vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-11 09:34:37 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Merge branch 'upstream' into concedo_experimental

Browse source 
# Conflicts:
#	.github/workflows/build.yml
#	.github/workflows/release.yml
#	.github/workflows/server.yml
#	README.md
#	docs/build.md
#	docs/install.md
#	ggml/src/ggml-cpu/CMakeLists.txt
#	ggml/src/ggml-opencl/CMakeLists.txt
#	ggml/src/ggml-opencl/ggml-opencl.cpp
#	ggml/src/ggml-sycl/ggml-sycl.cpp
#	ggml/src/ggml-sycl/mmvq.cpp
#	ggml/src/ggml-sycl/vecdotq.hpp
#	tests/test-backend-ops.cpp
#	tests/test-chat.cpp
This commit is contained in:
Concedo 2025-06-05 11:03:34 +08:00
commit bc89b465a8
35 changed files with 1070 additions and 288 deletions
Download patch file Download diff file Expand all files Collapse all files

80
ggml/src/ggml-vulkan/ggml-vulkan.cpp
View file

@ -412,6 +412,7 @@ struct vk_device_struct {
vk_pipeline pipeline_count_equal_i32;
vk_pipeline pipeline_im2col_f32, pipeline_im2col_f32_f16;
vk_pipeline pipeline_timestep_embedding_f32;
vk_pipeline pipeline_conv_transpose_1d_f32;
vk_pipeline pipeline_pool2d_f32;
vk_pipeline pipeline_rwkv_wkv6_f32;
vk_pipeline pipeline_rwkv_wkv7_f32;
@ -460,7 +461,7 @@ struct vk_device_struct {
// for GGML_VK_PERF_LOGGER
std::unique_ptr<vk_perf_logger> perf_logger;
vk::QueryPool query_pool;
uint32_t num_queries;
int32_t num_queries;
~vk_device_struct() {
VK_LOG_DEBUG("destroy device " << name);
@ -722,6 +723,21 @@ struct vk_op_timestep_embedding_push_constants {
uint32_t max_period;
};
struct vk_op_conv_transpose_1d_push_constants {
uint32_t Cout;
uint32_t Cin;
uint32_t K;
uint32_t L;
uint32_t KL;
uint32_t nb01;
uint32_t nb02;
uint32_t nb11;
uint32_t nb1;
int32_t s0;
};
struct vk_op_pool2d_push_constants {
uint32_t IW; uint32_t IH;
uint32_t OW; uint32_t OH;
@ -2742,6 +2758,8 @@ static void ggml_vk_load_shaders(vk_device& device) {
ggml_vk_create_pipeline(device, device->pipeline_timestep_embedding_f32, "timestep_embedding_f32", timestep_embedding_f32_len, timestep_embedding_f32_data, "main", 2, sizeof(vk_op_timestep_embedding_push_constants), {256, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_conv_transpose_1d_f32, "conv_transpose_1d_f32", conv_transpose_1d_f32_len, conv_transpose_1d_f32_data, "main", 3, sizeof(vk_op_conv_transpose_1d_push_constants), {1, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_pool2d_f32, "pool2d_f32", pool2d_f32_len, pool2d_f32_data, "main", 2, sizeof(vk_op_pool2d_push_constants), {512, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_rwkv_wkv6_f32, "rwkv_wkv6_f32", rwkv_wkv6_f32_len, rwkv_wkv6_f32_data, "main", 7, sizeof(vk_op_rwkv_wkv6_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
@ -6416,6 +6434,11 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
return ctx->device->pipeline_timestep_embedding_f32;
}
return nullptr;
case GGML_OP_CONV_TRANSPOSE_1D:
if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
return ctx->device->pipeline_conv_transpose_1d_f32;
}
return nullptr;
case GGML_OP_POOL_2D:
if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
return ctx->device->pipeline_pool2d_f32;
@ -6750,6 +6773,10 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
uint32_t half_ceil = (dim + 1) / 2;
elements = { half_ceil, (uint32_t)src0->ne[0], 1 };
} break;
case GGML_OP_CONV_TRANSPOSE_1D:
{
elements = {uint32_t(src0->ne[1]), 1, 1}; // parallelize in {Cout, 1, 1}
} break;
case GGML_OP_POOL_2D:
{
const uint32_t N = dst->ne[3];
@ -7553,6 +7580,37 @@ static void ggml_vk_timestep_embedding(ggml_backend_vk_context * ctx, vk_context
}, dryrun);
}
static void ggml_vk_conv_transpose_1d(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
// src0: (K, Cout, Cin, 1) -- kernel
// src1: (L, Cin, 1, 1) -- input
// dst: (*, Cout, 1, 1)
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(src1->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
GGML_TENSOR_BINARY_OP_LOCALS
GGML_ASSERT(nb00 == sizeof(float));
GGML_ASSERT(nb10 == sizeof(float));
const int32_t s0 = dst->op_params[0];
vk_op_conv_transpose_1d_push_constants p{};
p.Cout = static_cast<uint32_t>(ne01);
p.Cin = static_cast<uint32_t>(ne02);
p.K = static_cast<uint32_t>(ne00);
p.L = static_cast<uint32_t>(ne10);
p.KL = static_cast<uint32_t>(ne0);
p.nb01 = static_cast<uint32_t>(nb01 / nb00);
p.nb02 = static_cast<uint32_t>(nb02 / nb00);
p.nb11 = static_cast<uint32_t>(nb11 / nb10);
p.nb1 = static_cast<uint32_t>(nb1 / nb0);
p.s0 = static_cast<uint32_t>(s0);
ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_CONV_TRANSPOSE_1D, std::move(p), dryrun);
}
static void ggml_vk_pool_2d(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
uint32_t op = static_cast<uint32_t>(dst->op_params[0]);
const int32_t k1 = dst->op_params[1];
@ -8624,6 +8682,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
case GGML_OP_COUNT_EQUAL:
case GGML_OP_IM2COL:
case GGML_OP_TIMESTEP_EMBEDDING:
case GGML_OP_CONV_TRANSPOSE_1D:
case GGML_OP_POOL_2D:
case GGML_OP_CONV_2D_DW:
case GGML_OP_RWKV_WKV6:
@ -8688,6 +8747,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
case GGML_OP_COUNT_EQUAL:
case GGML_OP_IM2COL:
case GGML_OP_TIMESTEP_EMBEDDING:
case GGML_OP_CONV_TRANSPOSE_1D:
case GGML_OP_POOL_2D:
case GGML_OP_CONV_2D_DW:
case GGML_OP_LEAKY_RELU:
@ -8859,6 +8919,10 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
case GGML_OP_TIMESTEP_EMBEDDING:
ggml_vk_timestep_embedding(ctx, compute_ctx, src0, node, dryrun);
break;
case GGML_OP_CONV_TRANSPOSE_1D:
ggml_vk_conv_transpose_1d(ctx, compute_ctx, src0, src1, node, dryrun);
break;
case GGML_OP_POOL_2D:
ggml_vk_pool_2d(ctx, compute_ctx, src0, node, dryrun);
@ -8987,6 +9051,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor *
case GGML_OP_COUNT_EQUAL:
case GGML_OP_IM2COL:
case GGML_OP_TIMESTEP_EMBEDDING:
case GGML_OP_CONV_TRANSPOSE_1D:
case GGML_OP_POOL_2D:
case GGML_OP_CONV_2D_DW:
case GGML_OP_RWKV_WKV6:
@ -9537,8 +9602,8 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
if (ctx->device->query_pool) {
ctx->device->device.destroyQueryPool(ctx->device->query_pool);
}
VkQueryPoolCreateInfo query_create_info = { VK_STRUCTURE_TYPE_QUERY_POOL_CREATE_INFO };
query_create_info.queryType = VK_QUERY_TYPE_TIMESTAMP;
vk::QueryPoolCreateInfo query_create_info;
query_create_info.queryType = vk::QueryType::eTimestamp;
query_create_info.queryCount = cgraph->n_nodes + 100;
ctx->device->query_pool = ctx->device->device.createQueryPool(query_create_info);
ctx->device->num_queries = query_create_info.queryCount;
@ -9624,7 +9689,7 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
// Get the results and pass them to the logger
std::vector<uint64_t> timestamps(cgraph->n_nodes + 1);
ctx->device->device.getQueryPoolResults(ctx->device->query_pool, 0, cgraph->n_nodes + 1, (cgraph->n_nodes + 1)*sizeof(uint64_t), timestamps.data(), sizeof(uint64_t), vk::QueryResultFlagBits::e64 | vk::QueryResultFlagBits::eWait);
VK_CHECK(ctx->device->device.getQueryPoolResults(ctx->device->query_pool, 0, cgraph->n_nodes + 1, (cgraph->n_nodes + 1)*sizeof(uint64_t), timestamps.data(), sizeof(uint64_t), vk::QueryResultFlagBits::e64 | vk::QueryResultFlagBits::eWait), "get timestamp results");
for (int i = 0; i < cgraph->n_nodes; i++) {
if (!ggml_vk_is_empty(cgraph->nodes[i])) {
ctx->device->perf_logger->log_timing(cgraph->nodes[i], uint64_t((timestamps[i+1] - timestamps[i]) * ctx->device->properties.limits.timestampPeriod));
@ -10048,6 +10113,8 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
case GGML_OP_LEAKY_RELU:
case GGML_OP_OPT_STEP_ADAMW:
return true;
case GGML_OP_CONV_TRANSPOSE_1D:
return op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32;
default:
return false;
}
@ -10539,6 +10606,11 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) {
const int32_t dim = tensor->op_params[0];
const int32_t max_period = tensor->op_params[1];
tensor_clone = ggml_timestep_embedding(ggml_ctx, src_clone[0], dim, max_period);
} else if (tensor->op == GGML_OP_CONV_TRANSPOSE_1D){
const int32_t s0 = tensor->op_params[0];
const int32_t p0 = tensor->op_params[1];
const int32_t d0 = tensor->op_params[2];
tensor_clone = ggml_conv_transpose_1d(ggml_ctx, src_clone[0], src_clone[1], s0, p0, d0);
} else if (tensor->op == GGML_OP_POOL_2D) {
enum ggml_op_pool op = static_cast<ggml_op_pool>(tensor->op_params[0]);
const int32_t k0 = tensor->op_params[1];