Repository navigation
cuda: add experimental Tesla P4 8GB (sm_61) support - #336
Closed
CC-David-CC wants to merge 6 commits into
Closed
CC-David-CC wants to merge 6 commits into
CC-David-CC wants to merge 6 commits into
Conversation
Owner
|
Thanks! Please rebase onto 0.1.31's experimental flag instead of a new SM61 flag ( |
Closed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Adds experimental Tesla P4 8GB support using CUDA 12.x, with a bounded FP32 prefill path, numerical tests, configurations, and benchmark results.
Tested IQ3_S with Q8 KV, 8192 input tokens, up to 512 output, and 9216 allocated context. Tuned MTP generation: 22.09 tok/s counting, 20.24 coding, 14.94 writing. MTP increased total fresh-request time because prefill was slower. Full timings and limitations are in docs/DETAILS.md.
Separate from AMD PR #323. Includes shared non-MTP serving and benchmark support, but excludes the RX 5500 XT GPU port. Preserves upstream README.
Draft: measurements come from tested revision 9d393df. This isolated contribution branch has not yet been rebuilt and retested. Credit for Strata, expert caching, and MTP belongs to upstream.