lobes-cli 0.91.0__tar.gz → 0.92.0__tar.gz

{lobes_cli-0.91.0 → lobes_cli-0.92.0}/CHANGELOG.md

@@ -4,6 +4,16 @@ All notable changes to this project are documented here. The format is based on
 [Keep a Changelog](https://keepachangelog.com/en/1.1.0/), and this project
 adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [0.92.0] - 2026-10-26
+
+### Added
+
+- Catalog entry `nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB` (74B-A9B MoE hybrid, text+image+video, 262144 native, modelopt mixed IQ2/Q8/NVFP4, DFlash LiLiCorr draft) with its doc, a `qwen3_coder` parser rule, and a Spark GB10 acceptance transcript (text, image, tool call and a 231,935-token needle pass).
+
+### Changed
+
+- `worker` is now Nemotron 3.5 Bolt (operator decision, 2026-10-26), hosted on the DGX Spark; `nvidia/Qwen3.6-35B-A3B-NVFP4` is demoted to a kept candidate. The worker contract is unchanged. Bolt needs the card's vLLM nightly plus `b12x` and is sm_12x-only; the packaged `vllm-worker` lane cannot express its flags yet, so the Spark serves it through a full command in its hand-kept override.
+
 ## [0.91.0] - 2026-10-23
 
 ### Added

{lobes_cli-0.91.0 → lobes_cli-0.92.0}/CLAUDE.md

@@ -345,6 +345,19 @@ entry, and the `thor-muse` shape all **stay in-tree** (cite-don't-delete) —
 dormant, not deleted — and the tier vocabulary above still ranks `worker` <
 `muse`. See `docs/gemma-4-31b-nvfp4.md`.
 
+**`worker` re-checkpointed again (2026-10-26, operator decision): Nemotron 3.5
+Bolt 74B-A9B 24GB, hosted on the DGX Spark.** The catalog's `worker` role_hint
+moved from `nvidia/Qwen3.6-35B-A3B-NVFP4` (now a kept candidate) to
+`nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB`: a ~9B-active MoE hybrid with a
+RADIO ViT (image+video), 262144 native, DFlash from its LiLiCorr draft, the
+same multimodal-coder contract. Its IQ2 experts run only on sm_12x, so the
+Spark hosts it (cortex is dropped there and served by spark2) and the Thor
+stops hosting worker. The packaged `vllm-worker` lane cannot express its flags
+yet; the Spark serves it through a full command in its hand-kept override. See
+`docs/nemotron-3.5-bolt-74b-a9b-24gb.md` and
+`docs/evidence/2026-10-26-accept-nemotron-bolt-spark.txt`. The paragraphs
+below are the record of the Qwen and Lightning workers before it.
+
 **`worker` — the eighth role (opt-in hosting), the fast ground-work DOER —
 RE-CHECKPOINTED AGAIN, issue #244 (2026-09-10), off Lightning onto
 `nvidia/Qwen3.6-35B-A3B-NVFP4`.** Deviation d1 (2026-08-20) had put worker on
@@ -383,8 +396,8 @@ exactly (`WORKER_FEASIBLE` / `WORKER_PEER_ORIGIN` / `WORKER_PEER_PROXY` /
 like `muse`). Under pressure `worker` sheds (429) exactly like
 cortex/senses/muse.
 
-**Hosted on the Jetson AGX Thor again (#244); the DGX Spark reaches it by
-proxy (2026-09-11).** The Thor serves the lane locally through the
+**(Superseded 2026-10-26 — see the Bolt paragraph above.) Hosted on the Jetson
+AGX Thor again (#244); the DGX Spark reaches it by proxy (2026-09-11).** The Thor serves the lane locally through the
 `thor-worker` shape at `gpu_mem_util=0.45`, full `max_model_len=262144`,
 DFlash speculation, and `WORKER_MAX_NUM_SEQS=1`. One request at a time is
 **by design** and is still to be tested under a real multi-agent load. See

{lobes_cli-0.91.0 → lobes_cli-0.92.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: lobes-cli
-Version: 0.91.0
+Version: 0.92.0
 Summary: lobes — run, assess, and switch the local vLLM model.
 Project-URL: Homepage, https://github.com/agentculture/lobes-cli
 Project-URL: Issues, https://github.com/agentculture/lobes-cli/issues

lobes_cli-0.92.0/docs/evidence/2026-10-26-accept-nemotron-bolt-spark.txt

@@ -0,0 +1,117 @@
+ACCEPTANCE — nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB served standalone on the DGX Spark GB10
+Date: 2026-10-25 22:36Z - 22:47Z (UTC)
+Box: DGX Spark GB10 (sm_121, 128 GB unified memory)
+Scope: one standalone vLLM container via sparkrun, OUTSIDE the lobes fleet compose and gateway.
+       Proves the checkpoint boots and serves on this card with the recipe below. It does NOT
+       prove `lobes switch`, a fleet lane, a shape or a gateway route — none of those exist for it.
+
+== Checkpoint ==
+model:      nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB @ c6cc2e8ffa353403213d6b9b18aaf8eb684d705a
+draft:      nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-LiLiCorr @ 449d24140160b8e51871bfc0f0a7338abf640aa1
+config.json: NemotronHPuzzleForCausalLM in a nemotron_h_omni wrapper; RADIO vision_config; no sound_config;
+             llm_config.max_position_embeddings 262144
+quantization_config: quant_method modelopt, quant_algo MIXED_PRECISION
+             IQ2_XXS 27,728 layers / Q8_0 113 / W4A16_NVFP4 80
+
+== Engine ==
+image:  vllm/vllm-openai:nightly-77154ca6116ae34cbfe314beef1016bde9b63989 + b12x 1.5.1 baked in
+        (the card installs b12x at container start; baked so the boot runs offline)
+vllm:   0.32.0rc1.dev100+g77154ca61   torch 2.13.0+cu130
+Resolved architecture: NemotronH_Omni_Reasoning_V3 (target), LiLiCorrDraftModel (draft)
+
+== Serve flags (the card's vLLM command, with the deviations marked) ==
+--revision c6cc2e8f... --trust-remote-code --load-format auto --tensor-parallel-size 1
+--max-model-len 262144 --max-num-seqs 1 --enable-chunked-prefill --kv-cache-dtype fp8
+--linear-backend b12x --moe-backend b12x --attention-backend FLASHINFER
+--mamba-backend flashinfer --mamba-ssm-cache-dtype bfloat16 --mamba-cache-dtype bfloat16 --mamba-cache-mode align
+--limit-mm-per-prompt '{"video":1,"image":1}' --video-pruning-rate 0.5
+--media-io-kwargs '{"video":{"fps":2,"num_frames":256}}'
+--enable-auto-tool-choice --tool-call-parser qwen3_coder
+--reasoning-parser-plugin <checkpoint>/ultra_v3_reasoning_parser.py --reasoning-parser ultra_v3
+--speculative-config '{"method":"dflash","model":"...-LiLiCorr","revision":"449d2414...","num_speculative_tokens":7,
+                       "rejection_sample_method":"standard","attention_backend":"FLASHINFER","draft_load_config":{"load_format":"auto"}}'
+--seed 42
+DEVIATION  --kv-cache-memory-bytes 4294967296 (4 GiB) added; --gpu-memory-utilization 0.40 then only gates the
+           startup free-memory check. Reason: on this unified-memory box a util-sized KV pool is not
+           deterministic (vLLM's profiling counts other processes' memory growth), measured the same night on a
+           different checkpoint at util 0.70 / 0.60 / 0.63 -> 22.99 / 5.87 / 2.94 GiB.
+DEVIATION  --allowed-local-media-path / dropped (it lets any caller read any file on the box via file:// URLs).
+DEVIATION  revisions pinned; bound to 127.0.0.1.
+
+== Boot ==
+22:39:14  Loading weights took 125.55 seconds (target); 22:39:34 5.63 seconds (draft)
+22:39:35  Model loading took 24.09 GiB memory and 150.801226 seconds
+22:40:18  Initial free memory 105.18 GiB, reserved 4.0 GiB memory for KV Cache as specified by kv_cache_memory_bytes
+22:40:18  GPU KV cache size: 350,464 tokens, Maximum concurrency for 262,144 tokens per request: 1.34x
+22:40:30  Application startup complete; inference ready
+WARNING   max_num_scheduled_tokens is set to 2048 based on the speculative decoding settings (vLLM's own note)
+
+== Probes (temperature 1.0, top_p 0.95, thinking on by default) ==
+/v1/models: ['nemotron-3.5-bolt', 'nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB']
+known-answer  "What is 17*23?"                 -> '391', reasoning split off (149 chars)        0.8 s   PASS
+image         64x64 solid red PNG              -> 'Red'                                         2.0 s   PASS
+image         64x64 solid blue PNG             -> 'Blue'                                        3.1 s   PASS
+image         64x64 solid green PNG            -> 'Green'                                       2.2 s   PASS
+tool call     get_weather, "weather in Haifa?" -> finish_reason tool_calls,
