feat(tenstorrent): host-free decode forward — capture investigation + architecture - #694

Closed

lu-zero wants to merge 27 commits into

mudler:mainfrom

lu-zero:row/BACKEND-TENSTORRENT-HOST-FREE-R1

Commits on Aug 14, 2026

spec(tenstorrent): trace-runner spike — NO-GO for T=1 decode, MEASURED
lu-zero
committed
9d48a24
plan(tenstorrent): host-free decode forward — the capture prerequisite
lu-zero
committed
5181e1e
feat(tenstorrent): host-free decode investigation — R1-R3b + architecture answer
lu-zero
committed
6257a90
record(tenstorrent): host-free row ACTIVE + drop stray bootstrap .i32
lu-zero
committed
b770361
review(tenstorrent): address fresh-review findings on host-free branch
lu-zero
committed
315c97b
perf(tenstorrent): steady-state Qwen3 baselines — 7.3 default / 6.9 all-device
lu-zero
committed
6835333
feat(tenstorrent): item 5 — zero-cache (capture-safe fill) + upload bisection
lu-zero
committed
a1ff3f2
feat(tenstorrent): item 5 — rope cos/sin capture-safe (HIT measured on card)
lu-zero
committed
0112b8b
docs(env): document VT_TT_HOST_FREE_DECODE + VT_TT_TRACE_DEBUG
lu-zero
committed
eaa2a7c
fix(capi): hoist SpeechRegistry() out of extern "C" (main-red, aa3643b6)
lu-zero
committed
d579b10
spec(tenstorrent): RAC capture-safe design complete — the next step mapped
lu-zero
committed
bee5b7f
feat(tenstorrent): item 5 RAC — device branch + warm hook + shape-keyed cache
lu-zero
committed
e5ed189
feat(tenstorrent): item 5 RAC — shadow priming works, device branch executes
lu-zero
committed
05daeef
feat(tenstorrent): item 5 RAC — all fixes landed; paged_update_cache not capture-safe
lu-zero
committed
757036e
spec(tenstorrent): paged_update_cache IS in-place; writes from build_padded
lu-zero
committed
4f22487
fix(tenstorrent): sharded_zero crash — warm order (shadows before RAC idx)
lu-zero
committed
e667eb0
feat(tenstorrent): item 5 RAC — approach (b) probe: capture passes RAC
lu-zero
committed
a7f6947
feat(tenstorrent): item 5 PA — KV shadow skip works; next = metadata uploads
lu-zero
committed
9dbfb88
feat(tenstorrent): item 5 PA meta warm works; sdpa not compiled (cold bail)
lu-zero
committed
812ac3e
feat(tenstorrent): CAPTURE COMPLETE — 86 tok/s replay (12x speedup)
lu-zero
committed
0345788
feat(tenstorrent): RAC in-capture — paged_update_cache capture-safe
lu-zero
committed
6bbb4ac
diag(tenstorrent): RAC k/v copy wrong output — TILE→sharded layout mismatch
lu-zero
committed
c18895c
spec(tenstorrent): session checkpoint — capture works, 83 tok/s
lu-zero
committed
aa94a0f

Navigation Menu

feat(tenstorrent): host-free decode forward — capture investigation + architecture - #694

feat(tenstorrent): host-free decode forward — capture investigation + architecture#694
lu-zero wants to merge 27 commits into
mudler:mainfrom
lu-zero:row/BACKEND-TENSTORRENT-HOST-FREE-R1

Commits on Aug 14, 2026

spec(tenstorrent): trace-runner spike — NO-GO for T=1 decode, MEASURED

plan(tenstorrent): host-free decode forward — the capture prerequisite

feat(tenstorrent): host-free decode investigation — R1-R3b + architecture answer

record(tenstorrent): host-free row ACTIVE + drop stray bootstrap .i32

review(tenstorrent): address fresh-review findings on host-free branch

perf(tenstorrent): steady-state Qwen3 baselines — 7.3 default / 6.9 all-device

feat(tenstorrent): item 5 — zero-cache (capture-safe fill) + upload bisection

feat(tenstorrent): item 5 — rope cos/sin capture-safe (HIT measured on card)

docs(env): document VT_TT_HOST_FREE_DECODE + VT_TT_TRACE_DEBUG

fix(capi): hoist SpeechRegistry() out of extern "C" (main-red, aa3643b6)

spec(tenstorrent): RAC capture-safe design complete — the next step mapped

feat(tenstorrent): item 5 RAC — device branch + warm hook + shape-keyed cache

feat(tenstorrent): item 5 RAC — shadow priming works, device branch executes

feat(tenstorrent): item 5 RAC — all fixes landed; paged_update_cache not capture-safe

spec(tenstorrent): paged_update_cache IS in-place; writes from build_padded

fix(tenstorrent): sharded_zero crash — warm order (shadows before RAC idx)

feat(tenstorrent): item 5 RAC — approach (b) probe: capture passes RAC

feat(tenstorrent): item 5 PA — KV shadow skip works; next = metadata uploads

feat(tenstorrent): item 5 PA meta warm works; sdpa not compiled (cold bail)

feat(tenstorrent): CAPTURE COMPLETE — 86 tok/s replay (12x speedup)

feat(tenstorrent): RAC in-capture — paged_update_cache capture-safe

diag(tenstorrent): RAC k/v copy wrong output — TILE→sharded layout mismatch

spec(tenstorrent): session checkpoint — capture works, 83 tok/s

Commits on Aug 15, 2026

feat(tenstorrent): RAC flush approach — one-step lag diagnosed

diag(tenstorrent): in-region RAC — every input construction is a write

verify(tenstorrent): post-recovery — capture 87 tok/s confirmed on clean tt-metal

spec(tenstorrent): trace_region_size spike — conclusive