llama : fix platforms without mmap (#4578)

* llama : fix platforms without mmap

* win32 : limit prefetch size to the file size

* fix win32 error clobber, unnecessary std::string in std::runtime_error

Author: slaren

ggml-cuda.cu

@@ -7702,7 +7702,8 @@ inline void ggml_cuda_op_scale(
     GGML_ASSERT(src0->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
-    const float scale = ((float *) dst->op_params)[0];
+    float scale;
+    memcpy(&scale, dst->op_params, sizeof(float));
 
     scale_f32_cuda(src0_dd, dst_dd, scale, ggml_nelements(src0), main_stream);
     CUDA_CHECK(cudaGetLastError());

ggml.c

@@ -10335,7 +10335,8 @@ static void ggml_compute_forward_scale_f32(
     }
 
     // scale factor
-    const float v = *(float *) dst->op_params;
+    float v;
+    memcpy(&v, dst->op_params, sizeof(float));
 
     const int ith = params->ith;
     const int nth = params->nth;
@@ -15152,7 +15153,8 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
             {
                 // necessary for llama
                 if (src0->grad) {
-                    const float s = ((float *) tensor->op_params)[0];
+                    float s;
+                    memcpy(&s, tensor->op_params, sizeof(float));
 
                     src0->grad =
                         ggml_add_or_set(ctx,

llama.cpp

@@ -778,7 +778,7 @@ struct llama_file {
             throw std::runtime_error(format("read error: %s", strerror(errno)));
         }
         if (ret != 1) {
-            throw std::runtime_error(std::string("unexpectedly reached end of file"));
+            throw std::runtime_error("unexpectedly reached end of file");
         }
     }
 
@@ -931,29 +931,29 @@ struct llama_mmap {
 #elif defined(_WIN32)
     static constexpr bool SUPPORTED = true;
 
-    llama_mmap(struct llama_file * file, bool prefetch = true, bool numa = false) {
-        (void) numa;
+    llama_mmap(struct llama_file * file, size_t prefetch = (size_t) -1, bool numa = false) {
+        GGML_UNUSED(numa);
 
         size = file->size;
 
         HANDLE hFile = (HANDLE) _get_osfhandle(_fileno(file->fp));
 
         HANDLE hMapping = CreateFileMappingA(hFile, NULL, PAGE_READONLY, 0, 0, NULL);
-        DWORD error = GetLastError();
 
         if (hMapping == NULL) {
+            DWORD error = GetLastError();
             throw std::runtime_error(format("CreateFileMappingA failed: %s", llama_format_win_err(error).c_str()));
         }
 
         addr = MapViewOfFile(hMapping, FILE_MAP_READ, 0, 0, 0);
-        error = GetLastError();
+        DWORD error = GetLastError();
         CloseHandle(hMapping);
 
         if (addr == NULL) {
             throw std::runtime_error(format("MapViewOfFile failed: %s", llama_format_win_err(error).c_str()));
         }
 
-        if (prefetch) {
+        if (prefetch > 0) {
             // PrefetchVirtualMemory is only present on Windows 8 and above, so we dynamically load it
             BOOL (WINAPI *pPrefetchVirtualMemory) (HANDLE, ULONG_PTR, PWIN32_MEMORY_RANGE_ENTRY, ULONG);
             HMODULE hKernel32 = GetModuleHandleW(L"kernel32.dll");
@@ -965,9 +965,9 @@ struct llama_mmap {
                 // advise the kernel to preload the mapped memory
                 WIN32_MEMORY_RANGE_ENTRY range;
                 range.VirtualAddress = addr;
-                range.NumberOfBytes = (SIZE_T)size;
+                range.NumberOfBytes = (SIZE_T) std::min(size, prefetch);
                 if (!pPrefetchVirtualMemory(GetCurrentProcess(), 1, &range, 0)) {
-                    fprintf(stderr, "warning: PrefetchVirtualMemory failed: %s\n",
+                    LLAMA_LOG_WARN("warning: PrefetchVirtualMemory failed: %s\n",
                             llama_format_win_err(GetLastError()).c_str());
                 }
             }
@@ -982,26 +982,26 @@ struct llama_mmap {
 
     ~llama_mmap() {
         if (!UnmapViewOfFile(addr)) {
-            fprintf(stderr, "warning: UnmapViewOfFile failed: %s\n",
+            LLAMA_LOG_WARN("warning: UnmapViewOfFile failed: %s\n",
                     llama_format_win_err(GetLastError()).c_str());
         }
     }
 #else
     static constexpr bool SUPPORTED = false;
 
-    llama_mmap(struct llama_file * file, bool prefetch = true, bool numa = false) {
-        (void) file;
-        (void) prefetch;
-        (void) numa;
+    llama_mmap(struct llama_file * file, size_t prefetch = -1, bool numa = false) {
+        GGML_UNUSED(file);
+        GGML_UNUSED(prefetch);
+        GGML_UNUSED(numa);
 
-        throw std::runtime_error(std::string("mmap not supported"));
+        throw std::runtime_error("mmap not supported");
     }
 
-    void unmap(size_t offset, size_t len) {
-        (void) offset;
-        (void) len;
+    void unmap_fragment(size_t first, size_t last) {
+        GGML_UNUSED(first);
+        GGML_UNUSED(last);
 
-        throw std::runtime_error(std::string("mmap not supported"));
+        throw std::runtime_error("mmap not supported");
     }
 #endif
 };