ProcessMlaOptions Struct
process-MLA execution options. More...
Declaration
Included Headers
Public Member Attributes Index
| bool | async = true |
|
Enable the safe async processmla submit/emit path. More... | |
| int | output_pool_buffers = 0 |
|
Optional processmla output pool size override. More... | |
| bool | defer_output_invalidate = true |
|
For prepared MLASHM outputs, skip the immediate producer-side CPU invalidate and stamp the output metadata as device-produced/cpu-dirty. More... | |
Description
process-MLA execution options.
Definition at line 227 of file SessionOptions.h.
Public Member Attributes
async
|
Enable the safe async processmla submit/emit path.
Default true uses the optimized prepared async lane when the plugin/stage is eligible; set false, or use the plugin/env kill switches, to force the synchronous fallback.
Definition at line 231 of file SessionOptions.h.
defer_output_invalidate
|
For prepared MLASHM outputs, skip the immediate producer-side CPU invalidate and stamp the output metadata as device-produced/cpu-dirty.
The framework runtime default is enabled so MLA->CVU/postprocess routes pay the invalidate only at the actual CPU consumer boundary. Manual low-level pipelines that expose raw MLA outputs to legacy CPU readers can still override the element property to false.
Definition at line 245 of file SessionOptions.h.
output_pool_buffers
|
Optional processmla output pool size override.
A value <= 0 leaves the runtime default in place. The framework runtime default is 4, matching model-managed CVU/MLA buffering and avoiding artificial backpressure while downstream stages still hold previous tensor-set outputs.
Definition at line 237 of file SessionOptions.h.
The documentation for this struct was generated from the following file:
Generated via doxygen2docusaurus 2.0.0 by Doxygen 1.9.8.