[codex] Support GLM-5.2 744B Modal H100 run

[pawalt]

Summary

Adds a Modal/SLIME configuration and runtime support needed to run GLM-5.2-744B-A40B GRPO on a 32-node H100 allocation.

The working path uses colocated SGLang rollout engines instead of PD disaggregation because the available Modal/AWS transfer paths were not viable for this setup: Mooncake TCP failed CUDA buffer copies, NIXL/libfabric needed unavailable EFA memory-registration privileges, and NIXL/UCX only exposed loopback to the plugin.

Changes

• Add glm52_744b_a40b experiment config for 32 H100 nodes, FP8 GLM-5.2, DAPO-Math, colocated rollout, EFA disabled, and the validated memory settings.
• Add Modal runtime patches for GLM bridge config fixes, Megatron allgather-CP RoPE handling, SGLang KV CPU-copy compatibility, and the single-tensor weight-update path.
• Add helper wrappers so runtime patches are applied consistently inside Modal workers.
• Make Modal EFA experimental options conditional and add network diagnostics for Modal/Ray/NIXL/EFA debugging.
• Adjust checkpoint/launcher helpers used by the GLM run.

Validation

• PYTHONPATH=slime uv run python -c "from configs.glm52_744b_a40b import slime, modal; ..."
  • rollout_max_response_len 1024
  • max_tokens_per_gpu 4096
  • data_pad_size_multiplier 64
  • efa_enabled False
• uv run python -m py_compile on all changed Python files.
• Live Modal run on 32 H100 nodes completed a full step:
  • app ap-FhbQwfQSx1MpdtFmgeE7dL
  • W&B run q5nwqqfd
  • initial update_weights completed in 7429.1s
  • rollout completed 64/64, response length 1024, rollout time 75.38s
  • actor train completed 3/3 microbatches without the previous RoPE, MoE, or fused CE OOM failures
  • step 0 metrics logged successfully