Fixed the experts_num for MoE on HeteroFlow

update xdnn and pack to support fp8 gemm in prefill (#522)

remove self-host workflow (#527)

remove python wqdependencies (#526)

remove python wqdependencies (#525)

Bump torch from 2.7.0+cpu to 2.7.1 (#519)

Bump protobuf from 5.29.3 to 5.29.5 (#518)

Fixed build issue of oneccl & shm (#147)

balanced split for each expert in MoE (#143)

[Version] 2.1.2. (#146)

[Convert] Fix DeepSeek-R1 convert issue. (#145)

[XDNN] Upgrade xDNN (add new method of FP8 conversion) (#144)

Aligned on the topk method with SGlang&vLLM (#142)

[TIMELINE] Dump timeline.json for vllm-xft. (#141)

Add Batch gemm for BF16BF16BF16 with BA16a64b2a layout (#140)