* remove cpu and torch backends
* don't copy to cpu
* use clang instead of cpu
* multitensor gathers on the first device
* clang is cpu + use default
* fixup
* bugfix
* extra/gemm: add a simple_conv.py along with correctness check
The goal is to easily test tensor core triggering situations
* test: add tests for acc_dtype handling and fixed typing
* 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
* add onnx test_reduce_log_sum_exp
* more reuse
* more
* stuff
* good CenterCropPad
* imports
* good ArrayFeatureExtractor
* pretty good Pad
* stuff
* stuff
* onnx.py
* Atan
* pass int8 test
* dtype related
* fastmath stuff
* Resize linear
* fix CI
* move back
* try
* test: add logical_not tests
* gah im retarded, but this doesn't match types for const()
* fix: can't we jsut do this?
* big change: I don't actually know what I'm doing
* WOOO IM JUST CHANGING EVERYTHING WOW probably gon revert later
* BYE BYE noqa: E501
* fix: less lines and add test
* fix: rm 2 redundant tests
* fix: eq with False so we don't unintentionally implicit upcast, but it's bool anyways so w/e
* shard llama
* sharding works
* simpler
* simpler
* consume option
* disable that test
* save a line
---------
Co-authored-by: George Hotz <george@tinygrad.org>
* get basic ptx impl working
* test ops passing
* mypy
* dont hardcode target
* more walrus
* ptx in ci
* bool cast and f16 load/store
* weird numpy bug and f16 cast tolerance
* cast half to bool
* fix 1 byte load/store
* disable half for ptx
* fix args and enable xid
* fix non-ptr args
* allow bitcast
* mypy
* cleanups
* midcast use allclose
* add xor
* Revert "disable half for ptx"
This reverts commit 73391c05fde5f7811293f60d994417d97ab20613.
* enable float16
* mypy
* no more crashing in ci
* fix ci
* minor cleanups
* use new fn for ptx compiler
* no diskcache in ptx compile
* use rn instead of rz
* save some lines
* new DEFINE_GLOBAL syntax
* line length
* new llvm
* cmpeq
* minor fix
* cast in mulacc
* update test_recursive_add to check line count
* mypy
* remove llvmir.py
* fix bool const
* wip
* cleanups
* working
* llvm in separate pr
* cleanups
* more cleanups
* fix ci
* use in_features directly in nn.Linear.__init__ bound check (#3050)
* use in_features directly in nn.Linear.__init__ bound check
get rid of the unnecessary check of isinstance int
* that is always int
* long lines
* Device._buffers -> Device._devices (#3052)
backend devices used to be called buffers
* make Embedding device aware for multigpu (#3051)
* make Embedding device aware for multigpu
* split line instead of igore because that's cheating
* add test incomplete
* add test complete
* remove comment
* fix white space
* remove nn.Embedding
* remove unused reciprocal (#3053)
* remove unused reciprocal
* comment
* unit tests for Device.canonicalize (#3055)
* add multigpu test for RMSNorm (#3056)
* need all gather
* add two multigpu test scenarios for RMSNorm
* No extra vars call (#3054)
* remove unused reciprocal
* comment
* remove unneeded call to vars
* free speedup
* explicit lazybuffer caching (#3058)
* hotfix: remove useless slow assert from ShapeTracker
* Speed tweaks (#3059)
* base doesn't have to be a function
* no double fetch
* pop, don't check
* make the gc happy
* avoid hasattr
* cache canonicalize
* remove assert, faster base
* don't redefine that every time
* fix gpt2 attention with start_pos = 0 (#3061)
* fix gpt2 attention with start_pos size 1
test cases taken from ll_transformer branch
* fix interpreted
* Tensor.cat with 0 shape tensors (#3062)
* Tensor.cat with 0 shape tensors
supported both 0 in cat axis (for a subset of input), or 0 in non-cat axis (all needs to be 0)
* no shp
* test scaled dot product attention (#3063)
* add test
* add initial test for scaled dot product attention
* test pass for scaled dot product attention
* cached size (#3060)
* cached size
* simplify simplify
* 0 doesn't have base
* fix test
* cleaner cache
* hmm, metal is flaky on this...might be real(ish) but useless as test
* short circuit reshape/expand properly
* better reshape bypass
* hotfix: use is for enum compare
* hotfix: use is for enum compare, a few more
* speedtweaks3: apply shouldn't use the tensor constructor (#3065)
* speedtweaks3: apply shouldn't use the tensor constructor
* replace 0 size with CONST, not 0 in shape
* update gh actions (#3033)
* update checkout actions
* update upload artifact
* update setup python
---------
Co-authored-by: George Hotz <72895+geohot@users.noreply.github.com>
* unbind view or shapetracker also returns var_val (#3067)
* unbind view or shapetracker also returns var_val
4% faster for llama compile time
* one line less
* unbound_views
* hotfix: examples/transformer.py
* jit autorealizes output (#3069)
* early gate the graph (#3070)
* simpler idxs_to_idx (#3071)
* filter_strides -> canonicalize_strides (#3072)
* fix onehot and jit in examples/transformer (#3073)
trained to 0.999 in < 6 seconds on M1 Max consistently
* better test demonstration (#3077)
* a better test demonstration
* fix white space
* Tensor.expand resolves the new_shape before shortcut return (#3078)
similar to how reshape is done. also updated shrink shortcut criteria to read similar to pad
* minor cleanups of lazy.py (#3080)
* wmma: clean up device specific tensor core code (#3081)
* mem_estimate is always int, not symbolic (#3083)
* mem_estimate is always int, not symbolic
op_estimate can be symbolic, but mem_estimate is always int, thus we don't need to sym_infer it.
fixed some long lines too. update_stats is a very big function
* operator does not need underscores
* cat works (#3086)
* hotfix disable flaky mac runner wino cifar (#3087)
* remove the third merging state in view._merge_dims (#3085)
no logic depends on state == 0 or state == 2
* minor cleanup of View.reshape (#3088)
* minor cleanup of View.reshape
removed some redundant logic
* new_strides
* revert that
* use BEAM=2 instead of BEAM=4 in cuda ci gpt2 (#3089)
BEAM=2 is faster and less search time. investigating why BEAM2+BEAM4 is slower than BEAM2 alone
* use device from LinearizerOptions in kernel search (#3090)
* use device from LinearizerOptions in kernel search
removed all Device.DEFAULT in search.py
* pass device string for parallel pickle
* device for interpreted backends in LinearizerOptions
* update jit type annotation post lazy rewrite (#3091)
* add mutigpu support for llama attention (#3064)
* add llama attention test for multigpu
* test fails
* kv cache trying to shrink on sharded axis
* mask None works for scale dot product
* kv cache seems to be working but scale dot product breaks
* scaled dot product works, but the last linear layer failed
* running into the reshape case where it could be wrong for multigpu
* making sure it was the reshape
* adding contiguous doesn't solve
* need to shard more properly
* remove reshape test
* minor adjustment to scale dot product attention test
* weights are sharded wrong
* continue fix new weight sharding
* clean up
* fix attention when start_pos is 0
* remove print
* add TODOs for the best mutigpu interface
* bugfix do not reset shapetracker of 0 size lazybuffer (#3096)
it might be coming from an expand, and resetting results incorrect stride. caught by interpreted backend
* One hot in tensor.py (#3093)
* onehot in Tensor.py
* one_hot tests
* works for all shapes, not just 1
* pylint
* not a static method
* moved around, num_classes mandatory
* pylint
* pylint
* space & moving
* formatting
* moved tests
* fix broadcasted logic if there's 0 in shapes (#3097)
* fix broadcasted logic if there's 0 in shapes
should always expand into 0, not the other way around. fixed matmul with 0 in input shapes.
for forwards for now though, backward is more involved and would need to change 0 size shortcuts
* fix tests
* replace with tensor op (#3099)
* fix gpt2 with empty prompt (#3100)
logits would be empty so need to replace that with ones before sampling, also cannot reshape with -1 when there's 0 in other axes
* Revert "fix gpt2 with empty prompt" (#3101)
* fix gpt2 with empty prompt take 2 (#3102)
logits would be empty so need to replace that with ones before sampling, also cannot reshape with -1 when there's 0 in other axes
* wmma: enable METAL half tensor cores and clean up cstyle (#3095)
* wmma: enable METAL half tensor cores and clean up cstyle
* revert simple_matmul rand changes and break line in tensor
* added metal fp16->fp32 tensor core
* add half @ half to mac benchmark (#3103)
* flag to profile mixtral - 1.7 tok/s now (#3104)
* update NumNode.__hash__ to be hash(self.b) (#3105)
with this, `a:=NumNode(x) == b` implies `hash(a) == hash(b)`
* catch runtime error in search._time_program (#3106)
return inf if search encountered runtime errors.
* no exceptions in __del__ when module creation is failed in hip/cuda (#3107)
* failed test case due to cast resets shapetracker (#3109)
cast implicitly resets shapetracker and makes it contiguous (for disk tensor), which fails for Interpreted backend if inputs contain non-contiguous st.
* cleanup ops_disk type annotation and redundant str cast (#3110)
* minor cleanup of test_disk_tensor (#3112)
* add Tensor.var (#3114)
also updated MeanVarianceNormalization and made test_ops test tensors of var and std smaller
* move sample inside jit for beautiful_mnist (#3115)
also removed .realize() for jit functions since jit does it automatically now. a little more beautiful
* minor cleanups of onnx_ops (#3116)
* fix conversation: llama generates token not prob now (#3120)
* add device options for tests in multigpu (#3121)
* make DType a dataclass (#3111)
* remove np from DType
* convert to dataclass
* remove dunder hash, eq, ne overrides from ImageDType
* is dataclass required for PtrDType?
* fix GPU tests
* reduce lines
* revert changes to np
* minor cleanup
* hotfix: ptrdtype compare was broken
* move fromcpu out of lazy.py (#3122)
* move fromcpu out of lazy.py
* fix abstractions2
* remove numpy from device (#3123)
* remove numpy from device
* fix tests
* np item
* cleanups
* simplify with as_buffer
* no toCPU
* tinygradic
* cast to scalar
* remove numpy from ops_torch (#3124)
updated mnist test to cast label to int8 and avoid hacking cast issue of torch uint8
* Fix backward fn for `<` and `==` (#3037)
* fix no grad fn for < and ==
* remove 2 line breaks
* Remove deprecated autograd variable
---------
Co-authored-by: George Hotz <72895+geohot@users.noreply.github.com>
* separate try except blocks in onnx2torch in model benchmark (#3126)
exceptions can be raised from either model conversion or individual backend failed. openpilot on torch mps works, but does not work with torch cpu.
seperate the expcetion block so that the benchmark can inlcude torch mps for openpilot.
* update env_vars.md (#3127)
mostly removed deprecated ones. not clear how to maintain this especially for extra/examples
* update test_ptr_ne (#3130)
* remove np from metal graph (#3129)
* dtype fmt (#3132)
* dtype fmt
* three ways to access
* fix off-by-one error in st_equal (#3131)
* fix off by one error
* whitespace
* no numpy (#3134)
* fast resnet eval (#3135)
* fast resnet eval
* fix HIP multidevice graph
* neater expression for devices
* lines
* add decorator test
* remove LLVMOPT
* move ptx
* Update ops_cuda.py
---------
Co-authored-by: Christopher Milan <chrismilan@ucla.edu>
Co-authored-by: chenyu <chenyu@fastmail.com>
Co-authored-by: Yixiang Gao <yixiangg310573@gmail.com>
Co-authored-by: jxdv <virgoj@protonmail.com>
Co-authored-by: Francis Lam <flam@alum.mit.edu>
Co-authored-by: SnakeOnex <sheeproman@gmail.com>
Co-authored-by: nimlgen <138685161+nimlgen@users.noreply.github.com>
Co-authored-by: Jyotirmaya Mahanta <jyotirmaya.mahanta@gmail.com>
Co-authored-by: Guy Leroy <g.m.leroy@outlook.com>
Co-authored-by: Paul Gustafson <paul.gustafson@theambrusgroup.com>
* wmma: enable METAL half tensor cores and clean up cstyle
* revert simple_matmul rand changes and break line in tensor
* added metal fp16->fp32 tensor core
* add llama attention test for multigpu
* test fails
* kv cache trying to shrink on sharded axis
* mask None works for scale dot product
* kv cache seems to be working but scale dot product breaks
* scaled dot product works, but the last linear layer failed
* running into the reshape case where it could be wrong for multigpu
* making sure it was the reshape
* adding contiguous doesn't solve
* need to shard more properly
* remove reshape test
* minor adjustment to scale dot product attention test
* weights are sharded wrong
* continue fix new weight sharding
* clean up
* fix attention when start_pos is 0
* remove print
* add TODOs for the best mutigpu interface
* WebGL WIP
* 84% of ops passing test
* tests passing 100%
* Cleanup, refactor
* Shave off some lines
* Work on dtypes
* TestOps at 100% again
* Efficient net shaders compile in browser webgl2
* Compile all efficientnet shaders in browser
* Create empty textures for tensor buffers
* Run program. Up next weight loading
* Exported WebGL model working
* Add tests, refactor
* Explicit cast alu for GLSL
* Fix CI tests
* WebGL efficientnet demo
* Compile and run yolov8 in browser
* Fix imports
* Simplify yolo compile
* Fix bool*bool and cast cmplt to float
* More tests
* Do std tests pass on CI?
* Skip std tests on CI
* Remove explicit_cast_alu hack, and solve it in code_for_op
* Move to new dtype-less alloc api
* Remove local size hack: optimize local_size only if device has local
* Remove glsl.py, and move content to cstyle
* dont_use_locals in opts
* Fix dtype tests
* type_map in CStyleLanguage
* Make core changes smaller, cleaner, refactor export_model and demo
* Skip pad_slice
* Simplify: render_const, render_conditional
* solve bool alu for other binops, cleaner ops_webgl
* Fix noopt hack
* Remove some skipIfs
* WebGL image hack
* type_names is a better name
* global_max
* Fix dtype import
* Fix type_names -> type_map
* Fix lint
* Remove webgpu, back to 5k lines (#3040)
* remove webgpu
* max 5000 lines
* revert those to master
* retain that cstyle
---------
Co-authored-by: Ahmed Harmouche <ahmedharmouche92@gmail.com>
* add a failing test for LR scheduler when using multigpu
* fix calculation order and unnecessary tensor created for float
* min_lr is no longer tensor
* updated most dtype hacks in onnx_ops
* temporarily revert dequantizelinear change
* I think this is right...
* MORE FIXES WOOOO NEW DTYPE IS AWESOME
* ok
* oops missed a print
* half -> float32 for CI
* is npdtype
* some more
* fix if ordering
* more clean ups
* final cleanups
* casting to half not allowed
* k nvm
* revert ArgMax change
* only GPU
* llvm begone
* teeny tiny change
* fix: attempt to add cast tests
* try this
* fix dequantizelinear
* revert some stuff
* tests pass pls
* less lines in onnx_tests
* oops missed string tensor tests
* clean up
* try: revert default behavior changes
* fix: disabled Cast and Castlike tests
* docs: small changes
* fix: fixed isNaN op and enabled associated tests
* fix: forgot about float16
* done
* update disabled test
* gah missed another float16
* disable rest of failing tests
* rm extra line
* try...
---------
Co-authored-by: chenyu <chenyu@fastmail.com>
* cleanup llama apply_rotary_emb and other helpers
used ellipsis and other higher level tensor function.
disabled the half @ half -> half tensor core as it fails uop dtype checks
* keep hip 8x8->8 wmma
the correct condition is that PADTO cannot be applied to reduce axis, not Reduce.MAX in ops.
even for Reduce.SUM it's possible that the reduce axis had a div before, and the padded 0 became inf then sum over it is incorrect.
* these asserts should pass
* fix that assert
* ALU dtypes
* acc dtype for group_for_reduce
* cast image ALUs to the base dtype
* remove all casts from linearizer
* fix argmax
* fix multinomial
* fix __getitem__
* Revert "fix __getitem__"
This reverts commit 62ad719bfa5a2e1fcbfa931360f54897f8977602.
* fix MemBuffer outputs being wrong when there is an arange + ALU with a different dtype
eg. fancy slicing (int, float), bert embeddings (int, long)
this should be fixed in lazy instead of having to break the kernel
* cleanup argmax fix
* fix matmul in ints
cast in the end
* fix llama
* skip wrong hardcoded asts in the worlds dataset
* fix llama p2
* cleanup missing parts of the diff
---------
Co-authored-by: George Hotz <geohot@gmail.com>
* lazy rewrite, try 2
* min fix tests
* pass contig test
* put broken pads back
* move that to realize
* no contig child fixes array packing
* so wrong
* now that's correct
* base children
* fix bind issues
* disable to_image_idx
* fix tests
* that failure shouldn't break other tests
* more fixes
* fix torch
* skip failing tests in CI
* 1e-7
* half is broken
* 1e-6 margin of error
* bitcast renderers
* fast llama load
* make it one kernel
* regression testing p1: re-enable test_dtype for all backends
fix GPU
* regression testing p2: fuzz all possible cases against numpy
remove hancoded tests since the fuzzer covers them
* define ushort
* fix indent, probably need flake8 back for CI to catch
---------
Co-authored-by: George Hotz <72895+geohot@users.noreply.github.com>
* `global_load` and `global_store` using buffer dtype
* `UOps.PHI` in all dtypes
* `UOps.ALU` in all dtypes
* `UOps.CONST` & `UOps.DEFINE_ACC` in all dtypes
* -- endof implementation --
+tiny lint changes
* these tests require the fp16 extention
you can run them locally to confirm they're green: (GPT2 test is broken in master for mac, see [this](https://discord.com/channels/1068976834382925865/1069001075828469790/1177993277958533261)
`GPU=1 python3 -m pytest test/external/external_test_onnx_backend.py::OnnxBackendNodeModelTest::test_dequantizelinear_e4m3fn_float16_cpu test/external/external_test_onnx_backend.py::OnnxBackendNodeModelTest::test_max_float16_cpu test/external/external_test_onnx_backend.py::OnnxBackendNodeModelTest::test_min_float16_cpu test/models/test_real_world.py::TestRealWorld::test_llama test/models/test_real_world.py::TestRealWorld::test_gpt2 test/models/test_whisper.py test/test_specific_conv.py::TestSpecific::test_big_vec_mul`
skip the new test_linearizer_failures in CI GPU because of the fp16 extention
This passes on a real GPU since the extention is available:
`GPU=1 python3 -m pytest test/test_linearizer_failures.py::TestLinearizerFailures::test_failure_8`
see CI logs [here](https://github.com/tinygrad/tinygrad/actions/runs/6996590597/job/19032641427#step:14:644)
* these tests fail in CI due to segfaults and CPU crashes
To confirm they're green locally, you can run the following commands:
1. For the tests skipped in test_ops.py (note: CLANG is very slow)
`for var in GPU CUDA CLANG; do export $var=1; for test in test/test_ops.py::TestOps::test_slice_fancy_indexing_no_dim_collapse test/test_ops.py::TestOps::test_slice_fancy_indexing_dim_collapse_int test/test_ops.py::TestOps::test_slice_fancy_indexing_dim_inject_none test/test_ops.py::TestOps::test_slice_fancy_indexing_dim_inject_and_collapse; do python3 -m pytest $test; done; unset $var; done`
2. For the ONNX tests skipped in CLANG:
```
CLANG=1 python3 -m pytest test/external/external_test_onnx_backend.py::OnnxBackendNodeModelTest::test_ai_onnx_ml_array_feature_extractor_cpu \
test/external/external_test_onnx_backend.py::OnnxBackendNodeModelTest::test_gather_elements_0_cpu \
test/external/external_test_onnx_backend.py::OnnxBackendNodeModelTest::test_nllloss_NCd1_expanded_cpu \
test/external/external_test_onnx_backend.py::OnnxBackendNodeModelTest::test_sce_mean_weight_ii_3d_cpu \
test/external/external_test_onnx_backend.py::OnnxBackendNodeModelTest::test_gather_elements_1_cpu \
test/external/external_test_onnx_backend.py::OnnxBackendNodeModelTest::test_sce_NCd1_mean_weight_negative_ii_cpu \
test/external/external_test_onnx_backend.py::OnnxBackendNodeModelTest::test_nllloss_NCd1_weight_expanded_cpu \
test/external/external_test_onnx_backend.py::OnnxBackendNodeModelTest::test_nllloss_NCd1d2d3_none_no_weight_negative_ii_expanded_cpu \
test/external/external_test_onnx_backend.py::OnnxBackendNodeModelTest::test_nllloss_NCd1_ii_expanded_cpu \
test/external/external_test_onnx_backend.py::OnnxBackendNodeModelTest::test_sce_mean_weight_ii_4d_cpu \
test/external/external_test_onnx_backend.py::OnnxBackendNodeModelTest::test_sce_mean_weight_ii_3d_log_prob_cpu \
test/external/external_test_onnx_backend.py::OnnxBackendNodeModelTest::test_gather_elements_negative_indices_cpu \
test/external/external_test_onnx_backend.py::OnnxBackendNodeModelTest::test_sce_NCd1d2d3d4d5_mean_weight_log_prob_cpu \
test/external/external_test_onnx_backend.py::OnnxBackendNodeModelTest::test_sce_NCd1_mean_weight_negative_ii_log_prob_cpu \
test/external/external_test_onnx_backend.py::OnnxBackendNodeModelTest::test_nllloss_NCd1d2_no_weight_reduction_mean_ii_expanded_cpu \
test/external/external_test_onnx_backend.py::OnnxBackendNodeModelTest::test_sce_NCd1d2d3d4d5_mean_weight_cpu \
test/external/external_test_onnx_backend.py::OnnxBackendNodeModelTest::test_nllloss_NCd1d2d3d4d5_mean_weight_expanded_cpu \
test/external/external_test_onnx_backend.py::OnnxBackendNodeModelTest::test_nllloss_NCd1_mean_weight_negative_ii_expanded_cpu \
test/external/external_test_onnx_backend.py::OnnxBackendNodeModelTest::test_sce_mean_weight_ii_4d_log_prob_cpu \
test/external/external_test_onnx_backend.py::OnnxBackendNodeModelTest::test_nllloss_NCd1d2_with_weight_reduction_mean_expanded_cpu \
test/external/external_test_onnx_backend.py::OnnxBackendNodeModelTest::test_nllloss_NCd1_weight_ii_expanded_cpu \
test/external/external_test_onnx_backend.py::OnnxBackendNodeModelTest::test_nllloss_NCd1d2_with_weight_reduction_sum_ii_expanded_cpu \
test/external/external_test_onnx_backend.py::OnnxBackendNodeModelTest::test_nllloss_NCd1d2_with_weight_reduction_sum_expanded_cpu \
test/external/external_test_onnx_backend.py::OnnxBackendNodeModelTest::test_nllloss_NCd1d2_expanded_cpu \
test/external/external_test_onnx_backend.py::OnnxBackendNodeModelTest::test_nllloss_NCd1d2_reduction_sum_expanded_cpu \
test/external/external_test_onnx_backend.py::OnnxBackendNodeModelTest::test_nllloss_NCd1d2d3d4d5_none_no_weight_expanded_cpu \
test/external/external_test_onnx_backend.py::OnnxBackendNodeModelTest::test_nllloss_NCd1d2d3_sum_weight_high_ii_expanded_cpu \
test/external/external_test_onnx_backend.py::OnnxBackendNodeModelTest::test_nllloss_NCd1d2_reduction_mean_expanded_cpu \
test/external/external_test_onnx_backend.py::OnnxBackendNodeModelTest::test_nllloss_NCd1d2_with_weight_expanded_cpu
```
3. The LLVM test I skipped here is already [skipped in master for all backends](https://github.com/tinygrad/tinygrad/blob/master/test/external/external_test_onnx_backend.py#L186), I just made it more specific
`LLVM=1 python3 -m pytest test/external/external_test_onnx_backend.py::OnnxBackendNodeModelTest::test_dequantizelinear_e4m3fn_float16_cpu`
* Revert "these tests fail in CI due to segfaults and CPU crashes"
This reverts commit 15db57014381a4449d563526ac6c870e36257658.
* merge with cleanup-vectorized-hip-renders
* barely working HIP P1, ALU ops need a refactor?
* manage the fact that in HIP [half2 is actually an unsigned int vec](f921880387/hip/include/hip/amd_detail/amd_hip_fp16.h (L59)) and half is a totally different __half that [has an unsigned int element in it](f921880387/hip/include/hip/amd_detail/amd_hip_fp16.h (L50)) but can't be accessed [because it's private](f921880387/hip/include/hip/amd_detail/amd_hip_fp16.h (L86)). If you just do this:
```
half2 val0 = // ...
half val1 = // ...
```
then you can't do:
```
val0.x + val1 // error: use of overloaded operator '+' is ambiguous (with operand types 'unsigned short' and 'half' (aka '__half'))
```
* update the sign definition to avoid division by zero in all dtypes
* diff cleanup p1: why were these in the diff anyways
* less hacky HIP, enable CIFAR fp16 benchmark, test ops for HIP in CI!
add ALU ops overloads for HIP
this will make HIP max work
handle mod
Revert "handle mod"
This reverts commit 370fd4b3fbe99b6ae8cc293d005b106628205933.
update max to use hmax
add HIP GEP render logic
enable CIFAR fp16 benchmark
test ops for HIP
back to store as float because this only works for float4 grouping right now
test_ops for hip!!
always sign
* back to the sign we had before because we cant do a backward pass on a Less node
* remove old hacks
HIP compiling test_ops in CI takes ~9 mins, not doing it for now
new HIP ALUs
* reduce accs done right
* refactor to function
* no device hacks
hacks p2
the other way
* LLVM ALU ops
half, float and double are all float
update max
* update test_uops, cmplt is always a bool in the real linearizer. assertAlmostEqual is wrong when ret is bool
* cleanup LLVM wrong code
* dummy change for the CUDA install glitch
---------
Co-authored-by: George Hotz <72895+geohot@users.noreply.github.com>
* update cstyle renderers to take a dtype in code_for_op
* implement NEG for bools in LLVM
* update triton
---------
Co-authored-by: George Hotz <72895+geohot@users.noreply.github.com>
* cuda with gpuctypes
* hip gpuctypes
* graphs
* rename + linter happy
* use cpu_time_execution
* no ji in build_kernel_node_params
* remove hip_wrapper
* hip fix
* no arc
* smalle changes
* no clean moduke in cudacpu
* 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
* bring hip graph back
* share with metal
* fix linter
* remove hasattrs
* Update ops_hip.py
* hip wrapper does not use _buf
---------
Co-authored-by: George Hotz <72895+geohot@users.noreply.github.com>
* add name support
* use fetch in gpt2
* remove requests from main lib, networkx also optional
* umm, keep that assert
* updates to fetch
* i love the walrus so much
* stop bundling mnist with tinygrad
* err, https
* download cache names
* add DOWNLOAD_CACHE_VERSION
* need env.