+                                                  get_weather({"city": "Haifa"}), content None  1.2 s   PASS
+speed         palindrome function              -> 359 completion tokens in 5.2 s wall = 69.3 tok/s
+                                                  (one request; wall clock includes prefill and reasoning)
+long context  231,935-token prompt, needle mid-document
+                                               -> 'PELICAN-4471'                              200.3 s   PASS
+The image probes are synthetic solid colours: they prove the vision path is wired and the colour
+negative controls separate, not real-image quality.
+
+== Host after the probes ==
+host MemAvailable  70,146 MiB
+kernel NV_ERR_NO_MEMORY lines since the boot: 0
+hottest ACPI zone  86 C (just after the 200 s long-context prefill)
+
+== Not proven here ==
+- Any lobes path: `lobes switch` (the stock template's image lacks the nightly + b12x), a fleet lane, a gateway route.
+- Other cards: the IQ2 expert path is sm_12x-only in vLLM; the Orin (sm_87) refuses it, the Thor (sm_110) is untried.
+- Concurrency above 1, sustained decode speed under load, DFlash acceptance rates.
+
+======================================================================================================
+PART 2 — Bolt as THE `worker`, served by the Spark's lobes `vllm-worker` lane (2026-10-25 22:56Z - 23:20Z)
+======================================================================================================
+Operator decision: `worker` = Bolt, hosted on the Spark (sm_12x only); cortex dropped on the Spark and served
+by spark2 alone; the Thor stops hosting its nvidia/Qwen3.6-35B-A3B-NVFP4 worker and reaches `worker` by mesh.
+
+== Spark deployment (hand-kept files, each backed up first) ==
+.env:  WORKER_MODEL / WORKER_SERVED_NAME = nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB, WORKER_MAX_MODEL_LEN=262144,
+       WORKER_QUANTIZATION=modelopt_mixed (what vLLM resolves for this config), WORKER_GPU_MEM_UTIL=0.40,
+       WORKER_IMAGE=<the nightly + b12x image above>, WORKER_FEASIBLE=true, COMPOSE_PROFILES=worker,
+       PRIMARY_FEASIBLE=false
+docker-compose.override.yml: vllm-worker `command:` replaced with the PART 1 flag list (the packaged lane cannot
+       express --linear-backend, the mamba cache flags, --kv-cache-memory-bytes, the media limits or the
+       reasoning-parser plugin), `depends_on: !reset null` (the pooling lanes it waits on are not hosted here)
+docker-compose.shape.yml: vllm-primary parked in the inert shape-dropped profile
+`lobes up worker --apply`, then `lobes up gateway --apply`
+
+== Lane boot (cold compile cache) ==
+22:59:20  Model loading took 24.09 GiB memory and 155.386063 seconds
+23:01:24  GPU KV cache size: 350,464 tokens, Maximum concurrency for 262,144 tokens per request: 1.34x
+          Application startup complete; container healthy
+
+== Thor ==
+.env backed up; WORKER_FEASIBLE=false (explicit), worker dropped from COMPOSE_PROFILES, WORKER_MODEL /
+WORKER_SERVED_NAME set to the Bolt id (advert only). `lobes up worker --down --apply` (with worker still in
+COMPOSE_PROFILES: the verb refuses a role the .env does not activate), then `lobes up gateway --apply`.
+Host MemAvailable after the gateway was healthy: 62 GiB.
+
+== Through each member's gateway (API key on) ==
+                 /capabilities worker                                           model=worker   model=cortex
+spark (host)     feasible true, ready true, context 262144, quant modelopt      '391' local    '391' via spark2
+thor             feasible false, ready true, hosted_by spark, proxied true      '391' via spark '391' via spark2
+spark2           feasible false, ready true, hosted_by spark, proxied true      '391' via spark '391' local
+Every forward carried X-Lobes-Mesh-Member (spark for worker, spark2 for cortex) and X-Lobes-Proxied-By.
+
+Via the Thor's gateway (mesh forward to the Spark lane):
+image  64x64 solid red  -> 'red'   5.5 s   PASS
+image  64x64 solid blue -> 'Blue'  2.1 s   PASS
+tool   get_weather      -> finish_reason tool_calls, get_weather({"city": "Haifa"})  1.9 s   PASS
+
+== Not proven by PART 2 ==
+- The packaged template serving Bolt (it cannot yet; the Spark uses an override command).
+- Load: `--max-num-seqs 1`, so concurrent worker callers queue.
+- The Orin member's view (not probed).

lobes_cli-0.92.0/docs/nemotron-3.5-bolt-74b-a9b-24gb.md

@@ -0,0 +1,105 @@
+# NVIDIA Nemotron 3.5 Bolt 74B-A9B 24GB — the `worker` role
+
+> One entry in lobes's **supported catalog** (`lobes overview --list`). For the
+> catalog-vs-warm distinction — what you *can* load vs. what's loaded *now* —
+> see
+> [`gateway-fleet.md`](gateway-fleet.md#supported-catalog-vs-warm-backends).
+>
+> **Status: the `worker` role_hint (operator decision, 2026-10-26)**,
+> replacing `nvidia/Qwen3.6-35B-A3B-NVFP4` (now a kept candidate — see
+> [`nvidia-qwen3.6-35b-a3b-nvfp4.md`](nvidia-qwen3.6-35b-a3b-nvfp4.md)). The
+> worker contract does not move: Bolt is a fast multimodal (image+video)
+> coder/doer with tool calls, which is exactly what `roles.py` already
+> promises. It is hosted on the **DGX Spark** — the only card family its IQ2
+> experts run on — and every other mesh member reaches `worker` through the
+> mesh. Measured as a standalone container first
+> ([`evidence/2026-10-26-accept-nemotron-bolt-spark.txt`](evidence/2026-10-26-accept-nemotron-bolt-spark.txt)).
+>
+> **The packaged template cannot express Bolt's flags yet.** The fleet
+> `vllm-worker` lane has no knob for `--linear-backend`, the mamba cache
+> flags, `--kv-cache-memory-bytes`, the media limits or the reasoning-parser
+> plugin, and it always passes `--quantization`. The Spark serves it through a
+> full `command:` replacement in its hand-kept `docker-compose.override.yml`
+> (the [recipe](#serving-it-on-the-spark) below). `lobes switch` to this id
+> will not boot on the stock template either (see [Engine](#engine)).
+
+## What it is
+
+| | |
+|---|---|
+| id | `nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB` |
+| revision | `c6cc2e8ffa353403213d6b9b18aaf8eb684d705a` |
+| architecture | `NemotronHPuzzleForCausalLM` in a `nemotron_h_omni` wrapper: hybrid Mamba-2 / attention with sparse MoE, about 9B active of 74B total |
+| modalities | text, image and video (RADIO vision encoder); no audio |
+| context | 262144 native (`llm_config.max_position_embeddings`) |
+| quantization | `modelopt_mixed` (the checkpoint's `quant_method` is modelopt, `MIXED_PRECISION`): IQ2_XXS on the routed experts (~2.06 bits, ggml packing), Q8_0, and W4A16 NVFP4 on the mixer projections — 24.09 GiB loaded |
+| tool parser | `qwen3_coder` (the card's own serve command) |
+| reasoning parser | `ultra_v3`, shipped as a plugin file inside the checkpoint |
+| speculative decoding | DFlash, 7 draft tokens, from `nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-LiLiCorr` @ `449d2414…` |
+| licence | OpenMDW 1.1 (the card) |
+
+All of the above was read from the checkpoint's own `config.json` and
+`hf_quant_config.json` at the pinned revision, and from its card's vLLM
+command.
+
+## Engine
+
+- **sm_12x only.** vLLM's IQ2 expert path is gated to Blackwell consumer/GB10
+  silicon. The Jetson AGX Orin (sm_87) refuses it on v0.32.0 and on main; the
+  Jetson AGX Thor (sm_110) has not been tried.
+- **Nightly plus `b12x`.** The card names
+  `vllm/vllm-openai:nightly-77154ca6116ae34cbfe314beef1016bde9b63989` and
+  installs `b12x` at container start (`--linear-backend b12x --moe-backend
+  b12x`). The measured run used that nightly with `b12x==1.5.1` baked into a
+  derived image so the boot runs offline (vLLM `0.32.0rc1.dev100+g77154ca61`).
+
+## Serving it on the Spark
+
+The measured recipe is the card's vLLM command with three deviations:
+
+1. **`--kv-cache-memory-bytes 4294967296` (4 GiB).** On the Spark's unified
+   memory a KV pool sized by `--gpu-memory-utilization` is not
+   deterministic: vLLM's profiling counts other processes' memory growth, so
+   the same flag gives very different pools from boot to boot. Pinning the
+   pool makes the footprint predictable. 4 GiB holds 350,464 fp8 tokens = 1.34
+   full 262144-token requests; `--max-num-seqs` is 1 as on the card.
+2. **`--allowed-local-media-path /` dropped.** That flag lets any caller read
+   any file on the box through a `file://` URL.
+3. **Revisions pinned**, and the server bound to `127.0.0.1`.
+
+The full flag list is in the evidence transcript.
+
+## Measured — DGX Spark GB10, 2026-10-25
+
+From
+[`evidence/2026-10-26-accept-nemotron-bolt-spark.txt`](evidence/2026-10-26-accept-nemotron-bolt-spark.txt):
+
+| | |
+|---|---|
+| weights | 24.09 GiB, 150.8 s to load |
+| KV pool | 4.0 GiB pinned = 350,464 tokens (1.34× at 262144) |
+| host memory available after the probes | 70,146 MiB |
+| known-answer | PASS (`391`, reasoning split off) |
+| image | PASS on red / blue / green solid-colour controls (synthetic) |
+| tool call | PASS — `get_weather({"city": "Haifa"})`, `finish_reason: tool_calls` |
+| speed | 69.3 tok/s wall on one short request (includes prefill and reasoning) |
+| long context | PASS — 231,935-token needle retrieved in 200.3 s |
+
+## Measured — as the `worker` lane, 2026-10-25
+
+PART 2 of the same transcript: the Spark's lobes `vllm-worker` lane (cold
+compile, the same 350,464-token pinned pool), reached as `model=worker` from
+the spark, thor and spark2 gateways — the Thor and spark2 by mesh forward,
+with `X-Lobes-Mesh-Member: spark`. Through the Thor's forward the red/blue
+image controls and a `get_weather` tool call passed.
+
+## Not proven
+
+- The packaged template serving it (the Spark uses an override command), a
+  deployment shape that hosts it, and `lobes switch` to it — switch prints
+  the gear's serving note and, on `--apply`, writes `.env` without
+  restarting.
+- Any card other than the Spark GB10.
+- Concurrency above 1, sustained decode under load, and DFlash acceptance
+  rates.
+- Real-image and video quality (the image probes are synthetic).

{lobes_cli-0.91.0 → lobes_cli-0.92.0}/docs/nvidia-qwen3.6-35b-a3b-nvfp4.md

@@ -1,5 +1,13 @@
 # `nvidia/Qwen3.6-35B-A3B-NVFP4` — the Thor `worker` lane
 
+> **SUPERSEDED as the `worker` role_hint (2026-10-26, operator decision).**
+> The catalog's `worker` moved to `nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB`,
+> hosted on the DGX Spark (see
+> [`nemotron-3.5-bolt-74b-a9b-24gb.md`](nemotron-3.5-bolt-74b-a9b-24gb.md)).
+> This checkpoint is a kept `candidate` (cite-don't-delete); the
+> `thor-worker` shape still renders it. Read "the `worker` lane" below as
+> history.
+
 **Status: VALIDATED live on the physical Jetson AGX Thor, 2026-09-10, on the
 nightly `0.26.1rc1.dev942`; re-accepted on vLLM 0.31.0 (2026-10-17) for the
 items listed in [Engine relabel](#engine-relabel-310-t15) only** — see

{lobes_cli-0.91.0 → lobes_cli-0.92.0}/lobes/catalog.py

@@ -159,6 +159,12 @@ class SupportedModel:
     modalities: tuple[str, ...] = ()
     mrl_dims: tuple[int, ...] = ()
     normalization: str = ""
+    # A vLLM gear the STOCK lane cannot boot, in one sentence naming what it
+    # needs instead (an image, a lane). "" (the default) means the stock
+    # template serves it. When set, `lobes switch` prints this note INSTEAD of
+    # its stock-template edit notices (which would be wrong advice) and its
+    # `--apply` writes `.env` without restarting, exactly like a non-vLLM gear.
+    serving_note: str = ""
 
 
 SUPPORTED_MODELS: tuple[SupportedModel, ...] = (
@@ -1056,9 +1062,78 @@ SUPPORTED_MODELS: tuple[SupportedModel, ...] = (
         # declared, UNMEASURED, and evaluated separately by plan task t2.
         task="generate",
     ),
+    SupportedModel(
+        id="nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB",
+        # The `worker` gear (operator decision, 2026-10-26), replacing
+        # nvidia/Qwen3.6-35B-A3B-NVFP4 below (demoted to a kept candidate). It
+        # fits the worker contract unchanged — a fast multimodal (image+video)
+        # coder/doer with tool calls — so roles.py's responsibilities do not
+        # move. It is hosted on the DGX Spark (sm_12x is the only card its IQ2
+        # experts run on); other members reach `worker` through the mesh.
+        #
+        # Verified against the checkpoint's ACTUAL config files at the pinned
+        # revision below, not card prose:
+        #   config.json
+        #     - architectures ["NemotronHPuzzleForCausalLM"] inside a
+        #       `nemotron_h_omni` wrapper; llm_config.max_position_embeddings
+        #       = 262144 (native); a RADIO vision_config, image and video
+        #       context tokens, and no sound_config — text+image+video.
+        #     - quantization_config.quant_method = "modelopt",
+        #       quant_algo = "MIXED_PRECISION": 27,728 IQ2_XXS layers (the
+        #       routed experts, ggml packing, ~2.06 bits), 113 Q8_0 and 80
+        #       W4A16_NVFP4 (the mixer in/out projections). That is what
+        #       makes 74B fit in ~24 GiB.
+        #   README (vLLM serve command): --moe-backend b12x and
+        #   --linear-backend b12x, the `ultra_v3` reasoning parser shipped as
+        #   a plugin file in the checkpoint, --tool-call-parser qwen3_coder,
+        #   and a DFlash draft from the separate -LiLiCorr repo.
+        #
+        # ENGINE: the IQ2 expert path is gated to sm_12x (the Spark's GB10) —
+        # vLLM refuses it on the Orin (sm_87) and nothing has been tried on
+        # the Thor. It needs the card's nightly plus the `b12x` package,
+        # neither of which the fleet's pinned vLLM image carries, so
+        # `lobes switch` to this id will NOT boot on the stock template.
+        role_hint="worker",
+        shape="MoE (~9B active of 74B) hybrid Mamba-2/attention + RADIO ViT (text+image+video)",
+        context=_CONTEXT_256K_NATIVE,
+        native_max_model_len=262144,
+        tool_parser="qwen3_coder",
+        # quant_method is "modelopt"; vLLM resolves this MIXED_PRECISION config
+        # to "modelopt_mixed", which is what the Spark lane passes and what
+        # was measured to boot.
+        quantization="modelopt_mixed",
+        # Spark GB10 2026-10-25, docs/evidence/2026-10-26-accept-nemotron-bolt-spark.txt:
+        # PART 1 a standalone container (text/image/tools, 231,935-token needle
+        # PASS); PART 2 the Spark's lobes `vllm-worker` lane, reached as
+        # `model=worker` from the spark, thor and spark2 gateways.
+        status="load-tested",
+        doc="nemotron-3.5-bolt-74b-a9b-24gb.md",
+        moe_backend="b12x",
+        speculative_config=(
+            '{"method": "dflash",'
+            ' "model": "nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-LiLiCorr",'
+            ' "revision": "449d24140160b8e51871bfc0f0a7338abf640aa1",'
+            ' "num_speculative_tokens": 7}'
+        ),
+        hf_revision="c6cc2e8ffa353403213d6b9b18aaf8eb684d705a",
+        serving_note=(
+            "needs the card's vLLM nightly plus b12x and an sm_12x card (the"
+            " DGX Spark), with flags the stock lane does not pass — serve it on"
+            " the vllm-worker lane"
+        ),
+        task="generate",
+    ),
     SupportedModel(
         id="nvidia/Qwen3.6-35B-A3B-NVFP4",
-        # The NEW `worker` gear (issue #244, t1), taking the seat back from
+        # DEMOTED from role_hint="worker" to a kept candidate (2026-10-26,
+        # operator decision): the `worker` seat moves to
+        # nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB above, hosted on the DGX
+        # Spark. Kept, not deleted (cite-don't-delete): every 2026-09-10..
+        # 2026-10-26 worker-lane evidence transcript (the Thor, and the Spark's
+        # proxy to it) was measured against this checkpoint, and the
+        # `thor-worker` shape still renders it. Only role_hint moved.
+        #
+        # (History) The `worker` gear from issue #244, t1, taking the seat back from
         # nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 above (demoted to
         # role_hint="candidate", kept — cite-don't-delete). A DISTINCT entry
         # from BOTH other Qwen3.6-35B-A3B-NVFP4 catalog ids: unsloth/'s (the
@@ -1124,7 +1199,7 @@ SUPPORTED_MODELS: tuple[SupportedModel, ...] = (
         # is left empty (auto-select) rather than carrying an untested guess
         # forward — the outgoing unsloth 35B-A3B worker's own hard-won sm_110
         # lesson was that every FORCED NVFP4 MoE backend was refused there.
-        role_hint="worker",
+        role_hint="candidate",
         shape=(
             "hybrid Mamba/linear-attn MoE (~3B active) + ViT "
             "(text+image+video, self-hosted MTP draft)"
@@ -1133,7 +1208,9 @@ SUPPORTED_MODELS: tuple[SupportedModel, ...] = (
         native_max_model_len=262144,
         tool_parser="qwen3_coder",
         quantization="modelopt",
-        status="configured",  # declared 2026-09-10 (issue #244); no live boot yet
+        # VALIDATED live on the Jetson AGX Thor 2026-09-10 as the worker:
+        # docs/evidence/2026-09-10-accept-nvidia-35b-a3b-thor.txt.
+        status="load-tested",
         doc="nvidia-qwen3.6-35b-a3b-nvfp4.md",
         moe_backend="",
         # Self-hosted draft (no external "model"/"draft_model_id" key), same

{lobes_cli-0.91.0 → lobes_cli-0.92.0}/lobes/cli/_commands/switch.py

@@ -270,6 +270,25 @@ def _engine_notice(model) -> str | None:
     )
 
 
+def _serving_note_notice(model) -> str | None:
+    """A vLLM gear the STOCK lane cannot boot: say what it needs instead.
+
+    The stock-template notices (remove the MTP primary's command items, add
+    ``--moe-backend``) assume the stock image and the stock flags can serve
+    the gear. For a gear whose catalog entry carries a ``serving_note`` that
+    is false, and following them would mislead — e.g. removing
+    ``--trust-remote-code`` from a checkpoint that needs it. Like
+    :func:`_engine_notice`, this is a notice, so ``--apply`` writes ``.env``
+    and does not restart the container (:func:`_apply_env_only`).
+    """
+    if not model.serving_note:
+        return None
+    return (
+        f"the stock vLLM lane cannot boot this gear: it {model.serving_note}; "
+        f"see docs/{model.doc}."
+    )
+
+
 def _lock_staleness_notice(deploy_dir) -> str | None:
     """A committed ``deployment.lock.toml`` is present — say so before writing ``.env``.
 
@@ -312,6 +331,10 @@ def _serve_notices(model_id: str, args: argparse.Namespace | None = None) -> lis
         # "remove the MTP command items" / "add --moe-backend" are edits to a
         # compose file that does not serve this gear at all.
         return [notice for notice in [_engine_notice(model)] if notice]
+    if model is not None and model.serving_note:
+        # Same short-circuit, for a vLLM gear the stock lane cannot boot: the
+        # stock-template edit notices would be wrong advice for it.
+        return [notice for notice in [_serving_note_notice(model)] if notice]
     if model is not None:
         candidates += [
             _mtp_removal_notice(model),

{lobes_cli-0.91.0 → lobes_cli-0.92.0}/lobes/profiles/builtin_shapes/thor-worker.toml

@@ -1,3 +1,10 @@
+# NOT THE DEPLOYED STATE (2026-10-26): the mesh's `worker` is now Nemotron 3.5
+# Bolt, hosted on the DGX Spark, and the Thor stopped hosting worker by hand
+# (.env: WORKER_FEASIBLE=false, worker dropped from COMPOSE_PROFILES). This
+# shape still hosts nvidia/Qwen3.6-35B-A3B-NVFP4 as `worker` and is kept as the
+# rollback shape (cite-don't-delete). Re-rendering it on the Thor re-hosts that
+# Qwen worker beside the Spark's Bolt: the mesh then sees two disagreeing
+# `worker` members and exposes them only as worker-thor / worker-spark.
 # The Jetson AGX Thor WORKER variant of a mesh-brain deployment: hosts the
 # opt-in `worker` lobe (nvidia/Qwen3.6-35B-A3B-NVFP4 — the eighth Colleague
 # role, the fast MULTIMODAL ground-work DOER) + the two pooling gears + the

{lobes_cli-0.91.0 → lobes_cli-0.92.0}/lobes/runtime/_parser.py

@@ -123,6 +123,10 @@ _RULES: list[tuple[tuple[str, ...], str]] = [
     # retired NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4, a different family with
     # no such card claim).
     (("nemotron-3.5-lightning", "nemotron_3.5_lightning"), "qwen3_coder"),
+    # NVIDIA Nemotron 3.5 Bolt (the `worker` gear since 2026-10-26): its card's
+    # vLLM serve command passes `--tool-call-parser qwen3_coder`, the same
+    # cross-family claim as Lightning above. Scoped to the Bolt spelling only.
+    (("nemotron-3.5-bolt", "nemotron_3.5_bolt"), "qwen3_coder"),
 ]
 
