TT P150 at post-W2c main: the host-hybrid opt-out outperforms the host-free eager DEFAULT by 1.24x — #1604 flip premise inverted

[lu-zero]

What changed

#1604 flipped VT_TT_HOST_FREE_DECODE default-ON on 2026-08-21 after measuring
default 10.94–11.06 tok/s vs opt-out 5.34 (Qwen3-0.6B, b1, greedy,
thalia P150, base 52e328789). Re-measured 2026-08-26 at 21fe11cf1
(origin/main, post-BACKEND-TENSTORRENT-QWEN35 W2a–W2c):

arm	median tok/s	n	range
default (host-free eager)	10.822	12	10.51–11.03
VT_TT_HOST_FREE_DECODE=0 (host-hybrid)	13.369	12	13.09–13.55 (one outlier leg 11.49)

Method: order-alternated pairs ×3 (--repeat 5, in-process run 1 discarded),
one $HOME/gpu.lock hold, tt-smi -r first, box otherwise idle.
The opt-out wins 1.24×. The default arm is unchanged against its
2026-08-21 figures (within the band given no clock sampling — see caveats),
so the movement is ~2.5× improvement in the OPT-OUT arm since then,
mechanism UNATTRIBUTED. Next traceable step: per-op delta of the host-hybrid
path between b86e3705f and 21fe11cf1.

Why it matters

The shipped default now serves the slower of the two eager arms. Any gate
that uses the default as denominator inherits the slower figure.

Caveats stated plainly

• No clock window was sampled (tools/bench/gpu_clock_state.py is
  NVIDIA-only); all figures here are clock-unattributed. Ordering was
  alternated to cancel drift, but SM/board-clock attribution is owed.
• One model shape (b1, 96-token completions), one board (Blackhole P150,
  aarch64 host), tt-metal HEAD a3d33028975 (carries our copy_default_tilized
  patches).
• The captured opt-in arm (27.1 tok/s single-request on the old tree) is NOT
  retested; multi-request capture still hangs (Tenstorrent captured decode hangs on the first multi-request run (single-request captured legs work) #1625), and TT async readback
  remains TT backend never advertises SupportsAsyncSampledTokenReadback: async scheduling resolves OFF, test_qwen3_dense_async_serving FATALs on every cached checkpoint #1627.

Ownership

Row BACKEND-TENSTORRENT-HOST-FREE-FORWARD; recorded under the row spec's
## Owed in the same change that lands the refreshed records.

[lu-zero]

Re-adjudicated through #2005's clock harness (bench/tt-clock-state, P150 thalia): default host-free eager median 10.880 tok/s vs VT_TT_HOST_FREE_DECODE=0 median 13.645, ratio 1.254 (12 warm samples/arm, order-alternated ×3).

Clock attribution is now POSITIVE rather than absent: every busy sample of both arms sits at exactly 1350 MHz (busy-slice windows PASS all pairing rules, cross-arm offsets 0/0%), so the inversion is a genuine hybrid-path performance advantage, not a clock excursion. Owed next step unchanged: per-op delta of the host-hybrid path from b86e3705f to 21fe11cf1.