Deduplicate q4 quantization functions (#383)

* Deduplicate q4 quantization functions

* Use const; add basic test

* Re-enable quantization test

* Disable AVX2 flags in CI

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

.github/workflows/build.yml

@@ -89,7 +89,7 @@ jobs:
         run: |
           mkdir build
           cd build
-          cmake ..
+          cmake -DLLAMA_AVX2=OFF ..
           cmake --build . --config Release
           ctest --output-on-failure
 

CMakeLists.txt

@@ -217,6 +217,7 @@ add_library(utils OBJECT
 
 target_include_directories(utils PUBLIC .)
 target_compile_features(utils PUBLIC cxx_std_11) # don't bump
+target_link_libraries(utils PRIVATE ${LLAMA_EXTRA_LIBS})
 
 add_library(ggml OBJECT
             ggml.c
@@ -226,12 +227,13 @@ target_include_directories(ggml PUBLIC .)
 target_compile_features(ggml PUBLIC c_std_11) # don't bump
 target_link_libraries(ggml PRIVATE Threads::Threads ${LLAMA_EXTRA_LIBS})
 
-add_library(llama OBJECT
+add_library(llama
             llama.cpp
             llama.h)
 
 target_include_directories(llama PUBLIC .)
 target_compile_features(llama PUBLIC cxx_std_11) # don't bump
+target_link_libraries(llama PRIVATE utils ggml ${LLAMA_EXTRA_LIBS})
 
 #
 # Executables

README.md

@@ -240,6 +240,40 @@ or
 
 `shasum -a 256 --ignore-missing -c SHA256SUMS` on macOS
 
+### Perplexity (Measuring model quality)
+
+You can pass `--perplexity` as a command line option to measure perplexity over the given prompt.  For more background,
+see https://huggingface.co/docs/transformers/perplexity.  However, in general, lower perplexity is better for LLMs.
+
+#### Measurements
+
+https://github.com/ggerganov/llama.cpp/pull/270 is the unofficial tracking page for now.  llama.cpp is measuring very well
+compared to the baseline implementations.  Quantization has a small negative impact to quality, but, as you can see, running
+13B at q4_0 beats the 7B f16 model by a significant amount.
+
+All measurements are done against wikitext2 test dataset (https://paperswithcode.com/dataset/wikitext-2), with default options (512 length context).
+Note that the changing the context length will have a significant impact on perplexity (longer context = better perplexity).
+```
+Perplexity - model options
+5.5985 - 13B, q4_0
+5.9565 - 7B, f16
+6.3001 - 7B, q4_1
+6.5949 - 7B, q4_0
+6.5995 - 7B, q4_0, --memory_f16
+```
+
+#### How to run
+
+1. Download/extract: https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip?ref=salesforce-research
+2. Run `./main --perplexity -m models/7B/ggml-model-q4_0.bin -f wiki.test.raw`
+3. Output:
+```
+Calculating perplexity over 655 chunks
+24.43 seconds per pass - ETA 4.45 hours
+[1]4.5970,[2]5.1807,[3]6.0382,...
+```
+And after 4.45 hours, you will have the final perplexity.
+
 ### Android
 
 You can easily run `llama.cpp` on Android device with [termux](https://play.google.com/store/apps/details?id=com.termux).
@@ -290,7 +324,6 @@ docker run -v /llama/models:/models ghcr.io/ggerganov/llama.cpp:light -m /models
 
 ## Limitations
 
-- We don't know yet how much the quantization affects the quality of the generated text
 - Probably the token sampling can be improved
 - The Accelerate framework is actually currently unused since I found that for tensor shapes typical for the Decoder,
   there is no benefit compared to the ARM_NEON intrinsics implementation. Of course, it's possible that I simply don't

ggml.c

@@ -1,3 +1,6 @@
+// Defines CLOCK_MONOTONIC on Linux
+#define _POSIX_C_SOURCE 199309L
+
 #include "ggml.h"
 
 #if defined(_MSC_VER) || defined(__MINGW32__)
@@ -400,9 +403,55 @@ static inline __m128i packNibbles( __m256i bytes )
 // method 5
 // blocks of QK elements
 // represented with a single float (delta) and QK/2 8-bit ints (i.e QK 4-bit signed integer factors)
+
+// reference implementation for deterministic creation of model files
+static void quantize_row_q4_0_reference(const float * restrict x, void * restrict y, int k) {
+    assert(k % QK == 0);
+    const int nb = k / QK;
+
+    const size_t bs = sizeof(float) + QK/2;
+
+    uint8_t * restrict pd = ((uint8_t *)y + 0*bs);
+    uint8_t * restrict pb = ((uint8_t *)y + 0*bs + sizeof(float));
+
+    uint8_t pp[QK/2];
+
+    for (int i = 0; i < nb; i++) {
+        float amax = 0.0f; // absolute max
+
+        for (int l = 0; l < QK; l++) {
+            const float v = x[i*QK + l];
+            amax = MAX(amax, fabsf(v));
+        }
+
+        const float d = amax / ((1 << 3) - 1);
+        const float id = d ? 1.0f/d : 0.0f;
+
+        *(float *)pd = d;
+        pd += bs;
+
+        for (int l = 0; l < QK; l += 2) {
+            const float v0 = x[i*QK + l + 0]*id;
+            const float v1 = x[i*QK + l + 1]*id;
+
+            const uint8_t vi0 = ((int8_t) (round(v0))) + 8;
+            const uint8_t vi1 = ((int8_t) (round(v1))) + 8;
+
+            assert(vi0 >= 0 && vi0 < 16);
+            assert(vi1 >= 0 && vi1 < 16);
+
+            pp[l/2] = vi0 | (vi1 << 4);
+        }
+
+        memcpy(pb, pp, sizeof(pp));
+        pb += bs;
+    }
+}
+
 void quantize_row_q4_0(const float * restrict x, void * restrict y, int k) {
     assert(k % QK == 0);
 
+#if __ARM_NEON || defined(__AVX2__) || defined(__wasm_simd128__)
     const int nb = k / QK;
     const size_t bs = sizeof(float) + QK/2;
 
@@ -410,6 +459,7 @@ void quantize_row_q4_0(const float * restrict x, void * restrict y, int k) {
     uint8_t * restrict pb = ((uint8_t *)y + 0*bs + sizeof(float));
 
     uint8_t pp[QK/2];
+#endif
 
 #if __ARM_NEON
 #if QK == 32
@@ -566,36 +616,7 @@ void quantize_row_q4_0(const float * restrict x, void * restrict y, int k) {
 #endif
 #else
     // scalar
-    for (int i = 0; i < nb; i++) {
-        float amax = 0.0f; // absolute max
-
-        for (int l = 0; l < QK; l++) {
-            const float v = x[i*QK + l];
-            amax = MAX(amax, fabsf(v));
-        }
-
-        const float d = amax / ((1 << 3) - 1);
-        const float id = d ? 1.0f/d : 0.0f;
-
-        *(float *)pd = d;
-        pd += bs;
-
-        for (int l = 0; l < QK; l += 2) {
-            const float v0 = x[i*QK + l + 0]*id;
-            const float v1 = x[i*QK + l + 1]*id;
-
-            const uint8_t vi0 = ((int8_t) (round(v0))) + 8;
-            const uint8_t vi1 = ((int8_t) (round(v1))) + 8;
-
-            assert(vi0 >= 0 && vi0 < 16);
-            assert(vi1 >= 0 && vi1 < 16);
-
-            pp[l/2] = vi0 | (vi1 << 4);
-        }
-
-        memcpy(pb, pp, sizeof(pp));
-        pb += bs;
-    }
+    quantize_row_q4_0_reference(x, y, k);
 #endif
 }
 
@@ -10702,119 +10723,60 @@ enum ggml_opt_result ggml_opt(
 
 ////////////////////////////////////////////////////////////////////////////////
 
-size_t ggml_quantize_q4_0(float * src, void * dst, int n, int k, int qk, int64_t * hist) {
+size_t ggml_quantize_q4_0(const float * src, void * dst, int n, int k, int qk, int64_t * hist) {
     const int nb = k / qk;
     const size_t bs = (sizeof(float) + sizeof(uint8_t)*qk/2);
     const size_t row_size = nb*bs;
 
     assert(k % qk == 0);
 
-    const size_t pp_size = qk / 2;
-    uint8_t * pp = (uint8_t *) alloca(pp_size);
-
     char * pdst = (char *) dst;
 
     for (int j = 0; j < n; j += k) {
         uint8_t * pd = (uint8_t *) (pdst + (j/k)*row_size + 0*bs);
         uint8_t * pb = (uint8_t *) (pdst + (j/k)*row_size + 0*bs + sizeof(float));
 
+        quantize_row_q4_0_reference(src + j, pd, k);
+
         for (int i = 0; i < nb; i++) {
-            float amax = 0.0f; // absolute max
+            for (int l = 0; l < qk; l += 2) {
+                const uint8_t vi0 = pb[l/2] & 0xF;
+                const uint8_t vi1 = pb[l/2] >> 4;
 
-            {
-                for (int l = 0; l < qk; l++) {
-                    const float v = src[j + i*qk + l];
-                    amax = MAX(amax, fabsf(v));
-                }
-
-                const float d = amax / ((1 << 3) - 1);
-                const float id = d ? 1.0f/d : 0.0f;
-
-                *(float *) pd = d;
-                pd += bs;
-
-                for (int l = 0; l < qk; l += 2) {
-                    const float v0 = (src[j + i*qk + l + 0])*id;
-                    const float v1 = (src[j + i*qk + l + 1])*id;
-
-                    const uint8_t vi0 = ((int8_t) (round(v0))) + 8;
-                    const uint8_t vi1 = ((int8_t) (round(v1))) + 8;
-
-                    assert(vi0 >= 0 && vi0 < 16);
-                    assert(vi1 >= 0 && vi1 < 16);
-
-                    hist[vi0]++;
-                    hist[vi1]++;
-
-                    pp[l/2] = vi0 | (vi1 << 4);
-                }
-
-                memcpy(pb, pp, pp_size);
-                pb += bs;
+                hist[vi0]++;
+                hist[vi1]++;
             }
+            pb += bs;
         }
     }
 
     return (n/k)*row_size;
 }
 
-size_t ggml_quantize_q4_1(float * src, void * dst, int n, int k, int qk, int64_t * hist) {
+size_t ggml_quantize_q4_1(const float * src, void * dst, int n, int k, int qk, int64_t * hist) {
     const int nb = k / qk;
     const size_t bs = (2*sizeof(float) + sizeof(uint8_t)*qk/2);
     const size_t row_size = nb*bs;
 
     assert(k % qk == 0);
 
-    const size_t pp_size = qk / 2;
-    uint8_t * pp = (uint8_t *) alloca(pp_size);
-
     char * pdst = (char *) dst;
 
     for (int j = 0; j < n; j += k) {
         uint8_t * pd = (uint8_t *) (pdst + (j/k)*row_size + 0*bs);
-        uint8_t * pm = (uint8_t *) (pdst + (j/k)*row_size + 0*bs +   sizeof(float));
         uint8_t * pb = (uint8_t *) (pdst + (j/k)*row_size + 0*bs + 2*sizeof(float));
 
-        //printf("n = %d, k = %d, nb = %d, row_size = %d, j = %d, pm = %p, pd = %p, pb = %p\n", n, k, nb, row_size, j, pm, pd, pb);
+        quantize_row_q4_1(src + j, pd, k);
 
         for (int i = 0; i < nb; i++) {
-            float min = FLT_MAX;
-            float max = -FLT_MAX;
+            for (int l = 0; l < qk; l += 2) {
+                const uint8_t vi0 = pb[l/2] & 0xF;
+                const uint8_t vi1 = pb[l/2] >> 4;
 
-            {
-                for (int l = 0; l < qk; l++) {
-                    const float v = src[j + i*qk + l];
-                    if (v < min) min = v;
-                    if (v > max) max = v;
-                }
-
-                const float d = (max - min) / ((1 << 4) - 1);
-                const float id = d ? 1.0f/d : 0.0f;
-
-                *(float *) pd = d;
-                *(float *) pm = min;
-                pd += bs;
-                pm += bs;
-
-                for (int l = 0; l < qk; l += 2) {
-                    const float v0 = (src[j + i*qk + l + 0] - min)*id;
-                    const float v1 = (src[j + i*qk + l + 1] - min)*id;
-
-                    const uint8_t vi0 = round(v0);
-                    const uint8_t vi1 = round(v1);
-
-                    assert(vi0 >= 0 && vi0 < 16);
-                    assert(vi1 >= 0 && vi1 < 16);
-
-                    hist[vi0]++;
-                    hist[vi1]++;
-
-                    pp[l/2] = vi0 | (vi1 << 4);
-                }
-
-                memcpy(pb, pp, pp_size);
-                pb += bs;
+                hist[vi0]++;
+                hist[vi1]++;
             }
+            pb += bs;
         }
     }
 

ggml.h

@@ -745,8 +745,8 @@ enum ggml_opt_result ggml_opt(
 // quantization
 //
 
-size_t ggml_quantize_q4_0(float * src, void * dst, int n, int k, int qk, int64_t * hist);
-size_t ggml_quantize_q4_1(float * src, void * dst, int n, int k, int qk, int64_t * hist);
+size_t ggml_quantize_q4_0(const float * src, void * dst, int n, int k, int qk, int64_t * hist);
+size_t ggml_quantize_q4_1(const float * src, void * dst, int n, int k, int qk, int64_t * hist);
 
 //
 // system info

llama.cpp

@@ -9,6 +9,7 @@
 #include <queue>
 #include <regex>
 #include <cassert>
+#include <cstring>
 
 // determine number of model parts based on the dimension
 static const std::unordered_map<int, int> LLAMA_N_PARTS = {

main.cpp

@@ -85,7 +85,7 @@ void perplexity(llama_context * ctx, const gpt_params & params) {
     // Download: https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip?ref=salesforce-research
     // Run `./main --perplexity -m models/7B/ggml-model-q4_0.bin -f wiki.test.raw`
     // Output: `perplexity: 13.5106 [114/114]`
-    auto tokens = ::llama_tokenize(ctx, params.prompt.c_str(), true);
+    auto tokens = ::llama_tokenize(ctx, params.prompt, true);
 
     int count = 0;
     double nll = 0.0;
@@ -254,6 +254,10 @@ int main(int argc, char ** argv) {
         params.interactive = true;
     }
 
+    if (params.interactive_start) {
+        params.interactive = true;
+    }
+
     fprintf(stderr, "\n");
     fprintf(stderr, "%s: prompt: '%s'\n", __func__, params.prompt.c_str());
     fprintf(stderr, "%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
@@ -296,7 +300,7 @@ int main(int argc, char ** argv) {
 #endif
                " - Press Return to return control to LLaMa.\n"
                " - If you want to submit another line, end your input in '\\'.\n\n");
-        is_interacting = true;
+        is_interacting = params.interactive_start;
     }
 
     int input_consumed = 0;

tests/CMakeLists.txt

@@ -1,4 +1,9 @@
-set(TEST_TARGET test-tokenizer-0)
-add_executable(${TEST_TARGET} ${TEST_TARGET}.cpp)
-target_link_libraries(${TEST_TARGET} PRIVATE llama ggml utils)
-add_test(NAME ${TEST_TARGET} COMMAND $<TARGET_FILE:${TEST_TARGET}> ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab.bin)
+function(llama_add_test source)
+    get_filename_component(TEST_TARGET ${source} NAME_WE)
+    add_executable(${TEST_TARGET} ${source})
+    target_link_libraries(${TEST_TARGET} PRIVATE llama ggml utils)
+    add_test(NAME ${TEST_TARGET} COMMAND $<TARGET_FILE:${TEST_TARGET}> ${ARGN})
+endfunction()
+
+llama_add_test(test-quantize.c)
+llama_add_test(test-tokenizer-0.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab.bin)

tests/test-quantize.c

@@ -0,0 +1,42 @@
+#include "ggml.h"
+#undef NDEBUG
+#include <assert.h>
+#include <math.h>
+
+int main(void) {
+    #define QK 32
+    float src[QK];
+    uint8_t dst[24];
+    int64_t hist[16];
+
+    for (int i = 0; i < QK; i++) {
+        src[i] = (float)(i + 1);
+    }
+
+    size_t size = ggml_quantize_q4_0(src, dst, QK, QK, QK, hist);
+    assert(size == 20);
+    float max_result = ((float *)dst)[0];
+    float max_expected = src[31] / ((1 << 3) - 1);
+    assert(max_result == max_expected);
+    for (int i = 0; i < QK; i++) {
+        uint8_t q4_result = (i % 2) ? (dst[sizeof(float) + i/2] >> 4) : (dst[sizeof(float) + i/2] & 0xF);
+        uint8_t q4_expected = roundf(src[i] / max_expected) + 8;
+        assert(q4_result == q4_expected);
+    }
+
+    size = ggml_quantize_q4_1(src, dst, QK, QK, QK, hist);
+    assert(size == 24);
+    float delta_result = ((float *)dst)[0];
+    float delta_expected = (src[31] - src[0]) / ((1 << 4) - 1);
+    assert(delta_result == delta_expected);
+    float min_result = ((float *)dst)[1];
+    float min_expected = src[0];
+    assert(min_result == min_expected);
+    for (int i = 0; i < QK; i++) {
+        uint8_t q4_result = (i % 2) ? (dst[sizeof(float)*2 + i/2] >> 4) : (dst[sizeof(float)*2 + i/2] & 0xF);
+        uint8_t q4_expected = roundf((src[i] - min_expected) / delta_expected);
+        assert(q4_result == q4_expected);
+    }
+
+    return 0;
+}

utils.cpp

@@ -63,6 +63,8 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             params.model = argv[++i];
         } else if (arg == "-i" || arg == "--interactive") {
             params.interactive = true;
+        } else if (arg == "--interactive-first") {
+            params.interactive_start = true;
         } else if (arg == "-ins" || arg == "--instruct") {
             params.instruct = true;
         } else if (arg == "--color") {
@@ -96,9 +98,10 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     fprintf(stderr, "options:\n");
     fprintf(stderr, "  -h, --help            show this help message and exit\n");
     fprintf(stderr, "  -i, --interactive     run in interactive mode\n");
+    fprintf(stderr, "  --interactive-first   run in interactive mode and wait for input right away\n");
     fprintf(stderr, "  -ins, --instruct      run in instruction mode (use with Alpaca models)\n");
     fprintf(stderr, "  -r PROMPT, --reverse-prompt PROMPT\n");
-    fprintf(stderr, "                        in interactive mode, poll user input upon seeing PROMPT (can be\n");
+    fprintf(stderr, "                        run in interactive mode and poll user input upon seeing PROMPT (can be\n");
     fprintf(stderr, "                        specified more than once for multiple prompts).\n");
     fprintf(stderr, "  --color               colorise output to distinguish prompt and user input from generations\n");
     fprintf(stderr, "  -s SEED, --seed SEED  RNG seed (default: -1, use random seed for <= 0)\n");
@@ -146,8 +149,10 @@ std::string gpt_random_prompt(std::mt19937 & rng) {
 
 // TODO: not great allocating this every time
 std::vector<llama_token> llama_tokenize(struct llama_context * ctx, const std::string & text, bool add_bos) {
-    std::vector<llama_token> res(8096);
+    // initialize to prompt numer of chars, since n_tokens <= n_prompt_chars
+    std::vector<llama_token> res(text.size() + (int)add_bos);
     int n = llama_tokenize(ctx, text.c_str(), res.data(), res.size(), add_bos);
+    assert(n >= 0);
     res.resize(n);
 
     return res;

utils.h

@@ -38,7 +38,7 @@ struct gpt_params {
     bool random_prompt     = false; // do not randomize prompt if none provided
     bool use_color         = false; // use color to distinguish generations and inputs
     bool interactive       = false; // interactive mode
-    bool interactive_start = false; // reverse prompt immediately
+    bool interactive_start = false; // wait for user input immediately
     bool instruct          = false; // instruction mode (used for Alpaca models)
     bool ignore_eos        = false; // do not stop generating after eos
     bool perplexity        = false; // compute perplexity over the prompt