 

{lobes_cli-0.91.0 → lobes_cli-0.92.0}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "lobes-cli"
-version = "0.91.0"
+version = "0.92.0"
 description = "lobes — run, assess, and switch the local vLLM model."
 readme = "README.md"
 license = "Apache-2.0"

{lobes_cli-0.91.0 → lobes_cli-0.92.0}/tests/goldens/no-pool-gateway.json

@@ -267,10 +267,10 @@
           "security_decision"
         ],
         "loaded": false,
-        "model": "nvidia/Qwen3.6-35B-A3B-NVFP4",
+        "model": "nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB",
         "mtp": true,
         "path": "/v1/chat/completions",
-        "quant": "modelopt",
+        "quant": "modelopt_mixed",
         "ready": false,
         "responsibilities": [
           "execution",

{lobes_cli-0.91.0 → lobes_cli-0.92.0}/tests/goldens/switch-plans.txt

@@ -298,6 +298,22 @@
   msg: tool-call parser (auto-selected): qwen3_coder
   msg: quantization (from catalog): modelopt
   note: non-MTP model — the template ships the MTP default primary's flags; REMOVE these `command:` list items by hand to serve this model (see docs/qwen3.6-27b-text-nvfp4-mtp.md):\n      - '--speculative-config={"method": "mtp", "num_speculative_tokens": 2}'\n      - --trust-remote-code
+### nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB
+  VLLM_MODEL=nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB
+  VLLM_SERVED_NAME=nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB
+  VLLM_PORT=8000
+  VLLM_PURPOSE=balanced
+  VLLM_MACHINE=spark
+  VLLM_MAX_MODEL_LEN=262144
+  VLLM_GPU_MEM_UTIL=0.6
+  VLLM_ATTENTION_BACKEND=flashinfer
+  VLLM_MAX_NUM_SEQS=4
+  VLLM_MAX_NUM_BATCHED_TOKENS=8192
+  VLLM_TOOL_CALL_PARSER=qwen3_coder
+  VLLM_QUANTIZATION=modelopt_mixed
+  msg: tool-call parser (auto-selected): qwen3_coder
+  msg: quantization (from catalog): modelopt_mixed
+  note: the stock vLLM lane cannot boot this gear: it needs the card's vLLM nightly plus b12x and an sm_12x card (the DGX Spark), with flags the stock lane does not pass — serve it on the vllm-worker lane; see docs/nemotron-3.5-bolt-74b-a9b-24gb.md.
 ### nvidia/Qwen3.6-35B-A3B-NVFP4
   VLLM_MODEL=nvidia/Qwen3.6-35B-A3B-NVFP4
   VLLM_SERVED_NAME=nvidia/Qwen3.6-35B-A3B-NVFP4

{lobes_cli-0.91.0 → lobes_cli-0.92.0}/tests/test_associate_role.py

@@ -61,7 +61,9 @@ from lobes.profiles.shapes import DEFAULT_HOSTED_ROLES, OPT_IN_CORE_ROLES
 # catalog resolutions are independent as of t2: changing which checkpoint
 # carries role_hint="worker" cannot move associate's default.
 _LIGHTNING_ID = "nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4"
-_WORKER_ID = "nvidia/Qwen3.6-35B-A3B-NVFP4"
+# 2026-10-26: worker moved to Nemotron 3.5 Bolt (operator); the nvidia Qwen
+# export is a kept candidate.
+_WORKER_ID = "nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB"
 
 #: The nine role names that existed BEFORE associate — the no-regression set.
 _PRE_EXISTING_ROLES: tuple[str, ...] = (

