[P2] Long-sequence online-softmax logp & workspace sizing

[Flink-ddd]

Context: 8k-32k+ CoT responses require robust memory management.
Requirements:

• Optimize the online-softmax logp kernel for long sequences.
• Implement proactive workspace sizing and strict memory-safety boundary checks to prevent OOM/segfaults.

[JLiu4Coding]

Hi @Flink-ddd, if we still need help with this issue, I’m interested in working on this. I reviewed the current fused LogP implementation and found that the generic CUDA path already provides online_fp32, online_indexed_out, and online_indexed_fp32, including a basic launch-policy selector.
Would it make sense to split this into two stages? I could first implement long-sequence memory estimation, overflow-safe boundary checks, controlled failure behavior, and reproducible 8K/16K/32K benchmarks. Then, based on profiling, follow with a separate optimization PR.
Could you clarify whether “workspace” refers only to temporary memory allocated by the operator, or to the complete materialized-logits memory requirement, including inputs, outputs, and safety headroom?