EngineX-Ascend/enginex-ascend-910-llama.cpp
10
0
Fork 0
Code Issues Pull Requests Actions 4 Projects Releases Wiki Activity

llama : add gpt-oss (#15091)

Browse Source 
* oai moe

* compat with new checkpoint

* add attn sink impl

* add rope scaling yarn

* logits match with latest transformers code

* wip chat template

* rm trailing space

* use ggml_scale_bias

* rm redundant is_swa_all

* convert interleaved gate_up

* graph : fix activation function to match reference (#7)

* vocab : handle o200k_harmony special tokens

* ggml : add attention sinks support (#1)

* llama : add attn sinks

* ggml : add attn sinks

* cuda : add attn sinks

* vulkan : add support for sinks in softmax

remove unnecessary return

* ggml : add fused swiglu_oai op (#11)

* ggml : add fused swiglu_oai op

* Update ggml/src/ggml-cpu/ops.cpp

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* update CUDA impl

* cont : metal impl

* add vulkan impl

* test-backend-ops : more test cases, clean up

* llama : remove unfused impl

* remove extra lines

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

---------

Co-authored-by: slaren <slarengh@gmail.com>

* repack mxfp4 upon conversion

* clean up a bit

* enable thinking

* add quick hack to render only some special tokens

* fix bf16 conversion

* remove vocab hack

* webui ok

* support chat parsing for gpt-oss

* fix webui

* direct mapping mxfp4, FINALLY

* force using mxfp4

* properly use lazy tensor

* ggml : add mxfp4

ggml : use e8m0 conversion instead of powf

Co-authored-by: Diego Devesa <slarengh@gmail.com>

change kvalues_mxfp4 table to match e2m1 (#6)

metal : remove quantization for now (not used)

cuda : fix disabled CUDA graphs due to ffn moe bias

vulkan : add support for mxfp4

cont : add cm2 dequant

* ggml : add ggml_add_id (#13)

* ggml : add ggml_add_id

* add cuda impl

* llama : add weight support check for add_id

* perf opt

* add vulkan impl

* rename cuda files

* add metal impl

* allow in-place ggml_add_id

* llama : keep biases on CPU with --cpu-moe

* llama : fix compile error

ggml-ci

* cuda : add fallback for __nv_cvt_e8m0_to_bf16raw

ggml-ci

* cleanup

ggml-ci

* sycl : fix supports_op for MXFP4

ggml-ci

* fix Unknown reasoning format

* ggml-cpu : fix AVX build

ggml-ci

* fix hip build

ggml-ci

* cuda : add mxfp4 dequantization support for cuBLAS

ggml-ci

* ggml-cpu : fix mxfp4 fallback definitions for some architectures

ggml-ci

* cuda : fix version required for __nv_cvt_e8m0_to_bf16raw

---------

Co-authored-by: Xuan Son Nguyen <son@huggingface.co>
Co-authored-by: slaren <slarengh@gmail.com>
This commit is contained in:
Georgi Gerganov
2025-08-05 22:10:36 +03:00
committed by GitHub
parent f324a3b715
commit fd1234cb46
83 changed files with 2942 additions and 227 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
ggml/src/ggml-cpu/arch-fallback.h
View File

@@ -13,6 +13,7 @@
#define ggml_vec_dot_q5_0_q8_0_generic ggml_vec_dot_q5_0_q8_0
#define ggml_vec_dot_q5_1_q8_1_generic ggml_vec_dot_q5_1_q8_1
#define ggml_vec_dot_q8_0_q8_0_generic ggml_vec_dot_q8_0_q8_0
#define ggml_vec_dot_mxfp4_q8_0_generic ggml_vec_dot_mxfp4_q8_0
#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
#define ggml_vec_dot_q2_K_q8_K_generic ggml_vec_dot_q2_K_q8_K
@@ -68,6 +69,7 @@
#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
#define ggml_vec_dot_mxfp4_q8_0_generic ggml_vec_dot_mxfp4_q8_0
// repack.cpp
#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
@@ -90,6 +92,7 @@
#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
#define ggml_vec_dot_mxfp4_q8_0_generic ggml_vec_dot_mxfp4_q8_0
// repack.cpp
#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
@@ -120,6 +123,7 @@
#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
#define ggml_vec_dot_iq4_nl_q8_0_generic ggml_vec_dot_iq4_nl_q8_0
#define ggml_vec_dot_iq4_xs_q8_K_generic ggml_vec_dot_iq4_xs_q8_K
#define ggml_vec_dot_mxfp4_q8_0_generic ggml_vec_dot_mxfp4_q8_0
// repack.cpp
#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
@@ -149,6 +153,7 @@
#define ggml_vec_dot_iq3_s_q8_K_generic ggml_vec_dot_iq3_s_q8_K
#define ggml_vec_dot_iq1_s_q8_K_generic ggml_vec_dot_iq1_s_q8_K
#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
#define ggml_vec_dot_mxfp4_q8_0_generic ggml_vec_dot_mxfp4_q8_0
// repack.cpp
#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
@@ -179,6 +184,7 @@
#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
#define ggml_vec_dot_iq4_nl_q8_0_generic ggml_vec_dot_iq4_nl_q8_0
#define ggml_vec_dot_iq4_xs_q8_K_generic ggml_vec_dot_iq4_xs_q8_K
#define ggml_vec_dot_mxfp4_q8_0_generic ggml_vec_dot_mxfp4_q8_0
// repack.cpp
#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8

61
ggml/src/ggml-cpu/arch/arm/quants.c
View File

@@ -589,6 +589,67 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
*s = sumf;
}
void ggml_vec_dot_mxfp4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
assert(nrc == 1);
UNUSED(nrc);
UNUSED(bx);
UNUSED(by);
UNUSED(bs);
assert(n % QK_MXFP4 == 0);
static_assert(QK_MXFP4 == QK8_0, "QK_MXFP4 and QK8_0 must be the same");
const block_mxfp4 * GGML_RESTRICT x = vx;
const block_q8_0 * GGML_RESTRICT y = vy;
const int nb = n / QK_MXFP4;
int ib = 0;
float sumf = 0;
#if defined __ARM_NEON
const int8x16_t values = vld1q_s8(kvalues_mxfp4);
const uint8x16_t m4b = vdupq_n_u8(0x0f);
uint8x16x2_t q4bits;
int8x16x4_t q4b;
int8x16x4_t q8b;
int32x4_t prod_1;
int32x4_t prod_2;
for (; ib + 1 < nb; ib += 2) {
q4bits.val[0] = vld1q_u8(x[ib + 0].qs);
q4bits.val[1] = vld1q_u8(x[ib + 1].qs);
q8b.val[0] = vld1q_s8(y[ib + 0].qs);
q8b.val[1] = vld1q_s8(y[ib + 0].qs + 16);
q8b.val[2] = vld1q_s8(y[ib + 1].qs);
q8b.val[3] = vld1q_s8(y[ib + 1].qs + 16);
q4b.val[0] = ggml_vqtbl1q_s8(values, vandq_u8 (q4bits.val[0], m4b));
q4b.val[1] = ggml_vqtbl1q_s8(values, vshrq_n_u8(q4bits.val[0], 4));
q4b.val[2] = ggml_vqtbl1q_s8(values, vandq_u8 (q4bits.val[1], m4b));
q4b.val[3] = ggml_vqtbl1q_s8(values, vshrq_n_u8(q4bits.val[1], 4));
prod_1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q4b.val[0], q8b.val[0]), q4b.val[1], q8b.val[1]);
prod_2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q4b.val[2], q8b.val[2]), q4b.val[3], q8b.val[3]);
sumf +=
GGML_E8M0_TO_FP32_HALF(x[ib + 0].e) * GGML_CPU_FP16_TO_FP32(y[ib + 0].d) * vaddvq_s32(prod_1) +
GGML_E8M0_TO_FP32_HALF(x[ib + 1].e) * GGML_CPU_FP16_TO_FP32(y[ib + 1].d) * vaddvq_s32(prod_2);
}
#endif
for (; ib < nb; ++ib) {
const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_E8M0_TO_FP32_HALF(x[ib].e);
int sumi1 = 0;
int sumi2 = 0;
for (int j = 0; j < QK_MXFP4/2; ++j) {
sumi1 += y[ib].qs[j + 0] * kvalues_mxfp4[x[ib].qs[j] & 0xf];
sumi2 += y[ib].qs[j + QK_MXFP4/2] * kvalues_mxfp4[x[ib].qs[j] >> 4];
}
sumf += d * (sumi1 + sumi2);
}
*s = sumf;
}
void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
const int qk = QK8_0;
const int nb = n / qk;

