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/cuda/dequant_s2.cu - P2.S2's first naive kernel.
//
// S2 is Q2_0's canonical form and 31.64 GiB of the artifact is Q2_0, so this is the single hottest decode in
// the engine. The canonical plane layout is `docs/pack-format.md`'s: codes packed 4 per byte LSB-first
// (element i at byte i/4, bits (i%4)*2 - Q2_0's own convention, kept for every form) and one FP32 scale per
// group (64 elements for S2). The decode is `(code + code_bias) * scale` with `code_bias = -1`, the
// subtraction inside the INTEGER domain, which is what makes it bit-exact (docs/pack-format.md §3.1).
//
// Naive on purpose: one thread per block, no vectors, no shared memory, no unrolling. Phase 3 changes this
// file and the parity test in src/kernels/dequant_s2_parity.cpp is what says whether a change is still right.
#define DPCT_PROFILING_ENABLED
#include <sycl/sycl.hpp>
#include <dpct/dpct.hpp>
#include "strata/kernels/dequant_s2.hpp"
#include <cstdio>
#include <cstdlib>
namespace strata::kernels {
namespace {
constexpr int QK = 64; // S2 group = Q2_0 block = 64 elements
constexpr int CODES_PER_BYTE = 4;
__dpct_inline__ void dequant_s2_kernel(const uint8_t *__restrict__ codes,
const float *__restrict__ scales,
float *__restrict__ out,
long long n_blocks) {
auto item_ct1 = sycl::ext::oneapi::this_work_item::get_nd_item<3>();
const long long b =
(long long)item_ct1.get_group(2) * item_ct1.get_local_range(2) +
item_ct1.get_local_id(2);
if (b >= n_blocks) return;
const float d = scales[b];
const uint8_t* c = codes + b * (QK / CODES_PER_BYTE);
float* y = out + b * QK;
for (int j = 0; j < QK; ++j) {
const int code = (c[j / CODES_PER_BYTE] >> ((j % CODES_PER_BYTE) * 2)) & 0x03;
y[j] = (float)(code - 1) * d; // the -1 is applied to the CODE, then ONE multiply
}
}
} // namespace
void dequant_s2(const uint8_t *codes, const float *scales, float *out,
int64_t n_blocks) try {
if (n_blocks <= 0) return;
const int threads = 256;
const long long blocks = (n_blocks + threads - 1) / threads;
{
auto exp_props = sycl::ext::oneapi::experimental::properties{
sycl::ext::oneapi::experimental::use_root_sync};
dpct::get_in_order_queue()
.parallel_for<dpct_kernel_name<class dequant_s2_kernel_bdb833>>(
sycl::nd_range<3>(sycl::range(1, 1, (unsigned)blocks) *
sycl::range(1, 1, threads),
sycl::range(1, 1, threads)),
exp_props, [=](sycl::nd_item<3> item_ct1) {
dequant_s2_kernel(codes, scales, out, n_blocks);
});
}
const dpct::err0 e =
DPCT_CHECK_ERROR(dpct::get_current_device().queues_wait_and_throw());
}
catch (sycl::exception const &exc) {
std::cerr << exc.what() << "Exception caught at file:" << __FILE__
<< ", line:" << __LINE__ << std::endl;
std::exit(1);
}
} // namespace strata::kernels