diff --git a/csrc/quantization/awq/dequantize.cuh b/csrc/quantization/awq/dequantize.cuh
index 9106ea3779f4b..d1d926de18d78 100644
--- a/csrc/quantization/awq/dequantize.cuh
+++ b/csrc/quantization/awq/dequantize.cuh
@@ -16,7 +16,7 @@ namespace awq {
 
 __device__ uint4 dequantize_s4_to_fp16x2(uint32_t const& source)
 {
-#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 800
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 750
   assert(false);
 #else
     uint4 result;
diff --git a/csrc/quantization/awq/gemm_kernels.cu b/csrc/quantization/awq/gemm_kernels.cu
index 66f4d195f9eeb..3c5d08a18e0ab 100644
--- a/csrc/quantization/awq/gemm_kernels.cu
+++ b/csrc/quantization/awq/gemm_kernels.cu
@@ -29,7 +29,7 @@ __pack_half2(const half x, const half y) {
 
 __global__ void __launch_bounds__(64) gemm_forward_4bit_cuda_m16n128k32(int G, int split_k_iters, half* __restrict__ A, int* __restrict__ B, half* __restrict__ scaling_factors, int* __restrict__ zeros, int M, int IC, int OC, half* __restrict__ C) 
 {
-#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 800
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 750
   assert(false);
 #else
   static constexpr uint32_t ZERO = 0x0;
@@ -191,6 +191,39 @@ __global__ void __launch_bounds__(64) gemm_forward_4bit_cuda_m16n128k32(int G, i
         }
       }
       for (int j_0_4 = 0; j_0_4 < 4; ++j_0_4) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ == 750
+        {
+          __asm__ __volatile__(
+            "mma.sync.aligned.m16n8k8.row.col.f32.f16.f16.f32"
+            "{%0, %1, %2, %3}, {%4, %5}, {%6}, {%7, %8, %9, %10};\n"
+            :  "=f"(((float *)(C_warp + (j_0_4 * 8)))[0]), "=f"(((float *)(C_warp + (j_0_4 * 8)))[1]), "=f"(((float *)(C_warp + (j_0_4 * 8)))[2]), "=f"(((float *)(C_warp + (j_0_4 * 8)))[3])
+            : "r"(((unsigned *)(A_shared_warp + 0))[0]), "r"(((unsigned *)(A_shared_warp + 0))[1]), "r"(((unsigned *)(B_shared_warp + (j_0_4 * 8)))[0]), "f"(((float *)(C_warp + (j_0_4 * 8)))[0]), "f"(((float *)(C_warp + (j_0_4 * 8)))[1]), "f"(((float *)(C_warp + (j_0_4 * 8)))[2]), "f"(((float *)(C_warp + (j_0_4 * 8)))[3]));
+        }
+
+        {
+          __asm__ __volatile__(
+            "mma.sync.aligned.m16n8k8.row.col.f32.f16.f16.f32"
+            "{%0, %1, %2, %3}, {%4, %5}, {%6}, {%7, %8, %9, %10};\n"
+            :  "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[0]), "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[1]), "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[2]), "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[3])
+            : "r"(((unsigned *)(A_shared_warp + 0))[0]), "r"(((unsigned *)(A_shared_warp + 0))[1]), "r"(((unsigned *)(B_shared_warp + ((j_0_4 * 8) + 4)))[0]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[0]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[1]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[2]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[3]));
+        }
+
+        {
+          __asm__ __volatile__(
+            "mma.sync.aligned.m16n8k8.row.col.f32.f16.f16.f32"
+            "{%0, %1, %2, %3}, {%4, %5}, {%6}, {%7, %8, %9, %10};\n"
+            :  "=f"(((float *)(C_warp + (j_0_4 * 8)))[0]), "=f"(((float *)(C_warp + (j_0_4 * 8)))[1]), "=f"(((float *)(C_warp + (j_0_4 * 8)))[2]), "=f"(((float *)(C_warp + (j_0_4 * 8)))[3])
+            : "r"(((unsigned *)(A_shared_warp + 0))[2]), "r"(((unsigned *)(A_shared_warp + 0))[3]), "r"(((unsigned *)(B_shared_warp + (j_0_4 * 8)))[1]), "f"(((float *)(C_warp + (j_0_4 * 8)))[0]), "f"(((float *)(C_warp + (j_0_4 * 8)))[1]), "f"(((float *)(C_warp + (j_0_4 * 8)))[2]), "f"(((float *)(C_warp + (j_0_4 * 8)))[3]));
+        }
+
+        {
+          __asm__ __volatile__(
+            "mma.sync.aligned.m16n8k8.row.col.f32.f16.f16.f32"
+            "{%0, %1, %2, %3}, {%4, %5}, {%6}, {%7, %8, %9, %10};\n"
+            :  "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[0]), "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[1]), "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[2]), "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[3])
+            : "r"(((unsigned *)(A_shared_warp + 0))[2]), "r"(((unsigned *)(A_shared_warp + 0))[3]), "r"(((unsigned *)(B_shared_warp + ((j_0_4 * 8) + 4)))[1]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[0]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[1]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[2]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[3]));
+        }
+#else
         {
           __asm__ __volatile__(
             "mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32"
@@ -206,6 +239,8 @@ __global__ void __launch_bounds__(64) gemm_forward_4bit_cuda_m16n128k32(int G, i
             :  "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[0]), "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[1]), "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[2]), "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[3])
             : "r"(((unsigned *)(A_shared_warp + 0))[0]), "r"(((unsigned *)(A_shared_warp + 0))[1]), "r"(((unsigned *)(A_shared_warp + 0))[2]), "r"(((unsigned *)(A_shared_warp + 0))[3]), "r"(((unsigned *)(B_shared_warp + ((j_0_4 * 8) + 4)))[0]), "r"(((unsigned *)(B_shared_warp + ((j_0_4 * 8) + 4)))[1]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[0]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[1]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[2]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[3]));
         }
+
+#endif
       }
     }
   }
@@ -226,7 +261,7 @@ __global__ void __launch_bounds__(64) gemm_forward_4bit_cuda_m16n128k32(int G, i
 
 __global__ void __launch_bounds__(64) gemm_forward_4bit_cuda_m16n64k32(int G, int split_k_iters, half* __restrict__ A, int* __restrict__ B, half* __restrict__ scaling_factors, int* __restrict__ zeros, int M, int IC, int OC, half* __restrict__ C) 
 {
-#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 800
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 750
   assert(false);
 #else
   static constexpr uint32_t ZERO = 0x0;
@@ -392,7 +427,39 @@ __global__ void __launch_bounds__(64) gemm_forward_4bit_cuda_m16n64k32(int G, in
       
       for (int j_0_4 = 0; j_0_4 < 2; ++j_0_4) 
       {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ == 750
+        {
+          __asm__ __volatile__(
+            "mma.sync.aligned.m16n8k8.row.col.f32.f16.f16.f32"
+            "{%0, %1, %2, %3}, {%4, %5}, {%6}, {%7, %8, %9, %10};\n"
+            :  "=f"(((float *)(C_warp + (j_0_4 * 8)))[0]), "=f"(((float *)(C_warp + (j_0_4 * 8)))[1]), "=f"(((float *)(C_warp + (j_0_4 * 8)))[2]), "=f"(((float *)(C_warp + (j_0_4 * 8)))[3])
+            : "r"(((unsigned *)(A_shared_warp + 0))[0]), "r"(((unsigned *)(A_shared_warp + 0))[1]), "r"(((unsigned *)(B_shared_warp + (j_0_4 * 8)))[0]), "f"(((float *)(C_warp + (j_0_4 * 8)))[0]), "f"(((float *)(C_warp + (j_0_4 * 8)))[1]), "f"(((float *)(C_warp + (j_0_4 * 8)))[2]), "f"(((float *)(C_warp + (j_0_4 * 8)))[3]));
+        }
 
+        {
+          __asm__ __volatile__(
+            "mma.sync.aligned.m16n8k8.row.col.f32.f16.f16.f32"
+            "{%0, %1, %2, %3}, {%4, %5}, {%6}, {%7, %8, %9, %10};\n"
+            :  "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[0]), "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[1]), "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[2]), "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[3])
+            : "r"(((unsigned *)(A_shared_warp + 0))[0]), "r"(((unsigned *)(A_shared_warp + 0))[1]), "r"(((unsigned *)(B_shared_warp + ((j_0_4 * 8) + 4)))[0]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[0]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[1]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[2]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[3]));
+        }
+
+        {
+          __asm__ __volatile__(
+            "mma.sync.aligned.m16n8k8.row.col.f32.f16.f16.f32"
+            "{%0, %1, %2, %3}, {%4, %5}, {%6}, {%7, %8, %9, %10};\n"
+            :  "=f"(((float *)(C_warp + (j_0_4 * 8)))[0]), "=f"(((float *)(C_warp + (j_0_4 * 8)))[1]), "=f"(((float *)(C_warp + (j_0_4 * 8)))[2]), "=f"(((float *)(C_warp + (j_0_4 * 8)))[3])
+            : "r"(((unsigned *)(A_shared_warp + 0))[2]), "r"(((unsigned *)(A_shared_warp + 0))[3]), "r"(((unsigned *)(B_shared_warp + (j_0_4 * 8)))[1]), "f"(((float *)(C_warp + (j_0_4 * 8)))[0]), "f"(((float *)(C_warp + (j_0_4 * 8)))[1]), "f"(((float *)(C_warp + (j_0_4 * 8)))[2]), "f"(((float *)(C_warp + (j_0_4 * 8)))[3]));
+        }
+
+        {
+          __asm__ __volatile__(
+            "mma.sync.aligned.m16n8k8.row.col.f32.f16.f16.f32"
+            "{%0, %1, %2, %3}, {%4, %5}, {%6}, {%7, %8, %9, %10};\n"
+            :  "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[0]), "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[1]), "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[2]), "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[3])
+            : "r"(((unsigned *)(A_shared_warp + 0))[2]), "r"(((unsigned *)(A_shared_warp + 0))[3]), "r"(((unsigned *)(B_shared_warp + ((j_0_4 * 8) + 4)))[1]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[0]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[1]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[2]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[3]));
+        }
+#else
         {
           __asm__ __volatile__(
             "mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32"
@@ -408,6 +475,7 @@ __global__ void __launch_bounds__(64) gemm_forward_4bit_cuda_m16n64k32(int G, in
             :  "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[0]), "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[1]), "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[2]), "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[3])
             : "r"(((unsigned *)(A_shared_warp + 0))[0]), "r"(((unsigned *)(A_shared_warp + 0))[1]), "r"(((unsigned *)(A_shared_warp + 0))[2]), "r"(((unsigned *)(A_shared_warp + 0))[3]), "r"(((unsigned *)(B_shared_warp + ((j_0_4 * 8) + 4)))[0]), "r"(((unsigned *)(B_shared_warp + ((j_0_4 * 8) + 4)))[1]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[0]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[1]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[2]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[3]));
         }
+#endif
       }
     }
   }
diff --git a/vllm/model_executor/quantization_utils/awq.py b/vllm/model_executor/quantization_utils/awq.py
index 97a09ae6badff..9c6160fbe7c4b 100644
--- a/vllm/model_executor/quantization_utils/awq.py
+++ b/vllm/model_executor/quantization_utils/awq.py
@@ -42,8 +42,8 @@ def get_supported_act_dtypes(cls) -> List[torch.dtype]:
 
     @classmethod
     def get_min_capability(cls) -> int:
-        # The AWQ kernel only supports Ampere or newer GPUs.
-        return 80
+        # The AWQ kernel only supports Turing or newer GPUs.
+        return 75
 
     @classmethod
     def get_config_filenames(cls) -> List[str]: