@fugood/llama.node 1.4.14 → 1.5.0-rc.0

package/src/llama.cpp/common/common.h

@@ -80,6 +80,7 @@ int32_t cpu_get_num_math();
 //
 
 enum llama_example {
+    LLAMA_EXAMPLE_BATCHED,
     LLAMA_EXAMPLE_DEBUG,
     LLAMA_EXAMPLE_COMMON,
     LLAMA_EXAMPLE_SPECULATIVE,
@@ -118,6 +119,7 @@ enum common_sampler_type {
     COMMON_SAMPLER_TYPE_INFILL      = 9,
     COMMON_SAMPLER_TYPE_PENALTIES   = 10,
     COMMON_SAMPLER_TYPE_TOP_N_SIGMA = 11,
+    COMMON_SAMPLER_TYPE_ADAPTIVE_P  = 12,
 };
 
 // dimensionality reduction methods, used by cvector-generator
@@ -165,32 +167,34 @@ enum common_params_sampling_config : uint64_t {
 struct common_params_sampling {
     uint32_t seed = LLAMA_DEFAULT_SEED; // the seed used to initialize llama_sampler
 
-    int32_t n_prev             = 64;    // number of previous tokens to remember
-    int32_t n_probs            = 0;     // if greater than 0, output the probabilities of top n_probs tokens.
-    int32_t min_keep           = 0;     // 0 = disabled, otherwise samplers should return at least min_keep tokens
-    int32_t top_k              = 40;    // <= 0 to use vocab size
-    float   top_p              = 0.95f; // 1.0 = disabled
-    float   min_p              = 0.05f; // 0.0 = disabled
-    float   xtc_probability    = 0.00f; // 0.0 = disabled
-    float   xtc_threshold      = 0.10f; // > 0.5 disables XTC
-    float   typ_p              = 1.00f; // typical_p, 1.0 = disabled
-    float   temp               = 0.80f; // <= 0.0 to sample greedily, 0.0 to not output probabilities
-    float   dynatemp_range     = 0.00f; // 0.0 = disabled
-    float   dynatemp_exponent  = 1.00f; // controls how entropy maps to temperature in dynamic temperature sampler
-    int32_t penalty_last_n     = 64;    // last n tokens to penalize (0 = disable penalty, -1 = context size)
-    float   penalty_repeat     = 1.00f; // 1.0 = disabled
-    float   penalty_freq       = 0.00f; // 0.0 = disabled
-    float   penalty_present    = 0.00f; // 0.0 = disabled
-    float   dry_multiplier     = 0.0f;  // 0.0 = disabled;      DRY repetition penalty for tokens extending repetition:
-    float   dry_base           = 1.75f; // 0.0 = disabled;      multiplier * base ^ (length of sequence before token - allowed length)
-    int32_t dry_allowed_length = 2;     // tokens extending repetitions beyond this receive penalty
-    int32_t dry_penalty_last_n = -1;    // how many tokens to scan for repetitions (0 = disable penalty, -1 = context size)
-    int32_t mirostat           = 0;     // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
-    float   top_n_sigma        = -1.00f;// -1.0 = disabled
-    float   mirostat_tau       = 5.00f; // target entropy
-    float   mirostat_eta       = 0.10f; // learning rate
+    int32_t n_prev             = 64;     // number of previous tokens to remember
+    int32_t n_probs            = 0;      // if greater than 0, output the probabilities of top n_probs tokens.
+    int32_t min_keep           = 0;      // 0 = disabled, otherwise samplers should return at least min_keep tokens
+    int32_t top_k              = 40;     // <= 0 to use vocab size
+    float   top_p              = 0.95f;  // 1.0 = disabled
+    float   min_p              = 0.05f;  // 0.0 = disabled
+    float   xtc_probability    = 0.00f;  // 0.0 = disabled
+    float   xtc_threshold      = 0.10f;  // > 0.5 disables XTC
+    float   typ_p              = 1.00f;  // typical_p, 1.0 = disabled
+    float   temp               = 0.80f;  // <= 0.0 to sample greedily, 0.0 to not output probabilities
+    float   dynatemp_range     = 0.00f;  // 0.0 = disabled
+    float   dynatemp_exponent  = 1.00f;  // controls how entropy maps to temperature in dynamic temperature sampler
+    int32_t penalty_last_n     = 64;     // last n tokens to penalize (0 = disable penalty, -1 = context size)
+    float   penalty_repeat     = 1.00f;  // 1.0 = disabled
+    float   penalty_freq       = 0.00f;  // 0.0 = disabled
+    float   penalty_present    = 0.00f;  // 0.0 = disabled
+    float   dry_multiplier     = 0.0f;   // 0.0 = disabled;      DRY repetition penalty for tokens extending repetition:
+    float   dry_base           = 1.75f;  // 0.0 = disabled;      multiplier * base ^ (length of sequence before token - allowed length)
+    int32_t dry_allowed_length = 2;      // tokens extending repetitions beyond this receive penalty
+    int32_t dry_penalty_last_n = -1;     // how many tokens to scan for repetitions (0 = disable penalty, -1 = context size)
+    float   adaptive_target    = -1.0f;  // select tokens near this probability (valid range 0.0 to 1.0; negative = disabled)
+    float   adaptive_decay     = 0.90f;  // EMA decay for adaptation; history ≈ 1/(1-decay) tokens (0.0 - 0.99)
+    int32_t mirostat           = 0;      // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
+    float   top_n_sigma        = -1.00f; // -1.0 = disabled
+    float   mirostat_tau       = 5.00f;  // target entropy
+    float   mirostat_eta       = 0.10f;  // learning rate
     bool    ignore_eos         = false;
-    bool    no_perf            = false; // disable performance metrics
+    bool    no_perf            = false;  // disable performance metrics
     bool    timing_per_token   = false;
 
     uint64_t user_sampling_config = 0; // bitfield to track user-specified samplers
@@ -476,6 +480,7 @@ struct common_params {
     int32_t timeout_write     = timeout_read; // http write timeout in seconds
     int32_t n_threads_http    = -1;           // number of threads to process HTTP requests (TODO: support threadpool)
     int32_t n_cache_reuse     = 0;            // min chunk size to reuse from the cache via KV shifting
+    bool    cache_prompt      = true;         // whether to enable prompt caching
     int32_t n_ctx_checkpoints = 8;            // max number of context checkpoints per slot
     int32_t cache_ram_mib     = 8192;         // -1 = no limit, 0 - disable, 1 = 1 MiB, etc.
 

package/src/llama.cpp/common/debug.cpp

@@ -0,0 +1,165 @@
+#include "debug.h"
+
+#include "log.h"
+
+#include <cmath>
+#include <string>
+
+static std::string common_ggml_ne_string(const ggml_tensor * t) {
+    std::string str;
+    for (int i = 0; i < GGML_MAX_DIMS; ++i) {
+        str += std::to_string(t->ne[i]);
+        if (i + 1 < GGML_MAX_DIMS) {
+            str += ", ";
+        }
+    }
+    return str;
+}
+
+static float common_ggml_get_float_value(const uint8_t * data,
+                           ggml_type       type,
+                           const size_t *  nb,
+                           size_t          i0,
+                           size_t          i1,
+                           size_t          i2,
+                           size_t          i3) {
+    size_t i = i3 * nb[3] + i2 * nb[2] + i1 * nb[1] + i0 * nb[0];
+    float  v;
+    if (type == GGML_TYPE_F16) {
+        v = ggml_fp16_to_fp32(*(const ggml_fp16_t *) &data[i]);
+    } else if (type == GGML_TYPE_F32) {
+        v = *(const float *) &data[i];
+    } else if (type == GGML_TYPE_I64) {
+        v = (float) *(const int64_t *) &data[i];
+    } else if (type == GGML_TYPE_I32) {
+        v = (float) *(const int32_t *) &data[i];
+    } else if (type == GGML_TYPE_I16) {
+        v = (float) *(const int16_t *) &data[i];
+    } else if (type == GGML_TYPE_I8) {
+        v = (float) *(const int8_t *) &data[i];
+    } else if (type == GGML_TYPE_BF16) {
+        v = ggml_bf16_to_fp32(*(const ggml_bf16_t *) &data[i]);
+    } else {
+        GGML_ABORT("fatal error");
+    }
+    return v;
+}
+
+template <bool abort>
+void common_debug_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne, const size_t * nb, int64_t n) {
+    GGML_ASSERT(n > 0);
+    float sum = 0;
+    for (int64_t i3 = 0; i3 < ne[3]; i3++) {
+        for (int64_t i2 = 0; i2 < ne[2]; i2++) {
+            for (int64_t i1 = 0; i1 < ne[1]; i1++) {
+                for (int64_t i0 = 0; i0 < ne[0]; i0++) {
+                    const float v = common_ggml_get_float_value(data, type, nb, i0, i1, i2, i3);
+                    sum += v;
+                }
+            }
+        }
+    }
+    for (int64_t i3 = 0; i3 < ne[3]; i3++) {
+        LOG_ERR("                                     [\n");
+        for (int64_t i2 = 0; i2 < ne[2]; i2++) {
+            if (i2 == n && ne[2] > 2 * n) {
+                LOG_ERR("                                      ..., \n");
+                i2 = ne[2] - n;
+            }
+            LOG_ERR("                                      [\n");
+            for (int64_t i1 = 0; i1 < ne[1]; i1++) {
+                if (i1 == n && ne[1] > 2 * n) {
+                    LOG_ERR("                                       ..., \n");
+                    i1 = ne[1] - n;
+                }
+                LOG_ERR("                                       [");
+                for (int64_t i0 = 0; i0 < ne[0]; i0++) {
+                    if (i0 == n && ne[0] > 2 * n) {
+                        LOG_ERR("..., ");
+                        i0 = ne[0] - n;
+                    }
+                    const float v = common_ggml_get_float_value(data, type, nb, i0, i1, i2, i3);
+                    LOG_ERR("%12.4f", v);
+                    if (i0 < ne[0] - 1) {
+                        LOG_ERR(", ");
+                    }
+                }
+                LOG_ERR("],\n");
+            }
+            LOG_ERR("                                      ],\n");
+        }
+        LOG_ERR("                                     ]\n");
+        LOG_ERR("                                     sum = %f\n", sum);
+    }
+
+    if constexpr (abort) {
+        if (std::isnan(sum)) {
+            LOG_ERR("encountered NaN - aborting\n");
+            exit(0);
+        }
+    }
+}
+
+/**
+ * GGML operations callback during the graph execution.
+ *
+ * @param t current tensor
+ * @param ask when ask is true, the scheduler wants to know if we are interested in data from this tensor
+ *            if we return true, a follow-up call will be made with ask=false in which we can do the actual collection.
+ *            see ggml_backend_sched_eval_callback
+ * @param user_data user data to pass at each call back
+ * @return true to receive data or continue the graph, false otherwise
+ */
+template <bool abort_on_nan> bool common_debug_cb_eval(struct ggml_tensor * t, bool ask, void * user_data) {
+    auto * cb_data = (base_callback_data *) user_data;
+
+    const struct ggml_tensor * src0 = t->src[0];
+    const struct ggml_tensor * src1 = t->src[1];
+
+    if (ask) {
+        return true;  // Always retrieve data
+    }
+
+    bool matches_filter = cb_data->tensor_filters.empty();
+
+    if (!matches_filter) {
+        for (const auto & filter : cb_data->tensor_filters) {
+            if (std::regex_search(t->name, filter)) {
+                matches_filter = true;
+                break;
+            }
+        }
+    }
+
+    char src1_str[128] = { 0 };
+    if (src1) {
+        snprintf(src1_str, sizeof(src1_str), "%s{%s}", src1->name, common_ggml_ne_string(src1).c_str());
+    }
+
+    if (matches_filter) {
+        LOG_ERR("%s: %24s = (%s) %10s(%s{%s}, %s}) = {%s}\n", __func__, t->name, ggml_type_name(t->type),
+                ggml_op_desc(t), src0->name, common_ggml_ne_string(src0).c_str(), src1 ? src1_str : "",
+                common_ggml_ne_string(t).c_str());
+    }
+
+    const bool is_host = ggml_backend_buffer_is_host(t->buffer);
+
+    if (!is_host) {
+        auto n_bytes = ggml_nbytes(t);
+        cb_data->data.resize(n_bytes);
+        ggml_backend_tensor_get(t, cb_data->data.data(), 0, n_bytes);
+    }
+
+    if (!ggml_is_quantized(t->type) && matches_filter) {
+        uint8_t * data = is_host ? (uint8_t *) t->data : cb_data->data.data();
+        common_debug_print_tensor<abort_on_nan>(data, t->type, t->ne, t->nb, 3);
+    }
+
+    return true;
+}
+
+// Explicit template instantiations
+template bool common_debug_cb_eval<false>(ggml_tensor *, bool, void *);
+template bool common_debug_cb_eval<true>(ggml_tensor *, bool, void *);
+template void common_debug_print_tensor<false>(uint8_t *, ggml_type, const int64_t *, const size_t *, int64_t);
+template void common_debug_print_tensor<true>(uint8_t *, ggml_type, const int64_t *, const size_t *, int64_t);

package/src/llama.cpp/common/debug.h

@@ -0,0 +1,43 @@
+#pragma once
+#include "common.h"
+#include <string>
+#include <vector>
+#include <regex>
+
+// common debug functions and structs
+
+// Print a tensor's detailed data
+// data - the tensor's data in byte format
+// type - the tensor's quantization type
+// ne   - the tensor dimensions array
+// nb   - the tensor strides array
+// n    - the number of rows/columns to fully print
+template <bool abort_on_nan> void common_debug_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne, const size_t * nb, int64_t n);
+
+// Intended to use as callback for ggml_backend_sched_eval_callback
+// prints tensors that are processed in the computation graph
+// by default prints all tensors, but can be configured by creating a `base_callback_data` instance with
+// non-empty filter_patterns. See examples/debug.ccp for possible usage patterns
+// The template parameter determins whether an error should be thrown whenever a NaN is encountered
+// in a tensor (useful for stopping debug sessions on first erroneous tensor)
+// The callback data will be passed as the third parameter (user_data)
+template <bool abort_on_nan> bool common_debug_cb_eval(struct ggml_tensor * t, bool ask, void * user_data);
+struct base_callback_data {
+    std::vector<uint8_t>    data;
+    std::vector<std::regex> tensor_filters;
+
+    base_callback_data() = default;
+
+    base_callback_data(common_params & params, const std::vector<std::string> & filter_patterns) {
+        for (const auto & pattern : filter_patterns) {
+            try {
+                std::string anchored_pattern = "^" + pattern;
+                tensor_filters.emplace_back(anchored_pattern, std::regex::optimize);
+            } catch (const std::regex_error & e) {
+                throw std::runtime_error("Invalid regex pattern '" + pattern + "': " + e.what());
+            }
+        }
+        params.cb_eval           = common_debug_cb_eval<false>;
+        params.cb_eval_user_data = this;
+    }
+};