[BUG] ZeRO-3 with HuggingFace Transformers: Gathered Parameters in Forward Pass Not Released Promptly #7670
Description
Describe the bug
When training a Hugging Face Transformers model with DeepSpeed ZeRO-3 on 4 GPUs, memory profiling reveals that full-parameter tensors gathered during the forward pass are not released promptly after use. the params, once gathered in the forward pass, remain resident in GPU memory until the backward pass reaches that specific module—meaning they are held far longer than necessary and accumulate across layers during the forward pass.
To Reproduce
1.Run the code below with: deepspeed --include="localhost:0,1,2,3" "run.py"
import os
import random
import torch
import numpy as np
import psutil
import time
from pathlib import Path
from torch.utils.data import DataLoader
from transformers import GPT2LMHeadModel, GPT2Config, PreTrainedTokenizerFast
import deepspeed
import argparse
from apex.optimizers import FusedAdam
parser = argparse.ArgumentParser()
parser.add_argument("--local_rank", type=int, default=None)
args = parser.parse_args()
torch.cuda.memory._record_memory_history(max_entries=1000000)
config = GPT2Config(
vocab_size=50257,
n_embd=2560,
n_head=32,
n_layer=12,
n_positions=16*1024 + 1,
max_length=16*1024 + 1,
resid_pdrop=0.1,
embd_pdrop=0.0,
attn_pdrop=0.0,
use_cache=False,
)
model = GPT2LMHeadModel(config).to('cuda')
optimizer = FusedAdam(model.parameters(), lr=1e-4)
criterion = torch.nn.CrossEntropyLoss()
model_engine, optimizer, _, _ = deepspeed.initialize(
args=args, model=model, optimizer=optimizer, model_parameters=model.parameters(),
config="ds_config.json",
)
model_engine.train()
for i in range(6):
input = torch.randint(0, config.vocab_size,
size=(1, 16*1024),
dtype=torch.long, device='cuda')
label = torch.randint(0, config.vocab_size,
size=(1, 16*1024),
dtype=torch.long, device='cuda')
logits = model_engine(input).logits
print(f'logits.shape: {logits.shape}')
loss = criterion(logits.view(-1, config.vocab_size), label.view(-1))
print(f'loss: {loss}')
model_engine.backward(loss)
if args.local_rank==0:
snapshot_dir='./'
print(f"[INFO] Recording memory snapshot.")
torch.cuda.memory._dump_snapshot(f"{snapshot_dir}/test.pickle")
print(f"[INFO] Snapshot saved to {snapshot_dir}")
torch.cuda.memory._record_memory_history(enabled=None)use this ds_config:
{
"train_batch_size": 4,
"train_micro_batch_size_per_gpu": 1,
"zero_optimization": {
"stage": 3,
"allgather_bucket_size": 2e8,
"reduce_bucket_size": 2e8
},
"gradient_clipping": 0.0,
"prescale_gradients": false,
"fp16": {
"enabled": true,
"auto_cast": false,
"loss_scale": 1,
"initial_scale_power": 11,
"loss_scale_window": 500,
"hysteresis": 2,
"consecutive_hysteresis": false,
"min_loss_scale": 1
}
}
- Open the snap file "test.pickle" in https://docs.pytorch.org/memory_viz
Expected behavior
ZeRO-3 should release all-gathered full-parameter tensors immediately after they are used in forward.
ds_report output
[2025-11-05 15:16:35,511] [INFO] [real_accelerator.py:203:get_accelerator] Setting ds_accelerator to cuda (auto detect)
--------------------------------------------------
DeepSpeed C++/CUDA extension op report
--------------------------------------------------
NOTE: Ops not installed will be just-in-time (JIT) compiled at
runtime if needed. Op compatibility means that your system
meet the required dependencies to JIT install the op.
--------------------------------------------------
JIT compiled ops requires ninja
ninja .................. [OKAY]
--------------------------------------------------
op name ................ installed .. compatible
--------------------------------------------------
[WARNING] async_io requires the dev libaio .so object and headers but these were not found.
[WARNING] async_io: please install the libaio-dev package with apt
[WARNING] If libaio is already installed (perhaps from source), try setting the CFLAGS and LDFLAGS environment variables to where it can be found.
async_io ............... [NO] ....... [NO]
fused_adam ............. [NO] ....... [OKAY]
cpu_adam ............... [NO] ....... [OKAY]
cpu_adagrad ............ [NO] ....... [OKAY]
cpu_lion ............... [NO] ....... [OKAY]
[WARNING] Please specify the CUTLASS repo directory as environment variable $CUTLASS_PATH
evoformer_attn ......... [NO] ....... [NO]
[WARNING] FP Quantizer is using an untested triton version (3.4.0), only 2.3.(0, 1) and 3.0.0 are known to be compatible with these kernels
fp_quantizer ........... [NO] ....... [NO]
fused_lamb ............. [NO] ....... [OKAY]
fused_lion ............. [NO] ....... [OKAY]
/root/miniconda3/compiler_compat/ld: warning: librt.so.1, needed by /usr/local/cuda/lib64/libcufile.so, not found (try using -rpath or -rpath-link)
/root/miniconda3/compiler_compat/ld: warning: libpthread.so.0, needed by /usr/local/cuda/lib64/libcufile.so, not found (try using -rpath or -rpath-link)
/root/miniconda3/compiler_compat/ld: warning: libstdc++.so.6, needed by /usr/local/cuda/lib64/libcufile.so, not found (try using -rpath or -rpath-link)
/root/miniconda3/compiler_compat/ld: warning: libm.so.6, needed by /usr/local/cuda/lib64/libcufile.so, not found (try using -rpath or -rpath-link)
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::runtime_error::~runtime_error()@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `__gxx_personality_v0@CXXABI_1.3'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::ostream::tellp()@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::chrono::_V2::steady_clock::now()@GLIBCXX_3.4.19'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::string::_M_replace_aux(unsigned long, unsigned long, unsigned long, char)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `typeinfo for bool@CXXABI_1.3'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::__throw_logic_error(char const*)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `VTT for std::basic_ostringstream<char, std::char_traits<char>, std::allocator<char> >@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for std::logic_error@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::locale::~locale()@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_string<char, std::char_traits<char>, std::allocator<char> >::basic_string(std::string const&, unsigned long, unsigned long)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `__cxa_end_catch@CXXABI_1.3'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `VTT for std::basic_ofstream<char, std::char_traits<char> >@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::logic_error::~logic_error()@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for __cxxabiv1::__si_class_type_info@CXXABI_1.3'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_ios<char, std::char_traits<char> >::_M_cache_locale(std::locale const&)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `VTT for std::basic_stringstream<char, std::char_traits<char>, std::allocator<char> >@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `operator new[](unsigned long)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::string::_M_leak_hard()@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for std::basic_ifstream<char, std::char_traits<char> >@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_streambuf<wchar_t, std::char_traits<wchar_t> >::basic_streambuf(std::basic_streambuf<wchar_t, std::char_traits<wchar_t> > const&)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::string::append(char const*, unsigned long)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_string<char, std::char_traits<char>, std::allocator<char> >::basic_string(std::string const&)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `typeinfo for unsigned short@CXXABI_1.3'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::string::resize(unsigned long, char)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `typeinfo for char const*@CXXABI_1.3'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::ctype<char>::_M_widen_init() const@GLIBCXX_3.4.11'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::__throw_invalid_argument(char const*)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::locale::operator=(std::locale const&)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_ios<wchar_t, std::char_traits<wchar_t> >::_M_cache_locale(std::locale const&)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::_Rb_tree_decrement(std::_Rb_tree_node_base const*)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `__cxa_free_exception@CXXABI_1.3'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::condition_variable::notify_one()@GLIBCXX_3.4.11'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::ios_base::Init::~Init()@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_string<char, std::char_traits<char>, std::allocator<char> >::~basic_string()@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `__cxa_pure_virtual@CXXABI_1.3'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::ostream::flush()@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for __cxxabiv1::__class_type_info@CXXABI_1.3'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `__cxa_rethrow@CXXABI_1.3'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for std::basic_stringbuf<char, std::char_traits<char>, std::allocator<char> >@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_fstream<char, std::char_traits<char> >::~basic_fstream()@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::string::compare(char const*) const@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `VTT for std::basic_ostringstream<wchar_t, std::char_traits<wchar_t>, std::allocator<wchar_t> >@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::locale::locale()@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::chrono::_V2::system_clock::now()@GLIBCXX_3.4.19'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `VTT