feat: bump litertlm-android to 0.11.0 and add MTP toggle

[nmrenyi]

Summary

• Bumps com.google.ai.edge.litertlm:litertlm-android from 0.10.2 → 0.11.0.
• Adds a new Gradle property useMtpForLlm (default false) that opts the runtime into LiteRT-LM's Multi-token Prediction / speculative decoding (ExperimentalFlags.enableSpeculativeDecoding).
• Documents the smoke-test result for both changes in the README and inline near the new Gradle flag.

Supersedes #54 (same dependency bump, with extra wiring and the smoke-test write-up).

Smoke-test results (OPPO Snapdragon 8 Elite, 2026-05-14)

LiteRT-LM 0.11.0 was tested locally with useGpuForLlm=true. Same model file (gemma-4-E4B-it.litertlm), same RAG bundle.

Config	TTFT	Decode rate
0.10.2 CPU (README baseline)	11.7 s	13.8 tok/s
0.11.0 GPU, no MTP	~1–2 s	~12–14 tok/s
0.11.0 GPU, MTP on	~1–2 s	~10–12 tok/s (slower)

• 0.11.0 on GPU gives a ~7× TTFT win — the main payoff of this PR.
• MTP / speculative decoding was honoured at the native layer (enable_speculative_decoding: true, tf_lite_mtp_drafter subgraph loaded — confirmed in logcat) but produced a ~10–20% decode slowdown, 3/3 samples. Most likely cause: low drafter acceptance rate on long retrieved-context prompts + constrained medical prose. Off by default; re-test before re-enabling.

Risk notes

• API surface used in RagPipeline.kt (Engine, EngineConfig, ConversationConfig, Backend, SamplerConfig, Message, Contents) is source-compatible between 0.10.2 and 0.11.0 — verified via the v0.10.2…v0.11.0 commit log; the Kotlin API additions (filterChannelContentFromKvCache, file capability inspection, prefer_activation_type) are all additive.
• Per CLAUDE.md's eval-gated deployment rule, this is a smoke test, not a full eval. Full clinical eval still pending before any decision to flip GPU on by default in production.

Test plan

• Build APK (CPU default): flutter build apk — verify [BACKEND] *** LLM running on CPU *** in logcat.
• Build APK (GPU): flutter build apk -PuseGpuForLlm=true — verify [BACKEND] *** LLM running on GPU ***.
• Build APK (GPU + MTP): flutter build apk -PuseGpuForLlm=true -PuseMtpForLlm=true — verify [MTP] speculative decoding ENABLED and tf_lite_mtp_drafter lines in logcat.
• Run a slice of kenya_vignettes through the eval harness against 0.11.0 before considering for production rollout.

🤖 Generated with Claude Code

[Copilot]

Pull request overview

Bumps the on-device LLM runtime litertlm-android from 0.10.2 to 0.11.0 and exposes a new compile-time Gradle flag useMtpForLlm that toggles LiteRT-LM's experimental Multi-token Prediction / speculative decoding via ExperimentalFlags.enableSpeculativeDecoding. README and inline Gradle comments document the smoke-test results (GPU TTFT win, MTP regression on long RAG prompts).

Changes:

• Upgrade com.google.ai.edge.litertlm:litertlm-android to 0.11.0.
• Add useMtpForLlm Gradle property → BuildConfig.USE_MTP_FOR_LLM → applied via @OptIn(ExperimentalApi::class) in RagPipeline.kt.
• Update README and Gradle comments with smoke-test results and re-test guidance.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated no comments.

File	Description
app/android/app/build.gradle.kts	Bumps litertlm-android dependency and wires new useMtpForLlm Gradle property into BuildConfig.USE_MTP_FOR_LLM.
app/android/app/src/main/kotlin/com/example/app/RagPipeline.kt	Reads BuildConfig.USE_MTP_FOR_LLM and sets ExperimentalFlags.enableSpeculativeDecoding before engine init; opts into experimental API.
README.md	Updates benchmark notes — removes the "GPU blocked" sentence and documents 0.11.0 GPU TTFT win and MTP smoke-test regression.

---

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