[CUDA] sycl::aligned_alloc_host limited to 2**31 byte ?

[tdavidcl]

Describe the bug

Hi,

Today, I was trying to run fairly large simulations and got into the following error (internal reporting from the code)

<CUDA>[ERROR]: 
UR CUDA ERROR:
	Value:           1
	Name:            CUDA_ERROR_INVALID_VALUE
	Description:     invalid argument
	Function:        USMHostAllocImpl
	Source Location: /tmp/tmp.nlKu2FwFq5/intel-llvm-mirror/build/_deps/unified-runtime-src/source/adapters/cuda/usm.cpp:179

USM allocation failed, details : sz=2250476000, target=host, alignment=8, alloc result = 0x0
    World infos :
        World size = 8
        World rank = 2
    Device infos :
        Device name = NVIDIA H100
    Allocs :
        max_allocated_byte_host = 4.09 GB
        max_allocated_byte_device = 32.00 GB
        max_allocated_byte_shared = 0.00 B
        allocated_byte_host = 4.09 GB
        allocated_byte_device = 25.90 GB
        allocated_byte_shared = 0.00 B

I tried several parameters and all that lead to a crash (due to the returned ptr being 0x0) where with size above 2**31 bytes when using sycl::aligned_alloc_host, with CUDA backend. Also I did not find mention of such a limit when searching in Nvidia forums. Is that a known limit for intel/llvm / icpx or a legit issue ?

To reproduce

sycl::aligned_alloc_host(8, 2250476000, sycl_ctx);

it will return 0x0

Environment

uname -a
Linux kraken-gpu 6.11.7-amd64 #1 SMP PREEMPT_DYNAMIC Debian 6.11.7-1 (2024-11-09) x86_64 GNU/Linux

icpx --version
Intel(R) oneAPI DPC++/C++ Compiler 2025.0.4 (2025.0.4.20241205)
Target: x86_64-unknown-linux-gnu
Thread model: posix
InstalledDir: /softs/intel/2025/oneapi/compiler/2025.0/bin/compiler
Configuration file: /softs/intel/2025/oneapi/compiler/2025.0/bin/compiler/../icpx.cfg

cuda_12.9

Additional context

No response

[kekaczma]

Thank you for reporting this issue. I attempted to reproduce the problem but was unable to observe the failure in my test environment.

Test Environment

My Configuration:

• Intel LLVM: Current development build (January 2026)
• CUDA: 12.6
• GPU: NVIDIA H100 PCIe
• Test: sycl::aligned_alloc_host() with sizes from 2GB to 6GB, including your exact size (2,250,476,000 bytes)
• Result: All allocations succeeded

Your Configuration:

• Intel LLVM: 2025.0.4 (build date: December 5, 2024)
• CUDA: 12.9
• Allocation size: 2,250,476,000 bytes
• Result: Returns nullptr with CUDA_ERROR_INVALID_VALUE

Potentially Relevant Changes

Between your version (December 2024) and current builds, there have been updates to the Unified Memory Framework (UMF) CUDA provider, including:

• February 10, 2025 - Commit d36d5857: Migration from cuMemAllocHost() to cuMemHostAlloc() in the CUDA memory provider
• Additional context handling improvements in the same timeframe

These changes may address the issue you encountered.

Next Steps

Could you please test with a newer Intel LLVM version (released after February 2025) and let me know if the issue persists? This will help determine whether:

1. The issue is resolved in current builds, or
2. Further investigation is needed for your specific environment (CUDA 12.9 + specific driver version)

If the problem continues with a newer compiler version, please provide:

• Exact Intel LLVM version and build date
• CUDA driver version (nvidia-smi output)
• Any relevant error messages from the SYCL runtime