v1.2.0

New Features

• Added CoreX (Iluvatar) GPU platform support.
• Added CUDA Graph support through the TSG-based Level-3 preemption implementation.
• Added CUDA PTDS/PTSZ support and expanded CUDA interception coverage.
• Added new scheduling policies, including CFS and MLFQ.

Improvements

• Improved CUDA function handling by converting CUkernel to CUfunction.
• Improved CUDA extra argument handling using updated cuxtra.
• Improved Ascend command argument lifetime handling with deep copies.
• Improved environment handling, logging, IPC, process creation events, and scheduler protocol handling.
• Improved examples and documentation, including LLM serving docs and platform docs.
• Lowered the required CMake version.
• Improved teardown behavior to avoid hangs.

Bug Fixes

• Fixed CUDA pageable host-to-device copy handling.
• Fixed CUDA PTDS/PTSZ binding and interception issues.
• Fixed function address lookup and intercepted version handling.
• Fixed host memory release paths.
• Fixed Python binding issues.
• Fixed duplicate g_sched_agent initialization.
• Fixed minor include and error handling issues.