[hmellor]

Overview

Move Glm, Arcee, CWM and Mellum from native vLLM implementations to the Transformers modeling backend, following #53615. Removes 3 model files (~650 lines); all four architectures remain officially supported.

Claims

• GlmForCausalLM, ArceeForCausalLM, CwmForCausalLM → TransformersForCausalLM; MellumForCausalLM → TransformersMoEForCausalLM.
• Deletes glm.py, arcee.py and mellum.py. CWM was a registry alias of LlamaForCausalLM, so it only needs a registry change.
• gsm8k accuracy and throughput match main within run-to-run noise for all four models.
• Docs rows move into the "supported via the Transformers modeling backend" table.

Validation

1x B200 per model, main (d848c4ed3d) vs this branch, transformers 5.17.0.

Model	gsm8k main	gsm8k branch	Output tok/s main	Output tok/s branch
zai-org/glm-4-9b-chat-hf	0.637	0.650	7,481	7,591
arcee-ai/AFM-4.5B-Base	0.518	0.521	13,657	13,820
facebook/cwm	0.850	0.852	1,761	1,766
JetBrains/Mellum2-12B-A2.5B-Base	0.803	0.809	16,604	16,498

Throughput is from one run per cell, so the ±1.5% differences are noise, not a speedup or regression.

gsm8k (5-shot, 1319 questions):

vllm serve <model> --port 8200 --max-model-len 8192 --gpu-memory-utilization 0.6
python tests/evals/gsm8k/gsm8k_eval.py --port 8200

Throughput:

vllm bench throughput --model <model> \
  --max-model-len 8192 --gpu-memory-utilization 0.6 \
  --input-len 1024 --output-len 256 --num-prompts 1000

Tests:

pytest tests/models/test_initialization.py -k "ArceeForCausalLM or CwmForCausalLM or GlmForCausalLM or MellumForCausalLM"  # 4 passed (B200)
pytest tests/models/test_registry.py  # 390 passed, 18 skipped

---

Pull Request Checklist

• I used vLLM's /pr-checklist skill. (Mandatory for agents, optional for humans).

• AI assistance was used during the creation of this PR.

• Design Fit: Minimizes impact on core components, reuses existing functionality, and justifies added complexity.

• Testing and Validation: Validates the change and ensures any added tests are meaningful and reliable, with CI coverage or documented CI resource constraints and validation performed outside CI.

• Code Quality and Style: Keeps code and comments clear and concise, and updates relevant documentation and examples.

• Pull Request Contents: Includes a brief summary and relevant links, supports claims with evidence, explains root causes and implementation trade-offs, and follows the contributing guide.

🤖 Generated with Claude Code

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[mergify]

Documentation preview: https://vllm--59679.org.readthedocs.build/en/59679/

[DarkLight1337]

Thanks!

[DarkLight1337]

/ci run

[github-actions]

✅ Triggered Buildkite CI #92531 for commit 3b7dedd38bfc.