RTen ONNX model compatibility

hf-top-N.md

Hugging Face Top-200 ONNX Models — RTen Compatibility Sweep

Testing the 200 most-downloaded Hugging Face repos tagged onnx (by download count, snapshot 2026-07-25) against the RTen CLI. Repos 1-100 were covered by the first sweep; 101-200 by a follow-up sweep using the same method.

Method

• Models downloaded to hf_models/{repo} and run with target/release/rten <model.onnx> (random inputs).
• Dynamic dimensions set with --size, input value ranges with --range (e.g. token IDs must be in vocab range).
• Per repo: the main fp32 model (all parts, for encoder/decoder splits) plus one representative quantized variant. For repos 101-200 the ORT-preoptimized model_O4.onnx was also tested where present, since it is a structurally different graph.
• Models >2GB (including external .onnx_data) are skipped. The sweep initially used a 1GB limit; a second pass covered everything in the 1-2GB range.
• Success = model loads and runs end-to-end without error. Output values are not validated.
• Model files are deleted after testing.

Bugs fixed during the sweep

• use_cache_branch input shape. The CLI's random input generator produced a 0-dim scalar for this input regardless of the declared shape, so every Optimum-exported *_merged decoder failed to run with expected tensor with 1 dims but got 0. Now generates a tensor of the declared shape (rten-cli/src/input_generator.rs).

• Pointwise conv panicked on a channel mismatch. conv_impl validated the group count and input-vs-kernel channel counts after the 1x1 fast path had already returned, and conv_2d_pointwise reshapes assuming those invariants hold. Feeding SDXL's vae_decoder a latent with 3 channels instead of 4 aborted the process:

  thread 'rten-0' panicked at rten-tensor/src/tensor.rs:1866:17:
  element count mismatch reshaping [3, 64, 64] to [4, 4096]
     3: rten::ops::conv::conv_2d_pointwise::<f32, f32, f32>
     4: rten::ops::conv::conv_impl::<f32, f32, f32>
  

  The validation now runs before the fast path, so the same input returns Input channels (per group) does not match kernel input channels (src/ops/conv.rs).

• ConstantOfShape rejected most value dtypes. The ONNX deserializer parsed the value attribute with a bespoke helper that understood only float/int32/int64/bool, so f16 exports failed with error in attribute "value": unsupported data type for ConstantOfShape. It now deserializes the attribute through load_constant, the same path used for model weights, then narrows the result to a scalar. This adds f16, double, int8, uint8, external data and typed-field storage, and removes a duplicate deserializer that had drifted from the real one (src/op_registry/onnx_registry.rs, src/model/onnx_loader.rs). Both Jina fp16 models below now run.

• Empty f16 tensors failed to load. Reported by the sweep as f16 tensor data is not 2-byte aligned from convert_f16_constant (src/model/onnx_loader.rs), which reinterprets the protobuf bytes as &[f16] in place. The offending tensors turned out to be zero-element f16 initializers rather than misaligned ones: TensorProto::raw_data is decoded into a Vec<u8> and an empty Vec never allocates, so its pointer is the dangling sentinel 0x1, which fails the 2-byte alignment check. cast_slice and friends in rten-base/src/byte_cast.rs now short-circuit on an empty input and return a fresh empty slice of the destination type. Other dtypes were unaffected because they load via ArcSlice::from_bytes, which already worked around this. Fixes #131 briaai/RMBG-1.4, #156 Xenova/segformer-b0 and #185 inswapper_128.fp16, all of which now run.

• Graph optimization bound the wrong tensor when an operator input was an Identity alias of an initializer. Found in #107 (BERT), #148 (LLaMA), #163 (ViT) and #198 (Phi-3) — four different architectures, all from exports where many weights alias a single shared initializer via Identity (how PyTorch/Optimum emits tied or deduplicated tensors). The fused operator received an unrelated i32 tensor in place of its f32 weight or bias:

  operator "/model/layers.0/post_attention_layernorm/Mul_1" failed: conversion error for
  input 1: expected tensor with type tensor(f32) but has type tensor(i32).
  Inputs were (tensor(f32) [2, 32, 32], tensor(i32) [32])
  

  All four ran correctly with --no-optimize, placing the fault in the optimizer rather than the loader. Root cause: shape inference treats an all-integral f32 constant as a symbolic integer vector, and Identity propagates that known value to its output, so the optimizer materialized the inferred constant as i32 and substituted it for the f32 weight. Only untrained test models tripped it, since they need biases of exactly 0.0 or norm weights of exactly 1.0 to look integral. The replacement constant is now typed to match the value being replaced, and is skipped when the value is not representable in that type (src/optimize.rs). This was the only correctness bug in the sweep; everything under "Feature gaps" below is an unimplemented feature. All four repos now run with optimization enabled.

Feature gaps found (not fixed)

Gap	Affects	Detail
com.microsoft/Attention unsupported	#75, #130, #136, #147, #164, #172, #184, #187, #195	ORT-optimized graphs that fuse attention into the MS contrib Attention op fail to load.
MemcpyFromHost unsupported	#102, #109, #118, #125, #138, #145, #159, #169, #196	Optimum _O4 exports embed ORT's device-placement Memcpy nodes in the saved graph.
com.microsoft/FastGelu unsupported	#115, #116	Emitted by ORT's bias-Gelu fusion in _O4 exports.
com.microsoft/GatherBlockQuantized unsupported	#91, #101	Block-quantized embedding tables. Blocks every runnable variant of both repos.
com.microsoft/DynamicQuantizeLSTM unsupported	#68 onnx-community/Kokoro-82M	Only the quantized Kokoro export uses it; the fp32 model runs.
com.microsoft/MatMulBnb4 unsupported	#49 Snowflake/snowflake-arctic-embed-l-v2.0	bitsandbytes-4bit weights. Only the _bnb4 variants use it; int8/int4/fp16 exports of the same model all run.
Scan unsupported	#165 NX-AI/TiRex	The only standard ai.onnx operator gap in the sweep; everything else is a contrib op.
Native INT4 initializers unsupported	#173 onnx-community/embeddinggemma-300m	initializer has unsupported data type INT4 — the ONNX INT4 tensor type, distinct from the packed-uint8 MatMulNBits layout that other int4 exports use successfully.
MatMulNBits with K not a multiple of block_size	#63 llava-onevision (int4 vision encoder)	Weight is packed for a padded K (135 blocks x 32 = 4320) while the input has K=4304, and the scales-size check rejects it. The int8 export of the same model runs.

