deepspeedai / deepspeedai/DeepSpeed
[REQUEST] Support profiling backward using Flops Profiler as a standalone package
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 43.1k
- Forks
- 5k
- Avg merge
- 4d 15h
- Merged PRs (30d)
- 112
Description
Is your feature request related to a problem? Please describe.
I try to use flops profiler as a standalone package to profiling the backward of a custom model. I have followed both two instructions, here and here. However, both examples only consider profiling the forward stage. I check the code and it seems that it currently only prints the bwd_latency when using runtime as below
https://github.com/microsoft/DeepSpeed/blob/5e42cc8be22f58ef6be022fcc6d69fe11c7bfe25/deepspeed/profiling/flops_profiler/profiler.py#L313-L317
Describe the solution you'd like
Is there any approach to profiling the backward by the flops profiler?
Contributor guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
Start with deepspeed/profiling/flops_profiler/profiler.py around lines 313-317 and compare the standalone-package instructions in the linked DeepSpeed tutorial and API documentation. Trace how backward latency is currently reported through the runtime, then define and verify standalone backward profiling with a custom model, including its reported results.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python, pytorch
- Domain
- machine-learning, performance
- Issue type
- Feature
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Activity status
- Stale
- Clarity
- Mostly clear
- Newbie friendliness
- 35/100