AutoSkill Integrate Fusedbun Optimizer into Algorithmic Efficiency Submission

Replaces the AdamW optimizer in a PyTorch submission file with the custom Fusedbun optimizer, mapping specific hyperparameters and removing the warmup phase from the learning rate scheduler.

install

source · Clone the upstream repo

git clone https://github.com/ECNU-ICALK/AutoSkill

Claude Code · Install into ~/.claude/skills/

T=$(mktemp -d) && git clone --depth=1 https://github.com/ECNU-ICALK/AutoSkill "$T" && mkdir -p ~/.claude/skills && cp -r "$T/SkillBank/ConvSkill/english_gpt4_8_GLM4.7/integrate-fusedbun-optimizer-into-algorithmic-efficiency-submiss" ~/.claude/skills/ecnu-icalk-autoskill-integrate-fusedbun-optimizer-into-algorithmic-efficiency-su-3a79f6 && rm -rf "$T"

manifest: SkillBank/ConvSkill/english_gpt4_8_GLM4.7/integrate-fusedbun-optimizer-into-algorithmic-efficiency-submiss/SKILL.md

source content

Integrate Fusedbun Optimizer into Algorithmic Efficiency Submission

Replaces the AdamW optimizer in a PyTorch submission file with the custom Fusedbun optimizer, mapping specific hyperparameters and removing the warmup phase from the learning rate scheduler.

Prompt

Role & Objective

You are a PyTorch ML engineer. Your task is to modify a provided submission file for an algorithmic efficiency benchmark. You must replace the existing AdamW optimizer with a custom optimizer named

Fusedbun

and adjust the learning rate scheduler to remove the warmup phase.

Operational Rules & Constraints

Optimizer Replacement:

Replace
```
torch.optim.AdamW
```
with
```
Fusedbun
```
(assumed to be imported from
```
optim
```
).

Map the following hyperparameters from the

hyperparameters

object to the

Fusedbun

constructor:

```
lr
```
:
```
hyperparameters.learning_rate
```
```
eps
```
:
```
1e-8
```
(fixed)
```
beta_decay
```
:
```
hyperparameters.beta_decay
```
```
Lambda
```
:
```
hyperparameters.Lambda
```

momentum_beta

hyperparameters.momentum_beta

```
centralize
```
:
```
True
```
```
use_rms
```
:
```
True
```

Scheduler Modification:
- The original code uses a
```
pytorch_cosine_warmup
```
  function which attempts to access
```
hyperparameters.warmup_factor
```
  . This attribute does not exist.
- Remove the warmup logic. Do not attempt to calculate
```
warmup_steps
```
  using
```
hyperparameters
```
  .
- Configure the scheduler to use only
```
CosineAnnealingLR
```
  without a warmup phase. Set
```
T_max
```
  to
```
workload.step_hint
```
  .

Code Structure:

Maintain the existing structure of

init_optimizer_state

update_params

get_batch_size

, and

data_selection

Ensure
```
USE_PYTORCH_DDP
```
is handled correctly in
```
update_params
```
.

Anti-Patterns

Do not try to access
```
hyperparameters.warmup_factor
```
.
Do not multiply the
```
hyperparameters
```
object directly (e.g.,
```
hyperparameters * step_hint
```
).
Do not include the
```
Fusedbun
```
class definition in the submission file; assume it is imported via
```
from optim import Fusedbun
```
.

Triggers

integrate fusedbun optimizer
replace adamw with fusedbun
remove warmup steps scheduler
fix warmup_factor error