add decompress tests

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

@@ -0,0 +1,4 @@
+cadence: "commit"
+test_type: "regression"
+compressed_model_stub: horheynm/TinyLlama_1.1B_Chat_v1.0_FP8_Dynamic_compressed
+uncompressed_model_stub: horheynm/TinyLlama_1.1B_Chat_v1.0_FP8_Dynamic_uncompressed

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

@@ -0,0 +1,4 @@
+cadence: "commit"
+test_type: "regression"
+compressed_model_stub: horheynm/TinyLlama_1.1B_Chat_v1.0_W4A16_G128_compressed
+uncompressed_model_stub: horheynm/TinyLlama_1.1B_Chat_v1.0_W4A16_G128_uncompressed

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

@@ -0,0 +1,4 @@
+cadence: "commit"
+test_type: "regression"
+compressed_model_stub: horheynm/TinyLlama_1.1B_Chat_v1.0_W8A16_G128_compressed
+uncompressed_model_stub: horheynm/TinyLlama_1.1B_Chat_v1.0_W8A16_G128_uncompressed

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

@@ -0,0 +1,4 @@
+cadence: "commit"
+test_type: "regression"
+compressed_model_stub: horheynm/TinyLlama_1.1B_Chat_v1.0_W8A8_Dynamic_Per_Token_compressed
+uncompressed_model_stub: horheynm/TinyLlama_1.1B_Chat_v1.0_W8A8_Dynamic_Per_Token_uncompressed

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

@@ -1,4 +1,4 @@
 cadence: "commit"
 test_type: "regression"
-compressed_model_stub: horheynm/TinyLlama_1.1B_Chat_v1.0_FP8_Dynamic_compressed
-uncompressed_model_stub: horheynm/TinyLlama_1.1B_Chat_v1.0_FP8_Dynamic_uncompressed
+compressed_model_stub: "nm-testing/tinyllama-fp8-dynamic-compressed"
+skeleton_model_stub: "TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T"

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

@@ -1,4 +1,4 @@
 cadence: "commit"
 test_type: "regression"
-compressed_model_stub: horheynm/TinyLlama_1.1B_Chat_v1.0_W4A16_G128_compressed
-uncompressed_model_stub: horheynm/TinyLlama_1.1B_Chat_v1.0_W4A16_G128_uncompressed
+compressed_model_stub: "nm-testing/tinyllama-w4a16-compressed"
+skeleton_model_stub: "TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T"

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

@@ -1,4 +1,4 @@
 cadence: "commit"
 test_type: "regression"
-compressed_model_stub: horheynm/TinyLlama_1.1B_Chat_v1.0_W8A16_G128_compressed
-uncompressed_model_stub: horheynm/TinyLlama_1.1B_Chat_v1.0_W8A16_G128_uncompressed
+compressed_model_stub: "nm-testing/tinyllama-w8a16-dense"
+skeleton_model_stub: "TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T"

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

@@ -1,4 +1,4 @@
 cadence: "commit"
 test_type: "regression"
-compressed_model_stub: horheynm/TinyLlama_1.1B_Chat_v1.0_W8A8_Dynamic_Per_Token_compressed
-uncompressed_model_stub: horheynm/TinyLlama_1.1B_Chat_v1.0_W8A8_Dynamic_Per_Token_uncompressed
+compressed_model_stub: "nm-testing/tinyllama-w8a8-compressed"
+skeleton_model_stub: "TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T"

tests/llmcompressor/transformers/compression/test_decompress.py

@@ -0,0 +1,98 @@
+import copy
+import shutil
+import tempfile
+import unittest
+
+import torch
+from compressed_tensors import QUANTIZATION_CONFIG_NAME
+from compressed_tensors.compressors import ModelCompressor
+from compressed_tensors.quantization import QuantizationStatus
+from parameterized import parameterized_class
+from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer
+
+from tests.testing_utils import parse_params, requires_gpu
+
+CONFIG_DIR = "tests/llmcompressor/transformers/compression/run_compressed_configs"
+
+
+@requires_gpu
+@parameterized_class(parse_params(CONFIG_DIR))
+class TestQuantizationMatches(unittest.TestCase):
+    compressed_model_stub = None
+    skeleton_model_stub = None
+
+    SAMPLE_INPUTS = [
+        "I love 4-bit quantization because",
+        "What is the capital of France?",
+        "def fibonacci(n):",
+    ]
+
+    @classmethod
+    def setUpClass(self):
+        self.test_dir = tempfile.mkdtemp()
+        self.tokenizer = AutoTokenizer.from_pretrained(self.compressed_model_stub)
+
+        self.compressed_model = AutoModelForCausalLM.from_pretrained(
+            self.compressed_model_stub,
+            torch_dtype="auto",
+            device_map="auto",
+        )
+
+        self.dense_model = AutoModelForCausalLM.from_pretrained(
+            self.skeleton_model_stub,
+            torch_dtype=self.compressed_model.dtype,
+            device_map=self.compressed_model.device,
+        )
+
+        assert not hasattr(
+            self.dense_model.model.layers[0].self_attn.q_proj, "weight_scale"
+        )
+
+        self.decompressed_model = None
+        config = AutoConfig.from_pretrained(self.compressed_model_stub)
+
+        compression_config = getattr(config, QUANTIZATION_CONFIG_NAME, None)
+        self.compressor = ModelCompressor.from_compression_config(compression_config)
+        self.compressor.quantization_config.quantization_status = (
+            QuantizationStatus.FROZEN
+        )
+
+        # use the model_path to load the decompressed weights into dense_model
+        dense_model = copy.deepcopy(self.dense_model)
+
+        # overwrite the weights of the dense model
+        self.compressor.decompress(
+            model_path=self.compressed_model_stub,
+            model=self.dense_model,
+        )
+
+        # self.dense_model should be decompressed
+        assert dense_model is not self.dense_model
+
+        self.decompressed_model = self.dense_model
+
+        assert hasattr(
+            self.decompressed_model.model.layers[0].self_attn.q_proj, "weight_scale"
+        )
+
+    def test_compressed_matches_uncompressed(self):
+        for input in self.SAMPLE_INPUTS:
+            inputs = self.tokenizer(input, return_tensors="pt", padding=True).to(
+                self.compressed_model.device
+            )
+            compressed_output = self.tokenizer.batch_decode(
+                self.compressed_model.generate(**inputs, max_length=50)
+            )
+            uncompressed_output = self.tokenizer.batch_decode(
+                self.decompressed_model.generate(**inputs, max_length=50)
+            )
+
+            assert compressed_output == uncompressed_output
+
+    @classmethod
+    def tearDownClass(self):
+        shutil.rmtree(self.test_dir)
+        del self.compressed_model
+        del self.dense_model
+        del self.decompressed_model
+        torch.cuda.empty_cache()