feat: Thor hosts Qwen3.6-35B-A3B (unsloth) as the eighth role 'worker' — multimodal fast doer (0.54.6) - #162
Merged
Merged
Conversation
…devague /think) Converged frame: Thor moves off the Gemma 4 31B muse and hosts unsloth/Qwen3.6-35B-A3B-NVFP4 as 'worker' — a fast ground-work DOER (repo_action allowed) with self-draft MTP. Muse goes dormant in-tree. Budget/MoE-backend/MTP values are measurement-gated (decision c20). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VqwW6JvnPf4w7tydhFHwHj
…evague /spec-to-plan) t1 catalog+tiers → t2 roles / t3 gateway → t4 template / t5 shapes / t6 CLI / t8 docs → t7 live Thor measurement → t9 acceptance run. Four non-blocking risks on t7 (sm_110 MoE backend, MTP loadability, physical-box-only, nightly-pin kernel coverage). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VqwW6JvnPf4w7tydhFHwHj
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VqwW6JvnPf4w7tydhFHwHj
PR Summary by QodoSpec/plan: add Thor “worker” role for unsloth Qwen3.6-35B-A3B (0.54.6)
AI Description
Diagram
High-Level Assessment
Files changed (8)
|
Code Review by Qodo
1.
|
![qodo-code-review[bot]](https://avatars.githubusercontent.com/in/484649?s=60&v=4){kind=small}
…ear + tier Adds the `worker` capability tier (thor-worker-lobe plan, t1): a new catalog entry for unsloth/Qwen3.6-35B-A3B-NVFP4 — an MoE checkpoint that, unlike the existing mmangkad/Qwen3.6-35B-A3B-NVFP4 candidate, ships its own MTP draft module baked in (config.json: mtp_num_hidden_layers=1, quantization ignore pattern "re:^mtp.*"). Fields verified against the checkpoint's actual config.json (native_max_model_len=262144, quantization="compressed-tensors" from quant_method) rather than card prose; hf_quant_config.json 404s (that file is a modelopt-export convention, absent from this compressed-tensors export). speculative_config matches the README's documented serve command verbatim. `TIER_ROLE` (catalog.py) and its mirror in gateway/_pressure_policy.py both gain "worker": "worker", positioned so the derived ascending capability order becomes minor < multimodal < worker < muse < main. runtime._parser.infer_parser already resolves the id to qwen3_coder via the existing "qwen3.6" marker — no _parser.py change needed. Extends the existing MTP-carrier test guard (test_catalog.py) with an explicit, cited allow-list for checkpoints that carry their own MTP draft weights without an "-MTP" id suffix, since this entry is the first such case. Full suite: 2598 passed, 15 skipped. black/isort/flake8/bandit clean.
unsloth/Qwen3.6-35B-A3B-NVFP4 catalog entry (role_hint=worker, 262144 native, compressed-tensors, self-hosted MTP); TIER_ROLE + pressure-policy tiers gain worker at minor<multimodal<worker<muse<main. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VqwW6JvnPf4w7tydhFHwHj
Mirrors muse end-to-end in lobes/gateway/_config.py (thor-worker-lobe plan,
t3): _DEFAULT_WORKER, FEASIBLE_ENV/PEER_ORIGIN_ENV/PEER_PROXY_ENV/
PEER_API_KEY_ENV all gain a "worker" entry riding the same WORKER_* channel
convention, OPT_IN_BACKENDS becomes frozenset({"muse", "worker"}) so an
unwired worker is honestly infeasible by default (model=worker 404s
role_infeasible, never falls back), and build_config gains the
WORKER_BASE_URL-gated backend block.
Tests mirror muse's coverage across the four owned files: BASE_URL-gated
wiring (test_gateway_config_wiring.py), default-infeasible-when-unwired +
explicit-FEASIBLE-wins (test_gateway_feasibility.py), the four peer channels
(test_gateway_config_proxy.py), and — since worker's own backend wiring
already threads through the existing data-plane proxy machinery with no
server.py change needed — a wired-but-dropped forwarding case
(test_gateway_proxy.py).
t1 (catalog + TIER_ROLE) is already merged into this branch; t2 (roles.py
registry) is not, and is intentionally out of scope here.
