Qwen Support

[chrishayuk]

As you know one of the goals of this project is to get large models such as qwen working on consumer hardware.
I've done some work on model architecture for qwen but not tested it.. it has loads of small models so would really appreciate if anyone wants to take this on

[mvkorobkov]

Field report — we've been running the current Qwen arch (crates/larql-models/src/architectures/qwen.rs) in production for the LARQL vindexes that back our project, and it handles every Qwen size/variant we've thrown at it. Sharing concrete evidence in case it helps close the loop.

Successfully extracted to inference-level vindex via larql extract:

Variant	model_type	hidden / layers / kv	Vindex size	Notes
Qwen2 1.5B	qwen2	1536 / 28 / 2	small	b2_qwen2_1p5b.vindex
Qwen2.5 Coder 7B	qwen2	3584 / 28 / 4	(browse only — separate)	qwen25_coder_7b.vindex
Qwen3 1.7B (Bonsai)	qwen3	2048 / 28 / 8	OK	bonsai-1.7b unpacked
Qwen3 4B	qwen3	2560 / 36 / 8	OK	qwen3-4b-vindex-infer
Qwen3 14B	qwen3	5120 / 40 / 8	OK	qwen3-14b-vindex-infer
Qwen3.5 27B	qwen35	5120 / 64 / 8	OK	qwen35-27b-vindex
Qwen3.6 27B	qwen3	5120 / 40 / 8	9.1 GB	qwen36_27b.vindex

The arch dispatch (detect/mod.rs matching qwen2/qwen3/qwen3_moe/qwen35) and the per-family key generation (attention bias for Qwen2, QK norms for Qwen3, MoE expert routing for Qwen3-MoE) all work as written. Existing test_architectures.rs covers Qwen2 and Qwen3-MoE detection + key shapes.

What I think is still genuinely untested / open work for this issue:

1. qwen35 model_type — I don't see it in qwen.rs::family() or the parser tests; the dispatch in detect/mod.rs falls through to QwenArch::from_config based on the prefix match, but no explicit test pins it. Our Qwen3.5-27B vindex was extracted, so the runtime path is OK; a parser-level test would lock it in.
2. Qwen3 dense variants (4B / 14B / 27B / 27B-thinking) — covered behaviourally by extraction tests but not by config-detection unit tests.
3. Qwen3-Next 30B-A3B and Qwen-VL variants — outside what we run; flagging that someone else's report would be useful here.

If a 2-fixture PR adding qwen35 and qwen3 dense config-detection cases to test_architectures.rs would help close this issue, I'm happy to open one — let me know if that matches what you had in mind, or if you want broader coverage (CLI extraction integration tests against real configs).