for std::basic_ifstream<char, std::char_traits<char> >@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::_Hash_bytes(void const*, unsigned long, unsigned long)@CXXABI_1.3.5'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::ostream& std::ostream::_M_insert<long long>(long long)@GLIBCXX_3.4.9'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `typeinfo for char*@CXXABI_1.3'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::__detail::_Prime_rehash_policy::_M_need_rehash(unsigned long, unsigned long, unsigned long) const@GLIBCXX_3.4.18'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for std::out_of_range@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::ostream& std::ostream::_M_insert<unsigned long>(unsigned long)@GLIBCXX_3.4.9'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::_Rb_tree_increment(std::_Rb_tree_node_base const*)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::ios_base::~ios_base()@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::range_error::~range_error()@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::__basic_file<char>::~__basic_file()@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `__cxa_guard_acquire@CXXABI_1.3'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::ostream& std::ostream::_M_insert<bool>(bool)@GLIBCXX_3.4.9'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for std::overflow_error@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `VTT for std::basic_fstream<char, std::char_traits<char> >@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for std::range_error@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for std::basic_ios<char, std::char_traits<char> >@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for std::basic_filebuf<char, std::char_traits<char> >@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `operator delete[](void*)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for std::basic_stringstream<char, std::char_traits<char>, std::allocator<char> >@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_string<char, std::char_traits<char>, std::allocator<char> >::basic_string(unsigned long, char, std::allocator<char> const&)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::__detail::_List_node_base::_M_transfer(std::__detail::_List_node_base*, std::__detail::_List_node_base*)@GLIBCXX_3.4.15'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::string::replace(unsigned long, unsigned long, char const*, unsigned long)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `typeinfo for std::exception@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_string<wchar_t, std::char_traits<wchar_t>, std::allocator<wchar_t> >::_Rep::_M_destroy(std::allocator<wchar_t> const&)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::istream& std::istream::_M_extract<double>(double&)@GLIBCXX_3.4.9'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_filebuf<char, std::char_traits<char> >::close()@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for std::basic_fstream<char, std::char_traits<char> >@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_ifstream<char, std::char_traits<char> >::basic_ifstream(char const*, std::_Ios_Openmode)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::string::append(std::string const&)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `operator new(unsigned long)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `VTT for std::basic_istringstream<wchar_t, std::char_traits<wchar_t>, std::allocator<wchar_t> >@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `typeinfo for unsigned int@CXXABI_1.3'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::string::append(char const*)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for std::domain_error@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::string::find(char, unsigned long) const@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::ostream::put(char)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `typeinfo for int@CXXABI_1.3'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::__throw_bad_alloc()@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `__cxa_thread_atexit@CXXABI_1.3.7'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `typeinfo for unsigned int*@CXXABI_1.3'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::_Rb_tree_increment(std::_Rb_tree_node_base*)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_ifstream<char, std::char_traits<char> >::~basic_ifstream()@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::ios_base::Init::Init()@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::condition_variable::condition_variable()@GLIBCXX_3.4.11'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_filebuf<char, std::char_traits<char> >::basic_filebuf()@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `VTT for std::basic_istringstream<char, std::char_traits<char>, std::allocator<char> >@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::domain_error::~domain_error()@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::cerr@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::string::find(char const*, unsigned long, unsigned long) const@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for std::basic_istringstream<char, std::char_traits<char>, std::allocator<char> >@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_string<char, std::char_traits<char>, std::allocator<char> >::basic_string(std::allocator<char> const&)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_stringbuf<char, std::char_traits<char>, std::allocator<char> >::str() const@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for std::invalid_argument@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `typeinfo for void*@CXXABI_1.3'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::string::assign(std::string const&)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_ostringstream<char, std::char_traits<char>, std::allocator<char> >::~basic_ostringstream()@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::_Rb_tree_rebalance_for_erase(std::_Rb_tree_node_base*, std::_Rb_tree_node_base&)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `typeinfo for unsigned long@CXXABI_1.3'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::__detail::_List_node_base::_M_hook(std::__detail::_List_node_base*)@GLIBCXX_3.4.15'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::__detail::_List_node_base::_M_unhook()@GLIBCXX_3.4.15'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for std::basic_ostringstream<wchar_t, std::char_traits<wchar_t>, std::allocator<wchar_t> >@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_stringbuf<char, std::char_traits<char>, std::allocator<char> >::_M_sync(char*, unsigned long, unsigned long)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_iostream<char, std::char_traits<char> >::~basic_iostream()@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::locale::locale(std::locale const&)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for std::basic_istringstream<wchar_t, std::char_traits<wchar_t>, std::allocator<wchar_t> >@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `log2f@GLIBC_2.2.5'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::ostream::operator<<(std::basic_streambuf<char, std::char_traits<char> >*)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for std::basic_streambuf<wchar_t, std::char_traits<wchar_t> >@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::exception::~exception()@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::string::_Rep::_S_create(unsigned long, unsigned long, std::allocator<char> const&)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::__basic_file<char>::is_open() const@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_istringstream<char, std::char_traits<char>, std::allocator<char> >::~basic_istringstream()@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::string::swap(std::string&)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for std::basic_ostringstream<char, std::char_traits<char>, std::allocator<char> >@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_streambuf<char, std::char_traits<char> >::basic_streambuf(std::basic_streambuf<char, std::char_traits<char> > const&)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_ios<char, std::char_traits<char> >::init(std::basic_streambuf<char, std::char_traits<char> >*)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::__throw_bad_cast()@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_ios<char, std::char_traits<char> >::clear(std::_Ios_Iostate)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_streambuf<wchar_t, std::char_traits<wchar_t> >::operator=(std::basic_streambuf<wchar_t, std::char_traits<wchar_t> > const&)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `typeinfo for long*@CXXABI_1.3'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `operator delete(void*)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::ostream::operator<<(int)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::string::_Rep::_S_empty_rep_storage@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::string::_Rep::_M_destroy(std::allocator<char> const&)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_iostream<wchar_t, std::char_traits<wchar_t> >::~basic_iostream()@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for std::runtime_error@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for std::basic_ofstream<char, std::char_traits<char> >@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::_Rb_tree_insert_and_rebalance(bool, std::_Rb_tree_node_base*, std::_Rb_tree_node_base*, std::_Rb_tree_node_base&)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_stringstream<char, std::char_traits<char>, std::allocator<char> >::~basic_stringstream()@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `VTT for std::basic_stringstream<wchar_t, std::char_traits<wchar_t>, std::allocator<wchar_t> >@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::ostream& std::ostream::_M_insert<long>(long)@GLIBCXX_3.4.9'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::istream::get()@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `typeinfo for unsigned long long@CXXABI_1.3'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_ostream<char, std::char_traits<char> >& std::operator<< <std::char_traits<char> >(std::basic_ostream<char, std::char_traits<char> >&, char const*)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::out_of_range::~out_of_range()@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::length_error::~length_error()@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_ostream<char, std::char_traits<char> >& std::__ostream_insert<char, std::char_traits<char> >(std::basic_ostream<char, std::char_traits<char> >&, char const*, long)@GLIBCXX_3.4.9'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::invalid_argument::~invalid_argument()@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_string<wchar_t, std::char_traits<wchar_t>, std::allocator<wchar_t> >::swap(std::basic_string<wchar_t, std::char_traits<wchar_t>, std::allocator<wchar_t> >&)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::cout@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::ostream& std::ostream::_M_insert<unsigned long long>(unsigned long long)@GLIBCXX_3.4.9'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::ostream& std::ostream::_M_insert<void const*>(void const*)@GLIBCXX_3.4.9'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for std::underflow_error@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for std::basic_streambuf<char, std::char_traits<char> >@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `typeinfo for std::out_of_range@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `__cxa_allocate_exception@CXXABI_1.3'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for std::basic_ios<wchar_t, std::char_traits<wchar_t> >@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `typeinfo for void const*@CXXABI_1.3'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_ios<wchar_t, std::char_traits<wchar_t> >::init(std::basic_streambuf<wchar_t, std::char_traits<wchar_t> >*)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::string::reserve(unsigned long)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `__cxa_begin_catch@CXXABI_1.3'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `typeinfo for long@CXXABI_1.3'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_string<wchar_t, std::char_traits<wchar_t>, std::allocator<wchar_t> >::_Rep::_S_empty_rep_storage@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::string::_M_leak()@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_filebuf<char, std::char_traits<char> >::open(char const*, std::_Ios_Openmode)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_stringbuf<wchar_t, std::char_traits<wchar_t>, std::allocator<wchar_t> >::_M_sync(wchar_t*, unsigned long, unsigned long)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::istream::getline(char*, long, char)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_istream<char, std::char_traits<char> >& std::getline<char, std::char_traits<char>, std::allocator<char> >(std::basic_istream<char, std::char_traits<char> >&, std::basic_string<char, std::char_traits<char>, std::allocator<char> >&, char)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for std::basic_stringstream<wchar_t, std::char_traits<wchar_t>, std::allocator<wchar_t> >@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::condition_variable::~condition_variable()@GLIBCXX_3.4.11'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for std::basic_stringbuf<wchar_t, std::char_traits<wchar_t>, std::allocator<wchar_t> >@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::string::insert(unsigned long, char const*, unsigned long)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::string::assign(char const*, unsigned long)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `typeinfo for unsigned char@CXXABI_1.3'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::ios_base::ios_base()@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::__throw_out_of_range(char const*)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::overflow_error::~overflow_error()@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::__throw_length_error(char const*)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::__throw_system_error(int)@GLIBCXX_3.4.11'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_ofstream<char, std::char_traits<char> >::close()@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::ostream& std::ostream::_M_insert<double>(double)@GLIBCXX_3.4.9'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_streambuf<char, std::char_traits<char> >::operator=(std::basic_streambuf<char, std::char_traits<char> > const&)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `typeinfo for long long@CXXABI_1.3'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_string<char, std::char_traits<char>, std::allocator<char> >::basic_string(char const*, unsigned long, std::allocator<char> const&)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_ifstream<char, std::char_traits<char> >::close()@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `__cxa_guard_release@CXXABI_1.3'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `__cxa_throw@CXXABI_1.3'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::underflow_error::~underflow_error()@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::_Rb_tree_decrement(std::_Rb_tree_node_base*)@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `vtable for std::length_error@GLIBCXX_3.4'
/root/miniconda3/compiler_compat/ld: /usr/local/cuda/lib64/libcufile.so: undefined reference to `std::basic_filebuf<char, std::char_traits<char> >::~basic_filebuf()@GLIBCXX_3.4'
collect2: error: ld returned 1 exit status
gds .................... [NO] ....... [NO]
inference_core_ops ..... [NO] ....... [OKAY]
cutlass_ops ............ [NO] ....... [OKAY]
transformer_inference .. [NO] ....... [OKAY]
quantizer .............. [NO] ....... [OKAY]
ragged_device_ops ...... [NO] ....... [OKAY]
ragged_ops ............. [NO] ....... [OKAY]
random_ltd ............. [NO] ....... [OKAY]
[WARNING] sparse_attn requires a torch version >= 1.5 and < 2.0 but detected 2.8
[WARNING] using untested triton version (3.4.0), only 1.0.0 is known to be compatible
sparse_attn ............ [NO] ....... [NO]
spatial_inference ...... [NO] ....... [OKAY]
transformer ............ [NO] ....... [OKAY]
stochastic_transformer . [NO] ....... [OKAY]
--------------------------------------------------
DeepSpeed general environment info:
torch install path ............... ['/root/miniconda3/lib/python3.12/site-packages/torch']
torch version .................... 2.8.0+cu128
deepspeed install path ........... ['/root/miniconda3/lib/python3.12/site-packages/deepspeed']
deepspeed info ................... 0.15.1, unknown, unknown
torch cuda version ............... 12.8
torch hip version ................ None
nvcc version ..................... 12.8
deepspeed wheel compiled w. ...... torch 2.8, cuda 12.8
shared memory (/dev/shm) size .... 300.00 GB
Screenshots
37.5MB = 3*2560(hiddensize)*2560(hiddensize)*2Bytes/Element, which is param of qkv_proj.
There are also tensors of 12.5MB(param of attn_out_proj) and 50MB(param of FFN Linear).
Hidden states is: 16*1024(seqlen)*2560(hiddensize)*2Bytes/Element=80MB. Similarily, 40MB is dropout mask, and 320MB is FFN Linear Output.
From the memory snapshot we can see that parameters, once gathered in the forward pass, remain resident in GPU memory until the backward pass reaches that specific module—meaning they are held far longer than necessary and accumulate across layers during the forward pass.
System info (please complete the following information):
- OS: 22.04.5 LTS (Jammy Jellyfish)
- GPU count and types: One machine with 4 H20 96GB
- Interconnects: NVLink
- Python version: 3.12.3
- Other libs:
Name: torch
Version: 2.8.0+cu128
Name: transformers
Version: 4.57.0
Name: deepspeed
Version: 0.15.1
Name: apex
Version: 0.1
Launcher context
deepspeed --include="localhost:0,1,2,3" "run.py"