Current ROCm MTP status Production patch status (`eaman` branch) Use for the validated F16/off configuration: --pipeline-parallel off --cache-type-k-draft f16 --cache-type-v-draft f16 The current cumulative production patch is: bug/latest_rocm_improvement_c6824a9e.patch It contains the complete Eaman ROCm fitting/pipeline work, MTP Compact Rollback, ordered MTP multi-ubatch execution, HIP VEC control, and the ported rdna-boosts blocks 09/10 (meta-buffer compute headroom for recurrent n_rs_seq views; Q4_K/Q5_K/Q6_K/Q8_0 MMVQ VDR decode kernels). VEC forcing defaults to off and requires: --hip-fa-force-vec on The standalone mainline MTP Compact Rollback patch is: bug/mtp_compact_rollback_c6824a9e.patch Adaptive MTP is included in both current patches and remains opt-in with `--spec-draft-adaptive` (default off). The experimental adaptive p_min gate (`--spec-draft-p-min-adaptive`) was abandoned: at deep context it was ~14% slower than static 0.75 with ~2× the MTP-replay rate. Keep static `--spec-draft-p-min 0.75`. Archived as `old/patches/spec_draft_p_min_adaptive_eaman_9ee65ec3.patch` (plan/record: `old/adaptive_pmin_plan.md`); the `eaman` working tree was reverted to committed `9ee65ec3`. The controller now starts at n-max and changes its ceiling only after drafts that actually reach that ceiling; p-min confidence stops do not cause backoff. The focused parser/recurrent tests pass. A matched Vulkan n-max=7, p-min=0.75, n-min=0 run produced identical fixed/adaptive draft and replay counts at 30.76 and 30.73 t/s. `/home/eaman/llama/bin_vulkan` is now cumulative build 1493 / `9ee65ec3` (deployed 2026-09-10): the current `eaman` tip, same content as the canonical patch. The preserved prior deployment at `/home/eaman/llama/bin_vulkan_previous` and the canonical `llama.cpp/build/bin` are stale inconsistent copies that segfault; do not reuse them. The VEC Q4/Q4 MTP run initially reached 147,712 context tokens and completed startup plus short generation. Its former open-ended runtime follow-up is archived under `old/` and is not a current refresh acceptance criterion. Validation notes: - 2026-09-16: mainline advanced to `c6824a9e`; merged into `eaman` as `ddddf03d`. Recovery branch `eaman-pre-latest-c6824a9e` preserves the preceding tested tip `9ee65ec3`; `eaman-pre-latest-73a43d1` (tip `7f07026c`) verified as ancestor and pruned. Conflicts: upstream renamed `ggml_flash_attn_ext_set_prec` to `ggml_prec_set_acc` (kept the Eaman `set_force_vec` call, slot 5) in `llama-graph.cpp` and `minimax-m3.cpp`; upstream extended its own `test-recurrent-state-rollback.cpp` (cache fill, Kimi-K3, multi-seq split replay), so the Eaman tests were layered on as pure additions with `make_ctx` renamed to `make_ctx_rs`. HIP (188 objects, incl. fattn-vec) + Vulkan compiled and linked: build 1676 / `ddddf03d`, ROCm core-10.0 / HIP 7.15.26333, gfx1030;gfx1031. Logs: logs/build_c6824a9e.log. Patch applies cleanly to detached `c6824a9e` without `--3way` and exactly reproduces the `eaman` tree `6478afb9`. No binary deployed; `/home/eaman/llama/bin_vulkan` remains 1493 / `9ee65ec3`. - 2026-09-16: standalone `eaman-mtp-compact-rollback` merged to `c6824a9e` as `2586df93`; recovery branch `eaman-mtp-compact-rollback-pre-c6824a9e` preserves `2fe7df5f`. Only conflict was the same upstream test-file extension; resolved identically (Eaman tests as pure additions, `make_ctx_rs`). Build 1658 / `2586df93` compiled HIP + Vulkan and linked in the standalone worktree `build-standalone`; `test-arg-parser` passed. Full recurrent-model rollback tests not re-run (no recurrent/hybrid model locally). Patch applies cleanly to detached `c6824a9e` and reproduces the standalone tree `a9fb7032`. Log: logs/build_standalone_c6824a9e.log. - Latest mainline `73a43d1f` was merged into cumulative `eaman` as `9570cc08`; recovery branch `eaman-pre-latest-73a43d1` preserves the