[ravi9]

Moved from ravi9/llama.cpp-old#9 by @yzhou51 after the repository restore.

Overview

To add these NPUW dumps envs for profiling goal.

Additional information

• NPUW_DUMP_SUBS / NPUW_DUMP_SUBS_DIR: export selected subgraph IR to inspect partition boundaries, shapes, and operations after timing identifies a slow stage. IR does not itself measure runtime.
• NPUW_DUMP_IO: capture selected subgraph inputs and outputs when validating data flow around that stage.
• NPUW_DUMP_SUBS_ON_FAIL: capture IR for a subgraph that fails compilation; this supports failure diagnosis, not performance measurement.

All these env flags are opt-in, require an OpenVINO NPU developer build, and do not change llama.cpp’s KV-cache management.

Requirements

• I have read and agree with the contributing guidelines
  No touch the code logic, only add some opt-in dump flags.

• AI usage disclosure:
  No.

[yzhou51]

Hi @cavusmustafa @wine99 Can you help take a review for this? Thanks.

[wine99]

OK for me to add these debug env vars