@novastera-oss/llamarn 0.2.9 → 0.3.0

package/cpp/llama.cpp/ggml/src/ggml-cuda/scale.cu

@@ -1,18 +1,18 @@
 #include "scale.cuh"
 
-static __global__ void scale_f32(const float * x, float * dst, const float scale, const int k) {
+static __global__ void scale_f32(const float * x, float * dst, const float scale, const float bias, const int k) {
     const int i = blockDim.x*blockIdx.x + threadIdx.x;
 
     if (i >= k) {
         return;
     }
 
-    dst[i] = scale * x[i];
+    dst[i] = scale * x[i] + bias;
 }
 
-static void scale_f32_cuda(const float * x, float * dst, const float scale, const int k, cudaStream_t stream) {
+static void scale_f32_cuda(const float * x, float * dst, const float scale, const float bias, const int k, cudaStream_t stream) {
     const int num_blocks = (k + CUDA_SCALE_BLOCK_SIZE - 1) / CUDA_SCALE_BLOCK_SIZE;
-    scale_f32<<<num_blocks, CUDA_SCALE_BLOCK_SIZE, 0, stream>>>(x, dst, scale, k);
+    scale_f32<<<num_blocks, CUDA_SCALE_BLOCK_SIZE, 0, stream>>>(x, dst, scale, bias, k);
 }
 
 void ggml_cuda_op_scale(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
@@ -25,7 +25,9 @@ void ggml_cuda_op_scale(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
     float scale;
-    memcpy(&scale, dst->op_params, sizeof(float));
+    float bias;
+    memcpy(&scale, (float *) dst->op_params + 0, sizeof(float));
+    memcpy(&bias,  (float *) dst->op_params + 1, sizeof(float));
 
-    scale_f32_cuda(src0_d, dst_d, scale, ggml_nelements(src0), stream);
+    scale_f32_cuda(src0_d, dst_d, scale, bias, ggml_nelements(src0), stream);
 }

package/cpp/llama.cpp/ggml/src/ggml-cuda/softmax.cu

@@ -2,6 +2,7 @@
 #include "ggml.h"
 #include "softmax.cuh"
 #include <cstdint>
+#include <utility>
 
 template <typename T>
 static __device__ __forceinline__ float t2f32(T val) {
@@ -13,6 +14,29 @@ __device__ float __forceinline__ t2f32<half>(half val) {
     return __half2float(val);
 }
 
+struct soft_max_params {
+
+    int64_t nheads;
+    uint32_t n_head_log2;
+    int64_t ncols;
+    int64_t nrows_x;
+    int64_t nrows_y;
+    int64_t ne00;
+    int64_t ne01;
+    int64_t ne02;
+    int64_t ne03;
+    int64_t nb11;
+    int64_t nb12;
+    int64_t nb13;
+
+    int64_t ne12;
+    int64_t ne13;
+    float scale;
+    float max_bias;
+    float m0;
+    float m1;
+};
+
 // 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__
@@ -21,16 +45,24 @@ __device__ float __forceinline__ t2f32<half>(half val) {
 #endif // __clang__
 template <bool use_shared, int ncols_template, int block_size_template, typename T>
 static __global__ void soft_max_f32(
-        const float * x, const T * mask, float * dst, const int ncols_par, const int nrows_y,
-        const float scale, const float max_bias, const float m0, const float m1, uint32_t n_head_log2) {
-    const int ncols = ncols_template == 0 ? ncols_par : ncols_template;
+        const float * x, const T * mask, float * dst, const soft_max_params p) {
+    const int ncols = ncols_template == 0 ? p.ncols : ncols_template;
 
     const int tid  = threadIdx.x;
-    const int rowx = blockIdx.x;
-    const int rowy = rowx % nrows_y; // broadcast the mask in the row dimension
+
+    const int64_t i03 = blockIdx.z;
+    const int64_t i02 = blockIdx.y;
+    const int64_t i01 = blockIdx.x;
+
+    //TODO: noncontigous inputs/outputs
+    const int rowx = blockIdx.x + blockIdx.y * gridDim.x + blockIdx.z * gridDim.x * gridDim.y;
+
+    const int64_t i11 = i01;
+    const int64_t i12 = i02 % p.ne12;
+    const int64_t i13 = i03 % p.ne13;
 
     x    += int64_t(rowx)*ncols;
-    mask += int64_t(rowy)*ncols * (mask != nullptr);
+    mask += (i11*p.nb11 + i12*p.nb12 + i13*p.nb13) / sizeof(T) * (mask != nullptr);
     dst  += int64_t(rowx)*ncols;
 
     const int block_size = block_size_template == 0 ? blockDim.x : block_size_template;
@@ -38,7 +70,7 @@ static __global__ void soft_max_f32(
     const int warp_id = threadIdx.x / WARP_SIZE;
     const int lane_id = threadIdx.x % WARP_SIZE;
 
-    const float slope = get_alibi_slope(max_bias, rowx/nrows_y, n_head_log2, m0, m1);
+    const float slope = get_alibi_slope(p.max_bias, i02, p.n_head_log2, p.m0, p.m1);
 
     extern __shared__ float data_soft_max_f32[];
     float * buf_iw = data_soft_max_f32; // shared memory buffer for inter-warp communication
@@ -55,7 +87,7 @@ static __global__ void soft_max_f32(
             break;
         }
 
-        const float val = x[col]*scale + (mask ? slope*t2f32(mask[col]) : 0.0f);
+        const float val = x[col]*p.scale + (mask ? slope*t2f32(mask[col]) : 0.0f);
 
         vals[col] = val;
         max_val = max(max_val, val);
@@ -150,64 +182,58 @@ static __global__ void soft_max_back_f32(
     }
 }
 
+template<int... Ns, typename T>
+static void launch_soft_max_kernels(const float * x, const T * mask, float * dst,
+                             const soft_max_params & p, cudaStream_t stream, dim3 block_dims, dim3 block_nums, size_t nbytes_shared)
+{
+    const int id       = ggml_cuda_get_device();
+    const size_t smpbo = ggml_cuda_info().devices[id].smpbo;
+
+    auto launch_kernel = [=](auto I) -> bool {
+        constexpr int ncols = decltype(I)::value;
+        constexpr int block = (ncols > 1024 ? 1024 : ncols);
+
+        if (p.ncols == ncols) {
+            CUDA_SET_SHARED_MEMORY_LIMIT((soft_max_f32<true, ncols, block, T>), smpbo);
+            soft_max_f32<true, ncols, block><<<block_nums, block_dims, nbytes_shared, stream>>>
+                (x, mask, dst, p);
+            return true;
+        }
+        return false;
+    };
+
+    // unary fold over launch_kernel
+    if ((launch_kernel(std::integral_constant<int, Ns>{}) || ...)) {
+        return;
+    }
+
+    //default case
+    CUDA_SET_SHARED_MEMORY_LIMIT((soft_max_f32<true, 0, 0, T>), smpbo);
+    soft_max_f32<true, 0, 0><<<block_nums, block_dims, nbytes_shared, stream>>>(x, mask, dst, p);
+}
+
+
 template<typename T>
-static void soft_max_f32_cuda(const float * x, const T * mask, float * dst, const int ncols_x, const int nrows_x, const int nrows_y, const float scale, const float max_bias, cudaStream_t stream) {
+static void soft_max_f32_cuda(const float * x, const T * mask, float * dst, const soft_max_params & params, cudaStream_t stream) {
     int nth = WARP_SIZE;
+    const int64_t ncols_x = params.ncols;
+
     while (nth < ncols_x && nth < CUDA_SOFT_MAX_BLOCK_SIZE) nth *= 2;
     const dim3 block_dims(nth,     1, 1);
-    const dim3 block_nums(nrows_x, 1, 1);
+    const dim3 block_nums(params.ne01, params.ne02, params.ne03);
     const size_t nbytes_shared = (GGML_PAD(ncols_x, WARP_SIZE) + WARP_SIZE)*sizeof(float);
     static_assert(CUDA_SOFT_MAX_BLOCK_SIZE == 1024, "These values need to be adjusted.");
 
-    const uint32_t n_head      = nrows_x/nrows_y;
-    const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
 
-    const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
-    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
+    const int id       = ggml_cuda_get_device();
+    const size_t smpbo = ggml_cuda_info().devices[id].smpbo;
 
-    // FIXME: this limit could be raised by ~2-4x on Ampere or newer
-    if (nbytes_shared < ggml_cuda_info().devices[ggml_cuda_get_device()].smpb) {
-        switch (ncols_x) {
-            case 32:
-                soft_max_f32<true,   32,   32><<<block_nums, block_dims, nbytes_shared, stream>>>
-                    (x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
-                break;
-            case 64:
-                soft_max_f32<true,   64,   64><<<block_nums, block_dims, nbytes_shared, stream>>>
-                    (x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
-                break;
-            case 128:
-                soft_max_f32<true,  128,  128><<<block_nums, block_dims, nbytes_shared, stream>>>
-                    (x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
-                break;
-            case 256:
-                soft_max_f32<true,  256,  256><<<block_nums, block_dims, nbytes_shared, stream>>>
-                    (x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
-                break;
-            case 512:
-                soft_max_f32<true,  512,  512><<<block_nums, block_dims, nbytes_shared, stream>>>
-                    (x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
-                break;
-            case 1024:
-                soft_max_f32<true, 1024, 1024><<<block_nums, block_dims, nbytes_shared, stream>>>
-                    (x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
-                break;
-            case 2048:
-                soft_max_f32<true, 2048, 1024><<<block_nums, block_dims, nbytes_shared, stream>>>
-                    (x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
-                break;
-            case 4096:
-                soft_max_f32<true, 4096, 1024><<<block_nums, block_dims, nbytes_shared, stream>>>
-                    (x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
-                break;
-            default:
-                soft_max_f32<true,    0,    0><<<block_nums, block_dims, nbytes_shared, stream>>>
-                    (x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
-                break;
-        }
+
+    if (nbytes_shared <= smpbo) {
+        launch_soft_max_kernels<32, 64, 128, 256, 512, 1024, 2048, 4096>(x, mask, 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, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+        soft_max_f32<false, 0, 0><<<block_nums, block_dims, nbytes_shared_low, stream>>>(x, mask, dst, params);
     }
 }
 
@@ -235,10 +261,11 @@ void ggml_cuda_op_soft_max(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
 
     GGML_ASSERT(!src1 || src1->type == GGML_TYPE_F16 || src1->type == GGML_TYPE_F32); // src1 contains mask and it is optional
 
-    const int64_t ne00    = src0->ne[0];
     const int64_t nrows_x = ggml_nrows(src0);
     const int64_t nrows_y = src0->ne[1];
 
+    const int64_t ne00 = src0->ne[0];
+
     float scale    = 1.0f;
     float max_bias = 0.0f;
 
@@ -247,10 +274,44 @@ void ggml_cuda_op_soft_max(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
 
     const bool use_f16 = (src1 && src1->type == GGML_TYPE_F16);
 
+    const int64_t nb11 = src1 ? src1->nb[1] : 1;
+    const int64_t nb12 = src1 ? src1->nb[2] : 1;
+    const int64_t nb13 = src1 ? src1->nb[3] : 1;
+
+    const int64_t ne12 = src1 ? src1->ne[2] : 1;
+    const int64_t ne13 = src1 ? src1->ne[3] : 1;
+
+    const uint32_t n_head      = src0->ne[2];
+    const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
+
+    const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
+    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
+
+
+    soft_max_params params = {};
+    params.nheads = src0->ne[2];
+    params.n_head_log2 = n_head_log2;
+    params.ncols = ne00;
+    params.nrows_x = nrows_x;
+    params.nrows_y = nrows_y;
+    params.ne00 = src0->ne[0];
+    params.ne01 = src0->ne[1];
+    params.ne02 = src0->ne[2];
+    params.ne03 = src0->ne[3];
+    params.nb11 = nb11;
+    params.nb12 = nb12;
+    params.nb13 = nb13;
+    params.ne12 = ne12;
+    params.ne13 = ne13;
+    params.scale = scale;
+    params.max_bias = max_bias;
+    params.m0 = m0;
+    params.m1 = m1;
+
     if (use_f16) {
-        soft_max_f32_cuda(src0_d, (const half  *) src1_d, dst_d, ne00, nrows_x, nrows_y, scale, max_bias, stream);
+        soft_max_f32_cuda(src0_d, (const half  *) src1_d, dst_d, params, stream);
     } else {
-        soft_max_f32_cuda(src0_d, (const float *) src1_d, dst_d, ne00, nrows_x, nrows_y, scale, max_bias, stream);
+        soft_max_f32_cuda(src0_d, (const float *) src1_d, dst_d, params, stream);
     }
 }
 

package/cpp/llama.cpp/ggml/src/ggml-cuda/ssm-conv.cu

@@ -107,8 +107,11 @@ static void ssm_conv_f32_cuda(const float * src0, const float * src1, const int
         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);
         } else {
-            GGML_ABORT("Only support kernel size = 4  now.");
+            GGML_ABORT("Only support kernel size = 3 or size = 4 right now.");
         }
     } else {
         if (nc == 4) {
@@ -116,8 +119,13 @@ static void ssm_conv_f32_cuda(const float * src0, const float * src1, const int
             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>>>(
+                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 = 4 right now.");
+            GGML_ABORT("Only support kernel size = 3 or size = 4 right now.");
         }
     }
 }

package/cpp/llama.cpp/ggml/src/ggml-cuda/ssm-scan.cu

@@ -4,16 +4,15 @@ template <size_t splitD, size_t N>
 __global__ void __launch_bounds__(splitD, 2)
     ssm_scan_f32(const float * __restrict__ src0, const float * __restrict__ src1, const float * __restrict__ src2,
                  const float * __restrict__ src3, const float * __restrict__ src4, const float * __restrict__ src5,
-                 const int src0_nb1, const int src0_nb2, const int src1_nb0, const int src1_nb1, const int src1_nb2,
-                 const int src1_nb3, const int src2_nb0, const int src2_nb1, const int src2_nb2, const int src3_nb1,
-                 const int src4_nb1, const int src4_nb2, const int src5_nb1, const int src5_nb2,
-                 float * __restrict__ dst, const int64_t L) {
-    GGML_UNUSED(src1_nb0);
-    GGML_UNUSED(src2_nb0);
+                 const int32_t * __restrict__ src6, float * __restrict__ dst,
+                 const int src0_nb2, const int src0_nb3, const int src1_nb2, const int src1_nb3,
+                 const int src2_nb1, const int src2_nb2, const int src3_nb1,
+                 const int src4_nb2, const int src4_nb3, const int src5_nb2, const int src5_nb3,
+                 const int64_t s_off, const int64_t d_inner, const int64_t L) {
 
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
-    const int bidx = blockIdx.x;  // split along B
-    const int bidy = blockIdx.y;  // split along D
+    const int bidx = blockIdx.x;  // split along B (sequences)
+    const int bidy = blockIdx.y;  // split along D (d_inner)
     const int tid  = threadIdx.x;
     const int wid  = tid / 32;
     const int wtid = tid % 32;
@@ -24,23 +23,23 @@ __global__ void __launch_bounds__(splitD, 2)
     float *                 smem_A     = smem;
     float *                 smem_s0    = smem_A + splitD * stride_sA;
 
-    const float * s0_block = (const float *) ((const char *) src0 + bidx * src0_nb2 + bidy * splitD * src0_nb1);
-    const float * x_block  = (const float *) ((const char *) src1 + (bidx * src1_nb2) + bidy * splitD * sizeof(float));
+    const float * s0_block = (const float *) ((const char *) src0 + src6[bidx] * src0_nb3 + bidy * splitD * src0_nb2);
+    const float * x_block  = (const float *) ((const char *) src1 + (bidx * src1_nb3) + bidy * splitD * sizeof(float));
     const float * dt_block = (const float *) ((const char *) src2 + (bidx * src2_nb2) + bidy * splitD * sizeof(float));
     const float * A_block  = (const float *) ((const char *) src3 + bidy * splitD * src3_nb1);
-    const float * B_block  = (const float *) ((const char *) src4 + (bidx * src4_nb2));
-    const float * C_block  = (const float *) ((const char *) src5 + (bidx * src5_nb2));
-    float *       y_block  = (float *) ((char *) dst + (bidx * src1_nb2) + bidy * splitD * sizeof(float));
-    float *       s_block  = (float *) ((char *) dst + src1_nb3 + bidx * src0_nb2 + bidy * splitD * src0_nb1);
+    const float * B_block  = (const float *) ((const char *) src4 + (bidx * src4_nb3));
+    const float * C_block  = (const float *) ((const char *) src5 + (bidx * src5_nb3));
+    float *       y_block  = (float *) ((char *) dst + (bidx * d_inner * L * sizeof(float)) + bidy * splitD * sizeof(float));
+    float *       s_block  = (float *) ((char *) dst + s_off + bidx * src0_nb3 + bidy * splitD * src0_nb2);
 
-    const int stride_s0 = src0_nb1 / sizeof(float);
-    const int stride_x  = src1_nb1 / sizeof(float);
+    const int stride_s0 = src0_nb2 / sizeof(float);
+    const int stride_x  = src1_nb2 / sizeof(float);
     const int stride_dt = src2_nb1 / sizeof(float);
     const int stride_A  = src3_nb1 / sizeof(float);
-    const int stride_B  = src4_nb1 / sizeof(float);
-    const int stride_C  = src5_nb1 / sizeof(float);
+    const int stride_B  = src4_nb2 / sizeof(float);
+    const int stride_C  = src5_nb2 / sizeof(float);
     const int stride_s  = stride_s0;
-    const int stride_y  = stride_x;
+    const int stride_y  = d_inner;
 
     // can N not be 16? for example 32?
     if (N == 16) {
@@ -84,24 +83,167 @@ __global__ void __launch_bounds__(splitD, 2)
     }
 }
 
+// assumes as many threads as d_state
+template <int splitH, int d_state>
+__global__ void __launch_bounds__(d_state, 1)
+    ssm_scan_f32_group(
+        const float * __restrict__ src0, const float * __restrict__ src1, const float * __restrict__ src2,
+        const float * __restrict__ src3, const float * __restrict__ src4, const float * __restrict__ src5,
+        const int32_t * __restrict__ src6, float * __restrict__ dst,
+        const int src0_nb2, const int src0_nb3, const int src1_nb2, const int src1_nb3,
+        const int src2_nb1, const int src2_nb2, const int src3_nb1,
+        const int src4_nb2, const int src4_nb3, const int src5_nb2, const int src5_nb3,
+        const int64_t s_off, const int64_t n_head, const int64_t d_head, const int64_t n_group, const int64_t n_tok) {
+
+    const int head_idx = (blockIdx.x * splitH) / d_head;
+    const int head_off = ((blockIdx.x * splitH) % d_head) * sizeof(float);
+    const int seq_idx = blockIdx.y;
+
+    const int group_off = (head_idx & (n_group - 1)) * d_state * sizeof(float);
+
+    const float * s0_block = (const float *) ((const char *) src0 + src6[seq_idx] * src0_nb3 + head_idx * src0_nb2 + head_off * d_state);
+    const float * x_block  = (const float *) ((const char *) src1 + (seq_idx * src1_nb3) + blockIdx.x * splitH * sizeof(float));
+    const float * dt_block = (const float *) ((const char *) src2 + (seq_idx * src2_nb2) + head_idx * sizeof(float));
+    const float * A_block  = (const float *) ((const char *) src3 + head_idx * src3_nb1);
+    const float * B_block  = (const float *) ((const char *) src4 + (seq_idx * src4_nb3) + (group_off));
+    const float * C_block  = (const float *) ((const char *) src5 + (seq_idx * src5_nb3) + (group_off));
+    float *       y_block  = dst + (seq_idx * n_tok * n_head * d_head) + blockIdx.x * splitH;
+    float *       s_block  = (float *) ((char *) dst + s_off + seq_idx * src0_nb3 + head_idx * src0_nb2 + head_off * d_state);
+
+    // strides across n_seq_tokens
+    const int stride_x  = src1_nb2 / sizeof(float);
+    const int stride_dt = src2_nb1 / sizeof(float);
+    const int stride_B  = src4_nb2 / sizeof(float);
+    const int stride_C  = src5_nb2 / sizeof(float);
+    const int stride_y  = n_head * d_head;
+
+    float state[splitH];
+    // for the parallel accumulation
+    __shared__ float stateC[splitH * d_state];
+
+#pragma unroll
+    for (int j = 0; j < splitH; j++) {
+        state[j] = s0_block[j * d_state + threadIdx.x];
+    }
+
+    for (int64_t i = 0; i < n_tok; i++) {
+        // TODO: only calculate dA and dt_soft_plus once per head instead of every splitH head elements
+        // TODO: only calculate B and C once per head group
+        // NOTE: dt_soft_plus, dA and x_dt have the same value across threads here.
+        float dt_soft_plus = dt_block[i * stride_dt];
+        if (dt_soft_plus <= 20.0f) {
+            dt_soft_plus = log1pf(expf(dt_soft_plus));
+        }
+        const float dA = expf(dt_soft_plus * A_block[0]);
+        const float B = B_block[i * stride_B + threadIdx.x];
+        const float C = C_block[i * stride_C + threadIdx.x];
+
+        // across d_head
+#pragma unroll
+        for (int j = 0; j < splitH; j++) {
+            const float x_dt = x_block[i * stride_x + j] * dt_soft_plus;
+
+            state[j] = (state[j] * dA) + (B * x_dt);
+
+            stateC[j * d_state + threadIdx.x] = state[j] * C;
+        }
+
+        __syncthreads();
+
+        // parallel accumulation for stateC
+        // TODO: simplify
+        {
+            static_assert((d_state & -d_state) == d_state, "the state size has to be a power of 2");
+            static_assert((splitH & -splitH) == splitH, "splitH has to be a power of 2");
+
+            // reduce until w matches the warp size
+            // TODO: does this work even when the physical warp size is 64?
+#pragma unroll
+            for (int w = d_state; w > WARP_SIZE; w >>= 1) {
+                // (assuming there are d_state threads)
+#pragma unroll
+                for (int j = 0; j < ((w >> 1) * splitH + d_state - 1) / d_state; j++) {
+                    // TODO: check for bank conflicts
+                    const int k = (threadIdx.x % (w >> 1)) + (d_state * (threadIdx.x / (w >> 1))) + j * d_state * (d_state / (w >> 1));
+                    stateC[k] += stateC[k + (w >> 1)];
+
+                }
+                __syncthreads();
+            }
+
+            static_assert(splitH >= d_state / WARP_SIZE);
+
+#pragma unroll
+            for (int j = 0; j < splitH / (d_state / WARP_SIZE); j++) {
+                float y = stateC[(threadIdx.x % WARP_SIZE) + d_state * (threadIdx.x / WARP_SIZE) + j * d_state * (d_state / WARP_SIZE)];
+                y = warp_reduce_sum(y);
+
+                // store the above accumulations
+                if (threadIdx.x % WARP_SIZE == 0) {
+                    const int k = threadIdx.x / WARP_SIZE + j * (d_state / WARP_SIZE);
+                    y_block[i * stride_y + k] = y;
+                }
+            }
+        }
+    }
+
+    // write back the state
+#pragma unroll
+    for (int j = 0; j < splitH; j++) {
+        s_block[j * d_state + threadIdx.x] = state[j];
+    }
+}
+
 static void ssm_scan_f32_cuda(const float * src0, const float * src1, const float * src2, const float * src3,
-                              const float * src4, const float * src5, const int src0_nb1, const int src0_nb2,
-                              const int src1_nb0, const int src1_nb1, const int src1_nb2, const int src1_nb3,
-                              const int src2_nb0, const int src2_nb1, const int src2_nb2, const int src3_nb1,
-                              const int src4_nb1, const int src4_nb2, const int src5_nb1, const int src5_nb2,
-                              float * dst, const int64_t N, const int64_t D, const int64_t L, const int64_t B,
+                              const float * src4, const float * src5, const int32_t * src6, float * dst,
+                              const int src0_nb2, const int src0_nb3, const int src1_nb2, const int src1_nb3, const int src2_nb1,
+                              const int src2_nb2, const int src3_nb1, const int src4_nb2, const int src4_nb3, const int src5_nb2,
+                              const int src5_nb3, const int64_t s_off, const int64_t d_state, const int64_t head_dim,
+                              const int64_t n_head, const int64_t n_group, const int64_t n_tok, const int64_t n_seq,
                               cudaStream_t stream) {
-    const int threads = 128;
-    // todo: consider D cannot be divided,does this situation exist?
-    GGML_ASSERT(D % threads == 0);
-    const dim3 blocks(B, (D + threads - 1) / threads, 1);
-    const int  smem_size = (threads * (N + 1) * 2) * sizeof(float);
-    if (N == 16) {
-        ssm_scan_f32<128, 16><<<blocks, threads, smem_size, stream>>>(
-            src0, src1, src2, src3, src4, src5, src0_nb1, src0_nb2, src1_nb0, src1_nb1, src1_nb2, src1_nb3, src2_nb0,
-            src2_nb1, src2_nb2, src3_nb1, src4_nb1, src4_nb2, src5_nb1, src5_nb2, dst, L);
+    // NOTE: if you change conditions here, be sure to update the corresponding supports_op condition!
+    if (src3_nb1 == sizeof(float)) {
+        // Mamba-2
+        if (d_state == 128) {
+            const int threads = 128;
+            GGML_ASSERT(d_state % threads == 0);
+            // NOTE: can be any power of two between 4 and 64
+            const int splitH = 16;
+            GGML_ASSERT(head_dim % splitH == 0);
+            const dim3 blocks((n_head * head_dim + (splitH - 1)) / splitH, n_seq, 1);
+            ssm_scan_f32_group<16, 128><<<blocks, threads, 0, stream>>>(
+                    src0, src1, src2, src3, src4, src5, src6, dst,
+                    src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2, src3_nb1,
+                    src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, head_dim, n_group, n_tok);
+        } else if (d_state == 256) { // Falcon-H1
+            const int threads = 256;
+            // NOTE: can be any power of two between 8 and 64
+            const int splitH = 16;
+            GGML_ASSERT(head_dim % splitH == 0);
+            const dim3 blocks((n_head * head_dim + (splitH - 1)) / splitH, n_seq, 1);
+            ssm_scan_f32_group<16, 256><<<blocks, threads, 0, stream>>>(
+                    src0, src1, src2, src3, src4, src5, src6, dst,
+                    src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2, src3_nb1,
+                    src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, head_dim, n_group, n_tok);
+        } else {
+            GGML_ABORT("doesn't support d_state!=(128 or 256).");
+        }
     } else {
-        GGML_ABORT("doesn't support N!=16.");
+        const int threads = 128;
+        // Mamba-1
+        GGML_ASSERT(n_head % threads == 0);
+        GGML_ASSERT(head_dim == 1);
+        GGML_ASSERT(n_group == 1);
+        const dim3 blocks(n_seq, (n_head + threads - 1) / threads, 1);
+        const int  smem_size = (threads * (d_state + 1) * 2) * sizeof(float);
+        if (d_state == 16) {
+            ssm_scan_f32<128, 16><<<blocks, threads, smem_size, stream>>>(
+                src0, src1, src2, src3, src4, src5, src6, dst,
+                src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2,
+                src3_nb1, src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, n_tok);
+        } else {
+            GGML_ABORT("doesn't support d_state!=16.");
+        }
     }
 }
 
