* Switch to dawn, all tests passing locally
* Use dawn-python
* Skip failing test
* Skip midcast and fix timestamp on metal ci
* Autogen webgpu
* Try fetch dawn lib again
* /usr/lib
* Without lib prefix
* Test autogen diff
* Delete webgpu support, move everything to ops_webgpu
* mypy fix
* Simplify, refactor
* Line savings
* No ResultContainer
* Type annotation for result
* Some more simplifications
* Why was this explicit sync used at all?
* Refactor: delete functions that are only used once
* Create shader module inline
* Clear unit tests cache, maybe that solves it
* That wasn't it
* Try deleting cache to pass failing weight compare
* weights_only=False for pytorch 2.6
* Simplify ctype array creation
* Remove nanosecond precision timestamps
* Simplify error handling
* Refactor, add back type annotations
* Deleted custom submit function, refactor
* read_buffer simplify
* Fix use after free, refactor
* Simplify supported_features
* Runtime docs
---------
Co-authored-by: George Hotz <72895+geohot@users.noreply.github.com>
* LLVM JIT
* Autogen LLVM
* Update autogen
* Move things around
* even more non-determinism
* windows
* more autogen weirdness
* more windows stuff
* blind windows development try 2
* more blind windows development
* even more blind windows development
* maybe i should just set up a windows vm...
* why can't everyone just use sysv abi?
* cleanup debugging stuff
* unused import
* icache flushing isn't required on x86
* merge jit_nt and jit_unix
* more
* Temporary hack to not segfault
* better error
* bad conflict resolution
* Attempt to simplify support/llvm.py
* More refactoring
---------
Co-authored-by: George Hotz <72895+geohot@users.noreply.github.com>
* connect to gpu
* rlc init?
* gfx comp start init
* early init is hardoded, some progress with fw
* gart
* progress, next mqd
* ring setup, still does not execute anything
* ugh write correct reg
* pci2: vm
* pci2: start psp
* vm seems to work
* pci2: gfx start
* pci2: fix psp ring resp
* pci2: try ring
* pci2: mes and some fixes
* pci2: some progress
* pci2: progress
* pci2: mm
* pci2: discovery
* pci2: correct apertures
* pci2: b
* pci2: i
* pci2: l
* pci2: o
* pci2: cmu
* pci2: mes_kiq works
* pci2: mes
* pci2: kcq does not work(
* pci2: unhalt gfx
* ops_am
* minor
* check if amdgpu is there, or we will crash
* bring back graph, it just works
* less prints
* do not init mes (not used)
* remove unused files
* ops_am: start move into core
* ops_am: works
* clcks, but still slower
* faster + no mes_kiq
* vm frags + remove mes
* cleanup fw
* gmc tiny cleanup
* move to ops_amd
* comment out what we dont really need
* driverless
* close in speed
* am clean most of ips
* gmc to ips
* cleaner
* new vm walker
* comment old one
* remove unsued autogens
* last write ups
* remove psp hardcoded values
* more
* add logs
* ih
* p2p and sdma
* vfio hal and interrupts
* smth
* amd dev iface
* minor after rebase
* bind for sdma
* Revert "bind for sdma"
This reverts commit a90766514d.
* tmp
* debug new mm
* ugh, allreduce hangs fixed
* p1
* works
* no pci.py
* cleaner a bit
* smth
* tiny cleanups
* cleaner a bit
* pciiface
* linter
* linter 2
* linter 3
* linter
* pylint
* reverted unrelated changes
* unrelated
* cmp tool
* ugh wrong fw
* clockgating
* unrelated
* alloc smaller chunks
* this
* opt sigs
* collect stat
* ops
* upd
* proclogs
* proclogs2
* vfio
* ruff
* linter pylint
* oops
* mypy p1
* mem fix
* mypy p2
* mypy p3
* mypy p4
* correct
* minor
* more tests
* linter in tests
* pci_regs header
* minor write up
* setup
* do not require libs
---------
Co-authored-by: George Hotz <72895+geohot@users.noreply.github.com>
* start work on new gradient
* more correct
* working tests
* more tests
* work
* add (faliing) gradient test
* add view and reduce gradient
* test_add works, many failing test_ops
* add max and reduce max
* add max and reduce max
* 129 failing
* 108 failed
* better view drawing
* 101 failed
* i got 99 failures
* 94 failures
* it's tons of terrible code, but only 50 tests fail
* only 19 failures
* same 19 but shorter
* minimal doesn't matter
* shorter
* lil simpler
* simpler
* simpler
* simpler
* 13 test failures
* nine tests fail
* all ops tests pass
* add contiguous gradient + fix sched tests
* faster by removing toposort calls
* missed one
* add jax to testing
* explicit instantiation and narrowing asserts
* explicit cast
* bump
* one line assert
* handle case for no copy_queue_t
* Revert "handle case for no copy_queue_t"
This reverts commit 38347806ca.
* more readable control flow
---------
Co-authored-by: George Hotz <72895+geohot@users.noreply.github.com>
* Start from andredaprato:webgpu-clean
* Fix infs
* inf wgsl function is not needed
* Emulated ulong for threefry, more tests passing
* Randomness tests passing
* Update model export to support new changes in webgpu, efficientnet export works again
* Simplify shift emulation in wgsl
* Delete test file
* Fix bigger than u32 u32 literal
* Why was skip copies added here?
* Python3.12 for webgpu tests
* Fix model export syntax error
* Get test ops passing with some skips
* Fix lint
* Much simpler shift
* Run more tests
* Timestamp queries are not supported in CI, so skip search tests
* All fancy indexing passing
* r is ctx
* Run more dtype tests by using is_dtype_supported
* Cleanup ulong shift rendering
* UPat -> Pat, UOps -> Ops
* Pat -> UPat
* Refactor render_ushift if-else
* Pattern to avoid ulong mul
* Remove vals_dtype
* is_nan trick + rewrite, test_isnan passing
* Rewrite a * select(1, nan, gate) -> select(a, nan, gate)
* No arg, just op
* Support char, uchar, short, ushort
* Run test_index_mnis now that we have uint8
* Fix pyling
* Save 3 lines by using base Compiler
* No more long emulation
* Remove fixup_binops
* No more external_local_bufx wgsl specific cstyle modif, use base extra_pm
* Simpler, faster copyin/out
* Skip some new tests that use long
* Fix typo
* copyout touchup
* Save lines by using render_cast
* WebGL is not supported in core, delete it from is_dtype_supported
* More narrow test skips for some unary tests
* TernaryOps, UnaryOps -> Ops
* TinyGrad supports WebGPU
* StableDiffusion demo: f16tof32 gpu is a lib, update UI
* Packed load/store, no more scale_size, no core tinygrad changes
* Rename copyin, copyout
* Device -> dev
* Fix lint
* Pattern matcher rule for packed load/store
* Refactor
* Shorter packed load/store
* this should fix lint
* Fix mypy
* SD compile script working
* New SD webgpu UI
* New default prompt
* New SD weights
* Fix title when webgpu not available
* Run symbolic tests, simplify is_nan, use round_up
* Show step time on UI
* Bump minimum wgpu version to v0.19
* Fix latent
---------
Co-authored-by: George Hotz <72895+geohot@users.noreply.github.com>
* new cloud is cloudy [pr]
* waste lines to add security
* safety, with speed and less lines
* timing and del
* lines
* cleanups
* restore CloudSession
* bump to 3.10
* quotes
* renderer security
* basic loader, untested
* testing
* remove utils import in test
* q8_0
* q4_1
* end to end testing
* minor cleanup
* fix casting
* moved to state
* move tests
* move dequant to fn
* fix lint elif
* remove gguf from extra
* fix dict union
* q6_k simpler
* naming and spacing
* gpt2-gguf example
* cleanup
* move gguf example
* minor cleanup
---------