WORKER_BASE_URL-gated backend; worker joins OPT_IN_BACKENDS (unwired => infeasible); WORKER_* on the four peer channels. Mirrors muse. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VqwW6JvnPf4w7tydhFHwHj
Widen lobes.roles + lobes.profiles.schema to the eighth Colleague-facing role, `worker` (Qwen3.6-35B-A3B NVFP4, the fast ground-work DOER), added structurally exactly like the seventh role `muse`: - roles.ROLES gains "worker" (after muse, before embedder); every ROLE_* registry dict gets a worker entry: ROLE_BACKEND["worker"]="worker", ROLE_ROLE_HINT["worker"]="worker", ROLE_PATH["worker"]=/v1/chat/completions, ROLE_MAX_MODEL_LEN_ENV["worker"]="WORKER_MAX_MODEL_LEN". - build_role_registry's gateway-fronted generate loop gains "worker" — it flows through the SAME #92/#115 ready/feasible/loaded clamp as muse, so an unwired worker (its gateway backend is sibling task t3) is honestly loaded=False, ready=False (clamped even against a stray live True signal). - The doer contract: ROLE_RESPONSIBILITIES["worker"] carries execution/ground_work/bulk_transform/drafting PLUS tool_use AND repo_action — worker is the first role besides cortex permitted repo_action (it may ACT on the repo under cortex's direction). ROLE_FORBIDDEN["worker"] == ("final_decision", "security_decision") — it never decides. - schema.ROLES gains "worker" so a Profile may carry worker's full per-machine knob set; the unknown-role override rejection is unchanged. Tests added/updated in the three owned files (test_roles.py, test_roles_proxied.py, test_shapes.py schema-ROLES part). Sibling wiring (gateway backend + OPT_IN_BACKENDS t3, profiles render/shapes ROLE_ENV_PREFIX/ ROLE_SERVICE/OPT_IN_CORE_ROLES + thor-worker-lobe shape, measure, capabilities goldens) is intentionally out of scope and lands in later waves. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016hSuy5fv3zbCyjeAQ6b6oF
worker wired into every ROLE_* dict mirroring muse; doer contract (repo_action allowed, forbidden final_decision/security_decision); schema ROLES gains worker. NOTE: tree needs the opt-in-core + per-role map integration (next commit) to be green. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VqwW6JvnPf4w7tydhFHwHj
…imodal The 8th-role vocabulary widening (t2) is inseparable from its opt-in-core designation (t5): worker must be in OPT_IN_CORE_ROLES or it reads as a default-hosted role and the goldens break. Executed t5 inline so the tree greens atomically with t2: - shapes.OPT_IN_CORE_ROLES += worker; base.toml vetoes worker on unknown cards - per-role maps gain worker: render.ROLE_ENV_PREFIX, shape_render.ROLE_SERVICE + OPT_IN_CORE_ACTIVATION_ENV/COMPOSE_PROFILE, roles_measure family - goldens regenerated: recognized-card renders byte-identical; base-card renders gain WORKER_FEASIBLE=false exactly like MUSE_FEASIBLE=false Multimodal reconciliation (operator decision 2026-07-31): the checkpoint is image+video (config.json vision_config, image/video token ids, no audio), so worker is served MULTIMODAL — a seeing doer. catalog shape notes the ViT; worker responsibilities gain image_understanding/video_understanding; the compose lane must NOT pass --language-model-only (t4). Test-expectation updates + up.py service map follow in the next commit. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VqwW6JvnPf4w7tydhFHwHj
…llow-through) The worker checkpoint is multimodal (image+video); t9's acceptance was text-only. Added three honest-outcome live probes: ground-truth image (with negative control, like senses), video, and a thinking+coding task with MTP active — each recording the real verdict, degrading the multimodal claim honestly if vision/video does not serve (the #101 risk). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VqwW6JvnPf4w7tydhFHwHj
… length, context measurement t9: TTFT + decode_tps + derived ms/token + prefill_tps (MTP on vs off with draft-acceptance), open-ended writing-length + sustained decode, and a long-context prefill probe confirming the effective served window. t7: KV pool size + concurrency multiple at the served context (thor-muse pattern) so served context is a measured capacity, not a flag value. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VqwW6JvnPf4w7tydhFHwHj
The worker role (unsloth/Qwen3.6-35B-A3B-NVFP4, opt-in-core multimodal
doer) landed its plumbing (roles.py, schema.py, gateway _config.py,
catalog.py, shapes.py, render.py, roles_measure.py) but `lobes up`'s own
hardcoded ROLE_SERVICE dict and a few CLI help strings still lacked it,
and the widened 8-role vocabulary broke a set of test expectation
literals still pinned to 7 roles.
- lobes/cli/_commands/up.py: add "worker": "vllm-worker" to ROLE_SERVICE
(mirrors "muse": "vllm-muse"); add worker to the docstring's
role->service table and the `role` argument's help string.
