SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2503.11367 · ICML · 2025

Efficient Distributed MLLM Training with Cornstarch

Ang Chen, Mosharaf Chowdhury, Insu Jang, Runyu Lu, Nikhil Bansal

arXiv · PDF · Open in the Atlas

Code that ran

5 of 6 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (0 of them constructed an object rather than computing a result; 0 honoured the contract we drafted, 0 violated it and 4 have no contract verdict; 1 failure of our instrument; 1 no recorded run), measured 2026-11-22; repository links from the Papers with Code archive attributed under CC BY-SA, no affiliation; bodies inline only under a permissive licence as GitHub detects it. Not a reproduction of the paper's results.

"Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results. This record, dated, as JSON: /api/ran/2503.11367.json · badge: /api/ran/2503.11367.svg

Each repository's role is a label: a classifier's output recorded on the link, or, where no role is recorded, what the Papers with Code archive marked. The line under it names its basis. Neither says who wrote the code.

RepositoryRoleRan
cornstarch-org/Cornstarch canonical (classifier label)
classifier basis not recorded on the link
5 ran · 1 no recorded run
FunctionStatusWhere it lives
contrastive_loss Ran
ran; our contract draft was wrong, not the code
cornstarch-org/Cornstarch/cornstarch/models/evaclip/modeling_evaclip.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from microsoft/xpretrain, is pointer only (licence: NOASSERTION) · get_code("4675779aec92d6c0")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from microsoft/xpretrain, and GitHub could not classify the licence of that repository (NOASSERTION). The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

create_bitfield_attention_mask Ran
ran on a synthesized input
cornstarch-org/Cornstarch/cornstarch/kernel/interface.py
code served (permissive licence) · get_code("1d838b8bec0cdbad")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned Tensor, ndim 2, dtype family int, shape (2, 4) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

torch.manual_seed(0)

# Create a small input_ids tensor (batch_size=2, seq_length=4)
input_ids = torch.tensor([
    [101, 102, 103, 104],
    [101, 105, 103, 106]
], dtype=torch.int64)

# Create token_ids dict mapping modal keys to token IDs
token_ids = {
    "image": 103,
    "audio": 105,
    "video": 106
}

ARGS = {
    "input_ids": input_ids,
    "token_ids": token_ids
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 2, dtype family int, shape (2, 4) — shape (2, 4) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
default_num_feature_calculation_func_audio_static Ran
ran on a synthesized input
cornstarch-org/Cornstarch/cornstarch/models/multimodal_language_model/processing_multimodal_language_model.py
code served (permissive licence) · get_code("d5bb7e9a9f41e1be")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned list, ndim 1, dtype family int; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn
from transformers.configuration_utils import PretrainedConfig

torch.manual_seed(0)

# Create a minimal config class that has max_source_positions attribute
class MinimalConfig(PretrainedConfig):
    model_type = "test"
    
    def __init__(self, **kwargs):
        super().__init__(**kwargs)
        self.max_source_positions = 512

config = MinimalConfig()

# Create inputs dict (not directly used by the function, but required parameter)
inputs = {}

# Create outputs dict with input_features tensor
# Shape: [batch_size=2, feature_dim=4]
outputs = {
    'input_features': torch.randn(2, 4)
}

ARGS = {
    'inputs': inputs,
    'outputs': outputs,
    'config': config
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch, transformers — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 1, dtype family int — a list output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
default_num_feature_calculation_func_vision_clip Ran
ran on a synthesized input
cornstarch-org/Cornstarch/cornstarch/models/multimodal_language_model/processing_multimodal_language_model.py
code served (permissive licence) · get_code("0743dc96855ada0a")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned list, ndim 1, dtype family int; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
from transformers.configuration_utils import PretrainedConfig

torch.manual_seed(0)

# Create a minimal config class that has the required attributes
class MinimalVisionConfig(PretrainedConfig):
    def __init__(self):
        super().__init__()
        self.image_size = 224
        self.patch_size = 16

config = MinimalVisionConfig()

# Create inputs dict - the function doesn't actually use it
inputs = {}

# Create outputs dict with pixel_values tensor
# batch_size=2, channels=3, height=224, width=224
outputs = {
    'pixel_values': torch.randn(2, 3, 224, 224)
}

ARGS = {
    'inputs': inputs,
    'outputs': outputs,
    'config': config
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch, transformers — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 1, dtype family int — a list output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
default_num_feature_calculation_func_vision_static Ran
ran on a synthesized input
cornstarch-org/Cornstarch/cornstarch/models/multimodal_language_model/processing_multimodal_language_model.py
code served (permissive licence) · get_code("de6f68ef887497ad")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned list, ndim 1, dtype family int; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn
from transformers.configuration_utils import PretrainedConfig

torch.manual_seed(0)

# Create a minimal config class with required attributes
class MinimalVisionConfig(PretrainedConfig):
    model_type = "vision_model"
    
    def __init__(self, image_size=16, patch_size=4, **kwargs):
        super().__init__(**kwargs)
        self.image_size = image_size
        self.patch_size = patch_size

# Create config instance
config = MinimalVisionConfig(image_size=16, patch_size=4)

# Create inputs dict (function doesn't use it but requires it)
inputs = {}

# Create outputs dict with pixel_values tensor
# Batch size 2, with arbitrary shape for pixel_values
pixel_values = torch.randn(2, 3, 16, 16)
outputs = {
    'pixel_values': pixel_values
}

ARGS = {
    'inputs': inputs,
    'outputs': outputs,
    'config': config
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch, transformers — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 1, dtype family int — a list output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
clip_loss Not yet run
unverified (harvested, no recorded run)
cornstarch-org/Cornstarch/cornstarch/models/evaclip/modeling_evaclip.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from PeterGriffinJin/InstructG2I, is pointer only (licence: NONE) · get_code("f2ca34eca05d40cd")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from PeterGriffinJin/InstructG2I, and GitHub detected no licence in that repository when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. No invocation recorded.

Also present in other papers' code

Of 6 samples lifted from this paper's linked repositories, 6 carry a function identity and 0 do not. Of the 6 distinct functions with an identity, 2 are also present in other papers whose code we hold (2 with at least one paper dated earlier, 0 the same month, 2 later) and 4 are only here among the 52,977 papers whose code Syntology holds (as of 2026-09-26).

Only here means: among the 52,977 papers whose code Syntology holds (51,887 with lifted samples, 1,090 through their assemblies only), as of 2026-09-26, no other paper's lifted samples or current assembly bricks contain an identical function body after normalisation.

FunctionAlso present inEarliest-datedOthers named
contrastive_loss 27 other papers
22 dated earlier, 5 later · 3 only as an assembly brick
2206.00621 (2022-06)
Cross-View Language Modeling: Towards Unified Cross-Lingual Cross-Modal Pre-training
2209.06430 (2022-09)
CLIP-ViP: Adapting Pre-trained Image-Text Model to Video-Language Alignment
2210.00312 (2022-10)
Multimodal Analogical Reasoning over Knowledge Graphs
and 24 more not named here
clip_loss 16 other papers
12 dated earlier, 4 later · 1 only as an assembly brick
2305.13050 (2023-05)
AudioToken: Adaptation of Text-Conditioned Diffusion Models for Audio-to-Image Generation
2306.02583 (2023-06)
Stable Diffusion is Unstable
2309.16429 (2023-09)
Diverse and Aligned Audio-to-Video Generation via Text-to-Video Model Adaptation
and 13 more not named here

Presence, not direction: an identical function body in two papers' repositories does not say who wrote it, which came from which, or that either paper's method rests on it. Matching is exact after normalisation, so a renamed or edited function is not matched. Dates are when papers appeared, not when code was written. Only .py files the harvester lifted are covered; notebooks are not read. A new-style arXiv id YYMM.NNNN dates a paper to the month 20YY-MM; otherwise Paper.year dates it to the year; otherwise it is undated. Dates are when the paper appeared, not when the code was written. Read 2026-11-22.

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

Multimodal large language models (MLLMs) extend the capabilities of large language models (LLMs) by combining heterogeneous model architectures to handle diverse modalities like images and audio. However, this inherent heterogeneity in MLLM model structure and data types makes makeshift extensions to existing LLM training frameworks unsuitable for efficient MLLM training, especially in distributed training. In this paper, we present Cornstarch, an efficient distributed MLLM training framework that contemplates MLLM's unique characteristics in both model and data parallelization. Cornstarch introduces frozen-aware pipeline parallelism and workload-balanced context parallelism to improve MLLM training throughput. Our extensive evaluation shows that Cornstarch outperforms state-ofthe-art solutions by 2.26× on average in terms of MLLM training throughput. Cornstarch is an open-source project available on Github 1 .

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2503.11367")
get_code_for_paper("2503.11367")
have("2503.11367")

The run record, dated, one paper per request, free:

curl https://syntology.ai/api/ran/2503.11367.json

A badge for a README (the split and the date, never a ratio):

[![Syntology run record](https://syntology.ai/api/ran/2503.11367.svg)](https://syntology.ai/paper/2503.11367)

Connect an agent — have() is free.