github/tinygrad
1
1
Fork 0
mirror of https://github.com/tinygrad/tinygrad.git synced 2026-01-09 15:08:02 -05:00
Code Issues Packages Projects Releases Wiki Activity
4,766 Commits 203 Branches 13 Tags
7c3b877216cbe91867fa5111d4197723e168892c
Commit Graph

4766 Commits

This Branch
This Branch
All Branches
Author SHA1 Message Date
kormann
7c3b877216 rename uop [run_process_replay] (#5031) 
* rename

* fix unittests

* rename vin

* fix test

* fix type [run_process_replay]

* rm pre commit hook change
 2024-06-18 21:34:05 +03:00
chenyu
dc942bf1f6 jit sampling functionn in test_randomness.test_multinomial (#5034) 
* jit sampling functionn in test_randomness.test_multinomial

`THREEFRY=1 python3 -m pytest test/test_randomness.py::TestRandomness::test_multinomial --durations 1` 7 sec -> 1.2 sec

* skip that
 2024-06-18 14:21:05 -04:00
Elias Wahl
f31ef11537 Better default hparams for large BS (#5030) 
* better default hparams for large BS

* bf16 too

* use tuple
 2024-06-18 11:13:06 -04:00
Francis Lam
8d33998e0d [run_process_replay] linearizer: fix get_grouping_dims to respect global/local max (#4855) 
* linearizer: fix get_grouping_dims to respect global/local max

* fix lidx variable index offset and unrestrict clang/llvm global len

* test reverse variable indexing when reverse_dims is true

* change the collapse axis to be the right most if reversed
 2024-06-18 16:51:27 +03:00
joeshmoe0112358
7842559952 simplification of exp2 (#5023) 2024-06-18 06:51:16 -07:00
kormann
acc8f5e30e print_tree for uops (#5028) 2024-06-18 06:36:14 -07:00
Junjun Dong
c8cd6e725c Remove BinaryOps.SUB. Replace SUB by ADD and NEG in all tests. Regenerate dataset (#4977) 
* feat: remove BinaryOps.SUB

* remove SUB in test_early_end_local

* regenerate dataset. remove SUB in test_linearizer_*

* reenable overflow tests

* simplify tensor.sub function by returning a+(-b)

* remove whitespaces

---------

Co-authored-by: chenyu <chenyu@fastmail.com>
 2024-06-18 09:06:13 -04:00
chenyu
620fa6e5a2 check Tensor.reshape can have at most one -1 (#5026) 
raise RuntimeError to match torch. on master it throws weird errors from shapetracker
 2024-06-18 08:17:12 -04:00
Elias Wahl
7bfa9101c0 Float in scaled dot product attention (#4985) 
* Monkeypatch scaled-dot-product-attention

* Use dot instead of matmul

* new api

* imports

* least_upper_dtype
 2024-06-18 08:16:41 -04:00
nimlgen
194a168630 hcq signal scheduler (#5016) 
* faster hcq

* fix nv

* linter

* cleaner

* fix sync

* cleaner

* a bit cleaner
 2024-06-18 14:02:21 +03:00
chenyu
e9c6a36894 remove CACHELEVEL=0 in llama3 benchmark (#5025) 2024-06-17 22:43:16 -04:00
chenyu
acaf9a490d RECIP(-0.0) should be -inf (#5024) 
* RECIP(-0.0) should be -inf

added test_dtype_alu for PYTHON backend

* catcht that

* fix those two
 2024-06-17 22:26:58 -04:00
GabrielZCode
66760ae558 graph display floats rounded (#5021) 
Co-authored-by: gabrielsouza <gabriel.martins@perdcomp.com.br>
 2024-06-17 18:22:55 -07:00
chenyu
03b367c014 handle float16 overflow in PYTHON (#5022) 
* handle float16 overflow in PYTHON

use `truncate` when constructing tensor from list to make sure all values are packable (might be slow, but should be correct). add truncate_fp16 to cast overflowed values to inf/-inf.

