amd/deepseek_v4 integration 2/N - cuda graph 0426#23832

Merged

HaiShaw merged 19 commits intosgl-project:amd/deepseek_v4from

HaiShaw:amd/deepseek_v4_cuda-graph_0426

Apr 27, 2026

Commits on Apr 26, 2026

chore: bump sgl-kernel version to 0.4.1.post1 (#23720)

authored
Fix Qwen3 MoE: also guard EP all-reduce with not use_reduce_scatter (follow-up to #23731) (#23734)

authored
Apply should_use_dp_reduce_scatterv guard to remaining MoE models (follow-up to #23731) (#23732)

authored
docs(DeepSeek-V4): add GB200 platform to cookbook recipe (#23725)
yhyang201
authored
[CI] release-whl-kernel: clean root-owned build artifacts before checkout (#23747)

Kangyan-Zhou
and
claude
authored
docs(DeepSeek-V4): add h200|big verified recipes + tune H200 Pro parameters (#23742)
yushengsu-thu
authored
[CI] release-whl-kernel: strip +cu129 local version before PyPI upload (#23749)

Kangyan-Zhou
and
claude
authored
[CI] release-pypi-nightly: install protoc before building wheel (#23750)

Kangyan-Zhou
and
claude
authored
chore: bump sglang-kernel version to 0.4.1.post1 (#23733)

authored
[diffusion] refactor: make timestep scheduler request-local (#23716)
mickqian
authored
[MoE] Deprecate act_and_mul_triton; fold filter_expert into JIT silu/gelu_and_mul (#23707)

ch-wan
and
claude
authored
Add DeepSeek-V4 (compressed attention) model support
sogalin
committed
Fix pre-commit: ruff unused imports, codespell typos, formatting
sogalin
committed