-
Notifications
You must be signed in to change notification settings - Fork 1.6k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
adaption for moe models #2101
base: main
Are you sure you want to change the base?
adaption for moe models #2101
Commits on Sep 26, 2024
-
donghaoran committed
Sep 26, 2024 Configuration menu - View commit details
-
Copy full SHA for b75c001 - Browse repository at this point
Copy the full SHA b75c001View commit details
Commits on Sep 27, 2024
-
FIX: Change check if past_key_values is empty (huggingface#2106)
After transformers merged this PR: huggingface/transformers#33703 The bool of past_key_values (a Cache instance) would change from False to True in one of our checks. Use get_seq_length() method instead, which is consistent before and after that commit. I checked the tests with the new change for both transformers before and after that commit and they passed, so this change should be backwards compatible. Unrelated change: Mark X-LoRA scaling test as xfail-ing for now. This should be addressed in a separate PR. Marking it to xfail for now to get the original fix through CI.
Configuration menu - View commit details
-
Copy full SHA for c29810b - Browse repository at this point
Copy the full SHA c29810bView commit details
Commits on Sep 30, 2024
-
Configuration menu - View commit details
-
Copy full SHA for aa3bd8f - Browse repository at this point
Copy the full SHA aa3bd8fView commit details
Commits on Oct 1, 2024
-
FIX Refactor OFT, small changes to BOFT (huggingface#1996)
The previous OFT implementation contained a few errors, which are fixed now. Unfortunately, this makes previous OFT checkpoints invalid, which is why an error will be raised. Users are instructed to either retrain the OFT adapter or switch to an old PEFT version.
Configuration menu - View commit details
-
Copy full SHA for 2a80735 - Browse repository at this point
Copy the full SHA 2a80735View commit details
Commits on Oct 2, 2024
-
ENH: Improved attribute access for modules_to_save (huggingface#2117)
Resolves huggingface#2099 So far, if a module was wrapped due to modules_to_save, we handled access to the weight and bias attribute (albeit incorrectly in case of disabled adapters!). However, there could be more attributes than those that could be accessed, in which case we got an error so far. Instead of special properties, we now implement a generic __getattr__ method that can deal with any attribute. The implementation is a bit complex to take into account the way that torch.nn.Module handles __getattr__.
Configuration menu - View commit details
-
Copy full SHA for ae297f0 - Browse repository at this point
Copy the full SHA ae297f0View commit details -
FIX low_cpu_mem_usage consolidates devices (huggingface#2113)
See: huggingface/diffusers#9510 (comment) Right now, the low_cpu_mem_usage=True option does not consolidate the devices. E.g. when the model is on GPU and the state_dict on CPU, the adapter weight will be on CPU after loading, when it should be GPU. This fix ensures that the devices are consolidated.
Configuration menu - View commit details
-
Copy full SHA for ca8462b - Browse repository at this point
Copy the full SHA ca8462bView commit details -
TST Mark flaky X-LoRA test as xfail (huggingface#2114)
Currently, CI is failing constantly because one of the X-LoRA tests has become flaky lately, most likely caused by the transformers 4.45.0 release. Therefore, this test is now marked to non-strictly xfail. I cannot reproduce this error locally, neither on CPU nor GPU. It is thus unclear how to fix this test.
Configuration menu - View commit details
-
Copy full SHA for 534d361 - Browse repository at this point
Copy the full SHA 534d361View commit details -
ENH: Warn when from_pretrained misses PEFT keys (huggingface#2118)
After merging huggingface#2084, we now clean up the missing_keys when loading a PEFT adapter to remove all but the relevant keys (the fact that base model keys are missing is expected when loading a PEFT adapter). Since the presence of missing_keys now really means that something might have gone wrong during loading, we can now warn the user if they call PeftModel.from_pretrained. Note that load_adapter still does not warn, as here we return the load_result and users can already check, but for from_pretrained, they don't have that possibility.
Configuration menu - View commit details
-
Copy full SHA for d9d3059 - Browse repository at this point
Copy the full SHA d9d3059View commit details
Commits on Oct 3, 2024
-
FEAT: Adding exclude modules param(huggingface#2044) (huggingface#2102)
Allows to exclude target modules.
Configuration menu - View commit details
-
Copy full SHA for 8d9ecbe - Browse repository at this point
Copy the full SHA 8d9ecbeView commit details
Commits on Oct 7, 2024
-
Configuration menu - View commit details
-
Copy full SHA for e6f927b - Browse repository at this point
Copy the full SHA e6f927bView commit details -
FEAT: VeRA quantization using bitsandbytes (huggingface#2070) (huggin…
…gface#2076) VeRA can now be used with 4bit and 8bit bnb quantization.
