Milestones
List view
Q4 milestone for Speech models productization
Due by December 31, 2026•0/9 issues closedQ3 milestone for Speech models productization
Overdue by 12 day(s)•Due by September 30, 2026•14/20 issues closed- Overdue by 12 day(s)•Due by September 30, 2026•3/5 issues closed
- No due date•0/2 issues closed
- No due date•2/2 issues closed
- Overdue by 1 month(s)•Due by August 15, 2026•1/1 issues closed
- Overdue by 1 month(s)•Due by August 15, 2026•6/7 issues closed
- Overdue by 2 month(s)•Due by July 23, 2026•5/7 issues closed
- Overdue by 2 month(s)•Due by August 6, 2026•1/1 issues closed
- No due date•1/17 issues closed
- No due date•1/11 issues closed
- No due date•3/10 issues closed
- No due date•5/14 issues closed
- No due date•2/4 issues closed
- No due date•0/2 issues closed
- No due date•1/1 issues closed
To test owners, a lot of the test cases were unfortunately not guarded in CI (e.g., due to @pytest.mark.requires_grid_size((13, 10)), a few test cases have been skipped on P150b/P300 devices in CI); that is why I started to look at fixing the unit test coverage about 3 weeks ago. After TT-deploy, I hope test owners can spend some time now to fix these issues. One thing I noticed is that some test cases run successfully on bh_galaxy machine in IRD but fails consistently in CI. I added skips for those as well. We can talk about a proper solution to provide testing coverage for those test cases. --- Read detailed analysis here: https://gist.github.com/gwangTT/6715322f5f9d0c7c6f432e2c63769d83; copying important notes from the report to here: ## Function Rollup - Function status rules: `FAILED` if any case failed or errored, `MIXED` if there were both passed and skipped cases and no failures, `SKIPPED` if all cases were skipped, `PASSED` if all cases passed. - `Case summary` is `P/S/F` for passed, skipped, and failed case counts. - `Test case` uses the exact nodeid suffix when the terminal printed it. When the terminal only printed an aggregated skip summary, this file uses a synthetic case label and says so in `Notes`. - When the same exact test case appears in multiple included slices, this file keeps one row per unique case and uses the latest observed status. - Function rollup in included sources: `288` functions = `212` passed, `71` skipped, `5` mixed, `0` failed. - Case rollup in included sources: `805` unique case entries = `501` passed, `304` skipped, `0` failed.
No due date•28/28 issues closed- No due date•0/4 issues closed
- Overdue by 4 month(s)•Due by May 22, 2026•39/45 issues closed
- Overdue by 2 month(s)•Due by July 31, 2026
- Overdue by 1 month(s)•Due by August 28, 2026
Tracking issues for runtime CI test infrastructure and stability.
No due date•5/12 issues closed- Overdue by 3 month(s)•Due by June 30, 2026•4/4 issues closed
- No due date•8/9 issues closed
- No due date•4/16 issues closed
- Overdue by 2 month(s)•Due by July 31, 2026
- Overdue by 4 month(s)•Due by May 29, 2026•2/2 issues closed
- Overdue by 6 month(s)•Due by March 31, 2026•11/11 issues closed
- Overdue by 7 month(s)•Due by February 27, 2026•19/19 issues closed
- No due date•1/1 issues closed
Tracks productization/mainlining tasks for enabling "Fabric Channel Trimming" for consumption by blitzdecode
Overdue by 7 month(s)•Due by March 16, 2026•7/11 issues closed- No due date
- No due date•2/2 issues closed
- No due date•6/6 issues closed
- Overdue by 7 month(s)•Due by February 28, 2026
Captures work related to deployment, SLURM scheduling, topology scheduling for 36xBH GLX
Overdue by 6 month(s)•Due by March 31, 2026•3/3 issues closed- Overdue by 6 month(s)•Due by March 31, 2026•2/2 issues closed
- Overdue by 8 month(s)•Due by January 30, 2026•24/25 issues closed
- Overdue by 8 month(s)•Due by January 31, 2026•3/3 issues closed
Tracking baseline "zero-load" Fabric performance characterization for BH Galaxy Pod. Tested on 8x16 and 4x32 pods. zero-load latency, various bandwidth traffic patterns x all topologies
Overdue by 8 month(s)•Due by February 6, 2026•10/18 issues closedCatch all milestone for quality of life improvements
Overdue by 6 month(s)•Due by March 31, 2026•6/7 issues closed- No due date•0/1 issues closed
On completion of this milestone, all initial fabric manager + fabric control paths are expected to be complete. This means: - Fabric manager has control flows with fabric, allowing it to pause, resume, exit, issue retrain - Fabric is expected to always respond to the above messages (barring chip failure or noc hangs)
Overdue by 8 month(s)•Due by February 6, 2026•3/7 issues closedPerformance workitems scheduled for large LLM inference. Includes low-latency work for low-latency decode and other miscelaneous optimizations to bump up 2D perf to near linear fabric levels.
Overdue by 8 month(s)•Due by February 13, 2026•13/23 issues closed- No due date•16/41 issues closed
Time to do more model work with TTT (v1 and v2)!
No due date•11/16 issues closedThis milestone tracks all work needed to deliver DeepSeek v3 671B MoE Alpha1 on 2× Galaxy - [x] Prefill+decode on device at batch=512 - [ ] Supports prefill up to 128k and decode up to 64k - [x] Good top1/top5 accuracy - [ ] 8hr Stress Test Passes on CI - [ ] vLLM and tt-inference-server integration complete - [ ] Customer Eval Accuracy Tests Pass Use this milestone **only for DeepSeek Alpha1** tickets that directly contribute to the above criteria.
Overdue by 7 month(s)•Due by February 20, 2026•22/23 issues closed- Overdue by 7 month(s)•Due by February 28, 2026•18/24 issues closed
- Overdue by 7 month(s)•Due by March 15, 2026•17/25 issues closed
- Overdue by 6 month(s)•Due by March 31, 2026•3/5 issues closed