@fugood/llama.node 1.4.15 → 1.6.0-rc.0

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

@@ -630,10 +630,11 @@ extern "C" {
 
     // this tensor...
     enum ggml_tensor_flag {
-        GGML_TENSOR_FLAG_INPUT  =  1, // ...is an input for the GGML compute graph
-        GGML_TENSOR_FLAG_OUTPUT =  2, // ...is an output for the GGML compute graph
-        GGML_TENSOR_FLAG_PARAM  =  4, // ...contains trainable parameters
-        GGML_TENSOR_FLAG_LOSS   =  8, // ...defines loss for numerical optimization (multiple loss tensors add up)
+        GGML_TENSOR_FLAG_INPUT   =  1, // ...is an input for the GGML compute graph
+        GGML_TENSOR_FLAG_OUTPUT  =  2, // ...is an output for the GGML compute graph
+        GGML_TENSOR_FLAG_PARAM   =  4, // ...contains trainable parameters
+        GGML_TENSOR_FLAG_LOSS    =  8, // ...defines loss for numerical optimization (multiple loss tensors add up)
+        GGML_TENSOR_FLAG_COMPUTE = 16, // ...must be computed
     };
 
     enum ggml_tri_type {
@@ -2577,11 +2578,42 @@ extern "C" {
         struct ggml_tensor *  grad,
         struct ggml_tensor *  sgd_params); // alpha, weight decay
 
+    // build forward mutiple tensors and select one of them for computing
+    // this is useful for creating graphs that have constant topology but compute different things based on the input
+    // ref: https://github.com/ggml-org/llama.cpp/pull/18550
     //
-    // automatic differentiation
+    // nodes:
+    //   | - build forward into the graph but do not compute
+    //   c - build forward into the graph and compute
     //
+    //    |  |  ...  c  ...  |
+    //    |  |  ...  c  ...  |
+    //    |  |  ...  c  ...  |
+    //   [0  1  ... idx ...  n-1]        <-- ggml_build_forward_select(..., n, idx)
+    //               c
+    //               c
+    //
+    // example:
+    //   struct ggml_tensor * curs[3];
+    //
+    //   curs[0]  = compute0(...);
+    //   curs[1]  = compute1(...);
+    //   curs[2]  = compute2(...);
+    //
+    //   int idx = select_branch(some_input);
+    //
+    //   struct ggml_tensor * out = ggml_build_forward_select(cgraph, curs, 3, idx);
+    //
+    GGML_API struct ggml_tensor * ggml_build_forward_select(
+            struct ggml_cgraph  * cgraph,
+            struct ggml_tensor ** tensors,
+            int                   n_tensors,
+            int                   idx);
+
+    GGML_API void ggml_build_forward_expand(
+            struct ggml_cgraph * cgraph,
+            struct ggml_tensor * tensor);
 
-    GGML_API void ggml_build_forward_expand(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor);
     GGML_API void ggml_build_backward_expand(
         struct ggml_context *  ctx,        // context for gradient computation
         struct ggml_cgraph  *  cgraph,
@@ -2613,7 +2645,7 @@ extern "C" {
     GGML_API void ggml_graph_print(const struct ggml_cgraph * cgraph);
 
     // dump the graph into a file using the dot format
-    GGML_API void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph * gf, const char * filename);
+    GGML_API void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph * cgraph, const char * filename);
 
     // TODO these functions were sandwiched in the old optimization interface, is there a better place for them?
     typedef void (*ggml_log_callback)(enum ggml_log_level level, const char * text, void * user_data);

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

@@ -2943,6 +2943,10 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
             continue;
         }
 
+        if ((node->flags & GGML_TENSOR_FLAG_COMPUTE) == 0) {
+            continue;
+        }
+
         ggml_compute_forward(&params, node);
 
         if (state->ith == 0 && cplan->abort_callback &&

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

@@ -7,10 +7,9 @@
 #include "unary-ops.h"
 #include "vec.h"
 
-#include <cfloat>
 #include <algorithm>
+#include <cfloat>
 #include <cmath>
-#include <functional>
 
 // ggml_compute_forward_dup
 
@@ -7110,12 +7109,13 @@ void ggml_compute_forward_conv_2d_dw(
     }
 }
 
-// ggml_compute_forward_pool_1d_sk_p0
-
-static void ggml_compute_forward_pool_1d_sk_p0(
+// ggml_compute_forward_pool_1d_ksp
+static void ggml_compute_forward_pool_1d_ksp(
         const ggml_compute_params * params,
         const ggml_op_pool op,
         const int k,
+        const int s,
+        const int p,
         ggml_tensor * dst) {
 
     const ggml_tensor * src = dst->src[0];
@@ -7126,39 +7126,56 @@ static void ggml_compute_forward_pool_1d_sk_p0(
         return;
     }
 
