11,106 commits

This branch

This branch

All branches

Author	SHA1	Message	Date
chenyu	bad6adaf8c	add mixtral and 6 gpus cifar to tinybox ci (#3676) ... * add mixtral and 6 gpus cifar to tinybox ci * print total ram used at the end of loading	2024-03-10 18:25:31 -04:00
George Hotz	44a67bf783	constant folding (#3675) ... * constant fold * bool math * fix ptx	2024-03-10 14:47:24 -07:00
George Hotz	25aede6fd9	truncate for exec_alu (#3674)	2024-03-10 14:19:04 -07:00
Francis Lata	957ae9b594	Fix Tensor's __repr__ for printing out grad (#3673) ... * update check for Tensor's __repr__ with grad * add test for repr with grad bugfix	2024-03-10 17:04:29 -04:00
George Hotz	0f16729023	RDNA3: restore launch bounds (#3672) ... * bring launch bounds back * works * that second flag didn't do anything * fix linter	2024-03-10 10:27:52 -07:00
chenyu	d7452c2a20	clean up llvmir builder (#3671) ... ``` _block -> block builder._block.module -> builder.module var_dtype -> dtype ```	2024-03-09 21:19:36 -05:00
George Hotz	1143c62519	tensor.py touchups (#3667) ... * tensor.py touchups * put back	2024-03-09 16:12:20 -08:00
George Hotz	69ca7f7bf9	changes for teenygrad (#3665) ... * changes for teenygrad * upd * simpler test	2024-03-09 15:30:34 -08:00
Quentin Wach	89b8b5d549	Fix missing import. (#3666)	2024-03-09 14:55:23 -08:00
Maximilian Wolf	8ae85b2cf5	add inference_mode context manager with decorator support (#3621) ... * add inference_mode context manager with decorator support * change val to mode for train and inference_mode * fix wrong rename	2024-03-09 08:38:26 -08:00
Obada Khalili	b5cbf1792a	Fix Tensor.cumsum when axis of length 0 is selected (#3473) ... * fix Tensor.cumsum when axis of length 0 is selected * add cumsum regression test * define padding left size in a seperate line	2024-03-09 08:26:41 -08:00
chenyu	915f98791c	use custom KernelOptError in kernel opt (#3661) ... be more specific about invalid kernel opt, used that in test_linearizer_failures. make BEAM kernel search work even with assertion disabled. `BEAM=2 python3 -O examples/llama.py --temperature=0 --count=10 --prompt="Hello." --timing`	2024-03-08 15:36:16 -05:00
George Hotz	ac02e7347d	ptx timing vs cuda timing (#3659)	2024-03-08 10:17:49 -08:00
uuuvn	daa4034e80	No more metal flakiness (#3643)	2024-03-08 08:54:44 -08:00
chenyu	e25879d50e	don't get new var_val for the same ast in fuzz_linearizer (#3657) ... fixed result comparison for kernels with variables	2024-03-08 09:49:24 -05:00
chenyu	1130c73844	add FUZZ_NTH to fuzz_linearizer (#3656) ... * add FUZZ_NTH to fuzz_linearizer also update tests in test_linearizer_failures to not just run on METAL * update failures for HIP/HSA * test_failure_21 LLVM PADTO	2024-03-08 09:16:49 -05:00
David Hou	9f66dcf718	PolynomialDecayWithWarmup + tests (#3649) ... * working PolynomialDecayWithWarmup + tests....... add lars_util.py, oops * keep lars_util.py as intact as possible, simplify our interface * whitespace * clean up * clean up * asserts * test polylr for full resnet training run * add comment * rename * fix do_optim * don't cast lr * info * calculate from train_files * skip it	2024-03-07 18:53:36 -05:00
chenyu	57df8e8d82	update fuzz_linearizer (#3648) ... included non-reduce kernel and kernel with variables. green msg when everything passed it's possible that creating rawbufs failed due to memory error, included that in failure cases	2024-03-07 18:41:22 -05:00
chenyu	b282a45e39	fix direct store float4 with same vin (#3652) ... In a kernel that stores expanded value, the vin of float4 can come from same source, and we only remove once in that case.	2024-03-07 18:11:50 -05:00
chenyu	a66ffec6d3	update kernel dataset to exclude the disktensor ones (#3651) ... disk tensor load contains big offset and is not meant to be run by gpu. repro steps ``` time ./extra/optimization/generate_dataset.sh gzip /tmp/sops mv /tmp/sops.gz extra/datasets/ ```	2024-03-07 17:35:19 -05:00
chenyu	fcf4a5ccf2	fix example that calls Tensor.__bool__ (#3650) ... also removed `.cpu()` calls in mask_rcnn so `python3 examples/mlperf/model_spec.py` runs	2024-03-07 16:59:26 -05:00
George Hotz	6e50582e62	working to improve ptx (#3647) ... * working to improve ptx * fix compile fail	2024-03-07 12:39:31 -08:00
Zaffer	1853ec9a02	add tests for bfloat16 on HIP (#3638) ... * Fix bug in login functionality * Remove HSA backend test and add bfloat16 dtype tests that run in CI * Skip tests on HIPCPU * skip tests causing segfault on LLVM backend * Exclude bfloat16 tests causing segfaults in LLVM backend * move bf16 cast tests to only test on HIP	2024-03-07 10:45:36 -08:00
chenyu	0cef284aac	fix typing FlopCounter.flops can be sint (#3646)	2024-03-07 12:49:17 -05:00
