@fugood/llama.node 1.2.5 → 1.3.0-rc.0

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@fugood/llama.node",
   "access": "public",
-  "version": "1.2.5",
+  "version": "1.3.0-rc.0",
   "description": "An another Node binding of llama.cpp",
   "main": "lib/index.js",
   "scripts": {
@@ -72,19 +72,19 @@
     "CMakeLists.txt"
   ],
   "optionalDependencies": {
-    "@fugood/node-llama-linux-x64": "1.2.5",
-    "@fugood/node-llama-linux-x64-vulkan": "1.2.5",
-    "@fugood/node-llama-linux-x64-cuda": "1.2.5",
-    "@fugood/node-llama-linux-arm64": "1.2.5",
-    "@fugood/node-llama-linux-arm64-vulkan": "1.2.5",
-    "@fugood/node-llama-linux-arm64-cuda": "1.2.5",
-    "@fugood/node-llama-win32-x64": "1.2.5",
-    "@fugood/node-llama-win32-x64-vulkan": "1.2.5",
-    "@fugood/node-llama-win32-x64-cuda": "1.2.5",
-    "@fugood/node-llama-win32-arm64": "1.2.5",
-    "@fugood/node-llama-win32-arm64-vulkan": "1.2.5",
-    "@fugood/node-llama-darwin-x64": "1.2.5",
-    "@fugood/node-llama-darwin-arm64": "1.2.5"
+    "@fugood/node-llama-linux-x64": "1.3.0-rc.0",
+    "@fugood/node-llama-linux-x64-vulkan": "1.3.0-rc.0",
+    "@fugood/node-llama-linux-x64-cuda": "1.3.0-rc.0",
+    "@fugood/node-llama-linux-arm64": "1.3.0-rc.0",
+    "@fugood/node-llama-linux-arm64-vulkan": "1.3.0-rc.0",
+    "@fugood/node-llama-linux-arm64-cuda": "1.3.0-rc.0",
+    "@fugood/node-llama-win32-x64": "1.3.0-rc.0",
+    "@fugood/node-llama-win32-x64-vulkan": "1.3.0-rc.0",
+    "@fugood/node-llama-win32-x64-cuda": "1.3.0-rc.0",
+    "@fugood/node-llama-win32-arm64": "1.3.0-rc.0",
+    "@fugood/node-llama-win32-arm64-vulkan": "1.3.0-rc.0",
+    "@fugood/node-llama-darwin-x64": "1.3.0-rc.0",
+    "@fugood/node-llama-darwin-arm64": "1.3.0-rc.0"
   },
   "devDependencies": {
     "@babel/preset-env": "^7.24.4",

package/src/LlamaContext.cpp

@@ -168,6 +168,25 @@ void LlamaContext::Init(Napi::Env env, Napi::Object &exports) {
            static_cast<napi_property_attributes>(napi_enumerable)),
        InstanceMethod<&LlamaContext::DecodeAudioTokens>(
            "decodeAudioTokens",
+           static_cast<napi_property_attributes>(napi_enumerable)),
+       // Parallel decoding methods
+       InstanceMethod<&LlamaContext::EnableParallelMode>(
+           "enableParallelMode",
+           static_cast<napi_property_attributes>(napi_enumerable)),
+       InstanceMethod<&LlamaContext::DisableParallelMode>(
+           "disableParallelMode",
+           static_cast<napi_property_attributes>(napi_enumerable)),
+       InstanceMethod<&LlamaContext::QueueCompletion>(
+           "queueCompletion",
+           static_cast<napi_property_attributes>(napi_enumerable)),
+       InstanceMethod<&LlamaContext::QueueEmbedding>(
+           "queueEmbedding",
+           static_cast<napi_property_attributes>(napi_enumerable)),
+       InstanceMethod<&LlamaContext::QueueRerank>(
+           "queueRerank",
+           static_cast<napi_property_attributes>(napi_enumerable)),
+       InstanceMethod<&LlamaContext::CancelRequest>(
+           "cancelRequest",
            static_cast<napi_property_attributes>(napi_enumerable))});
   Napi::FunctionReference *constructor = new Napi::FunctionReference();
   *constructor = Napi::Persistent(func);
@@ -217,6 +236,7 @@ LlamaContext::LlamaContext(const Napi::CallbackInfo &info)
   params.n_ctx = get_option<int32_t>(options, "n_ctx", 512);
   params.n_batch = get_option<int32_t>(options, "n_batch", 2048);
   params.n_ubatch = get_option<int32_t>(options, "n_ubatch", 512);
+  params.n_parallel = get_option<int32_t>(options, "n_parallel", 1); // Default to 1 for compatibility
   params.embedding = get_option<bool>(options, "embedding", false);
   if (params.embedding) {
     // For non-causal models, batch size must be equal to ubatch size
@@ -288,6 +308,9 @@ LlamaContext::LlamaContext(const Napi::CallbackInfo &info)
       }
     }
   }
+  // Initialize validity flag for async callback safety
+  _context_valid = std::make_shared<std::atomic<bool>>(true);
+
   // Use rn-llama context instead of direct session
   _rn_ctx = new llama_rn_context();
   if (!_rn_ctx->loadModel(params)) {
@@ -305,6 +328,11 @@ LlamaContext::LlamaContext(const Napi::CallbackInfo &info)
 }
 
