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++
Merge catch-up: f-dsv41-win fd19f9d6b, f-dsv41-h100 c0f944e12, f-dsv41-cuk
Niko1221428950b96ce0
diff --cc src/dsv41/ds4_model.cuindex 154a8af48,df20a70da,b64d3934b,154a8af48..73a518bcb--- a/src/dsv41/ds4_model.cu+++ b/src/dsv41/ds4_model.cu@@@@@ -6336,18 -6325,6 -6267,49 -6336,18 +6370,49 @@@@@ struct Model::Implstd::fwrite(gb.data(), 4, gb.size(), dump_r);std::fflush(dump_r);}++ + if (mi == 0 && !l.mtp && (pred_f != nullptr || xin_f != nullptr)) {++ + std::lock_guard<std::mutex> lk(dump_mu);++ + if (pred_f != nullptr) {++ + float* dp = nullptr;++ + CK(cudaMalloc(&dp, (size_t) T * 256 * 4));++ + pred_proj_k<<<T, 256, 0, s>>>(c.x, (int) D, dp);++ + std::vector<float> hp((size_t) T * 256), hl((size_t) T * l.n_exp);++ + CK(cudaMemcpyAsync(hp.data(), dp, hp.size() * 4, cudaMemcpyDeviceToHost, s));++ + CK(cudaMemcpy2DAsync(hl.data(), (size_t) l.n_exp * 4, c.glog, 1024 * 4, (size_t) l.n_exp * 4, (size_t) T,++ + cudaMemcpyDeviceToHost, s));++ + CK(cudaStreamSynchronize(s));++ + CK(cudaFree(dp));++ + std::vector<uint16_t> o(hp.size() + hl.size());++ + for (size_t i = 0; i < hl.size(); ++i) o[i] = strata::kernels::f16_from_f32(hl[i]);++ + for (size_t i = 0; i < hp.size(); ++i) o[hl.size() + i] = strata::kernels::f16_from_f32(hp[i]);++ + const uint8_t hd[3] = {(uint8_t) l.il, (uint8_t) (T & 255), (uint8_t) (T >> 8)};++ + std::fwrite(hd, 1, 3, pred_f);++ + std::fwrite(o.data(), 2, o.size(), pred_f);++ + }++ + if (xin_f != nullptr && xin_got[(size_t) l.il] < xin_n) { // up to 32 evenly spaced tokens of the chunk++ + const int tk = std::min(std::min(T, 32), xin_n - xin_got[(size_t) l.il]);++ + const int stride = std::max(1, T / tk);++ + std::vector<float> hx((size_t) T * D);++ + CK(cudaMemcpyAsync(hx.data(), c.x, hx.size() * 4, cudaMemcpyDeviceToHost, s));++ + CK(cudaStreamSynchronize(s));++ + const uint8_t hd[3] = {(uint8_t) l.il, (uint8_t) (tk & 255), (uint8_t) (tk >> 8)};++ + std::fwrite(hd, 1, 3, xin_f);++ + for (int i = 0; i < tk; ++i) std::fwrite(hx.data() + (size_t) std::min(T - 1, i * stride) * D, 4, (size_t) D, xin_f);++ + xin_got[(size_t) l.il] += tk;++ + }++ + }+ if (mi == 0 && trace_f != nullptr && !l.mtp) { // DS4_TRACE: a chunk's picks, u8 (layer | 0x80), u16 T, T x k int16 ids+ std::vector<int> ri((size_t) T * kPicks);+ CK(cudaMemcpy(ri.data(), c.ridx, ri.size() * 4, cudaMemcpyDeviceToHost));+ std::lock_guard<std::mutex> lk(dump_mu);+ const uint8_t hd[3] = {(uint8_t) (l.il | 0x80), (uint8_t) (T & 255), (uint8_t) (T >> 8)};+ std::fwrite(hd, 1, 3, trace_f);+ for (int t = 0; t < T; ++t)+ for (int i = 0; i < l.k_exp; ++i) {+ const int16_t v = (int16_t) ri[(size_t) t * kPicks + i];+ std::fwrite(&v, 2, 1, trace_f);+ }+ }if (mi == 0) { // the profile (the CPU lane's experts: by the plan's order)uint32_t* pr = prof.data() + (size_t) l.il * g.n_exp;for (int e = 0; e < l.n_exp; ++e) {