`nemo_rl.modelopt.models.generation.vllm_modelopt_patch`#

vLLM ModelOpt NVFP4 patches for dense rollout weight reloads.

Module Contents#

`_unwrap_vllm_model`
`_canonicalize_nvfp4_weight_scale`
`_requests_w4a16_modelopt_config`
`_is_w4a16_modelopt_quant_config`
`_modelopt_nvfp4_config_from_config`
`_convert_nvfp4_linear_kernel_format`
`_convert_w4a16_linear_kernel_format`
`_capture_modelopt_dense_param_reload_meta`
`_modelopt_dense_process_w4a16_weights`	Convert dense ModelOpt NVFP4 W4A16 weights for Marlin weight-only GEMM.
`_modelopt_dense_process_weights`	Convert dense ModelOpt NVFP4 weights after initial load or refit.
`_modelopt_dense_apply`
`prepare_modelopt_for_weight_reload`	Prepare a dense ModelOpt-vLLM model for one weight reload cycle.
`modelopt_process_weights_after_loading`	Run vLLM ModelOpt post-load processing for dense quantized layers.
`apply_modelopt_nvfp4_patches`	Patch vLLM’s dense ModelOpt NVFP4 method for rollout refits.

`_DENSE_HF_PARAMS`
`_MODELOPT_W4A16_QUANT_MODES`
`_MODELOPT_W4A16_ATTR`
`_ORIGINAL_NVFP4_CONFIG_FROM_CONFIG_ATTR`
`_ORIGINAL_LINEAR_APPLY_ATTR`
`_patched`

nemo_rl.modelopt.models.generation.vllm_modelopt_patch._DENSE_HF_PARAMS#: (‘weight’, ‘weight_scale’, ‘weight_scale_2’)

nemo_rl.modelopt.models.generation.vllm_modelopt_patch._MODELOPT_W4A16_QUANT_MODES#: ‘frozenset(…)’

nemo_rl.modelopt.models.generation.vllm_modelopt_patch._MODELOPT_W4A16_ATTR#: ‘_nrl_weight_only_w4a16’

nemo_rl.modelopt.models.generation.vllm_modelopt_patch._ORIGINAL_NVFP4_CONFIG_FROM_CONFIG_ATTR#: ‘_nrl_original_from_config’

nemo_rl.modelopt.models.generation.vllm_modelopt_patch._ORIGINAL_LINEAR_APPLY_ATTR#: ‘_nrl_original_apply’

nemo_rl.modelopt.models.generation.vllm_modelopt_patch._unwrap_vllm_model(model: torch.nn.Module) → torch.nn.Module#

nemo_rl.modelopt.models.generation.vllm_modelopt_patch._canonicalize_nvfp4_weight_scale(layer: torch.nn.Module) → None#

nemo_rl.modelopt.models.generation.vllm_modelopt_patch._requests_w4a16_modelopt_config(config: dict) → bool#

nemo_rl.modelopt.models.generation.vllm_modelopt_patch._is_w4a16_modelopt_quant_config(quant_config) → bool#

nemo_rl.modelopt.models.generation.vllm_modelopt_patch._modelopt_nvfp4_config_from_config(cls, *args, **kwargs)#

nemo_rl.modelopt.models.generation.vllm_modelopt_patch._convert_nvfp4_linear_kernel_format( quant_method, layer: torch.nn.Module, ) → None#

nemo_rl.modelopt.models.generation.vllm_modelopt_patch._convert_w4a16_linear_kernel_format(layer: torch.nn.Module) → None#

nemo_rl.modelopt.models.generation.vllm_modelopt_patch._capture_modelopt_dense_param_reload_meta( layer: torch.nn.Module, ) → None#

nemo_rl.modelopt.models.generation.vllm_modelopt_patch._modelopt_dense_process_w4a16_weights( self, layer: torch.nn.Module, ) → None#: Convert dense ModelOpt NVFP4 W4A16 weights for Marlin weight-only GEMM.

nemo_rl.modelopt.models.generation.vllm_modelopt_patch._modelopt_dense_process_weights(self, layer: torch.nn.Module) → None#: Convert dense ModelOpt NVFP4 weights after initial load or refit.

nemo_rl.modelopt.models.generation.vllm_modelopt_patch._modelopt_dense_apply( self, layer: torch.nn.Module, x: torch.Tensor, bias: torch.Tensor | None = None, ) → torch.Tensor#

nemo_rl.modelopt.models.generation.vllm_modelopt_patch.prepare_modelopt_for_weight_reload(model, device=None) → None#: Prepare a dense ModelOpt-vLLM model for one weight reload cycle.

nemo_rl.modelopt.models.generation.vllm_modelopt_patch.modelopt_process_weights_after_loading(model) → None#: Run vLLM ModelOpt post-load processing for dense quantized layers.

nemo_rl.modelopt.models.generation.vllm_modelopt_patch.apply_modelopt_nvfp4_patches() → None#: Patch vLLM’s dense ModelOpt NVFP4 method for rollout refits.