imatrix: calculate activation-based statistics for new format (GGUF) imatrices

common/common.h

@@ -655,10 +655,11 @@ 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;

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 +70,32 @@ 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
+# analyze imatrix file and display summary statistics instead of running inference
 ./llama-imatrix --in-file imatrix.gguf --show-statistics
 ```
 
-`--show-statistics` will display the following statistics:
+## Statistics
+
+Please note that the L₂ Distance can only be calculated if the imatrix is in GGUF format. If a value lacks proper statistical interpretability, **nan** will be shown instead. The following statistics are computed:
 
 #### 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.
+* **Min / Max / μ / σ**: Tensor elements Min, Max, Mean, and Standard Deviation.
+* **H Norm**: Shannon Entropy normalized over log₂(N). Defined as $H Norm=\frac{-\sum_{i=1}^N p_i \log_2 p_i}{log_2 N}$. 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.
+* **Z-score Distribution (ZD)**: % of elements whose ZD-score is > 1.0 (an indicator of outliers), as described in _3.1 Layer Importance Scores_ of [Layer-Wise Quantization](https://arxiv.org/abs/2406.17415).
+* **∑ E[A²]**: 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 (e.g., Q6_K vs Q4_K).
+* **L₂ Distance**: Euclidean Distance from the tensor in the previous layer. Measure of transformation magnitude; higher values indicate more significant transformation on the data.
+* **CosSim**: Cosine Similarity with the tensor in the previous layer. _~1.0_, the tensor output points in the exact same direction as the previous layer's tensor (the layer is refining magnitude, not direction). _< 1.0_, the layer is rotating the vector space (changing semantic meaning).
+* **PCC**: Pearson Correlation Coefficient with the tensor in the previous layer. Checks for linear correlation excluding the mean shift. Similar to CosSim but centers geometric data first. Indicates if the pattern of activation changes or just the offset.
 
 #### Per layer
 
-Weighted averages of Σ(Act²), ZD Score and CosSim are also calculated.
+Aggregated metrics per block/layer:
 
-#### Important note on the computed Statistics
+* **Z-score Distribution (ZD)**: % of this layer's concatenated tensors' elements with |Z| > 1. Indicates general "spikiness" of the layer's activations.
+* **∑ E[A²]:** Total energy of the layer's concatenated tensors. Indicates the layer's overall contribution amplitude.
+* **L₂ Distance:** Euclidean Distance of the layer's concatenated tensors from the previous layer’s. Global measure of transformation magnitude.
+* **CosSim**: Cosine Similarity of this layer's concatenated tensors with the previous layer.
+* **PCC**: Average Pearson Correlation of the tensors in the layer.
 
-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.
+More information is available in https://github.com/ggml-org/llama.cpp/pull/14891