glm_moe_dsa (GLM-5.2) emits an invalid Squeeze; ORT cannot load the model

[justinchuby]

Summary

glm_moe_dsa (GLM-5.2, added in #555) builds an ONNX graph that ONNX Runtime refuses to load: the MLA attention emits a Squeeze whose target axis is not size 1.

Node (model/layers.0/self_attn/Squeeze_node_46) Op (Squeeze)
  [ShapeInferenceError] Dimension of input 1 must be 1 instead of 4

This also makes test_onnx_checker_passes[glm_moe_dsa] fail with ValidationError: Field 'type' of 'value_info' is required but missing, and leaves logits without a shape after inference.

Reproducing

pytest tests/synthetic_parity_test.py -k glm_moe_dsa
pytest tests/build_graph_test.py -k glm_moe_dsa

This needs #551 (or just its tests/_test_configs.py entry) on the branch. Before that, glm_moe_dsa had no tiny config at all, so it fell back to an auto-generated one with qk_nope_head_dim/qk_rope_head_dim unset and died earlier with:

TypeError: unsupported operand type(s) for +: 'NoneType' and 'NoneType'
  src/mobius/components/_deepseek_mla.py:57

I fixed that config gap in #551 (tiny config mirroring the model's own fixture in src/mobius/models/glm_moe_dsa_test.py), which takes glm_moe_dsa from 5 failing tests to 3. The invalid Squeeze is what remains, and it is a defect in the model rather than the config — the config fix only let the graph build far enough to expose it.

Why the co-located tests don't catch it

src/mobius/models/glm_moe_dsa_test.py passes because it asserts on graph structure and never loads the model in ORT. The failure only appears once something actually runs the graph.

Notes

• Not caused by symbolic shape inference: it reproduces with Normalize anonymous dims when propagating shapes into Scan/Loop bodies justinchuby/onnx-shape-inference#119 applied.
• L4/L5 cannot cover this — GLM-5.2 is 753B parameters across 282 safetensors shards — so L1/L3 is the only guard, which is why the missing tiny config mattered.

Fixing this needs GLM-5.2 DSA/MLA knowledge, so I've left it rather than guessing at the correct axis.