Merge pull request #810 from kvcache-ai/v0.2.3

V0.2.3

third_party/llamafile/tinyblas_cpu_sgemm.inc

@@ -323,20 +323,17 @@ bool llamafile_sgemm(long m, long n, long k, const void* A, long lda, const void
 #if QK_K == 256
 #if defined(__x86_64__) || defined(_M_X64)
 #if defined(__AVX2__) && (defined(__FMA__) || (defined(_MSC_VER) && (defined(__AVX2__) || defined(__AVX512F__))))
-    // if (X86_CHECK(AVX2) && X86_CHECK(FMA)) {
-    if (Btype == GGML_TYPE_Q8_K && Ctype == GGML_TYPE_F32) {
-        if (iqk_mul_mat(m, n, k * QK_K, Atype, A, B, (float*)C, ldc, ith, nth)) {
-            return true;
-        }
-    }
-    if ((Btype == GGML_TYPE_Q8_0 || Btype == GGML_TYPE_Q8_1) && Ctype == GGML_TYPE_F32) {
-        // assert(QK8_0 == QK8_1 == QK4_0 == QK4_1 == QK5_0 == QK5_1 == 32);
-        assert((QK8_0 == 32) && (QK8_1 == 32) && (QK4_0 == 32) && (QK4_1 == 32) && (QK5_0 == 32) && (QK5_1 == 32));
-        if (iqk_mul_mat(m, n, k * QK8_0, Atype, A, B, (float*)C, ldc, ith, nth)) {
+    /* 
+    moonll
+    more Btype accept
+    }*/
+
+    if (Ctype == GGML_TYPE_F32){
+        if (iqk_mul_mat(m, n, k * ggml_blck_size(ggml_type(Atype)), Atype, A,lda,Btype, B,ldb, (float*)C, ldc, ith, nth)) {
             return true;
         }
     }
-    // }
+
 #endif
 #elif defined __aarch64__ && defined __ARM_FEATURE_DOTPROD && !defined _MSC_VER
     if (Btype == GGML_TYPE_Q8_K && Ctype == GGML_TYPE_F32) {