Note the shape of the _O4 failures: Optimum's _O4 variant is "all ORT graph optimizations + fp16", so those exports are saturated with com.microsoft fusion ops and ORT-internal Memcpy nodes. In every affected repo the plain fp32 and int8 variants run, so no repo is blocked by this alone.

No model in either sweep failed because of a graph-structure problem: the transformer, CLIP/SigLIP, ViT, VLM, Whisper, T5, SegFormer, DETR, GAN-upscaler and TTS architectures in this list all execute.

Summary

	Count
Repositories total	200
Tested	200
Remaining	0
— of tested: all models passed	161
— of tested: some/all models failed	26
— of tested: skipped (gated / all variants >2GB)	13

Split by sweep:

	1-100	101-200	Total
✅ all models passed	87	74	161
⚠️ some models failed	3	18	21
❌ all models failed	2	3	5
⏭️ skipped	8	5	13

Counts are as of the fixes listed above; when the sweep was first run the totals were 154 / 24 / 9 / 13.

The remaining partial failures in 101-200 are entirely the _O4 variants: repos 101-200 contain far more Optimum exports that ship an ORT-preoptimized model_O4.onnx (or model_optimized.onnx) alongside the plain fp32/int8 files, and testing it turns an otherwise-green repo amber. All 18 are of exactly that form (9 MemcpyFromHost, 7 com.microsoft/Attention, 2 FastGelu); excluding them, 92 of the 95 tested repos in 101-200 pass every variant.

Repos where no variant runs, all now blocked on an unimplemented operator rather than a bug: #75 and #130 (com.microsoft/Attention), #91 and #101 (GatherBlockQuantized), #165 (Scan).

Results

Status: ✅ pass · ❌ fail · ⚠️ partial · ⏭️ skipped · ⬜ not yet tested

#	Repository	Status	Models tested	Notes
1	sentence-transformers/all-MiniLM-L6-v2	✅	onnx/model.onnx, onnx/model_qint8_arm64.onnx	Both run. BERT-style inputs (input_ids, attention_mask, token_type_ids).
2	google-bert/bert-base-uncased	✅	model.onnx	Runs. 133M params, output logits [2,128,30522].
3	cross-encoder/ms-marco-MiniLM-L6-v2	✅	onnx/model.onnx, onnx/model_qint8_arm64.onnx	Both run.
4	BAAI/bge-small-en-v1.5	✅	onnx/model.onnx	Runs.
5	sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2	✅	onnx/model.onnx, onnx/model_qint8_arm64.onnx	Both run.
6	BAAI/bge-m3	⏭️	—	Skipped: only variant is onnx/model.onnx + 2.27GB external data (>2GB).
7	sentence-transformers/all-mpnet-base-v2	✅	onnx/model.onnx, onnx/model_qint8_arm64.onnx	Both run. No token_type_ids input.
8	google-t5/t5-small	✅	encoder_model, decoder_model, decoder_with_past_model, decoder_model_merged, decoder_model_merged_quantized	All 5 run. Merged decoders initially failed with input "use_cache_branch" is invalid: expected tensor with 1 dims but got 0; fixed in the CLI (see notes below).
9	FacebookAI/xlm-roberta-base	✅	model.onnx (1.88GB)	Runs. Output logits [2,32,250002].
10	intfloat/multilingual-e5-small	✅	onnx/model.onnx, onnx/model_qint8_avx512_vnni.onnx	Both run (the avx512-named int8 model runs fine on ARM).
11	nomic-ai/nomic-embed-text-v1.5	✅	onnx/model.onnx, onnx/model_quantized.onnx	Both run.
12	openai-community/gpt2	✅	onnx/decoder_model.onnx, onnx/decoder_model_merged.onnx, onnx/decoder_with_past_model.onnx	All 3 run (after the use_cache_branch CLI fix).
13	BAAI/bge-large-en-v1.5	✅	onnx/model.onnx (1.34GB)	Runs.
14	intfloat/multilingual-e5-large	✅	onnx/model_qint8_avx512_vnni.onnx	Runs. fp32 onnx/model.onnx skipped (2.24GB external data, >2GB).
15	FacebookAI/roberta-large	✅	model.onnx (1.63GB)	Runs.
16	sentence-transformers/paraphrase-multilingual-mpnet-base-v2	✅	onnx/model.onnx (1.11GB), onnx/model_qint8_arm64.onnx	Both run.
17	cross-encoder/ms-marco-MiniLM-L4-v2	✅	onnx/model.onnx, onnx/model_qint8_arm64.onnx	Both run.
18	BAAI/bge-base-en-v1.5	✅	onnx/model.onnx	Runs.
19	FacebookAI/xlm-roberta-large	⏭️	—	Skipped: only variant is onnx/model.onnx + 2.24GB external data (>2GB).
20	intfloat/multilingual-e5-base	✅	onnx/model.onnx (1.11GB), onnx/model_qint8_avx512_vnni.onnx	Both run.
21	answerdotai/ModernBERT-base	✅	onnx/model.onnx, onnx/model_quantized.onnx	Both run. ModernBERT (RoPE, no token_type_ids).
22	BAAI/bge-reranker-base	✅	onnx/model.onnx (1.11GB)	Runs.
23	nomic-ai/nomic-embed-text-v1	✅	onnx/model.onnx, onnx/model_quantized.onnx	Both run.
24	mixedbread-ai/mxbai-embed-large-v1	✅	onnx/model.onnx (1.34GB), onnx/model_quantized.onnx	Both run.
25	distilbert/distilbert-base-uncased-finetuned-sst-2-english	✅	onnx/model.onnx	Runs. Output logits [2,2].
26	jinaai/jina-embeddings-v3	✅	onnx/model_fp16.onnx (1.15GB)	Runs, after the ConstantOfShape fp16 fix. fp32 onnx/model.onnx is 2.29GB (>2GB).
27	sentence-transformers/all-MiniLM-L12-v2	✅	onnx/model.onnx, onnx/model_qint8_arm64.onnx	Both run.
28	Snowflake/snowflake-arctic-embed-xs	✅	onnx/model.onnx, onnx/model_quantized.onnx	Both run.
29	patrickjohncyh/fashion-clip	✅	onnx/model.onnx	Runs. CLIP dual-tower: text + image inputs, 4 outputs.
30	colbert-ir/colbertv2.0	✅	model.onnx	Runs.
31	intfloat/e5-large-v2	✅	onnx/model.onnx (1.34GB), onnx/model_qint8_avx512_vnni.onnx	Both run.
32	cross-encoder/mmarco-mMiniLMv2-L12-H384-v1	✅	onnx/model.onnx, onnx/model_qint8_arm64.onnx	Both run.
33	Xenova/all-MiniLM-L6-v2	✅	onnx/model.onnx, onnx/model_quantized.onnx	Both run.
34	Alibaba-NLP/gte-reranker-modernbert-base	✅	onnx/model.onnx, onnx/model_quantized.onnx	Both run.
35	sentence-transformers/paraphrase-MiniLM-L6-v2	✅	onnx/model.onnx, onnx/model_qint8_arm64.onnx	Both run.
36	cross-encoder/ms-marco-MiniLM-L12-v2	✅	onnx/model.onnx, onnx/model_qint8_arm64.onnx	Both run.
37	sentence-transformers/all-distilroberta-v1	✅	onnx/model.onnx, onnx/model_qint8_arm64.onnx	Both run.
38	HuggingFaceTB/SmolLM2-135M-Instruct	✅	onnx/model.onnx, onnx/model_q4.onnx	Both run. Decoder LM with 30 KV-cache layer pairs; int4 (MatMulNBits) variant also runs.
39	BAAI/bge-reranker-large	⏭️	—	Skipped: only variant is onnx/model.onnx + 2.24GB external data (>2GB).
40	Xenova/bge-base-en-v1.5	✅	onnx/model.onnx, onnx/model_q4.onnx	Both run.
41	intfloat/multilingual-e5-large-instruct	⏭️	—	Skipped: only variant is onnx/model.onnx + 2.24GB external data (>2GB).
42	sentence-transformers/paraphrase-MiniLM-L3-v2	✅	onnx/model.onnx, onnx/model_qint8_arm64.onnx	Both run.
43	sentence-transformers/multi-qa-mpnet-base-dot-v1	✅	onnx/model.onnx, onnx/model_qint8_arm64.onnx	Both run.
44	answerdotai/ModernBERT-large	✅	onnx/model.onnx (1.58GB), onnx/model_quantized.onnx, onnx/model_q4.onnx	All 3 run.
