SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2305.02869 · ICLR · 2024

Masked Structural Growth for 2x Faster Language Model Pre-training

Zheng Zhang, Jing Li, Yiqun Yao, Yequan Wang

arXiv · PDF · Open in the Atlas

Code that ran

16 of 34 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (6 of them constructed an object rather than computing a result; 0 honoured the contract we drafted, 0 violated it and 10 have no contract verdict; 6 failures of our instrument; 18 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/2305.02869.json · badge: /api/ran/2305.02869.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
cofe-ai/msg canonical (classifier label)
classifier basis: classifier_2025
16 ran · 18 no recorded run
FunctionStatusWhere it lives
BertEmbeddingsEx Ran
ran on a synthesized input
cofe-ai/msg/model_ex/grow_ops_v2.py
code served (permissive licence) · get_code("12c02fdc7a3ef516")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, class construct-then-call prompt, model not recorded) and returned Tensor, ndim 3, dtype family float; 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.models.bert.modeling_bert import BertConfig
from typing import Optional

# Minimal config for BERT
config = BertConfig(
    vocab_size=100,
    hidden_size=4,
    num_hidden_layers=2,
    num_attention_heads=2,
    intermediate_size=8,
    hidden_act="gelu",
    hidden_dropout_prob=0.1,
    attention_probs_dropout_prob=0.1,
    max_position_embeddings=512,
    type_vocab_size=2,
    initializer_range=0.02,
    layer_norm_eps=1e-12,
    pad_token_id=0,
)

INIT_ARGS = {
    "config": config,
}

ARGS = {
    "input_ids": torch.randint(0, 100, (2, 3), dtype=torch.long),
    "token_type_ids": None,
    "position_ids": None,
    "inputs_embeds": None,
    "past_key_values_length": 0,
}

if __name__ == "__main__":
    from __main__ import LayerNormEx, BertEmbeddingsEx
    
    # Construct the class
    embeddings_ex = BertEmbeddingsEx(**INIT_ARGS)
    
    # Call the forward method
    output = embeddings_ex(**ARGS)
    
    print(f"Output shape: {output.shape}")
    print(f"Output dtype: {output.dtype}")
    print("Success!")
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
Imports
__main__, torch, transformers — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 3, dtype family float — no output shape is stated: observe_signature.py observed this class with its constructor sizes canonicalised where it could, which the node does not record, so its shape is not stated as this setup block's
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: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
BertIntermediateEx Ran
ran on a synthesized input · constructed an object; no result was computed
cofe-ai/msg/model_ex/grow_ops_v2.py
code served (permissive licence) · get_code("abc2a5159b60c88e")
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 an instance of its own class, BertIntermediateEx: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. The code mentions cuda, and the harness aliases cuda to the CPU for every sample; whether this run needed that is not recorded. If your copy returns the same type 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 torch import nn

torch.manual_seed(0)

class Config:
    def __init__(self):
        self.hidden_size = 4
        self.intermediate_size = 8
        self.hidden_act = "gelu"

config = Config()

ARGS = {
    "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 BertIntermediateEx — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a BertIntermediateEx output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
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
BertLMPredictionHeadEx Ran
ran on a synthesized input · constructed an object; no result was computed
cofe-ai/msg/model_ex/grow_ops_v2.py
code served (permissive licence) · get_code("8f34abe0617b5f97")
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 an instance of its own class, BertLMPredictionHeadEx: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type 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 torch import nn

torch.manual_seed(0)

# Define a minimal config class
class Config:
    def __init__(self):
        self.hidden_size = 4
        self.hidden_act = "gelu"  # string activation
        self.layer_norm_eps = 1e-12
        self.vocab_size = 5

# Create config instance
config = Config()

# Create a small hidden_states tensor
# Shape: (batch_size=2, sequence_length=3, hidden_size=4)
hidden_states = torch.randn(2, 3, 4)

ARGS = {
    "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 BertLMPredictionHeadEx — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a BertLMPredictionHeadEx output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
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
BertOutputEx Ran
ran on a synthesized input
cofe-ai/msg/model_ex/grow_ops_v2.py
code served (permissive licence) · get_code("ba319340870887ba")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, class construct-then-call prompt, model not recorded) and returned Tensor, ndim 2, dtype family float; 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 torch import nn

CONSTANT_MAX_MASK_VAL = 1.0

class LayerNormEx(torch.nn.LayerNorm):

    def forward(self, input_vec, ln_mask, in_growth=False, debug_print=False):
        if not in_growth:
            assert ln_mask is not None
            weighted_mean = torch.mean(input_vec, dim=-1, keepdim=True)
            weighted_var = torch.mean(torch.square(input_vec - weighted_mean), dim=-1, keepdim=True)
            output = (input_vec - weighted_mean) / torch.sqrt(weighted_var + self.eps)
            output = output * self.weight + self.bias
            return output
        else:
            weighted_mean = torch.sum(input_vec * ln_mask, dim=-1, keepdim=True) / torch.tensor(torch.sum(ln_mask).item())
            weighted_var = torch.sum(torch.square(input_vec - weighted_mean) * ln_mask, dim=-1, keepdim=True) / torch.tensor(torch.sum(ln_mask).item())
            output = (input_vec - weighted_mean) / torch.sqrt(weighted_var + self.eps)
            output = output * self.weight + self.bias
            return output

class BertOutputEx(nn.Module):

    def __init__(self, config):
        super().__init__()
        self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
        self.LayerNorm = LayerNormEx(config.hidden_size, eps=config.layer_norm_eps)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)
        self.hidden_size = config.hidden_size
        self.old_hidden_size = config.hidden_size
        self.intermediate_size = config.intermediate_size
        self.in_grow_dim = False
        self.in_grow_input = False
        self.grow_mask_vec = None
        self.debug_print = False
        self.grow_mask_vec = torch.ones(self.hidden_size, dtype=self.dense.weight.dtype, device=self.dense.weight.device)
        self.grow_mask_vec[:] = CONSTANT_MAX_MASK_VAL

    def set_mask(self, target_dim, val):
        if target_dim is not None:
            new_mask = torch.ones(target_dim, dtype=self.dense.weight.dtype)
            new_mask[self.old_hidden_size:] = val
            self.grow_mask_vec = new_mask
            self.grow_mask_vec = self.grow_mask_vec.to(self.dense.weight.device)
        else:
            self.grow_mask_vec[self.old_hidden_size:] = val

    def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor) -> torch.Tensor:
        hidden_states = self.dense(hidden_states)
        hidden_states = self.dropout(hidden_states)
        hidden_states = self.LayerNorm(hidden_states + input_tensor, self.grow_mask_vec, self.in_grow_dim)
        if self.in_grow_dim:
            hidden_states = hidden_states * self.grow_mask_vec
        return hidden_states


class Config:
    def __init__(self):
        self.intermediate_size = 6
        self.hidden_size = 4
        self.layer_norm_eps = 1e-12
        self.hidden_dropout_prob = 0.1


INIT_ARGS = {
    'config': Config()
}

ARGS = {
    'hidden_states': torch.randn(2, 6),
    'input_tensor': torch.randn(2, 4)
}

if __name__ == '__main__':
    model = BertOutputEx(**INIT_ARGS)
    output = model(**ARGS)
    print("Output shape:", output.shape)
    print("Output:", output)
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
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 float — no output shape is stated: observe_signature.py observed this class with its constructor sizes canonicalised where it could, which the node does not record, so its shape is not stated as this setup block's
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: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
BertPoolerEx Ran
ran on a synthesized input · constructed an object; no result was computed
cofe-ai/msg/model_ex/grow_ops_v2.py
code served (permissive licence) · get_code("f76bc6941b23b53c")
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 an instance of its own class, BertPoolerEx: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type 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 torch import nn

torch.manual_seed(0)

class Config:
    def __init__(self, hidden_size):
        self.hidden_size = hidden_size

config = Config(hidden_size=4)

ARGS = {
    'config': config
}
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 BertPoolerEx — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a BertPoolerEx output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
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
BertPreTrainingHeadsEx Ran
ran on a synthesized input · constructed an object; no result was computed
cofe-ai/msg/model_ex/grow_ops_v2.py
code served (permissive licence) · get_code("b3a39bfd6444ea7c")
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 an instance of its own class, BertPreTrainingHeadsEx: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type 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 torch import nn

torch.manual_seed(0)

class Config:
    def __init__(self):
        self.hidden_size = 4
        self.vocab_size = 8
        self.hidden_act = "gelu"
        self.layer_norm_eps = 1e-12

config = Config()

sequence_output = torch.randn(2, 3, 4)
pooled_output = torch.randn(2, 4)

ARGS = {
    "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 BertPreTrainingHeadsEx — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a BertPreTrainingHeadsEx output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
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
BertPredictionHeadTransformEx Ran
ran on a synthesized input · constructed an object; no result was computed
cofe-ai/msg/model_ex/grow_ops_v2.py
code served (permissive licence) · get_code("dcb1a915dcc1578a")
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 an instance of its own class, BertPredictionHeadTransformEx: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type 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 torch import nn

torch.manual_seed(0)

class ConfigStandIn:
    def __init__(self):
        self.hidden_size = 4
        self.hidden_act = "gelu"
        self.layer_norm_eps = 1e-12

config = ConfigStandIn()

hidden_states = torch.randn(2, 3, 4)

ARGS = {
    "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 BertPredictionHeadTransformEx — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a BertPredictionHeadTransformEx output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
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
BertSelfAttentionEx Ran
ran on a synthesized input · constructed an object; no result was computed
cofe-ai/msg/model_ex/grow_ops_v2.py
code served (permissive licence) · get_code("dc33918b0246bfa2")
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 an instance of its own class, BertSelfAttentionEx: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. The code mentions cuda, and the harness aliases cuda to the CPU for every sample; whether this run needed that is not recorded. If your copy returns the same type 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 torch import nn
import math

torch.manual_seed(0)

class MockConfig:
    def __init__(self):
        self.num_attention_heads = 2
        self.hidden_size = 128
        self.attention_probs_dropout_prob = 0.1
        self.max_position_embeddings = 512
        self.position_embedding_type = 'absolute'
        self.is_decoder = False

config = MockConfig()

hidden_states = torch.randn(2, 4, 128)
attention_mask = torch.ones(2, 1, 1, 4)
head_mask = None
encoder_hidden_states = None
encoder_attention_mask = None
past_key_value = None
output_attentions = False

ARGS = {
    'config': config,
    'position_embedding_type': None,
}
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 BertSelfAttentionEx — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a BertSelfAttentionEx output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
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
BertSelfOutputEx Ran
ran on a synthesized input
cofe-ai/msg/model_ex/grow_ops_v2.py
code served (permissive licence) · get_code("f3837d1329cfa0a9")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, class construct-then-call prompt, model not recorded) and returned Tensor, ndim 3, dtype family float; no output shape is stated. The image it ran in was not recorded. The code mentions cuda, and the harness aliases cuda to the CPU for every sample; whether this run needed that is 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 torch import nn

CONSTANT_ATTENTION_HEAD_SIZE = 64
CONSTANT_MAX_MASK_VAL = 1.0

class LayerNormEx(torch.nn.LayerNorm):

    def forward(self, input_vec, ln_mask, in_growth=False, debug_print=False):
        if not in_growth:
            assert ln_mask is not None
            weighted_mean = torch.mean(input_vec, dim=-1, keepdim=True)
            weighted_var = torch.mean(torch.square(input_vec - weighted_mean), dim=-1, keepdim=True)
            output = (input_vec - weighted_mean) / torch.sqrt(weighted_var + self.eps)
            output = output * self.weight + self.bias
            return output
        else:
            weighted_mean = torch.sum(input_vec * ln_mask, dim=-1, keepdim=True) / torch.tensor(torch.sum(ln_mask).item())
            weighted_var = torch.sum(torch.square(input_vec - weighted_mean) * ln_mask, dim=-1, keepdim=True) / torch.tensor(torch.sum(ln_mask).item())
            output = (input_vec - weighted_mean) / torch.sqrt(weighted_var + self.eps)
            output = output * self.weight + self.bias
            return output

class BertSelfOutputEx(nn.Module):

    def __init__(self, config):
        super().__init__()
        self.dense = nn.Linear(CONSTANT_ATTENTION_HEAD_SIZE * config.num_attention_heads, config.hidden_size)
        self.LayerNorm = LayerNormEx(config.hidden_size, eps=config.layer_norm_eps)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)
        self.in_grow_dim = False
        self.in_grow_input = False
        self.grow_mask_vec = None
        self.hidden_size = config.hidden_size
        self.old_hidden_size = config.hidden_size
        self.all_head_size = config.hidden_size
        self.debug_print = False
        self.grow_mask_vec = torch.ones(self.hidden_size, dtype=self.dense.weight.dtype, device=self.dense.weight.device)
        self.grow_mask_vec[:] = CONSTANT_MAX_MASK_VAL

    def set_mask(self, target_dim, val):
        if target_dim is not None:
            new_mask = torch.ones(target_dim, dtype=self.dense.weight.dtype)
            new_mask[self.old_hidden_size:] = val
            self.grow_mask_vec = new_mask
            self.grow_mask_vec = self.grow_mask_vec.to(self.dense.weight.device)
        else:
            self.grow_mask_vec[self.old_hidden_size:] = val

    def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor) -> torch.Tensor:
        if self.debug_print and hidden_states.device == torch.device(type='cuda', index=2):
            print('attention output')
            print(hidden_states[0][0][:10])
            print('attention input')
            print(input_tensor[0][0])
        hidden_states = self.dense(hidden_states)
        hidden_states = self.dropout(hidden_states)
        hidden_states = self.LayerNorm(hidden_states + input_tensor, self.grow_mask_vec, self.in_grow_dim)
        if self.in_grow_dim:
            hidden_states = hidden_states * self.grow_mask_vec
        if self.debug_print and hidden_states.device == torch.device(type='cuda', index=2):
            print('output of self_output')
            print(hidden_states[0][0])
        return hidden_states


class Config:
    def __init__(self):
        self.num_attention_heads = 2
        self.hidden_size = 8
        self.layer_norm_eps = 1e-12
        self.hidden_dropout_prob = 0.1


INIT_ARGS = {
    'config': Config()
}

ARGS = {
    'hidden_states': torch.randn(2, 3, CONSTANT_ATTENTION_HEAD_SIZE * 2),
    'input_tensor': torch.randn(2, 3, 8)
}

if __name__ == '__main__':
    model = BertSelfOutputEx(**INIT_ARGS)
    output = model(**ARGS)
    print("Output shape:", output.shape)
    print("Output:", output)
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
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 3, dtype family float — no output shape is stated: observe_signature.py observed this class with its constructor sizes canonicalised where it could, which the node does not record, so its shape is not stated as this setup block's
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: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
LayerNormEx Ran
ran on a synthesized input
cofe-ai/msg/model_ex/grow_ops_v2.py
code served (permissive licence) · get_code("a2517173045b652b")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, class construct-then-call prompt, model not recorded) and returned Tensor, ndim 3, dtype family float; 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

INIT_ARGS = {
    'normalized_shape': 4,
    'eps': 1e-5,
    'elementwise_affine': True,
}

ARGS = {
    'input_vec': torch.randn(2, 3, 4),
    'ln_mask': torch.ones(2, 3, 4),
    'in_growth': False,
    'debug_print': False,
}

if __name__ == '__main__':
    layer_norm_ex = torch.nn.LayerNorm(**INIT_ARGS)
    
    # Monkey-patch the forward method
    original_forward = layer_norm_ex.forward
    
    def new_forward(input_vec, ln_mask, in_growth=False, debug_print=False):
        if not in_growth:
            assert ln_mask is not None
            weighted_mean = torch.mean(input_vec, dim=-1, keepdim=True)
            weighted_var = torch.mean(torch.square(input_vec - weighted_mean), dim=-1, keepdim=True)
            output = (input_vec - weighted_mean) / torch.sqrt(weighted_var + layer_norm_ex.eps)
            output = output * layer_norm_ex.weight + layer_norm_ex.bias
            return output
        else:
            weighted_mean = torch.sum(input_vec * ln_mask, dim=-1, keepdim=True) / torch.tensor(torch.sum(ln_mask).item())
            weighted_var = torch.sum(torch.square(input_vec - weighted_mean) * ln_mask, dim=-1, keepdim=True) / torch.tensor(torch.sum(ln_mask).item())
            output = (input_vec - weighted_mean) / torch.sqrt(weighted_var + layer_norm_ex.eps)
            output = output * layer_norm_ex.weight + layer_norm_ex.bias
            return output
    
    layer_norm_ex.forward = new_forward
    
    result = layer_norm_ex(**ARGS)
    print("Output shape:", result.shape)
    print("Output:", result)
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
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 3, dtype family float — no output shape is stated: observe_signature.py observed this class with its constructor sizes canonicalised where it could, which the node does not record, so its shape is not stated as this setup block's
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: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
count_parameters Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
cofe-ai/msg/model_ex/grow_ops_v2.py
code served (permissive licence) · get_code("c03c6d119d2ac973")
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 int, ndim 0, dtype family int, shape () 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
import torch.nn as nn
from accelerate.utils.other import extract_model_from_parallel

torch.manual_seed(0)

class SimpleModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear1 = nn.Linear(4, 3)
        self.linear2 = nn.Linear(3, 2)
    
    def forward(self, x):
        x = self.linear1(x)
        x = self.linear2(x)
        return x

model = SimpleModel()

ARGS = {
    "model": model
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
accelerate, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type int, ndim 0, dtype family int, shape () — shape () 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
expand_embedding Ran
ran; our contract draft was wrong, not the code
cofe-ai/msg/model_ex/grow_ops_v2.py
code served (permissive licence) · get_code("a0b27279fea1a125")
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 Embedding. The image it ran in was not recorded. If your copy returns the same type 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 torch import nn

torch.manual_seed(0)

# Create a small old embedding to expand
vocab_size = 10
old_in_num = 4
old_embedding = nn.Embedding(vocab_size, old_in_num)
old_embedding.weight.data.normal_(mean=0.0, std=0.02)

# New embedding dimension (larger than old)
new_in_num = 7

# Create a simple args object with required attributes
class Args:
    def __init__(self):
        pass

args = Args()

ARGS = {
    "old_embedding": old_embedding,
    "new_in_num": new_in_num,
    "vocab_size": vocab_size,
    "args": args
}
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 Embedding — a Embedding output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a Embedding output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
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
expand_linear Ran
ran; our contract draft was wrong, not the code
cofe-ai/msg/model_ex/grow_ops_v2.py
code served (permissive licence) · get_code("8162611121d247e7")
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 Linear. The image it ran in was not recorded. If your copy returns the same type 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 torch import nn

torch.manual_seed(0)

class LayerNormEx(torch.nn.LayerNorm):
    def forward(self, input_vec, ln_mask, in_growth=False, debug_print=False):
        if not in_growth:
            assert ln_mask is not None
            weighted_mean = torch.mean(input_vec, dim=-1, keepdim=True)
            weighted_var = torch.mean(torch.square(input_vec - weighted_mean), dim=-1, keepdim=True)
            output = (input_vec - weighted_mean) / torch.sqrt(weighted_var + self.eps)
            output = output * self.weight + self.bias
            return output
        else:
            weighted_mean = torch.sum(input_vec * ln_mask, dim=-1, keepdim=True) / torch.tensor(torch.sum(ln_mask).item())
            weighted_var = torch.sum(torch.square(input_vec - weighted_mean) * ln_mask, dim=-1, keepdim=True) / torch.tensor(torch.sum(ln_mask).item())
            output = (input_vec - weighted_mean) / torch.sqrt(weighted_var + self.eps)
            output = output * self.weight + self.bias
            return output

def init_weight_ex(module, initializer_range=0.02, reference_module=None):
    if reference_module is not None:
        if isinstance(module, nn.Linear):
            std_w, mean_w = torch.std_mean(reference_module.weight)
            std_b, mean_b = torch.std_mean(reference_module.bias)
            std_w, mean_w, std_b, mean_b = (std_w.item(), mean_w.item(), std_b.item(), mean_b.item())
            module.weight.data.normal_(mean=mean_w, std=std_w)
            if module.bias is not None:
                module.bias.data.normal_(mean=mean_b, std=std_b)
        elif isinstance(module, nn.Embedding):
            std_w, mean_w = torch.std_mean(reference_module.weight)
            std_w, mean_w = (std_w.item(), mean_w.item())
            module.weight.data.normal_(mean=mean_w, std=std_w)
            if module.padding_idx is not None:
                module.weight.data[module.padding_idx].zero_()
        elif isinstance(module, LayerNormEx):
            std_w, mean_w = torch.std_mean(reference_module.weight)
            std_b, mean_b = torch.std_mean(reference_module.bias)
            std_w, mean_w, std_b, mean_b = (std_w.item(), mean_w.item(), std_b.item(), mean_b.item())
            module.bias.data.normal_(mean=mean_b, std=std_b)
            module.weight.data.normal_(mean=mean_w, std=std_w)
    elif isinstance(module, nn.Linear):
        module.weight.data.normal_(mean=0.0, std=initializer_range)
        if module.bias is not None:
            module.bias.data.zero_()
    elif isinstance(module, nn.Embedding):
        module.weight.data.normal_(mean=0.0, std=initializer_range)
        if module.padding_idx is not None:
            module.weight.data[module.padding_idx].zero_()
    elif isinstance(module, LayerNormEx) or isinstance(module, torch.nn.LayerNorm):
        module.bias.data.zero_()
        module.weight.data.fill_(1.0)

def expand_linear(old_linear, new_out_num, new_in_num, args, extra=None):
    old_out_num, old_in_num = old_linear.weight.size()
    if old_in_num > new_in_num or old_out_num > new_out_num:
        raise ValueError('New Linear smaller than old')
    has_bias = old_linear.bias is not None
    new_linear = nn.Linear(new_in_num, new_out_num, bias=has_bias)
    init_weight_ex(new_linear, reference_module=None)
    new_linear.weight.data[:old_out_num, :old_in_num] = old_linear.weight.data[:, :]
    if has_bias:
        new_linear.bias.data[:old_out_num] = old_linear.bias.data[:]
    return new_linear

old_linear = nn.Linear(3, 2, bias=True)
old_linear.weight.data.normal_(mean=0.0, std=0.02)
old_linear.bias.data.zero_()

new_out_num = 4
new_in_num = 5

args = {
    "initializer_range": 0.02
}

extra_dict = {"key": "value"}

ARGS = {
    "old_linear": old_linear,
    "new_out_num": new_out_num,
    "new_in_num": new_in_num,
    "args": args,
    "extra": extra_dict
}
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 Linear — a Linear output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a Linear output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
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
expand_norm Ran
ran; our contract draft was wrong, not the code
cofe-ai/msg/model_ex/grow_ops_v2.py
code served (permissive licence) · get_code("519d9bcfb16dcf6c")
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 LayerNormEx. The image it ran in was not recorded. If your copy returns the same type 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 torch import nn

torch.manual_seed(0)

class LayerNormEx(torch.nn.LayerNorm):
    def forward(self, input_vec, ln_mask, in_growth=False, debug_print=False):
        if not in_growth:
            assert ln_mask is not None
            weighted_mean = torch.mean(input_vec, dim=-1, keepdim=True)
            weighted_var = torch.mean(torch.square(input_vec - weighted_mean), dim=-1, keepdim=True)
            output = (input_vec - weighted_mean) / torch.sqrt(weighted_var + self.eps)
            output = output * self.weight + self.bias
            return output
        else:
            weighted_mean = torch.sum(input_vec * ln_mask, dim=-1, keepdim=True) / torch.tensor(torch.sum(ln_mask).item())
            weighted_var = torch.sum(torch.square(input_vec - weighted_mean) * ln_mask, dim=-1, keepdim=True) / torch.tensor(torch.sum(ln_mask).item())
            output = (input_vec - weighted_mean) / torch.sqrt(weighted_var + self.eps)
            output = output * self.weight + self.bias
            return output

def init_weight_ex(module, initializer_range=0.02, reference_module=None):
    if reference_module is not None:
        if isinstance(module, nn.Linear):
            std_w, mean_w = torch.std_mean(reference_module.weight)
            std_b, mean_b = torch.std_mean(reference_module.bias)
            std_w, mean_w, std_b, mean_b = (std_w.item(), mean_w.item(), std_b.item(), mean_b.item())
            module.weight.data.normal_(mean=mean_w, std=std_w)
            if module.bias is not None:
                module.bias.data.normal_(mean=mean_b, std=std_b)
        elif isinstance(module, nn.Embedding):
            std_w, mean_w = torch.std_mean(reference_module.weight)
            std_w, mean_w = (std_w.item(), mean_w.item())
            module.weight.data.normal_(mean=mean_w, std=std_w)
            if module.padding_idx is not None:
                module.weight.data[module.padding_idx].zero_()
        elif isinstance(module, LayerNormEx):
            std_w, mean_w = torch.std_mean(reference_module.weight)
            std_b, mean_b = torch.std_mean(reference_module.bias)
            std_w, mean_w, std_b, mean_b = (std_w.item(), mean_w.item(), std_b.item(), mean_b.item())
            module.bias.data.normal_(mean=mean_b, std=std_b)
            module.weight.data.normal_(mean=mean_w, std=std_w)
    elif isinstance(module, nn.Linear):
        module.weight.data.normal_(mean=0.0, std=initializer_range)
        if module.bias is not None:
            module.bias.data.zero_()
    elif isinstance(module, nn.Embedding):
        module.weight.data.normal_(mean=0.0, std=initializer_range)
        if module.padding_idx is not None:
            module.weight.data[module.padding_idx].zero_()
    elif isinstance(module, LayerNormEx) or isinstance(module, torch.nn.LayerNorm):
        module.bias.data.zero_()
        module.weight.data.fill_(1.0)