* ugh, wrong path
* replace get_child
* remove force_wait
* refactor
* get rid of stupid ASTRunner
* fix del in diskbuffer
* BufferOps.FROM_UNDERLYING
* put offset in the rawbuffer
* fix bugs
* use exec
* autopad shapetracker for BEAM
* OptOps.PADTO
* skip that test for now
* correct padding reduce axis
* just 32
* avoid more than double the FLOPs
* cleanups
* test case
* no support for triton and llvm yet
* typos
* symbolic shape would not work
* cannot PADTO with MAX kernel
* advance db version
* no breaking change - don't advance db version
* is triton just python?
* Revert "is triton just python?"
This reverts commit 17e776c25587615e33a3634c2fb0bb8591ce65d4.
* Revert "Revert "is triton just python?""
This reverts commit 6c434c01e1c4b0ea0431ec18632cd859fb3cf260.
* support llvm
* is it really passing in CI only?
* update tests
* oh triton test passed
* simpler
* revert that, with a test
* check if st are the same
* Revert "check if st are the same"
This reverts commit d2a5eac110a5da1af82a2728c883779ef69c3cad.
* update the db version
* rebase artifact
* replace all _dtypen with dtype.vec(n)
fix: print works
* conceptul refactor of cstyle render_load logic
* linearizer GEP is explicit that its dtype is the scalar version of localtype
* vectorized global_store and load don't need a conditional
* beautiful mnist
* beautiful mnist example
* from tinygrad import Tensor
* more beautiful
* the jit is super core tinygrad
* globalcounters reset on jit run
* symlinks and exclude
* beautiful_cartpole
* evaluate is it's own function
* no symlinks
* more beautiful
* jit reset for double speed
* type hinting for JIT
* beautiful_mnist gets 98%
* beautiful_mnist < 4s with BEAM=2
* better cartpole
* use actor critic
* zero_grad got lost
* delete double relu
* stable cartpole with PPO
* beautiful_cartpole is more beautiful
* REPLAY_BUFFER
* beautiful stuff typechecks
* None support in shape
* hp tuning
* add back as_strided, move rebuilt mops to extra
* negative stride for ops_cpu
* Revert "negative stride for ops_cpu"
This reverts commit a13b6815ac31478d31ae71c26f4d4e4d274bf155.
* skip that
* style
* very close
* remove comment
* negative strides working
* almost everything passes
* calculate offset with list comprehension
* some cleanup
* got disk load working
* review suggestions
* fix after merge
* overlap working
* did it
* clean
* fixed disk load
* lint
* mypy
* removed as_strided
* trying without simplify
* added back simplify
* make sure expanding to smaller shape
* cleanup
* removed comment
* removed env file
* trying whisper test again
* onnx test sqlite issue
* working on test
* finished test
* eliminate unnecessary shrink-then-pad
* don't shrink buffer
* added strides check
* added to ci under linters
* switch issue
* allow symbolic stride
* removed .env
* isinstance
* adjust strides for double expand
* cleanup
* needed to add type hint for mypy
* set pythonpath
* metal indirect command buffers
* sub 1ms gpt
* metal batch exec is good
* remove whitespace
* input_replace
* fix ci
* useResources
* very simple cacheallocator
* update_stats
* fix CI
* minor
* remove that from jit
* refactor/ci: delete many `# type: ignore`
* replace `axis.__class__ is int` with `isinstance(axis, int)` to make mypy happy
* add `--warn-unused-ignores` to mypy flag
refs #2240
* ci: move `--warn-unused-ignores` flag to mypy config
refs #2240
* var_vals are global
* working with global ish
* better
* fix export model
* fix tests
* better kv cache
* does it run?
* use where for kvmask
* fix excessive var_vals
* fix import
* how does multigpu use this?
* llama kinda work
* faster and simpler
* cleanup
* fix conversation mode
* test cleanups
* fix one more test
* test cleanup
---------
Co-authored-by: George Hotz <geohot@gmail.com>
* move metal+clang to compile api
* all to the new style
* remove binary arg
* fix triton
* fixup tests
* fix clang
* diskcache is generic
* __wrapped__
* compile_gpu
* fix thneed
* keep the src in the ASTRunner
* lib
* move compile_gpu
* compile_gpu in device
* put compiler in astrunner
* test reverts
* triton compiler
* ugh, that too
* remove arm64, caching for cuda
* caching in llvm
* switch cache_compiled to new cache
* fix clang
* caching for metal
* fix pylint
* cleanups
* perf_counter and binary
* merge kernel and optimizer
* linearize is reentrant
* move global/local size
* clean up linearizer copy
* remove unneeded lin copies
* stop linearizing twice
* oops, that should be None
* Enable Multi-Output Export
* Add test
* Update examples and lint
* fix padding
* test ops
* dummy commit to rerun test
* revert cuda lint
* Enforce tuple/list of tensors
* subscripted generics
* put back webgpu test
* Re-enable WebGPU Efficientnet test
* stable diffusion < 324ms
* revert swap action
* fix tests due to more sum splitting
* REDUCEOP_SPLIT_THRESHOLD env var
* added from unaligned np test (#2134)
* align cpu buffer before copy into cl buffer (#2135)
* remove shelve from handcode_resnet50_opt.py (#2139)
* Add dictionary keys to reduce db size (#2131)
* work
* ignore beam cache
* dictionary keys are generic
* minor db cleanups
* fix baseline and extract dataset
* fix training
* log likelihood
* more lin to feats
* sts
* training policynet
* net sort of works
* dedup
* refactor, stupid new actions
* fix uops deduping
* BEAM_ESTIMATE
---------
Co-authored-by: chenyu <chenyu@fastmail.com>
Co-authored-by: imaolo <56898718+imaolo@users.noreply.github.com>
* 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
* Revert "disable flaky triton test"
This reverts commit 1e15fdaee7.
* Update test.yml
* check if has shared for matvec
* disable ocelot cache for triton
* disable ocelot cache
* disable ocelot cache
* pass shared to triton uops tests
* temporary debugs for CI crash
* Revert "temporary debugs for CI crash"
This reverts commit fee3ea96c818e83c19b935c2f8482e0ccc91a542.
