Please ensure that you have installed the Cerebras Model Zoo package by going through the installation guide. Note that EEH version tested and packaged in the Cerebras Model Zoo is the official release v0.4.7.
This section covers the required steps for setting up an EEH run to perform standalone downstream validation on various tasks.
In particular, you will need to write a YAML configuration file to configure an instance of the Trainer callback.
The example in this section configures evaluation for LLaMA3 8B via the multiple choice (non-generative) eval harness task winogrande using a single CSX.
If you aren’t interested in seeing the break down of the configuration, feel free to skip ahead to the Putting it All Together section to see the full YAML configuration.
To run downstream validation harness, you must specify the name setting in the model configuration. Valid names corresponding to the supported models include:
The eeh_args section exposes the following settings to configure the EEH run:
Eleuther Eval Harness CLI Arguments
Description
--tasks
Comma separated string specifying Eleuther Eval Harness tasks. To get full list of tasks, use the command lm-eval --tasks list from within your python venv.
--num_fewshot
Number of examples to be added to the fewshot context string. Defaults to 0
--output_path
The path to the output file where the result metrics will be saved. If the path is a directory and log_samples is true, the results will be saved in the directory. Else the parent directory will be used.
--limit
Accepts an integer, or a float between 0.0 and 1.0. This limits the number of documents to evaluate per task to the first X documents (if an integer) or first X% of documents. This is useful for debugging.
--use_cache
A path to a sqlite db file for caching model responses. None if not caching.
--cache_requests {true,refresh,delete}
Speed up evaluation by caching the building of dataset requests. None if not caching.
--check_integrity
Whether to run the relevant part of the test suite for the tasks.
--write_out
Prints the prompt for the first few documents. Defaults to False.
--log_samples
If True, write out all model outputs and documents for per-sample measurement and post-hoc analysis. Defaults to False.
--show_config
If True, shows the the full config of all tasks at the end of the evaluation. Defaults to False.
--include_path
Additional path to include if there are external tasks to include.
--predict_only
Use with –log_samples. Only model outputs will be saved and metrics will not be evaluated.
--seed
Set seed for python’s random, numpy and torch.
--temperature
Sampling temperature used for generation (autoregressive, generate_until tasks only).
--top_p
Top-p parameter used for nucleus sampling (autoregressive, generate_until tasks only).
--top_k
Top-k parameter used for generation (autoregressive, generate_until tasks only).
You can either specify the settings here or pass them via CLI arguments to the standalone EEH run script.
The callback configuration also accepts dataloader settings that you must specify in the YAML to set up input data preprocessing for the run:
DataLoader Settings
Description
data_dir
This setting is required. Provide a path to the mounted directory visible to the worker containers where eval harness task data samples are dumped after preprocessing. Use the mount_dirs argument to specify a dir mount, similar to our existing flows.
tokenizer_file_path
Path to a custom tokenizer (JSON) file. If you provide a custom tokenizer, then you must also specify eos_id; otherwise, you must provide a pretrained tokenizer from Hugging Face in pretrained_model_name_or_path.
pretrained_model_name_or_path
Hugging Face (HF) pretrained model name or path. This setting is required if you do not specify tokenizer_file_path. For detailed description, see HF AutoTokenizers.
eos_id
End-of-sentence (eos) token ID to signal the termination of a sequence. This setting is required if you specify a custom tokenizer in tokenizer_file_path. You can set this by looking for the ID corresponding to the eos token in the custom tokenizer JSON file.
max_sequence_length
Maximum length of the input sequence. This setting is required for preprocessing input data samples from the specified eval harness tasks. You should align the max_sequence_length field to the max_position_embeddings value in the model configuration of the YAML. If you don’t specify max_sequence_length, the flow defaults to this max_position_embeddings setting.
Additionally, you may optionally specify the following, CSX-specific eval harness setting:
keep_data_dir: Use this to preserve the preprocessed eval harness task data samples, i.e. the directory specified under data_dir. Defaults to False, i.e. data samples are deleted after the run.
Now that the the YAML configuration is complete, use the Model Zoo CLI to run EEH on various tasks.
This script accepts the following arguments:
lm_eval
Copy
cszoo lm_eval --helpusage: cszoo lm_eval [-h] [--tasks task1,task2] [--num_fewshot N] [--output_path DIR|DIR/file.json] [--limit N|0<N<1] [--use_cache DIR] [--cache_requests {true,refresh,delete}] [--check_integrity] [--write_out] [--log_samples] [--system_instruction SYSTEM_INSTRUCTION] [--apply_chat_template] [--fewshot_as_multiturn] [--show_config] [--include_path DIR] [--predict_only] [--seed SEED] [--trust_remote_code] [--max_tokens MAX_TOKENS] [--temperature TEMPERATURE] [--top_p TOP_P] [--top_k TOP_K] [--keep_data_dir] [--target_device {CSX}] [-o MODEL_DIR] [--checkpoint_path CHECKPOINT_PATH] [--load_checkpoint_states LOAD_CHECKPOINT_STATES] [--logging LOGGING] [--compile_only] [--validate_only] [--job_labels JOB_LABELS [JOB_LABELS ...]] [--job_priority {p1,p2,p3}] [--mount_dirs MOUNT_DIRS [MOUNT_DIRS ...]] [--python_paths PYTHON_PATHS [PYTHON_PATHS ...]] [--credentials_path CREDENTIALS_PATH] [--mgmt_address MGMT_ADDRESS] [--disable_version_check] [--num_csx NUM_CSX] [--debug_args [DEBUG_ARGS [DEBUG_ARGS ...]]] [--debug_args_path DEBUG_ARGS_PATH] [--ini [INI [INI ...]]] [--transfer_processes TRANSFER_PROCESSES] [--config CONFIG] paramspositional arguments: params Path to .yaml file with model parameters.optional arguments: -h, --help show this help message and exit --system_instruction SYSTEM_INSTRUCTION System instruction to be used in the prompt --apply_chat_template If True, applies the chat template to the prompt --fewshot_as_multiturn If True, uses the fewshot as a multi-turn conversation --target_device {CSX} Target device to run on. Can be one of CSX. -o MODEL_DIR, --model_dir MODEL_DIR Model directory where checkpoints will be written. --checkpoint_path CHECKPOINT_PATH Checkpoint to initialize weights from. --load_checkpoint_states LOAD_CHECKPOINT_STATES Comma-separated string of keys to explicitly specify the components whose state should be loaded if present in a checkpoint. If this flag is used, then all component states that exist in a checkpoint, but are not specified to load via the flag will be ignored. For example, for fine-tuning runs on a different dataset, setting `--load_checkpoint_states="model" will only load the model state; any `optimizer` or `dataloader` state present in the checkpoint will not be loaded. By default, the config is `all`, i.e. everything present in the checkpoint is loaded. --logging LOGGING Specifies the default logging level. Defaults to INFO. --compile_only Enables compile only workflow. --validate_only Enables validate only workflowvalidate_only stops the compilation at ws_km stage for weight streaming mode. --job_labels JOB_LABELS [JOB_LABELS ...] A list of equal-sign-separated key value pairs served as job labels. --job_priority {p1,p2,p3} Priority of the job. When launching jobs, valid priority should be between p1 and p3, where p1 is highest priority. --mount_dirs MOUNT_DIRS [MOUNT_DIRS ...] A list of paths to be mounted to the appliance containers. It should generally contain path to the directory containing the Cerebras modelzoo. --python_paths PYTHON_PATHS [PYTHON_PATHS ...] A list of paths to be exported into PYTHONPATH for worker containers. It should generally contain path to the directory containing the Cerebras modelzoo, as well as any external python packages needed. --credentials_path CREDENTIALS_PATH Credentials for cluster access. Defaults to None. If None, the value from a pre-configured location will be used if available. --mgmt_address MGMT_ADDRESS <host>:<port> for cluster management. If None, the value from a pre-configured location will be used if available. Defaults to None. --disable_version_check Disable version check for local experimentation and debugging --num_csx NUM_CSX Number of CS nodes. Defaults to 1 --debug_args [DEBUG_ARGS [DEBUG_ARGS ...]] DebugArgs to pass to the Cerebras compile and execution, pass as --debug_args sub.object.key=value, where value can be bool int, float or str --debug_args_path DEBUG_ARGS_PATH Path to debugs args file. Defaults to None. --ini [INI [INI ...]] Debug INI settings to pass to the Cerebras compile and execution, pass as --ini key=value, where value can be bool, int, float or str --transfer_processes TRANSFER_PROCESSES Number of processes to use when transferring weights. --config CONFIG Specifies a specific key of the params file to return.Eleuther Eval Harness Arguments: --tasks task1,task2, -t task1,task2 Comma-separated list of task names or task groupings to evaluate on. To get full list of tasks, use one of the commands `lm-eval --tasks {{list_groups,list_subtasks,list_tags,list}}` to list out all available names for task groupings; only (sub)tasks; tags; or all of the above --num_fewshot N, -f N Number of examples in few-shot context --output_path DIR|DIR/file.json The path to the output file where the result metrics will be saved. If the path is a directory and log_samples is true, the results will be saved in the directory. Else the parent directory will be used. --limit N|0<N<1, -L N|0<N<1 Limit the number of examples per task. If <1, limit is a percentage of the total number of examples. --use_cache DIR A path to a sqlite db file for caching model responses. `None` if not caching. --cache_requests {true,refresh,delete} Speed up evaluation by caching the building of dataset requests. `None` if not caching. --check_integrity Whether to run the relevant part of the test suite for the tasks. --write_out, -w Prints the prompt for the first few documents. --log_samples, -s If True, write out all model outputs and documents for per-sample measurement and post-hoc analysis. Use with --output_path. --show_config If True, shows the the full config of all tasks at the end of the evaluation. --include_path DIR Additional path to include if there are external tasks to include. --predict_only, -x Use with --log_samples. Only model outputs will be saved and metrics will not be evaluated. --seed SEED Set seed for python's random, numpy, torch, and fewshot sampling. Accepts a comma-separated list of 4 values for python's random, numpy, torch, and fewshot sampling seeds, respectively, or a single integer to set the same seed for all four. The values are either an integer or 'None' to not set the seed. Default is `0,1234,1234,1234` (for backward compatibility). E.g. `--seed 0,None,8,52` sets `random.seed(0)`, `torch.manual_seed(8)`, and fewshot sampling seed to 52. Here numpy's seed is not set since the second value is `None`. E.g, `--seed 42` sets all four seeds to 42. --trust_remote_code Sets trust_remote_code to True to execute code to create HF Datasets from the Hub --max_tokens MAX_TOKENS Maximum number of tokens to generate. --temperature TEMPERATURE Sampling temperature used for generation. --top_p TOP_P Top-p parameter used for nucleus sampling. --top_k TOP_K Top-k parameter used for generation. --keep_data_dir Specifies whether dumped data samples should be kept for reuse. Defaults to False, i.e. data samples are deleted after the run.
We support a subset of Eleuther’s command line interface (CLI) arguments above. For a more detailed descrition of these supported arguments, see the EEH documentation.
You may also specify these arguments in the YAML under the eeh_args key of the EleutherEvalHarness configuration, but please note that the CLI setting will override the settings in the YAML.
The params CLI argument is required. Use it to specify the path to the YAML configuration file.
Use the --checkpoint_path CLI argument to specify the path to the checkpoint file to load model weights from. If a checkpoint path is not provided, we support checkpoint autoloading in this flow such that the latest checkpoint file will be picked up from the specified model_dir.
You may perform downstream validation on all EEH tasks with output_type: loglikelihood or output_type: multiple_choice in the task specification. See asdiv and arc_easy for respective examples. You may specify each of these types of tasks separately or together in a single EleutherEvalHarness callback.
We currently do not support eval harness tasks with output_type: loglikelihood_rolling.
We currently do not support running multiple generative eval harness tasks in the same callback.
EEH task groups, such as agieval, comprise multiple generative sub tasks that you will have to configure in the YAML via separate callbacks.
Please turn on grad accumulation and choose a small micro batch size (between 16 to 32) under the flags configuration of the EleutherEvalHarness callback of the YAML,
Let’s assume that the YAML configuration file above is written to ./llama3_8B_eeh.yaml. Then, to run evaluation for task winogrande, please set up a bash script as follows:
The EEH flow also supports running generative (autoregressive) eval harness tasks, i.e. tasks that specify output_type: generate_until in the task specification, such as triviaqa or drop. Refer to task triviqa for an example specification from the official EEH repository.
In order to run generative inference on CSX, you must specify the following inference settings in the model config of YAML file:
start_token - ID of the special token that indicates where to start inferring for each sample, as described above. You may specify a list of token IDs instead of a single ID. If you do, the model will start inference at the first token that matches any one of the provided IDs. The model will pad inferred predictions with the first ID in the list.
stop_sequences - List of sequences (each one being a list of token IDs). If any one of these sequences is emitted by the model, inference will stop for that sample. For example, suppose you would like to stop inferring after either a newline character (e.g. token id 1), or a combination of a period (e.g. token id 2) followed by a space (e.g. token id 3). In this case, set stop_sequences to [[1], [2, 3]]. To stop inferring after seeing a newline character only, set stop_sequences to [[1]]. To disable this feature, set stop_sequences to an empty list []. Additionally, the following optional parameters may be set:
max_tokens - Maximum tokens to infer for each sample.
loop_dim - Indicates the sequence dimension in the input and output data. Default value is 1. If set to 0, indicates that both input and output data is transposed (i.e. sequence X samples instead of samples X sequence).
For your LLaMA3 8B example, please update the YAML file ./llama3_8B_eeh.yaml to add these inference settings:
For start_token, it is ideal to choose a value that’s not going to be generated by the model, i.e. vocab_size in the example above.
The generative task itself defines stop_sequences under setting generation_kwargs.until of the task spec. For instance, triviqa specifies "\n", "." and "," as the stop tokens. The EEH flow will internally override the stop_sequences config with the value from the task, so you can also specify an arbitrary value in the YAML.
Finally, please update the bash script as follows:
By default, the model will perform greedy sampling of the inferred tokens, i.e. for all of the model’s outputs, pick the token with the highest probability.
In order to perform non-greedy sampling, you can pass in temperature, top_k or top_p to either the bash script or under eeh_args of the YAML. For example:
Since inference settings are baked into the model configuration and that inference requires different resources, there is a separate compile and execution for running downstream validation on generative tasks.
You may specify at most one generative task at a time (per callback).
In order to run multiple generative tasks, you may update the run script to the following:
