diff --git a/hrx-integration/hrx-release.env b/hrx-integration/hrx-release.env
index bfe65999..dd36b56e 100644
--- a/hrx-integration/hrx-release.env
+++ b/hrx-integration/hrx-release.env
@@ -3,10 +3,10 @@
 
 # Pinned HRX (amdxdna) release consumed by the HRX runtime path.
 HRX_RELEASE_REPO="jtuyls/hrx"
-HRX_RELEASE_TAG="flm-hrx-amdxdna-v2026.07.30"
+HRX_RELEASE_TAG="flm-hrx-amdxdna-v2026.09.09"
 # For Linux.
-HRX_RELEASE_ASSET="hrx-amdxdna-2026.07.30-amdxdna-hal-native-rel-eb0b39f-linux-x86_64.tar.zst"
-HRX_RELEASE_SHA256="661ed94051cc6ad04f53739b2df7a791aecb658bc435bd5a6ff3c46716696345"
+HRX_RELEASE_ASSET="hrx-amdxdna-2026.09.09-amdxdna-hal-native-rel-effecac-linux-x86_64.tar.zst"
+HRX_RELEASE_SHA256="16103aa1996e71f27ea355f707707ce09fbfd094b2e2989fd017f819d438a7e6"
 # For Windows.
-HRX_RELEASE_ASSET_WINDOWS="hrx-amdxdna-2026.07.30-amdxdna-hal-native-rel-eb0b39f-windows-x86_64.zip"
-HRX_RELEASE_SHA256_WINDOWS="b3af74c9b393ce49ef71a3cbec2cbe92be6d40601858b41009d2bda57eadeb43"
+HRX_RELEASE_ASSET_WINDOWS="hrx-amdxdna-2026.09.09-amdxdna-hal-native-rel-effecac-windows-x86_64.zip"
+HRX_RELEASE_SHA256_WINDOWS="850a13affb927cdba5c0ba279681939c156154a28a917e5c2fb72cad38bf7cf2"
diff --git a/src/common/AutoModel/modeling_hunyuan.cpp b/src/common/AutoModel/modeling_hunyuan.cpp
index be81e1c8..8ef4936f 100644
--- a/src/common/AutoModel/modeling_hunyuan.cpp
+++ b/src/common/AutoModel/modeling_hunyuan.cpp
@@ -10,9 +10,12 @@
 
 /************              hunyuan-dense family            **************/
 Hunyuan::Hunyuan(flm_rt::device* npu_device_inst) : AutoModel(npu_device_inst, "Hunyuan") {
-    // the translator emits one short line per turn and the caller already has it
-    // from the stream / return value, so the raw dump would only double the log
-    this->log_raw_output = false;
+    // The translator streams one short line per turn, so echoing the
+    // "Model RAW Output:" dump only doubles the interactive log; keep it off by
+    // default. Diagnostics (qualification, numerical-match) anchor per-turn
+    // parsing on that marker, so FLM_LOG_RAW_OUTPUT=1 forces it back on for them
+    // without changing the interactive default.
+    this->log_raw_output = env_forces_raw_output();
     // every turn either rewinds to the pinned prefix or clears, so the kv state
     // the trailing eos forward exists to preserve is discarded either way
     this->forward_on_eos = false;
diff --git a/src/include/AutoModel/automodel.hpp b/src/include/AutoModel/automodel.hpp
index 0237e1cf..e8285311 100644
--- a/src/include/AutoModel/automodel.hpp
+++ b/src/include/AutoModel/automodel.hpp
@@ -16,6 +16,8 @@
 #include <string>
 #include <type_traits>
 #include <any>
+#include <cstdlib>
+#include <cstring>
 #include "typedef.hpp"
 #include "causal_lm.hpp"
 #include "lm_config.hpp"
@@ -150,8 +152,18 @@ class AutoModel {
     std::vector<int> checkpoint_his;
 	/// \brief dump the undecorated model output to stdout once a turn ends
 	/// \note on by default; models whose turns are short and driven in bulk (the
-	///       hunyuan translator) turn it off so the log is not doubled.
+	///       hunyuan translator) turn it off so the log is not doubled. Those
+	///       models should seed this from env_forces_raw_output() so diagnostics
+	///       can opt back in without changing the interactive default.
 	bool log_raw_output = true;
+	/// \brief whether FLM_LOG_RAW_OUTPUT forces the "Model RAW Output:" dump on
+	/// \note the qualification harness and numerical-match anchor per-turn parsing
+	///       on that marker, so they set FLM_LOG_RAW_OUTPUT=1 to re-enable it for
+	///       models that quiet it by default; unset/empty/"0" leave the default.
+	static bool env_forces_raw_output() {
+		const char* value = std::getenv("FLM_LOG_RAW_OUTPUT");
+		return value != nullptr && value[0] != '\0' && std::strcmp(value, "0") != 0;
+	}
 	/// \brief run one more forward on the eos token once a turn ends
 	/// \note this keeps the kv cache aligned with token_history so a following
 	///       turn can append to it. Models that rewind or clear between turns
diff --git a/src/include/buffer.hpp b/src/include/buffer.hpp
index 53c14c61..d44e99af 100644
--- a/src/include/buffer.hpp
+++ b/src/include/buffer.hpp
@@ -11,6 +11,7 @@
 #include <cstring>
 #include <fstream>
 #include <iostream>
+#include <memory>
 #include <stdexcept>
 #include <string>
 #include <vector>
@@ -30,15 +31,25 @@
 /// \note A copy (or mapping) does not duplicate the underlying memory; it only maps the pointer.
 class bytes {
 protected:
+    // HRX engines (IRON) use shared_ptr so copies keep the BO mapped. Unique_ptr
+    // copies drop ownership and leave data_ dangling across the .so boundary.
