Add OpenVINO export support for glm4_moe_lite (GLM-4.7-Flash)

[openvino-agent]

What does this PR do?

Adds native OpenVINO IR export support for the glm4_moe_lite model type, which powers the GLM-4.7-Flash family of models (e.g. THUDM/GLM-4.7-Flash).

Architecture overview

Glm4MoeLiteForCausalLM is a decoder-only transformer (available in Transformers ≥ 5.0) that combines:

• Multi-head Latent Attention (MLA) — same LoRA-compressed KV projection used in MiniCPM3/DeepSeek, with separate qk_nope_head_dim/qk_rope_head_dim key dimensions and a smaller v_head_dim
• Hybrid MLP layers — alternating dense (Glm4MoeLiteMLP) and sparse Mixture-of-Experts (Glm4MoeLiteMoE / Glm4MoeLiteNaiveMoe) layers

Changes

optimum/exporters/openvino/model_configs.py

• Import Glm4MoeLitePatcher
• Add Glm4MoeLiteOpenVINOConfig registered under glm4_moe_lite for text-generation and text-generation-with-past tasks. Inherits from MiniCPM3OpenVINOConfig to reuse OVMiniCPM3DummyPastKeyValuesGenerator (which correctly handles MLA-style KV cache shapes: k_head_dim = qk_nope_head_dim + qk_rope_head_dim, v_head_dim).

optimum/exporters/openvino/model_patcher.py

• Add glm4_moe_lite_naive_moe_forward — a fully vectorized replacement for Glm4MoeLiteNaiveMoe.forward. The original implementation uses a nonzero()-driven Python for loop over active experts, which is incompatible with torch.jit.trace (loop count varies per input). The patch replaces it with batched matrix multiplications (torch.bmm) over all experts simultaneously, producing a consistent graph regardless of routing decisions.
• Add Glm4MoeLitePatcher — patches the experts sub-module inside every sparse MoE layer during export and restores the original on exit.

Tests

• tests/openvino/utils_tests.py — model ID entry for glm4_moe_lite and expected INT8 node count (42)
• tests/openvino/test_decoder.py — add glm4_moe_lite to SUPPORTED_ARCHITECTURES (Transformers ≥ 5.0) and EXPECTED_NUM_SDPA
• tests/openvino/test_export.py — add glm4_moe_lite export test (Transformers ≥ 5.0)
• tests/openvino/test_exporters_cli.py — add CLI export test for text-generation-with-past
• tests/openvino/test_quantization.py — add to SUPPORTED_ARCHITECTURES_WITH_AUTO_COMPRESSION

Docs

• docs/source/openvino/models.mdx — list Glm4MoeLite (GLM-4.7-Flash) as a supported model

Before submitting

• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests?