local-llm-rn 1.0.0

package/cpp/ggml/src/ggml-opencl/kernels/scale.cl

@@ -0,0 +1,27 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+kernel void kernel_scale_f32(
+        global float * src0,
+        ulong offset0,
+        global float * dst,
+        ulong offsetd,
+        float scale,
+        float bias
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst = (global float*)((global char*)dst + offsetd);
+    dst[get_global_id(0)] = src0[get_global_id(0)] * scale + bias;
+}
+
+kernel void kernel_scale_f32_4(
+        global float4 * src0,
+        ulong offset0,
+        global float4 * dst,
+        ulong offsetd,
+        float scale,
+        float bias
+) {
+    src0 = (global float4*)((global char*)src0 + offset0);
+    dst = (global float4*)((global char*)dst + offsetd);
+    dst[get_global_id(0)] = src0[get_global_id(0)] * scale + bias;
+}

package/cpp/ggml/src/ggml-opencl/kernels/set_rows.cl

@@ -0,0 +1,208 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+// v = { mp, L, d }
+inline uint fastdiv(uint n, uint4 v) {
+    uint msbs;
+    msbs = mul_hi(n, v.s0);
+    return (msbs + n) >> v.s1;
+}
+inline uint fastmod(uint n, uint4 v) {
+    uint q = fastdiv(n, v);
+    return n - q * v.s2;
+}
+
+kernel void kernel_set_rows_f32_i64(
+        global char * src0,
+        ulong         offset0,
+        global char * src1,
+        ulong         offset1,
+        global char * dst,
+        ulong         offsetd,
+        int           ne01,
+        ulong         nb01,
+        ulong         nb02,
+        ulong         nb03,
+        uint4         ne11,
+        uint4         ne12,
+        ulong         nb10,
+        ulong         nb11,
+        ulong         nb12,
+        int           nblk0,
+        ulong         nb1,
+        ulong         nb2,
+        ulong         nb3
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst  + offsetd;
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0)*get_local_size(1) + get_local_id(1);
+
+    if (i01 >= ne01) {
+        return;
+    }
+
+    //int i12 = i03%ne12;
+    //int i11 = i02%ne11;
+    int i12 = fastmod(i03, ne12);
+    int i11 = fastmod(i02, ne11);
+
+    int i10 = i01;
+    long i1 = ((global long *)(src1 + i10*nb10 + i11*nb11 + i12*nb12))[0];
+
+    global float * dst_row = (global float *) (dst  +  i1*nb1  + i02*nb2  + i03*nb3);
+    global float * src_row = (global float *) (src0 + i01*nb01 + i02*nb02 + i03*nb03);
+
+    for (int ind = get_local_id(0); ind < nblk0; ind += get_local_size(0)) {
+        dst_row[ind] = (float)src_row[ind];
+    }
+}
+
+kernel void kernel_set_rows_f16_i64(
+        global char * src0,
+        ulong         offset0,
+        global char * src1,
+        ulong         offset1,
+        global char * dst,
+        ulong         offsetd,
+        int           ne01,
+        ulong         nb01,
+        ulong         nb02,
+        ulong         nb03,
+        uint4         ne11,
+        uint4         ne12,
+        ulong         nb10,
+        ulong         nb11,
+        ulong         nb12,
+        int           nblk0,
+        ulong         nb1,
+        ulong         nb2,
+        ulong         nb3
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst  + offsetd;
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0)*get_local_size(1) + get_local_id(1);
+
+    if (i01 >= ne01) {
+        return;
+    }
+
+    //int i12 = i03%ne12;
+    //int i11 = i02%ne11;
+    int i12 = fastmod(i03, ne12);
+    int i11 = fastmod(i02, ne11);
+
+    int i10 = i01;
+    long i1 = ((global long *)(src1 + i10*nb10 + i11*nb11 + i12*nb12))[0];
+
+    global half  * dst_row = (global half  *) (dst  +  i1*nb1  + i02*nb2  + i03*nb3);
+    global float * src_row = (global float *) (src0 + i01*nb01 + i02*nb02 + i03*nb03);
+
+    for (int ind = get_local_id(0); ind < nblk0; ind += get_local_size(0)) {
+        dst_row[ind] = src_row[ind];
+    }
+}
+
+kernel void kernel_set_rows_f32_i32(
+        global char * src0,
+        ulong         offset0,
+        global char * src1,
+        ulong         offset1,
+        global char * dst,
+        ulong         offsetd,
+        int           ne01,
+        ulong         nb01,
+        ulong         nb02,
+        ulong         nb03,
+        uint4         ne11,
+        uint4         ne12,
+        ulong         nb10,
+        ulong         nb11,
+        ulong         nb12,
+        int           nblk0,
+        ulong         nb1,
+        ulong         nb2,
+        ulong         nb3
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst  + offsetd;
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0)*get_local_size(1) + get_local_id(1);
+
+    if (i01 >= ne01) {
+        return;
+    }
+
+    //int i12 = i03%ne12;
+    //int i11 = i02%ne11;
+    int i12 = fastmod(i03, ne12);
+    int i11 = fastmod(i02, ne11);
+
+    int i10 = i01;
+    int i1  = ((global int *)(src1 + i10*nb10 + i11*nb11 + i12*nb12))[0];
+
+    global float * dst_row = (global float *) (dst  +  i1*nb1  + i02*nb2  + i03*nb3);
+    global float * src_row = (global float *) (src0 + i01*nb01 + i02*nb02 + i03*nb03);
+
+    for (int ind = get_local_id(0); ind < nblk0; ind += get_local_size(0)) {
+        dst_row[ind] = (float)src_row[ind];
+    }
+}
+
+kernel void kernel_set_rows_f16_i32(
+        global char * src0,
+        ulong         offset0,
+        global char * src1,
+        ulong         offset1,
+        global char * dst,
+        ulong         offsetd,
+        int           ne01,
+        ulong         nb01,
+        ulong         nb02,
+        ulong         nb03,
+        uint4         ne11,
+        uint4         ne12,
+        ulong         nb10,
+        ulong         nb11,
+        ulong         nb12,
+        int           nblk0,
+        ulong         nb1,
+        ulong         nb2,
+        ulong         nb3
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst  + offsetd;
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0)*get_local_size(1) + get_local_id(1);
+
+    if (i01 >= ne01) {
+        return;
+    }
+
+    //int i12 = i03%ne12;
+    //int i11 = i02%ne11;
+    int i12 = fastmod(i03, ne12);
+    int i11 = fastmod(i02, ne11);
+
+    int i10 = i01;
+    int i1  = ((global int *)(src1 + i10*nb10 + i11*nb11 + i12*nb12))[0];
+
+    global half  * dst_row = (global half  *) (dst  +  i1*nb1  + i02*nb2  + i03*nb3);
+    global float * src_row = (global float *) (src0 + i01*nb01 + i02*nb02 + i03*nb03);
+
+    for (int ind = get_local_id(0); ind < nblk0; ind += get_local_size(0)) {
+        dst_row[ind] = src_row[ind];
+    }
+}

package/cpp/ggml/src/ggml-opencl/kernels/sigmoid.cl

@@ -0,0 +1,29 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// sigmoid
+//------------------------------------------------------------------------------
+
+kernel void kernel_sigmoid_f32(
+        global float * src0,
+        ulong offset0,
+        global float * dst,
+        ulong offsetd
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    dst[get_global_id(0)] = 1.0f / (1.0f + exp(-src0[get_global_id(0)]));
+}
+
+kernel void kernel_sigmoid_f16(
+        global half * src0,
+        ulong offset0,
+        global half * dst,
+        ulong offsetd
+) {
+    src0 = (global half*)((global char*)src0 + offset0);
+    dst = (global half*)((global char*)dst + offsetd);
+
+    dst[get_global_id(0)] = 1.0f / (1.0f + exp(-src0[get_global_id(0)]));
+}

package/cpp/ggml/src/ggml-opencl/kernels/silu.cl

@@ -0,0 +1,30 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// silu
+//------------------------------------------------------------------------------
+kernel void kernel_silu(
+        global float * src0,
+        ulong offset0,
+        global float * dst,
+        ulong offsetd
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    float x = src0[get_global_id(0)];
+    dst[get_global_id(0)] = x / (1.0f + exp(-x));
+}
+
+kernel void kernel_silu_4(
+        global float4 * src0,
+        ulong offset0,
+        global float4 * dst,
+        ulong offsetd
+) {
+    src0 = (global float4*)((global char*)src0 + offset0);
+    dst = (global float4*)((global char*)dst + offsetd);
+
+    float4 x = src0[get_global_id(0)];
+    dst[get_global_id(0)] = x / (1.0f + exp(-x));
+}

package/cpp/ggml/src/ggml-opencl/kernels/softmax_4_f16.cl

@@ -0,0 +1,108 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_soft_max_4_f16(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global char * src2,
+        ulong offset2,
+        global char * dst,
+        ulong offsetd,
+        int ne00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne12,
+        int ne13,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3,
+        float scale,
+        float max_bias,
+        float m0,
+        float m1,
+        int n_head_log2
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    src2 = src2 + offset2;
+    dst  = dst  + offsetd;
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    int i13 = i03%ne13;
+    int i12 = i02%ne12;
+    int i11 = i01;
+
+    global float4 * psrc4 = (global float4 *)(src0 + i01*nb01 + i02*nb02 + i03*nb03);
+    global half4  * pmask = src1 != src0 ? (global half4 *)(src1 + i11*nb11 + i12*nb12 + i13*nb13) : 0;
+    global float  * psrc2 = src2 != src0 ? (global float *)(src2) : 0;
+    global float4 * pdst4 = (global float4 *)(dst  + i01*nb1 + i02*nb2 + i03*nb3);
+
+    float slope = 1.0f;
+
+    // ALiBi
+    if (max_bias > 0.0f) {
+        int h = i02;
+
+        float base = h < n_head_log2 ? m0 : m1;
+        int   exp  = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+        slope = pow(base, exp);
+    }
+
+    // parallel max
+    float4 lmax4 = psrc2 ? psrc2[i02] : -INFINITY;
+    for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) {
+        lmax4 = fmax(lmax4, psrc4[i00]*scale + slope*(pmask ? convert_float4(pmask[i00]) : 0.0f));
+    }
+    float lmax = fmax(fmax(lmax4.s0, lmax4.s1), fmax(lmax4.s2, lmax4.s3));
+
+    const float max = sub_group_reduce_max(lmax);
+
+    // parallel sum
+    float4 lsum4 = 0.0f;
+    for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) {
+        const float4 exp_psrc4 = exp((psrc4[i00]*scale + slope*(pmask ? convert_float4(pmask[i00]) : 0.0f)) - max);
+        lsum4 += exp_psrc4;
+        pdst4[i00] = exp_psrc4;
+    }
+    float lsum = lsum4.s0 + lsum4.s1 + lsum4.s2 + lsum4.s3;
+
+    float sum = sub_group_reduce_add(lsum);
+
+    if (psrc2) {
+        sum += exp(psrc2[i02] - max);
+    }
+
+    for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) {
+        pdst4[i00] /= sum;
+    }
+}

package/cpp/ggml/src/ggml-opencl/kernels/softmax_4_f32.cl

@@ -0,0 +1,108 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_soft_max_4(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global char * src2,
+        ulong offset2,
+        global char * dst,
+        ulong offsetd,
+        int ne00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne12,
+        int ne13,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3,
+        float scale,
+        float max_bias,
+        float m0,
+        float m1,
+        int n_head_log2
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    src2 = src2 + offset2;
+    dst  = dst  + offsetd;
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    int i13 = i03%ne13;
+    int i12 = i02%ne12;
+    int i11 = i01;
+
+    global float4 * psrc4 = (global float4 *)(src0 + i01*nb01 + i02*nb02 + i03*nb03);
+    global float4 * pmask = src1 != src0 ? (global float4 *)(src1 + i11*nb11 + i12*nb12 + i13*nb13) : 0;
+    global float  * psrc2 = src2 != src0 ? (global float  *)(src2) : 0;
+    global float4 * pdst4 = (global float4 *)(dst  + i01*nb1 + i02*nb2 + i03*nb3);
+
+    float slope = 1.0f;
+
+    // ALiBi
+    if (max_bias > 0.0f) {
+        int h = i02;
+
+        float base = h < n_head_log2 ? m0 : m1;
+        int   exp  = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+        slope = pow(base, exp);
+    }
+
+    // parallel max
+    float4 lmax4 = psrc2 ? psrc2[i02] : -INFINITY;
+    for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) {
+        lmax4 = fmax(lmax4, psrc4[i00]*scale + (pmask ? slope*pmask[i00] : 0.0f));
+    }
+    float lmax = fmax(fmax(lmax4.s0, lmax4.s1), fmax(lmax4.s2, lmax4.s3));
+
+    const float max = sub_group_reduce_max(lmax);
+
+    // parallel sum
+    float4 lsum4 = 0.0f;
+    for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) {
+        const float4 exp_psrc4 = exp((psrc4[i00]*scale + (pmask ? slope*pmask[i00] : 0.0f)) - max);
+        lsum4 += exp_psrc4;
+        pdst4[i00] = exp_psrc4;
+    }
+    float lsum = lsum4.s0 + lsum4.s1 + lsum4.s2 + lsum4.s3;
+
+    float sum = sub_group_reduce_add(lsum);
+
+    if (psrc2) {
+        sum += exp(psrc2[i02] - max);
+    }
+
+    for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) {
+        pdst4[i00] /= sum;
+    }
+}

package/cpp/ggml/src/ggml-opencl/kernels/softmax_f16.cl

@@ -0,0 +1,107 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_soft_max_f16(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global char * src2,
+        ulong offset2,
+        global char * dst,
+        ulong offsetd,
+        int ne00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne12,
+        int ne13,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3,
+        float scale,
+        float max_bias,
+        float m0,
+        float m1,
+        int n_head_log2
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    src2 = src2 + offset2;
+    dst  = dst  + offsetd;
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    int i13 = i03%ne13;
+    int i12 = i02%ne12;
+    int i11 = i01;
+
+    global float * psrc0 = (global float *)(src0 + i01*nb01 + i02*nb02 + i03*nb03);
+    global half  * pmask = src1 != src0 ? (global half *)(src1 + i11*nb11 + i12*nb12 + i13*nb13) : 0;
+    global float * psrc2 = src2 != src0 ? (global float *)(src2) : 0;
+    global float * pdst  = (global float *)(dst  + i01*nb1 + i02*nb2 + i03*nb3);
+
+    float slope = 1.0f;
+
+    // ALiBi
+    if (max_bias > 0.0f) {
+        int h = i02;
+
+        float base = h < n_head_log2 ? m0 : m1;
+        int   exp  = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+        slope = pow(base, exp);
+    }
+
+    // parallel max
+    float lmax = psrc2 ? psrc2[i02] : -INFINITY;
+    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
+        lmax = fmax(lmax, psrc0[i00]*scale + (pmask ? slope*pmask[i00] : 0.0f));
+    }
+    float max = sub_group_reduce_max(lmax);
+
+    // parallel sum
+    float lsum = 0.0f;
+    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
+        float exp_psrc0 = exp((psrc0[i00]*scale + (pmask ? slope*pmask[i00] : 0.0f)) - max);
+        lsum += exp_psrc0;
+        // Remember the result of exp here. exp is expensive, so we really do not
+        // wish to compute it twice.
+        pdst[i00] = exp_psrc0;
+    }
+
+    float sum = sub_group_reduce_add(lsum);
+
+    if (psrc2) {
+        sum += exp(psrc2[i02] - max);
+    }
+
+    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
+        pdst[i00] /= sum;
+    }
+}