+#if defined(FLM_USE_HRX)
+    std::shared_ptr<uint8_t[]> owned_data_;
+#else
     std::unique_ptr<uint8_t[]> owned_data_;
+#endif
     uint8_t* data_;
     size_t size_;
     bool is_owner_;
 #ifdef FLM_DEVICE_BUFFER
     bool is_bo_owner_;
     flm_rt::bo* bo_;
+#if defined(FLM_USE_HRX)
+    std::shared_ptr<flm_rt::bo> owned_bo_;
+#else
     std::unique_ptr<flm_rt::bo> owned_bo_;
 #endif
+#endif
 
 public:
     /// \brief constructor
@@ -52,11 +63,19 @@ class bytes {
 
     /// \brief copy constructor
     /// \param other the other bytes
+#if defined(FLM_USE_HRX)
+    bytes(const bytes& other) : owned_data_(other.owned_data_), data_(other.data_), size_(other.size_), is_owner_(other.is_owner_)
+#ifdef FLM_DEVICE_BUFFER
+        , is_bo_owner_(other.is_bo_owner_), bo_(other.bo_), owned_bo_(other.owned_bo_)
+#endif
+    {}
+#else
     bytes(const bytes& other) : owned_data_(nullptr), data_(other.data_), size_(other.size_), is_owner_(false)
 #ifdef FLM_DEVICE_BUFFER
         , is_bo_owner_(false), bo_(other.bo_), owned_bo_(nullptr)
 #endif
     {}
+#endif
 
     /// \brief move constructor
     /// \param other the other bytes
@@ -86,7 +105,11 @@ class bytes {
     {
         if (size > 0 && size < 8ull * 1024 * 1024 * 1024){
             try {
+#if defined(FLM_USE_HRX)
+                owned_data_ = std::shared_ptr<uint8_t[]>(new uint8_t[size]());
+#else
                 owned_data_ = std::make_unique<uint8_t[]>(size);
+#endif
             }
             catch (const std::bad_alloc& e) {
                 throw std::runtime_error(std::string("Failed to allocate bytes of size ") + std::to_string(size) + ": " + e.what());
@@ -128,10 +151,14 @@ class bytes {
             throw std::runtime_error("Invalid size for bytes allocation");
         }
         size_t alignment = 1024 * 1024;
-        int padded_size = (size + alignment - 1) / alignment * alignment; // 1MB alignment
+        size_t padded_size = (size + alignment - 1) / alignment * alignment; // 1MB alignment
 
         try {
+#if defined(FLM_USE_HRX)
+            owned_bo_ = std::make_shared<flm_rt::ext::bo>(device, padded_size);
+#else
             owned_bo_ = std::make_unique<flm_rt::ext::bo>(device, padded_size);
+#endif
         }
         catch (const std::exception& e) {
             throw std::runtime_error(std::string("Failed to allocate flm_rt::ext::bo: ") + e.what());
@@ -169,6 +196,17 @@ class bytes {
     /// \param other the other bytes
     bytes& operator=(const bytes& other) {
         if (this != &other) {
+#if defined(FLM_USE_HRX)
+            owned_data_ = other.owned_data_;
+            data_ = other.data_;
+            size_ = other.size_;
+            is_owner_ = other.is_owner_;
+#ifdef FLM_DEVICE_BUFFER
+            owned_bo_ = other.owned_bo_;
+            is_bo_owner_ = other.is_bo_owner_;
+            bo_ = other.bo_;
+#endif
+#else
             if (is_owner_){
                 owned_data_.reset();
             }
@@ -181,6 +219,7 @@ class bytes {
             }
             is_bo_owner_ = false;
             bo_ = other.bo_;
+#endif
 #endif
         }
         return *this;
@@ -257,7 +296,11 @@ class bytes {
             throw std::runtime_error("Cannot resize to zero size");
         }
         try {
+#if defined(FLM_USE_HRX)
+            owned_data_.reset(new uint8_t[new_size]());
+#else
             owned_data_.reset(new uint8_t[new_size]);
+#endif
         }
         catch (const std::bad_alloc& e) {
             throw std::runtime_error(std::string("Failed to allocate bytes of size ") + std::to_string(new_size) + ": " + e.what());
@@ -302,6 +345,8 @@ class bytes {
     /// \return the is bo owner
     bool is_bo_owner() const { return is_bo_owner_; }
 
+    bool has_bo() const { return bo_ != nullptr; }
+
     /// \brief sync to device (host writes -> device)
 #if defined(FLM_USE_HRX)
     void sync_to_device() { assert(bo_); bo_->flush(); }
diff --git a/src/include/hrx_cpp/hrx_cpp.hpp b/src/include/hrx_cpp/hrx_cpp.hpp
index 13fbf79c..b4c4fe4d 100644
--- a/src/include/hrx_cpp/hrx_cpp.hpp
+++ b/src/include/hrx_cpp/hrx_cpp.hpp
@@ -113,10 +113,32 @@ inline bool hrx_report(hrx_status_t s, const char* where) {
 // Executable cache: build one HRX XADX executable per distinct executable
 // identity (xclbin + control program + host patch table) and resolve its export
 // ordinal once.
+//
+// Ownership / lifetime (context-exhaustion fix)
+// ---------------------------------------------------------------------------
+// Every executable is created with HRX_AMDXDNA_CONTEXT_MODE_CREATE, i.e. it is
+// backed by its own NPU hardware context. That pool is small: on the Windows
+// MCDM path the 314 driver runs out after enough distinct executables and
+// hrx_stream_wait then fails with
+//   D3DKMTCreateContextVirtual failed with 0xc01e0009
+// (the Linux amdxdna path fails the analogous BO allocation with errno 11 /
+// EAGAIN). A single long generation stays bounded because decode revisits the
+// same control sequences, but a `flm serve` sweep that loads many models used
+// to leak every model's executables forever: the cache was a process-wide map
+// that was never pruned on model unload, so contexts accumulated across the
+// sweep until the driver was exhausted -- exactly the "running out of contexts"
+// failure seen at gemma4-e4b after many prior models.
+//
+// Executables are therefore reference counted. Each npu_app that dispatches a
+// given (xclbin, ctrl_seq) holds exactly one reference for it; when the owning
+// model's engine objects are destroyed (model unload / model switch in serve)
+// the npu_app releases its references and, once the last owner drops, the
+// executable is destroyed and its hardware context returned to the driver.
 // ---------------------------------------------------------------------------
 struct CachedExe {
     hrx_executable_t exe = nullptr;
     uint32_t ord = 0;
+    size_t refs = 0;  // number of npu_app owners; released when it reaches 0
 };
 
 inline void append_key_bytes(std::string& key, const void* data, size_t byte_count) {
@@ -127,20 +149,37 @@ inline void append_key_bytes(std::string& key, const void* data, size_t byte_cou
     }
 }
 
-inline hrx_executable_t build_or_get_executable(
-    const std::vector<uint8_t>& xclbin_bytes, const uint32_t* cc, size_t n,
-    uint32_t* ord_out) {
-    static std::mutex mu;
-    static std::unordered_map<std::string, CachedExe> cache;
+inline std::mutex& exe_cache_mu() {
+    static std::mutex m;
+    return m;
+}
+inline std::unordered_map<std::string, CachedExe>& exe_cache() {
+    static std::unordered_map<std::string, CachedExe> c;
+    return c;
+}
+inline std::string make_exe_key(const std::vector<uint8_t>& xclbin_bytes,
+                                const uint32_t* cc, size_t n) {
     std::string key;
-    key.reserve(xclbin_bytes.size() + n * sizeof(uint32_t) +
-                2 * sizeof(uint64_t));
+    key.reserve(xclbin_bytes.size() + n * sizeof(uint32_t) + 2 * sizeof(uint64_t));
     append_key_bytes(key, xclbin_bytes.data(), xclbin_bytes.size());
     append_key_bytes(key, cc, n * sizeof(uint32_t));
-    std::lock_guard<std::mutex> lk(mu);
+    return key;
+}
+
+// Build (or fetch the cached) executable for (xclbin, cc). Does NOT change the
+// reference count; callers manage ownership with retain_executable /
+// release_executable below. Returns nullptr on create failure (never cached, so
+// a later call can retry instead of dispatching a silent no-op).
+inline hrx_executable_t lookup_or_build_executable(
+    const std::vector<uint8_t>& xclbin_bytes, const uint32_t* cc, size_t n,
+    uint32_t* ord_out, std::string* key_out) {
+    std::string key = make_exe_key(xclbin_bytes, cc, n);
+    std::lock_guard<std::mutex> lk(exe_cache_mu());
+    auto& cache = exe_cache();
     auto it = cache.find(key);
     if (it != cache.end()) {
         if (ord_out) *ord_out = it->second.ord;
+        if (key_out) *key_out = std::move(key);
         return it->second.exe;
     }
     hrx_const_byte_span_t xclbin = {xclbin_bytes.data(), xclbin_bytes.size()};
@@ -176,11 +215,51 @@ inline hrx_executable_t build_or_get_executable(
         hrx_executable_release(exe);
         exe = nullptr;
     }
-    cache.emplace(std::move(key), CachedExe{exe, ord});
+    if (exe) {
+        cache.emplace(key, CachedExe{exe, ord, 0});
+    }
     if (ord_out) *ord_out = ord;
+    if (key_out) *key_out = std::move(key);
     return exe;
 }
 
+// Add one owner reference to the cached executable identified by key.
+inline void retain_executable(const std::string& key) {
+    std::lock_guard<std::mutex> lk(exe_cache_mu());
+    auto& cache = exe_cache();
+    auto it = cache.find(key);
+    if (it != cache.end()) ++it->second.refs;
+}
+
+// Drop one owner reference; when the last owner drops the executable is
+// destroyed and its hardware context is returned to the driver.
+inline void release_executable(const std::string& key) {
+    hrx_executable_t to_free = nullptr;
+    {
+        std::lock_guard<std::mutex> lk(exe_cache_mu());
+        auto& cache = exe_cache();
+        auto it = cache.find(key);
+        if (it == cache.end()) return;
+        if (it->second.refs > 0) --it->second.refs;
+        if (it->second.refs == 0) {
+            to_free = it->second.exe;
+            cache.erase(it);
+        }
+    }
+    if (to_free) {
+        hrx_executable_release(to_free);
+    }
+}
+
+// Backward-compatible entry point (no ownership tracking). Retained for any
+// caller that does not manage executable lifetime; prefer the
+// lookup_or_build/retain/release trio used by npu_app.
+inline hrx_executable_t build_or_get_executable(
+    const std::vector<uint8_t>& xclbin_bytes, const uint32_t* cc, size_t n,
+    uint32_t* ord_out) {
+    return lookup_or_build_executable(xclbin_bytes, cc, n, ord_out, nullptr);
+}
+
 // ---------------------------------------------------------------------------
 // uuid / xclbin / device / hw_context
 // ---------------------------------------------------------------------------
diff --git a/src/include/npu_utils/npu_utils_hrx.hpp b/src/include/npu_utils/npu_utils_hrx.hpp
index 6268d698..1923ffdd 100644
--- a/src/include/npu_utils/npu_utils_hrx.hpp
+++ b/src/include/npu_utils/npu_utils_hrx.hpp
@@ -36,6 +36,7 @@
 #include <iostream>
 #include <sstream>
 #include <map>
+#include <unordered_set>
 #include <stdio.h>
 #ifndef __WINDOWS__
 #include <fcntl.h>
@@ -77,6 +78,30 @@ class npu_app {
     hrx_executable_t exe;  // HRX direct executable built from ctrl_seq->dump()
     uint32_t exe_ord;      // resolved "MLIR_AIE" export ordinal
     std::unique_ptr<npu_sequence> ctrl_seq;
+    // Executable-cache keys this app owns a reference for. Each distinct
+    // (xclbin, ctrl_seq) this app dispatches contributes one reference; they are
+    // all released in the destructor so the backing NPU hardware contexts are
+    // returned when the owning model is unloaded (see hrx_cpp.hpp).
+    std::unordered_set<std::string> exe_keys_;
+
+    ///@brief Transfer all state from `o` and leave it empty (owning no
+    ///       executable references) so its destructor releases nothing.
+    void move_from(npu_app&& o) noexcept {
+        device_gen = o.device_gen;
+        device = o.device;
+        context = o.context;
+        kernel_name = std::move(o.kernel_name);
+        enable_preemption = o.enable_preemption;
+        module_valid = o.module_valid;
+        module_version = o.module_version;
+        exe = o.exe;
+        exe_ord = o.exe_ord;
+        ctrl_seq = std::move(o.ctrl_seq);
+        exe_keys_ = std::move(o.exe_keys_);
+        o.exe = nullptr;
+        o.module_valid = false;
+        o.exe_keys_.clear();
+    }
 
     ///@brief Setup the kernel
     ///@note Builds (or fetches a cached) HRX XADX executable directly from the
@@ -86,13 +111,20 @@ class npu_app {
         assert(data.first != nullptr);
         assert(data.second > 0);
         assert(this->context != nullptr);
-        this->exe = hrx::build_or_get_executable(
+        std::string exe_key;
+        this->exe = hrx::lookup_or_build_executable(
             this->context->xclbin_bytes(), data.first, data.second,
-            &this->exe_ord);
+            &this->exe_ord, &exe_key);
         if (this->exe == nullptr){
             header_print_r("ERROR", "Failed to build HRX executable from ctrl_seq");
             exit(1);
         }
+        // Take one reference the first time this app uses this executable; the
+        // destructor drops all of them so the hardware context is freed on
+        // model unload instead of leaking for the process lifetime.
+        if (this->exe_keys_.insert(exe_key).second) {
+            hrx::retain_executable(exe_key);
+        }
         this->module_valid = true;
         this->module_version = this->ctrl_seq->sequence_version();
     }
@@ -154,6 +186,31 @@ class npu_app {
         this->module_version = 0xFF;
     }
 
+    ///@brief Destructor: release this app's executable references so their NPU
+    ///       hardware contexts are freed when the owning model is unloaded.
+    ~npu_app(){
+        for (const auto& k : this->exe_keys_) hrx::release_executable(k);
+    }
+
+    // npu_app owns HRX executable references, so it is move-only (copies are
+    // already deleted by the unique_ptr member). Declaring the destructor above
+    // suppresses the implicit move operations, so define them explicitly and
+    // transfer ownership of the executable keys, leaving the moved-from object
+    // with none (so it releases nothing).
+    npu_app(const npu_app&) = delete;
+    npu_app& operator=(const npu_app&) = delete;
+
+    npu_app(npu_app&& other) noexcept { this->move_from(std::move(other)); }
+
+    npu_app& operator=(npu_app&& other) noexcept {
+        if (this != &other) {
+            for (const auto& k : this->exe_keys_) hrx::release_executable(k);
+            this->exe_keys_.clear();
+            this->move_from(std::move(other));
+        }
+        return *this;
+    }
+
   
     void update_ctrl_seq(){
         assert(this->ctrl_seq != nullptr);
diff --git a/src/lib/hrx/dequant.dll b/src/lib/hrx/dequant.dll
index d6614115..ac666c5f 100644
Binary files a/src/lib/hrx/dequant.dll and b/src/lib/hrx/dequant.dll differ
diff --git a/src/lib/hrx/dequant.lib b/src/lib/hrx/dequant.lib
index 595bb742..420233d7 100644
Binary files a/src/lib/hrx/dequant.lib and b/src/lib/hrx/dequant.lib differ
diff --git a/src/lib/hrx/gemm.dll b/src/lib/hrx/gemm.dll
index ce51d217..ef544315 100644
Binary files a/src/lib/hrx/gemm.dll and b/src/lib/hrx/gemm.dll differ
diff --git a/src/lib/hrx/gemm.lib b/src/lib/hrx/gemm.lib
index 71708493..afd6c330 100644
Binary files a/src/lib/hrx/gemm.lib and b/src/lib/hrx/gemm.lib differ
diff --git a/src/lib/hrx/gemma4_12b_npu.dll b/src/lib/hrx/gemma4_12b_npu.dll
index 5eba1c4e..39285d9f 100644
Binary files a/src/lib/hrx/gemma4_12b_npu.dll and b/src/lib/hrx/gemma4_12b_npu.dll differ
diff --git a/src/lib/hrx/gemma4_12b_npu.lib b/src/lib/hrx/gemma4_12b_npu.lib
index d3f189e9..e9f9453f 100644
Binary files a/src/lib/hrx/gemma4_12b_npu.lib and b/src/lib/hrx/gemma4_12b_npu.lib differ
diff --git a/src/lib/hrx/gemma4e_npu.dll b/src/lib/hrx/gemma4e_npu.dll
index 5b4cdec1..d914ce7c 100644
Binary files a/src/lib/hrx/gemma4e_npu.dll and b/src/lib/hrx/gemma4e_npu.dll differ
diff --git a/src/lib/hrx/gemma4e_npu.lib b/src/lib/hrx/gemma4e_npu.lib
index bab48c3d..64061f53 100644
Binary files a/src/lib/hrx/gemma4e_npu.lib and b/src/lib/hrx/gemma4e_npu.lib differ
diff --git a/src/lib/hrx/gemma_embedding.dll b/src/lib/hrx/gemma_embedding.dll
index 476cf139..7d6d990c 100644
Binary files a/src/lib/hrx/gemma_embedding.dll and b/src/lib/hrx/gemma_embedding.dll differ
diff --git a/src/lib/hrx/gemma_embedding.lib b/src/lib/hrx/gemma_embedding.lib
index b5c9af6d..a54fbf35 100644
Binary files a/src/lib/hrx/gemma_embedding.lib and b/src/lib/hrx/gemma_embedding.lib differ
diff --git a/src/lib/hrx/gemma_npu.dll b/src/lib/hrx/gemma_npu.dll
index 50123227..3bc49456 100644
Binary files a/src/lib/hrx/gemma_npu.dll and b/src/lib/hrx/gemma_npu.dll differ
diff --git a/src/lib/hrx/gemma_npu.lib b/src/lib/hrx/gemma_npu.lib
index 79c76282..41b94e6d 100644
Binary files a/src/lib/hrx/gemma_npu.lib and b/src/lib/hrx/gemma_npu.lib differ
diff --git a/src/lib/hrx/gemma_text_npu.dll b/src/lib/hrx/gemma_text_npu.dll
index b7b3887c..2619345c 100644
Binary files a/src/lib/hrx/gemma_text_npu.dll and b/src/lib/hrx/gemma_text_npu.dll differ
diff --git a/src/lib/hrx/gemma_text_npu.lib b/src/lib/hrx/gemma_text_npu.lib
index ca0cc2fd..07e189d6 100644
Binary files a/src/lib/hrx/gemma_text_npu.lib and b/src/lib/hrx/gemma_text_npu.lib differ
diff --git a/src/lib/hrx/gpt_oss_npu.dll b/src/lib/hrx/gpt_oss_npu.dll
index 3e8f4f9f..66b45415 100644
Binary files a/src/lib/hrx/gpt_oss_npu.dll and b/src/lib/hrx/gpt_oss_npu.dll differ
diff --git a/src/lib/hrx/gpt_oss_npu.lib b/src/lib/hrx/gpt_oss_npu.lib
index c193c663..32ae0811 100644
Binary files a/src/lib/hrx/gpt_oss_npu.lib and b/src/lib/hrx/gpt_oss_npu.lib differ
diff --git a/src/lib/hrx/hunyuan_npu.dll b/src/lib/hrx/hunyuan_npu.dll
index 007154b7..532a02e2 100644
Binary files a/src/lib/hrx/hunyuan_npu.dll and b/src/lib/hrx/hunyuan_npu.dll differ
diff --git a/src/lib/hrx/hunyuan_npu.lib b/src/lib/hrx/hunyuan_npu.lib
index 44830bd1..aaba2b3b 100644
Binary files a/src/lib/hrx/hunyuan_npu.lib and b/src/lib/hrx/hunyuan_npu.lib differ
diff --git a/src/lib/hrx/lfm2_npu.dll b/src/lib/hrx/lfm2_npu.dll
index ac94c580..0d200917 100644
Binary files a/src/lib/hrx/lfm2_npu.dll and b/src/lib/hrx/lfm2_npu.dll differ
diff --git a/src/lib/hrx/lfm2_npu.lib b/src/lib/hrx/lfm2_npu.lib
index eefd034b..1d2f7f1e 100644
Binary files a/src/lib/hrx/lfm2_npu.lib and b/src/lib/hrx/lfm2_npu.lib differ
diff --git a/src/lib/hrx/libdequant.so b/src/lib/hrx/libdequant.so
index 42b859fa..e38037d1 100755
Binary files a/src/lib/hrx/libdequant.so and b/src/lib/hrx/libdequant.so differ
diff --git a/src/lib/hrx/libgemm.so b/src/lib/hrx/libgemm.so
index 4093c244..cf70041c 100755
Binary files a/src/lib/hrx/libgemm.so and b/src/lib/hrx/libgemm.so differ
diff --git a/src/lib/hrx/libgemma4_12b_npu.so b/src/lib/hrx/libgemma4_12b_npu.so
old mode 100644
new mode 100755
index aa347a5a..6774f4ec
Binary files a/src/lib/hrx/libgemma4_12b_npu.so and b/src/lib/hrx/libgemma4_12b_npu.so differ
diff --git a/src/lib/hrx/libgemma4e_npu.so b/src/lib/hrx/libgemma4e_npu.so
index 23353a2b..ec3070d1 100755
Binary files a/src/lib/hrx/libgemma4e_npu.so and b/src/lib/hrx/libgemma4e_npu.so differ
diff --git a/src/lib/hrx/libgemma_embedding.so b/src/lib/hrx/libgemma_embedding.so
index ddb74daa..acb8a60e 100755
Binary files a/src/lib/hrx/libgemma_embedding.so and b/src/lib/hrx/libgemma_embedding.so differ
diff --git a/src/lib/hrx/libgemma_npu.so b/src/lib/hrx/libgemma_npu.so
index 154f265b..1c09f7c1 100755
Binary files a/src/lib/hrx/libgemma_npu.so and b/src/lib/hrx/libgemma_npu.so differ
diff --git a/src/lib/hrx/libgemma_text_npu.so b/src/lib/hrx/libgemma_text_npu.so
index ad25e60c..13e8e1f5 100755
Binary files a/src/lib/hrx/libgemma_text_npu.so and b/src/lib/hrx/libgemma_text_npu.so differ
diff --git a/src/lib/hrx/libgpt_oss_npu.so b/src/lib/hrx/libgpt_oss_npu.so
index 26fc1985..a236a6b8 100755
Binary files a/src/lib/hrx/libgpt_oss_npu.so and b/src/lib/hrx/libgpt_oss_npu.so differ
diff --git a/src/lib/hrx/libhunyuan_npu.so b/src/lib/hrx/libhunyuan_npu.so
old mode 100644
new mode 100755
index 7079f0c1..f763759b
Binary files a/src/lib/hrx/libhunyuan_npu.so and b/src/lib/hrx/libhunyuan_npu.so differ
diff --git a/src/lib/hrx/liblfm2_npu.so b/src/lib/hrx/liblfm2_npu.so
index 8c568ff8..09e3427b 100755
Binary files a/src/lib/hrx/liblfm2_npu.so and b/src/lib/hrx/liblfm2_npu.so differ
diff --git a/src/lib/hrx/libllama_npu.so b/src/lib/hrx/libllama_npu.so
index 53a6e402..9056d74f 100755
Binary files a/src/lib/hrx/libllama_npu.so and b/src/lib/hrx/libllama_npu.so differ
diff --git a/src/lib/hrx/liblm_head.so b/src/lib/hrx/liblm_head.so
index 2666d258..ad7062a9 100755
Binary files a/src/lib/hrx/liblm_head.so and b/src/lib/hrx/liblm_head.so differ
diff --git a/src/lib/hrx/libmha.so b/src/lib/hrx/libmha.so
index daa108bb..979725a3 100755
Binary files a/src/lib/hrx/libmha.so and b/src/lib/hrx/libmha.so differ
diff --git a/src/lib/hrx/libnanbeige_npu.so b/src/lib/hrx/libnanbeige_npu.so
index 2bb6fa87..9ab2d184 100755
Binary files a/src/lib/hrx/libnanbeige_npu.so and b/src/lib/hrx/libnanbeige_npu.so differ
diff --git a/src/lib/hrx/libphi4_npu.so b/src/lib/hrx/libphi4_npu.so
index 67ca6e61..65aa9dc8 100755
Binary files a/src/lib/hrx/libphi4_npu.so and b/src/lib/hrx/libphi4_npu.so differ
diff --git a/src/lib/hrx/libq4_npu_eXpress.so b/src/lib/hrx/libq4_npu_eXpress.so
index 49074f5d..d692455f 100755
Binary files a/src/lib/hrx/libq4_npu_eXpress.so and b/src/lib/hrx/libq4_npu_eXpress.so differ
diff --git a/src/lib/hrx/libqwen2_npu.so b/src/lib/hrx/libqwen2_npu.so
index 340e5fa8..6230ed08 100755
Binary files a/src/lib/hrx/libqwen2_npu.so and b/src/lib/hrx/libqwen2_npu.so differ
diff --git a/src/lib/hrx/libqwen2vl_npu.so b/src/lib/hrx/libqwen2vl_npu.so
index d4845001..a2c93426 100755
Binary files a/src/lib/hrx/libqwen2vl_npu.so and b/src/lib/hrx/libqwen2vl_npu.so differ
diff --git a/src/lib/hrx/libqwen3_5_omni_npu.so b/src/lib/hrx/libqwen3_5_omni_npu.so
index 5e13c89f..4ba12039 100755
Binary files a/src/lib/hrx/libqwen3_5_omni_npu.so and b/src/lib/hrx/libqwen3_5_omni_npu.so differ
diff --git a/src/lib/hrx/libqwen3_5vl_npu.so b/src/lib/hrx/libqwen3_5vl_npu.so
index 3b7f9d8a..8144161c 100755
Binary files a/src/lib/hrx/libqwen3_5vl_npu.so and b/src/lib/hrx/libqwen3_5vl_npu.so differ
diff --git a/src/lib/hrx/libqwen3_6_moe_npu.so b/src/lib/hrx/libqwen3_6_moe_npu.so
index 6ea305cd..d6adba12 100755
Binary files a/src/lib/hrx/libqwen3_6_moe_npu.so and b/src/lib/hrx/libqwen3_6_moe_npu.so differ
diff --git a/src/lib/hrx/libqwen3_npu.so b/src/lib/hrx/libqwen3_npu.so
index 0842d680..3b47c7a5 100755
Binary files a/src/lib/hrx/libqwen3_npu.so and b/src/lib/hrx/libqwen3_npu.so differ
diff --git a/src/lib/hrx/libqwen3vl_npu.so b/src/lib/hrx/libqwen3vl_npu.so
index 5b47b72a..5468f363 100755
Binary files a/src/lib/hrx/libqwen3vl_npu.so and b/src/lib/hrx/libqwen3vl_npu.so differ
diff --git a/src/lib/hrx/libwhisper_npu.so b/src/lib/hrx/libwhisper_npu.so
index 71df18de..92d86174 100755
Binary files a/src/lib/hrx/libwhisper_npu.so and b/src/lib/hrx/libwhisper_npu.so differ
diff --git a/src/lib/hrx/llama_npu.dll b/src/lib/hrx/llama_npu.dll
index aa5157cb..94ae85f7 100644
Binary files a/src/lib/hrx/llama_npu.dll and b/src/lib/hrx/llama_npu.dll differ
diff --git a/src/lib/hrx/llama_npu.lib b/src/lib/hrx/llama_npu.lib
index a128ec96..c9a12e5e 100644
Binary files a/src/lib/hrx/llama_npu.lib and b/src/lib/hrx/llama_npu.lib differ
diff --git a/src/lib/hrx/lm_head.dll b/src/lib/hrx/lm_head.dll
index d83bee84..e30c24e4 100644
Binary files a/src/lib/hrx/lm_head.dll and b/src/lib/hrx/lm_head.dll differ
diff --git a/src/lib/hrx/lm_head.lib b/src/lib/hrx/lm_head.lib
index 311d63a8..dd3fb5f4 100644
Binary files a/src/lib/hrx/lm_head.lib and b/src/lib/hrx/lm_head.lib differ
diff --git a/src/lib/hrx/mha.dll b/src/lib/hrx/mha.dll
index fe511932..ff66b07d 100644
Binary files a/src/lib/hrx/mha.dll and b/src/lib/hrx/mha.dll differ
diff --git a/src/lib/hrx/mha.lib b/src/lib/hrx/mha.lib
index 81742777..98adbe51 100644
Binary files a/src/lib/hrx/mha.lib and b/src/lib/hrx/mha.lib differ
diff --git a/src/lib/hrx/nanbeige_npu.dll b/src/lib/hrx/nanbeige_npu.dll
index 59b5165f..7181c909 100644
Binary files a/src/lib/hrx/nanbeige_npu.dll and b/src/lib/hrx/nanbeige_npu.dll differ
diff --git a/src/lib/hrx/nanbeige_npu.lib b/src/lib/hrx/nanbeige_npu.lib
index 49a806e4..d5964499 100644
Binary files a/src/lib/hrx/nanbeige_npu.lib and b/src/lib/hrx/nanbeige_npu.lib differ
diff --git a/src/lib/hrx/phi4_npu.dll b/src/lib/hrx/phi4_npu.dll
index a21ac468..29b184d4 100644
Binary files a/src/lib/hrx/phi4_npu.dll and b/src/lib/hrx/phi4_npu.dll differ
diff --git a/src/lib/hrx/phi4_npu.lib b/src/lib/hrx/phi4_npu.lib
index d7860995..2f746e87 100644
Binary files a/src/lib/hrx/phi4_npu.lib and b/src/lib/hrx/phi4_npu.lib differ
diff --git a/src/lib/hrx/q4_npu_eXpress.dll b/src/lib/hrx/q4_npu_eXpress.dll
index b669c0c0..2061f4c1 100644
Binary files a/src/lib/hrx/q4_npu_eXpress.dll and b/src/lib/hrx/q4_npu_eXpress.dll differ
diff --git a/src/lib/hrx/q4_npu_eXpress.lib b/src/lib/hrx/q4_npu_eXpress.lib
index 963d31c2..07c41b52 100644
Binary files a/src/lib/hrx/q4_npu_eXpress.lib and b/src/lib/hrx/q4_npu_eXpress.lib differ
diff --git a/src/lib/hrx/qwen2_npu.dll b/src/lib/hrx/qwen2_npu.dll
index d83814da..5c1aa6a7 100644
Binary files a/src/lib/hrx/qwen2_npu.dll and b/src/lib/hrx/qwen2_npu.dll differ
diff --git a/src/lib/hrx/qwen2_npu.lib b/src/lib/hrx/qwen2_npu.lib
index 866af63f..9eb85b53 100644
Binary files a/src/lib/hrx/qwen2_npu.lib and b/src/lib/hrx/qwen2_npu.lib differ
diff --git a/src/lib/hrx/qwen2vl_npu.dll b/src/lib/hrx/qwen2vl_npu.dll
index a7519f36..998cc7e3 100644
Binary files a/src/lib/hrx/qwen2vl_npu.dll and b/src/lib/hrx/qwen2vl_npu.dll differ
diff --git a/src/lib/hrx/qwen2vl_npu.lib b/src/lib/hrx/qwen2vl_npu.lib
index 9d08e116..739beaa7 100644
Binary files a/src/lib/hrx/qwen2vl_npu.lib and b/src/lib/hrx/qwen2vl_npu.lib differ
diff --git a/src/lib/hrx/qwen3_5_omni_npu.dll b/src/lib/hrx/qwen3_5_omni_npu.dll
index d37c28be..d177150d 100644
Binary files a/src/lib/hrx/qwen3_5_omni_npu.dll and b/src/lib/hrx/qwen3_5_omni_npu.dll differ
diff --git a/src/lib/hrx/qwen3_5_omni_npu.lib b/src/lib/hrx/qwen3_5_omni_npu.lib
index a401cb88..e2e7f6f0 100644
Binary files a/src/lib/hrx/qwen3_5_omni_npu.lib and b/src/lib/hrx/qwen3_5_omni_npu.lib differ
diff --git a/src/lib/hrx/qwen3_5vl_npu.dll b/src/lib/hrx/qwen3_5vl_npu.dll
index a35449e3..a86ec377 100644
Binary files a/src/lib/hrx/qwen3_5vl_npu.dll and b/src/lib/hrx/qwen3_5vl_npu.dll differ
diff --git a/src/lib/hrx/qwen3_5vl_npu.lib b/src/lib/hrx/qwen3_5vl_npu.lib
index 9a05301d..9a133638 100644
Binary files a/src/lib/hrx/qwen3_5vl_npu.lib and b/src/lib/hrx/qwen3_5vl_npu.lib differ
diff --git a/src/lib/hrx/qwen3_6_moe_npu.dll b/src/lib/hrx/qwen3_6_moe_npu.dll
index 72cb8758..f151bb7d 100644
Binary files a/src/lib/hrx/qwen3_6_moe_npu.dll and b/src/lib/hrx/qwen3_6_moe_npu.dll differ
diff --git a/src/lib/hrx/qwen3_6_moe_npu.lib b/src/lib/hrx/qwen3_6_moe_npu.lib
index 204aa5be..f84b2cdb 100644
Binary files a/src/lib/hrx/qwen3_6_moe_npu.lib and b/src/lib/hrx/qwen3_6_moe_npu.lib differ
diff --git a/src/lib/hrx/qwen3_npu.dll b/src/lib/hrx/qwen3_npu.dll
index 5fb3d377..b168cd06 100644
Binary files a/src/lib/hrx/qwen3_npu.dll and b/src/lib/hrx/qwen3_npu.dll differ
diff --git a/src/lib/hrx/qwen3_npu.lib b/src/lib/hrx/qwen3_npu.lib
index 1352763e..6fdbf798 100644
Binary files a/src/lib/hrx/qwen3_npu.lib and b/src/lib/hrx/qwen3_npu.lib differ
diff --git a/src/lib/hrx/qwen3vl_npu.dll b/src/lib/hrx/qwen3vl_npu.dll
index cecb88fb..bd71adab 100644
Binary files a/src/lib/hrx/qwen3vl_npu.dll and b/src/lib/hrx/qwen3vl_npu.dll differ
diff --git a/src/lib/hrx/qwen3vl_npu.lib b/src/lib/hrx/qwen3vl_npu.lib
index 1ddbdfc4..0fc30d96 100644
Binary files a/src/lib/hrx/qwen3vl_npu.lib and b/src/lib/hrx/qwen3vl_npu.lib differ
diff --git a/src/lib/hrx/whisper_npu.dll b/src/lib/hrx/whisper_npu.dll
index dd7867e7..31206034 100644
Binary files a/src/lib/hrx/whisper_npu.dll and b/src/lib/hrx/whisper_npu.dll differ
diff --git a/src/lib/hrx/whisper_npu.lib b/src/lib/hrx/whisper_npu.lib
index 3fb1dd32..761313d4 100644
Binary files a/src/lib/hrx/whisper_npu.lib and b/src/lib/hrx/whisper_npu.lib differ