whispercpp 1.3.4 → 1.3.5

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

@@ -0,0 +1,39 @@
+#include "set.cuh"
+#include "cpy.cuh"
+
+void ggml_cuda_op_set(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+
+    GGML_ASSERT((src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_I32));
+    GGML_ASSERT(src1->type == src0->type);
+    GGML_ASSERT(dst ->type == src0->type);
+
+    GGML_ASSERT(ggml_is_contiguous(dst));
+    GGML_ASSERT(ggml_is_contiguous(src0));
+    GGML_ASSERT(ggml_is_contiguous(src1));
+
+    const size_t nb1    = ((int32_t *) dst->op_params)[0];
+    const size_t nb2    = ((int32_t *) dst->op_params)[1];
+    const size_t nb3    = ((int32_t *) dst->op_params)[2];
+    const size_t offset = ((int32_t *) dst->op_params)[3];
+    const bool   inplace= (bool)     ((int32_t *) dst->op_params)[4];
+
+    if (!inplace) {
+        ggml_cuda_cpy(ctx, src0, dst);
+    }
+
+    ggml_tensor dst_view = *dst;
+    dst_view.data  = (void *)((char *)dst->data + offset);
+    dst_view.ne[0] = src1->ne[0];
+    dst_view.ne[1] = src1->ne[1];
+    dst_view.ne[2] = src1->ne[2];
+    dst_view.ne[3] = src1->ne[3];
+
+    dst_view.nb[0] = ggml_element_size(dst);
+    dst_view.nb[1] = nb1;
+    dst_view.nb[2] = nb2;
+    dst_view.nb[3] = nb3;
+
+    ggml_cuda_cpy(ctx, src1, &dst_view);
+}

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

@@ -0,0 +1,7 @@
+#pragma once
+
+#include "common.cuh"
+
+#define CUDA_SET_BLOCK_SIZE 256
+
+void ggml_cuda_op_set(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

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

@@ -1,6 +1,14 @@
 #include "common.cuh"
 #include "ggml.h"
 #include "softmax.cuh"
+
+#ifdef GGML_USE_HIP
+#include <hip/hip_cooperative_groups.h>
+#else
+#include <cooperative_groups.h>
+#include <cooperative_groups/reduce.h>
+#endif // GGML_USE_HIP
+
 #include <cstdint>
 #include <utility>
 
@@ -160,6 +168,156 @@ static __global__ void soft_max_f32(
         dst[col] = vals[col] * inv_sum;
     }
 }
+
+
+// TODO: This is a common pattern used across kernels that could be moved to common.cuh + templated
+static __device__ float two_stage_warp_reduce_max(float val) {
+    val = warp_reduce_max(val);
+    if (blockDim.x > WARP_SIZE) {
+        assert((blockDim.x <= 1024) && (blockDim.x % WARP_SIZE) == 0);
+        __shared__ float local_vals[32];
+        const int        warp_id = threadIdx.x / WARP_SIZE;
+        const int        lane_id = threadIdx.x % WARP_SIZE;
+        if (lane_id == 0) {
+            local_vals[warp_id] = val;
+        }
+        __syncthreads();
+        val = -INFINITY;
+        if (lane_id < (static_cast<int>(blockDim.x) / WARP_SIZE)) {
+            val = local_vals[lane_id];
+        }
+        return warp_reduce_max(val);
+    } else {
+        return val;
+    }
+}
+
+static __device__ float two_stage_warp_reduce_sum(float val) {
+    val = warp_reduce_sum(val);
+    if (blockDim.x > WARP_SIZE) {
+        assert((blockDim.x <= 1024) && (blockDim.x % WARP_SIZE) == 0);
+        __shared__ float local_vals[32];
+        const int        warp_id = threadIdx.x / WARP_SIZE;
+        const int        lane_id = threadIdx.x % WARP_SIZE;
+        if (lane_id == 0) {
+            local_vals[warp_id] = val;
+        }
+        __syncthreads();
+        val = 0.0f;
+        if (lane_id < (static_cast<int>(blockDim.x) / WARP_SIZE)) {
+            val = local_vals[lane_id];
+        }
+        return warp_reduce_sum(val);
+    } else {
+        return val;
+    }
+}
+
+// TODO: Template to allow keeping ncols in registers if they fit
+static __device__ void soft_max_f32_parallelize_cols_single_row(const float * __restrict__ x,
+                                                                float * __restrict__ dst,
+                                                                float * __restrict__ tmp_maxs,
+                                                                float * __restrict__ tmp_sums,
+                                                                const soft_max_params p) {
+    namespace cg = cooperative_groups;
+
+    const cg::grid_group g = cg::this_grid();
+
+    const int tid               = threadIdx.x;
+    const int col_start         = blockIdx.x * blockDim.x + tid;
+    const int n_elem_per_thread = 4;
+
+    float     local_vals[n_elem_per_thread] = { -INFINITY, -INFINITY, -INFINITY, -INFINITY };
+    float     local_max                     = -INFINITY;
+    const int step_size                     = gridDim.x * blockDim.x;
+
+    // Compute thread-local max
+    for (int col = col_start; col < p.ncols;) {
+#pragma unroll
+        for (int i = 0; i < n_elem_per_thread; i++) {
+            const int idx = col + i * step_size;
+            local_vals[i] = idx < p.ncols ? x[idx] : -INFINITY;
+        }
+#pragma unroll
+        for (int i = 0; i < n_elem_per_thread; i++) {
+            local_max = fmaxf(local_max, local_vals[i]);
+        }
+        col += step_size * n_elem_per_thread;
+    }
+
+    // Compute CTA-level max
+    local_max = two_stage_warp_reduce_max(local_max);
+
+    // Store CTA-level max to GMEM
+    if (tid == 0) {
+        tmp_maxs[blockIdx.x] = local_max;
+    }
+    g.sync();
+
+    // Compute compute global max from CTA-level maxs
+    assert(gridDim.x < blockDim.x);  // currently we only support this case
+    if (tid < gridDim.x) {
+        local_max = tmp_maxs[tid];
+    } else {
+        local_max = -INFINITY;
+    }
+    local_max = two_stage_warp_reduce_max(local_max);
+
+    // Compute softmax dividends, accumulate divisor
+    float tmp_expf = 0.0f;
+    for (int col = col_start; col < p.ncols;) {
+#pragma unroll
+        for (int i = 0; i < n_elem_per_thread; i++) {
+            const int idx = col + i * step_size;
+            local_vals[i] = idx < p.ncols ? x[idx] : -INFINITY;
+        }
+#pragma unroll
+        for (int i = 0; i < n_elem_per_thread; i++) {
+            const int idx = col + i * step_size;
+            if (idx < p.ncols) {
+                const float tmp = expf(local_vals[i] - local_max);
+                tmp_expf += tmp;
+                dst[idx] = tmp;
+            }
+        }
+        col += step_size * n_elem_per_thread;
+    }
+
+    // Reduce divisor within CTA
+    tmp_expf = two_stage_warp_reduce_sum(tmp_expf);
+
+    // Store CTA-level sum to GMEM
+    if (tid == 0) {
+        tmp_sums[blockIdx.x] = tmp_expf;
+    }
+    g.sync();
+
+    // Compute global sum from CTA-level sums
+    if (tid < gridDim.x) {
+        tmp_expf = tmp_sums[tid];
+    } else {
+        tmp_expf = 0.0f;
+    }
+    tmp_expf = two_stage_warp_reduce_sum(tmp_expf);
+
+    // Divide dividend by global sum + store data
+    for (int col = col_start; col < p.ncols;) {
+#pragma unroll
+        for (int i = 0; i < n_elem_per_thread; i++) {
+            const int idx = col + i * step_size;
+            local_vals[i] = idx < p.ncols ? dst[idx] : -INFINITY;
+        }
+#pragma unroll
+        for (int i = 0; i < n_elem_per_thread; i++) {
+            const int idx = col + i * step_size;
+            if (idx < p.ncols) {
+                dst[idx] = local_vals[i] / tmp_expf;
+            }
+        }
+        col += step_size * n_elem_per_thread;
+    }
+}
+
 #ifdef __clang__
 #pragma clang diagnostic pop
 #endif // __clang__
@@ -216,9 +374,31 @@ static void launch_soft_max_kernels(const float * x, const T * mask, const float
     soft_max_f32<true, 0, 0><<<block_nums, block_dims, nbytes_shared, stream>>>(x, mask, sinks, dst, p);
 }
 
