Skip to content

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++

Dense dequant: the wide kernels by default on sm_80 too, bitwise (#3779, #3712)

linxuhao269c6d3e4ad9
diff --git a/src/kernels/cuda/dequant_bf16.cu b/src/kernels/cuda/dequant_bf16.cu
index f0aaa9250..00e1dd51e 100644
--- a/src/kernels/cuda/dequant_bf16.cu
+++ b/src/kernels/cuda/dequant_bf16.cu
@@ -285,14 +285,16 @@ bool dequant_wide_on() {
static const int env = [] { const char* v = std::getenv("STRATA_DEQUANT_WIDE"); return v != nullptr && v[0] != '\0' ? (std::atoi(v) != 0 ? 1 : 0) : -1; }();
if (g_dequant_wide >= 0) return g_dequant_wide != 0;
if (env >= 0) return env != 0;
- static int per_dev[64]; // 0 unknown, 1 no, 2 yes: sm_61 (measured on a Tesla P40) and sm_86 (RTX 3060 / A4000, 0.1.45: short-prompt TTFT -12..-17%, bitwise)
+ // 0 unknown, 1 no, 2 yes: sm_61 (measured on a Tesla P40), sm_86 (RTX 3060 / A4000, 0.1.45: short-prompt TTFT -12..-17%,
+ // bitwise) and sm_80 (CMP 170HX: a 3,259-token prompt 857 -> 792 ms of GPU time, prompts +4-7%, bitwise)
+ static int per_dev[64];
int dev = 0;
if (cudaGetDevice(&dev) != cudaSuccess || dev < 0 || dev >= 64) { (void) cudaGetLastError(); return false; }
if (!per_dev[dev]) {
int major = 0, minor = 0;
if (cudaDeviceGetAttribute(&major, cudaDevAttrComputeCapabilityMajor, dev) != cudaSuccess ||
cudaDeviceGetAttribute(&minor, cudaDevAttrComputeCapabilityMinor, dev) != cudaSuccess) { (void) cudaGetLastError(); return false; }
- per_dev[dev] = (major == 6 && minor == 1) || (major == 8 && minor == 6) ? 2 : 1;
+ per_dev[dev] = (major == 6 && minor == 1) || (major == 8 && (minor == 6 || minor == 0)) ? 2 : 1;
}
return per_dev[dev] == 2;
}