[ CI/Build ] Added E2E Test For Compressed Tensors (#5839)

Co-authored-by: Michael Goin <michael@neuralmagic.com>
Co-authored-by: Robert Shaw <rshaw@neuralmagic>

requirements-test.txt

@@ -14,6 +14,8 @@ peft
 requests
 ray
 sentence-transformers # required for embedding
+sparseml==1.8.0 # required for compressed-tensors
+compressed-tensors==0.4.0 # required for compressed-tensors
 
 # Benchmarking
 aiohttp

tests/conftest.py

@@ -176,6 +176,7 @@ class HfRunner:
         model_kwargs: Optional[Dict[str, Any]] = None,
         is_embedding_model: bool = False,
         is_vision_model: bool = False,
+        is_sparseml_model: bool = False,
     ) -> None:
         assert dtype in _STR_DTYPE_TO_TORCH_DTYPE
         torch_dtype = _STR_DTYPE_TO_TORCH_DTYPE[dtype]
@@ -193,6 +194,9 @@ class HfRunner:
         else:
             if is_vision_model:
                 auto_cls = AutoModelForVision2Seq
+            elif is_sparseml_model:
+                from sparseml.transformers import SparseAutoModelForCausalLM
+                auto_cls = SparseAutoModelForCausalLM
             else:
                 auto_cls = AutoModelForCausalLM
 

tests/models/test_compressed_tensors.py

@@ -0,0 +1,49 @@
+"""Compares vllm vs sparseml for compressed-tensors
+
+Note: vllm and sparseml do not have bitwise correctness, 
+so in this test, we just confirm that the top selected 
+tokens of the are in the top 5 selections of each other.
+"""
+
+import pytest
+
+from tests.quantization.utils import is_quant_method_supported
+
+from .utils import check_logprobs_close
+
+MODELS = [
+    "nm-testing/Meta-Llama-3-8B-Instruct-W8-Channel-A8-Dynamic-Per-Token-Test",
+]
+
+MAX_TOKENS = 32
+NUM_LOGPROBS = 5
+
+
+@pytest.mark.skipif(
+    not is_quant_method_supported("compressed-tensors"),
+    reason="compressed-tensors is not supported on this machine type.")
+@pytest.mark.parametrize("model_name", MODELS)
+def test_models(
+    vllm_runner,
+    hf_runner,
+    example_prompts,
+    model_name,
+) -> None:
+    # Run sparseml.
+    with hf_runner(model_name=model_name,
+                   is_sparseml_model=True) as sparseml_model:
+
+        sparseml_outputs = sparseml_model.generate_greedy_logprobs_limit(
+            example_prompts, MAX_TOKENS, NUM_LOGPROBS)
+
+    # Run vllm.
+    with vllm_runner(model_name=model_name) as vllm_model:
+        vllm_outputs = vllm_model.generate_greedy_logprobs(
+            example_prompts, MAX_TOKENS, NUM_LOGPROBS)
+
+    check_logprobs_close(
+        outputs_0_lst=sparseml_outputs,
+        outputs_1_lst=vllm_outputs,
+        name_0="sparseml",
+        name_1="vllm",
+    )

vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors.py

@@ -34,7 +34,8 @@ class CompressedTensorsConfig(QuantizationConfig):
         return [torch.float16, torch.bfloat16]
 
     # Need to figure it out
-    def get_min_capability(self) -> int:
+    @classmethod
+    def get_min_capability(cls) -> int:
         return 60
 
     def get_name(self) -> str: