whispercpp 1.3.5 → 1.3.7

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

@@ -65,6 +65,11 @@ static __device__ __forceinline__ float op_sqr(float x) {
     return x * x;
 }
 
+static __device__ __forceinline__ float op_relu_sqr(float x) {
+    const float r = fmaxf(x, 0.0f);
+    return r * r;
+}
+
 static __device__ __forceinline__ float op_sqrt(float x) {
     return sqrtf(x);
 }
@@ -111,19 +116,22 @@ static __device__ __forceinline__ float op_trunc(float x) {
 
 template <float (*op)(float), typename T>
 static __global__ void unary_op_kernel(const T * x, T * dst, const int k) {
+    ggml_cuda_pdl_lc();
     const int i = blockDim.x*blockIdx.x + threadIdx.x;
 
     if (i >= k) {
         return;
     }
 
+    ggml_cuda_pdl_sync();
     dst[i] = (T)op((float)x[i]);
 }
 
 template <float (*op)(float), typename T>
 static void unary_cuda(const T * x, T * dst, const int k, cudaStream_t stream) {
     const int num_blocks = (k + CUDA_NEG_BLOCK_SIZE - 1) / CUDA_NEG_BLOCK_SIZE;
-    unary_op_kernel<op><<<num_blocks, CUDA_NEG_BLOCK_SIZE, 0, stream>>>(x, dst, k);
+    const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params((dim3)num_blocks, CUDA_NEG_BLOCK_SIZE, 0, stream);
+    ggml_cuda_kernel_launch(unary_op_kernel<op, T>, launch_params, x, dst, k);
 }
 
 template <float (*op)(float)>
@@ -253,6 +261,7 @@ void ggml_cuda_op_softplus(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
 
 template <float (*op)(float), typename T>
 static __global__ void unary_gated_op_kernel(const T * x, const T * g, T * dst, const int64_t k, const int64_t n, const int64_t o0, const int64_t o1) {
+    ggml_cuda_pdl_lc();
     const int64_t i = int64_t(blockDim.x)*blockIdx.x + threadIdx.x;
 
     if (i >= k) {
@@ -263,13 +272,15 @@ static __global__ void unary_gated_op_kernel(const T * x, const T * g, T * dst,
     const int64_t j0 = (i / n) * o0 + (i % n);
     const int64_t j1 = o0 == o1 ? j0 : (i / n) * o1 + (i % n);
 
+    ggml_cuda_pdl_sync();
     dst[i] = (T)(op((float)x[j0]) * (float)g[j1]);
 }
 
 template <float (*op)(float), typename T>
 static void unary_gated_cuda(const T * x, const T * g, T * dst, const int64_t k, const int64_t n, const int64_t o0, const int64_t o1, cudaStream_t stream) {
     const int64_t num_blocks = (k + CUDA_GLU_BLOCK_SIZE - 1) / CUDA_GLU_BLOCK_SIZE;
-    unary_gated_op_kernel<op><<<num_blocks, CUDA_GLU_BLOCK_SIZE, 0, stream>>>(x, g, dst, k, n, o0, o1);
+    const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params((dim3)num_blocks, CUDA_GLU_BLOCK_SIZE, 0, stream);
+    ggml_cuda_kernel_launch(unary_gated_op_kernel<op, T>, launch_params, x, g, dst, k, n, o0, o1);
 }
 
