feat(server): add native mixed-backend draft placement

Author: weicj

dflash/CMakeLists.txt

@@ -671,6 +671,27 @@ if(DFLASH27B_TESTS)
         endif()
     endif()
 
+    if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/src/ipc/dflash_draft_ipc_main.cpp")
+        add_executable(dflash_draft_ipc_daemon
+            src/ipc/dflash_draft_ipc_main.cpp
+        )
+        target_include_directories(dflash_draft_ipc_daemon PRIVATE ${DFLASH27B_SRC_INCLUDE_DIRS})
+        if(DFLASH27B_GPU_BACKEND STREQUAL "hip")
+            target_compile_definitions(dflash_draft_ipc_daemon PRIVATE DFLASH27B_BACKEND_HIP=1 GGML_USE_HIP)
+        else()
+            target_compile_definitions(dflash_draft_ipc_daemon PRIVATE
+                DFLASH27B_BACKEND_CUDA=1
+                DFLASH27B_CUDA_MIN_SM=${_dflash_cuda_min_sm})
+        endif()
+        target_link_libraries(dflash_draft_ipc_daemon PRIVATE dflash_common ggml ${DFLASH27B_GGML_BACKEND_TARGET} pthread)
+        if(DFLASH27B_GPU_BACKEND STREQUAL "cuda")
+            find_package(CUDAToolkit REQUIRED)
+            target_link_libraries(dflash_draft_ipc_daemon PRIVATE CUDA::cudart)
+        else()
+            target_link_libraries(dflash_draft_ipc_daemon PRIVATE hip::host)
+        endif()
+    endif()
+
     # Tokenizer test harness (no GPU needed — links static lib for tokenizer + GGUF reader)
     if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/test/test_tokenizer_harness.cpp")
         add_executable(test_tokenizer_harness test/test_tokenizer_harness.cpp)

dflash/src/common/backend_factory.cpp

@@ -17,6 +17,10 @@ std::string detect_arch(const char * model_path) {
     return info.arch;
 }
 
+bool arch_supports_remote_draft(const std::string & arch) {
+    return arch == "qwen35";
+}
+
 std::unique_ptr<ModelBackend> create_backend(const BackendArgs & args) {
     if (!args.model_path) {
         std::fprintf(stderr, "[backend_factory] model_path is null\n");
@@ -38,6 +42,7 @@ std::unique_ptr<ModelBackend> create_backend(const BackendArgs & args) {
         cfg.draft_path         = args.draft_path;
         cfg.device             = args.device;
         cfg.draft_gpu          = args.draft_device.gpu;
+        cfg.remote_draft       = args.remote_draft;
         cfg.stream_fd          = args.stream_fd;
         cfg.fa_window          = args.fa_window;
         cfg.kq_stride_pad      = args.kq_stride_pad;

dflash/src/common/backend_factory.h

@@ -12,6 +12,7 @@
 
 #include "model_backend.h"
 #include "placement/placement_config.h"
+#include "placement/remote_draft_config.h"
 
 #include <memory>
 #include <string>
@@ -31,6 +32,7 @@ struct BackendArgs {
     // Device placement
     DevicePlacement device;
     DevicePlacement draft_device;
+    RemoteDraftConfig remote_draft;
 
     // I/O — only used when running under daemon_loop (legacy). The new
     // server passes -1 and uses on_token callbacks instead.
@@ -62,4 +64,6 @@ std::unique_ptr<ModelBackend> create_backend(const BackendArgs & args);
 // Useful for early dispatch (e.g. printing which backend will be used).
 std::string detect_arch(const char * model_path);
 
+bool arch_supports_remote_draft(const std::string & arch);
+
 }  // namespace dflash::common

dflash/src/common/model_backend.h

@@ -183,6 +183,12 @@ struct ModelBackend {
     // growth over time. Default is a no-op.
     virtual void release_scratch() {}
 
+    // Return true when the backend can route draft execution through the
+    // common remote-draft IPC transport. Model families that do not implement
+    // the DFlash feature boundary keep the default false and are rejected by
+    // the server before startup.
+    virtual bool supports_remote_draft() const { return false; }
+
     // ── Cleanup ──────────────────────────────────────────────────────
     // Release all resources (weights, cache, snapshots, drafter).
     // Called by run_daemon() before returning.

dflash/src/ipc/dflash_draft_ipc_main.cpp

@@ -0,0 +1,45 @@
+// Standalone DFlash draft IPC daemon entry point.
+
+#include "dflash_draft_ipc.h"
+
+#include <algorithm>
+#include <cstdio>
+#include <cstdlib>
+#include <cstring>
+
+using namespace dflash::common;
+
+int main(int argc, char ** argv) {
+    if (argc < 3 || std::strcmp(argv[1], "--draft-ipc-daemon") != 0) {
+        std::fprintf(stderr,
+            "usage: %s --draft-ipc-daemon <draft.safetensors|draft.gguf> "
+            "--ring-cap=N --stream-fd=FD [--draft-gpu=N]\n",
+            argv[0]);
+        return 2;
+    }
+
+    const char * draft_path = argv[2];
+    int ring_cap = 4096;
+    int draft_gpu = 0;
+    int stream_fd = -1;
+    for (int i = 3; i < argc; i++) {
+        if (std::strncmp(argv[i], "--ring-cap=", 11) == 0) {
+            ring_cap = std::atoi(argv[i] + 11);
+        } else if (std::strcmp(argv[i], "--ring-cap") == 0) {
+            if (i + 1 < argc) ring_cap = std::atoi(argv[++i]);
+        } else if (std::strncmp(argv[i], "--draft-gpu=", 12) == 0) {
+            draft_gpu = std::max(0, std::atoi(argv[i] + 12));
+        } else if (std::strcmp(argv[i], "--draft-gpu") == 0) {
+            if (i + 1 < argc) draft_gpu = std::max(0, std::atoi(argv[++i]));
+        } else if (std::strncmp(argv[i], "--stream-fd=", 12) == 0) {
+            stream_fd = std::atoi(argv[i] + 12);
+        } else if (std::strcmp(argv[i], "--stream-fd") == 0) {
+            if (i + 1 < argc) stream_fd = std::atoi(argv[++i]);
+        } else {
+            std::fprintf(stderr, "[draft-ipc-daemon] unknown option: %s\n", argv[i]);
+            return 2;
+        }
+    }
+
+    return run_dflash_draft_ipc_daemon(draft_path, ring_cap, draft_gpu, stream_fd);
+}

dflash/src/placement/remote_draft_config.h

@@ -0,0 +1,18 @@
+// Remote draft execution configuration for mixed-backend target/draft placement.
+
+#pragma once
+
+#include <string>
+
+namespace dflash::common {
+
+struct RemoteDraftConfig {
+    std::string ipc_bin;
+    std::string work_dir;
+    int ring_cap = 0;
+
+    bool enabled() const { return !ipc_bin.empty(); }
+    bool has_aux_options() const { return !work_dir.empty() || ring_cap > 0; }
+};
+
+}  // namespace dflash::common