whispercpp 1.3.0 → 1.3.2

data/ext/sources/ggml/src/ggml-cuda/vendors/cuda.h

@@ -0,0 +1,15 @@
+#pragma once
+
+#include <cuda_runtime.h>
+#include <cuda.h>
+#include <cublas_v2.h>
+#include <cuda_bf16.h>
+#include <cuda_fp16.h>
+
+#if CUDART_VERSION < 11020
+#define CU_DEVICE_ATTRIBUTE_VIRTUAL_MEMORY_MANAGEMENT_SUPPORTED CU_DEVICE_ATTRIBUTE_VIRTUAL_ADDRESS_MANAGEMENT_SUPPORTED
+#define CUBLAS_TF32_TENSOR_OP_MATH CUBLAS_TENSOR_OP_MATH
+#define CUBLAS_COMPUTE_16F CUDA_R_16F
+#define CUBLAS_COMPUTE_32F CUDA_R_32F
+#define cublasComputeType_t cudaDataType_t
+#endif // CUDART_VERSION < 11020

data/ext/sources/ggml/src/ggml-cuda/vendors/hip.h

@@ -0,0 +1,243 @@
+#pragma once
+
+#define HIP_ENABLE_WARP_SYNC_BUILTINS 1
+#include <hip/hip_runtime.h>
+#include <hipblas/hipblas.h>
+#include <hip/hip_fp16.h>
+#include <hip/hip_bfloat16.h>
+#ifdef __HIP_PLATFORM_AMD__
+// for rocblas_initialize()
+#include "rocblas/rocblas.h"
+#endif // __HIP_PLATFORM_AMD__
+
+#define CUBLAS_COMPUTE_16F HIPBLAS_R_16F
+#define CUBLAS_COMPUTE_32F HIPBLAS_R_32F
+#define CUBLAS_COMPUTE_32F_FAST_16F HIPBLAS_R_32F
+#define CUBLAS_GEMM_DEFAULT HIPBLAS_GEMM_DEFAULT
+#define CUBLAS_GEMM_DEFAULT_TENSOR_OP HIPBLAS_GEMM_DEFAULT
+#define CUBLAS_OP_N HIPBLAS_OP_N
+#define CUBLAS_OP_T HIPBLAS_OP_T
+#define CUBLAS_STATUS_SUCCESS HIPBLAS_STATUS_SUCCESS
+#define CUBLAS_TF32_TENSOR_OP_MATH 0
+#define CUDA_R_16F  HIPBLAS_R_16F
+#define CUDA_R_16BF HIPBLAS_R_16B
+#define CUDA_R_32F  HIPBLAS_R_32F
+#define CU_DEVICE_ATTRIBUTE_VIRTUAL_MEMORY_MANAGEMENT_SUPPORTED hipDeviceAttributeVirtualMemoryManagementSupported
+#define CU_MEM_ALLOC_GRANULARITY_RECOMMENDED hipMemAllocationGranularityRecommended
+#define CU_MEM_ALLOCATION_TYPE_PINNED hipMemAllocationTypePinned
+#define CU_MEM_LOCATION_TYPE_DEVICE hipMemLocationTypeDevice
+#define CU_MEM_ACCESS_FLAGS_PROT_READWRITE hipMemAccessFlagsProtReadWrite
+#define CU_CHECK(fn) {hipError_t err = fn; if(err != hipSuccess) { GGML_ABORT("HipVMM Failure: %s\n", hipGetErrorString(err)); }}
+#define __shfl_sync(mask, var, laneMask, width) __shfl(var, laneMask, width)
+#define __shfl_xor_sync(mask, var, laneMask, width) __shfl_xor(var, laneMask, width)
+#define cublasComputeType_t hipblasDatatype_t //deprecated, new hipblasComputeType_t not in 5.6
+#define cublasCreate hipblasCreate
+#define cublasDestroy hipblasDestroy
+#define cublasGemmEx hipblasGemmEx
+#define cublasGemmBatchedEx hipblasGemmBatchedEx
+#define cublasGemmStridedBatchedEx hipblasGemmStridedBatchedEx
+#define cublasHandle_t hipblasHandle_t
+#define cublasSetMathMode(handle, mode) CUBLAS_STATUS_SUCCESS
+#define cublasSetStream hipblasSetStream
+#define cublasSgemm hipblasSgemm
+#define cublasStatus_t hipblasStatus_t
+#define cublasOperation_t hipblasOperation_t
+#define cudaDataType_t hipblasDatatype_t //deprecated, new hipblasDatatype not in 5.6
+#define cudaDeviceCanAccessPeer hipDeviceCanAccessPeer
+#define cudaDeviceDisablePeerAccess hipDeviceDisablePeerAccess
+#define cudaDeviceEnablePeerAccess hipDeviceEnablePeerAccess
+#define cudaDeviceProp hipDeviceProp_t
+#define cudaDeviceSynchronize hipDeviceSynchronize
+#define cudaError_t hipError_t
+#define cudaErrorPeerAccessAlreadyEnabled hipErrorPeerAccessAlreadyEnabled
+#define cudaErrorPeerAccessNotEnabled hipErrorPeerAccessNotEnabled
+#define cudaEventCreateWithFlags hipEventCreateWithFlags
+#define cudaEventDisableTiming hipEventDisableTiming
+#define cudaEventRecord hipEventRecord
+#define cudaEventSynchronize hipEventSynchronize
+#define cudaEvent_t hipEvent_t
+#define cudaEventDestroy hipEventDestroy
+#define cudaFree hipFree
+#define cudaFreeHost hipHostFree
+#define cudaGetDevice hipGetDevice
+#define cudaGetDeviceCount hipGetDeviceCount
+#define cudaGetDeviceProperties hipGetDeviceProperties
+#define cudaGetErrorString hipGetErrorString
+#define cudaGetLastError hipGetLastError
+#define cudaHostRegister hipHostRegister
+#define cudaHostRegisterPortable hipHostRegisterPortable
+#define cudaHostRegisterReadOnly hipHostRegisterReadOnly
+#define cudaHostUnregister hipHostUnregister
+#define cudaLaunchHostFunc hipLaunchHostFunc
+#define cudaMalloc hipMalloc
+#define cudaMallocHost(ptr, size) hipHostMalloc(ptr, size, hipHostMallocDefault)
+#define cudaMallocManaged hipMallocManaged
+#define cudaMemAdvise hipMemAdvise
+#define cudaMemcpy hipMemcpy
+#define cudaMemcpyAsync hipMemcpyAsync
+#define cudaMemcpyPeerAsync hipMemcpyPeerAsync
+#define cudaMemcpy2DAsync hipMemcpy2DAsync
+#define cudaMemcpyDeviceToDevice hipMemcpyDeviceToDevice
+#define cudaMemcpyDeviceToHost hipMemcpyDeviceToHost
+#define cudaMemcpyHostToDevice hipMemcpyHostToDevice
+#define cudaMemcpyKind hipMemcpyKind
+#define cudaMemset hipMemset
+#define cudaMemsetAsync hipMemsetAsync
+#define cudaMemGetInfo hipMemGetInfo
+#define cudaOccupancyMaxPotentialBlockSize hipOccupancyMaxPotentialBlockSize
+#define cudaSetDevice hipSetDevice
+#define cuDeviceGet hipDeviceGet
+#define CUdevice hipDevice_t
+#define CUdeviceptr hipDeviceptr_t
+#define cuMemUnmap hipMemUnmap
+#define CUmemAccessDesc hipMemAccessDesc
+#define cuMemAddressFree hipMemAddressFree
+#define cuMemRelease hipMemRelease
+#define CUmemGenericAllocationHandle hipMemGenericAllocationHandle_t
+#define cuMemCreate hipMemCreate
+#define cuMemAddressReserve hipMemAddressReserve
+#define cuMemMap hipMemMap
+#define cuMemSetAccess hipMemSetAccess
+#define cuMemGetAllocationGranularity hipMemGetAllocationGranularity
+#define CUmemAllocationProp hipMemAllocationProp
+#define cuDeviceGetAttribute hipDeviceGetAttribute
+#define cudaStreamCreateWithFlags hipStreamCreateWithFlags
+#define cudaStreamDestroy hipStreamDestroy
+#define cudaStreamFireAndForget hipStreamFireAndForget
+#define cudaStreamNonBlocking hipStreamNonBlocking
+#define cudaStreamPerThread hipStreamPerThread
+#define cudaStreamSynchronize hipStreamSynchronize
+#define cudaStreamWaitEvent(stream, event, flags) hipStreamWaitEvent(stream, event, flags)
+#define cudaGraphExec_t hipGraphExec_t
+#define cudaGraphNode_t hipGraphNode_t
+#define cudaKernelNodeParams hipKernelNodeParams
+#define cudaKernelNodeParams hipKernelNodeParams
+#define cudaGraphExecDestroy hipGraphExecDestroy
+#define cudaGraphLaunch hipGraphLaunch
+#define cudaErrorGraphExecUpdateFailure hipErrorGraphExecUpdateFailure
+#define cudaGraphExecUpdateResult hipGraphExecUpdateResult
+#define cudaGraphNodeType hipGraphNodeType
+#define cudaGraphNodeTypeKernel hipGraphNodeTypeKernel
+#define cudaGraphInstantiate hipGraphInstantiate
+#define cudaStreamEndCapture hipStreamEndCapture
+#define cudaGraphDestroy hipGraphDestroy
+#define cudaGraphKernelNodeSetParams hipGraphKernelNodeSetParams
+#define cudaErrorInvalidDeviceFunction hipErrorInvalidDeviceFunction
+#define cudaGraphKernelNodeGetParams hipGraphKernelNodeGetParams
+#define cudaGraphNodeGetType hipGraphNodeGetType
+#define cudaGraphGetNodes hipGraphGetNodes
+#define cudaGraphExecUpdate hipGraphExecUpdate
+#define cudaStreamCaptureModeRelaxed hipStreamCaptureModeRelaxed
+#define cudaStreamBeginCapture hipStreamBeginCapture
+#define cudaGraph_t hipGraph_t
+#define cudaStream_t hipStream_t
+#define cudaSuccess hipSuccess
+#define cudaOccupancyMaxActiveBlocksPerMultiprocessor hipOccupancyMaxActiveBlocksPerMultiprocessor
+#define __trap() do { abort(); __builtin_unreachable(); } while(0)
+#define CUBLAS_STATUS_SUCCESS HIPBLAS_STATUS_SUCCESS
+#define CUBLAS_STATUS_NOT_INITIALIZED HIPBLAS_STATUS_NOT_INITIALIZED
+#define CUBLAS_STATUS_ALLOC_FAILED HIPBLAS_STATUS_ALLOC_FAILED
+#define CUBLAS_STATUS_INVALID_VALUE HIPBLAS_STATUS_INVALID_VALUE
+#define CUBLAS_STATUS_ARCH_MISMATCH HIPBLAS_STATUS_ARCH_MISMATCH
+#define CUBLAS_STATUS_MAPPING_ERROR HIPBLAS_STATUS_MAPPING_ERROR
+#define CUBLAS_STATUS_EXECUTION_FAILED HIPBLAS_STATUS_EXECUTION_FAILED
+#define CUBLAS_STATUS_INTERNAL_ERROR HIPBLAS_STATUS_INTERNAL_ERROR
+#define CUBLAS_STATUS_NOT_SUPPORTED HIPBLAS_STATUS_NOT_SUPPORTED
+
+#define __CUDA_ARCH__ 1300
+
+#if defined(__gfx803__) || defined(__gfx900__) || defined(__gfx906__)
+#define GCN
+#endif
+
+#if defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx942__)
+#define CDNA
+#endif
+
+#if defined(__GFX12__)
+#define RDNA4
+#endif
+
+#if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx1103__) || \
+    defined(__gfx1150__) || defined(__gfx1151__)
+#define RDNA3
+#endif
+
+#if defined(__gfx1030__) || defined(__gfx1031__) || defined(__gfx1032__) || defined(__gfx1033__) || \
+    defined(__gfx1034__) || defined(__gfx1035__) || defined(__gfx1036__) || defined(__gfx1037__)
+#define RDNA2
+#endif
+
+#if defined(__gfx1010__) || defined(__gfx1012__)
+#define RDNA1
+#endif
+
+#ifndef __has_builtin
+    #define __has_builtin(x) 0
+#endif
+
+typedef hip_bfloat16 nv_bfloat16;
+
+typedef int8_t int8x4_t __attribute__((ext_vector_type(4)));
+typedef uint8_t uint8x4_t __attribute__((ext_vector_type(4)));
+static __device__ __forceinline__ int __vsubss4(const int a, const int b) {
+    const int8x4_t va = reinterpret_cast<const int8x4_t&>(a);
+    const int8x4_t vb = reinterpret_cast<const int8x4_t&>(b);
+#if __has_builtin(__builtin_elementwise_sub_sat)
+    const int8x4_t c = __builtin_elementwise_sub_sat(va, vb);
+    return reinterpret_cast<const int &>(c);
+#else
+    int8x4_t c;
+    int16_t tmp;
+#pragma unroll
+    for (int i = 0; i < 4; i++) {
+        tmp = va[i] - vb[i];
+        if(tmp > std::numeric_limits<int8_t>::max()) tmp = std::numeric_limits<int8_t>::max();
