Elias Wahl
bb248a0dd1
Optional half matmul ( #4835 )
...
* half linear
* move weight cast back
* oops
* matmul dtype var
* todo comment
2024-06-04 17:53:41 -04:00
Elias Wahl
04e237328b
Refactor to class style ( #4804 )
2024-06-04 14:08:31 -07:00
nimlgen
1b8bed4a26
nv check cmdq overrun ( #4824 )
...
* nv check cmdq overrun
* fix assert
2024-06-04 23:22:58 +03:00
David Hou
cddce0e168
don't cast before view on shape changing bitcast ( #4833 )
...
* don't cast before view on shape changing bitcast
* make sure cast before view triggers
2024-06-04 16:04:52 -04:00
Alec Chen
0c3a996e64
Nest ifs for dtype and uop in pattern matcher ( #4834 )
2024-06-04 15:51:28 -04:00
Alec Chen
4909a0d16f
Fix arg set in pattern matcher ( #4830 )
2024-06-04 15:10:09 -04:00
Alec Chen
c96026ac65
Add arg set regression test for pattern matcher ( #4827 )
...
* Add arg set regression test for pattern matcher
* real regression
---------
Co-authored-by: qazalin <qazal.software@gmail.com >
2024-06-04 13:35:09 -04:00
chenyu
a70e8a80d7
test_ops test cmp with special floats ( #4826 )
...
prepare to fix nan, it did not work with ge and le before either
2024-06-04 12:10:21 -04:00
Szymon Ożóg
b6895dabaa
Remove ssa label ( #4823 )
...
* remove ssa label
* linting
2024-06-04 16:51:05 +02:00
George Hotz
052c928d06
hotfix: touchups from presentation
2024-06-04 16:31:03 +02:00
chenyu
1e02b4cae1
default skip all exception in beam ( #4822 )
...
added a flag `BEAM_STRICT_MODE` to catch compile error or other exceptions on demand
2024-06-03 18:21:36 -04:00
chenyu
3afc914617
CMPEQ -> CMPNE and make it safe to pad ( #4818 )
...
* CMPNE
* new dataset
2024-06-03 18:02:15 -04:00
qazal
79c7d402ee
improve augmented assign error message ( #4813 )
2024-06-03 16:57:22 -04:00
Szymon Ożóg
bb7b031c5c
Bitshift ( #4728 )
...
* WIP
* Cleanup
* Cleanup
* Fix variable, refactor to use set
* right shift should be signed/unsigned
* Test for bitshifts
* Allow a neg
2024-06-03 21:16:01 +02:00
nimlgen
e78a9bf3f2
support view in nv/amd ( #4812 )
...
* support view in nv/amd
* fix amd
* fix
* run test on nv/amd
2024-06-03 22:11:52 +03:00
chenyu
45083ccb43
canonicalize 0 in shape in View.create ( #4815 )
...
set strides to 0, offset to 0, mask to None, and contiguous to True with size 0 view.
2024-06-03 13:37:37 -04:00
Szymon Ożóg
d064bf6d8c
b2 is useless ( #4814 )
2024-06-03 18:29:53 +02:00
nimlgen
65f0071c4b
amd compute queue bind api ( #4732 )
...
* amd hcq bind api
* revert copy queue
* revert
2024-06-03 18:36:56 +03:00
chenyu
3cc6ae0d85
layernorm backward is indepedent of its mean ( #4806 )
2024-06-03 09:49:59 -04:00
George Hotz
2dae657415
improve readability ( #4809 )
2024-06-03 14:57:57 +02:00
George Hotz
eecfdd2f6e
hotfix: fix dataset reading for new llm.c
2024-06-03 14:10:05 +02:00
qazal
6e0c16dfb0
cleanup render_reduceop ( #4807 )
...
* update acc key
* refactor return type
* remove return type
* run all reduces
* set acc key [run_process_replay]
* local_idxs are copied in render_reduceop [run_process_replay]
2024-06-03 14:39:02 +03:00
George Hotz
dd84f7d35e
touchup: show process name in multiprocess assert
2024-06-03 13:09:40 +02:00
qazal
0db9674dea
skip process replay on master ( #4808 )
2024-06-03 12:29:28 +03:00
qazal
f64fa51a64
process replay for test/* ( #4799 )
...
* add input to unit tests [run_process_replay]
* add setup [run_process_replay]
* run tests [run_process_replay]
* add cuda and amd [run_process_replay]
* run everything but BEAM=2 [run_process_replay]
* skip export_model [run_process_replay]
* fix amd CI
* add concurrency back
2024-06-03 12:01:58 +03:00
nimlgen
e8b5f2040d
nv faster signal on dma queue ( #4789 )
2024-06-02 21:47:24 +03:00
Francis Lata
707099487a
Multiprocessing UNet3D dataloader ( #4801 )
...
