[GPU] Improve max-allocation error guidance and enable_large_allocations docs#34894
[GPU] Improve max-allocation error guidance and enable_large_allocations docs#34894Sahilbhatane wants to merge 7 commits intoopenvinotoolkit:masterfrom
Conversation
github-actions
bot
added
category: inference
There was a problem hiding this comment.
Pull request overview
Improves Intel GPU plugin user-facing guidance around failures caused by a single allocation exceeding the device’s maximum allocatable buffer size, and clarifies the behavior of the enable_large_allocations option/property in internal and public docs.
Changes:
- Update the max-allocation assertion message to fix formatting and point users to
GPU_ENABLE_LARGE_ALLOCATIONS. - Align the internal option description for
ov::intel_gpu::hint::enable_large_allocationswith the real check (device max allocation size rather than “>4GB”). - Clarify the public property documentation to mention bypassing the max-allocation-size check and its usefulness when exceeding
CL_DEVICE_MAX_MEM_ALLOC_SIZE.
Reviewed changes
Copilot reviewed 3 out of 3 changed files in this pull request and generated 1 comment.
| File | Description |
|---|---|
src/plugins/intel_gpu/src/runtime/engine.cpp |
Improves the assertion text for “single allocation exceeds device max alloc size” and adds a workaround hint. |
src/plugins/intel_gpu/include/intel_gpu/runtime/options.inl |
Updates the option help text to describe the behavior more accurately than the previous “>4GB” wording. |
src/inference/include/openvino/runtime/intel_gpu/properties.hpp |
Extends the public property doxygen to document bypassing the max-alloc-size check and relates it to CL_DEVICE_MAX_MEM_ALLOC_SIZE. |
|
build_jenkins |
|
@p-durandin Re-requesting build. Updated the branch, all checks passed on previous build. |
|
build_jenkins |
|
@p-durandin I'm unable to view failing checks. Does this concerns to my changes? My changes are only error log and doc related changes. |
|
This PR is still worth merging as-is. It’s low risk and improves the user-facing assertion message while also clarifying how Importantly, it doesn’t touch the actual allocation behavior or change when the hint is applied during compilation. The underlying issue- large constants getting allocated in In the meantime, merging this gives users better context when they hit the error and makes the current workaround easier to understand. Once the follow-up fix lands, that same workaround will start behaving correctly for early GPU allocations as well. |
Details:
engine::check_allocatable()fails with an assertion. The message had a small formatting issue (missing space before "Please") and didn’t point users to theGPU_ENABLE_LARGE_ALLOCATIONS/ov::intel_gpu::hint::enable_large_allocationsworkaround.src/plugins/intel_gpu/src/runtime/engine.cppto fix the spacing and include a clear suggestion to enableGPU_ENABLE_LARGE_ALLOCATIONSwhen applicable.src/plugins/intel_gpu/include/intel_gpu/runtime/options.inlwith the actual behavior. The check is against the device max allocation size, not just “>4GB”.src/inference/include/openvino/runtime/intel_gpu/properties.hppto clarify that enabling large allocations also bypasses the max-allocation-size check, which helps when a single buffer exceedsCL_DEVICE_MAX_MEM_ALLOC_SIZE.Tickets:
AI Assistance: