Skip to content

About

Stock Oh My Pi on a stock, local Strata inference server: one Windows 11 + RTX host, every component hash-pinned, loopback behind an API key, and real-host acceptance evidence for every claim. Nothing forked, no proxy.

alphastorm.github.io/omp-strata/

Topics

ai-agentcoding-agentcoding-assistantcudadeveloper-toolsgguflocal-ailocal-llmnvidiaoh-my-piopenai-compatible-apipythonqwenqwen3rtx-3090rtx-4090rtx-5090self-hostedstratawindows

Resources

Readme
MIT license

Code of conduct

Code of conduct

Contributing

Contributing

Security policy

Security policy
Activity

Stars

2 stars

Watchers

0 watching

Forks

0 forks
Report repository
BranchesTags

Repository files navigation

OMP Strata

Stock Oh My Pi on a stock, local Strata server. Nothing forked, everything pinned.

Current, qualified: stock OMP 18.18.3 with stock Strata v0.2.13 across the RTX PRO 6000 (four batch slots with per-slot stock DFlash2 drafters at window 8 + PLE-direct + KV in VRAM + stock parallel and GDN stacks; tuned-DFlash2 v0.2.13 alternate at --draft-conf 0.55, v0.2.12 DFlash2 rollback), the RTX 5090 (262K Coder IQ1_M default with KV in VRAM + KV growth + GDN stack, 131K Coder IQ1_M alternative, and full-model IQ3_S 131K low-RAM resident alternate; v0.2.12 262K envg and 131K rollbacks), the RTX 4090 + RTX 3090 pair (stock DFlash2 drafter with strata.drafting_args --draft-conf 0.5 + split-24 stage-weight trim + PLE-direct + engine sm_86 table, GDN stack and STRATA_SELECT_TWO_PASS=1; v0.2.12 env86gt two-pass DFlash2 rollback) and the Mac Studio route (plus its tuned-DFlash2 v0.2.13 alternate; v0.2.12 DFlash2 route rollback). No fork, no proxy, no cloud fallback.

Keep using Oh My Pi in your terminal. OMP Strata pins and verifies every download, installs Strata with its own unmodified setup.py from verified local inputs, starts it on loopback behind an API key, and launches stock OMP in an isolated profile where every model role points at that server: one Windows 11 + RTX host, one GPU you own, and real-host acceptance evidence for every claim.

Get started · How it works · Security model · Measurements · Decision · Website

CI Strata pin OMP pin License

Private by design: loopback-only server · API key required · isolated OMP profile · no cloud fallback · every byte hash-pinned

Qualified now

Qualified on 2026-11-09 (decision). Every local ledger below has 21 pass and 3 not applicable; the route ledgers passed their own G23.

Host Model and context Strata · OMP Ledger Rollback
RTX PRO 6000 Qwen3.8-Flash-Next IQ3_S, 262,144 tokens, every expert in VRAM; 4 batch slots with per-slot stock DFlash2 drafters (strata.drafting: "dflash2", one 85 MiB slot drafter per slot, --spec 8, no --batch-mtp, packed offline from pinned incoai/Qwen3.8-Flash-Next-DFlash2 files; qualified under the owner's override of its 11/90 window-8 output-burst and latency A/B misses against v0.2.12; tuned-pack alternate slots4-df2tuned at stock --draft-conf 0.55 also qualified), direct per-layer embedding (PLE) reads (--ple-io direct), KV in VRAM (--kv-streaming off, 78 GiB available-RAM start floor), stock parallel_stack + gdn_stack in strata.engine_env, conversation parking (8,192 MiB automatic cache), bare --prefill auto (32,768-token ceiling) v0.2.13 · 18.18.3 21 pass, 3 n/a v0.2.12 DFlash2
RTX 5090, default Coder IQ1_M, 262,144 tokens, KV in VRAM (--kv-streaming off), elastic KV growth (--kv-grow), stock gdn_stack in strata.engine_env; 34 GiB available-RAM start floor; engine's 4,608 MiB automatic conversation cache (requires the host's Docker Desktop and WSL, which run its production NInfer container, to be stopped; after WSL restarts, that container must be recreated by the owner's pinned relaunch script or returned through the owner's fleet controller) v0.2.13 · 18.18.3 21 pass, 3 n/a v0.2.12 262K envg
RTX 5090, qualified alternative Coder IQ1_M, 131,072 tokens; 34 GiB available-RAM start floor; engine's 4,096 MiB automatic conversation cache; starts with Docker Desktop and WSL running (production container stopped by the fleet controller; qualified on the same window's passing 262K A/B, as in v0.2.12, with every gate passing; full-model IQ3_S 131,072-token low-RAM resident alternate iq3s-131k-lowram — stock's planner pick for this host, 11,616 expert slots in VRAM at 22.04 GiB, 29.77 GiB of experts page-locked in RAM, and engine's 2,304 MiB automatic conversation cache — also qualified as a non-current alternate while the Coder stays the default) v0.2.13 · 18.18.3 21 pass, 3 n/a v0.2.12 131K
RTX 4090 + RTX 3090 IQ3_S, 262,144 tokens, stock roles (RTX 4090 as Main, staged last), manual layer split 24 (--layer-split 24), stage dense-weight trim (--trim-stage-weights, 19.40 / 17.90 GiB expert caches), stock DFlash2 drafter alone (strata.drafting: "dflash2", --draft dflash2 --draft-fallback none, packed offline from pinned incoai/Qwen3.8-Flash-Next-DFlash2 files; drops --pipeline-windows 2) with strata.drafting_args --draft-conf 0.5 (df2conf05, stock's slot-step value for the pair's per-card slots, written to stock user_args; passed all 96 plain checks against v0.2.12), no batch slots, display on the RTX 4090, direct PLE reads (--ple-io direct), engine sm_86 default table + stock gdn_stack + two-pass QSA block selection (STRATA_SELECT_TWO_PASS=1, select_two_pass) in strata.engine_env; engine's 8,192 MiB automatic conversation cache (v0.2.8 stock-residency roles draft stays measurement-only) v0.2.13 · 18.18.3 21 pass, 3 n/a v0.2.12 env86gt DFlash2
Mac Studio → RTX PRO 6000 SSH local forward to the PRO's 262K DFlash2 df2 profile (docs/REMOTE.md; tuned-pack alternate route df2tuned also qualified) v0.2.13 · 18.18.3 G23 pass, 2026-11-09 v0.2.12 DFlash2 route

Profiles and route, with the host labels the ledgers use:

What these ledgers establish, and what they do not:

  • New-release discovery is separate. The v0.2.13 campaign (opened with scripts/campaign.py open --release v0.2.13 --omp-tag v18.18.3) is complete across all five rows — the PRO slots4-df2 DFlash2 row (with a note naming the qualified slots4-df2tuned tuned-pack alternate), the RTX 5090 131K Coder alternative and 262K kvvram-kvgrow-envg default (with the full-model IQ3_S 131K low-RAM alternate also qualified), the pair df2conf05 --draft-conf 0.5 variant in place of the prepared env86gt base, and the Studio→PRO df2 route (with a note naming the qualified df2tuned alternate route). The v0.2.12 campaign (916f1d9) is complete across all five rows, the v0.2.11 campaign stays running with four of its five rows qualified (01b4d9b; its RTX 5090 131K row failed G15 twice), while earlier campaigns (v0.2.10, v0.2.9, v0.2.8, v0.2.7, v0.2.3 and v0.2.1) are complete across all five rows; the next campaign drafts each lane from the newest qualified profile across its listed variant families in scripts/campaign.py LANES (610be03, f3fed60, c7ba6c4, d10e538, 01b4d9b, 916f1d9). See the campaign lifecycle and prepared host commands.
  • Ledger counts and receipts. Each current local ledger has 21 pass and 3 not applicable; G22 (images), G23 (remote clients) and G25 (runtime comparison) are not applicable to these local, text-only profiles. Real-host v0.2.13 runs on 2026-11-09 are published at 2af52b8 on the PRO df2 DFlash2 current chain and the Studio df2 current route G23, 8ab88b0 on the PRO df2tuned tuned-pack alternate chain and its Studio df2tuned alternate route G23, d3c96fb on the RTX 5090 262K kvvram-kvgrow-envg default chain, d0c2e28 on the RTX 5090 131K Coder alternative chain, 6e2dda2 on the RTX 5090 full-model IQ3_S 131K low-RAM alternate chain, and 440fa1b (at fixed runner 7c2d25e, 18:43-19:26Z, after admitting engine 0.1.59's stage order at 6eb9f05 and fixing install verification for stock v0.2.13's carry_user_args at fb8740a) on the pair df2conf05 --draft-conf 0.5 current chain, with restored-state records across all three hosts and both Studio routes at ab96ab4 and "Not qualified" blockers on the five unqualified v0.2.13 drafts at b2e9730 (2026-11-09 windows measurements, decision; earlier v0.2.12 receipts at 243290e, 57dcd18, 2560f77 and 916f1d9 stay in their rollback ledgers). In three-pair cold-start A/B on 2026-11-09 (A1 B1 B2 A2 A3 B3, reporting B / A geometric means with A = v0.2.12 unless stated, under the owner's 5% non-regression rule), the RTX 5090 262K envg default passed all 88 checks against v0.2.12 (throughput w1 1.006 / w3 0.995, per-request 1.024 / 0.975, TTFT p50 0.989 / 1.002, TPOT p50 0.975 / 1.023, cold decode 0.986-1.018, prefill 0.998-1.075), while on the PRO stock v0.2.13 runs DFlash2 window 8 (--spec 8, engine cap 8, was 6) on one sm_120 card of 48 GB or more and missed 11 of 90 checks against v0.2.12 on output-event spacing and concurrent tail latency (inter-output-event latency p95 1.171-1.188 at every depth, w2/c1 largest gap p95 1.173, w2/c4 TTFT p95 1.118, w2/c4 TPOT max 1.067, while throughput 0.986-0.999, per-request 0.964-1.015, TTFT p50 1.009-1.036 and TPOT p50 0.981-1.033 passed and server decode reported 0.958-1.015), and the PRO tuned DFlash2 pack (slots4-df2tuned, which stock v0.2.13 runs at --draft-conf 0.55 instead of 0.35 on this card) missed 15 of 90 checks against v0.2.12 on the same burstiness (inter-output p95 1.169-1.172, w2/c1 largest gap p95 2.151, w2/c4 TTFT p50 1.129, with w2/c2 per-request 1.084 and TPOT p50 0.928); the owner overrode both on 2026-11-09 (~18:13Z) as "Override the base as current, and also qualify the tuned pack as an alternate", and the standard-pack --draft-conf 0.55 draft (df2conf055) was withdrawn before install. On the pair, the v0.2.13 env86gt base missed 6 of 96 checks against v0.2.12, all on one 32K request (w3/c1 TPOT p50 1.054, p95 1.069, max 1.070; largest gap p50 1.119, p95 1.239, max 1.252), and the STRATA_QSA_SPLIT_ATTN=1 (qsa_split_attn, ...-env86gts-...) arm missed 2 of 96 checks (w1 largest gap p95 1.075, max 1.092; targeted cold decode at 32K/100K/200K 1.003 < 1.02), whereas the --draft-conf 0.5 arm (df2conf05) passed all 96 plain checks against v0.2.12 (targeted cold decode 1.021 [1.020, 1.049, 0.993], neutral with one pair below 1.00; throughput 1.016 / 0.998, per-request 1.027 / 0.993, TTFT p50 1.005 / 1.001, TPOT p50 0.973 / 1.009; decode 100K 1.061, warm 32K 1.054, 32K 1.044) and became current under "Run the --draft-conf 0.5 main; it becomes current if every gate passes". On the RTX 5090, the STRATA_SPEC_COST=1 (spec_cost, ...-envgc-...) arm against v0.2.13 envg missed 3 of 88 checks (one-request window 1.119, largest gap p95 1.711, max 1.771) and 2 of 88 in three replacement pairs (largest gap p95 1.197, max 1.212; targeted cold decode 1.012 < 1.02, with throughput and TPOT 3-5% better in every pair and about one 4x largest-gap spike per three cold starts) and was not adopted, while the full-model IQ3_S 131K low-RAM resident profile (win11-rtx5090-iq3s-131k-lowram-strata0.2.13-omp18.18.3) beat the v0.2.13 Coder IQ1_M 131K profile on decode in its report-only screen (cold decode 1.046 at 8K, 1.093 at 32K, 1.071 at 100K and 1.094 on warm 32K vs 0.763 at the empty prompt; queue decode 1.086 / 1.086; per-request 1.134 / 1.129; TPOT p50 0.881 / 0.885; throughput w1 1.059, w3 0.917; TTFT p50 1.528 at 512 tokens and 1.122 at 32K; prefill 0.61-0.95; earlier private quality set 58/60 vs Coder 54/60) and was qualified as a non-current alternate under the owner's decision ("Qualify full-model IQ3_S 131K: prune old [rtx5090-win-a] roots for disk, then run its main gates"). Scored G24 on v0.2.13 passed 16/18 on the RTX 5090 262K envg default (tool-loop 1/3), the RTX 5090 131K Coder alternative, the RTX 5090 full-model IQ3_S 131K low-RAM alternate (tool-loop 1/3) and the pair df2conf05 --draft-conf 0.5 current profile (tool-loop 1/3), and 15/18 on the PRO df2 current profile (tool-loop 0/3, as v0.2.12) and the PRO df2tuned alternate (tool-loop 1/3), vs 15/18 across all qualified v0.2.12 server profiles (522 s on PRO df2 with pilot 6/6, 581 s on PRO df2tuned with pilot 5/6, 787 s on RTX 5090 262K envg with pilot 6/6, 726 s on RTX 5090 131K with pilot 5/6, 968 s on pair env86gt with pilot 6/6, and 906 s on pair env86g with pilot 5/6; OMP 18.18.3 changed delegation bias from eager to balanced for these models, added a system prompt instruction to read small files whole, and changed bash locale/fork/console handling, so G24 and task-set scores are not comparable with 18.18.0, just as OMP 18.17 was not comparable with 18.16.0). All protocol counters (including malformed arguments) and timeouts are zero on the qualified server profiles; all receipts and kept failures remain published.
  • The PRO pins four-slot stock DFlash2 drafting (now at stock window 8) alongside stock engine environment stacks, direct PLE reads and KV in VRAM on engine 0.1.59 (with the stock tuned pack at --draft-conf 0.55 qualified as an alternate), and the RTX 5090 pins the 262K envg Coder default, the 131K Coder alternative and the full-model IQ3_S 131K low-RAM resident alternate. On rtxpro6000-win-a, in addition to strata.ple_io: "direct" (--ple-io direct, OPERATIONS.md), setup_args.kv_streaming: "off" (--kv-streaming off, keeping KV in VRAM without --kv-resident 32768; 78 GiB start floor) and strata.engine_env with stock parallel_stack and gdn_stack (UPSTREAM.md; the launcher's parallel and GDN env stacks moved into engine 0.1.59 as Linux-only defaults, so our Windows groups stay explicit and are checked against the engine's tables), the current v0.2.13 slots4-df2-pledirect-kvvram-envpg-parking profile (2af52b8) pins strata.drafting: "dflash2" on all four batch slots (UPSTREAM.md and UPSTREAM.md: one 85 MiB slot drafter per slot under STRATA_BATCH_DF2, --spec 8 on stock v0.2.13 instead of --spec 6 on v0.2.12, no --batch-mtp, 93 GiB free-disk floor, every gate passing including G15, G17 and G18L, and G24 15/18, tool-loop 0/3) and follows stock's bare --prefill auto (32,768-token ceiling). Its qualified v0.2.13 non-current alternate win11-rtxpro6000-iq3s-262k-slots4-df2tuned-pledirect-kvvram-envpg-parking-strata0.2.13-omp18.18.3 (8ab88b0) pins stock's tuned DFlash2 pack (qwen3.8-flash-next-dflash2.strata1.q4_0-selq4.pack, 339,694,592 bytes, 94 GiB disk floor, which stock v0.2.13 now runs at --draft-conf 0.55 instead of 0.35 on this card; UPSTREAM.md; every gate passing, and G24 15/18, tool-loop 1/3), while its qualified rollback is the v0.2.12 DFlash2 profile win11-rtxpro6000-iq3s-262k-slots4-df2-pledirect-kvvram-envpg-parking-strata0.2.12-omp18.18.0 (--spec 6; G16 / G20 / G21 31.1-31.3 s, G15 kill recall 44.2-44.4 s, G10 126.94 GiB available RAM and 83.13 GiB commit headroom of 219.70 GiB with 73,526 MiB GPU memory used, the 8,192 MiB automatic conversation cache enabled with 24 parks in G19, and G24 15/18, pilot 6/6, 522 s; alongside the earlier v0.2.12 df2tuned alternate with G16 / G20 / G21 31.0-31.4 s, G15 recall 29.0 s, G10 126.99 GiB available RAM and 83.24 GiB commit headroom with 73,452 MiB GPU memory used, 36 parks in G19, and G24 15/18, pilot 5/6, 581 s, and the v0.2.11 df2 profile with G16 / G20 / G21 31.0-35.7 s, G15 rerun recall 29.2-29.3 s, G10 127.14 GiB available RAM and 83.55 GiB commit headroom with 73,496 MiB GPU memory used, 32 parks in G19, and G24 15/18, pilot 5/6, 627 s). On rtx5090-win-a, the v0.2.13 262K kvvram-kvgrow-envg default (d3c96fb, every gate passing including G15 and G18L, G24 16/18, tool-loop 1/3) pins strata.engine_env gdn_stack, setup_args.kv_streaming: "off" (34 GiB available-RAM start floor) and strata.kv_grow: true (--kv-grow), rolling back to the v0.2.12 262K envg profile (4,096 MiB automatic conversation cache, 16 parks in G19; G10 13.05 GiB available RAM and 19.53 GiB commit headroom of 85.15 GiB; G16 / G20 / G21 19.1-19.3 s; G24 15/18, pilot 6/6, 787 s), the v0.2.13 131K Coder alternative win11-rtx5090-coder-iq1m-131k-strata0.2.13-omp18.18.3 (d0c2e28, every gate passing, G24 16/18) rolls back to the v0.2.12 131K profile (3,584 MiB automatic conversation cache, 19 parks in G19; G10 11.37 GiB available RAM and 18.72 GiB commit headroom of 85.15 GiB; G15 kill recall 30.9-32.0 s; G16 / G20 / G21 16.6-18.9 s; G24 15/18, pilot 5/6, 726 s), and the v0.2.13 full-model IQ3_S 131K low-RAM resident profile win11-rtx5090-iq3s-131k-lowram-strata0.2.13-omp18.18.3 (6e2dda2, every gate passing including G15, G17, G18 and G18L, G24 16/18, tool-loop 1/3; 11,616 expert slots in VRAM at 22.04 GiB and 29.77 GiB of experts page-locked in RAM; UPSTREAM.md) is qualified as a non-current alternate while the Coder stays the default. GDN tensor-core numerics are not bit-exact upstream, and engine env pins are reviewed per engine release (ENGINE_ENV_REVIEWED_ENGINES, engines 0.1.55, 0.1.56, 0.1.57, 0.1.58 and 0.1.59, with all 22 grouped variables unchanged in 0.1.59 at 5c93890), so the next engine release needs a fresh review.
  • The pair pins strata.drafting_args (--draft-conf 0.5) and two-pass QSA block selection (STRATA_SELECT_TWO_PASS=1) alongside stock DFlash2 drafting alone, stock roles with the RTX 4090 as Main, manual split 24 with stage weight trim, direct PLE reads, and the engine's sm_86 table and GDN stack. On rtx3090-win-a, engine 0.1.59's stage order was admitted at 6eb9f05 after the 2026-11-09 pair stage-order measurement (following 0.1.58 at a5989fe): --gpu-roles 0=main,1=helper stages the RTX 3090 (GPU 1) first on layers 0-23 and the RTX 4090 (GPU 0) last on layers 24-47 at pinned manual split 24 (strata.setup_args.layer_split: "24"), with strata.trim_stage_weights: true (--trim-stage-weights, OPERATIONS.md) so each card loads only its own stage's dense weights (771 and 988 MiB), strata.ple_io: "direct" (--ple-io direct, keeping the 26.8 GiB PLE table on direct SSD reads and the 8,192 MiB automatic conversation cache enabled), strata.engine_env sm86_table + gdn_stack + select_two_pass (STRATA_SELECT_TWO_PASS=1, UPSTREAM.md and Niko1221/Strata#3313), stock DFlash2 drafting alone (strata.drafting: "dflash2", --draft dflash2 --draft-model <root>\data\dflash2\qwen3.8-flash-next-dflash2.q4_0-selq4.pack --draft-fallback none --draft-conf 0.5 --draft-vocab <root>\data\dflash2\draft_vocab.bin, packed offline during install by stock tools/dflash2_pack.py from the pinned incoai/Qwen3.8-Flash-Next-DFlash2 revision a22a1744e85c17364dcc54a6a420af7135d4c7f6 config.json and model.safetensors, with STRATA_NO_ACQUIRE=1 and no start-time fetch or silent MTP fallback; UPSTREAM.md), and strata.drafting_args pinning {"--draft-conf": "0.5"} (df2conf05, stock's slot-step value for the pair's per-card slots, which stock cannot compute on a split, written to stock user_args with stock write_config and required on the start's server line; UPSTREAM.md). The expert caches are 19.40 GiB on the RTX 3090 and 17.90 GiB on the RTX 4090 (vs 19.41 / 17.90 GiB on v0.2.11 df2, 19.41 / 18.31 GiB on v0.2.10 df2 and 19.14 / 17.75 GiB on v0.2.10 pw2 env86g MTP). After fb8740a fixed install verification for stock v0.2.13 setup re-runs over an installed root that merge user_args into args (carry_user_args), the df2conf05 profile passed all 96 plain checks against v0.2.12 (targeted cold decode 1.021 [1.020, 1.049, 0.993]) and its own requalification chain at 7c2d25e (440fa1b; every gate passing including G15, G17, G18 and G18L, G24 16/18, tool-loop 1/3) to become current, rolling back to the v0.2.12 env86gt two-pass profile (953a31f at 57dcd18, --draft-conf 0.35; G10 65.74 GiB available RAM and 28.21 GiB Windows commit headroom of 135.69 GiB, G16 / G20 / G21 55.4 / 55.3 / 55.5 s, G17 57.7 s, 22 parks in G19, G24 15/18, pilot 6/6, batch wall 968 s), while the v0.2.12 env86g-parking profile (G10 65.66 GiB available RAM and 28.18 GiB commit headroom, G16 / G20 / G21 55.2 / 55.3 / 55.6 s, G17 59.7 s, 14 parks in G19, G24 15/18, pilot 5/6, batch wall 906 s), the v0.2.11 and v0.2.10 df2 profiles and the v0.2.10 pw2 env86g MTP profile remain earlier qualified history. The v0.2.13 env86gt and env86gts (STRATA_QSA_SPLIT_ATTN=1) drafts (b2e9730), the v0.2.12 env86gw (STRATA_IQ4_W4A8=2) draft, the v0.2.8 stock-residency draft, the held v0.2.7 roles draft (PLE in RAM), the unmeasured v0.2.7 default-split draft, and the v0.2.3 default and split-32 drafts stay unqualified (decision).
  • Remote use. Through the tunnel on the current v0.2.13 DFlash2 route (G23 on 2026-11-09, published 2af52b8: missing key refused before launch, wrong key HTTP 401, health RTT 5.8-7.7 ms, stream TTFT 0.217 s, tunnel drop failing the turn in flight with the transcript preserved, and a one-turn launch-omp smoke answering ROUTE OK), on the qualified v0.2.13 tuned-pack alternate route (G23 on 2026-11-09, published 8ab88b0: health RTT 6.1-6.8 ms, stream TTFT 0.181 s, and ROUTE OK smoke), and on the qualified v0.2.12 DFlash2 rollback route (G23 on 2026-11-07 at 243290e, receipt 6b5c7e35: missing key refused before launch, wrong key 401, health RTT 5.9-7.4 ms, stream TTFT 0.247 s, tunnel drop failing the turn in flight with the prior transcript preserved, nonce recall from transcript replay after tunnel reopen 0.87 s, client restart 0.79 s and verified server restart 2.56 s, and a one-turn launch-omp smoke answering ROUTE OK with owned stop; alongside the earlier qualified v0.2.12 tuned-pack alternate route at 2560f77, receipt ca3919ce, health RTT 12.1-16.1 ms, stream TTFT 0.317 s, recalls 0.90 / 0.86 / 1.61 s, and the v0.2.11 DFlash2 route at b812df9, health RTT 6.0-10.1 ms, stream TTFT 0.270 s, recalls 1.83 / 1.11 / 2.34 s; docs/REMOTE.md).
  • Quality-first comparison drafts only. On the PRO, the one-slot IQ3_S (win11-rtxpro6000-iq3s-262k-df2-pledirect-kvvram-envg-parking-strata0.2.11-omp18.16.0), one-slot Unsloth UD-Q5_K_XL quantization (win11-rtxpro6000-ud-q5kxl-262k-df2-pledirect-kvvram-envg-parking-strata0.2.11-omp18.16.0) and six-expert DeepSeek-V4.1-Flash-1017 IQ2_XS (win11-rtxpro6000-dsv41-iq2xs-262k-experts6-strata0.2.11-omp18.16.0, separate pinned strata-dsv41 engine archive, one text conversation and no disk conversation cache) profiles stay comparison-only drafts after the 2026-11-06–07 22-task quality-first comparison kept the four-slot IQ3_S DFlash2 incumbent; no gate receipt or inherited qualification is claimed (decision, DeepSeek review, Qwen one-slot review).
  • Not supported. Images, servers on other operating systems, durable engine state, multiple tenants, fleets (no fleet route is published), other DeepSeek sets/layouts and other client routes; every profile other than the qualified ones above remains a draft or a rollback installation. Beside the three v0.2.11 quality-first PRO drafts above, the five v0.2.13 drafts carrying "Not qualified" blockers (b2e9730 — the pair env86gt draft that missed 6/96 A/B checks on one 32K request and env86gts STRATA_QSA_SPLIT_ATTN=1 draft that missed 2/96 checks with targeted cold decode at 1.003 < 1.02, the RTX 5090 envgc STRATA_SPEC_COST=1 draft that missed 3/88 checks and 2/88 in replacement pairs and the full-model IQ3_S 262K low-RAM ...-iq3s-262k-lowram-kvvram-kvgrow-envg-... draft whose first start refused when --kv-grow's elastic K/V line was absent, and the PRO df2conf055 --draft-conf 0.55 standard-pack draft withdrawn before install), the v0.2.12 PRO two-pass server and route drafts (envpgt, missed 3/96 A/B checks), the v0.2.12 RTX 5090 two-pass draft (envgt, failed G18), the v0.2.12 pair W4A8 draft (env86gw, missed cold TTFT, 32K decode and inter-token gap in its A/B), the PRO v0.2.11 UD-Q4_K_XL comparison screen (slower than the v0.2.11 IQ3_S profile in every measured workload: single-request decode 0.85-0.91x, total throughput 0.73-0.84x) and v0.2.9 UD-IQ4_XS and NVFP4-QAD model candidates (and their routes, while a 30-task Terminal-Bench comparison has not run), the v0.2.11 RTX 5090 131K profile (failed G15 twice), and the RTX 5090 v0.2.9 kvvram-kvgrow fallback (uninstalled) stay unqualified drafts.

Current profile results

Per gate, for the four current local lanes above (2026-11-09 windows measurements, 2026-11-07 windows measurements, decision).

Area RTX PRO 6000, 4-slot DFlash2 (df2, --spec 8) + PLE direct + KV VRAM + envpg (2026-11-09; v0.2.13 df2tuned alternate and v0.2.12 df2 rollback in parens) RTX 5090, 262K default, KV VRAM + KV grow + envg (2026-11-09; v0.2.12 rollback in parens) RTX 5090, 131K Coder alternative and full-model IQ3_S 131K low-RAM alternate (2026-11-09; v0.2.12 rollback in parens) RTX 4090 + RTX 3090, DFlash2 (df2conf05, --draft-conf 0.5) + split 24 + trim + PLE direct + env86gt two-pass (2026-11-09; v0.2.12 env86gt rollback in parens)
Typed tool loop (read/glob/edit/bash) through stock OMP → stock Strata pass: 3/3 tracer runs pass: 3/3 tracer runs pass: 3/3 tracer runs (Coder and IQ3_S low-RAM alternate) pass: 3/3 tracer runs
Live prefix reuse in a session (lost on engine restart) pass: 12 continuations, engine-reported (df2tuned alternate: 12; v0.2.12 df2 rollback: 12 continuations, engine-reported) pass: 15 continuations, engine-reported (v0.2.12 rollback: 15 continuations, engine-reported) pass: 13 continuations, engine-reported on Coder (IQ3_S low-RAM alternate: 12; v0.2.12 rollback: 14 continuations, engine-reported) pass: 12 continuations, engine-reported (v0.2.12 env86gt rollback: 12 continuations, engine-reported)
Interleaved conversations pass: A resumes after B; 48 uncached tokens on the controlled continuation, 34 parks (engine's 8,192 MiB cache; df2tuned alternate: 48 uncached tokens on the controlled continuation, 34 parks, 8,192 MiB cache; v0.2.12 df2 rollback: 48 uncached tokens on the controlled continuation, 24 parks), not an exact-output claim pass: A resumes after B; 46 uncached tokens on the controlled continuation, 25 parks (engine's 4,608 MiB cache with Docker Desktop/WSL stopped; v0.2.12 rollback: 48 uncached tokens, 16 parks), not an exact-output claim pass: A resumes after B; 47 uncached tokens on the controlled continuation, 33 parks on Coder (engine's 4,096 MiB cache on Coder; IQ3_S low-RAM alternate: 46 uncached tokens on the controlled continuation, 29 parks, 2,304 MiB cache; v0.2.12 rollback: 48 uncached tokens, 19 parks), not an exact-output claim pass: A resumes after B; 47 uncached tokens on the controlled continuation, 28 parks (engine's 8,192 MiB cache; v0.2.12 env86gt rollback: 47 uncached tokens, 22 parks), not an exact-output claim
Requests at once up to 4, decoded together with per-slot DFlash2 drafters (strata.drafting: "dflash2", --spec 8; tuned pack at --draft-conf 0.55 on df2tuned alternate; --spec 6 on v0.2.12 rollback) one at a time one at a time one at a time (DFlash2 verified windows up to 8 tokens, --draft-conf 0.5; --draft-conf 0.35 on v0.2.12 rollback)
Loopback, API key (including /status, /v1/telemetry and /tool-media/ on v0.2.13 / v0.2.12), fail-closed client, local-only routing, with an egress guard for OMP's startup catalog fetch pass (0 non-loopback / 8,702 ETW events, 2 compactions; df2tuned alternate: 0 non-loopback / 6,068 ETW events, 2 compactions; v0.2.12 df2 rollback: 0 non-loopback / 4,664 ETW events, 2 compactions) pass (0 non-loopback / 7,061 ETW events, 2 compactions; v0.2.12 rollback: 0 non-loopback / 9,775 ETW events, 2 compactions) pass (0 non-loopback / 7,265 ETW events, 1 compaction on Coder; IQ3_S low-RAM alternate: 0 non-loopback / 3,737 ETW events, 2 compactions; v0.2.12 rollback: 0 non-loopback / 6,461 ETW events, 2 compactions) pass (0 non-loopback / 8,390 ETW events, 2 compactions; v0.2.12 env86gt rollback: 0 non-loopback / 4,163 ETW events, 2 compactions)
Engine or client restart, then continue from OMP's transcript (an engine restart re-prefills it; no state restoration) pass, including G15 and G18L (readiness 31.1-31.4 s, G16 31.1 s, G20 31.2 s, G21 31.4 s, lowest available RAM 119.17 GiB in G19 / 119.14 GiB in G18L, G15 recall 44.8 s (idle kill), 44.8 s (mid-generation kill), G10 127.53 GiB available RAM, 83.66 GiB Windows commit headroom of a 219.70 GiB limit, PLE on direct SSD reads, KV in VRAM; df2tuned alternate: readiness 31.1-31.3 s, G16 31.1 s, G20 31.1 s, G21 31.3 s, lowest available RAM 119.26 GiB in G19 / 119.01 GiB in G18L, G15 recall 44.5 s (idle kill), 44.5 s (mid-generation kill), G10 127.83 GiB available RAM, 83.85 GiB Windows commit headroom of a 219.70 GiB limit; v0.2.12 df2 rollback: readiness 31.1-31.3 s, G16 31.1 s, G20 31.2 s, G21 31.3 s, lowest available RAM 113.51 GiB in G19 / 113.60 GiB in G18L, PLE on direct SSD reads, KV in VRAM) pass with Docker Desktop/WSL stopped, including G15 and G18L (readiness 19.1-19.3 s, G16 19.1 s, G20 19.1 s, G21 19.3 s, lowest available RAM 9.47 GiB in G19 / 8.25 GiB in G18L, G15 recall 18.7 s (idle kill), 18.1 s (mid-generation kill), G10 14.14 GiB available RAM, 16.28 GiB Windows commit headroom of a 81.15 GiB limit; v0.2.12 rollback: readiness 19.1-19.3 s, G16 19.2 s, G20 19.1 s, G21 19.3 s, lowest available RAM 7.77 GiB in G18L) pass on Coder and IQ3_S low-RAM alternate, including G15 and G18L (readiness 16.6-16.8 s, G16 16.6 s, G20 16.6 s, G21 16.8 s, lowest available RAM 8.33 GiB in G19 / 7.80 GiB in G18L, G15 recall 31.7 s (idle kill), 31.7 s (mid-generation kill), G10 12.63 GiB available RAM, 15.39 GiB Windows commit headroom of a 81.15 GiB limit on Coder; IQ3_S low-RAM alternate: readiness 25.4-27.7 s, G16 27.7 s, G20 25.4 s, G21 26.1 s, lowest available RAM 5.32 GiB in G19 / 4.96 GiB in G18L, G15 recall 22.4 s (idle kill), 20.8 s (mid-generation kill), G10 8.34 GiB available RAM, 10.41 GiB Windows commit headroom of a 81.15 GiB limit; v0.2.12 rollback: readiness 16.6-18.9 s, G16 16.6 s, G20 16.6 s, G21 18.9 s, G15 recall 30.9-32.0 s, lowest available RAM 7.08 GiB in G19) pass, including G15 and G18L (readiness 43.1-43.4 s, G16 43.2 s, G20 43.1 s, G21 43.4 s, lowest available RAM 58.22 GiB in G19 / 57.11 GiB in G18L, G15 recall 48.9 s (idle kill), 45.8 s (mid-generation kill), G10 65.95 GiB available RAM, 28.29 GiB Windows commit headroom of a 135.69 GiB limit; v0.2.12 env86gt rollback: readiness 55.3-55.5 s, G16 55.4 s, G20 55.3 s, G21 55.5 s, lowest available RAM 57.00 GiB in G18L)
Context window (262,144 tokens on the RTX PRO 6000, RTX 5090 default and pair; 131,072 on the RTX 5090 alternatives): exact limit, near-limit tool turns, explicit overflow (G17) pass (near-limit task 25.6 s; df2tuned alternate: near-limit task 24.9 s; v0.2.12 df2 rollback: near-limit task 25.6 s) pass (near-limit task 28.0 s; v0.2.12 rollback: near-limit task 30.7 s) pass (near-limit task 15.1 s on Coder; IQ3_S low-RAM alternate: near-limit task 16.7 s; v0.2.12 rollback: near-limit task 16.7 s) pass (near-limit task 55.9 s; v0.2.12 env86gt rollback: near-limit task 57.7 s)
OMP compaction (reduced threshold and production long session, G18) pass, with OMP's speculative compaction off (production 1 / reduced-threshold 6 compactions; df2tuned alternate: production 1 / reduced-threshold 6 compactions) pass, with OMP's speculative compaction off (production 1 / reduced-threshold 2 compactions) pass on Coder and IQ3_S low-RAM alternate, with OMP's speculative compaction off (production 1 / reduced-threshold 5 compactions on Coder; IQ3_S low-RAM alternate: production 1 / reduced-threshold 6 compactions) pass, with OMP's speculative compaction off (production 1 / reduced-threshold 4 compactions)
Cancelling a queued request pass, with all four slots generating pass pass pass
Tool call cut off mid-arguments (G04) pass: stock Strata and stock OMP refuse the unfinished call pass: stock Strata and stock OMP refuse the unfinished call pass: stock Strata and stock OMP refuse the unfinished call pass: stock Strata and stock OMP refuse the unfinished call
Six-task coding evaluation, 18 scored attempts (G24) 15/18 (tool-loop 0/3, pilot 5/6, batch wall 491 s; df2tuned alternate: 15/18, tool-loop 1/3, pilot 5/6, batch wall 485 s; v0.2.12 df2 rollback: 15/18, tool-loop 0/3, pilot 6/6, batch wall 522 s; not comparable across OMP 18.18.3) 16/18 (tool-loop 1/3, pilot 5/6, batch wall 798 s; v0.2.12 rollback: 15/18, tool-loop 0/3, pilot 6/6, batch wall 787 s; not comparable across OMP 18.18.3) 16/18 on Coder (pilot 5/6, batch wall 624 s) and 16/18 (tool-loop 1/3, pilot 6/6, batch wall 716 s) on IQ3_S low-RAM alternate (v0.2.12 rollback: 15/18, bugfix-a 2/3, tool-loop 1/3, pilot 5/6, batch wall 726 s; not comparable across OMP 18.18.3) 16/18 (tool-loop 1/3, pilot 5/6, batch wall 952 s; v0.2.12 env86gt rollback: 15/18, bugfix-b 2/3, tool-loop 1/3, pilot 6/6, batch wall 968 s; not comparable across OMP 18.18.3)
Guarded install and QUICKSTART example alongside earlier tuples (G26) pass on an already prepared root (install 26.2 s, start 52.2 s; df2tuned alternate: install 27.9 s, start 52.8 s; v0.2.12 df2 rollback: install 25.6 s, start 54.2 s; first install on v0.2.8 took 1,063.2 s) pass on an already prepared root (install 11.9 s, start 22.3 s; v0.2.12 rollback: install 12.7 s, start 22.4 s; first install on v0.2.8 took 365.6 s) pass on an already prepared root (install 12.4 s, start 37.9 s on Coder; IQ3_S low-RAM alternate: install 37.6 s, start 57.5 s; v0.2.12 rollback: install 12.7 s, start 36.0 s; first install on v0.2.8 took 297.8 s) pass on an already prepared root, including carry_user_args install verification (install 32.9 s, start 67.2 s; v0.2.12 env86gt rollback: install 33.4 s, start 67.0 s; first install on v0.2.8 took 820.3 s)
Earlier qualified profiles and rollbacks

Every profile and client route ledger is listed in docs/COMPATIBILITY.md; the owner's qualification decisions, retained failures, the G15 probe correction, the G16 predicate fix, the G18 fixture fix and the OMP speculative-compaction setting are in docs/DECISION.md, with full figures in docs/MEASUREMENTS.md.

  • 2026-11-07 and 2026-11-08 — Strata v0.2.12 + OMP 18.18.0 (rtxpro6000-win-a DFlash2 and tuned-DFlash2 alternate, rtx5090-win-a 262K and 131K, rtx3090-win-a pair two-pass and env86g DFlash2, Mac Studio routes): current rollbacks on the PRO (win11-rtxpro6000-iq3s-262k-slots4-df2-pledirect-kvvram-envpg-parking-strata0.2.12-omp18.18.0, ledger; G24 15/18, pilot 6/6, batch wall 522 s, G15 recall 44.2-44.4 s, restart readiness 31.1-31.3 s, 126.94 GiB available RAM and 83.13 GiB G10 commit headroom; alongside the qualified v0.2.12 tuned-pack alternate win11-rtxpro6000-iq3s-262k-slots4-df2tuned-pledirect-kvvram-envpg-parking-strata0.2.12-omp18.18.0, ledger; G24 15/18, pilot 5/6, batch wall 581 s, G15 recall 29.0 s, restart readiness 31.0-31.4 s, 126.99 GiB available RAM and 83.24 GiB G10 commit headroom), the RTX 5090 262K default (win11-rtx5090-coder-iq1m-262k-kvvram-kvgrow-envg-strata0.2.12-omp18.18.0, ledger; G24 15/18, pilot 6/6, batch wall 787 s, 34 GiB start floor, 4,096 MiB cache, restart readiness 19.1-19.3 s, 13.05 GiB available RAM and 19.53 GiB G10 commit headroom) and 131K alternative (win11-rtx5090-coder-iq1m-131k-strata0.2.12-omp18.18.0, ledger; G24 15/18, pilot 5/6, batch wall 726 s, 34 GiB start floor, 3,584 MiB cache, G15 recall 30.9-32.0 s, restart readiness 16.6-18.9 s, 11.37 GiB available RAM and 18.72 GiB G10 commit headroom), the pair (win11-rtx3090-rtx4090-iq3s-262k-df2-disp4090-roles-split24-trimweights-pledirect-env86gt-strata0.2.12-omp18.18.0, ledger; G24 15/18, pilot 6/6, batch wall 968 s, restart readiness 55.3-55.5 s, 19.40 / 17.90 GiB expert caches, 65.74 GiB available RAM and 28.21 GiB G10 commit headroom; alongside the qualified v0.2.12 env86g base profile win11-rtx3090-rtx4090-iq3s-262k-df2-disp4090-roles-split24-trimweights-pledirect-env86g-parking-strata0.2.12-omp18.18.0, ledger; G24 15/18, pilot 5/6, batch wall 906 s, restart readiness 55.2-55.6 s, 65.66 GiB available RAM and 28.18 GiB G10 commit headroom) and the Mac Studio client route (0.2.12 df2 route, ledger, stream TTFT 0.247 s; alongside the 0.2.12 df2tuned route, ledger, stream TTFT 0.317 s); measurements and decision.
  • 2026-11-06 — Strata v0.2.11 + OMP 18.16.0 (rtxpro6000-win-a DFlash2 and batch-MTP, rtx5090-win-a 262K, rtx3090-win-a pair DFlash2, Mac Studio routes): earlier qualified history on the PRO (win11-rtxpro6000-iq3s-262k-slots4-df2-pledirect-kvvram-envpg-parking-strata0.2.11-omp18.16.0, ledger; G24 15/18, pilot 5/6, batch wall 627 s, complete G15 rerun at b812df9 with 29.2-29.3 s recall, restart readiness 31.0-35.7 s, 127.14 GiB available RAM and 83.55 GiB G10 commit headroom; alongside the qualified v0.2.11 batch-MTP profile win11-rtxpro6000-iq3s-262k-slots4-batchmtp-pledirect-kvvram-envpg-parking-strata0.2.11-omp18.16.0, ledger; G24 16/18, pilot 5/6, batch wall 549 s, restart readiness 30.9-31.4 s, 127.23 GiB available RAM and 81.73 GiB G10 commit headroom), the RTX 5090 262K default (win11-rtx5090-coder-iq1m-262k-kvvram-kvgrow-envg-strata0.2.11-omp18.16.0, ledger; G24 17/18, pilot 6/6, batch wall 843 s, 34 GiB start floor, 4,352 MiB cache, restart readiness 18.5-19.3 s, 13.47 GiB available RAM and 20.30 GiB G10 commit headroom; the v0.2.11 131K draft failed G15 twice and stays unqualified), the Mac Studio client route (0.2.11 df2 route, ledger, stream TTFT 0.270 s; alongside the 0.2.11 batch-MTP route, ledger, stream TTFT 0.174 s) and the pair (win11-rtx3090-rtx4090-iq3s-262k-df2-disp4090-roles-split24-trimweights-pledirect-env86g-parking-strata0.2.11-omp18.16.0, ledger; G24 15/18, pilot 5/6, batch wall 896 s, complete G13 rerun at 611c342, restart readiness 40.5-40.8 s, 19.41 / 17.90 GiB expert caches, 66.35 GiB available RAM and 28.70 GiB G10 commit headroom); measurements and decision.
  • 2026-11-03 — Strata v0.2.10 + OMP 18.16.0 (rtxpro6000-win-a, rtx5090-win-a 262K and 131K, rtx3090-win-a pair DFlash2 and env86g, Mac Studio route): earlier qualified history on the RTX 5090 131K alternative (win11-rtx5090-coder-iq1m-131k-strata0.2.10-omp18.16.0, ledger; G24 16/18, pilot 6/6, 34 GiB start floor, 3,584 MiB cache, restart readiness 18.7-18.9 s), on the RTX 5090 262K default (win11-rtx5090-coder-iq1m-262k-kvvram-kvgrow-envg-strata0.2.10-omp18.16.0, ledger; G24 18/18, pilot 5/6, 34 GiB start floor, 4,352 MiB cache, restart readiness 19.2-19.5 s, 13.41 GiB available RAM and 20.31 GiB G10 commit headroom), on the pair (win11-rtx3090-rtx4090-iq3s-262k-df2-disp4090-roles-split24-trimweights-pledirect-env86g-parking-strata0.2.10-omp18.16.0, ledger; G24 17/18, pilot 5/6, restart readiness 43.1-43.4 s, 19.41 / 18.31 GiB expert caches, 65.65 GiB available RAM and 27.51 GiB G10 commit headroom; alongside the older qualified v0.2.10 pw2 env86g MTP alternative win11-rtx3090-rtx4090-iq3s-262k-pw2-disp4090-roles-split24-trimweights-pledirect-env86g-parking-strata0.2.10-omp18.16.0, ledger; G24 16/18, pilot 5/6, G20 rerun at 837bc89, restart readiness 55.3 / 43.2 / 55.7 s, 19.14 / 17.75 GiB expert caches, 65.37 GiB available RAM and 27.54 GiB commit headroom), on the PRO (win11-rtxpro6000-iq3s-262k-slots4-batchmtp-pledirect-kvvram-envpg-parking-strata0.2.10-omp18.16.0, ledger; G24 16/18, pilot 5/6, G14 rerun at 7971a1a, restart readiness 31.0-31.3 s, 129.12 GiB available RAM and 83.05 GiB G10 commit headroom) and on the Mac Studio client route (0.2.10 kvvram-envpg route, ledger; stream TTFT 0.219 s); measurements and decision.
  • 2026-11-02 and 2026-11-03 — Strata v0.2.9 + OMP 18.15.0 (rtxpro6000-win-a, rtx5090-win-a 262K and 131K, rtx3090-win-a pair, Mac Studio route): earlier qualified history on the RTX 5090 131K alternative (win11-rtx5090-coder-iq1m-131k-strata0.2.9-omp18.15.0, ledger; G24 14/18, pilot 5/6, 34 GiB start floor, 3,840 MiB cache, restart readiness 16.1-26.8 s), across the PRO ledger; G24 16/18, pilot 5/6, restart readiness 31.0-31.3 s, 129.57 GiB available RAM and 83.35 GiB G10 commit headroom; alongside the morning 2026-11-02 base profile win11-rtxpro6000-iq3s-262k-slots4-batchmtp-pledirect-parking-strata0.2.9-omp18.15.0, ledger; G24 16/18, pilot 5/6, restart readiness 35.1-39.7 s, 112.58 GiB available RAM and 81.28 GiB G10 commit headroom; both v0.2.9 PRO host roots were pruned on 2026-11-06 with the owner's approval), the RTX 5090 262K default (win11-rtx5090-coder-iq1m-262k-kvvram-kvgrow-envg-strata0.2.9-omp18.15.0, ledger; G24 17/18, pilot 5/6, 34 GiB start floor, 4,352 MiB cache, restart readiness 18.6-18.9 s; alongside the morning base profile win11-rtx5090-coder-iq1m-262k-strata0.2.9-omp18.15.0, ledger; G24 16/18, pilot 6/6, 36 GiB start floor, 3,072 MiB cache, restart readiness 16.6-16.8 s), the pair (win11-rtx3090-rtx4090-iq3s-262k-pw2-disp4090-roles-split24-trimweights-pledirect-env86g-parking-strata0.2.9-omp18.15.0, ledger; G24 15/18, pilot 5/6, restart readiness 43.5-43.9 s, 65.68 GiB available RAM and 27.77 GiB G10 commit headroom; and its morning base profile win11-rtx3090-rtx4090-iq3s-262k-pw2-disp4090-roles-split24-trimweights-pledirect-parking-strata0.2.9-omp18.15.0, ledger; G24 15/18, pilot 5/6, restart readiness 55.4-55.6 s, 19.15 / 17.76 GiB expert caches, 65.67 GiB available RAM and 27.74 GiB G10 commit headroom; both v0.2.9 pair host roots were pruned on 2026-11-06 with the owner's approval), and the Mac Studio client route (0.2.9 kvvram-envpg route, ledger; stream TTFT 0.172 s; alongside the morning base 0.2.9 batch-MTP route, ledger; stream TTFT 0.206 s); tuning measurements, windows measurements, tuning decision and morning decision.
  • 2026-10-30 and 2026-10-31 — Strata v0.2.8 + OMP 18.14.0 (rtxpro6000-win-a, rtx5090-win-a 262K and 131K, rtx3090-win-a pair, Mac Studio route): earlier qualified history across all lanes: on the PRO (win11-rtxpro6000-iq3s-262k-slots4-batchmtp-pledirect-parking-strata0.2.8-omp18.14.0, ledger; G24 15/18, pilot 6/6, restart readiness 35.5-51.5 s; alongside the earlier qualified PLE-direct-only alternative win11-rtxpro6000-iq3s-262k-slots4-pledirect-parking-strata0.2.8-omp18.14.0, ledger, G24 15/18, pilot 6/6, restart readiness 35.0-51.4 s, and the initial 2026-10-30 slots4-parking profile win11-rtxpro6000-iq3s-262k-slots4-parking-strata0.2.8-omp18.14.0, ledger, G24 16/18, pilot 5/6, restart readiness 73.6-88.2 s), the RTX 5090 262K default (win11-rtx5090-coder-iq1m-262k-strata0.2.8-omp18.14.0, ledger; G24 17/18, pilot 6/6, G10 rerun at f9e8c74 retained beside main-chain listener failure on Windows build 26300.9818) and 131K alternative (win11-rtx5090-coder-iq1m-131k-strata0.2.8-omp18.14.0, ledger; G24 18/18, pilot 5/6, G10 rerun at f9e8c74), the pair split-24 trim + PLE-direct profile (win11-rtx3090-rtx4090-iq3s-262k-pw2-disp4090-roles-split24-trimweights-pledirect-parking-strata0.2.8-omp18.14.0, ledger; G24 15/18, pilot 5/6, 19.15 / 17.76 GiB expert caches, 27.8 GiB G10 commit headroom; alongside the initial 2026-10-30 roles-pledirect profile win11-rtx3090-rtx4090-iq3s-262k-pw2-disp4090-roles-pledirect-parking-strata0.2.8-omp18.14.0, ledger, G24 15/18, pilot 5/6, 17.35 / 16.01 GiB expert caches, 27.9 GiB commit headroom), and the Mac Studio client route (0.2.8 batch-MTP route, ledger, alongside the initial 0.2.8 slots4-parking route, ledger); 2026-10-31 measurements, 2026-10-30 measurements and decision.
  • 2026-10-28 — Strata v0.2.7 + OMP 18.13.2 (rtxpro6000-win-a, rtx5090-win-a 262K and 131K, rtx3090-win-a pair PLE-direct, Mac Studio route): earlier qualified history across all lanes: on the PRO (ledger; G24 15/18), the RTX 5090 262K default (ledger; G24 16/18, afternoon rerun with Docker Desktop/WSL stopped retained beside morning failures), the RTX 5090 131K alternative (ledger; G24 16/18, G15 rerun retained beside main-run failure), the RTX 4090 + RTX 3090 pair in its PLE-direct stock-roles variant (--ple-io direct, 8,192 MiB automatic conversation cache, 27.3 GiB G10 commit headroom; ledger; G24 17/18; the v0.2.7 roles draft with PLE in RAM stays held on 1.03 GiB commit headroom, Strata#4518), and the Mac Studio client route (0.2.7 route, ledger); measurements and decision.
  • 2026-10-23 — Strata v0.2.3 + OMP 18.12.6 (rtxpro6000-win-a, rtx5090-win-a 131K and 262K, rtx3090-win-a pair roles, Mac Studio route): earlier qualified history on the PRO (ledger; G24 17/18, G10 rerun retained beside main-run deep-verification count), the RTX 5090 131K profile (ledger; G24 17/18, G20 rerun retained beside main-run failure) and 262K profile (ledger; G24 14/18, G19 rerun retained beside main-run failure; both v0.2.3 RTX 5090 integration roots were removed on 2026-11-02 with the owner's approval, while their ledgers and receipts stay in the repository), the RTX 4090 + RTX 3090 pair in its stock-roles variant with the engine's 8 GiB automatic conversation cache (ledger; G24 14/18, G18 rerun retained beside main-run failure), and the Mac Studio client route (0.2.3 route, ledger); measurements and decision.
  • 2026-10-21 — Strata v0.2.1 + OMP 18.10.1 (rtxpro6000-win-a, rtx5090-win-a 131K and 262K opt-in, rtx3090-win-a pair, Mac Studio route): earlier qualified history across all lanes (v0.2.1 integration roots on the hosts were removed on 2026-10-28, while their ledgers and receipts stay in the repository): on the PRO (ledger; G24 17/18), the RTX 5090 131K default (ledger; G24 17/18, G17 and G18 reruns retained beside main-run failures), the RTX 5090 262K opt-in (ledger; G24 15/18), the pair in its default automatic-split variant (ledger; G24 16/18, G17 rerun retained beside main-run failure), and the Mac Studio client route (0.2.1 route, ledger); measurements and decision.
  • 2026-10-14 (262K opt-in 2026-10-15) — Strata v0.1.44.1 + OMP 18.10.1 (rtxpro6000-win-a, rtx5090-win-a 131K and 262K opt-in, Mac Studio route): earlier qualified history on the PRO (ledger), the RTX 5090 131K default (ledger), the RTX 5090 262K opt-in (ledger; failed G26 receipt retained beside passing same-root rerun), and the Mac Studio client route (0.1.44.1 route, ledger); the pair's v0.1.44.1 release stays a draft (not qualified, not a rollback) after its K=45 split regression; 262K opt-in measurements and decision.
  • 2026-10-13 — Strata v0.1.43.1 + OMP 18.10.1 (rtx3090-win-a pair, rtxpro6000-win-a, rtx5090-win-a, Mac Studio route): earlier qualified history on the RTX 4090 + RTX 3090 pair at 262K and parking (ledger; G24 15/18, including one malformed-arguments counter on its unsuccessful third multifile-regression attempt), the PRO (ledger), the RTX 5090 (ledger), and the Mac Studio client route (0.1.43.1 route); measurements.
  • 2026-10-12 — Strata v0.1.43 + OMP 18.10.1 (rtx3090-win-a pair, rtxpro6000-win-a, rtx5090-win-a, Mac Studio route): earlier qualified history on the RTX 4090 + RTX 3090 pair at 262K and parking (ledger), the PRO (ledger), the RTX 5090 (ledger; failed G18 receipt retained beside passing full rerun), and the Mac Studio 262K route; measurements.
  • 2026-10-11 — Strata v0.1.42 + OMP 18.8.9, pair at 262K with parking (rtx3090-win-a): earlier qualified history, display on the RTX 4090 (ledger, measurements).
  • 2026-10-11 — Strata v0.1.42 + OMP 18.8.9, pair at 131K with display on RTX 4090 (rtx3090-win-a): earlier qualified history; retains its original startup refusal, pre-G16 degraded status and recovery, and failed G26 receipt beside the passing repeat (ledger, measurements).
  • 2026-10-11 — Strata v0.1.42 + OMP 18.8.9, pair at 131K with display on RTX 3090 (rtx3090-win-a): earlier qualified predecessor; requires moving the display back to the RTX 3090 (ledger, measurements).
  • 2026-10-11 — Strata v0.1.42 + OMP 18.8.9 (rtxpro6000-win-a at 262K, rtx5090-win-a at 131K, Mac Studio 262K route): earlier qualified history on the PRO (ledger), the RTX 5090 (ledger), and the Mac Studio 262K route; measurements.
  • 2026-10-10 — Strata v0.1.41 + OMP 18.8.6, PRO at 262K (rtxpro6000-win-a): earlier qualified history; retains its first G18 failure from 131K-sized fixtures beside the passing window-sized rerun (ledger, measurements).
  • 2026-10-09 — Strata v0.1.41 + OMP 18.8.6, RTX 5090 at 131K (rtx5090-win-a): earlier qualified history (ledger, measurements).
  • 2026-10-08 — Strata v0.1.41 + OMP 18.8.6, PRO at 131K (rtxpro6000-win-a, plus Mac Studio 131K route 2026-10-09): earlier qualified history (ledger, measurements, docs/REMOTE.md).
  • 2026-10-08 — Strata v0.1.40.3 + OMP 18.8.4 (rtxpro6000-win-a, rtx5090-win-a, single-card rtx3090-win-a): earlier qualified history on the PRO at 131K (G16 predicate-fix rerun; ledger) and the RTX 5090 (ledger); the single-card RTX 3090 Coder profile (ledger; 3/3 tracer, 16/16 live prefix reuse, 15/18 evaluation) stays qualified for single-card hosts, but is not a rollback while both cards are installed because single-card profiles name GPU 0, now the RTX 4090; measurements.
  • 2026-10-07 — Strata v0.1.40.3 + OMP 18.8.3, PRO at 131K with slots and parking (rtxpro6000-win-a): earlier qualified history (ledger).
  • 2026-10-07 — Strata v0.1.40.2 + OMP 18.8.0, PRO at 131K without slots and parking (rtxpro6000-win-a): earlier qualified history; first profile to turn OMP's speculative compaction off (ledger, docs/UPSTREAM.md OMP item 12).
  • 2026-10-02 (G15 repeats 2026-10-03) — Fourth tuple, Strata v0.1.34 + OMP 18.4.10, Coder IQ1_M at 131K: first tuple to pass G04 on stock releases; qualified on the RTX 5090 (ledger), the single-card RTX 3090 (ledger; not a rollback while both cards are installed), and the RTX 4090 in 32 GiB stock low-RAM mode (ledger; failed reduced-threshold G18 probe retained beside passing rerun).
  • 2026-09-30 to 2026-10-01 — Earlier unqualified rollback installations (G04 failed before both upstream fixes shipped): first RTX 5090 candidate on Strata v0.1.27 + OMP 18.4.0 (ledger), second RTX 5090 candidate on Strata v0.1.30 + OMP 18.4.6 (ledger), and third-tuple Strata v0.1.31 + OMP 18.4.8 on the 64 GiB RTX 3090 (ledger) and 32 GiB low-RAM RTX 4090 (ledger).
  • 2026-10-02 to 2026-10-06 — Unqualified comparison and variant-selection drafts (gates not run, except the G25 comparisons): Strata v0.1.36 + OMP 18.4.12 (128 GB RTX 3090 model, calibration and 262K/524K context probes; client and fleet route drafts), Strata v0.1.38 + OMP 18.5.0 (G25 same-host NInfer comparisons on the RTX 5090, 192 GB RTX 4090 and RTX 3090; docs/G25.md), and Strata v0.1.39 / v0.1.40.1 + OMP 18.5.0 (RTX PRO 6000 model and Terminal-Bench screens); see docs/DECISION.md and docs/MEASUREMENTS.md.