-    const char * cdata = (const char *)src->data;
-    const char * const data_end = cdata + ggml_nbytes(src);
-    float * drow = (float *)dst->data;
+    const int64_t IW = src->ne[0];
+    const int64_t OW = dst->ne[0];
 
-    const int64_t rs = dst->ne[0];
+    const int64_t nr = ggml_nrows(src);
 
-    while (cdata < data_end) {
-        const void * srow = (const void *)cdata;
-        int j = 0;
-        for (int64_t i = 0; i < rs; ++i) {
+    for (int64_t ir = 0; ir < nr; ++ir) {
+        const char * srow_bytes =            (const char *) src->data + ir * src->nb[1];
+        float      * drow       = (float *) ((      char *) dst->data + ir * dst->nb[1]);
+
+        for (int64_t ow = 0; ow < OW; ++ow) {
+            float res = 0;
             switch (op) {
-                case GGML_OP_POOL_AVG:   drow[i] = 0;        break;
-                case GGML_OP_POOL_MAX:   drow[i] = -FLT_MAX; break;
+                case GGML_OP_POOL_AVG: res = 0.0f;     break;
+                case GGML_OP_POOL_MAX: res = -FLT_MAX; break;
                 case GGML_OP_POOL_COUNT: GGML_ABORT("fatal error");
             }
+
+            int count = 0;
+            const int base = (int) ow * s - p;
+
             for (int ki = 0; ki < k; ++ki) {
-                const float srow_j = (src->type == GGML_TYPE_F32) ? ((const float*)srow)[j] : GGML_CPU_FP16_TO_FP32(((const ggml_fp16_t*)srow)[j]);
+                const int j = base + ki;
+                if (j < 0 || j >= (int) IW) {
+                    continue;
+                }
+
+                float v;
+                if (src->type == GGML_TYPE_F32) {
+                    v = ((const float *) srow_bytes)[j];
+                } else {
+                    v = GGML_CPU_FP16_TO_FP32(((const ggml_fp16_t *) srow_bytes)[j]);
+                }
+
                 switch (op) {
-                    case GGML_OP_POOL_AVG:                         drow[i] += srow_j; break;
-                    case GGML_OP_POOL_MAX:   if (srow_j > drow[i]) drow[i]  = srow_j; break;
-                    case GGML_OP_POOL_COUNT:                       GGML_ABORT("fatal error");
+                    case GGML_OP_POOL_AVG: res += v;                break;
+                    case GGML_OP_POOL_MAX: res =  std::max(v, res); break;
+                    case GGML_OP_POOL_COUNT: GGML_ABORT("fatal error");
                 }
-                ++j;
+
+                ++count;
             }
+
             switch (op) {
-                case GGML_OP_POOL_AVG:         drow[i] /= k; break;
-                case GGML_OP_POOL_MAX:                       break;
+                case GGML_OP_POOL_AVG: res = (count > 0) ? (res / count) : 0.0f; break;
+                case GGML_OP_POOL_MAX:                                           break;
                 case GGML_OP_POOL_COUNT: GGML_ABORT("fatal error");
             }
-        }
 
-        cdata += src->nb[1];
-        drow  += rs;
+            drow[ow] = res;
+        }
     }
 }
 
@@ -7173,10 +7190,8 @@ void ggml_compute_forward_pool_1d(
     const int k0 = opts[1];
     const int s0 = opts[2];
     const int p0 = opts[3];
-    GGML_ASSERT(p0 == 0); // padding not supported
-    GGML_ASSERT(k0 == s0); // only s = k supported
 
-    ggml_compute_forward_pool_1d_sk_p0(params, op, k0, dst);
+    ggml_compute_forward_pool_1d_ksp(params, op, k0, s0, p0, dst);
 }
 
 // ggml_compute_forward_pool_2d
@@ -7194,6 +7209,7 @@ void ggml_compute_forward_pool_2d(
     }
 
     const int32_t * opts = (const int32_t *)dst->op_params;
+
     ggml_op_pool op = static_cast<ggml_op_pool>(opts[0]);
     const int k0 = opts[1];
     const int k1 = opts[2];
@@ -7217,11 +7233,13 @@ void ggml_compute_forward_pool_2d(
     while (cdata < data_end) {
         for (int oy = 0; oy < py; ++oy) {
             float * const drow = dplane + oy * px;
+            float * const out  = drow;
+
             for (int ox = 0; ox < px; ++ox) {
-                float * const out =  drow + ox;
+                float res = 0;
                 switch (op) {
-                    case GGML_OP_POOL_AVG:     *out = 0;        break;
-                    case GGML_OP_POOL_MAX:     *out = -FLT_MAX; break;
+                    case GGML_OP_POOL_AVG: res = 0;        break;
+                    case GGML_OP_POOL_MAX: res = -FLT_MAX; break;
                     case GGML_OP_POOL_COUNT: GGML_ABORT("fatal error");
                 }
 
@@ -7229,24 +7247,32 @@ void ggml_compute_forward_pool_2d(
                 const int iy = offset1 + oy * s1;
 
                 for (int ky = 0; ky < k1; ++ky) {
-                    if (iy + ky < 0 || iy + ky >= src->ne[1]) continue;
+                    if (iy + ky < 0 || iy + ky >= src->ne[1]) {
+                        continue;
+                    }
+
                     const void * srow = (const void *)(cdata + src->nb[1] * (iy + ky));
                     for (int kx = 0; kx < k0; ++kx) {
                         int j = ix + kx;
-                        if (j < 0 || j >= src->ne[0]) continue;
+                        if (j < 0 || j >= src->ne[0]) {
+                            continue;
+                        }
+
                         const float srow_j = (src->type == GGML_TYPE_F32) ? ((const float*)srow)[j] : GGML_CPU_FP16_TO_FP32(((const ggml_fp16_t*)srow)[j]);
                         switch (op) {
-                            case GGML_OP_POOL_AVG:                     *out += srow_j; break;
-                            case GGML_OP_POOL_MAX: if (srow_j > *out)  *out  = srow_j; break;
+                            case GGML_OP_POOL_AVG: res += srow_j;                break;
+                            case GGML_OP_POOL_MAX: res =  std::max(srow_j, res); break;
                             case GGML_OP_POOL_COUNT:               GGML_ABORT("fatal error");
                         }
                     }
                 }
                 switch (op) {
-                    case GGML_OP_POOL_AVG:           *out /= ka; break;
-                    case GGML_OP_POOL_MAX:                       break;
+                    case GGML_OP_POOL_AVG:           res /= ka; break;
+                    case GGML_OP_POOL_MAX:                      break;
                     case GGML_OP_POOL_COUNT: GGML_ABORT("fatal error");
                 }
+
+                out[ox] = res;
             }
         }
 

package/src/llama.cpp/ggml/src/ggml-cpu/simd-mappings.h

@@ -654,6 +654,14 @@ static inline void __avx_f32cx8_store(ggml_fp16_t *x, __m256 y) {
           vec_extract(x[0], 2) +               \
           vec_extract(x[0], 3);                \
 }
+#define GGML_F32x4_REDUCE_4(res, s0, s1, s2, s3)        \
+{                                                       \
+    vector float v = vec_add(vec_add(s0, s1),           \
+                             vec_add(s2, s3));          \
+    v = vec_add(v, vec_sld(v, v, 8));                   \
+    v = vec_add(v, vec_sld(v, v, 4));                   \
+    res += (ggml_float) vec_extract(v, 0);              \
+}
 
 #define GGML_F32_VEC        GGML_F32x4
 #define GGML_F32_VEC_ZERO   GGML_F32x4_ZERO
@@ -690,6 +698,29 @@ static inline unsigned char ggml_endian_byte(int i) {
                                    r[i - GGML_ENDIAN_BYTE(0)]), \
             0, p - GGML_F16_EPR)
 
+//BF16 POWER9
+#define GGML_BF16_STEP 16
+#define GGML_BF16_EPR  8
+
+#define GGML_BF16x8         vector unsigned short
+#define GGML_BF16x8_ZERO    vec_splats((unsigned short)0)
+#define GGML_BF16x8_LOAD(p) vec_xl(0, (const unsigned short *)(p))
+
+#define GGML_BF16_VEC          GGML_BF16x8
+#define GGML_BF16_VEC_ZERO     GGML_BF16x8_ZERO
+#define GGML_BF16_VEC_LOAD     GGML_BF16x8_LOAD
+#if defined(__LITTLE_ENDIAN__)
+#define GGML_BF16_TO_F32_LO(v) ((vector float) vec_mergel(GGML_BF16_VEC_ZERO, (v)))
+#define GGML_BF16_TO_F32_HI(v) ((vector float) vec_mergeh(GGML_BF16_VEC_ZERO, (v)))
+#else
+#define GGML_BF16_TO_F32_LO(v) ((vector float) vec_mergel((v), GGML_BF16_VEC_ZERO))
+#define GGML_BF16_TO_F32_HI(v) ((vector float) vec_mergeh((v), GGML_BF16_VEC_ZERO))
+#endif
+#define GGML_BF16_FMA_LO(acc, x, y) \
+    (acc) = GGML_F32x4_FMA((acc), GGML_BF16_TO_F32_LO(x), GGML_BF16_TO_F32_LO(y))
+#define GGML_BF16_FMA_HI(acc, x, y) \
+    (acc) = GGML_F32x4_FMA((acc), GGML_BF16_TO_F32_HI(x), GGML_BF16_TO_F32_HI(y))
+
 #elif defined(__wasm_simd128__)
 
 #define GGML_SIMD

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

@@ -237,6 +237,24 @@ void ggml_vec_dot_bf16(int n, float * GGML_RESTRICT s, size_t bs, ggml_bf16_t *
     sumf += __riscv_vfmv_f_s_f32m1_f32(redsum);
 
 #endif
+#if defined(__POWER9_VECTOR__)
+    const int np = (n & ~(GGML_BF16_STEP - 1));
+    if (np > 0) {
+        GGML_F32_VEC sum[4] = {GGML_F32_VEC_ZERO};
+        for (; i < np; i += GGML_BF16_STEP) {
+            GGML_BF16_VEC vx0 = GGML_BF16_VEC_LOAD(x + i);
+            GGML_BF16_VEC vx1 = GGML_BF16_VEC_LOAD(x + i + 8);
+            GGML_BF16_VEC vy0 = GGML_BF16_VEC_LOAD(y + i);
+            GGML_BF16_VEC vy1 = GGML_BF16_VEC_LOAD(y + i + 8);
+            GGML_BF16_FMA_LO(sum[0], vx0, vy0);
+            GGML_BF16_FMA_HI(sum[1], vx0, vy0);
+            GGML_BF16_FMA_LO(sum[2], vx1, vy1);
+            GGML_BF16_FMA_HI(sum[3], vx1, vy1);
+        }
+        GGML_F32x4_REDUCE_4(sumf, sum[0], sum[1], sum[2], sum[3]);
+    }
+#endif
+
     for (; i < n; ++i) {
         sumf += (ggml_float)(GGML_BF16_TO_FP32(x[i]) *
                              GGML_BF16_TO_FP32(y[i]));

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

@@ -21,7 +21,9 @@ struct llama_sampler_deleter {
 };
 
 struct llama_adapter_lora_deleter {
-    void operator()(llama_adapter_lora * adapter) { llama_adapter_lora_free(adapter); }
+    void operator()(llama_adapter_lora *) {
+        // llama_adapter_lora_free is deprecated
+    }
 };
 
 typedef std::unique_ptr<llama_model, llama_model_deleter> llama_model_ptr;

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

@@ -646,7 +646,8 @@ extern "C" {
 
     // Manually free a LoRA adapter
     // NOTE: loaded adapters will be free when the associated model is deleted
-    LLAMA_API void llama_adapter_lora_free(struct llama_adapter_lora * adapter);
+    LLAMA_API DEPRECATED(void llama_adapter_lora_free(struct llama_adapter_lora * adapter),
+            "adapters are now freed together with the associated model");
 
     // Get the invocation tokens if the current lora is an alora
     LLAMA_API uint64_t            llama_adapter_get_alora_n_invocation_tokens(const struct llama_adapter_lora * adapter);
@@ -1255,7 +1256,6 @@ extern "C" {
     // [EXPERIMENTAL]
     // attach a sampler to the context
     // note: prefer initializing the context with llama_context_params.samplers when possible
-    // note: changing the samplers of a context can cause graph reallocations and degraded performance
     LLAMA_API bool llama_set_sampler(struct llama_context * ctx, llama_seq_id seq_id, struct llama_sampler * smpl);
 
     // mirror of llama_sampler_i:
@@ -1395,6 +1395,33 @@ extern "C" {
                           const char ** seq_breakers,
                               size_t    num_breakers);
 
+    /// adaptive-p: select tokens near a configurable target probability over time.
+    ///
+    /// the adaptive-p sampler transforms the token probability distribution to favor tokens
+    /// that fall near a user-configurable probability target.
+    ///
+    /// internally, the sampler maintains an exponential moving average of the *ORIGINAL*
+    /// probabilities of selected tokens at each sampling step. it uses this EMA to compute an
+    /// adapted target probability at each sampling step, thus maintaining the desired target
+    /// probability over time.
+    ///
+    /// adaptive-p selects a token ID rather than just mutating candidates, so it must be last
+    /// in the sampler chain (like mirostat, dist, greedy).
+    ///
+    /// only mild truncation before this sampler is recommended. we suggest applying min-p
+    /// before adaptive-p as the only other active sampler in the chain.
+    ///
+    /// @param target select tokens near this probability (valid range 0.0 to 1.0; negative = disabled)
+    /// @param decay  EMA decay for adaptation; history ≈ 1/(1-decay) tokens (valid range 0.0 - 0.99)
+    /// @param seed   RNG seed
+    ///
+    /// ref: https://github.com/ggml-org/llama.cpp/pull/17927
+    ///
+    LLAMA_API struct llama_sampler * llama_sampler_init_adaptive_p(
+                               float   target,
+                               float   decay,
+                            uint32_t   seed);
+
     LLAMA_API struct llama_sampler * llama_sampler_init_logit_bias(
                              int32_t   n_vocab,
                              int32_t   n_logit_bias,

package/src/llama.cpp/src/llama-adapter.cpp

@@ -146,11 +146,9 @@ llama_adapter_lora_weight * llama_adapter_lora::get_weight(ggml_tensor * w) {
     return nullptr;
 }
 
-static void llama_adapter_lora_init_impl(const char * path_lora, llama_adapter_lora & adapter) {
+static void llama_adapter_lora_init_impl(llama_model & model, const char * path_lora, llama_adapter_lora & adapter) {
     LLAMA_LOG_INFO("%s: loading lora adapter from '%s' ...\n", __func__, path_lora);
 
-    llama_model & model = adapter.model;
-
     ggml_context * ctx_init;
     gguf_init_params meta_gguf_params = {
         /* .no_alloc = */ true,
@@ -413,17 +411,17 @@ static void llama_adapter_lora_init_impl(const char * path_lora, llama_adapter_l
         }
     }
 
-    // update number of nodes used
-    model.n_lora_nodes += adapter.get_n_nodes();
+    // register adapter with model
+    model.loras.insert(&adapter);
 
     LLAMA_LOG_INFO("%s: loaded %zu tensors from lora file\n", __func__, adapter.ab_map.size()*2);
 }
 
 llama_adapter_lora * llama_adapter_lora_init(llama_model * model, const char * path_lora) {
-    llama_adapter_lora * adapter = new llama_adapter_lora(*model);
+    llama_adapter_lora * adapter = new llama_adapter_lora();
 
     try {
-        llama_adapter_lora_init_impl(path_lora, *adapter);
+        llama_adapter_lora_init_impl(*model, path_lora, *adapter);
         return adapter;
     } catch (const std::exception & err) {
         LLAMA_LOG_ERROR("%s: failed to apply lora adapter: %s\n", __func__, err.what());
@@ -473,12 +471,8 @@ int32_t llama_adapter_meta_val_str_by_index(const llama_adapter_lora * adapter,
     return snprintf(buf, buf_size, "%s", it->second.c_str());
 }
 
-void llama_adapter_lora_free(llama_adapter_lora * adapter) {
-    // update number of nodes used
-    GGML_ASSERT(adapter->model.n_lora_nodes >= adapter->get_n_nodes());
-    adapter->model.n_lora_nodes -= adapter->get_n_nodes();
-
-    delete adapter;
+void llama_adapter_lora_free(llama_adapter_lora *) {
+    // deprecated: adapters are freed by llama_model's destructor
 }
 
 uint64_t llama_adapter_get_alora_n_invocation_tokens(const struct llama_adapter_lora * adapter) {

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

@@ -59,8 +59,6 @@ struct llama_adapter_lora_weight {
 };
 
 struct llama_adapter_lora {
-    llama_model & model;
-
     // map tensor name to lora_a_b
     std::unordered_map<std::string, llama_adapter_lora_weight> ab_map;
 
@@ -75,7 +73,7 @@ struct llama_adapter_lora {
     // activated lora (aLoRA)
     std::vector<llama_token> alora_invocation_tokens;
 
-    llama_adapter_lora(llama_model & model) : model(model) {}
+    llama_adapter_lora() = default;
     ~llama_adapter_lora() = default;
 
     llama_adapter_lora_weight * get_weight(ggml_tensor * w);