def expand_norm(old_ln, new_out_num, args):
    old_out_num = old_ln.weight.size()[0]
    if old_out_num >= new_out_num:
        raise ValueError('New LN smaller than or equal to old')
    new_ln = LayerNormEx(new_out_num, eps=1e-12)
    init_weight_ex(new_ln, reference_module=None)
    new_ln.weight.data[:old_out_num] = old_ln.weight.data[:]
    new_ln.bias.data[:old_out_num] = old_ln.bias.data[:]
    return new_ln

old_ln = LayerNormEx(4, eps=1e-12)
init_weight_ex(old_ln, reference_module=None)

args = {}

ARGS = {
    "old_ln": old_ln,
    "new_out_num": 8,
    "args": args
}
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 LayerNormEx — a LayerNormEx output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a LayerNormEx output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
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
prepare_eval_per_device Ran
ran; our contract draft was wrong, not the code
cofe-ai/msg/run_grow_bert.py
code served (permissive licence) · get_code("ee96c6045e2257e2")
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) and returned tuple. The image it ran in was not recorded. If your copy returns the same type 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 small tensors for MLM predictions (batch_size=2, seq_len=3, vocab_size=4)
mlm_preds = torch.randn(2, 3, 4)

# Create small tensors for NSP predictions (batch_size=2, num_classes=2)
nsp_preds = torch.randn(2, 2)

# Create MLM labels (batch_size=2, seq_len=3) with some -100 values for masking
mlm_labels = torch.tensor([[1, -100, 2], [0, 1, -100]])

# Create NSP labels (batch_size=2)
nsp_labels = torch.tensor([0, 1])

ARGS = {
    "mlm_preds": mlm_preds,
    "nsp_preds": nsp_preds,
    "mlm_labels": mlm_labels,
    "nsp_labels": nsp_labels,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple — a tuple output has no array shape of its own; none is stated
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
static_mask_single Ran
ran; our contract draft was wrong, not the code
cofe-ai/msg/preprocess_bert_data.py
code served (permissive licence) · get_code("681bc51b62ed3855")
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) and returned tuple, ndim 2, 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 numpy as np

class MockTokenizer:
    def __init__(self):
        self.mask_token_id = 103
        self.vocab_size = 30522
    
    def __len__(self):
        return self.vocab_size
    
    def get_special_tokens_mask(self, token_ids, already_has_special_tokens=False):
        # Return a mask where special tokens (like [CLS], [SEP], [PAD]) are marked as 1
        # For this mock, mark positions 0 and len-1 as special tokens
        mask = [0] * len(token_ids)
        if len(token_ids) > 0:
            mask[0] = 1  # [CLS] token
            if len(token_ids) > 1:
                mask[-1] = 1  # [SEP] token
        return mask

# Create example input: a list of token IDs
example = np.array([101, 2054, 2003, 102], dtype=np.int64)  # [CLS] what is [SEP]

# Create tokenizer
tokenizer = MockTokenizer()

ARGS = {
    "example": example,
    "tokenizer": tokenizer
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple, ndim 2, dtype family int — a tuple 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
BertAttentionEx Not yet run
unverified (harvested, no recorded run)
cofe-ai/msg/model_ex/grow_ops_v2.py
code served (permissive licence) · get_code("ae85ec1873dd1703")
How it ran

no invocation recorded.

BertEncoderEx Not yet run
unverified (harvested, no recorded run)
cofe-ai/msg/model_ex/grow_ops_v2.py
code served (permissive licence) · get_code("1ccc2565fa59af1d")
How it ran

no invocation recorded.

BertLayerEx Not yet run
unverified (harvested, no recorded run)
cofe-ai/msg/model_ex/grow_ops_v2.py
code served (permissive licence) · get_code("86592733e07a8723")
How it ran

no invocation recorded.

grow_block Not yet run
unverified (harvested, no recorded run)
cofe-ai/msg/model_ex/grow_ops_v2.py
code served (permissive licence) · get_code("c969b61fcdd66216")
How it ran

no invocation recorded.

grow_dim_intermediate Not yet run
unverified (harvested, no recorded run)
cofe-ai/msg/model_ex/grow_ops_v2.py
code served (permissive licence) · get_code("2c3575cfa3435b09")
How it ran

no invocation recorded.

grow_dim_nsp Not yet run
unverified (harvested, no recorded run)
cofe-ai/msg/model_ex/grow_ops_v2.py
code served (permissive licence) · get_code("596cf53dd41ddb02")
How it ran

no invocation recorded.

grow_dim_output Not yet run
unverified (harvested, no recorded run)
cofe-ai/msg/model_ex/grow_ops_v2.py
code served (permissive licence) · get_code("fb20ae8f1b2151e9")
How it ran

no invocation recorded.

grow_dim_pooler Not yet run
unverified (harvested, no recorded run)
cofe-ai/msg/model_ex/grow_ops_v2.py
code served (permissive licence) · get_code("ac46a64ea2608233")
How it ran

no invocation recorded.

grow_dim_self_att Not yet run
unverified (harvested, no recorded run)
cofe-ai/msg/model_ex/grow_ops_v2.py
code served (permissive licence) · get_code("baaac4ac9de4e27a")
How it ran

no invocation recorded.

grow_dim_self_output Not yet run
unverified (harvested, no recorded run)
cofe-ai/msg/model_ex/grow_ops_v2.py
code served (permissive licence) · get_code("88e9fafdee923a80")
How it ran

no invocation recorded.

grow_dim_transform Not yet run
unverified (harvested, no recorded run)
cofe-ai/msg/model_ex/grow_ops_v2.py
code served (permissive licence) · get_code("06fa339e87bd096d")
How it ran

no invocation recorded.

grow_embedding Not yet run
unverified (harvested, no recorded run)
cofe-ai/msg/model_ex/grow_ops_v2.py
code served (permissive licence) · get_code("d1ada439198d07ca")
How it ran

no invocation recorded.

grow_head_num Not yet run
unverified (harvested, no recorded run)
cofe-ai/msg/model_ex/grow_ops_v2.py
code served (permissive licence) · get_code("42d5058f41beae5b")
How it ran

no invocation recorded.

grow_ops Not yet run
unverified (harvested, no recorded run)
cofe-ai/msg/model_ex/grow_ops_v2.py
code served (permissive licence) · get_code("109797f27864fd02")
How it ran

no invocation recorded.

init_weight_ex Not yet run
unverified (harvested, no recorded run)
cofe-ai/msg/model_ex/grow_ops_v2.py
code served (permissive licence) · get_code("ea16063f081619fc")
How it ran

no invocation recorded.

stack_block Not yet run
unverified (harvested, no recorded run)
cofe-ai/msg/model_ex/grow_ops_v2.py
code served (permissive licence) · get_code("18d7d645c49b39b9")
How it ran

no invocation recorded.

transfer_states Not yet run
unverified (harvested, no recorded run)
cofe-ai/msg/model_ex/grow_ops_v2.py
code served (permissive licence) · get_code("ef5ff778ebe66fa5")
How it ran

no invocation recorded.

vanilla_copy Not yet run
unverified (harvested, no recorded run)
cofe-ai/msg/model_ex/grow_ops_v2.py
code served (permissive licence) · get_code("8179d1035ab4a2dd")
How it ran

no invocation recorded.

Also present in other papers' code

Of 34 samples lifted from this paper's linked repositories, 34 carry a function identity and 0 do not. Of the 34 distinct functions with an identity, 1 are also present in other papers whose code we hold (0 with at least one paper dated earlier, 0 the same month, 1 later) and 33 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
prepare_eval_per_device 1 other paper
1 later
2407.04752 (2024-07)
SpikeLLM: Scaling up Spiking Neural Network to Large Language Models via Saliency-based Spiking

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

Accelerating large language model pre-training is a critical issue in present research. In this paper, we focus on speeding up pre-training by progressively growing from a small Transformer structure to a large one. There are two main research problems associated with progressive growth: determining the optimal growth schedule, and designing efficient growth operators. In terms of growth schedule, the impact of each single dimension on a schedule's efficiency is underexplored by existing work. Regarding the growth operators, existing methods rely on the initialization of new weights to inherit knowledge, and achieve only non-strict function preservation, limiting further improvements on training dynamics. To address these issues, we propose Masked Structural Growth (MSG), including (i) growth schedules involving all possible dimensions and (ii) strictly function-preserving growth operators that is independent of the initialization of new weights. Experiments show that MSG is significantly faster than related work: we achieve up to 2.2x speedup in pre-training different types of language models while maintaining comparable or better downstream performances. 1

For agents

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

get_harvested_code_for_paper("2305.02869")
get_code_for_paper("2305.02869")
have("2305.02869")

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

curl https://syntology.ai/api/ran/2305.02869.json

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

[![Syntology run record](https://syntology.ai/api/ran/2305.02869.svg)](https://syntology.ai/paper/2305.02869)

Connect an agent — have() is free.