SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2407.17465 · 2024

u-$μ$P: The Unit-Scaled Maximal Update Parametrization

arXiv · PDF · Open in the Atlas

Code that ran

12 of 13 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; 5 honoured the contract we drafted, 1 violated it and 0 have no contract verdict; 6 failures 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/2407.17465.json · badge: /api/ran/2407.17465.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
graphcore-research/unit-scaling canonical (classifier label)
classifier basis: census
12 ran · 1 no recorded run
FunctionStatusWhere it lives
Parameter Ran
ran; our contract draft was wrong, not the code · constructed an object; no result was computed
graphcore-research/unit-scaling/unit_scaling/parameter.py
code served (permissive licence) · get_code("8d7406030299bc46")
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, Parameter: 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)

# Create a small tensor for the parameter
data = torch.randn(4, 3)

# Set up the arguments dictionary
ARGS = {
    'data': data,
    'mup_type': 'weight',
    'mup_scaling_depth': None
}
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 Parameter — 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 Parameter output the recorded ndim 2 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
format_to_tuple Ran
ran, honoured the contract we drafted
graphcore-research/unit-scaling/unit_scaling/formats.py
code served (permissive licence) · get_code("070a400b0ece5434")
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 tuple, 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 dataclasses import dataclass

@dataclass
class FPFormat:
    """Generic representation of a floating-point number format."""
    exponent_bits: int
    mantissa_bits: int
    rounding: str = 'stochastic'
    srbits: int = 0

    def __post_init__(self) -> None:
        assert self.exponent_bits >= 2, 'FPFormat requires at least 2 exponent bits'
        assert self.srbits == 0 or self.rounding == 'stochastic', 'Nonzero srbits for non-stochastic rounding'
        if self.srbits == 0 and self.rounding == 'stochastic':
            self.srbits = 23 - self.mantissa_bits

    @property
    def bits(self) -> int:
        """The number of bits used by the format."""
        return 1 + self.exponent_bits + self.mantissa_bits

    def __str__(self) -> str:
        return f'E{self.exponent_bits}M{self.mantissa_bits}-' + dict(stochastic='SR', nearest='RN')[self.rounding]

    @property
    def max_absolute_value(self) -> float:
        """The maximum absolute value representable by the format."""
        max_exponent = 2 ** (self.exponent_bits - 1) - 1
        return float(2 ** max_exponent * (2 - 2 ** (-self.mantissa_bits)))

    @property
    def min_absolute_normal(self) -> float:
        """The minimum absolute normal value representable by the format."""
        min_exponent = 1 - 2 ** (self.exponent_bits - 1)
        return float(2 ** min_exponent)

    @property
    def min_absolute_subnormal(self) -> float:
        """The minimum absolute subnormal value representable by the format."""
        return self.min_absolute_normal * 2.0 ** (-self.mantissa_bits)


torch.manual_seed(0)

# Create a valid FPFormat instance with exponent_bits >= 2 and mantissa_bits > 0
fp_format = FPFormat(exponent_bits=5, mantissa_bits=10, rounding='stochastic')

ARGS = {
    'format': fp_format
}
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 tuple, ndim 1, 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
generate__all__ Ran
ran, honoured the contract we drafted
graphcore-research/unit-scaling/unit_scaling/_internal_utils.py
code served (permissive licence) · get_code("de8aa1c681baa450")
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 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 sys
import inspect

# Create a minimal test module to work with
import types

# Create a test module with some public and private members
test_module = types.ModuleType('test_module')

# Add some public items
def public_function():
    pass

class PublicClass:
    pass

def _private_function():
    pass

class _PrivateClass:
    pass

# Add members to the test module
test_module.public_function = public_function
test_module.PublicClass = PublicClass
test_module._private_function = _private_function
test_module._PrivateClass = _PrivateClass

# Register the module in sys.modules
sys.modules['test_module'] = test_module

# Create arguments for the function
ARGS = {
    'module_name': 'test_module',
    'include_imports': False
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — 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 float — 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
has_parameter_data Ran
ran, violated the contract we drafted
graphcore-research/unit-scaling/unit_scaling/parameter.py
code served (permissive licence) · get_code("d19cc5ff648afebd")
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 bool, ndim 0, dtype family float, 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
from torch import nn

torch.manual_seed(0)

# Create a parameter with the required ParameterData protocol attributes
parameter = nn.Parameter(torch.randn(4, 3))
parameter.mup_type = 'weight'
parameter.mup_scaling_depth = 2

ARGS = {
    'parameter': parameter
}
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 bool, ndim 0, dtype family float, 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
logarithmic_interpolation Ran
ran, honoured the contract we drafted
graphcore-research/unit-scaling/unit_scaling/core/functional.py
code served (permissive licence) · get_code("e0c95afd064e25ed")
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 float, ndim 0, dtype family float, 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 math

# Construct valid arguments for logarithmic_interpolation
ARGS = {
    "alpha": 0.5,
    "lower": 1/1000,
    "upper": 1/10
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type float, ndim 0, dtype family float, 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
mult_docstring Ran
ran; our contract draft was wrong, not the code
graphcore-research/unit-scaling/unit_scaling/docs.py
code served (permissive licence) · get_code("a1e2fc6500a07813")
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 str. 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

def mult_docstring(name: str='mult') -> str:
    return f"{name} (float, optional): a multiplier to be applied to change the shape of a nonlinear function. Typically, high multipliers (> 1) correspond to a 'sharper' (low temperature) function, while low multipliers (< 1) correspond to a 'flatter' (high temperature) function."

ARGS = {
    'name': 'mult'
}
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 str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str 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
rms Ran
ran; our contract draft was wrong, not the code
graphcore-research/unit-scaling/unit_scaling/core/functional.py
code served (permissive licence) · get_code("7f174b2d78b36837")
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 1, dtype family float, shape (2,) 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 small tensor for testing
x = torch.randn(2, 4, dtype=torch.float32)

# Define arguments for the rms function
ARGS = {
    'x': x,
    'dims': (1,),
    'keepdim': False,
    'eps': 1e-8
}
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 1, dtype family float, shape (2,) — shape (2,) 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
scale_bwd Ran
ran; our contract draft was wrong, not the code
graphcore-research/unit-scaling/unit_scaling/scale.py
code served (permissive licence) · get_code("af14bc3abf24163b")
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 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

torch.manual_seed(0)

# Create a small tensor for testing
input_tensor = torch.randn(2, 4, requires_grad=True)

# Define the scale factor
scale_factor = 2.0

# Create the ARGS dictionary with all parameters
ARGS = {
    "input": input_tensor,
    "scale": scale_factor,
}
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 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
scale_fwd Ran
ran; our contract draft was wrong, not the code
graphcore-research/unit-scaling/unit_scaling/scale.py
code served (permissive licence) · get_code("99e5d990030cdf9a")
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 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

torch.manual_seed(0)

# Create a small tensor for testing
input_tensor = torch.randn(2, 4, requires_grad=True)

# Define the scale factor
scale_factor = 2.5

ARGS = {
    "input": input_tensor,
    "scale": scale_factor,
}
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 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
to_grad_input_scale Ran
ran, honoured the contract we drafted
graphcore-research/unit-scaling/unit_scaling/constraints.py
code served (permissive licence) · get_code("311ddf75624e8833")
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 float, ndim 0, dtype family float, 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 numpy as np

torch.manual_seed(0)

ARGS = {
    "output_scale": 2.5,
    "grad_input_scale": 1.7,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type float, ndim 0, dtype family float, 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
to_left_grad_scale Ran
ran, honoured the contract we drafted
graphcore-research/unit-scaling/unit_scaling/constraints.py
code served (permissive licence) · get_code("e57d5ac2fdb91999")
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 float, ndim 0, dtype family float, 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 numpy as np

torch.manual_seed(0)

ARGS = {
    "output_scale": 1.5,
    "left_grad_scale": 2.5,
    "right_grad_scale": 3.5,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type float, ndim 0, dtype family float, 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
tuple_to_format Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
graphcore-research/unit-scaling/unit_scaling/formats.py
code served (permissive licence) · get_code("07ba3009b696a3c2")
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 FPFormat. 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 dataclasses import dataclass
from typing import Tuple

torch.manual_seed(0)

@dataclass
class FPFormat:
    """Generic representation of a floating-point number format."""
    exponent_bits: int
    mantissa_bits: int
    rounding: str = 'stochastic'
    srbits: int = 0

    def __post_init__(self) -> None:
        assert self.exponent_bits >= 2, 'FPFormat requires at least 2 exponent bits'
        assert self.srbits == 0 or self.rounding == 'stochastic', 'Nonzero srbits for non-stochastic rounding'
        if self.srbits == 0 and self.rounding == 'stochastic':
            self.srbits = 23 - self.mantissa_bits

    @property
    def bits(self) -> int:
        """The number of bits used by the format."""
        return 1 + self.exponent_bits + self.mantissa_bits

    def __str__(self) -> str:
        return f'E{self.exponent_bits}M{self.mantissa_bits}-' + dict(stochastic='SR', nearest='RN')[self.rounding]

    @property
    def max_absolute_value(self) -> float:
        """The maximum absolute value representable by the format."""
        max_exponent = 2 ** (self.exponent_bits - 1) - 1
        return float(2 ** max_exponent * (2 - 2 ** (-self.mantissa_bits)))

    @property
    def min_absolute_normal(self) -> float:
        """The minimum absolute normal value representable by the format."""
        min_exponent = 1 - 2 ** (self.exponent_bits - 1)
        return float(2 ** min_exponent)

    @property
    def min_absolute_subnormal(self) -> float:
        """The minimum absolute subnormal value representable by the format."""
        return self.min_absolute_normal * 2.0 ** (-self.mantissa_bits)

    def quantise(self, x: torch.Tensor) -> torch.Tensor:
        """Non-differentiably quantise the given tensor in this format."""
        absmax = self.max_absolute_value
        downscale = 2.0 ** (127 - 2 ** (self.exponent_bits - 1))
        mask = torch.tensor(2 ** (23 - self.mantissa_bits) - 1, device=x.device)
        if self.rounding == 'stochastic':
            srbitsbar = 23 - self.mantissa_bits - self.srbits
            offset = torch.randint(0, 2 ** self.srbits, x.shape, dtype=torch.int32, device=x.device) << srbitsbar
            if srbitsbar > 0:
                offset += 1 << (srbitsbar - 1)
        elif self.rounding == 'nearest':
            offset = mask // 2
        else:
            raise ValueError(f'Unexpected FPFormat(rounding="{self.rounding}"), expected "stochastic" or "nearest"')
        q = x.to(torch.float32)
        q = torch.clip(q, -absmax, absmax)
        q /= downscale
        q = (q.view(torch.int32) + offset & ~mask).view(torch.float32)
        q *= downscale
        return q.to(x.dtype)

    def quantise_fwd(self, x: torch.Tensor) -> torch.Tensor:
        """Quantise the given tensor in the forward pass only."""

        class QuantiseForward(torch.autograd.Function):

            @staticmethod
            def forward(ctx: torch.autograd.function.FunctionCtx, x: torch.Tensor) -> torch.Tensor:
                return self.quantise(x)

            @staticmethod
            def backward(ctx: torch.autograd.function.FunctionCtx, grad_y: torch.Tensor) -> torch.Tensor:
                return grad_y
        return QuantiseForward.apply(x)

    def quantise_bwd(self, x: torch.Tensor) -> torch.Tensor:
        """Quantise the given tensor in the backward pass only."""

        class QuantiseBackward(torch.autograd.Function):

            @staticmethod
            def forward(ctx: torch.autograd.function.FunctionCtx, x: torch.Tensor) -> torch.Tensor:
                return x

            @staticmethod
            def backward(ctx: torch.autograd.function.FunctionCtx, grad_y: torch.Tensor) -> torch.Tensor:
                return self.quantise(grad_y)
        return QuantiseBackward.apply(x)

def tuple_to_format(t: Tuple[int, int]) -> FPFormat:
    """Given a tuple of `(exponent_bits, mantissa_bits)` returns the corresponding
    :class:`FPFormat`"""
    return FPFormat(*t)

ARGS = {
    "t": (5, 10)
}
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 FPFormat — a FPFormat output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a FPFormat 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
to_output_scale Not yet run
unverified (harvested, no recorded run)
graphcore-research/unit-scaling/unit_scaling/constraints.py
code served (permissive licence) · get_code("873f0a22bcd0444f")
How it ran

no invocation recorded.

Also present in other papers' code

Of 13 samples lifted from this paper's linked repositories, 13 carry a function identity and 0 do not. Of the 13 distinct functions with an identity, 0 are also present in other papers whose code we hold (0 with at least one paper dated earlier, 0 the same month, 0 later) and 13 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.

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

The Maximal Update Parametrization ($μ$P) aims to make the optimal hyperparameters (HPs) of a model independent of its size, allowing them to be swept using a cheap proxy model rather than the full-size target model. We present a new scheme, u-$μ$P, which improves upon $μ$P by combining it with Unit Scaling, a method for designing models that makes them easy to train in low-precision. The two techniques have a natural affinity: $μ$P ensures that the scale of activations is independent of model size, and Unit Scaling ensures that activations, weights and gradients begin training with a scale of one. This synthesis opens the door to a simpler scheme, whose default values are near-optimal. This in turn facilitates a more efficient sweeping strategy, with u-$μ$P models reaching a loss that is equal to or lower than comparable $μ$P models and working out-of-the-box in FP8.

For agents

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

get_harvested_code_for_paper("2407.17465")
get_code_for_paper("2407.17465")
have("2407.17465")

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

curl https://syntology.ai/api/ran/2407.17465.json

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

[![Syntology run record](https://syntology.ai/api/ran/2407.17465.svg)](https://syntology.ai/paper/2407.17465)

Connect an agent — have() is free.