[Test Fix] Fix/update test_run_compressed

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

@@ -0,0 +1,4 @@
+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"

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

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

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

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

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

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

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

@@ -1,4 +1,4 @@
 cadence: "commit"
 test_type: "regression"
-model_stub: "nm-testing/tinyllama-fp8-dynamic-compressed"
-empty_model: "TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T" 
+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/run_compressed_configs/w4a16.yaml

@@ -1,4 +1,4 @@
 cadence: "commit"
 test_type: "regression"
-model_stub: "nm-testing/tinyllama-w4a16-compressed"
-empty_model: "TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T"
+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/run_compressed_configs/w8a16_dense.yaml

@@ -1,4 +1,4 @@
 cadence: "commit"
 test_type: "regression"
-model_stub: "nm-testing/tinyllama-w8a16-dense"
-empty_model: "TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T"
+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/run_compressed_configs/w8a8.yaml

@@ -1,4 +1,4 @@
 cadence: "commit"
 test_type: "regression"
-model_stub: "nm-testing/tinyllama-w8a8-compressed"
-empty_model: "TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T"
+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/test_decompress.py

@@ -0,0 +1,105 @@
+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/decompression_configs"
+
+
+@requires_gpu
+@parameterized_class(parse_params(CONFIG_DIR))
+class TestQuantizationMatches(unittest.TestCase):
+    """
+    Test decompression - given a skeleton model and path to the optimized model,
+    write the optimized model's safetensors to the skeleton model and decompress
+    Ex. write weight_scale to skeleton model and then fp4 -> fp16
+
+    """
+
+    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()

tests/llmcompressor/transformers/compression/test_run_compressed.py

@@ -3,53 +3,139 @@
 import unittest
 
 import torch
-from compressed_tensors import QUANTIZATION_CONFIG_NAME
-from compressed_tensors.compressors import ModelCompressor
-from compressed_tensors.quantization import QuantizationStatus
+from compressed_tensors.linear.compressed_linear import CompressedLinear
+from compressed_tensors.quantization.utils import iter_named_leaf_modules
 from parameterized import parameterized_class
-from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer
+from transformers import AutoModelForCausalLM, AutoTokenizer
+from transformers.utils.quantization_config import CompressedTensorsConfig
 
 from tests.testing_utils import parse_params, requires_gpu
 
-CONFIG_DIR = "tests/llmcompressor/transformers/compression/run_compressed_configs"
+COMPRESSED_LINEAR_CONFIG_DIR = (
+    "tests/llmcompressor/transformers/compression/run_compressed_configs"
+)
 
 
 @requires_gpu
-@parameterized_class(parse_params(CONFIG_DIR))
-class TestQuantizationMatches(unittest.TestCase):
-    model_stub = None
-    empty_model = None
+@parameterized_class(parse_params(COMPRESSED_LINEAR_CONFIG_DIR))
+class TestRunCompressedDecompression(unittest.TestCase):
+    """
+    Test the run_compressed input arg to AutoModelForCausalLM, where HFQuantizer is
+    responsible for decompressing if model is compressed.
+
+    Diagram flow https://tinyurl.com/2ynb6wbu
+
+        Given an optimized model that was saved (uncompressed),
+        and saved as run_compressed (compressed), decompress the compressed model
+        and check the outputs.
+
+        All modules should be linear, runs default foward calls
+
+    Test the run_compressed input arg to AutoModelForCausalLM, where HFQuantizer is
+    responsible for decompressing if model is compressed.
+
+    Diagram flow https://tinyurl.com/2ynb6wbu
+
+
+    """
+
+    compressed_model_stub = None
+    uncompressed_model_stub = None
 
     @classmethod
     def setUpClass(cls):
         cls.test_dir = tempfile.mkdtemp()
 
-        # TODO: Give option on HFQuantizer to run run_compressed True/False
-        # currently hardcoded to True
-        cls.compressed_model = AutoModelForCausalLM.from_pretrained(
-            cls.model_stub,
+        quantization_config = CompressedTensorsConfig(run_compressed=False)
+        cls.decompressed_model = AutoModelForCausalLM.from_pretrained(
+            cls.compressed_model_stub,
             torch_dtype="auto",
             device_map="auto",
-            # run_compressed=True, # TODO: Give option on HFQuantizer
+            quantization_config=quantization_config,
         )
-        # TODO: Use ModelCompressor until decompression is supported through
-        # HFQuant/run_compressed can be turned off.
+
         cls.uncompressed_model = AutoModelForCausalLM.from_pretrained(
-            cls.empty_model,
-            torch_dtype=cls.compressed_model.dtype,
-            device_map=cls.compressed_model.device,
+            cls.uncompressed_model_stub,
+            torch_dtype=cls.decompressed_model.dtype,
+            device_map=cls.decompressed_model.device,
+        )
+
+        cls.tokenizer = AutoTokenizer.from_pretrained(cls.compressed_model_stub)
+
+    def test_compressed_matches_decompressed(self):
+        SAMPLE_INPUT = [
+            "I love 4-bit quantization because",
+            "What is the capital of France?",
+            "def fibonacci(n):",
+        ]
+
+        inputs = self.tokenizer(SAMPLE_INPUT, return_tensors="pt", padding=True).to(
+            self.decompressed_model.device
+        )
+        decompressed_output = self.tokenizer.batch_decode(
+            self.decompressed_model.generate(**inputs, max_length=50)
         )
-        config = AutoConfig.from_pretrained(cls.model_stub)
-        compression_config = getattr(config, QUANTIZATION_CONFIG_NAME, None)
-        cls.compressor = ModelCompressor.from_compression_config(compression_config)
-        cls.compressor.quantization_config.quantization_status = (
-            QuantizationStatus.FROZEN
+        uncompressed_output = self.tokenizer.batch_decode(
+            self.uncompressed_model.generate(**inputs, max_length=50)
+        )
+
+        for idx in range(len(SAMPLE_INPUT)):
+            assert decompressed_output[idx] == uncompressed_output[idx]
+
+    @classmethod
+    def tearDownClass(cls):
+        shutil.rmtree(cls.test_dir)
+        del cls.decompressed_model
+        del cls.uncompressed_model
+        torch.cuda.empty_cache()
+
+
+@requires_gpu
+@parameterized_class(parse_params(COMPRESSED_LINEAR_CONFIG_DIR))
+class TestRunCompressedForward(unittest.TestCase):
+    """
+    Given an optimized model that was saved (uncompressed),
+    and saved as run_compressed (compressed), do not decompressed the compressed model
+    and check the outputs.
+
+    All compressed model should have CompressedLinear, which has its custom forward call
+
+    """
+
+    compressed_model_stub = None
+
+    @classmethod
+    def setUpClass(cls):
+        cls.test_dir = tempfile.mkdtemp()
+
+        # Should have CompressedLinear modules
+        cls.compressed_model = AutoModelForCausalLM.from_pretrained(
+            cls.compressed_model_stub,
+            torch_dtype="auto",
+            device_map="auto",
         )
-        cls.compressor.decompress(
-            model_path=cls.model_stub, model=cls.uncompressed_model
+
+        # Should just be linear modules
+        quantization_config = CompressedTensorsConfig(run_compressed=False)
+        cls.decompressed_model = AutoModelForCausalLM.from_pretrained(
+            cls.compressed_model_stub,
+            torch_dtype=cls.compressed_model.dtype,
+            device_map=cls.compressed_model.device,
+            quantization_config=quantization_config,
         )
 
-        cls.tokenizer = AutoTokenizer.from_pretrained(cls.model_stub)
+        cls.tokenizer = AutoTokenizer.from_pretrained(cls.compressed_model_stub)
+
+    def test_compressed_linear_modules_exist(self):
+        compressed_linear_counts = 0
+        for _, submodule in iter_named_leaf_modules(
+            self.compressed_model,
+        ):
+            if isinstance(submodule, CompressedLinear):
+                compressed_linear_counts += 1
+
+        # some linear models are not compressed - ex. lm_head
+        assert compressed_linear_counts > 0
 
     def test_compressed_matches_uncompressed(self):
         SAMPLE_INPUT = [
@@ -59,21 +145,21 @@ def test_compressed_matches_uncompressed(self):
         ]
 
         inputs = self.tokenizer(SAMPLE_INPUT, return_tensors="pt", padding=True).to(
-            self.compressed_model.device
+            self.decompressed_model.device
         )
-        compressed_output = self.tokenizer.batch_decode(
-            self.compressed_model.generate(**inputs, max_length=50)
+        compressed_model_out = self.tokenizer.batch_decode(
+            self.decompressed_model.generate(**inputs, max_length=50)
         )
-        uncompressed_output = self.tokenizer.batch_decode(
-            self.uncompressed_model.generate(**inputs, max_length=50)
+        decompressed_model_out = self.tokenizer.batch_decode(
+            self.decompressed_model.generate(**inputs, max_length=50)
         )
 
         for idx in range(len(SAMPLE_INPUT)):
-            assert compressed_output[idx] == uncompressed_output[idx]
+            assert compressed_model_out[idx] == decompressed_model_out[idx]
 
     @classmethod
     def tearDownClass(cls):
         shutil.rmtree(cls.test_dir)
+        del cls.decompressed_model
         del cls.compressed_model
-        del cls.uncompressed_model
         torch.cuda.empty_cache()