Skip to content

Actions: rayl15/Godwit

Actions

All workflows

Actions

Loading...
Loading

Showing runs from all workflows
38 workflow runs
38 workflow runs

Filter by Workflow

⌥ + click/return to exclude

Filter by Event

⌥ + click/return to exclude

Filter by Status

⌥ + click/return to exclude

Filter by Branch

⌥ + click/return to exclude

Filter by Actor

⌥ + click/return to exclude
Merge ui-conversation: the dashboard holds a conversation ci #30: Commit 094a91f pushed by rayl15
1m 23s main
main
1m 23s
Merge multiturn: reuse the KV cache prefix across chat turns ci #29: Commit 68b2599 pushed by rayl15
1m 14s main
main
1m 14s
Merge multiturn: reuse the KV cache prefix across chat turns pages #8: Commit 68b2599 pushed by rayl15
23s main
main
23s
Merge: prefill is at its floor; the cache stats were only ever tellin… ci #28: Commit 54846b3 pushed by rayl15
1m 3s main
main
1m 3s
Merge: the MXFP4 shrinkage is real, not the encoder, and not correctable ci #27: Commit 1f49106 pushed by rayl15
58s main
main
58s
Merge: GeGLU gets its own kernel ci #26: Commit ddeceb2 pushed by rayl15
1m 12s main
main
1m 12s
Merge: measure what quantisation costs Qwen3 ci #25: Commit 1a61a4a pushed by rayl15
1m 12s main
main
1m 12s
Flesh out .gitignore for people who are not me ci #24: Commit 01ec875 pushed by rayl15
1m 37s main
main
1m 37s
Merge: the range map predicts routing, pre-warming does not pay ci #23: Commit e105716 pushed by rayl15
1m 20s main
main
1m 20s
Fix what a first-time clone actually hits ci #22: Commit b7d706b pushed by rayl15
1m 10s main
main
1m 10s
Add lookahead prefetching: +3.1%, measured ci #21: Commit 89b381d pushed by rayl15
1m 49s main
main
1m 49s
Lookahead routing predicts 87-91% of experts a layer early ci #20: Commit 2c58c84 pushed by rayl15
1m 57s main
main
1m 57s
Measure HOBBIT-style precision tiering, and decline to build it ci #19: Commit f47bd50 pushed by rayl15
1m 39s main
main
1m 39s
Retract the claim that prefetching does not work pages #7: Commit 88cead1 pushed by rayl15
16s main
main
16s
Update the site for two model families pages #6: Commit f3e6a51 pushed by rayl15
15s main
main
15s
Build the range map from the dashboard ci #18: Commit 3bee808 pushed by rayl15
1m 44s main
main
1m 44s
Make the README true again pages #5: Commit 89877f4 pushed by rayl15
23s main
main
23s
Let serve take the sampling flags the README says it takes ci #17: Commit c2a95a5 pushed by rayl15
1m 26s main
main
1m 26s
Qwen3 answers: the loop was a dtype, not quantisation ci #16: Commit 13211dc pushed by rayl15
1m 18s main
main
1m 18s
Qwen3 runs, and is too degraded to use ci #15: Commit c13ad45 pushed by rayl15
1m 39s main
main
1m 39s
Add a verifier for a quantised install ci #14: Commit c68c021 pushed by rayl15
1m 4s main
main
1m 4s
Test the Qwen3 paths, and correct the README's claim about genericity ci #13: Commit b6d9e7a pushed by rayl15
1m 10s main
main
1m 10s
Teach the runtime the four things Qwen3 does differently ci #12: Commit 7352bb6 pushed by rayl15
1m 53s main
main
1m 53s
Install Qwen3: build expert contiguity instead of inheriting it ci #11: Commit 4dc0520 pushed by rayl15
1m 5s main
main
1m 5s
Describe a checkpoint's tensor names instead of hardcoding them ci #10: Commit 1eee65e pushed by rayl15
1m 47s main
main
1m 47s