[MLPerf] Add DLRM-DCNv2 #144

abheesht17 · 2025-08-18T08:48:56Z

TODOs:

Need to make it work for multi-host: Multi-host training does not work for distributed embeddings #143.
Currently uses dummy data, need to shift it to using the actual dataset. Already added code for loading actual data, but the path we are reading already has "batched" data. Can go ahead with this, or can load "unbatched" data and batch that up.
Eval dataset
Check if random seed is being correctly passed. We need to make sure that the initialiser returns different values for different layers (since we are passing SeedGenerator, I think we are good, but good to verify manually).
~~Instead of separate files for configs, maybe have separate functions for configs.~~
Is this a worry?

 08:44:12.046114: W jax_tpu_embedding/sparsecore/lib/core/input_preprocessing_util.cc:251] No Coo Buffer Size provided for table cat_14_table_cat_15_table_cat_23_table_cat_24_table_cat_25_table_cat_33_table_cat_34_table_cat_35_table_cat_36_table, the default value (6144) may be too large and can cause OOM. Utilize the stats returned from the sparse dense matmul preprocessing API.

Set up metric logging and checkpointing.

gemini-code-assist

Summary of Changes

Hello @abheesht17, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request introduces the DLRM-DCNv2 model as an MLPerf example, designed for efficient training on TPUs. It provides the model architecture, configuration files for dataset, model, and training hyperparameters, a dummy data loader, and a comprehensive shell script to set up the TPU environment and execute the training process. The current implementation uses dummy data and has known areas for future improvement, such as multi-host support and actual dataset integration.

Highlights

New Model Implementation: Adds the DLRM-DCNv2 model, a deep learning architecture combining deep neural networks with a cross-network for learning explicit feature interactions, commonly used in recommendation systems.
TPU Integration: Leverages Keras's distributed training capabilities with JAX backend and keras_rs.layers.DistributedEmbedding for efficient large-scale embedding table handling on TPUs, including SparseCore.
Modular Configuration: Introduces a structured configuration system using keras.utils.Config for defining dataset features, model parameters (embedding dimensions, MLP layers, DCN layers), and training hyperparameters.
Automated Setup Script: Provides a run.sh script to automate the entire setup process on Google Cloud TPU VMs, from VM creation and environment setup to dependency installation and model execution.
Dummy Data Support: Includes a dataloader.py to generate dummy data, enabling initial testing and development of the DLRM-DCNv2 model before integration with real datasets.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point in your pull request via creating an issue comment (i.e. comment on the pull request page) using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands.

Feature	Command	Description
Code Review	`/gemini review`	Performs a code review for the current pull request in its current state.
Pull Request Summary	`/gemini summary`	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in issue comments and review comments.
Help	`/gemini help`	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here.

You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.

Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

gemini-code-assist

Code Review

This pull request introduces a DLRM-DCNv2 model implementation for MLPerf benchmarks, including configurations, a dummy dataloader, the model definition, and a run script. The overall structure is good, but there are several critical and high-severity issues that need to be addressed. Specifically, there are bugs related to weight initialization due to seed reuse in loops, which will lead to layers having identical weights. There's also a potential runtime error in the model's forward pass due to improper handling of an empty list during tensor concatenation. Additionally, the main training script has a hardcoded number of epochs, ignoring the value from the configuration. I've also included some medium-severity suggestions to improve code maintainability and script robustness. Please review the detailed comments.

examples/ml_perf/model.py

examples/ml_perf/main.py

examples/ml_perf/model.py

examples/ml_perf/main.py

examples/ml_perf/run.sh

hertschuh · 2025-08-19T00:25:18Z

examples/ml_perf/main.py

+            vocabulary_size=vocabulary_size,
+            embedding_dim=embedding_dim,
+            # TODO(abheesht): Verify.
+            initializer=keras.initializers.VarianceScaling(


The default is keras.initializers.VarianceScaling(mode="fan_out"), which has a scale of 1 and a uniform distribution.

Why did you need to change it?

Oh, I did this because https://github.com/AI-Hypercomputer/RecML/blob/1821350b346b66479baaa0ab624aa67929305dea/examples/dlrm/dlrm_main.py#L232-L236

hertschuh · 2025-08-19T00:27:52Z

examples/ml_perf/main.py

+from .dataloader import DataLoader
+from .model import DLRMDCNV2
+
+SEED = 1337


It looks like you carefully passed the SEED everywhere, but it's hard to say if you missed one.

I would also do keras.config.set_random_seed(SEED) just to make sure things are consistent across hosts.

I would also do keras.config.set_random_seed(SEED) just to make sure things are consistent across hosts.

Won't this return the same output given the same input shape (for initialisers)? Does it take care to use a SeedGenerator() so that the seed is incremented for every call?

hertschuh · 2025-08-19T00:30:03Z

examples/ml_perf/main.py

+            ),
+            combiner="sum",
+            placement="sparsecore",
+            # TODO: These two args are not getting passed down to


You could try these XLA flags:

xla_sparse_core_max_ids_per_partition_per_sample

xla_sparse_core_max_unique_ids_per_partition_per_sample

hertschuh · 2025-08-19T00:31:35Z

examples/ml_perf/main.py

+            table=table_config,
+            # TODO: Verify whether it should be `(bsz, 1)` or
+            # `(bsz, multi_hot_size)`.
+            input_shape=(per_host_batch_size, multi_hot_size),


multi_hot as an input to embedding tables is not supported, it should be turned into a ragged or padded sequence of indices before being fed to the DistributedEmbedding.

Hmmm, I don't see happening here though: https://github.com/AI-Hypercomputer/RecML/blob/1821350b346b66479baaa0ab624aa67929305dea/examples/dlrm/dlrm_main.py#L245-L249. They use (bsz, 1). Not sure why, will check with them

multi_hot as an input to embedding tables is not supported, it should be turned into a ragged or padded sequence of indices before being fed to the DistributedEmbedding.

Oh, I think our understanding of multi-hot is different. I think what you mean by multi-hot is [[1, 0, 1, 0, 0], [0, 0, 1, 1, 1]]. I am passing indices here though. I should rename it to something else, it's confusing to keep multi_hot

np.random.randint( low=0, high=vocabulary_size, size=(self.batch_size, multi_hot_size), dtype=np.int64, )

abheesht17 added 12 commits August 18, 2025 12:03

Add DLRM-DCNv2 example for MLPerf

5ca57a7

Fix table_stacking arg

090098d

Rename dir

1dd9422

Add blank __init__ file to configs

25233c9

Fix imports

3d519d2

Fix imports

eef6568

Fix imports

fc77ad4

Fix imports

0a31e0f

Fix num_processes

53e9c23

Add bash script

f22b5ff

Add bash script

41f2977

Modify bash script to take in config name

09ca14c

abheesht17 requested a review from hertschuh August 18, 2025 08:48

gemini-code-assist bot reviewed Aug 18, 2025

View reviewed changes

abheesht17 added 9 commits August 18, 2025 19:54

Add way to load real dataset

2a1c759

Add way to load real dataset (1)

b50a7f5

Add dataset path

fe8dc41

Dataloader fixes (1)

ca297d4

Dataloader fixes (2)

35c3d61

Dataloader fixes (3)

9cc9b88

Feature naming edit

a0431ba

Feature naming edit

2b9538f

Feature naming edit

187ccd5

hertschuh reviewed Aug 19, 2025

View reviewed changes

abheesht17 added 5 commits August 19, 2025 06:04

Feature naming edit

a98d431

Actual dataset loading fixes (1)

d15957d

Fix feature spec dtypes

8870c8d

Fix feature spec dtypes

e971f19

Allow different batch sizes from file batch size

02c1881

abheesht17 added 11 commits August 20, 2025 11:18

Allow different batch sizes from file batch size (fixes)

b5db304

Fix feature naming

e98bdf9

Fix batching

73ca477

Fix batching

c2ad8a9

Fix dense features

a080143

Fix dense features concat

28b7189

Rename multi_hot_size to feature_list_length

a47817d

Rename sparse to lookup

9a33f09

Debug

a66e1c6

Try out XLA flags

a56532a

Format

0a9d00b

abheesht17 requested a review from hertschuh August 20, 2025 13:51

Format

42c4022

abheesht17 requested review from cantonios and silkyarora August 20, 2025 13:56

[MLPerf] Add DLRM-DCNv2 #144

Are you sure you want to change the base?

[MLPerf] Add DLRM-DCNv2 #144

Uh oh!

Conversation

abheesht17 commented Aug 18, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

gemini-code-assist bot left a comment

Choose a reason for hiding this comment

Summary of Changes

Highlights

Footnotes

Uh oh!

gemini-code-assist bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

hertschuh Aug 19, 2025

Choose a reason for hiding this comment

Uh oh!

abheesht17 Aug 19, 2025

Choose a reason for hiding this comment

Uh oh!

hertschuh Aug 19, 2025

Choose a reason for hiding this comment

Uh oh!

abheesht17 Aug 20, 2025 • 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!

hertschuh Aug 19, 2025

Choose a reason for hiding this comment

Uh oh!

hertschuh Aug 19, 2025

Choose a reason for hiding this comment

Uh oh!

abheesht17 Aug 19, 2025

Choose a reason for hiding this comment

Uh oh!

abheesht17 Aug 20, 2025 • 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!

abheesht17 Aug 20, 2025

Choose a reason for hiding this comment

Uh oh!

Uh oh!

abheesht17 commented Aug 18, 2025 •

edited

Loading

abheesht17 Aug 20, 2025 •

edited

Loading

abheesht17 Aug 20, 2025 •

edited

Loading