github/tinygrad
1
1
Fork 0
mirror of https://github.com/tinygrad/tinygrad.git synced 2026-01-10 07:28:15 -05:00
Code Issues Packages Projects Releases Wiki Activity
7,991 Commits 203 Branches 13 Tags
b4c3780df0cd729e2c8f0d95e53a7d8bb302bf79
Commit Graph

7991 Commits

This Branch
This Branch
All Branches
Author SHA1 Message Date
nimlgen
b4c3780df0 hotfix: interop example (#9237) 
* hotfix: interop example

* rm this

* fix

* fix ci mps

* atol rtol

* no uaf
 2025-02-25 10:32:00 +03:00
chenyu
8c7be428e5 update bert BS to 78 (#9236) 
fits 78 now. about 215 tflops on green
 2025-02-24 22:47:35 -05:00
Sieds Lykles
990c240b82 Stable pow gradient (#9226) 
* Stable gradient

* More efficient

* Fix and test for +-inf

* cleaner

* skip webgpu test

---------

Co-authored-by: chenyu <chenyu@fastmail.com>
 2025-02-24 20:54:26 -05:00
chenyu
731d14e718 hotfix bump testmetal2 timeout-minutes to 20 (#9235) 
setup is taking too long
 2025-02-24 20:23:56 -05:00
qazal
cbfe95d306 bring cast before view back (#9230) 
* bring cast before view back

* tune it to only trigger on expands

---------

Co-authored-by: chenyu <chenyu@fastmail.com>
 2025-02-25 01:50:39 +02:00
chenyu
90c3ed17c5 move cast to before softmax in attention (#9213) 
* move cast to before softmax in attention

saved some memory because exp (which is used for backward) are done in half. training bert seems fine and can fit BS=78 now (from 66)

* test
 2025-02-24 17:24:59 -05:00
geohotstan
f0b24d230c add test_onnx_ops.py (#8569) 
* boom

* fix webgpu

* use exact variable names in test so that AI can read easier

* add tag for specific test name like test a specific dtype

* fix ruff

* astype everything

* dtype in array creation

* just arange

* is 67% considered fixed?

* move test up

* small cleanups

* share function

* add qgemm as well

* add qgemm too

* make sure qgemm comes out as int

* take out qgemm for now

* fixed test

* add correct qgemm

* addressing feedback here too, early naive fix for now

* simplify bias and c to be minimalistic enough to test correctness

* refactored qlinearops

* maybe these asserts aren't the best..

* fix test

* updated tests to cover new ops

* try to add to CI

* move test_onnx_ops into testextra/

* more attention tests

* qlinear_add atol=1

* attention still not fullllllly correct

* it is what it is

---------

Co-authored-by: chenyu <chenyu@fastmail.com>
 2025-02-24 16:15:22 -05:00
nimlgen
56288243e6 metal PyTorch interop (#9229) 
* add from_blob support to mps cuda

* objc_id

* metal pytorch interop

* fix comments

---------

Co-authored-by: George Hotz <geohot@gmail.com>
 2025-02-24 22:36:08 +03:00
qazal
687d157906 delete cast early folding from ops [pr] (#9228) 2025-02-24 19:00:51 +01:00
George Hotz
c9493e41a6 reorder expand (#9051) 
* reorder expand

* symbolic ops needs resolve here

* s/arg/st + whitespace

* viz

---------

Co-authored-by: qazal <qazal.software@gmail.com>
 2025-02-24 13:55:47 +01:00
qazal
14aa2395d0 allow VIEW(BUFFER) in Tensor UOps [pr] (#9210) 
* allow VIEW(BUFFER) in Tensor UOps [pr]

* still reshapes

* update becomes_map tests

* bring copy folder to the scheduler

* lint

* only sgd left

* optimizer assign

* 13 kernels

* rename to test_reorder_expand + assert VIEW
 2025-02-24 13:06:15 +01:00
nimlgen
1d06d61b16 from_blob for cuda (#9223) 
* from_blob for cuda

* maybe docs?

* minor docs

* example

* waiting 9224

---------

Co-authored-by: George Hotz <72895+geohot@users.noreply.github.com>
 2025-02-24 14:02:06 +03:00
George Hotz
fc32ff80d6 torch and numpy dtype interop [pr] (#9224) 
* torch and numpy dtype interop [pr]

* less lines

* order
 2025-02-24 18:26:49 +08:00
George Hotz
24615db5f5 hotfix: torch cuda interop example 2025-02-24 09:02:48 +00:00
George Hotz
fd731e740a hotfix: add note on backend2.py 2025-02-24 11:23:03 +08:00
albanD
f2dd9c1562 simplify c++ code (#9221) 2025-02-24 11:04:41 +08:00
qazal
d12efc95d4 support custom name function in viz [pr] (#9219) 
* support custom name function in viz [pr]

* title case

* assert name count in test_track_rewrites_name_fxn
 2025-02-24 03:03:25 +02:00
chenyu
b3ae664d5d fix gradient of pow(t, int) (#9217) 
semi revert some pow logic back to tensor. added direct gradient check because the backward in test_ops passed by luck
 2025-02-23 17:42:09 -05:00
qazal
12b5b83821 set TRACK_MATCH_STATS=0 for real_strides [pr] (#9216) 2025-02-23 23:26:31 +02:00
qazal
9db0ec46a7 simpler buf_uop [pr] (#9215) 
* simpler buf_uop [pr]

* assert after realize it's buffer
 2025-02-23 19:23:14 +01:00
qazal
898aafe6fd move split_reduceop to scheduler + enable it for multi (#9214) 
* move split_reduceop to scheduler + enable it for multi

* merge r and _reduceop
 2025-02-23 17:30:04 +01:00
ShikChen
05e3202fba remove unused memsize_to_str and minor cleanups [pr] (#9211) 
* fix edge cases in memsize_to_str()

Inputs <= 1 now return "0.00 B" for 0 and "1.00 B" for 1, avoiding an
IndexError. Also, memsize_to_str(1000) now returns "1.00 KB" instead of
"1000.00 B".

Replaced the list comprehension with a next(...) generator for conciseness
and efficiency.

* simplify code using idiomatic python

- Remove the unused `memsize_to_str()` function in helpers.
- Use a tuple for checking multiple string prefixes/suffixes.
- Avoid unnecessary list construction by using iterables directly.
- Check None in @diskcache to ensure proper caching of falsy values.

* revert generators back to list comprehension

Sometimes building list first could be faster. Keep it as is.
 2025-02-23 09:58:37 -05:00
qazal
81a71ae0f6 hotfix: skip test_exclude_const_metadata (#9208) 2025-02-22 23:26:04 +02:00
chenyu
e0adb1fc76 really run test_ops with TINY_BACKEND in ci (#9206) 
was failing with `line 1: pytest: command not found`
 2025-02-22 15:51:24 -05:00
qazal
e6d20c47e3 simpler becomes_map update [pr] (#9201) 
* simpler becomes_map update

* err, no metadata for device

* simpler tensor metadata mapping + tests [pr]

* remove kernel metadata

* don't map nones

* pruning

* linter
 2025-02-22 20:50:58 +01:00
qazal
4578c3e8fd simpler tensor metadata mapping + tests [pr] (#9203) 
* simpler tensor metadata mapping + tests [pr]

* remove kernel metadata

* don't map nones
 2025-02-22 20:18:46 +01:00
qazal
b711c6343a no early return + allow childless const/bind/var in kernel graph [pr] (#9202) 2025-02-22 19:28:22 +01:00
George Hotz
97bc723538 torch backend works for ResNet-18 (#9200) 
* torch backend progress, a few more functions

* resnet works

* pillow

* tv
 2025-02-22 22:16:23 +08:00
George Hotz
f92820d30d torch backend tests (#9198) 
* torch backend tests

* pythonpath

* install ninja
 2025-02-22 16:01:49 +08:00
George Hotz
4e6665bda5 different way to write torch backend (#9197) 
* different way to write torch backend

* both backends

* more work

* simpler code

* more work

* test both

* imply unwrap/wrap

* FORWARD_ONLY=1 TINY_BACKEND=1 python3 test/test_ops.py TestOps.test_add works

* ready to start making test_ops work in torch backend

* backward pass, TINY_BACKEND=1 python3 test/test_ops.py TestOps.test_add works

* FORWARD_ONLY=1 TINY_BACKEND=1 python3 test/test_ops.py TestOps.test_simple_conv2d works

