RoPE length extrapolation with interpolation #7005

MaximumEntropy · 2023-07-10T17:29:11Z

What does this PR do ?

Add a one line overview of what this PR aims to accomplish.

Collection: NLP

Changelog

Added the ability to do length extrapolation with RoPE via interpolation.
Added a continue training script for GPT

Usage

You can potentially add a usage example below

# Add a code snippet demonstrating how to use this

Before your PR is "Ready for review"

Pre checks:

Make sure you read and followed Contributor guidelines
Did you write any new necessary tests?
Did you add or update any necessary documentation?
Does the PR affect components that are optional to install? (Ex: Numba, Pynini, Apex etc)
- Reviewer: Does the PR have correct import guards for all optional libraries?

PR Type:

New Feature
Bugfix
Documentation

If you haven't finished some of the above items you can still open "Draft" PR.

Who can review?

Anyone in the community is free to review the PR once the checks have passed.
Contributor guidelines contains specific people who can review PRs to various areas.

Additional Information

Related to # (issue)

Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca>

for more information, see https://pre-commit.ci

examples/nlp/language_modeling/megatron_gpt_continue_training.py

+        model = load_from_nemo(MegatronGPTModel, cfg, trainer, gpt_cfg, modify_confg_fn=_modify_config)
+    elif cfg.model.get("pretrained_checkpoint", None) is not None:
+        validate_checkpoint_loading_args(cfg.model.pretrained_checkpoint)
+        model = load_from_checkpoint_dir(MegatronGPTModel, cfg, trainer, gpt_cfg, modify_confg_fn=_modify_config)


examples/nlp/language_modeling/megatron_gpt_continue_training.py

+        model = load_from_nemo(MegatronGPTModel, cfg, trainer, gpt_cfg, modify_confg_fn=_modify_config)
+    elif cfg.model.get("pretrained_checkpoint", None) is not None:
+        validate_checkpoint_loading_args(cfg.model.pretrained_checkpoint)
+        model = load_from_checkpoint_dir(MegatronGPTModel, cfg, trainer, gpt_cfg, modify_confg_fn=_modify_config)


nemo/collections/nlp/modules/common/megatron/position_embedding/rotary_position_embedding.py

examples/nlp/language_modeling/conf/megatron_gpt_config.yaml

examples/nlp/language_modeling/megatron_gpt_continue_training.py

* sft with pi Signed-off-by: Evelina <ebakhturina@nvidia.com> * update values only if not None" Signed-off-by: Evelina <ebakhturina@nvidia.com> --------- Signed-off-by: Evelina <ebakhturina@nvidia.com>

Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca>

for more information, see https://pre-commit.ci

Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca>

…into sandeepsub/rope_interpolate

ekmb

Thank you!

Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca>

…into sandeepsub/rope_interpolate

yidong72 · 2023-07-11T18:20:18Z

examples/nlp/language_modeling/tuning/conf/megatron_gpt_sft.yaml

@@ -60,6 +60,8 @@ model:
  activations_checkpoint_num_layers: null # not used with 'selective'
  answer_only_loss: False # not used right now
  gradient_as_bucket_view: False
+  seq_len_interpolation_factor: null # if not None, seq_len_interpolation_factor will match the base model's value


add some explanation about how interpolation factor translate to longer sequences.
e.g. factor = 2, sequence length x 2

yidong72 · 2023-07-11T18:24:17Z

examples/nlp/language_modeling/megatron_gpt_continue_training.py

+from nemo.utils.model_utils import inject_model_parallel_rank
+
+
+def _modify_config(gpt_cfg, cfg, add_cfg_to_tree=False):


these modify config, load_from_nemo, load_from_checkpoint_dir, validate_checkpoint_loading_args functions are the same as in the SFT code. can we put them into an utility file?

They are almost the same not 100% identical. The issue is each one (SFT vs continued training) modifies some common attributes like data, optim, but also a few different things.

yidong72 · 2023-07-11T18:26:18Z

nemo/collections/nlp/modules/common/megatron/language_model.py

@@ -559,7 +562,9 @@ def __init__(
            assert 0 < rotary_percentage <= 1
            if rotary_percentage < 1:
                rotary_dim = int(rotary_dim * rotary_percentage)
-            self.rotary_pos_emb = RotaryEmbedding(rotary_dim)
+            self.rotary_pos_emb = RotaryEmbedding(


maybe not in this PR. we need to add seq_len_interpolation_factor for all the models that uses RoPe.

* Push changes Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca> * Fixes Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca> * add continue training script Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca> * [WIP] nonlinear interp Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca> * Fix Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca> * override encoder_seq_len Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca> * Remove nonlinear Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * sft with pi (#7006) * sft with pi Signed-off-by: Evelina <ebakhturina@nvidia.com> * update values only if not None" Signed-off-by: Evelina <ebakhturina@nvidia.com> --------- Signed-off-by: Evelina <ebakhturina@nvidia.com> * Address comments Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * Add info Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca> * Empty Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca> --------- Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca> Signed-off-by: Evelina <ebakhturina@nvidia.com> Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> Co-authored-by: Evelina <10428420+ekmb@users.noreply.github.com> Signed-off-by: Gerald Shen <geshen@nvidia.com>

* Add end_strings to SamplingParams Signed-off-by: Gerald Shen <geshen@nvidia.com> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci Signed-off-by: Gerald Shen <geshen@nvidia.com> * Add end_strings to megatron_gpt_inference.yaml Signed-off-by: Gerald Shen <geshen@nvidia.com> * Add end_strings to sampling params Signed-off-by: Gerald Shen <geshen@nvidia.com> * Remove extra_id_1 from default end_strings Signed-off-by: Gerald Shen <geshen@nvidia.com> * Fix require_grad typos (#6930) Signed-off-by: Sergii Dymchenko <sdym@fb.com> Signed-off-by: Gerald Shen <geshen@nvidia.com> * fix syntax error Signed-off-by: Gerald Shen <geshen@nvidia.com> * fix the mpt chatbot (#6957) (#6968) Signed-off-by: Yi Dong <yidong@nvidia.com> Co-authored-by: Yi Dong <43824965+yidong72@users.noreply.github.com> Signed-off-by: Gerald Shen <geshen@nvidia.com> * add support for max_total_length=4096 for 43b (#6763) * add support for max_total_length=4096 for 43b Signed-off-by: Zhilin Wang <wangzhilin12061996@hotmail.com> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci --------- Signed-off-by: Zhilin Wang <wangzhilin12061996@hotmail.com> Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> Signed-off-by: Gerald Shen <geshen@nvidia.com> * rnnt_greedy_decoding.py: typos? auto-repressively -> auto-regressively (#6989) Signed-off-by: Vadim Kantorov <vadimkantorov@gmail.com> Signed-off-by: Gerald Shen <geshen@nvidia.com> * Cache handling without input tensors mutation (#6980) (#6996) * Cache handling without input tensors mutation * Cleanup * Cleanup#2 * Cleanup#3 --------- Signed-off-by: Boris Fomitchev <bfomitchev@nvidia.com> Co-authored-by: Boris Fomitchev <borisfom@users.noreply.github.com> Co-authored-by: Somshubra Majumdar <titu1994@gmail.com> Signed-off-by: Gerald Shen <geshen@nvidia.com> * Hybrid conformer export (#6983) (#6995) * Implemented generic kv-pair setting of export_config from args * Hybrid conformer export * Hybrid decoder export * Cleanup * Changed from **kwargs * Docstring * Docs added * Stringify args * Added docs for ASR export configs * lowercase ctc --------- Signed-off-by: Boris Fomitchev <bfomitchev@nvidia.com> Co-authored-by: Boris Fomitchev <borisfom@users.noreply.github.com> Signed-off-by: Gerald Shen <geshen@nvidia.com> * Fixing an issue with confidence ensembles (#6987) (#7004) * Bug fix for the confidence ensembles * Relax constraints for the test --------- Signed-off-by: Igor Gitman <igitman@nvidia.com> Co-authored-by: Igor Gitman <igitman@nvidia.com> Signed-off-by: Gerald Shen <geshen@nvidia.com> * [TTS] Add cosine distance option to TTS aligner (#6806) * [TTS] Add cosine distance option to TTS aligner Signed-off-by: Ryan <rlangman@nvidia.com> * [TTS] Update aligner comments Signed-off-by: Ryan <rlangman@nvidia.com> --------- Signed-off-by: Ryan <rlangman@nvidia.com> Signed-off-by: Gerald Shen <geshen@nvidia.com> * Minor MPT-7B fixes and creation script update (#6982) * Initial commit of minor MPT-7B fixes Signed-off-by: Daniel Egert <degert@nvidia.com> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci --------- Signed-off-by: Daniel Egert <degert@nvidia.com> Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> Signed-off-by: Gerald Shen <geshen@nvidia.com> * Change Jenkins timeout (#6997) * change timeout Signed-off-by: ericharper <complex451@gmail.com> * change to 8 hours Signed-off-by: ericharper <complex451@gmail.com> --------- Signed-off-by: ericharper <complex451@gmail.com> Signed-off-by: Gerald Shen <geshen@nvidia.com> * remove hard coded input and output fields (#7008) * remove hard coded input and output fields Signed-off-by: arendu <adithya.r@gmail.com> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci --------- Signed-off-by: arendu <adithya.r@gmail.com> Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> Signed-off-by: Gerald Shen <geshen@nvidia.com> * RoPE length extrapolation with interpolation (#7005) * Push changes Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca> * Fixes Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca> * add continue training script Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca> * [WIP] nonlinear interp Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca> * Fix Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca> * override encoder_seq_len Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca> * Remove nonlinear Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * sft with pi (#7006) * sft with pi Signed-off-by: Evelina <ebakhturina@nvidia.com> * update values only if not None" Signed-off-by: Evelina <ebakhturina@nvidia.com> --------- Signed-off-by: Evelina <ebakhturina@nvidia.com> * Address comments Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * Add info Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca> * Empty Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca> --------- Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca> Signed-off-by: Evelina <ebakhturina@nvidia.com> Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> Co-authored-by: Evelina <10428420+ekmb@users.noreply.github.com> Signed-off-by: Gerald Shen <geshen@nvidia.com> * use proper config Signed-off-by: Gerald Shen <geshen@nvidia.com> * Add end_strings to SamplingParams Signed-off-by: Gerald Shen <geshen@nvidia.com> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci Signed-off-by: Gerald Shen <geshen@nvidia.com> * Add end_strings to megatron_gpt_inference.yaml Signed-off-by: Gerald Shen <geshen@nvidia.com> * Add end_strings to sampling params Signed-off-by: Gerald Shen <geshen@nvidia.com> * Remove extra_id_1 from default end_strings Signed-off-by: Gerald Shen <geshen@nvidia.com> * fix syntax error Signed-off-by: Gerald Shen <geshen@nvidia.com> * use proper config Signed-off-by: Gerald Shen <geshen@nvidia.com> --------- Signed-off-by: Gerald Shen <geshen@nvidia.com> Signed-off-by: Sergii Dymchenko <sdym@fb.com> Signed-off-by: Yi Dong <yidong@nvidia.com> Signed-off-by: Zhilin Wang <wangzhilin12061996@hotmail.com> Signed-off-by: Vadim Kantorov <vadimkantorov@gmail.com> Signed-off-by: Boris Fomitchev <bfomitchev@nvidia.com> Signed-off-by: Igor Gitman <igitman@nvidia.com> Signed-off-by: Ryan <rlangman@nvidia.com> Signed-off-by: Daniel Egert <degert@nvidia.com> Signed-off-by: ericharper <complex451@gmail.com> Signed-off-by: arendu <adithya.r@gmail.com> Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca> Signed-off-by: Evelina <ebakhturina@nvidia.com> Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> Co-authored-by: Sergii Dymchenko <kit1980@gmail.com> Co-authored-by: Gerald Shen <geshen@nvidia.com> Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Yi Dong <43824965+yidong72@users.noreply.github.com> Co-authored-by: Zhilin Wang <wangzhilin12061996@hotmail.com> Co-authored-by: Vadim Kantorov <vadimkantorov@gmail.com> Co-authored-by: Boris Fomitchev <borisfom@users.noreply.github.com> Co-authored-by: Somshubra Majumdar <titu1994@gmail.com> Co-authored-by: Igor Gitman <igitman@nvidia.com> Co-authored-by: Ryan Langman <rlangman@nvidia.com> Co-authored-by: trias702 <25867060+trias702@users.noreply.github.com> Co-authored-by: Eric Harper <complex451@gmail.com> Co-authored-by: Adi Renduchintala <adithyare@nvidia.com> Co-authored-by: Sandeep Subramanian <sandeep.subramanian.1@umontreal.ca> Co-authored-by: Evelina <10428420+ekmb@users.noreply.github.com>

* Push changes Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca> * Fixes Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca> * add continue training script Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca> * [WIP] nonlinear interp Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca> * Fix Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca> * override encoder_seq_len Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca> * Remove nonlinear Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * sft with pi (#7006) * sft with pi Signed-off-by: Evelina <ebakhturina@nvidia.com> * update values only if not None" Signed-off-by: Evelina <ebakhturina@nvidia.com> --------- Signed-off-by: Evelina <ebakhturina@nvidia.com> * Address comments Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * Add info Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca> * Empty Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca> --------- Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca> Signed-off-by: Evelina <ebakhturina@nvidia.com> Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> Co-authored-by: Evelina <10428420+ekmb@users.noreply.github.com>

MaximumEntropy added 7 commits June 28, 2023 14:50

Push changes

b9287e3

Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca>

Fixes

0a4574b

Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca>

add continue training script

6e79e92

Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca>

[WIP] nonlinear interp

363831e

Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca>

Fix

95282d4

Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca>

override encoder_seq_len

21739f0

Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca>

Remove nonlinear

9cf880a

Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca>

MaximumEntropy requested a review from ekmb July 10, 2023 17:29

github-actions bot added the NLP label Jul 10, 2023

MaximumEntropy requested a review from ericharper July 10, 2023 17:29

[pre-commit.ci] auto fixes from pre-commit.com hooks

f170ed2

for more information, see https://pre-commit.ci

github-advanced-security bot found potential problems Jul 10, 2023

View reviewed changes

ekmb reviewed Jul 10, 2023

View reviewed changes

ekmb and others added 6 commits July 10, 2023 13:29

sft with pi (#7006)

5daffb7

* sft with pi Signed-off-by: Evelina <ebakhturina@nvidia.com> * update values only if not None" Signed-off-by: Evelina <ebakhturina@nvidia.com> --------- Signed-off-by: Evelina <ebakhturina@nvidia.com>

Address comments

9a6fab2

Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca>

[pre-commit.ci] auto fixes from pre-commit.com hooks

6d5f75e

for more information, see https://pre-commit.ci

Add info

f1d7cbe

Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca>

Merge branch 'sandeepsub/rope_interpolate' of github.com:NVIDIA/NeMo …

61ac17f

…into sandeepsub/rope_interpolate

Merge branch 'main' into sandeepsub/rope_interpolate

bc6fe06

MaximumEntropy requested a review from ekmb July 10, 2023 20:39

ekmb approved these changes Jul 10, 2023

View reviewed changes

MaximumEntropy added 2 commits July 10, 2023 19:58

Empty

895d8bb

Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca>

Merge branch 'sandeepsub/rope_interpolate' of github.com:NVIDIA/NeMo …

b60be05

…into sandeepsub/rope_interpolate

yidong72 reviewed Jul 11, 2023

View reviewed changes

MaximumEntropy merged commit 41d8477 into main Jul 11, 2023

MaximumEntropy deleted the sandeepsub/rope_interpolate branch July 11, 2023 19:56

dimapihtar mentioned this pull request Jul 29, 2023

RoPE length extrapolation with interpolation (#7005) #7121

Closed

8 tasks

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

RoPE length extrapolation with interpolation #7005

RoPE length extrapolation with interpolation #7005

MaximumEntropy commented Jul 10, 2023

ekmb left a comment

yidong72 Jul 11, 2023

yidong72 Jul 11, 2023

MaximumEntropy Jul 11, 2023

yidong72 Jul 11, 2023

		from nemo.utils.model_utils import inject_model_parallel_rank


		def _modify_config(gpt_cfg, cfg, add_cfg_to_tree=False):

RoPE length extrapolation with interpolation #7005

RoPE length extrapolation with interpolation #7005

Conversation

MaximumEntropy commented Jul 10, 2023

What does this PR do ?

Changelog

Usage

Before your PR is "Ready for review"

Who can review?

Additional Information

ekmb left a comment

Choose a reason for hiding this comment

yidong72 Jul 11, 2023

Choose a reason for hiding this comment

yidong72 Jul 11, 2023

Choose a reason for hiding this comment

MaximumEntropy Jul 11, 2023

Choose a reason for hiding this comment

yidong72 Jul 11, 2023

Choose a reason for hiding this comment