deepseek-ai/DeepGEMM

DeepGEMM: clean and efficient BLAS kernel library on GPU

Stars: 7.8k
Forks: 1.3k
Open beginner issues: 0
Indexed issues: 69
Avg merge: 3d 7h
Merged PRs (30d): 3

Dominant language: Cuda
License: MIT
Last GitHub push: Sep 14, 2026
Latest indexed: Sep 15, 2026
Contributing guide: No contributing guide
Code of conduct: No code of conduct
Beginner labels: No beginner labels indexed

Open All

69 open issues indexed

Support for Ada Lovelace & Blackwell Open

deepseek-ai/DeepGEMM#6 · 13 comments · 10 reactions · 0 assignees · Indexed Sep 15, 2026
AssertionError: Errors of the kernel fp8_gemm in the profiling table Open

deepseek-ai/DeepGEMM#9 · 3 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
How to use the output of DeepEP as the input of DeepGemm Open

deepseek-ai/DeepGEMM#57 · 4 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
fp8_wgrad_gemm Support unaligned n,k and gmem stride? Open

deepseek-ai/DeepGEMM#96 · 1 comment · 0 reactions · 0 assignees · Indexed Sep 15, 2026
m_grouped_gemm_fp8_fp8_bf16_nt_contiguous's requiresment of 128 padding for each expert causes computation waste Open

deepseek-ai/DeepGEMM#98 · 11 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
[Help] Starter script for nn.Linear for fwd,bwd Open

deepseek-ai/DeepGEMM#102 · 8 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Wgrad gemm kernel illegal memory access with (m, n, k) in (512, 1024, 147456) Open

deepseek-ai/DeepGEMM#110 · 3 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Will swap ab tech improve small m (decode phase) performance? Open

deepseek-ai/DeepGEMM#114 · 3 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
The reduction of GPU core frequency leads to significant fluctuations in GEMM performance Open

deepseek-ai/DeepGEMM#120 · 1 comment · 0 reactions · 0 assignees · Indexed Sep 15, 2026
[Question] Do you have any plans to improve grouped wgrad gemm? Open

deepseek-ai/DeepGEMM#125 · 1 comment · 1 reaction · 0 assignees · Indexed Sep 15, 2026
TMA selection of get_best_configs in gemm.py Open

deepseek-ai/DeepGEMM#126 · 1 comment · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Can not install DeepGEMM with commit of 95e81b3 Open

deepseek-ai/DeepGEMM#127 · 1 comment · 2 reactions · 0 assignees · Indexed Sep 15, 2026
undefined symbol: _ZN3c106detail14torchCheckFailEPKcS2_jRKNSt7__cxx1112basic_stringIcSt11char_traitsIcESaIcEEE Open

deepseek-ai/DeepGEMM#142 · 2 comments · 1 reaction · 0 assignees · Indexed Sep 15, 2026
RuntimeError: Failed: Assertion error csrc/jit_kernels/impls/smxx_layout.hpp:89 'dim == 2 or dim == 3' Open

deepseek-ai/DeepGEMM#143 · 2 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Potential bug in get_num_blocks calculation for specific grouped GEMM types Open

deepseek-ai/DeepGEMM#150 · 1 comment · 0 reactions · 0 assignees · Indexed Sep 15, 2026
v2 seems to be slower than v1 on hopper on some cases Open

deepseek-ai/DeepGEMM#160 · 18 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Partner Listing Request：PaddlePaddle 的 PyTorch C++/Python 接口兼容项目 Open

deepseek-ai/DeepGEMM#170 · 6 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Support grouped gemm bf16 tn/nn contiguous? Open

deepseek-ai/DeepGEMM#171 · 1 comment · 0 reactions · 0 assignees · Indexed Sep 15, 2026
B200(sm=100a) FP8 accumulator bits Open

deepseek-ai/DeepGEMM#176 · 2 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Create Git tags and releases for versioning Open

deepseek-ai/DeepGEMM#177 · 3 comments · 2 reactions · 0 assignees · Indexed Sep 15, 2026
The current strategy in get_best_config is suboptimal in some cases on H20 Open

deepseek-ai/DeepGEMM#191 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Wheel is not supported in the platform Open

deepseek-ai/DeepGEMM#194 · 1 comment · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Perf regression since sha f85ec6 Open

deepseek-ai/DeepGEMM#195 · 2 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
[Feature] Speed up DeepGEMM v2 warm up Open

deepseek-ai/DeepGEMM#196 · 1 comment · 3 reactions · 0 assignees · Indexed Sep 15, 2026
Missing build requirements for 2.1.0 release package Open

deepseek-ai/DeepGEMM#207 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
[QUESTION] dsv32 mqa_logits kernel not considering causal masking? Open

deepseek-ai/DeepGEMM#209 · 1 comment · 1 reaction · 0 assignees · Indexed Sep 15, 2026
[Feature] Support Bias Open

deepseek-ai/DeepGEMM#210 · 5 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
[Question] Supported CUDA/Torch/Python combinations Open

deepseek-ai/DeepGEMM#212 · 21 comments · 0 reactions · 1 assignee · Indexed Sep 15, 2026
Question: planned support for Geforce GPUs (at least Blackwell sm_12x).. Open

deepseek-ai/DeepGEMM#216 · 1 comment · 2 reactions · 0 assignees · Indexed Sep 15, 2026
[QUESTION] Can fp8_paged_mqa_logits support next_n > 2? Open

deepseek-ai/DeepGEMM#222 · 4 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Question: Are ARM prebuilt wheels Open

deepseek-ai/DeepGEMM#223 · 1 comment · 0 reactions · 0 assignees · Indexed Sep 15, 2026
[Question] Contiguous Grouped Gemm for CUDA Graph Open

deepseek-ai/DeepGEMM#235 · 2 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Feature Request: Support sm_120 ( 5090 and blackwell 6000 pro ) Open

deepseek-ai/DeepGEMM#236 · 28 comments · 9 reactions · 0 assignees · Indexed Sep 15, 2026
[Bug] k_grouped_fp8_gemm_nt_contiguous crashes with n = 768 on H100 Open

deepseek-ai/DeepGEMM#237 · 3 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Question about k_grouped_gemm Open

deepseek-ai/DeepGEMM#243 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
[v3.2] deep_gemm::fp8_mqa_logits performs poorly on the H200 Open

deepseek-ai/DeepGEMM#279 · 3 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
weight gradient kernels for dense and MoE backward 没有grad的反向算子 Open

deepseek-ai/DeepGEMM#283 · 1 comment · 0 reactions · 0 assignees · Indexed Sep 15, 2026
[Bug]: NVRTC JIT compilation fails on CUDA 12.8 (smxx_clean_logits.cuh: expression must have a constant value) Open

deepseek-ai/DeepGEMM#295 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
[RFC] Using tvm-ffi binding instead of python api to remove the dependency for python binding Open

deepseek-ai/DeepGEMM#297 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
CUDA kernel errors Open

deepseek-ai/DeepGEMM#300 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
[Bug] JIT cache lacks cross-process synchronization, can cause failures under multi-process parallelism Open

deepseek-ai/DeepGEMM#301 · 1 comment · 0 reactions · 0 assignees · Indexed Sep 15, 2026
fix: operator precedence bug in pack_ue8m0_to_int assertion (mantissa check always passes) Open

deepseek-ai/DeepGEMM#309 · 1 comment · 0 reactions · 0 assignees · Indexed Sep 15, 2026
[Feature Request] Make binary artifacts portable on torch and Python versions Open

deepseek-ai/DeepGEMM#333 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Will DeepGEMM support GEMM + 1d1d quantization all in one fused kernel? Open

deepseek-ai/DeepGEMM#335 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Question about Nsight Compute results for sm100_fp8_fp4_mega_moe_impl on B300/Blackwell Open

deepseek-ai/DeepGEMM#336 · 1 comment · 1 reaction · 0 assignees · Indexed Sep 15, 2026
[RFC] Support FP4 (E2M1) activation dispatch in MegaMoE Open

deepseek-ai/DeepGEMM#338 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Does Mega MoE support RTX 50-series GPUs, and can it work in a PCIe-only setup without NVLink? Open

deepseek-ai/DeepGEMM#356 · 1 comment · 0 reactions · 0 assignees · Indexed Sep 15, 2026
[Feature request] Documented clangd setup for CUDA kernel indexing Open

deepseek-ai/DeepGEMM#358 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
[BUG] NVRTC build fails for `sm90_tf32_hc_prenorm_gemm` — `cudaGridDependencySynchronize` undefined (PDL intrinsic not available under NVRTC) Open

deepseek-ai/DeepGEMM#359 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
[BUG] Non-deterministic numerical problem of k_grouped_fp8_gemm_nt_contiguous Open

deepseek-ai/DeepGEMM#365 · 2 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026

A short digest of beginner-friendly GitHub issues.