sgl-deep-gemm 0.2.5.post2

Introduction

sgl-deep-gemm is a pypi package built from SGLang's customized branch of DeepGemm. Comparing with origina DeepGemm, it supports the following features to better support SGLang:

1. ABI support: ops are registered with TORCH_LIBRARY and the extension has no CPython dependency, so a single wheel runs on every Python version (it is still built against one PyTorch version).
2. pypi support: easy installation with pip install sgl-deep-gemm. No need to manually search for wheel links.
3. Fast iteration: add custom kernels and bump versions at no time.

Usage

Build requirements include CUDA 12.9 or newer, a C++20 compiler with std::format support (for example GCC 13), and the elfutils development headers (apt-get install libdw-dev on Debian/Ubuntu). Initialize the CUTLASS and DeepJIT submodules with git submodule update --init --recursive.

To build it locally, run bash build_sgl_deep_gemm.sh, then pip install the wheel generated under dist.

K-grouped FP8 GEMM and its scale-packing helper preserve compact per-group scaling factors by default. Pass use_padded_sf_layout=True to both packing and GEMM when packed scales cover each group's padded K extent. The flag is appended to existing arguments, so existing positional calls retain their behavior.

SM100 Mega MoE sizes BLOCK_M from the expected tokens per expert, assuming every EP rank sends as many tokens as the local rank. Under attention data parallelism the ranks can be uneven, and an idle rank would then pick the smallest tile while it still computes the experts other ranks route to it. Pass global_num_tokens (the real token count summed over all EP ranks, identical on every rank) to fp8_fp4_mega_moe, nvfp4_mega_moe, or bf16_mega_moe to size tiles from the EP-wide load; pass the same value to get_block_m_for_mega_moe when laying out shared-expert scales. The hint only affects tile selection, and omitting it keeps the previous heuristic.

Run wheel validation with bash sgl_deep_gemm/run_tests.sh. The Mega MoE reference comparisons require DeepEP with ElasticBuffer; Mega Gate and Mega mHC also require TileKernels and TileLang. Install the validation extras with pip install 'sgl-deep-gemm[dev]' (TileLang 0.1.9 and TileKernels 1.0.0). It runs legacy, lazy-init, and compute-sanitizer checks; --skip-sanitizer is available for ordinary development runs. Distributed MegaMoE runs include separate Hopper numerical accuracy coverage; the sanitizer's automatic discovery covers single-process tests and reports distributed entrypoints separately.

Run bash sgl_deep_gemm/run_tests.sh --release for reduced attention coverage and focused memcheck/synccheck, targeting a release gate under one hour. Selected cases retain their original assertions; other ordinary and distributed tests remain unchanged. The default command keeps the full suite; the table shows release / full attention case counts.

Use Compute Sanitizer 2025.4.1 or newer for Python tests; older releases can retain Python tensors while collecting host backtraces and exhaust GPU memory. Set COMPUTE_SANITIZER to select a separately installed executable. See NVIDIA's release notes.

To release a new set of wheels, please contact SGLang team and run the release workflow under SGLang repo

For each major version release (0.X.Y -> 0.(X+1).0), a new branch should be created (release/v0.(X+1).0) for stability purpose.

For any incoming pull requests, it should be rebased upon dev branch. Any newly added or modified tests should be put under sgl_deep_gemm/tests