preceding tested tip `7f07026c`. Same content as build 1447, rebased with zero conflicts. - The cumulative HIP/Vulkan build 1490 compiled and linked with ROCm core-10.0 / HIP 7.15.26333 for `gfx1030;gfx1031`. - rdna-boosts blocks 09/10 were cherry-picked from the RDNA3 7900 XTX optimization fork (original author Stew Forster) and merged into `eaman` as `9ee65ec3`. Worktree branch `eaman-rdna3-port` is kept for further candidate ports. Build 1492 compiled and linked with ROCm core-10.0 / HIP 7.15.26333 for `gfx1030;gfx1031`. - A/B on Qwen3.8-27B-UD-Q6_K_M, seeded synthetic MTP acceptance (`--spec-synth-len 4.5`): TG 24.27 -> 25.54 t/s (+5.2%), overall draft acceptance 0.733 -> 0.734 (flat), verify duration unchanged (0.534 ms), no new warnings. Logs: /tmp/llama.std (build 1282), /tmp/llama.rd (1492). - The regenerated cumulative patch applies cleanly to detached `73a43d1` and its applied diff exactly matches the `eaman` tree (tree `489357af`). Its SHA-256 is `1c67b3d62ab618a8cec4771618b5e9e8f8ad05d034dd66f907944bee42d1cab9`. - Obsolete branches pruned 2026-09-06: `eaman-mtp-compact-rollback-pre-662a0b01`, `eaman-mtp-cr-adaptive`, `sol-77918ca`, `eaman-rs-snapshot-depth`, `eaman-rs-snapshot-depth-pre-8144f31`, `fix2`, `latest-patched-d2f8305-recovery`, `eaman-mtp-compact-rollback-pre-3737e41`. Worktrees for `eaman-mtp-cr-adaptive` and `eaman-rs-snapshot-depth` removed. - Latest mainline advanced again to `de8656bd94f1163188125542534e4bcbc9f9fb1f`; this is now historical. Standalone MTP CR was merged onto that exact base as `f1b4a96f`; recovery branch `eaman-mtp-compact-rollback-pre-de8656bd` preserves `cdc3d024`. - The standalone CPU/default build completed. HIP/Vulkan build 1434 compiled and linked with pinned ROCm core-7.14 / HIP 7.14.60850 for `gfx1030;gfx1031`, enumerated both GPUs, and passed the parser, generated-model, dense, Nemotron-H, and DSV4 rollback tests. Cumulative-only pipeline and HIP VEC controls are absent as intended. - `mtp_compact_rollback_de8656bd.patch` applies cleanly to detached `de8656bd` and exactly reproduces the standalone delta. Its SHA-256 is `ff1eec3c4d0fc32812322a5a11fa4d78ceba3100456eaeb24af9300934162106`. - Latest mainline `0df017d6dd246edb1d06577976c759cf7a3c50d6` was merged into cumulative `eaman` as `7f07026c`; recovery branch `eaman-pre-latest-0df017d6` preserves the preceding tested tip `886fda24`. - The cumulative HIP/Vulkan build 1447 compiled and linked with pinned ROCm core-7.14 / HIP 7.14.60850 for `gfx1030;gfx1031`. Both ROCm and Vulkan GPUs enumerated; the parser, MTP multi-ubatch sync, generated-model, and all three recurrent rollback tests passed. The focused TOP_K suite passed 525/525 cases on each RDNA2 ROCm GPU. - The cumulative patch applies cleanly to detached `0df017d6` and its applied diff exactly matches the `eaman` tree. It has SHA-256 `ff848e448498fd8fdcb9c8817e975959d8ce22fc6a5c5be4ff4da4fe75e90e20`. - Upstream added ROCm radix TOP_K for long rows, concurrent streams per split, sparse FA, and fused MoE kernels. Sparse FA took op-parameter slot 4, so HIP VEC forcing was moved to slot 5 during the merge. - The preceding build at `886fda24` completed the user runtime check at 124,055 context tokens plus 2,160 generated tokens with MTP active, 0.90974 draft acceptance, and no OOM, NaN, decode error, or hang. - Latest mainline `662a0b0121a53c23b825a71e64ab6eff59b7f4d8` was merged into cumulative `eaman` as `9ee6674f`; this is now historical. - Standalone MTP CR was merged onto the same base as `cdc3d024`; recovery branch `eaman-mtp-compact-rollback-pre-662a0b01` preserved `6cb89357` (deleted 2026-09-06). - Obsolete cumulative recovery branches through `eaman-pre-latest-662a0b01` were verified as ancestors and pruned. Standalone recovery branch `eaman-mtp-compact-rollback-pre-3737e41` was retained because its tip is not an ancestor of the current standalone branch (deleted 2026-09-06 as superseded). - The standalone MTP CR patch applies cleanly without three-way fallback and produces exact tree `89b4b2266f368b5deaeb79cd398db88ef0e6de47`. - The cumulative patch applies cleanly without three-way fallback and produces exact tree `31dfa8e080bb3c3657979c59675661473ae73a90`. - Pinned ROCm core-7.14 / HIP 7.14.60850 and isolated ROCm core-10.0 / HIP 7.15.26333 combined builds both compiled HIP and Vulkan, linked `llama-server`, enumerated both GPUs, passed `test-arg-parser`, and passed all four registered generated-model/recurrent tests on both exact branches. - The fresh standalone binary is build 1362 / `cdc3d024`; the fresh cumulative binary is build 1376 / `9ee6674f`. The canonical cumulative source build remains the pinned 7.14 build; the separate 10.0 build is `build-rocm100`. - A build-1279 ROCm1 runtime check at fixed 32K context and MTP CR depth 1 generated 1,024 tokens and completed 11 deep-replay events / 25 replayed tokens. The preceding auto-fit startup also reserved the exact 50.25 MiB device checkpoint and reached listening at 241,408 fitted context. - Build 1282 / `527be7e6` is deployed at `/home/eaman/llama/bin_vulkan`. Local launchers use `--spec-mtp-cr-depth`; no retired public rollback flag remains there. - `test-arg-parser` passed. Dense Q6_K_L joint fitting selected 136,704 context at the same reported starting VRAM as the historical 136,448 matched trial, then completed 64-token MTP generation at 27.03 t/s with 41/44 drafts accepted. The separate historical 137,984 startup had 32 MiB more free on ROCm0, and its rendering/session mode is not established. - MoE Q5_K_S joint fitting measured the MTP extra context at every candidate without target-only fallback, selected 233,216 context, reached listening, and completed 64-token generation at 76.28 t/s with 43/47 drafts accepted. Automatic tensor-level placement remained active (25 ROCm0 layers with one partial UP layer, plus 17 full ROCm1 layers). - Target and MTP pipeline modes were disabled in both runtime checks. Runtime `auto` and `on` were not repeated during this refresh. - The source master branch was not modified; the validation worktree was removed. Historical d2f8305 IQ4_XS single-16GB-GPU fit results: patched ROCm: 76,544 mainline ROCm: 22,272 patched Vulkan: 79,616 mainline Vulkan: 69,120 Matched patched-ROCm MTP cache test on 2026-08-10: Q4_0/Q4_0 draft KV: 76,544 F16/F16 draft KV: 68,864 Q4_0 gain: 7,680 tokens (+11.15%) Both reached the listening state. This confirms a meaningful context-capacity advantage for Q4_0 on this supported HIP VEC model/shape, but generation, throughput, acceptance, delayed-OOM, and long-context stability remain untested. Matched current Q6_K_L dual-GPU ROCm cache test using think.sh_test: Q4_0/Q4_0 draft KV: 153,856 context, 333.00 MiB MTP on ROCm1 F16/F16 draft KV: 153,856 context, 764.70 MiB MTP on ROCm1 Q4_0 saved 431.70 MiB on ROCm1 but gained no context because the fixed per-device layout retained the same limiting 256-token-rounded fit. The cache choice is therefore model/topology dependent: meaningful context gain on the single-GPU IQ4_XS test, none on this dual-GPU Q6_K_L test. The Q6 Q4 ROCm1 headroom can be converted into context with a fine-grained target tensor override while retaining --tensor-split 0.6,0.4: -ot '^blk\.39\.ffn_(up|gate|down)\.weight$=ROCm1' This reached 165,632 context, +11,776 (+7.65%) over the 153,856 Q4 baseline. A whole-layer 0.59/0.41 split was worse at 145,152. The override has completed startup only; throughput, acceptance, delayed-OOM, and long-context stability remain pending. The 2026-08-11 Q8_0/Q8_0 target-KV recalibration of think.sh_test selected the same boundary FFN triplet after measuring it independently: -ot '^blk\.39\.ffn_(up|gate|down)\.weight$=ROCm1' With matched free VRAM, no override fitted 126,720 and the selected