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++
dsv41: --pack-info prints the DSpark and Engram geometry a pack needs at run time
Niko1221cf4bef2f787b
diff --git a/src/dsv41/ds4_packinfo.cpp b/src/dsv41/ds4_packinfo.cppindex a386aebd0..ee4783970 100644--- a/src/dsv41/ds4_packinfo.cpp+++ b/src/dsv41/ds4_packinfo.cpp@@ -4,6 +4,7 @@// pack's whole description (tools/dsv41/pack_gguf.py). The same report is the first line of every benchmark log.#include "ds4_packinfo.hpp"+#include <algorithm>#include <cstdio>#include <cstdlib>#include <cstring>@@ -211,6 +212,24 @@ int pack_info(const std::string& dir, std::FILE* out) {std::fprintf(out, " dense types:");for (const auto& d : dense_types) std::fprintf(out, " %s", d.c_str());std::fprintf(out, "\n");+ // the geometry the engine needs at run time but that is easy to lose in a pack built by another tool: DSpark's block+ // size / noise token / markov rank (0 there means it drafts one token a run, i.e. it is off) and the Engram's layers+ {+ auto kvn = [&](const char* k) {+ auto it = kv.find(k);+ return it == kv.end() ? 0 : std::atoll(it->second.c_str());+ };+ const int n_mtp = (int) std::count_if(t.begin(), t.end(), [](const auto& p) { return p.first.rfind("mtp.", 0) == 0; });+ const long long nl_eg = num("engram_layer_ids", -1);+ std::fprintf(out, " dspark: block_size %lld, noise_token_id %lld, markov_rank %lld, routed experts %lld of %lld%s\n",+ num("dspark_block_size", 0), num("dspark_noise_token_id", 0), num("dspark_markov_rank", 0),+ num("dspark_n_routed_experts", 0), num("dspark_num_experts_per_tok", 0),+ num("dspark_block_size", 0) == 0 ? " <- DSpark cannot draft (block size 0)" : "");+ std::fprintf(out, " mtp tensors: %d%s\n", n_mtp, n_mtp == 0 ? " <- no DSpark drafter in the pack" : "");+ auto it = kv.find("engram_layer_ids");+ std::fprintf(out, " engram layers: %s%s\n", it == kv.end() ? "(none)" : it->second.c_str(),+ it == kv.end() ? " <- the Engram tables are not used" : "");+ }if (no_kernel.empty()) {std::fprintf(out, " no GPU kernel: none (every tensor is a device type)\n");} else {