diff --git a/PRD_01_Device_MediaServer_RemoteMgmt.md b/PRD_01_Device_MediaServer_RemoteMgmt.md
index 0266bda..7d79418 100644
--- a/PRD_01_Device_MediaServer_RemoteMgmt.md
+++ b/PRD_01_Device_MediaServer_RemoteMgmt.md
@@ -158,12 +158,20 @@ Restart=always
 RestartSec=2
 LimitNOFILE=65535
 
+# （可选）性能调参：多路推理并发 / RGA 并发
+Environment=RK3588_RKNN_CTX_POOL_SIZE=3
+Environment=RK3588_RGA_MAX_INFLIGHT=2
+
 [Install]
 WantedBy=multi-user.target
 ```
 
 > V1 不要求 agent 控制 systemd，但要求两服务可开机自启。
 
+### 6.5 性能调参说明（新增）
+- `RK3588_RKNN_CTX_POOL_SIZE`：同一模型的 RKNN context 池大小（默认 3）。
+- `RK3588_RGA_MAX_INFLIGHT`：RGA 全局并发上限（默认 2）。
+
 ### 6.4 模型目录结构
 - `${models_dir}/manifest.json`
 - `${models_dir}/files/<name>__<sha256>.<ext>`（`ext` 默认 `rknn`，白名单控制）
diff --git a/PRD_04_Manager_ConfigGUI_AgentAPI.md b/PRD_04_Manager_ConfigGUI_AgentAPI.md
index ffc40ca..562eb2f 100644
--- a/PRD_04_Manager_ConfigGUI_AgentAPI.md
+++ b/PRD_04_Manager_ConfigGUI_AgentAPI.md
@@ -239,7 +239,28 @@ GUI 侧以 `instances[]` 为“通道列表”。
 
 ---
 
-## 8. 人脸库路径对齐建议（避免“上传了但识别不到”）
+## 8. 性能调优参数（可选，建议 GUI 以“高级设置”方式暴露）
+
+### 8.1 `preprocess`（AI 分支减拷贝 / RGA 并发）
+- `dst_packed`：bool，默认 `false`。
+  - 说明：当 `dst_format` 为 `rgb/bgr` 时，若开启则输出紧凑 packed（`stride = width * 3`），可避免下游（如 `ai_yolo`）对齐 stride 导致的逐行 memcpy。
+- `rga_max_inflight`：int，默认 `0`（表示不在配置里覆盖全局值）。
+  - 说明：RGA 全局并发上限（原先固定为 2）。
+  - 备注：也可通过环境变量 `RK3588_RGA_MAX_INFLIGHT` 设置。
+
+### 8.2 `publish/storage`（一次编码，多处复用）
+- `publish.attach_encoded_meta`：bool，默认 `true`（当 publish 有下游 output_queues 时）。
+  - 说明：将已编码视频包（含 codec extradata/pts/key）挂在 `frame.user_meta`，供下游复用（例如报警 clip、storage 复用码流）。
+- `storage.reuse_encoded_meta`：bool，默认 `true`。
+  - 说明：若收到 `EncodedVideoFrameMeta`，storage 将直接 remux 写文件（避免再次 MPP 编码）；若未收到则保持原逻辑（MPP 编码录制）。
+
+### 8.3 推理并发（环境变量，部署侧配置）
+- `RK3588_RKNN_CTX_POOL_SIZE`：默认 `3`。
+  - 说明：同一模型创建的 RKNN context 数量（多路同模型推理可并发，避免被单 context 串行化）。
+
+---
+
+## 9. 人脸库路径对齐建议（避免“上传了但识别不到”）
 
 1) 默认推荐（最省事）：
 - media-server work_dir：`/opt/rk3588sys`
@@ -251,7 +272,7 @@ GUI 侧以 `instances[]` 为“通道列表”。
 
 ---
 
-## 9. 验收标准
+## 10. 验收标准
 
 1) GUI 新增通道 → apply 成功后：`GET /v1/graphs` 出现该通道；`GET /v1/graphs/{name}` 能看到节点链路。
 2) 删除通道 → apply 成功后 graphs 中消失。
diff --git a/include/ai_scheduler.h b/include/ai_scheduler.h
index b1a07ce..0450537 100644
--- a/include/ai_scheduler.h
+++ b/include/ai_scheduler.h
@@ -93,17 +93,16 @@ public:
 #endif
     };
 
-    // Borrowed inference avoids per-call output allocations/copies by using per-model preallocated buffers.
-    // The returned result holds the per-model inference lock; it must be destroyed before the next InferBorrowed()
-    // on the same model handle.
+    // Borrowed inference avoids per-call output allocations/copies by using per-context preallocated buffers.
+    // The returned result holds the selected context's inference lock until it is destroyed.
     struct BorrowedInferResult {
         bool success = false;
         std::string error;
         std::vector<BorrowedOutput> outputs;
 
 #if defined(RK3588_ENABLE_RKNN)
-        std::shared_ptr<void> keepalive;               // keeps ModelContext alive
-        std::unique_lock<std::mutex> infer_lock;       // holds ModelContext::infer_mutex
+        std::shared_ptr<void> keepalive;               // keeps the selected ModelContext alive
+        std::unique_lock<std::mutex> infer_lock;       // holds ModelContext::infer_mutex for that context
 #endif
 
         BorrowedInferResult() = default;
@@ -133,7 +132,8 @@ private:
 #if defined(RK3588_ENABLE_RKNN)
     struct ModelContext {
         rknn_context ctx = 0;
-        std::vector<uint8_t> model_data;
+        // Shared model blob kept alive for RKNN runtime. Multiple contexts can share the same data.
+        std::shared_ptr<std::vector<uint8_t>> model_data;
         std::vector<rknn_tensor_attr> input_attrs;
         std::vector<rknn_tensor_attr> output_attrs;
         std::vector<std::vector<uint8_t>> output_buffers;  // preallocated output buffers
@@ -143,7 +143,7 @@ private:
         int input_h = 0;
         int input_c = 0;
         std::string path;
-        std::mutex infer_mutex;  // Per-model lock for inference
+        std::mutex infer_mutex;  // Per-context lock for inference
 
         ~ModelContext() {
             if (ctx) {
@@ -153,8 +153,14 @@ private:
         }
     };
 
-    std::unordered_map<ModelHandle, std::shared_ptr<ModelContext>> models_by_handle_;
-    std::unordered_map<std::string, std::weak_ptr<ModelContext>> models_by_path_;
+    struct ModelGroup {
+        std::string path;
+        std::vector<std::shared_ptr<ModelContext>> contexts;
+        std::atomic<uint32_t> rr{0};  // round-robin context selection
+    };
+
+    std::unordered_map<ModelHandle, std::shared_ptr<ModelGroup>> models_by_handle_;
+    std::unordered_map<std::string, std::weak_ptr<ModelGroup>> models_by_path_;
 #endif
 
     mutable std::mutex models_mutex_;  // Protects models_ map
diff --git a/plugins/preprocess/preprocess_node.cpp b/plugins/preprocess/preprocess_node.cpp
index c81a219..55c2094 100644
--- a/plugins/preprocess/preprocess_node.cpp
+++ b/plugins/preprocess/preprocess_node.cpp
@@ -2,6 +2,7 @@
 #include <atomic>
 #include <chrono>
 #include <condition_variable>
+#include <cstdlib>
 #include <cstring>
 #include <iostream>
 #include <memory>
@@ -43,9 +44,30 @@ int ToRgaFormat(PixelFormat fmt) {
     }
 }
 
-constexpr int GlobalRgaMaxInflight() {
-    // Fixed concurrency cap for RGA submissions.
-    return 2;
+std::atomic<int>& GlobalRgaMaxInflightRef() {
+    static std::atomic<int> v{2};
+    return v;
+}
+
+int GlobalRgaMaxInflight() {
+    static std::once_flag once;
+    std::call_once(once, []() {
+        const char* s = std::getenv("RK3588_RGA_MAX_INFLIGHT");
+        if (!s || !*s) return;
+        try {
+            int v = std::stoi(s);
+            if (v > 0 && v <= 32) GlobalRgaMaxInflightRef().store(v);
+        } catch (...) {
+        }
+    });
+    int v = GlobalRgaMaxInflightRef().load();
+    return v > 0 ? v : 1;
+}
+
+void SetGlobalRgaMaxInflight(int v) {
+    if (v <= 0) return;
+    if (v > 32) v = 32;
+    GlobalRgaMaxInflightRef().store(v);
 }
 
 class RgaGate {
@@ -254,10 +276,18 @@ public:
         dst_w_ = config.ValueOr<int>("dst_w", 640);
         dst_h_ = config.ValueOr<int>("dst_h", 640);
         keep_ratio_ = config.ValueOr<bool>("keep_ratio", false);
+        dst_packed_ = config.ValueOr<bool>("dst_packed", false);
         std::string fmt_str = config.ValueOr<std::string>("dst_format", "");
         if (!fmt_str.empty()) {
             dst_fmt_ = ParseFormat(fmt_str);
         }
+
+#if defined(RK3588_ENABLE_RGA)
+        const int rga_max_inflight = config.ValueOr<int>("rga_max_inflight", 0);
+        if (rga_max_inflight > 0) {
+            SetGlobalRgaMaxInflight(rga_max_inflight);
+        }
+#endif
         const bool requested_use_rga = config.ValueOr<bool>("use_rga", true);
         use_rga_ = requested_use_rga;
 
@@ -415,6 +445,11 @@ private:
             }
         }
         int dst_wstride = Align16(out_w);
+        // For AI input (RGB/BGR), allow a tightly packed output to avoid an extra per-frame memcpy
+        // in downstream nodes (e.g. ai_yolo).
