docs: Add Kvbm documentation for connector API integrations [DYN-1052] (#3578)

Signed-off-by: Suman Tatiraju <statiraju@nvidia.com>
Signed-off-by: akshatha-k <33278067+akshatha-k@users.noreply.github.com>
Co-authored-by: akshatha-k <33278067+akshatha-k@users.noreply.github.com>
Co-authored-by: coderabbitai[bot] <136622811+coderabbitai[bot]@users.noreply.github.com>
Co-authored-by: Ziqi Fan <ziqif@nvidia.com>

docs/architecture/architecture.md

@@ -92,11 +92,10 @@ Existing routing methods, including load-based routing, overlook the specific pr
 
 ### KV cache manager
 
-Dynamo's design enables KV cache offloading to system CPU memory. In accelerated servers, the CPU (system) memory is often larger than the GPU memory and fast enough to store and serve KV cache data. The following plot highlights the performance gains achieved through system memory offloading, even with prefix caching enabled via inference engine. In a scenario involving 10 multi-turn conversations with 80 users, system memory offloading resulted in a 40% improvement in TTFT, demonstrating benefits beyond basic prefix caching.
+The Dynamo KV Block Manager (KVBM) enables KV cache offloading to system CPU memory, local SSDs, and network-attached storage, allowing more KV blocks to be reused instead of recomputed. In many cases, KV transfer is faster than recomputation, so KVBM helps reduce time-to-first-token (TTFT). The following plot highlights the performance gains achieved through CPU memory offloading. In a scenario involving 20 multi-turn conversations with 15 users, KVBM with CPU memory offloading achieved a 2.2×–12× improvement in TTFT (depending on QPS), demonstrating benefits that extend beyond basic prefix caching.
+![Line graph comparing Pure GPU prefix caching with vLLM and KVBM host offloading for TTFT (Time To First Token)](../images/kvbm_agg_performance.png)
 
-![Line graph comparing Pure GPU prefix caching and Dynamo KV manager host offloading for TTFT (Time To First Token) across rounds with 80 users](../images/kv_manager.png)
-
-* Tested with 100K requests to R1 using R1 Distilled Llama 70B FP8 on 2 nodes of H100s. Avg 4K ISL / 800 OSL
+* Tested with different QPS using Qwen3-8B on H100. Avg 20K ISL / 100 OSL.
 
 ### NVIDIA Inference Transfer Library (NIXL)
 

docs/backends/trtllm/README.md

@@ -57,7 +57,7 @@ git checkout $(git describe --tags $(git rev-list --tags --max-count=1))
 | [**KV-Aware Routing**](../../../docs/architecture/kv_cache_routing.md) | ✅ |  |
 | [**SLA-Based Planner**](../../../docs/planner/sla_planner.md) | ✅ |  |
 | [**Load Based Planner**](../../../docs/planner/load_planner.md) | 🚧 | Planned |
-| [**KVBM**](../../../docs/kvbm/kvbm_architecture.md) | 🚧 | Planned |
+| [**KVBM**](../../../docs/kvbm/kvbm_architecture.md) | ✅ | |
 
 ### Large Scale P/D and WideEP Features
 

docs/kvbm/kvbm_architecture.md

@@ -20,7 +20,7 @@ limitations under the License.
 
 The KVBM serves as a critical infrastructure component for scaling LLM inference workloads efficiently. By cleanly separating runtime logic from memory management, and by enabling distributed block sharing, KVBM lays the foundation for high-throughput, multi-node, and memory-disaggregated AI systems.
 
-![A block diagram showing a layered architecture view of Dynamo KV Block manager.](../images/kvbm-arch.png)
+![A block diagram showing a layered architecture view of Dynamo KV Block manager.](../images/kvbm-architecture.png)
 **High level layered architecture view of Dynamo KV Block manager and how it interfaces with different components of LLM inference ecosystem**
 
 The KVBM has three primary logical layers. The top layer-the LLM inference runtimes (TRTLLM, vLLM and SGLang)-integrates through a dedicated connector module to the Dynamo KVBM module. These connectors act as translation layers, mapping runtime-specific operations and events into the KVBM’s block-oriented memory interface. This decouples memory management from the inference runtime, enabling backend portability and providing memory tiering.

docs/kvbm/kvbm_components.md

@@ -18,244 +18,53 @@ limitations under the License.
 
 # Understanding KVBM components
 
-The design of the KVBM is inspired from vLLM and SGLang KV block managers but with a twist from historical memory tiering design aspired in general GPU programming. [See KVBM Reading](kvbm_reading.md). The following figure shows the internal architecture of KVBM and how it works across workers using NIXL.
-
-![Internal architecture and key modules in the Dynamo KVBM. ](../images/kvbm-internal-arch.png)
-**Internal architecture and key modules in the Dynamo KVBM**
-
-## KvBlockManager as Orchestration Layer
-
-The `KvBlockManager <H, D>` acts as a coordinator across memory tiers—host (CPU), device (GPU), and remote—by managing per-backend block pools and exposing consistent block lifecycle APIs. It tracks KV block locations across device memory (G1), CPU memory within and across nodes (G2), local/pooled SSDs (G3), and remote storage (G4). G1-G4 are key tiers enabled by KVBM. Critical to note that KVBM treats G4 storage as an opaque blob store, unaware of internal layout optimizations.
-
-`KvBlockManager<H, D>` owns:
-
-* A device-side `BlockPool<Device>`
-* A host-side `BlockPool<Host>`
-* A remote NIXL agent that supports communication and memory sharing across nodes
-* A block set registry for remote lookup and import/export of block metadata
-
-Implementation-wise, `KvBlockManagerState` holds the logic: it's initialized by `KvBlockManagerConfig`, which merges runtime, model, and layout configurations. `NixlOptions` injects remote awareness.
-
-## Block Layout and Memory Mapping
-
-Each block is a 2D array `[num_layers][page_size × inner_dim]`. `BlockLayouttrait` abstracts the memory layout. The default implementation,`FullyContiguous`, stores all layers for all blocks in one region with alignment-aware stride computation:
-
-
-```none
-block_stride_in_bytes = align_up(num_layers × layer_stride, alignment);
-```
-
-
-Both CPU and GPU pools share this memory layout, but they use storage-specific backends:
-
-* `DeviceStorage` → CUDA device buffer
-* `PinnedStorage` → page-locked host memory
-* `SystemStorage` → CPU heap memory (fallback/test)
-* `NixlStorage` → remote memory through NIXL RDMA handles (includes storage)
-
-Each layout is constructed using a `LayoutConfig`, and storage is either passed directly or allocated using a StorageAllocator.
-
-## BlockPool and Memory Pools (Active and Inactive)
-
-Each `BlockPool<T>` (where `T` is `DeviceStorage`, `PinnedStorage`, and so forth) tracks two sub-pools:
-
-* `ActivePool`: Contains blocks currently in use by sequences
-* `InactivePool`: Recycled blocks ready for allocation; think free list
-
-When a token block is requested (for example, `get_mutable_block()`), the allocator pops from `InactivePool`, transitions its state, and returns a writable handle. On sequence commit or eviction, the system resets blocks and returns them to the inactive pool.
-
-The state machine (`BlockState`) that tracks the block lifecycle transitions includes:
-
-| State | Description | Ownership | Valid Actions/Transitions |
-| ----- | ----- | ----- | ----- |
-| Reset | Block hasn't been initialized or was reset. No associated sequence. | Held in InactivePool, reusable | init_sequence(salt_hash) → Partial |
-| Partial | Block is being filled with tokens for a new sequence. In-progress. | Owned by the sequence creator | add_token() / add_tokens() (accumulate)- commit() → Complete- reset() → Reset |
-| Complete | Block is fully filled with token data but not yet visible to others. | Still owned by creator thread | register() → Registered- reset() → Reset |
-| Registered | Block is finalized and visible for reuse. Available in the deduplication cache. Can use block for lookups | Shared ownership (global registry) | Auto drop() → triggers Remove event and transitions to Reset |
-
-This table lists the valid KVBM transitions:
-
-| From → To | Trigger | Validation |
-| ----- | ----- | ----- |
-| Reset → Partial | initsequence(salt_hash) | Must not be in use |
-| Partial → Complete | commit() | Must be full |
-| Complete → Registered | register() | Must be finalized |
-| Registered → Reset | Drop of RegistrationHandle | Automatic |
-| Partial → Reset | Aborted sequence | Explicit or drop |
-| Complete → Reset | Invalidated | Explicit or drop |
-
-Consider this example lifecycle of a block in the KVBM; in it, a sequence requests a new KV block:
-
-1. Allocator pops from InactivePool → Block is in Reset
-2. `init_sequence()` → Transitions to Partial
-3. Tokens are appended → State remains Partial
-4. On full → `commit()` → State becomes Complete
-5. `register()` → Block is hashed and moved to Registered. Blocks can now be used to lookup.
-6. On eviction or end-of-life → `drop()` of RAII handle returns block to Reset
-
-## Lifecycle Management using RAII and Event Plane
-
-The system uses RAII for memory lifecycle management. Every block holds metadata and registration state, and registration is coupled with an `EventManager`. On registration and drop:
-
-* `PublishHandle` triggers Register events
-* Dropping it triggers Remove events
-
-This pattern ensures consistency for shared memory tracking across workers without requiring explicit deallocation logic. The events are propagated in the Dynamo Events plane. Any Dynamo component subscribed to the events plane can listen to these changes. Note that even the storage provider can subscribe to the events plane and create an internal prefix tree representation that is tailored and optimized for the specific platform.
-
-## Remote Memory Integration using NIXL
-
-The NIXL agent exposes remote memory buffers using `NixlBlockSet`, `RemoteBlocks`, and layout descriptors. Key operations include:
-
-* `nixl_register()`: Registers memory region with NIXL runtime
-* `serialize() / deserialize()`: Converts layout and memory into transferable descriptors
-* `import_remote_blockset()`: Loads remote node's block layouts into the manager
-* `get_remote_blocks_mutable()`: Fetches transferable memory views from another node
-
-`RemoteBlocks` is a lightweight abstraction over shared memory for cross-node block usage (through UCX or other backends).
-
-The left side of the figure in [Understanding KVBM Components](#understanding-kvbm-components) illustrates a bidirectional remote memory registration and layout synchronization protocol between workers (for example, Worker 1 and Worker 2) using NIXL. The following steps break down the process:
-
-1. *Agent Creation & Memory Registration:*
-
-   Each worker independently sets up a NixlAgent:
-    * Registers its memory regions (that is, device memory) through `nixl_register()`.
-    * These regions correspond to blocks managed in the local BlockPool.
-      Once the worker registers the memory, NIXL creates remote-accessible descriptors, which it binds to the memory layout.
-
-2. *Metadata exchange:*
-
-   After memory registration, workers exchange serialized layout metadata, encapsulated in a `SerializedNixlBlockLayout`.
-
-    Why is this step critical?
-   * LLM inference workloads often differ in *tensor parallel (TP)* configurations:
-     * Worker 1 might have TP=4, while Worker 2 has TP=8.
-     * Thus, even if both systems use similar `FullyContiguous` layouts, their internal slicing and alignment assumptions differ.
-   * The metadata exchange bridges this semantic mismatch by sharing:
-     * LayoutConfig (num_layers, page_size, inner_dim, dtype)
-     * BlockSetID
-     * Base address + stride information (including alignment)
-     * Device ID + memory type (host/device)
-   * Once the workers share metadata, each can reconstruct the layout on its side using deserialize().
-  This enables NIXL to:
-      * Understand where each layer/block resides
-      * Perform correct gather-scatter operations during RDMA-like transfers
-
-    Without this step, remote fetches would result in data corruption or misaligned tokens.
-
-3. *Serialization & Deserialization: Making Layouts Portable*
-
-   In the serialization stage, KVBM exports and `FullyContiguous::serialize()` encodes:
-   * FullyContiguousConfig
-   * base_offset
-   * Physical memory descriptors (NixlStorage), including:
-     * Memory type (VRAM, DRAM)
-     * Address & size
-     * Device ID
-
-       The system sends this using NIXL transfer and then injects it into a KVBM scheduler state. In the deserialization stage, `SerializedNixlBlockLayout::deserialize()` rehydrates this into:
-       * A fully reconstructed memory layout view
-       * Local representation of a remote memory slice with correct offsets and size semantics
-       It also enables direct access to remote memory with consistent logical semantics
-       This guarantees that even across different system configurations (hardware or LLM shape), both parties agree on the memory view for each KV block.
-
-4. *Ownership handles and lifetime tracking*
-
-    Memory ownership in NIXL is tightly coupled with RAII-based handles:
-     * When a block is registered, it returns a `PublishHandle` which wraps a `RegistrationHandle`
-     * On drop of this handle, an automatic Remove event is published, which:
-       * Deregisters the block from the NIXL layer
-       * Removes it from the remote block registry
-     * This ensures that, once the block is evicted from the cache or no longer used in inference, all references are invalidated cleanly across nodes
-       This mechanism avoids:
-     * Stale memory access
-     * Dangling pointers on GPU or host
-     * Manual deregistration bugs
-       The system can batch and publish registration events using a Publisher, optimizing performance under high concurrency
-
-
-## Storage backends and pluggability
-
-You can integrate KVBM with a storage backend by extending or wrapping `NixlEnabledStorage` to support cross-node RDMA registration. All layouts and block pools are generic over these backends, allowing for fine-grained control over memory tiers. We defer detailed integration guidance, since we collaborate with storage partners to simplify and standardize these integration paths.
-
-```mermaid
----
-title: Example KVBM System Architecture
----
-flowchart TD
-    A["Distributed Inference Engine"] --> B["Dynamo KV Block Manager"]
-
-    B --> C["NIXL Storage Agent<br/>- Volume registration<br/>- get()/put() abstraction"]
-    B --> D["Event Plane<br/>- NATS-based Pub/Sub<br/>- StoreEvent / RemoveEvent"]
-
-    C --> E["G4 Storage Infrastructure<br/>(SSD, Object store, etc.)<br/>- Store KV blocks"]
-    D --> F["Storage Provider Subscriber<br/>- Parse Events<br/>- Build fast tree/index<br/>- Optimize G4 tiering"]
-```
-
-For now, the following breakdown provides a high-level understanding of how KVBM interacts with external storage using the NIXL storage interface and the Dynamo Event Plane:
-
-### NIXL Storage Interface (for Backend Integration)
-
-The NIXL interface abstracts volume interaction and decouples it from mounting, metadata tracking, or direct system I/O. It provides:
-
-* registerVolume(descriptor): Register a logical volume for KV cache data.
-* unregisterVolume(): Cleanly deregister and release volume mappings.
-* get() / put(): Block-level APIs used by KVBM to fetch and store token blocks.
-
-These abstractions allow backends to be integrated without tying into the host's file system stack, enabling safe interaction with block devices, local filesystems, and RDMA-capable volumes. Please note that these APIs are still being finalized.
-
-### Dynamo Event Plane (Pub/Sub Coordination Layer)
-
-To support external storage optimizations without modifying KVBM logic, we provide an **event plane** built on NATS.io that emits lifecycle events for all block operations. Particularly there are two events emitted.
-
-* StoreEvent: Emitted when a KV block is registered.
-* RemoveEvent: Emitted when a KV block is released or evicted.
-
-Each KVEvent (\~100 bytes) contains:
-
-* sequence_hash: Unique identifier of the KV block
-* prefix_hash: Prefix grouping for query-level aggregation
-* block_size: Size in bytes
-* storage_location: Logical volume identifier
-* event_type: Store or Remove
-* extra_metadata: Reserved fields for partner-specific optimization
-
-For scalability, the system batches and publishes these events periodically (for example, every \~10s, or dynamically based on system load).
-
-### A conceptual design of a storage advisor
-
-This section provides an overview for the storage provider who is interested in integrating as a custom backend to KVBM and providing optimized performance. **Please note, this is optional for KVBM integration with a backend.**
-
-External storage systems are not tightly coupled with Dynamo's execution pipeline. Instead, they passively observe KV block lifecycle events through a subscription model:
-
-* Storage volumes are pre-provisioned and mounted by the storage provider.
-* These volumes are then registered with Dynamo through the NIXL Storage Agent using registerVolume() APIs. Dynamo itself does not manage mounts or provisioning.
-* The Dynamo KV Block Manager interacts only with logical block-level APIs (that is, get() and put()).
-* In parallel, the Event Plane asynchronously broadcasts KV lifecycle events using a NATS-based pub/sub channel.
-* Storage vendors implement a lightweight subscriber process that listens to these events without interfering with the KV Manager's runtime behavior.
-* This decoupling ensures that external storage systems can optimize block placement and lifecycle tracking without modifying or instrumenting the core Dynamo codebase.
-
-Now, to enable fast lookup and dynamic tiering, storage vendors may build internal data structures using the received event stream. Here is a high level conceptual design:
-
-* On receiving a StoreEvent, the storage system:
-  * Inserts a record into an internal prefix tree, hash map, or LRU index.
-  * This record includes the prefix_hash and sequence_hash, which logically identify the token block and its grouping.
-  * Associated metadata (for example, block_size, storage_location) is also captured.
-* On receiving a RemoveEvent, the system:
-  * Deletes or prunes the corresponding record from its index.
-  * Optionally triggers cleanup or tier migration workflows.
-
-This event-driven indexing allows the storage system to track which KV blocks are live and where they belong—enabling low-latency lookup, efficient space reclamation, and multi-tier coordination. With real-time visibility into KV block usage patterns, the storage system can implement smart tiering policies, such as:
-
-* Hot block promotion: Frequently accessed KV blocks can be migrated to fast SSD volumes.
-* Cold block demotion: Infrequently used blocks can be demoted to slower storage (for example, HDDs, cloud object storage).
-* Proactive compaction: If block sizes or prefix patterns indicate fragmentation, the storage backend can coalesce or rewrite blocks.
-
-These optimizations are performed entirely outside of Dynamo, with the assumption that storage providers adhere to SLA guarantees and volume availability.
-
-Critically, this entire system is designed to be non-intrusive:
-
-* The Dynamo KV Block Manager remains agnostic to how data is stored or optimized.
-* The Event Plane doesn't block or intercept any critical path of inference.
-* Storage vendors are given the freedom to innovate and optimize without requiring changes to the inference runtime.
-
-This design ensures that performance, resilience, and extensibility scale independently across the KV layer and the storage backend layer.
+KVBM design takes inspiration from the KV block managers used in vLLM and SGLang, with an added influence from historical memory tiering strategies common in general GPU programming. For more details, [See KVBM Reading](kvbm_reading.md). The figure below illustrates the internal components of KVBM.
+
+![Internal Components of Dynamo KVBM. ](../images/kvbm-components.png)
+**Internal Components of Dynamo KVBM**
+
+## KVBM Components
+### Core
+- **KvBlockManager**: Public facade. Constructs/owns the internal state and exposes the pools and onboarding APIs.
+- **Scheduler**: Gates transfer execution relative to model progress (iteration/layer completion) when integrated with a framework connector (e.g., vLLM V1).
+- **Config (config.rs)**: Describes model dims, page size, layout choices, and runtime flags used to build pools and layouts.
+- **KvBlockManagerState**: Central object wiring together layouts, storage backends, and pools; owns the OffloadManager, metrics, and events hooks.
+- **Events/Metrics**: Observability components emitting counters/gauges and event hooks for integration/testing.
+
+### Layouts and Blocks
+- **LayoutConfig & LayoutType**: Translate tensor shapes into KV cache layouts (layer-separated or fully-contiguous), including block counts and geometry.
+- **Blocks & Metadata**: Typed block handles (mutable/immutable), metadata (e.g., priority), and views by layer/outer dims; used to allocate, register, and match by `sequence_hash`.
+
+### Transfer Manager
+- **TransferManager**: Asynchronous transfer orchestrator with per-path queues (Device→Host, Host→Disk, Host→Device, Disk→Device).
+
+### Storage & Pools
+- **Device Pool(G1)**: GPU-resident KV block pool. Allocates mutable GPU blocks, registers completed blocks (immutable), serves lookups by sequence hash, and is the target for onboarding (Host→Device, Disk→Device).
+- **Host Pool(G2)**: CPU pinned-memory KV block pool. Receives Device offloads (Device→Host), can onboard to Device (Host→Device), and offloads to Disk. Uses pinned (page-locked) memory for efficient CUDA transfers and NIXL I/O.
+- **Disk Pool(G3)**: Local SSD NVMe-backed KV block pool. Receives Host offloads (Host→Disk) and provides blocks for onboarding to Device (Disk→Device). NIXL descriptors expose file offsets/regions for zero-copy I/O and optional GDS.
+
+## KVBM DataFlows
+![KVBM Data Flows. ](../images/kvbm-data-flows.png)
+**KVBM Data Flows from device to other memory hierarchies**
+
+**Device → Host (Offload)**
+* Triggered explicitly requested to offload by the connector scheduler.
+* Worker allocates a Host block and performs CUDA D2H/Custom Kernel copy.
+* Host pool registers the new immutable block (dedup by sequence hash).
+
+**Host → Disk (Offload)**
+* Local Disk: NIXL Write via POSIX; GDS when available.
+* Remote Disk (Network FS like NFS/Lustre/GPFS): NIXL Write via POSIX to the mounted FS; batching/concurrency identical.
+* Triggered on registered host blocks or explicit offload requests.
+* Worker allocates a Disk block and performs NIXL Write (Host→Disk).
+* Disk pool registers the new immutable block (dedup by sequence hash).
+
+**Host → Device (Onboard)**
+* Called to bring a host block into GPU memory.
+* Worker uses provided Device targets and performs CUDA H2D/Custom Kernel copy.
+* Device pool registers the new immutable block.
+
+**Disk → Device (Onboard)**
+* Called to bring a disk block directly into GPU memory.
+* Worker uses provided Device targets and performs NIXL Read (Disk→Device), possibly via GDS.
+* Device pool registers the new immutable block.

docs/kvbm/kvbm_design_deepdive.md

+<!--
+SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES.
+All rights reserved.
+SPDX-License-Identifier: Apache-2.0
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+-->
+
+# KVBM components
+
+The design of the KVBM is inspired from vLLM and SGLang KV block managers but with a twist from historical memory tiering design aspired in general GPU programming. [See KVBM Reading](kvbm_reading.md). The following figure shows the internal architecture of KVBM and how it works across workers using NIXL.
+
+![Internal architecture and key modules in the Dynamo KVBM. ](../images/kvbm-internal-arch.png)
+**Internal architecture and key modules in the Dynamo KVBM**
+
+## KvBlockManager as Orchestration Layer
+
+The `KvBlockManager <H, D>` acts as a coordinator across memory tiers—host (CPU), device (GPU), and remote—by managing per-backend block pools and exposing consistent block lifecycle APIs. It tracks KV block locations across device memory (G1), CPU memory within and across nodes (G2), local/pooled SSDs (G3), and remote storage (G4). G1-G4 are key tiers enabled by KVBM. Critical to note that KVBM treats G4 storage as an opaque blob store, unaware of internal layout optimizations.
+
+`KvBlockManager<H, D>` owns:
+
+* A device-side `BlockPool<Device>`
+* A host-side `BlockPool<Host>`
+* A remote NIXL agent that supports communication and memory sharing across nodes
+* A block set registry for remote lookup and import/export of block metadata
+
+Implementation-wise, `KvBlockManagerState` holds the logic: it's initialized by `KvBlockManagerConfig`, which merges runtime, model, and layout configurations. `NixlOptions` injects remote awareness.
+
+## Block Layout and Memory Mapping
+
+Each block is a 2D array `[num_layers][page_size × inner_dim]`. `BlockLayouttrait` abstracts the memory layout. The default implementation,`FullyContiguous`, stores all layers for all blocks in one region with alignment-aware stride computation:
+
+
+```none
+block_stride_in_bytes = align_up(num_layers × layer_stride, alignment);
+```
+
+
+Both CPU and GPU pools share this memory layout, but they use storage-specific backends:
+
+* `DeviceStorage` → CUDA device buffer
+* `PinnedStorage` → page-locked host memory
+* `SystemStorage` → CPU heap memory (fallback/test)
+* `NixlStorage` → remote memory through NIXL RDMA handles (includes storage)
+
+Each layout is constructed using a `LayoutConfig`, and storage is either passed directly or allocated using a StorageAllocator.
+
+## BlockPool and Memory Pools (Active and Inactive)
+
+Each `BlockPool<T>` (where `T` is `DeviceStorage`, `PinnedStorage`, and so forth) tracks two sub-pools:
+
+* `ActivePool`: Contains blocks currently in use by sequences
+* `InactivePool`: Recycled blocks ready for allocation; think free list
+
+When a token block is requested (for example, `get_mutable_block()`), the allocator pops from `InactivePool`, transitions its state, and returns a writable handle. On sequence commit or eviction, the system resets blocks and returns them to the inactive pool.
+
+The state machine (`BlockState`) that tracks the block lifecycle transitions includes:
+
+| State | Description | Ownership | Valid Actions/Transitions |
+| ----- | ----- | ----- | ----- |
+| Reset | Block hasn't been initialized or was reset. No associated sequence. | Held in InactivePool, reusable | init_sequence(salt_hash) → Partial |
+| Partial | Block is being filled with tokens for a new sequence. In-progress. | Owned by the sequence creator | add_token() / add_tokens() (accumulate)- commit() → Complete- reset() → Reset |
+| Complete | Block is fully filled with token data but not yet visible to others. | Still owned by creator thread | register() → Registered- reset() → Reset |
+| Registered | Block is finalized and visible for reuse. Available in the deduplication cache. Can use block for lookups | Shared ownership (global registry) | Auto drop() → triggers Remove event and transitions to Reset |
+
+This table lists the valid KVBM transitions:
+
+| From → To | Trigger | Validation |
+| ----- | ----- | ----- |
+| Reset → Partial | initsequence(salt_hash) | Must not be in use |
+| Partial → Complete | commit() | Must be full |
+| Complete → Registered | register() | Must be finalized |
+| Registered → Reset | Drop of RegistrationHandle | Automatic |
+| Partial → Reset | Aborted sequence | Explicit or drop |
+| Complete → Reset | Invalidated | Explicit or drop |
+
+Consider this example lifecycle of a block in the KVBM; in it, a sequence requests a new KV block:
+
+1. Allocator pops from InactivePool → Block is in Reset
+2. `init_sequence()` → Transitions to Partial
+3. Tokens are appended → State remains Partial
+4. On full → `commit()` → State becomes Complete
+5. `register()` → Block is hashed and moved to Registered. Blocks can now be used to lookup.
+6. On eviction or end-of-life → `drop()` of RAII handle returns block to Reset
+
+## Lifecycle Management using RAII and Event Plane
+
+The system uses RAII for memory lifecycle management. Every block holds metadata and registration state, and registration is coupled with an `EventManager`. On registration and drop:
+
+* `PublishHandle` triggers Register events
+* Dropping it triggers Remove events
+
+This pattern ensures consistency for shared memory tracking across workers without requiring explicit deallocation logic. The events are propagated in the Dynamo Events plane. Any Dynamo component subscribed to the events plane can listen to these changes. Note that even the storage provider can subscribe to the events plane and create an internal prefix tree representation that is tailored and optimized for the specific platform.
+
+## Remote Memory Integration using NIXL
+
+The NIXL agent exposes remote memory buffers using `NixlBlockSet`, `RemoteBlocks`, and layout descriptors. Key operations include:
+
+* `nixl_register()`: Registers memory region with NIXL runtime
+* `serialize() / deserialize()`: Converts layout and memory into transferable descriptors
+* `import_remote_blockset()`: Loads remote node's block layouts into the manager
+* `get_remote_blocks_mutable()`: Fetches transferable memory views from another node
+
+`RemoteBlocks` is a lightweight abstraction over shared memory for cross-node block usage (through UCX or other backends).
+
+The left side of the figure in [Understanding KVBM Components](#understanding-kvbm-components) illustrates a bidirectional remote memory registration and layout synchronization protocol between workers (for example, Worker 1 and Worker 2) using NIXL. The following steps break down the process:
+
+1. *Agent Creation & Memory Registration:*
+
+   Each worker independently sets up a NixlAgent:
+    * Registers its memory regions (that is, device memory) through `nixl_register()`.
+    * These regions correspond to blocks managed in the local BlockPool.
+      Once the worker registers the memory, NIXL creates remote-accessible descriptors, which it binds to the memory layout.
+
+2. *Metadata exchange:*
+
+   After memory registration, workers exchange serialized layout metadata, encapsulated in a `SerializedNixlBlockLayout`.
+
+    Why is this step critical?
+   * LLM inference workloads often differ in *tensor parallel (TP)* configurations:
+     * Worker 1 might have TP=4, while Worker 2 has TP=8.
+     * Thus, even if both systems use similar `FullyContiguous` layouts, their internal slicing and alignment assumptions differ.
+   * The metadata exchange bridges this semantic mismatch by sharing:
+     * LayoutConfig (num_layers, page_size, inner_dim, dtype)
+     * BlockSetID
+     * Base address + stride information (including alignment)
+     * Device ID + memory type (host/device)
+   * Once the workers share metadata, each can reconstruct the layout on its side using deserialize().
+  This enables NIXL to:
+      * Understand where each layer/block resides
+      * Perform correct gather-scatter operations during RDMA-like transfers
+
+    Without this step, remote fetches would result in data corruption or misaligned tokens.
+
+3. *Serialization & Deserialization: Making Layouts Portable*
+
+   In the serialization stage, KVBM exports and `FullyContiguous::serialize()` encodes:
+   * FullyContiguousConfig
+   * base_offset
+   * Physical memory descriptors (NixlStorage), including:
+     * Memory type (VRAM, DRAM)
+     * Address & size
+     * Device ID
+
+       The system sends this using NIXL transfer and then injects it into a KVBM scheduler state. In the deserialization stage, `SerializedNixlBlockLayout::deserialize()` rehydrates this into:
+       * A fully reconstructed memory layout view
+       * Local representation of a remote memory slice with correct offsets and size semantics
+       It also enables direct access to remote memory with consistent logical semantics
+       This guarantees that even across different system configurations (hardware or LLM shape), both parties agree on the memory view for each KV block.
+
+4. *Ownership handles and lifetime tracking*
+
+    Memory ownership in NIXL is tightly coupled with RAII-based handles:
+     * When a block is registered, it returns a `PublishHandle` which wraps a `RegistrationHandle`
+     * On drop of this handle, an automatic Remove event is published, which:
+       * Deregisters the block from the NIXL layer
+       * Removes it from the remote block registry
+     * This ensures that, once the block is evicted from the cache or no longer used in inference, all references are invalidated cleanly across nodes
+       This mechanism avoids:
+     * Stale memory access
+     * Dangling pointers on GPU or host
+     * Manual deregistration bugs
+       The system can batch and publish registration events using a Publisher, optimizing performance under high concurrency
+
+
+## Storage backends and pluggability
+
+You can integrate KVBM with a storage backend by extending or wrapping `NixlEnabledStorage` to support cross-node RDMA registration. All layouts and block pools are generic over these backends, allowing for fine-grained control over memory tiers. We defer detailed integration guidance, since we collaborate with storage partners to simplify and standardize these integration paths.
+
+```mermaid
+---
+title: Example KVBM System Architecture
+---
+flowchart TD
+    A["Distributed Inference Engine"] --> B["Dynamo KV Block Manager"]
+
+    B --> C["NIXL Storage Agent<br/>- Volume registration<br/>- get()/put() abstraction"]
+    B --> D["Event Plane<br/>- NATS-based Pub/Sub<br/>- StoreEvent / RemoveEvent"]
+
+    C --> E["G4 Storage Infrastructure<br/>(SSD, Object store, etc.)<br/>- Store KV blocks"]
+    D --> F["Storage Provider Subscriber<br/>- Parse Events<br/>- Build fast tree/index<br/>- Optimize G4 tiering"]
+```
+
+For now, the following breakdown provides a high-level understanding of how KVBM interacts with external storage using the NIXL storage interface and the Dynamo Event Plane:
+
+### NIXL Storage Interface (for Backend Integration)
+
+The NIXL interface abstracts volume interaction and decouples it from mounting, metadata tracking, or direct system I/O. It provides:
+
+* registerVolume(descriptor): Register a logical volume for KV cache data.
+* unregisterVolume(): Cleanly deregister and release volume mappings.
+* get() / put(): Block-level APIs used by KVBM to fetch and store token blocks.
+
+These abstractions allow backends to be integrated without tying into the host's file system stack, enabling safe interaction with block devices, local filesystems, and RDMA-capable volumes. Please note that these APIs are still being finalized.
+
+### Dynamo Event Plane (Pub/Sub Coordination Layer)
+
+To support external storage optimizations without modifying KVBM logic, we provide an **event plane** built on NATS.io that emits lifecycle events for all block operations. Particularly there are two events emitted.
+
+* StoreEvent: Emitted when a KV block is registered.
+* RemoveEvent: Emitted when a KV block is released or evicted.
+
+Each KVEvent (\~100 bytes) contains:
+
+* sequence_hash: Unique identifier of the KV block
+* prefix_hash: Prefix grouping for query-level aggregation
+* block_size: Size in bytes
+* storage_location: Logical volume identifier
+* event_type: Store or Remove
+* extra_metadata: Reserved fields for partner-specific optimization
+
+For scalability, the system batches and publishes these events periodically (for example, every \~10s, or dynamically based on system load).
+
+### A conceptual design of a storage advisor
+
+This section provides an overview for the storage provider who is interested in integrating as a custom backend to KVBM and providing optimized performance. **Please note, this is optional for KVBM integration with a backend.**
+
+External storage systems are not tightly coupled with Dynamo's execution pipeline. Instead, they passively observe KV block lifecycle events through a subscription model:
+
+* Storage volumes are pre-provisioned and mounted by the storage provider.
+* These volumes are then registered with Dynamo through the NIXL Storage Agent using registerVolume() APIs. Dynamo itself does not manage mounts or provisioning.
+* The Dynamo KV Block Manager interacts only with logical block-level APIs (that is, get() and put()).
+* In parallel, the Event Plane asynchronously broadcasts KV lifecycle events using a NATS-based pub/sub channel.
+* Storage vendors implement a lightweight subscriber process that listens to these events without interfering with the KV Manager's runtime behavior.
+* This decoupling ensures that external storage systems can optimize block placement and lifecycle tracking without modifying or instrumenting the core Dynamo codebase.
+
+Now, to enable fast lookup and dynamic tiering, storage vendors may build internal data structures using the received event stream. Here is a high level conceptual design:
+
+* On receiving a StoreEvent, the storage system:
+  * Inserts a record into an internal prefix tree, hash map, or LRU index.
+  * This record includes the prefix_hash and sequence_hash, which logically identify the token block and its grouping.
+  * Associated metadata (for example, block_size, storage_location) is also captured.
+* On receiving a RemoveEvent, the system:
+  * Deletes or prunes the corresponding record from its index.
+  * Optionally triggers cleanup or tier migration workflows.
+
+This event-driven indexing allows the storage system to track which KV blocks are live and where they belong—enabling low-latency lookup, efficient space reclamation, and multi-tier coordination. With real-time visibility into KV block usage patterns, the storage system can implement smart tiering policies, such as:
+
+* Hot block promotion: Frequently accessed KV blocks can be migrated to fast SSD volumes.
+* Cold block demotion: Infrequently used blocks can be demoted to slower storage (for example, HDDs, cloud object storage).
+* Proactive compaction: If block sizes or prefix patterns indicate fragmentation, the storage backend can coalesce or rewrite blocks.
+
+These optimizations are performed entirely outside of Dynamo, with the assumption that storage providers adhere to SLA guarantees and volume availability.
+
+Critically, this entire system is designed to be non-intrusive:
+
+* The Dynamo KV Block Manager remains agnostic to how data is stored or optimized.
+* The Event Plane doesn't block or intercept any critical path of inference.
+* Storage vendors are given the freedom to innovate and optimize without requiring changes to the inference runtime.
+
+This design ensures that performance, resilience, and extensibility scale independently across the KV layer and the storage backend layer.

docs/kvbm/kvbm_integrations.md

+<!--
+SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES.
+All rights reserved.
+SPDX-License-Identifier: Apache-2.0
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+-->
+
+# KVBM Integrations
+
+KVBM Integrates with Inference frameworks (vLLM, TRTLLM, SGLang) via Connector APIs to influence KV caching behaviour, scheduling, and forward pass execution.
+There are two components of the interface, Scheduler and Worker. Scheduler(leader) is responsible for the orchestration of KV block offload/onboard, builds metadata specifying transfer data to the workers. It also maintains hooks for handling asynchronous transfer completion. Worker is responsible for reading metadata built by the scheduler(leader), does async onboarding/ offloading at the end of the forward pass.
+
+## Typical KVBM Integrations
+
+The following figure shows the typical integration of KVBM with inference frameworks (vLLM used as an example)
+
+![vLLM KVBM Integration ](../images/kvbm-integrations.png)
+**vLLM KVBM Integration**
+
+
+## How to run KVBM with Frameworks
+* Instructions to [run KVBM in vLLM](vllm-setup.md)
+* Instructions to [run KVBM with TRTLLM](trtllm-setup.md)
+
+## Onboarding
+![Onboarding blocks from Host to Device](../images/kvbm-onboard-host2device.png)
+**Onboarding blocks from Host to Device**
+![Onboarding blocks from Disk to Device](../images/kvbm-onboard-disk2device.png)
+**Onboarding blocks from Disk to Device**
+
+## Offloading
+![Offloading blocks from Device to Host&Disk](../images/kvbm-offload.png)
+**Offloading blocks from Device to Host&Disk**

