Add HTTPS support via cpp-httplib's mbedtls backend (#1011)

Author: aittalam

llama.cpp.patches/README.md

@@ -86,7 +86,29 @@ Cosmopolitan libc has specific behaviors with condition variables and signals th
 | `common_log.cpp.patch` | Adds `#include <csignal>`; blocks `SIGINT`/`SIGTERM` on logger thread via `pthread_sigmask` to prevent `EINTR` exceptions; replaces `cv.wait()` with `wait_for(30s)` loop to work around XNU futex timeout bug (~72 minute expiry) |
 | `tools_server_server-models.cpp.patch` | Adds `#include <csignal>`; blocks signals on the stopping thread via `pthread_sigmask`; replaces untimed `cv.wait()` with `wait_for(30s)` loops on every model-lifecycle wait (`unload_lru`, the reload-drain wait, `stopping_thread`, the `is_reloading` guard in `load`, and the generic `wait()` predicate helper) to work around the XNU futex timeout bug |
 | `tools_server_server-queue.cpp.patch` | Adds missing includes (`<cerrno>`, `<system_error>`, `<csignal>`); blocks `SIGINT`/`SIGTERM` on queue thread; replaces `wait()` with `wait_for()` loops in three locations (`wait_until_no_sleep`, main loop, `recv`) |
-| `vendor_cpp-httplib_httplib.cpp.patch` | Fixes httplib thread pool with `wait_for()` instead of `wait()` for XNU futex compatibility |
+| `vendor_cpp-httplib_httplib.cpp.patch` | Fixes httplib thread pool with `wait_for()` instead of `wait()` for XNU futex compatibility; also see HTTPS / TLS Support below |
+
+### HTTPS / TLS Support
+
+Upstream llama.cpp gets TLS from cpp-httplib's OpenSSL backend
+(`CPPHTTPLIB_OPENSSL_SUPPORT`, satisfied by system OpenSSL or vendored
+BoringSSL/LibreSSL at cmake time). None of those is available in the
+cosmocc make build, so llamafile instead enables cpp-httplib's **Mbed TLS
+backend** (`CPPHTTPLIB_MBEDTLS_SUPPORT`) against the mbedtls fork already
+vendored in `third_party/mbedtls` — the same TLS stack llamafile <= 0.9.3
+used. `third_party/mbedtls/include/` maps the canonical `<mbedtls/*.h>`
+include paths onto the fork's headers, and `BUILD.mk` sets the macro on
+every object that can reach `httplib.h` (the macro changes httplib class
+layouts, so all TUs must agree) and links `mbedtls.a` into `llama-server`.
+This enables HTTPS model downloads (`-hf`, `--model-url`), https clients
+in server-models, and TLS serving via `--ssl-cert-file`/`--ssl-key-file`.
+
+| Patch | Description |
+|-------|-------------|
+| `common_http.h.patch` | `#ifndef CPPHTTPLIB_OPENSSL_SUPPORT` -> `#ifndef CPPHTTPLIB_SSL_ENABLED` for the "HTTPS is not supported" guard, so any cpp-httplib TLS backend counts (candidate for upstreaming) |
+| `tools_server_server-http.cpp.patch` | Same macro fix for the `httplib::SSLServer` (`--ssl-cert-file`/`--ssl-key-file`) guard (candidate for upstreaming) |
+| `tools_server_server-models.cpp.patch` | Same macro fix for the direct `httplib::SSLClient` construction in `server_http_proxy` (candidate for upstreaming) |
+| `vendor_cpp-httplib_httplib.cpp.patch` | Under `__COSMOPOLITAN__`: appends `/zip/third_party/mbedtls/sslroot` to `system_ca_dirs()` as the trust-store fallback (essential on Windows hosts, where the `_WIN32` cert-store branches are not compiled into an APE), and `__static_yoink("ssl_root_support")` so the Mozilla root PEMs bundled by `third_party/mbedtls/BUILD.mk` are pulled into the executable's zip |
 
 ### TinyBLAS Integration
 

llama.cpp.patches/llamafile-files/BUILD.mk

@@ -460,6 +460,19 @@ $(TOOL_PERPLEXITY_OBJS) $(TOOL_BENCH_OBJS) $(TOOL_SERVER_OBJS) $(MTMD_OBJS): \
 		-iquote $(UI_GEN_DIR) \
 		-isystem llama.cpp/vendor
 
+# HTTPS support: build cpp-httplib with its Mbed TLS backend against the
+# mbedtls vendored in third_party/mbedtls (the same TLS stack llamafile
+# <= 0.9.3 used); third_party/mbedtls/include maps the canonical
+# <mbedtls/*.h> include paths onto it. The macro changes httplib class
+# layouts, so every object that includes httplib.h (directly or via
+# common/http.h, server-http.h, server-cors-proxy.h) must see it.
+$(LLAMA_CPP_OBJS) $(TOOL_QUANTIZE_OBJS) $(TOOL_IMATRIX_OBJS) \
+$(TOOL_PERPLEXITY_OBJS) $(TOOL_BENCH_OBJS) $(TOOL_SERVER_OBJS) $(MTMD_OBJS) \
+$(HTTPLIB_OBJS): \
+	private CPPFLAGS += \
+		-DCPPHTTPLIB_MBEDTLS_SUPPORT \
+		-isystem third_party/mbedtls/include
+
 # Server needs llamafile headers for Metal support.
 # The generated ui.h sits in $(UI_GEN_DIR); the -iquote above lets every
 # server source resolve `#include "ui.h"`.
@@ -576,9 +589,10 @@ o/$(MODE)/llama.cpp/server/llama-server: \
 	$(HTTPLIB_OBJS) \
 	$(TOOL_LLAMAFILE_OBJS) \
 	$$(TINYBLAS_CPU_OBJS) \
-	o/$(MODE)/llama.cpp/llama.cpp.a
+	o/$(MODE)/llama.cpp/llama.cpp.a \
+	o/$(MODE)/third_party/mbedtls/mbedtls.a
 	@mkdir -p $(dir $@)
-	$(LINK.o) $(TOOL_SERVER_OBJS) $(UI_GEN_OBJ) $(MTMD_OBJS) $(HTTPLIB_OBJS) $(TOOL_LLAMAFILE_OBJS) $(TINYBLAS_CPU_OBJS) o/$(MODE)/llama.cpp/llama.cpp.a $(LOADLIBES) $(LDLIBS) -o $@
+	$(LINK.o) $(TOOL_SERVER_OBJS) $(UI_GEN_OBJ) $(MTMD_OBJS) $(HTTPLIB_OBJS) $(TOOL_LLAMAFILE_OBJS) $(TINYBLAS_CPU_OBJS) o/$(MODE)/llama.cpp/llama.cpp.a o/$(MODE)/third_party/mbedtls/mbedtls.a $(LOADLIBES) $(LDLIBS) -o $@
 
 # ==============================================================================
 # Dependencies

llama.cpp.patches/patches/common_http.h.patch

@@ -0,0 +1,12 @@
+diff --git a/common/http.h b/common/http.h
+--- a/llama.cpp/common/http.h
++++ b/llama.cpp/common/http.h
+@@ -72,7 +72,7 @@ static std::pair<httplib::Client, common_http_url> common_http_client(const std:
+         throw std::runtime_error("error: invalid URL format");
+     }
+ 
+-#ifndef CPPHTTPLIB_OPENSSL_SUPPORT
++#ifndef CPPHTTPLIB_SSL_ENABLED
+     if (parts.scheme == "https") {
+         throw std::runtime_error(
+             "HTTPS is not supported. Please rebuild with one of:\n"

llama.cpp.patches/patches/tools_server_server-http.cpp.patch

@@ -0,0 +1,12 @@
+diff --git a/tools/server/server-http.cpp b/tools/server/server-http.cpp
+--- a/llama.cpp/tools/server/server-http.cpp
++++ b/llama.cpp/tools/server/server-http.cpp
+@@ -93,7 +93,7 @@ bool server_http_context::init(const common_params & params) {
+ 
+     auto & srv = pimpl->srv;
+ 
+-#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
++#ifdef CPPHTTPLIB_SSL_ENABLED
+     if (!params.ssl_file_key.empty() && !params.ssl_file_cert.empty()) {
+         SRV_INF("running with SSL: key = %s, cert = %s\n", params.ssl_file_key.c_str(), params.ssl_file_cert.c_str());
+         srv = std::make_unique<httplib::SSLServer>(

llama.cpp.patches/patches/tools_server_server-models.cpp.patch

@@ -120,3 +120,16 @@ diff --git a/tools/server/server-models.cpp b/tools/server/server-models.cpp
  }
 
