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 branch 'upstream' into concedo_experimental

Browse Source 
# Conflicts:
#	.devops/intel.Dockerfile
#	README.md
#	docs/backend/SYCL.md
#	docs/install.md
#	docs/ops.md
#	docs/ops/SYCL.csv
#	ggml/src/ggml-sycl/binbcast.cpp
#	ggml/src/ggml-sycl/concat.cpp
#	ggml/src/ggml-sycl/dmmv.cpp
#	ggml/src/ggml-sycl/element_wise.cpp
#	ggml/src/ggml-sycl/element_wise.hpp
#	ggml/src/ggml-sycl/ggml-sycl.cpp
#	ggml/src/ggml-sycl/mmvq.cpp
#	ggml/src/ggml-sycl/mmvq.hpp
#	tests/peg-parser/test-gbnf-generation.cpp
#	tests/test-backend-ops.cpp
#	tests/test-chat.cpp
#	tools/llama-bench/llama-bench.cpp
This commit is contained in:
Concedo
2026-06-16 17:55:04 +08:00
parent 05dbd9a195 a1824902b5
commit 382ce55fb7
18 changed files with 413 additions and 180 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tools/server/bench/bench.py
+3
View File
@@ -40,6 +40,7 @@ def main(args_in: list[str] | None = None) -> None:
required=True)
parser.add_argument("--hf-repo", type=str, help="Hugging Face model repository", required=True)
parser.add_argument("--hf-file", type=str, help="Hugging Face model file", required=True)
parser.add_argument("--offline", action="store_true", default=False, help="Offline mode: forces use of cache, prevents network access")
parser.add_argument("-ngl", "--n-gpu-layers", type=int, help="layers to the GPU for computation", required=True)
parser.add_argument("--ctx-size", type=int, help="Set the size of the prompt context", required=True)
parser.add_argument("--parallel", type=int, help="Set the number of slots for process requests", required=True)
@@ -268,6 +269,8 @@ def start_server_background(args):
]
server_args.extend(['--hf-repo', args.hf_repo])
server_args.extend(['--hf-file', args.hf_file])
if args.offline:
server_args.append('--offline')
server_args.extend(['--n-gpu-layers', args.n_gpu_layers])
server_args.extend(['--ctx-size', args.ctx_size])
server_args.extend(['--parallel', args.parallel])
tools/server/server-common.cpp
-31
View File
@@ -539,37 +539,6 @@ bool server_tokens::validate(const struct llama_context * ctx) const {
return true;
}
int32_t server_tokens::process_chunk(
llama_context * ctx,
mtmd_context * mctx,
size_t idx,
llama_pos pos,
int32_t seq_id,
size_t & n_tokens_out) const {
const auto & chunk = find_chunk(idx);
const char * name = mtmd_input_chunk_get_type(chunk.get()) == MTMD_INPUT_CHUNK_TYPE_IMAGE
? "image" : "audio";
SRV_INF("processing %s...\n", name);
int32_t n_batch = llama_n_batch(ctx);
int64_t t0 = ggml_time_ms();
llama_pos new_n_past; // unused for now
int32_t result = mtmd_helper_eval_chunk_single(mctx, ctx,
chunk.get(),
pos,
seq_id,
n_batch,
true, // logits last
&new_n_past);
SRV_INF("%s processed in %" PRId64 " ms\n", name, ggml_time_ms() - t0);
if (result != 0) {
LOG_ERR("mtmd_helper_eval failed with status %d", result);
n_tokens_out = 0;
return result;
}
n_tokens_out = mtmd_input_chunk_get_n_tokens(chunk.get());
return 0;
}
server_tokens server_tokens::clone() const {
server_tokens res;
res.has_mtmd = has_mtmd;
tools/server/server-common.h
-9
View File
@@ -221,15 +221,6 @@ public:
// make sure all text tokens are within the vocab range
bool validate(const struct llama_context * ctx) const;
// encode and decode the image chunk
int32_t process_chunk(
llama_context * ctx,
mtmd_context * mctx,
size_t idx,
llama_pos pos,
int32_t seq_id,
size_t & n_tokens_out) const;
server_tokens clone() const;
};
tools/server/server-context.cpp
+54 -80
View File
@@ -15,11 +15,6 @@
#include "mtmd.h"
#include "mtmd-helper.h"
#include "ggml-cpp.h"
// TODO: tmp until the mtmd draft processing is refactored [TAG_MTMD_DRAFT_PROCESSING]
#include "../../src/llama-ext.h"
#include <algorithm>
#include <cstddef>
#include <cinttypes>
@@ -81,7 +76,6 @@ struct server_slot {
// multimodal
mtmd_context * mctx = nullptr;
mtmd::batch_ptr mbatch = nullptr;
std::array<llama_context *, 2> mtgt = {nullptr, nullptr}; // [0] for main context, [1] for optional draft context
// speculative decoding
common_speculative * spec;
@@ -207,6 +201,8 @@ struct server_slot {
// Speculative decoding stats
int32_t n_draft_total = 0; // Total draft tokens generated
int32_t n_draft_accepted = 0; // Draft tokens actually accepted
int32_t n_draft_verif_steps = 0; // Total draft token verification steps by the target model
std::vector<int32_t> n_accepted_per_pos; // Accepted tokens per draft position
void reset() {
SLT_DBG(*this, "%s", "\n");
@@ -233,6 +229,8 @@ struct server_slot {
// clear speculative decoding stats
n_draft_total = 0;
n_draft_accepted = 0;
n_draft_verif_steps = 0;
n_accepted_per_pos.clear();
task_prev = std::move(task);
task.reset();
@@ -244,15 +242,6 @@ struct server_slot {
// clear multimodal state
mbatch.reset();
mtgt[0] = ctx_tgt;
mtgt[1] = nullptr;
if (ctx_dft && llama_get_ctx_other(ctx_dft) != ctx_tgt) {
// TODO: in the future, figure out how to infuse target embeddings to the images
// for now, we re-decode the same chunk in both ctx_tgt and ctx_dft
// maybe we simply need to call `common_speculative_process()` ?
// [TAG_MTMD_DRAFT_PROCESSING]
mtgt[1] = ctx_dft;
}
}
void init_sampler() const {
@@ -524,10 +513,22 @@ struct server_slot {
llama_perf_context(ctx_tgt).n_reused);
if (n_draft_total > 0) {
const float draft_ratio = (float) n_draft_accepted / n_draft_total;
const float draft_ratio = (float) n_draft_accepted / n_draft_total;
const double mean_acc_len = n_draft_verif_steps > 0 ? 1.0 + (double) n_draft_accepted / (double) n_draft_verif_steps : 1.0;
std::string acceptance_rates_per_pos;
if (n_draft_verif_steps > 0) {
for (size_t i = 0; i < n_accepted_per_pos.size(); ++i) {
if (i > 0) {
acceptance_rates_per_pos += ", ";
}
acceptance_rates_per_pos += string_format("%.3f", (double) n_accepted_per_pos[i] / (double) n_draft_verif_steps);
}
}
SLT_INF(*this,
"draft acceptance = %0.5f (%5d accepted / %5d generated)\n",
draft_ratio, n_draft_accepted, n_draft_total);
"draft acceptance = %0.5f (%5d accepted / %5d generated), mean acceptance length = %5.2f, acceptance rate per position = (%s)\n",
draft_ratio, n_draft_accepted, n_draft_total, mean_acc_len, acceptance_rates_per_pos.c_str());
}
common_speculative_print_stats(spec);
@@ -598,32 +599,38 @@ struct server_slot {
int process_mtmd_chunk(size_t idx, size_t & n_tokens_out) {
GGML_ASSERT(mctx);
const auto & input_tokens = task->tokens;
auto & chunk = input_tokens.find_chunk(idx);
const auto & chunk = input_tokens.find_chunk(idx);
int32_t res = 0;
auto try_decode = [&]() -> int32_t {
if (mbatch) {
float * embd = mtmd_batch_get_output_embd(mbatch.get(), chunk.get());
if (embd) {
for (auto * lctx : mtgt) {
if (lctx == nullptr) {
continue;
}
llama_pos new_n_past; // unused for now
res = mtmd_helper_decode_image_chunk(
mctx,
lctx,
chunk.get(),
embd,
prompt.tokens.pos_next(),
id,
llama_n_batch(lctx),
&new_n_past
);
if (res != 0) {
SLT_ERR(*this, "failed to decode mtmd chunk, idx = %zu, res = %d\n", idx, res);
return -1;
void * cb_data = spec;
static auto cb = [](llama_batch batch, void * user_data) {
common_speculative * spec = static_cast<common_speculative *>(user_data);
if (!common_speculative_process(spec, batch)) {
return 1;
}
return 0;
};
llama_pos new_n_past; // unused for now
res = mtmd_helper_decode_image_chunk(
mctx,
ctx_tgt,
chunk.get(),
embd,
prompt.tokens.pos_next(),
id,
llama_n_batch(ctx_tgt),
&new_n_past,
cb,
cb_data
);
if (res != 0) {
SLT_ERR(*this, "failed to decode mtmd chunk, idx = %zu, res = %d\n", idx, res);
return -1;
}
n_tokens_out = mtmd_input_chunk_get_n_tokens(chunk.get());
return 0; // success
@@ -636,7 +643,8 @@ struct server_slot {
res = try_decode();
if (res == 0) {
return 0;
} else if (res < 0) {
}
if (res < 0) {
// fatal error
return res;
}
@@ -3350,48 +3358,6 @@ private:
// TODO: avoid restoring the draft context and re-evaluating the drafted tokens when not needed [TAG_SPEC_AVOID_DRAFT_REEVAL]
// for now, always re-evaluate for simplicity
// ref: https://github.com/ggml-org/llama.cpp/pull/22728#issuecomment-4400925384
//
// | spec type | need re-eval |
// | --- | --- |
// | draft model | no | because the draft model does not use embeddings from the target
// | MTP (std) | yes |
// | MTP Gemma4 | no | because the KV cache is shared
// | Eagle3 | yes |
// | DFlash | yes | https://github.com/ggml-org/llama.cpp/pull/22728#issuecomment-4405406982
//
// note: this logic is now moved in `common_speculative_process()`
// keeping the sketch here until for a bit, until the logic is finalized
//
//if (ctx_dft) {
// // TODO: update as needed for MTP, Eagle3, etc.
// const bool need_tgt_embd = false;
// if (need_tgt_embd) {
// llama_synchronize(ctx_tgt);
// }
// // the logic here varies depending on the speculative decoding method
// // - some draft contexts require embeddings from the target context, others don't
// // - some draft contexts involve an encoder step to transform the target embeddings to draft embeddings
// // TODO: extract this in a function ?
// {
// // TODO: hook the embeddings from the last target batch here
// if (llama_model_has_encoder(model_dft.get())) {
// //llama_encode(ctx_dft, ...);
// GGML_ABORT("not implemented yet\n");
// }
// const int ret = llama_decode(ctx_dft.get(), batch_view);
// if (ret != 0) {
// SRV_ERR("failed to decode draft batch, ret = %d\n", ret);
// // TODO: handle error
// break;
// }
// }
//}
if (!common_speculative_process(spec.get(), batch_view)) {
SRV_ERR("%s", "failed to process speculative batch\n");
@@ -3593,6 +3559,14 @@ private:
// update how many tokens out of those tested were accepted
slot.n_draft_accepted += ids.size() - 1;
slot.n_draft_verif_steps += 1;
if (slot.n_accepted_per_pos.empty()) {
slot.n_accepted_per_pos.resize(common_speculative_n_max(&params_base.speculative), 0);
}
for (size_t i = 0; i < ids.size() - 1 && i < slot.n_accepted_per_pos.size(); ++i) {
slot.n_accepted_per_pos[i]++;
}
// add accepted tokens to the prompt
slot.prompt.tokens.keep_first(slot.prompt.n_tokens() - n_draft);