Windows ROCm 10.1 nightly runtime: hipHostRegister of pageable memory ≥1 MiB returns success but leaves corrupted state → background-thread segfault (gfx1100; 7.14 runtime fails cleanly)

Summary

On Windows with the ROCm 10.1 nightly runtime (pip rocm==10.1.0a20260822 from rocm.nightlies.amd.com/whl-multi-arch/), registering a ≥1 MiB pageable host buffer with hipHostRegister returns success but leaves broken state. Shortly afterwards a runtime background thread segfaults (0xC0000005) inside amdhip64_7.dll, killing the process with no catchable error.

The ROCm 7.14 runtime, on the same machine and same driver, handles the identical call by returning a clean hipErrorInvalidArgument (surfacing as a catchable exception) — no crash. Ordinary hipMemcpy from pageable memory works on both runtimes at every size we tested (≤14 MiB, sync/async, streams, double buffering). The defect is specific to the register path of the 10.1 nightly build.

Environment

• Windows 11 (build 26200), Radeon RX 7900 XTX (gfx1100), Adrenalin 26.8.1
• Runtime under test: pip rocm==10.1.0a20260822 (_rocm_sdk_core/bin/amdhip64_7.dll)
• Comparison runtime: ROCm 7.14 bundled with the lc700x/MIGraphX-Windows v7.14.0a20260615.post.2.2 wheel (same DLL name, older build)
• Workload triggering the registration: MIGraphX run() with a numpy-backed argument (its input path registers host memory for zero-copy)

Reproduction (minimal)

A single-op ONNX (ReLU) whose input buffer crosses the threshold. With the 10.1 nightly runtime:

With the 7.14 runtime: ≤768 KB RUN_OK; ≥1 MiB raises a clean RuntimeError from hip.cpp:128 register_on_gpu (hipErrorInvalidArgument), process survives.

Repro sketch (Python, migraphx wheel):

import migraphx, numpy as np
m = migraphx.parse_onnx("relu_2mb.onnx")          # single ReLU, input float32[524288]
m.compile(migraphx.get_target("gpu"))
shapes = m.get_parameter_shapes()
args = {k: migraphx.argument(np.random.rand(*shapes[k].lens()).astype(np.float32)) for k in shapes}
m.run(args)   # 10.1 nightly: SIGSEGV (0xC0000005) in amdhip64_7.dll background thread

Raw hipMemcpy micro-benchmarks (ctypes, pageable host memory, all sizes ≤14 MiB, sync/async/stream/double-buffered) pass on both runtimes, so plain copies are unaffected.

Workaround

Allocating the buffer with hipHostMalloc instead of registering existing pageable memory works on both runtimes — a 13.5 MB input transfers and runs fine, and with this shim our full 139 MB detection model compiles and runs end-to-end on the 7.14 runtime. But the fake-success-then-crash behavior of the 10.1 nightly register path looks like a genuine runtime (or runtime↔KMD) defect worth fixing: a failure here should be reported honestly, not deferred to a background-thread crash.

Cross-reference

microsoft/WindowsML#15 reports the same GPU (RX 7900 XTX) and the same crash signature — segfault in the bundled amdhip64_7.dll while compiling a large model — on AMD's official WinML MIGraphX EP. Possibly the same underlying Windows HIP runtime weakness.

Happy to provide the exact ONNX repro files, the threshold logs, and the full crash stacks.