mirrors/ollama-for-amd
1
0
Fork 0
mirror of https://github.com/likelovewant/ollama-for-amd.git synced 2025-12-22 06:43:57 +00:00
Code Issues Packages Projects Releases Wiki Activity

Update GGML to b6646 (#12245)

Browse Source 
Notable EOLs with this change:
- MacOS v12 and v13 are no longer supported (v14+ required)
- AMD gfx900 and gfx906 are no longer supported
This commit is contained in:
Daniel Hiltgen
2025-10-02 14:47:10 -07:00
committed by GitHub
parent fdb109469f
commit c68f367ef6
326 changed files with 30615 additions and 20624 deletions
Download Patch File Download Diff File Expand all files Collapse all files

58
llama/patches/0022-ggml-No-alloc-mode.patch
View File

@@ -16,10 +16,10 @@ must be recreated with no-alloc set to false before loading data.
5 files changed, 310 insertions(+), 44 deletions(-)
diff --git a/ggml/include/ggml-backend.h b/ggml/include/ggml-backend.h
index 2773cc310..ae94887dd 100644
index 48777212..d4352663 100644
--- a/ggml/include/ggml-backend.h
+++ b/ggml/include/ggml-backend.h
@@ -291,6 +291,7 @@ extern "C" {
@@ -303,6 +303,7 @@ extern "C" {
// Initialize a backend scheduler, backends with low index are given priority over backends with high index
GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size, bool parallel, bool op_offload);
@@ -28,7 +28,7 @@ index 2773cc310..ae94887dd 100644
// Initialize backend buffers from a measure graph
diff --git a/ggml/src/ggml-backend-impl.h b/ggml/src/ggml-backend-impl.h
index c36c12d65..369e9e25a 100644
index 07784d6f..869dc07d 100644
--- a/ggml/src/ggml-backend-impl.h
+++ b/ggml/src/ggml-backend-impl.h
@@ -26,12 +26,17 @@ extern "C" {
@@ -57,10 +57,10 @@ index c36c12d65..369e9e25a 100644
};
GGML_API ggml_backend_buffer_t ggml_backend_buffer_init(
@@ -114,6 +120,16 @@ extern "C" {
void (*event_record)(ggml_backend_t backend, ggml_backend_event_t event);
// wait for an event on on a different stream
void (*event_wait) (ggml_backend_t backend, ggml_backend_event_t event);
@@ -117,6 +123,16 @@ extern "C" {
// (optional) sort/optimize the nodes in the graph
void (*graph_optimize) (ggml_backend_t backend, struct ggml_cgraph * cgraph);
+
+ // (optional) reserves intermediate buffers needed for the compution
+ // if alloc is true, memory is actually allocated, otherwise the required amount is just returned by buffer_size
@@ -75,7 +75,7 @@ index c36c12d65..369e9e25a 100644
struct ggml_backend {
diff --git a/ggml/src/ggml-backend.cpp b/ggml/src/ggml-backend.cpp
index d02a40e60..6b4dee4c7 100644
index cb2b9956..6ef5eeaf 100644
--- a/ggml/src/ggml-backend.cpp
+++ b/ggml/src/ggml-backend.cpp
@@ -41,6 +41,19 @@ ggml_backend_buffer_t ggml_backend_buft_alloc_buffer(ggml_backend_buffer_type_t
@@ -95,10 +95,10 @@ index d02a40e60..6b4dee4c7 100644
+ return buf;
+ }
+
GGML_ASSERT(buft);
return buft->iface.alloc_buffer(buft, size);
}
@@ -89,7 +102,8 @@ ggml_backend_buffer_t ggml_backend_buffer_init(
@@ -95,7 +108,8 @@ ggml_backend_buffer_t ggml_backend_buffer_init(
/* .buft = */ buft,
/* .context = */ context,
/* .size = */ size,
@@ -108,7 +108,7 @@ index d02a40e60..6b4dee4c7 100644
};
return buffer;
@@ -119,6 +133,12 @@ void * ggml_backend_buffer_get_base(ggml_backend_buffer_t buffer) {
@@ -127,6 +141,12 @@ void * ggml_backend_buffer_get_base(ggml_backend_buffer_t buffer) {
return NULL;
}
@@ -121,7 +121,7 @@ index d02a40e60..6b4dee4c7 100644
void * base = buffer->iface.get_base(buffer);
GGML_ASSERT(base != NULL && "backend buffer base cannot be NULL");
@@ -663,6 +683,12 @@ struct ggml_backend_sched {
@@ -723,6 +743,12 @@ struct ggml_backend_sched {
bool op_offload;
int debug;
@@ -134,7 +134,7 @@ index d02a40e60..6b4dee4c7 100644
};
#define hash_id(tensor) ggml_hash_find_or_insert(&sched->hash_set, tensor)
@@ -1449,6 +1475,17 @@ ggml_backend_sched_t ggml_backend_sched_new(
@@ -1606,6 +1632,17 @@ ggml_backend_sched_t ggml_backend_sched_new(
size_t graph_size,
bool parallel,
bool op_offload) {
@@ -152,7 +152,7 @@ index d02a40e60..6b4dee4c7 100644
GGML_ASSERT(n_backends > 0);
GGML_ASSERT(n_backends <= GGML_SCHED_MAX_BACKENDS);
GGML_ASSERT(ggml_backend_dev_type(ggml_backend_get_device(backends[n_backends - 1])) == GGML_BACKEND_DEVICE_TYPE_CPU);
@@ -1490,10 +1527,13 @@ ggml_backend_sched_t ggml_backend_sched_new(
@@ -1647,10 +1684,13 @@ ggml_backend_sched_t ggml_backend_sched_new(
sched->events[b][c] = ggml_backend_event_new(backends[b]->device);
}
}
@@ -166,7 +166,7 @@ index d02a40e60..6b4dee4c7 100644
ggml_backend_sched_reset(sched);
@@ -1508,6 +1548,10 @@ void ggml_backend_sched_free(ggml_backend_sched_t sched) {
@@ -1665,6 +1705,10 @@ void ggml_backend_sched_free(ggml_backend_sched_t sched) {
for (int c = 0; c < sched->n_copies; c++) {
ggml_backend_event_free(sched->events[b][c]);
}
@@ -177,7 +177,7 @@ index d02a40e60..6b4dee4c7 100644
}
ggml_gallocr_free(sched->galloc);
ggml_free(sched->ctx);
@@ -1547,6 +1591,24 @@ bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph *
@@ -1708,6 +1752,24 @@ bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph *
return false;
}
@@ -202,7 +202,7 @@ index d02a40e60..6b4dee4c7 100644
ggml_backend_sched_reset(sched);
return true;
@@ -1635,7 +1697,13 @@ size_t ggml_backend_sched_get_attempted_buffer_size(ggml_backend_sched_t sched,
@@ -1813,7 +1875,13 @@ size_t ggml_backend_sched_get_attempted_buffer_size(ggml_backend_sched_t sched,
int backend_index = ggml_backend_sched_backend_id(sched, backend);
GGML_ASSERT(backend_index >= 0 && backend_index < sched->n_backends);
@@ -218,7 +218,7 @@ index d02a40e60..6b4dee4c7 100644
void ggml_backend_sched_set_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend) {
diff --git a/ggml/src/ggml-cuda/common.cuh b/ggml/src/ggml-cuda/common.cuh
index 2e5d48797..b915ee1b8 100644
index c4246b65..448badf0 100644
--- a/ggml/src/ggml-cuda/common.cuh
+++ b/ggml/src/ggml-cuda/common.cuh
@@ -35,6 +35,31 @@
@@ -253,7 +253,7 @@ index 2e5d48797..b915ee1b8 100644
#define STRINGIZE_IMPL(...) #__VA_ARGS__
#define STRINGIZE(...) STRINGIZE_IMPL(__VA_ARGS__)
@@ -771,6 +796,9 @@ struct ggml_cuda_pool {
@@ -880,6 +905,9 @@ struct ggml_cuda_pool {
virtual void * alloc(size_t size, size_t * actual_size) = 0;
virtual void free(void * ptr, size_t size) = 0;
@@ -263,7 +263,7 @@ index 2e5d48797..b915ee1b8 100644
};
template<typename T>
@@ -914,11 +942,11 @@ struct ggml_backend_cuda_context {
@@ -1023,11 +1051,11 @@ struct ggml_backend_cuda_context {
// pool
std::unique_ptr<ggml_cuda_pool> pools[GGML_CUDA_MAX_DEVICES];
@@ -277,7 +277,7 @@ index 2e5d48797..b915ee1b8 100644
}
return *pools[device];
}
@@ -926,4 +954,20 @@ struct ggml_backend_cuda_context {
@@ -1035,4 +1063,20 @@ struct ggml_backend_cuda_context {
ggml_cuda_pool & pool() {
return pool(device);
}
@@ -299,7 +299,7 @@ index 2e5d48797..b915ee1b8 100644
+ }
};
diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
index c7f9dc3a5..d5abe09e0 100644
index e51c5035..d324bc68 100644
--- a/ggml/src/ggml-cuda/ggml-cuda.cu
+++ b/ggml/src/ggml-cuda/ggml-cuda.cu
@@ -350,6 +350,8 @@ const ggml_cuda_device_info & ggml_cuda_info() {
@@ -540,7 +540,7 @@ index c7f9dc3a5..d5abe09e0 100644
};
ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device) {
@@ -2936,6 +2998,7 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
@@ -3008,6 +3070,7 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph * cgraph,
bool & graph_evaluated_or_captured, bool & use_cuda_graph, bool & cuda_graph_update_required) {
@@ -548,7 +548,7 @@ index c7f9dc3a5..d5abe09e0 100644
// flag used to determine whether it is an integrated_gpu
const bool integrated = ggml_cuda_info().devices[cuda_ctx->device].integrated;
@@ -2951,6 +3014,11 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
@@ -3023,6 +3086,11 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
continue;
}
@@ -559,8 +559,8 @@ index c7f9dc3a5..d5abe09e0 100644
+
static bool disable_fusion = (getenv("GGML_CUDA_DISABLE_FUSION") != nullptr);
if (!disable_fusion) {
if (ggml_cuda_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL }, {})) {
@@ -3022,6 +3090,7 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
@@ -3149,6 +3217,7 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
@@ -568,7 +568,7 @@ index c7f9dc3a5..d5abe09e0 100644
ggml_cuda_set_device(cuda_ctx->device);
@@ -3101,6 +3170,71 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
@@ -3228,6 +3297,71 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
return GGML_STATUS_SUCCESS;
}
@@ -640,10 +640,10 @@ index c7f9dc3a5..d5abe09e0 100644
static void ggml_backend_cuda_event_record(ggml_backend_t backend, ggml_backend_event_t event) {
ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
@@ -3140,6 +3274,9 @@ static const ggml_backend_i ggml_backend_cuda_interface = {
/* .graph_compute = */ ggml_backend_cuda_graph_compute,
@@ -3268,6 +3402,9 @@ static const ggml_backend_i ggml_backend_cuda_interface = {
/* .event_record = */ ggml_backend_cuda_event_record,
/* .event_wait = */ ggml_backend_cuda_event_wait,
/* .graph_optimize = */ NULL,
+ /* .graph_reserve = */ ggml_backend_cuda_graph_reserve,
+ /* .buffer_size = */ ggml_backend_cuda_buffer_size,
+ /* .reset = */ ggml_backend_cuda_reset,