## qualcomm/aimet — 2.36.0…2.37.0

_47 commits._

### Features
- **Add support for RotaryEmbedding Op in AIMET torch (#7545)** (2f50387)
- **Add ONNX quantization sensitivity analysis and result visualization (#7532)** (5da2808)
- **Add C++ FP8 QDQ path for AIMET-ONNX tensor quantizers (#7513)** (9bed6b2)
- **Add GenAI Lab recipe schema (#7392)** (ad454cf)
- **Add ONNX quant stats range visualizer (aimet_onnx.analysis.visualize_stats) (#7523)** (3284154)
- **Add apply_param_qdq opt-out to sim.export (#7472)** (c09a36f)
- **Add HTP quantsim config for v85 (#7442)** (3101924)

### Fixes
- **Fix OpOverloadPacket.op_overloads() call (#7577)** (e8c682e)
- **Fix stopping criteria for sequence lengths > 1** (b255967)
- **Fix flaky bfloat16 qc_quantize_op test** (c79f340)
- **Fix how model modifiers are reported (#7557)** (c22efc9)
- **Fix QuantizedLora attribute error** (d016c84)
- **Fix Qwen3.5 linear-attention mask patch for transformers >=5.13.0** (65a5b21)
- **Fix github action boolean vs string type mismatch bug** (093a105)
- **Fix bug in unit test surfaced by ORT 1.28** (a4e26b0)
- **Fix NightlyTest failure** (6c73009)

### Backend
- **Request 64 CPUs for unit test runners (#7611)** (a9dfd9c)
- **Update release notes for 2.37.0 (#7576)** (6ab858f)
- **Unblock torch nightly: pin deepspeed and stop test state leaks (#7595)** (eacb902)
- **Treat aten.as_strided as grid-preserving op** (988017b)
- **Avoid weight load and save during SeqMSE** (1ac2d92)
- **Install build deps at job time instead of a prebuilt docker image** (91d36f2)
- **Use onnx_ir shape inference in QuantSim init** (0c7ab35)
- **Only build dummy input if necessary** (b73aa0f)
- **Use smaller models for repeated to_onnx_qdq test** (49873e6)
- **Pre-compile config json schema validator** (5120fa6)
- **Harden qai-hub-models ignore-registration against non-import errors (#7563)** (246c77f)
- **Guard analytical bias-scale against fp16 export-dtype underflow (#7554)** (9728bba)
- **Re-apply exception rules after lite-MP precision change** (e80073a)
- **Refactor ONNX block_topology into llm_topology + crisp BlockTopology (#7559)** (42f3eca)
- **Align ATen IR with QNN IR before graph-based calibration** (fe538d1)
- **Refactor TfEncoding utils out of LPBQ utils** (802ec67)
- **Deprecate enable_recompute API** (c85549d)
- **Define QSpec dataclass for quantizer configuration** (29d94d4)
- **Treat  aten._unsafe_view as grid-preserving op** (1c66b17)
- **Migrate pypi-release-pipeline to ephemeral runners** (2a52ab2)
- **Replace aimet_common import with aimet_torch.common in unit test** (d5a04f1)
- **Revert "Temporarily disable slackapi Slack notifications (#7521)" (#7530)** (9600a51)
- **Temporarily disable slackapi Slack notifications (#7521)** (05c3953)
- **Temporarily skip flaky deepspeed unit test** (b0b230b)
- **Explicitly specify shell type as bash** (4c3c589)
- **Init commit (#7503)** (fcab868)
- **Remove hardcoded tensor names from AdaScale** (efc1e5c)
- **Temporarily skip test_torch_to_onnx** (a6a208d)
- **Remove legacy dependency for aimet_torch.v1** (cc1b580)
- **Raise error on shared bias tensor with analytic scale** (43ede58)
- **Recover pre-quantized weight grid scale via divisor search (#7411)** (8e91e22)

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