* 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
prepared bfloat16 change. added float() and cast(default_float) in whiteing, explicitly set dtype in various places that convert between numpy and Tensor
* 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 5e8a67a535abea2ff288b1b804a9aa95eba40732.
* Revert "update tests"
This reverts commit 7ebf65d89ace1d3a32c3b28ee323ddee253262d6.
* Revert "don't increment num_batches_tracked if not tracking running stats"
This reverts commit 78de0ea9ee8cbd65dce28bd4abcc131c98451aa2.
* Revert "replace nn.BatchNorm2d with UnsyncedBatchNorm"
This reverts commit d03da53da70f009338e95f2b46315ac02a30149a.
* 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
* cpu tests pass
* torch works
* works
* metal works
* fix ops_disk
* metal jit works
* fix openpilot
* llvm and clang work
* fix webgpu
* docs are rly broken
* LRU works on metal
* delete comment
* revert name to ._buf. LRU only on Compiled
* changes
* allocator
* allocator, getting closer
* lru alloc
* LRUAllocator
* all pass
* metal
* cuda
* test examples
* linearizer
* test fixes
* fix custom + clean realize
* fix hip
* skip tests
* fix tests
* fix size=0
* fix MOCKHIP
* fix thneed
* copy better
* simple
* old style metal copy
* fix thneed
* np reshape
* give cuda a device
* feat: move to hip
* feat: special path for RawBufferTransfer
* feat: initial rawbuffertransfer
* feat: hip ipc
* feat: working hip ipc
* feat: need to base device without args
* feat: close mem handle
* feat: modified test
* feat: more multihip stuff
* clean: cleanup
* feat: cleaner
* feat: don't crash
* feat: test more
* clean: way cleaner hip wrapper
* feat: barrier
* feat: barrier
* feat: this breaks stuff
* feat: we can use empty here
* feat: maybe fix tests
* feat: maybe fix tests again?
* fix: probably fix tests
* feat: no waiting here
* feat: wait here
* feat: much larger test
* feat: need to sync here
* feat: make this async
* feat: no waiting!
* feat: cut here
* feat: sync copy
* feat: random imports
* feat: much cleaner world
* feat: restore this
* feat: restore this
* clean: cleanup
* feat: set this
* load weights in fp16
* add dtype option in nn
* fix test
* no need for dtype in nn
* add option to load weights in FP16, but NaN
* change loss scaler
* cast to float32 for norm layer
* add a todo for the forward pass padding
* fix transform
* change reduceop heruistics
* add model ema and jit hack
* add ema eval
* have to create a duplicate eval function for jit
* remove manual seed
* 94% achieveable with normal eval
* ema is outputting the same results as normal
* fix ema bug
* ema achieves 94% with fix seed
* multigpu tested
* constant fold decay, fix jit, adjust message for multigpu
* pull SpeedyResNet out of train_cifar()
George Hotz
David Hou
chenyu
Francis Lam
Yixiang Gao
qazal
wozeparrot
Sean D'Souza
George Hotz
Yixiang Gao