furyhawk/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-07-21 10:15:35 +00:00
Code Issues Packages Projects Releases Wiki Activity

remove duplicated code and unnecesary refactor

Browse Source
This commit is contained in:
SamuelOliveirads
2026-06-14 16:02:02 -03:00
parent 3b1a0f88d5
commit 0d75eee35a
11 changed files with 2818 additions and 2963 deletions
Download Patch File Download Diff File Expand all files Collapse all files
examples/server/server-context.cpp
+16 -95
View File
@@ -6,6 +6,7 @@
#include "common.h"
#include "llama.h"
#include "llama-spec-features.h"
#include "log.h"
#include "sampling.h"
#include "speculative.h"
@@ -13,12 +14,8 @@
#include "mtmd-helper.h"
#include <fstream>
#include <atomic>
#include <cstring>
#include <cstdlib>
#include <iostream>
#include <regex>
#include <sstream>
#include <exception>
static void server_prompt_checkpoint_update(server_prompt_checkpoint & ckpt, llama_context * ctx, int id, int64_t n_tokens, llama_pos pos_min = -1, llama_pos pos_max = -1, int32_t offset = 0) {
@@ -49,83 +46,6 @@ static void log_text(const gpt_params & params_base, const std::string & text) {
}
}
static bool server_dflash_contract_log_enabled() {
const char * env = std::getenv("IK_DFLASH_CONTRACT_LOG");
if (env == nullptr || *env == '\0') {
return false;
}
return std::strcmp(env, "0") != 0 &&
std::strcmp(env, "false") != 0 &&
std::strcmp(env, "off") != 0;
}
static std::string server_dflash_contract_format_indices(
const std::vector<int32_t> & values,
size_t edge_count = 4) {
std::ostringstream oss;
oss << '[';
if (values.empty()) {
oss << ']';
return oss.str();
}
const size_t head = std::min(edge_count, values.size());
for (size_t i = 0; i < head; ++i) {
if (i > 0) {
oss << ',';
}
oss << values[i];
}
if (values.size() > edge_count * 2) {
oss << ",...,";
for (size_t i = values.size() - edge_count; i < values.size(); ++i) {
if (i > values.size() - edge_count) {
oss << ',';
}
oss << values[i];
}
} else {
for (size_t i = head; i < values.size(); ++i) {
oss << ',' << values[i];
}
}
oss << ']';
return oss.str();
}
static void server_dflash_contract_log_accept(
const server_slot & slot,
common_speculative_type spec_type_used,
const char * path,
bool any_rejected,
size_t n_draft,
const std::vector<llama_token> & ids,
llama_pos pos_base,
const std::vector<int32_t> & output_indices) {
if (!server_dflash_contract_log_enabled() || spec_type_used != COMMON_SPECULATIVE_TYPE_DFLASH) {
return;
}
static std::atomic<uint64_t> counter = 0;
const uint64_t ordinal = counter.fetch_add(1, std::memory_order_relaxed);
if (ordinal >= 8) {
return;
}
LLAMA_LOG_INFO("dflash contract accept[%llu]: slot=%d path=%s rejected=%s drafted=%zu accepted=%zu pos_base=%d output_indices=%s\n",
(unsigned long long) (ordinal + 1),
slot.id,
path,
any_rejected ? "true" : "false",
n_draft,
ids.size(),
(int) pos_base,
server_dflash_contract_format_indices(output_indices).c_str());
}
static bool server_slot_prompt_batch_overlaps(
const server_slot & slot,
int32_t batch_i0,
@@ -158,12 +78,13 @@ static bool server_response_needs_chat_parse(oaicompat_type oaicompat) {
oaicompat == OAICOMPAT_TYPE_RESP;
}
static bool server_speculative_has_dflash(const common_params_speculative & spec) {
return spec.has_stage_type(COMMON_SPECULATIVE_TYPE_DFLASH);
static bool server_speculative_uses_target_features(const common_params_speculative & spec) {
return spec.has_stage_type(COMMON_SPECULATIVE_TYPE_MTP) ||
spec.has_stage_type(COMMON_SPECULATIVE_TYPE_DFLASH);
}
static bool server_speculative_has_target_features(const common_params_speculative & spec) {
return spec.has_stage_type(COMMON_SPECULATIVE_TYPE_MTP) || server_speculative_has_dflash(spec);
static bool server_speculative_requires_single_slot(const common_params_speculative & spec) {
return spec.has_stage_chain();
}
static bool server_speculative_same_stage_types(
@@ -295,8 +216,8 @@ bool server_context::load_model(const gpt_params& params_) {
common_speculative_prepare_startup(params_base, false);
if (server_speculative_has_dflash(params_base.speculative) && params_base.n_parallel > 1) {
LOG_ERROR("DFlash is currently limited to a single server slot (-np 1).\n", {
if (server_speculative_requires_single_slot(params_base.speculative) && params_base.n_parallel > 1) {
LOG_ERROR("Speculative decoding is currently limited to a single server slot (-np 1).\n", {
{"n_parallel", params_base.n_parallel},
});
return false;
@@ -4127,7 +4048,7 @@ void server_context::speculative_decoding_accept() {
}
std::vector<int32_t> accepted_output_indices;
if (server_speculative_has_target_features(slot.params.speculative)) {
if (server_speculative_uses_target_features(slot.params.speculative)) {
if (!ids.empty()) {
accepted_output_indices.assign(slot.i_batch_dft.begin(), slot.i_batch_dft.begin() + ids.size());
}
@@ -4160,14 +4081,14 @@ void server_context::speculative_decoding_accept() {
const common_speculative_checkpoint * ckpt = common_speculative_get_checkpoint(slot.spec);
const bool will_restore = any_rejected && ckpt != nullptr && ckpt->valid;
if (server_speculative_has_target_features(slot.params.speculative) && !accepted_output_indices.empty()) {
server_dflash_contract_log_accept(
slot,
spec_type_used,
if (server_speculative_uses_target_features(slot.params.speculative) && !accepted_output_indices.empty()) {
llama_dflash_contract_log_accept(
slot.id,
spec_type_used == COMMON_SPECULATIVE_TYPE_DFLASH,
will_restore ? "restore" : "direct",
any_rejected,
n_draft,
ids,
ids.size(),
spec_pos_base,
accepted_output_indices);
}
@@ -4556,9 +4477,9 @@ void server_context::process_batch_tokens(int32_t & n_batch) {
continue; // continue loop of n_batch
}
if (server_speculative_has_target_features(params_base.speculative)) {
if (server_speculative_uses_target_features(params_base.speculative)) {
for (auto & slot : slots) {
if (!slot.spec || !server_speculative_has_target_features(slot.params.speculative)) {
if (!slot.spec || !server_speculative_uses_target_features(slot.params.speculative)) {
continue;
}