mirrors/ollama-for-amd
1
0
Fork 0
mirror of https://github.com/likelovewant/ollama-for-amd.git synced 2025-12-23 23:18:26 +00:00
Code Issues Packages Projects Releases Wiki Activity

ggml update to b6840 (#12791)

Browse Source
This commit is contained in:
Daniel Hiltgen
2025-11-06 10:19:22 -08:00
committed by GitHub
parent c4ba257c64
commit 544b6739dd
103 changed files with 3644 additions and 1215 deletions
Download Patch File Download Diff File Expand all files Collapse all files

38
llama/patches/0022-ggml-No-alloc-mode.patch
View File

@@ -219,7 +219,7 @@ index 41eef3b5f..c81a2e48a 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 e0abde542..e98044bd8 100644
index 41ff89c4d..2931c15ca 100644
--- a/ggml/src/ggml-cuda/common.cuh
+++ b/ggml/src/ggml-cuda/common.cuh
@@ -35,6 +35,41 @@
@@ -274,7 +274,7 @@ index e0abde542..e98044bd8 100644
};
template<typename T>
@@ -999,11 +1037,11 @@ struct ggml_backend_cuda_context {
@@ -992,11 +1030,11 @@ struct ggml_backend_cuda_context {
// pool
std::unique_ptr<ggml_cuda_pool> pools[GGML_CUDA_MAX_DEVICES];
@@ -288,7 +288,7 @@ index e0abde542..e98044bd8 100644
}
return *pools[device];
}
@@ -1011,4 +1049,20 @@ struct ggml_backend_cuda_context {
@@ -1004,4 +1042,20 @@ struct ggml_backend_cuda_context {
ggml_cuda_pool & pool() {
return pool(device);
}
@@ -310,10 +310,10 @@ index e0abde542..e98044bd8 100644
+ }
};
diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
index c555cd30f..eb3db0f19 100644
index 02d413467..f79e5d65c 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() {
@@ -359,6 +359,8 @@ const ggml_cuda_device_info & ggml_cuda_info() {
// #define DEBUG_CUDA_MALLOC
@@ -322,7 +322,7 @@ index c555cd30f..eb3db0f19 100644
// buffer pool for cuda (legacy)
struct ggml_cuda_pool_leg : public ggml_cuda_pool {
static const int MAX_BUFFERS = 256;
@@ -362,9 +364,12 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
@@ -371,9 +373,12 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
ggml_cuda_buffer buffer_pool[MAX_BUFFERS] = {};
size_t pool_size = 0;
@@ -337,7 +337,7 @@ index c555cd30f..eb3db0f19 100644
}
~ggml_cuda_pool_leg() {
@@ -372,7 +377,9 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
@@ -381,7 +386,9 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
for (int i = 0; i < MAX_BUFFERS; ++i) {
ggml_cuda_buffer & b = buffer_pool[i];
if (b.ptr != nullptr) {
@@ -348,7 +348,7 @@ index c555cd30f..eb3db0f19 100644
pool_size -= b.size;
}
}
@@ -420,8 +427,15 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
@@ -429,8 +436,15 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
void * ptr;
size_t look_ahead_size = (size_t) (1.05 * size);
look_ahead_size = 256 * ((look_ahead_size + 255)/256);
@@ -366,7 +366,7 @@ index c555cd30f..eb3db0f19 100644
*actual_size = look_ahead_size;
pool_size += look_ahead_size;
#ifdef DEBUG_CUDA_MALLOC
@@ -441,10 +455,20 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
@@ -450,10 +464,20 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
}
}
GGML_LOG_DEBUG(GGML_CUDA_NAME " buffer pool full, increase MAX_CUDA_BUFFERS\n");
@@ -389,7 +389,7 @@ index c555cd30f..eb3db0f19 100644
};
// pool with virtual memory
@@ -456,18 +480,24 @@ struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
@@ -465,18 +489,24 @@ struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
CUdeviceptr pool_addr = 0;
size_t pool_used = 0;
size_t pool_size = 0;
@@ -417,7 +417,7 @@ index c555cd30f..eb3db0f19 100644
#if defined(GGML_USE_HIP)
// Workaround for https://github.com/ROCm/ROCR-Runtime/issues/285
for (std::pair<CUdeviceptr, size_t> & mapping : mappings) {
@@ -494,35 +524,49 @@ struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
@@ -503,35 +533,49 @@ struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
GGML_ASSERT(pool_size + reserve_size <= CUDA_POOL_VMM_MAX_SIZE);
@@ -493,7 +493,7 @@ index c555cd30f..eb3db0f19 100644
// add to the pool
pool_size += reserve_size;
@@ -555,16 +599,24 @@ struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
@@ -564,16 +608,24 @@ struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
// all deallocations must be in reverse order of the allocations
GGML_ASSERT(ptr == (void *) ((char *)(pool_addr) + pool_used));
}
@@ -521,7 +521,7 @@ index c555cd30f..eb3db0f19 100644
}
// destroying a cuBLAS handle while a graph is being captured in a different thread can result in a CUDA error
@@ -748,11 +800,20 @@ static ggml_backend_buffer_t ggml_backend_cuda_buffer_type_alloc_buffer(ggml_bac
@@ -757,11 +809,20 @@ static ggml_backend_buffer_t ggml_backend_cuda_buffer_type_alloc_buffer(ggml_bac
}
static size_t ggml_backend_cuda_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
@@ -543,7 +543,7 @@ index c555cd30f..eb3db0f19 100644
static size_t ggml_backend_cuda_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) {
size_t size = ggml_nbytes(tensor);
int64_t ne0 = tensor->ne[0];
@@ -776,6 +837,7 @@ static const ggml_backend_buffer_type_i ggml_backend_cuda_buffer_type_interface
@@ -785,6 +846,7 @@ static const ggml_backend_buffer_type_i ggml_backend_cuda_buffer_type_interface
/* .get_max_size = */ NULL, // defaults to SIZE_MAX
/* .get_alloc_size = */ ggml_backend_cuda_buffer_type_get_alloc_size,
/* .is_host = */ NULL,
@@ -551,7 +551,7 @@ index c555cd30f..eb3db0f19 100644
};
ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device) {
@@ -3003,6 +3065,7 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
@@ -2986,6 +3048,7 @@ 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) {
@@ -559,7 +559,7 @@ index c555cd30f..eb3db0f19 100644
// flag used to determine whether it is an integrated_gpu
const bool integrated = ggml_cuda_info().devices[cuda_ctx->device].integrated;
@@ -3018,6 +3081,11 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
@@ -3001,6 +3064,11 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
continue;
}
@@ -571,7 +571,7 @@ index c555cd30f..eb3db0f19 100644
static bool disable_fusion = (getenv("GGML_CUDA_DISABLE_FUSION") != nullptr);
if (!disable_fusion) {
@@ -3144,6 +3212,7 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
@@ -3140,6 +3208,7 @@ 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, int batch_size) {
ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
@@ -579,7 +579,7 @@ index c555cd30f..eb3db0f19 100644
ggml_cuda_set_device(cuda_ctx->device);
@@ -3223,6 +3292,71 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
@@ -3215,6 +3284,71 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
return GGML_STATUS_SUCCESS;
}
@@ -651,7 +651,7 @@ index c555cd30f..eb3db0f19 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;
@@ -3263,6 +3397,9 @@ static const ggml_backend_i ggml_backend_cuda_interface = {
@@ -3255,6 +3389,9 @@ 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,