* re-enable gpt2 half/beam mac benchmark
from fuzzer it seems to be flaky due to numerical issue, not kernel bug. we used to have half in splitted reduce.
run this in M1 Max for 20 loops and it's fine
* that should be jitted
* add support for train/val datasets for kits19
* split dataset into train and val sets
* add tests for kits19 dataloader
* add MLPerf dataset tests to CI
* update unet3d model_eval script
* fix linting
* add nibabel
* fix how mock dataset gets created
* update ref implementation with permalink and no edits
* clean up test and update rand_flip implementation
* cleanups
* wmma: widen TC usage in search by using PADTO on TC axes when possible
* test: start tests for the new padding TC behavior
* search: upgrade padded TC search to TC_OPT >= 2
* test: add behavior and correctness test for padded TC
added optional argument to apply_tensor_core to set TC_OPT level
* linearizer: add tests for the PADTO behvaior and docs
* start
* fix err 93
* gpu
* ioctl mappings
* alloc like cuda
* semaphores
* wait for semaphores value
* start ops_nv
* very simple kernels work
* init several gpus
* qmd dumper
* dirty, but most of kernels work
* always all test_ops
* progress, more tests, stable
* test_ops passes, gpt2 works
but wth big fifo, wrap of fifo doesn't work, i think it's something coherency releated
* need better sync
* fix sync
* alloc2
* all tests pass!
* cleanup 1
* cleanup
* multigpu, simple transfer
* fix sync
* correct init
* nv_gpu autogen + sync bug fix
* clean extra/nv_gpu_driver
* p2p
* clean up
* remove old gen
* small fixes
* cleanup
* cleanup 2
* small fixes
* bigger queue size
* cleanups
* wait
* fixed signals for devs
* fix hang + parallel beam
* small fixes
* detect when local memory is big in kernel
* correct assert
* small fixes
* correct tls size est
* one va space
* less lines
* shorter
* save 2 lines
* save some lines
* remove type ignores
---------
Co-authored-by: George Hotz <72895+geohot@users.noreply.github.com>
* add DICE loss and metrics
* update dice to include reference implementation's link
* remove unused imports
* remove unnecessary test file and update pred + label for metrics and losses test
* add tests to CI + add exclusion of mlperf_unet3d
---------
Co-authored-by: chenyu <chenyu@fastmail.com>
* 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
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`
* tensor cores
* Merge from master
* faster program start in llvm (#3897)
* Fix the result permutation in einsum (#3895)
* Fix permutation of result indices in einsum.
* Delete stray line used for breaking tests
* Fix linter error by renaming twice-used variable
---------
Co-authored-by: chenyu <chenyu@fastmail.com>
* touchup einsum (#3900)
don't need rhs_letters
* hotfix check ckpts before writing achieved model (#3901)
this killed tinybox green run
* replace dtype.name str with render_dtype (#3903)
fixed some bf16 cast issue since it does not have `.name`.
also more robust if there are lang specific type override
* add --minimal flag to nvrtc (#3899)
* wmma: fix the AMD TC threads to split the first 16 threads (#3904)
previously it was incorrectly aliasing 16 into the size 8 upcast
on the store alias. now it splits it properly into 8 and the
remaining 2 into the correct local stride
* training cifar with BF16 on CUDA (#3905)
* 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
* include negative float in test_dtype (#3884)
* include negative float in test_dtype
* that is ub
* too annoying
* pack can overflow
* add to benchmark
* change var name to satisfy mypy
* spacing
* Update to new TensorCore format
* Spacing
---------
Co-authored-by: nimlgen <138685161+nimlgen@users.noreply.github.com>
Co-authored-by: Alejandro F Queiruga <33233447+afqueiruga@users.noreply.github.com>
Co-authored-by: chenyu <chenyu@fastmail.com>
Co-authored-by: sekstini <127142660+sekstini@users.noreply.github.com>
Co-authored-by: Francis Lam <flam@alum.mit.edu>
Co-authored-by: George Hotz <72895+geohot@users.noreply.github.com>
* kfd driver wip
* cleanups
* kfd almost ready to ring doorbell
* ding dong?
* issues with signals
* something
* works
* ops kfd
* add amd_signal_t
* works...sometimes
* program runs
* _gpu_alloc cleanup
* cleanups
* work
* header + enable profiling (#3959)
* header + enable profiling
* just cleaner
* measure
* only local time domain
* remove old comments
* fix with master
* elf parsing (#3965)
* elf parsing
* fix kernels with private
* not used
* clean up
* clean up 2
* add flags
* kfd sdma (#3970)
* working sdma
* remove driver, shorter
* all commands we might need
* svm
* kfd remove hardcoded values (#4007)
* remove hardcoded values
* match above line
* 7k lines + revert hsa
* update that from origin
* fix sdma reg gen
* not the updated SDMA
* compiler_opts
* don't require kfd_ioctl
* get ioctls from python
* get ioctls from python
* remove build_sdma_command
* merge into 64-bit fields
* shorter
* fix property spelling and off by one
---------
Co-authored-by: nimlgen <138685161+nimlgen@users.noreply.github.com>
* wmma: refactor to remove wmma_func and create TC funcs as needed
* test_linearizer: disable bf16 CUDA during emulation testing
* cstyle: clean up creation of CUDA vec dtypes
* extra/gemm: add option to accumulate to bfloat16
* cleanups
* benchmark: add CUDA bfloat16 matmul
* more cleanups
* fuzz_linearizer: reduce debug verbosity and make easier for CI usage
* rename FUZZ_BEAM to FUZZ_ALL_ACTIONS (not choosing a subset)
* skip simple ASTs (easier to use with LOGOPS output)
* don't fuzz a previously seen AST
* add options to allow non-zero --expected-failures
* clean up naming and use set
* 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
