Public/koboldcpp
1
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-09-19 01:05:09 +02:00
Code Issues Packages Projects Releases Wiki Activity

Merge commit '52e5f0a5c180248ffae41fe74060d21a07651a93' into concedo_experimental

Browse Source 
# Conflicts:
#	examples/gen-docs/gen-docs.cpp
#	examples/lookup/lookup-create.cpp
#	examples/lookup/lookup-stats.cpp
#	examples/lookup/lookup.cpp
#	examples/speculative-simple/speculative-simple.cpp
#	examples/speculative/speculative.cpp
#	ggml/src/CMakeLists.txt
#	ggml/src/ggml-cann/aclnn_ops.cpp
#	ggml/src/ggml-cann/aclnn_ops.h
#	ggml/src/ggml-cann/ggml-cann.cpp
#	ggml/src/ggml-rpc/ggml-rpc.cpp
#	ggml/src/ggml-vulkan/ggml-vulkan.cpp
#	ggml/src/ggml-webgpu/ggml-webgpu-shader-lib.hpp
#	ggml/src/ggml-webgpu/ggml-webgpu.cpp
#	ggml/src/ggml-webgpu/wgsl-shaders/binary.wgsl
#	ggml/src/ggml-webgpu/wgsl-shaders/get_rows.wgsl
#	ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_decls.tmpl
#	ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec.wgsl
#	ggml/src/ggml-webgpu/wgsl-shaders/rms_norm_mul.wgsl
#	ggml/src/ggml-webgpu/wgsl-shaders/ssm_scan.wgsl
#	tests/test-arg-parser.cpp
#	tests/test-backend-ops.cpp
#	tests/test-chat.cpp
#	tests/test-reasoning-budget.cpp
#	tools/llama-bench/llama-bench.cpp
#	tools/rpc/rpc-server.cpp
#	tools/server/webui/src/lib/components/app/chat/ChatScreen/ChatScreen.svelte
#	tools/server/webui/src/lib/components/app/chat/ChatSidebar/ChatSidebar.svelte
#	tools/server/webui/src/routes/(chat)/+page.svelte
This commit is contained in:
Concedo
2026-04-29 22:27:36 +08:00
parent 07b928386a 52e5f0a5c1
commit 45f8ff49bb
160 changed files with 12167 additions and 8718 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-backend-reg.cpp
+12
View File
@@ -181,6 +181,12 @@ struct ggml_backend_registry {
return;
}
for (auto & entry : backends) {
if (entry.reg == reg) {
return;
}
}
#ifndef NDEBUG
GGML_LOG_DEBUG("%s: registered backend %s (%zu devices)\n",
__func__, ggml_backend_reg_name(reg), ggml_backend_reg_dev_count(reg));
@@ -192,6 +198,12 @@ struct ggml_backend_registry {
}
void register_device(ggml_backend_dev_t device) {
for (auto & dev : devices) {
if (dev == device) {
return;
}
}
#ifndef NDEBUG
GGML_LOG_DEBUG("%s: registered device %s (%s)\n", __func__, ggml_backend_dev_name(device), ggml_backend_dev_description(device));
#endif
ggml/src/ggml-vulkan/ggml-vulkan.cpp
+3
View File
@@ -13041,6 +13041,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
if (vk_perf_logger_enabled && vk_perf_logger_concurrent) {
ctx->query_node_idx[ctx->query_idx] = node_idx;
compute_ctx->s->buffer->buf.writeTimestamp(vk::PipelineStageFlagBits::eAllCommands, ctx->query_pool, ctx->query_idx++);
ggml_vk_sync_buffers(ctx, compute_ctx);
}
}
// Add all fused nodes to the unsynchronized lists.
@@ -14530,6 +14531,7 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
compute_ctx = ggml_vk_get_compute_ctx(ctx);
ctx->query_idx = 0;
compute_ctx->s->buffer->buf.writeTimestamp(vk::PipelineStageFlagBits::eAllCommands, ctx->query_pool, ctx->query_idx++);
ggml_vk_sync_buffers(ctx, compute_ctx);
}
ctx->prealloc_y_last_pipeline_used = nullptr;
@@ -14766,6 +14768,7 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
ctx->query_nodes[ctx->query_idx] = cgraph->nodes[i];
ctx->query_fusion_names[ctx->query_idx] = fusion_string;
compute_ctx->s->buffer->buf.writeTimestamp(vk::PipelineStageFlagBits::eAllCommands, ctx->query_pool, ctx->query_idx++);
ggml_vk_sync_buffers(ctx, compute_ctx);
} else {
// track a fusion string and number of fused ops for the current node_idx
ctx->query_fusion_names[i] = fusion_string;
ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq_funcs.glsl
+16 -7
View File
@@ -296,13 +296,22 @@ vec2 get_dm_scale(uint ib, uint iqs) {
const uint ib_k = ib / 8;
const uint iqs_k = (ib % 8) * 8 + iqs;
const uint is = iqs_k / 8;
u8vec2 scale_dm;
if (is < 4) {
scale_dm = u8vec2(data_a[ib_k].scales[is] & 0x3F, data_a[ib_k].scales[is + 4] & 0x3F);
} else {
scale_dm = u8vec2((data_a[ib_k].scales[is+4] & 0xF) | ((data_a[ib_k].scales[is-4] & 0xC0) >> 2),
