## qualcomm/aimet — 2.40.0…2.41.0

_32 commits._

### Features
- **Support Qwen2.5(-VL), Qwen3-VL, Phi-3.5/4 and Gemma2/3 in analyze_llm_topology (#7916)** (49e3fba)

### Fixes
- **Fix stale CG products when an op output feeds a consumer twice (#7900)** (db7a805)
- **Fix bug in Pad op encoding propagation** (5f80b04)
- **Fix AdaScale silently overriding per-quantizer weight bitwidth (#7872)** (4591f60)
- **Fix AIMETExportedProgram onnx qdq export failure** (d3fc449)
- **Fix failing dependency bump workflow (#7858)** (3062f56)

### Backend
- **Update release notes for 2.41.0 release (#7914)** (973a7a4)
- **Validate encodings before converting them to ONNX QDQ (#7915)** (5fd6850)
- **Read head_dim and KV heads per layer for transformers 5.17 heterogeneous configs (#7921)** (0563f42)
- **Grace grader: deterministic kernels, fixed seed, cuDNN attention off (#7918)** (c261738)
- **Make analyze_llm_topology the single entry point (#7920)** (4775b80)
- **Suppress int32 encoding propagation during ATen/ONNX QDQ export** (c9f0cf7)
- **Find LLM topology by HF module name in analyze_llm_topology (#7903)** (b574e49)
- **Refactor param identification outside of CG** (2df88d9)
- **Separate ir_utils from fusion utils** (9e90b5d)
- **Remove quantizers on ir model during export** (6a6824d)
- **Represent control-flow subgraph nodes in ConnectedGraph** (59f4031)
- **Implement top-level configure_llm API** (7bcb318)
- **Remove quadratic GraphModule recompilation from AIMETExportedProgram** (89c8e80)
- **Delete unused _insert_data_movement_op_output_quantizer util** (c920c7c)
- **Delete redundant cg tensor_dict attribute** (5eb5c9a)
- **Consolidate ir_utils into one file** (c7727ad)
- **Use trivial encoding when an uninitialized quantizer receives empty input** (5561564)
- **Scope AdaScale mutation context to forward/backward** (342ed10)
- **Analyze LLM topology once up front in ONNX GenAI runner (#7870)** (a7f1c31)
- **Added one more dir to ignore during sync to mlops (#7869)** (f7f4abc)
- **Change to use uv pip in Gen AI Lab to speed up setup (#7868)** (1825415)
- **Derive linear-attention chunk size from sequence length (#7841)** (52c228c)
- **Decouple FP4 quantizer from MXFP4** (065e5a0)
- **Remove AIMETExportedProgram dependency on legacy meta data** (95a74e6)
- **Update HF deps (#7864)** (2b95bb9)
- **Work around backward-incompatible change in onnx 1.23.0** (88b4cb6)

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