[v1][torch.compile] support managing cudagraph buffer #10203
Conversation
Signed-off-by: youkaichao <youkaichao@gmail.com>
Signed-off-by: youkaichao <youkaichao@gmail.com>
Signed-off-by: youkaichao <youkaichao@gmail.com>
|
👋 Hi! Thank you for contributing to the vLLM project. Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging. To run CI, PR reviewers can do one of these:
🚀 |
| ): | ||
|
|
||
| with set_compile_context([1, 2]): | ||
| model(input_buffer) |
There was a problem hiding this comment.
To double check: Here, input_buffer is not directly cached and reused by the backend, right?
There was a problem hiding this comment.
right. let me rename the variable. it is not a persistent buffer anymore. it can be a newly created tensor.
Co-authored-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
Signed-off-by: youkaichao <youkaichao@gmail.com>
youkaichao
changed the title
…0203) Signed-off-by: youkaichao <youkaichao@gmail.com> Co-authored-by: Woosuk Kwon <woosuk.kwon@berkeley.edu> Signed-off-by: Sumit Dubey <sumit.dubey2@ibm.com>
…0203) Signed-off-by: youkaichao <youkaichao@gmail.com> Co-authored-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
advantage: caller does not need to explicitly manage the input buffer
disadvantage: the cost of one more input tensor, and one more copy kernel.
given the code simplification from user side, i think it should be worthwhile.