 template <float (*op)(float)>
@@ -560,3 +571,76 @@ void ggml_cuda_op_leaky_relu(ggml_backend_cuda_context & ctx, ggml_tensor * dst)
         leaky_relu_cuda((const float *)src0_d, (float *)dst_d, ggml_nelements(src0), negative_slope, stream);
     }
 }
+
+/* fused unary + mul */
+
+template <float (*op)(float)>
+static void ggml_cuda_op_unary_mul_impl(ggml_backend_cuda_context & ctx, ggml_tensor * unary_node, ggml_tensor * mul_node) {
+    // unary_node: UNARY op applied to unary_node->src[0]
+    // mul_node:   MUL(a, b) where one of a/b is unary_node
+    // Output goes to mul_node->data
+
+    const ggml_tensor * unary_src = unary_node->src[0];  // input to the unary op
+    const ggml_tensor * other_src = (mul_node->src[0] == unary_node) ? mul_node->src[1] : mul_node->src[0];
+
+    GGML_ASSERT(ggml_is_contiguous_1(unary_src));
+    GGML_ASSERT(unary_src->nb[0] == ggml_element_size(unary_src));
+    GGML_ASSERT(ggml_is_contiguous_1(other_src));
+    GGML_ASSERT(other_src->nb[0] == ggml_element_size(other_src));
+    GGML_ASSERT(ggml_are_same_shape(unary_src, other_src));
+
+    GGML_ASSERT(unary_src->type == GGML_TYPE_F32 || unary_src->type == GGML_TYPE_F16);
+    GGML_ASSERT(unary_src->type == other_src->type);
+    GGML_ASSERT(unary_src->type == mul_node->type);
+
+    cudaStream_t stream = ctx.stream();
+
+    const int64_t k  = ggml_nelements(mul_node);
+    const int64_t nc = unary_src->ne[0];
+    const int64_t unary_stride = unary_src->nb[1];
+    const int64_t other_stride = other_src->nb[1];
+
+    if (unary_src->type == GGML_TYPE_F16) {
+        unary_gated_cuda<op>((const half *) unary_src->data, (const half *) other_src->data,
+                             (half *) mul_node->data, k, nc,
+                             unary_stride / sizeof(half), other_stride / sizeof(half), stream);
+    } else {
+        unary_gated_cuda<op>((const float *) unary_src->data, (const float *) other_src->data,
+                             (float *) mul_node->data, k, nc,
+                             unary_stride / sizeof(float), other_stride / sizeof(float), stream);
+    }
+}
+
+void ggml_cuda_op_unary_mul(ggml_backend_cuda_context & ctx, ggml_tensor * unary_node, ggml_tensor * mul_node) {
+    switch (ggml_get_unary_op(unary_node)) {
+        case GGML_UNARY_OP_SILU:
+            ggml_cuda_op_unary_mul_impl<op_silu>(ctx, unary_node, mul_node);
+            break;
+        case GGML_UNARY_OP_SIGMOID:
+            ggml_cuda_op_unary_mul_impl<op_sigmoid>(ctx, unary_node, mul_node);
+            break;
+        case GGML_UNARY_OP_SOFTPLUS:
+            ggml_cuda_op_unary_mul_impl<op_softplus>(ctx, unary_node, mul_node);
+            break;
+        default:
+            GGML_ABORT("Unsupported unary op for fused unary+mul");
+    }
+}
+
+/* fused relu + sqr */
+
+void ggml_cuda_op_relu_sqr(ggml_backend_cuda_context & ctx, ggml_tensor * relu_node, ggml_tensor * sqr_node) {
+    const ggml_tensor * src = relu_node->src[0];
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(ggml_is_contiguous(src));
+    GGML_ASSERT(src->type == GGML_TYPE_F32 || src->type == GGML_TYPE_F16);
+    GGML_ASSERT(src->type == sqr_node->type);
+
+    const int k = ggml_nelements(src);
+    if (src->type == GGML_TYPE_F16) {
+        unary_cuda<op_relu_sqr>((const half *)src->data, (half *)sqr_node->data, k, stream);
+    } else {
+        unary_cuda<op_relu_sqr>((const float *)src->data, (float *)sqr_node->data, k, stream);
+    }
+}

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

@@ -89,6 +89,10 @@ void ggml_cuda_op_geglu_quick(ggml_backend_cuda_context & ctx, ggml_tensor * dst
 
 void ggml_cuda_op_xielu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
+void ggml_cuda_op_unary_mul(ggml_backend_cuda_context & ctx, ggml_tensor * unary_node, ggml_tensor * mul_node);
+
+void ggml_cuda_op_relu_sqr(ggml_backend_cuda_context & ctx, ggml_tensor * relu_node, ggml_tensor * sqr_node);
+
 __device__ __forceinline__ float ggml_cuda_op_silu_single(float x) {
     return x / (1.0f + expf(-x));
 }

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

@@ -94,9 +94,21 @@ static __device__ __forceinline__ int2 get_int_from_table_16(const int & q4, con
 #endif
 }
 
+static __device__ __forceinline__ uint32_t unpack_ksigns(const uint8_t v) {
+    // v is a 7 bit int, with the 8th sign being encodable as popcnt
+    // with xor we can "correct" the bit instead of having to mask
+    const uint32_t p = __popc(v) & 1;
+    const uint32_t s = v ^ p << 7;
+    // broadcast over uint to allow for 0x08040201 / 0x80402010 as selectors
+    return s * 0x01010101;
+}
+
 // VDR = vec dot ratio, how many contiguous integers each thread processes when the vec dot kernel is called
 // MMVQ = mul_mat_vec_q, MMQ = mul_mat_q
 
+#define VDR_Q1_0_Q8_1_MMVQ 1  // Process one 32-element chunk at a time for parallelism
+#define VDR_Q1_0_Q8_1_MMQ  4  // Q1_0 has 128 bits (4 ints) per block
+
 #define VDR_Q4_0_Q8_1_MMVQ 2
 #define VDR_Q4_0_Q8_1_MMQ  4
 
@@ -313,6 +325,38 @@ static __device__ __forceinline__ float vec_dot_mxfp4_q8_1(
     return d * sumi;
 }
 
+#define VDR_NVFP4_Q8_1_MMVQ 4
+#define VDR_NVFP4_Q8_1_MMQ  8
+
+static __device__ __forceinline__ float vec_dot_nvfp4_q8_1(
+                                        const void * __restrict__ vbq,
+                                        const block_q8_1 * __restrict__ bq8_1,
+                                        const int32_t & kbx,
+                                        const int32_t & iqs) {
+
+    const block_nvfp4 * bq4 = (const block_nvfp4 *) vbq + kbx;
+    float sum = 0.0f;
+#pragma unroll
+    for (int i = 0; i < VDR_NVFP4_Q8_1_MMVQ/2; i++) {
+        const int32_t iqs0 = iqs + 2*i;
+        const int32_t iqs1 = iqs0 + 1;
+        const int32_t is = iqs0 >> 1;
+        const int2 v0 = get_int_from_table_16(get_int_b4(bq4->qs, iqs0), kvalues_mxfp4);
+        const int2 v1 = get_int_from_table_16(get_int_b4(bq4->qs, iqs1), kvalues_mxfp4);
+        const block_q8_1 * bq8 = bq8_1 + (is >> 1);
+        const int32_t i8 = ((is & 1) << 2);
+
+        int sumi = ggml_cuda_dp4a(v0.x, get_int_b4(bq8->qs, i8 + 0), 0);
+        sumi = ggml_cuda_dp4a(v0.y, get_int_b4(bq8->qs, i8 + 2), sumi);
+        sumi = ggml_cuda_dp4a(v1.x, get_int_b4(bq8->qs, i8 + 1), sumi);
+        sumi = ggml_cuda_dp4a(v1.y, get_int_b4(bq8->qs, i8 + 3), sumi);
+
+        const float d = ggml_cuda_ue4m3_to_fp32(bq4->d[is]) * __low2float(bq8->ds);
+        sum += d * float(sumi);
+    }
+
+    return sum;
+}
 #define VDR_Q2_K_Q8_1_MMVQ 1
 #define VDR_Q2_K_Q8_1_MMQ  4
 
@@ -628,6 +672,51 @@ static __device__ __forceinline__ float vec_dot_q6_K_q8_1_impl_mmq(
     return d6 * sumf_d;
 }
 
+static __device__ __forceinline__ float vec_dot_q1_0_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) {
+
+    const block_q1_0 * bq1_0 = (const block_q1_0 *) vbq + kbx;
+
+    // Q1_0: 128 elements with ONE scale
+    // Q8_1: 32 elements per block with individual scales
+    // iqs selects which of the 4 chunks of 32 elements to process (0-3)
+
+    const float d1 = bq1_0->d;
+
+    // Process only the chunk specified by iqs
+    const block_q8_1 * bq8_1_chunk = bq8_1 + iqs;
+
+    // Load 32 bits (4 bytes) for this chunk from Q1_0
+    const int offset = iqs * 4;
+    const int v = bq1_0->qs[offset + 0] | (bq1_0->qs[offset + 1] << 8) |
+                  (bq1_0->qs[offset + 2] << 16) | (bq1_0->qs[offset + 3] << 24);
+
+    // Unpack 32 bits into 32 signed values (-1 or +1)
+    int vi_bytes[8];
+#pragma unroll
+    for (int j = 0; j < 8; ++j) {
+        const int shift = j * 4;
+        const int bits4 = (v >> shift) & 0x0F;
+        const int b0 = (bits4 & 0x01) ? 1 : -1;
+        const int b1 = (bits4 & 0x02) ? 1 : -1;
+        const int b2 = (bits4 & 0x04) ? 1 : -1;
+        const int b3 = (bits4 & 0x08) ? 1 : -1;
+        vi_bytes[j] = (b0 & 0xFF) | ((b1 & 0xFF) << 8) | ((b2 & 0xFF) << 16) | ((b3 & 0xFF) << 24);
+    }
+
+    // Compute dot product for this 32-element chunk
+    int sumi = 0;
+#pragma unroll
+    for (int j = 0; j < 8; ++j) {
+        const int u = get_int_b4(bq8_1_chunk->qs, j);
+        sumi = ggml_cuda_dp4a(vi_bytes[j], u, sumi);
+    }
+
+    // Apply Q1_0's single scale and this chunk's Q8_1 scale
+    const float d8 = __low2float(bq8_1_chunk->ds);
+    return d1 * d8 * sumi;
+}
+
 static __device__ __forceinline__ float vec_dot_q4_0_q8_1(
     const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) {
 
@@ -905,22 +994,22 @@ static __device__ __forceinline__ float vec_dot_iq2_xxs_q8_1(
     int sumi = 0;
 #pragma unroll
     for (int k0 = 0; k0 < 8; k0 += 2) {
-        const int * grid_pos = (const int *) (iq2xxs_grid + aux8[k0/2]);
-        const int signs_packed = ksigns_iq2xs[(aux32 >> (7*k0/2)) & 0x7F];
+        const uint2 grid_pos = ((const uint2*)iq2xxs_grid)[aux8[k0/2]];
+        const uint32_t signs = unpack_ksigns(aux32 >> (7 * k0 / 2));
 
-        const int signs0 = __vcmpne4(((signs_packed & 0x03) << 7) | ((signs_packed & 0x0C) << 21), 0x00000000);
-        const int grid0 = __vsub4(grid_pos[0] ^ signs0, signs0);
+        const int signs0 = __vcmpne4(signs & 0x08040201, 0);
+        const int grid0 = __vsub4(grid_pos.x ^ signs0, signs0);
         const int u0 = get_int_b4(bq8_1[iqs/2].qs, k0 + 0);
         sumi = ggml_cuda_dp4a(grid0, u0, sumi);
 
-        const int signs1 = __vcmpne4(((signs_packed & 0x30) << 3) | ((signs_packed & 0xC0) << 17), 0x00000000);
-        const int grid1 = __vsub4(grid_pos[1] ^ signs1, signs1);
+        const int signs1 = __vcmpne4(signs & 0x80402010, 0);
+        const int grid1 = __vsub4(grid_pos.y ^ signs1, signs1);
         const int u1 = get_int_b4(bq8_1[iqs/2].qs, k0 + 1);
         sumi = ggml_cuda_dp4a(grid1, u1, sumi);
     }
 
-    const int ls = aux32 >> 28;
-    sumi = (ls*sumi + sumi/2)/4;
+    const int ls = aux32 >> 27 | 1; // (scale * 2 + 1)
+    sumi = sumi * ls / 8;           // (sumi * scale + sumi / 2) / 4
     const float d = __half2float(bq2->d) * __low2float(bq8_1[iqs/2].ds);
     return d * sumi;
 }
@@ -942,13 +1031,15 @@ static __device__ __forceinline__ float vec_dot_iq2_xs_q8_1(
     int sumi1 = 0;
 #pragma unroll
     for (int l0 = 0; l0 < 8; l0 += 2) {
-        const uint32_t * grid_pos = (const uint32_t *)(iq2xs_grid + (q2[l0/2] & 0x000001FF));
-        const uint32_t * signs    = (const uint32_t *)(ksigns64   + (q2[l0/2] >> 9));
-
-        const int grid_l = __vsub4(grid_pos[0] ^ signs[0], signs[0]);
-        const int grid_h = __vsub4(grid_pos[1] ^ signs[1], signs[1]);
+        const uint2 grid_pos = ((const uint2*)iq2xs_grid)[q2[l0/2] & 0x1FF];
+        const uint32_t signs = unpack_ksigns(q2[l0/2] >> 9);
 
+        const int signs0 = __vcmpne4(signs & 0x08040201, 0);
+        const int grid_l = __vsub4(grid_pos.x ^ signs0, signs0);
         const int u0 = get_int_b4(bq8_1[iqs/2].qs, l0 + 0);
+
+        const int signs1 = __vcmpne4(signs & 0x80402010, 0);
+        const int grid_h = __vsub4(grid_pos.y ^ signs1, signs1);
         const int u1 = get_int_b4(bq8_1[iqs/2].qs, l0 + 1);
 
         if (l0 < 4) {
@@ -1028,13 +1119,16 @@ static __device__ __forceinline__ float vec_dot_iq3_xxs_q8_1(
 #pragma unroll
     for (int l0 = 0; l0 < 8; l0 += 2) {
         const int2 grid_pos = make_int2(iq3xxs_grid[q3[l0 + 0]], iq3xxs_grid[q3[l0 + 1]]);
+        const uint32_t signs = unpack_ksigns(aux32 >> (7*l0/2));
 
-        const int * signs = (const int *)(ksigns64 + ((aux32 >> (7*l0/2)) & 0x7F));
-
-        const int grid_l = __vsub4(grid_pos.x ^ signs[0], signs[0]);
-        const int grid_h = __vsub4(grid_pos.y ^ signs[1], signs[1]);
+        const int signs0 = __vcmpne4(signs & 0x08040201, 0);
+        const int grid_l = __vsub4(grid_pos.x ^ signs0, signs0);
 
         const int u0 = get_int_b4(bq8_1[iqs/2].qs, l0 + 0);
+
+        const int signs1 = __vcmpne4(signs & 0x80402010, 0);
+        const int grid_h = __vsub4(grid_pos.y ^ signs1, signs1);
+
         const int u1 = get_int_b4(bq8_1[iqs/2].qs, l0 + 1);
 
         sumi = ggml_cuda_dp4a(grid_l, u0, sumi);

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

@@ -6,9 +6,14 @@
 #include <cuda_bf16.h>
 #include <cuda_fp16.h>
 
-#if CUDART_VERSION >= 12050
+#ifdef GGML_USE_NCCL
+#include <nccl.h>
+#endif // GGML_USE_NCCL
+
+#if CUDART_VERSION >= 11080
 #include <cuda_fp8.h>
-#endif // CUDART_VERSION >= 12050
+#define FP8_AVAILABLE
+#endif // CUDART_VERSION >= 11080
 
 #if CUDART_VERSION >= 12080
 #include <cuda_fp4.h>

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

@@ -10,6 +10,11 @@
 #include <rocwmma/rocwmma-version.hpp>
 #endif // defined(GGML_HIP_ROCWMMA_FATTN)
 
+#ifdef GGML_USE_NCCL
+#include <rccl/rccl.h>
+#endif // GGML_USE_NCCL
+
+
 #define CUBLAS_GEMM_DEFAULT HIPBLAS_GEMM_DEFAULT
 #define CUBLAS_GEMM_DEFAULT_TENSOR_OP HIPBLAS_GEMM_DEFAULT
 #define CUBLAS_OP_N HIPBLAS_OP_N
@@ -43,6 +48,7 @@
 #define cublasSetMathMode(handle, mode) CUBLAS_STATUS_SUCCESS
 #define cublasSetStream hipblasSetStream
 #define cublasSgemm hipblasSgemm
+#define cublasSgemmStridedBatched hipblasSgemmStridedBatched
 #define cublasStatus_t hipblasStatus_t
 #define cublasOperation_t hipblasOperation_t
 #define cudaDevAttrCooperativeLaunch hipDeviceAttributeCooperativeLaunch
@@ -50,9 +56,11 @@
 #define cudaDeviceDisablePeerAccess hipDeviceDisablePeerAccess
 #define cudaDeviceEnablePeerAccess hipDeviceEnablePeerAccess
 #define cudaDeviceGetAttribute hipDeviceGetAttribute
+#define cudaDeviceGetPCIBusId hipDeviceGetPCIBusId
 #define cudaDeviceProp hipDeviceProp_t
 #define cudaDeviceSynchronize hipDeviceSynchronize
 #define cudaError_t hipError_t
+#define cudaErrorMemoryAllocation hipErrorOutOfMemory
 #define cudaErrorPeerAccessAlreadyEnabled hipErrorPeerAccessAlreadyEnabled
 #define cudaErrorPeerAccessNotEnabled hipErrorPeerAccessNotEnabled
 #define cudaEventCreateWithFlags hipEventCreateWithFlags
@@ -138,6 +146,8 @@
 #define cudaStream_t hipStream_t
 #define cudaSuccess hipSuccess
 #define cudaOccupancyMaxActiveBlocksPerMultiprocessor hipOccupancyMaxActiveBlocksPerMultiprocessor
+#define cudaFuncSetAttribute hipFuncSetAttribute
+#define cudaFuncAttributeMaxDynamicSharedMemorySize hipFuncAttributeMaxDynamicSharedMemorySize
 #define __trap() do { abort(); __builtin_unreachable(); } while(0)
 #define CUBLAS_STATUS_SUCCESS HIPBLAS_STATUS_SUCCESS
 #define CUBLAS_STATUS_NOT_INITIALIZED HIPBLAS_STATUS_NOT_INITIALIZED
@@ -181,6 +191,10 @@
 #define GCN
 #endif // defined(GCN5) || defined(GCN4)
 
+#if defined(__gfx950__)
+#define CDNA4
+#endif // defined(__gfx950__)
+
 #if defined(__gfx942__)
 #define CDNA3
 #endif // defined(__gfx942__)
@@ -193,9 +207,9 @@
 #define CDNA1
 #endif // defined(__gfx908__)
 
-#if defined(CDNA3) || defined(CDNA2) || defined(CDNA1)
+#if defined(CDNA4) || defined(CDNA3) || defined(CDNA2) || defined(CDNA1)
 #define CDNA // For the entire family
-#endif // defined(CDNA3) || defined(CDNA2) || defined(CDNA1)
+#endif // defined(CDNA4) || defined(CDNA3) || defined(CDNA2) || defined(CDNA1)
 