Configuration menu - View commit details
-
Copy full SHA for 859fd88 - Browse repository at this point
Copy the full SHA 859fd88View commit details
Commits on Oct 8, 2024
-
Configuration menu - View commit details
-
Copy full SHA for 5e91b54 - Browse repository at this point
Copy the full SHA 5e91b54View commit details -
FEAT: Support torchao (huggingface#2062)
Supports torch AO quantization. Currently supported: - int8_weight_only - int8_dynamic_activation_int8_weight --------- Co-authored-by: Marc Sun <[email protected]>
Configuration menu - View commit details
-
Copy full SHA for 9918977 - Browse repository at this point
Copy the full SHA 9918977View commit details -
FIX: PiSSA now works with Conv1D layers (huggingface#2103) (huggingfa…
…ce#2104) Transpose weight matrix based on fan_in_fan_out condition in PiSSA initialization. Co-authored-by: Yang Su <[email protected]>
Configuration menu - View commit details
-
Copy full SHA for a724834 - Browse repository at this point
Copy the full SHA a724834View commit details
Commits on Oct 9, 2024
-
FIX Type annoations in vera/bnb.py (huggingface#2139)
The file was missing the from __future__ import annotations part. As this code is only running nightly with GPU, the normal CI missed this omission.
Configuration menu - View commit details
-
Copy full SHA for 3b314cc - Browse repository at this point
Copy the full SHA 3b314ccView commit details -
ENH Make PEFT configs forward compatible (huggingface#2038)
Right now, loading a PEFT config saved with a more recent PEFT version than is currently installed will lead to errors when new arguments are added to the config in the newer PEFT version. The current workaround is for users to manually edit the adapter_config.json to remove those entries. With this PR, PEFT will make an attempt at removing these unknown keys by inspecting the signature. The user will be warned about these removed keys. This should generally be a safe measure because we will generally not introduce new config settings that change the default behavior. However, if a non-default is used, this could lead to wrong results. This is mentioned in the warning. While working on the tests, I also converted the unittest.TestCase to a normal pytest test in order to be able to use pytest fixtures. I also plan on adding the PEFT version to the adapter_config.json in the future. This will allow us to better handle compatibility issues in the future. As adding that new key to all PEFT configs could cause a lot of disruption, I want to get this PR in first to ensure forward compatibility. Note that this new mechanism will not help anyone using a PEFT version < 0.14.0, so this will be a slow transition.
Configuration menu - View commit details
-
Copy full SHA for 85e3202 - Browse repository at this point
Copy the full SHA 85e3202View commit details -
FIX Raise mixed adapter infer with missing adapter (huggingface#2090)
PEFT allows mixed batch adapter inference, i.e. when predicting, the same batch can use different adapters by passing the adapter_names argument. However, when users pass an adapter name that does not correspond to any of the existing adapters, these samples are currently being ignored (i.e. just the base model output is used). This is unexpected and can easily lead to errors, e.g. when users mistype the name of an adapter. This PR fixes this issue by checking all the existing adapter names first and comparing them to the adapter_names that the user passed. If there are unexpected entries, an error is raised. Due to this fix, an error in the test test_mixed_adapter_batches_lora_merged_raises was discovered and promptly fixed.
Configuration menu - View commit details
-
Copy full SHA for 8efa0cb - Browse repository at this point
Copy the full SHA 8efa0cbView commit details -
FIX Prompt learning with latest transformers error (huggingface#2140)
The error in PEFT is occurring after this transformers change: huggingface/transformers#33870 Now, in our tests, some model_kwargs no longer necessarily contain past_key_values, resulting in a KeyError. We now account for this possibility. Affected models were opt and gpt2.
Configuration menu - View commit details
-
Copy full SHA for 1eab9bd - Browse repository at this point
Copy the full SHA 1eab9bdView commit details
Commits on Oct 10, 2024
-
Configuration menu - View commit details
-
Copy full SHA for 5758a7e - Browse repository at this point
Copy the full SHA 5758a7eView commit details -
FIX TST NaN issue with HQQ GPU test (huggingface#2143)
This test calculates the correlation coefficient of HQQ model outputs. Although the model outputs are finite, the resulting matrix contains NaNs. Casting the outputs from 16 to 32 bit precision resolves the issue.
Configuration menu - View commit details
-
Copy full SHA for 0aa7e3a - Browse repository at this point
Copy the full SHA 0aa7e3aView commit details -
FIX Bug in target module optimization if suffix (huggingface#2144)
Solves the following bug: huggingface/diffusers#9622 (comment) The cause for the bug is as follows: When we have, say, a module called "bar.0.query" that we want to target and another module called "foo_bar.0.query" that we don't want to target, there was potential for an error. This is not caused by _find_minimal_target_modules directly, but rather the bug was inside of BaseTuner.inject_adapter and how the names_no_target were chosen. Those used to be chosen based on suffix. In our example, however, "bar.0.query" is a suffix of "foo_bar.0.query", therefore "foo_bar.0.query" was *not* added to names_no_target when it should have. As a consequence, during the optimization, it looks like "query" is safe to use as target_modules because we don't see that it wrongly matches "foo_bar.0.query".
Configuration menu - View commit details
-
Copy full SHA for c925d0a - Browse repository at this point
Copy the full SHA c925d0aView commit details
Commits on Oct 11, 2024
-
Bump version to 0.13.2.dev0 (huggingface#2145)
After the patch release of PEFT v0.13.2, let's bump the dev version of PEFT to v0.13.3.dev0 so that it stays ahead (the bugfix from the patch release is already contained in the main branch).
Configuration menu - View commit details
-
Copy full SHA for 749b924 - Browse repository at this point
Copy the full SHA 749b924View commit details
Commits on Oct 12, 2024
-
donghaoran committed
Oct 12, 2024 Configuration menu - View commit details
-
Copy full SHA for 669ce90 - Browse repository at this point
Copy the full SHA 669ce90View commit details