[Performance]: b200x8 deepseek-ai/DeepSeek-V3.2-Exp max perf #10278
Description
Proposal to improve performance
No response
Report of performance regression
Hi! Do you have any ideas on how to increase TPS on B200 ×8?
cuda_graph_config:
enable_padding: true
batch_sizes: [1, 2, 4, 8, 16, 32, 64, 128, 256]
moe_config:
backend: DEEPGEMM
max_num_tokens: 32768
kv_cache_config:
enable_block_reuse: false
dtype: fp8
tokens_per_block: 64
free_gpu_memory_fraction: 0.55
enable_chunked_prefill: true
stream_interval: 50
batch_wait_timeout_ms: 20
enable_attention_dp: true
attention_dp_config:
batching_wait_iters: 0
enable_balance: true
timeout_iters: 60
trtllm-serve deepseek-ai/DeepSeek-V3.2-Exp \
--backend pytorch \
--max_batch_size 256 \
--max_num_tokens 32768 \
--max_seq_len 32768 \
--tp_size 8 \
--ep_size 8 \
--pp_size 1 \
--tool_parser deepseek_v32 \
--extra_llm_api_options /app/extra-llm-api-config.yml \
--host 0.0.0.0 \
--port 12345
Misc discussion on performance
No response
Your current environment (if you think it is necessary)
System Information:
- OS: 24.04
- Python version:
- CUDA version: 1
- GPU model(s):
- Driver version:
- TensorRT version:
- PyTorch version:
- TensorRT-LLM version:
Detailed output:
Wed Dec 24 12:49:44 2025
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 580.95.05 Driver Version: 580.95.05 CUDA Version: 13.0 |
+-----------------------------------------+------------------------+----------------------+
| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|=========================================+========================+======================|
| 0 NVIDIA B200 On | 00000000:05:00.0 Off | 0 |
| N/A 33C P0 192W / 1000W | 166593MiB / 183359MiB | 0% Default |
| | | Disabled |
+-----------------------------------------+------------------------+----------------------+
| 1 NVIDIA B200 On | 00000000:06:00.0 Off | 0 |
| N/A 35C P0 196W / 1000W | 156996MiB / 183359MiB | 0% Default |
| | | Disabled |
+-----------------------------------------+------------------------+----------------------+
| 2 NVIDIA B200 On | 00000000:07:00.0 Off | 0 |
| N/A 34C P0 192W / 1000W | 156996MiB / 183359MiB | 0% Default |
| | | Disabled |
+-----------------------------------------+------------------------+----------------------+
| 3 NVIDIA B200 On | 00000000:08:00.0 Off | 0 |
| N/A 32C P0 194W / 1000W | 156996MiB / 183359MiB | 0% Default |
| | | Disabled |
+-----------------------------------------+------------------------+----------------------+
| 4 NVIDIA B200 On | 00000000:09:00.0 Off | 0 |
| N/A 33C P0 193W / 1000W | 156996MiB / 183359MiB | 0% Default |
| | | Disabled |
+-----------------------------------------+------------------------+----------------------+
| 5 NVIDIA B200 On | 00000000:0A:00.0 Off | 0 |
| N/A 35C P0 198W / 1000W | 156996MiB / 183359MiB | 0% Default |
| | | Disabled |
+-----------------------------------------+------------------------+----------------------+
| 6 NVIDIA B200 On | 00000000:0B:00.0 Off | 0 |
| N/A 34C P0 194W / 1000W | 156996MiB / 183359MiB | 0% Default |
| | | Disabled |
+-----------------------------------------+------------------------+----------------------+
| 7 NVIDIA B200 On | 00000000:0C:00.0 Off | 0 |
| N/A 32C P0 191W / 1000W | 156676MiB / 183359MiB | 0% Default |
| | | Disabled |
+-----------------------------------------+------------------------+----------------------+
+-----------------------------------------------------------------------------------------+
| Processes: |
| GPU GI CI PID Type Process name GPU Memory |
| ID ID Usage |
|=========================================================================================|
| 0 N/A N/A 769002 C /usr/bin/python 686MiB |
| 0 N/A N/A 770142 C /usr/bin/python 16588... |
| 1 N/A N/A 770143 C /usr/bin/python 15698... |
| 2 N/A N/A 770144 C /usr/bin/python 15698... |
| 3 N/A N/A 770145 C /usr/bin/python 15698... |
| 4 N/A N/A 770146 C /usr/bin/python 15698... |
| 5 N/A N/A 770147 C /usr/bin/python 15698... |
| 6 N/A N/A 770148 C /usr/bin/python 15698... |
| 7 N/A N/A 770149 C /usr/bin/python 15666... |
+-----------------------------------------------------------------------------------------+
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2025 NVIDIA Corporation
Built on Wed_Aug_20_01:58:59_PM_PDT_2025
Cuda compilation tools, release 13.0, V13.0.88
Build cuda_13.0.r13.0/compiler.36424714_0
python --version
pip show tensorrt_llm tensorrt torch
Python 3.12.3
Name: tensorrt_llm
Version: 1.2.0rc7
Summary: TensorRT LLM provides users with an easy-to-use Python API to define Large Language Models (LLMs) and supports state-of-the-art optimizations to perform inference efficiently on NVIDIA GPUs.
Home-page: https://github.com/NVIDIA/TensorRT-LLM
Author: NVIDIA Corporation
Author-email:
License: Apache License 2.0
Location: /usr/local/lib/python3.12/dist-packages
Requires: accelerate, aenum, apache-tvm-ffi, backoff, blake3, blobfile, build, click, click_option_group, colored, cuda-python, datasets, diffusers, einops, etcd3, evaluate, fastapi, flashinfer-python, h5py, jsonschema, lark, llguidance, matplotlib, meson, mistral-common, mpi4py, mpmath, ninja, numexpr, numpy, nvidia-cuda-nvrtc, nvidia-cutlass-dsl, nvidia-ml-py, nvidia-modelopt, nvidia-nccl-cu13, nvtx, omegaconf, onnx, onnx_graphsurgeon, openai, openai-harmony, opencv-python-headless, optimum, ordered-set, pandas, partial_json_parser, patchelf, peft, pillow, plotly, polygraphy, prometheus_client, prometheus_fastapi_instrumentator, psutil, pulp, pydantic, pydantic-settings, pyzmq, sentencepiece, setuptools, soundfile, starlette, StrEnum, tensorrt, tiktoken, torch, torch-c-dlpack-ext, torchao, torchvision, transformers, triton, urllib3, uvicorn, wheel, xgrammar
Required-by:
---
Name: tensorrt
Version: 10.13.3.9
Summary: A high performance deep learning inference library
Home-page: https://github.com/nvidia/tensorrt
Author: NVIDIA Corporation
Author-email:
License: Proprietary
Location: /usr/local/lib/python3.12/dist-packages
Requires:
Required-by: tensorrt_llm
---
Name: torch
Version: 2.9.0a0+145a3a7bda.nv25.10
Summary: Tensors and Dynamic neural networks in Python with strong GPU acceleration
Home-page: https://pytorch.org
Author:
Author-email: PyTorch Team <[email protected]>
License: BSD-3-Clause
Location: /usr/local/lib/python3.12/dist-packages
Requires: filelock, fsspec, jinja2, networkx, setuptools, sympy, typing-extensions
Required-by: accelerate, flash_attn, flashinfer-python, lightning-thunder, nvidia-modelopt, nvidia-resiliency-ext, optimum, peft, tensorrt_llm, torch_c_dlpack_ext, torchprofile, torchvision, transformer_engine, xgrammar
Before submitting a new issue...
- Make sure you already searched for relevant issues, and checked the documentation and examples for answers to frequently asked questions.