chore: update the compressed-tensors config

aphrodite/common/config.py

@@ -180,11 +180,8 @@ class ModelConfig:
     def _parse_quant_hf_config(self):
         quant_cfg = getattr(self.hf_config, "quantization_config", None)
         if quant_cfg is None:
-            # SparseML uses a "compression_config" with a "quantization_config".
-            compression_cfg = getattr(self.hf_config, "compression_config",
-                                      None)
-            if compression_cfg is not None:
-                quant_cfg = compression_cfg.get("quantization_config", None)
+            # compress-tensors uses a "compression_config" key
+            quant_cfg = getattr(self.hf_config, "compression_config", None)
 
         return quant_cfg
 

aphrodite/modeling/model_loader/weight_utils.py

@@ -119,11 +119,9 @@ def get_quant_config(model_config: ModelConfig,
     hf_quant_config = getattr(model_config.hf_config, "quantization_config",
                               None)
     if hf_quant_config is None:
-        compression_config = getattr(model_config.hf_config,
-                                     "compression_config", None)
-        if compression_config is not None:
-            hf_quant_config = compression_config.get("quantization_config",
-                                                     None)
+        # compressed-tensors uses a compressions_config
+        hf_quant_config = getattr(model_config.hf_config, "compression_config",
+                                  None)
     if hf_quant_config is not None:
         return quant_cls.from_config(hf_quant_config)
     # In case of bitsandbytes/QLoRA, get quant config from the adapter model.

aphrodite/quantization/__init__.py

@@ -36,7 +36,7 @@ QUANTIZATION_METHODS = {
     "gptq": GPTQConfig,
     "quip": QuipConfig,
     "squeezellm": SqueezeLLMConfig,
-    "sparseml": CompressedTensorsConfig,
+    "compressed-tensors": CompressedTensorsConfig,
     "bitsandbytes": BitsAndBytesConfig,
 }