chengscott

systemctl daemon-reload
systemctl enable --now job@{08:00,12:00,21:00}.timer

chengscott

#include <iostream>
#include <memory>
#include <mutex>
#include <thread>

class MeanTracker {
    int total_ = 0;
    float mean_ = 0.f;

public:

chengscott

#include <iostream>

__device__ int warpInclusiveScan(int val) {
  int laneId = threadIdx.x % warpSize;
  for (int offset = 1; offset < 32; offset <<= 1) {
    int v = __shfl_up_sync(0xffffffff, val, offset);
    if (laneId >= offset) val += v;
  }
  return val;
}

chengscott

diff --git a/python/aitemplate/backend/cuda/conv2d/common.py b/python/aitemplate/backend/cuda/conv2d/common.py
index 8cf7fb2..ca13a72 100644
--- a/python/aitemplate/backend/cuda/conv2d/common.py
+++ b/python/aitemplate/backend/cuda/conv2d/common.py
@@ -501,6 +501,7 @@ def emit_instance(op):
         emiter = cutlass_lib.conv2d_operation.EmitConv2dWithBroadcastInstance()
     else:
         emiter = cutlass_lib.conv2d_operation.EmitConv2dInstance()
+    op.tile_description.stages = 2
     op_def = emiter.emit(op)

chengscott

static size_t GLOBAL_WORKSPACE_SIZE_DeviceConvFwdInstance_0 = 0;


#include <cstdio>
#include <stdexcept>

#include "cutlass/cutlass.h"

#include "cutlass/conv/kernel/default_conv2d_fprop.h"

chengscott

# Maintainer: Leo Mao <leomaoyw at gmail dot com>

pkgname=python-apex-git
_pkgname=apex
pkgver=23.08
pkgrel=1
pkgdesc="A PyTorch Extension: Tools for easy mixed precision and distributed training in Pytorch"
arch=('x86_64')
url="https://github.com/NVIDIA/apex"
_github='NVIDIA/apex'

chengscott

diff --git a/megatron/arguments.py b/megatron/arguments.py
index 4bf1d72..bcea6ce 100644
--- a/megatron/arguments.py
+++ b/megatron/arguments.py
@@ -1090,7 +1090,7 @@ def _add_distributed_args(parser):
                        default=False, help='If set, use custom-built ring exchange '
                        'for p2p communications. Note that this option will require '
                        'a custom built image that support ring-exchange p2p.')
-    group.add_argument('--local_rank', type=int, default=None,
+    group.add_argument('--local-rank', type=int, default=None,

chengscott

#!/usr/bin/env python
import os
import sys
import shlex
os.putenv('RANK', os.getenv('PMI_RANK'))
os.putenv('WORLD_SIZE', os.getenv('PMI_SIZE'))
os.putenv('LOCAL_RANK', os.getenv('PMI_LOCAL_RANK'))
os.putenv('LOCAL_SIZE', os.getenv('PMI_LOCAL_SIZE'))
argv = []
if os.getenv('PMI_RANK') == '0':

chengscott

#include <fcntl.h>
#include <stdint.h>
#include <stdio.h>
#include <stdlib.h>
#include <sys/mman.h>
#include <time.h>
#include <unistd.h>

#define BASE_SIZE 128 // array base size = 1KB
#define MAX_SIZE 17   // array size ranges from 1KB to 64MB

chengscott

[Unit]
Description=Reverse SSH connection
After=network.target

[Service]
Type=simple
ExecStart=/usr/bin/ssh -N -T -o "ServerAliveInterval 10" -o "ExitOnForwardFailure yes" -R 2222:localhost:22 chengscott@host
Restart=always
RestartSec=5s