bf3583f9b2
use Buffer.ensure_allocated in search _ensure_buffer_alloc ( #4132 )
2024-04-10 13:11:50 -04:00
a35375df85
run_schedule is so simple now ( #4130 )
2024-04-10 09:49:30 -07:00
86bd2eb500
hotfix: update copy_from_fd for new DiskBuffer
2024-04-10 15:41:06 +00:00
ee457a4b20
no more underlying diskbuffer, that's just the device ( #4129 )
2024-04-10 08:32:25 -07:00
fe88591890
update onnx to 1.16.0 ( #4127 )
...
* update
* pass tests and skip tests
2024-04-10 11:19:13 -04:00
6bbbeb93ac
skip a few clang test that took > 30 seconds in CI ( #4126 )
...
* skip slow CLANG test test_train_cifar
* skip those too
* and that
* only CI
* one more
2024-04-10 02:00:34 -04:00
08ddeb5685
create schedule has global vars ( #4125 )
...
* abstractions3 is currently wishful thinking
* create_schedule_with_vars
2024-04-09 21:42:16 -07:00
216eb235e5
hotfix: cast mnist to float
2024-04-09 19:30:03 -07:00
fea774f669
spend 5 lines to bring mnist into the repo ( #4122 )
2024-04-09 19:24:57 -07:00
42edae8935
pickle schedules ( #4114 )
...
* pickle schedules
* Update test_pickle.py
* Update test_pickle.py
---------
Co-authored-by: George Hotz <72895+geohot@users.noreply.github.com>
2024-04-09 13:47:25 -07:00
38ae4194a6
Fixes for ops_kfd ( #4105 )
...
* kfd_ops: Fix GPU node discovery on NUMA systems
Ignore potentially multiple CPU NUMA nodes and any GPU nodes that are
not accessible because of device cgroups.
Signed-off-by: Felix Kuehling <felix.kuehling@amd.com>
* kfd_ops: Format the GFX arch target name correctly
The target version in sysfs properties is a decimal representation with
two digits per component.
The format for LLVM GFX target names is a bit quirky for historical
reasons. It uses one digit for the minor version and stepping. When it
ran out of decimal digits for the stepping on gfx90X it started using
hexadecimal there. But the major version is still decimal and went
double digit in GFX10.
Make sure to parse and format it accordingly for all supported GPUs.
Signed-off-by: Felix Kuehling <felix.kuehling@amd.com>
---------
Signed-off-by: Felix Kuehling <felix.kuehling@amd.com>
2024-04-09 13:21:21 -07:00
10dbf90b2c
hotfix: test speed
2024-04-09 13:20:39 -07:00
ae849d12d7
numpy device + pickle it ( #4120 )
2024-04-09 13:19:30 -07:00
1ef9c50fd7
Update ssa input order and annotate types in cstyle and assembly ( #4117 )
...
variable prefix is never optional (removed the default "t") and UOp can be optional (added the default None).
2024-04-09 13:10:29 -04:00
15f2f39658
conceptually simpler fancy index ( #3335 )
...
* init
* add failed case
* fix: temp comment out MULACC cast
* is this right?
* add test case
* oops, forgot to get rid of temp test
* WOOOOOO TOOK OUT 2 TRANSPOSES IN GATHER YAY
* cleaner
* comment cleanup
* update docs
* resolve conflict
* oops
* SUPA FAST
* comment out a test
* del some print statements
* use new broadcast stuff
* more clean up
* move try except
* skip fancy indexing for python backend test_ops
2024-04-09 11:18:04 -04:00
980124a605
add lerp operation to tensor ( #4102 )
...
* feat: add lerp operation to tensor
* fix
* style: fit in one line:
* tests: test backward for lerp
2024-04-08 17:03:27 -07:00
46850a0269
search: add a BEAM_COMPARE env to optionally not compare to hc/tc ( #4107 )
...
* search: add a BEAM_COMPARE env to optionally not compare to hc/tc
setting BEAM_COMPARE=0 will prevent additional memory allocation
needed to do the timing tests assuming the BEAM result is in
the diskcache.
* change to optionally use Buffer.allocate
2024-04-08 18:54:01 -04:00
c390828f61
refactor outbufs ( #4112 )
2024-04-08 14:54:10 -07:00
7fd12aba85
graph remove input buffer references ( #4100 )
...
Co-authored-by: chenyu <chenyu@fastmail.com>
2024-04-08 16:49:16 -04:00
078d841479
add SPLIT_REDUCEOP to disable reduce split ( #4115 )
...
verify with `SPLIT_REDUCEOP=0 BIG=2 MPS=1 python3 -m pytest -rA test/test_speed_v_torch.py -k sum`. 10X slower on mac
2024-04-08 16:31:08 -04:00
eea42d864f
account for all outputs ( #4113 )
2024-04-08 10:04:19 -07:00
dbd39ab78a
setitem support setting python const ( #4111 )
2024-04-08 11:37:50 -04:00
f8dc82a8a7
use single tensor for llama kv chache ( #4108 )
...
similar to optimization in gpt2
2024-04-08 00:38:32 -04:00
92c0675ccf
setitem initial support ( #4093 )
...
* wip setitem
it's an eager assign to output shapetracker view
* cleanups and tests
* more cleanups
2024-04-07 20:35:22 -04:00
183708b3fd
broadcast expand to match torch ( #4085 )
...
* initial version
* heh gimme grrrreen
* version 2
* clean ups
* some test confusion
* fix onnx
* rename to _broadcast_tensors
* improved errors and test
* fixed?
* some test fixup
* version 3 lol
* comments
* cleaner
* add failure test for expand to 0 test
* 1 more assertRaises test
* make err msg better
* also rewrite the expand onnx op? :s
2024-04-07 16:23:13 -04:00
2b81d9b334
Fix broken test ( #4104 )
2024-04-07 12:02:12 -04:00
9a95d87366
metal CI run llama with 4 shards ( #4103 )
...
this can catch multi tensor issue on mac.
2024-04-07 11:04:08 -04:00
444d2a7487
hotfix: fix SDMA read_pointer_address in KFD
2024-04-07 13:13:15 +00:00
bb7567b365
Fix metal ( #4101 )
2024-04-07 05:21:19 -07:00
bdbcac67f1
assign jit test case with other tensor as input ( #4098 )
...
hmm it works
2024-04-06 14:41:14 -04:00
e4a1858471
revert command queue ( #4097 )
2024-04-06 08:58:18 -07:00
97c402d69e
use imagenet spawn ( #4096 )
2024-04-06 08:34:10 -07:00
fffd9b05f5
mock mnist data for imagenet trainer ( #4095 )
...
* mock mnist data for imagenet
* move print and test
* needed to reshape
2024-04-06 08:08:40 -07:00
8739d33fe9
kfd: disable copy_from_fd while debugging ( #4091 )
...
* kfd: disable copy_from_fd while debugging
* increase timeout to a minute
2024-04-05 18:02:58 -07:00
93824e59eb
support MOCKDATA=1 for resnet ( #4090 )
...
* mockdata for resnet
* fix eval, revert hsa
2024-04-05 17:19:18 -07:00
164329a8ea
address kfd feedback ( #4087 )
...
* address kfd feedback
* signals cleanup
* signals cleanup
* handle 2 doorbell pages correctly
* signal reset cleanup
* signals cleanup
* more GTT
* cleanups
* minor cleanups
2024-04-05 15:24:41 -07:00
dafa42e864
clean up ( #4081 )
...
Co-authored-by: chenyu <chenyu@fastmail.com>
2024-04-05 11:57:44 -04:00
750ecf8fef
replace slice by pad/shrink in _pool ( #4082 )
2024-04-05 11:47:22 -04:00
a337922c44
more work on kfd ( #4079 )
...
* more work on kfd
* fix multitensor test on kfd
* stuff
2024-04-05 08:36:36 -07:00
e7ff5102cf
failed test in test_pattern_matcher ( #4080 )
...
something about the PTX rewrite is incorrect that it has duplicated rewritten uops
2024-04-05 02:53:50 -04:00
a023a1ed87
update github action to actions/cache@v4 ( #4077 )
...
get rid of warning `Node.js 16 actions are deprecated. Please update the following actions to use Node.js 20: actions/cache@v3.`
2024-04-04 22:24:26 -04:00
28ec6c67be
hotfix: hlb_cifar KFD works
2024-04-05 02:19:14 +00:00
1de9778949
import Buffer and BufferOption from tinygrad.buffer ( #4076 )
2024-04-04 22:12:23 -04:00
9e0ebf8979
remove dtype from FlopCounter ( #4075 )
...
the annoying thing to remove all FlopCounter is that for device that does not support local, matmul index alu is huge.
we can remove the dtype first.
sneak in updating `ruff` command to `ruff check`
2024-04-04 21:23:28 -04:00
3de855ea50
don't use SVM memory in KFD ( #4072 )
...
* don't use SVM memory in KFD
* copy from fd
* cleanups
* transfer
* hacks
* ops_hsa
* tighter API
2024-04-04 17:33:21 -07:00
5e6e6c9a67
use ConstType in various const function type hint ( #4074 )
2024-04-04 20:32:07 -04:00
c1cffed1df
add LazyOp.dtype ( #4073 )
...
an inferred cached_property.
removed all cases that use get_lazyop_info just to get the dtype of an op.
prereq to remove InterpretedFlopCounter
2024-04-04 17:38:19 -04:00
f836d6a03f
is_unrealized_unpadded_const -> is_unrealized_unmasked_const ( #4071 )
...
realized #3580 was doing the same thing. unmasked is more accurate
2024-04-04 14:25:17 -04:00
82b7b9655f
test for dtype set ( #4069 )
2024-04-04 11:24:33 -04:00
1a1dd1c1a7
add and enable tests for indexing const folding ( #4068 )
...
* enable test in test_indexing
* added tests
* rename stuff
* del a test case cuz it's loadops.copy
2024-04-04 10:46:28 -04:00
chenyu
George Hotz
George Hotz
geohotstan
qazal
Felix Kuehling
David González Martínez
Francis Lam
andresgit
uuuvn
Akshit Talwar
Szymon Ożóg