+        if(tmp < std::numeric_limits<int8_t>::min()) tmp = std::numeric_limits<int8_t>::min();
+        c[i] = tmp;
+    }
+    return reinterpret_cast<int &>(c);
+#endif // __has_builtin(__builtin_elementwise_sub_sat)
+}
+
+static __device__ __forceinline__ int __vsub4(const int a, const int b) {
+    return __vsubss4(a, b);
+}
+
+static __device__ __forceinline__ unsigned int __vcmpeq4(unsigned int a, unsigned int b) {
+    const uint8x4_t& va = reinterpret_cast<const uint8x4_t&>(a);
+    const uint8x4_t& vb = reinterpret_cast<const uint8x4_t&>(b);
+    unsigned int c;
+    uint8x4_t& vc = reinterpret_cast<uint8x4_t&>(c);
+#pragma unroll
+    for (int i = 0; i < 4; ++i) {
+        vc[i] = va[i] == vb[i] ? 0xff : 0x00;
+    }
+    return c;
+}
+
+static __device__ __forceinline__ unsigned int __vcmpne4(unsigned int a, unsigned int b) {
+    const uint8x4_t& va = reinterpret_cast<const uint8x4_t&>(a);
+    const uint8x4_t& vb = reinterpret_cast<const uint8x4_t&>(b);
+    unsigned int c;
+    uint8x4_t& vc = reinterpret_cast<uint8x4_t&>(c);
+#pragma unroll
+    for (int i = 0; i < 4; ++i) {
+        vc[i] = va[i] == vb[i] ? 0x00 : 0xff;
+    }
+    return c;
+}
+
+#if defined(__HIP_PLATFORM_AMD__) && HIP_VERSION < 50600000
+// __shfl_xor() for half2 was added in ROCm 5.6
+static __device__ __forceinline__ half2 __shfl_xor(half2 var, int laneMask, int width) {
+    typedef union half2_b32 {
+        half2 val;
+        int   b32;
+    } half2_b32_t;
+    half2_b32_t tmp;
+    tmp.val = var;
+    tmp.b32 = __shfl_xor(tmp.b32, laneMask, width);
+    return tmp.val;
+}
+#endif // defined(__HIP_PLATFORM_AMD__) && HIP_VERSION < 50600000

data/ext/sources/ggml/src/ggml-cuda/vendors/musa.h

@@ -0,0 +1,140 @@
+#pragma once
+
+#include <musa_runtime.h>
+#include <musa.h>
+#include <mublas.h>
+#include <musa_bf16.h>
+#include <musa_fp16.h>
+#define CUBLAS_COMPUTE_16F CUDA_R_16F
+#define CUBLAS_COMPUTE_32F CUDA_R_32F
+#define CUBLAS_COMPUTE_32F_FAST_16F MUBLAS_COMPUTE_32F_FAST_16F
+#define CUBLAS_GEMM_DEFAULT MUBLAS_GEMM_DEFAULT
+#define CUBLAS_GEMM_DEFAULT_TENSOR_OP MUBLAS_GEMM_DEFAULT
+#define CUBLAS_OP_N MUBLAS_OP_N
+#define CUBLAS_OP_T MUBLAS_OP_T
+#define CUBLAS_STATUS_SUCCESS MUBLAS_STATUS_SUCCESS
+#define CUBLAS_TF32_TENSOR_OP_MATH MUBLAS_MATH_MODE_DEFAULT
+#define CUDA_R_16F  MUSA_R_16F
+#define CUDA_R_16BF MUSA_R_16BF
+#define CUDA_R_32F  MUSA_R_32F
+#define cublasComputeType_t cudaDataType_t
+#define cublasCreate mublasCreate
+#define cublasDestroy mublasDestroy
+#define cublasGemmEx mublasGemmEx
+#define cublasGemmBatchedEx mublasGemmBatchedEx
+#define cublasGemmStridedBatchedEx mublasGemmStridedBatchedEx
+#define cublasHandle_t mublasHandle_t
+#define cublasSetMathMode mublasSetMathMode
+#define cublasSetStream mublasSetStream
+#define cublasSgemm mublasSgemm
+#define cublasStatus_t mublasStatus_t
+#define cublasOperation_t mublasOperation_t
+#define cublasGetStatusString mublasStatus_to_string
+#define cudaDataType_t musaDataType_t
+#define cudaDeviceCanAccessPeer musaDeviceCanAccessPeer
+#define cudaDeviceDisablePeerAccess musaDeviceDisablePeerAccess
+#define cudaDeviceEnablePeerAccess musaDeviceEnablePeerAccess
+#define cudaDeviceProp musaDeviceProp
+#define cudaDeviceSynchronize musaDeviceSynchronize
+#define cudaError_t musaError_t
+#define cudaErrorPeerAccessAlreadyEnabled musaErrorPeerAccessAlreadyEnabled
+#define cudaErrorPeerAccessNotEnabled musaErrorPeerAccessNotEnabled
+#define cudaEventCreateWithFlags musaEventCreateWithFlags
+#define cudaEventDisableTiming musaEventDisableTiming
+#define cudaEventRecord musaEventRecord
+#define cudaEventSynchronize musaEventSynchronize
+#define cudaEvent_t musaEvent_t
+#define cudaEventDestroy musaEventDestroy
+#define cudaFree musaFree
+#define cudaFreeHost musaFreeHost
+#define cudaGetDevice musaGetDevice
+#define cudaGetDeviceCount musaGetDeviceCount
+#define cudaGetDeviceProperties musaGetDeviceProperties
+#define cudaGetErrorString musaGetErrorString
+#define cudaGetLastError musaGetLastError
+#define cudaHostRegister musaHostRegister
+#define cudaHostRegisterPortable musaHostRegisterPortable
+#define cudaHostRegisterReadOnly musaHostRegisterReadOnly
+#define cudaHostUnregister musaHostUnregister
+#define cudaLaunchHostFunc musaLaunchHostFunc
+#define cudaMalloc musaMalloc
+#define cudaMallocHost musaMallocHost
+#define cudaMallocManaged musaMallocManaged
+#define cudaMemcpy musaMemcpy
+#define cudaMemcpyAsync musaMemcpyAsync
+#define cudaMemcpyPeerAsync musaMemcpyPeerAsync
+#define cudaMemcpy2DAsync musaMemcpy2DAsync
+#define cudaMemcpyDeviceToDevice musaMemcpyDeviceToDevice
+#define cudaMemcpyDeviceToHost musaMemcpyDeviceToHost
+#define cudaMemcpyHostToDevice musaMemcpyHostToDevice
+#define cudaMemcpyKind musaMemcpyKind
+#define cudaMemset musaMemset
+#define cudaMemsetAsync musaMemsetAsync
+#define cudaMemGetInfo musaMemGetInfo
+#define cudaOccupancyMaxPotentialBlockSize musaOccupancyMaxPotentialBlockSize
+#define cudaSetDevice musaSetDevice
+#define cudaStreamCreateWithFlags musaStreamCreateWithFlags
+#define cudaStreamDestroy musaStreamDestroy
+#define cudaStreamFireAndForget musaStreamFireAndForget
+#define cudaStreamNonBlocking musaStreamNonBlocking
+#define cudaStreamPerThread musaStreamPerThread
+#define cudaStreamSynchronize musaStreamSynchronize
+#define cudaStreamWaitEvent musaStreamWaitEvent
+#define cudaStream_t musaStream_t
+#define cudaSuccess musaSuccess
+
+// Additional mappings for MUSA virtual memory pool
+#define CU_DEVICE_ATTRIBUTE_VIRTUAL_MEMORY_MANAGEMENT_SUPPORTED MU_DEVICE_ATTRIBUTE_VIRTUAL_ADDRESS_MANAGEMENT_SUPPORTED
+#define CU_MEM_ACCESS_FLAGS_PROT_READWRITE MU_MEM_ACCESS_FLAGS_PROT_READWRITE
+#define CU_MEM_ALLOC_GRANULARITY_RECOMMENDED MU_MEM_ALLOC_GRANULARITY_RECOMMENDED
+#define CU_MEM_ALLOCATION_TYPE_PINNED MU_MEM_ALLOCATION_TYPE_PINNED
+#define CU_MEM_LOCATION_TYPE_DEVICE MU_MEM_LOCATION_TYPE_DEVICE
+#define CUdevice MUdevice
+#define CUdeviceptr MUdeviceptr
+#define CUmemAccessDesc MUmemAccessDesc
+#define CUmemAllocationProp MUmemAllocationProp
+#define CUmemGenericAllocationHandle MUmemGenericAllocationHandle
+#define cuDeviceGet muDeviceGet
+#define cuDeviceGetAttribute muDeviceGetAttribute
+#define cuMemAddressFree muMemAddressFree
+#define cuMemAddressReserve muMemAddressReserve
+#define cuMemCreate muMemCreate
+#define cuMemGetAllocationGranularity muMemGetAllocationGranularity
+#define cuMemMap muMemMap
+#define cuMemRelease muMemRelease
+#define cuMemSetAccess muMemSetAccess
+#define cuMemUnmap muMemUnmap
+#define cudaFuncAttributeMaxDynamicSharedMemorySize musaFuncAttributeMaxDynamicSharedMemorySize
+#define cudaFuncSetAttribute musaFuncSetAttribute
+#define cudaMemcpy3DPeerParms musaMemcpy3DPeerParms
+#define make_cudaExtent make_musaExtent
+#define make_cudaPitchedPtr make_musaPitchedPtr
+
+// Additional mappings for MUSA graphs
+#define CUDA_SUCCESS MUSA_SUCCESS
+#define CUresult MUresult
+#define cuGetErrorString muGetErrorString
+#define cudaErrorGraphExecUpdateFailure musaErrorGraphExecUpdateFailure
+#define cudaErrorInvalidDeviceFunction musaErrorInvalidDeviceFunction
+#define cudaGraphDestroy musaGraphDestroy
+#define cudaGraphExecDestroy musaGraphExecDestroy
+#define cudaGraphExec_t musaGraphExec_t
+#define cudaGraphExecUpdate musaGraphExecUpdate
+#define cudaGraphExecUpdateResult musaGraphExecUpdateResult
+#define cudaGraphGetNodes musaGraphGetNodes
+#define cudaGraphInstantiate musaGraphInstantiate
+#define cudaGraphKernelNodeGetParams musaGraphKernelNodeGetParams
+#define cudaGraphKernelNodeSetParams musaGraphKernelNodeSetParams
+#define cudaGraphLaunch musaGraphLaunch
+#define cudaGraphNodeGetType musaGraphNodeGetType
+#define cudaGraphNode_t musaGraphNode_t
+#define cudaGraphNodeType musaGraphNodeType
+#define cudaGraphNodeTypeKernel musaGraphNodeTypeKernel
+#define cudaGraph_t musaGraph_t
+#define cudaKernelNodeParams musaKernelNodeParams
+#define cudaStreamCaptureModeRelaxed musaStreamCaptureModeRelaxed
+#define cudaStreamBeginCapture musaStreamBeginCapture
+#define cudaStreamEndCapture musaStreamEndCapture
+#define cudaOccupancyMaxActiveBlocksPerMultiprocessor musaOccupancyMaxActiveBlocksPerMultiprocessor
+
+typedef mt_bfloat16 nv_bfloat16;

data/ext/sources/ggml/src/ggml-cuda/wkv.cu

@@ -0,0 +1,199 @@
+#include "common.cuh"
+#include "wkv.cuh"
+
+template <int block_size>
+static __global__ void rwkv_wkv_f32(const int B, const int T, const int C, const int H, const float * k, const float * v, const float * r, const float * tf, const float * td, const float * s, float * dst) {
+    const int tid = threadIdx.x;
+    const int bid = blockIdx.x;
+
+    const int head_size = block_size;
+    const int batch_i = bid / H;
+    const int head_i = bid % H;
+    const int state_size = C * head_size;
+    const int n_seq_tokens = T / B;
+
+    float state[head_size];
+    __shared__ float _k[head_size], _r[head_size], _tf[head_size], _td[head_size];
+
+    #pragma unroll
+    for (int i = 0; i < head_size; i++) {
+        state[i] = s[batch_i * state_size + head_i * head_size * head_size + i * head_size + tid];
+    }
+
+    __syncthreads();
+    _tf[tid] = tf[head_i * head_size + tid];
+    __syncthreads();
+
+    for (int t = batch_i * n_seq_tokens * C + head_i * head_size + tid; t < (batch_i + 1) * n_seq_tokens * C + head_i * head_size + tid; t += C) {
+        __syncthreads();
+        _k[tid] = k[t];
+        _r[tid] = r[t];
+        _td[tid] = td[t];
+        __syncthreads();
+
+        const float _v = v[t];
+        float y = 0;
+        for (int j = 0; j < head_size; j += 4) {
+            const float4& k = (float4&)(_k[j]);
+            const float4& r = (float4&)(_r[j]);
+            const float4& tf = (float4&)(_tf[j]);
+            const float4& td = (float4&)(_td[j]);
+            float4& s = (float4&)(state[j]);
+            float4 kv;
+
+            kv.x = k.x * _v;
+            kv.y = k.y * _v;
+            kv.z = k.z * _v;
+            kv.w = k.w * _v;
+
+            y += r.x * (tf.x * kv.x + s.x);
+            y += r.y * (tf.y * kv.y + s.y);
+            y += r.z * (tf.z * kv.z + s.z);
+            y += r.w * (tf.w * kv.w + s.w);
+
+            s.x = s.x * td.x + kv.x;
+            s.y = s.y * td.y + kv.y;
+            s.z = s.z * td.z + kv.z;
+            s.w = s.w * td.w + kv.w;
+        }
+        dst[t] = y;
+    }
+
+    #pragma unroll
+    for (int i = 0; i < head_size; i++) {
+        dst[T * C + batch_i * state_size + head_i * head_size * head_size + i * head_size + tid] = state[i];
+    }
+}
+
+template <int block_size>
+static __global__ void rwkv_wkv7_f32(const int B, const int T, const int C, const int H, const float * r, const float * w, const float * k, const float * v, const float * a, const float * b, const float * s, float * dst) {
+    const int tid = threadIdx.x;
+    const int bid = blockIdx.x;
+
+    const int head_size = block_size;
+    const int batch_i = bid / H;
+    const int head_i = bid % H;
+    const int state_size = C * head_size;
+    const int n_seq_tokens = T / B;
+
+    float state[head_size];
+    __shared__ float _r[head_size], _w[head_size], _k[head_size], _a[head_size], _b[head_size];
+
+#ifndef GGML_USE_MUSA
+    #pragma unroll
+#endif
+    for (int i = 0; i < head_size; i++) {
+        state[i] = s[batch_i * state_size + head_i * head_size * head_size + tid * head_size + i];
+    }
+
+    for (int t = batch_i * n_seq_tokens * C + head_i * head_size + tid; t < (batch_i + 1) * n_seq_tokens * C + head_i * head_size + tid; t += C) {
+        __syncthreads();
+        _r[tid] = r[t];
+        _w[tid] = w[t];
+        _k[tid] = k[t];
+        _a[tid] = a[t];
+        _b[tid] = b[t];
+        __syncthreads();
+
+        float sa = 0;
+        #pragma unroll
+        for (int j = 0; j < head_size; j += 4)
+        {
+            const float4& a = (float4&)(_a[j]);
+            const float4& s = (float4&)(state[j]);
+            sa += a.x * s.x;
+            sa += a.y * s.y;
+            sa += a.z * s.z;
+            sa += a.w * s.w;
+        }
+
+        const float _v = v[t];
+        float y = 0;
+        for (int j = 0; j < head_size; j += 4) {
+            const float4& r = (float4&)(_r[j]);
+            const float4& w = (float4&)(_w[j]);
+            const float4& k = (float4&)(_k[j]);
+            const float4& b = (float4&)(_b[j]);
+            float4& s = (float4&)(state[j]);
+            float4 kv;
+
+            kv.x = k.x * _v;
+            kv.y = k.y * _v;
+            kv.z = k.z * _v;
+            kv.w = k.w * _v;
+
+            s.x = s.x * w.x + kv.x + sa * b.x;
+            s.y = s.y * w.y + kv.y + sa * b.y;
+            s.z = s.z * w.z + kv.z + sa * b.z;
+            s.w = s.w * w.w + kv.w + sa * b.w;
+
+            y += s.x * r.x;
+            y += s.y * r.y;
+            y += s.z * r.z;
+            y += s.w * r.w;
+        }
+        dst[t] = y;
+    }
+
+    #pragma unroll
+    for (int i = 0; i < head_size; i++) {
+        dst[T * C + batch_i * state_size + head_i * head_size * head_size + tid * head_size + i] = state[i];
+    }
+}
+
+void ggml_cuda_op_rwkv_wkv6(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const float * k_d  = (const float *)dst->src[0]->data;
+    const float * v_d  = (const float *)dst->src[1]->data;
+    const float * r_d  = (const float *)dst->src[2]->data;
+    const float * tf_d = (const float *)dst->src[3]->data;
+    const float * td_d = (const float *)dst->src[4]->data;
+    const float * s_d  = (const float *)dst->src[5]->data;
+
+    const int64_t B = dst->src[5]->ne[1];
+    const int64_t T = dst->src[0]->ne[2];
+    const int64_t C = dst->ne[0];
+    const int64_t H = dst->src[0]->ne[1];
+
+    float * dst_d = (float *)dst->data;
+
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(dst->src[5]->type == GGML_TYPE_F32);
+    GGML_ASSERT(C % H == 0);
+    GGML_ASSERT(C / H == CUDA_WKV_BLOCK_SIZE || C / H == CUDA_WKV_BLOCK_SIZE * 2);
+
+    if (C / H == CUDA_WKV_BLOCK_SIZE) {
+        rwkv_wkv_f32<CUDA_WKV_BLOCK_SIZE><<<B * H, C / H, 0, stream>>>(B, T, C, H, k_d, v_d, r_d, tf_d, td_d, s_d, dst_d);
+    } else {
+        rwkv_wkv_f32<CUDA_WKV_BLOCK_SIZE * 2><<<B * H, C / H, 0, stream>>>(B, T, C, H, k_d, v_d, r_d, tf_d, td_d, s_d, dst_d);
+    }
+}
+
+void ggml_cuda_op_rwkv_wkv7(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const float * r_d = (const float *)dst->src[0]->data;
+    const float * w_d = (const float *)dst->src[1]->data;
+    const float * k_d = (const float *)dst->src[2]->data;
+    const float * v_d = (const float *)dst->src[3]->data;
+    const float * a_d = (const float *)dst->src[4]->data;
+    const float * b_d = (const float *)dst->src[5]->data;
+    const float * s_d = (const float *)dst->src[6]->data;
+
+    const int64_t B = dst->src[6]->ne[1];
+    const int64_t T = dst->src[0]->ne[2];
+    const int64_t C = dst->ne[0];
+    const int64_t H = dst->src[0]->ne[1];
+
+    float * dst_d = (float *)dst->data;
+
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(dst->src[6]->type == GGML_TYPE_F32);
+    GGML_ASSERT(C % H == 0);
+    GGML_ASSERT(C / H == CUDA_WKV_BLOCK_SIZE || C / H == CUDA_WKV_BLOCK_SIZE * 2);
+
+    if (C / H == CUDA_WKV_BLOCK_SIZE) {
+        rwkv_wkv7_f32<CUDA_WKV_BLOCK_SIZE><<<B * H, C / H, 0, stream>>>(B, T, C, H, r_d, w_d, k_d, v_d, a_d, b_d, s_d, dst_d);
+    } else {
+        rwkv_wkv7_f32<CUDA_WKV_BLOCK_SIZE * 2><<<B * H, C / H, 0, stream>>>(B, T, C, H, r_d, w_d, k_d, v_d, a_d, b_d, s_d, dst_d);
+    }
+}

data/ext/sources/ggml/src/ggml-cuda/wkv.cuh

@@ -0,0 +1,7 @@
+#include "common.cuh"
+
+#define CUDA_WKV_BLOCK_SIZE 64
+
+void ggml_cuda_op_rwkv_wkv6(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_rwkv_wkv7(ggml_backend_cuda_context & ctx, ggml_tensor * dst);