Loaders#
Dyana offers a set of loaders for different file types, each with its own arguments and examples. By default, each loader runs in an isolated, offline container.
To view the available loaders and their descriptions:
To show the help menu for a specific loader:
AutoModel#
Loads and profiles machine learning models compatible with AutoModel and AutoTokenizer.
Optional Build Arguments: --extra-requirements
Arguments#
| Argument | Description | Default | Required |
|---|---|---|---|
--model |
Path to the model to profile. | None |
yes |
--input |
Input for the model. | This is an example sentence. |
no |
--low-memory |
Perform tokenizer and model initialization without loading weights. | None |
no |
Examples#
dyana trace --model tohoku-nlp/bert-base-japanese --input "This is an example sentence." --extra-requirements "protobuf fugashi ipadic"
Ollama#
Loads and profiles models via an Ollama server. Local models on the host machine are mounted and shared with the container.
Requires Network: no
Arguments#
| Argument | Description | Default | Required |
|---|---|---|---|
--model |
Name of the Ollama model to profile. | None |
yes |
--input |
Input for the model. | This is an example sentence. |
no |
Example#
LoRA#
Loads LoRA adapters via PEFT.
Optional Build Arguments: --extra-requirements
Arguments#
| Argument | Description | Default | Required |
|---|---|---|---|
--adapter |
Path to the LoRA adapter to profile. | None |
yes |
Example#
ELF#
Loads and profiles ELF executable files.
Arguments#
| Argument | Description | Default | Required |
|---|---|---|---|
--elf |
Path to the executable file to run. | None |
yes |
Examples#
Pickle#
Loads and profiles Python pickle files.
Optional Build Arguments: --extra-requirements
Arguments#
| Argument | Description | Default | Required |
|---|---|---|---|
--pickle |
Path to the python pickle file to deserialize. | None |
yes |
--extra-requirements |
Extra pip requirements (comma-separated list). | None |
no |
Examples#
Python#
Loads and profiles serialized Python scripts.
Requires Network: no
Optional Build Arguments: --extra-requirements
Arguments#
| Argument | Description | Default | Required |
|---|---|---|---|
--script |
Path to the script to profile. | None |
yes |
Examples#
PIP#
Loads and profiles Python package installation via PIP.
Requires Network: yes
Optional Build Arguments: --extra-dependencies
Arguments#
| Argument | Description | Default | Required |
|---|---|---|---|
--package |
PIP compatible package name or expression. | None |
yes |
Examples#
JS#
Loads and profiles JavaScript files.
Optional Build Arguments: --extra-requirements
Arguments#
| Argument | Description | Default | Required |
|---|---|---|---|
--script |
Path to the script to profile. | None |
yes |
Examples#
NPM#
Loads and profiles NodeJS package installation via NPM.
Requires Network: yes
Optional Build Arguments: --extra-dependencies
Arguments#
| Argument | Description | Default | Required |
|---|---|---|---|
--package |
NPM compatible package name or expression. | None |
yes |
Examples#
Website#
Opens a website in a headless browser and profiles its performance.
Requires Network: yes
Arguments#
| Argument | Description | Default | Required |
|---|---|---|---|
--url |
URL to open. | None |
yes |
--screenshot |
Save a screenshot of the page. | None |
no |
--wait-for |
CSS selector to wait for before profiling. | None |
no |
--wait-for-timeout |
Timeout in seconds for page load and element wait. | None |
no |
--performance-log |
Enable advanced performance logging. | None |
no |
Examples#
Megatron#
Loads and profiles Megatron-LM DMC models for efficient inference.
Optional Build Arguments: --extra-requirements
Arguments#
| Argument | Description | Default | Required |
|---|---|---|---|
--model |
Path to model checkpoint. The tokenizer is auto-detected from the same directory when not provided explicitly. | None |
yes |
--size |
Model size such as 7B or 13B. |
None |
no |
--input |
Input text for inference. | This is an example prompt. |
no |
--tokenizer |
Explicit path to a tokenizer file. | None |
no |