Repository navigation
Strix profiles: measured fixes, and scope the port to Windows - #2
Merged
Merged
Conversation
…leOp off Measured on a Ryzen AI Max+ 395 with the released weights: - resident 64 took held-out chat 0.60 -> 0.92 and code 0.71 -> 1.19 nats; the model learned to choose within 32. strix_halo.yaml keeps 32 - context_end 8192 scored 1.66 nats against 0.62 at 4096: held-out is read at the ceiling, past positions the model was trained on. Kept at 4096 - TunableOp re-tunes nearly every call because the expert dispatch's matmul shapes change with every forward; a held-out pass ran past half an hour. Off in both profiles - README "What to expect" rewritten with these and the measured speeds - weights-*/ ignored (weights-strix/, backups) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The Linux instructions (GTT / ttm.pages_limit memory setup, Linux wheels) were written from AMD and community sources and never run on hardware. The port is tested on Windows only, so the README, the Strix profiles and minagi/device.py now say so, and CI runs on Windows. The code itself stays portable: nothing here stops the project running on Linux as upstream did. The profiles' memory note now records what was measured: torch reported 99.7 GB on the test machine while Windows kept 63.6 GB for itself. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
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 sugg
2F79
estion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Summary
Two commits pushed to
strix-haloafter #1 was merged. Without them,main'sconfigs/strix_halo.yamlstill carries the settings that measured badly on the released weights.9db226b: Strix profiles keep the released model's resident and context; TunableOp off. Measured on the 8060S with the released weights:resident64 took held-out chat from 0.60 to 0.92 nats and code from 0.71 to 1.19.context_end8192 scored 1.66 nats against 0.62 at 4096.strix_halo.yamlgoes back to 32 and 4096, and TunableOp is off in both profiles. The README's "What to expect" now reports these measurements.weights-*/is gitignored.4a2f138: scope the port to Windows. Drops the untested Linux setup notes and the Ubuntu CI job. The code stays portable.The full findings and the 20-minute GPU training session are written up in #1.
Test plan
pytest: 18/18 locally9db226b4a2f138(Windows only)🤖 Generated with Claude Code