biondizzle/DeepGEMM
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

[Public release 26/04] Introducing Mega MoE, FP4 Indexer and other features/fixes (#304)

Browse Source 
* Merge with private repo

* Update README

* Update README

* Update README

* Add PyTorch requirements

* Fix sync scopes for MQA logits (#256)

* Update README
This commit is contained in:
Chenggang Zhao
2026-04-17 09:45:14 +08:00
committed by GitHub
parent d30fc36c8f
commit 7f2a703ed5
109 changed files with 12101 additions and 3219 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
csrc/utils/math.hpp
View File

@@ -1,3 +1,4 @@
// TODO: merge this file with `math.cuh` (the device part)
#pragma once
#include <torch/python.h>
@@ -6,8 +7,8 @@
namespace deep_gemm {
// TODO: Use `torch::kFloat4_e2m1fn_x2`
constexpr auto kPackedFP4 = torch::kUInt8;
// TODO: use `torch::kFloat4_e2m1fn_x2`
constexpr auto kPackedFP4 = torch::kInt8;
template <typename T>
static T ceil_div(const T& a, const T& b) {