[QNN] Improve GPU StaticLLM pass

[qti-mattsinc]

Describe your changes

• Update search.max_length in genai_config.json to inform onnxruntime-genai when to stop decode, which is required for the QNN GPU exports with fixed max seq len.
• Remove the "update_genai_config" option.
• Handle new ModelBuilder behavior with symbolic "kv_cache_dim".
• Add "inherit_session_options" to pipeline models in the genai config.

Checklist before requesting a review

• Add unit tests for this change.
• Make sure all tests can pass.
• Update documents if necessary.
• Lint and apply fixes to your code by running lintrunner -a
• Is this a user-facing change? If yes, give a description of this change to be included in the release notes.

[azure-pipelines]

Azure Pipelines:
There may be pipelines that require an authorized user to comment /azp run to run.

[azure-pipelines]

Azure Pipelines:
There may be pipelines that require an authorized user to comment /azp run to run.

[Copilot]

Pull request overview

This PR updates the QNN GPU StaticLLM export flow so generated genai_config.json better reflects fixed-shape decoding requirements (fixed max sequence length), aligns with newer ModelBuilder symbolic dim behavior (kv_cache_dim), and standardizes session options inheritance across pipeline entries.

Changes:

• Update QNN GPU genai pipeline generation to set inherit_session_options and to derive model.context_length / search.max_length from the fixed KV-cache input shape.
• Remove the update_genai_config option and always perform genai_config updates in the relevant QNN GPU paths.
• Add QNN GPU-focused unit tests that validate prefill/decode pipeline structure and enforce static I/O shapes.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 3 comments.

File	Description
test/passes/onnx/test_static_llm.py	Adds QNN GPU coverage validating genai pipeline structure and ensuring exported model I/O shapes are fully static.
olive/passes/onnx/static_llm.py	Handles new ModelBuilder symbolic kv_cache_dim by mapping it using head_size from genai_config.json; removes update_genai_config gating.
olive/passes/onnx/context_binary.py	Removes update_genai_config option and updates QNN GPU context-binary path to update genai pipeline entries unconditionally.
olive/passes/onnx/common.py	Adds inherit_session_options to pipeline entries and updates genai max sequence length fields based on the fixed KV-cache input shape.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Copilot]

Pull request overview

Copilot reviewed 4 out of 4 changed files in this pull request and generated no new comments.

Suppressed comments (1)

olive/passes/onnx/context_binary.py:312

• update_llm_pipeline_genai_config_gpu_ctxbin is now called unconditionally for QNN GPU, but _generate_context_binary accepts model=None and ModelHandlers can have model_attributes=None (or missing additional_files). In those cases this call will raise before any genai_config update can be attempted. Guard the call so non-GenAI/non-LLM QNN GPU flows still work.

                update_llm_pipeline_genai_config_gpu_ctxbin(model, ctxbin_path)

[qti-mattsinc]

Jambay Kinley (@jambayk), PTAL. This is needed for QNN. Thanks!