Fix AMD APU unified memory VRAM detection using VIS_VRAM

On AMD APUs with unified memory architecture (e.g., AMD Ryzen AI MAX /
Strix Halo, GC 12.0.0), both amdsmi_get_gpu_vram_usage and
rsmi_dev_memory_total_get(RSMI_MEM_TYPE_VRAM) return only the BIOS
carveout (~512 MiB) instead of the actual usable unified memory.

VIS_VRAM (RSMI_MEM_TYPE_VIS_VRAM=2) correctly reports the full system
memory on APUs because all system memory is CPU-visible. Add the
constant to pyrocmsmi and use max(VRAM, VIS_VRAM) in both detection
paths so APUs report the correct memory capacity.

Co-authored-by: Readon <3614708+Readon@users.noreply.github.com>

Author: Copilot

gpustack_runtime/detector/amd.py

@@ -179,6 +179,16 @@ def detect(self) -> Devices | None:
                     dev_gpu_vram_usage = pyamdsmi.amdsmi_get_gpu_vram_usage(dev)
                     dev_mem = dev_gpu_vram_usage.get("vram_total")
                     dev_mem_used = dev_gpu_vram_usage.get("vram_used")
+                    # On APUs with unified memory (e.g., AMD Strix Halo), VRAM
+                    # reports only the BIOS carveout (~512 MiB); VIS_VRAM reports
+                    # the full usable system memory. Use VIS_VRAM when larger.
+                    with contextlib.suppress(pyrocmsmi.ROCMSMIError):
+                        dev_mem_vis_vram = byte_to_mebibyte(
+                            pyrocmsmi.rsmi_dev_memory_total_get(
+                                dev_idx, pyrocmsmi.RSMI_MEM_TYPE_VIS_VRAM,
+                            ),
+                        )
+                        dev_mem = max(dev_mem, dev_mem_vis_vram)
                     dev_ecc_count = pyamdsmi.amdsmi_get_gpu_ecc_count(
                         dev,
                         pyamdsmi.AmdSmiGpuBlock.UMC,
@@ -189,6 +199,16 @@ def detect(self) -> Devices | None:
                     dev_mem = byte_to_mebibyte(  # byte to MiB
                         pyrocmsmi.rsmi_dev_memory_total_get(dev_idx),
                     )
+                    # On APUs with unified memory (e.g., AMD Strix Halo), VRAM
+                    # reports only the BIOS carveout (~512 MiB); VIS_VRAM reports
+                    # the full usable system memory. Use VIS_VRAM when larger.
+                    with contextlib.suppress(pyrocmsmi.ROCMSMIError):
+                        dev_mem_vis_vram = byte_to_mebibyte(
+                            pyrocmsmi.rsmi_dev_memory_total_get(
+                                dev_idx, pyrocmsmi.RSMI_MEM_TYPE_VIS_VRAM,
+                            ),
+                        )
+                        dev_mem = max(dev_mem, dev_mem_vis_vram)
                     dev_mem_used = byte_to_mebibyte(  # byte to MiB
                         pyrocmsmi.rsmi_dev_memory_usage_get(dev_idx),
                     )

gpustack_runtime/detector/pyrocmsmi/__init__.py

@@ -62,6 +62,11 @@
 ROCMSMI_IOLINK_TYPE_XGMI = 2
 ROCMSMI_IOLINK_TYPE_NUMIOLINKTYPES = 3
 
+## Memory Types ##
+RSMI_MEM_TYPE_VRAM = 0
+RSMI_MEM_TYPE_GTT = 1
+RSMI_MEM_TYPE_VIS_VRAM = 2
+
 ## Error Codes ##
 ROCMSMI_ERROR_UNINITIALIZED = -99997
 ROCMSMI_ERROR_FUNCTION_NOT_FOUND = -99998