diff --git a/conversion/granite.py b/conversion/granite.py index 796d37cca..4cbeb7617 100644 --- a/conversion/granite.py +++ b/conversion/granite.py @@ -239,8 +239,21 @@ class GraniteSwitchModel(GraniteMoeModel): def __init__(self, *args, **kwargs): super().__init__(*args, **kwargs) - # the weightless switch reserves one cache slot: one fewer block than num_hidden_layers - self.block_count = self.block_count - 1 + # The weightless switch layer(s) are counted in num_hidden_layers but carry no tensors. + # The count is not fixed, nor is it present in hparams, so we derive it here based on tensor block names. + layer_ids = { + int(m.group(1)) + for name in self.model_tensors + if (m := re.match(r"(?:model\.)?layers\.(\d+)\.", name)) + } + if not layer_ids: + raise ValueError("graniteswitch: no 'model.layers.N' tensors found in checkpoint") + n_weight_layers = max(layer_ids) + 1 + logger.info( + "gguf: (graniteswitch) num_hidden_layers %s counts %d weightless switch layer(s); using %d blocks", + self.hparams["num_hidden_layers"], self.block_count - n_weight_layers, n_weight_layers, + ) + self.block_count = n_weight_layers self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count) self._n_adapters = int(self.hparams["num_adapters"])