| Ember |
OtherU's inference engine for DeepSeek-V4-Flash on AMD Strix Halo. See Ember. |
| Hermes Agent |
Software that connects a model to conversations, tools, scheduling, and memory integrations. |
| Hindsight |
A persistent memory service used by Hermes. See Hindsight Memory. |
| Memory bank |
A named grouping of memories in Hindsight. Grouping alone does not establish access control. |
| KV cache |
Model attention state reused to avoid processing an eligible prompt prefix again. It is distinct from stored conversation history. |
| Prefill |
Processing input tokens before generating a response. |
| Decode |
Generating response tokens after prompt processing. |
| Time to first token |
Elapsed time before the first output token, including relevant queueing and prompt-processing work. |
| DSpark |
A draft model used in Ember's speculative-decoding path. |
| Speculative decoding |
A smaller model proposes tokens that the target model verifies, potentially reducing the cost of generation. |
| Acceptance rate |
The fraction of draft tokens accepted during verification; it varies by model pair and workload. |
| Quantization |
Representing model weights at reduced precision to reduce storage and memory use. |
| bpw |
Bits per weight, a measure of effective model-weight storage cost. |
| imatrix |
An importance matrix collected from representative inputs to guide quantization. |
| Affine fp2 |
OtherU's custom Q2_0_ROCMFP2 format. See Quantization Pipeline. |
| Abliteration |
A model-weight edit intended to reduce refusal behavior. Its effect on other capabilities must be evaluated. |
| MoE |
Mixture of Experts: a model architecture that activates selected expert networks for each token. |
| GGUF |
A file format for model tensors and metadata. Runtime support depends on the architecture and tensor encodings inside it. |
| Strix Halo |
The AMD Ryzen AI Max platform used by the reference inference host. |
| gfx1151 |
The GPU architecture identifier targeted by Ember's Strix Halo backend. |
| Unified memory |
Physical memory shared by CPU and GPU, subject to platform limits and competing allocations. |
| Hardware KVM |
A device that captures display output and delivers keyboard and mouse input to another computer. |
| Loopback |
A network interface reachable from the same network namespace, commonly addressed as 127.0.0.1. |
| Air gap |
Network separation from external systems. A local model does not by itself create an air-gapped deployment. |