XProf GPU - Expose # of Chips to Profile

PiperOrigin-RevId: 829570660

Author: Becker-ZH

docs/profiling.md

@@ -363,6 +363,9 @@ The following options are available for GPU profiling:
 *   `gpu_pm_sample_interval_us`: Sets the sampling interval in microseconds
     for CUPTI PM sampling. Defaults to `500`.
 *   `gpu_pm_sample_buffer_size_per_gpu_mb`: Sets the system memory buffer size per device in MB for CUPTI PM sampling. Defaults to 64MB. The maximum supported value is 4GB.
+*   `gpu_num_chips_to_profile_per_task`: Specifies the number of GPU devices to
+    profile per task. If set to 0, all available GPUs will be profiled. This can
+    be used to decrease the trace collection size.
 *   `gpu_dump_graph_node_mapping`: If enabled, dumps CUDA graph node
     mapping information into the trace. Defaults to `False`.