Previously, we made backwards graph compilation lazy to avoid paying
for compilation if the user didn't actually end up using the backwards
graph. This was useful in the old days when a lot of things in Inductor
didn't work and we could bypass errors this way.
However, this has a bad implication for dynamic shapes: the backwards
graph compilation can trigger extra guards, which are too late to
install in the Dynamo context if we wait until backwards is being run.
So in this PR I move us back to compiling backwards graph immediately
if we capture any SymInts for backwards.
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/104971
Approved by: https://github.com/Chillee
Fix cpp wrapper failure on TorchBench model `hf_Reformer` with `randn`:
```
random_rotations = torch.randn(rotations_shape, device=vectors.device, dtype=vectors.dtype)
```
For cpp wrapper, when `kwargs` is not empty, for `OpOverloadPacket` kernel, we need to know the exact overload schema to handle the `kwargs` properly when calling the cpp kernel: including finding the correct order of the kwargs and getting the default value for optional args without provided value when calling the function (`layout` in the above case).
The current support in this PR is conservative and we'll extend the functionality in subsequent PRs.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/104575
Approved by: https://github.com/jgong5, https://github.com/desertfire
This allows `ops.minimum` and `ops.maximum` to be hoisted for indirect indexing
into direct indexing expressions. I also add support to the cpp printer for
Min/Max and fix the triton printer to support multi-argument Min/Max.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/105020
Approved by: https://github.com/lezcano
This PR handles inference. Will do similar thing for training later.
Some manual testing results shows this can improve inference perf by 2-3% (absolute improvement not relative one).
- convmixer: 4.285x -> 4.309x
- resnet50: 2.170x -> 2.203x
The PR is built upon freezing. Since without freezing, the weight input for a conv node may not be a parameter directly but be the output of precision converting ops. It's so much easier to implement this PR after freezing.
Commands
```
TORCHINDUCTOR_FREEZING=1 python benchmarks/dynamo/timm_models.py --backend inductor --amp --performance --only convmixer_768_32 --inference
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/103642
Approved by: https://github.com/eellison
This PR decouples the logic necessary to compute bounds on variables
from the logic that uses this info to perform the strenght analysis on
int64 variables. While doing so, it tries to minimize the number of
attributes of the class in favour of local variables.
This class is now accessible from any `LoopBody` object.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/100549
Approved by: https://github.com/eellison
Introduces two higher order operators
* run_and_save_rng_state - Saves the current rng state and then runs the op.
* run_with_rng_state - Runs the op with the rng state supplied as an input
Ideally, we would like to use torch.compile for these operators. But currently the plan is to introduce these operators at the partitioner level, obviating the need to support them fully through the torch.compile stack. To ensure that we have good enough debugging with minifiers, we have ensure that they work with make_fx. In future, we can move on torch.compile.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/102934
Approved by: https://github.com/jansel, https://github.com/zou3519
This PR just contains some mild gyrations necessary to appease mypy.
However, it is not complete; there are a number of legitimate bugs
and mistyping that I need to work out before I can actually turn this
on.
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/100712
Approved by: https://github.com/ngimel
Added helper functions to match nodes in the graph that are decomposed from their source (leaf modules, or functional ops), as a result of dynamo tracing.
`get_source_partitions(graph: torch.fx.Graph, wanted_sources: List[Any]) -> Dict[Any, SourcePartition]`
Args:
* graph: The graph we want to partition
* wanted_sources: List of sources of nodes that were decomposed from this source. This can be a function (ex. torch.nn.functional.linear) or a leaf module type (ex. torch.nn.Linear)
Returns:
* Dictionary mapping sources (ex. torch.nn.modules.linear.Linear) to a list of SourcePartitions that correspond to the list of nodes that were flattened from a module of that type.
```
@dataclass
class SourcePartition():
# Nodes in a particular partition
nodes: List[Node]
# Module type
module_type: Type
# Nodes in the graph that are needed as inputs to the partition
input_nodes: List[Node] = field(default_factory=list)
# Nodes in the partition that are being used by nodes outside of the partition
output_nodes: List[Node] = field(default_factory=list)
# Parameters that are being used
params: List[str] = field(default_factory=list)
```
Example:
Original:
```
x -> linear -> linear -> relu -> linear
```
Traced graph:
```
.graph():
%arg0 : [#users=1] = placeholder[target=arg0]
%_param_constant0 : [#users=1] = get_attr[target=_param_constant0]
%t_default : [#users=1] = call_function[target=torch.ops.aten.t.default](args = (%_param_constant0,), kwargs = {})
%_param_constant1 : [#users=1] = get_attr[target=_param_constant1]
%addmm_default : [#users=1] = call_function[target=torch.ops.aten.addmm.default](args = (%_param_constant1, %arg0, %t_default), kwargs = {})
%_param_constant0_1 : [#users=1] = get_attr[target=_param_constant0]
%t_default_1 : [#users=1] = call_function[target=torch.ops.aten.t.default](args = (%_param_constant0_1,), kwargs = {})
%_param_constant1_1 : [#users=1] = get_attr[target=_param_constant1]
%addmm_default_1 : [#users=1] = call_function[target=torch.ops.aten.addmm.default](args = (%_param_constant1_1, %addmm_default, %t_default_1), kwargs = {})
%relu_default : [#users=1] = call_function[target=torch.ops.aten.relu.default](args = (%addmm_default_1,), kwargs = {})
%_param_constant2 : [#users=1] = get_attr[target=_param_constant2]
%t_default_2 : [#users=1] = call_function[target=torch.ops.aten.t.default](args = (%_param_constant2,), kwargs = {})
%_param_constant3 : [#users=1] = get_attr[target=_param_constant3]
%addmm_default_2 : [#users=1] = call_function[target=torch.ops.aten.addmm.default](args = (%_param_constant3, %relu_default, %t_default_2), kwargs = {})
return [addmm_default_2]
```
Result of `get_module_partitions`:
```
{<class 'torch.nn.modules.linear.Linear'>: [
ModulePartition(nodes=[_param_constant0, t_default, _param_constant1, addmm_default], module_type=<class 'torch.nn.modules.linear.Linear'>, input_nodes=[arg0], output_nodes=[addmm_default], params=["_param_constant0", "_param_constant1"]),
ModulePartition(nodes=[_param_constant0_1, t_default_1, _param_constant1_1, addmm_default_1], module_type=<class 'torch.nn.modules.linear.Linear'>, input_nodes=[addmm_default], output_nodes=[addmm_default_1], params=["_param_constant0_1", "_param_constant1_1"]),
ModulePartition(nodes=[_param_constant2, t_default_2, _param_constant3, addmm_default_2], module_type=<class 'torch.nn.modules.linear.Linear'>, input_nodes=[relu_default], output_nodes=[addmm_default_2], params=["_param_constant2", "_param_constant3"])],
<class 'torch.nn.modules.activation.ReLU'>: [
ModulePartition(nodes=[relu_default], module_type=<class 'torch.nn.modules.activation.ReLU'>, input_nodes=[addmm_default_1], output_nodes=[relu_default], params=[])]}
```
Also added helper function to check if two module partitions are connected:
`check_subgraphs_connected(subgraph1: SourcePartition, subgraph2: SourcePartition) -> bool`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/98628
Approved by: https://github.com/cccclai
Command to run max autotune baseline:
```
TORCHINDUCTOR_MAX_AUTOTUNE=1 time python benchmarks/dynamo/torchbench.py --backend inductor --amp --performance --only ${MODEL_NAME} --training --batch-size-file $(realpath benchmarks/dynamo/torchbench_models_list.txt)
```
Command to do coordinate descent autotuning:
```
TORCHINDUCTOR_COORDINATE_DESCENT_TUNING=1 TORCHINDUCTOR_CACHE_DIR=/tmp/torchinductor_shunting_coordesc TORCHINDUCTOR_PERSISTENT_REDUCTIONS=0 TORCHINDUCTOR_MAX_AUTOTUNE=1 time python benchmarks/dynamo/torchbench.py --backend inductor --amp --performance --only ${MODEL_NAME} --training --batch-size-file $(realpath benchmarks/dynamo/torchbench_models_list.txt)
```
Explanation of the envvars show up on the command:
```
- TORCHINDUCTOR_COORDINATE_DESCENT_TUNING=1 : enable coordinate descent tuning
- TORCHINDUCTOR_PERSISTENT_REDUCTIONS=0 : disable persistent reduction. Need do this so we can tune RBLOCK for reductions
- TORCHINDUCTOR_MAX_AUTOTUNE=1: enable max autotune
- TORCHINDUCTOR_CACHE_DIR=/tmp/torchinductor_shunting_coordesc : use a separate cache dir for coordinate descent tuning. Optional.
```
Here are my experiments results for around 40 torchbench models: https://docs.google.com/spreadsheets/d/1G7i2whIf8Yu-HhN_WovNxwcE-iFDSAw4x3NK4uL4XhI/edit#gid=0
Some highlights
- We improve 2.2% further upon max-autotune on average (geomean)
- timm_resnest benefits most from coordinate descent tuning. There is 1.07x speedup
- We have descent speedup on transformer models
- BERT_pytorch: 1.056x
- timm_vision_transformer: 1.04x
- hf_Bert: 1.030x
- For resnet models, it looks like we have less gain as model get larger. My guess is larger model spend more time on mm/conv, so our tuning for pointwise/reduction helps less
- resnet18: 1.021x
- resnet50: 1.014x
- resnet152: 1.005x
This kind of coordinate descent autotuning can give us 'upper bound' of the gain we can get for tuning configs for pointwise/reduction. On the other hand, by spot checking, we roughly double the compilation time compared to max-autotune. Next steps can be
- we disable persistent reduction in coordinate descent autotune (it's still enabled in baseline) so we can tune RBLOCK for reduction. We can also try to use autotune to pick persistent reduction or not.
- pick good config without benchmarking (e.g. Natalia mentioned checking register spill)
- try the idea on matmul so we know what's the potential there.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/97203
Approved by: https://github.com/ngimel
Patterns based on https://github.com/pytorch/pytorch/pull/94729 mainly as a forcing function for implementing joint graph replacements.
Up until now, we had two places to do pattern matching
1) Pre-grad has janky infra (graph not normalized or functional), but is
desirable for many types of passes where you want your change to
affect grad formulas.
2) Post-grad has good infra, but cant change grad formulas.
This PR adds a third place to do pattern matching: the joint
forward+backwards graph. The idea is to take the patterns and lower
them to a joint graph and replace both the forwards+backwards before
we partition them. This allows us to do something similar to pre-grad
transforms, but run after normalization and functionalization.
Note that we don't seem to have kernels for all of these patterns, some get decomposed in the dispatcher.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/97741
Approved by: https://github.com/Chillee
Following metrics should be helpful:
- percent of time GPU is busy
- percent of time various category of kernels (e.g. pointwise/reduction triton kernel) takes
- percent of time each individual kernel takes compared to total wall time of the benchmark
This PR add those.
Example result from hf_Bert infernece graph:
```
== triton_pointwise category kernels ==
Kernel Self CUDA TIME (ms) Count Percent
------------------------------ --------------------- ------- ---------
triton_poi_fused_gelu_6_0d1d 0.48154 12.0 5.52%
triton_poi_fused_clone_1_0d1d2 0.29011 24.0 3.33%
triton_poi_fused_clone_2_0d1d2 0.17417 12.0 2.00%
triton_poi_fused_clone_4_0d1d2 0.10797 12.0 1.24%
Total 1.05379 12.08%
== triton_persistent_reduction category kernels ==
Kernel Self CUDA TIME (ms) Count Percent
------------------------------ --------------------- ------- ---------
triton_per_fused__softmax__to_ 0.97188 12.0 11.14%
triton_per_fused_add_native_la 0.37401 24.0 4.29%
triton_per_fused_gelu_native_l 0.02 1.0 0.23%
triton_per_fused_add_embedding 0.01718 1.0 0.20%
Total 1.38307 15.86%
== unknown category kernels ==
Kernel Self CUDA TIME (ms) Count Percent
------------------------------ --------------------- ------- ---------
ampere_fp16_s16816gemm_fp16_12 2.24514 24.0 25.74%
ampere_fp16_s16816gemm_fp16_25 1.39796 49.0 16.03%
void cutlass::Kernel<cutlass_8 1.36093 1.0 15.61%
ampere_fp16_s16816gemm_fp16_64 0.74591 12.0 8.55%
ampere_fp16_s16816gemm_fp16_12 0.61989 12.0 7.11%
Memset (Device) 0.024 12.0 0.28%
void at::native::(anonymous na 0.01543 2.03 0.18%
void at::native::vectorized_el 0.00011 0.03 0.00%
Total 6.40937 73.49%
Percent of time when GPU is busy: 101.44%
```
Note: the output shows total time GPU is busy is larger than total wall time. We measure total wall time disabling profiling while measure GPU time enabling profiling, that may distort the measurement a bit? But I assume the effect is not too large assuming the profiler mostly increase CPU time (rather than GPU).
