2.34.0

• Bug fixes and Improvements

  • ONNX

    • Added fold_param_quantizers() to QuantizationSimModel to bake param QDQ into initializers for faster inference (bc1b47e18)
    • Excluded SpinQuant R3 online rotations from sequential MSE to avoid shape mismatches under GQA (901cd936d)
    • Supported ScatterElements KV-cache pattern in _tie_quantizers_for_kv_cache (94551756e)
    • Aligned QcQuantizeOp.quantize_dequantize output dtype with input (9e1195254)
    • Emit per-channel axis for fused bias quantizers in ONNX QDQ export (725ffe724)

  • Torch

    • Implemented int32 bias overflow protection (d1b9b0c2b)
    • Removed graph breaks from almost all quantized modules for improved torch.compile performance (18501cbdb)
    • Precomputed transposed MatMul weight at ONNX QDQ export time, removing the intermediate transpose in nn.Linear export (a203e14fe)
    • Enabled exporting FloatQuantizeDequantize with the dynamo-based ONNX exporter (97e64111a)
    • Worked around torch.jit.trace bug on Gather-ScatterElements sequences (170ccf418)
    • Represented mxfp4 e8m0 scale with float32 (ccdeabc10)
    • Propagated block_size attribute through FloatEncoding.to() (1feeb5fc5)
    • Added optional producer attribute to EncodingBase (4eff161e5)
    • Skip importing InternVL on transformers < 4.52 without raising (3737b161f)
    • AdaScale: Enabled AdaScale for Gemma 4 (674805703)

  • Common

    • AdaScale: Made block loss function configurable (25aeac5a1)
    • AdaScale: Made scaled (sum-over-seq-dim) loss the default block loss function (c1a2c4444)

• Documentation

  • Refactored GenAI model results, added AIMET Torch companion recipe page, and added Qwen3 0.6B/1.7B recipe rows (b9d1a1034, b5452d769)