Talker Token Ids#

struct TalkerTokenIds#

Tokens a prefill layout writes as prefill rows or finds in the token stream.

Public Members

int32_t ttsPad = {-1}#

Every layout’s text-side rows. tts_eos is spelled <tts_text_eod>; <tts_text_bos_single> is another token.

int32_t ttsBos = {-1}#
int32_t ttsEos = {-1}#
int32_t imStart = {-1}#

Chat-template tokens, -1 unless asked for: the tts layout strips a fixed wrapper instead.

int32_t imEnd = {-1}#
int32_t systemRole = {-1}#
int32_t userRole = {-1}#
int32_t assistantRole = {-1}#
int32_t thinkOpen = {-1}#
int32_t thinkClose = {-1}#
int32_t newline = {-1}#
int32_t doubleNewline = {-1}#
int32_t audioPad = {-1}#

Media placeholders, -1 unless asked for: only the segments layout looks for them.

int32_t imagePad = {-1}#
int32_t videoPad = {-1}#
TalkerTokenIds trt_edgellm::rt::qwen3_omni::resolveTalkerTokenIds(
tokenizer::Tokenizer const &tokenizer,
bool withChatTemplate,
bool withMediaPads
)#

Look the tts tokens up in tokenizer, the chat-template tokens only if withChatTemplate and the media placeholders only if withMediaPads. Throws naming the first string that does not encode to exactly one token spelling all of it.