[ET][Memory planning] Improve greedy memory planning. #7926
Conversation
This diff replaces the old greedy algorithm. Older algorithm resulted in 35% worse compared to theoretical optimum. THis matter for long context even more since additional overhead can be few hundred MB. For example the theorical optimial for llama3_2 8B, 4-bit quantized modelw ith context length of 2k needs about 1G of memory. This theoretcial max can be observed by looking at the peaks in memory profile. Current agorithm resulted in about 1.6GB of planned memory. New algorithm reduce that to about 1.1G. Differential Revision: [D68448332](https://our.internmc.facebook.com/intern/diff/D68448332/) [ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/7926
Note: Links to docs will display an error until the docs builds have been completed. ❌ 2 New Failures, 1 Unrelated FailureAs of commit 816efe9 with merge base f73b8cf ( NEW FAILURES - The following jobs have failed:
BROKEN TRUNK - The following job failed but were present on the merge base:👉 Rebase onto the `viable/strict` branch to avoid these failures
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
facebook-github-bot
added
the
CLA Signed
|
This pull request was exported from Phabricator. Differential Revision: D68448332 |
This PR needs a
|
This diff replaces the old greedy algorithm. Older algorithm resulted in 35% worse compared to theoretical optimum. THis matter for long context even more since additional overhead can be few hundred MB. For example the theorical optimial for llama3_2 8B, 4-bit quantized modelw ith context length of 2k needs about 1G of memory. This theoretcial max can be observed by looking at the peaks in memory profile. Current agorithm resulted in about 1.6GB of planned memory. New algorithm reduce that to about 1.1G. Differential Revision: [D68448332](https://our.internmc.facebook.com/intern/diff/D68448332/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D68448332 |
Stack from ghstack (oldest at bottom):
This diff replaces the old greedy algorithm. Older algorithm resulted in 35%
worse compared to theoretical optimum. THis matter for long context even more
since additional overhead can be few hundred MB.
For example the theorical optimial for llama3_2 8B, 4-bit quantized modelw ith
context length of 2k needs about 1G of memory. This theoretcial max can be
observed by looking at the peaks in memory profile.
Current agorithm resulted in about 1.6GB of planned memory. New algorithm
reduce that to about 1.1G.
Differential Revision: D68448332