[roadmap] low precision support

[sophiayyya]

Rollout

FP8

• Doc: https://github.com/verl-project/verl/blob/main/docs/advance/fp8.md
• Recipe: https://github.com/verl-project/verl-recipe/tree/main/low_precision
• Supported backend: vllm/sglang/trtllm
• Supported scenario: colocated&disaggregated; blockwise fp8 recipe; only quantize fp8 linear layer;
• Hardware requirement: supported on Hopper and Blackwell
• Unsupported scenario: hidden/intermediate size must be an integer multiple of block_size(128) (related code)
• Related PR:
  • vllm: [rollout, vllm] feat: support blockwise fp8 rollout #3519
  • sglang: [rollout, sglang] feat: support blockwise fp8 rollout #4415
  • trtllm: [BREAKING][trtllm] feat: Add FP8 refit support for trtllm rollout #5374

MXFP8

• MXFP8 rollout

QAT

NVFP4 W4A16 QAT

• Doc: https://github.com/verl-project/verl/blob/main/docs/advance/nvfp4_qat.md
• Recipe: https://github.com/verl-project/verl-recipe/tree/main/qat
• Supported backend: vllm;
• Hardware requirement: only supported on Blackwell
• Related PR:
  • vllm:
    • FSDP [fsdp,algo] feat: add NVFP4 QAT (Quantization-Aware Training) support #5190
    • Megatron: feat(recipe/qat): Add Megatron backend support for NVFP4 W4A16 QAT recipe verl-recipe#55

Train

FP8

• Doc: https://github.com/verl-project/verl/blob/main/docs/advance/fp8.md
• Recipe: FP8 end-to-end RL: https://github.com/verl-project/verl-recipe/blob/main/low_precision/run_dapo_qwen3_moe_30b_megatron_fp8e2e.sh
• Supported backend: Megatron
• Supported scenario: quantize linear layer, all GEMMs (FPROP, DGRAD and WGRAD) for QKV, Proj, FC1 and FC2

• Hardware requirement: only supported on Hopper and Blackwell

MXFP8

• MXFP8 training in Megatron
• Hardware requirement: Blackwell

NVFP4

• NVFP4 training in Megatron
• Hardware requirement: Blackwell