Fine-tuning Llama Guard 3 1B takes much more VRAM than expected.

[jingzhaoou]

System Info

Collecting environment information...
PyTorch version: 2.6.0a0+df5bbc09d1.nv24.11
Is debug build: False
CUDA used to build PyTorch: 12.6
ROCM used to build PyTorch: N/A

OS: Ubuntu 24.04.1 LTS (x86_64)
GCC version: (Ubuntu 13.2.0-23ubuntu4) 13.2.0
Clang version: Could not collect
CMake version: version 3.31.0
Libc version: glibc-2.39

Python version: 3.12.3 (main, Sep 11 2024, 14:17:37) [GCC 13.2.0] (64-bit runtime)
Python platform: Linux-5.4.0-200-generic-x86_64-with-glibc2.39
Is CUDA available: True
CUDA runtime version: 12.6.85
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration:
GPU 0: Tesla V100-SXM2-32GB
GPU 1: Tesla V100-SXM2-32GB
GPU 2: Tesla V100-SXM2-32GB
GPU 3: Tesla V100-SXM2-32GB
GPU 4: Tesla V100-SXM2-32GB
GPU 5: Tesla V100-SXM2-32GB
GPU 6: Tesla V100-SXM2-32GB
GPU 7: Tesla V100-SXM2-32GB

Nvidia driver version: 550.90.12
cuDNN version: Probably one of the following:
/usr/lib/x86_64-linux-gnu/libcudnn.so.9.5.1
/usr/lib/x86_64-linux-gnu/libcudnn_adv.so.9.5.1
/usr/lib/x86_64-linux-gnu/libcudnn_cnn.so.9.5.1
/usr/lib/x86_64-linux-gnu/libcudnn_engines_precompiled.so.9.5.1
/usr/lib/x86_64-linux-gnu/libcudnn_engines_runtime_compiled.so.9.5.1
/usr/lib/x86_64-linux-gnu/libcudnn_graph.so.9.5.1
/usr/lib/x86_64-linux-gnu/libcudnn_heuristic.so.9.5.1
/usr/lib/x86_64-linux-gnu/libcudnn_ops.so.9.5.1
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True

CPU:
Architecture: x86_64
CPU op-mode(s): 32-bit, 64-bit
Address sizes: 46 bits physical, 48 bits virtual
Byte Order: Little Endian
CPU(s): 112
On-line CPU(s) list: 0-111
Vendor ID: GenuineIntel
Model name: Intel(R) Xeon(R) Platinum 8180M CPU @ 2.50GHz
CPU family: 6
Model: 85
Thread(s) per core: 2
Core(s) per socket: 28
Socket(s): 2
Stepping: 4
CPU(s) scaling MHz: 34%
CPU max MHz: 3800.0000
CPU min MHz: 1000.0000
BogoMIPS: 5000.00
Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid aperfmperf pni pclmulqdq dtes64 monitor ds_cpl vmx smx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid dca sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb cat_l3 cdp_l3 invpcid_single pti intel_ppin ssbd mba ibrs ibpb stibp tpr_shadow vnmi flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 hle avx2 smep bmi2 erms invpcid rtm cqm mpx rdt_a avx512f avx512dq rdseed adx smap clflushopt clwb intel_pt avx512cd avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local dtherm ida arat pln pts hwp hwp_act_window hwp_epp hwp_pkg_req pku ospke md_clear flush_l1d arch_capabilities
Virtualization: VT-x
L1d cache: 1.8 MiB (56 instances)
L1i cache: 1.8 MiB (56 instances)
L2 cache: 56 MiB (56 instances)
L3 cache: 77 MiB (2 instances)
NUMA node(s): 2
NUMA node0 CPU(s): 0-27,56-83
NUMA node1 CPU(s): 28-55,84-111
Vulnerability Gather data sampling: Mitigation; Microcode
Vulnerability Itlb multihit: KVM: Mitigation: Split huge pages
Vulnerability L1tf: Mitigation; PTE Inversion; VMX conditional cache flushes, SMT vulnerable
Vulnerability Mds: Mitigation; Clear CPU buffers; SMT vulnerable
Vulnerability Meltdown: Mitigation; PTI
Vulnerability Mmio stale data: Mitigation; Clear CPU buffers; SMT vulnerable
Vulnerability Retbleed: Mitigation; IBRS
Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl and seccomp
Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2: Mitigation; IBRS; IBPB conditional; STIBP conditional; RSB filling; PBRSB-eIBRS Not affected; BHI Not affected
Vulnerability Srbds: Not affected
Vulnerability Tsx async abort: Mitigation; Clear CPU buffers; SMT vulnerable

Versions of relevant libraries:
[pip3] mypy-extensions==1.0.0
[pip3] numpy==1.26.4
[pip3] onnx==1.17.0
[pip3] onnxruntime==1.20.1
[pip3] optree==0.13.1
[pip3] pytorch-triton==3.0.0+72734f086
[pip3] torch==2.6.0a0+df5bbc09d1.nv24.11
[pip3] torch_tensorrt==2.6.0a0
[pip3] torchao==0.7.0
[pip3] torchaudio==2.5.0a0+265bc5c
[pip3] torchaudio==2.5.0a0+265bc5c
[pip3] torchprofile==0.0.4
[pip3] torchtune==0.0.0
[pip3] torchvision==0.20.0a0
[pip3] triton==3.1.0

Information

• The official example scripts
• My own modified scripts

🐛 Describe the bug

I have a 8xV100 machine and tried to do a full fine-tune of meta-llama/Llama-Guard-3-1B. My code is like

        finetuning.main(
            model_name='meta-llama/Llama-Guard-3-1B',
            dataset="llamaguard_toxicchat_dataset",
            batch_size_training=1,
            use_peft=False,
            # None, '4bit' or '8bit'
            quantization='8bit',
            use_fp16=False,
            output_dir=output_dir,
        )

I have to change batch_size_training from 4 to 1 and set quantization = 8bit. Otherwise, if I set quantization to None, it will run out of CUDA memory. I am a bit surprised that fine-tuning a 1B model needs so much CUDA memory as each V100 has 32GB VRAM. Wonder if I miss anything here.

Also, I tried running things on a L4 GPU with 24GB VRAM. I have to set use_peft=True. No matter what configurations I tried, I have no luck doing a full fine-tune on a L4 GPU.

Error logs

torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 2.00 GiB. GPU 0 has a total capacity of 31.73 GiB of which 1.17 GiB is free. Process 2457833 has 30.56 GiB memory in use. Of the allocated memory 30.02 GiB is allocated by PyTorch, and 178.19 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation.  See documentation for Memory Management  (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables)

Expected behavior

I would expect things to pass with use_peft=True on most cases.