Subtalker paramater discrepancy between server and cli

[esdevver]

When the parameters are checked against defaults in tts-server.cpp, it sets subtalker_top_k to top_k, and subtalker_top_p to top_p. Keeping these at default parameters would have parity with the behaviour of the cli tool.

qwentts.cpp/tools/tts-server.cpp

Lines 268 to 297 in a8a7716

	// Sampling overrides ride straight into the ABI; the subtalker
	// mirrors the talker knobs so the HTTP surface stays a single
	// coherent set. A temperature of zero selects greedy decoding
	// on both.
	p.seed = req.seed;
	if (req.max_new_tokens != -1) {
	p.max_new_tokens = req.max_new_tokens;
	}
	if (req.top_k != -1) {
	p.top_k = req.top_k;
	p.subtalker_top_k = req.top_k;
	}
	if (!std::isnan(req.temperature)) {
	if (req.temperature == 0.0f) {
	p.do_sample = false;
	p.subtalker_do_sample = false;
	} else {
	p.temperature = req.temperature;
	p.subtalker_temperature = req.temperature;
	}
	}
	if (!std::isnan(req.top_p)) {
	p.top_p = req.top_p;
	p.subtalker_top_p = req.top_p;
	}
	if (!std::isnan(req.repetition_penalty)) {
	p.repetition_penalty = req.repetition_penalty;
	}

I recommend either removing lines 279 and 287, or properly exposing subtalker parameters (temp, top_k and top_p) to the REST API.