PreparedRunnerOptions Struct
Experimental prepared-route runner options. More...
Declaration
Included Headers
Public Member Attributes Index
| std::string | mode |
| int | ring_depth = 0 |
| bool | profile = false |
| std::string | dequant_flags |
|
Optional prepared-runner graph223/dequant optimization flags. More... | |
Description
Experimental prepared-route runner options.
Default mode is empty/"passthrough" and leaves the normal per-plugin pipeline untouched. mode="dequant" replaces eligible model-managed graph223 postprocess CVU stages with neatpreparedrunner mode=dequant; mode="route" is reserved for the full quant->MLA->dequant fused runner once all executor bodies are enabled.
Definition at line 257 of file SessionOptions.h.
Public Member Attributes
dequant_flags
|
Optional prepared-runner graph223/dequant optimization flags.
These are graph dequantize metadata tokens such as "fused,half,zpfold,bitmagic". Empty keeps model/runtime defaults. This is the model/session option equivalent of the old diagnostic env controls and is only consumed by the opt-in prepared runner.
Definition at line 269 of file SessionOptions.h.
mode
|
Definition at line 258 of file SessionOptions.h.
profile
|
Definition at line 260 of file SessionOptions.h.
ring_depth
|
Definition at line 259 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.