk_width
+        self.attend_to_chunk_stride = attend_to_chunk_stride
+
+    def prune_heads(self, heads):
+        if len(heads) == 0:
+            return
+        heads, index = find_pruneable_heads_and_indices(
+            heads, self.self.num_attention_heads, self.self.attention_head_size, self.pruned_heads
+        )
+
+        # Prune linear layers
+        self.self.query = prune_linear_layer(self.self.query, index)
+        self.self.key = prune_linear_layer(self.self.key, index)
+        self.self.value = prune_linear_layer(self.self.value, index)
+        self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)
+
+        # Update hyper params and store pruned heads
+        self.self.num_attention_heads = self.self.num_attention_heads - len(heads)
+        self.self.all_head_size = self.self.attention_head_size * self.self.num_attention_heads
+        self.pruned_heads = self.pruned_heads.union(heads)
+
+    def forward(
+        self,
+        hidden_states: Tuple[torch.FloatTensor],
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.FloatTensor, Optional[torch.FloatTensor]]:
+        if not self.local:
+            self_outputs = self.self(hidden_states, hidden_states, attention_mask, head_mask, output_attentions)
+            attention_output = self_outputs[0]
+        else:
+            from_seq_length = to_seq_length = hidden_states.shape[1]
+            from_tensor = to_tensor = hidden_states
+
+            # Create chunks (windows) that we will attend *from* and then concatenate them.
+            from_chunks = []
+            if self.first_position_attends_to_all:
+                from_chunks.append((0, 1))
+                # We must skip this first position so that our output sequence is the
+                # correct length (this matters in the *from* sequence only).
+                from_start = 1
+            else:
+                from_start = 0
+            for chunk_start in range(from_start, from_seq_length, self.attend_from_chunk_stride):
+                chunk_end = min(from_seq_length, chunk_start + self.attend_from_chunk_width)
+                from_chunks.append((chunk_start, chunk_end))
+
+            # Determine the chunks (windows) that will attend *to*.
+            to_chunks = []
+            if self.first_position_attends_to_all:
+                to_chunks.append((0, to_seq_length))
+            for chunk_start in range(0, to_seq_length, self.attend_to_chunk_stride):
+                chunk_end = min(to_seq_length, chunk_start + self.attend_to_chunk_width)
+                to_chunks.append((chunk_start, chunk_end))
+
+            if len(from_chunks) != len(to_chunks):
+                raise ValueError(
+                    f"Expected to have same number of `from_chunks` ({from_chunks}) and "
+                    f"`to_chunks` ({from_chunks}). Check strides."
+                )
+
+            # next, compute attention scores for each pair of windows and concatenate
+            attention_output_chunks = []
+            attention_probs_chunks = []
+            for (from_start, from_end), (to_start, to_end) in zip(from_chunks, to_chunks):
+                from_tensor_chunk = from_tensor[:, from_start:from_end, :]
+                to_tensor_chunk = to_tensor[:, to_start:to_end, :]
+                # `attention_mask`: <float>[batch_size, from_seq, to_seq]
+                # `attention_mask_chunk`: <float>[batch_size, from_seq_chunk, to_seq_chunk]
+                attention_mask_chunk = attention_mask[:, from_start:from_end, to_start:to_end]
+                if self.always_attend_to_first_position:
+                    cls_attention_mask = attention_mask[:, from_start:from_end, 0:1]
+                    attention_mask_chunk = torch.cat([cls_attention_mask, attention_mask_chunk], dim=2)
+
+                    cls_position = to_tensor[:, 0:1, :]
+                    to_tensor_chunk = torch.cat([cls_position, to_tensor_chunk], dim=1)
+
+                attention_outputs_chunk = self.self(
+                    from_tensor_chunk, to_tensor_chunk, attention_mask_chunk, head_mask, output_attentions
+                )
+                attention_output_chunks.append(attention_outputs_chunk[0])
+                if output_attentions:
+                    attention_probs_chunks.append(attention_outputs_chunk[1])
+
+            attention_output = torch.cat(attention_output_chunks, dim=1)
+
+        attention_output = self.output(attention_output, hidden_states)
+        outputs = (attention_output,)
+        if not self.local:
+            outputs = outputs + self_outputs[1:]  # add attentions if we output them
+        else:
+            outputs = outputs + tuple(attention_probs_chunks)  # add attentions if we output them
+        return outputs
+
+
+class CanineIntermediate(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.intermediate_size)
+        if isinstance(config.hidden_act, str):
+            self.intermediate_act_fn = ACT2FN[config.hidden_act]
+        else:
+            self.intermediate_act_fn = config.hidden_act
+
+    def forward(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.intermediate_act_fn(hidden_states)
+        return hidden_states
+
+
+class CanineOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(self, hidden_states: Tuple[torch.FloatTensor], input_tensor: torch.FloatTensor) -> torch.FloatTensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+class CanineLayer(nn.Module):
+    def __init__(
+        self,
+        config,
+        local,
+        always_attend_to_first_position,
+        first_position_attends_to_all,
+        attend_from_chunk_width,
+        attend_from_chunk_stride,
+        attend_to_chunk_width,
+        attend_to_chunk_stride,
+    ):
+        super().__init__()
+        self.chunk_size_feed_forward = config.chunk_size_feed_forward
+        self.seq_len_dim = 1
+        self.attention = CanineAttention(
+            config,
+            local,
+            always_attend_to_first_position,
+            first_position_attends_to_all,
+            attend_from_chunk_width,
+            attend_from_chunk_stride,
+            attend_to_chunk_width,
+            attend_to_chunk_stride,
+        )
+        self.intermediate = CanineIntermediate(config)
+        self.output = CanineOutput(config)
+
+    def forward(
+        self,
+        hidden_states: Tuple[torch.FloatTensor],
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.FloatTensor, Optional[torch.FloatTensor]]:
+        self_attention_outputs = self.attention(
+            hidden_states,
+            attention_mask,
+            head_mask,
+            output_attentions=output_attentions,
+        )
+        attention_output = self_attention_outputs[0]
+
+        outputs = self_attention_outputs[1:]  # add self attentions if we output attention weights
+
+        layer_output = apply_chunking_to_forward(
+            self.feed_forward_chunk, self.chunk_size_feed_forward, self.seq_len_dim, attention_output
+        )
+        outputs = (layer_output,) + outputs
+
+        return outputs
+
+    def feed_forward_chunk(self, attention_output):
+        intermediate_output = self.intermediate(attention_output)
+        layer_output = self.output(intermediate_output, attention_output)
+        return layer_output
+
+
+class CanineEncoder(nn.Module):
+    def __init__(
+        self,
+        config,
+        local=False,
+        always_attend_to_first_position=False,
+        first_position_attends_to_all=False,
+        attend_from_chunk_width=128,
+        attend_from_chunk_stride=128,
+        attend_to_chunk_width=128,
+        attend_to_chunk_stride=128,
+    ):
+        super().__init__()
+        self.config = config
+        self.layer = nn.ModuleList(
+            [
+                CanineLayer(
+                    config,
+                    local,
+                    always_attend_to_first_position,
+                    first_position_attends_to_all,
+                    attend_from_chunk_width,
+                    attend_from_chunk_stride,
+                    attend_to_chunk_width,
+                    attend_to_chunk_stride,
+                )
+                for _ in range(config.num_hidden_layers)
+            ]
+        )
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: Tuple[torch.FloatTensor],
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        output_attentions: Optional[bool] = False,
+        output_hidden_states: Optional[bool] = False,
+        return_dict: Optional[bool] = True,
+    ) -> Union[Tuple, BaseModelOutput]:
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attentions = () if output_attentions else None
+
+        for i, layer_module in enumerate(self.layer):
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            layer_head_mask = head_mask[i] if head_mask is not None else None
+
+            if self.gradient_checkpointing and self.training:
+                layer_outputs = self._gradient_checkpointing_func(
+                    layer_module.__call__,
+                    hidden_states,
+                    attention_mask,
+                    layer_head_mask,
+                    output_attentions,
+                )
+            else:
+                layer_outputs = layer_module(hidden_states, attention_mask, layer_head_mask, output_attentions)
+
+            hidden_states = layer_outputs[0]
+            if output_attentions:
+                all_self_attentions = all_self_attentions + (layer_outputs[1],)
+
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, all_hidden_states, all_self_attentions] if v is not None)
+        return BaseModelOutput(
+            last_hidden_state=hidden_states,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attentions,
+        )
+
+
+class CaninePooler(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.activation = nn.Tanh()
+
+    def forward(self, hidden_states: Tuple[torch.FloatTensor]) -> torch.FloatTensor:
+        # We "pool" the model by simply taking the hidden state corresponding
+        # to the first token.
+        first_token_tensor = hidden_states[:, 0]
+        pooled_output = self.dense(first_token_tensor)
+        pooled_output = self.activation(pooled_output)
+        return pooled_output
+
+
+class CaninePredictionHeadTransform(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        if isinstance(config.hidden_act, str):
+            self.transform_act_fn = ACT2FN[config.hidden_act]
+        else:
+            self.transform_act_fn = config.hidden_act
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+
+    def forward(self, hidden_states: Tuple[torch.FloatTensor]) -> torch.FloatTensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.transform_act_fn(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states)
+        return hidden_states
+
+
+class CanineLMPredictionHead(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.transform = CaninePredictionHeadTransform(config)
+
+        # The output weights are the same as the input embeddings, but there is
+        # an output-only bias for each token.
+        self.decoder = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        self.bias = nn.Parameter(torch.zeros(config.vocab_size))
+
+        # Need a link between the two variables so that the bias is correctly resized with `resize_token_embeddings`
+        self.decoder.bias = self.bias
+
+    def forward(self, hidden_states: Tuple[torch.FloatTensor]) -> torch.FloatTensor:
+        hidden_states = self.transform(hidden_states)
+        hidden_states = self.decoder(hidden_states)
+        return hidden_states
+
+
+class CanineOnlyMLMHead(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.predictions = CanineLMPredictionHead(config)
+
+    def forward(
+        self,
+        sequence_output: Tuple[torch.Tensor],
+    ) -> Tuple[torch.Tensor]:
+        prediction_scores = self.predictions(sequence_output)
+        return prediction_scores
+
+
+class CaninePreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = CanineConfig
+    load_tf_weights = load_tf_weights_in_canine
+    base_model_prefix = "canine"
+    supports_gradient_checkpointing = True
+
+    def _init_weights(self, module):
+        """Initialize the weights"""
+        if isinstance(module, (nn.Linear, nn.Conv1d)):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+        elif isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+
+
+CANINE_START_DOCSTRING = r"""
+    This model is a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) sub-class. Use
+    it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and
+    behavior.
+
+    Parameters:
+        config ([`CanineConfig`]): Model configuration class with all the parameters of the model.
+            Initializing with a config file does not load the weights associated with the model, only the
+            configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+CANINE_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `({0})`):
+            Indices of input sequence tokens in the vocabulary.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`torch.FloatTensor` of shape `({0})`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+        token_type_ids (`torch.LongTensor` of shape `({0})`, *optional*):
+            Segment token indices to indicate first and second portions of the inputs. Indices are selected in `[0,
+            1]`:
+
+            - 0 corresponds to a *sentence A* token,
+            - 1 corresponds to a *sentence B* token.
+
+            [What are token type IDs?](../glossary#token-type-ids)
+        position_ids (`torch.LongTensor` of shape `({0})`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.max_position_embeddings - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
+            Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        inputs_embeds (`torch.FloatTensor` of shape `({0}, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
+            is useful if you want more control over how to convert *input_ids* indices into associated vectors than the
+            model's internal embedding lookup matrix.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+@add_start_docstrings(
+    "The bare CANINE Model transformer outputting raw hidden-states without any specific head on top.",
+    CANINE_START_DOCSTRING,
+)
+class CanineModel(CaninePreTrainedModel):
+    def __init__(self, config, add_pooling_layer=True):
+        super().__init__(config)
+        self.config = config
+        shallow_config = copy.deepcopy(config)
+        shallow_config.num_hidden_layers = 1
+
+        self.char_embeddings = CanineEmbeddings(config)
+        # shallow/low-dim transformer encoder to get a initial character encoding
+        self.initial_char_encoder = CanineEncoder(
+            shallow_config,
+            local=True,
+            always_attend_to_first_position=False,
+            first_position_attends_to_all=False,
+            attend_from_chunk_width=config.local_transformer_stride,
+            attend_from_chunk_stride=config.local_transformer_stride,
+            attend_to_chunk_width=config.local_transformer_stride,
+            attend_to_chunk_stride=config.local_transformer_stride,
+        )
+        self.chars_to_molecules = CharactersToMolecules(config)
+        # deep transformer encoder
+        self.encoder = CanineEncoder(config)
+        self.projection = ConvProjection(config)
+        # shallow/low-dim transformer encoder to get a final character encoding
+        self.final_char_encoder = CanineEncoder(shallow_config)
+
+        self.pooler = CaninePooler(config) if add_pooling_layer else None
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def _prune_heads(self, heads_to_prune):
+        """
+        Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
+        class PreTrainedModel
+        """
+        for layer, heads in heads_to_prune.items():
+            self.encoder.layer[layer].attention.prune_heads(heads)
+
+    def _create_3d_attention_mask_from_input_mask(self, from_tensor, to_mask):
+        """
+        Create 3D attention mask from a 2D tensor mask.
+
+        Args:
+            from_tensor: 2D or 3D Tensor of shape [batch_size, from_seq_length, ...].
+            to_mask: int32 Tensor of shape [batch_size, to_seq_length].
+
+        Returns:
+            float Tensor of shape [batch_size, from_seq_length, to_seq_length].
+        """
+        batch_size, from_seq_length = from_tensor.shape[0], from_tensor.shape[1]
+
+        to_seq_length = to_mask.shape[1]
+
+        to_mask = torch.reshape(to_mask, (batch_size, 1, to_seq_length)).float()
+
+        # We don't assume that `from_tensor` is a mask (although it could be). We
+        # don't actually care if we attend *from* padding tokens (only *to* padding)
+        # tokens so we create a tensor of all ones.
+        broadcast_ones = torch.ones(size=(batch_size, from_seq_length, 1), dtype=torch.float32, device=to_mask.device)
+
+        # Here we broadcast along two dimensions to create the mask.
+        mask = broadcast_ones * to_mask
+
+        return mask
+
+    def _downsample_attention_mask(self, char_attention_mask: torch.Tensor, downsampling_rate: int):
+        """Downsample 2D character attention mask to 2D molecule attention mask using MaxPool1d layer."""
+
+        # first, make char_attention_mask 3D by adding a channel dim
+        batch_size, char_seq_len = char_attention_mask.shape
+        poolable_char_mask = torch.reshape(char_attention_mask, (batch_size, 1, char_seq_len))
+
+        # next, apply MaxPool1d to get pooled_molecule_mask of shape (batch_size, 1, mol_seq_len)
+        pooled_molecule_mask = torch.nn.MaxPool1d(kernel_size=downsampling_rate, stride=downsampling_rate)(
+            poolable_char_mask.float()
+        )
+
+        # finally, squeeze to get tensor of shape (batch_size, mol_seq_len)
+        molecule_attention_mask = torch.squeeze(pooled_molecule_mask, dim=-1)
+
+        return molecule_attention_mask
+
+    def _repeat_molecules(self, molecules: torch.Tensor, char_seq_length: torch.Tensor) -> torch.Tensor:
+        """Repeats molecules to make them the same length as the char sequence."""
+
+        rate = self.config.downsampling_rate
+
+        molecules_without_extra_cls = molecules[:, 1:, :]
+        # `repeated`: [batch_size, almost_char_seq_len, molecule_hidden_size]
+        repeated = torch.repeat_interleave(molecules_without_extra_cls, repeats=rate, dim=-2)
+
+        # So far, we've repeated the elements sufficient for any `char_seq_length`
+        # that's a multiple of `downsampling_rate`. Now we account for the last
+        # n elements (n < `downsampling_rate`), i.e. the remainder of floor
+        # division. We do this by repeating the last molecule a few extra times.
+        last_molecule = molecules[:, -1:, :]
+        remainder_length = torch.fmod(torch.tensor(char_seq_length), torch.tensor(rate)).item()
+        remainder_repeated = torch.repeat_interleave(
+            last_molecule,
+            # +1 molecule to compensate for truncation.
+            repeats=remainder_length + rate,
+            dim=-2,
+        )
+
+        # `repeated`: [batch_size, char_seq_len, molecule_hidden_size]
+        return torch.cat([repeated, remainder_repeated], dim=-2)
+
+    @add_start_docstrings_to_model_forward(CANINE_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=CanineModelOutputWithPooling,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, CanineModelOutputWithPooling]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attentions = () if output_attentions else None
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask)
+            input_shape = input_ids.size()
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        batch_size, seq_length = input_shape
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+
+        if attention_mask is None:
+            attention_mask = torch.ones(((batch_size, seq_length)), device=device)
+        if token_type_ids is None:
+            token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
+
+        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
+        # ourselves in which case we just need to make it broadcastable to all heads.
+        extended_attention_mask: torch.Tensor = self.get_extended_attention_mask(attention_mask, input_shape)
+        molecule_attention_mask = self._downsample_attention_mask(
+            attention_mask, downsampling_rate=self.config.downsampling_rate
+        )
+        extended_molecule_attention_mask: torch.Tensor = self.get_extended_attention_mask(
+            molecule_attention_mask, (batch_size, molecule_attention_mask.shape[-1])
+        )
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
+        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
+        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
+
+        # `input_char_embeddings`: shape (batch_size, char_seq, char_dim)
+        input_char_embeddings = self.char_embeddings(
+            input_ids=input_ids,
+            position_ids=position_ids,
+            token_type_ids=token_type_ids,
+            inputs_embeds=inputs_embeds,
+        )
+
+        # Contextualize character embeddings using shallow Transformer.
+        # We use a 3D attention mask for the local attention.
+        # `input_char_encoding`: shape (batch_size, char_seq_len, char_dim)
+        char_attention_mask = self._create_3d_attention_mask_from_input_mask(
+            input_ids if input_ids is not None else inputs_embeds, attention_mask
+        )
+        init_chars_encoder_outputs = self.initial_char_encoder(
+            input_char_embeddings,
+            attention_mask=char_attention_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+        )
+        input_char_encoding = init_chars_encoder_outputs.last_hidden_state
+
+        # Downsample chars to molecules.
+        # The following lines have dimensions: [batch, molecule_seq, molecule_dim].
+        # In this transformation, we change the dimensionality from `char_dim` to
+        # `molecule_dim`, but do *NOT* add a resnet connection. Instead, we rely on
+        # the resnet connections (a) from the final char transformer stack back into
+        # the original char transformer stack and (b) the resnet connections from
+        # the final char transformer stack back into the deep BERT stack of
+        # molecules.
+        #
+        # Empirically, it is critical to use a powerful enough transformation here:
+        # mean pooling causes training to diverge with huge gradient norms in this
+        # region of the model; using a convolution here resolves this issue. From
+        # this, it seems that molecules and characters require a very different
+        # feature space; intuitively, this makes sense.
+        init_molecule_encoding = self.chars_to_molecules(input_char_encoding)
+
+        # Deep BERT encoder
+        # `molecule_sequence_output`: shape (batch_size, mol_seq_len, mol_dim)
+        encoder_outputs = self.encoder(
+            init_molecule_encoding,
+            attention_mask=extended_molecule_attention_mask,
+            head_mask=head_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        molecule_sequence_output = encoder_outputs[0]
+        pooled_output = self.pooler(molecule_sequence_output) if self.pooler is not None else None
+
+        # Upsample molecules back to characters.
+        # `repeated_molecules`: shape (batch_size, char_seq_len, mol_hidden_size)
+        repeated_molecules = self._repeat_molecules(molecule_sequence_output, char_seq_length=input_shape[-1])
+
+        # Concatenate representations (contextualized char embeddings and repeated molecules):
+        # `concat`: shape [batch_size, char_seq_len, molecule_hidden_size+char_hidden_final]
+        concat = torch.cat([input_char_encoding, repeated_molecules], dim=-1)
+
+        # Project representation dimension back to hidden_size
+        # `sequence_output`: shape (batch_size, char_seq_len, hidden_size])
+        sequence_output = self.projection(concat)
+
+        # Apply final shallow Transformer
+        # `sequence_output`: shape (batch_size, char_seq_len, hidden_size])
+        final_chars_encoder_outputs = self.final_char_encoder(
+            sequence_output,
+            attention_mask=extended_attention_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+        )
+        sequence_output = final_chars_encoder_outputs.last_hidden_state
+
+        if output_hidden_states:
+            deep_encoder_hidden_states = encoder_outputs.hidden_states if return_dict else encoder_outputs[1]
+            all_hidden_states = (
+                all_hidden_states
+                + init_chars_encoder_outputs.hidden_states
+                + deep_encoder_hidden_states
+                + final_chars_encoder_outputs.hidden_states
+            )
+
+        if output_attentions:
+            deep_encoder_self_attentions = encoder_outputs.attentions if return_dict else encoder_outputs[-1]
+            all_self_attentions = (
+                all_self_attentions
+                + init_chars_encoder_outputs.attentions
+                + deep_encoder_self_attentions
+                + final_chars_encoder_outputs.attentions
+            )
+
+        if not return_dict:
+            output = (sequence_output, pooled_output)
+            output += tuple(v for v in [all_hidden_states, all_self_attentions] if v is not None)
+            return output
+
+        return CanineModelOutputWithPooling(
+            last_hidden_state=sequence_output,
+            pooler_output=pooled_output,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    CANINE Model transformer with a sequence classification/regression head on top (a linear layer on top of the pooled
+    output) e.g. for GLUE tasks.
+    """,
+    CANINE_START_DOCSTRING,
+)
+class CanineForSequenceClassification(CaninePreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+
+        self.canine = CanineModel(config)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        self.classifier = nn.Linear(config.hidden_size, config.num_labels)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @add_start_docstrings_to_model_forward(CANINE_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=SequenceClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, SequenceClassifierOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.canine(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        pooled_output = outputs[1]
+
+        pooled_output = self.dropout(pooled_output)
+        logits = self.classifier(pooled_output)
+
+        loss = None
+        if labels is not None:
+            if self.config.problem_type is None:
+                if self.num_labels == 1:
+                    self.config.problem_type = "regression"
+                elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
+                    self.config.problem_type = "single_label_classification"
+                else:
+                    self.config.problem_type = "multi_label_classification"
+
+            if self.config.problem_type == "regression":
+                loss_fct = MSELoss()
+                if self.num_labels == 1:
+                    loss = loss_fct(logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(logits, labels)
+            elif self.config.problem_type == "single_label_classification":
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            elif self.config.problem_type == "multi_label_classification":
+                loss_fct = BCEWithLogitsLoss()
+                loss = loss_fct(logits, labels)
+        if not return_dict:
+            output = (logits,) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return SequenceClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    CANINE Model with a multiple choice classification head on top (a linear layer on top of the pooled output and a
+    softmax) e.g. for RocStories/SWAG tasks.
+    """,
+    CANINE_START_DOCSTRING,
+)
+class CanineForMultipleChoice(CaninePreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.canine = CanineModel(config)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        self.classifier = nn.Linear(config.hidden_size, 1)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @add_start_docstrings_to_model_forward(CANINE_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length"))
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=MultipleChoiceModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, MultipleChoiceModelOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the multiple choice classification loss. Indices should be in `[0, ...,
+            num_choices-1]` where `num_choices` is the size of the second dimension of the input tensors. (See
+            `input_ids` above)
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        num_choices = input_ids.shape[1] if input_ids is not None else inputs_embeds.shape[1]
+
+        input_ids = input_ids.view(-1, input_ids.size(-1)) if input_ids is not None else None
+        attention_mask = attention_mask.view(-1, attention_mask.size(-1)) if attention_mask is not None else None
+        token_type_ids = token_type_ids.view(-1, token_type_ids.size(-1)) if token_type_ids is not None else None
+        position_ids = position_ids.view(-1, position_ids.size(-1)) if position_ids is not None else None
+        inputs_embeds = (
+            inputs_embeds.view(-1, inputs_embeds.size(-2), inputs_embeds.size(-1))
+            if inputs_embeds is not None
+            else None
+        )
+
+        outputs = self.canine(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        pooled_output = outputs[1]
+
+        pooled_output = self.dropout(pooled_output)
+        logits = self.classifier(pooled_output)
+        reshaped_logits = logits.view(-1, num_choices)
+
+        loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(reshaped_logits, labels)
+
+        if not return_dict:
+            output = (reshaped_logits,) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return MultipleChoiceModelOutput(
+            loss=loss,
+            logits=reshaped_logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    CANINE Model with a token classification head on top (a linear layer on top of the hidden-states output) e.g. for
+    Named-Entity-Recognition (NER) tasks.
+    """,
+    CANINE_START_DOCSTRING,
+)
+class CanineForTokenClassification(CaninePreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+
+        self.canine = CanineModel(config)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        self.classifier = nn.Linear(config.hidden_size, config.num_labels)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @add_start_docstrings_to_model_forward(CANINE_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @replace_return_docstrings(output_type=TokenClassifierOutput, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, TokenClassifierOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the token classification loss. Indices should be in `[0, ..., config.num_labels - 1]`.
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> from transformers import AutoTokenizer, CanineForTokenClassification
+        >>> import torch
+
+        >>> tokenizer = AutoTokenizer.from_pretrained("google/canine-s")
+        >>> model = CanineForTokenClassification.from_pretrained("google/canine-s")
+
+        >>> inputs = tokenizer(
+        ...     "HuggingFace is a company based in Paris and New York", add_special_tokens=False, return_tensors="pt"
+        ... )
+
+        >>> with torch.no_grad():
+        ...     logits = model(**inputs).logits
+
+        >>> predicted_token_class_ids = logits.argmax(-1)
+
+        >>> # Note that tokens are classified rather then input words which means that
+        >>> # there might be more predicted token classes than words.
+        >>> # Multiple token classes might account for the same word
+        >>> predicted_tokens_classes = [model.config.id2label[t.item()] for t in predicted_token_class_ids[0]]
+        >>> predicted_tokens_classes  # doctest: +SKIP
+        ```
+
+        ```python
+        >>> labels = predicted_token_class_ids
+        >>> loss = model(**inputs, labels=labels).loss
+        >>> round(loss.item(), 2)  # doctest: +SKIP
+        ```"""
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.canine(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        sequence_output = outputs[0]
+
+        sequence_output = self.dropout(sequence_output)
+        logits = self.classifier(sequence_output)
+
+        loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+
+        if not return_dict:
+            output = (logits,) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TokenClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    CANINE Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear
+    layers on top of the hidden-states output to compute `span start logits` and `span end logits`).
+    """,
+    CANINE_START_DOCSTRING,
+)
+class CanineForQuestionAnswering(CaninePreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+
+        self.canine = CanineModel(config)
+        self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @add_start_docstrings_to_model_forward(CANINE_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        checkpoint="Splend1dchan/canine-c-squad",
+        output_type=QuestionAnsweringModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+        expected_output="'nice puppet'",
+        expected_loss=8.81,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        start_positions: Optional[torch.LongTensor] = None,
+        end_positions: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, QuestionAnsweringModelOutput]:
+        r"""
+        start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for position (index) of the start of the labelled span for computing the token classification loss.
+            Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
+            are not taken into account for computing the loss.
+        end_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for position (index) of the end of the labelled span for computing the token classification loss.
+            Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
+            are not taken into account for computing the loss.
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.canine(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        sequence_output = outputs[0]
+
+        logits = self.qa_outputs(sequence_output)
+        start_logits, end_logits = logits.split(1, dim=-1)
+        start_logits = start_logits.squeeze(-1)
+        end_logits = end_logits.squeeze(-1)
+
+        total_loss = None
+        if start_positions is not None and end_positions is not None:
+            # If we are on multi-GPU, split add a dimension
+            if len(start_positions.size()) > 1:
+                start_positions = start_positions.squeeze(-1)
+            if len(end_positions.size()) > 1:
+                end_positions = end_positions.squeeze(-1)
+            # sometimes the start/end positions are outside our model inputs, we ignore these terms
+            ignored_index = start_logits.size(1)
+            start_positions.clamp_(0, ignored_index)
+            end_positions.clamp_(0, ignored_index)
+
+            loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
+            start_loss = loss_fct(start_logits, start_positions)
+            end_loss = loss_fct(end_logits, end_positions)
+            total_loss = (start_loss + end_loss) / 2
+
+        if not return_dict:
+            output = (start_logits, end_logits) + outputs[2:]
+            return ((total_loss,) + output) if total_loss is not None else output
+
+        return QuestionAnsweringModelOutput(
+            loss=total_loss,
+            start_logits=start_logits,
+            end_logits=end_logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )

parrot/lib/python3.10/site-packages/transformers/models/canine/tokenization_canine.py

@@ -0,0 +1,241 @@
+# coding=utf-8
+# Copyright Google AI and The HuggingFace Inc. team. All rights reserved.
+#
+# 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.
+"""Tokenization classes for CANINE."""
+
+from typing import Dict, List, Optional
+
+from ...tokenization_utils import AddedToken, PreTrainedTokenizer
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+# Unicode defines 1,114,112 total “codepoints”
+UNICODE_VOCAB_SIZE = 1114112
+
+# Below: Constants defining canonical codepoints for special, pseudo-characters.
+# Copied from https://github.com/google-research/language/blob/master/language/canine/special_codepoints.py
+PAD = 0
+CLS = 0xE000
+SEP = 0xE001
+BOS = 0xE002
+MASK = 0xE003
+RESERVED = 0xE004
+
+# Maps special codepoints to human-readable names.
+SPECIAL_CODEPOINTS: Dict[int, str] = {
+    # Special symbols are represented using codepoints values that are valid,
+    # but designated as "Private Use", meaning that they will never be assigned
+    # characters by the Unicode Consortium, and are thus safe for use here.
+    #
+    # NOTE: Do *NOT* add any sort of [UNK_CHAR] here. They are explicitly
+    # excluded and should fail with a hard error.
+    CLS: "[CLS]",
+    SEP: "[SEP]",
+    BOS: "[BOS]",
+    MASK: "[MASK]",
+    PAD: "[PAD]",
+    RESERVED: "[RESERVED]",
+}
+
+# Maps special codepoint human-readable names to their codepoint values.
+SPECIAL_CODEPOINTS_BY_NAME: Dict[str, int] = {name: codepoint for codepoint, name in SPECIAL_CODEPOINTS.items()}
+
+
+class CanineTokenizer(PreTrainedTokenizer):
+    r"""
+    Construct a CANINE tokenizer (i.e. a character splitter). It turns text into a sequence of characters, and then
+    converts each character into its Unicode code point.
+
+    [`CanineTokenizer`] inherits from [`PreTrainedTokenizer`].
+
+    Refer to superclass [`PreTrainedTokenizer`] for usage examples and documentation concerning parameters.
+
+    Args:
+        model_max_length (`int`, *optional*, defaults to 2048):
+                The maximum sentence length the model accepts.
+    """
+
+    def __init__(
+        self,
+        bos_token=chr(CLS),
+        eos_token=chr(SEP),
+        sep_token=chr(SEP),
+        cls_token=chr(CLS),
+        pad_token=chr(PAD),
+        mask_token=chr(MASK),
+        add_prefix_space=False,
+        model_max_length=2048,
+        **kwargs,
+    ):
+        bos_token = AddedToken(bos_token, lstrip=False, rstrip=False) if isinstance(bos_token, str) else bos_token
+        eos_token = AddedToken(eos_token, lstrip=False, rstrip=False) if isinstance(eos_token, str) else eos_token
+        sep_token = AddedToken(sep_token, lstrip=False, rstrip=False) if isinstance(sep_token, str) else sep_token
+        cls_token = AddedToken(cls_token, lstrip=False, rstrip=False) if isinstance(cls_token, str) else cls_token
+        pad_token = AddedToken(pad_token, lstrip=False, rstrip=False) if isinstance(pad_token, str) else pad_token
+
+        # Mask token behave like a normal word, i.e. include the space before it
+        mask_token = AddedToken(mask_token, lstrip=True, rstrip=False) if isinstance(mask_token, str) else mask_token
+
+        # Creates a mapping for looking up the IDs of special symbols.
+        self._special_codepoints: Dict[str, int] = {}
+        for codepoint, name in SPECIAL_CODEPOINTS.items():
+            self._special_codepoints[name] = codepoint
+
+        # Creates a mapping for looking up the string forms of special symbol IDs.
+        self._special_codepoint_strings: Dict[int, str] = {
+            codepoint: name for name, codepoint in self._special_codepoints.items()
+        }
+
+        self._unicode_vocab_size = UNICODE_VOCAB_SIZE
+        self._num_special_tokens = len(self._special_codepoints)
+
+        super().__init__(
+            bos_token=bos_token,
+            eos_token=eos_token,
+            sep_token=sep_token,
+            cls_token=cls_token,
+            pad_token=pad_token,
+            mask_token=mask_token,
+            add_prefix_space=add_prefix_space,
+            model_max_length=model_max_length,
+            **kwargs,
+        )
+
+    @property
+    def vocab_size(self) -> int:
+        return self._unicode_vocab_size
+
+    def get_vocab(self):
+        vocab = {chr(i): i for i in range(self.vocab_size)}
+        vocab.update(self.added_tokens_encoder)
+        return vocab
+
+    def _tokenize(self, text: str) -> List[str]:
+        """Tokenize a string (i.e. perform character splitting)."""
+        return list(text)
+
+    def _convert_token_to_id(self, token: str) -> int:
+        """Converts a token (i.e. a Unicode character) in an id (i.e. its integer Unicode code point value)."""
+        try:
+            return ord(token)
+        except TypeError:
+            raise ValueError(f"invalid token: '{token}'")
+
+    def _convert_id_to_token(self, index: int) -> str:
+        """
+        Converts a Unicode code point (integer) in a token (str). In case it's a special code point, convert to
+        human-readable format.
+        """
+        try:
+            if index in SPECIAL_CODEPOINTS:
+                return SPECIAL_CODEPOINTS[index]
+            return chr(index)
+        except TypeError:
+            raise ValueError(f"invalid id: {index}")
+
+    def convert_tokens_to_string(self, tokens):
+        return "".join(tokens)
+
+    def build_inputs_with_special_tokens(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
+    ) -> List[int]:
+        """
+        Build model inputs from a sequence or a pair of sequence for sequence classification tasks by concatenating and
+        adding special tokens. A CANINE sequence has the following format:
+
+        - single sequence: `[CLS] X [SEP]`
+        - pair of sequences: `[CLS] A [SEP] B [SEP]`
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs to which the special tokens will be added.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            `List[int]`: List of [input IDs](../glossary#input-ids) with the appropriate special tokens.
+        """
+        sep = [self.sep_token_id]
+        cls = [self.cls_token_id]
+
+        result = cls + token_ids_0 + sep
+        if token_ids_1 is not None:
+            result += token_ids_1 + sep
+        return result
+
+    def get_special_tokens_mask(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None, already_has_special_tokens: bool = False
+    ) -> List[int]:
+        """
+        Retrieve sequence ids from a token list that has no special tokens added. This method is called when adding
+        special tokens using the tokenizer `prepare_for_model` method.
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+            already_has_special_tokens (`bool`, *optional*, defaults to `False`):
+                Whether or not the token list is already formatted with special tokens for the model.
+
+        Returns:
+            `List[int]`: A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.
+        """
+        if already_has_special_tokens:
+            return super().get_special_tokens_mask(
+                token_ids_0=token_ids_0, token_ids_1=token_ids_1, already_has_special_tokens=True
+            )
+
+        result = [1] + ([0] * len(token_ids_0)) + [1]
+        if token_ids_1 is not None:
+            result += ([0] * len(token_ids_1)) + [1]
+        return result
+
+    def create_token_type_ids_from_sequences(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
+    ) -> List[int]:
+        """
+        Create a mask from the two sequences passed to be used in a sequence-pair classification task. A CANINE
+        sequence pair mask has the following format:
+
+        ```
+        0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1
+        | first sequence    | second sequence |
+        ```
+
+        If `token_ids_1` is `None`, this method only returns the first portion of the mask (0s).
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            `List[int]`: List of [token type IDs](../glossary#token-type-ids) according to the given sequence(s).
+        """
+        sep = [self.sep_token_id]
+        cls = [self.cls_token_id]
+
+        result = len(cls + token_ids_0 + sep) * [0]
+        if token_ids_1 is not None:
+            result += len(token_ids_1 + sep) * [1]
+        return result
+
+    # CanineTokenizer has no vocab file
+    def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None):
+        return ()

parrot/lib/python3.10/site-packages/transformers/models/clipseg/processing_clipseg.py

@@ -0,0 +1,161 @@
+# coding=utf-8
+# Copyright 2022 The HuggingFace Inc. 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.
+"""
+Image/Text processor class for CLIPSeg
+"""
+
+import warnings
+
+from ...processing_utils import ProcessorMixin
+from ...tokenization_utils_base import BatchEncoding
+
+
+class CLIPSegProcessor(ProcessorMixin):
+    r"""
+    Constructs a CLIPSeg processor which wraps a CLIPSeg image processor and a CLIP tokenizer into a single processor.
+
+    [`CLIPSegProcessor`] offers all the functionalities of [`ViTImageProcessor`] and [`CLIPTokenizerFast`]. See the
+    [`~CLIPSegProcessor.__call__`] and [`~CLIPSegProcessor.decode`] for more information.
+
+    Args:
+        image_processor ([`ViTImageProcessor`], *optional*):
+            The image processor is a required input.
+        tokenizer ([`CLIPTokenizerFast`], *optional*):
+            The tokenizer is a required input.
+    """
+
+    attributes = ["image_processor", "tokenizer"]
+    image_processor_class = "ViTImageProcessor"
+    tokenizer_class = ("CLIPTokenizer", "CLIPTokenizerFast")
+
+    def __init__(self, image_processor=None, tokenizer=None, **kwargs):
+        feature_extractor = None
+        if "feature_extractor" in kwargs:
+            warnings.warn(
+                "The `feature_extractor` argument is deprecated and will be removed in v5, use `image_processor`"
+                " instead.",
+                FutureWarning,
+            )
+            feature_extractor = kwargs.pop("feature_extractor")
+
+        image_processor = image_processor if image_processor is not None else feature_extractor
+        if image_processor is None:
+            raise ValueError("You need to specify an `image_processor`.")
+        if tokenizer is None:
+            raise ValueError("You need to specify a `tokenizer`.")
+
+        super().__init__(image_processor, tokenizer)
+
+    def __call__(self, text=None, images=None, visual_prompt=None, return_tensors=None, **kwargs):
+        """
+        Main method to prepare for the model one or several sequences(s) and image(s). This method forwards the `text`
+        and `kwargs` arguments to CLIPTokenizerFast's [`~CLIPTokenizerFast.__call__`] if `text` is not `None` to encode
+        the text. To prepare the image(s), this method forwards the `images` and `kwrags` arguments to
+        ViTImageProcessor's [`~ViTImageProcessor.__call__`] if `images` is not `None`. Please refer to the doctsring of
+        the above two methods for more information.
+
+        Args:
+            text (`str`, `List[str]`, `List[List[str]]`):
+                The sequence or batch of sequences to be encoded. Each sequence can be a string or a list of strings
+                (pretokenized string). If the sequences are provided as list of strings (pretokenized), you must set
+                `is_split_into_words=True` (to lift the ambiguity with a batch of sequences).
+            images (`PIL.Image.Image`, `np.ndarray`, `torch.Tensor`, `List[PIL.Image.Image]`, `List[np.ndarray]`, `List[torch.Tensor]`):
+                The image or batch of images to be prepared. Each image can be a PIL image, NumPy array or PyTorch
+                tensor. Both channels-first and channels-last formats are supported.
+            visual_prompt (`PIL.Image.Image`, `np.ndarray`, `torch.Tensor`, `List[PIL.Image.Image]`, `List[np.ndarray]`, `List[torch.Tensor]`):
+                The visual prompt image or batch of images to be prepared. Each visual prompt image can be a PIL image,
+                NumPy array or PyTorch tensor. In case of a NumPy array/PyTorch tensor, each image should be of shape
+                (C, H, W), where C is a number of channels, H and W are image height and width.
+
+            return_tensors (`str` or [`~utils.TensorType`], *optional*):
+                If set, will return tensors of a particular framework. Acceptable values are:
+
+                - `'tf'`: Return TensorFlow `tf.constant` objects.
+                - `'pt'`: Return PyTorch `torch.Tensor` objects.
+                - `'np'`: Return NumPy `np.ndarray` objects.
+                - `'jax'`: Return JAX `jnp.ndarray` objects.
+
+        Returns:
+            [`BatchEncoding`]: A [`BatchEncoding`] with the following fields:
+
+            - **input_ids** -- List of token ids to be fed to a model. Returned when `text` is not `None`.
+            - **attention_mask** -- List of indices specifying which tokens should be attended to by the model (when
+              `return_attention_mask=True` or if *"attention_mask"* is in `self.model_input_names` and if `text` is not
+              `None`).
+            - **pixel_values** -- Pixel values to be fed to a model. Returned when `images` is not `None`.
+        """
+        if text is None and visual_prompt is None and images is None:
+            raise ValueError("You have to specify either text, visual prompt or images.")
+
+        if text is not None and visual_prompt is not None:
+            raise ValueError("You have to specify exactly one type of prompt. Either text or visual prompt.")
+
+        if text is not None:
+            encoding = self.tokenizer(text, return_tensors=return_tensors, **kwargs)
+
+        if visual_prompt is not None:
+            prompt_features = self.image_processor(visual_prompt, return_tensors=return_tensors, **kwargs)
+
+        if images is not None:
+            image_features = self.image_processor(images, return_tensors=return_tensors, **kwargs)
+
+        if visual_prompt is not None and images is not None:
+            encoding = {
+                "pixel_values": image_features.pixel_values,
+                "conditional_pixel_values": prompt_features.pixel_values,
+            }
+            return encoding
+        elif text is not None and images is not None:
+            encoding["pixel_values"] = image_features.pixel_values
+            return encoding
+        elif text is not None:
+            return encoding
+        elif visual_prompt is not None:
+            encoding = {
+                "conditional_pixel_values": prompt_features.pixel_values,
+            }
+            return encoding
+        else:
+            return BatchEncoding(data=dict(**image_features), tensor_type=return_tensors)
+
+    def batch_decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to CLIPTokenizerFast's [`~PreTrainedTokenizer.batch_decode`]. Please
+        refer to the docstring of this method for more information.
+        """
+        return self.tokenizer.batch_decode(*args, **kwargs)
+
+    def decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to CLIPTokenizerFast's [`~PreTrainedTokenizer.decode`]. Please refer to
+        the docstring of this method for more information.
+        """
+        return self.tokenizer.decode(*args, **kwargs)
+
+    @property
+    def feature_extractor_class(self):
+        warnings.warn(
+            "`feature_extractor_class` is deprecated and will be removed in v5. Use `image_processor_class` instead.",
+            FutureWarning,
+        )
+        return self.image_processor_class
+
+    @property
+    def feature_extractor(self):
+        warnings.warn(
+            "`feature_extractor` is deprecated and will be removed in v5. Use `image_processor` instead.",
+            FutureWarning,
+        )
+        return self.image_processor

parrot/lib/python3.10/site-packages/transformers/models/codegen/__init__.py

@@ -0,0 +1,71 @@
+# Copyright 2022 Salesforce authors, The EleutherAI, and HuggingFace Teams. All rights reserved.
+#
+# 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.
+from typing import TYPE_CHECKING
+
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_tokenizers_available, is_torch_available
+
+
+_import_structure = {
+    "configuration_codegen": ["CodeGenConfig", "CodeGenOnnxConfig"],
+    "tokenization_codegen": ["CodeGenTokenizer"],
+}
+
+try:
+    if not is_tokenizers_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["tokenization_codegen_fast"] = ["CodeGenTokenizerFast"]
+
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_codegen"] = [
+        "CodeGenForCausalLM",
+        "CodeGenModel",
+        "CodeGenPreTrainedModel",
+    ]
+
+if TYPE_CHECKING:
+    from .configuration_codegen import CodeGenConfig, CodeGenOnnxConfig
+    from .tokenization_codegen import CodeGenTokenizer
+
+    try:
+        if not is_tokenizers_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .tokenization_codegen_fast import CodeGenTokenizerFast
+
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_codegen import (
+            CodeGenForCausalLM,
+            CodeGenModel,
+            CodeGenPreTrainedModel,
+        )
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)

parrot/lib/python3.10/site-packages/transformers/models/codegen/configuration_codegen.py

@@ -0,0 +1,226 @@
+# coding=utf-8
+# Copyright 2022 Salesforce authors, The EleutherAI, and HuggingFace Teams. All rights reserved.
