InternLM/internlm/utils/gputest.py

341 lines
12 KiB
Python

#!/usr/bin/env python
# -*- encoding: utf-8 -*-
import gc
import math
import socket
import torch
import torch.distributed as dist
from flash_attn.modules.mha import FlashSelfAttention, SelfAttention
from torch.utils import benchmark
from internlm.monitor import send_alert_message
from internlm.utils.logger import get_logger
from internlm.utils.megatron_timers import megatron_timer as timer
try:
import GPUtil
import psutil
except ImportError:
GPUtil, psutil = None, None
from internlm.core.context import ParallelMode
from internlm.core.context import global_context as gpc
from internlm.utils.common import get_current_device
logger = get_logger(__file__)
# Gloabl cuda cache flush counter
n_caching_allocator_flushes = 0
def empty_cache_and_diag(batch_count, interval=50):
"""empty cuda cache and run diag bench or tests."""
if interval <= 0:
interval = 50
cuda_memory_analyze(batch_count, batch_count % int(interval) == 0 or batch_count <= 5)
if batch_count % int(interval) == 0:
# there is no need to do diag on the first batch
if batch_count > 0:
if gpc.is_rank_for_log():
logger.info("Empty Cache and Diagnosis GPU/NCCL/Timer ...")
with torch.no_grad():
timer_diagnosis()
bench_gpu()
# FIXME: Runtime benchmark diagnosis can easily cause the training process
# to exit due to NCCL errors.
# bench_net()
# do empty_cache after the bench
torch.cuda.empty_cache()
# do garbage collection
gc.collect()
def benchmark_forward(
test_fn,
*inputs,
repeats=100,
amp=True,
amp_dtype=torch.float16,
**kwinputs,
):
"""Use Pytorch Benchmark on the forward pass of an arbitrary function."""
def amp_wrapper(*inputs, **kwinputs):
with torch.autocast(device_type="cuda", dtype=amp_dtype, enabled=amp):
test_fn(*inputs, **kwinputs)
bench_timer = benchmark.Timer(
stmt="test_fn_amp(*inputs, **kwinputs)",
globals={"test_fn_amp": amp_wrapper, "inputs": inputs, "kwinputs": kwinputs},
num_threads=torch.get_num_threads(),
)
used_time = bench_timer.timeit(repeats)
return used_time.mean
def flops(batch, seqlen, headdim, nheads, time_f):
"""Compute the flops value of a GPU with give flashattention function"""
flop = 4 * batch * seqlen**2 * nheads * headdim
return (flop / time_f / 10**12) if not math.isnan(time_f) else 0.0
def get_gpu_temperature():
"""Get current GPU temperature."""
try:
gpu_id = torch.cuda.current_device()
except AssertionError:
gpu_id = -1
if GPUtil is not None and gpu_id >= 0:
gpus = GPUtil.getGPUs()
gpu_temperature = gpus[gpu_id].temperature
else:
gpu_temperature = -1
return gpu_temperature
def get_cpu_temperature():
"""Get current CPU temperature."""
if psutil is not None:
cpu_temperature = psutil.sensors_temperatures()["coretemp"][0].current
else:
cpu_temperature = -1
return cpu_temperature
def timer_diagnosis():
"""Diagnosis running time"""
if len(timer.names) == 0 or len(timer.times) == 0:
return
world_size = gpc.get_world_size(ParallelMode.DATA)
if world_size < 2:
return
# if gpc.is_rank_for_log():
# logger.info("Diagnosis running timers ...")
# detect slow rank compared to other ranks in the same DP group
running_time = torch.Tensor(timer.times).to(device=get_current_device())
avg_time = running_time.detach().clone()
if world_size <= 4:
dist.all_reduce(avg_time, op=torch.distributed.ReduceOp.AVG, group=gpc.get_group(ParallelMode.DATA))
else:
running_time_max = avg_time.detach().clone()
running_time_min = avg_time.detach().clone()
dist.all_reduce(running_time_max, op=torch.distributed.ReduceOp.MAX, group=gpc.get_group(ParallelMode.DATA))
dist.all_reduce(running_time_min, op=torch.distributed.ReduceOp.MIN, group=gpc.get_group(ParallelMode.DATA))
dist.all_reduce(avg_time, op=torch.distributed.ReduceOp.SUM, group=gpc.get_group(ParallelMode.DATA))
avg_time = (avg_time - running_time_max - running_time_min) / (world_size - 2)
diag_result = running_time > avg_time * gpc.config.data.diag_outlier_ratio
diag_result = diag_result.tolist()
avg_time = avg_time.tolist()
for slow, name, time, avg in zip(diag_result, timer.names, timer.times, avg_time):
if slow is False or avg < 0.5:
continue
msg = (
f"Rank {gpc.get_local_rank(ParallelMode.GLOBAL)} is slower than avg on {name}, "
f"Hostname {socket.gethostname()}, "
f"its time {time:.2f}, avg {avg:.2f}, "
f"CPU temp {get_cpu_temperature()}, GPU temp { get_gpu_temperature()}"
)
logger.warning(msg)
send_alert_message(
address=gpc.config.monitor.alert.feishu_alert_address,
message=msg,
)
# detect slow rank compared to historical timer data
for name, time in zip(timer.names, timer.times):
if name not in timer.hist or len(timer.hist[name]) < 5:
continue
hist_avg = sum(timer.hist[name]) / len(timer.hist[name])
if time > hist_avg * gpc.config.data.diag_outlier_ratio and time > 0.5:
msg = (
f"Rank {gpc.get_local_rank(ParallelMode.GLOBAL)} is slower than hist avg on {name}, "
f"Hostname {socket.gethostname()}, "
f"its time {time:.2f}, hist_avg {hist_avg:.2f}, "
f"CPU temp {get_cpu_temperature()}, GPU temp { get_gpu_temperature()}"
)
logger.warning(msg)
send_alert_message(
address=gpc.config.monitor.alert.feishu_alert_address,
message=msg,
)
def bench_net():
"""Benchmark nccl performance for slow node detection."""
if gpc.get_world_size(ParallelMode.GLOBAL) <= 1:
return
# if gpc.is_rank_for_log():
# logger.info("benchmarking network speed ...")
repeats = 100
input_data = torch.randn(
8 * 1024 * 1024,
device=get_current_device(),
dtype=torch.bfloat16,
)
def allreduce_fn(inputs):
dist.all_reduce(inputs, op=torch.distributed.ReduceOp.AVG, group=gpc.get_group(ParallelMode.NETTEST))
bench_timer = benchmark.Timer(
stmt="test_fn_amp(inputs)",
globals={"test_fn_amp": allreduce_fn, "inputs": input_data},
num_threads=torch.get_num_threads(),
)
allreduce_time = bench_timer.timeit(repeats).mean
allreduce_time = allreduce_time * 10**3
allreduce_time_this = allreduce_time
allreduce_time = torch.Tensor([allreduce_time]).to(device=get_current_device())
dist.all_reduce(allreduce_time, group=gpc.get_group(ParallelMode.GLOBAL))
allreduce_time_avg = allreduce_time / gpc.get_world_size(ParallelMode.GLOBAL)
allreduce_time_avg = float(allreduce_time_avg.item())
if allreduce_time_this >= allreduce_time_avg * gpc.config.data.diag_outlier_ratio:
msg = (
f"Rank {gpc.get_local_rank(ParallelMode.GLOBAL)} NCCL test is slower than avg, "
f"Hostname {socket.gethostname()}, "
f"allreduce_time {allreduce_time_this:.2f}, avg {allreduce_time_avg:.2f}, "
f"CPU temp {get_cpu_temperature()}, GPU temp { get_gpu_temperature()}"
)
logger.warning(msg)
send_alert_message(
address=gpc.config.monitor.alert.feishu_alert_address,
message=msg,
)
def bench_gpu(use_flash_attn=True):
"""Benchmark single GPU performance for slow node detection."""
# if gpc.is_rank_for_log():
# logger.info("benchmarking gpu speed ...")
headdim = 64
dim = 2048
batch_size, seqlen = 2, 1024
nheads = dim // headdim
inner_attn = FlashSelfAttention if use_flash_attn else SelfAttention
inner_attn = inner_attn(causal=True, softmax_scale=None, attention_dropout=0)
qkv = torch.randn(
batch_size,
seqlen,
3,
dim // headdim,
headdim,
device=get_current_device(),
dtype=torch.float16,
requires_grad=True,
)
time_f = benchmark_forward(inner_attn, qkv)
speed = flops(batch_size, seqlen, headdim, nheads, time_f)
speed_this = speed
speed = torch.Tensor([speed]).to(device=get_current_device())
dist.all_reduce(speed, group=gpc.get_group(ParallelMode.GLOBAL))
speed_avg = speed / gpc.get_world_size(ParallelMode.GLOBAL)
speed_avg = float(speed_avg.item())
if speed_this <= speed_avg / gpc.config.data.diag_outlier_ratio:
msg = (
f"Rank {gpc.get_local_rank(ParallelMode.GLOBAL)} GPU is slower than avg, "
f"Hostname {socket.gethostname()}, "
f"tflops {speed_this:.2f}, avg {speed_avg:.2f}, "
f"CPU temp {get_cpu_temperature()}, GPU temp { get_gpu_temperature()}"
)
logger.warning(msg)
send_alert_message(
address=gpc.config.monitor.alert.feishu_alert_address,
message=msg,
)
"""
Useful utility functions migrated from deepseped.
"""
def warmup_process_group():
# Prevent OOM from nccl communication.
if dist.is_initialized():
buffer = torch.ones([64]).cuda()
if gpc.is_initialized(ParallelMode.DATA):
dist.all_reduce(buffer, group=gpc.get_group(ParallelMode.DATA))
if gpc.is_initialized(ParallelMode.TENSOR):
dist.all_reduce(buffer, group=gpc.get_group(ParallelMode.TENSOR))
if gpc.is_initialized(ParallelMode.PIPELINE):
dist.all_reduce(buffer, group=gpc.get_group(ParallelMode.PIPELINE))
if gpc.is_initialized(ParallelMode.ZERO1):
dist.all_reduce(buffer, group=gpc.get_group(ParallelMode.ZERO1))
if gpc.is_initialized(ParallelMode.MODEL):
dist.all_reduce(buffer, group=gpc.get_group(ParallelMode.MODEL))
if gpc.is_initialized(ParallelMode.ZERO3_DP):
dist.all_reduce(buffer, group=gpc.get_group(ParallelMode.ZERO3_DP))
if gpc.is_initialized(ParallelMode.EXPERT_DATA):
dist.all_reduce(buffer, group=gpc.get_group(ParallelMode.EXPERT_DATA))
if gpc.is_initialized(ParallelMode.EXPERT):
dist.all_reduce(buffer, group=gpc.get_group(ParallelMode.EXPERT))
dist.barrier()
del buffer
torch.cuda.empty_cache()
def cuda_memory_analyze(step=0, print_mm_suage=False):
global n_caching_allocator_flushes
torch.cuda.synchronize()
g_rank = gpc.get_global_rank()
tp_rank = gpc.get_local_rank(ParallelMode.TENSOR)
pp_rank = gpc.get_local_rank(ParallelMode.PIPELINE)
dp_rank = gpc.get_local_rank(ParallelMode.DATA)
rank_id = f"Rank:{g_rank}-tp{tp_rank}-pp{pp_rank}-dp{dp_rank}"
if print_mm_suage and gpc.get_local_rank(ParallelMode.DATA) == 0:
logger.info(
f"{rank_id}: Step {step}: "
f"Allocated {round(torch.cuda.memory_allocated() / (1024 * 1024 * 1024),4 )} GB, "
f"Max_Allocated {round(torch.cuda.max_memory_allocated() / (1024 * 1024 * 1024),4)} GB, "
f"Reserved {round(torch.cuda.memory_reserved()/ (1024 * 1024 * 1024),4)} GB, "
f"Max_Reserved {round(torch.cuda.max_memory_reserved()/ (1024 * 1024 * 1024),4)} GB "
)
torch.cuda.reset_peak_memory_stats()
# warn user about caching allocator flushes
memory_stats = torch.cuda.memory_stats()
alloc_retries = memory_stats.get("num_alloc_retries")
if alloc_retries is None:
alloc_retries = 0
if alloc_retries > n_caching_allocator_flushes:
retry_count = alloc_retries - n_caching_allocator_flushes
if gpc.get_global_rank() == 0:
logger.warning(
f"{rank_id}: pytorch allocator cache flushes {retry_count} times since last step."
"this happens when there is high memory pressure and is detrimental to "
"performance. if this is happening frequently consider adjusting "
"settings to reduce memory consumption. If you are unable to "
"make the cache flushes go away consider adding "
"torch.cuda.empty_cache() calls in your training loop to ensure "
"that all ranks flush their caches at the same time"
)
n_caching_allocator_flushes = alloc_retries