Bounded experiments (comparison-only)

scripts/experiment_loop.py freezes two checked-in RTX 5090 or RTX PRO 6000 profiles, their artifact/model/template/configuration identities, harness file digests, workload, repetitions, stop limits and the predecessor restoration target. It does not qualify a release, change defaults or route development work to a local model.

python3 scripts/experiment_loop.py plan \
  --baseline win11-rtx5090-coder-iq1m-131k-strata0.2.3-omp18.12.6 \
  --candidate win11-rtx5090-coder-iq1m-262k-strata0.2.3-omp18.12.6 \
  --rationale "Compare the separately qualified opt-in context without changing model or quantization" \
  --repetitions 3 --out <private-manifest.json>
python3 scripts/experiment_loop.py report --directory releases/experiments/<key>
python3 scripts/experiment_loop.py next --directory releases/experiments/<key> \
  --candidate <another-checked-in-5090-profile-id>

plan refuses identities already recorded under releases/experiments/, including blocked or failed trials. Changing repetitions, workload or prose does not evade the source/model/quant/configuration dedupe key. next writes exactly one validated proposal to queue.json; it never executes the queue. An authorized retry of a blocked trial uses plan --baseline ... --candidate ... --retry-of <blocked-key>. The referenced receipt must be either an unstarted blocked trial or a failed trial with a successfully restored predecessor; it remains untouched, and the retry has a distinct append-only key. Completed trials cannot retry. --resume-candidate can reuse a failed trial's complete baseline only when no candidate repetition started and the workload/repetition count still match. Every reused measurement names its source receipt; it is not a new run. --host-profile rtxpro6000-win-a selects the PRO host contract, requiring both profiles to name that GPU; it never rewrites a 5090 profile's identity. The comparison-only PRO Coder profiles keep the same model and 131K/262K contexts, with their own host floors and stock hybrid-CPU worker setting; neither is qualified.

For an explicitly authorized exclusive window, sync the frozen tooling to the GPU host and launch python scripts/experiment_loop.py run --manifest <private-manifest.json> --bindings <private-bindings.json> through scripts/windows/launch-detached.ps1, not a foreground SSH process. Bindings contain roots.baseline, roots.candidate (distinct integration roots) and host_lock_id, the same private host identifier used by eval_taskset's HOST_LOCKS on that machine. Keep a corresponding client-side host claim for the whole detached window if other client schedulers share this host: PID-file locks are machine-local, not a distributed lease. Poll the four JSON records with bounded waits; a disconnect is not permission to start another worker or delete a claim. An existing claim, including an ambiguous stale one, is refused. python scripts/experiment_loop.py preflight --manifest <file> --bindings <file> prints the loop's admission JSON without installing or starting either arm. Display-attached profiles tolerate graphics clients under the existing idle-VRAM ceiling, but never tolerate foreign compute processes.

Admission requires the predecessor's installed fingerprint, an idle GPU, no competing tenant or occupied server port, and RAM/disk above both profiles' floors. A currently serving predecessor must first be released by its owner; the experiment will not evict it to satisfy admission. The candidate installs in isolation using existing pinned lifecycle tooling. Both arms run perf_probe.py for every repetition; disk/owned-GPU checks continue during probes. The wall/failure limits stop new work and terminate the owned probe tree. Restoration runs in finally with a separate bounded safety budget: stop owned arms, start the predecessor, then require G10 and authenticated healthy status. Raw child logs stay private. Each public write goes through the existing publication scrubber; missing measurements remain missing. After releasing the baseline, a read-only RAM-settle check polls at five-second intervals for at most 90 seconds inside the trial's existing wall budget. It never lowers either profile's floor or changes host services.

The probe records TTFT, server-reported prefill/decode rates and two-request FIFO bounds where available. It does not score task correctness (eval_taskset.py remains the separate macOS sandboxed task evaluator) or expose true inter-token latency. Reports therefore hold adoption pending paired task-quality evidence, never select on throughput alone, and refuse --engine-only when model, quantization, template, protocol, harness or configuration changed. record-blocked --manifest <file> --observation <file> can preserve a read-only failed preflight without starting any runtime; it cannot authorize a run or manufacture a restoration pass. The initial 262K trial was blocked by occupancy and available RAM. The stale queue's predecessor-engine proposal was not executed: both arms retain Strata 0.1.44.1 and OMP 18.10.1.

Authorized context comparison, 2026-10-16 (three repetitions per arm). Both hosts completed the Coder IQ1_M 131K-versus-262K workload, passed predecessor restoration G10, then stopped the experiment roots and restored the same original serving containers with healthy endpoints. The RTX 5090 baseline comes from the preserved failed attempt: an externally reactivated owned CPU encoder exhausted the 36 GiB admission floor before the candidate started. Its candidate-only retry explicitly references the original baseline measurements; the temporary encoder pause was removed and its health restored. That host-state difference makes small deltas descriptive, not a causal performance claim.

Values below are median / p95, with p95 linearly interpolated across the three repetitions. TTFT/prefill/decode use the 102,400-token unique-prefix samples; FIFO bounds use the separate two-request queue probe. Full depth and warm-repeat results, preflight, failures, exact scrubbed execution references and restoration identities are in the RTX 5090 receipt and RTX PRO 6000 receipt.

Host Metric 131K baseline 262K candidate
RTX 5090 TTFT, ms 13,718.77 / 13,770.88 13,696.39 / 13,697.49
RTX 5090 Prefill, tokens/s 7,503.26 / 7,516.30 7,521.34 / 7,521.78
RTX 5090 Decode, tokens/s 229.02 / 238.38 225.12 / 238.33
RTX 5090 FIFO lower bound, ms 2,261.47 / 2,262.14 2,233.47 / 2,277.33
RTX 5090 FIFO upper bound, ms 2,331.94 / 2,359.19 2,302.83 / 2,360.18
RTX PRO 6000 TTFT, ms 12,522.93 / 12,543.47 12,558.12 / 12,568.19
RTX PRO 6000 Prefill, tokens/s 8,221.66 / 8,256.22 8,193.31 / 8,221.03
RTX PRO 6000 Decode, tokens/s 215.20 / 228.64 216.39 / 234.65
RTX PRO 6000 FIFO lower bound, ms 2,371.90 / 2,375.97 2,347.23 / 2,358.74
RTX PRO 6000 FIFO upper bound, ms 2,437.83 / 2,467.51 2,374.17 / 2,418.26

These probes stop below the baseline's context limit and do not measure task quality. No existing profile, qualification or default selection changes; the queue contains no automatic follow-up.

Host-free regression command (standard library only): python3 -m unittest discover -s tests -t . -v. Real-client/frontend tests require the documented dev-env prerequisites below; skips are not live gates.

How it works

Oh My Pi (coding-agent client) → OMP Strata (profile, lifecycle tool, evidence) → Strata (inference engine) → Qwen3.8-Flash-Next (served model: IQ3_S on the RTX PRO 6000 and the RTX 4090 + RTX 3090 pair, Coder IQ1_M on the RTX 5090 and the fourth tuple's GPUs). Stock OMP talks directly to stock Strata over OpenAI Chat Completions on 127.0.0.1:18090. OMP owns transcripts, tools, compaction and resume; Strata owns inference, templating and its live cache. After any restart OMP's transcript is authoritative and the engine re-prefills it; nothing claims restored GPU state.

docs/QUICKSTART.md has the exact tested commands: fetch → install → keygen → start → launch-omp → stop.

  • profiles/<id>.json is the single nonsecret source of truth: host expectations, stock setup choices, every artifact pinned by URL, size and SHA-256, and the OMP route. A changed component or setting is a new profile with its own integration root and release ledger; the predecessor stays as the rollback installation.
  • fetch downloads the pinned OMP binary, Strata engine and llama.cpp archives and model shards, resumable, and verifies every byte before anything is used.
  • install runs stock Strata's own setup.py from those local inputs only, with its data folder redirected into the integration root, and checks the generated engine flags against the profile.
  • keygen writes a random API key to a user-only file under the root; start binds Strata to loopback with that key, refuses to start when the port, GPU or RAM budget is not free, and returns once authenticated identity checks pass.
  • launch-omp runs the pinned stock OMP binary with an isolated home and the omp-strata profile: every model role on the local server, as many requests in flight as the server has batch slots (one without slots), discovery off, retries, model fallback and speculative compaction off, ambient provider credentials scrubbed, and OMP's startup catalog fetch held to loopback by the launcher's proxy settings.
  • Gate probes, a tracer and a frozen six-task evaluation run on the real host; scrubbed receipts under releases/<profile>/ bind each result to the exact profile fingerprint.

Documents

Operate

Verify

Layout

Path Purpose
profiles/ The server profiles (current, first, and every later tuple's candidates and drafts): every artifact pinned by URL, size and SHA-256
routes/, examples/ Client routes (server profile ids and fingerprints, ports, roles) and a neutral private-binding example
locks/ Hash-locked Python wheels for stock setup.py
omp_strata/ Stdlib-only tooling: fetch/verify, install, lifecycle, OMP configuration, SSH client routes, transcripts
scripts/omp_strata.py The operator CLI
scripts/tracer.py, scripts/realhost_gates.py, scripts/requalify.py Real-host qualification probes and the unattended sequence that runs them (on the GPU host)
scripts/evaluate.py, eval/ The frozen six-task evaluation
scripts/pull_run.py, scripts/publish_run.py Copy a finished run's results from the GPU host, then scrub them into receipts, ledger and manifest
scripts/verify_release.py Checks that the profile or route, ledger and receipts bind together
scripts/upstream_watch.py, upstream-watch.json New Strata/OMP releases, tracked upstream issues, and drafting the next tuple's profiles
scripts/remote_gates.py, scripts/fanout_proof.py G23 probe for a client route, and the subagent fan-out proof for a fleet (from the client)
scripts/perf_probe.py Prefill, decode, TTFT and draft acceptance by context depth, to compare variants on one host
scripts/compare_g25.py, omp_strata/comparison*.py G25: the frozen evaluation on Strata and on NInfer in alternating exclusive windows, paired and scored
scripts/g25_host.py, scripts/g25_host_probe.py, scripts/probe_g25_speed.py G25 on a Windows host: prepare, engine switches and the read-only host probe; plus a client-side speed probe for both engines
scripts/render_compatibility.py, scripts/documented_route.py Generate docs/COMPATIBILITY.md; check QUICKSTART's commands against the CLI
releases/<profile or route>/ Manifest, gate ledger, receipts and scrubbed evidence
tests/ Host-free unit tests and mock-tier tests with the real stock OMP client
assets/, scripts/render_assets.py The mark, artwork sources and their rendered PNGs (docs/BRAND.md)
site/ The one-page public site, deployed by .github/workflows/pages.yml

Host-free checks (no GPU). dev-env provides what the client and composed tests need: this platform's pinned OMP binary, the pinned Strata source, and a Python 3.13 env (the lock's interpreter, which must be on PATH) with stock setup.py's packages at the lock's versions. The CUDA wheels are left out, and versions are pinned but not hash-checked because the lock hashes Windows wheels. It prints the OMP_STRATA_* variables, or runs a command with them set; CI runs the same command. Without those variables the client and composed tests skip explicitly.

python3 scripts/omp_strata.py dev-env --profile profiles/win11-rtxpro6000-iq3s-262k-slots4-parking-strata0.1.43-omp18.10.1.json -- \
  python3 -m unittest discover -s tests -t .
python3 scripts/verify_release.py --manifest releases/win11-rtxpro6000-iq3s-262k-slots4-parking-strata0.1.43-omp18.10.1/manifest.json

The OMP family

  • OMP NInfer: durable local inference, session state that survives an engine restart
  • OMP Session Gateway: every live session, one private mobile page
  • OMP Oracle: asynchronous web oracles for hard questions

License

MIT. See LICENSE. Oh My Pi, Strata, llama.cpp and the model retain their own licenses and notices.

Community project; not affiliated with or endorsed by the Oh My Pi maintainers, the Strata maintainer, Qwen or NVIDIA.

About

Stock Oh My Pi on a stock, local Strata inference server: one Windows 11 + RTX host, every component hash-pinned, loopback behind an API key, and real-host acceptance evidence for every claim. Nothing forked, no proxy.

alphastorm.github.io/omp-strata/

Topics

ai-agentcoding-agentcoding-assistantcudadeveloper-toolsgguflocal-ailocal-llmnvidiaoh-my-piopenai-compatible-apipythonqwenqwen3rtx-3090rtx-4090rtx-5090self-hostedstratawindows

Resources

Readme
MIT license

Code of conduct

Code of conduct

Contributing

Contributing

Security policy

Security policy
Activity

Stars

2 stars

Watchers

0 watching

Forks

0 forks
Report repository

Releases

Packages

Used by

Contributors

Languages