llm_runner: add Engine and Session interfaces

[mergennachin]

The LLM serving path needs a stable contract between generic serving code and model-specific execution code. TextLLM, Qwen, Gemma, CUDA, and future backends all differ in how they own weights and mutable state, but the server should not know those details or grow a new Python binding for every model.

This introduces the minimal runner-level split needed for that contract. LLMEngine represents the loaded physical model and its serving capacity; LLMSession represents one logical conversation state and exposes reset/prefill/decode-style operations. That shape lets a worker drive different model implementations through one interface while keeping KV/recurrent/cache ownership inside C++.

This commit is only the interface and build export. It deliberately does not add a concrete adapter or change existing runner behavior, so model migrations and serving can be reviewed as downstream uses of the contract rather than hidden side effects.

#20001

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/20243

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

❗ 1 Active SEVs

There are 1 currently active SEVs. If your PR is affected, please view them below:

• [ROCm] MI350 CI jobs will have longer queue times due to CI migration

⏳ No Failures, 4 Pending

As of commit e94d710 with merge base d7ca5db ():
💚 Looks good so far! There are no failures yet. 💚

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[github-actions]

This PR needs a release notes: label

If your change should be included in the release notes (i.e. would users of this library care about this change?), please use a label starting with release notes:. This helps us keep track and include your important work in the next release notes.

To add a label, you can comment to pytorchbot, for example
@pytorchbot label "release notes: none"

For more information, see
https://github.com/pytorch/pytorch/wiki/PyTorch-AutoLabel-Bot#why-categorize-for-release-notes-and-how-does-it-work.

[Copilot]

Pull request overview

This PR introduces a new, model-agnostic C++ contract for the LLM serving path by defining LLMEngine (loaded model/resources + serving capacity) and LLMSession (per-conversation mutable state with reset/prefill/decode operations). This establishes an interface boundary so serving/worker code can depend on a stable API while backends keep ownership of KV/cache/state internally.

Changes:

• Adds new LLMEngine / LLMSession interfaces plus supporting structs (SamplingConfig, DecodeResult, LLMServingCapacity).
• Exports the new header via the Bazel runner library target.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 6 comments.

File	Description
extension/llm/runner/llm_session.h	Adds new engine/session interfaces and associated data types for a stable serving contract.
extension/llm/runner/targets.bzl	Exports the new header in the runner library’s public headers.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[mergennachin]

@claude Review this PR

[claude]

Claude finished @mergennachin's task in 1m 16s —— View job

---

PR Review: llm_runner: add Engine and Session interfaces

• Read changed files (llm_session.h, targets.bzl)
• Analyze interface design & contract consistency
• Check against Copilot review comments
• Post review feedback

---

Reviewed the two changed files. This is a clean, well-scoped interface-only PR — no concrete adapters, no behavior change. The documentation is unusually thorough for a header, and the Engine/Session split (immutable model resources vs. per-conversation mutable state) is a sound contract for the serving path. Overall looks good to merge; comments below are minor.

On the Copilot comments — all already addressed ✅

The Copilot review was generated at an earlier snapshot. The current head of the branch already resolves every one of its six comments:

• SamplingConfig, DecodeResult, LLMServingCapacity are all marked ET_EXPERIMENTAL (lines 31, 51, 62).
• The cooperative-stop() terminal-step contract is now fully specified in three consistent places: DecodeResult (lines 47–50: "token_id must be 0, text_piece must be empty, position() must not advance"), decode_one() (lines 108–110), and stop() (lines 129–135).
• The decode_one() doc no longer contradicts stop() — it explicitly states the short-circuit path.

No action needed there.

Minor observations

1. DecodeResult has no default member initializers (lines 52–55). The other three structs (SamplingConfig, LLMServingCapacity) initialize every field, but DecodeResult's four members are left uninitialized. An adapter that forgets a field yields a garbage is_eos/is_terminal — a nasty bug since is_terminal drives the decode loop. Recommend defaulting for consistency and safety:

   struct ET_EXPERIMENTAL DecodeResult {
     uint64_t token_id = 0;
     std::string text_piece;
     bool is_eos = false;
     bool is_terminal = false;
   };

2. Namespace style differs from the sibling header. This file uses C++17 nested namespace executorch::extension::llm { (line 25), while irunner.h in the same directory uses the older namespace executorch { namespace extension { namespace llm { form. Both compile fine; the new form is cleaner. Just flagging the inconsistency within one directory in case the codebase has a convention.

3. Relationship to GenerationConfig/IRunner is worth a one-line note. SamplingConfig overlaps conceptually with GenerationConfig::temperature from irunner.h, but operates at a different altitude (per-decode-step vs. per-generation). The header comment (lines 9–12) explains the adapter relationship well; a sentence clarifying that this Session contract sits below the existing IRunner/GenerationConfig layer (rather than replacing it) would help future readers who land on both files.

4. metadata() returning const std::unordered_map<std::string, int64_t>& ties the engine contract to a specific container type. That's consistent with how runner metadata is passed elsewhere in this codebase, so it's reasonable — just noting it constrains adapters to store metadata in exactly that shape.

None of these are blocking. The only one I'd actively recommend acting on is #1 (default-initialize DecodeResult) since uninitialized is_terminal is a real footgun. Fix this →

targets.bzl: the single-line addition correctly exports llm_session.h from the runner_lib target alongside the other public headers. ✓
· branch llm-pr-c-llm-engine-session-interface

[Copilot]

Pull request overview

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

[Gasoonjia]

why we have to have seek function in the top-level llm_session? For speculative decoding?