fix: import GLM5 fused experts under transformers>=5.13

[Linxiushen]

Fixes #382

Problem

angelslim/models/llm/glm5_1.py does

from transformers.models.glm_moe_dsa.modeling_glm_moe_dsa import GlmMoeDsaNaiveMoe

and angelslim/models/llm/__init__.py imports GLM5_1 eagerly. transformers renamed that class to GlmMoeDsaExperts in 5.13.0 (5.12.x still has GlmMoeDsaNaiveMoe; 5.13.0 onwards only GlmMoeDsaExperts), while requirements/requirements.txt allows transformers>=5.6.0,<6.0. So with transformers 5.13–5.17 (current release 5.17.0) every entry point fails at import time, for any model:

$ python tools/run.py --help
ImportError: cannot import name 'GlmMoeDsaNaiveMoe' from 'transformers.models.glm_moe_dsa.modeling_glm_moe_dsa'
$ python -c "from angelslim import Engine"
ImportError: cannot import name 'GlmMoeDsaNaiveMoe' ...

Fix

Try the old name first and fall back to the new one:

try:
    from transformers.models.glm_moe_dsa.modeling_glm_moe_dsa import GlmMoeDsaNaiveMoe
except ImportError:
    from transformers.models.glm_moe_dsa.modeling_glm_moe_dsa import GlmMoeDsaExperts as GlmMoeDsaNaiveMoe

The alias is exact, not just name-compatible: diffing GlmMoeDsaNaiveMoe (v5.12.0) against GlmMoeDsaExperts (v5.17.0) in transformers shows the only change is the class name — same num_experts / hidden_dim / intermediate_dim / act_fn attributes, same 3-D gate_up_proj / down_proj parameters, same forward(hidden_states, top_k_index, top_k_weights). Those are exactly what _is_glm_naive_moe() and GlmExpertsWithLinear rely on, so the GLM5 quantization path keeps working rather than silently skipping the expert replacement.

Test

tests/test_glm5_experts_compat.py:

• angelslim.models.llm imports, and glm5_1.GlmMoeDsaNaiveMoe is whichever fused-experts class the installed transformers provides;
• a tiny GlmMoeDsaConfig (hidden 8, intermediate 6, 4 experts) fused experts module wrapped in GlmExpertsWithLinear gives the same forward output as the original module (torch.testing.assert_close), i.e. the linearization still matches the renamed class.

Checked with transformers 5.17.0 / torch 2.14: on main both import angelslim.models and tools/run.py --help raise the ImportError above and the new test errors at collection; with this change they succeed, the two new tests pass, and tests/test_config_parser.py + tests/test_install_extras.py still pass (8 passed). black (99) / isort (black profile) / flake8 (99) clean on the changed lines (the E231 flake8 reports in glm5_1.py are pre-existing, on lines this PR does not touch).

Related: #330 is a different, runtime-level transformers-5.x incompatibility (rope / tied-weights); this PR only restores importability.

Written with Claude Code (AI-assisted) and submitted under the account owner's authorization.