metal : null-check buffer alloc to fix OOM crash (#25371)

* metal : null-check ggml_metal_buffer_init result to avoid OOM crash

ggml_backend_metal_buffer_type_alloc_buffer used the result of
ggml_metal_buffer_init without checking for NULL. ggml_metal_buffer_init
returns NULL when the underlying Metal allocation fails (e.g. an
out-of-memory condition), and the following ggml_metal_buffer_is_shared(res)
call dereferences it, turning a recoverable allocation failure into a hard
crash (EXC_BAD_ACCESS). This is easy to hit on memory-constrained devices
such as iOS when a model/context exceeds the available Metal budget.

Log the failure using the existing GGML_LOG_ERROR convention and return
NULL so the allocator surfaces a diagnosable error up the stack instead of
crashing.

* cont : fix log

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

.github/workflows/make-release.yml

@@ -84,11 +84,13 @@ jobs:
 
             New version has been released.
 
+            ## Assets
+
             ${{ steps.desc.outputs.nightly }}
 
-            **Web UI:** the `nightly-tag.txt` asset contains the tag of the corresponding nightly release
+            ## More info
 
-            **More info:** [dist : releases and versioning of ggml-org projects](https://github.com/ggml-org/ggml/discussions/1579)
+            - [Releases and versioning of `ggml-org` projects](https://github.com/ggml-org/ggml/discussions/1579)
 
             ## ${{ steps.desc.outputs.changelog_title }}
 

CMakeLists.txt

@@ -4,7 +4,7 @@ include(CheckIncludeFileCXX)
 
 ### llama.cpp version
 set(LLAMA_VERSION_MAJOR 0)
-set(LLAMA_VERSION_MINOR 2)
+set(LLAMA_VERSION_MINOR 3)
 set(LLAMA_VERSION_PATCH 0)
 set(LLAMA_VERSION_BASE "${LLAMA_VERSION_MAJOR}.${LLAMA_VERSION_MINOR}.${LLAMA_VERSION_PATCH}")
 

ggml/CMakeLists.txt

@@ -4,7 +4,7 @@ project("ggml" C CXX ASM)
 
 ### GGML Version
 set(GGML_VERSION_MAJOR 0)
-set(GGML_VERSION_MINOR 21)
+set(GGML_VERSION_MINOR 22)
 set(GGML_VERSION_PATCH 0)
 set(GGML_VERSION_BASE "${GGML_VERSION_MAJOR}.${GGML_VERSION_MINOR}.${GGML_VERSION_PATCH}")
 

ggml/src/ggml-metal/ggml-metal.cpp

@@ -204,6 +204,11 @@ static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buffer(ggml_ba
     ggml_metal_device_t ctx_dev = (ggml_metal_device_t)buft->device->context;
     ggml_metal_buffer_t res = ggml_metal_buffer_init(ctx_dev, size, shared);
 
+    if (res == NULL) {
+        GGML_LOG_ERROR("%s: failed to allocate Metal buffer of %zu bytes (out of memory)\n", __func__, size);
+        return NULL;
+    }
+
     ggml_backend_buffer_i buf_i = ggml_metal_buffer_is_shared(res)
         ? ggml_backend_metal_buffer_shared_i
         : ggml_backend_metal_buffer_private_i;

ggml/src/ggml-sycl/ggml-sycl.cpp

@@ -6018,6 +6018,11 @@ static bool do_ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, cons
                     a->ne[0] > 128 && a->ne[2] == 1 && src0_type == GGML_TYPE_F16) {
                     return false;
                 }
+
+                if (src0_type == GGML_TYPE_TQ2_0) {
+                    return false;
+                }
+
                 return true;
             }
         case GGML_OP_OUT_PROD:
@@ -6068,6 +6073,9 @@ static bool do_ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, cons
 
         case GGML_OP_SET_ROWS:
             {
+                if (op->type == GGML_TYPE_TQ2_0) {
+                    return false;
+                }
                 auto res = (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16 ||
                             op->src[0]->type == GGML_TYPE_BF16) &&
                            (op->src[1]->type == GGML_TYPE_I64 || op->src[1]->type == GGML_TYPE_I32);
@@ -6186,11 +6194,16 @@ static bool do_ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, cons
                         src1_type == GGML_TYPE_IQ3_XXS ||
                         src1_type == GGML_TYPE_IQ3_S ||
                         src1_type == GGML_TYPE_IQ1_S ||
-                        src1_type == GGML_TYPE_IQ1_M) {
+                        src1_type == GGML_TYPE_IQ1_M ||
+                        src1_type == GGML_TYPE_TQ2_0) {
                         return false;
                     }
                 }
 
+                if (src0_type == GGML_TYPE_TQ2_0 || src1_type == GGML_TYPE_TQ2_0) {
+                    return false;
+                }
+
                 return true;
             }
         case GGML_OP_REPEAT_BACK:

ggml/src/ggml-sycl/mmvq.cpp

@@ -2559,7 +2559,7 @@ void ggml_sycl_op_mul_mat_vec_q(ggml_backend_sycl_context & ctx, const ggml_tens
                 }
                 break;
             default:
-                GGML_ABORT("fatal error: unsupport data type=%s\n", ggml_type_name(src0->type));
+                GGML_ABORT("fatal error: unsupport src0 data type %s\n", ggml_type_name(src0->type));
         }
     }
     GGML_UNUSED(src1);

ggml/src/ggml-sycl/set_rows.cpp

@@ -546,7 +546,8 @@ static void set_rows_sycl(ggml_backend_sycl_context & ctx, const ggml_tensor * s
                 stream);
             break;
         default:
-            GGML_ABORT("Unsupported tensor type!");
+            GGML_ABORT("Unsupported tensor type: src0 %s src1 %s dst %s", ggml_type_name(dst->src[0]->type),
+                ggml_type_name(dst->src[1]->type), ggml_type_name(dst->type));
             break;
     }
 }

ggml/src/ggml-webgpu/wgsl-shaders/argsort.wgsl

@@ -34,11 +34,9 @@ var<uniform> params: Params;
 var<workgroup> shmem_idx: array<u32, WG_SIZE>;
 
 #if ORDER == 0
-#define EXTREME_VALUE 1e30
 #define SWAP_COMPARE_UP >
 #define SWAP_COMPARE_DOWN <
 #else
-#define EXTREME_VALUE -1e30
 #define SWAP_COMPARE_UP <
 #define SWAP_COMPARE_DOWN >
 #endif
@@ -78,11 +76,9 @@ fn main(@builtin(workgroup_id) wid: vec3<u32>,
                 let dir_up = (lid.x & k) == 0;
                 let a_idx = shmem_idx[lid.x];
                 let b_idx = shmem_idx[ixj];
-                let a_val = select(EXTREME_VALUE, src[row_base + a_idx], a_idx < params.src_ne0);
-                let b_val = select(EXTREME_VALUE, src[row_base + b_idx], b_idx < params.src_ne0);
                 let should_swap = select(
-                    (a_val SWAP_COMPARE_DOWN b_val),
-                    (a_val SWAP_COMPARE_UP b_val),
+                    b_idx >= params.src_ne0 || (a_idx < params.src_ne0 && src[row_base + a_idx] SWAP_COMPARE_DOWN src[row_base + b_idx]),
+                    a_idx >= params.src_ne0 || (b_idx < params.src_ne0 && src[row_base + a_idx] SWAP_COMPARE_UP src[row_base + b_idx]),
                     dir_up);
                 if (should_swap) {
                     shmem_idx[lid.x] = b_idx;

scripts/make-release-summary.txt

@@ -0,0 +1,47 @@
+Take a look at the changelog between the current version and the previous version - use the `./scripts/make-release-desc.sh [current-version]` to obtain it.
+
+Write a summary of the change log in a few sections:
+
+```
+## Overview
+
+[an overview using 1 to 3 sentences (no line breaks)]
+
+### API changes (if applicable)
+
+[summarize any API changes to `/include/*`, `/tools/mtmd/mtmd.h` and `/tools/server`]
+
+### New models (if applicable)
+
+[summarize new models added to the `src/models/` directory]
+
+### Core changes (if applicable)
+
+[summarize the changes to `/src/*`
+
+### Multi-modality changes (if applicable)
+
+[summarize the changes to `/tools/mtmd/`]
+
+### Server changes (if applicable)
+
+[summarize the changes to `/tools/server/`]
+
+### UI changes (if applicable)
+
+[summarize the changes to `/tools/ui/`]
+
+### ggml changes (if applicable)
+
+[if the version has been updated, link to the respective `ggml` releases on Github, f.ex `https://github.com/ggml-org/ggml/releases/tag/v0.22.0`. for each version bump, lookup the release description and copy the summary here]
+
+```
+
+Guidelines:
+
+- All bullet point in the summary should be concise and rarely exceed a single line of 120 characters
+- Avoid repeating `ggml`-specific changes - these should already be covered by the `ggml` release links
+- Provide PR link for each bullet point where possible
+- Don't add bullet point to state that there are no API changes in some module
+
+Output just the summary in a markdown block, without any extra text.

scripts/sync-ggml.last

@@ -1 +1 @@
-8599e0ea3756c4bac4ef813af2241cb1a8bbfb0b
+36da57138425487184aa1da2eee2cde155909c6f

src/llama-grammar.cpp

@@ -172,6 +172,7 @@ static std::pair<uint32_t, const char *> parse_char(const char * src) {
             case '"':
             case '[':
             case ']':
+            case '-':
                       return std::make_pair(src[1], src + 2);
             default:
                       throw std::runtime_error(std::string("unknown escape at ") + src);