github/tinygrad
1
1
Fork 0
mirror of https://github.com/tinygrad/tinygrad.git synced 2026-02-12 23:54:58 -05:00
Code Issues Packages Projects Releases Wiki Activity
7,810 Commits 221 Branches 14 Tags
cd77e518108eede2d91caedaf233690f36731c80
Commit Graph

280 Commits

Author SHA1 Message Date
qazal
cd77e51810 fix tensor realization bug in #8975 (#8984) 
* fix tensor realization bug in #8975

* that's a reshape now

* work

* works

* give those tests better names

* test when multiple mops result in the same ShapeTracker

* test_become_existing_buf_complex is enough

* that too
 2025-02-10 13:51:30 +01:00
qazal
fd9f9ec772 realized base tensors become RESHAPE(BUFFER) [pr] (#8994) 2025-02-10 10:17:54 +01:00
qazal
7eba5fb413 Tensor.empty is RESHAPE(BUFFER) (#8987) 
* empty is RESHAPE(BUFFER)

* eh

* add test_empty_buf

* can we unsupport this

* linter

* Revert "can we unsupport this"

This reverts commit 0f71e1aadb.
 2025-02-09 18:42:51 +01:00
qazal
55351ebb31 minimal failing test for #8975 [pr] (#8982) 2025-02-09 14:10:37 +01:00
chenyu
cfd28517df move pow folding tests to test_schedule [pr] (#8955) 
not really belongs to test_const_folding
 2025-02-07 12:51:43 -05:00
chenyu
488200f16c move more pow const to rewrite (#8916) 
* move more pow const to rewrite

one less use of _to_const_val

* fix
 2025-02-05 20:30:12 -05:00
qazal
af4f9d1aa9 use matchers to verify AST shape [pr] (#8828) 
* use matchers to verify kernel AST [pr]

* work

* use swizzle_cnt

* add comment

* imports

* modified_ast comment

* brief
 2025-01-31 09:17:42 +02:00
George Hotz
643c09a6c6 tensor uop spec should be in spec.py [pr] (#8827) 
* tensor uop spec should be in spec.py [pr]

* err, spec.py

* print uops can stay
 2025-01-31 13:54:04 +08:00
qazal
a78f0f85d3 remove support for checking tensor uops in FUSE_ARANGE [pr] (#8829) 2025-01-31 07:48:28 +02:00
qazal
1fce864a6d delete multi output support (#8822) 
* delete multioutput for now

* test_schedule

* test_assign too

* linter

* 515 for sd

* update tests and ctx

* update that assign check
 2025-01-30 22:45:50 -05:00
qazal
530961f7d5 realized only exists on base (#8815) 
* realized only exists on base [pr]

* shorter

* update that too
 2025-01-30 23:02:25 +02:00
qazal
5643429c17 give BUFFER UOp a ShapeTracker [pr] (#8811) 
* give BUFFER UOp a ShapeTracker [pr]

* move that

* update contiguous

* test_advancedindex should use movement ops
 2025-01-30 22:33:32 +02:00
qazal
ba17786068 do not construct unmasked VALID (#8759) 
* new lines that exist in codegen/ops

* update tests

* update sops.gz (13071 -> 13070 asts)

* fix viz too

* remove that TODO

* diff pruning

* mask assert + device

* work

* diff pruning

* re: fix viz too

---------

Co-authored-by: George Hotz <72895+geohot@users.noreply.github.com>
 2025-01-28 20:51:21 +02:00
qazal
3417bc1814 fix ShapeTracker spec for const [pr] (#8791) 2025-01-28 19:53:36 +02:00
George Hotz
96bff0b4f7 contiguous is no longer needed in SGD [pr] (#8760) 
* contiguous is no longer needed in SGD [pr]

* add allow condition
 2025-01-27 15:19:11 +09:00
qazal
ac70f63d4b tensor_map cleanups [pr] (#8754) 
* tensor_map cleanups [pr]

* update test_schedule too
 2025-01-26 11:41:54 +02:00
George Hotz
b4bf6a7dea switch backward to use gradient [pr] (#8235) 
* switch backward to use gradient [pr]

* set device correctly, dedup

* why does that fail?

* add noop cast

* simple backward

* fix beautiful_mnist

* touchups

* set in compute_gradient

* uop_count

* uop_count was wrong

* collections

* no note

* skip that test

* update sched kernel counts

* train mnist is 65

* fix metadata and gc

* fixes

* materialize_grads

* no pathlib stuff

* add contiguous_backward, fix bugs

* add some realize

* fix multi
 2025-01-26 09:12:16 +09:00
qazal
8e5bd0cd7a fix buffer init and skip test_swizzle_failure_permute [pr] (#8732) 
* fix buffer init and skip test_swizzle_failure_permute [pr]

* replace preload with just load

* add
 2025-01-23 17:21:38 +02:00
qazal
07ec99001a keep VIEW in big_sink + copy of buffer view spec [pr] (#8727) 
* keep views in sink [pr]

* tests

* things from the gpt2 bug
 2025-01-23 11:29:30 +02:00
qazal
e3d1464ba4 move assign preload out of schedule item [pr] (#8710) 
* move assign preload out of schedule item [pr]

* fix that
 2025-01-22 12:43:57 +02:00
qazal
d6bf1feaab remove the "no copy" line from copy_to_device (#8702) 
* delete the no copy one

* add tests
 2025-01-21 17:09:33 +02:00
qazal
f0d424ecdf Tensor UOps can become a buffer or const after scheduling (#8698) 
* spec

* work

* update test_viewed_consts_do_not_realize

* remove
 2025-01-21 12:33:19 +02:00
qazal
e2008c98c3 allow symbolic shape in tensor const parents [pr] (#8699) 2025-01-21 12:01:25 +02:00
qazal
66ac0087e8 more high level contiguous tests + scheduler deletions [pr] (#8695) 
* delete those

* move the upat too

* rename ops_folding to just sym

* keep that
 2025-01-21 01:52:58 +02:00
qazal
08eb1f1f56 simplify tensors before scheduling [pr] (#8580) 
* delete forced_realize

* put that back

* work

* remove forced_realize

* expectedFailures

* contiguous(buffer)

* multi

* expectedFailures

* cleaner create_subbuffer

* more comments

* remove that

* note

* realizes

* work

* one upat and image is back

* remove

* cleaner

* fix test_complex_backward for now

---------

Co-authored-by: George Hotz <geohot@gmail.com>
 2025-01-20 23:42:42 +02:00
chenyu
679b1ad058 move softmax upcast to after subtracting max (#8684) 
* move softmax upcast to after subtracting max

max can always be done in the same dtype without any numerical loss, so this is better when explicitly upcasting in softmax

* skipUnless half
 2025-01-20 12:16:32 -05:00
qazal
9e55495b4d fold double contiguous [pr] (#8687) 2025-01-20 14:38:33 +02:00
qazal
ed63ff2372 Remove contiguous on buffer (#8676) 
* remove contiguous on buffer

* spec

* make things that can't be images not images
 2025-01-20 13:48:33 +02:00
George Hotz
168c16646a change create_schedule_with_vars api to big_sink [pr] (#8677) 2025-01-19 13:30:26 -08:00
chenyu
beba490ba8 update mask in scaled_dot_product_attention (#8674) 
built is_causal mask with ones_like and start with boolean, and reversed the mask -inf order
 2025-01-19 15:19:23 -05:00
chenyu
5842ee56c6 raise if attn_mask is set when is_causal=True in sdpa [pr] (#8675) 
matches torch, also fixed incorrect usage in tests
 2025-01-19 12:55:04 -05:00
qazal
2faf8774fe replace DEVICE of CONST after copy folding (#8673) 2025-01-19 11:33:39 -05:00
qazal
d957a4f108 add tests for div buffer collapsing in the scheduler [pr] (#8671) 
* add tests for mul/div buffer collapsing in the scheduler [pr]

* lint

* merge with test_linearizer's version of this

* 4*3
 2025-01-18 14:15:29 -05:00
qazal
2b7db9b45d delete unused cast/bitcast lines from ops.py [pr] (#8651) 
* move cast and bitcast out

* more deletion of bitcast arg

* fix test_bitcast_fuses

* update tests

* work
 2025-01-17 03:04:18 -05:00
qazal
81a84aa85a remove is_unrealized_unmasked_const [pr] (#8644) 2025-01-16 05:27:47 -05:00
qazal
a1f70ce7d0 only use BUFFER_VIEW in disk [pr] (#8629) 
* only use BUFFER_VIEW in disk [pr]

* delete can_view

* BUFFER_VIEW op on DISK

* remove that allow_buffer_view=False

* notes

* bitcast is a low-level op too

* this passes on AMD and LLVM
 2025-01-15 12:34:15 -05:00
George Hotz
504ad08e73 hotfix: add test_example_matmul_same 2025-01-14 19:03:17 -08:00
George Hotz
bfbe81df71 remove cast before view (#8613) 
* remove cast before view

* greener

* indexing

* that passes too

* openpilot too

* ack

---------

Co-authored-by: qazal <qazal.software@gmail.com>
 2025-01-14 15:04:58 -05:00
qazal
ae2229d727 assert kernel buffer limit at compile time [pr] (#8595) 
* remove the BUF_LIMIT assert

* skip the base one
 2025-01-13 16:32:07 -05:00
qazal
7562cc0399 better test for reduce swizzle + don't use double dtype [pr] (#8586) 
* better test_permute_rewrite

* use float32
 2025-01-13 05:02:21 -05:00
qazal
cff1ee9038 add SINK folding from the tensor_map branch [pr] (#8562) 
* delete is_constant from the scheduler

* add sink folding

* always give BUFFER uops Buffers [pr]

* spec for view, var (bind) and const

* add test_buffer_only_after_realize

* work

* 3 lines

* more work
 2025-01-12 03:39:34 -05:00
qazal
87cbff3ac0 always give BUFFER uops Buffers [pr] (#8572) 
* always give BUFFER uops Buffers [pr]

* add test_buffer_only_after_realize
 2025-01-11 23:17:09 +02:00
chenyu
d09897c2aa allow double copy [pr] (#8559) 
fixed ring allreduce pattern and recovered most of the bert step time regression (10% faster), will double check all benchmark
 2025-01-10 18:21:01 -05:00
qazal
2fd068ffc0 delete empty op (#8544) 
* simple delete EMPTY op

* there's no schedule for empty
 2025-01-09 14:10:15 -05:00
qazal
f6eb0574f2 start tests for putting the tensor graph in a single kernel [pr] (#8542) 
* start tests for putting the tensor graph in a single kernel [pr]

* parallel actually

* better view_left test

* test a softmax

* put all that in sym
 2025-01-09 13:33:21 -05:00
qazal
947de23cac add VIEW(DEVICE) to tensor variable [pr] (#8529) 
* add VIEW(DEVICE) to tensor variable [pr]

* bind 2

* restrict shapetracker

* move var and bind closer

* one less line
 2025-01-08 01:39:42 -05:00
qazal
b22494b710 restrict tensor const ShapeTracker in spec [pr] (#8447) 
* restrict tensor const ShapeTracker in spec [pr]

* pass sink srcs

* reject if any of the specs disagree

* deceive mypy

* viz

* default to float

* just check the view

* create_schedule is gone

* test_verify_arg is flaky
 2025-01-07 19:05:11 -05:00
qazal
0e97f807e0 test fixup prereqs for delete_buffer_view [pr] (#8523) 2025-01-07 11:52:18 +02:00
qazal
ed618a72e7 do not use subbuffer for bitcast (#8514) 
* do not use subbuffer for bitcast

* edit that test

* explicit test for ptx

* ptx
 2025-01-06 18:40:46 +02:00
qazal
ed121d235c spec for CAST_BEFORE_VIEW=1 [pr] (#8512) 2025-01-06 10:43:58 +02:00