+        if (dst_packed_ && (out_fmt == PixelFormat::RGB || out_fmt == PixelFormat::BGR)) {
+            dst_wstride = out_w;
+        }
         int dst_hstride = Align16(out_h);
 
         if (src_fmt_rga == RK_FORMAT_UNKNOWN || dst_fmt_rga == RK_FORMAT_UNKNOWN) {
@@ -802,6 +837,7 @@ private:
     int dst_h_ = 640;
     bool keep_ratio_ = false;
     PixelFormat dst_fmt_ = PixelFormat::UNKNOWN;
+    bool dst_packed_ = false;
     bool use_rga_ = true;
 
     bool stats_log_ = false;
diff --git a/plugins/storage/storage_node.cpp b/plugins/storage/storage_node.cpp
index 1ea1856..9301619 100644
--- a/plugins/storage/storage_node.cpp
+++ b/plugins/storage/storage_node.cpp
@@ -1,3 +1,4 @@
+#include <algorithm>
 #include <atomic>
 #include <chrono>
 #include <cstring>
@@ -12,6 +13,7 @@
 #include <mutex>
 
 #include "node.h"
+#include "media/encoded_video_meta.h"
 
 #if defined(RK3588_ENABLE_FFMPEG)
 extern "C" {
@@ -42,6 +44,118 @@ namespace rk3588 {
 namespace {
 inline int Align16(int v) { return (v + 15) & ~15; }
 
+std::shared_ptr<EncodedVideoFrameMeta> TryGetEncodedMeta(const FramePtr& frame) {
+    if (!frame || !frame->user_meta) return nullptr;
+    auto meta = std::static_pointer_cast<EncodedVideoFrameMeta>(frame->user_meta);
+    if (!meta || meta->magic != EncodedVideoFrameMeta::kMagic) return nullptr;
+    if (!meta->codec || meta->pkt.data.empty() || meta->pkt.pts_ms <= 0) return nullptr;
+    return meta;
+}
+
+#if HAS_FFMPEG
+static bool HasAnnexBStartCode(const uint8_t* d, size_t n) {
+    if (!d || n < 4) return false;
+    for (size_t i = 0; i + 3 < n; ++i) {
+        if (d[i] == 0 && d[i + 1] == 0 && d[i + 2] == 1) return true;
+        if (i + 4 < n && d[i] == 0 && d[i + 1] == 0 && d[i + 2] == 0 && d[i + 3] == 1) return true;
+    }
+    return false;
+}
+
+static size_t FindStartCode(const uint8_t* d, size_t n, size_t from, size_t* sc_len) {
+    for (size_t i = from; i + 3 < n; ++i) {
+        if (d[i] == 0 && d[i + 1] == 0 && d[i + 2] == 1) {
+            if (sc_len) *sc_len = 3;
+            return i;
+        }
+        if (i + 4 < n && d[i] == 0 && d[i + 1] == 0 && d[i + 2] == 0 && d[i + 3] == 1) {
+            if (sc_len) *sc_len = 4;
+            return i;
+        }
+    }
+    if (sc_len) *sc_len = 0;
+    return n;
+}
+
+static std::vector<std::vector<uint8_t>> SplitAnnexBNals(const uint8_t* d, size_t n) {
+    std::vector<std::vector<uint8_t>> out;
+    if (!d || n < 4) return out;
+    size_t pos = 0;
+    while (true) {
+        size_t sc_len = 0;
+        size_t sc = FindStartCode(d, n, pos, &sc_len);
+        if (sc >= n) break;
+        size_t nal_start = sc + sc_len;
+        size_t next_sc = FindStartCode(d, n, nal_start, nullptr);
+        size_t nal_end = (next_sc >= n) ? n : next_sc;
+
+        while (nal_end > nal_start && d[nal_end - 1] == 0) --nal_end;
+        if (nal_end > nal_start) {
+            out.emplace_back(d + nal_start, d + nal_end);
+        }
+        pos = nal_end;
+    }
+    return out;
+}
+
+static bool BuildAvccFromAnnexB(const std::vector<uint8_t>& annexb, std::vector<uint8_t>& avcc_out) {
+    avcc_out.clear();
+    if (annexb.empty()) return false;
+    auto nals = SplitAnnexBNals(annexb.data(), annexb.size());
+
+    const uint8_t* sps = nullptr;
+    size_t sps_len = 0;
+    const uint8_t* pps = nullptr;
+    size_t pps_len = 0;
+
+    for (const auto& nal : nals) {
+        if (nal.empty()) continue;
+        const uint8_t t = nal[0] & 0x1F;
+        if (t == 7 && !sps) {
+            sps = nal.data();
+            sps_len = nal.size();
+        } else if (t == 8 && !pps) {
+            pps = nal.data();
+            pps_len = nal.size();
+        }
+    }
+    if (!sps || sps_len < 4 || !pps || pps_len < 1) return false;
+
+    avcc_out.reserve(11 + sps_len + pps_len);
+    avcc_out.push_back(1);
+    avcc_out.push_back(sps[1]);
+    avcc_out.push_back(sps[2]);
+    avcc_out.push_back(sps[3]);
+    avcc_out.push_back(0xFF);
+    avcc_out.push_back(0xE1);
+    avcc_out.push_back(static_cast<uint8_t>((sps_len >> 8) & 0xFF));
+    avcc_out.push_back(static_cast<uint8_t>((sps_len) & 0xFF));
+    avcc_out.insert(avcc_out.end(), sps, sps + sps_len);
+    avcc_out.push_back(1);