@@ -112,30 +254,25 @@ void ggml_cuda_op_ssm_scan(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const struct ggml_tensor * src3 = dst->src[3];  // A
     const struct ggml_tensor * src4 = dst->src[4];  // B
     const struct ggml_tensor * src5 = dst->src[5];  // C
-
-    //   const int64_t d_state = src0->ne[0];
-    //   const int64_t d_inner = src0->ne[1];
-    //   const int64_t l = src1->ne[1];
-    //   const int64_t b = src0->ne[2];
+    const struct ggml_tensor * src6 = dst->src[6];  // ids
 
     const int64_t nc  = src0->ne[0];  // d_state
-    const int64_t nr  = src0->ne[1];  // d_inner
-    const int64_t n_t = src1->ne[1];  // number of tokens per sequence
-    const int64_t n_s = src0->ne[2];  // number of sequences in the batch
+    const int64_t nr  = src0->ne[1];  // head_dim or 1
+    const int64_t nh  = src1->ne[1];  // n_head
+    const int64_t ng  = src4->ne[1];  // n_group
+    const int64_t n_t = src1->ne[2];  // number of tokens per sequence
+    const int64_t n_s = src1->ne[3];  // number of sequences in the batch
+
+    const int64_t s_off = ggml_nelements(src1) * sizeof(float);
 
-    GGML_ASSERT(ggml_nelements(src1) + ggml_nelements(src0) == ggml_nelements(dst));
+    GGML_ASSERT(ggml_nelements(src1) + nc*nr*nh*n_s == ggml_nelements(dst));
     GGML_ASSERT(src0->nb[0] == sizeof(float));
     GGML_ASSERT(src1->nb[0] == sizeof(float));
     GGML_ASSERT(src2->nb[0] == sizeof(float));
     GGML_ASSERT(src3->nb[0] == sizeof(float));
     GGML_ASSERT(src4->nb[0] == sizeof(float));
     GGML_ASSERT(src5->nb[0] == sizeof(float));
-    // required for the dot product between s and C
-    GGML_ASSERT(src0->nb[1] == src0->ne[0] * sizeof(float));
-    // required for per-sequence offsets for states
-    GGML_ASSERT(src0->nb[2] == src0->ne[0] * src0->ne[1] * sizeof(float));
-    // required to get correct offset for state destination (i.e. src1->nb[3])
-    GGML_ASSERT(src1->nb[3] == src1->ne[0] * src1->ne[1] * src1->ne[2] * sizeof(float));
+    GGML_ASSERT(src6->nb[0] == sizeof(int32_t));
 
     const float * src0_d = (const float *) src0->data;
     const float * src1_d = (const float *) src1->data;
@@ -143,13 +280,16 @@ void ggml_cuda_op_ssm_scan(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const float * src3_d = (const float *) src3->data;
     const float * src4_d = (const float *) src4->data;
     const float * src5_d = (const float *) src5->data;
+    const int32_t * src6_d = (const int32_t *) src6->data;
     float *       dst_d  = (float *) dst->data;
     cudaStream_t  stream = ctx.stream();
 
     GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src6->type == GGML_TYPE_I32);
     GGML_ASSERT(dst->type == GGML_TYPE_F32);
 
-    ssm_scan_f32_cuda(src0_d, src1_d, src2_d, src3_d, src4_d, src5_d, src0->nb[1], src0->nb[2], src1->nb[0],
-                      src1->nb[1], src1->nb[2], src1->nb[3], src2->nb[0], src2->nb[1], src2->nb[2], src3->nb[1],
-                      src4->nb[1], src4->nb[2], src5->nb[1], src5->nb[2], dst_d, nc, nr, n_t, n_s, stream);
+    ssm_scan_f32_cuda(src0_d, src1_d, src2_d, src3_d, src4_d, src5_d, src6_d, dst_d,
+                      src0->nb[2], src0->nb[3], src1->nb[2], src1->nb[3], src2->nb[1], src2->nb[2],
+                      src3->nb[1], src4->nb[2], src4->nb[3], src5->nb[2], src5->nb[3],
+                      s_off, nc, nr, nh, ng, n_t, n_s, stream);
 }