+#
+# 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.
+""" CodeGen model configuration"""
+from collections import OrderedDict
+from typing import Any, List, Mapping, Optional
+
+from ... import PreTrainedTokenizer, TensorType, is_torch_available
+from ...configuration_utils import PretrainedConfig
+from ...onnx import OnnxConfigWithPast, PatchingSpec
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class CodeGenConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`CodeGenModel`]. It is used to instantiate a
+    CodeGen model according to the specified arguments, defining the model architecture. Instantiating a configuration
+    with the defaults will yield a similar configuration to that of the CodeGen
+    [Salesforce/codegen-2B-mono](https://huggingface.co/Salesforce/codegen-2B-mono) architecture. Configuration objects
+    inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the documentation from
+    [`PretrainedConfig`] for more information.
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 50400):
+            Vocabulary size of the CodeGen model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`CodeGenModel`].
+        n_positions (`int`, *optional*, defaults to 2048):
+            The maximum sequence length that this model might ever be used with. Typically set this to something large
+            just in case (e.g., 512 or 1024 or 2048).
+        n_ctx (`int`, *optional*, defaults to 2048):
+            This attribute is used in `CodeGenModel.__init__` without any real effect.
+        n_embd (`int`, *optional*, defaults to 4096):
+            Dimensionality of the embeddings and hidden states.
+        n_layer (`int`, *optional*, defaults to 28):
+            Number of hidden layers in the Transformer encoder.
+        n_head (`int`, *optional*, defaults to 16):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        rotary_dim (`int`, *optional*, defaults to 64):
+            Number of dimensions in the embedding that Rotary Position Embedding is applied to.
+        n_inner (`int`, *optional*):
+            Dimensionality of the inner feed-forward layers. `None` will set it to 4 times n_embd
+        activation_function (`str`, *optional*, defaults to `"gelu_new"`):
+            Activation function, to be selected in the list `["relu", "silu", "gelu", "tanh", "gelu_new"]`.
+        resid_pdrop (`float`, *optional*, defaults to 0.0):
+            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
+        embd_pdrop (`int`, *optional*, defaults to 0.0):
+            The dropout ratio for the embeddings.
+        attn_pdrop (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention.
+        layer_norm_epsilon (`float`, *optional*, defaults to 1e-05):
+            The epsilon to use in the layer normalization layers.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models).
+        bos_token_id (`int`, *optional*, defaults to 50256):
+            Beginning of stream token id.
+        eos_token_id (`int`, *optional*, defaults to 50256):
+            End of stream token id.
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether the model's input and output word embeddings should be tied. Note that this is only relevant if the
+            model has a output word embedding layer.
+
+    Example:
+
+    ```python
+    >>> from transformers import CodeGenConfig, CodeGenModel
+
+    >>> # Initializing a CodeGen 6B configuration
+    >>> configuration = CodeGenConfig()
+
+    >>> # Initializing a model (with random weights) from the configuration
+    >>> model = CodeGenModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "codegen"
+    attribute_map = {
+        "max_position_embeddings": "n_positions",
+        "hidden_size": "n_embd",
+        "num_attention_heads": "n_head",
+        "num_hidden_layers": "n_layer",
+    }
+
+    def __init__(
+        self,
+        vocab_size=50400,
+        n_positions=2048,
+        n_ctx=2048,
+        n_embd=4096,
+        n_layer=28,
+        n_head=16,
+        rotary_dim=64,
+        n_inner=None,
+        activation_function="gelu_new",
+        resid_pdrop=0.0,
+        embd_pdrop=0.0,
+        attn_pdrop=0.0,
+        layer_norm_epsilon=1e-5,
+        initializer_range=0.02,
+        use_cache=True,
+        bos_token_id=50256,
+        eos_token_id=50256,
+        tie_word_embeddings=False,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.n_ctx = n_ctx
+        self.n_positions = n_positions
+        self.n_embd = n_embd
+        self.n_layer = n_layer
+        self.n_head = n_head
+        self.n_inner = n_inner
+        self.rotary_dim = rotary_dim
+        self.activation_function = activation_function
+        self.resid_pdrop = resid_pdrop
+        self.embd_pdrop = embd_pdrop
+        self.attn_pdrop = attn_pdrop
+        self.layer_norm_epsilon = layer_norm_epsilon
+        self.initializer_range = initializer_range
+        self.use_cache = use_cache
+
+        self.bos_token_id = bos_token_id
+        self.eos_token_id = eos_token_id
+
+        super().__init__(
+            bos_token_id=bos_token_id, eos_token_id=eos_token_id, tie_word_embeddings=tie_word_embeddings, **kwargs
+        )
+
+
+# Copied from transformers.models.gpt2.configuration_gpt2.GPT2OnnxConfig
+class CodeGenOnnxConfig(OnnxConfigWithPast):
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        task: str = "default",
+        patching_specs: List[PatchingSpec] = None,
+        use_past: bool = False,
+    ):
+        super().__init__(config, task=task, patching_specs=patching_specs, use_past=use_past)
+        if not getattr(self._config, "pad_token_id", None):
+            # TODO: how to do that better?
+            self._config.pad_token_id = 0
+
+    @property
+    def inputs(self) -> Mapping[str, Mapping[int, str]]:
+        common_inputs = OrderedDict({"input_ids": {0: "batch", 1: "sequence"}})
+        if self.use_past:
+            self.fill_with_past_key_values_(common_inputs, direction="inputs")
+            common_inputs["attention_mask"] = {0: "batch", 1: "past_sequence + sequence"}
+        else:
+            common_inputs["attention_mask"] = {0: "batch", 1: "sequence"}
+
+        return common_inputs
+
+    @property
+    def num_layers(self) -> int:
+        return self._config.n_layer
+
+    @property
+    def num_attention_heads(self) -> int:
+        return self._config.n_head
+
+    def generate_dummy_inputs(
+        self,
+        tokenizer: PreTrainedTokenizer,
+        batch_size: int = -1,
+        seq_length: int = -1,
+        is_pair: bool = False,
+        framework: Optional[TensorType] = None,
+    ) -> Mapping[str, Any]:
+        common_inputs = super(OnnxConfigWithPast, self).generate_dummy_inputs(
+            tokenizer, batch_size=batch_size, seq_length=seq_length, is_pair=is_pair, framework=framework
+        )
+
+        # We need to order the input in the way they appears in the forward()
+        ordered_inputs = OrderedDict({"input_ids": common_inputs["input_ids"]})
+
+        # Need to add the past_keys
+        if self.use_past:
+            if not is_torch_available():
+                raise ValueError("Cannot generate dummy past_keys inputs without PyTorch installed.")
+            else:
+                import torch
+
+                batch, seqlen = common_inputs["input_ids"].shape
+                # Not using the same length for past_key_values
+                past_key_values_length = seqlen + 2
+                past_shape = (
+                    batch,
+                    self.num_attention_heads,
+                    past_key_values_length,
+                    self._config.hidden_size // self.num_attention_heads,
+                )
+                ordered_inputs["past_key_values"] = [
+                    (torch.zeros(past_shape), torch.zeros(past_shape)) for _ in range(self.num_layers)
+                ]
+
+        ordered_inputs["attention_mask"] = common_inputs["attention_mask"]
+        if self.use_past:
+            mask_dtype = ordered_inputs["attention_mask"].dtype
+            ordered_inputs["attention_mask"] = torch.cat(
+                [ordered_inputs["attention_mask"], torch.ones(batch, past_key_values_length, dtype=mask_dtype)], dim=1
+            )
+
+        return ordered_inputs
+
+    @property
+    def default_onnx_opset(self) -> int:
+        return 13

parrot/lib/python3.10/site-packages/transformers/models/codegen/modeling_codegen.py

@@ -0,0 +1,724 @@
+# coding=utf-8
+# Copyright 2022 Salesforce authors, The EleutherAI, and HuggingFace Teams. All rights reserved.
+#
+# 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.
+""" PyTorch CodeGen model."""
+
+from typing import Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import CrossEntropyLoss
+
+from ...activations import ACT2FN
+from ...modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
+from ...modeling_utils import PreTrainedModel
+from ...utils import add_code_sample_docstrings, add_start_docstrings, add_start_docstrings_to_model_forward, logging
+from .configuration_codegen import CodeGenConfig
+
+
+logger = logging.get_logger(__name__)
+
+_CHECKPOINT_FOR_DOC = "Salesforce/codegen-2B-mono"
+_CONFIG_FOR_DOC = "CodeGenConfig"
+
+
+# Copied from transformers.models.gptj.modeling_gptj.create_sinusoidal_positions
+def create_sinusoidal_positions(num_pos: int, dim: int) -> torch.Tensor:
+    inv_freq = 1.0 / (10000 ** (torch.arange(0, dim, 2, dtype=torch.int64) / dim))
+    sinusoid_inp = torch.einsum("i , j -> i j", torch.arange(num_pos, dtype=torch.int64).float(), inv_freq).float()
+    return torch.cat((torch.sin(sinusoid_inp), torch.cos(sinusoid_inp)), dim=1)
+
+
+# Copied from transformers.models.gptj.modeling_gptj.rotate_every_two
+def rotate_every_two(x: torch.Tensor) -> torch.Tensor:
+    x1 = x[:, :, :, ::2]
+    x2 = x[:, :, :, 1::2]
+    x = torch.stack((-x2, x1), dim=-1)
+    return x.flatten(-2)  # in einsum notation: rearrange(x, '... d j -> ... (d j)')
+
+
+# Copied from transformers.models.gptj.modeling_gptj.apply_rotary_pos_emb
+def apply_rotary_pos_emb(tensor: torch.Tensor, sin: torch.Tensor, cos: torch.Tensor) -> torch.Tensor:
+    sin = torch.repeat_interleave(sin[:, :, None, :], 2, 3)
+    cos = torch.repeat_interleave(cos[:, :, None, :], 2, 3)
+    return (tensor * cos) + (rotate_every_two(tensor) * sin)
+
+
+class CodeGenAttention(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+
+        max_positions = config.max_position_embeddings
+        self.register_buffer(
+            "causal_mask",
+            torch.tril(torch.ones((max_positions, max_positions), dtype=torch.bool)).view(
+                1, 1, max_positions, max_positions
+            ),
+            persistent=False,
+        )
+
+        self.attn_dropout = nn.Dropout(config.attn_pdrop)
+        self.resid_dropout = nn.Dropout(config.resid_pdrop)
+
+        self.embed_dim = config.hidden_size
+        self.num_attention_heads = config.num_attention_heads
+        self.head_dim = self.embed_dim // self.num_attention_heads
+        if self.head_dim * self.num_attention_heads != self.embed_dim:
+            raise ValueError(
+                f"embed_dim must be divisible by num_attention_heads (got `embed_dim`: {self.embed_dim} and"
+                f" `num_attention_heads`: {self.num_attention_heads})."
+            )
+        self.scale_attn = torch.sqrt(torch.tensor(self.head_dim, dtype=torch.float32)).to(torch.get_default_dtype())
+        self.qkv_proj = nn.Linear(self.embed_dim, self.embed_dim * 3, bias=False)
+
+        self.out_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=False)
+        self.rotary_dim = config.rotary_dim
+        pos_embd_dim = self.rotary_dim or self.embed_dim
+        self.embed_positions = create_sinusoidal_positions(max_positions, pos_embd_dim)
+
+    def _split_heads(self, x, n_head, dim_head, mp_num):
+        reshaped = x.reshape(x.shape[:-1] + (n_head // mp_num, dim_head))
+        reshaped = reshaped.reshape(x.shape[:-2] + (-1,) + reshaped.shape[-1:])
+        return reshaped
+
+    def _merge_heads(self, tensor, num_attention_heads, attn_head_size):
+        """
+        Merges attn_head_size dim and num_attn_heads dim into n_ctx
+        """
+        if len(tensor.shape) == 5:
+            tensor = tensor.permute(0, 1, 3, 2, 4).contiguous()
+        elif len(tensor.shape) == 4:
+            tensor = tensor.permute(0, 2, 1, 3).contiguous()
+        else:
+            raise ValueError(f"Input tensor rank should be one of [4, 5], but is: {len(tensor.shape)}")
+        new_shape = tensor.size()[:-2] + (num_attention_heads * attn_head_size,)
+        return tensor.view(new_shape)
+
+    def _attn(
+        self,
+        query,
+        key,
+        value,
+        attention_mask=None,
+        head_mask=None,
+    ):
+        # compute causal mask from causal mask buffer
+        query_length, key_length = query.size(-2), key.size(-2)
+        causal_mask = self.causal_mask[:, :, key_length - query_length : key_length, :key_length]
+
+        # Keep the attention weights computation in fp32 to avoid overflow issues
+        query = query.to(torch.float32)
+        key = key.to(torch.float32)
+
+        attn_weights = torch.matmul(query, key.transpose(-1, -2))
+
+        attn_weights = attn_weights / self.scale_attn
+        mask_value = torch.finfo(attn_weights.dtype).min
+        # Need to be a tensor, otherwise we get error: `RuntimeError: expected scalar type float but found double`.
+        # Need to be on the same device, otherwise `RuntimeError: ..., x and y to be on the same device`
+        mask_value = torch.tensor(mask_value, dtype=attn_weights.dtype).to(attn_weights.device)
+        attn_weights = torch.where(causal_mask, attn_weights, mask_value)
+
+        if attention_mask is not None:
+            # Apply the attention mask
+            attn_weights = attn_weights + attention_mask
+
+        attn_weights = nn.Softmax(dim=-1)(attn_weights)
+        attn_weights = attn_weights.to(value.dtype)
+        attn_weights = self.attn_dropout(attn_weights)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            attn_weights = attn_weights * head_mask
+
+        attn_output = torch.matmul(attn_weights, value)
+
+        return attn_output, attn_weights
+
+    def forward(
+        self,
+        hidden_states: Optional[torch.FloatTensor],
+        layer_past: Optional[Tuple[torch.Tensor]] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = False,
+        output_attentions: Optional[bool] = False,
+    ) -> Union[
+        Tuple[torch.Tensor, Tuple[torch.Tensor]],
+        Optional[Tuple[torch.Tensor, Tuple[torch.Tensor], Tuple[torch.Tensor, ...]]],
+    ]:
+        qkv = self.qkv_proj(hidden_states)
+        # TODO(enijkamp): factor out number of logical TPU-v4 cores or make forward pass agnostic
+        mp_num = 4
+        qkv_split = qkv.reshape(qkv.shape[:-1] + (mp_num, -1))
+
+        local_dim = self.head_dim * self.num_attention_heads // mp_num
+        query, value, key = torch.split(qkv_split, local_dim, dim=-1)
+        query = self._split_heads(query, self.num_attention_heads, self.head_dim, mp_num=mp_num)
+        key = self._split_heads(key, self.num_attention_heads, self.head_dim, mp_num=mp_num)
+
+        value = self._split_heads(value, self.num_attention_heads, self.head_dim, mp_num=mp_num)
+        value = value.permute(0, 2, 1, 3)
+
+        embed_positions = self.embed_positions
+        if embed_positions.device != position_ids.device:
+            embed_positions = embed_positions.to(position_ids.device)
+            self.embed_positions = embed_positions
+
+        sincos = embed_positions[position_ids]
+        sin, cos = torch.split(sincos, sincos.shape[-1] // 2, dim=-1)
+
+        if self.rotary_dim is not None:
+            k_rot = key[:, :, :, : self.rotary_dim]
+            k_pass = key[:, :, :, self.rotary_dim :]
+
+            q_rot = query[:, :, :, : self.rotary_dim]
+            q_pass = query[:, :, :, self.rotary_dim :]
+
+            k_rot = apply_rotary_pos_emb(k_rot, sin, cos)
+            q_rot = apply_rotary_pos_emb(q_rot, sin, cos)
+
+            key = torch.cat([k_rot, k_pass], dim=-1)
+            query = torch.cat([q_rot, q_pass], dim=-1)
+        else:
+            key = apply_rotary_pos_emb(key, sin, cos)
+            query = apply_rotary_pos_emb(query, sin, cos)
+
+        key = key.permute(0, 2, 1, 3)
+        query = query.permute(0, 2, 1, 3)
+
+        if layer_past is not None:
+            past_key = layer_past[0]
+            past_value = layer_past[1]
+            key = torch.cat((past_key, key), dim=-2)
+            value = torch.cat((past_value, value), dim=-2)
+
+        if use_cache is True:
+            # Note that this cast is quite ugly, but is not implemented before ROPE as k_rot in the original codebase is always in fp32.
+            # Reference: https://github.com/salesforce/CodeGen/blob/f210c3bb1216c975ad858cd4132c0fdeabf4bfc2/codegen1/jaxformer/hf/codegen/modeling_codegen.py#L38
+            present = (key.to(hidden_states.dtype), value)
+        else:
+            present = None
+
+        # compute self-attention: V x Softmax(QK^T)
+        attn_output, attn_weights = self._attn(query, key, value, attention_mask, head_mask)
+
+        attn_output = self._merge_heads(attn_output, self.num_attention_heads, self.head_dim)
+        attn_output = self.out_proj(attn_output)
+        attn_output = self.resid_dropout(attn_output)
+
+        outputs = (attn_output, present)
+        if output_attentions:
+            outputs += (attn_weights,)
+
+        return outputs  # a, present, (attentions)
+
+
+# Copied from transformers.models.gptj.modeling_gptj.GPTJMLP with GPTJ->CodeGen
+class CodeGenMLP(nn.Module):
+    def __init__(self, intermediate_size, config):  # in MLP: intermediate_size= 4 * embed_dim
+        super().__init__()
+        embed_dim = config.n_embd
+
+        self.fc_in = nn.Linear(embed_dim, intermediate_size)
+        self.fc_out = nn.Linear(intermediate_size, embed_dim)
+
+        self.act = ACT2FN[config.activation_function]
+        self.dropout = nn.Dropout(config.resid_pdrop)
+
+    def forward(self, hidden_states: Optional[torch.FloatTensor]) -> torch.FloatTensor:
+        hidden_states = self.fc_in(hidden_states)
+        hidden_states = self.act(hidden_states)
+        hidden_states = self.fc_out(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        return hidden_states
+
+
+# Copied from transformers.models.gptj.modeling_gptj.GPTJBlock with GPTJ->CodeGen
+class CodeGenBlock(nn.Module):
+    # Ignore copy
+    def __init__(self, config):
+        super().__init__()
+        inner_dim = config.n_inner if config.n_inner is not None else 4 * config.n_embd
+        self.ln_1 = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
+        self.attn = CodeGenAttention(config)
+        self.mlp = CodeGenMLP(inner_dim, config)
+
+    def forward(
+        self,
+        hidden_states: Optional[torch.FloatTensor],
+        layer_past: Optional[Tuple[torch.Tensor]] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = False,
+        output_attentions: Optional[bool] = False,
+    ) -> Union[Tuple[torch.Tensor], Optional[Tuple[torch.Tensor, Tuple[torch.FloatTensor, ...]]]]:
+        residual = hidden_states
+        hidden_states = self.ln_1(hidden_states)
+        attn_outputs = self.attn(
+            hidden_states=hidden_states,
+            layer_past=layer_past,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+        )
+        attn_output = attn_outputs[0]  # output_attn: a, present, (attentions)
+        outputs = attn_outputs[1:]
+
+        feed_forward_hidden_states = self.mlp(hidden_states)
+        hidden_states = attn_output + feed_forward_hidden_states + residual
+
+        if use_cache:
+            outputs = (hidden_states,) + outputs
+        else:
+            outputs = (hidden_states,) + outputs[1:]
+
+        return outputs  # hidden_states, present, (attentions)
+
+
+class CodeGenPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = CodeGenConfig
+    base_model_prefix = "transformer"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["CodeGenBlock"]
+    _skip_keys_device_placement = "past_key_values"
+
+    def __init__(self, *inputs, **kwargs):
+        super().__init__(*inputs, **kwargs)
+
+    def _init_weights(self, module):
+        """Initialize the weights."""
+        if isinstance(module, (nn.Linear,)):
+            # Slightly different from Mesh Transformer JAX which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+        elif isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+
+
+CODEGEN_START_DOCSTRING = r"""
+    This model is a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) sub-class. Use
+    it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and
+    behavior.
+
+    Parameters:
+        config ([`CodeGenConfig`]): Model configuration class with all the parameters of the model.
+            Initializing with a config file does not load the weights associated with the model, only the
+            configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+CODEGEN_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `({0})`):
+            Indices of input sequence tokens in the vocabulary.
+
+            Indices can be obtained using [`AutoProcenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`torch.FloatTensor` of shape `({0})`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+        token_type_ids (`torch.LongTensor` of shape `({0})`, *optional*):
+            Segment token indices to indicate first and second portions of the inputs. Indices are selected in `[0,
+            1]`:
+
+            - 0 corresponds to a *sentence A* token,
+            - 1 corresponds to a *sentence B* token.
+
+            [What are token type IDs?](../glossary#token-type-ids)
+        position_ids (`torch.LongTensor` of shape `({0})`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.n_positions - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        head_mask (`torch.FloatTensor` of shape `(num_attention_heads,)` or `(n_layer, num_attention_heads)`, *optional*):
+            Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        inputs_embeds (`torch.FloatTensor` of shape `({0}, hidden_dim)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
+            is useful if you want more control over how to convert *input_ids* indices into associated vectors than the
+            model's internal embedding lookup matrix.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+@add_start_docstrings(
+    "The bare CodeGen Model transformer outputting raw hidden-states without any specific head on top.",
+    CODEGEN_START_DOCSTRING,
+)
+class CodeGenModel(CodeGenPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.embed_dim = config.n_embd
+        self.vocab_size = config.vocab_size
+        self.wte = nn.Embedding(config.vocab_size, self.embed_dim)
+        self.drop = nn.Dropout(config.embd_pdrop)
+        self.h = nn.ModuleList([CodeGenBlock(config) for _ in range(config.n_layer)])
+        self.ln_f = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_epsilon)
+        self.rotary_dim = min(config.rotary_dim, config.n_ctx // config.num_attention_heads)
+
+        self.gradient_checkpointing = False
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.wte
+
+    def set_input_embeddings(self, new_embeddings):
+        self.wte = new_embeddings
+
+    @add_start_docstrings_to_model_forward(CODEGEN_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=BaseModelOutputWithPast,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPast]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask)
+            input_shape = input_ids.size()
+            input_ids = input_ids.view(-1, input_shape[-1])
+            batch_size = input_ids.shape[0]
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+            batch_size = inputs_embeds.shape[0]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+
+        if token_type_ids is not None:
+            token_type_ids = token_type_ids.view(-1, input_shape[-1])
+
+        if past_key_values is None:
+            past_length = 0
+            past_key_values = tuple([None] * len(self.h))
+        else:
+            past_length = past_key_values[0][0].size(-2)
+
+        if position_ids is None:
+            position_ids = torch.arange(past_length, input_shape[-1] + past_length, dtype=torch.long, device=device)
+            position_ids = position_ids.unsqueeze(0)
+
+        # Attention mask.
+        if attention_mask is not None:
+            if batch_size <= 0:
+                raise ValueError("batch_size has to be defined and > 0")
+            attention_mask = attention_mask.view(batch_size, -1)
+            # We create a 3D attention mask from a 2D tensor mask.
+            # Sizes are [batch_size, 1, 1, to_seq_length]
+            # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]
+            # this attention mask is more simple than the triangular masking of causal attention
+            # used in OpenAI GPT, we just need to prepare the broadcast dimension here.
+            attention_mask = attention_mask[:, None, None, :]
+
+            # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
+            # masked positions, this operation will create a tensor which is 0.0 for
+            # positions we want to attend and the dtype's smallest value for masked positions.
+            # Since we are adding it to the raw scores before the softmax, this is
+            # effectively the same as removing these entirely.
+            attention_mask = attention_mask.to(dtype=self.dtype)  # fp16 compatibility
+            attention_mask = (1.0 - attention_mask) * torch.finfo(self.dtype).min
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x num_attention_heads x N x N
+        # head_mask has shape n_layer x batch x num_attention_heads x N x N
+        head_mask = self.get_head_mask(head_mask, self.config.n_layer)
+
+        if inputs_embeds is None:
+            inputs_embeds = self.wte(input_ids)
+
+        hidden_states = inputs_embeds
+
+        if token_type_ids is not None:
+            token_type_embeds = self.wte(token_type_ids)
+            hidden_states = hidden_states + token_type_embeds
+
+        hidden_states = self.drop(hidden_states)
+
+        output_shape = input_shape + (hidden_states.size(-1),)
+
+        if self.gradient_checkpointing and self.training:
+            if use_cache:
+                logger.warning_once(
+                    "`use_cache=True` is incompatible with `config.gradient_checkpointing=True`. Setting "
+                    "`use_cache=False`..."
+                )
+                use_cache = False
+
+        presents = () if use_cache else None
+        all_self_attentions = () if output_attentions else None
+        all_hidden_states = () if output_hidden_states else None
+        for i, (block, layer_past) in enumerate(zip(self.h, past_key_values)):
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+                outputs = self._gradient_checkpointing_func(
+                    block.__call__,
+                    hidden_states,
+                    None,
+                    attention_mask,
+                    position_ids,
+                    head_mask[i],
+                    use_cache,
+                    output_attentions,
+                )
+            else:
+                outputs = block(
+                    hidden_states=hidden_states,
+                    layer_past=layer_past,
+                    attention_mask=attention_mask,
+                    position_ids=position_ids,
+                    head_mask=head_mask[i],
+                    use_cache=use_cache,
+                    output_attentions=output_attentions,
+                )
+
+            hidden_states = outputs[0]
+            if use_cache is True:
+                presents = presents + (outputs[1],)
+
+            if output_attentions:
+                all_self_attentions = all_self_attentions + (outputs[2 if use_cache else 1],)
+
+        hidden_states = self.ln_f(hidden_states)
+
+        hidden_states = hidden_states.view(output_shape)
+        # Add last hidden state
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, presents, all_hidden_states, all_self_attentions] if v is not None)
+
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=presents,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    The CodeGen Model transformer with a language modeling head on top.
+    """,
+    CODEGEN_START_DOCSTRING,
+)
+class CodeGenForCausalLM(CodeGenPreTrainedModel):
+    _tied_weights_keys = ["lm_head.weight"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.transformer = CodeGenModel(config)
+        self.lm_head = nn.Linear(config.n_embd, config.vocab_size)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def prepare_inputs_for_generation(self, input_ids, inputs_embeds=None, past_key_values=None, **kwargs):
+        token_type_ids = kwargs.get("token_type_ids", None)
+        # Omit tokens covered by past_key_values
+        if past_key_values:
+            past_length = past_key_values[0][0].shape[2]
+
+            # Some generation methods already pass only the last input ID
+            if input_ids.shape[1] > past_length:
+                remove_prefix_length = past_length
+            else:
+                # Default to old behavior: keep only final ID
+                remove_prefix_length = input_ids.shape[1] - 1
+
+            input_ids = input_ids[:, remove_prefix_length:]
+            if token_type_ids is not None:
+                token_type_ids = token_type_ids[:, -input_ids.shape[1] :]
+
+        attention_mask = kwargs.get("attention_mask", None)
+        position_ids = kwargs.get("position_ids", None)
+
+        if attention_mask is not None and position_ids is None:
+            # create position_ids on the fly for batch generation
+            position_ids = attention_mask.long().cumsum(-1) - 1
+            position_ids.masked_fill_(attention_mask == 0, 1)
+            if past_key_values:
+                position_ids = position_ids[:, -input_ids.shape[1] :]
+
+        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
+        if inputs_embeds is not None and past_key_values is None:
+            model_inputs = {"inputs_embeds": inputs_embeds}
+        else:
+            model_inputs = {"input_ids": input_ids.contiguous()}
+
+        model_inputs.update(
+            {
+                "past_key_values": past_key_values,
+                "use_cache": kwargs.get("use_cache"),
+                "position_ids": position_ids,
+                "attention_mask": attention_mask,
+                "token_type_ids": token_type_ids,
+            }
+        )
+        return model_inputs
+
+    @add_start_docstrings_to_model_forward(CODEGEN_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=CausalLMOutputWithPast,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
+            `labels = input_ids` Indices are selected in `[-100, 0, ..., config.vocab_size]` All labels set to `-100`
+            are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]`
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        transformer_outputs = self.transformer(
+            input_ids,
+            past_key_values=past_key_values,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        hidden_states = transformer_outputs[0]
+
+        # make sure sampling in fp16 works correctly and
+        # compute loss in fp32 to match with mesh-tf version
+        # https://github.com/EleutherAI/gpt-neo/blob/89ce74164da2fb16179106f54e2269b5da8db333/models/gpt2/gpt2.py#L179
+        lm_logits = self.lm_head(hidden_states).to(torch.float32)
+
+        loss = None
+        if labels is not None:
+            # move labels to correct device to enable model parallelism
+            labels = labels.to(lm_logits.device)
+            # Shift so that tokens < n predict n
+            shift_logits = lm_logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
+
+            loss = loss.to(hidden_states.dtype)
+
+        if not return_dict:
+            output = (lm_logits,) + transformer_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=lm_logits,
+            past_key_values=transformer_outputs.past_key_values,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
+
+    @staticmethod
+    def _reorder_cache(
+        past_key_values: Tuple[Tuple[torch.Tensor]], beam_idx: torch.Tensor
+    ) -> Tuple[Tuple[torch.Tensor]]:
+        """
+        This function is used to re-order the `past_key_values` cache if [`~PretrainedModel.beam_search`] or
+        [`~PretrainedModel.beam_sample`] is called. This is required to match `past_key_values` with the correct
+        beam_idx at every generation step.
+        """
+        return tuple(
+            tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past)
+            for layer_past in past_key_values
+        )

parrot/lib/python3.10/site-packages/transformers/models/codegen/tokenization_codegen.py

@@ -0,0 +1,417 @@
+# coding=utf-8
+# Copyright 2022 The Salesforce authors, The Open AI Team Authors and The HuggingFace Inc. 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.
+"""Tokenization classes for CodeGen"""
+
+
+import json
+import os
+from functools import lru_cache
+from typing import TYPE_CHECKING, List, Optional, Tuple, Union
+
+import numpy as np
+import regex as re
+
+from ...utils import is_tf_available, is_torch_available, logging, to_py_obj
+
+
+if TYPE_CHECKING:
+    if is_torch_available():
+        import torch
+    if is_tf_available():
+        import tensorflow as tf
+
+from ...tokenization_utils import AddedToken, PreTrainedTokenizer
+
+
+logger = logging.get_logger(__name__)
+
+VOCAB_FILES_NAMES = {
+    "vocab_file": "vocab.json",
+    "merges_file": "merges.txt",
+}
+
+
+@lru_cache()
+def bytes_to_unicode():
+    """
+    Returns list of utf-8 byte and a mapping to unicode strings. We specifically avoids mapping to whitespace/control
+    characters the bpe code barfs on.
+
+    The reversible bpe codes work on unicode strings. This means you need a large # of unicode characters in your vocab
+    if you want to avoid UNKs. When you're at something like a 10B token dataset you end up needing around 5K for
+    decent coverage. This is a significant percentage of your normal, say, 32K bpe vocab. To avoid that, we want lookup
+    tables between utf-8 bytes and unicode strings.
+    """
+    bs = (
+        list(range(ord("!"), ord("~") + 1)) + list(range(ord("¡"), ord("¬") + 1)) + list(range(ord("®"), ord("ÿ") + 1))
+    )
+    cs = bs[:]
+    n = 0
+    for b in range(2**8):
+        if b not in bs:
+            bs.append(b)
+            cs.append(2**8 + n)
+            n += 1
+    cs = [chr(n) for n in cs]
+    return dict(zip(bs, cs))
+
+
+def get_pairs(word):
+    """
+    Return set of symbol pairs in a word.
+
+    Word is represented as tuple of symbols (symbols being variable-length strings).
+    """
+    pairs = set()
+    prev_char = word[0]
+    for char in word[1:]:
+        pairs.add((prev_char, char))
+        prev_char = char
+    return pairs
+
+
+class CodeGenTokenizer(PreTrainedTokenizer):
+    """
+    Construct a CodeGen tokenizer. Based on byte-level Byte-Pair-Encoding.
+
+    This tokenizer has been trained to treat spaces like parts of the tokens (a bit like sentencepiece) so a word will
+    be encoded differently whether it is at the beginning of the sentence (without space) or not:
+
+    ```python
+    >>> from transformers import CodeGenTokenizer
+
+    >>> tokenizer = CodeGenTokenizer.from_pretrained("Salesforce/codegen-350M-mono")
+    >>> tokenizer("Hello world")["input_ids"]
+    [15496, 995]
+
+    >>> tokenizer(" Hello world")["input_ids"]
+    [18435, 995]
+    ```
+
+    You can get around that behavior by passing `add_prefix_space=True` when instantiating this tokenizer or when you
+    call it on some text, but since the model was not pretrained this way, it might yield a decrease in performance.
+
+    <Tip>
+
+    When used with `is_split_into_words=True`, this tokenizer will add a space before each word (even the first one).
+
+    </Tip>
+
+    This tokenizer inherits from [`PreTrainedTokenizer`] which contains most of the main methods. Users should refer to
+    this superclass for more information regarding those methods.
+
+    Args:
+        vocab_file (`str`):
+            Path to the vocabulary file.
+        merges_file (`str`):
+            Path to the merges file.
+        errors (`str`, *optional*, defaults to `"replace"`):
+            Paradigm to follow when decoding bytes to UTF-8. See
+            [bytes.decode](https://docs.python.org/3/library/stdtypes.html#bytes.decode) for more information.
+        unk_token (`str`, *optional*, defaults to `"<|endoftext|>"`):
+            The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this
+            token instead.
+        bos_token (`str`, *optional*, defaults to `"<|endoftext|>"`):
+            The beginning of sequence token.
+        eos_token (`str`, *optional*, defaults to `"<|endoftext|>"`):
+            The end of sequence token.
+        pad_token (`str`, *optional*):
+            The token used for padding, for example when batching sequences of different lengths.
+        add_prefix_space (`bool`, *optional*, defaults to `False`):
+            Whether or not to add an initial space to the input. This allows to treat the leading word just as any
+            other word. (CodeGen tokenizer detect beginning of words by the preceding space).
+        add_bos_token (`bool`, *optional*, defaults to `False`):
+            Whether to add a beginning of sequence token at the start of sequences.
+        return_token_type_ids (`bool`, *optional*, defaults to `False`):
+            Whether to return token type IDs.
+    """
+
+    vocab_files_names = VOCAB_FILES_NAMES
+    model_input_names = ["input_ids", "attention_mask"]
+
+    def __init__(
+        self,
+        vocab_file,
+        merges_file,
+        errors="replace",
+        unk_token="<|endoftext|>",
+        bos_token="<|endoftext|>",
+        eos_token="<|endoftext|>",
+        pad_token=None,
+        add_prefix_space=False,
+        add_bos_token=False,
+        return_token_type_ids=False,
+        **kwargs,
+    ):
+        bos_token = AddedToken(bos_token, special=True) if isinstance(bos_token, str) else bos_token
+        eos_token = AddedToken(eos_token, special=True) if isinstance(eos_token, str) else eos_token
+        unk_token = AddedToken(unk_token, special=True) if isinstance(unk_token, str) else unk_token
+        pad_token = AddedToken(pad_token, special=True) if isinstance(pad_token, str) else pad_token
+        self.add_bos_token = add_bos_token
+        self.return_token_type_ids = return_token_type_ids
+        if self.return_token_type_ids:
+            self.model_input_names.append("token_type_ids")
+
+        with open(vocab_file, encoding="utf-8") as vocab_handle:
+            self.encoder = json.load(vocab_handle)
+        self.decoder = {v: k for k, v in self.encoder.items()}
+        self.errors = errors  # how to handle errors in decoding
+        self.byte_encoder = bytes_to_unicode()
+        self.byte_decoder = {v: k for k, v in self.byte_encoder.items()}
+        with open(merges_file, encoding="utf-8") as merges_handle:
+            bpe_merges = merges_handle.read().split("\n")[1:-1]
+        bpe_merges = [tuple(merge.split()) for merge in bpe_merges]
+        self.bpe_ranks = dict(zip(bpe_merges, range(len(bpe_merges))))
+        self.cache = {}
+        self.add_prefix_space = add_prefix_space
+
+        # Should have added re.IGNORECASE so BPE merges can happen for capitalized versions of contractions
+        self.pat = re.compile(r"""'s|'t|'re|'ve|'m|'ll|'d| ?\p{L}+| ?\p{N}+| ?[^\s\p{L}\p{N}]+|\s+(?!\S)|\s+""")
+        super().__init__(
+            errors=errors,
+            unk_token=unk_token,
+            bos_token=bos_token,
+            eos_token=eos_token,
+            pad_token=pad_token,
+            add_prefix_space=add_prefix_space,
+            add_bos_token=add_bos_token,
+            return_token_type_ids=return_token_type_ids,
+            **kwargs,
+        )
+
+    @property
+    def vocab_size(self):
+        return len(self.encoder)
+
+    def get_vocab(self):
+        return dict(self.encoder, **self.added_tokens_encoder)
+
+    def bpe(self, token):
+        if token in self.cache:
+            return self.cache[token]
+        word = tuple(token)
+        pairs = get_pairs(word)
+
+        if not pairs:
+            return token
+
+        while True:
+            bigram = min(pairs, key=lambda pair: self.bpe_ranks.get(pair, float("inf")))
+            if bigram not in self.bpe_ranks:
+                break
+            first, second = bigram
+            new_word = []
+            i = 0
+            while i < len(word):
+                try:
+                    j = word.index(first, i)
+                except ValueError:
+                    new_word.extend(word[i:])
+                    break
+                else:
+                    new_word.extend(word[i:j])
+                    i = j
+
+                if word[i] == first and i < len(word) - 1 and word[i + 1] == second:
+                    new_word.append(first + second)
+                    i += 2
+                else:
+                    new_word.append(word[i])
+                    i += 1
+            new_word = tuple(new_word)
+            word = new_word
+            if len(word) == 1:
+                break
+            else:
+                pairs = get_pairs(word)
+        word = " ".join(word)
+        self.cache[token] = word
+        return word
+
+    def build_inputs_with_special_tokens(self, token_ids_0, token_ids_1=None):
+        if self.add_bos_token:
+            bos_token_ids = [self.bos_token_id]
+        else:
+            bos_token_ids = []
+
+        output = bos_token_ids + token_ids_0
+
+        if token_ids_1 is None:
+            return output
+
+        return output + bos_token_ids + token_ids_1
+
+    def _tokenize(self, text):
+        """Tokenize a string."""
+        bpe_tokens = []
+        for token in re.findall(self.pat, text):
+            token = "".join(
+                self.byte_encoder[b] for b in token.encode("utf-8")
+            )  # Maps all our bytes to unicode strings, avoiding control tokens of the BPE (spaces in our case)
+            bpe_tokens.extend(bpe_token for bpe_token in self.bpe(token).split(" "))
+        return bpe_tokens
+
+    def _convert_token_to_id(self, token):
+        """Converts a token (str) in an id using the vocab."""
+        return self.encoder.get(token, self.encoder.get(self.unk_token))
+
+    def _convert_id_to_token(self, index):
+        """Converts an index (integer) in a token (str) using the vocab."""
+        return self.decoder.get(index)
+
+    def convert_tokens_to_string(self, tokens):
+        """Converts a sequence of tokens (string) in a single string."""
+        text = "".join(tokens)
+        text = bytearray([self.byte_decoder[c] for c in text]).decode("utf-8", errors=self.errors)
+        return text
+
+    def create_token_type_ids_from_sequences(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
+    ) -> List[int]:
+        """
+        Create a mask from the two sequences passed to be used in a sequence-pair classification task. A sequence
+        pair mask has the following format:
+
+        ```
+        0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1
+        | first sequence    | second sequence |
+        ```
+
+        If `token_ids_1` is `None`, this method only returns the first portion of the mask (0s).
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            `List[int]`: List of [token type IDs](../glossary#token-type-ids) according to the given sequence(s).
+        """
+        sep = [self.sep_token_id] if self.sep_token_id is not None else []
+        cls = [self.cls_token_id] if self.sep_token_id is not None else []
+        if token_ids_1 is None:
+            return len(cls + token_ids_0 + sep) * [0]
+        return len(cls + token_ids_0 + sep) * [0] + len(token_ids_1 + sep) * [1]
+
+    def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
+        if not os.path.isdir(save_directory):
+            logger.error(f"Vocabulary path ({save_directory}) should be a directory")
+            return
+        vocab_file = os.path.join(
+            save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["vocab_file"]
+        )
+        merge_file = os.path.join(
+            save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["merges_file"]
+        )
+
+        with open(vocab_file, "w", encoding="utf-8") as f:
+            f.write(json.dumps(self.encoder, indent=2, sort_keys=True, ensure_ascii=False) + "\n")
+
+        index = 0
+        with open(merge_file, "w", encoding="utf-8") as writer:
+            writer.write("#version: 0.2\n")
+            for bpe_tokens, token_index in sorted(self.bpe_ranks.items(), key=lambda kv: kv[1]):
+                if index != token_index:
+                    logger.warning(
+                        f"Saving vocabulary to {merge_file}: BPE merge indices are not consecutive."
+                        " Please check that the tokenizer is not corrupted!"
+                    )
+                    index = token_index
+                writer.write(" ".join(bpe_tokens) + "\n")
+                index += 1
+
+        return vocab_file, merge_file
+
+    def prepare_for_tokenization(self, text, is_split_into_words=False, **kwargs):
+        add_prefix_space = kwargs.pop("add_prefix_space", self.add_prefix_space)
+        if is_split_into_words or add_prefix_space:
+            text = " " + text
+        return (text, kwargs)
+
+    def decode(
+        self,
+        token_ids: Union[int, List[int], "np.ndarray", "torch.Tensor", "tf.Tensor"],
+        skip_special_tokens: bool = False,
+        clean_up_tokenization_spaces: bool = None,
+        truncate_before_pattern: Optional[List[str]] = None,
+        **kwargs,
+    ) -> str:
+        """
+        Converts a sequence of ids in a string, using the tokenizer and vocabulary with options to remove special
+        tokens and clean up tokenization spaces.
+
+        Similar to doing `self.convert_tokens_to_string(self.convert_ids_to_tokens(token_ids))`.
+
+        Args:
+            token_ids (`Union[int, List[int], np.ndarray, torch.Tensor, tf.Tensor]`):
+                List of tokenized input ids. Can be obtained using the `__call__` method.
+            skip_special_tokens (`bool`, *optional*, defaults to `False`):
+                Whether or not to remove special tokens in the decoding.
+            clean_up_tokenization_spaces (`bool`, *optional*):
+                Whether or not to clean up the tokenization spaces. If `None`, will default to
+                `self.clean_up_tokenization_spaces` (available in the `tokenizer_config`).
+            truncate_before_pattern (`List[str]`, *optional*, defaults to `None`):
+                A list of regular expression strings that will be used to truncate the returned string. This can be
+                used to remove extra pieces of code (e.g. truncate if observing a comment symbol "#" at the beginning
+                of a new line). An example pattern could be `["^#", re.escape("<|endoftext|>"), "^'''", "\n\n\n"]`.
+            kwargs (additional keyword arguments, *optional*):
+                Will be passed to the underlying model specific decode method.
+
+        Returns:
+            `str`: The decoded sentence.
+        """
+
+        token_ids = to_py_obj(token_ids)
+
+        decoded_text = super()._decode(
+            token_ids=token_ids,
+            skip_special_tokens=skip_special_tokens,
+            clean_up_tokenization_spaces=clean_up_tokenization_spaces,
+            **kwargs,
+        )
+
+        if truncate_before_pattern is not None and len(truncate_before_pattern) > 0:
+            decoded_text = self.truncate(decoded_text, truncate_before_pattern)
+
+        return decoded_text
+
+    def truncate(self, completion, truncate_before_pattern):
+        def find_re(string, pattern, start_pos):
+            m = pattern.search(string, start_pos)
+            return m.start() if m else -1
+
+        terminals = [re.compile(pattern, re.MULTILINE) for pattern in truncate_before_pattern]
+
+        prints = list(re.finditer("^print", completion, re.MULTILINE))
+
+        if len(prints) > 1:
+            completion = completion[: prints[1].start()]
+
+        defs = list(re.finditer("^def", completion, re.MULTILINE))
+
+        if len(defs) > 1:
+            completion = completion[: defs[1].start()]
+
+        start_pos = 0
+
+        terminals_pos = [
+            pos for pos in [find_re(completion, terminal, start_pos) for terminal in terminals] if pos != -1
+        ]
+
+        if len(terminals_pos) > 0:
+            return completion[: min(terminals_pos)]
+        else:
+            return completion