From d4baa5da9fb4c39f3e54bdd13e23195d95981e6f Mon Sep 17 00:00:00 2001 From: hrx-compat lane Date: Sun, 15 Nov 2026 10:13:12 -0400 Subject: [PATCH 1/4] census: two classes reviewed as not-servable, with the evidence in the entry (engine#2528) The daily new-model watcher reported two classes as UNCOVERED, which reads as "needs engine support" when the honest answer for both is "not a conversion target", and left them to be re-reported every day. MegatronGPTMoeForCausalLM: a training-framework checkpoint format, not a serving topology. HF tags carry megatron_gpt_moe and custom_code (checked 2026-11-17, nagohachi/tiny-lm-v3-japanese-1.5b-a500m-sft). A gguf conversion lands on whichever arch its exporter targets, so there is no single arch to map it to and aliasing it to qwen*moe / llama / deepseek2 would serve the model with the wrong tensor names. RusMindAIForCausalLM: HF tags carry rusmind_ai, from-scratch and custom_code (checked 2026-11-17, DICPlay/RusMind-Zernyshko-2.0.1). from-scratch is the load-bearing tag: it rules out the "renamed finetune" case in which a mapping would otherwise be plausible, and a class string is not evidence of a tensor layout. Both go into registry/significant.json with watch: report (report only, not a mapping) and a section each in docs/arch-gaps.md, which is where the registry's own `about` says these decisions live. Checked: tools/registry_build.py --check-gaps reports "none of the 44 reviewed significant classes is mapped without a reason", and tests/census_watch_test.py passes all checks -- including the one that failed first, "every one of the 44 reviewed classes is cited in docs/arch-gaps.md", which is the guard that made this a two-file change rather than a one-file one. Not addressed here: the UNVERIFIABLE (gated, no config) rows, which need a terminal state of their own so they stop being re-reported; that is a separate change to the watcher, not a registry entry. --- docs/arch-gaps.md | 30 +++++ registry/significant.json | 258 ++++++++++++++++++++------------------ 2 files changed, 165 insertions(+), 123 deletions(-) diff --git a/docs/arch-gaps.md b/docs/arch-gaps.md index c53592247..63e955403 100644 --- a/docs/arch-gaps.md +++ b/docs/arch-gaps.md @@ -1224,3 +1224,33 @@ The one `!! UNCOVERED` line of the 2026-11-14 watcher run. Read from the repo's not alerted (`"watch": "report"`), so the daily watcher prints `. REVIEWED` instead of `!! UNCOVERED` and does not exit 1 on it. If a backend starts accepting the class, record why as `mapped_ok` in `registry/significant.json` rather than deleting this entry. + +## `megatron_gpt_moe` — Megatron-LM GPT-MoE checkpoints + +* **Class / model_type**: `MegatronGPTMoeForCausalLM` / `megatron_gpt_moe`; example + `nagohachi/tiny-lm-v3-japanese-1.5b-a500m-sft`. +* **Evidence**: the repo's HF tags carry `megatron_gpt_moe` and `custom_code` + (checked 2026-11-17) — the modeling code ships in the repository rather than in + transformers, which is what the census calls `REMOTE-CODE`. +* **Why an alias is wrong**: this is a **training-framework checkpoint format**, + not a serving topology. A gguf conversion of one lands on whichever + architecture its exporter targets, so there is no single engine arch to map it + to, and treating it as an alias of `qwen*moe` / `llama` / `deepseek2` would + serve the model with the wrong tensor names — the failure mode that produces a + load error at best and wrong numbers at worst. +* **Real support needs**: an exporter (Megatron → gguf) and then whichever arch + it emits; the engine side would be that arch's path, not this class. + +## `rusmind_ai` — RusMind AI (from-scratch) + +* **Class / model_type**: `RusMindAIForCausalLM` / `rusmind_ai`; example + `DICPlay/RusMind-Zernyshko-2.0.1`. +* **Evidence**: the repo's HF tags carry `rusmind_ai`, `from-scratch` and + `custom_code` (checked 2026-11-17). The `from-scratch` tag is the load-bearing + one: it says this is not a finetune of a known family under a new name, which + is the case where a mapping would otherwise be plausible. +* **Why an alias is wrong**: the class string of a renamed finetune is not + evidence of the underlying architecture, and here the repository says it is a + new architecture. Mapping on the name would be guessing at the tensor layout. +* **Real support needs**: the architecture itself (attention/FFN/embedding + layout) plus a converter; nothing to map until that exists. diff --git a/registry/significant.json b/registry/significant.json index 4b72ce065..cff785fbd 100644 --- a/registry/significant.json +++ b/registry/significant.json @@ -1,88 +1,124 @@ { "about": "HF architecture classes reviewed as NOT aliases of a supported family (docs/arch-gaps.md). tools/registry_build.py --check-gaps fails if one becomes mapped without a 'mapped_ok' reason.", "classes": { - "DeepseekV41ForCausalLM": { - "model_type": "deepseek_v41", - "family": "DeepSeek V4.1 (engram embeddings + candidate-block sparse attention; NOT an alias of deepseek4 - different layer/head/expert shape, and the V4 path runs neither, which is why it needed its own port)", - "watch": "report", - "mapped_ok": "the pinned backend implements it as its own architecture: third_party/llama.cpp 5e6388a5a maps DeepseekV41ForCausalLM -> deepseek41 (conversion/__init__.py), registers LLM_ARCH_DEEPSEEK41 (src/llama-arch.cpp) and implements hparams/tensors/graph/engram in src/models/deepseek41.cpp - fork #156, the port of upstream #31086 + #31300. tools/registry_build.py --check-gaps stays quiet for this class; docs/arch-gaps.md carries the dated note." + "BagelForConditionalGeneration": { + "model_type": "bagel", + "family": "BAGEL-style unified understand+generate model (first seen: benson123dc/tmp-4461ae1a1900, 4 models in the 2026-11-12 census). Composite config: llm_config Qwen2 (hidden 3584, 28 layers, 28 heads / 4 KV) with qk_norm, a SigLIP vit_config (27 layers, patch 14) and a VAE latent path (visual_gen true). Every LLM layer carries a second Mixture-of-Transformers copy (*_moe_gen q/k/v/o, MLP and norms) for the image-generation tokens. Not an alias of Qwen2ForCausalLM: the qwen2 graph has no q_norm/k_norm and no _moe_gen weights, so even an understanding-only text path needs its own converter.", + "watch": "report" }, - "Qwen3Mamba3ForCausalLM": { - "model_type": "qwen3_mamba3", - "family": "Qwen3 + Mamba-3 hybrid" + "BlindForCausalLM": { + "model_type": "blind_decoder", + "family": "KenT Compressed-QK attention + gated key refinement (Vi-LAB)", + "watch": "report" }, - "MolmoAct2ForConditionalGeneration": { - "model_type": "molmoact2", - "family": "MolmoAct2, vision-language-action" + "CDMMoELMPoincare": { + "model_type": "cdm_moe_poincare", + "family": "CDM Poincare-geometry MoE (hyperbolic attention + telepathy gate, 629M, one uploader)", + "watch": "report" }, "Cosmos3EdgeForConditionalGeneration": { "model_type": "cosmos3_edge", "family": "Cosmos3-Edge, video world model" }, - "QwenDriveForPlanning": { - "model_type": "qwen_drive", - "family": "Qwen-Drive, driving planner" + "DFlash2DraftModel": { + "model_type": "dflash2", + "family": "DFlash2 speculative drafter (speculators; selector + conv heads the pinned dflash does not run)", + "watch": "report", + "mapped_ok": "the pinned backend now runs DFlash2, the condition docs/arch-gaps.md set: third_party/llama.cpp 2c178f44b (llama.cpp #327, engine#2138) registers DFlash2DraftModel with DFlashDraftModel in conversion/qwen.py, which writes conv_kernel_size, conv_group_size, selector_rank, selector_top_k and the candidate_selector codebooks, and src/models/dflash.cpp loads selector_prev/next/hidden and builds them (build_dflash2_conv, build_dflash2_selector). docs/arch-gaps.md carries the dated note." }, - "Qwen3_5MLAForConditionalGeneration": { - "model_type": "qwen3_5_mla", - "family": "Qwen3.5 MLA attention variant" + "DeepseekV41ForCausalLM": { + "model_type": "deepseek_v41", + "family": "DeepSeek V4.1 (engram embeddings + candidate-block sparse attention; NOT an alias of deepseek4 - different layer/head/expert shape, and the V4 path runs neither, which is why it needed its own port)", + "watch": "report", + "mapped_ok": "the pinned backend implements it as its own architecture: third_party/llama.cpp 5e6388a5a maps DeepseekV41ForCausalLM -> deepseek41 (conversion/__init__.py), registers LLM_ARCH_DEEPSEEK41 (src/llama-arch.cpp) and implements hparams/tensors/graph/engram in src/models/deepseek41.cpp - fork #156, the port of upstream #31086 + #31300. tools/registry_build.py --check-gaps stays quiet for this class; docs/arch-gaps.md carries the dated note." + }, + "DeepseekV4FlashSuperExpertForCausalLM": { + "model_type": "deepseek_v4_flash_superexpert", + "family": "DeepSeek V4.1 Flash SuperExpert, published as remote code and with AW-SECD-compressed weights (8.5 GB, claimed 0.962 retention); MLA-style geometry (kv_lora_rank 512, q_lora_rank 1536, qk_rope_head_dim 64, v_head_dim 128, 32 heads / 16 KV, IM 4096). Not an alias of deepseek_v41: nothing at the pin decodes AW-SECD weights, so mapping it would name a backend that cannot read it", + "watch": "report" + }, + "DiacritizerModel": { + "model_type": "tashkeel_nano", + "family": "tashkeel_nano Arabic diacritizer (bidirectional tagger, not a decoder)", + "watch": "report" }, "EnglishBaseForCausalLM": { "model_type": "fabryka_english_base", "family": "relu2 MLP + parameter-free QK-norm (implemented in the old engine only)" }, - "GDN2ForCausalLM": { - "model_type": "gdn2", - "family": "GDN2 hybrid (fla config)" - }, - "HaikuForCausalLM": { - "model_type": "haiku", - "family": "Haiku" + "Ernie5_5_MoeForCausalLM": { + "model_type": "ernie5_5_moe", + "family": "ERNIE 5.5 Flash MoE (hybrid MLA + DSA indexer, hyper-connections, latent MoE, situ activation; not an alias of ernie4_5-moe, nothing at the pin registers it)", + "watch": "report" }, - "VaporForCausalLM": { - "model_type": "vapor", - "family": "Vapor (lfm2-shaped config, different tensors)" + "Factorial": { + "model_type": "Factorial", + "family": "Factorial selection/memmap LM (repo-local training and inference code, .pt + weights.dat, SentencePiece; not a Transformer decoder, so nothing to alias)", + "watch": "report" }, "FidelForCausalLM": { "model_type": "fidel", "family": "Mamba/MoE hybrid" }, - "Zarya": { - "model_type": "zarya", - "family": "Zarya, diffusion LM sampler" + "GDN2ForCausalLM": { + "model_type": "gdn2", + "family": "GDN2 hybrid (fla config)" }, - "Needle3ForCausalLM": { - "model_type": "needle3", - "family": "Sakura Needle-3 (ONNX export)" + "GPT": { + "model_type": "nanogpt-custom", + "family": "nanoGPT-style one-offs (first seen: learner-subodh/Delivery-Rider-Assistant-SLM). 'GPT' is the class name nanoGPT's model.py exports, so this key covers any upload that saves nanoGPT's own state dict. Not an alias of GPT2LMHeadModel: the config uses nanoGPT's keys (n_embd 384, n_layer 6, n_head 8, block_size 256, bias, dropout) rather than GPT2Config's n_positions/n_ctx, the repo carries only a pickled pytorch_model.bin (no safetensors, no tokenizer files; the card names tiktoken gpt2), and nanoGPT stores nn.Linear weights where GPT-2's checkpoint stores Conv1D (transposed), so the gpt2 converter would read every projection the wrong way round. Nothing at the pin registers the class. Declined on 2026-11-09 under engine #1667; the census reports it as . REVIEWED. Remove this entry if a backend gains a nanoGPT import.", + "watch": "report" }, - "SepiaCharMLP": { - "model_type": "sepia-char-mlp", - "family": "SEPIA character-level MLP" + "GRTForCausalLM": { + "model_type": "grt", + "family": "GRT rotational/coupling decoder (MatterWaveAI grt-30m-code, 8 models in registry/census.json). Not an alias of anything mapped: the config reads model_type grt with a rotational coupling mixer (n_qkv_rotations 11, n_mixer_rotations 1, mor_router_mode rotational, mor_delta_scale 0.5, coupling_init_mode diagonal, angle_init_scale 0.1), learnable_rope with rope_placement stream, residual_gate_mode bounded_depth, no residual norm, and it declares transformers_version 5.23.0.dev0 - a research decoder, not a llama/qwen shape with extra keys. Nothing at the pin registers the class or a 'grt' gguf architecture (conversion/ has no grt module, src/llama-arch.cpp has no grt arch), so mapping it would claim a graph the engine cannot run. Declined with this reason on 2026-11-07 under engine #1328; the census reports it as . REVIEWED instead of !! UNCOVERED. Remove this entry and watch: report if a backend ever ports it.", + "watch": "report" }, - "KairoForCausalLM": { - "model_type": "kairo", - "family": "Kairo (crypto-model research)" + "Glm5NextForConditionalGeneration": { + "model_type": "glm5_next", + "family": "GLM-5.3-Flash (KDA linear-attention + DeepSeek sparse attention hybrid, hyper-connections; upstream llama.cpp 649dcb10, after the pin)", + "watch": "report", + "mapped_ok": "the pinned backend implements it as its own architecture: third_party/llama.cpp 2c178f44b (llama.cpp #327, engine#2138) maps Glm5NextForConditionalGeneration -> glm5-next (conversion/__init__.py; text in conversion/glm.py, vision tower in conversion/qwen3vl.py), registers LLM_ARCH_GLM5_NEXT (src/llama-arch.cpp, with its KDA + MLA cache layout in src/llama-model.cpp) and builds its graph in src/models/glm5-next.cpp. docs/arch-gaps.md carries the dated note." + }, + "GraniteSwitchForCausalLM": { + "model_type": "granite_switch", + "family": "Granite Switch 4.1 (dense all-attention Granite with N embedded per-token LoRA adapters, a zero adapter slot and weightless switch layers; upstream llama.cpp 157b81fe, not in the pin)", + "watch": "report", + "mapped_ok": "the pinned backend implements it as its own architecture: third_party/llama.cpp 2c178f44b (llama.cpp #327, engine#2138) maps GraniteSwitchForCausalLM -> graniteswitch (conversion/__init__.py, conversion/granite.py), registers LLM_ARCH_GRANITE_SWITCH (src/llama-arch.cpp) and builds its graph in src/models/granite-switch.cpp. docs/arch-gaps.md carries the dated note." }, "GrugMoeForCausalLM": { "model_type": "grug_moe", "family": "Grug MoE (marin-community research MoE)", "watch": "report" }, + "HaikuForCausalLM": { + "model_type": "haiku", + "family": "Haiku" + }, "HrmTextForCausalLM": { "model_type": "hrm_text", "family": "HRM-Text hierarchical-reasoning LM (Sapient)", "watch": "report", "mapped_ok": "the pinned backend implements it as its own architecture: third_party/llama.cpp 2c178f44b (llama.cpp #327, engine#2138) maps HrmTextForCausalLM -> hrm_text (conversion/__init__.py, conversion/hrm_text.py), registers LLM_ARCH_HRM_TEXT (src/llama-arch.cpp) and builds its graph in src/models/hrm-text.cpp. docs/arch-gaps.md carries the dated note." }, - "DFlash2DraftModel": { - "model_type": "dflash2", - "family": "DFlash2 speculative drafter (speculators; selector + conv heads the pinned dflash does not run)", + "IndonesianChatLM": { + "model_type": "indonesian_chat_lm", + "family": "Liwet (fadhilahkhairogi, a from-scratch Indonesian coursework MoE: Liwet-400M-A188M-id-{base,instruct,chat}). Not an alias of anything mapped. The repo ships config.json, model.safetensors and a tokenizer but no modeling code and no auto_map, and the class is not registered at the pin (conversion/ has no module for it, src/llama-arch.cpp has no arch). Its tensors are read from the safetensors header: no 'model.' prefix; per layer, q/k/v/o (16Q/4KV, head_dim 64, no q/k norm), a fused gate_up_proj [8, 1024, 720] and down_proj [8, 360, 1024] over 8 experts top-2, a router.weight [8, 1024], and two non-standard buffers, expert_bias [8] and load_ema [8], alongside router_bias_update_speed / router_load_decay / moe_capacity_factor in the config. Whether expert_bias enters the selection only (aux-loss-free balancing) or also the weights, and whether routing is softmax or sigmoid, is defined nowhere the engine can read, and the config's intermediate_size 2880 has no tensor behind it. Mapping it to a Mixtral/Qwen-MoE family would claim a routing rule the uploader never published. Declined on 2026-11-09 under engine #1667; the census reports it as . REVIEWED. Remove this entry if the uploader publishes modeling code a backend ports.", + "watch": "report" + }, + "KairoForCausalLM": { + "model_type": "kairo", + "family": "Kairo (crypto-model research)" + }, + "Lfm3ForCausalLM": { + "model_type": "lfm3", + "family": "LiquidAI LFM3 (sandwich norms, sliding-attention layers with partial rotary, 7-tap conv cache, tied embeddings; not an alias of lfm2/lfm2moe, nothing at the pin registers it)", "watch": "report", - "mapped_ok": "the pinned backend now runs DFlash2, the condition docs/arch-gaps.md set: third_party/llama.cpp 2c178f44b (llama.cpp #327, engine#2138) registers DFlash2DraftModel with DFlashDraftModel in conversion/qwen.py, which writes conv_kernel_size, conv_group_size, selector_rank, selector_top_k and the candidate_selector codebooks, and src/models/dflash.cpp loads selector_prev/next/hidden and builds them (build_dflash2_conv, build_dflash2_selector). docs/arch-gaps.md carries the dated note." + "mapped_ok": "the pinned backend implements it as its own architecture: third_party/llama.cpp 2c178f44b (llama.cpp #327, engine#2138) maps Lfm3ForCausalLM -> lfm3 (conversion/__init__.py, conversion/lfm3.py), registers LLM_ARCH_LFM3 (src/llama-arch.cpp) and builds its graph in src/models/lfm3.cpp. docs/arch-gaps.md carries the dated note." }, - "DiacritizerModel": { - "model_type": "tashkeel_nano", - "family": "tashkeel_nano Arabic diacritizer (bidirectional tagger, not a decoder)", + "LiveMemForCausalLM": { + "model_type": "livemem", + "family": "LiveMem RL backbone (Qwen2.5-shaped dense decoder, MLX q4 uploads)", "watch": "report" }, "MBartForConditionalGeneration": { @@ -90,15 +126,15 @@ "family": "mBART encoder-decoder seq2seq (no decoder-only route)", "watch": "report" }, - "BlindForCausalLM": { - "model_type": "blind_decoder", - "family": "KenT Compressed-QK attention + gated key refinement (Vi-LAB)", - "watch": "report" + "MegatronGPTMoeForCausalLM": { + "model_type": "megatron_gpt_moe", + "family": "Megatron-LM GPT-MoE: a training-framework checkpoint format with its own modeling code, not a serving topology. A gguf conversion of one lands on whichever architecture its exporter targets, so there is no single engine arch this class maps to and treating it as an alias of qwen*moe / llama / deepseek2 would serve the model with the wrong tensor names. Checked 2026-11-17 against nagohachi/tiny-lm-v3-japanese-1.5b-a500m-sft: HF tags include megatron_gpt_moe and custom_code.", + "watch": "report", + "checked": "2026-11-17" }, - "NanochronoForCausalLM": { - "model_type": "sn38-nanochrono", - "family": "sn38-nanochrono (ChronoLLM tap/aux-gate stack)", - "watch": "report" + "MolmoAct2ForConditionalGeneration": { + "model_type": "molmoact2", + "family": "MolmoAct2, vision-language-action" }, "MuseSparkForConditionalGeneration": { "model_type": "muse_spark", @@ -106,103 +142,79 @@ "watch": "report", "mapped_ok": "the pinned backend implements it as its own architecture: third_party/llama.cpp 2c178f44b (llama.cpp #327, engine#2138) maps MuseSparkForConditionalGeneration -> muse_spark (conversion/__init__.py, conversion/muse_spark.py), registers LLM_ARCH_MUSE_SPARK (src/llama-arch.cpp) and builds its graph in src/models/muse-spark.cpp. docs/arch-gaps.md carries the dated note." }, - "Glm5NextForConditionalGeneration": { - "model_type": "glm5_next", - "family": "GLM-5.3-Flash (KDA linear-attention + DeepSeek sparse attention hybrid, hyper-connections; upstream llama.cpp 649dcb10, after the pin)", - "watch": "report", - "mapped_ok": "the pinned backend implements it as its own architecture: third_party/llama.cpp 2c178f44b (llama.cpp #327, engine#2138) maps Glm5NextForConditionalGeneration -> glm5-next (conversion/__init__.py; text in conversion/glm.py, vision tower in conversion/qwen3vl.py), registers LLM_ARCH_GLM5_NEXT (src/llama-arch.cpp, with its KDA + MLA cache layout in src/llama-model.cpp) and builds its graph in src/models/glm5-next.cpp. docs/arch-gaps.md carries the dated note." - }, - "LiveMemForCausalLM": { - "model_type": "livemem", - "family": "LiveMem RL backbone (Qwen2.5-shaped dense decoder, MLX q4 uploads)", + "NanochronoForCausalLM": { + "model_type": "sn38-nanochrono", + "family": "sn38-nanochrono (ChronoLLM tap/aux-gate stack)", "watch": "report" }, - "CDMMoELMPoincare": { - "model_type": "cdm_moe_poincare", - "family": "CDM Poincare-geometry MoE (hyperbolic attention + telepathy gate, 629M, one uploader)", - "watch": "report" + "Needle3ForCausalLM": { + "model_type": "needle3", + "family": "Sakura Needle-3 (ONNX export)" }, - "Senryu2ForCausalLM": { - "model_type": "senryu2", - "family": "Senryu-2.2 GDN2 + parallax-attention hybrid with repo-local n-gram tables", + "NeedleForToolCalling": { + "model_type": "needle", + "family": "Cactus-Compute needle3 tool-calling model (hadamard MLP, engram slots, confidence head, cactus-quants, .cact container for its own on-device runtime; not a Transformer a pinned backend can load)", "watch": "report" }, - "GraniteSwitchForCausalLM": { - "model_type": "granite_switch", - "family": "Granite Switch 4.1 (dense all-attention Granite with N embedded per-token LoRA adapters, a zero adapter slot and weightless switch layers; upstream llama.cpp 157b81fe, not in the pin)", - "watch": "report", - "mapped_ok": "the pinned backend implements it as its own architecture: third_party/llama.cpp 2c178f44b (llama.cpp #327, engine#2138) maps GraniteSwitchForCausalLM -> graniteswitch (conversion/__init__.py, conversion/granite.py), registers LLM_ARCH_GRANITE_SWITCH (src/llama-arch.cpp) and builds its graph in src/models/granite-switch.cpp. docs/arch-gaps.md carries the dated note." - }, - "Ernie5_5_MoeForCausalLM": { - "model_type": "ernie5_5_moe", - "family": "ERNIE 5.5 Flash MoE (hybrid MLA + DSA indexer, hyper-connections, latent MoE, situ activation; not an alias of ernie4_5-moe, nothing at the pin registers it)", + "OleanderForCausalLM": { + "model_type": "oleander", + "family": "Oleander-150M, a from-scratch training export (first seen: CyanMonkey/Test-Subject-3; its own config says step 2000 of 28820, 6.94 %). Hybrid 28 layers in a gdn2,gdn2,gdn2,gqa pattern, hidden 576, vocab 32770, F32. Ships no modelling code and no auto_map. Not an alias of the Qwen3.5 / Qwen3-Next linear-attention families despite the same layer_types: gdn2 layers carry per-q/k/v conv1d and low-rank f/g gate projections, every sublayer has attention-residual mixers (attn_res_*), and GQA layers a fused qg_proj and a value-residual v_mix; no pinned graph has those tensors.", "watch": "report" }, - "Factorial": { - "model_type": "Factorial", - "family": "Factorial selection/memmap LM (repo-local training and inference code, .pt + weights.dat, SentencePiece; not a Transformer decoder, so nothing to alias)", + "OpensciForCausalLM": { + "model_type": "opensci", + "family": "OpenSci (1.7B 'nemotron-hq' recipe). A REMOTE-CODE class: its config carries an auto_map pointing at modeling_opensci.py / configuration_opensci.py, so the pinned backends cannot accept the architecture without that custom code. Not an alias of any supported family, and not mappable until the code is vendored or upstreamed.", "watch": "report" }, - "Lfm3ForCausalLM": { - "model_type": "lfm3", - "family": "LiquidAI LFM3 (sandwich norms, sliding-attention layers with partial rotary, 7-tap conv cache, tied embeddings; not an alias of lfm2/lfm2moe, nothing at the pin registers it)", - "watch": "report", - "mapped_ok": "the pinned backend implements it as its own architecture: third_party/llama.cpp 2c178f44b (llama.cpp #327, engine#2138) maps Lfm3ForCausalLM -> lfm3 (conversion/__init__.py, conversion/lfm3.py), registers LLM_ARCH_LFM3 (src/llama-arch.cpp) and builds its graph in src/models/lfm3.cpp. docs/arch-gaps.md carries the dated note." + "Qwen3Mamba3ForCausalLM": { + "model_type": "qwen3_mamba3", + "family": "Qwen3 + Mamba-3 hybrid" }, - "NeedleForToolCalling": { - "model_type": "needle", - "family": "Cactus-Compute needle3 tool-calling model (hadamard MLP, engram slots, confidence head, cactus-quants, .cact container for its own on-device runtime; not a Transformer a pinned backend can load)", - "watch": "report" + "Qwen3_5MLAForConditionalGeneration": { + "model_type": "qwen3_5_mla", + "family": "Qwen3.5 MLA attention variant" }, - "GRTForCausalLM": { - "model_type": "grt", - "family": "GRT rotational/coupling decoder (MatterWaveAI grt-30m-code, 8 models in registry/census.json). Not an alias of anything mapped: the config reads model_type grt with a rotational coupling mixer (n_qkv_rotations 11, n_mixer_rotations 1, mor_router_mode rotational, mor_delta_scale 0.5, coupling_init_mode diagonal, angle_init_scale 0.1), learnable_rope with rope_placement stream, residual_gate_mode bounded_depth, no residual norm, and it declares transformers_version 5.23.0.dev0 - a research decoder, not a llama/qwen shape with extra keys. Nothing at the pin registers the class or a 'grt' gguf architecture (conversion/ has no grt module, src/llama-arch.cpp has no grt arch), so mapping it would claim a graph the engine cannot run. Declined with this reason on 2026-11-07 under engine #1328; the census reports it as . REVIEWED instead of !! UNCOVERED. Remove this entry and watch: report if a backend ever ports it.", - "watch": "report" + "QwenDriveForPlanning": { + "model_type": "qwen_drive", + "family": "Qwen-Drive, driving planner" }, "RMSN": { "model_type": "rmsn", "family": "RMSN, Resonant Metriplectic Sheaf Network (DuoNeural physical-informed research decoder: model_type rmsn with dim/depth/slots/seq_len and a Lyapunov diagnostic in the config, card tags port-hamiltonian/metriplectic/lyapunov/sheaf, its own GitHub runtime - a continuous-time dynamical system, not a Transformer with attention/FFN tensors). Not an alias of anything mapped: config.json declares architectures [RMSN] with vocab_size 49152 on the HuggingFaceTB/cosmo2-tokenizer, dim 1024, depth 16, slots 64, seq_len 512, 243.7 M parameters and no attention/FFN keys, and nothing at the pin registers the class or an 'rmsn' gguf architecture (conversion/ has no module for it, src/llama-arch.cpp has no arch), so mapping it would claim a graph the engine cannot run. Declined with this reason on 2026-11-08 under engine #1451; the census reports it as . REVIEWED instead of !! UNCOVERED. Remove this entry and watch: report if a backend ever ports it.", "watch": "report" }, - "IndonesianChatLM": { - "model_type": "indonesian_chat_lm", - "family": "Liwet (fadhilahkhairogi, a from-scratch Indonesian coursework MoE: Liwet-400M-A188M-id-{base,instruct,chat}). Not an alias of anything mapped. The repo ships config.json, model.safetensors and a tokenizer but no modeling code and no auto_map, and the class is not registered at the pin (conversion/ has no module for it, src/llama-arch.cpp has no arch). Its tensors are read from the safetensors header: no 'model.' prefix; per layer, q/k/v/o (16Q/4KV, head_dim 64, no q/k norm), a fused gate_up_proj [8, 1024, 720] and down_proj [8, 360, 1024] over 8 experts top-2, a router.weight [8, 1024], and two non-standard buffers, expert_bias [8] and load_ema [8], alongside router_bias_update_speed / router_load_decay / moe_capacity_factor in the config. Whether expert_bias enters the selection only (aux-loss-free balancing) or also the weights, and whether routing is softmax or sigmoid, is defined nowhere the engine can read, and the config's intermediate_size 2880 has no tensor behind it. Mapping it to a Mixtral/Qwen-MoE family would claim a routing rule the uploader never published. Declined on 2026-11-09 under engine #1667; the census reports it as . REVIEWED. Remove this entry if the uploader publishes modeling code a backend ports.", - "watch": "report" - }, - "GPT": { - "model_type": "nanogpt-custom", - "family": "nanoGPT-style one-offs (first seen: learner-subodh/Delivery-Rider-Assistant-SLM). 'GPT' is the class name nanoGPT's model.py exports, so this key covers any upload that saves nanoGPT's own state dict. Not an alias of GPT2LMHeadModel: the config uses nanoGPT's keys (n_embd 384, n_layer 6, n_head 8, block_size 256, bias, dropout) rather than GPT2Config's n_positions/n_ctx, the repo carries only a pickled pytorch_model.bin (no safetensors, no tokenizer files; the card names tiktoken gpt2), and nanoGPT stores nn.Linear weights where GPT-2's checkpoint stores Conv1D (transposed), so the gpt2 converter would read every projection the wrong way round. Nothing at the pin registers the class. Declined on 2026-11-09 under engine #1667; the census reports it as . REVIEWED. Remove this entry if a backend gains a nanoGPT import.", - "watch": "report" - }, - "DeepseekV4FlashSuperExpertForCausalLM": { - "model_type": "deepseek_v4_flash_superexpert", - "family": "DeepSeek V4.1 Flash SuperExpert, published as remote code and with AW-SECD-compressed weights (8.5 GB, claimed 0.962 retention); MLA-style geometry (kv_lora_rank 512, q_lora_rank 1536, qk_rope_head_dim 64, v_head_dim 128, 32 heads / 16 KV, IM 4096). Not an alias of deepseek_v41: nothing at the pin decodes AW-SECD weights, so mapping it would name a backend that cannot read it", - "watch": "report" + "RusMindAIForCausalLM": { + "model_type": "rusmind_ai", + "family": "RusMind AI: a from-scratch architecture with its own modeling code, not a rename of a supported family. Checked 2026-11-17 against DICPlay/RusMind-Zernyshko-2.0.1: HF tags include rusmind_ai, from-scratch and custom_code, which is the evidence that it is neither an alias nor a conversion target today.", + "watch": "report", + "checked": "2026-11-17" }, - "OpensciForCausalLM": { - "model_type": "opensci", - "family": "OpenSci (1.7B 'nemotron-hq' recipe). A REMOTE-CODE class: its config carries an auto_map pointing at modeling_opensci.py / configuration_opensci.py, so the pinned backends cannot accept the architecture without that custom code. Not an alias of any supported family, and not mappable until the code is vendored or upstreamed.", + "SegaForCausalLM": { + "model_type": "sega_tot_moe", + "family": "Sega 7.0 'Singularity-Infinity' (the123manaic) - 40 layers, hidden 4096, 64 heads. Reviewed as NOT an alias of a family we serve: model_type `sega_tot_moe` is not supported, the config carries non-standard `reasoning_passes` / `reasoning_speed_*` keys, and there is no auto_map (so it is not the remote-code case either - it is a custom architecture with a plain config).", "watch": "report" }, - "BagelForConditionalGeneration": { - "model_type": "bagel", - "family": "BAGEL-style unified understand+generate model (first seen: benson123dc/tmp-4461ae1a1900, 4 models in the 2026-11-12 census). Composite config: llm_config Qwen2 (hidden 3584, 28 layers, 28 heads / 4 KV) with qk_norm, a SigLIP vit_config (27 layers, patch 14) and a VAE latent path (visual_gen true). Every LLM layer carries a second Mixture-of-Transformers copy (*_moe_gen q/k/v/o, MLP and norms) for the image-generation tokens. Not an alias of Qwen2ForCausalLM: the qwen2 graph has no q_norm/k_norm and no _moe_gen weights, so even an understanding-only text path needs its own converter.", + "Senryu2ForCausalLM": { + "model_type": "senryu2", + "family": "Senryu-2.2 GDN2 + parallax-attention hybrid with repo-local n-gram tables", "watch": "report" }, - "OleanderForCausalLM": { - "model_type": "oleander", - "family": "Oleander-150M, a from-scratch training export (first seen: CyanMonkey/Test-Subject-3; its own config says step 2000 of 28820, 6.94 %). Hybrid 28 layers in a gdn2,gdn2,gdn2,gqa pattern, hidden 576, vocab 32770, F32. Ships no modelling code and no auto_map. Not an alias of the Qwen3.5 / Qwen3-Next linear-attention families despite the same layer_types: gdn2 layers carry per-q/k/v conv1d and low-rank f/g gate projections, every sublayer has attention-residual mixers (attn_res_*), and GQA layers a fused qg_proj and a value-residual v_mix; no pinned graph has those tensors.", - "watch": "report" + "SepiaCharMLP": { + "model_type": "sepia-char-mlp", + "family": "SEPIA character-level MLP" }, "Skylar2ForCausalLM": { "model_type": "skylar2", "family": "Skylar 2 (first seen: Skyl4r-Ai/Skylar2-1B-Instruct, the 2026-11-13 census), a from-scratch 1B hybrid that loads through its own `skylar` pip package (no auto_map, no modelling code in the repo). 36 layers, hidden 1536, vocab 64000: 27 KDA linear-attention layers (A_log, dt_bias, per-q/k/v short conv, low-rank f_a/f_b and g_a/g_b gates) and 9 GQA layers (12 heads / 4 KV, qk_norm, a per-head output gate W_gate) in a 3:1 pattern; every norm is a gated RMSNorm with rank-16 gate_up/gate_down, every sublayer has an attention-residual mixer (res_attn/res_ffn query + key_weight), and the FFN uses a situ_glu activation. Not an alias of KimiLinearForCausalLM (different KDA gate, norms and residual mixers) nor of Kimi K3 in llama.cpp #327, which shares situ and attention residuals but uses MLA attention, a full-rank KDA gate and a latent MoE. No pinned graph has these tensors.", "watch": "report" }, - "SegaForCausalLM": { - "model_type": "sega_tot_moe", - "family": "Sega 7.0 'Singularity-Infinity' (the123manaic) - 40 layers, hidden 4096, 64 heads. Reviewed as NOT an alias of a family we serve: model_type `sega_tot_moe` is not supported, the config carries non-standard `reasoning_passes` / `reasoning_speed_*` keys, and there is no auto_map (so it is not the remote-code case either - it is a custom architecture with a plain config).", - "watch": "report" + "VaporForCausalLM": { + "model_type": "vapor", + "family": "Vapor (lfm2-shaped config, different tensors)" + }, + "Zarya": { + "model_type": "zarya", + "family": "Zarya, diffusion LM sampler" } } } From f0d7afea235c8cd7fc98f33ea3488754e61d756b Mon Sep 17 00:00:00 2001 From: hrx-compat lane Date: Sun, 15 Nov 2026 10:24:12 -0400 Subject: [PATCH 2/4] census: regenerate the summary for the two reviewed classes (engine#2528) Adding MegatronGPTMoeForCausalLM and RusMindAIForCausalLM to registry/significant.json changes what the summary derives from the raw block: declined_models 434 -> 442 and declined_architectures 21 -> 23, while registry/census.json still carried the old numbers. That is exactly the invariant engine#2540's arm enforces -- the committed summary must be what its own raw block implies -- and it caught this PR before review, which is the arm doing its job on a real change rather than in a test. `census.coverage(raw)` is pure and `raw` is committed, so this is a regeneration, not a re-sweep: no network, no new data, just the summary the file's own raw block already implies. Checked: tests/census_watch_test.py all checks passed (it failed before this commit with "declined_models: (442, 434), declined_architectures: (23, 21)"), and tools/registry_build.py --check-gaps still reports none of the 44 reviewed classes mapped without a reason. --- registry/census.json | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/registry/census.json b/registry/census.json index ec0dd910f..e6ed43d0a 100644 --- a/registry/census.json +++ b/registry/census.json @@ -7,8 +7,8 @@ "declined": "a human reviewed the class and deliberately did not map it (registry/significant.json, `watch: \"report\"`). Counted separately and excluded from top_unmapped: it is a recorded decision, not work remaining. The denominator is unchanged -- declined models are still `with_architecture`", "with_architecture": "models whose HF card names an architecture this census can look up; models with no architecture are excluded from every percentage above" }, - "declined_models": 434, - "declined_architectures": 21, + "declined_models": 442, + "declined_architectures": 23, "models": 438457, "with_architecture": 345360, "architectures_seen": 3071, From c4d3f19c4f8c38f816d764a2dafb97e09b797538 Mon Sep 17 00:00:00 2001 From: hrx-compat lane Date: Sun, 15 Nov 2026 10:32:31 -0400 Subject: [PATCH 3/4] census: an UNVERIFIABLE row is recorded, not retried -- say so (engine#2528) The line said "retried next run", which is not what happens. remember() records an unverifiable model in state["seen"] exactly like every other group, and classify() skips anything already there, so the row is reported ONCE and then goes quiet. Promising a retry that never comes is how a daily report teaches its readers to skim: they wait for a resolution, the line stops appearing, and they cannot tell "resolved" from "remembered". The message now says which, and names the way back in (DEFAULT_STATE, or add a token and re-run). Found while working engine#2528: I had asserted in the issue that the gated rows were "retried and re-reported forever", and reading remember() showed the mechanism was already there -- the misleading text was the whole defect, not the behaviour. --- tools/census_watch.py | 8 +++++++- 1 file changed, 7 insertions(+), 1 deletion(-) diff --git a/tools/census_watch.py b/tools/census_watch.py index 45f6247b3..50e5c1756 100644 --- a/tools/census_watch.py +++ b/tools/census_watch.py @@ -973,8 +973,14 @@ def report(models, res, previously_seen): # run that sent one and still got no config back hint = " (HF_TOKEN sent, still no config)" if hf_token() else "" for mid in sorted(res[UNVERIFIABLE]): + # engine#2528: this said "retried next run", which is not what happens. remember() records an + # unverifiable model in state["seen"] like every other group, and classify() skips anything already + # there -- so the row is reported ONCE and then goes quiet. Promising a retry that never comes is how a + # daily report teaches its readers to skim: they wait for a resolution, stop seeing the line, and cannot + # tell "resolved" from "remembered". Say which one it is, and name the way back in. print(f" ? UNVERIFIABLE {mid} ({res['reasons'][mid]}) — gated repos need a token{hint}; " - f"retried next run") + f"recorded, not retried automatically (remove it from {DEFAULT_STATE}, or add a token and " + f"re-run, to see it again)") return 1 if any(c not in quiet for c in res["classes"]) else 0 From 19f281885df4e1b9d1d249f6bd35556d954736b5 Mon Sep 17 00:00:00 2001 From: hrx-compat lane Date: Sun, 15 Nov 2026 10:43:34 -0400 Subject: [PATCH 4/4] census: pin the UNVERIFIABLE line's claim, not just its behaviour (engine#2528) The test suite already had the behavioural arm -- "unverifiable remembered -> quiet" -- and it passed for as long as the message said "retried next run", which is why the wrong promise survived: the behaviour was correct, and nothing checked that the line describing it was. This adds that arm. Two-sided, checked by reverting the message locally: old line ("retried next run") -> FAIL unverifiable line says it is recorded, not retried this line ("recorded, not retried ...") -> all checks passed That is the same distinction as the registry guards on this issue (a documented decision, a derived summary): pinning the mechanism does not pin the report, and the report is what a reader acts on. --- tests/census_watch_test.py | 13 +++++++++++++ 1 file changed, 13 insertions(+) diff --git a/tests/census_watch_test.py b/tests/census_watch_test.py index 23c3527c5..9196f059f 100644 --- a/tests/census_watch_test.py +++ b/tests/census_watch_test.py @@ -206,6 +206,19 @@ def own_code(mid, arch, files=("config.json", "model.safetensors", "modeling_mic rc, res = rc_of([model("f/seven", tags=["text-generation"])], seen={"f/seven"}) check("unverifiable remembered -> quiet", rc == 0 and res[cw.UNVERIFIABLE] == []) +# 7b. engine#2528: the LINE must say what the behaviour does. The arm above pins that an unverifiable +# model is remembered and not re-reported; this one pins that the message does not promise a retry that +# the code will never perform. It said "retried next run" for as long as the behaviour was already +# correct, so pinning the behaviour alone could not catch it. +buf = io.StringIO() +with contextlib.redirect_stdout(buf): + cw.report([model("f/eight", tags=["text-generation"])], + cw.classify([model("f/eight", tags=["text-generation"])], ARCHS, SIG, ()), ()) +line = [l for l in buf.getvalue().splitlines() if "UNVERIFIABLE" in l] +check("unverifiable line says it is recorded, not retried", + len(line) == 1 and "retried next run" not in line[0] + and "not retried automatically" in line[0]) + # 8. A mixed batch: the alert fires, and the summary counts each group. rc, res = rc_of([ model("g/eight", "Qwen3ForCausalLM"),