feat: SM-constraint-GEMM by triton persistent kernel (#982)

perf: prefetch page indices for mla kernel (#991)

misc: fix devcontainer conda path (#989)

ci: add torch 2.6+cu126 wheel (#985)

misc: update devcontainer (#986)

ci: switch to on-demand instances if spot instance is interrupted (#987)

misc: Rename output_emitted_token_num -> `output_emitted_draft_toke…

feat: Allow passing workspace base directory via environment variable (…

triton: Triton rms_norm kernels (#983)

misc: Use environment variable to control JIT verbose flag (#981)

Triton rms_norm kernels

bugfix: Fix compilation with FP16_QK_REDUCTION enabled. (#962)

release: bump version to v0.2.4 (#980)

perf: Use 2WG pipeline design for MLA implementation on Hopper (#952)

perf: dual pivot top-p/top-k renorm (#974)

benchmark: add sampling.renorm benchmarks (#970)

bugfix: Fix POD JIT bugs (#971)

perf: Fix python API overhead when CUDAGraph is not enabled (#969)

feat: Added tvm binding for sampling kernel (#958)

perf: reduce torch.library dispatch overhead (#968)

doc: remove misleading docstring about non_blocking (#966)

remove-non-blocking-docstring

bugfix: Fix compilation on cuda 12.2 (#961)

ci: improve jenkins (#943)

misc: Temporarily disable POD from AOT wheels (#956)

misc: Temporarily disable POD from AOT wheels

bugfix: bugfix to #949 (#951)