- lobes/cli/_commands/capabilities.py, overview.py, measure.py: bump the
hardcoded "seven first-class roles (cortex/senses/muse/embedder/...)"
help/docstring prose to eight roles with worker in capability order.
- tests/test_profile_schema.py, test_profile_render.py: ROLES /
ROLE_ENV_PREFIX literals gain worker.
- tests/test_shapes.py: COLLEAGUE_ROLES / SHAPE_ROLES gain worker;
OPT_IN_CORE_ROLES == ("muse", "worker"); DEFAULT_HOSTED_ROLES stays
worker-free (opt-in-core is excluded from the default-hosted set).
- tests/test_dropped_lobe_honesty.py: spark-lobe/thor-lobe CLI/gateway
agreement dicts gain "worker": False (unwired on both shapes).
- tests/test_roles.py: worker's feasible flag flips True -> False (an
unwired opt-in-core role is honestly infeasible by default, exactly
like muse - OPT_IN_BACKENDS); the responsibilities assertions gain
image_understanding/video_understanding (worker is multimodal).
- tests/test_cli_capabilities.py: `lobes endpoint` role-coverage dict
gains worker (resolves to the gateway URL, same as muse, despite
being infeasible here).
- tests/test_cli_up.py: rename the ROLE_SERVICE-coverage test off a
stale "six roles" count now that adding worker makes it pass again.
- tests/test_cli_measure.py: _FAMILY_BY_ROLE gains worker (family
"llm"); rename off the stale "six roles" count.
- tests/test_shape_contract_matrix.py: worker is opt-in-core with no
hosting shape built yet, so every built-in shape now drops it too -
add a (<shape>, "worker") cell per shape alongside each existing
(<shape>, "muse") cell, plus worker's model id and *_BASE_URL/
*_SERVED_NAME wire-env keys in the shared helper dicts.
- tests/test_proxy_integration.py: regenerate the golden /capabilities
byte literal with a "worker" entry (infeasible, unloaded, endpoint
resolves to the gateway origin) inserted in ROLES order between
"muse" and "embedder". Verified against the real
build_role_registry() payload for the golden env before pasting it
in, not just green-until-it-passes.
Left red on purpose: test_shape_goldens.py::
test_role_service_constants_exist_in_compose_templates still fails
because the vllm-worker compose service doesn't exist yet - that's a
parallel task's (t4) scope, not touched here.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016hSuy5fv3zbCyjeAQ6b6oF
… suite up.py ROLE_SERVICE gains worker; capabilities/overview/measure prose bumped to eight roles; all shape/vocab/contract/CLI test expectations updated for the worker vocabulary + multimodal responsibilities; proxy capabilities golden regenerated from the real registry. Only test_shape_goldens' compose-service check remains red (t4 adds the vllm-worker service). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VqwW6JvnPf4w7tydhFHwHj
…3.5 MoE)
The eighth Colleague role `worker` (unsloth/Qwen3.6-35B-A3B-NVFP4) now has a
real compose service. Mirrors the vllm-muse opt-in-core precedent (profile
`worker`, no host port, embed/rerank health-gated boot ordering, gateway
backend wired only via WORKER_BASE_URL) but adapted for the Qwen multimodal
MoE:
- Rides the SAME Qwen/vLLM nightly lane (VLLM_NIGHTLY_IMAGE) as the primary,
overridable via WORKER_IMAGE — NOT the Gemma lobes/vllm-gemma4:local tag.
- Served MULTIMODAL (image+video): deliberately NO --language-model-only.
- --quantization=compressed-tensors (the checkpoint's own quant_method).
- Baked-in self-draft MTP (--speculative-config method=mtp, no external draft).
- Qwen parser pair: --tool-call-parser=qwen3_coder --reasoning-parser=qwen3.
- --moe-backend overridable (flashinfer_b12x is Spark's sm_121 pick; Thor
sm_110 UNCONFIRMED until the t7 live boot). GPU util 0.30 is a PLACEHOLDER
measured live at t7, like thor-muse's 0.55.
Gateway environment gains the WORKER_* passthroughs (BASE_URL / SERVED_NAME /
MAX_MODEL_LEN / FEASIBLE / PEER_ORIGIN / PEER_PROXY / PEER_API_KEY), mirroring
muse. env.example documents every WORKER_* knob (opt-in, commented-out block).
Tests: new tests/test_worker_compose.py (profile gating, Qwen-lane image,
parser pair, self-draft MTP, no --language-model-only, gateway wiring, plus a
live `docker compose config` render). Greens
test_shape_goldens::test_role_service_constants_exist_in_compose_templates.
Recomputed the tool-parser tripwire (new vllm-worker + changed gateway hash;
every other service byte-identical) and regenerated the template-defaults
golden for the new ${WORKER_*:-default} surface.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016hSuy5fv3zbCyjeAQ6b6oF
vllm-worker on the Qwen nightly lane (NOT gemma image): compressed-tensors, --max-model-len 262144, self-draft MTP, qwen3_coder+qwen3 parser pair, NO --language-model-only (serves image+video). WORKER_* knobs in env.example + gateway passthroughs. Greens test_shape_goldens; tripwires (tool-parser service set, template-defaults golden) reconciled. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VqwW6JvnPf4w7tydhFHwHj
…(t8) Documents the thor-worker-lobe plan's eighth Colleague role (unsloth/ Qwen3.6-35B-A3B-NVFP4, the fast ground-work DOER, first non-cortex role permitted repo_action) across colleague-stack.md, deployment-shapes.md, gateway-fleet.md, openai-api.md, machine-profiles.md, qwen3.6-35b-a3b-nvfp4.md, gemma-4-31b-nvfp4.md, and CLAUDE.md. Every "cortex is the only lobe that acts" style claim is updated to reflect worker's narrower repo_action grant. Muse is documented dormant/unhosted mesh-wide (the Thor moved to hosting worker instead) while its role, catalog entry, and thor-muse shape stay in-tree. CLAUDE.md's stale "pressure degrades to minor" wording is corrected to match the shipped code (429 shed, no degrade-to-minor substitution). Worker's shape/compose-service/live-measured-budget pieces are described as forthcoming (thor-worker-lobe plan tasks t4/t6/t7) — no number is invented. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016hSuy5fv3zbCyjeAQ6b6oF
colleague-stack/deployment-shapes/gateway-fleet/qwen3.6-35b-a3b docs + CLAUDE.md updated to eight roles; worker's widened division of labour (may act) recorded; muse documented dormant/unhosted (in-tree, cite-don't-delete); stale 'degrades to minor' wording fixed to the 429-shed behaviour. Budget/MoE/MTP kept declared/UNVALIDATED (t7). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VqwW6JvnPf4w7tydhFHwHj
lobes up worker is wired (t6) — only the thor-worker hosting shape is forthcoming (t7); CHANGELOG 0.54.6 now describes the shipped multimodal worker role + muse dormancy + the pressure-shed doc fix. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VqwW6JvnPf4w7tydhFHwHj
Contributor
Author
Implementation landed — code side complete, fully greenThe fannable waves (t1–t6, t8) are merged on this branch; the suite is 2678 passed / 0 failed, lint clean (black/isort/flake8/bandit), and the What shippedThe
Mid-flight correction (operator catch)The checkpoint is multimodal (image+video) — MuseDormant/unhosted mesh-wide ( NOT in this PR — the on-box phase (DECLARED/UNVALIDATED, #108)
No performance number or budget value is committed until t7/t9 measure it live. 🤖 Generated with Claude Code |
OriNachum
changed the title
Live boot on the physical Jetson AGX Thor (2026-07-31, docs/evidence/2026-07-31-accept-worker-thor.txt): - boots + serves at util 0.45 / 262144: KV 41.78 GiB = 14.07x concurrency, weights 24.81 GiB; MTP self-draft loads + accepts 89.1%; decode 50.8 tok/s (thinking) / 73.5 sustained; image intake correct (red/blue + neg control); model=worker routes via gateway, model=muse 404s role_infeasible. - KEY FIX: the vllm-worker lane must NOT force --moe-backend on sm_110 (every forced NVFP4 backend refused; vLLM auto-selects TRITON + a modular NVFP4 kernel per path). Template no longer forces it; catalog moe_backend=. thor-worker.toml lands with MEASURED budget; catalog status load-tested; init --shape gains thor-worker; goldens + contract-matrix + compose tests reconciled. Video intake still unverified (image proven). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VqwW6JvnPf4w7tydhFHwHj
Real Spark-webcam clip through worker's video_url path: 1572 video tokens decoded, accurate scene/subject/motion description (finish=stop), NOT a #101-style drop. Image was already proven; video was the last open capability. worker is now a fully-validated multimodal (image+video) seeing doer. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VqwW6JvnPf4w7tydhFHwHj
The 4 genuinely-open S8997 issues: test_chatterbox_pcm16.py (numpy-available toggle) and test_readiness_peer_probe.py (capturing-handler class attrs) now use monkeypatch (auto-restored) instead of manual save/restore. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VqwW6JvnPf4w7tydhFHwHj
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VqwW6JvnPf4w7tydhFHwHj
python:S3776 was the only genuinely open SonarCloud issue on lobes/realtime/tts_client.py (18 against the 15 limit, line 330); the other four issues in the same historical batch (S8513 x2, S8572, S8786) were already fixed and merged on main in #157 (0.54.3) — confirmed via a live SonarCloud query before touching anything, so this change is scoped to S3776 only. Extracted the per-attempt try/except tree out of _synthesize_single into a new _attempt_synthesize() helper, plus a tiny _retry_or_give_up() shared by its two transient-network-error arms (ConnectError, ReadTimeout). The retry loop in _synthesize_single now just awaits the helper and checks the _Retry sentinel — no other control flow moved. Verified behavior-identical: same log messages, same t0/elapsed placement (client construction before the timer, exactly as before), same retry-once-then-give-up semantics, same log.exception in the catch-all arm. Verified locally with flake8-cognitive-complexity (matches SonarCloud's 18 exactly pre-change; 6 post-change for _synthesize_single, all other functions unchanged) since a live Sonar rescan isn't available from this worktree. black/isort/flake8/bandit clean; 2609 passed / 11 skipped on the full suite (skips are all live-deployment gated, unrelated), including the httpx-gated tts_client tests run with `--with httpx --with numpy` to exercise this file's own logic directly rather than relying on the (correct) offline skip.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VqwW6JvnPf4w7tydhFHwHj
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VqwW6JvnPf4w7tydhFHwHj
…r) into the worker PR # Conflicts: # CHANGELOG.md # pyproject.toml
…ale thor-muse anchor)
SonarCloud python:S9073 flagged eight `assert A and B` composite assertions in test_readiness_peer_probe.py and test_proxy_integration.py. Split each into separate top-level asserts so a failure pinpoints which half tripped, with no change in what's actually being verified. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016hSuy5fv3zbCyjeAQ6b6oF
SonarCloud python:S9073 flags `assert A and B` — split each top-level `and` in an assert statement into separate asserts across test_roles.py, test_worker_compose.py, and test_init_shape.py so a failure pinpoints which half broke. Nested `and`s inside `not (...)` or `any(... for ...)` are untouched — they aren't the composite top-level assertion the rule targets. python:S9083 flags empty-parens decorators; drop the `()` from `@pytest.fixture()` in test_cli_measure.py. Behavior-identical: same conditions asserted, same fixture. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016hSuy5fv3zbCyjeAQ6b6oF
Each `assert A and B` becomes two separate asserts so a failure pinpoints which half failed. Behavior-identical; all 47 tests still pass.
|
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.
What
Converged devague spec + plan for the worker role: Thor moves off the Gemma 4 31B
museand hostsunsloth/Qwen3.6-35B-A3B-NVFP4— a fast ground-work doer lobe with self-draft MTP.docs/specs/2026-07-31-thor-worker-lobe-qwen3-6-35b-a3b.md(frame: 19 confirmed claims, 17 confirmed honesty conditions, 8 scope entries with provenance)docs/plans/2026-07-31-thor-worker-lobe-qwen3-6-35b-a3b.md(9 tasks, 5 waves, 4 first-class risks)Key decisions (operator-confirmed)
repo_actionallowed — the first role besides cortex permitted to act; forbidden onlyfinal_decision/security_decision.minor < multimodal < worker < muse < main; worker's role name is its tier, like muse.MUSE_PEER_ORIGIN,model=muse404srole_infeasiblewith no referral. Code/shape/catalog stay in-tree (cite-don't-delete).gpu_mem_util/max_model_len, the sm_110 MoE backend, and the MTP verdict are measured by a live boot on the physical Thor before any value is committed — the thor-muse 0.40→0.55 lesson. Role ships DECLARED/UNVALIDATED until the acceptance transcript lands underdocs/evidence/(spec+plan: lobes fits the machine it lands on — per-machine hardware profiles #108).Checkpoint facts (verified on HF 2026-07-31)
MoE 35B / ~3B active, 262144-token native (→1M), ships its own MTP draft module (
--speculative-config '{"method": "mtp", "num_speculative_tokens": 2}'). Distinct from the in-catalogmmangkadsibling (32K native, MTP not carried). The card'sflashinfer_b12x+sm_121aMoE advice is Spark-specific — Thor's sm_110 backend is a live measurement (risk r1).Not in this PR
Implementation (waves t1–t9) follows via /assign-to-workforce; t7/t9 run on the physical Thor.
🤖 Generated with Claude Code
https://claude.ai/code/session_01VqwW6JvnPf4w7tydhFHwHj