[Bug]: v0.8.0 Causes OOM and Performance Degradation for Deepseek R1

[linyu09-oss]

⚙️ Your current environment

The output of python collect_env.py

### Environment Information ###
Operating System: `Linux-5.4.250-4-velinux1u1-amd64-x86_64-with-glibc2.35`
Python Version: `3.11.13 (main, Jun  5 2025, 13:12:00) [GCC 11.2.0]`
llm-compressor Version: `0.7.1`
compressed-tensors Version: `0.11.0`
transformers Version: `4.55.2`
torch Version: `2.8.0`
CUDA Devices: `['NVIDIA H20', 'NVIDIA H20', 'NVIDIA H20', 'NVIDIA H20', 'NVIDIA H20', 'NVIDIA H20', 'NVIDIA H20', 'NVIDIA H20']`
AMD Devices: `None`

🐛 Describe the bug

When executing examples/quantizing_moe/deepseek_r1_example.py, an OOM (Out of Memory) error occurs at step 235 (model.layers.3.mlp.experts.226.down_proj), with the process taking approximately 1 hour. However, when using version v0.7.1, the memory usage is normal, and it only takes 25 minutes to reach step 235.

🛠️ Steps to reproduce

No response

[kylesayrs]

Hi @linyu09-oss!

I assume the OOM you're referring to is a CPU OOM? Note that in both versions, deepseek 685B requires ~1.3TB of CPU memory.

Could you please post the memory graph/ maximum memory used by your LLM-Compressor 0.7.1 run?

[linyu09-oss]

Hi @linyu09-oss!

I assume the OOM you're referring to is a CPU OOM? Note that in both versions, deepseek 685B requires ~1.3TB of CPU memory.

Could you please post the memory graph/ maximum memory used by your LLM-Compressor 0.7.1 run?

Thanks for your reply. I test the script on a machine with 1.8TB RAM. I am currently using v0.7.1 to process DeepSeek R1. I have completed 600 steps and used 170GB, and everything is going well.