whispercpp 1.3.6 → 1.3.8

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

@@ -1,13 +1,16 @@
 #include "gated_delta_net.cuh"
+#include "ggml-cuda/common.cuh"
 
-template <int S_v, bool KDA>
-__global__ void gated_delta_net_cuda(const float * q,
+template <int S_v, bool KDA, bool keep_rs_t>
+__global__ void __launch_bounds__((ggml_cuda_get_physical_warp_size() < S_v ? ggml_cuda_get_physical_warp_size() : S_v) * 4, 2)
+gated_delta_net_cuda(const float * q,
                                      const float * k,
                                      const float * v,
                                      const float * g,
                                      const float * beta,
                                      const float * curr_state,
                                      float *       dst,
+                                     float *       state,
                                      int64_t       H,
                                      int64_t       n_tokens,
                                      int64_t       n_seqs,
@@ -22,7 +25,9 @@ __global__ void gated_delta_net_cuda(const float * q,
                                      int64_t       sb3,
                                      const uint3   neqk1_magic,
                                      const uint3   rq3_magic,
-                                     float         scale) {
+                                     float         scale,
+                                     int64_t       state_slot_stride,
+                                     int           K) {
     const uint32_t h_idx    = blockIdx.x;
     const uint32_t sequence = blockIdx.y;
     // each warp owns one column, using warp-level primitives to reduce across rows
@@ -32,13 +37,14 @@ __global__ void gated_delta_net_cuda(const float * q,
     const uint32_t iq1 = fastmodulo(h_idx, neqk1_magic);
     const uint32_t iq3 = fastdiv(sequence, rq3_magic);
 
-    const int64_t attn_score_elems = S_v * H * n_tokens * n_seqs;
     float *       attn_data        = dst;
-    float *       state            = dst + attn_score_elems;
 
-    const int64_t state_offset = (sequence * H + h_idx) * S_v * S_v;
-    state += state_offset;
-    curr_state += state_offset;
+    // input state holds s0 only: [S_v, S_v, H, n_seqs] — seq stride is D = H * S_v * S_v.
+    // output state layout (per-slot D * n_seqs) — same per-(seq,head) offset as before.
+    const int64_t state_in_offset      = sequence * H * S_v * S_v + h_idx * S_v * S_v;
+    const int64_t state_out_offset     = (sequence * H + h_idx) * S_v * S_v;
+    state += state_out_offset;
+    curr_state += state_in_offset + col * S_v;
     attn_data += (sequence * n_tokens * H + h_idx) * S_v;
 
     constexpr int warp_size = ggml_cuda_get_physical_warp_size() < S_v ? ggml_cuda_get_physical_warp_size() : S_v;
@@ -46,10 +52,12 @@ __global__ void gated_delta_net_cuda(const float * q,
     constexpr int rows_per_lane = (S_v + warp_size - 1) / warp_size;
     float         s_shard[rows_per_lane];
     // state is stored transposed: M[col][i] = S[i][col], row col is contiguous
+
+    ggml_cuda_pdl_sync();
 #pragma unroll
     for (int r = 0; r < rows_per_lane; r++) {
         const int i = r * warp_size + lane;
-        s_shard[r]  = curr_state[col * S_v + i];
+        s_shard[r]  = curr_state[i];
     }
 
     for (int t = 0; t < n_tokens; t++) {
@@ -63,6 +71,16 @@ __global__ void gated_delta_net_cuda(const float * q,
 
         const float beta_val = *beta_t;
 
+        // Cache k and q in registers
+        float k_reg[rows_per_lane];
+        float q_reg[rows_per_lane];
+#pragma unroll
+        for (int r = 0; r < rows_per_lane; r++) {
+            const int i = r * warp_size + lane;
+            k_reg[r] = k_t[i];
+            q_reg[r] = q_t[i];
+        }
+
         if constexpr (!KDA) {
             const float g_val = expf(*g_t);
 
@@ -70,8 +88,7 @@ __global__ void gated_delta_net_cuda(const float * q,
             float kv_shard = 0.0f;
 #pragma unroll
             for (int r = 0; r < rows_per_lane; r++) {
-                const int i = r * warp_size + lane;
-                kv_shard += s_shard[r] * k_t[i];
+                kv_shard += s_shard[r] * k_reg[r];
             }
             float kv_col = warp_reduce_sum<warp_size>(kv_shard);
 
@@ -83,9 +100,8 @@ __global__ void gated_delta_net_cuda(const float * q,
             float attn_partial = 0.0f;
 #pragma unroll
             for (int r = 0; r < rows_per_lane; r++) {
-                const int i = r * warp_size + lane;
-                s_shard[r]  = g_val * s_shard[r] + k_t[i] * delta_col;
-                attn_partial += s_shard[r] * q_t[i];
+                s_shard[r]  = g_val * s_shard[r] + k_reg[r] * delta_col;
+                attn_partial += s_shard[r] * q_reg[r];
             }
 
             float attn_col = warp_reduce_sum<warp_size>(attn_partial);
@@ -99,7 +115,7 @@ __global__ void gated_delta_net_cuda(const float * q,
 #pragma unroll
             for (int r = 0; r < rows_per_lane; r++) {
                 const int i = r * warp_size + lane;
-                kv_shard += expf(g_t[i]) * s_shard[r] * k_t[i];
+                kv_shard += expf(g_t[i]) * s_shard[r] * k_reg[r];
             }
 
             float kv_col = warp_reduce_sum<warp_size>(kv_shard);
@@ -113,8 +129,8 @@ __global__ void gated_delta_net_cuda(const float * q,
 #pragma unroll
             for (int r = 0; r < rows_per_lane; r++) {
                 const int i = r * warp_size + lane;
-                s_shard[r]  = expf(g_t[i]) * s_shard[r] + k_t[i] * delta_col;
-                attn_partial += s_shard[r] * q_t[i];
+                s_shard[r]  = expf(g_t[i]) * s_shard[r] + k_reg[r] * delta_col;
+                attn_partial += s_shard[r] * q_reg[r];
             }
 
             float attn_col = warp_reduce_sum<warp_size>(attn_partial);
@@ -125,27 +141,42 @@ __global__ void gated_delta_net_cuda(const float * q,
         }
 
         attn_data += S_v * H;
+
+        if constexpr (keep_rs_t) {
+            // snapshot slot mapping: slot 0 = most recent state, slot s = s tokens back.
+            // When n_tokens < K only slots 0..n_tokens-1 are written; older slots are caller-owned.
+            const int target_slot = (int) n_tokens - 1 - t;
+            if (target_slot >= 0 && target_slot < K) {
+                float * curr_state = state + target_slot * state_slot_stride;
+#pragma unroll
+                for (int r = 0; r < rows_per_lane; r++) {
+                    const int i = r * warp_size + lane;
+                    curr_state[col * S_v + i] = s_shard[r];
+                }
+            }
+        }
     }
 
-    // Write state back to global memory (transposed layout)
+    if constexpr (!keep_rs_t) {
 #pragma unroll
-    for (int r = 0; r < rows_per_lane; r++) {
-        const int i          = r * warp_size + lane;
-        state[col * S_v + i] = s_shard[r];
+        for (int r = 0; r < rows_per_lane; r++) {
+            const int i          = r * warp_size + lane;
+            state[col * S_v + i] = s_shard[r];
+        }
     }
 }
 
-template <bool KDA>
+template <bool KDA, bool keep_rs_t>
 static void launch_gated_delta_net(
         const float * q_d, const float * k_d, const float * v_d,
         const float * g_d, const float * b_d, const float * s_d,
-        float * dst_d,
+        float * dst_d, float * state_d,
         int64_t S_v,   int64_t H, int64_t n_tokens, int64_t n_seqs,
         int64_t sq1,   int64_t sq2, int64_t sq3,
         int64_t sv1,   int64_t sv2, int64_t sv3,
         int64_t sb1,   int64_t sb2, int64_t sb3,
         int64_t neqk1, int64_t rq3,
-        float scale, cudaStream_t stream) {
+        float scale, int64_t state_slot_stride, int K, cudaStream_t stream) {
     //TODO: Add chunked kernel for even faster pre-fill
     const int warp_size = ggml_cuda_info().devices[ggml_cuda_get_device()].warp_size;
     const int num_warps = 4;
@@ -155,33 +186,32 @@ static void launch_gated_delta_net(
     const uint3 neqk1_magic = init_fastdiv_values(neqk1);
     const uint3 rq3_magic   = init_fastdiv_values(rq3);
 
-    int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
-
+    const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(grid_dims, block_dims, 0, stream);
     switch (S_v) {
         case 16:
-            gated_delta_net_cuda<16, KDA><<<grid_dims, block_dims, 0, stream>>>(
-                q_d, k_d, v_d, g_d, b_d, s_d, dst_d, H,
+            ggml_cuda_kernel_launch(gated_delta_net_cuda<16, KDA, keep_rs_t>, launch_params,
+                q_d, k_d, v_d, g_d, b_d, s_d, dst_d, state_d, H,
                 n_tokens, n_seqs, sq1, sq2, sq3, sv1, sv2, sv3,
-                sb1, sb2, sb3, neqk1_magic, rq3_magic, scale);
+                sb1, sb2, sb3, neqk1_magic, rq3_magic, scale, state_slot_stride, K);
             break;
         case 32:
-            gated_delta_net_cuda<32, KDA><<<grid_dims, block_dims, 0, stream>>>(
-                q_d, k_d, v_d, g_d, b_d, s_d, dst_d, H,
+            ggml_cuda_kernel_launch(gated_delta_net_cuda<32, KDA, keep_rs_t>, launch_params,
+                q_d, k_d, v_d, g_d, b_d, s_d, dst_d, state_d, H,
                 n_tokens, n_seqs, sq1, sq2, sq3, sv1, sv2, sv3,
-                sb1, sb2, sb3, neqk1_magic, rq3_magic, scale);
+                sb1, sb2, sb3, neqk1_magic, rq3_magic, scale, state_slot_stride, K);
             break;
         case 64: {
-            gated_delta_net_cuda<64, KDA><<<grid_dims, block_dims, 0, stream>>>(
-                q_d, k_d, v_d, g_d, b_d, s_d, dst_d, H,
+            ggml_cuda_kernel_launch(gated_delta_net_cuda<64, KDA, keep_rs_t>, launch_params,
+                q_d, k_d, v_d, g_d, b_d, s_d, dst_d, state_d, H,
                 n_tokens, n_seqs, sq1, sq2, sq3, sv1, sv2, sv3,
-                sb1, sb2, sb3, neqk1_magic, rq3_magic, scale);
+                sb1, sb2, sb3, neqk1_magic, rq3_magic, scale, state_slot_stride, K);
             break;
         }
         case 128: {
-            gated_delta_net_cuda<128, KDA><<<grid_dims, block_dims, 0, stream>>>(
-                q_d, k_d, v_d, g_d, b_d, s_d, dst_d, H,
+            ggml_cuda_kernel_launch(gated_delta_net_cuda<128, KDA, keep_rs_t>, launch_params,
+                q_d, k_d, v_d, g_d, b_d, s_d, dst_d, state_d, H,
                 n_tokens, n_seqs, sq1, sq2, sq3, sv1, sv2, sv3,
-                sb1, sb2, sb3, neqk1_magic, rq3_magic, scale);
+                sb1, sb2, sb3, neqk1_magic, rq3_magic, scale, state_slot_stride, K);
             break;
         }
         default:
@@ -190,7 +220,8 @@ static void launch_gated_delta_net(
     }
 }
 
-void ggml_cuda_op_gated_delta_net(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+static void ggml_cuda_op_gated_delta_net_impl(
+        ggml_backend_cuda_context & ctx, ggml_tensor * dst, const ggml_cuda_gated_delta_net_fused_cache * cache) {
     ggml_tensor * src_q     = dst->src[0];
     ggml_tensor * src_k     = dst->src[1];
     ggml_tensor * src_v     = dst->src[2];
@@ -251,13 +282,46 @@ void ggml_cuda_op_gated_delta_net(ggml_backend_cuda_context & ctx, ggml_tensor *
 
     cudaStream_t stream = ctx.stream();
 
+    // K (snapshot slot count) is an op param; state holds s0 only [S_v, S_v, H, n_seqs].
+    const int K = ggml_get_op_params_i32(dst, 0);
+    const bool keep_rs = K > 1;
+
+    // recurrent state -> gdn_out tail (after attention scores), or the cache when fusing
+    float * state_d           = dst_d + S_v * H * n_tokens * n_seqs;
+    int64_t state_slot_stride = S_v * S_v * H * n_seqs;
+    if (cache != nullptr) {
+        state_d           = cache->data;
+        state_slot_stride = cache->slot_stride;
+    }
+
     if (kda) {
-        launch_gated_delta_net<true>(q_d, k_d, v_d, g_d, b_d, s_d, dst_d,
-            S_v, H, n_tokens, n_seqs, sq1, sq2, sq3, sv1, sv2, sv3,
-            sb1, sb2, sb3, neqk1, rq3, scale, stream);
+        if (keep_rs) {
+            launch_gated_delta_net<true, true>(q_d, k_d, v_d, g_d, b_d, s_d, dst_d, state_d,
+                S_v, H, n_tokens, n_seqs, sq1, sq2, sq3, sv1, sv2, sv3,
+                sb1, sb2, sb3, neqk1, rq3, scale, state_slot_stride, K, stream);
+        } else {
+            launch_gated_delta_net<true, false>(q_d, k_d, v_d, g_d, b_d, s_d, dst_d, state_d,
+                S_v, H, n_tokens, n_seqs, sq1, sq2, sq3, sv1, sv2, sv3,
+                sb1, sb2, sb3, neqk1, rq3, scale, state_slot_stride, K, stream);
+        }
     } else {
-        launch_gated_delta_net<false>(q_d, k_d, v_d, g_d, b_d, s_d, dst_d,
-            S_v, H, n_tokens, n_seqs, sq1, sq2, sq3, sv1, sv2, sv3,
-            sb1, sb2, sb3, neqk1, rq3, scale, stream);
+        if (keep_rs) {
+            launch_gated_delta_net<false, true>(q_d, k_d, v_d, g_d, b_d, s_d, dst_d, state_d,
+                S_v, H, n_tokens, n_seqs, sq1, sq2, sq3, sv1, sv2, sv3,
+                sb1, sb2, sb3, neqk1, rq3, scale, state_slot_stride, K, stream);
+        } else {
+            launch_gated_delta_net<false, false>(q_d, k_d, v_d, g_d, b_d, s_d, dst_d, state_d,
+                S_v, H, n_tokens, n_seqs, sq1, sq2, sq3, sv1, sv2, sv3,
+                sb1, sb2, sb3, neqk1, rq3, scale, state_slot_stride, K, stream);
+        }
     }
 }
+
+void ggml_cuda_op_gated_delta_net(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_gated_delta_net_impl(ctx, dst, nullptr);
+}
+
+void ggml_cuda_op_gated_delta_net_fused_cache(
+        ggml_backend_cuda_context & ctx, ggml_tensor * dst, ggml_cuda_gated_delta_net_fused_cache cache) {
+    ggml_cuda_op_gated_delta_net_impl(ctx, dst, &cache);
+}

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

@@ -1,4 +1,14 @@
 #include "common.cuh"
 #include "ggml.h"
 
+// fused-kernel recurrent-state output; strides in elements (per-seq stride is always D, set in-kernel)
+struct ggml_cuda_gated_delta_net_fused_cache {
+    float * data;        // rollback slot 0
+    int64_t slot_stride; // between rollback slots (0 when K==1)
+};
+
 void ggml_cuda_op_gated_delta_net(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+// same op, but writes the snapshot(s) into the cache instead of dst (see ggml_cuda_try_gdn_cache_fusion)
+void ggml_cuda_op_gated_delta_net_fused_cache(ggml_backend_cuda_context & ctx, ggml_tensor * dst,
+                                              ggml_cuda_gated_delta_net_fused_cache cache);

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

@@ -6,17 +6,19 @@ template<int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
 static __global__ void k_get_rows(
         const void * __restrict__ src0, const int32_t * __restrict__ src1, dst_t * __restrict__ dst,
         const int64_t ne00, /*const int64_t ne01, const int64_t ne02, const int64_t ne03,*/
-        /*const int64_t ne10,*/ const int64_t ne11, const int64_t ne12, /*const int64_t ne13,*/
+        /*const int64_t ne10,*/ const int64_t ne11, const uint3 ne12_fdv, /*const int64_t ne13,*/
         /*const size_t s0,*/ const size_t s1, const size_t s2, const size_t s3,
         /*const size_t nb00,*/ const size_t nb01, const size_t nb02, const size_t nb03,
         const size_t s10, const size_t s11, const size_t s12/*, const size_t s13*/) {
 
-    for (int64_t z = blockIdx.z; z < ne11*ne12; z += gridDim.z) {
+    ggml_cuda_pdl_sync();
+    for (int64_t z = blockIdx.z; z < ne11*(int64_t)ne12_fdv.z; z += gridDim.z) {
         for (int64_t i00 = 2*(blockIdx.y*blockDim.x + threadIdx.x); i00 < ne00; i00 += gridDim.y*blockDim.x) {
             // The x and y dimensions of the grid are swapped because the maximum allowed grid size for x is higher.
             const int i10 =  blockIdx.x;
-            const int i11 =  z / ne12; // TODO fastdiv
-            const int i12 =  z % ne12;
+            const uint2 dm  = fast_div_modulo((uint32_t)z, ne12_fdv);
+            const int i11 =  dm.x;
+            const int i12 =  dm.y;
 
             const int i01 = src1[i10*s10 + i11*s11 + i12*s12];
 
@@ -40,19 +42,25 @@ static __global__ void k_get_rows(
 
 template<typename src0_t, typename dst_t>
 static __global__ void k_get_rows_float(
-        const src0_t * __restrict__ src0, const int32_t * __restrict__ src1, dst_t * __restrict__ dst,
+        const src0_t * src0_ptr, const int32_t * src1_ptr, dst_t * dst_ptr,
         const int64_t ne00, /*const int64_t ne01, const int64_t ne02, const int64_t ne03,*/
-        /*const int64_t ne10,*/ const int64_t ne11, const int64_t ne12, /*const int64_t ne13,*/
+        /*const int64_t ne10,*/ const int64_t ne11, const uint3 ne12_fdv, /*const int64_t ne13,*/
         /*const size_t s0,*/ const size_t s1, const size_t s2, const size_t s3,
         /*const size_t nb00,*/ const size_t nb01, const size_t nb02, const size_t nb03,
         const size_t s10, const size_t s11, const size_t s12/*, const size_t s13*/) {
 
-    for (int64_t z = blockIdx.z; z < ne11*ne12; z += gridDim.z) {
+    ggml_cuda_pdl_lc();
+    const src0_t  * GGML_CUDA_RESTRICT src0 = src0_ptr;
+    const int32_t * GGML_CUDA_RESTRICT src1 = src1_ptr;
+    dst_t         * GGML_CUDA_RESTRICT dst  = dst_ptr;
+    ggml_cuda_pdl_sync();
+    for (int64_t z = blockIdx.z; z < ne11*(int64_t)ne12_fdv.z; z += gridDim.z) {
         for (int64_t i00 = blockIdx.y*blockDim.x + threadIdx.x; i00 < ne00; i00 += gridDim.y*blockDim.x) {
             // The x and y dimensions of the grid are swapped because the maximum allowed grid size for x is higher.
             const int i10 = blockIdx.x;
-            const int i11 = z / ne12; // TODO fastdiv
-            const int i12 = z % ne12;
+            const uint2 dm = fast_div_modulo((uint32_t)z, ne12_fdv);
+            const int i11 = dm.x;
+            const int i12 = dm.y;
 
             if (i00 >= ne00) {
                 return;
@@ -70,25 +78,29 @@ static __global__ void k_get_rows_float(
 
 template<typename grad_t, typename dst_t>
 static __global__ void k_get_rows_back_float(
-        const grad_t * __restrict__ grad, const int32_t * __restrict__ rows, dst_t * __restrict__ dst, const int64_t ncols, const int64_t nrows_grad) {
+        const grad_t * __restrict__ grad, const int32_t * __restrict__ rows, dst_t * __restrict__ dst,
+        const int64_t ncols, const int64_t nrows_grad, const int64_t nrows_dst) {
     const int col = blockIdx.x*blockDim.x + threadIdx.x;
 
     if (col >= ncols) {
         return;
     }
 
-    const int dst_row = blockIdx.y*blockDim.y + threadIdx.y;
+    ggml_cuda_pdl_sync();
 
-    float sum = 0.0f;
+    // grid.y is clamped to the CUDA grid limit, so stride over the destination rows
+    for (int64_t dst_row = blockIdx.y; dst_row < nrows_dst; dst_row += gridDim.y) {
+        float sum = 0.0f;
 
-    for (int64_t i = 0; i < nrows_grad; ++i) {
-        if (rows[i] != dst_row) {
-            continue;
+        for (int64_t i = 0; i < nrows_grad; ++i) {
+            if (rows[i] != dst_row) {
+                continue;
+            }
+            sum += grad[i*ncols + col];
         }
-        sum += grad[i*ncols + col];
-    }
 
-    dst[dst_row*ncols + col] = sum;
+        dst[dst_row*ncols + col] = sum;
+    }
 }
 
 template<int qk, int qr, dequantize_kernel_t dq, typename dst_t>
@@ -115,10 +127,14 @@ static void get_rows_cuda_q(
 
     GGML_ASSERT(ne00 % 2 == 0);
 
+    GGML_ASSERT(ne12 > 0);
+    GGML_ASSERT(ne11 <= std::numeric_limits<uint32_t>::max() / ne12);
+    const uint3 ne12_fdv = init_fastdiv_values(ne12);
+
     k_get_rows<qk, qr, dq><<<block_nums, block_dims, 0, stream>>>(
         src0_d, src1_d, dst_d,
         ne00, /*ne01, ne02, ne03,*/
-        /*ne10,*/ ne11, ne12, /*ne13,*/
+        /*ne10,*/ ne11, ne12_fdv, /*ne13,*/
         /* s0,*/ s1, s2, s3,
         /* nb00,*/ nb01, nb02, nb03,
         s10, s11, s12/*, s13*/);
@@ -146,10 +162,15 @@ static void get_rows_cuda_float(
     const size_t s12 = nb12 / sizeof(int32_t);
     // const size_t s13 = nb13 / sizeof(int32_t);
 
-    k_get_rows_float<<<block_nums, block_dims, 0, stream>>>(
+    GGML_ASSERT(ne12 > 0);
+    GGML_ASSERT(ne11 <= std::numeric_limits<uint32_t>::max() / ne12);
+    const uint3 ne12_fdv = init_fastdiv_values(ne12);
+
+    const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params{block_nums, block_dims, 0, stream};
+    ggml_cuda_kernel_launch(k_get_rows_float<src0_t, dst_t>, launch_params,
         src0_d, src1_d, dst_d,
         ne00, /*ne01, ne02, ne03,*/
-        /*ne10,*/ ne11, ne12, /*ne13,*/
+        /*ne10,*/ ne11, ne12_fdv, /*ne13,*/
         /* s0,*/ s1, s2, s3,
         /* nb00,*/ nb01, nb02, nb03,
         s10, s11, s12/*, s13*/);
@@ -179,6 +200,10 @@ static void ggml_cuda_get_rows_switch_src0_type(
             get_rows_cuda_float((const nv_bfloat16 *) src0_d, src1_d, dst_d,
                 ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
             break;
+        case GGML_TYPE_Q1_0:
+            get_rows_cuda_q<QK1_0, QR1_0, dequantize_q1_0>(src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
         case GGML_TYPE_Q4_0:
             get_rows_cuda_q<QK4_0, QR4_0, dequantize_q4_0>(src0_d, src1_d, dst_d,
                 ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
@@ -280,7 +305,7 @@ void ggml_cuda_op_get_rows_back(ggml_backend_cuda_context & ctx, ggml_tensor * d
 
     const dim3 block_dims(CUDA_GET_ROWS_BACK_BLOCK_SIZE, 1, 1);
     const int block_num_x = (ne00 + CUDA_GET_ROWS_BACK_BLOCK_SIZE - 1) / CUDA_GET_ROWS_BACK_BLOCK_SIZE;
-    const dim3 block_nums(block_num_x, ne1, 1);
+    const dim3 block_nums(block_num_x, MIN(ne1, (int64_t)UINT16_MAX), 1);
 
-    k_get_rows_back_float<<<block_nums, block_dims, 0, stream>>>(src0_d, src1_d, dst_d, ne00, ne10);
+    k_get_rows_back_float<<<block_nums, block_dims, 0, stream>>>(src0_d, src1_d, dst_d, ne00, ne10, ne1);
 }