chenyu 1891ebb655 make ring allreduce chunks a multiple of 2^n if possible (#4302) ...

in resnet, instead of chunking as [43691, 43691, 43691, 43691, 43690, 43690], chunk as [43712, 43712, 43680, 43680, 43680, 43680] and those can have 32 local.

more than 2X faster for the applicable kernels and overall 1% for resnet

2024-04-25 23:45:28 -04:00

..

codegen

Optimize ptx loops (#4263)

2024-04-23 12:20:14 +04:00

engine

save the schedule graph in SAVE_SCHEDULE (#4248)

2024-04-24 12:08:51 +03:00

features

make ring allreduce chunks a multiple of 2^n if possible (#4302)

2024-04-25 23:45:28 -04:00

nn

use at least float32 for optim.lr (#4297)

2024-04-25 14:42:28 -04:00

renderer

Ptx beam fix (#4296)

2024-04-25 15:39:39 -04:00

runtime

fix lds size for amd (#4287)

2024-04-24 22:54:42 +03:00

shape

clean up update stats (#4226)

2024-04-19 15:41:30 +04:00

__init__.py

add mnist tutorial

2024-04-25 16:08:32 +08:00

buffer.py

new memory scheduler with explicit refcounts (#4198)

2024-04-17 08:46:47 +04:00

device.py

more docs (#4271)

2024-04-24 10:52:42 +08:00

dtype.py

rename Scalar to ConstType and cast_scalar to as_const (#3946)

2024-03-26 22:39:58 -04:00

function.py

New docs are in mkdocs (#4178)

2024-04-16 10:59:51 +04:00

helpers.py

search: add additional 7 factors to the action space (#4256)

2024-04-22 19:14:23 -04:00

lazy.py

new split reduce heuristic try 2 (#4294)

2024-04-25 18:14:15 -04:00

ops.py

si.inputs+outputs -> bufs (#4279)

2024-04-24 15:12:34 +08:00

tensor.py

setitem no return value (#4266)

2024-04-23 16:28:14 -04:00