Files
yyj e80614fc2d [feat](kt-sft): Activation reuse & Int8 kernel refactor & native block-FP8 LoRA SFT (#2141)
* feat(sft): support distributed activation reuse policies

* feat(sft): add frozen-base INT8 LoRA training

* fix(sft): make INT8 expert LoRA rank-zero authoritative

* fix(sft): preserve DeepSeek router LoRA routing

* feat(sft): enable persistent INT8 LoRA training

* perf(sft): accelerate INT8 VNNI with oneDNN BRGEMM

* perf(int8): fuse oneDNN compensation into backward repack

* [feat]: support BF16 expert LoRA training

* [fix]: honor forwarded activation policy in SFT workers

* feat(sft): add native block-FP8 routed expert LoRA

* feat(sft): expose explicit expert placeholder ownership

* fix(sft): publish fused adapter artifacts atomically

* feat(sft): own artifact and adapter lifecycle contracts

* fix(sft): harden artifact and rank-local contracts

* fix(sft): auto-adapt owner before adapter restore

* style(sft): keep lifecycle comments concise

* fix(sft): require fused adapter manifests

* test(sft): use spawn for distributed workers

* fix(sft): preserve runtime checkpoint metadata

* fix(sft): validate wrapped runtime configuration

* fix(sft): preserve expert format provenance

* fix(sft): own routed experts during device dispatch

* test(sft): lock explicit quantization conflict

* fix(cpu): make shared memory buffers lifetime-safe

* release: prepare v0.7.0
2026-08-17 16:25:14 +08:00

100 lines
2.6 KiB
C++

/**
* @Description :
* @Author : chenht2022
* @Date : 2024-07-17 12:25:51
* @Version : 1.0.0
* @LastEditors : chenht2022
* @LastEditTime : 2024-10-09 11:08:10
* @Copyright (c) 2024 by KVCache.AI, All Rights Reserved.
**/
#include "task_queue.h"
#include <pthread.h>
#include <sched.h>
#include <chrono>
#include <iostream>
#include <thread>
TaskQueue::TaskQueue() : done(false), pending(0) {
Node* dummy = new Node();
head.store(dummy, std::memory_order_relaxed);
tail.store(dummy, std::memory_order_relaxed);
workerThread = std::thread(&TaskQueue::worker, this);
}
TaskQueue::~TaskQueue() {
{
std::lock_guard<std::mutex> lock(mtx);
done.store(true, std::memory_order_release);
}
cv.notify_all();
if (workerThread.joinable()) workerThread.join();
Node* node = head.load(std::memory_order_relaxed);
while (node) {
Node* next = node->next.load(std::memory_order_relaxed);
delete node;
node = next;
}
}
void TaskQueue::enqueue(std::function<void()> task) {
pending.fetch_add(1, std::memory_order_acq_rel);
Node* node = new Node(task);
Node* prev = tail.exchange(node, std::memory_order_acq_rel);
prev->next.store(node, std::memory_order_release);
{
std::lock_guard<std::mutex> lock(mtx);
}
cv.notify_one();
}
void TaskQueue::sync(size_t allow_n_pending) {
std::exception_ptr task_exception;
{
std::unique_lock<std::mutex> lock(mtx);
cv.wait(lock, [&] {
return pending.load(std::memory_order_acquire) <= allow_n_pending
|| done.load(std::memory_order_acquire);
});
task_exception = first_exception;
first_exception = nullptr;
}
if (task_exception) std::rethrow_exception(task_exception);
}
void TaskQueue::worker() {
Node* curr = head.load(std::memory_order_relaxed);
while (!done.load(std::memory_order_acquire)) {
Node* next = curr->next.load(std::memory_order_acquire);
if (next) {
std::exception_ptr task_exception;
if (next->task) {
try {
next->task();
} catch (...) {
task_exception = std::current_exception();
}
}
delete curr;
curr = next;
head.store(curr, std::memory_order_release);
{
std::lock_guard<std::mutex> lock(mtx);
if (task_exception && !first_exception) {
first_exception = task_exception;
}
pending.fetch_sub(1, std::memory_order_acq_rel);
}
cv.notify_all();
} else {
std::unique_lock<std::mutex> lock(mtx);
cv.wait(lock, [&] {
return curr->next.load(std::memory_order_acquire) != nullptr
|| done.load(std::memory_order_acquire);
});
}
}
}