feat: integrate FP8 weight transfer format

[S1ro1]

Note

Medium Risk
Adds a new NCCL weight transfer mode that changes the on-the-wire weight format and in-place loading behavior in vLLM workers, which could break updates for unsupported models or mismatched tensor shapes. Guardrails/validators reduce misconfiguration risk but correctness depends on model-specific conversion logic.

Overview
Adds an opt-in quantize_in_weight_transfer flag for NCCL weight broadcast across shared RLConfig, TrainerConfig, and OrchestratorConfig, and wires it through orchestrator → HTTP /init_broadcaster → vLLM worker initialization.

When enabled, the trainer broadcasts vLLM kernel-format FP8 (e4m3) weights instead of HF checkpoint-format tensors, using new per-layer conversion (PreTrainedModelPrimeRL.convert_layer_to_vllm_kernel) and FP8 blockwise quantization utilities. The inference-side NCCL worker now selects between checkpoint loading and a new kernel-mode in-place copy loader (weight_transfer.py) with extra post-processing for certain kernels (e.g., MLA absorbed KV weights).

Also tightens config behavior: validates that quantized transfer is only used with weight_broadcast.type='nccl', with inference configured, and with trainer.model.impl='custom', and adjusts shared model name propagation/validation to keep inference and orchestrator model names consistent.

^{Written by Cursor Bugbot for commit a997302. This will update automatically on new commits. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

[cursor]

Missing CHANGELOG entry for new config field

Low Severity

A new quantize_in_weight_transfer config field is added to NCCLWeightBroadcastConfig in the orchestrator, trainer, and RL config files, as well as SharedWeightBroadcastConfig, but CHANGELOG.md has no corresponding entry. This violates the rule requiring changelog updates when configuration structures are modified.

Additional Locations (2)

• src/prime_rl/configs/orchestrator.py#L644-L648
• src/prime_rl/configs/trainer.py#L586-L595

 

^{Triggered by project rule: BugBot Instructions}