Repository navigation

All pull requests

Labels (9)Milestones (0)New pull request

Open

(28)

Closed

(89)

fix(pyutils): raise on cudaFuncSetAttribute and launch errors in bind_kernel
#215TanishGudise opened Oct 26, 2026
tests: build and pass on SM80-class GPUs; add ARCH=SM87 for Jetson Orin
#212phillip-vu opened Sep 21, 20261 of 3 tasks
Implemented DSV4 CSA block compression kernel on H100
#207emanueleromito opened Aug 31, 2026
1
Add initial TVM-FFI bindings for Rotary and LayerNorm
#205mshr-h opened Jul 23, 2026
2
Enable warpgroup-MMA on sm_121 (GB10): WGMMA shim + SM121 build target
#204chauhang opened Jun 23, 2026
fix: guard destroy() against moved-from TKParallelTensor
#202mstkyvz opened Jun 10, 2026
1
Pseudo wgmma
#200KevinLikesDrinkingCoffee opened May 30, 2026
Explicitly cast char to int8_t
#197save-buffer opened May 21, 2026
m3 + benchmark harness
#195Lazarus-931 opened Apr 20, 2026
2
fp8 shared vector support (e4m3/e5m2/e8m0)
#193Emre-Dinc opened Apr 9, 2026
fix MMA Pipeline theoretical race
#190Topologized opened Mar 29, 2026
Possible solution of using DLPack to handle TVM FFI
#181haoran35-jpg opened Mar 17, 2026
Fix narrowing conversion from char to signed char in packing #116
#179florianmattana opened Mar 10, 2026
Fixes H100 attn & matmul kernel
#155symlons opened Oct 28, 2025
added a docker container (only try on GPU wont work on CPU)
#146Kushagra481 opened Aug 24, 2025
Fix unit test Makefile bugs for A100 and 4090 results in linker failure due to cuda libraries missing
#145tomflinda opened Aug 20, 2025Contributor
Document and simplify fp16 -> fp8 conversion
#140melonedo opened Aug 3, 2025
Add Implementation of Native Sparse Attention
#137yukavio opened Jul 22, 2025
Added LCSF template implementation support for FlashAttention Backward.
#135KuangjuX opened Jul 22, 20252 tasks done
5
Remove redundant register declarations.
#134KuangjuX opened Jul 22, 2025
Remove unnecessary device and stream syncs
#129Edenzzzz opened Jun 15, 2025
10
Use -gencode instead of -arch in mla_decode
#126lucifer1004 opened Jun 6, 2025
1
40% build speedup: get torch and python include paths without subprocesses
#119technillogue opened May 7, 2025
Gpt2
#103oleitersdorf opened Mar 17, 2025
1
global_to_shared.cuh row accessing fixes
#95dylanllim opened Feb 21, 2025Collaborator

Navigation Menu

Repository navigation

All pull requests

fix(pyutils): raise on cudaFuncSetAttribute and launch errors in bind_kernel

tests: build and pass on SM80-class GPUs; add ARCH=SM87 for Jetson Orin

Implemented DSV4 CSA block compression kernel on H100

Add initial TVM-FFI bindings for Rotary and LayerNorm

Enable warpgroup-MMA on sm_121 (GB10): WGMMA shim + SM121 build target

fix: guard destroy() against moved-from TKParallelTensor

Pseudo wgmma

Explicitly cast char to int8_t

m3 + benchmark harness

fp8 shared vector support (e4m3/e5m2/e8m0)

fix MMA Pipeline theoretical race

Possible solution of using DLPack to handle TVM FFI

Fix narrowing conversion from char to signed char in packing #116

Fixes H100 attn & matmul kernel

added a docker container (only try on GPU wont work on CPU)

Fix unit test Makefile bugs for A100 and 4090 results in linker failure due to cuda libraries missing

Document and simplify fp16 -> fp8 conversion

Add Implementation of Native Sparse Attention

Added `LCSF` template implementation support for FlashAttention Backward.

Remove redundant register declarations.

Remove unnecessary device and stream syncs

Use `-gencode` instead of `-arch` in mla_decode

40% build speedup: get torch and python include paths without subprocesses

Gpt2

global_to_shared.cuh row accessing fixes

Pull requests