Co-authored-by: George Hotz <72895+geohot@users.noreply.github.com>
* Add CDLL interface for metal
* remove two unused functions
* Cover most of the API methods
* switch to cdll
* directly call objc message in ops_metal
* keep only obj interface
* Use direct message sending for graph
* may have found a solution to the memoryview on ctypes pointer
* buf indexing bug fixed
* fix c_int
* fix c int to bytes
* fix gpu time bug
* line savings for cdll metal core
* wip
* c int bug
* fix buf casting
* dedup for c_void_p
* dedup for c_void_p
* linter fix
* remove unused stuff
* my py fix
* more mypy error fix
* line savings
* line savings
* rename send_message to msg; add __hash__ and __eq__ for dedup
* wip
* refactor
* refactor
* remove named import from ctypes
* forgot to change variable name
* file reorg, put support.py to ops_metal
* refactor
* hash error
* remove to_ns_array
* test oom exception, fix exception change
* typevar for msg
* add back dedup
* test for compile error
* move constant to graph
* move header constant around
* get label for icb buffer
* check icb label using "in"
* wip fixing mypy reported error
* fixed mypy error
* code formatting
* all_resources dedup match previous
* code formatting
* code formatting; buffer set to objc_id
* revert changes on buf for the manual release, seems like _free is not always called
* skip unless on metal, for test_metal
* fix premature mem release causing seg fault
* test_metal check for device before importing
* Buffer should only be released under _free explicitly
* mypy fixes
* change object ownership
* test compile success
* lint fixes
* remove load_library
* wrap sel_register in cache
* simplify to_struct
* swap lines
* fix type error in to_struct
* bump line to 9800
* remove pyobjc from setup.py
* command buffer should be objc_instance and get released
* stringWithUTF8String: returns objc_instance
* Use constant for MTLPipelineOptionNone
* better explanation for [MTLBuffer contents:] return
* Use dyld_find in case the path differs
* trailing whitespace
* handle exception for methods that take error:
* load /System/Library instead of /Library
* Init c_void_p with None instead of zero for error objects
---------
Co-authored-by: Mesozoic Egg <mesozoic.egg@proton.me>
Co-authored-by: George Hotz <72895+geohot@users.noreply.github.com>
* tqdm replacement almost
* formatting
* formatting
* imports
* line len
* fix
* removed set description :(
* removed set description :(
* fix
* fix
* green check?
* rewrote as class, fixed several bugs
* types spacing
* removed imports
* fix
* iterable
* typing
* mypy disagreement
* imports
* more e2e tests vs tqdm
* removed seed setting
* robustness against time.sleep() flakiness
* flaky fix
* automatic bar closing when count==total
* cleanup
* clang error with tqdm
* tqdm back
* use os lib, print to stderr (fixes the clang bug, where the bar was leaking into the generated c program
* back to shutil
* unit_scale + unit_scale test
* custom unit to tests
* pretty
* clean
* removed flaky test
* less test iters
* empty line
* remove disable
* 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
* 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>
* move gpuctypes in tree
* fix mypy
* regex exclude
* autogen sh
* mypy exclude
* does that fix it
* fix mypy
* add hip confirm
* verify all autogens
* build clang2py
* opencl headers
* gpu on 22.04
* dtypes alu test
* those types don't exist in torch
* floats
* more tests
* disable those
* a couple unary tests
* skip float16 tests in CI for GPU
* fix LLVM bool add True+True=1+1=2 which truncates to False in native LLVM
* remove hardcoded float for LLVM ALU fns
* less sensitive atol for fp32, 1e-10 is flaky and sometimes failed even if you revert the merge commit for non-fp32 math, nothing has changed in our kernels for fp32.
* return on overflows
* fix CUDA exp2
* compute results of op regardless of bounds in a python backend
* skip fp16 in GPU and CUDACPU
* fuzz a smaller range in the float_midcast_int32 test
I sampled this and we overflow ~70% of the time.
because numpy behaves differently on different devices for overflows and Metal seems to do the same, I'm opting to eliminate the non-determinism here
* remove CUDA exp2 overload it's already there now
---------
Co-authored-by: George Hotz <geohot@gmail.com>
