pytorch - 💡(How to fix) Fix [ROCm] RX 9060 XT 16GB hard-capped at ~8GB usable VRAM in PyTorch — segfault on allocation beyond cap

🐛 Describe the bug

I was trying to make images with comfy on Linux Mint 22.2 with my rx 9060 xt 16 gb. When loading models, they would get stuck at 0% with the gpu at 100% usage, no matter how long I waited, still stuck at 0%.

I am no programmer so I asked claude sonnet 4.6 to help me with this and it thinks it's a bug and I asked it to write the following bug report, I was just trying to make images for fun, to see if it could work.

Bug Report Summary On an AMD Radeon RX 9060 XT 16GB (gfx1200, Navi 44), PyTorch/ROCm reports and enforces a hard VRAM limit of ~7915 MB (~8GB), despite the kernel driver and rocminfo correctly seeing the full ~16GB. Attempting to allocate beyond this limit results in a segmentation fault (not a graceful OOM error). This makes it impossible to use the full VRAM of the card for AI workloads such as ComfyUI with large models (e.g. Lumina2), forcing unnecessary CPU offloading and dramatically degrading performance.

Environment GPU: AMD Radeon RX 9060 XT 16GB (Navi 44) GPU Architecture: gfx1200 (RDNA 4) OS: Linux Mint 22.2 Kernel: 6.17.0-29-generic PyTorch: 2.13.0a0+rocm7.13.0a20260416 ROCm: 7.13.0a20260416 (nightly) Python: 3.12.3 Driver: amdgpu (kernel built-in)

Steps to Reproduce

Check what the kernel driver sees (correct):

cat /sys/class/drm/card*/device/mem_info_vram_total Output: 17095983104 (~16GB ✅)

Check what rocminfo sees (correct):

rocminfo | grep -A3 "Pool 1" Output: Size: 16211708(0xf75efc) KB (~15.8GB ✅)

Check what rocm-smi sees (correct):

rocm-smi --showmeminfo vram Output: VRAM Total Memory (B): 17095983104 (~16GB ✅)

Check what PyTorch sees (WRONG):

import torch print(torch.cuda.get_device_properties(0).total_memory // 1024**2, 'MB') Output: 7915 MB (~8GB ❌) Note also that the card is misidentified: print(torch.cuda.get_device_properties(0).name) Output: 'AMD Radeon Graphics' (should be 'AMD Radeon RX 9060 XT') ❌

Attempt to allocate beyond the ~8GB cap — segfault instead of graceful OOM:

import torch t = torch.zeros(14000, 1024, 1024, dtype=torch.float16, device='cuda') Output: Segmentation fault (core dumped) ❌

Expected Behavior torch.cuda.get_device_properties(0).total_memory should report ~16GB (~16384 MB). The card should be correctly identified as AMD Radeon RX 9060 XT. Allocations up to ~15GB should succeed or fail with a graceful torch.cuda.OutOfMemoryError, not a segfault.

Actual Behavior PyTorch hard-caps usable VRAM at ~7915 MB (~8GB), exactly half the physical VRAM. Card is misidentified as AMD Radeon Graphics (generic name). Allocating beyond the cap causes a segmentation fault instead of a proper OOM error.

Attempted Workarounds (none worked) All of the following env vars were tested individually and in combination — none changed the reported VRAM: export GPU_MAX_ALLOC_PERCENT=100 export GPU_SINGLE_ALLOC_PERCENT=100 export HSA_ENABLE_SDMA=0 export ROCR_VISIBLE_DEVICES=0 export HSA_FORCE_FINE_GRAIN_PCIE=1 export PYTORCH_HIP_ALLOC_CONF=expandable_segments:True,max_split_size_mb:512 export HIP_VISIBLE_DEVICES=0 export ROCM_VISIBLE_DEVICES=0

Root Cause Hypothesis The RX 9060 XT 16GB uses two 8GB GDDR6 modules on a 128-bit bus. It is suspected that PyTorch's HIP memory allocator or the ROCm topology reader is only enumerating one of the two memory modules, resulting in exactly half the VRAM being visible. Evidence:

The cap is exactly 50% of physical VRAM (8GB out of 16GB) rocminfo and the kernel driver both correctly see the full 16GB cat /sys/class/kfd/kfd/topology/nodes//mem_banks//size_in_bytes returns empty output, suggesting the KFD topology layer is not properly enumerating memory banks for this card

Related Issues https://github.com/deepbeepmeep/Wan2GP/issues/1341 https://github.com/ggml-org/llama.cpp/issues/21376 https://github.com/ROCm/ROCm/issues/5657

