feat(dflash): VRAM management improvements for C++ server

dflash/src/common/model_backend.h

@@ -174,6 +174,10 @@ struct ModelBackend {
     // supports_dflash_spec_decode() returns true. Default returns nullptr.
     virtual class DFlashTarget * dflash_target() { return nullptr; }
 
+    // Release oversized scratch buffers between requests to prevent VRAM
+    // growth over time. Default is a no-op.
+    virtual void release_scratch() {}
+
     // ── Cleanup ──────────────────────────────────────────────────────
     // Release all resources (weights, cache, snapshots, drafter).
     // Called by run_daemon() before returning.

dflash/src/qwen35/qwen35_backend.cpp

@@ -13,6 +13,8 @@
 
 #include "ggml-cuda.h"
 #include "common/snapshot_backend.h"
+#include "pflash_ggml_adapter.h"
+#include "flashprefill.h"
 
 #include <algorithm>
 #include <chrono>
@@ -436,6 +438,32 @@ void Qwen35Backend::shutdown() {
     }
 }
 
+// ── Release scratch buffers between requests ────────────────────────────
+
+void Qwen35Backend::release_scratch() {
+    // Target graph allocator: grows during large prefill batches, not needed
+    // between requests. Will be lazily recreated on next build_target_step().
+    if (sg_.alloc) {
+        ggml_gallocr_free(sg_.alloc);
+        sg_.alloc = nullptr;
+    }
+    step_graph_free(sg_);
+
+    // LM-head projection allocator (same pattern).
+    if (proj_sg_.alloc) {
+        ggml_gallocr_free(proj_sg_.alloc);
+        proj_sg_.alloc = nullptr;
+    }
+    step_graph_free(proj_sg_);
+
+    // BSA persistent CUDA buffers (blockmask, head_mask_type, softmax_lse).
+#ifdef DFLASH27B_HAVE_BSA
+    flashprefill::dflash_bsa_free_persistent();
+#endif
+
+    std::fprintf(stderr, "[vram] released scratch buffers\n");
+}
+
 // ── Generate (speculative decode) ───────────────────────────────────────
 
 GenerateResult Qwen35Backend::generate(const GenerateRequest & req,

dflash/src/qwen35/qwen35_backend.h

@@ -109,6 +109,10 @@ class Qwen35Backend : public ModelBackend {
 
     void shutdown() override;
 
+    // Release oversized scratch buffers (gallocr, BSA cache) between requests
+    // to prevent VRAM growth over time.
+    void release_scratch() override;
+
 private:
     // ── Configuration ────────────────────────────────────────────────
     Qwen35Config cfg_;

dflash/src/server/http_server.cpp

@@ -712,13 +712,28 @@ void HttpServer::worker_loop() {
         };
 
         // Run generation (with or without restore).
+        // Lazy-draft: ensure decode draft is loaded before generate.
+        if (config_.lazy_draft) {
+            backend_.free_drafter();    // free pflash drafter (~1.4 GB) if loaded
+            backend_.unpark("draft");   // reload decode draft (~3.3 GB)
+        }
+
         GenerateResult result;
         if (using_restore) {
             result = backend_.restore_and_generate(cache_slot, gen_req, io);
         } else {
             result = backend_.generate(gen_req, io);
         }
 
+        // Lazy-draft: park decode draft after generate to free VRAM.
+        if (config_.lazy_draft) {
+            backend_.park("draft");
+        }
+
+        // Release oversized scratch buffers (gallocr, BSA cache) so VRAM
+        // doesn't grow monotonically across requests with different sizes.
+        backend_.release_scratch();
+
         // Confirm or abort the inline snapshot.
         if (snap_prepared) {
             if (completion_tokens > 0 && !client_disconnected) {

dflash/src/server/http_server.h

@@ -56,6 +56,7 @@ struct ServerConfig {
     float       pflash_keep_ratio = 0.05f;  // fraction of tokens to keep
     std::string pflash_drafter_path;        // path to drafter GGUF (Qwen3-0.6B)
     bool        pflash_skip_park = false;   // skip park/unpark for ≥32GB GPUs
+    bool        lazy_draft      = false;   // park decode draft when idle to save VRAM
 
     // Disk prefix cache
     std::string disk_cache_dir;             // empty = disabled

dflash/src/server/server_main.cpp

@@ -68,6 +68,7 @@ static void print_usage(const char * prog) {
         "  --prefill-keep-ratio <F>    Fraction of tokens to keep (default: 0.05)\n"
         "  --prefill-drafter <path>    Drafter GGUF for compression (Qwen3-0.6B)\n"
         "  --prefill-skip-park         Skip park/unpark (for >=32GB GPUs)\n"
+        "  --lazy-draft                Park decode draft when idle to save VRAM\n"
         "\n"
         "Disk KV cache:\n"
         "  --kv-cache-dir <path>       Directory for ondisk KV cache (enables feature)\n"
@@ -140,6 +141,8 @@ int main(int argc, char ** argv) {
             sconfig.pflash_drafter_path = argv[++i];
         } else if (std::strcmp(argv[i], "--prefill-skip-park") == 0) {
             sconfig.pflash_skip_park = true;
+        } else if (std::strcmp(argv[i], "--lazy-draft") == 0) {
+            sconfig.lazy_draft = true;
         } else if (std::strcmp(argv[i], "--kv-cache-dir") == 0 && i + 1 < argc) {
             sconfig.disk_cache_dir = argv[++i];
         } else if (std::strcmp(argv[i], "--kv-cache-budget") == 0 && i + 1 < argc) {
@@ -194,6 +197,12 @@ int main(int argc, char ** argv) {
         setenv("DFLASH27B_FA_WINDOW", "0", 0);
     }
 
+    // Lazy-draft requires both prefill-drafter AND decode draft to be useful.
+    if (sconfig.lazy_draft && !(pflash_enabled && bargs.draft_path)) {
+        std::fprintf(stderr, "[server] --lazy-draft ignored: requires both --prefill-drafter and --draft\n");
+        sconfig.lazy_draft = false;
+    }
+
     // Load tokenizer.
     std::fprintf(stderr, "[server] loading tokenizer from %s\n", bargs.model_path);
     Tokenizer tokenizer;
@@ -269,6 +278,9 @@ int main(int argc, char ** argv) {
     std::fprintf(stderr, "[server] │  fp_use_bsa      = %s\n", getenv("DFLASH_FP_USE_BSA") ? "ON" : "off");
     std::fprintf(stderr, "[server] │  fp_alpha        = %s\n", getenv("DFLASH_FP_ALPHA") ? getenv("DFLASH_FP_ALPHA") : "0.12 (default)");
     }
+    if (bargs.draft_path) {
+    std::fprintf(stderr, "[server] │  lazy_draft      = %s\n", sconfig.lazy_draft ? "ON" : "off");
+    }
     std::fprintf(stderr, "[server] ╰─────────────────────────────────────────────────────╯\n\n");
 
     HttpServer server(*backend, tokenizer, sconfig);
@@ -278,6 +290,12 @@ int main(int argc, char ** argv) {
     if (pflash_enabled) {
         server.set_drafter_tokenizer(&drafter_tokenizer);
     }
+
+    // Lazy-draft: park decode draft at startup to free VRAM (~3.3 GB).
+    if (sconfig.lazy_draft && bargs.draft_path) {
+        backend->park("draft");
+    }
+
     int ret = server.run();
 
     // Cleanup.