alexarmbr

"""
test performance and correctness of ulysses parallel attention vs single gpu attention
torchrun --nproc-per-node 2 benchmark_attn.py

using two H100s I get:
Rank 0 single gpu attention: 1698.14 ms
Rank 0 ulysses attention: 912.84 ms

running pip install para-attn should install everything needed
"""

alexarmbr

"""
test performance and correctness of ring attention vs. single gpu attention
torchrun --nproc-per-node 4 ring_attn.py
using 4 H100s I get:
Rank 0 single gpu attention: 261.78 ms
Rank 0 ring attention: 73.34 ms
"""

import os
import math

alexarmbr

Nsight Systems

add torch.cuda.cudart().cudaProfilerStart() and torch.cuda.cudart().cudaProfilerStop() where profiling should start and stop. launch profiler with

CUDA_VISIBLE_DEVICES=0,1,2,3 \
nsys profile \
-w true \
-t cuda,nvtx,osrt,cudnn,cublas \

alexarmbr

import functools
import torch
import math

def taylor_seer_approximation(WARMUP_STEPS=1, SKIP_INTERVAL_STEPS=1, compute_step_map=None, n_derivatives = 2):
    """
    A decorator that approximates the forward pass of an nn.Module to reduce computation.
    
    Args:
        warmup: Number of steps to compute the actual forward pass before starting approximation

alexarmbr

import os
from tqdm import tqdm

import torch
import torch.distributed as dist
import torch.nn.functional as F
from torch.distributed.tensor.experimental import context_parallel
from torch.distributed.tensor.experimental._attention import _cp_options
from torch.nn.attention import SDPBackend, sdpa_kernel

alexarmbr

import torch
import torch.nn.functional as F
import torch.profiler

def benchmark_forward_pass(q, k, v, num_warmup=10, num_timed_runs=20):
    """
    Benchmarks the forward pass of torch.nn.functional.scaled_dot_product_attention.
    """
    start_event = torch.cuda.Event(enable_timing=True)
    end_event = torch.cuda.Event(enable_timing=True)

alexarmbr

# torchrun --nproc-per-node 1 benchmark_sdpa.py
# torchrun --nproc-per-node 2 benchmark_sdpa.py
# torchrun --nproc-per-node 4 benchmark_sdpa.py
# torchrun --nproc-per-node 8 benchmark_sdpa.py

import torch
import torch.nn.functional as F

from torch.distributed.tensor.experimental import context_parallel
from torch.distributed.tensor.experimental._attention import _cp_options

alexarmbr

# mypy: allow-untyped-decorators
# mypy: allow-untyped-defs
import functools
import itertools
import logging
import operator
from collections import Counter, defaultdict
from typing import Any, Callable, Optional, TypeVar, Union
from typing_extensions import ParamSpec