Versions

(venv) vinicius@vinicius-pc:~/ComfyUI$ curl -sL https://raw.githubusercontent.com/pytorch/pytorch/main/torch/utils/collect_env.py | python Collecting environment information... PyTorch version: 2.13.0a0+rocm7.13.0a20260416 Is debug build: False CUDA used to build PyTorch: N/A ROCM used to build PyTorch: 7.13.61040

OS: Linux Mint 22.2 (x86_64) GCC version: (Ubuntu 15.2.0-14ubuntu124ppa1) 15.2.0 Clang version: 22.1.2 (https://github.com/llvm/llvm-project.git 1ab49a973e210e97d61e5db6557180dcb92c3e98) CMake version: version 3.28.3 Libc version: glibc-2.39

Python version: 3.12.3 (main, Mar 23 2026, 19:04:32) [GCC 13.3.0] (64-bit runtime) Python platform: Linux-6.17.0-29-generic-x86_64-with-glibc2.39 Is CUDA available: True CUDA runtime version: Could not collect CUDA_MODULE_LOADING set to: GPU models and configuration: AMD Radeon Graphics (gfx1200) Nvidia driver version: Could not collect cuDNN version: Could not collect Is XPU available: False HIP runtime version: 7.13.61040 MIOpen runtime version: 3.5.1 Is XNNPACK available: True Caching allocator config: {'PYTORCH_HIP_ALLOC_CONF': 'expandable_segments:True'}

CPU: Architecture: x86_64 CPU op-mode(s): 32-bit, 64-bit Address sizes: 39 bits physical, 48 bits virtual Byte Order: Little Endian CPU(s): 8 On-line CPU(s) list: 0-7 Vendor ID: GenuineIntel Model name: 12th Gen Intel(R) Core(TM) i3-12100F CPU family: 6 Model: 151 Thread(s) per core: 2 Core(s) per socket: 4 Socket(s): 1 Stepping: 5 CPU(s) scaling MHz: 15% CPU max MHz: 5500,0000 CPU min MHz: 800,0000 BogoMIPS: 6604,80 Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid aperfmperf tsc_known_freq pni pclmulqdq dtes64 monitor ds_cpl vmx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault ssbd ibrs ibpb stibp ibrs_enhanced tpr_shadow flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid rdseed adx smap clflushopt clwb intel_pt sha_ni xsaveopt xsavec xgetbv1 xsaves split_lock_detect user_shstk avx_vnni dtherm ida arat pln pts hwp hwp_notify hwp_act_window hwp_epp hwp_pkg_req hfi vnmi umip pku ospke waitpkg gfni vaes vpclmulqdq rdpid movdiri movdir64b fsrm md_clear serialize arch_lbr ibt flush_l1d arch_capabilities Virtualization: VT-x L1d cache: 192 KiB (4 instances) L1i cache: 128 KiB (4 instances) L2 cache: 5 MiB (4 instances) L3 cache: 12 MiB (1 instance) NUMA node(s): 1 NUMA node0 CPU(s): 0-7 Vulnerability Gather data sampling: Not affected Vulnerability Ghostwrite: Not affected Vulnerability Indirect target selection: Not affected Vulnerability Itlb multihit: Not affected Vulnerability L1tf: Not affected Vulnerability Mds: Not affected Vulnerability Meltdown: Not affected Vulnerability Mmio stale data: Not affected Vulnerability Old microcode: Not affected Vulnerability Reg file data sampling: Not affected Vulnerability Retbleed: Not affected Vulnerability Spec rstack overflow: Not affected Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization Vulnerability Spectre v2: Mitigation; Enhanced / Automatic IBRS; IBPB conditional; PBRSB-eIBRS SW sequence; BHI BHI_DIS_S Vulnerability Srbds: Not affected Vulnerability Tsa: Not affected Vulnerability Tsx async abort: Not affected Vulnerability Vmscape: Mitigation; IBPB before exit to userspace

Versions of relevant libraries: [pip3] numpy==2.4.3 [pip3] torch==2.13.0a0+rocm7.13.0a20260416 [pip3] torchaudio==2.11.0+rocm7.13.0a20260425 [pip3] torchsde==0.2.6 [pip3] torchvision==0.27.0a0+rocm7.13.0a20260416 [pip3] triton==3.7.0+git9e2c158e.rocm7.13.0a20260416 [pip3] triton-rocm==3.7.0 [conda] Could not collect (venv) vinicius@vinicius-pc:~/ComfyUI$

cc @ptrblck @msaroufim @eqy @jerryzh168 @tinglvv @nWEIdia @jeffdaily @sunway513 @jithunnair-amd @pruthvistony @ROCmSupport @jataylo @hongxiayang @naromero77amd @pragupta @jerrymannil @xinyazhang