google-deepmind / google-deepmind/alphafold3
Autotuning cache miss for PallasTritonGatedLinearUnit and Inference time increases significantly.
- Dominant language
- Python
- Stars
- 8.6k
- Forks
- 1.4k
- PR merge metrics
- No merged PRs in 30d
Description
After updating to the latest version of the code, I noticed the following Warning when running on an RTX A6000: `Autotuning cache miss for PallasTritonGatedLinearUnit(config=None, vjp=GatedLinearUnitVjp(config=None, vjp=None)) on NVIDIA RTX A6000 with key immutabledict.` Additionally, for a protein with a length of 898 amino acids, the GPU inference time for a single seed is at least 80 seconds longer compared to version V3.0.1.
Contributor guide
Research direction
Reproduce the warning and slowdown on an NVIDIA RTX A6000 using the latest code, comparing it with V3.0.1 for a protein sequence of length 898. Start from the PallasTritonGatedLinearUnit autotuning path and measure single-seed GPU inference time; done means identifying and resolving the cache miss or regression and confirming the runtime comparison.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python
- Domain
- machine-learning, performance
- Issue type
- Bug
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Activity status
- Active
- Clarity
- Needs clarification
- Newbie friendliness
- 35/100