## qualcomm/aimet — 2.35.1…2.36.0

_40 commits._

### Features
- **Add minor updates to 2.36 release note** (db9e710)
- **Add R2 per-head rotation to aimet-torch SpinQuant** (0ba884a)
- **Support Gemm layers in model converter** (ea71d76)
- **add fp16 support in GenAILab (#7439)** (fec7c2c)
- **Add SplitFusedLayers adaptation to GenAI Lab (#7445)** (0c1729d)
- **Add tensor-level set_precision API to aimet-onnx QuantSim** (ad61748)
- **Add tqdm progress bar to Sequential MSE (#7422)** (26f4c25)
- **Add MMLU Pro dataset/metric to GenAI Lab (#7408)** (c16100b)

### Fixes
- **Fix bug in unit test surfaced by ORT 1.28** (3ac30da)
- **Fix NightlyTest failure** (6fa2c2d)
- **Fix export omitting bias encodings for dynamic convs** (326167c)
- **Fix bugged or missing op types in onnx2torch** (6f3255b)
- **Fix broken tokenizers version (#7412)** (229210c)
- **Fix bfloat16 test failure in release pipeline** (fcede8a)

### Backend
- **Temporarily skip test_torch_to_onnx** (b6e5ff5)
- **Update version and release notes for 2.36.0 release** (6a1caac)
- **Onboard gemma4 e4b qat model on GenAI Lab (#7461)** (eabacc4)
- **Rebind instance-bound methods in seq_mse FP shared-weights copy** (83c3454)
- **Resolve false positive unit test failure with deepspeed 0.19.3** (001e1fb)
- **Document onnx mixed-precision APIs in apiref** (a489e5d)
- **Skip weightless dynamic MatMuls in ONNX sequential MSE** (8413aea)
- **Detect decoder residual writers by graph walk in role map** (22e6052)
- **Absorb decoder-forwarded kwargs in exportable GatedDeltaNet forward** (abd4863)
- **Fall back to chained sampling for cached-region resume in BlockwiseSampler (#7441)** (5341997)
- **Onboard gemma4 qat model to GenAILab (#7449)** (f65cf66)
- **Migrate qai-hub-models evaluate imports to .helpers (0.58.0) (#7457)** (dc3bca5)
- **Update version and release notes for 2.35.1 release** (ccbe21b)
- **Skip deriving data movement op output encoding for MaskedSoftmax subgraph** (75bf0f7)
- **Fall back from Triton to PyTorch if input size exceeds 2^31** (a8636b5)
- **Move aimet-onnx/torch joint tests to NightlyTests/common** (4f14a01)
- **Propagate through float-to-float casts to find effective quantizer** (9ca5115)
- **Prevent HTP-specific overflow/underflow issues** (a052923)
- **Enable AdaScale for Qwen3.5 (torch)** (e5f5462)
- **Remove erroneous QuantizationMixin.ignore (#7420)** (dcfe9de)
- **Remove outdated aimet_tf note from install docs** (a880422)
- **Enable AdaScale for Qwen3.5 in aimet-onnx** (03df10a)
- **Define QcQuantizeOp interface to set precision via qtype** (d0c1e4b)
- **Rebased original VLM refactor branch (#7399)** (cf0627d)
- **Update pinned package dependencies (#7402)** (7b3f0fb)
- **Consolidate ONNX decoder block detection into block_topology package** (f441f2f)

_Recap by [Repo Wrapped](https://repowrapped.com/gh/qualcomm/aimet?utm_source=github-action)._