feat(qwen36): name a checkpoint by its geometry, and a converter that refuses what it cannot place

[kreuzzelg]

Fixes the banner defect confirmed in #1045 (comment 5458211750), and lands the pieces that make the Qwen3.8-2.4T-A95B shape a verified one rather than a guessed one — without a single weight of it.

The defect

Qwen/Qwen3.8-2.4T-A95B declares the same model_type as Qwen3.6-35B-A3B (qwen3_5_moe_text) and resolves to the qwen36 engine today, where the banner announces it as "Qwen3.6-35B-A3B · 35B". A loader that misreports what it loaded costs someone a day.

Registry: display by geometry

A model_type resolves to exactly one family (_BY_TYPE), so a second descriptor is not available. Instead the descriptor carries display_variants: a DisplayVariant pairs a geometry (config keys that must match exactly) with a name, and display_for(resolved) picks the first match.

geometry	banner
40 layers / 256 experts / hidden 2048	Qwen3.6-35B-A3B · 35B MoE · 21 GB on disk
92 / 512 / 8192	Qwen3.8-2.4T-A95B · 2.4T MoE
anything else (a tiny fixture, a third sibling)	qwen3_5_moe_text · 8L x 8E MoE — its own model_type and measured numbers, never a sibling's parameter count

The registry refuses a variant that matches everything, and a family whose static display_name is not among its variants (the READMEs are held to that name by the existing contract test). coli's banner reads the family config, so the VL-wrapped 35B repo (text_config) and the flat 2.4T config both display right. public_metadata carries the variants.

Converter: every tensor name is placed, or it stops

While checking the 2.4T index I found the class of defect you described for GLM-5.3, live in this converter. It selected <prefix>layers.<i>. plus four globals and let the rest fall through unmentioned. That skipped the mtp.* head and the visual.* tower by accident, and would have copied an unfamiliar tensor inside a layer as f16 without a word.

tools/qwen36_tensor_kinds.py now classifies every name before the first shard is read: layer tensors by an exact suffix list, mtp.* / visual.* skipped on purpose with a count and a reason (skipping 19 mtp.* tensor(s): multi-token-prediction head …, config mtp_num_hidden_layers=1), anything else refused with the names. Torch-free, so tests/test_qwen36_tensor_kinds.py runs wherever the registry tests run; it pins the contract to the safetensors indexes of both real checkpoints (1609 and 1045 tensors) and to the per-expert layout transformers saves.

One correction on the way: a comment claimed the real 35B ships per-expert tensors. Both real checkpoints ship the fused layout (experts.gate_up_proj / down_proj, one tensor per layer) and both carry the mtp head; only the tiny fixture is per-expert. So the "five-minute index check" I proposed in #1045 was aimed at the wrong regex — expert_inventory describes the container, not the checkpoint — and the real thing to verify was the fused split and the deliberate skip. Both are now under test.

Fixture: the 2.4T shape at toy widths, in the real layout

make_qwen36_tiny.py --geometry qwen38-2p4t: 92 layers / interval 4 / 512 experts top-10 / 16:1 attention heads / 8:1 DeltaNet heads, and the saved shard rewritten into what the checkpoints ship — fused experts plus a one-layer mtp.* head with mtp_num_hidden_layers: 1. The converter has to split the one and skip the other; the engine has to match the transformers reference token for token. The default preset reproduces the shipped fixture's reference exactly, so the existing gate is unchanged.

CI (qwen36-tiny-check) gains a step: build the 2.4T fixture, assert the converter said it skipped 19 mtp tensors and placed 1587 in 92 layers, assert no mtp.* survives in the container, token-exact at cap 1 / 2 / 512; the sanitizer step now runs both fixtures.

Verified

• old fixture: reference identical, converted shard bit-identical (cmp), 16/16 at cap 1/2/8
• 2.4T geometry: 19 mtp tensors skipped and said so, 1587 placed in 92 layers, 16/16 at cap 1/2/512, 0.35 s per run; generator 17 s, converter 9 s
• ASan/UBSan (nm confirms 31 __asan symbols — the lesson of fix(pilot): the prefetch worker outlives the Model it dereferences #1277): no diagnostic on either fixture under PILOT=1 WIDE=2
• registry planner on the real 2.4T config: KV 1.44 GiB @ 8k, 46.0 GiB @ 256k, DeltaNet state 0.55 GiB context-free — the numbers quoted in [Feature]: Can you please consider adding qwen 3.8 #1045. The engine's KV cache is F32 (ensure_kv → falloc), so they stand; the F16 caveat I raised there is closed
• python suite: 743 tests, OK (skipped=35); the new CI step run locally word for word
• --seed 3 for the 2.4T rung: the default seed collapses this geometry's reference to one repeated token, which a shape error could still reproduce; seed 3 gives twelve distinct tokens over sixteen

Not in this PR, and said in docs/qwen36.md

Running the 2.4T. The warmstart keeps every expert in RAM by design, ~1.4 TB of int4 for this checkpoint; that needs the disk-streaming design, not this one. What this PR guarantees is that the shape is right before anyone spends the bandwidth.

Thanks to claude code!