Граф коммитов

1132 Коммитов

Автор SHA1 Сообщение Дата
Gustavo Rosa 8761385958 fix(hf_opt): Uses better bounds for search configuration. 2022-12-16 16:51:21 -03:00
Gustavo Rosa 853af076ba chore(hf_opt): Adjusts default parameters to OPT-350M. 2022-12-16 16:51:21 -03:00
Gustavo Rosa 9c65ca9ae4 chore(root): Bumps version for release. 2022-12-16 16:51:13 -03:00
Gustavo Rosa 98801d9e89 chore(root): Only triggers docs building when docs are changed. 2022-12-16 16:50:58 -03:00
Gustavo Rosa 446312b62b chore(root): Adjusts pull request template description. 2022-12-16 16:50:58 -03:00
Chris Lovett 6de3e1bcc1 fix: quantizer requirements 2022-12-16 16:50:58 -03:00
Chris Lovett 7911f72361 publish new quantizer. 2022-12-16 16:50:58 -03:00
Chris Lovett cef423b4a6 Make device runner more robust to Azure connection reset exceptions (#99)
* fix: snpe output folder cleanup script, more general.

* fix typo

* feat: give status.py an optional --name argument.

* switch to snpe-1.64.0.3605

* publish new docker image for quantization

* find snpe_target_arch dynamically.

* fix bug

* fix: make it possible for reset.py to reset everything.

* fix: remove unnecessary diagnostic output.

* add logging

* fix: open log file after setting working dir.

* fix: catch and continue after azure ConnectionResetErrors.

* fix versions

* fix for invalid cast from EntityProperty.

* remove noise logging
2022-12-16 16:50:58 -03:00
Chris Lovett 81617e3606 break down utilization by day. (#93) 2022-12-16 16:50:57 -03:00
Gustavo Rosa 118f76f459 chore(nlp): Updates the version of transformers to allow latest implemented models. 2022-12-16 16:50:57 -03:00
Gustavo Rosa 0bdcf3b72e fix(model_utils): Fixes the CodeGen analytical parameters formula. 2022-12-16 16:50:57 -03:00
Gustavo Rosa 4450e5e984 fix(hf_codegen): Fixes word embeddings being tied all the time. 2022-12-16 16:50:57 -03:00
Gustavo de Rosa 45b163623c chore(models): Adds support for CodeGen. 2022-12-16 16:50:57 -03:00
Gustavo de Rosa 4b49a359fa chore(models): Adds support for OPT. 2022-12-16 16:50:56 -03:00
Gustavo de Rosa 718677f4a0 fix(nlp): Fixes train.py restart not producing same loss from checkpoint. 2022-12-16 16:50:56 -03:00
Gustavo Rosa 5e29ac001d fix(config_base): Uses Caio's fix for all configuration classes. 2022-12-16 16:50:56 -03:00
Caio Mendes 20449f2a6f Fixed onnx export config 2022-12-16 16:50:56 -03:00
Gustavo Rosa f6085b4b6f fix(model_utils): Fixes analytical calculation of GPT-2 parameters. 2022-12-16 16:50:55 -03:00
Gustavo de Rosa 72bb0370f1 fix(metrics): Fixes text predict when using model from amlt. 2022-12-16 16:50:55 -03:00
Gustavo de Rosa e4337c780b fix(nlp): Reverts find_unused_parameters back to original value. 2022-12-16 16:50:55 -03:00
Gustavo Rosa 009b18051a chore(nlp): Adds support for bfloat16. 2022-12-16 16:50:55 -03:00
Caio Mendes a454858799 Reverts config mapping 2022-12-16 16:50:54 -03:00
Gustavo Rosa 8dc9c9f8bb fix(nlp): Fixes lm_iterators multi-file number of chunks. 2022-12-16 16:50:54 -03:00
Gustavo Rosa 9116d90d7e fix(hf_gpt2): Fixes the attribute map on HfGPT2Config. 2022-12-16 16:50:54 -03:00
Gustavo Rosa 537623a514 chore(nas): Improves outputs when perplexity or char accept rate pipeline gives errors. 2022-12-16 16:50:53 -03:00
Gustavo Rosa 28970500e1 chore(nlp): Ensures json is being loaded as utf-8 in Text Predict. 2022-12-16 16:50:53 -03:00
Gustavo Rosa aacc8cff89 chore(constraints): Changes measure_torch_perplexity to measure_torch_val_ppl. 2022-12-16 16:50:53 -03:00
Gustavo Rosa 57019465dc chore(constraints): Adds initial draft of Text Predict scoring during search. 2022-12-16 16:50:53 -03:00
Gustavo Rosa c53653cdd5 chore(evolution): Removes some unused ` training_` prefixes and adds scoring_file argument. 2022-12-16 16:50:52 -03:00
Gustavo Rosa 0a3318d5e6 chore(nlp): Adds scoring_file argument to the search. 2022-12-16 16:50:52 -03:00
Gustavo Rosa 3cff60d520 chore(nas): Changes `use_training_proxy` to `training_strategy`. 2022-12-16 16:50:52 -03:00
Gustavo de Rosa 865130acf9 fix(nlp): Sync workers before saving summaries.tsv to prevent multiple/null outputs. 2022-12-16 16:50:47 -03:00
Gustavo de Rosa 773b90ee55 fix(setup.py): Setting onnxruntime to 1.10.0 for now, because there were some API changes. 2022-12-16 16:50:39 -03:00
Gustavo de Rosa 66b8f9f200 fix(mem_transformer): Adds missing **kwargs to prevent additional keywords raising errors. 2022-12-16 16:50:33 -03:00
Gustavo Rosa d73903f279 fix(nlp): Improves logic to disable multiple dlloggers. 2022-12-16 16:50:26 -03:00
Gustavo de Rosa 141946f711 fix(config_base): Fixes Config not adding attribute_map information when exported as dict. 2022-12-16 16:50:25 -03:00
Gustavo de Rosa ed7fb59ea8 fix(config_base): Fixes Config class not being inheritted by PretrainedConfig. 2022-12-16 16:50:25 -03:00
Gustavo Rosa e898dd35f5 chore(nlp): Improves ONNX configuration classes. 2022-12-16 16:50:24 -03:00
Gustavo Rosa cd75dff11f fix(nlp): Fixes ONNX exports. 2022-12-16 16:50:20 -03:00
Gustavo Rosa 490f4ef5f3 chore(nas): Checking if cuda is available to allow training during search. 2022-12-16 16:50:10 -03:00
Gustavo Rosa 4b7ead5d59 fix(search_space_profiler): Asserts that it uses the training proxy. 2022-12-16 16:50:09 -03:00
Gustavo Rosa f3cf52c966 fix(nlp): Puts n_iter argument from search.py back to 10. 2022-12-16 16:50:09 -03:00
Gustavo Rosa c051151057 chore(nas): Enables command line arguments for training-related arguments. 2022-12-16 16:50:09 -03:00
Gustavo Rosa 4c972fceec fix(evolution): Fixes use_proxy to use_training_proxy. 2022-12-16 16:50:08 -03:00
Gustavo Rosa 03b288cedc chore(nas): Improves constraint_proxy nomenclature to use_training_proxy. 2022-12-16 16:50:08 -03:00
Gustavo Rosa 1152112c53 chore(nas): Removes unused hard-coded variables and improves legends of output plots. 2022-12-16 16:50:08 -03:00
Gustavo de Rosa 6d8d827ba5 fix(config_hf_gpt2): Fixes weight initialization range value. 2022-12-16 16:50:07 -03:00
Gustavo de Rosa cbad680dfa fix(analytical_params_formulae): Fixes keys not being found on dictionary. 2022-12-16 16:50:07 -03:00
Gustavo Rosa 8e38013e55 chore(nas): Adds option to choose for proxy when running search script. 2022-12-16 16:50:07 -03:00
Gustavo Rosa 2cbf3ed35d chore(nas_utils): Adds first draft of perplexity calculation during search. 2022-12-16 16:50:06 -03:00