+    avcc_out.push_back(static_cast<uint8_t>((pps_len >> 8) & 0xFF));
+    avcc_out.push_back(static_cast<uint8_t>((pps_len) & 0xFF));
+    avcc_out.insert(avcc_out.end(), pps, pps + pps_len);
+    return true;
+}
+
+static std::vector<uint8_t> AnnexBToLengthPrefixed(const uint8_t* d, size_t n) {
+    std::vector<uint8_t> out;
+    auto nals = SplitAnnexBNals(d, n);
+    size_t total = 0;
+    for (const auto& nal : nals) total += 4 + nal.size();
+    out.reserve(total);
+    for (const auto& nal : nals) {
+        uint32_t len = static_cast<uint32_t>(nal.size());
+        out.push_back(static_cast<uint8_t>((len >> 24) & 0xFF));
+        out.push_back(static_cast<uint8_t>((len >> 16) & 0xFF));
+        out.push_back(static_cast<uint8_t>((len >> 8) & 0xFF));
+        out.push_back(static_cast<uint8_t>((len) & 0xFF));
+        out.insert(out.end(), nal.begin(), nal.end());
+    }
+    return out;
+}
+#endif
+
 std::string FormatTime(const std::string& pattern) {
     auto now = std::chrono::system_clock::now();
     auto time_t_now = std::chrono::system_clock::to_time_t(now);
@@ -264,6 +378,7 @@ public:
         filename_pattern_ = config.ValueOr<std::string>("filename_pattern", "%Y%m%d/%H%M%S");
         fps_ = config.ValueOr<int>("fps", 25);
         bitrate_kbps_ = config.ValueOr<int>("bitrate_kbps", 2000);
+        reuse_encoded_meta_ = config.ValueOr<bool>("reuse_encoded_meta", true);
 
         input_queue_ = ctx.input_queue;
         if (!input_queue_) {
@@ -373,6 +488,11 @@ private:
 
     void OpenNewFile(FramePtr frame) {
 #if HAS_FFMPEG
+        auto meta = reuse_encoded_meta_ ? TryGetEncodedMeta(frame) : nullptr;
+        using_encoded_meta_ = (meta != nullptr);
+        segment_base_pts_ms_ = using_encoded_meta_ ? meta->pkt.pts_ms : 0;
+        last_pts90k_ = -1;
+
         std::string filename = FormatTime(filename_pattern_) + "." + format_;
         current_path_ = base_path_ + "/" + filename;
 
@@ -390,41 +510,78 @@ private:
             return;
         }
 
-        AVCodecID codec_id = (codec_ == "h265" || codec_ == "hevc")
-                                 ? AV_CODEC_ID_HEVC : AV_CODEC_ID_H264;
-        const AVCodec* encoder = avcodec_find_encoder(codec_id);
-        stream_ = avformat_new_stream(fmt_ctx_, encoder);
+        AVCodecID codec_id = AV_CODEC_ID_H264;
+        if (using_encoded_meta_ && meta && meta->codec) {
+            codec_id = (meta->codec->codec == VideoCodec::H265) ? AV_CODEC_ID_HEVC : AV_CODEC_ID_H264;
+        } else {
+            codec_id = (codec_ == "h265" || codec_ == "hevc") ? AV_CODEC_ID_HEVC : AV_CODEC_ID_H264;
+        }
+
+        stream_ = avformat_new_stream(fmt_ctx_, nullptr);
         if (!stream_) {
             avformat_free_context(fmt_ctx_);
             fmt_ctx_ = nullptr;
             return;
         }
 
-        stream_->time_base = AVRational{1, fps_};
+        if (using_encoded_meta_) {
+            stream_->time_base = AVRational{1, 90000};
+            const int fps_eff = (meta && meta->codec && meta->codec->fps > 0) ? meta->codec->fps : fps_;
+            stream_->avg_frame_rate = AVRational{std::max(1, fps_eff), 1};
+            stream_->r_frame_rate = stream_->avg_frame_rate;
+        } else {
+            stream_->time_base = AVRational{1, fps_};
+        }
         stream_->codecpar->codec_type = AVMEDIA_TYPE_VIDEO;
         stream_->codecpar->codec_id = codec_id;
-        stream_->codecpar->width = frame->width;
-        stream_->codecpar->height = frame->height;
+        stream_->codecpar->width = (using_encoded_meta_ && meta && meta->codec && meta->codec->width > 0)
+                                       ? meta->codec->width
+                                       : frame->width;
+        stream_->codecpar->height = (using_encoded_meta_ && meta && meta->codec && meta->codec->height > 0)
+                                        ? meta->codec->height
+                                        : frame->height;
         stream_->codecpar->format = AV_PIX_FMT_YUV420P;
 
 #if HAS_MPP
-        if (!mpp_encoder_) {
-            mpp_encoder_ = std::make_unique<MppStorageEncoder>();
-            if (!mpp_encoder_->Init(frame->width, frame->height, frame->format,
-                                    codec_, fps_, bitrate_kbps_)) {
-                mpp_encoder_.reset();
+        if (!using_encoded_meta_) {
+            if (!mpp_encoder_) {
+                mpp_encoder_ = std::make_unique<MppStorageEncoder>();
+                if (!mpp_encoder_->Init(frame->width, frame->height, frame->format,
+                                        codec_, fps_, bitrate_kbps_)) {
+                    mpp_encoder_.reset();
+                }
+            }
+
+            if (mpp_encoder_ && !mpp_encoder_->Header().empty()) {
+                const auto& hdr = mpp_encoder_->Header();
+                stream_->codecpar->extradata_size = static_cast<int>(hdr.size());
+                stream_->codecpar->extradata = static_cast<uint8_t*>(
+                    av_mallocz(hdr.size() + AV_INPUT_BUFFER_PADDING_SIZE));
+                std::memcpy(stream_->codecpar->extradata, hdr.data(), hdr.size());
             }
         }
-
-        if (mpp_encoder_ && !mpp_encoder_->Header().empty()) {
-            const auto& hdr = mpp_encoder_->Header();
-            stream_->codecpar->extradata_size = static_cast<int>(hdr.size());
-            stream_->codecpar->extradata = static_cast<uint8_t*>(
-                av_mallocz(hdr.size() + AV_INPUT_BUFFER_PADDING_SIZE));
-            std::memcpy(stream_->codecpar->extradata, hdr.data(), hdr.size());
-        }
 #endif
 
+        if (using_encoded_meta_ && meta && meta->codec && !meta->codec->extradata.empty()) {
+            std::vector<uint8_t> ex = meta->codec->extradata;
+            const bool is_mp4 = (format_ != "ts");
+
+            // MP4 prefers AVCC/HVCC. Do a best-effort conversion for H264 AnnexB extradata.
+            if (is_mp4 && codec_id == AV_CODEC_ID_H264) {
+                if (!(ex.size() >= 1 && ex[0] == 1) && HasAnnexBStartCode(ex.data(), ex.size())) {
+                    std::vector<uint8_t> avcc;
+                    if (BuildAvccFromAnnexB(ex, avcc)) {
+                        ex = std::move(avcc);
+                    }
+                }
+            }
+
+            stream_->codecpar->extradata_size = static_cast<int>(ex.size());
+            stream_->codecpar->extradata = static_cast<uint8_t*>(
+                av_mallocz(ex.size() + AV_INPUT_BUFFER_PADDING_SIZE));
+            std::memcpy(stream_->codecpar->extradata, ex.data(), ex.size());
+        }
+
         if (!(fmt_ctx_->oformat->flags & AVFMT_NOFILE)) {
             if (avio_open(&fmt_ctx_->pb, current_path_.c_str(), AVIO_FLAG_WRITE) < 0) {
                 avformat_free_context(fmt_ctx_);
@@ -450,6 +607,52 @@ private:
     }
 
     void WriteFrame(FramePtr frame) {
+#if HAS_FFMPEG
+        if (!file_open_ || !fmt_ctx_ || !stream_) return;
+
+        if (using_encoded_meta_) {
+            auto meta = TryGetEncodedMeta(frame);
+            if (!meta || !meta->codec) return;
+
+            const bool is_mp4 = (format_ != "ts");
+            std::vector<uint8_t> sample = meta->pkt.data;
+
+            if (is_mp4 && stream_->codecpar->codec_id == AV_CODEC_ID_H264 &&
+                HasAnnexBStartCode(sample.data(), sample.size())) {
+                sample = AnnexBToLengthPrefixed(sample.data(), sample.size());
+            }
+            if (sample.empty()) return;
+
+            const int fps_eff = (meta->codec->fps > 0) ? meta->codec->fps : fps_;
+            const int64_t frame_dur = fps_eff > 0 ? (90000 / std::max(1, fps_eff)) : 0;
+
+            int64_t rel_ms = meta->pkt.pts_ms - segment_base_pts_ms_;
+            if (rel_ms < 0) rel_ms = 0;
+            int64_t pts90k = rel_ms * 90;
+            if (last_pts90k_ >= 0 && pts90k <= last_pts90k_) {
+                pts90k = last_pts90k_ + (frame_dur > 0 ? frame_dur : 1);
+            }
+            last_pts90k_ = pts90k;
+
+            AVPacket* pkt = av_packet_alloc();
+            if (!pkt) return;
+            if (av_new_packet(pkt, static_cast<int>(sample.size())) < 0) {
+                av_packet_free(&pkt);
+                return;
+            }
+            std::memcpy(pkt->data, sample.data(), sample.size());
+            pkt->stream_index = stream_->index;
+            pkt->flags = meta->pkt.key ? AV_PKT_FLAG_KEY : 0;
+            pkt->pts = pts90k;
+            pkt->dts = pts90k;
+            pkt->duration = frame_dur;
+            (void)av_interleaved_write_frame(fmt_ctx_, pkt);
+            av_packet_free(&pkt);
+            ++segment_frames_;
+            return;
+        }
+#endif
+
 #if HAS_FFMPEG && HAS_MPP
         if (!file_open_ || !fmt_ctx_ || !mpp_encoder_) return;
 
@@ -476,8 +679,7 @@ private:
         av_interleaved_write_frame(fmt_ctx_, pkt);
         av_packet_free(&pkt);
         ++segment_frames_;
-#elif HAS_FFMPEG
-        // Software encoding fallback would go here
+#else
         (void)frame;
 #endif
     }
@@ -494,6 +696,9 @@ private:
         fmt_ctx_ = nullptr;
         stream_ = nullptr;
         file_open_ = false;
+        using_encoded_meta_ = false;
+        segment_base_pts_ms_ = 0;
+        last_pts90k_ = -1;
 
         std::cout << "[storage] closed: " << current_path_
                   << " (" << segment_frames_ << " frames)\n";
@@ -509,6 +714,7 @@ private:
     std::string filename_pattern_;
     int fps_ = 25;
     int bitrate_kbps_ = 2000;
+    bool reuse_encoded_meta_ = true;
 
     std::mutex mu_;
 
@@ -520,6 +726,9 @@ private:
     std::string current_path_;
     std::chrono::steady_clock::time_point segment_start_;
     int64_t segment_frames_ = 0;
+    bool using_encoded_meta_ = false;
+    int64_t segment_base_pts_ms_ = 0;
+    int64_t last_pts90k_ = -1;
 
 #if HAS_FFMPEG
     AVFormatContext* fmt_ctx_ = nullptr;
diff --git a/src/ai_scheduler.cpp b/src/ai_scheduler.cpp
index bdf0646..d45f112 100644
--- a/src/ai_scheduler.cpp
+++ b/src/ai_scheduler.cpp
@@ -1,5 +1,6 @@
 #include "ai_scheduler.h"
 
+#include <cstdlib>
 #include <cstring>
 #include <fstream>
 
@@ -10,6 +11,30 @@ namespace rk3588 {
 #if defined(RK3588_ENABLE_RKNN)
 namespace {
 
+int GetEnvInt(const char* name, int default_value) {
+    if (!name) return default_value;
+    const char* v = std::getenv(name);
+    if (!v || !*v) return default_value;
+    try {
+        return std::stoi(v);
+    } catch (...) {
+        return default_value;
+    }
+}
+
+int ClampInt(int v, int lo, int hi) {
+    if (v < lo) return lo;
+    if (v > hi) return hi;
+    return v;
+}
+
+int DefaultContextPoolSize() {
+    // Default to 3 contexts to better utilize RK3588 NPU (3 cores).
+    // Can be overridden by env: RK3588_RKNN_CTX_POOL_SIZE.
+    const int v = GetEnvInt("RK3588_RKNN_CTX_POOL_SIZE", 3);
+    return ClampInt(v, 1, 16);
+}
+
 uint32_t TensorTypeSizeBytes(rknn_tensor_type t) {
     switch (t) {
         case RKNN_TENSOR_INT8:
@@ -83,98 +108,118 @@ ModelHandle AiScheduler::LoadModel(const std::string& model_path, std::string& e
     size_t model_size = file.tellg();
     file.seekg(0, std::ios::beg);
 
-    auto ctx = std::make_shared<ModelContext>();
-    ctx->model_data.resize(model_size);
-    ctx->path = model_path;
-
-    if (!file.read(reinterpret_cast<char*>(ctx->model_data.data()), model_size)) {
+    auto model_data = std::make_shared<std::vector<uint8_t>>();
+    model_data->resize(model_size);
+    if (!file.read(reinterpret_cast<char*>(model_data->data()), model_size)) {
         err = "Failed to read model file: " + model_path;
         return kInvalidModelHandle;
     }
 
-    // Initialize RKNN context
-    int ret = rknn_init(&ctx->ctx, ctx->model_data.data(), model_size, 0, nullptr);
-    if (ret < 0) {
-        err = "rknn_init failed with code: " + std::to_string(ret);
+    auto group = std::make_shared<ModelGroup>();
+    group->path = model_path;
+
+    const int pool_size = DefaultContextPoolSize();
+    group->contexts.reserve(static_cast<size_t>(pool_size));
+
+    for (int i = 0; i < pool_size; ++i) {
+        auto ctx = std::make_shared<ModelContext>();
+        ctx->model_data = model_data;
+        ctx->path = model_path;
+
+        int ret = rknn_init(&ctx->ctx, ctx->model_data->data(), model_size, 0, nullptr);
+        if (ret < 0) {
+            err = "rknn_init failed with code: " + std::to_string(ret);
+            return kInvalidModelHandle;
+        }
+
+        // If we create multiple contexts, bind them to different NPU cores when possible.
+        // This reduces contention and avoids the single-context serialization bottleneck.
+        {
+            int mask = RKNN_NPU_CORE_0_1_2;
+#if defined(RKNN_NPU_CORE_0) && defined(RKNN_NPU_CORE_1) && defined(RKNN_NPU_CORE_2)
+            if (pool_size >= 3) {
+                const int idx = i % 3;
+                mask = (idx == 0) ? RKNN_NPU_CORE_0 : (idx == 1 ? RKNN_NPU_CORE_1 : RKNN_NPU_CORE_2);
+            }
+#endif
+            ret = rknn_set_core_mask(ctx->ctx, mask);
+            if (ret < 0) {
+                LogWarn("[AiScheduler] rknn_set_core_mask failed: " + std::to_string(ret));
+            }
+        }
+
+        rknn_input_output_num io_num;
+        ret = rknn_query(ctx->ctx, RKNN_QUERY_IN_OUT_NUM, &io_num, sizeof(io_num));
+        if (ret < 0) {
+            err = "rknn_query IO num failed";
+            return kInvalidModelHandle;
+        }
+
+        ctx->n_input = io_num.n_input;
+        ctx->n_output = io_num.n_output;
+
+        ctx->input_attrs.resize(ctx->n_input);
+        for (uint32_t j = 0; j < ctx->n_input; ++j) {
+            ctx->input_attrs[j].index = j;
+            rknn_query(ctx->ctx, RKNN_QUERY_INPUT_ATTR, &ctx->input_attrs[j], sizeof(rknn_tensor_attr));
+        }
+
+        ctx->output_attrs.resize(ctx->n_output);
+        for (uint32_t j = 0; j < ctx->n_output; ++j) {
+            ctx->output_attrs[j].index = j;
+            rknn_query(ctx->ctx, RKNN_QUERY_OUTPUT_ATTR, &ctx->output_attrs[j], sizeof(rknn_tensor_attr));
+        }
+
+        ctx->output_buffers.resize(ctx->n_output);
+        for (uint32_t j = 0; j < ctx->n_output; ++j) {
+            uint32_t out_sz = ctx->output_attrs[j].size;
+            if (out_sz == 0 && ctx->output_attrs[j].size_with_stride > 0) {
+                out_sz = ctx->output_attrs[j].size_with_stride;
+            }
+            if (out_sz == 0 && ctx->output_attrs[j].n_elems > 0) {
+                out_sz = ctx->output_attrs[j].n_elems * TensorTypeSizeBytes(ctx->output_attrs[j].type);
+            }
+            if (out_sz > 0) {
+                ctx->output_buffers[j].resize(out_sz);
+            } else {
+                ctx->output_buffers[j].clear();
+            }
+        }
+
+        if (!ctx->input_attrs.empty()) {
+            if (ctx->input_attrs[0].fmt == RKNN_TENSOR_NCHW) {
+                ctx->input_c = ctx->input_attrs[0].dims[1];
+                ctx->input_h = ctx->input_attrs[0].dims[2];
+                ctx->input_w = ctx->input_attrs[0].dims[3];
+            } else {
+                ctx->input_h = ctx->input_attrs[0].dims[1];
+                ctx->input_w = ctx->input_attrs[0].dims[2];
+                ctx->input_c = ctx->input_attrs[0].dims[3];
+            }
+        }
+
+        group->contexts.push_back(ctx);
+    }
+
+    if (group->contexts.empty()) {
+        err = "No RKNN contexts created";
         return kInvalidModelHandle;
     }
 
-    // Prefer using all NPU cores (RK3588).
-    // This does not change the shared-context behavior; it only hints RKNN runtime scheduling.
-    ret = rknn_set_core_mask(ctx->ctx, RKNN_NPU_CORE_0_1_2);
-    if (ret < 0) {
-        LogWarn("[AiScheduler] rknn_set_core_mask failed: " + std::to_string(ret));
-    }
-
-    // Query input/output info
-    rknn_input_output_num io_num;
-    ret = rknn_query(ctx->ctx, RKNN_QUERY_IN_OUT_NUM, &io_num, sizeof(io_num));
-    if (ret < 0) {
-        err = "rknn_query IO num failed";
-        rknn_destroy(ctx->ctx);
-        ctx->ctx = 0;
-        return kInvalidModelHandle;
-    }
-
-    ctx->n_input = io_num.n_input;
-    ctx->n_output = io_num.n_output;
-
-    // Query input attributes
-    ctx->input_attrs.resize(ctx->n_input);
-    for (uint32_t i = 0; i < ctx->n_input; ++i) {
-        ctx->input_attrs[i].index = i;
-        rknn_query(ctx->ctx, RKNN_QUERY_INPUT_ATTR, &ctx->input_attrs[i], sizeof(rknn_tensor_attr));
-    }
-
-    // Query output attributes
-    ctx->output_attrs.resize(ctx->n_output);
-    for (uint32_t i = 0; i < ctx->n_output; ++i) {
-        ctx->output_attrs[i].index = i;
-        rknn_query(ctx->ctx, RKNN_QUERY_OUTPUT_ATTR, &ctx->output_attrs[i], sizeof(rknn_tensor_attr));
-    }
-
-    // Preallocate output buffers for borrowed inference.
-    ctx->output_buffers.resize(ctx->n_output);
-    for (uint32_t i = 0; i < ctx->n_output; ++i) {
-        uint32_t out_sz = ctx->output_attrs[i].size;
-        if (out_sz == 0 && ctx->output_attrs[i].size_with_stride > 0) {
-            out_sz = ctx->output_attrs[i].size_with_stride;
-        }
-        if (out_sz == 0 && ctx->output_attrs[i].n_elems > 0) {
-            out_sz = ctx->output_attrs[i].n_elems * TensorTypeSizeBytes(ctx->output_attrs[i].type);
-        }
-        if (out_sz > 0) {
-            ctx->output_buffers[i].resize(out_sz);
-        } else {
-            ctx->output_buffers[i].clear();
-        }
-    }
-
-    // Extract input dimensions
-    if (ctx->input_attrs[0].fmt == RKNN_TENSOR_NCHW) {
-        ctx->input_c = ctx->input_attrs[0].dims[1];
-        ctx->input_h = ctx->input_attrs[0].dims[2];
-        ctx->input_w = ctx->input_attrs[0].dims[3];
-    } else {
-        // NHWC
-        ctx->input_h = ctx->input_attrs[0].dims[1];
-        ctx->input_w = ctx->input_attrs[0].dims[2];
-        ctx->input_c = ctx->input_attrs[0].dims[3];
-    }
-
     ModelHandle handle = next_handle_.fetch_add(1);
-
     {
         std::lock_guard<std::mutex> lock(models_mutex_);
-        models_by_handle_[handle] = ctx;
-        models_by_path_[model_path] = ctx;
+        models_by_handle_[handle] = group;
+        models_by_path_[model_path] = group;
     }
 
+    const auto& first = group->contexts.front();
     LogInfo("[AiScheduler] loaded model: " + model_path +
             " (handle=" + std::to_string(handle) +
-            ", input=" + std::to_string(ctx->input_w) + "x" + std::to_string(ctx->input_h) +
-            "x" + std::to_string(ctx->input_c) +
-            ", outputs=" + std::to_string(ctx->n_output) + ")");
+            ", ctx_pool=" + std::to_string(group->contexts.size()) +
+            ", input=" + std::to_string(first->input_w) + "x" + std::to_string(first->input_h) +
+            "x" + std::to_string(first->input_c) +
+            ", outputs=" + std::to_string(first->n_output) + ")");
 
     return handle;
 #else
@@ -205,16 +250,19 @@ void AiScheduler::UnloadModel(ModelHandle handle) {
 
 bool AiScheduler::GetModelInfo(ModelHandle handle, ModelInfo& info) const {
 #if defined(RK3588_ENABLE_RKNN)
-    std::shared_ptr<ModelContext> ctx;
+    std::shared_ptr<ModelGroup> group;
     {
         std::lock_guard<std::mutex> lock(models_mutex_);
         auto it = models_by_handle_.find(handle);
         if (it == models_by_handle_.end() || !it->second) {
             return false;
         }
-        ctx = it->second;
+        group = it->second;
     }
 
+    if (!group || group->contexts.empty() || !group->contexts[0]) return false;
+    auto ctx = group->contexts[0];
+
     info.input_width = ctx->input_w;
     info.input_height = ctx->input_h;
     info.input_channels = ctx->input_c;
@@ -233,7 +281,7 @@ InferResult AiScheduler::Infer(ModelHandle handle, const InferInput& input) {
     InferResult result;
 
 #if defined(RK3588_ENABLE_RKNN)
-    std::shared_ptr<ModelContext> ctx;
+    std::shared_ptr<ModelGroup> group;
     {
         std::lock_guard<std::mutex> lock(models_mutex_);
         auto it = models_by_handle_.find(handle);
@@ -242,7 +290,20 @@ InferResult AiScheduler::Infer(ModelHandle handle, const InferInput& input) {
             total_errors_.fetch_add(1);
             return result;
         }
-        ctx = it->second;
+        group = it->second;
+    }
+
+    if (!group || group->contexts.empty()) {
+        result.error = "Invalid model context group";
+        total_errors_.fetch_add(1);
+        return result;
+    }
+
+    auto ctx = group->contexts[group->rr.fetch_add(1) % group->contexts.size()];
+    if (!ctx) {
+        result.error = "Invalid model context";
+        total_errors_.fetch_add(1);
+        return result;
     }
 
     // Lock this specific model for inference.
@@ -332,7 +393,7 @@ AiScheduler::BorrowedInferResult AiScheduler::InferBorrowed(ModelHandle handle,
     BorrowedInferResult result;
 
 #if defined(RK3588_ENABLE_RKNN)
-    std::shared_ptr<ModelContext> ctx;
+    std::shared_ptr<ModelGroup> group;
     {
         std::lock_guard<std::mutex> lock(models_mutex_);
         auto it = models_by_handle_.find(handle);
@@ -341,7 +402,20 @@ AiScheduler::BorrowedInferResult AiScheduler::InferBorrowed(ModelHandle handle,
             total_errors_.fetch_add(1);
             return result;
         }
-        ctx = it->second;
+        group = it->second;
+    }
+
+    if (!group || group->contexts.empty()) {
+        result.error = "Invalid model context group";
+        total_errors_.fetch_add(1);
+        return result;
+    }
+
+    auto ctx = group->contexts[group->rr.fetch_add(1) % group->contexts.size()];
+    if (!ctx) {
+        result.error = "Invalid model context";
+        total_errors_.fetch_add(1);
+        return result;
     }
 
     if (!input.data || input.size == 0) {