  bool server_models::ensure_model_ready(const std::string & name) {
+@@ -1894,10 +1928,10 @@ server_http_proxy::server_http_proxy(
+     auto pipe = std::make_shared<pipe_t<msg_t>>();
+ 
+     if (scheme == "https") {
+-#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
++#ifdef CPPHTTPLIB_SSL_ENABLED
+         cli.reset(new httplib::SSLClient(host, port));
+ #else
+-        throw std::runtime_error("HTTPS requested but CPPHTTPLIB_OPENSSL_SUPPORT is not defined");
++        throw std::runtime_error("HTTPS requested but the server is built without SSL support");
+ #endif
+     }
+ 

llama.cpp.patches/patches/vendor_cpp-httplib_httplib.cpp.patch

@@ -39,3 +39,28 @@ diff --git a/vendor/cpp-httplib/httplib.cpp b/vendor/cpp-httplib/httplib.cpp
        }
 
        idle_thread_count_--;
+@@ -12625,6 +12647,13 @@ bool enumerate_macos_keychain_certs(Callback cb) {
+ 
+ #if !defined(_WIN32) && !(defined(__APPLE__) &&                                \
+                           defined(CPPHTTPLIB_USE_CERTS_FROM_MACOSX_KEYCHAIN))
++#ifdef __COSMOPOLITAN__
++// Reference the Mozilla CA roots zipped into the APE (see llamafile's
++// third_party/mbedtls/sslroot) so the linker keeps them; they serve as
++// the trust-store fallback below on hosts without a system CA store
++// (notably Windows, where the _WIN32 branches above are not compiled).
++__static_yoink("ssl_root_support");
++#endif
+ // Common CA certificate file paths on Linux/Unix
+ const char **system_ca_paths() {
+   static const char *paths[] = {
+@@ -12642,6 +12671,10 @@ const char **system_ca_dirs() {
+   static const char *dirs[] = {"/etc/ssl/certs",             // Debian/Ubuntu
+                                "/etc/pki/tls/certs",         // RHEL/CentOS
+                                "/usr/share/ca-certificates", // Other
++#ifdef __COSMOPOLITAN__
++                               // CA bundle zipped into the APE (llamafile)
++                               "/zip/third_party/mbedtls/sslroot",
++#endif
+                                nullptr};
+   return dirs;
+ }

llamafile/BUILD.mk

@@ -62,6 +62,20 @@ LLAMAFILE_CPPFLAGS := \
 	-DLLAMAFILE_TUI \
 	-DCOSMOCC=1
 
+# Flags for every TU that includes httplib.h: cpp-httplib is built with its
+# Mbed TLS backend (see llama.cpp/BUILD.mk), and CPPHTTPLIB_MBEDTLS_SUPPORT
+# changes httplib class layouts (e.g. httplib::Result, httplib::Client), so
+# an includer compiled without it corrupts memory when it exchanges those
+# types with httplib.cpp. If you add an httplib include to a llamafile
+# source, add its object below next to the chatbot ones.
+# -iquote . and -mcosmo are for the vendored mbedtls headers themselves
+# (repo-rooted internal includes, cosmo-extension macros).
+LLAMAFILE_HTTPLIB_TLS_FLAGS := \
+	-DCPPHTTPLIB_MBEDTLS_SUPPORT \
+	-isystem third_party/mbedtls/include \
+	-iquote . \
+	-mcosmo
+
 # ==============================================================================
 # Source files - Highlight library
 # ==============================================================================
@@ -285,17 +299,21 @@ LLAMAFILE_DEPS = \
 	$(LLAMAFILE_HIGHLIGHT_KEYWORDS) \
 	$(LLAMAFILE_METAL_SOURCES) \
 	$(TINYBLAS_CPU_OBJS) \
-	o/$(MODE)/third_party/stb/stb_image_resize2.o
+	o/$(MODE)/third_party/stb/stb_image_resize2.o \
+	o/$(MODE)/third_party/mbedtls/mbedtls.a
 
 # ==============================================================================
 # Server integration
 # ==============================================================================
 
-# Include paths needed for server compilation
+# Include paths needed for server compilation. server.cpp reaches
+# httplib.h via server-cors-proxy.h, so it needs the httplib TLS flags
+# (see LLAMAFILE_HTTPLIB_TLS_FLAGS above).
 LLAMAFILE_SERVER_INCS := \
 	$(LLAMAFILE_INCLUDES) \
 	-iquote llama.cpp/tools/server \
-	-iquote o/$(MODE)/llama.cpp/tools/server
+	-iquote o/$(MODE)/llama.cpp/tools/server \
+	$(LLAMAFILE_HTTPLIB_TLS_FLAGS)
 
 # Compile server.cpp
 o/$(MODE)/llamafile/server.cpp.o: llama.cpp/tools/server/server.cpp
@@ -317,7 +335,7 @@ o/$(MODE)/llamafile/llamafile: \
 		$(LLAMAFILE_OBJS) \
 		$(LLAMAFILE_DEPS)
 	@mkdir -p $(@D)
-	$(CXX) $(LDFLAGS) -o $@ $(filter %.o,$^) $(LDLIBS)
+	$(CXX) $(LDFLAGS) -o $@ $(filter %.o %.a,$^) $(LDLIBS)
 
 # ==============================================================================
 # Pattern rules for llamafile sources
@@ -354,6 +372,13 @@ LLAMAFILE_GPU_OBJS := \
 
 o/$(MODE)/llamafile/gpu.a: $(LLAMAFILE_GPU_OBJS)
 
+# The TUI chatbot talks to the server through httplib as an HTTP client,
+# so its objects must agree with httplib.cpp on the httplib class layouts
+# (see LLAMAFILE_HTTPLIB_TLS_FLAGS).
+o/$(MODE)/llamafile/chatbot_api.o \
+o/$(MODE)/llamafile/chatbot_main.o: private \
+	LLAMAFILE_CPPFLAGS += $(LLAMAFILE_HTTPLIB_TLS_FLAGS)
+
 o/$(MODE)/llamafile/%.o: llamafile/%.cpp
 	@mkdir -p $(@D)
 	$(CXX) $(CXXFLAGS) $(LLAMAFILE_CPPFLAGS) -c -o $@ $<

llamafile/args.cpp

@@ -57,6 +57,11 @@ LlamafileArgs parse_llamafile_args(int argc, char** argv) {
         if ((strcmp(argv[i], "-m") == 0 || strcmp(argv[i], "--model") == 0) && i + 1 < argc) {
             args.model_path = argv[i + 1];
         }
+        if ((strcmp(argv[i], "-hf") == 0 || strcmp(argv[i], "-hfr") == 0 ||
+             strcmp(argv[i], "--hf-repo") == 0 || strcmp(argv[i], "-mu") == 0 ||
+             strcmp(argv[i], "--model-url") == 0) && i + 1 < argc) {
+            args.remote_model = argv[i + 1];
+        }
     }
 
     // Determine execution mode from flags

llamafile/args.h

@@ -45,6 +45,11 @@ struct LlamafileArgs {
     // Model path captured from -m (for display in combined mode TUI)
     std::string model_path;
 
+    // Remote model reference captured from -hf/--hf-repo/-mu/--model-url.
+    // llama.cpp downloads it over HTTPS at load time, so -m is not required
+    // when one of these is present; also used as the TUI display fallback.
+    std::string remote_model;
+
     // Note: Llamafile-specific flags are stored in FLAG_* globals (llamafile.h):
     //   --verbose  -> FLAG_verbose
     //   --nothink  -> FLAG_nothink

llamafile/main.cpp

@@ -240,7 +240,8 @@ static int combined_main(const LlamafileArgs &args) {
         // stack is too small for the httplib + SSE + JSON parsing call chain)
         auto *ctx = new TuiThreadCtx{
             &shutdown_fn, &mu, &cv, &shutdown_ready,
-            listen_addr, args.system_prompt, args.model_path
+            listen_addr, args.system_prompt,
+            args.model_path.empty() ? args.remote_model : args.model_path
         };
 
         pthread_attr_t attr;
@@ -314,8 +315,10 @@ int main(int argc, char **argv) {
         }
     }
 
-    // All modes require a model file (but let --server --help pass through).
-    if (args.model_path.empty() &&
+    // All modes require a model: either a local file (-m) or a remote
+    // reference (-hf/--model-url) that llama.cpp downloads over HTTPS.
+    // --server --help still passes through.
+    if (args.model_path.empty() && args.remote_model.empty() &&
         !llamafile_has(argv, "--help") && !llamafile_has(argv, "-h")) {
         fprintf(stderr, "error: missing required -m MODEL.gguf\n\n");
         switch (args.mode) {