add logger.warning to avoid possible oom conditions (#2489)

### What this PR does / why we need it?
after capturing graph, the available memory on each NPU would be scarce,
so we introduce this warning log to remind users.

### Does this PR introduce _any_ user-facing change?
no

### How was this patch tested?
by ci.

---------

Signed-off-by: zouyida <[email protected]>
Co-authored-by: zouyida <[email protected]>

Author: zouyida2052

vllm_ascend/worker/model_runner_v1.py

@@ -2189,6 +2189,16 @@ def capture_model(self) -> None:
         elapsed_time = end_time - start_time
         npu_graph_size = start_free_npu_memory - end_free_npu_memory
         # This usually takes 5~20 seconds.
+        # TODO: remove when aclgraph is ready for deepseek and uses more reasonable memory handling.
+        LOW_MEMORY_THRESHOLD_GB = 3
+        end_free_npu_memory_GB = end_free_npu_memory / (1 << 30)
+        if end_free_npu_memory_GB < LOW_MEMORY_THRESHOLD_GB:
+            logger.warning(
+                f"Post graph compilation, the available memory on each NPU is reduced to merely {end_free_npu_memory_GB:.2f} GB. "
+                "During inference operations, subsequent memory allocations may potentially trigger Out-of-Memory(OOM) conditions. "
+                "Should you encounter this warning followed by service termination, "
+                "consider decreasing gpu-memory-utilization configuration parameter."
+            )
         logger.info("Graph capturing finished in %.0f secs, took %.2f GiB",
                     elapsed_time, npu_graph_size / (1 << 30))