ml/backend/ggml: fix crash on dlopen for non-AVX systems (#8976)

llama/patches/0016-remove-sgemm-global-variables.patch

@@ -0,0 +1,55 @@
+From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
+From: jmorganca <jmorganca@gmail.com>
+Date: Sun, 9 Feb 2025 17:22:15 -0800
+Subject: [PATCH] remove sgemm global variables
+
+removes the 'iq4nlt' global variable in sgemm.cpp that causes
+a runtime crash when calling dlopen on ggml-cpu libraries as
+its initialization depends on AVX instructions the host machine
+may not have
+---
+ ggml/src/ggml-cpu/llamafile/sgemm.cpp | 17 +++++++++--------
+ 1 file changed, 9 insertions(+), 8 deletions(-)
+
+diff --git a/ggml/src/ggml-cpu/llamafile/sgemm.cpp b/ggml/src/ggml-cpu/llamafile/sgemm.cpp
+index 8fce576c..3f260ce5 100644
+--- a/ggml/src/ggml-cpu/llamafile/sgemm.cpp
++++ b/ggml/src/ggml-cpu/llamafile/sgemm.cpp
+@@ -279,14 +279,6 @@ template <> inline __m256bh load(const float *p) {
+ }
+ #endif
+ 
+-////////////////////////////////////////////////////////////////////////////////////////////////////
+-// CONSTANTS
+-
+-#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
+-static const int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113};
+-static const __m128i iq4nlt = _mm_loadu_si128((const __m128i *) kvalues_iq4nl);
+-#endif
+-
+ ////////////////////////////////////////////////////////////////////////////////////////////////////
+ // FLOATING POINT MATRIX MULTIPLICATION
+ 
+@@ -613,6 +605,14 @@ class tinyBLAS_Q0_AVX {
+                     TC *C, int64_t ldc,
+                     int ith, int nth)
+         : A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) {
++        const int8_t kvalues_iq4nl[16] = {
++            -127, -104, -83, -65,
++            -49,  -35,  -22, -10,
++              1,   13,   25,  38,
++             53,   69,   89, 113
++        };
++
++        iq4nlt = _mm_loadu_si128((const __m128i *)kvalues_iq4nl);
+     }
+ 
+     void matmul(int64_t m, int64_t n) {
+@@ -1037,6 +1037,7 @@ class tinyBLAS_Q0_AVX {
+     const int64_t ldc;
+     const int ith;
+     const int nth;
++    __m128i iq4nlt;
+ };
+ #endif // __AVX__
+ 

ml/backend/ggml/ggml/src/ggml-cpu/llamafile/sgemm.cpp

@@ -279,14 +279,6 @@ template <> inline __m256bh load(const float *p) {
 }
 #endif
 
-////////////////////////////////////////////////////////////////////////////////////////////////////
-// CONSTANTS
-
-#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
-static const int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113};
-static const __m128i iq4nlt = _mm_loadu_si128((const __m128i *) kvalues_iq4nl);
-#endif
-
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 // FLOATING POINT MATRIX MULTIPLICATION
 
@@ -613,6 +605,14 @@ class tinyBLAS_Q0_AVX {
                     TC *C, int64_t ldc,
                     int ith, int nth)
         : A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) {
+        const int8_t kvalues_iq4nl[16] = {
+            -127, -104, -83, -65,
+            -49,  -35,  -22, -10,
+              1,   13,   25,  38,
+             53,   69,   89, 113
+        };
+
+        iq4nlt = _mm_loadu_si128((const __m128i *)kvalues_iq4nl);
     }
 
     void matmul(int64_t m, int64_t n) {
@@ -1037,6 +1037,7 @@ class tinyBLAS_Q0_AVX {
     const int64_t ldc;
     const int ith;
     const int nth;
+    __m128i iq4nlt;
 };
 #endif // __AVX__