$ the-wire · showcase
verl lands native NVFP4 training and rollout, open-instruct overlaps MILES startup
By RepoJournal · Filed · About Alignment & Post-training · Composed from the cited sources · methodology
verl shipped end-to-end NVFP4 training and rollout on Blackwell with live refit, while its QAT path moved from a static per-tensor activation scale to per-token scaling, a change flagged breaking for existing NVFP4 QAT configs.
[megatron, vllm, env] feat: add native NVFP4 training and rollout with live refit verl-project/verl
verl now runs native NVFP4 W4A4 training and rollout end to end on NVIDIA Blackwell GPUs, demonstrated with Qwen3-30B-A3B-Base DAPO through Megatron/Transformer Engine and vLLM, including live refit between trainer and rollout engine. Transformer Engine runs selected routed-expert MLPs in NVFP4 with per-token activation scaling.
[BREAKING][megatron, vllm] feat: per-token activations for NVFP4 QAT verl-project/verl
NVFP4 QAT now quantizes each token with a scale from its own amax, matching the NVFP4 rollout scheme, and the static per-tensor activation scale is removed. The PR is tagged breaking, so existing NVFP4 QAT configs that relied on that static scale will not carry over unchanged.
[megatron, rollout, perf] feat: add grouped TileLang quantization for QAT and export verl-project/verl
Grouped TileLang fake-quantization and weight-export paths for FP8, MXFP8 and MXFP4 quantize Transformer Engine grouped expert weights through pointer tables without materializing a stacked copy, and export handles per-expert w1/w2/w3, gate_proj/up_proj/down_proj and fused gate_up_proj/down_proj naming. Fused 3D Megatron-Bridge expert tensors are quantized once and split into the wire format ro...
[megatron] feat: add TileLang FP8 and FP4 QAT kernels verl-project/verl
Shared TileLang quantization kernels now cover FP8_BLOCK, MXFP8 and MXFP4 for QAT and parameter export: QAT uses fused quantize-dequantize with straight-through gradients, while export runs the same kernels without dequantization so values and scales follow one numerical recipe. NVFP4 QAT stays on the Triton kernel from the per-token change.
Overlap independent MILES startup stages allenai/open-instruct
Managed GRPO startup no longer waits for judge readiness before Ray and workflow preparation, and no longer initializes inference before the trainer; independent branches start concurrently with explicit barriers kept before worker placement, initial weight publication and production generation/evaluation. The other open-instruct work was narrower in reach: the TPU image layer retries Orbax's o...
Action items
