fix(pytorch): add CPU affinity to optimize Qwen3-0.6B and LLaMA-3.1-8B training
Co-authored-by: HanhuiChen<chenhanhui1@h-partners.com>
# message auto-generated for no-merge-commit merge:
!4327 merge master into master
fix(pytorch): add CPU affinity to optimize Qwen3-0.6B and LLaMA-3.1-8B training
Created-by: HANHU1CHEN
Commit-by: HanhuiChen
Merged-by: ascend-robot
Description: ## What this PR does / why we need it?
Bind inference processes to specific CPU cores via affinity settings for Qwen3-0.6B and LLaMA-3.1-8B scripts to reduce cross-core scheduling overhead and stabilize throughput.
## Does this PR introduce any user-facing change?
No interface changes; users may optionally configure core binding parameters in the scripts as needed.
## How was this patch tested?
Ran repeated inference benchmarks on both models before and after the change, confirming reduced latency jitter and improved throughput consistency.
See merge request: Ascend/MindSpeed-LLM!4327
