CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-04 09:32:00 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'master' into compilade/mamba2

Browse Source
This commit is contained in:
Francis Couture-Harpin
2025-06-10 19:22:15 -04:00
parent e94f3932f2 dad5c44398
commit 9864bfcd01
581 changed files with 159374 additions and 37809 deletions
Download Patch File Download Diff File Expand all files Collapse all files

360
ggml/src/ggml-cpu/ops.cpp
View File

@@ -8,19 +8,6 @@
#include <float.h>
#if defined(_MSC_VER)
// disable "possible loss of data" to avoid hundreds of casts
// we should just be careful :)
#pragma warning(disable: 4244 4267)
// disable POSIX deprecation warnings
// these functions are never going away, anyway
#pragma warning(disable: 4996)
// unreachable code because of multiple instances of code after GGML_ABORT
#pragma warning(disable: 4702)
#endif
// ggml_compute_forward_dup
static void ggml_compute_forward_dup_same_cont(
@@ -2704,6 +2691,109 @@ static void ggml_compute_forward_gelu(
}
}
// ggml_compute_forward_gelu_erf
static void ggml_compute_forward_gelu_erf_f32(
const ggml_compute_params * params,
ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
assert(ggml_is_contiguous_1(src0));
assert(ggml_is_contiguous_1(dst));
assert(ggml_are_same_shape(src0, dst));
const int ith = params->ith;
const int nth = params->nth;
const int nc = src0->ne[0];
const int nr = ggml_nrows(src0);
// rows per thread
const int dr = (nr + nth - 1)/nth;
// row range for this thread
const int ir0 = dr*ith;
const int ir1 = MIN(ir0 + dr, nr);
for (int i1 = ir0; i1 < ir1; i1++) {
ggml_vec_gelu_erf_f32(nc,
(float *) ((char *) dst->data + i1*( dst->nb[1])),
(float *) ((char *) src0->data + i1*(src0->nb[1])));
#ifndef NDEBUG
for (int k = 0; k < nc; k++) {
const float x = ((float *) ((char *) dst->data + i1*( dst->nb[1])))[k];
GGML_UNUSED(x);
assert(!isnan(x));
assert(!isinf(x));
}
#endif
}
}
static void ggml_compute_forward_gelu_erf_f16(
const ggml_compute_params * params,
ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
assert(ggml_is_contiguous_1(src0));
assert(ggml_is_contiguous_1(dst));
assert(ggml_are_same_shape(src0, dst));
const int ith = params->ith;
const int nth = params->nth;
const int nc = src0->ne[0];
const int nr = ggml_nrows(src0);
// rows per thread
const int dr = (nr + nth - 1)/nth;
// row range for this thread
const int ir0 = dr*ith;
const int ir1 = MIN(ir0 + dr, nr);
for (int i1 = ir0; i1 < ir1; i1++) {
ggml_vec_gelu_erf_f16(nc,
(ggml_fp16_t *) ((char *) dst->data + i1*( dst->nb[1])),
(ggml_fp16_t *) ((char *) src0->data + i1*(src0->nb[1])));
#ifndef NDEBUG
for (int k = 0; k < nc; k++) {
const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i1*( dst->nb[1])))[k];
const float v = GGML_FP16_TO_FP32(x);
GGML_UNUSED(v);
assert(!isnan(v));
assert(!isinf(v));
}
#endif
}
}
static void ggml_compute_forward_gelu_erf(
const ggml_compute_params * params,
ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
switch (src0->type) {
case GGML_TYPE_F32:
{
ggml_compute_forward_gelu_erf_f32(params, dst);
} break;
case GGML_TYPE_F16:
{
ggml_compute_forward_gelu_erf_f16(params, dst);
} break;
default:
{
GGML_ABORT("fatal error");
}
}
}
// ggml_compute_forward_gelu_quick
static void ggml_compute_forward_gelu_quick_f32(
@@ -7632,6 +7722,30 @@ static void ggml_compute_forward_ssm_scan_f32(
for (int i1 = 0; i1 < nr; ++i1) {
const int ii = i1 + h*nr;
const float x_dt = x[ii] * dt_soft_plus;
#ifdef __ARM_FEATURE_SVE
svfloat32_t vx_dt = GGML_F32_VEC_SET1(x_dt);
svfloat32_t vdt_soft_plus = GGML_F32_VEC_SET1(dt_soft_plus);
svfloat32_t r1_vector = GGML_F32_VEC_ZERO;
// d_state
// TODO: what happens when (d_state % svcntw()) != 0?
for (int64_t k = 0; k < nc; k += svcntw()) {
svfloat32_t vA = GGML_F32_VEC_LOAD(&A[h*nc + k]);
svfloat32_t vB = GGML_F32_VEC_LOAD(&B[k + (h & (ng - 1))*nc]);
svfloat32_t vC = GGML_F32_VEC_LOAD(&C[k + (h & (ng - 1))*nc]);
svfloat32_t vs0 = GGML_F32_VEC_LOAD(&s0[ii*nc + k]);
svfloat32_t t1 = GGML_F32_VEC_MUL(vdt_soft_plus, vA);
t1 = exp_ps_sve(svptrue_b32(), t1);
svfloat32_t t2 = GGML_F32_VEC_MUL(vx_dt, vB);
vs0 = GGML_F32_VEC_FMA(vs0, t1, t2);
r1_vector = GGML_F32_VEC_ADD(GGML_F32_VEC_MUL(vs0, vC), r1_vector);
GGML_F32_VEC_STORE(&s[ii*nc + k], vs0);
}
y[ii] = GGML_F32xt_REDUCE_ONE(r1_vector);
#else
float sumf = 0.0f;
// NOTE: can't really use GGML_SIMD here because d_state is usually 16
// and also because expf is used within the loop.
@@ -7646,6 +7760,7 @@ static void ggml_compute_forward_ssm_scan_f32(
s[i] = state;
}
y[ii] = sumf;
#endif
}
}
}
@@ -7839,6 +7954,10 @@ void ggml_compute_forward_unary(
{
ggml_compute_forward_gelu(params, dst);
} break;
case GGML_UNARY_OP_GELU_ERF:
{
ggml_compute_forward_gelu_erf(params, dst);
} break;
case GGML_UNARY_OP_GELU_QUICK:
{
ggml_compute_forward_gelu_quick(params, dst);
@@ -8053,6 +8172,14 @@ static void ggml_compute_forward_rwkv_wkv6_f32(
#define GGML_F32X_MUL GGML_F32x16_MUL
#define GGML_F32X_FMA GGML_F32x16_FMA
#define WKV_VECTOR_SIZE 16
#elif defined(__ARM_FEATURE_SVE) && defined(__aarch64__)
#define GGML_F32X GGML_F32xt
#define GGML_F32X_SET1 GGML_F32xt_SET1
#define GGML_F32X_LOAD GGML_F32xt_LOAD
#define GGML_F32X_STORE GGML_F32xt_STORE
#define GGML_F32X_MUL GGML_F32xt_MUL
#define GGML_F32X_FMA GGML_F32xt_FMA
#define WKV_VECTOR_SIZE 8
#elif defined(__ARM_NEON) && defined(__aarch64__)
#define GGML_F32X GGML_F32x4
#define GGML_F32X_SET1 GGML_F32x4_SET1
@@ -8064,7 +8191,13 @@ static void ggml_compute_forward_rwkv_wkv6_f32(
#endif
#ifdef WKV_VECTOR_SIZE
const int64_t vec_count = head_size / WKV_VECTOR_SIZE;
int wkv_vector_size;
#if defined(__ARM_FEATURE_SVE)
wkv_vector_size = svcntw();
#else
wkv_vector_size = WKV_VECTOR_SIZE;
#endif
const int64_t vec_count = head_size / wkv_vector_size;
for (int64_t t = 0; t < T; t++) {
size_t t_offset = t * t_stride;
@@ -8094,7 +8227,7 @@ static void ggml_compute_forward_rwkv_wkv6_f32(
GGML_F32X time_decay_vec = GGML_F32X_SET1(time_decay_val);
for (int64_t j = 0; j < vec_count; j++) {
size_t base_j = j * WKV_VECTOR_SIZE;
size_t base_j = j * wkv_vector_size;
size_t t_h_j_offset = t_h_offset + base_j;
size_t h_2d_i_j_offset = h_2d_i_offset + base_j;
@@ -8119,7 +8252,7 @@ static void ggml_compute_forward_rwkv_wkv6_f32(
}
// Handle remaining elements, this will not be used.
for (int64_t j = vec_count * WKV_VECTOR_SIZE; j < head_size; j++) {
for (int64_t j = vec_count * wkv_vector_size; j < head_size; j++) {
size_t t_h_j_offset = t_h_offset + j;
size_t h_2d_i_j_offset = h_2d_i_offset + j;
float v_val = v[t_h_j_offset];
@@ -8255,6 +8388,14 @@ static void ggml_compute_forward_gla_f32(
#define GGML_F32X_MUL GGML_F32x16_MUL
#define GGML_F32X_FMA GGML_F32x16_FMA
#define GLA_VECTOR_SIZE 16
#elif defined(__ARM_FEATURE_SVE) && defined(__aarch64__)
#define GGML_F32X GGML_F32xt
#define GGML_F32X_SET1 GGML_F32xt_SET1
#define GGML_F32X_LOAD GGML_F32xt_LOAD
#define GGML_F32X_STORE GGML_F32xt_STORE
#define GGML_F32X_MUL GGML_F32xt_MUL
#define GGML_F32X_FMA GGML_F32xt_FMA
#define GLA_VECTOR_SIZE 8
#elif defined(__ARM_NEON) && defined(__aarch64__)
#define GGML_F32X GGML_F32x4
#define GGML_F32X_SET1 GGML_F32x4_SET1
@@ -8266,7 +8407,13 @@ static void ggml_compute_forward_gla_f32(
#endif
#ifdef GLA_VECTOR_SIZE
const int64_t vec_count = head_size / GLA_VECTOR_SIZE;
int gla_vector_size;
#if defined(__ARM_FEATURE_SVE)
gla_vector_size = svcntw();
#else
gla_vector_size = GLA_VECTOR_SIZE;
#endif
const int64_t vec_count = head_size / gla_vector_size;
for (int64_t t = 0; t < T; t++) {
size_t t_offset = t * t_stride;
@@ -8293,7 +8440,7 @@ static void ggml_compute_forward_gla_f32(
GGML_F32X g_vec = GGML_F32X_SET1(g_val);
for (int64_t j = 0; j < vec_count; j++) {
size_t base_j = j * GLA_VECTOR_SIZE;
size_t base_j = j * gla_vector_size;
size_t t_h_j_offset = t_h_offset + base_j;
size_t h_2d_i_j_offset = h_2d_i_offset + base_j;
@@ -8317,7 +8464,7 @@ static void ggml_compute_forward_gla_f32(
}
// Handle remaining elements, this will not be used.
for (int64_t j = vec_count * GLA_VECTOR_SIZE; j < head_size; j++) {
for (int64_t j = vec_count * gla_vector_size; j < head_size; j++) {
size_t t_h_j_offset = t_h_offset + j;
size_t h_2d_i_j_offset = h_2d_i_offset + j;
float v_val = v[t_h_j_offset];
@@ -8426,83 +8573,126 @@ static void ggml_compute_forward_rwkv_wkv7_f32(
int64_t h_stride_2d = head_size * head_size;
#if defined(GGML_SIMD)
for (int64_t t = 0; t < T; t++) {
int64_t t_offset = t * t_stride;
int64_t state_offset = head_size * C * (t / (T / n_seqs));
float * state_cur = state + state_offset;
float * state_prev = t % (T / n_seqs) ? state_cur : (float*)dst->src[6]->data + state_offset;
#if defined(__ARM_FEATURE_SVE)
// scalar Route to scalar implementation //TODO: Write SVE code
for (int64_t t = 0; t < T; t++) {
int64_t t_offset = t * t_stride;
int64_t state_offset = head_size * C * (t / (T / n_seqs));
float * state_cur = state + state_offset;
float * state_prev = t % (T / n_seqs) ? state_cur : (float*)dst->src[6]->data + state_offset;
for (int64_t h = h_start; h < h_end; h++) {
