nemo_rl.models.huggingface.common#

Module Contents#

Classes#

ModelFlag

Enum that defines special flags for model-specific behaviors.

Functions#

API#

class nemo_rl.models.huggingface.common.ModelFlag(*args, **kwds)[source]#

Bases: enum.Enum

Enum that defines special flags for model-specific behaviors.

This enum provides a way to identify models that require special handling or configuration in different parts of the NeMo RL codebase.

Flags: SKIP_DTENSOR_TIED_WEIGHTS_CHECK: Models that should skip the tied weights check for the DTensor Policy even without setting the NRL_SKIP_TIED_WEIGHT_CHECK flag. VLLM_LOAD_FORMAT_AUTO: Models that should use the “auto” load format when initializing VLLM.

Each flag has a matches method that determines if the flag applies to a given model_name.

Initialization

SKIP_DTENSOR_TIED_WEIGHTS_CHECK#

‘auto(…)’

VLLM_LOAD_FORMAT_AUTO#

‘auto(…)’

matches(model_name: str) bool[source]#
nemo_rl.models.huggingface.common.is_gemma3_model(model_name: str) bool[source]#