* all valid fmt supports truncate
 2024-06-17 21:12:52 -04:00
chenyu
c0139b05d8 python_alu sin(inf) is nan (#5020) 
* python_alu sin(inf) is nan

without special handling, it throws ValueError: math domain error

* skip CUDACPU
 2024-06-17 19:47:30 -04:00
chenyu
4296507021 Tensor.sum returns in acc_dtype if specified (#5012) 
* Tensor.sum returns in acc_dtype if specified

* skip PYTHON for now

* revert that

* relax that
 2024-06-17 16:35:52 -04:00
chenyu
013c73c3b3 minor refactor overflow handing in python backend (#5015) 
made it clear that it's only handing int now. need to handle float inf next
 2024-06-17 12:18:38 -04:00
Ray
1ad3b25461 fix einsum output str (#4998) 
* fix einsum output str

* new line to satisfy linter

* removed redundant cast (satisfy linter)
 2024-06-17 12:18:14 -04:00
nimlgen
794acefbf3 hcq update waits and signals in place (#4984) 
* hcq update waits and signals in place

* start amd

* amd works

* prettier

* test

* normal messages

* linetr

* linter 2
 2024-06-17 17:19:07 +03:00
qazal
603a4a0ce1 process replay contributor docs (#5010) 2024-06-17 09:38:59 -04:00
qazal
026c59543c allow keyword args in UOp.store [run_process_replay] (#5008) 
* allow keyword args in UOp.store [run_process_replay]

* same for load

* typing can stay
 2024-06-17 15:42:27 +03:00
uuuvn
f1de8cd8cf Convert a bunch more rules [run_process_replay] (#5007) 
* Convert a bunch more rules [run_process_replay]

* more rules, narrow down CMPLT rule

* smart linter cut two lines

* nope, the linter is dumb

* make dumb linter shut up

* revert two rules

* Revert "revert two rules"

This reverts commit 585688da17.

* fix
 2024-06-17 15:16:31 +03:00
chenyu
c52352bd9a fix yolov8 example (#5003) 
it was creating Tensor from a list of numpy arrays, which is not supported after moving creating from a list not using numpy.
 2024-06-16 20:47:29 -04:00
nimlgen
8bc0cbf67b nv tiny cleanups (#5001) 
* nv tiny cleanups

* gpfifo rework

* return type
 2024-06-17 00:43:44 +03:00
qazal
04feeb37e6 look for unsafe pad ops in multiview ShapeTrackers (#5002) 2024-06-17 00:28:12 +03:00
George Hotz
bee8fc29ee add GPT2 half/half+beam to AMD (#5000) 
* add GPT2 half/half+beam to AMD

* winograd in training. half and half/beam file upload
 2024-06-16 14:07:14 -07:00
chenyu
72c9b22833 sort vars in jit when building expected input args (#4990) 
* sort vars in jit when building expected input args

fixed symbolic jit bugs with two variables.

* sort in clanggraph

* space

* one more
 2024-06-16 15:55:51 -04:00
qazal
71aad183fd check Program from HEAD [run_process_replay] (#4996) 
* use the same prg [run_process_replay]

* put var back
 2024-06-16 20:12:30 +03:00
chenyu
2b07847f2b matmul returns in acc_dtype if specified (#4994) 
more flexible to not automatically downcast, can fix bert mixed precision training with this
 2024-06-16 12:56:15 -04:00
George Hotz
1d6f1a15e1 add lt and ge uop methods [run_process_replay] (#4995) 
* add lt and ge uop methods [run_process_replay]

* more correct (should still run process replay)
 2024-06-16 09:33:53 -07:00
uuuvn
1b3f27565a Boring UOps to UPat compiler [run_process_replay] (#4991) 
* Boring UOps to UPat compiler

* ruff

* weirdness

* dtype fix

* Revert "weirdness"

This reverts commit 4bc213a157.

* weirdness

* end weirdness?

* a bunch more rules

* more patterns
 2024-06-16 09:03:41 -07:00
George Hotz
dac96f177e ignore indexing in the flopcounter (#4993) 2024-06-16 08:59:55 -07:00
Timmy
01b26756d6 Multireduce Scheduler Tests (#4972) 
* scheduler tests

* linters

* cleaning up tests

* fixing tests

* syntax

* fixing metal
 2024-06-16 16:30:22 +03:00
chenyu
5eb8001514 minor cleanup in jit (#4989) 
found a non-deterministic bug in jit with multiple variables. but first cleanup some variable names.
[run_process_replay]
 2024-06-15 23:43:17 -04:00
chenyu
44dfa37c70 use threefry in stable diffusion benchmark (#4988) 
also updated default steps to 10. easier to tell the image is following the prompt.
 2024-06-15 20:25:29 -04:00
chenyu
20b50d8d64 doc: manual_seed (#4987) 
there was a docstring just not linked to the doc page. also updated the example to show re-seed instead of a internal variable
 2024-06-15 19:57:26 -04:00
wozeparrot
ce1ed374c9 more tinychat fixes (#4971) 2024-06-15 16:29:39 -07:00
chenyu
50bc14d186 re-enable test that loads torch pkl format (#4986) 2024-06-15 14:11:30 -04:00
qazal
ff8e9eefc3 hotfix: don't use ASSERT_COMPILE for benchmarks process replay (#4981) 
* use replay_codegen [run_process_replay]

* disable for now [run_process_replay]
 2024-06-15 16:57:47 +03:00
uuuvn
92f49efd06 Trigger process replay from pull request title [run_process_replay] (#4980) 
* Trigger process replay from pull request title

* idk how this thing works btw

* test if it will work

* try 2

* Revert "idk how this thing works btw"

This reverts commit 580da51b07.

* Revert "try 2"

This reverts commit 7ff1e86d5d.

* test if it works

* meh

* Reapply "idk how this thing works btw"

This reverts commit dd33ad7c14.

* revert
 2024-06-15 16:21:00 +03:00
uuuvn
033fb53f9e Incomplete/buggy rule breaks process replay on #4976 (#4978) 
* Incomplete/buggy rule breaks process replay on #4976

* test passes

---------

Co-authored-by: qazal <qazal.software@gmail.com>
 2024-06-15 15:18:35 +03:00
qazal
d91f0ee85b add regression test for the neg folding pattern (#4979) 2024-06-15 15:08:28 +03:00
nimlgen
dfadf82e10 hcq optimize enqueue time (#4973) 
* hcq optimize enqueue time

* linter
 2024-06-15 10:47:25 +03:00
chenyu
5f7dd74655 docs: update wording for unflatten (#4974) 
it was using `Expands`, the same in torch doc, but we also have expand so it's confusing
 2024-06-14 23:12:41 -04:00
Cyril Roumégous
efbf4fca05 perf: graph_rewrite line reduction and make it a little bit faster [run_process_replay] (#4958) 2024-06-14 16:37:27 -07:00
wozeparrot
8209cd3c55 easier llama3 + fetch subdir (#4938) 2024-06-14 13:47:27 -07:00
chenyu
64cda3c481 raise TypeError calling len() on a 0-d tensor (#4970) 
matched numpy and torch
 2024-06-14 16:34:27 -04:00
chenyu
67e8df4969 remove numpy from dtype (#4969) 
replaced all dtype.np with _to_np_dtype defined in tensor.py.

after this, the only numpy usages are (1) Tensor(np.ndarray), (2) construct .numpy() output, (3) numpy random buffer
 2024-06-14 15:38:45 -04:00
wozeparrot
62dc36d371 autogen _try_dlopen (#4949) 2024-06-14 12:12:18 -07:00
qazal
3e297d8216 delete Linearizer.const [run_process_replay] (#4967) 2024-06-14 21:51:37 +03:00