Block Hash

struct Hash128

Two-word representation used for stable context-cache identities.

Public Members

uint64_t hi = {}

uint64_t lo = {}

struct AdapterKey

Public Members

Hash128 id = {}

uint64_t generation = {}

struct MediaSpanKey

Public Members

Hash128 contentDigest = {}

Hash128 placementDigest = {}

int32_t offsetWithinBlock = {}

int32_t itemOrder = {}

uint8_t modality = {}

struct BlockKeyExtras

Public Members

std::vector<MediaSpanKey> media

Non-token inputs whose content or placement changes the state produced for this block.

std::optional<AdapterKey> adapter

std::optional<Hash128> positionDigest

std::optional<Hash128> customEmbeddingDigest

std::optional<Hash128> isolationDigest

template<>
struct hash<trt_edgellm::rt::Hash128>

Public Functions

inline size_t operator()(

trt_edgellm::rt::Hash128 const &value

) const noexcept

Hash128 trt_edgellm::rt::hashOpaqueIdentity(std::string_view bytes)

Deterministically hash exact opaque bytes for adapter, media, or isolation identity. This uses the same non-cryptographic 128-bit FNV-1a primitive as block chaining.

BlockHash trt_edgellm::rt::hashBlock(

BlockHash parent,

int32_t const *tokens,

size_t count,

BlockKeyExtras const &extras,

Hash128 const *perPositionMediaHash

)

Hash one caller-supplied token block and its non-token identity into its parent chain.

The implementation uses deterministic 128-bit Fowler-Noll-Vo (FNV-1a) over a tagged byte stream. It is a non-cryptographic hash; this layer does not perform a secondary token comparison when two hashes are equal.

When perPositionMediaHash is non-null, at each token position where the entry is non-zero the 128-bit content hash is fed into the byte stream instead of the 4-byte token ID. A zero Hash128 means “use the token ID”. Passing nullptr (default) uses token IDs for all positions.

std::vector<BlockHash> trt_edgellm::rt::hashFullBlocks(

int32_t const *tokens,

size_t tokenCount,

int32_t pageSize,

std::vector<BlockKeyExtras> const &extrasPerBlock,

Hash128 const *perPositionMediaHash

)

Hash each complete token block in order, chaining from kCHAIN_ROOT and excluding any partial tail.

When perPositionMediaHash is non-null it must cover the full token range (tokenCount entries). The appropriate slice is passed to each block.

BlockHash trt_edgellm::rt::hashExactPrefix(

int32_t const *tokens,

size_t tokenCount,

int32_t pageSize,

std::vector<BlockKeyExtras> const &extrasPerBlock

)

Hash an exact token prefix, including a partial final block when present.

When extras are supplied, one entry is required for every touched block (ceil(tokenCount / pageSize)). An aligned prefix returns the same digest as its final complete BlockHash; an empty prefix returns kCHAIN_ROOT.