Fix inaccuracies and add NPU constraints in ascend_npu_profiling.mdx. (#28283)

docs_new/docs/hardware-platforms/ascend-npus/ascend_npu_profiling.mdx

@@ -44,12 +44,6 @@ sglang serve \
   --max-running-requests 128
 ```
 
-<Note>
-On Ascend NPU, SGLang uses `torch_npu._apply_patches()` to automatically
-redirect PyTorch Profiler's CUDA activity to NPU, so
-`activities: ["CPU", "GPU"]` actually captures NPU operator events.
-</Note>
-
 **Profiling-related environment variables:**
 
 <table>
@@ -79,6 +73,10 @@ redirect PyTorch Profiler's CUDA activity to NPU, so
   </tbody>
 </table>
 
+<Note>
+Priority order: **API parameter > Environment variable > Default value**.
+</Note>
+
 ### 2. Collection Methods
 
 SGLang provides four collection methods. The core differences are **whether you
@@ -227,6 +225,12 @@ sending `/stop_profile`.
 shown in the server log as `Profiling done. Traces are saved to: <path>`.
 </Note>
 
+<Note>
+On Ascend NPU, SGLang uses `torch_npu._apply_patches()` to automatically
+redirect PyTorch Profiler's CUDA activity to NPU, so
+`activities: ["CPU", "GPU"]` actually captures NPU operator events.
+</Note>
+
 **`bench_serving --profile` parameters:**
 
 <table>
@@ -256,7 +260,9 @@ python3 -m sglang.profiler \
   --cpu --gpu &
 ```
 ```bash Command
-# Terminal 2: Immediately send inference requests to provide steps for profiling
+# Terminal 2: Wait for "Waiting for N steps" output from Terminal 1, then send requests.
+# The profiler starts recording once /start_profile is received by the server.
+# Requests sent before the server receives /start_profile will not be captured.
 curl http://127.0.0.1:30000/generate \
   -H "Content-Type: application/json" \
   -d '{"text": "Hello", "sampling_params": {"max_new_tokens": 32}}'
@@ -352,7 +358,8 @@ set of supported parameters:
     <tr>
       <td><code>activities</code></td>
       <td>
-        Activity types: CPU, GPU, MEM, RPD. On Ascend NPU, primarily CPU and GPU
+        Activity types: CPU, GPU, MEM, RPD. On Ascend NPU, only CPU and GPU are supported.
+        MEM depends on CUDA memory APIs and will be silently ignored. RPD requires ROCm and will cause an error.
       </td>
       <td><code>["CPU", "GPU"]</code></td>
     </tr>
@@ -426,7 +433,7 @@ Waiting for 10 steps and the trace to be flushed.... (profile_by_stage=False)
 The directory structure is
 `<output_dir>/<hostname>_<pid>_<timestamp>_ascend_pt/`. When using Method C
 (`bench_serving --profile`), a timestamp subdirectory is added:
-`<output_dir>/<timestamp>/`. Always check the server log for the exact path:
+`<output_dir>/<timestamp>/<hostname>_<pid>_<timestamp>_ascend_pt/`. Always check the server log for the exact path:
 `Profiling done. Traces are saved to: <path>`.
 
 ### 5. Viewing Results
@@ -465,12 +472,11 @@ files — **no need to manually call `analyse()`**:
 </Note>
 
 <Note>
-If you need to merge distributed trace files in a multi-node deployment, set
-`"merge_profiles": true` in the `/start_profile` request. Note: on Ascend NPU,
-the merger has limited support for the `*_ascend_pt` format — check
-`trace_view.json` on each node individually. See
-[Benchmark and Profiling](/docs/developer_guide/benchmark_and_profiling#profiler-trace-merger-for-distributed-traces)
-for details.
+In a multi-node deployment, each node generates its own trace files.
+To merge them into a single unified trace, set `"merge_profiles": true` in the `/start_profile` request.
+However, on Ascend NPU, the merge feature does **not** fully support the `*_ascend_pt` format — merged results may be incomplete or incorrect.
+It is recommended to view `trace_view.json` on each node individually instead.
+For more details, see [Benchmark and Profiling](/docs/developer_guide/benchmark_and_profiling#profiler-trace-merger-for-distributed-traces).
 </Note>
 
 ### 6. Re-parsing Raw Data (Optional)
@@ -487,6 +493,8 @@ analyse("./sglang_profile/<hostname>_*_ascend_pt/")
 <Note>
 Normally **no need** to manually run `analyse()` — the server already parses
 data automatically. Only use this for re-parsing or handling interrupted data.
+Running `analyse()` when ASCEND_PROFILER_OUTPUT already exists will overwrite the original directory.
+If the original data is still needed, back it up before running `analyse()`.
 </Note>
 
 ## Best Practices