* testing dataloader
* matching dataloader implementation for unet3d
* remove comments
* clean up dataloader
* add cookie and cleanup
* use shm_path when creating SharedMemory
* add support for testing resnet and unet3d dataloaders
* update dataset test to return preprocesed data directory in prep for dataloader testing
* pass preprocessed dataset directory properly
* update loader function for dataloader
* add shuffling on indices
* update shm name
* more cleanup for unet3d dataloader
* remove changes to tests
---------
Co-authored-by: chenyu <chenyu@fastmail.com >
2024-06-02 11:30:47 -04:00
Timmy
ca32921f84
Multireduce PADTO Test ( #4785 )
...
* padto test
* expanded multireduce padto tests
* cuda doesnt run on ci
* moving padto_where_multireduce test to SUM so that we can check the reduce axis
* cleaning up tests some more
* add wanna_outputs
* refactor test_padto_sum_multireduce
* fix max and refactor where
* fix axis
---------
Co-authored-by: qazal <qazal.software@gmail.com >
2024-06-02 13:46:53 +03:00
qazal
231ed2c656
compute aliased buffer idxs pre reduce ( #4788 )
2024-06-01 16:46:52 -04:00
nimlgen
1b18ebb133
minor cleanups ( #4802 )
2024-06-01 20:11:43 +03:00
chenyu
1ffa5ec492
unit test ShapeTracker.consecutive ( #4800 )
2024-06-01 10:10:51 -04:00
nimlgen
7384ee08a0
amd cleanup sdma ( #4796 )
...
* amd cleanup sdma
* faster enqueue for sdma
* typo
* remove commnted lines
* fix overrun check
* flushhdp better command
2024-06-01 17:06:44 +03:00
qazal
240d6b5bc0
process replay benchmarks ( #4668 )
2024-06-01 14:36:21 +03:00
Alec Chen
b377db7f0d
Refactor UOps pattern matcher to UPat instead of dicts ( #4791 )
2024-06-01 10:55:51 +02:00
qazal
de8c8abbd8
define indexes pre reduce ( #4795 )
2024-05-31 18:53:27 -04:00
nimlgen
bd2e7c8b31
amd registers from file ( #4778 )
...
* amd registers from file
* remove commentes
* linetr
* no off
2024-05-31 18:48:57 +03:00
chenyu
8942230b1f
minor cleanups of test_tensor and extend some cases ( #4794 )
2024-05-31 10:43:22 -04:00
qazal
637f482588
configure derandomizing CI tests ( #4793 )
2024-05-31 17:06:58 +03:00
wozeparrot
ed0a740fe4
greater chat api endpoint compat ( #4792 )
2024-05-30 22:47:31 -07:00
chenyu
7cc883ecee
CMPLT is safe to pad ( #4790 )
...
0 < 0 evals to False
2024-05-30 22:50:48 -04:00
chenyu
236390aafb
fix lazy r const folding with variable shape ( #4783 )
...
currently not supporting const fold symbolic shape. I think it's possible with a refactor to Tensor.from_node.
also added some failed required tests for symbolic arange.
2024-05-30 15:19:28 -04:00
chenyu
c4d1283049
simplify _cumsum with _first_zero=True ( #4782 )
...
handled the case with 0 in shape output of _cumsum, and _cumsum returns the correct shape with _first_zero=True
2024-05-30 13:19:33 -04:00
chenyu
4921de1945
fix cumsum of 0-d tensor ( #4781 )
...
* fix cumsum of 0-d tensor
* _resolve_dim for all
2024-05-30 12:41:09 -04:00
chenyu
4cf0eadf8f
failed test case for ellipsis in einsum ( #4779 )
...
from #4156
2024-05-30 11:14:42 -04:00
Alec Chen
e89bc42cc7
Add UOps pattern matcher regression tests ( #4725 )
...
* add pattern matcher regression tests
* Remove test for dtype str after rebasing
* Make test uops match type spec
* leave const const, add const alu vin test
* correct uops
* actually correct uops
2024-05-30 17:12:20 +03:00
qazal
c2945be0a3
add fused tensor core opts tests ( #4775 )
...
* add fused tc opts tests
* n=64
2024-05-30 13:50:00 +03:00
chenyu
f1bf916b8a
apply NOOPT in test_arange complexity ( #4774 )
...
with hcopt, arange(2560) uses less ops than arange(256)
2024-05-29 23:12:35 -04:00
chenyu
cde7a7cda7
isolate the 134ms kernel in train_gpt2.py ( #4773 )
...
133ms on tinybox red with BEAM=2
2024-05-29 17:26:24 -04:00
nimlgen
57204c4014
amd cleanup pm4 queue ( #4772 )
2024-05-29 22:59:06 +03:00
lopusz
b2c408912c
Add docs link to README ( #4768 )
2024-05-29 17:47:47 +00:00