int64_t h_offset = h * h_stride;
int64_t t_h_offset = t_offset + h_offset;
int64_t h_2d_offset = h * h_stride_2d;
for (int64_t h = h_start; h < h_end; h++) {
int64_t h_offset = h * h_stride;
int64_t t_h_offset = t_offset + h_offset;
int64_t h_2d_offset = h * h_stride_2d;
for (int64_t ii = 0; ii < head_size; ii++) {
int64_t t_h_i_offset = t_h_offset + ii;
int64_t h_2d_i_offset = h_2d_offset + ii * h_stride;
for (int64_t i = 0; i < head_size; i++) {
int64_t t_h_i_offset = t_h_offset + i;
int64_t h_2d_i_offset = h_2d_offset + i * h_stride;
GGML_F32_VEC v_vec = GGML_F32_VEC_SET1(v[t_h_i_offset]);
float v_val = v[t_h_i_offset];
float sa = 0;
{
GGML_F32_VEC sum[GGML_F32_ARR] = { GGML_F32_VEC_ZERO };
GGML_F32_VEC ax[GGML_F32_ARR];
GGML_F32_VEC ay[GGML_F32_ARR];
for (int64_t j = 0; j < head_size; j += GGML_F32_STEP) {
for (int64_t kk = 0; kk < GGML_F32_ARR; kk++) {
ax[kk] = GGML_F32_VEC_LOAD(&a[t_h_offset + j + kk * GGML_F32_EPR]);
ay[kk] = GGML_F32_VEC_LOAD(&state_prev[h_2d_i_offset + j + kk * GGML_F32_EPR]);
sum[kk] = GGML_F32_VEC_FMA(sum[kk], ax[kk], ay[kk]);
}
float sa = 0, result = 0;
for (int64_t j = 0; j < head_size; j++) {
sa += a[t_h_offset + j] * state_prev[h_2d_i_offset + j];
}
GGML_F32_VEC_REDUCE(sa, sum);
}
GGML_F32_VEC sa_vec = GGML_F32_VEC_SET1(sa);
for (int64_t j = 0; j < head_size; j++) {
int64_t t_h_j_offset = t_h_offset + j;
int64_t h_2d_i_j_offset = h_2d_i_offset + j;
int64_t j = 0;
GGML_F32_VEC result_vec[GGML_F32_ARR] = { GGML_F32_VEC_ZERO };
for (; j < head_size; j += GGML_F32_STEP) {
for (int64_t kk = 0; kk < GGML_F32_ARR; kk++) {
int64_t t_h_j_offset = t_h_offset + j + kk * GGML_F32_EPR;
int64_t h_2d_i_j_offset = h_2d_i_offset + j + kk * GGML_F32_EPR;
GGML_F32_VEC r_vec = GGML_F32_VEC_LOAD(&r[t_h_j_offset]);
GGML_F32_VEC w_vec = GGML_F32_VEC_LOAD(&w[t_h_j_offset]);
GGML_F32_VEC k_vec = GGML_F32_VEC_LOAD(&k[t_h_j_offset]);
GGML_F32_VEC b_vec = GGML_F32_VEC_LOAD(&b[t_h_j_offset]);
k_vec = GGML_F32_VEC_MUL(v_vec, k_vec);
GGML_F32_VEC state_vec = GGML_F32_VEC_LOAD(&state_prev[h_2d_i_j_offset]);
// kv + s * decay + sa * b
state_vec = GGML_F32_VEC_FMA(k_vec, state_vec, w_vec);
state_vec = GGML_F32_VEC_FMA(state_vec, sa_vec, b_vec);
GGML_F32_VEC_STORE(&state_cur[h_2d_i_j_offset], state_vec);
result_vec[kk] = GGML_F32_VEC_FMA(result_vec[kk], state_vec, r_vec);
float r_val = r[t_h_j_offset];
float w_val = w[t_h_j_offset];
float k_val = k[t_h_j_offset];
float b_val = b[t_h_j_offset];
float kv_val = v_val * k_val;
float prev_state_val = state_prev[h_2d_i_j_offset];
state_cur[h_2d_i_j_offset] = prev_state_val * w_val + kv_val + sa * b_val;
result += state_cur[h_2d_i_j_offset] * r_val;
}
}
GGML_F32_VEC_REDUCE(dst_data[t_h_i_offset], result_vec);
// There shouldn't be left-overs though.
for (; j < head_size; j++) {
int64_t t_h_j_offset = t_h_offset + j;
int64_t h_2d_i_j_offset = h_2d_i_offset + j;
float r_val = r[t_h_j_offset];
float w_val = w[t_h_j_offset];
float k_val = k[t_h_j_offset];
float b_val = b[t_h_j_offset];
float kv_val = v[t_h_i_offset] * k_val;
float prev_state_val = state_prev[h_2d_i_j_offset];
state_cur[h_2d_i_j_offset] = prev_state_val * w_val + kv_val + sa * b_val;
dst_data[t_h_i_offset] += state_cur[h_2d_i_j_offset] * r_val;
dst_data[t_h_i_offset] = result;
}
}
}
}
#else
for (int64_t t = 0; t < T; t++) {
int64_t t_offset = t * t_stride;
int64_t state_offset = head_size * C * (t / (T / n_seqs));
float * state_cur = state + state_offset;
float * state_prev = t % (T / n_seqs) ? state_cur : (float*)dst->src[6]->data + state_offset;
for (int64_t h = h_start; h < h_end; h++) {
int64_t h_offset = h * h_stride;
int64_t t_h_offset = t_offset + h_offset;
int64_t h_2d_offset = h * h_stride_2d;
for (int64_t ii = 0; ii < head_size; ii++) {
int64_t t_h_i_offset = t_h_offset + ii;
int64_t h_2d_i_offset = h_2d_offset + ii * h_stride;
GGML_F32_VEC v_vec = GGML_F32_VEC_SET1(v[t_h_i_offset]);
float sa = 0;
{
GGML_F32_VEC sum[GGML_F32_ARR] = { GGML_F32_VEC_ZERO };
GGML_F32_VEC ax[GGML_F32_ARR];
GGML_F32_VEC ay[GGML_F32_ARR];
for (int64_t j = 0; j < head_size; j += GGML_F32_STEP) {
for (int64_t kk = 0; kk < GGML_F32_ARR; kk++) {
ax[kk] = GGML_F32_VEC_LOAD(&a[t_h_offset + j + kk * GGML_F32_EPR]);
ay[kk] = GGML_F32_VEC_LOAD(&state_prev[h_2d_i_offset + j + kk * GGML_F32_EPR]);
sum[kk] = GGML_F32_VEC_FMA(sum[kk], ax[kk], ay[kk]);
}
}
GGML_F32_VEC_REDUCE(sa, sum);
}
GGML_F32_VEC sa_vec = GGML_F32_VEC_SET1(sa);
int64_t j = 0;
GGML_F32_VEC result_vec[GGML_F32_ARR] = { GGML_F32_VEC_ZERO };
for (; j < head_size; j += GGML_F32_STEP) {
for (int64_t kk = 0; kk < GGML_F32_ARR; kk++) {
int64_t t_h_j_offset = t_h_offset + j + kk * GGML_F32_EPR;
int64_t h_2d_i_j_offset = h_2d_i_offset + j + kk * GGML_F32_EPR;
GGML_F32_VEC r_vec = GGML_F32_VEC_LOAD(&r[t_h_j_offset]);
GGML_F32_VEC w_vec = GGML_F32_VEC_LOAD(&w[t_h_j_offset]);
GGML_F32_VEC k_vec = GGML_F32_VEC_LOAD(&k[t_h_j_offset]);
GGML_F32_VEC b_vec = GGML_F32_VEC_LOAD(&b[t_h_j_offset]);
k_vec = GGML_F32_VEC_MUL(v_vec, k_vec);
GGML_F32_VEC state_vec = GGML_F32_VEC_LOAD(&state_prev[h_2d_i_j_offset]);
// kv + s * decay + sa * b
state_vec = GGML_F32_VEC_FMA(k_vec, state_vec, w_vec);
state_vec = GGML_F32_VEC_FMA(state_vec, sa_vec, b_vec);
GGML_F32_VEC_STORE(&state_cur[h_2d_i_j_offset], state_vec);
result_vec[kk] = GGML_F32_VEC_FMA(result_vec[kk], state_vec, r_vec);
}
}
GGML_F32_VEC_REDUCE(dst_data[t_h_i_offset], result_vec);
// There shouldn't be left-overs though.
for (; j < head_size; j++) {
int64_t t_h_j_offset = t_h_offset + j;
int64_t h_2d_i_j_offset = h_2d_i_offset + j;
float r_val = r[t_h_j_offset];
float w_val = w[t_h_j_offset];
float k_val = k[t_h_j_offset];
float b_val = b[t_h_j_offset];
float kv_val = v[t_h_i_offset] * k_val;
float prev_state_val = state_prev[h_2d_i_j_offset];
state_cur[h_2d_i_j_offset] = prev_state_val * w_val + kv_val + sa * b_val;
dst_data[t_h_i_offset] += state_cur[h_2d_i_j_offset] * r_val;
}
}
}
}
#endif
#else
for (int64_t t = 0; t < T; t++) {
int64_t t_offset = t * t_stride;