[EXPERIMENT][WIP][OpenVINO] Add support for Cohere2 MoE (cohere2_moe)

[mlukasze]

⚠️ AUTOMATICALLY GENERATED BY OMEGA AGENT — REQUIRES HUMAN REVIEW ⚠️
This PR was created by an AI agent as part of automated model enablement.
A human maintainer must review and approve it before it can be considered for merge.
Do NOT merge without human review and sign-off.

What does this PR do?

Add OpenVINO export support for the Cohere2 MoE architecture (model_type cohere2_moe, Cohere2MoeForCausalLM; e.g. CohereLabs/North-Mini-Code-1.0 — 30B total / 3B active, 128 experts / 8 active, sigmoid router, hybrid sliding-window+RoPE / global-NoPE attention, Cohere2 parallel block).

Changes (minimal, scoped to cohere2_moe):

• Cohere2MoeOpenVINOConfig(Cohere2OpenVINOConfig) registered for cohere2_moe (text-generation[-with-past]).
• Cohere2MoePatcher: replaces the untraceable Cohere2MoeExperts.forward (fused gate_up_proj, per-hit-expert loop) with the existing vectorized lfm2_moe_experts_forward (torch.bmm, chunk(2, dim=-2)), mirroring the Qwen3.5-MoE approach (PR [OpenVINO] Support Qwen3.5, Qwen3.5-MoE and Qwen3.6 #1689/[OpenVINO] Use performant 3GEMM MoE for Qwen3.5 #1728). Patcher numerically verified vs the original forward (max-abs diff ~3e-7 in fp32).
• Default weight-compression config keeping embed_tokens/lm_head (tied) and per-layer MoE routers out of compression via ignored_scope — required for accuracy on this sparse-MoE model.
• Tiny random cohere2_moe added to the OpenVINO test suite.

Validation: tiny-model greedy export matches PyTorch 25/25 tokens. Full model WWB text-similarity vs PyTorch baseline (CPU, greedy, 8 samples): INT8 = 0.925 (PASS, ≥0.9).

Known limitations (flagged for human review, NOT addressed here):

• INT4 weight-only PTQ has a proven accuracy ceiling (~0.78) on this model (swept group_size/ratio/AWQ/scale-estimation) — INT8 is the recommended precision.
• OpenVINO GPU: the compressed-MoE op MOECompressed is not fused on the GPU plugin for cohere2_moe routing (FuseMOE3GemmCompressed only matches activation-before-topk + normalized routers; cohere2_moe applies sigmoid after top-k with no normalization). GPU inference therefore fails (MOECompressed reached the GPU backend without being fused). Additionally INT8 (29GB) exceeds available dGPU VRAM (22.7GB). Recommended owner: OpenVINO GPU plugin / MoE team. CPU INT8 is the supported path.

Installation instructions

pip install "git+https://github.com/huggingface/optimum-intel.git@main#egg=optimum-intel[openvino]"
# transformers from main is required for cohere2_moe (PR #46115):
pip install "git+https://github.com/huggingface/transformers.git"

Exporting cmd-line

# INT8 is the recommended/validated precision (CPU). The per-model default config
# automatically keeps embeddings/lm_head + MoE routers uncompressed for accuracy.
optimum-cli export openvino -m CohereLabs/North-Mini-Code-1.0 --weight-format int8 --task text-generation-with-past ./North-Mini-Code-1.0-ov-int8

Inference script

import openvino_genai as ov_genai

pipe = ov_genai.LLMPipeline("./North-Mini-Code-1.0-ov-int8", "CPU")
config = ov_genai.GenerationConfig()
config.max_new_tokens = 128
config.apply_chat_template = False  # Cohere Jinja template uses constructs unsupported by GenAI Minja
print(pipe.generate("def fibonacci(n):\n    ", config))

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests?