sampler : remove "full-context windows" from history-based samplers (#26524)

* Resolve -1 to 1024 instead of ctx-len for samplers

Because of backend-sampling we initialize samplers before the complete
llama_context is there. Therefore, we cannot infer the resolved context
length yet at the time we construct the samplers.

* Shared default of 64 for history-based samplers, remove context_size
This commit is contained in:
Oliver Simons
2026-08-04 21:28:55 +03:00
committed by GitHub
parent 76c956c137
commit a6aa6f5450
18 changed files with 52 additions and 80 deletions
+2 -7
View File
@@ -186,8 +186,7 @@ std::string common_params_sampling::print() const {
struct common_sampler * common_sampler_init(
const struct llama_model * model,
struct common_params_sampling & params,
int32_t n_ctx) {
struct common_params_sampling & params) {
if (!std::isfinite(params.penalty_repeat) ||
params.penalty_repeat <= 0.0f ||
!std::isfinite(1.0f/params.penalty_repeat)) {
@@ -199,10 +198,6 @@ struct common_sampler * common_sampler_init(
if (!std::isfinite(params.penalty_present)) {
throw std::invalid_argument("penalty_present must be finite");
}
if (params.penalty_last_n == -1) {
params.penalty_last_n = n_ctx > 0 ? n_ctx : llama_model_n_ctx_train(model);
}
const llama_vocab * vocab = llama_model_get_vocab(model);
llama_sampler_chain_params lparams = llama_sampler_chain_default_params();
@@ -355,7 +350,7 @@ struct common_sampler * common_sampler_init(
for (const auto & str : params.dry_sequence_breakers) {
c_breakers.push_back(str.c_str());
}
samplers.push_back(llama_sampler_init_dry(vocab, llama_model_n_ctx_train(model), params.dry_multiplier, params.dry_base, params.dry_allowed_length, params.dry_penalty_last_n, c_breakers.data(), c_breakers.size()));
samplers.push_back(llama_sampler_init_dry(vocab, params.dry_multiplier, params.dry_base, params.dry_allowed_length, params.dry_penalty_last_n, c_breakers.data(), c_breakers.size()));
}
break;
case COMMON_SAMPLER_TYPE_TOP_K: