Merge e83d2707d3 into a5b57b08ce

* ggml: CUDA unary op EXP

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* ggml: rwkv_wkv op CUDA impl

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

---------

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

.github/workflows/build.yml

@@ -967,6 +967,7 @@ jobs:
           name: llama-bin-win-sycl-x64.zip
 
   windows-latest-cmake-hip:
+    if: ${{ github.event.inputs.create_release != 'true' }}
     runs-on: windows-latest
 
     steps:
@@ -994,8 +995,72 @@ jobs:
         run: |
           $env:HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)
           $env:CMAKE_PREFIX_PATH="${env:HIP_PATH}"
-          cmake -G "Unix Makefiles" -B build -S . -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" -DGGML_HIPBLAS=ON
-          cmake --build build --config Release
+          cmake -G "Unix Makefiles" -B build -S . -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" -DGGML_HIPBLAS=ON -DCMAKE_BUILD_TYPE=Release -DGGML_RPC=ON
+          cmake --build build -j ${env:NUMBER_OF_PROCESSORS}
+
+  windows-latest-cmake-hip-release:
+    if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+    runs-on: windows-latest
+
+    strategy:
+      matrix:
+        gpu_target: [gfx1100, gfx1101, gfx1030]
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: Install
+        id: depends
+        run: |
+          $ErrorActionPreference = "Stop"
+          write-host "Downloading AMD HIP SDK Installer"
+          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q3-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
+          write-host "Installing AMD HIP SDK"
+          Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -Wait
+          write-host "Completed AMD HIP SDK installation"
+
+      - name: Verify ROCm
+        id: verify
+        run: |
+          & 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' --version
+
+      - name: Build
+        id: cmake_build
+        run: |
+          $env:HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)
+          $env:CMAKE_PREFIX_PATH="${env:HIP_PATH}"
+          cmake -G "Unix Makefiles" -B build -S . -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" -DGGML_HIPBLAS=ON -DCMAKE_BUILD_TYPE=Release -DGPU_TARGETS=${{ matrix.gpu_target }} -DGGML_RPC=ON
+          cmake --build build -j ${env:NUMBER_OF_PROCESSORS}
+          md "build\bin\rocblas\library\"
+          cp "${env:HIP_PATH}\bin\hipblas.dll" "build\bin\"
+          cp "${env:HIP_PATH}\bin\rocblas.dll" "build\bin\"
+          cp "${env:HIP_PATH}\bin\rocblas\library\*" "build\bin\rocblas\library\"
+
+      - name: Determine tag name
+        id: tag
+        shell: bash
+        run: |
+          BUILD_NUMBER="$(git rev-list --count HEAD)"
+          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
+          if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
+            echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
+          else
+            SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
+            echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
+          fi
+
+      - name: Pack artifacts
+        id: pack_artifacts
+        run: |
+          7z a llama-${{ steps.tag.outputs.name }}-bin-win-hip-x64-${{ matrix.gpu_target }}.zip .\build\bin\*
+
+      - name: Upload artifacts
+        uses: actions/upload-artifact@v4
+        with:
+          path: llama-${{ steps.tag.outputs.name }}-bin-win-hip-x64-${{ matrix.gpu_target }}.zip
+          name: llama-bin-win-hip-x64-${{ matrix.gpu_target }}.zip
 
   ios-xcode-build:
     runs-on: macos-latest
@@ -1060,6 +1125,7 @@ jobs:
       - macOS-latest-cmake
       - windows-latest-cmake
       - windows-latest-cmake-cuda
+      - windows-latest-cmake-hip-release
       - macOS-latest-cmake-arm64
       - macOS-latest-cmake-x64
 

ggml/src/ggml-alloc.c

@@ -294,6 +294,12 @@ static void ggml_dyn_tallocr_reset(struct ggml_dyn_tallocr * alloc) {
     alloc->free_blocks[0].offset = 0;
     alloc->free_blocks[0].size = SIZE_MAX/2; // restrict maximum size of a measure allocator to half size_t max to avoid overflows
     alloc->max_size = 0;
+
+#ifdef GGML_ALLOCATOR_DEBUG
+    for (int i = 0; i < 1024; i++) {
+        alloc->allocated_tensors[i].tensor = NULL;
+    }
+#endif
 }
 
 static struct ggml_dyn_tallocr * ggml_dyn_tallocr_new(size_t alignment) {

ggml/src/ggml-cuda.cu

@@ -34,6 +34,7 @@
 #include "ggml-cuda/tsembd.cuh"
 #include "ggml-cuda/unary.cuh"
 #include "ggml-cuda/upscale.cuh"
+#include "ggml-cuda/rwkv-wkv.cuh"
 
 #include <algorithm>
 #include <array>
@@ -2243,6 +2244,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
                 case GGML_UNARY_OP_HARDSWISH:
                     ggml_cuda_op_hardswish(ctx, dst);
                     break;
+                case GGML_UNARY_OP_EXP:
+                    ggml_cuda_op_exp(ctx, dst);
+                    break;
                 default:
                     return false;
             }
@@ -2345,6 +2349,8 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_CROSS_ENTROPY_LOSS:
             ggml_cuda_cross_entropy_loss(ctx, dst);
             break;
+        case GGML_OP_RWKV_WKV:
+            ggml_cuda_op_rwkv_wkv(ctx, dst);
         case GGML_OP_CROSS_ENTROPY_LOSS_BACK:
             ggml_cuda_cross_entropy_loss_back(ctx, dst);
             break;
@@ -2478,6 +2484,7 @@ static void set_ggml_graph_node_properties(ggml_tensor * node, ggml_graph_node_p
     for (int i = 0; i < GGML_MAX_SRC; i++) {
         graph_node_properties->src_address[i] = node->src[i] ? node->src[i]->data : nullptr;
     }
+    memcpy(graph_node_properties->op_params, node->op_params, GGML_MAX_OP_PARAMS);
 }
 
 static bool ggml_graph_node_has_matching_properties(ggml_tensor * node, ggml_graph_node_properties * graph_node_properties) {
@@ -2509,6 +2516,12 @@ static bool ggml_graph_node_has_matching_properties(ggml_tensor * node, ggml_gra
             return false;
         }
     }
+
+    if (node->op == GGML_OP_SCALE &&
+        memcmp(graph_node_properties->op_params, node->op_params, GGML_MAX_OP_PARAMS) != 0) {
+        return false;
+    }
+
     return true;
 }
 
@@ -2720,7 +2733,9 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
             // First call with null argument gets number of nodes in graph
             CUDA_CHECK(cudaGraphGetNodes(cuda_ctx->cuda_graph->graph, nullptr, &cuda_ctx->cuda_graph->num_nodes));
             // Subsequent call with non-null argument gets nodes
+            cuda_ctx->cuda_graph->nodes.clear();
             cuda_ctx->cuda_graph->nodes.resize(cuda_ctx->cuda_graph->num_nodes);
+            cuda_ctx->cuda_graph->params.clear();
             cuda_ctx->cuda_graph->params.resize(cuda_ctx->cuda_graph->num_nodes);
             if (cuda_ctx->cuda_graph->num_nodes > 0) {
                 CUDA_CHECK(cudaGraphGetNodes(cuda_ctx->cuda_graph->graph, cuda_ctx->cuda_graph->nodes.data(), &cuda_ctx->cuda_graph->num_nodes));
@@ -2797,6 +2812,7 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
                 case GGML_UNARY_OP_HARDSWISH:
                 case GGML_UNARY_OP_GELU_QUICK:
                 case GGML_UNARY_OP_TANH:
+                case GGML_UNARY_OP_EXP:
                     return ggml_is_contiguous(op->src[0]);
                 default:
                     return false;
@@ -2958,20 +2974,21 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
         case GGML_OP_ARANGE:
         case GGML_OP_TIMESTEP_EMBEDDING:
         case GGML_OP_LEAKY_RELU:
+        case GGML_OP_RWKV_WKV:
             return true;
-        case GGML_OP_FLASH_ATTN_EXT:
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-            return (op->src[0]->ne[0] == 64 && op->src[1]->type == GGML_TYPE_F16) || op->src[0]->ne[0] == 128;
-#else
-            if (op->src[0]->ne[0] == 128) {
-                return true;
-            }
+        case GGML_OP_FLASH_ATTN_EXT: {
             if (op->src[0]->ne[0] ==  64 && op->src[1]->type == GGML_TYPE_F16) {
                 return true;
             }
-            return ggml_cuda_info().devices[cuda_ctx->device].cc >= CC_VOLTA &&
-                op->src[1]->type == GGML_TYPE_F16 && op->src[2]->type == GGML_TYPE_F16;
-#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+            if (op->src[0]->ne[0] == 128) {
+                return true;
+            }
+            if (op->src[0]->ne[0] == 256 && op->src[1]->type == GGML_TYPE_F16 && op->src[2]->type == GGML_TYPE_F16) {
+                return true;
+            }
+            const int cc = ggml_cuda_info().devices[cuda_ctx->device].cc;
+            return cc >= CC_VOLTA && cc < CC_OFFSET_AMD && op->src[1]->type == GGML_TYPE_F16 && op->src[2]->type == GGML_TYPE_F16;
+        }
         case GGML_OP_CROSS_ENTROPY_LOSS:
         case GGML_OP_CROSS_ENTROPY_LOSS_BACK:
         case GGML_OP_OPT_STEP_ADAMW:

ggml/src/ggml-cuda/common.cuh

@@ -569,6 +569,7 @@ struct ggml_graph_node_properties {
     int64_t ne[GGML_MAX_DIMS];
     size_t nb[GGML_MAX_DIMS];
     void * src_address[GGML_MAX_SRC];
+    int32_t op_params[GGML_MAX_OP_PARAMS / sizeof(int32_t)];
 };
 
 struct ggml_cuda_graph {

ggml/src/ggml-cuda/fattn.cu

@@ -314,7 +314,7 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst
     }
 
     if (!fast_fp16_available(cc)) {
-        if (Q->ne[1] <= 8) {
+        if (Q->ne[1] <= 8 || Q->ne[0] == 256) {
             ggml_cuda_flash_attn_ext_vec_f32(ctx, dst);
         } else {
             ggml_cuda_flash_attn_ext_tile_f32(ctx, dst);

ggml/src/ggml-cuda/rwkv-wkv.cu

@@ -0,0 +1,89 @@
+#include "common.cuh"
+#include "rwkv-wkv.cuh"
+
+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 = CUDA_WKV_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];
+    }
+}
+
+void ggml_cuda_op_rwkv_wkv(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[3];
+    const int64_t C = dst->ne[0];
+    const int64_t H = dst->src[0]->ne[2];
+
+    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);
+
+    rwkv_wkv_f32<<<B * H, C / H, 0, stream>>>(B, T, C, H, k_d, v_d, r_d, tf_d, td_d, s_d, dst_d);
+}

ggml/src/ggml-cuda/rwkv-wkv.cuh

@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_WKV_BLOCK_SIZE 64
+
+void ggml_cuda_op_rwkv_wkv(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml/src/ggml-cuda/unary.cu

@@ -95,6 +95,15 @@ static __global__ void hardswish_f32(const float * x, float * dst, const int k)
     dst[i] = x[i] * fminf(1.0f, fmaxf(0.0f, (x[i] + 3.0f) / 6.0f));
 }
 
+static __global__ void exp_f32(const float * x, float * dst, const int k) {
+    const int i = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+    dst[i] = expf(x[i]);
+}
+
 static __global__ void leaky_relu_f32(const float * x, float * dst, const int k, const float negative_slope) {
     const int i  = blockDim.x*blockIdx.x + threadIdx.x;
     if (i >= k) {
@@ -189,6 +198,11 @@ static void hardswish_f32_cuda(const float * x, float * dst, const int k, cudaSt
     hardswish_f32<<<num_blocks, CUDA_HARDSWISH_BLOCK_SIZE, 0, stream>>>(x, dst, k);
 }
 
+static void exp_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_EXP_BLOCK_SIZE - 1) / CUDA_EXP_BLOCK_SIZE;
+    exp_f32<<<num_blocks, CUDA_EXP_BLOCK_SIZE, 0, stream>>>(x, dst, k);
+}
+
 static void leaky_relu_f32_cuda(const float * x, float * dst, const int k, const float negative_slope, cudaStream_t stream) {
     const int num_blocks = (k + CUDA_RELU_BLOCK_SIZE - 1) / CUDA_RELU_BLOCK_SIZE;
     leaky_relu_f32<<<num_blocks, CUDA_RELU_BLOCK_SIZE, 0, stream>>>(x, dst, k, negative_slope);
@@ -354,6 +368,20 @@ void ggml_cuda_op_hardswish(ggml_backend_cuda_context & ctx, ggml_tensor * dst)
     hardswish_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
 }
 
+void ggml_cuda_op_exp(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    exp_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
+}
+
 void ggml_cuda_op_leaky_relu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0 = dst->src[0];
     const float * src0_d = (const float *)src0->data;

ggml/src/ggml-cuda/unary.cuh

@@ -8,6 +8,7 @@
 #define CUDA_RELU_BLOCK_SIZE 256
 #define CUDA_SIGMOID_BLOCK_SIZE 256
 #define CUDA_HARDSIGMOID_BLOCK_SIZE 256
+#define CUDA_EXP_BLOCK_SIZE 256
 #define CUDA_HARDSWISH_BLOCK_SIZE 256
 #define CUDA_SQR_BLOCK_SIZE 256
 #define CUDA_SQRT_BLOCK_SIZE 256
@@ -32,6 +33,8 @@ void ggml_cuda_op_sigmoid(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
 void ggml_cuda_op_hardsigmoid(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
+void ggml_cuda_op_exp(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
 void ggml_cuda_op_hardswish(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
 void ggml_cuda_op_leaky_relu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

src/llama.cpp

@@ -3056,18 +3056,14 @@ struct llama_sbatch {
         } else {
             // simple split
             if (batch->n_seq_id) {
-                for (size_t i = 0; i < length; ++i) {
-                    ubatch.n_seq_id = batch->n_seq_id + seq.offset;
-                }
+                ubatch.n_seq_id = batch->n_seq_id + seq.offset;
             } else {
                 for (size_t i = 0; i < length; ++i) {
                     ubatch.n_seq_id[ubatch.n_seqs + i] = 1;
                 }
             }
             if (batch->seq_id) {
-                for (size_t i = 0; i < length; ++i) {
-                    ubatch.seq_id = batch->seq_id + seq.offset;
-                }
+                ubatch.seq_id = batch->seq_id + seq.offset;
             } else {
                 for (size_t i = 0; i < length; ++i) {
                     ubatch.seq_id[ubatch.n_seqs + i] = &seq.all_seq_id;

tests/test-backend-ops.cpp

@@ -1543,6 +1543,36 @@ struct test_ssm_scan : public test_case {
     }
 };
 
+// GGML_OP_RWKV_WKV
+struct test_rwkv_wkv : public test_case {
+    const ggml_type type;
+
+    const int64_t head_count;
+    const int64_t head_size;
+    const int64_t n_seq_tokens;
+    const int64_t n_seqs;
+
+    std::string vars() override {
+        return VARS_TO_STR5(type, head_count, head_size, n_seq_tokens, n_seqs);
+    }
+
+    test_rwkv_wkv(ggml_type type = GGML_TYPE_F32,
+            int64_t head_count = 32, int64_t head_size = 64, int64_t n_seq_tokens = 32, int64_t n_seqs = 32)
+        : type(type), head_count(head_count), head_size(head_size), n_seq_tokens(n_seq_tokens), n_seqs(n_seqs) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        const int64_t n_tokens = n_seq_tokens * n_seqs;
+        ggml_tensor * r   = ggml_new_tensor(ctx, type, 4, std::vector<int64_t>{ 1, head_size, head_count, n_tokens }.data());
+        ggml_tensor * k   = ggml_new_tensor(ctx, type, 4, std::vector<int64_t>{ head_size, 1, head_count, n_tokens }.data());
+        ggml_tensor * v   = ggml_new_tensor(ctx, type, 4, std::vector<int64_t>{ 1, head_size, head_count, n_tokens }.data());
+        ggml_tensor * tf  = ggml_new_tensor(ctx, type, 2, std::vector<int64_t>{ head_size, head_count }.data());
+        ggml_tensor * td  = ggml_new_tensor(ctx, type, 4, std::vector<int64_t>{ 1, head_size, head_count, n_tokens }.data());
+        ggml_tensor * s   = ggml_new_tensor(ctx, type, 2, std::vector<int64_t>{ head_size * head_size * head_count, n_seqs }.data());
+        ggml_tensor * out = ggml_rwkv_wkv(ctx, k, v, r, tf, td, s);
+        return out;
+    }
+};
+
 // GGML_OP_MUL_MAT
 struct test_mul_mat : public test_case {
     const ggml_type type_a;
@@ -3337,6 +3367,11 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
 
     test_cases.emplace_back(new test_ssm_scan(GGML_TYPE_F32, 16, 1024, 32, 4));
 
+    test_cases.emplace_back(new test_rwkv_wkv(GGML_TYPE_F32, 32, 64, 1, 1));
+    test_cases.emplace_back(new test_rwkv_wkv(GGML_TYPE_F32, 32, 64, 32, 1));
+    test_cases.emplace_back(new test_rwkv_wkv(GGML_TYPE_F32, 32, 64, 32, 4));
+    test_cases.emplace_back(new test_rwkv_wkv(GGML_TYPE_F32, 32, 64, 128, 4));
+
 #if 1
     for (ggml_type type_a : base_types) {
         for (ggml_type type_b : {GGML_TYPE_F32, GGML_TYPE_F16}) {
@@ -3564,7 +3599,7 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
                     if (hs != 128 && logit_softcap != 0.0f) continue;
                     for (int nh : { 32, }) {
                         for (int kv : { 512, 1024, }) {
-                            for (int nb : { 1, 2, 4, 8, }) {
+                            for (int nb : { 1, 3, 32, 35, }) {
                                 for (ggml_type type_KV : {GGML_TYPE_F16, GGML_TYPE_Q8_0, GGML_TYPE_Q4_0}) {
                                     test_cases.emplace_back(new test_flash_attn_ext(hs, nh, kv, nb, mask, max_bias, logit_softcap, type_KV));
                                 }