Add switchable draft LoRA variants for Laguna

server/src/common/backend_factory.cpp

@@ -146,6 +146,7 @@ std::unique_ptr<ModelBackend> create_backend(const BackendArgs & args) {
         LagunaBackendArgs lcfg;
         lcfg.target_path = args.model_path;
         lcfg.draft_path  = args.draft_path ? args.draft_path : "";
+        lcfg.draft_loras = args.draft_loras;
         lcfg.draft_gpu   = args.draft_device.gpu;
         lcfg.draft_ctx_max = args.draft_ctx_max;
         lcfg.ddtree_mode = args.ddtree_mode;

server/src/common/backend_factory.h

@@ -11,12 +11,14 @@
 #pragma once
 
 #include "model_backend.h"
+#include "internal.h"
 #include "placement/placement_config.h"
 #include "placement/remote_draft_config.h"
 #include "placement/remote_target_shard_config.h"
 
 #include <memory>
 #include <string>
+#include <vector>
 
 namespace dflash::common {
 
@@ -29,6 +31,7 @@ struct BackendArgs {
 
     // Optional: speculative decode draft model (qwen35 only)
     const char *    draft_path   = nullptr;
+    std::vector<DraftLoraSpec> draft_loras;
 
     // Device placement
     DevicePlacement device;

server/src/common/model_backend.h

@@ -124,6 +124,9 @@ struct GenerateRequest {
     // path returns success but emits no tokens, so each backend can route the
     // retry through its existing AR path without copying retry policy.
     bool                       force_ar_decode = false;
+    // Optional request-selected DFlash draft LoRA variant. Empty uses the
+    // backend default; "base" selects the unadapted draft when available.
+    std::string                draft_lora;
 };
 
 struct GenerateResult {