Move the MIC coverage page from tarekziade/mic-coverage

README.md

@@ -1,10 +1,11 @@
 ---
-title: Mic Coverage
-emoji: ⚡
+title: MIC coverage
+emoji: 📋
 colorFrom: blue
-colorTo: purple
+colorTo: green
 sdk: static
 pinned: false
+short_description: MIC contract coverage of Transformers model types
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+Sortable list of every Transformers model type: whether it has a MIC contract (Done), can be reached with the generic shrink rule (Reachable), or not (Unreachable, with the sweep's reason), plus Hub downloads.

data.json

@@ -0,0 +1 @@
+{"rows": [["EvollaModel", "Reachable", "", 732, "westlake-repl/Evolla-10B-hf"], ["afmoe", "Reachable", "", 84004, "optimum-intel-internal-testing/tiny-random-trinity"], ["aimv2", "Reachable", "", 10627, "apple/aimv2-large-patch14-224-lit"], ["aimv2_text_model", "Unreachable", "component config only", 0, ""], ["aimv2_vision_model", "Reachable", "", 3412, "apple/aimv2-large-patch14-224"], ["albert", "Done", "", 2009876, "albert/albert-base-v2"], ["align", "Unreachable", "forward fails after shrink", 21429, "kakaobrain/align-base"], ["align_text_model", "Unreachable", "component config only", 0, ""], ["align_vision_model", "Unreachable", "component config only", 0, ""], ["altclip", "Reachable", "", 7652, "BAAI/AltCLIP"], ["altclip_text_model", "Unreachable", "component config only", 0, ""], ["altclip_vision_model", "Unreachable", "component config only", 0, ""], ["apertus", "Done", "", 457103, "swiss-ai/Apertus-8B-Instruct-2509"], ["arcee", "Reachable", "", 64187, "optimum-intel-internal-testing/tiny-random-ArceeForCausalLM"], ["aria", "Unreachable", "forward fails after shrink", 29760, "rhymes-ai/Aria"], ["aria_text", "Unreachable", "forward fails after shrink", 88, "hf-tiny-v2/tiny-random-AriaTextForCausalLM"], ["audio-spectrogram-transformer", "Unreachable", "forward fails after shrink", 1651174, "MIT/ast-finetuned-audioset-10-10-0.4593"], ["audioflamingo3", "Done", "", 144463, "nvidia/audio-flamingo-3-hf"], ["audioflamingo3_encoder", "Unreachable", "forward fails after shrink", 27, "hf-tiny-v2/tiny-random-AudioFlamingo3Encoder"], ["autoformer", "Unreachable", "synthetic inputs rejected", 79649, "huggingface/autoformer-tourism-monthly"], ["axk1", "Reachable", "", 5094, "skt/A.X-K1"], ["axk2", "Reachable", "", 90892, "skt/A.X-K2-NVFP4"], ["aya_vision", "Reachable", "", 3689, "CohereLabs/aya-vision-8b"], ["bailing_hybrid", "Unreachable", "forward fails after shrink", 727122, "rapid-mlx/Ling-3.0-tiny-MLX-4bit"], ["bamba", "Reachable", "", 11728, "ibm-ai-platform/Bamba-9B-v1"], ["bark", "Unreachable", "forward fails after shrink", 28264, "suno/bark"], ["bart", "Done", "", 6328395, "facebook/bart-large-mnli"], ["beam", "Reachable", "", 20667, "reflection/Beam-1.0-501B-A23B"], ["beit", "Done", "", 169127, "microsoft/dit-base-finetuned-rvlcdip"], ["bert", "Done", "", 527874026, "sentence-transformers/all-MiniLM-L6-v2"], ["bert-generation", "Reachable", "", 3386, "ybelkada/random-tiny-BertGenerationModel"], ["big_bird", "Done", "", 194505, "google/bigbird-roberta-base"], ["bigbird_pegasus", "Reachable", "", 71110, "optimum-intel-internal-testing/tiny-random-bigbird_pegasus"], ["biogpt", "Reachable", "", 56866, "optimum-intel-internal-testing/tiny-random-BioGptForCausalLM"], ["bit", "Unreachable", "model build fails", 8676, "google/bit-50"], ["bitnet", "Reachable", "", 35767, "optimum-intel-internal-testing/tiny-random-bitnet"], ["blenderbot", "Done", "", 235225, "optimum-intel-internal-testing/tiny-random-BlenderbotModel"], ["blenderbot-small", "Reachable", "", 5857, "facebook/blenderbot_small-90M"], ["blip", "Unreachable", "forward fails after shrink", 2359364, "Salesforce/blip-image-captioning-base"], ["blip-2", "Unreachable", "forward fails after shrink", 457477, "Salesforce/blip2-opt-2.7b"], ["blip_2_qformer", "Unreachable", "forward fails after shrink", 145, "hf-tiny-v2/tiny-random-Blip2QFormerModel"], ["blip_2_vision_model", "Unreachable", "component config only", 24, "tmnam20/blip2-opt-2.7b-vision"], ["blip_text_model", "Unreachable", "component config only", 0, ""], ["blip_vision_model", "Unreachable", "component config only", 0, ""], ["bloom", "Done", "", 1527823, "bigscience/bloomz-560m"], ["blt", "Unreachable", "forward fails after shrink", 6999, "itazap/blt-1b-hf"], ["blt_global_transformer", "No checkpoint", "no Hub checkpoint has this model_type (documented itazap/blt-1b-hf is another type)", 0, ""], ["blt_local_decoder", "No checkpoint", "no Hub checkpoint has this model_type (documented itazap/blt-1b-hf is another type)", 0, ""], ["blt_local_encoder", "No checkpoint", "no Hub checkpoint has this model_type (documented itazap/blt-1b-hf is another type)", 0, ""], ["blt_patcher", "No checkpoint", "no Hub checkpoint has this model_type (documented itazap/blt-1b-hf is another type)", 0, ""], ["bridgetower", "Reachable", "", 80, "regisss/bridgetower-newyorker-a100-8x"], ["bridgetower_text_model", "No checkpoint", "no Hub checkpoint has this model_type (documented BridgeTower/bridgetower-base is another type)", 0, ""], ["bridgetower_vision_model", "No checkpoint", "no Hub checkpoint has this model_type (documented BridgeTower/bridgetower-base is another type)", 0, ""], ["bros", "Unreachable", "forward fails after shrink", 42661, "naver-clova-ocr/bros-base-uncased"], ["camembert", "Done", "", 2226781, "almanach/camembert-base"], ["canary", "Unreachable", "forward fails after shrink", 237192, "nvidia/canary-1b-v2"], ["canary_decoder", "Unreachable", "component config only", 0, ""], ["canine", "Done", "", 167492, "google/canine-c"], ["chameleon", "Reachable", "", 8504, "facebook/chameleon-7b"], ["chameleon_vqgan", "Unreachable", "component config only", 33, "leloy/Anole-7b-v0.1-vqvae-hf"], ["chinese_clip", "Done", "", 276376, "OFA-Sys/chinese-clip-vit-base-patch16"], ["chinese_clip_text_model", "Unreachable", "component config only", 0, ""], ["chinese_clip_vision_model", "Reachable", "", 57, "hf-tiny-v2/tiny-random-ChineseCLIPVisionModel"], ["chmv2", "Unreachable", "forward fails after shrink", 1526, "facebook/dinov3-vitl16-chmv2-dpt-head"], ["clap", "Done", "", 9104856, "laion/clap-htsat-fused"], ["clap_audio_model", "Unreachable", "component config only", 257, "NGDtuanh/abook-music-student"], ["clap_text_model", "Unreachable", "component config only", 61, "werent4/larger_clap_general_text_model"], ["clip", "Done", "", 37612455, "openai/clip-vit-base-patch32"], ["clip_text_model", "Done", "", 940714, "peft-internal-testing/tiny-clip-text-2"], ["clip_vision_model", "Done", "", 432156, "xingpng/OneIG-StyleEncoder"], ["clipseg", "Done", "", 545844, "CIDAS/clipseg-rd64-refined"], ["clipseg_text_model", "Unreachable", "component config only", 0, ""], ["clipseg_vision_model", "Unreachable", "component config only", 0, ""], ["clvp", "Unreachable", "forward fails after shrink", 5530, "susnato/clvp_dev"], ["clvp_decoder", "Unreachable", "component config only", 0, ""], ["clvp_encoder", "Unreachable", "component config only", 0, ""], ["codegen", "Unreachable", "forward fails after shrink", 168810, "Salesforce/codegen2-16B_P"], ["cohere", "Done", "", 166589, "optimum-intel-internal-testing/tiny-random-CohereForCausalLM"], ["cohere2", "Done", "", 162977, "optimum-intel-internal-testing/tiny-random-aya-base"], ["cohere2_moe", "Reachable", "", 15203, "CohereLabs/North-Mini-Code-1.0"], ["cohere2_vision", "Reachable", "", 37679, "CohereLabs/command-a-plus-05-2026-bf16"], ["cohere_asr", "Unreachable", "forward fails after shrink", 350010, "CohereLabs/cohere-transcribe-arabic-07-2026"], ["cohere_compass", "Reachable", "", 34840, "CohereLabs/North-Micro-Vision-Instruct"], ["cohere_compass_text", "Reachable", "", 41, "hf-tiny-v2/tiny-random-CohereCompassTextForSequenceClassification"], ["cohere_compass_vision", "Unreachable", "forward fails after shrink", 15, "hf-tiny-v2/tiny-random-CohereCompassVisionModel"], ["colmodernvbert", "Unreachable", "no usable checkpoint", 1649, "Qdrant/colmodernvbert"], ["colpali", "Reachable", "", 8568, "vidore/colpali-v1.3-hf"], ["colqwen2", "Reachable", "", 36700, "vidore/colqwen2-v1.0-hf"], ["conditional_detr", "Reachable", "", 62673, "yainage90/fashion-object-detection"], ["convbert", "Done", "", 509373, "YituTech/conv-bert-base"], ["convnext", "Done", "", 192305, "facebook/convnext-tiny-224"], ["convnextv2", "Done", "", 330963, "facebook/convnextv2-base-22k-384"], ["cosmos3_edge", "Reachable", "", 83288, "nvidia/Cosmos3-Edge"], ["cosmos3_edge_text", "Reachable", "", 14, "hf-tiny-v2/tiny-random-Cosmos3EdgeTextModel"], ["cosmos3_edge_vision", "Unreachable", "forward fails after shrink", 12, "hf-tiny-v2/tiny-random-Cosmos3EdgeVisionModel"], ["cosmos3_omni", "Done", "", 244937, "nvidia/Cosmos3-Nano"], ["cpmant", "Reachable", "", 3742, "openbmb/cpm-ant-10b"], ["csm", "Done", "", 145078, "sesame/csm-1b"], ["csm_depth_decoder_model", "Unreachable", "checkpoint gated", 0, "sesame/csm-1b"], ["ctrl", "Reachable", "", 17032, "Salesforce/ctrl"], ["cvt", "Reachable", "", 12319, "microsoft/cvt-13"], ["cwm", "Unreachable", "no usable checkpoint", 1007, "facebook/cwm"], ["d_fine", "Unreachable", "forward fails after shrink", 126298, "ustc-community/dfine-xlarge-obj2coco"], ["dab-detr", "Unreachable", "config load fails", 7050, "IDEA-Research/dab-detr-resnet-50"], ["dac", "Unreachable", "forward fails after shrink", 72000, "descript/dac_44khz"], ["data2vec-audio", "Reachable", "", 39013, "optimum-intel-internal-testing/tiny-random-Data2VecAudioModel"], ["data2vec-text", "Reachable", "", 7219, "optimum-intel-internal-testing/tiny-random-Data2VecTextModel"], ["data2vec-vision", "Unreachable", "shrink fails", 17284, "optimum-intel-internal-testing/tiny-random-Data2VecVisionModel"], ["dbrx", "Reachable", "", 16983, "trl-internal-testing/tiny-DbrxForCausalLM"], ["deberta", "Done", "", 692267, "microsoft/deberta-large-mnli"], ["deberta-v2", "Done", "", 21746039, "microsoft/mdeberta-v3-base"], ["decision_transformer", "Unreachable", "synthetic inputs rejected", 6158, "edbeeching/decision-transformer-gym-hopper-medium"], ["deepseek_ocr2", "Unreachable", "synthetic inputs rejected", 22922, "optimum-intel-internal-testing/tiny-random-deepseek-ocr-2"], ["deepseek_ocr2_encoder", "Unreachable", "component config only", 0, ""], ["deepseek_ocr2_sam_vision_model", "Unreachable", "component config only", 0, ""], ["deepseek_ocr2_text", "Unreachable", "component config only", 0, ""], ["deepseek_ocr2_vision", "Unreachable", "component config only", 0, ""], ["deepseek_v2", "Done", "", 668975, "deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct"], ["deepseek_v3", "Done", "", 4950570, "deepseek-ai/DeepSeek-R1"], ["deepseek_v32", "Done", "", 6161743, "deepseek-ai/DeepSeek-V3.2"], ["deepseek_v4", "Done", "", 6213673, "deepseek-ai/DeepSeek-V4-Flash-0731"], ["deepseek_vl", "Reachable", "", 8828, "deepseek-community/deepseek-vl-1.3b-chat"], ["deepseek_vl_hybrid", "Reachable", "", 8249, "deepseek-community/deepseek-vl-7b-chat"], ["deformable_detr", "Reachable", "", 23013, "SenseTime/deformable-detr"], ["deimv2", "Unreachable", "forward fails after shrink", 4393, "alessioarcara/deimv2-deimv2_dinov3_x_coco"], ["deit", "Reachable", "", 50030, "facebook/deit-base-distilled-patch16-384"], ["depth_anything", "Unreachable", "forward fails after shrink", 2447209, "depth-anything/Depth-Anything-V2-Small-hf"], ["depth_pro", "Unreachable", "forward fails after shrink", 52957, "apple/DepthPro-hf"], ["detr", "Done", "", 385370, "TahaDouaji/detr-doc-table-detection"], ["dia", "Unreachable", "no usable checkpoint", 16396, "nari-labs/Dia-1.6B-0626"], ["dia_decoder", "No checkpoint", "no Hub checkpoint has this model_type (documented nari-labs/Dia-1.6B is another type)", 0, ""], ["dia_encoder", "No checkpoint", "no Hub checkpoint has this model_type (documented nari-labs/Dia-1.6B is another type)", 0, ""], ["diffllama", "Reachable", "", 4593, "kajuma/DiffLlama-0.3B-handcut"], ["diffusion_gemma", "Unreachable", "forward fails after shrink", 1034537, "google/diffusiongemma-26B-A4B-it"], ["diffusion_gemma_text", "Unreachable", "component config only", 0, ""], ["dinat", "Unreachable", "model build fails", 209, "shi-labs/dinat-mini-in1k-224"], ["dinov2", "Done", "", 7235906, "facebook/dinov2-base"], ["dinov2_with_registers", "Done", "", 354429, "facebook/dinov2-with-registers-base"], ["dinov3_convnext", "Reachable", "", 95750, "facebook/dinov3-convnext-tiny-pretrain-lvd1689m"], ["dinov3_vit", "Done", "", 2089303, "facebook/dinov3-vitl16-pretrain-lvd1689m"], ["distilbert", "Done", "", 20324420, "distilbert/distilbert-base-uncased"], ["doge", "Reachable", "", 8362, "SmallDoge/Doge-320M"], ["donut-swin", "Reachable", "", 2963, "optimum-intel-internal-testing/tiny-random-DonutSwinModel"], ["dots1", "Reachable", "", 6731, "rednote-hilab/dots.llm1.base"], ["dpr", "Reachable", "", 53829, "facebook/dpr-question_encoder-single-nq-base"], ["dpt", "Done", "", 620365, "Intel/dpt-hybrid-midas"], ["edgetam", "Unreachable", "no usable checkpoint", 9579, "onnx-community/EdgeTAM-ONNX"], ["edgetam_mask_decoder", "No checkpoint", "no Hub checkpoint has this model_type (documented yonigozlan/EdgeTAM-hf is another type)", 0, ""], ["edgetam_prompt_encoder", "No checkpoint", "no Hub checkpoint has this model_type (documented yonigozlan/EdgeTAM-hf is another type)", 0, ""], ["edgetam_video", "Unreachable", "forward fails after shrink", 11709, "yonigozlan/EdgeTAM-hf"], ["edgetam_video_mask_decoder", "No checkpoint", "no Hub checkpoint has this model_type (documented yonigozlan/EdgeTAM-hf is another type)", 0, ""], ["edgetam_video_prompt_encoder", "No checkpoint", "no Hub checkpoint has this model_type (documented yonigozlan/EdgeTAM-hf is another type)", 0, ""], ["edgetam_vision_model", "Reachable", "", 53, "hf-tiny-v2/tiny-random-EdgeTamVisionModel"], ["efficientloftr", "Unreachable", "shrink fails", 16245, "zju-community/efficientloftr"], ["efficientnet", "Unreachable", "forward fails after shrink", 168152, "docling-project/DocumentFigureClassifier-v2.5"], ["electra", "Done", "", 28465435, "google/electra-base-discriminator"], ["embedding_gemma2", "Done", "", 719174, "onnx-community/embeddinggemma-2-ONNX"], ["embedding_gemma2_text", "Reachable", "", 171, "HukukcuApp/tr-legal-embedding-gemma-2"], ["emu3", "Reachable", "", 20741, "BAAI/Emu3-Chat-hf"], ["emu3_text_model", "Unreachable", "component config only", 118, "hf-tiny-v2/tiny-random-Emu3TextModel"], ["emu3_vqgan", "Unreachable", "component config only", 0, ""], ["encodec", "Unreachable", "forward fails after shrink", 105065, "facebook/encodec_24khz"], ["encoder-decoder", "Done", "", 29649, "optimum-internal-testing/tiny-random-encoder-decoder-gpt2-bert"], ["eomt", "Unreachable", "forward fails after shrink", 67972, "tue-mps/coco_panoptic_eomt_large_640"], ["eomt_dinov3", "Unreachable", "forward fails after shrink", 21640, "tue-mps/eomt-dinov3-ade-semantic-large-512"], ["ernie", "Reachable", "", 73981, "shibing624/text2vec-base-chinese-paraphrase"], ["ernie4_5", "Reachable", "", 14731, "baidu/ERNIE-4.5-0.3B-PT"], ["ernie4_5_moe", "Unreachable", "forward fails after shrink", 123822, "baidu/ERNIE-4.5-21B-A3B-Thinking"], ["ernie4_5_vl_moe", "Unreachable", "forward fails after shrink", 45, "hf-tiny-v2/tiny-random-Ernie4_5_VLMoeForConditionalGeneration"], ["ernie4_5_vl_moe_text", "No checkpoint", "no Hub checkpoint has this model_type (documented baidu/ERNIE-4.5-VL-28B-A3B-PT is another type)", 0, ""], ["ernie4_5_vl_moe_vision", "No checkpoint", "no Hub checkpoint has this model_type (documented baidu/ERNIE-4.5-VL-28B-A3B-PT is another type)", 0, ""], ["esm", "Done", "", 7030784, "facebook/esm2_t33_650M_UR50D"], ["esmc", "Done", "", 393143, "biohub/ESMC-300M"], ["esmfold2", "Unreachable", "shrink fails", 388538, "biohub/ESMFold2"], ["eurobert", "Done", "", 847801, "OpenMed/OpenMed-NER-ChemicalDetect-EuroMed-212M"], ["evolla", "No checkpoint", "no Hub checkpoint has this model_type (none documented)", 0, ""], ["exaone4", "Reachable", "", 64834, "LGAI-EXAONE/EXAONE-4.0-32B"], ["exaone4_5", "Reachable", "", 39320, "LGAI-EXAONE/EXAONE-4.5-33B"], ["exaone4_5_vision", "Unreachable", "forward fails after shrink", 31, "hf-tiny-v2/tiny-random-Exaone4_5_VisionModel"], ["exaone_moe", "Reachable", "", 18548, "LGAI-EXAONE/K-EXAONE-236B-A23B"], ["falcon", "Done", "", 350854, "tiiuae/falcon-7b"], ["falcon_h1", "Done", "", 104762, "tiiuae/Falcon-H1-3B-Instruct"], ["falcon_mamba", "Reachable", "", 61205, "trl-internal-testing/tiny-FalconMambaForCausalLM"], ["fast_vlm", "Reachable", "", 5841, "KamilaMila/FastVLM-7B"], ["fastspeech2_conformer", "Unreachable", "forward fails after shrink", 2880, "espnet/fastspeech2_conformer"], ["fastspeech2_conformer_hifigan", "No checkpoint", "no Hub checkpoint has this model_type (documented espnet/fastspeech2_conformer_with_hifigan is another type)", 0, ""], ["fastspeech2_conformer_with_hifigan", "Unreachable", "forward fails after shrink", 2800, "espnet/fastspeech2_conformer_with_hifigan"], ["flaubert", "Reachable", "", 74651, "flaubert/flaubert_base_uncased"], ["flava", "Reachable", "", 32370, "facebook/flava-full"], ["flava_image_codebook", "Unreachable", "component config only", 23, "facebook/flava-image-codebook"], ["flava_image_model", "Unreachable", "component config only", 0, ""], ["flava_multimodal_model", "Unreachable", "component config only", 0, ""], ["flava_text_model", "Unreachable", "component config only", 0, ""], ["flex_olmo", "Reachable", "", 12810, "allenai/FlexOlmo-7x7B-1T"], ["flint", "Reachable", "", 60545, "RecursiveSI/flint"], ["florence2", "Done", "", 4293112, "microsoft/Florence-2-base"], ["florence_vision", "Unreachable", "component config only", 0, ""], ["fnet", "Done", "", 478381, "google/fnet-base"], ["focalnet", "Reachable", "", 11569, "microsoft/focalnet-tiny"], ["fsmt", "Reachable", "", 57368, "facebook/wmt19-en-de"], ["fun_asr_nano", "Unreachable", "shrink fails", 162557, "FunAudioLLM/Fun-ASR-Nano-2512-hf"], ["fun_asr_nano_encoder", "Unreachable", "forward fails after shrink", 21, "hf-tiny-v2/tiny-random-FunAsrNanoEncoder"], ["funnel", "Reachable", "", 11894, "funnel-transformer/small"], ["fuyu", "Reachable", "", 12098, "adept/fuyu-8b"], ["gemma", "Done", "", 611743, "google/gemma-2b"], ["gemma2", "Done", "", 2138107, "hugging-quants/gemma-2-9b-it-AWQ-INT4"], ["gemma3", "Done", "", 6276380, "google/gemma-3-4b-it"], ["gemma3_text", "Done", "", 3985142, "google/embeddinggemma-300m"], ["gemma3n", "Done", "", 360811, "unsloth/gemma-3n-E4B-it"], ["gemma3n_audio", "Unreachable", "synthetic inputs rejected", 63, "hf-tiny-v2/tiny-random-Gemma3nAudioEncoder"], ["gemma3n_text", "Reachable", "", 2175, "optimum-intel-internal-testing/tiny-random-gemma3n-text"], ["gemma3n_vision", "Reachable", "", 0, ""], ["gemma4", "Done", "", 47591452, "google/gemma-4-31B-it"], ["gemma4_assistant", "Unreachable", "forward fails after shrink", 683069, "google/gemma-4-26B-A4B-it-assistant"], ["gemma4_audio", "Unreachable", "forward fails after shrink", 138, "rnagabh/gemma4-audio-encoder"], ["gemma4_text", "Reachable", "", 41837, "AlexKoff88/gemma4_tinystories"], ["gemma4_unified", "Done", "", 4749964, "google/gemma-4-12B-it"], ["gemma4_unified_assistant", "Unreachable", "forward fails after shrink", 37821, "google/gemma-4-12B-it-assistant"], ["gemma4_unified_audio", "Unreachable", "component config only", 0, ""], ["gemma4_unified_text", "Reachable", "", 2960, "jgeuter/Jeff-1.0-Medium"], ["gemma4_unified_vision", "Unreachable", "component config only", 0, ""], ["gemma4_vision", "Unreachable", "forward fails after shrink", 214, "rnagabh/gemma4-vision-encoder"], ["git", "Reachable", "", 46165, "microsoft/git-base"], ["git_vision_model", "Unreachable", "component config only", 0, ""], ["glm", "Unreachable", "no usable checkpoint", 43219, "optimum-intel-internal-testing/tiny-random-glm-edge"], ["glm4", "Done", "", 139087, "optimum-intel-internal-testing/tiny-random-glm4"], ["glm46v", "Unreachable", "no usable checkpoint", 170, "hf-tiny-v2/tiny-random-Glm46VModel"], ["glm4_moe", "Done", "", 1607130, "zai-org/GLM-4.5-Air"], ["glm4_moe_lite", "Done", "", 1802094, "zai-org/GLM-4.7-Flash"], ["glm4v", "Unreachable", "forward fails after shrink", 277739, "zai-org/GLM-4.1V-9B-Thinking"], ["glm4v_moe", "Unreachable", "forward fails after shrink", 35179, "zai-org/GLM-4.5V"], ["glm4v_moe_text", "Unreachable", "forward fails after shrink", 58, "hf-tiny-v2/tiny-random-Glm4vMoeTextModel"], ["glm4v_moe_vision", "Unreachable", "no usable checkpoint", 38, "hf-tiny-v2/tiny-random-Glm4vMoeVisionModel"], ["glm4v_text", "Unreachable", "forward fails after shrink", 58, "hf-tiny-v2/tiny-random-Glm4vTextModel"], ["glm4v_vision", "Unreachable", "forward fails after shrink", 38, "hf-tiny-v2/tiny-random-Glm4vVisionModel"], ["glm5_next", "Done", "", 7619607, "zai-org/GLM-5.3-Flash"], ["glm5_next_text", "Reachable", "", 23, "hf-tiny-v2/tiny-random-Glm5NextTextModel"], ["glm5_next_vision", "Unreachable", "forward fails after shrink", 65, "immanuelpeter/GLM-5.3-Flash-Vision"], ["glm_image", "Unreachable", "no usable checkpoint", 111, "hf-tiny-v2/tiny-random-GlmImageModel"], ["glm_image_text", "Unreachable", "no usable checkpoint", 55, "hf-tiny-v2/tiny-random-GlmImageTextModel"], ["glm_image_vision", "Unreachable", "no usable checkpoint", 35, "hf-tiny-v2/tiny-random-GlmImageVisionModel"], ["glm_image_vqmodel", "Unreachable", "no usable checkpoint", 6, "hf-tiny-v2/tiny-random-GlmImageVQVAE"], ["glm_moe_dsa", "Done", "", 7133417, "zai-org/GLM-5.3"], ["glm_ocr", "Done", "", 1798225, "zai-org/GLM-OCR"], ["glm_ocr_text", "Reachable", "", 51, "hf-tiny-v2/tiny-random-GlmOcrTextModel"], ["glm_ocr_vision", "Unreachable", "forward fails after shrink", 33, "hf-tiny-v2/tiny-random-GlmOcrVisionModel"], ["glmasr", "Reachable", "", 31331, "zai-org/GLM-ASR-Nano-2512"], ["glmasr_encoder", "Reachable", "", 50, "hf-tiny-v2/tiny-random-GlmAsrEncoder"], ["glmga", "No checkpoint", "no Hub checkpoint has this model_type (documented zai-org/GLM-4.1V-9B-Thinking is another type)", 0, ""], ["glpn", "Reachable", "", 8196, "vinvino02/glpn-kitti"], ["got_ocr2", "Done", "", 56531, "stepfun-ai/GOT-OCR-2.0-hf"], ["got_ocr2_vision", "No checkpoint", "no Hub checkpoint has this model_type (documented facebook/sam-vit-huge is another type)", 0, ""], ["gpt-sw3", "No checkpoint", "no Hub checkpoint has this model_type (none documented)", 0, ""], ["gpt2", "Done", "", 30545353, "openai-community/gpt2"], ["gpt_bigcode", "Done", "", 188077, "bigcode/tiny_starcoder_py"], ["gpt_neo", "Unreachable", "shrink fails", 865506, "EleutherAI/gpt-neo-125m"], ["gpt_neox", "Done", "", 7013890, "EleutherAI/pythia-160m"], ["gpt_neox_japanese", "Reachable", "", 46916, "optimum-intel-internal-testing/tiny-random-GPTNeoXJapaneseForCausalLM"], ["gpt_oss", "Done", "", 10738031, "openai/gpt-oss-20b"], ["gptj", "Unreachable", "forward fails after shrink", 188360, "EleutherAI/gpt-j-6b"], ["gradium_phonon_tts", "Unreachable", "shrink fails", 114118, "gradium/phonon-31m-en"], ["granite", "Done", "", 1699090, "ibm-granite/granite-4.1-3b"], ["granite4_vision", "Done", "", 194823, "ibm-granite/granite-vision-4.1-4b"], ["granite4_vision_text", "No checkpoint", "no Hub checkpoint has this model_type (documented ibm-granite4_vision_text/granite4_vision_text-3.0-8b-base does not exist)", 0, ""], ["granite_speech", "Done", "", 494648, "ibm-granite/granite-4.0-1b-speech"], ["granite_speech5_ctc", "Unreachable", "forward fails after shrink", 99951, "ibm-granite/granite-speech-5.0-470m-turboctc"], ["granite_speech5_encoder", "Unreachable", "forward fails after shrink", 17, "hf-tiny-v2/tiny-random-GraniteSpeech5Encoder"], ["granite_speech_encoder", "Unreachable", "component config only", 0, ""], ["granite_speech_plus", "Unreachable", "shrink fails", 29187, "ibm-granite/granite-speech-4.1-2b-plus"], ["granite_speech_plus_encoder", "Unreachable", "component config only", 0, ""], ["granite_swa", "Reachable", "", 7057, "ibm-granite/granite-swash-2b"], ["granitemoe", "Done", "", 244089, "ibm-granite/granite-3.1-3b-a800m-instruct"], ["granitemoe_swa", "Reachable", "", 6507, "ibm-granite/granite-swash-3b-a600m"], ["granitemoehybrid", "Done", "", 332014, "optimum-intel-internal-testing/tiny-random-granitemoehybrid"], ["granitemoeshared", "Unreachable", "no usable checkpoint", 1648, "ibm-research/moe-7b-1b-active-shared-experts"], ["grounding-dino", "Done", "", 2682931, "IDEA-Research/grounding-dino-base"], ["groupvit", "Reachable", "", 6660, "nvidia/groupvit-gcc-yfcc"], ["groupvit_text_model", "Unreachable", "component config only", 0, ""], ["groupvit_vision_model", "Unreachable", "component config only", 0, ""], ["gte", "Reachable", "", 56943, "Snowflake/snowflake-arctic-embed-m-v2.0"], ["helium", "Reachable", "", 6264, "kyutai/helium-1-preview-2b"], ["hgnet_v2", "Unreachable", "forward fails after shrink", 149, "ustc-community/hgnet-v2"], ["hiera", "Reachable", "", 11635, "facebook/hiera-tiny-224-hf"], ["higgs_audio_v2", "Reachable", "", 21331, "bosonai/higgs-tts-2-3b-base"], ["higgs_audio_v2_tokenizer", "Unreachable", "model build fails", 235573, "eustlb/higgs-audio-v2-tokenizer"], ["hrm_text", "Unreachable", "forward fails after shrink", 14676, "sapientinc/HRM-Text-1B"], ["hubert", "Done", "", 1327557, "facebook/hubert-large-ls960-ft"], ["hunyuan_v1_dense", "Done", "", 197468, "optimum-intel-internal-testing/tiny-random-hunyuan-v1-dense"], ["hunyuan_v1_moe", "Unreachable", "forward fails after shrink", 64830, "tencent/Hunyuan-A13B-Instruct"], ["hunyuan_vl", "Done", "", 807278, "tencent/HunyuanOCR"], ["hunyuan_vl_text", "Unreachable", "no usable checkpoint", 9, "hf-tiny-v2/tiny-random-HunYuanVLTextModel"], ["hunyuan_vl_vision", "Unreachable", "no usable checkpoint", 10, "hf-tiny-v2/tiny-random-HunYuanVLVisionTransformer"], ["hy_v3", "Done", "", 1408804, "tencent/Hy3"], ["hy_v4", "Reachable", "", 32339, "tencent/Hy4-preview-FP8"], ["hyperclovax", "Reachable", "", 20512, "naver-hyperclovax/HyperCLOVAX-SEED-Think-14B"], ["hyperclovax_vision_v2", "Reachable", "", 10279, "naver-hyperclovax/HyperCLOVAX-SEED-Think-32B"], ["ibert", "Reachable", "", 40201, "optimum-intel-internal-testing/tiny-random-ibert"], ["idefics", "Unreachable", "forward fails after shrink", 21036, "HuggingFaceM4/idefics-80b-instruct"], ["idefics2", "Reachable", "", 18780, "HuggingFaceM4/idefics2-8b"], ["idefics2_perceiver", "No checkpoint", "no Hub checkpoint has this model_type (documented HuggingFaceM4/idefics2-8b is another type)", 0, ""], ["idefics2_vision", "No checkpoint", "no Hub checkpoint has this model_type (documented HuggingFaceM4/idefics2-8b is another type)", 0, ""], ["idefics3", "Done", "", 776490, "ibm-granite/granite-docling-258M"], ["idefics3_vision", "Unreachable", "no usable checkpoint", 53, "hf-tiny-v2/tiny-random-Idefics3VisionTransformer"], ["idefics_perciever", "No checkpoint", "no Hub checkpoint has this model_type (documented HuggingFaceM4/idefics-9b is another type)", 0, ""], ["idefics_vision", "No checkpoint", "no Hub checkpoint has this model_type (documented HuggingFaceM4/idefics-9b is another type)", 0, ""], ["ijepa", "Reachable", "", 23643, "facebook/ijepa_vith14_1k"], ["imagegpt", "Reachable", "", 7911, "openai/imagegpt-small"], ["informer", "Unreachable", "synthetic inputs rejected", 30807, "huggingface/informer-tourism-monthly"], ["inkling_audio", "No checkpoint", "no Hub checkpoint has this model_type (documented thinkingmachines/Inkling is another type)", 0, ""], ["inkling_mm_model", "Done", "", 3938529, "thinkingmachines/Inkling"], ["inkling_text", "Unreachable", "no usable checkpoint", 33, "hf-tiny-v2/tiny-random-InklingTextModel"], ["inkling_vision", "No checkpoint", "no Hub checkpoint has this model_type (documented thinkingmachines/Inkling is another type)", 0, ""], ["instructblip", "Unreachable", "forward fails after shrink", 28294, "Salesforce/instructblip-flan-t5-xl"], ["instructblip_qformer", "Unreachable", "component config only", 0, ""], ["instructblip_vision_model", "Unreachable", "component config only", 0, ""], ["instructblipvideo", "Unreachable", "no usable checkpoint", 64, "hf-tiny-v2/tiny-random-InstructBlipVideoForConditionalGeneration"], ["instructblipvideo_qformer", "No checkpoint", "no Hub checkpoint has this model_type (documented Salesforce/instructblip-flan-t5-xl is another type)", 0, ""], ["instructblipvideo_vision_model", "No checkpoint", "no Hub checkpoint has this model_type (documented Salesforce/instructblip-flan-t5-xl is another type)", 0, ""], ["internvl", "Done", "", 873854, "OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview-HF"], ["internvl_vision", "Reachable", "", 105, "hf-tiny-v2/tiny-random-InternVLVisionModel"], ["jais2", "Reachable", "", 5408, "inception42/Jais-2-8B-Chat"], ["jamba", "Done", "", 210865, "ai21labs/Jamba-tiny-dev"], ["janus", "Done", "", 9275, "deepseek-community/Janus-Pro-1B"], ["janus_vision_model", "Unreachable", "component config only", 0, ""], ["janus_vqgan", "Unreachable", "component config only", 0, ""], ["jetmoe", "Reachable", "", 5517, "jetmoe/jetmoe-8b"], ["jina_embeddings_v3", "Done", "", 245085, "jinaai/jina-embeddings-v3-hf"], ["kimi_k25", "Reachable", "", 1381590, "moonshotai/Kimi-K2.5"], ["kimi_k25_vision", "Unreachable", "no usable checkpoint", 26, "Aquiles-ai/MoonViT-3D"], ["kimi_linear", "Unreachable", "shrink fails", 214668, "moonshotai/Kimi-Linear-48B-A3B-Instruct"], ["kolibri1", "Reachable", "", 52602, "Aleph-Alpha/Kolibri-1"], ["kosmos-2", "Unreachable", "forward fails after shrink", 24555, "microsoft/kosmos-2-patch14-224"], ["kosmos-2.5", "Reachable", "", 7720, "microsoft/kosmos-2.5"], ["kosmos_2_5_text_model", "Unreachable", "component config only", 14, "Fireblossom/kosmos-2.5-4bit-text"], ["kosmos_2_5_vision_model", "Unreachable", "component config only", 15, "Fireblossom/kosmos-2.5-4bit-vision"], ["kosmos_2_text_model", "Unreachable", "component config only", 0, ""], ["kosmos_2_vision_model", "Unreachable", "component config only", 0, ""], ["kyutai_speech_to_text", "Unreachable", "forward fails after shrink", 8225, "kyutai/stt-2.6b-en-trfs"], ["laguna", "Done", "", 3313579, "poolside/Laguna-XS.2"], ["lasr_ctc", "Unreachable", "forward fails after shrink", 13038, "google/medasr"], ["lasr_encoder", "Unreachable", "no usable checkpoint", 40, "hf-tiny-v2/tiny-random-LasrEncoder"], ["layoutlm", "Done", "", 518133, "impira/layoutlm-document-qa"], ["layoutlmv2", "Unreachable", "model build fails", 440458, "microsoft/layoutxlm-base"], ["layoutlmv3", "Unreachable", "forward fails after shrink", 970269, "microsoft/layoutlmv3-base"], ["layoutxlm", "No checkpoint", "no Hub checkpoint has this model_type (documented microsoft/layoutxlm-base is another type)", 0, ""], ["led", "Unreachable", "model build fails", 33694, "allenai/led-base-16384"], ["levit", "Reachable", "", 19387, "optimum-intel-internal-testing/tiny-random-LevitModel"], ["lfm2", "Done", "", 1255733, "LiquidAI/LFM2.5-230M-Base"], ["lfm2_moe", "Done", "", 688657, "lmstudio-community/LFM2-24B-A2B-MLX-4bit"], ["lfm2_vl", "Done", "", 199290, "LiquidAI/LFM2.5-VL-450M"], ["lightglue", "Unreachable", "forward fails after shrink", 22972, "ETH-CVG/lightglue_superpoint"], ["lighton_ocr", "Reachable", "", 606, "onnx-community/LightOnOCR-2-1B-ONNX"], ["lilt", "Unreachable", "forward fails after shrink", 81714, "SCUT-DLVCLab/lilt-roberta-en-base"], ["llama", "Done", "", 50186457, "meta-llama/Llama-3.1-8B-Instruct"], ["llama4", "Done", "", 373138, "meta-llama/Llama-4-Scout-17B-16E-Instruct"], ["llama4_text", "Reachable", "", 28667, "trl-internal-testing/tiny-Llama4ForCausalLM"], ["llama4_vision_model", "Unreachable", "checkpoint gated", 0, "meta-llama/Llama-4-Scout-17B-16E"], ["llava", "Done", "", 1766490, "llava-hf/llava-1.5-7b-hf"], ["llava_next", "Done", "", 667783, "llava-hf/llava-v1.6-mistral-7b-hf"], ["llava_next_video", "Done", "", 53334, "llava-hf/LLaVA-NeXT-Video-7B-hf"], ["llava_onevision", "Done", "", 262658, "llava-hf/llava-onevision-qwen2-0.5b-ov-hf"], ["longcat_flash", "Unreachable", "forward fails after shrink", 1566, "wangbo981016/longcat2.0-nvfp4"], ["longformer", "Done", "", 1478599, "allenai/longformer-base-4096"], ["longt5", "Done", "", 243258, "pszemraj/long-t5-tglobal-base-sci-simplify"], ["luke", "Done", "", 321239, "studio-ousia/luke-large-finetuned-tacred"], ["lw_detr", "Reachable", "", 13080, "AnnaZhang/lwdetr_small_60e_coco"], ["lw_detr_vit", "Unreachable", "backward fails", 57, "hf-tiny-v2/tiny-random-LwDetrViTBackbone"], ["lxmert", "Unreachable", "forward fails after shrink", 4977, "unc-nlp/lxmert-base-uncased"], ["m2m_100", "Done", "", 2425200, "facebook/nllb-200-distilled-600M"], ["mamba", "Done", "", 138353, "state-spaces/mamba-130m-hf"], ["mamba2", "Unreachable", "shrink fails", 1401, "limloop/whiff-mamba2-50M-v0.1"], ["marian", "Done", "", 5061374, "Helsinki-NLP/opus-mt-nl-en"], ["markuplm", "Reachable", "", 10766, "microsoft/markuplm-base"], ["mask2former", "Done", "", 1060987, "facebook/mask2former-swin-large-ade-semantic"], ["maskformer", "Reachable", "", 16159, "facebook/maskformer-swin-base-ade"], ["maskformer-swin", "No checkpoint", "no Hub checkpoint has this model_type (documented microsoft/swin-tiny-patch4-window7-224 is another type)", 0, ""], ["mbart", "Done", "", 579384, "ai4bharat/IndicBART"], ["megatron-bert", "Reachable", "", 21960, "KBLab/robust-swedish-sentiment-multiclass"], ["mellum", "Reachable", "", 51107, "JetBrains/Mellum2-12B-A2.5B-Instruct"], ["metaclip_2", "Reachable", "", 92623, "facebook/metaclip-2-worldwide-b16"], ["metaclip_2_text_model", "Unreachable", "component config only", 0, ""], ["metaclip_2_vision_model", "Unreachable", "component config only", 0, ""], ["mgp-str", "Unreachable", "forward fails after shrink", 13719, "alibaba-damo/mgp-str-base"], ["mimi", "Unreachable", "forward fails after shrink", 184266, "kyutai/mimi"], ["mimi_decoder", "No checkpoint", "no Hub checkpoint has this model_type (none documented)", 0, ""], ["mimi_encoder", "No checkpoint", "no Hub checkpoint has this model_type (none documented)", 0, ""], ["mimo_v2_flash", "Reachable", "", 49711, "XiaomiMiMo/MiMo-V2-Flash"], ["minicpm3", "Done", "", 17559, "openbmb/MiniCPM3-4B"], ["minicpm_ocr", "Reachable", "", 2508, "openbmb/MiniCPM-OCR-1.0"], ["minicpm_ocr_vision", "Unreachable", "forward fails after shrink", 0, ""], ["minicpmv4_6", "Done", "", 372489, "openbmb/MiniCPM-V-4.6"], ["minicpmv4_6_vision", "Unreachable", "component config only", 0, ""], ["minicpmv4_7", "Unreachable", "no usable checkpoint", 775231, "openbmb/MiniCPM-V-4.7-35B-A3B"], ["minicpmv4_7_vision", "Unreachable", "no usable checkpoint", 28, "hf-tiny-v2/tiny-random-MiniCPMV4_7VisionModel"], ["minimax", "Reachable", "", 19547, "MiniMaxAI/MiniMax-Text-01-hf"], ["minimax_m2", "Done", "", 1886356, "MiniMaxAI/MiniMax-M2.5"], ["minimax_m3_vl", "Done", "", 384244, "MiniMaxAI/MiniMax-M3"], ["minimax_m3_vl_text", "Reachable", "", 314, "ProprietaryLegal/minimax-m3-legal-reap-0.25"], ["minimax_m3_vl_vision", "Unreachable", "component config only", 30, "immanuelpeter/MiniMax-M3-Vision"], ["ministral", "Unreachable", "no usable checkpoint", 1857, "saidutta69/Ministral-8B-Instruct-heretic"], ["ministral3", "Done", "", 500650, "nvidia/Nemotron-3-Embed-1B-BF16"], ["mistral", "Done", "", 10522940, "mistralai/Mistral-7B-Instruct-v0.2"], ["mistral3", "Done", "", 2910198, "lightonai/LightOnOCR-2-1B"], ["mistral4", "Reachable", "", 971, "onnx-internal-testing/tiny-random-Mistral4ForCausalLM"], ["mixtral", "Done", "", 1221154, "mistralai/Mixtral-8x7B-Instruct-v0.1"], ["mlcd", "Reachable", "", 6623, "DeepGlint-AI/mlcd-vit-bigG-patch14-336"], ["mlcd_vision_model", "Unreachable", "no usable checkpoint", 107, "hf-tiny-v2/tiny-random-MLCDVisionModel"], ["mllama", "Done", "", 128886, "meta-llama/Llama-3.2-11B-Vision-Instruct"], ["mllama_text_model", "Unreachable", "component config only", 637, "hf-tiny-v2/tiny-random-MllamaTextModel"], ["mllama_vision_model", "Unreachable", "component config only", 65, "hf-tiny-v2/tiny-random-MllamaVisionModel"], ["mm-grounding-dino", "Done", "", 77467, "openmmlab-community/mm_grounding_dino_large_all"], ["mobilebert", "Unreachable", "model build fails", 378729, "google/mobilebert-uncased"], ["mobilenet_v1", "Reachable", "", 22769, "optimum-intel-internal-testing/mobilenet_v1_0.75_192"], ["mobilenet_v2", "Done", "", 116493, "google/mobilenet_v2_1.0_224"], ["mobilevit", "Done", "", 62083, "apple/mobilevit-xx-small"], ["mobilevitv2", "Reachable", "", 4054, "apple/mobilevitv2-1.0-imagenet1k-256"], ["modernbert", "Done", "", 22009085, "ibm-granite/granite-embedding-small-english-r2"], ["modernbert-decoder", "Reachable", "", 13311, "blab-jhu/test-32m-dec"], ["modernvbert", "Reachable", "", 477, "naver/v-splade-quality"], ["molmo2", "Unreachable", "shrink fails", 124222, "allenai/Molmo2-8B"], ["molmo2_adapter", "No checkpoint", "no Hub checkpoint has this model_type (documented allenai/Molmo2-8B is another type)", 0, ""], ["molmo2_text", "Reachable", "", 0, ""], ["molmo2_vision", "No checkpoint", "no Hub checkpoint has this model_type (documented allenai/Molmo2-8B is another type)", 0, ""], ["moonshine", "Unreachable", "forward fails after shrink", 120877, "moonshine-ai/moonshine-tiny"], ["moonshine_streaming", "Unreachable", "forward fails after shrink", 167472, "moonshine-ai/moonshine-streaming-tiny"], ["moonshine_streaming_encoder", "Unreachable", "component config only", 0, ""], ["moshi", "Reachable", "", 8294, "kyutai/personaplex-rl-seamless"], ["moshi_depth", "Unreachable", "component config only", 0, ""], ["moss_transcribe_diarize", "Done", "", 139891, "OpenMOSS-Team/MOSS-Transcribe-Diarize"], ["moss_transcribe_diarize_audio", "Reachable", "", 0, ""], ["mpnet", "Done", "", 23509384, "sentence-transformers/all-mpnet-base-v2"], ["mpt", "Reachable", "", 84471, "echarlaix/tiny-mpt-random-remote-code"], ["mra", "Reachable", "", 3305, "uw-madison/mra-base-512-4"], ["mt5", "Done", "", 725133, "google/mt5-small"], ["muse_glimmer", "Done", "", 688700, "RadixArk/Muse-Glimmer-NVFP4"], ["muse_glimmer_assistant", "Unreachable", "synthetic inputs rejected", 220523, "meta-models/Muse-Glimmer-30B-assistant"], ["muse_glimmer_text", "Reachable", "", 210, "OpensourceWTF/Muse-Glimmer-30B-MLX-q4"], ["muse_glimmer_vision", "Unreachable", "forward fails after shrink", 54, "immanuelpeter/Muse-Glimmer-Vision"], ["muse_spark", "Reachable", "", 48269, "meta-models/Muse-Spark-Open-816B-A42B-NVFP4"], ["muse_spark_audio", "Unreachable", "checkpoint gated", 0, "meta-models/Muse-Spark-Open-816B-A42B"], ["muse_spark_text", "Unreachable", "checkpoint gated", 0, "meta-models/Muse-Spark-Open-816B-A42B"], ["muse_spark_vision", "Unreachable", "checkpoint gated", 0, "meta-models/Muse-Spark-Open-816B-A42B"], ["musicflamingo", "Reachable", "", 18191, "nvidia/music-flamingo-2601-hf"], ["musicgen", "Unreachable", "model build fails", 2150786, "facebook/musicgen-medium"], ["musicgen_decoder", "Unreachable", "component config only", 13, "hf-tiny-v2/tiny-random-MusicgenForCausalLM"], ["musicgen_melody", "Unreachable", "model build fails", 12068, "facebook/musicgen-melody"], ["musicgen_melody_decoder", "Unreachable", "component config only", 39, "ylacombe/musicgen-melody-bella-ciao"], ["mvp", "Reachable", "", 4701, "RUCAIBox/mvp"], ["nanochat", "Reachable", "", 5316, "l0ulan/chinese-babylm-nanochat-d22-step10000"], ["nemo_speechlm", "Unreachable", "shrink fails", 5799, "nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B"], ["nemotron", "Reachable", "", 48968, "nvidia/Nemotron-Mini-4B-Instruct"], ["nemotron3_5_asr", "Unreachable", "forward fails after shrink", 842520, "nvidia/nemotron-3.5-asr-streaming-0.6b"], ["nemotron3_5_transcribe_asr_audio", "Unreachable", "forward fails after shrink", 0, ""], ["nemotron3_5_transcribe_audio", "Unreachable", "shrink fails", 0, ""], ["nemotron3_5_transcribe_word_timestamp", "Unreachable", "synthetic inputs rejected", 0, ""], ["nemotron3_diarization", "Unreachable", "shrink fails", 146475, "nvidia/Nemotron-3-Diarization"], ["nemotron3_diarization_audio", "Unreachable", "forward fails after shrink", 34, "hf-tiny-v2/tiny-random-Nemotron3DiarizationAudioModel"], ["nemotron_asr_streaming", "Unreachable", "forward fails after shrink", 103140, "nvidia/nemotron-speech-streaming-en-0.6b"], ["nemotron_asr_streaming_encoder", "Unreachable", "forward fails after shrink", 6, "hf-tiny-v2/tiny-random-NemotronAsrStreamingEncoder"], ["nemotron_h", "Done", "", 12146358, "nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16"], ["nemotron_h_omni", "Unreachable", "no usable checkpoint", 113462, "nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16"], ["neomme", "Reachable", "", 6699, "Hcompany/NeoMME-260M"], ["neucodec", "Unreachable", "forward fails after shrink", 48553, "neuphonic/neucodec"], ["nllb-moe", "Reachable", "", 3303, "facebook/nllb-moe-54b"], ["nomic_bert", "Done", "", 14371774, "nomic-ai/nomic-embed-text-v1.5"], ["nougat", "Unreachable", "config load fails", 8, "zphilip48/nougat-middle-cn"], ["nystromformer", "Reachable", "", 5529, "uw-madison/nystromformer-512"], ["olmo", "Done", "", 125508, "optimum-intel-internal-testing/tiny-random-olmo-hf"], ["olmo2", "Done", "", 442211, "allenai/OLMo-2-0425-1B"], ["olmo3", "Done", "", 1055647, "allenai/Olmo-3-7B-Think"], ["olmo_hybrid", "Reachable", "", 19286, "allenai/Olmo-Hybrid-7B"], ["olmoe", "Done", "", 132339, "allenai/OLMoE-1B-7B-0924"], ["omdet-turbo", "Unreachable", "config load fails", 35542, "omlab/omdet-turbo-swin-tiny-hf"], ["oneformer", "Unreachable", "synthetic inputs rejected", 390959, "shi-labs/oneformer_cityscapes_swin_large"], ["openai-gpt", "Reachable", "", 85250, "openai-community/openai-gpt"], ["openai_privacy_filter", "Done", "", 634682, "OpenMed/privacy-filter-nemotron-v2"], ["opt", "Done", "", 5616003, "facebook/opt-125m"], ["ovis2", "Reachable", "", 6898, "thisisiron/Ovis2-1B-hf"], ["ovis2_vision", "No checkpoint", "no Hub checkpoint has this model_type (documented thisisiron/Ovis2-1B-hf is another type)", 0, ""], ["owlv2", "Done", "", 1501028, "google/owlv2-base-patch16-ensemble"], ["owlv2_text_model", "No checkpoint", "no Hub checkpoint has this model_type (none documented)", 0, ""], ["owlv2_vision_model", "No checkpoint", "no Hub checkpoint has this model_type (none documented)", 0, ""], ["owlvit", "Done", "", 220839, "google/owlvit-base-patch32"], ["owlvit_text_model", "Unreachable", "component config only", 0, ""], ["owlvit_vision_model", "Unreachable", "component config only", 0, ""], ["paddleocr_vl", "Done", "", 227189, "BDRC/tibetan-ocr"], ["paddleocr_vl_text", "No checkpoint", "no Hub checkpoint has this model_type (documented PaddlePaddle/PaddleOCR-VL is another type)", 0, ""], ["paddleocr_vl_vision", "No checkpoint", "no Hub checkpoint has this model_type (documented PaddlePaddle/PaddleOCR-VL is another type)", 0, ""], ["paligemma", "Done", "", 490686, "google/paligemma-3b-pt-224"], ["parakeet_ctc", "Unreachable", "forward fails after shrink", 105992, "nvidia/parakeet-ctc-1.1b"], ["parakeet_encoder", "Unreachable", "forward fails after shrink", 125, "nithinraok/parakeet-tdt-v3-encoder"], ["parakeet_rnnt", "Unreachable", "forward fails after shrink", 40081, "nvidia/parakeet-rnnt-0.6b"], ["parakeet_tdt", "Unreachable", "forward fails after shrink", 694911, "nvidia/parakeet-tdt-0.6b-v3"], ["patchtsmixer", "Unreachable", "synthetic inputs rejected", 15970, "ibm-granite/granite-timeseries-patchtsmixer"], ["patchtst", "Unreachable", "synthetic inputs rejected", 28739, "GilpinLab/panda"], ["pe_audio", "Unreachable", "no usable checkpoint", 89572, "facebook/pe-a-frame-large"], ["pe_audio_encoder", "Unreachable", "forward fails after shrink", 31, "hf-tiny-v2/tiny-random-PeAudioEncoder"], ["pe_audio_video", "Unreachable", "forward fails after shrink", 21786, "facebook/pe-av-large"], ["pe_audio_video_encoder", "Unreachable", "forward fails after shrink", 21, "hf-tiny-v2/tiny-random-PeAudioVideoEncoder"], ["pe_video", "Unreachable", "no usable checkpoint", 25, "hf-tiny-v2/tiny-random-PeVideoModel"], ["pe_video_encoder", "Unreachable", "synthetic inputs rejected", 25, "hf-tiny-v2/tiny-random-PeVideoEncoder"], ["pegasus", "Done", "", 120276, "optimum-intel-internal-testing/tiny-random-pegasus"], ["pegasus_x", "Reachable", "", 6499, "google/pegasus-x-large"], ["perceiver", "Done", "", 41482, "optimum-intel-internal-testing/tiny-random-language_perceiver"], ["perception_lm", "Reachable", "", 1028, "facebook/Perception-LM-8B"], ["persimmon", "Reachable", "", 50277, "optimum-intel-internal-testing/tiny-random-PersimmonForCausalLM"], ["phi", "Done", "", 575882, "microsoft/phi-2"], ["phi3", "Done", "", 2440907, "microsoft/Phi-3.5-mini-instruct"], ["phi4_multimodal", "Done", "", 549, "huggingface-mic/Phi-4-multimodal-instruct-mic"], ["phi4_multimodal_audio", "No checkpoint", "no Hub checkpoint has this model_type (documented microsoft/Phi-4-multimodal-instruct is another type)", 0, ""], ["phi4_multimodal_vision", "No checkpoint", "no Hub checkpoint has this model_type (documented microsoft/Phi-4-multimodal-instruct is another type)", 0, ""], ["phimoe", "Done", "", 169903, "optimum-intel-internal-testing/phi-3.5-moe-tiny-random"], ["pi0", "Unreachable", "synthetic inputs rejected", 181, "yinchenghust/openpi_libero_all"], ["pix2struct", "Unreachable", "synthetic inputs rejected", 116238, "optimum-intel-internal-testing/pix2struct-tiny-random"], ["pix2struct_text_model", "Unreachable", "component config only", 0, ""], ["pix2struct_vision_model", "Unreachable", "component config only", 22, "UiPath/pix2struct-vision-base"], ["pixio", "Reachable", "", 163, "LiheYoung/pixio-vith16"], ["pixtral", "Reachable", "", 940, "mlx-community/pixtral-12b-4bit"], ["plbart", "Reachable", "", 11136, "uclanlp/plbart-base"], ["pocket_tts", "Unreachable", "shrink fails", 2949, "mlx-community/pocket-tts"], ["pocket_tts_audio_tokenizer", "Unreachable", "forward fails after shrink", 471, "gradium/phonon-mimi"], ["pocket_tts_audio_tokenizer_decoder", "No checkpoint", "no Hub checkpoint has this model_type (none documented)", 0, ""], ["pocket_tts_audio_tokenizer_encoder", "No checkpoint", "no Hub checkpoint has this model_type (none documented)", 0, ""], ["poolformer", "Reachable", "", 22043, "optimum-intel-internal-testing/tiny-random-PoolFormerModel"], ["pop2piano", "Reachable", "", 105, "hf-tiny-v2/tiny-random-Pop2PianoForConditionalGeneration"], ["pp_chart2table", "Reachable", "", 272, "PaddlePaddle/PP-Chart2Table_safetensors"], ["pp_chart2table_vision", "No checkpoint", "no Hub checkpoint has this model_type (documented facebook/sam-vit-huge is another type)", 0, ""], ["pp_doclayout_v2", "Unreachable", "forward fails after shrink", 585, "PaddlePaddle/PP-DocLayoutV2_safetensors"], ["pp_doclayout_v3", "Unreachable", "forward fails after shrink", 932166, "PaddlePaddle/PP-DocLayoutV3_safetensors"], ["pp_doclayout_v4", "No checkpoint", "no Hub checkpoint has this model_type (documented PaddlePaddle/PP-DocLayoutV4_safetensors does not exist)", 0, ""], ["pp_formulanet", "Unreachable", "forward fails after shrink", 3829, "PaddlePaddle/PP-FormulaNet_plus-L_safetensors"], ["pp_formulanet_text", "No checkpoint", "no Hub checkpoint has this model_type (documented PaddlePaddle/PP-FormulaNet_plus-L_safetensors is another type)", 0, ""], ["pp_formulanet_vision", "No checkpoint", "no Hub checkpoint has this model_type (documented PaddlePaddle/PP-FormulaNet_plus-L_safetensors is another type)", 0, ""], ["pp_lcnet", "Reachable", "", 865, "PaddlePaddle/PP-LCNet_x1_0_doc_ori_safetensors"], ["pp_lcnet_v3", "Unreachable", "no usable checkpoint", 112, "hf-tiny-v2/tiny-random-PPLCNetV3Backbone"], ["pp_lcnet_v4", "Unreachable", "no usable checkpoint", 21, "hf-tiny-v2/tiny-random-PPLCNetV4Backbone"], ["pp_ocrv5_mobile_det", "Reachable", "", 530, "PaddlePaddle/PP-OCRv5_mobile_det_safetensors"], ["pp_ocrv5_mobile_rec", "Reachable", "", 3902, "PaddlePaddle/latin_PP-OCRv5_mobile_rec_safetensors"], ["pp_ocrv5_server_det", "Reachable", "", 509, "PaddlePaddle/PP-OCRv5_server_det_safetensors"], ["pp_ocrv5_server_rec", "Reachable", "", 585, "PaddlePaddle/PP-OCRv5_server_rec_safetensors"], ["pp_ocrv6_medium_det", "Reachable", "", 2694, "PaddlePaddle/PP-OCRv6_medium_det_safetensors"], ["pp_ocrv6_small_det", "Reachable", "", 1617, "PaddlePaddle/PP-OCRv6_small_det_safetensors"], ["pp_ocrv6_small_rec", "Reachable", "", 3559, "PaddlePaddle/PP-OCRv6_medium_rec_safetensors"], ["pp_ocrv6_tiny_rec", "Reachable", "", 693, "PaddlePaddle/PP-OCRv6_tiny_rec_safetensors"], ["prompt_depth_anything", "Unreachable", "forward fails after shrink", 1530, "depth-anything/prompt-depth-anything-vits-hf"], ["prophetnet", "Reachable", "", 7258, "microsoft/prophetnet-large-uncased"], ["pvt", "Done", "", 8712, "Xrenya/pvt-tiny-224"], ["pvt_v2", "Reachable", "", 9824, "OpenGVLab/pvt_v2_b0"], ["qianfan_ocr", "Reachable", "", 38183, "baidu/Qianfan-OCR"], ["qianfan_ocr_vision", "Reachable", "", 77, "hf-tiny-v2/tiny-random-QianfanOCRVisionModel"], ["qwen2", "Done", "", 52739906, "Qwen/Qwen2.5-7B-Instruct"], ["qwen2_5_omni", "Done", "", 456813, "Qwen/Qwen2.5-Omni-7B"], ["qwen2_5_omni_audio_encoder", "Unreachable", "component config only", 23, "FlexiSLM/Qwen2_5-Omni-Audio_Encoder"], ["qwen2_5_omni_bigvgan", "Unreachable", "component config only", 24, "Ryan241005/scail-2"], ["qwen2_5_omni_dit", "Unreachable", "component config only", 0, ""], ["qwen2_5_omni_talker", "Unreachable", "component config only", 0, ""], ["qwen2_5_omni_text", "Unreachable", "component config only", 59, "AlexHung29629/Qwen-2.5-Omni-ThinkerTextModel"], ["qwen2_5_omni_thinker", "Done", "", 128795, "laion/voiceclap-large-v2"], ["qwen2_5_omni_token2wav", "Unreachable", "component config only", 0, ""], ["qwen2_5_omni_vision_encoder", "Unreachable", "component config only", 0, ""], ["qwen2_5_vl", "Done", "", 15195814, "Qwen/Qwen2.5-VL-7B-Instruct"], ["qwen2_5_vl_text", "Reachable", "", 83, "hf-tiny-v2/tiny-random-Qwen2_5_VLTextModel"], ["qwen2_5_vl_vision", "Unreachable", "no usable checkpoint", 22, "hf-tiny-v2/tiny-random-Qwen2_5_VisionTransformerPretrainedModel"], ["qwen2_audio", "Done", "", 199606, "Qwen/Qwen2-Audio-7B-Instruct"], ["qwen2_audio_encoder", "Reachable", "", 200, "Atotti/qwen2-audio-encoder"], ["qwen2_moe", "Done", "", 194932, "Qwen/Qwen1.5-MoE-A2.7B"], ["qwen2_vl", "Done", "", 3757282, "Qwen/Qwen2-VL-7B-Instruct-AWQ"], ["qwen2_vl_text", "Reachable", "", 72, "hf-tiny-v2/tiny-random-Qwen2VLTextModel"], ["qwen2_vl_vision", "No checkpoint", "no Hub checkpoint has this model_type (documented Qwen/Qwen2-VL-7B-Instruct is another type)", 0, ""], ["qwen3", "Done", "", 108746675, "Qwen/Qwen3-0.6B"], ["qwen3_5", "Done", "", 92274452, "Qwen/Qwen3.8-27B"], ["qwen3_5_moe", "Done", "", 34351908, "Qwen/Qwen3.6-35B-A3B"], ["qwen3_5_moe_text", "Done", "", 717113, "Bahushruth/Qwen3.6-35B-A3B-abliterated-v4"], ["qwen3_5_moe_vision", "Unreachable", "no usable checkpoint", 16, "hf-tiny-v2/tiny-random-Qwen3_5MoeVisionModel"], ["qwen3_5_text", "Done", "", 2017293, "autotrust/JEV-9B"], ["qwen3_5_vision", "Unreachable", "forward fails after shrink", 135, "immanuelpeter/Qwen3.8-27B-Vision"], ["qwen3_asr", "Done", "", 4105092, "Qwen/Qwen3-ASR-1.7B"], ["qwen3_asr_encoder", "Unreachable", "forward fails after shrink", 17, "hf-tiny-v2/tiny-random-Qwen3ASREncoder"], ["qwen3_moe", "Done", "", 8500460, "Qwen/Qwen3-Coder-30B-A3B-Instruct-FP8"], ["qwen3_next", "Done", "", 2235316, "Qwen/Qwen3-Coder-Next-FP8"], ["qwen3_omni_moe", "Unreachable", "forward fails after shrink", 1249447, "Qwen/Qwen3-Omni-30B-A3B-Instruct"], ["qwen3_omni_moe_audio_encoder", "Unreachable", "component config only", 1968, "Atotti/Qwen3-Omni-AudioTransformer"], ["qwen3_omni_moe_talker_code_predictor", "Unreachable", "component config only", 0, ""], ["qwen3_omni_moe_talker_text", "No checkpoint", "no Hub checkpoint has this model_type (documented Qwen/Qwen3-Omni-30B-A3B-Instruct is another type)", 0, ""], ["qwen3_omni_moe_text", "Unreachable", "component config only", 0, ""], ["qwen3_omni_moe_thinker", "Reachable", "", 298, "hf-tiny-v2/tiny-random-Qwen3OmniMoeThinkerForConditionalGeneration"], ["qwen3_omni_moe_vision_encoder", "Unreachable", "component config only", 0, ""], ["qwen3_vl", "Done", "", 19337885, "Qwen/Qwen3-VL-8B-Instruct"], ["qwen3_vl_moe", "Done", "", 1686056, "Qwen/Qwen3-VL-30B-A3B-Instruct"], ["qwen3_vl_moe_text", "Reachable", "", 63, "hf-tiny-v2/tiny-random-Qwen3VLMoeTextModel"], ["qwen3_vl_moe_vision", "Unreachable", "no usable checkpoint", 54, "hf-tiny-v2/tiny-random-Qwen3VLMoeVisionModel"], ["qwen3_vl_text", "Reachable", "", 587, "DuoNeural/Cosmos3-Nano-GPTQ-4bit"], ["qwen3_vl_vision", "Unreachable", "forward fails after shrink", 221, "madrisight/madrimed1.2-vision-embed"], ["qwen4_exp", "Done", "", 5763811, "Qwen/Qwen3.8-Flash-Next"], ["qwen4_exp_text", "Reachable", "", 17917, "Jab1718/qwen3.8-flash-coder-44gb-selective-int8"], ["qwen4_exp_vision", "No checkpoint", "no Hub checkpoint has this model_type (documented Qwen/Qwen4-Exp does not exist)", 0, ""], ["radio", "Unreachable", "no usable checkpoint", 3, "hf-tiny-v2/tiny-random-RadioModel"], ["rag", "Unreachable", "no usable checkpoint", 3924, "facebook/rag-token-nq"], ["recurrent_gemma", "Done", "", 4729, "google/recurrentgemma-2b-it"], ["reformer", "Unreachable", "model build fails", 17043, "google/reformer-crime-and-punishment"], ["regnet", "Reachable", "", 7237, "facebook/regnet-y-040"], ["rembert", "Reachable", "", 11890, "google/rembert"], ["resnet", "Done", "", 1174379, "microsoft/resnet-18"], ["rf_detr", "Unreachable", "forward fails after shrink", 28618, "Roboflow/rf-detr-base"], ["rf_detr_dinov2", "Unreachable", "forward fails after shrink", 70, "hf-tiny-v2/tiny-random-RfDetrDinov2Backbone"], ["roberta", "Done", "", 46597117, "FacebookAI/roberta-large"], ["roberta-prelayernorm", "Reachable", "", 4655, "andreasmadsen/efficient_mlm_m0.40"], ["roc_bert", "Reachable", "", 12437, "weiweishi/roc-bert-base-zh"], ["roformer", "Reachable", "", 45159, "optimum-intel-internal-testing/tiny-random-roformer"], ["rt_detr", "Done", "", 1687337, "PekingU/rtdetr_r101vd_coco_o365"], ["rt_detr_resnet", "Unreachable", "no usable checkpoint", 54, "hf-tiny-v2/tiny-random-RTDetrResNetBackbone"], ["rt_detr_v2", "Done", "forward fails after shrink", 2526763, "docling-project/docling-layout-heron"], ["rwkv", "Reachable", "", 14762, "RWKV/rwkv-4-169m-pile"], ["sam", "Unreachable", "forward fails after shrink", 976537, "facebook/sam-vit-base"], ["sam2", "Done", "", 19375, "onnx-community/sam2.1-hiera-tiny-ONNX"], ["sam2_hiera_det_model", "Unreachable", "forward fails after shrink", 59, "hf-tiny-v2/tiny-random-Sam2HieraDetModel"], ["sam2_mask_decoder", "No checkpoint", "no Hub checkpoint has this model_type (documented facebook/sam2.1-hiera-tiny is another type)", 0, ""], ["sam2_prompt_encoder", "No checkpoint", "no Hub checkpoint has this model_type (documented facebook/sam2.1-hiera-tiny is another type)", 0, ""], ["sam2_video", "Unreachable", "forward fails after shrink", 701203, "facebook/sam2.1-hiera-large"], ["sam2_video_mask_decoder", "No checkpoint", "no Hub checkpoint has this model_type (documented facebook/sam2_video.1-hiera-tiny does not exist)", 0, ""], ["sam2_video_prompt_encoder", "No checkpoint", "no Hub checkpoint has this model_type (documented facebook/sam2_video.1-hiera-tiny does not exist)", 0, ""], ["sam2_vision_model", "Unreachable", "forward fails after shrink", 57, "hf-tiny-v2/tiny-random-Sam2VisionModel"], ["sam3", "Unreachable", "forward fails after shrink", 3530, "unsloth/sam3.1"], ["sam3_detr_decoder", "Unreachable", "checkpoint gated", 0, "facebook/sam3"], ["sam3_detr_encoder", "Unreachable", "checkpoint gated", 0, "facebook/sam3"], ["sam3_geometry_encoder", "Unreachable", "checkpoint gated", 0, "facebook/sam3"], ["sam3_lite_text", "Unreachable", "forward fails after shrink", 14775, "vil-uob/sam3-litetext-s0"], ["sam3_lite_text_detr_decoder", "Unreachable", "component config only", 0, ""], ["sam3_lite_text_detr_encoder", "Unreachable", "component config only", 0, ""], ["sam3_lite_text_geometry_encoder", "Unreachable", "component config only", 0, ""], ["sam3_lite_text_mask_decoder", "Unreachable", "component config only", 0, ""], ["sam3_lite_text_text_model", "Unreachable", "forward fails after shrink", 39, "hf-tiny-v2/tiny-random-Sam3LiteTextTextModel"], ["sam3_mask_decoder", "Unreachable", "checkpoint gated", 0, "facebook/sam3"], ["sam3_tracker", "Unreachable", "no usable checkpoint", 1300, "onnx-community/sam3-tracker-ONNX"], ["sam3_tracker_mask_decoder", "Unreachable", "checkpoint gated", 0, "facebook/sam3"], ["sam3_tracker_prompt_encoder", "Unreachable", "checkpoint gated", 0, "facebook/sam3"], ["sam3_tracker_video", "Unreachable", "checkpoint gated", 0, "facebook/sam3"], ["sam3_tracker_video_mask_decoder", "Unreachable", "checkpoint gated", 0, "facebook/sam3"], ["sam3_tracker_video_prompt_encoder", "Unreachable", "checkpoint gated", 0, "facebook/sam3"], ["sam3_video", "Unreachable", "forward fails after shrink", 2087329, "facebook/sam3"], ["sam3_vision_model", "Unreachable", "no usable checkpoint", 38, "hf-tiny-v2/tiny-random-Sam3VisionModel"], ["sam3_vit_model", "Unreachable", "no usable checkpoint", 37, "hf-tiny-v2/tiny-random-Sam3ViTModel"], ["sam_hq", "Unreachable", "forward fails after shrink", 16860, "syscv-community/sam-hq-vit-base"], ["sam_hq_mask_decoder", "No checkpoint", "no Hub checkpoint has this model_type (documented syscv-community/sam-hq-vit-base is another type)", 0, ""], ["sam_hq_prompt_encoder", "No checkpoint", "no Hub checkpoint has this model_type (documented syscv-community/sam-hq-vit-base is another type)", 0, ""], ["sam_hq_vision_model", "Reachable", "", 138, "hf-tiny-v2/tiny-random-SamHQVisionModel"], ["sam_mask_decoder", "No checkpoint", "no Hub checkpoint has this model_type (documented facebook/sam-vit-huge is another type)", 0, ""], ["sam_prompt_encoder", "No checkpoint", "no Hub checkpoint has this model_type (documented facebook/sam-vit-huge is another type)", 0, ""], ["sam_vision_model", "Reachable", "", 138, "hf-tiny-v2/tiny-random-SamVisionModel"], ["sapiens2", "Unreachable", "forward fails after shrink", 103690, "facebook/sapiens2-pose-1b"], ["sapiens2_head", "Unreachable", "component config only", 0, ""], ["seamless_m4t", "Reachable", "", 13797, "facebook/hf-seamless-m4t-medium"], ["seamless_m4t_v2", "Done", "", 316927, "facebook/seamless-m4t-v2-large"], ["seed_oss", "Done", "", 128943, "ByteDance-Seed/Seed-OSS-36B-Instruct"], ["segformer", "Done", "", 2291917, "mattmdjaga/segformer_b2_clothes"], ["seggpt", "Unreachable", "forward fails after shrink", 21384, "BAAI/seggpt-vit-large"], ["sew", "Reachable", "", 15680, "optimum-intel-internal-testing/tiny-random-SEWModel"], ["sew-d", "Reachable", "", 18152, "optimum-intel-internal-testing/sew-d-tiny-100k-ft-ls100h"], ["shieldgemma2", "Reachable", "", 6627, "google/shieldgemma-2-4b-it"], ["siglip", "Done", "", 18018379, "prithivMLmods/Gender-Classifier-Mini"], ["siglip2", "Unreachable", "forward fails after shrink", 1622065, "google/siglip2-base-patch16-naflex"], ["siglip2_text_model", "Unreachable", "component config only", 56, "sortify-ai/siglip2-so400m-patch14-384-text"], ["siglip2_vision_model", "Unreachable", "forward fails after shrink", 58, "hf-tiny-v2/tiny-random-Siglip2VisionModel"], ["siglip_text_model", "Unreachable", "component config only", 19, "carsonpoole/binary-siglip-text"], ["siglip_vision_model", "Reachable", "", 1842, "zchuning/pi05-siglip-vision-tower"], ["slanet", "Reachable", "", 6957, "PaddlePaddle/SLANet_plus_safetensors"], ["slanext", "Reachable", "", 212, "PaddlePaddle/SLANeXt_wired_safetensors"], ["slanext_vision", "No checkpoint", "no Hub checkpoint has this model_type (documented PaddlePaddle/SLANeXt_wired_safetensors is another type)", 0, ""], ["smollm3", "Done", "", 1135357, "HuggingFaceTB/SmolLM3-3B"], ["smolvlm", "Done", "", 1742362, "HuggingFaceTB/SmolVLM2-500M-Video-Instruct"], ["smolvlm_vision", "Unreachable", "forward fails after shrink", 5, "hf-tiny-v2/tiny-random-SmolVLMVisionTransformer"], ["solar_open", "Reachable", "", 32622, "upstage/Solar-Open-100B"], ["speech-encoder-decoder", "Unreachable", "synthetic inputs rejected", 5244, "hf-tiny-v2/tiny-random-SpeechEncoderDecoderModel-wav2vec2-bert"], ["speech_to_text", "Unreachable", "forward fails after shrink", 23544, "facebook/s2t-small-librispeech-asr"], ["speecht5", "Done", "", 194296, "optimum-intel-internal-testing/tiny-random-SpeechT5ForTextToSpeech"], ["speecht5_hifigan", "No checkpoint", "no Hub checkpoint has this model_type (documented microsoft/speecht5_asr is another type)", 0, ""], ["splinter", "Reachable", "", 3741, "tau/splinter-base"], ["squeezebert", "Unreachable", "model build fails", 33330, "optimum-intel-internal-testing/tiny-random-squeezebert"], ["stablelm", "Done", "", 169690, "optimum-intel-internal-testing/tiny-random-StableLmForCausalLM"], ["starcoder2", "Done", "", 152451, "optimum-intel-internal-testing/tiny-random-Starcoder2ForCausalLM"], ["step3p5", "Unreachable", "shrink fails", 266885, "stepfun-ai/Step-3.5-Flash"], ["step3p5_vision", "No checkpoint", "no Hub checkpoint has this model_type (documented stepfun-ai/Step-3.7-Flash is another type)", 0, ""], ["step3p7", "Unreachable", "shrink fails", 26135, "stepfun-ai/Step-3.7-Flash"], ["superglue", "Unreachable", "forward fails after shrink", 8964, "magic-leap-community/superglue_indoor"], ["superpoint", "Unreachable", "backward fails", 9952, "magic-leap-community/superpoint"], ["swiftformer", "Reachable", "", 7504, "MBZUAI/swiftformer-xs"], ["swin", "Done", "", 238869, "Organika/sdxl-detector"], ["swin2sr", "Reachable", "", 56057, "caidas/swin2SR-classical-sr-x2-64"], ["swinv2", "Done", "", 434756, "microsoft/swinv2-tiny-patch4-window16-256"], ["switch_transformers", "Reachable", "", 10063, "google/switch-base-8"], ["t5", "Done", "", 94678669, "google-t5/t5-small"], ["t5_gemma_module", "Unreachable", "component config only", 70, "RE-N-Y/t5gemma-2b-2b-ul2-it"], ["t5gemma", "Unreachable", "no usable checkpoint", 176067, "google/t5gemma-2b-2b-ul2-it"], ["t5gemma2", "Unreachable", "forward fails after shrink", 114374, "google/t5gemma-2-270m-270m"], ["t5gemma2_decoder", "Unreachable", "component config only", 0, ""], ["t5gemma2_encoder", "Reachable", "", 149, "hf-tiny-v2/tiny-random-T5Gemma2Encoder"], ["t5gemma2_text", "Unreachable", "component config only", 30, "amanchoudhri/t5-gemma-2-270m-text-encoder"], ["table-transformer", "Done", "", 1872765, "microsoft/table-transformer-structure-recognition"], ["tapas", "Reachable", "", 37821, "google/tapas-base-finetuned-wtq"], ["textnet", "Unreachable", "forward fails after shrink", 7475, "czczup/textnet-base"], ["time_series_transformer", "Unreachable", "synthetic inputs rejected", 4534, "huggingface/time-series-transformer-tourism-monthly"], ["timesfm", "Done", "", 1854129, "google/timesfm-2.5-200m-pytorch"], ["timesfm2_5", "Unreachable", "synthetic inputs rejected", 37239, "google/timesfm-2.5-200m-transformers"], ["timesformer", "Unreachable", "forward fails after shrink", 25737, "facebook/timesformer-base-finetuned-k400"], ["timm_backbone", "Unreachable", "no usable checkpoint", 133, "hf-tiny-v2/tiny-random-TimmBackbone"], ["timm_wrapper", "Unreachable", "forward fails after shrink", 383060, "Freepik/nsfw_image_detector"], ["tipsv2", "Done", "", 289583, "google/tipsv2-so400m14"], ["tipsv2_dpt", "Unreachable", "forward fails after shrink", 8014, "google/tipsv2-b14-dpt"], ["tipsv2_text_model", "Reachable", "", 4, "hf-tiny-v2/tiny-random-Tipsv2TextModel"], ["tipsv2_vision_model", "Reachable", "", 8, "hf-tiny-v2/tiny-random-Tipsv2VisionBackbone"], ["trocr", "Reachable", "", 53, "sasmithasarasi/trocr-sinhala-handwriting"], ["tvp", "Unreachable", "forward fails after shrink", 13994, "Intel/tvp-base"], ["udop", "Unreachable", "forward fails after shrink", 9051, "microsoft/udop-large"], ["ultrabert", "Reachable", "", 29547, "dmis-lab/UltraBERT-380M"], ["ultravbert", "Reachable", "", 1133, "athrael-soju/ColUltraVBERT"], ["umt5", "Done", "", 5305, "huggingface-mic/umt5-small-mic"], ["unispeech", "Reachable", "", 24130, "optimum-intel-internal-testing/tiny-random-unispeech"], ["unispeech-sat", "Reachable", "", 38482, "optimum-intel-internal-testing/tiny-random-UnispeechSatModel"], ["univnet", "Unreachable", "forward fails after shrink", 5412, "dg845/univnet-dev"], ["unlimited_ocr", "Done", "", 3535621, "baidu/Unlimited-OCR"], ["unlimited_ocr_sam_vision_model", "No checkpoint", "no Hub checkpoint has this model_type (documented baidu/Unlimited-OCR is another type)", 0, ""], ["unlimited_ocr_text", "No checkpoint", "no Hub checkpoint has this model_type (documented baidu/Unlimited-OCR is another type)", 0, ""], ["unlimited_ocr_vision", "No checkpoint", "no Hub checkpoint has this model_type (documented baidu/Unlimited-OCR is another type)", 0, ""], ["unlimited_ocr_vision_encoder", "No checkpoint", "no Hub checkpoint has this model_type (documented baidu/Unlimited-OCR is another type)", 0, ""], ["upernet", "Unreachable", "forward fails after shrink", 46968, "openmmlab/upernet-convnext-small"], ["uvdoc", "Reachable", "", 431, "PaddlePaddle/UVDoc_safetensors"], ["uvdoc_backbone", "Unreachable", "backward fails", 39, "hf-tiny-v2/tiny-random-UVDocBackbone"], ["vaultgemma", "Reachable", "", 2790, "google/vaultgemma-1b"], ["vibevoice", "Done", "", 1070414, "microsoft/VibeVoice-ASR"], ["vibevoice_acoustic_tokenizer", "Unreachable", "forward fails after shrink", 2151, "microsoft/VibeVoice-AcousticTokenizer"], ["vibevoice_acoustic_tokenizer_decoder", "Unreachable", "no usable checkpoint", 33, "hf-tiny-v2/tiny-random-VibeVoiceAcousticTokenizerDecoderModel"], ["vibevoice_acoustic_tokenizer_encoder", "Unreachable", "no usable checkpoint", 49, "hf-tiny-v2/tiny-random-VibeVoiceAcousticTokenizerEncoderModel"], ["vibevoice_asr", "Reachable", "", 59115, "microsoft/VibeVoice-ASR-HF"], ["vibevoice_diffusion", "No checkpoint", "no Hub checkpoint has this model_type (documented vibevoice/VibeVoice-1.5B-hf is another type)", 0, ""], ["video_llama_3", "Reachable", "", 11087, "lkhl/VideoLLaMA3-2B-Image-HF"], ["video_llama_3_vision", "Unreachable", "forward fails after shrink", 32, "hf-tiny-v2/tiny-random-VideoLlama3VisionModel"], ["video_llava", "Reachable", "", 10045, "LanguageBind/Video-LLaVA-7B-hf"], ["videomae", "Unreachable", "forward fails after shrink", 475037, "MCG-NJU/videomae-base"], ["videomt", "Unreachable", "synthetic inputs rejected", 5328, "tue-mps/videomt-dinov2-small-ytvis2019"], ["videoprism", "Unreachable", "synthetic inputs rejected", 545, "MHRDYN7/videoprism-lvt-base-f16r288"], ["videoprism_text_model", "No checkpoint", "no Hub checkpoint has this model_type (documented google/videoprism-lvt-base-f16r288 has no config.json)", 0, ""], ["videoprism_vision_model", "Unreachable", "synthetic inputs rejected", 191, "MHRDYN7/videoprism-base-f16r288"], ["vilt", "Unreachable", "forward fails after shrink", 89943, "dandelin/vilt-b32-finetuned-vqa"], ["vipllava", "Reachable", "", 7898, "llava-hf/vip-llava-7b-hf"], ["vision-encoder-decoder", "Unreachable", "forward fails after shrink", 2482977, "kha-white/manga-ocr-base"], ["vision-text-dual-encoder", "Unreachable", "model build fails", 3250, "hf-tiny-v2/tiny-random-VisionTextDualEncoderModel-vit-bert"], ["visual_bert", "Reachable", "", 5237, "uclanlp/visualbert-vqa-coco-pre"], ["vit", "Done", "", 23571533, "google/vit-base-patch16-224"], ["vit_mae", "Reachable", "", 43367, "facebook/vit-mae-base"], ["vit_msn", "Reachable", "", 1782, "facebook/vit-msn-small"], ["vitdet", "Unreachable", "no usable checkpoint", 214, "hf-tiny-v2/tiny-random-VitDetBackbone"], ["vitmatte", "Done", "", 383080, "hustvl/vitmatte-small-composition-1k"], ["vitpose", "Unreachable", "forward fails after shrink", 392468, "usyd-community/vitpose-base-simple"], ["vitpose_backbone", "Unreachable", "forward fails after shrink", 124, "hf-tiny-v2/tiny-random-VitPoseBackbone"], ["vits", "Done", "", 455247, "facebook/mms-tts-eng"], ["vivit", "Unreachable", "forward fails after shrink", 89874, "google/vivit-b-16x2"], ["vjepa2", "Unreachable", "synthetic inputs rejected", 426419, "facebook/vjepa2-vitl-fpc64-256"], ["voxtral", "Done", "", 319265, "mzbac/voxtral-mini-3b-4bit-mixed"], ["voxtral_encoder", "Reachable", "", 108, "hf-tiny-v2/tiny-random-VoxtralEncoder"], ["voxtral_realtime", "Unreachable", "forward fails after shrink", 1948459, "mistralai/Voxtral-Mini-4B-Realtime-2602"], ["voxtral_realtime_encoder", "Reachable", "", 80, "hf-tiny-v2/tiny-random-VoxtralRealtimeEncoder"], ["voxtral_realtime_text", "Unreachable", "model build fails", 10, "hf-tiny-v2/tiny-random-VoxtralRealtimeTextModel"], ["wav2vec2", "Done", "", 100317588, "jonatasgrosman/wav2vec2-large-xlsr-53-japanese"], ["wav2vec2-bert", "Unreachable", "model build fails", 1320062, "facebook/w2v-bert-2.0"], ["wav2vec2-conformer", "Unreachable", "model build fails", 160365, "facebook/wav2vec2-conformer-rope-large-960h-ft"], ["wavlm", "Done", "", 1990346, "microsoft/wavlm-large"], ["whisper", "Done", "", 22332636, "openai/whisper-large-v3-turbo"], ["xclip", "Unreachable", "forward fails after shrink", 453798, "microsoft/xclip-base-patch16-zero-shot"], ["xclip_text_model", "Unreachable", "component config only", 0, ""], ["xclip_vision_model", "Unreachable", "component config only", 0, ""], ["xcodec", "Unreachable", "config load fails", 14050, "hf-audio/xcodec-hubert-general"], ["xcodec2", "Unreachable", "forward fails after shrink", 9249, "HKUSTAudio/xcodec2-hf"], ["xglm", "Reachable", "", 64850, "optimum-intel-internal-testing/tiny-random-XGLMForCausalLM"], ["xlm", "Reachable", "", 19454, "FacebookAI/xlm-mlm-en-2048"], ["xlm-roberta", "Done", "", 104687870, "BAAI/bge-m3"], ["xlm-roberta-xl", "Done", "", 101085, "facebook/xlm-roberta-xl"], ["xlnet", "Unreachable", "shrink fails", 166190, "xlnet/xlnet-base-cased"], ["xlstm", "Unreachable", "config load fails", 4922, "NX-AI/xLSTM-7b"], ["xmod", "Unreachable", "forward fails after shrink", 29093, "facebook/xmod-base"], ["yolos", "Done", "", 859794, "hustvl/yolos-small"], ["yoso", "Reachable", "", 5281, "uw-madison/yoso-4096"], ["youtu", "Reachable", "", 12971, "tencent/Youtu-LLM-2B"], ["zamba", "Unreachable", "model build fails", 5935, "Zyphra/Zamba-7B-v1"], ["zamba2", "Unreachable", "forward fails after shrink", 70083, "Zyphra/Zamba2-1.2B-instruct"], ["zaya", "Reachable", "", 8978, "Zyphra/ZAYA1-8B"], ["zoedepth", "Unreachable", "forward fails after shrink", 48545, "Intel/zoedepth-nyu-kitti"]], "meta": {"generated": "2026-10-31", "fleet": "666d54fcc782e473ff276ec9626df0b0a3f0cafd", "transformers": "420dca92a4cf7d96dfbc1cc46f00375b0681ff8f", "sweep": "98a8d1d90b64baf9cbca7197d95f139901ae28f6"}, "shrink": {"albert": [["albert-base-v2-mic", 0]], "apertus": [["Apertus-v1.1-0.5B-Instruct-mic", 0]], "audioflamingo3": [["audio-flamingo-3-hf-mic", 0]], "bart": [["bart-base-mic", 0]], "beit": [["beit-base-patch16-224-mic", 0]], "bert": [["bert-base-uncased-mic", 0], ["bert-mic", 0]], "big_bird": [["bigbird-roberta-base-mic", 0]], "blenderbot": [["blenderbot-400M-distill-mic", 0]], "bloom": [["bloomz-560m-mic", 0]], "camembert": [["camembert-base-mic", 0]], "canine": [["canine-c-mic", 0]], "chinese_clip": [["chinese-clip-vit-base-patch16-mic", 0]], "clap": [["clap-htsat-fused-mic", 0]], "clip": [["clip-mic", 0], ["clip-vit-base-patch32-mic", 0]], "clip_text_model": [["tiny-clip-text-2-mic", 0]], "clip_vision_model": [["clip-vit-base-patch32-vision-mic", 0]], "clipseg": [["clipseg-rd64-refined-mic", 0]], "cohere": [["aya-expanse-8b-ungated-mic", 0]], "cohere2": [["tiny-random-aya-base-mic", 0]], "convbert": [["conv-bert-base-mic", 0]], "convnext": [["convnext-tiny-224-mic", 0]], "convnextv2": [["convnextv2-nano-1k-224-mic", 0]], "cosmos3_omni": [["Cosmos3-Nano-mic", 0]], "csm": [["csm-1b-mic", 0]], "deberta": [["deberta-large-mnli-mic", 0]], "deberta-v2": [["mDeBERTa-v3-base-xnli-multilingual-nli-2mil7-mic", 0]], "deepseek_v2": [["DeepSeek-Coder-V2-Lite-Instruct-mic", 0]], "deepseek_v3": [["DeepSeek-R1-mic", 0]], "deepseek_v32": [["DeepSeek-V3.2-mic", 0]], "deepseek_v4": [["DeepSeek-V4-Flash-0731-mic", 0]], "detr": [["detr-mic", 0]], "dinov2": [["dinov2-small-mic", 0]], "dinov2_with_registers": [["dinov2-with-registers-base-mic", 0]], "dinov3_vit": [["dinov3-vitl16-pretrain-lvd1689m-mic", 0]], "distilbert": [["distilbert-base-uncased-mic", 0]], "dpt": [["dpt-large-mic", 0]], "electra": [["electra-small-discriminator-mic", 0]], "embedding_gemma2": [["embeddinggemma-2-mic", 0]], "encoder-decoder": [["bert2bert-cnn_dailymail-fp16-mic", 0]], "esm": [["esm2_t6_8M_UR50D-mic", 0]], "esmc": [["ESMC-300M-hf-mic", 0]], "eurobert": [["EuroBERT-210m-mic", 0]], "falcon": [["falcon-7b-mic", 0]], "falcon_h1": [["Falcon-H1-Tiny-90M-Instruct-mic", 0]], "florence2": [["Florence-2-base-mic", 0]], "fnet": [["fnet-base-mic", 0]], "gemma": [["gemma-2b-mic", 0]], "gemma2": [["gemma-2-2b-it-mic", 0]], "gemma3": [["gemma3-mic", 0]], "gemma3n": [["gemma-3n-E2B-it-mic", 0]], "gemma4": [["gemma-4-E2B-it-mic", 0]], "gemma4_unified": [["gemma-4-12B-it-mic", 0]], "glm4": [["GLM-4-9B-0414-mic", 0]], "glm4_moe": [["GLM-4.5-Air-mic", 0]], "glm4_moe_lite": [["GLM-4.7-Flash-mic", 0]], "glm5_next": [["GLM-5.3-Flash-mic", 1]], "glm_moe_dsa": [["GLM-5.3-mic", 1]], "glm_ocr": [["GLM-OCR-mic", 0]], "got_ocr2": [["got_ocr2-mic", 0]], "gpt2": [["gpt2-mic", 0]], "gpt_bigcode": [["tiny_starcoder_py-mic", 0]], "gpt_neox": [["pythia-160m-mic", 0]], "gpt_oss": [["gpt_oss-mic", 0]], "granite": [["granite-3.1-2b-instruct-mic", 0]], "granite4_vision": [["granite-vision-4.1-4b-mic", 0]], "granite_speech": [["granite-4.0-1b-speech-mic", 0]], "granitemoe": [["granite-3.0-1b-a400m-instruct-mic", 0]], "granitemoehybrid": [["granite-4.0-h-350m-mic", 0]], "grounding-dino": [["grounding-dino-tiny-mic", 0]], "hubert": [["hubert-large-ls960-ft-mic", 0]], "hunyuan_v1_dense": [["Hunyuan-0.5B-Instruct-mic", 0]], "hunyuan_vl": [["HunyuanOCR-mic", 0]], "hy_v3": [["Hy-MT2-30B-A3B-mic", 0]], "idefics3": [["granite-docling-258M-mic", 0]], "inkling_mm_model": [["Inkling-mic", 0]], "internvl": [["internvl-mic", 0]], "jamba": [["Jamba-tiny-dev-mic", 0]], "janus": [["Janus-Pro-1B-mic", 0]], "jina_embeddings_v3": [["jina-embeddings-v3-hf-mic", 0]], "laguna": [["Laguna-XS.2-mic", 0]], "layoutlm": [["layoutlm-base-uncased-mic", 0]], "lfm2": [["LFM2.5-230M-Base-mic", 0]], "lfm2_moe": [["LFM2-8B-A1B-mic", 0]], "lfm2_vl": [["LFM2.5-VL-450M-mic", 0]], "llama": [["llama-mic", 0], ["smollm-1.7b-mic", 0], ["smollm-mic", 0]], "llama4": [["Llama-4-Scout-17B-16E-Instruct-mic", 0]], "llava": [["llava-mic", 0]], "llava_next": [["llava-v1.6-mistral-7b-hf-mic", 0]], "llava_next_video": [["LLaVA-NeXT-Video-7B-hf-mic", 0]], "llava_onevision": [["llava-onevision-qwen2-0.5b-ov-hf-mic", 0]], "longformer": [["longformer-base-4096-mic", 0]], "longt5": [["long-t5-tglobal-base-mic", 0]], "luke": [["luke-base-mic", 0]], "m2m_100": [["nllb-200-distilled-600M-mic", 0]], "mamba": [["mamba-130m-hf-mic", 0]], "marian": [["opus-mt-fr-en-mic", 0]], "mask2former": [["mask2former-swin-tiny-coco-instance-mic", 0]], "mbart": [["mbart-large-50-mic", 0]], "minicpm3": [["MiniCPM3-4B-mic", 0]], "minicpmv4_6": [["MiniCPM-V-4.6-mic", 0]], "minimax_m2": [["MiniMax-M2.5-mic", 0]], "minimax_m3_vl": [["MiniMax-M3-MXFP8-mic", 0]], "ministral3": [["Nemotron-3-Embed-1B-BF16-mic", 0]], "mistral": [["Mistral-7B-Instruct-v0.2-mic", 1]], "mistral3": [["Ministral-3-3B-Instruct-2512-mic", 0], ["mistral3-mic", 0]], "mixtral": [["Mixtral-8x7B-Instruct-v0.1-mic", 0]], "mllama": [["Llama-3.2-11B-Vision-Instruct-mic", 0]], "mm-grounding-dino": [["mm_grounding_dino_tiny_o365v1_goldg_v3det-mic", 0]], "mobilenet_v2": [["mobilenet_v2_1.0_224-mic", 0]], "mobilevit": [["mobilevit-xx-small-mic", 0]], "modernbert": [["modernbert-mic", 0]], "moss_transcribe_diarize": [["MOSS-Transcribe-Diarize-HF-mic", 0]], "mpnet": [["mpnet-base-mic", 0]], "mt5": [["mt5-small-mic", 0]], "muse_glimmer": [["tiny-random-muse-glimmer-mic", 0]], "nemotron_h": [["NVIDIA-Nemotron-3-Nano-4B-BF16-mic", 0]], "nomic_bert": [["nomic-embed-text-v1.5-mic", 0]], "olmo": [["OLMo-1B-hf-mic", 0]], "olmo2": [["OLMo-2-0425-1B-mic", 0]], "olmo3": [["Olmo-3-7B-Instruct-mic", 0]], "olmoe": [["OLMoE-1B-7B-0924-mic", 0]], "openai_privacy_filter": [["privacy-filter-mic", 0]], "opt": [["opt-125m-mic", 0]], "owlv2": [["owlv2-base-patch16-ensemble-mic", 0]], "owlvit": [["owlvit-base-patch32-mic", 0]], "paddleocr_vl": [["tibetan-ocr-mic", 0]], "paligemma": [["paligemma-3b-pt-224-mic", 0]], "pegasus": [["pegasus-xsum-mic", 0]], "perceiver": [["language-perceiver-mic", 0]], "phi": [["phi-2-mic", 0]], "phi3": [["Phi-3.5-mini-instruct-mic", 0]], "phi4_multimodal": [["Phi-4-multimodal-instruct-mic", 0]], "phimoe": [["Phi-3.5-MoE-0.8B-A0.2B-mic", 0]], "pvt": [["pvt-tiny-224-mic", 0]], "qwen2": [["qwen2-mic", 0]], "qwen2_5_omni": [["qwen2_5_omni-mic", 0]], "qwen2_5_omni_thinker": [["LCO-Embedding-Omni-3B-mic", 0]], "qwen2_5_vl": [["qwen2_5_vl-mic", 0]], "qwen2_audio": [["qwen2_audio-mic", 0]], "qwen2_moe": [["Qwen1.5-MoE-A2.7B-mic", 0]], "qwen2_vl": [["Qwen2-VL-2B-Instruct-mic", 0]], "qwen3": [["Qwen3-0.6B-mic", 0]], "qwen3_5_moe": [["Qwen3.5-35B-A3B-mic", 0]], "qwen3_5_moe_text": [["Qwen3.8-1.0B-A0.6B-mic", 0]], "qwen3_5_text": [["Qwen3.5-0.8B-text-only-mic", 0]], "qwen3_asr": [["Qwen3-ASR-0.6B-hf-mic", 0]], "qwen3_moe": [["Qwen3-30B-A3B-mic", 0]], "qwen3_next": [["Qwen3-Next-80B-A3B-Instruct-mic", 0]], "qwen3_vl": [["Qwen3-VL-2B-Instruct-mic", 0]], "qwen3_vl_moe": [["Qwen3-VL-30B-A3B-Instruct-mic", 0]], "qwen4_exp": [["Qwen3.8-Flash-Next-mic", 0]], "recurrent_gemma": [["recurrentgemma-2b-mic", 0]], "resnet": [["resnet-18-mic", 0]], "roberta": [["roberta-base-mic", 0]], "rt_detr": [["rtdetr_r50vd-mic", 0]], "rt_detr_v2": [["docling-layout-heron-mic", 0]], "sam2": [["sam2.1-hiera-tiny-mic", 4]], "seamless_m4t_v2": [["seamless-m4t-v2-large-mic", 0]], "seed_oss": [["Seed-OSS-36B-Instruct-mic", 0]], "segformer": [["segformer-b0-finetuned-ade-512-512-mic", 0]], "siglip": [["siglip-base-patch16-224-mic", 0]], "smollm3": [["SmolLM3-3B-mic", 1]], "smolvlm": [["smolvlm-mic", 0]], "speecht5": [["speech_t5-mic", 0]], "stablelm": [["stablelm-2-1_6b-mic", 0]], "starcoder2": [["starcoder2-3b-mic", 0]], "swin": [["swin-tiny-patch4-window7-224-mic", 0]], "swinv2": [["swinv2-tiny-patch4-window16-256-mic", 0]], "t5": [["t5-mic", 0], ["t5-small-mic", 0]], "table-transformer": [["table_transformer-mic", 0]], "timesfm": [["timesfm-2.0-500m-pytorch-mic", 0]], "tipsv2": [["tipsv2-b14-mic", 0]], "umt5": [["umt5-small-mic", 0]], "unlimited_ocr": [["Unlimited-OCR-mic", 0]], "vibevoice": [["VibeVoice-1.5B-hf-mic", 0]], "vit": [["vit-mic", 0]], "vitmatte": [["vitmatte-small-composition-1k-mic", 0]], "vits": [["mms-tts-eng-mic", 0]], "voxtral": [["Voxtral-Mini-3B-2507-mic", 0]], "wav2vec2": [["wav2vec2-base-960h-mic", 0], ["wav2vec2-mic", 0]], "wavlm": [["wavlm-base-plus-mic", 0]], "whisper": [["whisper-mic", 0], ["whisper-tiny-mic", 0]], "xlm-roberta": [["xlm-roberta-base-mic", 0]], "xlm-roberta-xl": [["xlm-roberta-xl-mic", 0]], "yolos": [["yolos-small-mic", 0]]}, "gated": {"aya_vision": "auto", "chameleon": "manual", "chmv2": "manual", "cohere_asr": "auto", "csm_depth_decoder_model": "auto", "cwm": "manual", "dinov3_convnext": "manual", "gemma": "manual", "gemma2": "manual", "gemma3": "manual", "gemma3_text": "manual", "gemma3n": "manual", "jais2": "auto", "lasr_ctc": "auto", "llama4_vision_model": "manual", "moshi": "auto", "muse_spark": "manual", "muse_spark_audio": "manual", "muse_spark_text": "manual", "muse_spark_vision": "manual", "neucodec": "auto", "perception_lm": "manual", "recurrent_gemma": "manual", "sam3_detr_decoder": "manual", "sam3_detr_encoder": "manual", "sam3_geometry_encoder": "manual", "sam3_mask_decoder": "manual", "sam3_tracker_mask_decoder": "manual", "sam3_tracker_prompt_encoder": "manual", "sam3_tracker_video": "manual", "sam3_tracker_video_mask_decoder": "manual", "sam3_tracker_video_prompt_encoder": "manual", "sam3_video": "manual", "shieldgemma2": "manual", "t5gemma": "manual", "t5gemma2": "manual", "vaultgemma": "manual"}}

index.html

@@ -1,19 +1,205 @@
 <!doctype html>
-<html>
-	<head>
-		<meta charset="utf-8" />
-		<meta name="viewport" content="width=device-width" />
-		<title>My static Space</title>
-		<link rel="stylesheet" href="style.css" />
-	</head>
-	<body>
-		<div class="card">
-			<h1>Welcome to your static Space!</h1>
-			<p>You can modify this app directly by editing <i>index.html</i> in the Files and versions tab.</p>
-			<p>
-				Also don't forget to check the
-				<a href="https://huggingface.co/docs/hub/spaces" target="_blank">Spaces documentation</a>.
-			</p>
-		</div>
-	</body>
+<html lang="en">
+<head>
+<meta charset="utf-8">
+<meta name="viewport" content="width=device-width,initial-scale=1">
+<title>MIC coverage</title>
+<style>
+:root{--bg:#fff;--fg:#1f2328;--muted:#656d76;--line:#d0d7de;--soft:#f6f8fa;--accent:#0969da;--done:#1a7f37;--reach:#bf8700;--no:#cf222e;--goal:#8250df;--goal-bg:#fbefff;--none:#8c959f}
+@media (prefers-color-scheme:dark){:root{--bg:#0d1117;--fg:#e6edf3;--muted:#8d96a0;--line:#30363d;--soft:#161b22;--accent:#4493f8;--done:#3fb950;--reach:#d29922;--no:#f85149;--goal:#ab7df8;--goal-bg:#2a1f3d;--none:#6e7681}}
+*{box-sizing:border-box}
+body{margin:0;background:var(--bg);color:var(--fg);font:14px/1.45 system-ui,-apple-system,"Segoe UI",sans-serif}
+main{max-width:1200px;margin:0 auto;padding:24px 16px 48px}
+h1{font-size:22px;margin:0 0 6px}
+.lead{color:var(--muted);margin:0 0 18px;max-width:85ch}
+.bars{display:grid;gap:14px;margin-bottom:18px}
+.bar-head{display:flex;justify-content:space-between;gap:12px;flex-wrap:wrap;margin-bottom:5px}
+.bar-head strong{font-size:14px}.bar-head span{color:var(--muted);font-size:13px}
+.bar{position:relative;display:flex;height:22px;border-radius:6px;overflow:hidden;background:var(--soft);border:1px solid var(--line)}
+.bar i{display:block;height:100%;transition:flex-grow .3s}
+.bar .Done{background:var(--done)}.bar .Reachable{background:var(--reach)}.bar .Unreachable{background:var(--no);opacity:.55}.bar .NoCheckpoint{background:var(--none)}
+.bar .mark{position:absolute;top:-3px;bottom:-3px;width:2px;background:var(--fg)}
+.legend{display:flex;gap:16px;flex-wrap:wrap;color:var(--muted);font-size:12px;margin-top:5px}
+.legend b{display:inline-block;width:10px;height:10px;border-radius:2px;margin-right:5px;vertical-align:-1px}
+.stats{display:flex;gap:10px;flex-wrap:wrap;margin-bottom:14px}
+.stat{border:1px solid var(--line);border-radius:6px;padding:8px 12px;background:var(--soft);cursor:pointer;font:inherit;color:inherit;text-align:left}
+.stat strong{display:block;font-size:20px}
+.stat[aria-pressed=true]{outline:2px solid var(--accent)}
+.stat.goal strong{color:var(--goal)}
+.controls{display:flex;gap:10px;align-items:center;margin-bottom:10px;flex-wrap:wrap}
+input[type=search]{flex:1;min-width:200px;max-width:360px;padding:6px 10px;border:1px solid var(--line);border-radius:6px;background:var(--bg);color:var(--fg);font:inherit}
+.count{color:var(--muted)}
+.wrap{overflow-x:auto;border:1px solid var(--line);border-radius:6px}
+table{width:100%;border-collapse:collapse}
+th,td{padding:6px 10px;border-bottom:1px solid var(--line);text-align:left;white-space:nowrap}
+th{position:sticky;top:0;background:var(--soft);cursor:pointer;user-select:none;font-weight:600}
+th:hover{color:var(--accent)}
+th .arrow{color:var(--muted);font-size:11px;margin-left:4px}
+td.num,th.num{text-align:right;font-variant-numeric:tabular-nums}
+tbody tr:hover{background:var(--soft)}
+tr.goal td:first-child{box-shadow:inset 3px 0 0 var(--goal)}
+tr.goal.miss{background:var(--goal-bg)}
+.badge{display:inline-block;margin-left:6px;padding:0 6px;border-radius:10px;font-size:11px;font-weight:600;color:var(--goal);border:1px solid var(--goal);vertical-align:1px}
+td.top{white-space:normal;overflow-wrap:anywhere;min-width:12ch}
+td.gated{white-space:nowrap;font-size:12px}td.why{white-space:normal;overflow-wrap:anywhere;min-width:14ch;max-width:36ch}
+@media (max-width:700px){main{padding:16px 12px 40px}.wrap{border:0;overflow:visible}table,tbody,tr,td{display:block}thead tr{display:flex;flex-wrap:wrap;gap:6px;margin-bottom:8px}thead th{position:static;border:1px solid var(--line);border-radius:12px;padding:2px 8px;font-size:12px}tbody tr{border:1px solid var(--line);border-radius:6px;margin-bottom:8px;padding:6px 10px}td{border:0;padding:1px 0;white-space:normal;overflow-wrap:anywhere;text-align:left!important}td:empty{display:none}td::before{content:attr(data-label);display:inline-block;min-width:11em;color:var(--muted);font-size:12px}td.why,td.top{min-width:0;max-width:none}tr.goal td:first-child{box-shadow:none}tr.goal{border-left:3px solid var(--goal)}}.legend span{min-width:0}
+a{color:var(--accent);text-decoration:none}a:hover{text-decoration:underline}
+.state{font-weight:600}td.Done{color:var(--done)}td.Reachable{color:var(--reach)}td.Unreachable{color:var(--no)}td.NoCheckpoint{color:var(--muted)}
+.foot{color:var(--muted);font-size:12px;margin-top:14px}
+</style>
+</head>
+<body>
+<main>
+<h1>MIC coverage of Transformers model types</h1>
+<p class="lead" id="lead">Loading…</p>
+<div class="bars" id="bars"></div>
+<div class="stats" id="stats"></div>
+<div class="controls">
+<input type="search" id="q" placeholder="Filter by model type, reason or checkpoint" aria-label="Filter rows">
+<span class="count" id="count"></span>
+</div>
+<div class="wrap"><table>
+<thead><tr>
+<th data-col="5" scope="col" class="num" title="Rank by downloads">#<span class="arrow"></span></th>
+<th data-col="0" scope="col">Model type<span class="arrow"></span></th>
+<th data-col="1" scope="col">State<span class="arrow"></span></th>
+<th data-col="2" scope="col">Why not reachable<span class="arrow"></span></th>
+<th data-col="3" scope="col" class="num">Downloads (30 days)<span class="arrow"></span></th>
+<th data-col="4" scope="col">Top checkpoint<span class="arrow"></span></th>
+<th scope="col" title="Source checkpoint gated on the Hub (auto or manual approval)">Gated</th>
+<th scope="col" title="How the fixture was shrunk (tmic shrink)">Shrink</th>
+</tr></thead>
+<tbody id="rows"></tbody>
+</table></div>
+<p class="foot" id="foot"></p>
+</main>
+<script>
+(async () => {
+  const {rows, meta, shrink = {}, gated = {}} = await (await fetch('data.json')).json();
+  const STATES = ['Done', 'Reachable', 'Unreachable', 'No checkpoint'];
+  const cls = s => s.replace(' ', '');
+  const ORDER = {Done: 0, Reachable: 1, Unreachable: 2, 'No checkpoint': 3};
+  const GOAL_SHARE = 0.3;
+  const total = rows.length, goalSize = Math.ceil(total * GOAL_SHARE);
+  // Rank by downloads (ties by name); the top 30% are the goal.
+  [...rows].sort((a, b) => b[3] - a[3] || a[0].localeCompare(b[0])).forEach((r, i) => { r[5] = i + 1; r[6] = i < goalSize; });
+  const tally = list => { const c = {Done: 0, Reachable: 0, Unreachable: 0, 'No checkpoint': 0}; list.forEach(r => c[r[1]]++); return c; };
+  const all = tally(rows), goal = tally(rows.filter(r => r[6]));
+  const pct = (n, of) => (100 * n / of).toFixed(1) + '%';
+  const covered = c => c.Done + c.Reachable;
+
+  document.getElementById('lead').textContent =
+    `${all.Done} of ${total} model types have a MIC contract; ${covered(all)} (${pct(covered(all), total)}) are done or reachable ` +
+    `with the generic shrink rule. Goal: cover the ${goalSize} most downloaded model types (the top ${GOAL_SHARE * 100}%), ` +
+    `marked in purple below. ${all['No checkpoint']} model types have no checkpoint on the Hub and cannot be treated. Click a column to sort; click again to reverse.`;
+
+  const bars = document.getElementById('bars');
+  function bar(title, counts, of, target) {
+    const box = document.createElement('div');
+    const head = document.createElement('div'); head.className = 'bar-head';
+    const t = document.createElement('strong'); t.textContent = title;
+    const s = document.createElement('span');
+    s.textContent = `${counts.Done} done · ${counts.Reachable} reachable · ${counts.Unreachable} unreachable · ${counts['No checkpoint']} no checkpoint — ${pct(covered(counts), of)} done or reachable`;
+    head.append(t, s);
+    const b = document.createElement('div'); b.className = 'bar'; b.setAttribute('role', 'img');
+    b.setAttribute('aria-label', `${title}: ${s.textContent}`);
+    for (const k of STATES) {
+      const seg = document.createElement('i'); seg.className = cls(k); seg.style.flexGrow = counts[k];
+      seg.title = `${k}: ${counts[k]} (${pct(counts[k], of)})`;
+      b.appendChild(seg);
+    }
+    if (target) {
+      const m = document.createElement('span'); m.className = 'mark'; m.style.left = `calc(${target * 100}% - 1px)`;
+      m.title = `${target * 100}% target`; b.appendChild(m);
+    }
+    box.append(head, b);
+    bars.appendChild(box);
+  }
+  bar(`Goal: top ${GOAL_SHARE * 100}% most downloaded (${goalSize} model types)`, goal, goalSize);
+  bar(`All ${total} model types`, all, total);
+  const legend = document.createElement('div'); legend.className = 'legend';
+  legend.innerHTML = '<span><b style="background:var(--done)"></b>Done: has a MIC contract</span>' +
+    '<span><b style="background:var(--reach)"></b>Reachable: generic shrink rule builds and runs it</span>' +
+    '<span><b style="background:var(--no);opacity:.55"></b>Unreachable</span>' +
+    '<span><b style="background:var(--none)"></b>No checkpoint: no Hub checkpoint has this model type, so it cannot be treated</span>';
+  bars.appendChild(legend);
+
+  let sort = {col: 3, dir: -1}, only = '';
+  const stats = document.getElementById('stats');
+  const filters = [['', 'All model types', total], ['goal', `Top ${GOAL_SHARE * 100}% (goal)`, goalSize], ...STATES.map(s => [s, s, all[s]])];
+  for (const [v, label, n] of filters) {
+    const b = document.createElement('button');
+    b.className = 'stat' + (v === 'goal' ? ' goal' : ''); b.type = 'button'; b.dataset.v = v;
+    b.innerHTML = `<strong>${n}</strong>${label}`;
+    b.onclick = () => { only = v; render(); };
+    stats.appendChild(b);
+  }
+  const cmp = (a, b) => {
+    const c = sort.col;
+    let d;
+    if (c === 3 || c === 5) d = a[c] - b[c];
+    else if (c === 1) d = ORDER[a[1]] - ORDER[b[1]] || b[3] - a[3];
+    else d = String(a[c]).localeCompare(String(b[c]), undefined, {sensitivity: 'base'});
+    return d * sort.dir || a[0].localeCompare(b[0]);
+  };
+  document.querySelectorAll('th[data-col]').forEach(th => th.onclick = () => {
+    const col = +th.dataset.col;
+    sort = sort.col === col ? {col, dir: -sort.dir} : {col, dir: col === 3 ? -1 : 1};
+    render();
+  });
+  const LABELS = [...document.querySelectorAll('thead th')].map(th => th.firstChild.textContent.trim());
+  const q = document.getElementById('q');
+  q.oninput = render;
+  const keep = r => !only || (only === 'goal' ? r[6] : r[1] === only);
+  function render() {
+    const f = q.value.trim().toLowerCase();
+    const shown = rows.filter(r => keep(r) && (!f || r.slice(0, 5).join(' ').toLowerCase().includes(f))).sort(cmp);
+    document.getElementById('rows').replaceChildren(...shown.map(r => {
+      const tr = document.createElement('tr');
+      if (r[6]) tr.className = 'goal' + (r[1] === 'Unreachable' ? ' miss' : '');
+      const td = (text, cls) => { const c = document.createElement('td'); if (cls) c.className = cls; c.textContent = text; tr.appendChild(c); return c; };
+      td(r[5], 'num');
+      const name = td(r[0]);
+      if (r[6]) { const b = document.createElement('span'); b.className = 'badge'; b.textContent = `Top ${GOAL_SHARE * 100}%`; name.appendChild(b); }
+      td(r[1], `state ${cls(r[1])}`);
+      td(r[2], 'why');
+      td(r[3].toLocaleString('en-US'), 'num');
+      const top = td('', 'top');
+      if (r[4]) {
+        const a = document.createElement('a');
+        a.href = `https://huggingface.co/${r[4]}`; a.target = '_blank'; a.rel = 'noopener'; a.textContent = r[4];
+        top.appendChild(a);
+      }
+      td(gated[r[0]] ? `🔒 ${gated[r[0]]}` : '', 'gated');
+      const view = td('', 'top');
+      if (r[1] === 'Done') for (const [page, findings] of shrink[r[0]] || []) {
+        const a = document.createElement('a');
+        a.href = `shrink/${page}.html`; a.target = '_blank'; a.rel = 'noopener';
+        a.textContent = findings ? `⚠ ${findings}` : 'view';
+        a.title = `${page}: ${findings ? findings + ' finding(s) to review' : 'same architecture, smaller'}`;
+        if (view.childNodes.length) view.append(' · ');
+        view.appendChild(a);
+      }
+      tr.querySelectorAll('td').forEach((c, i) => c.dataset.label = LABELS[i]);
+      return tr;
+    }));
+    document.querySelectorAll('th[data-col]').forEach(th => {
+      const on = +th.dataset.col === sort.col;
+      th.querySelector('.arrow').textContent = on ? (sort.dir > 0 ? '▲' : '▼') : '';
+      th.setAttribute('aria-sort', on ? (sort.dir > 0 ? 'ascending' : 'descending') : 'none');
+    });
+    stats.querySelectorAll('.stat').forEach(b => b.setAttribute('aria-pressed', String(b.dataset.v === only)));
+    document.getElementById('count').textContent = `${shown.length} of ${total} shown`;
+  }
+  const short = s => s.slice(0, 7);
+  document.getElementById('foot').innerHTML =
+    `Snapshot ${meta.generated}. Downloads = Hub 30-day count summed over each type's top 1,000 checkpoints whose ` +
+    `<code>config.model_type</code> matches. Sources: ` +
+    `<a href="https://huggingface.co/datasets/huggingface-mic/transformers-fleet/tree/${meta.fleet}">MIC fleet @${short(meta.fleet)}</a>, ` +
+    `<a href="https://github.com/huggingface/transformers/blob/${meta.transformers}/src/transformers/models/auto/auto_mappings.py">Transformers model types @${short(meta.transformers)}</a>, ` +
+    `<a href="https://github.com/huggingface/transformers-mic/blob/${meta.sweep}/experiments/results/sweep.json">shrink sweep @${short(meta.sweep)}</a>.`;
+  render();
+})();
+</script>
+</body>
 </html>

shrink/Apertus-v1.1-0.5B-Instruct-mic.html

@@ -0,0 +1,34 @@
+<!doctype html>
+<html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
+<title>Shrunk architecture</title>
+<style>
+:root{--bg:#fcfcfb;--fg:#0b0b0b;--muted:#52514e;--line:#e4e3df;--ok:#1d7a43;--warn:#b35c00;--s0:#2a78d6;--s1:#eb6834;--s2:#1baf7a;--s3:#eda100;--s4:#e87ba4;--s5:#008300;--s6:#6250d6;--s7:#e34948;--s8:#8a8984;}
+@media (prefers-color-scheme:dark){:root:not([data-theme="light"]){--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}}
+:root[data-theme="dark"]{--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}
+body{margin:0;background:var(--bg);color:var(--fg);font:15px/1.5 system-ui,sans-serif}
+main{max-width:1040px;margin:0 auto;padding:24px 16px 64px}
+h1{font-size:24px;margin:0 0 4px}h2{font-size:18px;margin:36px 0 8px}h3{font-size:14px;margin:0 0 8px;font-family:ui-monospace,monospace}
+.meta,.note{color:var(--muted);margin:0 0 12px;overflow-wrap:anywhere}
+.stats{display:grid;grid-template-columns:repeat(auto-fit,minmax(150px,1fr));gap:12px;margin:20px 0}
+.stat{border:1px solid var(--line);border-radius:8px;padding:12px;display:flex;flex-direction:column}
+.stat .v{font-size:26px;font-weight:600}.stat .k{color:var(--muted);font-size:13px}
+.findings{border-radius:8px;padding:4px 16px;border:1px solid var(--line)}
+.findings.ok{border-left:4px solid var(--ok)}.findings.warn{border-left:4px solid var(--warn)}
+.findings h2{margin:12px 0 4px;font-size:16px}
+.stack{border:1px solid var(--line);border-radius:8px;padding:12px;margin:12px 0}
+.row{display:flex;align-items:center;gap:12px;margin:6px 0}
+.lab{flex:0 0 110px;color:var(--muted);font-size:13px}
+.cells{display:flex;flex-wrap:wrap;gap:2px}
+.cells i{width:12px;height:12px;border-radius:2px;display:block}
+.bar{flex:1;display:flex;gap:2px;height:16px}.bar i{display:block;border-radius:2px;min-width:2px}
+table{border-collapse:collapse;width:100%;font-size:13px;margin:8px 0;display:block;overflow-x:auto}
+th,td{text-align:left;padding:4px 10px 4px 0;border-bottom:1px solid var(--line);vertical-align:top}
+th{color:var(--muted);font-weight:500}
+td:first-child{font-family:ui-monospace,monospace;overflow-wrap:anywhere}
+td i{display:inline-block;width:10px;height:10px;border-radius:2px;margin-right:6px}
+tr.bad td{color:var(--warn)}
+.c0{background:var(--s0)}.c1{background:var(--s1)}.c2{background:var(--s2)}.c3{background:var(--s3)}.c4{background:var(--s4)}.c5{background:var(--s5)}.c6{background:var(--s6)}.c7{background:var(--s7)}.c8{background:var(--s8)}
+</style></head><body><main>
+<h1>Shrunk architecture</h1><p class="meta">huggingface-mic/Apertus-v1.1-0.5B-Instruct-mic@55ae98cbccd9da534828bccc9b1f8bee4974731d ← swiss-ai/Apertus-v1.1-0.5B-Instruct@a140fd61fb57422c36a26301caea17edee799874 · ApertusForCausalLM · random weights</p>
+<div class="stats"><div class="stat"><span class="k">Original</span><span class="v">438.3M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Fixture</span><span class="v">8.4M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Layer stacks</span><span class="v">1</span><span class="k">compared</span></div><div class="stat"><span class="k">Findings</span><span class="v">0</span><span class="k">to review</span></div></div><section class="findings ok"><h2>✓ Same architecture, smaller</h2><p>Every module class, layer kind and checked property of the original is in the fixture.</p></section><h2>Layer stacks</h2><p class="note">One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.</p><div class="stack"><h3>model.layers</h3><div class="row"><span class="lab">Original · 20</span><div class="cells"><i class="c0" title="layer 0: DecoderLayer"></i><i class="c0" title="layer 1: DecoderLayer"></i><i class="c0" title="layer 2: DecoderLayer"></i><i class="c0" title="layer 3: DecoderLayer"></i><i class="c0" title="layer 4: DecoderLayer"></i><i class="c0" title="layer 5: DecoderLayer"></i><i class="c0" title="layer 6: DecoderLayer"></i><i class="c0" title="layer 7: DecoderLayer"></i><i class="c0" title="layer 8: DecoderLayer"></i><i class="c0" title="layer 9: DecoderLayer"></i><i class="c0" title="layer 10: DecoderLayer"></i><i class="c0" title="layer 11: DecoderLayer"></i><i class="c0" title="layer 12: DecoderLayer"></i><i class="c0" title="layer 13: DecoderLayer"></i><i class="c0" title="layer 14: DecoderLayer"></i><i class="c0" title="layer 15: DecoderLayer"></i><i class="c0" title="layer 16: DecoderLayer"></i><i class="c0" title="layer 17: DecoderLayer"></i><i class="c0" title="layer 18: DecoderLayer"></i><i class="c0" title="layer 19: DecoderLayer"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: DecoderLayer"></i><i class="c0" title="layer 1: DecoderLayer"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>DecoderLayer</td><td>20</td><td>2</td><td>✓ kept</td></tr></table></div><h2>Where the parameters are</h2><p class="note">Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.</p><div class="row"><span class="lab">Original</span><div class="bar"><i class="c0" style="flex:0.693809" title="model.layers: 304.1M (69.4%)"></i><i class="c1" style="flex:0.306189" title="model.embed_tokens: 134.2M (30.6%)"></i><i class="c2" style="flex:0.000002" title="model.norm: 1.0K (0.0%)"></i></div></div><div class="row"><span class="lab">Fixture</span><div class="bar"><i class="c0" style="flex:0.006828" title="model.layers: 57.7K (0.7%)"></i><i class="c1" style="flex:0.993165" title="model.embed_tokens: 8.4M (99.3%)"></i><i class="c2" style="flex:0.000008" title="model.norm: 64 (0.0%)"></i></div></div><table class="legend"><tr><th>Component</th><th>Original</th><th>Share</th><th>Fixture</th><th>Share</th></tr><tr><td><i class="c0"></i>model.layers</td><td>304.1M</td><td>69.4%</td><td>57.7K</td><td>0.7%</td></tr><tr><td><i class="c1"></i>model.embed_tokens</td><td>134.2M</td><td>30.6%</td><td>8.4M</td><td>99.3%</td></tr><tr><td><i class="c2"></i>model.norm</td><td>1.0K</td><td>0.0%</td><td>64</td><td>0.0%</td></tr></table><h2>Preserved properties</h2><table><tr><th>Config</th><th>Property</th><th>Original</th><th>Fixture</th><th></th></tr><tr class=""><td>model</td><td>attention style</td><td>GQA (16 heads, 4 KV)</td><td>GQA (4 heads, 2 KV)</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>RoPE type</td><td>linear</td><td>linear</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>model_type</td><td>apertus</td><td>apertus</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>vocab_size</td><td>131072</td><td>131072</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>tie_word_embeddings</td><td>True</td><td>True</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>hidden_act</td><td>xielu</td><td>xielu</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>max_position_embeddings</td><td>4096</td><td>4096</td><td>✓ kept</td></tr></table><h2>Module classes</h2><p class="note">Instances of each class in the built model (prefix Apertus omitted).</p><table><tr><th>Class</th><th>Original</th><th>Fixture</th></tr><tr class=""><td>Linear</td><td>121</td><td>13</td></tr><tr class=""><td>RMSNorm</td><td>81</td><td>9</td></tr><tr class=""><td>Attention</td><td>20</td><td>2</td></tr><tr class=""><td>DecoderLayer</td><td>20</td><td>2</td></tr><tr class=""><td>MLP</td><td>20</td><td>2</td></tr><tr class=""><td>XIELUActivation</td><td>20</td><td>2</td></tr><tr class=""><td>ForCausalLM</td><td>1</td><td>1</td></tr><tr class=""><td>Model</td><td>1</td><td>1</td></tr><tr class=""><td>RotaryEmbedding</td><td>1</td><td>1</td></tr><tr class=""><td>Embedding</td><td>1</td><td>1</td></tr><tr class=""><td>ModuleList</td><td>1</td><td>1</td></tr></table><h2>Config changes</h2><p class="note">5 fields changed, 26 unchanged.</p><table><tr><th>Field</th><th>Original</th><th>Fixture</th></tr><tr><td>hidden_size</td><td>1024</td><td>64</td></tr><tr><td>intermediate_size</td><td>6144</td><td>128</td></tr><tr><td>num_attention_heads</td><td>16</td><td>4</td></tr><tr><td>num_hidden_layers</td><td>20</td><td>2</td></tr><tr><td>num_key_value_heads</td><td>4</td><td>2</td></tr></table>
+</main></body></html>

shrink/Cosmos3-Nano-mic.html

@@ -0,0 +1,34 @@
+<!doctype html>
+<html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
+<title>Shrunk architecture</title>
+<style>
+:root{--bg:#fcfcfb;--fg:#0b0b0b;--muted:#52514e;--line:#e4e3df;--ok:#1d7a43;--warn:#b35c00;--s0:#2a78d6;--s1:#eb6834;--s2:#1baf7a;--s3:#eda100;--s4:#e87ba4;--s5:#008300;--s6:#6250d6;--s7:#e34948;--s8:#8a8984;}
+@media (prefers-color-scheme:dark){:root:not([data-theme="light"]){--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}}
+:root[data-theme="dark"]{--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}
+body{margin:0;background:var(--bg);color:var(--fg);font:15px/1.5 system-ui,sans-serif}
+main{max-width:1040px;margin:0 auto;padding:24px 16px 64px}
+h1{font-size:24px;margin:0 0 4px}h2{font-size:18px;margin:36px 0 8px}h3{font-size:14px;margin:0 0 8px;font-family:ui-monospace,monospace}
+.meta,.note{color:var(--muted);margin:0 0 12px;overflow-wrap:anywhere}
+.stats{display:grid;grid-template-columns:repeat(auto-fit,minmax(150px,1fr));gap:12px;margin:20px 0}
+.stat{border:1px solid var(--line);border-radius:8px;padding:12px;display:flex;flex-direction:column}
+.stat .v{font-size:26px;font-weight:600}.stat .k{color:var(--muted);font-size:13px}
+.findings{border-radius:8px;padding:4px 16px;border:1px solid var(--line)}
+.findings.ok{border-left:4px solid var(--ok)}.findings.warn{border-left:4px solid var(--warn)}
+.findings h2{margin:12px 0 4px;font-size:16px}
+.stack{border:1px solid var(--line);border-radius:8px;padding:12px;margin:12px 0}
+.row{display:flex;align-items:center;gap:12px;margin:6px 0}
+.lab{flex:0 0 110px;color:var(--muted);font-size:13px}
+.cells{display:flex;flex-wrap:wrap;gap:2px}
+.cells i{width:12px;height:12px;border-radius:2px;display:block}
+.bar{flex:1;display:flex;gap:2px;height:16px}.bar i{display:block;border-radius:2px;min-width:2px}
+table{border-collapse:collapse;width:100%;font-size:13px;margin:8px 0;display:block;overflow-x:auto}
+th,td{text-align:left;padding:4px 10px 4px 0;border-bottom:1px solid var(--line);vertical-align:top}
+th{color:var(--muted);font-weight:500}
+td:first-child{font-family:ui-monospace,monospace;overflow-wrap:anywhere}
+td i{display:inline-block;width:10px;height:10px;border-radius:2px;margin-right:6px}
+tr.bad td{color:var(--warn)}
+.c0{background:var(--s0)}.c1{background:var(--s1)}.c2{background:var(--s2)}.c3{background:var(--s3)}.c4{background:var(--s4)}.c5{background:var(--s5)}.c6{background:var(--s6)}.c7{background:var(--s7)}.c8{background:var(--s8)}
+</style></head><body><main>
+<h1>Shrunk architecture</h1><p class="meta">huggingface-mic/Cosmos3-Nano-mic@4c4a117fd8d6b13993505f86a24b15580e8b25d6 ← nvidia/Cosmos3-Nano@e59a53c25979a090fa8706c9acc0c254a6e89b92 · Cosmos3OmniForConditionalGeneration · random weights</p>
+<div class="stats"><div class="stat"><span class="k">Original</span><span class="v">8.8B</span><span class="k">parameters</span></div><div class="stat"><span class="k">Fixture</span><span class="v">20.2M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Layer stacks</span><span class="v">3</span><span class="k">compared</span></div><div class="stat"><span class="k">Findings</span><span class="v">0</span><span class="k">to review</span></div></div><section class="findings ok"><h2>✓ Same architecture, smaller</h2><p>Every module class, layer kind and checked property of the original is in the fixture.</p></section><h2>Layer stacks</h2><p class="note">One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.</p><div class="stack"><h3>model.visual.blocks</h3><div class="row"><span class="lab">Original · 27</span><div class="cells"><i class="c0" title="layer 0: Qwen3VLVisionBlock"></i><i class="c0" title="layer 1: Qwen3VLVisionBlock"></i><i class="c0" title="layer 2: Qwen3VLVisionBlock"></i><i class="c0" title="layer 3: Qwen3VLVisionBlock"></i><i class="c0" title="layer 4: Qwen3VLVisionBlock"></i><i class="c0" title="layer 5: Qwen3VLVisionBlock"></i><i class="c0" title="layer 6: Qwen3VLVisionBlock"></i><i class="c0" title="layer 7: Qwen3VLVisionBlock"></i><i class="c0" title="layer 8: Qwen3VLVisionBlock"></i><i class="c0" title="layer 9: Qwen3VLVisionBlock"></i><i class="c0" title="layer 10: Qwen3VLVisionBlock"></i><i class="c0" title="layer 11: Qwen3VLVisionBlock"></i><i class="c0" title="layer 12: Qwen3VLVisionBlock"></i><i class="c0" title="layer 13: Qwen3VLVisionBlock"></i><i class="c0" title="layer 14: Qwen3VLVisionBlock"></i><i class="c0" title="layer 15: Qwen3VLVisionBlock"></i><i class="c0" title="layer 16: Qwen3VLVisionBlock"></i><i class="c0" title="layer 17: Qwen3VLVisionBlock"></i><i class="c0" title="layer 18: Qwen3VLVisionBlock"></i><i class="c0" title="layer 19: Qwen3VLVisionBlock"></i><i class="c0" title="layer 20: Qwen3VLVisionBlock"></i><i class="c0" title="layer 21: Qwen3VLVisionBlock"></i><i class="c0" title="layer 22: Qwen3VLVisionBlock"></i><i class="c0" title="layer 23: Qwen3VLVisionBlock"></i><i class="c0" title="layer 24: Qwen3VLVisionBlock"></i><i class="c0" title="layer 25: Qwen3VLVisionBlock"></i><i class="c0" title="layer 26: Qwen3VLVisionBlock"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: Qwen3VLVisionBlock"></i><i class="c0" title="layer 1: Qwen3VLVisionBlock"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>Qwen3VLVisionBlock</td><td>27</td><td>2</td><td>✓ kept</td></tr></table></div><div class="stack"><h3>model.visual.deepstack_merger_list</h3><div class="row"><span class="lab">Original · 3</span><div class="cells"><i class="c0" title="layer 0: GELU"></i><i class="c0" title="layer 1: GELU"></i><i class="c0" title="layer 2: GELU"></i></div></div><div class="row"><span class="lab">Fixture · 3</span><div class="cells"><i class="c0" title="layer 0: GELU"></i><i class="c0" title="layer 1: GELU"></i><i class="c0" title="layer 2: GELU"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>GELU</td><td>3</td><td>3</td><td>✓ kept</td></tr></table></div><div class="stack"><h3>model.language_model.layers</h3><div class="row"><span class="lab">Original · 36</span><div class="cells"><i class="c0" title="layer 0: Qwen3VLTextDecoderLayer"></i><i class="c0" title="layer 1: Qwen3VLTextDecoderLayer"></i><i class="c0" title="layer 2: Qwen3VLTextDecoderLayer"></i><i class="c0" title="layer 3: Qwen3VLTextDecoderLayer"></i><i class="c0" title="layer 4: Qwen3VLTextDecoderLayer"></i><i class="c0" title="layer 5: Qwen3VLTextDecoderLayer"></i><i class="c0" title="layer 6: Qwen3VLTextDecoderLayer"></i><i class="c0" title="layer 7: Qwen3VLTextDecoderLayer"></i><i class="c0" title="layer 8: Qwen3VLTextDecoderLayer"></i><i class="c0" title="layer 9: Qwen3VLTextDecoderLayer"></i><i class="c0" title="layer 10: Qwen3VLTextDecoderLayer"></i><i class="c0" title="layer 11: Qwen3VLTextDecoderLayer"></i><i class="c0" title="layer 12: Qwen3VLTextDecoderLayer"></i><i class="c0" title="layer 13: Qwen3VLTextDecoderLayer"></i><i class="c0" title="layer 14: Qwen3VLTextDecoderLayer"></i><i class="c0" title="layer 15: Qwen3VLTextDecoderLayer"></i><i class="c0" title="layer 16: Qwen3VLTextDecoderLayer"></i><i class="c0" title="layer 17: Qwen3VLTextDecoderLayer"></i><i class="c0" title="layer 18: Qwen3VLTextDecoderLayer"></i><i class="c0" title="layer 19: Qwen3VLTextDecoderLayer"></i><i class="c0" title="layer 20: Qwen3VLTextDecoderLayer"></i><i class="c0" title="layer 21: Qwen3VLTextDecoderLayer"></i><i class="c0" title="layer 22: Qwen3VLTextDecoderLayer"></i><i class="c0" title="layer 23: Qwen3VLTextDecoderLayer"></i><i class="c0" title="layer 24: Qwen3VLTextDecoderLayer"></i><i class="c0" title="layer 25: Qwen3VLTextDecoderLayer"></i><i class="c0" title="layer 26: Qwen3VLTextDecoderLayer"></i><i class="c0" title="layer 27: Qwen3VLTextDecoderLayer"></i><i class="c0" title="layer 28: Qwen3VLTextDecoderLayer"></i><i class="c0" title="layer 29: Qwen3VLTextDecoderLayer"></i><i class="c0" title="layer 30: Qwen3VLTextDecoderLayer"></i><i class="c0" title="layer 31: Qwen3VLTextDecoderLayer"></i><i class="c0" title="layer 32: Qwen3VLTextDecoderLayer"></i><i class="c0" title="layer 33: Qwen3VLTextDecoderLayer"></i><i class="c0" title="layer 34: Qwen3VLTextDecoderLayer"></i><i class="c0" title="layer 35: Qwen3VLTextDecoderLayer"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: Qwen3VLTextDecoderLayer"></i><i class="c0" title="layer 1: Qwen3VLTextDecoderLayer"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>Qwen3VLTextDecoderLayer</td><td>36</td><td>2</td><td>✓ kept</td></tr></table></div><h2>Where the parameters are</h2><p class="note">Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.</p><div class="row"><span class="lab">Original</span><div class="bar"><i class="c0" style="flex:0.792286" title="model.language_model.layers: 6.9B (79.2%)"></i><i class="c1" style="flex:0.070985" title="model.language_model: 622.3M (7.1%)"></i><i class="c2" style="flex:0.070984" title="lm_head: 622.3M (7.1%)"></i><i class="c3" style="flex:0.046933" title="model.visual.blocks: 411.5M (4.7%)"></i><i class="c4" style="flex:0.013731" title="model.visual.deepstack_merger_list: 120.4M (1.4%)"></i><i class="c5" style="flex:0.005081" title="model.visual: 44.5M (0.5%)"></i></div></div><div class="row"><span class="lab">Fixture</span><div class="bar"><i class="c0" style="flex:0.003672" title="model.language_model.layers: 74.0K (0.4%)"></i><i class="c1" style="flex:0.482213" title="model.language_model: 9.7M (48.2%)"></i><i class="c2" style="flex:0.482209" title="lm_head: 9.7M (48.2%)"></i><i class="c3" style="flex:0.003320" title="model.visual.blocks: 66.9K (0.3%)"></i><i class="c4" style="flex:0.012311" title="model.visual.deepstack_merger_list: 248.3K (1.2%)"></i><i class="c5" style="flex:0.016275" title="model.visual: 328.2K (1.6%)"></i></div></div><table class="legend"><tr><th>Component</th><th>Original</th><th>Share</th><th>Fixture</th><th>Share</th></tr><tr><td><i class="c0"></i>model.language_model.layers</td><td>6.9B</td><td>79.2%</td><td>74.0K</td><td>0.4%</td></tr><tr><td><i class="c1"></i>model.language_model</td><td>622.3M</td><td>7.1%</td><td>9.7M</td><td>48.2%</td></tr><tr><td><i class="c2"></i>lm_head</td><td>622.3M</td><td>7.1%</td><td>9.7M</td><td>48.2%</td></tr><tr><td><i class="c3"></i>model.visual.blocks</td><td>411.5M</td><td>4.7%</td><td>66.9K</td><td>0.3%</td></tr><tr><td><i class="c4"></i>model.visual.deepstack_merger_list</td><td>120.4M</td><td>1.4%</td><td>248.3K</td><td>1.2%</td></tr><tr><td><i class="c5"></i>model.visual</td><td>44.5M</td><td>0.5%</td><td>328.2K</td><td>1.6%</td></tr></table><h2>Preserved properties</h2><table><tr><th>Config</th><th>Property</th><th>Original</th><th>Fixture</th><th></th></tr><tr class=""><td>model</td><td>model_type</td><td>cosmos3_omni</td><td>cosmos3_omni</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>tie_word_embeddings</td><td>False</td><td>False</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>attention style</td><td>GQA (32 heads, 8 KV)</td><td>GQA (4 heads, 2 KV)</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>RoPE type</td><td>default</td><td>default</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>model_type</td><td>qwen3_vl_text</td><td>qwen3_vl_text</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>vocab_size</td><td>151936</td><td>151936</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>hidden_act</td><td>silu</td><td>silu</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>max_position_embeddings</td><td>262144</td><td>262144</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>RoPE type</td><td>axial</td><td>axial</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>model_type</td><td>qwen3_vl_vision</td><td>qwen3_vl_vision</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>hidden_act</td><td>gelu_pytorch_tanh</td><td>gelu_pytorch_tanh</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>patch_size</td><td>16</td><td>16</td><td>✓ kept</td></tr></table><h2>Module classes</h2><p class="note">Instances of each class in the built model (prefix Cosmos3Omni omitted).</p><table><tr><th>Class</th><th>Original</th><th>Fixture</th></tr><tr class=""><td>Linear</td><td>369</td><td>31</td></tr><tr class=""><td>Qwen3VLTextRMSNorm</td><td>145</td><td>9</td></tr><tr class=""><td>LayerNorm</td><td>58</td><td>8</td></tr><tr class=""><td>Qwen3VLTextAttention</td><td>36</td><td>2</td></tr><tr class=""><td>Qwen3VLTextDecoderLayer</td><td>36</td><td>2</td></tr><tr class=""><td>Qwen3VLTextMLP</td><td>36</td><td>2</td></tr><tr class=""><td>SiLUActivation</td><td>36</td><td>2</td></tr><tr class=""><td>GELUTanh</td><td>27</td><td>2</td></tr><tr class=""><td>Qwen3VLVisionAttention</td><td>27</td><td>2</td></tr><tr class=""><td>Qwen3VLVisionBlock</td><td>27</td><td>2</td></tr><tr class=""><td>Qwen3VLVisionMLP</td><td>27</td><td>2</td></tr><tr class=""><td>GELU</td><td>4</td><td>4</td></tr><tr class=""><td>Qwen3VLVisionPatchMerger</td><td>4</td><td>4</td></tr><tr class=""><td>ModuleList</td><td>3</td><td>3</td></tr><tr class=""><td>Embedding</td><td>2</td><td>2</td></tr><tr class=""><td>Conv3d</td><td>1</td><td>1</td></tr><tr class=""><td>ForConditionalGeneration</td><td>1</td><td>1</td></tr><tr class=""><td>Model</td><td>1</td><td>1</td></tr><tr class=""><td>Qwen3VLTextModel</td><td>1</td><td>1</td></tr><tr class=""><td>Qwen3VLTextRotaryEmbedding</td><td>1</td><td>1</td></tr><tr class=""><td>Qwen3VLVisionModel</td><td>1</td><td>1</td></tr><tr class=""><td>Qwen3VLVisionPatchEmbed</td><td>1</td><td>1</td></tr><tr class=""><td>Qwen3VLVisionRotaryEmbedding</td><td>1</td><td>1</td></tr></table><h2>Config changes</h2><p class="note">12 fields changed, 54 unchanged.</p><table><tr><th>Field</th><th>Original</th><th>Fixture</th></tr><tr><td>text_config.head_dim</td><td>128</td><td>16</td></tr><tr><td>text_config.hidden_size</td><td>4096</td><td>64</td></tr><tr><td>text_config.intermediate_size</td><td>12288</td><td>128</td></tr><tr><td>text_config.num_attention_heads</td><td>32</td><td>4</td></tr><tr><td>text_config.num_hidden_layers</td><td>36</td><td>2</td></tr><tr><td>text_config.num_key_value_heads</td><td>8</td><td>2</td></tr><tr><td>text_config.rope_parameters</td><td>{&quot;mrope_interleaved&quot;: true, &quot;mrope_section&quot;: [24, 20, 20], &quot;rope_type&quot;: &quot;defa…</td><td>{&quot;mrope_interleaved&quot;: true, &quot;mrope_section&quot;: [3, 3, 2], &quot;rope_theta&quot;: 5000000…</td></tr><tr><td>vision_config.depth</td><td>27</td><td>2</td></tr><tr><td>vision_config.hidden_size</td><td>1152</td><td>64</td></tr><tr><td>vision_config.intermediate_size</td><td>4304</td><td>128</td></tr><tr><td>vision_config.num_heads</td><td>16</td><td>4</td></tr><tr><td>vision_config.out_hidden_size</td><td>4096</td><td>64</td></tr></table>
+</main></body></html>

shrink/DeepSeek-Coder-V2-Lite-Instruct-mic.html

@@ -0,0 +1,34 @@
+<!doctype html>
+<html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
+<title>Shrunk architecture</title>
+<style>
+:root{--bg:#fcfcfb;--fg:#0b0b0b;--muted:#52514e;--line:#e4e3df;--ok:#1d7a43;--warn:#b35c00;--s0:#2a78d6;--s1:#eb6834;--s2:#1baf7a;--s3:#eda100;--s4:#e87ba4;--s5:#008300;--s6:#6250d6;--s7:#e34948;--s8:#8a8984;}
+@media (prefers-color-scheme:dark){:root:not([data-theme="light"]){--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}}
+:root[data-theme="dark"]{--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}
+body{margin:0;background:var(--bg);color:var(--fg);font:15px/1.5 system-ui,sans-serif}
+main{max-width:1040px;margin:0 auto;padding:24px 16px 64px}
+h1{font-size:24px;margin:0 0 4px}h2{font-size:18px;margin:36px 0 8px}h3{font-size:14px;margin:0 0 8px;font-family:ui-monospace,monospace}
+.meta,.note{color:var(--muted);margin:0 0 12px;overflow-wrap:anywhere}
+.stats{display:grid;grid-template-columns:repeat(auto-fit,minmax(150px,1fr));gap:12px;margin:20px 0}
+.stat{border:1px solid var(--line);border-radius:8px;padding:12px;display:flex;flex-direction:column}
+.stat .v{font-size:26px;font-weight:600}.stat .k{color:var(--muted);font-size:13px}
+.findings{border-radius:8px;padding:4px 16px;border:1px solid var(--line)}
+.findings.ok{border-left:4px solid var(--ok)}.findings.warn{border-left:4px solid var(--warn)}
+.findings h2{margin:12px 0 4px;font-size:16px}
+.stack{border:1px solid var(--line);border-radius:8px;padding:12px;margin:12px 0}
+.row{display:flex;align-items:center;gap:12px;margin:6px 0}
+.lab{flex:0 0 110px;color:var(--muted);font-size:13px}
+.cells{display:flex;flex-wrap:wrap;gap:2px}
+.cells i{width:12px;height:12px;border-radius:2px;display:block}
+.bar{flex:1;display:flex;gap:2px;height:16px}.bar i{display:block;border-radius:2px;min-width:2px}
+table{border-collapse:collapse;width:100%;font-size:13px;margin:8px 0;display:block;overflow-x:auto}
+th,td{text-align:left;padding:4px 10px 4px 0;border-bottom:1px solid var(--line);vertical-align:top}
+th{color:var(--muted);font-weight:500}
+td:first-child{font-family:ui-monospace,monospace;overflow-wrap:anywhere}
+td i{display:inline-block;width:10px;height:10px;border-radius:2px;margin-right:6px}
+tr.bad td{color:var(--warn)}
+.c0{background:var(--s0)}.c1{background:var(--s1)}.c2{background:var(--s2)}.c3{background:var(--s3)}.c4{background:var(--s4)}.c5{background:var(--s5)}.c6{background:var(--s6)}.c7{background:var(--s7)}.c8{background:var(--s8)}
+</style></head><body><main>
+<h1>Shrunk architecture</h1><p class="meta">huggingface-mic/DeepSeek-Coder-V2-Lite-Instruct-mic@90601ffd1320fd0db167a3ade1c6cd0ed35c2229 ← deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct@e434a23f91ba5b4923cf6c9d9a238eb4a08e3a11 · DeepseekV2ForCausalLM · random weights</p>
+<div class="stats"><div class="stat"><span class="k">Original</span><span class="v">15.7B</span><span class="k">parameters</span></div><div class="stat"><span class="k">Fixture</span><span class="v">13.2M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Layer stacks</span><span class="v">1</span><span class="k">compared</span></div><div class="stat"><span class="k">Findings</span><span class="v">0</span><span class="k">to review</span></div></div><section class="findings ok"><h2>✓ Same architecture, smaller</h2><p>Every module class, layer kind and checked property of the original is in the fixture.</p></section><h2>Layer stacks</h2><p class="note">One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.</p><div class="stack"><h3>model.layers</h3><div class="row"><span class="lab">Original · 27</span><div class="cells"><i class="c0" title="layer 0: no Experts"></i><i class="c1" title="layer 1: Experts, Moe, TopkRouter"></i><i class="c1" title="layer 2: Experts, Moe, TopkRouter"></i><i class="c1" title="layer 3: Experts, Moe, TopkRouter"></i><i class="c1" title="layer 4: Experts, Moe, TopkRouter"></i><i class="c1" title="layer 5: Experts, Moe, TopkRouter"></i><i class="c1" title="layer 6: Experts, Moe, TopkRouter"></i><i class="c1" title="layer 7: Experts, Moe, TopkRouter"></i><i class="c1" title="layer 8: Experts, Moe, TopkRouter"></i><i class="c1" title="layer 9: Experts, Moe, TopkRouter"></i><i class="c1" title="layer 10: Experts, Moe, TopkRouter"></i><i class="c1" title="layer 11: Experts, Moe, TopkRouter"></i><i class="c1" title="layer 12: Experts, Moe, TopkRouter"></i><i class="c1" title="layer 13: Experts, Moe, TopkRouter"></i><i class="c1" title="layer 14: Experts, Moe, TopkRouter"></i><i class="c1" title="layer 15: Experts, Moe, TopkRouter"></i><i class="c1" title="layer 16: Experts, Moe, TopkRouter"></i><i class="c1" title="layer 17: Experts, Moe, TopkRouter"></i><i class="c1" title="layer 18: Experts, Moe, TopkRouter"></i><i class="c1" title="layer 19: Experts, Moe, TopkRouter"></i><i class="c1" title="layer 20: Experts, Moe, TopkRouter"></i><i class="c1" title="layer 21: Experts, Moe, TopkRouter"></i><i class="c1" title="layer 22: Experts, Moe, TopkRouter"></i><i class="c1" title="layer 23: Experts, Moe, TopkRouter"></i><i class="c1" title="layer 24: Experts, Moe, TopkRouter"></i><i class="c1" title="layer 25: Experts, Moe, TopkRouter"></i><i class="c1" title="layer 26: Experts, Moe, TopkRouter"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: no Experts"></i><i class="c1" title="layer 1: Experts, Moe, TopkRouter"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>no Experts</td><td>1</td><td>1</td><td>✓ kept</td></tr><tr><td><i class="c1"></i>Experts, Moe, TopkRouter</td><td>26</td><td>1</td><td>✓ kept</td></tr></table></div><h2>Where the parameters are</h2><p class="note">Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.</p><div class="row"><span class="lab">Original</span><div class="bar"><i class="c0" style="flex:0.973296" title="model.layers: 15.3B (97.3%)"></i><i class="c1" style="flex:0.013352" title="lm_head: 209.7M (1.3%)"></i><i class="c2" style="flex:0.013352" title="model.embed_tokens: 209.7M (1.3%)"></i><i class="c3" style="flex:0.000000" title="model.norm: 2.0K (0.0%)"></i></div></div><div class="row"><span class="lab">Fixture</span><div class="bar"><i class="c0" style="flex:0.006792" title="model.layers: 89.6K (0.7%)"></i><i class="c1" style="flex:0.496602" title="lm_head: 6.6M (49.7%)"></i><i class="c2" style="flex:0.496602" title="model.embed_tokens: 6.6M (49.7%)"></i><i class="c3" style="flex:0.000005" title="model.norm: 64 (0.0%)"></i></div></div><table class="legend"><tr><th>Component</th><th>Original</th><th>Share</th><th>Fixture</th><th>Share</th></tr><tr><td><i class="c0"></i>model.layers</td><td>15.3B</td><td>97.3%</td><td>89.6K</td><td>0.7%</td></tr><tr><td><i class="c1"></i>lm_head</td><td>209.7M</td><td>1.3%</td><td>6.6M</td><td>49.7%</td></tr><tr><td><i class="c2"></i>model.embed_tokens</td><td>209.7M</td><td>1.3%</td><td>6.6M</td><td>49.7%</td></tr><tr><td><i class="c3"></i>model.norm</td><td>2.0K</td><td>0.0%</td><td>64</td><td>0.0%</td></tr></table><h2>Preserved properties</h2><table><tr><th>Config</th><th>Property</th><th>Original</th><th>Fixture</th><th></th></tr><tr class=""><td>model</td><td>attention style</td><td>MHA (16 heads, 16 KV)</td><td>MHA (4 heads, 4 KV)</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>MoE routing</td><td>top-6 of 64 experts</td><td>top-2 of 4 experts</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>RoPE type</td><td>yarn</td><td>yarn</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>model_type</td><td>deepseek_v2</td><td>deepseek_v2</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>vocab_size</td><td>102400</td><td>102400</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>tie_word_embeddings</td><td>False</td><td>False</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>hidden_act</td><td>silu</td><td>silu</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>max_position_embeddings</td><td>163840</td><td>163840</td><td>✓ kept</td></tr></table><h2>Module classes</h2><p class="note">Instances of each class in the built model (prefix DeepseekV2 omitted).</p><table><tr><th>Class</th><th>Original</th><th>Fixture</th></tr><tr class=""><td>Linear</td><td>190</td><td>15</td></tr><tr class=""><td>RMSNorm</td><td>82</td><td>7</td></tr><tr class=""><td>SiLUActivation</td><td>53</td><td>3</td></tr><tr class=""><td>Attention</td><td>27</td><td>2</td></tr><tr class=""><td>DecoderLayer</td><td>27</td><td>2</td></tr><tr class=""><td>MLP</td><td>27</td><td>2</td></tr><tr class=""><td>Experts</td><td>26</td><td>1</td></tr><tr class=""><td>Moe</td><td>26</td><td>1</td></tr><tr class=""><td>TopkRouter</td><td>26</td><td>1</td></tr><tr class=""><td>ForCausalLM</td><td>1</td><td>1</td></tr><tr class=""><td>Model</td><td>1</td><td>1</td></tr><tr class=""><td>RotaryEmbedding</td><td>1</td><td>1</td></tr><tr class=""><td>Embedding</td><td>1</td><td>1</td></tr><tr class=""><td>ModuleList</td><td>1</td><td>1</td></tr></table><h2>Config changes</h2><p class="note">13 fields changed, 39 unchanged.</p><table><tr><th>Field</th><th>Original</th><th>Fixture</th></tr><tr><td>head_dim</td><td>64</td><td>8</td></tr><tr><td>hidden_size</td><td>2048</td><td>64</td></tr><tr><td>intermediate_size</td><td>10944</td><td>128</td></tr><tr><td>kv_lora_rank</td><td>512</td><td>16</td></tr><tr><td>moe_intermediate_size</td><td>1408</td><td>32</td></tr><tr><td>n_routed_experts</td><td>64</td><td>4</td></tr><tr><td>num_attention_heads</td><td>16</td><td>4</td></tr><tr><td>num_experts_per_tok</td><td>6</td><td>2</td></tr><tr><td>num_hidden_layers</td><td>27</td><td>2</td></tr><tr><td>num_key_value_heads</td><td>16</td><td>4</td></tr><tr><td>qk_nope_head_dim</td><td>128</td><td>16</td></tr><tr><td>qk_rope_head_dim</td><td>64</td><td>8</td></tr><tr><td>v_head_dim</td><td>128</td><td>16</td></tr></table>
+</main></body></html>

shrink/DeepSeek-R1-mic.html

@@ -0,0 +1,34 @@
+<!doctype html>
+<html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
+<title>Shrunk architecture</title>
+<style>
+:root{--bg:#fcfcfb;--fg:#0b0b0b;--muted:#52514e;--line:#e4e3df;--ok:#1d7a43;--warn:#b35c00;--s0:#2a78d6;--s1:#eb6834;--s2:#1baf7a;--s3:#eda100;--s4:#e87ba4;--s5:#008300;--s6:#6250d6;--s7:#e34948;--s8:#8a8984;}
+@media (prefers-color-scheme:dark){:root:not([data-theme="light"]){--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}}
+:root[data-theme="dark"]{--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}
+body{margin:0;background:var(--bg);color:var(--fg);font:15px/1.5 system-ui,sans-serif}
+main{max-width:1040px;margin:0 auto;padding:24px 16px 64px}
+h1{font-size:24px;margin:0 0 4px}h2{font-size:18px;margin:36px 0 8px}h3{font-size:14px;margin:0 0 8px;font-family:ui-monospace,monospace}
+.meta,.note{color:var(--muted);margin:0 0 12px;overflow-wrap:anywhere}
+.stats{display:grid;grid-template-columns:repeat(auto-fit,minmax(150px,1fr));gap:12px;margin:20px 0}
+.stat{border:1px solid var(--line);border-radius:8px;padding:12px;display:flex;flex-direction:column}
+.stat .v{font-size:26px;font-weight:600}.stat .k{color:var(--muted);font-size:13px}
+.findings{border-radius:8px;padding:4px 16px;border:1px solid var(--line)}
+.findings.ok{border-left:4px solid var(--ok)}.findings.warn{border-left:4px solid var(--warn)}
+.findings h2{margin:12px 0 4px;font-size:16px}
+.stack{border:1px solid var(--line);border-radius:8px;padding:12px;margin:12px 0}
+.row{display:flex;align-items:center;gap:12px;margin:6px 0}
+.lab{flex:0 0 110px;color:var(--muted);font-size:13px}
+.cells{display:flex;flex-wrap:wrap;gap:2px}
+.cells i{width:12px;height:12px;border-radius:2px;display:block}
+.bar{flex:1;display:flex;gap:2px;height:16px}.bar i{display:block;border-radius:2px;min-width:2px}
+table{border-collapse:collapse;width:100%;font-size:13px;margin:8px 0;display:block;overflow-x:auto}
+th,td{text-align:left;padding:4px 10px 4px 0;border-bottom:1px solid var(--line);vertical-align:top}
+th{color:var(--muted);font-weight:500}
+td:first-child{font-family:ui-monospace,monospace;overflow-wrap:anywhere}
+td i{display:inline-block;width:10px;height:10px;border-radius:2px;margin-right:6px}
+tr.bad td{color:var(--warn)}
+.c0{background:var(--s0)}.c1{background:var(--s1)}.c2{background:var(--s2)}.c3{background:var(--s3)}.c4{background:var(--s4)}.c5{background:var(--s5)}.c6{background:var(--s6)}.c7{background:var(--s7)}.c8{background:var(--s8)}
+</style></head><body><main>
+<h1>Shrunk architecture</h1><p class="meta">huggingface-mic/DeepSeek-R1-mic@89eddb3d8f9419e63b667eb7d5244fec8c768ed9 ← deepseek-ai/DeepSeek-R1@56d4cbbb4d29f4355bab4b9a39ccb717a14ad5ad · DeepseekV3ForCausalLM · random weights · built before the shrink rule (no recorded recipe)</p>
+<div class="stats"><div class="stat"><span class="k">Original</span><span class="v">671.0B</span><span class="k">parameters</span></div><div class="stat"><span class="k">Fixture</span><span class="v">16.6M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Layer stacks</span><span class="v">1</span><span class="k">compared</span></div><div class="stat"><span class="k">Findings</span><span class="v">0</span><span class="k">to review</span></div></div><section class="findings ok"><h2>✓ Same architecture, smaller</h2><p>Every module class, layer kind and checked property of the original is in the fixture.</p></section><h2>Layer stacks</h2><p class="note">One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.</p><div class="stack"><h3>model.layers</h3><div class="row"><span class="lab">Original · 61</span><div class="cells"><i class="c0" title="layer 0: no Experts"></i><i class="c0" title="layer 1: no Experts"></i><i class="c0" title="layer 2: no Experts"></i><i class="c1" title="layer 3: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 4: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 5: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 6: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 7: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 8: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 9: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 10: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 11: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 12: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 13: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 14: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 15: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 16: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 17: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 18: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 19: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 20: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 21: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 22: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 23: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 24: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 25: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 26: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 27: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 28: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 29: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 30: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 31: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 32: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 33: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 34: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 35: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 36: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 37: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 38: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 39: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 40: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 41: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 42: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 43: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 44: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 45: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 46: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 47: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 48: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 49: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 50: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 51: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 52: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 53: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 54: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 55: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 56: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 57: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 58: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 59: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 60: Experts, MoE, TopkRouter"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: no Experts"></i><i class="c1" title="layer 1: Experts, MoE, TopkRouter"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>no Experts</td><td>3</td><td>1</td><td>✓ kept</td></tr><tr><td><i class="c1"></i>Experts, MoE, TopkRouter</td><td>58</td><td>1</td><td>✓ kept</td></tr></table></div><h2>Where the parameters are</h2><p class="note">Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.</p><div class="row"><span class="lab">Original</span><div class="bar"><i class="c0" style="flex:0.997238" title="model.layers: 669.2B (99.7%)"></i><i class="c1" style="flex:0.001381" title="model.embed_tokens: 926.7M (0.1%)"></i><i class="c2" style="flex:0.001381" title="lm_head: 926.7M (0.1%)"></i><i class="c3" style="flex:0.000000" title="model.norm: 7.2K (0.0%)"></i></div></div><div class="row"><span class="lab">Fixture</span><div class="bar"><i class="c0" style="flex:0.004593" title="model.layers: 76.4K (0.5%)"></i><i class="c1" style="flex:0.497702" title="model.embed_tokens: 8.3M (49.8%)"></i><i class="c2" style="flex:0.497702" title="lm_head: 8.3M (49.8%)"></i><i class="c3" style="flex:0.000004" title="model.norm: 64 (0.0%)"></i></div></div><table class="legend"><tr><th>Component</th><th>Original</th><th>Share</th><th>Fixture</th><th>Share</th></tr><tr><td><i class="c0"></i>model.layers</td><td>669.2B</td><td>99.7%</td><td>76.4K</td><td>0.5%</td></tr><tr><td><i class="c1"></i>model.embed_tokens</td><td>926.7M</td><td>0.1%</td><td>8.3M</td><td>49.8%</td></tr><tr><td><i class="c2"></i>lm_head</td><td>926.7M</td><td>0.1%</td><td>8.3M</td><td>49.8%</td></tr><tr><td><i class="c3"></i>model.norm</td><td>7.2K</td><td>0.0%</td><td>64</td><td>0.0%</td></tr></table><h2>Preserved properties</h2><table><tr><th>Config</th><th>Property</th><th>Original</th><th>Fixture</th><th></th></tr><tr class=""><td>model</td><td>attention style</td><td>MHA (128 heads, 128 KV)</td><td>MHA (4 heads, 4 KV)</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>MoE routing</td><td>top-8 of 256 experts</td><td>top-2 of 4 experts</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>RoPE type</td><td>yarn</td><td>yarn</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>model_type</td><td>deepseek_v3</td><td>deepseek_v3</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>vocab_size</td><td>129280</td><td>129280</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>tie_word_embeddings</td><td>False</td><td>False</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>hidden_act</td><td>silu</td><td>silu</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>max_position_embeddings</td><td>163840</td><td>163840</td><td>✓ kept</td></tr></table><h2>Module classes</h2><p class="note">Instances of each class in the built model (prefix DeepseekV3 omitted).</p><table><tr><th>Class</th><th>Original</th><th>Fixture</th></tr><tr class=""><td>Linear</td><td>489</td><td>17</td></tr><tr class=""><td>RMSNorm</td><td>245</td><td>9</td></tr><tr class=""><td>SiLUActivation</td><td>119</td><td>3</td></tr><tr class=""><td>Attention</td><td>61</td><td>2</td></tr><tr class=""><td>DecoderLayer</td><td>61</td><td>2</td></tr><tr class=""><td>MLP</td><td>61</td><td>2</td></tr><tr class=""><td>Experts</td><td>58</td><td>1</td></tr><tr class=""><td>MoE</td><td>58</td><td>1</td></tr><tr class=""><td>TopkRouter</td><td>58</td><td>1</td></tr><tr class=""><td>ForCausalLM</td><td>1</td><td>1</td></tr><tr class=""><td>Model</td><td>1</td><td>1</td></tr><tr class=""><td>RotaryEmbedding</td><td>1</td><td>1</td></tr><tr class=""><td>Embedding</td><td>1</td><td>1</td></tr><tr class=""><td>ModuleList</td><td>1</td><td>1</td></tr></table><h2>Config changes</h2><p class="note">18 fields changed, 35 unchanged.</p><table><tr><th>Field</th><th>Original</th><th>Fixture</th></tr><tr><td>first_k_dense_replace</td><td>3</td><td>1</td></tr><tr><td>head_dim</td><td>64</td><td>8</td></tr><tr><td>hidden_size</td><td>7168</td><td>64</td></tr><tr><td>intermediate_size</td><td>18432</td><td>128</td></tr><tr><td>kv_lora_rank</td><td>512</td><td>16</td></tr><tr><td>moe_intermediate_size</td><td>2048</td><td>32</td></tr><tr><td>n_group</td><td>8</td><td>1</td></tr><tr><td>n_routed_experts</td><td>256</td><td>4</td></tr><tr><td>num_attention_heads</td><td>128</td><td>4</td></tr><tr><td>num_experts_per_tok</td><td>8</td><td>2</td></tr><tr><td>num_hidden_layers</td><td>61</td><td>2</td></tr><tr><td>num_key_value_heads</td><td>128</td><td>4</td></tr><tr><td>q_lora_rank</td><td>1536</td><td>16</td></tr><tr><td>qk_nope_head_dim</td><td>128</td><td>16</td></tr><tr><td>qk_rope_head_dim</td><td>64</td><td>8</td></tr><tr><td>quantization_config</td><td>{&quot;activation_scheme&quot;: &quot;dynamic&quot;, &quot;fmt&quot;: &quot;e4m3&quot;, &quot;quant_method&quot;: &quot;fp8&quot;, &quot;weigh…</td><td>—</td></tr><tr><td>topk_group</td><td>4</td><td>1</td></tr><tr><td>v_head_dim</td><td>128</td><td>16</td></tr></table>
+</main></body></html>

shrink/DeepSeek-V3.2-mic.html

@@ -0,0 +1,34 @@
+<!doctype html>
+<html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
+<title>Shrunk architecture</title>
+<style>
+:root{--bg:#fcfcfb;--fg:#0b0b0b;--muted:#52514e;--line:#e4e3df;--ok:#1d7a43;--warn:#b35c00;--s0:#2a78d6;--s1:#eb6834;--s2:#1baf7a;--s3:#eda100;--s4:#e87ba4;--s5:#008300;--s6:#6250d6;--s7:#e34948;--s8:#8a8984;}
+@media (prefers-color-scheme:dark){:root:not([data-theme="light"]){--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}}
+:root[data-theme="dark"]{--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}
+body{margin:0;background:var(--bg);color:var(--fg);font:15px/1.5 system-ui,sans-serif}
+main{max-width:1040px;margin:0 auto;padding:24px 16px 64px}
+h1{font-size:24px;margin:0 0 4px}h2{font-size:18px;margin:36px 0 8px}h3{font-size:14px;margin:0 0 8px;font-family:ui-monospace,monospace}
+.meta,.note{color:var(--muted);margin:0 0 12px;overflow-wrap:anywhere}
+.stats{display:grid;grid-template-columns:repeat(auto-fit,minmax(150px,1fr));gap:12px;margin:20px 0}
+.stat{border:1px solid var(--line);border-radius:8px;padding:12px;display:flex;flex-direction:column}
+.stat .v{font-size:26px;font-weight:600}.stat .k{color:var(--muted);font-size:13px}
+.findings{border-radius:8px;padding:4px 16px;border:1px solid var(--line)}
+.findings.ok{border-left:4px solid var(--ok)}.findings.warn{border-left:4px solid var(--warn)}
+.findings h2{margin:12px 0 4px;font-size:16px}
+.stack{border:1px solid var(--line);border-radius:8px;padding:12px;margin:12px 0}
+.row{display:flex;align-items:center;gap:12px;margin:6px 0}
+.lab{flex:0 0 110px;color:var(--muted);font-size:13px}
+.cells{display:flex;flex-wrap:wrap;gap:2px}
+.cells i{width:12px;height:12px;border-radius:2px;display:block}
+.bar{flex:1;display:flex;gap:2px;height:16px}.bar i{display:block;border-radius:2px;min-width:2px}
+table{border-collapse:collapse;width:100%;font-size:13px;margin:8px 0;display:block;overflow-x:auto}
+th,td{text-align:left;padding:4px 10px 4px 0;border-bottom:1px solid var(--line);vertical-align:top}
+th{color:var(--muted);font-weight:500}
+td:first-child{font-family:ui-monospace,monospace;overflow-wrap:anywhere}
+td i{display:inline-block;width:10px;height:10px;border-radius:2px;margin-right:6px}
+tr.bad td{color:var(--warn)}
+.c0{background:var(--s0)}.c1{background:var(--s1)}.c2{background:var(--s2)}.c3{background:var(--s3)}.c4{background:var(--s4)}.c5{background:var(--s5)}.c6{background:var(--s6)}.c7{background:var(--s7)}.c8{background:var(--s8)}
+</style></head><body><main>
+<h1>Shrunk architecture</h1><p class="meta">huggingface-mic/DeepSeek-V3.2-mic@2156444e97b761624604df50de0ac677d402f289 ← deepseek-ai/DeepSeek-V3.2@a7e62ac04ecb2c0a54d736dc46601c5606cf10a6 · DeepseekV32ForCausalLM · random weights · built before the shrink rule (no recorded recipe)</p>
+<div class="stats"><div class="stat"><span class="k">Original</span><span class="v">671.9B</span><span class="k">parameters</span></div><div class="stat"><span class="k">Fixture</span><span class="v">16.9M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Layer stacks</span><span class="v">1</span><span class="k">compared</span></div><div class="stat"><span class="k">Findings</span><span class="v">0</span><span class="k">to review</span></div></div><section class="findings ok"><h2>✓ Same architecture, smaller</h2><p>Every module class, layer kind and checked property of the original is in the fixture.</p></section><h2>Layer stacks</h2><p class="note">One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.</p><div class="stack"><h3>model.layers</h3><div class="row"><span class="lab">Original · 61</span><div class="cells"><i class="c0" title="layer 0: no Experts"></i><i class="c0" title="layer 1: no Experts"></i><i class="c0" title="layer 2: no Experts"></i><i class="c1" title="layer 3: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 4: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 5: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 6: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 7: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 8: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 9: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 10: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 11: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 12: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 13: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 14: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 15: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 16: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 17: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 18: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 19: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 20: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 21: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 22: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 23: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 24: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 25: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 26: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 27: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 28: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 29: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 30: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 31: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 32: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 33: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 34: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 35: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 36: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 37: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 38: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 39: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 40: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 41: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 42: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 43: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 44: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 45: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 46: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 47: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 48: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 49: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 50: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 51: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 52: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 53: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 54: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 55: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 56: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 57: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 58: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 59: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 60: Experts, MoE, TopkRouter"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: no Experts"></i><i class="c1" title="layer 1: Experts, MoE, TopkRouter"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>no Experts</td><td>3</td><td>1</td><td>✓ kept</td></tr><tr><td><i class="c1"></i>Experts, MoE, TopkRouter</td><td>58</td><td>1</td><td>✓ kept</td></tr></table></div><h2>Where the parameters are</h2><p class="note">Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.</p><div class="row"><span class="lab">Original</span><div class="bar"><i class="c0" style="flex:0.997242" title="model.layers: 670.0B (99.7%)"></i><i class="c1" style="flex:0.001379" title="model.embed_tokens: 926.7M (0.1%)"></i><i class="c2" style="flex:0.001379" title="lm_head: 926.7M (0.1%)"></i><i class="c3" style="flex:0.000000" title="model.norm: 7.2K (0.0%)"></i></div></div><div class="row"><span class="lab">Fixture</span><div class="bar"><i class="c0" style="flex:0.021499" title="model.layers: 363.6K (2.1%)"></i><i class="c1" style="flex:0.489248" title="model.embed_tokens: 8.3M (48.9%)"></i><i class="c2" style="flex:0.489248" title="lm_head: 8.3M (48.9%)"></i><i class="c3" style="flex:0.000004" title="model.norm: 64 (0.0%)"></i></div></div><table class="legend"><tr><th>Component</th><th>Original</th><th>Share</th><th>Fixture</th><th>Share</th></tr><tr><td><i class="c0"></i>model.layers</td><td>670.0B</td><td>99.7%</td><td>363.6K</td><td>2.1%</td></tr><tr><td><i class="c1"></i>model.embed_tokens</td><td>926.7M</td><td>0.1%</td><td>8.3M</td><td>48.9%</td></tr><tr><td><i class="c2"></i>lm_head</td><td>926.7M</td><td>0.1%</td><td>8.3M</td><td>48.9%</td></tr><tr><td><i class="c3"></i>model.norm</td><td>7.2K</td><td>0.0%</td><td>64</td><td>0.0%</td></tr></table><h2>Preserved properties</h2><table><tr><th>Config</th><th>Property</th><th>Original</th><th>Fixture</th><th></th></tr><tr class=""><td>model</td><td>attention style</td><td>MHA (128 heads, 128 KV)</td><td>MHA (4 heads, 4 KV)</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>MoE routing</td><td>top-8 of 256 experts</td><td>top-2 of 4 experts</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>RoPE type</td><td>yarn</td><td>yarn</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>model_type</td><td>deepseek_v32</td><td>deepseek_v32</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>vocab_size</td><td>129280</td><td>129280</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>tie_word_embeddings</td><td>False</td><td>False</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>hidden_act</td><td>silu</td><td>silu</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>max_position_embeddings</td><td>163840</td><td>163840</td><td>✓ kept</td></tr></table><h2>Module classes</h2><p class="note">Instances of each class in the built model (prefix DeepseekV32 omitted).</p><table><tr><th>Class</th><th>Original</th><th>Fixture</th></tr><tr class=""><td>Linear</td><td>672</td><td>23</td></tr><tr class=""><td>RMSNorm</td><td>245</td><td>9</td></tr><tr class=""><td>SiLUActivation</td><td>119</td><td>3</td></tr><tr class=""><td>Attention</td><td>61</td><td>2</td></tr><tr class=""><td>DecoderLayer</td><td>61</td><td>2</td></tr><tr class=""><td>Indexer</td><td>61</td><td>2</td></tr><tr class=""><td>MLP</td><td>61</td><td>2</td></tr><tr class=""><td>LayerNorm</td><td>61</td><td>2</td></tr><tr class=""><td>Experts</td><td>58</td><td>1</td></tr><tr class=""><td>MoE</td><td>58</td><td>1</td></tr><tr class=""><td>TopkRouter</td><td>58</td><td>1</td></tr><tr class=""><td>ForCausalLM</td><td>1</td><td>1</td></tr><tr class=""><td>Model</td><td>1</td><td>1</td></tr><tr class=""><td>RotaryEmbedding</td><td>1</td><td>1</td></tr><tr class=""><td>Embedding</td><td>1</td><td>1</td></tr><tr class=""><td>ModuleList</td><td>1</td><td>1</td></tr></table><h2>Config changes</h2><p class="note">20 fields changed, 36 unchanged.</p><table><tr><th>Field</th><th>Original</th><th>Fixture</th></tr><tr><td>first_k_dense_replace</td><td>3</td><td>1</td></tr><tr><td>head_dim</td><td>64</td><td>8</td></tr><tr><td>hidden_size</td><td>7168</td><td>64</td></tr><tr><td>intermediate_size</td><td>18432</td><td>128</td></tr><tr><td>kv_lora_rank</td><td>512</td><td>16</td></tr><tr><td>layer_types</td><td>61 entries: 61× indexed_attention</td><td>[&quot;indexed_attention&quot;, &quot;indexed_attention&quot;]</td></tr><tr><td>mlp_layer_types</td><td>61 entries: 58× sparse, 3× dense</td><td>[&quot;dense&quot;, &quot;sparse&quot;]</td></tr><tr><td>moe_intermediate_size</td><td>2048</td><td>32</td></tr><tr><td>n_group</td><td>8</td><td>1</td></tr><tr><td>n_routed_experts</td><td>256</td><td>4</td></tr><tr><td>num_attention_heads</td><td>128</td><td>4</td></tr><tr><td>num_experts_per_tok</td><td>8</td><td>2</td></tr><tr><td>num_hidden_layers</td><td>61</td><td>2</td></tr><tr><td>num_key_value_heads</td><td>128</td><td>4</td></tr><tr><td>q_lora_rank</td><td>1536</td><td>16</td></tr><tr><td>qk_nope_head_dim</td><td>128</td><td>16</td></tr><tr><td>qk_rope_head_dim</td><td>64</td><td>8</td></tr><tr><td>quantization_config</td><td>{&quot;activation_scheme&quot;: &quot;dynamic&quot;, &quot;fmt&quot;: &quot;e4m3&quot;, &quot;quant_method&quot;: &quot;fp8&quot;, &quot;scale…</td><td>—</td></tr><tr><td>topk_group</td><td>4</td><td>1</td></tr><tr><td>v_head_dim</td><td>128</td><td>16</td></tr></table>
+</main></body></html>

shrink/DeepSeek-V4-Flash-0731-mic.html

@@ -0,0 +1,34 @@
+<!doctype html>
+<html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
+<title>Shrunk architecture</title>
+<style>
+:root{--bg:#fcfcfb;--fg:#0b0b0b;--muted:#52514e;--line:#e4e3df;--ok:#1d7a43;--warn:#b35c00;--s0:#2a78d6;--s1:#eb6834;--s2:#1baf7a;--s3:#eda100;--s4:#e87ba4;--s5:#008300;--s6:#6250d6;--s7:#e34948;--s8:#8a8984;}
+@media (prefers-color-scheme:dark){:root:not([data-theme="light"]){--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}}
+:root[data-theme="dark"]{--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}
+body{margin:0;background:var(--bg);color:var(--fg);font:15px/1.5 system-ui,sans-serif}
+main{max-width:1040px;margin:0 auto;padding:24px 16px 64px}
+h1{font-size:24px;margin:0 0 4px}h2{font-size:18px;margin:36px 0 8px}h3{font-size:14px;margin:0 0 8px;font-family:ui-monospace,monospace}
+.meta,.note{color:var(--muted);margin:0 0 12px;overflow-wrap:anywhere}
+.stats{display:grid;grid-template-columns:repeat(auto-fit,minmax(150px,1fr));gap:12px;margin:20px 0}
+.stat{border:1px solid var(--line);border-radius:8px;padding:12px;display:flex;flex-direction:column}
+.stat .v{font-size:26px;font-weight:600}.stat .k{color:var(--muted);font-size:13px}
+.findings{border-radius:8px;padding:4px 16px;border:1px solid var(--line)}
+.findings.ok{border-left:4px solid var(--ok)}.findings.warn{border-left:4px solid var(--warn)}
+.findings h2{margin:12px 0 4px;font-size:16px}
+.stack{border:1px solid var(--line);border-radius:8px;padding:12px;margin:12px 0}
+.row{display:flex;align-items:center;gap:12px;margin:6px 0}
+.lab{flex:0 0 110px;color:var(--muted);font-size:13px}
+.cells{display:flex;flex-wrap:wrap;gap:2px}
+.cells i{width:12px;height:12px;border-radius:2px;display:block}
+.bar{flex:1;display:flex;gap:2px;height:16px}.bar i{display:block;border-radius:2px;min-width:2px}
+table{border-collapse:collapse;width:100%;font-size:13px;margin:8px 0;display:block;overflow-x:auto}
+th,td{text-align:left;padding:4px 10px 4px 0;border-bottom:1px solid var(--line);vertical-align:top}
+th{color:var(--muted);font-weight:500}
+td:first-child{font-family:ui-monospace,monospace;overflow-wrap:anywhere}
+td i{display:inline-block;width:10px;height:10px;border-radius:2px;margin-right:6px}
+tr.bad td{color:var(--warn)}
+.c0{background:var(--s0)}.c1{background:var(--s1)}.c2{background:var(--s2)}.c3{background:var(--s3)}.c4{background:var(--s4)}.c5{background:var(--s5)}.c6{background:var(--s6)}.c7{background:var(--s7)}.c8{background:var(--s8)}
+</style></head><body><main>
+<h1>Shrunk architecture</h1><p class="meta">huggingface-mic/DeepSeek-V4-Flash-0731-mic@7c36fb38ddaf52e554d784d9e6a2a7ed435b8311 ← deepseek-ai/DeepSeek-V4-Flash-0731@7872f01b1d1fe23eabc4c98b48bffcef5a386062 · DeepseekV4ForCausalLM · random weights</p>
+<div class="stats"><div class="stat"><span class="k">Original</span><span class="v">284.3B</span><span class="k">parameters</span></div><div class="stat"><span class="k">Fixture</span><span class="v">20.1M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Layer stacks</span><span class="v">1</span><span class="k">compared</span></div><div class="stat"><span class="k">Findings</span><span class="v">0</span><span class="k">to review</span></div></div><section class="findings ok"><h2>✓ Same architecture, smaller</h2><p>Every module class, layer kind and checked property of the original is in the fixture.</p></section><h2>Layer stacks</h2><p class="note">One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.</p><div class="stack"><h3>model.layers</h3><div class="row"><span class="lab">Original · 43</span><div class="cells"><i class="c0" title="layer 0: sliding_attention, HashRouter"></i><i class="c0" title="layer 1: sliding_attention, HashRouter"></i><i class="c1" title="layer 2: compressed_sparse_attention, CSACompressor, HashRouter …"></i><i class="c2" title="layer 3: heavily_compressed_attention, HCACompressor, RotaryEmbedding …"></i><i class="c3" title="layer 4: compressed_sparse_attention, CSACompressor, Indexer …"></i><i class="c2" title="layer 5: heavily_compressed_attention, HCACompressor, RotaryEmbedding …"></i><i class="c3" title="layer 6: compressed_sparse_attention, CSACompressor, Indexer …"></i><i class="c2" title="layer 7: heavily_compressed_attention, HCACompressor, RotaryEmbedding …"></i><i class="c3" title="layer 8: compressed_sparse_attention, CSACompressor, Indexer …"></i><i class="c2" title="layer 9: heavily_compressed_attention, HCACompressor, RotaryEmbedding …"></i><i class="c3" title="layer 10: compressed_sparse_attention, CSACompressor, Indexer …"></i><i class="c2" title="layer 11: heavily_compressed_attention, HCACompressor, RotaryEmbedding …"></i><i class="c3" title="layer 12: compressed_sparse_attention, CSACompressor, Indexer …"></i><i class="c2" title="layer 13: heavily_compressed_attention, HCACompressor, RotaryEmbedding …"></i><i class="c3" title="layer 14: compressed_sparse_attention, CSACompressor, Indexer …"></i><i class="c2" title="layer 15: heavily_compressed_attention, HCACompressor, RotaryEmbedding …"></i><i class="c3" title="layer 16: compressed_sparse_attention, CSACompressor, Indexer …"></i><i class="c2" title="layer 17: heavily_compressed_attention, HCACompressor, RotaryEmbedding …"></i><i class="c3" title="layer 18: compressed_sparse_attention, CSACompressor, Indexer …"></i><i class="c2" title="layer 19: heavily_compressed_attention, HCACompressor, RotaryEmbedding …"></i><i class="c3" title="layer 20: compressed_sparse_attention, CSACompressor, Indexer …"></i><i class="c2" title="layer 21: heavily_compressed_attention, HCACompressor, RotaryEmbedding …"></i><i class="c3" title="layer 22: compressed_sparse_attention, CSACompressor, Indexer …"></i><i class="c2" title="layer 23: heavily_compressed_attention, HCACompressor, RotaryEmbedding …"></i><i class="c3" title="layer 24: compressed_sparse_attention, CSACompressor, Indexer …"></i><i class="c2" title="layer 25: heavily_compressed_attention, HCACompressor, RotaryEmbedding …"></i><i class="c3" title="layer 26: compressed_sparse_attention, CSACompressor, Indexer …"></i><i class="c2" title="layer 27: heavily_compressed_attention, HCACompressor, RotaryEmbedding …"></i><i class="c3" title="layer 28: compressed_sparse_attention, CSACompressor, Indexer …"></i><i class="c2" title="layer 29: heavily_compressed_attention, HCACompressor, RotaryEmbedding …"></i><i class="c3" title="layer 30: compressed_sparse_attention, CSACompressor, Indexer …"></i><i class="c2" title="layer 31: heavily_compressed_attention, HCACompressor, RotaryEmbedding …"></i><i class="c3" title="layer 32: compressed_sparse_attention, CSACompressor, Indexer …"></i><i class="c2" title="layer 33: heavily_compressed_attention, HCACompressor, RotaryEmbedding …"></i><i class="c3" title="layer 34: compressed_sparse_attention, CSACompressor, Indexer …"></i><i class="c2" title="layer 35: heavily_compressed_attention, HCACompressor, RotaryEmbedding …"></i><i class="c3" title="layer 36: compressed_sparse_attention, CSACompressor, Indexer …"></i><i class="c2" title="layer 37: heavily_compressed_attention, HCACompressor, RotaryEmbedding …"></i><i class="c3" title="layer 38: compressed_sparse_attention, CSACompressor, Indexer …"></i><i class="c2" title="layer 39: heavily_compressed_attention, HCACompressor, RotaryEmbedding …"></i><i class="c3" title="layer 40: compressed_sparse_attention, CSACompressor, Indexer …"></i><i class="c2" title="layer 41: heavily_compressed_attention, HCACompressor, RotaryEmbedding …"></i><i class="c3" title="layer 42: compressed_sparse_attention, CSACompressor, Indexer …"></i></div></div><div class="row"><span class="lab">Fixture · 5</span><div class="cells"><i class="c0" title="layer 0: sliding_attention, HashRouter"></i><i class="c0" title="layer 1: sliding_attention, HashRouter"></i><i class="c1" title="layer 2: compressed_sparse_attention, CSACompressor, HashRouter …"></i><i class="c2" title="layer 3: heavily_compressed_attention, HCACompressor, RotaryEmbedding …"></i><i class="c3" title="layer 4: compressed_sparse_attention, CSACompressor, Indexer …"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>sliding_attention, HashRouter</td><td>2</td><td>2</td><td>✓ kept</td></tr><tr><td><i class="c1"></i>compressed_sparse_attention, CSACompressor, HashRouter …</td><td>1</td><td>1</td><td>✓ kept</td></tr><tr><td><i class="c2"></i>heavily_compressed_attention, HCACompressor, RotaryEmbedding …</td><td>20</td><td>1</td><td>✓ kept</td></tr><tr><td><i class="c3"></i>compressed_sparse_attention, CSACompressor, Indexer …</td><td>20</td><td>1</td><td>✓ kept</td></tr></table></div><h2>Where the parameters are</h2><p class="note">Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.</p><div class="row"><span class="lab">Original</span><div class="bar"><i class="c0" style="flex:0.996275" title="model.layers: 283.3B (99.6%)"></i><i class="c1" style="flex:0.001862" title="lm_head: 529.5M (0.2%)"></i><i class="c2" style="flex:0.001862" title="model.embed_tokens: 529.5M (0.2%)"></i><i class="c3" style="flex:0.000000" title="model.hc_head: 65.5K (0.0%)"></i><i class="c4" style="flex:0.000000" title="model.norm: 4.1K (0.0%)"></i></div></div><div class="row"><span class="lab">Fixture</span><div class="bar"><i class="c0" style="flex:0.175915" title="model.layers: 3.5M (17.6%)"></i><i class="c1" style="flex:0.412015" title="lm_head: 8.3M (41.2%)"></i><i class="c2" style="flex:0.412015" title="model.embed_tokens: 8.3M (41.2%)"></i><i class="c3" style="flex:0.000051" title="model.hc_head: 1.0K (0.0%)"></i><i class="c4" style="flex:0.000003" title="model.norm: 64 (0.0%)"></i></div></div><table class="legend"><tr><th>Component</th><th>Original</th><th>Share</th><th>Fixture</th><th>Share</th></tr><tr><td><i class="c0"></i>model.layers</td><td>283.3B</td><td>99.6%</td><td>3.5M</td><td>17.6%</td></tr><tr><td><i class="c1"></i>lm_head</td><td>529.5M</td><td>0.2%</td><td>8.3M</td><td>41.2%</td></tr><tr><td><i class="c2"></i>model.embed_tokens</td><td>529.5M</td><td>0.2%</td><td>8.3M</td><td>41.2%</td></tr><tr><td><i class="c3"></i>model.hc_head</td><td>65.5K</td><td>0.0%</td><td>1.0K</td><td>0.0%</td></tr><tr><td><i class="c4"></i>model.norm</td><td>4.1K</td><td>0.0%</td><td>64</td><td>0.0%</td></tr></table><h2>Preserved properties</h2><table><tr><th>Config</th><th>Property</th><th>Original</th><th>Fixture</th><th></th></tr><tr class=""><td>model</td><td>attention style</td><td>MQA (64 heads, 1 KV)</td><td>MQA (4 heads, 1 KV)</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>MoE routing</td><td>top-6 of 256 experts</td><td>top-2 of 4 experts</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>model_type</td><td>deepseek_v4</td><td>deepseek_v4</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>vocab_size</td><td>129280</td><td>129280</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>tie_word_embeddings</td><td>False</td><td>False</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>hidden_act</td><td>silu</td><td>silu</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>max_position_embeddings</td><td>1048576</td><td>1048576</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>rope_theta</td><td>10000</td><td>10000</td><td>✓ kept</td></tr></table><h2>Module classes</h2><p class="note">Instances of each class in the built model (prefix DeepseekV4 omitted).</p><table><tr><th>Class</th><th>Original</th><th>Fixture</th></tr><tr class=""><td>Linear</td><td>468</td><td>50</td></tr><tr class=""><td>RMSNorm</td><td>235</td><td>26</td></tr><tr class=""><td>UnweightedRMSNorm</td><td>130</td><td>16</td></tr><tr class=""><td>HyperConnection</td><td>86</td><td>10</td></tr><tr class=""><td>SiLUActivation</td><td>86</td><td>10</td></tr><tr class=""><td>RotaryEmbedding</td><td>63</td><td>6</td></tr><tr class=""><td>Attention</td><td>43</td><td>5</td></tr><tr class=""><td>DecoderLayer</td><td>43</td><td>5</td></tr><tr class=""><td>Experts</td><td>43</td><td>5</td></tr><tr class=""><td>GroupedLinear</td><td>43</td><td>5</td></tr><tr class=""><td>MLP</td><td>43</td><td>5</td></tr><tr class=""><td>SparseMoeBlock</td><td>43</td><td>5</td></tr><tr class=""><td>SqrtSoftplusActivation</td><td>43</td><td>5</td></tr><tr class=""><td>TopKRouter</td><td>40</td><td>2</td></tr><tr class=""><td>CSACompressor</td><td>21</td><td>2</td></tr><tr class=""><td>Indexer</td><td>21</td><td>2</td></tr><tr class=""><td>IndexerScorer</td><td>21</td><td>2</td></tr><tr class=""><td>HCACompressor</td><td>20</td><td>1</td></tr><tr class=""><td>HashRouter</td><td>3</td><td>3</td></tr><tr class=""><td>ForCausalLM</td><td>1</td><td>1</td></tr><tr class=""><td>HyperHead</td><td>1</td><td>1</td></tr><tr class=""><td>Model</td><td>1</td><td>1</td></tr><tr class=""><td>Embedding</td><td>1</td><td>1</td></tr><tr class=""><td>ModuleList</td><td>1</td><td>1</td></tr></table><h2>Config changes</h2><p class="note">12 fields changed, 52 unchanged.</p><table><tr><th>Field</th><th>Original</th><th>Fixture</th></tr><tr><td>head_dim</td><td>512</td><td>16</td></tr><tr><td>hidden_size</td><td>4096</td><td>64</td></tr><tr><td>layer_types</td><td>43 entries: 21× compressed_sparse_attention, 20× heavily_compressed_attention, 2× sliding_attention</td><td>[&quot;sliding_attention&quot;, &quot;sliding_attention&quot;, &quot;compressed_sparse_attention&quot;, &quot;he…</td></tr><tr><td>mlp_layer_types</td><td>43 entries: 40× moe, 3× hash_moe</td><td>[&quot;hash_moe&quot;, &quot;hash_moe&quot;, &quot;hash_moe&quot;, &quot;moe&quot;, &quot;moe&quot;]</td></tr><tr><td>moe_intermediate_size</td><td>2048</td><td>32</td></tr><tr><td>n_routed_experts</td><td>256</td><td>4</td></tr><tr><td>num_attention_heads</td><td>64</td><td>4</td></tr><tr><td>num_experts_per_tok</td><td>6</td><td>2</td></tr><tr><td>num_hidden_layers</td><td>43</td><td>5</td></tr><tr><td>q_lora_rank</td><td>1024</td><td>16</td></tr><tr><td>qk_rope_head_dim</td><td>64</td><td>2</td></tr><tr><td>quantization_config</td><td>{&quot;activation_scheme&quot;: &quot;dynamic&quot;, &quot;fmt&quot;: &quot;e4m3&quot;, &quot;quant_method&quot;: &quot;fp8&quot;, &quot;scale…</td><td>—</td></tr></table>
+</main></body></html>

shrink/ESMC-300M-hf-mic.html

@@ -0,0 +1,34 @@
+<!doctype html>
+<html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
+<title>Shrunk architecture</title>
+<style>
+:root{--bg:#fcfcfb;--fg:#0b0b0b;--muted:#52514e;--line:#e4e3df;--ok:#1d7a43;--warn:#b35c00;--s0:#2a78d6;--s1:#eb6834;--s2:#1baf7a;--s3:#eda100;--s4:#e87ba4;--s5:#008300;--s6:#6250d6;--s7:#e34948;--s8:#8a8984;}
+@media (prefers-color-scheme:dark){:root:not([data-theme="light"]){--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}}
+:root[data-theme="dark"]{--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}
+body{margin:0;background:var(--bg);color:var(--fg);font:15px/1.5 system-ui,sans-serif}
+main{max-width:1040px;margin:0 auto;padding:24px 16px 64px}
+h1{font-size:24px;margin:0 0 4px}h2{font-size:18px;margin:36px 0 8px}h3{font-size:14px;margin:0 0 8px;font-family:ui-monospace,monospace}
+.meta,.note{color:var(--muted);margin:0 0 12px;overflow-wrap:anywhere}
+.stats{display:grid;grid-template-columns:repeat(auto-fit,minmax(150px,1fr));gap:12px;margin:20px 0}
+.stat{border:1px solid var(--line);border-radius:8px;padding:12px;display:flex;flex-direction:column}
+.stat .v{font-size:26px;font-weight:600}.stat .k{color:var(--muted);font-size:13px}
+.findings{border-radius:8px;padding:4px 16px;border:1px solid var(--line)}
+.findings.ok{border-left:4px solid var(--ok)}.findings.warn{border-left:4px solid var(--warn)}
+.findings h2{margin:12px 0 4px;font-size:16px}
+.stack{border:1px solid var(--line);border-radius:8px;padding:12px;margin:12px 0}
+.row{display:flex;align-items:center;gap:12px;margin:6px 0}
+.lab{flex:0 0 110px;color:var(--muted);font-size:13px}
+.cells{display:flex;flex-wrap:wrap;gap:2px}
+.cells i{width:12px;height:12px;border-radius:2px;display:block}
+.bar{flex:1;display:flex;gap:2px;height:16px}.bar i{display:block;border-radius:2px;min-width:2px}
+table{border-collapse:collapse;width:100%;font-size:13px;margin:8px 0;display:block;overflow-x:auto}
+th,td{text-align:left;padding:4px 10px 4px 0;border-bottom:1px solid var(--line);vertical-align:top}
+th{color:var(--muted);font-weight:500}
+td:first-child{font-family:ui-monospace,monospace;overflow-wrap:anywhere}
+td i{display:inline-block;width:10px;height:10px;border-radius:2px;margin-right:6px}
+tr.bad td{color:var(--warn)}
+.c0{background:var(--s0)}.c1{background:var(--s1)}.c2{background:var(--s2)}.c3{background:var(--s3)}.c4{background:var(--s4)}.c5{background:var(--s5)}.c6{background:var(--s6)}.c7{background:var(--s7)}.c8{background:var(--s8)}
+</style></head><body><main>
+<h1>Shrunk architecture</h1><p class="meta">huggingface-mic/ESMC-300M-hf-mic@7c11c22ceffcbddb7fb1bb3216255d34ff38a43c ← biohub/ESMC-300M-hf@5554e4b3ab19eb66f04d7ff8e8c2d36c69b52fa0 · EsmcForMaskedLM · random weights</p>
+<div class="stats"><div class="stat"><span class="k">Original</span><span class="v">333.0M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Fixture</span><span class="v">95.3K</span><span class="k">parameters</span></div><div class="stat"><span class="k">Layer stacks</span><span class="v">1</span><span class="k">compared</span></div><div class="stat"><span class="k">Findings</span><span class="v">0</span><span class="k">to review</span></div></div><section class="findings ok"><h2>✓ Same architecture, smaller</h2><p>Every module class, layer kind and checked property of the original is in the fixture.</p></section><h2>Layer stacks</h2><p class="note">One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.</p><div class="stack"><h3>esmc.layers</h3><div class="row"><span class="lab">Original · 30</span><div class="cells"><i class="c0" title="layer 0: Layer"></i><i class="c0" title="layer 1: Layer"></i><i class="c0" title="layer 2: Layer"></i><i class="c0" title="layer 3: Layer"></i><i class="c0" title="layer 4: Layer"></i><i class="c0" title="layer 5: Layer"></i><i class="c0" title="layer 6: Layer"></i><i class="c0" title="layer 7: Layer"></i><i class="c0" title="layer 8: Layer"></i><i class="c0" title="layer 9: Layer"></i><i class="c0" title="layer 10: Layer"></i><i class="c0" title="layer 11: Layer"></i><i class="c0" title="layer 12: Layer"></i><i class="c0" title="layer 13: Layer"></i><i class="c0" title="layer 14: Layer"></i><i class="c0" title="layer 15: Layer"></i><i class="c0" title="layer 16: Layer"></i><i class="c0" title="layer 17: Layer"></i><i class="c0" title="layer 18: Layer"></i><i class="c0" title="layer 19: Layer"></i><i class="c0" title="layer 20: Layer"></i><i class="c0" title="layer 21: Layer"></i><i class="c0" title="layer 22: Layer"></i><i class="c0" title="layer 23: Layer"></i><i class="c0" title="layer 24: Layer"></i><i class="c0" title="layer 25: Layer"></i><i class="c0" title="layer 26: Layer"></i><i class="c0" title="layer 27: Layer"></i><i class="c0" title="layer 28: Layer"></i><i class="c0" title="layer 29: Layer"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: Layer"></i><i class="c0" title="layer 1: Layer"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>Layer</td><td>30</td><td>2</td><td>✓ kept</td></tr></table></div><h2>Where the parameters are</h2><p class="note">Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.</p><div class="row"><span class="lab">Original</span><div class="bar"><i class="c0" style="flex:0.996852" title="esmc.layers: 331.9M (99.7%)"></i><i class="c1" style="flex:0.002770" title="lm_head.dense: 922.6K (0.3%)"></i><i class="c2" style="flex:0.000185" title="lm_head.decoder: 61.5K (0.0%)"></i><i class="c3" style="flex:0.000185" title="esmc.embed_tokens: 61.4K (0.0%)"></i><i class="c4" style="flex:0.000006" title="lm_head.layer_norm: 1.9K (0.0%)"></i><i class="c5" style="flex:0.000003" title="esmc.norm: 960 (0.0%)"></i></div></div><div class="row"><span class="lab">Fixture</span><div class="bar"><i class="c0" style="flex:0.867696" title="esmc.layers: 82.7K (86.8%)"></i><i class="c1" style="flex:0.043653" title="lm_head.dense: 4.2K (4.4%)"></i><i class="c2" style="flex:0.043653" title="lm_head.decoder: 4.2K (4.4%)"></i><i class="c3" style="flex:0.042982" title="esmc.embed_tokens: 4.1K (4.3%)"></i><i class="c4" style="flex:0.001343" title="lm_head.layer_norm: 128 (0.1%)"></i><i class="c5" style="flex:0.000672" title="esmc.norm: 64 (0.1%)"></i></div></div><table class="legend"><tr><th>Component</th><th>Original</th><th>Share</th><th>Fixture</th><th>Share</th></tr><tr><td><i class="c0"></i>esmc.layers</td><td>331.9M</td><td>99.7%</td><td>82.7K</td><td>86.8%</td></tr><tr><td><i class="c1"></i>lm_head.dense</td><td>922.6K</td><td>0.3%</td><td>4.2K</td><td>4.4%</td></tr><tr><td><i class="c2"></i>lm_head.decoder</td><td>61.5K</td><td>0.0%</td><td>4.2K</td><td>4.4%</td></tr><tr><td><i class="c3"></i>esmc.embed_tokens</td><td>61.4K</td><td>0.0%</td><td>4.1K</td><td>4.3%</td></tr><tr><td><i class="c4"></i>lm_head.layer_norm</td><td>1.9K</td><td>0.0%</td><td>128</td><td>0.1%</td></tr><tr><td><i class="c5"></i>esmc.norm</td><td>960</td><td>0.0%</td><td>64</td><td>0.1%</td></tr></table><h2>Preserved properties</h2><table><tr><th>Config</th><th>Property</th><th>Original</th><th>Fixture</th><th></th></tr><tr class=""><td>model</td><td>attention style</td><td>MHA (15 heads, 15 KV)</td><td>MHA (4 heads, 4 KV)</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>RoPE type</td><td>default</td><td>default</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>model_type</td><td>esmc</td><td>esmc</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>vocab_size</td><td>64</td><td>64</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>tie_word_embeddings</td><td>False</td><td>False</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>hidden_act</td><td>silu</td><td>silu</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>max_position_embeddings</td><td>2048</td><td>2048</td><td>✓ kept</td></tr></table><h2>Module classes</h2><p class="note">Instances of each class in the built model (prefix Esmc omitted).</p><table><tr><th>Class</th><th>Original</th><th>Fixture</th></tr><tr class=""><td>Linear</td><td>212</td><td>16</td></tr><tr class=""><td>LayerNorm</td><td>121</td><td>9</td></tr><tr class=""><td>Attention</td><td>30</td><td>2</td></tr><tr class=""><td>Layer</td><td>30</td><td>2</td></tr><tr class=""><td>MLP</td><td>30</td><td>2</td></tr><tr class=""><td>SiLUActivation</td><td>30</td><td>2</td></tr><tr class=""><td>Embedding</td><td>1</td><td>1</td></tr><tr class=""><td>ForMaskedLM</td><td>1</td><td>1</td></tr><tr class=""><td>MaskedLMHead</td><td>1</td><td>1</td></tr><tr class=""><td>Model</td><td>1</td><td>1</td></tr><tr class=""><td>RotaryEmbedding</td><td>1</td><td>1</td></tr><tr class=""><td>LayerNorm</td><td>1</td><td>1</td></tr><tr class=""><td>ModuleList</td><td>1</td><td>1</td></tr></table><h2>Config changes</h2><p class="note">6 fields changed, 23 unchanged.</p><table><tr><th>Field</th><th>Original</th><th>Fixture</th></tr><tr><td>head_dim</td><td>64</td><td>16</td></tr><tr><td>hidden_size</td><td>960</td><td>64</td></tr><tr><td>intermediate_size</td><td>2560</td><td>128</td></tr><tr><td>num_attention_heads</td><td>15</td><td>4</td></tr><tr><td>num_hidden_layers</td><td>30</td><td>2</td></tr><tr><td>num_key_value_heads</td><td>15</td><td>4</td></tr></table>
+</main></body></html>

shrink/EuroBERT-210m-mic.html

@@ -0,0 +1,34 @@
+<!doctype html>
+<html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
+<title>Shrunk architecture</title>
+<style>
+:root{--bg:#fcfcfb;--fg:#0b0b0b;--muted:#52514e;--line:#e4e3df;--ok:#1d7a43;--warn:#b35c00;--s0:#2a78d6;--s1:#eb6834;--s2:#1baf7a;--s3:#eda100;--s4:#e87ba4;--s5:#008300;--s6:#6250d6;--s7:#e34948;--s8:#8a8984;}
+@media (prefers-color-scheme:dark){:root:not([data-theme="light"]){--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}}
+:root[data-theme="dark"]{--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}
+body{margin:0;background:var(--bg);color:var(--fg);font:15px/1.5 system-ui,sans-serif}
+main{max-width:1040px;margin:0 auto;padding:24px 16px 64px}
+h1{font-size:24px;margin:0 0 4px}h2{font-size:18px;margin:36px 0 8px}h3{font-size:14px;margin:0 0 8px;font-family:ui-monospace,monospace}
+.meta,.note{color:var(--muted);margin:0 0 12px;overflow-wrap:anywhere}
+.stats{display:grid;grid-template-columns:repeat(auto-fit,minmax(150px,1fr));gap:12px;margin:20px 0}
+.stat{border:1px solid var(--line);border-radius:8px;padding:12px;display:flex;flex-direction:column}
+.stat .v{font-size:26px;font-weight:600}.stat .k{color:var(--muted);font-size:13px}
+.findings{border-radius:8px;padding:4px 16px;border:1px solid var(--line)}
+.findings.ok{border-left:4px solid var(--ok)}.findings.warn{border-left:4px solid var(--warn)}
+.findings h2{margin:12px 0 4px;font-size:16px}
+.stack{border:1px solid var(--line);border-radius:8px;padding:12px;margin:12px 0}
+.row{display:flex;align-items:center;gap:12px;margin:6px 0}
+.lab{flex:0 0 110px;color:var(--muted);font-size:13px}
+.cells{display:flex;flex-wrap:wrap;gap:2px}
+.cells i{width:12px;height:12px;border-radius:2px;display:block}
+.bar{flex:1;display:flex;gap:2px;height:16px}.bar i{display:block;border-radius:2px;min-width:2px}
+table{border-collapse:collapse;width:100%;font-size:13px;margin:8px 0;display:block;overflow-x:auto}
+th,td{text-align:left;padding:4px 10px 4px 0;border-bottom:1px solid var(--line);vertical-align:top}
+th{color:var(--muted);font-weight:500}
+td:first-child{font-family:ui-monospace,monospace;overflow-wrap:anywhere}
+td i{display:inline-block;width:10px;height:10px;border-radius:2px;margin-right:6px}
+tr.bad td{color:var(--warn)}
+.c0{background:var(--s0)}.c1{background:var(--s1)}.c2{background:var(--s2)}.c3{background:var(--s3)}.c4{background:var(--s4)}.c5{background:var(--s5)}.c6{background:var(--s6)}.c7{background:var(--s7)}.c8{background:var(--s8)}
+</style></head><body><main>
+<h1>Shrunk architecture</h1><p class="meta">huggingface-mic/EuroBERT-210m-mic@514cea1b679f118f051124ef88272fb809f25f16 ← EuroBERT/EuroBERT-210m@39b51e15dd1f1a06f58b5cbf6a8a188cec60bd0e · EuroBertForMaskedLM · random weights</p>
+<div class="stats"><div class="stat"><span class="k">Original</span><span class="v">310.3M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Fixture</span><span class="v">16.5M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Layer stacks</span><span class="v">1</span><span class="k">compared</span></div><div class="stat"><span class="k">Findings</span><span class="v">0</span><span class="k">to review</span></div></div><section class="findings ok"><h2>✓ Same architecture, smaller</h2><p>Every module class, layer kind and checked property of the original is in the fixture.</p></section><h2>Layer stacks</h2><p class="note">One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.</p><div class="stack"><h3>model.layers</h3><div class="row"><span class="lab">Original · 12</span><div class="cells"><i class="c0" title="layer 0: DecoderLayer"></i><i class="c0" title="layer 1: DecoderLayer"></i><i class="c0" title="layer 2: DecoderLayer"></i><i class="c0" title="layer 3: DecoderLayer"></i><i class="c0" title="layer 4: DecoderLayer"></i><i class="c0" title="layer 5: DecoderLayer"></i><i class="c0" title="layer 6: DecoderLayer"></i><i class="c0" title="layer 7: DecoderLayer"></i><i class="c0" title="layer 8: DecoderLayer"></i><i class="c0" title="layer 9: DecoderLayer"></i><i class="c0" title="layer 10: DecoderLayer"></i><i class="c0" title="layer 11: DecoderLayer"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: DecoderLayer"></i><i class="c0" title="layer 1: DecoderLayer"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>DecoderLayer</td><td>12</td><td>2</td><td>✓ kept</td></tr></table></div><h2>Where the parameters are</h2><p class="note">Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.</p><div class="row"><span class="lab">Original</span><div class="bar"><i class="c0" style="flex:0.365056" title="model.layers: 113.3M (36.5%)"></i><i class="c1" style="flex:0.317471" title="model.embed_tokens: 98.5M (31.7%)"></i><i class="c2" style="flex:0.317471" title="lm_head: 98.5M (31.7%)"></i><i class="c3" style="flex:0.000002" title="model.norm: 768 (0.0%)"></i></div></div><div class="row"><span class="lab">Fixture</span><div class="bar"><i class="c0" style="flex:0.004981" title="model.layers: 82.2K (0.5%)"></i><i class="c1" style="flex:0.497508" title="model.embed_tokens: 8.2M (49.8%)"></i><i class="c2" style="flex:0.497508" title="lm_head: 8.2M (49.8%)"></i><i class="c3" style="flex:0.000004" title="model.norm: 64 (0.0%)"></i></div></div><table class="legend"><tr><th>Component</th><th>Original</th><th>Share</th><th>Fixture</th><th>Share</th></tr><tr><td><i class="c0"></i>model.layers</td><td>113.3M</td><td>36.5%</td><td>82.2K</td><td>0.5%</td></tr><tr><td><i class="c1"></i>model.embed_tokens</td><td>98.5M</td><td>31.7%</td><td>8.2M</td><td>49.8%</td></tr><tr><td><i class="c2"></i>lm_head</td><td>98.5M</td><td>31.7%</td><td>8.2M</td><td>49.8%</td></tr><tr><td><i class="c3"></i>model.norm</td><td>768</td><td>0.0%</td><td>64</td><td>0.0%</td></tr></table><h2>Preserved properties</h2><table><tr><th>Config</th><th>Property</th><th>Original</th><th>Fixture</th><th></th></tr><tr class=""><td>model</td><td>attention style</td><td>MHA (12 heads, 12 KV)</td><td>MHA (4 heads, 4 KV)</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>RoPE type</td><td>default</td><td>default</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>model_type</td><td>eurobert</td><td>eurobert</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>vocab_size</td><td>128256</td><td>128256</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>tie_word_embeddings</td><td>False</td><td>False</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>hidden_act</td><td>silu</td><td>silu</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>max_position_embeddings</td><td>8192</td><td>8192</td><td>✓ kept</td></tr></table><h2>Module classes</h2><p class="note">Instances of each class in the built model (prefix EuroBert omitted).</p><table><tr><th>Class</th><th>Original</th><th>Fixture</th></tr><tr class=""><td>Linear</td><td>85</td><td>15</td></tr><tr class=""><td>RMSNorm</td><td>25</td><td>5</td></tr><tr class=""><td>Attention</td><td>12</td><td>2</td></tr><tr class=""><td>DecoderLayer</td><td>12</td><td>2</td></tr><tr class=""><td>MLP</td><td>12</td><td>2</td></tr><tr class=""><td>SiLUActivation</td><td>12</td><td>2</td></tr><tr class=""><td>Embedding</td><td>1</td><td>1</td></tr><tr class=""><td>ForMaskedLM</td><td>1</td><td>1</td></tr><tr class=""><td>Model</td><td>1</td><td>1</td></tr><tr class=""><td>RotaryEmbedding</td><td>1</td><td>1</td></tr><tr class=""><td>ModuleList</td><td>1</td><td>1</td></tr></table><h2>Config changes</h2><p class="note">6 fields changed, 33 unchanged.</p><table><tr><th>Field</th><th>Original</th><th>Fixture</th></tr><tr><td>head_dim</td><td>64</td><td>16</td></tr><tr><td>hidden_size</td><td>768</td><td>64</td></tr><tr><td>intermediate_size</td><td>3072</td><td>128</td></tr><tr><td>num_attention_heads</td><td>12</td><td>4</td></tr><tr><td>num_hidden_layers</td><td>12</td><td>2</td></tr><tr><td>num_key_value_heads</td><td>12</td><td>4</td></tr></table>
+</main></body></html>

shrink/Falcon-H1-Tiny-90M-Instruct-mic.html

@@ -0,0 +1,34 @@
+<!doctype html>
+<html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
+<title>Shrunk architecture</title>
+<style>
+:root{--bg:#fcfcfb;--fg:#0b0b0b;--muted:#52514e;--line:#e4e3df;--ok:#1d7a43;--warn:#b35c00;--s0:#2a78d6;--s1:#eb6834;--s2:#1baf7a;--s3:#eda100;--s4:#e87ba4;--s5:#008300;--s6:#6250d6;--s7:#e34948;--s8:#8a8984;}
+@media (prefers-color-scheme:dark){:root:not([data-theme="light"]){--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}}
+:root[data-theme="dark"]{--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}
+body{margin:0;background:var(--bg);color:var(--fg);font:15px/1.5 system-ui,sans-serif}
+main{max-width:1040px;margin:0 auto;padding:24px 16px 64px}
+h1{font-size:24px;margin:0 0 4px}h2{font-size:18px;margin:36px 0 8px}h3{font-size:14px;margin:0 0 8px;font-family:ui-monospace,monospace}
+.meta,.note{color:var(--muted);margin:0 0 12px;overflow-wrap:anywhere}
+.stats{display:grid;grid-template-columns:repeat(auto-fit,minmax(150px,1fr));gap:12px;margin:20px 0}
+.stat{border:1px solid var(--line);border-radius:8px;padding:12px;display:flex;flex-direction:column}
+.stat .v{font-size:26px;font-weight:600}.stat .k{color:var(--muted);font-size:13px}
+.findings{border-radius:8px;padding:4px 16px;border:1px solid var(--line)}
+.findings.ok{border-left:4px solid var(--ok)}.findings.warn{border-left:4px solid var(--warn)}
+.findings h2{margin:12px 0 4px;font-size:16px}
+.stack{border:1px solid var(--line);border-radius:8px;padding:12px;margin:12px 0}
+.row{display:flex;align-items:center;gap:12px;margin:6px 0}
+.lab{flex:0 0 110px;color:var(--muted);font-size:13px}
+.cells{display:flex;flex-wrap:wrap;gap:2px}
+.cells i{width:12px;height:12px;border-radius:2px;display:block}
+.bar{flex:1;display:flex;gap:2px;height:16px}.bar i{display:block;border-radius:2px;min-width:2px}
+table{border-collapse:collapse;width:100%;font-size:13px;margin:8px 0;display:block;overflow-x:auto}
+th,td{text-align:left;padding:4px 10px 4px 0;border-bottom:1px solid var(--line);vertical-align:top}
+th{color:var(--muted);font-weight:500}
+td:first-child{font-family:ui-monospace,monospace;overflow-wrap:anywhere}
+td i{display:inline-block;width:10px;height:10px;border-radius:2px;margin-right:6px}
+tr.bad td{color:var(--warn)}
+.c0{background:var(--s0)}.c1{background:var(--s1)}.c2{background:var(--s2)}.c3{background:var(--s3)}.c4{background:var(--s4)}.c5{background:var(--s5)}.c6{background:var(--s6)}.c7{background:var(--s7)}.c8{background:var(--s8)}
+</style></head><body><main>
+<h1>Shrunk architecture</h1><p class="meta">huggingface-mic/Falcon-H1-Tiny-90M-Instruct-mic@0db75b996c3e47fea9cc1a9c213983988b24495a ← tiiuae/Falcon-H1-Tiny-90M-Instruct@e6389502a0b12cd8da894b395ba5bf7436873b16 · FalconH1ForCausalLM · random weights</p>
+<div class="stats"><div class="stat"><span class="k">Original</span><span class="v">91.1M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Fixture</span><span class="v">2.5M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Layer stacks</span><span class="v">1</span><span class="k">compared</span></div><div class="stat"><span class="k">Findings</span><span class="v">0</span><span class="k">to review</span></div></div><section class="findings ok"><h2>✓ Same architecture, smaller</h2><p>Every module class, layer kind and checked property of the original is in the fixture.</p></section><h2>Layer stacks</h2><p class="note">One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.</p><div class="stack"><h3>model.layers</h3><div class="row"><span class="lab">Original · 24</span><div class="cells"><i class="c0" title="layer 0: DecoderLayer"></i><i class="c0" title="layer 1: DecoderLayer"></i><i class="c0" title="layer 2: DecoderLayer"></i><i class="c0" title="layer 3: DecoderLayer"></i><i class="c0" title="layer 4: DecoderLayer"></i><i class="c0" title="layer 5: DecoderLayer"></i><i class="c0" title="layer 6: DecoderLayer"></i><i class="c0" title="layer 7: DecoderLayer"></i><i class="c0" title="layer 8: DecoderLayer"></i><i class="c0" title="layer 9: DecoderLayer"></i><i class="c0" title="layer 10: DecoderLayer"></i><i class="c0" title="layer 11: DecoderLayer"></i><i class="c0" title="layer 12: DecoderLayer"></i><i class="c0" title="layer 13: DecoderLayer"></i><i class="c0" title="layer 14: DecoderLayer"></i><i class="c0" title="layer 15: DecoderLayer"></i><i class="c0" title="layer 16: DecoderLayer"></i><i class="c0" title="layer 17: DecoderLayer"></i><i class="c0" title="layer 18: DecoderLayer"></i><i class="c0" title="layer 19: DecoderLayer"></i><i class="c0" title="layer 20: DecoderLayer"></i><i class="c0" title="layer 21: DecoderLayer"></i><i class="c0" title="layer 22: DecoderLayer"></i><i class="c0" title="layer 23: DecoderLayer"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: DecoderLayer"></i><i class="c0" title="layer 1: DecoderLayer"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>DecoderLayer</td><td>24</td><td>2</td><td>✓ kept</td></tr></table></div><h2>Where the parameters are</h2><p class="note">Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.</p><div class="row"><span class="lab">Original</span><div class="bar"><i class="c0" style="flex:0.815895" title="model.layers: 74.4M (81.6%)"></i><i class="c1" style="flex:0.184100" title="model.embed_tokens: 16.8M (18.4%)"></i><i class="c2" style="flex:0.000006" title="model.final_layernorm: 512 (0.0%)"></i></div></div><div class="row"><span class="lab">Fixture</span><div class="bar"><i class="c0" style="flex:0.154103" title="model.layers: 382.1K (15.4%)"></i><i class="c1" style="flex:0.845871" title="model.embed_tokens: 2.1M (84.6%)"></i><i class="c2" style="flex:0.000026" title="model.final_layernorm: 64 (0.0%)"></i></div></div><table class="legend"><tr><th>Component</th><th>Original</th><th>Share</th><th>Fixture</th><th>Share</th></tr><tr><td><i class="c0"></i>model.layers</td><td>74.4M</td><td>81.6%</td><td>382.1K</td><td>15.4%</td></tr><tr><td><i class="c1"></i>model.embed_tokens</td><td>16.8M</td><td>18.4%</td><td>2.1M</td><td>84.6%</td></tr><tr><td><i class="c2"></i>model.final_layernorm</td><td>512</td><td>0.0%</td><td>64</td><td>0.0%</td></tr></table><h2>Preserved properties</h2><table><tr><th>Config</th><th>Property</th><th>Original</th><th>Fixture</th><th></th></tr><tr class=""><td>model</td><td>attention style</td><td>GQA (8 heads, 2 KV)</td><td>GQA (4 heads, 2 KV)</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>RoPE type</td><td>default</td><td>default</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>model_type</td><td>falcon_h1</td><td>falcon_h1</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>vocab_size</td><td>32768</td><td>32768</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>tie_word_embeddings</td><td>True</td><td>True</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>hidden_act</td><td>silu</td><td>silu</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>max_position_embeddings</td><td>262144</td><td>262144</td><td>✓ kept</td></tr></table><h2>Module classes</h2><p class="note">Instances of each class in the built model (prefix FalconH1 omitted).</p><table><tr><th>Class</th><th>Original</th><th>Fixture</th></tr><tr class=""><td>Linear</td><td>217</td><td>19</td></tr><tr class=""><td>RMSNorm</td><td>49</td><td>5</td></tr><tr class=""><td>SiLUActivation</td><td>48</td><td>4</td></tr><tr class=""><td>Conv1d</td><td>24</td><td>2</td></tr><tr class=""><td>Attention</td><td>24</td><td>2</td></tr><tr class=""><td>DecoderLayer</td><td>24</td><td>2</td></tr><tr class=""><td>MLP</td><td>24</td><td>2</td></tr><tr class=""><td>Mixer</td><td>24</td><td>2</td></tr><tr class=""><td>Identity</td><td>24</td><td>2</td></tr><tr class=""><td>Embedding</td><td>1</td><td>1</td></tr><tr class=""><td>ForCausalLM</td><td>1</td><td>1</td></tr><tr class=""><td>Model</td><td>1</td><td>1</td></tr><tr class=""><td>RotaryEmbedding</td><td>1</td><td>1</td></tr><tr class=""><td>ModuleList</td><td>1</td><td>1</td></tr></table><h2>Config changes</h2><p class="note">9 fields changed, 53 unchanged.</p><table><tr><th>Field</th><th>Original</th><th>Fixture</th></tr><tr><td>head_dim</td><td>64</td><td>16</td></tr><tr><td>hidden_size</td><td>512</td><td>64</td></tr><tr><td>intermediate_size</td><td>768</td><td>128</td></tr><tr><td>mamba_chunk_size</td><td>128</td><td>8</td></tr><tr><td>mamba_d_head</td><td>32</td><td>96</td></tr><tr><td>mamba_d_state</td><td>64</td><td>16</td></tr><tr><td>mamba_n_heads</td><td>24</td><td>8</td></tr><tr><td>num_attention_heads</td><td>8</td><td>4</td></tr><tr><td>num_hidden_layers</td><td>24</td><td>2</td></tr></table>
+</main></body></html>

shrink/Florence-2-base-mic.html

@@ -0,0 +1,34 @@
+<!doctype html>
+<html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
+<title>Shrunk architecture</title>
+<style>
+:root{--bg:#fcfcfb;--fg:#0b0b0b;--muted:#52514e;--line:#e4e3df;--ok:#1d7a43;--warn:#b35c00;--s0:#2a78d6;--s1:#eb6834;--s2:#1baf7a;--s3:#eda100;--s4:#e87ba4;--s5:#008300;--s6:#6250d6;--s7:#e34948;--s8:#8a8984;}
+@media (prefers-color-scheme:dark){:root:not([data-theme="light"]){--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}}
+:root[data-theme="dark"]{--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}
+body{margin:0;background:var(--bg);color:var(--fg);font:15px/1.5 system-ui,sans-serif}
+main{max-width:1040px;margin:0 auto;padding:24px 16px 64px}
+h1{font-size:24px;margin:0 0 4px}h2{font-size:18px;margin:36px 0 8px}h3{font-size:14px;margin:0 0 8px;font-family:ui-monospace,monospace}
+.meta,.note{color:var(--muted);margin:0 0 12px;overflow-wrap:anywhere}
+.stats{display:grid;grid-template-columns:repeat(auto-fit,minmax(150px,1fr));gap:12px;margin:20px 0}
+.stat{border:1px solid var(--line);border-radius:8px;padding:12px;display:flex;flex-direction:column}
+.stat .v{font-size:26px;font-weight:600}.stat .k{color:var(--muted);font-size:13px}
+.findings{border-radius:8px;padding:4px 16px;border:1px solid var(--line)}
+.findings.ok{border-left:4px solid var(--ok)}.findings.warn{border-left:4px solid var(--warn)}
+.findings h2{margin:12px 0 4px;font-size:16px}
+.stack{border:1px solid var(--line);border-radius:8px;padding:12px;margin:12px 0}
+.row{display:flex;align-items:center;gap:12px;margin:6px 0}
+.lab{flex:0 0 110px;color:var(--muted);font-size:13px}
+.cells{display:flex;flex-wrap:wrap;gap:2px}
+.cells i{width:12px;height:12px;border-radius:2px;display:block}
+.bar{flex:1;display:flex;gap:2px;height:16px}.bar i{display:block;border-radius:2px;min-width:2px}
+table{border-collapse:collapse;width:100%;font-size:13px;margin:8px 0;display:block;overflow-x:auto}
+th,td{text-align:left;padding:4px 10px 4px 0;border-bottom:1px solid var(--line);vertical-align:top}
+th{color:var(--muted);font-weight:500}
+td:first-child{font-family:ui-monospace,monospace;overflow-wrap:anywhere}
+td i{display:inline-block;width:10px;height:10px;border-radius:2px;margin-right:6px}
+tr.bad td{color:var(--warn)}
+.c0{background:var(--s0)}.c1{background:var(--s1)}.c2{background:var(--s2)}.c3{background:var(--s3)}.c4{background:var(--s4)}.c5{background:var(--s5)}.c6{background:var(--s6)}.c7{background:var(--s7)}.c8{background:var(--s8)}
+</style></head><body><main>
+<h1>Shrunk architecture</h1><p class="meta">huggingface-mic/Florence-2-base-mic@85c1b12e1d9565763f6f2b9466669c58bc7aac33 ← florence-community/Florence-2-base@00921df66db728a9ceb750f5eca43e5c203a2051 · Florence2ForConditionalGeneration · overfit weights · built before the shrink rule (no recorded recipe)</p>
+<div class="stats"><div class="stat"><span class="k">Original</span><span class="v">231.4M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Fixture</span><span class="v">6.4M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Layer stacks</span><span class="v">4</span><span class="k">compared</span></div><div class="stat"><span class="k">Findings</span><span class="v">0</span><span class="k">to review</span></div></div><section class="findings ok"><h2>✓ Same architecture, smaller</h2><p>Every module class, layer kind and checked property of the original is in the fixture.</p></section><h2>Layer stacks</h2><p class="note">One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.</p><div class="stack"><h3>model.vision_tower.convs</h3><div class="row"><span class="lab">Original · 4</span><div class="cells"><i class="c0" title="layer 0: Conv2d"></i><i class="c0" title="layer 1: Conv2d"></i><i class="c0" title="layer 2: Conv2d"></i><i class="c0" title="layer 3: Conv2d"></i></div></div><div class="row"><span class="lab">Fixture · 4</span><div class="cells"><i class="c0" title="layer 0: Conv2d"></i><i class="c0" title="layer 1: Conv2d"></i><i class="c0" title="layer 2: Conv2d"></i><i class="c0" title="layer 3: Conv2d"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>Conv2d</td><td>4</td><td>4</td><td>✓ kept</td></tr></table></div><div class="stack"><h3>model.vision_tower.blocks</h3><div class="row"><span class="lab">Original · 4</span><div class="cells"><i class="c0" title="layer 0: VisionBlock"></i><i class="c0" title="layer 1: VisionBlock"></i><i class="c0" title="layer 2: VisionBlock"></i><i class="c0" title="layer 3: VisionBlock"></i></div></div><div class="row"><span class="lab">Fixture · 4</span><div class="cells"><i class="c0" title="layer 0: VisionBlock"></i><i class="c0" title="layer 1: VisionBlock"></i><i class="c0" title="layer 2: VisionBlock"></i><i class="c0" title="layer 3: VisionBlock"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>VisionBlock</td><td>4</td><td>4</td><td>✓ kept</td></tr></table></div><div class="stack"><h3>model.language_model.encoder.layers</h3><div class="row"><span class="lab">Original · 6</span><div class="cells"><i class="c0" title="layer 0: BartEncoderLayer"></i><i class="c0" title="layer 1: BartEncoderLayer"></i><i class="c0" title="layer 2: BartEncoderLayer"></i><i class="c0" title="layer 3: BartEncoderLayer"></i><i class="c0" title="layer 4: BartEncoderLayer"></i><i class="c0" title="layer 5: BartEncoderLayer"></i></div></div><div class="row"><span class="lab">Fixture · 1</span><div class="cells"><i class="c0" title="layer 0: BartEncoderLayer"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>BartEncoderLayer</td><td>6</td><td>1</td><td>✓ kept</td></tr></table></div><div class="stack"><h3>model.language_model.decoder.layers</h3><div class="row"><span class="lab">Original · 6</span><div class="cells"><i class="c0" title="layer 0: BartDecoderLayer"></i><i class="c0" title="layer 1: BartDecoderLayer"></i><i class="c0" title="layer 2: BartDecoderLayer"></i><i class="c0" title="layer 3: BartDecoderLayer"></i><i class="c0" title="layer 4: BartDecoderLayer"></i><i class="c0" title="layer 5: BartDecoderLayer"></i></div></div><div class="row"><span class="lab">Fixture · 1</span><div class="cells"><i class="c0" title="layer 0: BartDecoderLayer"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>BartDecoderLayer</td><td>6</td><td>1</td><td>✓ kept</td></tr></table></div><h2>Where the parameters are</h2><p class="note">Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.</p><div class="row"><span class="lab">Original</span><div class="bar"><i class="c0" style="flex:0.363596" title="model.vision_tower.blocks: 84.2M (36.4%)"></i><i class="c1" style="flex:0.245030" title="model.language_model.decoder.layers: 56.7M (24.5%)"></i><i class="c2" style="flex:0.183747" title="model.language_model.encoder.layers: 42.5M (18.4%)"></i><i class="c3" style="flex:0.177144" title="model.language_model: 41.0M (17.7%)"></i><i class="c4" style="flex:0.026857" title="model.vision_tower.convs: 6.2M (2.7%)"></i><i class="c5" style="flex:0.003626" title="model.multi_modal_projector: 839.2K (0.4%)"></i></div></div><div class="row"><span class="lab">Fixture</span><div class="bar"><i class="c0" style="flex:0.391378" title="model.vision_tower.blocks: 2.5M (39.1%)"></i><i class="c1" style="flex:0.007796" title="model.language_model.decoder.layers: 50.2K (0.8%)"></i><i class="c2" style="flex:0.005194" title="model.language_model.encoder.layers: 33.5K (0.5%)"></i><i class="c3" style="flex:0.530140" title="model.language_model: 3.4M (53.0%)"></i><i class="c4" style="flex:0.060945" title="model.vision_tower.convs: 392.8K (6.1%)"></i><i class="c5" style="flex:0.004548" title="model.multi_modal_projector: 29.3K (0.5%)"></i></div></div><table class="legend"><tr><th>Component</th><th>Original</th><th>Share</th><th>Fixture</th><th>Share</th></tr><tr><td><i class="c0"></i>model.vision_tower.blocks</td><td>84.2M</td><td>36.4%</td><td>2.5M</td><td>39.1%</td></tr><tr><td><i class="c1"></i>model.language_model.decoder.layers</td><td>56.7M</td><td>24.5%</td><td>50.2K</td><td>0.8%</td></tr><tr><td><i class="c2"></i>model.language_model.encoder.layers</td><td>42.5M</td><td>18.4%</td><td>33.5K</td><td>0.5%</td></tr><tr><td><i class="c3"></i>model.language_model</td><td>41.0M</td><td>17.7%</td><td>3.4M</td><td>53.0%</td></tr><tr><td><i class="c4"></i>model.vision_tower.convs</td><td>6.2M</td><td>2.7%</td><td>392.8K</td><td>6.1%</td></tr><tr><td><i class="c5"></i>model.multi_modal_projector</td><td>839.2K</td><td>0.4%</td><td>29.3K</td><td>0.5%</td></tr></table><h2>Preserved properties</h2><table><tr><th>Config</th><th>Property</th><th>Original</th><th>Fixture</th><th></th></tr><tr class=""><td>model</td><td>model_type</td><td>florence2</td><td>florence2</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>tie_word_embeddings</td><td>True</td><td>True</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>model_type</td><td>bart</td><td>bart</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>vocab_size</td><td>51328</td><td>51328</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>tie_word_embeddings</td><td>True</td><td>True</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>max_position_embeddings</td><td>1024</td><td>1024</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>model_type</td><td>florence_vision</td><td>florence_vision</td><td>�� kept</td></tr><tr class=""><td>vision_config</td><td>max_position_embeddings</td><td>50</td><td>50</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>patch_size</td><td>[7, 3, 3, 3]</td><td>[7, 3, 3, 3]</td><td>✓ kept</td></tr></table><h2>Module classes</h2><p class="note">Instances of each class in the built model (prefix Florence2 omitted).</p><table><tr><th>Class</th><th>Original</th><th>Fixture</th></tr><tr class=""><td>Linear</td><td>194</td><td>58</td></tr><tr class=""><td>LayerNorm</td><td>85</td><td>32</td></tr><tr class=""><td>Conv2d</td><td>52</td><td>24</td></tr><tr class=""><td>DropPath</td><td>46</td><td>18</td></tr><tr class=""><td>GELUActivation</td><td>36</td><td>12</td></tr><tr class=""><td>VisionMLP</td><td>24</td><td>10</td></tr><tr class=""><td>BartAttention</td><td>18</td><td>3</td></tr><tr class=""><td>VisionBlock</td><td>12</td><td>5</td></tr><tr class=""><td>VisionChannelAttention</td><td>12</td><td>5</td></tr><tr class=""><td>VisionChannelBlock</td><td>12</td><td>5</td></tr><tr class=""><td>VisionSpatialBlock</td><td>12</td><td>5</td></tr><tr class=""><td>VisionWindowAttention</td><td>12</td><td>5</td></tr><tr class=""><td>ModuleList</td><td>8</td><td>8</td></tr><tr class=""><td>BartDecoderLayer</td><td>6</td><td>1</td></tr><tr class=""><td>BartEncoderLayer</td><td>6</td><td>1</td></tr><tr class=""><td>VisionConvEmbed</td><td>4</td><td>4</td></tr><tr class=""><td>BartScaledWordEmbedding</td><td>3</td><td>3</td></tr><tr class=""><td>BartLearnedPositionalEmbedding</td><td>2</td><td>2</td></tr><tr class=""><td>Embedding</td><td>2</td><td>2</td></tr><tr class=""><td>Identity</td><td>2</td><td>2</td></tr><tr class=""><td>BartDecoder</td><td>1</td><td>1</td></tr><tr class=""><td>BartEncoder</td><td>1</td><td>1</td></tr><tr class=""><td>BartModel</td><td>1</td><td>1</td></tr><tr class=""><td>ForConditionalGeneration</td><td>1</td><td>1</td></tr><tr class=""><td>Model</td><td>1</td><td>1</td></tr><tr class=""><td>MultiModalProjector</td><td>1</td><td>1</td></tr><tr class=""><td>VisionBackbone</td><td>1</td><td>1</td></tr><tr class=""><td>VisionLearnedAbsolutePositionEmbedding2D</td><td>1</td><td>1</td></tr><tr class=""><td>VisionPositionalEmbeddingCosine1D</td><td>1</td><td>1</td></tr></table><h2>Config changes</h2><p class="note">12 fields changed, 72 unchanged.</p><table><tr><th>Field</th><th>Original</th><th>Fixture</th></tr><tr><td>text_config.d_model</td><td>768</td><td>64</td></tr><tr><td>text_config.decoder_attention_heads</td><td>12</td><td>2</td></tr><tr><td>text_config.decoder_ffn_dim</td><td>3072</td><td>128</td></tr><tr><td>text_config.decoder_layers</td><td>6</td><td>1</td></tr><tr><td>text_config.encoder_attention_heads</td><td>12</td><td>2</td></tr><tr><td>text_config.encoder_ffn_dim</td><td>3072</td><td>128</td></tr><tr><td>text_config.encoder_layers</td><td>6</td><td>1</td></tr><tr><td>vision_config.depths</td><td>[1, 1, 9, 1]</td><td>[1, 1, 2, 1]</td></tr><tr><td>vision_config.embed_dim</td><td>[128, 256, 512, 1024]</td><td>[32, 64, 128, 256]</td></tr><tr><td>vision_config.num_groups</td><td>[4, 8, 16, 32]</td><td>[2, 4, 8, 16]</td></tr><tr><td>vision_config.num_heads</td><td>[4, 8, 16, 32]</td><td>[2, 4, 8, 16]</td></tr><tr><td>vision_config.projection_dim</td><td>768</td><td>64</td></tr></table>
+</main></body></html>

shrink/GLM-4-9B-0414-mic.html

@@ -0,0 +1,34 @@
+<!doctype html>
+<html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
+<title>Shrunk architecture</title>
+<style>
+:root{--bg:#fcfcfb;--fg:#0b0b0b;--muted:#52514e;--line:#e4e3df;--ok:#1d7a43;--warn:#b35c00;--s0:#2a78d6;--s1:#eb6834;--s2:#1baf7a;--s3:#eda100;--s4:#e87ba4;--s5:#008300;--s6:#6250d6;--s7:#e34948;--s8:#8a8984;}
+@media (prefers-color-scheme:dark){:root:not([data-theme="light"]){--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}}
+:root[data-theme="dark"]{--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}
+body{margin:0;background:var(--bg);color:var(--fg);font:15px/1.5 system-ui,sans-serif}
+main{max-width:1040px;margin:0 auto;padding:24px 16px 64px}
+h1{font-size:24px;margin:0 0 4px}h2{font-size:18px;margin:36px 0 8px}h3{font-size:14px;margin:0 0 8px;font-family:ui-monospace,monospace}
+.meta,.note{color:var(--muted);margin:0 0 12px;overflow-wrap:anywhere}
+.stats{display:grid;grid-template-columns:repeat(auto-fit,minmax(150px,1fr));gap:12px;margin:20px 0}
+.stat{border:1px solid var(--line);border-radius:8px;padding:12px;display:flex;flex-direction:column}
+.stat .v{font-size:26px;font-weight:600}.stat .k{color:var(--muted);font-size:13px}
+.findings{border-radius:8px;padding:4px 16px;border:1px solid var(--line)}
+.findings.ok{border-left:4px solid var(--ok)}.findings.warn{border-left:4px solid var(--warn)}
+.findings h2{margin:12px 0 4px;font-size:16px}
+.stack{border:1px solid var(--line);border-radius:8px;padding:12px;margin:12px 0}
+.row{display:flex;align-items:center;gap:12px;margin:6px 0}
+.lab{flex:0 0 110px;color:var(--muted);font-size:13px}
+.cells{display:flex;flex-wrap:wrap;gap:2px}
+.cells i{width:12px;height:12px;border-radius:2px;display:block}
+.bar{flex:1;display:flex;gap:2px;height:16px}.bar i{display:block;border-radius:2px;min-width:2px}
+table{border-collapse:collapse;width:100%;font-size:13px;margin:8px 0;display:block;overflow-x:auto}
+th,td{text-align:left;padding:4px 10px 4px 0;border-bottom:1px solid var(--line);vertical-align:top}
+th{color:var(--muted);font-weight:500}
+td:first-child{font-family:ui-monospace,monospace;overflow-wrap:anywhere}
+td i{display:inline-block;width:10px;height:10px;border-radius:2px;margin-right:6px}
+tr.bad td{color:var(--warn)}
+.c0{background:var(--s0)}.c1{background:var(--s1)}.c2{background:var(--s2)}.c3{background:var(--s3)}.c4{background:var(--s4)}.c5{background:var(--s5)}.c6{background:var(--s6)}.c7{background:var(--s7)}.c8{background:var(--s8)}
+</style></head><body><main>
+<h1>Shrunk architecture</h1><p class="meta">huggingface-mic/GLM-4-9B-0414-mic@140a76e1cd7001bc3d0aed3b715a2a902214ece6 ← zai-org/GLM-4-9B-0414@645b8482494e31b6b752272bf7f7f273ef0f3caf · Glm4ForCausalLM · random weights</p>
+<div class="stats"><div class="stat"><span class="k">Original</span><span class="v">9.4B</span><span class="k">parameters</span></div><div class="stat"><span class="k">Fixture</span><span class="v">19.5M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Layer stacks</span><span class="v">1</span><span class="k">compared</span></div><div class="stat"><span class="k">Findings</span><span class="v">0</span><span class="k">to review</span></div></div><section class="findings ok"><h2>✓ Same architecture, smaller</h2><p>Every module class, layer kind and checked property of the original is in the fixture.</p></section><h2>Layer stacks</h2><p class="note">One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.</p><div class="stack"><h3>model.layers</h3><div class="row"><span class="lab">Original · 40</span><div class="cells"><i class="c0" title="layer 0: DecoderLayer"></i><i class="c0" title="layer 1: DecoderLayer"></i><i class="c0" title="layer 2: DecoderLayer"></i><i class="c0" title="layer 3: DecoderLayer"></i><i class="c0" title="layer 4: DecoderLayer"></i><i class="c0" title="layer 5: DecoderLayer"></i><i class="c0" title="layer 6: DecoderLayer"></i><i class="c0" title="layer 7: DecoderLayer"></i><i class="c0" title="layer 8: DecoderLayer"></i><i class="c0" title="layer 9: DecoderLayer"></i><i class="c0" title="layer 10: DecoderLayer"></i><i class="c0" title="layer 11: DecoderLayer"></i><i class="c0" title="layer 12: DecoderLayer"></i><i class="c0" title="layer 13: DecoderLayer"></i><i class="c0" title="layer 14: DecoderLayer"></i><i class="c0" title="layer 15: DecoderLayer"></i><i class="c0" title="layer 16: DecoderLayer"></i><i class="c0" title="layer 17: DecoderLayer"></i><i class="c0" title="layer 18: DecoderLayer"></i><i class="c0" title="layer 19: DecoderLayer"></i><i class="c0" title="layer 20: DecoderLayer"></i><i class="c0" title="layer 21: DecoderLayer"></i><i class="c0" title="layer 22: DecoderLayer"></i><i class="c0" title="layer 23: DecoderLayer"></i><i class="c0" title="layer 24: DecoderLayer"></i><i class="c0" title="layer 25: DecoderLayer"></i><i class="c0" title="layer 26: DecoderLayer"></i><i class="c0" title="layer 27: DecoderLayer"></i><i class="c0" title="layer 28: DecoderLayer"></i><i class="c0" title="layer 29: DecoderLayer"></i><i class="c0" title="layer 30: DecoderLayer"></i><i class="c0" title="layer 31: DecoderLayer"></i><i class="c0" title="layer 32: DecoderLayer"></i><i class="c0" title="layer 33: DecoderLayer"></i><i class="c0" title="layer 34: DecoderLayer"></i><i class="c0" title="layer 35: DecoderLayer"></i><i class="c0" title="layer 36: DecoderLayer"></i><i class="c0" title="layer 37: DecoderLayer"></i><i class="c0" title="layer 38: DecoderLayer"></i><i class="c0" title="layer 39: DecoderLayer"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: DecoderLayer"></i><i class="c0" title="layer 1: DecoderLayer"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>DecoderLayer</td><td>40</td><td>2</td><td>✓ kept</td></tr></table></div><h2>Where the parameters are</h2><p class="note">Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.</p><div class="row"><span class="lab">Original</span><div class="bar"><i class="c0" style="flex:0.867928" title="model.layers: 8.2B (86.8%)"></i><i class="c1" style="flex:0.066036" title="model.embed_tokens: 620.8M (6.6%)"></i><i class="c2" style="flex:0.066036" title="lm_head: 620.8M (6.6%)"></i><i class="c3" style="flex:0.000000" title="model.norm: 4.1K (0.0%)"></i></div></div><div class="row"><span class="lab">Fixture</span><div class="bar"><i class="c0" style="flex:0.003826" title="model.layers: 74.5K (0.4%)"></i><i class="c1" style="flex:0.498086" title="model.embed_tokens: 9.7M (49.8%)"></i><i class="c2" style="flex:0.498086" title="lm_head: 9.7M (49.8%)"></i><i class="c3" style="flex:0.000003" title="model.norm: 64 (0.0%)"></i></div></div><table class="legend"><tr><th>Component</th><th>Original</th><th>Share</th><th>Fixture</th><th>Share</th></tr><tr><td><i class="c0"></i>model.layers</td><td>8.2B</td><td>86.8%</td><td>74.5K</td><td>0.4%</td></tr><tr><td><i class="c1"></i>model.embed_tokens</td><td>620.8M</td><td>6.6%</td><td>9.7M</td><td>49.8%</td></tr><tr><td><i class="c2"></i>lm_head</td><td>620.8M</td><td>6.6%</td><td>9.7M</td><td>49.8%</td></tr><tr><td><i class="c3"></i>model.norm</td><td>4.1K</td><td>0.0%</td><td>64</td><td>0.0%</td></tr></table><h2>Preserved properties</h2><table><tr><th>Config</th><th>Property</th><th>Original</th><th>Fixture</th><th></th></tr><tr class=""><td>model</td><td>attention style</td><td>GQA (32 heads, 2 KV)</td><td>GQA (4 heads, 2 KV)</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>RoPE type</td><td>default</td><td>default</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>model_type</td><td>glm4</td><td>glm4</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>vocab_size</td><td>151552</td><td>151552</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>tie_word_embeddings</td><td>False</td><td>False</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>hidden_act</td><td>silu</td><td>silu</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>max_position_embeddings</td><td>32768</td><td>32768</td><td>✓ kept</td></tr></table><h2>Module classes</h2><p class="note">Instances of each class in the built model (prefix Glm4 omitted).</p><table><tr><th>Class</th><th>Original</th><th>Fixture</th></tr><tr class=""><td>Linear</td><td>241</td><td>13</td></tr><tr class=""><td>RMSNorm</td><td>161</td><td>9</td></tr><tr class=""><td>Attention</td><td>40</td><td>2</td></tr><tr class=""><td>DecoderLayer</td><td>40</td><td>2</td></tr><tr class=""><td>MLP</td><td>40</td><td>2</td></tr><tr class=""><td>SiLUActivation</td><td>40</td><td>2</td></tr><tr class=""><td>Embedding</td><td>1</td><td>1</td></tr><tr class=""><td>ForCausalLM</td><td>1</td><td>1</td></tr><tr class=""><td>Model</td><td>1</td><td>1</td></tr><tr class=""><td>RotaryEmbedding</td><td>1</td><td>1</td></tr><tr class=""><td>ModuleList</td><td>1</td><td>1</td></tr></table><h2>Config changes</h2><p class="note">5 fields changed, 23 unchanged.</p><table><tr><th>Field</th><th>Original</th><th>Fixture</th></tr><tr><td>head_dim</td><td>128</td><td>16</td></tr><tr><td>hidden_size</td><td>4096</td><td>64</td></tr><tr><td>intermediate_size</td><td>13696</td><td>128</td></tr><tr><td>num_attention_heads</td><td>32</td><td>4</td></tr><tr><td>num_hidden_layers</td><td>40</td><td>2</td></tr></table>
+</main></body></html>

shrink/GLM-4.5-Air-mic.html

@@ -0,0 +1,34 @@
+<!doctype html>
+<html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
+<title>Shrunk architecture</title>
+<style>
+:root{--bg:#fcfcfb;--fg:#0b0b0b;--muted:#52514e;--line:#e4e3df;--ok:#1d7a43;--warn:#b35c00;--s0:#2a78d6;--s1:#eb6834;--s2:#1baf7a;--s3:#eda100;--s4:#e87ba4;--s5:#008300;--s6:#6250d6;--s7:#e34948;--s8:#8a8984;}
+@media (prefers-color-scheme:dark){:root:not([data-theme="light"]){--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}}
+:root[data-theme="dark"]{--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}
+body{margin:0;background:var(--bg);color:var(--fg);font:15px/1.5 system-ui,sans-serif}
+main{max-width:1040px;margin:0 auto;padding:24px 16px 64px}
+h1{font-size:24px;margin:0 0 4px}h2{font-size:18px;margin:36px 0 8px}h3{font-size:14px;margin:0 0 8px;font-family:ui-monospace,monospace}
+.meta,.note{color:var(--muted);margin:0 0 12px;overflow-wrap:anywhere}
+.stats{display:grid;grid-template-columns:repeat(auto-fit,minmax(150px,1fr));gap:12px;margin:20px 0}
+.stat{border:1px solid var(--line);border-radius:8px;padding:12px;display:flex;flex-direction:column}
+.stat .v{font-size:26px;font-weight:600}.stat .k{color:var(--muted);font-size:13px}
+.findings{border-radius:8px;padding:4px 16px;border:1px solid var(--line)}
+.findings.ok{border-left:4px solid var(--ok)}.findings.warn{border-left:4px solid var(--warn)}
+.findings h2{margin:12px 0 4px;font-size:16px}
+.stack{border:1px solid var(--line);border-radius:8px;padding:12px;margin:12px 0}
+.row{display:flex;align-items:center;gap:12px;margin:6px 0}
+.lab{flex:0 0 110px;color:var(--muted);font-size:13px}
+.cells{display:flex;flex-wrap:wrap;gap:2px}
+.cells i{width:12px;height:12px;border-radius:2px;display:block}
+.bar{flex:1;display:flex;gap:2px;height:16px}.bar i{display:block;border-radius:2px;min-width:2px}
+table{border-collapse:collapse;width:100%;font-size:13px;margin:8px 0;display:block;overflow-x:auto}
+th,td{text-align:left;padding:4px 10px 4px 0;border-bottom:1px solid var(--line);vertical-align:top}
+th{color:var(--muted);font-weight:500}
+td:first-child{font-family:ui-monospace,monospace;overflow-wrap:anywhere}
+td i{display:inline-block;width:10px;height:10px;border-radius:2px;margin-right:6px}
+tr.bad td{color:var(--warn)}
+.c0{background:var(--s0)}.c1{background:var(--s1)}.c2{background:var(--s2)}.c3{background:var(--s3)}.c4{background:var(--s4)}.c5{background:var(--s5)}.c6{background:var(--s6)}.c7{background:var(--s7)}.c8{background:var(--s8)}
+</style></head><body><main>
+<h1>Shrunk architecture</h1><p class="meta">huggingface-mic/GLM-4.5-Air-mic@abc7ea97d6d633936cfd54900edce3b610755f86 ← zai-org/GLM-4.5-Air@a24ceef6ce4f3536971efe9b778bdaa1bab18daa · Glm4MoeForCausalLM · random weights</p>
+<div class="stats"><div class="stat"><span class="k">Original</span><span class="v">106.9B</span><span class="k">parameters</span></div><div class="stat"><span class="k">Fixture</span><span class="v">19.5M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Layer stacks</span><span class="v">1</span><span class="k">compared</span></div><div class="stat"><span class="k">Findings</span><span class="v">0</span><span class="k">to review</span></div></div><section class="findings ok"><h2>✓ Same architecture, smaller</h2><p>Every module class, layer kind and checked property of the original is in the fixture.</p></section><h2>Layer stacks</h2><p class="note">One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.</p><div class="stack"><h3>model.layers</h3><div class="row"><span class="lab">Original · 46</span><div class="cells"><i class="c0" title="layer 0: no Experts"></i><i class="c1" title="layer 1: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 2: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 3: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 4: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 5: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 6: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 7: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 8: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 9: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 10: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 11: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 12: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 13: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 14: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 15: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 16: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 17: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 18: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 19: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 20: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 21: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 22: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 23: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 24: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 25: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 26: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 27: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 28: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 29: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 30: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 31: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 32: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 33: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 34: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 35: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 36: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 37: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 38: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 39: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 40: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 41: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 42: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 43: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 44: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 45: Experts, MoE, TopkRouter"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: no Experts"></i><i class="c1" title="layer 1: Experts, MoE, TopkRouter"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>no Experts</td><td>1</td><td>1</td><td>✓ kept</td></tr><tr><td><i class="c1"></i>Experts, MoE, TopkRouter</td><td>45</td><td>1</td><td>✓ kept</td></tr></table></div><h2>Where the parameters are</h2><p class="note">Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.</p><div class="row"><span class="lab">Original</span><div class="bar"><i class="c0" style="flex:0.988381" title="model.layers: 105.6B (98.8%)"></i><i class="c1" style="flex:0.005809" title="model.embed_tokens: 620.8M (0.6%)"></i><i class="c2" style="flex:0.005809" title="lm_head: 620.8M (0.6%)"></i><i class="c3" style="flex:0.000000" title="model.norm: 4.1K (0.0%)"></i></div></div><div class="row"><span class="lab">Fixture</span><div class="bar"><i class="c0" style="flex:0.004140" title="model.layers: 80.6K (0.4%)"></i><i class="c1" style="flex:0.497928" title="model.embed_tokens: 9.7M (49.8%)"></i><i class="c2" style="flex:0.497928" title="lm_head: 9.7M (49.8%)"></i><i class="c3" style="flex:0.000003" title="model.norm: 64 (0.0%)"></i></div></div><table class="legend"><tr><th>Component</th><th>Original</th><th>Share</th><th>Fixture</th><th>Share</th></tr><tr><td><i class="c0"></i>model.layers</td><td>105.6B</td><td>98.8%</td><td>80.6K</td><td>0.4%</td></tr><tr><td><i class="c1"></i>model.embed_tokens</td><td>620.8M</td><td>0.6%</td><td>9.7M</td><td>49.8%</td></tr><tr><td><i class="c2"></i>lm_head</td><td>620.8M</td><td>0.6%</td><td>9.7M</td><td>49.8%</td></tr><tr><td><i class="c3"></i>model.norm</td><td>4.1K</td><td>0.0%</td><td>64</td><td>0.0%</td></tr></table><h2>Preserved properties</h2><table><tr><th>Config</th><th>Property</th><th>Original</th><th>Fixture</th><th></th></tr><tr class=""><td>model</td><td>attention style</td><td>GQA (96 heads, 8 KV)</td><td>GQA (4 heads, 2 KV)</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>MoE routing</td><td>top-8 of 128 experts</td><td>top-2 of 4 experts</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>RoPE type</td><td>default</td><td>default</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>model_type</td><td>glm4_moe</td><td>glm4_moe</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>vocab_size</td><td>151552</td><td>151552</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>tie_word_embeddings</td><td>False</td><td>False</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>hidden_act</td><td>silu</td><td>silu</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>max_position_embeddings</td><td>131072</td><td>131072</td><td>✓ kept</td></tr></table><h2>Module classes</h2><p class="note">Instances of each class in the built model (prefix Glm4Moe omitted).</p><table><tr><th>Class</th><th>Original</th><th>Fixture</th></tr><tr class=""><td>Linear</td><td>323</td><td>15</td></tr><tr class=""><td>RMSNorm</td><td>93</td><td>5</td></tr><tr class=""><td>SiLUActivation</td><td>91</td><td>3</td></tr><tr class=""><td>Attention</td><td>46</td><td>2</td></tr><tr class=""><td>DecoderLayer</td><td>46</td><td>2</td></tr><tr class=""><td>MLP</td><td>46</td><td>2</td></tr><tr class=""><td>Experts</td><td>45</td><td>1</td></tr><tr class=""><td>MoE</td><td>45</td><td>1</td></tr><tr class=""><td>TopkRouter</td><td>45</td><td>1</td></tr><tr class=""><td>Embedding</td><td>1</td><td>1</td></tr><tr class=""><td>ForCausalLM</td><td>1</td><td>1</td></tr><tr class=""><td>Model</td><td>1</td><td>1</td></tr><tr class=""><td>RotaryEmbedding</td><td>1</td><td>1</td></tr><tr class=""><td>ModuleList</td><td>1</td><td>1</td></tr></table><h2>Config changes</h2><p class="note">9 fields changed, 31 unchanged.</p><table><tr><th>Field</th><th>Original</th><th>Fixture</th></tr><tr><td>head_dim</td><td>128</td><td>16</td></tr><tr><td>hidden_size</td><td>4096</td><td>64</td></tr><tr><td>intermediate_size</td><td>10944</td><td>128</td></tr><tr><td>moe_intermediate_size</td><td>1408</td><td>32</td></tr><tr><td>n_routed_experts</td><td>128</td><td>4</td></tr><tr><td>num_attention_heads</td><td>96</td><td>4</td></tr><tr><td>num_experts_per_tok</td><td>8</td><td>2</td></tr><tr><td>num_hidden_layers</td><td>46</td><td>2</td></tr><tr><td>num_key_value_heads</td><td>8</td><td>2</td></tr></table>
+</main></body></html>

shrink/GLM-4.7-Flash-mic.html

@@ -0,0 +1,34 @@
+<!doctype html>
+<html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
+<title>Shrunk architecture</title>
+<style>
+:root{--bg:#fcfcfb;--fg:#0b0b0b;--muted:#52514e;--line:#e4e3df;--ok:#1d7a43;--warn:#b35c00;--s0:#2a78d6;--s1:#eb6834;--s2:#1baf7a;--s3:#eda100;--s4:#e87ba4;--s5:#008300;--s6:#6250d6;--s7:#e34948;--s8:#8a8984;}
+@media (prefers-color-scheme:dark){:root:not([data-theme="light"]){--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}}
+:root[data-theme="dark"]{--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}
+body{margin:0;background:var(--bg);color:var(--fg);font:15px/1.5 system-ui,sans-serif}
+main{max-width:1040px;margin:0 auto;padding:24px 16px 64px}
+h1{font-size:24px;margin:0 0 4px}h2{font-size:18px;margin:36px 0 8px}h3{font-size:14px;margin:0 0 8px;font-family:ui-monospace,monospace}
+.meta,.note{color:var(--muted);margin:0 0 12px;overflow-wrap:anywhere}
+.stats{display:grid;grid-template-columns:repeat(auto-fit,minmax(150px,1fr));gap:12px;margin:20px 0}
+.stat{border:1px solid var(--line);border-radius:8px;padding:12px;display:flex;flex-direction:column}
+.stat .v{font-size:26px;font-weight:600}.stat .k{color:var(--muted);font-size:13px}
+.findings{border-radius:8px;padding:4px 16px;border:1px solid var(--line)}
+.findings.ok{border-left:4px solid var(--ok)}.findings.warn{border-left:4px solid var(--warn)}
+.findings h2{margin:12px 0 4px;font-size:16px}
+.stack{border:1px solid var(--line);border-radius:8px;padding:12px;margin:12px 0}
+.row{display:flex;align-items:center;gap:12px;margin:6px 0}
+.lab{flex:0 0 110px;color:var(--muted);font-size:13px}
+.cells{display:flex;flex-wrap:wrap;gap:2px}
+.cells i{width:12px;height:12px;border-radius:2px;display:block}
+.bar{flex:1;display:flex;gap:2px;height:16px}.bar i{display:block;border-radius:2px;min-width:2px}
+table{border-collapse:collapse;width:100%;font-size:13px;margin:8px 0;display:block;overflow-x:auto}
+th,td{text-align:left;padding:4px 10px 4px 0;border-bottom:1px solid var(--line);vertical-align:top}
+th{color:var(--muted);font-weight:500}
+td:first-child{font-family:ui-monospace,monospace;overflow-wrap:anywhere}
+td i{display:inline-block;width:10px;height:10px;border-radius:2px;margin-right:6px}
+tr.bad td{color:var(--warn)}
+.c0{background:var(--s0)}.c1{background:var(--s1)}.c2{background:var(--s2)}.c3{background:var(--s3)}.c4{background:var(--s4)}.c5{background:var(--s5)}.c6{background:var(--s6)}.c7{background:var(--s7)}.c8{background:var(--s8)}
+</style></head><body><main>
+<h1>Shrunk architecture</h1><p class="meta">huggingface-mic/GLM-4.7-Flash-mic@8b4c9af3fc9e6a8325e3dade3558514edd2cd694 ← zai-org/GLM-4.7-Flash@7dd20894a642a0aa287e9827cb1a1f7f91386b67 · Glm4MoeLiteForCausalLM · random weights</p>
+<div class="stats"><div class="stat"><span class="k">Original</span><span class="v">29.9B</span><span class="k">parameters</span></div><div class="stat"><span class="k">Fixture</span><span class="v">19.9M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Layer stacks</span><span class="v">1</span><span class="k">compared</span></div><div class="stat"><span class="k">Findings</span><span class="v">0</span><span class="k">to review</span></div></div><section class="findings ok"><h2>✓ Same architecture, smaller</h2><p>Every module class, layer kind and checked property of the original is in the fixture.</p></section><h2>Layer stacks</h2><p class="note">One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.</p><div class="stack"><h3>model.layers</h3><div class="row"><span class="lab">Original · 47</span><div class="cells"><i class="c0" title="layer 0: no Experts"></i><i class="c1" title="layer 1: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 2: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 3: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 4: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 5: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 6: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 7: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 8: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 9: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 10: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 11: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 12: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 13: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 14: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 15: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 16: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 17: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 18: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 19: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 20: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 21: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 22: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 23: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 24: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 25: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 26: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 27: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 28: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 29: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 30: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 31: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 32: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 33: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 34: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 35: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 36: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 37: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 38: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 39: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 40: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 41: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 42: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 43: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 44: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 45: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 46: Experts, MoE, TopkRouter"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: no Experts"></i><i class="c1" title="layer 1: Experts, MoE, TopkRouter"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>no Experts</td><td>1</td><td>1</td><td>✓ kept</td></tr><tr><td><i class="c1"></i>Experts, MoE, TopkRouter</td><td>46</td><td>1</td><td>✓ kept</td></tr></table></div><h2>Where the parameters are</h2><p class="note">Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.</p><div class="row"><span class="lab">Original</span><div class="bar"><i class="c0" style="flex:0.978814" title="model.layers: 29.3B (97.9%)"></i><i class="c1" style="flex:0.010593" title="model.embed_tokens: 317.2M (1.1%)"></i><i class="c2" style="flex:0.010593" title="lm_head: 317.2M (1.1%)"></i><i class="c3" style="flex:0.000000" title="model.norm: 2.0K (0.0%)"></i></div></div><div class="row"><span class="lab">Fixture</span><div class="bar"><i class="c0" style="flex:0.003837" title="model.layers: 76.4K (0.4%)"></i><i class="c1" style="flex:0.498080" title="model.embed_tokens: 9.9M (49.8%)"></i><i class="c2" style="flex:0.498080" title="lm_head: 9.9M (49.8%)"></i><i class="c3" style="flex:0.000003" title="model.norm: 64 (0.0%)"></i></div></div><table class="legend"><tr><th>Component</th><th>Original</th><th>Share</th><th>Fixture</th><th>Share</th></tr><tr><td><i class="c0"></i>model.layers</td><td>29.3B</td><td>97.9%</td><td>76.4K</td><td>0.4%</td></tr><tr><td><i class="c1"></i>model.embed_tokens</td><td>317.2M</td><td>1.1%</td><td>9.9M</td><td>49.8%</td></tr><tr><td><i class="c2"></i>lm_head</td><td>317.2M</td><td>1.1%</td><td>9.9M</td><td>49.8%</td></tr><tr><td><i class="c3"></i>model.norm</td><td>2.0K</td><td>0.0%</td><td>64</td><td>0.0%</td></tr></table><h2>Preserved properties</h2><table><tr><th>Config</th><th>Property</th><th>Original</th><th>Fixture</th><th></th></tr><tr class=""><td>model</td><td>attention style</td><td>MHA (20 heads, 20 KV)</td><td>MHA (4 heads, 4 KV)</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>MoE routing</td><td>top-4 of 64 experts</td><td>top-2 of 4 experts</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>RoPE type</td><td>default</td><td>default</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>model_type</td><td>glm4_moe_lite</td><td>glm4_moe_lite</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>vocab_size</td><td>154880</td><td>154880</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>tie_word_embeddings</td><td>False</td><td>False</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>hidden_act</td><td>silu</td><td>silu</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>max_position_embeddings</td><td>202752</td><td>202752</td><td>✓ kept</td></tr></table><h2>Module classes</h2><p class="note">Instances of each class in the built model (prefix Glm4MoeLite omitted).</p><table><tr><th>Class</th><th>Original</th><th>Fixture</th></tr><tr class=""><td>Linear</td><td>377</td><td>17</td></tr><tr class=""><td>RMSNorm</td><td>189</td><td>9</td></tr><tr class=""><td>SiLUActivation</td><td>93</td><td>3</td></tr><tr class=""><td>Attention</td><td>47</td><td>2</td></tr><tr class=""><td>DecoderLayer</td><td>47</td><td>2</td></tr><tr class=""><td>MLP</td><td>47</td><td>2</td></tr><tr class=""><td>Experts</td><td>46</td><td>1</td></tr><tr class=""><td>MoE</td><td>46</td><td>1</td></tr><tr class=""><td>TopkRouter</td><td>46</td><td>1</td></tr><tr class=""><td>Embedding</td><td>1</td><td>1</td></tr><tr class=""><td>ForCausalLM</td><td>1</td><td>1</td></tr><tr class=""><td>Model</td><td>1</td><td>1</td></tr><tr class=""><td>RotaryEmbedding</td><td>1</td><td>1</td></tr><tr class=""><td>ModuleList</td><td>1</td><td>1</td></tr></table><h2>Config changes</h2><p class="note">14 fields changed, 34 unchanged.</p><table><tr><th>Field</th><th>Original</th><th>Fixture</th></tr><tr><td>hidden_size</td><td>2048</td><td>64</td></tr><tr><td>intermediate_size</td><td>10240</td><td>128</td></tr><tr><td>kv_lora_rank</td><td>512</td><td>16</td></tr><tr><td>mlp_layer_types</td><td>47 entries: 46× sparse, 1× dense</td><td>[&quot;dense&quot;, &quot;sparse&quot;]</td></tr><tr><td>moe_intermediate_size</td><td>1536</td><td>32</td></tr><tr><td>n_routed_experts</td><td>64</td><td>4</td></tr><tr><td>num_attention_heads</td><td>20</td><td>4</td></tr><tr><td>num_experts_per_tok</td><td>4</td><td>2</td></tr><tr><td>num_hidden_layers</td><td>47</td><td>2</td></tr><tr><td>num_key_value_heads</td><td>20</td><td>4</td></tr><tr><td>q_lora_rank</td><td>768</td><td>16</td></tr><tr><td>qk_nope_head_dim</td><td>192</td><td>16</td></tr><tr><td>qk_rope_head_dim</td><td>64</td><td>8</td></tr><tr><td>v_head_dim</td><td>256</td><td>16</td></tr></table>
+</main></body></html>

shrink/GLM-5.3-Flash-mic.html

@@ -0,0 +1,34 @@
+<!doctype html>
+<html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
+<title>Shrunk architecture</title>
+<style>
+:root{--bg:#fcfcfb;--fg:#0b0b0b;--muted:#52514e;--line:#e4e3df;--ok:#1d7a43;--warn:#b35c00;--s0:#2a78d6;--s1:#eb6834;--s2:#1baf7a;--s3:#eda100;--s4:#e87ba4;--s5:#008300;--s6:#6250d6;--s7:#e34948;--s8:#8a8984;}
+@media (prefers-color-scheme:dark){:root:not([data-theme="light"]){--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}}
+:root[data-theme="dark"]{--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}
+body{margin:0;background:var(--bg);color:var(--fg);font:15px/1.5 system-ui,sans-serif}
+main{max-width:1040px;margin:0 auto;padding:24px 16px 64px}
+h1{font-size:24px;margin:0 0 4px}h2{font-size:18px;margin:36px 0 8px}h3{font-size:14px;margin:0 0 8px;font-family:ui-monospace,monospace}
+.meta,.note{color:var(--muted);margin:0 0 12px;overflow-wrap:anywhere}
+.stats{display:grid;grid-template-columns:repeat(auto-fit,minmax(150px,1fr));gap:12px;margin:20px 0}
+.stat{border:1px solid var(--line);border-radius:8px;padding:12px;display:flex;flex-direction:column}
+.stat .v{font-size:26px;font-weight:600}.stat .k{color:var(--muted);font-size:13px}
+.findings{border-radius:8px;padding:4px 16px;border:1px solid var(--line)}
+.findings.ok{border-left:4px solid var(--ok)}.findings.warn{border-left:4px solid var(--warn)}
+.findings h2{margin:12px 0 4px;font-size:16px}
+.stack{border:1px solid var(--line);border-radius:8px;padding:12px;margin:12px 0}
+.row{display:flex;align-items:center;gap:12px;margin:6px 0}
+.lab{flex:0 0 110px;color:var(--muted);font-size:13px}
+.cells{display:flex;flex-wrap:wrap;gap:2px}
+.cells i{width:12px;height:12px;border-radius:2px;display:block}
+.bar{flex:1;display:flex;gap:2px;height:16px}.bar i{display:block;border-radius:2px;min-width:2px}
+table{border-collapse:collapse;width:100%;font-size:13px;margin:8px 0;display:block;overflow-x:auto}
+th,td{text-align:left;padding:4px 10px 4px 0;border-bottom:1px solid var(--line);vertical-align:top}
+th{color:var(--muted);font-weight:500}
+td:first-child{font-family:ui-monospace,monospace;overflow-wrap:anywhere}
+td i{display:inline-block;width:10px;height:10px;border-radius:2px;margin-right:6px}
+tr.bad td{color:var(--warn)}
+.c0{background:var(--s0)}.c1{background:var(--s1)}.c2{background:var(--s2)}.c3{background:var(--s3)}.c4{background:var(--s4)}.c5{background:var(--s5)}.c6{background:var(--s6)}.c7{background:var(--s7)}.c8{background:var(--s8)}
+</style></head><body><main>
+<h1>Shrunk architecture</h1><p class="meta">huggingface-mic/GLM-5.3-Flash-mic@53f3e5bd10a67d2c75a170a3c2a28009e338c256 ← zai-org/GLM-5.3-Flash@46a050bec13002ad29ee795c16e61cf4fa3c4a9c · Glm5NextForConditionalGeneration · random weights · built before the shrink rule (no recorded recipe)</p>
+<div class="stats"><div class="stat"><span class="k">Original</span><span class="v">313.9B</span><span class="k">parameters</span></div><div class="stat"><span class="k">Fixture</span><span class="v">40.7M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Layer stacks</span><span class="v">2</span><span class="k">compared</span></div><div class="stat"><span class="k">Findings</span><span class="v">1</span><span class="k">to review</span></div></div><section class="findings warn"><h2>⚠ To review</h2><ul><li>vision_config: image_size 448 became 56</li></ul></section><h2>Layer stacks</h2><p class="note">One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.</p><div class="stack"><h3>model.visual.blocks</h3><div class="row"><span class="lab">Original · 24</span><div class="cells"><i class="c0" title="layer 0: VisionBlock"></i><i class="c0" title="layer 1: VisionBlock"></i><i class="c0" title="layer 2: VisionBlock"></i><i class="c0" title="layer 3: VisionBlock"></i><i class="c0" title="layer 4: VisionBlock"></i><i class="c0" title="layer 5: VisionBlock"></i><i class="c0" title="layer 6: VisionBlock"></i><i class="c0" title="layer 7: VisionBlock"></i><i class="c0" title="layer 8: VisionBlock"></i><i class="c0" title="layer 9: VisionBlock"></i><i class="c0" title="layer 10: VisionBlock"></i><i class="c0" title="layer 11: VisionBlock"></i><i class="c0" title="layer 12: VisionBlock"></i><i class="c0" title="layer 13: VisionBlock"></i><i class="c0" title="layer 14: VisionBlock"></i><i class="c0" title="layer 15: VisionBlock"></i><i class="c0" title="layer 16: VisionBlock"></i><i class="c0" title="layer 17: VisionBlock"></i><i class="c0" title="layer 18: VisionBlock"></i><i class="c0" title="layer 19: VisionBlock"></i><i class="c0" title="layer 20: VisionBlock"></i><i class="c0" title="layer 21: VisionBlock"></i><i class="c0" title="layer 22: VisionBlock"></i><i class="c0" title="layer 23: VisionBlock"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: VisionBlock"></i><i class="c0" title="layer 1: VisionBlock"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>VisionBlock</td><td>24</td><td>2</td><td>✓ kept</td></tr></table></div><div class="stack"><h3>model.language_model.layers</h3><div class="row"><span class="lab">Original · 45</span><div class="cells"><i class="c0" title="layer 0: linear_attention, Conv1d, TextForgetGate …"></i><i class="c0" title="layer 1: linear_attention, Conv1d, TextForgetGate …"></i><i class="c0" title="layer 2: linear_attention, Conv1d, TextForgetGate …"></i><i class="c1" title="layer 3: is_causal=True, TextAttention, TextExperts …"></i><i class="c2" title="layer 4: linear_attention, Conv1d, TextExperts …"></i><i class="c2" title="layer 5: linear_attention, Conv1d, TextExperts …"></i><i class="c2" title="layer 6: linear_attention, Conv1d, TextExperts …"></i><i class="c1" title="layer 7: is_causal=True, TextAttention, TextExperts …"></i><i class="c2" title="layer 8: linear_attention, Conv1d, TextExperts …"></i><i class="c2" title="layer 9: linear_attention, Conv1d, TextExperts …"></i><i class="c2" title="layer 10: linear_attention, Conv1d, TextExperts …"></i><i class="c1" title="layer 11: is_causal=True, TextAttention, TextExperts …"></i><i class="c2" title="layer 12: linear_attention, Conv1d, TextExperts …"></i><i class="c2" title="layer 13: linear_attention, Conv1d, TextExperts …"></i><i class="c2" title="layer 14: linear_attention, Conv1d, TextExperts …"></i><i class="c1" title="layer 15: is_causal=True, TextAttention, TextExperts …"></i><i class="c2" title="layer 16: linear_attention, Conv1d, TextExperts …"></i><i class="c2" title="layer 17: linear_attention, Conv1d, TextExperts …"></i><i class="c2" title="layer 18: linear_attention, Conv1d, TextExperts …"></i><i class="c1" title="layer 19: is_causal=True, TextAttention, TextExperts …"></i><i class="c2" title="layer 20: linear_attention, Conv1d, TextExperts …"></i><i class="c2" title="layer 21: linear_attention, Conv1d, TextExperts …"></i><i class="c2" title="layer 22: linear_attention, Conv1d, TextExperts …"></i><i class="c1" title="layer 23: is_causal=True, TextAttention, TextExperts …"></i><i class="c2" title="layer 24: linear_attention, Conv1d, TextExperts …"></i><i class="c2" title="layer 25: linear_attention, Conv1d, TextExperts …"></i><i class="c2" title="layer 26: linear_attention, Conv1d, TextExperts …"></i><i class="c1" title="layer 27: is_causal=True, TextAttention, TextExperts …"></i><i class="c2" title="layer 28: linear_attention, Conv1d, TextExperts …"></i><i class="c2" title="layer 29: linear_attention, Conv1d, TextExperts …"></i><i class="c2" title="layer 30: linear_attention, Conv1d, TextExperts …"></i><i class="c1" title="layer 31: is_causal=True, TextAttention, TextExperts …"></i><i class="c2" title="layer 32: linear_attention, Conv1d, TextExperts …"></i><i class="c2" title="layer 33: linear_attention, Conv1d, TextExperts …"></i><i class="c2" title="layer 34: linear_attention, Conv1d, TextExperts …"></i><i class="c1" title="layer 35: is_causal=True, TextAttention, TextExperts …"></i><i class="c2" title="layer 36: linear_attention, Conv1d, TextExperts …"></i><i class="c2" title="layer 37: linear_attention, Conv1d, TextExperts …"></i><i class="c2" title="layer 38: linear_attention, Conv1d, TextExperts …"></i><i class="c1" title="layer 39: is_causal=True, TextAttention, TextExperts …"></i><i class="c2" title="layer 40: linear_attention, Conv1d, TextExperts …"></i><i class="c2" title="layer 41: linear_attention, Conv1d, TextExperts …"></i><i class="c2" title="layer 42: linear_attention, Conv1d, TextExperts …"></i><i class="c1" title="layer 43: is_causal=True, TextAttention, TextExperts …"></i><i class="c2" title="layer 44: linear_attention, Conv1d, TextExperts …"></i></div></div><div class="row"><span class="lab">Fixture · 3</span><div class="cells"><i class="c0" title="layer 0: linear_attention, Conv1d, TextForgetGate …"></i><i class="c1" title="layer 1: is_causal=True, TextAttention, TextExperts …"></i><i class="c2" title="layer 2: linear_attention, Conv1d, TextExperts …"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>linear_attention, Conv1d, TextForgetGate …</td><td>3</td><td>1</td><td>✓ kept</td></tr><tr><td><i class="c1"></i>is_causal=True, TextAttention, TextExperts …</td><td>11</td><td>1</td><td>✓ kept</td></tr><tr><td><i class="c2"></i>linear_attention, Conv1d, TextExperts …</td><td>31</td><td>1</td><td>✓ kept</td></tr></table></div><h2>Where the parameters are</h2><p class="note">Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.</p><div class="row"><span class="lab">Original</span><div class="bar"><i class="c0" style="flex:0.994162" title="model.language_model.layers: 312.1B (99.4%)"></i><i class="c1" style="flex:0.002021" title="model.language_model: 634.4M (0.2%)"></i><i class="c2" style="flex:0.002021" title="lm_head: 634.4M (0.2%)"></i><i class="c3" style="flex:0.001284" title="model.visual.blocks: 403.0M (0.1%)"></i><i class="c4" style="flex:0.000512" title="model.visual: 160.6M (0.1%)"></i></div></div><div class="row"><span class="lab">Fixture</span><div class="bar"><i class="c0" style="flex:0.019919" title="model.language_model.layers: 811.0K (2.0%)"></i><i class="c1" style="flex:0.486881" title="model.language_model: 19.8M (48.7%)"></i><i class="c2" style="flex:0.486877" title="lm_head: 19.8M (48.7%)"></i><i class="c3" style="flex:0.002048" title="model.visual.blocks: 83.4K (0.2%)"></i><i class="c4" style="flex:0.004275" title="model.visual: 174.1K (0.4%)"></i></div></div><table class="legend"><tr><th>Component</th><th>Original</th><th>Share</th><th>Fixture</th><th>Share</th></tr><tr><td><i class="c0"></i>model.language_model.layers</td><td>312.1B</td><td>99.4%</td><td>811.0K</td><td>2.0%</td></tr><tr><td><i class="c1"></i>model.language_model</td><td>634.4M</td><td>0.2%</td><td>19.8M</td><td>48.7%</td></tr><tr><td><i class="c2"></i>lm_head</td><td>634.4M</td><td>0.2%</td><td>19.8M</td><td>48.7%</td></tr><tr><td><i class="c3"></i>model.visual.blocks</td><td>403.0M</td><td>0.1%</td><td>83.4K</td><td>0.2%</td></tr><tr><td><i class="c4"></i>model.visual</td><td>160.6M</td><td>0.1%</td><td>174.1K</td><td>0.4%</td></tr></table><h2>Preserved properties</h2><table><tr><th>Config</th><th>Property</th><th>Original</th><th>Fixture</th><th></th></tr><tr class=""><td>model</td><td>model_type</td><td>glm5_next</td><td>glm5_next</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>tie_word_embeddings</td><td>False</td><td>False</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>attention style</td><td>MHA (64 heads, 64 KV)</td><td>MHA (4 heads, 4 KV)</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>MoE routing</td><td>top-8 of 288 experts</td><td>top-2 of 8 experts</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>model_type</td><td>glm5_next_text</td><td>glm5_next_text</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>vocab_size</td><td>154880</td><td>154880</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>tie_word_embeddings</td><td>False</td><td>False</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>hidden_act</td><td>silu</td><td>silu</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>max_position_embeddings</td><td>1048576</td><td>1048576</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>RoPE type</td><td>axial</td><td>axial</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>model_type</td><td>glm5_next_vision</td><td>glm5_next_vision</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>hidden_act</td><td>silu</td><td>silu</td><td>✓ kept</td></tr><tr class="bad"><td>vision_config</td><td>image_size</td><td>448</td><td>56</td><td>⚠ changed</td></tr><tr class=""><td>vision_config</td><td>patch_size</td><td>14</td><td>14</td><td>✓ kept</td></tr></table><h2>Module classes</h2><p class="note">Instances of each class in the built model (prefix Glm5Next omitted).</p><table><tr><th>Class</th><th>Original</th><th>Fixture</th></tr><tr class=""><td>Linear</td><td>654</td><td>50</td></tr><tr class=""><td>TextRMSNorm</td><td>113</td><td>9</td></tr><tr class=""><td>RMSNorm</td><td>97</td><td>9</td></tr><tr class=""><td>TextHyperConnection</td><td>90</td><td>6</td></tr><tr class=""><td>TextUnweightedRMSNorm</td><td>90</td><td>6</td></tr><tr class=""><td>SiLUActivation</td><td>70</td><td>6</td></tr><tr class=""><td>TextDecoderLayer</td><td>45</td><td>3</td></tr><tr class=""><td>TextMLP</td><td>45</td><td>3</td></tr><tr class=""><td>TextExperts</td><td>42</td><td>2</td></tr><tr class=""><td>TextMoE</td><td>42</td><td>2</td></tr><tr class=""><td>TextTopkRouter</td><td>42</td><td>2</td></tr><tr class=""><td>Conv1d</td><td>34</td><td>2</td></tr><tr class=""><td>TextForgetGate</td><td>34</td><td>2</td></tr><tr class=""><td>TextLinearAttention</td><td>34</td><td>2</td></tr><tr class=""><td>TextRMSNormGated</td><td>34</td><td>2</td></tr><tr class=""><td>VisionAttention</td><td>24</td><td>2</td></tr><tr class=""><td>VisionBlock</td><td>24</td><td>2</td></tr><tr class=""><td>VisionMLP</td><td>24</td><td>2</td></tr><tr class=""><td>LayerNorm</td><td>12</td><td>2</td></tr><tr class=""><td>TextAttention</td><td>11</td><td>1</td></tr><tr class=""><td>TextIndexer</td><td>11</td><td>1</td></tr><tr class=""><td>ModuleList</td><td>2</td><td>2</td></tr><tr class=""><td>Conv2d</td><td>1</td><td>1</td></tr><tr class=""><td>Conv3d</td><td>1</td><td>1</td></tr><tr class=""><td>Embedding</td><td>1</td><td>1</td></tr><tr class=""><td>GELU</td><td>1</td><td>1</td></tr><tr class=""><td>ForConditionalGeneration</td><td>1</td><td>1</td></tr><tr class=""><td>Model</td><td>1</td><td>1</td></tr><tr class=""><td>TextHyperHead</td><td>1</td><td>1</td></tr><tr class=""><td>TextModel</td><td>1</td><td>1</td></tr><tr class=""><td>VisionModel</td><td>1</td><td>1</td></tr><tr class=""><td>VisionPatchEmbed</td><td>1</td><td>1</td></tr><tr class=""><td>VisionPatchMerger</td><td>1</td><td>1</td></tr><tr class=""><td>VisionRotaryEmbedding</td><td>1</td><td>1</td></tr></table><h2>Config changes</h2><p class="note">33 fields changed, 86 unchanged.</p><table><tr><th>Field</th><th>Original</th><th>Fixture</th></tr><tr><td>quantization_config</td><td>{&quot;activation_scheme&quot;: &quot;dynamic&quot;, &quot;fmt&quot;: &quot;e4m3&quot;, &quot;modules_to_not_convert&quot;: [&quot;a…</td><td>—</td></tr><tr><td>text_config.first_k_dense_replace</td><td>3</td><td>1</td></tr><tr><td>text_config.hidden_size</td><td>4096</td><td>128</td></tr><tr><td>text_config.index_head_dim</td><td>128</td><td>32</td></tr><tr><td>text_config.index_n_heads</td><td>32</td><td>4</td></tr><tr><td>text_config.index_topk</td><td>2048</td><td>64</td></tr><tr><td>text_config.indexer_types</td><td>45 entries: 45× full</td><td>[&quot;full&quot;, &quot;full&quot;, &quot;full&quot;]</td></tr><tr><td>text_config.intermediate_size</td><td>12288</td><td>128</td></tr><tr><td>text_config.kv_lora_rank</td><td>512</td><td>64</td></tr><tr><td>text_config.layer_types</td><td>45 entries: 34× linear_attention, 11× indexed_attention</td><td>[&quot;linear_attention&quot;, &quot;indexed_attention&quot;, &quot;linear_attention&quot;]</td></tr><tr><td>text_config.linear_attn_config.full_attn_layers</td><td>11 entries</td><td>[1]</td></tr><tr><td>text_config.linear_attn_config.head_dim</td><td>128</td><td>32</td></tr><tr><td>text_config.linear_attn_config.kda_layers</td><td>34 entries</td><td>[0, 2]</td></tr><tr><td>text_config.linear_attn_config.num_heads</td><td>64</td><td>4</td></tr><tr><td>text_config.linear_head_dim</td><td>128</td><td>32</td></tr><tr><td>text_config.linear_num_heads</td><td>64</td><td>4</td></tr><tr><td>text_config.mlp_layer_types</td><td>45 entries: 42× sparse, 3× dense</td><td>[&quot;dense&quot;, &quot;sparse&quot;, &quot;sparse&quot;]</td></tr><tr><td>text_config.moe_intermediate_size</td><td>2048</td><td>64</td></tr><tr><td>text_config.n_routed_experts</td><td>288</td><td>8</td></tr><tr><td>text_config.num_attention_heads</td><td>64</td><td>4</td></tr><tr><td>text_config.num_experts_per_tok</td><td>8</td><td>2</td></tr><tr><td>text_config.num_hidden_layers</td><td>45</td><td>3</td></tr><tr><td>text_config.num_key_value_heads</td><td>64</td><td>4</td></tr><tr><td>text_config.q_lora_rank</td><td>1536</td><td>64</td></tr><tr><td>text_config.qk_nope_head_dim</td><td>256</td><td>32</td></tr><tr><td>text_config.v_head_dim</td><td>256</td><td>32</td></tr><tr><td>vision_config.depth</td><td>24</td><td>2</td></tr><tr><td>vision_config.hidden_size</td><td>1024</td><td>64</td></tr><tr><td>vision_config.image_size</td><td>448</td><td>56</td></tr><tr><td>vision_config.intermediate_size</td><td>4096</td><td>128</td></tr><tr><td>vision_config.num_heads</td><td>16</td><td>4</td></tr><tr><td>vision_config.out_hidden_size</td><td>4096</td><td>128</td></tr><tr><td>vision_config.projection_intermediate_size</td><td>10240</td><td>128</td></tr></table>
+</main></body></html>

shrink/GLM-5.3-mic.html

@@ -0,0 +1,34 @@
+<!doctype html>
+<html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
+<title>Shrunk architecture</title>
+<style>
+:root{--bg:#fcfcfb;--fg:#0b0b0b;--muted:#52514e;--line:#e4e3df;--ok:#1d7a43;--warn:#b35c00;--s0:#2a78d6;--s1:#eb6834;--s2:#1baf7a;--s3:#eda100;--s4:#e87ba4;--s5:#008300;--s6:#6250d6;--s7:#e34948;--s8:#8a8984;}
+@media (prefers-color-scheme:dark){:root:not([data-theme="light"]){--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}}
+:root[data-theme="dark"]{--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}
+body{margin:0;background:var(--bg);color:var(--fg);font:15px/1.5 system-ui,sans-serif}
+main{max-width:1040px;margin:0 auto;padding:24px 16px 64px}
+h1{font-size:24px;margin:0 0 4px}h2{font-size:18px;margin:36px 0 8px}h3{font-size:14px;margin:0 0 8px;font-family:ui-monospace,monospace}
+.meta,.note{color:var(--muted);margin:0 0 12px;overflow-wrap:anywhere}
+.stats{display:grid;grid-template-columns:repeat(auto-fit,minmax(150px,1fr));gap:12px;margin:20px 0}
+.stat{border:1px solid var(--line);border-radius:8px;padding:12px;display:flex;flex-direction:column}
+.stat .v{font-size:26px;font-weight:600}.stat .k{color:var(--muted);font-size:13px}
+.findings{border-radius:8px;padding:4px 16px;border:1px solid var(--line)}
+.findings.ok{border-left:4px solid var(--ok)}.findings.warn{border-left:4px solid var(--warn)}
+.findings h2{margin:12px 0 4px;font-size:16px}
+.stack{border:1px solid var(--line);border-radius:8px;padding:12px;margin:12px 0}
+.row{display:flex;align-items:center;gap:12px;margin:6px 0}
+.lab{flex:0 0 110px;color:var(--muted);font-size:13px}
+.cells{display:flex;flex-wrap:wrap;gap:2px}
+.cells i{width:12px;height:12px;border-radius:2px;display:block}
+.bar{flex:1;display:flex;gap:2px;height:16px}.bar i{display:block;border-radius:2px;min-width:2px}
+table{border-collapse:collapse;width:100%;font-size:13px;margin:8px 0;display:block;overflow-x:auto}
+th,td{text-align:left;padding:4px 10px 4px 0;border-bottom:1px solid var(--line);vertical-align:top}
+th{color:var(--muted);font-weight:500}
+td:first-child{font-family:ui-monospace,monospace;overflow-wrap:anywhere}
+td i{display:inline-block;width:10px;height:10px;border-radius:2px;margin-right:6px}
+tr.bad td{color:var(--warn)}
+.c0{background:var(--s0)}.c1{background:var(--s1)}.c2{background:var(--s2)}.c3{background:var(--s3)}.c4{background:var(--s4)}.c5{background:var(--s5)}.c6{background:var(--s6)}.c7{background:var(--s7)}.c8{background:var(--s8)}
+</style></head><body><main>
+<h1>Shrunk architecture</h1><p class="meta">huggingface-mic/GLM-5.3-mic@1088c4f5a18bf3ba097240a67f6c391be8072006 ← zai-org/GLM-5.3@ba32f8c19c8738e2c57c71798c601d477a7bf844 · GlmMoeDsaForCausalLM · random weights</p>
+<div class="stats"><div class="stat"><span class="k">Original</span><span class="v">743.4B</span><span class="k">parameters</span></div><div class="stat"><span class="k">Fixture</span><span class="v">41.0M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Layer stacks</span><span class="v">1</span><span class="k">compared</span></div><div class="stat"><span class="k">Findings</span><span class="v">1</span><span class="k">to review</span></div></div><section class="findings warn"><h2>⚠ To review</h2><ul><li>model: max_position_embeddings 1048576 became 512</li></ul></section><h2>Layer stacks</h2><p class="note">One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.</p><div class="stack"><h3>model.layers</h3><div class="row"><span class="lab">Original · 78</span><div class="cells"><i class="c0" title="layer 0: Indexer, LayerNorm"></i><i class="c0" title="layer 1: Indexer, LayerNorm"></i><i class="c0" title="layer 2: Indexer, LayerNorm"></i><i class="c1" title="layer 3: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 4: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 5: Experts, MoE, TopkRouter"></i><i class="c2" title="layer 6: Experts, Indexer, MoE …"></i><i class="c1" title="layer 7: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 8: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 9: Experts, MoE, TopkRouter"></i><i class="c2" title="layer 10: Experts, Indexer, MoE …"></i><i class="c1" title="layer 11: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 12: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 13: Experts, MoE, TopkRouter"></i><i class="c2" title="layer 14: Experts, Indexer, MoE …"></i><i class="c1" title="layer 15: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 16: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 17: Experts, MoE, TopkRouter"></i><i class="c2" title="layer 18: Experts, Indexer, MoE ���"></i><i class="c1" title="layer 19: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 20: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 21: Experts, MoE, TopkRouter"></i><i class="c2" title="layer 22: Experts, Indexer, MoE …"></i><i class="c1" title="layer 23: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 24: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 25: Experts, MoE, TopkRouter"></i><i class="c2" title="layer 26: Experts, Indexer, MoE …"></i><i class="c1" title="layer 27: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 28: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 29: Experts, MoE, TopkRouter"></i><i class="c2" title="layer 30: Experts, Indexer, MoE …"></i><i class="c1" title="layer 31: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 32: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 33: Experts, MoE, TopkRouter"></i><i class="c2" title="layer 34: Experts, Indexer, MoE …"></i><i class="c1" title="layer 35: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 36: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 37: Experts, MoE, TopkRouter"></i><i class="c2" title="layer 38: Experts, Indexer, MoE …"></i><i class="c1" title="layer 39: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 40: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 41: Experts, MoE, TopkRouter"></i><i class="c2" title="layer 42: Experts, Indexer, MoE …"></i><i class="c1" title="layer 43: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 44: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 45: Experts, MoE, TopkRouter"></i><i class="c2" title="layer 46: Experts, Indexer, MoE …"></i><i class="c1" title="layer 47: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 48: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 49: Experts, MoE, TopkRouter"></i><i class="c2" title="layer 50: Experts, Indexer, MoE …"></i><i class="c1" title="layer 51: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 52: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 53: Experts, MoE, TopkRouter"></i><i class="c2" title="layer 54: Experts, Indexer, MoE …"></i><i class="c1" title="layer 55: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 56: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 57: Experts, MoE, TopkRouter"></i><i class="c2" title="layer 58: Experts, Indexer, MoE …"></i><i class="c1" title="layer 59: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 60: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 61: Experts, MoE, TopkRouter"></i><i class="c2" title="layer 62: Experts, Indexer, MoE …"></i><i class="c1" title="layer 63: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 64: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 65: Experts, MoE, TopkRouter"></i><i class="c2" title="layer 66: Experts, Indexer, MoE …"></i><i class="c1" title="layer 67: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 68: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 69: Experts, MoE, TopkRouter"></i><i class="c2" title="layer 70: Experts, Indexer, MoE …"></i><i class="c1" title="layer 71: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 72: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 73: Experts, MoE, TopkRouter"></i><i class="c2" title="layer 74: Experts, Indexer, MoE …"></i><i class="c1" title="layer 75: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 76: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 77: Experts, MoE, TopkRouter"></i></div></div><div class="row"><span class="lab">Fixture · 7</span><div class="cells"><i class="c0" title="layer 0: Indexer, LayerNorm"></i><i class="c0" title="layer 1: Indexer, LayerNorm"></i><i class="c0" title="layer 2: Indexer, LayerNorm"></i><i class="c1" title="layer 3: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 4: Experts, MoE, TopkRouter"></i><i class="c1" title="layer 5: Experts, MoE, TopkRouter"></i><i class="c2" title="layer 6: Experts, Indexer, MoE …"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>Indexer, LayerNorm</td><td>3</td><td>3</td><td>✓ kept</td></tr><tr><td><i class="c1"></i>Experts, MoE, TopkRouter</td><td>57</td><td>3</td><td>✓ kept</td></tr><tr><td><i class="c2"></i>Experts, Indexer, MoE …</td><td>18</td><td>1</td><td>✓ kept</td></tr></table></div><h2>Where the parameters are</h2><p class="note">Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.</p><div class="row"><span class="lab">Original</span><div class="bar"><i class="c0" style="flex:0.997440" title="model.layers: 741.5B (99.7%)"></i><i class="c1" style="flex:0.001280" title="lm_head: 951.6M (0.1%)"></i><i class="c2" style="flex:0.001280" title="model.embed_tokens: 951.6M (0.1%)"></i><i class="c3" style="flex:0.000000" title="model.norm: 6.1K (0.0%)"></i></div></div><div class="row"><span class="lab">Fixture</span><div class="bar"><i class="c0" style="flex:0.033438" title="model.layers: 1.4M (3.3%)"></i><i class="c1" style="flex:0.483279" title="lm_head: 19.8M (48.3%)"></i><i class="c2" style="flex:0.483279" title="model.embed_tokens: 19.8M (48.3%)"></i><i class="c3" style="flex:0.000003" title="model.norm: 128 (0.0%)"></i></div></div><table class="legend"><tr><th>Component</th><th>Original</th><th>Share</th><th>Fixture</th><th>Share</th></tr><tr><td><i class="c0"></i>model.layers</td><td>741.5B</td><td>99.7%</td><td>1.4M</td><td>3.3%</td></tr><tr><td><i class="c1"></i>lm_head</td><td>951.6M</td><td>0.1%</td><td>19.8M</td><td>48.3%</td></tr><tr><td><i class="c2"></i>model.embed_tokens</td><td>951.6M</td><td>0.1%</td><td>19.8M</td><td>48.3%</td></tr><tr><td><i class="c3"></i>model.norm</td><td>6.1K</td><td>0.0%</td><td>128</td><td>0.0%</td></tr></table><h2>Preserved properties</h2><table><tr><th>Config</th><th>Property</th><th>Original</th><th>Fixture</th><th></th></tr><tr class=""><td>model</td><td>attention style</td><td>MHA (64 heads, 64 KV)</td><td>MHA (4 heads, 4 KV)</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>MoE routing</td><td>top-8 of 256 experts</td><td>top-2 of 8 experts</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>RoPE type</td><td>default</td><td>default</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>model_type</td><td>glm_moe_dsa</td><td>glm_moe_dsa</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>vocab_size</td><td>154880</td><td>154880</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>tie_word_embeddings</td><td>False</td><td>False</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>hidden_act</td><td>silu</td><td>silu</td><td>✓ kept</td></tr><tr class="bad"><td>model</td><td>max_position_embeddings</td><td>1048576</td><td>512</td><td>⚠ changed</td></tr></table><h2>Module classes</h2><p class="note">Instances of each class in the built model (prefix GlmMoeDsa omitted).</p><table><tr><th>Class</th><th>Original</th><th>Fixture</th></tr><tr class=""><td>Linear</td><td>688</td><td>69</td></tr><tr class=""><td>RMSNorm</td><td>313</td><td>29</td></tr><tr class=""><td>SiLUActivation</td><td>153</td><td>11</td></tr><tr class=""><td>Attention</td><td>78</td><td>7</td></tr><tr class=""><td>DecoderLayer</td><td>78</td><td>7</td></tr><tr class=""><td>MLP</td><td>78</td><td>7</td></tr><tr class=""><td>Experts</td><td>75</td><td>4</td></tr><tr class=""><td>MoE</td><td>75</td><td>4</td></tr><tr class=""><td>TopkRouter</td><td>75</td><td>4</td></tr><tr class=""><td>Indexer</td><td>21</td><td>4</td></tr><tr class=""><td>LayerNorm</td><td>21</td><td>4</td></tr><tr class=""><td>Embedding</td><td>1</td><td>1</td></tr><tr class=""><td>ForCausalLM</td><td>1</td><td>1</td></tr><tr class=""><td>Model</td><td>1</td><td>1</td></tr><tr class=""><td>RotaryEmbedding</td><td>1</td><td>1</td></tr><tr class=""><td>ModuleList</td><td>1</td><td>1</td></tr></table><h2>Config changes</h2><p class="note">23 fields changed, 42 unchanged.</p><table><tr><th>Field</th><th>Original</th><th>Fixture</th></tr><tr><td>first_k_dense_replace</td><td>3</td><td>1</td></tr><tr><td>head_dim</td><td>64</td><td>8</td></tr><tr><td>hidden_size</td><td>6144</td><td>128</td></tr><tr><td>index_head_dim</td><td>128</td><td>16</td></tr><tr><td>index_n_heads</td><td>32</td><td>2</td></tr><tr><td>index_topk</td><td>2048</td><td>8</td></tr><tr><td>indexer_types</td><td>78 entries: 57× shared, 21× full</td><td>7 entries: 4× full, 3× shared</td></tr><tr><td>intermediate_size</td><td>12288</td><td>128</td></tr><tr><td>kv_lora_rank</td><td>512</td><td>32</td></tr><tr><td>layer_types</td><td>78 entries: 78× indexed_attention</td><td>7 entries: 7× indexed_attention</td></tr><tr><td>max_position_embeddings</td><td>1048576</td><td>512</td></tr><tr><td>mlp_layer_types</td><td>78 entries: 75× sparse, 3× dense</td><td>7 entries: 4× sparse, 3× dense</td></tr><tr><td>moe_intermediate_size</td><td>2048</td><td>64</td></tr><tr><td>n_routed_experts</td><td>256</td><td>8</td></tr><tr><td>num_attention_heads</td><td>64</td><td>4</td></tr><tr><td>num_experts_per_tok</td><td>8</td><td>2</td></tr><tr><td>num_hidden_layers</td><td>78</td><td>7</td></tr><tr><td>num_key_value_heads</td><td>64</td><td>4</td></tr><tr><td>q_lora_rank</td><td>2048</td><td>64</td></tr><tr><td>qk_nope_head_dim</td><td>192</td><td>24</td></tr><tr><td>qk_rope_head_dim</td><td>64</td><td>8</td></tr><tr><td>quantization_config</td><td>{&quot;activation_scheme&quot;: &quot;dynamic&quot;, &quot;fmt&quot;: &quot;e4m3&quot;, &quot;quant_method&quot;: &quot;fp8&quot;, &quot;weigh…</td><td>—</td></tr><tr><td>v_head_dim</td><td>256</td><td>32</td></tr></table>
+</main></body></html>

shrink/GLM-OCR-mic.html

@@ -0,0 +1,34 @@
+<!doctype html>
+<html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
+<title>Shrunk architecture</title>
+<style>
+:root{--bg:#fcfcfb;--fg:#0b0b0b;--muted:#52514e;--line:#e4e3df;--ok:#1d7a43;--warn:#b35c00;--s0:#2a78d6;--s1:#eb6834;--s2:#1baf7a;--s3:#eda100;--s4:#e87ba4;--s5:#008300;--s6:#6250d6;--s7:#e34948;--s8:#8a8984;}
+@media (prefers-color-scheme:dark){:root:not([data-theme="light"]){--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}}
+:root[data-theme="dark"]{--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}
+body{margin:0;background:var(--bg);color:var(--fg);font:15px/1.5 system-ui,sans-serif}
+main{max-width:1040px;margin:0 auto;padding:24px 16px 64px}
+h1{font-size:24px;margin:0 0 4px}h2{font-size:18px;margin:36px 0 8px}h3{font-size:14px;margin:0 0 8px;font-family:ui-monospace,monospace}
+.meta,.note{color:var(--muted);margin:0 0 12px;overflow-wrap:anywhere}
+.stats{display:grid;grid-template-columns:repeat(auto-fit,minmax(150px,1fr));gap:12px;margin:20px 0}
+.stat{border:1px solid var(--line);border-radius:8px;padding:12px;display:flex;flex-direction:column}
+.stat .v{font-size:26px;font-weight:600}.stat .k{color:var(--muted);font-size:13px}
+.findings{border-radius:8px;padding:4px 16px;border:1px solid var(--line)}
+.findings.ok{border-left:4px solid var(--ok)}.findings.warn{border-left:4px solid var(--warn)}
+.findings h2{margin:12px 0 4px;font-size:16px}
+.stack{border:1px solid var(--line);border-radius:8px;padding:12px;margin:12px 0}
+.row{display:flex;align-items:center;gap:12px;margin:6px 0}
+.lab{flex:0 0 110px;color:var(--muted);font-size:13px}
+.cells{display:flex;flex-wrap:wrap;gap:2px}
+.cells i{width:12px;height:12px;border-radius:2px;display:block}
+.bar{flex:1;display:flex;gap:2px;height:16px}.bar i{display:block;border-radius:2px;min-width:2px}
+table{border-collapse:collapse;width:100%;font-size:13px;margin:8px 0;display:block;overflow-x:auto}
+th,td{text-align:left;padding:4px 10px 4px 0;border-bottom:1px solid var(--line);vertical-align:top}
+th{color:var(--muted);font-weight:500}
+td:first-child{font-family:ui-monospace,monospace;overflow-wrap:anywhere}
+td i{display:inline-block;width:10px;height:10px;border-radius:2px;margin-right:6px}
+tr.bad td{color:var(--warn)}
+.c0{background:var(--s0)}.c1{background:var(--s1)}.c2{background:var(--s2)}.c3{background:var(--s3)}.c4{background:var(--s4)}.c5{background:var(--s5)}.c6{background:var(--s6)}.c7{background:var(--s7)}.c8{background:var(--s8)}
+</style></head><body><main>
+<h1>Shrunk architecture</h1><p class="meta">huggingface-mic/GLM-OCR-mic@08eddf7ae838d5d3b6b135b984c0e3ec53c28441 ← zai-org/GLM-OCR@2e85a62840ccac27daa451df36c736c4636b8628 · GlmOcrForConditionalGeneration · random weights</p>
+<div class="stats"><div class="stat"><span class="k">Original</span><span class="v">1.1B</span><span class="k">parameters</span></div><div class="stat"><span class="k">Fixture</span><span class="v">8.8M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Layer stacks</span><span class="v">2</span><span class="k">compared</span></div><div class="stat"><span class="k">Findings</span><span class="v">0</span><span class="k">to review</span></div></div><section class="findings ok"><h2>✓ Same architecture, smaller</h2><p>Every module class, layer kind and checked property of the original is in the fixture.</p></section><h2>Layer stacks</h2><p class="note">One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.</p><div class="stack"><h3>model.visual.blocks</h3><div class="row"><span class="lab">Original · 24</span><div class="cells"><i class="c0" title="layer 0: VisionBlock"></i><i class="c0" title="layer 1: VisionBlock"></i><i class="c0" title="layer 2: VisionBlock"></i><i class="c0" title="layer 3: VisionBlock"></i><i class="c0" title="layer 4: VisionBlock"></i><i class="c0" title="layer 5: VisionBlock"></i><i class="c0" title="layer 6: VisionBlock"></i><i class="c0" title="layer 7: VisionBlock"></i><i class="c0" title="layer 8: VisionBlock"></i><i class="c0" title="layer 9: VisionBlock"></i><i class="c0" title="layer 10: VisionBlock"></i><i class="c0" title="layer 11: VisionBlock"></i><i class="c0" title="layer 12: VisionBlock"></i><i class="c0" title="layer 13: VisionBlock"></i><i class="c0" title="layer 14: VisionBlock"></i><i class="c0" title="layer 15: VisionBlock"></i><i class="c0" title="layer 16: VisionBlock"></i><i class="c0" title="layer 17: VisionBlock"></i><i class="c0" title="layer 18: VisionBlock"></i><i class="c0" title="layer 19: VisionBlock"></i><i class="c0" title="layer 20: VisionBlock"></i><i class="c0" title="layer 21: VisionBlock"></i><i class="c0" title="layer 22: VisionBlock"></i><i class="c0" title="layer 23: VisionBlock"></i></div></div><div class="row"><span class="lab">Fixture · 24</span><div class="cells"><i class="c0" title="layer 0: VisionBlock"></i><i class="c0" title="layer 1: VisionBlock"></i><i class="c0" title="layer 2: VisionBlock"></i><i class="c0" title="layer 3: VisionBlock"></i><i class="c0" title="layer 4: VisionBlock"></i><i class="c0" title="layer 5: VisionBlock"></i><i class="c0" title="layer 6: VisionBlock"></i><i class="c0" title="layer 7: VisionBlock"></i><i class="c0" title="layer 8: VisionBlock"></i><i class="c0" title="layer 9: VisionBlock"></i><i class="c0" title="layer 10: VisionBlock"></i><i class="c0" title="layer 11: VisionBlock"></i><i class="c0" title="layer 12: VisionBlock"></i><i class="c0" title="layer 13: VisionBlock"></i><i class="c0" title="layer 14: VisionBlock"></i><i class="c0" title="layer 15: VisionBlock"></i><i class="c0" title="layer 16: VisionBlock"></i><i class="c0" title="layer 17: VisionBlock"></i><i class="c0" title="layer 18: VisionBlock"></i><i class="c0" title="layer 19: VisionBlock"></i><i class="c0" title="layer 20: VisionBlock"></i><i class="c0" title="layer 21: VisionBlock"></i><i class="c0" title="layer 22: VisionBlock"></i><i class="c0" title="layer 23: VisionBlock"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>VisionBlock</td><td>24</td><td>24</td><td>✓ kept</td></tr></table></div><div class="stack"><h3>model.language_model.layers</h3><div class="row"><span class="lab">Original · 16</span><div class="cells"><i class="c0" title="layer 0: TextDecoderLayer"></i><i class="c0" title="layer 1: TextDecoderLayer"></i><i class="c0" title="layer 2: TextDecoderLayer"></i><i class="c0" title="layer 3: TextDecoderLayer"></i><i class="c0" title="layer 4: TextDecoderLayer"></i><i class="c0" title="layer 5: TextDecoderLayer"></i><i class="c0" title="layer 6: TextDecoderLayer"></i><i class="c0" title="layer 7: TextDecoderLayer"></i><i class="c0" title="layer 8: TextDecoderLayer"></i><i class="c0" title="layer 9: TextDecoderLayer"></i><i class="c0" title="layer 10: TextDecoderLayer"></i><i class="c0" title="layer 11: TextDecoderLayer"></i><i class="c0" title="layer 12: TextDecoderLayer"></i><i class="c0" title="layer 13: TextDecoderLayer"></i><i class="c0" title="layer 14: TextDecoderLayer"></i><i class="c0" title="layer 15: TextDecoderLayer"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: TextDecoderLayer"></i><i class="c0" title="layer 1: TextDecoderLayer"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>TextDecoderLayer</td><td>16</td><td>2</td><td>✓ kept</td></tr></table></div><h2>Where the parameters are</h2><p class="note">Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.</p><div class="row"><span class="lab">Original</span><div class="bar"><i class="c0" style="flex:0.443227" title="model.language_model.layers: 490.8M (44.3%)"></i><i class="c1" style="flex:0.363936" title="model.visual.blocks: 403.0M (36.4%)"></i><i class="c2" style="flex:0.082380" title="model.language_model: 91.2M (8.2%)"></i><i class="c3" style="flex:0.082378" title="lm_head: 91.2M (8.2%)"></i><i class="c4" style="flex:0.028079" title="model.visual: 31.1M (2.8%)"></i></div></div><div class="row"><span class="lab">Fixture</span><div class="bar"><i class="c0" style="flex:0.008427" title="model.language_model.layers: 74.2K (0.8%)"></i><i class="c1" style="flex:0.113586" title="model.visual.blocks: 1.0M (11.4%)"></i><i class="c2" style="flex:0.431453" title="model.language_model: 3.8M (43.1%)"></i><i class="c3" style="flex:0.431446" title="lm_head: 3.8M (43.1%)"></i><i class="c4" style="flex:0.015088" title="model.visual: 132.9K (1.5%)"></i></div></div><table class="legend"><tr><th>Component</th><th>Original</th><th>Share</th><th>Fixture</th><th>Share</th></tr><tr><td><i class="c0"></i>model.language_model.layers</td><td>490.8M</td><td>44.3%</td><td>74.2K</td><td>0.8%</td></tr><tr><td><i class="c1"></i>model.visual.blocks</td><td>403.0M</td><td>36.4%</td><td>1.0M</td><td>11.4%</td></tr><tr><td><i class="c2"></i>model.language_model</td><td>91.2M</td><td>8.2%</td><td>3.8M</td><td>43.1%</td></tr><tr><td><i class="c3"></i>lm_head</td><td>91.2M</td><td>8.2%</td><td>3.8M</td><td>43.1%</td></tr><tr><td><i class="c4"></i>model.visual</td><td>31.1M</td><td>2.8%</td><td>132.9K</td><td>1.5%</td></tr></table><h2>Preserved properties</h2><table><tr><th>Config</th><th>Property</th><th>Original</th><th>Fixture</th><th></th></tr><tr class=""><td>model</td><td>model_type</td><td>glm_ocr</td><td>glm_ocr</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>tie_word_embeddings</td><td>False</td><td>False</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>attention style</td><td>GQA (16 heads, 8 KV)</td><td>GQA (4 heads, 2 KV)</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>RoPE type</td><td>default</td><td>default</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>model_type</td><td>glm_ocr_text</td><td>glm_ocr_text</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>vocab_size</td><td>59392</td><td>59392</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>tie_word_embeddings</td><td>False</td><td>False</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>hidden_act</td><td>silu</td><td>silu</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>max_position_embeddings</td><td>131072</td><td>131072</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>RoPE type</td><td>axial</td><td>axial</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>model_type</td><td>glm_ocr_vision</td><td>glm_ocr_vision</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>hidden_act</td><td>silu</td><td>silu</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>image_size</td><td>336</td><td>336</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>patch_size</td><td>14</td><td>14</td><td>✓ kept</td></tr></table><h2>Module classes</h2><p class="note">Instances of each class in the built model (prefix GlmOcr omitted).</p><table><tr><th>Class</th><th>Original</th><th>Fixture</th></tr><tr class=""><td>Linear</td><td>221</td><td>137</td></tr><tr class=""><td>RMSNorm</td><td>162</td><td>106</td></tr><tr class=""><td>SiLUActivation</td><td>41</td><td>27</td></tr><tr class=""><td>VisionAttention</td><td>24</td><td>24</td></tr><tr class=""><td>VisionBlock</td><td>24</td><td>24</td></tr><tr class=""><td>VisionMlp</td><td>24</td><td>24</td></tr><tr class=""><td>TextAttention</td><td>16</td><td>2</td></tr><tr class=""><td>TextDecoderLayer</td><td>16</td><td>2</td></tr><tr class=""><td>TextMLP</td><td>16</td><td>2</td></tr><tr class=""><td>ModuleList</td><td>2</td><td>2</td></tr><tr class=""><td>Conv2d</td><td>1</td><td>1</td></tr><tr class=""><td>Conv3d</td><td>1</td><td>1</td></tr><tr class=""><td>Embedding</td><td>1</td><td>1</td></tr><tr class=""><td>GELU</td><td>1</td><td>1</td></tr><tr class=""><td>ForConditionalGeneration</td><td>1</td><td>1</td></tr><tr class=""><td>Model</td><td>1</td><td>1</td></tr><tr class=""><td>TextModel</td><td>1</td><td>1</td></tr><tr class=""><td>TextRotaryEmbedding</td><td>1</td><td>1</td></tr><tr class=""><td>VisionModel</td><td>1</td><td>1</td></tr><tr class=""><td>VisionPatchEmbed</td><td>1</td><td>1</td></tr><tr class=""><td>VisionPatchMerger</td><td>1</td><td>1</td></tr><tr class=""><td>VisionRotaryEmbedding</td><td>1</td><td>1</td></tr><tr class=""><td>LayerNorm</td><td>1</td><td>1</td></tr></table><h2>Config changes</h2><p class="note">11 fields changed, 59 unchanged.</p><table><tr><th>Field</th><th>Original</th><th>Fixture</th></tr><tr><td>text_config.head_dim</td><td>128</td><td>16</td></tr><tr><td>text_config.hidden_size</td><td>1536</td><td>64</td></tr><tr><td>text_config.intermediate_size</td><td>4608</td><td>128</td></tr><tr><td>text_config.num_attention_heads</td><td>16</td><td>4</td></tr><tr><td>text_config.num_hidden_layers</td><td>16</td><td>2</td></tr><tr><td>text_config.num_key_value_heads</td><td>8</td><td>2</td></tr><tr><td>text_config.rope_parameters</td><td>{&quot;mrope_section&quot;: [16, 24, 24], &quot;partial_rotary_factor&quot;: 1.0, &quot;rope_theta&quot;: 1…</td><td>{&quot;mrope_section&quot;: [2, 3, 3], &quot;partial_rotary_factor&quot;: 1.0, &quot;rope_theta&quot;: 1000…</td></tr><tr><td>vision_config.hidden_size</td><td>1024</td><td>64</td></tr><tr><td>vision_config.intermediate_size</td><td>4096</td><td>128</td></tr><tr><td>vision_config.num_heads</td><td>16</td><td>4</td></tr><tr><td>vision_config.out_hidden_size</td><td>1536</td><td>64</td></tr></table>
+</main></body></html>

shrink/Hunyuan-0.5B-Instruct-mic.html

@@ -0,0 +1,34 @@
+<!doctype html>
+<html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
+<title>Shrunk architecture</title>
+<style>
+:root{--bg:#fcfcfb;--fg:#0b0b0b;--muted:#52514e;--line:#e4e3df;--ok:#1d7a43;--warn:#b35c00;--s0:#2a78d6;--s1:#eb6834;--s2:#1baf7a;--s3:#eda100;--s4:#e87ba4;--s5:#008300;--s6:#6250d6;--s7:#e34948;--s8:#8a8984;}
+@media (prefers-color-scheme:dark){:root:not([data-theme="light"]){--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}}
+:root[data-theme="dark"]{--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}
+body{margin:0;background:var(--bg);color:var(--fg);font:15px/1.5 system-ui,sans-serif}
+main{max-width:1040px;margin:0 auto;padding:24px 16px 64px}
+h1{font-size:24px;margin:0 0 4px}h2{font-size:18px;margin:36px 0 8px}h3{font-size:14px;margin:0 0 8px;font-family:ui-monospace,monospace}
+.meta,.note{color:var(--muted);margin:0 0 12px;overflow-wrap:anywhere}
+.stats{display:grid;grid-template-columns:repeat(auto-fit,minmax(150px,1fr));gap:12px;margin:20px 0}
+.stat{border:1px solid var(--line);border-radius:8px;padding:12px;display:flex;flex-direction:column}
+.stat .v{font-size:26px;font-weight:600}.stat .k{color:var(--muted);font-size:13px}
+.findings{border-radius:8px;padding:4px 16px;border:1px solid var(--line)}
+.findings.ok{border-left:4px solid var(--ok)}.findings.warn{border-left:4px solid var(--warn)}
+.findings h2{margin:12px 0 4px;font-size:16px}
+.stack{border:1px solid var(--line);border-radius:8px;padding:12px;margin:12px 0}
+.row{display:flex;align-items:center;gap:12px;margin:6px 0}
+.lab{flex:0 0 110px;color:var(--muted);font-size:13px}
+.cells{display:flex;flex-wrap:wrap;gap:2px}
+.cells i{width:12px;height:12px;border-radius:2px;display:block}
+.bar{flex:1;display:flex;gap:2px;height:16px}.bar i{display:block;border-radius:2px;min-width:2px}
+table{border-collapse:collapse;width:100%;font-size:13px;margin:8px 0;display:block;overflow-x:auto}
+th,td{text-align:left;padding:4px 10px 4px 0;border-bottom:1px solid var(--line);vertical-align:top}
+th{color:var(--muted);font-weight:500}
+td:first-child{font-family:ui-monospace,monospace;overflow-wrap:anywhere}
+td i{display:inline-block;width:10px;height:10px;border-radius:2px;margin-right:6px}
+tr.bad td{color:var(--warn)}
+.c0{background:var(--s0)}.c1{background:var(--s1)}.c2{background:var(--s2)}.c3{background:var(--s3)}.c4{background:var(--s4)}.c5{background:var(--s5)}.c6{background:var(--s6)}.c7{background:var(--s7)}.c8{background:var(--s8)}
+</style></head><body><main>
+<h1>Shrunk architecture</h1><p class="meta">huggingface-mic/Hunyuan-0.5B-Instruct-mic@d7fe47dc3288bba155076d3747959365159073b0 ← tencent/Hunyuan-0.5B-Instruct@2359fb220c010e9d6d62c62d466f0eda179c2cf3 · HunYuanDenseV1ForCausalLM · random weights</p>
+<div class="stats"><div class="stat"><span class="k">Original</span><span class="v">539.0M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Fixture</span><span class="v">7.8M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Layer stacks</span><span class="v">1</span><span class="k">compared</span></div><div class="stat"><span class="k">Findings</span><span class="v">0</span><span class="k">to review</span></div></div><section class="findings ok"><h2>✓ Same architecture, smaller</h2><p>Every module class, layer kind and checked property of the original is in the fixture.</p></section><h2>Layer stacks</h2><p class="note">One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.</p><div class="stack"><h3>model.layers</h3><div class="row"><span class="lab">Original · 24</span><div class="cells"><i class="c0" title="layer 0: DecoderLayer"></i><i class="c0" title="layer 1: DecoderLayer"></i><i class="c0" title="layer 2: DecoderLayer"></i><i class="c0" title="layer 3: DecoderLayer"></i><i class="c0" title="layer 4: DecoderLayer"></i><i class="c0" title="layer 5: DecoderLayer"></i><i class="c0" title="layer 6: DecoderLayer"></i><i class="c0" title="layer 7: DecoderLayer"></i><i class="c0" title="layer 8: DecoderLayer"></i><i class="c0" title="layer 9: DecoderLayer"></i><i class="c0" title="layer 10: DecoderLayer"></i><i class="c0" title="layer 11: DecoderLayer"></i><i class="c0" title="layer 12: DecoderLayer"></i><i class="c0" title="layer 13: DecoderLayer"></i><i class="c0" title="layer 14: DecoderLayer"></i><i class="c0" title="layer 15: DecoderLayer"></i><i class="c0" title="layer 16: DecoderLayer"></i><i class="c0" title="layer 17: DecoderLayer"></i><i class="c0" title="layer 18: DecoderLayer"></i><i class="c0" title="layer 19: DecoderLayer"></i><i class="c0" title="layer 20: DecoderLayer"></i><i class="c0" title="layer 21: DecoderLayer"></i><i class="c0" title="layer 22: DecoderLayer"></i><i class="c0" title="layer 23: DecoderLayer"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: DecoderLayer"></i><i class="c0" title="layer 1: DecoderLayer"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>DecoderLayer</td><td>24</td><td>2</td><td>✓ kept</td></tr></table></div><h2>Where the parameters are</h2><p class="note">Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.</p><div class="row"><span class="lab">Original</span><div class="bar"><i class="c0" style="flex:0.770471" title="model.layers: 415.3M (77.0%)"></i><i class="c1" style="flex:0.229528" title="model.embed_tokens: 123.7M (23.0%)"></i><i class="c2" style="flex:0.000002" title="model.norm: 1.0K (0.0%)"></i></div></div><div class="row"><span class="lab">Fixture</span><div class="bar"><i class="c0" style="flex:0.009485" title="model.layers: 74.0K (0.9%)"></i><i class="c1" style="flex:0.990506" title="model.embed_tokens: 7.7M (99.1%)"></i><i class="c2" style="flex:0.000008" title="model.norm: 64 (0.0%)"></i></div></div><table class="legend"><tr><th>Component</th><th>Original</th><th>Share</th><th>Fixture</th><th>Share</th></tr><tr><td><i class="c0"></i>model.layers</td><td>415.3M</td><td>77.0%</td><td>74.0K</td><td>0.9%</td></tr><tr><td><i class="c1"></i>model.embed_tokens</td><td>123.7M</td><td>23.0%</td><td>7.7M</td><td>99.1%</td></tr><tr><td><i class="c2"></i>model.norm</td><td>1.0K</td><td>0.0%</td><td>64</td><td>0.0%</td></tr></table><h2>Preserved properties</h2><table><tr><th>Config</th><th>Property</th><th>Original</th><th>Fixture</th><th></th></tr><tr class=""><td>model</td><td>attention style</td><td>GQA (16 heads, 8 KV)</td><td>GQA (4 heads, 2 KV)</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>RoPE type</td><td>dynamic</td><td>dynamic</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>model_type</td><td>hunyuan_v1_dense</td><td>hunyuan_v1_dense</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>vocab_size</td><td>120818</td><td>120818</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>tie_word_embeddings</td><td>True</td><td>True</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>hidden_act</td><td>silu</td><td>silu</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>max_position_embeddings</td><td>262144</td><td>262144</td><td>✓ kept</td></tr></table><h2>Module classes</h2><p class="note">Instances of each class in the built model (prefix HunYuanDenseV1 omitted).</p><table><tr><th>Class</th><th>Original</th><th>Fixture</th></tr><tr class=""><td>Linear</td><td>169</td><td>15</td></tr><tr class=""><td>RMSNorm</td><td>97</td><td>9</td></tr><tr class=""><td>Attention</td><td>24</td><td>2</td></tr><tr class=""><td>DecoderLayer</td><td>24</td><td>2</td></tr><tr class=""><td>MLP</td><td>24</td><td>2</td></tr><tr class=""><td>SiLUActivation</td><td>24</td><td>2</td></tr><tr class=""><td>Embedding</td><td>1</td><td>1</td></tr><tr class=""><td>ForCausalLM</td><td>1</td><td>1</td></tr><tr class=""><td>Model</td><td>1</td><td>1</td></tr><tr class=""><td>RotaryEmbedding</td><td>1</td><td>1</td></tr><tr class=""><td>ModuleList</td><td>1</td><td>1</td></tr></table><h2>Config changes</h2><p class="note">7 fields changed, 43 unchanged.</p><table><tr><th>Field</th><th>Original</th><th>Fixture</th></tr><tr><td>attention_head_dim</td><td>128</td><td>16</td></tr><tr><td>head_dim</td><td>128</td><td>16</td></tr><tr><td>hidden_size</td><td>1024</td><td>64</td></tr><tr><td>intermediate_size</td><td>3584</td><td>128</td></tr><tr><td>num_attention_heads</td><td>16</td><td>4</td></tr><tr><td>num_hidden_layers</td><td>24</td><td>2</td></tr><tr><td>num_key_value_heads</td><td>8</td><td>2</td></tr></table>
+</main></body></html>

shrink/HunyuanOCR-mic.html

@@ -0,0 +1,34 @@
+<!doctype html>
+<html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
+<title>Shrunk architecture</title>
+<style>
+:root{--bg:#fcfcfb;--fg:#0b0b0b;--muted:#52514e;--line:#e4e3df;--ok:#1d7a43;--warn:#b35c00;--s0:#2a78d6;--s1:#eb6834;--s2:#1baf7a;--s3:#eda100;--s4:#e87ba4;--s5:#008300;--s6:#6250d6;--s7:#e34948;--s8:#8a8984;}
+@media (prefers-color-scheme:dark){:root:not([data-theme="light"]){--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}}
+:root[data-theme="dark"]{--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}
+body{margin:0;background:var(--bg);color:var(--fg);font:15px/1.5 system-ui,sans-serif}
+main{max-width:1040px;margin:0 auto;padding:24px 16px 64px}
+h1{font-size:24px;margin:0 0 4px}h2{font-size:18px;margin:36px 0 8px}h3{font-size:14px;margin:0 0 8px;font-family:ui-monospace,monospace}
+.meta,.note{color:var(--muted);margin:0 0 12px;overflow-wrap:anywhere}
+.stats{display:grid;grid-template-columns:repeat(auto-fit,minmax(150px,1fr));gap:12px;margin:20px 0}
+.stat{border:1px solid var(--line);border-radius:8px;padding:12px;display:flex;flex-direction:column}
+.stat .v{font-size:26px;font-weight:600}.stat .k{color:var(--muted);font-size:13px}
+.findings{border-radius:8px;padding:4px 16px;border:1px solid var(--line)}
+.findings.ok{border-left:4px solid var(--ok)}.findings.warn{border-left:4px solid var(--warn)}
+.findings h2{margin:12px 0 4px;font-size:16px}
+.stack{border:1px solid var(--line);border-radius:8px;padding:12px;margin:12px 0}
+.row{display:flex;align-items:center;gap:12px;margin:6px 0}
+.lab{flex:0 0 110px;color:var(--muted);font-size:13px}
+.cells{display:flex;flex-wrap:wrap;gap:2px}
+.cells i{width:12px;height:12px;border-radius:2px;display:block}
+.bar{flex:1;display:flex;gap:2px;height:16px}.bar i{display:block;border-radius:2px;min-width:2px}
+table{border-collapse:collapse;width:100%;font-size:13px;margin:8px 0;display:block;overflow-x:auto}
+th,td{text-align:left;padding:4px 10px 4px 0;border-bottom:1px solid var(--line);vertical-align:top}
+th{color:var(--muted);font-weight:500}
+td:first-child{font-family:ui-monospace,monospace;overflow-wrap:anywhere}
+td i{display:inline-block;width:10px;height:10px;border-radius:2px;margin-right:6px}
+tr.bad td{color:var(--warn)}
+.c0{background:var(--s0)}.c1{background:var(--s1)}.c2{background:var(--s2)}.c3{background:var(--s3)}.c4{background:var(--s4)}.c5{background:var(--s5)}.c6{background:var(--s6)}.c7{background:var(--s7)}.c8{background:var(--s8)}
+</style></head><body><main>
+<h1>Shrunk architecture</h1><p class="meta">huggingface-mic/HunyuanOCR-mic@2b69a85316ef4335c59c1a82d624555a1dfd7e71 ← tencent/HunyuanOCR@47644ecc4fc854efa4f505155158831f36773ee4 · HunYuanVLForConditionalGeneration · random weights</p>
+<div class="stats"><div class="stat"><span class="k">Original</span><span class="v">996.2M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Fixture</span><span class="v">9.1M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Layer stacks</span><span class="v">2</span><span class="k">compared</span></div><div class="stat"><span class="k">Findings</span><span class="v">0</span><span class="k">to review</span></div></div><section class="findings ok"><h2>✓ Same architecture, smaller</h2><p>Every module class, layer kind and checked property of the original is in the fixture.</p></section><h2>Layer stacks</h2><p class="note">One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.</p><div class="stack"><h3>model.language_model.layers</h3><div class="row"><span class="lab">Original · 24</span><div class="cells"><i class="c0" title="layer 0: DenseV1DecoderLayer"></i><i class="c0" title="layer 1: DenseV1DecoderLayer"></i><i class="c0" title="layer 2: DenseV1DecoderLayer"></i><i class="c0" title="layer 3: DenseV1DecoderLayer"></i><i class="c0" title="layer 4: DenseV1DecoderLayer"></i><i class="c0" title="layer 5: DenseV1DecoderLayer"></i><i class="c0" title="layer 6: DenseV1DecoderLayer"></i><i class="c0" title="layer 7: DenseV1DecoderLayer"></i><i class="c0" title="layer 8: DenseV1DecoderLayer"></i><i class="c0" title="layer 9: DenseV1DecoderLayer"></i><i class="c0" title="layer 10: DenseV1DecoderLayer"></i><i class="c0" title="layer 11: DenseV1DecoderLayer"></i><i class="c0" title="layer 12: DenseV1DecoderLayer"></i><i class="c0" title="layer 13: DenseV1DecoderLayer"></i><i class="c0" title="layer 14: DenseV1DecoderLayer"></i><i class="c0" title="layer 15: DenseV1DecoderLayer"></i><i class="c0" title="layer 16: DenseV1DecoderLayer"></i><i class="c0" title="layer 17: DenseV1DecoderLayer"></i><i class="c0" title="layer 18: DenseV1DecoderLayer"></i><i class="c0" title="layer 19: DenseV1DecoderLayer"></i><i class="c0" title="layer 20: DenseV1DecoderLayer"></i><i class="c0" title="layer 21: DenseV1DecoderLayer"></i><i class="c0" title="layer 22: DenseV1DecoderLayer"></i><i class="c0" title="layer 23: DenseV1DecoderLayer"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: DenseV1DecoderLayer"></i><i class="c0" title="layer 1: DenseV1DecoderLayer"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>DenseV1DecoderLayer</td><td>24</td><td>2</td><td>✓ kept</td></tr></table></div><div class="stack"><h3>model.vision_tower.layers</h3><div class="row"><span class="lab">Original · 27</span><div class="cells"><i class="c0" title="layer 0: VisionBlock"></i><i class="c0" title="layer 1: VisionBlock"></i><i class="c0" title="layer 2: VisionBlock"></i><i class="c0" title="layer 3: VisionBlock"></i><i class="c0" title="layer 4: VisionBlock"></i><i class="c0" title="layer 5: VisionBlock"></i><i class="c0" title="layer 6: VisionBlock"></i><i class="c0" title="layer 7: VisionBlock"></i><i class="c0" title="layer 8: VisionBlock"></i><i class="c0" title="layer 9: VisionBlock"></i><i class="c0" title="layer 10: VisionBlock"></i><i class="c0" title="layer 11: VisionBlock"></i><i class="c0" title="layer 12: VisionBlock"></i><i class="c0" title="layer 13: VisionBlock"></i><i class="c0" title="layer 14: VisionBlock"></i><i class="c0" title="layer 15: VisionBlock"></i><i class="c0" title="layer 16: VisionBlock"></i><i class="c0" title="layer 17: VisionBlock"></i><i class="c0" title="layer 18: VisionBlock"></i><i class="c0" title="layer 19: VisionBlock"></i><i class="c0" title="layer 20: VisionBlock"></i><i class="c0" title="layer 21: VisionBlock"></i><i class="c0" title="layer 22: VisionBlock"></i><i class="c0" title="layer 23: VisionBlock"></i><i class="c0" title="layer 24: VisionBlock"></i><i class="c0" title="layer 25: VisionBlock"></i><i class="c0" title="layer 26: VisionBlock"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: VisionBlock"></i><i class="c0" title="layer 1: VisionBlock"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>VisionBlock</td><td>27</td><td>2</td><td>✓ kept</td></tr></table></div><h2>Where the parameters are</h2><p class="note">Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.</p><div class="row"><span class="lab">Original</span><div class="bar"><i class="c0" style="flex:0.416872" title="model.language_model.layers: 415.3M (41.7%)"></i><i class="c1" style="flex:0.413033" title="model.vision_tower.layers: 411.5M (41.3%)"></i><i class="c2" style="flex:0.124190" title="model.language_model: 123.7M (12.4%)"></i><i class="c3" style="flex:0.045906" title="model.vision_tower: 45.7M (4.6%)"></i></div></div><div class="row"><span class="lab">Fixture</span><div class="bar"><i class="c0" style="flex:0.008178" title="model.language_model.layers: 74.0K (0.8%)"></i><i class="c1" style="flex:0.007394" title="model.vision_tower.layers: 66.9K (0.7%)"></i><i class="c2" style="flex:0.854014" title="model.language_model: 7.7M (85.4%)"></i><i class="c3" style="flex:0.130414" title="model.vision_tower: 1.2M (13.0%)"></i></div></div><table class="legend"><tr><th>Component</th><th>Original</th><th>Share</th><th>Fixture</th><th>Share</th></tr><tr><td><i class="c0"></i>model.language_model.layers</td><td>415.3M</td><td>41.7%</td><td>74.0K</td><td>0.8%</td></tr><tr><td><i class="c1"></i>model.vision_tower.layers</td><td>411.5M</td><td>41.3%</td><td>66.9K</td><td>0.7%</td></tr><tr><td><i class="c2"></i>model.language_model</td><td>123.7M</td><td>12.4%</td><td>7.7M</td><td>85.4%</td></tr><tr><td><i class="c3"></i>model.vision_tower</td><td>45.7M</td><td>4.6%</td><td>1.2M</td><td>13.0%</td></tr></table><h2>Preserved properties</h2><table><tr><th>Config</th><th>Property</th><th>Original</th><th>Fixture</th><th></th></tr><tr class=""><td>model</td><td>model_type</td><td>hunyuan_vl</td><td>hunyuan_vl</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>tie_word_embeddings</td><td>True</td><td>True</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>attention style</td><td>GQA (16 heads, 8 KV)</td><td>GQA (4 heads, 2 KV)</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>RoPE type</td><td>dynamic</td><td>dynamic</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>model_type</td><td>hunyuan_vl_text</td><td>hunyuan_vl_text</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>vocab_size</td><td>120818</td><td>120818</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>tie_word_embeddings</td><td>True</td><td>True</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>hidden_act</td><td>silu</td><td>silu</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>max_position_embeddings</td><td>131072</td><td>131072</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>attention style</td><td>MHA (16 heads, 16 KV)</td><td>MHA (4 heads, 4 KV)</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>model_type</td><td>hunyuan_vl_vision</td><td>hunyuan_vl_vision</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>hidden_act</td><td>gelu</td><td>gelu</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>patch_size</td><td>16</td><td>16</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>num_channels</td><td>3</td><td>3</td><td>✓ kept</td></tr></table><h2>Module classes</h2><p class="note">Instances of each class in the built model (prefix HunYuanVL omitted).</p><table><tr><th>Class</th><th>Original</th><th>Fixture</th></tr><tr class=""><td>Linear</td><td>332</td><td>28</td></tr><tr class=""><td>RMSNorm</td><td>99</td><td>11</td></tr><tr class=""><td>LayerNorm</td><td>54</td><td>4</td></tr><tr class=""><td>GELUActivation</td><td>28</td><td>3</td></tr><tr class=""><td>VisionAttention</td><td>27</td><td>2</td></tr><tr class=""><td>VisionBlock</td><td>27</td><td>2</td></tr><tr class=""><td>VisionMLP</td><td>27</td><td>2</td></tr><tr class=""><td>DenseV1Attention</td><td>24</td><td>2</td></tr><tr class=""><td>DenseV1DecoderLayer</td><td>24</td><td>2</td></tr><tr class=""><td>MLP</td><td>24</td><td>2</td></tr><tr class=""><td>SiLUActivation</td><td>24</td><td>2</td></tr><tr class=""><td>Conv2d</td><td>3</td><td>3</td></tr><tr class=""><td>Embedding</td><td>2</td><td>2</td></tr><tr class=""><td>ModuleList</td><td>2</td><td>2</td></tr><tr class=""><td>ForConditionalGeneration</td><td>1</td><td>1</td></tr><tr class=""><td>Model</td><td>1</td><td>1</td></tr><tr class=""><td>RotaryEmbedding</td><td>1</td><td>1</td></tr><tr class=""><td>TextModel</td><td>1</td><td>1</td></tr><tr class=""><td>VisionPatchEmbed</td><td>1</td><td>1</td></tr><tr class=""><td>VisionPatchMerger</td><td>1</td><td>1</td></tr><tr class=""><td>VisionTransformer</td><td>1</td><td>1</td></tr></table><h2>Config changes</h2><p class="note">14 fields changed, 68 unchanged.</p><table><tr><th>Field</th><th>Original</th><th>Fixture</th></tr><tr><td>text_config.head_dim</td><td>128</td><td>16</td></tr><tr><td>text_config.hidden_size</td><td>1024</td><td>64</td></tr><tr><td>text_config.intermediate_size</td><td>3584</td><td>128</td></tr><tr><td>text_config.num_attention_heads</td><td>16</td><td>4</td></tr><tr><td>text_config.num_hidden_layers</td><td>24</td><td>2</td></tr><tr><td>text_config.num_key_value_heads</td><td>8</td><td>2</td></tr><tr><td>text_config.rope_parameters</td><td>{&quot;alpha&quot;: 1000.0, &quot;beta_fast&quot;: 32, &quot;beta_slow&quot;: 1, &quot;factor&quot;: 1.0, &quot;mscale&quot;: 1…</td><td>{&quot;alpha&quot;: 1000.0, &quot;beta_fast&quot;: 32, &quot;beta_slow&quot;: 1, &quot;factor&quot;: 1.0, &quot;mrope_sect…</td></tr><tr><td>vision_config.hidden_size</td><td>1152</td><td>64</td></tr><tr><td>vision_config.intermediate_size</td><td>4304</td><td>128</td></tr><tr><td>vision_config.num_attention_heads</td><td>16</td><td>4</td></tr><tr><td>vision_config.num_hidden_layers</td><td>27</td><td>2</td></tr><tr><td>vision_config.num_key_value_heads</td><td>16</td><td>4</td></tr><tr><td>vision_config.out_hidden_size</td><td>1024</td><td>64</td></tr><tr><td>vision_config.text_hidden_size</td><td>1024</td><td>64</td></tr></table>
+</main></body></html>

shrink/Hy-MT2-30B-A3B-mic.html

@@ -0,0 +1,34 @@
+<!doctype html>
+<html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
+<title>Shrunk architecture</title>
+<style>
+:root{--bg:#fcfcfb;--fg:#0b0b0b;--muted:#52514e;--line:#e4e3df;--ok:#1d7a43;--warn:#b35c00;--s0:#2a78d6;--s1:#eb6834;--s2:#1baf7a;--s3:#eda100;--s4:#e87ba4;--s5:#008300;--s6:#6250d6;--s7:#e34948;--s8:#8a8984;}
+@media (prefers-color-scheme:dark){:root:not([data-theme="light"]){--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}}
+:root[data-theme="dark"]{--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}
+body{margin:0;background:var(--bg);color:var(--fg);font:15px/1.5 system-ui,sans-serif}
+main{max-width:1040px;margin:0 auto;padding:24px 16px 64px}
+h1{font-size:24px;margin:0 0 4px}h2{font-size:18px;margin:36px 0 8px}h3{font-size:14px;margin:0 0 8px;font-family:ui-monospace,monospace}
+.meta,.note{color:var(--muted);margin:0 0 12px;overflow-wrap:anywhere}
+.stats{display:grid;grid-template-columns:repeat(auto-fit,minmax(150px,1fr));gap:12px;margin:20px 0}
+.stat{border:1px solid var(--line);border-radius:8px;padding:12px;display:flex;flex-direction:column}
+.stat .v{font-size:26px;font-weight:600}.stat .k{color:var(--muted);font-size:13px}
+.findings{border-radius:8px;padding:4px 16px;border:1px solid var(--line)}
+.findings.ok{border-left:4px solid var(--ok)}.findings.warn{border-left:4px solid var(--warn)}
+.findings h2{margin:12px 0 4px;font-size:16px}
+.stack{border:1px solid var(--line);border-radius:8px;padding:12px;margin:12px 0}
+.row{display:flex;align-items:center;gap:12px;margin:6px 0}
+.lab{flex:0 0 110px;color:var(--muted);font-size:13px}
+.cells{display:flex;flex-wrap:wrap;gap:2px}
+.cells i{width:12px;height:12px;border-radius:2px;display:block}
+.bar{flex:1;display:flex;gap:2px;height:16px}.bar i{display:block;border-radius:2px;min-width:2px}
+table{border-collapse:collapse;width:100%;font-size:13px;margin:8px 0;display:block;overflow-x:auto}
+th,td{text-align:left;padding:4px 10px 4px 0;border-bottom:1px solid var(--line);vertical-align:top}
+th{color:var(--muted);font-weight:500}
+td:first-child{font-family:ui-monospace,monospace;overflow-wrap:anywhere}
+td i{display:inline-block;width:10px;height:10px;border-radius:2px;margin-right:6px}
+tr.bad td{color:var(--warn)}
+.c0{background:var(--s0)}.c1{background:var(--s1)}.c2{background:var(--s2)}.c3{background:var(--s3)}.c4{background:var(--s4)}.c5{background:var(--s5)}.c6{background:var(--s6)}.c7{background:var(--s7)}.c8{background:var(--s8)}
+</style></head><body><main>
+<h1>Shrunk architecture</h1><p class="meta">huggingface-mic/Hy-MT2-30B-A3B-mic@8357a369b5aa965d0f486cbe4fe63329d0a1d00e ← tencent/Hy-MT2-30B-A3B@d3ead4dba61c09aac60a261a96ad1df3e705febb · HYV3ForCausalLM · random weights</p>
+<div class="stats"><div class="stat"><span class="k">Original</span><span class="v">30.1B</span><span class="k">parameters</span></div><div class="stat"><span class="k">Fixture</span><span class="v">15.5M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Layer stacks</span><span class="v">1</span><span class="k">compared</span></div><div class="stat"><span class="k">Findings</span><span class="v">0</span><span class="k">to review</span></div></div><section class="findings ok"><h2>✓ Same architecture, smaller</h2><p>Every module class, layer kind and checked property of the original is in the fixture.</p></section><h2>Layer stacks</h2><p class="note">One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.</p><div class="stack"><h3>model.layers</h3><div class="row"><span class="lab">Original · 48</span><div class="cells"><i class="c0" title="layer 0: no Experts"></i><i class="c1" title="layer 1: Experts, MoE, TopKRouter"></i><i class="c1" title="layer 2: Experts, MoE, TopKRouter"></i><i class="c1" title="layer 3: Experts, MoE, TopKRouter"></i><i class="c1" title="layer 4: Experts, MoE, TopKRouter"></i><i class="c1" title="layer 5: Experts, MoE, TopKRouter"></i><i class="c1" title="layer 6: Experts, MoE, TopKRouter"></i><i class="c1" title="layer 7: Experts, MoE, TopKRouter"></i><i class="c1" title="layer 8: Experts, MoE, TopKRouter"></i><i class="c1" title="layer 9: Experts, MoE, TopKRouter"></i><i class="c1" title="layer 10: Experts, MoE, TopKRouter"></i><i class="c1" title="layer 11: Experts, MoE, TopKRouter"></i><i class="c1" title="layer 12: Experts, MoE, TopKRouter"></i><i class="c1" title="layer 13: Experts, MoE, TopKRouter"></i><i class="c1" title="layer 14: Experts, MoE, TopKRouter"></i><i class="c1" title="layer 15: Experts, MoE, TopKRouter"></i><i class="c1" title="layer 16: Experts, MoE, TopKRouter"></i><i class="c1" title="layer 17: Experts, MoE, TopKRouter"></i><i class="c1" title="layer 18: Experts, MoE, TopKRouter"></i><i class="c1" title="layer 19: Experts, MoE, TopKRouter"></i><i class="c1" title="layer 20: Experts, MoE, TopKRouter"></i><i class="c1" title="layer 21: Experts, MoE, TopKRouter"></i><i class="c1" title="layer 22: Experts, MoE, TopKRouter"></i><i class="c1" title="layer 23: Experts, MoE, TopKRouter"></i><i class="c1" title="layer 24: Experts, MoE, TopKRouter"></i><i class="c1" title="layer 25: Experts, MoE, TopKRouter"></i><i class="c1" title="layer 26: Experts, MoE, TopKRouter"></i><i class="c1" title="layer 27: Experts, MoE, TopKRouter"></i><i class="c1" title="layer 28: Experts, MoE, TopKRouter"></i><i class="c1" title="layer 29: Experts, MoE, TopKRouter"></i><i class="c1" title="layer 30: Experts, MoE, TopKRouter"></i><i class="c1" title="layer 31: Experts, MoE, TopKRouter"></i><i class="c1" title="layer 32: Experts, MoE, TopKRouter"></i><i class="c1" title="layer 33: Experts, MoE, TopKRouter"></i><i class="c1" title="layer 34: Experts, MoE, TopKRouter"></i><i class="c1" title="layer 35: Experts, MoE, TopKRouter"></i><i class="c1" title="layer 36: Experts, MoE, TopKRouter"></i><i class="c1" title="layer 37: Experts, MoE, TopKRouter"></i><i class="c1" title="layer 38: Experts, MoE, TopKRouter"></i><i class="c1" title="layer 39: Experts, MoE, TopKRouter"></i><i class="c1" title="layer 40: Experts, MoE, TopKRouter"></i><i class="c1" title="layer 41: Experts, MoE, TopKRouter"></i><i class="c1" title="layer 42: Experts, MoE, TopKRouter"></i><i class="c1" title="layer 43: Experts, MoE, TopKRouter"></i><i class="c1" title="layer 44: Experts, MoE, TopKRouter"></i><i class="c1" title="layer 45: Experts, MoE, TopKRouter"></i><i class="c1" title="layer 46: Experts, MoE, TopKRouter"></i><i class="c1" title="layer 47: Experts, MoE, TopKRouter"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: no Experts"></i><i class="c1" title="layer 1: Experts, MoE, TopKRouter"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>no Experts</td><td>1</td><td>1</td><td>✓ kept</td></tr><tr><td><i class="c1"></i>Experts, MoE, TopKRouter</td><td>47</td><td>1</td><td>✓ kept</td></tr></table></div><h2>Where the parameters are</h2><p class="note">Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.</p><div class="row"><span class="lab">Original</span><div class="bar"><i class="c0" style="flex:0.983538" title="model.layers: 29.6B (98.4%)"></i><i class="c1" style="flex:0.008231" title="lm_head: 247.5M (0.8%)"></i><i class="c2" style="flex:0.008231" title="model.embed_tokens: 247.5M (0.8%)"></i><i class="c3" style="flex:0.000000" title="model.norm: 2.0K (0.0%)"></i></div></div><div class="row"><span class="lab">Fixture</span><div class="bar"><i class="c0" style="flex:0.005175" title="model.layers: 80.4K (0.5%)"></i><i class="c1" style="flex:0.497411" title="lm_head: 7.7M (49.7%)"></i><i class="c2" style="flex:0.497411" title="model.embed_tokens: 7.7M (49.7%)"></i><i class="c3" style="flex:0.000004" title="model.norm: 64 (0.0%)"></i></div></div><table class="legend"><tr><th>Component</th><th>Original</th><th>Share</th><th>Fixture</th><th>Share</th></tr><tr><td><i class="c0"></i>model.layers</td><td>29.6B</td><td>98.4%</td><td>80.4K</td><td>0.5%</td></tr><tr><td><i class="c1"></i>lm_head</td><td>247.5M</td><td>0.8%</td><td>7.7M</td><td>49.7%</td></tr><tr><td><i class="c2"></i>model.embed_tokens</td><td>247.5M</td><td>0.8%</td><td>7.7M</td><td>49.7%</td></tr><tr><td><i class="c3"></i>model.norm</td><td>2.0K</td><td>0.0%</td><td>64</td><td>0.0%</td></tr></table><h2>Preserved properties</h2><table><tr><th>Config</th><th>Property</th><th>Original</th><th>Fixture</th><th></th></tr><tr class=""><td>model</td><td>attention style</td><td>GQA (32 heads, 4 KV)</td><td>GQA (4 heads, 2 KV)</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>MoE routing</td><td>top-8 of 128 experts</td><td>top-2 of 4 experts</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>RoPE type</td><td>default</td><td>default</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>model_type</td><td>hy_v3</td><td>hy_v3</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>vocab_size</td><td>120832</td><td>120832</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>tie_word_embeddings</td><td>False</td><td>False</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>hidden_act</td><td>silu</td><td>silu</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>max_position_embeddings</td><td>262144</td><td>262144</td><td>✓ kept</td></tr></table><h2>Module classes</h2><p class="note">Instances of each class in the built model (prefix HYV3 omitted).</p><table><tr><th>Class</th><th>Original</th><th>Fixture</th></tr><tr class=""><td>Linear</td><td>337</td><td>15</td></tr><tr class=""><td>RMSNorm</td><td>193</td><td>9</td></tr><tr class=""><td>SiLUActivation</td><td>95</td><td>3</td></tr><tr class=""><td>Attention</td><td>48</td><td>2</td></tr><tr class=""><td>DecoderLayer</td><td>48</td><td>2</td></tr><tr class=""><td>MLP</td><td>48</td><td>2</td></tr><tr class=""><td>Experts</td><td>47</td><td>1</td></tr><tr class=""><td>MoE</td><td>47</td><td>1</td></tr><tr class=""><td>TopKRouter</td><td>47</td><td>1</td></tr><tr class=""><td>Embedding</td><td>1</td><td>1</td></tr><tr class=""><td>ForCausalLM</td><td>1</td><td>1</td></tr><tr class=""><td>Model</td><td>1</td><td>1</td></tr><tr class=""><td>RotaryEmbedding</td><td>1</td><td>1</td></tr><tr class=""><td>ModuleList</td><td>1</td><td>1</td></tr></table><h2>Config changes</h2><p class="note">10 fields changed, 39 unchanged.</p><table><tr><th>Field</th><th>Original</th><th>Fixture</th></tr><tr><td>head_dim</td><td>128</td><td>16</td></tr><tr><td>hidden_size</td><td>2048</td><td>64</td></tr><tr><td>intermediate_size</td><td>6912</td><td>128</td></tr><tr><td>mlp_layer_types</td><td>48 entries: 47× sparse, 1× dense</td><td>[&quot;dense&quot;, &quot;sparse&quot;]</td></tr><tr><td>moe_intermediate_size</td><td>768</td><td>32</td></tr><tr><td>num_attention_heads</td><td>32</td><td>4</td></tr><tr><td>num_experts</td><td>128</td><td>4</td></tr><tr><td>num_experts_per_tok</td><td>8</td><td>2</td></tr><tr><td>num_hidden_layers</td><td>48</td><td>2</td></tr><tr><td>num_key_value_heads</td><td>4</td><td>2</td></tr></table>
+</main></body></html>

shrink/Inkling-mic.html

@@ -0,0 +1,34 @@
+<!doctype html>
+<html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
+<title>Shrunk architecture</title>
+<style>
+:root{--bg:#fcfcfb;--fg:#0b0b0b;--muted:#52514e;--line:#e4e3df;--ok:#1d7a43;--warn:#b35c00;--s0:#2a78d6;--s1:#eb6834;--s2:#1baf7a;--s3:#eda100;--s4:#e87ba4;--s5:#008300;--s6:#6250d6;--s7:#e34948;--s8:#8a8984;}
+@media (prefers-color-scheme:dark){:root:not([data-theme="light"]){--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}}
+:root[data-theme="dark"]{--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}
+body{margin:0;background:var(--bg);color:var(--fg);font:15px/1.5 system-ui,sans-serif}
+main{max-width:1040px;margin:0 auto;padding:24px 16px 64px}
+h1{font-size:24px;margin:0 0 4px}h2{font-size:18px;margin:36px 0 8px}h3{font-size:14px;margin:0 0 8px;font-family:ui-monospace,monospace}
+.meta,.note{color:var(--muted);margin:0 0 12px;overflow-wrap:anywhere}
+.stats{display:grid;grid-template-columns:repeat(auto-fit,minmax(150px,1fr));gap:12px;margin:20px 0}
+.stat{border:1px solid var(--line);border-radius:8px;padding:12px;display:flex;flex-direction:column}
+.stat .v{font-size:26px;font-weight:600}.stat .k{color:var(--muted);font-size:13px}
+.findings{border-radius:8px;padding:4px 16px;border:1px solid var(--line)}
+.findings.ok{border-left:4px solid var(--ok)}.findings.warn{border-left:4px solid var(--warn)}
+.findings h2{margin:12px 0 4px;font-size:16px}
+.stack{border:1px solid var(--line);border-radius:8px;padding:12px;margin:12px 0}
+.row{display:flex;align-items:center;gap:12px;margin:6px 0}
+.lab{flex:0 0 110px;color:var(--muted);font-size:13px}
+.cells{display:flex;flex-wrap:wrap;gap:2px}
+.cells i{width:12px;height:12px;border-radius:2px;display:block}
+.bar{flex:1;display:flex;gap:2px;height:16px}.bar i{display:block;border-radius:2px;min-width:2px}
+table{border-collapse:collapse;width:100%;font-size:13px;margin:8px 0;display:block;overflow-x:auto}
+th,td{text-align:left;padding:4px 10px 4px 0;border-bottom:1px solid var(--line);vertical-align:top}
+th{color:var(--muted);font-weight:500}
+td:first-child{font-family:ui-monospace,monospace;overflow-wrap:anywhere}
+td i{display:inline-block;width:10px;height:10px;border-radius:2px;margin-right:6px}
+tr.bad td{color:var(--warn)}
+.c0{background:var(--s0)}.c1{background:var(--s1)}.c2{background:var(--s2)}.c3{background:var(--s3)}.c4{background:var(--s4)}.c5{background:var(--s5)}.c6{background:var(--s6)}.c7{background:var(--s7)}.c8{background:var(--s8)}
+</style></head><body><main>
+<h1>Shrunk architecture</h1><p class="meta">huggingface-mic/Inkling-mic@a7c61810959524051418bc1a49a70ac559307c0b ← thinkingmachines/Inkling@828496eeae4c243ff1a22f7f28ff83694f2f7bc9 · InklingForConditionalGeneration · random weights</p>
+<div class="stats"><div class="stat"><span class="k">Original</span><span class="v">947.1B</span><span class="k">parameters</span></div><div class="stat"><span class="k">Fixture</span><span class="v">33.8M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Layer stacks</span><span class="v">2</span><span class="k">compared</span></div><div class="stat"><span class="k">Findings</span><span class="v">0</span><span class="k">to review</span></div></div><section class="findings ok"><h2>✓ Same architecture, smaller</h2><p>Every module class, layer kind and checked property of the original is in the fixture.</p></section><h2>Layer stacks</h2><p class="note">One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.</p><div class="stack"><h3>model.language_model.layers</h3><div class="row"><span class="lab">Original · 66</span><div class="cells"><i class="c0" title="layer 0: hybrid_sliding, MLP"></i><i class="c0" title="layer 1: hybrid_sliding, MLP"></i><i class="c1" title="layer 2: hybrid_sliding, Experts, MoE …"></i><i class="c1" title="layer 3: hybrid_sliding, Experts, MoE …"></i><i class="c1" title="layer 4: hybrid_sliding, Experts, MoE …"></i><i class="c2" title="layer 5: hybrid, Experts, MoE …"></i><i class="c1" title="layer 6: hybrid_sliding, Experts, MoE …"></i><i class="c1" title="layer 7: hybrid_sliding, Experts, MoE …"></i><i class="c1" title="layer 8: hybrid_sliding, Experts, MoE …"></i><i class="c1" title="layer 9: hybrid_sliding, Experts, MoE …"></i><i class="c1" title="layer 10: hybrid_sliding, Experts, MoE …"></i><i class="c2" title="layer 11: hybrid, Experts, MoE …"></i><i class="c1" title="layer 12: hybrid_sliding, Experts, MoE …"></i><i class="c1" title="layer 13: hybrid_sliding, Experts, MoE …"></i><i class="c1" title="layer 14: hybrid_sliding, Experts, MoE …"></i><i class="c1" title="layer 15: hybrid_sliding, Experts, MoE …"></i><i class="c1" title="layer 16: hybrid_sliding, Experts, MoE …"></i><i class="c2" title="layer 17: hybrid, Experts, MoE …"></i><i class="c1" title="layer 18: hybrid_sliding, Experts, MoE …"></i><i class="c1" title="layer 19: hybrid_sliding, Experts, MoE …"></i><i class="c1" title="layer 20: hybrid_sliding, Experts, MoE …"></i><i class="c1" title="layer 21: hybrid_sliding, Experts, MoE …"></i><i class="c1" title="layer 22: hybrid_sliding, Experts, MoE …"></i><i class="c2" title="layer 23: hybrid, Experts, MoE …"></i><i class="c1" title="layer 24: hybrid_sliding, Experts, MoE …"></i><i class="c1" title="layer 25: hybrid_sliding, Experts, MoE …"></i><i class="c1" title="layer 26: hybrid_sliding, Experts, MoE …"></i><i class="c1" title="layer 27: hybrid_sliding, Experts, MoE …"></i><i class="c1" title="layer 28: hybrid_sliding, Experts, MoE …"></i><i class="c2" title="layer 29: hybrid, Experts, MoE …"></i><i class="c1" title="layer 30: hybrid_sliding, Experts, MoE …"></i><i class="c1" title="layer 31: hybrid_sliding, Experts, MoE …"></i><i class="c1" title="layer 32: hybrid_sliding, Experts, MoE …"></i><i class="c1" title="layer 33: hybrid_sliding, Experts, MoE …"></i><i class="c1" title="layer 34: hybrid_sliding, Experts, MoE …"></i><i class="c2" title="layer 35: hybrid, Experts, MoE …"></i><i class="c1" title="layer 36: hybrid_sliding, Experts, MoE …"></i><i class="c1" title="layer 37: hybrid_sliding, Experts, MoE …"></i><i class="c1" title="layer 38: hybrid_sliding, Experts, MoE …"></i><i class="c1" title="layer 39: hybrid_sliding, Experts, MoE …"></i><i class="c1" title="layer 40: hybrid_sliding, Experts, MoE …"></i><i class="c2" title="layer 41: hybrid, Experts, MoE …"></i><i class="c1" title="layer 42: hybrid_sliding, Experts, MoE …"></i><i class="c1" title="layer 43: hybrid_sliding, Experts, MoE …"></i><i class="c1" title="layer 44: hybrid_sliding, Experts, MoE …"></i><i class="c1" title="layer 45: hybrid_sliding, Experts, MoE …"></i><i class="c1" title="layer 46: hybrid_sliding, Experts, MoE …"></i><i class="c2" title="layer 47: hybrid, Experts, MoE …"></i><i class="c1" title="layer 48: hybrid_sliding, Experts, MoE …"></i><i class="c1" title="layer 49: hybrid_sliding, Experts, MoE …"></i><i class="c1" title="layer 50: hybrid_sliding, Experts, MoE …"></i><i class="c1" title="layer 51: hybrid_sliding, Experts, MoE …"></i><i class="c1" title="layer 52: hybrid_sliding, Experts, MoE …"></i><i class="c2" title="layer 53: hybrid, Experts, MoE …"></i><i class="c1" title="layer 54: hybrid_sliding, Experts, MoE …"></i><i class="c1" title="layer 55: hybrid_sliding, Experts, MoE …"></i><i class="c1" title="layer 56: hybrid_sliding, Experts, MoE …"></i><i class="c1" title="layer 57: hybrid_sliding, Experts, MoE …"></i><i class="c1" title="layer 58: hybrid_sliding, Experts, MoE …"></i><i class="c2" title="layer 59: hybrid, Experts, MoE …"></i><i class="c1" title="layer 60: hybrid_sliding, Experts, MoE …"></i><i class="c1" title="layer 61: hybrid_sliding, Experts, MoE …"></i><i class="c1" title="layer 62: hybrid_sliding, Experts, MoE …"></i><i class="c1" title="layer 63: hybrid_sliding, Experts, MoE …"></i><i class="c1" title="layer 64: hybrid_sliding, Experts, MoE …"></i><i class="c2" title="layer 65: hybrid, Experts, MoE …"></i></div></div><div class="row"><span class="lab">Fixture · 6</span><div class="cells"><i class="c0" title="layer 0: hybrid_sliding, MLP"></i><i class="c0" title="layer 1: hybrid_sliding, MLP"></i><i class="c1" title="layer 2: hybrid_sliding, Experts, MoE …"></i><i class="c1" title="layer 3: hybrid_sliding, Experts, MoE …"></i><i class="c1" title="layer 4: hybrid_sliding, Experts, MoE …"></i><i class="c2" title="layer 5: hybrid, Experts, MoE …"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>hybrid_sliding, MLP</td><td>2</td><td>2</td><td>✓ kept</td></tr><tr><td><i class="c1"></i>hybrid_sliding, Experts, MoE …</td><td>53</td><td>3</td><td>✓ kept</td></tr><tr><td><i class="c2"></i>hybrid, Experts, MoE …</td><td>11</td><td>1</td><td>✓ kept</td></tr></table></div><div class="stack"><h3>model.vision_tower.encoder_layers</h3><div class="row"><span class="lab">Original · 4</span><div class="cells"><i class="c0" title="layer 0: RMSNorm"></i><i class="c0" title="layer 1: RMSNorm"></i><i class="c0" title="layer 2: RMSNorm"></i><i class="c1" title="layer 3: no RMSNorm"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: RMSNorm"></i><i class="c1" title="layer 1: no RMSNorm"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>RMSNorm</td><td>3</td><td>1</td><td>✓ kept</td></tr><tr><td><i class="c1"></i>no RMSNorm</td><td>1</td><td>1</td><td>✓ kept</td></tr></table></div><h2>Where the parameters are</h2><p class="note">Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.</p><div class="row"><span class="lab">Original</span><div class="bar"><i class="c0" style="flex:0.997295" title="model.language_model.layers: 944.6B (99.7%)"></i><i class="c1" style="flex:0.001304" title="model.language_model: 1.2B (0.1%)"></i><i class="c2" style="flex:0.001304" title="lm_head: 1.2B (0.1%)"></i><i class="c3" style="flex:0.000088" title="model.vision_tower.encoder_layers: 83.7M (0.0%)"></i><i class="c4" style="flex:0.000008" title="model.audio_tower: 7.9M (0.0%)"></i><i class="c5" style="flex:0.000000" title="model.vision_tower: 6.1K (0.0%)"></i></div></div><div class="row"><span class="lab">Fixture</span><div class="bar"><i class="c0" style="flex:0.214181" title="model.language_model.layers: 7.2M (21.4%)"></i><i class="c1" style="flex:0.380580" title="model.language_model: 12.9M (38.1%)"></i><i class="c2" style="flex:0.380576" title="lm_head: 12.9M (38.1%)"></i><i class="c3" style="flex:0.022235" title="model.vision_tower.encoder_layers: 751.7K (2.2%)"></i><i class="c4" style="flex:0.002425" title="model.audio_tower: 82.0K (0.2%)"></i><i class="c5" style="flex:0.000002" title="model.vision_tower: 64 (0.0%)"></i></div></div><table class="legend"><tr><th>Component</th><th>Original</th><th>Share</th><th>Fixture</th><th>Share</th></tr><tr><td><i class="c0"></i>model.language_model.layers</td><td>944.6B</td><td>99.7%</td><td>7.2M</td><td>21.4%</td></tr><tr><td><i class="c1"></i>model.language_model</td><td>1.2B</td><td>0.1%</td><td>12.9M</td><td>38.1%</td></tr><tr><td><i class="c2"></i>lm_head</td><td>1.2B</td><td>0.1%</td><td>12.9M</td><td>38.1%</td></tr><tr><td><i class="c3"></i>model.vision_tower.encoder_layers</td><td>83.7M</td><td>0.0%</td><td>751.7K</td><td>2.2%</td></tr><tr><td><i class="c4"></i>model.audio_tower</td><td>7.9M</td><td>0.0%</td><td>82.0K</td><td>0.2%</td></tr><tr><td><i class="c5"></i>model.vision_tower</td><td>6.1K</td><td>0.0%</td><td>64</td><td>0.0%</td></tr></table><h2>Preserved properties</h2><table><tr><th>Config</th><th>Property</th><th>Original</th><th>Fixture</th><th></th></tr><tr class=""><td>model</td><td>model_type</td><td>inkling_mm_model</td><td>inkling_mm_model</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>attention style</td><td>GQA (64 heads, 8 KV)</td><td>GQA (4 heads, 2 KV)</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>MoE routing</td><td>top-6 of 256 experts</td><td>top-2 of 4 experts</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>model_type</td><td>inkling_text</td><td>inkling_text</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>vocab_size</td><td>201024</td><td>201024</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>hidden_act</td><td>silu</td><td>silu</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>max_position_embeddings</td><td>1048576</td><td>1048576</td><td>✓ kept</td></tr><tr class=""><td>audio_config</td><td>model_type</td><td>inkling_audio</td><td>inkling_audio</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>model_type</td><td>inkling_vision</td><td>inkling_vision</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>patch_size</td><td>40</td><td>40</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>num_channels</td><td>3</td><td>3</td><td>✓ kept</td></tr></table><h2>Module classes</h2><p class="note">Instances of each class in the built model (prefix Inkling omitted).</p><table><tr><th>Class</th><th>Original</th><th>Fixture</th></tr><tr class=""><td>Linear</td><td>341</td><td>39</td></tr><tr class=""><td>RMSNorm</td><td>271</td><td>29</td></tr><tr class=""><td>Conv1d</td><td>264</td><td>24</td></tr><tr class=""><td>ShortConvolution</td><td>264</td><td>24</td></tr><tr class=""><td>SiLUActivation</td><td>130</td><td>10</td></tr><tr class=""><td>Attention</td><td>66</td><td>6</td></tr><tr class=""><td>DecoderLayer</td><td>66</td><td>6</td></tr><tr class=""><td>RelativeLogits</td><td>66</td><td>6</td></tr><tr class=""><td>Experts</td><td>64</td><td>4</td></tr><tr class=""><td>MoE</td><td>64</td><td>4</td></tr><tr class=""><td>SharedExperts</td><td>64</td><td>4</td></tr><tr class=""><td>TopkRouter</td><td>64</td><td>4</td></tr><tr class=""><td>VisionEncoderLayer</td><td>4</td><td>2</td></tr><tr class=""><td>MLP</td><td>2</td><td>2</td></tr><tr class=""><td>ModuleList</td><td>2</td><td>2</td></tr><tr class=""><td>Embedding</td><td>1</td><td>1</td></tr><tr class=""><td>AudioModel</td><td>1</td><td>1</td></tr><tr class=""><td>AudioModelEmbeddings</td><td>1</td><td>1</td></tr><tr class=""><td>ForConditionalGeneration</td><td>1</td><td>1</td></tr><tr class=""><td>Model</td><td>1</td><td>1</td></tr><tr class=""><td>NormedEmbedding</td><td>1</td><td>1</td></tr><tr class=""><td>TextModel</td><td>1</td><td>1</td></tr><tr class=""><td>VisionModel</td><td>1</td><td>1</td></tr></table><h2>Config changes</h2><p class="note">16 fields changed, 102 unchanged.</p><table><tr><th>Field</th><th>Original</th><th>Fixture</th></tr><tr><td>audio_config.text_hidden_size</td><td>6144</td><td>64</td></tr><tr><td>text_config.head_dim</td><td>128</td><td>16</td></tr><tr><td>text_config.hidden_size</td><td>6144</td><td>64</td></tr><tr><td>text_config.intermediate_size</td><td>24576</td><td>128</td></tr><tr><td>text_config.layer_types</td><td>66 entries: 55× hybrid_sliding, 11× hybrid</td><td>[&quot;hybrid_sliding&quot;, &quot;hybrid_sliding&quot;, &quot;hybrid_sliding&quot;, &quot;hybrid_sliding&quot;, &quot;hyb…</td></tr><tr><td>text_config.mlp_layer_types</td><td>66 entries: 64× sparse, 2× dense</td><td>[&quot;dense&quot;, &quot;dense&quot;, &quot;sparse&quot;, &quot;sparse&quot;, &quot;sparse&quot;, &quot;sparse&quot;]</td></tr><tr><td>text_config.moe_intermediate_size</td><td>3072</td><td>32</td></tr><tr><td>text_config.n_routed_experts</td><td>256</td><td>4</td></tr><tr><td>text_config.num_attention_heads</td><td>64</td><td>4</td></tr><tr><td>text_config.num_experts_per_tok</td><td>6</td><td>2</td></tr><tr><td>text_config.num_hidden_layers</td><td>66</td><td>6</td></tr><tr><td>text_config.num_key_value_heads</td><td>8</td><td>2</td></tr><tr><td>vision_config.hidden_size</td><td>1024</td><td>64</td></tr><tr><td>vision_config.n_layers</td><td>4</td><td>2</td></tr><tr><td>vision_config.num_attention_heads</td><td>16</td><td>4</td></tr><tr><td>vision_config.text_hidden_size</td><td>6144</td><td>64</td></tr></table>
+</main></body></html>

shrink/Jamba-tiny-dev-mic.html

@@ -0,0 +1,34 @@
+<!doctype html>
+<html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
+<title>Shrunk architecture</title>
+<style>
+:root{--bg:#fcfcfb;--fg:#0b0b0b;--muted:#52514e;--line:#e4e3df;--ok:#1d7a43;--warn:#b35c00;--s0:#2a78d6;--s1:#eb6834;--s2:#1baf7a;--s3:#eda100;--s4:#e87ba4;--s5:#008300;--s6:#6250d6;--s7:#e34948;--s8:#8a8984;}
+@media (prefers-color-scheme:dark){:root:not([data-theme="light"]){--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}}
+:root[data-theme="dark"]{--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}
+body{margin:0;background:var(--bg);color:var(--fg);font:15px/1.5 system-ui,sans-serif}
+main{max-width:1040px;margin:0 auto;padding:24px 16px 64px}
+h1{font-size:24px;margin:0 0 4px}h2{font-size:18px;margin:36px 0 8px}h3{font-size:14px;margin:0 0 8px;font-family:ui-monospace,monospace}
+.meta,.note{color:var(--muted);margin:0 0 12px;overflow-wrap:anywhere}
+.stats{display:grid;grid-template-columns:repeat(auto-fit,minmax(150px,1fr));gap:12px;margin:20px 0}
+.stat{border:1px solid var(--line);border-radius:8px;padding:12px;display:flex;flex-direction:column}
+.stat .v{font-size:26px;font-weight:600}.stat .k{color:var(--muted);font-size:13px}
+.findings{border-radius:8px;padding:4px 16px;border:1px solid var(--line)}
+.findings.ok{border-left:4px solid var(--ok)}.findings.warn{border-left:4px solid var(--warn)}
+.findings h2{margin:12px 0 4px;font-size:16px}
+.stack{border:1px solid var(--line);border-radius:8px;padding:12px;margin:12px 0}
+.row{display:flex;align-items:center;gap:12px;margin:6px 0}
+.lab{flex:0 0 110px;color:var(--muted);font-size:13px}
+.cells{display:flex;flex-wrap:wrap;gap:2px}
+.cells i{width:12px;height:12px;border-radius:2px;display:block}
+.bar{flex:1;display:flex;gap:2px;height:16px}.bar i{display:block;border-radius:2px;min-width:2px}
+table{border-collapse:collapse;width:100%;font-size:13px;margin:8px 0;display:block;overflow-x:auto}
+th,td{text-align:left;padding:4px 10px 4px 0;border-bottom:1px solid var(--line);vertical-align:top}
+th{color:var(--muted);font-weight:500}
+td:first-child{font-family:ui-monospace,monospace;overflow-wrap:anywhere}
+td i{display:inline-block;width:10px;height:10px;border-radius:2px;margin-right:6px}
+tr.bad td{color:var(--warn)}
+.c0{background:var(--s0)}.c1{background:var(--s1)}.c2{background:var(--s2)}.c3{background:var(--s3)}.c4{background:var(--s4)}.c5{background:var(--s5)}.c6{background:var(--s6)}.c7{background:var(--s7)}.c8{background:var(--s8)}
+</style></head><body><main>
+<h1>Shrunk architecture</h1><p class="meta">huggingface-mic/Jamba-tiny-dev-mic@d03ef8cc4d1f81b8af2ae68030ecaa9bc1d51a6d ← ai21labs/Jamba-tiny-dev@ed303361004ac875426a61675edecf8e9d976882 · JambaForCausalLM · random weights</p>
+<div class="stats"><div class="stat"><span class="k">Original</span><span class="v">318.7M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Fixture</span><span class="v">8.8M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Layer stacks</span><span class="v">1</span><span class="k">compared</span></div><div class="stat"><span class="k">Findings</span><span class="v">0</span><span class="k">to review</span></div></div><section class="findings ok"><h2>✓ Same architecture, smaller</h2><p>Every module class, layer kind and checked property of the original is in the fixture.</p></section><h2>Layer stacks</h2><p class="note">One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.</p><div class="stack"><h3>model.layers</h3><div class="row"><span class="lab">Original · 16</span><div class="cells"><i class="c0" title="layer 0: Conv1d, MLP, MambaDecoderLayer …"></i><i class="c1" title="layer 1: Conv1d, Experts, MambaDecoderLayer …"></i><i class="c0" title="layer 2: Conv1d, MLP, MambaDecoderLayer …"></i><i class="c1" title="layer 3: Conv1d, Experts, MambaDecoderLayer …"></i><i class="c2" title="layer 4: is_causal=True, Attention, AttentionDecoderLayer …"></i><i class="c1" title="layer 5: Conv1d, Experts, MambaDecoderLayer …"></i><i class="c0" title="layer 6: Conv1d, MLP, MambaDecoderLayer …"></i><i class="c1" title="layer 7: Conv1d, Experts, MambaDecoderLayer …"></i><i class="c0" title="layer 8: Conv1d, MLP, MambaDecoderLayer …"></i><i class="c1" title="layer 9: Conv1d, Experts, MambaDecoderLayer …"></i><i class="c0" title="layer 10: Conv1d, MLP, MambaDecoderLayer …"></i><i class="c1" title="layer 11: Conv1d, Experts, MambaDecoderLayer …"></i><i class="c2" title="layer 12: is_causal=True, Attention, AttentionDecoderLayer …"></i><i class="c1" title="layer 13: Conv1d, Experts, MambaDecoderLayer …"></i><i class="c0" title="layer 14: Conv1d, MLP, MambaDecoderLayer …"></i><i class="c1" title="layer 15: Conv1d, Experts, MambaDecoderLayer …"></i></div></div><div class="row"><span class="lab">Fixture · 5</span><div class="cells"><i class="c0" title="layer 0: Conv1d, MLP, MambaDecoderLayer …"></i><i class="c1" title="layer 1: Conv1d, Experts, MambaDecoderLayer …"></i><i class="c0" title="layer 2: Conv1d, MLP, MambaDecoderLayer …"></i><i class="c1" title="layer 3: Conv1d, Experts, MambaDecoderLayer …"></i><i class="c2" title="layer 4: is_causal=True, Attention, AttentionDecoderLayer …"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>Conv1d, MLP, MambaDecoderLayer …</td><td>6</td><td>2</td><td>✓ kept</td></tr><tr><td><i class="c1"></i>Conv1d, Experts, MambaDecoderLayer …</td><td>8</td><td>2</td><td>✓ kept</td></tr><tr><td><i class="c2"></i>is_causal=True, Attention, AttentionDecoderLayer …</td><td>2</td><td>1</td><td>✓ kept</td></tr></table></div><h2>Where the parameters are</h2><p class="note">Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.</p><div class="row"><span class="lab">Original</span><div class="bar"><i class="c0" style="flex:0.789420" title="model.layers: 251.6M (78.9%)"></i><i class="c1" style="flex:0.105289" title="lm_head: 33.6M (10.5%)"></i><i class="c2" style="flex:0.105289" title="model.embed_tokens: 33.6M (10.5%)"></i><i class="c3" style="flex:0.000002" title="model.final_layernorm: 512 (0.0%)"></i></div></div><div class="row"><span class="lab">Fixture</span><div class="bar"><i class="c0" style="flex:0.050189" title="model.layers: 443.3K (5.0%)"></i><i class="c1" style="flex:0.474902" title="lm_head: 4.2M (47.5%)"></i><i class="c2" style="flex:0.474902" title="model.embed_tokens: 4.2M (47.5%)"></i><i class="c3" style="flex:0.000007" title="model.final_layernorm: 64 (0.0%)"></i></div></div><table class="legend"><tr><th>Component</th><th>Original</th><th>Share</th><th>Fixture</th><th>Share</th></tr><tr><td><i class="c0"></i>model.layers</td><td>251.6M</td><td>78.9%</td><td>443.3K</td><td>5.0%</td></tr><tr><td><i class="c1"></i>lm_head</td><td>33.6M</td><td>10.5%</td><td>4.2M</td><td>47.5%</td></tr><tr><td><i class="c2"></i>model.embed_tokens</td><td>33.6M</td><td>10.5%</td><td>4.2M</td><td>47.5%</td></tr><tr><td><i class="c3"></i>model.final_layernorm</td><td>512</td><td>0.0%</td><td>64</td><td>0.0%</td></tr></table><h2>Preserved properties</h2><table><tr><th>Config</th><th>Property</th><th>Original</th><th>Fixture</th><th></th></tr><tr class=""><td>model</td><td>attention style</td><td>GQA (8 heads, 2 KV)</td><td>GQA (4 heads, 2 KV)</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>MoE routing</td><td>top-2 of 8 experts</td><td>top-2 of 4 experts</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>model_type</td><td>jamba</td><td>jamba</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>vocab_size</td><td>65536</td><td>65536</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>tie_word_embeddings</td><td>False</td><td>False</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>hidden_act</td><td>silu</td><td>silu</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>max_position_embeddings</td><td>262144</td><td>262144</td><td>✓ kept</td></tr></table><h2>Module classes</h2><p class="note">Instances of each class in the built model (prefix Jamba omitted).</p><table><tr><th>Class</th><th>Original</th><th>Fixture</th></tr><tr class=""><td>Linear</td><td>89</td><td>30</td></tr><tr class=""><td>RMSNorm</td><td>75</td><td>23</td></tr><tr class=""><td>SiLUActivation</td><td>30</td><td>9</td></tr><tr class=""><td>Conv1d</td><td>14</td><td>4</td></tr><tr class=""><td>MambaDecoderLayer</td><td>14</td><td>4</td></tr><tr class=""><td>MambaMixer</td><td>14</td><td>4</td></tr><tr class=""><td>Experts</td><td>8</td><td>2</td></tr><tr class=""><td>MLP</td><td>8</td><td>3</td></tr><tr class=""><td>SparseMoeBlock</td><td>8</td><td>2</td></tr><tr class=""><td>TopKRouter</td><td>8</td><td>2</td></tr><tr class=""><td>Attention</td><td>2</td><td>1</td></tr><tr class=""><td>AttentionDecoderLayer</td><td>2</td><td>1</td></tr><tr class=""><td>Embedding</td><td>1</td><td>1</td></tr><tr class=""><td>ForCausalLM</td><td>1</td><td>1</td></tr><tr class=""><td>Model</td><td>1</td><td>1</td></tr><tr class=""><td>ModuleList</td><td>1</td><td>1</td></tr></table><h2>Config changes</h2><p class="note">5 fields changed, 39 unchanged.</p><table><tr><th>Field</th><th>Original</th><th>Fixture</th></tr><tr><td>hidden_size</td><td>512</td><td>64</td></tr><tr><td>intermediate_size</td><td>2048</td><td>128</td></tr><tr><td>num_attention_heads</td><td>8</td><td>4</td></tr><tr><td>num_experts</td><td>8</td><td>4</td></tr><tr><td>num_hidden_layers</td><td>16</td><td>5</td></tr></table>
+</main></body></html>

shrink/Janus-Pro-1B-mic.html

@@ -0,0 +1,34 @@
+<!doctype html>
+<html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
+<title>Shrunk architecture</title>
+<style>
+:root{--bg:#fcfcfb;--fg:#0b0b0b;--muted:#52514e;--line:#e4e3df;--ok:#1d7a43;--warn:#b35c00;--s0:#2a78d6;--s1:#eb6834;--s2:#1baf7a;--s3:#eda100;--s4:#e87ba4;--s5:#008300;--s6:#6250d6;--s7:#e34948;--s8:#8a8984;}
+@media (prefers-color-scheme:dark){:root:not([data-theme="light"]){--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}}
+:root[data-theme="dark"]{--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}
+body{margin:0;background:var(--bg);color:var(--fg);font:15px/1.5 system-ui,sans-serif}
+main{max-width:1040px;margin:0 auto;padding:24px 16px 64px}
+h1{font-size:24px;margin:0 0 4px}h2{font-size:18px;margin:36px 0 8px}h3{font-size:14px;margin:0 0 8px;font-family:ui-monospace,monospace}
+.meta,.note{color:var(--muted);margin:0 0 12px;overflow-wrap:anywhere}
+.stats{display:grid;grid-template-columns:repeat(auto-fit,minmax(150px,1fr));gap:12px;margin:20px 0}
+.stat{border:1px solid var(--line);border-radius:8px;padding:12px;display:flex;flex-direction:column}
+.stat .v{font-size:26px;font-weight:600}.stat .k{color:var(--muted);font-size:13px}
+.findings{border-radius:8px;padding:4px 16px;border:1px solid var(--line)}
+.findings.ok{border-left:4px solid var(--ok)}.findings.warn{border-left:4px solid var(--warn)}
+.findings h2{margin:12px 0 4px;font-size:16px}
+.stack{border:1px solid var(--line);border-radius:8px;padding:12px;margin:12px 0}
+.row{display:flex;align-items:center;gap:12px;margin:6px 0}
+.lab{flex:0 0 110px;color:var(--muted);font-size:13px}
+.cells{display:flex;flex-wrap:wrap;gap:2px}
+.cells i{width:12px;height:12px;border-radius:2px;display:block}
+.bar{flex:1;display:flex;gap:2px;height:16px}.bar i{display:block;border-radius:2px;min-width:2px}
+table{border-collapse:collapse;width:100%;font-size:13px;margin:8px 0;display:block;overflow-x:auto}
+th,td{text-align:left;padding:4px 10px 4px 0;border-bottom:1px solid var(--line);vertical-align:top}
+th{color:var(--muted);font-weight:500}
+td:first-child{font-family:ui-monospace,monospace;overflow-wrap:anywhere}
+td i{display:inline-block;width:10px;height:10px;border-radius:2px;margin-right:6px}
+tr.bad td{color:var(--warn)}
+.c0{background:var(--s0)}.c1{background:var(--s1)}.c2{background:var(--s2)}.c3{background:var(--s3)}.c4{background:var(--s4)}.c5{background:var(--s5)}.c6{background:var(--s6)}.c7{background:var(--s7)}.c8{background:var(--s8)}
+</style></head><body><main>
+<h1>Shrunk architecture</h1><p class="meta">huggingface-mic/Janus-Pro-1B-mic@5d0bfc05499ea83be331a434ea79583ca970581a ← deepseek-community/Janus-Pro-1B@1655280bb75959cc1cb85529a2a8b26e7016072e · JanusForConditionalGeneration · random weights · built before the shrink rule (no recorded recipe)</p>
+<div class="stats"><div class="stat"><span class="k">Original</span><span class="v">1.9B</span><span class="k">parameters</span></div><div class="stat"><span class="k">Fixture</span><span class="v">81.9M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Layer stacks</span><span class="v">4</span><span class="k">compared</span></div><div class="stat"><span class="k">Findings</span><span class="v">0</span><span class="k">to review</span></div></div><section class="findings ok"><h2>✓ Same architecture, smaller</h2><p>Every module class, layer kind and checked property of the original is in the fixture.</p></section><h2>Layer stacks</h2><p class="note">One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.</p><div class="stack"><h3>model.vision_model.encoder.layers</h3><div class="row"><span class="lab">Original · 24</span><div class="cells"><i class="c0" title="layer 0: VisionEncoderLayer"></i><i class="c0" title="layer 1: VisionEncoderLayer"></i><i class="c0" title="layer 2: VisionEncoderLayer"></i><i class="c0" title="layer 3: VisionEncoderLayer"></i><i class="c0" title="layer 4: VisionEncoderLayer"></i><i class="c0" title="layer 5: VisionEncoderLayer"></i><i class="c0" title="layer 6: VisionEncoderLayer"></i><i class="c0" title="layer 7: VisionEncoderLayer"></i><i class="c0" title="layer 8: VisionEncoderLayer"></i><i class="c0" title="layer 9: VisionEncoderLayer"></i><i class="c0" title="layer 10: VisionEncoderLayer"></i><i class="c0" title="layer 11: VisionEncoderLayer"></i><i class="c0" title="layer 12: VisionEncoderLayer"></i><i class="c0" title="layer 13: VisionEncoderLayer"></i><i class="c0" title="layer 14: VisionEncoderLayer"></i><i class="c0" title="layer 15: VisionEncoderLayer"></i><i class="c0" title="layer 16: VisionEncoderLayer"></i><i class="c0" title="layer 17: VisionEncoderLayer"></i><i class="c0" title="layer 18: VisionEncoderLayer"></i><i class="c0" title="layer 19: VisionEncoderLayer"></i><i class="c0" title="layer 20: VisionEncoderLayer"></i><i class="c0" title="layer 21: VisionEncoderLayer"></i><i class="c0" title="layer 22: VisionEncoderLayer"></i><i class="c0" title="layer 23: VisionEncoderLayer"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: VisionEncoderLayer"></i><i class="c0" title="layer 1: VisionEncoderLayer"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>VisionEncoderLayer</td><td>24</td><td>2</td><td>✓ kept</td></tr></table></div><div class="stack"><h3>model.vqmodel.encoder.down</h3><div class="row"><span class="lab">Original · 5</span><div class="cells"><i class="c0" title="layer 0: VQVAEConvDownsample"></i><i class="c0" title="layer 1: VQVAEConvDownsample"></i><i class="c0" title="layer 2: VQVAEConvDownsample"></i><i class="c0" title="layer 3: VQVAEConvDownsample"></i><i class="c1" title="layer 4: VQVAEAttnBlock"></i></div></div><div class="row"><span class="lab">Fixture · 5</span><div class="cells"><i class="c0" title="layer 0: VQVAEConvDownsample"></i><i class="c0" title="layer 1: VQVAEConvDownsample"></i><i class="c0" title="layer 2: VQVAEConvDownsample"></i><i class="c0" title="layer 3: VQVAEConvDownsample"></i><i class="c1" title="layer 4: VQVAEAttnBlock"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>VQVAEConvDownsample</td><td>4</td><td>4</td><td>✓ kept</td></tr><tr><td><i class="c1"></i>VQVAEAttnBlock</td><td>1</td><td>1</td><td>✓ kept</td></tr></table></div><div class="stack"><h3>model.vqmodel.decoder.up</h3><div class="row"><span class="lab">Original · 5</span><div class="cells"><i class="c0" title="layer 0: VQVAEAttnBlock, VQVAEConvUpsample"></i><i class="c1" title="layer 1: VQVAEConvUpsample"></i><i class="c1" title="layer 2: VQVAEConvUpsample"></i><i class="c1" title="layer 3: VQVAEConvUpsample"></i><i class="c2" title="layer 4: no VQVAEAttnBlock"></i></div></div><div class="row"><span class="lab">Fixture · 5</span><div class="cells"><i class="c0" title="layer 0: VQVAEAttnBlock, VQVAEConvUpsample"></i><i class="c1" title="layer 1: VQVAEConvUpsample"></i><i class="c1" title="layer 2: VQVAEConvUpsample"></i><i class="c1" title="layer 3: VQVAEConvUpsample"></i><i class="c2" title="layer 4: no VQVAEAttnBlock"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>VQVAEAttnBlock, VQVAEConvUpsample</td><td>1</td><td>1</td><td>✓ kept</td></tr><tr><td><i class="c1"></i>VQVAEConvUpsample</td><td>3</td><td>3</td><td>✓ kept</td></tr><tr><td><i class="c2"></i>no VQVAEAttnBlock</td><td>1</td><td>1</td><td>✓ kept</td></tr></table></div><div class="stack"><h3>model.language_model.layers</h3><div class="row"><span class="lab">Original · 24</span><div class="cells"><i class="c0" title="layer 0: LlamaDecoderLayer"></i><i class="c0" title="layer 1: LlamaDecoderLayer"></i><i class="c0" title="layer 2: LlamaDecoderLayer"></i><i class="c0" title="layer 3: LlamaDecoderLayer"></i><i class="c0" title="layer 4: LlamaDecoderLayer"></i><i class="c0" title="layer 5: LlamaDecoderLayer"></i><i class="c0" title="layer 6: LlamaDecoderLayer"></i><i class="c0" title="layer 7: LlamaDecoderLayer"></i><i class="c0" title="layer 8: LlamaDecoderLayer"></i><i class="c0" title="layer 9: LlamaDecoderLayer"></i><i class="c0" title="layer 10: LlamaDecoderLayer"></i><i class="c0" title="layer 11: LlamaDecoderLayer"></i><i class="c0" title="layer 12: LlamaDecoderLayer"></i><i class="c0" title="layer 13: LlamaDecoderLayer"></i><i class="c0" title="layer 14: LlamaDecoderLayer"></i><i class="c0" title="layer 15: LlamaDecoderLayer"></i><i class="c0" title="layer 16: LlamaDecoderLayer"></i><i class="c0" title="layer 17: LlamaDecoderLayer"></i><i class="c0" title="layer 18: LlamaDecoderLayer"></i><i class="c0" title="layer 19: LlamaDecoderLayer"></i><i class="c0" title="layer 20: LlamaDecoderLayer"></i><i class="c0" title="layer 21: LlamaDecoderLayer"></i><i class="c0" title="layer 22: LlamaDecoderLayer"></i><i class="c0" title="layer 23: LlamaDecoderLayer"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: LlamaDecoderLayer"></i><i class="c0" title="layer 1: LlamaDecoderLayer"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>LlamaDecoderLayer</td><td>24</td><td>2</td><td>✓ kept</td></tr></table></div><h2>Where the parameters are</h2><p class="note">Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.</p><div class="row"><span class="lab">Original</span><div class="bar"><i class="c0" style="flex:0.660565" title="model.language_model.layers: 1.2B (66.1%)"></i><i class="c1" style="flex:0.161929" title="model.vision_model.encoder.layers: 302.3M (16.2%)"></i><i class="c2" style="flex:0.112333" title="model.language_model: 209.7M (11.2%)"></i><i class="c3" style="flex:0.020230" title="model.generation_head: 37.8M (2.0%)"></i><i class="c4" style="flex:0.016482" title="model.vqmodel.decoder.up: 30.8M (1.6%)"></i><i class="c5" style="flex:0.012584" title="model.vqmodel: 23.5M (1.3%)"></i><i class="c6" style="flex:0.009437" title="model.vqmodel.encoder.down: 17.6M (0.9%)"></i><i class="c7" style="flex:0.006439" title="4 others: 12.0M (0.6%)"></i></div></div><div class="row"><span class="lab">Fixture</span><div class="bar"><i class="c0" style="flex:0.001003" title="model.language_model.layers: 82.2K (0.1%)"></i><i class="c1" style="flex:0.001220" title="model.vision_model.encoder.layers: 100.0K (0.1%)"></i><i class="c2" style="flex:0.080008" title="model.language_model: 6.6M (8.0%)"></i><i class="c3" style="flex:0.014602" title="model.generation_head: 1.2M (1.5%)"></i><i class="c4" style="flex:0.375654" title="model.vqmodel.decoder.up: 30.8M (37.6%)"></i><i class="c5" style="flex:0.298366" title="model.vqmodel: 24.4M (29.8%)"></i><i class="c6" style="flex:0.215090" title="model.vqmodel.encoder.down: 17.6M (21.5%)"></i><i class="c7" style="flex:0.014057" title="4 others: 1.2M (1.4%)"></i></div></div><table class="legend"><tr><th>Component</th><th>Original</th><th>Share</th><th>Fixture</th><th>Share</th></tr><tr><td><i class="c0"></i>model.language_model.layers</td><td>1.2B</td><td>66.1%</td><td>82.2K</td><td>0.1%</td></tr><tr><td><i class="c1"></i>model.vision_model.encoder.layers</td><td>302.3M</td><td>16.2%</td><td>100.0K</td><td>0.1%</td></tr><tr><td><i class="c2"></i>model.language_model</td><td>209.7M</td><td>11.2%</td><td>6.6M</td><td>8.0%</td></tr><tr><td><i class="c3"></i>model.generation_head</td><td>37.8M</td><td>2.0%</td><td>1.2M</td><td>1.5%</td></tr><tr><td><i class="c4"></i>model.vqmodel.decoder.up</td><td>30.8M</td><td>1.6%</td><td>30.8M</td><td>37.6%</td></tr><tr><td><i class="c5"></i>model.vqmodel</td><td>23.5M</td><td>1.3%</td><td>24.4M</td><td>29.8%</td></tr><tr><td><i class="c6"></i>model.vqmodel.encoder.down</td><td>17.6M</td><td>0.9%</td><td>17.6M</td><td>21.5%</td></tr><tr><td><i class="c7"></i>4 others</td><td>12.0M</td><td>0.6%</td><td>1.2M</td><td>1.4%</td></tr></table><h2>Preserved properties</h2><table><tr><th>Config</th><th>Property</th><th>Original</th><th>Fixture</th><th></th></tr><tr class=""><td>model</td><td>model_type</td><td>janus</td><td>janus</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>tie_word_embeddings</td><td>True</td><td>True</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>attention style</td><td>MHA (16 heads, 16 KV)</td><td>MHA (4 heads, 4 KV)</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>RoPE type</td><td>default</td><td>default</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>model_type</td><td>llama</td><td>llama</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>vocab_size</td><td>102400</td><td>102400</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>tie_word_embeddings</td><td>False</td><td>False</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>hidden_act</td><td>silu</td><td>silu</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>max_position_embeddings</td><td>16384</td><td>16384</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>model_type</td><td>janus_vision_model</td><td>janus_vision_model</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>hidden_act</td><td>gelu</td><td>gelu</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>image_size</td><td>384</td><td>384</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>patch_size</td><td>16</td><td>16</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>num_channels</td><td>3</td><td>3</td><td>✓ kept</td></tr><tr class=""><td>vq_config</td><td>model_type</td><td>janus_vqgan</td><td>janus_vqgan</td><td>✓ kept</td></tr><tr class=""><td>vq_config</td><td>hidden_act</td><td>gelu</td><td>gelu</td><td>✓ kept</td></tr></table><h2>Module classes</h2><p class="note">Instances of each class in the built model (prefix Janus omitted).</p><table><tr><th>Class</th><th>Original</th><th>Fixture</th></tr><tr class=""><td>Linear</td><td>319</td><td>33</td></tr><tr class=""><td>Conv2d</td><td>105</td><td>105</td></tr><tr class=""><td>Dropout</td><td>77</td><td>33</td></tr><tr class=""><td>Identity</td><td>72</td><td>6</td></tr><tr class=""><td>GroupNorm</td><td>67</td><td>67</td></tr><tr class=""><td>LayerNorm</td><td>49</td><td>5</td></tr><tr class=""><td>LlamaRMSNorm</td><td>49</td><td>5</td></tr><tr class=""><td>VQVAEResnetBlock</td><td>29</td><td>29</td></tr><tr class=""><td>GELUActivation</td><td>27</td><td>5</td></tr><tr class=""><td>ModuleList</td><td>26</td><td>26</td></tr><tr class=""><td>VisionAttention</td><td>24</td><td>2</td></tr><tr class=""><td>VisionEncoderLayer</td><td>24</td><td>2</td></tr><tr class=""><td>VisionMLP</td><td>24</td><td>2</td></tr><tr class=""><td>LlamaAttention</td><td>24</td><td>2</td></tr><tr class=""><td>LlamaDecoderLayer</td><td>24</td><td>2</td></tr><tr class=""><td>LlamaMLP</td><td>24</td><td>2</td></tr><tr class=""><td>SiLUActivation</td><td>24</td><td>2</td></tr><tr class=""><td>Module</td><td>10</td><td>10</td></tr><tr class=""><td>VQVAEAttnBlock</td><td>7</td><td>7</td></tr><tr class=""><td>Embedding</td><td>4</td><td>4</td></tr><tr class=""><td>VQVAEConvDownsample</td><td>4</td><td>4</td></tr><tr class=""><td>VQVAEConvUpsample</td><td>4</td><td>4</td></tr><tr class=""><td>VQVAEMidBlock</td><td>2</td><td>2</td></tr><tr class=""><td>ForConditionalGeneration</td><td>1</td><td>1</td></tr><tr class=""><td>Model</td><td>1</td><td>1</td></tr><tr class=""><td>VQVAE</td><td>1</td><td>1</td></tr><tr class=""><td>VQVAEAlignerMLP</td><td>1</td><td>1</td></tr><tr class=""><td>VQVAEDecoder</td><td>1</td><td>1</td></tr><tr class=""><td>VQVAEEncoder</td><td>1</td><td>1</td></tr><tr class=""><td>VQVAEHead</td><td>1</td><td>1</td></tr><tr class=""><td>VQVAEVectorQuantizer</td><td>1</td><td>1</td></tr><tr class=""><td>VisionAlignerMLP</td><td>1</td><td>1</td></tr><tr class=""><td>VisionEmbeddings</td><td>1</td><td>1</td></tr><tr class=""><td>VisionEncoder</td><td>1</td><td>1</td></tr><tr class=""><td>VisionModel</td><td>1</td><td>1</td></tr><tr class=""><td>LlamaModel</td><td>1</td><td>1</td></tr><tr class=""><td>LlamaRotaryEmbedding</td><td>1</td><td>1</td></tr></table><h2>Config changes</h2><p class="note">12 fields changed, 82 unchanged.</p><table><tr><th>Field</th><th>Original</th><th>Fixture</th></tr><tr><td>text_config.head_dim</td><td>128</td><td>16</td></tr><tr><td>text_config.hidden_size</td><td>2048</td><td>64</td></tr><tr><td>text_config.intermediate_size</td><td>5632</td><td>128</td></tr><tr><td>text_config.num_attention_heads</td><td>16</td><td>4</td></tr><tr><td>text_config.num_hidden_layers</td><td>24</td><td>2</td></tr><tr><td>text_config.num_key_value_heads</td><td>16</td><td>4</td></tr><tr><td>vision_config.hidden_size</td><td>1024</td><td>64</td></tr><tr><td>vision_config.num_attention_heads</td><td>16</td><td>4</td></tr><tr><td>vision_config.num_hidden_layers</td><td>24</td><td>2</td></tr><tr><td>vision_config.projection_dim</td><td>2048</td><td>64</td></tr><tr><td>vq_config.embed_dim</td><td>8</td><td>64</td></tr><tr><td>vq_config.projection_dim</td><td>2048</td><td>64</td></tr></table>
+</main></body></html>

shrink/LCO-Embedding-Omni-3B-mic.html

@@ -0,0 +1,34 @@
+<!doctype html>
+<html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
+<title>Shrunk architecture</title>
+<style>
+:root{--bg:#fcfcfb;--fg:#0b0b0b;--muted:#52514e;--line:#e4e3df;--ok:#1d7a43;--warn:#b35c00;--s0:#2a78d6;--s1:#eb6834;--s2:#1baf7a;--s3:#eda100;--s4:#e87ba4;--s5:#008300;--s6:#6250d6;--s7:#e34948;--s8:#8a8984;}
+@media (prefers-color-scheme:dark){:root:not([data-theme="light"]){--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}}
+:root[data-theme="dark"]{--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}
+body{margin:0;background:var(--bg);color:var(--fg);font:15px/1.5 system-ui,sans-serif}
+main{max-width:1040px;margin:0 auto;padding:24px 16px 64px}
+h1{font-size:24px;margin:0 0 4px}h2{font-size:18px;margin:36px 0 8px}h3{font-size:14px;margin:0 0 8px;font-family:ui-monospace,monospace}
+.meta,.note{color:var(--muted);margin:0 0 12px;overflow-wrap:anywhere}
+.stats{display:grid;grid-template-columns:repeat(auto-fit,minmax(150px,1fr));gap:12px;margin:20px 0}
+.stat{border:1px solid var(--line);border-radius:8px;padding:12px;display:flex;flex-direction:column}
+.stat .v{font-size:26px;font-weight:600}.stat .k{color:var(--muted);font-size:13px}
+.findings{border-radius:8px;padding:4px 16px;border:1px solid var(--line)}
+.findings.ok{border-left:4px solid var(--ok)}.findings.warn{border-left:4px solid var(--warn)}
+.findings h2{margin:12px 0 4px;font-size:16px}
+.stack{border:1px solid var(--line);border-radius:8px;padding:12px;margin:12px 0}
+.row{display:flex;align-items:center;gap:12px;margin:6px 0}
+.lab{flex:0 0 110px;color:var(--muted);font-size:13px}
+.cells{display:flex;flex-wrap:wrap;gap:2px}
+.cells i{width:12px;height:12px;border-radius:2px;display:block}
+.bar{flex:1;display:flex;gap:2px;height:16px}.bar i{display:block;border-radius:2px;min-width:2px}
+table{border-collapse:collapse;width:100%;font-size:13px;margin:8px 0;display:block;overflow-x:auto}
+th,td{text-align:left;padding:4px 10px 4px 0;border-bottom:1px solid var(--line);vertical-align:top}
+th{color:var(--muted);font-weight:500}
+td:first-child{font-family:ui-monospace,monospace;overflow-wrap:anywhere}
+td i{display:inline-block;width:10px;height:10px;border-radius:2px;margin-right:6px}
+tr.bad td{color:var(--warn)}
+.c0{background:var(--s0)}.c1{background:var(--s1)}.c2{background:var(--s2)}.c3{background:var(--s3)}.c4{background:var(--s4)}.c5{background:var(--s5)}.c6{background:var(--s6)}.c7{background:var(--s7)}.c8{background:var(--s8)}
+</style></head><body><main>
+<h1>Shrunk architecture</h1><p class="meta">huggingface-mic/LCO-Embedding-Omni-3B-mic@1e770f7015e573f2c4287893af7c6eb0c1d45c33 ← LCO-Embedding/LCO-Embedding-Omni-3B@73434565c33d032551639362f7fc91fa3c6f065f · Qwen2_5OmniThinkerForConditionalGeneration · random weights</p>
+<div class="stats"><div class="stat"><span class="k">Original</span><span class="v">4.7B</span><span class="k">parameters</span></div><div class="stat"><span class="k">Fixture</span><span class="v">21.1M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Layer stacks</span><span class="v">3</span><span class="k">compared</span></div><div class="stat"><span class="k">Findings</span><span class="v">0</span><span class="k">to review</span></div></div><section class="findings ok"><h2>✓ Same architecture, smaller</h2><p>Every module class, layer kind and checked property of the original is in the fixture.</p></section><h2>Layer stacks</h2><p class="note">One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.</p><div class="stack"><h3>model.audio_tower.layers</h3><div class="row"><span class="lab">Original · 32</span><div class="cells"><i class="c0" title="layer 0: Qwen2_5OmniAudioEncoderLayer"></i><i class="c0" title="layer 1: Qwen2_5OmniAudioEncoderLayer"></i><i class="c0" title="layer 2: Qwen2_5OmniAudioEncoderLayer"></i><i class="c0" title="layer 3: Qwen2_5OmniAudioEncoderLayer"></i><i class="c0" title="layer 4: Qwen2_5OmniAudioEncoderLayer"></i><i class="c0" title="layer 5: Qwen2_5OmniAudioEncoderLayer"></i><i class="c0" title="layer 6: Qwen2_5OmniAudioEncoderLayer"></i><i class="c0" title="layer 7: Qwen2_5OmniAudioEncoderLayer"></i><i class="c0" title="layer 8: Qwen2_5OmniAudioEncoderLayer"></i><i class="c0" title="layer 9: Qwen2_5OmniAudioEncoderLayer"></i><i class="c0" title="layer 10: Qwen2_5OmniAudioEncoderLayer"></i><i class="c0" title="layer 11: Qwen2_5OmniAudioEncoderLayer"></i><i class="c0" title="layer 12: Qwen2_5OmniAudioEncoderLayer"></i><i class="c0" title="layer 13: Qwen2_5OmniAudioEncoderLayer"></i><i class="c0" title="layer 14: Qwen2_5OmniAudioEncoderLayer"></i><i class="c0" title="layer 15: Qwen2_5OmniAudioEncoderLayer"></i><i class="c0" title="layer 16: Qwen2_5OmniAudioEncoderLayer"></i><i class="c0" title="layer 17: Qwen2_5OmniAudioEncoderLayer"></i><i class="c0" title="layer 18: Qwen2_5OmniAudioEncoderLayer"></i><i class="c0" title="layer 19: Qwen2_5OmniAudioEncoderLayer"></i><i class="c0" title="layer 20: Qwen2_5OmniAudioEncoderLayer"></i><i class="c0" title="layer 21: Qwen2_5OmniAudioEncoderLayer"></i><i class="c0" title="layer 22: Qwen2_5OmniAudioEncoderLayer"></i><i class="c0" title="layer 23: Qwen2_5OmniAudioEncoderLayer"></i><i class="c0" title="layer 24: Qwen2_5OmniAudioEncoderLayer"></i><i class="c0" title="layer 25: Qwen2_5OmniAudioEncoderLayer"></i><i class="c0" title="layer 26: Qwen2_5OmniAudioEncoderLayer"></i><i class="c0" title="layer 27: Qwen2_5OmniAudioEncoderLayer"></i><i class="c0" title="layer 28: Qwen2_5OmniAudioEncoderLayer"></i><i class="c0" title="layer 29: Qwen2_5OmniAudioEncoderLayer"></i><i class="c0" title="layer 30: Qwen2_5OmniAudioEncoderLayer"></i><i class="c0" title="layer 31: Qwen2_5OmniAudioEncoderLayer"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: Qwen2_5OmniAudioEncoderLayer"></i><i class="c0" title="layer 1: Qwen2_5OmniAudioEncoderLayer"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>Qwen2_5OmniAudioEncoderLayer</td><td>32</td><td>2</td><td>✓ kept</td></tr></table></div><div class="stack"><h3>model.visual.blocks</h3><div class="row"><span class="lab">Original · 32</span><div class="cells"><i class="c0" title="layer 0: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 1: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 2: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 3: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 4: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 5: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 6: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 7: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 8: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 9: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 10: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 11: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 12: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 13: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 14: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 15: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 16: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 17: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 18: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 19: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 20: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 21: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 22: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 23: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 24: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 25: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 26: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 27: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 28: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 29: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 30: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 31: Qwen2_5OmniVisionBlock"></i></div></div><div class="row"><span class="lab">Fixture · 32</span><div class="cells"><i class="c0" title="layer 0: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 1: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 2: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 3: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 4: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 5: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 6: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 7: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 8: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 9: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 10: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 11: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 12: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 13: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 14: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 15: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 16: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 17: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 18: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 19: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 20: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 21: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 22: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 23: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 24: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 25: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 26: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 27: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 28: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 29: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 30: Qwen2_5OmniVisionBlock"></i><i class="c0" title="layer 31: Qwen2_5OmniVisionBlock"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>Qwen2_5OmniVisionBlock</td><td>32</td><td>32</td><td>✓ kept</td></tr></table></div><div class="stack"><h3>model.language_model.layers</h3><div class="row"><span class="lab">Original · 36</span><div class="cells"><i class="c0" title="layer 0: Qwen2_5OmniDecoderLayer"></i><i class="c0" title="layer 1: Qwen2_5OmniDecoderLayer"></i><i class="c0" title="layer 2: Qwen2_5OmniDecoderLayer"></i><i class="c0" title="layer 3: Qwen2_5OmniDecoderLayer"></i><i class="c0" title="layer 4: Qwen2_5OmniDecoderLayer"></i><i class="c0" title="layer 5: Qwen2_5OmniDecoderLayer"></i><i class="c0" title="layer 6: Qwen2_5OmniDecoderLayer"></i><i class="c0" title="layer 7: Qwen2_5OmniDecoderLayer"></i><i class="c0" title="layer 8: Qwen2_5OmniDecoderLayer"></i><i class="c0" title="layer 9: Qwen2_5OmniDecoderLayer"></i><i class="c0" title="layer 10: Qwen2_5OmniDecoderLayer"></i><i class="c0" title="layer 11: Qwen2_5OmniDecoderLayer"></i><i class="c0" title="layer 12: Qwen2_5OmniDecoderLayer"></i><i class="c0" title="layer 13: Qwen2_5OmniDecoderLayer"></i><i class="c0" title="layer 14: Qwen2_5OmniDecoderLayer"></i><i class="c0" title="layer 15: Qwen2_5OmniDecoderLayer"></i><i class="c0" title="layer 16: Qwen2_5OmniDecoderLayer"></i><i class="c0" title="layer 17: Qwen2_5OmniDecoderLayer"></i><i class="c0" title="layer 18: Qwen2_5OmniDecoderLayer"></i><i class="c0" title="layer 19: Qwen2_5OmniDecoderLayer"></i><i class="c0" title="layer 20: Qwen2_5OmniDecoderLayer"></i><i class="c0" title="layer 21: Qwen2_5OmniDecoderLayer"></i><i class="c0" title="layer 22: Qwen2_5OmniDecoderLayer"></i><i class="c0" title="layer 23: Qwen2_5OmniDecoderLayer"></i><i class="c0" title="layer 24: Qwen2_5OmniDecoderLayer"></i><i class="c0" title="layer 25: Qwen2_5OmniDecoderLayer"></i><i class="c0" title="layer 26: Qwen2_5OmniDecoderLayer"></i><i class="c0" title="layer 27: Qwen2_5OmniDecoderLayer"></i><i class="c0" title="layer 28: Qwen2_5OmniDecoderLayer"></i><i class="c0" title="layer 29: Qwen2_5OmniDecoderLayer"></i><i class="c0" title="layer 30: Qwen2_5OmniDecoderLayer"></i><i class="c0" title="layer 31: Qwen2_5OmniDecoderLayer"></i><i class="c0" title="layer 32: Qwen2_5OmniDecoderLayer"></i><i class="c0" title="layer 33: Qwen2_5OmniDecoderLayer"></i><i class="c0" title="layer 34: Qwen2_5OmniDecoderLayer"></i><i class="c0" title="layer 35: Qwen2_5OmniDecoderLayer"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: Qwen2_5OmniDecoderLayer"></i><i class="c0" title="layer 1: Qwen2_5OmniDecoderLayer"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>Qwen2_5OmniDecoderLayer</td><td>36</td><td>2</td><td>✓ kept</td></tr></table></div><h2>Where the parameters are</h2><p class="note">Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.</p><div class="row"><span class="lab">Original</span><div class="bar"><i class="c0" style="flex:0.589942" title="model.language_model.layers: 2.8B (59.0%)"></i><i class="c1" style="flex:0.134044" title="model.visual.blocks: 630.5M (13.4%)"></i><i class="c2" style="flex:0.133867" title="model.audio_tower.layers: 629.6M (13.4%)"></i><i class="c3" style="flex:0.066157" title="model.language_model: 311.2M (6.6%)"></i><i class="c4" style="flex:0.066157" title="lm_head: 311.2M (6.6%)"></i><i class="c5" style="flex:0.008125" title="model.visual: 38.2M (0.8%)"></i><i class="c6" style="flex:0.001709" title="model.audio_tower: 8.0M (0.2%)"></i></div></div><div class="row"><span class="lab">Fixture</span><div class="bar"><i class="c0" style="flex:0.003515" title="model.language_model.layers: 74.2K (0.4%)"></i><i class="c1" style="flex:0.063124" title="model.visual.blocks: 1.3M (6.3%)"></i><i class="c2" style="flex:0.003163" title="model.audio_tower.layers: 66.8K (0.3%)"></i><i class="c3" style="flex:0.460390" title="model.language_model: 9.7M (46.0%)"></i><i class="c4" style="flex:0.460387" title="lm_head: 9.7M (46.0%)"></i><i class="c5" style="flex:0.007460" title="model.visual: 157.6K (0.7%)"></i><i class="c6" style="flex:0.001960" title="model.audio_tower: 41.4K (0.2%)"></i></div></div><table class="legend"><tr><th>Component</th><th>Original</th><th>Share</th><th>Fixture</th><th>Share</th></tr><tr><td><i class="c0"></i>model.language_model.layers</td><td>2.8B</td><td>59.0%</td><td>74.2K</td><td>0.4%</td></tr><tr><td><i class="c1"></i>model.visual.blocks</td><td>630.5M</td><td>13.4%</td><td>1.3M</td><td>6.3%</td></tr><tr><td><i class="c2"></i>model.audio_tower.layers</td><td>629.6M</td><td>13.4%</td><td>66.8K</td><td>0.3%</td></tr><tr><td><i class="c3"></i>model.language_model</td><td>311.2M</td><td>6.6%</td><td>9.7M</td><td>46.0%</td></tr><tr><td><i class="c4"></i>lm_head</td><td>311.2M</td><td>6.6%</td><td>9.7M</td><td>46.0%</td></tr><tr><td><i class="c5"></i>model.visual</td><td>38.2M</td><td>0.8%</td><td>157.6K</td><td>0.7%</td></tr><tr><td><i class="c6"></i>model.audio_tower</td><td>8.0M</td><td>0.2%</td><td>41.4K</td><td>0.2%</td></tr></table><h2>Preserved properties</h2><table><tr><th>Config</th><th>Property</th><th>Original</th><th>Fixture</th><th></th></tr><tr class=""><td>model</td><td>model_type</td><td>qwen2_5_omni_thinker</td><td>qwen2_5_omni_thinker</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>tie_word_embeddings</td><td>False</td><td>False</td><td>✓ kept</td></tr><tr class=""><td>audio_config</td><td>model_type</td><td>qwen2_5_omni_audio_encoder</td><td>qwen2_5_omni_audio_encoder</td><td>✓ kept</td></tr><tr class=""><td>audio_config</td><td>num_mel_bins</td><td>128</td><td>128</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>RoPE type</td><td>axial</td><td>axial</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>model_type</td><td>qwen2_5_omni_vision_encoder</td><td>qwen2_5_omni_vision_encoder</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>hidden_act</td><td>silu</td><td>silu</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>patch_size</td><td>14</td><td>14</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>attention style</td><td>GQA (16 heads, 2 KV)</td><td>GQA (4 heads, 2 KV)</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>RoPE type</td><td>default</td><td>default</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>model_type</td><td>qwen2_5_omni_text</td><td>qwen2_5_omni_text</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>vocab_size</td><td>151936</td><td>151936</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>tie_word_embeddings</td><td>True</td><td>True</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>hidden_act</td><td>silu</td><td>silu</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>max_position_embeddings</td><td>32768</td><td>32768</td><td>✓ kept</td></tr></table><h2>Module classes</h2><p class="note">Instances of each class in the built model (prefix Qwen2_5OmniThinker omitted).</p><table><tr><th>Class</th><th>Original</th><th>Fixture</th></tr><tr class=""><td>Linear</td><td>672</td><td>254</td></tr><tr class=""><td>Qwen2_5OmniRMSNorm</td><td>138</td><td>70</td></tr><tr class=""><td>SiLUActivation</td><td>68</td><td>34</td></tr><tr class=""><td>LayerNorm</td><td>65</td><td>5</td></tr><tr class=""><td>Qwen2MLP</td><td>36</td><td>2</td></tr><tr class=""><td>Qwen2_5OmniAttention</td><td>36</td><td>2</td></tr><tr class=""><td>Qwen2_5OmniDecoderLayer</td><td>36</td><td>2</td></tr><tr class=""><td>GELUActivation</td><td>32</td><td>2</td></tr><tr class=""><td>Qwen2_5OmniAudioAttention</td><td>32</td><td>2</td></tr><tr class=""><td>Qwen2_5OmniAudioEncoderLayer</td><td>32</td><td>2</td></tr><tr class=""><td>Qwen2_5OmniMLP</td><td>32</td><td>32</td></tr><tr class=""><td>Qwen2_5OmniVisionAttention</td><td>32</td><td>32</td></tr><tr class=""><td>Qwen2_5OmniVisionBlock</td><td>32</td><td>32</td></tr><tr class=""><td>ModuleList</td><td>3</td><td>3</td></tr><tr class=""><td>Conv1d</td><td>2</td><td>2</td></tr><tr class=""><td>Embedding</td><td>2</td><td>2</td></tr><tr class=""><td>AvgPool1d</td><td>1</td><td>1</td></tr><tr class=""><td>Conv3d</td><td>1</td><td>1</td></tr><tr class=""><td>GELU</td><td>1</td><td>1</td></tr><tr class=""><td>Qwen2_5OmniAudioEncoder</td><td>1</td><td>1</td></tr><tr class=""><td>Qwen2_5OmniPatchMerger</td><td>1</td><td>1</td></tr><tr class=""><td>Qwen2_5OmniRotaryEmbedding</td><td>1</td><td>1</td></tr><tr class=""><td>ForConditionalGeneration</td><td>1</td><td>1</td></tr><tr class=""><td>Model</td><td>1</td><td>1</td></tr><tr class=""><td>TextModel</td><td>1</td><td>1</td></tr><tr class=""><td>Qwen2_5OmniVisionEncoder</td><td>1</td><td>1</td></tr><tr class=""><td>Qwen2_5OmniVisionRotaryEmbedding</td><td>1</td><td>1</td></tr><tr class=""><td>Qwen2_5_VisionPatchEmbed</td><td>1</td><td>1</td></tr><tr class=""><td>Sequential</td><td>1</td><td>1</td></tr><tr class=""><td>SinusoidsPositionEmbedding</td><td>1</td><td>1</td></tr></table><h2>Config changes</h2><p class="note">16 fields changed, 102 unchanged.</p><table><tr><th>Field</th><th>Original</th><th>Fixture</th></tr><tr><td>audio_config.d_model</td><td>1280</td><td>64</td></tr><tr><td>audio_config.encoder_attention_heads</td><td>20</td><td>4</td></tr><tr><td>audio_config.encoder_ffn_dim</td><td>5120</td><td>128</td></tr><tr><td>audio_config.encoder_layers</td><td>32</td><td>2</td></tr><tr><td>audio_config.output_dim</td><td>2048</td><td>64</td></tr><tr><td>text_config.hidden_size</td><td>2048</td><td>64</td></tr><tr><td>text_config.intermediate_size</td><td>11008</td><td>128</td></tr><tr><td>text_config.layer_types</td><td>36 entries: 36× full_attention</td><td>[&quot;full_attention&quot;, &quot;full_attention&quot;]</td></tr><tr><td>text_config.num_attention_heads</td><td>16</td><td>4</td></tr><tr><td>text_config.num_hidden_layers</td><td>36</td><td>2</td></tr><tr><td>text_config.rope_parameters</td><td>{&quot;mrope_section&quot;: [16, 24, 24], &quot;rope_type&quot;: &quot;default&quot;, &quot;type&quot;: &quot;default&quot;, &quot;r…</td><td>{&quot;mrope_section&quot;: [2, 3, 3], &quot;rope_theta&quot;: 1000000.0, &quot;rope_type&quot;: &quot;default&quot;,…</td></tr><tr><td>vision_config.embed_dim</td><td>1280</td><td>64</td></tr><tr><td>vision_config.hidden_size</td><td>1280</td><td>64</td></tr><tr><td>vision_config.intermediate_size</td><td>3420</td><td>128</td></tr><tr><td>vision_config.num_heads</td><td>16</td><td>4</td></tr><tr><td>vision_config.out_hidden_size</td><td>2048</td><td>64</td></tr></table>
+</main></body></html>

shrink/LFM2-8B-A1B-mic.html

@@ -0,0 +1,34 @@
+<!doctype html>
+<html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
+<title>Shrunk architecture</title>
+<style>
+:root{--bg:#fcfcfb;--fg:#0b0b0b;--muted:#52514e;--line:#e4e3df;--ok:#1d7a43;--warn:#b35c00;--s0:#2a78d6;--s1:#eb6834;--s2:#1baf7a;--s3:#eda100;--s4:#e87ba4;--s5:#008300;--s6:#6250d6;--s7:#e34948;--s8:#8a8984;}
+@media (prefers-color-scheme:dark){:root:not([data-theme="light"]){--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}}
+:root[data-theme="dark"]{--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}
+body{margin:0;background:var(--bg);color:var(--fg);font:15px/1.5 system-ui,sans-serif}
+main{max-width:1040px;margin:0 auto;padding:24px 16px 64px}
+h1{font-size:24px;margin:0 0 4px}h2{font-size:18px;margin:36px 0 8px}h3{font-size:14px;margin:0 0 8px;font-family:ui-monospace,monospace}
+.meta,.note{color:var(--muted);margin:0 0 12px;overflow-wrap:anywhere}
+.stats{display:grid;grid-template-columns:repeat(auto-fit,minmax(150px,1fr));gap:12px;margin:20px 0}
+.stat{border:1px solid var(--line);border-radius:8px;padding:12px;display:flex;flex-direction:column}
+.stat .v{font-size:26px;font-weight:600}.stat .k{color:var(--muted);font-size:13px}
+.findings{border-radius:8px;padding:4px 16px;border:1px solid var(--line)}
+.findings.ok{border-left:4px solid var(--ok)}.findings.warn{border-left:4px solid var(--warn)}
+.findings h2{margin:12px 0 4px;font-size:16px}
+.stack{border:1px solid var(--line);border-radius:8px;padding:12px;margin:12px 0}
+.row{display:flex;align-items:center;gap:12px;margin:6px 0}
+.lab{flex:0 0 110px;color:var(--muted);font-size:13px}
+.cells{display:flex;flex-wrap:wrap;gap:2px}
+.cells i{width:12px;height:12px;border-radius:2px;display:block}
+.bar{flex:1;display:flex;gap:2px;height:16px}.bar i{display:block;border-radius:2px;min-width:2px}
+table{border-collapse:collapse;width:100%;font-size:13px;margin:8px 0;display:block;overflow-x:auto}
+th,td{text-align:left;padding:4px 10px 4px 0;border-bottom:1px solid var(--line);vertical-align:top}
+th{color:var(--muted);font-weight:500}
+td:first-child{font-family:ui-monospace,monospace;overflow-wrap:anywhere}
+td i{display:inline-block;width:10px;height:10px;border-radius:2px;margin-right:6px}
+tr.bad td{color:var(--warn)}
+.c0{background:var(--s0)}.c1{background:var(--s1)}.c2{background:var(--s2)}.c3{background:var(--s3)}.c4{background:var(--s4)}.c5{background:var(--s5)}.c6{background:var(--s6)}.c7{background:var(--s7)}.c8{background:var(--s8)}
+</style></head><body><main>
+<h1>Shrunk architecture</h1><p class="meta">huggingface-mic/LFM2-8B-A1B-mic@02493cd540778260b98790f375fb2070e8acd824 ← LiquidAI/LFM2-8B-A1B@c1c44ff9fc00db3ebf4516970563f5f383d23670 · Lfm2MoeForCausalLM · random weights</p>
+<div class="stats"><div class="stat"><span class="k">Original</span><span class="v">8.3B</span><span class="k">parameters</span></div><div class="stat"><span class="k">Fixture</span><span class="v">4.4M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Layer stacks</span><span class="v">1</span><span class="k">compared</span></div><div class="stat"><span class="k">Findings</span><span class="v">0</span><span class="k">to review</span></div></div><section class="findings ok"><h2>✓ Same architecture, smaller</h2><p>Every module class, layer kind and checked property of the original is in the fixture.</p></section><h2>Layer stacks</h2><p class="note">One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.</p><div class="stack"><h3>model.layers</h3><div class="row"><span class="lab">Original · 24</span><div class="cells"><i class="c0" title="layer 0: conv, Conv1d, MLP …"></i><i class="c0" title="layer 1: conv, Conv1d, MLP …"></i><i class="c1" title="layer 2: is_causal=True, Attention, Experts …"></i><i class="c2" title="layer 3: conv, Conv1d, Experts …"></i><i class="c2" title="layer 4: conv, Conv1d, Experts …"></i><i class="c2" title="layer 5: conv, Conv1d, Experts …"></i><i class="c1" title="layer 6: is_causal=True, Attention, Experts …"></i><i class="c2" title="layer 7: conv, Conv1d, Experts …"></i><i class="c2" title="layer 8: conv, Conv1d, Experts …"></i><i class="c2" title="layer 9: conv, Conv1d, Experts …"></i><i class="c1" title="layer 10: is_causal=True, Attention, Experts …"></i><i class="c2" title="layer 11: conv, Conv1d, Experts …"></i><i class="c2" title="layer 12: conv, Conv1d, Experts …"></i><i class="c2" title="layer 13: conv, Conv1d, Experts …"></i><i class="c1" title="layer 14: is_causal=True, Attention, Experts …"></i><i class="c2" title="layer 15: conv, Conv1d, Experts …"></i><i class="c2" title="layer 16: conv, Conv1d, Experts …"></i><i class="c2" title="layer 17: conv, Conv1d, Experts …"></i><i class="c1" title="layer 18: is_causal=True, Attention, Experts …"></i><i class="c2" title="layer 19: conv, Conv1d, Experts …"></i><i class="c2" title="layer 20: conv, Conv1d, Experts …"></i><i class="c1" title="layer 21: is_causal=True, Attention, Experts …"></i><i class="c2" title="layer 22: conv, Conv1d, Experts …"></i><i class="c2" title="layer 23: conv, Conv1d, Experts …"></i></div></div><div class="row"><span class="lab">Fixture · 4</span><div class="cells"><i class="c0" title="layer 0: conv, Conv1d, MLP …"></i><i class="c0" title="layer 1: conv, Conv1d, MLP …"></i><i class="c1" title="layer 2: is_causal=True, Attention, Experts …"></i><i class="c2" title="layer 3: conv, Conv1d, Experts …"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>conv, Conv1d, MLP …</td><td>2</td><td>2</td><td>✓ kept</td></tr><tr><td><i class="c1"></i>is_causal=True, Attention, Experts …</td><td>6</td><td>1</td><td>✓ kept</td></tr><tr><td><i class="c2"></i>conv, Conv1d, Experts …</td><td>16</td><td>1</td><td>✓ kept</td></tr></table></div><h2>Where the parameters are</h2><p class="note">Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.</p><div class="row"><span class="lab">Original</span><div class="bar"><i class="c0" style="flex:0.983906" title="model.layers: 8.2B (98.4%)"></i><i class="c1" style="flex:0.016093" title="model.embed_tokens: 134.2M (1.6%)"></i><i class="c2" style="flex:0.000000" title="model.embedding_norm: 2.0K (0.0%)"></i></div></div><div class="row"><span class="lab">Fixture</span><div class="bar"><i class="c0" style="flex:0.037049" title="model.layers: 161.4K (3.7%)"></i><i class="c1" style="flex:0.962936" title="model.embed_tokens: 4.2M (96.3%)"></i><i class="c2" style="flex:0.000015" title="model.embedding_norm: 64 (0.0%)"></i></div></div><table class="legend"><tr><th>Component</th><th>Original</th><th>Share</th><th>Fixture</th><th>Share</th></tr><tr><td><i class="c0"></i>model.layers</td><td>8.2B</td><td>98.4%</td><td>161.4K</td><td>3.7%</td></tr><tr><td><i class="c1"></i>model.embed_tokens</td><td>134.2M</td><td>1.6%</td><td>4.2M</td><td>96.3%</td></tr><tr><td><i class="c2"></i>model.embedding_norm</td><td>2.0K</td><td>0.0%</td><td>64</td><td>0.0%</td></tr></table><h2>Preserved properties</h2><table><tr><th>Config</th><th>Property</th><th>Original</th><th>Fixture</th><th></th></tr><tr class=""><td>model</td><td>attention style</td><td>GQA (32 heads, 8 KV)</td><td>GQA (4 heads, 2 KV)</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>MoE routing</td><td>top-4 of 32 experts</td><td>top-2 of 4 experts</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>RoPE type</td><td>default</td><td>default</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>model_type</td><td>lfm2_moe</td><td>lfm2_moe</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>vocab_size</td><td>65536</td><td>65536</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>tie_word_embeddings</td><td>True</td><td>True</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>max_position_embeddings</td><td>128000</td><td>128000</td><td>✓ kept</td></tr></table><h2>Module classes</h2><p class="note">Instances of each class in the built model (prefix Lfm2Moe omitted).</p><table><tr><th>Class</th><th>Original</th><th>Fixture</th></tr><tr class=""><td>Linear</td><td>67</td><td>17</td></tr><tr class=""><td>RMSNorm</td><td>61</td><td>11</td></tr><tr class=""><td>DecoderLayer</td><td>24</td><td>4</td></tr><tr class=""><td>Experts</td><td>22</td><td>2</td></tr><tr class=""><td>SparseMoeBlock</td><td>22</td><td>2</td></tr><tr class=""><td>TopKRouter</td><td>22</td><td>2</td></tr><tr class=""><td>Conv1d</td><td>18</td><td>3</td></tr><tr class=""><td>ShortConv</td><td>18</td><td>3</td></tr><tr class=""><td>Attention</td><td>6</td><td>1</td></tr><tr class=""><td>MLP</td><td>2</td><td>2</td></tr><tr class=""><td>Embedding</td><td>1</td><td>1</td></tr><tr class=""><td>ForCausalLM</td><td>1</td><td>1</td></tr><tr class=""><td>Model</td><td>1</td><td>1</td></tr><tr class=""><td>RotaryEmbedding</td><td>1</td><td>1</td></tr><tr class=""><td>ModuleList</td><td>1</td><td>1</td></tr></table><h2>Config changes</h2><p class="note">9 fields changed, 26 unchanged.</p><table><tr><th>Field</th><th>Original</th><th>Fixture</th></tr><tr><td>hidden_size</td><td>2048</td><td>64</td></tr><tr><td>intermediate_size</td><td>7168</td><td>128</td></tr><tr><td>layer_types</td><td>24 entries: 18× conv, 6× full_attention</td><td>[&quot;conv&quot;, &quot;conv&quot;, &quot;full_attention&quot;, &quot;conv&quot;]</td></tr><tr><td>moe_intermediate_size</td><td>1792</td><td>32</td></tr><tr><td>num_attention_heads</td><td>32</td><td>4</td></tr><tr><td>num_experts</td><td>32</td><td>4</td></tr><tr><td>num_experts_per_tok</td><td>4</td><td>2</td></tr><tr><td>num_hidden_layers</td><td>24</td><td>4</td></tr><tr><td>num_key_value_heads</td><td>8</td><td>2</td></tr></table>
+</main></body></html>

shrink/LFM2.5-230M-Base-mic.html

@@ -0,0 +1,34 @@
+<!doctype html>
+<html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
+<title>Shrunk architecture</title>
+<style>
+:root{--bg:#fcfcfb;--fg:#0b0b0b;--muted:#52514e;--line:#e4e3df;--ok:#1d7a43;--warn:#b35c00;--s0:#2a78d6;--s1:#eb6834;--s2:#1baf7a;--s3:#eda100;--s4:#e87ba4;--s5:#008300;--s6:#6250d6;--s7:#e34948;--s8:#8a8984;}
+@media (prefers-color-scheme:dark){:root:not([data-theme="light"]){--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}}
+:root[data-theme="dark"]{--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}
+body{margin:0;background:var(--bg);color:var(--fg);font:15px/1.5 system-ui,sans-serif}
+main{max-width:1040px;margin:0 auto;padding:24px 16px 64px}
+h1{font-size:24px;margin:0 0 4px}h2{font-size:18px;margin:36px 0 8px}h3{font-size:14px;margin:0 0 8px;font-family:ui-monospace,monospace}
+.meta,.note{color:var(--muted);margin:0 0 12px;overflow-wrap:anywhere}
+.stats{display:grid;grid-template-columns:repeat(auto-fit,minmax(150px,1fr));gap:12px;margin:20px 0}
+.stat{border:1px solid var(--line);border-radius:8px;padding:12px;display:flex;flex-direction:column}
+.stat .v{font-size:26px;font-weight:600}.stat .k{color:var(--muted);font-size:13px}
+.findings{border-radius:8px;padding:4px 16px;border:1px solid var(--line)}
+.findings.ok{border-left:4px solid var(--ok)}.findings.warn{border-left:4px solid var(--warn)}
+.findings h2{margin:12px 0 4px;font-size:16px}
+.stack{border:1px solid var(--line);border-radius:8px;padding:12px;margin:12px 0}
+.row{display:flex;align-items:center;gap:12px;margin:6px 0}
+.lab{flex:0 0 110px;color:var(--muted);font-size:13px}
+.cells{display:flex;flex-wrap:wrap;gap:2px}
+.cells i{width:12px;height:12px;border-radius:2px;display:block}
+.bar{flex:1;display:flex;gap:2px;height:16px}.bar i{display:block;border-radius:2px;min-width:2px}
+table{border-collapse:collapse;width:100%;font-size:13px;margin:8px 0;display:block;overflow-x:auto}
+th,td{text-align:left;padding:4px 10px 4px 0;border-bottom:1px solid var(--line);vertical-align:top}
+th{color:var(--muted);font-weight:500}
+td:first-child{font-family:ui-monospace,monospace;overflow-wrap:anywhere}
+td i{display:inline-block;width:10px;height:10px;border-radius:2px;margin-right:6px}
+tr.bad td{color:var(--warn)}
+.c0{background:var(--s0)}.c1{background:var(--s1)}.c2{background:var(--s2)}.c3{background:var(--s3)}.c4{background:var(--s4)}.c5{background:var(--s5)}.c6{background:var(--s6)}.c7{background:var(--s7)}.c8{background:var(--s8)}
+</style></head><body><main>
+<h1>Shrunk architecture</h1><p class="meta">huggingface-mic/LFM2.5-230M-Base-mic@d6697c673a52ed87ae4e8ed607e5b6a1ea6f7a8d ← LiquidAI/LFM2.5-230M-Base@9d2be5519834990d30996f878b6771cccbd24f2c · Lfm2ForCausalLM · random weights</p>
+<div class="stats"><div class="stat"><span class="k">Original</span><span class="v">229.7M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Fixture</span><span class="v">4.3M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Layer stacks</span><span class="v">1</span><span class="k">compared</span></div><div class="stat"><span class="k">Findings</span><span class="v">0</span><span class="k">to review</span></div></div><section class="findings ok"><h2>✓ Same architecture, smaller</h2><p>Every module class, layer kind and checked property of the original is in the fixture.</p></section><h2>Layer stacks</h2><p class="note">One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.</p><div class="stack"><h3>model.layers</h3><div class="row"><span class="lab">Original · 14</span><div class="cells"><i class="c0" title="layer 0: conv, Conv1d, ShortConv"></i><i class="c0" title="layer 1: conv, Conv1d, ShortConv"></i><i class="c1" title="layer 2: is_causal=True, Attention"></i><i class="c0" title="layer 3: conv, Conv1d, ShortConv"></i><i class="c1" title="layer 4: is_causal=True, Attention"></i><i class="c0" title="layer 5: conv, Conv1d, ShortConv"></i><i class="c1" title="layer 6: is_causal=True, Attention"></i><i class="c0" title="layer 7: conv, Conv1d, ShortConv"></i><i class="c1" title="layer 8: is_causal=True, Attention"></i><i class="c0" title="layer 9: conv, Conv1d, ShortConv"></i><i class="c1" title="layer 10: is_causal=True, Attention"></i><i class="c0" title="layer 11: conv, Conv1d, ShortConv"></i><i class="c1" title="layer 12: is_causal=True, Attention"></i><i class="c0" title="layer 13: conv, Conv1d, ShortConv"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: conv, Conv1d, ShortConv"></i><i class="c1" title="layer 1: is_causal=True, Attention"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>conv, Conv1d, ShortConv</td><td>8</td><td>1</td><td>✓ kept</td></tr><tr><td><i class="c1"></i>is_causal=True, Attention</td><td>6</td><td>1</td><td>✓ kept</td></tr></table></div><h2>Where the parameters are</h2><p class="note">Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.</p><div class="row"><span class="lab">Original</span><div class="bar"><i class="c0" style="flex:0.707828" title="model.layers: 162.6M (70.8%)"></i><i class="c1" style="flex:0.292167" title="model.embed_tokens: 67.1M (29.2%)"></i><i class="c2" style="flex:0.000004" title="model.embedding_norm: 1.0K (0.0%)"></i></div></div><div class="row"><span class="lab">Fixture</span><div class="bar"><i class="c0" style="flex:0.018327" title="model.layers: 78.3K (1.8%)"></i><i class="c1" style="flex:0.981658" title="model.embed_tokens: 4.2M (98.2%)"></i><i class="c2" style="flex:0.000015" title="model.embedding_norm: 64 (0.0%)"></i></div></div><table class="legend"><tr><th>Component</th><th>Original</th><th>Share</th><th>Fixture</th><th>Share</th></tr><tr><td><i class="c0"></i>model.layers</td><td>162.6M</td><td>70.8%</td><td>78.3K</td><td>1.8%</td></tr><tr><td><i class="c1"></i>model.embed_tokens</td><td>67.1M</td><td>29.2%</td><td>4.2M</td><td>98.2%</td></tr><tr><td><i class="c2"></i>model.embedding_norm</td><td>1.0K</td><td>0.0%</td><td>64</td><td>0.0%</td></tr></table><h2>Preserved properties</h2><table><tr><th>Config</th><th>Property</th><th>Original</th><th>Fixture</th><th></th></tr><tr class=""><td>model</td><td>attention style</td><td>GQA (16 heads, 8 KV)</td><td>GQA (4 heads, 2 KV)</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>RoPE type</td><td>default</td><td>default</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>model_type</td><td>lfm2</td><td>lfm2</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>vocab_size</td><td>65536</td><td>65536</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>tie_word_embeddings</td><td>True</td><td>True</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>max_position_embeddings</td><td>128000</td><td>128000</td><td>✓ kept</td></tr></table><h2>Module classes</h2><p class="note">Instances of each class in the built model (prefix Lfm2 omitted).</p><table><tr><th>Class</th><th>Original</th><th>Fixture</th></tr><tr class=""><td>Linear</td><td>83</td><td>13</td></tr><tr class=""><td>RMSNorm</td><td>41</td><td>7</td></tr><tr class=""><td>DecoderLayer</td><td>14</td><td>2</td></tr><tr class=""><td>MLP</td><td>14</td><td>2</td></tr><tr class=""><td>Conv1d</td><td>8</td><td>1</td></tr><tr class=""><td>ShortConv</td><td>8</td><td>1</td></tr><tr class=""><td>Attention</td><td>6</td><td>1</td></tr><tr class=""><td>Embedding</td><td>1</td><td>1</td></tr><tr class=""><td>ForCausalLM</td><td>1</td><td>1</td></tr><tr class=""><td>Model</td><td>1</td><td>1</td></tr><tr class=""><td>RotaryEmbedding</td><td>1</td><td>1</td></tr><tr class=""><td>ModuleList</td><td>1</td><td>1</td></tr></table><h2>Config changes</h2><p class="note">7 fields changed, 41 unchanged.</p><table><tr><th>Field</th><th>Original</th><th>Fixture</th></tr><tr><td>hidden_size</td><td>1024</td><td>64</td></tr><tr><td>intermediate_size</td><td>2560</td><td>128</td></tr><tr><td>layer_types</td><td>14 entries: 8× conv, 6× full_attention</td><td>[&quot;conv&quot;, &quot;full_attention&quot;]</td></tr><tr><td>num_attention_heads</td><td>16</td><td>4</td></tr><tr><td>num_heads</td><td>16</td><td>4</td></tr><tr><td>num_hidden_layers</td><td>14</td><td>2</td></tr><tr><td>num_key_value_heads</td><td>8</td><td>2</td></tr></table>
+</main></body></html>

shrink/LFM2.5-VL-450M-mic.html

@@ -0,0 +1,34 @@
+<!doctype html>
+<html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
+<title>Shrunk architecture</title>
+<style>
+:root{--bg:#fcfcfb;--fg:#0b0b0b;--muted:#52514e;--line:#e4e3df;--ok:#1d7a43;--warn:#b35c00;--s0:#2a78d6;--s1:#eb6834;--s2:#1baf7a;--s3:#eda100;--s4:#e87ba4;--s5:#008300;--s6:#6250d6;--s7:#e34948;--s8:#8a8984;}
+@media (prefers-color-scheme:dark){:root:not([data-theme="light"]){--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}}
+:root[data-theme="dark"]{--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}
+body{margin:0;background:var(--bg);color:var(--fg);font:15px/1.5 system-ui,sans-serif}
+main{max-width:1040px;margin:0 auto;padding:24px 16px 64px}
+h1{font-size:24px;margin:0 0 4px}h2{font-size:18px;margin:36px 0 8px}h3{font-size:14px;margin:0 0 8px;font-family:ui-monospace,monospace}
+.meta,.note{color:var(--muted);margin:0 0 12px;overflow-wrap:anywhere}
+.stats{display:grid;grid-template-columns:repeat(auto-fit,minmax(150px,1fr));gap:12px;margin:20px 0}
+.stat{border:1px solid var(--line);border-radius:8px;padding:12px;display:flex;flex-direction:column}
+.stat .v{font-size:26px;font-weight:600}.stat .k{color:var(--muted);font-size:13px}
+.findings{border-radius:8px;padding:4px 16px;border:1px solid var(--line)}
+.findings.ok{border-left:4px solid var(--ok)}.findings.warn{border-left:4px solid var(--warn)}
+.findings h2{margin:12px 0 4px;font-size:16px}
+.stack{border:1px solid var(--line);border-radius:8px;padding:12px;margin:12px 0}
+.row{display:flex;align-items:center;gap:12px;margin:6px 0}
+.lab{flex:0 0 110px;color:var(--muted);font-size:13px}
+.cells{display:flex;flex-wrap:wrap;gap:2px}
+.cells i{width:12px;height:12px;border-radius:2px;display:block}
+.bar{flex:1;display:flex;gap:2px;height:16px}.bar i{display:block;border-radius:2px;min-width:2px}
+table{border-collapse:collapse;width:100%;font-size:13px;margin:8px 0;display:block;overflow-x:auto}
+th,td{text-align:left;padding:4px 10px 4px 0;border-bottom:1px solid var(--line);vertical-align:top}
+th{color:var(--muted);font-weight:500}
+td:first-child{font-family:ui-monospace,monospace;overflow-wrap:anywhere}
+td i{display:inline-block;width:10px;height:10px;border-radius:2px;margin-right:6px}
+tr.bad td{color:var(--warn)}
+.c0{background:var(--s0)}.c1{background:var(--s1)}.c2{background:var(--s2)}.c3{background:var(--s3)}.c4{background:var(--s4)}.c5{background:var(--s5)}.c6{background:var(--s6)}.c7{background:var(--s7)}.c8{background:var(--s8)}
+</style></head><body><main>
+<h1>Shrunk architecture</h1><p class="meta">huggingface-mic/LFM2.5-VL-450M-mic@27b8f6b75e00ddd9758961d5abac08622f2b6d73 ← LiquidAI/LFM2.5-VL-450M@fc6221ca597f3315e4f82fc2df606783267b34ba · Lfm2VlForConditionalGeneration · random weights</p>
+<div class="stats"><div class="stat"><span class="k">Original</span><span class="v">448.7M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Fixture</span><span class="v">5.1M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Layer stacks</span><span class="v">2</span><span class="k">compared</span></div><div class="stat"><span class="k">Findings</span><span class="v">0</span><span class="k">to review</span></div></div><section class="findings ok"><h2>✓ Same architecture, smaller</h2><p>Every module class, layer kind and checked property of the original is in the fixture.</p></section><h2>Layer stacks</h2><p class="note">One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.</p><div class="stack"><h3>model.vision_tower.encoder.layers</h3><div class="row"><span class="lab">Original · 12</span><div class="cells"><i class="c0" title="layer 0: Siglip2EncoderLayer"></i><i class="c0" title="layer 1: Siglip2EncoderLayer"></i><i class="c0" title="layer 2: Siglip2EncoderLayer"></i><i class="c0" title="layer 3: Siglip2EncoderLayer"></i><i class="c0" title="layer 4: Siglip2EncoderLayer"></i><i class="c0" title="layer 5: Siglip2EncoderLayer"></i><i class="c0" title="layer 6: Siglip2EncoderLayer"></i><i class="c0" title="layer 7: Siglip2EncoderLayer"></i><i class="c0" title="layer 8: Siglip2EncoderLayer"></i><i class="c0" title="layer 9: Siglip2EncoderLayer"></i><i class="c0" title="layer 10: Siglip2EncoderLayer"></i><i class="c0" title="layer 11: Siglip2EncoderLayer"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: Siglip2EncoderLayer"></i><i class="c0" title="layer 1: Siglip2EncoderLayer"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>Siglip2EncoderLayer</td><td>12</td><td>2</td><td>✓ kept</td></tr></table></div><div class="stack"><h3>model.language_model.layers</h3><div class="row"><span class="lab">Original · 16</span><div class="cells"><i class="c0" title="layer 0: conv, Conv1d, Lfm2ShortConv"></i><i class="c0" title="layer 1: conv, Conv1d, Lfm2ShortConv"></i><i class="c1" title="layer 2: is_causal=True, Lfm2Attention"></i><i class="c0" title="layer 3: conv, Conv1d, Lfm2ShortConv"></i><i class="c0" title="layer 4: conv, Conv1d, Lfm2ShortConv"></i><i class="c1" title="layer 5: is_causal=True, Lfm2Attention"></i><i class="c0" title="layer 6: conv, Conv1d, Lfm2ShortConv"></i><i class="c0" title="layer 7: conv, Conv1d, Lfm2ShortConv"></i><i class="c1" title="layer 8: is_causal=True, Lfm2Attention"></i><i class="c0" title="layer 9: conv, Conv1d, Lfm2ShortConv"></i><i class="c1" title="layer 10: is_causal=True, Lfm2Attention"></i><i class="c0" title="layer 11: conv, Conv1d, Lfm2ShortConv"></i><i class="c1" title="layer 12: is_causal=True, Lfm2Attention"></i><i class="c0" title="layer 13: conv, Conv1d, Lfm2ShortConv"></i><i class="c1" title="layer 14: is_causal=True, Lfm2Attention"></i><i class="c0" title="layer 15: conv, Conv1d, Lfm2ShortConv"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: conv, Conv1d, Lfm2ShortConv"></i><i class="c1" title="layer 1: is_causal=True, Lfm2Attention"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>conv, Conv1d, Lfm2ShortConv</td><td>10</td><td>1</td><td>✓ kept</td></tr><tr><td><i class="c1"></i>is_causal=True, Lfm2Attention</td><td>6</td><td>1</td><td>✓ kept</td></tr></table></div><h2>Where the parameters are</h2><p class="note">Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.</p><div class="row"><span class="lab">Original</span><div class="bar"><i class="c0" style="flex:0.640432" title="model.language_model.layers: 287.4M (64.0%)"></i><i class="c1" style="flex:0.189550" title="model.vision_tower.encoder.layers: 85.1M (19.0%)"></i><i class="c2" style="flex:0.149559" title="model.language_model: 67.1M (15.0%)"></i><i class="c3" style="flex:0.018701" title="model.multi_modal_projector: 8.4M (1.9%)"></i><i class="c4" style="flex:0.001758" title="model.vision_tower: 788.7K (0.2%)"></i></div></div><div class="row"><span class="lab">Fixture</span><div class="bar"><i class="c0" style="flex:0.024933" title="model.language_model.layers: 127.5K (2.5%)"></i><i class="c1" style="flex:0.013096" title="model.vision_tower.encoder.layers: 66.9K (1.3%)"></i><i class="c2" style="flex:0.820500" title="model.language_model: 4.2M (82.0%)"></i><i class="c3" style="flex:0.128614" title="model.multi_modal_projector: 657.5K (12.9%)"></i><i class="c4" style="flex:0.012858" title="model.vision_tower: 65.7K (1.3%)"></i></div></div><table class="legend"><tr><th>Component</th><th>Original</th><th>Share</th><th>Fixture</th><th>Share</th></tr><tr><td><i class="c0"></i>model.language_model.layers</td><td>287.4M</td><td>64.0%</td><td>127.5K</td><td>2.5%</td></tr><tr><td><i class="c1"></i>model.vision_tower.encoder.layers</td><td>85.1M</td><td>19.0%</td><td>66.9K</td><td>1.3%</td></tr><tr><td><i class="c2"></i>model.language_model</td><td>67.1M</td><td>15.0%</td><td>4.2M</td><td>82.0%</td></tr><tr><td><i class="c3"></i>model.multi_modal_projector</td><td>8.4M</td><td>1.9%</td><td>657.5K</td><td>12.9%</td></tr><tr><td><i class="c4"></i>model.vision_tower</td><td>788.7K</td><td>0.2%</td><td>65.7K</td><td>1.3%</td></tr></table><h2>Preserved properties</h2><table><tr><th>Config</th><th>Property</th><th>Original</th><th>Fixture</th><th></th></tr><tr class=""><td>model</td><td>model_type</td><td>lfm2_vl</td><td>lfm2_vl</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>tie_word_embeddings</td><td>True</td><td>True</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>model_type</td><td>siglip2_vision_model</td><td>siglip2_vision_model</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>hidden_act</td><td>gelu_pytorch_tanh</td><td>gelu_pytorch_tanh</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>patch_size</td><td>16</td><td>16</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>num_channels</td><td>3</td><td>3</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>attention style</td><td>GQA (16 heads, 8 KV)</td><td>GQA (4 heads, 2 KV)</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>RoPE type</td><td>default</td><td>default</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>model_type</td><td>lfm2</td><td>lfm2</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>vocab_size</td><td>65536</td><td>65536</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>tie_word_embeddings</td><td>True</td><td>True</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>max_position_embeddings</td><td>128000</td><td>128000</td><td>✓ kept</td></tr></table><h2>Module classes</h2><p class="note">Instances of each class in the built model (prefix Lfm2Vl omitted).</p><table><tr><th>Class</th><th>Original</th><th>Fixture</th></tr><tr class=""><td>Linear</td><td>168</td><td>28</td></tr><tr class=""><td>Lfm2RMSNorm</td><td>45</td><td>7</td></tr><tr class=""><td>LayerNorm</td><td>25</td><td>5</td></tr><tr class=""><td>Lfm2DecoderLayer</td><td>16</td><td>2</td></tr><tr class=""><td>Lfm2MLP</td><td>16</td><td>2</td></tr><tr class=""><td>GELUTanh</td><td>12</td><td>2</td></tr><tr class=""><td>Siglip2Attention</td><td>12</td><td>2</td></tr><tr class=""><td>Siglip2EncoderLayer</td><td>12</td><td>2</td></tr><tr class=""><td>Siglip2MLP</td><td>12</td><td>2</td></tr><tr class=""><td>Conv1d</td><td>10</td><td>1</td></tr><tr class=""><td>Lfm2ShortConv</td><td>10</td><td>1</td></tr><tr class=""><td>Lfm2Attention</td><td>6</td><td>1</td></tr><tr class=""><td>Embedding</td><td>2</td><td>2</td></tr><tr class=""><td>ModuleList</td><td>2</td><td>2</td></tr><tr class=""><td>GELUActivation</td><td>1</td><td>1</td></tr><tr class=""><td>Lfm2Model</td><td>1</td><td>1</td></tr><tr class=""><td>Lfm2RotaryEmbedding</td><td>1</td><td>1</td></tr><tr class=""><td>ForConditionalGeneration</td><td>1</td><td>1</td></tr><tr class=""><td>Model</td><td>1</td><td>1</td></tr><tr class=""><td>MultiModalProjector</td><td>1</td><td>1</td></tr><tr class=""><td>Siglip2Encoder</td><td>1</td><td>1</td></tr><tr class=""><td>Siglip2VisionEmbeddings</td><td>1</td><td>1</td></tr><tr class=""><td>Siglip2VisionModel</td><td>1</td><td>1</td></tr></table><h2>Config changes</h2><p class="note">11 fields changed, 77 unchanged.</p><table><tr><th>Field</th><th>Original</th><th>Fixture</th></tr><tr><td>text_config.hidden_size</td><td>1024</td><td>64</td></tr><tr><td>text_config.intermediate_size</td><td>6656</td><td>128</td></tr><tr><td>text_config.layer_types</td><td>16 entries: 10× conv, 6× full_attention</td><td>[&quot;conv&quot;, &quot;full_attention&quot;]</td></tr><tr><td>text_config.num_attention_heads</td><td>16</td><td>4</td></tr><tr><td>text_config.num_heads</td><td>16</td><td>4</td></tr><tr><td>text_config.num_hidden_layers</td><td>16</td><td>2</td></tr><tr><td>text_config.num_key_value_heads</td><td>8</td><td>2</td></tr><tr><td>vision_config.hidden_size</td><td>768</td><td>64</td></tr><tr><td>vision_config.intermediate_size</td><td>3072</td><td>128</td></tr><tr><td>vision_config.num_attention_heads</td><td>12</td><td>4</td></tr><tr><td>vision_config.num_hidden_layers</td><td>12</td><td>2</td></tr></table>
+</main></body></html>

shrink/LLaVA-NeXT-Video-7B-hf-mic.html

@@ -0,0 +1,34 @@
+<!doctype html>
+<html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
+<title>Shrunk architecture</title>
+<style>
+:root{--bg:#fcfcfb;--fg:#0b0b0b;--muted:#52514e;--line:#e4e3df;--ok:#1d7a43;--warn:#b35c00;--s0:#2a78d6;--s1:#eb6834;--s2:#1baf7a;--s3:#eda100;--s4:#e87ba4;--s5:#008300;--s6:#6250d6;--s7:#e34948;--s8:#8a8984;}
+@media (prefers-color-scheme:dark){:root:not([data-theme="light"]){--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}}
+:root[data-theme="dark"]{--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}
+body{margin:0;background:var(--bg);color:var(--fg);font:15px/1.5 system-ui,sans-serif}
+main{max-width:1040px;margin:0 auto;padding:24px 16px 64px}
+h1{font-size:24px;margin:0 0 4px}h2{font-size:18px;margin:36px 0 8px}h3{font-size:14px;margin:0 0 8px;font-family:ui-monospace,monospace}
+.meta,.note{color:var(--muted);margin:0 0 12px;overflow-wrap:anywhere}
+.stats{display:grid;grid-template-columns:repeat(auto-fit,minmax(150px,1fr));gap:12px;margin:20px 0}
+.stat{border:1px solid var(--line);border-radius:8px;padding:12px;display:flex;flex-direction:column}
+.stat .v{font-size:26px;font-weight:600}.stat .k{color:var(--muted);font-size:13px}
+.findings{border-radius:8px;padding:4px 16px;border:1px solid var(--line)}
+.findings.ok{border-left:4px solid var(--ok)}.findings.warn{border-left:4px solid var(--warn)}
+.findings h2{margin:12px 0 4px;font-size:16px}
+.stack{border:1px solid var(--line);border-radius:8px;padding:12px;margin:12px 0}
+.row{display:flex;align-items:center;gap:12px;margin:6px 0}
+.lab{flex:0 0 110px;color:var(--muted);font-size:13px}
+.cells{display:flex;flex-wrap:wrap;gap:2px}
+.cells i{width:12px;height:12px;border-radius:2px;display:block}
+.bar{flex:1;display:flex;gap:2px;height:16px}.bar i{display:block;border-radius:2px;min-width:2px}
+table{border-collapse:collapse;width:100%;font-size:13px;margin:8px 0;display:block;overflow-x:auto}
+th,td{text-align:left;padding:4px 10px 4px 0;border-bottom:1px solid var(--line);vertical-align:top}
+th{color:var(--muted);font-weight:500}
+td:first-child{font-family:ui-monospace,monospace;overflow-wrap:anywhere}
+td i{display:inline-block;width:10px;height:10px;border-radius:2px;margin-right:6px}
+tr.bad td{color:var(--warn)}
+.c0{background:var(--s0)}.c1{background:var(--s1)}.c2{background:var(--s2)}.c3{background:var(--s3)}.c4{background:var(--s4)}.c5{background:var(--s5)}.c6{background:var(--s6)}.c7{background:var(--s7)}.c8{background:var(--s8)}
+</style></head><body><main>
+<h1>Shrunk architecture</h1><p class="meta">huggingface-mic/LLaVA-NeXT-Video-7B-hf-mic@e025a3d15d2cc35a893c1362a237f82b09b90410 ← llava-hf/LLaVA-NeXT-Video-7B-hf@f417d9edeb69a9eaa88283780332d00ce3c66593 · LlavaNextVideoForConditionalGeneration · random weights · built before the shrink rule (no recorded recipe)</p>
+<div class="stats"><div class="stat"><span class="k">Original</span><span class="v">7.1B</span><span class="k">parameters</span></div><div class="stat"><span class="k">Fixture</span><span class="v">4.3M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Layer stacks</span><span class="v">2</span><span class="k">compared</span></div><div class="stat"><span class="k">Findings</span><span class="v">0</span><span class="k">to review</span></div></div><section class="findings ok"><h2>✓ Same architecture, smaller</h2><p>Every module class, layer kind and checked property of the original is in the fixture.</p></section><h2>Layer stacks</h2><p class="note">One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.</p><div class="stack"><h3>model.vision_tower.encoder.layers</h3><div class="row"><span class="lab">Original · 24</span><div class="cells"><i class="c0" title="layer 0: CLIPEncoderLayer"></i><i class="c0" title="layer 1: CLIPEncoderLayer"></i><i class="c0" title="layer 2: CLIPEncoderLayer"></i><i class="c0" title="layer 3: CLIPEncoderLayer"></i><i class="c0" title="layer 4: CLIPEncoderLayer"></i><i class="c0" title="layer 5: CLIPEncoderLayer"></i><i class="c0" title="layer 6: CLIPEncoderLayer"></i><i class="c0" title="layer 7: CLIPEncoderLayer"></i><i class="c0" title="layer 8: CLIPEncoderLayer"></i><i class="c0" title="layer 9: CLIPEncoderLayer"></i><i class="c0" title="layer 10: CLIPEncoderLayer"></i><i class="c0" title="layer 11: CLIPEncoderLayer"></i><i class="c0" title="layer 12: CLIPEncoderLayer"></i><i class="c0" title="layer 13: CLIPEncoderLayer"></i><i class="c0" title="layer 14: CLIPEncoderLayer"></i><i class="c0" title="layer 15: CLIPEncoderLayer"></i><i class="c0" title="layer 16: CLIPEncoderLayer"></i><i class="c0" title="layer 17: CLIPEncoderLayer"></i><i class="c0" title="layer 18: CLIPEncoderLayer"></i><i class="c0" title="layer 19: CLIPEncoderLayer"></i><i class="c0" title="layer 20: CLIPEncoderLayer"></i><i class="c0" title="layer 21: CLIPEncoderLayer"></i><i class="c0" title="layer 22: CLIPEncoderLayer"></i><i class="c0" title="layer 23: CLIPEncoderLayer"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: CLIPEncoderLayer"></i><i class="c0" title="layer 1: CLIPEncoderLayer"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>CLIPEncoderLayer</td><td>24</td><td>2</td><td>✓ kept</td></tr></table></div><div class="stack"><h3>model.language_model.layers</h3><div class="row"><span class="lab">Original · 32</span><div class="cells"><i class="c0" title="layer 0: LlamaDecoderLayer"></i><i class="c0" title="layer 1: LlamaDecoderLayer"></i><i class="c0" title="layer 2: LlamaDecoderLayer"></i><i class="c0" title="layer 3: LlamaDecoderLayer"></i><i class="c0" title="layer 4: LlamaDecoderLayer"></i><i class="c0" title="layer 5: LlamaDecoderLayer"></i><i class="c0" title="layer 6: LlamaDecoderLayer"></i><i class="c0" title="layer 7: LlamaDecoderLayer"></i><i class="c0" title="layer 8: LlamaDecoderLayer"></i><i class="c0" title="layer 9: LlamaDecoderLayer"></i><i class="c0" title="layer 10: LlamaDecoderLayer"></i><i class="c0" title="layer 11: LlamaDecoderLayer"></i><i class="c0" title="layer 12: LlamaDecoderLayer"></i><i class="c0" title="layer 13: LlamaDecoderLayer"></i><i class="c0" title="layer 14: LlamaDecoderLayer"></i><i class="c0" title="layer 15: LlamaDecoderLayer"></i><i class="c0" title="layer 16: LlamaDecoderLayer"></i><i class="c0" title="layer 17: LlamaDecoderLayer"></i><i class="c0" title="layer 18: LlamaDecoderLayer"></i><i class="c0" title="layer 19: LlamaDecoderLayer"></i><i class="c0" title="layer 20: LlamaDecoderLayer"></i><i class="c0" title="layer 21: LlamaDecoderLayer"></i><i class="c0" title="layer 22: LlamaDecoderLayer"></i><i class="c0" title="layer 23: LlamaDecoderLayer"></i><i class="c0" title="layer 24: LlamaDecoderLayer"></i><i class="c0" title="layer 25: LlamaDecoderLayer"></i><i class="c0" title="layer 26: LlamaDecoderLayer"></i><i class="c0" title="layer 27: LlamaDecoderLayer"></i><i class="c0" title="layer 28: LlamaDecoderLayer"></i><i class="c0" title="layer 29: LlamaDecoderLayer"></i><i class="c0" title="layer 30: LlamaDecoderLayer"></i><i class="c0" title="layer 31: LlamaDecoderLayer"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: LlamaDecoderLayer"></i><i class="c0" title="layer 1: LlamaDecoderLayer"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>LlamaDecoderLayer</td><td>32</td><td>2</td><td>✓ kept</td></tr></table></div><h2>Where the parameters are</h2><p class="note">Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.</p><div class="row"><span class="lab">Original</span><div class="bar"><i class="c0" style="flex:0.916873" title="model.language_model.layers: 6.5B (91.7%)"></i><i class="c1" style="flex:0.042799" title="model.vision_tower.encoder.layers: 302.3M (4.3%)"></i><i class="c2" style="flex:0.018594" title="model.language_model: 131.3M (1.9%)"></i><i class="c3" style="flex:0.018594" title="lm_head: 131.3M (1.9%)"></i><i class="c4" style="flex:0.002970" title="model.multi_modal_projector: 21.0M (0.3%)"></i><i class="c5" style="flex:0.000170" title="model.vision_tower: 1.2M (0.0%)"></i><i class="c6" style="flex:0.000001" title="model: 4.1K (0.0%)"></i></div></div><div class="row"><span class="lab">Fixture</span><div class="bar"><i class="c0" style="flex:0.018949" title="model.language_model.layers: 82.2K (1.9%)"></i><i class="c1" style="flex:0.015437" title="model.vision_tower.encoder.layers: 66.9K (1.5%)"></i><i class="c2" style="flex:0.473214" title="model.language_model: 2.1M (47.3%)"></i><i class="c3" style="flex:0.473200" title="lm_head: 2.1M (47.3%)"></i><i class="c4" style="flex:0.001919" title="model.multi_modal_projector: 8.3K (0.2%)"></i><i class="c5" style="flex:0.017267" title="model.vision_tower: 74.9K (1.7%)"></i><i class="c6" style="flex:0.000015" title="model: 64 (0.0%)"></i></div></div><table class="legend"><tr><th>Component</th><th>Original</th><th>Share</th><th>Fixture</th><th>Share</th></tr><tr><td><i class="c0"></i>model.language_model.layers</td><td>6.5B</td><td>91.7%</td><td>82.2K</td><td>1.9%</td></tr><tr><td><i class="c1"></i>model.vision_tower.encoder.layers</td><td>302.3M</td><td>4.3%</td><td>66.9K</td><td>1.5%</td></tr><tr><td><i class="c2"></i>model.language_model</td><td>131.3M</td><td>1.9%</td><td>2.1M</td><td>47.3%</td></tr><tr><td><i class="c3"></i>lm_head</td><td>131.3M</td><td>1.9%</td><td>2.1M</td><td>47.3%</td></tr><tr><td><i class="c4"></i>model.multi_modal_projector</td><td>21.0M</td><td>0.3%</td><td>8.3K</td><td>0.2%</td></tr><tr><td><i class="c5"></i>model.vision_tower</td><td>1.2M</td><td>0.0%</td><td>74.9K</td><td>1.7%</td></tr><tr><td><i class="c6"></i>model</td><td>4.1K</td><td>0.0%</td><td>64</td><td>0.0%</td></tr></table><h2>Preserved properties</h2><table><tr><th>Config</th><th>Property</th><th>Original</th><th>Fixture</th><th></th></tr><tr class=""><td>model</td><td>model_type</td><td>llava_next_video</td><td>llava_next_video</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>tie_word_embeddings</td><td>False</td><td>False</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>model_type</td><td>clip_vision_model</td><td>clip_vision_model</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>vocab_size</td><td>32000</td><td>32000</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>hidden_act</td><td>quick_gelu</td><td>quick_gelu</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>image_size</td><td>336</td><td>336</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>patch_size</td><td>14</td><td>14</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>num_channels</td><td>3</td><td>3</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>attention style</td><td>MHA (32 heads, 32 KV)</td><td>MHA (4 heads, 4 KV)</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>RoPE type</td><td>linear</td><td>linear</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>model_type</td><td>llama</td><td>llama</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>vocab_size</td><td>32064</td><td>32064</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>tie_word_embeddings</td><td>False</td><td>False</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>hidden_act</td><td>silu</td><td>silu</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>max_position_embeddings</td><td>4096</td><td>4096</td><td>✓ kept</td></tr></table><h2>Module classes</h2><p class="note">Instances of each class in the built model (prefix LlavaNextVideo omitted).</p><table><tr><th>Class</th><th>Original</th><th>Fixture</th></tr><tr class=""><td>Linear</td><td>371</td><td>29</td></tr><tr class=""><td>LlamaRMSNorm</td><td>65</td><td>5</td></tr><tr class=""><td>LayerNorm</td><td>50</td><td>6</td></tr><tr class=""><td>LlamaAttention</td><td>32</td><td>2</td></tr><tr class=""><td>LlamaDecoderLayer</td><td>32</td><td>2</td></tr><tr class=""><td>LlamaMLP</td><td>32</td><td>2</td></tr><tr class=""><td>SiLUActivation</td><td>32</td><td>2</td></tr><tr class=""><td>CLIPAttention</td><td>24</td><td>2</td></tr><tr class=""><td>CLIPEncoderLayer</td><td>24</td><td>2</td></tr><tr class=""><td>CLIPMLP</td><td>24</td><td>2</td></tr><tr class=""><td>QuickGELUActivation</td><td>24</td><td>2</td></tr><tr class=""><td>Embedding</td><td>2</td><td>2</td></tr><tr class=""><td>ModuleList</td><td>2</td><td>2</td></tr><tr class=""><td>AvgPool2d</td><td>1</td><td>1</td></tr><tr class=""><td>CLIPEncoder</td><td>1</td><td>1</td></tr><tr class=""><td>CLIPVisionEmbeddings</td><td>1</td><td>1</td></tr><tr class=""><td>CLIPVisionModel</td><td>1</td><td>1</td></tr><tr class=""><td>Conv2d</td><td>1</td><td>1</td></tr><tr class=""><td>GELUActivation</td><td>1</td><td>1</td></tr><tr class=""><td>LlamaModel</td><td>1</td><td>1</td></tr><tr class=""><td>LlamaRotaryEmbedding</td><td>1</td><td>1</td></tr><tr class=""><td>ForConditionalGeneration</td><td>1</td><td>1</td></tr><tr class=""><td>Model</td><td>1</td><td>1</td></tr><tr class=""><td>MultiModalProjector</td><td>1</td><td>1</td></tr><tr class=""><td>Pooler</td><td>1</td><td>1</td></tr></table><h2>Config changes</h2><p class="note">11 fields changed, 67 unchanged.</p><table><tr><th>Field</th><th>Original</th><th>Fixture</th></tr><tr><td>text_config.head_dim</td><td>128</td><td>16</td></tr><tr><td>text_config.hidden_size</td><td>4096</td><td>64</td></tr><tr><td>text_config.intermediate_size</td><td>11008</td><td>128</td></tr><tr><td>text_config.num_attention_heads</td><td>32</td><td>4</td></tr><tr><td>text_config.num_hidden_layers</td><td>32</td><td>2</td></tr><tr><td>text_config.num_key_value_heads</td><td>32</td><td>4</td></tr><tr><td>vision_config.hidden_size</td><td>1024</td><td>64</td></tr><tr><td>vision_config.intermediate_size</td><td>4096</td><td>128</td></tr><tr><td>vision_config.num_attention_heads</td><td>16</td><td>4</td></tr><tr><td>vision_config.num_hidden_layers</td><td>24</td><td>2</td></tr><tr><td>vision_config.projection_dim</td><td>768</td><td>32</td></tr></table>
+</main></body></html>

shrink/Laguna-XS.2-mic.html

@@ -0,0 +1,34 @@
+<!doctype html>
+<html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
+<title>Shrunk architecture</title>
+<style>
+:root{--bg:#fcfcfb;--fg:#0b0b0b;--muted:#52514e;--line:#e4e3df;--ok:#1d7a43;--warn:#b35c00;--s0:#2a78d6;--s1:#eb6834;--s2:#1baf7a;--s3:#eda100;--s4:#e87ba4;--s5:#008300;--s6:#6250d6;--s7:#e34948;--s8:#8a8984;}
+@media (prefers-color-scheme:dark){:root:not([data-theme="light"]){--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}}
+:root[data-theme="dark"]{--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}
+body{margin:0;background:var(--bg);color:var(--fg);font:15px/1.5 system-ui,sans-serif}
+main{max-width:1040px;margin:0 auto;padding:24px 16px 64px}
+h1{font-size:24px;margin:0 0 4px}h2{font-size:18px;margin:36px 0 8px}h3{font-size:14px;margin:0 0 8px;font-family:ui-monospace,monospace}
+.meta,.note{color:var(--muted);margin:0 0 12px;overflow-wrap:anywhere}
+.stats{display:grid;grid-template-columns:repeat(auto-fit,minmax(150px,1fr));gap:12px;margin:20px 0}
+.stat{border:1px solid var(--line);border-radius:8px;padding:12px;display:flex;flex-direction:column}
+.stat .v{font-size:26px;font-weight:600}.stat .k{color:var(--muted);font-size:13px}
+.findings{border-radius:8px;padding:4px 16px;border:1px solid var(--line)}
+.findings.ok{border-left:4px solid var(--ok)}.findings.warn{border-left:4px solid var(--warn)}
+.findings h2{margin:12px 0 4px;font-size:16px}
+.stack{border:1px solid var(--line);border-radius:8px;padding:12px;margin:12px 0}
+.row{display:flex;align-items:center;gap:12px;margin:6px 0}
+.lab{flex:0 0 110px;color:var(--muted);font-size:13px}
+.cells{display:flex;flex-wrap:wrap;gap:2px}
+.cells i{width:12px;height:12px;border-radius:2px;display:block}
+.bar{flex:1;display:flex;gap:2px;height:16px}.bar i{display:block;border-radius:2px;min-width:2px}
+table{border-collapse:collapse;width:100%;font-size:13px;margin:8px 0;display:block;overflow-x:auto}
+th,td{text-align:left;padding:4px 10px 4px 0;border-bottom:1px solid var(--line);vertical-align:top}
+th{color:var(--muted);font-weight:500}
+td:first-child{font-family:ui-monospace,monospace;overflow-wrap:anywhere}
+td i{display:inline-block;width:10px;height:10px;border-radius:2px;margin-right:6px}
+tr.bad td{color:var(--warn)}
+.c0{background:var(--s0)}.c1{background:var(--s1)}.c2{background:var(--s2)}.c3{background:var(--s3)}.c4{background:var(--s4)}.c5{background:var(--s5)}.c6{background:var(--s6)}.c7{background:var(--s7)}.c8{background:var(--s8)}
+</style></head><body><main>
+<h1>Shrunk architecture</h1><p class="meta">huggingface-mic/Laguna-XS.2-mic@21bbfb128fc8f6764e660376a3482ec8b6e8c93b ← poolside/Laguna-XS.2@a397bde04501591d0ac3d74a5d9d0a4ec298ffb8 · LagunaForCausalLM · random weights · built before the shrink rule (no recorded recipe)</p>
+<div class="stats"><div class="stat"><span class="k">Original</span><span class="v">33.4B</span><span class="k">parameters</span></div><div class="stat"><span class="k">Fixture</span><span class="v">13.2M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Layer stacks</span><span class="v">1</span><span class="k">compared</span></div><div class="stat"><span class="k">Findings</span><span class="v">0</span><span class="k">to review</span></div></div><section class="findings ok"><h2>✓ Same architecture, smaller</h2><p>Every module class, layer kind and checked property of the original is in the fixture.</p></section><h2>Layer stacks</h2><p class="note">One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.</p><div class="stack"><h3>model.layers</h3><div class="row"><span class="lab">Original · 40</span><div class="cells"><i class="c0" title="layer 0: no Experts"></i><i class="c1" title="layer 1: Experts, SparseMoeBlock, TopKRouter"></i><i class="c1" title="layer 2: Experts, SparseMoeBlock, TopKRouter"></i><i class="c1" title="layer 3: Experts, SparseMoeBlock, TopKRouter"></i><i class="c1" title="layer 4: Experts, SparseMoeBlock, TopKRouter"></i><i class="c1" title="layer 5: Experts, SparseMoeBlock, TopKRouter"></i><i class="c1" title="layer 6: Experts, SparseMoeBlock, TopKRouter"></i><i class="c1" title="layer 7: Experts, SparseMoeBlock, TopKRouter"></i><i class="c1" title="layer 8: Experts, SparseMoeBlock, TopKRouter"></i><i class="c1" title="layer 9: Experts, SparseMoeBlock, TopKRouter"></i><i class="c1" title="layer 10: Experts, SparseMoeBlock, TopKRouter"></i><i class="c1" title="layer 11: Experts, SparseMoeBlock, TopKRouter"></i><i class="c1" title="layer 12: Experts, SparseMoeBlock, TopKRouter"></i><i class="c1" title="layer 13: Experts, SparseMoeBlock, TopKRouter"></i><i class="c1" title="layer 14: Experts, SparseMoeBlock, TopKRouter"></i><i class="c1" title="layer 15: Experts, SparseMoeBlock, TopKRouter"></i><i class="c1" title="layer 16: Experts, SparseMoeBlock, TopKRouter"></i><i class="c1" title="layer 17: Experts, SparseMoeBlock, TopKRouter"></i><i class="c1" title="layer 18: Experts, SparseMoeBlock, TopKRouter"></i><i class="c1" title="layer 19: Experts, SparseMoeBlock, TopKRouter"></i><i class="c1" title="layer 20: Experts, SparseMoeBlock, TopKRouter"></i><i class="c1" title="layer 21: Experts, SparseMoeBlock, TopKRouter"></i><i class="c1" title="layer 22: Experts, SparseMoeBlock, TopKRouter"></i><i class="c1" title="layer 23: Experts, SparseMoeBlock, TopKRouter"></i><i class="c1" title="layer 24: Experts, SparseMoeBlock, TopKRouter"></i><i class="c1" title="layer 25: Experts, SparseMoeBlock, TopKRouter"></i><i class="c1" title="layer 26: Experts, SparseMoeBlock, TopKRouter"></i><i class="c1" title="layer 27: Experts, SparseMoeBlock, TopKRouter"></i><i class="c1" title="layer 28: Experts, SparseMoeBlock, TopKRouter"></i><i class="c1" title="layer 29: Experts, SparseMoeBlock, TopKRouter"></i><i class="c1" title="layer 30: Experts, SparseMoeBlock, TopKRouter"></i><i class="c1" title="layer 31: Experts, SparseMoeBlock, TopKRouter"></i><i class="c1" title="layer 32: Experts, SparseMoeBlock, TopKRouter"></i><i class="c1" title="layer 33: Experts, SparseMoeBlock, TopKRouter"></i><i class="c1" title="layer 34: Experts, SparseMoeBlock, TopKRouter"></i><i class="c1" title="layer 35: Experts, SparseMoeBlock, TopKRouter"></i><i class="c1" title="layer 36: Experts, SparseMoeBlock, TopKRouter"></i><i class="c1" title="layer 37: Experts, SparseMoeBlock, TopKRouter"></i><i class="c1" title="layer 38: Experts, SparseMoeBlock, TopKRouter"></i><i class="c1" title="layer 39: Experts, SparseMoeBlock, TopKRouter"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: no Experts"></i><i class="c1" title="layer 1: Experts, SparseMoeBlock, TopKRouter"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>no Experts</td><td>1</td><td>1</td><td>✓ kept</td></tr><tr><td><i class="c1"></i>Experts, SparseMoeBlock, TopKRouter</td><td>39</td><td>1</td><td>✓ kept</td></tr></table></div><h2>Where the parameters are</h2><p class="note">Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.</p><div class="row"><span class="lab">Original</span><div class="bar"><i class="c0" style="flex:0.987709" title="model.layers: 33.0B (98.8%)"></i><i class="c1" style="flex:0.006145" title="lm_head: 205.5M (0.6%)"></i><i class="c2" style="flex:0.006145" title="model.embed_tokens: 205.5M (0.6%)"></i><i class="c3" style="flex:0.000000" title="model.norm: 2.0K (0.0%)"></i></div></div><div class="row"><span class="lab">Fixture</span><div class="bar"><i class="c0" style="flex:0.024236" title="model.layers: 319.0K (2.4%)"></i><i class="c1" style="flex:0.487880" title="lm_head: 6.4M (48.8%)"></i><i class="c2" style="flex:0.487880" title="model.embed_tokens: 6.4M (48.8%)"></i><i class="c3" style="flex:0.000005" title="model.norm: 64 (0.0%)"></i></div></div><table class="legend"><tr><th>Component</th><th>Original</th><th>Share</th><th>Fixture</th><th>Share</th></tr><tr><td><i class="c0"></i>model.layers</td><td>33.0B</td><td>98.8%</td><td>319.0K</td><td>2.4%</td></tr><tr><td><i class="c1"></i>lm_head</td><td>205.5M</td><td>0.6%</td><td>6.4M</td><td>48.8%</td></tr><tr><td><i class="c2"></i>model.embed_tokens</td><td>205.5M</td><td>0.6%</td><td>6.4M</td><td>48.8%</td></tr><tr><td><i class="c3"></i>model.norm</td><td>2.0K</td><td>0.0%</td><td>64</td><td>0.0%</td></tr></table><h2>Preserved properties</h2><table><tr><th>Config</th><th>Property</th><th>Original</th><th>Fixture</th><th></th></tr><tr class=""><td>model</td><td>attention style</td><td>GQA (48 heads, 8 KV)</td><td>GQA (4 heads, 2 KV)</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>MoE routing</td><td>top-8 of 256 experts</td><td>top-2 of 4 experts</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>model_type</td><td>laguna</td><td>laguna</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>vocab_size</td><td>100352</td><td>100352</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>tie_word_embeddings</td><td>False</td><td>False</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>hidden_act</td><td>silu</td><td>silu</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>max_position_embeddings</td><td>262144</td><td>262144</td><td>✓ kept</td></tr></table><h2>Module classes</h2><p class="note">Instances of each class in the built model (prefix Laguna omitted).</p><table><tr><th>Class</th><th>Original</th><th>Fixture</th></tr><tr class=""><td>Linear</td><td>321</td><td>17</td></tr><tr class=""><td>RMSNorm</td><td>161</td><td>9</td></tr><tr class=""><td>SiLUActivation</td><td>79</td><td>3</td></tr><tr class=""><td>Attention</td><td>40</td><td>2</td></tr><tr class=""><td>DecoderLayer</td><td>40</td><td>2</td></tr><tr class=""><td>MLP</td><td>40</td><td>2</td></tr><tr class=""><td>Experts</td><td>39</td><td>1</td></tr><tr class=""><td>SparseMoeBlock</td><td>39</td><td>1</td></tr><tr class=""><td>TopKRouter</td><td>39</td><td>1</td></tr><tr class=""><td>Embedding</td><td>1</td><td>1</td></tr><tr class=""><td>ForCausalLM</td><td>1</td><td>1</td></tr><tr class=""><td>Model</td><td>1</td><td>1</td></tr><tr class=""><td>RotaryEmbedding</td><td>1</td><td>1</td></tr><tr class=""><td>ModuleList</td><td>1</td><td>1</td></tr></table><h2>Config changes</h2><p class="note">13 fields changed, 31 unchanged.</p><table><tr><th>Field</th><th>Original</th><th>Fixture</th></tr><tr><td>head_dim</td><td>128</td><td>16</td></tr><tr><td>hidden_size</td><td>2048</td><td>64</td></tr><tr><td>intermediate_size</td><td>8192</td><td>128</td></tr><tr><td>layer_types</td><td>40 entries: 30× sliding_attention, 10× full_attention</td><td>[&quot;full_attention&quot;, &quot;sliding_attention&quot;]</td></tr><tr><td>mlp_layer_types</td><td>40 entries: 39× sparse, 1× dense</td><td>[&quot;dense&quot;, &quot;sparse&quot;]</td></tr><tr><td>moe_intermediate_size</td><td>512</td><td>32</td></tr><tr><td>num_attention_heads</td><td>48</td><td>4</td></tr><tr><td>num_attention_heads_per_layer</td><td>40 entries: 30× 64, 10× 48</td><td>[48, 64]</td></tr><tr><td>num_experts</td><td>256</td><td>4</td></tr><tr><td>num_experts_per_tok</td><td>8</td><td>2</td></tr><tr><td>num_hidden_layers</td><td>40</td><td>2</td></tr><tr><td>num_key_value_heads</td><td>8</td><td>2</td></tr><tr><td>shared_expert_intermediate_size</td><td>512</td><td>128</td></tr></table>
+</main></body></html>

shrink/Llama-3.2-11B-Vision-Instruct-mic.html

@@ -0,0 +1,34 @@
+<!doctype html>
+<html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
+<title>Shrunk architecture</title>
+<style>
+:root{--bg:#fcfcfb;--fg:#0b0b0b;--muted:#52514e;--line:#e4e3df;--ok:#1d7a43;--warn:#b35c00;--s0:#2a78d6;--s1:#eb6834;--s2:#1baf7a;--s3:#eda100;--s4:#e87ba4;--s5:#008300;--s6:#6250d6;--s7:#e34948;--s8:#8a8984;}
+@media (prefers-color-scheme:dark){:root:not([data-theme="light"]){--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}}
+:root[data-theme="dark"]{--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}
+body{margin:0;background:var(--bg);color:var(--fg);font:15px/1.5 system-ui,sans-serif}
+main{max-width:1040px;margin:0 auto;padding:24px 16px 64px}
+h1{font-size:24px;margin:0 0 4px}h2{font-size:18px;margin:36px 0 8px}h3{font-size:14px;margin:0 0 8px;font-family:ui-monospace,monospace}
+.meta,.note{color:var(--muted);margin:0 0 12px;overflow-wrap:anywhere}
+.stats{display:grid;grid-template-columns:repeat(auto-fit,minmax(150px,1fr));gap:12px;margin:20px 0}
+.stat{border:1px solid var(--line);border-radius:8px;padding:12px;display:flex;flex-direction:column}
+.stat .v{font-size:26px;font-weight:600}.stat .k{color:var(--muted);font-size:13px}
+.findings{border-radius:8px;padding:4px 16px;border:1px solid var(--line)}
+.findings.ok{border-left:4px solid var(--ok)}.findings.warn{border-left:4px solid var(--warn)}
+.findings h2{margin:12px 0 4px;font-size:16px}
+.stack{border:1px solid var(--line);border-radius:8px;padding:12px;margin:12px 0}
+.row{display:flex;align-items:center;gap:12px;margin:6px 0}
+.lab{flex:0 0 110px;color:var(--muted);font-size:13px}
+.cells{display:flex;flex-wrap:wrap;gap:2px}
+.cells i{width:12px;height:12px;border-radius:2px;display:block}
+.bar{flex:1;display:flex;gap:2px;height:16px}.bar i{display:block;border-radius:2px;min-width:2px}
+table{border-collapse:collapse;width:100%;font-size:13px;margin:8px 0;display:block;overflow-x:auto}
+th,td{text-align:left;padding:4px 10px 4px 0;border-bottom:1px solid var(--line);vertical-align:top}
+th{color:var(--muted);font-weight:500}
+td:first-child{font-family:ui-monospace,monospace;overflow-wrap:anywhere}
+td i{display:inline-block;width:10px;height:10px;border-radius:2px;margin-right:6px}
+tr.bad td{color:var(--warn)}
+.c0{background:var(--s0)}.c1{background:var(--s1)}.c2{background:var(--s2)}.c3{background:var(--s3)}.c4{background:var(--s4)}.c5{background:var(--s5)}.c6{background:var(--s6)}.c7{background:var(--s7)}.c8{background:var(--s8)}
+</style></head><body><main>
+<h1>Shrunk architecture</h1><p class="meta">huggingface-mic/Llama-3.2-11B-Vision-Instruct-mic@8071268efa52824fc95b35bb2c1499946ec1bb23 ← unsloth/Llama-3.2-11B-Vision-Instruct@677b0c1b7008230a0fb88708c5550748e72b9a83 · MllamaForConditionalGeneration · overfit weights</p>
+<div class="stats"><div class="stat"><span class="k">Original</span><span class="v">10.7B</span><span class="k">parameters</span></div><div class="stat"><span class="k">Fixture</span><span class="v">20.7M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Layer stacks</span><span class="v">3</span><span class="k">compared</span></div><div class="stat"><span class="k">Findings</span><span class="v">0</span><span class="k">to review</span></div></div><section class="findings ok"><h2>✓ Same architecture, smaller</h2><p>Every module class, layer kind and checked property of the original is in the fixture.</p></section><h2>Layer stacks</h2><p class="note">One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.</p><div class="stack"><h3>model.vision_model.transformer.layers</h3><div class="row"><span class="lab">Original · 32</span><div class="cells"><i class="c0" title="layer 0: VisionEncoderLayer"></i><i class="c0" title="layer 1: VisionEncoderLayer"></i><i class="c0" title="layer 2: VisionEncoderLayer"></i><i class="c0" title="layer 3: VisionEncoderLayer"></i><i class="c0" title="layer 4: VisionEncoderLayer"></i><i class="c0" title="layer 5: VisionEncoderLayer"></i><i class="c0" title="layer 6: VisionEncoderLayer"></i><i class="c0" title="layer 7: VisionEncoderLayer"></i><i class="c0" title="layer 8: VisionEncoderLayer"></i><i class="c0" title="layer 9: VisionEncoderLayer"></i><i class="c0" title="layer 10: VisionEncoderLayer"></i><i class="c0" title="layer 11: VisionEncoderLayer"></i><i class="c0" title="layer 12: VisionEncoderLayer"></i><i class="c0" title="layer 13: VisionEncoderLayer"></i><i class="c0" title="layer 14: VisionEncoderLayer"></i><i class="c0" title="layer 15: VisionEncoderLayer"></i><i class="c0" title="layer 16: VisionEncoderLayer"></i><i class="c0" title="layer 17: VisionEncoderLayer"></i><i class="c0" title="layer 18: VisionEncoderLayer"></i><i class="c0" title="layer 19: VisionEncoderLayer"></i><i class="c0" title="layer 20: VisionEncoderLayer"></i><i class="c0" title="layer 21: VisionEncoderLayer"></i><i class="c0" title="layer 22: VisionEncoderLayer"></i><i class="c0" title="layer 23: VisionEncoderLayer"></i><i class="c0" title="layer 24: VisionEncoderLayer"></i><i class="c0" title="layer 25: VisionEncoderLayer"></i><i class="c0" title="layer 26: VisionEncoderLayer"></i><i class="c0" title="layer 27: VisionEncoderLayer"></i><i class="c0" title="layer 28: VisionEncoderLayer"></i><i class="c0" title="layer 29: VisionEncoderLayer"></i><i class="c0" title="layer 30: VisionEncoderLayer"></i><i class="c0" title="layer 31: VisionEncoderLayer"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: VisionEncoderLayer"></i><i class="c0" title="layer 1: VisionEncoderLayer"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>VisionEncoderLayer</td><td>32</td><td>2</td><td>✓ kept</td></tr></table></div><div class="stack"><h3>model.vision_model.global_transformer.layers</h3><div class="row"><span class="lab">Original · 8</span><div class="cells"><i class="c0" title="layer 0: VisionEncoderLayer"></i><i class="c0" title="layer 1: VisionEncoderLayer"></i><i class="c0" title="layer 2: VisionEncoderLayer"></i><i class="c0" title="layer 3: VisionEncoderLayer"></i><i class="c0" title="layer 4: VisionEncoderLayer"></i><i class="c0" title="layer 5: VisionEncoderLayer"></i><i class="c0" title="layer 6: VisionEncoderLayer"></i><i class="c0" title="layer 7: VisionEncoderLayer"></i></div></div><div class="row"><span class="lab">Fixture · 8</span><div class="cells"><i class="c0" title="layer 0: VisionEncoderLayer"></i><i class="c0" title="layer 1: VisionEncoderLayer"></i><i class="c0" title="layer 2: VisionEncoderLayer"></i><i class="c0" title="layer 3: VisionEncoderLayer"></i><i class="c0" title="layer 4: VisionEncoderLayer"></i><i class="c0" title="layer 5: VisionEncoderLayer"></i><i class="c0" title="layer 6: VisionEncoderLayer"></i><i class="c0" title="layer 7: VisionEncoderLayer"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>VisionEncoderLayer</td><td>8</td><td>8</td><td>✓ kept</td></tr></table></div><div class="stack"><h3>model.language_model.layers</h3><div class="row"><span class="lab">Original · 40</span><div class="cells"><i class="c0" title="layer 0: is_causal=True, SelfAttentionDecoderLayer, TextSelfAttention"></i><i class="c0" title="layer 1: is_causal=True, SelfAttentionDecoderLayer, TextSelfAttention"></i><i class="c0" title="layer 2: is_causal=True, SelfAttentionDecoderLayer, TextSelfAttention"></i><i class="c1" title="layer 3: CrossAttentionDecoderLayer, TextCrossAttention"></i><i class="c0" title="layer 4: is_causal=True, SelfAttentionDecoderLayer, TextSelfAttention"></i><i class="c0" title="layer 5: is_causal=True, SelfAttentionDecoderLayer, TextSelfAttention"></i><i class="c0" title="layer 6: is_causal=True, SelfAttentionDecoderLayer, TextSelfAttention"></i><i class="c0" title="layer 7: is_causal=True, SelfAttentionDecoderLayer, TextSelfAttention"></i><i class="c1" title="layer 8: CrossAttentionDecoderLayer, TextCrossAttention"></i><i class="c0" title="layer 9: is_causal=True, SelfAttentionDecoderLayer, TextSelfAttention"></i><i class="c0" title="layer 10: is_causal=True, SelfAttentionDecoderLayer, TextSelfAttention"></i><i class="c0" title="layer 11: is_causal=True, SelfAttentionDecoderLayer, TextSelfAttention"></i><i class="c0" title="layer 12: is_causal=True, SelfAttentionDecoderLayer, TextSelfAttention"></i><i class="c1" title="layer 13: CrossAttentionDecoderLayer, TextCrossAttention"></i><i class="c0" title="layer 14: is_causal=True, SelfAttentionDecoderLayer, TextSelfAttention"></i><i class="c0" title="layer 15: is_causal=True, SelfAttentionDecoderLayer, TextSelfAttention"></i><i class="c0" title="layer 16: is_causal=True, SelfAttentionDecoderLayer, TextSelfAttention"></i><i class="c0" title="layer 17: is_causal=True, SelfAttentionDecoderLayer, TextSelfAttention"></i><i class="c1" title="layer 18: CrossAttentionDecoderLayer, TextCrossAttention"></i><i class="c0" title="layer 19: is_causal=True, SelfAttentionDecoderLayer, TextSelfAttention"></i><i class="c0" title="layer 20: is_causal=True, SelfAttentionDecoderLayer, TextSelfAttention"></i><i class="c0" title="layer 21: is_causal=True, SelfAttentionDecoderLayer, TextSelfAttention"></i><i class="c0" title="layer 22: is_causal=True, SelfAttentionDecoderLayer, TextSelfAttention"></i><i class="c1" title="layer 23: CrossAttentionDecoderLayer, TextCrossAttention"></i><i class="c0" title="layer 24: is_causal=True, SelfAttentionDecoderLayer, TextSelfAttention"></i><i class="c0" title="layer 25: is_causal=True, SelfAttentionDecoderLayer, TextSelfAttention"></i><i class="c0" title="layer 26: is_causal=True, SelfAttentionDecoderLayer, TextSelfAttention"></i><i class="c0" title="layer 27: is_causal=True, SelfAttentionDecoderLayer, TextSelfAttention"></i><i class="c1" title="layer 28: CrossAttentionDecoderLayer, TextCrossAttention"></i><i class="c0" title="layer 29: is_causal=True, SelfAttentionDecoderLayer, TextSelfAttention"></i><i class="c0" title="layer 30: is_causal=True, SelfAttentionDecoderLayer, TextSelfAttention"></i><i class="c0" title="layer 31: is_causal=True, SelfAttentionDecoderLayer, TextSelfAttention"></i><i class="c0" title="layer 32: is_causal=True, SelfAttentionDecoderLayer, TextSelfAttention"></i><i class="c1" title="layer 33: CrossAttentionDecoderLayer, TextCrossAttention"></i><i class="c0" title="layer 34: is_causal=True, SelfAttentionDecoderLayer, TextSelfAttention"></i><i class="c0" title="layer 35: is_causal=True, SelfAttentionDecoderLayer, TextSelfAttention"></i><i class="c0" title="layer 36: is_causal=True, SelfAttentionDecoderLayer, TextSelfAttention"></i><i class="c0" title="layer 37: is_causal=True, SelfAttentionDecoderLayer, TextSelfAttention"></i><i class="c1" title="layer 38: CrossAttentionDecoderLayer, TextCrossAttention"></i><i class="c0" title="layer 39: is_causal=True, SelfAttentionDecoderLayer, TextSelfAttention"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: is_causal=True, SelfAttentionDecoderLayer, TextSelfAttention"></i><i class="c1" title="layer 1: CrossAttentionDecoderLayer, TextCrossAttention"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>is_causal=True, SelfAttentionDecoderLayer, TextSelfAttention</td><td>32</td><td>1</td><td>✓ kept</td></tr><tr><td><i class="c1"></i>CrossAttentionDecoderLayer, TextCrossAttention</td><td>8</td><td>1</td><td>✓ kept</td></tr></table></div><h2>Where the parameters are</h2><p class="note">Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.</p><div class="row"><span class="lab">Original</span><div class="bar"><i class="c0" style="flex:0.817648" title="model.language_model.layers: 8.7B (81.8%)"></i><i class="c1" style="flex:0.058997" title="model.vision_model.transformer.layers: 629.5M (5.9%)"></i><i class="c2" style="flex:0.049237" title="model.language_model: 525.4M (4.9%)"></i><i class="c3" style="flex:0.049234" title="lm_head: 525.3M (4.9%)"></i><i class="c4" style="flex:0.014749" title="model.vision_model.global_transformer.layers: 157.4M (1.5%)"></i><i class="c5" style="flex:0.007186" title="model.vision_model: 76.7M (0.7%)"></i><i class="c6" style="flex:0.002949" title="model.multi_modal_projector: 31.5M (0.3%)"></i></div></div><div class="row"><span class="lab">Fixture</span><div class="bar"><i class="c0" style="flex:0.003581" title="model.language_model.layers: 74.0K (0.4%)"></i><i class="c1" style="flex:0.003214" title="model.vision_model.transformer.layers: 66.4K (0.3%)"></i><i class="c2" style="flex:0.397151" title="model.language_model: 8.2M (39.7%)"></i><i class="c3" style="flex:0.397123" title="lm_head: 8.2M (39.7%)"></i><i class="c4" style="flex:0.012857" title="model.vision_model.global_transformer.layers: 265.7K (1.3%)"></i><i class="c5" style="flex:0.185477" title="model.vision_model: 3.8M (18.5%)"></i><i class="c6" style="flex:0.000598" title="model.multi_modal_projector: 12.4K (0.1%)"></i></div></div><table class="legend"><tr><th>Component</th><th>Original</th><th>Share</th><th>Fixture</th><th>Share</th></tr><tr><td><i class="c0"></i>model.language_model.layers</td><td>8.7B</td><td>81.8%</td><td>74.0K</td><td>0.4%</td></tr><tr><td><i class="c1"></i>model.vision_model.transformer.layers</td><td>629.5M</td><td>5.9%</td><td>66.4K</td><td>0.3%</td></tr><tr><td><i class="c2"></i>model.language_model</td><td>525.4M</td><td>4.9%</td><td>8.2M</td><td>39.7%</td></tr><tr><td><i class="c3"></i>lm_head</td><td>525.3M</td><td>4.9%</td><td>8.2M</td><td>39.7%</td></tr><tr><td><i class="c4"></i>model.vision_model.global_transformer.layers</td><td>157.4M</td><td>1.5%</td><td>265.7K</td><td>1.3%</td></tr><tr><td><i class="c5"></i>model.vision_model</td><td>76.7M</td><td>0.7%</td><td>3.8M</td><td>18.5%</td></tr><tr><td><i class="c6"></i>model.multi_modal_projector</td><td>31.5M</td><td>0.3%</td><td>12.4K</td><td>0.1%</td></tr></table><h2>Preserved properties</h2><table><tr><th>Config</th><th>Property</th><th>Original</th><th>Fixture</th><th></th></tr><tr class=""><td>model</td><td>model_type</td><td>mllama</td><td>mllama</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>model_type</td><td>mllama_vision_model</td><td>mllama_vision_model</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>tie_word_embeddings</td><td>True</td><td>True</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>hidden_act</td><td>gelu</td><td>gelu</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>image_size</td><td>560</td><td>560</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>patch_size</td><td>14</td><td>14</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>num_channels</td><td>3</td><td>3</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>attention style</td><td>GQA (32 heads, 8 KV)</td><td>GQA (4 heads, 2 KV)</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>RoPE type</td><td>llama3</td><td>llama3</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>model_type</td><td>mllama_text_model</td><td>mllama_text_model</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>vocab_size</td><td>128256</td><td>128256</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>tie_word_embeddings</td><td>False</td><td>False</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>hidden_act</td><td>silu</td><td>silu</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>max_position_embeddings</td><td>131072</td><td>131072</td><td>✓ kept</td></tr></table><h2>Module classes</h2><p class="note">Instances of each class in the built model (prefix Mllama omitted).</p><table><tr><th>Class</th><th>Original</th><th>Fixture</th></tr><tr class=""><td>Linear</td><td>522</td><td>76</td></tr><tr class=""><td>TextRMSNorm</td><td>97</td><td>7</td></tr><tr class=""><td>LayerNorm</td><td>82</td><td>22</td></tr><tr class=""><td>GELUActivation</td><td>40</td><td>10</td></tr><tr class=""><td>TextMLP</td><td>40</td><td>2</td></tr><tr class=""><td>VisionAttention</td><td>40</td><td>10</td></tr><tr class=""><td>VisionEncoderLayer</td><td>40</td><td>10</td></tr><tr class=""><td>VisionMLP</td><td>40</td><td>10</td></tr><tr class=""><td>SiLUActivation</td><td>40</td><td>2</td></tr><tr class=""><td>SelfAttentionDecoderLayer</td><td>32</td><td>1</td></tr><tr class=""><td>TextSelfAttention</td><td>32</td><td>1</td></tr><tr class=""><td>CrossAttentionDecoderLayer</td><td>8</td><td>1</td></tr><tr class=""><td>TextCrossAttention</td><td>8</td><td>1</td></tr><tr class=""><td>Embedding</td><td>4</td><td>4</td></tr><tr class=""><td>ModuleList</td><td>3</td><td>3</td></tr><tr class=""><td>PrecomputedAspectRatioEmbedding</td><td>2</td><td>2</td></tr><tr class=""><td>VisionEncoder</td><td>2</td><td>2</td></tr><tr class=""><td>Conv2d</td><td>1</td><td>1</td></tr><tr class=""><td>ForConditionalGeneration</td><td>1</td><td>1</td></tr><tr class=""><td>Model</td><td>1</td><td>1</td></tr><tr class=""><td>PrecomputedPositionEmbedding</td><td>1</td><td>1</td></tr><tr class=""><td>RotaryEmbedding</td><td>1</td><td>1</td></tr><tr class=""><td>TextModel</td><td>1</td><td>1</td></tr><tr class=""><td>VisionModel</td><td>1</td><td>1</td></tr></table><h2>Config changes</h2><p class="note">11 fields changed, 85 unchanged.</p><table><tr><th>Field</th><th>Original</th><th>Fixture</th></tr><tr><td>text_config.cross_attention_layers</td><td>8 entries</td><td>[1]</td></tr><tr><td>text_config.hidden_size</td><td>4096</td><td>64</td></tr><tr><td>text_config.intermediate_size</td><td>14336</td><td>128</td></tr><tr><td>text_config.num_attention_heads</td><td>32</td><td>4</td></tr><tr><td>text_config.num_hidden_layers</td><td>40</td><td>2</td></tr><tr><td>text_config.num_key_value_heads</td><td>8</td><td>2</td></tr><tr><td>vision_config.hidden_size</td><td>1280</td><td>64</td></tr><tr><td>vision_config.intermediate_layers_indices</td><td>[3, 7, 15, 23, 30]</td><td>[0, 1]</td></tr><tr><td>vision_config.intermediate_size</td><td>5120</td><td>128</td></tr><tr><td>vision_config.num_hidden_layers</td><td>32</td><td>2</td></tr><tr><td>vision_config.vision_output_dim</td><td>7680</td><td>192</td></tr></table>
+</main></body></html>

shrink/Llama-4-Scout-17B-16E-Instruct-mic.html

@@ -0,0 +1,34 @@
+<!doctype html>
+<html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
+<title>Shrunk architecture</title>
+<style>
+:root{--bg:#fcfcfb;--fg:#0b0b0b;--muted:#52514e;--line:#e4e3df;--ok:#1d7a43;--warn:#b35c00;--s0:#2a78d6;--s1:#eb6834;--s2:#1baf7a;--s3:#eda100;--s4:#e87ba4;--s5:#008300;--s6:#6250d6;--s7:#e34948;--s8:#8a8984;}
+@media (prefers-color-scheme:dark){:root:not([data-theme="light"]){--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}}
+:root[data-theme="dark"]{--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}
+body{margin:0;background:var(--bg);color:var(--fg);font:15px/1.5 system-ui,sans-serif}
+main{max-width:1040px;margin:0 auto;padding:24px 16px 64px}
+h1{font-size:24px;margin:0 0 4px}h2{font-size:18px;margin:36px 0 8px}h3{font-size:14px;margin:0 0 8px;font-family:ui-monospace,monospace}
+.meta,.note{color:var(--muted);margin:0 0 12px;overflow-wrap:anywhere}
+.stats{display:grid;grid-template-columns:repeat(auto-fit,minmax(150px,1fr));gap:12px;margin:20px 0}
+.stat{border:1px solid var(--line);border-radius:8px;padding:12px;display:flex;flex-direction:column}
+.stat .v{font-size:26px;font-weight:600}.stat .k{color:var(--muted);font-size:13px}
+.findings{border-radius:8px;padding:4px 16px;border:1px solid var(--line)}
+.findings.ok{border-left:4px solid var(--ok)}.findings.warn{border-left:4px solid var(--warn)}
+.findings h2{margin:12px 0 4px;font-size:16px}
+.stack{border:1px solid var(--line);border-radius:8px;padding:12px;margin:12px 0}
+.row{display:flex;align-items:center;gap:12px;margin:6px 0}
+.lab{flex:0 0 110px;color:var(--muted);font-size:13px}
+.cells{display:flex;flex-wrap:wrap;gap:2px}
+.cells i{width:12px;height:12px;border-radius:2px;display:block}
+.bar{flex:1;display:flex;gap:2px;height:16px}.bar i{display:block;border-radius:2px;min-width:2px}
+table{border-collapse:collapse;width:100%;font-size:13px;margin:8px 0;display:block;overflow-x:auto}
+th,td{text-align:left;padding:4px 10px 4px 0;border-bottom:1px solid var(--line);vertical-align:top}
+th{color:var(--muted);font-weight:500}
+td:first-child{font-family:ui-monospace,monospace;overflow-wrap:anywhere}
+td i{display:inline-block;width:10px;height:10px;border-radius:2px;margin-right:6px}
+tr.bad td{color:var(--warn)}
+.c0{background:var(--s0)}.c1{background:var(--s1)}.c2{background:var(--s2)}.c3{background:var(--s3)}.c4{background:var(--s4)}.c5{background:var(--s5)}.c6{background:var(--s6)}.c7{background:var(--s7)}.c8{background:var(--s8)}
+</style></head><body><main>
+<h1>Shrunk architecture</h1><p class="meta">huggingface-mic/Llama-4-Scout-17B-16E-Instruct-mic@18ed0f73576cf32d488b1b776bba7891007dcf12 ← RedHatAI/Llama-4-Scout-17B-16E-Instruct@5996646120a69f5f182866bb56a8f1573f367604 · Llama4ForConditionalGeneration · random weights</p>
+<div class="stats"><div class="stat"><span class="k">Original</span><span class="v">108.6B</span><span class="k">parameters</span></div><div class="stat"><span class="k">Fixture</span><span class="v">44.7M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Layer stacks</span><span class="v">2</span><span class="k">compared</span></div><div class="stat"><span class="k">Findings</span><span class="v">0</span><span class="k">to review</span></div></div><section class="findings ok"><h2>✓ Same architecture, smaller</h2><p>Every module class, layer kind and checked property of the original is in the fixture.</p></section><h2>Layer stacks</h2><p class="note">One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.</p><div class="stack"><h3>vision_model.model.layers</h3><div class="row"><span class="lab">Original · 34</span><div class="cells"><i class="c0" title="layer 0: VisionEncoderLayer"></i><i class="c0" title="layer 1: VisionEncoderLayer"></i><i class="c0" title="layer 2: VisionEncoderLayer"></i><i class="c0" title="layer 3: VisionEncoderLayer"></i><i class="c0" title="layer 4: VisionEncoderLayer"></i><i class="c0" title="layer 5: VisionEncoderLayer"></i><i class="c0" title="layer 6: VisionEncoderLayer"></i><i class="c0" title="layer 7: VisionEncoderLayer"></i><i class="c0" title="layer 8: VisionEncoderLayer"></i><i class="c0" title="layer 9: VisionEncoderLayer"></i><i class="c0" title="layer 10: VisionEncoderLayer"></i><i class="c0" title="layer 11: VisionEncoderLayer"></i><i class="c0" title="layer 12: VisionEncoderLayer"></i><i class="c0" title="layer 13: VisionEncoderLayer"></i><i class="c0" title="layer 14: VisionEncoderLayer"></i><i class="c0" title="layer 15: VisionEncoderLayer"></i><i class="c0" title="layer 16: VisionEncoderLayer"></i><i class="c0" title="layer 17: VisionEncoderLayer"></i><i class="c0" title="layer 18: VisionEncoderLayer"></i><i class="c0" title="layer 19: VisionEncoderLayer"></i><i class="c0" title="layer 20: VisionEncoderLayer"></i><i class="c0" title="layer 21: VisionEncoderLayer"></i><i class="c0" title="layer 22: VisionEncoderLayer"></i><i class="c0" title="layer 23: VisionEncoderLayer"></i><i class="c0" title="layer 24: VisionEncoderLayer"></i><i class="c0" title="layer 25: VisionEncoderLayer"></i><i class="c0" title="layer 26: VisionEncoderLayer"></i><i class="c0" title="layer 27: VisionEncoderLayer"></i><i class="c0" title="layer 28: VisionEncoderLayer"></i><i class="c0" title="layer 29: VisionEncoderLayer"></i><i class="c0" title="layer 30: VisionEncoderLayer"></i><i class="c0" title="layer 31: VisionEncoderLayer"></i><i class="c0" title="layer 32: VisionEncoderLayer"></i><i class="c0" title="layer 33: VisionEncoderLayer"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: VisionEncoderLayer"></i><i class="c0" title="layer 1: VisionEncoderLayer"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>VisionEncoderLayer</td><td>34</td><td>2</td><td>✓ kept</td></tr></table></div><div class="stack"><h3>language_model.model.layers</h3><div class="row"><span class="lab">Original · 48</span><div class="cells"><i class="c0" title="layer 0: TextL2Norm"></i><i class="c0" title="layer 1: TextL2Norm"></i><i class="c0" title="layer 2: TextL2Norm"></i><i class="c1" title="layer 3: no TextL2Norm"></i><i class="c0" title="layer 4: TextL2Norm"></i><i class="c0" title="layer 5: TextL2Norm"></i><i class="c0" title="layer 6: TextL2Norm"></i><i class="c1" title="layer 7: no TextL2Norm"></i><i class="c0" title="layer 8: TextL2Norm"></i><i class="c0" title="layer 9: TextL2Norm"></i><i class="c0" title="layer 10: TextL2Norm"></i><i class="c1" title="layer 11: no TextL2Norm"></i><i class="c0" title="layer 12: TextL2Norm"></i><i class="c0" title="layer 13: TextL2Norm"></i><i class="c0" title="layer 14: TextL2Norm"></i><i class="c1" title="layer 15: no TextL2Norm"></i><i class="c0" title="layer 16: TextL2Norm"></i><i class="c0" title="layer 17: TextL2Norm"></i><i class="c0" title="layer 18: TextL2Norm"></i><i class="c1" title="layer 19: no TextL2Norm"></i><i class="c0" title="layer 20: TextL2Norm"></i><i class="c0" title="layer 21: TextL2Norm"></i><i class="c0" title="layer 22: TextL2Norm"></i><i class="c1" title="layer 23: no TextL2Norm"></i><i class="c0" title="layer 24: TextL2Norm"></i><i class="c0" title="layer 25: TextL2Norm"></i><i class="c0" title="layer 26: TextL2Norm"></i><i class="c1" title="layer 27: no TextL2Norm"></i><i class="c0" title="layer 28: TextL2Norm"></i><i class="c0" title="layer 29: TextL2Norm"></i><i class="c0" title="layer 30: TextL2Norm"></i><i class="c1" title="layer 31: no TextL2Norm"></i><i class="c0" title="layer 32: TextL2Norm"></i><i class="c0" title="layer 33: TextL2Norm"></i><i class="c0" title="layer 34: TextL2Norm"></i><i class="c1" title="layer 35: no TextL2Norm"></i><i class="c0" title="layer 36: TextL2Norm"></i><i class="c0" title="layer 37: TextL2Norm"></i><i class="c0" title="layer 38: TextL2Norm"></i><i class="c1" title="layer 39: no TextL2Norm"></i><i class="c0" title="layer 40: TextL2Norm"></i><i class="c0" title="layer 41: TextL2Norm"></i><i class="c0" title="layer 42: TextL2Norm"></i><i class="c1" title="layer 43: no TextL2Norm"></i><i class="c0" title="layer 44: TextL2Norm"></i><i class="c0" title="layer 45: TextL2Norm"></i><i class="c0" title="layer 46: TextL2Norm"></i><i class="c1" title="layer 47: no TextL2Norm"></i></div></div><div class="row"><span class="lab">Fixture · 4</span><div class="cells"><i class="c0" title="layer 0: TextL2Norm"></i><i class="c0" title="layer 1: TextL2Norm"></i><i class="c0" title="layer 2: TextL2Norm"></i><i class="c1" title="layer 3: no TextL2Norm"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>TextL2Norm</td><td>36</td><td>3</td><td>✓ kept</td></tr><tr><td><i class="c1"></i>no TextL2Norm</td><td>12</td><td>1</td><td>✓ kept</td></tr></table></div><h2>Where the parameters are</h2><p class="note">Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.</p><div class="row"><span class="lab">Original</span><div class="bar"><i class="c0" style="flex:0.972930" title="language_model.model.layers: 105.7B (97.3%)"></i><i class="c1" style="flex:0.009522" title="language_model.model: 1.0B (1.0%)"></i><i class="c2" style="flex:0.009522" title="language_model.lm_head: 1.0B (1.0%)"></i><i class="c3" style="flex:0.007451" title="vision_model.model.layers: 809.5M (0.7%)"></i><i class="c4" style="flex:0.000367" title="vision_model.vision_adapter: 39.8M (0.0%)"></i><i class="c5" style="flex:0.000193" title="multi_modal_projector.linear_1: 21.0M (0.0%)"></i><i class="c6" style="flex:0.000008" title="vision_model.patch_embedding: 827.9K (0.0%)"></i><i class="c7" style="flex:0.000008" title="3 others: 819.5K (0.0%)"></i></div></div><div class="row"><span class="lab">Fixture</span><div class="bar"><i class="c0" style="flex:0.012139" title="language_model.model.layers: 542.2K (1.2%)"></i><i class="c1" style="flex:0.289500" title="language_model.model: 12.9M (28.9%)"></i><i class="c2" style="flex:0.289498" title="language_model.lm_head: 12.9M (28.9%)"></i><i class="c3" style="flex:0.002238" title="vision_model.model.layers: 100.0K (0.2%)"></i><i class="c4" style="flex:0.399080" title="vision_model.vision_adapter: 17.8M (39.9%)"></i><i class="c5" style="flex:0.005869" title="multi_modal_projector.linear_1: 262.1K (0.6%)"></i><i class="c6" style="flex:0.000842" title="vision_model.patch_embedding: 37.6K (0.1%)"></i><i class="c7" style="flex:0.000834" title="3 others: 37.2K (0.1%)"></i></div></div><table class="legend"><tr><th>Component</th><th>Original</th><th>Share</th><th>Fixture</th><th>Share</th></tr><tr><td><i class="c0"></i>language_model.model.layers</td><td>105.7B</td><td>97.3%</td><td>542.2K</td><td>1.2%</td></tr><tr><td><i class="c1"></i>language_model.model</td><td>1.0B</td><td>1.0%</td><td>12.9M</td><td>28.9%</td></tr><tr><td><i class="c2"></i>language_model.lm_head</td><td>1.0B</td><td>1.0%</td><td>12.9M</td><td>28.9%</td></tr><tr><td><i class="c3"></i>vision_model.model.layers</td><td>809.5M</td><td>0.7%</td><td>100.0K</td><td>0.2%</td></tr><tr><td><i class="c4"></i>vision_model.vision_adapter</td><td>39.8M</td><td>0.0%</td><td>17.8M</td><td>39.9%</td></tr><tr><td><i class="c5"></i>multi_modal_projector.linear_1</td><td>21.0M</td><td>0.0%</td><td>262.1K</td><td>0.6%</td></tr><tr><td><i class="c6"></i>vision_model.patch_embedding</td><td>827.9K</td><td>0.0%</td><td>37.6K</td><td>0.1%</td></tr><tr><td><i class="c7"></i>3 others</td><td>819.5K</td><td>0.0%</td><td>37.2K</td><td>0.1%</td></tr></table><h2>Preserved properties</h2><table><tr><th>Config</th><th>Property</th><th>Original</th><th>Fixture</th><th></th></tr><tr class=""><td>model</td><td>model_type</td><td>llama4</td><td>llama4</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>tie_word_embeddings</td><td>False</td><td>False</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>RoPE type</td><td>default</td><td>default</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>model_type</td><td>llama4_vision_model</td><td>llama4_vision_model</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>hidden_act</td><td>gelu</td><td>gelu</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>image_size</td><td>336</td><td>336</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>patch_size</td><td>14</td><td>14</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>num_channels</td><td>3</td><td>3</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>attention style</td><td>GQA (40 heads, 8 KV)</td><td>GQA (4 heads, 2 KV)</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>MoE routing</td><td>top-1 of 16 experts</td><td>top-1 of 4 experts</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>RoPE type</td><td>llama3</td><td>llama3</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>model_type</td><td>llama4_text</td><td>llama4_text</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>vocab_size</td><td>202048</td><td>202048</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>tie_word_embeddings</td><td>False</td><td>False</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>hidden_act</td><td>silu</td><td>silu</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>max_position_embeddings</td><td>10485760</td><td>10485760</td><td>✓ kept</td></tr></table><h2>Module classes</h2><p class="note">Instances of each class in the built model (prefix Llama4 omitted).</p><table><tr><th>Class</th><th>Original</th><th>Fixture</th></tr><tr class=""><td>Linear</td><td>545</td><td>45</td></tr><tr class=""><td>TextRMSNorm</td><td>97</td><td>9</td></tr><tr class=""><td>SiLUActivation</td><td>96</td><td>8</td></tr><tr class=""><td>LayerNorm</td><td>70</td><td>6</td></tr><tr class=""><td>Router</td><td>48</td><td>4</td></tr><tr class=""><td>TextAttention</td><td>48</td><td>4</td></tr><tr class=""><td>TextDecoderLayer</td><td>48</td><td>4</td></tr><tr class=""><td>TextExperts</td><td>48</td><td>4</td></tr><tr class=""><td>TextMLP</td><td>48</td><td>4</td></tr><tr class=""><td>TextMoe</td><td>48</td><td>4</td></tr><tr class=""><td>TextL2Norm</td><td>36</td><td>3</td></tr><tr class=""><td>GELU</td><td>35</td><td>3</td></tr><tr class=""><td>VisionAttention</td><td>34</td><td>2</td></tr><tr class=""><td>VisionEncoderLayer</td><td>34</td><td>2</td></tr><tr class=""><td>VisionMLP</td><td>34</td><td>2</td></tr><tr class=""><td>ModuleList</td><td>2</td><td>2</td></tr><tr class=""><td>Embedding</td><td>1</td><td>1</td></tr><tr class=""><td>ForCausalLM</td><td>1</td><td>1</td></tr><tr class=""><td>ForConditionalGeneration</td><td>1</td><td>1</td></tr><tr class=""><td>MultiModalProjector</td><td>1</td><td>1</td></tr><tr class=""><td>TextModel</td><td>1</td><td>1</td></tr><tr class=""><td>TextRotaryEmbedding</td><td>1</td><td>1</td></tr><tr class=""><td>UnfoldConvolution</td><td>1</td><td>1</td></tr><tr class=""><td>VisionEncoder</td><td>1</td><td>1</td></tr><tr class=""><td>VisionMLP2</td><td>1</td><td>1</td></tr><tr class=""><td>VisionModel</td><td>1</td><td>1</td></tr><tr class=""><td>VisionPixelShuffleMLP</td><td>1</td><td>1</td></tr><tr class=""><td>VisionRotaryEmbedding</td><td>1</td><td>1</td></tr><tr class=""><td>Unfold</td><td>1</td><td>1</td></tr></table><h2>Config changes</h2><p class="note">14 fields changed, 75 unchanged.</p><table><tr><th>Field</th><th>Original</th><th>Fixture</th></tr><tr><td>text_config.head_dim</td><td>128</td><td>16</td></tr><tr><td>text_config.hidden_size</td><td>5120</td><td>64</td></tr><tr><td>text_config.intermediate_size</td><td>8192</td><td>128</td></tr><tr><td>text_config.layer_types</td><td>48 entries: 36× chunked_attention, 12× full_attention</td><td>[&quot;chunked_attention&quot;, &quot;chunked_attention&quot;, &quot;chunked_attention&quot;, &quot;full_attenti…</td></tr><tr><td>text_config.moe_layers</td><td>48 entries</td><td>[0, 1, 2, 3]</td></tr><tr><td>text_config.no_rope_layers</td><td>48 entries: 36× 1, 12× 0</td><td>[1, 1, 1, 0]</td></tr><tr><td>text_config.num_attention_heads</td><td>40</td><td>4</td></tr><tr><td>text_config.num_hidden_layers</td><td>48</td><td>4</td></tr><tr><td>text_config.num_key_value_heads</td><td>8</td><td>2</td></tr><tr><td>text_config.num_local_experts</td><td>16</td><td>4</td></tr><tr><td>vision_config.hidden_size</td><td>1408</td><td>64</td></tr><tr><td>vision_config.intermediate_size</td><td>5632</td><td>256</td></tr><tr><td>vision_config.num_attention_heads</td><td>16</td><td>4</td></tr><tr><td>vision_config.num_hidden_layers</td><td>34</td><td>2</td></tr></table>
+</main></body></html>

shrink/MOSS-Transcribe-Diarize-HF-mic.html

@@ -0,0 +1,34 @@
+<!doctype html>
+<html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
+<title>Shrunk architecture</title>
+<style>
+:root{--bg:#fcfcfb;--fg:#0b0b0b;--muted:#52514e;--line:#e4e3df;--ok:#1d7a43;--warn:#b35c00;--s0:#2a78d6;--s1:#eb6834;--s2:#1baf7a;--s3:#eda100;--s4:#e87ba4;--s5:#008300;--s6:#6250d6;--s7:#e34948;--s8:#8a8984;}
+@media (prefers-color-scheme:dark){:root:not([data-theme="light"]){--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}}
+:root[data-theme="dark"]{--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}
+body{margin:0;background:var(--bg);color:var(--fg);font:15px/1.5 system-ui,sans-serif}
+main{max-width:1040px;margin:0 auto;padding:24px 16px 64px}
+h1{font-size:24px;margin:0 0 4px}h2{font-size:18px;margin:36px 0 8px}h3{font-size:14px;margin:0 0 8px;font-family:ui-monospace,monospace}
+.meta,.note{color:var(--muted);margin:0 0 12px;overflow-wrap:anywhere}
+.stats{display:grid;grid-template-columns:repeat(auto-fit,minmax(150px,1fr));gap:12px;margin:20px 0}
+.stat{border:1px solid var(--line);border-radius:8px;padding:12px;display:flex;flex-direction:column}
+.stat .v{font-size:26px;font-weight:600}.stat .k{color:var(--muted);font-size:13px}
+.findings{border-radius:8px;padding:4px 16px;border:1px solid var(--line)}
+.findings.ok{border-left:4px solid var(--ok)}.findings.warn{border-left:4px solid var(--warn)}
+.findings h2{margin:12px 0 4px;font-size:16px}
+.stack{border:1px solid var(--line);border-radius:8px;padding:12px;margin:12px 0}
+.row{display:flex;align-items:center;gap:12px;margin:6px 0}
+.lab{flex:0 0 110px;color:var(--muted);font-size:13px}
+.cells{display:flex;flex-wrap:wrap;gap:2px}
+.cells i{width:12px;height:12px;border-radius:2px;display:block}
+.bar{flex:1;display:flex;gap:2px;height:16px}.bar i{display:block;border-radius:2px;min-width:2px}
+table{border-collapse:collapse;width:100%;font-size:13px;margin:8px 0;display:block;overflow-x:auto}
+th,td{text-align:left;padding:4px 10px 4px 0;border-bottom:1px solid var(--line);vertical-align:top}
+th{color:var(--muted);font-weight:500}
+td:first-child{font-family:ui-monospace,monospace;overflow-wrap:anywhere}
+td i{display:inline-block;width:10px;height:10px;border-radius:2px;margin-right:6px}
+tr.bad td{color:var(--warn)}
+.c0{background:var(--s0)}.c1{background:var(--s1)}.c2{background:var(--s2)}.c3{background:var(--s3)}.c4{background:var(--s4)}.c5{background:var(--s5)}.c6{background:var(--s6)}.c7{background:var(--s7)}.c8{background:var(--s8)}
+</style></head><body><main>
+<h1>Shrunk architecture</h1><p class="meta">huggingface-mic/MOSS-Transcribe-Diarize-HF-mic@10010f1a0105941db4ce5363f4e724a85f053c64 ← itazap/MOSS-Transcribe-Diarize-HF@df4bdbe3c5376de5029005b69a301286b1bb7f35 · MossTranscribeDiarizeForConditionalGeneration · random weights</p>
+<div class="stats"><div class="stat"><span class="k">Original</span><span class="v">908.5M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Fixture</span><span class="v">10.0M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Layer stacks</span><span class="v">2</span><span class="k">compared</span></div><div class="stat"><span class="k">Findings</span><span class="v">0</span><span class="k">to review</span></div></div><section class="findings ok"><h2>✓ Same architecture, smaller</h2><p>Every module class, layer kind and checked property of the original is in the fixture.</p></section><h2>Layer stacks</h2><p class="note">One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.</p><div class="stack"><h3>model.audio_tower.layers</h3><div class="row"><span class="lab">Original · 24</span><div class="cells"><i class="c0" title="layer 0: AudioLayer"></i><i class="c0" title="layer 1: AudioLayer"></i><i class="c0" title="layer 2: AudioLayer"></i><i class="c0" title="layer 3: AudioLayer"></i><i class="c0" title="layer 4: AudioLayer"></i><i class="c0" title="layer 5: AudioLayer"></i><i class="c0" title="layer 6: AudioLayer"></i><i class="c0" title="layer 7: AudioLayer"></i><i class="c0" title="layer 8: AudioLayer"></i><i class="c0" title="layer 9: AudioLayer"></i><i class="c0" title="layer 10: AudioLayer"></i><i class="c0" title="layer 11: AudioLayer"></i><i class="c0" title="layer 12: AudioLayer"></i><i class="c0" title="layer 13: AudioLayer"></i><i class="c0" title="layer 14: AudioLayer"></i><i class="c0" title="layer 15: AudioLayer"></i><i class="c0" title="layer 16: AudioLayer"></i><i class="c0" title="layer 17: AudioLayer"></i><i class="c0" title="layer 18: AudioLayer"></i><i class="c0" title="layer 19: AudioLayer"></i><i class="c0" title="layer 20: AudioLayer"></i><i class="c0" title="layer 21: AudioLayer"></i><i class="c0" title="layer 22: AudioLayer"></i><i class="c0" title="layer 23: AudioLayer"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: AudioLayer"></i><i class="c0" title="layer 1: AudioLayer"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>AudioLayer</td><td>24</td><td>2</td><td>✓ kept</td></tr></table></div><div class="stack"><h3>model.language_model.layers</h3><div class="row"><span class="lab">Original · 28</span><div class="cells"><i class="c0" title="layer 0: Qwen3DecoderLayer"></i><i class="c0" title="layer 1: Qwen3DecoderLayer"></i><i class="c0" title="layer 2: Qwen3DecoderLayer"></i><i class="c0" title="layer 3: Qwen3DecoderLayer"></i><i class="c0" title="layer 4: Qwen3DecoderLayer"></i><i class="c0" title="layer 5: Qwen3DecoderLayer"></i><i class="c0" title="layer 6: Qwen3DecoderLayer"></i><i class="c0" title="layer 7: Qwen3DecoderLayer"></i><i class="c0" title="layer 8: Qwen3DecoderLayer"></i><i class="c0" title="layer 9: Qwen3DecoderLayer"></i><i class="c0" title="layer 10: Qwen3DecoderLayer"></i><i class="c0" title="layer 11: Qwen3DecoderLayer"></i><i class="c0" title="layer 12: Qwen3DecoderLayer"></i><i class="c0" title="layer 13: Qwen3DecoderLayer"></i><i class="c0" title="layer 14: Qwen3DecoderLayer"></i><i class="c0" title="layer 15: Qwen3DecoderLayer"></i><i class="c0" title="layer 16: Qwen3DecoderLayer"></i><i class="c0" title="layer 17: Qwen3DecoderLayer"></i><i class="c0" title="layer 18: Qwen3DecoderLayer"></i><i class="c0" title="layer 19: Qwen3DecoderLayer"></i><i class="c0" title="layer 20: Qwen3DecoderLayer"></i><i class="c0" title="layer 21: Qwen3DecoderLayer"></i><i class="c0" title="layer 22: Qwen3DecoderLayer"></i><i class="c0" title="layer 23: Qwen3DecoderLayer"></i><i class="c0" title="layer 24: Qwen3DecoderLayer"></i><i class="c0" title="layer 25: Qwen3DecoderLayer"></i><i class="c0" title="layer 26: Qwen3DecoderLayer"></i><i class="c0" title="layer 27: Qwen3DecoderLayer"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: Qwen3DecoderLayer"></i><i class="c0" title="layer 1: Qwen3DecoderLayer"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>Qwen3DecoderLayer</td><td>28</td><td>2</td><td>✓ kept</td></tr></table></div><h2>Where the parameters are</h2><p class="note">Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.</p><div class="row"><span class="lab">Original</span><div class="bar"><i class="c0" style="flex:0.484821" title="model.language_model.layers: 440.5M (48.5%)"></i><i class="c1" style="flex:0.332725" title="model.audio_tower.layers: 302.3M (33.3%)"></i><i class="c2" style="flex:0.171251" title="model.language_model: 155.6M (17.1%)"></i><i class="c3" style="flex:0.005775" title="model.multi_modal_projector: 5.2M (0.6%)"></i><i class="c4" style="flex:0.005428" title="model.audio_tower: 4.9M (0.5%)"></i></div></div><div class="row"><span class="lab">Fixture</span><div class="bar"><i class="c0" style="flex:0.007398" title="model.language_model.layers: 74.0K (0.7%)"></i><i class="c1" style="flex:0.006675" title="model.audio_tower.layers: 66.8K (0.7%)"></i><i class="c2" style="flex:0.971477" title="model.language_model: 9.7M (97.1%)"></i><i class="c3" style="flex:0.002072" title="model.multi_modal_projector: 20.7K (0.2%)"></i><i class="c4" style="flex:0.012379" title="model.audio_tower: 123.9K (1.2%)"></i></div></div><table class="legend"><tr><th>Component</th><th>Original</th><th>Share</th><th>Fixture</th><th>Share</th></tr><tr><td><i class="c0"></i>model.language_model.layers</td><td>440.5M</td><td>48.5%</td><td>74.0K</td><td>0.7%</td></tr><tr><td><i class="c1"></i>model.audio_tower.layers</td><td>302.3M</td><td>33.3%</td><td>66.8K</td><td>0.7%</td></tr><tr><td><i class="c2"></i>model.language_model</td><td>155.6M</td><td>17.1%</td><td>9.7M</td><td>97.1%</td></tr><tr><td><i class="c3"></i>model.multi_modal_projector</td><td>5.2M</td><td>0.6%</td><td>20.7K</td><td>0.2%</td></tr><tr><td><i class="c4"></i>model.audio_tower</td><td>4.9M</td><td>0.5%</td><td>123.9K</td><td>1.2%</td></tr></table><h2>Preserved properties</h2><table><tr><th>Config</th><th>Property</th><th>Original</th><th>Fixture</th><th></th></tr><tr class=""><td>model</td><td>model_type</td><td>moss_transcribe_diarize</td><td>moss_transcribe_diarize</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>tie_word_embeddings</td><td>True</td><td>True</td><td>✓ kept</td></tr><tr class=""><td>audio_config</td><td>model_type</td><td>moss_transcribe_diarize_audio</td><td>moss_transcribe_diarize_audio</td><td>✓ kept</td></tr><tr class=""><td>audio_config</td><td>num_mel_bins</td><td>80</td><td>80</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>attention style</td><td>GQA (16 heads, 8 KV)</td><td>GQA (4 heads, 2 KV)</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>RoPE type</td><td>default</td><td>default</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>model_type</td><td>qwen3</td><td>qwen3</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>vocab_size</td><td>151936</td><td>151936</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>tie_word_embeddings</td><td>True</td><td>True</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>hidden_act</td><td>silu</td><td>silu</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>max_position_embeddings</td><td>131072</td><td>131072</td><td>✓ kept</td></tr></table><h2>Module classes</h2><p class="note">Instances of each class in the built model (prefix MossTranscribeDiarize omitted).</p><table><tr><th>Class</th><th>Original</th><th>Fixture</th></tr><tr class=""><td>Linear</td><td>343</td><td>29</td></tr><tr class=""><td>Qwen3RMSNorm</td><td>113</td><td>9</td></tr><tr class=""><td>LayerNorm</td><td>50</td><td>6</td></tr><tr class=""><td>SiLUActivation</td><td>29</td><td>3</td></tr><tr class=""><td>Qwen3Attention</td><td>28</td><td>2</td></tr><tr class=""><td>Qwen3DecoderLayer</td><td>28</td><td>2</td></tr><tr class=""><td>Qwen3MLP</td><td>28</td><td>2</td></tr><tr class=""><td>GELUActivation</td><td>24</td><td>2</td></tr><tr class=""><td>AudioAttention</td><td>24</td><td>2</td></tr><tr class=""><td>AudioLayer</td><td>24</td><td>2</td></tr><tr class=""><td>Conv1d</td><td>2</td><td>2</td></tr><tr class=""><td>Embedding</td><td>2</td><td>2</td></tr><tr class=""><td>ModuleList</td><td>2</td><td>2</td></tr><tr class=""><td>AudioModel</td><td>1</td><td>1</td></tr><tr class=""><td>ForConditionalGeneration</td><td>1</td><td>1</td></tr><tr class=""><td>Model</td><td>1</td><td>1</td></tr><tr class=""><td>MultiModalProjector</td><td>1</td><td>1</td></tr><tr class=""><td>Qwen3Model</td><td>1</td><td>1</td></tr><tr class=""><td>Qwen3RotaryEmbedding</td><td>1</td><td>1</td></tr></table><h2>Config changes</h2><p class="note">11 fields changed, 60 unchanged.</p><table><tr><th>Field</th><th>Original</th><th>Fixture</th></tr><tr><td>audio_config.hidden_size</td><td>1024</td><td>64</td></tr><tr><td>audio_config.intermediate_size</td><td>4096</td><td>128</td></tr><tr><td>audio_config.num_attention_heads</td><td>16</td><td>4</td></tr><tr><td>audio_config.num_hidden_layers</td><td>24</td><td>2</td></tr><tr><td>text_config.head_dim</td><td>128</td><td>16</td></tr><tr><td>text_config.hidden_size</td><td>1024</td><td>64</td></tr><tr><td>text_config.intermediate_size</td><td>3072</td><td>128</td></tr><tr><td>text_config.layer_types</td><td>28 entries: 28× full_attention</td><td>[&quot;full_attention&quot;, &quot;full_attention&quot;]</td></tr><tr><td>text_config.num_attention_heads</td><td>16</td><td>4</td></tr><tr><td>text_config.num_hidden_layers</td><td>28</td><td>2</td></tr><tr><td>text_config.num_key_value_heads</td><td>8</td><td>2</td></tr></table>
+</main></body></html>

shrink/MiniCPM-V-4.6-mic.html

@@ -0,0 +1,34 @@
+<!doctype html>
+<html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
+<title>Shrunk architecture</title>
+<style>
+:root{--bg:#fcfcfb;--fg:#0b0b0b;--muted:#52514e;--line:#e4e3df;--ok:#1d7a43;--warn:#b35c00;--s0:#2a78d6;--s1:#eb6834;--s2:#1baf7a;--s3:#eda100;--s4:#e87ba4;--s5:#008300;--s6:#6250d6;--s7:#e34948;--s8:#8a8984;}
+@media (prefers-color-scheme:dark){:root:not([data-theme="light"]){--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}}
+:root[data-theme="dark"]{--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}
+body{margin:0;background:var(--bg);color:var(--fg);font:15px/1.5 system-ui,sans-serif}
+main{max-width:1040px;margin:0 auto;padding:24px 16px 64px}
+h1{font-size:24px;margin:0 0 4px}h2{font-size:18px;margin:36px 0 8px}h3{font-size:14px;margin:0 0 8px;font-family:ui-monospace,monospace}
+.meta,.note{color:var(--muted);margin:0 0 12px;overflow-wrap:anywhere}
+.stats{display:grid;grid-template-columns:repeat(auto-fit,minmax(150px,1fr));gap:12px;margin:20px 0}
+.stat{border:1px solid var(--line);border-radius:8px;padding:12px;display:flex;flex-direction:column}
+.stat .v{font-size:26px;font-weight:600}.stat .k{color:var(--muted);font-size:13px}
+.findings{border-radius:8px;padding:4px 16px;border:1px solid var(--line)}
+.findings.ok{border-left:4px solid var(--ok)}.findings.warn{border-left:4px solid var(--warn)}
+.findings h2{margin:12px 0 4px;font-size:16px}
+.stack{border:1px solid var(--line);border-radius:8px;padding:12px;margin:12px 0}
+.row{display:flex;align-items:center;gap:12px;margin:6px 0}
+.lab{flex:0 0 110px;color:var(--muted);font-size:13px}
+.cells{display:flex;flex-wrap:wrap;gap:2px}
+.cells i{width:12px;height:12px;border-radius:2px;display:block}
+.bar{flex:1;display:flex;gap:2px;height:16px}.bar i{display:block;border-radius:2px;min-width:2px}
+table{border-collapse:collapse;width:100%;font-size:13px;margin:8px 0;display:block;overflow-x:auto}
+th,td{text-align:left;padding:4px 10px 4px 0;border-bottom:1px solid var(--line);vertical-align:top}
+th{color:var(--muted);font-weight:500}
+td:first-child{font-family:ui-monospace,monospace;overflow-wrap:anywhere}
+td i{display:inline-block;width:10px;height:10px;border-radius:2px;margin-right:6px}
+tr.bad td{color:var(--warn)}
+.c0{background:var(--s0)}.c1{background:var(--s1)}.c2{background:var(--s2)}.c3{background:var(--s3)}.c4{background:var(--s4)}.c5{background:var(--s5)}.c6{background:var(--s6)}.c7{background:var(--s7)}.c8{background:var(--s8)}
+</style></head><body><main>
+<h1>Shrunk architecture</h1><p class="meta">huggingface-mic/MiniCPM-V-4.6-mic@0ab6a30ad4aeae45c8ad56b7859397c24bca47e8 ← openbmb/MiniCPM-V-4.6@36f34a661a4bd35d0dc2294cb044d2584646c7d3 · MiniCPMV4_6ForConditionalGeneration · random weights</p>
+<div class="stats"><div class="stat"><span class="k">Original</span><span class="v">1.3B</span><span class="k">parameters</span></div><div class="stat"><span class="k">Fixture</span><span class="v">17.3M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Layer stacks</span><span class="v">3</span><span class="k">compared</span></div><div class="stat"><span class="k">Findings</span><span class="v">0</span><span class="k">to review</span></div></div><section class="findings ok"><h2>✓ Same architecture, smaller</h2><p>Every module class, layer kind and checked property of the original is in the fixture.</p></section><h2>Layer stacks</h2><p class="note">One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.</p><div class="stack"><h3>model.vision_tower.encoder.layers</h3><div class="row"><span class="lab">Original · 27</span><div class="cells"><i class="c0" title="layer 0: VisionEncoderLayer"></i><i class="c0" title="layer 1: VisionEncoderLayer"></i><i class="c0" title="layer 2: VisionEncoderLayer"></i><i class="c0" title="layer 3: VisionEncoderLayer"></i><i class="c0" title="layer 4: VisionEncoderLayer"></i><i class="c0" title="layer 5: VisionEncoderLayer"></i><i class="c0" title="layer 6: VisionEncoderLayer"></i><i class="c0" title="layer 7: VisionEncoderLayer"></i><i class="c0" title="layer 8: VisionEncoderLayer"></i><i class="c0" title="layer 9: VisionEncoderLayer"></i><i class="c0" title="layer 10: VisionEncoderLayer"></i><i class="c0" title="layer 11: VisionEncoderLayer"></i><i class="c0" title="layer 12: VisionEncoderLayer"></i><i class="c0" title="layer 13: VisionEncoderLayer"></i><i class="c0" title="layer 14: VisionEncoderLayer"></i><i class="c0" title="layer 15: VisionEncoderLayer"></i><i class="c0" title="layer 16: VisionEncoderLayer"></i><i class="c0" title="layer 17: VisionEncoderLayer"></i><i class="c0" title="layer 18: VisionEncoderLayer"></i><i class="c0" title="layer 19: VisionEncoderLayer"></i><i class="c0" title="layer 20: VisionEncoderLayer"></i><i class="c0" title="layer 21: VisionEncoderLayer"></i><i class="c0" title="layer 22: VisionEncoderLayer"></i><i class="c0" title="layer 23: VisionEncoderLayer"></i><i class="c0" title="layer 24: VisionEncoderLayer"></i><i class="c0" title="layer 25: VisionEncoderLayer"></i><i class="c0" title="layer 26: VisionEncoderLayer"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: VisionEncoderLayer"></i><i class="c0" title="layer 1: VisionEncoderLayer"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>VisionEncoderLayer</td><td>27</td><td>2</td><td>✓ kept</td></tr></table></div><div class="stack"><h3>model.language_model.layers</h3><div class="row"><span class="lab">Original · 24</span><div class="cells"><i class="c0" title="layer 0: linear_attention, Conv1d, Qwen3_5GatedDeltaNet …"></i><i class="c0" title="layer 1: linear_attention, Conv1d, Qwen3_5GatedDeltaNet …"></i><i class="c0" title="layer 2: linear_attention, Conv1d, Qwen3_5GatedDeltaNet …"></i><i class="c1" title="layer 3: is_causal=True, Qwen3_5Attention"></i><i class="c0" title="layer 4: linear_attention, Conv1d, Qwen3_5GatedDeltaNet …"></i><i class="c0" title="layer 5: linear_attention, Conv1d, Qwen3_5GatedDeltaNet …"></i><i class="c0" title="layer 6: linear_attention, Conv1d, Qwen3_5GatedDeltaNet …"></i><i class="c1" title="layer 7: is_causal=True, Qwen3_5Attention"></i><i class="c0" title="layer 8: linear_attention, Conv1d, Qwen3_5GatedDeltaNet …"></i><i class="c0" title="layer 9: linear_attention, Conv1d, Qwen3_5GatedDeltaNet …"></i><i class="c0" title="layer 10: linear_attention, Conv1d, Qwen3_5GatedDeltaNet …"></i><i class="c1" title="layer 11: is_causal=True, Qwen3_5Attention"></i><i class="c0" title="layer 12: linear_attention, Conv1d, Qwen3_5GatedDeltaNet …"></i><i class="c0" title="layer 13: linear_attention, Conv1d, Qwen3_5GatedDeltaNet …"></i><i class="c0" title="layer 14: linear_attention, Conv1d, Qwen3_5GatedDeltaNet …"></i><i class="c1" title="layer 15: is_causal=True, Qwen3_5Attention"></i><i class="c0" title="layer 16: linear_attention, Conv1d, Qwen3_5GatedDeltaNet …"></i><i class="c0" title="layer 17: linear_attention, Conv1d, Qwen3_5GatedDeltaNet …"></i><i class="c0" title="layer 18: linear_attention, Conv1d, Qwen3_5GatedDeltaNet …"></i><i class="c1" title="layer 19: is_causal=True, Qwen3_5Attention"></i><i class="c0" title="layer 20: linear_attention, Conv1d, Qwen3_5GatedDeltaNet …"></i><i class="c0" title="layer 21: linear_attention, Conv1d, Qwen3_5GatedDeltaNet …"></i><i class="c0" title="layer 22: linear_attention, Conv1d, Qwen3_5GatedDeltaNet …"></i><i class="c1" title="layer 23: is_causal=True, Qwen3_5Attention"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: linear_attention, Conv1d, Qwen3_5GatedDeltaNet …"></i><i class="c1" title="layer 1: is_causal=True, Qwen3_5Attention"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>linear_attention, Conv1d, Qwen3_5GatedDeltaNet …</td><td>18</td><td>1</td><td>✓ kept</td></tr><tr><td><i class="c1"></i>is_causal=True, Qwen3_5Attention</td><td>6</td><td>1</td><td>✓ kept</td></tr></table></div><div class="stack"><h3>model.merger.mlp</h3><div class="row"><span class="lab">Original · 1</span><div class="cells"><i class="c0" title="layer 0: GELU"></i></div></div><div class="row"><span class="lab">Fixture · 1</span><div class="cells"><i class="c0" title="layer 0: GELU"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>GELU</td><td>1</td><td>1</td><td>✓ kept</td></tr></table></div><h2>Where the parameters are</h2><p class="note">Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.</p><div class="row"><span class="lab">Original</span><div class="bar"><i class="c0" style="flex:0.383037" title="model.language_model.layers: 498.1M (38.3%)"></i><i class="c1" style="flex:0.316409" title="model.vision_tower.encoder.layers: 411.5M (31.6%)"></i><i class="c2" style="flex:0.195358" title="model.language_model: 254.0M (19.5%)"></i><i class="c3" style="flex:0.085228" title="model.vision_tower: 110.8M (8.5%)"></i><i class="c4" style="flex:0.019968" title="model.merger.mlp: 26.0M (2.0%)"></i></div></div><div class="row"><span class="lab">Fixture</span><div class="bar"><i class="c0" style="flex:0.043213" title="model.language_model.layers: 748.0K (4.3%)"></i><i class="c1" style="flex:0.003868" title="model.vision_tower.encoder.layers: 66.9K (0.4%)"></i><i class="c2" style="flex:0.917338" title="model.language_model: 15.9M (91.7%)"></i><i class="c3" style="flex:0.030800" title="model.vision_tower: 533.1K (3.1%)"></i><i class="c4" style="flex:0.004781" title="model.merger.mlp: 82.8K (0.5%)"></i></div></div><table class="legend"><tr><th>Component</th><th>Original</th><th>Share</th><th>Fixture</th><th>Share</th></tr><tr><td><i class="c0"></i>model.language_model.layers</td><td>498.1M</td><td>38.3%</td><td>748.0K</td><td>4.3%</td></tr><tr><td><i class="c1"></i>model.vision_tower.encoder.layers</td><td>411.5M</td><td>31.6%</td><td>66.9K</td><td>0.4%</td></tr><tr><td><i class="c2"></i>model.language_model</td><td>254.0M</td><td>19.5%</td><td>15.9M</td><td>91.7%</td></tr><tr><td><i class="c3"></i>model.vision_tower</td><td>110.8M</td><td>8.5%</td><td>533.1K</td><td>3.1%</td></tr><tr><td><i class="c4"></i>model.merger.mlp</td><td>26.0M</td><td>2.0%</td><td>82.8K</td><td>0.5%</td></tr></table><h2>Preserved properties</h2><table><tr><th>Config</th><th>Property</th><th>Original</th><th>Fixture</th><th></th></tr><tr class=""><td>model</td><td>model_type</td><td>minicpmv4_6</td><td>minicpmv4_6</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>tie_word_embeddings</td><td>True</td><td>True</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>image_size</td><td>1120</td><td>1120</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>patch_size</td><td>14</td><td>14</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>attention style</td><td>GQA (8 heads, 2 KV)</td><td>GQA (4 heads, 2 KV)</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>RoPE type</td><td>default</td><td>default</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>model_type</td><td>qwen3_5_text</td><td>qwen3_5_text</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>vocab_size</td><td>248094</td><td>248094</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>tie_word_embeddings</td><td>True</td><td>True</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>hidden_act</td><td>silu</td><td>silu</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>max_position_embeddings</td><td>262144</td><td>262144</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>model_type</td><td>minicpmv4_6_vision</td><td>minicpmv4_6_vision</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>hidden_act</td><td>gelu_pytorch_tanh</td><td>gelu_pytorch_tanh</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>image_size</td><td>980</td><td>980</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>patch_size</td><td>14</td><td>14</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>num_channels</td><td>3</td><td>3</td><td>✓ kept</td></tr></table><h2>Module classes</h2><p class="note">Instances of each class in the built model (prefix MiniCPMV4_6 omitted).</p><table><tr><th>Class</th><th>Original</th><th>Fixture</th></tr><tr class=""><td>Linear</td><td>357</td><td>36</td></tr><tr class=""><td>Qwen3_5RMSNorm</td><td>61</td><td>7</td></tr><tr class=""><td>LayerNorm</td><td>58</td><td>8</td></tr><tr class=""><td>GELUTanh</td><td>28</td><td>3</td></tr><tr class=""><td>VisionAttention</td><td>28</td><td>3</td></tr><tr class=""><td>VisionEncoderLayer</td><td>27</td><td>2</td></tr><tr class=""><td>VisionMLP</td><td>27</td><td>2</td></tr><tr class=""><td>Qwen3_5DecoderLayer</td><td>24</td><td>2</td></tr><tr class=""><td>Qwen3_5MLP</td><td>24</td><td>2</td></tr><tr class=""><td>SiLUActivation</td><td>24</td><td>2</td></tr><tr class=""><td>Conv1d</td><td>18</td><td>1</td></tr><tr class=""><td>Qwen3_5GatedDeltaNet</td><td>18</td><td>1</td></tr><tr class=""><td>Qwen3_5RMSNormGated</td><td>18</td><td>1</td></tr><tr class=""><td>Qwen3_5Attention</td><td>6</td><td>1</td></tr><tr class=""><td>ModuleList</td><td>3</td><td>3</td></tr><tr class=""><td>Embedding</td><td>2</td><td>2</td></tr><tr class=""><td>Conv2d</td><td>1</td><td>1</td></tr><tr class=""><td>GELU</td><td>1</td><td>1</td></tr><tr class=""><td>DownsampleMLP</td><td>1</td><td>1</td></tr><tr class=""><td>ForConditionalGeneration</td><td>1</td><td>1</td></tr><tr class=""><td>Merger</td><td>1</td><td>1</td></tr><tr class=""><td>Model</td><td>1</td><td>1</td></tr><tr class=""><td>ViTWindowAttentionMerger</td><td>1</td><td>1</td></tr><tr class=""><td>VisionEmbeddings</td><td>1</td><td>1</td></tr><tr class=""><td>VisionEncoder</td><td>1</td><td>1</td></tr><tr class=""><td>VisionModel</td><td>1</td><td>1</td></tr><tr class=""><td>Qwen3_5TextModel</td><td>1</td><td>1</td></tr><tr class=""><td>Qwen3_5TextRotaryEmbedding</td><td>1</td><td>1</td></tr></table><h2>Config changes</h2><p class="note">10 fields changed, 73 unchanged.</p><table><tr><th>Field</th><th>Original</th><th>Fixture</th></tr><tr><td>text_config.head_dim</td><td>256</td><td>16</td></tr><tr><td>text_config.hidden_size</td><td>1024</td><td>64</td></tr><tr><td>text_config.intermediate_size</td><td>3584</td><td>128</td></tr><tr><td>text_config.layer_types</td><td>24 entries: 18× linear_attention, 6× full_attention</td><td>[&quot;linear_attention&quot;, &quot;full_attention&quot;]</td></tr><tr><td>text_config.num_attention_heads</td><td>8</td><td>4</td></tr><tr><td>text_config.num_hidden_layers</td><td>24</td><td>2</td></tr><tr><td>vision_config.hidden_size</td><td>1152</td><td>64</td></tr><tr><td>vision_config.intermediate_size</td><td>4304</td><td>128</td></tr><tr><td>vision_config.num_attention_heads</td><td>16</td><td>4</td></tr><tr><td>vision_config.num_hidden_layers</td><td>27</td><td>2</td></tr></table>
+</main></body></html>

shrink/MiniCPM3-4B-mic.html

@@ -0,0 +1,34 @@
+<!doctype html>
+<html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
+<title>Shrunk architecture</title>
+<style>
+:root{--bg:#fcfcfb;--fg:#0b0b0b;--muted:#52514e;--line:#e4e3df;--ok:#1d7a43;--warn:#b35c00;--s0:#2a78d6;--s1:#eb6834;--s2:#1baf7a;--s3:#eda100;--s4:#e87ba4;--s5:#008300;--s6:#6250d6;--s7:#e34948;--s8:#8a8984;}
+@media (prefers-color-scheme:dark){:root:not([data-theme="light"]){--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}}
+:root[data-theme="dark"]{--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}
+body{margin:0;background:var(--bg);color:var(--fg);font:15px/1.5 system-ui,sans-serif}
+main{max-width:1040px;margin:0 auto;padding:24px 16px 64px}
+h1{font-size:24px;margin:0 0 4px}h2{font-size:18px;margin:36px 0 8px}h3{font-size:14px;margin:0 0 8px;font-family:ui-monospace,monospace}
+.meta,.note{color:var(--muted);margin:0 0 12px;overflow-wrap:anywhere}
+.stats{display:grid;grid-template-columns:repeat(auto-fit,minmax(150px,1fr));gap:12px;margin:20px 0}
+.stat{border:1px solid var(--line);border-radius:8px;padding:12px;display:flex;flex-direction:column}
+.stat .v{font-size:26px;font-weight:600}.stat .k{color:var(--muted);font-size:13px}
+.findings{border-radius:8px;padding:4px 16px;border:1px solid var(--line)}
+.findings.ok{border-left:4px solid var(--ok)}.findings.warn{border-left:4px solid var(--warn)}
+.findings h2{margin:12px 0 4px;font-size:16px}
+.stack{border:1px solid var(--line);border-radius:8px;padding:12px;margin:12px 0}
+.row{display:flex;align-items:center;gap:12px;margin:6px 0}
+.lab{flex:0 0 110px;color:var(--muted);font-size:13px}
+.cells{display:flex;flex-wrap:wrap;gap:2px}
+.cells i{width:12px;height:12px;border-radius:2px;display:block}
+.bar{flex:1;display:flex;gap:2px;height:16px}.bar i{display:block;border-radius:2px;min-width:2px}
+table{border-collapse:collapse;width:100%;font-size:13px;margin:8px 0;display:block;overflow-x:auto}
+th,td{text-align:left;padding:4px 10px 4px 0;border-bottom:1px solid var(--line);vertical-align:top}
+th{color:var(--muted);font-weight:500}
+td:first-child{font-family:ui-monospace,monospace;overflow-wrap:anywhere}
+td i{display:inline-block;width:10px;height:10px;border-radius:2px;margin-right:6px}
+tr.bad td{color:var(--warn)}
+.c0{background:var(--s0)}.c1{background:var(--s1)}.c2{background:var(--s2)}.c3{background:var(--s3)}.c4{background:var(--s4)}.c5{background:var(--s5)}.c6{background:var(--s6)}.c7{background:var(--s7)}.c8{background:var(--s8)}
+</style></head><body><main>
+<h1>Shrunk architecture</h1><p class="meta">huggingface-mic/MiniCPM3-4B-mic@874f11d622cb165e995468f15d2f43f242b52b4a ← openbmb/MiniCPM3-4B@d6b14ddaefdb11c624dd75c3c779549bc90b08cb · MiniCPM3ForCausalLM · random weights · built before the shrink rule (no recorded recipe)</p>
+<div class="stats"><div class="stat"><span class="k">Original</span><span class="v">4.1B</span><span class="k">parameters</span></div><div class="stat"><span class="k">Fixture</span><span class="v">4.8M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Layer stacks</span><span class="v">1</span><span class="k">compared</span></div><div class="stat"><span class="k">Findings</span><span class="v">0</span><span class="k">to review</span></div></div><section class="findings ok"><h2>✓ Same architecture, smaller</h2><p>Every module class, layer kind and checked property of the original is in the fixture.</p></section><h2>Layer stacks</h2><p class="note">One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.</p><div class="stack"><h3>model.layers</h3><div class="row"><span class="lab">Original · 62</span><div class="cells"><i class="c0" title="layer 0: DecoderLayer"></i><i class="c0" title="layer 1: DecoderLayer"></i><i class="c0" title="layer 2: DecoderLayer"></i><i class="c0" title="layer 3: DecoderLayer"></i><i class="c0" title="layer 4: DecoderLayer"></i><i class="c0" title="layer 5: DecoderLayer"></i><i class="c0" title="layer 6: DecoderLayer"></i><i class="c0" title="layer 7: DecoderLayer"></i><i class="c0" title="layer 8: DecoderLayer"></i><i class="c0" title="layer 9: DecoderLayer"></i><i class="c0" title="layer 10: DecoderLayer"></i><i class="c0" title="layer 11: DecoderLayer"></i><i class="c0" title="layer 12: DecoderLayer"></i><i class="c0" title="layer 13: DecoderLayer"></i><i class="c0" title="layer 14: DecoderLayer"></i><i class="c0" title="layer 15: DecoderLayer"></i><i class="c0" title="layer 16: DecoderLayer"></i><i class="c0" title="layer 17: DecoderLayer"></i><i class="c0" title="layer 18: DecoderLayer"></i><i class="c0" title="layer 19: DecoderLayer"></i><i class="c0" title="layer 20: DecoderLayer"></i><i class="c0" title="layer 21: DecoderLayer"></i><i class="c0" title="layer 22: DecoderLayer"></i><i class="c0" title="layer 23: DecoderLayer"></i><i class="c0" title="layer 24: DecoderLayer"></i><i class="c0" title="layer 25: DecoderLayer"></i><i class="c0" title="layer 26: DecoderLayer"></i><i class="c0" title="layer 27: DecoderLayer"></i><i class="c0" title="layer 28: DecoderLayer"></i><i class="c0" title="layer 29: DecoderLayer"></i><i class="c0" title="layer 30: DecoderLayer"></i><i class="c0" title="layer 31: DecoderLayer"></i><i class="c0" title="layer 32: DecoderLayer"></i><i class="c0" title="layer 33: DecoderLayer"></i><i class="c0" title="layer 34: DecoderLayer"></i><i class="c0" title="layer 35: DecoderLayer"></i><i class="c0" title="layer 36: DecoderLayer"></i><i class="c0" title="layer 37: DecoderLayer"></i><i class="c0" title="layer 38: DecoderLayer"></i><i class="c0" title="layer 39: DecoderLayer"></i><i class="c0" title="layer 40: DecoderLayer"></i><i class="c0" title="layer 41: DecoderLayer"></i><i class="c0" title="layer 42: DecoderLayer"></i><i class="c0" title="layer 43: DecoderLayer"></i><i class="c0" title="layer 44: DecoderLayer"></i><i class="c0" title="layer 45: DecoderLayer"></i><i class="c0" title="layer 46: DecoderLayer"></i><i class="c0" title="layer 47: DecoderLayer"></i><i class="c0" title="layer 48: DecoderLayer"></i><i class="c0" title="layer 49: DecoderLayer"></i><i class="c0" title="layer 50: DecoderLayer"></i><i class="c0" title="layer 51: DecoderLayer"></i><i class="c0" title="layer 52: DecoderLayer"></i><i class="c0" title="layer 53: DecoderLayer"></i><i class="c0" title="layer 54: DecoderLayer"></i><i class="c0" title="layer 55: DecoderLayer"></i><i class="c0" title="layer 56: DecoderLayer"></i><i class="c0" title="layer 57: DecoderLayer"></i><i class="c0" title="layer 58: DecoderLayer"></i><i class="c0" title="layer 59: DecoderLayer"></i><i class="c0" title="layer 60: DecoderLayer"></i><i class="c0" title="layer 61: DecoderLayer"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: DecoderLayer"></i><i class="c0" title="layer 1: DecoderLayer"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>DecoderLayer</td><td>62</td><td>2</td><td>✓ kept</td></tr></table></div><h2>Where the parameters are</h2><p class="note">Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.</p><div class="row"><span class="lab">Original</span><div class="bar"><i class="c0" style="flex:0.953845" title="model.layers: 3.9B (95.4%)"></i><i class="c1" style="flex:0.046154" title="model.embed_tokens: 188.0M (4.6%)"></i><i class="c2" style="flex:0.000001" title="model.norm: 2.6K (0.0%)"></i></div></div><div class="row"><span class="lab">Fixture</span><div class="bar"><i class="c0" style="flex:0.014663" title="model.layers: 70.0K (1.5%)"></i><i class="c1" style="flex:0.985324" title="model.embed_tokens: 4.7M (98.5%)"></i><i class="c2" style="flex:0.000013" title="model.norm: 64 (0.0%)"></i></div></div><table class="legend"><tr><th>Component</th><th>Original</th><th>Share</th><th>Fixture</th><th>Share</th></tr><tr><td><i class="c0"></i>model.layers</td><td>3.9B</td><td>95.4%</td><td>70.0K</td><td>1.5%</td></tr><tr><td><i class="c1"></i>model.embed_tokens</td><td>188.0M</td><td>4.6%</td><td>4.7M</td><td>98.5%</td></tr><tr><td><i class="c2"></i>model.norm</td><td>2.6K</td><td>0.0%</td><td>64</td><td>0.0%</td></tr></table><h2>Preserved properties</h2><table><tr><th>Config</th><th>Property</th><th>Original</th><th>Fixture</th><th></th></tr><tr class=""><td>model</td><td>attention style</td><td>MHA (40 heads, 40 KV)</td><td>MHA (4 heads, 4 KV)</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>RoPE type</td><td>longrope</td><td>longrope</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>model_type</td><td>minicpm3</td><td>minicpm3</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>vocab_size</td><td>73448</td><td>73448</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>tie_word_embeddings</td><td>True</td><td>True</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>hidden_act</td><td>silu</td><td>silu</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>max_position_embeddings</td><td>32768</td><td>32768</td><td>✓ kept</td></tr></table><h2>Module classes</h2><p class="note">Instances of each class in the built model (prefix MiniCPM3 omitted).</p><table><tr><th>Class</th><th>Original</th><th>Fixture</th></tr><tr class=""><td>Linear</td><td>497</td><td>17</td></tr><tr class=""><td>RMSNorm</td><td>249</td><td>9</td></tr><tr class=""><td>Attention</td><td>62</td><td>2</td></tr><tr class=""><td>DecoderLayer</td><td>62</td><td>2</td></tr><tr class=""><td>MLP</td><td>62</td><td>2</td></tr><tr class=""><td>SiLUActivation</td><td>62</td><td>2</td></tr><tr class=""><td>ForCausalLM</td><td>1</td><td>1</td></tr><tr class=""><td>Model</td><td>1</td><td>1</td></tr><tr class=""><td>RotaryEmbedding</td><td>1</td><td>1</td></tr><tr class=""><td>ScaledWordEmbedding</td><td>1</td><td>1</td></tr><tr class=""><td>ModuleList</td><td>1</td><td>1</td></tr></table><h2>Config changes</h2><p class="note">12 fields changed, 28 unchanged.</p><table><tr><th>Field</th><th>Original</th><th>Fixture</th></tr><tr><td>head_dim</td><td>32</td><td>8</td></tr><tr><td>hidden_size</td><td>2560</td><td>64</td></tr><tr><td>intermediate_size</td><td>6400</td><td>128</td></tr><tr><td>kv_lora_rank</td><td>256</td><td>16</td></tr><tr><td>num_attention_heads</td><td>40</td><td>4</td></tr><tr><td>num_hidden_layers</td><td>62</td><td>2</td></tr><tr><td>num_key_value_heads</td><td>40</td><td>4</td></tr><tr><td>q_lora_rank</td><td>768</td><td>16</td></tr><tr><td>qk_nope_head_dim</td><td>64</td><td>16</td></tr><tr><td>qk_rope_head_dim</td><td>32</td><td>8</td></tr><tr><td>rope_parameters</td><td>{&quot;type&quot;: &quot;longrope&quot;, &quot;long_factor&quot;: [1.0591234137867171, 1.1241891283591912, …</td><td>{&quot;long_factor&quot;: [1.0591234137867171, 2.093982484148734, 10.475458000005451, 1…</td></tr><tr><td>v_head_dim</td><td>64</td><td>16</td></tr></table>
+</main></body></html>

shrink/MiniMax-M2.5-mic.html

@@ -0,0 +1,34 @@
+<!doctype html>
+<html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
+<title>Shrunk architecture</title>
+<style>
+:root{--bg:#fcfcfb;--fg:#0b0b0b;--muted:#52514e;--line:#e4e3df;--ok:#1d7a43;--warn:#b35c00;--s0:#2a78d6;--s1:#eb6834;--s2:#1baf7a;--s3:#eda100;--s4:#e87ba4;--s5:#008300;--s6:#6250d6;--s7:#e34948;--s8:#8a8984;}
+@media (prefers-color-scheme:dark){:root:not([data-theme="light"]){--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}}
+:root[data-theme="dark"]{--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}
+body{margin:0;background:var(--bg);color:var(--fg);font:15px/1.5 system-ui,sans-serif}
+main{max-width:1040px;margin:0 auto;padding:24px 16px 64px}
+h1{font-size:24px;margin:0 0 4px}h2{font-size:18px;margin:36px 0 8px}h3{font-size:14px;margin:0 0 8px;font-family:ui-monospace,monospace}
+.meta,.note{color:var(--muted);margin:0 0 12px;overflow-wrap:anywhere}
+.stats{display:grid;grid-template-columns:repeat(auto-fit,minmax(150px,1fr));gap:12px;margin:20px 0}
+.stat{border:1px solid var(--line);border-radius:8px;padding:12px;display:flex;flex-direction:column}
+.stat .v{font-size:26px;font-weight:600}.stat .k{color:var(--muted);font-size:13px}
+.findings{border-radius:8px;padding:4px 16px;border:1px solid var(--line)}
+.findings.ok{border-left:4px solid var(--ok)}.findings.warn{border-left:4px solid var(--warn)}
+.findings h2{margin:12px 0 4px;font-size:16px}
+.stack{border:1px solid var(--line);border-radius:8px;padding:12px;margin:12px 0}
+.row{display:flex;align-items:center;gap:12px;margin:6px 0}
+.lab{flex:0 0 110px;color:var(--muted);font-size:13px}
+.cells{display:flex;flex-wrap:wrap;gap:2px}
+.cells i{width:12px;height:12px;border-radius:2px;display:block}
+.bar{flex:1;display:flex;gap:2px;height:16px}.bar i{display:block;border-radius:2px;min-width:2px}
+table{border-collapse:collapse;width:100%;font-size:13px;margin:8px 0;display:block;overflow-x:auto}
+th,td{text-align:left;padding:4px 10px 4px 0;border-bottom:1px solid var(--line);vertical-align:top}
+th{color:var(--muted);font-weight:500}
+td:first-child{font-family:ui-monospace,monospace;overflow-wrap:anywhere}
+td i{display:inline-block;width:10px;height:10px;border-radius:2px;margin-right:6px}
+tr.bad td{color:var(--warn)}
+.c0{background:var(--s0)}.c1{background:var(--s1)}.c2{background:var(--s2)}.c3{background:var(--s3)}.c4{background:var(--s4)}.c5{background:var(--s5)}.c6{background:var(--s6)}.c7{background:var(--s7)}.c8{background:var(--s8)}
+</style></head><body><main>
+<h1>Shrunk architecture</h1><p class="meta">huggingface-mic/MiniMax-M2.5-mic@5e9d65abcedbbc965c77d1f07ffe7fba04f5d512 ← MiniMaxAI/MiniMax-M2.5@2515a3ea0afec1dd22d0c7794b3a91557804d195 · MiniMaxM2ForCausalLM · random weights</p>
+<div class="stats"><div class="stat"><span class="k">Original</span><span class="v">228.7B</span><span class="k">parameters</span></div><div class="stat"><span class="k">Fixture</span><span class="v">25.8M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Layer stacks</span><span class="v">1</span><span class="k">compared</span></div><div class="stat"><span class="k">Findings</span><span class="v">0</span><span class="k">to review</span></div></div><section class="findings ok"><h2>✓ Same architecture, smaller</h2><p>Every module class, layer kind and checked property of the original is in the fixture.</p></section><h2>Layer stacks</h2><p class="note">One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.</p><div class="stack"><h3>model.layers</h3><div class="row"><span class="lab">Original · 62</span><div class="cells"><i class="c0" title="layer 0: DecoderLayer"></i><i class="c0" title="layer 1: DecoderLayer"></i><i class="c0" title="layer 2: DecoderLayer"></i><i class="c0" title="layer 3: DecoderLayer"></i><i class="c0" title="layer 4: DecoderLayer"></i><i class="c0" title="layer 5: DecoderLayer"></i><i class="c0" title="layer 6: DecoderLayer"></i><i class="c0" title="layer 7: DecoderLayer"></i><i class="c0" title="layer 8: DecoderLayer"></i><i class="c0" title="layer 9: DecoderLayer"></i><i class="c0" title="layer 10: DecoderLayer"></i><i class="c0" title="layer 11: DecoderLayer"></i><i class="c0" title="layer 12: DecoderLayer"></i><i class="c0" title="layer 13: DecoderLayer"></i><i class="c0" title="layer 14: DecoderLayer"></i><i class="c0" title="layer 15: DecoderLayer"></i><i class="c0" title="layer 16: DecoderLayer"></i><i class="c0" title="layer 17: DecoderLayer"></i><i class="c0" title="layer 18: DecoderLayer"></i><i class="c0" title="layer 19: DecoderLayer"></i><i class="c0" title="layer 20: DecoderLayer"></i><i class="c0" title="layer 21: DecoderLayer"></i><i class="c0" title="layer 22: DecoderLayer"></i><i class="c0" title="layer 23: DecoderLayer"></i><i class="c0" title="layer 24: DecoderLayer"></i><i class="c0" title="layer 25: DecoderLayer"></i><i class="c0" title="layer 26: DecoderLayer"></i><i class="c0" title="layer 27: DecoderLayer"></i><i class="c0" title="layer 28: DecoderLayer"></i><i class="c0" title="layer 29: DecoderLayer"></i><i class="c0" title="layer 30: DecoderLayer"></i><i class="c0" title="layer 31: DecoderLayer"></i><i class="c0" title="layer 32: DecoderLayer"></i><i class="c0" title="layer 33: DecoderLayer"></i><i class="c0" title="layer 34: DecoderLayer"></i><i class="c0" title="layer 35: DecoderLayer"></i><i class="c0" title="layer 36: DecoderLayer"></i><i class="c0" title="layer 37: DecoderLayer"></i><i class="c0" title="layer 38: DecoderLayer"></i><i class="c0" title="layer 39: DecoderLayer"></i><i class="c0" title="layer 40: DecoderLayer"></i><i class="c0" title="layer 41: DecoderLayer"></i><i class="c0" title="layer 42: DecoderLayer"></i><i class="c0" title="layer 43: DecoderLayer"></i><i class="c0" title="layer 44: DecoderLayer"></i><i class="c0" title="layer 45: DecoderLayer"></i><i class="c0" title="layer 46: DecoderLayer"></i><i class="c0" title="layer 47: DecoderLayer"></i><i class="c0" title="layer 48: DecoderLayer"></i><i class="c0" title="layer 49: DecoderLayer"></i><i class="c0" title="layer 50: DecoderLayer"></i><i class="c0" title="layer 51: DecoderLayer"></i><i class="c0" title="layer 52: DecoderLayer"></i><i class="c0" title="layer 53: DecoderLayer"></i><i class="c0" title="layer 54: DecoderLayer"></i><i class="c0" title="layer 55: DecoderLayer"></i><i class="c0" title="layer 56: DecoderLayer"></i><i class="c0" title="layer 57: DecoderLayer"></i><i class="c0" title="layer 58: DecoderLayer"></i><i class="c0" title="layer 59: DecoderLayer"></i><i class="c0" title="layer 60: DecoderLayer"></i><i class="c0" title="layer 61: DecoderLayer"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: DecoderLayer"></i><i class="c0" title="layer 1: DecoderLayer"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>DecoderLayer</td><td>62</td><td>2</td><td>✓ kept</td></tr></table></div><h2>Where the parameters are</h2><p class="note">Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.</p><div class="row"><span class="lab">Original</span><div class="bar"><i class="c0" style="flex:0.994625" title="model.layers: 227.5B (99.5%)"></i><i class="c1" style="flex:0.002687" title="lm_head: 614.6M (0.3%)"></i><i class="c2" style="flex:0.002687" title="model.embed_tokens: 614.6M (0.3%)"></i><i class="c3" style="flex:0.000000" title="model.norm: 3.1K (0.0%)"></i></div></div><div class="row"><span class="lab">Fixture</span><div class="bar"><i class="c0" style="flex:0.008600" title="model.layers: 222.1K (0.9%)"></i><i class="c1" style="flex:0.495699" title="lm_head: 12.8M (49.6%)"></i><i class="c2" style="flex:0.495699" title="model.embed_tokens: 12.8M (49.6%)"></i><i class="c3" style="flex:0.000002" title="model.norm: 64 (0.0%)"></i></div></div><table class="legend"><tr><th>Component</th><th>Original</th><th>Share</th><th>Fixture</th><th>Share</th></tr><tr><td><i class="c0"></i>model.layers</td><td>227.5B</td><td>99.5%</td><td>222.1K</td><td>0.9%</td></tr><tr><td><i class="c1"></i>lm_head</td><td>614.6M</td><td>0.3%</td><td>12.8M</td><td>49.6%</td></tr><tr><td><i class="c2"></i>model.embed_tokens</td><td>614.6M</td><td>0.3%</td><td>12.8M</td><td>49.6%</td></tr><tr><td><i class="c3"></i>model.norm</td><td>3.1K</td><td>0.0%</td><td>64</td><td>0.0%</td></tr></table><h2>Preserved properties</h2><table><tr><th>Config</th><th>Property</th><th>Original</th><th>Fixture</th><th></th></tr><tr class=""><td>model</td><td>attention style</td><td>GQA (48 heads, 8 KV)</td><td>GQA (4 heads, 2 KV)</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>MoE routing</td><td>top-8 of 256 experts</td><td>top-2 of 4 experts</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>RoPE type</td><td>default</td><td>default</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>model_type</td><td>minimax_m2</td><td>minimax_m2</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>vocab_size</td><td>200064</td><td>200064</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>tie_word_embeddings</td><td>False</td><td>False</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>hidden_act</td><td>silu</td><td>silu</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>max_position_embeddings</td><td>196608</td><td>196608</td><td>✓ kept</td></tr></table><h2>Module classes</h2><p class="note">Instances of each class in the built model (prefix MiniMaxM2 omitted).</p><table><tr><th>Class</th><th>Original</th><th>Fixture</th></tr><tr class=""><td>Linear</td><td>249</td><td>9</td></tr><tr class=""><td>RMSNorm</td><td>249</td><td>9</td></tr><tr class=""><td>Attention</td><td>62</td><td>2</td></tr><tr class=""><td>DecoderLayer</td><td>62</td><td>2</td></tr><tr class=""><td>Experts</td><td>62</td><td>2</td></tr><tr class=""><td>SparseMoeBlock</td><td>62</td><td>2</td></tr><tr class=""><td>TopKRouter</td><td>62</td><td>2</td></tr><tr class=""><td>SiLUActivation</td><td>62</td><td>2</td></tr><tr class=""><td>Embedding</td><td>1</td><td>1</td></tr><tr class=""><td>ForCausalLM</td><td>1</td><td>1</td></tr><tr class=""><td>Model</td><td>1</td><td>1</td></tr><tr class=""><td>RotaryEmbedding</td><td>1</td><td>1</td></tr><tr class=""><td>ModuleList</td><td>1</td><td>1</td></tr></table><h2>Config changes</h2><p class="note">10 fields changed, 34 unchanged.</p><table><tr><th>Field</th><th>Original</th><th>Fixture</th></tr><tr><td>attn_type_list</td><td>62 entries: 62× 1</td><td>[1, 1]</td></tr><tr><td>head_dim</td><td>128</td><td>16</td></tr><tr><td>hidden_size</td><td>3072</td><td>64</td></tr><tr><td>intermediate_size</td><td>1536</td><td>128</td></tr><tr><td>num_attention_heads</td><td>48</td><td>4</td></tr><tr><td>num_experts_per_tok</td><td>8</td><td>2</td></tr><tr><td>num_hidden_layers</td><td>62</td><td>2</td></tr><tr><td>num_key_value_heads</td><td>8</td><td>2</td></tr><tr><td>num_local_experts</td><td>256</td><td>4</td></tr><tr><td>quantization_config</td><td>{&quot;activation_scheme&quot;: &quot;dynamic&quot;, &quot;fmt&quot;: &quot;float8_e4m3fn&quot;, &quot;quant_method&quot;: &quot;fp8…</td><td>—</td></tr></table>
+</main></body></html>

shrink/MiniMax-M3-MXFP8-mic.html

@@ -0,0 +1,34 @@
+<!doctype html>
+<html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
+<title>Shrunk architecture</title>
+<style>
+:root{--bg:#fcfcfb;--fg:#0b0b0b;--muted:#52514e;--line:#e4e3df;--ok:#1d7a43;--warn:#b35c00;--s0:#2a78d6;--s1:#eb6834;--s2:#1baf7a;--s3:#eda100;--s4:#e87ba4;--s5:#008300;--s6:#6250d6;--s7:#e34948;--s8:#8a8984;}
+@media (prefers-color-scheme:dark){:root:not([data-theme="light"]){--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}}
+:root[data-theme="dark"]{--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}
+body{margin:0;background:var(--bg);color:var(--fg);font:15px/1.5 system-ui,sans-serif}
+main{max-width:1040px;margin:0 auto;padding:24px 16px 64px}
+h1{font-size:24px;margin:0 0 4px}h2{font-size:18px;margin:36px 0 8px}h3{font-size:14px;margin:0 0 8px;font-family:ui-monospace,monospace}
+.meta,.note{color:var(--muted);margin:0 0 12px;overflow-wrap:anywhere}
+.stats{display:grid;grid-template-columns:repeat(auto-fit,minmax(150px,1fr));gap:12px;margin:20px 0}
+.stat{border:1px solid var(--line);border-radius:8px;padding:12px;display:flex;flex-direction:column}
+.stat .v{font-size:26px;font-weight:600}.stat .k{color:var(--muted);font-size:13px}
+.findings{border-radius:8px;padding:4px 16px;border:1px solid var(--line)}
+.findings.ok{border-left:4px solid var(--ok)}.findings.warn{border-left:4px solid var(--warn)}
+.findings h2{margin:12px 0 4px;font-size:16px}
+.stack{border:1px solid var(--line);border-radius:8px;padding:12px;margin:12px 0}
+.row{display:flex;align-items:center;gap:12px;margin:6px 0}
+.lab{flex:0 0 110px;color:var(--muted);font-size:13px}
+.cells{display:flex;flex-wrap:wrap;gap:2px}
+.cells i{width:12px;height:12px;border-radius:2px;display:block}
+.bar{flex:1;display:flex;gap:2px;height:16px}.bar i{display:block;border-radius:2px;min-width:2px}
+table{border-collapse:collapse;width:100%;font-size:13px;margin:8px 0;display:block;overflow-x:auto}
+th,td{text-align:left;padding:4px 10px 4px 0;border-bottom:1px solid var(--line);vertical-align:top}
+th{color:var(--muted);font-weight:500}
+td:first-child{font-family:ui-monospace,monospace;overflow-wrap:anywhere}
+td i{display:inline-block;width:10px;height:10px;border-radius:2px;margin-right:6px}
+tr.bad td{color:var(--warn)}
+.c0{background:var(--s0)}.c1{background:var(--s1)}.c2{background:var(--s2)}.c3{background:var(--s3)}.c4{background:var(--s4)}.c5{background:var(--s5)}.c6{background:var(--s6)}.c7{background:var(--s7)}.c8{background:var(--s8)}
+</style></head><body><main>
+<h1>Shrunk architecture</h1><p class="meta">huggingface-mic/MiniMax-M3-MXFP8-mic@314c24fd31f3764227458c8fe5aa7750eb65f6b8 ← MiniMaxAI/MiniMax-M3-MXFP8@d2be410f483297907301ec94fdf0ca5385f9b169 · MiniMaxM3SparseForConditionalGeneration · random weights</p>
+<div class="stats"><div class="stat"><span class="k">Original</span><span class="v">427.0B</span><span class="k">parameters</span></div><div class="stat"><span class="k">Fixture</span><span class="v">31.0M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Layer stacks</span><span class="v">2</span><span class="k">compared</span></div><div class="stat"><span class="k">Findings</span><span class="v">0</span><span class="k">to review</span></div></div><section class="findings ok"><h2>✓ Same architecture, smaller</h2><p>Every module class, layer kind and checked property of the original is in the fixture.</p></section><h2>Layer stacks</h2><p class="note">One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.</p><div class="stack"><h3>model.vision_tower.layers</h3><div class="row"><span class="lab">Original · 32</span><div class="cells"><i class="c0" title="layer 0: MiniMaxM3VLVisionEncoderLayer"></i><i class="c0" title="layer 1: MiniMaxM3VLVisionEncoderLayer"></i><i class="c0" title="layer 2: MiniMaxM3VLVisionEncoderLayer"></i><i class="c0" title="layer 3: MiniMaxM3VLVisionEncoderLayer"></i><i class="c0" title="layer 4: MiniMaxM3VLVisionEncoderLayer"></i><i class="c0" title="layer 5: MiniMaxM3VLVisionEncoderLayer"></i><i class="c0" title="layer 6: MiniMaxM3VLVisionEncoderLayer"></i><i class="c0" title="layer 7: MiniMaxM3VLVisionEncoderLayer"></i><i class="c0" title="layer 8: MiniMaxM3VLVisionEncoderLayer"></i><i class="c0" title="layer 9: MiniMaxM3VLVisionEncoderLayer"></i><i class="c0" title="layer 10: MiniMaxM3VLVisionEncoderLayer"></i><i class="c0" title="layer 11: MiniMaxM3VLVisionEncoderLayer"></i><i class="c0" title="layer 12: MiniMaxM3VLVisionEncoderLayer"></i><i class="c0" title="layer 13: MiniMaxM3VLVisionEncoderLayer"></i><i class="c0" title="layer 14: MiniMaxM3VLVisionEncoderLayer"></i><i class="c0" title="layer 15: MiniMaxM3VLVisionEncoderLayer"></i><i class="c0" title="layer 16: MiniMaxM3VLVisionEncoderLayer"></i><i class="c0" title="layer 17: MiniMaxM3VLVisionEncoderLayer"></i><i class="c0" title="layer 18: MiniMaxM3VLVisionEncoderLayer"></i><i class="c0" title="layer 19: MiniMaxM3VLVisionEncoderLayer"></i><i class="c0" title="layer 20: MiniMaxM3VLVisionEncoderLayer"></i><i class="c0" title="layer 21: MiniMaxM3VLVisionEncoderLayer"></i><i class="c0" title="layer 22: MiniMaxM3VLVisionEncoderLayer"></i><i class="c0" title="layer 23: MiniMaxM3VLVisionEncoderLayer"></i><i class="c0" title="layer 24: MiniMaxM3VLVisionEncoderLayer"></i><i class="c0" title="layer 25: MiniMaxM3VLVisionEncoderLayer"></i><i class="c0" title="layer 26: MiniMaxM3VLVisionEncoderLayer"></i><i class="c0" title="layer 27: MiniMaxM3VLVisionEncoderLayer"></i><i class="c0" title="layer 28: MiniMaxM3VLVisionEncoderLayer"></i><i class="c0" title="layer 29: MiniMaxM3VLVisionEncoderLayer"></i><i class="c0" title="layer 30: MiniMaxM3VLVisionEncoderLayer"></i><i class="c0" title="layer 31: MiniMaxM3VLVisionEncoderLayer"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: MiniMaxM3VLVisionEncoderLayer"></i><i class="c0" title="layer 1: MiniMaxM3VLVisionEncoderLayer"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>MiniMaxM3VLVisionEncoderLayer</td><td>32</td><td>2</td><td>✓ kept</td></tr></table></div><div class="stack"><h3>model.language_model.layers</h3><div class="row"><span class="lab">Original · 60</span><div class="cells"><i class="c0" title="layer 0: no MiniMaxM3VLExperts"></i><i class="c0" title="layer 1: no MiniMaxM3VLExperts"></i><i class="c0" title="layer 2: no MiniMaxM3VLExperts"></i><i class="c1" title="layer 3: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 4: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 5: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 6: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 7: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 8: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 9: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 10: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 11: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 12: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 13: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 14: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 15: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 16: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 17: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 18: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 19: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 20: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 21: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 22: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 23: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 24: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 25: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 26: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 27: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 28: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 29: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 30: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 31: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 32: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 33: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 34: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 35: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 36: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 37: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 38: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 39: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 40: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 41: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 42: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 43: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 44: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 45: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 46: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 47: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 48: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 49: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 50: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 51: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 52: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 53: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 54: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 55: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 56: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 57: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 58: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i><i class="c1" title="layer 59: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: no MiniMaxM3VLExperts"></i><i class="c1" title="layer 1: MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>no MiniMaxM3VLExperts</td><td>3</td><td>1</td><td>✓ kept</td></tr><tr><td><i class="c1"></i>MiniMaxM3VLExperts, MiniMaxM3VLIndexer, MiniMaxM3VLSparseMoeBlock …</td><td>57</td><td>1</td><td>✓ kept</td></tr></table></div><h2>Where the parameters are</h2><p class="note">Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.</p><div class="row"><span class="lab">Original</span><div class="bar"><i class="c0" style="flex:0.992216" title="model.language_model.layers: 423.7B (99.2%)"></i><i class="c1" style="flex:0.002878" title="model.language_model: 1.2B (0.3%)"></i><i class="c2" style="flex:0.002878" title="lm_head: 1.2B (0.3%)"></i><i class="c3" style="flex:0.001475" title="model.vision_tower.layers: 629.7M (0.1%)"></i><i class="c4" style="flex:0.000549" title="model.multi_modal_projector: 234.4M (0.1%)"></i><i class="c5" style="flex:0.000004" title="model.vision_tower: 1.5M (0.0%)"></i></div></div><div class="row"><span class="lab">Fixture</span><div class="bar"><i class="c0" style="flex:0.081557" title="model.language_model.layers: 2.5M (8.2%)"></i><i class="c1" style="flex:0.412403" title="model.language_model: 12.8M (41.2%)"></i><i class="c2" style="flex:0.412401" title="lm_head: 12.8M (41.2%)"></i><i class="c3" style="flex:0.002156" title="model.vision_tower.layers: 66.9K (0.2%)"></i><i class="c4" style="flex:0.089054" title="model.multi_modal_projector: 2.8M (8.9%)"></i><i class="c5" style="flex:0.002428" title="model.vision_tower: 75.4K (0.2%)"></i></div></div><table class="legend"><tr><th>Component</th><th>Original</th><th>Share</th><th>Fixture</th><th>Share</th></tr><tr><td><i class="c0"></i>model.language_model.layers</td><td>423.7B</td><td>99.2%</td><td>2.5M</td><td>8.2%</td></tr><tr><td><i class="c1"></i>model.language_model</td><td>1.2B</td><td>0.3%</td><td>12.8M</td><td>41.2%</td></tr><tr><td><i class="c2"></i>lm_head</td><td>1.2B</td><td>0.3%</td><td>12.8M</td><td>41.2%</td></tr><tr><td><i class="c3"></i>model.vision_tower.layers</td><td>629.7M</td><td>0.1%</td><td>66.9K</td><td>0.2%</td></tr><tr><td><i class="c4"></i>model.multi_modal_projector</td><td>234.4M</td><td>0.1%</td><td>2.8M</td><td>8.9%</td></tr><tr><td><i class="c5"></i>model.vision_tower</td><td>1.5M</td><td>0.0%</td><td>75.4K</td><td>0.2%</td></tr></table><h2>Preserved properties</h2><table><tr><th>Config</th><th>Property</th><th>Original</th><th>Fixture</th><th></th></tr><tr class=""><td>model</td><td>model_type</td><td>minimax_m3_vl</td><td>minimax_m3_vl</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>tie_word_embeddings</td><td>False</td><td>False</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>RoPE type</td><td>axial</td><td>axial</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>model_type</td><td>minimax_m3_vl_vision</td><td>minimax_m3_vl_vision</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>vocab_size</td><td>32000</td><td>32000</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>hidden_act</td><td>gelu</td><td>gelu</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>image_size</td><td>2016</td><td>2016</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>patch_size</td><td>14</td><td>14</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>num_channels</td><td>3</td><td>3</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>attention style</td><td>GQA (64 heads, 4 KV)</td><td>GQA (4 heads, 2 KV)</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>MoE routing</td><td>top-4 of 128 experts</td><td>top-2 of 4 experts</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>RoPE type</td><td>default</td><td>default</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>model_type</td><td>minimax_m3_vl_text</td><td>minimax_m3_vl_text</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>vocab_size</td><td>200064</td><td>200064</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>tie_word_embeddings</td><td>False</td><td>False</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>hidden_act</td><td>silu</td><td>silu</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>max_position_embeddings</td><td>1048576</td><td>1048576</td><td>✓ kept</td></tr></table><h2>Module classes</h2><p class="note">Instances of each class in the built model (prefix MiniMaxM3Sparse omitted).</p><table><tr><th>Class</th><th>Original</th><th>Fixture</th></tr><tr class=""><td>Linear</td><td>671</td><td>31</td></tr><tr class=""><td>MiniMaxM3VLRMSNorm</td><td>355</td><td>11</td></tr><tr class=""><td>LayerNorm</td><td>65</td><td>5</td></tr><tr class=""><td>MiniMaxM3VLAttention</td><td>60</td><td>2</td></tr><tr class=""><td>MiniMaxM3VLDecoderLayer</td><td>60</td><td>2</td></tr><tr class=""><td>MiniMaxM3VLDenseMLP</td><td>60</td><td>2</td></tr><tr class=""><td>MiniMaxM3VLExperts</td><td>57</td><td>1</td></tr><tr class=""><td>MiniMaxM3VLIndexer</td><td>57</td><td>1</td></tr><tr class=""><td>MiniMaxM3VLSparseMoeBlock</td><td>57</td><td>1</td></tr><tr class=""><td>MiniMaxM3VLTopKRouter</td><td>57</td><td>1</td></tr><tr class=""><td>GELUActivation</td><td>34</td><td>4</td></tr><tr class=""><td>MiniMaxM3VLVisionAttention</td><td>32</td><td>2</td></tr><tr class=""><td>MiniMaxM3VLVisionEncoderLayer</td><td>32</td><td>2</td></tr><tr class=""><td>MiniMaxM3VLVisionMLP</td><td>32</td><td>2</td></tr><tr class=""><td>ModuleList</td><td>2</td><td>2</td></tr><tr class=""><td>Conv3d</td><td>1</td><td>1</td></tr><tr class=""><td>Embedding</td><td>1</td><td>1</td></tr><tr class=""><td>ForConditionalGeneration</td><td>1</td><td>1</td></tr><tr class=""><td>MiniMaxM3VLModel</td><td>1</td><td>1</td></tr><tr class=""><td>MiniMaxM3VLMultiModalProjector</td><td>1</td><td>1</td></tr><tr class=""><td>MiniMaxM3VLRotaryEmbedding</td><td>1</td><td>1</td></tr><tr class=""><td>MiniMaxM3VLTextModel</td><td>1</td><td>1</td></tr><tr class=""><td>MiniMaxM3VLVisionEmbeddings</td><td>1</td><td>1</td></tr><tr class=""><td>MiniMaxM3VLVisionModel</td><td>1</td><td>1</td></tr><tr class=""><td>MiniMaxM3VLVisionRotaryEmbedding</td><td>1</td><td>1</td></tr></table><h2>Config changes</h2><p class="note">21 fields changed, 91 unchanged.</p><table><tr><th>Field</th><th>Original</th><th>Fixture</th></tr><tr><td>merged_hidden_size</td><td>24576</td><td>256</td></tr><tr><td>quantization_config</td><td>{&quot;quant_method&quot;: &quot;mxfp8&quot;, &quot;activation_scheme&quot;: &quot;dynamic&quot;, &quot;weight_block_size&quot;…</td><td>—</td></tr><tr><td>text_config.head_dim</td><td>128</td><td>16</td></tr><tr><td>text_config.hidden_size</td><td>6144</td><td>64</td></tr><tr><td>text_config.index_block_size</td><td>128</td><td>8</td></tr><tr><td>text_config.index_n_heads</td><td>4</td><td>2</td></tr><tr><td>text_config.index_topk_blocks</td><td>16</td><td>4</td></tr><tr><td>text_config.intermediate_size</td><td>3072</td><td>128</td></tr><tr><td>text_config.layer_types</td><td>60 entries: 57× minimax_m3_sparse, 3× full_attention</td><td>[&quot;full_attention&quot;, &quot;minimax_m3_sparse&quot;]</td></tr><tr><td>text_config.mlp_layer_types</td><td>60 entries: 57× sparse, 3× dense</td><td>[&quot;dense&quot;, &quot;sparse&quot;]</td></tr><tr><td>text_config.num_attention_heads</td><td>64</td><td>4</td></tr><tr><td>text_config.num_experts_per_tok</td><td>4</td><td>2</td></tr><tr><td>text_config.num_hidden_layers</td><td>60</td><td>2</td></tr><tr><td>text_config.num_key_value_heads</td><td>4</td><td>2</td></tr><tr><td>text_config.num_local_experts</td><td>128</td><td>4</td></tr><tr><td>text_config.shared_intermediate_size</td><td>3072</td><td>128</td></tr><tr><td>vision_config.hidden_size</td><td>1280</td><td>64</td></tr><tr><td>vision_config.intermediate_size</td><td>5120</td><td>128</td></tr><tr><td>vision_config.num_attention_heads</td><td>16</td><td>4</td></tr><tr><td>vision_config.num_hidden_layers</td><td>32</td><td>2</td></tr><tr><td>vision_config.projection_dim</td><td>6144</td><td>64</td></tr></table>
+</main></body></html>

shrink/Ministral-3-3B-Instruct-2512-mic.html

@@ -0,0 +1,34 @@
+<!doctype html>
+<html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
+<title>Shrunk architecture</title>
+<style>
+:root{--bg:#fcfcfb;--fg:#0b0b0b;--muted:#52514e;--line:#e4e3df;--ok:#1d7a43;--warn:#b35c00;--s0:#2a78d6;--s1:#eb6834;--s2:#1baf7a;--s3:#eda100;--s4:#e87ba4;--s5:#008300;--s6:#6250d6;--s7:#e34948;--s8:#8a8984;}
+@media (prefers-color-scheme:dark){:root:not([data-theme="light"]){--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}}
+:root[data-theme="dark"]{--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}
+body{margin:0;background:var(--bg);color:var(--fg);font:15px/1.5 system-ui,sans-serif}
+main{max-width:1040px;margin:0 auto;padding:24px 16px 64px}
+h1{font-size:24px;margin:0 0 4px}h2{font-size:18px;margin:36px 0 8px}h3{font-size:14px;margin:0 0 8px;font-family:ui-monospace,monospace}
+.meta,.note{color:var(--muted);margin:0 0 12px;overflow-wrap:anywhere}
+.stats{display:grid;grid-template-columns:repeat(auto-fit,minmax(150px,1fr));gap:12px;margin:20px 0}
+.stat{border:1px solid var(--line);border-radius:8px;padding:12px;display:flex;flex-direction:column}
+.stat .v{font-size:26px;font-weight:600}.stat .k{color:var(--muted);font-size:13px}
+.findings{border-radius:8px;padding:4px 16px;border:1px solid var(--line)}
+.findings.ok{border-left:4px solid var(--ok)}.findings.warn{border-left:4px solid var(--warn)}
+.findings h2{margin:12px 0 4px;font-size:16px}
+.stack{border:1px solid var(--line);border-radius:8px;padding:12px;margin:12px 0}
+.row{display:flex;align-items:center;gap:12px;margin:6px 0}
+.lab{flex:0 0 110px;color:var(--muted);font-size:13px}
+.cells{display:flex;flex-wrap:wrap;gap:2px}
+.cells i{width:12px;height:12px;border-radius:2px;display:block}
+.bar{flex:1;display:flex;gap:2px;height:16px}.bar i{display:block;border-radius:2px;min-width:2px}
+table{border-collapse:collapse;width:100%;font-size:13px;margin:8px 0;display:block;overflow-x:auto}
+th,td{text-align:left;padding:4px 10px 4px 0;border-bottom:1px solid var(--line);vertical-align:top}
+th{color:var(--muted);font-weight:500}
+td:first-child{font-family:ui-monospace,monospace;overflow-wrap:anywhere}
+td i{display:inline-block;width:10px;height:10px;border-radius:2px;margin-right:6px}
+tr.bad td{color:var(--warn)}
+.c0{background:var(--s0)}.c1{background:var(--s1)}.c2{background:var(--s2)}.c3{background:var(--s3)}.c4{background:var(--s4)}.c5{background:var(--s5)}.c6{background:var(--s6)}.c7{background:var(--s7)}.c8{background:var(--s8)}
+</style></head><body><main>
+<h1>Shrunk architecture</h1><p class="meta">huggingface-mic/Ministral-3-3B-Instruct-2512-mic@444a971de7a3a92574df14dcdd5dde549e0d660f ← mistralai/Ministral-3-3B-Instruct-2512@b35d4dfe56c142746f54dbd64f579faab2744308 · Mistral3ForConditionalGeneration · random weights · built before the shrink rule (no recorded recipe)</p>
+<div class="stats"><div class="stat"><span class="k">Original</span><span class="v">3.8B</span><span class="k">parameters</span></div><div class="stat"><span class="k">Fixture</span><span class="v">8.6M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Layer stacks</span><span class="v">2</span><span class="k">compared</span></div><div class="stat"><span class="k">Findings</span><span class="v">0</span><span class="k">to review</span></div></div><section class="findings ok"><h2>✓ Same architecture, smaller</h2><p>Every module class, layer kind and checked property of the original is in the fixture.</p></section><h2>Layer stacks</h2><p class="note">One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.</p><div class="stack"><h3>model.vision_tower.transformer.layers</h3><div class="row"><span class="lab">Original · 24</span><div class="cells"><i class="c0" title="layer 0: PixtralAttentionLayer"></i><i class="c0" title="layer 1: PixtralAttentionLayer"></i><i class="c0" title="layer 2: PixtralAttentionLayer"></i><i class="c0" title="layer 3: PixtralAttentionLayer"></i><i class="c0" title="layer 4: PixtralAttentionLayer"></i><i class="c0" title="layer 5: PixtralAttentionLayer"></i><i class="c0" title="layer 6: PixtralAttentionLayer"></i><i class="c0" title="layer 7: PixtralAttentionLayer"></i><i class="c0" title="layer 8: PixtralAttentionLayer"></i><i class="c0" title="layer 9: PixtralAttentionLayer"></i><i class="c0" title="layer 10: PixtralAttentionLayer"></i><i class="c0" title="layer 11: PixtralAttentionLayer"></i><i class="c0" title="layer 12: PixtralAttentionLayer"></i><i class="c0" title="layer 13: PixtralAttentionLayer"></i><i class="c0" title="layer 14: PixtralAttentionLayer"></i><i class="c0" title="layer 15: PixtralAttentionLayer"></i><i class="c0" title="layer 16: PixtralAttentionLayer"></i><i class="c0" title="layer 17: PixtralAttentionLayer"></i><i class="c0" title="layer 18: PixtralAttentionLayer"></i><i class="c0" title="layer 19: PixtralAttentionLayer"></i><i class="c0" title="layer 20: PixtralAttentionLayer"></i><i class="c0" title="layer 21: PixtralAttentionLayer"></i><i class="c0" title="layer 22: PixtralAttentionLayer"></i><i class="c0" title="layer 23: PixtralAttentionLayer"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: PixtralAttentionLayer"></i><i class="c0" title="layer 1: PixtralAttentionLayer"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>PixtralAttentionLayer</td><td>24</td><td>2</td><td>✓ kept</td></tr></table></div><div class="stack"><h3>model.language_model.layers</h3><div class="row"><span class="lab">Original · 26</span><div class="cells"><i class="c0" title="layer 0: Ministral3DecoderLayer"></i><i class="c0" title="layer 1: Ministral3DecoderLayer"></i><i class="c0" title="layer 2: Ministral3DecoderLayer"></i><i class="c0" title="layer 3: Ministral3DecoderLayer"></i><i class="c0" title="layer 4: Ministral3DecoderLayer"></i><i class="c0" title="layer 5: Ministral3DecoderLayer"></i><i class="c0" title="layer 6: Ministral3DecoderLayer"></i><i class="c0" title="layer 7: Ministral3DecoderLayer"></i><i class="c0" title="layer 8: Ministral3DecoderLayer"></i><i class="c0" title="layer 9: Ministral3DecoderLayer"></i><i class="c0" title="layer 10: Ministral3DecoderLayer"></i><i class="c0" title="layer 11: Ministral3DecoderLayer"></i><i class="c0" title="layer 12: Ministral3DecoderLayer"></i><i class="c0" title="layer 13: Ministral3DecoderLayer"></i><i class="c0" title="layer 14: Ministral3DecoderLayer"></i><i class="c0" title="layer 15: Ministral3DecoderLayer"></i><i class="c0" title="layer 16: Ministral3DecoderLayer"></i><i class="c0" title="layer 17: Ministral3DecoderLayer"></i><i class="c0" title="layer 18: Ministral3DecoderLayer"></i><i class="c0" title="layer 19: Ministral3DecoderLayer"></i><i class="c0" title="layer 20: Ministral3DecoderLayer"></i><i class="c0" title="layer 21: Ministral3DecoderLayer"></i><i class="c0" title="layer 22: Ministral3DecoderLayer"></i><i class="c0" title="layer 23: Ministral3DecoderLayer"></i><i class="c0" title="layer 24: Ministral3DecoderLayer"></i><i class="c0" title="layer 25: Ministral3DecoderLayer"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: Ministral3DecoderLayer"></i><i class="c0" title="layer 1: Ministral3DecoderLayer"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>Ministral3DecoderLayer</td><td>26</td><td>2</td><td>✓ kept</td></tr></table></div><h2>Where the parameters are</h2><p class="note">Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.</p><div class="row"><span class="lab">Original</span><div class="bar"><i class="c0" style="flex:0.786251" title="model.language_model.layers: 3.0B (78.6%)"></i><i class="c1" style="flex:0.104623" title="model.vision_tower.transformer.layers: 402.7M (10.5%)"></i><i class="c2" style="flex:0.104611" title="model.language_model: 402.7M (10.5%)"></i><i class="c3" style="flex:0.004359" title="model.multi_modal_projector: 16.8M (0.4%)"></i><i class="c4" style="flex:0.000157" title="model.vision_tower: 603.1K (0.0%)"></i></div></div><div class="row"><span class="lab">Fixture</span><div class="bar"><i class="c0" style="flex:0.008596" title="model.language_model.layers: 74.0K (0.9%)"></i><i class="c1" style="flex:0.009547" title="model.vision_tower.transformer.layers: 82.2K (1.0%)"></i><i class="c2" style="flex:0.974615" title="model.language_model: 8.4M (97.5%)"></i><i class="c3" style="flex:0.002863" title="model.multi_modal_projector: 24.6K (0.3%)"></i><i class="c4" style="flex:0.004380" title="model.vision_tower: 37.7K (0.4%)"></i></div></div><table class="legend"><tr><th>Component</th><th>Original</th><th>Share</th><th>Fixture</th><th>Share</th></tr><tr><td><i class="c0"></i>model.language_model.layers</td><td>3.0B</td><td>78.6%</td><td>74.0K</td><td>0.9%</td></tr><tr><td><i class="c1"></i>model.vision_tower.transformer.layers</td><td>402.7M</td><td>10.5%</td><td>82.2K</td><td>1.0%</td></tr><tr><td><i class="c2"></i>model.language_model</td><td>402.7M</td><td>10.5%</td><td>8.4M</td><td>97.5%</td></tr><tr><td><i class="c3"></i>model.multi_modal_projector</td><td>16.8M</td><td>0.4%</td><td>24.6K</td><td>0.3%</td></tr><tr><td><i class="c4"></i>model.vision_tower</td><td>603.1K</td><td>0.0%</td><td>37.7K</td><td>0.4%</td></tr></table><h2>Preserved properties</h2><table><tr><th>Config</th><th>Property</th><th>Original</th><th>Fixture</th><th></th></tr><tr class=""><td>model</td><td>model_type</td><td>mistral3</td><td>mistral3</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>tie_word_embeddings</td><td>True</td><td>True</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>RoPE type</td><td>axial</td><td>axial</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>model_type</td><td>pixtral</td><td>pixtral</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>hidden_act</td><td>silu</td><td>silu</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>image_size</td><td>1540</td><td>1540</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>patch_size</td><td>14</td><td>14</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>num_channels</td><td>3</td><td>3</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>attention style</td><td>GQA (32 heads, 8 KV)</td><td>GQA (4 heads, 2 KV)</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>RoPE type</td><td>yarn</td><td>yarn</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>model_type</td><td>ministral3</td><td>ministral3</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>vocab_size</td><td>131072</td><td>131072</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>tie_word_embeddings</td><td>True</td><td>True</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>hidden_act</td><td>silu</td><td>silu</td><td>✓ kept</td></tr><tr class=""><td>text_config</td><td>max_position_embeddings</td><td>262144</td><td>262144</td><td>✓ kept</td></tr></table><h2>Module classes</h2><p class="note">Instances of each class in the built model (prefix Mistral3 omitted).</p><table><tr><th>Class</th><th>Original</th><th>Fixture</th></tr><tr class=""><td>Linear</td><td>354</td><td>32</td></tr><tr class=""><td>Ministral3RMSNorm</td><td>53</td><td>5</td></tr><tr class=""><td>SiLUActivation</td><td>50</td><td>4</td></tr><tr class=""><td>PixtralRMSNorm</td><td>49</td><td>5</td></tr><tr class=""><td>Ministral3Attention</td><td>26</td><td>2</td></tr><tr class=""><td>Ministral3DecoderLayer</td><td>26</td><td>2</td></tr><tr class=""><td>Ministral3MLP</td><td>26</td><td>2</td></tr><tr class=""><td>PixtralAttention</td><td>24</td><td>2</td></tr><tr class=""><td>PixtralAttentionLayer</td><td>24</td><td>2</td></tr><tr class=""><td>PixtralMLP</td><td>24</td><td>2</td></tr><tr class=""><td>ModuleList</td><td>2</td><td>2</td></tr><tr class=""><td>Conv2d</td><td>1</td><td>1</td></tr><tr class=""><td>Embedding</td><td>1</td><td>1</td></tr><tr class=""><td>GELUActivation</td><td>1</td><td>1</td></tr><tr class=""><td>Ministral3Model</td><td>1</td><td>1</td></tr><tr class=""><td>Ministral3RotaryEmbedding</td><td>1</td><td>1</td></tr><tr class=""><td>ForConditionalGeneration</td><td>1</td><td>1</td></tr><tr class=""><td>Model</td><td>1</td><td>1</td></tr><tr class=""><td>MultiModalProjector</td><td>1</td><td>1</td></tr><tr class=""><td>PatchMerger</td><td>1</td><td>1</td></tr><tr class=""><td>RMSNorm</td><td>1</td><td>1</td></tr><tr class=""><td>PixtralTransformer</td><td>1</td><td>1</td></tr><tr class=""><td>PixtralVisionModel</td><td>1</td><td>1</td></tr><tr class=""><td>PixtralVisionRotaryEmbedding</td><td>1</td><td>1</td></tr></table><h2>Config changes</h2><p class="note">12 fields changed, 53 unchanged.</p><table><tr><th>Field</th><th>Original</th><th>Fixture</th></tr><tr><td>quantization_config</td><td>{&quot;activation_scheme&quot;: &quot;static&quot;, &quot;dequantize&quot;: false, &quot;modules_to_not_convert&quot;…</td><td>—</td></tr><tr><td>text_config.head_dim</td><td>128</td><td>16</td></tr><tr><td>text_config.hidden_size</td><td>3072</td><td>64</td></tr><tr><td>text_config.intermediate_size</td><td>9216</td><td>128</td></tr><tr><td>text_config.num_attention_heads</td><td>32</td><td>4</td></tr><tr><td>text_config.num_hidden_layers</td><td>26</td><td>2</td></tr><tr><td>text_config.num_key_value_heads</td><td>8</td><td>2</td></tr><tr><td>vision_config.head_dim</td><td>64</td><td>16</td></tr><tr><td>vision_config.hidden_size</td><td>1024</td><td>64</td></tr><tr><td>vision_config.intermediate_size</td><td>4096</td><td>128</td></tr><tr><td>vision_config.num_attention_heads</td><td>16</td><td>4</td></tr><tr><td>vision_config.num_hidden_layers</td><td>24</td><td>2</td></tr></table>
+</main></body></html>

shrink/Mistral-7B-Instruct-v0.2-mic.html

@@ -0,0 +1,34 @@
+<!doctype html>
+<html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
+<title>Shrunk architecture</title>
+<style>
+:root{--bg:#fcfcfb;--fg:#0b0b0b;--muted:#52514e;--line:#e4e3df;--ok:#1d7a43;--warn:#b35c00;--s0:#2a78d6;--s1:#eb6834;--s2:#1baf7a;--s3:#eda100;--s4:#e87ba4;--s5:#008300;--s6:#6250d6;--s7:#e34948;--s8:#8a8984;}
+@media (prefers-color-scheme:dark){:root:not([data-theme="light"]){--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}}
+:root[data-theme="dark"]{--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}
+body{margin:0;background:var(--bg);color:var(--fg);font:15px/1.5 system-ui,sans-serif}
+main{max-width:1040px;margin:0 auto;padding:24px 16px 64px}
+h1{font-size:24px;margin:0 0 4px}h2{font-size:18px;margin:36px 0 8px}h3{font-size:14px;margin:0 0 8px;font-family:ui-monospace,monospace}
+.meta,.note{color:var(--muted);margin:0 0 12px;overflow-wrap:anywhere}
+.stats{display:grid;grid-template-columns:repeat(auto-fit,minmax(150px,1fr));gap:12px;margin:20px 0}
+.stat{border:1px solid var(--line);border-radius:8px;padding:12px;display:flex;flex-direction:column}
+.stat .v{font-size:26px;font-weight:600}.stat .k{color:var(--muted);font-size:13px}
+.findings{border-radius:8px;padding:4px 16px;border:1px solid var(--line)}
+.findings.ok{border-left:4px solid var(--ok)}.findings.warn{border-left:4px solid var(--warn)}
+.findings h2{margin:12px 0 4px;font-size:16px}
+.stack{border:1px solid var(--line);border-radius:8px;padding:12px;margin:12px 0}
+.row{display:flex;align-items:center;gap:12px;margin:6px 0}
+.lab{flex:0 0 110px;color:var(--muted);font-size:13px}
+.cells{display:flex;flex-wrap:wrap;gap:2px}
+.cells i{width:12px;height:12px;border-radius:2px;display:block}
+.bar{flex:1;display:flex;gap:2px;height:16px}.bar i{display:block;border-radius:2px;min-width:2px}
+table{border-collapse:collapse;width:100%;font-size:13px;margin:8px 0;display:block;overflow-x:auto}
+th,td{text-align:left;padding:4px 10px 4px 0;border-bottom:1px solid var(--line);vertical-align:top}
+th{color:var(--muted);font-weight:500}
+td:first-child{font-family:ui-monospace,monospace;overflow-wrap:anywhere}
+td i{display:inline-block;width:10px;height:10px;border-radius:2px;margin-right:6px}
+tr.bad td{color:var(--warn)}
+.c0{background:var(--s0)}.c1{background:var(--s1)}.c2{background:var(--s2)}.c3{background:var(--s3)}.c4{background:var(--s4)}.c5{background:var(--s5)}.c6{background:var(--s6)}.c7{background:var(--s7)}.c8{background:var(--s8)}
+</style></head><body><main>
+<h1>Shrunk architecture</h1><p class="meta">huggingface-mic/Mistral-7B-Instruct-v0.2-mic@90938461b4cb2497a78ea7f817f89df453996848 ← mistralai/Mistral-7B-Instruct-v0.2@63a8b081895390a26e140280378bc85ec8bce07a · MistralForCausalLM · random weights · built before the shrink rule (no recorded recipe)</p>
+<div class="stats"><div class="stat"><span class="k">Original</span><span class="v">7.2B</span><span class="k">parameters</span></div><div class="stat"><span class="k">Fixture</span><span class="v">8.4M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Layer stacks</span><span class="v">1</span><span class="k">compared</span></div><div class="stat"><span class="k">Findings</span><span class="v">1</span><span class="k">to review</span></div></div><section class="findings warn"><h2>⚠ To review</h2><ul><li>model: max_position_embeddings 32768 became 512</li></ul></section><h2>Layer stacks</h2><p class="note">One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.</p><div class="stack"><h3>model.layers</h3><div class="row"><span class="lab">Original · 32</span><div class="cells"><i class="c0" title="layer 0: DecoderLayer"></i><i class="c0" title="layer 1: DecoderLayer"></i><i class="c0" title="layer 2: DecoderLayer"></i><i class="c0" title="layer 3: DecoderLayer"></i><i class="c0" title="layer 4: DecoderLayer"></i><i class="c0" title="layer 5: DecoderLayer"></i><i class="c0" title="layer 6: DecoderLayer"></i><i class="c0" title="layer 7: DecoderLayer"></i><i class="c0" title="layer 8: DecoderLayer"></i><i class="c0" title="layer 9: DecoderLayer"></i><i class="c0" title="layer 10: DecoderLayer"></i><i class="c0" title="layer 11: DecoderLayer"></i><i class="c0" title="layer 12: DecoderLayer"></i><i class="c0" title="layer 13: DecoderLayer"></i><i class="c0" title="layer 14: DecoderLayer"></i><i class="c0" title="layer 15: DecoderLayer"></i><i class="c0" title="layer 16: DecoderLayer"></i><i class="c0" title="layer 17: DecoderLayer"></i><i class="c0" title="layer 18: DecoderLayer"></i><i class="c0" title="layer 19: DecoderLayer"></i><i class="c0" title="layer 20: DecoderLayer"></i><i class="c0" title="layer 21: DecoderLayer"></i><i class="c0" title="layer 22: DecoderLayer"></i><i class="c0" title="layer 23: DecoderLayer"></i><i class="c0" title="layer 24: DecoderLayer"></i><i class="c0" title="layer 25: DecoderLayer"></i><i class="c0" title="layer 26: DecoderLayer"></i><i class="c0" title="layer 27: DecoderLayer"></i><i class="c0" title="layer 28: DecoderLayer"></i><i class="c0" title="layer 29: DecoderLayer"></i><i class="c0" title="layer 30: DecoderLayer"></i><i class="c0" title="layer 31: DecoderLayer"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: DecoderLayer"></i><i class="c0" title="layer 1: DecoderLayer"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>DecoderLayer</td><td>32</td><td>2</td><td>✓ kept</td></tr></table></div><h2>Where the parameters are</h2><p class="note">Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.</p><div class="row"><span class="lab">Original</span><div class="bar"><i class="c0" style="flex:0.963800" title="model.layers: 7.0B (96.4%)"></i><i class="c1" style="flex:0.018100" title="model.embed_tokens: 131.1M (1.8%)"></i><i class="c2" style="flex:0.018100" title="lm_head: 131.1M (1.8%)"></i><i class="c3" style="flex:0.000001" title="model.norm: 4.1K (0.0%)"></i></div></div><div class="row"><span class="lab">Fixture</span><div class="bar"><i class="c0" style="flex:0.029185" title="model.layers: 246.3K (2.9%)"></i><i class="c1" style="flex:0.485400" title="model.embed_tokens: 4.1M (48.5%)"></i><i class="c2" style="flex:0.485400" title="lm_head: 4.1M (48.5%)"></i><i class="c3" style="flex:0.000015" title="model.norm: 128 (0.0%)"></i></div></div><table class="legend"><tr><th>Component</th><th>Original</th><th>Share</th><th>Fixture</th><th>Share</th></tr><tr><td><i class="c0"></i>model.layers</td><td>7.0B</td><td>96.4%</td><td>246.3K</td><td>2.9%</td></tr><tr><td><i class="c1"></i>model.embed_tokens</td><td>131.1M</td><td>1.8%</td><td>4.1M</td><td>48.5%</td></tr><tr><td><i class="c2"></i>lm_head</td><td>131.1M</td><td>1.8%</td><td>4.1M</td><td>48.5%</td></tr><tr><td><i class="c3"></i>model.norm</td><td>4.1K</td><td>0.0%</td><td>128</td><td>0.0%</td></tr></table><h2>Preserved properties</h2><table><tr><th>Config</th><th>Property</th><th>Original</th><th>Fixture</th><th></th></tr><tr class=""><td>model</td><td>attention style</td><td>GQA (32 heads, 8 KV)</td><td>GQA (4 heads, 2 KV)</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>RoPE type</td><td>default</td><td>default</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>model_type</td><td>mistral</td><td>mistral</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>vocab_size</td><td>32000</td><td>32000</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>tie_word_embeddings</td><td>False</td><td>False</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>hidden_act</td><td>silu</td><td>silu</td><td>✓ kept</td></tr><tr class="bad"><td>model</td><td>max_position_embeddings</td><td>32768</td><td>512</td><td>⚠ changed</td></tr></table><h2>Module classes</h2><p class="note">Instances of each class in the built model (prefix Mistral omitted).</p><table><tr><th>Class</th><th>Original</th><th>Fixture</th></tr><tr class=""><td>Linear</td><td>225</td><td>15</td></tr><tr class=""><td>RMSNorm</td><td>65</td><td>5</td></tr><tr class=""><td>Attention</td><td>32</td><td>2</td></tr><tr class=""><td>DecoderLayer</td><td>32</td><td>2</td></tr><tr class=""><td>MLP</td><td>32</td><td>2</td></tr><tr class=""><td>SiLUActivation</td><td>32</td><td>2</td></tr><tr class=""><td>Embedding</td><td>1</td><td>1</td></tr><tr class=""><td>ForCausalLM</td><td>1</td><td>1</td></tr><tr class=""><td>Model</td><td>1</td><td>1</td></tr><tr class=""><td>RotaryEmbedding</td><td>1</td><td>1</td></tr><tr class=""><td>ModuleList</td><td>1</td><td>1</td></tr></table><h2>Config changes</h2><p class="note">7 fields changed, 21 unchanged.</p><table><tr><th>Field</th><th>Original</th><th>Fixture</th></tr><tr><td>head_dim</td><td>128</td><td>16</td></tr><tr><td>hidden_size</td><td>4096</td><td>128</td></tr><tr><td>intermediate_size</td><td>14336</td><td>256</td></tr><tr><td>max_position_embeddings</td><td>32768</td><td>512</td></tr><tr><td>num_attention_heads</td><td>32</td><td>4</td></tr><tr><td>num_hidden_layers</td><td>32</td><td>2</td></tr><tr><td>num_key_value_heads</td><td>8</td><td>2</td></tr></table>
+</main></body></html>

shrink/Mixtral-8x7B-Instruct-v0.1-mic.html

@@ -0,0 +1,34 @@
+<!doctype html>
+<html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
+<title>Shrunk architecture</title>
+<style>
+:root{--bg:#fcfcfb;--fg:#0b0b0b;--muted:#52514e;--line:#e4e3df;--ok:#1d7a43;--warn:#b35c00;--s0:#2a78d6;--s1:#eb6834;--s2:#1baf7a;--s3:#eda100;--s4:#e87ba4;--s5:#008300;--s6:#6250d6;--s7:#e34948;--s8:#8a8984;}
+@media (prefers-color-scheme:dark){:root:not([data-theme="light"]){--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}}
+:root[data-theme="dark"]{--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}
+body{margin:0;background:var(--bg);color:var(--fg);font:15px/1.5 system-ui,sans-serif}
+main{max-width:1040px;margin:0 auto;padding:24px 16px 64px}
+h1{font-size:24px;margin:0 0 4px}h2{font-size:18px;margin:36px 0 8px}h3{font-size:14px;margin:0 0 8px;font-family:ui-monospace,monospace}
+.meta,.note{color:var(--muted);margin:0 0 12px;overflow-wrap:anywhere}
+.stats{display:grid;grid-template-columns:repeat(auto-fit,minmax(150px,1fr));gap:12px;margin:20px 0}
+.stat{border:1px solid var(--line);border-radius:8px;padding:12px;display:flex;flex-direction:column}
+.stat .v{font-size:26px;font-weight:600}.stat .k{color:var(--muted);font-size:13px}
+.findings{border-radius:8px;padding:4px 16px;border:1px solid var(--line)}
+.findings.ok{border-left:4px solid var(--ok)}.findings.warn{border-left:4px solid var(--warn)}
+.findings h2{margin:12px 0 4px;font-size:16px}
+.stack{border:1px solid var(--line);border-radius:8px;padding:12px;margin:12px 0}
+.row{display:flex;align-items:center;gap:12px;margin:6px 0}
+.lab{flex:0 0 110px;color:var(--muted);font-size:13px}
+.cells{display:flex;flex-wrap:wrap;gap:2px}
+.cells i{width:12px;height:12px;border-radius:2px;display:block}
+.bar{flex:1;display:flex;gap:2px;height:16px}.bar i{display:block;border-radius:2px;min-width:2px}
+table{border-collapse:collapse;width:100%;font-size:13px;margin:8px 0;display:block;overflow-x:auto}
+th,td{text-align:left;padding:4px 10px 4px 0;border-bottom:1px solid var(--line);vertical-align:top}
+th{color:var(--muted);font-weight:500}
+td:first-child{font-family:ui-monospace,monospace;overflow-wrap:anywhere}
+td i{display:inline-block;width:10px;height:10px;border-radius:2px;margin-right:6px}
+tr.bad td{color:var(--warn)}
+.c0{background:var(--s0)}.c1{background:var(--s1)}.c2{background:var(--s2)}.c3{background:var(--s3)}.c4{background:var(--s4)}.c5{background:var(--s5)}.c6{background:var(--s6)}.c7{background:var(--s7)}.c8{background:var(--s8)}
+</style></head><body><main>
+<h1>Shrunk architecture</h1><p class="meta">huggingface-mic/Mixtral-8x7B-Instruct-v0.1-mic@9c86dd7f51c866a4511616932943f95f6d163e66 ← mistralai/Mixtral-8x7B-Instruct-v0.1@eba92302a2861cdc0098cc54bc9f17cb2c47eb61 · MixtralForCausalLM · random weights</p>
+<div class="stats"><div class="stat"><span class="k">Original</span><span class="v">46.7B</span><span class="k">parameters</span></div><div class="stat"><span class="k">Fixture</span><span class="v">4.3M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Layer stacks</span><span class="v">1</span><span class="k">compared</span></div><div class="stat"><span class="k">Findings</span><span class="v">0</span><span class="k">to review</span></div></div><section class="findings ok"><h2>✓ Same architecture, smaller</h2><p>Every module class, layer kind and checked property of the original is in the fixture.</p></section><h2>Layer stacks</h2><p class="note">One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.</p><div class="stack"><h3>model.layers</h3><div class="row"><span class="lab">Original · 32</span><div class="cells"><i class="c0" title="layer 0: DecoderLayer"></i><i class="c0" title="layer 1: DecoderLayer"></i><i class="c0" title="layer 2: DecoderLayer"></i><i class="c0" title="layer 3: DecoderLayer"></i><i class="c0" title="layer 4: DecoderLayer"></i><i class="c0" title="layer 5: DecoderLayer"></i><i class="c0" title="layer 6: DecoderLayer"></i><i class="c0" title="layer 7: DecoderLayer"></i><i class="c0" title="layer 8: DecoderLayer"></i><i class="c0" title="layer 9: DecoderLayer"></i><i class="c0" title="layer 10: DecoderLayer"></i><i class="c0" title="layer 11: DecoderLayer"></i><i class="c0" title="layer 12: DecoderLayer"></i><i class="c0" title="layer 13: DecoderLayer"></i><i class="c0" title="layer 14: DecoderLayer"></i><i class="c0" title="layer 15: DecoderLayer"></i><i class="c0" title="layer 16: DecoderLayer"></i><i class="c0" title="layer 17: DecoderLayer"></i><i class="c0" title="layer 18: DecoderLayer"></i><i class="c0" title="layer 19: DecoderLayer"></i><i class="c0" title="layer 20: DecoderLayer"></i><i class="c0" title="layer 21: DecoderLayer"></i><i class="c0" title="layer 22: DecoderLayer"></i><i class="c0" title="layer 23: DecoderLayer"></i><i class="c0" title="layer 24: DecoderLayer"></i><i class="c0" title="layer 25: DecoderLayer"></i><i class="c0" title="layer 26: DecoderLayer"></i><i class="c0" title="layer 27: DecoderLayer"></i><i class="c0" title="layer 28: DecoderLayer"></i><i class="c0" title="layer 29: DecoderLayer"></i><i class="c0" title="layer 30: DecoderLayer"></i><i class="c0" title="layer 31: DecoderLayer"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: DecoderLayer"></i><i class="c0" title="layer 1: DecoderLayer"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>DecoderLayer</td><td>32</td><td>2</td><td>✓ kept</td></tr></table></div><h2>Where the parameters are</h2><p class="note">Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.</p><div class="row"><span class="lab">Original</span><div class="bar"><i class="c0" style="flex:0.994387" title="model.layers: 46.4B (99.4%)"></i><i class="c1" style="flex:0.002807" title="lm_head: 131.1M (0.3%)"></i><i class="c2" style="flex:0.002807" title="model.embed_tokens: 131.1M (0.3%)"></i><i class="c3" style="flex:0.000000" title="model.norm: 4.1K (0.0%)"></i></div></div><div class="row"><span class="lab">Fixture</span><div class="bar"><i class="c0" style="flex:0.051401" title="model.layers: 222.0K (5.1%)"></i><i class="c1" style="flex:0.474292" title="lm_head: 2.0M (47.4%)"></i><i class="c2" style="flex:0.474292" title="model.embed_tokens: 2.0M (47.4%)"></i><i class="c3" style="flex:0.000015" title="model.norm: 64 (0.0%)"></i></div></div><table class="legend"><tr><th>Component</th><th>Original</th><th>Share</th><th>Fixture</th><th>Share</th></tr><tr><td><i class="c0"></i>model.layers</td><td>46.4B</td><td>99.4%</td><td>222.0K</td><td>5.1%</td></tr><tr><td><i class="c1"></i>lm_head</td><td>131.1M</td><td>0.3%</td><td>2.0M</td><td>47.4%</td></tr><tr><td><i class="c2"></i>model.embed_tokens</td><td>131.1M</td><td>0.3%</td><td>2.0M</td><td>47.4%</td></tr><tr><td><i class="c3"></i>model.norm</td><td>4.1K</td><td>0.0%</td><td>64</td><td>0.0%</td></tr></table><h2>Preserved properties</h2><table><tr><th>Config</th><th>Property</th><th>Original</th><th>Fixture</th><th></th></tr><tr class=""><td>model</td><td>attention style</td><td>GQA (32 heads, 8 KV)</td><td>GQA (4 heads, 2 KV)</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>MoE routing</td><td>top-2 of 8 experts</td><td>top-2 of 4 experts</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>RoPE type</td><td>default</td><td>default</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>model_type</td><td>mixtral</td><td>mixtral</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>vocab_size</td><td>32000</td><td>32000</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>tie_word_embeddings</td><td>False</td><td>False</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>hidden_act</td><td>silu</td><td>silu</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>max_position_embeddings</td><td>32768</td><td>32768</td><td>✓ kept</td></tr></table><h2>Module classes</h2><p class="note">Instances of each class in the built model (prefix Mixtral omitted).</p><table><tr><th>Class</th><th>Original</th><th>Fixture</th></tr><tr class=""><td>Linear</td><td>129</td><td>9</td></tr><tr class=""><td>RMSNorm</td><td>65</td><td>5</td></tr><tr class=""><td>Attention</td><td>32</td><td>2</td></tr><tr class=""><td>DecoderLayer</td><td>32</td><td>2</td></tr><tr class=""><td>Experts</td><td>32</td><td>2</td></tr><tr class=""><td>SparseMoeBlock</td><td>32</td><td>2</td></tr><tr class=""><td>TopKRouter</td><td>32</td><td>2</td></tr><tr class=""><td>SiLUActivation</td><td>32</td><td>2</td></tr><tr class=""><td>Embedding</td><td>1</td><td>1</td></tr><tr class=""><td>ForCausalLM</td><td>1</td><td>1</td></tr><tr class=""><td>Model</td><td>1</td><td>1</td></tr><tr class=""><td>RotaryEmbedding</td><td>1</td><td>1</td></tr><tr class=""><td>ModuleList</td><td>1</td><td>1</td></tr></table><h2>Config changes</h2><p class="note">6 fields changed, 27 unchanged.</p><table><tr><th>Field</th><th>Original</th><th>Fixture</th></tr><tr><td>hidden_size</td><td>4096</td><td>64</td></tr><tr><td>intermediate_size</td><td>14336</td><td>128</td></tr><tr><td>num_attention_heads</td><td>32</td><td>4</td></tr><tr><td>num_hidden_layers</td><td>32</td><td>2</td></tr><tr><td>num_key_value_heads</td><td>8</td><td>2</td></tr><tr><td>num_local_experts</td><td>8</td><td>4</td></tr></table>
+</main></body></html>

shrink/NVIDIA-Nemotron-3-Nano-4B-BF16-mic.html

@@ -0,0 +1,34 @@
+<!doctype html>
+<html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
+<title>Shrunk architecture</title>
+<style>
+:root{--bg:#fcfcfb;--fg:#0b0b0b;--muted:#52514e;--line:#e4e3df;--ok:#1d7a43;--warn:#b35c00;--s0:#2a78d6;--s1:#eb6834;--s2:#1baf7a;--s3:#eda100;--s4:#e87ba4;--s5:#008300;--s6:#6250d6;--s7:#e34948;--s8:#8a8984;}
+@media (prefers-color-scheme:dark){:root:not([data-theme="light"]){--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}}
+:root[data-theme="dark"]{--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}
+body{margin:0;background:var(--bg);color:var(--fg);font:15px/1.5 system-ui,sans-serif}
+main{max-width:1040px;margin:0 auto;padding:24px 16px 64px}
+h1{font-size:24px;margin:0 0 4px}h2{font-size:18px;margin:36px 0 8px}h3{font-size:14px;margin:0 0 8px;font-family:ui-monospace,monospace}
+.meta,.note{color:var(--muted);margin:0 0 12px;overflow-wrap:anywhere}
+.stats{display:grid;grid-template-columns:repeat(auto-fit,minmax(150px,1fr));gap:12px;margin:20px 0}
+.stat{border:1px solid var(--line);border-radius:8px;padding:12px;display:flex;flex-direction:column}
+.stat .v{font-size:26px;font-weight:600}.stat .k{color:var(--muted);font-size:13px}
+.findings{border-radius:8px;padding:4px 16px;border:1px solid var(--line)}
+.findings.ok{border-left:4px solid var(--ok)}.findings.warn{border-left:4px solid var(--warn)}
+.findings h2{margin:12px 0 4px;font-size:16px}
+.stack{border:1px solid var(--line);border-radius:8px;padding:12px;margin:12px 0}
+.row{display:flex;align-items:center;gap:12px;margin:6px 0}
+.lab{flex:0 0 110px;color:var(--muted);font-size:13px}
+.cells{display:flex;flex-wrap:wrap;gap:2px}
+.cells i{width:12px;height:12px;border-radius:2px;display:block}
+.bar{flex:1;display:flex;gap:2px;height:16px}.bar i{display:block;border-radius:2px;min-width:2px}
+table{border-collapse:collapse;width:100%;font-size:13px;margin:8px 0;display:block;overflow-x:auto}
+th,td{text-align:left;padding:4px 10px 4px 0;border-bottom:1px solid var(--line);vertical-align:top}
+th{color:var(--muted);font-weight:500}
+td:first-child{font-family:ui-monospace,monospace;overflow-wrap:anywhere}
+td i{display:inline-block;width:10px;height:10px;border-radius:2px;margin-right:6px}
+tr.bad td{color:var(--warn)}
+.c0{background:var(--s0)}.c1{background:var(--s1)}.c2{background:var(--s2)}.c3{background:var(--s3)}.c4{background:var(--s4)}.c5{background:var(--s5)}.c6{background:var(--s6)}.c7{background:var(--s7)}.c8{background:var(--s8)}
+</style></head><body><main>
+<h1>Shrunk architecture</h1><p class="meta">huggingface-mic/NVIDIA-Nemotron-3-Nano-4B-BF16-mic@9e2cf8579314aef2d07fb651a0e47989c49ce98d ← nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16@dfaf35de3e30f1867dd8dbc38a7fc9fb52d3914f · NemotronHForCausalLM · random weights · built before the shrink rule (no recorded recipe)</p>
+<div class="stats"><div class="stat"><span class="k">Original</span><span class="v">4.0B</span><span class="k">parameters</span></div><div class="stat"><span class="k">Fixture</span><span class="v">49.5M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Layer stacks</span><span class="v">1</span><span class="k">compared</span></div><div class="stat"><span class="k">Findings</span><span class="v">0</span><span class="k">to review</span></div></div><section class="findings ok"><h2>✓ Same architecture, smaller</h2><p>Every module class, layer kind and checked property of the original is in the fixture.</p></section><h2>Layer stacks</h2><p class="note">One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.</p><div class="stack"><h3>model.layers</h3><div class="row"><span class="lab">Original · 42</span><div class="cells"><i class="c0" title="layer 0: linear_attention, Conv1d, Mamba2Mixer …"></i><i class="c1" title="layer 1: MLP, ReLUSquaredActivation"></i><i class="c0" title="layer 2: linear_attention, Conv1d, Mamba2Mixer …"></i><i class="c1" title="layer 3: MLP, ReLUSquaredActivation"></i><i class="c0" title="layer 4: linear_attention, Conv1d, Mamba2Mixer …"></i><i class="c1" title="layer 5: MLP, ReLUSquaredActivation"></i><i class="c0" title="layer 6: linear_attention, Conv1d, Mamba2Mixer …"></i><i class="c0" title="layer 7: linear_attention, Conv1d, Mamba2Mixer …"></i><i class="c1" title="layer 8: MLP, ReLUSquaredActivation"></i><i class="c0" title="layer 9: linear_attention, Conv1d, Mamba2Mixer …"></i><i class="c1" title="layer 10: MLP, ReLUSquaredActivation"></i><i class="c0" title="layer 11: linear_attention, Conv1d, Mamba2Mixer …"></i><i class="c2" title="layer 12: is_causal=True, Attention"></i><i class="c1" title="layer 13: MLP, ReLUSquaredActivation"></i><i class="c0" title="layer 14: linear_attention, Conv1d, Mamba2Mixer …"></i><i class="c1" title="layer 15: MLP, ReLUSquaredActivation"></i><i class="c0" title="layer 16: linear_attention, Conv1d, Mamba2Mixer …"></i><i class="c2" title="layer 17: is_causal=True, Attention"></i><i class="c1" title="layer 18: MLP, ReLUSquaredActivation"></i><i class="c0" title="layer 19: linear_attention, Conv1d, Mamba2Mixer …"></i><i class="c1" title="layer 20: MLP, ReLUSquaredActivation"></i><i class="c0" title="layer 21: linear_attention, Conv1d, Mamba2Mixer …"></i><i class="c1" title="layer 22: MLP, ReLUSquaredActivation"></i><i class="c0" title="layer 23: linear_attention, Conv1d, Mamba2Mixer …"></i><i class="c2" title="layer 24: is_causal=True, Attention"></i><i class="c1" title="layer 25: MLP, ReLUSquaredActivation"></i><i class="c0" title="layer 26: linear_attention, Conv1d, Mamba2Mixer …"></i><i class="c1" title="layer 27: MLP, ReLUSquaredActivation"></i><i class="c0" title="layer 28: linear_attention, Conv1d, Mamba2Mixer …"></i><i class="c1" title="layer 29: MLP, ReLUSquaredActivation"></i><i class="c0" title="layer 30: linear_attention, Conv1d, Mamba2Mixer …"></i><i class="c0" title="layer 31: linear_attention, Conv1d, Mamba2Mixer …"></i><i class="c2" title="layer 32: is_causal=True, Attention"></i><i class="c1" title="layer 33: MLP, ReLUSquaredActivation"></i><i class="c0" title="layer 34: linear_attention, Conv1d, Mamba2Mixer …"></i><i class="c0" title="layer 35: linear_attention, Conv1d, Mamba2Mixer …"></i><i class="c0" title="layer 36: linear_attention, Conv1d, Mamba2Mixer …"></i><i class="c1" title="layer 37: MLP, ReLUSquaredActivation"></i><i class="c0" title="layer 38: linear_attention, Conv1d, Mamba2Mixer …"></i><i class="c1" title="layer 39: MLP, ReLUSquaredActivation"></i><i class="c0" title="layer 40: linear_attention, Conv1d, Mamba2Mixer …"></i><i class="c1" title="layer 41: MLP, ReLUSquaredActivation"></i></div></div><div class="row"><span class="lab">Fixture · 42</span><div class="cells"><i class="c0" title="layer 0: linear_attention, Conv1d, Mamba2Mixer …"></i><i class="c1" title="layer 1: MLP, ReLUSquaredActivation"></i><i class="c0" title="layer 2: linear_attention, Conv1d, Mamba2Mixer …"></i><i class="c1" title="layer 3: MLP, ReLUSquaredActivation"></i><i class="c0" title="layer 4: linear_attention, Conv1d, Mamba2Mixer …"></i><i class="c1" title="layer 5: MLP, ReLUSquaredActivation"></i><i class="c0" title="layer 6: linear_attention, Conv1d, Mamba2Mixer …"></i><i class="c0" title="layer 7: linear_attention, Conv1d, Mamba2Mixer …"></i><i class="c1" title="layer 8: MLP, ReLUSquaredActivation"></i><i class="c0" title="layer 9: linear_attention, Conv1d, Mamba2Mixer …"></i><i class="c1" title="layer 10: MLP, ReLUSquaredActivation"></i><i class="c0" title="layer 11: linear_attention, Conv1d, Mamba2Mixer …"></i><i class="c2" title="layer 12: is_causal=True, Attention"></i><i class="c1" title="layer 13: MLP, ReLUSquaredActivation"></i><i class="c0" title="layer 14: linear_attention, Conv1d, Mamba2Mixer …"></i><i class="c1" title="layer 15: MLP, ReLUSquaredActivation"></i><i class="c0" title="layer 16: linear_attention, Conv1d, Mamba2Mixer …"></i><i class="c2" title="layer 17: is_causal=True, Attention"></i><i class="c1" title="layer 18: MLP, ReLUSquaredActivation"></i><i class="c0" title="layer 19: linear_attention, Conv1d, Mamba2Mixer …"></i><i class="c1" title="layer 20: MLP, ReLUSquaredActivation"></i><i class="c0" title="layer 21: linear_attention, Conv1d, Mamba2Mixer …"></i><i class="c1" title="layer 22: MLP, ReLUSquaredActivation"></i><i class="c0" title="layer 23: linear_attention, Conv1d, Mamba2Mixer …"></i><i class="c2" title="layer 24: is_causal=True, Attention"></i><i class="c1" title="layer 25: MLP, ReLUSquaredActivation"></i><i class="c0" title="layer 26: linear_attention, Conv1d, Mamba2Mixer …"></i><i class="c1" title="layer 27: MLP, ReLUSquaredActivation"></i><i class="c0" title="layer 28: linear_attention, Conv1d, Mamba2Mixer …"></i><i class="c1" title="layer 29: MLP, ReLUSquaredActivation"></i><i class="c0" title="layer 30: linear_attention, Conv1d, Mamba2Mixer …"></i><i class="c0" title="layer 31: linear_attention, Conv1d, Mamba2Mixer …"></i><i class="c2" title="layer 32: is_causal=True, Attention"></i><i class="c1" title="layer 33: MLP, ReLUSquaredActivation"></i><i class="c0" title="layer 34: linear_attention, Conv1d, Mamba2Mixer …"></i><i class="c0" title="layer 35: linear_attention, Conv1d, Mamba2Mixer …"></i><i class="c0" title="layer 36: linear_attention, Conv1d, Mamba2Mixer …"></i><i class="c1" title="layer 37: MLP, ReLUSquaredActivation"></i><i class="c0" title="layer 38: linear_attention, Conv1d, Mamba2Mixer …"></i><i class="c1" title="layer 39: MLP, ReLUSquaredActivation"></i><i class="c0" title="layer 40: linear_attention, Conv1d, Mamba2Mixer …"></i><i class="c1" title="layer 41: MLP, ReLUSquaredActivation"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>linear_attention, Conv1d, Mamba2Mixer …</td><td>21</td><td>21</td><td>✓ kept</td></tr><tr><td><i class="c1"></i>MLP, ReLUSquaredActivation</td><td>17</td><td>17</td><td>✓ kept</td></tr><tr><td><i class="c2"></i>is_causal=True, Attention</td><td>4</td><td>4</td><td>✓ kept</td></tr></table></div><h2>Where the parameters are</h2><p class="note">Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.</p><div class="row"><span class="lab">Original</span><div class="bar"><i class="c0" style="flex:0.793111" title="model.layers: 3.2B (79.3%)"></i><i class="c1" style="flex:0.103444" title="lm_head: 411.0M (10.3%)"></i><i class="c2" style="flex:0.103444" title="model.embeddings: 411.0M (10.3%)"></i><i class="c3" style="flex:0.000001" title="model.norm_f: 3.1K (0.0%)"></i></div></div><div class="row"><span class="lab">Fixture</span><div class="bar"><i class="c0" style="flex:0.661387" title="model.layers: 32.8M (66.1%)"></i><i class="c1" style="flex:0.169306" title="lm_head: 8.4M (16.9%)"></i><i class="c2" style="flex:0.169306" title="model.embeddings: 8.4M (16.9%)"></i><i class="c3" style="flex:0.000001" title="model.norm_f: 64 (0.0%)"></i></div></div><table class="legend"><tr><th>Component</th><th>Original</th><th>Share</th><th>Fixture</th><th>Share</th></tr><tr><td><i class="c0"></i>model.layers</td><td>3.2B</td><td>79.3%</td><td>32.8M</td><td>66.1%</td></tr><tr><td><i class="c1"></i>lm_head</td><td>411.0M</td><td>10.3%</td><td>8.4M</td><td>16.9%</td></tr><tr><td><i class="c2"></i>model.embeddings</td><td>411.0M</td><td>10.3%</td><td>8.4M</td><td>16.9%</td></tr><tr><td><i class="c3"></i>model.norm_f</td><td>3.1K</td><td>0.0%</td><td>64</td><td>0.0%</td></tr></table><h2>Preserved properties</h2><table><tr><th>Config</th><th>Property</th><th>Original</th><th>Fixture</th><th></th></tr><tr class=""><td>model</td><td>attention style</td><td>GQA (40 heads, 8 KV)</td><td>GQA (4 heads, 2 KV)</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>MoE routing</td><td>top-2 of 8 experts</td><td>top-2 of 4 experts</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>model_type</td><td>nemotron_h</td><td>nemotron_h</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>vocab_size</td><td>131072</td><td>131072</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>tie_word_embeddings</td><td>False</td><td>False</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>max_position_embeddings</td><td>262144</td><td>262144</td><td>✓ kept</td></tr></table><h2>Module classes</h2><p class="note">Instances of each class in the built model (prefix NemotronH omitted).</p><table><tr><th>Class</th><th>Original</th><th>Fixture</th></tr><tr class=""><td>Linear</td><td>93</td><td>93</td></tr><tr class=""><td>RMSNorm</td><td>43</td><td>43</td></tr><tr class=""><td>Block</td><td>42</td><td>42</td></tr><tr class=""><td>Conv1d</td><td>21</td><td>21</td></tr><tr class=""><td>Mamba2Mixer</td><td>21</td><td>21</td></tr><tr class=""><td>RMSNormGated</td><td>21</td><td>21</td></tr><tr class=""><td>SiLUActivation</td><td>21</td><td>21</td></tr><tr class=""><td>MLP</td><td>17</td><td>17</td></tr><tr class=""><td>ReLUSquaredActivation</td><td>17</td><td>17</td></tr><tr class=""><td>Attention</td><td>4</td><td>4</td></tr><tr class=""><td>Embedding</td><td>1</td><td>1</td></tr><tr class=""><td>ModuleList</td><td>1</td><td>1</td></tr><tr class=""><td>ForCausalLM</td><td>1</td><td>1</td></tr><tr class=""><td>Model</td><td>1</td><td>1</td></tr></table><h2>Config changes</h2><p class="note">9 fields changed, 58 unchanged.</p><table><tr><th>Field</th><th>Original</th><th>Fixture</th></tr><tr><td>chunk_size</td><td>256</td><td>8</td></tr><tr><td>head_dim</td><td>128</td><td>16</td></tr><tr><td>hidden_size</td><td>3136</td><td>64</td></tr><tr><td>intermediate_size</td><td>12544</td><td>128</td></tr><tr><td>moe_intermediate_size</td><td>7688</td><td>32</td></tr><tr><td>n_routed_experts</td><td>8</td><td>4</td></tr><tr><td>num_attention_heads</td><td>40</td><td>4</td></tr><tr><td>num_key_value_heads</td><td>8</td><td>2</td></tr><tr><td>ssm_state_size</td><td>128</td><td>16</td></tr></table>
+</main></body></html>

shrink/Nemotron-3-Embed-1B-BF16-mic.html

@@ -0,0 +1,34 @@
+<!doctype html>
+<html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
+<title>Shrunk architecture</title>
+<style>
+:root{--bg:#fcfcfb;--fg:#0b0b0b;--muted:#52514e;--line:#e4e3df;--ok:#1d7a43;--warn:#b35c00;--s0:#2a78d6;--s1:#eb6834;--s2:#1baf7a;--s3:#eda100;--s4:#e87ba4;--s5:#008300;--s6:#6250d6;--s7:#e34948;--s8:#8a8984;}
+@media (prefers-color-scheme:dark){:root:not([data-theme="light"]){--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}}
+:root[data-theme="dark"]{--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}
+body{margin:0;background:var(--bg);color:var(--fg);font:15px/1.5 system-ui,sans-serif}
+main{max-width:1040px;margin:0 auto;padding:24px 16px 64px}
+h1{font-size:24px;margin:0 0 4px}h2{font-size:18px;margin:36px 0 8px}h3{font-size:14px;margin:0 0 8px;font-family:ui-monospace,monospace}
+.meta,.note{color:var(--muted);margin:0 0 12px;overflow-wrap:anywhere}
+.stats{display:grid;grid-template-columns:repeat(auto-fit,minmax(150px,1fr));gap:12px;margin:20px 0}
+.stat{border:1px solid var(--line);border-radius:8px;padding:12px;display:flex;flex-direction:column}
+.stat .v{font-size:26px;font-weight:600}.stat .k{color:var(--muted);font-size:13px}
+.findings{border-radius:8px;padding:4px 16px;border:1px solid var(--line)}
+.findings.ok{border-left:4px solid var(--ok)}.findings.warn{border-left:4px solid var(--warn)}
+.findings h2{margin:12px 0 4px;font-size:16px}
+.stack{border:1px solid var(--line);border-radius:8px;padding:12px;margin:12px 0}
+.row{display:flex;align-items:center;gap:12px;margin:6px 0}
+.lab{flex:0 0 110px;color:var(--muted);font-size:13px}
+.cells{display:flex;flex-wrap:wrap;gap:2px}
+.cells i{width:12px;height:12px;border-radius:2px;display:block}
+.bar{flex:1;display:flex;gap:2px;height:16px}.bar i{display:block;border-radius:2px;min-width:2px}
+table{border-collapse:collapse;width:100%;font-size:13px;margin:8px 0;display:block;overflow-x:auto}
+th,td{text-align:left;padding:4px 10px 4px 0;border-bottom:1px solid var(--line);vertical-align:top}
+th{color:var(--muted);font-weight:500}
+td:first-child{font-family:ui-monospace,monospace;overflow-wrap:anywhere}
+td i{display:inline-block;width:10px;height:10px;border-radius:2px;margin-right:6px}
+tr.bad td{color:var(--warn)}
+.c0{background:var(--s0)}.c1{background:var(--s1)}.c2{background:var(--s2)}.c3{background:var(--s3)}.c4{background:var(--s4)}.c5{background:var(--s5)}.c6{background:var(--s6)}.c7{background:var(--s7)}.c8{background:var(--s8)}
+</style></head><body><main>
+<h1>Shrunk architecture</h1><p class="meta">huggingface-mic/Nemotron-3-Embed-1B-BF16-mic@645864f9759d320c610786a2270ecb18e982534a ← nvidia/Nemotron-3-Embed-1B-BF16@c0c9fea93ea424587517f2c59e20db9f1d6bf615 · Ministral3Model · random weights</p>
+<div class="stats"><div class="stat"><span class="k">Original</span><span class="v">1.1B</span><span class="k">parameters</span></div><div class="stat"><span class="k">Fixture</span><span class="v">8.5M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Layer stacks</span><span class="v">1</span><span class="k">compared</span></div><div class="stat"><span class="k">Findings</span><span class="v">0</span><span class="k">to review</span></div></div><section class="findings ok"><h2>✓ Same architecture, smaller</h2><p>Every module class, layer kind and checked property of the original is in the fixture.</p></section><h2>Layer stacks</h2><p class="note">One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.</p><div class="stack"><h3>layers</h3><div class="row"><span class="lab">Original · 16</span><div class="cells"><i class="c0" title="layer 0: DecoderLayer"></i><i class="c0" title="layer 1: DecoderLayer"></i><i class="c0" title="layer 2: DecoderLayer"></i><i class="c0" title="layer 3: DecoderLayer"></i><i class="c0" title="layer 4: DecoderLayer"></i><i class="c0" title="layer 5: DecoderLayer"></i><i class="c0" title="layer 6: DecoderLayer"></i><i class="c0" title="layer 7: DecoderLayer"></i><i class="c0" title="layer 8: DecoderLayer"></i><i class="c0" title="layer 9: DecoderLayer"></i><i class="c0" title="layer 10: DecoderLayer"></i><i class="c0" title="layer 11: DecoderLayer"></i><i class="c0" title="layer 12: DecoderLayer"></i><i class="c0" title="layer 13: DecoderLayer"></i><i class="c0" title="layer 14: DecoderLayer"></i><i class="c0" title="layer 15: DecoderLayer"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: DecoderLayer"></i><i class="c0" title="layer 1: DecoderLayer"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>DecoderLayer</td><td>16</td><td>2</td><td>✓ kept</td></tr></table></div><h2>Where the parameters are</h2><p class="note">Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.</p><div class="row"><span class="lab">Original</span><div class="bar"><i class="c0" style="flex:0.764718" title="layers: 872.5M (76.5%)"></i><i class="c1" style="flex:0.235280" title="embed_tokens: 268.4M (23.5%)"></i><i class="c2" style="flex:0.000002" title="norm: 2.0K (0.0%)"></i></div></div><div class="row"><span class="lab">Fixture</span><div class="bar"><i class="c0" style="flex:0.008742" title="layers: 74.0K (0.9%)"></i><i class="c1" style="flex:0.991250" title="embed_tokens: 8.4M (99.1%)"></i><i class="c2" style="flex:0.000008" title="norm: 64 (0.0%)"></i></div></div><table class="legend"><tr><th>Component</th><th>Original</th><th>Share</th><th>Fixture</th><th>Share</th></tr><tr><td><i class="c0"></i>layers</td><td>872.5M</td><td>76.5%</td><td>74.0K</td><td>0.9%</td></tr><tr><td><i class="c1"></i>embed_tokens</td><td>268.4M</td><td>23.5%</td><td>8.4M</td><td>99.1%</td></tr><tr><td><i class="c2"></i>norm</td><td>2.0K</td><td>0.0%</td><td>64</td><td>0.0%</td></tr></table><h2>Preserved properties</h2><table><tr><th>Config</th><th>Property</th><th>Original</th><th>Fixture</th><th></th></tr><tr class=""><td>model</td><td>attention style</td><td>GQA (24 heads, 8 KV)</td><td>GQA (4 heads, 2 KV)</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>RoPE type</td><td>yarn</td><td>yarn</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>model_type</td><td>ministral3</td><td>ministral3</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>vocab_size</td><td>131072</td><td>131072</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>tie_word_embeddings</td><td>True</td><td>True</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>hidden_act</td><td>silu</td><td>silu</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>max_position_embeddings</td><td>262144</td><td>262144</td><td>✓ kept</td></tr></table><h2>Module classes</h2><p class="note">Instances of each class in the built model (prefix Ministral3 omitted).</p><table><tr><th>Class</th><th>Original</th><th>Fixture</th></tr><tr class=""><td>Linear</td><td>112</td><td>14</td></tr><tr class=""><td>RMSNorm</td><td>33</td><td>5</td></tr><tr class=""><td>Attention</td><td>16</td><td>2</td></tr><tr class=""><td>DecoderLayer</td><td>16</td><td>2</td></tr><tr class=""><td>MLP</td><td>16</td><td>2</td></tr><tr class=""><td>SiLUActivation</td><td>16</td><td>2</td></tr><tr class=""><td>Embedding</td><td>1</td><td>1</td></tr><tr class=""><td>Model</td><td>1</td><td>1</td></tr><tr class=""><td>RotaryEmbedding</td><td>1</td><td>1</td></tr><tr class=""><td>ModuleList</td><td>1</td><td>1</td></tr></table><h2>Config changes</h2><p class="note">6 fields changed, 26 unchanged.</p><table><tr><th>Field</th><th>Original</th><th>Fixture</th></tr><tr><td>head_dim</td><td>128</td><td>16</td></tr><tr><td>hidden_size</td><td>2048</td><td>64</td></tr><tr><td>intermediate_size</td><td>6144</td><td>128</td></tr><tr><td>num_attention_heads</td><td>24</td><td>4</td></tr><tr><td>num_hidden_layers</td><td>16</td><td>2</td></tr><tr><td>num_key_value_heads</td><td>8</td><td>2</td></tr></table>
+</main></body></html>

shrink/OLMo-1B-hf-mic.html

@@ -0,0 +1,34 @@
+<!doctype html>
+<html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
+<title>Shrunk architecture</title>
+<style>
+:root{--bg:#fcfcfb;--fg:#0b0b0b;--muted:#52514e;--line:#e4e3df;--ok:#1d7a43;--warn:#b35c00;--s0:#2a78d6;--s1:#eb6834;--s2:#1baf7a;--s3:#eda100;--s4:#e87ba4;--s5:#008300;--s6:#6250d6;--s7:#e34948;--s8:#8a8984;}
+@media (prefers-color-scheme:dark){:root:not([data-theme="light"]){--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}}
+:root[data-theme="dark"]{--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}
+body{margin:0;background:var(--bg);color:var(--fg);font:15px/1.5 system-ui,sans-serif}
+main{max-width:1040px;margin:0 auto;padding:24px 16px 64px}
+h1{font-size:24px;margin:0 0 4px}h2{font-size:18px;margin:36px 0 8px}h3{font-size:14px;margin:0 0 8px;font-family:ui-monospace,monospace}
+.meta,.note{color:var(--muted);margin:0 0 12px;overflow-wrap:anywhere}
+.stats{display:grid;grid-template-columns:repeat(auto-fit,minmax(150px,1fr));gap:12px;margin:20px 0}
+.stat{border:1px solid var(--line);border-radius:8px;padding:12px;display:flex;flex-direction:column}
+.stat .v{font-size:26px;font-weight:600}.stat .k{color:var(--muted);font-size:13px}
+.findings{border-radius:8px;padding:4px 16px;border:1px solid var(--line)}
+.findings.ok{border-left:4px solid var(--ok)}.findings.warn{border-left:4px solid var(--warn)}
+.findings h2{margin:12px 0 4px;font-size:16px}
+.stack{border:1px solid var(--line);border-radius:8px;padding:12px;margin:12px 0}
+.row{display:flex;align-items:center;gap:12px;margin:6px 0}
+.lab{flex:0 0 110px;color:var(--muted);font-size:13px}
+.cells{display:flex;flex-wrap:wrap;gap:2px}
+.cells i{width:12px;height:12px;border-radius:2px;display:block}
+.bar{flex:1;display:flex;gap:2px;height:16px}.bar i{display:block;border-radius:2px;min-width:2px}
+table{border-collapse:collapse;width:100%;font-size:13px;margin:8px 0;display:block;overflow-x:auto}
+th,td{text-align:left;padding:4px 10px 4px 0;border-bottom:1px solid var(--line);vertical-align:top}
+th{color:var(--muted);font-weight:500}
+td:first-child{font-family:ui-monospace,monospace;overflow-wrap:anywhere}
+td i{display:inline-block;width:10px;height:10px;border-radius:2px;margin-right:6px}
+tr.bad td{color:var(--warn)}
+.c0{background:var(--s0)}.c1{background:var(--s1)}.c2{background:var(--s2)}.c3{background:var(--s3)}.c4{background:var(--s4)}.c5{background:var(--s5)}.c6{background:var(--s6)}.c7{background:var(--s7)}.c8{background:var(--s8)}
+</style></head><body><main>
+<h1>Shrunk architecture</h1><p class="meta">huggingface-mic/OLMo-1B-hf-mic@87c81bd7d9f93663d9906bbde0737125e4d85c54 ← allenai/OLMo-1B-hf@aee7752d9c08ee4775e9b0091426d8410e8f6a89 · OlmoForCausalLM · random weights</p>
+<div class="stats"><div class="stat"><span class="k">Original</span><span class="v">1.2B</span><span class="k">parameters</span></div><div class="stat"><span class="k">Fixture</span><span class="v">3.3M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Layer stacks</span><span class="v">1</span><span class="k">compared</span></div><div class="stat"><span class="k">Findings</span><span class="v">0</span><span class="k">to review</span></div></div><section class="findings ok"><h2>✓ Same architecture, smaller</h2><p>Every module class, layer kind and checked property of the original is in the fixture.</p></section><h2>Layer stacks</h2><p class="note">One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.</p><div class="stack"><h3>model.layers</h3><div class="row"><span class="lab">Original · 16</span><div class="cells"><i class="c0" title="layer 0: DecoderLayer"></i><i class="c0" title="layer 1: DecoderLayer"></i><i class="c0" title="layer 2: DecoderLayer"></i><i class="c0" title="layer 3: DecoderLayer"></i><i class="c0" title="layer 4: DecoderLayer"></i><i class="c0" title="layer 5: DecoderLayer"></i><i class="c0" title="layer 6: DecoderLayer"></i><i class="c0" title="layer 7: DecoderLayer"></i><i class="c0" title="layer 8: DecoderLayer"></i><i class="c0" title="layer 9: DecoderLayer"></i><i class="c0" title="layer 10: DecoderLayer"></i><i class="c0" title="layer 11: DecoderLayer"></i><i class="c0" title="layer 12: DecoderLayer"></i><i class="c0" title="layer 13: DecoderLayer"></i><i class="c0" title="layer 14: DecoderLayer"></i><i class="c0" title="layer 15: DecoderLayer"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: DecoderLayer"></i><i class="c0" title="layer 1: DecoderLayer"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>DecoderLayer</td><td>16</td><td>2</td><td>✓ kept</td></tr></table></div><h2>Where the parameters are</h2><p class="note">Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.</p><div class="row"><span class="lab">Original</span><div class="bar"><i class="c0" style="flex:0.912453" title="model.layers: 1.1B (91.2%)"></i><i class="c1" style="flex:0.087547" title="model.embed_tokens: 103.0M (8.8%)"></i></div></div><div class="row"><span class="lab">Fixture</span><div class="bar"><i class="c0" style="flex:0.024814" title="model.layers: 81.9K (2.5%)"></i><i class="c1" style="flex:0.975186" title="model.embed_tokens: 3.2M (97.5%)"></i></div></div><table class="legend"><tr><th>Component</th><th>Original</th><th>Share</th><th>Fixture</th><th>Share</th></tr><tr><td><i class="c0"></i>model.layers</td><td>1.1B</td><td>91.2%</td><td>81.9K</td><td>2.5%</td></tr><tr><td><i class="c1"></i>model.embed_tokens</td><td>103.0M</td><td>8.8%</td><td>3.2M</td><td>97.5%</td></tr></table><h2>Preserved properties</h2><table><tr><th>Config</th><th>Property</th><th>Original</th><th>Fixture</th><th></th></tr><tr class=""><td>model</td><td>attention style</td><td>MHA (16 heads, 16 KV)</td><td>MHA (4 heads, 4 KV)</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>RoPE type</td><td>default</td><td>default</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>model_type</td><td>olmo</td><td>olmo</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>vocab_size</td><td>50304</td><td>50304</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>tie_word_embeddings</td><td>True</td><td>True</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>hidden_act</td><td>silu</td><td>silu</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>max_position_embeddings</td><td>2048</td><td>2048</td><td>✓ kept</td></tr></table><h2>Module classes</h2><p class="note">Instances of each class in the built model (prefix Olmo omitted).</p><table><tr><th>Class</th><th>Original</th><th>Fixture</th></tr><tr class=""><td>Linear</td><td>113</td><td>15</td></tr><tr class=""><td>LayerNorm</td><td>33</td><td>5</td></tr><tr class=""><td>Attention</td><td>16</td><td>2</td></tr><tr class=""><td>DecoderLayer</td><td>16</td><td>2</td></tr><tr class=""><td>MLP</td><td>16</td><td>2</td></tr><tr class=""><td>SiLUActivation</td><td>16</td><td>2</td></tr><tr class=""><td>Embedding</td><td>1</td><td>1</td></tr><tr class=""><td>ModuleList</td><td>1</td><td>1</td></tr><tr class=""><td>ForCausalLM</td><td>1</td><td>1</td></tr><tr class=""><td>Model</td><td>1</td><td>1</td></tr><tr class=""><td>RotaryEmbedding</td><td>1</td><td>1</td></tr></table><h2>Config changes</h2><p class="note">5 fields changed, 23 unchanged.</p><table><tr><th>Field</th><th>Original</th><th>Fixture</th></tr><tr><td>hidden_size</td><td>2048</td><td>64</td></tr><tr><td>intermediate_size</td><td>8192</td><td>128</td></tr><tr><td>num_attention_heads</td><td>16</td><td>4</td></tr><tr><td>num_hidden_layers</td><td>16</td><td>2</td></tr><tr><td>num_key_value_heads</td><td>16</td><td>4</td></tr></table>
+</main></body></html>

shrink/OLMo-2-0425-1B-mic.html

@@ -0,0 +1,34 @@
+<!doctype html>
+<html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
+<title>Shrunk architecture</title>
+<style>
+:root{--bg:#fcfcfb;--fg:#0b0b0b;--muted:#52514e;--line:#e4e3df;--ok:#1d7a43;--warn:#b35c00;--s0:#2a78d6;--s1:#eb6834;--s2:#1baf7a;--s3:#eda100;--s4:#e87ba4;--s5:#008300;--s6:#6250d6;--s7:#e34948;--s8:#8a8984;}
+@media (prefers-color-scheme:dark){:root:not([data-theme="light"]){--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}}
+:root[data-theme="dark"]{--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}
+body{margin:0;background:var(--bg);color:var(--fg);font:15px/1.5 system-ui,sans-serif}
+main{max-width:1040px;margin:0 auto;padding:24px 16px 64px}
+h1{font-size:24px;margin:0 0 4px}h2{font-size:18px;margin:36px 0 8px}h3{font-size:14px;margin:0 0 8px;font-family:ui-monospace,monospace}
+.meta,.note{color:var(--muted);margin:0 0 12px;overflow-wrap:anywhere}
+.stats{display:grid;grid-template-columns:repeat(auto-fit,minmax(150px,1fr));gap:12px;margin:20px 0}
+.stat{border:1px solid var(--line);border-radius:8px;padding:12px;display:flex;flex-direction:column}
+.stat .v{font-size:26px;font-weight:600}.stat .k{color:var(--muted);font-size:13px}
+.findings{border-radius:8px;padding:4px 16px;border:1px solid var(--line)}
+.findings.ok{border-left:4px solid var(--ok)}.findings.warn{border-left:4px solid var(--warn)}
+.findings h2{margin:12px 0 4px;font-size:16px}
+.stack{border:1px solid var(--line);border-radius:8px;padding:12px;margin:12px 0}
+.row{display:flex;align-items:center;gap:12px;margin:6px 0}
+.lab{flex:0 0 110px;color:var(--muted);font-size:13px}
+.cells{display:flex;flex-wrap:wrap;gap:2px}
+.cells i{width:12px;height:12px;border-radius:2px;display:block}
+.bar{flex:1;display:flex;gap:2px;height:16px}.bar i{display:block;border-radius:2px;min-width:2px}
+table{border-collapse:collapse;width:100%;font-size:13px;margin:8px 0;display:block;overflow-x:auto}
+th,td{text-align:left;padding:4px 10px 4px 0;border-bottom:1px solid var(--line);vertical-align:top}
+th{color:var(--muted);font-weight:500}
+td:first-child{font-family:ui-monospace,monospace;overflow-wrap:anywhere}
+td i{display:inline-block;width:10px;height:10px;border-radius:2px;margin-right:6px}
+tr.bad td{color:var(--warn)}
+.c0{background:var(--s0)}.c1{background:var(--s1)}.c2{background:var(--s2)}.c3{background:var(--s3)}.c4{background:var(--s4)}.c5{background:var(--s5)}.c6{background:var(--s6)}.c7{background:var(--s7)}.c8{background:var(--s8)}
+</style></head><body><main>
+<h1>Shrunk architecture</h1><p class="meta">huggingface-mic/OLMo-2-0425-1B-mic@7a6eea1e741066708957df80693b23000613b907 ← allenai/OLMo-2-0425-1B@a1847dff35000b4271fa70afc5db10fd29fedbdf · Olmo2ForCausalLM · random weights</p>
+<div class="stats"><div class="stat"><span class="k">Original</span><span class="v">1.5B</span><span class="k">parameters</span></div><div class="stat"><span class="k">Fixture</span><span class="v">12.9M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Layer stacks</span><span class="v">1</span><span class="k">compared</span></div><div class="stat"><span class="k">Findings</span><span class="v">0</span><span class="k">to review</span></div></div><section class="findings ok"><h2>✓ Same architecture, smaller</h2><p>Every module class, layer kind and checked property of the original is in the fixture.</p></section><h2>Layer stacks</h2><p class="note">One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.</p><div class="stack"><h3>model.layers</h3><div class="row"><span class="lab">Original · 16</span><div class="cells"><i class="c0" title="layer 0: DecoderLayer"></i><i class="c0" title="layer 1: DecoderLayer"></i><i class="c0" title="layer 2: DecoderLayer"></i><i class="c0" title="layer 3: DecoderLayer"></i><i class="c0" title="layer 4: DecoderLayer"></i><i class="c0" title="layer 5: DecoderLayer"></i><i class="c0" title="layer 6: DecoderLayer"></i><i class="c0" title="layer 7: DecoderLayer"></i><i class="c0" title="layer 8: DecoderLayer"></i><i class="c0" title="layer 9: DecoderLayer"></i><i class="c0" title="layer 10: DecoderLayer"></i><i class="c0" title="layer 11: DecoderLayer"></i><i class="c0" title="layer 12: DecoderLayer"></i><i class="c0" title="layer 13: DecoderLayer"></i><i class="c0" title="layer 14: DecoderLayer"></i><i class="c0" title="layer 15: DecoderLayer"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: DecoderLayer"></i><i class="c0" title="layer 1: DecoderLayer"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>DecoderLayer</td><td>16</td><td>2</td><td>✓ kept</td></tr></table></div><h2>Where the parameters are</h2><p class="note">Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.</p><div class="row"><span class="lab">Original</span><div class="bar"><i class="c0" style="flex:0.723187" title="model.layers: 1.1B (72.3%)"></i><i class="c1" style="flex:0.138406" title="model.embed_tokens: 205.5M (13.8%)"></i><i class="c2" style="flex:0.138406" title="lm_head: 205.5M (13.8%)"></i><i class="c3" style="flex:0.000001" title="model.norm: 2.0K (0.0%)"></i></div></div><div class="row"><span class="lab">Fixture</span><div class="bar"><i class="c0" style="flex:0.006376" title="model.layers: 82.4K (0.6%)"></i><i class="c1" style="flex:0.496809" title="model.embed_tokens: 6.4M (49.7%)"></i><i class="c2" style="flex:0.496809" title="lm_head: 6.4M (49.7%)"></i><i class="c3" style="flex:0.000005" title="model.norm: 64 (0.0%)"></i></div></div><table class="legend"><tr><th>Component</th><th>Original</th><th>Share</th><th>Fixture</th><th>Share</th></tr><tr><td><i class="c0"></i>model.layers</td><td>1.1B</td><td>72.3%</td><td>82.4K</td><td>0.6%</td></tr><tr><td><i class="c1"></i>model.embed_tokens</td><td>205.5M</td><td>13.8%</td><td>6.4M</td><td>49.7%</td></tr><tr><td><i class="c2"></i>lm_head</td><td>205.5M</td><td>13.8%</td><td>6.4M</td><td>49.7%</td></tr><tr><td><i class="c3"></i>model.norm</td><td>2.0K</td><td>0.0%</td><td>64</td><td>0.0%</td></tr></table><h2>Preserved properties</h2><table><tr><th>Config</th><th>Property</th><th>Original</th><th>Fixture</th><th></th></tr><tr class=""><td>model</td><td>attention style</td><td>MHA (16 heads, 16 KV)</td><td>MHA (4 heads, 4 KV)</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>RoPE type</td><td>default</td><td>default</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>model_type</td><td>olmo2</td><td>olmo2</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>vocab_size</td><td>100352</td><td>100352</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>tie_word_embeddings</td><td>False</td><td>False</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>hidden_act</td><td>silu</td><td>silu</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>max_position_embeddings</td><td>4096</td><td>4096</td><td>✓ kept</td></tr></table><h2>Module classes</h2><p class="note">Instances of each class in the built model (prefix Olmo2 omitted).</p><table><tr><th>Class</th><th>Original</th><th>Fixture</th></tr><tr class=""><td>Linear</td><td>113</td><td>15</td></tr><tr class=""><td>RMSNorm</td><td>65</td><td>9</td></tr><tr class=""><td>Attention</td><td>16</td><td>2</td></tr><tr class=""><td>DecoderLayer</td><td>16</td><td>2</td></tr><tr class=""><td>MLP</td><td>16</td><td>2</td></tr><tr class=""><td>SiLUActivation</td><td>16</td><td>2</td></tr><tr class=""><td>Embedding</td><td>1</td><td>1</td></tr><tr class=""><td>ModuleList</td><td>1</td><td>1</td></tr><tr class=""><td>ForCausalLM</td><td>1</td><td>1</td></tr><tr class=""><td>Model</td><td>1</td><td>1</td></tr><tr class=""><td>RotaryEmbedding</td><td>1</td><td>1</td></tr></table><h2>Config changes</h2><p class="note">5 fields changed, 22 unchanged.</p><table><tr><th>Field</th><th>Original</th><th>Fixture</th></tr><tr><td>hidden_size</td><td>2048</td><td>64</td></tr><tr><td>intermediate_size</td><td>8192</td><td>128</td></tr><tr><td>num_attention_heads</td><td>16</td><td>4</td></tr><tr><td>num_hidden_layers</td><td>16</td><td>2</td></tr><tr><td>num_key_value_heads</td><td>16</td><td>4</td></tr></table>
+</main></body></html>

shrink/OLMoE-1B-7B-0924-mic.html

@@ -0,0 +1,34 @@
+<!doctype html>
+<html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
+<title>Shrunk architecture</title>
+<style>
+:root{--bg:#fcfcfb;--fg:#0b0b0b;--muted:#52514e;--line:#e4e3df;--ok:#1d7a43;--warn:#b35c00;--s0:#2a78d6;--s1:#eb6834;--s2:#1baf7a;--s3:#eda100;--s4:#e87ba4;--s5:#008300;--s6:#6250d6;--s7:#e34948;--s8:#8a8984;}
+@media (prefers-color-scheme:dark){:root:not([data-theme="light"]){--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}}
+:root[data-theme="dark"]{--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}
+body{margin:0;background:var(--bg);color:var(--fg);font:15px/1.5 system-ui,sans-serif}
+main{max-width:1040px;margin:0 auto;padding:24px 16px 64px}
+h1{font-size:24px;margin:0 0 4px}h2{font-size:18px;margin:36px 0 8px}h3{font-size:14px;margin:0 0 8px;font-family:ui-monospace,monospace}
+.meta,.note{color:var(--muted);margin:0 0 12px;overflow-wrap:anywhere}
+.stats{display:grid;grid-template-columns:repeat(auto-fit,minmax(150px,1fr));gap:12px;margin:20px 0}
+.stat{border:1px solid var(--line);border-radius:8px;padding:12px;display:flex;flex-direction:column}
+.stat .v{font-size:26px;font-weight:600}.stat .k{color:var(--muted);font-size:13px}
+.findings{border-radius:8px;padding:4px 16px;border:1px solid var(--line)}
+.findings.ok{border-left:4px solid var(--ok)}.findings.warn{border-left:4px solid var(--warn)}
+.findings h2{margin:12px 0 4px;font-size:16px}
+.stack{border:1px solid var(--line);border-radius:8px;padding:12px;margin:12px 0}
+.row{display:flex;align-items:center;gap:12px;margin:6px 0}
+.lab{flex:0 0 110px;color:var(--muted);font-size:13px}
+.cells{display:flex;flex-wrap:wrap;gap:2px}
+.cells i{width:12px;height:12px;border-radius:2px;display:block}
+.bar{flex:1;display:flex;gap:2px;height:16px}.bar i{display:block;border-radius:2px;min-width:2px}
+table{border-collapse:collapse;width:100%;font-size:13px;margin:8px 0;display:block;overflow-x:auto}
+th,td{text-align:left;padding:4px 10px 4px 0;border-bottom:1px solid var(--line);vertical-align:top}
+th{color:var(--muted);font-weight:500}
+td:first-child{font-family:ui-monospace,monospace;overflow-wrap:anywhere}
+td i{display:inline-block;width:10px;height:10px;border-radius:2px;margin-right:6px}
+tr.bad td{color:var(--warn)}
+.c0{background:var(--s0)}.c1{background:var(--s1)}.c2{background:var(--s2)}.c3{background:var(--s3)}.c4{background:var(--s4)}.c5{background:var(--s5)}.c6{background:var(--s6)}.c7{background:var(--s7)}.c8{background:var(--s8)}
+</style></head><body><main>
+<h1>Shrunk architecture</h1><p class="meta">huggingface-mic/OLMoE-1B-7B-0924-mic@abd7fb026a08bdb27dfa8ff4d6a1680721f1e85a ← allenai/OLMoE-1B-7B-0924@6d84c48581ece794365f2b8e9cfb043c68ade9c5 · OlmoeForCausalLM · random weights</p>
+<div class="stats"><div class="stat"><span class="k">Original</span><span class="v">6.9B</span><span class="k">parameters</span></div><div class="stat"><span class="k">Fixture</span><span class="v">6.7M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Layer stacks</span><span class="v">1</span><span class="k">compared</span></div><div class="stat"><span class="k">Findings</span><span class="v">0</span><span class="k">to review</span></div></div><section class="findings ok"><h2>✓ Same architecture, smaller</h2><p>Every module class, layer kind and checked property of the original is in the fixture.</p></section><h2>Layer stacks</h2><p class="note">One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.</p><div class="stack"><h3>model.layers</h3><div class="row"><span class="lab">Original · 16</span><div class="cells"><i class="c0" title="layer 0: DecoderLayer"></i><i class="c0" title="layer 1: DecoderLayer"></i><i class="c0" title="layer 2: DecoderLayer"></i><i class="c0" title="layer 3: DecoderLayer"></i><i class="c0" title="layer 4: DecoderLayer"></i><i class="c0" title="layer 5: DecoderLayer"></i><i class="c0" title="layer 6: DecoderLayer"></i><i class="c0" title="layer 7: DecoderLayer"></i><i class="c0" title="layer 8: DecoderLayer"></i><i class="c0" title="layer 9: DecoderLayer"></i><i class="c0" title="layer 10: DecoderLayer"></i><i class="c0" title="layer 11: DecoderLayer"></i><i class="c0" title="layer 12: DecoderLayer"></i><i class="c0" title="layer 13: DecoderLayer"></i><i class="c0" title="layer 14: DecoderLayer"></i><i class="c0" title="layer 15: DecoderLayer"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: DecoderLayer"></i><i class="c0" title="layer 1: DecoderLayer"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>DecoderLayer</td><td>16</td><td>2</td><td>✓ kept</td></tr></table></div><h2>Where the parameters are</h2><p class="note">Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.</p><div class="row"><span class="lab">Original</span><div class="bar"><i class="c0" style="flex:0.970221" title="model.layers: 6.7B (97.0%)"></i><i class="c1" style="flex:0.014889" title="model.embed_tokens: 103.0M (1.5%)"></i><i class="c2" style="flex:0.014889" title="lm_head: 103.0M (1.5%)"></i><i class="c3" style="flex:0.000000" title="model.norm: 2.0K (0.0%)"></i></div></div><div class="row"><span class="lab">Fixture</span><div class="bar"><i class="c0" style="flex:0.034546" title="model.layers: 230.4K (3.5%)"></i><i class="c1" style="flex:0.482722" title="model.embed_tokens: 3.2M (48.3%)"></i><i class="c2" style="flex:0.482722" title="lm_head: 3.2M (48.3%)"></i><i class="c3" style="flex:0.000010" title="model.norm: 64 (0.0%)"></i></div></div><table class="legend"><tr><th>Component</th><th>Original</th><th>Share</th><th>Fixture</th><th>Share</th></tr><tr><td><i class="c0"></i>model.layers</td><td>6.7B</td><td>97.0%</td><td>230.4K</td><td>3.5%</td></tr><tr><td><i class="c1"></i>model.embed_tokens</td><td>103.0M</td><td>1.5%</td><td>3.2M</td><td>48.3%</td></tr><tr><td><i class="c2"></i>lm_head</td><td>103.0M</td><td>1.5%</td><td>3.2M</td><td>48.3%</td></tr><tr><td><i class="c3"></i>model.norm</td><td>2.0K</td><td>0.0%</td><td>64</td><td>0.0%</td></tr></table><h2>Preserved properties</h2><table><tr><th>Config</th><th>Property</th><th>Original</th><th>Fixture</th><th></th></tr><tr class=""><td>model</td><td>attention style</td><td>MHA (16 heads, 16 KV)</td><td>MHA (4 heads, 4 KV)</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>MoE routing</td><td>top-8 of 64 experts</td><td>top-2 of 4 experts</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>RoPE type</td><td>default</td><td>default</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>model_type</td><td>olmoe</td><td>olmoe</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>vocab_size</td><td>50304</td><td>50304</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>tie_word_embeddings</td><td>False</td><td>False</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>hidden_act</td><td>silu</td><td>silu</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>max_position_embeddings</td><td>4096</td><td>4096</td><td>✓ kept</td></tr></table><h2>Module classes</h2><p class="note">Instances of each class in the built model (prefix Olmoe omitted).</p><table><tr><th>Class</th><th>Original</th><th>Fixture</th></tr><tr class=""><td>Linear</td><td>65</td><td>9</td></tr><tr class=""><td>RMSNorm</td><td>65</td><td>9</td></tr><tr class=""><td>Attention</td><td>16</td><td>2</td></tr><tr class=""><td>DecoderLayer</td><td>16</td><td>2</td></tr><tr class=""><td>Experts</td><td>16</td><td>2</td></tr><tr class=""><td>SparseMoeBlock</td><td>16</td><td>2</td></tr><tr class=""><td>TopKRouter</td><td>16</td><td>2</td></tr><tr class=""><td>SiLUActivation</td><td>16</td><td>2</td></tr><tr class=""><td>Embedding</td><td>1</td><td>1</td></tr><tr class=""><td>ModuleList</td><td>1</td><td>1</td></tr><tr class=""><td>ForCausalLM</td><td>1</td><td>1</td></tr><tr class=""><td>Model</td><td>1</td><td>1</td></tr><tr class=""><td>RotaryEmbedding</td><td>1</td><td>1</td></tr></table><h2>Config changes</h2><p class="note">7 fields changed, 26 unchanged.</p><table><tr><th>Field</th><th>Original</th><th>Fixture</th></tr><tr><td>hidden_size</td><td>2048</td><td>64</td></tr><tr><td>intermediate_size</td><td>1024</td><td>128</td></tr><tr><td>num_attention_heads</td><td>16</td><td>4</td></tr><tr><td>num_experts</td><td>64</td><td>4</td></tr><tr><td>num_experts_per_tok</td><td>8</td><td>2</td></tr><tr><td>num_hidden_layers</td><td>16</td><td>2</td></tr><tr><td>num_key_value_heads</td><td>16</td><td>4</td></tr></table>
+</main></body></html>

shrink/Olmo-3-7B-Instruct-mic.html

@@ -0,0 +1,34 @@
+<!doctype html>
+<html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
+<title>Shrunk architecture</title>
+<style>
+:root{--bg:#fcfcfb;--fg:#0b0b0b;--muted:#52514e;--line:#e4e3df;--ok:#1d7a43;--warn:#b35c00;--s0:#2a78d6;--s1:#eb6834;--s2:#1baf7a;--s3:#eda100;--s4:#e87ba4;--s5:#008300;--s6:#6250d6;--s7:#e34948;--s8:#8a8984;}
+@media (prefers-color-scheme:dark){:root:not([data-theme="light"]){--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}}
+:root[data-theme="dark"]{--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}
+body{margin:0;background:var(--bg);color:var(--fg);font:15px/1.5 system-ui,sans-serif}
+main{max-width:1040px;margin:0 auto;padding:24px 16px 64px}
+h1{font-size:24px;margin:0 0 4px}h2{font-size:18px;margin:36px 0 8px}h3{font-size:14px;margin:0 0 8px;font-family:ui-monospace,monospace}
+.meta,.note{color:var(--muted);margin:0 0 12px;overflow-wrap:anywhere}
+.stats{display:grid;grid-template-columns:repeat(auto-fit,minmax(150px,1fr));gap:12px;margin:20px 0}
+.stat{border:1px solid var(--line);border-radius:8px;padding:12px;display:flex;flex-direction:column}
+.stat .v{font-size:26px;font-weight:600}.stat .k{color:var(--muted);font-size:13px}
+.findings{border-radius:8px;padding:4px 16px;border:1px solid var(--line)}
+.findings.ok{border-left:4px solid var(--ok)}.findings.warn{border-left:4px solid var(--warn)}
+.findings h2{margin:12px 0 4px;font-size:16px}
+.stack{border:1px solid var(--line);border-radius:8px;padding:12px;margin:12px 0}
+.row{display:flex;align-items:center;gap:12px;margin:6px 0}
+.lab{flex:0 0 110px;color:var(--muted);font-size:13px}
+.cells{display:flex;flex-wrap:wrap;gap:2px}
+.cells i{width:12px;height:12px;border-radius:2px;display:block}
+.bar{flex:1;display:flex;gap:2px;height:16px}.bar i{display:block;border-radius:2px;min-width:2px}
+table{border-collapse:collapse;width:100%;font-size:13px;margin:8px 0;display:block;overflow-x:auto}
+th,td{text-align:left;padding:4px 10px 4px 0;border-bottom:1px solid var(--line);vertical-align:top}
+th{color:var(--muted);font-weight:500}
+td:first-child{font-family:ui-monospace,monospace;overflow-wrap:anywhere}
+td i{display:inline-block;width:10px;height:10px;border-radius:2px;margin-right:6px}
+tr.bad td{color:var(--warn)}
+.c0{background:var(--s0)}.c1{background:var(--s1)}.c2{background:var(--s2)}.c3{background:var(--s3)}.c4{background:var(--s4)}.c5{background:var(--s5)}.c6{background:var(--s6)}.c7{background:var(--s7)}.c8{background:var(--s8)}
+</style></head><body><main>
+<h1>Shrunk architecture</h1><p class="meta">huggingface-mic/Olmo-3-7B-Instruct-mic@3d2269c8c387b9dec2f5a987d2739464ad990224 ← allenai/Olmo-3-7B-Instruct@6e5971d9eba42665f5bd5a0fcf047f299ce1dccc · Olmo3ForCausalLM · random weights</p>
+<div class="stats"><div class="stat"><span class="k">Original</span><span class="v">7.3B</span><span class="k">parameters</span></div><div class="stat"><span class="k">Fixture</span><span class="v">12.9M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Layer stacks</span><span class="v">1</span><span class="k">compared</span></div><div class="stat"><span class="k">Findings</span><span class="v">0</span><span class="k">to review</span></div></div><section class="findings ok"><h2>✓ Same architecture, smaller</h2><p>Every module class, layer kind and checked property of the original is in the fixture.</p></section><h2>Layer stacks</h2><p class="note">One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.</p><div class="stack"><h3>model.layers</h3><div class="row"><span class="lab">Original · 32</span><div class="cells"><i class="c0" title="layer 0: sliding_attention"></i><i class="c0" title="layer 1: sliding_attention"></i><i class="c0" title="layer 2: sliding_attention"></i><i class="c1" title="layer 3: full_attention"></i><i class="c0" title="layer 4: sliding_attention"></i><i class="c0" title="layer 5: sliding_attention"></i><i class="c0" title="layer 6: sliding_attention"></i><i class="c1" title="layer 7: full_attention"></i><i class="c0" title="layer 8: sliding_attention"></i><i class="c0" title="layer 9: sliding_attention"></i><i class="c0" title="layer 10: sliding_attention"></i><i class="c1" title="layer 11: full_attention"></i><i class="c0" title="layer 12: sliding_attention"></i><i class="c0" title="layer 13: sliding_attention"></i><i class="c0" title="layer 14: sliding_attention"></i><i class="c1" title="layer 15: full_attention"></i><i class="c0" title="layer 16: sliding_attention"></i><i class="c0" title="layer 17: sliding_attention"></i><i class="c0" title="layer 18: sliding_attention"></i><i class="c1" title="layer 19: full_attention"></i><i class="c0" title="layer 20: sliding_attention"></i><i class="c0" title="layer 21: sliding_attention"></i><i class="c0" title="layer 22: sliding_attention"></i><i class="c1" title="layer 23: full_attention"></i><i class="c0" title="layer 24: sliding_attention"></i><i class="c0" title="layer 25: sliding_attention"></i><i class="c0" title="layer 26: sliding_attention"></i><i class="c1" title="layer 27: full_attention"></i><i class="c0" title="layer 28: sliding_attention"></i><i class="c0" title="layer 29: sliding_attention"></i><i class="c0" title="layer 30: sliding_attention"></i><i class="c1" title="layer 31: full_attention"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: sliding_attention"></i><i class="c1" title="layer 1: full_attention"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>sliding_attention</td><td>24</td><td>1</td><td>✓ kept</td></tr><tr><td><i class="c1"></i>full_attention</td><td>8</td><td>1</td><td>✓ kept</td></tr></table></div><h2>Where the parameters are</h2><p class="note">Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.</p><div class="row"><span class="lab">Original</span><div class="bar"><i class="c0" style="flex:0.887438" title="model.layers: 6.5B (88.7%)"></i><i class="c1" style="flex:0.056281" title="model.embed_tokens: 410.7M (5.6%)"></i><i class="c2" style="flex:0.056281" title="lm_head: 410.7M (5.6%)"></i><i class="c3" style="flex:0.000001" title="model.norm: 4.1K (0.0%)"></i></div></div><div class="row"><span class="lab">Fixture</span><div class="bar"><i class="c0" style="flex:0.006381" title="model.layers: 82.4K (0.6%)"></i><i class="c1" style="flex:0.496807" title="model.embed_tokens: 6.4M (49.7%)"></i><i class="c2" style="flex:0.496807" title="lm_head: 6.4M (49.7%)"></i><i class="c3" style="flex:0.000005" title="model.norm: 64 (0.0%)"></i></div></div><table class="legend"><tr><th>Component</th><th>Original</th><th>Share</th><th>Fixture</th><th>Share</th></tr><tr><td><i class="c0"></i>model.layers</td><td>6.5B</td><td>88.7%</td><td>82.4K</td><td>0.6%</td></tr><tr><td><i class="c1"></i>model.embed_tokens</td><td>410.7M</td><td>5.6%</td><td>6.4M</td><td>49.7%</td></tr><tr><td><i class="c2"></i>lm_head</td><td>410.7M</td><td>5.6%</td><td>6.4M</td><td>49.7%</td></tr><tr><td><i class="c3"></i>model.norm</td><td>4.1K</td><td>0.0%</td><td>64</td><td>0.0%</td></tr></table><h2>Preserved properties</h2><table><tr><th>Config</th><th>Property</th><th>Original</th><th>Fixture</th><th></th></tr><tr class=""><td>model</td><td>attention style</td><td>MHA (32 heads, 32 KV)</td><td>MHA (4 heads, 4 KV)</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>model_type</td><td>olmo3</td><td>olmo3</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>vocab_size</td><td>100278</td><td>100278</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>tie_word_embeddings</td><td>False</td><td>False</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>hidden_act</td><td>silu</td><td>silu</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>max_position_embeddings</td><td>65536</td><td>65536</td><td>✓ kept</td></tr></table><h2>Module classes</h2><p class="note">Instances of each class in the built model (prefix Olmo3 omitted).</p><table><tr><th>Class</th><th>Original</th><th>Fixture</th></tr><tr class=""><td>Linear</td><td>225</td><td>15</td></tr><tr class=""><td>RMSNorm</td><td>129</td><td>9</td></tr><tr class=""><td>Attention</td><td>32</td><td>2</td></tr><tr class=""><td>DecoderLayer</td><td>32</td><td>2</td></tr><tr class=""><td>MLP</td><td>32</td><td>2</td></tr><tr class=""><td>SiLUActivation</td><td>32</td><td>2</td></tr><tr class=""><td>Embedding</td><td>1</td><td>1</td></tr><tr class=""><td>ModuleList</td><td>1</td><td>1</td></tr><tr class=""><td>ForCausalLM</td><td>1</td><td>1</td></tr><tr class=""><td>Model</td><td>1</td><td>1</td></tr><tr class=""><td>RotaryEmbedding</td><td>1</td><td>1</td></tr></table><h2>Config changes</h2><p class="note">6 fields changed, 23 unchanged.</p><table><tr><th>Field</th><th>Original</th><th>Fixture</th></tr><tr><td>hidden_size</td><td>4096</td><td>64</td></tr><tr><td>intermediate_size</td><td>11008</td><td>128</td></tr><tr><td>layer_types</td><td>32 entries: 24× sliding_attention, 8× full_attention</td><td>[&quot;sliding_attention&quot;, &quot;full_attention&quot;]</td></tr><tr><td>num_attention_heads</td><td>32</td><td>4</td></tr><tr><td>num_hidden_layers</td><td>32</td><td>2</td></tr><tr><td>num_key_value_heads</td><td>32</td><td>4</td></tr></table>
+</main></body></html>

shrink/Phi-3.5-MoE-0.8B-A0.2B-mic.html

@@ -0,0 +1,34 @@
+<!doctype html>
+<html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
+<title>Shrunk architecture</title>
+<style>
+:root{--bg:#fcfcfb;--fg:#0b0b0b;--muted:#52514e;--line:#e4e3df;--ok:#1d7a43;--warn:#b35c00;--s0:#2a78d6;--s1:#eb6834;--s2:#1baf7a;--s3:#eda100;--s4:#e87ba4;--s5:#008300;--s6:#6250d6;--s7:#e34948;--s8:#8a8984;}
+@media (prefers-color-scheme:dark){:root:not([data-theme="light"]){--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}}
+:root[data-theme="dark"]{--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}
+body{margin:0;background:var(--bg);color:var(--fg);font:15px/1.5 system-ui,sans-serif}
+main{max-width:1040px;margin:0 auto;padding:24px 16px 64px}
+h1{font-size:24px;margin:0 0 4px}h2{font-size:18px;margin:36px 0 8px}h3{font-size:14px;margin:0 0 8px;font-family:ui-monospace,monospace}
+.meta,.note{color:var(--muted);margin:0 0 12px;overflow-wrap:anywhere}
+.stats{display:grid;grid-template-columns:repeat(auto-fit,minmax(150px,1fr));gap:12px;margin:20px 0}
+.stat{border:1px solid var(--line);border-radius:8px;padding:12px;display:flex;flex-direction:column}
+.stat .v{font-size:26px;font-weight:600}.stat .k{color:var(--muted);font-size:13px}
+.findings{border-radius:8px;padding:4px 16px;border:1px solid var(--line)}
+.findings.ok{border-left:4px solid var(--ok)}.findings.warn{border-left:4px solid var(--warn)}
+.findings h2{margin:12px 0 4px;font-size:16px}
+.stack{border:1px solid var(--line);border-radius:8px;padding:12px;margin:12px 0}
+.row{display:flex;align-items:center;gap:12px;margin:6px 0}
+.lab{flex:0 0 110px;color:var(--muted);font-size:13px}
+.cells{display:flex;flex-wrap:wrap;gap:2px}
+.cells i{width:12px;height:12px;border-radius:2px;display:block}
+.bar{flex:1;display:flex;gap:2px;height:16px}.bar i{display:block;border-radius:2px;min-width:2px}
+table{border-collapse:collapse;width:100%;font-size:13px;margin:8px 0;display:block;overflow-x:auto}
+th,td{text-align:left;padding:4px 10px 4px 0;border-bottom:1px solid var(--line);vertical-align:top}
+th{color:var(--muted);font-weight:500}
+td:first-child{font-family:ui-monospace,monospace;overflow-wrap:anywhere}
+td i{display:inline-block;width:10px;height:10px;border-radius:2px;margin-right:6px}
+tr.bad td{color:var(--warn)}
+.c0{background:var(--s0)}.c1{background:var(--s1)}.c2{background:var(--s2)}.c3{background:var(--s3)}.c4{background:var(--s4)}.c5{background:var(--s5)}.c6{background:var(--s6)}.c7{background:var(--s7)}.c8{background:var(--s8)}
+</style></head><body><main>
+<h1>Shrunk architecture</h1><p class="meta">huggingface-mic/Phi-3.5-MoE-0.8B-A0.2B-mic@38df14197e6a9ed50ccfe370498b9764d8dcadda ← inference-optimization/Phi-3.5-MoE-0.8B-A0.2B@f6258a2026894c90bffd3c0cfb11c42490935bb8 · PhimoeForCausalLM · random weights</p>
+<div class="stats"><div class="stat"><span class="k">Original</span><span class="v">802.6M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Fixture</span><span class="v">4.4M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Layer stacks</span><span class="v">1</span><span class="k">compared</span></div><div class="stat"><span class="k">Findings</span><span class="v">0</span><span class="k">to review</span></div></div><section class="findings ok"><h2>✓ Same architecture, smaller</h2><p>Every module class, layer kind and checked property of the original is in the fixture.</p></section><h2>Layer stacks</h2><p class="note">One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.</p><div class="stack"><h3>model.layers</h3><div class="row"><span class="lab">Original · 4</span><div class="cells"><i class="c0" title="layer 0: DecoderLayer"></i><i class="c0" title="layer 1: DecoderLayer"></i><i class="c0" title="layer 2: DecoderLayer"></i><i class="c0" title="layer 3: DecoderLayer"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: DecoderLayer"></i><i class="c0" title="layer 1: DecoderLayer"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>DecoderLayer</td><td>4</td><td>2</td><td>✓ kept</td></tr></table></div><h2>Where the parameters are</h2><p class="note">Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.</p><div class="row"><span class="lab">Original</span><div class="bar"><i class="c0" style="flex:0.836315" title="model.layers: 671.2M (83.6%)"></i><i class="c1" style="flex:0.081860" title="lm_head: 65.7M (8.2%)"></i><i class="c2" style="flex:0.081820" title="model.embed_tokens: 65.7M (8.2%)"></i><i class="c3" style="flex:0.000005" title="model.norm: 4.1K (0.0%)"></i></div></div><div class="row"><span class="lab">Fixture</span><div class="bar"><i class="c0" style="flex:0.051065" title="model.layers: 222.6K (5.1%)"></i><i class="c1" style="flex:0.478131" title="lm_head: 2.1M (47.8%)"></i><i class="c2" style="flex:0.470775" title="model.embed_tokens: 2.1M (47.1%)"></i><i class="c3" style="flex:0.000029" title="model.norm: 128 (0.0%)"></i></div></div><table class="legend"><tr><th>Component</th><th>Original</th><th>Share</th><th>Fixture</th><th>Share</th></tr><tr><td><i class="c0"></i>model.layers</td><td>671.2M</td><td>83.6%</td><td>222.6K</td><td>5.1%</td></tr><tr><td><i class="c1"></i>lm_head</td><td>65.7M</td><td>8.2%</td><td>2.1M</td><td>47.8%</td></tr><tr><td><i class="c2"></i>model.embed_tokens</td><td>65.7M</td><td>8.2%</td><td>2.1M</td><td>47.1%</td></tr><tr><td><i class="c3"></i>model.norm</td><td>4.1K</td><td>0.0%</td><td>128</td><td>0.0%</td></tr></table><h2>Preserved properties</h2><table><tr><th>Config</th><th>Property</th><th>Original</th><th>Fixture</th><th></th></tr><tr class=""><td>model</td><td>attention style</td><td>GQA (16 heads, 4 KV)</td><td>GQA (4 heads, 2 KV)</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>MoE routing</td><td>top-2 of 8 experts</td><td>top-2 of 4 experts</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>RoPE type</td><td>longrope</td><td>longrope</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>model_type</td><td>phimoe</td><td>phimoe</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>vocab_size</td><td>32064</td><td>32064</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>tie_word_embeddings</td><td>False</td><td>False</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>hidden_act</td><td>silu</td><td>silu</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>max_position_embeddings</td><td>131072</td><td>131072</td><td>✓ kept</td></tr></table><h2>Module classes</h2><p class="note">Instances of each class in the built model (prefix Phimoe omitted).</p><table><tr><th>Class</th><th>Original</th><th>Fixture</th></tr><tr class=""><td>Linear</td><td>17</td><td>9</td></tr><tr class=""><td>LayerNorm</td><td>9</td><td>5</td></tr><tr class=""><td>Attention</td><td>4</td><td>2</td></tr><tr class=""><td>DecoderLayer</td><td>4</td><td>2</td></tr><tr class=""><td>Experts</td><td>4</td><td>2</td></tr><tr class=""><td>SparseMoeBlock</td><td>4</td><td>2</td></tr><tr class=""><td>TopKRouter</td><td>4</td><td>2</td></tr><tr class=""><td>SiLUActivation</td><td>4</td><td>2</td></tr><tr class=""><td>Embedding</td><td>1</td><td>1</td></tr><tr class=""><td>ModuleList</td><td>1</td><td>1</td></tr><tr class=""><td>ForCausalLM</td><td>1</td><td>1</td></tr><tr class=""><td>Model</td><td>1</td><td>1</td></tr><tr class=""><td>RotaryEmbedding</td><td>1</td><td>1</td></tr></table><h2>Config changes</h2><p class="note">7 fields changed, 31 unchanged.</p><table><tr><th>Field</th><th>Original</th><th>Fixture</th></tr><tr><td>hidden_size</td><td>2048</td><td>64</td></tr><tr><td>intermediate_size</td><td>3200</td><td>128</td></tr><tr><td>num_attention_heads</td><td>16</td><td>4</td></tr><tr><td>num_hidden_layers</td><td>4</td><td>2</td></tr><tr><td>num_key_value_heads</td><td>4</td><td>2</td></tr><tr><td>num_local_experts</td><td>8</td><td>4</td></tr><tr><td>rope_parameters</td><td>{&quot;long_factor&quot;: [1.0199999809265137, 1.0299999713897705, 1.0399999618530273, …</td><td>{&quot;long_factor&quot;: [1.0199999809265137, 1.059999942779541, 1.4899998903274536, 1…</td></tr></table>
+</main></body></html>

shrink/Phi-3.5-mini-instruct-mic.html

@@ -0,0 +1,34 @@
+<!doctype html>
+<html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
+<title>Shrunk architecture</title>
+<style>
+:root{--bg:#fcfcfb;--fg:#0b0b0b;--muted:#52514e;--line:#e4e3df;--ok:#1d7a43;--warn:#b35c00;--s0:#2a78d6;--s1:#eb6834;--s2:#1baf7a;--s3:#eda100;--s4:#e87ba4;--s5:#008300;--s6:#6250d6;--s7:#e34948;--s8:#8a8984;}
+@media (prefers-color-scheme:dark){:root:not([data-theme="light"]){--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}}
+:root[data-theme="dark"]{--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}
+body{margin:0;background:var(--bg);color:var(--fg);font:15px/1.5 system-ui,sans-serif}
+main{max-width:1040px;margin:0 auto;padding:24px 16px 64px}
+h1{font-size:24px;margin:0 0 4px}h2{font-size:18px;margin:36px 0 8px}h3{font-size:14px;margin:0 0 8px;font-family:ui-monospace,monospace}
+.meta,.note{color:var(--muted);margin:0 0 12px;overflow-wrap:anywhere}
+.stats{display:grid;grid-template-columns:repeat(auto-fit,minmax(150px,1fr));gap:12px;margin:20px 0}
+.stat{border:1px solid var(--line);border-radius:8px;padding:12px;display:flex;flex-direction:column}
+.stat .v{font-size:26px;font-weight:600}.stat .k{color:var(--muted);font-size:13px}
+.findings{border-radius:8px;padding:4px 16px;border:1px solid var(--line)}
+.findings.ok{border-left:4px solid var(--ok)}.findings.warn{border-left:4px solid var(--warn)}
+.findings h2{margin:12px 0 4px;font-size:16px}
+.stack{border:1px solid var(--line);border-radius:8px;padding:12px;margin:12px 0}
+.row{display:flex;align-items:center;gap:12px;margin:6px 0}
+.lab{flex:0 0 110px;color:var(--muted);font-size:13px}
+.cells{display:flex;flex-wrap:wrap;gap:2px}
+.cells i{width:12px;height:12px;border-radius:2px;display:block}
+.bar{flex:1;display:flex;gap:2px;height:16px}.bar i{display:block;border-radius:2px;min-width:2px}
+table{border-collapse:collapse;width:100%;font-size:13px;margin:8px 0;display:block;overflow-x:auto}
+th,td{text-align:left;padding:4px 10px 4px 0;border-bottom:1px solid var(--line);vertical-align:top}
+th{color:var(--muted);font-weight:500}
+td:first-child{font-family:ui-monospace,monospace;overflow-wrap:anywhere}
+td i{display:inline-block;width:10px;height:10px;border-radius:2px;margin-right:6px}
+tr.bad td{color:var(--warn)}
+.c0{background:var(--s0)}.c1{background:var(--s1)}.c2{background:var(--s2)}.c3{background:var(--s3)}.c4{background:var(--s4)}.c5{background:var(--s5)}.c6{background:var(--s6)}.c7{background:var(--s7)}.c8{background:var(--s8)}
+</style></head><body><main>
+<h1>Shrunk architecture</h1><p class="meta">huggingface-mic/Phi-3.5-mini-instruct-mic@0c7964e6c77b965e4cccc04ccb4f4d4a20a9ff81 ← microsoft/Phi-3.5-mini-instruct@2fe192450127e6a83f7441aef6e3ca586c338b77 · Phi3ForCausalLM · random weights · built before the shrink rule (no recorded recipe)</p>
+<div class="stats"><div class="stat"><span class="k">Original</span><span class="v">3.8B</span><span class="k">parameters</span></div><div class="stat"><span class="k">Fixture</span><span class="v">4.2M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Layer stacks</span><span class="v">1</span><span class="k">compared</span></div><div class="stat"><span class="k">Findings</span><span class="v">0</span><span class="k">to review</span></div></div><section class="findings ok"><h2>✓ Same architecture, smaller</h2><p>Every module class, layer kind and checked property of the original is in the fixture.</p></section><h2>Layer stacks</h2><p class="note">One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.</p><div class="stack"><h3>model.layers</h3><div class="row"><span class="lab">Original · 32</span><div class="cells"><i class="c0" title="layer 0: DecoderLayer"></i><i class="c0" title="layer 1: DecoderLayer"></i><i class="c0" title="layer 2: DecoderLayer"></i><i class="c0" title="layer 3: DecoderLayer"></i><i class="c0" title="layer 4: DecoderLayer"></i><i class="c0" title="layer 5: DecoderLayer"></i><i class="c0" title="layer 6: DecoderLayer"></i><i class="c0" title="layer 7: DecoderLayer"></i><i class="c0" title="layer 8: DecoderLayer"></i><i class="c0" title="layer 9: DecoderLayer"></i><i class="c0" title="layer 10: DecoderLayer"></i><i class="c0" title="layer 11: DecoderLayer"></i><i class="c0" title="layer 12: DecoderLayer"></i><i class="c0" title="layer 13: DecoderLayer"></i><i class="c0" title="layer 14: DecoderLayer"></i><i class="c0" title="layer 15: DecoderLayer"></i><i class="c0" title="layer 16: DecoderLayer"></i><i class="c0" title="layer 17: DecoderLayer"></i><i class="c0" title="layer 18: DecoderLayer"></i><i class="c0" title="layer 19: DecoderLayer"></i><i class="c0" title="layer 20: DecoderLayer"></i><i class="c0" title="layer 21: DecoderLayer"></i><i class="c0" title="layer 22: DecoderLayer"></i><i class="c0" title="layer 23: DecoderLayer"></i><i class="c0" title="layer 24: DecoderLayer"></i><i class="c0" title="layer 25: DecoderLayer"></i><i class="c0" title="layer 26: DecoderLayer"></i><i class="c0" title="layer 27: DecoderLayer"></i><i class="c0" title="layer 28: DecoderLayer"></i><i class="c0" title="layer 29: DecoderLayer"></i><i class="c0" title="layer 30: DecoderLayer"></i><i class="c0" title="layer 31: DecoderLayer"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: DecoderLayer"></i><i class="c0" title="layer 1: DecoderLayer"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>DecoderLayer</td><td>32</td><td>2</td><td>✓ kept</td></tr></table></div><h2>Where the parameters are</h2><p class="note">Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.</p><div class="row"><span class="lab">Original</span><div class="bar"><i class="c0" style="flex:0.948443" title="model.layers: 3.6B (94.8%)"></i><i class="c1" style="flex:0.025778" title="lm_head: 98.5M (2.6%)"></i><i class="c2" style="flex:0.025778" title="model.embed_tokens: 98.5M (2.6%)"></i><i class="c3" style="flex:0.000001" title="model.norm: 3.1K (0.0%)"></i></div></div><div class="row"><span class="lab">Fixture</span><div class="bar"><i class="c0" style="flex:0.019629" title="model.layers: 82.2K (2.0%)"></i><i class="c1" style="flex:0.490178" title="lm_head: 2.1M (49.0%)"></i><i class="c2" style="flex:0.490178" title="model.embed_tokens: 2.1M (49.0%)"></i><i class="c3" style="flex:0.000015" title="model.norm: 64 (0.0%)"></i></div></div><table class="legend"><tr><th>Component</th><th>Original</th><th>Share</th><th>Fixture</th><th>Share</th></tr><tr><td><i class="c0"></i>model.layers</td><td>3.6B</td><td>94.8%</td><td>82.2K</td><td>2.0%</td></tr><tr><td><i class="c1"></i>lm_head</td><td>98.5M</td><td>2.6%</td><td>2.1M</td><td>49.0%</td></tr><tr><td><i class="c2"></i>model.embed_tokens</td><td>98.5M</td><td>2.6%</td><td>2.1M</td><td>49.0%</td></tr><tr><td><i class="c3"></i>model.norm</td><td>3.1K</td><td>0.0%</td><td>64</td><td>0.0%</td></tr></table><h2>Preserved properties</h2><table><tr><th>Config</th><th>Property</th><th>Original</th><th>Fixture</th><th></th></tr><tr class=""><td>model</td><td>attention style</td><td>MHA (32 heads, 32 KV)</td><td>MHA (4 heads, 4 KV)</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>RoPE type</td><td>longrope</td><td>longrope</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>model_type</td><td>phi3</td><td>phi3</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>vocab_size</td><td>32064</td><td>32064</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>tie_word_embeddings</td><td>False</td><td>False</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>hidden_act</td><td>silu</td><td>silu</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>max_position_embeddings</td><td>131072</td><td>131072</td><td>✓ kept</td></tr></table><h2>Module classes</h2><p class="note">Instances of each class in the built model (prefix Phi3 omitted).</p><table><tr><th>Class</th><th>Original</th><th>Fixture</th></tr><tr class=""><td>Linear</td><td>129</td><td>9</td></tr><tr class=""><td>RMSNorm</td><td>65</td><td>5</td></tr><tr class=""><td>Dropout</td><td>64</td><td>4</td></tr><tr class=""><td>Attention</td><td>32</td><td>2</td></tr><tr class=""><td>DecoderLayer</td><td>32</td><td>2</td></tr><tr class=""><td>MLP</td><td>32</td><td>2</td></tr><tr class=""><td>SiLUActivation</td><td>32</td><td>2</td></tr><tr class=""><td>Embedding</td><td>1</td><td>1</td></tr><tr class=""><td>ModuleList</td><td>1</td><td>1</td></tr><tr class=""><td>ForCausalLM</td><td>1</td><td>1</td></tr><tr class=""><td>Model</td><td>1</td><td>1</td></tr><tr class=""><td>RotaryEmbedding</td><td>1</td><td>1</td></tr></table><h2>Config changes</h2><p class="note">6 fields changed, 26 unchanged.</p><table><tr><th>Field</th><th>Original</th><th>Fixture</th></tr><tr><td>hidden_size</td><td>3072</td><td>64</td></tr><tr><td>intermediate_size</td><td>8192</td><td>128</td></tr><tr><td>num_attention_heads</td><td>32</td><td>4</td></tr><tr><td>num_hidden_layers</td><td>32</td><td>2</td></tr><tr><td>num_key_value_heads</td><td>32</td><td>4</td></tr><tr><td>rope_parameters</td><td>{&quot;long_factor&quot;: [1.0800000429153442, 1.1100000143051147, 1.1399999856948853, …</td><td>{&quot;long_factor&quot;: [1.0800000429153442, 1.6200000047683716, 7.700000286102295, 3…</td></tr></table>
+</main></body></html>

shrink/Phi-4-multimodal-instruct-mic.html

@@ -0,0 +1,34 @@
+<!doctype html>
+<html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
+<title>Shrunk architecture</title>
+<style>
+:root{--bg:#fcfcfb;--fg:#0b0b0b;--muted:#52514e;--line:#e4e3df;--ok:#1d7a43;--warn:#b35c00;--s0:#2a78d6;--s1:#eb6834;--s2:#1baf7a;--s3:#eda100;--s4:#e87ba4;--s5:#008300;--s6:#6250d6;--s7:#e34948;--s8:#8a8984;}
+@media (prefers-color-scheme:dark){:root:not([data-theme="light"]){--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}}
+:root[data-theme="dark"]{--bg:#1a1a19;--fg:#f0efec;--muted:#c3c2b7;--line:#3a3936;--ok:#2f9e5a;--warn:#d9822b;--s0:#3987e5;--s1:#d95926;--s2:#199e70;--s3:#c98500;--s4:#d55181;--s5:#008300;--s6:#9085e9;--s7:#e66767;--s8:#6f6e69;}
+body{margin:0;background:var(--bg);color:var(--fg);font:15px/1.5 system-ui,sans-serif}
+main{max-width:1040px;margin:0 auto;padding:24px 16px 64px}
+h1{font-size:24px;margin:0 0 4px}h2{font-size:18px;margin:36px 0 8px}h3{font-size:14px;margin:0 0 8px;font-family:ui-monospace,monospace}
+.meta,.note{color:var(--muted);margin:0 0 12px;overflow-wrap:anywhere}
+.stats{display:grid;grid-template-columns:repeat(auto-fit,minmax(150px,1fr));gap:12px;margin:20px 0}
+.stat{border:1px solid var(--line);border-radius:8px;padding:12px;display:flex;flex-direction:column}
+.stat .v{font-size:26px;font-weight:600}.stat .k{color:var(--muted);font-size:13px}
+.findings{border-radius:8px;padding:4px 16px;border:1px solid var(--line)}
+.findings.ok{border-left:4px solid var(--ok)}.findings.warn{border-left:4px solid var(--warn)}
+.findings h2{margin:12px 0 4px;font-size:16px}
+.stack{border:1px solid var(--line);border-radius:8px;padding:12px;margin:12px 0}
+.row{display:flex;align-items:center;gap:12px;margin:6px 0}
+.lab{flex:0 0 110px;color:var(--muted);font-size:13px}
+.cells{display:flex;flex-wrap:wrap;gap:2px}
+.cells i{width:12px;height:12px;border-radius:2px;display:block}
+.bar{flex:1;display:flex;gap:2px;height:16px}.bar i{display:block;border-radius:2px;min-width:2px}
+table{border-collapse:collapse;width:100%;font-size:13px;margin:8px 0;display:block;overflow-x:auto}
+th,td{text-align:left;padding:4px 10px 4px 0;border-bottom:1px solid var(--line);vertical-align:top}
+th{color:var(--muted);font-weight:500}
+td:first-child{font-family:ui-monospace,monospace;overflow-wrap:anywhere}
+td i{display:inline-block;width:10px;height:10px;border-radius:2px;margin-right:6px}
+tr.bad td{color:var(--warn)}
+.c0{background:var(--s0)}.c1{background:var(--s1)}.c2{background:var(--s2)}.c3{background:var(--s3)}.c4{background:var(--s4)}.c5{background:var(--s5)}.c6{background:var(--s6)}.c7{background:var(--s7)}.c8{background:var(--s8)}
+</style></head><body><main>
+<h1>Shrunk architecture</h1><p class="meta">huggingface-mic/Phi-4-multimodal-instruct-mic@721c204bb9a6c7113710f11eb53f6798c3b87039 ← microsoft/Phi-4-multimodal-instruct@b5b5a5f60533515297d0e14ece86ed80d4b13b19 · Phi4MultimodalForCausalLM · random weights · built before the shrink rule (no recorded recipe)</p>
+<div class="stats"><div class="stat"><span class="k">Original</span><span class="v">4.7B</span><span class="k">parameters</span></div><div class="stat"><span class="k">Fixture</span><span class="v">16.6M</span><span class="k">parameters</span></div><div class="stat"><span class="k">Layer stacks</span><span class="v">3</span><span class="k">compared</span></div><div class="stat"><span class="k">Findings</span><span class="v">0</span><span class="k">to review</span></div></div><section class="findings ok"><h2>✓ Same architecture, smaller</h2><p>Every module class, layer kind and checked property of the original is in the fixture.</p></section><h2>Layer stacks</h2><p class="note">One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.</p><div class="stack"><h3>model.layers</h3><div class="row"><span class="lab">Original · 32</span><div class="cells"><i class="c0" title="layer 0: DecoderLayer"></i><i class="c0" title="layer 1: DecoderLayer"></i><i class="c0" title="layer 2: DecoderLayer"></i><i class="c0" title="layer 3: DecoderLayer"></i><i class="c0" title="layer 4: DecoderLayer"></i><i class="c0" title="layer 5: DecoderLayer"></i><i class="c0" title="layer 6: DecoderLayer"></i><i class="c0" title="layer 7: DecoderLayer"></i><i class="c0" title="layer 8: DecoderLayer"></i><i class="c0" title="layer 9: DecoderLayer"></i><i class="c0" title="layer 10: DecoderLayer"></i><i class="c0" title="layer 11: DecoderLayer"></i><i class="c0" title="layer 12: DecoderLayer"></i><i class="c0" title="layer 13: DecoderLayer"></i><i class="c0" title="layer 14: DecoderLayer"></i><i class="c0" title="layer 15: DecoderLayer"></i><i class="c0" title="layer 16: DecoderLayer"></i><i class="c0" title="layer 17: DecoderLayer"></i><i class="c0" title="layer 18: DecoderLayer"></i><i class="c0" title="layer 19: DecoderLayer"></i><i class="c0" title="layer 20: DecoderLayer"></i><i class="c0" title="layer 21: DecoderLayer"></i><i class="c0" title="layer 22: DecoderLayer"></i><i class="c0" title="layer 23: DecoderLayer"></i><i class="c0" title="layer 24: DecoderLayer"></i><i class="c0" title="layer 25: DecoderLayer"></i><i class="c0" title="layer 26: DecoderLayer"></i><i class="c0" title="layer 27: DecoderLayer"></i><i class="c0" title="layer 28: DecoderLayer"></i><i class="c0" title="layer 29: DecoderLayer"></i><i class="c0" title="layer 30: DecoderLayer"></i><i class="c0" title="layer 31: DecoderLayer"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: DecoderLayer"></i><i class="c0" title="layer 1: DecoderLayer"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>DecoderLayer</td><td>32</td><td>2</td><td>✓ kept</td></tr></table></div><div class="stack"><h3>model.embed_tokens_extend.image_embed.img_processor.encoder.layers</h3><div class="row"><span class="lab">Original · 27</span><div class="cells"><i class="c0" title="layer 0: VisionEncoderLayer"></i><i class="c0" title="layer 1: VisionEncoderLayer"></i><i class="c0" title="layer 2: VisionEncoderLayer"></i><i class="c0" title="layer 3: VisionEncoderLayer"></i><i class="c0" title="layer 4: VisionEncoderLayer"></i><i class="c0" title="layer 5: VisionEncoderLayer"></i><i class="c0" title="layer 6: VisionEncoderLayer"></i><i class="c0" title="layer 7: VisionEncoderLayer"></i><i class="c0" title="layer 8: VisionEncoderLayer"></i><i class="c0" title="layer 9: VisionEncoderLayer"></i><i class="c0" title="layer 10: VisionEncoderLayer"></i><i class="c0" title="layer 11: VisionEncoderLayer"></i><i class="c0" title="layer 12: VisionEncoderLayer"></i><i class="c0" title="layer 13: VisionEncoderLayer"></i><i class="c0" title="layer 14: VisionEncoderLayer"></i><i class="c0" title="layer 15: VisionEncoderLayer"></i><i class="c0" title="layer 16: VisionEncoderLayer"></i><i class="c0" title="layer 17: VisionEncoderLayer"></i><i class="c0" title="layer 18: VisionEncoderLayer"></i><i class="c0" title="layer 19: VisionEncoderLayer"></i><i class="c0" title="layer 20: VisionEncoderLayer"></i><i class="c0" title="layer 21: VisionEncoderLayer"></i><i class="c0" title="layer 22: VisionEncoderLayer"></i><i class="c0" title="layer 23: VisionEncoderLayer"></i><i class="c0" title="layer 24: VisionEncoderLayer"></i><i class="c0" title="layer 25: VisionEncoderLayer"></i><i class="c0" title="layer 26: VisionEncoderLayer"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: VisionEncoderLayer"></i><i class="c0" title="layer 1: VisionEncoderLayer"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>VisionEncoderLayer</td><td>27</td><td>2</td><td>✓ kept</td></tr></table></div><div class="stack"><h3>model.embed_tokens_extend.audio_embed.encoder.encoders</h3><div class="row"><span class="lab">Original · 24</span><div class="cells"><i class="c0" title="layer 0: AudioConformerEncoderLayer"></i><i class="c0" title="layer 1: AudioConformerEncoderLayer"></i><i class="c0" title="layer 2: AudioConformerEncoderLayer"></i><i class="c0" title="layer 3: AudioConformerEncoderLayer"></i><i class="c0" title="layer 4: AudioConformerEncoderLayer"></i><i class="c0" title="layer 5: AudioConformerEncoderLayer"></i><i class="c0" title="layer 6: AudioConformerEncoderLayer"></i><i class="c0" title="layer 7: AudioConformerEncoderLayer"></i><i class="c0" title="layer 8: AudioConformerEncoderLayer"></i><i class="c0" title="layer 9: AudioConformerEncoderLayer"></i><i class="c0" title="layer 10: AudioConformerEncoderLayer"></i><i class="c0" title="layer 11: AudioConformerEncoderLayer"></i><i class="c0" title="layer 12: AudioConformerEncoderLayer"></i><i class="c0" title="layer 13: AudioConformerEncoderLayer"></i><i class="c0" title="layer 14: AudioConformerEncoderLayer"></i><i class="c0" title="layer 15: AudioConformerEncoderLayer"></i><i class="c0" title="layer 16: AudioConformerEncoderLayer"></i><i class="c0" title="layer 17: AudioConformerEncoderLayer"></i><i class="c0" title="layer 18: AudioConformerEncoderLayer"></i><i class="c0" title="layer 19: AudioConformerEncoderLayer"></i><i class="c0" title="layer 20: AudioConformerEncoderLayer"></i><i class="c0" title="layer 21: AudioConformerEncoderLayer"></i><i class="c0" title="layer 22: AudioConformerEncoderLayer"></i><i class="c0" title="layer 23: AudioConformerEncoderLayer"></i></div></div><div class="row"><span class="lab">Fixture · 2</span><div class="cells"><i class="c0" title="layer 0: AudioConformerEncoderLayer"></i><i class="c0" title="layer 1: AudioConformerEncoderLayer"></i></div></div><table class="legend"><tr><th>Kind</th><th>Original</th><th>Fixture</th><th></th></tr><tr><td><i class="c0"></i>AudioConformerEncoderLayer</td><td>24</td><td>2</td><td>✓ kept</td></tr></table></div><h2>Where the parameters are</h2><p class="note">Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.</p><div class="row"><span class="lab">Original</span><div class="bar"><i class="c0" style="flex:0.679054" title="model.layers: 3.2B (67.9%)"></i><i class="c1" style="flex:0.129553" title="model.embed_tokens: 614.6M (13.0%)"></i><i class="c2" style="flex:0.090347" title="model.embed_tokens_extend.audio_embed.encoder.encoders: 428.6M (9.0%)"></i><i class="c3" style="flex:0.086734" title="model.embed_tokens_extend.image_embed.img_processor.encoder.layers: 411.5M (8.7%)"></i><i class="c4" style="flex:0.014312" title="model.embed_tokens_extend: 67.9M (1.4%)"></i><i class="c5" style="flex:0.000001" title="model.norm: 3.1K (0.0%)"></i></div></div><div class="row"><span class="lab">Fixture</span><div class="bar"><i class="c0" style="flex:0.004466" title="model.layers: 74.0K (0.4%)"></i><i class="c1" style="flex:0.772843" title="model.embed_tokens: 12.8M (77.3%)"></i><i class="c2" style="flex:0.040515" title="model.embed_tokens_extend.audio_embed.encoder.encoders: 671.2K (4.1%)"></i><i class="c3" style="flex:0.004041" title="model.embed_tokens_extend.image_embed.img_processor.encoder.layers: 66.9K (0.4%)"></i><i class="c4" style="flex:0.178132" title="model.embed_tokens_extend: 3.0M (17.8%)"></i><i class="c5" style="flex:0.000004" title="model.norm: 64 (0.0%)"></i></div></div><table class="legend"><tr><th>Component</th><th>Original</th><th>Share</th><th>Fixture</th><th>Share</th></tr><tr><td><i class="c0"></i>model.layers</td><td>3.2B</td><td>67.9%</td><td>74.0K</td><td>0.4%</td></tr><tr><td><i class="c1"></i>model.embed_tokens</td><td>614.6M</td><td>13.0%</td><td>12.8M</td><td>77.3%</td></tr><tr><td><i class="c2"></i>model.embed_tokens_extend.audio_embed.encoder.encoders</td><td>428.6M</td><td>9.0%</td><td>671.2K</td><td>4.1%</td></tr><tr><td><i class="c3"></i>model.embed_tokens_extend.image_embed.img_processor.encoder.layers</td><td>411.5M</td><td>8.7%</td><td>66.9K</td><td>0.4%</td></tr><tr><td><i class="c4"></i>model.embed_tokens_extend</td><td>67.9M</td><td>1.4%</td><td>3.0M</td><td>17.8%</td></tr><tr><td><i class="c5"></i>model.norm</td><td>3.1K</td><td>0.0%</td><td>64</td><td>0.0%</td></tr></table><h2>Preserved properties</h2><table><tr><th>Config</th><th>Property</th><th>Original</th><th>Fixture</th><th></th></tr><tr class=""><td>model</td><td>attention style</td><td>GQA (24 heads, 8 KV)</td><td>GQA (4 heads, 2 KV)</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>RoPE type</td><td>longrope</td><td>longrope</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>model_type</td><td>phi4_multimodal</td><td>phi4_multimodal</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>vocab_size</td><td>200064</td><td>200064</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>tie_word_embeddings</td><td>True</td><td>True</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>hidden_act</td><td>silu</td><td>silu</td><td>✓ kept</td></tr><tr class=""><td>model</td><td>max_position_embeddings</td><td>131072</td><td>131072</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>model_type</td><td>phi4_multimodal_vision</td><td>phi4_multimodal_vision</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>hidden_act</td><td>gelu_pytorch_tanh</td><td>gelu_pytorch_tanh</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>image_size</td><td>448</td><td>448</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>patch_size</td><td>14</td><td>14</td><td>✓ kept</td></tr><tr class=""><td>vision_config</td><td>num_channels</td><td>3</td><td>3</td><td>✓ kept</td></tr><tr class=""><td>audio_config</td><td>model_type</td><td>phi4_multimodal_audio</td><td>phi4_multimodal_audio</td><td>✓ kept</td></tr></table><h2>Module classes</h2><p class="note">Instances of each class in the built model (prefix Phi4Multimodal omitted).</p><table><tr><th>Class</th><th>Original</th><th>Fixture</th></tr><tr class=""><td>Linear</td><td>492</td><td>46</td></tr><tr class=""><td>LayerNorm</td><td>176</td><td>16</td></tr><tr class=""><td>Dropout</td><td>139</td><td>13</td></tr><tr class=""><td>Conv1d</td><td>96</td><td>8</td></tr><tr class=""><td>SiLU</td><td>96</td><td>8</td></tr><tr class=""><td>RMSNorm</td><td>65</td><td>5</td></tr><tr class=""><td>AudioMLP</td><td>48</td><td>4</td></tr><tr class=""><td>Attention</td><td>32</td><td>2</td></tr><tr class=""><td>DecoderLayer</td><td>32</td><td>2</td></tr><tr class=""><td>MLP</td><td>32</td><td>2</td></tr><tr class=""><td>SiLUActivation</td><td>32</td><td>2</td></tr><tr class=""><td>GELUTanh</td><td>28</td><td>3</td></tr><tr class=""><td>VisionMLP</td><td>28</td><td>3</td></tr><tr class=""><td>VisionAttention</td><td>27</td><td>2</td></tr><tr class=""><td>VisionEncoderLayer</td><td>27</td><td>2</td></tr><tr class=""><td>AudioAttention</td><td>24</td><td>2</td></tr><tr class=""><td>AudioConformerEncoderLayer</td><td>24</td><td>2</td></tr><tr class=""><td>AudioConvModule</td><td>24</td><td>2</td></tr><tr class=""><td>AudioDepthWiseSeparableConv1d</td><td>24</td><td>2</td></tr><tr class=""><td>AudioGluPointWiseConv</td><td>24</td><td>2</td></tr><tr class=""><td>Conv2d</td><td>6</td><td>6</td></tr><tr class=""><td>Embedding</td><td>3</td><td>3</td></tr><tr class=""><td>ModuleList</td><td>3</td><td>3</td></tr><tr class=""><td>AvgPool2d</td><td>1</td><td>1</td></tr><tr class=""><td>MultiheadAttention</td><td>1</td><td>1</td></tr><tr class=""><td>NonDynamicallyQuantizableLinear</td><td>1</td><td>1</td></tr><tr class=""><td>AudioEmbedding</td><td>1</td><td>1</td></tr><tr class=""><td>AudioMeanVarianceNormLayer</td><td>1</td><td>1</td></tr><tr class=""><td>AudioModel</td><td>1</td><td>1</td></tr><tr class=""><td>AudioNemoConvSubsampling</td><td>1</td><td>1</td></tr><tr class=""><td>AudioRelativeAttentionBias</td><td>1</td><td>1</td></tr><tr class=""><td>FeatureEmbedding</td><td>1</td><td>1</td></tr><tr class=""><td>ForCausalLM</td><td>1</td><td>1</td></tr><tr class=""><td>ImageEmbedding</td><td>1</td><td>1</td></tr><tr class=""><td>Model</td><td>1</td><td>1</td></tr><tr class=""><td>RotaryEmbedding</td><td>1</td><td>1</td></tr><tr class=""><td>VisionEmbeddings</td><td>1</td><td>1</td></tr><tr class=""><td>VisionEncoder</td><td>1</td><td>1</td></tr><tr class=""><td>VisionModel</td><td>1</td><td>1</td></tr><tr class=""><td>VisionMultiheadAttentionPoolingHead</td><td>1</td><td>1</td></tr><tr class=""><td>ReLU</td><td>1</td><td>1</td></tr><tr class=""><td>Sequential</td><td>1</td><td>1</td></tr></table><h2>Config changes</h2><p class="note">15 fields changed, 78 unchanged.</p><table><tr><th>Field</th><th>Original</th><th>Fixture</th></tr><tr><td>audio_config.chunk_size</td><td>-1</td><td>8</td></tr><tr><td>audio_config.hidden_size</td><td>1024</td><td>64</td></tr><tr><td>audio_config.intermediate_size</td><td>1536</td><td>128</td></tr><tr><td>audio_config.num_attention_heads</td><td>16</td><td>4</td></tr><tr><td>audio_config.num_blocks</td><td>24</td><td>2</td></tr><tr><td>hidden_size</td><td>3072</td><td>64</td></tr><tr><td>intermediate_size</td><td>8192</td><td>128</td></tr><tr><td>num_attention_heads</td><td>24</td><td>4</td></tr><tr><td>num_hidden_layers</td><td>32</td><td>2</td></tr><tr><td>num_key_value_heads</td><td>8</td><td>2</td></tr><tr><td>rope_parameters</td><td>{&quot;long_factor&quot;: [1, 1.118320672, 1.250641126, 1.398617824, 1.564103225, 1.749…</td><td>{&quot;long_factor&quot;: [1, 2.446418898, 5.984965424, 14.64173252, 32.1, 32.71], &quot;ori…</td></tr><tr><td>vision_config.hidden_size</td><td>1152</td><td>64</td></tr><tr><td>vision_config.intermediate_size</td><td>4304</td><td>128</td></tr><tr><td>vision_config.num_attention_heads</td><td>16</td><td>4</td></tr><tr><td>vision_config.num_hidden_layers</td><td>27</td><td>2</td></tr></table>
+</main></body></html>