github/tinygrad
1
1
Fork 0
mirror of https://github.com/tinygrad/tinygrad.git synced 2026-01-09 15:08:02 -05:00
Code Issues Packages Projects Releases Wiki Activity

remove CAST_BEFORE_VIEW (#4152)

Browse Source 
* remove CAST_BEFORE_VIEW

testing perf, also this might have issue with assign?

* remove all
This commit is contained in:
chenyu
2024-04-13 01:05:08 -04:00
committed by GitHub
parent ebc94c9d6c
commit a7c6864260
4 changed files with 6 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
.github/workflows/benchmark.yml vendored
View File

@@ -53,7 +53,7 @@ jobs:
- name: Run GPT2 w HALF
run: JIT=1 HALF=1 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half.txt
- name: Run GPT2 w HALF/BEAM
run: JIT=1 HALF=1 BEAM=2 CACHELEVEL=0 CAST_BEFORE_VIEW=0 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half_beam.txt
run: JIT=1 HALF=1 BEAM=2 CACHELEVEL=0 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half_beam.txt
- name: Train MNIST
run: time PYTHONPATH=. TARGET_EVAL_ACC_PCT=97.3 python3 examples/beautiful_mnist.py | tee beautiful_mnist.txt
- name: Run 10 CIFAR training steps
@@ -129,7 +129,7 @@ jobs:
- name: Run GPT2 w HALF
run: CUDA=1 JIT=1 HALF=1 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half.txt
- name: Run GPT2 w HALF/BEAM
run: CUDA=1 JIT=1 HALF=1 BEAM=2 CACHELEVEL=0 CAST_BEFORE_VIEW=0 JIT_BATCH_SIZE=4 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half_beam.txt
run: CUDA=1 JIT=1 HALF=1 BEAM=2 CACHELEVEL=0 JIT_BATCH_SIZE=4 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half_beam.txt
- name: Train MNIST
run: time PYTHONPATH=. CUDA=1 TARGET_EVAL_ACC_PCT=97.3 python3 examples/beautiful_mnist.py | tee beautiful_mnist.txt
- name: Run 10 CIFAR training steps

2
test/external/verify_kernel.py vendored
View File

@@ -9,7 +9,7 @@ from tinygrad.features.search import time_linearizer
# Use this with the LOGKERN options to verify that all executed kernels are valid and evaluate to the same ground truth results
# Example for GPT2:
# 1) Run the model to log all kernels: `PYTHONPATH=. LOGKERN=/tmp/gpt2_kerns.txt JIT=1 HALF=1 BEAM=2 CACHELEVEL=0 CAST_BEFORE_VIEW=0 python3 examples/gpt2.py --count 10 --temperature 0 --timing` # noqa: E501
# 1) Run the model to log all kernels: `PYTHONPATH=. LOGKERN=/tmp/gpt2_kerns.txt JIT=1 HALF=1 BEAM=2 CACHELEVEL=0 python3 examples/gpt2.py --count 10 --temperature 0 --timing` # noqa: E501
# 2) Validate the kernel correctness: `PYTHONPATH=. python3 ./test/external/verify_kernel.py --file /tmp/gpt2_kerns.txt`
if __name__ == "__main__":

8
test/test_const_folding.py
View File

@@ -106,14 +106,12 @@ class TestMovedConstFolding(unittest.TestCase):
_check_ast_count(1, Tensor([1.0, 2, 3, 4]) * Tensor.ones(2).pad(((1, 1),)))
def test_cast_padded(self):
# NOTE: this is folded due to CAST_BEFORE_VIEW
_check_ast_count(0, Tensor.ones(4).pad(((1, 1),)).cast(dtypes.int16))
_check_ast_count(1, Tensor.ones(4).pad(((1, 1),)).cast(dtypes.int16))
np.testing.assert_equal(Tensor.ones(4).pad(((1, 1),)).cast(dtypes.int16).numpy(), [0, 1, 1, 1, 1, 0])
_check_ast_count(0, Tensor.full(4, fill_value=-1).pad(((1, 1),)).cast(dtypes.uint16))
np.testing.assert_equal(Tensor.full(4, fill_value=-1).pad(((1, 1),)).cast(dtypes.uint16).numpy(), [0, 65535, 65535, 65535, 65535, 0])
# not folded
_check_ast_count(1, Tensor.ones(4).pad(((1, 1),)).cast(dtypes.int64))
np.testing.assert_equal(Tensor.ones(4).pad(((1, 1),)).cast(dtypes.int64).numpy(), [0, 1, 1, 1, 1, 0])
_check_ast_count(1, Tensor.full(4, fill_value=-1).pad(((1, 1),)).cast(dtypes.uint16))
np.testing.assert_equal(Tensor.full(4, fill_value=-1).pad(((1, 1),)).cast(dtypes.uint16).numpy(), [0, 65535, 65535, 65535, 65535, 0])
class TestReduceOpsConstFolding(unittest.TestCase):
def test_const_sum(self):

3
tinygrad/lazy.py
View File

@@ -84,9 +84,6 @@ class LazyBuffer:
if self.device.startswith("DISK") and not bitcast: raise RuntimeError("attempted to cast disk buffer (bitcast only)")
if self.is_unrealized_unmasked_const() and not bitcast:
return create_lazybuffer(self.device, self.st, dtype, LoadOps.CONST, dtypes.as_const(self.base.arg, dtype))
# TODO: applying this makes gpt2 slower
if getenv("CAST_BEFORE_VIEW", 1) and dtype.itemsize <= self.dtype.itemsize and self != self.base:
return self.base.cast(dtype, bitcast)._view(self.st)
new_shape = self.shape
if bitcast and self.dtype.itemsize != dtype.itemsize:
if not self.device.startswith("DISK"): raise RuntimeError("shape changing bitcast only supported on DISK right now")