diff --git a/registry/census.json b/registry/census.json index f92ef3de8..8d610006b 100644 --- a/registry/census.json +++ b/registry/census.json @@ -1,24 +1,32 @@ { - "date": "2026-11-09", + "date": "2026-11-10", "coverage": { - "models": 435467, - "with_architecture": 344003, - "architectures_seen": 3030, - "mapped_models": 324692, - "mapped_pct": 94.39, - "checked_models": 218943, - "checked_pct": 63.65, + "definitions": { + "mapped": "the pinned backend's code accepts this HF architecture class (registry/architectures.json names a backend for it); it is a property of the ARCHITECTURE, so every model of a mapped class counts as mapped whether or not that model was ever run", + "checked": "the architecture has a PASSING serve_e2e at the pin -- not that this model was run. It is the same per-architecture property, which is why the npu figure is essentially one architecture's model population rather than a count of models anyone exercised", + "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": 687, + "declined_architectures": 21, + "models": 435913, + "with_architecture": 344207, + "architectures_seen": 3037, + "mapped_models": 324863, + "mapped_pct": 94.38, + "checked_models": 219019, + "checked_pct": 63.63, "per_backend": { "hrx": { - "mapped": 324692, - "checked": 218943, - "mapped_pct": 94.39, - "checked_pct": 63.65 + "mapped": 324863, + "checked": 219019, + "mapped_pct": 94.38, + "checked_pct": 63.63 }, "npu": { - "mapped": 189569, - "checked": 32493, - "mapped_pct": 55.11, + "mapped": 189670, + "checked": 32514, + "mapped_pct": 55.1, "checked_pct": 9.45 } }, @@ -64,16 +72,16 @@ "models": 174 }, { - "architecture": "MixFormerSequentialForCausalLM", + "architecture": "K2HorizonForCausalLM", "models": 156 }, { - "architecture": "K2HorizonForCausalLM", - "models": 154 + "architecture": "MixFormerSequentialForCausalLM", + "models": 156 }, { "architecture": "Spark2_5ForCausalLM", - "models": 146 + "models": 148 }, { "architecture": "RobertaForCausalLM", @@ -83,22 +91,10 @@ "architecture": "NanochatGPTForCausalLM", "models": 126 }, - { - "architecture": "Lfm3ForCausalLM", - "models": 121 - }, { "architecture": "OpensciForCausalLM", "models": 121 }, - { - "architecture": "MBartForConditionalGeneration", - "models": 112 - }, - { - "architecture": "Glm5NextForConditionalGeneration", - "models": 111 - }, { "architecture": "BartForCausalLM", "models": 100 @@ -117,7 +113,7 @@ }, { "architecture": "BailingMoeV3ForCausalLM", - "models": 84 + "models": 85 }, { "architecture": "GPT2LMHeadCustomModel", @@ -126,59 +122,59 @@ ] }, "raw": { - "total": 435467, - "no_arch": 91464, + "total": 435913, + "no_arch": 91706, "pages": 436, "complete": true, "counts": { - "LlamaForCausalLM": 101752, - "Qwen2ForCausalLM": 47664, - "GPT2LMHeadModel": 33352, - "Qwen3ForCausalLM": 32493, + "LlamaForCausalLM": 101774, + "Qwen2ForCausalLM": 47692, + "GPT2LMHeadModel": 33384, + "Qwen3ForCausalLM": 32514, "MistralForCausalLM": 28244, "GPTNeoXForCausalLM": 8122, "StableLmForCausalLM": 6683, - "Gemma3ForCausalLM": 6226, - "Gemma2ForCausalLM": 5675, + "Gemma3ForCausalLM": 6228, + "Gemma2ForCausalLM": 5673, "GemmaForCausalLM": 5114, - "Qwen3_5ForConditionalGeneration": 3749, - "Phi3ForCausalLM": 3258, + "Qwen3_5ForConditionalGeneration": 3766, + "Phi3ForCausalLM": 3257, "Step1MoEForCausalLM": 2882, "MixtralForCausalLM": 2855, "OPTForCausalLM": 2079, - "GptOssForCausalLM": 1958, - "Qwen3MoeForCausalLM": 1822, - "PhiForCausalLM": 1660, - "GPTNeoForCausalLM": 1609, + "GptOssForCausalLM": 1961, + "Qwen3MoeForCausalLM": 1827, + "PhiForCausalLM": 1658, + "GPTNeoForCausalLM": 1610, "ParlerTTSForConditionalGeneration": 1585, - "Qwen3_5ForCausalLM": 1568, - "Qwen3_5MoeForConditionalGeneration": 1558, - "Gemma4ForConditionalGeneration": 1513, - "Lfm2ForCausalLM": 1235, + "Qwen3_5ForCausalLM": 1576, + "Qwen3_5MoeForConditionalGeneration": 1556, + "Gemma4ForConditionalGeneration": 1520, + "Lfm2ForCausalLM": 1231, "BloomForCausalLM": 1194, "Gemma3ForConditionalGeneration": 1130, "Olmo3ForCausalLM": 1125, - "T5ForConditionalGeneration": 1059, + "T5ForConditionalGeneration": 1062, "ChessForCausalLM": 970, "FalconForCausalLM": 923, - "GlmMoeDsaForCausalLM": 731, + "GlmMoeDsaForCausalLM": 733, "Olmo2ForCausalLM": 725, - "GraniteForCausalLM": 637, + "GraniteForCausalLM": 639, "GPTJForCausalLM": 614, "DeepseekV4ForCausalLM": 611, + "NemotronHForCausalLM": 604, "OlmoForCausalLM": 601, - "NemotronHForCausalLM": 594, "DeepseekV3ForCausalLM": 537, "CohereForCausalLM": 534, "GraniteMoeHybridForCausalLM": 532, "PicoDecoderHF": 503, - "SmolLM3ForCausalLM": 462, + "SmolLM3ForCausalLM": 463, "LlavaLlamaForCausalLM": 453, "GPTBigCodeForCausalLM": 433, "CodeGenForCausalLM": 387, "MT5ForConditionalGeneration": 364, - "Qwen3NextForCausalLM": 357, - "Lfm2MoeForCausalLM": 350, + "Qwen3NextForCausalLM": 356, + "Lfm2MoeForCausalLM": 353, "QWenLMHeadModel": 347, "Glm4MoeForCausalLM": 340, "MiniMaxM2ForCausalLM": 319, @@ -192,10 +188,10 @@ "Qwen3_5MoeForCausalLM": 235, "SeedOssForCausalLM": 231, "Glm4ForCausalLM": 228, - "Mistral3ForConditionalGeneration": 214, + "Mistral3ForConditionalGeneration": 215, "BertLMHeadModel": 212, + "Cohere2ForCausalLM": 191, "ExaoneForCausalLM": 186, - "Cohere2ForCausalLM": 184, "Glm4MoeLiteForCausalLM": 179, "OlmoeForCausalLM": 178, "MobileLLMForCausalLM": 176, @@ -208,29 +204,29 @@ "OpenELMForCausalLM": 166, "DFlashDraftModel": 163, "MambaForCausalLM": 157, + "K2HorizonForCausalLM": 156, "MixFormerSequentialForCausalLM": 156, - "K2HorizonForCausalLM": 154, "LLaMAForCausalLM": 153, - "Qwen4ExpForConditionalGeneration": 146, - "Spark2_5ForCausalLM": 146, - "Gemma4ForCausalLM": 145, + "Spark2_5ForCausalLM": 148, + "Qwen4ExpForConditionalGeneration": 147, + "Gemma4ForCausalLM": 146, "ApertusForCausalLM": 143, "RobertaForCausalLM": 139, "BaichuanForCausalLM": 132, "DeciLMForCausalLM": 130, "LlamaForCausalLMEagle3": 130, + "Lfm3ForCausalLM": 126, "NanochatGPTForCausalLM": 126, "InternLM2ForCausalLM": 122, - "Lfm3ForCausalLM": 121, "OpensciForCausalLM": 121, "MiniCPMForCausalLM": 118, + "Glm5NextForConditionalGeneration": 115, "HunYuanDenseV1ForCausalLM": 114, "CambrianQwenForCausalLM": 113, "MBartForConditionalGeneration": 112, - "Glm5NextForConditionalGeneration": 111, "HYV3ForCausalLM": 107, + "DeepseekV41ForCausalLM": 106, "MiMoV2ForCausalLM": 105, - "DeepseekV41ForCausalLM": 103, "Qwen2MoeForCausalLM": 103, "BartForCausalLM": 100, "LlavaQwen2ForCausalLM": 97, @@ -238,14 +234,14 @@ "JAISLMHeadModel": 89, "BioGptForCausalLM": 86, "RwkvForCausalLM": 86, + "BailingMoeV3ForCausalLM": 85, "AfmoeForCausalLM": 84, - "BailingMoeV3ForCausalLM": 84, "GPT2LMHeadCustomModel": 84, "Llama4ForCausalLM": 82, "NemotronForCausalLM": 81, "DFlash2DraftModel": 80, + "FlintForCausalLM": 80, "DynamicSlidingWindowForCausalLM": 79, - "FlintForCausalLM": 79, "XGLMForCausalLM": 78, "MarianForCausalLM": 76, "NanoChatForCausalLM": 75, @@ -363,6 +359,7 @@ "GPT2MoEForCausalLM": 23, "M2M100ForConditionalGeneration": 23, "MiMoV2FlashForCausalLM": 23, + "RicursiveForCausalLM": 23, "GPT2LMAndValueHeadModel": 22, "HunYuanMoEV1ForCausalLM": 22, "MiniMaxM3SparseForConditionalGeneration": 22, @@ -375,7 +372,6 @@ "LlavaPhiForCausalLM": 21, "MotifForCausalLM": 21, "ProGenForCausalLM": 21, - "RicursiveForCausalLM": 21, "SkipMiddleModel": 21, "ZayaForCausalLM": 21, "BeamForCausalLM": 20, @@ -396,6 +392,7 @@ "BVVForCausalLM": 18, "BailingMoeLinearV2ForCausalLM": 18, "BaseLoopLlamaForCausalLM": 18, + "DoryForCausalLM": 18, "FalconMambaForCausalLM": 18, "GraniteSwitchForCausalLM": 18, "MFuyuForCausalLM": 18, @@ -408,7 +405,6 @@ "StickbreakingForCausalLM": 18, "AlibiForCausalLM": 17, "CubeLM": 17, - "DoryForCausalLM": 17, "GPTForCausalLM": 17, "InternLMForCausalLM": 17, "NCPOlmo3ForCausalLM": 17, @@ -452,6 +448,7 @@ "BertModel": 13, "DiscreteDiffusionModel": 13, "Eagle3DeepseekV2ForCausalLM": 13, + "Eagle3DraftModel": 13, "EncoderDecoderModel": 13, "Ernie4_5_ForCausalLM": 13, "MobilintQwen3ForCausalLM": 13, @@ -493,7 +490,6 @@ "Zamba2ForCausalLM": 12, "A2DQwen3LMHeadModel": 11, "BananaMind2ProForCausalLM": 11, - "Eagle3DraftModel": 11, "ElectraForCausalLM": 11, "Ernie4_5ForCausalLM": 11, "GRTForCausalLM": 11, @@ -516,6 +512,7 @@ "BabyLlamaForCausalLM": 10, "BambaForCausalLM": 10, "EmoForCausalLM": 10, + "FHN_T4Max_150M_SingleHead_GAU": 10, "FinanceDecoder": 10, "GPTBERTForCausalLM": 10, "Gemma4UnifiedAssistantForCausalLM": 10, @@ -552,7 +549,6 @@ "CognicaPoEForCausalLM": 9, "Dots1ForCausalLM": 9, "Evo2ForCausalLM": 9, - "FHN_T4Max_150M_SingleHead_GAU": 9, "GLAForCausalLM": 9, "GPTPanguForCausalLM": 9, "HelpingAIForCausalLM": 9, @@ -588,6 +584,7 @@ "DFlareDraftModel": 8, "DUO": 8, "DeepForCausalLM": 8, + "DomynEdgeForCausalLM": 8, "DuchifatCore": 8, "FP8Qwen2ForCausalLM": 8, "GPT3DevLMHeadModel": 8, @@ -636,6 +633,7 @@ "FWKVLanguageModel": 7, "Int8OPTForCausalLM": 7, "K3DSparkModel": 7, + "KimiK31ForConditionalGeneration": 7, "LLaMAModel": 7, "LizzyForCausalLM": 7, "LlamaPixarForCausalLM": 7, @@ -690,7 +688,6 @@ "InklingForConditionalGeneration": 6, "InternLMXComposer2ForCausalLM": 6, "JugnuVRForCausalLM": 6, - "KimiK31ForConditionalGeneration": 6, "LSGBartForConditionalGeneration": 6, "LilleForCausalLM": 6, "LlavaMambaForCausalLM": 6, @@ -784,6 +781,7 @@ "RWKV6Qwen2ForCausalLM": 5, "ReplitLM": 5, "RosettaForCausalLM": 5, + "RusMindAIForCausalLM": 5, "RwkvHybridForCausalLM": 5, "SansarForCausalLM": 5, "Share4VLlamaForCausalLM": 5, @@ -823,7 +821,6 @@ "DLiteDraftModel": 4, "DeltaNetForCausalLM": 4, "DiffuQwen35": 4, - "DomynEdgeForCausalLM": 4, "DuoLagunaForCausalLM": 4, "EffT5ForConditionalGeneration": 4, "EfficientDLM": 4, @@ -863,6 +860,7 @@ "LoamForCausalLM": 4, "LoopMoEForCausalLM": 4, "LoopedGDNForCausalLM": 4, + "MM200ForCausalLM": 4, "MPLUGDocOwlForConditionalGeneration": 4, "MambaModelForCausalLM": 4, "MegrezMoeForCausalLM": 4, @@ -910,7 +908,6 @@ "RecursiveCompressorLM": 4, "RoseX1ForCausalLM": 4, "RuGPT3XLForCausalLM": 4, - "RusMindAIForCausalLM": 4, "SLM": 4, "SLMForCausalLM": 4, "SeerAttnLlamaForCausalLM": 4, @@ -958,7 +955,6 @@ "BigBirdPegasusForCausalLM": 3, "BlazeForCausalLM": 3, "CFRDForCausalLM": 3, - "CascadeForCausalLM": 3, "CausalLM": 3, "CloverMoeModelLM": 3, "CoRMForCausalLM": 3, @@ -1067,7 +1063,6 @@ "LumeesForCausalLM": 3, "LuminaNextForCausalLM": 3, "LummaForCausalLM": 3, - "MM200ForCausalLM": 3, "MabaSparseForCausalLM": 3, "Maira2ForConditionalGeneration": 3, "MambaModel": 3, @@ -1203,6 +1198,7 @@ "AceRAG": 2, "AdditionForCausalLM": 2, "AdelicLlamaForCausalLM": 2, + "AffineUUM8DS4ForCausalLM": 2, "AleaMoeForCausalLM": 2, "AlexaLlamaForCausalLM": 2, "AmadablamForCausalLM": 2, @@ -1215,7 +1211,6 @@ "AresMath1.0ForCausalLM": 2, "ArjunaV1ForCausalLM": 2, "ArlowForCausalLM": 2, - "AsterForCausalLM": 2, "AudioQwen2VLForConditionalGeneration": 2, "AuroraGPT2ForCausalLM": 2, "BDH": 2, @@ -1255,6 +1250,7 @@ "CMAForCausalLM": 2, "CPTForConditionalGeneration": 2, "CS336ForCausalLM": 2, + "CascadeForCausalLM": 2, "CheXagentForCausalLM": 2, "CheXagentForConditionalGeneration": 2, "ChemQ3MTPForCausalLM": 2, @@ -1279,6 +1275,7 @@ "DDLlamaForCausalLM": 2, "DNikudModel": 2, "DSHybridForCausalLM": 2, + "DSpineDraftModel": 2, "DashQQwen3_5MoeForCausalLM": 2, "Data2VecTextForCausalLM": 2, "DeCodon": 2, @@ -1581,6 +1578,7 @@ "Qwen3LoopForCausalLM": 2, "Qwen3MoeFusedForCausalLM": 2, "Qwen3MoeModel": 2, + "Qwen3MoePlusPlusForCausalLM": 2, "Qwen3QuaRotForCausalLM": 2, "Qwen3ReasoningForCausalLM": 2, "Qwen3TerminatorForCausalLM": 2, @@ -1761,6 +1759,7 @@ "ArthurFriedenreichForCausalLM": 1, "ArzLMForCausalLM": 1, "AscleLMForCausalLM": 1, + "AsterForCausalLM": 1, "AstralLMForCausalLM": 1, "AtomInima": 1, "AtomK3DSparkModel": 1, @@ -1958,6 +1957,7 @@ "DeepseekOcrForConditionalGeneration": 1, "DeepseekV2MoBEForCausalLM": 1, "DeepseekV2SparseMoBEForCausalLM": 1, + "DeepseekV4FlashSuperExpertForCausalLM": 1, "DeepseekV4ForCausalLMDSpark": 1, "DeepseekVLV2ForCausalLM": 1, "DendroForCausalLM": 1, @@ -2042,7 +2042,6 @@ "FastPlusForCausalLM": 1, "FastyForCausalLM": 1, "FelaForCausalLM": 1, - "Fern3BModel": 1, "FiPhi-NeuralArk-3.9-Ultra": 1, "FiPhi-NeuralMark-V3": 1, "FidelForCausalLM": 1, @@ -2178,6 +2177,7 @@ "HanseForCausalLM": 1, "HanziForCausalLM": 1, "HelionOSCForCausalLM": 1, + "HeliosForCausalLM": 1, "HelixMindForCausalLM": 1, "HelixmRNAForCausalLM": 1, "HelloAgentForCausalLM": 1, @@ -2275,6 +2275,7 @@ "KiyoDiffusion": 1, "KmoshiForConditionalGeneration": 1, "KohakuForCausalLM": 1, + "KonkanGPT": 1, "KopriaLMForCausalLM": 1, "Kosmos2_5TextForCausalLM": 1, "KotodamaForCausalLM": 1, @@ -2448,6 +2449,7 @@ "MiMoV2ASRForCausalLM": 1, "MicroBananaForCausalLM": 1, "MicroLlama": 1, + "MicroLlamaForCausalLM": 1, "MicroStoryBananaMindModel": 1, "MightyLlamaForCausalLM": 1, "MijatovicForCausalLM": 1, @@ -2701,6 +2703,7 @@ "PrismCustomModel": 1, "PrivateLLMForCausalLM": 1, "PrivateWhisperForConditionalGeneration": 1, + "ProjectKairoMiniForCausalLM": 1, "ProximaStarKVLlamaForCausalLM": 1, "PsdelKdaForCausalLM": 1, "PurrenceForCausalLM": 1, @@ -2760,7 +2763,6 @@ "Qwen3MTPForCausalLM": 1, "Qwen3Mamba2ForCausalLM": 1, "Qwen3MoePTEAdapterForCausalLM": 1, - "Qwen3MoePlusPlusForCausalLM": 1, "Qwen3OmniForCausalLM": 1, "Qwen3ScaleSeqForCausalLM": 1, "Qwen3VLModel": 1, @@ -2977,6 +2979,7 @@ "TernaryCoreForCausalLM": 1, "TernaryQwen3_5ForConditionalGeneration": 1, "TernovaHG": 1, + "Test5MForCausalLM": 1, "TextGeneratorC2KForCausalLM": 1, "ThalamusReasoningQwen3_5ForConditionalGeneration": 1, "ThanatosForCausalLM": 1, diff --git a/tools/census.py b/tools/census.py index d51d6ced4..44abcc935 100644 --- a/tools/census.py +++ b/tools/census.py @@ -93,6 +93,10 @@ def checked_on(hf): out.append(b) return out + def a_get(table, key): + a = table.get(key) + return bool(a and a.get("backends")) + with_arch = raw["total"] - raw["no_arch"] mapped = checked = 0 per_backend = {b: {"mapped": 0, "checked": 0} for b in ("hrx", "npu")} @@ -110,8 +114,36 @@ def checked_on(hf): checked += n for b in c: per_backend[b]["checked"] += n + # engine#1929 review: `declined` is a THIRD state, not work remaining. registry/significant.json lists the classes a + # human reviewed and deliberately did not map (`watch: "report"`, the same predicate tools/census_watch.py uses for + # "reviewed"), so a census that counts their models as unmapped reports work that is actually a recorded decision -- + # and one that subtracts them silently reports a percentage over an unwritten denominator. They are counted here, and + # kept out of `top_unmapped`, which is the list of things still to do. + declined_archs = set() + try: + sig = json.load(open(os.path.join(REG, "significant.json")))["classes"] + declined_archs = {hf for hf, e in sig.items() if e.get("watch") == "report" and not (a_get(archs, hf))} + except (OSError, ValueError, KeyError): + declined_archs = set() # no registry, no declines: say zero rather than fail the census + declined = sum(n for hf, n in raw["counts"].items() if hf in declined_archs) + unmapped = [(hf, n) for hf, n in unmapped if hf not in declined_archs] pct = lambda x: round(100.0 * x / with_arch, 2) if with_arch else 0.0 return { + "definitions": { + "mapped": "the pinned backend's code accepts this HF architecture class (registry/architectures.json " + "names a backend for it); it is a property of the ARCHITECTURE, so every model of a mapped " + "class counts as mapped whether or not that model was ever run", + "checked": "the architecture has a PASSING serve_e2e at the pin -- not that this model was run. It is " + "the same per-architecture property, which is why the npu figure is essentially one " + "architecture's model population rather than a count of models anyone exercised", + "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": declined, "declined_architectures": len(declined_archs), "models": raw["total"], "with_architecture": with_arch, "architectures_seen": len(raw["counts"]), "mapped_models": mapped, "mapped_pct": pct(mapped),