diff --git a/csrc/sparse/cutlass/sparse_scaled_mm_entry.cu b/csrc/sparse/cutlass/sparse_scaled_mm_entry.cu
index 38b929be4..dbed5fa4e 100644
--- a/csrc/sparse/cutlass/sparse_scaled_mm_entry.cu
+++ b/csrc/sparse/cutlass/sparse_scaled_mm_entry.cu
@@ -6,11 +6,11 @@
 #include "cutlass_extensions/common.hpp"
 
 bool cutlass_sparse_scaled_mm_supported(int64_t cuda_device_capability) {
-  // sparse CUTLASS kernels need at least
+  // sparse CUTLASS kernels need exactly hopper and are not forward compatible
   //   CUDA 12.2 and SM90 (Hopper)
 
 #if defined CUDA_VERSION
-  return CUDA_VERSION >= 12020 && cuda_device_capability >= 90;
+  return CUDA_VERSION >= 12020 && cuda_device_capability == 90;
 #endif
 
   return false;
@@ -98,7 +98,7 @@ std::vector<torch::Tensor> cutlass_sparse_compress(torch::Tensor const& a) {
 
   TORCH_CHECK_NOT_IMPLEMENTED(
       false,
-      "No compiled cutlass_sparse_compress for a compute capability less than "
+      "No compiled cutlass_sparse_compress for a compute capability equal to "
       "CUDA device capability: ",
       version_num);
 }
diff --git a/vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors.py b/vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors.py
index df3d733b7..9de2228b7 100644
--- a/vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors.py
+++ b/vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors.py
@@ -207,18 +207,19 @@ class CompressedTensorsConfig(QuantizationConfig):
         # because Attention quantization on its own is not supported by vLLM.
         # It is coupled with KV-cache quantization, and if scales are present in the
         # checkpoint, they will be used properly.
-        grps_without_attn_quant = {}
-        for k, v in config["config_groups"].items():
-            # e.g. LlamaAttention, Qwen3Attention, etc.
-            if len(v["targets"]) == 1 and v["targets"][0].endswith("Attention"):
-                logger.warning(
-                    "Skipping CompressedTensors config group for %s. Attention quant "
-                    "is coupled with KV-cache quantization in vLLM.",
-                    v["targets"][0],
-                )
-                continue
-            grps_without_attn_quant[k] = v
-        config["config_groups"] = grps_without_attn_quant
+        if "config_groups" in config:
+            grps_without_attn_quant = {}
+            for k, v in config["config_groups"].items():
+                # e.g. LlamaAttention, Qwen3Attention, etc.
+                if len(v["targets"]) == 1 and v["targets"][0].endswith("Attention"):
+                    logger.warning(
+                        "Skipping CompressedTensors config group for %s. Attention "
+                        "quant is coupled with KV-cache quantization in vLLM.",
+                        v["targets"][0],
+                    )
+                    continue
+                grps_without_attn_quant[k] = v
+            config["config_groups"] = grps_without_attn_quant
 
         ignore: list[str] = cast(list[str], config.get("ignore", []))
         quant_format = cast(str, config.get("format"))
diff --git a/vllm/model_executor/model_loader/weight_utils.py b/vllm/model_executor/model_loader/weight_utils.py
index 7025efd1c..43ea6f285 100644
--- a/vllm/model_executor/model_loader/weight_utils.py
+++ b/vllm/model_executor/model_loader/weight_utils.py
@@ -261,6 +261,7 @@ def get_quant_config(
     if (
         hf_quant_config is not None
         and hf_quant_config.get("quant_method") == "compressed-tensors"
+        and "config_groups" in hf_quant_config
     ):
         if hf_text_config is not None:
             n_heads = getattr(hf_text_config, "num_attention_heads", None)