MLM adaptation and Multitask Finetuning #284

lintangsutawika · 2022-06-12T10:06:42Z

No description provided.

thomasw21

Had a quick look at MLMDataset. Thanks again for tackling this issue! Will read more when I have more time.

thomasw21 · 2022-06-18T21:26:48Z

megatron/arguments.py

                            'They are used for span masking in the T5 model')
    group.add_argument('--seq-length', type=int, default=None,
                       help='Maximum sequence length to process.')
+    group.add_argument('--input-length', type=int, default=None,


This feels like a weird argument given that you already set seq-length

As far as I've understood, seq-length also initializes the max position embeddings. So for MLM, a seq-length of 512 would result in a final sequence length of 626.

I didn't know how to handle this discrepancy so I opted to make a new input-length to handle 512 seq length and seq-length to handle the final length of 626.

thomasw21 · 2022-06-18T21:27:02Z

megatron/data/non_causal_mlm_dataset.py

+# coding=utf-8
+# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.


thomasw21 · 2022-06-18T21:28:45Z

megatron/data/non_causal_mlm_dataset.py

+from megatron.data.indexed_dataset import make_dataset as make_indexed_dataset
+
+
+def build_train_valid_test_datasets(data_prefix, data_impl, splits_string,


What's the difference with

Megatron-DeepSpeed/megatron/data/dataset_utils.py

Line 421 in e23393f

def build_train_valid_test_datasets(data_prefix, data_impl, splits_string,

?

If the goal is to replace GPTDataset with NonCausalMLMDataset it seems like a lot of duplication.

Also, Megatron-LM already has some sort of T5 dataset, why not use that one? https://github.com/NVIDIA/Megatron-LM/blob/d898a8991d1a08d29074f87819d1bf41517e35f5/megatron/data/t5_dataset.py#L29

I adjusted build_train_valid_test_datasets input arguments.

And the T5 dataset didn't count for prefix length. Also, my understanding is that get_samples_mapping in T5Dataset doesn't always return the required sequence length without paddings. So I adapted from the original T5 codebase and also made a get_samples_mapping that would concat every obtained sample to the desired input length.

thomasw21 · 2022-06-18T21:39:21Z

prepare_tokenizer.py

+
+tokenizer = AutoTokenizer.from_pretrained('bigscience/tokenizer')
+
+tokenizer.add_special_tokens({


Maybe this could be additional config to _AutoTokenizer?

I didn't want to changed that since each user might have different preference for their tokenizer.

Hum but you are making some assumptions when building the dataset self.sentinel_tokens = tokenizer.additional_special_tokens_ids ...

Oh yeah, which didn't get used. Maybe it's better to remove that line.

They did get used (otherwise your MLM implementation sounds wrong, if you don't provide a list of sentinel tokens)

thomasw21 · 2022-06-18T22:13:21Z

4B8-en-ND-MLM.sh

+INPUT_LEN=1675
+TARGET_LEN=373


I'm thinking we need a way to compute these values from a given SEQ_LEN. Typically given a noise_density, mean_noise_span_length, and sequence_length we should be able to compute an input and target no? The reason why, is because what we really care about is that SEQ_LEN is 2048 (for performance), the rest we don';t really care as they are implementation details.

Yes, I agree. But I'm not sure where to put this function in.

thomasw21 · 2022-06-18T22:29:37Z

megatron/data/non_causal_mlm_dataset.py

+    tokens_length = inputs_length
+
+    while _tokens_length_to_inputs_length_targets_length(tokens_length + 1)[0] <= inputs_length:
+        tokens_length += 1
+
+    inputs_length, targets_length = _tokens_length_to_inputs_length_targets_length(tokens_length)
+
+    # minor hack to get the targets length to be equal to inputs length
+    # which is more likely to have been set to a nice round number.
+    if noise_density == 0.5 and targets_length > inputs_length:
+        tokens_length -= 1
+        targets_length -= 1
+    return tokens_length, targets_length


Suggested change

tokens_length = inputs_length

while _tokens_length_to_inputs_length_targets_length(tokens_length + 1)[0] <= inputs_length:

tokens_length += 1

inputs_length, targets_length = _tokens_length_to_inputs_length_targets_length(tokens_length)

# minor hack to get the targets length to be equal to inputs length

# which is more likely to have been set to a nice round number.

if noise_density == 0.5 and targets_length > inputs_length:

tokens_length -= 1

targets_length -= 1

return tokens_length, targets_length

tokens_length = inputs_length

while sum(_tokens_length_to_inputs_length_targets_length(tokens_length)) > inputs_length:

tokens_length -= 1

inputs_length, targets_length = tokens_length_to_inputs_length_targets_length(tokens_length)

# minor hack to get the targets length to be equal to inputs length

# which is more likely to have been set to a nice round number.

if noise_density == 0.5 and targets_length > inputs_length:

tokens_length -= 1

targets_length -= 1

# tokens_length is the number of raw tokens we need to get

# inputs_length will be the input

# targets_length will be the target

return tokens_length, inputs_length, targets_length

So typically:

>>> compute_input_and_target_lengths(2048, 0.15, 3) (1860, 1675, 373)

And so you really need to only pass only the sequence length in argument.py

Thanks, actually this solves the above arguments discussion. Thanks for the input.

thomasw21 · 2022-06-18T22:33:29Z

megatron/tokenizer/tokenizer.py

    @property
    def eod(self):
        return self.tokenizer.eos_token_id


so we've overloaded <eos> as it signals:

end of document

seperation between input and target in case of MLM.

Since we already use eos as eod, can we create a new special token then? Maybe use sep instead? cc @SaulLu as she's our guru.

thomasw21 · 2022-06-18T22:34:09Z

megatron/data/non_causal_mlm_dataset.py

+    labels_sentinel = create_sentinel_ids(labels_mask.astype(np.int8), vocab_len=len(vocab_id_list))
+
+    tokens = np.asarray([tokens])
+    input_tokens_ids = filter_input_ids(tokens, input_ids_sentinel, eos_id)[0]


Typically this shouldn't use eos_id as we use it already to seperate documents.

thomasw21 · 2022-06-18T22:41:23Z

megatron/data/non_causal_mlm_dataset.py

+                    current_len += _len
+
+        print_rank_0(' > done building sapmles index maping')
+        np.save(indexmap_filename, samples_mapping, allow_pickle=True)


Unless I'm mistaken, you're missing a shuffling step here. I think it's quite important to shuffle it, otherwise you have a single document within a single batch and I believe @leandro and @loubnabnl have seen poor performances when you don't actually chuffle..

I was under the assumption that any pretraining data would already be shuffled or atleast not have any continuity between samples? Wouldn't shuffling during training also happen within the dataloader and not during data preparation in the dataset object?

I'm not sure how to shuffle in this part during initialization of the dataset.

It's shuffled here for gpt dataset:

Megatron-DeepSpeed/megatron/data/gpt_dataset.py

Line 376 in e23393f

shuffle_idx = _build_shuffle_idx(num_samples_,

So there's a few things:

document level shuffling, which can be assumed I guess

row level shuffling, ie once you create your segments of 2048 in length. Since you process documents sequentially, consecutive rows are often from the same document and that's usually an issue because you build a batch from consecutive rows. So in order to change that, gpt dataset had that shuffling index that allows you shuffle everything up together (should be the link I shared) before saving.

DataLoader should not be loading things in random order:

Megatron-DeepSpeed/megatron/data/data_samplers.py

Lines 93 to 106 in e23393f

def __iter__(self):

batch = []

# Last batch will be dropped if drop_last is not set False

for idx in range(self.consumed_samples, self.total_samples):

batch.append(idx)

if len(batch) == self.micro_batch_times_data_parallel_size:

start_idx, end_idx = self.get_start_end_idx()

yield batch[start_idx:end_idx]

batch = []

# Check the last partial batch and see drop_last is set

if len(batch) > 0 and not self.drop_last:

start_idx, end_idx = self.get_start_end_idx()

yield batch[start_idx:end_idx]

Okay, I can add a shuffling process then.

@thomasw21 I guess you mentioned by here by accident.

Oops @leandro my bad, I meant @lvwerra

Yes, we saw that very long documents in the pretraining corpus could dominate a single batch leading to significantly slower convergence.

thomasw21 · 2022-06-18T22:44:45Z

megatron/data/non_causal_mlm_dataset.py

+        for doc_idx, sample_len in zip(indexed_dataset.doc_idx, indexed_dataset.sizes):
+            _idx = 0
+
+            if current_len + sample_len > max_len:
+                end_idx = max_len - current_len
+                sample_indices.append([doc_idx, 0, end_idx])
+                samples_mapping.append(sample_indices)
+                sample_indices = []
+                current_len = 0
+                sample_len -= end_idx
+                _idx = end_idx
+
+            break_len = current_len + sample_len
+
+            indices = breakdown(sample_len, max_len=max_len)
+            for _start_idx, _end_idx in indices:
+                _len = _end_idx - _start_idx
+                if _len == max_len:
+                    samples_mapping.append([[doc_idx, _start_idx+_idx, _end_idx+_idx]])
+                else:
+                    sample_indices.append([doc_idx, _start_idx+_idx, _end_idx+_idx])
+                    current_len += _len


Are you doing the same thing as helpers.build_sample_idx?

I'm concatenating sequences up the desired sequence length like that in run_t5_mlm_flax.py

https://github.com/huggingface/transformers/blob/6589e510fa4e6c442059de2fab84752535de9b23/examples/flax/language-modeling/run_t5_mlm_flax.py#L678-L691

I noticed that get_samples_mapping from dataset_utils.py doesn't always return a sequence with the desired length which then requires padding but that would mess the t5-style MLM.

Hum this feels very close to what I expect helpers.build_sample_idx does. There's a python version of that same code:

Megatron-DeepSpeed/megatron/data/gpt_dataset.py

Line 448 in e23393f

def _build_sample_idx(sizes, doc_idx, seq_length,

thomasw21 · 2022-06-18T23:01:01Z

megatron/data/non_causal_mlm_dataset.py

+    input_ids = np.concatenate(
+        [input_ids, np.full((batch_size, 1), eos_id, dtype=np.int32)], axis=-1
+    )


Wait this feels like input_ids would be 1676 instead of 1675. I don't see where you make sure that it's 1675 so that the final sequence length is 2048

I can re-run to make sure.

thomasw21 · 2022-06-22T09:22:06Z

megatron/data/non_causal_mlm_dataset.py

+    indexmap_filename = data_prefix
+    indexmap_filename += '_{}_indexmap'.format(name)
+    indexmap_filename += '.npy'


Can we change the naming, something that mentions it's mlm dataset? Ideally we could have a index for gpt dataset and one for mlm dataset based on the same dataset.

Also stupid question, why do you have only one file, GPT-dataset seems to generate 3 files: samples / index / shuffle.

thomasw21 · 2022-06-22T09:58:51Z

megatron/data/non_causal_mlm_dataset.py

+        for doc_idx, start_index, end_index in indices:
+            sample.append(self.indexed_dataset.get(doc_idx)[start_index:end_index])
+
+        return build_training_sample(


I think we should cache that result. ie create a number_of_samples x 2048 array with already preprocessed MLM. I can clearly see that step being a bit slow, and so loading directly from an array should make things much faster? Actually scratch that it requires some thinking as you might not be able to load the whole dataset in memory, so you need to read partially from disk ...

Muennighoff · 2022-06-24T19:50:23Z

This PR has been fully superseded by mlm & mtf, correct?

lintangsutawika · 2022-06-25T05:51:25Z

Yes, we can close this if needed.

thomasw21 · 2022-06-26T16:00:40Z

Let's keep it. There's still the code from train scripts. And we can close it when everything is merged.

lintangsutawika · 2022-07-22T03:28:22Z

Closing as the MTF and MLM has been merged in separate PRs

Lintang Sutawika and others added 30 commits May 7, 2022 18:08

added train script but with prefix manually declared

d2c35fc

made new dataset

f977b85

minor adjustments

fcfbf17

added capabilities for padding and prefix lm index

870dfd8

added finetune script

791bbd0

removed script

0f44b92

added adjustments and new dataset

2ff0815

try mlm dataset

f0a79f6

minor changes

eb416c7

minor addition of import packages

c0bc21b

minor error fix

82e824c

minor error fix

7bb17ec

samples follow how gpt dataset is loaded

9929766

added masked_lm_prob

861c41f

fixed tokenizer abstractions for HF tokenizer

fe95115

added mask id

8ea5943

added mask id

aa0d146

added mask id

215e8cc

added mask id

b6eef43

added fix

bfc73a5

added bos and eos token id

1890f87

no need for sentinal token

01392a9

add aux functions

923decb

add aux functions

4611d67

add aux functions

4356de3

add pad_id

f31c686

changed lm predictions to t5

a3951e8

changed lm predictions to t5

97b9a92

changed lm predictions to t5

fe73a73

changed lm predictions to t5

6a9cb75

Lintang Sutawika added 19 commits June 3, 2022 00:26

minor fix

920343f

minor fix

a68873d

minor fix

5d43986

set correct seq len

79e8c1a

refined sampling method

786d252

refined sampling method

9110520

refined sampling method

7db34b9

refined sampling method

d946515

refined sampling method

bb4e656

refined sampling method

2e7161d

first commit, adding non causal mlm dataset

00473e4

fixed mlm dataset

5992776

fixed mlm dataset

83f5dee

fixed mlm dataset

3235c2d

fixed mlm dataset

5449978

fixed mlm dataset

95c9851

Merge branch 'bigscience-workshop:main' into mt0

9ff6172

minor changes

451318f

Merge branch 'mt0' of https://github.com/lintangsutawika/Megatron-Dee…

edfaa19

…pSpeed into mt0

thomasw21 reviewed Jun 18, 2022

View reviewed changes

thomasw21 reviewed Jun 22, 2022

View reviewed changes

lintangsutawika closed this Jul 22, 2022

lintangsutawika deleted the mt0 branch July 22, 2022 03:28

adammoody pushed a commit to adammoody/Megatron-DeepSpeed that referenced this pull request Dec 18, 2023

fix circular dependencies when using Lora (bigscience-workshop#284)

27d7865

		from megatron.data.indexed_dataset import make_dataset as make_indexed_dataset


		def build_train_valid_test_datasets(data_prefix, data_impl, splits_string,


		tokenizer = AutoTokenizer.from_pretrained('bigscience/tokenizer')

		tokenizer.add_special_tokens({

	def __iter__(self):
	batch = []
	# Last batch will be dropped if drop_last is not set False
	for idx in range(self.consumed_samples, self.total_samples):
	batch.append(idx)
	if len(batch) == self.micro_batch_times_data_parallel_size:
	start_idx, end_idx = self.get_start_end_idx()
	yield batch[start_idx:end_idx]
	batch = []

	# Check the last partial batch and see drop_last is set
	if len(batch) > 0 and not self.drop_last:
	start_idx, end_idx = self.get_start_end_idx()
	yield batch[start_idx:end_idx]

		INPUT_LEN=1675
		TARGET_LEN=373

MLM adaptation and Multitask Finetuning #284

MLM adaptation and Multitask Finetuning #284

Uh oh!

Conversation

lintangsutawika commented Jun 12, 2022

Uh oh!

thomasw21 left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

thomasw21 Jun 18, 2022 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

thomasw21 Jun 18, 2022 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

thomasw21 Jun 22, 2022 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

thomasw21 Jun 18, 2022 •

edited

Loading

thomasw21 Jun 18, 2022 •

edited

Loading

thomasw21 Jun 22, 2022 •

edited

Loading

thomasw21 commented Jun 26, 2022 •

edited

Loading