+__launch_bounds__(8*WARP_SIZE, 1) static __global__ void soft_max_f32_parallelize_cols(const float * __restrict__ x,
+                                                     float * __restrict__ dst,
+                                                     float * __restrict__ tmp_maxs,
+                                                     float * __restrict__ tmp_sums,
+                                                     const soft_max_params p)
+// We loop over all instead of parallelizing across gridDim.y as cooperative groups
+// currently only support synchronizing the complete grid if not launched as a cluster group
+// (which requires CC > 9.0)
+// https://docs.nvidia.com/cuda/cuda-programming-guide/05-appendices/device-callable-apis.html#grid-synchronization
+// https://docs.nvidia.com/cuda/cuda-programming-guide/05-appendices/device-callable-apis.html#class-cluster-group
+{
+    for (int rowx = 0; rowx < p.ne01 * p.ne02 * p.ne03; rowx++) {
+        soft_max_f32_parallelize_cols_single_row(x + int64_t(rowx) * p.ncols, dst + int64_t(rowx) * p.ncols, tmp_maxs,
+                                                 tmp_sums, p);
+    }
+}
 
-template<typename T>
-static void soft_max_f32_cuda(const float * x, const T * mask, const float * sinks, float * dst, const soft_max_params & params, cudaStream_t stream) {
+template <typename T>
+static void soft_max_f32_cuda(const float *                                x,
+                              const T *                                    mask,
+                              const float *                                sinks,
+                              float *                                      dst,
+                              const soft_max_params &                      params,
+                              cudaStream_t                                 stream,
+                              [[maybe_unused]] ggml_backend_cuda_context & ctx) {
     int nth = WARP_SIZE;
     const int64_t ncols_x = params.ncols;
 
@@ -236,8 +416,25 @@ static void soft_max_f32_cuda(const float * x, const T * mask, const float * sin
     if (nbytes_shared <= smpbo) {
         launch_soft_max_kernels<32, 64, 128, 256, 512, 1024, 2048, 4096>(x, mask, sinks, dst, params, stream, block_dims, block_nums, nbytes_shared);
     } else {
-        const size_t nbytes_shared_low = WARP_SIZE*sizeof(float);
-        soft_max_f32<false, 0, 0><<<block_nums, block_dims, nbytes_shared_low, stream>>>(x, mask, sinks, dst, params);
+        // Parallelize across SMs for top-p/dist-sampling
+        // The heuristic for parallelizing rows across SMs vs parallelizing single row & looping over all rows was done on the basis of a B6000 GPU and
+        // Can be adapted further for lower-SM-count GPUs, though keeping data in registers should be implemented first as that is the optimal solution.
+        if (ggml_cuda_info().devices[id].supports_cooperative_launch &&
+            ncols_x / (params.ne01 * params.ne02 * params.ne03) > 8192 && mask == nullptr && sinks == nullptr &&
+            params.scale == 1.0f && params.max_bias == 0.0f) {
+            ggml_cuda_pool_alloc<float> tmp_maxs_alloc(ctx.pool(), ggml_cuda_info().devices[id].nsm * sizeof(float));
+            ggml_cuda_pool_alloc<float> tmp_sums_alloc(ctx.pool(), ggml_cuda_info().devices[id].nsm * sizeof(float));
+
+            void * kernel_args[] = { (void *) &x, (void *) &dst, (void *) &tmp_maxs_alloc.ptr,
+                                     (void *) &tmp_sums_alloc.ptr, (void *) const_cast<soft_max_params *>(&params) };
+            CUDA_CHECK(cudaLaunchCooperativeKernel((void *) soft_max_f32_parallelize_cols,
+                                                   dim3(ggml_cuda_info().devices[id].nsm, 1, 1),
+                                                   dim3(WARP_SIZE * 8, 1, 1), kernel_args, 0, stream));
+        } else {
+            const size_t nbytes_shared_low = WARP_SIZE * sizeof(float);
+            soft_max_f32<false, 0, 0>
+                <<<block_nums, block_dims, nbytes_shared_low, stream>>>(x, mask, sinks, dst, params);
+        }
     }
 }
 
@@ -315,9 +512,9 @@ void ggml_cuda_op_soft_max(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     params.m1 = m1;
 
     if (use_f16) {
-        soft_max_f32_cuda(src0_d, (const half  *) src1_d, (const float *) src2_d, dst_d, params, stream);
+        soft_max_f32_cuda(src0_d, (const half *) src1_d, (const float *) src2_d, dst_d, params, stream, ctx);
     } else {
-        soft_max_f32_cuda(src0_d, (const float *) src1_d, (const float *) src2_d, dst_d, params, stream);
+        soft_max_f32_cuda(src0_d, (const float *) src1_d, (const float *) src2_d, dst_d, params, stream, ctx);
     }
 }
 

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

@@ -0,0 +1,275 @@
+#include "common.cuh"
+#include "ggml.h"
+#include "solve_tri.cuh"
+
+#define MAX_N_FAST 64
+#define MAX_K_FAST 32
+
+static __global__ void get_batch_pointers(const float *  A,
+                                          float *        X,
+                                          const float ** A_ptrs,
+                                          float **       X_ptrs,
+                                          int64_t        ne02,
+                                          int64_t        total_batches,
+                                          size_t         s02,
+                                          size_t         s03,
+                                          size_t         s2,
+                                          size_t         s3) {
+    const int idx = blockIdx.x * blockDim.x + threadIdx.x;
+    if (idx >= total_batches) {
+        return;
+    }
+
+    const int64_t i3 = idx / ne02;
+    const int64_t i2 = idx % ne02;
+
+    A_ptrs[idx] = A + i3 * s03 + i2 * s02;
+    X_ptrs[idx] = X + i3 * s3 + i2 * s2;
+}
+
+static void solve_tri_f32_cublas(ggml_backend_cuda_context & ctx,
+                                 const float *               A,
+                                 const float *               B,
+                                 float *                     X,
+                                 int                         n,
+                                 int                         k,
+                                 int64_t                     ne02,
+                                 int64_t                     ne03,
+                                 size_t                      s02,
+                                 size_t                      s03,
+                                 size_t                      s12,
+                                 size_t                      s13,
+                                 size_t                      s2,
+                                 size_t                      s3,
+                                 cudaStream_t                stream) {
+    const float   alpha         = 1.0f;
+    const int64_t total_batches = ne02 * ne03;
+    if (total_batches == 0) {
+        return;
+    }
+
+    // Bulk copy B -> X (contiguous tensors)
+    if (X != B) {
+        const int64_t total_elements_BX = n * k * total_batches;
+        CUDA_CHECK(cudaMemcpyAsync(X, B, total_elements_BX * sizeof(float), cudaMemcpyDeviceToDevice, stream));
+    }
+
+    const int id = ggml_cuda_get_device();
+
+    ggml_cuda_pool_alloc<const float *> A_ptrs_alloc(ctx.pool(id), total_batches);
+    ggml_cuda_pool_alloc<float *>       X_ptrs_alloc(ctx.pool(id), total_batches);
+
+    const float ** A_ptrs_dev = A_ptrs_alloc.get();
+    float **       X_ptrs_dev = X_ptrs_alloc.get();
+
+    get_batch_pointers<<<(total_batches + 255) / 256, 256, 0, stream>>>(A, X, A_ptrs_dev, X_ptrs_dev, ne02,
+                                                                        total_batches, s02, s03, s2, s3);
+
+    CUBLAS_CHECK(cublasSetStream(ctx.cublas_handle(id), stream));
+
+    // Yes, this is necessary, without this we get RMSE errors
+    CUBLAS_CHECK(cublasSetMathMode(ctx.cublas_handle(id), CUBLAS_DEFAULT_MATH));
+    CUBLAS_CHECK(cublasStrsmBatched(ctx.cublas_handle(id), CUBLAS_SIDE_RIGHT, CUBLAS_FILL_MODE_UPPER, CUBLAS_OP_N,
+                                    CUBLAS_DIAG_NON_UNIT, k, n, &alpha, A_ptrs_dev, n, X_ptrs_dev, k, total_batches));
+
+    // revert to standard mode from common.cuh
+    CUBLAS_CHECK(cublasSetMathMode(ctx.cublas_handle(id), CUBLAS_TF32_TENSOR_OP_MATH));
+
+    GGML_UNUSED_VARS(s12, s13);
+}
+
+// ======================
+// Fast Kernel (n <= 64, k <= 32) - Warp-based parallel reduction
+// ======================
+// When ncols_template == 0 the bounds for the loops in this function are not
+// known and can't be unrolled. As we want to keep pragma unroll for all other
+// cases we supress the clang transformation warning here.
+#ifdef __clang__
+#    pragma clang diagnostic push
+#    pragma clang diagnostic ignored "-Wpass-failed"
+#endif  // __clang__
+template <int n_template, int k_template>
+static __global__ void solve_tri_f32_fast(const float * __restrict__ A,
+                                          const float * __restrict__ B,
+                                          float * __restrict__ X,
+                                          const uint3  ne02,
+                                          const size_t nb02,
+                                          const size_t nb03,
+                                          const size_t nb12,
+                                          const size_t nb13,
+                                          const size_t nb2,
+                                          const size_t nb3,
+                                          const int    n_arg,
+                                          const int    k_arg) {
+    const int n = n_template == 0 ? n_arg : n_template;
+    const int k = k_template == 0 ? k_arg : k_template;
+
+    const int batch_idx = blockIdx.x;
+    const int lane      = threadIdx.x;
+    const int col_idx   = threadIdx.y;
+
+    if (col_idx >= k) {
+        return;
+    }
+
+    const uint2   i02_i03 = fast_div_modulo(batch_idx, ne02);
+    const int64_t i02     = i02_i03.y;
+    const int64_t i03     = i02_i03.x;
+
+    const float * const A_batch = (const float *) (A + i02 * nb02 + i03 * nb03);
+    const float * const B_batch = (const float *) (B + i02 * nb12 + i03 * nb13);
+    float *             X_batch = (float *) (X + i02 * nb2 + i03 * nb3);
+
+    __shared__ float sA[MAX_N_FAST * MAX_N_FAST];
+
+    const int offset = threadIdx.x + threadIdx.y * blockDim.x;
+
+#pragma unroll
+    for (int i = 0; i < n * n; i += k * WARP_SIZE) {
+        const int i0 = i + offset;
+        if (i0 < n * n) {
+            sA[i0] = A_batch[i0];
+        }
+    }
+
+    __syncthreads();
+
+    float x_low  = (lane < n) ? B_batch[lane * k + col_idx] : 0.0f;
+    float x_high = (WARP_SIZE + lane < n) ? B_batch[(WARP_SIZE + lane) * k + col_idx] : 0.0f;
+
+    const int half      = WARP_SIZE;
+    const int nrows_low = (n < half) ? n : half;
+
+#pragma unroll
+    for (int row = 0; row < nrows_low; ++row) {
+        float sum = 0.0f;
+        if (lane < row) {
+            sum += sA[row * n + lane] * x_low;
+        }
+        sum = warp_reduce_sum(sum);
+
+        if (lane == row) {
+            x_low = (x_low - sum) / sA[row * n + row];
+        }
+    }
+
+#pragma unroll
+    for (int row = half; row < n; ++row) {
+        float     sum = sA[row * n + lane] * x_low;
+        const int j   = half + lane;
+        if (j < row) {
+            sum += sA[row * n + j] * x_high;
+        }
+        sum = warp_reduce_sum(sum);
+
+        if (lane == row - half) {
+            x_high = (x_high - sum) / sA[row * n + row];
+        }
+    }
+
+#pragma unroll
+    for (int rr = 0; rr < 2; ++rr) {
+        const int row = rr * WARP_SIZE + lane;
+        if (row < n) {
+            const float val            = (row < half) ? x_low : x_high;
+            X_batch[row * k + col_idx] = val;
+        }
+    }
+}
+#ifdef __clang__
+#    pragma clang diagnostic pop
+#endif  // __clang__
+
+static void solve_tri_f32_cuda(const float * A,
+                               const float * B,
+                               float *       X,
+                               int           n,
+                               int           k,
+                               int64_t       ne02,
+                               int64_t       ne03,
+                               size_t        nb02,
