mirror of
https://github.com/ggml-org/llama.cpp.git
synced 2025-11-18 11:46:58 +00:00
llama : add gpt-oss (#15091)
* 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
@@ -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;
|
||||
|
||||
@@ -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);
|
||||
|
||||
Reference in New Issue
Block a user