104
ggml/src/ggml-cpu/arch/x86/quants.c
View File

@@ -66,6 +66,12 @@ static inline int hsum_i32_4(const __m128i a) {
}
#if defined(__AVX2__) || defined(__AVX512F__)
static inline __m256i mul_add_epi8(const __m256i x, const __m256i y) {
const __m256i ax = _mm256_sign_epi8(x, x);
const __m256i sy = _mm256_sign_epi8(y, x);
return _mm256_maddubs_epi16(ax, sy);
}
// spread 32 bits to 32 bytes { 0x00, 0xFF }
static inline __m256i bytes_from_bits_32(const uint8_t * x) {
uint32_t x32;
@@ -261,6 +267,11 @@ static inline __m256 quad_fp16_delta_float(const float x0, const float y0, const
return _mm256_set_m128(_mm_set1_ps(GGML_CPU_FP16_TO_FP32(x1) * GGML_CPU_FP16_TO_FP32(y1)),
_mm_set1_ps(GGML_CPU_FP16_TO_FP32(x0) * GGML_CPU_FP16_TO_FP32(y0)));
}
static inline __m256 quad_mx_delta_float(const int8_t x0, const float y0, const int8_t x1, const float y1) {
return _mm256_set_m128(_mm_set1_ps(GGML_E8M0_TO_FP32_HALF(x1) * GGML_CPU_FP16_TO_FP32(y1)),
_mm_set1_ps(GGML_E8M0_TO_FP32_HALF(x0) * GGML_CPU_FP16_TO_FP32(y0)));
}
#endif
#elif defined(__SSSE3__)
// horizontally add 4x4 floats
@@ -746,6 +757,91 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
#endif
}
void ggml_vec_dot_mxfp4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
assert(nrc == 1);
UNUSED(nrc);
UNUSED(bx);
UNUSED(by);
UNUSED(bs);
assert(n % QK_MXFP4 == 0);
static_assert(QK_MXFP4 == QK8_0, "QK_MXFP4 and QK8_0 must be the same");
const block_mxfp4 * GGML_RESTRICT x = vx;
const block_q8_0 * GGML_RESTRICT y = vy;
const int nb = n / QK_MXFP4;
int ib = 0;
float sumf = 0;
#if defined __AVX2__
const __m128i values128 = _mm_loadu_si128((const __m128i*)kvalues_mxfp4);
const __m128i m4b = _mm_set1_epi8(0x0f);
const __m256i mone = _mm256_set1_epi16(1);
__m256 accum1 = _mm256_setzero_ps();
__m256 accum2 = _mm256_setzero_ps();
for (; ib + 1 < nb; ib += 2) {
const __m128i q4bits_1 = _mm_loadu_si128((const __m128i*)x[ib + 0].qs);
const __m128i q4bits_2 = _mm_loadu_si128((const __m128i*)x[ib + 1].qs);
const __m256i q8b_1 = _mm256_loadu_si256((const __m256i *)y[ib + 0].qs);
const __m256i q8b_2 = _mm256_loadu_si256((const __m256i *)y[ib + 1].qs);
const __m256i q4b_1 = MM256_SET_M128I(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4b)),
_mm_shuffle_epi8(values128, _mm_and_si128(q4bits_1, m4b)));
const __m256i q4b_2 = MM256_SET_M128I(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_2, 4), m4b)),
_mm_shuffle_epi8(values128, _mm_and_si128(q4bits_2, m4b)));
const __m256i p16_1 = mul_add_epi8(q4b_1, q8b_1);
const __m256i p16_2 = mul_add_epi8(q4b_2, q8b_2);
const __m256i p_1 = _mm256_madd_epi16(p16_1, mone);
const __m256i p_2 = _mm256_madd_epi16(p16_2, mone);
accum1 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib + 0].d)*GGML_E8M0_TO_FP32_HALF(x[ib + 0].e)),
_mm256_cvtepi32_ps(p_1), accum1);
accum2 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib + 1].d)*GGML_E8M0_TO_FP32_HALF(x[ib + 1].e)),
_mm256_cvtepi32_ps(p_2), accum2);
}
sumf = hsum_float_8(_mm256_add_ps(accum1, accum2));
#elif defined __AVX__
const __m128i values128 = _mm_loadu_si128((const __m128i*)kvalues_mxfp4);
const __m128i m4b = _mm_set1_epi8(0x0f);
__m256 accum = _mm256_setzero_ps();
for (; ib + 1 < nb; ib += 2) {
const __m128i q4bits_1 = _mm_loadu_si128((const __m128i *)x[ib + 0].qs);
const __m128i q4bits_2 = _mm_loadu_si128((const __m128i *)x[ib + 1].qs);
const __m128i q8b_1_0 = _mm_loadu_si128((const __m128i *)y[ib + 0].qs);
const __m128i q8b_1_1 = _mm_loadu_si128((const __m128i *)y[ib + 0].qs + 1);
const __m128i q8b_2_0 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs);
const __m128i q8b_2_1 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs + 1);
const __m128i q4b_1_0 = _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_1, m4b));
const __m128i q4b_1_1 = _mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4b));
const __m128i q4b_2_0 = _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_2, m4b));
const __m128i q4b_2_1 = _mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_2, 4), m4b));
const __m256 p = mul_sum_i8_quad_float(q4b_1_0, q4b_1_1, q4b_2_0, q4b_2_1, q8b_1_0, q8b_1_1, q8b_2_0, q8b_2_1);
const __m256 deltas = quad_mx_delta_float(x[ib].e, y[ib].d, x[ib + 1].e, y[ib + 1].d);
accum = _mm256_add_ps(_mm256_mul_ps(deltas, p), accum);
}
sumf = hsum_float_8(accum);
#endif
for (; ib < nb; ++ib) {
const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_E8M0_TO_FP32_HALF(x[ib].e);
int sumi1 = 0;
int sumi2 = 0;
for (int j = 0; j < QK_MXFP4/2; ++j) {
sumi1 += y[ib].qs[j + 0] * kvalues_mxfp4[x[ib].qs[j] & 0xf];
sumi2 += y[ib].qs[j + QK_MXFP4/2] * kvalues_mxfp4[x[ib].qs[j] >> 4];
}
sumf += d * (sumi1 + sumi2);
}
*s = sumf;
}
void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
const int qk = QK8_0;
const int nb = n / qk;
@@ -3206,14 +3302,6 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
#endif
}
#if defined(__AVX2__)
static inline __m256i mul_add_epi8(const __m256i x, const __m256i y) {
const __m256i ax = _mm256_sign_epi8(x, x);
const __m256i sy = _mm256_sign_epi8(y, x);
return _mm256_maddubs_epi16(ax, sy);
}
#endif
void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
assert(n % QK_K == 0);
assert(nrc == 1);

15
ggml/src/ggml-cpu/ggml-cpu.c
View File

@@ -253,6 +253,12 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
.vec_dot_type = GGML_TYPE_Q8_1,
.nrows = 1,
},
[GGML_TYPE_MXFP4] = {
.from_float = quantize_row_mxfp4,
.vec_dot = ggml_vec_dot_mxfp4_q8_0,
.vec_dot_type = GGML_TYPE_Q8_0,
.nrows = 1,
},
[GGML_TYPE_Q2_K] = {
.from_float = quantize_row_q2_K,
.vec_dot = ggml_vec_dot_q2_K_q8_K,
@@ -1670,6 +1676,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
{
ggml_compute_forward_add(params, tensor);
} break;
case GGML_OP_ADD_ID:
{
ggml_compute_forward_add_id(params, tensor);
} break;
case GGML_OP_ADD1:
{
ggml_compute_forward_add1(params, tensor);
@@ -1924,7 +1934,7 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
} break;
case GGML_OP_FLASH_ATTN_EXT:
{
ggml_compute_forward_flash_attn_ext(params, tensor->src[0], tensor->src[1], tensor->src[2], tensor->src[3], tensor);
ggml_compute_forward_flash_attn_ext(params, tensor);
} break;
case GGML_OP_FLASH_ATTN_BACK:
{
@@ -2111,6 +2121,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
case GGML_OP_DUP:
case GGML_OP_CONT:
case GGML_OP_ADD:
case GGML_OP_ADD_ID:
case GGML_OP_ADD1:
case GGML_OP_ACC:
{
@@ -2172,6 +2183,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
case GGML_GLU_OP_REGLU:
case GGML_GLU_OP_GEGLU:
case GGML_GLU_OP_SWIGLU:
case GGML_GLU_OP_SWIGLU_OAI:
case GGML_GLU_OP_GEGLU_ERF:
case GGML_GLU_OP_GEGLU_QUICK:
{
@@ -2673,6 +2685,7 @@ struct ggml_cplan ggml_graph_plan(
}
} break;
case GGML_OP_ADD:
case GGML_OP_ADD_ID:
case GGML_OP_ADD1:
{
if (ggml_is_quantized(node->src[0]->type)) {

216
ggml/src/ggml-cpu/ops.cpp
View File

@@ -8,6 +8,7 @@
#include "vec.h"
#include <float.h>
#include <algorithm>
// ggml_compute_forward_dup
@@ -1283,6 +1284,7 @@ void ggml_compute_forward_add(
case GGML_TYPE_Q5_0:
case GGML_TYPE_Q5_1:
case GGML_TYPE_Q8_0:
case GGML_TYPE_MXFP4:
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q3_K:
case GGML_TYPE_Q4_K:
@@ -1309,6 +1311,77 @@ void ggml_compute_forward_add(
}
}
// ggml_compute_forward_add_id
static void ggml_compute_forward_add_id_f32(
const ggml_compute_params * params,
ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src1 = dst->src[1];
const ggml_tensor * src2 = dst->src[2];
GGML_ASSERT(dst->type == GGML_TYPE_F32);
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(src1->type == GGML_TYPE_F32);
GGML_ASSERT(src2->type == GGML_TYPE_I32);
GGML_ASSERT(src0->nb[0] == sizeof(float));
GGML_ASSERT(src1->nb[0] == sizeof(float));
const int ith = params->ith;
const int nth = params->nth;
const int nr = ggml_nrows(src0);
GGML_TENSOR_TERNARY_OP_LOCALS
GGML_ASSERT( nb0 == sizeof(float));
GGML_ASSERT(nb10 == sizeof(float));
// 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 ir = ir0; ir < ir1; ++ir) {
// src0 indices
const int i3 = ir/(ne2*ne1);
const int i2 = (ir - i3*ne2*ne1)/ne1;
const int i1 = (ir - i3*ne2*ne1 - i2*ne1);
// src1 indices
const int i11 = *(int32_t *) ((char *) src2->data + i1*nb20 + i2*nb21);
GGML_ASSERT(i11 >= 0 && i11 < ne11);
ggml_vec_add_f32(ne0,
(float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ),
(float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01),
(float *) ((char *) src1->data + i11*nb11));
}
}
void ggml_compute_forward_add_id(
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_add_id_f32(params, dst);
} break;
default:
{
GGML_ABORT("unsupported type for ggml_compute_forward_add_id: %s", ggml_type_name(src0->type));
}
}
}
// ggml_compute_forward_add1
static void ggml_compute_forward_add1_f32(
@@ -1660,6 +1733,7 @@ void ggml_compute_forward_add1(
case GGML_TYPE_Q5_1:
case GGML_TYPE_Q8_0:
case GGML_TYPE_Q8_1:
case GGML_TYPE_MXFP4:
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q3_K:
case GGML_TYPE_Q4_K:
@@ -1787,6 +1861,7 @@ void ggml_compute_forward_acc(
case GGML_TYPE_Q5_1:
case GGML_TYPE_Q8_0:
case GGML_TYPE_Q8_1:
case GGML_TYPE_MXFP4:
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q3_K:
case GGML_TYPE_Q4_K:
@@ -3614,6 +3689,93 @@ static void ggml_compute_forward_swiglu(
}
}
// ggml_compute_forward_swiglu_oai
static void ggml_compute_forward_swiglu_oai_f32(
const ggml_compute_params * params,
ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src1 = dst->src[1];
char * src0_d = (char *) src0->data;
char * src1_d = (char *) (src1 ? src1->data : src0->data);
const size_t src0_o = src0->nb[1];
const size_t src1_o = src1 ? src1->nb[1] : src0->nb[1];
GGML_ASSERT(ggml_is_contiguous_1(src0));
GGML_ASSERT(ggml_is_contiguous_1(dst));
if (src1) {
GGML_ASSERT(ggml_is_contiguous_1(src1));
GGML_ASSERT(src0->type == src1->type);
}
const int ith = params->ith;
const int nth = params->nth;
const int nc = src1 ? src0->ne[0] : src0->ne[0] / 2;
const int nr = ggml_nrows(src0);
GGML_ASSERT(dst->ne[0] == nc);
GGML_ASSERT(ggml_nrows(dst) == nr);
const int32_t swapped = ggml_get_op_params_i32(dst, 1);
const float alpha = ggml_get_op_params_f32(dst, 2);
const float limit = ggml_get_op_params_f32(dst, 3);
// 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++) {
float * src0_p = (float *) (src0_d + i1*src0_o);
float * src1_p = (float *) (src1_d + i1*src1_o);
float * dst_p = (float *) ((char *) dst->data + i1*(dst->nb[1]));
if (!src1) {
src0_p += swapped ? nc : 0;
src1_p += swapped ? 0 : nc;
}
for (int k = 0; k < nc; k++) {
const float x = std::min(src0_p[k], limit);
const float y = std::clamp(src1_p[k], -limit, limit);
const float out_glu = x / (1.f + expf(alpha * (-x)));
dst_p[k] = out_glu * (y + 1.f);
}
#ifndef NDEBUG
for (int k = 0; k < nc; k++) {
const float x = dst_p[k];
GGML_UNUSED(x);
