Skip to content

Qwen export harness (StableHLO + parameters) for the IREE path — none exists; FunctionGemma is the only tool-calling model with a compiled leg #411

Description

@michalharakal

Summary

Export harnesses exist for SmolLM2 and FunctionGemma only. Qwen3-0.6B is the model that actually reaches the accuracy bar on our German TV tool-call golden sets (7/8 golden-8, 36/48 golden-40, greedy, thinking off, 0.53.0 AgentLoop), while FunctionGemma-270M sits at 1–2/8 regardless of catalog wording. On the device, eager Qwen3 on ART is 3,210 s per token (no kernel pack for armeabi-v7a), so without a compiled leg Qwen cannot ship at all.

Ask

A QwenExportHarness mirroring FunctionGemmaExportHarness (redecode + prefill + with-past graphs, external parameters to safetensors → .irpa), with the same contract as the KV proposal, for Qwen2.5/Qwen3 on the DSL path (#340/#352 already put Qwen on the engine loader). A conformance row in skainet-iree-conformance next to functiongemma-270m would make the parity check reproducible.

Measured device context: FunctionGemma-270M FP32 redecode steps in 10.4 s (CPU) / 3.2 s (Vulkan) at SEQ 64 on the MagentaTV One; Qwen3-0.6B is ≈ 2.2× the parameters, so ≈ 7 s per step on Vulkan at SEQ 64 is the expectation to verify.

Context. Measured on 2026-09-03 while bringing FunctionGemma-270M up as an NLU cartridge on a MagentaTV One (Amlogic S905X5L, Android 14, armeabi-v7a-only 32-bit process, 4× ARMv8 @ 2.0 GHz, Mali GPU) with SKaiNET 0.53.0 + SKaiNET-transformers 0.53.0 from Maven Central. Harness and raw result files: tvv/nlu-llm-harness (local; can share on request).

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions