fix(refit): release trainer memory before collective export

[seonjinn]

What does this PR do ?

Adds an opt-in trainer-memory release before default non-colocated vLLM collective refit. This avoids large MXFP8 exports, including Qwen3.5 397B, competing with completed gradient buffers, optimizer state, and Transformer Engine workspaces for trainer GPU memory.

The implementation reuses policy.offload_before_refit(), supports GRPO and distillation, preserves the existing path by default, and rejects unsupported backend, topology, and transport combinations.

Issues

N/A

Usage

policy:
  release_grads_before_refit: true
  offload_optimizer_for_refit: true
  megatron_cfg:
    fp8_cfg:
      enabled: true
      force_clear_fp8_caches: true
  generation:
    backend: vllm
    colocated:
      enabled: false
    refit_transport: null

Before your PR is "Ready for review"

Pre checks:

• Read and followed the contributor guidelines.
• Added focused unit coverage for configuration, routing, and refit ordering.
• Run the full unit and functional suites locally. The project lockfile supports Linux only; focused lifecycle checks, Ruff, Pyrefly, compile, and YAML/reference validation passed on macOS.
• Updated the refit guide and exemplar/reference configurations.

Additional Information

The option is disabled by default. It is supported for Megatron policy training with default non-colocated vLLM NCCL collective refit.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[copy-pr-bot]

Auto-sync is disabled for ready for review pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[seonjinn]

/ok to test ca43e3d

[seonjinn]

/ok to test 32b0467