* Revert "triton isn't tested, and allows this refactor (#2007)"
This reverts commit dea8bb0938.
* add runtime_args to every renderer, move triton local size override to runtime args
* Add binary to args, correct type returned
* update to new loops
* Update test.yml
* some cleanup
* move continue back
* more more more
* added to CI
* try
* try intentionally break some tests
* wtf
* del True for test
* yay tests broke, now pls no break
* try AGAIN
* gahy
* lol
* try
* move over constant
* moved over MORE
* move shrink over
* trailing lines
* try CUDA CI
* try again
* boom
* oops
* improved comments
* try: disable some flags and disable CUDA
* try breaking tests
* traceback has too much info so add --tb=no
* revert forced CI failure
* add comments and del unused imports
* oooooooo using regular debug try enable tb
* intentionally break tests
* added tb back. Maybe not too verbose
* strip whitespcae
* missed something
* Shape op int32 -> int64
* oops missed something
* add some types
* get rid of crazy 1 liners in pad op
* actually test Split this time LOL
* strip that whitespace
* create cache for q learning
* make linter happy
* global beam
* where it belongs
* bugfix
* ditch the kopt, use the beam
* faster lin and DEBUG=2 okay
* remove kopt, move search to features
* start compile2
* tweak
* why are there two more kernels?
* minor cleanups
* don't break onnx tests
* add __metadata__ support to safetensors
* no early realize in onnx
* cleanups
* bugfix
* clean up image type, add optimize
* opt to match old
* try that
* opt work
* run compile2
* optimizer
* prt more
* prerealize
* imp
* NOLOCALS works
* no locals means no locals
* support fractional globals
* all locals welcome
* int that
* cleanups
* show gemv regression
* clean up diff
* use idx for the cond
* nolocals
---------
Co-authored-by: Comma Device <device@comma.ai>
* start work on auto opt
* lin failure
* not beating hcopt
* greedy
* timing is fast
* codegen.search
* greedy search in handcode_opt
* track running gflops
* clean up those files
* no failure
* Allow multi-input model export
* Add model export unit test
* Fix efficientnet compilation
* Only run model export test on JIT supported devices
* Skip export model test if not EXPORT_SUPPORTED_DEVICE
* small changes
* expand in terms of substitute, directly expand g_idxs g_valid
* delete expand_ops
* don't compare using hash
* any instead of in
thanks gijskoning
Co-authored-by: Gijs Koning <gijs-koning@live.nl>
* support tc
* testing code
* no more create_rednode
* maxsize none in view/node
* oops
* undo
* typing
* oops
* oops
* lmao
* lmao
* add expand multi test
* Node.iter_idxs
* type
* type
* delete checks!
* clean up a little?
* expand_idx in symbolic
* un-golf
* play around with types >.>
* test_substitute and also remove an incorrect test?
* get rid of range
* Update symbolic.py
* split out view cache change
* split out flat components change
* reduce diff
* reduce diff
* add some float4 tests
* fix
---------
Co-authored-by: Gijs Koning <gijs-koning@live.nl>
* lazy cleanups
* ast functions take in LazyOps
* op instead of self.op
* _base for mops
* fix contiguous
* start schedule
* test_schedule
* fix openpilot
* more tests
* bugfix and test skip
* work
* make sure things get freed
* fix zerosized tensors
* fix failing test
* fix ceil and friends
* fix openpilot
* disable training
* disable test collectives
* init hip graph
* optimize args update
* cache symbolic in jit
* remove NOSTAT
* init BasicBatchExecutor
* symbolic infer cache per jit instance
* basicbatchexec is defualt for compiled
* batch_exec is taken from ASTRunner
* no infer cache
* batched execution of hip graph
* add comment about hip graph batches
* readable hip graph
* Move ops_triton to runtime and remove errors from deprecated code
* Remove deprecated AST Kernel
* Remove deprecated buffer
* Add TritonProgram
* Triton Buffer
* Use RawCUDABuffer
* triton_compile
* Added new parameter
* pass _buf to program
* remove deprecated include
* Added triton tests
* Deprecated includes removed
* remove double print
* Disable float4 support
* Disable float4 support
* variable load fix
* Track local size
* Add pycuda to triton dependencies
* Merge test.yml
* install cuda packages for testing
* merge double package install
* remove emulated from triton tests
* upscale local index to power of 2 and add masking
* cuda envs
* Add TernaryOps
* ConstOp loading
* proper function name
* remove deprecated variables
* get global program from name
* const ops match local shape
* Enable test_nn
* remove deprecated import
* fix linter error
* Add wait logic
* Add local size override
* accumulate local shapes instead of using max shape
* Merge triton tests into global tests
* fix envs in testing
* Old testing routine
* split file into renderer and program
* remove print and starting whitespace
* pretty ptx print on debug 5
* linter errors
* ignore triton saturation tests
* ignore test example
* remove pytorch cpu extra index
* Add triton to existing testing routine
* use triton tests
* disable cuda backend in triton tests
* use cudacpu in tests
* print used device
* Print device default
* Remove print
* ensure we are running triton backend
* update variable signatures
* update dtypes for load
* infinity render fixed
* limit global size
* negative infinity now properly rendered
* split chain with parentheses for and node
* Add option to disable shared memory, disable for triton
* missing import
* Properly index and mask conditional load
* use mask only if not loading a block pointer
* nan support
* fix symbolic tests to include chain split
* proper masking for stores
* Implemented bool dtype
* Add mod
* fix loads for variables with valid range
* merge triton with cuda runtime
* merge from master
* run triton tests with cuda
* Correct target when running from triton
* conftest with triton compiler config
* use triton nightly
* verbose tests for triton
* capture stdout
* fix function depth when exiting multiple loops
* add render valid function for readabilty
* fix mask for local loops
* add _arg_int32 datatype
* fix dims for conditional loads
* enable non float stores
* correct variable dtypes
* fix type for arg_int32
* remove junk
* Added get max function for range based var.max
* remove deprecated code
* Fix triton ptxas path
* Fix testing for CI
* clamp local size by max local size instead of always running max
* Disable matmul test in triton cpu
* rerun tests
* Disable broken test in triton cpu
* whitespace removed
* rerun tests again
* Disable TestSymbolicOps for triton
* update to new uops
* linter fix
* ignore test/extra
* linting fix
* Update tinygrad/renderer/triton.py
Co-authored-by: Gijs Koning <gijs-koning@live.nl>
* remove deprecated line
* quotes type fix
* linter
* Remove unnecesary lines
* UnaryOps.NEG
* dont define constants
* Linting fix
* Disable tests that are broken in ocelot
* remove trailing whitespace
* reduce line count
* linting fix
* update to new uast
* New looping style
* Update to new uast
* make AST runner work with triton
* linting fix
* set renderer var for testing
* disable local for ocelot
* reenable all tests for ocelot
* Pass shared to cuda
* Don't group if the backend doesn't support shared mem
* use working gpuocelot branch
* enable all tests
* enable local for ocelot
* cleanup
* Update test.yml
* update cache key
* reenable test symbolic and extra
* Update test.yml
* Revert "Update test.yml" (rerun tests)
This reverts commit 98c0630ee5da4379e5c6b2437a5145fe87058c35.
* Revert "fix symbolic tests to include chain split"
This reverts commit 22a9a4c9cd14d23735e6540c8d90ee005ac4ea17.
* Revert "split chain with parentheses for and node"
This reverts commit 7499a7004ef4db785d0cd05cf292fdeff65ca90d.
* use global size from linearizer
* rename newvar to dtype to match other renderers
* join program start lines
* simplify code that adds axis to local dims
* assign r[u] in ssa
* We no longer need to replace target in src
* we no longer need to cast indices to int by hand
* Update triton.py(rerun tests)
* Update triton.py(rerun tests)
* Update triton.py(rerun tests)
---------
Co-authored-by: Gijs Koning <gijs-koning@live.nl>
Co-authored-by: George Hotz <72895+geohot@users.noreply.github.com>
* 1
* 83 failed
* learning how git works
* lol idk
* zero shape aaaa
* space lol
* aaa
* test check
* haha
* fixed gather
* 73 failing
* 71 failing
* 68 failing
* added some debug
* fking resize
* lol
* 62 failing
* 58 failling fucking did nearest resize hell yeah
* clean up
* 56 failing
* janitor duty
* lol
* 53 failing
* hi mom
* 50 failing
* added linear interp, but coord_trans is wrong
* did lin interpolation woohoo
* 43 failing
* 40 failing
* temporary Gather fix
* 39 failing
* fixed slice onnxver<10
* 37 failing
* 35 failing
* excluded tests that use float64
* 32 failing with hacks
* added _batchnorm() for 3D 5D batchnorm, 29 failing
* changed ALLOWED_KERNEL_COUNT from 199 to 207
* added improved Gather op, reverted ALLOWED_KERNEL_COUNT commit
* support Round op
* added storage_order/indices maxpool, 27 failing
* support maxunpool, 25 failures
* support Gradient, 23 failures
* merged new where
* added Adam
* cleanups
* added Momentum and Nesterov Momentum
* added Adagrad
* support sequence_type, 20 failing
* ugh git
* I give up on cubic interp :D, 9 failing
* sexy 1 liner gather, much improved, wow
* polished gather to make it shine bright like a diamond
* clean 1 liner for gather
* improved readability of gather
* uhh
* clean up
* more clean up
* WHITEspace
* implemented SoftmaxCrossEntropyLoss op
* added comments and cleaned up if statements
* update
* thank based wozeparrot for pow and new GatherElements
* CPU and TORCH all pass | cast float64 -> float32 for all fromCPU()
* _nearest_gather() failing on yolo
* reverted ops_cpu change and added assert in Resize
* added comments for resize for multiple channels
* oops
* merge
* test
* switched np.pad to Tensor.pad for constant padding
* gah
* gah2
* sexy reflect pad with movementops -> add
* delete commented out lines
* edge mode pad sexy as well
* trying out model_benchmark
* revert gitignore change lol
* init
* Revert "init"
This reverts commit 682bf2073a8b4eca111596c67cf6ebd79f59e585.
* wrote cast workaround for CPU, CPU and TORCH all pass
* wrote cast workaround for CPU, CPU and TORCH all pass
* skipped tests w/ 0 shape for METAL and GPU
* excluded tests for CLANG, CPU, TORCH, CLANG pass
* fixed hacky ConvTranspose
* gotta figure out autopad
* UOps.STORE support cast bool -> float
* small fix for fast gather
* reverted 0 shape skipped tests
* oops missed a file
* added comment
* fixed slice op hack
* First commit to pr
* More trig ops
* More trig ops
* format
* isinf support
* More ops
* changed onnx_ops to use our new gather :D
* Det op bug fix
* rebase
* fixed some tests
* det broken and slow
* fixed compress to use new gather
* implemented argmax argmin
* support variable types in type_proto
* support Upsample and Identity sequence
* we support float64 now and tinygrad support automatic broadcasting
* added EyeLike op
* resize does support multiple channels now actually
* yolov8 onnx runs successfully
* added batch size 1
* oops
* finally fixed type_proto I think
* fixed some llvm bugs
* del whitespaces
* added ZenginU Format PR
* test
* oops
* added float64 exclude tests back
* more skipped tests
* try
* ok openpilot pass
* flake8 pass
* woooooohooo
* revert external_model_benchmark changes
* perf tested gather
* removed promote types from ops_cpu
* numerical errors from 1681 is fixed
---------
Co-authored-by: ZenginU <umutzengin00@gmail.com>
* Symbolic Shape JIT
update tests
2 variables symbolic ops, adding more tests
test passing
cleanup
* more test cases
* single flag
* review update
* jit attention one piece
* realize
* symbolic_jit test for cuda
* old artifact
* works with cuda gpu but failed ci
* CUDACPU
* feat: train cifar using multigpu
* feat: split eval batch across 5
* feat: cleaner allreduce
* feat: 93.88%
* feat: cleaner batch chunking from bert
* feat: cleaner grad sync
* feat: tinygrad argmax
* feat: make it work with different gpu counts
* feat: move some stuff into the normal __init__
* feat: autodetect gpu count
* feat: move import inside
* move assembly, assembly_ptx
* successful but broken rendering of ptx asm
* clear ins before render asm
* slightly less broken :')
* we needed thread syncs
* fix float16 loading, rounding modifiers and other casting stuff, passing casts_from_half
* Fix runtime_args for gpuocelot
* our casts were flipped on both ends
* more casting
* add ternary where op
* dealing with storing/loading bool
* add test for casting to bool from negative
* Fix args.valid on ConstOp
* add to CI, TODO: fix runtime_args for test_uops
* fix placement of runtime_args to work with lazy.Device
* undo ci changes so I can push
* fix lints
* start cleanup and fix things we broke fixing lints
* add checks for PTX specifc asm instructions
* revert added test -- doesn't pass on llvm
* skip tests for underflow,overflow
* another fix for how we're setting runtime args
* Less broken cleanup
* add to CI
* add more env variables for ci test
* fix ci to install pycuda for ptx
* ci: copy cuda test command
* cleanup
* assert to make sure we're actually running ptx in ci
* remove test assert
* move is_ptx arg
* move assembly, assembly_ptx back to extras
* fix imports
* initial merge fixes
* clear registers, fix UOps.LOAD with invalid value
* draft merge fixes
* remove prints
* quick lint and merge fixes
* cleanup
* remove PTXProgram wrapper
* final cleanup
* temp change for ci rerun
* ci rerun
* rollback ISA version
* try to run commavq
* fix 0 dim, start implementing new ops
- Implement EmbedLayerNormalization
- Implement Attention
* SkipLayerNormalization and FastGelu
* use original torch model, cast inputs
* fix some ops:
- properly do Cast
- Attention: bi- and unidirectional
- FastGelu: add bias before gelu
* cleanup onnx_ops.py
* add validation option to benchmark
* cleanup imports
* add checks incase onnx2torch implements ops in future
* run onnx instead of original torch
* just skip gpu on m1
* reactivate the other models
* check for strange params & squash whitespace
* cleanup
* fix causal mask Attention
* Range doesn't need int cast
* embedding vocab_counter same dtype as input
* no need to cast
* always validate, fix PosixPath ort
---------
Co-authored-by: George Hotz <george@comma.ai>
* testing new memops
* better debugging
* testing padded conv
* branching with load
* refactoring a bit
* first try
* fixing bugs
* fixing some
* eq
* eq2
* do not use x's
* working
* fixing imm
* getting things working
* refactor
* pow not working
* working except one
* refactor: one store mem
* refactor: global load
* refactor: imm
* refactor: cleaning
* fixing big offsets
* refactor with ci
* try ci
* typo
* another typo
* ubuntu default
* forgot git
* do i need git?
