Compare "GGUF" with file header char by char

1.  Set GGUF_MAGIC to "GGUF" string instead of int value
2. Compare "GGUF" char by char to ensure its byte order
3. Move bytes swap code from convert.py to gguf.py write_tensor_data

convert.py

@@ -932,8 +932,6 @@ class OutputFile:
             elapsed = time.time() - start
             size = ' x '.join(f"{dim:6d}" for dim in lazy_tensor.shape)
             padi = len(str(len(model)))
-            if endianess==gguf.GGUFEndian.BIG:
-                ndarray.byteswap(inplace=True)
             print(f"[{i+1:{padi}d}/{len(model)}] Writing tensor {name:38s} | size {size:16} | type {lazy_tensor.data_type.name:4} | T+{int(elapsed):4}")
             of.gguf.write_tensor_data(ndarray)
 

examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp

@@ -536,7 +536,7 @@ static bool is_ggml_file(const char * filename) {
     if (file.size < 4) {
         return false;
     }
-    uint32_t magic = file.read_u32();
+    std::string magic = file.read_string(4);
     return magic == GGUF_MAGIC;
 }
 

ggml.c

@@ -20813,7 +20813,7 @@ struct gguf_kv {
 };
 
 struct gguf_header {
-    uint32_t magic;
+    char magic[4];
     uint32_t version;
     uint64_t n_tensors; // GGUFv2
     uint64_t n_kv;      // GGUFv2
@@ -20883,7 +20883,7 @@ static bool gguf_fread_str_v1(FILE * file, struct gguf_str * p, size_t * offset)
 struct gguf_context * gguf_init_empty(void) {
     struct gguf_context * ctx = GGML_ALIGNED_MALLOC(sizeof(struct gguf_context));
 
-    ctx->header.magic     = GGUF_MAGIC;
+    memcpy(ctx->header.magic, GGUF_MAGIC, sizeof(ctx->header.magic));
     ctx->header.version   = GGUF_VERSION;
     ctx->header.n_tensors = 0;
     ctx->header.n_kv      = 0;
@@ -20909,16 +20909,18 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
     // offset from start of file
     size_t offset = 0;
 
-    uint32_t magic = 0;
+    char magic[4];
 
     // check the magic before making allocations
     {
         gguf_fread_el(file, &magic, sizeof(magic), &offset);
 
-        if (magic != GGUF_MAGIC) {
-            fprintf(stderr, "%s: invalid magic number %08x\n", __func__, magic);
-            fclose(file);
-            return NULL;
+        for (uint32_t i = 0; i < sizeof(magic); i++) {
+            if (magic[i] != GGUF_MAGIC[i]) {
+                fprintf(stderr, "%s: invalid magic characters %s.\n", __func__, magic);
+                fclose(file);
+                return NULL;
+            }
         }
     }
 
@@ -20928,7 +20930,8 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
 
     // read the header
     {
-        ctx->header.magic = magic;
+        strncpy(ctx->header.magic, magic, 4);
+
 
         ctx->kv    = NULL;
         ctx->infos = NULL;

ggml.h

@@ -231,17 +231,7 @@
 #define GGML_EXIT_SUCCESS 0
 #define GGML_EXIT_ABORTED 1
 
-#if defined(__linux__) 
-    #include <endian.h>
-    #if BYTE_ORDER == LITTLE_ENDIAN
-        #define GGUF_MAGIC 0x46554747
-    #elif BYTE_ORDER == BIG_ENDIAN
-        #define GGUF_MAGIC 0x47475546
-    #endif
-#else
-    // Use little endian magic uint_32 value
-    #define GGUF_MAGIC 0x46554747
-#endif
+#define GGUF_MAGIC "GGUF"
 
 #define GGUF_VERSION 3
 

gguf-py/gguf/gguf.py

@@ -804,6 +804,8 @@ class GGUFWriter:
             fp.write(bytes([0] * pad))
 
     def write_tensor_data(self, tensor: np.ndarray[Any, Any]):
+        if self.endianess==GGUFEndian.BIG:
+            tensor.byteswap(inplace=True)
         self.write_padding(self.fout, self.fout.tell())
         tensor.tofile(self.fout)
         self.write_padding(self.fout, tensor.nbytes)