[ckpt] feat: implement large tensor slicing in vllm rollout and CheckpointEngine for weight updating#5378

Open

jianjunzhong wants to merge 19 commits intoverl-project:mainfrom

jianjunzhong:feat/chunked_weight_update

Commits on Feb 24, 2026

[vllm] feat: implement chunked weight handling in vllm rollout for large tensors
jianjunzhong
committed
fix: handle large weight tensors in chunking logic to prevent ZeroDivisionError
jianjunzhong
committed

Commits on Feb 25, 2026

[vllm] feat: implement chunked weight handling in vllm rollout for large tensors
jianjunzhong
committed
fix: handle large weight tensors in chunking logic to prevent ZeroDivisionError
jianjunzhong
committed
feat: implement chunked weight handling in checkpoint engines to support large tensors
jianjunzhong
committed
refactor: move tensor yielding logic to CheckpointEngine for better code reuse
jianjunzhong
committed
feat: implement collective checkpoint engine for NCCL and HCCL with chunked weight handling
jianjunzhong
committed
refactor: implement chunk slicing in NIXLCheckpointEngine using base class method
jianjunzhong
committed
fix: add properties for bucket size, rank, and buffers in HCCL and NCCL checkpoint engines
jianjunzhong
committed
fix: add update_weights_bucket_megabytes parameter to checkpoint engine configs and clone tensors in pending chunks
jianjunzhong
committed
refactor: streamline HCCL and NCCL checkpoint engines
jianjunzhong
committed
fix: update total_params calculation in NIXL and Collective checkpoint engines
jianjunzhong
committed
[ckpt] feat: implement large tensor slicing in CheckpointEngine
jianjunzhong
authored
refactor: utilize compute_weight_chunks utility for weight tensor slicing in ServerAdapter and CheckpointEngine
jianjunzhong
committed
fix: add bucket_size property with getter and setter
jianjunzhong
committed