youkaichao

#include <cstdio>
#include <iostream>
#include <cuda_runtime.h>

__global__ void waitKernel(volatile bool *flag) {
    // Busy-wait loop
    while (!*flag) {
        // The use of volatile ensures that the GPU fetches the flag value from memory each time
        // This is necessary because without volatile, the compiler might optimize the memory read
        __threadfence_system(); // Optional for system-wide memory coherence

youkaichao

from vllm import LLM, SamplingParams
prompts = [
    "Hello, my name is",
    "The president of the United States is",
    "The capital of France is",
    "The future of AI is",
]
sampling_params = SamplingParams(temperature=0.8, top_p=0.95)

from contextlib import nullcontext

youkaichao

#define _GNU_SOURCE
#include <stdio.h>
#include <link.h>
#include <stdbool.h>
#include <string.h>
#include <stdlib.h>

typedef int cudaError_t;
typedef void* cudaGraph_t;

youkaichao

# run the code with `torchrun --nproc-per-node 4 test.py`
import os
os.environ['NCCL_DEBUG'] = 'TRACE'
import torch
import torch.distributed as dist

# nccl communicators are lazily created
dist.init_process_group(backend='nccl')
print("init done")

youkaichao

import torch
import torch.distributed as dist
import os
import multiprocessing
import multiprocessing.shared_memory
import io
import pickle

N_warmup = 10 # warmup N_warmup times
N = 100 # repeat N times

youkaichao

import torch
from torch import nn
import copy

class BackboneModel(nn.Module):
    def __init__(self, *args, **kwargs) -> None:
        super().__init__(*args, **kwargs)
        self.conv1 = nn.Conv2d(16, 16, 6)
        self.bn1 = nn.BatchNorm2d(16)
    def forward(self, x):

youkaichao

import torch
from torch import nn
import copy

from torch.fx.experimental.efficient_conv_bn_eval import turn_on_efficient_conv_bn_eval


class BackboneModel(nn.Module):
    def __init__(self, *args, **kwargs) -> None:
        super().__init__(*args, **kwargs)

youkaichao

# --------
# Hardware
# --------

# Opcode - operational code
# Assebly mnemonic - abbreviation for an operation

# Instruction Code Format (IA-32)
# - Optional instruction prefix
# - Operational code