biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Docs] add docs for cuda graph v1 (#24374)

Browse Source 
Signed-off-by: fhl <2410591650@qq.com>
Signed-off-by: fhl2000 <63384265+fhl2000@users.noreply.github.com>
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Co-authored-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Co-authored-by: Luka Govedič <ProExpertProg@users.noreply.github.com>
This commit is contained in:
fhl2000
2025-10-07 20:25:05 +08:00
committed by GitHub
parent 883b42896a
commit 63773a6200
6 changed files with 246 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
docs/design/torch_compile.md
View File

@@ -4,6 +4,9 @@ In vLLM's V1 architecture, `torch.compile` is enabled by default and is a critic
Throughout the example, we will run a common Llama model, and turn on debug level logging to show all the details. The command to be used is `VLLM_LOGGING_LEVEL=DEBUG vllm serve meta-llama/Llama-3.2-1B`.
!!! note
For more information and the latest progress of `torch.compile` integration, see this [Blog Post](https://blog.vllm.ai/2025/08/20/torch-compile.html).
## Compilation Cache
In the very verbose logs, we can see:
@@ -133,7 +136,7 @@ Unfortunately, because auto-tuning takes quite a long time (from seconds to minu
## Cudagraph Capture
vLLM's V1 architecture uses piecewise cudagraph. The full computation graph is split as mentioned above, and we only capture the cudagraph for the piece of graph between attention operations (including the first graph before any attention operation, and the last graph after all the attention operation). This is based on a common observation: computation between attentions are usually token-wise and easy to deal with for cudagraph; while the attention operation is non-trivial to be cudagraph compatible. Thus, by running the attention operation in eager mode while the rest operations in cudagraph, we keep the flexibility of the attention operation.
vLLM's V1 architecture uses piecewise cudagraph that aligns with the piecewise compilation. The full computation graph is split as mentioned above, and we only capture the cudagraph for the piece of graph between attention operations (including the first graph before any attention operation, and the last graph after all the attention operation). This is based on a common observation: computation between attentions are usually token-wise and easy to deal with for cudagraph; while the attention operation is non-trivial to be cudagraph compatible. Thus, by running the attention operation in eager mode while the rest operations in cudagraph, we keep the flexibility of the attention operation.
The piecewise cudagraph also has fine-grained memory management. The purpose is to only exclude the attention kernel from cudagraph, while keeping all the rest modules and the memory allocation operations in the cudagraph. This is why the attention operation in V1 has the output tensor as the input of the attention.
@@ -150,6 +153,4 @@ Then it will only capture cudagraph for the specified sizes. It can be useful to
### Full Cudagraph capture
It is possible to include attention as part of the cudagraph if using an attention backend that is cudagraph compatible. This can improve performance in some cases such as decode speed for smaller models. Enable this using `--compilation-config '{"full_cuda_graph": true}'`.
Currently only FlashAttention 3 is compatible, and only when cascade attention is disabled.
It is possible to include attention as part of the cudagraph if using an attention backend that is cudagraph compatible. This can improve performance in some cases such as decode speed for smaller models or MOEs. See [CUDA Graphs](cuda_graphs.md) for more details.