[inureyes]

Summary

Adds model_type: "cohere_compass" (CohereLabs North-Micro-Vision-Instruct, 2.5B): a Qwen3-VL deepstack vision tower in front of a Command-style parallel text decoder. The vision tower, image processor, <|IMAGE_PAD|> expansion, MRoPE position-id rule and weight-prefix remap are the existing Qwen3-VL ones reused; the decoder and its rotation are the new code.

The rotation is not Qwen3-VL's interleaved MRoPE

The issue specified InterleavedMRoPE on the sliding layers. That is wrong, and it is the one thing worth reading the diff for. The checkpoint carries mrope_interleaved: true, but upstream lists that key under ignore_keys_at_rope_validation and never reads it. CohereCompassRotaryEmbedding pre-permutes inv_freq over the section[0] + section[1] head block into evens-then-odds, then assigns axes in contiguous [H, W, T] sections. Over the published [24, 20, 20] that gives H 24 channels; Qwen3-VL's step-3 interleave over the same list gives T 24 channels, over unpermuted frequencies. Measured against the transformers class itself at (t, h, w) = (7, 11, 13): this implementation is within 2.2e-7 in cos, the specified interleave is off by 1.96. For the same reason there is no fast_rope text-only shortcut here, since a permuted inv_freq does not collapse to natural-order 1-D RoPE when the three axes agree.

What changed

• src/models/cohere_compass{,_config,_layers,_rope}.rs: the text decoder. Parallel block (one input LayerNorm feeds both attention and the SwiGLU MLP, both land on the residual, no post_attention_layernorm), tied head times logit_scale 0.25, per-layer-type RoPE where a JSON null entry means no positional encoding at all (the seven full-attention layers are NoPE), and the sliding/full split enforced by two prefill masks plus the SDPA window over a dense KVCache, the same contract cohere2.rs and olmo3.rs use.
• src/vision/cohere_compass.rs: the VLM wrapper over Qwen3VLVisionEncoder, DeepStack state, visual position mask and MRoPE state, plus the QwenVlRuntime impl so CLI and server image paths route through the shared Qwen-VL machinery.
• src/loading/vlm_cohere_compass.rs: loader. Both published weight layouts land on one key space through the existing remap_qwen3_vl_weights. Resize bounds come from the nested image_processor block of processor_config.json, which is what HF's AutoProcessor reads; the standalone preprocessor_config.json on this checkpoint disagrees and is only the fallback.
• Detection, ModelType, LoadedModel, registry, metadata and docs rows, plus the bilingual technical report.

Test plan

Oracle is transformers 5.18.0.dev0 reading the same weights: the MLX 4-bit checkpoint dequantized offline into one float32 file in the CohereLabs key layout, so both sides load byte-identical weights and any divergence is the forward pass. Every run was done on both the 4-bit and the float32 checkpoint, which covers both published weight-prefix layouts.

• cargo test --profile test-fast --features cuda --lib cohere_compass on GB10: test result: ok. 23 passed; 0 failed.
• Rotary table against CohereCompassRotaryEmbedding: max|cos diff| 2.23e-07, max|sin diff| 3.80e-07; the interleave the issue specified is off by 1.96.
• MRoPE position ids bit-identical to get_rope_index, one image (rope_delta = -182) and two images (rope_delta = -224).
• Text-only token-exact: server /v1/chat/completions with logprobs: true gives 32/32 identical tokens against the oracle on both checkpoints; the CLI produces the identical string on three prompts, one of them a 79-token prompt.
• Server image path: prompt_tokens 210 for one image and 245 image tokens for two, both exactly the oracle's counts, completion_tokens non-zero, EOS reached. The server text equals the CLI text on every prompt. This is the family's known zero-decoder-layer trap, checked rather than assumed.
• mlxcel arch lists the family, mlxcel list shows the checkpoint, mlxcel inspect reports sliding-window: 21 layer(s) capped at 4096 tokens, 7 global.
• cargo clippy --workspace --all-targets -- -D warnings and cargo fmt --all -- --check: green in CI.
• Image greedy ids do not match the oracle. Everything up to the decoder input matches, and mlxcel picks the oracle's rank-2 token, but the flip is real: single image diverges at step 2 (' displays' -1.5065 against ' features' -1.7449, 0.238 nats), two images at step 0 ('3' -0.511 against "Let's" -1.3771, 0.866 nats). Ruled out: TF32 (MLX_ENABLE_TF32=0 changes nothing), both vision activations, DeepStack order and depth, quantization (4-bit and float32 outputs are byte-identical). What remains is the shared Qwen3-VL tower, which this PR does not touch. Filed as Qwen3-VL vision tower is not token-exact against transformers on the image path #1738 with the numbers.
• cargo test --workspace --profile test-fast --features metal,accelerate was not run. It cannot run on this host, which is Linux with CUDA and has neither Metal nor Accelerate. The CUDA equivalent above was run instead.

Closes #1354