Remove deprecated components for 0.6.0 (#2293)

Author: RdoubleA

torchtune/modules/peft/_utils.py

@@ -9,7 +9,6 @@
 
 import torch
 from torch import nn
-from torchtune.utils._logging import deprecated
 
 # Modules from MultiHeadAttention that LoRA can be applied to
 LORA_ATTN_MODULES = Literal["q_proj", "k_proj", "v_proj", "output_proj"]
@@ -313,17 +312,3 @@ def validate_missing_and_unexpected_for_lora(
                 raise AssertionError(f"Missing LoRA key {k} from adapter state dict")
     if lora_unexpected:
         raise AssertionError("Unexpected key loading adapter")
-
-
-@deprecated(
-    msg="load_dora_magnitudes will be deprecated in 0.6.0. Please use DoRALinear.initialize_dora_magnitude instead."
-)
-def load_dora_magnitudes(model: nn.Module) -> None:
-    """
-    For DoRA magnitude we use setattr to move from meta device
-    """
-    dora_parents = {
-        n: p for n, p in model.named_modules() if hasattr(p, "adapter_params")
-    }
-    sd = {f"{n}.magnitude": p.magnitude for n, p in dora_parents.items()}
-    model.load_state_dict(sd, strict=False, assign=True)

torchtune/training/quantization.py

@@ -5,7 +5,6 @@
 # LICENSE file in the root directory of this source tree.
 
 from typing import Callable, Optional
-from warnings import warn
 
 from torch import nn
 from torchtune.modules.peft.lora import LoRALinear, QATLoRALinear
@@ -144,6 +143,7 @@ class Int4WeightOnlyQATQuantizerModuleSwap(Int4WeightOnlyQATQuantizer):
     "4w-qat-module-swap"
 ] = enable_4w_fake_quant_module_swap
 
+
 # int8 dynamic activations + int4 weight
 class Int8DynActInt4WeightQATQuantizerModuleSwap(Int8DynActInt4WeightQATQuantizer):
     pass
@@ -179,12 +179,7 @@ def get_quantizer_mode(quantizer: Optional[Callable]) -> Optional[str]:
     Returns:
         Optional[str]: The quantization mode.
     """
-    mode = _quantizer_to_mode.get(type(quantizer), None)
-    if mode is not None and "module-swap" in mode:
-        warn(
-            "*QuantizerModuleSwap is deprecated. Please use the version without 'ModuleSwap' instead"
-        )
-    return mode
+    return _quantizer_to_mode.get(type(quantizer), None)
 
 
 def _get_disable_fake_quant(quantizer_mode: str) -> Callable: