# 0.11.0

## Release Information

- OS & Framework support - Tested with the following configurations:

> 
> 
> - Non-Gen AI:
> 
> 
> 
> > 
> > 
> > - Frameworks: ONNX, TFLite, Pytorch
> >         - Host Platforms: Linux-x86\_64 (Ubuntu 22.04, Ubuntu 24.04), Linux-aarch64 (Ubuntu 24.04, Python 3.12), Windows-x86\_64 (10+), Windows-arm64 (10+)
> >         - Target Platforms: Linux-x86\_64 (Ubuntu 22.04, Ubuntu 24.04), Linux-aarch64 (Ubuntu 24.04), Windows-arm64 (10+), Android-arm64, QNX
>     - Gen AI:
> 
> 
> 
> > 
> > 
> > - Frameworks: ONNX, GGUF
> >         - Host Platforms: Linux-x86\_64 (Ubuntu 22.04, Ubuntu 24.04)
> >         - Target Platforms: Android-arm64, Linux-aarch64
- QAIRT SDK - Tested with version (2.46.0, 2.47.0, 2.48.0, 2.49.0 (Default))

Note

For QAIRT SDK, other supported SDK versions can be listed using `qairt-vm fetch --list`.

## Highlights

- Default QAIRT SDK updated to 2.49.0.
- Minimum supported QAIRT SDK updated to 2.46.0.
- New features in QAIRT Dev Python APIs (compatible with QAIRT SDK 2.49.0+, unless noted otherwise)

> 
> 
> - Compiler
> 
> 
> 
> > 
> > 
> > - `qairt.compile` can now produce a model that can be saved as a DLC by setting
> > `CompileConfig.module_format="dlc"`, letting a single `compile()`/`CompiledModel`
> > workflow emit either a raw context binary or a DLC container without a separate
> > conversion step.
> >         - `CompileConfig.enable_backend_trace` generates backend operation mapping artifacts
> > alongside the compiled output, embedding the mapping directly into the DLC when
> > `module_format="dlc"` and the installed SDK supports it.
>     - Gen AI Builder
> 
> 
> 
> > 
> > 
> > - `sliding_context_length` is now propagated through the Gen AI Builder for
> > sliding-window-attention (SWA) models, rewriting the SWA KV cache and attention mask via
> > `qairt.optimizer.onnx.change_sliding_context_length`. Defaults to
> > `ceil(sliding_window / 256) * 256` when unset, and raises `RuntimeError` if set on a
> > non-SWA model.
> >         - The residual-Add heuristic used to locate layer split boundaries when splitting a model
> > is now configurable through Gen AI Builder via `split.layer_boundary_name_regex`.
> >         - Embedding-related APIs are consolidated behind a single `GenAIBuilderHTP.embeddings`
> > facade (`(source, mode)` pair, typed via the new `EmbeddingMode` enum), replacing the
> > previous separate `.embedding`/`attach_model_for_embedding`/`.prepare_embedding_lut`
> > entry points with one consistent interface.
> >         - `GenAIBuilderFactory` now supports Qwen3.5 and Qwen3.5 MoE checkpoints via dedicated
> > `Qwen3_5BuilderHTP`/`Qwen3_5MoeBuilderHTP` builders, with `precision_compensation`
> > enabled by default.
> >         - `ImageT2TExecutor` now accepts a standard image file (JPEG, PNG, WEBP) in addition to
> > a pre-processed `.raw` file, automatically preprocessing it through the vision
> > encoder’s processor.
>     - Pipeline
> 
> 
> 
> > 
> > 
> > - The `quantization_opt` stage now supports evaluation (`can_evaluate = True`), so
> > `pipe.evaluate()` can be called immediately after pre-quantization optimization runs,
> > without first advancing to the `quantization` stage.
> >         - [Preview] The pipeline now supports **speculative decoding** via the new Eaglet
> > draft-model architecture, validated with Qwen3-4B (see [Eaglet](https://docs.qualcomm.com/doc/80-87189-2/topic/eaglet.html#pipeline-eaglet)).
> >         - Pipeline recipes can register custom stages from a `.py` file via the top-level
> > `custom_stage` field and the new `register_stage_plugin()` — no separate Python
> > registration call needed.
> >         - Pipeline recipes now support repeating the same stage type via list-format `stages:`
> > YAML (e.g. `model_loader` → re-authoring → `model_loader_1`), auto-wired and
> > disambiguated by instance key.
> >         - `QuantizationStage` supports an opt-in `TestVectorConfig` that records matched
> > floating-point/quantized tensor pairs (`fp_<N>.pkl`/`qt_<N>.pkl`) for downstream
> > QuantSim validation (see [Test Vector Generation](https://docs.qualcomm.com/doc/80-87189-2/topic/pipeline_configuration.html#pipeline-test-vectors)).

## Resolved Issues

> 
> 
> - Fixed Gen AI Builder SCL (sliding context length) variant generation raising on SWA models
> whose `sliding_window` was already &gt;= the minimum context length (e.g. Phi-3.5-mini with
> `sliding_window=262144` and `context_length=4096`). `_scl_variants` now short-circuits to
> `[None]` in this case instead of attempting an invalid variant sweep.
> - Fixed `LLMPipeline.generate()` returning a plain string for stages upstream of
> `genai_builder` (e.g. `model_loader`, `quantization`) instead of the documented
> `TextGenerationResult`, breaking callers that expect `result.print()` / `result.generated_text`
> regardless of which stage the pipeline was advanced to.
> - Fixed `transposed_key_cache` and other QC-specific flags not round-tripping through
> `to_dict()` for reauthored Llama/Phi3/Qwen3/Qwen3.5 configs, by giving `QcConfigMixin` its
> own `to_dict()` override and reordering the config base classes so the mixin takes precedence.

## Deprecation Notices

> 
> 
> - The `OnnxModel` interface (`OnnxModel.load()`, `OnnxModel.split()`,
> `OnnxModel.mha2sha_v2()`), deprecated in 0.10.0, has now been removed. Use `GraphContext`
> together with the simple optimizer APIs (`convert_mha_to_sha`, `split_llm`, `adapt_moe`)
> instead.

To be removed in 0.12.0:

> 
> 
> - `GenAIConfig.allow_async_init` and `GenAIConfig.enable_graph_switching` are deprecated. Set
> these through `EngineConfig.htp.allow_async_init` and `EngineConfig.htp.enable_graph_switching`
> and pass the `EngineConfig` to `get_executor()` / `T2TExecutor` instead.
> - `GenAIExecutor` (`qairt.gen_ai_api.executors.gen_ai_executor`) is deprecated and has been
> renamed to `GenAIExecutable` (`qairt.gen_ai_api.executors.gen_ai_executable`). Update imports
> to the new module.

## Known Issues

> 
> 
> - Gen AI Builder is temporarily unsupported on Windows-arm64. Use Linux-x86\_64 as the host platform
> for Gen AI workflows.
> - Windows on ARM64 is not functional. This is a known issue (AISW-188015).
> - Models with batch size greater than 1
> 
> 
> 
> > 
> > 
> > - A discrepancy exists between the qnn-net-run CLI and the Python API regarding data preparation
> > for batch sizes greater than 1. While the CLI automatically groups individual tensor paths from
> > an input list into a batch, the Python API requires inputs to be pre-batched (concatenated into
> > a single raw file) prior to execution

Last Published: Oct 05, 2026

[Previous Topic
Resolved Issues](https://docs.qualcomm.com/bundle/publicresource/80-87189-2/topics/0_10_1.md) [Next Topic
0.12.0](https://docs.qualcomm.com/bundle/publicresource/80-87189-2/topics/0_12_0.md)