⚠ This notebook has stopped taking notes: newest record is 23 days old (eng-0280), against an expected cadence of 14 days.
clm-0105
communitymed ●●○superseded
citable URL: https://halobench.com/records/clm-0105/ — this address never moves; the anchor /records/#clm-0105 keeps resolving
SUPERSEDED (operative claim) as of strix-halo-llamacpp v0.6.10. As written against v0.6.9 (2026-08-22): on hybrid GDN (qwen35moe) targets the then-stable Strix Halo Vulkan fork did NOT guarantee token-exact MTP rollback after a rejected draft. v0.6.8 had introduced a one-line change forcing MTP rollback through full sequence-state checkpoints; v0.6.9 (2026-08-22T04:09:43Z) reverted that line because the full-checkpoint save/restore path deadlocked deterministically on Vulkan with a hybrid target (Qwen3.6-35B-A3B stalled a few hundred tokens into a long response, all threads parked in futex_do_wait, GPU idle, GTT flat). With the revert, rollback returned to the v0.6.4-v0.6.7 fast snapshot-plane restore, which the vendor stated "can diverge slightly from a no-draft run after a rejected draft" - a subtle distribution drift after rejection, not garbled output. That was a deliberate availability / exactness tradeoff by the fork, pending a state-save-path fix. v0.6.10 (clm-0107) re-lands token-exact full-checkpoint rollback; the operative tradeoff claim above no longer holds on the current stable fork.
Historical fourth arm of the MTP-correctness dossier (with con-0009/clm-0107). Fork-vendor release-note context only: single-payload verification on gfx1151 Vulkan, speed figures explicit single runs NOT the BENCHMARKS.md protocol, so nothing here inherits into a HaloBench number. The v0.6.9-era exactness tradeoff is superseded by v0.6.10 (token-exact rollback restored via 9c5d899 + f25eefe). Reassessment outcome (see clm-0107): the house varied-prompt/raw-token MTP-invariance requirement does NOT relax - the v0.6.9 tradeoff was a fork-rollback-path issue, whereas the #25618 quantized-target divergence (clm-0094/clm-0104/clm-0106) and #26750 spec-path collapse (clm-0102) are separate upstream verify/target-side issues that still stand. And HO-009 plain-vs-n2 results (clm-0097, clm-0101) were measured on UPSTREAM ggml-org/llama.cpp 7077abb on ROCm0/gfx1151 - which never used the fork's snapshot-plane rollback path - so this tradeoff never directly applied to those numbers and v0.6.10 changes no comparison boundary there.