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++
fix(sm75 merge): stray brace after hf() in prefill/kernels.cu (#2093 #1821 resolution)
Niko1221f260925adc87
diff --git a/src/prefill/kernels.cu b/src/prefill/kernels.cuindex 29c9176cb..a637797ae 100644--- a/src/prefill/kernels.cu+++ b/src/prefill/kernels.cu@@ -42,7 +42,6 @@ __device__ __forceinline__ uint16_t bf(float f) {__device__ __forceinline__ uint16_t bf_lo(float f, uint16_t hi) { return bf(f - __uint_as_float((uint32_t) hi << 16)); }__device__ __forceinline__ float sigm(float x) { return 1.0f / (1.0f + __expf(-x)); }__device__ __forceinline__ uint16_t hf(float f) { return __half_as_ushort(__float2half_rn(f)); }-}// A SwiGLU product for an FP16 GEMM: saturated, so a token with a massive activation cannot turn into inf and then// NaN in the down projection (decode's q8_1 has room to ~8e6; FP16 ends at 65504). A NaN stays NaN (fminf/fmaxf// would make it -65504 and hide where it came from); finite values below 65504 round exactly as before.