## qualcomm/aimet — 2.34.0…2.35.0

_38 commits._

### Features
- **Add stopping technique criteria (#7395)** (9d82ce9)
- **update python and ubuntu support list (#7393)** (7e67530)
- **Add GenAI Lab Dashboard SQL Queries (#7382)** (1b6b069)
- **Add ScatterElements converter to onnx2torch_ext (#7223)** (71300be)
- **Add GenAI Lab metric versioning + bump MMMU version (#7381)** (c70529a)
- **Add fixes & workarounds for torch 2.13.0** (7637a27)
- **Support lossless calibration of prequantized weight given min(|W.nonzero()|) = 1 * scale** (5fc1007)
- **Support internal bfloat16 tensors in onnx QuantSim** (2fa6fe9)
- **Add architecture-coverage tests for decoder block detection in ONNX** (8636ee0)
- **GenAILab: add SelfGenerated calibration dataset (#7281)** (058d0ee)
- **Add Gather output-quantizer override to HTP V68 config** (2f7dedb)
- **Support rsqrt-based RMSNorm pattern in ONNX graph passes** (f02eed2)

### Fixes
- **Fix AMP phase 1 bug with tied quantizers** (6eef3ba)
- **Fix Windows ONNX CI: ARM64 import error and AMD64 disk exhaustion** (b4eb348)
- **Fix skipped/flooding nightly Slack failure notifications** (2379e75)
- **Fix protobuf 2GB limit by using external data for large models for aimet_torch QDQ export (#7373)** (1b7dbab)
- **Fix onnx-gpu and torch pipelines** (d23b82c)
- **Pin huggingface-hub<1.0 in regression job; drop dead requirements.txt (#7372)** (0165bde)
- **Authenticate to HuggingFace in regression runs to avoid crashes (#7353)** (94c35e2)

### Backend
- **Update version and release notes for 2.35.0 release** (948a663)
- **Relax posenet_mobilenet test threshold** (bc8b20e)
- **Revert "Make export_int32_bias True during export and to_onnx_qdq (#7133)" (#7400)** (4c6acfe)
- **Update repo references from quic to qualcomm (#7348)** (4041fd8)
- **Unpin pytorch** (7d3b050)
- **Make export_int32_bias True during export and to_onnx_qdq (#7133)** (e2a7cd7)
- **Export mxfp4-int8 weights on mxfp4 grid** (2f45120)
- **Allow spinquant R1 for models with no lm_head** (ce5933d)
- **Refactor supported-platforms into per-package matrix** (0909f8a)
- **Make PR jobs runnable with manual workflow dispatch** (198f4d9)
- **Enable SpinQuant R1 for Qwen3.5 (torch) (#7358)** (6a47c31)
- **disable Mobilenetv2 AMP test (#7356)** (528ad88)
- **Implement lossless encoding calibration for pre-quantized int2 weights** (9f8c52e)
- **Let _remove_intermedate_identity_nodes pass remove redundant QDQ sequences** (5b4f9ee)
- **Improve baseline comparison and sourcing (#7310)** (95747c2)
- **Temporarily pin aimet-onnx CI pipeline's aimet-torch version to <2.34** (ce46c65)
- **Unify the number of evaluation samples across fp, quantsim, and onnx qdq** (081f7c0)
- **Work around torch.jit.trace input order magling issue for custom.Gather** (d261af1)
- **Duplicate QDQ input when tensor consumed by non QDQ ops** (0d269bb)

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