Public/koboldcpp
1
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-09-18 16:55:14 +02:00
Code Issues Packages Projects Releases Wiki Activity

Merge commit '19f4decae0ead52debe56095ba8d693b4f14e4df' into concedo_experimental

Browse Source 
# Conflicts:
#	.github/workflows/build.yml
#	.github/workflows/release.yml
#	tests/test-backend-ops.cpp
This commit is contained in:
Concedo
2025-08-19 15:00:32 +08:00
parent 140ae92886 19f4decae0
commit a6bbd449c4
5 changed files with 104 additions and 58 deletions
Download Patch File Download Diff File Expand all files Collapse all files
convert_hf_to_gguf.py
+12 -16
View File
@@ -1334,6 +1334,12 @@ class MmprojModel(ModelBase):
return None
raise KeyError(f"could not find any of: {keys}")
def tensor_force_quant(self, name, new_name, bid, n_dims):
del bid, name, n_dims # unused
if ".patch_embd.weight" in new_name:
return gguf.GGMLQuantizationType.F16 if self.ftype == gguf.LlamaFileType.MOSTLY_F16 else gguf.GGMLQuantizationType.F32
return False
@ModelBase.register("GPTNeoXForCausalLM")
class GPTNeoXModel(TextModel):
@@ -2305,10 +2311,9 @@ class SmolVLMModel(MmprojModel):
self.gguf_writer.add_vision_use_gelu(True)
def tensor_force_quant(self, name, new_name, bid, n_dims):
del bid, new_name, n_dims # unused
if ".embeddings." in name:
return gguf.GGMLQuantizationType.F32
return False
return super().tensor_force_quant(name, new_name, bid, n_dims)
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
del bid # unused
@@ -3296,12 +3301,9 @@ class Qwen2VLVisionModel(MmprojModel):
self.gguf_writer.add_vision_attention_layernorm_eps(self.global_config.get("rms_norm_eps", 1e-6))
def tensor_force_quant(self, name, new_name, bid, n_dims):
del bid, name, n_dims # unused
if ".patch_embd." in new_name:
return gguf.GGMLQuantizationType.F16
if ".position_embd." in new_name:
return gguf.GGMLQuantizationType.F32
return False
return super().tensor_force_quant(name, new_name, bid, n_dims)
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
del bid # unused
@@ -3374,10 +3376,9 @@ class Qwen25OmniModel(Qwen2VLVisionModel):
yield ("audio_tower.embed_positions.weight", pos_embd)
def tensor_force_quant(self, name, new_name, bid, n_dims):
del bid, new_name, n_dims # unused
if ".conv" in name and ".weight" in name:
return gguf.GGMLQuantizationType.F16
return False
return super().tensor_force_quant(name, new_name, bid, n_dims)
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
if name.startswith("thinker."):
@@ -3423,12 +3424,9 @@ class InternVisionModel(MmprojModel):
self.gguf_writer.add_vision_projector_scale_factor(int(1.0 / downsample_ratio))
def tensor_force_quant(self, name, new_name, bid, n_dims):
del bid, name, n_dims # unused
if ".patch_embd." in new_name:
return gguf.GGMLQuantizationType.F16
if ".position_embd." in new_name:
return gguf.GGMLQuantizationType.F32
return False
return super().tensor_force_quant(name, new_name, bid, n_dims)
def _mapping_interns1_name(self, name):
names_map = {
@@ -5062,13 +5060,12 @@ class Gemma3VisionModel(MmprojModel):
self.gguf_writer.add_vision_projector_scale_factor(proj_scale_factor)
def tensor_force_quant(self, name, new_name, bid, n_dims):
del bid, new_name, n_dims # unused
# related to https://github.com/ggml-org/llama.cpp/issues/13025
if "input_projection" in name:
return gguf.GGMLQuantizationType.F16
if ".embeddings." in name:
return gguf.GGMLQuantizationType.F32
return False
return super().tensor_force_quant(name, new_name, bid, n_dims)
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
del bid # unused
@@ -7727,10 +7724,9 @@ class WhisperEncoderModel(MmprojModel):
self.gguf_writer.add_audio_attention_layernorm_eps(self.hparams.get("layer_norm_eps", 1e-5))
def tensor_force_quant(self, name, new_name, bid, n_dims):
del bid, new_name, n_dims # unused
if ".conv" in name and ".weight" in name:
return gguf.GGMLQuantizationType.F16
return False
return super().tensor_force_quant(name, new_name, bid, n_dims)
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
del bid # unused
ggml/src/ggml-vulkan/ggml-vulkan.cpp
+34 -13
View File
@@ -361,6 +361,9 @@ enum vk_conv_shapes {
CONV_SHAPE_COUNT,
};
static constexpr uint32_t num_argsort_pipelines = 11;
static constexpr uint32_t max_argsort_cols = 1 << (num_argsort_pipelines-1);
struct vk_device_struct {
std::recursive_mutex mutex;
@@ -477,6 +480,7 @@ struct vk_device_struct {
vk_pipeline pipeline_upscale_nearest_f32, pipeline_upscale_bilinear_f32, pipeline_upscale_bilinear_ac_f32;
vk_pipeline pipeline_scale_f32;
vk_pipeline pipeline_sqr_f32;
vk_pipeline pipeline_sqrt_f32;
vk_pipeline pipeline_sin_f32;
vk_pipeline pipeline_cos_f32;
vk_pipeline pipeline_clamp_f32;
@@ -521,7 +525,7 @@ struct vk_device_struct {
vk_pipeline pipeline_rope_neox_f32, pipeline_rope_neox_f16;
vk_pipeline pipeline_rope_multi_f32, pipeline_rope_multi_f16;
vk_pipeline pipeline_rope_vision_f32, pipeline_rope_vision_f16;
vk_pipeline pipeline_argsort_f32;
vk_pipeline pipeline_argsort_f32[num_argsort_pipelines];
vk_pipeline pipeline_sum_rows_f32;
vk_pipeline pipeline_argmax_f32;
vk_pipeline pipeline_count_equal_i32;
@@ -886,7 +890,6 @@ struct vk_op_soft_max_push_constants {
struct vk_op_argsort_push_constants {
uint32_t ncols;
uint32_t ncols_pad;
int32_t order;
};
@@ -3045,6 +3048,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
ggml_vk_create_pipeline(device, device->pipeline_scale_f32, "scale_f32", scale_f32_len, scale_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_sqr_f32, "sqr_f32", sqr_f32_len, sqr_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_sqrt_f32, "sqrt_f32", sqrt_f32_len, sqrt_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_sin_f32, "sin_f32", sin_f32_len, sin_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_cos_f32, "cos_f32", cos_f32_len, cos_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
@@ -3115,7 +3119,9 @@ static void ggml_vk_load_shaders(vk_device& device) {
ggml_vk_create_pipeline(device, device->pipeline_rope_vision_f16, "rope_vision_f16", rope_vision_f16_len, rope_vision_f16_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
}
ggml_vk_create_pipeline(device, device->pipeline_argsort_f32, "argsort_f32", argsort_f32_len, argsort_f32_data, "main", 2, sizeof(vk_op_argsort_push_constants), {1024, 1, 1}, {}, 1);
for (uint32_t i = 0; i < num_argsort_pipelines; ++i) {
ggml_vk_create_pipeline(device, device->pipeline_argsort_f32[i], "argsort_f32_"+std::to_string(i), argsort_f32_len, argsort_f32_data, "main", 2, sizeof(vk_op_argsort_push_constants), {1u<<i, 1, 1}, {1u<<i, i}, 1, true);
}
ggml_vk_create_pipeline(device, device->pipeline_argmax_f32, "argmax_f32", argmax_f32_len, argmax_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
@@ -7007,6 +7013,11 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
return ctx->device->pipeline_sqr_f32;
}
return nullptr;
case GGML_OP_SQRT:
if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
return ctx->device->pipeline_sqrt_f32;
}
return nullptr;
case GGML_OP_SIN:
if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
return ctx->device->pipeline_sin_f32;
@@ -7190,7 +7201,8 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
}
case GGML_OP_ARGSORT:
if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_I32) {
return ctx->device->pipeline_argsort_f32;
uint32_t idx = (uint32_t)ceilf(log2f(float(dst->ne[0])));
return ctx->device->pipeline_argsort_f32[idx];
}
return nullptr;
case GGML_OP_SUM:
@@ -7315,6 +7327,7 @@ static bool ggml_vk_op_supports_incontiguous(ggml_op op) {
case GGML_OP_CONCAT:
case GGML_OP_UPSCALE:
case GGML_OP_SQR:
case GGML_OP_SQRT:
case GGML_OP_SIN:
case GGML_OP_COS:
case GGML_OP_CLAMP:
@@ -7620,6 +7633,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
case GGML_OP_MUL:
case GGML_OP_SCALE:
case GGML_OP_SQR:
case GGML_OP_SQRT:
case GGML_OP_SIN:
case GGML_OP_COS:
case GGML_OP_CLAMP:
@@ -8267,6 +8281,10 @@ static void ggml_vk_sqr(ggml_backend_vk_context * ctx, vk_context& subctx, const
ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SQR, vk_op_unary_push_constants_init(src0, dst), dryrun);
}
static void ggml_vk_sqrt(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SQRT, vk_op_unary_push_constants_init(src0, dst), dryrun);
}
static void ggml_vk_sin(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SIN, vk_op_unary_push_constants_init(src0, dst), dryrun);
}
@@ -8515,16 +8533,8 @@ static void ggml_vk_argsort(ggml_backend_vk_context * ctx, vk_context& subctx, c
uint32_t ncols = src0->ne[0];
uint32_t ncols_pad = 1;
while (ncols_pad < ncols) {
ncols_pad *= 2;
}
GGML_ASSERT(ncols_pad <= 1024);
ggml_vk_op_f32<vk_op_argsort_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_ARGSORT, {
ncols,
ncols_pad,
op_params[0],
}, dryrun);
}
@@ -9730,6 +9740,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
case GGML_OP_UPSCALE:
case GGML_OP_SCALE:
case GGML_OP_SQR:
case GGML_OP_SQRT:
case GGML_OP_SIN:
case GGML_OP_COS:
case GGML_OP_CLAMP:
@@ -9799,6 +9810,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
case GGML_OP_UPSCALE:
case GGML_OP_SCALE:
case GGML_OP_SQR:
case GGML_OP_SQRT:
case GGML_OP_SIN:
case GGML_OP_COS:
case GGML_OP_CLAMP:
@@ -9900,6 +9912,10 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
case GGML_OP_SQR:
ggml_vk_sqr(ctx, compute_ctx, src0, node, dryrun);
break;
case GGML_OP_SQRT:
ggml_vk_sqrt(ctx, compute_ctx, src0, node, dryrun);
break;
case GGML_OP_SIN:
ggml_vk_sin(ctx, compute_ctx, src0, node, dryrun);
@@ -10151,6 +10167,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_cgraph *
case GGML_OP_UPSCALE:
case GGML_OP_SCALE:
case GGML_OP_SQR:
case GGML_OP_SQRT:
case GGML_OP_SIN:
case GGML_OP_COS:
case GGML_OP_CLAMP:
@@ -11390,6 +11407,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
case GGML_OP_SILU_BACK:
case GGML_OP_RMS_NORM_BACK:
case GGML_OP_SQR:
case GGML_OP_SQRT:
case GGML_OP_SIN:
case GGML_OP_COS:
case GGML_OP_CLAMP:
@@ -11397,6 +11415,8 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
case GGML_OP_OPT_STEP_ADAMW:
case GGML_OP_OPT_STEP_SGD:
return op->src[0]->type == GGML_TYPE_F32;
case GGML_OP_ARGSORT:
return op->ne[0] <= max_argsort_cols;
case GGML_OP_UPSCALE:
case GGML_OP_ACC:
case GGML_OP_CONCAT:
@@ -11406,7 +11426,6 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
case GGML_OP_DIAG_MASK_INF:
case GGML_OP_SOFT_MAX:
case GGML_OP_SOFT_MAX_BACK:
case GGML_OP_ARGSORT:
case GGML_OP_SUM:
case GGML_OP_SUM_ROWS:
case GGML_OP_ARGMAX:
@@ -11833,6 +11852,8 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
tensor_clone = ggml_scale_bias(ggml_ctx, src_clone[0], params[0], params[1]);
} else if (tensor->op == GGML_OP_SQR) {
tensor_clone = ggml_sqr(ggml_ctx, src_clone[0]);
} else if (tensor->op == GGML_OP_SQRT) {
tensor_clone = ggml_sqrt(ggml_ctx, src_clone[0]);
} else if (tensor->op == GGML_OP_SIN) {
tensor_clone = ggml_sin(ggml_ctx, src_clone[0]);
} else if (tensor->op == GGML_OP_COS) {
ggml/src/ggml-vulkan/vulkan-shaders/argsort.comp
+39 -29
View File
@@ -1,22 +1,24 @@
#version 450
#extension GL_EXT_control_flow_attributes : enable
#include "types.comp"
#define BLOCK_SIZE 1024
layout(constant_id = 0) const int BLOCK_SIZE = 1024;
layout(constant_id = 1) const int BLOCK_SIZE_LOG2 = 10;
#define ASC 0
layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
layout (binding = 1) buffer D {int data_d[];};
layout (push_constant) uniform parameter {
uint ncols;
uint ncols_pad;
uint order;
} p;
shared int dst_row[BLOCK_SIZE];
shared A_TYPE a_sh[BLOCK_SIZE];
void swap(uint idx0, uint idx1) {
int tmp = dst_row[idx0];
@@ -24,7 +26,7 @@ void swap(uint idx0, uint idx1) {
dst_row[idx1] = tmp;
}
void main() {
void argsort(bool needs_bounds_check) {
// bitonic sort
const int col = int(gl_LocalInvocationID.x);
const uint row = gl_WorkGroupID.y;
@@ -32,38 +34,46 @@ void main() {
const uint row_offset = row * p.ncols;
// initialize indices
if (col < p.ncols_pad) {
dst_row[col] = col;
}
dst_row[col] = col;
a_sh[col] = data_a[row_offset + col];
barrier();
for (uint k = 2; k <= p.ncols_pad; k *= 2) {
for (uint j = k / 2; j > 0; j /= 2) {
const uint ixj = col ^ j;
if (col < p.ncols_pad && ixj > col) {
if ((col & k) == 0) {
if (dst_row[col] >= p.ncols ||
(dst_row[ixj] < p.ncols && (p.order == ASC ?
data_a[row_offset + dst_row[col]] > data_a[row_offset + dst_row[ixj]] :
data_a[row_offset + dst_row[col]] < data_a[row_offset + dst_row[ixj]]))
) {
swap(col, ixj);
}
} else {
if (dst_row[ixj] >= p.ncols ||
(dst_row[col] < p.ncols && (p.order == ASC ?
data_a[row_offset + dst_row[col]] < data_a[row_offset + dst_row[ixj]] :
data_a[row_offset + dst_row[col]] > data_a[row_offset + dst_row[ixj]]))
) {
swap(col, ixj);
}
}
uint num_outer_loop_iters = BLOCK_SIZE_LOG2;
[[unroll]] for (uint k = 2, outer_idx = 0; outer_idx < num_outer_loop_iters; k *= 2, outer_idx++) {
uint num_inner_loop_iters = outer_idx + 1;
[[unroll]] for (uint j = k / 2, inner_idx = 0; inner_idx < num_inner_loop_iters; j /= 2, inner_idx++) {
const int ixj = int(col ^ j);
int idx_0 = (col & k) == 0 ? col : ixj;
int idx_1 = (col & k) == 0 ? ixj : col;
int sh_idx_0 = dst_row[idx_0];
int sh_idx_1 = dst_row[idx_1];
bool idx_0_oob = needs_bounds_check ? sh_idx_0 >= p.ncols : false;
bool idx_1_oob = needs_bounds_check ? sh_idx_1 >= p.ncols : false;
if ((idx_0_oob ||
(!idx_1_oob && a_sh[sh_idx_0] > a_sh[sh_idx_1])) && (ixj > col)) {
swap(idx_0, idx_1);
}
barrier();
}
}
if (col < p.ncols) {
data_d[row_offset + col] = dst_row[col];
if (p.order == ASC) {
data_d[row_offset + col] = dst_row[col];
} else {
data_d[row_offset + p.ncols - col - 1] = dst_row[col];
}
}
}
void main() {
if (p.ncols == BLOCK_SIZE) {
argsort(false);
} else {
argsort(true);
}
}
ggml/src/ggml-vulkan/vulkan-shaders/sqrt.comp
+17
View File
@@ -0,0 +1,17 @@
#version 450
#include "types.comp"
#include "generic_unary_head.comp"
layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
void main() {
const uint idx = get_idx();
if (idx >= p.ne) {
return;
}
const FLOAT_TYPE val = FLOAT_TYPE(data_a[get_aoffset() + src0_idx(idx)]);
data_d[get_doffset() + dst_idx(idx)] = D_TYPE(sqrt(val));
}
ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
+2
View File
@@ -580,6 +580,8 @@ void process_shaders() {
string_to_spv("sqr_f32", "square.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
string_to_spv("sqrt_f32", "sqrt.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
string_to_spv("sin_f32", "sin.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
string_to_spv("cos_f32", "cos.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});