override fitted 136,448 (+9,728, +7.68%) while retaining 3 target scheduler graph splits. Adding attention Q/output reached 139,008 but raised those splits to 7, so it was not selected without throughput validation. A clean final-launcher repeat fitted 137,984 and reached the listening state. Detailed measurements: bug/logs/q6_kl_q8_target_tensor_balance_20260811.log The helper was also applied to /home/eaman/models/think.sh_small with the Q6_K model. Its original Q4 and F16 layouts both fitted 189,440 context even though Q4 saved 531.61 MiB of MTP allocation on ROCm1. The max-context override is: -ot '^blk\.39\.(ffn_(up|gate|down)|attn_(q|output))\.weight$=ROCm1' It fitted 205,312 context, +15,872 (+8.38%), and passed startup, health, and one short MTP request. It increases target graph splits from 35 to 39. The FFN-only alternative reaches 201,216 while retaining 35 graph splits: -ot '^blk\.39\.ffn_(up|gate|down)\.weight$=ROCm1' Detailed measurements: bug/logs/q6_k_small_rocm_tensor_balance_20260810.log All four reached the server listening state. Against the corrected historical results, patched ROCm changed from 76,032 to 76,544 and patched Vulkan from 78,592 to 79,616. The patch remained necessary on that tested mainline. Current-mainline patches: bug/latest_rocm_improvement_c6824a9e.patch base c6824a9e, cumulative eaman ddddf03d, tested llama.cpp version 1676 SHA-256 ab3ee93974930b239fd62aa61c2bfc1926e8dc3e5ba6e31872e453d4e2a7a4cf bug/mtp_compact_rollback_c6824a9e.patch base c6824a9e, MTP CR 2586df93, tested llama.cpp version 1658 SHA-256 f9894a567a2d6cd930366bb6bc634cf64bbce080dceb42e447f03aa6c676654c Previous (now historical): bug/mtp_compact_rollback_73a43d1.patch base 73a43d1, MTP CR 2fe7df5f, same content as build 1434 SHA-256 82a9b537e04074e17f7c1caa5c908ade9d4d2f10800bae5110bbcdbf4e9acf7a Previous (now historical): bug/mtp_compact_rollback_de8656bd.patch base de8656bd, MTP CR f1b4a96f, tested llama.cpp version 1434 SHA-256 ff1eec3c4d0fc32812322a5a11fa4d78ceba3100456eaeb24af9300934162106 bug/latest_rocm_improvement_73a43d1.patch base 73a43d1, cumulative eaman 9570cc08, tested llama.cpp version 1490 SHA-256 8149508206fa581f8af17811d7379c3ca0e2de7184065c5f72ae6ab1730caabb Previous (now historical): bug/latest_rocm_improvement_73a43d1.patch base 73a43d1, cumulative eaman 9ee65ec3 (includes rdna-boosts blocks 09/10), tested llama.cpp version 1492 SHA-256 1c67b3d62ab618a8cec4771618b5e9e8f8ad05d034dd66f907944bee42d1cab9 Previous (now historical): bug/latest_rocm_improvement_0df017d6.patch base 0df017d6, cumulative eaman 7f07026c, tested llama.cpp version 1447 SHA-256 ff848e448498fd8fdcb9c8817e975959d8ce22fc6a5c5be4ff4da4fe75e90e20 For Vulkan, initially retain: --pipeline-parallel auto Build 1180 controlled Qwen3.6-35B-A3B comparison at 32,768 fixed context: - ROCm F16: pipeline auto 70.38 t/s versus off 69.67 t/s; identical 96.283% acceptance. The 1.02% generation difference is inconclusive from one matched request. - Vulkan auto: Q4_0 41.45 t/s versus F16 39.00 t/s (+6.27%); acceptance was 97.133% versus 96.198%. Q4_0 is provisional for this model pending repeats and fitted-context comparison. - Verbose startup confirmed target pipeline enabled under ROCm auto and disabled for MTP. - Existing MoE n-max testing fitted 203,776 context with 4 versus 195,584 with 5. Four has the better capacity tradeoff, but the active launcher retains 5 for its small potential long-context speculative benefit. - Raw logs: bug/logs/controlled_*_20260820.log; earlier n-max notes: /home/eaman/launch/bench/moe_large. - Build 1180 Qwopus3.5 9B measurements confirmed recurrent state = 50.25 MiB * (1 + n_rs_seq): 50.25 MiB at 0, 100.50 MiB at 1, and 251.25 MiB at 4. - The operational rollback path produced byte-identical 128-token greedy output to target-only generation at both n-max 1 and 4. The n-max 1 run rejected 18 of 72 drafts; n-max 4 rejected 100 of 181. The server `PARTIAL_ONLY` path is validated from the minimum useful depth through the active depth. Historical pre-integration MTP rollback validation (`eaman-rs-snapshot-depth` branch, deleted 2026-09-06): - The opt-in RS branch is rebased onto production `a4e47992` at `fff1fc7c`; production `eaman` remains unchanged. CPU and combined HIP/Vulkan build 1249, parser tests, and registered Qwen3.5/Nemotron-H rollback tests pass. - Matched dense depth 3/1 fitted 136,704/149,504 context (+12,800, +9.36%); the seeded 192-token outputs were byte-identical at 24.79/24.12 t/s. - Matched MoE depth 4/1 fitted 233,216/262,144 context (+28,928, +12.4%). Depth 1 generated 192 tokens at 71.30 t/s with 2 replay events / 5 tokens. Real-GPU MoE outputs diverged late even for target-only generation, so long output byte identity is not used as a rollback-correctness criterion. - Direct `-np 2` startup captured a 100.50 MiB incremental fitter reservation and a 201.00 MiB final two-sequence RS buffer. Two concurrent 192-token requests completed without OOM/cross-talk and both exercised replay. - A 1,024-token 9B depth-1 run completed at 62.94 t/s with 523/612 accepted drafts and 24 replay events / 69 replayed tokens. - A controlled 60,550-token ROCm MoE prefill sweep at fixed 224,256 context measured microbatch 256/512/750 at 383.02/385.33/372.27 t/s; 1,024 failed with `CUBLAS_STATUS_INTERNAL_ERROR`. Microbatch 512 is the best tested value, but improves only 3.51% over 750. - At fixed 120,064 context and microbatch 512, F16/F16 target KV reached 835.97 prompt t/s versus 385.13 for Q8_0/Q8_0 (2.17x). This isolates the major ROCm long-prefill loss to the patched policy that forces supported quantized-KV HIP shapes, including wide prompt batches, onto VEC to avoid upstream's large F16 conversion workspace. F16 takes the faster large-query MMA path but remains diagnostic because its context-capacity cost was not measured. - An unrestricted microbatch-512 fit selected 262,144 but OOMed on the first prompt in `launch_fattn<256,2,1>`; do not treat the model-ceiling fit as safe for the active batch-2,250/n-max-5/target-70 launcher. Microbatch 256 did complete at that ceiling. Q8 with flash attention off failed during pre-fit context creation. - The F16 target-KV `moe.sh_opti` profile selected 151,552 at target 70 and twice OOMed after a 60,550-token prefill while allocating its first on-device recurrent replay checkpoint. The old RS fitter proxy split 62.81 MiB as 43.97/18.84 MiB, under-reserving limiting ROCm1. Commit `37c29f3c` measures the exact backend layout at 37.69/25.12 MiB; commit `fff1fc7c` now allocates and pins those buffers for every slot before evaluation and makes host fallback sticky for the run. Final production-equivalent target-70 startup retained 151,552 and reached listening with the reservation allocated. The older 140,288 startup used target 200 and was not comparable. Focused CPU and HIP/Vulkan tests pass; the user will combine the long-prompt confirmation with performance testing. Detailed issue record: `rs_checkpoint_fit_reservation_oom.md`. - In this historical build-1249 comparison, `--hip-fa-force-vec on|off` defaulted to `on`; the current build defaults to `off`. The historical `off` run resumed upstream HIP dispatch for Q8_0 KV fast-kernel/workspace testing. - The matched Q8_0 target-KV MoE test fitted 224,256 context with forced VEC versus 163,072 with `--hip-fa-force-vec off` (-27.3%). `off` processed the 60,550-token prompt at 973.90 t/s versus 361.95 t/s (2.69x), then generated 3,383 tokens at 76.21 t/s with 85.17% draft acceptance and 27 replay events / 208 replayed tokens. It completed without checkpoint-allocation failure and is the preferred tested MoE compromise when prompt speed matters more than maximum context. This is one successful long run, not a repeated stress test. - The complete historical build 1249 directory was formerly deployed at `/home/eaman/llama/bin_vulkan`; the current deployment is build 1282. Its preceding deployment was temporarily backed up at `/tmp/bin_vulkan_pre_1249`. - Current experimental exports: `latest_rs_cumulative_8144f31.patch` (SHA-256 `339317ad294c33ba6e9c28f2a64cc839c69d899140202586c41de7abd8554209`) and `mtp_rs_only_8144f31.patch` (SHA-256 `902eaa2424444fc3afc414b8ca68f83c0dbf5bc4681fb8f95c06ffd466feea59`). Both cleanly apply to `8144f31`; the cumulative tree exactly matches the RS branch. Historical `a3b1eff` exports remain pinned historical artifacts. - Remaining RS follow-up: cache repeated scratch-fit measurements and perform near-limit long-prompt/delayed-OOM testing before considering production merge. - Separate findings remain: the dedicated test fails only when a full checkpoint is loaded over a dirty hybrid context, and MTP n-max 0 aborts on its first request. Neither affects normal n >= 1 MTP or the validated server `PARTIAL_ONLY` path. - A matched RX 6800 Qwopus3.5 9B test found no benefit from `--no-spec-draft-backend-sampling`: disabled versus enabled changed mean generation from 66.98 to 66.13 t/s and median from 65.43 to 65.22 t/s, with byte-identical output and identical 95.187% draft acceptance. The enabled path reported unsupported ROCm TOP_K and fell back; the measured overhead was below run-to-run noise. Continue comparing Q4_0 and F16 draft KV per model and monitor fitted context, pipeline mode, buffer sizes, allocation failures, throughput, draft acceptance, and long-prompt stability. Author and maintainer: eaman. Historical Qwen and Sol labels identify LLM-assisted development series, not patch authorship. Historical notes and logs are under bug/logs/; archived patches are under bug/old/patches/. Manual per-model dual-GPU placement procedure: bug/dual_gpu_context_balancing_guide.md 2026-08-18 MoE MTP auto-fit result: - Unpatched `--fit-target 50` selected 235,008 context but OOMed on a 12,309-token prompt because provisional MoE placement contaminated later fits. - `mtp_fit_scratch_placement.patch` isolates estimation placement from the final fit. Patched target 50 selected the same 235,008 context and passed both a 32,327-token prompt plus generation and a separate 12,308 + 64-token run. - All four patched fit passes succeeded; HIP/Vulkan server compiled and linked. - The fix is committed on source branch `eaman` as `773364c` and included in the current standalone patch; the validated HIP/Vulkan build is deployed to `/home/eaman/llama/bin_vulkan`. - `moe_tester.sh` now uses no explicit context or placement, target 50, and MTP n-max 4. Binary and launcher rollback copies use the suffix `.pre-moe-fit-20260818`. - A later production run reached 96,484 cached tokens plus at least 6,632 new tokens, then SIGSEGV occurred inside ROCm `hipGraphExecUpdate`, not in the fitter or checkpoint path. `moe_tester.sh` now exports `GGML_CUDA_DISABLE_GRAPHS=1` only for this MoE process; dense launchers are unaffected. - The no-HIP-graphs regression test retained 235,008 fitted context, ingested 32,318 tokens, generated 4,718 tokens, restored the prompt checkpoint, then generated a forced 7,000 tokens. It completed without a crash at 41.73 and 47.31 t/s respectively. Full-context stability remains pending. - Production validation then completed 10,385 generated tokens at a 104,837-token slot, followed by checkpoint restore, a 143,851-token prompt, and 1,698 tokens generated at 30.26 t/s. The server released cleanly at 145,550 slot tokens. - Final target-50 stress validation restored a 197,327-token checkpoint, processed a 203,372-token prompt, generated 11,145 tokens at 25.78 t/s with 99.252% draft acceptance, and released cleanly at 214,518 slot tokens. No OOM or segfault occurred; this is 20,490 tokens below the fitted 235,008 limit.