diff --git a/common/common.h b/common/common.h index 2adb310b83fe..1e68bad8eeab 100644 --- a/common/common.h +++ b/common/common.h @@ -695,10 +695,11 @@ struct common_params { int32_t i_chunk = 0; // start processing from this chunk int8_t imat_dat = 0; // whether the legacy imatrix.dat format should be output (gguf <= 0 < dat) - bool process_output = false; // collect data for the output tensor - bool compute_ppl = true; // whether to compute perplexity - bool show_statistics = false; // show imatrix statistics per tensor - bool parse_special = false; // whether to parse special tokens during imatrix tokenization + bool process_output = false; // collect data for the output tensor + bool compute_ppl = true; // whether to compute perplexity + bool show_statistics = false; // show imatrix statistics per tensor + bool activation_statistics = false; // generate data to calculate activation based statistics + bool parse_special = false; // whether to parse special tokens during imatrix tokenization // cvector-generator params int n_pca_batch = 100; diff --git a/common/imatrix-loader.cpp b/common/imatrix-loader.cpp index efe9aecee3f8..f48650edfd98 100644 --- a/common/imatrix-loader.cpp +++ b/common/imatrix-loader.cpp @@ -110,33 +110,41 @@ bool common_imatrix_load(const std::string & fname, common_imatrix & imatrix) { } } - imatrix.has_metadata = (datasets_key != -1 && chunk_count_key != -1 && chunk_size_key != -1); - imatrix.chunk_count = (chunk_count_key != -1) ? gguf_get_val_u32(ctx_gguf, chunk_count_key) : 0; - imatrix.chunk_size = (chunk_size_key != -1) ? gguf_get_val_u32(ctx_gguf, chunk_size_key) : 0; + imatrix.has_metadata = datasets_key != -1 && chunk_count_key != -1 && chunk_size_key != -1; + imatrix.chunk_count = chunk_count_key != -1 ? gguf_get_val_u32(ctx_gguf, chunk_count_key) : 0; + imatrix.chunk_size = chunk_size_key != -1 ? gguf_get_val_u32(ctx_gguf, chunk_size_key) : 0; + const std::string in_sum_suffix{ ".in_sum" }; const std::string in_sum2_suffix{ ".in_sum2" }; const std::string counts_suffix{ ".counts" }; - std::map> sums_counts_for; + struct sum_tensors { + struct ggml_tensor * in_sum = nullptr; + struct ggml_tensor * in_sum2 = nullptr; + struct ggml_tensor * counts = nullptr; + }; + std::map sums_counts_for; for (struct ggml_tensor * cur = ggml_get_first_tensor(ctx); cur; cur = ggml_get_next_tensor(ctx, cur)) { std::string name = cur->name; - if (name.empty()) { continue; } - if (string_remove_suffix(name, in_sum2_suffix)) { - sums_counts_for[std::move(name)].first = cur; + if (string_remove_suffix(name, in_sum_suffix)) { + sums_counts_for[std::move(name)].in_sum = cur; + } else if (string_remove_suffix(name, in_sum2_suffix)) { + sums_counts_for[std::move(name)].in_sum2 = cur; } else if (string_remove_suffix(name, counts_suffix)) { - sums_counts_for[std::move(name)].second = cur; + sums_counts_for[std::move(name)].counts = cur; } } for (const auto & sc : sums_counts_for) { const std::string & name = sc.first; - const struct ggml_tensor * in_sum2 = sc.second.first; - const struct ggml_tensor * counts = sc.second.second; + const struct ggml_tensor * in_sum = sc.second.in_sum; + const struct ggml_tensor * in_sum2 = sc.second.in_sum2; + const struct ggml_tensor * counts = sc.second.counts; - if (!in_sum2 || !counts) { + if (!in_sum2 || !counts || (in_sum != nullptr && ggml_nelements(in_sum) != ggml_nelements(in_sum2))) { LOG_ERR("%s: mismatched sums and counts for %s\n", __func__, name.c_str()); gguf_free(ctx_gguf); ggml_free(ctx); @@ -157,6 +165,13 @@ bool common_imatrix_load(const std::string & fname, common_imatrix & imatrix) { for (int64_t j = 0; j < ncounts; ++j) { e.counts[j] = std::lround(((const float *) counts->data)[j]); } + + if (in_sum && ggml_nelements(in_sum) == nval) { + e.activations.resize(nval); + for (int64_t j = 0; j < nval; ++j) { + e.activations[j] = ((const float *) in_sum->data)[j]; + } + } } gguf_free(ctx_gguf); diff --git a/common/imatrix-loader.h b/common/imatrix-loader.h index ed00d724ac85..023383130d37 100644 --- a/common/imatrix-loader.h +++ b/common/imatrix-loader.h @@ -8,9 +8,11 @@ inline constexpr const char * LLM_KV_IMATRIX_DATASETS = "imatrix.datasets"; inline constexpr const char * LLM_KV_IMATRIX_CHUNK_COUNT = "imatrix.chunk_count"; inline constexpr const char * LLM_KV_IMATRIX_CHUNK_SIZE = "imatrix.chunk_size"; +inline constexpr const char * LLM_KV_IMATRIX_STATS_SCHEMA = "imatrix.stats_schema"; struct common_imatrix_entry { std::vector sums; + std::vector activations; std::vector counts; }; diff --git a/tools/imatrix/README.md b/tools/imatrix/README.md index 4cbe4fd0cf75..a92f861126ec 100644 --- a/tools/imatrix/README.md +++ b/tools/imatrix/README.md @@ -20,13 +20,13 @@ The parameters in square brackets are optional and have the following meaning: * `-lv | --verbosity` specifies the verbosity level. If set to `0`, no output other than the perplexity of the processed chunks will be generated. If set to `1`, each time the results are saved a message is written to `stderr`. If `>=2`, a message is output each time data is collected for any tensor. Default verbosity level is `1`. * `-o | --output-file` specifies the name of the file where the computed data will be stored. If missing `imatrix.gguf` is used. * `-ofreq | --output-frequency` specifies how often the so far computed result is saved to disk. Default is 10 (i.e., every 10 chunks) -* `--output-format` specifies the output format of the generated imatrix file. Either "gguf", or "dat" (the legacy format). Defaults to "gguf". +* `--output-format` specifies the output format of the generated imatrix file. Either `gguf`, or `dat` (the legacy format). Defaults to `gguf`. * `--save-frequency` specifies how often to save a copy of the imatrix in a separate file. Default is 0 (i.e., never) * `--process-output` specifies if data will be collected for the `output.weight` tensor. Typically, it is better not to utilize the importance matrix when quantizing `output.weight`, so this is set to `false` by default. * `--in-file` one or more existing imatrix files to load and combine. Useful for merging files from multiple runs/datasets. * `--parse-special` enables parsing of special tokens (e.g., `<|im_start|>` in some models). Useful for models with custom tokenizers. * `--chunk | --from-chunk` to skip the first `n` chunks of tokens from the input data. Useful for resuming or skipping initial low-quality data. -* `--chunks` maximum number of chunks to process. Default is -1 for all available chunks. +* `--chunks` maximum number of chunks to process. Default is `-1` for all available chunks. * `--no-ppl` disables the calculation of perplexity for the processed chunks. Useful if you want to speed up the processing and do not care about perplexity. * `--show-statistics` displays imatrix file's statistics. @@ -70,29 +70,47 @@ Recent versions of `llama-imatrix` store data in GGUF format by default. For the ``` ```bash -# analyse imatrix file and display summary statistics instead of running inference -./llama-imatrix --in-file imatrix.gguf --show-statistics +# analyze imatrix file and display summary statistics instead of running inference +./llama-imatrix -m ggml-model-f16.gguf --in-file imatrix.gguf --show-statistics ``` -`--show-statistics` will display the following statistics: +## Statistics +Please note that if a value lacks statistical interpretability, **nan** will be shown instead. #### Per tensor - -* Σ(Act²): sum of all squared activations (the importance scores) -* Min & Max: minimum and maximum squared activations values -* μ & σ: Squared activations' mean and standard deviation -* % Active: proportion of elements whose average squared activation exceeds a small threshold (1e-5). Helpful to determine how alive/dormant the tensor is during inference -* N: number of squared activations -* Entropy: entropy of the squared activation distribution, in bits (standard Shannon entropy measurement) $S = -\sum_{i=1}^N p_i \log_2 p_i$ -* E (norm): Normalized entropy. $E(norm)=\frac{-\sum_{i=1}^N p_i \log_2 p_i}{log_2 N}$. These two metrics can be used to determine how well a prompt "exercises" the model's capabilities -* ZD Score: z-score distribution as described in _3.1 Layer Importance Scores_ of [Layer-Wise Quantization](https://arxiv.org/abs/2406.17415) -* CosSim: cosine similarity with respect to the previous layer's tensor. Useful to determine how similar the squared activations of the current layer are to the previous layer's squared activations. +Statistical properties of a single tensor's average activation or activation energy (squared magnitude). + +* **Mean / StdDev**: $\mu = \frac{1}{N} \sum v_i$ and $\sigma = \sqrt{\frac{1}{N} \sum (v_i - \mu)^2}$ + - Establishes the baseline distribution of the tensor's outputs. Low variance means the tensor outputs a mostly constant projection; high variance implies high information density across dimensions. +* **Skewness & Kurtosis**: $skew = \frac{\frac{1}{N} \sum (v_i - \mu)^3}{\sigma^3}$ and $kurt = \frac{\frac{1}{N} \sum (v_i - \mu)^4}{\sigma^4} - 3.0$ + - Skewness measures the asymmetry of a distribution around its mean. Kurtosis measures the "tailedness" of the feature activations. A high kurtosis indicates a highly sparse/heavy-tailed activation distribution (e.g., outlier features). High-kurtosis tensors typically require higher precision quantization to prevent outlier degradation. +* **H Norm**: $H = -\sum_{i} P_i \log_2(P_i)$, where $P_i = \frac{v_{val_i}}{\sum v_{val_i}}$ + - Shannon Entropy normalized over log₂(N). Used to determine how well a prompt "exercises" the model's capabilities. Higher values indicate more uniform distribution of activations. Every neuron is firing equally; hard to prune. +* **$\sum E[A^2]$**: $\sum E[x_i^2]$ + - The sum of squares of activations (Energy) for the tensor. Tensors with high "energy" contribute most to the final output. Quantization errors here propagate strongly. These tensors usually need higher precision. + +#### Intra-layer +These statistics compare identical tensor between the current layer $L$ and the previsou layer $L-1$ (e.g., `blk.1.attn_v` vs `blk.0.attn_v`). + +* **Gain**: $G = \frac{\sqrt{\sum C_i^2 / N_{curr}}}{\sqrt{\sum P_i^2 / N_{prev}}}$ + - Indicates if a layer acts as an "amplifier" ($G > 1$) or a "dampener" ($G < 1$). +* **L2 Distance**: $L2 = \sqrt{ \sum (C_i - P_i)^2 }$ where $C$ is the current layer and $P$ is the previous layer's tensor. + - Measures absolute representational shift. Huge leaps in L2 distance indicate that a layer fundamentally transforms the hidden states. +* **Pearson Correlation Coefficient (PCC)**: $r = \frac{\sum (C_i - \bar{C})(P_i - \bar{P})}{\sqrt{\sum (C_i - \bar{C})^2} \sqrt{\sum (P_i - \bar{P})^2}}$ + - Similar to Cosine Similarity, but invariant to scalar or offset biases (centers the data first). Highly correlated adjacent layers signify structural repetition. +* **Covariance (Cov)**: $\frac{1}{N}\sum (c_i-\bar c)(p_i-\bar p)$ + - The **unnormalized covariance** between current and previous layer activations. Captures both the correlation structure and the magnitude of the joint variation. Large absolute covariance indicates the layers are jointly processing strong, correlated signals. #### Per layer - -Weighted averages of Σ(Act²), ZD Score and CosSim are also calculated. - -#### Important note on the computed Statistics - -When using these statistics, please note that they are computed on the squared activations, **not on the actual (raw) activations**. -Whilst the results are still useful, they're less reliable than using the raw values, and in the case of the cosine similarity, could be misleading if the tensor contains opposite vectors. +Aggregated metrics per block/layer: + +* **∑ E[A²]:** Total energy of the layer's concatenated tensors. Indicates the layer's overall contribution amplitude. +* **Gain**: Indicates if a layer acts as an "amplifier" ($G > 1$) or a "dampener" ($G < 1$). +* **L₂ Distance:** Euclidean Distance of the layer's concatenated tensors from the previous layer’s. Global measure of transformation magnitude. +* **CosSim**: $\text{CosSim}_{Layer} = \frac{\sum_{\text{tensors}} (\text{Dot Prod})}{\sqrt{\sum_{\text{tensors}} (\text{Norm1}^2)} \sqrt{\sum_{\text{tensors}} (\text{Norm2}^2)}}$ + - Cosine Similarity of the current layer's concatenated tensors with the previous layer. +* **PCC**: $\text{PCC}_{Layer} = \frac{\sum \text{Covariance}^2}{\sqrt{\sum \text{Var}_{curr}} \sqrt{\sum \text{Var}_{prev}}}$ + - Average Pearson Correlation of the tensors in the layer. +* **Cov**: The **unnormalized covariance** between current layer's concatenated tensors and the previous layer. + +More information is available in https://github.com/ggml-org/llama.cpp/pull/14891 diff --git a/tools/imatrix/imatrix.cpp b/tools/imatrix/imatrix.cpp index 3431a4eca84b..0cdd1208fdd7 100644 --- a/tools/imatrix/imatrix.cpp +++ b/tools/imatrix/imatrix.cpp @@ -1,30 +1,27 @@ #include "arg.h" #include "common.h" +#include "gguf.h" #include "imatrix-loader.h" -#include "log.h" #include "llama.h" -#include "gguf.h" +#include "log.h" -#include #include -#include +#include #include -#include -#include -#include -#include -#include -#include #include +#include +#include #include -#include -#include -#include #if defined(_MSC_VER) #pragma warning(disable: 4244 4267) // possible loss of data #endif +#if defined(__linux__) || defined(__APPLE__) || defined(__FreeBSD__) +#include +#include +#endif + static void print_usage(int, char ** argv) { LOG("\nexample usage:\n"); LOG("\n %s \\\n" @@ -36,6 +33,7 @@ static void print_usage(int, char ** argv) { } struct Stats { + std::vector activations; std::vector values; std::vector counts; }; @@ -43,16 +41,28 @@ struct Stats { struct tensor_statistics { std::string tensor; Stats stats; - float total_sqract = 0.0f; - float mean_sqract = 0.0f; - float max_sqract = 0.0f; - float min_sqract = 0.0f; - int elements = 0; - float stddev = 0.0f; - float active = 0.0f; - float entropy = 0.0f; - float zd = 0.0f; - float cossim = 0.0f; + double sum = 0.0f; + float mean = 0.0f; + int64_t elements = 0; + float std_deviation = 0.0f; + float skewness = 0.0f; + float kurtosis = 0.0f; + float gain = std::numeric_limits::quiet_NaN(); + float entropy = 0.0f; + float l2_dist = std::numeric_limits::quiet_NaN(); + float cossim = std::numeric_limits::quiet_NaN(); + float pearson = std::numeric_limits::quiet_NaN(); + float covariance = std::numeric_limits::quiet_NaN(); + double cov_sum = 0.0; + double var_c_sum = 0.0; + double var_p_sum = 0.0; + double dot_prod = 0.0; + double norm1_sq = 0.0; + double norm2_sq = 0.0; + double l2_dist_sq = 0.0; + double sum_prev = 0.0; + int64_t elements_prev = 0; + int64_t n_features = 0; }; class IMatrixCollector { @@ -94,6 +104,9 @@ static std::string filter_tensor_name(const char * name) { } static void process_tensor_name(const std::string & input, std::string & layer, std::string & tensor) { + layer.clear(); + tensor.clear(); + std::vector name; std::istringstream stream(input); std::string item; @@ -109,115 +122,374 @@ static void process_tensor_name(const std::string & input, std::string & layer, } for (size_t i = 0; i < name.size(); ++i) { if (name[i] == "weight" && i > 0) { - tensor = name[i - 1]; + for (size_t j = 0; j < name.size(); ++j) { + if (name[j] == "blk") { + j+=2; + continue; + } + if (j == i) { break; } + if (!tensor.empty()) { tensor += "."; } + tensor += name[j]; + } break; } } - if (tensor.empty()) { - tensor = input; - } - if (layer.empty()) { - layer = "-"; - } + if (tensor.empty()) { tensor = input; } + if (layer.empty()) { layer = "-"; } } -static void compute_statistics(std::vector & tstats, const std::string & name, const Stats & e) { - if (e.values.size() % e.counts.size() != 0) { - LOG_ERR("%s: activation size mismatch for tensor %s (%zu vs %zu)\n", __func__, name.c_str(), e.counts.size(), e.values.size()); - return; - } - if (e.counts.empty()) { - LOG_ERR("%s: there are no activations for tensor %s. The imatrix may be suboptimal\n", __func__, name.c_str()); - return; - } +static std::vector compute_tensor_averages(const Stats & tstats) { + if (tstats.counts.empty()) { return {}; } - const int n_mat = e.counts.size(); - const int row_size = e.values.size() / n_mat; + const size_t n_mat = tstats.counts.size(); + const size_t len = !tstats.activations.empty() ? tstats.activations.size() : tstats.values.size(); + if (len == 0 || n_mat == 0 || len % n_mat != 0) { return {}; } - std::vector activations; - activations.reserve(e.values.size()); + const size_t row = len / n_mat; + std::vector vec(len, std::numeric_limits::quiet_NaN()); - for (int i = 0; i < n_mat; ++i) { - if (e.counts[i] == 0) { - LOG_DBG("%s: skipping tensor %s due to zero count at index %d\n", __func__, name.c_str(), i); - continue; - } - for (int j = 0; j < row_size; ++j) { - activations.push_back(e.values[i*row_size + j] / e.counts[i]); - } + bool has_valid = false; + const bool use_activations = !tstats.activations.empty(); + + for (size_t m = 0; m < n_mat; ++m) { + const auto c = (float) tstats.counts[m]; + const size_t off = m * row; + + if (c <= 0.0f) { continue; } + + has_valid = true; + const float scale = 1.0f / c; + const float * src = use_activations ? &tstats.activations[off] : &tstats.values[off]; + float * dst = & vec[off]; + + for (size_t j = 0; j < row; ++j) { dst[j] = src[j] * scale; } } - if (activations.empty()) { - LOG_ERR("%s: all counts are zero for tensor %s, skipping statistics computation\n", __func__, name.c_str()); - return; + if (!has_valid) { return {}; } + return vec; +} + +static bool compute_vector_statistics(std::vector & tstats, const std::string & name, const Stats & e, bool & legacy) { + constexpr auto fnan = std::numeric_limits::quiet_NaN(); + legacy = e.activations.empty(); + const size_t n_mat = e.counts.size(); + const size_t len = legacy ? e.values.size() : e.activations.size(); + + if (n_mat == 0 || len == 0 || len % n_mat != 0) { + LOG_ERR("%s: data size mismatch or empty for tensor %s\n", __func__, name.c_str()); + return false; + } + if (!legacy && e.values.size() != len) { + LOG_ERR("%s: activations/values size mismatch for %s\n", __func__, name.c_str()); + return false; } - const float act_total = std::accumulate(activations.begin(), activations.end(), 0.0f); - const float act_max = *std::max_element(activations.begin(), activations.end()); - const float act_min = *std::min_element(activations.begin(), activations.end()); - const float act_mean = act_total / activations.size(); - const float act_sqr_total = std::inner_product(activations.begin(), activations.end(), activations.begin(), 0.0f); - const float act_var = (act_sqr_total / activations.size()) - (act_mean * act_mean); - const float act_dev = std::sqrt(std::max(0.0f, act_var)); - float threshold = 1e-5f; - const int inactive_count = std::count_if(activations.begin(), activations.end(), - [threshold](const float v) { return fabsf(v) <= threshold; }); - const float active_ratio = 1 - static_cast(inactive_count) / activations.size(); - - float entropy = 0; - if (act_total > 0) { - for (const auto act : activations) { - if (const float p = act / act_total; p > 0) { - entropy -= p * std::log2(p); - } + const size_t row_size = len / n_mat; + double sum = 0.0; + double mean = 0.0; + double sum_sq_diff = 0.0; + double sum_cu_diff = 0.0; + double sum_qd_diff = 0.0; + double sum_energy = 0.0; + size_t valid_n = 0; + + // Mean + for (size_t i = 0; i < n_mat; ++i) { + const auto c = (float)e.counts[i]; + if (c <= 0.0f) { continue; } + + const double inv_c = 1.0 / (double)c; + const size_t off = i * row_size; + + for (size_t j = 0; j < row_size; ++j) { + const double v_act = legacy ? 0.0 : (double)e.activations[off + j] * inv_c; + const double v_val = (double)e.values[off + j] * inv_c; + const double v = legacy ? v_val : v_act; // Use activation average for non-legacy + if (!std::isfinite(v) || !std::isfinite(v_val)) { continue; } + + sum += v_val; + valid_n++; + const double delta = v - mean; + mean += delta / (double)valid_n; + + if (v_val > 0.0) { sum_energy += v_val; } } } - int z_score = 0; - if (act_dev > 0.0f) { - for (const auto act : activations) { - if (const float p = (act - act_mean) / act_dev; p > 1) { - z_score++; + if (valid_n == 0) { return false; } + + float std_deviation = 0.0f; + float entropy = 0.0f; + + // Std Dev, Skew, Kurtosis, Entropy + const double inv_sum_energy = sum_energy > 0.0 ? 1.0 / sum_energy : 0.0; + const double log2_inv = 1.0 / std::log(2.0); + + for (size_t i = 0; i < n_mat; ++i) { + const auto c = (float)e.counts[i]; + if (c <= 0.0f) { continue; } + const double inv_c = 1.0 / (double)c; + const size_t off = i * row_size; + + for (size_t j = 0; j < row_size; ++j) { + const double v_act = legacy ? 0.0 : (double)e.activations[off + j] * inv_c; + const double v_val = (double)e.values[off + j] * inv_c; + const double v = legacy ? v_val : v_act; + if (!std::isfinite(v) || !std::isfinite(v_val)) { continue; } + const double diff = v - mean; + + sum_sq_diff += diff * diff; + sum_cu_diff += diff * diff * diff; + sum_qd_diff += diff * diff * diff * diff; + + // Entropy (Distribution of Energy) + if (inv_sum_energy > 0.0) { + const double v_energy = (double)e.values[off + j] * inv_c; + const double p = std::max(0.0, v_energy) * inv_sum_energy; + if (p > 1e-10) { entropy -= (float)(p * std::log(p) * log2_inv); } } } } + const double variance = valid_n > 1 ? sum_sq_diff / (double)valid_n : 0.0; + std_deviation = std::sqrt((float)std::max(variance, 0.0)); + float skewness = 0.0f; + float kurtosis = 0.0f; + if (std_deviation > 1e-10f) { + const double m2 = sum_sq_diff / (double)valid_n; + skewness = (float)(sum_cu_diff / (double)valid_n / (m2 * std::sqrt(m2))); + kurtosis = (float)(sum_qd_diff / (double)valid_n / (m2 * m2) - 3.0); + } + auto & ts = tstats.emplace_back(); - ts.tensor = name; - ts.stats = e; - ts.total_sqract = act_total; - ts.mean_sqract = act_mean; - ts.max_sqract = act_max; - ts.min_sqract = act_min; - ts.elements = static_cast(activations.size()); - ts.stddev = act_dev; - ts.active = active_ratio; - ts.entropy = entropy; - ts.zd = static_cast(z_score) / ts.elements; + ts.tensor = name; + ts.stats = e; + ts.sum = sum; + ts.mean = (float)mean; + ts.elements = (int64_t)valid_n; + ts.std_deviation = std_deviation; + ts.skewness = skewness; + ts.kurtosis = kurtosis; + ts.gain = fnan; + ts.entropy = entropy; + ts.l2_dist = fnan; + ts.cossim = fnan; + ts.pearson = fnan; + ts.covariance = fnan; + + return true; } -static void compute_cossim(std::vector & tstats) { - static const std::regex pattern(R"(blk\.(\d+)\.)"); +static void compute_tensor_statistics(std::vector & tstats) { + constexpr auto fnan = std::numeric_limits::quiet_NaN(); + std::unordered_map tensor_map; + tensor_map.reserve(tstats.size()); + for (size_t i = 0; i < tstats.size(); ++i) { tensor_map[tstats[i].tensor] = i; } + for (auto & ts : tstats) { - if (std::smatch match; std::regex_search(ts.tensor, match, pattern)) { - const int blk = std::stoi(match[1]); - std::string tname(ts.tensor); - tname.replace(match.position(1), match.length(1), std::to_string(blk-1)); - auto prev = std::find_if(tstats.begin(), tstats.end(), - [tname](const tensor_statistics & t) { return t.tensor == tname; }); - if (prev != tstats.end()) { - const float dp = std::inner_product(ts.stats.values.begin(), ts.stats.values.end(), - prev->stats.values.begin(), 0.0f); - const float curr_mag = std::sqrt(std::inner_product(ts.stats.values.begin(), ts.stats.values.end(), - ts.stats.values.begin(), 0.0f)); - const float prev_mag = std::sqrt(std::inner_product(prev->stats.values.begin(), prev->stats.values.end(), - prev->stats.values.begin(), 0.0f)); - const float cs = dp / (curr_mag * prev_mag); - ts.cossim = cs; + std::string layer_str; + std::string dummy_tensor; + process_tensor_name(ts.tensor, layer_str, dummy_tensor); + + int blk = -1; + try { blk = std::stoi(layer_str); } catch (...) { continue; } + if (blk <= 0) { continue; } + + const size_t blk_start_pos = ts.tensor.find("blk." + layer_str); + if (blk_start_pos == std::string::npos) { continue; } + + std::string tname = ts.tensor; + tname.replace(blk_start_pos, layer_str.length() + 4, "blk." + std::to_string(blk - 1)); + + auto it = tensor_map.find(tname); + if (it == tensor_map.end()) { + LOG_WRN("%s: missing previous-layer tensor '%s'\n", __func__, tname.c_str()); + continue; + } + + const auto & prev_ts = tstats[it->second]; + const auto curr_avg = compute_tensor_averages(ts.stats); + const auto prev_avg = compute_tensor_averages(prev_ts.stats); + + if (curr_avg.empty() || curr_avg.size() != prev_avg.size()) { continue; } + + double dot_prod = 0.0; + double norm1_sq = 0.0; + double norm2_sq = 0.0; + double l2_dist_sq = 0.0; + double sum_c = 0.0; + double sum_p = 0.0; + const size_t n = curr_avg.size(); + size_t valid_n = 0; + + // Sums for Means + for (size_t i = 0; i < n; ++i) { + const double c_val = curr_avg[i]; + const double p_val = prev_avg[i]; + if (std::isfinite(c_val) && std::isfinite(p_val)) { + sum_c += c_val; + sum_p += p_val; + valid_n++; } + } + + if (valid_n == 0) { continue; } + const double mean_c = sum_c / valid_n; + const double mean_p = sum_p / valid_n; + + double cov_sum = 0.0; + double var_c_sum = 0.0; + double var_p_sum = 0.0; + + // Metrics + for (size_t i = 0; i < n; ++i) { + const double c_val = curr_avg[i]; + const double p_val = prev_avg[i]; + if (!std::isfinite(c_val) || !std::isfinite(p_val)) { continue; } + + // Cosine Similarity & L2 Distance + dot_prod += c_val * p_val; + norm1_sq += c_val * c_val; + norm2_sq += p_val * p_val; + const double diff = c_val - p_val; + l2_dist_sq += diff * diff; + + // Pearson (Centered stats) + const double dc = c_val - mean_c; + const double dp = p_val - mean_p; + cov_sum += dc * dp; + var_c_sum += dc * dc; + var_p_sum += dp * dp; + } + + ts.n_features = (int64_t)n; + ts.dot_prod = dot_prod; + ts.norm1_sq = norm1_sq; + ts.norm2_sq = norm2_sq; + ts.cov_sum = cov_sum; + ts.var_c_sum = var_c_sum; + ts.var_p_sum = var_p_sum; + ts.l2_dist_sq = l2_dist_sq; + ts.l2_dist = (float)std::sqrt(l2_dist_sq); + + if (valid_n > 1) { + ts.covariance = (float)(cov_sum / (double)valid_n); + } + + if (norm1_sq > 1e-12 && norm2_sq > 1e-12) { + ts.cossim = (float)(dot_prod / (std::sqrt(norm1_sq) * std::sqrt(norm2_sq))); + ts.cossim = std::clamp(ts.cossim, -1.0f, 1.0f); + } else { + ts.cossim = (norm1_sq == 0.0 && norm2_sq == 0.0) ? fnan : 0.0f; + } + + if (var_c_sum > 1e-12 && var_p_sum > 1e-12) { + ts.pearson = (float)(cov_sum / (std::sqrt(var_c_sum) * std::sqrt(var_p_sum))); + ts.pearson = std::clamp(ts.pearson, -1.0f, 1.0f); + } else { + ts.pearson = (var_c_sum == 0.0 && var_p_sum == 0.0) ? fnan : 0.0f; + } + + if (prev_ts.sum > 1e-10f) { + ts.gain = std::sqrt(ts.sum / ts.elements) / std::sqrt(prev_ts.sum / prev_ts.elements); + ts.sum_prev = prev_ts.sum; + ts.elements_prev = prev_ts.elements; + } else { + ts.gain = ts.sum <= 1e-10f ? 1.0f : fnan; + } + } +} + +static void compute_layer_statistics(const std::vector & tstats, + std::map & layer_cossim, + std::map & layer_l2_dist, + std::map & layer_pearson, + std::map & layer_covariance, + std::map & layer_gain +) +{ + struct layer_aggregation { + double sum_dot_prod = 0.0; + double sum_norm1_sq = 0.0; + double sum_norm2_sq = 0.0; + double sum_l2_dist_sq = 0.0; + double sum_cov = 0.0; + double sum_var_c = 0.0; + double sum_var_p = 0.0; + double sum_total_energy_curr = 0.0; + double sum_total_energy_prev = 0.0; + int64_t sum_valid_n = 0; + int64_t sum_n_features = 0; + int n_tensors = 0; + }; + + constexpr auto fnan = std::numeric_limits::quiet_NaN(); + std::map laggr; + + for (const auto & ts : tstats) { + std::string layer_str; + std::string dummy; + process_tensor_name(ts.tensor, layer_str, dummy); + int blk = -1; + try { + blk = std::stoi(layer_str); + } catch(...) { + if (layer_str == "-") { blk = -1; } + } + + if (blk <= 0) { continue; } + + if (ts.norm1_sq == 0.0 && ts.norm2_sq == 0.0 && ts.l2_dist_sq == 0.0) { continue; } + auto & entry = laggr[blk]; + entry.sum_dot_prod += ts.dot_prod; + entry.sum_norm1_sq += ts.norm1_sq; + entry.sum_norm2_sq += ts.norm2_sq; + entry.sum_l2_dist_sq += ts.l2_dist_sq; + entry.sum_cov += ts.cov_sum; + entry.sum_var_c += ts.var_c_sum; + entry.sum_var_p += ts.var_p_sum; + entry.n_tensors++; + if (ts.elements > 0) { entry.sum_valid_n += ts.elements; } + if (ts.n_features > 0) { entry.sum_n_features += ts.n_features; } + + // Accumulate Energy for correct Layer Gain calculation + entry.sum_total_energy_curr += ts.sum; + if (ts.sum_prev > 0.0) { entry.sum_total_energy_prev += ts.sum_prev; } + } + + for (const auto & [layer, agg] : laggr) { + if (agg.n_tensors == 0) { continue; } + + float cossim = 0.0f; + if (agg.sum_norm1_sq > 0.0 && agg.sum_norm2_sq > 0.0) { + cossim = (float)(agg.sum_dot_prod / (std::sqrt(agg.sum_norm1_sq) * std::sqrt(agg.sum_norm2_sq))); + cossim = std::clamp(cossim, -1.0f, 1.0f); + } else if (agg.sum_norm1_sq == 0.0 && agg.sum_norm2_sq == 0.0) { + cossim = fnan; + } + + float gain = 0.0f; + if (agg.sum_total_energy_prev > 0.0) { + gain = (float)(std::sqrt(agg.sum_total_energy_curr) / std::sqrt(agg.sum_total_energy_prev)); } else { - ts.cossim = 0; + gain = fnan; + } + + layer_cossim[layer] = cossim; + layer_l2_dist[layer] = (float)std::sqrt(agg.sum_l2_dist_sq); + layer_gain[layer] = gain; + + if (agg.sum_n_features > 0) { layer_covariance[layer] = (float)(agg.sum_cov / (double)agg.sum_n_features); } + else { layer_covariance[layer] = fnan; } + + if (agg.sum_var_c > 1e-12 && agg.sum_var_p > 1e-12) { + auto pearson = (float)(agg.sum_cov / (std::sqrt(agg.sum_var_c) * std::sqrt(agg.sum_var_p))); + layer_pearson[layer] = std::clamp(pearson, -1.0f, 1.0f); + } else if (agg.sum_var_c == 0.0 && agg.sum_var_p == 0.0) { + layer_pearson[layer] = fnan; + } else { + layer_pearson[layer] = 0.0f; } } } @@ -287,6 +559,7 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void * e.counts.resize(n_as, e.counts[0]); } if (e.values.empty()) { + e.activations.resize(src1->ne[0]*n_as, 0); e.values.resize(src1->ne[0]*n_as, 0); e.counts.resize(n_as, 0); } @@ -318,6 +591,7 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void * e.counts[ex]++; for (int64_t j = 0; j < src1->ne[0]; ++j) { + e.activations[e_start + j] += x[j]; e.values[e_start + j] += x[j] * x[j]; if (!std::isfinite((float)e.values[e_start + j])) { LOG_ERR("%f detected in %s\n", (float)e.values[e_start + j], wname.c_str()); @@ -357,6 +631,7 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void * } } if (e.values.empty()) { + e.activations.resize(src1->ne[0] * n_mat, 0); e.values.resize(src1->ne[0] * n_mat, 0); e.counts.resize(1, 0); } @@ -375,6 +650,7 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void * for (int64_t row = 0; row < src1->ne[1]; ++row) { const float * x = (const float *) (data + row * src1->nb[1] + i2 * src1->nb[2] + i3 * src1->nb[3]); for (int64_t j = 0; j < src1->ne[0]; ++j) { + e.activations[mat_start + j] += x[j]; e.values[mat_start + j] += x[j] * x[j]; if (!std::isfinite((float)e.values[j])) { LOG_ERR("%f detected in %s\n", (float)e.values[j], wname.c_str()); @@ -401,6 +677,16 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void * } } +#if defined(__linux__) || defined(__APPLE__) || defined(__FreeBSD__) + if (m_params.use_mmap && src0->buffer && ggml_backend_buffer_is_host(src0->buffer)) { + const size_t page_size = sysconf(_SC_PAGESIZE); + const uintptr_t addr = (uintptr_t)src0->data; + const uintptr_t aligned_addr = addr & ~(page_size - 1); + const size_t size = ggml_nbytes(src0) + (addr - aligned_addr); + madvise((void *)aligned_addr, size, MADV_DONTNEED); + } +#endif + return true; } @@ -556,13 +842,29 @@ void IMatrixCollector::save_imatrix(int32_t n_chunk) const { } to_store.push_back(kv.first); + data_size += GGML_PAD(ggml_tensor_overhead() + sizeof(float) * kv.second.activations.size(), GGML_MEM_ALIGN); data_size += GGML_PAD(ggml_tensor_overhead() + sizeof(float) * kv.second.values.size(), GGML_MEM_ALIGN); data_size += GGML_PAD(ggml_tensor_overhead() + sizeof(float) * kv.second.counts.size(), GGML_MEM_ALIGN); + data_size += GGML_PAD(ggml_tensor_overhead() + sizeof(float) * 10, GGML_MEM_ALIGN); } // deterministic tensor name order std::sort(to_store.begin(), to_store.end()); + // Compute per-tensor statistics + std::vector tstats; + tstats.reserve(m_stats.size()); + bool legacy; + for (const auto & kv : m_stats) { + compute_vector_statistics(tstats, kv.first, kv.second, legacy); + } + if (!tstats.empty()) { compute_tensor_statistics(tstats); } + + // index by tensor name + std::unordered_map tstat_index; + tstat_index.reserve(tstats.size()); + for (const auto & ts : tstats) { tstat_index[ts.tensor] = &ts; } + struct ggml_init_params params = { /* .mem_size = */ data_size, /* .mem_buffer = */ NULL, @@ -587,6 +889,12 @@ void IMatrixCollector::save_imatrix(int32_t n_chunk) const { // Write the number of chunks the matrix was computed with gguf_set_val_u32(ctx_gguf, LLM_KV_IMATRIX_CHUNK_COUNT, m_last_chunk); gguf_set_val_u32(ctx_gguf, LLM_KV_IMATRIX_CHUNK_SIZE, m_params.n_ctx / m_params.n_parallel); + + // Define the schema for the tensor statistics (for use in quantize.cpp) + const char * stats_schema[] = {"sum_sq", "mean", "elements", "std_deviation", "skewness", "kurtosis", "gain", + "h_norm", "l2_dist", "cossim", "pearson", "covariance" + }; + gguf_set_arr_str(ctx_gguf, LLM_KV_IMATRIX_STATS_SCHEMA, stats_schema, 12); } for (const auto & name : to_store) { @@ -608,6 +916,67 @@ void IMatrixCollector::save_imatrix(int32_t n_chunk) const { gguf_add_tensor(ctx_gguf, in_sum2); gguf_add_tensor(ctx_gguf, counts); + + if (!stat.activations.empty()) { + const int32_t nact = (int32_t) stat.activations.size(); + struct ggml_tensor * in_sum = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, nact / nmat, nmat); + ggml_format_name(in_sum, "%s.in_sum", name.c_str()); + for (int32_t j = 0; j < nact; ++j) { + ((float *) in_sum->data)[j] = (float) stat.activations[j]; + } + gguf_add_tensor(ctx_gguf, in_sum); + } + } else { + LOG_WRN("%s: no data for tensor %s\n", __func__, name.c_str()); + } + + // Store per-tensor statistics as a small 1D tensor + { + float fnan = std::numeric_limits::quiet_NaN(); + double sum_sq = 0.0f; + float mean = 0.0f; + float elements = 0.0f; + float std_deviation = 0.0f; + float skewness = 0.0f; + float kurtosis = 0.0f; + float gain = 0.0f; + float h_norm = 0.0f; + float l2_dist = 0.0f; + float cossim = 0.0f; + float pearson = 0.0f; + float covariance = 0.0f; + auto ts = tstat_index.find(name); + if (ts != tstat_index.end() && ts->second != nullptr) { + sum_sq = ts->second->sum; + mean = ts->second->mean; + elements = (float)ts->second->elements; + std_deviation = ts->second->std_deviation; + skewness = ts->second->skewness; + kurtosis = ts->second->kurtosis; + gain = ts->second->gain; + h_norm = ts->second->elements > 1 ? 100.0f * (ts->second->entropy / std::log2f((float)ts->second->elements)) : fnan; + l2_dist = ts->second->l2_dist; + cossim = ts->second->cossim; + pearson = ts->second->pearson; + covariance = ts->second->covariance; + } + + struct ggml_tensor * stats = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 12); + ggml_format_name(stats, "%s.stats", name.c_str()); + // Store the statistics in the same order as defined in stats_schema[] + ((float *)stats->data)[0] = (float)sum_sq; + ((float *)stats->data)[1] = mean; + ((float *)stats->data)[2] = elements; + ((float *)stats->data)[3] = std_deviation; + ((float *)stats->data)[4] = skewness; + ((float *)stats->data)[5] = kurtosis; + ((float *)stats->data)[6] = gain; + ((float *)stats->data)[7] = h_norm; + ((float *)stats->data)[8] = l2_dist; + ((float *)stats->data)[9] = cossim; + ((float *)stats->data)[10] = pearson; + ((float *)stats->data)[11] = covariance; + gguf_add_tensor(ctx_gguf, stats); } } @@ -652,7 +1021,7 @@ bool IMatrixCollector::load_imatrix(const char * file_name) { if (e.values.empty()) { e.values.resize(nval, 0.0f); - } else if ((size_t) nval != e.values.size()) { + } else if ((size_t)nval != e.values.size()) { LOG_ERR("%s: mismatched sums size for %s: %zu != %zu\n", __func__, name.c_str(), (size_t) nval, e.values.size()); return false; } @@ -947,111 +1316,209 @@ static bool compute_imatrix(llama_context * ctx, const common_params & params, c } static bool show_statistics(const common_params & params) { + constexpr auto fnan = std::numeric_limits::quiet_NaN(); + g_collector.set_params(params); std::vector ts; - if (params.in_files.empty() || params.in_files.size() > 1) { - LOG_ERR("\nError: a single imatrix file is required to compute tensor statistics\n\n"); - return false; - } + + if (params.in_files.empty()) { return false; } + + // Load and process data if (g_collector.load_imatrix(params.in_files[0].c_str())) { - for (const auto & [name, stats] :g_collector.get_mstats()) { - compute_statistics(ts, name, stats); + ts.reserve(g_collector.get_mstats().size()); + for (const auto & [name, stats] : g_collector.get_mstats()) { + bool legacy_imatrix = true; + if (!compute_vector_statistics(ts, name, stats, legacy_imatrix)) { continue; } } } else { - LOG_ERR("\nError: %s is not a valid imatrix file\n\n", params.in_files[0].c_str()); - return false; - } - if (!ts.empty()) { - compute_cossim(ts); - } else { - LOG_ERR("Error: cannot compute statistics for %s\n\n", params.in_files[0].c_str()); return false; } + if (ts.empty()) { return false; } + + bool legacy = ts.empty() ? true : ts[0].stats.activations.empty(); + compute_tensor_statistics(ts); + + // Sorting logic (Layer index -> Tensor Name) struct tensor_comparer { bool operator()(const tensor_statistics & a, const tensor_statistics & b) const { - std::string layer, name_a, name_b; - ; - process_tensor_name(a.tensor, layer, name_a); - process_tensor_name(b.tensor, layer, name_b); - return name_a < name_b || (name_a == name_b && a.total_sqract > b.total_sqract); + std::string lay_a; + std::string lay_b; + std::string name_a; + std::string name_b; + process_tensor_name(a.tensor, lay_a, name_a); + process_tensor_name(b.tensor, lay_b, name_b); + + // Handle non-numeric layers (e.g., "output") + int blk_a = INT_MAX - 1; + int blk_b = INT_MAX - 1; + try { + blk_a = std::stoi(lay_a); + } catch(...) { + if (a.tensor.find("output") != std::string::npos) { blk_a = INT_MAX; } + } + try { + blk_b = std::stoi(lay_b); + } catch(...) { + if (b.tensor.find("output") != std::string::npos) { blk_b = INT_MAX; } + } + + if (blk_a != blk_b) { return blk_a < blk_b; } + return name_a < name_b; } }; std::sort(ts.begin(), ts.end(), tensor_comparer()); - struct weighted_stats { - float weighted_bias = 0.0f; - float weighted_zd = 0.0f; - float weighted_cossim = 0.0f; - int total_elements = 0; + struct layer_stats { + float layer_sum = 0.0f; + int n = 0; + }; + std::map ls; + + // Shorten names for table formatting + auto label_fmt = [](std::string s, size_t w) -> std::string { + if (s.length() <= w) { return s; } + return ".." + s.substr(s.length() - (w - 2)); }; - std::map ws; - - LOG_INF("\nComputing statistics for %s (%d tensors)\n", params.in_files[0].c_str(), static_cast(ts.size())); - LOG_INF("\n%s\t%s\t%s\t%s\t%s\t%s\t%s\t%s\t%s\t%s\t%s\t%s\t%s\n", " Layer", " Tensor", " Σ(Act²)", - " Min", " Max", " μ", " σ", " % Active", "N", " Entropy", "E (norm)", "ZD", - " CosSim"); - LOG_INF( - "==============================================================================================================" - "===========================================================\n"); + + constexpr int w_lay = 6; + constexpr int w_nam = 40; // Should be wide enough for most tensor names + const auto * sep = " | "; + + printf("\nComputing tensor statistics for %s (%d tensors)\n", params.in_files[0].c_str(), static_cast(ts.size())); + + if (legacy) { + printf("\n%*s%s%-*s%s%10s%10s%12s%12s%9s%s%17s%8s%s%10s%10s\n", + w_lay, "Layer", sep, + w_nam, "Tensor", sep, + "Mean", "StdDev", "Skew", "Kurt", "H Norm", sep, + "∑ E[A²]", "Gain", sep, + "PCC", "Cov"); + printf("%s\n", std::string(153, '-').c_str()); + } else { + printf("\n%*s%s%-*s%s%10s%10s%12s%12s%9s%s%17s%8s%s%12s%10s%10s\n", + w_lay, "Layer", sep, + w_nam, "Tensor", sep, + "Mean", "StdDev", "Skew", "Kurt", "H Norm", sep, + "∑ E[A²]", "Gain", sep, + "L2 Dist", "PCC", "Cov"); + printf("%s\n", std::string(165, '-').c_str()); + } + + // Tensor Statistics for (const auto & tstat : ts) { - std::string layer, name; + std::string layer; + std::string name; + process_tensor_name(tstat.tensor, layer, name); + const float h_norm = tstat.elements > 1 ? 100.0f * (tstat.entropy / std::log2f((float)tstat.elements)) : fnan; int blk; try { blk = std::stoi(layer); - } catch (const std::exception & e) { - blk = -1; // not a block layer + } catch (...) { + if (tstat.tensor.find("output") != std::string::npos) { blk = INT_MAX; } + else { blk = -1; } } - const float entropy_norm = (tstat.elements > 0) ? 100.0f * (tstat.entropy / std::log2(tstat.elements)) : 0.0f; + if (legacy) { + printf("%*s%s%-*s%s%10.4f%10.4f%12.4f%12.4f%8.2f%%%s%14.4f%8.2f%s%10.4f%10.4f\n", + w_lay, layer.c_str(), sep, + w_nam, label_fmt(tstat.tensor, w_nam).c_str(), sep, + tstat.mean, tstat.std_deviation, tstat.skewness, tstat.kurtosis, h_norm, sep, + tstat.sum, tstat.gain, sep, + tstat.pearson, tstat.covariance + ); + } else { + printf("%*s%s%-*s%s%10.4f%10.4f%12.4f%12.4f%8.2f%%%s%14.4f%8.2f%s%12.4f%10.4f%10.4f\n", + w_lay, layer.c_str(), sep, + w_nam, label_fmt(tstat.tensor, w_nam).c_str(), sep, + tstat.mean, tstat.std_deviation, tstat.skewness, tstat.kurtosis, h_norm, sep, + tstat.sum, tstat.gain, sep, + tstat.l2_dist, tstat.pearson, tstat.covariance + ); + } - LOG_INF("%5s\t%-20s\t%10.2f\t%8.4f\t%11.4f\t%6.2f\t%6.2f\t%8.2f%%\t%6d\t%10.4f\t%6.2f%%\t%10.2f%%\t%8.4f\n", - layer.c_str(), name.c_str(), tstat.total_sqract, tstat.min_sqract, tstat.max_sqract, tstat.mean_sqract, - tstat.stddev, tstat.active * 100.0f, tstat.elements, tstat.entropy, - entropy_norm, 100.0f * tstat.zd, tstat.cossim); + // Aggregate Layer Stats + auto & l = ls[blk]; + l.layer_sum += tstat.sum; + l.n += tstat.elements; + } - const float weighted_bias = tstat.elements * tstat.total_sqract; - const float weighted_zd = tstat.elements * tstat.zd; - const float weighted_cossim = tstat.elements * tstat.cossim; + // Layer Statistics + std::map layer_cossim; + std::map layer_l2_dist; + std::map layer_pearson; + std::map layer_covariance; + std::map layer_gain; + compute_layer_statistics(ts, layer_cossim, layer_l2_dist, layer_pearson, layer_covariance, layer_gain); + + size_t layers = 0; + int min = std::numeric_limits::max(); + int max = -1; + + for (const auto & [layer, stats] : ls) { + if (layer >= 0 && layer < 9999 && stats.n > 0) { + layers++; + min = std::min(layer, min); + max = std::max(layer, max); + } + } - if (ws.find(blk) != ws.end()) { - ws[blk].weighted_bias += weighted_bias; - ws[blk].weighted_zd += weighted_zd; - ws[blk].weighted_cossim += weighted_cossim; - ws[blk].total_elements += tstat.elements; - } else { - weighted_stats temp_ws; - temp_ws.weighted_bias = weighted_bias; - temp_ws.weighted_zd = weighted_zd; - temp_ws.weighted_cossim = weighted_cossim; - temp_ws.total_elements = tstat.elements; - ws[blk] = temp_ws; + if (layers > 0) { + const auto expected = (size_t)(max - min + 1); + if (layers != expected) { + LOG_WRN("\n%s: layer sequence gap detected (found %zu layers in range %d-%d, expected %zu); layer statistics will not be shown\n", + __func__, layers, min, max, expected); + return false; } } - const int layers = std::count_if(ws.begin(), ws.end(), [](const auto & kv) { return kv.first >= 0; }); - LOG_INF("\nComputing weighted average statistics per layer (%d layers)\n", layers); - LOG_INF("\n%s\t%s\t%s\t%s\n", " Layer", " μΣ(Act²)", " μZD", "μCosSim"); - LOG_INF("================================================\n"); - for (const auto & [first, second] : ws) { - const auto & layer = first; - const auto & stats = second; + printf("\nComputing layer statistics for %s (%zu layers)\n\n", params.in_files[0].c_str(), layers); - if (stats.total_elements == 0) { - continue; - } + if (legacy) { + printf("%*s%s%17s%8s%s%9s%9s%12s\n", + w_lay, "Layer", sep, + "∑ E[A²]", "Gain", sep, + "CosSim", "PCC", "Cov"); + printf("%s\n", std::string(64, '-').c_str()); + } else { + printf("%*s%s%17s%8s%s%12s%9s%9s%12s\n", + w_lay, "Layer", sep, + "∑ E[A²]", "Gain", sep, + "L2 Dist", "CosSim", "PCC", "Cov"); + printf("%s\n", std::string(76, '-').c_str()); + } - if (layer >= 0) { - const float bias = stats.weighted_bias / stats.total_elements; - const float zd = stats.weighted_zd / stats.total_elements; - const float cossim = stats.weighted_cossim / stats.total_elements; + auto get_layer_stat = [&](const std::map& map, const int layer) -> float { + const auto it = map.find(layer); + return it != map.end() ? it->second : fnan; + }; - LOG_INF("%5d\t%14.2f\t%10.4f%%\t%6.4f\n", layer, bias, 100.0f * zd, cossim); + for (const auto & [layer, stats] : ls) { + if (layer < 0 || stats.n == 0) { continue; } + + float lgn = layer == 0 || layer == INT_MAX ? fnan : get_layer_stat(layer_gain, layer); + float ll2 = layer == 0 || layer == INT_MAX ? fnan : get_layer_stat(layer_l2_dist, layer); + float lcs = layer == 0 || layer == INT_MAX ? fnan : get_layer_stat(layer_cossim, layer); + float lpc = layer == 0 || layer == INT_MAX ? fnan : get_layer_stat(layer_pearson, layer); + float lcv = layer == 0 || layer == INT_MAX ? fnan : get_layer_stat(layer_covariance, layer); + auto str = std::to_string(layer); + const auto *lyr = layer == INT_MAX ? "-" : str.c_str(); + + if (legacy) { + printf("%*s%s%14.4f%8.2f%s%9.4f%9.4f%12.4f\n", + w_lay, lyr, sep, + stats.layer_sum, lgn, sep, + lcs, lpc, lcv); + } else { + printf("%*s%s%14.4f%8.2f%s%12.4f%9.4f%9.4f%12.4f\n", + w_lay, lyr, sep, + stats.layer_sum, lgn, sep, + ll2, lcs, lpc, lcv); } } - LOG_INF("\n"); + printf("\n"); return true; } diff --git a/uv.lock b/uv.lock new file mode 100644 index 000000000000..7518fc90bf78 --- /dev/null +++ b/uv.lock @@ -0,0 +1,3 @@ +version = 1 +revision = 3 +requires-python = ">=3.12"