chenyu	906cc3a69b	cleanup tests Device[Device.DEFAULT] is always Compiled (#3645)	2024-03-07 11:15:42 -05:00
qazal	bdd62c7fd8	make the bf16 include dynamic (#3642) ... * dynamic prefix * add common ones above these are common dtypes aesthetics * regression test fuzz it test * run in CI * use .append * faster	2024-03-07 10:31:35 -05:00
chenyu	4552248c84	fix Tensor.to preserves grad.data (#3636)	2024-03-06 21:44:49 -05:00
chenyu	d33311ebe0	remove parens of ALU if it has associative property (#3635) ... need to remove SUB since it's possible to have (const - (const - const)) in test/test_ops.py::TestOps::test_cos, in which case cannot remove the parens of children	2024-03-06 21:12:11 -05:00
chenyu	fe6b6e38c1	remove parentheses of GEP if it's from SSA (#3634) ... fixed some bracket nesting level exceeded maximum of 256 errors	2024-03-06 20:22:46 -05:00
David Hou	0afaf70d57	lars optimizer + tests (#3631) ... * 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>	2024-03-06 18:11:01 -05:00
chenyu	b2e92d44fa	skip METAL sin test in test_dtype_alu (#3633) ... revert this part of #3629. this is flaky	2024-03-06 17:29:19 -05:00
chenyu	8f10bfa2ff	ban __bool__ on Tensor (#3632) ... * ban __bool__ on Tensor avoid misuse * test case * fix tests * fix more tests	2024-03-06 17:12:35 -05:00
George Hotz	81baf3eed3	bring ptx back (#3623) ... * bring ptx back * ptx back * fix define var * fix a few bugs * bugfixes * fixes * fix llvm bug * fix test bug	2024-03-06 13:34:21 -08:00
chenyu	c270d54c32	update test_dtype_alu for METAL (#3629)	2024-03-06 14:55:19 -05:00
qazal	abc5f3a6a0	hip bf16 hotfix (#3630) ... * hip bf16 * remu dev mac * Revert "remu dev mac" This reverts commit 465069a0dc3c7f2045f3348b312a1dcbf1587acd. * skip disk tests in CI * bring float8 back	2024-03-06 11:42:30 -08:00
chenyu	bc2a13a5f7	test case to show clang and python doing math in double (#3628)	2024-03-06 13:49:03 -05:00
George Hotz	568353fa84	hotfix: bump line count to 6500	2024-03-06 07:52:18 -08:00
Elias Wahl	a1507c7fd4	Fix Tensor.dropout() with multigpu (#3619) ... * Tensor.rand with multilazybuffer * remove recursive + test * whitespace * another whitespace. Sorry * remove else * Conconicalize multidevice tuple + Remove src	2024-03-05 18:26:21 -05:00
Jungwan Woo	e5ee6bb2bd	fix outdated url in showcase doc (#3624)	2024-03-05 14:44:40 -08:00
George Hotz	8500265561	this mem fault still happening (#3620) ... * this mem fault still happening * smaller * that print doesn't work * overflows test * hip doesn't uses_ptr_arithmetic * only with locals * test overflow new name * it's not ptr arith * simpler * simple repro * old compiler * simpler * put that back	2024-03-05 10:39:32 -08:00
chenyu	3c3f846c45	tinybox benchmark with HSA (#3603) ... * tinybox benchmark with HSA * torch cuda init can fail * no TORCHCUDA * print torch version * LD_PRELOAD="/opt/rocm/lib/libhsa-runtime64.so"	2024-03-05 11:03:52 -05:00
George Hotz	f500be1313	out of bounds access caused by launch bounds (#3615) ... * lin overflow * remove launch bounds * remove launch bounds infra * oops, fix bufs type	2024-03-05 06:34:00 -08:00
qazal	eb83e2d3a0	decouple buffer mutability from cstyle (#3617) ... * buffer mutability as an arg * update test_uops	2024-03-05 06:20:59 -08:00
chenyu	3275260c98	Revert "test: add failing bfloat16 test case for metal backend (#3481)" (#3618) ... This reverts commit 1e12a2ae80.	2024-03-05 09:08:42 -05:00
Skosh	1e12a2ae80	test: add failing bfloat16 test case for metal backend (#3481) ... * test: add failing bfloat16 test case for metal backend * test: move bfloat 16 test to dtypes test	2024-03-05 08:44:54 -05:00
chenyu	957e9800f1	llama + beam to mac benchmark, full cifar to nvidia benchmark (#3612) ... would merge if it's also ~1 minute. btw why is gpt2 beam not slower in the first beam run?	2024-03-04 21:35:57 -05:00
chenyu	282bbd5acb	check the input length into argfix (#3610) ... * check the input length into argfix it's possible to overlook setting keyword for kwargs and argfix silently truncates input * add test	2024-03-04 19:50:17 -05:00
Elias Wahl	7db6dd725d	multilazybuffer fix (#3609)	2024-03-04 17:36:23 -05:00
chenyu	c3b8d285aa	cleanup uops (#3605) ... using `is` to compare with enums, remove long lines and slightly more compact	2024-03-04 11:03:14 -05:00
qazal	94679322a3	simpler float4 direct store and locals support (#3592) ... * swap vins instead * delete the upcast * leave it to remove_childless try 1 * Revert "leave it to remove_childless try 1" This reverts commit bf25e935f8. * try 2, simpler * Revert "try 2, simpler" This reverts commit d2472af711. * add note	2024-03-04 06:28:28 -08:00