**Summary:** Previously, we call `assert_and_get_unqiue_device` once per node in both prepare and convert. This is expensive and unnecessary since the model device is the same across all nodes, so we should just call this once in the beginning and reuse the same model device across all the nodes.
**Test Plan:**
python test/test_quantization.py -k TestQuantizePT2E
Pull Request resolved: https://github.com/pytorch/pytorch/pull/159901
Approved by: https://github.com/jerryzh168
Over time, a large number of the existing type ignores have become irrelevant/unused/dead as a result of improvements in annotations and type checking.
Having these `# type: ignore` linger around is not ideal for two reasons:
- They are verbose/ugly syntatically.
- They could hide genuine bugs in the future, if a refactoring would actually introduce a bug but it gets hidden by the ignore.
I'm counting over 1500 unused ignores already. This is a first PR that removes some of them. Note that I haven't touched type ignores that looked "conditional" like the import challenge mentioned in https://github.com/pytorch/pytorch/pull/60006#issuecomment-2480604728. I will address these at a later point, and eventually would enable `warn_unused_ignores = True` in the mypy configuration as discussed in that comment to prevent accumulating more dead ignores going forward.
This PR should have no effect on runtime at all.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/142325
Approved by: https://github.com/Skylion007, https://github.com/janeyx99
Summary:
There were two problems with the HistogramObserver:
1. It does not work when someone passes a batch_size 1, tensor_size 1 data-point.
2. The Histogram doesn't seem to actually update if the range of the new x falls within the old one
These issues were both fixed.
On top of this, I greatly simplified the logic for the histogram updating. Now, it doesn't do the downsampling anymore, which saves a ton of memory and code. The accuracy can still be controlled with the upsampling ratio. This ratio was also too high for the accuracy we generally need here, I reduced the default for this.
Also the code is cleaner now, much easier to follow what's happening.
test_histogram_observer_same_inputs was likely wrong - If I pass 0s and 1s to my histogramobserver, I want them to actually count! The current test now thinks it's good to discard and ignore these values.
Test Plan: You can run the included tests.
Differential Revision: D58931336
Pull Request resolved: https://github.com/pytorch/pytorch/pull/129387
Approved by: https://github.com/jerryzh168
This commit enables float8_e5m2 and float8_e4m3fn dtypes in fx quantization and PT2E.
Motivation for using fp8 quantization instead of int8:
- it works better to run inference with the same datatype the model was trained with,
- fp8 can handle outliers better, which is one of the problems in LLMs activations.
The numerical recipe we want to use it for is fp8 inference:
- bgemms/gemms running in float8_e4m3fn,
- Per-Tensor-Quantization/Scaling,
- amax observer for measurement with input_backoff and weight_backoff.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/123161
Approved by: https://github.com/jgong5, https://github.com/jerryzh168
Adds a ruff lint rule to ban raising raw exceptions. Most of these should at the very least be runtime exception, value errors, type errors or some other errors. There are hundreds of instance of these bad exception types already in the codebase, so I have noqa'd most of them. Hopefully this error code will get commiters to rethink what exception type they should raise when they submit a PR.
I also encourage people to gradually go and fix all the existing noqas that have been added so they can be removed overtime and our exception typing can be improved.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/124570
Approved by: https://github.com/ezyang
Summary:
Even with changes in D55347133, it is still possible to OOM in histogram observer, because the size of allocated tensor also depends on *downsample_rate*.
For example, I still see OOM due to the attempt of allocating a 10GB+ histogram tensor in multi-task model.
To fix OOM issue better, we use *try-catch* clause to avoid OOM.
Empirically, we set the max size of a single histogram tensor size to 1 GB.
Test Plan: Test the change for Multi-Task model (depth + segmentation)
Differential Revision: D55567292
Pull Request resolved: https://github.com/pytorch/pytorch/pull/123309
Approved by: https://github.com/jerryzh168
Summary:
When `version` is missing in the metadata, use `min_val/max_val` as keys instead of `max_vals/min_vals`
## Reasons
1. It's been almost 2 years since this change D30003700, which means now most checkpoints are using the `max_val/min_val` keys
2. most checkpoints dumps using `model.state_dict()` don't have version info, which will lead a fake `missing keys` error when loading state_dict
Test Plan: CI
Differential Revision: D53233012
Pull Request resolved: https://github.com/pytorch/pytorch/pull/118659
Approved by: https://github.com/jerryzh168
Summary: This PR does 2 things:
1) Previously this would simply error, now it will ignore any
torch.inf values that it recieves. note: The code checks for torch.inf after
aminmax that way if there are no torch.inf values found, the perf is a
relatively unchanged
2) as mentioned in https://github.com/pytorch/pytorch/issues/100051,
values close to (but not quite at) the maximum/minimum float value could
overflow to infinity in the course of _adjust_min_max() (when this large
value would be multiplied by something in the middle of a calculation
that would otherwise result in a non inf value). This was fixed by
rearranging the order of operations for the lines in question without
altering the actual equations. Specifically, where operations in lines
1095, 1098 and 1100 have multiplication and division of large values,
its better to divide the two large values before multiplying, rather
than multiplying the two large values together (creating overflow) before dividing like it had been.
Test Plan: python test/test_quantization.py
TestObserver.test_histogram_observer_ignore_infinity
python test/test_quantization.py TestObserver.test_histogram_observer_handle_close_to_infinity
Reviewers:
Subscribers:
Tasks:
Tags:
Differential Revision: [D51489345](https://our.internmc.facebook.com/intern/diff/D51489345)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/103467
Approved by: https://github.com/andrewor14
Summary:
FX graph mode quant workflow and also pt2e flow relies on the `is_dynamic` flag in observer/quantizationspec to
convert an observer to dynamic quantization patterns (choose_qparams -> q -> dq), this PR added is_dynamic flag
for all observers so that it's possible to convert these observers to the pattern.
However, this dynamic quantization pattern (choose_qparams -> q -> dq) is actually only valid for MovingAverageObserver(averaging_constant=1)
for the computation before convert and after convert to match in the context of QAT. So we'll have some sanity
checks in other observers to make sure the is_dynamic is False.
Test Plan:
python test/test_quantization.py TestXNNPACKQuantizer.test_qat_dynamic_linear
Reviewers:
Subscribers:
Tasks:
Tags:
Differential Revision: [D51124725](https://our.internmc.facebook.com/intern/diff/D51124725)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/113288
Approved by: https://github.com/kimishpatel
Summary:
Previously we can only use native pytorch int dtypes that has corresponding quantized dtypes (e.g. quint8, qint8), this
PR removes this assumption in observers/fake_quants so that users can use all pytorch native dtypes (except for int64, we can add it later if need)
the main addition here is int16.
Test Plan:
python test/test_quantization.py TestQuantizePT2E
Reviewers:
Subscribers:
Tasks:
Tags:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/108453
Approved by: https://github.com/kimishpatel
Summary: D41985889 removed the cast to int for the inputs to torch.histc below, allowing the inputs to still be tensors. These tensors still have require_grad_ set to True, causing issues with the call to torch.histc.
Test Plan: buck2 test 'fbcode//mode/opt' fbcode//dper3/dper3/modules/low_level_modules/tests:stat_collector_test -- --exact 'dper3/dper3/modules/low_level_modules/tests:stat_collector_test - test_scripted_module (dper3.dper3.modules.low_level_modules.tests.stat_collector_test.StatCollectorTest_1)'
Differential Revision: D48800879
Pull Request resolved: https://github.com/pytorch/pytorch/pull/108232
Approved by: https://github.com/jerryzh168
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/99220
Previously we have two places we need to decide whether to insert observer or fake quantizer or not:
(1) input arguments of a node (2) output of a node, and right now we have separate code to do this
in this PR, the logic is unified in `_needs_obs_or_fq` helper function that takes the target_dtype and is_dynamic from previous output
and target_dtype and is_dynamic for the current Tensor we are looking at
let's use an example for conv node:
```
conv = convolution(input, weight, bias, ...)
```
let's say we have `input_node` object for argument `input`, and `conv_node` for `conv` node in the graph
(1) input arguments, e.g. `input`
the target_dtype/is_dynamic from previous output is the node that produces `input`, we get this from
input_node.meta["target_dtype_info"]["output_act_obs_or_fq"]
the taregt_dtype/is_dynamic for the current argument `input`, comes from conv_node.meta["target_dtype_info"]["input_act_obs_or_fq"]
similarly for weight it comes from conv_node.meta["target"]["weightobs_or_fq"] etc.
(2) output for conv node
the target_dtype/is_dynamic from previous output will be the floating point output from the fp32 convolution operator, so it
is hardcoded to be (torch.float, False), however, technically we should get this from node.meta["val"], but since the
current code base is shared by fx graph mode quantization and pytorch 2.0 export quantization, we cannot do that, we can revisit
after we decide to deprecate fx graph mode quantization
the target_dtype/is_dynamic for the current output comes from conv_node.meta["target_dtype_info"]["output_act_obs_or_fq"]
there is one caveat here about dynamic quantization, that is explained in the comment, so I won't repeat here
Note: also fixed some places in `_get_arg_target_dtype_as_input_to_node` and `_get_arg_target_is_dynamic_as_input_to_node` to make sure "not specified" == specifying a fp32 placeholder observer as well
Next: we can merge the two get target dtype and get is_dynamic function to reduce code duplication
Test Plan:
python test/test_quantization.py TestQuantizeFx
python test/test_quantization.py TestQuantizeFxOps
python test/test_quantization.py TestQuantizeFxModels
python test/test_quantization.py TestQuantizePT2E
python test/test_quantization.py TestQuantizePT2EModels
Imported from OSS
Differential Revision: D45198323
Pull Request resolved: https://github.com/pytorch/pytorch/pull/99767
Approved by: https://github.com/kimishpatel
Summary:
Previously we have two places we need to decide whether to insert observer or fake quantizer or not:
(1) input arguments of a node (2) output of a node, and right now we have separate code to do this
in this PR, the logic is unified in `_needs_obs_or_fq` helper function that takes the target_dtype and is_dynamic from previous output
and target_dtype and is_dynamic for the current Tensor we are looking at
let's use an example for conv node:
```
conv = convolution(input, weight, bias, ...)
```
let's say we have `input_node` object for argument `input`, and `conv_node` for `conv` node in the graph
(1) input arguments, e.g. `input`
the target_dtype/is_dynamic from previous output is the node that produces `input`, we get this from
input_node.meta["target_dtype_info"]["output_act_obs_or_fq"]
the taregt_dtype/is_dynamic for the current argument `input`, comes from conv_node.meta["target_dtype_info"]["input_act_obs_or_fq"]
similarly for weight it comes from conv_node.meta["target"]["weightobs_or_fq"] etc.
(2) output for conv node
the target_dtype/is_dynamic from previous output will be the floating point output from the fp32 convolution operator, so it
is hardcoded to be (torch.float, False), however, technically we should get this from node.meta["val"], but since the
current code base is shared by fx graph mode quantization and pytorch 2.0 export quantization, we cannot do that, we can revisit
after we decide to deprecate fx graph mode quantization
the target_dtype/is_dynamic for the current output comes from conv_node.meta["target_dtype_info"]["output_act_obs_or_fq"]
there is one caveat here about dynamic quantization, that is explained in the comment, so I won't repeat here
Note: also fixed some places in `_get_arg_target_dtype_as_input_to_node` and `_get_arg_target_is_dynamic_as_input_to_node` to make sure "not specified" == specifying a fp32 placeholder observer as well
Next: we can merge the two get target dtype and get is_dynamic function to reduce code duplication
Test Plan:
python test/test_quantization.py TestQuantizeFx
python test/test_quantization.py TestQuantizeFxOps
python test/test_quantization.py TestQuantizeFxModels
python test/test_quantization.py TestQuantizePT2E
python test/test_quantization.py TestQuantizePT2EModels
Reviewers:
Subscribers:
Tasks:
Tags:
Differential Revision: [D45167585](https://our.internmc.facebook.com/intern/diff/D45167585)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/99220
Approved by: https://github.com/kimishpatel
Summary:
Previously we assumed asymmetric quantization for dynamic quantization, this diff adds the support of symmetric quantization
for the input in dynamic quantization
Test Plan: buck run executorch/exir/tests:quant_lowering_custom_backend_pass -- "executorch.exir.tests.test_quant_lowering_custom_backend_pass.TestQuantLoweringCustomBackendPass.test_quantized_linear_dynamic"
Reviewed By: digantdesai
Differential Revision: D43134794
Pull Request resolved: https://github.com/pytorch/pytorch/pull/94854
Approved by: https://github.com/digantdesai
