Crate atomr_accel_agents 

Agentic / LLM GPU actor blueprints on atomr-accel-cuda.

use atomr_accel_agents::prelude::*;

• shared_state::SharedGpuStateCoordinator — write-token coordination over a ManagedRef<f32> world buffer.
• embedding_cache::EmbeddingCache — LRU cache of (input hash) -> Vec<f32>.
• rag::RagPipeline — retrieval + generate harness.
• vector_index::CpuVectorIndex — top-k cosine similarity over a flat host index.
• langgraph_nodes::LangGraphGpuActor — DAG executor with cycle detection.

Modules

embedding_cache

EmbeddingCache — LRU cache of (input hash) -> Vec<f32>.

langgraph_nodes

LangGraphGpuNodes — composable agent graph nodes.

prelude

Canonical re-exports. use atomr_accel_agents::prelude::*;.

rag

RagPipeline — query → embedding → vector search → context assembly → LLM call.

shared_state

SharedGpuStateCoordinator — issues write tokens to N agents sharing a ManagedRef<f32> world-state buffer.

vector_index

GpuVectorIndex — similarity-search index abstraction.