Allow s390x to load little endian models unmodified

examples/gguf-split/gguf-split.cpp

@@ -328,14 +328,20 @@ struct split_strategy {
                 const char * t_name = gguf_get_tensor_name(ctx_out, i);
                 struct ggml_tensor * t = ggml_get_tensor(ctx_meta, t_name);
                 auto n_bytes = ggml_nbytes(t);
+                auto n_elements = ggml_nelements(t) / ggml_blck_size(t->type);
                 read_buf.resize(n_bytes);
 
                 // calculate offset
                 auto i_tensor_in = gguf_find_tensor(ctx_gguf, t_name); // idx of tensor in the input file
                 auto offset = gguf_get_data_offset(ctx_gguf) + gguf_get_tensor_offset(ctx_gguf, i_tensor_in);
 
+                ggml_byteswap_t byteswap_func = nullptr;
+                if (gguf_needs_byteswap(ctx_gguf)) {
+                    byteswap_func = ggml_get_type_traits(t->type)->byteswap;
+                }
+
                 // copy tensor from input to output file
-                copy_file_to_file(f_input, fout, offset, n_bytes);
+                copy_file_to_file(f_input, fout, offset, n_bytes, n_elements, byteswap_func);
                 zeros(fout, GGML_PAD(n_bytes, GGUF_DEFAULT_ALIGNMENT) - n_bytes);
             }
 
@@ -346,13 +352,18 @@ struct split_strategy {
         }
     }
 
-    void copy_file_to_file(std::ifstream & f_in, std::ofstream & f_out, const size_t in_offset, const size_t len) {
+    void copy_file_to_file(std::ifstream & f_in, std::ofstream & f_out, const size_t in_offset, const size_t len, const size_t elements, ggml_byteswap_t byteswap_func) {
         // TODO: detect OS and use copy_file_range() here for better performance
         if (read_buf.size() < len) {
             read_buf.resize(len);
         }
         f_in.seekg(in_offset);
         f_in.read((char *)read_buf.data(), len);
+
+        if (byteswap_func != nullptr) {
+            byteswap_func(read_buf.data(), elements);
+        }
+
         f_out.write((const char *)read_buf.data(), len);
     }
 };
@@ -541,6 +552,13 @@ static void gguf_merge(const split_params & split_params) {
             f_input.seekg(offset);
             f_input.read((char *)read_data.data(), n_bytes);
 
+            if (gguf_needs_byteswap(ctx_gguf)) {
+                auto byteswap = ggml_get_type_traits(t->type)->byteswap;
+                if (byteswap != nullptr) {
+                    byteswap(read_data.data(), ggml_nelements(t) / ggml_blck_size(t->type));
+                }
+            }
+
             // write tensor data + padding
             fout.write((const char *)read_data.data(), n_bytes);
             zeros(fout, GGML_PAD(n_bytes, GGUF_DEFAULT_ALIGNMENT) - n_bytes);

ggml/include/ggml.h

@@ -2144,6 +2144,7 @@ extern "C" {
 #endif
     typedef void (*ggml_to_float_t)  (const void  * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
     typedef void (*ggml_from_float_t)(const float * GGML_RESTRICT x, void  * GGML_RESTRICT y, int64_t k);
+    typedef void (*ggml_byteswap_t)  (      void  * GGML_RESTRICT buffer, size_t elements);
 
     struct ggml_type_traits {
         const char             * type_name;
@@ -2153,6 +2154,7 @@ extern "C" {
         bool                     is_quantized;
         ggml_to_float_t          to_float;
         ggml_from_float_t        from_float_ref;
+        ggml_byteswap_t          byteswap;
     };
 
     GGML_API const struct ggml_type_traits * ggml_get_type_traits(enum ggml_type type);

ggml/include/gguf.h

@@ -197,6 +197,9 @@ extern "C" {
     // writes the meta data to pointer "data"
     GGML_API void   gguf_get_meta_data(const struct gguf_context * ctx, void * data);
 
+    // returns true if gguf file needs byteswapping when reading. byteswapping for writing not implemented
+    GGML_API bool gguf_needs_byteswap(const struct gguf_context * ctx);
+
 #ifdef  __cplusplus
 }
 #endif

ggml/src/ggml-impl.h

@@ -28,6 +28,14 @@
 #include <immintrin.h>
 #endif
 
+#if defined(__gnu_linux__)
+#include <byteswap.h>
+#else // defined(__gnu_linux__)
+#define bswap_16(x) (x)
+#define bswap_32(x) (x)
+#define bswap_64(x) (x)
+#endif // defined(__gnu_linux__)
+
 #ifdef __cplusplus
 extern "C" {
 #endif
@@ -553,12 +561,47 @@ static inline ggml_bf16_t ggml_compute_fp32_to_bf16(float s) {
 #define GGML_FP32_TO_BF16(x) ggml_compute_fp32_to_bf16(x)
 #define GGML_BF16_TO_FP32(x) ggml_compute_bf16_to_fp32(x)
 
+// endianness conversion
+static inline void ggml_bswap16(void * value) {
+    *((uint16_t*)value) = bswap_16(*((uint16_t*)value));
+}
+
+static inline void ggml_bswap32(void * value) {
+    *((uint32_t*)value) = bswap_32(*((uint32_t*)value));
+}
+
+static inline void ggml_bswap64(void * value) {
+    *((uint64_t*)value) = bswap_64(*((uint64_t*)value));
+}
+
 #ifdef __cplusplus
 }
 #endif
 
 #ifdef __cplusplus
 #include <vector>
+#include <type_traits>
+
+template <typename T, std::enable_if_t<sizeof(T) == 1, int> = 0>
+static inline void ggml_bswap(T * value)
+{
+    GGML_UNUSED(value);
+}
+
+template <typename T, std::enable_if_t<sizeof(T) == 2, int> = 0>
+static inline void ggml_bswap(T * value) {
+    ggml_bswap16(value);
+}
+
+template <typename T, std::enable_if_t<sizeof(T) == 4, int> = 0>
+static inline void ggml_bswap(T * value) {
+    ggml_bswap32(value);
+}
+
+template <typename T, std::enable_if_t<sizeof(T) == 8, int> = 0>
+static inline void ggml_bswap(T * value) {
+    ggml_bswap64(value);
+}
 
 // expose GGUF internals for test code
 GGML_API size_t gguf_type_size(enum gguf_type type);