Small param fixes (#1854)

SUMMARY:
Two small fixes:
1. `splits` isn't included in the `oneshot` function signature even
though it is used that way in multiple tests. There are also a few other
args that are missing (which aren't used by tests currently), opened
#1850 to track.
2. `sequential_update` is deprecated but was still listed in a test and
a couple examples. Removed.

---------

Signed-off-by: Fynn Schmitt-Ulms <[email protected]>

Author: fynnsu

examples/sparse_2of4_quantization_fp8/README.md

@@ -57,7 +57,6 @@ recipe = [
     SparseGPTModifier(
         sparsity=0.5,
         mask_structure="2:4",
-        sequential_update=True,
         targets=[r"re:model.layers.\d*$"],
     )
 ]

examples/sparse_2of4_quantization_fp8/llama3_8b_2of4.py

@@ -45,7 +45,6 @@ def get_recipe(fp8_enabled):
         SparseGPTModifier(
             sparsity=0.5,
             mask_structure="2:4",
-            sequential_update=True,
             targets=[r"re:model.layers.\d*$"],
         )
     ]

src/llmcompressor/entrypoints/oneshot.py

@@ -9,7 +9,7 @@
 
 import os
 from datetime import datetime
-from typing import TYPE_CHECKING, List, Optional, Union
+from typing import TYPE_CHECKING, Dict, List, Optional, Union
 
 from loguru import logger
 from torch.utils.data import DataLoader
@@ -230,6 +230,7 @@ def oneshot(
     dataset: Optional[Union[str, "Dataset", "DatasetDict"]] = None,
     dataset_config_name: Optional[str] = None,
     dataset_path: Optional[str] = None,
+    splits: Optional[Union[str, List, Dict]] = None,
     num_calibration_samples: int = 512,
     shuffle_calibration_samples: bool = True,
     max_seq_length: int = 384,
@@ -288,6 +289,7 @@ def oneshot(
     :param dataset_config_name: The configuration name of the dataset
         to use.
     :param dataset_path: Path to a custom dataset. Supports json, csv, dvc.
+    :param splits: Optional percentages of each split to download.
     :param num_calibration_samples: Number of samples to use for one-shot
         calibration.
     :param shuffle_calibration_samples: Whether to shuffle the dataset before

src/llmcompressor/modifiers/obcq/sgpt_base.py

@@ -1,4 +1,3 @@
-import warnings
 from abc import abstractmethod
 from collections import defaultdict
 from functools import partial
@@ -34,7 +33,6 @@ class SparsityModifierBase(Modifier):
     owl_lmbda: Optional[float] = None
 
     # data pipeline arguments
-    sequential_update: Optional[bool] = False  # deprecated
     sequential_targets: Union[str, List[str], None] = None
     targets: Union[str, List[str]] = ["Linear"]
     ignore: List[str] = Field(default_factory=list)
@@ -46,17 +44,6 @@ class SparsityModifierBase(Modifier):
     _target_layers: Dict[str, torch.nn.Module] = PrivateAttr(default_factory=dict)
     _module_sparsities: Dict[torch.nn.Module, str] = PrivateAttr(default_factory=dict)
 
-    @field_validator("sequential_update", mode="before")
-    def validate_sequential_update(cls, value: bool) -> bool:
-        if not value:
-            warnings.warn(
-                "`sequential_update=False` is no longer supported, setting "
-                "sequential_update=True",
-                DeprecationWarning,
-            )
-
-        return True
-
     @field_validator("sparsity_profile", mode="before")
     def validate_sparsity_profile(cls, value: Optional[str]) -> bool:
         if value is None:

src/llmcompressor/modifiers/quantization/gptq/base.py

@@ -1,5 +1,4 @@
 import contextlib
-import warnings
 from typing import Dict, List, Optional, Tuple, Union
 
 import torch
@@ -17,7 +16,7 @@
     update_offload_parameter,
 )
 from loguru import logger
-from pydantic import PrivateAttr, field_validator
+from pydantic import PrivateAttr
 
 from llmcompressor.core import Event, EventType, State
 from llmcompressor.modifiers import Modifier
@@ -108,7 +107,6 @@ class GPTQModifier(Modifier, QuantizationMixin):
     """
 
     # gptq modifier arguments
-    sequential_update: bool = True  # DEPRECATED
     sequential_targets: Union[str, List[str], None] = None
     block_size: int = 128
     dampening_frac: Optional[float] = 0.01
@@ -310,14 +308,3 @@ def _maybe_onload_hessian(self, module: torch.nn.Module):
         if self.offload_hessians:
             if module in self._hessians:  # may have been deleted in context
                 self._hessians[module] = self._hessians[module].to(device="cpu")
-
-    @field_validator("sequential_update", mode="before")
-    def validate_sequential_update(cls, value: bool) -> bool:
-        if not value:
-            warnings.warn(
-                "`sequential_update=False` is no longer supported, setting "
-                "sequential_update=True",
-                DeprecationWarning,
-            )
-
-        return True

tests/e2e/vLLM/recipes/kv_cache/gptq.yaml

@@ -1,7 +1,6 @@
 quant_stage:
   quant_modifiers:
     GPTQModifier:
-      sequential_update: false
       ignore: ["lm_head"]
       actorder: null
       config_groups: