ORT 1.19.2 Release: Cherry Pick Round 1 #21861
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
### Description Fixed #21775 ### Motivation and Context The dlls should be signed with Keycode CP-230012. The default is the test code sign.
### Description <!-- Describe your changes. --> ### Motivation and Context The training wheel size limit should be 400M
### Description <!-- Describe your changes. --> ### Motivation and Context <!-- - Why is this change required? What problem does it solve? - If it fixes an open issue, please link to the issue here. --> --------- Co-authored-by: Your Name <[email protected]>
Softmax (formula 1) is like the following:
```math
y_{i} = \frac{exp(x_{i})}{\sum_{i} exp(x_{i})}
```
After applying softmax, each element will be in the range of $(0, 1)$,
and the elements will add up to 1, so that they can be interpreted as
probabilities.
However, in language model, softmax has two issues:
* When all elements are -inf (for example, a whole row is masked when a
query token is padding), the result is not defined since exp(-inf)=0 and
divided-by-zero is encountered in the above formula.
* Why do we need normalize in a way that each query word are treated as
equal important (each row has sum equals to1)?
**Smooth Softmax** (formula 2) is a modified version that introduces a
smooth factor like the following:
```math
s_{i} = \frac{exp(x_{i})}{1+ \sum_{i} exp(x_{i})}
```
This formula could tackle the above two issues:
* It could handle the special case that all elements are -inf: the
result $s_{i}$ is 0 for every element in such case.
* Sum of all elements $\sum_{i}{s_{i}} = \frac{\sum_{i}{exp(x_{i})}}{1+
\sum_{i} exp(x_{i})}$ is in the range of (0, 1), so that we can train
the model to assign different importance to different query words.
Since exponential is prone to overflow or underflow, to get stable
result, formula 3 can be used:
```math
s_{i} = \frac{exp(x_{i} + c)}{exp(c)+ \sum_{i} exp(x_{i} +c)}
```
c can be any value in theory. In practical, choice of constant c shall
avoid $exp(c)$ and $exp(x_{i} +c)$ overflow (or underflow) at the same
time. A reasonable choice is like formula 4:
```math
c=-\max_{i} \{ x_i \}
```
or apply a constraint that c <=0 like the following formula 5:
```math
c=-\max(0, \max_{i} \{ x_i \})
```
The latter one (formula 5) ensures that $s_{i}$ will fallback to formula
2 when all elements are negative.
For CPU provider, smooth softmax is implemented in MLAS. CPU
implementation uses formula 5.
@wangyems implemented the smooth softmax in flash attention for CUDA,
which requires Ampere or newer GPU. The implementation of smooth softmax
in flash attention uses formula 4.
---------
Co-authored-by: Ye Wang
<!-- Describe your changes. --> refer to #21867 <!-- - Why is this change required? What problem does it solve? - If it fixes an open issue, please link to the issue here. --> --------- Co-authored-by: Your Name <[email protected]>
tianleiwu
force-pushed
the
prathikrao/cherry-pick-r1-1.19.2
branch
from
18002b5 to
37f896d
Compare
Enable causal in MultiHeadAttention cuda operator. All formats (Q_K_V_BSNH_BSNH_BSNH, Q_K_V_BSNH_BNSH_BNSH, Q_KV_BSNH_BSN2H and QKV_BSN3H) supports causal for now. Internally, casual will be dispatch to flash attention, efficient attention or unfused attention kernel. Currently, MultiHeadAttention has causal enabled in CPU ep, but not in CUDA ep. It could cause issues in onnx conversion, like some model can run in CPU but not in CUDA. Enable causal in CUDA will reduce the difference of support matrix of CPU/CUDA.
### Description Found a bug with num splits where the heuristic isn't being performed properly due to incorrect passing of sequence length to heuristic function. ### Motivation and Context We were experiencing significant performance issues with long sequence length with flash attention due to this misconfiguration.
…line (#21789) ### Description Upgrade pytorch_lightning to fix orttraining_amd_gpu_ci_pipeline ``` #24 1.838 WARNING: Ignoring version 1.6.0 of pytorch_lightning since it has invalid metadata: #24 1.838 Requested pytorch_lightning==1.6.0 from https://files.pythonhosted.org/packages/09/18/cee67f4849dea9a29b7af7cdf582246bcba9eaa73d9443e138a4172ec786/pytorch_lightning-1.6.0-py3-none-any.whl has invalid metadata: .* suffix can only be used with `==` or `!=` operators #24 1.838 torch (>=1.8.*) #24 1.838 ~~~~~~^ #24 1.838 Please use pip<24.1 if you need to use this version. #24 1.838 ERROR: Ignored the following versions that require a different python version: 1.14.0 Requires-Python >=3.10; 1.14.0rc1 Requires-Python >=3.10; 1.14.0rc2 Requires-Python >=3.10; 2.1.0 Requires-Python >=3.10; 2.1.0rc1 Requires-Python >=3.10 #24 1.838 ERROR: Could not find a version that satisfies the requirement pytorch_lightning==1.6.0 (from versions: 0.0.2, 0.2, 0.2.2, 0.2.3, 0.2.4, 0.2.4.1, 0.2.5, 0.2.5.1, 0.2.5.2, 0.2.6, 0.3, 0.3.1, 0.3.2, 0.3.3, 0.3.4, 0.3.4.1, 0.3.5, 0.3.6, 0.3.6.1, 0.3.6.3, 0.3.6.4, 0.3.6.5, 0.3.6.6, 0.3.6.7, 0.3.6.8, 0.3.6.9, 0.4.0, 0.4.1, 0.4.2, 0.4.3, 0.4.4, 0.4.5, 0.4.6, 0.4.7, 0.4.8, 0.4.9, 0.5.0, 0.5.1, 0.5.1.2, 0.5.1.3, 0.5.2, 0.5.2.1, 0.5.3, 0.5.3.1, 0.5.3.2, 0.5.3.3, 0.6.0, 0.7.1, 0.7.3, 0.7.5, 0.7.6, 0.8.1, 0.8.3, 0.8.4, 0.8.5, 0.9.0, 0.10.0, 1.0.0, 1.0.1, 1.0.2, 1.0.3, 1.0.4, 1.0.5, 1.0.6, 1.0.7, 1.0.8, 1.1.0, 1.1.1, 1.1.2, 1.1.3, 1.1.4, 1.1.5, 1.1.6, 1.1.7, 1.1.8, 1.2.0rc0, 1.2.0rc1, 1.2.0rc2, 1.2.0, 1.2.1, 1.2.2, 1.2.3, 1.2.4, 1.2.5, 1.2.6, 1.2.7, 1.2.8, 1.2.9, 1.2.10, 1.3.0rc1, 1.3.0rc2, 1.3.0rc3, 1.3.0, 1.3.1, 1.3.2, 1.3.3, 1.3.4, 1.3.5, 1.3.6, 1.3.7, 1.3.7.post0, 1.3.8, 1.4.0rc0, 1.4.0rc1, 1.4.0rc2, 1.4.0, 1.4.1, 1.4.2, 1.4.3, 1.4.4, 1.4.5, 1.4.6, 1.4.7, 1.4.8, 1.4.9, 1.5.0rc0, 1.5.0rc1, 1.5.0, 1.5.1, 1.5.2, 1.5.3, 1.5.4, 1.5.5, 1.5.6, 1.5.7, 1.5.8, 1.5.9, 1.5.10, 1.6.0rc0, 1.6.0rc1, 1.6.0, 1.6.1, 1.6.2, 1.6.3, 1.6.4, 1.6.5, 1.7.0rc0, 1.7.0rc1, 1.7.0, 1.7.1, 1.7.2, 1.7.3, 1.7.4, 1.7.5, 1.7.6, 1.7.7, 1.8.0rc0, 1.8.0rc1, 1.8.0rc2, 1.8.0, 1.8.0.post1, 1.8.1, 1.8.2, 1.8.3, 1.8.3.post0, 1.8.3.post1, 1.8.3.post2, 1.8.4, 1.8.4.post0, 1.8.5, 1.8.5.post0, 1.8.6, 1.9.0rc0, 1.9.0, 1.9.1, 1.9.2, 1.9.3, 1.9.4, 1.9.5, 2.0.0rc0, 2.0.0, 2.0.1, 2.0.1.post0, 2.0.2, 2.0.3, 2.0.4, 2.0.5, 2.0.6, 2.0.7, 2.0.8, 2.0.9, 2.0.9.post0, 2.1.0rc0, 2.1.0rc1, 2.1.0, 2.1.1, 2.1.2, 2.1.3, 2.1.4, 2.2.0rc0, 2.2.0, 2.2.0.post0, 2.2.1, 2.2.2, 2.2.3, 2.2.4, 2.2.5, 2.3.0, 2.3.1, 2.3.2, 2.3.3, 2.4.0) #24 1.838 ERROR: No matching distribution found for pytorch_lightning==1.6.0 ```
### Description Fix `Orttraining Linux Lazy Tensor CI Pipeline` - Remove unused import of `torch.onnx._internal.exporter`, whose path is changed in newer torch (pytorch/pytorch#132429). - Move import of `register_custom_op_symbolic` from `torch.onnx` into local function, which causes circular import when running `import torch.onnx` (at least in the CI environment).
### Description This change disables Abseil's symbolize functionality in Windows non-debug builds. ### Motivation and Context To solve #21826. Avoid having a dependency on dbghelp.dll.
baijumeswani
approved these changes
Aug 30, 2024
MaanavD
approved these changes
Aug 30, 2024
tianleiwu
approved these changes
Aug 30, 2024
edgchen1
approved these changes
Aug 30, 2024
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Approved cherry picks for ORT 1.19.2 release.