Struct FTryllContextOverrides¶
ClassList > FTryllContextOverrides
#include <TryllContextOverrides.h>
Public Attributes¶
| Type | Name |
|---|---|
| ETryllKvCacheType | KvCacheType = ETryllKvCacheType::Inherit |
| int32 | NBatch = 0 |
| int32 | NOutputsMax = 0 |
| int32 | NUbatch = 0 |
| bool | OffloadKqv = 0 |
| bool | bOverrideNBatch = false |
| bool | bOverrideNOutputsMax = false |
| bool | bOverrideNUbatch = false |
| bool | bOverrideOffloadKqv = false |
Detailed Description¶
Backend context construction. Null fields inherit ModelVariant then server config, except n_outputs_max which CreateContext defaults to 1. Structural at the field: these are baked into llama_context_params at context creation, so a change means a new context and a lost KV cache.
Public Attributes Documentation¶
variable KvCacheType¶
Number format of the context memory (KV cache). Smaller types save memory at some quality cost. Inherit = the model's setting, else q8_0.
variable NBatch¶
variable NOutputsMax¶
variable NUbatch¶
variable OffloadKqv¶
variable bOverrideNBatch¶
Most tokens sent to the model in one call while reading the prompt. Unset = the model's setting.
variable bOverrideNOutputsMax¶
Leave unset. Advanced: how many positions return scores per step.
variable bOverrideNUbatch¶
Tokens processed per GPU step; sizes the working buffers. Unset = the model's setting.
variable bOverrideOffloadKqv¶
Keep the context memory on the GPU. Can only turn it off, not on for a model placed in RAM. Unset = server and model placement decide.
The documentation for this class was generated from the following file C:/_tryll/_monorepo3/tryll/clients/unreal/Source/TryllClient/Public/Generated/Nodes/TryllContextOverrides.h