docs/kvbm/kvbm_intro.rst

@@ -20,7 +20,7 @@ The Dynamo KV Block Manager (KVBM) is a scalable runtime component designed to h
 
 It offers:
 
-* A **unified memory API** that spans GPU memory, pinned host memory, remote RDMA-accessible memory, local or distributed pool of SSDs and remote file/object/cloud storage systems.
+* A **unified memory API** that spans GPU memory(in future) , pinned host memory, remote RDMA-accessible memory, local or distributed pool of SSDs and remote file/object/cloud storage systems.
 * Support for evolving **block lifecycles** (allocate → register → match) with event-based state transitions that storage can subscribe to.
 * Integration with **NIXL**, a dynamic memory exchange layer used for remote registration, sharing, and access of memory blocks over RDMA/NVLink.
 
@@ -43,7 +43,7 @@ The Dynamo KV Block Manager serves as a reference implementation that emphasizes
      - ✅
      - vLLM
    * -
-     - ❌
+     - ✅
      - TensorRT-LLM
    * -
      - ❌
@@ -60,7 +60,11 @@ The Dynamo KV Block Manager serves as a reference implementation that emphasizes
 
    Overview <self>
    Motivation <kvbm_motivation.md>
-   KVBM Architecture <kvbm_architecture.md>
-   Understanding KVBM components <kvbm_components.md>
-   KVBM Further Reading <kvbm_reading>
+   Architecture <kvbm_architecture.md>
+   Components <kvbm_components.md>
+   Design Deep Dive <kvbm_design_deepdive.md>
+   Integrations <kvbm_integrations.md>
+   KVBM in vLLM <vllm-setup.md>
+   KVBM in TRTLLM <trtllm-setup.md>
    LMCache Integration <../backends/vllm/LMCache_Integration>
+   Further Reading <kvbm_reading.md>
\ No newline at end of file

docs/kvbm/kvbm_motivation.md

@@ -28,3 +28,16 @@ Large language models (LLMs) and other AI workloads increasingly rely on KV cach
 * Difficulty in optimizing storage placement across heterogeneous storage tiers (for example, SSDs, object storage, and cloud storage).
 
 Conventional systems either lack dynamic feedback mechanisms or require deep integration into core storage paths, which both increases complexity and reduces portability.
+
+## Benefits of KV Cache offloading
+
+KV Cache offloading avoids expensive KV Cache recomputation, resulting in faster response times and a better user experience. In the end, providers benefit from higher throughput and lower cost per token, making their inference services more scalable and efficient.
+
+## When to offload KV Cache for reuse
+
+Offloading KV cache to CPU or storage is most effective when KV Cache exceeds GPU memory and cache reuse outweighs the overhead of transferring data. It is especially valuable in long-context, high-concurrency, or resource-constrained inference environments such as:
+
+* **Long sessions and multi-turn conversations:** Offloading preserves large prompt prefixes, avoids recomputation, and improves first-token latency and throughput.
+* **High concurrency (future):** Idle or partial conversations can be moved out of GPU memory, allowing active requests to proceed without hitting memory limits.
+* **Shared or repeated content (future):** Reuse across users or sessions (for example, system prompts and templates) increases cache hits, especially with remote or cross-instance sharing.
+* **Memory- or cost-constrained deployments:** Offloading to RAM or SSD reduces GPU demand, allowing longer prompts or more users without adding hardware.

docs/kvbm/kvbm_reading.md

@@ -18,6 +18,6 @@ limitations under the License.
 
 # KVBM Further Reading
 
-- [vLLM](https://docs.vllm.ai/en/latest/design/automatic_prefix_caching.html)
-- [SGLang](https://github.com/sgl-project/sglang/pull/2693)
+- [vLLM](https://docs.vllm.ai/en/latest/features/automatic_prefix_caching.html)
+- [SGLang](https://github.com/sgl-project/sglang/tree/main/benchmark/hicache)
 - [EMOGI](https://arxiv.org/abs/2006.06890)
\ No newline at end of file

docs/kvbm/trtllm-setup.md

@@ -17,9 +17,9 @@ limitations under the License.
 
 # Running KVBM in TensorRT-LLM
 
-This guide explains how to leverage KVBM (KV Block Manager) to mange KV cache and do KV offloading in TensorRT-LLM (trtllm).
+This guide explains how to leverage KVBM (KV Block Manager) to manage KV cache and do KV offloading in TensorRT-LLM (trtllm).
 
-To learn what KVBM is, please check [here](https://docs.nvidia.com/dynamo/latest/kvbm/kvbm_intro.html)
+To learn what KVBM is, please check [here](kvbm_architecture.md)
 
 > [!Note]
 > - Ensure that `etcd` and `nats` are running before starting.