fix(BACKEND-TENSTORRENT-QWEN35): EnsureDevice2D consumed stale host bytes under host-free decode; Qwen3.5-0.8B sacred gate 16/16 on ambient

.agents/backend-matrix.md

@@ -247,6 +247,7 @@ this repository. State remains `ACTIVE`; no lifecycle transition is claimed.
 | `BACKEND-TENSTORRENT-RESIDUAL-GOLDEN` | Child of `BACKEND-TENSTORRENT` — the owed op-level numerics evidence at the residual-RMS device boundary (`kDeviceResidualMinRows == 32`): device path does `ttnn::add`+`ttnn::rms_norm` in bf16; host/CPU path accumulates in f32. Bot-flagged on #289; never measured at the boundary. | vllm.cpp CPU oracle `RmsNormKernel` mirrors vLLM `fused_add_rms_norm` (add in model dtype, variance in f32); `src/vt/cpu/cpu_ops.cpp:371-398` | `src/vt/tenstorrent/tenstorrent_ops.cpp:1067-1117` (host/device split, `kDeviceResidualMinRows=32`) | [test_tenstorrent_backend.cpp](../tests/vt/test_tenstorrent_backend.cpp) `kRmsNorm residual: device vs CPU f32 oracle across the rows=32 boundary`: 22/22 cases on real Blackhole P150. **Measured 2026-08-11:** host path `rows<32` bit-identical to CPU (`max_abs=0`); device bf16 path `rows>=32` diverges by constant **0.0459 abs** (1.9–2.6× rel on near-zero outputs) — bf16 rounding signature, not accumulation. Decision pending the e2e golden tie-break | [tenstorrent-residual-golden.md](specs/tenstorrent-residual-golden.md) | `SPIKE` | `CLAIM-BACKEND-TENSTORRENT-RESIDUAL-GOLDEN` |
 | `BACKEND-TENSTORRENT-MISTRAL` | Child of `BACKEND-TENSTORRENT` — allowlist `MistralForCausalLM` (Mistral-7B-v0.3: GQA 32/8, head_dim 128, plain rope theta 1e6, untied lm_head, full attention) on the TT platform + device-aware SACRED gate. Mistral reuses the Qwen3-dense forward verbatim (qk-norm skipped); every op already registered. No new kernel. | vLLM `mistral.py::MistralForCausalLM(LlamaForCausalLM)` (already ported to the shared dense machinery); gate pattern mirrored from `test_qwen3_paged_engine.cpp:221-296` | `src/vllm/platforms/tenstorrent.cpp:52-54` (allowlist) + `tests/parity/test_mistral_paged_engine.cpp` (device-aware wiring + Backend Proof) | **Gate PASSED on Blackhole P150 (2026-08-12):** [test_mistral_paged_engine.cpp](../tests/parity/test_mistral_paged_engine.cpp) 16/16 prompts PASS (12/16 strict-exact, 4/16 near-tie, 0 forward-divergent), max gap **0.062 nats**, BACKEND PROOF 0 declines (kMatmul selections=256 = untied lm_head on device, kPagedAttention=8192). Goldens `our_ids_tenstorrent.npy` + `neartie_gap_mnats_tenstorrent.npy` (transformers alternative-oracle; POL-ORACLE deviation recorded, same as Qwen3-0.6B TT precedent). Exit SIGSEGV 139 is the known MeshDevice teardown crash, not a gate failure | [tenstorrent-mistral.md](specs/tenstorrent-mistral.md) | `ACTIVE` | `CLAIM-BACKEND-TENSTORRENT-MISTRAL` |
 | `BACKEND-TENSTORRENT-GDN` | Child of `BACKEND-TENSTORRENT` — the GDN linear-attention op chain as native TT kernels, the hard prerequisite for the Qwen3.5/3.8 family (#1715): `kGdnPrefill`, `kGdnDecode`, `kL2Norm`, `kRmsNormGated`, `kCausalConv1dFwd`/`kCausalConv1dUpdate`, `kGdnStateGather`/`kGdnStateScatter`. The P150 is discrete, so an op miss refuses by name — the ops must land before any `Qwen3_5*` arch registration. Correctness oracle is our own CPU f32 arm (residual-golden precedent); no vLLM mirror exists for TT | Substrate: pinned tt-metal `ttnn::transformer::chunk_gated_delta_rule` (FLA chunked GDN forward, on-core recurrent state, `initial_state`/`final_state`) behind a varlen+state-permute adapter for `kGdnPrefill`; decode = rank-1 update composed from ttnn matmul+eltwise with a device shadow keyed by host pointer (`PagedKvShadow` pattern); contracts at `src/vt/ops.cpp:1823-2500`, CPU reference `src/vt/cpu/cpu_ops.cpp:1537-1740` | `src/vt/tenstorrent/tenstorrent_ops.cpp` (kernels + `RegisterOp` block `:3165-3199`) + `tests/vt/test_tenstorrent_backend.cpp` (op-level cases vs CPU f32 oracle, T-sweep, indexed-`state_idx` form, prefill↔decode state round-trip) | Spec committed 2026-08-22; implementation waves W1 (prefill set) then W2 (decode+state-I/O set) owed; lands production-unreached until the wiring row registers `Qwen3_5*` (named in `## Owed` per Nothing-lands-dead) | [tenstorrent-gdn.md](specs/tenstorrent-gdn.md) | `ACTIVE` | `CLAIM-BACKEND-TENSTORRENT-GDN` |
+| `BACKEND-TENSTORRENT-QWEN35` | Child of `BACKEND-TENSTORRENT` — the wiring row: `Qwen3_5ForConditionalGeneration` (dense text GDN hybrid) on the TT allow-list, the op delta the family forward refuses by name (`kGdnPostConv`, `kSigmoidGateBf16`, `kAttnQkNormRopeGate`, `kAttnQkNormRope` — pinned empirically by a W0 refusal sweep), and the first e2e gate. Makes the GDN row's ops production-reached (#1715 stays open until the family runs; GDN row lifecycle moves in the same change this lands) | Substrate: the GDN row's 8 kernels + the 27-op TT registry; e2e mirrors the ratified Mistral TT golden treatment (`VT_DUMP_IDS` bootstrap → `scripts/qwen3-neartie-gap-transformers.py` teacher-forced near-tie gaps → committed device-golden pair, POL-ORACLE deviation); op-level oracle is the CPU f32 arm. Capacity: `Qwen/Qwen3.5-0.8B` bf16 ≈1.6 GB fits (proven envelope Mistral-7B ≈14.5 GB); 27B bf16 ≈53.8 GB and GGUF k-quant arms (no TT kernels) refused by name, owed | `src/vllm/platforms/tenstorrent.cpp` (allow-list, lands last) + `src/vt/tenstorrent/tenstorrent_ops.cpp` (op delta) + `tests/parity/test_qwen35_paged_engine.cpp` (TT arm) | Spec committed 2026-08-23; W0 sweep, W1 op delta, W2 e2e, W3 GDN-reviewer leftovers (d2h counter completeness, `conv_transposed` fast-path check) owed | [tenstorrent-qwen35.md](specs/tenstorrent-qwen35.md) | `ACTIVE` | `CLAIM-BACKEND-TENSTORRENT-QWEN35` |
 | `BACKEND-TENSTORRENT-TRACE-RUNNER` | Child of `BACKEND-TENSTORRENT` — wire the landed graph-capture foundation (#354 / `59568772`) into a capturable forward region. Handoff §8/§9 "not done". **Decision (2026-08-13): NO-GO for pure T=1 decode capture — MEASURED, not assumed.** | CUDA is the ONLY backend with `SupportsGraphCapture()==true` (`cuda_backend.cu:184-240`); Metal/Vulkan `false`. Shared decode-graph framework `Qwen3DenseDecodeGraph` (qwen3.cpp:489, used by Qwen3/Mistral/Llama/InternLM2) gated on `support_static_graph_mode()` which TT does NOT override (base `false`) | `src/vt/tenstorrent/tenstorrent_backend.cpp:70-76` (capture surface, landed) + `tenstorrent_ops.cpp` `Trace*` (landed); NOT wired into a dense forward | **Measured on Blackhole P150 (2026-08-13):** Q1 no host-free region at T=1; Q2 all-device-at-T=1 costs 12.5→10.7 tok/s; Q2b capture attempt with both overrides flipped → ttnn **`TT_FATAL: Reads are not supported during trace capture`** (backtrace through `to_vector<float>`), `0 replays`. The T=1 forward does device→host readbacks that ttnn trace prohibits; flipping the two thresholds is insufficient. Requires a host-free `ForwardLayers` redesign, not a threshold tweak. Next: prefill capture (separate row) must first audit its `to_vector` readbacks | [tenstorrent-trace-runner.md](specs/tenstorrent-trace-runner.md) | `SPIKE` (decision record complete) | `CLAIM-BACKEND-TENSTORRENT-TRACE-RUNNER` |
 | `BACKEND-TENSTORRENT-HOST-FREE-FORWARD` | Child of `BACKEND-TENSTORRENT` — make the per-decode-layer forward region host-free (zero `to_vector` readbacks) so mesh-trace capture can run. Decomposes into R1 RmsNorm+RoPE all-device, R2 QkvSplit+ReshapeAndCache device, R3 PA decode device-resident metadata, R4 capture wire+measure. Prerequisite revealed by the trace-runner spike (capture aborts on `to_vector`). | CUDA decode-graph capture contract (`cuda_backend.cu:184-197`): captured region is async, no host sync, no malloc, fixed ptrs. TT must match: no `to_vector` between Begin/EndCapture | `src/vt/tenstorrent/tenstorrent_ops.cpp` (RmsNorm:1067, PreferDeviceRope:1344, QkvSplit:1460, ReshapeAndCache:1527, PagedAttention:2009) | **R1-R3b MEASURED on P150** (env-gated `VT_TT_HOST_FREE_DECODE`, inert by default; 23/23 TT tests incl. a dedicated default-path inertness guard, M1-mutation-proved): R1 threshold flip landed; R2 device->device copy (`CopyDeviceDeviceIfCapture`, ttnn::copy+empty) landed; R3 program-cache warm (`enable_program_cache` + eager-warm) landed; R3b device zero-fill (`MemsetDeviceIfCapture`) landed. Capture enters the forward and reaches layer ops (CastBf16/RmsNorm fire). Remaining item-5 blocker = per-op enqueue_write; answer = persistent device tensors + before-replay populate (tt-metal vLLM plugin design). Full blocker map + architecture in [tenstorrent-host-free-r1.md](specs/tenstorrent-host-free-r1.md). **R5 (2026-08-21, #1604): host-free decode is the DEFAULT** — `HostFreeDecodeEnabled()` centralizes the polarity (exact `0` = pre-flip opt-out), both device golden pairs re-dumped and re-adjudicated under the new default (Qwen3 max 375 mnats, Mistral max 250 mnats, 0 outside top-K), both paged-engine gates 16/16 green (125/125 + 128/128 assertions), default leg 10.94-11.06 tok/s vs 5.34 opt-out (2.1x, Qwen3-0.6B b1). Capture declined by default: multi-request captured hangs (#1625); TT async scheduling stays off, no `SupportsAsyncSampledTokenReadback` override (#1627) | [tenstorrent-host-free-forward.md](specs/tenstorrent-host-free-forward.md) | `ACTIVE` (R1-R3b + R5 default flip gated on `row/BACKEND-TENSTORRENT-HOST-FREE-1604`; capture declined per #1625, TT async readback owed per #1627) | `CLAIM-BACKEND-TENSTORRENT-HOST-FREE-FORWARD` |
 ## Native competitor and performance gates

.agents/claims/CLAIM-BACKEND-TENSTORRENT-QWEN35.md

@@ -0,0 +1,5 @@
+# CLAIM-BACKEND-TENSTORRENT-QWEN35
+
+| Claim | Row IDs | Agent | Worktree / remote dir | Branch | Owned scope | State | Last update |
+|---|---|---|---|---|---|---|---|
+| `CLAIM-BACKEND-TENSTORRENT-QWEN35` | `BACKEND-TENSTORRENT-QWEN35` (`ACTIVE`, new) | Maki (zai-glm-5.3), helper role per `scripts/agent-role.py`, working from the committed spec `.agents/specs/tenstorrent-qwen35.md` | isolated worktree `/home/lu_zero/Sources/vllmcpp-tt-qwen35`; real Blackhole P150 (`thalia`, local board, `flock` mutex per GPU rule — not an `rc` fleet device) | `row/BACKEND-TENSTORRENT-QWEN35`, base `origin/main` `175733000`, issue [#1715](https://github.com/mudler/vllm.cpp/issues/1715) (tracks the wiring row per the GDN spec's `## Owed`) | Owns ONLY: the `Qwen3_5ForConditionalGeneration` TT allow-list entry (lands last), the W0 refusal sweep that pins the op-delta scope, the W1 op delta (`kGdnPostConv`, `kSigmoidGateBf16`, `kAttnQkNormRopeGate`, `kAttnQkNormRope`, plus whatever the sweep adds), the W2 e2e TT golden-pair gate for `tests/parity/test_qwen35_paged_engine.cpp` + the `docs/USAGE.md` weights entry (`Qwen/Qwen3.5-0.8B` @ `2fc0636471`, download authorized 2026-08-23), the W3 GDN-row reviewer leftovers (d2h counter completeness for `EnsureGdnCacheDevice` slow path + `CommitConvTransposed` fallback; the `conv_transposed` fast-path check), the spec, the backend-matrix row, and the two checker re-pins. EXCLUDES: MoE arches, GGUF k-quant arms (refused by name, owed), the VL tower, 27B-on-TT (no fitting arm), `src/vt/cpu/` (the oracle stays untouched), `src/vt/cuda/`, #1625/#1627 | `ACTIVE` | 2026-08-24 — W2b: pool-tenancy defect fixed (mutation-pinned), gemma `+1` defect fixed (mutation-pinned, op-level case added); ambient collapse healed — fixed engine matches the oracle on p0 and sits near-tie on 247/256 steps. Open: 2b (six out-of-band steps, leg-independent, intrinsic; layer bisect needs a GDN dump hook). Goldens in worktree describe the fixed engine. Nothing committed yet |