SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2406.08903 · NeurIPS · 2024

Delta-CoMe: Training-Free Delta-Compression with Mixed-Precision for Large Language Models

Shuo Wang, Maosong Sun, Zhiyuan Liu, Xu Han, Hanqing Wang, Baobao Chang, Yun Chen, Yuzhuang Xu, Bowen Ping, Yukun Yan

arXiv · PDF · Open in the Atlas

Code that ran

5 of 5 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (1 of them constructed an object rather than computing a result; 0 honoured the contract we drafted, 0 violated it and 1 has no contract verdict; 4 failures of our instrument; 0 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/2406.08903.json · badge: /api/ran/2406.08903.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
thunlp/Delta-CoMe canonical (classifier label)
classifier basis: classifier_2025
2 ran · 0 no recorded run
copy not recorded community
not in the paper's declared repositories
2 ran · 0 no recorded run
thunlp/delta-come marked official in Papers with Code
Papers with Code archive
1 ran · 0 no recorded run
FunctionStatusWhere it lives
Delta Ran
ran on a synthesized input · constructed an object; no result was computed
thunlp/delta-come/llama.py
code served (permissive licence) · get_code("1ac6df1d312759df")
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, Delta: 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
import torch.nn as nn

torch.manual_seed(0)

# Construct tensor arguments
# base: weight matrix to be updated
base = torch.randn(5, 4)

# U, S, V for SVD decomposition
# U: (output_dim, rank)
# S: (rank,)
# V: (input_dim, rank)
rank = 3
U = torch.randn(5, rank)
S = torch.randn(rank)
V = torch.randn(4, rank)

ARGS = {
    'base': base,
    'U': U,
    'S': S,
    'V': V,
}
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 Delta — 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 Delta 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
batch_data Ran
ran; our contract draft was wrong, not the code
this paper's copy was not recorded; identical code first harvested from meta-math/MetaMath
pointer only · get_code("66b3b5450182da3a")
How it ran

The body of this sample is withheld: this paper's copy of the code was not recorded (its link names no repository or source path), so no licence was read for it. 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.

llama_quant_mlp_forward Ran
ran; our contract draft was wrong, not the code
thunlp/Delta-CoMe/llama.py
code served (permissive licence) · get_code("80029b25298d1d86")
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 Tensor, ndim 2, dtype family float, 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
import torch.nn as nn

torch.manual_seed(0)

# Create a mock quantized projection layer that accepts gptq parameter
class QuantizedLinear(nn.Module):
    def __init__(self, in_features, out_features):
        super().__init__()
        self.weight = nn.Parameter(torch.randn(out_features, in_features))
        self.bias = nn.Parameter(torch.randn(out_features))
    
    def forward(self, x, gptq=None, quant_type=None):
        return torch.nn.functional.linear(x, self.weight, self.bias)

# Create the MLP module with the required projections
class LlamaQuantMLP(nn.Module):
    def __init__(self):
        super().__init__()
        self.gate_proj = QuantizedLinear(4, 8)
        self.up_proj = QuantizedLinear(4, 8)
        self.down_proj = QuantizedLinear(8, 4)
        self.act_fn = nn.GELU()
    
    def forward(self, x):
        quant_type = gptqs.get('quant_type', None)
        gptq_up, gptq_gate, gptq_down = (None, None, None)
        if len(gptqs.keys()) == 3:
            for k, v in gptqs.items():
                if 'up_proj' in k:
                    gptq_up = v
                elif 'gate_proj' in k:
                    gptq_gate = v
        else:
            for k, v in gptqs.items():
                if 'down_proj' in k:
                    gptq_down = v
        return self.down_proj(self.act_fn(self.gate_proj(x, gptq_gate, quant_type)) * self.up_proj(x, gptq_up, quant_type), gptq_down, quant_type)

# Set up the global gptqs dictionary with quantization parameters
gptqs = {
    'quant_type': 'gptq',
    'gate_proj': torch.tensor([0.5, 0.6, 0.7, 0.8]),
    'up_proj': torch.tensor([0.4, 0.5, 0.6, 0.7])
}

# Create the model instance
mlp_model = LlamaQuantMLP()

# Create input tensor (batch_size=2, input_dim=4)
x = torch.randn(2, 4)

# Define ARGS dictionary
ARGS = {
    'self': mlp_model,
    'x': x
}
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 Tensor, ndim 2, dtype family float, 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
mistral_quant_mlp_forward Ran
ran; our contract draft was wrong, not the code
thunlp/Delta-CoMe/mistral.py
code served (permissive licence) · get_code("3276320f10b08020")
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 Tensor, ndim 2, dtype family float, 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
import torch.nn as nn

torch.manual_seed(0)

# Create a mock quantized linear layer that accepts gptq parameter
class QuantizedLinear(nn.Module):
    def __init__(self, in_features, out_features):
        super().__init__()
        self.weight = nn.Parameter(torch.randn(out_features, in_features))
        self.bias = nn.Parameter(torch.randn(out_features))
    
    def forward(self, x, gptq=None, quant_type=None):
        return torch.nn.functional.linear(x, self.weight, self.bias)

# Create the MLP module with the three projections
class MistralQuantMLP(nn.Module):
    def __init__(self):
        super().__init__()
        self.gate_proj = QuantizedLinear(4, 8)
        self.up_proj = QuantizedLinear(4, 8)
        self.down_proj = QuantizedLinear(8, 4)
        self.act_fn = nn.GELU()
    
    def forward(self, x):
        return mistral_quant_mlp_forward(self, x)

def mistral_quant_mlp_forward(self, x):
    quant_type = gptqs.get('quant_type', None)
    gptq_up, gptq_gate, gptq_down = (None, None, None)
    if len(gptqs.keys()) == 3:
        for k, v in gptqs.items():
            if 'up_proj' in k:
                gptq_up = v
            elif 'gate_proj' in k:
                gptq_gate = v
    else:
        for k, v in gptqs.items():
            if 'down_proj' in k:
                gptq_down = v
    return self.down_proj(self.act_fn(self.gate_proj(x, gptq_gate, quant_type)) * self.up_proj(x, gptq_up, quant_type), gptq_down, quant_type)

# Set up the global gptqs dict with quantization info
gptqs = {
    'quant_type': 'gptq',
    'gate_proj': torch.randn(8, 4),
    'up_proj': torch.randn(8, 4),
    'down_proj': torch.randn(4, 8)
}

# Create input tensor: batch_size=2, hidden_dim=4
x = torch.randn(2, 4)

# Create the MLP module
mlp = MistralQuantMLP()

ARGS = {
    'self': mlp,
    'x': x
}
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 Tensor, ndim 2, dtype family float, 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
remove_boxed Ran
ran; our contract draft was wrong, not the code
this paper's copy was not recorded; identical code first harvested from meta-math/MetaMath
pointer only · get_code("f3bbe264b05aadd3")
How it ran

The body of this sample is withheld: this paper's copy of the code was not recorded (its link names no repository or source path), so no licence was read for it. 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.

Also present in other papers' code

Of 5 samples lifted from this paper's linked repositories, 5 carry a function identity and 0 do not. Of the 5 distinct functions with an identity, 2 are also present in other papers whose code we hold (2 with at least one paper dated earlier, 2 the same month, 1 later) and 3 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
remove_boxed 20 other papers
9 dated earlier, 1 the same month, 9 later, 1 same year, order not determinable · 1 only as an assembly brick
2103.03874 (2021-03)
Measuring Mathematical Problem Solving With the MATH Dataset
2309.06275 (2023-09)
Re-Reading Improves Reasoning in Large Language Models
2309.12284 (2023-09)
MetaMath: Bootstrap Your Own Mathematical Questions for Large Language Models
and 17 more not named here
batch_data 7 other papers
5 dated earlier, 2 the same month
2309.12284 (2023-09)
MetaMath: Bootstrap Your Own Mathematical Questions for Large Language Models
2404.10346 (2024-04)
Self-Explore: Enhancing Mathematical Reasoning in Language Models with Fine-grained Rewards
2405.06680 (2024-05)
Exploring the Compositional Deficiency of Large Language Models in Mathematical Reasoning Through Trap Problems
and 4 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

Fine-tuning is a crucial process for adapting large language models (LLMs) to diverse applications. In certain scenarios, such as multi-tenant serving, deploying multiple LLMs becomes necessary to meet complex demands. Recent studies suggest decomposing a fine-tuned LLM into a base model and corresponding delta weights, which are then compressed using low-rank or low-bit approaches to reduce costs. In this work, we observe that existing low-rank and low-bit compression methods can significantly harm the model performance for taskspecific fine-tuned LLMs (e.g., WizardMath for math problems). Motivated by the long-tail distribution of singular values in the delta weights, we propose a delta quantization approach using mixed-precision. This method employs higher-bit representation for singular vectors corresponding to larger singular values. We evaluate our approach on various fine-tuned LLMs, including math LLMs, code LLMs, chat LLMs, and even VLMs. Experimental results demonstrate that our approach performs comparably to full fine-tuned LLMs, surpassing both low-rank and low-bit baselines by a considerable margin. Additionally, we show that our method is compatible with various backbone LLMs, such as Llama-2, Llama-3, and Mistral, highlighting its generalizability. 3 * Equal contribution.

For agents

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

get_harvested_code_for_paper("2406.08903")
get_code_for_paper("2406.08903")
have("2406.08903")

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

curl https://syntology.ai/api/ran/2406.08903.json

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

[![Syntology run record](https://syntology.ai/api/ran/2406.08903.svg)](https://syntology.ai/paper/2406.08903)

Connect an agent — have() is free.