feat: Jetson AGX Orin becomes a first-class card — senses on unsloth/gemma-4-12B-it-qat-w4a16, VALIDATED live (0.55.0) - #176
Merged
Conversation
Scope-explored (13 surfaces), converged, and exported. User decisions: orin-only pin, new orin-lobe shape, drop MTP if unsupported, model card guides context (re-measured live). Capability goals image/video/audio/ reasoning are probe-gated; audio rides on vLLM #101, video unknown. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HzuzZKukvUXJooAZSBn2XD
Rigorous pass (hardware + hard-to-reverse + distributed-state signals). Findings: config-verified 262144 ctx + declared video/audio + int4 quant (c3 precedent narrowed); incumbent-first benchmark ordering; Tegra iowait threshold persistence; snapshot rollback; evidence-key redaction; the reasoning-field probe trap (h11 rejected, h20 replaces). q3 resolved: accept the served-id break + role-name migration follow-up. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HzuzZKukvUXJooAZSBn2XD
…e /spec-to-plan) Frame extended mid-plan with c29/h26 (worker->Thor proxy new, cortex->Spark preserved) and re-exported. Plan covers all 48 targets; risks: KV-at-256K, mesh-visible downtime, video/audio verdicts, operator-provided Thor key. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HzuzZKukvUXJooAZSBn2XD
t4: ~/.lobes snapshotted and verified byte-identical (rollback point). t3: incumbent coolthor baseline on the current engine — decode 9.99 tok/s (medium) / 6.39 (long) from usage.completion_tokens, MTP acceptance 69-83%, KV 11.29 GiB / 480,431 tok / 3.67x. d1 (approved): docs/orin-profiles.md overstates that KV pool 1.67x — boot ORDER, not knobs, sets it on unified memory. Filed as #171; frame c20 amended to measured reality per h16. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HzuzZKukvUXJooAZSBn2XD
…a16 (t1) Adds a candidate SupportedModel entry for the Gemma 4 12B QAT int4 W4A16 checkpoint (unsloth-qat-senses-first-class-orin-variation plan, t1), mirroring the coolthor gemma-4-12B-it-NVFP4A16 gear in architecture (Gemma4UnifiedForConditionalGeneration, gemma4 tool parser, compressed-tensors) but citing its own config.json for the knobs that differ: 262144 native context (double the incumbent 131072), INT4 pack-quantized weight-only quantization (not FP4), and a natively declared video_token_id the repo's prior "text+image+audio" Gemma capability line predates. role_hint is deliberately "candidate", not the literal "multimodal" the covering plan names, to preserve test_exactly_one_gemma_multimodal_gear's singular-tier-owner invariant -- the same reason the sakamakismile coder entry is "candidate" rather than "multimodal". This keeps the fleet-wide multimodal/senses/normal tier default, and every thor/spark profile that pins the raw coolthor id, untouched. docs/gemma-4-12b-qat-w4a16.md documents the checkpoint with a capability table whose rows are all marked pending-live-probe -- nothing is claimed validated, and issue #101's audio-drop gap is flagged as an expected risk carried over from the vLLM path, not resolved by this checkpoint's own audio_config declaration. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Add lobes/machines/orin.py — the Jetson AGX Orin 64GB Developer Kit (Ampere sm_87), following the thor.py CardStrategy convention exactly. One new module + one register() import line in machines/__init__.py (slotted after thor; its sole "orin" marker never overlaps spark/thor/blackwell's markers, verified by test). Ground truth (docs/orin-profiles.md, live 2026-07-16/17): compute capability 8.7 (sm_87), hostname "orin", device-tree model "NVIDIA Jetson AGX Orin Developer Kit". Legacy single-model defaults use the measured 0.45 gpu_mem_util and TRITON_ATTN. The embedder/reranker pooling divergences (TRITON_ATTN, reranker enforce_eager) are declared as orin's OWN role_overrides rather than composing the shared SM_110 trait — the doc is explicit these are a conservative carry-over from Thor's sm_110 finding, not independently proven on sm_87, so reusing a trait named "sm_110" would misattribute the cause. Detection-only scope: no lobes/profiles/builtin/orin.toml (that's a later task) and no shapes/goldens/catalog/compose changes. TDD: failing tests added first to test_detect.py (fact-set resolution, device-tree fallback, hostname-alone resolution, still-UNKNOWN for unrelated hardware) and test_machines.py (signature, non-shadowing precedence, role_overrides provenance, no SM_110 composition), then the implementation made them pass. Verified on the physical Orin dev box: `pytest tests/ -k "detect or machine or profile" -n auto` is green (268 passed). The unfiltered full suite has 20 pre-existing failures, all in test_init.py/ test_cli_logs.py's `main(["init", ...])` helpers that don't mock hardware detection (unlike the neighboring _inject_spark_detection fixture) — now that this literal box resolves to the real "orin" card instead of UNKNOWN, `lobes init`'s fleet path 404s on the not-yet-created orin.toml profile. Confirmed via git-stash that these same tests pass pre-orin on this exact box; they exercise the `init`/ `logs` verbs, not detection, and creating orin.toml is out of this task's scope (it would break the builtin_names() goldens this task is committed to leaving alone). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
gpu_mem_util is a fraction of the whole device, so co-resident engines are deducted one-for-one from KV — the only elastic term. Arithmetic reconciles to within ~0.5 GiB. Also records the 31% CUDA-graph under-estimate (~0.37 GiB of unused headroom) vLLM's own log flags. Detail added to #171. Records d2 (t1 role_hint precedent) and d3 (host-dependent tests, #172). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HzuzZKukvUXJooAZSBn2XD
…tion The 20 failures the coordinator reproduced were a pre-existing test-isolation gap: main(["init", ...]) helpers in both files called real lobes.runtime._detect.detect_card() with no injection, unlike the neighboring _fake_card/_patch_detect pattern already established in test_init_profile.py, test_init_shape.py, test_cli_fleet.py and test_upgrade_compat.py. That worked by luck as long as the box running pytest resolved UNKNOWN -> the conservative 'base' profile; landing the new orin CardStrategy exposed it, because this dev box IS a real Jetson AGX Orin (hostname "orin") that now correctly resolves an "orin" card with no built-in profile yet (still t6's job). Fix: a file-local autouse pytest fixture (_pin_spark_detection) in each file, monkeypatching _detect.detect_card to a fixed "spark" card for every test. Chose autouse over the repo's more common per-test _patch_detect() call because every test in these two files wants the exact same neutral, resolvable card — the card's identity is incidental to what's being asserted (scaffold mechanics / lobes logs plumbing), unlike test_init_profile.py/test_init_shape.py where the injected card IS the thing under test. Mirrors tests/conftest.py's existing offline_runtime autouse fixture, which neutralises docker/ health/capabilities probes for the identical reason. No assertions weakened or tests skipped — same expectations, now driven by an injected fact instead of whatever hardware happens to run pytest. Removed the now-fully-redundant explicit _inject_spark_detection(monkeypatch) call sites in test_init.py (5) now that the autouse fixture covers them. Host-independence is proven, not just asserted: a new parametrized test_detection_pin_drives_the_resolved_profile_not_the_host overrides the fixture per-test with two different injected cards (spark, thor) and checks the dry-run output's "Profile: <name> (...)" line tracks the injected card — direct evidence the outcome is driven by the fixture, never by socket.gethostname()/nvidia-smi/device-tree on the box actually running pytest. Does not create lobes/profiles/builtin/orin.toml (still out of scope for t2/this fix) — the fix works by never letting detection resolve a real card at all in these two files. Verified on the physical Orin dev box: `uv run pytest -n auto` is now fully green — 2730 passed, 15 skipped, 0 failed (up from 20 failed pre-fix; +22 = the 20 previously-failing tests plus 2 new parametrize cases). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Adds lobes/machines/orin.py (Ampere sm_87, 64GB) with orin's own role_overrides (TRITON_ATTN pooling, reranker enforce_eager) carrying provenance that flags them as unproven-on-sm_87 carry-overs rather than composing the sm_110 trait. Also makes test_init.py / test_cli_logs.py hermetic (autouse detection pin) — they previously inherited the runner's real GPU card and passed only because CI is not an Orin. See #172.
…off-switch) The senses swap targets unsloth/gemma-4-12B-it-qat-w4a16, which ships no MTP draft head. The recorded operator decision is to attempt the incumbent google/gemma-4-12B-it-assistant draft at boot and, if vLLM refuses it, drop speculative decoding rather than block the swap. That was impossible: the flag was a hardcoded literal in the vllm-multimodal command. MECHANISM — a shell-lexed STRING command:, scoped to this lane only. `docker compose` cannot conditionally omit a `command:` LIST item. An item that substitutes to "" renders as an empty argv element, and `vllm serve` reads that as a second positional and exits 2 with `error: unrecognized arguments:` — VERIFIED by parsing that exact argv through the served build's own serve parser (vLLM 0.23.1rc1.dev672+g93d8f834d, image lobes/vllm-gemma4:local). So folding the flag into one list item and blanking it is dead, as suspected. A string `command:` is shell-lexed AFTER substitution, so a variable expanding to nothing leaves no token at all — the flag is omitted ENTIRELY, the only "off" rendering vLLM accepts. Verified end to end with real `docker compose config` (Compose v5.3.1): unset renders argv byte-identical to before; MULTIMODAL_ SPECULATIVE_CONFIG= renders 13 tokens with no --speculative-config and no empty element. The dash-only default operator is deliberate — the colon-dash form every other knob uses treats an empty value as unset and would put the draft straight back. Every other lane keeps its list; the two other Gemma lanes are byte-identical. - ZERO golden changes (template-defaults.env and friends untouched): the knob uses the dash-only operator, which regen's extractor documents as outside the tracked defaults surface. - test_catalog.py's byte-guard is STRENGTHENED, not deleted: it now pins the catalog item as the knob's full substitution rather than as a bare substring — the item string also appears in the lane's new prose comments, which a substring check would match after someone deleted the actual flag. The same hole is closed in test_cli_fleet.py. - test_tool_parser_plugin.py's vllm-multimodal service hash recomputed (the rendered argv is unchanged; only the template's representation of it moved). - env.example documents unset / empty / retarget, including the two-layer quoting a custom value needs (bare single quotes silently mangle the JSON). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HzuzZKukvUXJooAZSBn2XD
…-switch)
String command: form so an empty MULTIMODAL_SPECULATIVE_CONFIG omits the
flag entirely; dash-only default operator (${VAR-default}) is load-bearing
so the off-switch stays reachable. Zero golden changes. Byte-guards
strengthened rather than deleted.
Checkpoint on disk (9.56 GiB, +1.91 vs incumbent), config.json read locally: int4 W4A16, 262144 native, video_token_id declared. Projection from measured incumbent facts: 262144 reachable at util 0.45 either way, but senses-first gives 2.67x vs gears-first 1.52x. Also records Thor peer state (keyless, worker live) and this box's 0.45.0 gateway gap (no worker role) — risk r5 / deviation d5. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HzuzZKukvUXJooAZSBn2XD
… lobes vetoed Adds `lobes/profiles/builtin/orin.toml`, the fleet-shaped half of the Jetson AGX Orin 64GB card (Ampere sm_87) whose detection landed in lobes/machines/orin.py (t2). It supersedes the hand-authored operator profile docs/orin-profiles.md records. - `senses` = unsloth/gemma-4-12B-it-qat-w4a16 (compressed-tensors, TRITON_ATTN). Every checkpoint fact in the comment block is read off the checkpoint's own config.json, not the model card: Gemma4UnifiedForConditionalGeneration, 262144 native, num_bits=4/type=int/strategy=group (INT4 weight-only, NOT FP4), vision_config + audio_config both present. - `gpu_mem_util = 0.45` / `max_model_len = 262144` are marked MEASURED-PENDING in a prominent block, with the reason they cannot simply be inherited: the 2026-08-04 live re-measurement on this box (approved deviation d1, issue #171) contradicts docs/orin-profiles.md — the incumbent at 0.45/131072 holds 11.29 GiB KV = 480,431 tokens = 3.67x, not 18.86 GiB / 802,644 / 6.12x, because gpu_memory_utilization is a fraction of the WHOLE device and the doc's figure came from a senses-FIRST boot into an empty one. At 480,431 tokens the 256K window is ~1.83x and may be refused outright. The live boot (t9) backfills both; a refusal is the more useful datum (thor-muse's 0.40 was refused, thor-worker's 0.45 booted first try). - `cortex` / `muse` / `worker` feasible=false: all three are NVFP4 exports that quantize ACTIVATIONS to FP4, which needs Blackwell tensor cores. A hard architecture line, not a memory tradeoff — contrast senses's W4A16, whose weights dequantize through Marlin (cc >= 8.0) with 16-bit activations. - The pooling gears' carried-over Jetson divergences (embedder/reranker TRITON_ATTN, reranker enforce_eager) are deliberately ABSENT from the TOML: `orin` joins `thor` in `_MACHINE_DERIVED_BUILTINS`, so the loader overlays them from lobes/machines/orin.py's role_overrides and the literals are never re-typed. A test pins both halves (values match the registry; the knob names do not appear in the file). Goldens regenerated with tests/goldens/regen.py: six NEW files only (tests/goldens/orin.env + five shapes/*__orin.env). Zero byte changes to the existing base/spark/thor goldens or to spark.toml/thor.toml, and the spark/thor asserts in tests/test_profile_schema.py are untouched. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
senses = unsloth/gemma-4-12B-it-qat-w4a16 at util 0.45 / 262144 marked MEASURED-PENDING (d1/#171 cited). cortex/muse/worker vetoed on the FP4-activations-need-Blackwell line. loader now derives orin's role overrides from lobes/machines/orin.py rather than re-typing them, exactly as thor does.
image PASS (control held), video INGESTED but no directional discrimination, audio encoded yet unusable, reasoning opt-in via enable_thinking, tools PASS. Refines #101: audio tokens DO scale with duration (~25 tok/s), so the documented '~19 placeholder tokens' signature is wrong on this build — the honest test is discrimination, which the model fails (it reports hearing nothing while carrying 112 audio tokens). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HzuzZKukvUXJooAZSBn2XD
All four stream through the gateway with usage in the final chunk. Records the consumer trap: with thinking on, the trace streams on delta.reasoning while delta.content stays empty for the whole trace. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HzuzZKukvUXJooAZSBn2XD
…n card Two live-found regressions on the Jetson AGX Orin 64GB, both previously survivable only by hand-patching the deployed box. 1. The budget clobber. The box ran `--profile orin --shape thor-lobe`, and a shape's [overrides.senses] WINS over the card profile at render time, so Thor's measured sm_110 values (0.30 / 131072) replaced the Orin card's own (0.45 / 262144) — visible in the wave-1 goldens: thor-lobe__orin.env vs orin.env. The new builtin `orin-lobe` shape hosts senses + embedder + reranker, drops cortex, and carries the Orin budget as its own declaration: orin-lobe__orin.env renders 0.45 / 262144. Unlike every other builtin shape it hosts NO stt/tts — the Parakeet base image (scitrera/dgx-spark-vllm) ships a torch with no sm_87 kernels (measured live: 8 restarts, "no kernel image is available"), so audio forwards to a peer via AUDIO_URL. DECLARED/UNVALIDATED per #108: no box has booted it, and its senses values are the card profile's MEASURED-PENDING hypothesis, pending the t9 boot. A lockstep test fails CI if the card is backfilled and the shape is not. 2. The Tegra iowait quirk, now persistent. /proc/stat here reports ~59% iowait with zero disk I/O (the sugov:* cpufreq-governor kthreads flicker in D state and inflate nr_iowait), so at the shipped LOBES_IOWAIT_DEGRADED_THRESHOLD=50 the gateway 429-sheds every full-tier request indefinitely — on this card, all of senses. The live box survived on an ephemeral shell-env override any `docker compose up` reverts. Mechanism: neither render.py nor shape_render.py could emit a non-role env key (only shape_render's hardcoded opt-in activation tables), so Profile grows one optional top-level `[host_env]` table — string-valued, validated, rendered BEFORE the role keys so a role knob always wins a collision — and builtin/orin.toml declares LOBES_IOWAIT_DEGRADED_THRESHOLD = "100" with the accounting quirk written out so nobody "fixes" it back to 50. It sits on the CARD, not the shape: it is a fact about the board, so every shape over `orin` inherits it, machine-as-brain (what a bare `lobes init` renders) included — a shape-scoped fix would leave the default path shedding. No other card declares a host_env, so base/spark/thor renderings are byte-unchanged; the six orin goldens each gain exactly one line. `lobes doctor` now names the key on a deployment still carrying 50. Also: --shape metavar derived from builtin_shape_names() instead of a hand-listed string that had already gone stale (thor-worker was missing), and orin-lobe added to scripts/accept-shape.sh so t9 can produce its transcript. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…t_env orin-lobe hosts senses+embedder+reranker (no local stt/tts on sm_87) and carries the orin senses budget so thor-lobe's Thor-measured 0.30/131072 no longer clobbers it — pinned by a lockstep test against builtin/orin.toml. New optional [host_env] table on Profile renders the Tegra iowait threshold (100) from the CARD, so a bare machine-as-brain init is fixed too, not just this shape. DECLARED/UNVALIDATED per #108.
…pu_access` (0.55.0)
This Jetson AGX Orin's NVIDIA container toolkit (1.19.1, `mode = "auto"`)
resolves to legacy **csv** mode, where the compose templates'
`deploy.resources.reservations.devices` GPU request fails at container CREATE:
"invoking the NVIDIA Container Runtime Hook directly … is not supported."
The live fix was a HAND EDIT of every GPU service's `deploy:` stanza into
`runtime: nvidia` — and a re-init reverts it (docs/orin-profiles.md
divergence 1), which is precisely what the next live boot is about to do.
Mechanism: a card profile declares `gpu_access = "devices"` (the default,
today's behaviour) or `"runtime"`, and `lobes init --apply` GENERATES two
compose overrides that `!reset` each GPU service's `deploy:` stanza and set
`runtime: nvidia` — on EVERY render, and removed when the resolved card no
longer declares it.
Why a generated override rather than a template `${VAR}`: docker-compose has no
conditional-block syntax, so a service needs one form or the other and no
substitution can pick between them — only a second compose file can. That is
the mechanism `docker-compose.shape.yml` already proved.
Why TWO files: verified with real `docker compose config` on this box, an
override naming a service no file in the same `-f` chain declares is a hard
error ("service chatterbox has neither an image nor a build context
specified"). The audio sidecars live in the opt-in audio overlay that
`lobes up <non-audio-role>` deliberately leaves out, so their half is paired
with that overlay and the base half with docker-compose.yml.
Why on the CARD and not a shape: which syntax the runtime accepts is a fact
about the board, true of every shape rendered over it — a shape-scoped fix
would leave a bare `lobes init` on the same board broken. Same reasoning as
the `[host_env]` iowait declaration.
`gpu_access` renders NO .env key, so every profile and shape golden — orin's
included — is byte-unchanged (regen.py reproduces them bit for bit).
UNVALIDATED (#108): `docker compose config` on the real csv-mode board renders
and merges the overrides correctly (all 12 GPU services convert, zero legacy
stanzas remain). That proves the compose merge, not a container create; only
the live boot can do that.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HzuzZKukvUXJooAZSBn2XD
…ed overrides
Profile.gpu_access ('devices' default | 'runtime') drives a generated
docker-compose.gpu.yml pair using 'deploy: !reset null' to drop the base
devices stanza. Card-scoped, so a bare machine-as-brain init on a csv-mode
board works too. Renders no .env key — every golden byte-unchanged.
Compose MERGE verified live; container create is UNVALIDATED until t9.
d6: operator profile shadowed the new builtin (#175, archived on-box). d7: live re-render blocked by the environment's permission classifier — t9/t11 and t12's measured backfill become operator follow-ups. r6: help-surface follow-up (#173). r7: MEASURED-PENDING budgets unvalidated. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HzuzZKukvUXJooAZSBn2XD
Accountability artifact for the assign-to-workforce run: every plan task accounted for, 7 drift entries classified, delivery claims carrying evidence or an explicit unverified marker. The four capability claims and the budget hypothesis are marked UNVERIFIED — no boot occurred (d7). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HzuzZKukvUXJooAZSBn2XD
Adds the orin card/profile/shape entries, the speculative-config knob, the host-dependent-test fix, and a 'Known issues found while building this release' section pointing at #171/#173/#174/#175 and the #101 correction. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HzuzZKukvUXJooAZSBn2XD
PR Summary by QodoJetson AGX Orin first-class card + orin-lobe + QAT senses profile
AI Description
Diagram
High-Level Assessment
Files changed (58)
|
Escapes the pipes in <|channel> / <|tool_call> inside table cells (they were splitting the rows), and stops a line starting with '#108' being read as an ATX heading. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HzuzZKukvUXJooAZSBn2XD
Code Review by Qodo
1.
|
![qodo-code-review[bot]](https://avatars.githubusercontent.com/in/484649?s=60&v=4){kind=small}
…t11) 0.45/262144 booted FIRST TRY at the full native window: weights 9.07 GiB, KV 11.81 GiB = 609,266 tokens = 2.32x. csv-mode runtime:nvidia proven at container CREATE (t8's mechanism, live). orin-lobe is now VALIDATED. Capability matrix vs the incumbent, same box/engine/day: VIDEO now PASSES the reversed-motion control the incumbent FAILS; image/reasoning/tools hold; audio still unusable (#101, vLLM-side). Decode +8% medium, +79% long; the incumbent's long-generation decay does not reproduce. Three blockers hit and recorded: operator profile shadowing (#175), --force destroying operator .env lines (#174), and the checkpoint shipping without vision_config.num_soft_tokens (#177, patched locally to boot). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HzuzZKukvUXJooAZSBn2XD
OriNachum
changed the title
Contributor
Author
✅ Live-validated on the physical Jetson AGX Orin — the UNVALIDATED caveat in the description is now supersededThe re-render, boot, and probe matrix all ran on the real board (2026-08-04). The hypothesis booted first try
That last line is the live proof Does it fit the machine better? Mostly yes — with one honest exception
Lighter, cheaper per KV token, faster — and the incumbent's decode decay on long generations (9.99 → 6.39) simply does not reproduce (10.81 → 11.45). The exception is concurrency headroom: ~6.7 GiB of the budget goes to peak activation that vLLM profiles "with 1 video items of the maximum feature size". That is the price of the ViT actually working — see below. Capability matrix, same box and engine as the baseline
Video is the headline. The forward clip and its reversed control now get different, correct answers — a single-frame read cannot pass that. The incumbent answered Audio remains unusable and is honestly reported as such. Tokens scale with duration (~25 tok/s, so the audio is encoded), but the discrimination control fails 0/3 — the model answers Three blockers hit on the way — all filed
Live fleet state nowStill outstanding: the gateway image is cached at 0.45.0, so |
…lease cortex->Spark verified proxied=True after the --force .env rewrite. worker->Thor wiring is in .env (keyless, peer verified live) but the gateway image installs the PUBLISHED wheel, so 0.55.0 is undeployable until #176 merges (d8). The failed build left the running fleet healthy. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HzuzZKukvUXJooAZSBn2XD
1. Orin marker false positives (lobes/machines/orin.py) — the 'orin' marker matches every Jetson Orin variant, and DetectionSignature.matches ignored the declared compute_capability/total_memory_gb, so an Orin Nano/NX would resolve to the 64GB card and inherit a senses budget (util 0.45 @ 256K) it cannot hold. Declared traits are now CONSTRAINTS when the caller supplies the corresponding probed fact; unprobed facts constrain nothing, so the legacy name-only behaviour and every existing call site are unchanged. Memory is a band, not equality — a '64GB' board reports 61.3. This also caught three tests asserting a physically impossible board: they described an Orin while passing the helper's 128 GB default. Corrected to the real 61.3, plus a new Orin-Nano regression test. 2. host_env newline corrupts .env (lobes/profiles/schema.py) — values are written verbatim into a line-oriented file, so an embedded newline SPLITS the entry rather than escaping. Rejected at profile load (where the operator can see the cause) with a last-resort guardrail in _env.set_env so no code path can write a malformed .env. Live re-verified on the physical box: detect_card() still resolves 'orin' (sm_87, 61.3 GB). Suite 2866 passed. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HzuzZKukvUXJooAZSBn2XD
The live boot the classifier initially blocked was operator-approved and ran: t9/t11/t12 move blocked -> delivered, and the four unverified capability claims are re-stated with evidence (video PASS is now a high-confidence claim; audio is marked DISPROVEN rather than merely unverified). t10 stays partial and honest: cortex verified, worker wired but gated on releasing 0.55.0 (d8). Adds d8 and the #177 checkpoint-config finding. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HzuzZKukvUXJooAZSBn2XD
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HzuzZKukvUXJooAZSBn2XD
S5799 (init.py) — implicit string concatenation that read as a missing comma; merged into one literal. S3776 (init.py) — _emit_dry_run cognitive complexity 17 > 15; split its two independent jobs into _dry_run_payload / _dry_run_lines / _dry_run_scope. Behaviour unchanged (106 init tests green). S5958 (test_init_gpu_access.py) — pytest.raises(Exception) would have passed on an incidental TypeError; narrowed to ModelGearError + EXIT_USER_ERROR + message. S6353 (schema.py) — PUSHBACK, not fixed. Python's \w is Unicode-aware, so the suggested [A-Za-z_]\w* also accepts CAFÉ_VAR and Aπ (verified). These names get written verbatim into .env, which must stay ASCII. Documented at the regex and pinned by a new test so the 'concise' refactor cannot land silently. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HzuzZKukvUXJooAZSBn2XD
|
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Makes the Jetson AGX Orin a first-class card alongside
spark/thor: a detected card, a built-in profile, a neworin-lobeshape, andsenseson the QAT checkpointunsloth/gemma-4-12B-it-qat-w4a16. The portability goal is the point — moving this setup to another architecture stays a profile pick, not a rework.What landed (8 of 13 plan tasks)
orincard detectionlobes/machines/orin.py— sm_87, verified live:detect_card()resolvesorinon the real board with no--profile. Pooling-gear divergences are orin's own overrides, deliberately not theSM_110trait — those values are an untested Thor carry-over, and borrowing an sm_110 trait would misstate why they hold.orinprofilesenses= the QAT checkpoint (compressed-tensors int4 W4A16 — weight-only, so Ampere qualifies without Blackwell FP4 tensor cores).cortex/muse/workervetoed on the hard architecture line.orin-lobeshapesenses+embedder+reranker, no localstt/tts(the Parakeet image ships no sm_87 kernels). Fixes a real clobber —thor-lobeover the orin card rendered Thor's0.30/131072, which the deployed box was hand-patching in.env.Profile.host_envlobes initis fixed too — not just this shape. Without it the gateway 429-sheds allsensestraffic indefinitely on a board whose/proc/statreports ~59% iowait with zero disk I/O.gpu_accessruntime: nvidiasurvives a re-render instead of being hand-edited back each time.MULTIMODAL_SPECULATIVE_CONFIGconfig.json. All capability rows marked pending-live-probe.Boundary proof
spark.toml,thor.toml,base.tomland every pre-existing golden are byte-identical —git diff mainover those paths is empty, and the golden diff is additions-only (neworin*files). This release adds a card; it does not move the Blackwell fleet.Verification
Every merge was TDD-gated: the suite ran green before and after each one (2725 → 2736 → 2745 → 2760 → 2816 → 2860). One branch was deliberately held back from merge until its failures were fixed rather than merged red.
Pre-existing defects this work surfaced — filed, not fixed here
These are the reviewable substance as much as the feature is. Each was found by measurement, not inspection:
docs/orin-profiles.mdoverstates the Orin KV pool 1.67× (18.86 GiB documented vs 11.29 GiB measured at identical knobs). Cause:gpu_mem_utilis a fraction of the whole device, so co-resident engines are deducted one-for-one from KV. The documented figure was measured senses-first into an empty device. Boot order changes the served budget, not just OOM risk.lobes init --forcesilently destroys operator-typed .env lines (peer wiring, API keys) — 12 keys lost on a real deployment #174 —lobes init --forcesilently destroys operator-typed.envkeys: 12 lost on a real deployment, including the cortex peer-proxy origin, flag, and credential. After it,model=cortexstops proxying and 404s with no error and no diff — and the credential cannot be re-derived. Found by rehearsing the render against a copy rather than running it live.lobes help <topic>+ aGET /helpgateway endpoint — surface the field traps that currently only live in evidence transcripts #173 — proposeslobes help <role>/GET /help, because the field traps below are discoverable only in evidence transcripts.Measured on the incumbent, before the swap (unrecoverable afterwards)
docs/evidence/2026-08-04-accept-senses-unsloth-orin.txtrecords the baseline the playbook demands, plus a capability baseline that validated the probe harness:usage.completion_tokens; the chunk-count trap reproduced (~1.8× under-report), both figures kept so the difference is auditableRIGHT-TO-LEFT— no directional discrimination, which only the control revealedenable_thinking), and streaming sends the trace ondelta.reasoningwhiledelta.contentstays empty for its entire duration — a client reading onlycontentsees a hung laneTo finish this on hardware
Boot senses first (#171 — order changes the KV pool), restore the 12 operator
.envlines immediately after (#174), then run the probe harness and backfill the measured budget into bothorin.tomlandorin-lobe.toml— a lockstep test fails CI if only one moves. Record the value that booted and any value refused; the refusal is the more useful datum.🤖 Generated with Claude Code
https://claude.ai/code/session_01HzuzZKukvUXJooAZSBn2XD