* missing packages
* adding python-dev
* with cache?
* buildx action
* buildx name issue?
* maybe now?
* python3
* newline warning
* maybe now
* i actually need this
* ci should work now
* improved caching
* fixing cache
* maybe now it will cache
* this
* testing cache
* trying again
* load
* missing platform
* caching gha
* testing cache
* full testing
* typo
* now?
* why
* adding checkout back
* bad formatting
* fixing convention issues
* supporting python
* adding CI flag
* testing all
* better comments
* adding debugging
* takes 12x longer
* does it output progress now?
* ignore models for speed
* fixing merge
* excluding conv_transpose2d
* only 2 test cuz is to slow
* another approach
* let's see
* faster duh
* my bad
* T_T
* typo
* sup
* with output?
* comment test
* comment test
* comment test
* :?
* no comment
* with cache
* back to normal
* testing that ci works
* back to passing
* trying again
* does it create another entry
* does it create another entry?
* build local
* hey
* Revert "excluding conv_transpose2d"
This reverts commit cc7348de03033e032f47d69caff174e2f1a7bfea.
* does it cache if done before?
* does it cache?
* done
* adding test ops
* bad formatting
* no need for this
* working static mem
* sum 1d
* add ndim
* better reg import
* fix stack
* back to np
* working except for softmax
* 5 failing
* no pogress
* remove keystone
* remove keystone
* testops passing
* cleanups
* more cleanup
* typo
* ci
* ci2
* cond import
* ci3
* ci4
* ci4
* ci5
* ci5
* ci6
* aligment
* test all
* correct test
* err read_unmapped
* passing test
* ignore for speed
* ignore for speed
* ci7
* cleanup
* remove docker
* fixing merge
* fixing bugs
* add skipload for const ops
* comments
* First merge to master: Renderer
* fix emulation
* passing all tests arm64
* cleaning
* fix handcoded binary
* cleaning
* fix errs
* fix runtime arg binary
* clean git diff
* fix and clean
* fixing metal test
* cleaning
* fix metal test
* ci ~8 min
* fix pylint and clang
* cache the files in ops_clang
---------
Co-authored-by: George Hotz <72895+geohot@users.noreply.github.com>
* feat: world
* feat: tests
* feat: no more backwards
* feat: recv into
* feat: whoops
* feat: test in ci
* feat: some debug logging
* feat: workflow naming
* feat: need to set pythonpath
* feat: just send to same device
* feat: allreduce
* feat: test
* feat: need contiguous
* feat: test in ci
* feat: exit with correct code
* feat: don't need that
* feat: opencl wait_for just doesn't work
* feat: synchronize on out
* feat: try?
* feat: try again?
* feat: add extra realizes
* feat: print
* feat: seed
* feat: tol
* feat: test ones and zeros
* feat: remove print
* feat: are you just flaky
* feat: seperate scatter and gather?
* feat: just try synchronizing
* feat: remove print again
* feat: bring back difference
* feat: no sync
* feat: revert that
* feat: back to wait_for
* fix: typo
* feat: world
* feat: tests
* feat: no more backwards
* feat: recv into
* feat: whoops
* feat: test in ci
* feat: some debug logging
* feat: workflow naming
* feat: need to set pythonpath
* feat: just send to same device
* add disk_tensor
* fix jit
* new baseline before whitening
* whitening through torch
* whiting done currently at 91.65%
* 91.99%
* clean up mixup and 92.3%
* clean up 92.30%
* 92.49% before searching for new hyper-parameters
* fix CI
* fix white space
* add whitening init in test
* refactor, update hyperpara, 92.72%
* converting whiting to tinygrad operation
* update CI kernels count for CIFAR
* add pad reflect
* add random crop 92.53%
* update hyperpara 93%
* 93.15% on docker container, need to refactor the assignment for hyper param
* print out weights and bias to be separated
* bias/non-bias params separated
* fix whitespace
* clean up
* refactor hyper-param with dict
* refactor lr schedular params
* fix whitespace
* fix cross entropy loss
* fix whitespace
* move opt hyp to hyp dict
* minor fixup
* adjust model, loss scaling
* 92.74% while using half of compute as before
* update hyp for cutmix
* random shuffle during batches
* clean up
* updating the model
* update ConvGroup
* disable gradients for batchnorm layer weights
* whitespace
* 93.92%
* clean up
* finally 94%git add .!
* rewrite whitening to remove dependency on torch
* whitespace
* remove dependency on torch, 93.91%
* back to 94.03%
* clean up
* update test_real_world
chenyu
George Hotz
Elias Wahl
Caleb Bunch
wozeparrot
geohotstan
Yoshinori Sano
terafo
Francis Lam
Ahmed Harmouche
Max-We
George Hotz
SnakeOnex
Yixiang Gao
qazal
Rory Clear
Nguyen Nguyen Phuong
Yixiang Gao
nimlgen
Matthias Kronberg
Davi Silva
Jake
forcefieldsovereign
rodfer
valar
Comma Device
Akshay Kashyap
Sean D'Souza
Szymon Ożóg
Roelof van Dijk
David Hou
kormann
Pavol Rusnak
Karan Handa
Umut Zengin
Ethan Sorrell
JaSpa99
Steven Anderson
Diogo
Cole Sutyak
Jacob Pradels
madt2709