(data_a[ib_k].scales[is+4] >> 4) | ((data_a[ib_k].scales[is ] & 0xC0) >> 2));
}
const uvec3 scales = uvec3(data_a_packed32[ib_k].scales[0],
data_a_packed32[ib_k].scales[1],
data_a_packed32[ib_k].scales[2]);
const uint scalesoffs = (is & 3) * 8;
const uint scidx0 = (is < 4) ? 0 : 2;
const uint scidxshift0 = scalesoffs;
const uint scidxshift1 = (is < 4) ? scalesoffs : scalesoffs + 2;
const uint mbidx0 = (is < 4) ? 1 : 2;
const uint mbidxshift0 = (is < 4) ? scalesoffs : scalesoffs + 4;
const uint mbidxshift1 = (is < 4) ? scalesoffs : scalesoffs + 2;
const uint8_t sc = uint8_t(((scales[scidx0] >> scidxshift0) & 0xF) | ((scales[0] >> scidxshift1) & 0x30));
const uint8_t mbyte = uint8_t(((scales[mbidx0] >> mbidxshift0) & 0xF) | ((scales[1] >> mbidxshift1) & 0x30));
u8vec2 scale_dm = u8vec2(sc, mbyte);
return FLOAT_TYPEV2(data_a_packed32[ib_k].dm) * FLOAT_TYPEV2(scale_dm);
}
ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_funcs.glsl
+26 -24
View File
@@ -201,19 +201,20 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
const vec2 loadd = vec2(data_a[ib].dm);
const uint scidx0 = (is < 4) ? is : (is + 4);
const uint scidx1 = (is < 4) ? is : (is - 4);
const uint scidxmask1 = (is < 4) ? 0x30 : 0xC0;
const uint scidxshift1 = (is < 4) ? 0 : 2;
const uint mbidx0 = is + 4;
const uint mbidx1 = (is < 4) ? is + 4 : is;
const uint mbidxmask0 = (is < 4) ? 0xF : 0xF0;
const uint mbidxshift0 = (is < 4) ? 0 : 4;
const uint mbidxmask1 = (is < 4) ? 0x30 : 0xC0;
const uint mbidxshift1 = (is < 4) ? 0 : 2;
const uvec3 scales = uvec3(data_a_packed32[ib].scales[0],
data_a_packed32[ib].scales[1],
data_a_packed32[ib].scales[2]);
const uint scalesoffs = (is & 3) * 8;
const uint8_t sc = uint8_t((data_a[ib].scales[scidx0] & 0xF) | ((data_a[ib].scales[scidx1] & scidxmask1) >> scidxshift1));
const uint8_t mbyte = uint8_t((data_a[ib].scales[mbidx0] & mbidxmask0) >> mbidxshift0 | ((data_a[ib].scales[mbidx1] & mbidxmask1) >> mbidxshift1));
const uint scidx0 = (is < 4) ? 0 : 2;
const uint scidxshift0 = scalesoffs;
const uint scidxshift1 = (is < 4) ? scalesoffs : scalesoffs + 2;
const uint mbidx0 = (is < 4) ? 1 : 2;
const uint mbidxshift0 = (is < 4) ? scalesoffs : scalesoffs + 4;
const uint mbidxshift1 = (is < 4) ? scalesoffs : scalesoffs + 2;
const uint8_t sc = uint8_t(((scales[scidx0] >> scidxshift0) & 0xF) | ((scales[0] >> scidxshift1) & 0x30));
const uint8_t mbyte = uint8_t(((scales[mbidx0] >> mbidxshift0) & 0xF) | ((scales[1] >> mbidxshift1) & 0x30));
const float d = loadd.x * sc;
const float m = -loadd.y * mbyte;
@@ -237,19 +238,20 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
const vec2 loadd = vec2(data_a[ib].dm);
const uint scidx0 = (is < 4) ? is : (is + 4);
const uint scidx1 = (is < 4) ? is : (is - 4);
const uint scidxmask1 = (is < 4) ? 0x30 : 0xC0;
const uint scidxshift1 = (is < 4) ? 0 : 2;
const uint mbidx0 = is + 4;
const uint mbidx1 = (is < 4) ? is + 4 : is;
const uint mbidxmask0 = (is < 4) ? 0xF : 0xF0;
const uint mbidxshift0 = (is < 4) ? 0 : 4;
const uint mbidxmask1 = (is < 4) ? 0x30 : 0xC0;
const uint mbidxshift1 = (is < 4) ? 0 : 2;
const uvec3 scales = uvec3(data_a_packed32[ib].scales[0],
data_a_packed32[ib].scales[1],
data_a_packed32[ib].scales[2]);
const uint scalesoffs = (is & 3) * 8;
const uint8_t sc = uint8_t((data_a[ib].scales[scidx0] & 0xF) | ((data_a[ib].scales[scidx1] & scidxmask1) >> scidxshift1));
const uint8_t mbyte = uint8_t(((data_a[ib].scales[mbidx0] & mbidxmask0) >> mbidxshift0) | ((data_a[ib].scales[mbidx1] & mbidxmask1) >> mbidxshift1));
const uint scidx0 = (is < 4) ? 0 : 2;
const uint scidxshift0 = scalesoffs;
const uint scidxshift1 = (is < 4) ? scalesoffs : scalesoffs + 2;
const uint mbidx0 = (is < 4) ? 1 : 2;
const uint mbidxshift0 = (is < 4) ? scalesoffs : scalesoffs + 4;
const uint mbidxshift1 = (is < 4) ? scalesoffs : scalesoffs + 2;
const uint8_t sc = uint8_t(((scales[scidx0] >> scidxshift0) & 0xF) | ((scales[0] >> scidxshift1) & 0x30));
const uint8_t mbyte = uint8_t(((scales[mbidx0] >> mbidxshift0) & 0xF) | ((scales[1] >> mbidxshift1) & 0x30));
const float d = loadd.x * sc;
const float m = -loadd.y * mbyte;