 #if defined(__GFX12__)
 #define RDNA4
@@ -205,6 +219,14 @@
 #define RDNA3
 #endif // defined(__GFX11__)
 
+#if defined(__gfx1150__) || defined(__gfx1151__) || defined(__gfx1152__) || defined(__gfx1153__)
+#define RDNA3_5
+#endif // defined(__gfx1150__) || defined(__gfx1151__) || defined(__gfx1152__) || defined(__gfx1153__)
+
+#if defined(RDNA3) && !defined(RDNA3_5)
+#define RDNA3_0
+#endif // defined(RDNA3) && !defined(RDNA3_5)
+
 #if defined(__gfx1030__) || defined(__gfx1031__) || defined(__gfx1032__) || defined(__gfx1033__) || \
     defined(__gfx1034__) || defined(__gfx1035__) || defined(__gfx1036__) || defined(__gfx1037__)
 #define RDNA2
@@ -225,6 +247,12 @@
 typedef __hip_bfloat16 nv_bfloat16;
 typedef __hip_bfloat162 nv_bfloat162;
 
+#if HIP_VERSION >= 60200000
+#include <hip/hip_fp8.h>
+typedef __hip_fp8_e4m3 __nv_fp8_e4m3;
+#define FP8_AVAILABLE
+#endif // HIP_VERSION >= 60200000
+
 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) {

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

@@ -32,6 +32,7 @@
 #define cublasSetMathMode mublasSetMathMode
 #define cublasSetStream mublasSetStream
 #define cublasSgemm mublasSgemm
+#define cublasSgemmStridedBatched mublasSgemmStridedBatched
 #define cublasStatus_t mublasStatus_t
 #define cublasOperation_t mublasOperation_t
 #define cublasGetStatusString mublasGetStatusString
@@ -39,9 +40,11 @@
 #define cudaDeviceCanAccessPeer musaDeviceCanAccessPeer
 #define cudaDeviceDisablePeerAccess musaDeviceDisablePeerAccess
 #define cudaDeviceEnablePeerAccess musaDeviceEnablePeerAccess
+#define cudaDeviceGetPCIBusId musaDeviceGetPCIBusId
 #define cudaDeviceProp musaDeviceProp
 #define cudaDeviceSynchronize musaDeviceSynchronize
 #define cudaError_t musaError_t
+#define cudaErrorMemoryAllocation musaErrorMemoryAllocation
 #define cudaErrorPeerAccessAlreadyEnabled musaErrorPeerAccessAlreadyEnabled
 #define cudaErrorPeerAccessNotEnabled musaErrorPeerAccessNotEnabled
 #define cudaEventCreateWithFlags musaEventCreateWithFlags

data/ext/sources/ggml/src/ggml-hexagon/CMakeLists.txt

@@ -1,7 +1,30 @@
+file(TO_CMAKE_PATH "${HEXAGON_SDK_ROOT}"   HEXAGON_SDK_ROOT)
+file(TO_CMAKE_PATH "${HEXAGON_TOOLS_ROOT}" HEXAGON_TOOLS_ROOT)
+
+if (NOT IS_DIRECTORY "${HEXAGON_SDK_ROOT}")
+    message(FATAL_ERROR "Make sure HEXAGON_SDK_ROOT point to the correct Hexagon SDK installation.")
+endif()
+
+if (NOT IS_DIRECTORY "${HEXAGON_TOOLS_ROOT}")
+    message("Try to read HEXAGON_TOOLS_ROOT from hexagon_sdk.json")
+    file(READ "${HEXAGON_SDK_ROOT}/hexagon_sdk.json" HEXAGON_SDK_CONFIG_PATH)
+    string(JSON HEXAGON_TOOLS_PATH GET ${HEXAGON_SDK_CONFIG_PATH} "root" "tools" "info" 0 "path")
+    message("Found HEXAGON_TOOLS_PATH: ${HEXAGON_TOOLS_PATH}")
+    set(HEXAGON_TOOLS_ROOT "${HEXAGON_SDK_ROOT}/${HEXAGON_TOOLS_PATH}")
+    file(TO_CMAKE_PATH "${HEXAGON_TOOLS_ROOT}" HEXAGON_TOOLS_ROOT)
+    if (NOT IS_DIRECTORY "${HEXAGON_TOOLS_ROOT}")
+        message(FATAL_ERROR "Make sure HEXAGON_TOOLS_ROOT point to the correct Hexagon SDK installation.")
+    endif()
+endif()
+
+message(STATUS "hexagon: using ${HEXAGON_SDK_ROOT} and ${HEXAGON_TOOLS_ROOT} for building libggml-htp skels")
+
 include(${HEXAGON_SDK_ROOT}/build/cmake/hexagon_fun.cmake)
 include(ExternalProject)
 
-option(GGML_HEXAGON_HTP_DEBUG "ggml-hexagon: enable HTP debug output" OFF)
+option(GGML_HEXAGON_HTP_DEBUG  "ggml-hexagon: enable HTP debug output" OFF)
+option(GGML_HEXAGON_FA_EXP2_HF "ggml-hexagon: use FP16 exp2 polynomial in FA softmax instead of F32 exp round-trip" OFF)
+set(GGML_HEXAGON_HTP_CERT  "$ENV{HEXAGON_HTP_CERT}" CACHE PATH "ggml-hexagon: enable HTP library signing using certificate")
 set(GGML_HEXAGON_FP32_QUANTIZE_GROUP_SIZE 128 CACHE STRING "ggml-hexagon: quantize group size (32, 64, or 128)")
 
 add_library(htp_iface OBJECT
@@ -25,56 +48,71 @@ else()
     target_link_options(htp_iface PUBLIC -ldl)
 endif()
 
-link_custom_library(htp_iface cdsprpc)
-link_custom_library(htp_iface rpcmem)
-
 set(TARGET_NAME ggml-hexagon)
 ggml_add_backend_library(${TARGET_NAME}
-    ggml-hexagon.cpp htp-utils.c htp-utils.h ../../include/ggml-hexagon.h)
+    ggml-hexagon.cpp
+    htp-drv.cpp
+    htp-drv.h
+    libdl.h
+    ../../include/ggml-hexagon.h)
 
 target_link_libraries(${TARGET_NAME} PRIVATE htp_iface)
 target_include_directories(${TARGET_NAME} PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/htp ${CMAKE_CURRENT_BINARY_DIR})
 
-# Build HTP bits
-set(HTP_CMAKE_ARGS
-    -DCMAKE_TOOLCHAIN_FILE=${CMAKE_CURRENT_SOURCE_DIR}/htp/cmake-toolchain.cmake
-    -DCMAKE_BUILD_TYPE=Release
-    -DCMAKE_INSTALL_LIBDIR=${CMAKE_CURRENT_BINARY_DIR}
-    -DHEXAGON_SDK_ROOT=$ENV{HEXAGON_SDK_ROOT}
-    -DHEXAGON_TOOLS_ROOT=$ENV{HEXAGON_TOOLS_ROOT}
-    -DHEXAGON_HTP_DEBUG=${GGML_HEXAGON_HTP_DEBUG}
-    -DGGML_HEXAGON_FP32_QUANTIZE_GROUP_SIZE=${GGML_HEXAGON_FP32_QUANTIZE_GROUP_SIZE})
-
-ExternalProject_Add(htp-v68
-    SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/htp BUILD_ALWAYS ON
-    CMAKE_ARGS ${HTP_CMAKE_ARGS} -DDSP_VERSION=v68 -DPREBUILT_LIB_DIR="toolv19_v68")
-
-ExternalProject_Add(htp-v69
-    SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/htp BUILD_ALWAYS ON
-    CMAKE_ARGS ${HTP_CMAKE_ARGS} -DDSP_VERSION=v69 -DPREBUILT_LIB_DIR="toolv19_v69")
-
-ExternalProject_Add(htp-v73
-    SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/htp BUILD_ALWAYS ON
-    CMAKE_ARGS ${HTP_CMAKE_ARGS} -DDSP_VERSION=v73 -DPREBUILT_LIB_DIR="toolv19_v73")
-
-ExternalProject_Add(htp-v75
-    SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/htp BUILD_ALWAYS ON
-    CMAKE_ARGS ${HTP_CMAKE_ARGS} -DDSP_VERSION=v75 -DPREBUILT_LIB_DIR="toolv19_v75")
-
-ExternalProject_Add(htp-v79
-    SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/htp BUILD_ALWAYS ON
-    CMAKE_ARGS ${HTP_CMAKE_ARGS} -DDSP_VERSION=v79 -DPREBUILT_LIB_DIR="toolv19_v79")
-
-ExternalProject_Add(htp-v81
-    SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/htp BUILD_ALWAYS ON
-    CMAKE_ARGS ${HTP_CMAKE_ARGS} -DDSP_VERSION=v81 -DPREBUILT_LIB_DIR="toolv19_v81")
+# Build HTP skels
+set(HTP_SKELS)
+function(build_htp_skel V)
+    ExternalProject_Add(htp-${V}
+        SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/htp BUILD_ALWAYS ON
+        BUILD_BYPRODUCTS ${CMAKE_CURRENT_BINARY_DIR}/libggml-htp-${V}.so
+        CMAKE_ARGS
+            -DCMAKE_BUILD_TYPE=Release
+            -DCMAKE_TOOLCHAIN_FILE=${CMAKE_CURRENT_SOURCE_DIR}/htp/cmake-toolchain.cmake
+            -DCMAKE_INSTALL_LIBDIR=${CMAKE_CURRENT_BINARY_DIR}
+            -DHEXAGON_SDK_ROOT=${HEXAGON_SDK_ROOT}
+            -DHEXAGON_TOOLS_ROOT=${HEXAGON_TOOLS_ROOT}
+            -DHEXAGON_HTP_DEBUG=${GGML_HEXAGON_HTP_DEBUG}
+            -DGGML_HEXAGON_FP32_QUANTIZE_GROUP_SIZE=${GGML_HEXAGON_FP32_QUANTIZE_GROUP_SIZE}
+            -DDSP_VERSION=${V}
+            -DPREBUILT_LIB_DIR="toolv19_${V}")
+    list(APPEND HTP_SKELS ${CMAKE_CURRENT_BINARY_DIR}/libggml-htp-${V}.so)
+    set(HTP_SKELS ${HTP_SKELS} PARENT_SCOPE)
+endfunction()
+
+build_htp_skel(v68)
+build_htp_skel(v69)
+build_htp_skel(v73)
+build_htp_skel(v75)
+build_htp_skel(v79)
+build_htp_skel(v81)
 
 # Install Hexagon skels required at runtime
-install(FILES
-    ${CMAKE_CURRENT_BINARY_DIR}/libggml-htp-v68.so
-    ${CMAKE_CURRENT_BINARY_DIR}/libggml-htp-v69.so
-    ${CMAKE_CURRENT_BINARY_DIR}/libggml-htp-v73.so
-    ${CMAKE_CURRENT_BINARY_DIR}/libggml-htp-v75.so
-    ${CMAKE_CURRENT_BINARY_DIR}/libggml-htp-v79.so
-    ${CMAKE_CURRENT_BINARY_DIR}/libggml-htp-v81.so
-    TYPE LIB)
+install(FILES ${HTP_SKELS} TYPE LIB)
+
+if (CMAKE_SYSTEM_NAME MATCHES Windows AND GGML_HEXAGON_HTP_CERT)
+    file(TO_CMAKE_PATH "$ENV{WINDOWS_SDK_BIN}/arm64"      WINSDK_BIN0_ARM64)
+    file(TO_CMAKE_PATH "$ENV{WINDOWS_SDK_BIN}/x86"        WINSDK_BIN0_X86)
+    file(TO_CMAKE_PATH "$ENV{WindowsSdkVerBinPath}/arm64" WINSDK_BIN1_ARM64)
+    file(TO_CMAKE_PATH "$ENV{WindowsSdkVerBinPath}/x86"   WINSDK_BIN1_X86)
+
+    set(WINSDK_PATHS ${WINSDK_BIN0_ARM64} ${WINSDK_BIN0_X86} ${WINSDK_BIN1_ARM64} ${WINSDK_BIN1_X86})
+
+    find_program(INF2CAT  NAMES inf2cat.exe  PATHS ${WINSDK_PATHS} REQUIRED)
+    find_program(SIGNTOOL NAMES signtool.exe PATHS ${WINSDK_PATHS} REQUIRED)
+
+    message(STATUS "hexagon: using ${GGML_HEXAGON_HTP_CERT} to sign libggml-htp skels")
+
+    set(LIBGGML_HTP_CAT ${CMAKE_CURRENT_BINARY_DIR}/libggml-htp.cat)
+    add_custom_target(libggml-htp-cat
+        BYPRODUCTS ${LIBGGML_HTP_CAT}
+        DEPENDS libggml-htp.inf ${HTP_SKELS}
+        COMMAND ${CMAKE_COMMAND} -E copy ${CMAKE_CURRENT_SOURCE_DIR}/libggml-htp.inf ${CMAKE_CURRENT_BINARY_DIR}
+        COMMAND ${INF2CAT} /driver:${CMAKE_CURRENT_BINARY_DIR} /os:10_25H2_ARM64
+        COMMAND ${SIGNTOOL} sign /fd sha256 /f ${GGML_HEXAGON_HTP_CERT} ${LIBGGML_HTP_CAT}
+        COMMENT "generating and signing libggml-htp.cat file"
+        VERBATIM
+    )
+
+    add_dependencies(${TARGET_NAME} libggml-htp-cat)
+    install(FILES ${LIBGGML_HTP_CAT} TYPE LIB)
+endif()