[ADD] support multi-gpu qlen>1 q5_k

pyproject.toml

@@ -3,7 +3,8 @@ requires = [
   "setuptools",
   "torch >= 2.3.0", 
   "ninja",
-  "packaging"
+  "packaging",
+  "cpufeature"
   ]
 build-backend = "setuptools.build_meta"
 

third_party/llamafile/iqk_mul_mat.inc

@@ -22,7 +22,7 @@
 
 #include <cstring>
 #include <type_traits>
-#if defined __x86_64__ || defined __aarch64__
+#if defined __x86_64__ || defined __aarch64__ || defined(_M_X64)
 
 #include "llama.cpp/ggml-impl.h"
 #include "llama.cpp/ggml-quants.h"
@@ -225,7 +225,7 @@ bool iqk_mul_mat_moe(long Nx, long Ny, long ne00, int ne11, int typeA, const voi
     return true;
 }
 
-#if defined __x86_64__
+#if defined __x86_64__ || defined(_M_X64)
 
 #if defined HAVE_FANCY_SIMD
     #undef HAVE_FANCY_SIMD
@@ -1412,7 +1412,8 @@ template <typename Dequantizer> void MulMat::set_functions(MulMat& m) {
 
 bool MulMat::set_mul_mat(int typeA, int ne00, MulMat& mm, int& row_size_q8, int) {
 
-    row_size_q8 = ggml_row_size(GGML_TYPE_Q8_K, ne00);
+    if (ne00 % ggml_blck_size(GGML_TYPE_Q8_K) == 0)
+        row_size_q8 = ggml_row_size(GGML_TYPE_Q8_K, ne00);
 
     switch (typeA) {
         case GGML_TYPE_Q2_K: