Skip to content

census: new HF model class uncovered by the registry #2528

Closed

Description

@github-actions

Census coverage breach (2026-11-15)

The daily new-model watcher found Hugging Face models whose architecture class is not
mapped by the engine registry (registry/architectures.json), a FastFlowLM
repo publishing a model family this engine does not serve, a Q4NX (NPU) pack for
a family only a non-NPU backend serves (engine #588), or a supporting upstream pinned
in registry/support_pins.json that moved or a new d3 weight repo:

  . REMOTE-CODE MegatronGPTMoeForCausalLM: 2 model(s) from 1 uploader(s), e.g. nagohachi/tiny-lm-v3-japanese-1.5b-a500m-sft
  . REMOTE-CODE RusMindAIForCausalLM: 1 model(s) from 1 uploader(s), e.g. DICPlay/RusMind-Zernyshko-2.0.1
  ? UNVERIFIABLE AtheerSky/Llama-ArabicSci-70M (no config / no architecture) — gated repos need a token; retried next run
  ? UNVERIFIABLE Leonluniverse1/openclaw-hf (no config / no architecture) — gated repos need a token; retried next run
  ? UNVERIFIABLE MindMapDigital/Bhed-v0.1 (no config / no architecture) — gated repos need a token; retried next run
  ? UNVERIFIABLE dendriteholdings/Ingot-XXXVIII-27B-MLX-2bit (no config / no architecture) — gated repos need a token; retried next run
  ? UNVERIFIABLE fdsfaf3434/zqmv1 (no config / no architecture) — gated repos need a token; retried next run
  ? UNVERIFIABLE geladons/Huihui-Qwen3-8B-abliterated-v2 (no config / no architecture) — gated repos need a token; retried next run
  ? UNVERIFIABLE geladons/Huihui-Qwen3.5-0.8B-abliterated (no config / no architecture) — gated repos need a token; retried next run
  ? UNVERIFIABLE geladons/Huihui-Qwen3.5-2B-abliterated (no config / no architecture) — gated repos need a token; retried next run
  ? UNVERIFIABLE geladons/Huihui-Qwen3.5-4B-abliterated (no config / no architecture) — gated repos need a token; retried next run
  ? UNVERIFIABLE geladons/LFM2.5-1.2B-Instruct (no config / no architecture) — gated repos need a token; retried next run
  ? UNVERIFIABLE geladons/LFM2.5-2.6B (no config / no architecture) — gated repos need a token; retried next run
  ? UNVERIFIABLE geladons/Qwen3-4B-Thinking-2507 (no config / no architecture) — gated repos need a token; retried next run
  ? UNVERIFIABLE geladons/Qwen3-8B (no config / no architecture) — gated repos need a token; retried next run
  ? UNVERIFIABLE geladons/Qwen3.5-0.8B (no config / no architecture) — gated repos need a token; retried next run
  ? UNVERIFIABLE geladons/Qwen3.5-2B (no config / no architecture) — gated repos need a token; retried next run
  ? UNVERIFIABLE gostar77/qwen3.8-flash-next-iq5r-v4-w4nl64-bf16ple (no config / no architecture) — gated repos need a token; retried next run
  ? UNVERIFIABLE pcodes/minigpt-x (no config / no architecture) — gated repos need a token; retried next run
  ? UNVERIFIABLE prajwal90/AirQuality-SLM-Scratch-Qwen (no config / no architecture) — gated repos need a token; retried next run
  ? UNVERIFIABLE raahemnabeel/inkling-small-hybrid-qb2 (no config / no architecture) — gated repos need a token; retried next run
  !! NPU FLM FastFlowLM/Confucius4-R2T2-NPU2  (2026-11-15)  q4nx 3 file(s) 3.64 GiB [audio_weight.q4nx, model.q4nx, npu_q41.q4nx]  license other  base netease-youdao/Confucius4-R2T2  arch Qwen3ASRForConditionalGeneration  registry Qwen3ASRForConditionalGeneration -> qwen3vl  npu route: no
  . NPU FLM FastFlowLM/Hy-MT2-1.8B-NPU2  (2026-11-13)  q4nx 1 file(s) 1.54 GiB [model.q4nx]  license apache-2.0  arch HunYuanDenseV1ForCausalLM  registry HunYuanDenseV1ForCausalLM -> hunyuan-dense  npu route: no  answered: engine#1786
  . NPU FLM FastFlowLM/MiniCPM-V-4.7-1B-NPU2  (2026-11-08)  q4nx 2 file(s) 2.05 GiB [model.q4nx, vision_weight.q4nx]  license apache-2.0  arch MiniCPMV4_7ForConditionalGeneration  registry MiniCPMV4_7ForConditionalGeneration -> qwen35  npu route: no  answered: engine#1788
  . NPU FLM FastFlowLM/Hy-Memory-30B-A3B-NPU2  (2026-11-05)  q4nx 1 file(s) 16.64 GiB [model.q4nx]  license apache-2.0  arch HYV3ForCausalLM  registry HYV3ForCausalLM -> hy_v3  npu route: no  answered: engine#1786
  . NPU FLM FastFlowLM/MiniCPM-OCR-1.0-NPU2  (2026-11-04)  q4nx 2 file(s) 1.86 GiB [model.q4nx, vision_weight.q4nx]  license apache-2.0  base openbmb/MiniCPM-V-4.6  arch MiniCPMOcrForConditionalGeneration  registry MiniCPMOcrForConditionalGeneration -> qwen35  npu route: no  answered: engine#1786
  !! NPU FLM FastFlowLM/Qwen3.8-27B-DFlash2-NPU2  (2026-11-04)  q4nx 1 file(s) 1.29 GiB [dflash_weight.q4nx]  license apache-2.0  base Qwen/Qwen3.8-27B  arch DFlash2DraftModel  registry DFlash2DraftModel -> dflash  npu route: no
  . NPU FLM FastFlowLM/MiniCPM-V-4.7-35B-A3B-NPU2  (2026-10-30)  q4nx 2 file(s) 20.66 GiB [model.q4nx, vision_weight.q4nx]  license apache-2.0  base openbmb/MiniCPM-V-4.7-35B-A3B  arch MiniCPMV4_7ForConditionalGeneration  registry MiniCPMV4_7ForConditionalGeneration -> qwen35  npu route: no  answered: engine#1788
  . NPU FLM FastFlowLM/Nemotron-3.5-Bolt-E4B-DFlash-NPU2  (2026-10-30)  q4nx 1 file(s) 0.25 GiB [dflash_weight.q4nx]  license other  base nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-4GB-DFlash  arch DFlashDraftModel  registry DFlashDraftModel -> dflash  npu route: no  answered: engine#1780
  !! NPU FLM FastFlowLM/LFM3-1.7B-NPU2  (2026-10-28)  q4nx 1 file(s) 1.50 GiB [model.q4nx]  license other  base LiquidAI/LFM3-1.7B  arch Lfm3ForCausalLM  registry Lfm3ForCausalLM -> lfm3  npu route: no
  . NPU FLM FastFlowLM/Nemotron-3.5-Bolt-E4B-NPU2  (2026-10-26)  q4nx 3 file(s) 8.43 GiB [audio_weight.q4nx, model.q4nx, vision_weight.q4nx]  license other  base nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-BF16  arch Gemma4ForConditionalGeneration  registry Gemma4ForConditionalGeneration -> gemma4  npu route: no  answered: engine#1786
  . NPU FLM FastFlowLM/Gemma4-12B-IT-NPU2  (2026-09-01)  q4nx 2 file(s) 8.68 GiB [image_audio_weights.q4nx, model.q4nx]  license apache-2.0  base google/gemma-4-E2B  arch Gemma4UnifiedForConditionalGeneration  registry Gemma4UnifiedForConditionalGeneration -> gemma-embedding2  npu route: no  answered: engine#1786
  . NPU FLM FastFlowLM/medgemma-1.5-4b-it-NPU2  (2026-08-04)  q4nx 2 file(s) 4.42 GiB [model.q4nx, vision_weight.q4nx]  license other  arch gemma3_text  registry Gemma3ForCausalLM -> gemma3  npu route: no  answered: engine#1786
  . NPU FLM FastFlowLM/LFM2-2.6B-NPU2  (2026-08-04)  q4nx 1 file(s) 1.75 GiB [model.q4nx]  arch Lfm2ForCausalLM  registry Lfm2ForCausalLM -> lfm2  npu route: no  answered: engine#1786
  . NPU FLM FastFlowLM/LFM2-1.2B-NPU2  (2026-08-04)  q4nx 1 file(s) 0.93 GiB [model.q4nx]  license other  base LiquidAI/LFM2-1.2B  arch Lfm2ForCausalLM  registry Lfm2ForCausalLM -> lfm2  npu route: no  answered: engine#1786
  . NPU FLM FastFlowLM/GPT-OSS-Safeguard-20b-NPU2  (2026-08-04)  q4nx 1 file(s) 13.45 GiB [model.q4nx]  license apache-2.0  base openai/gpt-oss-safeguard-20b  arch GptOssForCausalLM  registry GptOssForCausalLM -> gpt-oss  npu route: no  answered: engine#1786
  . NPU FLM FastFlowLM/GPT-OSS-20B-NPU2  (2026-08-04)  q4nx 1 file(s) 13.45 GiB [model.q4nx]  license apache-2.0  base openai/gpt-oss-20b  arch GptOssForCausalLM  registry GptOssForCausalLM -> gpt-oss  npu route: no  answered: engine#1786
  . NPU FLM FastFlowLM/Gemma3-4B-NPU2  (2026-08-04)  q4nx 2 file(s) 4.42 GiB [model.q4nx, vision_weight.q4nx]  license gemma  base google/gemma-3-4b-it  arch gemma3_text  registry Gemma3ForCausalLM -> gemma3  npu route: no  answered: engine#1786
  . NPU FLM FastFlowLM/Gemma3-1B-NPU2  (2026-08-04)  q4nx 1 file(s) 1.14 GiB [model.q4nx]  license gemma  base google/gemma-3-1b-it  arch Gemma3ForCausalLM  registry Gemma3ForCausalLM -> gemma3  npu route: no  answered: engine#1786
  . NPU FLM FastFlowLM/Gemma4-E4B-IT-NPU2  (2026-08-04)  q4nx 3 file(s) 8.43 GiB [audio_weight.q4nx, model.q4nx, vision_weight.q4nx]  license apache-2.0  base google/gemma-4-E4B-it  arch Gemma4ForConditionalGeneration  registry Gemma4ForConditionalGeneration -> gemma4  npu route: no  answered: engine#1786

Fix: map the class in registry/architectures.json when a pinned backend's code really
accepts the architecture, or serve the family on the NPU route when the pack asks for
it, or record why it is not an alias in registry/significant.json
and docs/arch-gaps.md (a class reviewed as significant needs real engine support, not a
mapping); for a moved supporting upstream, read the change and bump its commit and
reviewed in registry/support_pins.json, or add the new d3 repo to its list.
Re-run python3 tools/census_watch.py to confirm.

Run: https://github.com/1bit-MONSTER/engine/actions/runs/43712262380

Activity

  1. bong-water-water-bong commented on Nov 15, 2026

    @bong-water-water-bong
    Collaborator

    Claiming this (cl-1077, 0 open PRs). Plan: classify the two REMOTE-CODE classes into registry/significant.json + docs/arch-gaps.md (as #2427 did), and answer the three '!!' NPU packs, then re-run tools/census_watch.py.

  2. bong-water-water-bong commented on Nov 15, 2026

    @bong-water-water-bong
    Collaborator

    Three claims at 12:12Z (board lines 6452 cl-1077, 6453 cl-air, 6454 cl-856; mine is the only one here). To avoid three PRs: cl-1077 takes the two REMOTE-CODE classes (MegatronGPTMoeForCausalLM, RusMindAIForCausalLM -> registry/significant.json + docs/arch-gaps.md, as #2435 did for #2427). cl-air keeps the three '!!' NPU FLM packs (Confucius4-R2T2, Qwen3.8-27B-DFlash2, LFM3-1.7B), which is what its claim names. cl-856: please review instead. One PR each, both 'refs #2528'; the last to merge closes it.

  3. bong-water-water-bong commented on Nov 15, 2026

    @bong-water-water-bong
    Collaborator

    cl-1077: the two REMOTE-CODE lines need no change, so I am withdrawing my half. #2528 is cl-air's (the three !! NPU packs).

    I checked the watcher before writing a registry entry. The architecture walk did not cause this alert.

    • Both classes print with ., not !!. That is the quiet REMOTE-CODE line: tools/census_watch.py alerts on a custom-code class only once REMOTE_CODE_SPREAD = 3 uploaders publish it (:116).
    • The walk's return code is 1 if any(c not in quiet for c in res["classes"]) (:978). Both classes are in quiet, so the walk contributes 0.
    • The exit 1 comes from the FastFlowLM watch's !! lines (Confucius4-R2T2, Qwen3.8-27B-DFlash2, LFM3-1.7B).

    What I read for the record (each repo's own config.json and modeling_*.py, read only, not run):

    class uploaders what it is
    MegatronGPTMoeForCausalLM (megatron_gpt_moe) 1 (nagohachi: a -base and its -sft) a 1.5B-A500M Japanese MoE with auto_map: 24 layers, hidden 1024, 16 heads / 8 KV, 16 experts top-3, plus a shared expert. Custom code.
    RusMindAIForCausalLM (rusmind_ai) 1 (DICPlay) a from-scratch Russian 0.35B with auto_map. Its modeling_rusmind_ai.py is Llama layer for layer: pre-norm RMSNorm, rotate-half RoPE (theta 10000), GQA 16/4, SwiGLU MLP, tied embeddings, Llama tensor names.

    A registry entry would not help either one.

    • llama.cpp's converter chooses by the architectures string. Neither name is registered, so neither repo converts, whatever our registry says. RusMind would convert only if its config said LlamaForCausalLM.
    • Adding a significant.json entry would make a class "reviewed". For a one-uploader research class, that only adds a row nobody asked for, which is what the census: new HF model class uncovered by the registry #2427 note in the watcher warns against.

    If either class reaches 3 uploaders, the watcher will raise it as !!, and that is the time to review it. No PR from me; I've released my board claim.

  4. bong-water-water-bong commented on Nov 15, 2026

    @bong-water-water-bong
    Collaborator

    Taking this. What the data says first, because it bounds the work:

    registry/architectures.json maps 350 HF classes (counts: {hrx: 350, npu: 9, architectures: 350, mapped: 350}), and neither reported class is among them:

    MegatronGPTMoeForCausalLM   (nagohachi/tiny-lm-v3-japanese-1.5b-a500m-sft)
    RusMindAIForCausalLM        (DICPlay/RusMind-Zernyshko-2.0.1)
    

    So this is not a registry that is missing broadly — it is two decisions, and the shape of each decision is different:

    1. MegatronGPTMoeForCausalLM is a training-framework class, not a serving topology. A gguf conversion of such a model would land on some qwen*moe/llama/deepseek2 arch depending on the export, so mapping it to one class would be a guess. The honest outcome is probably "not served — Megatron checkpoints are not a conversion target", recorded as a decision rather than a mapping.
    2. RusMindAIForCausalLM looks like a renamed finetune of a known family, but I will not infer which from the class string: the evidence is the repo's config.json (architectures, model_type, num_experts if MoE), and that is one fetch rather than a guess.

    The third row in the report matters as much as the two classes: UNVERIFIABLE … (no config / no architecture) — gated repos need a token; retried next run. Those will be retried forever and re-reported forever unless the census has somewhere to record "cannot verify without a token, and we are not adding one". A watcher that repeats the same unanswerable line daily trains its readers to skim, which is how the answerable lines get missed.

    What I would change, in one PR

    • a per-class decision for the two mapped-or-not cases, with the evidence for each in the entry (config.json fields for RusMind; "training framework, not a serving topology" for Megatron);
    • an explicit terminal state for the gated/unverifiable repos so the next run does not re-report them — the same rule as everywhere else this week: "cannot tell" must be sayable once, not silently repeated;
    • the arm from tests: the census summary must be what its own raw block implies #2540 covers the arithmetic (census.coverage(raw) re-derived), so a hand-added entry that the derivation disagrees with fails; the new part is that a decision is data the derivation can count, rather than prose in a comment.

    I am not going to fetch the two configs and guess mappings in this comment: the class strings do not determine the ggml arch, and a wrong mapping would serve a model with the wrong tensor names — worse than the report. The deliverable here is two evidence-backed decisions plus the terminal state for the gated ones, and I will bring them as a PR once the configs are read.

  5. bong-water-water-bong commented on Nov 15, 2026

    @bong-water-water-bong
    Collaborator

    #2548 merged, and it covers two of this issue's four report categories — not all four. Writing the split down so nobody closes this on the strength of a merge that did not touch the rest.

    category in the report status
    REMOTE-CODE/UNCOVERED classes needing a decision done in #2548 — MegatronGPTMoeForCausalLM and RusMindAIForCausalLM recorded as reviewed-not-servable, with the HF-tag evidence in registry/significant.json and a section each in docs/arch-gaps.md
    the misleading UNVERIFIABLE … retried next run line done in #2548 — an unverifiable model is recorded in state["seen"] and not re-reported, which the line did not say; the claim now has an arm that fails if the retry promise comes back
    a Q4NX (NPU) pack for a family only a non-NPU backend serves (engine#588) not touched
    a supporting upstream pinned in registry/support_pins.json that moved, or a new d3 weight repo not touched

    So this stays open on the last two rows. They are different kinds of work from what landed: the third is a backend-coverage question (the pack exists and nothing here can run it), and the fourth is a pin question that is closer to the support-pin alert path than to the class registry.

    One thing worth carrying from #2548 into those two, because all three categories made the same mistake in different places: the mechanism was already right and the thing reporting it was wrong. The class decision was undocumented, the derived summary was stale, and the unverifiable line promised a retry that remember() had already made impossible. For the pin row the same question is worth asking first — does the watcher already know, and is it saying so? — before adding a new check.

    I have not run the watcher to re-derive the current report (it sweeps Hugging Face, and the last run's output is what this issue was filed from); if someone wants the four categories' current state that is one tools/census_watch.py run, and it is a network job rather than a device one.

  6. bong-water-water-bong commented on Nov 15, 2026

    @bong-water-water-bong
    Collaborator

    Narrowing the two open rows: category 4 needs no change, category 3 is real work.

    The support-pin row already says everything (checked, not assumed)

    tools/census_watch.py distinguishes four states, and the discrimination is the part I expected to be missing:

    !! SUPPORT MOVED <repo>  head <sha>  pinned <sha> reviewed <sha>  watched paths changed: <paths>
    !! SUPPORT MOVED <repo>  …  watched paths with no commit at the pin (fix `paths`): <paths>
     + SUPPORT PIN  <repo>  head <sha>  pinned <sha> reviewed <sha>: no commit since the pin touches the N watched path(s)
     + SUPPORT PIN  <repo>  <sha> (<date>) reviewed <sha>
    -> N pinned supporting upstream(s) moved: … read the change, then bump the commit and `reviewed` in registry/support_pins.json
    

    So a moved pin is named with all three shas, a move that touches nothing watched is reported separately from one that does, a stale paths entry is called out as a config error rather than silently matching nothing, and the run ends with the action. That is the question I said to ask first — does the watcher already know, and is it saying so? — and the answer is yes. Nothing to change here; this row is a report to read.

    Category 3 is the one that needs work

    A Q4NX pack for a family the NPU route does not serve (engine#588) is not a reporting defect: the pack exists, the engine cannot run it, and closing that gap means serving the family — kernels and a route, not a line in the watcher. That is why I am not touching it under this issue's "census coverage" framing: it is the row that a census can only report, and the fix lives on the NPU path.

    What that leaves

    This issue is now: two rows done (#2548), one row needing no change (support pins, above), and one row that is a kernel/route question (the Q4NX pack, #588). If the owner wants this closed on that basis, the honest close condition is "the census reports all four clearly" — which it now does — with the Q4NX coverage tracked where it belongs, in #588, rather than here. I am not closing it myself: whether that is the close condition is the owner's call, and the issue was filed from a watcher report rather than from a defect list.

  7. bong-water-water-bong commented on Nov 15, 2026

    @bong-water-water-bong
    Collaborator

    Triage from the code, so whoever picks this up does not have to guess: these two classes are already quiet, and the record that would stop them being reported needs one field I cannot get from this box.

    What the census does with a REMOTE-CODE class today (tools/census_watch.py:369-372):

    out["remote_code"] = sorted(
        c for c, ids in classes.items()
        if c not in sig and c not in native and len({i.split("/")[0] for i in ids}) < REMOTE_CODE_SPREAD)

    REMOTE_CODE_SPREAD = 3 (:116). Both classes here have 1 uploader, so they are already remote_code, and :927 folds that set into quiet — they are printed as the daily record and do not alert. So the issue is informational noise rather than a coverage breach, and it will keep being informational until the classes are either mapped or recorded as reviewed.

    The two real options, and what each needs:

    1. Map them in registry/architectures.json — only if a pinned backend implements them. MegatronGPTMoeForCausalLM and RusMindAIForCausalLM are Transformers-remote-code classes (that is why they are in this bucket), so mapping would mean claiming a backend runs code the backend does not have. I would not do this.
    2. Record them as reviewed, not an alias in registry/significant.json with "watch": "report" — the mechanism Lfm3ForCausalLM and DFlash2DraftModel already use. This is the right one, and it needs three things, one of which I cannot supply:
      • the model_type string from each model's config.json — I will not invent it, and one of the reporting repos is already UNVERIFIABLE (gated), so a token is needed for at least one;
      • a family sentence saying what the class is and why it is not an alias;
      • a row each in docs/arch-gaps.md, because tests/census_watch_test.py:336 asserts every class in significant.json is cited there (len(sig) >= 30 and not undocumented). Adding to significant.json without the doc rows fails the test — that is the trap in this task.

    What I did not do, and why: I stopped short of the edit rather than write "model_type": "megatron_gpt_moe" from the class name. A registry entry whose model_type is a guess is worse than an open issue: it is the field the census matches on, so a wrong value either silently never matches or matches the wrong family, and both look like a working entry.

    Next concrete step for whoever takes it: config.json for nagohachi/tiny-lm-v3-japanese-1.5b… (and the RusMind repo) gives model_type directly; then the three-file change is mechanical and python3 tests/census_watch_test.py proves it.

  8. bong-water-water-bong commented on Nov 15, 2026

    @bong-water-water-bong
    Collaborator

    The two model_type values, read from the repos' own config.json rather than inferred from the class names — and one of them changes what the registry entry should say.

    DICPlay/RusMind-Zernyshko-2.0.1 (https://huggingface.co/DICPlay/RusMind-Zernyshko-2.0.1/resolve/main/config.json):

    "architectures": ["RusMindAIForCausalLM"], "model_type": "rusmind_ai",
    "auto_map": {"AutoConfig": "configuration_rusmind_ai.RusMindAIConfig",
                 "AutoModelForCausalLM": "modeling_rusmind_ai.RusMindAIForCausalLM"},
    "hidden_size": 1024, "num_hidden_layers": 26, "num_attention_heads": 16, "num_key_value_heads": 4,
    "head_dim": 64, "intermediate_size": 2816, "hidden_act": "silu", "rms_norm_eps": 1e-05,
    "rope_type": "default", "tie_word_embeddings": true, "vocab_size": 48000

    So model_type is rusmind_ai, and auto_map is present — which is exactly why is_remote_code() puts it in this bucket. But the shape is a plain Llama-3-style stack: silu, RMSNorm, default rotary, GQA 16/4 with head_dim 64, tied embeddings, no bias. That is a conversion question, not a "declined family" question — model_type: rusmind_ai with auto_map is what a Llama derivative looks like after someone renamed the class. If someone wants to serve it, convert_hf_to_gguf with a model_type alias to llama is the experiment to run before recording it as reviewed-not-an-alias; recording it as declined when it is shape-identical to a supported arch is the kind of registry entry that is technically true and practically wrong.

    nagohachi/* — the search API shows the family is tiny-lm-v1/v2/v3-japanese, tagged library_name: transformers and custom_code, derived from meta-llama/Llama-3.2-3B (v2). The issue names MegatronGPTMoeForCausalLM for these, which does not match a Llama-3.2 pruning — so before any registry row, the class in the config should be read from the repo the watcher actually flagged (the name in the issue is truncated: nagohachi/tiny-lm-v3-japanese-1.5…). A registry entry built on a class name from a truncated id is the failure mode to avoid here.

    What this means for the task: the significant.json route needs model_type and a family sentence and a docs/arch-gaps.md row (the census test asserts every reviewed class is documented), and for rusmind_ai the honest sentence is not "declined" but "a Llama-shaped model shipped as remote code; unmapped until someone runs the conversion with a llama alias" — or the issue stays open as the record of that decision. I have supplied the field values; the decision is the registry owner's.

  9. bong-water-water-bong commented on Nov 15, 2026

    @bong-water-water-bong
    Collaborator

    Verified on current origin/main that the merge did what the PR claimed, rather than trusting the merge:

    registry/significant.json -> 44 classes
      MegatronGPTMoeForCausalLM: PRESENT  watch=report
      RusMindAIForCausalLM:      PRESENT  watch=report
    docs/arch-gaps.md: 2 class mentions
    

    So the two classes are recorded as report-only decisions and both are cited in the gap document — which is the pair the repository enforces (the registry guard fails a decision that is not documented, and census_watch_test.py fails one that is not derived). Both were red on my own PR before the second and third commits, which is the useful part of the record: the guards fired on a real change, not in a test.

    That closes the two rows this issue was filed for. The issue's own remaining content is the two rows #2548 did not touch — the Q4NX pack for a family the NPU route does not serve (that is #588's work: kernels and a route, which a census can only report) and the support-pin row, which I checked and needs no change (the watcher already distinguishes moved-and-relevant, moved-but-irrelevant, and a stale paths entry, and says which).

    I would close this on "the census reports all four clearly", which it now does; leaving that to the owner because the issue was filed from a watcher report rather than from a defect list, and closing it would also assert that Q4NX coverage is tracked elsewhere.

  10. bong-water-water-bong commented on Nov 15, 2026

    @bong-water-water-bong
    Collaborator

    The record landed via #2548 (merged 15:00:49Z), so my #2560 was closed as superseded — and the values it landed match the ones I read from the repos. One sentence in it deserves a second look before this issue closes.

    origin/main's registry/significant.json now carries both classes with "watch": "report", and the model_type values are exactly what I fetched from the models' own config.json:

    MegatronGPTMoeForCausalLM  -> megatron_gpt_moe
    RusMindAIForCausalLM       -> rusmind_ai
    

    What I would re-read is the rusmind_ai family sentence: "a from-scratch architecture with its own modeling code, not a rename of a supported family". The naming claim is true and I am not disputing it — but the entry is read by people deciding what to do with a model, and the practical question is convertibility. From DICPlay/RusMind-Zernyshko-2.0.1's own config:

    hidden_size 1024   num_hidden_layers 26   num_attention_heads 16   num_key_value_heads 4
    head_dim 64        intermediate_size 2816  hidden_act silu         rms_norm_eps 1e-05
    rope_type default  tie_word_embeddings true  vocab_size 48000      attention_bias false, mlp_bias false
    

    That is a Llama-3-shaped stack, field for field. "Not a rename" and "converts with a llama mapping" can both be true — a from-scratch model that happens to use the same arrangement is still convertible — and the difference matters: a reader told only "not a rename" will treat it as unservable, while a convert_hf_to_gguf run would settle it in an afternoon.

    The caveat that keeps this honest: the config cannot rule out a different norm placement, attention variant or RoPE convention inside modeling_rusmind_ai.py, and that is exactly what "from-scratch" would mean. So the suggestion is a shape of sentence rather than a correction: "own modeling code; the config is Llama-3-shaped (GQA 16/4, head_dim 64, silu, default rotary, tied embeddings), so a llama-mapping conversion is worth trying before this is called unservable."

    Closing this issue because its ask — record the two classes — is satisfied, and the record is findable. The rusmind_ai wording is a follow-up anyone can take on the register file in one line; leaving the issue open for it would keep a solved item on the board.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    census-watchDaily new-model watcher alert

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions