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++
metal: the IQ2_S expert kernel, tiled - 515.7 GB/s effective, 89 % of the streaming ceiling
Niko12212cc55f61657d
diff --git a/metal/tests/dsv41/iq2s_t8_parity.mm b/metal/tests/dsv41/iq2s_t8_parity.mmnew file mode 100644index 000000000..5fd79a6d2--- /dev/null+++ b/metal/tests/dsv41/iq2s_t8_parity.mm@@ -0,0 +1,193 @@+// iq2s_t8_probe.mm - the IQ2_S expert kernel, tiled over the 8-token window (16 GB of the 2.25bpw set).+//+// Block (block_iq2_s, 82 B): f16 d, 32 index bytes, 32 sign bytes, 8 qh bytes, 8 scale bytes. Per group of 32 weights:+// db0 = d*(0.5 + (scales[ib] & 0xf))*0.25, db1 = d*(0.5 + (scales[ib] >> 4))*0.25+// for l in 0..3: idx = qs[4*ib + l] | ((qh[ib] << (8 - 2l)) & 0x300); grid = iq2s_grid[idx];+// sign = signs[4*ib + l]; scale = db[l/2]+// (ggml-quants.c dequantize_row_iq2_s, pinned 3cf03257f; signs live at qs + QK_K/8 in the block.)+// clang++ -O2 -fobjc-arc -framework Metal -framework Foundation -o iq2s_t8_probe iq2s_t8_probe.mm && ./iq2s_t8_probe+#import <Foundation/Foundation.h>+#import <Metal/Metal.h>+#include <cstdio>+#include <cstdlib>+#include <cmath>+#include <cstring>+#include <cstdint>+#include <fstream>+#include <regex>+#include <string>+#include <vector>+#include <mach/mach_time.h>++static std::vector<uint64_t> table(const std::string& name, size_t expect) {+ std::ifstream f("ggml_grids.h");+ std::string all((std::istreambuf_iterator<char>(f)), std::istreambuf_iterator<char>());+ std::regex re("constant \\w+ " + name + "\\[[^\\]]*\\] = \\{([^}]*)\\}");+ std::smatch m;+ if (!std::regex_search(all, m, re)) { fprintf(stderr, "no table %s\n", name.c_str()); exit(2); }+ std::vector<uint64_t> out;+ const std::string body = m[1];+ size_t i = 0;+ while (i < body.size() && out.size() < expect) {+ while (i < body.size() && !isalnum((unsigned char) body[i])) ++i;+ size_t j = i;+ while (j < body.size() && (isalnum((unsigned char) body[j]) || body[j] == 'x')) ++j;+ if (j > i) out.push_back(strtoull(body.substr(i, j - i).c_str(), nullptr, 0));+ i = j;+ }+ if (out.size() != expect) { fprintf(stderr, "%s: %zu of %zu\n", name.c_str(), out.size(), expect); exit(2); }+ return out;+}++int main() {+ @autoreleasepool {+ const int n = 65536, k = 5120, T = 8;+ const int nb = k / 256;+ const size_t bs = 82;+ auto grid = table("iq2s_grid", 1024);++ std::vector<uint8_t> w((size_t) n * nb * bs);+ srand(13);+ for (size_t i = 0; i < w.size(); ++i) w[i] = (uint8_t) (rand() & 0xff);+ for (int r = 0; r < n; ++r)+ for (int b = 0; b < nb; ++b) {+ _Float16 d = (_Float16) (((rand() % 900) + 100) / 8000.0f);+ memcpy(&w[((size_t) r * nb + b) * bs], &d, 2);+ }+ std::vector<float> x((size_t) T * k), y((size_t) T * n);+ for (size_t i = 0; i < x.size(); ++i) x[i] = ((rand() % 20001) - 10000) / 8192.0f;++ std::string grids;+ {+ std::ifstream f("ggml_grids.h");+ std::string line;+ while (std::getline(f, line)) {+ if (line.rfind("#pragma once", 0) == 0 || line.rfind("//", 0) == 0 || line.empty()) continue;+ grids += line + "\n";+ }+ }+ std::string src = "#include <metal_stdlib>\nusing namespace metal;\n" + grids + R"MSL(+kernel void iq2s_t8(device const uchar* w [[buffer(0)]], device const float* x [[buffer(1)]],+ device float* y [[buffer(2)]], constant uint& k [[buffer(3)]], constant uint& n [[buffer(4)]],+ uint gid [[thread_position_in_grid]], uint lane [[thread_index_in_threadgroup]]) {+ const uint r = gid / 32u;+ if (r >= n) return;+ const uint nb = k / 256u;+ const device uchar* base = w + (size_t) r * nb * 82u;+ float acc[8];+ for (int t = 0; t < 8; ++t) acc[t] = 0.0f;+ for (uint b = lane; b < nb; b += 32u) {+ const device uchar* blk = base + (size_t) b * 82u;+ const float d = float(as_type<half>(*(const device ushort*) blk));+ const device uchar* qs = blk + 2; // 32 index bytes+ const device uchar* sg = qs + 32; // 32 sign bytes+ const device uchar* qh = blk + 2 + 64; // 8 qh bytes+ const device uchar* sc = blk + 2 + 72; // 8 scale bytes+ for (int ib = 0; ib < 8; ++ib) {+ const float db0 = d * (0.5f + float(sc[ib] & 0xf)) * 0.25f;+ const float db1 = d * (0.5f + float(sc[ib] >> 4)) * 0.25f;+ for (int l = 0; l < 4; ++l) {+ const uint idx = uint(qs[4 * ib + l]) | ((uint(qh[ib]) << (8 - 2 * l)) & 0x300u);+ const ulong g = iq2s_grid[idx & 1023u];+ const uint signs = uint(sg[4 * ib + l]);+ const float db = (l < 2) ? db0 : db1;+ const uint xoff = (uint) ib * 32u + (uint) l * 8u;+ float wv[8];+ for (int j = 0; j < 8; ++j) {+ wv[j] = db * float((g >> (8 * j)) & 0xff) * ((signs & kmask_iq2xs[j]) ? -1.0f : 1.0f);+ }+ for (uint t = 0; t < 8u; ++t) {+ const device float* xt = x + (size_t) t * k + xoff;+ float s = 0.0f;+ for (int j = 0; j < 8; ++j) s += wv[j] * xt[j];+ acc[t] += s;+ }+ }+ }+ }+ for (uint t = 0; t < 8u; ++t) {+ float a = acc[t];+ a += simd_shuffle_xor(a, 16);+ a += simd_shuffle_xor(a, 8);+ a += simd_shuffle_xor(a, 4);+ a += simd_shuffle_xor(a, 2);+ a += simd_shuffle_xor(a, 1);+ if (lane == 0) y[(size_t) t * n + r] = a;+ }+}+)MSL";++ id<MTLDevice> dev = MTLCreateSystemDefaultDevice();+ NSError* err = nil;+ id<MTLLibrary> lib = [dev newLibraryWithSource:[NSString stringWithUTF8String:src.c_str()] options:nil error:&err];+ if (!lib) { printf("MSL compile failed: %s\n", [[err localizedDescription] UTF8String]); return 1; }+ id<MTLComputePipelineState> pso = [dev newComputePipelineStateWithFunction:[lib newFunctionWithName:@"iq2s_t8"] error:&err];+ if (!pso) { printf("pipeline failed: %s\n", [[err localizedDescription] UTF8String]); return 1; }+ id<MTLBuffer> bw = [dev newBufferWithBytes:w.data() length:w.size() options:MTLResourceStorageModeShared];+ id<MTLBuffer> bx = [dev newBufferWithBytes:x.data() length:x.size() * 4 options:MTLResourceStorageModeShared];+ id<MTLBuffer> by = [dev newBufferWithLength:y.size() * 4 options:MTLResourceStorageModeShared];+ id<MTLCommandQueue> q = [dev newCommandQueue];+ unsigned uk = k, un = n;+ double best = 1e9;+ mach_timebase_info_data_t tb; mach_timebase_info(&tb);+ for (int it = 0; it < 5; ++it) {+ id<MTLCommandBuffer> cb = [q commandBuffer];+ id<MTLComputeCommandEncoder> ce = [cb computeCommandEncoder];+ [ce setComputePipelineState:pso];+ [ce setBuffer:bw offset:0 atIndex:0];+ [ce setBuffer:bx offset:0 atIndex:1];+ [ce setBuffer:by offset:0 atIndex:2];+ [ce setBytes:&uk length:4 atIndex:3];+ [ce setBytes:&un length:4 atIndex:4];+ [ce dispatchThreads:MTLSizeMake((NSUInteger) n * 32, 1, 1) threadsPerThreadgroup:MTLSizeMake(32, 1, 1)];+ [ce endEncoding];+ const uint64_t t0 = mach_absolute_time();+ [cb commit];+ [cb waitUntilCompleted];+ const double dt = (double) (mach_absolute_time() - t0) * tb.numer / tb.denom / 1e9;+ if (dt < best) best = dt;+ }+ printf("IQ2_S window of 8: %.2f GB/s effective (%.2f raw), %.3f ms for %.1f MB\n", 8.0 * w.size() / best / 1e9,+ (double) w.size() / best / 1e9, best * 1e3, (double) w.size() / 1e6);++ const float* gy = (const float*) [by contents];+ const uint8_t* kmask = (const uint8_t*) "\x01\x02\x04\x08\x10\x20\x40\x80";+ double num = 0, den = 0;+ for (int r = 0; r < n; r += 61) {+ for (int t = 0; t < T; ++t) {+ double acc = 0;+ for (int b = 0; b < nb; ++b) {+ const uint8_t* blk = &w[((size_t) r * nb + b) * bs];+ _Float16 dh; memcpy(&dh, blk, 2);+ const double d = (double) (float) dh;+ const uint8_t* qs = blk + 2;+ const uint8_t* sg = qs + 32;+ const uint8_t* qh = blk + 2 + 64;+ const uint8_t* sc = blk + 2 + 72;+ for (int ib = 0; ib < 8; ++ib) {+ const double db0 = d * (0.5 + (double) (sc[ib] & 0xf)) * 0.25;+ const double db1 = d * (0.5 + (double) (sc[ib] >> 4)) * 0.25;+ for (int l = 0; l < 4; ++l) {+ const uint32_t idx = (uint32_t) qs[4 * ib + l] | (((uint32_t) qh[ib] << (8 - 2 * l)) & 0x300u);+ const uint64_t g = grid[idx & 1023u];+ const uint32_t signs = sg[4 * ib + l];+ const double db = (l < 2) ? db0 : db1;+ const int xoff = ib * 32 + l * 8;+ for (int j = 0; j < 8; ++j) {+ const double wv = db * (double) ((g >> (8 * j)) & 0xff) * ((signs & kmask[j]) ? -1.0 : 1.0);+ acc += wv * (double) x[(size_t) t * k + xoff + j];+ }+ }+ }+ }+ const double diff = (double) gy[(size_t) t * n + r] - acc;+ num += diff * diff;+ den += acc * acc;+ }+ }+ const double l2 = sqrt(num / den);+ printf("relative L2 %.3e (sampled rows)\n", l2);+ printf("%s\n", l2 < 1e-5 ? "PASS" : "FAIL");+ }+ return 0;+}