| 1 | #include "common.cuh" |
| 2 | #include "gla.cuh" |
| 3 | |
| 4 | template<int HEAD_SIZE> |
| 5 | static __global__ void gated_linear_attn_f32(const int B, const int T, const int C, const int H, const float scale, |
| 6 | const float * k, const float * v, const float * r, const float * td, const float * s, float * dst) { |
| 7 | const int tid = threadIdx.x; |
| 8 | const int bid = blockIdx.x; |
| 9 | |
| 10 | const int head_size = HEAD_SIZE; |
| 11 | const int batch_i = bid / H; |
| 12 | const int head_i = bid % H; |
| 13 | const int state_size = C * head_size; |
| 14 | const int n_seq_tokens = T / B; |
| 15 | |
| 16 | float state[head_size]; |
| 17 | __shared__ float _k[head_size], _r[head_size], _td[head_size]; |
| 18 | |
| 19 | #pragma unroll |
| 20 | for (int i = 0; i < head_size; i++) { |
| 21 | state[i] = s[batch_i * state_size + head_i * head_size * head_size + i * head_size + tid]; |
| 22 | } |
| 23 | |
| 24 | 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) { |
| 25 | __syncthreads(); |
| 26 | _k[tid] = k[t]; |
| 27 | _r[tid] = r[t]; |
| 28 | _td[tid] = td[t]; |
| 29 | __syncthreads(); |
| 30 | |
| 31 | const float _v = v[t]; |
| 32 | float y = 0; |
| 33 | for (int j = 0; j < head_size; j += 4) { |
| 34 | const float4 & k = (float4 &)(_k[j]); |
| 35 | const float4 & r = (float4 &)(_r[j]); |
| 36 | const float4 & td = (float4 &)(_td[j]); |
| 37 | float4 & s = (float4 &)(state[j]); |
| 38 | float4 kv; |
| 39 | |
| 40 | kv.x = k.x * _v; |
| 41 | kv.y = k.y * _v; |
| 42 | kv.z = k.z * _v; |
| 43 | kv.w = k.w * _v; |
| 44 | |
| 45 | s.x = s.x * td.x + kv.x; |
| 46 | s.y = s.y * td.y + kv.y; |
| 47 | s.z = s.z * td.z + kv.z; |
| 48 | s.w = s.w * td.w + kv.w; |
| 49 | |
| 50 | y += r.x * s.x; |
| 51 | y += r.y * s.y; |
| 52 | y += r.z * s.z; |
| 53 | y += r.w * s.w; |
| 54 | } |
| 55 | dst[t] = y * scale; |
| 56 | } |
| 57 | |
| 58 | #pragma unroll |
| 59 | for (int i = 0; i < head_size; i++) { |
| 60 | dst[T * C + batch_i * state_size + head_i * head_size * head_size + i * head_size + tid] = state[i]; |
| 61 | } |
| 62 | } |
| 63 | |
| 64 | void ggml_cuda_op_gated_linear_attn(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { |
| 65 | const float * k_d = (const float *)dst->src[0]->data; |
| 66 | const float * v_d = (const float *)dst->src[1]->data; |
| 67 | const float * r_d = (const float *)dst->src[2]->data; |
| 68 | const float * td_d = (const float *)dst->src[3]->data; |
| 69 | const float * s_d = (const float *)dst->src[4]->data; |
| 70 | |
| 71 | const int64_t B = dst->src[4]->ne[1]; |
| 72 | const int64_t T = dst->src[0]->ne[2]; |
| 73 | const int64_t C = dst->ne[0]; |
| 74 | const int64_t H = dst->src[0]->ne[1]; |
| 75 | |
| 76 | float scale; |
| 77 | memcpy(dest: &scale, src: (float*)dst->op_params, n: sizeof(float)); |
| 78 | |
| 79 | float * dst_d = (float *)dst->data; |
| 80 | |
| 81 | cudaStream_t stream = ctx.stream(); |
| 82 | |
| 83 | GGML_ASSERT(dst->src[4]->type == GGML_TYPE_F32); |
| 84 | GGML_ASSERT(C % H == 0); |
| 85 | GGML_ASSERT(C / H == 64 || C / H == 128); |
| 86 | |
| 87 | |
| 88 | if (C / H == 64) { |
| 89 | gated_linear_attn_f32<64><<<gridDim: B * H, blockDim: C / H, sharedMem: 0, stream>>>(B, T, C, H, scale, k: k_d, v: v_d, r: r_d, td: td_d, s: s_d, dst: dst_d); |
| 90 | } else { |
| 91 | gated_linear_attn_f32<128><<<gridDim: B * H, blockDim: C / H, sharedMem: 0, stream>>>(B, T, C, H, scale, k: k_d, v: v_d, r: r_d, td: td_d, s: s_d, dst: dst_d); |
| 92 | } |
| 93 | } |
| 94 | |