 LlamaContext::~LlamaContext() {
+  // Invalidate the context to prevent use-after-free in async callbacks
+  if (_context_valid) {
+    _context_valid->store(false);
+  }
+
   // The DisposeWorker is responsible for cleanup of _rn_ctx
   // If _rn_ctx is still not null here, it means disposal was not properly initiated
   if (_rn_ctx) {
@@ -579,7 +607,7 @@ Napi::Value LlamaContext::GetFormattedChat(const Napi::CallbackInfo &info) {
     // grammar: string
     result.Set("grammar", chatParams.grammar);
     // grammar_lazy: boolean
-    result.Set("grammea_lazy", chatParams.grammar_lazy);
+    result.Set("grammar_lazy", chatParams.grammar_lazy);
     // grammar_triggers: [{ value: string, token: number }]
     Napi::Array grammar_triggers = Napi::Array::New(env);
     for (size_t i = 0; i < chatParams.grammar_triggers.size(); i++) {
@@ -1135,6 +1163,11 @@ Napi::Value LlamaContext::Release(const Napi::CallbackInfo &info) {
     _wip->SetStop();
   }
 
+  // stop_processing_loop
+  if (_rn_ctx && _rn_ctx->slot_manager) {
+    _rn_ctx->slot_manager->stop_processing_loop();
+  }
+
   if (_rn_ctx == nullptr) {
     auto promise = Napi::Promise::Deferred(env);
     promise.Resolve(env.Undefined());

package/src/LlamaContext.h

@@ -4,6 +4,10 @@
 #include "rn-llama/rn-llama.h"
 #include "rn-llama/rn-completion.h"
 #include "rn-llama/rn-tts.h"
+#include "rn-llama/rn-slot.h"
+#include "rn-llama/rn-slot-manager.h"
+#include <atomic>
+#include <memory>
 
 using namespace rnllama;
 
@@ -55,10 +59,22 @@ private:
   Napi::Value GetAudioCompletionGuideTokens(const Napi::CallbackInfo &info);
   Napi::Value DecodeAudioTokens(const Napi::CallbackInfo &info);
 
+  // Parallel decoding methods
+  Napi::Value EnableParallelMode(const Napi::CallbackInfo &info);
+  void DisableParallelMode(const Napi::CallbackInfo &info);
+  Napi::Value QueueCompletion(const Napi::CallbackInfo &info);
+  Napi::Value QueueEmbedding(const Napi::CallbackInfo &info);
+  Napi::Value QueueRerank(const Napi::CallbackInfo &info);
+  void CancelRequest(const Napi::CallbackInfo &info);
+
   std::string _info;
   Napi::Object _meta;
   LlamaCompletionWorker *_wip = nullptr;
 
   // Use rn-llama context instead of direct llama.cpp types
   llama_rn_context *_rn_ctx = nullptr;
+
+  // Validity flag for async callbacks to prevent use-after-free
+  // Shared pointer ensures callbacks can safely check if context is still alive
+  std::shared_ptr<std::atomic<bool>> _context_valid;
 };

package/src/common.hpp

@@ -16,11 +16,12 @@ static bool is_nil(const Napi::Value &value) {
   return value.IsNull() || value.IsUndefined();
 }
 
-static std::string json_stringify(const Napi::Object &obj) {
-  Napi::Env env = obj.Env();
+// Overload for Napi::Value to handle both arrays and objects
+static std::string json_stringify(const Napi::Value &value) {
+  Napi::Env env = value.Env();
   Napi::Object json = env.Global().Get("JSON").As<Napi::Object>();
   Napi::Function stringify = json.Get("stringify").As<Napi::Function>();
-  return stringify.Call(json, {obj}).As<Napi::String>().ToString();
+  return stringify.Call(json, {value}).As<Napi::String>().ToString();
 }
 
 static void console_log(Napi::Env env, const std::string &message) {

package/src/llama.cpp/common/arg.cpp

@@ -1760,7 +1760,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_LOOKUP}));
     add_opt(common_arg(
         {"-t", "--threads"}, "N",
-        string_format("number of threads to use during generation (default: %d)", params.cpuparams.n_threads),
+        string_format("number of CPU threads to use during generation (default: %d)", params.cpuparams.n_threads),
         [](common_params & params, int value) {
             params.cpuparams.n_threads = value;
             if (params.cpuparams.n_threads <= 0) {

package/src/llama.cpp/ggml/include/ggml.h

@@ -577,6 +577,10 @@ extern "C" {
         GGML_UNARY_OP_EXP,
         GGML_UNARY_OP_GELU_ERF,
         GGML_UNARY_OP_XIELU,
+        GGML_UNARY_OP_FLOOR,
+        GGML_UNARY_OP_CEIL,
+        GGML_UNARY_OP_ROUND,
+        GGML_UNARY_OP_TRUNC,
 
         GGML_UNARY_OP_COUNT,
     };
@@ -1151,6 +1155,46 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    GGML_API struct ggml_tensor * ggml_floor(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_floor_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_ceil(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_ceil_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_round(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_round_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+     /**
+     * Truncates the fractional part of each element in the tensor (towards zero).
+     * For example: trunc(3.7) = 3.0, trunc(-2.9) = -2.0
+     * Similar to std::trunc in C/C++.
+     */
+
+    GGML_API struct ggml_tensor * ggml_trunc(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_trunc_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+
+
     // xIELU activation function
     // x = x * (c_a(alpha_n) + c_b(alpha_p, beta) * sigmoid(beta * x)) + eps * (x > 0)
     // where c_a = softplus and c_b(a, b) = softplus(a) + b are constraining functions

package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu-impl.h

@@ -68,7 +68,7 @@ struct ggml_compute_params {
 #endif  // __VXE2__
 #endif  // __s390x__ && __VEC__
 
-#if defined(__ARM_FEATURE_SVE)
+#if defined(__ARM_FEATURE_SVE) && defined(__linux__)
 #include <sys/prctl.h>
 #endif
 

package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu.c

@@ -689,8 +689,13 @@ bool ggml_is_numa(void) {
 #endif
 
 static void ggml_init_arm_arch_features(void) {
-#if defined(__linux__) && defined(__aarch64__) && defined(__ARM_FEATURE_SVE)
+#if defined(__aarch64__) && defined(__ARM_FEATURE_SVE)
+#if defined(__linux__)
     ggml_arm_arch_features.sve_cnt = PR_SVE_VL_LEN_MASK & prctl(PR_SVE_GET_VL);
+#else
+    // TODO: add support of SVE for non-linux systems
+#error "TODO: SVE is not supported on this platform. To use SVE, sve_cnt needs to be initialized here."
+#endif
 #endif
 }
 
@@ -2179,6 +2184,10 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
                 case GGML_UNARY_OP_HARDSWISH:
                 case GGML_UNARY_OP_HARDSIGMOID:
                 case GGML_UNARY_OP_EXP:
+                case GGML_UNARY_OP_FLOOR:
+                case GGML_UNARY_OP_CEIL:
+                case GGML_UNARY_OP_ROUND:
+                case GGML_UNARY_OP_TRUNC:
                     {
                         n_tasks = 1;
                     } break;
@@ -3558,13 +3567,17 @@ void ggml_cpu_init(void) {
 #ifdef GGML_USE_OPENMP
             //if (!getenv("OMP_WAIT_POLICY")) {
             //    // set the wait policy to active, so that OpenMP threads don't sleep
-            //    putenv("OMP_WAIT_POLICY=active");
+            //    setenv("OMP_WAIT_POLICY", "active", 0)
             //}
 
             if (!getenv("KMP_BLOCKTIME")) {
                 // set the time to wait before sleeping a thread
                 // this is less aggressive than setting the wait policy to active, but should achieve similar results in most cases
-                putenv("KMP_BLOCKTIME=200"); // 200ms
+#ifdef _WIN32
+                _putenv_s("KMP_BLOCKTIME", "200"); // 200ms
+#else
+                setenv("KMP_BLOCKTIME", "200", 0); // 200ms
+#endif
             }
 #endif
         }

package/src/llama.cpp/ggml/src/ggml-cpu/ops.cpp

@@ -8993,6 +8993,22 @@ void ggml_compute_forward_unary(
             {
                 ggml_compute_forward_exp(params, dst);
             } break;
+        case GGML_UNARY_OP_FLOOR:
+            {
+                ggml_compute_forward_floor(params, dst);
+            } break;
+        case GGML_UNARY_OP_CEIL:
+            {
+                ggml_compute_forward_ceil(params, dst);
+            } break;
+        case GGML_UNARY_OP_ROUND:
+            {
+                ggml_compute_forward_round(params, dst);
+            } break;
+        case GGML_UNARY_OP_TRUNC:
+            {
+                ggml_compute_forward_trunc(params, dst);
+            } break;
         case GGML_UNARY_OP_XIELU:
             {
                 ggml_compute_forward_xielu(params, dst);

package/src/llama.cpp/ggml/src/ggml-cpu/unary-ops.cpp

@@ -73,6 +73,22 @@ static inline float op_log(float x) {
     return logf(x);
 }
 
+static inline float op_floor(float x) {
+    return floorf(x);
+}
+
+static inline float op_ceil(float x) {
+    return ceilf(x);
+}
+
+static inline float op_round(float x) {
+    return roundf(x);
+}
+
+static inline float op_trunc(float x) {
+    return truncf(x);
+}
+
 template <float (*op)(float), typename src0_t, typename dst_t>
 static inline void vec_unary_op(int64_t n, dst_t * y, const src0_t * x) {
     constexpr auto src0_to_f32 = type_conversion_table<src0_t>::to_f32;
@@ -274,6 +290,22 @@ void ggml_compute_forward_log(const ggml_compute_params * params, ggml_tensor *
     unary_op<op_log>(params, dst);
 }
 
+void ggml_compute_forward_floor(const ggml_compute_params * params, ggml_tensor * dst) {
+    unary_op<op_floor>(params, dst);
+}
+
+void ggml_compute_forward_ceil(const ggml_compute_params * params, ggml_tensor * dst) {
+    unary_op<op_ceil>(params, dst);
+}
+
+void ggml_compute_forward_round(const ggml_compute_params * params, ggml_tensor * dst) {
+    unary_op<op_round>(params, dst);
+}
+
+void ggml_compute_forward_trunc(const ggml_compute_params * params, ggml_tensor * dst) {
+    unary_op<op_trunc>(params, dst);
+}
+
 void ggml_compute_forward_xielu(const ggml_compute_params * params, ggml_tensor * dst) {
     const float alpha_n = ggml_get_op_params_f32(dst, 1);
     const float alpha_p = ggml_get_op_params_f32(dst, 2);

package/src/llama.cpp/ggml/src/ggml-cpu/unary-ops.h

@@ -22,6 +22,10 @@ void ggml_compute_forward_sqrt(const struct ggml_compute_params * params, struct
 void ggml_compute_forward_sin(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_cos(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_log(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_floor(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_ceil(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_round(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_trunc(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_xielu(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 
 #ifdef __cplusplus

package/src/llama.cpp/ggml/src/ggml-cpu/vec.cpp

@@ -463,9 +463,9 @@ ggml_float ggml_vec_cvar_f32(const int n, float * y, const float * x, const floa
 #endif
     for (; i < n; ++i) {
         float val = x[i] - mean;
+        y[i] = val;
         val *= val;
         sum += (ggml_float)val;
-        y[i] = val;
     }
     return sum/n;
 }

package/src/llama.cpp/src/llama-arch.cpp

@@ -5,6 +5,7 @@
 #include <map>
 
 static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
+    { LLM_ARCH_CLIP,             "clip"             }, // dummy, only used by llama-quantize
     { LLM_ARCH_LLAMA,            "llama"            },
     { LLM_ARCH_LLAMA4,           "llama4"           },
     { LLM_ARCH_DECI,             "deci"             },
@@ -275,6 +276,10 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
 };
 
 static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_NAMES = {
+    {
+        LLM_ARCH_CLIP,
+        {},
+    },
     {
         LLM_ARCH_LLAMA,
         {

package/src/llama.cpp/src/llama-arch.h

@@ -9,6 +9,7 @@
 //
 
 enum llm_arch {
+    LLM_ARCH_CLIP,
     LLM_ARCH_LLAMA,
     LLM_ARCH_LLAMA4,
     LLM_ARCH_DECI,

package/src/llama.cpp/src/llama-graph.cpp

@@ -261,12 +261,17 @@ void llm_graph_input_cross_embd::set_input(const llama_ubatch * ubatch) {
     }
 }
 
-static void print_mask(float * data, int64_t n_tokens, int64_t n_kv, int64_t n_swa, llama_swa_type swa_type) {
+static void print_mask(const float * data, int64_t n_tokens, int64_t n_kv, int64_t n_swa, llama_swa_type swa_type) {
     LLAMA_LOG_DEBUG("%s: === Attention mask ===\n", __func__);
-    const char * swa_type_str = (swa_type == LLAMA_SWA_TYPE_NONE) ? "LLAMA_SWA_TYPE_NONE" :
-                          (swa_type == LLAMA_SWA_TYPE_STANDARD) ? "LLAMA_SWA_TYPE_STANDARD" :
-                          (swa_type == LLAMA_SWA_TYPE_CHUNKED) ? "LLAMA_SWA_TYPE_CHUNKED" :
-                          (swa_type == LLAMA_SWA_TYPE_SYMMETRIC) ? "LLAMA_SWA_TYPE_SYMMETRIC" : "unknown";
+    const char * swa_type_str = "unknown";
+
+    switch (swa_type) {
+        case LLAMA_SWA_TYPE_NONE:      swa_type_str = "LLAMA_SWA_TYPE_NONE"; break;
+        case LLAMA_SWA_TYPE_STANDARD:  swa_type_str = "LLAMA_SWA_TYPE_STANDARD"; break;
+        case LLAMA_SWA_TYPE_CHUNKED:   swa_type_str = "LLAMA_SWA_TYPE_CHUNKED"; break;
+        case LLAMA_SWA_TYPE_SYMMETRIC: swa_type_str = "LLAMA_SWA_TYPE_SYMMETRIC"; break;
+    };
+
     LLAMA_LOG_DEBUG("%s: n_swa : %d, n_kv: %d, swq_type: %s\n", __func__, (int)n_swa, (int)n_kv, swa_type_str);
     LLAMA_LOG_DEBUG("%s: '0' = can attend, '∞' = masked\n", __func__);
     LLAMA_LOG_DEBUG("%s: Rows = query tokens, Columns = key/value tokens\n\n", __func__);
@@ -295,50 +300,67 @@ void llm_graph_input_attn_no_cache::set_input(const llama_ubatch * ubatch) {
     const int64_t n_kv     = ubatch->n_tokens;
     const int64_t n_tokens = ubatch->n_tokens;
 
-    GGML_ASSERT(kq_mask);
-    GGML_ASSERT(ggml_backend_buffer_is_host(kq_mask->buffer));
-
-    float * data = (float *) kq_mask->data;
-
-    // [TAG_NO_CACHE_ISWA]
-    GGML_ASSERT(hparams.swa_type == LLAMA_SWA_TYPE_NONE && "TODO: implement");
+    const auto fill_mask = [&](float * data, int n_swa, llama_swa_type swa_type) {
+        for (int h = 0; h < 1; ++h) {
+            for (int i1 = 0; i1 < n_tokens; ++i1) {
+                const llama_seq_id s1 = ubatch->seq_id[i1][0];
+                const llama_pos    p1 = ubatch->pos[i1];
 
-    for (int h = 0; h < 1; ++h) {
-        for (int i1 = 0; i1 < n_tokens; ++i1) {
-            const llama_seq_id s1 = ubatch->seq_id[i1][0];
+                const uint64_t idst = h*(n_kv*n_tokens) + i1*n_kv;
 
-            for (int i0 = 0; i0 < n_tokens; ++i0) {
-                float f = -INFINITY;
-
-                for (int s = 0; s < ubatch->n_seq_id[i0]; ++s) {
+                for (int i0 = 0; i0 < n_tokens; ++i0) {
                     const llama_seq_id s0 = ubatch->seq_id[i0][0];
+                    const llama_pos p0    = ubatch->pos[i0];
 
+                    // mask different sequences
                     if (s0 != s1) {
-                        continue; // skip different sequences
+                        continue;
                     }
 
-                    if (cparams.causal_attn && ubatch->pos[i0] > ubatch->pos[i1]) {
-                        continue; // skip future tokens for causal attention
+                    // mask future tokens
+                    if (cparams.causal_attn && p0 > p1) {
+                        continue;
                     }
 
-                    // TODO: this does not take into account that some layers are SWA and others are note (i.e. iSWA) [TAG_NO_CACHE_ISWA]
-                    //if (hparams.is_masked_swa(ubatch->pos[i0], ubatch->pos[i1])) {
-                    //    continue; // skip masked tokens for SWA
-                    //}
-
-                    // TODO: reimplement this like in llama_kv_cache_unified
-                    if (hparams.use_alibi) {
-                        f = -std::abs(ubatch->pos[i0] - ubatch->pos[i1]);
-                    } else {
-                        f = 0.0f;
+                    // apply SWA if any
+                    if (llama_hparams::is_masked_swa(n_swa, swa_type, p0, p1)) {
+                        continue;
                     }
+
+                    data[idst + i0] = hparams.use_alibi ? -std::abs(p0 - p1) : 0.0f;
                 }
-                data[h*(n_kv*n_tokens) + i1*n_kv + i0] = f;
             }
         }
+    };
+
+    {
+        GGML_ASSERT(self_kq_mask);
+        GGML_ASSERT(ggml_backend_buffer_is_host(self_kq_mask->buffer));
+
+        float * data = (float *) self_kq_mask->data;
+
+        std::fill(data, data + ggml_nelements(self_kq_mask), -INFINITY);
+
+        fill_mask(data, 0, LLAMA_SWA_TYPE_NONE);
+
+        if (debug) {
+            print_mask(data, n_tokens, n_kv, 0, LLAMA_SWA_TYPE_NONE);
+        }
     }
-    if (debug) {
-        print_mask(data, n_tokens, n_kv, hparams.n_swa, hparams.swa_type);
+
+    if (hparams.swa_type != LLAMA_SWA_TYPE_NONE) {
+        GGML_ASSERT(self_kq_mask_swa);
+        GGML_ASSERT(ggml_backend_buffer_is_host(self_kq_mask_swa->buffer));
+
+        float * data = (float *) self_kq_mask_swa->data;
+
+        std::fill(data, data + ggml_nelements(self_kq_mask_swa), -INFINITY);
+
+        fill_mask(data, hparams.n_swa, hparams.swa_type);
+
+        if (debug) {
+            print_mask(data, n_tokens, n_kv, hparams.n_swa, hparams.swa_type);
+        }
     }
 }
 
