Qwen3.8-Flash-Next on any consumer hardware: one-click install for Windows / Linux. Strata inference engine, OpenAI/Anthropic API on localhost, optional image input.
C++
// src/kernels/cpu/tq_avx2.cpp - the trellis expert rows (TQ2_T / TQK6 / TQK7) on AVX2 + FMA (tq_cpu.hpp).
// Compiled for AVX2 only and reached behind cpu_avx2_ok(), so nothing here can fault on an older CPU.
//
// Per 128-weight block: the 32 states in scalar code (tq_states.hpp: one 32-bit read, a shift and a mask each, used
// as table indices straight away), four 8-byte state-table rows per YMM register as scalar loads, then per token
// VPMADDWD against the unpacked int16 activations and a VPHADDD tree that leaves the four exact sub-block sums
// S[k] in one XMM, so the float part is the scalar kernel's, lane for lane. The weights are decoded once per block
// for all the tokens of the group.
//
// Measured on Zen 5 (cycles per block, one token, AVX2): ~48 with this scheme; computing the states with byte
// shuffles instead (VPSHUFB / VPSRLVD per 8 states) and moving them to GPRs or through memory was slower (60-80),
// and ggml-cpu's AVX2 vec_dot for these types (agentionai llama.cpp, the same state table) takes about 2.3x as long.
#include "strata/kernels/cpu/tq_cpu.hpp"
#include "tq_states.hpp"
#include <immintrin.h>
namespace strata::kernels::cpu::tq_detail {
namespace {
// four state-table rows (8 bytes each) into one YMM: a load and three broadcast loads blended in. Broadcast loads
// and VPBLENDD avoid the shuffle port that VPINSRQ / VINSERTI128 need (the two measure the same on Zen 5).
inline __m256i rows4(const long long* tab, uint32_t s0, uint32_t s1, uint32_t s2, uint32_t s3) {
__m256i v = _mm256_castsi128_si256(_mm_loadl_epi64((const __m128i*) (tab + s0)));
v = _mm256_blend_epi32(v, _mm256_set1_epi64x(tab[s1]), 0x0C);
v = _mm256_blend_epi32(v, _mm256_set1_epi64x(tab[s2]), 0x30);
return _mm256_blend_epi32(v, _mm256_set1_epi64x(tab[s3]), 0xC0);
}
template <int KB>
void rows_t(const int16_t* table, const uint8_t* w, size_t row_bytes, int n, const TqAct* const* a, int nt, float* const* out, int r0,
int r1) {
const int nb = n / 128;
constexpr size_t bsz = 2 + 4 * KB;
const long long* tab = (const long long*) table;
for (int r = r0; r < r1; ++r) {
const uint8_t* x = w + (size_t) r * row_bytes;
__m128 acc[kTqMaxTokens];
for (int t = 0; t < nt; ++t) acc[t] = _mm_setzero_ps();
for (int i = 0; i < nb; ++i) {
const uint8_t* blk = x + (size_t) i * bsz;
const uint8_t* qs = blk + 2;
const uint64_t wr = tq_wrap<KB>(qs);
__m256i wv[8]; // wv[m] = weights 16m .. 16m+15 (states 4m .. 4m+3)
#if defined(__GNUC__)
#pragma GCC unroll 8
#endif
for (int m = 0; m < 8; ++m)
wv[m] = rows4(tab, tq_state<KB>(qs, wr, 4 * m), tq_state<KB>(qs, wr, 4 * m + 1),
tq_state<KB>(qs, wr, 4 * m + 2), tq_state<KB>(qs, wr, 4 * m + 3));
const __m128 dx = _mm_set1_ps(tq_half_to_float(tq_ld16(blk)));
for (int t = 0; t < nt; ++t) {
const __m256i* q = (const __m256i*) (a[t]->q + 128 * i);
const __m256i p0 = _mm256_add_epi32(_mm256_madd_epi16(wv[0], _mm256_load_si256(q + 0)),
_mm256_madd_epi16(wv[1], _mm256_load_si256(q + 1)));
const __m256i p1 = _mm256_add_epi32(_mm256_madd_epi16(wv[2], _mm256_load_si256(q + 2)),
_mm256_madd_epi16(wv[3], _mm256_load_si256(q + 3)));
const __m256i p2 = _mm256_add_epi32(_mm256_madd_epi16(wv[4], _mm256_load_si256(q + 4)),
_mm256_madd_epi16(wv[5], _mm256_load_si256(q + 5)));
const __m256i p3 = _mm256_add_epi32(_mm256_madd_epi16(wv[6], _mm256_load_si256(q + 6)),
_mm256_madd_epi16(wv[7], _mm256_load_si256(q + 7)));
// per 128-bit lane [p0, p1, p2, p3] partial sums, then the two lanes added: S[0..3], exact
const __m256i h = _mm256_hadd_epi32(_mm256_hadd_epi32(p0, p1), _mm256_hadd_epi32(p2, p3));
const __m128i s = _mm_add_epi32(_mm256_castsi256_si128(h), _mm256_extracti128_si256(h, 1));
const __m128 sc = _mm_mul_ps(dx, _mm_loadu_ps(a[t]->d + 4 * i));
acc[t] = _mm_fmadd_ps(_mm_cvtepi32_ps(s), sc, acc[t]);
}
}
for (int t = 0; t < nt; ++t) {
alignas(16) float v[4];
_mm_store_ps(v, acc[t]);
out[t][r] = ((v[0] + v[1]) + (v[2] + v[3])) * (1.0f / 8192.0f);
}
}
}
// TQP (147-149): 64 pair-steps per 128-value block. A step's two weights are the FIRST pair of its state
// table row, so 8 states make one YMM of 16 int16 weights, and a 32-value group (16 steps) is two YMM of
// weights against two YMM of activations: VPMADDWD, one exact int32 sum per group, then the four group sums
// go into the same per-token __m128 the other kernels use, so the float part is the scalar kernel's.
inline __m256i pairs8(const int16_t* tab, uint32_t s0, uint32_t s1, uint32_t s2, uint32_t s3, uint32_t s4,
uint32_t s5, uint32_t s6, uint32_t s7) {
const __m128i a0 = _mm_unpacklo_epi32(_mm_loadl_epi64((const __m128i*) (tab + 4 * s0)),
_mm_loadl_epi64((const __m128i*) (tab + 4 * s1)));
const __m128i a1 = _mm_unpacklo_epi32(_mm_loadl_epi64((const __m128i*) (tab + 4 * s2)),
_mm_loadl_epi64((const __m128i*) (tab + 4 * s3)));
const __m128i a2 = _mm_unpacklo_epi32(_mm_loadl_epi64((const __m128i*) (tab + 4 * s4)),
_mm_loadl_epi64((const __m128i*) (tab + 4 * s5)));
const __m128i a3 = _mm_unpacklo_epi32(_mm_loadl_epi64((const __m128i*) (tab + 4 * s6)),
_mm_loadl_epi64((const __m128i*) (tab + 4 * s7)));
return _mm256_set_m128i(_mm_unpacklo_epi64(a2, a3), _mm_unpacklo_epi64(a0, a1));
}
template <int KB>
void rows_tqp_t(const uint8_t* w, size_t row_bytes, int n, const TqAct* const* a, int nt, float* const* out, int r0,
int r1) {
const int nb = n / 128;
constexpr size_t bsz = 2 + 8 * KB;
const int16_t* tab = tq_state_table();
for (int r = r0; r < r1; ++r) {
const uint8_t* x = w + (size_t) r * row_bytes;
__m128 acc[kTqMaxTokens];
for (int t = 0; t < nt; ++t) acc[t] = _mm_setzero_ps();
for (int i = 0; i < nb; ++i) {
const uint8_t* blk = x + (size_t) i * bsz;
uint32_t st[64];
tq_states_tqp<KB>(blk + 2, st);
__m256i wv[8]; // wv[2g], wv[2g+1]: the weights of the 32-value group g
#pragma GCC unroll 4
for (int g = 0; g < 4; ++g) {
const uint32_t* s = st + 16 * g;
wv[2 * g] = pairs8(tab, s[0], s[1], s[2], s[3], s[4], s[5], s[6], s[7]);
wv[2 * g + 1] = pairs8(tab, s[8], s[9], s[10], s[11], s[12], s[13], s[14], s[15]);
}
const __m128 dx = _mm_set1_ps(tq_half_to_float(tq_ld16(blk)));
for (int t = 0; t < nt; ++t) {
const __m256i* q = (const __m256i*) (a[t]->q + 128 * i);
int32_t sum[4];
#pragma GCC unroll 4
for (int g = 0; g < 4; ++g) {
const __m256i p = _mm256_add_epi32(_mm256_madd_epi16(wv[2 * g], _mm256_load_si256(q + 2 * g)),
_mm256_madd_epi16(wv[2 * g + 1], _mm256_load_si256(q + 2 * g + 1)));
__m128i s = _mm_add_epi32(_mm256_castsi256_si128(p), _mm256_extracti128_si256(p, 1));
s = _mm_hadd_epi32(s, s);
s = _mm_hadd_epi32(s, s);
sum[g] = _mm_cvtsi128_si32(s);
}
const __m128 sc = _mm_mul_ps(dx, _mm_loadu_ps(a[t]->d + 4 * i));
const __m128i s4 = _mm_set_epi32(sum[3], sum[2], sum[1], sum[0]);
acc[t] = _mm_fmadd_ps(_mm_cvtepi32_ps(s4), sc, acc[t]);
}
}
for (int t = 0; t < nt; ++t) {
alignas(16) float v[4];
_mm_store_ps(v, acc[t]);
out[t][r] = ((v[0] + v[1]) + (v[2] + v[3])) * (1.0f / 8192.0f);
}
}
}
} // namespace
void tq_rows_avx2_tqp(int kb, const uint8_t* w, size_t row_bytes, int n, const TqAct* const* a, int nt,
float* const* out, int r0, int r1) {
if (kb == 5) rows_tqp_t<5>(w, row_bytes, n, a, nt, out, r0, r1);
else if (kb == 6) rows_tqp_t<6>(w, row_bytes, n, a, nt, out, r0, r1);
else rows_tqp_t<7>(w, row_bytes, n, a, nt, out, r0, r1);
}
void tq_rows_avx2(int kb, const int16_t* table, const uint8_t* w, size_t row_bytes, int n, const TqAct* const* a, int nt,
float* const* out, int r0, int r1) {
if (kb == 8) rows_t<8>(table, w, row_bytes, n, a, nt, out, r0, r1);
else if (kb == 6) rows_t<6>(table, w, row_bytes, n, a, nt, out, r0, r1);
else if (kb == 5) rows_t<5>(table, w, row_bytes, n, a, nt, out, r0, r1); // experimental TQK5X
else if (kb == 4) rows_t<4>(table, w, row_bytes, n, a, nt, out, r0, r1); // experimental TQK4X
else rows_t<7>(table, w, row_bytes, n, a, nt, out, r0, r1);
}
} // namespace strata::kernels::cpu::tq_detail