45	jhgan/ko-sroberta-multitask	✅	onnx/model.onnx, onnx/model_qint8_avx512_vnni.onnx	Both run.
46	distil-whisper/distil-large-v3	✅	decoder_model, decoder_model_merged, decoder_with_past_model, encoder_model_quantized	All 4 run. Encoder needs realistic mel dims (-s feature_size=128 -s encoder_sequence_length=3000); the harness default of 1 mel bin produced a Conv channel-mismatch error (correctly reported, not a crash). fp32 encoder skipped (2.55GB).
47	sentence-transformers/distiluse-base-multilingual-cased-v2	✅	onnx/model.onnx, onnx/model_qint8_arm64.onnx	Both run.
48	Alibaba-NLP/gte-large-en-v1.5	✅	onnx/model.onnx (1.75GB), onnx/model_quantized.onnx, onnx/model_q4.onnx	All 3 run.
49	Snowflake/snowflake-arctic-embed-l-v2.0	⚠️	onnx/model_int8.onnx, onnx/model_q4.onnx (1.25GB), onnx/model_fp16.onnx (1.13GB) ✅ / onnx/model_bnb4.onnx ❌	int8, int4 and fp16 all run. bnb4 fails to load: in node ".../query/MatMul_Bnb4": operator error: com.microsoft/MatMulBnb4 operator not supported or not enabled. fp32 onnx/model.onnx is 2.27GB (>2GB).
50	dslim/bert-base-NER	✅	onnx/model.onnx	Runs. Token classification, output logits [2,32,9].
51	stabilityai/stable-diffusion-xl-base-1.0	✅	text_encoder/model.onnx, vae_encoder/model.onnx, vae_decoder/model.onnx	All 3 run (vae_decoder needs -s num_channels_latent=4). unet (10.3GB) and text_encoder_2 (2.8GB) skipped for size. Note: passing the wrong latent channel count panics instead of erroring — see robustness issues below.
52	intfloat/e5-base-v2	✅	onnx/model.onnx, onnx/model_qint8_avx512_vnni.onnx	Both run.
53	sentence-transformers/paraphrase-mpnet-base-v2	✅	onnx/model.onnx, onnx/model_qint8_arm64.onnx	Both run.
54	openai-community/gpt2-large	⏭️	—	Skipped: every variant has 3.35GB external data (>2GB).
55	Qdrant/all-MiniLM-L6-v2-onnx	✅	model.onnx	Runs.
56	microsoft/layoutlmv3-base	✅	model.onnx	Runs. LayoutLMv3 (text + bbox + pixel_values).
57	cross-encoder/ms-marco-MiniLM-L2-v2	✅	onnx/model.onnx, onnx/model_qint8_arm64.onnx	Both run.
58	TaylorAI/bge-micro-v2	✅	onnx/model.onnx, onnx/model_quantized.onnx	Both run.
59	sentence-transformers/distiluse-base-multilingual-cased-v1	✅	onnx/model.onnx, onnx/model_qint8_arm64.onnx	Both run.
60	jinaai/jina-embeddings-v2-small-en	✅	model.onnx, model-w-mean-pooling.onnx	Both run (ALiBi-based Jina BERT).
61	jinaai/jina-reranker-v2-base-multilingual	✅	onnx/model.onnx (1.11GB), onnx/model_int8.onnx, onnx/model_fp16.onnx	All 3 run. fp16 previously failed to load with the ConstantOfShape fp16 attribute error, fixed during the sweep.
62	HuggingFaceTB/SmolVLM-256M-Instruct	✅	decoder_model_merged, decoder_model_merged_q4, embed_tokens, vision_encoder	All 4 run. VLM split into vision encoder + token embedder + merged decoder.
63	llava-hf/llava-onevision-qwen2-0.5b-ov-hf	⚠️	decoder_model_merged (1.99GB), decoder_model_merged_q4, embed_tokens_int8, vision_encoder (1.6GB), vision_encoder_int8 ✅ / vision_encoder_q4 ❌	Everything runs except the int4 vision encoder: MatMul_Q4 ... Expected 1D \scales` size to match columns * block_size. Inputs were (tensor(f32) [2, 729, 4304], tensor(u8) [1152, 135, 16], tensor(f32) [155520])` — K=4304 is not a multiple of block_size 32, so the packed weight has 135 blocks covering a padded K of 4320.
64	stabilityai/sdxl-turbo	✅	text_encoder/model.onnx, vae_encoder/model.onnx, vae_decoder/model.onnx	Runs (same layout as SDXL base). unet (10.3GB) and text_encoder_2 (2.8GB) still >2GB and skipped.
65	sentence-transformers/all-roberta-large-v1	✅	onnx/model.onnx (1.42GB), onnx/model_qint8_arm64.onnx	Both run.
66	HuggingFaceTB/SmolVLM2-500M-Video-Instruct	✅	decoder_model_merged (1.45GB), decoder_model_merged_q4, embed_tokens, vision_encoder	All 4 run.
67	lxyuan/distilbert-base-multilingual-cased-sentiments-student	✅	onnx/model.onnx	Runs.
68	onnx-community/Kokoro-82M-v1.0-ONNX	⚠️	onnx/model.onnx ✅ / onnx/model_quantized.onnx ❌	fp32 TTS model runs (needs -r input_ids=0:177; the phoneme vocab is only 178 entries, and the CLI default range correctly reported an out-of-range Gather). Quantized variant fails to load: in node "/encoder/text_encoder/lstm/LSTM_quant": operator error: com.microsoft/DynamicQuantizeLSTM operator not supported or not enabled.
69	sentence-transformers/multi-qa-MiniLM-L6-cos-v1	✅	onnx/model.onnx, onnx/model_qint8_arm64.onnx	Both run.
70	MoritzLaurer/mDeBERTa-v3-base-xnli-multilingual-nli-2mil7	✅	onnx/model.onnx (1.12GB), onnx/model_quantized.onnx	Both run. DeBERTa-v3 (disentangled attention).
71	shibing624/text2vec-base-chinese	✅	onnx/model.onnx, onnx/model_qint8_avx512_vnni.onnx	Both run.
72	sentence-transformers/LaBSE	✅	onnx/model.onnx (1.88GB)	Runs.
73	thenlper/gte-large	✅	onnx/model.onnx (1.34GB), onnx/model_qint8_avx512_vnni.onnx	Both run.
74	intfloat/e5-small-v2	✅	model.onnx, onnx/model_qint8_avx512_vnni.onnx	Both run.
75	Qdrant/bge-small-en-v1.5-onnx-Q	❌	model_optimized.onnx	Fails to load: in node "Attention_0": operator error: com.microsoft/Attention operator not supported or not enabled. This is the only ONNX file in the repo (an ORT-optimized BGE-small using the MS contrib Attention fusion).
76	distilbert/distilbert-base-multilingual-cased	✅	model.onnx	Runs.
77	nomic-ai/nomic-embed-vision-v1.5	✅	onnx/model.onnx, onnx/model_quantized.onnx	Both run. Vision tower (ViT).
78	livekit/turn-detector	✅	model_quantized.onnx	Runs. Only ONNX file in the repo.
79	cross-encoder/nli-deberta-v3-small	✅	onnx/model.onnx, onnx/model_qint8_arm64.onnx	Both run. DeBERTa-v3.
80	minishlab/potion-base-8M	✅	onnx/model.onnx	Runs. Static-embedding model (model2vec).
81	oliverguhr/fullstop-punctuation-multilang-large	⏭️	—	Skipped: only variant is onnx/model.onnx + 2.24GB external data (>2GB).
82	briaai/RMBG-2.0	⏭️	—	Skipped: gated repo — 403 Forbidden / GatedRepoError when downloading without an accepted licence.
83	BAAI/bge-base-en	✅	onnx/model.onnx	Runs.
84	WhereIsAI/UAE-Large-V1	✅	onnx/model.onnx (1.34GB), onnx/model_quantized.onnx	Both run.
85	AdamCodd/vit-base-nsfw-detector	✅	onnx/model.onnx, onnx/model_quantized.onnx	Both run at 384x384 (-s height=384 -s width=384); the 224x224 default gave a correctly-reported broadcast error against the 577-position embedding table.
86	bigscience/bloom-560m	⏭️	—	Skipped: every decoder variant has 3.26GB external data (>2GB).
87	jinaai/jina-clip-v2	✅	onnx/model_int8.onnx	Runs. CLIP-style dual tower, 4 outputs. fp32 (3.45GB), fp16/q4/bnb4 all >1GB and skipped.
88	sentence-transformers/msmarco-bert-base-dot-v5	✅	onnx/model.onnx, onnx/model_qint8_arm64.onnx	Both run.
89	thenlper/gte-small	✅	onnx/model.onnx, onnx/model_qint8_avx512_vnni.onnx	Both run.
90	openai-community/gpt2-medium	✅	onnx/decoder_model.onnx, onnx/decoder_model_merged.onnx, onnx/decoder_with_past_model.onnx (1.63GB each)	All 3 run.
91	perplexity-ai/pplx-embed-v1-0.6b	❌	onnx/model_q4.onnx, onnx/model_quantized.onnx	Both fail to load: in node "/model/embed_tokens/Gather_Quant": operator error: com.microsoft/GatherBlockQuantized operator not supported or not enabled. fp32 skipped (2.1GB external data), so no variant of this repo runs.
92	Xenova/whisper-tiny.en	✅	encoder_model, decoder_model, decoder_model_merged, decoder_with_past_model, decoder_model_merged_quantized	All 5 run. Encoder needs -s feature_size=80 -s encoder_sequence_length=3000.
93	openbmb/MiniCPM-o-4_5	✅	assets/token2wav/campplus.onnx, assets/token2wav/speech_tokenizer_v2_25hz.onnx	Both run. campplus needs -s sequence_length=400 (short inputs correctly error out in AveragePool). Only ONNX files in this repo; the main LLM is not exported to ONNX.
94	mixedbread-ai/mxbai-rerank-xsmall-v1	✅	onnx/model.onnx, onnx/model_quantized.onnx	Both run.
95	shibing624/macbert4csc-base-chinese	✅	onnx/model.onnx	Runs.
96	thenlper/gte-base	✅	onnx/model.onnx, onnx/model_qint8_avx512_vnni.onnx	Both run.
97	segment-any-text/sat-3l-sm	✅	model.onnx, model_optimized.onnx	Both run. Sentence-segmentation (XLM-R based).
98	chopratejas/technique-router-onnx	✅	model_quantized.onnx	Runs. Only ONNX file in the repo.
99	google-bert/bert-base-german-cased	✅	onnx/model.onnx	Runs.
100	Marqo/marqo-fashionSigLIP	✅	onnx/text_model.onnx, onnx/vision_model.onnx, onnx/text_model_quantized.onnx, onnx/vision_model_quantized.onnx	All 4 run. SigLIP dual tower.
101	openai/privacy-filter	❌	onnx/model_q4.onnx, onnx/model_quantized.onnx	Both fail to load: in node "/model/embed_tokens/Gather_Quant": com.microsoft/GatherBlockQuantized operator not supported. fp32 (5.6GB over 3 .onnx_data_N shards) and fp16 (2.8GB) are >2GB, so no variant of this repo runs.
102	cross-encoder/stsb-roberta-large	⚠️	onnx/model.onnx (1.42GB), onnx/model_O4.onnx, onnx/model_qint8_arm64.onnx	fp32 and int8 run. model_O4 fails to load: in node "Memcpy": MemcpyFromHost operator not supported.
103	Xenova/bge-small-en-v1.5	✅	onnx/model.onnx, onnx/model_fp16.onnx, onnx/model_q4.onnx	All 3 run.
104	SmilingWolf/wd-swinv2-tagger-v3	✅	model.onnx	Runs. Swin-V2 tagger with NHWC input [2,448,448,3], output [2,10861].
105	MoritzLaurer/mDeBERTa-v3-base-mnli-xnli	✅	onnx/model.onnx (1.12GB), onnx/model_quantized.onnx	Both run. DeBERTa-v3 (disentangled attention).
106	Snowflake/snowflake-arctic-embed-m	✅	onnx/model.onnx, onnx/model_q4.onnx	Both run.
107	optimum-intel-internal-testing/tiny-random-bert	✅	onnx/model.onnx	Runs, output [1,8,32]. Originally hit the Identity-alias graph-optimization bug ("/encoder/layer.0/attention/self/query/Add" failed: conversion error for input 2: expected tensor(f32) but has tensor(i32)) — every linear bias in this model is an Identity alias of one shared initializer. Fixed above.
108	HuggingFaceTB/SmolLM2-360M-Instruct	✅	onnx/model.onnx (1.45GB), onnx/model_q4.onnx	Both run. 32 KV-cache layer pairs.
109	cross-encoder/ms-marco-TinyBERT-L2-v2	⚠️	onnx/model.onnx, onnx/model_O4.onnx, onnx/model_qint8_arm64.onnx	fp32 and int8 run. model_O4 fails to load with the same MemcpyFromHost error as #102.
110	gravitee-io/bert-small-pii-detection	✅	model.onnx, model.quant.onnx	Both run.
111	DeepBeepMeep/Wan2.1	✅	pose/dw-ll_ucoco_384.onnx, pose/yolox_l.onnx	Both run. Only ONNX files in this repo (DWPose + YOLOX preprocessors for Wan2.1); the video model itself is not ONNX.
112	Alibaba-NLP/gte-base-en-v1.5	✅	onnx/model.onnx, onnx/model_q4.onnx	Both run. New-GTE architecture (unpadded attention, RoPE).
113	1-800-BAD-CODE/xlm-roberta_punctuation_fullstop_truecase	✅	model.onnx (1.11GB)	Runs. 4 heads (pre/post punctuation, truecasing, sentence segmentation). Dynamic dims are named input_ids_dynamic_axes_1/2.
114	jinaai/jina-embeddings-v2-base-en	✅	model.onnx, model-w-mean-pooling.onnx	Both run. JinaBERT (ALiBi).
115	sentence-transformers/paraphrase-albert-small-v2	⚠️	onnx/model.onnx, onnx/model_O4.onnx, onnx/model_qint8_arm64.onnx	fp32 and int8 run. model_O4 fails to load: in node "FastGelu": com.microsoft/FastGelu operator not supported.
116	sentence-transformers/nli-mpnet-base-v2	⚠️	onnx/model.onnx, onnx/model_O4.onnx, onnx/model_qint8_arm64.onnx	fp32 and int8 run. model_O4 fails to load with the same com.microsoft/FastGelu error as #115.
117	jinaai/jina-embeddings-v2-base-code	✅	onnx/model.onnx, onnx/model_fp16.onnx, onnx/model_quantized.onnx	All 3 run.
118	cross-encoder/nli-deberta-v3-base	⚠️	onnx/model.onnx, onnx/model_O4.onnx, onnx/model_qint8_arm64.onnx	fp32 and int8 run. model_O4 fails to load with the same MemcpyFromHost error as #102.
119	ibm-granite/granite-embedding-97m-multilingual-r2	✅	onnx/model.onnx, onnx/model_quint8_avx2.onnx	Both run.
120	nomic-ai/modernbert-embed-base	✅	onnx/model.onnx, onnx/model_q4.onnx	Both run. Two outputs (token + pooled sentence embedding).
121	jinaai/jina-embeddings-v2-base-de	✅	onnx/model.onnx, onnx/model_quantized.onnx	Both run.
122	pnnbao-ump/VieNeu-TTS-v3-Turbo	✅	denoiser.onnx, speaker_encoder.onnx, onnx/vieneu_prefill.onnx, onnx/vieneu_decode_step.onnx, onnx/vieneu_acoustic_cached.onnx	All 5 run. Speaker encoder needs -s sequence_length=400; prefill/decode share one external-data blob (onnx/vieneu_backbone_shared.data) referenced by both models.
123	distilbert/distilbert-base-cased	✅	model.onnx	Runs.
124	HuggingFaceTB/SmolVLM-500M-Instruct	✅	onnx/embed_tokens.onnx, onnx/vision_encoder.onnx, onnx/decoder_model_merged.onnx (1.45GB), onnx/decoder_model_merged_q4.onnx	All 4 run. VLM split into embed / vision / decoder; vision encoder takes [2,4,3,512,512] pixel values.
125	sentence-transformers/paraphrase-MiniLM-L12-v2	⚠️	onnx/model.onnx, onnx/model_O4.onnx, onnx/model_qint8_arm64.onnx	fp32 and int8 run. model_O4 fails to load with the same MemcpyFromHost error as #102.
126	answerdotai/answerai-colbert-small-v1	✅	onnx/model.onnx, onnx/model_q4.onnx	Both run.
127	google/mt5-small	✅	onnx/encoder_model.onnx, onnx/decoder_model.onnx (1.13GB), onnx/decoder_model_merged.onnx (1.13GB), onnx/decoder_with_past_model.onnx (1.11GB)	All 4 run. mT5 encoder-decoder with cross-attention KV cache.
128	Snowflake/snowflake-arctic-embed-m-v1.5	✅	onnx/model.onnx, onnx/model_q4.onnx	Both run.
129	jinaai/jina-embeddings-v5-text-small-retrieval	⏭️	—	Skipped: only variant is onnx/model.onnx + 2.38GB external data (>2GB).
130	Qdrant/paraphrase-multilingual-MiniLM-L12-v2-onnx-Q	❌	model_optimized.onnx	Fails to load: in node "Attention_0": com.microsoft/Attention operator not supported. Only ONNX file in the repo, same gap as #75.
131	briaai/RMBG-1.4	✅	onnx/model.onnx, onnx/model_fp16.onnx, onnx/model_quantized.onnx	All three run ([2,3,1024,1024] input). model_fp16 initially failed to load with f16 tensor data is not 2-byte aligned; fixed by the empty-slice cast fix above.
132	DeepBeepMeep/LTX-2	✅	sherpa/3dspeaker_speech_eres2net_base_sv_zh-cn_3dspeaker_16k.onnx, sherpa/sherpa-onnx-pyannote-segmentation-3-0/model.onnx	Both run (speaker embedding at -s T=400, diarisation at -s T=32000). Only ONNX files in the repo; the video model is not ONNX.
133	onecxi/open-vakgyata	✅	onnx/model_quantized.onnx	Runs at -s num_samples=16000. Wav2Vec2 classifier; short inputs correctly error out in the feature-extractor Conv stack. Only ONNX file in the repo.
134	protectai/deberta-v3-base-prompt-injection-v2	✅	onnx/model.onnx	Runs.
135	dslim/distilbert-NER	✅	onnx/model.onnx	Runs.
136	sentence-transformers/stsb-roberta-base-v2	⚠️	onnx/model.onnx, onnx/model_O4.onnx, onnx/model_qint8_arm64.onnx	fp32 and int8 run. model_O4 fails to load: in node "Attention_0": com.microsoft/Attention operator not supported.
137	sentence-transformers/msmarco-MiniLM-L6-v3	✅	onnx/model.onnx, onnx/model_qint8_arm64.onnx	Both run.
138	cross-encoder/nli-MiniLM2-L6-H768	⚠️	onnx/model.onnx, onnx/model_O4.onnx, onnx/model_qint8_arm64.onnx	fp32 and int8 run. model_O4 fails to load with the same MemcpyFromHost error as #102.
139	LazarusNLP/all-indo-e5-small-v4	✅	onnx/model.onnx	Runs.
140	HuggingFaceTB/SmolLM-135M	✅	onnx/model.onnx, onnx/model_q4.onnx	Both run.
141	Davlan/bert-base-multilingual-cased-ner-hrl	✅	onnx/model.onnx	Runs.
142	Qdrant/all_miniLM_L6_v2_with_attentions	✅	model.onnx	Runs. Returns per-layer attention matrices alongside the embedding (8 outputs).
143	Qdrant/Splade_PP_en_v1	✅	model.onnx	Runs. SPLADE sparse retrieval head, output [2,32,30522].
144	Qdrant/multilingual-e5-large-onnx	⏭️	—	Skipped: only variant is model.onnx + 2.24GB external data (>2GB).
145	sentence-transformers/bert-base-nli-mean-tokens	⚠️	onnx/model.onnx, onnx/model_O4.onnx, onnx/model_qint8_arm64.onnx	fp32 and int8 run. model_O4 fails to load with the same MemcpyFromHost error as #102.
146	cross-encoder/nli-deberta-v3-large	✅	onnx/model.onnx (1.74GB), onnx/model_qint8_arm64.onnx	Both run.
147	sentence-transformers/msmarco-distilbert-base-tas-b	⚠️	onnx/model.onnx, onnx/model_O4.onnx, onnx/model_qint8_arm64.onnx	fp32 and int8 run. model_O4 fails to load with the same com.microsoft/Attention error as #136.
148	Maykeye/TinyLLama-v0	✅	model.onnx	Runs, output [1,8,32000]. Originally hit the same Identity-alias bug as #107 ("/model/model/layers.1/post_attention_layernorm/Mul_1" failed: conversion error for input 1); all 8 RMSNorm weights are Identity aliases of layers.1.input_layernorm.weight.
149	mixedbread-ai/deepset-mxbai-embed-de-large-v1	✅	onnx/model.onnx (1.94GB), onnx/model_q4.onnx	Both run.
150	mattmdjaga/segformer_b2_clothes	✅	onnx/model.onnx	Runs. SegFormer-B2, output [2,18,56,56].
151	MoritzLaurer/DeBERTa-v3-large-mnli-fever-anli-ling-wanli	✅	onnx/model.onnx (1.74GB), onnx/model_quantized.onnx	Both run.
152	jonathandinu/face-parsing	✅	onnx/model.onnx, onnx/model_quantized.onnx	Both run. SegFormer face parser.
153	notmax123/Zonos-Hebrew	✅	phonikud-1.0.onnx	Runs. Hebrew diacritics tagger (3 heads); only ONNX file in the repo.
154	jinaai/jina-colbert-v2	⏭️	—	Skipped: only variant is onnx/model.onnx + 2.23GB external data (>2GB).
155	Xenova/multilingual-e5-small	✅	onnx/model.onnx, onnx/model_q4.onnx	Both run.
156	Xenova/segformer-b0-finetuned-ade-512-512	✅	onnx/model.onnx, onnx/model_fp16.onnx, onnx/model_quantized.onnx	All three run (-s height=512 -s width=512 -s num_channels=3, output [1,150,128,128]). model_fp16 initially failed to load with the same empty-f16-tensor error as #131.
157	oliverguhr/german-sentiment-bert	✅	onnx/model.onnx	Runs.
158	Xenova/bge-reranker-base	✅	onnx/model.onnx (1.11GB), onnx/model_q4.onnx	Both run.
159	cross-encoder/stsb-roberta-base	⚠️	onnx/model.onnx, onnx/model_O4.onnx, onnx/model_qint8_arm64.onnx	fp32 and int8 run. model_O4 fails to load with the same MemcpyFromHost error as #102.
160	Xenova/bge-m3	✅	onnx/model_fp16.onnx (1.13GB), onnx/model_int8.onnx	Both run. fp32 is 2.27GB external data (>2GB).
161	Snowflake/snowflake-arctic-embed-m-v2.0	✅	onnx/model.onnx (1.23GB), onnx/model_q4.onnx	Both run.
162	HuggingFaceTB/SmolLM2-1.7B-Instruct	✅	onnx/model_q4f16.onnx (1.11GB), onnx/model_int8.onnx (1.71GB)	Both run. 24 KV-cache layer pairs. fp32 is 6.85GB external data (>2GB).
163	optimum-intel-internal-testing/tiny-random-vit	✅	onnx/model.onnx	Runs with -s height=30 -s width=30 -s num_channels=3 (this test model has image_size: 30), output [1,2]. Originally hit the same Identity-alias bug as #107 and #148 ("/vit/encoder/layer.0/layernorm_before/Add_1" failed: conversion error for input 2).
164	sentence-transformers/clip-ViT-B-32-multilingual-v1	⚠️	onnx/model.onnx, onnx/model_O4.onnx, onnx/model_qint8_arm64.onnx	fp32 and int8 run. model_O4 fails to load with the same com.microsoft/Attention error as #136.
165	NX-AI/TiRex	❌	tirex.onnx	Fails to load: Scan operator not supported or not enabled. Time-series forecasting model; the only ONNX file in the repo. Scan is an ai.onnx standard op, not a contrib op.
166	Xenova/paraphrase-multilingual-MiniLM-L12-v2	✅	onnx/model.onnx, onnx/model_q4.onnx	Both run.
167	mixedbread-ai/mxbai-embed-2d-large-v1	✅	onnx/model.onnx (1.34GB), onnx/model_quantized.onnx	Both run.
168	intfloat/e5-small	✅	model.onnx	Runs.
169	cross-encoder/nli-deberta-v3-xsmall	⚠️	onnx/model.onnx, onnx/model_O4.onnx, onnx/model_qint8_arm64.onnx	fp32 and int8 run. model_O4 fails to load with the same MemcpyFromHost error as #102.
170	unsloth/bge-small-en-v1.5	✅	onnx/model.onnx	Runs.
171	dslim/bert-large-NER	✅	onnx/model.onnx (1.33GB)	Runs.
172	sentence-transformers/distilbert-base-nli-mean-tokens	⚠️	onnx/model.onnx, onnx/model_O4.onnx, onnx/model_qint8_arm64.onnx	fp32 and int8 run. model_O4 fails to load with the same com.microsoft/Attention error as #136.
173	onnx-community/embeddinggemma-300m-ONNX	✅	onnx/model.onnx (1.23GB), onnx/model_no_gather_q4.onnx	fp32 and the int4 no_gather export run. model_q4/model_q4f16 fail to load: they quantize the embedding table with com.microsoft/GatherBlockQuantized over a native ONNX INT4 initializer, and hit the INT4 gap first. model_no_gather_q4 keeps the embedding table in fp32 and uses only MatMulNBits, so the repo has a working int4 variant.
174	Alibaba-NLP/gte-modernbert-base	✅	onnx/model.onnx, onnx/model_q4.onnx	Both run.
175	sentence-transformers/msmarco-MiniLM-L12-cos-v5	✅	onnx/model.onnx, onnx/model_qint8_arm64.onnx	Both run.
176	avsolatorio/GIST-all-MiniLM-L6-v2	✅	onnx/model.onnx, onnx/model_quantized.onnx	Both run.
177	Xenova/clip-vit-base-patch32	✅	onnx/model.onnx, onnx/text_model.onnx, onnx/vision_model.onnx, onnx/model_q4.onnx	All 4 run. Combined dual tower plus the separate text/vision towers.
178	HuggingFaceTB/SmolLM3-3B-Base	⏭️	—	Skipped: smallest variant is onnx/model_q4f16.onnx at 2.12GB external data (>2GB).
179	shibing624/text2vec-base-multilingual	✅	onnx/model.onnx	Runs.
180	immich-app/ViT-B-32__openai	✅	textual/model.onnx, visual/model.onnx	Both run. Fixed shapes ([1,77] text, [1,3,224,224] image).
181	Organika/sdxl-detector	✅	onnx/model.onnx	Runs. Swin-based detector.
182	protectai/unbiased-toxic-roberta-onnx	✅	model.onnx, model_quantized.onnx	Both run.
183	TahaDouaji/detr-doc-table-detection	✅	onnx/model.onnx	Runs. DETR detector, outputs logits [2,100,3] and pred_boxes [2,100,4].
184	protectai/xlm-roberta-base-language-detection-onnx	⚠️	model.onnx (1.11GB), model_optimized.onnx, model_quantized.onnx	fp32 and int8 run. model_optimized fails to load with the same com.microsoft/Attention error as #136.
185	MonsterMMORPG/Wan_GGUF	✅	RMBG2/onnx/model.onnx, IDM-VTON/.../parsing_atr.onnx, Viso_Master_Models/{det_10g,GFPGANv1.4,inswapper_128.fp16,w600k_r50,yoloface_8n,realesr-general-x4v3,occluder,2d106det}.onnx	A 96-file model dump; 10 mainstream models tested, all 10 run — RMBG-2.0 matting (at -s height=1024 -s width=1024), human parsing, GFPGAN restoration, ArcFace recognition, YOLO face detection, Real-ESRGAN 4x upscaling, occlusion masking, SCRFD detection (-s ?=640), 106-point landmarks and inswapper_128.fp16 face swapping. The last of those initially failed to load with the same empty-f16-tensor error as #131.
186	Jean-Baptiste/camembert-ner	✅	model.onnx	Runs. CamemBERT NER.
187	sentence-transformers/paraphrase-xlm-r-multilingual-v1	⚠️	onnx/model.onnx (1.11GB), onnx/model_O4.onnx, onnx/model_qint8_arm64.onnx	fp32 and int8 run. model_O4 fails to load with the same com.microsoft/Attention error as #136.
188	SimianLuo/LCM_Dreamshaper_v7	✅	text_encoder/model.onnx, vae_decoder/model.onnx, vae_encoder/model.onnx	All 3 run. The VAE decoder needs a 4-channel latent (-s num_channels_latent=4). The UNet is 3.44GB external data (>2GB).
189	MoritzLaurer/deberta-v3-large-zeroshot-v2.0	✅	onnx/model.onnx (1.74GB)	Runs.
190	MoritzLaurer/bge-m3-zeroshot-v2.0	⏭️	—	Skipped: only variant is onnx/model.onnx + 2.27GB external data (>2GB).
191	onnx-community/whisper-small	✅	onnx/encoder_model.onnx, onnx/decoder_model.onnx, onnx/decoder_model_merged.onnx, onnx/decoder_with_past_model.onnx, onnx/decoder_model_merged_q4.onnx	All 5 run. Encoder needs -s feature_size=80 -s encoder_sequence_length=3000.
192	wan-world/Wan2.1-2.2	✅	pose/dw-ll_ucoco_384.onnx, pose/yolox_l.onnx	Both run. Same two DWPose/YOLOX preprocessors as #111; the video model is not ONNX.
193	Xenova/clip-vit-base-patch16	✅	onnx/model.onnx, onnx/text_model.onnx, onnx/vision_model.onnx, onnx/model_quantized.onnx	All 4 run.
194	Xenova/whisper-base.en	✅	onnx/encoder_model.onnx, onnx/decoder_model.onnx, onnx/decoder_model_merged.onnx, onnx/decoder_with_past_model.onnx, onnx/decoder_model_merged_q4.onnx	All 5 run. Encoder needs -s feature_size=80 -s encoder_sequence_length=3000. This export also returns per-layer decoder and cross attentions.
195	sentence-transformers/msmarco-distilbert-base-v4	⚠️	onnx/model.onnx, onnx/model_O4.onnx, onnx/model_qint8_arm64.onnx	fp32 and int8 run. model_O4 fails to load with the same com.microsoft/Attention error as #136.
196	cross-encoder/stsb-distilroberta-base	⚠️	onnx/model.onnx, onnx/model_O4.onnx, onnx/model_qint8_arm64.onnx	fp32 and int8 run. model_O4 fails to load with the same MemcpyFromHost error as #102.
197	redis/langcache-embed-v1	✅	onnx/model.onnx	Runs.
198	optimum-intel-internal-testing/tiny-random-Phi3ForCausalLM	✅	onnx/model.onnx, onnx/model_q4.onnx	Both run, output [1,8,32064] plus 2 KV-cache layer pairs. Both originally hit the same Identity-alias bug as #107 ("/model/layers.0/post_attention_layernorm/Mul_1" failed: conversion error for input 1).
199	sentence-transformers/msmarco-distilbert-cos-v5	✅	onnx/model.onnx, onnx/model_qint8_arm64.onnx	Both run.
200	segment-any-text/sat-12l-sm	✅	model.onnx, model_optimized.onnx	Both run.