make it read nicer and cleanup some movement methods and math simplification.
790m, 1.4b, 2.8b model does not really run.
sampling is not implemented.
jit is incorrect.
some deadcode / wrong code path and copied from torch stuff stuff.
* first commit
* state back to orig
* mamba comparisions
* rm file
* rename file
* use Tensor.einsum and mke default model 370M
* Cleaned code and made a comparision test
* Simplyfy pull request. Only has 1 mamba implementation now.
* Update prompt
* rm whitespaces
* last space
* remove Einops dependency
* rm unused code
* add tests
* rm print statement
* rm imports
* skip CLANG
* Update skipIf description
* skip model test in CI and add CLANG fix
* rm Device import
* don't be stupid
* Fix conv assign
When the prompt is too short, the logic for conv_state assign messes up. This can be fixed when padding the tokenized array to min length of 4. I padded using the empty string token, but idk if proper practice is to use the PAD token
* fix p1
* temp
* fix jit import
---------
Co-authored-by: schlimeszn <schlimeszn@gmail.com>
Co-authored-by: reddyn <nikidsniper@gmail.com>
Co-authored-by: George Hotz <72895+geohot@users.noreply.github.com>
* fp16 resnet
* cast running mean and var back to default float
* extra cast
* check symbolic no overflow
* add linearizer failure
* loss scaler after grad contig
* oops
* i think this works
* don't loss scale fp32
* remove overflow test case
* remove symbolic bounds check
* loss scaler should be float
* temporarily disable padto cuz bug
shruggie
* make running stats in batchnorm float32?
* calculate lars stuff in fp32?
* oops
* remove most changes
* move loss scaler out of optimizer
* no more FP16 var
* oops
---------
Co-authored-by: chenyu <chenyu@fastmail.com>
* env var to change default float to fp16 or bf16
looking for standard names for these. we have FLOAT16 that does something to IMAGE and HALF to convert weights.
working on default bf16 too.
```
RuntimeError: compile failed: <null>(6): error: identifier "__bf16" is undefined
__bf16 cast0 = (nv_bfloat16)(val0);
```
remove that in cifar
* DEFAULT_FLOAT
* default of default
* unit test
* don't check default
* tests work on linux
* training cifar with BF16 on CUDA
memory usage is between float and half due to numpy calls on dataset preprocessing, which converts into float.
* simpler bf16 functions
* bf16 cifar works for HSA too just very slow
* simpler bf16 functions, we love cuda
copy scale on all device for now. naive sharding does not work because scale needs expand to really save memory.
70B does not work due to HSA_STATUS_ERROR_OUT_OF_RESOURCES.
`python3 examples/llama.py --gen 2 --size 13B --shard 6 --prompt "Hello." --count 10 --temperature 0 --timing --quantize`
13B on 6 gpus uses 47 GB v.s. 34 GB quantized
This is how bf16 load is tested in test_bf16_disk_write_read now and it should fix#2775.
I tested that it fixed loading coder using PYTHON backend.
Will separate this special bf16 load v.s. regular bf16 support
* simple LoadOps.ASSIGN
* skip that test
* don't assign in onnx ops gemm
* track cache usage
* recreate the lazybuffer to avoid the cache
* fix contigs
* skip that test
* lol
* better letters
* this is a lot of stuff
TEST_TRAIN env for less data
don't diskcache get_train_files
debug message
no lr_scaler for fp32
comment, typo
type stuff
don't destructure proc
make batchnorm parameters float
make batchnorm parameters float
resnet18, checkpointing
hack up checkpointing to keep the names in there
oops
wandb_resume
lower lr
eval/ckpt use e+1
lars
report top_1_acc
some wandb stuff
split fw and bw steps to save memory
oops
save model when reach target
formatting
make sgd hparams consistent
just always write the cats tag...
pass X and Y into backward_step to trigger input replace
shuffle eval set to fix batchnorm eval
dataset is sorted by class, so the means and variances are all wrong
small cleanup
hack restore only one copy of each tensor
do bufs from lin after cache check (lru should handle it fine)
record epoch in wandb
more digits for topk in eval
more env vars
small cleanup
cleanup hack tricks
cleanup hack tricks
don't save ckpt for testeval
cleanup
diskcache train file glob
clean up a little
device_str
SCE into tensor
small
small
log_softmax out of resnet.py
oops
hack :(
comments
HeNormal, track gradient norm
oops
log SYNCBN to wandb
real truncnorm
less samples for truncated normal
custom init for Linear
log layer stats
small
Revert "small"
This reverts commit 988f4c1cf3.
Revert "log layer stats"
This reverts commit 9d98224585.
rename BNSYNC to SYNCBN to be consistent with cifar
optional TRACK_NORMS
fix label smoothing :/
lars skip list
only weight decay if not in skip list
comment
default 0 TRACK_NORMS
don't allocate beam scratch buffers if in cache
clean up data pipeline, unsplit train/test, put back a hack
remove print
run test_indexing on remu (#3404)
* emulated ops_hip infra
* add int4
* include test_indexing in remu
* Revert "Merge branch 'remu-dev-mac'"
This reverts commit 6870457e57, reversing
changes made to 3c4c8c9e16.
fix bad seeding
UnsyncBatchNorm2d but with synced trainable weights
label downsample batchnorm in Bottleneck
:/
:/
i mean... it runs... its hits the acc... its fast...
new unsyncbatchnorm for resnet
small fix
don't do assign buffer reuse for axis change
* remove changes
* remove changes
* move LARS out of tinygrad/
* rand_truncn rename
* whitespace
* stray whitespace
* no more gnorms
* delete some dataloading stuff
* remove comment
* clean up train script
* small comments
* move checkpointing stuff to mlperf helpers
* if WANDB
* small comments
* remove whitespace change
* new unsynced bn
* clean up prints / loop vars
* whitespace
* undo nn changes
* clean up loops
* rearrange getenvs
* cpu_count()
* PolynomialLR whitespace
* move he_normal out
* cap warmup in polylr
* rearrange wandb log
* realize both x and y in data_get
* use double quotes
* combine prints in ckpts resume
* take UBN from cifar
* running_var
* whitespace
* whitespace
* typo
* if instead of ternary for resnet downsample
* clean up dataloader cleanup a little?
* separate rng for shuffle
* clean up imports in model_train
* clean up imports
* don't realize copyin in data_get
* remove TESTEVAL (train dataloader didn't get freed every loop)
* adjust wandb_config entries a little
* clean up wandb config dict
* reduce lines
* whitespace
* shorter lines
* put shm unlink back, but it doesn't seem to do anything
* don't pass seed per task
* monkeypatch batchnorm
* the reseed was wrong
* add epoch number to desc
* don't unsyncedbatchnorm is syncbn=1
* put back downsample name
* eval every epoch
* Revert "the reseed was wrong"
This reverts commit 3440a07dff3f40e8a8d156ca3f1938558a59249f.
* cast lr in onecycle
* support fp16
* cut off kernel if expand after reduce
* test polynomial lr
* move polynomiallr to examples/mlperf
* working PolynomialDecayWithWarmup + tests.......
add lars_util.py, oops
* keep lars_util.py as intact as possible, simplify our interface
* no more half
* polylr and lars were merged
* undo search change
* override Linear init
* remove half stuff from model_train
* update scheduler init with new args
* don't divide by input mean
* mistake in resnet.py
* restore whitespace in resnet.py
* add test_data_parallel_resnet_train_step
* move initializers out of resnet.py
* unused imports
* log_softmax to model output in test to fix precision flakiness
* log_softmax to model output in test to fix precision flakiness
* oops, don't realize here
* is None
* realize initializations in order for determinism
* BENCHMARK flag for number of steps
* add resnet to bechmark.yml
* return instead of break
* missing return
* cpu_count, rearrange benchmark.yml
* unused variable
* disable tqdm if BENCHMARK
* getenv WARMUP_EPOCHS
* unlink disktensor shm file if exists
* terminate instead of join
* properly shut down queues
* use hip in benchmark for now
---------
Co-authored-by: George Hotz <72895+geohot@users.noreply.github.com>
prepared bfloat16 change. added float() and cast(default_float) in whiteing, explicitly set dtype in various places that convert between numpy and Tensor
* examples/stable_diffusion: support model checkpoints without alphas_cumprod key
(which is most models on civitai)
* fix indent
---------
Co-authored-by: a <a@a.aa>
* working PolynomialDecayWithWarmup + tests.......
add lars_util.py, oops
* keep lars_util.py as intact as possible, simplify our interface
* whitespace
* clean up
* clean up
* asserts
* test polylr for full resnet training run
* add comment
* rename
* fix do_optim
* don't cast lr
* info
* calculate from train_files
* skip it
* lars optimizer + tests
* fix skip list!
* use id to compare in skip list
* go back to using set
* Tensor(bool) * Tensor(bool) is and
* don't lint external/mlperf_resnet
* whitespace
* add external_test_optim to opencl tests
* give mlperf task a name
* mlperf under onnx
* remove track_gnorm
* contiguous instead of realize
* assert momentum and weight decay positive
---------
Co-authored-by: chenyu <chenyu@fastmail.com>
* allow LB <- MLB assign, but don't reuse buffer
* update test
* update test
* assign assert axes are the same
* update tests to manually shard running stats
* unused import
* UnsyncedBatchNorm with synced trainable weights for hlb cifar
* multitensor reshape tests
* test mlb assign change axis
* E501
* argfix axis
* don't import batchnorm from hlb_cifar in test_multitensor
* pass num_devices to UnsyncedBatchNorm in test, allow UnsyncedBatchNorm to be used with LB
* add backprop test for UnsyncedBatchNorm
* break out MLB assign and reshape changes
* manually shard running mean and running var
* don't shard unless syncbn=0
* replace nn.BatchNorm2d with UnsyncedBatchNorm
* don't increment num_batches_tracked if not tracking running stats
* update tests
* oops
* Revert "oops"
This reverts commit 5e8a67a535.
* Revert "update tests"
This reverts commit 7ebf65d89a.
* Revert "don't increment num_batches_tracked if not tracking running stats"
This reverts commit 78de0ea9ee.
* Revert "replace nn.BatchNorm2d with UnsyncedBatchNorm"
This reverts commit d03da53da7.
* don't increment num_batched_tracked if not tracking running stats
* oops
* test_batchnorm_axis
* compare against torch
* types
---------
Co-authored-by: chenyu <chenyu@fastmail.com>
* shrink MLB on sharded axis
use onehot structure to store the real partition. goal is unsynced batchnorm2d that can be run on multigpu for training.
draft version in https://github.com/chenyuxyz/tinygrad/pull/109
* SYNCBN flag
* test unclean shrinks
* UnsyncedBatchNorm reuses BatchNorm
* more robust pad arg check
* better types
* more tests!
* 6 gpus in benchmark
* disable slow GPUS=6 benchmark