## interesting usages
1. I pick a model that cudagraphs improve perf significantly like densenet121 and run the tool on it's forward graph. It's no surprise that quite a lot of time GPU is idle:
```
(Forward graph) Percent of time when GPU is busy: 32.69%
Total wall time 17.307 ms
```
Its backward graph has less percent of GPU idle time, but it's still high:
```
(Backward graph) Percent of time when GPU is busy: 46.70%
Total wall time 17.422 ms
```
2. I profile a subset of torchbench models and plot a table to show the percent of execution time for pointwise/reduction/persistent_reduction/unknown_category . Since I plan to explore using coordinate descent tuner to improve reduction, those models with high percent of time spending on reduction should be good caididates (e.g. resnet50, mobilenet_v2 ).
NOTE: a same model appears twice. The first rows is for the fwd graph and the second for the bwd graph. We profile different graphs for a model separately.
```
benchmark_name pointwise_percent reduction_percent persistent_reduction_percent unknown_category_percent GPU_busy_percent wall_time_ms
----------------------- ------------------- ------------------- ------------------------------ -------------------------- ------------------ --------------
resnet18 19.73% 7.86% 4.81% 41.25% 73.65% 2.549ms
resnet18 18.59% 7.13% 3.35% 67.35% 96.41% 3.467ms
resnet50 29.57% 22.13% 2.07% 51.68% 105.46% 6.834ms
resnet50 26.42% 15.27% 0.94% 59.68% 102.31% 13.346ms
vgg16 26.23% 0.00% 0.00% 74.20% 100.43% 18.212ms
vgg16 15.63% 5.61% 0.10% 79.42% 100.75% 33.485ms
BERT_pytorch 28.62% 4.82% 14.88% 33.32% 81.64% 7.162ms
BERT_pytorch 14.43% 13.41% 18.19% 49.24% 95.27% 10.395ms
densenet121 11.89% 2.14% 3.86% 16.36% 34.25% 16.531ms
densenet121 10.37% 2.06% 4.09% 31.46% 47.98% 16.934ms
hf_Bert 23.94% 0.00% 29.88% 46.09% 99.90% 7.766ms
hf_Bert 11.65% 10.54% 20.26% 61.66% 104.11% 11.892ms
nvidia_deeprecommender 42.92% 0.00% 0.00% 56.75% 99.67% 3.476ms
nvidia_deeprecommender 31.36% 3.44% 0.46% 65.20% 100.45% 3.872ms
alexnet 30.99% 0.00% 0.00% 69.16% 100.14% 3.169ms
alexnet 24.41% 4.83% 0.17% 71.09% 100.50% 4.709ms
mobilenet_v2 29.21% 27.79% 2.49% 44.00% 103.49% 10.160ms
mobilenet_v2 17.50% 15.05% 1.06% 69.68% 103.29% 20.715ms
resnext50_32x4d 18.96% 9.28% 2.31% 28.79% 59.33% 5.899ms
resnext50_32x4d 18.48% 11.01% 1.86% 53.80% 85.14% 7.167ms
mnasnet1_0 19.07% 14.52% 3.01% 35.43% 72.03% 6.028ms
mnasnet1_0 14.17% 12.00% 1.87% 67.56% 95.60% 9.225ms
squeezenet1_1 38.56% 0.00% 1.77% 56.21% 96.53% 2.221ms
squeezenet1_1 21.26% 7.57% 1.05% 67.30% 97.18% 4.942ms
timm_vision_transformer 17.05% 0.00% 18.80% 65.79% 101.64% 9.608ms
timm_vision_transformer 9.31% 9.07% 10.32% 73.25% 101.96% 16.814ms
```
## how to use
`python {compiled_module_wrapper.py} -p`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/97723
Approved by: https://github.com/jansel
data type: float32
Input size: torch.Size([64, 4, 128, 128])
single socket (32cores):
```
Before: bernoulli 0.001327775239944458 s dropout 0.0014216173489888509 s
After: bernoulli 0.0002424612840016683 s dropout 0.00039757410685221353 s
```
single core:
```
Before: bernoulli 0.04154032731056213 s dropout 0.04382548745473226 s
After: bernoulli 0.006143261671066284 s dropout 0.0065830423831939695 s
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/97002
Approved by: https://github.com/jgong5, https://github.com/jansel