{lobes_cli-0.91.0 → lobes_cli-0.92.0}/tests/test_catalog.py

@@ -31,7 +31,10 @@ from lobes.runtime._parser import infer_parser
 
 _DOCS = Path(__file__).resolve().parents[1] / "docs"
 _TEMPLATES = Path(__file__).resolve().parents[1] / "lobes" / "templates"
-_WORKER_ID = "nvidia/Qwen3.6-35B-A3B-NVFP4"
+# 2026-10-26: worker moved to Nemotron 3.5 Bolt (operator decision).
+_WORKER_ID = "nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB"
+# The previous worker (issue #244, t1), now a kept candidate.
+_NVIDIA_QWEN_WORKER_ID = "nvidia/Qwen3.6-35B-A3B-NVFP4"
 # The DEMOTED former worker gear (issue #244, t1) — kept in the catalog as a
 # candidate (cite-don't-delete), no longer the role_hint="worker" holder.
 _NEMOTRON_ID = "nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4"
@@ -173,7 +176,7 @@ def test_moe_backend_aligns_with_shape() -> None:
 # module, no external draft repo, and no "-MTP" marker in its upstream id.
 _SELF_HOSTED_MTP_WITHOUT_ID_MARKER = frozenset(
     {
-        _WORKER_ID,
+        _NVIDIA_QWEN_WORKER_ID,
         _QWEN_WORKER_ID,
         "unsloth/Qwen3.6-27B-NVFP4",
         "unsloth/Qwen3.8-27B-NVFP4",
@@ -767,10 +770,16 @@ def test_worker_gear_exists_with_correct_fields() -> None:
     worker = next((m for m in SUPPORTED_MODELS if m.id == _WORKER_ID), None)
     assert worker is not None, f"{_WORKER_ID} not found in catalog"
     assert worker.role_hint == "worker"
-    assert worker.native_max_model_len == 262144  # config.json max_position_embeddings
-    assert worker.quantization == "modelopt"  # hf_quant_config.json producer/quant_algo
-    assert worker.tool_parser == "qwen3_coder"  # infer_parser("...qwen3.6...")
-    assert worker.status == "configured"  # declared 2026-09-10, no live boot yet
+    assert worker.native_max_model_len == 262144  # llm_config.max_position_embeddings
+    # quant_method is modelopt; vLLM resolves MIXED_PRECISION to modelopt_mixed,
+    # the value the Spark lane passes and was measured to boot with.
+    assert worker.quantization == "modelopt_mixed"
+    # The stock lane cannot boot it, so `lobes switch` must say so.
+    assert "b12x" in worker.serving_note
+    assert worker.tool_parser == "qwen3_coder"  # the card's vLLM command
+    # Spark GB10 2026-10-25, docs/evidence/2026-10-26-accept-nemotron-bolt-spark.txt
+    assert worker.status == "load-tested"
+    assert worker.hf_revision == "c6cc2e8ffa353403213d6b9b18aaf8eb684d705a"
     assert worker.task == "generate"
     assert worker.dimension == 0
     assert worker.hf_overrides == ""
@@ -778,34 +787,43 @@ def test_worker_gear_exists_with_correct_fields() -> None:
 
 
 def test_worker_gear_is_multimodal() -> None:
-    # config.json carries a vision_config (image+video intake), unlike the
-    # demoted Nemotron worker, which was text-only.
+    # config.json carries a RADIO vision_config (image+video intake), which the
+    # worker contract's image_understanding/video_understanding rely on.
     worker = next(m for m in SUPPORTED_MODELS if m.id == _WORKER_ID)
     shape = worker.shape.lower()
     assert "vit" in shape or "vision" in shape or "image" in shape or "video" in shape
     assert "text-only" not in shape
 
 
-def test_worker_gear_has_self_hosted_mtp_speculative_config() -> None:
-    # config.json's text_config.mtp_num_hidden_layers=1 plus the ignore/
-    # exclude_modules ["mtp.layers.0*", "mtp*"] pattern confirm a physically
-    # present, self-hosted MTP draft module — declared here as a MEASUREMENT
-    # of the checkpoint's own config, not the card's separate MTP/DSpark prose.
+def test_worker_gear_drafts_with_the_external_lilicorr_dflash_head() -> None:
+    # The card's speculative decoding is DFlash from a SEPARATE draft repo, so
+    # the config names that repo (and pins its revision) instead of `mtp`.
     worker = next(m for m in SUPPORTED_MODELS if m.id == _WORKER_ID)
     cfg = json.loads(worker.speculative_config)
-    assert cfg["method"] == "mtp"
-    assert "model" not in cfg
-    assert "draft_model_id" not in cfg
+    assert cfg["method"] == "dflash"
+    assert cfg["model"] == "nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-LiLiCorr"
+    assert cfg["revision"] == "449d24140160b8e51871bfc0f0a7338abf640aa1"
+    assert cfg["num_speculative_tokens"] == 7
 
 
-def test_worker_gear_is_moe_with_auto_selected_backend() -> None:
+def test_worker_gear_is_moe_with_the_cards_b12x_backend() -> None:
     worker = next(m for m in SUPPORTED_MODELS if m.id == _WORKER_ID)
-    assert worker.shape.lower().startswith("hybrid")
-    assert "moe" in worker.shape.lower()
-    # NO forced moe_backend: mirrors the demoted unsloth worker's own
-    # hard-won sm_110 lesson (forced NVFP4 MoE backends were refused there)
-    # rather than carrying an untested guess forward.
-    assert worker.moe_backend == ""
+    assert worker.shape.lower().startswith("moe")
+    # The card's own --moe-backend; the IQ2 expert path is sm_12x-only.
+    assert worker.moe_backend == "b12x"
+
+
+def test_previous_nvidia_qwen_worker_is_demoted_to_candidate_and_kept() -> None:
+    # nvidia/Qwen3.6-35B-A3B-NVFP4 (issue #244, t1) is KEPT (cite-don't-delete):
+    # the thor-worker shape still renders it and the 2026-09-10..2026-10-26
+    # worker evidence was measured against it. Only role_hint moved.
+    outgoing = next((m for m in SUPPORTED_MODELS if m.id == _NVIDIA_QWEN_WORKER_ID), None)
+    assert outgoing is not None, f"{_NVIDIA_QWEN_WORKER_ID}: expected to remain in the catalog"
+    assert outgoing.role_hint == "candidate"
+    assert outgoing.native_max_model_len == 262144
+    assert outgoing.quantization == "modelopt"
+    assert outgoing.tool_parser == "qwen3_coder"
+    assert json.loads(outgoing.speculative_config)["method"] == "mtp"
 
 
 def test_worker_gear_tool_parser_matches_infer_parser() -> None: