llama : require first token to be BOS

.gitignore

@@ -42,5 +42,6 @@ zig-out/
 zig-cache/
 
 ppl-*.txt
+qnt-*.txt
 
 examples/jeopardy/results.txt

README.md

@@ -294,17 +294,25 @@ Several quantization methods are supported. They differ in the resulting model d
 
 | Model | Measure      | F16    | Q4_0   | Q4_1   | Q4_2   | Q5_0   | Q5_1   | Q8_0   |
 |------:|--------------|-------:|-------:|-------:|-------:|-------:|-------:|-------:|
-|    7B | perplexity   | 5.9565 | 6.2103 | 6.1286 | 6.1698 | 6.0139 | 5.9934 | 5.9571 |
+|    7B | perplexity   | 5.9066 | 6.1620 | 6.0910 | 6.1466 | 5.9862 | 5.9481 | 5.9069 |
 |    7B | file size    |  13.0G |   4.0G |   4.8G |   4.0G |   4.4G |   4.8G |   7.1G |
 |    7B | ms/tok @ 4th |    128 |     56 |     61 |     84 |     91 |     95 |     75 |
 |    7B | ms/tok @ 8th |    128 |     47 |     55 |     48 |     53 |     59 |     75 |
 |    7B | bits/weight  |   16.0 |    5.0 |    6.0 |    5.0 |    5.5 |    6.0 |    9.0 |
-|   13B | perplexity   | 5.2455 | 5.3748 | 5.3471 | 5.3433 | 5.2768 | 5.2582 | 5.2458 |
+|   13B | perplexity   | 5.2543 | 5.3863 | 5.3607 | 5.3513 | 5.2856 | 5.2706 | 5.2548 |
 |   13B | file size    |  25.0G |   7.6G |   9.1G |   7.6G |   8.4G |   9.1G |    14G |
 |   13B | ms/tok @ 4th |    239 |    104 |    113 |    160 |    176 |    185 |    141 |
 |   13B | ms/tok @ 8th |    240 |     85 |     99 |     97 |    108 |    117 |    147 |
 |   13B | bits/weight  |   16.0 |    5.0 |    6.0 |    5.0 |    5.5 |    6.0 |    9.0 |
 
+### Perplexity (measuring model quality)
+
+You can use the `perplexity` example to measure perplexity over a given prompt (lower perplexity is better).
+For more information, see [https://huggingface.co/docs/transformers/perplexity](https://huggingface.co/docs/transformers/perplexity).
+
+The perplexity measurements in table above are done against the `wikitext2` test dataset (https://paperswithcode.com/dataset/wikitext-2), with context length of 512.
+The time per token is measured on a MacBook M1 Pro 32GB RAM using 4 and 8 threads.
+
 ### Interactive mode
 
 If you want a more ChatGPT-like experience, you can run in interactive mode by passing `-i` as a parameter.
@@ -403,26 +411,6 @@ If your issue is with model generation quality, then please at least scan the fo
     - [Aligning language models to follow instructions](https://openai.com/research/instruction-following)
     - [Training language models to follow instructions with human feedback](https://arxiv.org/abs/2203.02155)
 
-### Perplexity (measuring model quality)
-
-You can use the `perplexity` example to measure perplexity over the given prompt. For more background, see [https://huggingface.co/docs/transformers/perplexity](https://huggingface.co/docs/transformers/perplexity). However, in general, lower perplexity is better for LLMs.
-
-#### Latest measurements
-
-The latest perplexity scores for the various model sizes and quantizations are being tracked in [discussion #406](https://github.com/ggerganov/llama.cpp/discussions/406). `llama.cpp` is measuring very well compared to the baseline implementations. Quantization has a small negative impact on 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 the wikitext2 test dataset (https://paperswithcode.com/dataset/wikitext-2), with default options (512 length context).
-Note that 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

examples/common.cpp

@@ -424,8 +424,8 @@ 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) {
     // 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);
+    std::vector<llama_token> res(text.size() + (int) add_bos);
+    const int n = llama_tokenize(ctx, text.c_str(), res.data(), res.size(), add_bos);
     assert(n >= 0);
     res.resize(n);
 

examples/main/main.cpp

@@ -309,7 +309,8 @@ int main(int argc, char ** argv) {
             if (n_past + (int) embd.size() > n_ctx) {
                 const int n_left = n_past - params.n_keep;
 
-                n_past = params.n_keep;
+                // always keep the first token - BOS
+                n_past = std::max(1, params.n_keep);
 
                 // insert n_left/2 tokens at the start of embd from last_n_tokens
                 embd.insert(embd.begin(), last_n_tokens.begin() + n_ctx - n_left/2 - embd.size(), last_n_tokens.end() - embd.size());
@@ -327,7 +328,6 @@ int main(int argc, char ** argv) {
             }
 
             // try to reuse a matching prefix from the loaded session instead of re-eval (via n_past)
-            // REVIEW
             if (n_session_consumed < (int) session_tokens.size()) {
                 size_t i = 0;
                 for ( ; i < embd.size(); i++) {

examples/perplexity/perplexity.cpp

@@ -25,46 +25,68 @@ 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 `./perplexity -m models/7B/ggml-model-q4_0.bin -f wiki.test.raw`
     // Output: `perplexity: 13.5106 [114/114]`
+    // BOS tokens will be added for each chunk before eval
     auto tokens = ::llama_tokenize(ctx, params.prompt, true);
 
-    int count = 0;
-    int seq_count = tokens.size() / params.n_ctx;
-    int n_vocab = llama_n_vocab(ctx);
+    int count   = 0;
+
+    const int n_chunk = tokens.size() / params.n_ctx;
+    const int n_vocab = llama_n_vocab(ctx);
+    const int n_batch = params.n_batch;
 
     double nll = 0.0;
-    fprintf(stderr, "%s : calculating perplexity over %d chunks, batch_size=%d\n", __func__, seq_count, params.n_batch);
+    fprintf(stderr, "%s: calculating perplexity over %d chunks, batch_size=%d\n", __func__, n_chunk, n_batch);
+
+    for (int i = 0; i < n_chunk; ++i) {
+        const int start =     i * params.n_ctx;
+        const int end   = start + params.n_ctx;
 
-    for (int i = 0; i < seq_count; ++i) {
-        int start = i * params.n_ctx;
-        int end = start + params.n_ctx;
+        const int num_batches = (params.n_ctx + n_batch - 1) / n_batch;
 
         std::vector<float> logits;
-        int num_batches = (params.n_ctx + params.n_batch - 1) / params.n_batch;
-        auto start_t = std::chrono::high_resolution_clock::now();
+
+        const auto t_start = std::chrono::high_resolution_clock::now();
+
         for (int j = 0; j < num_batches; ++j) {
-            int batch_start = start + j * params.n_batch;
-            int batch_size = std::min(end - batch_start, params.n_batch);
-            if (llama_eval(ctx, tokens.data() + batch_start, batch_size, j * params.n_batch, params.n_threads)) {
+            const int batch_start = start + j * n_batch;
+            const int batch_size  = std::min(end - batch_start, n_batch);
+
+            // save original token and restore it after eval
+            const auto token_org = tokens[batch_start];
+
+            // add BOS token for the first batch of each chunk
+            if (j == 0) {
+                tokens[batch_start] = llama_token_bos();
+            }
+
+            if (llama_eval(ctx, tokens.data() + batch_start, batch_size, j * n_batch, params.n_threads)) {
                 fprintf(stderr, "%s : failed to eval\n", __func__);
                 return;
             }
-            auto batch_logits = llama_get_logits(ctx);
+
+            // restore the original token in case it was set to BOS
+            tokens[batch_start] = token_org;
+
+            const auto batch_logits = llama_get_logits(ctx);
             logits.insert(logits.end(), batch_logits, batch_logits + batch_size * n_vocab);
         }
-        auto end_t = std::chrono::high_resolution_clock::now();
+
+        const auto t_end = std::chrono::high_resolution_clock::now();
+
         if (i == 0) {
-            const float seconds = std::chrono::duration<float>(end_t - start_t).count();
-            printf("%.2f seconds per pass - ETA ", seconds);
-            int total_seconds = (int)(seconds * seq_count);
+            const float t_total = std::chrono::duration<float>(t_end - t_start).count();
+            fprintf(stderr, "%s: %.2f seconds per pass - ETA ", __func__, t_total);
+            int total_seconds = (int)(t_total * n_chunk);
             if (total_seconds >= 60*60) {
-                printf("%d hours ", total_seconds / (60*60));
+                fprintf(stderr, "%d hours ", total_seconds / (60*60));
                 total_seconds = total_seconds % (60*60);
             }
-            printf("%d minutes\n", total_seconds / 60);
+            fprintf(stderr, "%d minutes\n", total_seconds / 60);
         }
+
         // We get the logits for all the tokens in the context window (params.n_ctx)
         // from llama_eval above.  Now, based on https://huggingface.co/docs/transformers/perplexity,
-        // calculate the perplexity over the last half the window (so the model always has
+        // calculate the perplexity over the last half of the window (so the model always has
         // some context to predict the token).
         //
         // We rely on the fact that attention in the forward pass only looks at previous
@@ -76,10 +98,12 @@ void perplexity(llama_context * ctx, const gpt_params & params) {
         // process the entire prompt.
         for (int j = std::min(512, params.n_ctx / 2); j < params.n_ctx - 1; ++j) {
             // Calculate probability of next token, given the previous ones.
-            std::vector<float> tok_logits(
-                logits.begin() + j * n_vocab,
+            const std::vector<float> tok_logits(
+                logits.begin() + (j + 0) * n_vocab,
                 logits.begin() + (j + 1) * n_vocab);
-            float prob = softmax(tok_logits)[tokens[start + j + 1]];
+
+            const float prob = softmax(tok_logits)[tokens[start + j + 1]];
+
             nll += -std::log(prob);
             ++count;
         }

llama.cpp

@@ -1052,6 +1052,13 @@ static bool llama_eval_internal(
             const int   n_tokens,
             const int   n_past,
             const int   n_threads) {
+
+    // enforce that the first token is BOS
+    if (n_past == 0 && tokens[0] != llama_token_bos()) {
+        fprintf(stderr, "%s: first token must be BOS\n", __func__);
+        return false;
+    }
+
     const int64_t t_start_us = ggml_time_us();
 
     const int N = n_tokens;
@@ -1482,7 +1489,7 @@ static std::vector<llama_vocab::id> llama_tokenize(const llama_vocab & vocab, co
     }
 
     if (bos) {
-        output.push_back(1);
+        output.push_back(llama_token_bos());
     }
 
     tokenizer.tokenize(text, output);
@@ -2727,11 +2734,14 @@ int llama_eval(
         fprintf(stderr, "%s: failed to eval\n", __func__);
         return 1;
     }
+
     // get a more accurate load time, upon first eval
+    // TODO: fix this
     if (!ctx->has_evaluated_once) {
         ctx->t_load_us = ggml_time_us() - ctx->t_start_us;
         ctx->has_evaluated_once = true;
     }
+
     return 0;
 }
 

scripts/ppl-run-all.sh

@@ -0,0 +1,43 @@
+#!/bin/bash
+
+#
+# quantize
+#
+
+# 7B
+time ./bin/quantize ../models/7B/ggml-model-f16.bin ../models/7B/ggml-model-q4_0.bin q4_0 2>&1 | tee ../qnt-7b-q4_0.txt
+time ./bin/quantize ../models/7B/ggml-model-f16.bin ../models/7B/ggml-model-q4_1.bin q4_1 2>&1 | tee ../qnt-7b-q4_1.txt
+time ./bin/quantize ../models/7B/ggml-model-f16.bin ../models/7B/ggml-model-q4_2.bin q4_2 2>&1 | tee ../qnt-7b-q4_2.txt
+time ./bin/quantize ../models/7B/ggml-model-f16.bin ../models/7B/ggml-model-q5_0.bin q5_0 2>&1 | tee ../qnt-7b-q5_0.txt
+time ./bin/quantize ../models/7B/ggml-model-f16.bin ../models/7B/ggml-model-q5_1.bin q5_1 2>&1 | tee ../qnt-7b-q5_1.txt
+time ./bin/quantize ../models/7B/ggml-model-f16.bin ../models/7B/ggml-model-q8_0.bin q8_0 2>&1 | tee ../qnt-7b-q8_0.txt
+
+# 13B
+time ./bin/quantize ../models/13B/ggml-model-f16.bin ../models/13B/ggml-model-q4_0.bin q4_0 2>&1 | tee ../qnt-13b-q4_0.txt
+time ./bin/quantize ../models/13B/ggml-model-f16.bin ../models/13B/ggml-model-q4_1.bin q4_1 2>&1 | tee ../qnt-13b-q4_1.txt
+time ./bin/quantize ../models/13B/ggml-model-f16.bin ../models/13B/ggml-model-q4_2.bin q4_2 2>&1 | tee ../qnt-13b-q4_2.txt
+time ./bin/quantize ../models/13B/ggml-model-f16.bin ../models/13B/ggml-model-q5_0.bin q5_0 2>&1 | tee ../qnt-13b-q5_0.txt
+time ./bin/quantize ../models/13B/ggml-model-f16.bin ../models/13B/ggml-model-q5_1.bin q5_1 2>&1 | tee ../qnt-13b-q5_1.txt
+time ./bin/quantize ../models/13B/ggml-model-f16.bin ../models/13B/ggml-model-q8_0.bin q8_0 2>&1 | tee ../qnt-13b-q8_0.txt
+
+#
+# perplexity
+#
+
+# 7B
+time ./bin/perplexity -m ../models/7B/ggml-model-f16.bin  -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-7b-f16.txt
+time ./bin/perplexity -m ../models/7B/ggml-model-q4_0.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-7b-q4_0.txt
+time ./bin/perplexity -m ../models/7B/ggml-model-q4_1.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-7b-q4_1.txt
+time ./bin/perplexity -m ../models/7B/ggml-model-q4_2.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-7b-q4_2.txt
+time ./bin/perplexity -m ../models/7B/ggml-model-q5_0.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-7b-q5_0.txt
+time ./bin/perplexity -m ../models/7B/ggml-model-q5_1.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-7b-q5_1.txt
+time ./bin/perplexity -m ../models/7B/ggml-model-q8_0.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-7b-q8_0.txt
+
+# 13B
+time ./bin/perplexity -m ../models/13B/ggml-model-f16.bin  -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-13b-f16.txt
+time ./bin/perplexity -m ../models/13B/ggml-model-q4_0.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-13b-q4_0.txt
+time ./bin/perplexity -m ../models/13B/ggml-model-q4_1.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-13b-q4_1.txt
+time ./bin/perplexity -m ../models/13B/ggml-model-q4_2.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-13b-q4_2.txt
+time ./bin/perplexity -m ../models/13B/ggml-model-q5_0.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-13b-q5_0.txt
+time ./bin/perplexity -m ../models/13B/ggml-model-q5_1.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-13b-q5_1.txt
+time ./bin/perplexity -m ../models/13B/ggml-model-q8_0.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-13b-q8_0.txt