diff --git a/src/cuda/memory.c b/src/cuda/memory.c
index 00857f30..bdfddf82 100755
--- a/src/cuda/memory.c
+++ b/src/cuda/memory.c
@@ -512,16 +512,17 @@ CUresult cuMemGetInfo_v2(size_t* free, size_t* total) {
         *free = *total - usage;
         LOG_INFO("after free=%ld total=%ld", *free, *total);
         return CUDA_SUCCESS;
-    } else if (limit < usage) {
-        LOG_WARN("limit < usage; usage=%ld, limit=%ld", usage, limit);
-        return CUDA_ERROR_INVALID_VALUE;
     } else {
         CUDA_OVERRIDE_CALL(cuda_library_entry,cuMemGetInfo_v2, free, total);
         LOG_INFO("orig free=%ld total=%ld limit=%ld usage=%ld",
             *free, *total, limit, usage);
         // Ensure total memory does not exceed the physical or imposed limit.
         size_t actual_limit = (limit > *total) ? *total : limit;
-        *free = (actual_limit > usage) ? (actual_limit - usage) : 0;
+        size_t clamped = (usage > limit) ? limit : usage;
+        if (usage > limit) {
+            LOG_WARN("CUDA meminfo: usage %lu exceeds limit %lu, clamping", usage, limit);
+        }
+        *free = (actual_limit > clamped) ? (actual_limit - clamped) : 0;
         *total = actual_limit;
         LOG_INFO("after free=%ld total=%ld limit=%ld usage=%ld",
             *free, *total, limit, usage);
diff --git a/src/nvml/hook.c b/src/nvml/hook.c
index 0bc16d31..71b37281 100644
--- a/src/nvml/hook.c
+++ b/src/nvml/hook.c
@@ -353,18 +353,34 @@ nvmlReturn_t _nvmlDeviceGetMemoryInfo(nvmlDevice_t device,void* memory,int versi
             return NVML_SUCCESS;
         }
     } else {
+        size_t physical_total;
         switch (version) {
         case 1:
-             ((nvmlMemory_t*)memory)->free = (limit-usage);
-             ((nvmlMemory_t*)memory)->total = limit;
-             ((nvmlMemory_t*)memory)->used = usage;
+            physical_total = ((nvmlMemory_t*)memory)->total;
+            break;
+        case 2:
+            physical_total = ((nvmlMemory_v2_t *)memory)->total;
+            break;
+        default:
+            return NVML_ERROR_INVALID_ARGUMENT;
+        }
+        size_t actual_limit = (limit > physical_total) ? physical_total : limit;
+        size_t clamped = (usage > actual_limit) ? actual_limit : usage;
+        if (usage > actual_limit) {
+            LOG_WARN("NVML meminfo: usage %lu exceeds limit %lu, clamping", usage, actual_limit);
+        }
+        switch (version) {
+        case 1:
+            ((nvmlMemory_t*)memory)->free = (actual_limit > clamped) ? (actual_limit - clamped) : 0;
+            ((nvmlMemory_t*)memory)->total = actual_limit;
+            ((nvmlMemory_t*)memory)->used = clamped;
             return NVML_SUCCESS;
         case 2:
-            ((nvmlMemory_v2_t *)memory)->free = (limit-usage);
-            ((nvmlMemory_v2_t *)memory)->total = limit;
-            ((nvmlMemory_v2_t *)memory)->used = usage;
+            ((nvmlMemory_v2_t *)memory)->free = (actual_limit > clamped) ? (actual_limit - clamped) : 0;
+            ((nvmlMemory_v2_t *)memory)->total = actual_limit;
+            ((nvmlMemory_v2_t *)memory)->used = clamped;
             return NVML_SUCCESS;
-        } 
+        }
     }
     return NVML_SUCCESS;
 }