diff --git a/ggml/src/ggml-cuda.cu b/ggml/src/ggml-cuda.cu index e06a7826..6331bc17 100644 --- a/ggml/src/ggml-cuda.cu +++ b/ggml/src/ggml-cuda.cu @@ -3470,6 +3470,7 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons case GGML_TYPE_IQ6_K: case GGML_TYPE_IQ1_BN: case GGML_TYPE_IQ2_BN: + case GGML_TYPE_IQ2_K_R4: case GGML_TYPE_IQ3_K_R4: case GGML_TYPE_IQ4_K_R4: case GGML_TYPE_IQ5_K_R4: diff --git a/ggml/src/ggml-cuda/convert.cu b/ggml/src/ggml-cuda/convert.cu index a94c00e9..2ccca01b 100644 --- a/ggml/src/ggml-cuda/convert.cu +++ b/ggml/src/ggml-cuda/convert.cu @@ -886,6 +886,53 @@ static __global__ void dequantize_block_iq5_k_r4(const void * __restrict__ vx, d } } +template +static __global__ void dequantize_block_iq2_k_r4(const void * __restrict__ vx, dst_t * __restrict__ yy, int64_t n_per_row, int64_t row_size) { + + int64_t ii = blockIdx.x; + + int64_t nblock = n_per_row/256; + int64_t row = ii/nblock; + int64_t row4 = row/4; + int64_t ir = row%4; + int64_t ibl = row4*nblock + ii%nblock; + + const int tid = threadIdx.x; + const int il = tid/8; // 0...3 + const int ib = tid%8; // 0...7 + + const block_iq2_k_r4 * x = (const block_iq2_k_r4 *)vx; + dst_t * y = yy + 256*ii + 32*ib; + + const float d = __half2float(x[ibl].d[ir]); + int is = 8*ib + ir; + float dl1 = d * (((x[ibl].scales[is%32] >> 4*(is/32)) & 0xf) - 8); + is += 4; + float dl2 = d * (((x[ibl].scales[is%32] >> 4*(is/32)) & 0xf) - 8); + auto values1 = iq2nl_values + (((x[ibl].extra[ir+0] >> ib) & 1) << 2); + auto values2 = iq2nl_values + (((x[ibl].extra[ir+4] >> ib) & 1) << 2); + auto ql = x[ibl].qs + 32*ib + 4*ir; + if constexpr (std::is_same_v) { + y[il+ 0] = __float2bfloat16(dl1 * values1[(ql[il+ 0] >> 0) & 3]); + y[il+ 4] = __float2bfloat16(dl1 * values1[(ql[il+ 0] >> 2) & 3]); + y[il+ 8] = __float2bfloat16(dl1 * values1[(ql[il+ 0] >> 4) & 3]); + y[il+12] = __float2bfloat16(dl1 * values1[(ql[il+ 0] >> 6) & 3]); + y[il+16] = __float2bfloat16(dl2 * values2[(ql[il+16] >> 0) & 3]); + y[il+20] = __float2bfloat16(dl2 * values2[(ql[il+16] >> 2) & 3]); + y[il+24] = __float2bfloat16(dl2 * values2[(ql[il+16] >> 4) & 3]); + y[il+28] = __float2bfloat16(dl2 * values2[(ql[il+16] >> 6) & 3]); + } else { + y[il+ 0] = dl1 * values1[(ql[il+ 0] >> 0) & 3]; + y[il+ 4] = dl1 * values1[(ql[il+ 0] >> 2) & 3]; + y[il+ 8] = dl1 * values1[(ql[il+ 0] >> 4) & 3]; + y[il+12] = dl1 * values1[(ql[il+ 0] >> 6) & 3]; + y[il+16] = dl2 * values2[(ql[il+16] >> 0) & 3]; + y[il+20] = dl2 * values2[(ql[il+16] >> 2) & 3]; + y[il+24] = dl2 * values2[(ql[il+16] >> 4) & 3]; + y[il+28] = dl2 * values2[(ql[il+16] >> 6) & 3]; + } +} + template static __global__ void dequantize_block_iq3_k_r4(const void * __restrict__ vx, dst_t * __restrict__ yy, int64_t n_per_row, int64_t row_size) { @@ -1353,6 +1400,14 @@ static void dequantize_row_iq3_k_r4_cuda(const void * vx, dst_t * y, const int64 dequantize_block_iq3_k_r4<<>>(vx, y, n_per_row, row_size); } +template +static void dequantize_row_iq2_k_r4_cuda(const void * vx, dst_t * y, const int64_t nrows, const int64_t n_per_row, cudaStream_t stream) { + const int64_t k = nrows * n_per_row; + const int64_t row_size = ggml_row_size(GGML_TYPE_IQ4_K, n_per_row); + const int nb = (k + QK_K - 1) / QK_K; + dequantize_block_iq2_k_r4<<>>(vx, y, n_per_row, row_size); +} + template static void dequantize_row_iq4_k_cuda(const void * vx, dst_t * y, const int64_t nrows, const int64_t n_per_row, cudaStream_t stream) { const int64_t k = nrows * n_per_row; @@ -1479,6 +1534,8 @@ to_bf16_cuda_t ggml_get_to_bf16_cuda(ggml_type type) { return dequantize_row_iq5_k_cuda; case GGML_TYPE_IQ6_K: return dequantize_row_iq6_k_cuda; + case GGML_TYPE_IQ2_K_R4: + return dequantize_row_iq2_k_r4_cuda; case GGML_TYPE_IQ3_K_R4: return dequantize_row_iq3_k_r4_cuda; case GGML_TYPE_IQ4_K_R4: @@ -1567,6 +1624,8 @@ to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) { return convert_unary_cuda; case GGML_TYPE_BF16: return convert_from_bf16_cuda; + case GGML_TYPE_IQ2_K_R4: + return dequantize_row_iq2_k_r4_cuda; case GGML_TYPE_IQ3_K_R4: return dequantize_row_iq3_k_r4_cuda; case GGML_TYPE_IQ4_K_R4: @@ -1652,6 +1711,8 @@ to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) { return convert_unary_cuda; case GGML_TYPE_BF16: return convert_from_bf16_cuda; + case GGML_TYPE_IQ2_K_R4: + return dequantize_row_iq2_k_r4_cuda; case GGML_TYPE_IQ3_K_R4: return dequantize_row_iq3_k_r4_cuda; case GGML_TYPE_IQ4_K_R4: diff --git a/ggml/src/ggml-cuda/iqk_mmvq.cu b/ggml/src/ggml-cuda/iqk_mmvq.cu index df72e0c2..23f019ec 100644 --- a/ggml/src/ggml-cuda/iqk_mmvq.cu +++ b/ggml/src/ggml-cuda/iqk_mmvq.cu @@ -8,6 +8,13 @@ typedef void (*vec_dot_q_cuda_t)(const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs, float *); +template<> +struct ggml_cuda_type_traits { + static constexpr int qk = QK_K; + static constexpr int qr = QR4_XS; + static constexpr int qi = QI4_XS; +}; + template<> struct ggml_cuda_type_traits { static constexpr int qk = QK_K; @@ -494,7 +501,7 @@ __device__ __forceinline__ void vec_dot_iq3_k_r4_q8_1( // This is not faster. Why? //scales[1] = __vcmpeq4((scales_h[is] >> ib32) & 0x01010101, 0x01010101); //scales[0] = __vsub4(scales[0] ^ scales[1], scales[1]); - const int8_t * s8 = (const int8_t *)&scales; + const int8_t * s8 = (const int8_t *)scales; int2 val1; const int * q2 = (const int *)bq3->qs + 8*ib32 + 4*is; const int * qh = (const int *)bq3->qh + 4*ib32; @@ -520,6 +527,45 @@ __device__ __forceinline__ void vec_dot_iq3_k_r4_q8_1( } } +__device__ __forceinline__ void vec_dot_iq2_k_r4_q8_1( + const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs, float * result) { + + const block_iq2_k_r4 * bq2 = (const block_iq2_k_r4 *)vbq + kbx; + + // iqs is 0...30 in steps of 2 + const int ib16 = iqs/2; + const float d8 = __low2float(bq8_1[ib16/2].ds); + const int32_t * q8 = (const int *)bq8_1[ib16/2].qs + 4*(ib16%2); + + int ib32 = ib16/2; + int is = ib16%2; + const int * scales_l = (const int *)bq2->scales; + + int scales = __vsub4(((scales_l[2*(ib32%4)+is] >> 4*(ib32/4)) & 0x0f0f0f0f), 0x08080808); + const int8_t * s8 = (const int8_t *)&scales; + int2 val1; + const int * q2 = (const int *)bq2->qs + 8*ib32 + 4*is; + int aux32[2]; + const uint8_t * aux8 = (const uint8_t *)aux32; + for (int i = 0; i < 4; ++i) { + auto values1 = iq2nl_values + (((bq2->extra[i+4*is] >> ib32) & 1) << 2); + int sumi1 = 0; + aux32[0] = ((q2[i] >> 0) & 0x03030303); + aux32[1] = ((q2[i] >> 2) & 0x03030303); + // TODO: int_from_table_4 + val1.x = int_from_table(aux8+0, (const uint8_t *)values1); + val1.y = int_from_table(aux8+4, (const uint8_t *)values1); + sumi1 = ggml_cuda_dp4a(val1.x, q8[0], ggml_cuda_dp4a(val1.y, q8[1], sumi1)); + aux32[0] = ((q2[i] >> 4) & 0x03030303); + aux32[1] = ((q2[i] >> 6) & 0x03030303); + val1.x = int_from_table(aux8+0, (const uint8_t *)values1); + val1.y = int_from_table(aux8+4, (const uint8_t *)values1); + sumi1 = ggml_cuda_dp4a(val1.x, q8[2], ggml_cuda_dp4a(val1.y, q8[3], sumi1)); + const float d = __half2float(bq2->d[i]) * d8; + result[i] += d * sumi1 * s8[i]; + } +} + #define VDR_IQ6_K_Q8_1_MMVQ 4 #define VDR_IQ6_K_Q8_1_MMQ 4 @@ -973,6 +1019,14 @@ void mul_mat_vec_iq5_k_r4_q8_1_cuda( iqk_mul_mat_vec_q_cuda(vx, vy, dst, ids_data, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, ne2, nb02, nb12, nb2, ids_nb0, stream); } +void mul_mat_vec_iq2_k_r4_q8_1_cuda( + const void * vx, const void * vy, float * dst, const char * ids_data, + const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, + const int ne2, const uint64_t nb02, const uint64_t nb12, const uint64_t nb2, int64_t ids_nb0, cudaStream_t stream) { + + iqk_mul_mat_vec_q_cuda(vx, vy, dst, ids_data, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, ne2, nb02, nb12, nb2, ids_nb0, stream); +} + void mul_mat_vec_iq3_k_r4_q8_1_cuda( const void * vx, const void * vy, float * dst, const char * ids_data, const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, diff --git a/ggml/src/ggml-cuda/iqk_mmvq.cuh b/ggml/src/ggml-cuda/iqk_mmvq.cuh index 73c1a946..228c513b 100644 --- a/ggml/src/ggml-cuda/iqk_mmvq.cuh +++ b/ggml/src/ggml-cuda/iqk_mmvq.cuh @@ -61,6 +61,11 @@ void mul_mat_vec_iq2_bn_q8_1_cuda( const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, const int ne2, const uint64_t nb02, const uint64_t nb12, const uint64_t nb2, const int64_t ids_nb0, cudaStream_t stream); +void mul_mat_vec_iq2_k_r4_q8_1_cuda( + const void * vx, const void * vy, float * dst, const char * ids_data, + const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, + const int ne2, const uint64_t nb02, const uint64_t nb12, const uint64_t nb2, const int64_t ids_nb0, cudaStream_t stream); + void mul_mat_vec_iq3_k_r4_q8_1_cuda( const void * vx, const void * vy, float * dst, const char * ids_data, const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, diff --git a/ggml/src/ggml-cuda/mmvq.cu b/ggml/src/ggml-cuda/mmvq.cu index e1cc9bc0..d7bed266 100644 --- a/ggml/src/ggml-cuda/mmvq.cu +++ b/ggml/src/ggml-cuda/mmvq.cu @@ -542,6 +542,9 @@ static void ggml_cuda_op_mul_mat_vec_q_impl(ggml_backend_cuda_context & ctx, ggm case GGML_TYPE_IQ3_S: mul_mat_vec_iq3_s_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ids_data, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, ne2, nb02, nb12, nb2, ids_nb0, stream); break; + case GGML_TYPE_IQ2_K_R4: + mul_mat_vec_iq2_k_r4_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ids_data, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, ne2, nb02, nb12, nb2, ids_nb0, stream); + break; case GGML_TYPE_IQ3_K_R4: mul_mat_vec_iq3_k_r4_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ids_data, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, ne2, nb02, nb12, nb2, ids_nb0, stream); break; @@ -664,6 +667,7 @@ bool ggml_cuda_mmvq_type_supported(ggml_type src0_type) { case GGML_TYPE_IQ5_KS: case GGML_TYPE_IQ6_K: case GGML_TYPE_IQ3_S: + case GGML_TYPE_IQ2_K_R4: case GGML_TYPE_IQ3_K_R4: case GGML_TYPE_IQ4_K_R4: case GGML_TYPE_IQ5_K_R4: