deepseek-ai / DeepGEMM

DeepGEMM: clean and efficient FP8 GEMM kernels with fine-grained scaling

Repository from Github https://github.comdeepseek-ai/DeepGEMM

deepseek-ai/DeepGEMM Issues

[Bug]TMA Multicast code issue
Closed 5 hours ago1
The current strategy in get_best_config is suboptimal in some cases on H20
Updated 4 days ago
关于Wgrad 计算中 fp8_gemm_nt 输入量化方式的疑问与实验观察
Updated 5 days ago1
performance optimizations on CUDA 12.9?
Closed 5 days ago1
How can I use DeepGemm with sm120a（5060）?
Updated 5 days ago1
deepseek TP+DP+EP PD Disaggregation prefill nodes hang, seems related to deepgemm
Closed 7 days ago
Question about the sf layout
Closed 14 days ago2
m_grouped_fp8_gemm_nt_contiguous stuck on matrix shape (1, 1,24576, 1536)
Closed 17 days ago3
Does DeepGEMM support batched GEMM?
Closed 17 days ago1
Create Git tags and releases for versioning
Updated 17 days ago2
Question: it seems that DeepSeek V3.1 is faster than DeepSeek V3(0324 version) when decoding under SGLang and H800.
Closed 18 days ago2
B200(sm=100a) FP8 accumulator bits
Updated 18 days ago1
Bug: ImportError: /home/tiger/.local/lib/python3.12/site-packages/deep_gemm_cpp.cpython-312-x86_64-linux-gnu.so: undefined symbol: cudaLibraryGetKernel, version libcudart.so.12
Closed 18 days ago5
Support grouped gemm bf16 tn/nn contiguous?
Updated 19 days ago1
[Bug] cu128 test_bf16 error
Closed 19 days ago5
RuntimeError: Failed: CUDA runtime error csrc/jit/kernel_runtime.hpp:108 '98'
Closed 20 days ago3
Partner Listing Request：PaddlePaddle 的 PyTorch C++/Python 接口兼容项目
Updated 21 days ago6
Suspicious condition in is_block_size_legal
Closed 24 days ago1
Support CUDA 13
Closed 24 days ago1
[Question] The clamp value in per_token_cast_to_fp8
Closed 25 days ago11
v2 seems to be slower than v1 on hopper on some cases
Updated a month ago18
wgrad_gemm_fp8_fp8_fp32_nt not supporting arbitrary m?
Closed a month ago1
BF16 Gemms
Closed a month ago4
Question about overlapping between promotion and wgmma
Closed a month ago2
Will swap ab tech improve small m (decode phase) performance?
Updated a month ago3
从训练的角度出发，这个可以用在megatron里面吗
Closed a month ago8
[Question] How to call DeepGemm kernel from C++
Closed a month ago1
~5% performance regression for grouped GEMMs on H100
Closed a month ago2
VLLM Integration: Assertion error csrc/jit_kernels/impls/smxx_layout.hpp:136 'mn % 4 == 0 and num_groups == 1'
Closed a month ago2
Potential bug in get_num_blocks calculation for specific grouped GEMM types
Updated a month ago1
undefined symbol: _ZN3c106detail14torchCheckFailEPKcS2_jRKNSt7__cxx1112basic_stringIcSt11char_traitsIcESaIcEEE
Updated a month ago2
[bug] Unspecified Launch Failure of Normal GEMM when n=16
Closed a month ago2
RuntimeError: Failed: Assertion error csrc/jit_kernels/impls/smxx_layout.hpp:89 'dim == 2 or dim == 3'
Updated 2 months ago2
[Question] Why sm100 does not support block_m == 64 ?
Closed 2 months ago3
Question about restriction on BLK_N on SM100 when type is 1d2d
Closed 2 months ago3
weight gradient kernels are deleted?
Closed 2 months ago2
VLLM Integration: `Assertion error csrc/jit_kernels/impls/smxx_layout.hpp:136 'mn % 4 == 0 and num_groups == 1'`
Closed 2 months ago1
The replacement of deep_gemm.wgrad_gemm_fp8_fp8_fp32_nt after Add more GPU architectures support (#112)
Updated 2 months ago1
[Question] Do you have any plans to improve grouped wgrad gemm?
Updated 2 months ago1
[QST]Some questions about unit testing
Closed 2 months ago2
TMA selection of get_best_configs in gemm.py
Updated 2 months ago1
Low E2E throughput performance on SM100
Closed 2 months ago6
Can not install DeepGEMM with commit of 95e81b3
Updated 2 months ago1
find out why wgrad mma will cause illegal memory access in specific shape mnk
Closed 2 months ago6
通过数据内存预处理可能会大幅提升性能
Closed 2 months ago6
The reduction of GPU core frequency leads to significant fluctuations in GEMM performance
Updated 2 months ago1
Wgrad gemm kernel illegal memory access with (m, n, k) in (512, 1024, 147456)
Updated 3 months ago3
how to change input data type as float
Closed 3 months ago1
Bug: AssertionError in `gemm_fp8_fp8_bf16_nt` due to new test case in `test_core.py`
Closed 3 months ago8
[Perf] Grouped W-Grad GEMM is 10–18 % slower than the released benchmark on H200 (#95)
Closed 3 months ago4