Add files using upload-large-folder tool

.gitattributes

@@ -65,3 +65,4 @@ tuning-competition-baseline/.venv/lib/python3.11/site-packages/Cython/Compiler/_
 tuning-competition-baseline/.venv/lib/python3.11/site-packages/torchgen/__pycache__/model.cpython-311.pyc filter=lfs diff=lfs merge=lfs -text
 tuning-competition-baseline/.venv/lib/python3.11/site-packages/mpmath/tests/__pycache__/test_fp.cpython-311.pyc filter=lfs diff=lfs merge=lfs -text
 tuning-competition-baseline/.venv/lib/python3.11/site-packages/mpmath/__pycache__/function_docs.cpython-311.pyc filter=lfs diff=lfs merge=lfs -text
+tuning-competition-baseline/.venv/lib/python3.11/site-packages/pip/_vendor/distlib/w64-arm.exe filter=lfs diff=lfs merge=lfs -text

tuning-competition-baseline/.venv/lib/python3.11/site-packages/nvidia/curand/include/curand.h

@@ -0,0 +1,1077 @@
+
+ /* Copyright 2010-2014 NVIDIA Corporation.  All rights reserved.
+  *
+  * NOTICE TO LICENSEE:
+  *
+  * The source code and/or documentation ("Licensed Deliverables") are
+  * subject to NVIDIA intellectual property rights under U.S. and
+  * international Copyright laws.
+  *
+  * The Licensed Deliverables contained herein are PROPRIETARY and
+  * CONFIDENTIAL to NVIDIA and are 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.  THEY ARE
+  * 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 are 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(CURAND_H_)
+#define CURAND_H_
+
+/**
+ * \defgroup HOST Host API
+ *
+ * @{
+ */
+#ifndef __CUDACC_RTC__
+#include <cuda_runtime.h>
+#endif
+
+#ifndef CURANDAPI
+#ifdef _WIN32
+#define CURANDAPI __stdcall
+#else
+#define CURANDAPI
+#endif
+#endif
+
+#if defined(__cplusplus)
+extern "C" {
+#endif /* __cplusplus */
+
+#define CURAND_VER_MAJOR 10
+#define CURAND_VER_MINOR 3
+#define CURAND_VER_PATCH 0
+#define CURAND_VER_BUILD 86
+#define CURAND_VERSION (CURAND_VER_MAJOR * 1000 + \
+                        CURAND_VER_MINOR *  100 + \
+                        CURAND_VER_PATCH)
+/* CURAND Host API datatypes */
+
+/**
+ * @{
+ */
+
+/**
+ * CURAND function call status types
+ */
+enum curandStatus {
+    CURAND_STATUS_SUCCESS = 0, ///< No errors
+    CURAND_STATUS_VERSION_MISMATCH = 100, ///< Header file and linked library version do not match
+    CURAND_STATUS_NOT_INITIALIZED = 101, ///< Generator not initialized
+    CURAND_STATUS_ALLOCATION_FAILED = 102, ///< Memory allocation failed
+    CURAND_STATUS_TYPE_ERROR = 103, ///< Generator is wrong type
+    CURAND_STATUS_OUT_OF_RANGE = 104, ///< Argument out of range
+    CURAND_STATUS_LENGTH_NOT_MULTIPLE = 105, ///< Length requested is not a multple of dimension
+    CURAND_STATUS_DOUBLE_PRECISION_REQUIRED = 106, ///< GPU does not have double precision required by MRG32k3a
+    CURAND_STATUS_LAUNCH_FAILURE = 201, ///< Kernel launch failure
+    CURAND_STATUS_PREEXISTING_FAILURE = 202, ///< Preexisting failure on library entry
+    CURAND_STATUS_INITIALIZATION_FAILED = 203, ///< Initialization of CUDA failed
+    CURAND_STATUS_ARCH_MISMATCH = 204, ///< Architecture mismatch, GPU does not support requested feature
+    CURAND_STATUS_INTERNAL_ERROR = 999 ///< Internal library error
+};
+
+/*
+ * CURAND function call status types
+*/
+/** \cond UNHIDE_TYPEDEFS */
+typedef enum curandStatus curandStatus_t;
+/** \endcond */
+
+/**
+ * CURAND generator types
+ */
+enum curandRngType {
+    CURAND_RNG_TEST = 0,
+    CURAND_RNG_PSEUDO_DEFAULT = 100, ///< Default pseudorandom generator
+    CURAND_RNG_PSEUDO_XORWOW = 101, ///< XORWOW pseudorandom generator
+    CURAND_RNG_PSEUDO_MRG32K3A = 121, ///< MRG32k3a pseudorandom generator
+    CURAND_RNG_PSEUDO_MTGP32 = 141, ///< Mersenne Twister MTGP32 pseudorandom generator
+    CURAND_RNG_PSEUDO_MT19937 = 142, ///< Mersenne Twister MT19937 pseudorandom generator
+    CURAND_RNG_PSEUDO_PHILOX4_32_10 = 161, ///< PHILOX-4x32-10 pseudorandom generator
+    CURAND_RNG_QUASI_DEFAULT = 200, ///< Default quasirandom generator
+    CURAND_RNG_QUASI_SOBOL32 = 201, ///< Sobol32 quasirandom generator
+    CURAND_RNG_QUASI_SCRAMBLED_SOBOL32 = 202,  ///< Scrambled Sobol32 quasirandom generator
+    CURAND_RNG_QUASI_SOBOL64 = 203, ///< Sobol64 quasirandom generator
+    CURAND_RNG_QUASI_SCRAMBLED_SOBOL64 = 204  ///< Scrambled Sobol64 quasirandom generator
+};
+
+/*
+ * CURAND generator types
+ */
+/** \cond UNHIDE_TYPEDEFS */
+typedef enum curandRngType curandRngType_t;
+/** \endcond */
+
+/**
+ * CURAND ordering of results in memory
+ */
+enum curandOrdering {
+    CURAND_ORDERING_PSEUDO_BEST = 100, ///< Best ordering for pseudorandom results
+    CURAND_ORDERING_PSEUDO_DEFAULT = 101, ///< Specific default thread sequence for pseudorandom results, same as CURAND_ORDERING_PSEUDO_BEST
+    CURAND_ORDERING_PSEUDO_SEEDED = 102, ///< Specific seeding pattern for fast lower quality pseudorandom results
+    CURAND_ORDERING_PSEUDO_LEGACY = 103, ///< Specific legacy sequence for pseudorandom results, guaranteed to remain the same for all cuRAND release
+    CURAND_ORDERING_PSEUDO_DYNAMIC = 104, ///< Specific ordering adjusted to the device it is being executed on, provides the best performance
+    CURAND_ORDERING_QUASI_DEFAULT = 201 ///< Specific n-dimensional ordering for quasirandom results
+};
+
+/*
+ * CURAND ordering of results in memory
+ */
+/** \cond UNHIDE_TYPEDEFS */
+typedef enum curandOrdering curandOrdering_t;
+/** \endcond */
+
+/**
+ * CURAND choice of direction vector set
+ */
+enum curandDirectionVectorSet {
+    CURAND_DIRECTION_VECTORS_32_JOEKUO6 = 101, ///< Specific set of 32-bit direction vectors generated from polynomials recommended by S. Joe and F. Y. Kuo, for up to 20,000 dimensions
+    CURAND_SCRAMBLED_DIRECTION_VECTORS_32_JOEKUO6 = 102, ///< Specific set of 32-bit direction vectors generated from polynomials recommended by S. Joe and F. Y. Kuo, for up to 20,000 dimensions, and scrambled
+    CURAND_DIRECTION_VECTORS_64_JOEKUO6 = 103, ///< Specific set of 64-bit direction vectors generated from polynomials recommended by S. Joe and F. Y. Kuo, for up to 20,000 dimensions
+    CURAND_SCRAMBLED_DIRECTION_VECTORS_64_JOEKUO6 = 104 ///< Specific set of 64-bit direction vectors generated from polynomials recommended by S. Joe and F. Y. Kuo, for up to 20,000 dimensions, and scrambled
+};
+
+/*
+ * CURAND choice of direction vector set
+ */
+/** \cond UNHIDE_TYPEDEFS */
+typedef enum curandDirectionVectorSet curandDirectionVectorSet_t;
+/** \endcond */
+
+/**
+ * CURAND array of 32-bit direction vectors
+ */
+/** \cond UNHIDE_TYPEDEFS */
+typedef unsigned int curandDirectionVectors32_t[32];
+/** \endcond */
+
+ /**
+ * CURAND array of 64-bit direction vectors
+ */
+/** \cond UNHIDE_TYPEDEFS */
+typedef unsigned long long curandDirectionVectors64_t[64];
+/** \endcond **/
+
+/**
+ * CURAND generator (opaque)
+ */
+struct curandGenerator_st;
+
+/**
+ * CURAND generator
+ */
+/** \cond UNHIDE_TYPEDEFS */
+typedef struct curandGenerator_st *curandGenerator_t;
+/** \endcond */
+
+/**
+ * CURAND distribution
+ */
+/** \cond UNHIDE_TYPEDEFS */
+typedef double curandDistribution_st;
+typedef curandDistribution_st *curandDistribution_t;
+typedef struct curandDistributionShift_st *curandDistributionShift_t;
+/** \endcond */
+/**
+ * CURAND distribution M2
+ */
+/** \cond UNHIDE_TYPEDEFS */
+typedef struct curandDistributionM2Shift_st *curandDistributionM2Shift_t;
+typedef struct curandHistogramM2_st *curandHistogramM2_t;
+typedef unsigned int curandHistogramM2K_st;
+typedef curandHistogramM2K_st *curandHistogramM2K_t;
+typedef curandDistribution_st curandHistogramM2V_st;
+typedef curandHistogramM2V_st *curandHistogramM2V_t;
+
+typedef struct curandDiscreteDistribution_st *curandDiscreteDistribution_t;
+/** \endcond */
+
+/*
+ * CURAND METHOD
+ */
+/** \cond UNHIDE_ENUMS */
+enum curandMethod {
+    CURAND_CHOOSE_BEST = 0, // choose best depends on args
+    CURAND_ITR = 1,
+    CURAND_KNUTH = 2,
+    CURAND_HITR = 3,
+    CURAND_M1 = 4,
+    CURAND_M2 = 5,
+    CURAND_BINARY_SEARCH = 6,
+    CURAND_DISCRETE_GAUSS = 7,
+    CURAND_REJECTION = 8,
+    CURAND_DEVICE_API = 9,
+    CURAND_FAST_REJECTION = 10,
+    CURAND_3RD = 11,
+    CURAND_DEFINITION = 12,
+    CURAND_POISSON = 13
+};
+
+typedef enum curandMethod curandMethod_t;
+/** \endcond */
+
+
+#ifndef __CUDACC_RTC__
+
+/**
+ * @}
+ */
+
+/**
+ * \brief Create new random number generator.
+ *
+ * Creates a new random number generator of type \p rng_type
+ * and returns it in \p *generator.
+ *
+ * Legal values for \p rng_type are:
+ * - CURAND_RNG_PSEUDO_DEFAULT
+ * - CURAND_RNG_PSEUDO_XORWOW
+ * - CURAND_RNG_PSEUDO_MRG32K3A
+ * - CURAND_RNG_PSEUDO_MTGP32
+ * - CURAND_RNG_PSEUDO_MT19937
+ * - CURAND_RNG_PSEUDO_PHILOX4_32_10
+ * - CURAND_RNG_QUASI_DEFAULT
+ * - CURAND_RNG_QUASI_SOBOL32
+ * - CURAND_RNG_QUASI_SCRAMBLED_SOBOL32
+ * - CURAND_RNG_QUASI_SOBOL64
+ * - CURAND_RNG_QUASI_SCRAMBLED_SOBOL64
+ *
+ * When \p rng_type is CURAND_RNG_PSEUDO_DEFAULT, the type chosen
+ * is CURAND_RNG_PSEUDO_XORWOW.  \n
+ * When \p rng_type is CURAND_RNG_QUASI_DEFAULT,
+ * the type chosen is CURAND_RNG_QUASI_SOBOL32.
+ *
+ * The default values for \p rng_type = CURAND_RNG_PSEUDO_XORWOW are:
+ * - \p seed = 0
+ * - \p offset = 0
+ * - \p ordering = CURAND_ORDERING_PSEUDO_DEFAULT
+ *
+ * The default values for \p rng_type = CURAND_RNG_PSEUDO_MRG32K3A are:
+ * - \p seed = 0
+ * - \p offset = 0
+ * - \p ordering = CURAND_ORDERING_PSEUDO_DEFAULT
+ *
+ * The default values for \p rng_type = CURAND_RNG_PSEUDO_MTGP32 are:
+ * - \p seed = 0
+ * - \p offset = 0
+ * - \p ordering = CURAND_ORDERING_PSEUDO_DEFAULT
+ *
+ * The default values for \p rng_type = CURAND_RNG_PSEUDO_MT19937 are:
+ * - \p seed = 0
+ * - \p offset = 0
+ * - \p ordering = CURAND_ORDERING_PSEUDO_DEFAULT
+ *
+ * * The default values for \p rng_type = CURAND_RNG_PSEUDO_PHILOX4_32_10 are:
+ * - \p seed = 0
+ * - \p offset = 0
+ * - \p ordering = CURAND_ORDERING_PSEUDO_DEFAULT
+ *
+ * The default values for \p rng_type = CURAND_RNG_QUASI_SOBOL32 are:
+ * - \p dimensions = 1
+ * - \p offset = 0
+ * - \p ordering = CURAND_ORDERING_QUASI_DEFAULT
+ *
+ * The default values for \p rng_type = CURAND_RNG_QUASI_SOBOL64 are:
+ * - \p dimensions = 1
+ * - \p offset = 0
+ * - \p ordering = CURAND_ORDERING_QUASI_DEFAULT
+ *
+ * The default values for \p rng_type = CURAND_RNG_QUASI_SCRAMBBLED_SOBOL32 are:
+ * - \p dimensions = 1
+ * - \p offset = 0
+ * - \p ordering = CURAND_ORDERING_QUASI_DEFAULT
+ *
+ * The default values for \p rng_type = CURAND_RNG_QUASI_SCRAMBLED_SOBOL64 are:
+ * - \p dimensions = 1
+ * - \p offset = 0
+ * - \p ordering = CURAND_ORDERING_QUASI_DEFAULT
+ *
+ * \param generator - Pointer to generator
+ * \param rng_type - Type of generator to create
+ *
+ * \return
+ * - CURAND_STATUS_ALLOCATION_FAILED, if memory could not be allocated \n
+ * - CURAND_STATUS_INITIALIZATION_FAILED if there was a problem setting up the GPU \n
+ * - CURAND_STATUS_VERSION_MISMATCH if the header file version does not match the
+ *   dynamically linked library version \n
+ * - CURAND_STATUS_TYPE_ERROR if the value for \p rng_type is invalid \n
+ * - CURAND_STATUS_SUCCESS if generator was created successfully \n
+ *
+ */
+curandStatus_t CURANDAPI
+curandCreateGenerator(curandGenerator_t *generator, curandRngType_t rng_type);
+
+/**
+ * \brief Create new host CPU random number generator.
+ *
+ * Creates a new host CPU random number generator of type \p rng_type
+ * and returns it in \p *generator.
+ *
+ * Legal values for \p rng_type are:
+ * - CURAND_RNG_PSEUDO_DEFAULT
+ * - CURAND_RNG_PSEUDO_XORWOW
+ * - CURAND_RNG_PSEUDO_MRG32K3A
+ * - CURAND_RNG_PSEUDO_MTGP32
+ * - CURAND_RNG_PSEUDO_MT19937
+ * - CURAND_RNG_PSEUDO_PHILOX4_32_10
+ * - CURAND_RNG_QUASI_DEFAULT
+ * - CURAND_RNG_QUASI_SOBOL32
+ *
+ * When \p rng_type is CURAND_RNG_PSEUDO_DEFAULT, the type chosen
+ * is CURAND_RNG_PSEUDO_XORWOW.  \n
+ * When \p rng_type is CURAND_RNG_QUASI_DEFAULT,
+ * the type chosen is CURAND_RNG_QUASI_SOBOL32.
+ *
+ * The default values for \p rng_type = CURAND_RNG_PSEUDO_XORWOW are:
+ * - \p seed = 0
+ * - \p offset = 0
+ * - \p ordering = CURAND_ORDERING_PSEUDO_DEFAULT
+ *
+ * The default values for \p rng_type = CURAND_RNG_PSEUDO_MRG32K3A are:
+ * - \p seed = 0
+ * - \p offset = 0
+ * - \p ordering = CURAND_ORDERING_PSEUDO_DEFAULT
+ *
+ * The default values for \p rng_type = CURAND_RNG_PSEUDO_MTGP32 are:
+ * - \p seed = 0
+ * - \p offset = 0
+ * - \p ordering = CURAND_ORDERING_PSEUDO_DEFAULT
+ *
+ * The default values for \p rng_type = CURAND_RNG_PSEUDO_MT19937 are:
+ * - \p seed = 0
+ * - \p offset = 0
+ * - \p ordering = CURAND_ORDERING_PSEUDO_DEFAULT
+ *
+ * * The default values for \p rng_type = CURAND_RNG_PSEUDO_PHILOX4_32_10 are:
+ * - \p seed = 0
+ * - \p offset = 0
+ * - \p ordering = CURAND_ORDERING_PSEUDO_DEFAULT
+ *
+ * The default values for \p rng_type = CURAND_RNG_QUASI_SOBOL32 are:
+ * - \p dimensions = 1
+ * - \p offset = 0
+ * - \p ordering = CURAND_ORDERING_QUASI_DEFAULT
+ *
+ * The default values for \p rng_type = CURAND_RNG_QUASI_SOBOL64 are:
+ * - \p dimensions = 1
+ * - \p offset = 0
+ * - \p ordering = CURAND_ORDERING_QUASI_DEFAULT
+ *
+ * The default values for \p rng_type = CURAND_RNG_QUASI_SCRAMBLED_SOBOL32 are:
+ * - \p dimensions = 1
+ * - \p offset = 0
+ * - \p ordering = CURAND_ORDERING_QUASI_DEFAULT
+ *
+ * The default values for \p rng_type = CURAND_RNG_QUASI_SCRAMBLED_SOBOL64 are:
+ * - \p dimensions = 1
+ * - \p offset = 0
+ * - \p ordering = CURAND_ORDERING_QUASI_DEFAULT
+ *
+ * \param generator - Pointer to generator
+ * \param rng_type - Type of generator to create
+ *
+ * \return
+ * - CURAND_STATUS_ALLOCATION_FAILED if memory could not be allocated \n
+ * - CURAND_STATUS_INITIALIZATION_FAILED if there was a problem setting up the GPU \n
+ * - CURAND_STATUS_VERSION_MISMATCH if the header file version does not match the
+ *   dynamically linked library version \n
+ * - CURAND_STATUS_TYPE_ERROR if the value for \p rng_type is invalid \n
+ * - CURAND_STATUS_SUCCESS if generator was created successfully \n
+ */
+curandStatus_t CURANDAPI
+curandCreateGeneratorHost(curandGenerator_t *generator, curandRngType_t rng_type);
+
+/**
+ * \brief Destroy an existing generator.
+ *
+ * Destroy an existing generator and free all memory associated with its state.
+ *
+ * \param generator - Generator to destroy
+ *
+ * \return
+ * - CURAND_STATUS_NOT_INITIALIZED if the generator was never created \n
+ * - CURAND_STATUS_SUCCESS if generator was destroyed successfully \n
+ */
+curandStatus_t CURANDAPI
+curandDestroyGenerator(curandGenerator_t generator);
+
+/**
+ * \brief Return the version number of the library.
+ *
+ * Return in \p *version the version number of the dynamically linked CURAND
+ * library.  The format is the same as CUDART_VERSION from the CUDA Runtime.
+ * The only supported configuration is CURAND version equal to CUDA Runtime
+ * version.
+ *
+ * \param version - CURAND library version
+ *
+ * \return
+ * - CURAND_STATUS_SUCCESS if the version number was successfully returned \n
+ */
+curandStatus_t CURANDAPI
+curandGetVersion(int *version);
+
+/**
+* \brief Return the value of the curand property.
+*
+* Return in \p *value the number for the property described by \p type of the
+* dynamically linked CURAND library.
+*
+* \param type - CUDA library property
+* \param value - integer value for the requested property
+*
+* \return
+* - CURAND_STATUS_SUCCESS if the property value was successfully returned \n
+* - CURAND_STATUS_OUT_OF_RANGE if the property type is not recognized \n
+*/
+curandStatus_t CURANDAPI
+curandGetProperty(libraryPropertyType type, int *value);
+
+
+/**
+ * \brief Set the current stream for CURAND kernel launches.
+ *
+ * Set the current stream for CURAND kernel launches.  All library functions
+ * will use this stream until set again.
+ *
+ * \param generator - Generator to modify
+ * \param stream - Stream to use or ::NULL for null stream
+ *
+ * \return
+ * - CURAND_STATUS_NOT_INITIALIZED if the generator was never created \n
+ * - CURAND_STATUS_SUCCESS if stream was set successfully \n
+ */
+curandStatus_t CURANDAPI
+curandSetStream(curandGenerator_t generator, cudaStream_t stream);
+
+/**
+ * \brief Set the seed value of the pseudo-random number generator.
+ *
+ * Set the seed value of the pseudorandom number generator.
+ * All values of seed are valid.  Different seeds will produce different sequences.
+ * Different seeds will often not be statistically correlated with each other,
+ * but some pairs of seed values may generate sequences which are statistically correlated.
+ *
+ * \param generator - Generator to modify
+ * \param seed - Seed value
+ *
+ * \return
+ * - CURAND_STATUS_NOT_INITIALIZED if the generator was never created \n
+ * - CURAND_STATUS_TYPE_ERROR if the generator is not a pseudorandom number generator \n
+ * - CURAND_STATUS_SUCCESS if generator seed was set successfully \n
+ */
+curandStatus_t CURANDAPI
+curandSetPseudoRandomGeneratorSeed(curandGenerator_t generator, unsigned long long seed);
+
+/**
+ * \brief Set the absolute offset of the pseudo or quasirandom number generator.
+ *
+ * Set the absolute offset of the pseudo or quasirandom number generator.
+ *
+ * All values of offset are valid.  The offset position is absolute, not
+ * relative to the current position in the sequence.
+ *
+ * \param generator - Generator to modify
+ * \param offset - Absolute offset position
+ *
+ * \return
+ * - CURAND_STATUS_NOT_INITIALIZED if the generator was never created \n
+ * - CURAND_STATUS_SUCCESS if generator offset was set successfully \n
+ */
+curandStatus_t CURANDAPI
+curandSetGeneratorOffset(curandGenerator_t generator, unsigned long long offset);
+
+/**
+ * \brief Set the ordering of results of the pseudo or quasirandom number generator.
+ *
+ * Set the ordering of results of the pseudo or quasirandom number generator.
+ *
+ * Legal values of \p order for pseudorandom generators are:
+ * - CURAND_ORDERING_PSEUDO_DEFAULT
+ * - CURAND_ORDERING_PSEUDO_BEST
+ * - CURAND_ORDERING_PSEUDO_SEEDED
+ * - CURAND_ORDERING_PSEUDO_LEGACY
+ *
+ * Legal values of \p order for quasirandom generators are:
+ * - CURAND_ORDERING_QUASI_DEFAULT
+ *
+ * \param generator - Generator to modify
+ * \param order - Ordering of results
+ *
+ * \return
+ * - CURAND_STATUS_NOT_INITIALIZED if the generator was never created \n
+ * - CURAND_STATUS_OUT_OF_RANGE if the ordering is not valid \n
+ * - CURAND_STATUS_SUCCESS if generator ordering was set successfully \n
+ */
+curandStatus_t CURANDAPI
+curandSetGeneratorOrdering(curandGenerator_t generator, curandOrdering_t order);
+
+/**
+ * \brief Set the number of dimensions.
+ *
+ * Set the number of dimensions to be generated by the quasirandom number
+ * generator.
+ *
+ * Legal values for \p num_dimensions are 1 to 20000.
+ *
+ * \param generator - Generator to modify
+ * \param num_dimensions - Number of dimensions
+ *
+ * \return
+ * - CURAND_STATUS_NOT_INITIALIZED if the generator was never created \n
+ * - CURAND_STATUS_OUT_OF_RANGE if num_dimensions is not valid \n
+ * - CURAND_STATUS_TYPE_ERROR if the generator is not a quasirandom number generator \n
+ * - CURAND_STATUS_SUCCESS if generator ordering was set successfully \n
+ */
+curandStatus_t CURANDAPI
+curandSetQuasiRandomGeneratorDimensions(curandGenerator_t generator, unsigned int num_dimensions);
+
+/**
+ * \brief Generate 32-bit pseudo or quasirandom numbers.
+ *
+ * Use \p generator to generate \p num 32-bit results into the device memory at
+ * \p outputPtr.  The device memory must have been previously allocated and be
+ * large enough to hold all the results.  Launches are done with the stream
+ * set using ::curandSetStream(), or the null stream if no stream has been set.
+ *
+ * Results are 32-bit values with every bit random.
+ *
+ * \param generator - Generator to use
+ * \param outputPtr - Pointer to device memory to store CUDA-generated results, or
+ *                 Pointer to host memory to store CPU-generated results
+ * \param num - Number of random 32-bit values to generate
+ *
+ * \return
+ * - CURAND_STATUS_ALLOCATION_FAILED if memory could not be allocated \n
+ * - CURAND_STATUS_NOT_INITIALIZED if the generator was never created \n
+ * - CURAND_STATUS_PREEXISTING_FAILURE if there was an existing error from
+ *     a previous kernel launch \n
+ * - CURAND_STATUS_LENGTH_NOT_MULTIPLE if the number of output samples is
+ *    not a multiple of the quasirandom dimension \n
+ * - CURAND_STATUS_LAUNCH_FAILURE if the kernel launch failed for any reason \n
+ * - CURAND_STATUS_TYPE_ERROR if the generator is a 64 bit quasirandom generator.
+ * (use ::curandGenerateLongLong() with 64 bit quasirandom generators)
+ * - CURAND_STATUS_SUCCESS if the results were generated successfully \n
+ */
+curandStatus_t CURANDAPI
+curandGenerate(curandGenerator_t generator, unsigned int *outputPtr, size_t num);
+
+/**
+ * \brief Generate 64-bit quasirandom numbers.
+ *
+ * Use \p generator to generate \p num 64-bit results into the device memory at
+ * \p outputPtr.  The device memory must have been previously allocated and be
+ * large enough to hold all the results.  Launches are done with the stream
+ * set using ::curandSetStream(), or the null stream if no stream has been set.
+ *
+ * Results are 64-bit values with every bit random.
+ *
+ * \param generator - Generator to use
+ * \param outputPtr - Pointer to device memory to store CUDA-generated results, or
+ *                 Pointer to host memory to store CPU-generated results
+ * \param num - Number of random 64-bit values to generate
+ *
+ * \return
+ * - CURAND_STATUS_NOT_INITIALIZED if the generator was never created \n
+ * - CURAND_STATUS_PREEXISTING_FAILURE if there was an existing error from
+ *     a previous kernel launch \n
+ * - CURAND_STATUS_LENGTH_NOT_MULTIPLE if the number of output samples is
+ *    not a multiple of the quasirandom dimension \n
+ * - CURAND_STATUS_LAUNCH_FAILURE if the kernel launch failed for any reason \n
+ * - CURAND_STATUS_TYPE_ERROR if the generator is not a 64 bit quasirandom generator\n
+ * - CURAND_STATUS_SUCCESS if the results were generated successfully \n
+ */
+curandStatus_t CURANDAPI
+curandGenerateLongLong(curandGenerator_t generator, unsigned long long *outputPtr, size_t num);
+
+/**
+ * \brief Generate uniformly distributed floats.
+ *
+ * Use \p generator to generate \p num float results into the device memory at
+ * \p outputPtr.  The device memory must have been previously allocated and be
+ * large enough to hold all the results.  Launches are done with the stream
+ * set using ::curandSetStream(), or the null stream if no stream has been set.
+ *
+ * Results are 32-bit floating point values between \p 0.0f and \p 1.0f,
+ * excluding \p 0.0f and including \p 1.0f.
+ *
+ * \param generator - Generator to use
+ * \param outputPtr - Pointer to device memory to store CUDA-generated results, or
+ *                 Pointer to host memory to store CPU-generated results
+ * \param num - Number of floats to generate
+ *
+ * \return
+ * - CURAND_STATUS_ALLOCATION_FAILED if memory could not be allocated \n
+ * - CURAND_STATUS_NOT_INITIALIZED if the generator was never created \n
+ * - CURAND_STATUS_PREEXISTING_FAILURE if there was an existing error from
+ *    a previous kernel launch \n
+ * - CURAND_STATUS_LAUNCH_FAILURE if the kernel launch failed for any reason \n
+ * - CURAND_STATUS_LENGTH_NOT_MULTIPLE if the number of output samples is
+ *    not a multiple of the quasirandom dimension \n
+ * - CURAND_STATUS_SUCCESS if the results were generated successfully \n
+ */
+curandStatus_t CURANDAPI
+curandGenerateUniform(curandGenerator_t generator, float *outputPtr, size_t num);
+
+/**
+ * \brief Generate uniformly distributed doubles.
+ *
+ * Use \p generator to generate \p num double results into the device memory at
+ * \p outputPtr.  The device memory must have been previously allocated and be
+ * large enough to hold all the results.  Launches are done with the stream
+ * set using ::curandSetStream(), or the null stream if no stream has been set.
+ *
+ * Results are 64-bit double precision floating point values between
+ * \p 0.0 and \p 1.0, excluding \p 0.0 and including \p 1.0.
+ *
+ * \param generator - Generator to use
+ * \param outputPtr - Pointer to device memory to store CUDA-generated results, or
+ *                 Pointer to host memory to store CPU-generated results
+ * \param num - Number of doubles to generate
+ *
+ * \return
+ * - CURAND_STATUS_ALLOCATION_FAILED if memory could not be allocated \n
+ * - CURAND_STATUS_NOT_INITIALIZED if the generator was never created \n
+ * - CURAND_STATUS_PREEXISTING_FAILURE if there was an existing error from
+ *    a previous kernel launch \n
+ * - CURAND_STATUS_LAUNCH_FAILURE if the kernel launch failed for any reason \n
+ * - CURAND_STATUS_LENGTH_NOT_MULTIPLE if the number of output samples is
+ *    not a multiple of the quasirandom dimension \n
+ * - CURAND_STATUS_DOUBLE_PRECISION_REQUIRED if the GPU does not support double precision \n
+ * - CURAND_STATUS_SUCCESS if the results were generated successfully \n
+ */
+curandStatus_t CURANDAPI
+curandGenerateUniformDouble(curandGenerator_t generator, double *outputPtr, size_t num);
+
+/**
+ * \brief Generate normally distributed doubles.
+ *
+ * Use \p generator to generate \p n float results into the device memory at
+ * \p outputPtr.  The device memory must have been previously allocated and be
+ * large enough to hold all the results.  Launches are done with the stream
+ * set using ::curandSetStream(), or the null stream if no stream has been set.
+ *
+ * Results are 32-bit floating point values with mean \p mean and standard
+ * deviation \p stddev.
+ *
+ * Normally distributed results are generated from pseudorandom generators
+ * with a Box-Muller transform, and so require \p n to be even.
+ * Quasirandom generators use an inverse cumulative distribution
+ * function to preserve dimensionality.
+ *
+ * There may be slight numerical differences between results generated
+ * on the GPU with generators created with ::curandCreateGenerator()
+ * and results calculated on the CPU with generators created with
+ * ::curandCreateGeneratorHost().  These differences arise because of
+ * differences in results for transcendental functions.  In addition,
+ * future versions of CURAND may use newer versions of the CUDA math
+ * library, so different versions of CURAND may give slightly different
+ * numerical values.
+ *
+ * \param generator - Generator to use
+ * \param outputPtr - Pointer to device memory to store CUDA-generated results, or
+ *                 Pointer to host memory to store CPU-generated results
+ * \param n - Number of floats to generate
+ * \param mean - Mean of normal distribution
+ * \param stddev - Standard deviation of normal distribution
+ *
+ * \return
+ * - CURAND_STATUS_ALLOCATION_FAILED if memory could not be allocated \n
+ * - CURAND_STATUS_NOT_INITIALIZED if the generator was never created \n
+ * - CURAND_STATUS_PREEXISTING_FAILURE if there was an existing error from
+ *    a previous kernel launch \n
+ * - CURAND_STATUS_LAUNCH_FAILURE if the kernel launch failed for any reason \n
+ * - CURAND_STATUS_LENGTH_NOT_MULTIPLE if the number of output samples is
+ *    not a multiple of the quasirandom dimension, or is not a multiple
+ *    of two for pseudorandom generators \n
+ * - CURAND_STATUS_SUCCESS if the results were generated successfully \n
+ */
+curandStatus_t CURANDAPI
+curandGenerateNormal(curandGenerator_t generator, float *outputPtr,
+                     size_t n, float mean, float stddev);
+
+/**
+ * \brief Generate normally distributed doubles.
+ *
+ * Use \p generator to generate \p n double results into the device memory at
+ * \p outputPtr.  The device memory must have been previously allocated and be
+ * large enough to hold all the results.  Launches are done with the stream
+ * set using ::curandSetStream(), or the null stream if no stream has been set.
+ *
+ * Results are 64-bit floating point values with mean \p mean and standard
+ * deviation \p stddev.
+ *
+ * Normally distributed results are generated from pseudorandom generators
+ * with a Box-Muller transform, and so require \p n to be even.
+ * Quasirandom generators use an inverse cumulative distribution
+ * function to preserve dimensionality.
+ *
+ * There may be slight numerical differences between results generated
+ * on the GPU with generators created with ::curandCreateGenerator()
+ * and results calculated on the CPU with generators created with
+ * ::curandCreateGeneratorHost().  These differences arise because of
+ * differences in results for transcendental functions.  In addition,
+ * future versions of CURAND may use newer versions of the CUDA math
+ * library, so different versions of CURAND may give slightly different
+ * numerical values.
+ *
+ * \param generator - Generator to use
+ * \param outputPtr - Pointer to device memory to store CUDA-generated results, or
+ *                 Pointer to host memory to store CPU-generated results
+ * \param n - Number of doubles to generate
+ * \param mean - Mean of normal distribution
+ * \param stddev - Standard deviation of normal distribution
+ *
+ * \return
+ * - CURAND_STATUS_ALLOCATION_FAILED if memory could not be allocated \n
+ * - CURAND_STATUS_NOT_INITIALIZED if the generator was never created \n
+ * - CURAND_STATUS_PREEXISTING_FAILURE if there was an existing error from
+ *    a previous kernel launch \n
+ * - CURAND_STATUS_LAUNCH_FAILURE if the kernel launch failed for any reason \n
+ * - CURAND_STATUS_LENGTH_NOT_MULTIPLE if the number of output samples is
+ *    not a multiple of the quasirandom dimension, or is not a multiple
+ *    of two for pseudorandom generators \n
+ * - CURAND_STATUS_DOUBLE_PRECISION_REQUIRED if the GPU does not support double precision \n
+ * - CURAND_STATUS_SUCCESS if the results were generated successfully \n
+ */
+curandStatus_t CURANDAPI
+curandGenerateNormalDouble(curandGenerator_t generator, double *outputPtr,
+                     size_t n, double mean, double stddev);
+
+/**
+ * \brief Generate log-normally distributed floats.
+ *
+ * Use \p generator to generate \p n float results into the device memory at
+ * \p outputPtr.  The device memory must have been previously allocated and be
+ * large enough to hold all the results.  Launches are done with the stream
+ * set using ::curandSetStream(), or the null stream if no stream has been set.
+ *
+ * Results are 32-bit floating point values with log-normal distribution based on
+ * an associated normal distribution with mean \p mean and standard deviation \p stddev.
+ *
+ * Normally distributed results are generated from pseudorandom generators
+ * with a Box-Muller transform, and so require \p n to be even.
+ * Quasirandom generators use an inverse cumulative distribution
+ * function to preserve dimensionality.
+ * The normally distributed results are transformed into log-normal distribution.
+ *
+ * There may be slight numerical differences between results generated
+ * on the GPU with generators created with ::curandCreateGenerator()
+ * and results calculated on the CPU with generators created with
+ * ::curandCreateGeneratorHost().  These differences arise because of
+ * differences in results for transcendental functions.  In addition,
+ * future versions of CURAND may use newer versions of the CUDA math
+ * library, so different versions of CURAND may give slightly different
+ * numerical values.
+ *
+ * \param generator - Generator to use
+ * \param outputPtr - Pointer to device memory to store CUDA-generated results, or
+ *                 Pointer to host memory to store CPU-generated results
+ * \param n - Number of floats to generate
+ * \param mean - Mean of associated normal distribution
+ * \param stddev - Standard deviation of associated normal distribution
+ *
+ * \return
+ * - CURAND_STATUS_ALLOCATION_FAILED if memory could not be allocated \n
+ * - CURAND_STATUS_NOT_INITIALIZED if the generator was never created \n
+ * - CURAND_STATUS_PREEXISTING_FAILURE if there was an existing error from
+ *    a previous kernel launch \n
+ * - CURAND_STATUS_LAUNCH_FAILURE if the kernel launch failed for any reason \n
+ * - CURAND_STATUS_LENGTH_NOT_MULTIPLE if the number of output samples is
+ *    not a multiple of the quasirandom dimension, or is not a multiple
+ *    of two for pseudorandom generators \n
+ * - CURAND_STATUS_SUCCESS if the results were generated successfully \n
+ */
+curandStatus_t CURANDAPI
+curandGenerateLogNormal(curandGenerator_t generator, float *outputPtr,
+                     size_t n, float mean, float stddev);
+
+/**
+ * \brief Generate log-normally distributed doubles.
+ *
+ * Use \p generator to generate \p n double results into the device memory at
+ * \p outputPtr.  The device memory must have been previously allocated and be
+ * large enough to hold all the results.  Launches are done with the stream
+ * set using ::curandSetStream(), or the null stream if no stream has been set.
+ *
+ * Results are 64-bit floating point values with log-normal distribution based on
+ * an associated normal distribution with mean \p mean and standard deviation \p stddev.
+ *
+ * Normally distributed results are generated from pseudorandom generators
+ * with a Box-Muller transform, and so require \p n to be even.
+ * Quasirandom generators use an inverse cumulative distribution
+ * function to preserve dimensionality.
+ * The normally distributed results are transformed into log-normal distribution.
+ *
+ * There may be slight numerical differences between results generated
+ * on the GPU with generators created with ::curandCreateGenerator()
+ * and results calculated on the CPU with generators created with
+ * ::curandCreateGeneratorHost().  These differences arise because of
+ * differences in results for transcendental functions.  In addition,
+ * future versions of CURAND may use newer versions of the CUDA math
+ * library, so different versions of CURAND may give slightly different
+ * numerical values.
+ *
+ * \param generator - Generator to use
+ * \param outputPtr - Pointer to device memory to store CUDA-generated results, or
+ *                 Pointer to host memory to store CPU-generated results
+ * \param n - Number of doubles to generate
+ * \param mean - Mean of normal distribution
+ * \param stddev - Standard deviation of normal distribution
+ *
+ * \return
+ * - CURAND_STATUS_ALLOCATION_FAILED if memory could not be allocated \n
+ * - CURAND_STATUS_NOT_INITIALIZED if the generator was never created \n
+ * - CURAND_STATUS_PREEXISTING_FAILURE if there was an existing error from
+ *    a previous kernel launch \n
+ * - CURAND_STATUS_LAUNCH_FAILURE if the kernel launch failed for any reason \n
+ * - CURAND_STATUS_LENGTH_NOT_MULTIPLE if the number of output samples is
+ *    not a multiple of the quasirandom dimension, or is not a multiple
+ *    of two for pseudorandom generators \n
+ * - CURAND_STATUS_DOUBLE_PRECISION_REQUIRED if the GPU does not support double precision \n
+ * - CURAND_STATUS_SUCCESS if the results were generated successfully \n
+ */
+curandStatus_t CURANDAPI
+curandGenerateLogNormalDouble(curandGenerator_t generator, double *outputPtr,
+                     size_t n, double mean, double stddev);
+
+/**
+ * \brief Construct the histogram array for a Poisson distribution.
+ *
+ * Construct the histogram array for the Poisson distribution with lambda \p lambda.
+ * For lambda greater than 2000, an approximation with a normal distribution is used.
+ *
+ * \param lambda - lambda for the Poisson distribution
+ *
+ *
+ * \param discrete_distribution - pointer to the histogram in device memory
+ *
+ * \return
+ * - CURAND_STATUS_ALLOCATION_FAILED if memory could not be allocated \n
+ * - CURAND_STATUS_DOUBLE_PRECISION_REQUIRED if the GPU does not support double precision \n
+ * - CURAND_STATUS_INITIALIZATION_FAILED if there was a problem setting up the GPU \n
+ * - CURAND_STATUS_NOT_INITIALIZED if the distribution pointer was null \n
+ * - CURAND_STATUS_PREEXISTING_FAILURE if there was an existing error from
+ *    a previous kernel launch \n
+ * - CURAND_STATUS_OUT_OF_RANGE if lambda is non-positive or greater than 400,000 \n
+ * - CURAND_STATUS_SUCCESS if the histogram was generated successfully \n
+ */
+
+curandStatus_t CURANDAPI
+curandCreatePoissonDistribution(double lambda, curandDiscreteDistribution_t *discrete_distribution);
+
+
+
+/**
+ * \brief Destroy the histogram array for a discrete distribution (e.g. Poisson).
+ *
+ * Destroy the histogram array for a discrete distribution created by curandCreatePoissonDistribution.
+ *
+ * \param discrete_distribution - pointer to device memory where the histogram is stored
+ *
+ * \return
+ * - CURAND_STATUS_NOT_INITIALIZED if the histogram was never created \n
+ * - CURAND_STATUS_SUCCESS if the histogram was destroyed successfully \n
+ */
+curandStatus_t CURANDAPI
+curandDestroyDistribution(curandDiscreteDistribution_t discrete_distribution);
+
+
+/**
+ * \brief Generate Poisson-distributed unsigned ints.
+ *
+ * Use \p generator to generate \p n unsigned int results into device memory at
+ * \p outputPtr.  The device memory must have been previously allocated and must be
+ * large enough to hold all the results.  Launches are done with the stream
+ * set using ::curandSetStream(), or the null stream if no stream has been set.
+ *
+ * Results are 32-bit unsigned int point values with Poisson distribution, with lambda \p lambda.
+ *
+ * \param generator - Generator to use
+ * \param outputPtr - Pointer to device memory to store CUDA-generated results, or
+ *                 Pointer to host memory to store CPU-generated results
+ * \param n - Number of unsigned ints to generate
+ * \param lambda - lambda for the Poisson distribution
+ *
+ * \return
+ * - CURAND_STATUS_ALLOCATION_FAILED if memory could not be allocated \n
+ * - CURAND_STATUS_NOT_INITIALIZED if the generator was never created \n
+ * - CURAND_STATUS_PREEXISTING_FAILURE if there was an existing error from
+ *    a previous kernel launch \n
+ * - CURAND_STATUS_LAUNCH_FAILURE if the kernel launch failed for any reason \n
+ * - CURAND_STATUS_LENGTH_NOT_MULTIPLE if the number of output samples is
+ *    not a multiple of the quasirandom dimension\n
+ * - CURAND_STATUS_DOUBLE_PRECISION_REQUIRED if the GPU or sm does not support double precision \n
+ * - CURAND_STATUS_OUT_OF_RANGE if lambda is non-positive or greater than 400,000 \n
+ * - CURAND_STATUS_SUCCESS if the results were generated successfully \n
+ */
+
+curandStatus_t CURANDAPI
+curandGeneratePoisson(curandGenerator_t generator, unsigned int *outputPtr,
+                     size_t n, double lambda);
+// just for internal usage
+curandStatus_t CURANDAPI
+curandGeneratePoissonMethod(curandGenerator_t generator, unsigned int *outputPtr,
+                     size_t n, double lambda, curandMethod_t method);
+
+
+curandStatus_t CURANDAPI
+curandGenerateBinomial(curandGenerator_t generator, unsigned int *outputPtr,
+                       size_t num, unsigned int n, double p);
+// just for internal usage
+curandStatus_t CURANDAPI
+curandGenerateBinomialMethod(curandGenerator_t generator,
+                             unsigned int *outputPtr,
+                             size_t num, unsigned int n, double p,
+                             curandMethod_t method);
+
+
+/**
+ * \brief Setup starting states.
+ *
+ * Generate the starting state of the generator.  This function is
+ * automatically called by generation functions such as
+ * ::curandGenerate() and ::curandGenerateUniform().
+ * It can be called manually for performance testing reasons to separate
+ * timings for starting state generation and random number generation.
+ *
+ * \param generator - Generator to update
+ *
+ * \return
+ * - CURAND_STATUS_ALLOCATION_FAILED if memory could not be allocated \n
+ * - CURAND_STATUS_NOT_INITIALIZED if the generator was never created \n
+ * - CURAND_STATUS_PREEXISTING_FAILURE if there was an existing error from
+ *     a previous kernel launch \n
+ * - CURAND_STATUS_LAUNCH_FAILURE if the kernel launch failed for any reason \n
+ * - CURAND_STATUS_SUCCESS if the seeds were generated successfully \n
+ */
+curandStatus_t CURANDAPI
+curandGenerateSeeds(curandGenerator_t generator);
+
+/**
+ * \brief Get direction vectors for 32-bit quasirandom number generation.
+ *
+ * Get a pointer to an array of direction vectors that can be used
+ * for quasirandom number generation.  The resulting pointer will
+ * reference an array of direction vectors in host memory.
+ *
+ * The array contains vectors for many dimensions.  Each dimension
+ * has 32 vectors.  Each individual vector is an unsigned int.
+ *
+ * Legal values for \p set are:
+ * - CURAND_DIRECTION_VECTORS_32_JOEKUO6 (20,000 dimensions)
+ * - CURAND_SCRAMBLED_DIRECTION_VECTORS_32_JOEKUO6 (20,000 dimensions)
+ *
+ * \param vectors - Address of pointer in which to return direction vectors
+ * \param set - Which set of direction vectors to use
+ *
+ * \return
+ * - CURAND_STATUS_OUT_OF_RANGE if the choice of set is invalid \n
+ * - CURAND_STATUS_SUCCESS if the pointer was set successfully \n
+ */
+curandStatus_t CURANDAPI
+curandGetDirectionVectors32(curandDirectionVectors32_t *vectors[], curandDirectionVectorSet_t set);
+
+/**
+ * \brief Get scramble constants for 32-bit scrambled Sobol' .
+ *
+ * Get a pointer to an array of scramble constants that can be used
+ * for quasirandom number generation.  The resulting pointer will
+ * reference an array of unsinged ints in host memory.
+ *
+ * The array contains constants for many dimensions.  Each dimension
+ * has a single unsigned int constant.
+ *
+ * \param constants - Address of pointer in which to return scramble constants
+ *
+ * \return
+ * - CURAND_STATUS_SUCCESS if the pointer was set successfully \n
+ */
+curandStatus_t CURANDAPI
+curandGetScrambleConstants32(unsigned int * * constants);
+
+/**
+ * \brief Get direction vectors for 64-bit quasirandom number generation.
+ *
+ * Get a pointer to an array of direction vectors that can be used
+ * for quasirandom number generation.  The resulting pointer will
+ * reference an array of direction vectors in host memory.
+ *
+ * The array contains vectors for many dimensions.  Each dimension
+ * has 64 vectors.  Each individual vector is an unsigned long long.
+ *
+ * Legal values for \p set are:
+ * - CURAND_DIRECTION_VECTORS_64_JOEKUO6 (20,000 dimensions)
+ * - CURAND_SCRAMBLED_DIRECTION_VECTORS_64_JOEKUO6 (20,000 dimensions)
+ *
+ * \param vectors - Address of pointer in which to return direction vectors
+ * \param set - Which set of direction vectors to use
+ *
+ * \return
+ * - CURAND_STATUS_OUT_OF_RANGE if the choice of set is invalid \n
+ * - CURAND_STATUS_SUCCESS if the pointer was set successfully \n
+ */
+curandStatus_t CURANDAPI
+curandGetDirectionVectors64(curandDirectionVectors64_t *vectors[], curandDirectionVectorSet_t set);
+
+/**
+ * \brief Get scramble constants for 64-bit scrambled Sobol' .
+ *
+ * Get a pointer to an array of scramble constants that can be used
+ * for quasirandom number generation.  The resulting pointer will
+ * reference an array of unsinged long longs in host memory.
+ *
+ * The array contains constants for many dimensions.  Each dimension
+ * has a single unsigned long long constant.
+ *
+ * \param constants - Address of pointer in which to return scramble constants
+ *
+ * \return
+ * - CURAND_STATUS_SUCCESS if the pointer was set successfully \n
+ */
+curandStatus_t CURANDAPI
+curandGetScrambleConstants64(unsigned long long * * constants);
+
+/** @} */
+
+#endif // __CUDACC_RTC__
+
+#if defined(__cplusplus)
+}
+#endif /* __cplusplus */
+
+
+#endif /* !defined(CURAND_H_) */

