v3.6.0

Summary

Added the most adapters in a single release, we now support over 100 architectures. We also bumped our transformers version up to 5.13.0 to support newer architectures

What's Changed

• Add Arcee squared-ReLU MLP adapter (ArceeForCausalLM) by @pablocs116 in #1478
• Redo of #1477: Qwen2-MoE bridge support (retarget to dev) by @jlarson4 in #1482
• SSM Interp Support by @jlarson4 in #1481
• Add Falcon-H1 parallel hybrid architecture adapter (FalconH1ForCausalLM) by @udapy in #1480
• Add BD3LM architecture adapter by @puranikyashaswin in #1479
• Add RecurrentGemma (Griffin) architecture adapter by @pablocs116 in #1483
• Add Cohere2 architecture adapter by @hsusul in #1485
• feat: add Zamba2ForCausalLM architecture adapter by @mukund1985 in #1486
• Centralize soft-cap logic behind a shared helper and a named disabled… by @delaidam in #1491
• Added Ouro architecture adapter and integration tests for OuroForCausalLM by @Canonik in #1488
• fix: migrate jupyter dependency group to Notebook 7 by @chandrudp29 in #1494
• add t5gemma2 adapter and tests by @zinichakraborty in #1495
• Add OLMo 3 base model aliases by @abhinav-bellapu in #1500
• Fix n_devices max_memory caps by @abhinav-bellapu in #1501
• Add Apertus adapter tests and fix nested key translation by @abhinav-bellapu in #1503
• Keep input_tokens on the model device in generate() by @danielxmed in #1509
• Document all four adapter registration locations by @abhinav-bellapu in #1514
• Add DeepSeek V4 architecture adapter by @abhinav-bellapu in #1513
• Add forward-only LLaDA architecture adapter by @hsusul in #1512
• Fix 4-bit split QKV projection by @abhinav-bellapu in #1502
• feat: enable meta device targets in TransformerBridge device_map by @hiro-nikaitou in #1493
• Add Lfm2 support by @TensorCruncher in #1510
• Multigpu boot transformers by @jlarson4 in #1518
• Added HRM-Text two-timescale recurrent adapter by @MdSadiqMd in #1487
• Add Jacobian lens (J-lens): published-artifact loading, readout, native fitting, and interventions by @danielxmed in #1507
• Add Zamba2 architecture adapter unit tests by @abhinav-bellapu in #1522
• Add dedicated GraniteMoe adapter tests (#1302) by @delaidam in #1524
• Updating documentation for future deprecation by @jlarson4 in #1523
• feat: add RWKV-7 (Goose) TransformerBridge adapter (RWKV7ForCausalLM) by @mukund1985 in #1521
• feat: add RavenForCausalLM (Huginn) depth-recurrent adapter by @mukund1985 in #1520
• Fix Tracr demo output_label: compiled labels are auto-numbered by @dewstend in #1525
• Add TransformerBridge adapter for lightweight decoder-only pretraining models by @Cacapice in #1519
• Add config tests for DeepSeek V2 architecture adapter by @SanjidMzi in #1529
• Remove dead _last_input_before_norm write in NormalizationBridge by @Hetens in #1528
• Fix silent hook_normalized/hook_scale no-op on NormalizationBridge native-autograd path by @Hetens in #1527
• Update integration testing explainer docs to better outline our system by @jlarson4 in #1534
• Add config tests for DeepSeek V3 architecture adapter by @SanjidMzi in #1531
• Add Jamba attention+Mamba interleave adapter (JambaForCausalLM) by @udapy in #1535
• Add config tests for mamba and mamba2 architecture adapter by @MdSadiqMd in #1532
• Mistral: use fork-capable attention bridge (enable hook_attn_in / set_use_attn_in) by @almogtavor in #1497
• Add lfm2 integration test by @TensorCruncher in #1530
• feat(jacobian_lens): add artifact registry + short-name resolution in from_pretrained by @mukund1985 in #1537
• Architecture Gaps Fable Sweep by @jlarson4 in #1542
• Release 3.6.0 by @jlarson4 in #1548

New Contributors

• @pablocs116 made their first contribution in #1478
• @udapy made their first contribution in #1480
• @delaidam made their first contribution in #1491
• @zinichakraborty made their first contribution in #1495
• @danielxmed made their first contribution in #1509
• @hiro-nikaitou made their first contribution in #1493
• @dewstend made their first contribution in #1525
• @Cacapice made their first contribution in #1519
• @SanjidMzi made their first contribution in #1529
• @Hetens made their first contribution in #1528
• @almogtavor made their first contribution in #1497

Full Changelog: v3.5.1...3.6.0