## qualcomm/aimet — 2.33.0…2.34.0

_36 commits._

### Features
- **Add internvl to adascale onnx types (#7331)** (7aebfe6)
- **Support Mul/Div ops between R3 rotation and QK matmul** (44add4d)
- **Support SHA graphs in spinquant R2** (6962b91)
- **Support SHA+GQA graphs in spinquant R3** (cd43000)
- **Add producer as optional attribute of EncodingBase** (4eff161)
- **Add fold_param_quantizers() to aimet-onnx QuantizationSimModel (#7256)** (bc1b47e)
- **Support ScatterElements kvcache in _tie_quantizers_for_kv_cache** (9455175)

### Fixes
- **Fix sliding window attn mask computation for seq len > 1 (#7203)** (8a73c01)
- **Fix major bugs in aimet-torch int32 bias overflow protection** (60c53ff)
- **Work around torch.jit.trace bug upon Gather-ScatterElements sequence** (170ccf4)
- **Fix trivially passing test cases** (1c7e3d1)

### Backend
- **Duplicate QDQ input when tensor consumed by non QDQ ops (#7340)** (1babd98)
- **release 2.34 notes and version bump (#7326)** (d9ab55b)
- **Make scaled (sum-over-seq-dim) loss the default for AdaScale** (c1a2c44)
- **Emit per-channel axis for fused bias quantizers in ONNX QDQ export (#7314)** (725ffe7)
- **Implement int32 bias overflow protection in aimet-torch** (d1b9b0c)
- **Make Adascale block loss function configurable** (25aeac5)
- **Precompute transposed MatMul weight in aimet-torch ONNX QDQ export time** (a203e14)
- **Enable exporting FloatQuantizeDequantize with dynamo-based onnx exporter** (97e6411)
- **Re-implement Qwen 3.5 PR (#7265)** (e81e841)
- **Refactor GenAI model results in AIMET docs (#7294)** (b9d1a10)
- **Align QcQuantizeOp.quantize_dequantize's output dtype with input** (9e11952)
- **Remove graph breaks from almost all quantized modules** (18501cb)
- **Enable AdaScale for gemma4 in aimet-torch (#7279)** (6748057)
- **Catch internvl import error (#7246)** (3737b16)
- **CLI arg to assume a different role when invoking stop_pod.sh (#7266)** (9b93c8f)
- **Run GenAILab torch SpinQuant before quantsim creation (#7263)** (4289289)
- **Represent mxfp4 e8m0 scale with float32** (ccdeabc)
- **Pass down block_size attribute through FloatEncoding.to()** (1feeb5f)
- **Exclude SpinQuant R3 online rotations from sequential MSE (#7259)** (901cd93)
- **GenAILab: mark pre-sim SpinQuant in ONNX recipe results (#7262)** (b084991)
- **Refactor test_default_kernels function for more readible failure message** (22766c7)
- **Update release notes and version for 2.33.0 release (#7252)** (26fdc1a)
- **Quiet S3 upload progress logs with --no-progress (#7250)** (9e71d31)

### Docs
- **Docs: add Qwen3-8B, Gemma3 recipe rows and label SpinQuant r1 (#7300)** (7ea4644)
- **Docs: add Qwen3 0.6B/1.7B recipe rows and fix per-model table banding (#7296)** (b5452d7)

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