tuning-competition-baseline/.venv/lib/python3.11/site-packages/nvidia_cuda_nvrtc_cu11-11.8.89.dist-info/License.txt

@@ -0,0 +1,1568 @@
+End User License Agreement
+--------------------------
+
+
+Preface
+-------
+
+The Software License Agreement in Chapter 1 and the Supplement
+in Chapter 2 contain license terms and conditions that govern
+the use of NVIDIA software. By accepting this agreement, you
+agree to comply with all the terms and conditions applicable
+to the product(s) included herein.
+
+
+NVIDIA Driver
+
+
+Description
+
+This package contains the operating system driver and
+fundamental system software components for NVIDIA GPUs.
+
+
+NVIDIA CUDA Toolkit
+
+
+Description
+
+The NVIDIA CUDA Toolkit provides command-line and graphical
+tools for building, debugging and optimizing the performance
+of applications accelerated by NVIDIA GPUs, runtime and math
+libraries, and documentation including programming guides,
+user manuals, and API references.
+
+
+Default Install Location of CUDA Toolkit
+
+Windows platform:
+
+%ProgramFiles%\NVIDIA GPU Computing Toolkit\CUDA\v#.#
+
+Linux platform:
+
+/usr/local/cuda-#.#
+
+Mac platform:
+
+/Developer/NVIDIA/CUDA-#.#
+
+
+NVIDIA CUDA Samples
+
+
+Description
+
+This package includes over 100+ CUDA examples that demonstrate
+various CUDA programming principles, and efficient CUDA
+implementation of algorithms in specific application domains.
+
+
+Default Install Location of CUDA Samples
+
+Windows platform:
+
+%ProgramData%\NVIDIA Corporation\CUDA Samples\v#.#
+
+Linux platform:
+
+/usr/local/cuda-#.#/samples
+
+and
+
+$HOME/NVIDIA_CUDA-#.#_Samples
+
+Mac platform:
+
+/Developer/NVIDIA/CUDA-#.#/samples
+
+
+NVIDIA Nsight Visual Studio Edition (Windows only)
+
+
+Description
+
+NVIDIA Nsight Development Platform, Visual Studio Edition is a
+development environment integrated into Microsoft Visual
+Studio that provides tools for debugging, profiling, analyzing
+and optimizing your GPU computing and graphics applications.
+
+
+Default Install Location of Nsight Visual Studio Edition
+
+Windows platform:
+
+%ProgramFiles(x86)%\NVIDIA Corporation\Nsight Visual Studio Edition #.#
+
+
+1. License Agreement for NVIDIA Software Development Kits
+---------------------------------------------------------
+
+
+Release Date: July 26, 2018
+---------------------------
+
+
+Important NoticeRead before downloading, installing,
+copying or using the licensed software:
+-------------------------------------------------------
+
+This license agreement, including exhibits attached
+("Agreement”) is a legal agreement between you and NVIDIA
+Corporation ("NVIDIA") and governs your use of a NVIDIA
+software development kit (“SDK”).
+
+Each SDK has its own set of software and materials, but here
+is a description of the types of items that may be included in
+a SDK: source code, header files, APIs, data sets and assets
+(examples include images, textures, models, scenes, videos,
+native API input/output files), binary software, sample code,
+libraries, utility programs, programming code and
+documentation.
+
+This Agreement can be accepted only by an adult of legal age
+of majority in the country in which the SDK is used.
+
+If you are entering into this Agreement on behalf of a company
+or other legal entity, you represent that you have the legal
+authority to bind the entity to this Agreement, in which case
+“you” will mean the entity you represent.
+
+If you don’t have the required age or authority to accept
+this Agreement, or if you don’t accept all the terms and
+conditions of this Agreement, do not download, install or use
+the SDK.
+
+You agree to use the SDK only for purposes that are permitted
+by (a) this Agreement, and (b) any applicable law, regulation
+or generally accepted practices or guidelines in the relevant
+jurisdictions.
+
+
+1.1. License
+
+
+1.1.1. License Grant
+
+Subject to the terms of this Agreement, NVIDIA hereby grants
+you a non-exclusive, non-transferable license, without the
+right to sublicense (except as expressly provided in this
+Agreement) to:
+
+  1. Install and use the SDK,
+
+  2. Modify and create derivative works of sample source code
+    delivered in the SDK, and
+
+  3. Distribute those portions of the SDK that are identified
+    in this Agreement as distributable, as incorporated in
+    object code format into a software application that meets
+    the distribution requirements indicated in this Agreement.
+
+
+1.1.2. Distribution Requirements
+
+These are the distribution requirements for you to exercise
+the distribution grant:
+
+  1. Your application must have material additional
+    functionality, beyond the included portions of the SDK.
+
+  2. The distributable portions of the SDK shall only be
+    accessed by your application.
+
+  3. The following notice shall be included in modifications
+    and derivative works of sample source code distributed:
+    “This software contains source code provided by NVIDIA
+    Corporation.”
+
+  4. Unless a developer tool is identified in this Agreement
+    as distributable, it is delivered for your internal use
+    only.
+
+  5. The terms under which you distribute your application
+    must be consistent with the terms of this Agreement,
+    including (without limitation) terms relating to the
+    license grant and license restrictions and protection of
+    NVIDIA’s intellectual property rights. Additionally, you
+    agree that you will protect the privacy, security and
+    legal rights of your application users.
+
+  6. You agree to notify NVIDIA in writing of any known or
+    suspected distribution or use of the SDK not in compliance
+    with the requirements of this Agreement, and to enforce
+    the terms of your agreements with respect to distributed
+    SDK.
+
+
+1.1.3. Authorized Users
+
+You may allow employees and contractors of your entity or of
+your subsidiary(ies) to access and use the SDK from your
+secure network to perform work on your behalf.
+
+If you are an academic institution you may allow users
+enrolled or employed by the academic institution to access and
+use the SDK from your secure network.
+
+You are responsible for the compliance with the terms of this
+Agreement by your authorized users. If you become aware that
+your authorized users didn’t follow the terms of this
+Agreement, you agree to take reasonable steps to resolve the
+non-compliance and prevent new occurrences.
+
+
+1.1.4. Pre-Release SDK
+
+The SDK versions identified as alpha, beta, preview or
+otherwise as pre-release, may not be fully functional, may
+contain errors or design flaws, and may have reduced or
+different security, privacy, accessibility, availability, and
+reliability standards relative to commercial versions of
+NVIDIA software and materials. Use of a pre-release SDK may
+result in unexpected results, loss of data, project delays or
+other unpredictable damage or loss.
+
+You may use a pre-release SDK at your own risk, understanding
+that pre-release SDKs are not intended for use in production
+or business-critical systems.
+
+NVIDIA may choose not to make available a commercial version
+of any pre-release SDK. NVIDIA may also choose to abandon
+development and terminate the availability of a pre-release
+SDK at any time without liability.
+
+
+1.1.5. Updates
+
+NVIDIA may, at its option, make available patches, workarounds
+or other updates to this SDK. Unless the updates are provided
+with their separate governing terms, they are deemed part of
+the SDK licensed to you as provided in this Agreement. You
+agree that the form and content of the SDK that NVIDIA
+provides may change without prior notice to you. While NVIDIA
+generally maintains compatibility between versions, NVIDIA may
+in some cases make changes that introduce incompatibilities in
+future versions of the SDK.
+
+
+1.1.6. Third Party Licenses
+
+The SDK may come bundled with, or otherwise include or be
+distributed with, third party software licensed by a NVIDIA
+supplier and/or open source software provided under an open
+source license. Use of third party software is subject to the
+third-party license terms, or in the absence of third party
+terms, the terms of this Agreement. Copyright to third party
+software is held by the copyright holders indicated in the
+third-party software or license.
+
+
+1.1.7. Reservation of Rights
+
+NVIDIA reserves all rights, title, and interest in and to the
+SDK, not expressly granted to you under this Agreement.
+
+
+1.2. Limitations
+
+The following license limitations apply to your use of the
+SDK:
+
+  1. You may not reverse engineer, decompile or disassemble,
+    or remove copyright or other proprietary notices from any
+    portion of the SDK or copies of the SDK.
+
+  2. Except as expressly provided in this Agreement, you may
+    not copy, sell, rent, sublicense, transfer, distribute,
+    modify, or create derivative works of any portion of the
+    SDK. For clarity, you may not distribute or sublicense the
+    SDK as a stand-alone product.
+
+  3. Unless you have an agreement with NVIDIA for this
+    purpose, you may not indicate that an application created
+    with the SDK is sponsored or endorsed by NVIDIA.
+
+  4. You may not bypass, disable, or circumvent any
+    encryption, security, digital rights management or
+    authentication mechanism in the SDK.
+
+  5. You may not use the SDK in any manner that would cause it
+    to become subject to an open source software license. As
+    examples, licenses that require as a condition of use,
+    modification, and/or distribution that the SDK be:
+
+      a. Disclosed or distributed in source code form;
+
+      b. Licensed for the purpose of making derivative works;
+        or
+
+      c. Redistributable at no charge.
+
+  6. Unless you have an agreement with NVIDIA for this
+    purpose, you may not use the SDK with any system or
+    application where the use or failure of the system or
+    application can reasonably be expected to threaten or
+    result in personal injury, death, or catastrophic loss.
+    Examples include use in avionics, navigation, military,
+    medical, life support or other life critical applications.
+    NVIDIA does not design, test or manufacture the SDK for
+    these critical uses and NVIDIA shall not be liable to you
+    or any third party, in whole or in part, for any claims or
+    damages arising from such uses.
+
+  7. You agree to defend, indemnify and hold harmless NVIDIA
+    and its affiliates, and their respective employees,
+    contractors, agents, officers and directors, from and
+    against any and all claims, damages, obligations, losses,
+    liabilities, costs or debt, fines, restitutions and
+    expenses (including but not limited to attorney’s fees
+    and costs incident to establishing the right of
+    indemnification) arising out of or related to your use of
+    the SDK outside of the scope of this Agreement, or not in
+    compliance with its terms.
+
+
+1.3. Ownership
+
+  1.  NVIDIA or its licensors hold all rights, title and
+    interest in and to the SDK and its modifications and
+    derivative works, including their respective intellectual
+    property rights, subject to your rights described in this
+    section. This SDK may include software and materials from
+    NVIDIA’s licensors, and these licensors are intended
+    third party beneficiaries that may enforce this Agreement
+    with respect to their intellectual property rights.
+
+  2.  You hold all rights, title and interest in and to your
+    applications and your derivative works of the sample
+    source code delivered in the SDK, including their
+    respective intellectual property rights, subject to
+    NVIDIA’s rights described in this section.
+
+  3. You may, but don’t have to, provide to NVIDIA
+    suggestions, feature requests or other feedback regarding
+    the SDK, including possible enhancements or modifications
+    to the SDK. For any feedback that you voluntarily provide,
+    you hereby grant NVIDIA and its affiliates a perpetual,
+    non-exclusive, worldwide, irrevocable license to use,
+    reproduce, modify, license, sublicense (through multiple
+    tiers of sublicensees), and distribute (through multiple
+    tiers of distributors) it without the payment of any
+    royalties or fees to you. NVIDIA will use feedback at its
+    choice. NVIDIA is constantly looking for ways to improve
+    its products, so you may send feedback to NVIDIA through
+    the developer portal at https://developer.nvidia.com.
+
+
+1.4. No Warranties
+
+THE SDK IS PROVIDED BY NVIDIA “AS IS” AND “WITH ALL
+FAULTS.” TO THE MAXIMUM EXTENT PERMITTED BY LAW, NVIDIA AND
+ITS AFFILIATES EXPRESSLY DISCLAIM ALL WARRANTIES OF ANY KIND
+OR NATURE, WHETHER EXPRESS, IMPLIED OR STATUTORY, INCLUDING,
+BUT NOT LIMITED TO, ANY WARRANTIES OF MERCHANTABILITY, FITNESS
+FOR A PARTICULAR PURPOSE, TITLE, NON-INFRINGEMENT, OR THE
+ABSENCE OF ANY DEFECTS THEREIN, WHETHER LATENT OR PATENT. NO
+WARRANTY IS MADE ON THE BASIS OF TRADE USAGE, COURSE OF
+DEALING OR COURSE OF TRADE.
+
+
+1.5. Limitation of Liability
+
+TO THE MAXIMUM EXTENT PERMITTED BY LAW, NVIDIA AND ITS
+AFFILIATES SHALL NOT BE LIABLE FOR ANY SPECIAL, INCIDENTAL,
+PUNITIVE OR CONSEQUENTIAL DAMAGES, OR ANY LOST PROFITS, LOSS
+OF USE, LOSS OF DATA OR LOSS OF GOODWILL, OR THE COSTS OF
+PROCURING SUBSTITUTE PRODUCTS, ARISING OUT OF OR IN CONNECTION
+WITH THIS AGREEMENT OR THE USE OR PERFORMANCE OF THE SDK,
+WHETHER SUCH LIABILITY ARISES FROM ANY CLAIM BASED UPON BREACH
+OF CONTRACT, BREACH OF WARRANTY, TORT (INCLUDING NEGLIGENCE),
+PRODUCT LIABILITY OR ANY OTHER CAUSE OF ACTION OR THEORY OF
+LIABILITY. IN NO EVENT WILL NVIDIA’S AND ITS AFFILIATES
+TOTAL CUMULATIVE LIABILITY UNDER OR ARISING OUT OF THIS
+AGREEMENT EXCEED US$10.00. THE NATURE OF THE LIABILITY OR THE
+NUMBER OF CLAIMS OR SUITS SHALL NOT ENLARGE OR EXTEND THIS
+LIMIT.
+
+These exclusions and limitations of liability shall apply
+regardless if NVIDIA or its affiliates have been advised of
+the possibility of such damages, and regardless of whether a
+remedy fails its essential purpose. These exclusions and
+limitations of liability form an essential basis of the
+bargain between the parties, and, absent any of these
+exclusions or limitations of liability, the provisions of this
+Agreement, including, without limitation, the economic terms,
+would be substantially different.
+
+
+1.6. Termination
+
+  1. This Agreement will continue to apply until terminated by
+    either you or NVIDIA as described below.
+
+  2. If you want to terminate this Agreement, you may do so by
+    stopping to use the SDK.
+
+  3. NVIDIA may, at any time, terminate this Agreement if:
+
+      a. (i) you fail to comply with any term of this
+        Agreement and the non-compliance is not fixed within
+        thirty (30) days following notice from NVIDIA (or
+        immediately if you violate NVIDIA’s intellectual
+        property rights);
+
+      b. (ii) you commence or participate in any legal
+        proceeding against NVIDIA with respect to the SDK; or
+
+      c. (iii) NVIDIA decides to no longer provide the SDK in
+        a country or, in NVIDIA’s sole discretion, the
+        continued use of it is no longer commercially viable.
+
+  4. Upon any termination of this Agreement, you agree to
+    promptly discontinue use of the SDK and destroy all copies
+    in your possession or control. Your prior distributions in
+    accordance with this Agreement are not affected by the
+    termination of this Agreement. Upon written request, you
+    will certify in writing that you have complied with your
+    commitments under this section. Upon any termination of
+    this Agreement all provisions survive except for the
+    license grant provisions.
+
+
+1.7. General
+
+If you wish to assign this Agreement or your rights and
+obligations, including by merger, consolidation, dissolution
+or operation of law, contact NVIDIA to ask for permission. Any
+attempted assignment not approved by NVIDIA in writing shall
+be void and of no effect. NVIDIA may assign, delegate or
+transfer this Agreement and its rights and obligations, and if
+to a non-affiliate you will be notified.
+
+You agree to cooperate with NVIDIA and provide reasonably
+requested information to verify your compliance with this
+Agreement.
+
+This Agreement will be governed in all respects by the laws of
+the United States and of the State of Delaware as those laws
+are applied to contracts entered into and performed entirely
+within Delaware by Delaware residents, without regard to the
+conflicts of laws principles. The United Nations Convention on
+Contracts for the International Sale of Goods is specifically
+disclaimed. You agree to all terms of this Agreement in the
+English language.
+
+The state or federal courts residing in Santa Clara County,
+California shall have exclusive jurisdiction over any dispute
+or claim arising out of this Agreement. Notwithstanding this,
+you agree that NVIDIA shall still be allowed to apply for
+injunctive remedies or an equivalent type of urgent legal
+relief in any jurisdiction.
+
+If any court of competent jurisdiction determines that any
+provision of this Agreement is illegal, invalid or
+unenforceable, such provision will be construed as limited to
+the extent necessary to be consistent with and fully
+enforceable under the law and the remaining provisions will
+remain in full force and effect. Unless otherwise specified,
+remedies are cumulative.
+
+Each party acknowledges and agrees that the other is an
+independent contractor in the performance of this Agreement.
+
+The SDK has been developed entirely at private expense and is
+“commercial items” consisting of “commercial computer
+software” and “commercial computer software
+documentation” provided with RESTRICTED RIGHTS. Use,
+duplication or disclosure by the U.S. Government or a U.S.
+Government subcontractor is subject to the restrictions in
+this Agreement pursuant to DFARS 227.7202-3(a) or as set forth
+in subparagraphs (c)(1) and (2) of the Commercial Computer
+Software - Restricted Rights clause at FAR 52.227-19, as
+applicable. Contractor/manufacturer is NVIDIA, 2788 San Tomas
+Expressway, Santa Clara, CA 95051.
+
+The SDK is subject to United States export laws and
+regulations. You agree that you will not ship, transfer or
+export the SDK into any country, or use the SDK in any manner,
+prohibited by the United States Bureau of Industry and
+Security or economic sanctions regulations administered by the
+U.S. Department of Treasury’s Office of Foreign Assets
+Control (OFAC), or any applicable export laws, restrictions or
+regulations. These laws include restrictions on destinations,
+end users and end use. By accepting this Agreement, you
+confirm that you are not a resident or citizen of any country
+currently embargoed by the U.S. and that you are not otherwise
+prohibited from receiving the SDK.
+
+Any notice delivered by NVIDIA to you under this Agreement
+will be delivered via mail, email or fax. You agree that any
+notices that NVIDIA sends you electronically will satisfy any
+legal communication requirements. Please direct your legal
+notices or other correspondence to NVIDIA Corporation, 2788
+San Tomas Expressway, Santa Clara, California 95051, United
+States of America, Attention: Legal Department.
+
+This Agreement and any exhibits incorporated into this
+Agreement constitute the entire agreement of the parties with
+respect to the subject matter of this Agreement and supersede
+all prior negotiations or documentation exchanged between the
+parties relating to this SDK license. Any additional and/or
+conflicting terms on documents issued by you are null, void,
+and invalid. Any amendment or waiver under this Agreement
+shall be in writing and signed by representatives of both
+parties.
+
+
+2. CUDA Toolkit Supplement to Software License Agreement for
+NVIDIA Software Development Kits
+------------------------------------------------------------
+
+
+Release date: August 16, 2018
+-----------------------------
+
+The terms in this supplement govern your use of the NVIDIA
+CUDA Toolkit SDK under the terms of your license agreement
+(“Agreement”) as modified by this supplement. Capitalized
+terms used but not defined below have the meaning assigned to
+them in the Agreement.
+
+This supplement is an exhibit to the Agreement and is
+incorporated as an integral part of the Agreement. In the
+event of conflict between the terms in this supplement and the
+terms in the Agreement, the terms in this supplement govern.
+
+
+2.1. License Scope
+
+The SDK is licensed for you to develop applications only for
+use in systems with NVIDIA GPUs.
+
+
+2.2. Distribution
+
+The portions of the SDK that are distributable under the
+Agreement are listed in Attachment A.
+
+
+2.3. Operating Systems
+
+Those portions of the SDK designed exclusively for use on the
+Linux or FreeBSD operating systems, or other operating systems
+derived from the source code to these operating systems, may
+be copied and redistributed for use in accordance with this
+Agreement, provided that the object code files are not
+modified in any way (except for unzipping of compressed
+files).
+
+
+2.4. Audio and Video Encoders and Decoders
+
+You acknowledge and agree that it is your sole responsibility
+to obtain any additional third-party licenses required to
+make, have made, use, have used, sell, import, and offer for
+sale your products or services that include or incorporate any
+third-party software and content relating to audio and/or
+video encoders and decoders from, including but not limited
+to, Microsoft, Thomson, Fraunhofer IIS, Sisvel S.p.A.,
+MPEG-LA, and Coding Technologies. NVIDIA does not grant to you
+under this Agreement any necessary patent or other rights with
+respect to any audio and/or video encoders and decoders.
+
+
+2.5. Licensing
+
+If the distribution terms in this Agreement are not suitable
+for your organization, or for any questions regarding this
+Agreement, please contact NVIDIA at
+nvidia-compute-license-questions@nvidia.com.
+
+
+2.6. Attachment A
+
+The following portions of the SDK are distributable under the
+Agreement:
+
+Component
+
+CUDA Runtime
+
+Windows
+
+cudart.dll, cudart_static.lib, cudadevrt.lib
+
+Mac OSX
+
+libcudart.dylib, libcudart_static.a, libcudadevrt.a
+
+Linux
+
+libcudart.so, libcudart_static.a, libcudadevrt.a
+
+Android
+
+libcudart.so, libcudart_static.a, libcudadevrt.a
+
+Component
+
+CUDA FFT Library
+
+Windows
+
+cufft.dll, cufftw.dll, cufft.lib, cufftw.lib
+
+Mac OSX
+
+libcufft.dylib, libcufft_static.a, libcufftw.dylib,
+libcufftw_static.a
+
+Linux
+
+libcufft.so, libcufft_static.a, libcufftw.so,
+libcufftw_static.a
+
+Android
+
+libcufft.so, libcufft_static.a, libcufftw.so,
+libcufftw_static.a
+
+Component
+
+CUDA BLAS Library
+
+Windows
+
+cublas.dll, cublasLt.dll
+
+Mac OSX
+
+libcublas.dylib, libcublasLt.dylib, libcublas_static.a,
+libcublasLt_static.a
+
+Linux
+
+libcublas.so, libcublasLt.so, libcublas_static.a,
+libcublasLt_static.a
+
+Android
+
+libcublas.so, libcublasLt.so, libcublas_static.a,
+libcublasLt_static.a
+
+Component
+
+NVIDIA "Drop-in" BLAS Library
+
+Windows
+
+nvblas.dll
+
+Mac OSX
+
+libnvblas.dylib
+
+Linux
+
+libnvblas.so
+
+Component
+
+CUDA Sparse Matrix Library
+
+Windows
+
+cusparse.dll, cusparse.lib
+
+Mac OSX
+
+libcusparse.dylib, libcusparse_static.a
+
+Linux
+
+libcusparse.so, libcusparse_static.a
+
+Android
+
+libcusparse.so, libcusparse_static.a
+
+Component
+
+CUDA Linear Solver Library
+
+Windows
+
+cusolver.dll, cusolver.lib
+
+Mac OSX
+
+libcusolver.dylib, libcusolver_static.a
+
+Linux
+
+libcusolver.so, libcusolver_static.a
+
+Android
+
+libcusolver.so, libcusolver_static.a
+
+Component
+
+CUDA Random Number Generation Library
+
+Windows
+
+curand.dll, curand.lib
+
+Mac OSX
+
+libcurand.dylib, libcurand_static.a
+
+Linux
+
+libcurand.so, libcurand_static.a
+
+Android
+
+libcurand.so, libcurand_static.a
+
+Component
+
+CUDA Accelerated Graph Library
+
+Component
+
+NVIDIA Performance Primitives Library
+
+Windows
+
+nppc.dll, nppc.lib, nppial.dll, nppial.lib, nppicc.dll,
+nppicc.lib, nppicom.dll, nppicom.lib, nppidei.dll,
+nppidei.lib, nppif.dll, nppif.lib, nppig.dll, nppig.lib,
+nppim.dll, nppim.lib, nppist.dll, nppist.lib, nppisu.dll,
+nppisu.lib, nppitc.dll, nppitc.lib, npps.dll, npps.lib
+
+Mac OSX
+
+libnppc.dylib, libnppc_static.a, libnppial.dylib,
+libnppial_static.a, libnppicc.dylib, libnppicc_static.a,
+libnppicom.dylib, libnppicom_static.a, libnppidei.dylib,
+libnppidei_static.a, libnppif.dylib, libnppif_static.a,
+libnppig.dylib, libnppig_static.a, libnppim.dylib,
+libnppisu_static.a, libnppitc.dylib, libnppitc_static.a,
+libnpps.dylib, libnpps_static.a
+
+Linux
+
+libnppc.so, libnppc_static.a, libnppial.so,
+libnppial_static.a, libnppicc.so, libnppicc_static.a,
+libnppicom.so, libnppicom_static.a, libnppidei.so,
+libnppidei_static.a, libnppif.so, libnppif_static.a
+libnppig.so, libnppig_static.a, libnppim.so,
+libnppim_static.a, libnppist.so, libnppist_static.a,
+libnppisu.so, libnppisu_static.a, libnppitc.so
+libnppitc_static.a, libnpps.so, libnpps_static.a
+
+Android
+
+libnppc.so, libnppc_static.a, libnppial.so,
+libnppial_static.a, libnppicc.so, libnppicc_static.a,
+libnppicom.so, libnppicom_static.a, libnppidei.so,
+libnppidei_static.a, libnppif.so, libnppif_static.a
+libnppig.so, libnppig_static.a, libnppim.so,
+libnppim_static.a, libnppist.so, libnppist_static.a,
+libnppisu.so, libnppisu_static.a, libnppitc.so
+libnppitc_static.a, libnpps.so, libnpps_static.a
+
+Component
+
+NVIDIA JPEG Library
+
+Linux
+
+libnvjpeg.so, libnvjpeg_static.a
+
+Component
+
+Internal common library required for statically linking to
+cuBLAS, cuSPARSE, cuFFT, cuRAND, nvJPEG and NPP
+
+Mac OSX
+
+libculibos.a
+
+Linux
+
+libculibos.a
+
+Component
+
+NVIDIA Runtime Compilation Library and Header
+
+All
+
+nvrtc.h
+
+Windows
+
+nvrtc.dll, nvrtc-builtins.dll
+
+Mac OSX
+
+libnvrtc.dylib, libnvrtc-builtins.dylib
+
+Linux
+
+libnvrtc.so, libnvrtc-builtins.so
+
+Component
+
+NVIDIA Optimizing Compiler Library
+
+Windows
+
+nvvm.dll
+
+Mac OSX
+
+libnvvm.dylib
+
+Linux
+
+libnvvm.so
+
+Component
+
+NVIDIA Common Device Math Functions Library
+
+Windows
+
+libdevice.10.bc
+
+Mac OSX
+
+libdevice.10.bc
+
+Linux
+
+libdevice.10.bc
+
+Component
+
+CUDA Occupancy Calculation Header Library
+
+All
+
+cuda_occupancy.h
+
+Component
+
+CUDA Half Precision Headers
+
+All
+
+cuda_fp16.h, cuda_fp16.hpp
+
+Component
+
+CUDA Profiling Tools Interface (CUPTI) Library
+
+Windows
+
+cupti.dll
+
+Mac OSX
+
+libcupti.dylib
+
+Linux
+
+libcupti.so
+
+Component
+
+NVIDIA Tools Extension Library
+
+Windows
+
+nvToolsExt.dll, nvToolsExt.lib
+
+Mac OSX
+
+libnvToolsExt.dylib
+
+Linux
+
+libnvToolsExt.so
+
+Component
+
+NVIDIA CUDA Driver Libraries
+
+Linux
+
+libcuda.so, libnvidia-fatbinaryloader.so,
+libnvidia-ptxjitcompiler.so
+
+The NVIDIA CUDA Driver Libraries are only distributable in
+applications that meet this criteria:
+
+  1. The application was developed starting from a NVIDIA CUDA
+    container obtained from Docker Hub or the NVIDIA GPU
+    Cloud, and
+
+  2. The resulting application is packaged as a Docker
+    container and distributed to users on Docker Hub or the
+    NVIDIA GPU Cloud only.
+
+
+2.7. Attachment B
+
+
+Additional Licensing Obligations
+
+The following third party components included in the SOFTWARE
+are licensed to Licensee pursuant to the following terms and
+conditions:
+
+  1. Licensee's use of the GDB third party component is
+    subject to the terms and conditions of GNU GPL v3:
+
+    This product includes copyrighted third-party software licensed
+    under the terms of the GNU General Public License v3 ("GPL v3").
+    All third-party software packages are copyright by their respective
+    authors. GPL v3 terms and conditions are hereby incorporated into
+    the Agreement by this reference:     http://www.gnu.org/licenses/gpl.txt
+
+    Consistent with these licensing requirements, the software
+    listed below is provided under the terms of the specified
+    open source software licenses. To obtain source code for
+    software provided under licenses that require
+    redistribution of source code, including the GNU General
+    Public License (GPL) and GNU Lesser General Public License
+    (LGPL), contact oss-requests@nvidia.com. This offer is
+    valid for a period of three (3) years from the date of the
+    distribution of this product by NVIDIA CORPORATION.
+
+    Component          License
+    CUDA-GDB           GPL v3
+
+  2. Licensee represents and warrants that any and all third
+    party licensing and/or royalty payment obligations in
+    connection with Licensee's use of the H.264 video codecs
+    are solely the responsibility of Licensee.
+
+  3. Licensee's use of the Thrust library is subject to the
+    terms and conditions of the Apache License Version 2.0.
+    All third-party software packages are copyright by their
+    respective authors. Apache License Version 2.0 terms and
+    conditions are hereby incorporated into the Agreement by
+    this reference.
+    http://www.apache.org/licenses/LICENSE-2.0.html
+
+    In addition, Licensee acknowledges the following notice:
+    Thrust includes source code from the Boost Iterator,
+    Tuple, System, and Random Number libraries.
+
+    Boost Software License - Version 1.0 - August 17th, 2003
+    . . . .
+
+    Permission is hereby granted, free of charge, to any person or
+    organization obtaining a copy of the software and accompanying
+    documentation covered by this license (the "Software") to use,
+    reproduce, display, distribute, execute, and transmit the Software,
+    and to prepare derivative works of the Software, and to permit
+    third-parties to whom the Software is furnished to do so, all
+    subject to the following:
+
+    The copyright notices in the Software and this entire statement,
+    including the above license grant, this restriction and the following
+    disclaimer, must be included in all copies of the Software, in whole
+    or in part, and all derivative works of the Software, unless such
+    copies or derivative works are solely in the form of machine-executable
+    object code generated by a source language processor.
+
+    THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+    EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+    MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE, TITLE AND
+    NON-INFRINGEMENT. IN NO EVENT SHALL THE COPYRIGHT HOLDERS OR
+    ANYONE DISTRIBUTING THE SOFTWARE BE LIABLE FOR ANY DAMAGES OR
+    OTHER LIABILITY, WHETHER IN CONTRACT, TORT OR OTHERWISE, ARISING
+    FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+    OTHER DEALINGS IN THE SOFTWARE.
+
+  4. Licensee's use of the LLVM third party component is
+    subject to the following terms and conditions:
+
+    ======================================================
+    LLVM Release License
+    ======================================================
+    University of Illinois/NCSA
+    Open Source License
+
+    Copyright (c) 2003-2010 University of Illinois at Urbana-Champaign.
+    All rights reserved.
+
+    Developed by:
+
+        LLVM Team
+
+        University of Illinois at Urbana-Champaign
+
+        http://llvm.org
+
+    Permission is hereby granted, free of charge, to any person obtaining a copy
+    of this software and associated documentation files (the "Software"), to
+    deal with the Software without restriction, including without limitation the
+    rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+    sell copies of the Software, and to permit persons to whom the Software is
+    furnished to do so, subject to the following conditions:
+
+    *  Redistributions of source code must retain the above copyright notice,
+       this list of conditions and the following disclaimers.
+
+    *  Redistributions in binary form must reproduce the above copyright
+       notice, this list of conditions and the following disclaimers in the
+       documentation and/or other materials provided with the distribution.
+
+    *  Neither the names of the LLVM Team, University of Illinois at Urbana-
+       Champaign, nor the names of its contributors may be used to endorse or
+       promote products derived from this Software without specific prior
+       written permission.
+
+    THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+    IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+    FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+    THE CONTRIBUTORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
+    OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+    ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+    DEALINGS WITH THE SOFTWARE.
+
+  5. Licensee's use (e.g. nvprof) of the PCRE third party
+    component is subject to the following terms and
+    conditions:
+
+    ------------
+    PCRE LICENCE
+    ------------
+    PCRE is a library of functions to support regular expressions whose syntax
+    and semantics are as close as possible to those of the Perl 5 language.
+    Release 8 of PCRE is distributed under the terms of the "BSD" licence, as
+    specified below. The documentation for PCRE, supplied in the "doc"
+    directory, is distributed under the same terms as the software itself. The
+    basic library functions are written in C and are freestanding. Also
+    included in the distribution is a set of C++ wrapper functions, and a just-
+    in-time compiler that can be used to optimize pattern matching. These are
+    both optional features that can be omitted when the library is built.
+
+    THE BASIC LIBRARY FUNCTIONS
+    ---------------------------
+    Written by:       Philip Hazel
+    Email local part: ph10
+    Email domain:     cam.ac.uk
+    University of Cambridge Computing Service,
+    Cambridge, England.
+    Copyright (c) 1997-2012 University of Cambridge
+    All rights reserved.
+
+    PCRE JUST-IN-TIME COMPILATION SUPPORT
+    -------------------------------------
+    Written by:       Zoltan Herczeg
+    Email local part: hzmester
+    Emain domain:     freemail.hu
+    Copyright(c) 2010-2012 Zoltan Herczeg
+    All rights reserved.
+
+    STACK-LESS JUST-IN-TIME COMPILER
+    --------------------------------
+    Written by:       Zoltan Herczeg
+    Email local part: hzmester
+    Emain domain:     freemail.hu
+    Copyright(c) 2009-2012 Zoltan Herczeg
+    All rights reserved.
+
+    THE C++ WRAPPER FUNCTIONS
+    -------------------------
+    Contributed by:   Google Inc.
+    Copyright (c) 2007-2012, Google Inc.
+    All rights reserved.
+
+    THE "BSD" LICENCE
+    -----------------
+    Redistribution and use in source and binary forms, with or without
+    modification, are permitted provided that the following conditions are met:
+
+      * Redistributions of source code must retain the above copyright notice,
+        this list of conditions and the following disclaimer.
+
+      * Redistributions in binary form must reproduce the above copyright
+        notice, this list of conditions and the following disclaimer in the
+        documentation and/or other materials provided with the distribution.
+
+      * Neither the name of the University of Cambridge nor the name of Google
+        Inc. nor the names of their contributors may be used to endorse or
+        promote products derived from this software without specific prior
+        written permission.
+
+    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+    AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+    IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+    ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
+    LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+    CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+    SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+    INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+    CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+    ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+    POSSIBILITY OF SUCH DAMAGE.
+
+  6. Some of the cuBLAS library routines were written by or
+    derived from code written by Vasily Volkov and are subject
+    to the Modified Berkeley Software Distribution License as
+    follows:
+
+    Copyright (c) 2007-2009, Regents of the University of California
+
+    All rights reserved.
+
+    Redistribution and use in source and binary forms, with or without
+    modification, are permitted provided that the following conditions are
+    met:
+        * Redistributions of source code must retain the above copyright
+          notice, this list of conditions and the following disclaimer.
+        * Redistributions in binary form must reproduce the above
+          copyright notice, this list of conditions and the following
+          disclaimer in the documentation and/or other materials provided
+          with the distribution.
+        * Neither the name of the University of California, Berkeley nor
+          the names of its contributors may be used to endorse or promote
+          products derived from this software without specific prior
+          written permission.
+
+    THIS SOFTWARE IS PROVIDED BY THE AUTHOR "AS IS" AND ANY EXPRESS OR
+    IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+    WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+    DISCLAIMED. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT,
+    INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+    (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+    SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
+    HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+    STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING
+    IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+    POSSIBILITY OF SUCH DAMAGE.
+
+  7. Some of the cuBLAS library routines were written by or
+    derived from code written by Davide Barbieri and are
+    subject to the Modified Berkeley Software Distribution
+    License as follows:
+
+    Copyright (c) 2008-2009 Davide Barbieri @ University of Rome Tor Vergata.
+
+    All rights reserved.
+
+    Redistribution and use in source and binary forms, with or without
+    modification, are permitted provided that the following conditions are
+    met:
+        * Redistributions of source code must retain the above copyright
+          notice, this list of conditions and the following disclaimer.
+        * Redistributions in binary form must reproduce the above
+          copyright notice, this list of conditions and the following
+          disclaimer in the documentation and/or other materials provided
+          with the distribution.
+        * The name of the author may not be used to endorse or promote
+          products derived from this software without specific prior
+          written permission.
+
+    THIS SOFTWARE IS PROVIDED BY THE AUTHOR "AS IS" AND ANY EXPRESS OR
+    IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+    WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+    DISCLAIMED. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT,
+    INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+    (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+    SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
+    HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+    STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING
+    IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+    POSSIBILITY OF SUCH DAMAGE.
+
+  8. Some of the cuBLAS library routines were derived from
+    code developed by the University of Tennessee and are
+    subject to the Modified Berkeley Software Distribution
+    License as follows:
+
+    Copyright (c) 2010 The University of Tennessee.
+
+    All rights reserved.
+
+    Redistribution and use in source and binary forms, with or without
+    modification, are permitted provided that the following conditions are
+    met:
+        * Redistributions of source code must retain the above copyright
+          notice, this list of conditions and the following disclaimer.
+        * Redistributions in binary form must reproduce the above
+          copyright notice, this list of conditions and the following
+          disclaimer listed in this license in the documentation and/or
+          other materials provided with the distribution.
+        * Neither the name of the copyright holders nor the names of its
+          contributors may be used to endorse or promote products derived
+          from this software without specific prior written permission.
+
+    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+    "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+    A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+    OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+    SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+    LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+    DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+    THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+    (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+    OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+  9. Some of the cuBLAS library routines were written by or
+    derived from code written by Jonathan Hogg and are subject
+    to the Modified Berkeley Software Distribution License as
+    follows:
+
+    Copyright (c) 2012, The Science and Technology Facilities Council (STFC).
+
+    All rights reserved.
+
+    Redistribution and use in source and binary forms, with or without
+    modification, are permitted provided that the following conditions are
+    met:
+        * Redistributions of source code must retain the above copyright
+          notice, this list of conditions and the following disclaimer.
+        * Redistributions in binary form must reproduce the above
+          copyright notice, this list of conditions and the following
+          disclaimer in the documentation and/or other materials provided
+          with the distribution.
+        * Neither the name of the STFC nor the names of its contributors
+          may be used to endorse or promote products derived from this
+          software without specific prior written permission.
+
+    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+    "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+    A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE STFC BE
+    LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+    CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+    SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR
+    BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
+    WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE
+    OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN
+    IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+  10. Some of the cuBLAS library routines were written by or
+    derived from code written by Ahmad M. Abdelfattah, David
+    Keyes, and Hatem Ltaief, and are subject to the Apache
+    License, Version 2.0, as follows:
+
+     -- (C) Copyright 2013 King Abdullah University of Science and Technology
+      Authors:
+      Ahmad Abdelfattah (ahmad.ahmad@kaust.edu.sa)
+      David Keyes (david.keyes@kaust.edu.sa)
+      Hatem Ltaief (hatem.ltaief@kaust.edu.sa)
+
+      Redistribution  and  use  in  source and binary forms, with or without
+      modification,  are  permitted  provided  that the following conditions
+      are met:
+
+      * Redistributions  of  source  code  must  retain  the above copyright
+        notice,  this  list  of  conditions  and  the  following  disclaimer.
+      * Redistributions  in  binary  form must reproduce the above copyright
+        notice,  this list of conditions and the following disclaimer in the
+        documentation  and/or other materials provided with the distribution.
+      * Neither  the  name of the King Abdullah University of Science and
+        Technology nor the names of its contributors may be used to endorse
+        or promote products derived from this software without specific prior
+        written permission.
+
+      THIS  SOFTWARE  IS  PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+      ``AS IS''  AND  ANY  EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+      LIMITED  TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+      A  PARTICULAR  PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+      HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+      SPECIAL,  EXEMPLARY,  OR  CONSEQUENTIAL  DAMAGES  (INCLUDING,  BUT NOT
+      LIMITED  TO,  PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+      DATA,  OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+      THEORY  OF  LIABILITY,  WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+      (INCLUDING  NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+      OF  THIS  SOFTWARE,  EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE
+
+  11. Some of the cuSPARSE library routines were written by or
+    derived from code written by Li-Wen Chang and are subject
+    to the NCSA Open Source License as follows:
+
+    Copyright (c) 2012, University of Illinois.
+
+    All rights reserved.
+
+    Developed by: IMPACT Group, University of Illinois, http://impact.crhc.illinois.edu
+
+    Permission is hereby granted, free of charge, to any person obtaining
+    a copy of this software and associated documentation files (the
+    "Software"), to deal with the Software without restriction, including
+    without limitation the rights to use, copy, modify, merge, publish,
+    distribute, sublicense, and/or sell copies of the Software, and to
+    permit persons to whom the Software is furnished to do so, subject to
+    the following conditions:
+        * Redistributions of source code must retain the above copyright
+          notice, this list of conditions and the following disclaimer.
+        * Redistributions in binary form must reproduce the above
+          copyright notice, this list of conditions and the following
+          disclaimers in the documentation and/or other materials provided
+          with the distribution.
+        * Neither the names of IMPACT Group, University of Illinois, nor
+          the names of its contributors may be used to endorse or promote
+          products derived from this Software without specific prior
+          written permission.
+
+    THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+    EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+    MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+    NONINFRINGEMENT. IN NO EVENT SHALL THE CONTRIBUTORS OR COPYRIGHT
+    HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER
+    IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR
+    IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS WITH THE
+    SOFTWARE.
+
+  12. Some of the cuRAND library routines were written by or
+    derived from code written by Mutsuo Saito and Makoto
+    Matsumoto and are subject to the following license:
+
+    Copyright (c) 2009, 2010 Mutsuo Saito, Makoto Matsumoto and Hiroshima
+    University. All rights reserved.
+
+    Copyright (c) 2011 Mutsuo Saito, Makoto Matsumoto, Hiroshima
+    University and University of Tokyo.  All rights reserved.
+
+    Redistribution and use in source and binary forms, with or without
+    modification, are permitted provided that the following conditions are
+    met:
+        * Redistributions of source code must retain the above copyright
+          notice, this list of conditions and the following disclaimer.
+        * Redistributions in binary form must reproduce the above
+          copyright notice, this list of conditions and the following
+          disclaimer in the documentation and/or other materials provided
+          with the distribution.
+        * Neither the name of the Hiroshima University nor the names of
+          its contributors may be used to endorse or promote products
+          derived from this software without specific prior written
+          permission.
+
+    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+    "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+    A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+    OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+    SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+    LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+    DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+    THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+    (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+    OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+  13. Some of the cuRAND library routines were derived from
+    code developed by D. E. Shaw Research and are subject to
+    the following license:
+
+    Copyright 2010-2011, D. E. Shaw Research.
+
+    All rights reserved.
+
+    Redistribution and use in source and binary forms, with or without
+    modification, are permitted provided that the following conditions are
+    met:
+        * Redistributions of source code must retain the above copyright
+          notice, this list of conditions, and the following disclaimer.
+        * Redistributions in binary form must reproduce the above
+          copyright notice, this list of conditions, and the following
+          disclaimer in the documentation and/or other materials provided
+          with the distribution.
+        * Neither the name of D. E. Shaw Research nor the names of its
+          contributors may be used to endorse or promote products derived
+          from this software without specific prior written permission.
+
+    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+    "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+    A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+    OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+    SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+    LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+    DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+    THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+    (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+    OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+  14. Some of the Math library routines were written by or
+    derived from code developed by Norbert Juffa and are
+    subject to the following license:
+
+    Copyright (c) 2015-2017, Norbert Juffa
+    All rights reserved.
+
+    Redistribution and use in source and binary forms, with or without
+    modification, are permitted provided that the following conditions
+    are met:
+
+    1. Redistributions of source code must retain the above copyright
+       notice, this list of conditions and the following disclaimer.
+
+    2. Redistributions in binary form must reproduce the above copyright
+       notice, this list of conditions and the following disclaimer in the
+       documentation and/or other materials provided with the distribution.
+
+    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+    "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+    A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+    HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+    SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+    LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+    DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+    THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+    (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+    OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+  15. Licensee's use of the lz4 third party component is
+    subject to the following terms and conditions:
+
+    Copyright (C) 2011-2013, Yann Collet.
+    BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+    Redistribution and use in source and binary forms, with or without
+    modification, are permitted provided that the following conditions are
+    met:
+
+        * Redistributions of source code must retain the above copyright
+    notice, this list of conditions and the following disclaimer.
+        * Redistributions in binary form must reproduce the above
+    copyright notice, this list of conditions and the following disclaimer
+    in the documentation and/or other materials provided with the
+    distribution.
+
+    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+    "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+    A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+    OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+    SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+    LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+    DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+    THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+    (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+    OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+  16. The NPP library uses code from the Boost Math Toolkit,
+    and is subject to the following license:
+
+    Boost Software License - Version 1.0 - August 17th, 2003
+    . . . .
+
+    Permission is hereby granted, free of charge, to any person or
+    organization obtaining a copy of the software and accompanying
+    documentation covered by this license (the "Software") to use,
+    reproduce, display, distribute, execute, and transmit the Software,
+    and to prepare derivative works of the Software, and to permit
+    third-parties to whom the Software is furnished to do so, all
+    subject to the following:
+
+    The copyright notices in the Software and this entire statement,
+    including the above license grant, this restriction and the following
+    disclaimer, must be included in all copies of the Software, in whole
+    or in part, and all derivative works of the Software, unless such
+    copies or derivative works are solely in the form of machine-executable
+    object code generated by a source language processor.
+
+    THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+    EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+    MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE, TITLE AND
+    NON-INFRINGEMENT. IN NO EVENT SHALL THE COPYRIGHT HOLDERS OR
+    ANYONE DISTRIBUTING THE SOFTWARE BE LIABLE FOR ANY DAMAGES OR
+    OTHER LIABILITY, WHETHER IN CONTRACT, TORT OR OTHERWISE, ARISING
+    FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+    OTHER DEALINGS IN THE SOFTWARE.
+
+  17. Portions of the Nsight Eclipse Edition is subject to the
+    following license:
+
+    The Eclipse Foundation makes available all content in this plug-in
+    ("Content"). Unless otherwise indicated below, the Content is provided
+    to you under the terms and conditions of the Eclipse Public License
+    Version 1.0 ("EPL"). A copy of the EPL is available at http://
+    www.eclipse.org/legal/epl-v10.html. For purposes of the EPL, "Program"
+    will mean the Content.
+
+    If you did not receive this Content directly from the Eclipse
+    Foundation, the Content is being redistributed by another party
+    ("Redistributor") and different terms and conditions may apply to your
+    use of any object code in the Content. Check the Redistributor's
+    license that was provided with the Content. If no such license exists,
+    contact the Redistributor. Unless otherwise indicated below, the terms
+    and conditions of the EPL still apply to any source code in the
+    Content and such source code may be obtained at http://www.eclipse.org.
+
+  18. Some of the cuBLAS library routines uses code from
+    OpenAI, which is subject to the following license:
+
+    License URL
+    https://github.com/openai/openai-gemm/blob/master/LICENSE
+
+    License Text
+    The MIT License
+
+    Copyright (c) 2016 OpenAI (http://openai.com), 2016 Google Inc.
+
+    Permission is hereby granted, free of charge, to any person obtaining a copy
+    of this software and associated documentation files (the "Software"), to deal
+    in the Software without restriction, including without limitation the rights
+    to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+    copies of the Software, and to permit persons to whom the Software is
+    furnished to do so, subject to the following conditions:
+
+    The above copyright notice and this permission notice shall be included in
+    all copies or substantial portions of the Software.
+
+    THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+    IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+    FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+    AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+    LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+    OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+    THE SOFTWARE.
+
+  19. Licensee's use of the Visual Studio Setup Configuration
+    Samples is subject to the following license:
+
+    The MIT License (MIT)
+    Copyright (C) Microsoft Corporation. All rights reserved.
+
+    Permission is hereby granted, free of charge, to any person
+    obtaining a copy of this software and associated documentation
+    files (the "Software"), to deal in the Software without restriction,
+    including without limitation the rights to use, copy, modify, merge,
+    publish, distribute, sublicense, and/or sell copies of the Software,
+    and to permit persons to whom the Software is furnished to do so,
+    subject to the following conditions:
+
+    The above copyright notice and this permission notice shall be included
+    in all copies or substantial portions of the Software.
+
+    THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
+    OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+    FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+    AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+    LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+    OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+  20. Licensee's use of linmath.h header for CPU functions for
+    GL vector/matrix operations from lunarG is subject to the
+    Apache License Version 2.0.
+
+  21. The DX12-CUDA sample uses the d3dx12.h header, which is
+    subject to the MIT license .
+
+-----------------

tuning-competition-baseline/.venv/lib/python3.11/site-packages/nvidia_cuda_runtime_cu11-11.8.89.dist-info/License.txt

@@ -0,0 +1,1568 @@
+End User License Agreement
+--------------------------
+
+
+Preface
+-------
+
+The Software License Agreement in Chapter 1 and the Supplement
+in Chapter 2 contain license terms and conditions that govern
+the use of NVIDIA software. By accepting this agreement, you
+agree to comply with all the terms and conditions applicable
+to the product(s) included herein.
+
+
+NVIDIA Driver
+
+
+Description
+
+This package contains the operating system driver and
+fundamental system software components for NVIDIA GPUs.
+
+
+NVIDIA CUDA Toolkit
+
+
+Description
+
+The NVIDIA CUDA Toolkit provides command-line and graphical
+tools for building, debugging and optimizing the performance
+of applications accelerated by NVIDIA GPUs, runtime and math
+libraries, and documentation including programming guides,
+user manuals, and API references.
+
+
+Default Install Location of CUDA Toolkit
+
+Windows platform:
+
+%ProgramFiles%\NVIDIA GPU Computing Toolkit\CUDA\v#.#
+
+Linux platform:
+
+/usr/local/cuda-#.#
+
+Mac platform:
+
+/Developer/NVIDIA/CUDA-#.#
+
+
+NVIDIA CUDA Samples
+
+
+Description
+
+This package includes over 100+ CUDA examples that demonstrate
+various CUDA programming principles, and efficient CUDA
+implementation of algorithms in specific application domains.
+
+
+Default Install Location of CUDA Samples
+
+Windows platform:
+
+%ProgramData%\NVIDIA Corporation\CUDA Samples\v#.#
+
+Linux platform:
+
+/usr/local/cuda-#.#/samples
+
+and
+
+$HOME/NVIDIA_CUDA-#.#_Samples
+
+Mac platform:
+
+/Developer/NVIDIA/CUDA-#.#/samples
+
+
+NVIDIA Nsight Visual Studio Edition (Windows only)
+
+
+Description
+
+NVIDIA Nsight Development Platform, Visual Studio Edition is a
+development environment integrated into Microsoft Visual
+Studio that provides tools for debugging, profiling, analyzing
+and optimizing your GPU computing and graphics applications.
+
+
+Default Install Location of Nsight Visual Studio Edition
+
+Windows platform:
+
+%ProgramFiles(x86)%\NVIDIA Corporation\Nsight Visual Studio Edition #.#
+
+
+1. License Agreement for NVIDIA Software Development Kits
+---------------------------------------------------------
+
+
+Release Date: July 26, 2018
+---------------------------
+
+
+Important NoticeRead before downloading, installing,
+copying or using the licensed software:
+-------------------------------------------------------
+
+This license agreement, including exhibits attached
+("Agreement”) is a legal agreement between you and NVIDIA
+Corporation ("NVIDIA") and governs your use of a NVIDIA
+software development kit (“SDK”).
+
+Each SDK has its own set of software and materials, but here
+is a description of the types of items that may be included in
+a SDK: source code, header files, APIs, data sets and assets
+(examples include images, textures, models, scenes, videos,
+native API input/output files), binary software, sample code,
+libraries, utility programs, programming code and
+documentation.
+
+This Agreement can be accepted only by an adult of legal age
+of majority in the country in which the SDK is used.
+
+If you are entering into this Agreement on behalf of a company
+or other legal entity, you represent that you have the legal
+authority to bind the entity to this Agreement, in which case
+“you” will mean the entity you represent.
+
+If you don’t have the required age or authority to accept
+this Agreement, or if you don’t accept all the terms and
+conditions of this Agreement, do not download, install or use
+the SDK.
+
+You agree to use the SDK only for purposes that are permitted
+by (a) this Agreement, and (b) any applicable law, regulation
+or generally accepted practices or guidelines in the relevant
+jurisdictions.
+
+
+1.1. License
+
+
+1.1.1. License Grant
+
+Subject to the terms of this Agreement, NVIDIA hereby grants
+you a non-exclusive, non-transferable license, without the
+right to sublicense (except as expressly provided in this
+Agreement) to:
+
+  1. Install and use the SDK,
+
+  2. Modify and create derivative works of sample source code
+    delivered in the SDK, and
+
+  3. Distribute those portions of the SDK that are identified
+    in this Agreement as distributable, as incorporated in
+    object code format into a software application that meets
+    the distribution requirements indicated in this Agreement.
+
+
+1.1.2. Distribution Requirements
+
+These are the distribution requirements for you to exercise
+the distribution grant:
+
+  1. Your application must have material additional
+    functionality, beyond the included portions of the SDK.
+
+  2. The distributable portions of the SDK shall only be
+    accessed by your application.
+
+  3. The following notice shall be included in modifications
+    and derivative works of sample source code distributed:
+    “This software contains source code provided by NVIDIA
+    Corporation.”
+
+  4. Unless a developer tool is identified in this Agreement
+    as distributable, it is delivered for your internal use
+    only.
+
+  5. The terms under which you distribute your application
+    must be consistent with the terms of this Agreement,
+    including (without limitation) terms relating to the
+    license grant and license restrictions and protection of
+    NVIDIA’s intellectual property rights. Additionally, you
+    agree that you will protect the privacy, security and
+    legal rights of your application users.
+
+  6. You agree to notify NVIDIA in writing of any known or
+    suspected distribution or use of the SDK not in compliance
+    with the requirements of this Agreement, and to enforce
+    the terms of your agreements with respect to distributed
+    SDK.
+
+
+1.1.3. Authorized Users
+
+You may allow employees and contractors of your entity or of
+your subsidiary(ies) to access and use the SDK from your
+secure network to perform work on your behalf.
+
+If you are an academic institution you may allow users
+enrolled or employed by the academic institution to access and
+use the SDK from your secure network.
+
+You are responsible for the compliance with the terms of this
+Agreement by your authorized users. If you become aware that
+your authorized users didn’t follow the terms of this
+Agreement, you agree to take reasonable steps to resolve the
+non-compliance and prevent new occurrences.
+
+
+1.1.4. Pre-Release SDK
+
+The SDK versions identified as alpha, beta, preview or
+otherwise as pre-release, may not be fully functional, may
+contain errors or design flaws, and may have reduced or
+different security, privacy, accessibility, availability, and
+reliability standards relative to commercial versions of
+NVIDIA software and materials. Use of a pre-release SDK may
+result in unexpected results, loss of data, project delays or
+other unpredictable damage or loss.
+
+You may use a pre-release SDK at your own risk, understanding
+that pre-release SDKs are not intended for use in production
+or business-critical systems.
+
+NVIDIA may choose not to make available a commercial version
+of any pre-release SDK. NVIDIA may also choose to abandon
+development and terminate the availability of a pre-release
+SDK at any time without liability.
+
+
+1.1.5. Updates
+
+NVIDIA may, at its option, make available patches, workarounds
+or other updates to this SDK. Unless the updates are provided
+with their separate governing terms, they are deemed part of
+the SDK licensed to you as provided in this Agreement. You
+agree that the form and content of the SDK that NVIDIA
+provides may change without prior notice to you. While NVIDIA
+generally maintains compatibility between versions, NVIDIA may
+in some cases make changes that introduce incompatibilities in
+future versions of the SDK.
+
+
+1.1.6. Third Party Licenses
+
+The SDK may come bundled with, or otherwise include or be
+distributed with, third party software licensed by a NVIDIA
+supplier and/or open source software provided under an open
+source license. Use of third party software is subject to the
+third-party license terms, or in the absence of third party
+terms, the terms of this Agreement. Copyright to third party
+software is held by the copyright holders indicated in the
+third-party software or license.
+
+
+1.1.7. Reservation of Rights
+
+NVIDIA reserves all rights, title, and interest in and to the
+SDK, not expressly granted to you under this Agreement.
+
+
+1.2. Limitations
+
+The following license limitations apply to your use of the
+SDK:
+
+  1. You may not reverse engineer, decompile or disassemble,
+    or remove copyright or other proprietary notices from any
+    portion of the SDK or copies of the SDK.
+
+  2. Except as expressly provided in this Agreement, you may
+    not copy, sell, rent, sublicense, transfer, distribute,
+    modify, or create derivative works of any portion of the
+    SDK. For clarity, you may not distribute or sublicense the
+    SDK as a stand-alone product.
+
+  3. Unless you have an agreement with NVIDIA for this
+    purpose, you may not indicate that an application created
+    with the SDK is sponsored or endorsed by NVIDIA.
+
+  4. You may not bypass, disable, or circumvent any
+    encryption, security, digital rights management or
+    authentication mechanism in the SDK.
+
+  5. You may not use the SDK in any manner that would cause it
+    to become subject to an open source software license. As
+    examples, licenses that require as a condition of use,
+    modification, and/or distribution that the SDK be:
+
+      a. Disclosed or distributed in source code form;
+
+      b. Licensed for the purpose of making derivative works;
+        or
+
+      c. Redistributable at no charge.
+
+  6. Unless you have an agreement with NVIDIA for this
+    purpose, you may not use the SDK with any system or
+    application where the use or failure of the system or
+    application can reasonably be expected to threaten or
+    result in personal injury, death, or catastrophic loss.
+    Examples include use in avionics, navigation, military,
+    medical, life support or other life critical applications.
+    NVIDIA does not design, test or manufacture the SDK for
+    these critical uses and NVIDIA shall not be liable to you
+    or any third party, in whole or in part, for any claims or
+    damages arising from such uses.
+
+  7. You agree to defend, indemnify and hold harmless NVIDIA
+    and its affiliates, and their respective employees,
+    contractors, agents, officers and directors, from and
+    against any and all claims, damages, obligations, losses,
+    liabilities, costs or debt, fines, restitutions and
+    expenses (including but not limited to attorney’s fees
+    and costs incident to establishing the right of
+    indemnification) arising out of or related to your use of
+    the SDK outside of the scope of this Agreement, or not in
+    compliance with its terms.
+
+
+1.3. Ownership
+
+  1.  NVIDIA or its licensors hold all rights, title and
+    interest in and to the SDK and its modifications and
+    derivative works, including their respective intellectual
+    property rights, subject to your rights described in this
+    section. This SDK may include software and materials from
+    NVIDIA’s licensors, and these licensors are intended
+    third party beneficiaries that may enforce this Agreement
+    with respect to their intellectual property rights.
+
+  2.  You hold all rights, title and interest in and to your
+    applications and your derivative works of the sample
+    source code delivered in the SDK, including their
+    respective intellectual property rights, subject to
+    NVIDIA’s rights described in this section.
+
+  3. You may, but don’t have to, provide to NVIDIA
+    suggestions, feature requests or other feedback regarding
+    the SDK, including possible enhancements or modifications
+    to the SDK. For any feedback that you voluntarily provide,
+    you hereby grant NVIDIA and its affiliates a perpetual,
+    non-exclusive, worldwide, irrevocable license to use,
+    reproduce, modify, license, sublicense (through multiple
+    tiers of sublicensees), and distribute (through multiple
+    tiers of distributors) it without the payment of any
+    royalties or fees to you. NVIDIA will use feedback at its
+    choice. NVIDIA is constantly looking for ways to improve
+    its products, so you may send feedback to NVIDIA through
+    the developer portal at https://developer.nvidia.com.
+
+
+1.4. No Warranties
+
+THE SDK IS PROVIDED BY NVIDIA “AS IS” AND “WITH ALL
+FAULTS.” TO THE MAXIMUM EXTENT PERMITTED BY LAW, NVIDIA AND
+ITS AFFILIATES EXPRESSLY DISCLAIM ALL WARRANTIES OF ANY KIND
+OR NATURE, WHETHER EXPRESS, IMPLIED OR STATUTORY, INCLUDING,
+BUT NOT LIMITED TO, ANY WARRANTIES OF MERCHANTABILITY, FITNESS
+FOR A PARTICULAR PURPOSE, TITLE, NON-INFRINGEMENT, OR THE
+ABSENCE OF ANY DEFECTS THEREIN, WHETHER LATENT OR PATENT. NO
+WARRANTY IS MADE ON THE BASIS OF TRADE USAGE, COURSE OF
+DEALING OR COURSE OF TRADE.
+
+
+1.5. Limitation of Liability
+
+TO THE MAXIMUM EXTENT PERMITTED BY LAW, NVIDIA AND ITS
+AFFILIATES SHALL NOT BE LIABLE FOR ANY SPECIAL, INCIDENTAL,
+PUNITIVE OR CONSEQUENTIAL DAMAGES, OR ANY LOST PROFITS, LOSS
+OF USE, LOSS OF DATA OR LOSS OF GOODWILL, OR THE COSTS OF
+PROCURING SUBSTITUTE PRODUCTS, ARISING OUT OF OR IN CONNECTION
+WITH THIS AGREEMENT OR THE USE OR PERFORMANCE OF THE SDK,
+WHETHER SUCH LIABILITY ARISES FROM ANY CLAIM BASED UPON BREACH
+OF CONTRACT, BREACH OF WARRANTY, TORT (INCLUDING NEGLIGENCE),
+PRODUCT LIABILITY OR ANY OTHER CAUSE OF ACTION OR THEORY OF
+LIABILITY. IN NO EVENT WILL NVIDIA’S AND ITS AFFILIATES
+TOTAL CUMULATIVE LIABILITY UNDER OR ARISING OUT OF THIS
+AGREEMENT EXCEED US$10.00. THE NATURE OF THE LIABILITY OR THE
+NUMBER OF CLAIMS OR SUITS SHALL NOT ENLARGE OR EXTEND THIS
+LIMIT.
+
+These exclusions and limitations of liability shall apply
+regardless if NVIDIA or its affiliates have been advised of
+the possibility of such damages, and regardless of whether a
+remedy fails its essential purpose. These exclusions and
+limitations of liability form an essential basis of the
+bargain between the parties, and, absent any of these
+exclusions or limitations of liability, the provisions of this
+Agreement, including, without limitation, the economic terms,
+would be substantially different.
+
+
+1.6. Termination
+
+  1. This Agreement will continue to apply until terminated by
+    either you or NVIDIA as described below.
+
+  2. If you want to terminate this Agreement, you may do so by
+    stopping to use the SDK.
+
+  3. NVIDIA may, at any time, terminate this Agreement if:
+
+      a. (i) you fail to comply with any term of this
+        Agreement and the non-compliance is not fixed within
+        thirty (30) days following notice from NVIDIA (or
+        immediately if you violate NVIDIA’s intellectual
+        property rights);
+
+      b. (ii) you commence or participate in any legal
+        proceeding against NVIDIA with respect to the SDK; or
+
+      c. (iii) NVIDIA decides to no longer provide the SDK in
+        a country or, in NVIDIA’s sole discretion, the
+        continued use of it is no longer commercially viable.
+
+  4. Upon any termination of this Agreement, you agree to
+    promptly discontinue use of the SDK and destroy all copies
+    in your possession or control. Your prior distributions in
+    accordance with this Agreement are not affected by the
+    termination of this Agreement. Upon written request, you
+    will certify in writing that you have complied with your
+    commitments under this section. Upon any termination of
+    this Agreement all provisions survive except for the
+    license grant provisions.
+
+
+1.7. General
+
+If you wish to assign this Agreement or your rights and
+obligations, including by merger, consolidation, dissolution
+or operation of law, contact NVIDIA to ask for permission. Any
+attempted assignment not approved by NVIDIA in writing shall
+be void and of no effect. NVIDIA may assign, delegate or
+transfer this Agreement and its rights and obligations, and if
+to a non-affiliate you will be notified.
+
+You agree to cooperate with NVIDIA and provide reasonably
+requested information to verify your compliance with this
+Agreement.
+
+This Agreement will be governed in all respects by the laws of
+the United States and of the State of Delaware as those laws
+are applied to contracts entered into and performed entirely
+within Delaware by Delaware residents, without regard to the
+conflicts of laws principles. The United Nations Convention on
+Contracts for the International Sale of Goods is specifically
+disclaimed. You agree to all terms of this Agreement in the
+English language.
+
+The state or federal courts residing in Santa Clara County,
+California shall have exclusive jurisdiction over any dispute
+or claim arising out of this Agreement. Notwithstanding this,
+you agree that NVIDIA shall still be allowed to apply for
+injunctive remedies or an equivalent type of urgent legal
+relief in any jurisdiction.
+
+If any court of competent jurisdiction determines that any
+provision of this Agreement is illegal, invalid or
+unenforceable, such provision will be construed as limited to
+the extent necessary to be consistent with and fully
+enforceable under the law and the remaining provisions will
+remain in full force and effect. Unless otherwise specified,
+remedies are cumulative.
+
+Each party acknowledges and agrees that the other is an
+independent contractor in the performance of this Agreement.
+
+The SDK has been developed entirely at private expense and is
+“commercial items” consisting of “commercial computer
+software” and “commercial computer software
+documentation” provided with RESTRICTED RIGHTS. Use,
+duplication or disclosure by the U.S. Government or a U.S.
+Government subcontractor is subject to the restrictions in
+this Agreement pursuant to DFARS 227.7202-3(a) or as set forth
+in subparagraphs (c)(1) and (2) of the Commercial Computer
+Software - Restricted Rights clause at FAR 52.227-19, as
+applicable. Contractor/manufacturer is NVIDIA, 2788 San Tomas
+Expressway, Santa Clara, CA 95051.
+
+The SDK is subject to United States export laws and
+regulations. You agree that you will not ship, transfer or
+export the SDK into any country, or use the SDK in any manner,
+prohibited by the United States Bureau of Industry and
+Security or economic sanctions regulations administered by the
+U.S. Department of Treasury’s Office of Foreign Assets
+Control (OFAC), or any applicable export laws, restrictions or
+regulations. These laws include restrictions on destinations,
+end users and end use. By accepting this Agreement, you
+confirm that you are not a resident or citizen of any country
+currently embargoed by the U.S. and that you are not otherwise
+prohibited from receiving the SDK.
+
+Any notice delivered by NVIDIA to you under this Agreement
+will be delivered via mail, email or fax. You agree that any
+notices that NVIDIA sends you electronically will satisfy any
+legal communication requirements. Please direct your legal
+notices or other correspondence to NVIDIA Corporation, 2788
+San Tomas Expressway, Santa Clara, California 95051, United
+States of America, Attention: Legal Department.
+
+This Agreement and any exhibits incorporated into this
+Agreement constitute the entire agreement of the parties with
+respect to the subject matter of this Agreement and supersede
+all prior negotiations or documentation exchanged between the
+parties relating to this SDK license. Any additional and/or
+conflicting terms on documents issued by you are null, void,
+and invalid. Any amendment or waiver under this Agreement
+shall be in writing and signed by representatives of both
+parties.
+
+
+2. CUDA Toolkit Supplement to Software License Agreement for
+NVIDIA Software Development Kits
+------------------------------------------------------------
+
+
+Release date: August 16, 2018
+-----------------------------
+
+The terms in this supplement govern your use of the NVIDIA
+CUDA Toolkit SDK under the terms of your license agreement
+(“Agreement”) as modified by this supplement. Capitalized
+terms used but not defined below have the meaning assigned to
+them in the Agreement.
+
+This supplement is an exhibit to the Agreement and is
+incorporated as an integral part of the Agreement. In the
+event of conflict between the terms in this supplement and the
+terms in the Agreement, the terms in this supplement govern.
+
+
+2.1. License Scope
+
+The SDK is licensed for you to develop applications only for
+use in systems with NVIDIA GPUs.
+
+
+2.2. Distribution
+
+The portions of the SDK that are distributable under the
+Agreement are listed in Attachment A.
+
+
+2.3. Operating Systems
+
+Those portions of the SDK designed exclusively for use on the
+Linux or FreeBSD operating systems, or other operating systems
+derived from the source code to these operating systems, may
+be copied and redistributed for use in accordance with this
+Agreement, provided that the object code files are not
+modified in any way (except for unzipping of compressed
+files).
+
+
+2.4. Audio and Video Encoders and Decoders
+
+You acknowledge and agree that it is your sole responsibility
+to obtain any additional third-party licenses required to
+make, have made, use, have used, sell, import, and offer for
+sale your products or services that include or incorporate any
+third-party software and content relating to audio and/or
+video encoders and decoders from, including but not limited
+to, Microsoft, Thomson, Fraunhofer IIS, Sisvel S.p.A.,
+MPEG-LA, and Coding Technologies. NVIDIA does not grant to you
+under this Agreement any necessary patent or other rights with
+respect to any audio and/or video encoders and decoders.
+
+
+2.5. Licensing
+
+If the distribution terms in this Agreement are not suitable
+for your organization, or for any questions regarding this
+Agreement, please contact NVIDIA at
+nvidia-compute-license-questions@nvidia.com.
+
+
+2.6. Attachment A
+
+The following portions of the SDK are distributable under the
+Agreement:
+
+Component
+
+CUDA Runtime
+
+Windows
+
+cudart.dll, cudart_static.lib, cudadevrt.lib
+
+Mac OSX
+
+libcudart.dylib, libcudart_static.a, libcudadevrt.a
+
+Linux
+
+libcudart.so, libcudart_static.a, libcudadevrt.a
+
+Android
+
+libcudart.so, libcudart_static.a, libcudadevrt.a
+
+Component
+
+CUDA FFT Library
+
+Windows
+
+cufft.dll, cufftw.dll, cufft.lib, cufftw.lib
+
+Mac OSX
+
+libcufft.dylib, libcufft_static.a, libcufftw.dylib,
+libcufftw_static.a
+
+Linux
+
+libcufft.so, libcufft_static.a, libcufftw.so,
+libcufftw_static.a
+
+Android
+
+libcufft.so, libcufft_static.a, libcufftw.so,
+libcufftw_static.a
+
+Component
+
+CUDA BLAS Library
+
+Windows
+
+cublas.dll, cublasLt.dll
+
+Mac OSX
+
+libcublas.dylib, libcublasLt.dylib, libcublas_static.a,
+libcublasLt_static.a
+
+Linux
+
+libcublas.so, libcublasLt.so, libcublas_static.a,
+libcublasLt_static.a
+
+Android
+
+libcublas.so, libcublasLt.so, libcublas_static.a,
+libcublasLt_static.a
+
+Component
+
+NVIDIA "Drop-in" BLAS Library
+
+Windows
+
+nvblas.dll
+
+Mac OSX
+
+libnvblas.dylib
+
+Linux
+
+libnvblas.so
+
+Component
+
+CUDA Sparse Matrix Library
+
+Windows
+
+cusparse.dll, cusparse.lib
+
+Mac OSX
+
+libcusparse.dylib, libcusparse_static.a
+
+Linux
+
+libcusparse.so, libcusparse_static.a
+
+Android
+
+libcusparse.so, libcusparse_static.a
+
+Component
+
+CUDA Linear Solver Library
+
+Windows
+
+cusolver.dll, cusolver.lib
+
+Mac OSX
+
+libcusolver.dylib, libcusolver_static.a
+
+Linux
+
+libcusolver.so, libcusolver_static.a
+
+Android
+
+libcusolver.so, libcusolver_static.a
+
+Component
+
+CUDA Random Number Generation Library
+
+Windows
+
+curand.dll, curand.lib
+
+Mac OSX
+
+libcurand.dylib, libcurand_static.a
+
+Linux
+
+libcurand.so, libcurand_static.a
+
+Android
+
+libcurand.so, libcurand_static.a
+
+Component
+
+CUDA Accelerated Graph Library
+
+Component
+
+NVIDIA Performance Primitives Library
+
+Windows
+
+nppc.dll, nppc.lib, nppial.dll, nppial.lib, nppicc.dll,
+nppicc.lib, nppicom.dll, nppicom.lib, nppidei.dll,
+nppidei.lib, nppif.dll, nppif.lib, nppig.dll, nppig.lib,
+nppim.dll, nppim.lib, nppist.dll, nppist.lib, nppisu.dll,
+nppisu.lib, nppitc.dll, nppitc.lib, npps.dll, npps.lib
+
+Mac OSX
+
+libnppc.dylib, libnppc_static.a, libnppial.dylib,
+libnppial_static.a, libnppicc.dylib, libnppicc_static.a,
+libnppicom.dylib, libnppicom_static.a, libnppidei.dylib,
+libnppidei_static.a, libnppif.dylib, libnppif_static.a,
+libnppig.dylib, libnppig_static.a, libnppim.dylib,
+libnppisu_static.a, libnppitc.dylib, libnppitc_static.a,
+libnpps.dylib, libnpps_static.a
+
+Linux
+
+libnppc.so, libnppc_static.a, libnppial.so,
+libnppial_static.a, libnppicc.so, libnppicc_static.a,
+libnppicom.so, libnppicom_static.a, libnppidei.so,
+libnppidei_static.a, libnppif.so, libnppif_static.a
+libnppig.so, libnppig_static.a, libnppim.so,
+libnppim_static.a, libnppist.so, libnppist_static.a,
+libnppisu.so, libnppisu_static.a, libnppitc.so
+libnppitc_static.a, libnpps.so, libnpps_static.a
+
+Android
+
+libnppc.so, libnppc_static.a, libnppial.so,
+libnppial_static.a, libnppicc.so, libnppicc_static.a,
+libnppicom.so, libnppicom_static.a, libnppidei.so,
+libnppidei_static.a, libnppif.so, libnppif_static.a
+libnppig.so, libnppig_static.a, libnppim.so,
+libnppim_static.a, libnppist.so, libnppist_static.a,
+libnppisu.so, libnppisu_static.a, libnppitc.so
+libnppitc_static.a, libnpps.so, libnpps_static.a
+
+Component
+
+NVIDIA JPEG Library
+
+Linux
+
+libnvjpeg.so, libnvjpeg_static.a
+
+Component
+
+Internal common library required for statically linking to
+cuBLAS, cuSPARSE, cuFFT, cuRAND, nvJPEG and NPP
+
+Mac OSX
+
+libculibos.a
+
+Linux
+
+libculibos.a
+
+Component
+
+NVIDIA Runtime Compilation Library and Header
+
+All
+
+nvrtc.h
+
+Windows
+
+nvrtc.dll, nvrtc-builtins.dll
+
+Mac OSX
+
+libnvrtc.dylib, libnvrtc-builtins.dylib
+
+Linux
+
+libnvrtc.so, libnvrtc-builtins.so
+
+Component
+
+NVIDIA Optimizing Compiler Library
+
+Windows
+
+nvvm.dll
+
+Mac OSX
+
+libnvvm.dylib
+
+Linux
+
+libnvvm.so
+
+Component
+
+NVIDIA Common Device Math Functions Library
+
+Windows
+
+libdevice.10.bc
+
+Mac OSX
+
+libdevice.10.bc
+
+Linux
+
+libdevice.10.bc
+
+Component
+
+CUDA Occupancy Calculation Header Library
+
+All
+
+cuda_occupancy.h
+
+Component
+
+CUDA Half Precision Headers
+
+All
+
+cuda_fp16.h, cuda_fp16.hpp
+
+Component
+
+CUDA Profiling Tools Interface (CUPTI) Library
+
+Windows
+
+cupti.dll
+
+Mac OSX
+
+libcupti.dylib
+
+Linux
+
+libcupti.so
+
+Component
+
+NVIDIA Tools Extension Library
+
+Windows
+
+nvToolsExt.dll, nvToolsExt.lib
+
+Mac OSX
+
+libnvToolsExt.dylib
+
+Linux
+
+libnvToolsExt.so
+
+Component
+
+NVIDIA CUDA Driver Libraries
+
+Linux
+
+libcuda.so, libnvidia-fatbinaryloader.so,
+libnvidia-ptxjitcompiler.so
+
+The NVIDIA CUDA Driver Libraries are only distributable in
+applications that meet this criteria:
+
+  1. The application was developed starting from a NVIDIA CUDA
+    container obtained from Docker Hub or the NVIDIA GPU
+    Cloud, and
+
+  2. The resulting application is packaged as a Docker
+    container and distributed to users on Docker Hub or the
+    NVIDIA GPU Cloud only.
+
+
+2.7. Attachment B
+
+
+Additional Licensing Obligations
+
+The following third party components included in the SOFTWARE
+are licensed to Licensee pursuant to the following terms and
+conditions:
+
+  1. Licensee's use of the GDB third party component is
+    subject to the terms and conditions of GNU GPL v3:
+
+    This product includes copyrighted third-party software licensed
+    under the terms of the GNU General Public License v3 ("GPL v3").
+    All third-party software packages are copyright by their respective
+    authors. GPL v3 terms and conditions are hereby incorporated into
+    the Agreement by this reference:     http://www.gnu.org/licenses/gpl.txt
+
+    Consistent with these licensing requirements, the software
+    listed below is provided under the terms of the specified
+    open source software licenses. To obtain source code for
+    software provided under licenses that require
+    redistribution of source code, including the GNU General
+    Public License (GPL) and GNU Lesser General Public License
+    (LGPL), contact oss-requests@nvidia.com. This offer is
+    valid for a period of three (3) years from the date of the
+    distribution of this product by NVIDIA CORPORATION.
+
+    Component          License
+    CUDA-GDB           GPL v3
+
+  2. Licensee represents and warrants that any and all third
+    party licensing and/or royalty payment obligations in
+    connection with Licensee's use of the H.264 video codecs
+    are solely the responsibility of Licensee.
+
+  3. Licensee's use of the Thrust library is subject to the
+    terms and conditions of the Apache License Version 2.0.
+    All third-party software packages are copyright by their
+    respective authors. Apache License Version 2.0 terms and
+    conditions are hereby incorporated into the Agreement by
+    this reference.
+    http://www.apache.org/licenses/LICENSE-2.0.html
+
+    In addition, Licensee acknowledges the following notice:
+    Thrust includes source code from the Boost Iterator,
+    Tuple, System, and Random Number libraries.
+
+    Boost Software License - Version 1.0 - August 17th, 2003
+    . . . .
+
+    Permission is hereby granted, free of charge, to any person or
+    organization obtaining a copy of the software and accompanying
+    documentation covered by this license (the "Software") to use,
+    reproduce, display, distribute, execute, and transmit the Software,
+    and to prepare derivative works of the Software, and to permit
+    third-parties to whom the Software is furnished to do so, all
+    subject to the following:
+
+    The copyright notices in the Software and this entire statement,
+    including the above license grant, this restriction and the following
+    disclaimer, must be included in all copies of the Software, in whole
+    or in part, and all derivative works of the Software, unless such
+    copies or derivative works are solely in the form of machine-executable
+    object code generated by a source language processor.
+
+    THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+    EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+    MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE, TITLE AND
+    NON-INFRINGEMENT. IN NO EVENT SHALL THE COPYRIGHT HOLDERS OR
+    ANYONE DISTRIBUTING THE SOFTWARE BE LIABLE FOR ANY DAMAGES OR
+    OTHER LIABILITY, WHETHER IN CONTRACT, TORT OR OTHERWISE, ARISING
+    FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+    OTHER DEALINGS IN THE SOFTWARE.
+
+  4. Licensee's use of the LLVM third party component is
+    subject to the following terms and conditions:
+
+    ======================================================
+    LLVM Release License
+    ======================================================
+    University of Illinois/NCSA
+    Open Source License
+
+    Copyright (c) 2003-2010 University of Illinois at Urbana-Champaign.
+    All rights reserved.
+
+    Developed by:
+
+        LLVM Team
+
+        University of Illinois at Urbana-Champaign
+
+        http://llvm.org
+
+    Permission is hereby granted, free of charge, to any person obtaining a copy
+    of this software and associated documentation files (the "Software"), to
+    deal with the Software without restriction, including without limitation the
+    rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+    sell copies of the Software, and to permit persons to whom the Software is
+    furnished to do so, subject to the following conditions:
+
+    *  Redistributions of source code must retain the above copyright notice,
+       this list of conditions and the following disclaimers.
+
+    *  Redistributions in binary form must reproduce the above copyright
+       notice, this list of conditions and the following disclaimers in the
+       documentation and/or other materials provided with the distribution.
+
+    *  Neither the names of the LLVM Team, University of Illinois at Urbana-
+       Champaign, nor the names of its contributors may be used to endorse or
+       promote products derived from this Software without specific prior
+       written permission.
+
+    THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+    IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+    FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+    THE CONTRIBUTORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
+    OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+    ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+    DEALINGS WITH THE SOFTWARE.
+
+  5. Licensee's use (e.g. nvprof) of the PCRE third party
+    component is subject to the following terms and
+    conditions:
+
+    ------------
+    PCRE LICENCE
+    ------------
+    PCRE is a library of functions to support regular expressions whose syntax
+    and semantics are as close as possible to those of the Perl 5 language.
+    Release 8 of PCRE is distributed under the terms of the "BSD" licence, as
+    specified below. The documentation for PCRE, supplied in the "doc"
+    directory, is distributed under the same terms as the software itself. The
+    basic library functions are written in C and are freestanding. Also
+    included in the distribution is a set of C++ wrapper functions, and a just-
+    in-time compiler that can be used to optimize pattern matching. These are
+    both optional features that can be omitted when the library is built.
+
+    THE BASIC LIBRARY FUNCTIONS
+    ---------------------------
+    Written by:       Philip Hazel
+    Email local part: ph10
+    Email domain:     cam.ac.uk
+    University of Cambridge Computing Service,
+    Cambridge, England.
+    Copyright (c) 1997-2012 University of Cambridge
+    All rights reserved.
+
+    PCRE JUST-IN-TIME COMPILATION SUPPORT
+    -------------------------------------
+    Written by:       Zoltan Herczeg
+    Email local part: hzmester
+    Emain domain:     freemail.hu
+    Copyright(c) 2010-2012 Zoltan Herczeg
+    All rights reserved.
+
+    STACK-LESS JUST-IN-TIME COMPILER
+    --------------------------------
+    Written by:       Zoltan Herczeg
+    Email local part: hzmester
+    Emain domain:     freemail.hu
+    Copyright(c) 2009-2012 Zoltan Herczeg
+    All rights reserved.
+
+    THE C++ WRAPPER FUNCTIONS
+    -------------------------
+    Contributed by:   Google Inc.
+    Copyright (c) 2007-2012, Google Inc.
+    All rights reserved.
+
+    THE "BSD" LICENCE
+    -----------------
+    Redistribution and use in source and binary forms, with or without
+    modification, are permitted provided that the following conditions are met:
+
+      * Redistributions of source code must retain the above copyright notice,
+        this list of conditions and the following disclaimer.
+
+      * Redistributions in binary form must reproduce the above copyright
+        notice, this list of conditions and the following disclaimer in the
+        documentation and/or other materials provided with the distribution.
+
+      * Neither the name of the University of Cambridge nor the name of Google
+        Inc. nor the names of their contributors may be used to endorse or
+        promote products derived from this software without specific prior
+        written permission.
+
+    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+    AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+    IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+    ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
+    LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+    CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+    SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+    INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+    CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+    ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+    POSSIBILITY OF SUCH DAMAGE.
+
+  6. Some of the cuBLAS library routines were written by or
+    derived from code written by Vasily Volkov and are subject
+    to the Modified Berkeley Software Distribution License as
+    follows:
+
+    Copyright (c) 2007-2009, Regents of the University of California
+
+    All rights reserved.
+
+    Redistribution and use in source and binary forms, with or without
+    modification, are permitted provided that the following conditions are
+    met:
+        * Redistributions of source code must retain the above copyright
+          notice, this list of conditions and the following disclaimer.
+        * Redistributions in binary form must reproduce the above
+          copyright notice, this list of conditions and the following
+          disclaimer in the documentation and/or other materials provided
+          with the distribution.
+        * Neither the name of the University of California, Berkeley nor
+          the names of its contributors may be used to endorse or promote
+          products derived from this software without specific prior
+          written permission.
+
+    THIS SOFTWARE IS PROVIDED BY THE AUTHOR "AS IS" AND ANY EXPRESS OR
+    IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+    WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+    DISCLAIMED. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT,
+    INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+    (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+    SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
+    HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+    STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING
+    IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+    POSSIBILITY OF SUCH DAMAGE.
+
+  7. Some of the cuBLAS library routines were written by or
+    derived from code written by Davide Barbieri and are
+    subject to the Modified Berkeley Software Distribution
+    License as follows:
+
+    Copyright (c) 2008-2009 Davide Barbieri @ University of Rome Tor Vergata.
+
+    All rights reserved.
+
+    Redistribution and use in source and binary forms, with or without
+    modification, are permitted provided that the following conditions are
+    met:
+        * Redistributions of source code must retain the above copyright
+          notice, this list of conditions and the following disclaimer.
+        * Redistributions in binary form must reproduce the above
+          copyright notice, this list of conditions and the following
+          disclaimer in the documentation and/or other materials provided
+          with the distribution.
+        * The name of the author may not be used to endorse or promote
+          products derived from this software without specific prior
+          written permission.
+
+    THIS SOFTWARE IS PROVIDED BY THE AUTHOR "AS IS" AND ANY EXPRESS OR
+    IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+    WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+    DISCLAIMED. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT,
+    INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+    (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+    SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
+    HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+    STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING
+    IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+    POSSIBILITY OF SUCH DAMAGE.
+
+  8. Some of the cuBLAS library routines were derived from
+    code developed by the University of Tennessee and are
+    subject to the Modified Berkeley Software Distribution
+    License as follows:
+
+    Copyright (c) 2010 The University of Tennessee.
+
+    All rights reserved.
+
+    Redistribution and use in source and binary forms, with or without
+    modification, are permitted provided that the following conditions are
+    met:
+        * Redistributions of source code must retain the above copyright
+          notice, this list of conditions and the following disclaimer.
+        * Redistributions in binary form must reproduce the above
+          copyright notice, this list of conditions and the following
+          disclaimer listed in this license in the documentation and/or
+          other materials provided with the distribution.
+        * Neither the name of the copyright holders nor the names of its
+          contributors may be used to endorse or promote products derived
+          from this software without specific prior written permission.
+
+    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+    "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+    A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+    OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+    SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+    LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+    DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+    THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+    (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+    OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+  9. Some of the cuBLAS library routines were written by or
+    derived from code written by Jonathan Hogg and are subject
+    to the Modified Berkeley Software Distribution License as
+    follows:
+
+    Copyright (c) 2012, The Science and Technology Facilities Council (STFC).
+
+    All rights reserved.
+
+    Redistribution and use in source and binary forms, with or without
+    modification, are permitted provided that the following conditions are
+    met:
+        * Redistributions of source code must retain the above copyright
+          notice, this list of conditions and the following disclaimer.
+        * Redistributions in binary form must reproduce the above
+          copyright notice, this list of conditions and the following
+          disclaimer in the documentation and/or other materials provided
+          with the distribution.
+        * Neither the name of the STFC nor the names of its contributors
+          may be used to endorse or promote products derived from this
+          software without specific prior written permission.
+
+    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+    "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+    A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE STFC BE
+    LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+    CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+    SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR
+    BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
+    WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE
+    OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN
+    IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+  10. Some of the cuBLAS library routines were written by or
+    derived from code written by Ahmad M. Abdelfattah, David
+    Keyes, and Hatem Ltaief, and are subject to the Apache
+    License, Version 2.0, as follows:
+
+     -- (C) Copyright 2013 King Abdullah University of Science and Technology
+      Authors:
+      Ahmad Abdelfattah (ahmad.ahmad@kaust.edu.sa)
+      David Keyes (david.keyes@kaust.edu.sa)
+      Hatem Ltaief (hatem.ltaief@kaust.edu.sa)
+
+      Redistribution  and  use  in  source and binary forms, with or without
+      modification,  are  permitted  provided  that the following conditions
+      are met:
+
+      * Redistributions  of  source  code  must  retain  the above copyright
+        notice,  this  list  of  conditions  and  the  following  disclaimer.
+      * Redistributions  in  binary  form must reproduce the above copyright
+        notice,  this list of conditions and the following disclaimer in the
+        documentation  and/or other materials provided with the distribution.
+      * Neither  the  name of the King Abdullah University of Science and
+        Technology nor the names of its contributors may be used to endorse
+        or promote products derived from this software without specific prior
+        written permission.
+
+      THIS  SOFTWARE  IS  PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+      ``AS IS''  AND  ANY  EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+      LIMITED  TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+      A  PARTICULAR  PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+      HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+      SPECIAL,  EXEMPLARY,  OR  CONSEQUENTIAL  DAMAGES  (INCLUDING,  BUT NOT
+      LIMITED  TO,  PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+      DATA,  OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+      THEORY  OF  LIABILITY,  WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+      (INCLUDING  NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+      OF  THIS  SOFTWARE,  EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE
+
+  11. Some of the cuSPARSE library routines were written by or
+    derived from code written by Li-Wen Chang and are subject
+    to the NCSA Open Source License as follows:
+
+    Copyright (c) 2012, University of Illinois.
+
+    All rights reserved.
+
+    Developed by: IMPACT Group, University of Illinois, http://impact.crhc.illinois.edu
+
+    Permission is hereby granted, free of charge, to any person obtaining
+    a copy of this software and associated documentation files (the
+    "Software"), to deal with the Software without restriction, including
+    without limitation the rights to use, copy, modify, merge, publish,
+    distribute, sublicense, and/or sell copies of the Software, and to
+    permit persons to whom the Software is furnished to do so, subject to
+    the following conditions:
+        * Redistributions of source code must retain the above copyright
+          notice, this list of conditions and the following disclaimer.
+        * Redistributions in binary form must reproduce the above
+          copyright notice, this list of conditions and the following
+          disclaimers in the documentation and/or other materials provided
+          with the distribution.
+        * Neither the names of IMPACT Group, University of Illinois, nor
+          the names of its contributors may be used to endorse or promote
+          products derived from this Software without specific prior
+          written permission.
+
+    THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+    EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+    MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+    NONINFRINGEMENT. IN NO EVENT SHALL THE CONTRIBUTORS OR COPYRIGHT
+    HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER
+    IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR
+    IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS WITH THE
+    SOFTWARE.
+
+  12. Some of the cuRAND library routines were written by or
+    derived from code written by Mutsuo Saito and Makoto
+    Matsumoto and are subject to the following license:
+
+    Copyright (c) 2009, 2010 Mutsuo Saito, Makoto Matsumoto and Hiroshima
+    University. All rights reserved.
+
+    Copyright (c) 2011 Mutsuo Saito, Makoto Matsumoto, Hiroshima
+    University and University of Tokyo.  All rights reserved.
+
+    Redistribution and use in source and binary forms, with or without
+    modification, are permitted provided that the following conditions are
+    met:
+        * Redistributions of source code must retain the above copyright
+          notice, this list of conditions and the following disclaimer.
+        * Redistributions in binary form must reproduce the above
+          copyright notice, this list of conditions and the following
+          disclaimer in the documentation and/or other materials provided
+          with the distribution.
+        * Neither the name of the Hiroshima University nor the names of
+          its contributors may be used to endorse or promote products
+          derived from this software without specific prior written
+          permission.
+
+    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+    "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+    A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+    OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+    SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+    LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+    DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+    THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+    (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+    OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+  13. Some of the cuRAND library routines were derived from
+    code developed by D. E. Shaw Research and are subject to
+    the following license:
+
+    Copyright 2010-2011, D. E. Shaw Research.
+
+    All rights reserved.
+
+    Redistribution and use in source and binary forms, with or without
+    modification, are permitted provided that the following conditions are
+    met:
+        * Redistributions of source code must retain the above copyright
+          notice, this list of conditions, and the following disclaimer.
+        * Redistributions in binary form must reproduce the above
+          copyright notice, this list of conditions, and the following
+          disclaimer in the documentation and/or other materials provided
+          with the distribution.
+        * Neither the name of D. E. Shaw Research nor the names of its
+          contributors may be used to endorse or promote products derived
+          from this software without specific prior written permission.
+
+    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+    "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+    A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+    OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+    SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+    LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+    DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+    THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+    (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+    OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+  14. Some of the Math library routines were written by or
+    derived from code developed by Norbert Juffa and are
+    subject to the following license:
+
+    Copyright (c) 2015-2017, Norbert Juffa
+    All rights reserved.
+
+    Redistribution and use in source and binary forms, with or without
+    modification, are permitted provided that the following conditions
+    are met:
+
+    1. Redistributions of source code must retain the above copyright
+       notice, this list of conditions and the following disclaimer.
+
+    2. Redistributions in binary form must reproduce the above copyright
+       notice, this list of conditions and the following disclaimer in the
+       documentation and/or other materials provided with the distribution.
+
+    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+    "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+    A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+    HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+    SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+    LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+    DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+    THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+    (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+    OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+  15. Licensee's use of the lz4 third party component is
+    subject to the following terms and conditions:
+
+    Copyright (C) 2011-2013, Yann Collet.
+    BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+    Redistribution and use in source and binary forms, with or without
+    modification, are permitted provided that the following conditions are
+    met:
+
+        * Redistributions of source code must retain the above copyright
+    notice, this list of conditions and the following disclaimer.
+        * Redistributions in binary form must reproduce the above
+    copyright notice, this list of conditions and the following disclaimer
+    in the documentation and/or other materials provided with the
+    distribution.
+
+    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+    "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+    A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+    OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+    SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+    LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+    DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+    THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+    (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+    OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+  16. The NPP library uses code from the Boost Math Toolkit,
+    and is subject to the following license:
+
+    Boost Software License - Version 1.0 - August 17th, 2003
+    . . . .
+
+    Permission is hereby granted, free of charge, to any person or
+    organization obtaining a copy of the software and accompanying
+    documentation covered by this license (the "Software") to use,
+    reproduce, display, distribute, execute, and transmit the Software,
+    and to prepare derivative works of the Software, and to permit
+    third-parties to whom the Software is furnished to do so, all
+    subject to the following:
+
+    The copyright notices in the Software and this entire statement,
+    including the above license grant, this restriction and the following
+    disclaimer, must be included in all copies of the Software, in whole
+    or in part, and all derivative works of the Software, unless such
+    copies or derivative works are solely in the form of machine-executable
+    object code generated by a source language processor.
+
+    THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+    EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+    MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE, TITLE AND
+    NON-INFRINGEMENT. IN NO EVENT SHALL THE COPYRIGHT HOLDERS OR
+    ANYONE DISTRIBUTING THE SOFTWARE BE LIABLE FOR ANY DAMAGES OR
+    OTHER LIABILITY, WHETHER IN CONTRACT, TORT OR OTHERWISE, ARISING
+    FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+    OTHER DEALINGS IN THE SOFTWARE.
+
+  17. Portions of the Nsight Eclipse Edition is subject to the
+    following license:
+
+    The Eclipse Foundation makes available all content in this plug-in
+    ("Content"). Unless otherwise indicated below, the Content is provided
+    to you under the terms and conditions of the Eclipse Public License
+    Version 1.0 ("EPL"). A copy of the EPL is available at http://
+    www.eclipse.org/legal/epl-v10.html. For purposes of the EPL, "Program"
+    will mean the Content.
+
+    If you did not receive this Content directly from the Eclipse
+    Foundation, the Content is being redistributed by another party
+    ("Redistributor") and different terms and conditions may apply to your
+    use of any object code in the Content. Check the Redistributor's
+    license that was provided with the Content. If no such license exists,
+    contact the Redistributor. Unless otherwise indicated below, the terms
+    and conditions of the EPL still apply to any source code in the
+    Content and such source code may be obtained at http://www.eclipse.org.
+
+  18. Some of the cuBLAS library routines uses code from
+    OpenAI, which is subject to the following license:
+
+    License URL
+    https://github.com/openai/openai-gemm/blob/master/LICENSE
+
+    License Text
+    The MIT License
+
+    Copyright (c) 2016 OpenAI (http://openai.com), 2016 Google Inc.
+
+    Permission is hereby granted, free of charge, to any person obtaining a copy
+    of this software and associated documentation files (the "Software"), to deal
+    in the Software without restriction, including without limitation the rights
+    to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+    copies of the Software, and to permit persons to whom the Software is
+    furnished to do so, subject to the following conditions:
+
+    The above copyright notice and this permission notice shall be included in
+    all copies or substantial portions of the Software.
+
+    THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+    IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+    FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+    AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+    LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+    OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+    THE SOFTWARE.
+
+  19. Licensee's use of the Visual Studio Setup Configuration
+    Samples is subject to the following license:
+
+    The MIT License (MIT)
+    Copyright (C) Microsoft Corporation. All rights reserved.
+
+    Permission is hereby granted, free of charge, to any person
+    obtaining a copy of this software and associated documentation
+    files (the "Software"), to deal in the Software without restriction,
+    including without limitation the rights to use, copy, modify, merge,
+    publish, distribute, sublicense, and/or sell copies of the Software,
+    and to permit persons to whom the Software is furnished to do so,
+    subject to the following conditions:
+
+    The above copyright notice and this permission notice shall be included
+    in all copies or substantial portions of the Software.
+
+    THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
+    OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+    FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+    AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+    LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+    OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+  20. Licensee's use of linmath.h header for CPU functions for
+    GL vector/matrix operations from lunarG is subject to the
+    Apache License Version 2.0.
+
+  21. The DX12-CUDA sample uses the d3dx12.h header, which is
+    subject to the MIT license .
+
+-----------------

tuning-competition-baseline/.venv/lib/python3.11/site-packages/pip/_vendor/distlib/w64-arm.exe

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c5dc9884a8f458371550e09bd396e5418bf375820a31b9899f6499bf391c7b2e
+size 168448

tuning-competition-baseline/.venv/lib/python3.11/site-packages/pip/_vendor/distro/__main__.py

@@ -0,0 +1,4 @@
+from .distro import main
+
+if __name__ == "__main__":
+    main()

tuning-competition-baseline/.venv/lib/python3.11/site-packages/pip/_vendor/urllib3/contrib/_securetransport/bindings.py

@@ -0,0 +1,519 @@
+"""
+This module uses ctypes to bind a whole bunch of functions and constants from
+SecureTransport. The goal here is to provide the low-level API to
+SecureTransport. These are essentially the C-level functions and constants, and
+they're pretty gross to work with.
+
+This code is a bastardised version of the code found in Will Bond's oscrypto
+library. An enormous debt is owed to him for blazing this trail for us. For
+that reason, this code should be considered to be covered both by urllib3's
+license and by oscrypto's:
+
+    Copyright (c) 2015-2016 Will Bond <will@wbond.net>
+
+    Permission is hereby granted, free of charge, to any person obtaining a
+    copy of this software and associated documentation files (the "Software"),
+    to deal in the Software without restriction, including without limitation
+    the rights to use, copy, modify, merge, publish, distribute, sublicense,
+    and/or sell copies of the Software, and to permit persons to whom the
+    Software is furnished to do so, subject to the following conditions:
+
+    The above copyright notice and this permission notice shall be included in
+    all copies or substantial portions of the Software.
+
+    THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+    IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+    FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+    AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+    LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+    FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+    DEALINGS IN THE SOFTWARE.
+"""
+from __future__ import absolute_import
+
+import platform
+from ctypes import (
+    CDLL,
+    CFUNCTYPE,
+    POINTER,
+    c_bool,
+    c_byte,
+    c_char_p,
+    c_int32,
+    c_long,
+    c_size_t,
+    c_uint32,
+    c_ulong,
+    c_void_p,
+)
+from ctypes.util import find_library
+
+from ...packages.six import raise_from
+
+if platform.system() != "Darwin":
+    raise ImportError("Only macOS is supported")
+
+version = platform.mac_ver()[0]
+version_info = tuple(map(int, version.split(".")))
+if version_info < (10, 8):
+    raise OSError(
+        "Only OS X 10.8 and newer are supported, not %s.%s"
+        % (version_info[0], version_info[1])
+    )
+
+
+def load_cdll(name, macos10_16_path):
+    """Loads a CDLL by name, falling back to known path on 10.16+"""
+    try:
+        # Big Sur is technically 11 but we use 10.16 due to the Big Sur
+        # beta being labeled as 10.16.
+        if version_info >= (10, 16):
+            path = macos10_16_path
+        else:
+            path = find_library(name)
+        if not path:
+            raise OSError  # Caught and reraised as 'ImportError'
+        return CDLL(path, use_errno=True)
+    except OSError:
+        raise_from(ImportError("The library %s failed to load" % name), None)
+
+
+Security = load_cdll(
+    "Security", "/System/Library/Frameworks/Security.framework/Security"
+)
+CoreFoundation = load_cdll(
+    "CoreFoundation",
+    "/System/Library/Frameworks/CoreFoundation.framework/CoreFoundation",
+)
+
+
+Boolean = c_bool
+CFIndex = c_long
+CFStringEncoding = c_uint32
+CFData = c_void_p
+CFString = c_void_p
+CFArray = c_void_p
+CFMutableArray = c_void_p
+CFDictionary = c_void_p
+CFError = c_void_p
+CFType = c_void_p
+CFTypeID = c_ulong
+
+CFTypeRef = POINTER(CFType)
+CFAllocatorRef = c_void_p
+
+OSStatus = c_int32
+
+CFDataRef = POINTER(CFData)
+CFStringRef = POINTER(CFString)
+CFArrayRef = POINTER(CFArray)
+CFMutableArrayRef = POINTER(CFMutableArray)
+CFDictionaryRef = POINTER(CFDictionary)
+CFArrayCallBacks = c_void_p
+CFDictionaryKeyCallBacks = c_void_p
+CFDictionaryValueCallBacks = c_void_p
+
+SecCertificateRef = POINTER(c_void_p)
+SecExternalFormat = c_uint32
+SecExternalItemType = c_uint32
+SecIdentityRef = POINTER(c_void_p)
+SecItemImportExportFlags = c_uint32
+SecItemImportExportKeyParameters = c_void_p
+SecKeychainRef = POINTER(c_void_p)
+SSLProtocol = c_uint32
+SSLCipherSuite = c_uint32
+SSLContextRef = POINTER(c_void_p)
+SecTrustRef = POINTER(c_void_p)
+SSLConnectionRef = c_uint32
+SecTrustResultType = c_uint32
+SecTrustOptionFlags = c_uint32
+SSLProtocolSide = c_uint32
+SSLConnectionType = c_uint32
+SSLSessionOption = c_uint32
+
+
+try:
+    Security.SecItemImport.argtypes = [
+        CFDataRef,
+        CFStringRef,
+        POINTER(SecExternalFormat),
+        POINTER(SecExternalItemType),
+        SecItemImportExportFlags,
+        POINTER(SecItemImportExportKeyParameters),
+        SecKeychainRef,
+        POINTER(CFArrayRef),
+    ]
+    Security.SecItemImport.restype = OSStatus
+
+    Security.SecCertificateGetTypeID.argtypes = []
+    Security.SecCertificateGetTypeID.restype = CFTypeID
+
+    Security.SecIdentityGetTypeID.argtypes = []
+    Security.SecIdentityGetTypeID.restype = CFTypeID
+
+    Security.SecKeyGetTypeID.argtypes = []
+    Security.SecKeyGetTypeID.restype = CFTypeID
+
+    Security.SecCertificateCreateWithData.argtypes = [CFAllocatorRef, CFDataRef]
+    Security.SecCertificateCreateWithData.restype = SecCertificateRef
+
+    Security.SecCertificateCopyData.argtypes = [SecCertificateRef]
+    Security.SecCertificateCopyData.restype = CFDataRef
+
+    Security.SecCopyErrorMessageString.argtypes = [OSStatus, c_void_p]
+    Security.SecCopyErrorMessageString.restype = CFStringRef
+
+    Security.SecIdentityCreateWithCertificate.argtypes = [
+        CFTypeRef,
+        SecCertificateRef,
+        POINTER(SecIdentityRef),
+    ]
+    Security.SecIdentityCreateWithCertificate.restype = OSStatus
+
+    Security.SecKeychainCreate.argtypes = [
+        c_char_p,
+        c_uint32,
+        c_void_p,
+        Boolean,
+        c_void_p,
+        POINTER(SecKeychainRef),
+    ]
+    Security.SecKeychainCreate.restype = OSStatus
+
+    Security.SecKeychainDelete.argtypes = [SecKeychainRef]
+    Security.SecKeychainDelete.restype = OSStatus
+
+    Security.SecPKCS12Import.argtypes = [
+        CFDataRef,
+        CFDictionaryRef,
+        POINTER(CFArrayRef),
+    ]
+    Security.SecPKCS12Import.restype = OSStatus
+
+    SSLReadFunc = CFUNCTYPE(OSStatus, SSLConnectionRef, c_void_p, POINTER(c_size_t))
+    SSLWriteFunc = CFUNCTYPE(
+        OSStatus, SSLConnectionRef, POINTER(c_byte), POINTER(c_size_t)
+    )
+
+    Security.SSLSetIOFuncs.argtypes = [SSLContextRef, SSLReadFunc, SSLWriteFunc]
+    Security.SSLSetIOFuncs.restype = OSStatus
+
+    Security.SSLSetPeerID.argtypes = [SSLContextRef, c_char_p, c_size_t]
+    Security.SSLSetPeerID.restype = OSStatus
+
+    Security.SSLSetCertificate.argtypes = [SSLContextRef, CFArrayRef]
+    Security.SSLSetCertificate.restype = OSStatus
+
+    Security.SSLSetCertificateAuthorities.argtypes = [SSLContextRef, CFTypeRef, Boolean]
+    Security.SSLSetCertificateAuthorities.restype = OSStatus
+
+    Security.SSLSetConnection.argtypes = [SSLContextRef, SSLConnectionRef]
+    Security.SSLSetConnection.restype = OSStatus
+
+    Security.SSLSetPeerDomainName.argtypes = [SSLContextRef, c_char_p, c_size_t]
+    Security.SSLSetPeerDomainName.restype = OSStatus
+
+    Security.SSLHandshake.argtypes = [SSLContextRef]
+    Security.SSLHandshake.restype = OSStatus
+
+    Security.SSLRead.argtypes = [SSLContextRef, c_char_p, c_size_t, POINTER(c_size_t)]
+    Security.SSLRead.restype = OSStatus
+
+    Security.SSLWrite.argtypes = [SSLContextRef, c_char_p, c_size_t, POINTER(c_size_t)]
+    Security.SSLWrite.restype = OSStatus
+
+    Security.SSLClose.argtypes = [SSLContextRef]
+    Security.SSLClose.restype = OSStatus
+
+    Security.SSLGetNumberSupportedCiphers.argtypes = [SSLContextRef, POINTER(c_size_t)]
+    Security.SSLGetNumberSupportedCiphers.restype = OSStatus
+
+    Security.SSLGetSupportedCiphers.argtypes = [
+        SSLContextRef,
+        POINTER(SSLCipherSuite),
+        POINTER(c_size_t),
+    ]
+    Security.SSLGetSupportedCiphers.restype = OSStatus
+
+    Security.SSLSetEnabledCiphers.argtypes = [
+        SSLContextRef,
+        POINTER(SSLCipherSuite),
+        c_size_t,
+    ]
+    Security.SSLSetEnabledCiphers.restype = OSStatus
+
+    Security.SSLGetNumberEnabledCiphers.argtype = [SSLContextRef, POINTER(c_size_t)]
+    Security.SSLGetNumberEnabledCiphers.restype = OSStatus
+
+    Security.SSLGetEnabledCiphers.argtypes = [
+        SSLContextRef,
+        POINTER(SSLCipherSuite),
+        POINTER(c_size_t),
+    ]
+    Security.SSLGetEnabledCiphers.restype = OSStatus
+
+    Security.SSLGetNegotiatedCipher.argtypes = [SSLContextRef, POINTER(SSLCipherSuite)]
+    Security.SSLGetNegotiatedCipher.restype = OSStatus
+
+    Security.SSLGetNegotiatedProtocolVersion.argtypes = [
+        SSLContextRef,
+        POINTER(SSLProtocol),
+    ]
+    Security.SSLGetNegotiatedProtocolVersion.restype = OSStatus
+
+    Security.SSLCopyPeerTrust.argtypes = [SSLContextRef, POINTER(SecTrustRef)]
+    Security.SSLCopyPeerTrust.restype = OSStatus
+
+    Security.SecTrustSetAnchorCertificates.argtypes = [SecTrustRef, CFArrayRef]
+    Security.SecTrustSetAnchorCertificates.restype = OSStatus
+
+    Security.SecTrustSetAnchorCertificatesOnly.argstypes = [SecTrustRef, Boolean]
+    Security.SecTrustSetAnchorCertificatesOnly.restype = OSStatus
+
+    Security.SecTrustEvaluate.argtypes = [SecTrustRef, POINTER(SecTrustResultType)]
+    Security.SecTrustEvaluate.restype = OSStatus
+
+    Security.SecTrustGetCertificateCount.argtypes = [SecTrustRef]
+    Security.SecTrustGetCertificateCount.restype = CFIndex
+
+    Security.SecTrustGetCertificateAtIndex.argtypes = [SecTrustRef, CFIndex]
+    Security.SecTrustGetCertificateAtIndex.restype = SecCertificateRef
+
+    Security.SSLCreateContext.argtypes = [
+        CFAllocatorRef,
+        SSLProtocolSide,
+        SSLConnectionType,
+    ]
+    Security.SSLCreateContext.restype = SSLContextRef
+
+    Security.SSLSetSessionOption.argtypes = [SSLContextRef, SSLSessionOption, Boolean]
+    Security.SSLSetSessionOption.restype = OSStatus
+
+    Security.SSLSetProtocolVersionMin.argtypes = [SSLContextRef, SSLProtocol]
+    Security.SSLSetProtocolVersionMin.restype = OSStatus
+
+    Security.SSLSetProtocolVersionMax.argtypes = [SSLContextRef, SSLProtocol]
+    Security.SSLSetProtocolVersionMax.restype = OSStatus
+
+    try:
+        Security.SSLSetALPNProtocols.argtypes = [SSLContextRef, CFArrayRef]
+        Security.SSLSetALPNProtocols.restype = OSStatus
+    except AttributeError:
+        # Supported only in 10.12+
+        pass
+
+    Security.SecCopyErrorMessageString.argtypes = [OSStatus, c_void_p]
+    Security.SecCopyErrorMessageString.restype = CFStringRef
+
+    Security.SSLReadFunc = SSLReadFunc
+    Security.SSLWriteFunc = SSLWriteFunc
+    Security.SSLContextRef = SSLContextRef
+    Security.SSLProtocol = SSLProtocol
+    Security.SSLCipherSuite = SSLCipherSuite
+    Security.SecIdentityRef = SecIdentityRef
+    Security.SecKeychainRef = SecKeychainRef
+    Security.SecTrustRef = SecTrustRef
+    Security.SecTrustResultType = SecTrustResultType
+    Security.SecExternalFormat = SecExternalFormat
+    Security.OSStatus = OSStatus
+
+    Security.kSecImportExportPassphrase = CFStringRef.in_dll(
+        Security, "kSecImportExportPassphrase"
+    )
+    Security.kSecImportItemIdentity = CFStringRef.in_dll(
+        Security, "kSecImportItemIdentity"
+    )
+
+    # CoreFoundation time!
+    CoreFoundation.CFRetain.argtypes = [CFTypeRef]
+    CoreFoundation.CFRetain.restype = CFTypeRef
+
+    CoreFoundation.CFRelease.argtypes = [CFTypeRef]
+    CoreFoundation.CFRelease.restype = None
+
+    CoreFoundation.CFGetTypeID.argtypes = [CFTypeRef]
+    CoreFoundation.CFGetTypeID.restype = CFTypeID
+
+    CoreFoundation.CFStringCreateWithCString.argtypes = [
+        CFAllocatorRef,
+        c_char_p,
+        CFStringEncoding,
+    ]
+    CoreFoundation.CFStringCreateWithCString.restype = CFStringRef
+
+    CoreFoundation.CFStringGetCStringPtr.argtypes = [CFStringRef, CFStringEncoding]
+    CoreFoundation.CFStringGetCStringPtr.restype = c_char_p
+
+    CoreFoundation.CFStringGetCString.argtypes = [
+        CFStringRef,
+        c_char_p,
+        CFIndex,
+        CFStringEncoding,
+    ]
+    CoreFoundation.CFStringGetCString.restype = c_bool
+
+    CoreFoundation.CFDataCreate.argtypes = [CFAllocatorRef, c_char_p, CFIndex]
+    CoreFoundation.CFDataCreate.restype = CFDataRef
+
+    CoreFoundation.CFDataGetLength.argtypes = [CFDataRef]
+    CoreFoundation.CFDataGetLength.restype = CFIndex
+
+    CoreFoundation.CFDataGetBytePtr.argtypes = [CFDataRef]
+    CoreFoundation.CFDataGetBytePtr.restype = c_void_p
+
+    CoreFoundation.CFDictionaryCreate.argtypes = [
+        CFAllocatorRef,
+        POINTER(CFTypeRef),
+        POINTER(CFTypeRef),
+        CFIndex,
+        CFDictionaryKeyCallBacks,
+        CFDictionaryValueCallBacks,
+    ]
+    CoreFoundation.CFDictionaryCreate.restype = CFDictionaryRef
+
+    CoreFoundation.CFDictionaryGetValue.argtypes = [CFDictionaryRef, CFTypeRef]
+    CoreFoundation.CFDictionaryGetValue.restype = CFTypeRef
+
+    CoreFoundation.CFArrayCreate.argtypes = [
+        CFAllocatorRef,
+        POINTER(CFTypeRef),
+        CFIndex,
+        CFArrayCallBacks,
+    ]
+    CoreFoundation.CFArrayCreate.restype = CFArrayRef
+
+    CoreFoundation.CFArrayCreateMutable.argtypes = [
+        CFAllocatorRef,
+        CFIndex,
+        CFArrayCallBacks,
+    ]
+    CoreFoundation.CFArrayCreateMutable.restype = CFMutableArrayRef
+
+    CoreFoundation.CFArrayAppendValue.argtypes = [CFMutableArrayRef, c_void_p]
+    CoreFoundation.CFArrayAppendValue.restype = None
+
+    CoreFoundation.CFArrayGetCount.argtypes = [CFArrayRef]
+    CoreFoundation.CFArrayGetCount.restype = CFIndex
+
+    CoreFoundation.CFArrayGetValueAtIndex.argtypes = [CFArrayRef, CFIndex]
+    CoreFoundation.CFArrayGetValueAtIndex.restype = c_void_p
+
+    CoreFoundation.kCFAllocatorDefault = CFAllocatorRef.in_dll(
+        CoreFoundation, "kCFAllocatorDefault"
+    )
+    CoreFoundation.kCFTypeArrayCallBacks = c_void_p.in_dll(
+        CoreFoundation, "kCFTypeArrayCallBacks"
+    )
+    CoreFoundation.kCFTypeDictionaryKeyCallBacks = c_void_p.in_dll(
+        CoreFoundation, "kCFTypeDictionaryKeyCallBacks"
+    )
+    CoreFoundation.kCFTypeDictionaryValueCallBacks = c_void_p.in_dll(
+        CoreFoundation, "kCFTypeDictionaryValueCallBacks"
+    )
+
+    CoreFoundation.CFTypeRef = CFTypeRef
+    CoreFoundation.CFArrayRef = CFArrayRef
+    CoreFoundation.CFStringRef = CFStringRef
+    CoreFoundation.CFDictionaryRef = CFDictionaryRef
+
+except (AttributeError):
+    raise ImportError("Error initializing ctypes")
+
+
+class CFConst(object):
+    """
+    A class object that acts as essentially a namespace for CoreFoundation
+    constants.
+    """
+
+    kCFStringEncodingUTF8 = CFStringEncoding(0x08000100)
+
+
+class SecurityConst(object):
+    """
+    A class object that acts as essentially a namespace for Security constants.
+    """
+
+    kSSLSessionOptionBreakOnServerAuth = 0
+
+    kSSLProtocol2 = 1
+    kSSLProtocol3 = 2
+    kTLSProtocol1 = 4
+    kTLSProtocol11 = 7
+    kTLSProtocol12 = 8
+    # SecureTransport does not support TLS 1.3 even if there's a constant for it
+    kTLSProtocol13 = 10
+    kTLSProtocolMaxSupported = 999
+
+    kSSLClientSide = 1
+    kSSLStreamType = 0
+
+    kSecFormatPEMSequence = 10
+
+    kSecTrustResultInvalid = 0
+    kSecTrustResultProceed = 1
+    # This gap is present on purpose: this was kSecTrustResultConfirm, which
+    # is deprecated.
+    kSecTrustResultDeny = 3
+    kSecTrustResultUnspecified = 4
+    kSecTrustResultRecoverableTrustFailure = 5
+    kSecTrustResultFatalTrustFailure = 6
+    kSecTrustResultOtherError = 7
+
+    errSSLProtocol = -9800
+    errSSLWouldBlock = -9803
+    errSSLClosedGraceful = -9805
+    errSSLClosedNoNotify = -9816
+    errSSLClosedAbort = -9806
+
+    errSSLXCertChainInvalid = -9807
+    errSSLCrypto = -9809
+    errSSLInternal = -9810
+    errSSLCertExpired = -9814
+    errSSLCertNotYetValid = -9815
+    errSSLUnknownRootCert = -9812
+    errSSLNoRootCert = -9813
+    errSSLHostNameMismatch = -9843
+    errSSLPeerHandshakeFail = -9824
+    errSSLPeerUserCancelled = -9839
+    errSSLWeakPeerEphemeralDHKey = -9850
+    errSSLServerAuthCompleted = -9841
+    errSSLRecordOverflow = -9847
+
+    errSecVerifyFailed = -67808
+    errSecNoTrustSettings = -25263
+    errSecItemNotFound = -25300
+    errSecInvalidTrustSettings = -25262
+
+    # Cipher suites. We only pick the ones our default cipher string allows.
+    # Source: https://developer.apple.com/documentation/security/1550981-ssl_cipher_suite_values
+    TLS_ECDHE_ECDSA_WITH_AES_256_GCM_SHA384 = 0xC02C
+    TLS_ECDHE_RSA_WITH_AES_256_GCM_SHA384 = 0xC030
+    TLS_ECDHE_ECDSA_WITH_AES_128_GCM_SHA256 = 0xC02B
+    TLS_ECDHE_RSA_WITH_AES_128_GCM_SHA256 = 0xC02F
+    TLS_ECDHE_ECDSA_WITH_CHACHA20_POLY1305_SHA256 = 0xCCA9
+    TLS_ECDHE_RSA_WITH_CHACHA20_POLY1305_SHA256 = 0xCCA8
+    TLS_DHE_RSA_WITH_AES_256_GCM_SHA384 = 0x009F
+    TLS_DHE_RSA_WITH_AES_128_GCM_SHA256 = 0x009E
+    TLS_ECDHE_ECDSA_WITH_AES_256_CBC_SHA384 = 0xC024
+    TLS_ECDHE_RSA_WITH_AES_256_CBC_SHA384 = 0xC028
+    TLS_ECDHE_ECDSA_WITH_AES_256_CBC_SHA = 0xC00A
+    TLS_ECDHE_RSA_WITH_AES_256_CBC_SHA = 0xC014
+    TLS_DHE_RSA_WITH_AES_256_CBC_SHA256 = 0x006B
+    TLS_DHE_RSA_WITH_AES_256_CBC_SHA = 0x0039
+    TLS_ECDHE_ECDSA_WITH_AES_128_CBC_SHA256 = 0xC023
+    TLS_ECDHE_RSA_WITH_AES_128_CBC_SHA256 = 0xC027
+    TLS_ECDHE_ECDSA_WITH_AES_128_CBC_SHA = 0xC009
+    TLS_ECDHE_RSA_WITH_AES_128_CBC_SHA = 0xC013
+    TLS_DHE_RSA_WITH_AES_128_CBC_SHA256 = 0x0067
+    TLS_DHE_RSA_WITH_AES_128_CBC_SHA = 0x0033
+    TLS_RSA_WITH_AES_256_GCM_SHA384 = 0x009D
+    TLS_RSA_WITH_AES_128_GCM_SHA256 = 0x009C
+    TLS_RSA_WITH_AES_256_CBC_SHA256 = 0x003D
+    TLS_RSA_WITH_AES_128_CBC_SHA256 = 0x003C
+    TLS_RSA_WITH_AES_256_CBC_SHA = 0x0035
+    TLS_RSA_WITH_AES_128_CBC_SHA = 0x002F
+    TLS_AES_128_GCM_SHA256 = 0x1301
+    TLS_AES_256_GCM_SHA384 = 0x1302
+    TLS_AES_128_CCM_8_SHA256 = 0x1305
+    TLS_AES_128_CCM_SHA256 = 0x1304

tuning-competition-baseline/.venv/lib/python3.11/site-packages/pip/_vendor/urllib3/contrib/_securetransport/low_level.py

@@ -0,0 +1,397 @@
+"""
+Low-level helpers for the SecureTransport bindings.
+
+These are Python functions that are not directly related to the high-level APIs
+but are necessary to get them to work. They include a whole bunch of low-level
+CoreFoundation messing about and memory management. The concerns in this module
+are almost entirely about trying to avoid memory leaks and providing
+appropriate and useful assistance to the higher-level code.
+"""
+import base64
+import ctypes
+import itertools
+import os
+import re
+import ssl
+import struct
+import tempfile
+
+from .bindings import CFConst, CoreFoundation, Security
+
+# This regular expression is used to grab PEM data out of a PEM bundle.
+_PEM_CERTS_RE = re.compile(
+    b"-----BEGIN CERTIFICATE-----\n(.*?)\n-----END CERTIFICATE-----", re.DOTALL
+)
+
+
+def _cf_data_from_bytes(bytestring):
+    """
+    Given a bytestring, create a CFData object from it. This CFData object must
+    be CFReleased by the caller.
+    """
+    return CoreFoundation.CFDataCreate(
+        CoreFoundation.kCFAllocatorDefault, bytestring, len(bytestring)
+    )
+
+
+def _cf_dictionary_from_tuples(tuples):
+    """
+    Given a list of Python tuples, create an associated CFDictionary.
+    """
+    dictionary_size = len(tuples)
+
+    # We need to get the dictionary keys and values out in the same order.
+    keys = (t[0] for t in tuples)
+    values = (t[1] for t in tuples)
+    cf_keys = (CoreFoundation.CFTypeRef * dictionary_size)(*keys)
+    cf_values = (CoreFoundation.CFTypeRef * dictionary_size)(*values)
+
+    return CoreFoundation.CFDictionaryCreate(
+        CoreFoundation.kCFAllocatorDefault,
+        cf_keys,
+        cf_values,
+        dictionary_size,
+        CoreFoundation.kCFTypeDictionaryKeyCallBacks,
+        CoreFoundation.kCFTypeDictionaryValueCallBacks,
+    )
+
+
+def _cfstr(py_bstr):
+    """
+    Given a Python binary data, create a CFString.
+    The string must be CFReleased by the caller.
+    """
+    c_str = ctypes.c_char_p(py_bstr)
+    cf_str = CoreFoundation.CFStringCreateWithCString(
+        CoreFoundation.kCFAllocatorDefault,
+        c_str,
+        CFConst.kCFStringEncodingUTF8,
+    )
+    return cf_str
+
+
+def _create_cfstring_array(lst):
+    """
+    Given a list of Python binary data, create an associated CFMutableArray.
+    The array must be CFReleased by the caller.
+
+    Raises an ssl.SSLError on failure.
+    """
+    cf_arr = None
+    try:
+        cf_arr = CoreFoundation.CFArrayCreateMutable(
+            CoreFoundation.kCFAllocatorDefault,
+            0,
+            ctypes.byref(CoreFoundation.kCFTypeArrayCallBacks),
+        )
+        if not cf_arr:
+            raise MemoryError("Unable to allocate memory!")
+        for item in lst:
+            cf_str = _cfstr(item)
+            if not cf_str:
+                raise MemoryError("Unable to allocate memory!")
+            try:
+                CoreFoundation.CFArrayAppendValue(cf_arr, cf_str)
+            finally:
+                CoreFoundation.CFRelease(cf_str)
+    except BaseException as e:
+        if cf_arr:
+            CoreFoundation.CFRelease(cf_arr)
+        raise ssl.SSLError("Unable to allocate array: %s" % (e,))
+    return cf_arr
+
+
+def _cf_string_to_unicode(value):
+    """
+    Creates a Unicode string from a CFString object. Used entirely for error
+    reporting.
+
+    Yes, it annoys me quite a lot that this function is this complex.
+    """
+    value_as_void_p = ctypes.cast(value, ctypes.POINTER(ctypes.c_void_p))
+
+    string = CoreFoundation.CFStringGetCStringPtr(
+        value_as_void_p, CFConst.kCFStringEncodingUTF8
+    )
+    if string is None:
+        buffer = ctypes.create_string_buffer(1024)
+        result = CoreFoundation.CFStringGetCString(
+            value_as_void_p, buffer, 1024, CFConst.kCFStringEncodingUTF8
+        )
+        if not result:
+            raise OSError("Error copying C string from CFStringRef")
+        string = buffer.value
+    if string is not None:
+        string = string.decode("utf-8")
+    return string
+
+
+def _assert_no_error(error, exception_class=None):
+    """
+    Checks the return code and throws an exception if there is an error to
+    report
+    """
+    if error == 0:
+        return
+
+    cf_error_string = Security.SecCopyErrorMessageString(error, None)
+    output = _cf_string_to_unicode(cf_error_string)
+    CoreFoundation.CFRelease(cf_error_string)
+
+    if output is None or output == u"":
+        output = u"OSStatus %s" % error
+
+    if exception_class is None:
+        exception_class = ssl.SSLError
+
+    raise exception_class(output)
+
+
+def _cert_array_from_pem(pem_bundle):
+    """
+    Given a bundle of certs in PEM format, turns them into a CFArray of certs
+    that can be used to validate a cert chain.
+    """
+    # Normalize the PEM bundle's line endings.
+    pem_bundle = pem_bundle.replace(b"\r\n", b"\n")
+
+    der_certs = [
+        base64.b64decode(match.group(1)) for match in _PEM_CERTS_RE.finditer(pem_bundle)
+    ]
+    if not der_certs:
+        raise ssl.SSLError("No root certificates specified")
+
+    cert_array = CoreFoundation.CFArrayCreateMutable(
+        CoreFoundation.kCFAllocatorDefault,
+        0,
+        ctypes.byref(CoreFoundation.kCFTypeArrayCallBacks),
+    )
+    if not cert_array:
+        raise ssl.SSLError("Unable to allocate memory!")
+
+    try:
+        for der_bytes in der_certs:
+            certdata = _cf_data_from_bytes(der_bytes)
+            if not certdata:
+                raise ssl.SSLError("Unable to allocate memory!")
+            cert = Security.SecCertificateCreateWithData(
+                CoreFoundation.kCFAllocatorDefault, certdata
+            )
+            CoreFoundation.CFRelease(certdata)
+            if not cert:
+                raise ssl.SSLError("Unable to build cert object!")
+
+            CoreFoundation.CFArrayAppendValue(cert_array, cert)
+            CoreFoundation.CFRelease(cert)
+    except Exception:
+        # We need to free the array before the exception bubbles further.
+        # We only want to do that if an error occurs: otherwise, the caller
+        # should free.
+        CoreFoundation.CFRelease(cert_array)
+        raise
+
+    return cert_array
+
+
+def _is_cert(item):
+    """
+    Returns True if a given CFTypeRef is a certificate.
+    """
+    expected = Security.SecCertificateGetTypeID()
+    return CoreFoundation.CFGetTypeID(item) == expected
+
+
+def _is_identity(item):
+    """
+    Returns True if a given CFTypeRef is an identity.
+    """
+    expected = Security.SecIdentityGetTypeID()
+    return CoreFoundation.CFGetTypeID(item) == expected
+
+
+def _temporary_keychain():
+    """
+    This function creates a temporary Mac keychain that we can use to work with
+    credentials. This keychain uses a one-time password and a temporary file to
+    store the data. We expect to have one keychain per socket. The returned
+    SecKeychainRef must be freed by the caller, including calling
+    SecKeychainDelete.
+
+    Returns a tuple of the SecKeychainRef and the path to the temporary
+    directory that contains it.
+    """
+    # Unfortunately, SecKeychainCreate requires a path to a keychain. This
+    # means we cannot use mkstemp to use a generic temporary file. Instead,
+    # we're going to create a temporary directory and a filename to use there.
+    # This filename will be 8 random bytes expanded into base64. We also need
+    # some random bytes to password-protect the keychain we're creating, so we
+    # ask for 40 random bytes.
+    random_bytes = os.urandom(40)
+    filename = base64.b16encode(random_bytes[:8]).decode("utf-8")
+    password = base64.b16encode(random_bytes[8:])  # Must be valid UTF-8
+    tempdirectory = tempfile.mkdtemp()
+
+    keychain_path = os.path.join(tempdirectory, filename).encode("utf-8")
+
+    # We now want to create the keychain itself.
+    keychain = Security.SecKeychainRef()
+    status = Security.SecKeychainCreate(
+        keychain_path, len(password), password, False, None, ctypes.byref(keychain)
+    )
+    _assert_no_error(status)
+
+    # Having created the keychain, we want to pass it off to the caller.
+    return keychain, tempdirectory
+
+
+def _load_items_from_file(keychain, path):
+    """
+    Given a single file, loads all the trust objects from it into arrays and
+    the keychain.
+    Returns a tuple of lists: the first list is a list of identities, the
+    second a list of certs.
+    """
+    certificates = []
+    identities = []
+    result_array = None
+
+    with open(path, "rb") as f:
+        raw_filedata = f.read()
+
+    try:
+        filedata = CoreFoundation.CFDataCreate(
+            CoreFoundation.kCFAllocatorDefault, raw_filedata, len(raw_filedata)
+        )
+        result_array = CoreFoundation.CFArrayRef()
+        result = Security.SecItemImport(
+            filedata,  # cert data
+            None,  # Filename, leaving it out for now
+            None,  # What the type of the file is, we don't care
+            None,  # what's in the file, we don't care
+            0,  # import flags
+            None,  # key params, can include passphrase in the future
+            keychain,  # The keychain to insert into
+            ctypes.byref(result_array),  # Results
+        )
+        _assert_no_error(result)
+
+        # A CFArray is not very useful to us as an intermediary
+        # representation, so we are going to extract the objects we want
+        # and then free the array. We don't need to keep hold of keys: the
+        # keychain already has them!
+        result_count = CoreFoundation.CFArrayGetCount(result_array)
+        for index in range(result_count):
+            item = CoreFoundation.CFArrayGetValueAtIndex(result_array, index)
+            item = ctypes.cast(item, CoreFoundation.CFTypeRef)
+
+            if _is_cert(item):
+                CoreFoundation.CFRetain(item)
+                certificates.append(item)
+            elif _is_identity(item):
+                CoreFoundation.CFRetain(item)
+                identities.append(item)
+    finally:
+        if result_array:
+            CoreFoundation.CFRelease(result_array)
+
+        CoreFoundation.CFRelease(filedata)
+
+    return (identities, certificates)
+
+
+def _load_client_cert_chain(keychain, *paths):
+    """
+    Load certificates and maybe keys from a number of files. Has the end goal
+    of returning a CFArray containing one SecIdentityRef, and then zero or more
+    SecCertificateRef objects, suitable for use as a client certificate trust
+    chain.
+    """
+    # Ok, the strategy.
+    #
+    # This relies on knowing that macOS will not give you a SecIdentityRef
+    # unless you have imported a key into a keychain. This is a somewhat
+    # artificial limitation of macOS (for example, it doesn't necessarily
+    # affect iOS), but there is nothing inside Security.framework that lets you
+    # get a SecIdentityRef without having a key in a keychain.
+    #
+    # So the policy here is we take all the files and iterate them in order.
+    # Each one will use SecItemImport to have one or more objects loaded from
+    # it. We will also point at a keychain that macOS can use to work with the
+    # private key.
+    #
+    # Once we have all the objects, we'll check what we actually have. If we
+    # already have a SecIdentityRef in hand, fab: we'll use that. Otherwise,
+    # we'll take the first certificate (which we assume to be our leaf) and
+    # ask the keychain to give us a SecIdentityRef with that cert's associated
+    # key.
+    #
+    # We'll then return a CFArray containing the trust chain: one
+    # SecIdentityRef and then zero-or-more SecCertificateRef objects. The
+    # responsibility for freeing this CFArray will be with the caller. This
+    # CFArray must remain alive for the entire connection, so in practice it
+    # will be stored with a single SSLSocket, along with the reference to the
+    # keychain.
+    certificates = []
+    identities = []
+
+    # Filter out bad paths.
+    paths = (path for path in paths if path)
+
+    try:
+        for file_path in paths:
+            new_identities, new_certs = _load_items_from_file(keychain, file_path)
+            identities.extend(new_identities)
+            certificates.extend(new_certs)
+
+        # Ok, we have everything. The question is: do we have an identity? If
+        # not, we want to grab one from the first cert we have.
+        if not identities:
+            new_identity = Security.SecIdentityRef()
+            status = Security.SecIdentityCreateWithCertificate(
+                keychain, certificates[0], ctypes.byref(new_identity)
+            )
+            _assert_no_error(status)
+            identities.append(new_identity)
+
+            # We now want to release the original certificate, as we no longer
+            # need it.
+            CoreFoundation.CFRelease(certificates.pop(0))
+
+        # We now need to build a new CFArray that holds the trust chain.
+        trust_chain = CoreFoundation.CFArrayCreateMutable(
+            CoreFoundation.kCFAllocatorDefault,
+            0,
+            ctypes.byref(CoreFoundation.kCFTypeArrayCallBacks),
+        )
+        for item in itertools.chain(identities, certificates):
+            # ArrayAppendValue does a CFRetain on the item. That's fine,
+            # because the finally block will release our other refs to them.
+            CoreFoundation.CFArrayAppendValue(trust_chain, item)
+
+        return trust_chain
+    finally:
+        for obj in itertools.chain(identities, certificates):
+            CoreFoundation.CFRelease(obj)
+
+
+TLS_PROTOCOL_VERSIONS = {
+    "SSLv2": (0, 2),
+    "SSLv3": (3, 0),
+    "TLSv1": (3, 1),
+    "TLSv1.1": (3, 2),
+    "TLSv1.2": (3, 3),
+}
+
+
+def _build_tls_unknown_ca_alert(version):
+    """
+    Builds a TLS alert record for an unknown CA.
+    """
+    ver_maj, ver_min = TLS_PROTOCOL_VERSIONS[version]
+    severity_fatal = 0x02
+    description_unknown_ca = 0x30
+    msg = struct.pack(">BB", severity_fatal, description_unknown_ca)
+    msg_len = len(msg)
+    record_type_alert = 0x15
+    record = struct.pack(">BBBH", record_type_alert, ver_maj, ver_min, msg_len) + msg
+    return record

tuning-competition-baseline/.venv/lib/python3.11/site-packages/pip/_vendor/urllib3/contrib/appengine.py

@@ -0,0 +1,314 @@
+"""
+This module provides a pool manager that uses Google App Engine's
+`URLFetch Service <https://cloud.google.com/appengine/docs/python/urlfetch>`_.
+
+Example usage::
+
+    from pip._vendor.urllib3 import PoolManager
+    from pip._vendor.urllib3.contrib.appengine import AppEngineManager, is_appengine_sandbox
+
+    if is_appengine_sandbox():
+        # AppEngineManager uses AppEngine's URLFetch API behind the scenes
+        http = AppEngineManager()
+    else:
+        # PoolManager uses a socket-level API behind the scenes
+        http = PoolManager()
+
+    r = http.request('GET', 'https://google.com/')
+
+There are `limitations <https://cloud.google.com/appengine/docs/python/\
+urlfetch/#Python_Quotas_and_limits>`_ to the URLFetch service and it may not be
+the best choice for your application. There are three options for using
+urllib3 on Google App Engine:
+
+1. You can use :class:`AppEngineManager` with URLFetch. URLFetch is
+   cost-effective in many circumstances as long as your usage is within the
+   limitations.
+2. You can use a normal :class:`~urllib3.PoolManager` by enabling sockets.
+   Sockets also have `limitations and restrictions
+   <https://cloud.google.com/appengine/docs/python/sockets/\
+   #limitations-and-restrictions>`_ and have a lower free quota than URLFetch.
+   To use sockets, be sure to specify the following in your ``app.yaml``::
+
+        env_variables:
+            GAE_USE_SOCKETS_HTTPLIB : 'true'
+
+3. If you are using `App Engine Flexible
+<https://cloud.google.com/appengine/docs/flexible/>`_, you can use the standard
+:class:`PoolManager` without any configuration or special environment variables.
+"""
+
+from __future__ import absolute_import
+
+import io
+import logging
+import warnings
+
+from ..exceptions import (
+    HTTPError,
+    HTTPWarning,
+    MaxRetryError,
+    ProtocolError,
+    SSLError,
+    TimeoutError,
+)
+from ..packages.six.moves.urllib.parse import urljoin
+from ..request import RequestMethods
+from ..response import HTTPResponse
+from ..util.retry import Retry
+from ..util.timeout import Timeout
+from . import _appengine_environ
+
+try:
+    from google.appengine.api import urlfetch
+except ImportError:
+    urlfetch = None
+
+
+log = logging.getLogger(__name__)
+
+
+class AppEnginePlatformWarning(HTTPWarning):
+    pass
+
+
+class AppEnginePlatformError(HTTPError):
+    pass
+
+
+class AppEngineManager(RequestMethods):
+    """
+    Connection manager for Google App Engine sandbox applications.
+
+    This manager uses the URLFetch service directly instead of using the
+    emulated httplib, and is subject to URLFetch limitations as described in
+    the App Engine documentation `here
+    <https://cloud.google.com/appengine/docs/python/urlfetch>`_.
+
+    Notably it will raise an :class:`AppEnginePlatformError` if:
+        * URLFetch is not available.
+        * If you attempt to use this on App Engine Flexible, as full socket
+          support is available.
+        * If a request size is more than 10 megabytes.
+        * If a response size is more than 32 megabytes.
+        * If you use an unsupported request method such as OPTIONS.
+
+    Beyond those cases, it will raise normal urllib3 errors.
+    """
+
+    def __init__(
+        self,
+        headers=None,
+        retries=None,
+        validate_certificate=True,
+        urlfetch_retries=True,
+    ):
+        if not urlfetch:
+            raise AppEnginePlatformError(
+                "URLFetch is not available in this environment."
+            )
+
+        warnings.warn(
+            "urllib3 is using URLFetch on Google App Engine sandbox instead "
+            "of sockets. To use sockets directly instead of URLFetch see "
+            "https://urllib3.readthedocs.io/en/1.26.x/reference/urllib3.contrib.html.",
+            AppEnginePlatformWarning,
+        )
+
+        RequestMethods.__init__(self, headers)
+        self.validate_certificate = validate_certificate
+        self.urlfetch_retries = urlfetch_retries
+
+        self.retries = retries or Retry.DEFAULT
+
+    def __enter__(self):
+        return self
+
+    def __exit__(self, exc_type, exc_val, exc_tb):
+        # Return False to re-raise any potential exceptions
+        return False
+
+    def urlopen(
+        self,
+        method,
+        url,
+        body=None,
+        headers=None,
+        retries=None,
+        redirect=True,
+        timeout=Timeout.DEFAULT_TIMEOUT,
+        **response_kw
+    ):
+
+        retries = self._get_retries(retries, redirect)
+
+        try:
+            follow_redirects = redirect and retries.redirect != 0 and retries.total
+            response = urlfetch.fetch(
+                url,
+                payload=body,
+                method=method,
+                headers=headers or {},
+                allow_truncated=False,
+                follow_redirects=self.urlfetch_retries and follow_redirects,
+                deadline=self._get_absolute_timeout(timeout),
+                validate_certificate=self.validate_certificate,
+            )
+        except urlfetch.DeadlineExceededError as e:
+            raise TimeoutError(self, e)
+
+        except urlfetch.InvalidURLError as e:
+            if "too large" in str(e):
+                raise AppEnginePlatformError(
+                    "URLFetch request too large, URLFetch only "
+                    "supports requests up to 10mb in size.",
+                    e,
+                )
+            raise ProtocolError(e)
+
+        except urlfetch.DownloadError as e:
+            if "Too many redirects" in str(e):
+                raise MaxRetryError(self, url, reason=e)
+            raise ProtocolError(e)
+
+        except urlfetch.ResponseTooLargeError as e:
+            raise AppEnginePlatformError(
+                "URLFetch response too large, URLFetch only supports"
+                "responses up to 32mb in size.",
+                e,
+            )
+
+        except urlfetch.SSLCertificateError as e:
+            raise SSLError(e)
+
+        except urlfetch.InvalidMethodError as e:
+            raise AppEnginePlatformError(
+                "URLFetch does not support method: %s" % method, e
+            )
+
+        http_response = self._urlfetch_response_to_http_response(
+            response, retries=retries, **response_kw
+        )
+
+        # Handle redirect?
+        redirect_location = redirect and http_response.get_redirect_location()
+        if redirect_location:
+            # Check for redirect response
+            if self.urlfetch_retries and retries.raise_on_redirect:
+                raise MaxRetryError(self, url, "too many redirects")
+            else:
+                if http_response.status == 303:
+                    method = "GET"
+
+                try:
+                    retries = retries.increment(
+                        method, url, response=http_response, _pool=self
+                    )
+                except MaxRetryError:
+                    if retries.raise_on_redirect:
+                        raise MaxRetryError(self, url, "too many redirects")
+                    return http_response
+
+                retries.sleep_for_retry(http_response)
+                log.debug("Redirecting %s -> %s", url, redirect_location)
+                redirect_url = urljoin(url, redirect_location)
+                return self.urlopen(
+                    method,
+                    redirect_url,
+                    body,
+                    headers,
+                    retries=retries,
+                    redirect=redirect,
+                    timeout=timeout,
+                    **response_kw
+                )
+
+        # Check if we should retry the HTTP response.
+        has_retry_after = bool(http_response.headers.get("Retry-After"))
+        if retries.is_retry(method, http_response.status, has_retry_after):
+            retries = retries.increment(method, url, response=http_response, _pool=self)
+            log.debug("Retry: %s", url)
+            retries.sleep(http_response)
+            return self.urlopen(
+                method,
+                url,
+                body=body,
+                headers=headers,
+                retries=retries,
+                redirect=redirect,
+                timeout=timeout,
+                **response_kw
+            )
+
+        return http_response
+
+    def _urlfetch_response_to_http_response(self, urlfetch_resp, **response_kw):
+
+        if is_prod_appengine():
+            # Production GAE handles deflate encoding automatically, but does
+            # not remove the encoding header.
+            content_encoding = urlfetch_resp.headers.get("content-encoding")
+
+            if content_encoding == "deflate":
+                del urlfetch_resp.headers["content-encoding"]
+
+        transfer_encoding = urlfetch_resp.headers.get("transfer-encoding")
+        # We have a full response's content,
+        # so let's make sure we don't report ourselves as chunked data.
+        if transfer_encoding == "chunked":
+            encodings = transfer_encoding.split(",")
+            encodings.remove("chunked")
+            urlfetch_resp.headers["transfer-encoding"] = ",".join(encodings)
+
+        original_response = HTTPResponse(
+            # In order for decoding to work, we must present the content as
+            # a file-like object.
+            body=io.BytesIO(urlfetch_resp.content),
+            msg=urlfetch_resp.header_msg,
+            headers=urlfetch_resp.headers,
+            status=urlfetch_resp.status_code,
+            **response_kw
+        )
+
+        return HTTPResponse(
+            body=io.BytesIO(urlfetch_resp.content),
+            headers=urlfetch_resp.headers,
+            status=urlfetch_resp.status_code,
+            original_response=original_response,
+            **response_kw
+        )
+
+    def _get_absolute_timeout(self, timeout):
+        if timeout is Timeout.DEFAULT_TIMEOUT:
+            return None  # Defer to URLFetch's default.
+        if isinstance(timeout, Timeout):
+            if timeout._read is not None or timeout._connect is not None:
+                warnings.warn(
+                    "URLFetch does not support granular timeout settings, "
+                    "reverting to total or default URLFetch timeout.",
+                    AppEnginePlatformWarning,
+                )
+            return timeout.total
+        return timeout
+
+    def _get_retries(self, retries, redirect):
+        if not isinstance(retries, Retry):
+            retries = Retry.from_int(retries, redirect=redirect, default=self.retries)
+
+        if retries.connect or retries.read or retries.redirect:
+            warnings.warn(
+                "URLFetch only supports total retries and does not "
+                "recognize connect, read, or redirect retry parameters.",
+                AppEnginePlatformWarning,
+            )
+
+        return retries
+
+
+# Alias methods from _appengine_environ to maintain public API interface.
+
+is_appengine = _appengine_environ.is_appengine
+is_appengine_sandbox = _appengine_environ.is_appengine_sandbox
+is_local_appengine = _appengine_environ.is_local_appengine
+is_prod_appengine = _appengine_environ.is_prod_appengine
+is_prod_appengine_mvms = _appengine_environ.is_prod_appengine_mvms

tuning-competition-baseline/.venv/lib/python3.11/site-packages/pip/_vendor/urllib3/util/request.py

@@ -0,0 +1,137 @@
+from __future__ import absolute_import
+
+from base64 import b64encode
+
+from ..exceptions import UnrewindableBodyError
+from ..packages.six import b, integer_types
+
+# Pass as a value within ``headers`` to skip
+# emitting some HTTP headers that are added automatically.
+# The only headers that are supported are ``Accept-Encoding``,
+# ``Host``, and ``User-Agent``.
+SKIP_HEADER = "@@@SKIP_HEADER@@@"
+SKIPPABLE_HEADERS = frozenset(["accept-encoding", "host", "user-agent"])
+
+ACCEPT_ENCODING = "gzip,deflate"
+
+_FAILEDTELL = object()
+
+
+def make_headers(
+    keep_alive=None,
+    accept_encoding=None,
+    user_agent=None,
+    basic_auth=None,
+    proxy_basic_auth=None,
+    disable_cache=None,
+):
+    """
+    Shortcuts for generating request headers.
+
+    :param keep_alive:
+        If ``True``, adds 'connection: keep-alive' header.
+
+    :param accept_encoding:
+        Can be a boolean, list, or string.
+        ``True`` translates to 'gzip,deflate'.
+        List will get joined by comma.
+        String will be used as provided.
+
+    :param user_agent:
+        String representing the user-agent you want, such as
+        "python-urllib3/0.6"
+
+    :param basic_auth:
+        Colon-separated username:password string for 'authorization: basic ...'
+        auth header.
+
+    :param proxy_basic_auth:
+        Colon-separated username:password string for 'proxy-authorization: basic ...'
+        auth header.
+
+    :param disable_cache:
+        If ``True``, adds 'cache-control: no-cache' header.
+
+    Example::
+
+        >>> make_headers(keep_alive=True, user_agent="Batman/1.0")
+        {'connection': 'keep-alive', 'user-agent': 'Batman/1.0'}
+        >>> make_headers(accept_encoding=True)
+        {'accept-encoding': 'gzip,deflate'}
+    """
+    headers = {}
+    if accept_encoding:
+        if isinstance(accept_encoding, str):
+            pass
+        elif isinstance(accept_encoding, list):
+            accept_encoding = ",".join(accept_encoding)
+        else:
+            accept_encoding = ACCEPT_ENCODING
+        headers["accept-encoding"] = accept_encoding
+
+    if user_agent:
+        headers["user-agent"] = user_agent
+
+    if keep_alive:
+        headers["connection"] = "keep-alive"
+
+    if basic_auth:
+        headers["authorization"] = "Basic " + b64encode(b(basic_auth)).decode("utf-8")
+
+    if proxy_basic_auth:
+        headers["proxy-authorization"] = "Basic " + b64encode(
+            b(proxy_basic_auth)
+        ).decode("utf-8")
+
+    if disable_cache:
+        headers["cache-control"] = "no-cache"
+
+    return headers
+
+
+def set_file_position(body, pos):
+    """
+    If a position is provided, move file to that point.
+    Otherwise, we'll attempt to record a position for future use.
+    """
+    if pos is not None:
+        rewind_body(body, pos)
+    elif getattr(body, "tell", None) is not None:
+        try:
+            pos = body.tell()
+        except (IOError, OSError):
+            # This differentiates from None, allowing us to catch
+            # a failed `tell()` later when trying to rewind the body.
+            pos = _FAILEDTELL
+
+    return pos
+
+
+def rewind_body(body, body_pos):
+    """
+    Attempt to rewind body to a certain position.
+    Primarily used for request redirects and retries.
+
+    :param body:
+        File-like object that supports seek.
+
+    :param int pos:
+        Position to seek to in file.
+    """
+    body_seek = getattr(body, "seek", None)
+    if body_seek is not None and isinstance(body_pos, integer_types):
+        try:
+            body_seek(body_pos)
+        except (IOError, OSError):
+            raise UnrewindableBodyError(
+                "An error occurred when rewinding request body for redirect/retry."
+            )
+    elif body_pos is _FAILEDTELL:
+        raise UnrewindableBodyError(
+            "Unable to record file position for rewinding "
+            "request body during a redirect/retry."
+        )
+    else:
+        raise ValueError(
+            "body_pos must be of type integer, instead it was %s." % type(body_pos)
+        )

tuning-competition-baseline/.venv/lib/python3.11/site-packages/pip/_vendor/urllib3/util/ssl_match_hostname.py

@@ -0,0 +1,159 @@
+"""The match_hostname() function from Python 3.3.3, essential when using SSL."""
+
+# Note: This file is under the PSF license as the code comes from the python
+# stdlib.   http://docs.python.org/3/license.html
+
+import re
+import sys
+
+# ipaddress has been backported to 2.6+ in pypi.  If it is installed on the
+# system, use it to handle IPAddress ServerAltnames (this was added in
+# python-3.5) otherwise only do DNS matching.  This allows
+# util.ssl_match_hostname to continue to be used in Python 2.7.
+try:
+    import ipaddress
+except ImportError:
+    ipaddress = None
+
+__version__ = "3.5.0.1"
+
+
+class CertificateError(ValueError):
+    pass
+
+
+def _dnsname_match(dn, hostname, max_wildcards=1):
+    """Matching according to RFC 6125, section 6.4.3
+
+    http://tools.ietf.org/html/rfc6125#section-6.4.3
+    """
+    pats = []
+    if not dn:
+        return False
+
+    # Ported from python3-syntax:
+    # leftmost, *remainder = dn.split(r'.')
+    parts = dn.split(r".")
+    leftmost = parts[0]
+    remainder = parts[1:]
+
+    wildcards = leftmost.count("*")
+    if wildcards > max_wildcards:
+        # Issue #17980: avoid denials of service by refusing more
+        # than one wildcard per fragment.  A survey of established
+        # policy among SSL implementations showed it to be a
+        # reasonable choice.
+        raise CertificateError(
+            "too many wildcards in certificate DNS name: " + repr(dn)
+        )
+
+    # speed up common case w/o wildcards
+    if not wildcards:
+        return dn.lower() == hostname.lower()
+
+    # RFC 6125, section 6.4.3, subitem 1.
+    # The client SHOULD NOT attempt to match a presented identifier in which
+    # the wildcard character comprises a label other than the left-most label.
+    if leftmost == "*":
+        # When '*' is a fragment by itself, it matches a non-empty dotless
+        # fragment.
+        pats.append("[^.]+")
+    elif leftmost.startswith("xn--") or hostname.startswith("xn--"):
+        # RFC 6125, section 6.4.3, subitem 3.
+        # The client SHOULD NOT attempt to match a presented identifier
+        # where the wildcard character is embedded within an A-label or
+        # U-label of an internationalized domain name.
+        pats.append(re.escape(leftmost))
+    else:
+        # Otherwise, '*' matches any dotless string, e.g. www*
+        pats.append(re.escape(leftmost).replace(r"\*", "[^.]*"))
+
+    # add the remaining fragments, ignore any wildcards
+    for frag in remainder:
+        pats.append(re.escape(frag))
+
+    pat = re.compile(r"\A" + r"\.".join(pats) + r"\Z", re.IGNORECASE)
+    return pat.match(hostname)
+
+
+def _to_unicode(obj):
+    if isinstance(obj, str) and sys.version_info < (3,):
+        # ignored flake8 # F821 to support python 2.7 function
+        obj = unicode(obj, encoding="ascii", errors="strict")  # noqa: F821
+    return obj
+
+
+def _ipaddress_match(ipname, host_ip):
+    """Exact matching of IP addresses.
+
+    RFC 6125 explicitly doesn't define an algorithm for this
+    (section 1.7.2 - "Out of Scope").
+    """
+    # OpenSSL may add a trailing newline to a subjectAltName's IP address
+    # Divergence from upstream: ipaddress can't handle byte str
+    ip = ipaddress.ip_address(_to_unicode(ipname).rstrip())
+    return ip == host_ip
+
+
+def match_hostname(cert, hostname):
+    """Verify that *cert* (in decoded format as returned by
+    SSLSocket.getpeercert()) matches the *hostname*.  RFC 2818 and RFC 6125
+    rules are followed, but IP addresses are not accepted for *hostname*.
+
+    CertificateError is raised on failure. On success, the function
+    returns nothing.
+    """
+    if not cert:
+        raise ValueError(
+            "empty or no certificate, match_hostname needs a "
+            "SSL socket or SSL context with either "
+            "CERT_OPTIONAL or CERT_REQUIRED"
+        )
+    try:
+        # Divergence from upstream: ipaddress can't handle byte str
+        host_ip = ipaddress.ip_address(_to_unicode(hostname))
+    except (UnicodeError, ValueError):
+        # ValueError: Not an IP address (common case)
+        # UnicodeError: Divergence from upstream: Have to deal with ipaddress not taking
+        # byte strings.  addresses should be all ascii, so we consider it not
+        # an ipaddress in this case
+        host_ip = None
+    except AttributeError:
+        # Divergence from upstream: Make ipaddress library optional
+        if ipaddress is None:
+            host_ip = None
+        else:  # Defensive
+            raise
+    dnsnames = []
+    san = cert.get("subjectAltName", ())
+    for key, value in san:
+        if key == "DNS":
+            if host_ip is None and _dnsname_match(value, hostname):
+                return
+            dnsnames.append(value)
+        elif key == "IP Address":
+            if host_ip is not None and _ipaddress_match(value, host_ip):
+                return
+            dnsnames.append(value)
+    if not dnsnames:
+        # The subject is only checked when there is no dNSName entry
+        # in subjectAltName
+        for sub in cert.get("subject", ()):
+            for key, value in sub:
+                # XXX according to RFC 2818, the most specific Common Name
+                # must be used.
+                if key == "commonName":
+                    if _dnsname_match(value, hostname):
+                        return
+                    dnsnames.append(value)
+    if len(dnsnames) > 1:
+        raise CertificateError(
+            "hostname %r "
+            "doesn't match either of %s" % (hostname, ", ".join(map(repr, dnsnames)))
+        )
+    elif len(dnsnames) == 1:
+        raise CertificateError("hostname %r doesn't match %r" % (hostname, dnsnames[0]))
+    else:
+        raise CertificateError(
+            "no appropriate commonName or subjectAltName fields were found"
+        )

tuning-competition-baseline/.venv/lib/python3.11/site-packages/pip/_vendor/urllib3/util/ssltransport.py

@@ -0,0 +1,221 @@
+import io
+import socket
+import ssl
+
+from ..exceptions import ProxySchemeUnsupported
+from ..packages import six
+
+SSL_BLOCKSIZE = 16384
+
+
+class SSLTransport:
+    """
+    The SSLTransport wraps an existing socket and establishes an SSL connection.
+
+    Contrary to Python's implementation of SSLSocket, it allows you to chain
+    multiple TLS connections together. It's particularly useful if you need to
+    implement TLS within TLS.
+
+    The class supports most of the socket API operations.
+    """
+
+    @staticmethod
+    def _validate_ssl_context_for_tls_in_tls(ssl_context):
+        """
+        Raises a ProxySchemeUnsupported if the provided ssl_context can't be used
+        for TLS in TLS.
+
+        The only requirement is that the ssl_context provides the 'wrap_bio'
+        methods.
+        """
+
+        if not hasattr(ssl_context, "wrap_bio"):
+            if six.PY2:
+                raise ProxySchemeUnsupported(
+                    "TLS in TLS requires SSLContext.wrap_bio() which isn't "
+                    "supported on Python 2"
+                )
+            else:
+                raise ProxySchemeUnsupported(
+                    "TLS in TLS requires SSLContext.wrap_bio() which isn't "
+                    "available on non-native SSLContext"
+                )
+
+    def __init__(
+        self, socket, ssl_context, server_hostname=None, suppress_ragged_eofs=True
+    ):
+        """
+        Create an SSLTransport around socket using the provided ssl_context.
+        """
+        self.incoming = ssl.MemoryBIO()
+        self.outgoing = ssl.MemoryBIO()
+
+        self.suppress_ragged_eofs = suppress_ragged_eofs
+        self.socket = socket
+
+        self.sslobj = ssl_context.wrap_bio(
+            self.incoming, self.outgoing, server_hostname=server_hostname
+        )
+
+        # Perform initial handshake.
+        self._ssl_io_loop(self.sslobj.do_handshake)
+
+    def __enter__(self):
+        return self
+
+    def __exit__(self, *_):
+        self.close()
+
+    def fileno(self):
+        return self.socket.fileno()
+
+    def read(self, len=1024, buffer=None):
+        return self._wrap_ssl_read(len, buffer)
+
+    def recv(self, len=1024, flags=0):
+        if flags != 0:
+            raise ValueError("non-zero flags not allowed in calls to recv")
+        return self._wrap_ssl_read(len)
+
+    def recv_into(self, buffer, nbytes=None, flags=0):
+        if flags != 0:
+            raise ValueError("non-zero flags not allowed in calls to recv_into")
+        if buffer and (nbytes is None):
+            nbytes = len(buffer)
+        elif nbytes is None:
+            nbytes = 1024
+        return self.read(nbytes, buffer)
+
+    def sendall(self, data, flags=0):
+        if flags != 0:
+            raise ValueError("non-zero flags not allowed in calls to sendall")
+        count = 0
+        with memoryview(data) as view, view.cast("B") as byte_view:
+            amount = len(byte_view)
+            while count < amount:
+                v = self.send(byte_view[count:])
+                count += v
+
+    def send(self, data, flags=0):
+        if flags != 0:
+            raise ValueError("non-zero flags not allowed in calls to send")
+        response = self._ssl_io_loop(self.sslobj.write, data)
+        return response
+
+    def makefile(
+        self, mode="r", buffering=None, encoding=None, errors=None, newline=None
+    ):
+        """
+        Python's httpclient uses makefile and buffered io when reading HTTP
+        messages and we need to support it.
+
+        This is unfortunately a copy and paste of socket.py makefile with small
+        changes to point to the socket directly.
+        """
+        if not set(mode) <= {"r", "w", "b"}:
+            raise ValueError("invalid mode %r (only r, w, b allowed)" % (mode,))
+
+        writing = "w" in mode
+        reading = "r" in mode or not writing
+        assert reading or writing
+        binary = "b" in mode
+        rawmode = ""
+        if reading:
+            rawmode += "r"
+        if writing:
+            rawmode += "w"
+        raw = socket.SocketIO(self, rawmode)
+        self.socket._io_refs += 1
+        if buffering is None:
+            buffering = -1
+        if buffering < 0:
+            buffering = io.DEFAULT_BUFFER_SIZE
+        if buffering == 0:
+            if not binary:
+                raise ValueError("unbuffered streams must be binary")
+            return raw
+        if reading and writing:
+            buffer = io.BufferedRWPair(raw, raw, buffering)
+        elif reading:
+            buffer = io.BufferedReader(raw, buffering)
+        else:
+            assert writing
+            buffer = io.BufferedWriter(raw, buffering)
+        if binary:
+            return buffer
+        text = io.TextIOWrapper(buffer, encoding, errors, newline)
+        text.mode = mode
+        return text
+
+    def unwrap(self):
+        self._ssl_io_loop(self.sslobj.unwrap)
+
+    def close(self):
+        self.socket.close()
+
+    def getpeercert(self, binary_form=False):
+        return self.sslobj.getpeercert(binary_form)
+
+    def version(self):
+        return self.sslobj.version()
+
+    def cipher(self):
+        return self.sslobj.cipher()
+
+    def selected_alpn_protocol(self):
+        return self.sslobj.selected_alpn_protocol()
+
+    def selected_npn_protocol(self):
+        return self.sslobj.selected_npn_protocol()
+
+    def shared_ciphers(self):
+        return self.sslobj.shared_ciphers()
+
+    def compression(self):
+        return self.sslobj.compression()
+
+    def settimeout(self, value):
+        self.socket.settimeout(value)
+
+    def gettimeout(self):
+        return self.socket.gettimeout()
+
+    def _decref_socketios(self):
+        self.socket._decref_socketios()
+
+    def _wrap_ssl_read(self, len, buffer=None):
+        try:
+            return self._ssl_io_loop(self.sslobj.read, len, buffer)
+        except ssl.SSLError as e:
+            if e.errno == ssl.SSL_ERROR_EOF and self.suppress_ragged_eofs:
+                return 0  # eof, return 0.
+            else:
+                raise
+
+    def _ssl_io_loop(self, func, *args):
+        """Performs an I/O loop between incoming/outgoing and the socket."""
+        should_loop = True
+        ret = None
+
+        while should_loop:
+            errno = None
+            try:
+                ret = func(*args)
+            except ssl.SSLError as e:
+                if e.errno not in (ssl.SSL_ERROR_WANT_READ, ssl.SSL_ERROR_WANT_WRITE):
+                    # WANT_READ, and WANT_WRITE are expected, others are not.
+                    raise e
+                errno = e.errno
+
+            buf = self.outgoing.read()
+            self.socket.sendall(buf)
+
+            if errno is None:
+                should_loop = False
+            elif errno == ssl.SSL_ERROR_WANT_READ:
+                buf = self.socket.recv(SSL_BLOCKSIZE)
+                if buf:
+                    self.incoming.write(buf)
+                else:
+                    self.incoming.write_eof()
+        return ret

tuning-competition-baseline/.venv/lib/python3.11/site-packages/torch/ao/nn/intrinsic/quantized/dynamic/modules/__init__.py

@@ -0,0 +1,6 @@
+import torch
+from .linear_relu import LinearReLU
+
+__all__ = [
+    'LinearReLU',
+]

tuning-competition-baseline/.venv/lib/python3.11/site-packages/torch/ao/nn/qat/dynamic/__init__.py

@@ -0,0 +1 @@
+from .modules import *  # noqa: F403

tuning-competition-baseline/.venv/lib/python3.11/site-packages/torch/ao/nn/quantizable/__init__.py

@@ -0,0 +1 @@
+from .modules import *  # noqa: F403

tuning-competition-baseline/.venv/lib/python3.11/site-packages/torch/ao/nn/quantizable/modules/rnn.py

@@ -0,0 +1,411 @@
+import numbers
+from typing import Optional, Tuple
+import warnings
+
+import torch
+from torch import Tensor
+
+"""
+We will recreate all the RNN modules as we require the modules to be decomposed
+into its building blocks to be able to observe.
+"""
+
+__all__ = [
+    "LSTMCell",
+    "LSTM"
+]
+
+class LSTMCell(torch.nn.Module):
+    r"""A quantizable long short-term memory (LSTM) cell.
+
+    For the description and the argument types, please, refer to :class:`~torch.nn.LSTMCell`
+
+    Examples::
+
+        >>> import torch.ao.nn.quantizable as nnqa
+        >>> rnn = nnqa.LSTMCell(10, 20)
+        >>> input = torch.randn(6, 10)
+        >>> hx = torch.randn(3, 20)
+        >>> cx = torch.randn(3, 20)
+        >>> output = []
+        >>> for i in range(6):
+        ...     hx, cx = rnn(input[i], (hx, cx))
+        ...     output.append(hx)
+    """
+    _FLOAT_MODULE = torch.nn.LSTMCell
+
+    def __init__(self, input_dim: int, hidden_dim: int, bias: bool = True,
+                 device=None, dtype=None) -> None:
+        factory_kwargs = {'device': device, 'dtype': dtype}
+        super().__init__()
+        self.input_size = input_dim
+        self.hidden_size = hidden_dim
+        self.bias = bias
+
+        self.igates = torch.nn.Linear(input_dim, 4 * hidden_dim, bias=bias, **factory_kwargs)
+        self.hgates = torch.nn.Linear(hidden_dim, 4 * hidden_dim, bias=bias, **factory_kwargs)
+        self.gates = torch.ao.nn.quantized.FloatFunctional()
+
+        self.input_gate = torch.nn.Sigmoid()
+        self.forget_gate = torch.nn.Sigmoid()
+        self.cell_gate = torch.nn.Tanh()
+        self.output_gate = torch.nn.Sigmoid()
+
+        self.fgate_cx = torch.ao.nn.quantized.FloatFunctional()
+        self.igate_cgate = torch.ao.nn.quantized.FloatFunctional()
+        self.fgate_cx_igate_cgate = torch.ao.nn.quantized.FloatFunctional()
+
+        self.ogate_cy = torch.ao.nn.quantized.FloatFunctional()
+
+        self.initial_hidden_state_qparams: Tuple[float, int] = (1.0, 0)
+        self.initial_cell_state_qparams: Tuple[float, int] = (1.0, 0)
+        self.hidden_state_dtype: torch.dtype = torch.quint8
+        self.cell_state_dtype: torch.dtype = torch.quint8
+
+    def forward(self, x: Tensor, hidden: Optional[Tuple[Tensor, Tensor]] = None) -> Tuple[Tensor, Tensor]:
+        if hidden is None or hidden[0] is None or hidden[1] is None:
+            hidden = self.initialize_hidden(x.shape[0], x.is_quantized)
+        hx, cx = hidden
+
+        igates = self.igates(x)
+        hgates = self.hgates(hx)
+        gates = self.gates.add(igates, hgates)
+
+        input_gate, forget_gate, cell_gate, out_gate = gates.chunk(4, 1)
+
+        input_gate = self.input_gate(input_gate)
+        forget_gate = self.forget_gate(forget_gate)
+        cell_gate = self.cell_gate(cell_gate)
+        out_gate = self.output_gate(out_gate)
+
+        fgate_cx = self.fgate_cx.mul(forget_gate, cx)
+        igate_cgate = self.igate_cgate.mul(input_gate, cell_gate)
+        fgate_cx_igate_cgate = self.fgate_cx_igate_cgate.add(fgate_cx, igate_cgate)
+        cy = fgate_cx_igate_cgate
+
+        # TODO: make this tanh a member of the module so its qparams can be configured
+        tanh_cy = torch.tanh(cy)
+        hy = self.ogate_cy.mul(out_gate, tanh_cy)
+        return hy, cy
+
+    def initialize_hidden(self, batch_size: int, is_quantized: bool = False) -> Tuple[Tensor, Tensor]:
+        h, c = torch.zeros((batch_size, self.hidden_size)), torch.zeros((batch_size, self.hidden_size))
+        if is_quantized:
+            (h_scale, h_zp) = self.initial_hidden_state_qparams
+            (c_scale, c_zp) = self.initial_cell_state_qparams
+            h = torch.quantize_per_tensor(h, scale=h_scale, zero_point=h_zp, dtype=self.hidden_state_dtype)
+            c = torch.quantize_per_tensor(c, scale=c_scale, zero_point=c_zp, dtype=self.cell_state_dtype)
+        return h, c
+
+    def _get_name(self):
+        return 'QuantizableLSTMCell'
+
+    @classmethod
+    def from_params(cls, wi, wh, bi=None, bh=None):
+        """Uses the weights and biases to create a new LSTM cell.
+
+        Args:
+            wi, wh: Weights for the input and hidden layers
+            bi, bh: Biases for the input and hidden layers
+        """
+        assert (bi is None) == (bh is None)  # Either both None or both have values
+        input_size = wi.shape[1]
+        hidden_size = wh.shape[1]
+        cell = cls(input_dim=input_size, hidden_dim=hidden_size,
+                   bias=(bi is not None))
+        cell.igates.weight = torch.nn.Parameter(wi)
+        if bi is not None:
+            cell.igates.bias = torch.nn.Parameter(bi)
+        cell.hgates.weight = torch.nn.Parameter(wh)
+        if bh is not None:
+            cell.hgates.bias = torch.nn.Parameter(bh)
+        return cell
+
+    @classmethod
+    def from_float(cls, other):
+        assert type(other) == cls._FLOAT_MODULE
+        assert hasattr(other, 'qconfig'), "The float module must have 'qconfig'"
+        observed = cls.from_params(other.weight_ih, other.weight_hh,
+                                   other.bias_ih, other.bias_hh)
+        observed.qconfig = other.qconfig
+        observed.igates.qconfig = other.qconfig
+        observed.hgates.qconfig = other.qconfig
+        return observed
+
+
+class _LSTMSingleLayer(torch.nn.Module):
+    r"""A single one-directional LSTM layer.
+
+    The difference between a layer and a cell is that the layer can process a
+    sequence, while the cell only expects an instantaneous value.
+    """
+    def __init__(self, input_dim: int, hidden_dim: int, bias: bool = True,
+                 device=None, dtype=None) -> None:
+        factory_kwargs = {'device': device, 'dtype': dtype}
+        super().__init__()
+        self.cell = LSTMCell(input_dim, hidden_dim, bias=bias, **factory_kwargs)
+
+    def forward(self, x: Tensor, hidden: Optional[Tuple[Tensor, Tensor]] = None):
+        result = []
+        seq_len = x.shape[0]
+        for i in range(seq_len):
+            hidden = self.cell(x[i], hidden)
+            result.append(hidden[0])  # type: ignore[index]
+        result_tensor = torch.stack(result, 0)
+        return result_tensor, hidden
+
+    @classmethod
+    def from_params(cls, *args, **kwargs):
+        cell = LSTMCell.from_params(*args, **kwargs)
+        layer = cls(cell.input_size, cell.hidden_size, cell.bias)
+        layer.cell = cell
+        return layer
+
+
+class _LSTMLayer(torch.nn.Module):
+    r"""A single bi-directional LSTM layer."""
+    def __init__(self, input_dim: int, hidden_dim: int, bias: bool = True,
+                 batch_first: bool = False, bidirectional: bool = False,
+                 device=None, dtype=None) -> None:
+        factory_kwargs = {'device': device, 'dtype': dtype}
+        super().__init__()
+        self.batch_first = batch_first
+        self.bidirectional = bidirectional
+        self.layer_fw = _LSTMSingleLayer(input_dim, hidden_dim, bias=bias, **factory_kwargs)
+        if self.bidirectional:
+            self.layer_bw = _LSTMSingleLayer(input_dim, hidden_dim, bias=bias, **factory_kwargs)
+
+    def forward(self, x: Tensor, hidden: Optional[Tuple[Tensor, Tensor]] = None):
+        if self.batch_first:
+            x = x.transpose(0, 1)
+        if hidden is None:
+            hx_fw, cx_fw = (None, None)
+        else:
+            hx_fw, cx_fw = hidden
+        hidden_bw: Optional[Tuple[Tensor, Tensor]] = None
+        if self.bidirectional:
+            if hx_fw is None:
+                hx_bw = None
+            else:
+                hx_bw = hx_fw[1]
+                hx_fw = hx_fw[0]
+            if cx_fw is None:
+                cx_bw = None
+            else:
+                cx_bw = cx_fw[1]
+                cx_fw = cx_fw[0]
+            if hx_bw is not None and cx_bw is not None:
+                hidden_bw = hx_bw, cx_bw
+        if hx_fw is None and cx_fw is None:
+            hidden_fw = None
+        else:
+            hidden_fw = torch.jit._unwrap_optional(hx_fw), torch.jit._unwrap_optional(cx_fw)
+        result_fw, hidden_fw = self.layer_fw(x, hidden_fw)
+
+        if hasattr(self, 'layer_bw') and self.bidirectional:
+            x_reversed = x.flip(0)
+            result_bw, hidden_bw = self.layer_bw(x_reversed, hidden_bw)
+            result_bw = result_bw.flip(0)
+
+            result = torch.cat([result_fw, result_bw], result_fw.dim() - 1)
+            if hidden_fw is None and hidden_bw is None:
+                h = None
+                c = None
+            elif hidden_fw is None:
+                (h, c) = torch.jit._unwrap_optional(hidden_bw)
+            elif hidden_bw is None:
+                (h, c) = torch.jit._unwrap_optional(hidden_fw)
+            else:
+                h = torch.stack([hidden_fw[0], hidden_bw[0]], 0)  # type: ignore[list-item]
+                c = torch.stack([hidden_fw[1], hidden_bw[1]], 0)  # type: ignore[list-item]
+        else:
+            result = result_fw
+            h, c = torch.jit._unwrap_optional(hidden_fw)  # type: ignore[assignment]
+
+        if self.batch_first:
+            result.transpose_(0, 1)
+
+        return result, (h, c)
+
+    @classmethod
+    def from_float(cls, other, layer_idx=0, qconfig=None, **kwargs):
+        r"""
+        There is no FP equivalent of this class. This function is here just to
+        mimic the behavior of the `prepare` within the `torch.ao.quantization`
+        flow.
+        """
+        assert hasattr(other, 'qconfig') or (qconfig is not None)
+
+        input_size = kwargs.get('input_size', other.input_size)
+        hidden_size = kwargs.get('hidden_size', other.hidden_size)
+        bias = kwargs.get('bias', other.bias)
+        batch_first = kwargs.get('batch_first', other.batch_first)
+        bidirectional = kwargs.get('bidirectional', other.bidirectional)
+
+        layer = cls(input_size, hidden_size, bias, batch_first, bidirectional)
+        layer.qconfig = getattr(other, 'qconfig', qconfig)
+        wi = getattr(other, f'weight_ih_l{layer_idx}')
+        wh = getattr(other, f'weight_hh_l{layer_idx}')
+        bi = getattr(other, f'bias_ih_l{layer_idx}', None)
+        bh = getattr(other, f'bias_hh_l{layer_idx}', None)
+
+        layer.layer_fw = _LSTMSingleLayer.from_params(wi, wh, bi, bh)
+
+        if other.bidirectional:
+            wi = getattr(other, f'weight_ih_l{layer_idx}_reverse')
+            wh = getattr(other, f'weight_hh_l{layer_idx}_reverse')
+            bi = getattr(other, f'bias_ih_l{layer_idx}_reverse', None)
+            bh = getattr(other, f'bias_hh_l{layer_idx}_reverse', None)
+            layer.layer_bw = _LSTMSingleLayer.from_params(wi, wh, bi, bh)
+        return layer
+
+
+class LSTM(torch.nn.Module):
+    r"""A quantizable long short-term memory (LSTM).
+
+    For the description and the argument types, please, refer to :class:`~torch.nn.LSTM`
+
+    Attributes:
+        layers : instances of the `_LSTMLayer`
+
+    .. note::
+        To access the weights and biases, you need to access them per layer.
+        See examples below.
+
+    Examples::
+
+        >>> import torch.ao.nn.quantizable as nnqa
+        >>> rnn = nnqa.LSTM(10, 20, 2)
+        >>> input = torch.randn(5, 3, 10)
+        >>> h0 = torch.randn(2, 3, 20)
+        >>> c0 = torch.randn(2, 3, 20)
+        >>> output, (hn, cn) = rnn(input, (h0, c0))
+        >>> # To get the weights:
+        >>> # xdoctest: +SKIP
+        >>> print(rnn.layers[0].weight_ih)
+        tensor([[...]])
+        >>> print(rnn.layers[0].weight_hh)
+        AssertionError: There is no reverse path in the non-bidirectional layer
+    """
+    _FLOAT_MODULE = torch.nn.LSTM
+
+    def __init__(self, input_size: int, hidden_size: int,
+                 num_layers: int = 1, bias: bool = True,
+                 batch_first: bool = False, dropout: float = 0.,
+                 bidirectional: bool = False,
+                 device=None, dtype=None) -> None:
+        factory_kwargs = {'device': device, 'dtype': dtype}
+        super().__init__()
+        self.input_size = input_size
+        self.hidden_size = hidden_size
+        self.num_layers = num_layers
+        self.bias = bias
+        self.batch_first = batch_first
+        self.dropout = float(dropout)
+        self.bidirectional = bidirectional
+        self.training = False  # Default to eval mode. If we want to train, we will explicitly set to training.
+        num_directions = 2 if bidirectional else 1
+
+        if not isinstance(dropout, numbers.Number) or not 0 <= dropout <= 1 or \
+                isinstance(dropout, bool):
+            raise ValueError("dropout should be a number in range [0, 1] "
+                             "representing the probability of an element being "
+                             "zeroed")
+        if dropout > 0:
+            warnings.warn("dropout option for quantizable LSTM is ignored. "
+                          "If you are training, please, use nn.LSTM version "
+                          "followed by `prepare` step.")
+            if num_layers == 1:
+                warnings.warn("dropout option adds dropout after all but last "
+                              "recurrent layer, so non-zero dropout expects "
+                              f"num_layers greater than 1, but got dropout={dropout} "
+                              f"and num_layers={num_layers}")
+
+        layers = [_LSTMLayer(self.input_size, self.hidden_size,
+                             self.bias, batch_first=False,
+                             bidirectional=self.bidirectional, **factory_kwargs)]
+        for layer in range(1, num_layers):
+            layers.append(_LSTMLayer(self.hidden_size, self.hidden_size,
+                                     self.bias, batch_first=False,
+                                     bidirectional=self.bidirectional,
+                                     **factory_kwargs))
+        self.layers = torch.nn.ModuleList(layers)
+
+    def forward(self, x: Tensor, hidden: Optional[Tuple[Tensor, Tensor]] = None):
+        if self.batch_first:
+            x = x.transpose(0, 1)
+
+        max_batch_size = x.size(1)
+        num_directions = 2 if self.bidirectional else 1
+        if hidden is None:
+            zeros = torch.zeros(num_directions, max_batch_size,
+                                self.hidden_size, dtype=torch.float,
+                                device=x.device)
+            zeros.squeeze_(0)
+            if x.is_quantized:
+                zeros = torch.quantize_per_tensor(zeros, scale=1.0,
+                                                  zero_point=0, dtype=x.dtype)
+            hxcx = [(zeros, zeros) for _ in range(self.num_layers)]
+        else:
+            hidden_non_opt = torch.jit._unwrap_optional(hidden)
+            if isinstance(hidden_non_opt[0], Tensor):
+                hx = hidden_non_opt[0].reshape(self.num_layers, num_directions,
+                                               max_batch_size,
+                                               self.hidden_size)
+                cx = hidden_non_opt[1].reshape(self.num_layers, num_directions,
+                                               max_batch_size,
+                                               self.hidden_size)
+                hxcx = [(hx[idx].squeeze(0), cx[idx].squeeze(0)) for idx in range(self.num_layers)]
+            else:
+                hxcx = hidden_non_opt
+
+        hx_list = []
+        cx_list = []
+        for idx, layer in enumerate(self.layers):
+            x, (h, c) = layer(x, hxcx[idx])
+            hx_list.append(torch.jit._unwrap_optional(h))
+            cx_list.append(torch.jit._unwrap_optional(c))
+        hx_tensor = torch.stack(hx_list)
+        cx_tensor = torch.stack(cx_list)
+
+        # We are creating another dimension for bidirectional case
+        # need to collapse it
+        hx_tensor = hx_tensor.reshape(-1, hx_tensor.shape[-2], hx_tensor.shape[-1])
+        cx_tensor = cx_tensor.reshape(-1, cx_tensor.shape[-2], cx_tensor.shape[-1])
+
+        if self.batch_first:
+            x = x.transpose(0, 1)
+
+        return x, (hx_tensor, cx_tensor)
+
+    def _get_name(self):
+        return 'QuantizableLSTM'
+
+    @classmethod
+    def from_float(cls, other, qconfig=None):
+        assert isinstance(other, cls._FLOAT_MODULE)
+        assert (hasattr(other, 'qconfig') or qconfig)
+        observed = cls(other.input_size, other.hidden_size, other.num_layers,
+                       other.bias, other.batch_first, other.dropout,
+                       other.bidirectional)
+        observed.qconfig = getattr(other, 'qconfig', qconfig)
+        for idx in range(other.num_layers):
+            observed.layers[idx] = _LSTMLayer.from_float(other, idx, qconfig,
+                                                         batch_first=False)
+
+        # Prepare the model
+        if other.training:
+            observed.train()
+            observed = torch.ao.quantization.prepare_qat(observed, inplace=True)
+        else:
+            observed.eval()
+            observed = torch.ao.quantization.prepare(observed, inplace=True)
+        return observed
+
+    @classmethod
+    def from_observed(cls, other):
+        # The whole flow is float -> observed -> quantized
+        # This class does float -> observed only
+        raise NotImplementedError("It looks like you are trying to convert a "
+                                  "non-quantizable LSTM module. Please, see "
+                                  "the examples on quantizable LSTMs.")

tuning-competition-baseline/.venv/lib/python3.11/site-packages/torch/ao/nn/quantized/__init__.py

@@ -0,0 +1,38 @@
+from . import functional
+from .modules import *  # noqa: F403
+from .modules import MaxPool2d
+
+__all__ = [
+    'BatchNorm2d',
+    'BatchNorm3d',
+    'Conv1d',
+    'Conv2d',
+    'Conv3d',
+    'ConvTranspose1d',
+    'ConvTranspose2d',
+    'ConvTranspose3d',
+    'DeQuantize',
+    'ELU',
+    'Embedding',
+    'EmbeddingBag',
+    'GroupNorm',
+    'Hardswish',
+    'InstanceNorm1d',
+    'InstanceNorm2d',
+    'InstanceNorm3d',
+    'LayerNorm',
+    'LeakyReLU',
+    'Linear',
+    'LSTM',
+    'MultiheadAttention',
+    'Quantize',
+    'ReLU6',
+    'Sigmoid',
+    'Softmax',
+    'Dropout',
+    'PReLU',
+    # Wrapper modules
+    'FloatFunctional',
+    'FXFloatFunctional',
+    'QFunctional',
+]

tuning-competition-baseline/.venv/lib/python3.11/site-packages/torch/ao/nn/quantized/dynamic/modules/__init__.py

@@ -0,0 +1,19 @@
+
+from .linear import Linear
+from .rnn import LSTM, GRU, LSTMCell, RNNCell, GRUCell
+from .conv import Conv1d, Conv2d, Conv3d, ConvTranspose1d, ConvTranspose2d, ConvTranspose3d
+
+__all__ = [
+    'Linear',
+    'LSTM',
+    'GRU',
+    'LSTMCell',
+    'RNNCell',
+    'GRUCell',
+    'Conv1d',
+    'Conv2d',
+    'Conv3d',
+    'ConvTranspose1d',
+    'ConvTranspose2d',
+    'ConvTranspose3d',
+]

tuning-competition-baseline/.venv/lib/python3.11/site-packages/torch/ao/nn/quantized/dynamic/modules/conv.py

@@ -0,0 +1,399 @@
+r"""Dynamically quantized convolution modules."""
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from torch import Tensor
+from torch._ops import ops
+from torch.nn.common_types import _size_1_t
+from torch.nn.modules.utils import _single, _pair, _triple
+from torch.ao.nn.quantized.modules.conv import _reverse_repeat_padding
+import torch.ao.nn.quantized as nnq
+import warnings
+
+__all__ = ['Conv1d', 'Conv2d', 'Conv3d', 'ConvTranspose1d', 'ConvTranspose2d', 'ConvTranspose3d']
+
+
+class Conv1d(nnq.Conv1d):
+    r"""A dynamically quantized conv module with floating point tensors as inputs and outputs.
+
+    For details on input arguments, parameters, and implementation see
+    :class:`~torch.nn.Conv1d` and :class:`~torch.ao.nn.quantized.dynamic.Conv1d` and
+
+    Attributes:
+        weight (Tensor):     packed tensor derived from the learnable weight
+                             parameter.
+        scale (Tensor):      scalar for the output scale
+        zero_point (Tensor): scalar for the output zero point
+
+    See :class:`~torch.nn.Conv1d` for other attributes.
+
+    Examples::
+
+        >>> # xdoctest: +SKIP
+        >>> m = nn.quantized.dynamic.Conv1d(16, 33, 3, stride=2)
+        >>> input = torch.randn(20, 16, 100)
+        >>> output = m(input)
+
+    """
+
+    _FLOAT_MODULE = nn.Conv1d
+    _NNIQAT_CONV_BN_MODULE = None  # type: ignore[assignment]
+    _NNI_CONV_RELU_MODULE = None  # type: ignore[assignment]
+
+    def __init__(self,
+                 in_channels: int,
+                 out_channels: int,
+                 kernel_size: _size_1_t,
+                 stride: _size_1_t = 1,
+                 padding: _size_1_t = 0,
+                 dilation: _size_1_t = 1,
+                 groups: int = 1,
+                 bias: bool = True,
+                 padding_mode: str = 'zeros',
+                 device=None,
+                 dtype=None,
+                 reduce_range=True):
+        warnings.warn(
+            "The current implementation of the {} module has poor numerical accuracy and its use is not recommended".format(
+                self._get_name()
+            )
+        )
+        factory_kwargs = {'device': device, 'dtype': dtype}
+        kernel_size = _single(kernel_size)
+        stride = _single(stride)
+        padding = padding if isinstance(padding, str) else _single(padding)
+        dilation = _single(dilation)
+
+        super().__init__(
+            in_channels, out_channels, kernel_size, stride, padding, dilation,
+            groups, bias, padding_mode, **factory_kwargs)
+
+    def _get_name(self):
+        return 'DynamicQuantizedConv1d'
+
+    def forward(self, input: Tensor, reduce_range: bool = True) -> Tensor:
+        # Temporarily using len(shape) instead of ndim due to JIT issue
+        # https://github.com/pytorch/pytorch/issues/23890
+        if len(input.shape) != 3:
+            raise ValueError("Input shape must be `(N, C, L)`!")
+        if self.padding_mode != 'zeros':
+            # Padding in Conv1d is stored as (p, p), need to get (p,)
+            _reversed_padding_repeated_twice = _reverse_repeat_padding(self.padding[:1])
+            input = F.pad(input, _reversed_padding_repeated_twice,
+                          mode=self.padding_mode)
+        return ops.quantized.conv1d_dynamic(input, self._packed_params, reduce_range)
+
+
+class Conv2d(nnq.Conv2d):
+    r"""A dynamically quantized conv module with floating point tensors as inputs and outputs.
+
+    For details on input arguments, parameters, and implementation see
+    :class:`~torch.nn.Conv2d` and :class:`~torch.ao.nn.quantized.dynamic.Conv2d` and
+
+    Attributes:
+        weight (Tensor):     packed tensor derived from the learnable weight
+                             parameter.
+        scale (Tensor):      scalar for the output scale
+        zero_point (Tensor): scalar for the output zero point
+
+    See :class:`~torch.nn.Conv2d` for other attributes.
+
+    Examples::
+
+        >>> # xdoctest: +SKIP
+        >>> # With square kernels and equal stride
+        >>> m = nn.quantized.dynamic.Conv2d(16, 33, 3, stride=2)
+        >>> # non-square kernels and unequal stride and with padding
+        >>> m = nn.quantized.dynamic.Conv2d(16, 33, (3, 5), stride=(2, 1), padding=(4, 2))
+        >>> # non-square kernels and unequal stride and with padding and dilation
+        >>> m = nn.quantized.dynamic.Conv2d(16, 33, (3, 5), stride=(2, 1), padding=(4, 2), dilation=(3, 1))
+        >>> input = torch.randn(20, 16, 50, 100)
+        >>> output = m(input)
+
+    """
+    _FLOAT_MODULE = nn.Conv2d
+    _NNIQAT_CONV_BN_MODULE = None  # type: ignore[assignment]
+    _NNI_CONV_RELU_MODULE = None  # type: ignore[assignment]
+
+    def __init__(self, in_channels, out_channels, kernel_size, stride=1,
+                 padding=0, dilation=1, groups=1, bias=True,
+                 padding_mode='zeros', device=None, dtype=None):
+        warnings.warn(
+            "The current implementation of the {} module has poor numerical accuracy and its use is not recommended".format(
+                self._get_name()
+            )
+        )
+        factory_kwargs = {'device': device, 'dtype': dtype}
+        kernel_size = _pair(kernel_size)
+        stride = _pair(stride)
+        padding = _pair(padding)
+        dilation = _pair(dilation)
+
+        super().__init__(
+            in_channels, out_channels, kernel_size, stride, padding, dilation,
+            groups, bias, padding_mode, **factory_kwargs)
+
+    def _get_name(self):
+        return 'DynamicQuantizedConv2d'
+
+    def forward(self, input: Tensor, reduce_range: bool = True) -> Tensor:
+        # Temporarily using len(shape) instead of ndim due to JIT issue
+        # https://github.com/pytorch/pytorch/issues/23890
+        if len(input.shape) != 4:
+            raise ValueError("Input shape must be `(N, C, H, W)`!")
+        if self.padding_mode != 'zeros':
+            _reversed_padding_repeated_twice = _reverse_repeat_padding(self.padding)
+            input = F.pad(input, _reversed_padding_repeated_twice,
+                          mode=self.padding_mode)
+        return ops.quantized.conv2d_dynamic(
+            input, self._packed_params, reduce_range)
+
+
+class Conv3d(nnq.Conv3d):
+    r"""A dynamically quantized conv module with floating point tensors as inputs and outputs.
+
+    For details on input arguments, parameters, and implementation see
+    :class:`~torch.nn.Conv3d` and :class:`~torch.ao.nn.quantized.dynamic.Conv3d` and
+
+    Attributes:
+        weight (Tensor):     packed tensor derived from the learnable weight
+                             parameter.
+        scale (Tensor):      scalar for the output scale
+        zero_point (Tensor): scalar for the output zero point
+
+    See :class:`~torch.nn.Conv3d` for other attributes.
+
+    Examples::
+
+        >>> # xdoctest: +SKIP
+        >>> # With square kernels and equal stride
+        >>> m = nn.quantized.dynamic.Conv3d(16, 33, 3, stride=2)
+        >>> # non-square kernels and unequal stride and with padding
+        >>> m = nn.quantized.dynamic.Conv3d(16, 33, (3, 5, 5), stride=(1, 2, 2), padding=(1, 2, 2))
+        >>> # non-square kernels and unequal stride and with padding and dilation
+        >>> m = nn.quantized.dynamic.Conv3d(16, 33, (3, 5, 5), stride=(1, 2, 2), padding=(1, 2, 2), dilation=(1, 2, 2))
+        >>> input = torch.randn(20, 16, 56, 56, 56)
+        >>> output = m(input)
+
+    """
+    _FLOAT_MODULE = nn.Conv3d
+    _NNIQAT_CONV_BN_MODULE = None  # type: ignore[assignment]
+    _NNI_CONV_RELU_MODULE = None  # type: ignore[assignment]
+
+    def __init__(self, in_channels, out_channels, kernel_size, stride=1,
+                 padding=0, dilation=1, groups=1, bias=True,
+                 padding_mode='zeros', device=None, dtype=None):
+        warnings.warn(
+            "The current implementation of the {} module has poor numerical accuracy and its use is not recommended".format(
+                self._get_name()
+            )
+        )
+        assert padding_mode != 'reflect', "Conv3d does not support reflection padding"
+        factory_kwargs = {'device': device, 'dtype': dtype}
+        kernel_size = _triple(kernel_size)
+        stride = _triple(stride)
+        padding = _triple(padding)
+        dilation = _triple(dilation)
+        super()._init(
+            in_channels, out_channels, kernel_size, stride, padding, dilation,
+            False, _triple(0), groups, bias, padding_mode, **factory_kwargs)
+
+    def _get_name(self):
+        return 'DynamicQuantizedConv3d'
+
+    def forward(self, input: Tensor, reduce_range: bool = True) -> Tensor:
+        # Temporarily using len(shape) instead of ndim due to JIT issue
+        # https://github.com/pytorch/pytorch/issues/23890
+        if len(input.shape) != 5:
+            raise ValueError("Input shape must be `(N, C, D, H, W)`!")
+        if self.padding_mode != 'zeros':
+            _reversed_padding_repeated_twice = _reverse_repeat_padding(self.padding)
+            input = F.pad(input, _reversed_padding_repeated_twice,
+                          mode=self.padding_mode)
+        return ops.quantized.conv3d_dynamic(
+            input, self._packed_params, reduce_range)
+
+
+class ConvTranspose1d(nnq.ConvTranspose1d):
+    r"""A dynamically quantized transposed convolution module with floating point tensors as inputs and outputs.
+
+    For details on input arguments, parameters, and implementation see
+    :class:`~torch.nn.ConvTranspose1d`.
+
+    For special notes, please, see :class:`~torch.ao.nn.quantized.dynamic.Conv1d`
+
+    Attributes:
+        weight (Tensor):     packed tensor derived from the learnable weight
+                             parameter.
+        scale (Tensor):      scalar for the output scale
+        zero_point (Tensor): scalar for the output zero point
+    See :class:`~torch.nn.ConvTranspose1d` for other attributes.
+
+    Examples::
+
+        >>> # xdoctest: +SKIP
+        >>> # With square kernels and equal stride
+        >>> m = nndq.ConvTranspose1d(16, 33, 3, stride=2)
+        >>> # non-square kernels and unequal stride and with padding
+        >>> m = nndq.ConvTranspose1d(16, 33, (3, 5), stride=(2, 1), padding=(4, 2))
+        >>> output = m(input)
+        >>> # exact output size can be also specified as an argument
+        >>> downsample = nndq.Conv1d(16, 16, 3, stride=2, padding=1)
+        >>> upsample = nndq.ConvTranspose1d(16, 16, 3, stride=2, padding=1)
+        >>> h = downsample(input)
+        >>> h.size()
+        torch.Size([1, 16, 6])
+        >>> output = upsample(h, output_size=input.size())
+        >>> output.size()
+        torch.Size([1, 16, 12])
+    """
+
+    _FLOAT_MODULE = nn.ConvTranspose1d
+
+    def __init__(self, in_channels, out_channels, kernel_size, stride=1,
+                 padding=0, output_padding=0, groups=1, bias=True,
+                 dilation=1, padding_mode='zeros', device=None, dtype=None):
+        warnings.warn(
+            "The current implementation of the {} module has poor numerical accuracy and its use is not recommended".format(
+                self._get_name()
+            )
+        )
+        factory_kwargs = {'device': device, 'dtype': dtype}
+        super().__init__(
+            in_channels, out_channels, kernel_size, stride, padding, output_padding,
+            groups, bias, dilation, padding_mode, **factory_kwargs)
+
+    def _get_name(self):
+        return 'DynamicQuantizedConvTranspose1d'
+
+    def forward(self, input: Tensor, reduce_range: bool = True) -> Tensor:
+        # Temporarily using len(shape) instead of ndim due to JIT issue
+        # https://github.com/pytorch/pytorch/issues/23890
+        if len(input.shape) != 3:
+            raise ValueError("Input shape must be `(N, C, L)`!")
+        return torch.ops.quantized.conv_transpose1d_dynamic(
+            input, self._packed_params, reduce_range)
+
+
+class ConvTranspose2d(nnq.ConvTranspose2d):
+    r"""A dynamically quantized transposed convolution module with floating point tensors as inputs and outputs.
+
+    For details on input arguments, parameters, and implementation see
+    :class:`~torch.nn.ConvTranspose2d`.
+
+    For special notes, please, see :class:`~torch.ao.nn.quantized.dynamic.Conv2d`
+
+    Attributes:
+        weight (Tensor):     packed tensor derived from the learnable weight
+                             parameter.
+        scale (Tensor):      scalar for the output scale
+        zero_point (Tensor): scalar for the output zero point
+    See :class:`~torch.nn.ConvTranspose2d` for other attributes.
+
+    Examples::
+
+        >>> # xdoctest: +SKIP
+        >>> # With square kernels and equal stride
+        >>> m = nnq.ConvTranspose2d(16, 33, 3, stride=2)
+        >>> # non-square kernels and unequal stride and with padding
+        >>> m = nnq.ConvTranspose2d(16, 33, (3, 5), stride=(2, 1), padding=(4, 2))
+        >>> output = m(input)
+        >>> # exact output size can be also specified as an argument
+        >>> downsample = nnq.Conv2d(16, 16, 3, stride=2, padding=1)
+        >>> upsample = nnq.ConvTranspose2d(16, 16, 3, stride=2, padding=1)
+        >>> h = downsample(input)
+        >>> h.size()
+        torch.Size([1, 16, 6, 6])
+        >>> output = upsample(h, output_size=input.size())
+        >>> output.size()
+        torch.Size([1, 16, 12, 12])
+    """
+
+    _FLOAT_MODULE = nn.ConvTranspose2d
+
+    def __init__(self, in_channels, out_channels, kernel_size, stride=1,
+                 padding=0, output_padding=0, groups=1, bias=True,
+                 dilation=1, padding_mode='zeros', device=None, dtype=None):
+        warnings.warn(
+            "The current implementation of the {} module has poor numerical accuracy and its use is not recommended".format(
+                self._get_name()
+            )
+        )
+        factory_kwargs = {'device': device, 'dtype': dtype}
+        super().__init__(
+            in_channels, out_channels, kernel_size, stride, padding, output_padding,
+            groups, bias, dilation, padding_mode, **factory_kwargs)
+
+    def _get_name(self):
+        return 'DynamicQuantizedConvTranspose2d'
+
+    def forward(self, input: Tensor, reduce_range: bool = True) -> Tensor:
+        # Temporarily using len(shape) instead of ndim due to JIT issue
+        # https://github.com/pytorch/pytorch/issues/23890
+        if len(input.shape) != 4:
+            raise ValueError("Input shape must be `(N, C, H, W)`!")
+        return ops.quantized.conv_transpose2d_dynamic(
+            input, self._packed_params, reduce_range)
+
+
+class ConvTranspose3d(nnq.ConvTranspose3d):
+    r"""A dynamically quantized transposed convolution module with floating point tensors as inputs and outputs.
+
+    For details on input arguments, parameters, and implementation see
+    :class:`~torch.nn.ConvTranspose3d`.
+
+    For special notes, please, see :class:`~torch.ao.nn.quantized.dynamic.Conv3d`
+
+    Attributes:
+        weight (Tensor):     packed tensor derived from the learnable weight
+                             parameter.
+        scale (Tensor):      scalar for the output scale
+        zero_point (Tensor): scalar for the output zero point
+    See :class:`~torch.nn.ConvTranspose3d` for other attributes.
+
+    Examples::
+
+        >>> # xdoctest: +SKIP
+        >>> # With cubic kernels and equal stride
+        >>> m = nnq.ConvTranspose3d(16, 33, 3, stride=2)
+        >>> # non-cubic kernels and unequal stride and with padding
+        >>> m = nnq.ConvTranspose3d(16, 33, (3, 3, 5), stride=(2, 1, 1), padding=(4, 2, 2))
+        >>> output = m(input)
+        >>> # exact output size can be also specified as an argument
+        >>> downsample = nnq.Conv3d(16, 16, 3, stride=2, padding=1)
+        >>> upsample = nnq.ConvTranspose3d(16, 16, 3, stride=2, padding=1)
+        >>> h = downsample(input)
+        >>> h.size()
+        torch.Size([1, 16, 6, 6, 6])
+        >>> output = upsample(h, output_size=input.size())
+        >>> output.size()
+        torch.Size([1, 16, 12, 12, 12])
+    """
+
+    _FLOAT_MODULE = nn.ConvTranspose3d
+
+    def __init__(self, in_channels, out_channels, kernel_size, stride=1,
+                 padding=0, output_padding=0, groups=1, bias=True,
+                 dilation=1, padding_mode='zeros', device=None, dtype=None):
+        warnings.warn(
+            "The current implementation of the {} module has poor numerical accuracy and its use is not recommended".format(
+                self._get_name()
+            )
+        )
+        factory_kwargs = {'device': device, 'dtype': dtype}
+        super().__init__(
+            in_channels, out_channels, kernel_size, stride, padding, output_padding,
+            groups, bias, dilation, padding_mode, **factory_kwargs)
+
+    def _get_name(self):
+        return 'DynamicQuantizedConvTranspose3d'
+
+    def forward(self, input: Tensor, reduce_range: bool = True) -> Tensor:
+        # Temporarily using len(shape) instead of ndim due to JIT issue
+        # https://github.com/pytorch/pytorch/issues/23890
+        if len(input.shape) != 5:
+            raise ValueError("Input shape must be `(N, C, T, H, W)`!")
+        return ops.quantized.conv_transpose3d_dynamic(
+            input, self._packed_params, reduce_range)

tuning-competition-baseline/.venv/lib/python3.11/site-packages/torch/ao/nn/quantized/dynamic/modules/linear.py

@@ -0,0 +1,132 @@
+import torch
+import torch.ao.nn.quantized as nnq
+from torch.ao.nn.quantized.modules.utils import _quantize_weight
+import torch.ao.nn.intrinsic as nni
+
+__all__ = [
+    "Linear",
+]
+
+
+class Linear(nnq.Linear):
+    r"""
+    A dynamic quantized linear module with floating point tensor as inputs and outputs.
+    We adopt the same interface as `torch.nn.Linear`, please see
+    https://pytorch.org/docs/stable/nn.html#torch.nn.Linear for documentation.
+
+    Similar to :class:`torch.nn.Linear`, attributes will be randomly
+    initialized at module creation time and will be overwritten later
+
+    Attributes:
+        weight (Tensor): the non-learnable quantized weights of the module which are of
+                         shape :math:`(\text{out\_features}, \text{in\_features})`.
+        bias (Tensor): the non-learnable floating point bias of the module of shape
+                       :math:`(\text{out\_features})`. If :attr:`bias` is ``True``,
+                       the values are initialized to zero.
+
+    Examples::
+
+        >>> # xdoctest: +SKIP
+        >>> m = nn.quantized.dynamic.Linear(20, 30)
+        >>> input = torch.randn(128, 20)
+        >>> output = m(input)
+        >>> print(output.size())
+        torch.Size([128, 30])
+    """
+    # version used in this class is different from the parent class nnq.Linear
+    _version = 4
+
+    def __init__(self, in_features, out_features, bias_=True, dtype=torch.qint8):
+        super().__init__(in_features, out_features, bias_, dtype=dtype)
+        # We don't muck around with buffers or attributes or anything here
+        # to keep the module simple. *everything* is simply a Python attribute.
+        # Serialization logic is explicitly handled in the below serialization and
+        # deserialization modules
+        self.version = 4
+
+    def forward(self, x):
+        # Note that we can handle self.bias == None case.
+        if self._packed_params.dtype == torch.qint8:
+            if self.version is None or self.version < 4:
+                Y = torch.ops.quantized.linear_dynamic(
+                    x, self._packed_params._packed_params)
+            else:
+                Y = torch.ops.quantized.linear_dynamic(
+                    x, self._packed_params._packed_params, reduce_range=True)
+        elif self._packed_params.dtype == torch.float16:
+            Y = torch.ops.quantized.linear_dynamic_fp16(
+                x, self._packed_params._packed_params)
+        else:
+            raise RuntimeError('Unsupported dtype on dynamic quantized linear!')
+        return Y.to(x.dtype)
+
+    def _get_name(self):
+        return 'DynamicQuantizedLinear'
+
+    def extra_repr(self):
+        extra_repr_str = 'in_features={}, out_features={}, dtype={}'.format(
+            self.in_features, self.out_features, self._packed_params.dtype
+        )
+        if self._packed_params.dtype == torch.qint8:
+            extra_repr_str += f', qscheme={self.weight().qscheme()}'
+        return extra_repr_str
+
+    def _load_from_state_dict(self, state_dict, prefix, local_metadata, strict,
+                              missing_keys, unexpected_keys, error_msgs):
+        version = local_metadata.get('version', None)
+        self.version = version
+        super()._load_from_state_dict(state_dict, prefix, local_metadata, False,
+                                      missing_keys, unexpected_keys, error_msgs)
+
+    @classmethod
+    def from_float(cls, mod):
+        r"""Create a dynamic quantized module from a float module or qparams_dict
+
+        Args:
+            mod (Module): a float module, either produced by torch.ao.quantization
+                          utilities or provided by the user
+        """
+        float_modules = [torch.nn.Linear, torch.nn.modules.linear.NonDynamicallyQuantizableLinear,
+                         torch.ao.nn.intrinsic.modules.fused.LinearReLU, torch.ao.nn.qat.dynamic.Linear]
+
+        assert type(mod) in float_modules, \
+            'nn.quantized.dynamic.Linear.from_float only works for one of' + \
+            str([float_mod.__name__ for float_mod in float_modules])
+        assert hasattr(mod, 'qconfig'), 'Input float module must have qconfig defined'
+        if type(mod) == nni.LinearReLU:
+            mod = mod[0]
+        if mod.qconfig is not None and mod.qconfig.weight is not None:
+            weight_observer = mod.qconfig.weight()
+        else:
+            # We have the circular import issues if we import the qconfig in the beginning of this file:
+            # https://github.com/pytorch/pytorch/pull/24231. The current workaround is to postpone the
+            # import until we need it.
+            from torch.ao.quantization.qconfig import default_dynamic_qconfig
+            weight_observer = default_dynamic_qconfig.weight()
+        dtype = weight_observer.dtype
+        assert dtype in [torch.qint8, torch.float16], "The only supported dtypes for " \
+            f"dynamic quantized linear are qint8 and float16 got: {dtype}"
+        weight_observer(mod.weight)
+        if dtype == torch.qint8:
+            qweight = _quantize_weight(mod.weight.float(), weight_observer)
+        elif dtype == torch.float16:
+            qweight = mod.weight.float()
+        else:
+            raise RuntimeError('Unsupported dtype specified for dynamic quantized Linear!')
+        qlinear = cls(mod.in_features, mod.out_features, dtype=dtype)
+        qlinear.set_weight_bias(qweight, mod.bias)
+        return qlinear
+
+    @classmethod
+    def from_reference(cls, ref_qlinear):
+        """ Create a (fbgemm/qnnpack) dynamic quantized module from a reference quantized
+        module
+        Args:
+            ref_qlinear (Module): a reference quantized  module, either produced by
+            torch.ao.quantization functions or provided by the user
+        """
+        qlinear = cls(ref_qlinear.in_features, ref_qlinear.out_features, dtype=ref_qlinear.weight_dtype)
+        qweight = ref_qlinear.get_quantized_weight()
+        bias = ref_qlinear.bias
+        qlinear.set_weight_bias(qweight, bias)
+        return qlinear

tuning-competition-baseline/.venv/lib/python3.11/site-packages/torch/ao/nn/quantized/dynamic/modules/rnn.py

@@ -0,0 +1,1096 @@
+import numbers
+import warnings
+
+import torch
+import torch.nn as nn
+from torch import Tensor  # noqa: F401
+from torch._jit_internal import Tuple, Optional, List, Union, Dict  # noqa: F401
+from torch.nn.utils.rnn import PackedSequence
+from torch.ao.nn.quantized.modules.utils import _quantize_weight
+
+__all__ = ['pack_weight_bias', 'PackedParameter', 'RNNBase', 'LSTM', 'GRU', 'RNNCellBase', 'RNNCell', 'LSTMCell',
+           'GRUCell', "apply_permutation"]
+
+
+def _apply_permutation(tensor: Tensor, permutation: Tensor, dim: int = 1) -> Tensor:
+    return tensor.index_select(dim, permutation)
+
+
+def apply_permutation(tensor: Tensor, permutation: Tensor, dim: int = 1) -> Tensor:
+    warnings.warn("apply_permutation is deprecated, please use tensor.index_select(dim, permutation) instead")
+    return _apply_permutation(tensor, permutation, dim)
+
+
+def pack_weight_bias(qweight, bias, dtype):
+
+    if dtype == torch.qint8:
+        # for each layer, for each direction we need to quantize and pack
+        # weights and pack parameters in this order:
+        #
+        #   w_ih, w_hh
+        packed_weight = \
+            torch.ops.quantized.linear_prepack(qweight, bias)
+
+        return packed_weight
+    else:
+        # for each layer, for each direction we need to quantize and pack
+        # weights and pack parameters in this order:
+        #
+        #   packed_ih, packed_hh, b_ih, b_hh
+        packed_weight = torch.ops.quantized.linear_prepack_fp16(
+            qweight, bias)
+
+        return packed_weight
+
+
+class PackedParameter(torch.nn.Module):
+    def __init__(self, param):
+        super().__init__()
+        self.param = param
+
+    def _save_to_state_dict(self, destination, prefix, keep_vars):
+        super()._save_to_state_dict(destination, prefix, keep_vars)
+        destination[prefix + 'param'] = self.param
+
+    def _load_from_state_dict(self, state_dict, prefix, local_metadata, strict,
+                              missing_keys, unexpected_keys, error_msgs):
+        self.param = state_dict[prefix + 'param']
+        super()._load_from_state_dict(state_dict, prefix, local_metadata, False,
+                                      missing_keys, unexpected_keys, error_msgs)
+
+
+class RNNBase(torch.nn.Module):
+
+    _FLOAT_MODULE = nn.RNNBase
+
+    _version = 2
+
+    def __init__(self, mode, input_size, hidden_size,
+                 num_layers=1, bias=True, batch_first=False,
+                 dropout=0., bidirectional=False, dtype=torch.qint8):
+        super().__init__()
+
+        self.mode = mode
+        self.input_size = input_size
+        self.hidden_size = hidden_size
+        self.num_layers = num_layers
+        self.bias = bias
+        self.batch_first = batch_first
+        self.dropout = float(dropout)
+        self.bidirectional = bidirectional
+        self.dtype = dtype
+        self.version = 2
+        self.training = False
+        num_directions = 2 if bidirectional else 1
+
+        # "type: ignore" is required since ints and Numbers are not fully comparable
+        # https://github.com/python/mypy/issues/8566
+        if not isinstance(dropout, numbers.Number) \
+                or not 0 <= dropout <= 1 or isinstance(dropout, bool):  # type: ignore[operator]
+            raise ValueError("dropout should be a number in range [0, 1] "
+                             "representing the probability of an element being "
+                             "zeroed")
+        if dropout > 0 and num_layers == 1:  # type: ignore[operator]
+            warnings.warn("dropout option adds dropout after all but last "
+                          "recurrent layer, so non-zero dropout expects "
+                          f"num_layers greater than 1, but got dropout={dropout} and "
+                          f"num_layers={num_layers}")
+
+        if mode == 'LSTM':
+            gate_size = 4 * hidden_size
+        elif mode == 'GRU':
+            gate_size = 3 * hidden_size
+        else:
+            raise ValueError("Unrecognized RNN mode: " + mode)
+
+        _all_weight_values = []
+        for layer in range(num_layers):
+            for direction in range(num_directions):
+                layer_input_size = input_size if layer == 0 else hidden_size * num_directions
+
+                w_ih = torch.randn(gate_size, layer_input_size).to(torch.float)
+                w_hh = torch.randn(gate_size, hidden_size).to(torch.float)
+                b_ih = torch.randn(gate_size).to(torch.float)
+                b_hh = torch.randn(gate_size).to(torch.float)
+                if dtype == torch.qint8:
+                    w_ih = torch.quantize_per_tensor(w_ih, scale=0.1, zero_point=0, dtype=torch.qint8)
+                    w_hh = torch.quantize_per_tensor(w_hh, scale=0.1, zero_point=0, dtype=torch.qint8)
+                    packed_ih = \
+                        torch.ops.quantized.linear_prepack(w_ih, b_ih)
+                    packed_hh = \
+                        torch.ops.quantized.linear_prepack(w_hh, b_hh)
+                    if self.version is None or self.version < 2:
+                        cell_params = torch.ops.quantized.make_quantized_cell_params_dynamic(
+                            packed_ih, packed_hh, b_ih, b_hh)
+                    else:
+                        cell_params = torch.ops.quantized.make_quantized_cell_params_dynamic(
+                            packed_ih, packed_hh, b_ih, b_hh, True)
+                else:
+                    packed_ih = torch.ops.quantized.linear_prepack_fp16(w_ih, b_ih)
+                    packed_hh = torch.ops.quantized.linear_prepack_fp16(w_hh, b_hh)
+                    cell_params = torch.ops.quantized.make_quantized_cell_params_fp16(
+                        packed_ih, packed_hh)
+
+                _all_weight_values.append(PackedParameter(cell_params))
+        self._all_weight_values = torch.nn.ModuleList(_all_weight_values)
+
+    def _get_name(self):
+        return 'DynamicQuantizedRNN'
+
+    def extra_repr(self):
+        s = '{input_size}, {hidden_size}'
+        if self.num_layers != 1:
+            s += ', num_layers={num_layers}'
+        if self.bias is not True:
+            s += ', bias={bias}'
+        if self.batch_first is not False:
+            s += ', batch_first={batch_first}'
+        if self.dropout != 0:
+            s += ', dropout={dropout}'
+        if self.bidirectional is not False:
+            s += ', bidirectional={bidirectional}'
+        return s.format(**self.__dict__)
+
+    def __repr__(self):
+        # We don't want to show `ModuleList` children, hence custom
+        # `__repr__`. This is the same as nn.Module.__repr__, except the check
+        # for the `PackedParameter` and `nn.ModuleList`.
+        # You should still override `extra_repr` to add more info.
+        extra_lines = []
+        extra_repr = self.extra_repr()
+        # empty string will be split into list ['']
+        if extra_repr:
+            extra_lines = extra_repr.split('\n')
+        child_lines = []
+        for key, module in self._modules.items():
+            if isinstance(module, (PackedParameter, nn.ModuleList)):
+                continue
+            mod_str = repr(module)
+            mod_str = nn.modules.module._addindent(mod_str, 2)
+            child_lines.append('(' + key + '): ' + mod_str)
+        lines = extra_lines + child_lines
+
+        main_str = self._get_name() + '('
+        if lines:
+            # simple one-liner info, which most builtin Modules will use
+            if len(extra_lines) == 1 and not child_lines:
+                main_str += extra_lines[0]
+            else:
+                main_str += '\n  ' + '\n  '.join(lines) + '\n'
+
+        main_str += ')'
+        return main_str
+
+    def check_input(self, input: Tensor, batch_sizes: Optional[Tensor]) -> None:
+        expected_input_dim = 2 if batch_sizes is not None else 3
+        if input.dim() != expected_input_dim:
+            raise RuntimeError(
+                f'input must have {expected_input_dim} dimensions, got {input.dim()}')
+        if self.input_size != input.size(-1):
+            raise RuntimeError(
+                f'input.size(-1) must be equal to input_size. Expected {self.input_size}, got {input.size(-1)}')
+
+    def get_expected_hidden_size(self, input: Tensor, batch_sizes: Optional[Tensor]) -> Tuple[int, int, int]:
+        if batch_sizes is not None:
+            mini_batch = int(batch_sizes[0])
+        else:
+            mini_batch = input.size(0) if self.batch_first else input.size(1)
+        num_directions = 2 if self.bidirectional else 1
+        expected_hidden_size = (self.num_layers * num_directions,
+                                mini_batch, self.hidden_size)
+        return expected_hidden_size
+
+    def check_hidden_size(
+        self, hx: Tensor, expected_hidden_size: Tuple[int, int, int],
+        msg: str = 'Expected hidden size {}, got {}'
+    ) -> None:
+        if hx.size() != expected_hidden_size:
+            raise RuntimeError(msg.format(
+                expected_hidden_size, list(hx.size())))
+
+    def check_forward_args(self, input: Tensor, hidden: Tensor, batch_sizes: Optional[Tensor]) -> None:
+        self.check_input(input, batch_sizes)
+        expected_hidden_size = self.get_expected_hidden_size(input, batch_sizes)
+        self.check_hidden_size(hidden, expected_hidden_size,
+                               msg='Expected hidden size {}, got {}')
+
+    def permute_hidden(self, hx: Tensor, permutation: Optional[Tensor]) -> Tensor:
+        if permutation is None:
+            return hx
+        return _apply_permutation(hx, permutation)
+
+    def _load_from_state_dict(self, state_dict, prefix, local_metadata, strict,
+                              missing_keys, unexpected_keys, error_msgs):
+        version = local_metadata.get('version', None)
+        self.version = version
+        super()._load_from_state_dict(state_dict, prefix, local_metadata, False,
+                                      missing_keys, unexpected_keys, error_msgs)
+
+    def set_weight_bias(self, weight_bias_dict):
+
+        def weight_bias_name(ihhh, layer, suffix):
+            weight_name = f"weight_{ihhh}_l{layer}{suffix}"
+            bias_name = f"bias_{ihhh}_l{layer}{suffix}"
+            return weight_name, bias_name
+
+        num_directions = 2 if self.bidirectional else 1
+        # TODO: dedup with __init__ of RNNBase
+        _all_weight_values = []
+        for layer in range(self.num_layers):
+            for direction in range(num_directions):
+                suffix = "_reverse" if direction == 1 else ""
+                w_ih_name, b_ih_name = weight_bias_name("ih", layer, suffix)
+                w_hh_name, b_hh_name = weight_bias_name("hh", layer, suffix)
+                w_ih = weight_bias_dict[w_ih_name]
+                b_ih = weight_bias_dict[b_ih_name]
+                w_hh = weight_bias_dict[w_hh_name]
+                b_hh = weight_bias_dict[b_hh_name]
+                if w_ih.dtype == torch.qint8:
+                    packed_ih = torch.ops.quantized.linear_prepack(w_ih, b_ih)
+                    packed_hh = torch.ops.quantized.linear_prepack(w_hh, b_hh)
+                    if self.version is None or self.version < 2:
+                        cell_params = torch.ops.quantized.make_quantized_cell_params_dynamic(
+                            packed_ih, packed_hh, b_ih, b_hh)
+                    else:
+                        cell_params = torch.ops.quantized.make_quantized_cell_params_dynamic(
+                            packed_ih, packed_hh, b_ih, b_hh, True)
+                else:
+                    packed_ih = torch.ops.quantized.linear_prepack_fp16(w_ih, b_ih)
+                    packed_hh = torch.ops.quantized.linear_prepack_fp16(w_hh, b_hh)
+                    cell_params = torch.ops.quantized.make_quantized_cell_params_fp16(
+                        packed_ih, packed_hh)
+
+                _all_weight_values.append(PackedParameter(cell_params))
+        self._all_weight_values = torch.nn.ModuleList(_all_weight_values)
+
+    @classmethod
+    def from_float(cls, mod):
+        assert type(mod) in {torch.nn.LSTM,
+                             torch.nn.GRU}, 'nn.quantized.dynamic.RNNBase.from_float only works for nn.LSTM and nn.GRU'
+        assert hasattr(
+            mod,
+            'qconfig'
+        ), 'Input float module must have qconfig defined'
+
+        if mod.qconfig is not None and mod.qconfig.weight is not None:
+            weight_observer_method = mod.qconfig.weight
+        else:
+            # We have the circular import issues if we import the qconfig in the beginning of this file:
+            # https://github.com/pytorch/pytorch/pull/24231. The current workaround is to postpone the
+            # import until we need it.
+            from torch.ao.quantization.qconfig import default_dynamic_qconfig
+            weight_observer_method = default_dynamic_qconfig.weight
+
+        dtype = weight_observer_method().dtype
+        supported_scalar_types = [torch.qint8, torch.float16]
+        if dtype not in supported_scalar_types:
+            raise RuntimeError(f'Unsupported dtype for dynamic RNN quantization: {dtype}')
+        # RNNBase can be either LSTM or GRU
+        qRNNBase: Union[LSTM, GRU]
+        if mod.mode == 'LSTM':
+            qRNNBase = LSTM(mod.input_size, mod.hidden_size, mod.num_layers,
+                            mod.bias, mod.batch_first, mod.dropout, mod.bidirectional, dtype)
+        elif mod.mode == 'GRU':
+            qRNNBase = GRU(mod.input_size, mod.hidden_size, mod.num_layers,
+                           mod.bias, mod.batch_first, mod.dropout, mod.bidirectional, dtype)
+        else:
+            raise NotImplementedError('Only LSTM/GRU is supported for QuantizedRNN for now')
+
+        num_directions = 2 if mod.bidirectional else 1
+
+        assert mod.bias
+
+        _all_weight_values = []
+        for layer in range(qRNNBase.num_layers):
+            for direction in range(num_directions):
+                suffix = '_reverse' if direction == 1 else ''
+
+                def retrieve_weight_bias(ihhh):
+                    weight_name = f'weight_{ihhh}_l{layer}{suffix}'
+                    bias_name = f'bias_{ihhh}_l{layer}{suffix}'
+                    weight = getattr(mod, weight_name)
+                    bias = getattr(mod, bias_name)
+                    return weight, bias
+
+                weight_ih, bias_ih = retrieve_weight_bias('ih')
+                weight_hh, bias_hh = retrieve_weight_bias('hh')
+
+                if dtype == torch.qint8:
+                    def quantize_and_pack(w, b):
+                        weight_observer = weight_observer_method()
+                        weight_observer(w)
+                        qweight = _quantize_weight(w.float(), weight_observer)
+                        packed_weight = \
+                            torch.ops.quantized.linear_prepack(qweight, b)
+                        return packed_weight
+                    packed_ih = quantize_and_pack(weight_ih, bias_ih)
+                    packed_hh = quantize_and_pack(weight_hh, bias_hh)
+                    if qRNNBase.version is None or qRNNBase.version < 2:
+                        cell_params = torch.ops.quantized.make_quantized_cell_params_dynamic(
+                            packed_ih, packed_hh, bias_ih, bias_hh)
+                    else:
+                        cell_params = torch.ops.quantized.make_quantized_cell_params_dynamic(
+                            packed_ih, packed_hh, bias_ih, bias_hh, True)
+
+                elif dtype == torch.float16:
+                    packed_ih = torch.ops.quantized.linear_prepack_fp16(
+                        weight_ih.float(), bias_ih)
+                    packed_hh = torch.ops.quantized.linear_prepack_fp16(
+                        weight_hh.float(), bias_hh)
+
+                    cell_params = torch.ops.quantized.make_quantized_cell_params_fp16(
+                        packed_ih, packed_hh)
+                else:
+                    raise RuntimeError('Unsupported dtype specified for dynamic quantized LSTM!')
+
+                _all_weight_values.append(PackedParameter(cell_params))
+        qRNNBase._all_weight_values = torch.nn.ModuleList(_all_weight_values)
+
+        return qRNNBase
+
+    def _weight_bias(self):
+        # Returns a dict of weights and biases
+        weight_bias_dict: Dict[str, Dict] = {'weight' : {}, 'bias' : {}}
+        count = 0
+        num_directions = 2 if self.bidirectional else 1
+        for layer in range(self.num_layers):
+            for direction in range(num_directions):
+                suffix = '_reverse' if direction == 1 else ''
+                key_name1 = f'weight_ih_l{layer}{suffix}'
+                key_name2 = f'weight_hh_l{layer}{suffix}'
+                # packed weights are part of torchbind class, CellParamsSerializationType
+                # Within the packed weight class, the weight and bias are accessible as Tensors
+                packed_weight_bias = self._all_weight_values[count].param.__getstate__()[0][4]
+                weight_bias_dict['weight'][key_name1] = packed_weight_bias[0].__getstate__()[0][0]
+                weight_bias_dict['weight'][key_name2] = packed_weight_bias[1].__getstate__()[0][0]
+                key_name1 = f'bias_ih_l{layer}{suffix}'
+                key_name2 = f'bias_hh_l{layer}{suffix}'
+                weight_bias_dict['bias'][key_name1] = packed_weight_bias[0].__getstate__()[0][1]
+                weight_bias_dict['bias'][key_name2] = packed_weight_bias[1].__getstate__()[0][1]
+                count = count + 1
+        return weight_bias_dict
+
+    def get_weight(self):
+        return self._weight_bias()['weight']
+
+    def get_bias(self):
+        return self._weight_bias()['bias']
+
+
+class LSTM(RNNBase):
+    r"""
+    A dynamic quantized LSTM module with floating point tensor as inputs and outputs.
+    We adopt the same interface as `torch.nn.LSTM`, please see
+    https://pytorch.org/docs/stable/nn.html#torch.nn.LSTM for documentation.
+
+    Examples::
+
+        >>> # xdoctest: +SKIP
+        >>> rnn = nn.LSTM(10, 20, 2)
+        >>> input = torch.randn(5, 3, 10)
+        >>> h0 = torch.randn(2, 3, 20)
+        >>> c0 = torch.randn(2, 3, 20)
+        >>> output, (hn, cn) = rnn(input, (h0, c0))
+    """
+    _FLOAT_MODULE = nn.LSTM
+
+    __overloads__ = {'forward': ['forward_packed', 'forward_tensor']}
+
+    def __init__(self, *args, **kwargs):
+        super().__init__('LSTM', *args, **kwargs)
+
+    def _get_name(self):
+        return 'DynamicQuantizedLSTM'
+
+    def forward_impl(
+        self, input: Tensor, hx: Optional[Tuple[Tensor, Tensor]],
+        batch_sizes: Optional[Tensor], max_batch_size: int,
+        sorted_indices: Optional[Tensor]
+    ) -> Tuple[Tensor, Tuple[Tensor, Tensor]]:
+        if hx is None:
+            num_directions = 2 if self.bidirectional else 1
+            zeros = torch.zeros(self.num_layers * num_directions,
+                                max_batch_size, self.hidden_size,
+                                dtype=input.dtype, device=input.device)
+            hx = (zeros, zeros)
+        else:
+            # Each batch of the hidden state should match the input sequence that
+            # the user believes he/she is passing in.
+            hx = self.permute_hidden(hx, sorted_indices)
+
+        self.check_forward_args(input, hx, batch_sizes)
+
+        _all_params = ([m.param for m in self._all_weight_values])
+        if batch_sizes is None:
+            result = torch.quantized_lstm(input, hx, _all_params, self.bias, self.num_layers,
+                                          float(self.dropout), self.training, self.bidirectional,
+                                          self.batch_first, dtype=self.dtype, use_dynamic=True)
+        else:
+            result = torch.quantized_lstm(input, batch_sizes, hx, _all_params, self.bias,
+                                          self.num_layers, float(self.dropout), self.training,
+                                          self.bidirectional, dtype=self.dtype, use_dynamic=True)
+        output = result[0]
+        hidden = result[1:]
+
+        return output, hidden
+
+    @torch.jit.export
+    def forward_tensor(
+        self, input: Tensor, hx: Optional[Tuple[Tensor, Tensor]] = None
+    ) -> Tuple[Tensor, Tuple[Tensor, Tensor]]:
+        batch_sizes = None
+        max_batch_size = input.size(0) if self.batch_first else input.size(1)
+        sorted_indices = None
+        unsorted_indices = None
+
+        output, hidden = self.forward_impl(
+            input, hx, batch_sizes, max_batch_size, sorted_indices)
+
+        return output, self.permute_hidden(hidden, unsorted_indices)
+
+    @torch.jit.export
+    def forward_packed(
+        self, input: PackedSequence, hx: Optional[Tuple[Tensor, Tensor]] = None
+    ) -> Tuple[PackedSequence, Tuple[Tensor, Tensor]]:
+        input_, batch_sizes, sorted_indices, unsorted_indices = input
+        max_batch_size = int(batch_sizes[0])
+
+        output_, hidden = self.forward_impl(
+            input_, hx, batch_sizes, max_batch_size, sorted_indices
+        )
+
+        output = PackedSequence(output_, batch_sizes,
+                                sorted_indices, unsorted_indices)
+        return output, self.permute_hidden(hidden, unsorted_indices)
+
+    # "type: ignore" is required due to issue #43072
+    def permute_hidden(  # type: ignore[override]
+        self, hx: Tuple[Tensor, Tensor], permutation: Optional[Tensor]
+    ) -> Tuple[Tensor, Tensor]:
+        if permutation is None:
+            return hx
+        return _apply_permutation(hx[0], permutation), _apply_permutation(hx[1], permutation)
+
+    # "type: ignore" is required due to issue #43072
+    def check_forward_args(  # type: ignore[override]
+        self, input: Tensor, hidden: Tuple[Tensor, Tensor], batch_sizes: Optional[Tensor]
+    ) -> None:
+        self.check_input(input, batch_sizes)
+        expected_hidden_size = self.get_expected_hidden_size(input, batch_sizes)
+
+        self.check_hidden_size(hidden[0], expected_hidden_size,
+                               'Expected hidden[0] size {}, got {}')
+        self.check_hidden_size(hidden[1], expected_hidden_size,
+                               'Expected hidden[1] size {}, got {}')
+
+    @torch.jit.ignore
+    def forward(self, input, hx=None):
+        if isinstance(input, PackedSequence):
+            return self.forward_packed(input, hx)
+        else:
+            return self.forward_tensor(input, hx)
+
+    @classmethod
+    def from_float(cls, mod):
+        return super().from_float(mod)
+
+    @classmethod
+    def from_reference(cls, ref_mod):
+        assert hasattr(ref_mod, "weight_ih_l0_dtype"), "We are assuming weight_ih_l0 "
+        "exists in LSTM, may need to relax the assumption to support the use case"
+        qmod = cls(
+            ref_mod.input_size,
+            ref_mod.hidden_size,
+            ref_mod.num_layers,
+            ref_mod.bias,
+            ref_mod.batch_first,
+            ref_mod.dropout,
+            ref_mod.bidirectional,
+            # assuming there is layer 0, which should be OK
+            ref_mod.weight_ih_l0_dtype,
+        )
+        qmod.set_weight_bias(ref_mod.get_quantized_weight_bias_dict())
+        return qmod
+
+
+class GRU(RNNBase):
+    r"""Applies a multi-layer gated recurrent unit (GRU) RNN to an input sequence.
+
+
+    For each element in the input sequence, each layer computes the following
+    function:
+
+    .. math::
+        \begin{array}{ll}
+            r_t = \sigma(W_{ir} x_t + b_{ir} + W_{hr} h_{(t-1)} + b_{hr}) \\
+            z_t = \sigma(W_{iz} x_t + b_{iz} + W_{hz} h_{(t-1)} + b_{hz}) \\
+            n_t = \tanh(W_{in} x_t + b_{in} + r_t \odot (W_{hn} h_{(t-1)}+ b_{hn})) \\
+            h_t = (1 - z_t) \odot n_t + z_t \odot h_{(t-1)}
+        \end{array}
+
+    where :math:`h_t` is the hidden state at time `t`, :math:`x_t` is the input
+    at time `t`, :math:`h_{(t-1)}` is the hidden state of the layer
+    at time `t-1` or the initial hidden state at time `0`, and :math:`r_t`,
+    :math:`z_t`, :math:`n_t` are the reset, update, and new gates, respectively.
+    :math:`\sigma` is the sigmoid function, and :math:`\odot` is the Hadamard product.
+
+    In a multilayer GRU, the input :math:`x^{(l)}_t` of the :math:`l` -th layer
+    (:math:`l >= 2`) is the hidden state :math:`h^{(l-1)}_t` of the previous layer multiplied by
+    dropout :math:`\delta^{(l-1)}_t` where each :math:`\delta^{(l-1)}_t` is a Bernoulli random
+    variable which is :math:`0` with probability :attr:`dropout`.
+
+    Args:
+        input_size: The number of expected features in the input `x`
+        hidden_size: The number of features in the hidden state `h`
+        num_layers: Number of recurrent layers. E.g., setting ``num_layers=2``
+            would mean stacking two GRUs together to form a `stacked GRU`,
+            with the second GRU taking in outputs of the first GRU and
+            computing the final results. Default: 1
+        bias: If ``False``, then the layer does not use bias weights `b_ih` and `b_hh`.
+            Default: ``True``
+        batch_first: If ``True``, then the input and output tensors are provided
+            as (batch, seq, feature). Default: ``False``
+        dropout: If non-zero, introduces a `Dropout` layer on the outputs of each
+            GRU layer except the last layer, with dropout probability equal to
+            :attr:`dropout`. Default: 0
+        bidirectional: If ``True``, becomes a bidirectional GRU. Default: ``False``
+
+    Inputs: input, h_0
+        - **input** of shape `(seq_len, batch, input_size)`: tensor containing the features
+          of the input sequence. The input can also be a packed variable length
+          sequence. See :func:`torch.nn.utils.rnn.pack_padded_sequence`
+          for details.
+        - **h_0** of shape `(num_layers * num_directions, batch, hidden_size)`: tensor
+          containing the initial hidden state for each element in the batch.
+          Defaults to zero if not provided. If the RNN is bidirectional,
+          num_directions should be 2, else it should be 1.
+
+    Outputs: output, h_n
+        - **output** of shape `(seq_len, batch, num_directions * hidden_size)`: tensor
+          containing the output features h_t from the last layer of the GRU,
+          for each `t`. If a :class:`torch.nn.utils.rnn.PackedSequence` has been
+          given as the input, the output will also be a packed sequence.
+          For the unpacked case, the directions can be separated
+          using ``output.view(seq_len, batch, num_directions, hidden_size)``,
+          with forward and backward being direction `0` and `1` respectively.
+
+          Similarly, the directions can be separated in the packed case.
+        - **h_n** of shape `(num_layers * num_directions, batch, hidden_size)`: tensor
+          containing the hidden state for `t = seq_len`
+
+          Like *output*, the layers can be separated using
+          ``h_n.view(num_layers, num_directions, batch, hidden_size)``.
+
+    Shape:
+        - Input1: :math:`(L, N, H_{in})` tensor containing input features where
+          :math:`H_{in}=\text{input\_size}` and `L` represents a sequence length.
+        - Input2: :math:`(S, N, H_{out})` tensor
+          containing the initial hidden state for each element in the batch.
+          :math:`H_{out}=\text{hidden\_size}`
+          Defaults to zero if not provided. where :math:`S=\text{num\_layers} * \text{num\_directions}`
+          If the RNN is bidirectional, num_directions should be 2, else it should be 1.
+        - Output1: :math:`(L, N, H_{all})` where :math:`H_{all}=\text{num\_directions} * \text{hidden\_size}`
+        - Output2: :math:`(S, N, H_{out})` tensor containing the next hidden state
+          for each element in the batch
+
+    Attributes:
+        weight_ih_l[k] : the learnable input-hidden weights of the :math:`\text{k}^{th}` layer
+            (W_ir|W_iz|W_in), of shape `(3*hidden_size, input_size)` for `k = 0`.
+            Otherwise, the shape is `(3*hidden_size, num_directions * hidden_size)`
+        weight_hh_l[k] : the learnable hidden-hidden weights of the :math:`\text{k}^{th}` layer
+            (W_hr|W_hz|W_hn), of shape `(3*hidden_size, hidden_size)`
+        bias_ih_l[k] : the learnable input-hidden bias of the :math:`\text{k}^{th}` layer
+            (b_ir|b_iz|b_in), of shape `(3*hidden_size)`
+        bias_hh_l[k] : the learnable hidden-hidden bias of the :math:`\text{k}^{th}` layer
+            (b_hr|b_hz|b_hn), of shape `(3*hidden_size)`
+
+    .. note::
+        All the weights and biases are initialized from :math:`\mathcal{U}(-\sqrt{k}, \sqrt{k})`
+        where :math:`k = \frac{1}{\text{hidden\_size}}`
+
+    .. note::
+        The calculation of new gate :math:`n_t` subtly differs from the original paper and other frameworks.
+        In the original implementation, the Hadamard product :math:`(\odot)` between :math:`r_t` and the
+        previous hidden state :math:`h_{(t-1)}` is done before the multiplication with the weight matrix
+        `W` and addition of bias:
+
+        .. math::
+            \begin{aligned}
+                n_t = \tanh(W_{in} x_t + b_{in} + W_{hn} ( r_t \odot h_{(t-1)} ) + b_{hn})
+            \end{aligned}
+
+        This is in contrast to PyTorch implementation, which is done after :math:`W_{hn} h_{(t-1)}`
+
+        .. math::
+            \begin{aligned}
+                n_t = \tanh(W_{in} x_t + b_{in} + r_t \odot (W_{hn} h_{(t-1)}+ b_{hn}))
+            \end{aligned}
+
+        This implementation differs on purpose for efficiency.
+
+    .. include:: ../cudnn_persistent_rnn.rst
+
+    Examples::
+
+        >>> # xdoctest: +SKIP
+        >>> rnn = nn.GRU(10, 20, 2)
+        >>> input = torch.randn(5, 3, 10)
+        >>> h0 = torch.randn(2, 3, 20)
+        >>> output, hn = rnn(input, h0)
+    """
+    _FLOAT_MODULE = nn.GRU
+
+    __overloads__ = {'forward': ['forward_packed', 'forward_tensor']}
+
+    def __init__(self, *args, **kwargs):
+        super().__init__('GRU', *args, **kwargs)
+
+    def _get_name(self):
+        return 'DynamicQuantizedGRU'
+
+    def check_forward_args(self, input: Tensor, hidden: Tensor, batch_sizes: Optional[Tensor]) -> None:
+        self.check_input(input, batch_sizes)
+        expected_hidden_size = self.get_expected_hidden_size(input, batch_sizes)
+
+        self.check_hidden_size(hidden, expected_hidden_size,
+                               'Expected hidden size {}, got {}')
+
+    def forward_impl(
+        self, input: Tensor, hx: Optional[Tensor],
+        batch_sizes: Optional[Tensor], max_batch_size: int,
+        sorted_indices: Optional[Tensor]
+    ) -> Tuple[Tensor, Tensor]:
+        if hx is None:
+            num_directions = 2 if self.bidirectional else 1
+            zeros = torch.zeros(self.num_layers * num_directions,
+                                max_batch_size, self.hidden_size,
+                                dtype=input.dtype, device=input.device)
+            hx = zeros
+        else:
+            # Each batch of the hidden state should match the input sequence that
+            # the user believes he/she is passing in.
+            hx = self.permute_hidden(hx, sorted_indices)
+
+        self.check_forward_args(input, hx, batch_sizes)
+
+        _all_params = ([m.param for m in self._all_weight_values])
+        if batch_sizes is None:
+            result = torch.quantized_gru(input,
+                                         hx,
+                                         _all_params,
+                                         self.bias,
+                                         self.num_layers,
+                                         self.dropout,
+                                         self.training,
+                                         self.bidirectional,
+                                         self.batch_first)
+        else:
+            result = torch.quantized_gru(input,
+                                         batch_sizes,
+                                         hx,
+                                         _all_params,
+                                         self.bias,
+                                         self.num_layers,
+                                         self.dropout,
+                                         self.training,
+                                         self.bidirectional)
+        output = result[0]
+        hidden = result[1]
+
+        return output, hidden
+
+
+    @torch.jit.export
+    def forward_tensor(
+        self, input: Tensor, hx: Optional[Tensor] = None
+    ) -> Tuple[Tensor, Tensor]:
+        batch_sizes = None
+        max_batch_size = input.size(0) if self.batch_first else input.size(1)
+        sorted_indices = None
+        unsorted_indices = None
+
+        output, hidden = self.forward_impl(
+            input, hx, batch_sizes, max_batch_size, sorted_indices)
+
+        return output, self.permute_hidden(hidden, unsorted_indices)
+
+    @torch.jit.export
+    def forward_packed(
+        self, input: PackedSequence, hx: Optional[Tensor] = None
+    ) -> Tuple[PackedSequence, Tensor]:
+        input_, batch_sizes, sorted_indices, unsorted_indices = input
+        max_batch_size = int(batch_sizes[0])
+        output_, hidden = self.forward_impl(
+            input_, hx, batch_sizes, max_batch_size, sorted_indices
+        )
+
+        output = PackedSequence(output_, batch_sizes,
+                                sorted_indices, unsorted_indices)
+        return output, self.permute_hidden(hidden, unsorted_indices)
+
+    def permute_hidden(
+        self, hx: Tensor, permutation: Optional[Tensor]
+    ) -> Tensor:
+        if permutation is None:
+            return hx
+        return _apply_permutation(hx, permutation)
+
+    @torch.jit.ignore
+    def forward(self, input, hx=None):
+        if isinstance(input, PackedSequence):
+            return self.forward_packed(input, hx)
+        else:
+            return self.forward_tensor(input, hx)
+
+    @classmethod
+    def from_float(cls, mod):
+        return super().from_float(mod)
+
+    @classmethod
+    def from_reference(cls, ref_mod):
+        assert hasattr(ref_mod, "weight_ih_l0_dtype"), "We are assuming weight_ih_l0 "
+        "exists in LSTM, may need to relax the assumption to support the use case"
+        qmod = cls(
+            ref_mod.input_size,
+            ref_mod.hidden_size,
+            ref_mod.num_layers,
+            ref_mod.bias,
+            ref_mod.batch_first,
+            ref_mod.dropout,
+            ref_mod.bidirectional,
+            # assuming there is layer 0, which should be OK
+            ref_mod.weight_ih_l0_dtype,
+        )
+        qmod.set_weight_bias(ref_mod.get_quantized_weight_bias_dict())
+        return qmod
+
+class RNNCellBase(torch.nn.Module):
+    # _FLOAT_MODULE = nn.CellRNNBase
+    __constants__ = ['input_size', 'hidden_size', 'bias']
+
+    def __init__(self, input_size, hidden_size, bias=True, num_chunks=4, dtype=torch.qint8):
+        super().__init__()
+        self.input_size = input_size
+        self.hidden_size = hidden_size
+        self.bias = bias
+        self.weight_dtype = dtype
+        if bias:
+            self.bias_ih = torch.randn(num_chunks * hidden_size).to(dtype=torch.float)
+            self.bias_hh = torch.randn(num_chunks * hidden_size).to(dtype=torch.float)
+        else:
+            self.register_parameter('bias_ih', None)
+            self.register_parameter('bias_hh', None)
+
+        weight_ih = torch.randn(num_chunks * hidden_size, input_size).to(torch.float)
+        weight_hh = torch.randn(num_chunks * hidden_size, hidden_size).to(torch.float)
+        if dtype == torch.qint8:
+            weight_ih = torch.quantize_per_tensor(weight_ih, scale=1, zero_point=0, dtype=torch.qint8)
+            weight_hh = torch.quantize_per_tensor(weight_hh, scale=1, zero_point=0, dtype=torch.qint8)
+
+        if dtype == torch.qint8:
+            # for each layer, for each direction we need to quantize and pack
+            # weights and pack parameters in this order:
+            #
+            #   w_ih, w_hh
+            packed_weight_ih = \
+                torch.ops.quantized.linear_prepack(weight_ih, self.bias_ih)
+            packed_weight_hh = \
+                torch.ops.quantized.linear_prepack(weight_hh, self.bias_hh)
+        else:
+            # for each layer, for each direction we need to quantize and pack
+            # weights and pack parameters in this order:
+            #
+            #   packed_ih, packed_hh, b_ih, b_hh
+            packed_weight_ih = torch.ops.quantized.linear_prepack_fp16(
+                weight_ih, self.bias_ih)
+            packed_weight_hh = torch.ops.quantized.linear_prepack_fp16(
+                weight_hh, self.bias_hh)
+
+        self._packed_weight_ih = packed_weight_ih
+        self._packed_weight_hh = packed_weight_hh
+
+    def _get_name(self):
+        return 'DynamicQuantizedRNNBase'
+
+    def extra_repr(self):
+        s = '{input_size}, {hidden_size}'
+        if 'bias' in self.__dict__ and self.bias is not True:
+            s += ', bias={bias}'
+        if 'nonlinearity' in self.__dict__ and self.nonlinearity != "tanh":
+            s += ', nonlinearity={nonlinearity}'
+        return s.format(**self.__dict__)
+
+    def check_forward_input(self, input):
+        if input.size(1) != self.input_size:
+            raise RuntimeError(
+                f"input has inconsistent input_size: got {input.size(1)}, expected {self.input_size}")
+
+    def check_forward_hidden(self, input: Tensor, hx: Tensor, hidden_label: str = '') -> None:
+        if input.size(0) != hx.size(0):
+            raise RuntimeError(
+                f"Input batch size {input.size(0)} doesn't match hidden{hidden_label} batch size {hx.size(0)}")
+
+        if hx.size(1) != self.hidden_size:
+            raise RuntimeError(
+                f"hidden{hidden_label} has inconsistent hidden_size: got {hx.size(1)}, expected {self.hidden_size}")
+
+    @classmethod
+    def from_float(cls, mod):
+        assert type(mod) in {torch.nn.LSTMCell,
+                             torch.nn.GRUCell,
+                             torch.nn.RNNCell}, 'nn.quantized.dynamic.RNNCellBase.from_float \
+                                 only works for nn.LSTMCell, nn.GRUCell and nn.RNNCell'
+        assert hasattr(
+            mod, 'qconfig'), 'Input float module must have qconfig defined'
+
+        if mod.qconfig is not None and mod.qconfig.weight is not None:
+            weight_observer_method = mod.qconfig.weight
+        else:
+            # We have the circular import issues if we import the qconfig in the beginning of this file:
+            # https://github.com/pytorch/pytorch/pull/24231. The current workaround is to postpone the
+            # import until we need it.
+            from torch.ao.quantization.qconfig import default_dynamic_qconfig
+            weight_observer_method = default_dynamic_qconfig.weight
+
+        dtype = weight_observer_method().dtype
+        supported_scalar_types = [torch.qint8, torch.float16]
+        if dtype not in supported_scalar_types:
+            raise RuntimeError(f'Unsupported dtype for dynamic RNN quantization: {dtype}')
+
+        qRNNCellBase: Union[LSTMCell, GRUCell, RNNCell]
+
+        if type(mod) == torch.nn.LSTMCell:
+            qRNNCellBase = LSTMCell(mod.input_size, mod.hidden_size, bias=mod.bias, dtype=dtype)
+        elif type(mod) == torch.nn.GRUCell:
+            qRNNCellBase = GRUCell(mod.input_size, mod.hidden_size, bias=mod.bias, dtype=dtype)
+        elif type(mod) == torch.nn.RNNCell:
+            qRNNCellBase = RNNCell(mod.input_size, mod.hidden_size, bias=mod.bias, nonlinearity=mod.nonlinearity, dtype=dtype)
+        else:
+            raise NotImplementedError('Only LSTMCell, GRUCell and RNNCell \
+            are supported for QuantizedRNN for now')
+
+        assert mod.bias
+
+        def _observe_and_quantize_weight(weight):
+            if dtype == torch.qint8:
+                weight_observer = weight_observer_method()
+                weight_observer(weight)
+                qweight = _quantize_weight(weight.float(), weight_observer)
+                return qweight
+            else:
+                return weight.float()
+
+        qRNNCellBase._packed_weight_ih = pack_weight_bias(_observe_and_quantize_weight(mod.weight_ih), mod.bias_ih, dtype)
+        qRNNCellBase._packed_weight_hh = pack_weight_bias(_observe_and_quantize_weight(mod.weight_hh), mod.bias_hh, dtype)
+        return qRNNCellBase
+
+    @classmethod
+    def from_reference(cls, ref_mod):
+        assert hasattr(ref_mod, "weight_ih_dtype"), "We are assuming weight_ih "
+        "exists in reference module, may need to relax the assumption to support the use case"
+        if hasattr(ref_mod, "nonlinearity"):
+            qmod = cls(
+                ref_mod.input_size,
+                ref_mod.hidden_size,
+                ref_mod.bias,
+                ref_mod.nonlinearity,
+                dtype=ref_mod.weight_ih_dtype
+            )
+        else:
+            qmod = cls(
+                ref_mod.input_size,
+                ref_mod.hidden_size,
+                ref_mod.bias,
+                dtype=ref_mod.weight_ih_dtype
+            )
+        weight_bias_dict = {
+            "weight": {
+                "weight_ih": ref_mod.get_quantized_weight_ih(),
+                "weight_hh": ref_mod.get_quantized_weight_hh(),
+            },
+            "bias": {
+                "bias_ih": ref_mod.bias_ih,
+                "bias_hh": ref_mod.bias_hh,
+            }
+        }
+        qmod.set_weight_bias(weight_bias_dict)
+        return qmod
+
+    def _weight_bias(self):
+        # Returns a dict of weights and biases
+        weight_bias_dict: Dict[str, Dict] = {'weight' : {}, 'bias' : {}}
+        w1, b1 = self._packed_weight_ih.__getstate__()[0]
+        w2, b2 = self._packed_weight_hh.__getstate__()[0]
+        # TODO: these can be simplified to one level? e.g. using weight_ih as key
+        # directly
+        weight_bias_dict['weight']['weight_ih'] = w1
+        weight_bias_dict['weight']['weight_hh'] = w2
+        weight_bias_dict['bias']['bias_ih'] = b1
+        weight_bias_dict['bias']['bias_hh'] = b2
+        return weight_bias_dict
+
+    def get_weight(self):
+        return self._weight_bias()['weight']
+
+    def get_bias(self):
+        return self._weight_bias()['bias']
+
+    def set_weight_bias(self, weight_bias_dict):
+        # TODO: these can be simplified to one level? e.g. using weight_ih as key
+        # directly
+        self._packed_weight_ih = pack_weight_bias(
+            weight_bias_dict["weight"]["weight_ih"],
+            weight_bias_dict["bias"]["bias_ih"],
+            self.weight_dtype)
+        self._packed_weight_hh = pack_weight_bias(
+            weight_bias_dict["weight"]["weight_hh"],
+            weight_bias_dict["bias"]["bias_hh"],
+            self.weight_dtype)
+
+    def _save_to_state_dict(self, destination, prefix, keep_vars):
+        super()._save_to_state_dict(destination, prefix, keep_vars)
+        destination[prefix + '_packed_weight_ih'] = self._packed_weight_ih
+        destination[prefix + '_packed_weight_hh'] = self._packed_weight_hh
+
+    def _load_from_state_dict(self, state_dict, prefix, local_metadata, strict,
+                              missing_keys, unexpected_keys, error_msgs):
+        self._packed_weight_ih = state_dict.pop(prefix + '_packed_weight_ih')
+        self._packed_weight_hh = state_dict.pop(prefix + '_packed_weight_hh')
+        super()._load_from_state_dict(state_dict, prefix, local_metadata, False,
+                                      missing_keys, unexpected_keys, error_msgs)
+
+
+class RNNCell(RNNCellBase):
+    r"""An Elman RNN cell with tanh or ReLU non-linearity.
+    A dynamic quantized RNNCell module with floating point tensor as inputs and outputs.
+    Weights are quantized to 8 bits. We adopt the same interface as `torch.nn.RNNCell`,
+    please see https://pytorch.org/docs/stable/nn.html#torch.nn.RNNCell for documentation.
+
+    Examples::
+
+        >>> # xdoctest: +SKIP
+        >>> rnn = nn.RNNCell(10, 20)
+        >>> input = torch.randn(6, 3, 10)
+        >>> hx = torch.randn(3, 20)
+        >>> output = []
+        >>> for i in range(6):
+        ...     hx = rnn(input[i], hx)
+        ...     output.append(hx)
+    """
+    __constants__ = ['input_size', 'hidden_size', 'bias', 'nonlinearity']
+
+    def __init__(self, input_size, hidden_size, bias=True, nonlinearity="tanh", dtype=torch.qint8):
+        super().__init__(input_size, hidden_size, bias, num_chunks=1, dtype=dtype)
+        self.nonlinearity = nonlinearity
+
+    def _get_name(self):
+        return 'DynamicQuantizedRNNCell'
+
+    def forward(self, input: Tensor, hx: Optional[Tensor] = None) -> Tensor:
+        self.check_forward_input(input)
+        if hx is None:
+            hx = torch.zeros(input.size(0), self.hidden_size, dtype=input.dtype, device=input.device)
+        self.check_forward_hidden(input, hx, '')
+        if self.nonlinearity == "tanh":
+            ret = torch.ops.quantized.quantized_rnn_tanh_cell_dynamic(
+                input, hx,
+                self._packed_weight_ih, self._packed_weight_hh,
+                self.bias_ih, self.bias_hh)
+        elif self.nonlinearity == "relu":
+            ret = torch.ops.quantized.quantized_rnn_relu_cell_dynamic(
+                input, hx,
+                self._packed_weight_ih, self._packed_weight_hh,
+                self.bias_ih, self.bias_hh)
+        else:
+            ret = input  # TODO: remove when jit supports exception flow
+            raise RuntimeError(
+                f"Unknown nonlinearity: {self.nonlinearity}")
+        return ret
+
+    @classmethod
+    def from_float(cls, mod):
+        return super().from_float(mod)
+
+
+class LSTMCell(RNNCellBase):
+    r"""A long short-term memory (LSTM) cell.
+
+    A dynamic quantized LSTMCell module with floating point tensor as inputs and outputs.
+    Weights are quantized to 8 bits. We adopt the same interface as `torch.nn.LSTMCell`,
+    please see https://pytorch.org/docs/stable/nn.html#torch.nn.LSTMCell for documentation.
+
+    Examples::
+
+        >>> # xdoctest: +SKIP
+        >>> rnn = nn.LSTMCell(10, 20)
+        >>> input = torch.randn(6, 3, 10)
+        >>> hx = torch.randn(3, 20)
+        >>> cx = torch.randn(3, 20)
+        >>> output = []
+        >>> for i in range(6):
+        ...     hx, cx = rnn(input[i], (hx, cx))
+        ...     output.append(hx)
+    """
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, num_chunks=4, **kwargs)  # type: ignore[misc]
+
+    def _get_name(self):
+        return 'DynamicQuantizedLSTMCell'
+
+    def forward(self, input: Tensor, hx: Optional[Tuple[Tensor, Tensor]] = None) -> Tuple[Tensor, Tensor]:
+        self.check_forward_input(input)
+        if hx is None:
+            zeros = torch.zeros(input.size(0), self.hidden_size, dtype=input.dtype, device=input.device)
+            hx = (zeros, zeros)
+        self.check_forward_hidden(input, hx[0], '[0]')
+        self.check_forward_hidden(input, hx[1], '[1]')
+        return torch.ops.quantized.quantized_lstm_cell_dynamic(
+            input, hx,
+            self._packed_weight_ih, self._packed_weight_hh,
+            self.bias_ih, self.bias_hh)
+
+    @classmethod
+    def from_float(cls, mod):
+        return super().from_float(mod)
+
+
+class GRUCell(RNNCellBase):
+    r"""A gated recurrent unit (GRU) cell
+
+    A dynamic quantized GRUCell module with floating point tensor as inputs and outputs.
+    Weights are quantized to 8 bits. We adopt the same interface as `torch.nn.GRUCell`,
+    please see https://pytorch.org/docs/stable/nn.html#torch.nn.GRUCell for documentation.
+
+    Examples::
+
+        >>> # xdoctest: +SKIP
+        >>> rnn = nn.GRUCell(10, 20)
+        >>> input = torch.randn(6, 3, 10)
+        >>> hx = torch.randn(3, 20)
+        >>> output = []
+        >>> for i in range(6):
+        ...     hx = rnn(input[i], hx)
+        ...     output.append(hx)
+    """
+
+    def __init__(self, input_size, hidden_size, bias=True, dtype=torch.qint8):
+        super().__init__(input_size, hidden_size, bias, num_chunks=3, dtype=dtype)
+
+    def _get_name(self):
+        return 'DynamicQuantizedGRUCell'
+
+    def forward(self, input: Tensor, hx: Optional[Tensor] = None) -> Tensor:
+        self.check_forward_input(input)
+        if hx is None:
+            hx = torch.zeros(input.size(0), self.hidden_size, dtype=input.dtype, device=input.device)
+        self.check_forward_hidden(input, hx, '')
+        return torch.ops.quantized.quantized_gru_cell_dynamic(
+            input, hx,
+            self._packed_weight_ih, self._packed_weight_hh,
+            self.bias_ih, self.bias_hh,
+        )
+
+    @classmethod
+    def from_float(cls, mod):
+        return super().from_float(mod)