Hacktoberfest 2026：维护者为十月标记出来的 issue，仍然开放、适合新手。 浏览 Hacktoberfest issue

Qwen3 0.6B fails to load after flutter_edge_ai 2.3 / litertlm 1.13 bump

未关闭
  • 哪里不对？
#286 2 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

维护者通常 1 天内回复

还没有人认领这个 Issue。

评估

难度
4/5
预计耗时
3-5 天
新手友好度
18/100
Issue 类型
缺陷
描述清晰度
基本清楚
活跃度
活跃
技术栈
android, dart
领域
machine-learning, mobile

调研方向

Start with the failing path in the LiteRT-LM client at src/ffi/litert_lm_client.dart (initialize, around line 956) and LiteRtLmEngine.createModel in litert_lm_engine.dart (line 167), then reproduce with integration_test/llm_stability_e2e_test.dart, which is listed in all_tests.dart. The open question is whether Qwen3-0.6B needs re-conversion for LiteRT-LM 0.18 or a different load setting, and the issue still lacks those probe results. Done means either the model loads on the bumped packages or AIModelId.qwen25_05b is gated behind a load check with a fallback to Gemma.

由索引模型根据 Issue 内容生成。

描述

bug P1-High

Summary

After bumping to flutter_edge_ai 2.3.2 / flutter_edge_ai_litertlm 1.13.2 (LiteRT-LM 0.18.0, see #257), Qwen3 0.6B (Qwen3-0.6B.litertlm) no longer loads on device. The same file loads and answers on our current packages. Gemma 4 E2B is unaffected.

Error

BackendInitException: all FFI backends failed. Last attempted cpu. Attempts:
  gpu: Exception: Failed to create engine. Model may be invalid: .../app_flutter/Qwen3-0.6B.litertlm;
  cpu: Exception: Failed to create engine. Model may be invalid: .../app_flutter/Qwen3-0.6B.litertlm
#0 LiteRtLmFfiClient.initialize (flutter_edge_ai_litertlm/src/ffi/litert_lm_client.dart:956)
#1 LiteRtLmEngine.createModel (litert_lm_engine.dart:167)

Logged just before: [LiteRtLmFfi] litert_lm_engine_create took 91ms, then [LiteRtLmEngine] cpu backend failed, no candidates are left.

Evidence it is the package, not our code or the file

Result
Old: flutter_gemma 1.8.3 + _litertlm 1.6.3 (native 0.16.0), same phone, same file loads, replies (<think> ... )
New: flutter_edge_ai 2.3.2 + _litertlm 1.13.2 (native 0.18.0-e) Failed to create engine. Model may be invalid on GPU and CPU
File on device 614,236,160 bytes, identical size to the pushed fixture
Gemma 4 E2B (gemma-4-E2B-it.litertlm) on the new packages all cases pass

Device: vivo 1933, Android 11 (API 30), arm64. The native bundle downloaded and verified its checksum (native-v0.18.0-e), so it is not a corrupt download.

How it was found

New device test integration_test/llm_stability_e2e_test.dart (runs every model fixture). Before it, every baseline and post-bump device test ran on Gemma, so the Qwen regression was invisible.

Impact

Qwen3 0.6B is in our local catalog (AIModelId.qwen25_05b). On the bumped packages it cannot be activated, so anyone who picked it would get no local model.

Open

  • Probe in progress: does a different maxTokens (256 / 1024 / 4096) or an explicit CPU backend load it? Results will be added here.
  • Whether the file needs re-converting for LiteRT-LM 0.18 (the upstream "Model may be invalid" text suggests a format check). I could not confirm without upstream.

Proposed handling

  1. Report upstream (flutter_edge_ai) with the table above.
  2. Until fixed, do not ship the bump with Qwen3 0.6B in the catalog, or gate it behind a load check with a clear message and fall back to Gemma.
  3. Keep llm_stability_e2e_test in all_tests.dart so this is caught on every device run.

Related: #257 (the bump), #88 (device tests).

主要语言
Dart
星标
2
派生
2
平均合并
1 天 35 分钟
30 天内合并 PR
36

环境准备

  • 没有 Dockerfile 或 Docker Compose 文件
  • 没有 Pull Request 模板
  • 阅读贡献指南

从这里开始

  1. 先读完整个 Issue，再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库，在一个分支上完成修改。
  4. 提交 Pull Request，并在描述里引用这个 Issue 编号。

basictech01/uniun 的其他 Issue

查看 basictech01/uniun 的全部 Issue

相似的 Issue

更多 Dart Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。