+                               size_t        nb03,
+                               size_t        nb12,
+                               size_t        nb13,
+                               size_t        nb2,
+                               size_t        nb3,
+                               cudaStream_t  stream) {
+    const uint3 ne02_fd = init_fastdiv_values((uint32_t) ne02);
+    dim3        threads(WARP_SIZE, k);
+    dim3        grid(ne02 * ne03);
+    if (n == 64) {
+        switch (k) {
+            case 32:
+                solve_tri_f32_fast<64, 32>
+                    <<<grid, threads, 0, stream>>>(A, B, X, ne02_fd, nb02, nb03, nb12, nb13, nb2, nb3, 0, 0);
+                break;
+            case 16:
+                solve_tri_f32_fast<64, 16>
+                    <<<grid, threads, 0, stream>>>(A, B, X, ne02_fd, nb02, nb03, nb12, nb13, nb2, nb3, 0, 0);
+                break;
+            case 14:
+                solve_tri_f32_fast<64, 14>
+                    <<<grid, threads, 0, stream>>>(A, B, X, ne02_fd, nb02, nb03, nb12, nb13, nb2, nb3, 0, 0);
+                break;
+            case 12:
+                solve_tri_f32_fast<64, 12>
+                    <<<grid, threads, 0, stream>>>(A, B, X, ne02_fd, nb02, nb03, nb12, nb13, nb2, nb3, 0, 0);
+                break;
+            case 10:
+                solve_tri_f32_fast<64, 10>
+                    <<<grid, threads, 0, stream>>>(A, B, X, ne02_fd, nb02, nb03, nb12, nb13, nb2, nb3, 0, 0);
+                break;
+            case 8:
+                solve_tri_f32_fast<64, 8>
+                    <<<grid, threads, 0, stream>>>(A, B, X, ne02_fd, nb02, nb03, nb12, nb13, nb2, nb3, 0, 0);
+                break;
+            case 6:
+                solve_tri_f32_fast<64, 6>
+                    <<<grid, threads, 0, stream>>>(A, B, X, ne02_fd, nb02, nb03, nb12, nb13, nb2, nb3, 0, 0);
+                break;
+            case 4:
+                solve_tri_f32_fast<64, 4>
+                    <<<grid, threads, 0, stream>>>(A, B, X, ne02_fd, nb02, nb03, nb12, nb13, nb2, nb3, 0, 0);
+                break;
+            case 2:
+                solve_tri_f32_fast<64, 2>
+                    <<<grid, threads, 0, stream>>>(A, B, X, ne02_fd, nb02, nb03, nb12, nb13, nb2, nb3, 0, 0);
+                break;
+            case 1:
+                solve_tri_f32_fast<64, 1>
+                    <<<grid, threads, 0, stream>>>(A, B, X, ne02_fd, nb02, nb03, nb12, nb13, nb2, nb3, 0, 0);
+                break;
+            default:
+                solve_tri_f32_fast<0, 0>
+                    <<<grid, threads, 0, stream>>>(A, B, X, ne02_fd, nb02, nb03, nb12, nb13, nb2, nb3, n, k);
+        }
+    } else {  // run general case
+        solve_tri_f32_fast<0, 0>
+            <<<grid, threads, 0, stream>>>(A, B, X, ne02_fd, nb02, nb03, nb12, nb13, nb2, nb3, n, k);
+    }
+}
+
+void ggml_cuda_op_solve_tri(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];  // A (n×n, lower triangular)
+    const ggml_tensor * src1 = dst->src[1];  // B (n×k)
+
+    ggml_is_contiguous(src0);
+    ggml_is_contiguous(src1);
+
+    const int64_t n    = src0->ne[0];
+    const int64_t k    = src1->ne[0];
+    const int64_t ne02 = src0->ne[2];
+    const int64_t ne03 = src0->ne[3];
+
+    if (n <= MAX_N_FAST && k <= MAX_K_FAST) {
+        solve_tri_f32_cuda((const float *) src0->data, (const float *) src1->data, (float *) dst->data, n, k,
+                           src0->ne[2], src0->ne[3], src0->nb[2] / sizeof(float), src0->nb[3] / sizeof(float),
+                           src1->nb[2] / sizeof(float), src1->nb[3] / sizeof(float), dst->nb[2] / sizeof(float),
+                           dst->nb[3] / sizeof(float), ctx.stream());
+    } else {
+        solve_tri_f32_cublas(ctx, (const float *) src0->data, (const float *) src1->data, (float *) dst->data, n, k,
+                             ne02, ne03, src0->nb[2] / sizeof(float), src0->nb[3] / sizeof(float),
+                             src1->nb[2] / sizeof(float), src1->nb[3] / sizeof(float), dst->nb[2] / sizeof(float),
+                             dst->nb[3] / sizeof(float), ctx.stream());
+    }
+}

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

@@ -0,0 +1,3 @@
+#include "common.cuh"
+
+void ggml_cuda_op_solve_tri(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

data/ext/sources/ggml/src/ggml-cuda/ssm-conv.cu

@@ -102,31 +102,25 @@ static void ssm_conv_f32_cuda(const float * src0, const float * src1, const int
     const int threads = 128;
     GGML_ASSERT(nr % threads == 0);
 
-    if (n_t <= 32) {
-        const dim3 blocks(n_s, (nr + threads - 1) / threads, 1);
-        if (nc == 4) {
-            ssm_conv_f32<threads, 4><<<blocks, threads, 0, stream>>>(src0, src1, src0_nb0, src0_nb1, src0_nb2, src1_nb1,
-                                                                     dst, dst_nb0, dst_nb1, dst_nb2, n_t);
-        } else if (nc == 3) {
-            ssm_conv_f32<threads, 3><<<blocks, threads, 0, stream>>>(src0, src1, src0_nb0, src0_nb1, src0_nb2, src1_nb1,
-                                                                     dst, dst_nb0, dst_nb1, dst_nb2, n_t);
+    auto launch_kernel = [&](auto NC) {
+        constexpr int kNC = decltype(NC)::value;
+        if (n_t <= 32) {
+            const dim3 blocks(n_s, (nr + threads - 1) / threads, 1);
+            ssm_conv_f32<threads, kNC><<<blocks, threads, 0, stream>>>(src0, src1, src0_nb0, src0_nb1, src0_nb2, src1_nb1,
+                                                                       dst, dst_nb0, dst_nb1, dst_nb2, n_t);
         } else {
-            GGML_ABORT("Only support kernel size = 3 or size = 4 right now.");
-        }
-    } else {
-        if (nc == 4) {
-            const int64_t split_n_t = 32;
-            dim3          blocks(n_s, (nr + threads - 1) / threads, (n_t + split_n_t - 1) / split_n_t);
-            ssm_conv_long_token_f32<threads, 4, split_n_t><<<blocks, threads, 0, stream>>>(
-                src0, src1, src0_nb0, src0_nb1, src0_nb2, src1_nb1, dst, dst_nb0, dst_nb1, dst_nb2, n_t);
-        } else if (nc == 3) {
             const int64_t split_n_t = 32;
             dim3          blocks(n_s, (nr + threads - 1) / threads, (n_t + split_n_t - 1) / split_n_t);
-            ssm_conv_long_token_f32<threads, 3, split_n_t><<<blocks, threads, 0, stream>>>(
+            ssm_conv_long_token_f32<threads, kNC, split_n_t><<<blocks, threads, 0, stream>>>(
                 src0, src1, src0_nb0, src0_nb1, src0_nb2, src1_nb1, dst, dst_nb0, dst_nb1, dst_nb2, n_t);
-        } else {
-            GGML_ABORT("Only support kernel size = 3 or size = 4 right now.");
         }
+    };
+
+    switch (nc) {
+        case 3: launch_kernel(std::integral_constant<int, 3>{}); break;
+        case 4: launch_kernel(std::integral_constant<int, 4>{}); break;
+        case 9: launch_kernel(std::integral_constant<int, 9>{}); break;
+        default: GGML_ABORT("Only support kernel sizes 3, 4, 9 right now.");
     }
 }