mirrors/tinygrad
2
0
Fork 0
mirror of https://github.com/tinygrad/tinygrad.git synced 2026-06-24 02:14:17 +00:00
Code Issues Projects Releases Packages Wiki Activity
8,422 commits 280 branches 15 tags 325 MiB
Commit graph

8,422 commits

This branch
This branch
All branches
Author SHA1 Message Date
qazal
16afe04f45
 		move process replay to grouper (#9830) 
* simpler

* sched
 2025-04-10 18:27:42 +08:00
chenyu
c8f47c1d07
 		not_support_multi_device helper (#9831) 
unify the test helper to skip ci device that does not support multi
 2025-04-10 05:25:29 -04:00
chenyu
817746b30e
 		add contiguous to EmbeddingBert output (#9829) 
for some reason with random dropout it creates different ast on each device. And search embedding is slow. This workaround saved 6 minutes setup time on mi300x (25->19) and resulted in similar speed
 2025-04-10 04:31:19 -04:00
qazal
fd4f06e623
 		kernelize prereqs [pr] (#9811) 
* kernelize prereqs [pr]

* work

* tensor maps to assign

* unwrap st

* process replay

* grouper changes

* replay
 2025-04-10 15:22:20 +08:00
chenyu
c462162db8
 		update benchmark bert scripts with BS and ACC_DTYPE (#9826) 
BS=16, ACC_DTYPE=half for tinybox, BS=128, ACC_DTYPE=float for mi300x
 2025-04-10 02:06:02 -04:00
qazal
498a2bf738
 		add err handling tests to viz + cleanups (#9825) 
* cleanup

* add err handling tests to viz + cleanups

* lint
 2025-04-10 14:05:05 +08:00
chenyu
a0b72f066a
 		don't free intermediate for bert mi300x (#9824) 2025-04-10 01:48:34 -04:00
chenyu
566e389585
 		more relaxed ATOL for HALF=1 simple_matmul test (#9823) 
it's a function of N so only updated in the test command
 2025-04-10 00:46:16 -04:00
Francis Lata
eb2e59db42
 		RetinaNet model type annotations and loss functions (#9822) 
* add type annotations and loss functions for training

* combine sum of multiple dims inside loss functions
 2025-04-10 00:31:37 -04:00
chenyu
06a928b341
 		higher ATOL for half input TC test (#9821) 
flaky
 2025-04-09 23:57:25 -04:00
Francis Lata
7bb36d71b2
 		remove openimages iterate (#9820) 2025-04-09 22:54:12 -04:00
chenyu
2e1002e179
 		EVAL_BS=96 and BEAM=3 for bert green (#9819) 
19m -> 13m setup and same end to end time
 2025-04-09 22:37:27 -04:00
uuuvn
3ee317ffed
 		Fix kfd autogen and verify it in ci (#9818) 
Had to autogen newer uapi headers for #9746 (dmabuf export ioctl missing),
submitting just the fix without updating to newer headers as they are only
needed for infiniband stuff
 2025-04-10 09:53:42 +08:00
nimlgen
d7330ea6ad
 		amd: refactor sqtt into sep functions (#9816) 
* amd: refactor sqtt into sep functions

* fix
 2025-04-10 00:39:45 +03:00
nimlgen
0ca98b9f20
 		amd: gfx9 use cache ctrls in acquire_mem (#9815) 2025-04-09 20:17:02 +03:00
George Hotz
fce432d2e3
 		Ops.FUSE makes softmax a single kernel (#9808) 
* KERNELIZE makes softmax a single kernel

* single kernel works

* softmax works

* broken

* correct

* skip that test

* kernelize tests

* rename to fuse

* better reduce_push_add_ones code

* correct now

* cleanups

* oops

* return None if we can't push ones

* rename + docs

* atol fixes group

* flash attention broken test
 2025-04-09 22:56:28 +08:00
nimlgen
1798ce7e52
 		amd: faster xcc sync (#9783) 
* amd: faster xcc sync

* move to same cacheline

* comment

* keep it uncached + better poll timings

* revert this, should be fine

* fixed now?

* minor
 2025-04-09 15:56:50 +03:00
qazal
3bd992dc95
 		multi stage graph_rewrite_map (#9803) 
* multistage graph_rewrite_map

* s/merge_map/input_map

* build up kernel_map from the tensor_map
 2025-04-09 15:59:45 +08:00
chenyu
57f4bc3fbb
 		add numpy to setup linting (#9806) 
this would have caught the mypy error in fp8 pr. keep ignore_missing_imports to true as we also import torch which is fat
 2025-04-09 03:47:03 -04:00
George Hotz
bf769fa5c5
 		label ranges with their number (#9805) 2025-04-09 14:31:18 +08:00
chenyu
c5db5b83b9
 		add SHOULD_USE_TC=1 check to simple_matmul (#9802) 
* add SHOULD_USE_TC=1 check to simple_matmul

also zero centered the random input and update atol for tf32

* ATOL=2e-2 for HALF
 2025-04-09 02:24:42 -04:00
qazal
f27dbc8c35
 		becomes_map cleanups [pr] (#9790) 
* cleanup becomes_map [pr]

* source
 2025-04-09 14:11:53 +08:00
qazal
7d2349c827
 		track_rewrites in scheduler [pr] (#9801) 2025-04-09 12:48:14 +08:00
George Hotz
bb18adb0d5
 		reduce with a mul chain (#9799) 
* reduce with a mul chain

* inside is just 1
 2025-04-09 12:42:32 +08:00
George Hotz
78caf55154 Revert "FP8 support on NVIDIA (#8631)" 
This reverts commit 2c8e4ea865.
 2025-04-09 12:27:41 +08:00
George Hotz
d1505137ad Revert "move TestOpsFp8s skipTest (#9797)" 
This reverts commit a3aaf92b21.
 2025-04-09 12:27:40 +08:00
George Hotz
14928fecff Revert "fix TF32 tensor core dropped in tc_sm89 (#9798)" 
This reverts commit 7c9a96824f.
 2025-04-09 12:27:39 +08:00
qazal
1ed4eae510
 		hotfix: don't add shape to SINK viz node (#9800) 2025-04-09 12:04:33 +08:00
chenyu
7c9a96824f
 		fix TF32 tensor core dropped in tc_sm89 (#9798) 
also add `SHOULD_USE_TC=1` to verify TC is applied in simple_matmul
 2025-04-08 23:20:50 -04:00
chenyu
a3aaf92b21
 		move TestOpsFp8s skipTest (#9797) 
so get_available_devices is not called when running other tests
 2025-04-08 22:44:07 -04:00
pkotzbach
2c8e4ea865
 		FP8 support on NVIDIA (#8631) 
* squashed fp8 commits

* tensorcore start

* minor changes

* pre-commit

* pylint

* Delete fp8mul.cu

* clean

* small bugfix

* fix test_dtype

* fix test_dtype_alu

* add EMULATE_CUDA_SM89

* fix ci

* fix test_linearizer

* fix test_linearizer

* fix swizzle

* add debug to simple_matmul

* fixed swizzle

* python emulator

* refactor python emulator

* setup fix

* numpy setup

* ml_dtypes only in emulate_cuda_sm89

* fix pylint

* fix tests

* fix mypy

* fix mypy

* fix ruff

* done python emulator

* add acc type

* tests

* mypy

* clean code

* add cuda tensor core tests to CI

* minor fix

* clean test_dtype.py

* clean cstyle.py

* clean test_ops.py

* fix test

* fix test

* whitespaces

* pylint

* pylint

* amd?

* amd?

* amd

* reduce lines

* mockgpu remove

* fix

* ruff

* ruff

* fix mypy

* ruff

* test only for cuda

* fixed formatting

* small fixes

* small fix

* least_upper_dtype if fp8s not supported

* log and reciprocal are supported for fp8s

* ops python fixes

* dtypes.fp8s use

* e4m3 + e5m2 result dtype test

* truncate linter fix

---------

Co-authored-by: pkotzbach <pawkotz@gmail.com>
Co-authored-by: George Hotz <72895+geohot@users.noreply.github.com>
Co-authored-by: chenyu <chenyu@fastmail.com>
 2025-04-08 21:54:04 -04:00
hooved
5d85765327
 		types for WebGPU runtime (#9791) 
* add type annotations to ops_webgpu

* rerun CI

* add types to some _run params
 2025-04-08 22:52:17 +03:00
chenyu
4c8582a7ce
 		pipe allow_test_size in _time_program (#9789) 
* pipe allow_test_size in _time_program

it was dropped long time ago and BEAM_ESTIMATE is doing nothing

* revert BEAM_ESTIMATE
 2025-04-08 09:07:40 -04:00
chenyu
8fe83385ec
 		add system json for mi300x mlperf (#9786) 
* add system json for mi300x mlperf

```
python3 -m mlperf_logging.system_desc_checker examples/mlperf/training_submission_v5.0/tinycorp/systems/tinybox_8xMI300X.json training 4.1.0
INFO -   System description checker passed for tinybox 8xMI300X
```

also removed the rocm from tinybox_red since we are not using it

* update mlperf-logging version
 2025-04-08 06:36:44 -04:00
chenyu
4a807ee952
 		remove duplicated z3-solver in setup.py (#9787) 2025-04-08 06:12:58 -04:00
qazal
21e872df44
 		remove consts from sched_sink [pr] (#9782) 2025-04-08 16:08:24 +08:00
qazal
f13e9cf2d9
 		move view_left to grouper.py + tiny reorders [pr] (#9780) 
* move view_left to grouper.py [pr]

* reorder grouper

* test_schedule
 2025-04-08 15:39:28 +08:00
chenyu
7a28133b37
 		failed test for single softmax backward (#9778) 
getting RecursionError with DONT_GROUP_REDUCES=1
 2025-04-08 02:36:32 -04:00
George Hotz
fefee5d3ab
 		single kernel softmax (#9776) 
* real single kernel softmax

* cleanup

* fix blockend insertion

* add to bert test
 2025-04-08 12:35:48 +08:00
qazal
9963bb51e0
 		grouper tests cleanups [pr] (#9777) 
* grouper tests cleanups [pr]

* viz

* tuple

* whitespace
 2025-04-08 12:33:11 +08:00
chenyu
4cc7422769
 		use AM driver in bert mlperf (#9775) 
we should commit to use AM. it's 7ms slower python time now
 2025-04-07 23:40:27 -04:00
George Hotz
db22094d35 hotfix: update softmax fusion test 2025-04-08 11:23:19 +08:00
Francis Lata
f8fe15e64e
 		move BoxCoder to mlperf helpers (#9773) 2025-04-07 20:27:06 -04:00
Eitan Turok
bb7922b95f
 		Vectorize Transcendental Regression Tests (#9753) 
* init test

* cleanup
 2025-04-08 01:27:39 +08:00
chenyu
7c4a739fe4
 		full script for bert mi300x (#9772) 2025-04-07 11:41:31 -04:00
Sieds Lykles
07d1aefaf4
 		fast idiv (#9755) 
* fast idiv with tests and fuzzer

* Add todo comment

* Add env variable to toggle fast_idiv

* Move env check

* Add fuzz fast_idiv to ci

---------

Co-authored-by: chenyu <chenyu@fastmail.com>
 2025-04-07 08:32:24 -04:00
nimlgen
fa888ee077
 		minor test cleanups (#9770) 
* fix test_graph on max

* pcie5
 2025-04-07 15:29:12 +03:00
chenyu
3069ebfad1
 		use BERT_LAYERS=2 in bert init (#9769) 
save 5 minut scheduling in setup so we can fit more search
 2025-04-07 07:46:37 -04:00
qazal
891322fd51
 		split into grouper.py (#9768) 
* split into grouper.py

* update tests

* reorder
 2025-04-07 18:40:59 +08:00
qazal
219b8c9e8b
 		return becomes_map in scheduler [pr] (#9766) 
* add a graph_rewrite pass for creating asts [pr]

* disk

* benchmark

* return becomes_map in scheduler

* reorder schedule.py into grouper and linearizer [pr]

* comments
 2025-04-07 17:36:23 +08:00