assert(!isnan(x));
assert(!isinf(x));
}
#endif
}
}
static void ggml_compute_forward_swiglu_oai(
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_swiglu_oai_f32(params, dst);
} break;
default:
{
GGML_ABORT("fatal error");
}
}
}
// ggml_compute_forward_geglu_erf
static void ggml_compute_forward_geglu_erf_f32(
@@ -4599,6 +4761,7 @@ void ggml_compute_forward_out_prod(
case GGML_TYPE_Q5_0:
case GGML_TYPE_Q5_1:
case GGML_TYPE_Q8_0:
case GGML_TYPE_MXFP4:
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q3_K:
case GGML_TYPE_Q4_K:
@@ -4873,6 +5036,7 @@ void ggml_compute_forward_set(
case GGML_TYPE_Q5_1:
case GGML_TYPE_Q8_0:
case GGML_TYPE_Q8_1:
case GGML_TYPE_MXFP4:
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q3_K:
case GGML_TYPE_Q4_K:
@@ -5134,6 +5298,7 @@ void ggml_compute_forward_get_rows(
case GGML_TYPE_Q5_1:
case GGML_TYPE_Q8_0:
case GGML_TYPE_Q8_1:
case GGML_TYPE_MXFP4:
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q3_K:
case GGML_TYPE_Q4_K:
@@ -5523,6 +5688,7 @@ static void ggml_compute_forward_soft_max_f32(
const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src1 = dst->src[1];
const ggml_tensor * src2 = dst->src[2];
assert(ggml_is_contiguous(dst));
assert(ggml_are_same_shape(src0, dst));
@@ -5557,6 +5723,9 @@ static void ggml_compute_forward_soft_max_f32(
const bool use_f16 = (src1 && src1->type == GGML_TYPE_F16);
// sinks
const float * sk = src2 ? (float *)((char *) src2->data) : nullptr;
for (int64_t i03 = 0; i03 < ne03; i03++) {
for (int64_t i02 = 0; i02 < ne02; i02++) {
for (int64_t i01 = ith; i01 < ne01; i01 += nth) {
@@ -5599,9 +5768,18 @@ static void ggml_compute_forward_soft_max_f32(
float max = -INFINITY;
ggml_vec_max_f32(ne00, &max, wp);
// if we have sinks, make a correction as if they were included in the softmax
if (sk) {
max = MAX(max, sk[i02]);
}
ggml_float sum = ggml_vec_soft_max_f32(ne00, dp, wp, max);
assert(sum > 0.0);
if (sk) {
sum += (ggml_float) expf(sk[i02] - max);
}
sum = 1.0/sum;
ggml_vec_scale_f32(ne00, dp, sum);
@@ -5836,6 +6014,7 @@ void ggml_compute_forward_clamp(
case GGML_TYPE_Q5_1:
case GGML_TYPE_Q8_0:
case GGML_TYPE_Q8_1:
case GGML_TYPE_MXFP4:
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q3_K:
case GGML_TYPE_Q4_K:
@@ -7989,12 +8168,14 @@ void ggml_compute_forward_argsort(
static void ggml_compute_forward_flash_attn_ext_f16(
const ggml_compute_params * params,
const ggml_tensor * q,
const ggml_tensor * k,
const ggml_tensor * v,
const ggml_tensor * mask,
ggml_tensor * dst) {
const ggml_tensor * q = dst->src[0];
const ggml_tensor * k = dst->src[1];
const ggml_tensor * v = dst->src[2];
const ggml_tensor * mask = dst->src[3];
const ggml_tensor * sinks = dst->src[4];
GGML_TENSOR_LOCALS(int64_t, neq, q, ne)
GGML_TENSOR_LOCALS(size_t, nbq, q, nb)
GGML_TENSOR_LOCALS(int64_t, nek, k, ne)
@@ -8189,6 +8370,23 @@ static void ggml_compute_forward_flash_attn_ext_f16(
}
}
// sinks
if (sinks) {
const float s = ((float *)((char *) sinks->data))[h];
float ms = 1.0f;
float vs = 1.0f;
if (s > M) {
ms = expf(M - s);
ggml_vec_scale_f32(DV, VKQ32, ms);
} else {
vs = expf(s - M);
}
S = S*ms + vs;
}
// V /= S
const float S_inv = 1.0f/S;
ggml_vec_scale_f32(DV, VKQ32, S_inv);
@@ -8208,17 +8406,13 @@ static void ggml_compute_forward_flash_attn_ext_f16(
void ggml_compute_forward_flash_attn_ext(
const ggml_compute_params * params,
const ggml_tensor * q,
const ggml_tensor * k,
const ggml_tensor * v,
const ggml_tensor * mask,
ggml_tensor * dst) {
switch (dst->op_params[3]) {
case GGML_PREC_DEFAULT:
case GGML_PREC_F32:
{
// uses F32 accumulators
ggml_compute_forward_flash_attn_ext_f16(params, q, k, v, mask, dst);
ggml_compute_forward_flash_attn_ext_f16(params, dst);
} break;
default:
{
@@ -9080,6 +9274,10 @@ void ggml_compute_forward_glu(
{
ggml_compute_forward_swiglu(params, dst);
} break;
case GGML_GLU_OP_SWIGLU_OAI:
{
ggml_compute_forward_swiglu_oai(params, dst);
} break;
case GGML_GLU_OP_GEGLU_ERF:
{
ggml_compute_forward_geglu_erf(params, dst);

9
ggml/src/ggml-cpu/ops.h
View File

@@ -29,6 +29,7 @@ extern "C" {
void ggml_compute_forward_dup(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_add(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_add_id(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_add1(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_acc(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_sum(const struct ggml_compute_params * params, struct ggml_tensor * dst);
@@ -82,13 +83,7 @@ void ggml_compute_forward_arange(const struct ggml_compute_params * params, stru
void ggml_compute_forward_timestep_embedding(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_argsort(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_leaky_relu(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_flash_attn_ext(
const struct ggml_compute_params * params,
const struct ggml_tensor * q,
const struct ggml_tensor * k,
const struct ggml_tensor * v,
const struct ggml_tensor * mask,
struct ggml_tensor * dst);
void ggml_compute_forward_flash_attn_ext(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_flash_attn_back(
const struct ggml_compute_params * params,
const bool masked,

35
ggml/src/ggml-cpu/quants.c
View File

@@ -46,6 +46,10 @@ void quantize_row_q8_1_generic(const float * GGML_RESTRICT x, void * GGML_RESTRI
quantize_row_q8_1_ref(x, y, k);
}
void quantize_row_mxfp4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) {
quantize_row_mxfp4_ref(x, y, k);
}
//
// 2-6 bit quantization in super-blocks
//
@@ -181,6 +185,37 @@ void ggml_vec_dot_q4_1_q8_1_generic(int n, float * GGML_RESTRICT s, size_t bs, c
*s = sumf;
}
void ggml_vec_dot_mxfp4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
assert(nrc == 1);
UNUSED(nrc);
UNUSED(bx);
UNUSED(by);
UNUSED(bs);
assert(n % QK_MXFP4 == 0);
static_assert(QK_MXFP4 == QK8_0, "QK_MXFP4 and QK8_0 must be the same");
const block_mxfp4 * GGML_RESTRICT x = vx;
const block_q8_0 * GGML_RESTRICT y = vy;
const int nb = n / QK_MXFP4;
int ib = 0;
float sumf = 0;
for (; ib < nb; ++ib) {
const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_E8M0_TO_FP32_HALF(x[ib].e);
int sumi1 = 0;
int sumi2 = 0;
for (int j = 0; j < QK_MXFP4/2; ++j) {
sumi1 += y[ib].qs[j + 0] * kvalues_mxfp4[x[ib].qs[j] & 0xf];
sumi2 += y[ib].qs[j + QK_MXFP4/2] * kvalues_mxfp4[x[ib].qs[j] >> 4];
}
sumf += d * (sumi1 + sumi2);
}
*s = sumf;
}
void ggml_vec_dot_q5_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
const int qk = QK8_0;
const int nb = n / qk;

8
ggml/src/ggml-cpu/quants.h
View File

@@ -19,6 +19,8 @@ void quantize_row_q5_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, in
void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_mxfp4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q2_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q3_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q4_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
@@ -39,6 +41,8 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_mxfp4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
@@ -67,8 +71,12 @@ void ggml_vec_dot_q4_1_q8_1_generic(int n, float * GGML_RESTRICT s, size_t bs, c
void ggml_vec_dot_q5_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q5_1_q8_1_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q8_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_mxfp4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_tq1_0_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_tq2_0_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q2_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q3_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q4_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);

23
ggml/src/ggml-cpu/vec.h
View File

@@ -55,7 +55,22 @@ inline static void ggml_vec_cpy_i32(const int n, int32_t * y, const int32_t * x)
inline static void ggml_vec_set_f16(const int n, ggml_fp16_t * x, const ggml_fp16_t v) { for (int i = 0; i < n; ++i) x[i] = v; }
inline static void ggml_vec_set_bf16(const int n, ggml_bf16_t * x, const ggml_bf16_t v) { for (int i = 0; i < n; ++i) x[i] = v; }
inline static void ggml_vec_add_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i] = x[i] + y[i]; }
inline static void ggml_vec_add_f32 (const int n, float * z, const float * x, const float * y) {
int i = 0;
#if defined(__AVX2__)
for (; i + 7 < n; i += 8) {
__m256 vx = _mm256_loadu_ps(x + i);
__m256 vy = _mm256_loadu_ps(y + i);
__m256 vz = _mm256_add_ps(vx, vy);
_mm256_storeu_ps(z + i, vz);
}
#endif
for (; i < n; ++i) {
z[i] = x[i] + y[i];
}
}
inline static void ggml_vec_add_f16 (const int n, ggml_fp16_t * z, const ggml_fp16_t * x, const ggml_fp16_t * y) {
for (int i = 0; i < n; ++i) {
z[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(x[i]) + GGML_CPU_FP16_TO_FP32(y[i]));
@@ -992,9 +1007,9 @@ void ggml_vec_swiglu_f32(const int n, float * y, const float * x, const float *
inline static void ggml_vec_swiglu_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x, const ggml_fp16_t * g) {
for (int i = 0; i < n; ++i) {
float v = GGML_CPU_FP16_TO_FP32(x[i]);
float w = GGML_CPU_FP16_TO_FP32(g[i]);
y[i] = GGML_CPU_FP32_TO_FP16((v/(1.0f + expf(-v))) * w);
float xi = GGML_CPU_FP16_TO_FP32(x[i]);
float gi = GGML_CPU_FP16_TO_FP32(g[i]);
y[i] = GGML_CPU_FP32_TO_FP16((xi/(1.0f + expf(-xi))) * gi);
}
}