build: disable cuda compression

Most of the time this is not an error.

CMakeLists.txt

@@ -24,7 +24,7 @@ set(GGML_LLAMAFILE ON)
 set(GGML_CUDA_PEER_MAX_BATCH_SIZE 128)
 set(GGML_CUDA_GRAPHS ON)
 set(GGML_CUDA_FA ON)
-set(GGML_CUDA_COMPRESSION_MODE default)
+set(GGML_CUDA_COMPRESSION_MODE none)
 
 if((CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_OSX_ARCHITECTURES MATCHES "arm64")
     OR (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_SYSTEM_PROCESSOR MATCHES "arm|aarch64|ARM64|ARMv[0-9]+"))

convert/convert_test.go

@@ -130,6 +130,7 @@ func TestConvertModel(t *testing.T) {
 			if err != nil {
 				t.Fatal(err)
 			}
+			defer expectFile.Close()
 
 			var expect map[string]string
 			if err := json.NewDecoder(expectFile).Decode(&expect); err != nil {

discover/gpu_info_cudart.c

@@ -58,7 +58,7 @@ void cudart_init(char *cudart_lib_path, cudart_init_resp_t *resp) {
     LOG(resp->ch.verbose, "cudaSetDevice err: %d\n", ret);
     UNLOAD_LIBRARY(resp->ch.handle);
     resp->ch.handle = NULL;
-    if (ret == CUDA_ERROR_INSUFFICIENT_DRIVER) {
+    if (ret == CUDART_ERROR_INSUFFICIENT_DRIVER) {
       resp->err = strdup("your nvidia driver is too old or missing.  If you have a CUDA GPU please upgrade to run ollama");
       return;
     }
@@ -168,9 +168,9 @@ void cudart_bootstrap(cudart_handle_t h, int i, mem_info_t *resp) {
   resp->free = memInfo.free;
   resp->used = memInfo.used;
 
-  LOG(h.verbose, "[%s] CUDA totalMem %lu\n", resp->gpu_id, resp->total);
-  LOG(h.verbose, "[%s] CUDA freeMem %lu\n", resp->gpu_id, resp->free);
-  LOG(h.verbose, "[%s] CUDA usedMem %lu\n", resp->gpu_id, resp->used);
+  LOG(h.verbose, "[%s] CUDA totalMem %llu\n", resp->gpu_id, resp->total);
+  LOG(h.verbose, "[%s] CUDA freeMem %llu\n", resp->gpu_id, resp->free);
+  LOG(h.verbose, "[%s] CUDA usedMem %llu\n", resp->gpu_id, resp->used);
   LOG(h.verbose, "[%s] Compute Capability %d.%d\n", resp->gpu_id, resp->major, resp->minor);
 }
 

discover/gpu_info_nvcuda.c

@@ -193,8 +193,8 @@ void nvcuda_bootstrap(nvcuda_handle_t h, int i, mem_info_t *resp) {
   resp->total = memInfo.total;
   resp->free = memInfo.free;
 
-  LOG(h.verbose, "[%s] CUDA totalMem %lu mb\n", resp->gpu_id, resp->total / 1024 / 1024);
-  LOG(h.verbose, "[%s] CUDA freeMem %lu mb\n", resp->gpu_id, resp->free / 1024 / 1024);
+  LOG(h.verbose, "[%s] CUDA totalMem %llu mb\n", resp->gpu_id, resp->total / 1024 / 1024);
+  LOG(h.verbose, "[%s] CUDA freeMem %llu mb\n", resp->gpu_id, resp->free / 1024 / 1024);
   LOG(h.verbose, "[%s] Compute Capability %d.%d\n", resp->gpu_id, resp->major, resp->minor);
 
   

fs/ggml/ggml.go

@@ -149,7 +149,7 @@ func keyValue[T valueTypes | arrayValueTypes](kv KV, key string, defaultValue ..
 		return val.(T)
 	}
 
-	slog.Warn("key not found", "key", key, "default", defaultValue[0])
+	slog.Debug("key not found", "key", key, "default", defaultValue[0])
 	return defaultValue[0]
 }
 

integration/utils_test.go

@@ -217,6 +217,7 @@ func InitServerConnection(ctx context.Context, t *testing.T) (*api.Client, strin
 					slog.Error("failed to open server log", "logfile", lifecycle.ServerLogFile, "error", err)
 					return
 				}
+				defer fp.Close()
 				data, err := io.ReadAll(fp)
 				if err != nil {
 					slog.Error("failed to read server log", "logfile", lifecycle.ServerLogFile, "error", err)

llama/llama.go

@@ -2,6 +2,7 @@ package llama
 
 /*
 #cgo CFLAGS: -std=c11
+#cgo windows CFLAGS: -Wno-dll-attribute-on-redeclaration
 #cgo CXXFLAGS: -std=c++17
 #cgo CPPFLAGS: -I${SRCDIR}/llama.cpp/include
 #cgo CPPFLAGS: -I${SRCDIR}/llama.cpp/common

llm/llm_windows.go

@@ -7,6 +7,7 @@ import (
 const (
 	CREATE_DEFAULT_ERROR_MODE   = 0x04000000
 	ABOVE_NORMAL_PRIORITY_CLASS = 0x00008000
+	CREATE_NO_WINDOW            = 0x08000000
 )
 
 var LlamaServerSysProcAttr = &syscall.SysProcAttr{
@@ -18,5 +19,5 @@ var LlamaServerSysProcAttr = &syscall.SysProcAttr{
 	//
 	// Setting Above Normal priority class ensures when running as a "background service"
 	// with "programs" given best priority, we aren't starved of cpu cycles
-	CreationFlags: CREATE_DEFAULT_ERROR_MODE | ABOVE_NORMAL_PRIORITY_CLASS,
+	CreationFlags: CREATE_DEFAULT_ERROR_MODE | ABOVE_NORMAL_PRIORITY_CLASS | CREATE_NO_WINDOW,
 }

ml/backend/ggml/ggml/src/ggml.go

@@ -3,6 +3,7 @@ package ggml
 // #cgo CXXFLAGS: -std=c++17
 // #cgo CPPFLAGS: -DNDEBUG -DGGML_USE_CPU
 // #cgo CPPFLAGS: -I${SRCDIR}/../include -I${SRCDIR}/ggml-cpu
+// #cgo windows CFLAGS: -Wno-dll-attribute-on-redeclaration
 // #cgo windows LDFLAGS: -lmsvcrt -static -static-libgcc -static-libstdc++
 // #include <stdlib.h>
 // #include "ggml-backend.h"

server/sched.go

@@ -691,11 +691,13 @@ func (runner *runnerRef) LogValue() slog.Value {
 	attrs = append(attrs,
 		slog.String("size", format.HumanBytes2(runner.estimatedTotal)),
 		slog.String("vram", format.HumanBytes2(runner.estimatedVRAM)),
-		slog.Int("num_ctx", runner.NumCtx),
 		slog.Int("parallel", runner.numParallel),
 		slog.Int("pid", runner.pid),
 		slog.String("model", runner.modelPath),
 	)
+	if runner.Options != nil {
+		attrs = append(attrs, slog.Int("num_ctx", runner.Options.NumCtx))
+	}
 	return slog.GroupValue(attrs...)
 }