Skip to content

[Doc] update the document of qwen3 and cuda_graph usage #5703

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 2 commits into from
Jul 7, 2025
Merged

[Doc] update the document of qwen3 and cuda_graph usage #5703

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
71abb9b
update the document of qwen3 and cuda_graph usage
byshiue Jul 3, 2025
eb1581d
revert change
byshiue Jul 7, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
64 changes: 36 additions & 28 deletions examples/models/core/qwen/README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -42,33 +42,35 @@ In addition, there are two shared files in the parent folder [`examples`](../../
* [`summarize.py`](../../../summarize.py) to summarize the articles in the [cnn_dailymail](https://huggingface.co/datasets/abisee/cnn_dailymail) dataset.

## Support Matrix
| Model Name | FP16/BF16 | FP8 | WO | AWQ | GPTQ | SQ | TP | PP | Arch |
| :-------------: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :-----: |
| Qwen-1_8B(-Chat) | Y | Y | Y | Y* | Y | Y | Y | Y | Ampere+ |
| Qwen-7B(-Chat) | Y | Y | Y | Y | Y | Y | Y | Y | Ampere+ |
| Qwen-14B(-Chat) | Y | Y | Y | Y* | Y | Y | Y | Y | Ampere+ |
| Qwen-72B(-Chat) | Y | Y | Y | Y | Y | Y | Y | Y | Ampere+ |
| Qwen1.5-0.5B(-Chat)| Y | Y | Y | Y* | Y | Y | Y | Y | Ampere+ |
| Qwen1.5-1.8B(-Chat)| Y | Y | Y | Y* | Y | Y | Y | Y | Ampere+ |
| Qwen1.5-4B(-Chat) | Y | Y | Y | Y* | Y | Y | Y | Y | Ampere+ |
| Qwen1.5-7B(-Chat) | Y | Y | Y | Y | Y | Y | Y | Y | Ampere+ |
| Qwen1.5-14B(-Chat) | Y | Y | Y | Y* | Y | Y | Y | Y | Ampere+ |
| Qwen1.5-32B(-Chat) | Y | Y | Y | Y | Y | Y | Y | Y | Ampere+ |
| Qwen1.5-72B(-Chat) | Y | Y | Y | Y | Y | Y | Y | Y | Ampere+ |
| Qwen1.5-110B(-Chat)| Y | Y | Y | Y | Y | Y | Y | Y | Ampere+ |
| Qwen1.5-MoE-A2.7B(-Chat)| Y | - | Y | - | - | - | Y | Y | Ampere+ |
| Qwen2-0.5B(-Instruct)| Y | Y | Y | Y* | Y | Y | Y | Y | Ampere+ |
| Qwen2-1.5B(-Instruct)| Y | Y | Y | Y* | Y | Y | Y | Y | Ampere+ |
| Qwen2-7B(-Instruct)| Y | Y | Y | Y | Y | Y | Y | Y | Ampere+ |
| Qwen2-57B-A14B(-Instruct)| Y | - | Y | - | - | - | Y | Y | Ampere+ |
| Qwen2-72B(-Instruct)| Y | Y | Y | Y* | Y | Y | Y | Y | Ampere+ |
| Qwen2.5-0.5B(-Instruct)| Y | Y | Y | Y* | Y | Y | Y | Y | Ampere+ |
| Qwen2.5-3B(-Instruct)| Y | Y | Y | Y* | Y | Y | Y | Y | Ampere+ |
| Qwen2.5-1.5B(-Instruct)| Y | Y | Y | Y | Y | Y | Y | Y | Ampere+ |
| Qwen2.5-7B(-Instruct)| Y | Y | Y | Y | Y | Y | Y | Y | Ampere+ |
| Qwen2.5-32B(-Instruct)| Y | Y | Y | Y | Y | Y | Y | Y | Ampere+ |
| Qwen2.5-72B(-Instruct)| Y | Y | Y | Y* | Y | Y | Y | Y | Ampere+ |
| QwQ-32B | Y | Y | Y | Y | Y | Y | Y | Y | Ampere+ |
| Model Name | FP16/BF16 | FP8 | nvfp4 | WO | AWQ | GPTQ | SQ | TP | PP | EP | Arch |
| :-------------: | :---: | :---: | :-----: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :-----: |
| Qwen-1_8B(-Chat) | Y | Y | - | Y | Y* | Y | Y | Y | Y | - | Ampere+ |
| Qwen-7B(-Chat) | Y | Y | - | Y | Y | Y | Y | Y | Y | - | Ampere+ |
| Qwen-14B(-Chat) | Y | Y | - | Y | Y* | Y | Y | Y | Y | - | Ampere+ |
| Qwen-72B(-Chat) | Y | Y | - | Y | Y | Y | Y | Y | Y | - | Ampere+ |
| Qwen1.5-0.5B(-Chat)| Y | Y | - | Y | Y* | Y | Y | Y | Y | - | Ampere+ |
| Qwen1.5-1.8B(-Chat)| Y | Y | - | Y | Y* | Y | Y | Y | Y | - | Ampere+ |
| Qwen1.5-4B(-Chat) | Y | Y | - | Y | Y* | Y | Y | Y | Y | - | Ampere+ |
| Qwen1.5-7B(-Chat) | Y | Y | - | Y | Y | Y | Y | Y | Y | - | Ampere+ |
| Qwen1.5-14B(-Chat) | Y | Y | - | Y | Y* | Y | Y | Y | Y | - | Ampere+ |
| Qwen1.5-32B(-Chat) | Y | Y | - | Y | Y | Y | Y | Y | Y | - | Ampere+ |
| Qwen1.5-72B(-Chat) | Y | Y | - | Y | Y | Y | Y | Y | Y | - | Ampere+ |
| Qwen1.5-110B(-Chat)| Y | Y | - | Y | Y | Y | Y | Y | Y | - | Ampere+ |
| Qwen1.5-MoE-A2.7B(-Chat)| Y | - | - | Y | - | - | - | Y | Y | - | Ampere+ |
| Qwen2-0.5B(-Instruct)| Y | Y | - | Y | Y* | Y | Y | Y | Y | - | Ampere+ |
| Qwen2-1.5B(-Instruct)| Y | Y | - | Y | Y* | Y | Y | Y | Y | - | Ampere+ |
| Qwen2-7B(-Instruct)| Y | Y | - | Y | Y | Y | Y | Y | Y | - | Ampere+ |
| Qwen2-57B-A14B(-Instruct)| Y | - | - | Y | - | - | - | Y | Y | - | Ampere+ |
| Qwen2-72B(-Instruct)| Y | Y | - | Y | Y* | Y | Y | Y | Y | - | Ampere+ |
| Qwen2.5-0.5B(-Instruct)| Y | Y | - | Y | Y* | Y | Y | Y | Y | - | Ampere+ |
| Qwen2.5-3B(-Instruct)| Y | Y | - | Y | Y* | Y | Y | Y | Y | - | Ampere+ |
| Qwen2.5-1.5B(-Instruct)| Y | Y | - | Y | Y | Y | Y | Y | Y | - | Ampere+ |
| Qwen2.5-7B(-Instruct)| Y | Y | - | Y | Y | Y | Y | Y | Y | - | Ampere+ |
| Qwen2.5-32B(-Instruct)| Y | Y | - | Y | Y | Y | Y | Y | Y | - | Ampere+ |
| Qwen2.5-72B(-Instruct)| Y | Y | - | Y | Y* | Y | Y | Y | Y | - | Ampere+ |
| QwQ-32B | Y | Y | - | Y | Y | Y | Y | Y | Y | - | Ampere+ |
| Qwen3-32B | Y | Y | Y | - | - | - | - | Y | - | Y | Hopper+ |
| Qwen3-235B-A3B | Y | Y | Y | - | - | - | - | Y | - | Y | Hopper+ |

Please note that Y* sign means that the model does not support all the AWQ + TP combination.

Expand All @@ -82,6 +84,8 @@ Please note that Y* sign means that the model does not support all the AWQ + TP

Currently Qwen1 models does not support dynamic NTK and logn attention. Therefore, accuracy on long sequence input for the Qwen-7B and Qwen-14B model is not promised.

For Qwen3 models, we only list the largest models for dense and MoE architectures, but models of other sizes follow similar patterns.

## Usage

The TensorRT-LLM Qwen example code locates at [examples/models/core/qwen](./). It takes HF weights as input, and builds the corresponding TensorRT engines. The number of TensorRT engines depends on the number of GPUs used to run inference.
Expand Down Expand Up @@ -692,7 +696,11 @@ concurrency=128
path_data=./aa_prompt_isl_1k_osl_2k_qwen3_10000samples.txt

# Setup the extra configuration for llm-api
echo -e "disable_overlap_scheduler: false\cuda_graph_config: {}\nprint_iter_log: true\ncuda_graph_batch_sizes: [1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,32,64,128]\nenable_attention_dp: true " > ${path_config}
echo -e "disable_overlap_scheduler: false
print_iter_log: true
cuda_graph_config:
batch_sizes: [1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,32,64,128]
enable_attention_dp: true " > ${path_config}

# Run trtllm-bench with pytorch backend
mpirun --allow-run-as-root --oversubscribe -n 1 \
Expand Down