marifamd · Jul 28, 2026 · Jul 29, 2026 · Jul 29, 2026 · Jul 28, 2026 · Jul 28, 2026
@@ -4,13 +4,64 @@ Full documentation for amd_smi_lib is available at [https://rocm.docs.amd.com/pr
 
 ***All information listed below is for reference and subject to change.***
 
-## amd_smi_lib for ROCm 7.15.0
+## amd_smi_lib for ROCm 10.0
+
+### Added
+
+- **Added accelerator partition memory allocation mode API**.  
+  - New APIs: `amdsmi_get_gpu_accelerator_partition_mem_alloc_mode()`, `amdsmi_set_gpu_accelerator_partition_mem_alloc_mode()`.
+  - New enum: `amdsmi_accelerator_partition_mem_alloc_mode_t` (`AMDSMI_ACCELERATOR_PARTITION_MEM_ALLOC_CAPPING`, `AMDSMI_ACCELERATOR_PARTITION_MEM_ALLOC_ALL`).
+  - Supersedes the equivalent `compute_partition` memory allocation mode APIs, which are now deprecated.
 
 ### Changed
 
 - **Bumped the library major version to 27.0.0** (breaking).  
   - The shared library SONAME is now `libamd_smi.so.27`. Consumers linked against `libamd_smi.so.26` must relink; no source changes are required beyond the API changes listed elsewhere in this release.
 
+- **Widened five `amdsmi_gpu_metrics_t` accumulator counters from 32-bit to 64-bit** (breaking).  
+  - `gfx_activity_acc`, `mem_activity_acc`, `pcie_nak_sent_count_acc`, `pcie_nak_rcvd_count_acc`, and `pcie_lc_perf_other_end_recovery` are now `uint64_t` to match the amdgpu pmfw metrics header. Recompile callers that read these fields; the struct layout and field offsets have changed.
+
+- **Aligned the fabric telemetry and NIC firmware API surface with the unified ABI** (breaking).  
+  - `amdsmi_fabric_telem_id_to_string()` now returns `amdsmi_status_t` and writes the name through a `const char**` out-parameter, instead of returning a `const char*` directly.
+  - Renamed `amdsmi_nic_fw_t` to `amdsmi_nic_fw_entry_t`.
+  - Added the `amdsmi_fabric_type_t` enumerator `AMDSMI_FABRIC_TYPE_UALINK`; the previous spelling `AMDSMI_FABRIC_TYPE_UALLINK` is retained as a deprecated alias with the same value and is slated for removal in a future ROCm release.
+
+- **Flattened the `amdsmi_fabric_info_t` structure and removed `amdsmi_fabric_info_ver_t`**.  
+  - The intermediate `amdsmi_fabric_info_ver_t` type was removed from the public header. Its payload union is now the `fabric_info` member of `amdsmi_fabric_info_t`, and its version field is exposed directly as the top-level `fabric_version` field.
+  - Field access simplifies from `fabric_info.fabric_version.v1.<field>` to `fabric_info.v1.<field>`, and `fabric_info.version` becomes `fabric_version`.
+  - The change is ABI-preserving: field offsets and the overall structure size are unchanged. The Python `amdsmi_get_gpu_fabric_info()` dictionary keys are also unchanged.
+
+- **Prefixed public preprocessor macros with `AMDSMI_` in `amdsmi.h`** (breaking).  
+  - `MAX_SVI3_RAIL_INDEX`, `MAX_SVI3_RAIL_SELECTION`, `POWER_EFFICIENCY_MODE_4`, `POWER_EFFICIENCY_MODE_5`, and `MAX_NUMBER_OF_AFIDS_PER_RECORD` are now `AMDSMI_MAX_SVI3_RAIL_INDEX`, `AMDSMI_MAX_SVI3_RAIL_SELECTION`, `AMDSMI_POWER_EFFICIENCY_MODE_4`, `AMDSMI_POWER_EFFICIENCY_MODE_5`, and `AMDSMI_MAX_NUMBER_OF_AFIDS_PER_RECORD`. The unused `CENTRIGRADE_TO_MILLI_CENTIGRADE` macro was removed. Update references to the new names.
+
+- **Deprecated the `compute_partition` APIs in favor of `accelerator_partition` equivalents**.  
+  - `amdsmi_get_gpu_compute_partition()`, `amdsmi_set_gpu_compute_partition()`, `amdsmi_get_gpu_compute_partition_mem_alloc_mode()`, and `amdsmi_set_gpu_compute_partition_mem_alloc_mode()` are slated for removal in a future ROCm release. They now emit a `DeprecationWarning` from the Python interface and delegate to their `accelerator_partition` counterparts.
+  - The `amdsmi_compute_partition_type_t` and `amdsmi_compute_partition_mem_alloc_mode_t` enums are deprecated in favor of `amdsmi_accelerator_partition_type_t` and `amdsmi_accelerator_partition_mem_alloc_mode_t`.
+
+- **Deprecated `amdsmi_set_gpu_memory_partition()` in favor of `amdsmi_set_gpu_memory_partition_mode()`**.  
+  - `amdsmi_set_gpu_memory_partition` is slated for removal in a future ROCm release. It now emits a `DeprecationWarning` from the Python interface and functions as a wrapper of `amdsmi_set_gpu_memory_partition_mode()`.
+
+- **Namespaced `amdsmi_clk_limit_type_t` and `amdsmi_io_bw_encoding_t` enumerators with an `AMDSMI_` prefix**.  
+  - Added `AMDSMI_CLK_LIMIT_MIN`/`AMDSMI_CLK_LIMIT_MAX` and `AMDSMI_AGG_BW0`/`AMDSMI_RD_BW0`/`AMDSMI_WR_BW0`.
+  - The unprefixed names (`CLK_LIMIT_MIN`, `CLK_LIMIT_MAX`, `AGG_BW0`, `RD_BW0`, `WR_BW0`) are retained as deprecated aliases with unchanged values and are slated for removal in a future ROCm release.
+
+- **Removed the `amdsmi_gpu_driver_reload()` API and its Python binding** (breaking).  
+  - Reload the amdgpu driver out of band with `sudo modprobe -r amdgpu && sudo modprobe amdgpu` to apply memory partition changes instead.
+
+- **Removed unused public macros to match the unified ABI** (breaking).  
+  - Dropped the `AMDSMI_MAX_VF_COUNT`, `AMDSMI_MAX_DRIVER_NUM`, `AMDSMI_DFC_FW_NUMBER_OF_ENTRIES`, `AMDSMI_MAX_WHITE_LIST_ELEMENTS`, `AMDSMI_MAX_BLACK_LIST_ELEMENTS`, `AMDSMI_MAX_TA_WHITE_LIST_ELEMENTS`, `AMDSMI_MAX_ERR_RECORDS`, `AMDSMI_MAX_PROFILE_COUNT`, and `AMDSMI_PF_INDEX` defines, which were unreferenced by any API or struct.
+
+- **Deduplicated the `amdsmi_fabric_telemetry_category_t` enumerators**.  
+  - `AMDSMI_FABRIC_TELEMETRY_CATEGORY_INVALID` is now the canonical `0xFFFFFFFF` value; the previous `AMDSMI_FABRIC_TELEMETRY_CATEGORY_UNKNOWN` name (same value) is retained as a deprecated alias and is slated for removal in a future ROCm release.
+
+- **Deprecated `amdsmi_get_gpu_vram_vendor()`**. Use `amdsmi_get_gpu_vram_info()` and read the `vram_vendor` field instead. The function is retained (it now delegates to `amdsmi_get_gpu_vram_info()`, and the Python binding emits a `DeprecationWarning`) and is slated for removal in a future ROCm release.
+
+- **Removed `amdsmi_get_cpusocket_handles()` from the Python interface** (breaking). Use `amdsmi_get_cpu_handles()` instead.
+
+- **Removed the `plpds` key from `amdsmi_get_xgmi_plpd()` Python output** (breaking). Use the `policies` key instead.
+
+- **Removed `amdsmi_set_gpu_clk_range()`** (breaking). Use `amdsmi_set_gpu_clk_limit()` instead.
+
 - **Restructured AMD SMI C++ tests into unit and functional suites**.  
   - The `amdsmitst` source tree now separates unit tests from hardware-backed functional tests under `tests/amd_smi_test/unit/` and `tests/amd_smi_test/functional/`.
   - GTest suite names now follow a `<Component><Type>[<Operation>]` scheme: functional tests are `<Component>FunctionalReadOnly`/`<Component>FunctionalReadWrite` (e.g. `GpuFunctionalReadOnly`) and unit tests are `<Component>Unit` (e.g. `GpuUnit`). This replaces the old `amdsmitstReadOnly`/`amdsmitstReadWrite` and `AmdSmiDynamicMetricTest` names.
@@ -127,6 +178,7 @@ Full documentation for amd_smi_lib is available at [https://rocm.docs.amd.com/pr
   - The module is additionally staged under `share/amd_smi` (its historical location) for consumers that resolve it via `ROCM_PATH/share/amd_smi` rather than `site-packages` (e.g. rocprofiler-compute, omnistat) and for the TheRock artifact layout, whose manifest captures `share/amd_smi` but not the interpreter's `site-packages`. `site-packages` remains the primary importable location.
   - Added new CMake option `-DBUILD_PYTHON_WHEEL=ON` (default `OFF`) which builds the standalone Python wheel and an isolated `libamd_smi_python.so` (distinct SONAME) bundled inside it, so the wheel-shipped library can coexist in-process with the system `libamd_smi.so` without symbol collisions. With `-DBUILD_PYTHON_WHEEL=OFF` (the default used by ROCm CI) only the system-package layout is built; no wheel artifact is produced.
   - `py-interface/amdsmi_wrapper.py` now loads the shared library in this order: the `AMDSMI_LIB_OVERRIDE` env var (development / ABI-test escape hatch), a `libamd_smi_python.so` bundled next to the wrapper (pip wheel), then the system `libamd_smi.so` via the dynamic linker. A `_MissingLibrary` sentinel defers `OSError` to the first API call when no candidate is loadable, so import-time tooling (docs, lint) still works without a runtime library.
+
 - **Normalized JSON/CSV key casing in `amd-smi metric` clock and temperature sections**.  
   - The `uclk_aid`, `socclks_mid`, and temperature `xcd` keys are now lowercase (`aid_<N>`, `mid_<N>`, `xcp_<N>`) in JSON and CSV output, matching the existing `xcp_<N>` usage keys; they were previously uppercase (`AID_<N>`, `MID_<N>`, `XCP_<N>`).
   - Human-readable output is unchanged, since it uppercases all keys.

@@ -1254,7 +1254,9 @@ def format_fan_error(message, include_driver_note=False):
                     memory_partition = amdsmi_interface.AmdSmiMemoryPartitionType[
                         args.memory_partition
                     ]
-                    amdsmi_interface.amdsmi_set_gpu_memory_partition(args.gpu, memory_partition)
+                    amdsmi_interface.amdsmi_set_gpu_memory_partition_mode(
+                        args.gpu, memory_partition
+                    )
                     out = f"Successfully set memory partition to {args.memory_partition}, use `sudo modprobe -r amdgpu && sudo modprobe amdgpu` to reload driver"
                 except amdsmi_exception.AmdSmiLibraryException as e:
                     out = f"[{e.get_error_info(detailed=False)}] Unable to set memory partition to {args.memory_partition}"
@@ -1805,34 +1807,34 @@ def format_fan_error(message, include_driver_note=False):
 
         if getattr(args, "compute_partition_mem_alloc_mode", None):
             try:
-                mode = amdsmi_interface.AmdSmiComputePartitionMemAllocModeType[
+                mode = amdsmi_interface.AmdSmiAcceleratorPartitionMemAllocModeType[
                     args.compute_partition_mem_alloc_mode
                 ]
-                amdsmi_interface.amdsmi_set_gpu_compute_partition_mem_alloc_mode(args.gpu, mode)
-                out = f"Successfully set compute partition memory allocation mode to {args.compute_partition_mem_alloc_mode}"
+                amdsmi_interface.amdsmi_set_gpu_accelerator_partition_mem_alloc_mode(args.gpu, mode)
+                out = f"Successfully set accelerator partition memory allocation mode to {args.compute_partition_mem_alloc_mode}"
             except KeyError:
                 out = (
-                    "Invalid compute partition memory allocation mode "
+                    "Invalid accelerator partition memory allocation mode "
                     f"{args.compute_partition_mem_alloc_mode}"
                 )
-                self.logger.store_output(args.gpu, "compute_partition_mem_alloc_mode", out)
+                self.logger.store_output(args.gpu, "accelerator_partition_mem_alloc_mode", out)
                 self.logger.print_output()
                 self.logger.clear_multiple_devices_output()
                 return
             except amdsmi_exception.AmdSmiLibraryException as e:
-                out = f"[{e.get_error_info(detailed=False)}] Unable to set compute partition memory allocation mode to {args.compute_partition_mem_alloc_mode}"
+                out = f"[{e.get_error_info(detailed=False)}] Unable to set accelerator partition memory allocation mode to {args.compute_partition_mem_alloc_mode}"
                 if e.get_error_code() == amdsmi_interface.amdsmi_wrapper.AMDSMI_STATUS_NO_PERM:
                     out = "[AMDSMI_STATUS_NO_PERM] Command requires elevation"
-                    self.logger.store_output(args.gpu, "compute_partition_mem_alloc_mode", out)
+                    self.logger.store_output(args.gpu, "accelerator_partition_mem_alloc_mode", out)
                     self.logger.print_output()
                     self.logger.clear_multiple_devices_output()
                     raise PermissionError("Command requires elevation") from e
                 else:
-                    self.logger.store_output(args.gpu, "compute_partition_mem_alloc_mode", out)
+                    self.logger.store_output(args.gpu, "accelerator_partition_mem_alloc_mode", out)
                     self.logger.print_output()
                     self.logger.clear_multiple_devices_output()
                     return
-            self.logger.store_output(args.gpu, "compute_partition_mem_alloc_mode", out)
+            self.logger.store_output(args.gpu, "accelerator_partition_mem_alloc_mode", out)
             self.logger.print_output()
             self.logger.clear_multiple_devices_output()
             return

@@ -836,13 +836,13 @@ def static_gpu(
                     e.get_error_info(),
                 )
             try:
-                mem_alloc_mode = amdsmi_interface.amdsmi_get_gpu_compute_partition_mem_alloc_mode(
-                    args.gpu
+                mem_alloc_mode = (
+                    amdsmi_interface.amdsmi_get_gpu_accelerator_partition_mem_alloc_mode(args.gpu)
                 )
             except amdsmi_exception.AmdSmiLibraryException as e:
                 mem_alloc_mode = "N/A"
                 logging.debug(
-                    "Failed to get compute partition mem alloc mode for gpu %s | %s",
+                    "Failed to get accelerator partition mem alloc mode for gpu %s | %s",
                     gpu_id,
                     e.get_error_info(),
                 )

@@ -460,8 +460,6 @@ Changing partition settings has strict requirements:
   sudo modprobe -r amdgpu && sudo modprobe amdgpu
   ```
 
-  Alternatively, call `amdsmi_gpu_driver_reload()` from the library.
-
 - A driver reload affects **all GPUs in the hive** -- every GPU in the system is reconfigured
   to the new memory partition configuration at once.
 
@@ -475,7 +473,7 @@ set API automatically triggered an immediate driver reload on invocation; the CL
 reloaded after the user explicitly requested the partition change. The API-level reload was
 separated to give applications control over when the disruptive reload occurs. Additionally,
 as of [**ROCm 7.13.0**](https://github.com/ROCm/rocm-systems/blob/develop/projects/amdsmi/CHANGELOG.md#amd_smi_lib-for-rocm-7130), `amd-smi reset -r` is no longer available for driver reloading — use
-`sudo modprobe -r amdgpu && sudo modprobe amdgpu` or `amdsmi_gpu_driver_reload()` instead.
+`sudo modprobe -r amdgpu && sudo modprobe amdgpu` instead.
 ```
 
 ## Workload isolation and assignment
@@ -695,7 +693,8 @@ int main() {
 
     // Step 4: Reload the driver -- required to apply the memory partition change.
     // Stop all GPU workloads first. The reload may reset the accelerator partition.
-    amdsmi_gpu_driver_reload();
+    // The reload needs to occur out of band via calls to `modprobe -r amdgpu` and
+    // then `modprobe amdgpu`
 
     // Step 5: Re-initialize to pull in the updated topology (new device count/handles)
     amdsmi_shut_down();
@@ -731,7 +730,6 @@ APIs, see
   index (obtained from {c:func}`amdsmi_get_gpu_accelerator_partition_profile_config`).
 - {c:func}`amdsmi_get_gpu_memory_partition_config` -- Query the current NPS mode and supported NPS modes.
 - {c:func}`amdsmi_set_gpu_memory_partition_mode` -- Set the NPS memory partition mode.
-- {c:func}`amdsmi_gpu_driver_reload` -- Reload the amdgpu driver to apply memory partition changes.
 
 **Bare metal only:**
 - {c:func}`amdsmi_get_gpu_compute_partition` -- Query the current compute partition setting as a string.
@@ -769,7 +767,8 @@ amdsmi.amdsmi_set_gpu_memory_partition_mode(gpu, amdsmi.AmdSmiMemoryPartitionTyp
 
 # Step 4: Reload the driver -- required to apply the memory partition change.
 # Stop all GPU workloads first. The reload may reset the accelerator partition.
-amdsmi.amdsmi_gpu_driver_reload()
+# The reload needs to occur out of band via calls to `modprobe -r amdgpu` and
+# then `modprobe amdgpu`
 
 # Step 5: Re-initialize to pull in the updated topology (new device count/handles)
 amdsmi.amdsmi_shut_down()