Bump llama.cpp to b2474

The release just before ggml-cuda.cu refactoring

llm/ext_server/server.cpp

@@ -343,6 +343,12 @@ struct llama_server_context
 
     ~llama_server_context()
     {
+        if (clp_ctx)
+        {
+            LOG_INFO("freeing clip model", {});
+            clip_free(clp_ctx);
+            clp_ctx = nullptr;
+        }
         if (ctx)
         {
             llama_free(ctx);

llm/llama.cpp

@@ -1 +1 @@
-Subproject commit ceca1aef0738b57951cd12c603c3477e75312dec
+Subproject commit 91f8ad167dcd24b54615b468d9dd764ebe1d37ad

llm/patches/01-cache.diff

@@ -1,21 +0,0 @@
-diff --git a/examples/server/server.cpp b/examples/server/server.cpp
-index 8fe5e0b1..3e82acb9 100644
---- a/examples/server/server.cpp
-+++ b/examples/server/server.cpp
-@@ -997,13 +997,15 @@ struct llama_server_context
-                 slot.n_sent_text += result.text_to_send.size();
-                 // add the token to slot queue and cache
-             }
--            slot.add_token_string(result);
-+
-             if (slot.params.stream)
-             {
-                 send_partial_response(slot, result);
-             }
-         }
- 
-+        slot.add_token_string(result);
-+
-         if (incomplete)
-         {
-             slot.has_next_token = true;

llm/patches/02-cudaleaks.diff

@@ -1,117 +0,0 @@
-diff --git a/examples/server/server.cpp b/examples/server/server.cpp
-index 8fe5e0b1..53bf39c1 100644
---- a/examples/server/server.cpp
-+++ b/examples/server/server.cpp
-@@ -31,6 +31,10 @@
- #include <atomic>
- #include <signal.h>
- 
-+#ifdef GGML_USE_CUBLAS
-+extern "C" GGML_CALL void ggml_free_cublas(void);
-+#endif
-+
- using json = nlohmann::json;
- 
- struct server_params {
-@@ -363,6 +367,10 @@ struct llama_server_context
-             llama_free_model(model);
-             model = nullptr;
-         }
-+
-+#ifdef GGML_USE_CUBLAS
-+        ggml_free_cublas();
-+#endif
-     }
- 
-     bool load_model(const gpt_params &params_)
-@@ -3543,6 +3551,7 @@ int main(int argc, char **argv)
-     sigemptyset (&sigint_action.sa_mask);
-     sigint_action.sa_flags = 0;
-     sigaction(SIGINT, &sigint_action, NULL);
-+    sigaction(SIGUSR1, &sigint_action, NULL);
- #elif defined (_WIN32)
-     auto console_ctrl_handler = +[](DWORD ctrl_type) -> BOOL {
-         return (ctrl_type == CTRL_C_EVENT) ? (signal_handler(SIGINT), true) : false;
-diff --git a/ggml-cuda.cu b/ggml-cuda.cu
-index 72bcec8c..6c934e8c 100644
---- a/ggml-cuda.cu
-+++ b/ggml-cuda.cu
-@@ -43,6 +43,7 @@
- #define __shfl_xor_sync(mask, var, laneMask, width) __shfl_xor(var, laneMask, width)
- #define cublasComputeType_t hipblasDatatype_t //deprecated, new hipblasComputeType_t not in 5.6
- #define cublasCreate hipblasCreate
-+#define cublasDestroy hipblasDestroy
- #define cublasGemmEx hipblasGemmEx
- #define cublasGemmBatchedEx hipblasGemmBatchedEx
- #define cublasGemmStridedBatchedEx hipblasGemmStridedBatchedEx
-@@ -8751,10 +8752,10 @@ GGML_CALL bool ggml_cublas_loaded(void) {
-     return g_cublas_loaded;
- }
- 
--GGML_CALL void ggml_init_cublas() {
--    static bool initialized = false;
-+static bool g_cublas_initialized = false;
- 
--    if (!initialized) {
-+GGML_CALL void ggml_init_cublas() {
-+    if (!g_cublas_initialized) {
- 
- #ifdef __HIP_PLATFORM_AMD__
-         // Workaround for a rocBLAS bug when using multiple graphics cards:
-@@ -8764,7 +8765,7 @@ GGML_CALL void ggml_init_cublas() {
- #endif
- 
-         if (cudaGetDeviceCount(&g_device_count) != cudaSuccess) {
--            initialized = true;
-+            g_cublas_initialized = true;
-             g_cublas_loaded = false;
-             fprintf(stderr, "%s: no " GGML_CUDA_NAME " devices found, " GGML_CUDA_NAME " will be disabled\n", __func__);
-             return;
-@@ -8835,7 +8836,7 @@ GGML_CALL void ggml_init_cublas() {
-         // configure logging to stdout
-         // CUBLAS_CHECK(cublasLoggerConfigure(1, 1, 0, nullptr));
- 
--        initialized = true;
-+        g_cublas_initialized = true;
-         g_cublas_loaded = true;
-     }
- }
-@@ -12490,3 +12491,23 @@ GGML_CALL int ggml_backend_cuda_reg_devices() {
-     }
-     return device_count;
- }
-+
-+
-+extern "C" GGML_CALL void ggml_free_cublas(void);
-+GGML_CALL void ggml_free_cublas(void) {
-+    for (int id = 0; id < g_device_count; ++id) {
-+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
-+        if (g_device_caps[id].vmm) {
-+            CU_CHECK(cuMemUnmap(g_cuda_pool_addr[id], g_cuda_pool_size[id]));
-+            g_cuda_pool_size[id] = 0;
-+            g_cuda_pool_addr[id] = 0;
-+        }
-+#endif
-+        // TODO: free legacy non-vmm memory
-+        // destroy cublas handle
-+        CUBLAS_CHECK(cublasDestroy(g_cublas_handles[id]));
-+        g_cublas_handles[id] = nullptr;
-+    }
-+
-+    g_cublas_initialized = false;
-+}
-\ No newline at end of file
-diff --git a/ggml-cuda.h b/ggml-cuda.h
-index b1ebd61d..6dd58ddf 100644
---- a/ggml-cuda.h
-+++ b/ggml-cuda.h
-@@ -23,6 +23,9 @@ GGML_API GGML_CALL void   ggml_init_cublas(void);
- // Returns `true` if there are available CUDA devices and cublas loads successfully; otherwise, it returns `false`.
- GGML_API GGML_CALL bool   ggml_cublas_loaded(void);
- 
-+// Release CUDA resources
-+GGML_API GGML_CALL void   ggml_free_cublas(void);
-+
- GGML_API GGML_CALL void * ggml_cuda_host_malloc(size_t size);
- GGML_API GGML_CALL void   ggml_cuda_host_free(void * ptr);
- 

llm/patches/05-fix-clip-free.diff

@@ -1,45 +0,0 @@
-From 9192432daf90b1bfec75577434a99b4ea70d54c8 Mon Sep 17 00:00:00 2001
-From: Michael Yang <mxyng@pm.me>
-Date: Thu, 14 Mar 2024 12:09:50 -0700
-Subject: [PATCH] fix clip free
-
----
- examples/llava/clip.cpp    | 4 ++++
- examples/server/server.cpp | 6 ++++++
- 2 files changed, 10 insertions(+)
-
-diff --git a/examples/llava/clip.cpp b/examples/llava/clip.cpp
-index ef9e4ba7..b4ddfe6b 100644
---- a/examples/llava/clip.cpp
-+++ b/examples/llava/clip.cpp
-@@ -1673,6 +1673,10 @@ void clip_free(clip_ctx * ctx) {
-     ggml_free(ctx->ctx_data);
-     gguf_free(ctx->ctx_gguf);
- 
-+    ggml_backend_buffer_free(ctx->params_buffer);
-+    ggml_backend_buffer_free(ctx->compute_buffer);
-+    ggml_backend_free(ctx->backend);
-+    ggml_gallocr_free(ctx->compute_alloc);
-     delete ctx;
- }
- 
-diff --git a/examples/server/server.cpp b/examples/server/server.cpp
-index 8fe5e0b1..f927336b 100644
---- a/examples/server/server.cpp
-+++ b/examples/server/server.cpp
-@@ -353,6 +353,12 @@ struct llama_server_context
- 
-     ~llama_server_context()
-     {
-+        if (clp_ctx)
-+        {
-+            LOG_INFO("freeing clip model", {});
-+            clip_free(clp_ctx);
-+            clp_ctx = nullptr;
-+        }
-         if (ctx)
-         {
-             llama_free(ctx);
--- 
-2.43.2
-