OLMo in-loop evals change with torch.compile() in 2.7.0 #150516
Description
🐛 Describe the bug
OLMo-core is the LLM trainer used for the OLMo series of models. It features in-loop evals that compute perplexity on held-out validation sets. With torch 2.7.0, these evals start the same as with torch 2.6.0, but start diverging at some point.
After a brief discussion on the PyTorch Slack, I have put together a self-contained repro in the OLMo-core codebase. It takes about three minutes to reproduce on one H100. Please don't be alarmed by how much code there is. OLMo-core has a lot of features, but most of it doesn't run in this example. Most of the flags needed below are just there to turn stuff off and force the trainer to just run the eval, instead of training.
To reproduce the problem:
- Check out https://github.com/allenai/OLMo-core
- Switch to the
1B-ReproForTorchbranch pip install -e .[all]- To see the bug, install torch 2.7.0 at this point. For the baseline / expected behavior, skip this step.
- Run this gnarly command:
torchrun --standalone src/scripts/train/OLMo2-1B.py train titan-baseline-5T-eval-local local --train_module.optim.compile=true --trainer.callbacks.lm_evaluator.eval_on_startup=true --trainer.load_path=s3://ai2-llm-public/checkpoints/dirkg/titan-baseline-5T/step200000 --trainer.callbacks.comet.enabled=false --trainer.hard_stop.unit=steps --trainer.hard_stop.value=200001 --trainer.callbacks.lm_evaluator.eval_interval=2 --trainer.callbacks.downstream_evaluator.enabled=false --trainer.load_strategy=always --trainer.save_folder=./runs/test --dataset.mix_base_dir=http://olmo-data.org --trainer.callbacks.lm_evaluator.eval_dataset.mix_base_dir=http://olmo-data.org - The command starts up the trainer, loads the model and data (from the internet the first time, cached after that), and performs an evaluation right away. Then runs out of memory because you can't train with these settings on a single GPU, but we don't care about that. We just care about the evaluation. It will print some lines that look like the following:
pile-validation/CE loss=2.230
pile-validation/PPL=9.296
A CE loss around 2.25 is expected. CE loss of 2.90 or worse shows the bug.
More notes:
- In the command, you can turn off compile with
--train_module.compile_model=False. - The model checkpoint this is loading was trained with torch 2.7.0. This seems to be an eval-only issue.
Error logs
No response
Versions
Collecting environment information...
PyTorch version: 2.7.0+cu128
Is debug build: False
CUDA used to build PyTorch: 12.8
ROCM used to build PyTorch: N/A
OS: Ubuntu 20.04.6 LTS (x86_64)
GCC version: (Ubuntu 9.4.0-1ubuntu1~20.04.2) 9.4.0
Clang version: Could not collect
CMake version: Could not collect
Libc version: glibc-2.31
Python version: 3.11.11 (main, Dec 11 2024, 16:28:39) [GCC 11.2.0] (64-bit runtime)
Python platform: Linux-5.15.0-135-generic-x86_64-with-glibc2.31
Is CUDA available: True
CUDA runtime version: Could not collect
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration: GPU 0: NVIDIA H100 80GB HBM3
Nvidia driver version: 570.124.06
cuDNN version: Could not collect
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
Byte Order: Little Endian
Address sizes: 46 bits physical, 57 bits virtual
CPU(s): 192
On-line CPU(s) list: 0-191
Thread(s) per core: 2
Core(s) per socket: 48
Socket(s): 2
NUMA node(s): 2
Vendor ID: GenuineIntel
CPU family: 6
Model: 143
Model name: Intel(R) Xeon(R) Platinum 8468
Stepping: 8
Frequency boost: enabled
CPU MHz: 3800.010
CPU max MHz: 2101.0000
CPU min MHz: 800.0000
BogoMIPS: 4200.00
Virtualization: VT-x
L1d cache: 4.5 MiB
L1i cache: 3 MiB
L2 cache: 192 MiB
L3 cache: 210 MiB
NUMA node0 CPU(s): 0-47,96-143
NUMA node1 CPU(s): 48-95,144-191
Vulnerability Gather data sampling: Not affected
Vulnerability Itlb multihit: Not affected
Vulnerability L1tf: Not affected
Vulnerability Mds: Not affected
Vulnerability Meltdown: Not affected
Vulnerability Mmio stale data: Not affected
Vulnerability Reg file data sampling: Not affected
Vulnerability Retbleed: Not affected
Vulnerability Spec rstack overflow: Not affected
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; Enhanced / Automatic IBRS; IBPB conditional; RSB filling; PBRSB-eIBRS SW sequence; BHI BHI_DIS_S
Vulnerability Srbds: Not affected
Vulnerability Tsx async abort: Not affected
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 tsc_known_freq 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 cat_l2 cdp_l3 invpcid_single intel_ppin cdp_l2 ssbd mba ibrs ibpb stibp ibrs_enhanced tpr_shadow vnmi flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb intel_pt avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local split_lock_detect avx_vnni avx512_bf16 wbnoinvd dtherm ida arat pln pts avx512vbmi umip pku ospke waitpkg avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg tme avx512_vpopcntdq la57 rdpid bus_lock_detect cldemote movdiri movdir64b enqcmd fsrm md_clear serialize tsxldtrk pconfig arch_lbr amx_bf16 avx512_fp16 amx_tile amx_int8 flush_l1d arch_capabilities
Versions of relevant libraries:
[pip3] numpy==1.26.4
[pip3] nvidia-cublas-cu12==12.8.3.14
[pip3] nvidia-cuda-cupti-cu12==12.8.57
[pip3] nvidia-cuda-nvrtc-cu12==12.8.61
[pip3] nvidia-cuda-runtime-cu12==12.8.57
[pip3] nvidia-cudnn-cu12==9.7.1.26
[pip3] nvidia-cufft-cu12==11.3.3.41
[pip3] nvidia-curand-cu12==10.3.9.55
[pip3] nvidia-cusolver-cu12==11.7.2.55
[pip3] nvidia-cusparse-cu12==12.5.7.53
[pip3] nvidia-cusparselt-cu12==0.6.3
[pip3] nvidia-nccl-cu12==2.26.2
[pip3] nvidia-nvjitlink-cu12==12.8.61
[pip3] nvidia-nvtx-cu12==12.8.55
[pip3] torch==2.7.0+cu128
[pip3] torchaudio==2.7.0+cu128
[pip3] torchmetrics==1.7.0
[pip3] torchvision==0.22.0+cu128
[pip3] triton==3.3.0
[conda] numpy 1.26.4 pypi_0 pypi
[conda] nvidia-cublas-cu12 12.8.3.14 pypi_0 pypi
[conda] nvidia-cuda-cupti-cu12 12.8.57 pypi_0 pypi
[conda] nvidia-cuda-nvrtc-cu12 12.8.61 pypi_0 pypi
[conda] nvidia-cuda-runtime-cu12 12.8.57 pypi_0 pypi
[conda] nvidia-cudnn-cu12 9.7.1.26 pypi_0 pypi
[conda] nvidia-cufft-cu12 11.3.3.41 pypi_0 pypi
[conda] nvidia-curand-cu12 10.3.9.55 pypi_0 pypi
[conda] nvidia-cusolver-cu12 11.7.2.55 pypi_0 pypi
[conda] nvidia-cusparse-cu12 12.5.7.53 pypi_0 pypi
[conda] nvidia-cusparselt-cu12 0.6.3 pypi_0 pypi
[conda] nvidia-nccl-cu12 2.26.2 pypi_0 pypi
[conda] nvidia-nvjitlink-cu12 12.8.61 pypi_0 pypi
[conda] nvidia-nvtx-cu12 12.8.55 pypi_0 pypi
[conda] torch 2.7.0+cu128 pypi_0 pypi
[conda] torchaudio 2.7.0+cu128 pypi_0 pypi
[conda] torchmetrics 1.7.0 pypi_0 pypi
[conda] torchvision 0.22.0+cu128 pypi_0 pypi
[conda] triton 3.3.0 pypi_0 pypi
cc @ezyang @gchanan @zou3519 @kadeng @msaroufim @chauhang @penguinwu