forgotten wip changes

src/solanaceae/llama-cpp-web/llama_cpp_web_impl.cpp

@@ -77,7 +77,7 @@ int64_t LlamaCppWeb::completeSelect(const std::string_view prompt, const std::ve
 	}
 	//grammar += ")";
 
-	//std::cout << "generated grammar:\n" << grammar << "\n";
+	std::cerr << "generated grammar:\n" << grammar << "\n";
 
 	auto ret = complete(nlohmann::json{
 		{"prompt", prompt},
@@ -89,19 +89,23 @@ int64_t LlamaCppWeb::completeSelect(const std::string_view prompt, const std::ve
 		{"top_p", 1.0}, // disable
 		{"n_predict", 256}, // unlikely to ever be so high
 		{"seed", _rng()},
+		{"ignore_eos", true},
 		{"cache_prompt", static_cast<bool>(_use_server_cache)},
 	});
 
 	if (ret.empty()) {
+		assert("ret empty" && false);
 		return -2;
 	}
 
 	if (!ret.count("content")) {
+		assert("no content" && false);
 		return -3;
 	}
 
 	std::string selected = ret.at("content");
 	if (selected.empty()) {
+		assert("content empty" && false);
 		return -4;
 	}
 
@@ -111,6 +115,7 @@ int64_t LlamaCppWeb::completeSelect(const std::string_view prompt, const std::ve
 		}
 	}
 
+	std::cerr << "content does not contain match\n";
 	std::cerr << "complete failed j:'" << ret.dump() << "'\n";
 	return -5;
 }
@@ -125,7 +130,7 @@ std::string LlamaCppWeb::completeLine(const std::string_view prompt) {
 		{"top_p", 1.0}, // disable
 		{"n_predict", 400},
 		{"seed", _rng()},
-		{"stop", {"\n"}},
+		{"stop", nlohmann::json::array({"\n"})},
 		{"cache_prompt", static_cast<bool>(_use_server_cache)},
 	});
 
@@ -147,7 +152,7 @@ nlohmann::json LlamaCppWeb::complete(const nlohmann::json& request_j) {
 	// steaming instead would be better
 	_cli.set_read_timeout(std::chrono::minutes(10));
 
-	//std::cout << "j dump: '" << request_j.dump(-1, ' ', true) << "'\n";
+	std::cerr << "j dump: '" << request_j.dump(-1, ' ', true) << "'\n";
 
 	auto res = _cli.Post("/completion", request_j.dump(-1, ' ', true), "application/json");
 
@@ -159,7 +164,7 @@ nlohmann::json LlamaCppWeb::complete(const nlohmann::json& request_j) {
 		//res->body.empty() ||
 		//res->get_header_value("Content-Type") != "application/json"
 	) {
-		std::cerr << "error posting\n";
+		std::cerr << "error posting: '" << res->body << "'\n";
 		return {};
 	}
 

src/solanaceae/llama-cpp-web/llama_cpp_web_impl.hpp

@@ -19,7 +19,7 @@ struct LlamaCppWeb : public TextCompletionI {
 	// this is a bad idea
 	static std::minstd_rand thread_local _rng;
 
-	std::atomic<bool> _use_server_cache {true};
+	std::atomic<bool> _use_server_cache {false};
 
 	LlamaCppWeb(
 		ConfigModelI& conf

src/test1.cpp

@@ -20,7 +20,7 @@ int main(void) {
 	}
 	std::cerr << lcw._cli.host() << " " << lcw._cli.port() << " endpoint healthy\n";
 
-	std::cout << "The meaning of life is to"
+	std::cerr << "The meaning of life is to"
 		<< lcw.complete(nlohmann::json{
 			{"prompt", "The meaning of life is to"},
 			{"min_p", 0.1}, // model dependent
@@ -34,9 +34,9 @@ int main(void) {
 		})
 		<< "\n";
 
-	std::cout << "-------------------------\n";
+	std::cerr << "-------------------------\n";
 
-	std::cout << "complete from select:\n";
+	std::cerr << "complete from select:\n";
 	std::vector<std::string_view> possible {
 		" die",
 		" die.",
@@ -46,12 +46,12 @@ int main(void) {
 		" Hi",
 	};
 	for (size_t i = 0; i < 10; i++) {
-		std::cout << "The meaning of life is to";
+		std::cerr << "The meaning of life is to";
 		auto res = lcw.completeSelect("The meaning of life is to", possible);
 		if (res < 0) {
-			std::cout << " error--\n";
+			std::cerr << " error\n";
 		} else {
-			std::cout << possible[res] << "\n";
+			std::cerr << possible[res] << "\n";
 		}
 	}