* matmul backward is broken with as_strided
 2025-02-22 14:42:26 +08:00
nimlgen
041b6d5678 am: load fw in batches (#9185) 
* am: load fw in batches

* am: 1mb less fw copies

* mypy

* list
 2025-02-21 23:21:31 +03:00
qazal
1db4341e9f move viz graph to lib/graph [pr] (#9196) 
* move viz graph to lib/graph [pr]

* add package

* share with program
 2025-02-21 21:04:07 +01:00
geohotstan
6587c7879b simple fixes to onnx (#9195) 
* uncontroversial changes

* cleaner _prepare_quantize
 2025-02-21 13:10:06 -05:00
Simon R
2318d7ac51 Add missing tinygrad.runtime.autogen.am to packages (#9194) 2025-02-21 15:38:24 +02:00
qazal
8bb80b6e5e reorder AST matchers + comments [pr] (#9193) 2025-02-21 14:31:15 +01:00
qazal
2eab8021fb remove inputs+outputs attributes from ScheduleItem [pr] (#9192) 
* remove inputs/outputs from ScheduleItem

* fix test_linearizer

* fix test_conv_shapetracker

* fix test_schedule + lint

* test_image_dtype + multitensor + search
 2025-02-21 13:48:11 +01:00
George Hotz
e87be0131e torch backend start (#9191) 
* start torch backend

* progress

* ugh, you need cpp crap

* 1+1 works

* 1+1 works

* becoming a real backend

* ready to merge?
 2025-02-21 16:57:28 +08:00
George Hotz
d3a21cced2 hotfix: bump version to 0.10.2 v0.10.2 2025-02-21 10:43:49 +08:00
chenyu
2e7c2780a9 CLANG -> CPU (#9189) 2025-02-20 18:03:09 -05:00
nimlgen
f986e12f91 metal: choose compile spec based on macos (#9188) 
* metal: choose compile spec based on macos

* correction
 2025-02-21 00:43:39 +03:00
chenyu
3e22747799 run unit test on windows ci (#9187) 
* factor out testing_minimal in setup.py [pr]

* testing_unit + windows
 2025-02-20 14:40:41 -05:00
chenyu
287de4ecc6 use torch in test_gradient (#9186) 
used torch.autograd.grad, but not sure if it can be a template like jax
 2025-02-20 12:26:11 -05:00
qazal
574a905291 Fix running VIZ=1 after package installation + test (#9183) 
* test running viz from pip install

* add pkg

* do 10 connection attempts

* include assets in package_data

* quiet curl

* better print
 2025-02-20 15:02:00 +01:00
chenyu
1692087db5 _one_hot_along_dim input needs to be int (#9179) 
* _one_hot_along_dim input needs to be int

indexing and onehot compare with arange, and non-int dtype is likely a bug
 2025-02-20 09:00:43 -05:00
George Hotz
bf36967883 cuda hooking (#9180) 
* cuda hooking

* progress

* more hook cuda

* fix params

* compile + cuMemHostAlloc hook

* work

* revert that
 2025-02-20 19:20:01 +08:00
chenyu
3b37cc898b add bert tiny config (#9177) 
set with BERT_SIZE=tiny. easier to study embedding and fusion
 2025-02-19 14:57:03 -05:00
qazal
5662c898f1 correctly step through bottom_up_rewrites in viz [pr] (#9176) 2025-02-19 19:20:57 +01:00
peppingdore
b1ddb2a1a6 fix win32 CPUProgram missing cache flush (#9171) 
* win32: fix missing inst cache flush, rename ptr->self.mem for consistency with posix code

* fix types, remove assert

* fix memory leak

* rm whitespace
 2025-02-19 21:38:51 +08:00
qazal
1bb9d78c7a hotfix: add output buffer back to kernel parents + comment [pr] (#9174) 2025-02-19 14:22:01 +01:00
chenyu
975c318dbc bert use int32 for input ids (#9173) 
original data was int32 for these. float might have caused precision issues
 2025-02-19 08:17:27 -05:00