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 routing trio ported ahead of its turn - moe_sort counts, bounds scan, fill
Niko1221154c2990edb2
diff --git a/metal/include/metal_compat/ds4_metal_host.h b/metal/include/metal_compat/ds4_metal_host.hindex fba08b9af..d26707802 100644--- a/metal/include/metal_compat/ds4_metal_host.h+++ b/metal/include/metal_compat/ds4_metal_host.h@@ -32,6 +32,9 @@ struct FinishIn {uint8_t* pdst;long long pdst_stride;};+// ds4_kernels.cu:1938 moe_sort: counts, prefix, fill+bool moe_sort(const int* idx, int T, int k, const int* vs, int n_vs, int* bounds, int* row_tok, int* pos);+// ds4_kernels.cu:1908 hc_split (the sinkhorn normalisation)bool hc_split(const float* mixes, const float* scale, const float* base, int hc, int iters, float hc_eps, int T,float* pre, float* post, float* comb);diff --git a/metal/src/engine/ds4_finish.cpp b/metal/src/engine/ds4_finish.cppindex 37a5fe52e..e0a01b7a0 100644--- a/metal/src/engine/ds4_finish.cpp+++ b/metal/src/engine/ds4_finish.cpp@@ -240,4 +240,62 @@ bool hc_split(const float* mixes, const float* scale, const float* base, int hc,return true;}++// ds4_kernels.cu:1938 - moe_sort: count each virtual slot's picks, prefix-sum the counts, fill in token order.+// Three dispatches on one command buffer, so their order is the order of the calls (serial dispatch).+bool moe_sort(const int* idx, int T, int k, const int* vs, int n_vs, int* bounds, int* row_tok, int* pos) {+ if (!ensure()) return false;+ void* hi = strata::gpu::host_view((void*) idx);+ void* hv = strata::gpu::host_view((void*) vs);+ void* hb = strata::gpu::host_view((void*) bounds);+ void* hr = strata::gpu::host_view((void*) row_tok);+ void* hp = strata::gpu::host_view((void*) pos);+ if (!hi || !hv || !hb || !hr || !hp) { std::fprintf(stderr, "dsv41 metal: moe_sort has no host view\n"); return false; }+ const size_t page = 16384;+ auto wrap = [&](void* host, size_t bytes) {+ return g_cx->wrap((void*) (((uintptr_t) host) & ~(uintptr_t) (page - 1)), ((bytes + page - 1) / page) * page);+ };+ auto off = [&](void* host) { return (size_t) ((uintptr_t) host - (((uintptr_t) host) & ~(uintptr_t) (page - 1))); };+ const size_t t_bytes = (size_t) T * k * 4 + 65536;+ mtl::Buffer bi = wrap(hi, t_bytes);+ mtl::Buffer bv = wrap(hv, (size_t) n_vs * 4 + 65536);+ mtl::Buffer bb = wrap(hb, (size_t) (n_vs + 2) * 4 + 65536);+ mtl::Buffer br = wrap(hr, t_bytes);+ mtl::Buffer bp = wrap(hp, t_bytes);+ if (!bi.handle || !bv.handle || !bb.handle || !br.handle || !bp.handle) return false;+ mtl::Pipeline pc, pf, ps;+ try {+ pc = g_cx->pipeline(g_lib, "ds4_moe_sort_count");+ pf = g_cx->pipeline(g_lib, "ds4_moe_sort_fill");+ ps = g_cx->pipeline(g_lib, "ds4_bounds_scan");+ } catch (...) { return false; }+ // 1. counts+ mtl::Dispatch d1;+ d1.p = &pc;+ d1.args = {mtl::Arg::B(bi, off(hi)), mtl::Arg::V((unsigned) T), mtl::Arg::V((unsigned) k),+ mtl::Arg::B(bv, off(hv)), mtl::Arg::B(bb, off(hb))};+ d1.grid[0] = (uint32_t) n_vs;+ d1.grid[1] = d1.grid[2] = 1;+ d1.tg[0] = 256; d1.tg[1] = d1.tg[2] = 1;+ g_cx->run({d1});+ // 2. the exclusive prefix over the slots+ mtl::Dispatch d2;+ d2.p = &ps;+ d2.args = {mtl::Arg::B(bb, off(hb)), mtl::Arg::V((unsigned) n_vs)};+ d2.grid[0] = 1;+ d2.grid[1] = d2.grid[2] = 1;+ d2.tg[0] = 32; d2.tg[1] = d2.tg[2] = 1;+ g_cx->run({d2});+ // 3. the fill+ mtl::Dispatch d3;+ d3.p = &pf;+ d3.args = {mtl::Arg::B(bi, off(hi)), mtl::Arg::V((unsigned) T), mtl::Arg::V((unsigned) k),+ mtl::Arg::B(bv, off(hv)), mtl::Arg::B(bb, off(hb)), mtl::Arg::B(br, off(hr)), mtl::Arg::B(bp, off(hp))};+ d3.grid[0] = (uint32_t) n_vs;+ d3.grid[1] = d3.grid[2] = 1;+ d3.tg[0] = 256; d3.tg[1] = d3.tg[2] = 1;+ g_cx->run({d3});+ return true;+}+} // namespace strata::dsv41::metal_hostdiff --git a/metal/src/engine/msl/ds4_init.metal b/metal/src/engine/msl/ds4_init.metalindex 6e7094c90..c2550903d 100644--- a/metal/src/engine/msl/ds4_init.metal+++ b/metal/src/engine/msl/ds4_init.metal@@ -239,3 +239,73 @@ kernel void ds4_hc_split(device const float* mixes [[buffer(0)]], device const ffor (uint k = 0; k < 8u; ++k) { if (k >= hc) break; cb[j * hc + k] = c[j][k]; }}}++// ds4_kernels.cu:1938 moe_sort - count each virtual slot's picks, prefix-sum the counts, then fill in token order.+// kPicks = 8 (ds4_kernels.cuh:19): idx is [T][8], so q -> (t = q/8, i = q%8).+kernel void ds4_bounds_scan(device int* bounds [[buffer(0)]], constant uint& n_vs [[buffer(1)]],+ uint tid [[thread_index_in_threadgroup]]) {+ if (tid != 0u) return;+ bounds[0] = 0;+ for (uint s = 0; s < n_vs; ++s) bounds[s + 1] += bounds[s];+}++// simd_ballot returns a simd_vote, which does not cast to an integer in this toolchain; the same counts come from the+// prefix-sum and sum primitives, which give a lane its offset inside its simdgroup and the group its total.+static inline int ds4_warp_prefix(bool hit) { return simd_prefix_exclusive_sum(hit ? 1 : 0); }+static inline int ds4_warp_total(bool hit) { return simd_sum(hit ? 1 : 0); }++kernel void ds4_moe_sort_count(device const int* idx [[buffer(0)]], constant uint& T [[buffer(1)]],+ constant uint& kk [[buffer(2)]], device const int* vs [[buffer(3)]],+ device int* bounds [[buffer(4)]], uint slot [[threadgroup_position_in_grid]],+ uint tid [[thread_index_in_threadgroup]]) {+ threadgroup int warp_n[8];+ threadgroup int base_s;+ const uint wp = tid / 32u, lane = tid % 32u;+ if (tid == 0u) base_s = 0;+ threadgroup_barrier(mem_flags::mem_threadgroup);+ const uint n = T * kk;+ for (uint q = tid; q < n; q += 256u) {+ const uint t = q / kk, i = q - t * kk;+ const bool hit = vs[idx[t * 8u + i]] == int(slot);+ // one vote a warp: the running offset needs each warp's count, accumulated in slot order+ const int wtot = ds4_warp_total(hit);+ if ((tid % 32u) == 0u) warp_n[wp] = wtot;+ threadgroup_barrier(mem_flags::mem_threadgroup);+ int total = 0;+ for (uint w = 0; w < 8u; ++w) total += warp_n[w];+ threadgroup_barrier(mem_flags::mem_threadgroup);+ if (tid == 0u) base_s += total;+ threadgroup_barrier(mem_flags::mem_threadgroup);+ }+ if (tid == 0u) bounds[slot + 1] = base_s;+}++kernel void ds4_moe_sort_fill(device const int* idx [[buffer(0)]], constant uint& T [[buffer(1)]],+ constant uint& kk [[buffer(2)]], device const int* vs [[buffer(3)]],+ device int* bounds [[buffer(4)]], device int* row_tok [[buffer(5)]],+ device int* pos [[buffer(6)]], uint slot [[threadgroup_position_in_grid]],+ uint tid [[thread_index_in_threadgroup]]) {+ threadgroup int warp_n[8];+ threadgroup int base_s;+ const uint wp = tid / 32u, lane = tid % 32u;+ if (tid == 0u) base_s = bounds[slot];+ threadgroup_barrier(mem_flags::mem_threadgroup);+ const uint n = T * kk;+ for (uint q = tid; q < n; q += 256u) {+ const uint t = q / kk, i = q - t * kk;+ const bool hit = vs[idx[t * 8u + i]] == int(slot);+ const int wpre = ds4_warp_prefix(hit), wtot = ds4_warp_total(hit);+ if ((tid % 32u) == 0u) warp_n[wp] = wtot;+ threadgroup_barrier(mem_flags::mem_threadgroup);+ int before = 0, total = 0;+ for (uint w = 0; w < 8u; ++w) { if (w < wp) before += warp_n[w]; total += warp_n[w]; }+ if (hit) {+ const int r = base_s + before + wpre;+ row_tok[r] = int(t);+ pos[t * 8u + i] = r;+ }+ threadgroup_barrier(mem_flags::mem_threadgroup);+ if (tid == 0u) base_s += total;+ threadgroup_barrier(mem_flags::mem_threadgroup);+ }+}diff --git a/src/dsv41/ds4_kernels.cu b/src/dsv41/ds4_kernels.cuindex e387ae1f4..21343c8fd 100644--- a/src/dsv41/ds4_kernels.cu+++ b/src/dsv41/ds4_kernels.cu@@ -1937,6 +1937,9 @@ void engram_dequant_tok(const uint8_t* src, size_t per, int cols, int T, int hd,}void moe_sort(const int* idx, int T, int k, const int* vs, int n_vs, int* bounds, int* row_tok, int* pos,cudaStream_t s) {+#if defined(STRATA_METAL)+ if (strata::dsv41::metal_host::moe_sort(idx, T, k, vs, n_vs, bounds, row_tok, pos)) return;+#endifstrata_ds4_launch(moe_sort_kernel<false>, n_vs, kThreads, 0, s, idx, T, k, vs, bounds, row_tok, pos);strata_ds4_launch(bounds_scan_kernel, 1, 32, 0, s, bounds, n_vs);strata_ds4_launch(moe_sort_kernel<true>, n_vs, kThreads, 0, s, idx, T, k, vs, bounds, row_tok, pos);