[Compressors] Remove sparse compression

src/llmcompressor/entrypoints/model_free/lifecycle.py

@@ -1,6 +1,4 @@
 import torch
-from compressed_tensors.compressors import BaseCompressor
-from compressed_tensors.config.format import _get_quant_compression_format
 from compressed_tensors.quantization import (
     QuantizationScheme,
     initialize_module_for_quantization,
@@ -20,7 +18,6 @@
     "validate_weight_for_quantization",
     "calibrate_global_scale",
     "calibrate_scale_zp",
-    "compress_module",
 ]
 
 
@@ -64,29 +61,3 @@ def calibrate_scale_zp(module: torch.nn.Linear):
     apply_calibration_status(module)
     update_weight_zp_scale(module)
     freeze_module_quantization(module)
-
-
-def compress_module(module: torch.nn.Linear):
-    scheme: QuantizationScheme = getattr(module, "quantization_scheme")
-
-    format = _get_quant_compression_format(scheme.input_activations, scheme.weights)
-    scheme.format = format.value
-
-    compressor = BaseCompressor.load_from_registry(format.value)
-    data = compressor.compress_weight(
-        module.weight,
-        quantization_args=scheme.weights,
-        scale=getattr(module, "weight_scale"),
-        zero_point=getattr(module, "weight_zero_point", None),
-        global_scale=getattr(module, "weight_global_scale", None),
-    )
-
-    # `compress_weight` is a messy api
-    delattr(module, "weight")
-    for key, value in data.items():
-        if hasattr(module, key):
-            getattr(module, key).data = value
-        else:
-            module.register_parameter(
-                key, torch.nn.Parameter(value, requires_grad=False)
-            )

src/llmcompressor/entrypoints/model_free/process.py

@@ -4,6 +4,7 @@
 from typing import Iterable
 
 import torch
+from compressed_tensors.compressors import compress_module
 from compressed_tensors.quantization import QuantizationScheme
 from compressed_tensors.utils.match import match_name
 from safetensors.torch import load_file, save_file
@@ -12,7 +13,6 @@
 from llmcompressor.entrypoints.model_free.lifecycle import (
     calibrate_global_scale,
     calibrate_scale_zp,
-    compress_module,
     initialize_quantized_linear,
     validate_weight_for_quantization,
 )

src/llmcompressor/transformers/compression/compressed_tensors_utils.py

@@ -3,7 +3,6 @@
 from functools import wraps
 
 import torch
-from accelerate.accelerator import get_state_dict_offloaded_model
 from compressed_tensors import (
     ModelCompressor,
     SparsityCompressionConfig,
@@ -15,9 +14,6 @@
 
 from llmcompressor.core import active_session
 from llmcompressor.pytorch.model_load.helpers import copy_python_files_from_model_cache
-from llmcompressor.transformers.compression.sparsity_metadata_config import (
-    SparsityConfigMetadata,
-)
 from llmcompressor.transformers.utils import RECIPE_FILE_NAME
 from llmcompressor.transformers.utils.helpers import infer_recipe_from_model_path
 
@@ -143,59 +139,14 @@ def get_model_compressor(
     :param disable_sparse_compression: bool to skip sparse compression
     """
 
-    if sparsity_config is None:
-        """
-        Case 1: No sparsity config is provided
-            1. Will either skip sparsity compression
-            2. Or we will infer sparsity from the model directly
-
-        Check recipe for applied sparsity:
-            - Set skip_sparsity_compression_stats to False if don't find a
-                sparsity structure from the recipe
-            - If we identify sparsity based on the recipe or the user
-                set skip_sparsity_compression_stats to False, generate config
-        """
-        sparsity_structure = SparsityConfigMetadata.infer_sparsity_structure(
-            model, check_only_modifiers=True
+    if (
+        sparsity_config is not None
+        or not skip_sparsity_compression_stats
+        or disable_sparse_compression
+    ):
+        logger.warning(
+            "Sparse compression is no longer supported by compressed-tensors"
         )
-        if sparsity_structure is not None:
-            skip_sparsity_compression_stats = False
-
-        if skip_sparsity_compression_stats:
-            logger.info(
-                "skip_sparsity_compression_stats set to True. Skipping sparsity "
-                "compression statistic calculations. No sparsity compressor will "
-                "be applied."
-            )
-            sparsity_config = None
-        else:
-            state_dict = get_state_dict_offloaded_model(model)
-
-            sparsity_config = SparsityConfigMetadata.from_pretrained(
-                model,
-                state_dict=state_dict,
-                compress=save_compressed,
-                quantization_format=quantization_format,
-                disable_sparse_compression=disable_sparse_compression,
-                sparsity_structure=sparsity_structure,
-            )
-    else:
-        """
-        # Case 2: User provides a Sparsity Config
-            - This is the case when there is existing sparsity in the
-                model that we'd like to account for while compressing
-            - Users should provide a SparsityConfig, conveying the model's
-                sparsity structure when saving the model
-        """
-        if sparsity_config.sparsity_structure is None:
-            logger.info(
-                "SparsityConfigMetadata provided without indicating ",
-                "the sparsity structure. Sparisty will be inferred from the model. "
-                "Consider providing the structure to skip this step ",
-            )
-            sparsity_config.sparsity_structure = (
-                SparsityConfigMetadata.infer_sparsity_structure(model)
-            )
 
     if not save_compressed:
         if quantization_format not in (None, CompressionFormat.dense.value):
@@ -209,7 +160,6 @@ def get_model_compressor(
 
     return ModelCompressor.from_pretrained_model(
         model,
-        sparsity_config_or_format=sparsity_config,
         quantization_format=quantization_format,
     )
 

tests/llmcompressor/transformers/compression/decompression_configs/fp8_dynamic.yaml

@@ -1,4 +1,3 @@
 cadence: "commit"
 test_type: "regression"
-compressed_model_stub: "nm-testing/tinyllama-fp8-dynamic-compressed"
-skeleton_model_stub: "TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T"
+compressed_model_stub: "nm-testing/tinyllama-fp8-dynamic-compressed"

tests/llmcompressor/transformers/compression/decompression_configs/w4a16.yaml

@@ -1,4 +1,3 @@
 cadence: "nightly"
 test_type: "regression"
-compressed_model_stub: "nm-testing/tinyllama-w4a16-compressed"
-skeleton_model_stub: "TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T"
+compressed_model_stub: "nm-testing/tinyllama-w4a16-compressed"

tests/llmcompressor/transformers/compression/decompression_configs/w8a16_dense.yaml

@@ -1,4 +1,3 @@
 cadence: "nightly"
 test_type: "regression"
-compressed_model_stub: "nm-testing/tinyllama-w8a16-dense"
-skeleton_model_stub: "TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T"
+compressed_model_stub: "nm-testing/tinyllama-w8a16-dense"

tests/llmcompressor/transformers/compression/decompression_configs/w8a8.yaml

@@ -1,4 +1,3 @@
 cadence: "commit"
 test_type: "regression"
-compressed_model_stub: "nm-testing/tinyllama-w8a8-compressed"
-skeleton_model_stub: "TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T"
+compressed_model_stub: "nm-testing/tinyllama-w8a8-compressed"