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: test includes point at src/dsv41
Niko1221a5f54c97965b
diff --git a/tests/dsv41/ds4_cpu_bench.cpp b/tests/dsv41/ds4_cpu_bench.cppindex 605b8d01c..d30980d4f 100644--- a/tests/dsv41/ds4_cpu_bench.cpp+++ b/tests/dsv41/ds4_cpu_bench.cpp@@ -8,7 +8,7 @@// that node's physical cores (CPU i is on node i % 2 here, CPUs 0..43 the physical cores), each call computing// --per-call experts' gate and up (all columns) from one activation: the time of a matrix pass and its GB/s.#include "strata/dsv41/exl3_cpu.hpp"-#include "../src/ds4/ds4_pool.hpp"+#include "../../src/dsv41/ds4_pool.hpp"#include <sys/mman.h>#include <sys/syscall.h>diff --git a/tests/dsv41/ds4_gguf_test.cu b/tests/dsv41/ds4_gguf_test.cuindex 0bc0c4f51..74179c154 100644--- a/tests/dsv41/ds4_gguf_test.cu+++ b/tests/dsv41/ds4_gguf_test.cu@@ -6,11 +6,11 @@// shared expert), a window's grouped experts (bit for bit the token path's), the prompt path's expert sets and the CPU// lane's rows; and the KV caches' packed rows (ds4_kernels.cu) against the BF16 rows bit for bit. Needs a GPU, no// model. ds4_gguf_test [n_embd n_ff] | --kv (the caches alone)-#include "../src/ds4/ds4_gguf.cuh"-#include "../src/ds4/ds4_gguf_cpu.hpp"-#include "../src/ds4/ds4_iq.cuh"-#include "../src/ds4/ds4_kernels.cuh"-#include "../src/ds4/ds4_q6k.cuh"+#include "../../src/dsv41/ds4_gguf.cuh"+#include "../../src/dsv41/ds4_gguf_cpu.hpp"+#include "../../src/dsv41/ds4_iq.cuh"+#include "../../src/dsv41/ds4_kernels.cuh"+#include "../../src/dsv41/ds4_q6k.cuh"#include "ggml.h" // (the low-bit types' reference values: ggml's own dequantize_row_*)diff --git a/tests/dsv41/ds4_mx_bench.cpp b/tests/dsv41/ds4_mx_bench.cppindex 9d17e2fdd..a381898f7 100644--- a/tests/dsv41/ds4_mx_bench.cpp+++ b/tests/dsv41/ds4_mx_bench.cpp@@ -12,8 +12,8 @@// experts, so the weights come from RAM, not the cache. Prints the weights' GB/s for each kernel x tokens x threads// (T tokens a call: each row read once for up to 8 of them), the int8 tiers' agreement with the scalar tier, and the// cost of the activation quant and of madvise(WILLNEED) on cached pages.-#include "../src/ds4/ds4_gguf_cpu.hpp"-#include "../src/ds4/ds4_pool.hpp"+#include "../../src/dsv41/ds4_gguf_cpu.hpp"+#include "../../src/dsv41/ds4_pool.hpp"#if defined(STRATA_NATIVE_EXPERTS)#include "ggml-cpu.h"