@@ -1299,12 +1321,9 @@ ggml_tensor * llm_graph_context::build_attn_mha(
     k = ggml_permute(ctx0, k, 0, 2, 1, 3);
     v = ggml_permute(ctx0, v, 0, 2, 1, 3);
 
-    const auto n_kv = k->ne[1];
-
     ggml_tensor * cur;
 
-    // TODO: replace hardcoded padding with ggml-provided padding
-    if (cparams.flash_attn && (n_kv % 256 == 0) && kq_b == nullptr) {
+    if (cparams.flash_attn && kq_b == nullptr) {
         GGML_ASSERT(kq_b == nullptr && "Flash attention does not support KQ bias yet");
 
         if (v_trans) {
@@ -1419,10 +1438,20 @@ llm_graph_input_attn_no_cache * llm_graph_context::build_attn_inp_no_cache() con
     auto inp = std::make_unique<llm_graph_input_attn_no_cache>(hparams, cparams);
 
     // note: there is no KV cache, so the number of KV values is equal to the number of tokens in the batch
-    inp->kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_tokens, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD), 1, 1);
-    ggml_set_input(inp->kq_mask);
+    inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_tokens, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD), 1, 1);
+    ggml_set_input(inp->self_kq_mask);
+
+    inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
 
-    inp->kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->kq_mask, GGML_TYPE_F16) : inp->kq_mask;
+    if (hparams.swa_type != LLAMA_SWA_TYPE_NONE) {
+        inp->self_kq_mask_swa = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_tokens, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD), 1, 1);
+        ggml_set_input(inp->self_kq_mask_swa);
+
+        inp->self_kq_mask_swa_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask_swa, GGML_TYPE_F16) : inp->self_kq_mask_swa;
+    } else {
+        inp->self_kq_mask_swa     = nullptr;
+        inp->self_kq_mask_swa_cnv = nullptr;
+    }
 
     return (llm_graph_input_attn_no_cache *) res->add_input(std::move(inp));
 }
@@ -1447,7 +1476,9 @@ ggml_tensor * llm_graph_context::build_attn(
     ggml_build_forward_expand(gf, k_cur);
     ggml_build_forward_expand(gf, v_cur);
 
-    const auto & kq_mask = inp->get_kq_mask();
+    const bool is_swa = hparams.is_swa(il);
+
+    const auto & kq_mask = is_swa ? inp->get_kq_mask_swa() : inp->get_kq_mask();
 
     // [TAG_NO_CACHE_PAD]
     // TODO: if ubatch.equal_seqs() == true, we can split the three tensors below into ubatch.n_seqs_unq streams

package/src/llama.cpp/src/llama-graph.h

@@ -257,10 +257,14 @@ public:
 
     void set_input(const llama_ubatch * ubatch) override;
 
-    ggml_tensor * get_kq_mask() const { return kq_mask_cnv; }
+    ggml_tensor * get_kq_mask()     const { return self_kq_mask_cnv; }
+    ggml_tensor * get_kq_mask_swa() const { return self_kq_mask_swa_cnv; }
 
-    ggml_tensor * kq_mask     = nullptr; // F32 [n_tokens, n_batch, 1, 1]
-    ggml_tensor * kq_mask_cnv = nullptr; //     [n_tokens, n_batch, 1, 1]
+    // n_tokens == n_batch
+    ggml_tensor * self_kq_mask         = nullptr; // F32 [n_tokens, n_batch/n_stream, 1, n_stream]
+    ggml_tensor * self_kq_mask_cnv     = nullptr; //     [n_tokens, n_batch/n_stream, 1, n_stream]
+    ggml_tensor * self_kq_mask_swa     = nullptr; // F32 [n_tokens, n_batch/n_stream, 1, n_stream]
+    ggml_tensor * self_kq_mask_swa_cnv = nullptr; //     [n_tokens, n_batch/n_stream, 1, n_stream]
 
     const llama_hparams hparams;
     const llama_cparams cparams;