* start
* tiny clean up
* whoops, didn't mean to accidentally fix this
* fix .to(device), kinda hacky and this fix makes it slower?
* merge properly
* FINALLY figured out slowness, also hack pylint for now
* add DEBUGONNX print for subgraph
* oops
* WOOOOOOOO SHAPE CACHE 50% SPEED INCREASE
* small fix, but maybe all deterministic Tensor creation in fp should be cached
* cache condition
* sliiiightly cleaner
* better abstraction?
* remove sam from model_benchmark
* remove shape cache speed up for now
* less lines
* isinstance fix
---------
Co-authored-by: chenyu <chenyu@fastmail.com>
* start
* add reference
* this is so much slower
* this makes sense but differs from official impl, but results are still correct..?
* add a comment
* Just keep it simple for now since I don't fully get it yet
* address comments
* correct
* teeny clean up
* another small comment improvement lol
* bump
* thou hast implement functions
* hacked in domain support
* some clean ups
* hack quantize_onnx_test too
* add helper lol, why onnx tests why
* better dispatcher, but need tests and better naming
* flaky ci
* change some names
* small clean ups
* make it easier to clean up tests once ORT supports 1.18.0
* nits
* fix bug of Softmax_1 being registered in onnx_ops
* need a default value
* resolve_const is better name
* fix OnnxRunner.to
* use proper domain names
* file path as input and have parse be in OnnxRunner.__init__
* modelproto_to_onnxrunner -> modelproto_to_runner
* whoops, fix import
* oh flakiness again, is it because it's getting gc-ed?
* small changes
* CI flaky so just move compile4 fix in
* copy typing of onnx_load
* actually can just import onnx_load instead of onnx.load
* fix external_benchmark_openpilot
* fix onnx_runner test to use onnx_helper
* rerun CI
* try run_modelproto
* spam CI a few times
* revert run_modelproto since that's flaky also
* no external onnx_load usage except onnx.py
* cursor tab complete is evil. Snuck a darn sorted in. But does order change result? Why?
* model_benchmark 193s -> 80s, add OnnxRunner.to()...
* minimize diff and clean up
* device can be None, weird but eh
---------
Co-authored-by: chenyu <chenyu@fastmail.com>
* try
* ruff check --fix
* no skip test
* hmmmmmmm I don't get this D:
* run CI again
* why is PYTHON device faster than CPU?
* run ci again and fix lint
* actually doesn't PYTHON device make sense here?
* see cpu speed again
* Revert "see cpu speed again"
This reverts commit 1e366f2256.
* trigger CI
* pretty good
---------
Co-authored-by: chenyu <chenyu@fastmail.com>
* squash commits
* temp fix for const tensor
* actually realizing float16 can only happen in raw_data
* .float -> cast(float) to rerun CI
---------
Co-authored-by: chenyu <chenyu@fastmail.com>
* why does max_unpool2d feel slower than out.gradient ...
* slightly cleaner
* what happened to ruff
* need to think about this some more
* slightly faster now?
* clean up, 1 more failing edge case
* ok good
* working TINY_BACKEND
* nit doc wording
* retry CI
* terrible but somewhat working impl
* linux behaves differently than macos?
* slightly better impl
* small clean up; haven't figured this out yet
* better
* torch has different behavior on linux and macos for duplicated values
* add sum docs
* fix test
* add torch return_type test
* add an exception test
* wrap_fxn instead, and move op lower in order
* better repeated values test
* rerun ci
* add DynamicDequantizeLinear and corresponding tests
* wow qlinearops are round away from zero
* this passes locally...
* again
* try
* try separate test
* round to even again
* also add QLinearMul
---------
Co-authored-by: chenyu <chenyu@fastmail.com>
* pytorch scatter -> scatter_reduce
* WIP scatter_reduce implementation
* _pre_scatter return type hint
* split out src, mask to satisfy linter
* Add src cast back in
* dict of lambdas instead of ifs
* sum and prod reduction ops with include_self
* add reduce arg error message
* add amax and amin reduction ops
* Fix include_self for higher dims
* Simplify
* Simplify amax and amin too
* Pull include_self logic out into _inv_mask function
* reduce arg cannot be None for scatter_reduce
* Fix self-mask issue
* Add mean reduce op
* Add tests
* any() not needed here
* remove comment
* End support for Tensor src with reduce arg in tinygrad scatter
* Process index, dim inside actual functions
* Add scatter_reduce to onnx
* Add excluded onnx ScatterElements reduction tests back in
* Save 2 lines on the mask helpers
* Update docs
* Add include_self=False tests
* cleanup
* Remove unneeded helper function
---------
Co-authored-by: chenyu <chenyu@fastmail.com>
* this
* clean up
* more clean ups and improve debug msg
* more correct training toggler
* remove manual training toggling
* change some variable names
* actually just add the training toggle for LIMIT envvar too
* more refinement
* __call__ and OnnxRunner
* fix half pylint, other half is importing from onnx while this file is onnx.py, figure out later
* ahhhh found another mistake
* remove limit from __call__
---------
Co-authored-by: chenyu <chenyu@fastmail.com>
* is 67% considered fixed?
* move test up
* share function
* add qgemm too
* make sure qgemm comes out as int
* actually that note is not right
* remove qgemm (I did it wrong) and add it later lol.
* QLinearEverything
* ok ort verify passes
* this should be int instead
* cast to int then char to do wraparound
* cleaner
* move contrib ops to microsoft ops
---------
Co-authored-by: George Hotz <72895+geohot@users.noreply.github.com>
* implemented in tensor
* apply onnx tests to asymmetrical pads
* better onnx op ordering
* correct ceil_mode asymmetrical
* fix onnx_ops comments
* a few more TODOs and fix some stupidity
* fix some typing
* fix test
* mypy still a little messed up
* refactor out pad struct transformation
* add simple docs for now
* add whatever tests possible
* add tests for _resolve_pool_pads
* better err msg
* whoops didn't mean to include this
* retry CI
* enable asymmetric pads onnx tests
* better docs
---------
Co-authored-by: chenyu <chenyu@fastmail.com>
it's a python style mod. possibily can be cleaner with a floor div
relaxed the vmin for MOD slightly for cstyle negatives mod, it's more correct and might fix other bugs
* implemented
* this implementation is now correct
* this is fine I guess
* better variable names
* finally correct gathernd
* add a note
* eh just leave it at this for now
* teeny adjustment
* 1 is simpler than 2
* variable name
* change error wording
* shapes for sequence type must be homogeneous
* bug fix for model benchmark
* fix comments too
---------
Co-authored-by: chenyu <chenyu@fastmail.com>
* simple clean ups first
* more work
* kinda have adam
* ooo momentum worked nicely
* almost there
* wow.. is the onnx test wrong
* nicer optim stuff
* just skip that test
* small comment changes
* use naming convention from other parts of codebase
---------
Co-authored-by: chenyu <chenyu@fastmail.com>
* working I think
* where are my onnx scatter tests??
* forward_only for now
* try if nan hack fix NV
* looks like issue is different... CUDA WHY
* oops that was wrong. Try if this fixes CUDA
* simpler multiply
* actually finish this up tmrw morning :x
* fix tests?
* improve tests
* improve test and implementation
* fix ruff
* complete but lots of expected failure...
* reviewed tests
* add onnx tests
* is this a processing op?
* add return type to indicate that it's not in-place
* final cleanups
* use or and improve tests a little
* add masked_index_select
* call it masked_setitem instead
* try
* FIXED
---------
Co-authored-by: chenyu <chenyu@fastmail.com>
