vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-10 09:04:36 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Merge branch 'upstream' into concedo_experimental

Browse source 
# Conflicts:
#	.github/workflows/build.yml
#	examples/llava/clip.cpp
This commit is contained in:
Concedo 2025-04-20 17:59:58 +08:00
commit 17360a3b32
6 changed files with 527 additions and 475 deletions
Download patch file Download diff file Expand all files Collapse all files

922
examples/llava/clip.cpp
View file

File diff suppressed because it is too large Load diff

15
ggml/src/ggml-metal/ggml-metal.m
View file

@ -481,6 +481,7 @@ enum ggml_metal_kernel_type {
GGML_METAL_KERNEL_TYPE_SQRT,
GGML_METAL_KERNEL_TYPE_SIN,
GGML_METAL_KERNEL_TYPE_COS,
GGML_METAL_KERNEL_TYPE_NEG,
GGML_METAL_KERNEL_TYPE_SUM_ROWS,
GGML_METAL_KERNEL_TYPE_POOL_2D_AVG_F32,
GGML_METAL_KERNEL_TYPE_POOL_2D_MAX_F32,
@ -1159,6 +1160,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SQRT, sqrt, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SIN, sin, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_COS, cos, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_NEG, neg, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUM_ROWS, sum_rows, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGMAX, argmax, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_POOL_2D_AVG_F32, pool_2d_avg_f32, true);
@ -1320,6 +1322,7 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
case GGML_UNARY_OP_GELU_QUICK:
case GGML_UNARY_OP_SILU:
case GGML_UNARY_OP_ELU:
case GGML_UNARY_OP_NEG:
return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
default:
return false;
@ -2010,6 +2013,18 @@ static void ggml_metal_encode_node(
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
case GGML_UNARY_OP_NEG:
{
id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_NEG].pipeline;
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
const int64_t n = ggml_nelements(dst);
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
default:
{
GGML_LOG_WARN("%s: node %3d, op = %8s not implemented\n", __func__, idx, ggml_op_name(dst->op));

7
ggml/src/ggml-metal/ggml-metal.metal
View file

@ -949,6 +949,13 @@ kernel void kernel_cos(
dst[tpig] = cos(src0[tpig]);
}
kernel void kernel_neg(
device const float * src0,
device float * dst,
uint tpig[[thread_position_in_grid]]) {
dst[tpig] = -src0[tpig];
}
kernel void kernel_sum_rows(
device const float * src0,
device float * dst,

22
ggml/src/ggml-vulkan/ggml-vulkan.cpp
View file

@ -2413,7 +2413,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
ggml_vk_create_pipeline(device, device->pipeline_norm_f32, "norm_f32", norm_f32_len, norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_group_norm_f32, "group_norm_f32", group_norm_f32_len, group_norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_rms_norm_f32, "rms_norm_f32", rms_norm_f32_len, rms_norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_rms_norm_f32, "rms_norm_f32", rms_norm_f32_len, rms_norm_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {1, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_rms_norm_back_f32, "rms_norm_back_f32", rms_norm_back_f32_len, rms_norm_back_f32_data, "main", 3, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_l2_norm_f32, "l2_norm_f32", l2_norm_f32_len, l2_norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
@ -6030,6 +6030,7 @@ static bool ggml_vk_op_supports_incontiguous(ggml_op op) {
case GGML_OP_REPEAT:
case GGML_OP_REPEAT_BACK:
case GGML_OP_ROPE:
case GGML_OP_RMS_NORM:
return true;
default:
return false;
@ -6240,7 +6241,6 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
switch (op) {
case GGML_OP_NORM:
case GGML_OP_RMS_NORM:
case GGML_OP_RMS_NORM_BACK:
case GGML_OP_L2_NORM:
case GGML_OP_SOFT_MAX:
@ -6257,6 +6257,10 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
elements = { nr, 1, 1 };
}
} break;
case GGML_OP_RMS_NORM:
elements = { (uint32_t)ne01, (uint32_t)ne02, (uint32_t)ne03 };
break;
case GGML_OP_SUM:
// We use GGML_OP_SUM_ROWS with 1 row.
elements = { 1, 1, 1 };
@ -6907,7 +6911,17 @@ static void ggml_vk_group_norm(ggml_backend_vk_context * ctx, vk_context& subctx
static void ggml_vk_rms_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
float * op_params = (float *)dst->op_params;
ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_RMS_NORM, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f }, dryrun);
const uint32_t src0_type_size = ggml_type_size(src0->type);
const uint32_t dst_type_size = ggml_type_size(dst->type);
ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_RMS_NORM, {
(uint32_t)ggml_nelements(src0),
(uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
(uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] / dst_type_size, (uint32_t) dst->nb[1] / dst_type_size, (uint32_t) dst->nb[2] / dst_type_size, (uint32_t) dst->nb[3] / dst_type_size,
0,
op_params[0], 0.0f,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
}, dryrun);
}
static void ggml_vk_rms_norm_back(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
@ -9412,10 +9426,10 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
case GGML_OP_VIEW:
case GGML_OP_PERMUTE:
case GGML_OP_TRANSPOSE:
case GGML_OP_RMS_NORM:
return true;
case GGML_OP_NORM:
case GGML_OP_GROUP_NORM:
case GGML_OP_RMS_NORM:
case GGML_OP_L2_NORM:
return ggml_is_contiguous(op->src[0]);
case GGML_OP_ADD:

32
ggml/src/ggml-vulkan/vulkan-shaders/rms_norm.comp
View file

@ -1,6 +1,6 @@
#version 450
#include "generic_head.comp"
#include "generic_unary_head.comp"
#include "types.comp"
#extension GL_EXT_control_flow_attributes : enable
@ -8,19 +8,29 @@
layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
shared FLOAT_TYPE sum[BLOCK_SIZE];
void main() {
const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
const uint tid = gl_LocalInvocationID.x;
const uint ncols = p.ne00;
const uint nrows = gl_NumWorkGroups.x;
const uint nchannels = gl_NumWorkGroups.y;
const uint row = gl_WorkGroupID.x;
const uint channel = gl_WorkGroupID.y;
const uint samp = gl_WorkGroupID.z;
const uint tid = gl_LocalInvocationID.x;
const uint stride_row = p.nb01;
const uint stride_channel = p.nb02;
const uint stride_sample = p.nb03;
uint32_t a_offset = samp*stride_sample + channel*stride_channel + row*stride_row + get_aoffset();
uint32_t d_offset = ((samp*nchannels + channel)*nrows + row)*ncols + get_doffset();
sum[tid] = FLOAT_TYPE(0.0f); // partial sum for thread in warp
[[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) {
const FLOAT_TYPE xi = FLOAT_TYPE(data_a[row*p.KX + col]);
[[unroll]] for (uint col = tid; col < ncols; col += BLOCK_SIZE) {
const FLOAT_TYPE xi = FLOAT_TYPE(data_a[a_offset + col]);
sum[tid] += xi * xi;
}
@ -33,10 +43,10 @@ void main() {
barrier();
}
const FLOAT_TYPE mean = sum[0] / FLOAT_TYPE(p.KX);
const FLOAT_TYPE mean = sum[0] / FLOAT_TYPE(ncols);
const FLOAT_TYPE scale = inversesqrt(mean + FLOAT_TYPE(p.param1));
[[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) {
data_d[row*p.KX + col] = D_TYPE(scale * FLOAT_TYPE(data_a[row*p.KX + col]));
[[unroll]] for (uint col = tid; col < ncols; col += BLOCK_SIZE) {
data_d[d_offset + col] = D_TYPE(scale * FLOAT_TYPE(data_a[a_offset + col]));
}
}

4
gguf-py/pyproject.toml
View file

@ -1,6 +1,6 @@
[tool.poetry]
name = "gguf"
version = "0.16.1"
version = "0.16.2"
description = "Read and write ML models in GGUF for GGML"
authors = ["GGML <ggml@ggml.ai>"]
packages = [
@ -23,7 +23,7 @@ numpy = ">=1.17"
tqdm = ">=4.27"
pyyaml = ">=5.1"
sentencepiece = ">=0.1.98,<=0.2.0"
PySide6 = { version = "^6.9", optional = true }
PySide6 = { version = "^6.9", python = ">=3.9,<3.14", optional = true }
[tool.poetry.dev-dependencies]
pytest = "^5.2"