What's Changed

• ci: fix whell build script by @guocuimi in https://github.com/vectorch-ai/ScaleLLM/pull/418
• kernel: added attention combine kernel to support split kv by @guocuimi in https://github.com/vectorch-ai/ScaleLLM/pull/419
• kernel: refactor and added more unittests for attn combine kernel by @guocuimi in https://github.com/vectorch-ai/ScaleLLM/pull/420
• moe: added token dispatcher interface for MOE layer. by @guocuimi in https://github.com/vectorch-ai/ScaleLLM/pull/421
• moe: added local token dispatcher pytorch implementation for testing by @guocuimi in https://github.com/vectorch-ai/ScaleLLM/pull/422
• nccl: added all2all for nccl process group by @guocuimi in https://github.com/vectorch-ai/ScaleLLM/pull/423
• moe: added all to all token dispatcher pytorch implementation by @guocuimi in https://github.com/vectorch-ai/ScaleLLM/pull/424
• upgrade cutlass to 3.9 by @guocuimi in https://github.com/vectorch-ai/ScaleLLM/pull/425
• kernel: added fused gate for moe by @guocuimi in https://github.com/vectorch-ai/ScaleLLM/pull/426
• chore: added pre-commit-config by @guocuimi in https://github.com/vectorch-ai/ScaleLLM/pull/427
• kernel: added moe permute kernels by @guocuimi in https://github.com/vectorch-ai/ScaleLLM/pull/428
• chore: clean up attn dependencies by @guocuimi in https://github.com/vectorch-ai/ScaleLLM/pull/429
• chore: clean up JinjaChatTemplate by @guocuimi in https://github.com/vectorch-ai/ScaleLLM/pull/430
• test: added different dtype unittests for moe permute kernels by @guocuimi in https://github.com/vectorch-ai/ScaleLLM/pull/431
• refactor: use __ldlu to load/store data and refactor code for moe permute kernels by @guocuimi in https://github.com/vectorch-ai/ScaleLLM/pull/432
• upgrade pytorch to 2.7 by @guocuimi in https://github.com/vectorch-ai/ScaleLLM/pull/434
• chore: build manylinux2_28 builder image by @guocuimi in https://github.com/vectorch-ai/ScaleLLM/pull/435
• fix: fix manylinux2_28 build by @guocuimi in https://github.com/vectorch-ai/ScaleLLM/pull/436
• upgrade vcpkg after switch to manylinux_2_28 by @guocuimi in https://github.com/vectorch-ai/ScaleLLM/pull/437
• chore: add option to install py module into scalellm folder by @guocuimi in https://github.com/vectorch-ai/ScaleLLM/pull/438
• chore: add script to install zsh for devbox by @guocuimi in https://github.com/vectorch-ai/ScaleLLM/pull/439
• ci: enable docker cache by @guocuimi in https://github.com/vectorch-ai/ScaleLLM/pull/441
• kenerl: add kernel for moe permutation with mask map by @guocuimi in https://github.com/vectorch-ai/ScaleLLM/pull/433
• kernel: added align block permutation kernel for moe by @guocuimi in https://github.com/vectorch-ai/ScaleLLM/pull/442
• build: added build for blackwell by @guocuimi in https://github.com/vectorch-ai/ScaleLLM/pull/459
• chore: upgrade cutlass to v4.0 by @guocuimi in https://github.com/vectorch-ai/ScaleLLM/pull/460
• ci: change self-hosted runner tags by @guocuimi in https://github.com/vectorch-ai/ScaleLLM/pull/461

Full Changelog: https://github.com/vectorch-ai/ScaleLLM/compare/v0.2.4...v0.2.5