Merge branch 'ggerganov:master' into systemd-units

README.md

@@ -11,6 +11,10 @@ Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
 
 ### Hot topics
 
+- #### IMPORTANT: Tokenizer fixes and API change (developers and projects using `llama.cpp` built-in tokenization must read): https://github.com/ggerganov/llama.cpp/pull/2810
+
+- GGUFv2 adds support for 64-bit sizes + backwards compatible: https://github.com/ggerganov/llama.cpp/pull/2821
+
 - Added support for Falcon models: https://github.com/ggerganov/llama.cpp/pull/2717
 
 - A new file format has been introduced: [GGUF](https://github.com/ggerganov/llama.cpp/pull/2398)

ci/run.sh

@@ -196,17 +196,17 @@ function gg_run_open_llama_3b_v2 {
     (time ./bin/main --model ${model_q5_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
     (time ./bin/main --model ${model_q6_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
 
-    (time ./bin/perplexity --model ${model_f16}  -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/perplexity --model ${model_f16}  -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/perplexity --model ${model_q8_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/perplexity --model ${model_q8_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
-    (time ./bin/perplexity --model ${model_q4_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/perplexity --model ${model_q4_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
-    (time ./bin/perplexity --model ${model_q4_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/perplexity --model ${model_q4_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
-    (time ./bin/perplexity --model ${model_q5_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/perplexity --model ${model_q5_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
-    (time ./bin/perplexity --model ${model_q5_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/perplexity --model ${model_q5_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
-    (time ./bin/perplexity --model ${model_q2_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/perplexity --model ${model_q2_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
-    (time ./bin/perplexity --model ${model_q3_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/perplexity --model ${model_q3_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
-    (time ./bin/perplexity --model ${model_q4_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/perplexity --model ${model_q4_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
-    (time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
-    (time ./bin/perplexity --model ${model_q6_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+    (time ./bin/perplexity --model ${model_q6_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
 
     function check_ppl {
         qnt="$1"
@@ -233,6 +233,48 @@ function gg_run_open_llama_3b_v2 {
     check_ppl "q5_k" "$(cat $OUT/${ci}-tg-q5_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
     check_ppl "q6_k" "$(cat $OUT/${ci}-tg-q6_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
 
+    # lora
+    function compare_ppl {
+        qnt="$1"
+        ppl1=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
+        ppl2=$(echo "$3" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
+
+        if [ $(echo "$ppl1 < $ppl2" | bc) -eq 1 ]; then
+            printf '  - %s @ %s (FAIL: %s > %s)\n' "$qnt" "$ppl" "$ppl1" "$ppl2"
+            return 20
+        fi
+
+        printf '  - %s @ %s %s OK\n' "$qnt" "$ppl1" "$ppl2"
+        return 0
+    }
+
+    path_lora="../models-mnt/open-llama/3B-v2/lora"
+    path_shakespeare="../models-mnt/shakespeare"
+
+    shakespeare="${path_shakespeare}/shakespeare.txt"
+    lora_shakespeare="${path_lora}/ggml-adapter-model.bin"
+
+    gg_wget ${path_lora} https://huggingface.co/slaren/open_llama_3b_v2_shakespeare_lora/resolve/main/adapter_config.json
+    gg_wget ${path_lora} https://huggingface.co/slaren/open_llama_3b_v2_shakespeare_lora/resolve/main/adapter_model.bin
+    gg_wget ${path_shakespeare} https://huggingface.co/slaren/open_llama_3b_v2_shakespeare_lora/resolve/main/shakespeare.txt
+
+    python3 ../convert-lora-to-ggml.py ${path_lora}
+
+    # f16
+    (time ./bin/perplexity --model ${model_f16} -f ${shakespeare}                            -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-f16.log
+    (time ./bin/perplexity --model ${model_f16} -f ${shakespeare} --lora ${lora_shakespeare} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-f16.log
+    compare_ppl "f16 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-f16.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-f16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
+
+    # q8_0
+    (time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare}                            -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-q8_0.log
+    (time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-q8_0.log
+    compare_ppl "q8_0 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
+
+    # q8_0 + f16 lora-base
+    (time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} --lora-base ${model_f16} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log
+    compare_ppl "q8_0 / f16 base shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
+
+
     set +e
 }
 
@@ -242,6 +284,7 @@ function gg_sum_open_llama_3b_v2 {
     gg_printf 'OpenLLaMA 3B-v2:\n'
     gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
     gg_printf '- perplexity:\n%s\n' "$(cat $OUT/${ci}-ppl.log)"
+    gg_printf '- lora:\n%s\n' "$(cat $OUT/${ci}-lora-ppl.log)"
     gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
     gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
     gg_printf '- q4_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_0.log)"
@@ -253,6 +296,11 @@ function gg_sum_open_llama_3b_v2 {
     gg_printf '- q4_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_k.log)"
     gg_printf '- q5_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_k.log)"
     gg_printf '- q6_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q6_k.log)"
+    gg_printf '- shakespeare (f16):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-f16.log)"
+    gg_printf '- shakespeare (f16 lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-f16.log)"
+    gg_printf '- shakespeare (q8_0):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log)"
+    gg_printf '- shakespeare (q8_0 lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0.log)"
+    gg_printf '- shakespeare (q8_0 / f16 base lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log)"
 }
 
 # open_llama_7b_v2
@@ -310,17 +358,17 @@ function gg_run_open_llama_7b_v2 {
     ./bin/quantize ${model_f16} ${model_q5_k} q5_k
     ./bin/quantize ${model_f16} ${model_q6_k} q6_k
 
-    (time ./bin/main --model ${model_f16}  -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/main --model ${model_f16}  -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/main --model ${model_q8_0} -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/main --model ${model_q8_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
-    (time ./bin/main --model ${model_q4_0} -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/main --model ${model_q4_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
-    (time ./bin/main --model ${model_q4_1} -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/main --model ${model_q4_1} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
-    (time ./bin/main --model ${model_q5_0} -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/main --model ${model_q5_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
-    (time ./bin/main --model ${model_q5_1} -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/main --model ${model_q5_1} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
-    (time ./bin/main --model ${model_q2_k} -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/main --model ${model_q2_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
-    (time ./bin/main --model ${model_q3_k} -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/main --model ${model_q3_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
-    (time ./bin/main --model ${model_q4_k} -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/main --model ${model_q4_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
-    (time ./bin/main --model ${model_q5_k} -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/main --model ${model_q5_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
-    (time ./bin/main --model ${model_q6_k} -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+    (time ./bin/main --model ${model_q6_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
 
     (time ./bin/perplexity --model ${model_f16}  -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
     (time ./bin/perplexity --model ${model_q8_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
@@ -359,6 +407,48 @@ function gg_run_open_llama_7b_v2 {
     check_ppl "q5_k" "$(cat $OUT/${ci}-tg-q5_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
     check_ppl "q6_k" "$(cat $OUT/${ci}-tg-q6_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
 
+    # lora
+    function compare_ppl {
+        qnt="$1"
+        ppl1=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
+        ppl2=$(echo "$3" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
+
+        if [ $(echo "$ppl1 < $ppl2" | bc) -eq 1 ]; then
+            printf '  - %s @ %s (FAIL: %s > %s)\n' "$qnt" "$ppl" "$ppl1" "$ppl2"
+            return 20
+        fi
+
+        printf '  - %s @ %s %s OK\n' "$qnt" "$ppl1" "$ppl2"
+        return 0
+    }
+
+    path_lora="../models-mnt/open-llama/7B-v2/lora"
+    path_shakespeare="../models-mnt/shakespeare"
+
+    shakespeare="${path_shakespeare}/shakespeare.txt"
+    lora_shakespeare="${path_lora}/ggml-adapter-model.bin"
+
+    gg_wget ${path_lora} https://huggingface.co/slaren/open_llama_7b_v2_shakespeare_lora/resolve/main/adapter_config.json
+    gg_wget ${path_lora} https://huggingface.co/slaren/open_llama_7b_v2_shakespeare_lora/resolve/main/adapter_model.bin
+    gg_wget ${path_shakespeare} https://huggingface.co/slaren/open_llama_7b_v2_shakespeare_lora/resolve/main/shakespeare.txt
+
+    python3 ../convert-lora-to-ggml.py ${path_lora}
+
+    # f16
+    (time ./bin/perplexity --model ${model_f16} -f ${shakespeare}                            -t 1 -ngl 999 -c 2048 -b 512 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-f16.log
+    (time ./bin/perplexity --model ${model_f16} -f ${shakespeare} --lora ${lora_shakespeare} -t 1 -ngl 999 -c 2048 -b 512 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-f16.log
+    compare_ppl "f16 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-f16.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-f16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
+
+    # currently not supported by the CUDA backend
+    # q8_0
+    #(time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare}                            -t 1 -ngl 999 -c 2048 -b 512 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-q8_0.log
+    #(time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} -t 1 -ngl 999 -c 2048 -b 512 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-q8_0.log
+    #compare_ppl "q8_0 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
+
+    # q8_0 + f16 lora-base
+    #(time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} --lora-base ${model_f16} -t 1 -ngl 999 -c 2048 -b 512 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log
+    #compare_ppl "q8_0 / f16 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
+
     set +e
 }
 
@@ -368,6 +458,7 @@ function gg_sum_open_llama_7b_v2 {
     gg_printf 'OpenLLaMA 7B-v2:\n'
     gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
     gg_printf '- perplexity:\n%s\n' "$(cat $OUT/${ci}-ppl.log)"
+    gg_printf '- lora:\n%s\n' "$(cat $OUT/${ci}-lora-ppl.log)"
     gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
     gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
     gg_printf '- q4_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_0.log)"
@@ -379,6 +470,11 @@ function gg_sum_open_llama_7b_v2 {
     gg_printf '- q4_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_k.log)"
     gg_printf '- q5_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_k.log)"
     gg_printf '- q6_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q6_k.log)"
+    gg_printf '- shakespeare (f16):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-f16.log)"
+    gg_printf '- shakespeare (f16 lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-f16.log)"
+    #gg_printf '- shakespeare (q8_0):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log)"
+    #gg_printf '- shakespeare (q8_0 lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0.log)"
+    #gg_printf '- shakespeare (q8_0 / f16 base lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log)"
 }
 
 ## main

common/common.cpp

@@ -733,12 +733,12 @@ std::vector<llama_token> llama_tokenize(
     return result;
 }
 
-std::string llama_token_to_str(const struct llama_context * ctx, llama_token token) {
+std::string llama_token_to_piece(const struct llama_context * ctx, llama_token token) {
     std::vector<char> result(8, 0);
-    const int n_tokens = llama_token_to_str(ctx, token, result.data(), result.size());
+    const int n_tokens = llama_token_to_piece(ctx, token, result.data(), result.size());
     if (n_tokens < 0) {
         result.resize(-n_tokens);
-        int check = llama_token_to_str(ctx, token, result.data(), result.size());
+        int check = llama_token_to_piece(ctx, token, result.data(), result.size());
         GGML_ASSERT(check == -n_tokens);
     } else {
         result.resize(n_tokens);
@@ -746,3 +746,36 @@ std::string llama_token_to_str(const struct llama_context * ctx, llama_token tok
 
     return std::string(result.data(), result.size());
 }
+
+std::string llama_detokenize_spm(llama_context * ctx, const std::vector<llama_token> & tokens) {
+    const llama_token bos_id = llama_token_bos(ctx);
+
+    std::string piece;
+    std::string result;
+
+    for (size_t i = 0; i < tokens.size(); ++i) {
+        piece = llama_token_to_piece(ctx, tokens[i]);
+
+        // remove the leading space of the first non-BOS token
+        if (((tokens[0] == bos_id && i == 1) || (tokens[0] != bos_id && i == 0)) && piece[0] == ' ') {
+            piece = piece.substr(1);
+        }
+
+        result += piece;
+    }
+
+    return result;
+}
+
+std::string llama_detokenize_bpe(llama_context * ctx, const std::vector<llama_token> & tokens) {
+    std::string piece;
+    std::string result;
+
+    for (size_t i = 0; i < tokens.size(); ++i) {
+        piece = llama_token_to_piece(ctx, tokens[i]);
+
+        result += piece;
+    }
+
+    return result;
+}

common/common.h

@@ -28,6 +28,7 @@ struct gpt_params {
     int32_t main_gpu                        = 0;    // the GPU that is used for scratch and small tensors
     float   tensor_split[LLAMA_MAX_DEVICES] = {0};  // how split tensors should be distributed across GPUs
     int32_t n_probs                         = 0;    // if greater than 0, output the probabilities of top n_probs tokens.
+    int32_t n_beams                         = 0;    // if non-zero then use beam search of given width.
     float   rope_freq_base                  = 10000.0f; // RoPE base frequency
     float   rope_freq_scale                 = 1.0f;     // RoPE frequency scaling factor
 
@@ -115,11 +116,31 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
 // Vocab utils
 //
 
+// tokenizes a string into a vector of tokens
+// should work similar to Python's `tokenizer.encode`
 std::vector<llama_token> llama_tokenize(
         struct llama_context * ctx,
            const std::string & text,
                         bool   add_bos);
 
-std::string llama_token_to_str(
+// tokenizes a token into a piece
+// should work similar to Python's `tokenizer.id_to_piece`
+std::string llama_token_to_piece(
         const struct llama_context * ctx,
                        llama_token   token);
+
+// TODO: these should be moved in llama.h C-style API under single `llama_detokenize` function
+//       that takes into account the tokenizer type and decides how to handle the leading space
+//
+// detokenizes a vector of tokens into a string
+// should work similar to Python's `tokenizer.decode`
+// removes the leading space from the first non-BOS token
+std::string llama_detokenize_spm(
+                         llama_context * ctx,
+        const std::vector<llama_token> & tokens);
+
+// detokenizes a vector of tokens into a string
+// should work similar to Python's `tokenizer.decode`
+std::string llama_detokenize_bpe(
+                         llama_context * ctx,
+        const std::vector<llama_token> & tokens);

convert.py

@@ -3,6 +3,7 @@
 import gguf
 import argparse
 import concurrent.futures
+from concurrent.futures import ThreadPoolExecutor, ProcessPoolExecutor
 import copy
 import enum
 import faulthandler
@@ -17,13 +18,14 @@ import re
 import signal
 import struct
 import sys
+import time
 import zipfile
 import numpy as np
 
 from abc import ABCMeta, abstractmethod
 from dataclasses import dataclass
 from pathlib import Path
-from typing import (IO, TYPE_CHECKING, Any, Callable, Dict, Iterable, List, Literal, Optional, Sequence, Tuple, TypeVar, Union)
+from typing import (IO, TYPE_CHECKING, Any, Callable, Dict, Generator, Iterable, List, Literal, Optional, Sequence, Set, Tuple, TypeVar, Union)
 from sentencepiece import SentencePieceProcessor  # type: ignore
 
 if TYPE_CHECKING:
@@ -37,30 +39,70 @@ NDArray: 'TypeAlias' = 'np.ndarray[Any, Any]'
 ARCH=gguf.MODEL_ARCH.LLAMA
 NAMES=gguf.MODEL_TENSOR_NAMES[ARCH]
 
+DEFAULT_CONCURRENCY = 8
 #
 # data types
 #
 
 @dataclass(frozen=True)
-class UnquantizedDataType:
+class DataType:
     name: str
+    dtype: 'np.dtype[Any]'
+    valid_conversions: List[str]
 
-DT_F16  = UnquantizedDataType('F16')
+    def elements_to_bytes(self, n_elements: int) -> int:
-DT_F32  = UnquantizedDataType('F32')
+        return n_elements * self.dtype.itemsize
-DT_I32  = UnquantizedDataType('I32')
-DT_BF16 = UnquantizedDataType('BF16')
 
-DataType = Union[UnquantizedDataType]
+@dataclass(frozen=True)
+class UnquantizedDataType(DataType):
+    pass
 
-DATA_TYPE_TO_NUMPY: Dict[DataType, 'np.dtype[Any]'] = {
+DT_F16  = UnquantizedDataType('F16', dtype = np.dtype(np.float16), valid_conversions = ['F32', 'Q8_0'])
-    DT_BF16: np.dtype(np.uint16),
+DT_F32  = UnquantizedDataType('F32', dtype = np.dtype(np.float32), valid_conversions = ['F16', 'Q8_0'])
-    DT_F16:  np.dtype(np.float16),
+DT_I32  = UnquantizedDataType('I32', dtype = np.dtype(np.int16), valid_conversions = [])
-    DT_F32:  np.dtype(np.float32),
+DT_BF16 = UnquantizedDataType('BF16', dtype = np.dtype(np.uint16), valid_conversions = ['F32', 'F16', 'Q8_0'])
-    DT_I32:  np.dtype(np.int32),
-}
 
-NUMPY_TYPE_TO_DATA_TYPE: Dict['np.dtype[Any]', DataType] = \
+@dataclass(frozen=True)
-    {dtype: data_type for (data_type, dtype) in DATA_TYPE_TO_NUMPY.items()}
+class QuantizedDataType(DataType):
+    block_size: int
+    quantized_dtype: 'np.dtype[Any]'
+    ggml_type: gguf.GGMLQuantizationType
+
+    def quantize(self, arr: NDArray) -> NDArray:
+        raise NotImplementedError(f'Quantization for {self.name} not implemented')
+
+    def elements_to_bytes(self, n_elements: int) -> int:
+        assert n_elements % self.block_size == 0, f'Invalid number of elements {n_elements} for {self.name} with block size {self.block_size}'
+        return self.quantized_dtype.itemsize * (n_elements // self.block_size)
+
+@dataclass(frozen=True)
+class Q8_0QuantizedDataType(QuantizedDataType):
+    # Mini Q8_0 quantization in Python!
+    def quantize(self, arr: NDArray) -> NDArray:
+        assert arr.size % self.block_size == 0 and arr.size != 0, f'Bad array size {arr.size}'
+        assert arr.dtype == np.float32, f'Bad array type {arr.dtype}'
+        n_blocks = arr.size // self.block_size
+        blocks = arr.reshape((n_blocks, self.block_size))
+        # Much faster implementation of block quantization contributed by @Cebtenzzre
+        def quantize_blocks_q8_0(blocks: NDArray) -> Iterable[Tuple[Any, Any]]:
+            d = abs(blocks).max(axis = 1) / np.float32(127)
+            with np.errstate(divide = 'ignore'):
+                qs = (blocks / d[:, None]).round()
+            qs[d == 0] = 0
+            yield from zip(d, qs)
+        return np.fromiter(quantize_blocks_q8_0(blocks), count = n_blocks, dtype = self.quantized_dtype)
+
+DT_Q8_0 = Q8_0QuantizedDataType('Q8_0',
+    dtype = np.dtype(np.float32), valid_conversions = [],
+    ggml_type = gguf.GGMLQuantizationType.Q8_0, block_size = 32,
+    quantized_dtype = np.dtype([('d', '<f2'), ('qs', 'i1', (32,))]))
+
+# Quantized types skipped here because they may also map to np.float32
+NUMPY_TYPE_TO_DATA_TYPE: Dict['np.dtype[Any]', DataType] = {}
+for dt in (DT_BF16, DT_F16, DT_F32, DT_I32):
+    if dt.dtype in NUMPY_TYPE_TO_DATA_TYPE:
+        raise ValueError(f'Invalid duplicate data type {dt}')
+    NUMPY_TYPE_TO_DATA_TYPE[dt.dtype] = dt
 
 SAFETENSORS_DATA_TYPES: Dict[str, DataType] = {
     'BF16': DT_BF16,
@@ -73,20 +115,22 @@ SAFETENSORS_DATA_TYPES: Dict[str, DataType] = {
 # TODO: rename to LLAMAFileType
 # TODO: move to `gguf.py`
 class GGMLFileType(enum.IntEnum):
-    AllF32    = 0
+    AllF32     = 0
-    MostlyF16 = 1  # except 1d tensors
+    MostlyF16  = 1  # except 1d tensors
+    MostlyQ8_0 = 7  # except 1d tensors
 
     def type_for_tensor(self, name: str, tensor: 'LazyTensor') -> DataType:
-        if len(tensor.shape) == 1:
+        dt = GGML_FILE_TYPE_TO_DATA_TYPE.get(self)
-            # 1D tensors are always F32.
+        if dt is None:
-            return DT_F32
-        elif self == GGMLFileType.AllF32:
-            return DT_F32
-        elif self == GGMLFileType.MostlyF16:
-            return DT_F16
-        else:
             raise ValueError(self)
+        # 1D tensors are always F32.
+        return dt if len(tensor.shape) > 1 else DT_F32
 
+GGML_FILE_TYPE_TO_DATA_TYPE: Dict[GGMLFileType, DataType] = {
+    GGMLFileType.AllF32    : DT_F32,
+    GGMLFileType.MostlyF16 : DT_F16,
+    GGMLFileType.MostlyQ8_0: DT_Q8_0,
+}
 
 #
 # hparams loading
@@ -105,6 +149,7 @@ class Params:
     f_norm_eps: float
 
     f_rope_freq_base: Optional[float] = None
+    f_rope_scale: Optional[float] = None
 
     ftype: Optional[GGMLFileType] = None
 
@@ -160,13 +205,20 @@ class Params:
     def loadHFTransformerJson(model: 'LazyModel', config_path: 'Path') -> 'Params':
         config = json.load(open(config_path))
 
-        n_vocab    = config["vocab_size"]
+        n_vocab          = config["vocab_size"]
-        n_embd     = config["hidden_size"]
+        n_embd           = config["hidden_size"]
-        n_layer    = config["num_hidden_layers"]
+        n_layer          = config["num_hidden_layers"]
-        n_ff       = config["intermediate_size"]
+        n_ff             = config["intermediate_size"]
-        n_head     = config["num_attention_heads"]
+        n_head           = config["num_attention_heads"]
-        n_head_kv  = config["num_key_value_heads"] if "num_key_value_heads" in config else n_head
+        n_head_kv        = config["num_key_value_heads"] if "num_key_value_heads" in config else n_head
-        f_norm_eps = config["rms_norm_eps"]
+        f_norm_eps       = config["rms_norm_eps"]
+        f_rope_freq_base = config["rope_theta"] if "rope_theta" in config else None
+
+        rope_scaling = config.get("rope_scaling")
+        if isinstance(rope_scaling, dict) and rope_scaling.get("type") == "linear":
+            f_rope_scale = config["rope_scaling"].get("factor")
+        else:
+            f_rope_scale = None
 
         n_mult = Params.find_n_mult(n_ff, n_embd)
 
@@ -179,15 +231,17 @@ class Params:
                             "Suggestion: provide 'config.json' of the model in the same directory containing model files.")
 
         return Params(
-            n_vocab    = n_vocab,
+            n_vocab          = n_vocab,
-            n_embd     = n_embd,
+            n_embd           = n_embd,
-            n_mult     = n_mult,
+            n_mult           = n_mult,
-            n_layer    = n_layer,
+            n_layer          = n_layer,
-            n_ctx      = n_ctx,
+            n_ctx            = n_ctx,
-            n_ff       = n_ff,
+            n_ff             = n_ff,
-            n_head     = n_head,
+            n_head           = n_head,
-            n_head_kv  = n_head_kv,
+            n_head_kv        = n_head_kv,
-            f_norm_eps = f_norm_eps,
+            f_norm_eps       = f_norm_eps,
+            f_rope_freq_base = f_rope_freq_base,
+            f_rope_scale     = f_rope_scale,
         )
 
     # LLaMA v2 70B params.json
@@ -405,7 +459,7 @@ class UnquantizedTensor(Tensor):
         self.data_type = NUMPY_TYPE_TO_DATA_TYPE[ndarray.dtype]
 
     def astype(self, data_type: DataType) -> Tensor:
-        dtype = DATA_TYPE_TO_NUMPY[data_type]
+        dtype = data_type.dtype
         if self.data_type == DT_BF16:
             self.ndarray = bf16_to_fp32(self.ndarray)
         return UnquantizedTensor(self.ndarray.astype(dtype))
@@ -444,22 +498,6 @@ def load_unquantized(lazy_tensor: 'LazyTensor', expected_dtype: Any = None, conv
 GGMLCompatibleTensor = Union[UnquantizedTensor]
 
 
-class DeferredPermutedTensor(Tensor):
-    def __init__(self, base: Tensor, n_head: int, n_head_kv: int) -> None:
-        self.base = base
-        self.n_head = n_head
-        self.data_type = self.base.data_type
-
-    def astype(self, data_type: DataType) -> Tensor:
-        return self.base.astype(data_type).permute(self.n_head, self.n_head_kv)
-
-    def to_ggml(self) -> GGMLCompatibleTensor:
-        return self.base.to_ggml().permute(self.n_head, self.n_head_kv)
-
-    def permute(self, n_head: int, n_head_kv: int) -> Tensor:
-        raise Exception("shouldn't permute twice")
-
-
 @dataclass
 class LazyTensor:
     _load: Callable[[], Tensor]
@@ -469,7 +507,9 @@ class LazyTensor:
 
     def load(self) -> Tensor:
         ret = self._load()
-        assert ret.data_type == self.data_type, (self.data_type, ret.data_type, self.description)
+        # Should be okay if it maps to the same numpy type?
+        assert ret.data_type == self.data_type or (self.data_type.dtype == ret.data_type.dtype), \
+                (self.data_type, ret.data_type, self.description)
         return ret
 
     def astype(self, data_type: DataType) -> 'LazyTensor':
@@ -480,8 +520,8 @@ class LazyTensor:
         return LazyTensor(load, self.shape, data_type, f'convert({data_type}) {self.description}')
 
     def validate_conversion_to(self, data_type: DataType) -> None:
-        if data_type == self.data_type:
+        if data_type != self.data_type and data_type.name not in self.data_type.valid_conversions:
-            return
+            raise ValueError(f'Cannot validate conversion from {self.data_type} to {data_type}.')
 
 
 LazyModel = Dict[str, LazyTensor]
@@ -607,9 +647,7 @@ class LazyUnpickler(pickle.Unpickler):
         info = self.zip_file.getinfo(filename)
 
         def load(offset: int, elm_count: int) -> NDArray:
-            dtype = DATA_TYPE_TO_NUMPY.get(data_type)
+            dtype = data_type.dtype
-            if dtype is None:
-                raise Exception("tensor stored in unsupported format")
             fp = self.zip_file.open(info)
             fp.seek(offset * dtype.itemsize)
             size = elm_count * dtype.itemsize
@@ -673,7 +711,7 @@ def lazy_load_safetensors_file(fp: IO[bytes], path: Path) -> ModelPlus:
 
     def convert(info: Dict[str, Any]) -> LazyTensor:
         data_type = SAFETENSORS_DATA_TYPES[info['dtype']]
-        numpy_dtype = DATA_TYPE_TO_NUMPY[data_type]
+        numpy_dtype = data_type.dtype
         shape: List[int] = info['shape']
         begin, end = info['data_offsets']
         assert 0 <= begin <= end <= len(byte_buf)
@@ -713,23 +751,35 @@ def lazy_load_file(path: Path) -> ModelPlus:
 In = TypeVar('In')
 Out = TypeVar('Out')
 
-def bounded_parallel_map(func: Callable[[In], Out], iterable: Iterable[In], concurrency: int) -> Iterable[Out]:
+def bounded_parallel_map(func: Callable[[In], Out], iterable: Iterable[In], concurrency: int, max_workers: Optional[int] = None, factory: Callable = ThreadPoolExecutor) -> Iterable[Out]:
     '''Parallel map, but with backpressure.  If the caller doesn't call `next`
     fast enough, this will stop calling `func` at some point rather than
     letting results pile up in memory.  Specifically, there is a max of one
     output value buffered per thread.'''
-    with concurrent.futures.ThreadPoolExecutor() as executor:
+    if concurrency < 2:
+        yield from map(func, iterable)
+        # Not reached.
+    iterable = iter(iterable)
+    with factory(max_workers = max_workers) as executor:
         futures: List[concurrent.futures.Future[Out]] = []
-        items_rev = list(iterable)[::-1]
+        done = False
-        for i in range(min(concurrency, len(items_rev))):
+        for _ in range(concurrency):
-            futures.append(executor.submit(func, items_rev.pop()))
+            try:
+                futures.append(executor.submit(func, next(iterable)))
+            except StopIteration:
+                done = True
+                break
+
         while futures:
             result = futures.pop(0).result()
-            if items_rev:
+            while not done and len(futures) < concurrency:
-                futures.append(executor.submit(func, items_rev.pop()))
+                try:
+                    futures.append(executor.submit(func, next(iterable)))
+                except StopIteration:
+                    done = True
+                    break
             yield result
 
-
 def check_vocab_size(params: Params, vocab: Vocab) -> None:
     if params.n_vocab != vocab.vocab_size:
         assert isinstance(vocab, BpeVocab) or isinstance(vocab, SentencePieceVocab)
@@ -771,6 +821,9 @@ class OutputFile:
         if params.f_rope_freq_base:
             self.gguf.add_rope_freq_base(params.f_rope_freq_base)
 
+        if params.f_rope_scale:
+            self.gguf.add_rope_scale_linear(params.f_rope_scale)
+
         if params.ftype:
             self.gguf.add_file_type(params.ftype)
 
@@ -791,12 +844,11 @@ class OutputFile:
         self.gguf.add_token_types(toktypes)
 
     def add_tensor_info(self, name: str, tensor: LazyTensor) -> None:
-        n_elements = 1
+        n_elements = int(np.prod(tensor.shape))
-        for dim in tensor.shape:
+        raw_dtype = getattr(tensor.data_type, 'ggml_type', None)
-            n_elements *= dim
+        data_type = getattr(tensor.data_type, 'quantized_type', None) or tensor.data_type.dtype
-        data_type = DATA_TYPE_TO_NUMPY[tensor.data_type]
+        data_nbytes = tensor.data_type.elements_to_bytes(n_elements)
-        data_nbytes = n_elements * data_type.itemsize
+        self.gguf.add_tensor_info(name, tensor.shape, data_type, data_nbytes, raw_dtype = raw_dtype)
-        self.gguf.add_tensor_info(name, tensor.shape, data_type, data_nbytes)
 
     def write_meta(self) -> None:
         self.gguf.write_header_to_file()
@@ -822,7 +874,20 @@ class OutputFile:
         of.close()
 
     @staticmethod
-    def write_all(fname_out: Path, params: Params, model: LazyModel, vocab: Vocab) -> None:
+    def do_item(item: Tuple[str, LazyTensor]) -> Tuple[DataType, NDArray]:
+        name, lazy_tensor = item
+        tensor = lazy_tensor.load().to_ggml()
+        return (lazy_tensor.data_type, tensor.ndarray)
+
+    @staticmethod
+    def maybe_do_quantize(item: Tuple[DataType, NDArray]) -> NDArray:
+        dt, arr = item
+        if not isinstance(dt, QuantizedDataType):
+            return arr
+        return dt.quantize(arr)
+
+    @staticmethod
+    def write_all(fname_out: Path, ftype: GGMLFileType, params: Params, model: LazyModel, vocab: Vocab, concurrency: int = DEFAULT_CONCURRENCY) -> None:
         check_vocab_size(params, vocab)
 
         of = OutputFile(fname_out)
@@ -838,16 +903,19 @@ class OutputFile:
         of.write_meta()
         of.write_tensor_info()
 
-        def do_item(item: Tuple[str, LazyTensor]) -> NDArray:
-            name, lazy_tensor = item
-            return lazy_tensor.load().to_ggml().ndarray
-
         # tensor data
-        ndarrays = bounded_parallel_map(do_item, model.items(), concurrency=8)
+        ndarrays_inner = bounded_parallel_map(OutputFile.do_item, model.items(), concurrency = concurrency)
+        if ftype == GGMLFileType.MostlyQ8_0:
+            ndarrays = bounded_parallel_map(OutputFile.maybe_do_quantize, ndarrays_inner, concurrency = concurrency, max_workers = concurrency, factory = ProcessPoolExecutor)
+        else:
+            ndarrays = map(OutputFile.maybe_do_quantize, ndarrays_inner)
+
+        start = time.time()
         for i, ((name, lazy_tensor), ndarray) in enumerate(zip(model.items(), ndarrays)):
+            elapsed = time.time() - start
             size = ' x '.join(f"{dim:6d}" for dim in lazy_tensor.shape)
             padi = len(str(len(model)))
-            print(f"[{i+1:{padi}d}/{len(model)}] Writing tensor {name:38s} | size {size:16} | type {lazy_tensor.data_type}")
+            print(f"[{i+1:{padi}d}/{len(model)}] Writing tensor {name:38s} | size {size:16} | type {lazy_tensor.data_type.name:4} | T+{int(elapsed):4}")
             of.gguf.write_tensor_data(ndarray)
 
         of.close()
@@ -859,6 +927,8 @@ def pick_output_type(model: LazyModel, output_type_str: Optional[str]) -> GGMLFi
         return GGMLFileType.AllF32
     if output_type_str == "f16" or (output_type_str is None and wq_type in (DT_F16, DT_BF16)):
         return GGMLFileType.MostlyF16
+    if output_type_str == "q8_0":
+        return GGMLFileType.MostlyQ8_0
 
     name_to_type = {name: lazy_tensor.data_type for (name, lazy_tensor) in model.items()}
 
@@ -905,7 +975,7 @@ def convert_model_names(model: LazyModel, params: Params) -> LazyModel:
             print(f"skipping tensor {name_new}")
             continue
         else:
-            print(f"{name:48s} -> {name_new:40s} | {lazy_tensor.data_type} | {lazy_tensor.shape}")
+            print(f"{name:48s} -> {name_new:40s} | {lazy_tensor.data_type.name:6s} | {lazy_tensor.shape}")
             out[name_new] = lazy_tensor
 
     return out
@@ -1010,6 +1080,7 @@ def default_outfile(model_paths: List[Path], file_type: GGMLFileType) -> Path:
     namestr = {
         GGMLFileType.AllF32:    "f32",
         GGMLFileType.MostlyF16: "f16",
+        GGMLFileType.MostlyQ8_0:"q8_0",
     }[file_type]
     ret = model_paths[0].parent / f"ggml-model-{namestr}.gguf"
     if ret in model_paths:
@@ -1033,12 +1104,13 @@ def main(args_in: Optional[List[str]] = None) -> None:
     parser.add_argument("--dump",        action="store_true",    help="don't convert, just show what's in the model")
     parser.add_argument("--dump-single", action="store_true",    help="don't convert, just show what's in a single model file")
     parser.add_argument("--vocab-only",  action="store_true",    help="extract only the vocab")
-    parser.add_argument("--outtype",     choices=["f32", "f16"], help="output format (default: based on input)")
+    parser.add_argument("--outtype",     choices=["f32", "f16", "q8_0"], help="output format - note: q8_0 may be very slow (default: f16 or f32 based on input)")
     parser.add_argument("--vocab-dir",   type=Path,              help="directory containing tokenizer.model, if separate from model file")
     parser.add_argument("--outfile",     type=Path,              help="path to write to; default: based on input")
     parser.add_argument("model",         type=Path,              help="directory containing model file, or model file itself (*.pth, *.pt, *.bin)")
     parser.add_argument("--vocabtype",   choices=["spm", "bpe"], help="vocab format (default: spm)", default="spm")
     parser.add_argument("--ctx",         type=int,               help="model training context (default: based on input)")
+    parser.add_argument("--concurrency", type=int,               help=f"concurrency used for conversion (default: {DEFAULT_CONCURRENCY})", default = DEFAULT_CONCURRENCY)
     args = parser.parse_args(args_in)
 
     if args.dump_single:
@@ -1060,6 +1132,7 @@ def main(args_in: Optional[List[str]] = None) -> None:
         params.ftype = {
             "f32": GGMLFileType.AllF32,
             "f16": GGMLFileType.MostlyF16,
+            "q8_0": GGMLFileType.MostlyQ8_0,
         }[args.outtype]
 
     print(f"params = {params}")
@@ -1091,7 +1164,7 @@ def main(args_in: Optional[List[str]] = None) -> None:
         params.ftype = ftype
         print(f"Writing {outfile}, format {ftype}")
 
-        OutputFile.write_all(outfile, params, model, vocab)
+        OutputFile.write_all(outfile, ftype, params, model, vocab, concurrency = args.concurrency)
         print(f"Wrote {outfile}")
 
 

examples/CMakeLists.txt

@@ -25,6 +25,7 @@ else()
     add_subdirectory(simple)
     add_subdirectory(embd-input)
     add_subdirectory(llama-bench)
+    add_subdirectory(beam_search)
     if (LLAMA_METAL)
         add_subdirectory(metal)
     endif()

examples/beam_search/CMakeLists.txt

@@ -0,0 +1,8 @@
+set(TARGET beam_search)
+add_executable(${TARGET} beam_search.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)
+if(TARGET BUILD_INFO)
+  add_dependencies(${TARGET} BUILD_INFO)
+endif()

examples/beam_search/beam_search.cpp

@@ -0,0 +1,188 @@
+#ifndef _GNU_SOURCE
+#define _GNU_SOURCE
+#endif
+
+#include "common.h"
+#include "llama.h"
+#include "build-info.h"
+
+#include <cassert>
+#include <cinttypes>
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <ctime>
+#include <fstream>
+#include <iostream>
+#include <string>
+#include <vector>
+
+#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
+#include <signal.h>
+#include <unistd.h>
+#elif defined (_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#define NOMINMAX
+#include <windows.h>
+#include <signal.h>
+#endif
+
+// Used for debugging to print out beam tokens.
+struct ostream_beam_view {
+    llama_context * ctx;
+    llama_beam_view beam_view;
+};
+std::ostream& operator<<(std::ostream& os, const ostream_beam_view & obv) {
+    os << "p(" << obv.beam_view.p << ") eob(" << std::boolalpha << obv.beam_view.eob << ") tokens(";
+    for (size_t i = 0 ; i < obv.beam_view.n_tokens ; ++i) {
+        os << llama_token_to_piece(obv.ctx, obv.beam_view.tokens[i]);
+    }
+    return os << ')';
+}
+
+// Put here anything you want back in beam_search_callback().
+struct beam_search_callback_data {
+    llama_context * ctx;
+    std::vector<llama_token> response;
+};
+
+// In this case, end-of-beam (eob) is equivalent to end-of-sentence (eos) but this need not always be the same.
+// For example, eob can be flagged due to maximum token length, stop words, etc.
+bool is_at_eob(const beam_search_callback_data & callback_data, const llama_token * tokens, const size_t n_tokens) {
+    return n_tokens && tokens[n_tokens-1] == llama_token_eos(callback_data.ctx);
+}
+
+// Function matching type llama_beam_search_callback_fn_t.
+// Custom callback example is called each time the beams lengths increase:
+//  * Show progress by printing ',' following by number of convergent beam tokens if any.
+//  * When all beams converge to a common prefix, they are made available in beams_state.beams[0].
+//    This is also called when the stop condition is met.
+//    Collect tokens into std::vector<llama_token> response which is pointed to by callback_data.
+void beam_search_callback(void * callback_data_ptr, llama_beams_state beams_state) {
+    auto& callback_data = *static_cast<beam_search_callback_data*>(callback_data_ptr);
+    // Mark beams as EOS as needed.
+    for (size_t i = 0 ; i < beams_state.n_beams ; ++i) {
+        llama_beam_view& beam_view = beams_state.beam_views[i];
+        if (!beam_view.eob && is_at_eob(callback_data, beam_view.tokens, beam_view.n_tokens)) {
+            beam_view.eob = true;
+        }
+    }
+    printf(",");  // Show progress
+    if (const size_t n = beams_state.common_prefix_length) {
+        callback_data.response.resize(callback_data.response.size() + n);
+        assert(0u < beams_state.n_beams);
+        const llama_token * tokens = beams_state.beam_views[0].tokens;
+        std::copy(tokens, tokens + n, callback_data.response.end() - n);
+        printf("%lu", n);
+    }
+    fflush(stdout);
+#if 1 // DEBUG: print current beams for this iteration
+    std::cout << "\n\nCurrent beams (last_call=" << beams_state.last_call << "):\n";
+    for (size_t i = 0 ; i < beams_state.n_beams ; ++i) {
+        std::cout << "beams["<<i<<"]: " << ostream_beam_view{callback_data.ctx,beams_state.beam_views[i]} << std::endl;
+    }
+#endif
+}
+
+int main(int argc, char ** argv)
+{
+    gpt_params params;
+    //params.n_gpu_layers = 200;
+
+    //---------------------------------
+    // Print help :
+    //---------------------------------
+
+    if ( argc < 2 || argv[1][0] == '-' )
+    {
+        printf( "Usage: %s MODEL_PATH [BEAM_WIDTH=2] [PROMPT]\n" , argv[0] );
+        return 1 ;
+    }
+
+    //---------------------------------
+    // Load parameters :
+    //---------------------------------
+
+    params.model = argv[1];
+
+    params.n_beams = 2 < argc ? std::stoi(argv[2]) : 2;
+
+    if ( argc > 3 )
+    {
+        params.prompt = argv[3];
+    }
+
+    if ( params.prompt.empty() )
+    {
+        params.prompt = "### Request:\nHow many countries are there?\n\n### Response:\n";
+    }
+
+    //---------------------------------
+    // Init LLM :
+    //---------------------------------
+
+    llama_backend_init(params.numa);
+
+    llama_model * model;
+    llama_context * ctx;
+
+    std::tie(model, ctx) = llama_init_from_gpt_params( params );
+
+    if ( model == NULL )
+    {
+        fprintf( stderr , "%s: error: unable to load model\n" , __func__ );
+        return 1;
+    }
+
+    //---------------------------------
+    // Tokenize the prompt :
+    //---------------------------------
+
+    std::vector<llama_token> tokens_list = llama_tokenize(ctx, params.prompt, true);
+
+    const size_t max_context_size     = llama_n_ctx( ctx );
+    const size_t max_tokens_list_size = max_context_size - 4 ;
+
+    if (tokens_list.size() > max_tokens_list_size)
+    {
+        fprintf( stderr , "%s: error: prompt too long (%lu tokens, max %lu)\n" ,
+             __func__ , tokens_list.size() , max_tokens_list_size );
+        return 1;
+    }
+
+    fprintf( stderr, "\n\n" );
+
+    // Print the tokens from the prompt :
+
+    for( auto id : tokens_list )
+    {
+        std::cout << llama_token_to_piece(ctx, id);
+    }
+    std::cout << std::flush;
+
+    int n_past = llama_get_kv_cache_token_count(ctx);
+    if (llama_eval(ctx, tokens_list.data(), tokens_list.size(), n_past, params.n_threads))
+    {
+        fprintf(stderr, "%s : failed to eval prompt.\n" , __func__ );
+        return 1;
+    }
+    n_past += tokens_list.size();
+
+    beam_search_callback_data callback_data{ctx, {}};
+    size_t const beam_width = static_cast<size_t>(params.n_beams);
+    int const n_predict = 256;
+    llama_beam_search(ctx, beam_search_callback, &callback_data, beam_width, n_past, n_predict, params.n_threads);
+
+    std::cout << "\n\n";
+    for (llama_token const token_id : callback_data.response) {
+        std::cout << llama_token_to_piece(ctx,token_id);
+    }
+    std::cout << std::endl;
+
+    llama_free( ctx );
+    llama_free_model( model );
+
+    llama_backend_free();
+
+    return 0;
+}

examples/convert-llama2c-to-ggml/README.md

@@ -12,18 +12,14 @@ usage: ./convert-llama2c-to-ggml [options]
 
 options:
   -h, --help                       show this help message and exit
-  --copy-vocab-from-model FNAME    model path from which to copy vocab (default 'tokenizer.bin')
+  --copy-vocab-from-model FNAME    path of gguf llama model or llama2.c vocabulary from which to copy vocab (default 'models/7B/ggml-model-f16.gguf')
   --llama2c-model FNAME            [REQUIRED] model path from which to load Karpathy's llama2.c model
   --llama2c-output-model FNAME     model path to save the converted llama2.c model (default ak_llama_model.bin')
 ```
 
 An example command using a model from [karpathy/tinyllamas](https://huggingface.co/karpathy/tinyllamas) is as follows:
 
-`$ ./convert-llama2c-to-ggml --copy-vocab-from-model ../llama2.c/tokenizer.bin --llama2c-model stories42M.bin --llama2c-output-model stories42M.ggmlv3.bin`
+`$ ./convert-llama2c-to-ggml --copy-vocab-from-model llama-2-7b-chat.gguf.q2_K.bin --llama2c-model stories42M.bin --llama2c-output-model stories42M.gguf.bin`
-
-For now the generated model is in the legacy GGJTv3 format, so you need to convert it to gguf manually:
-
-`$ python ./convert-llama-ggmlv3-to-gguf.py --eps 1e-5 --input stories42M.ggmlv3.bin --output stories42M.gguf.bin`
 
 Now you can use the model with a command like:
 

examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp

@@ -10,9 +10,48 @@
 #include <ctime>
 #include <random>
 #include <stdexcept>
+#include <sstream>
 #include <algorithm>
 #include <string>
 
+// GGUF keys & tensor names.
+
+#define KV_GENERAL_ARCHITECTURE          "general.architecture"
+#define KV_GENERAL_NAME                  "general.name"
+
+#define KV_TOKENIZER_MODEL               "tokenizer.ggml.model"
+#define KV_TOKENIZER_LIST                "tokenizer.ggml.tokens"
+#define KV_TOKENIZER_TOKEN_TYPE          "tokenizer.ggml.token_type"
+#define KV_TOKENIZER_SCORES              "tokenizer.ggml.scores"
+#define KV_TOKENIZER_BOS_ID              "tokenizer.ggml.bos_token_id"
+#define KV_TOKENIZER_EOS_ID              "tokenizer.ggml.eos_token_id"
+#define KV_TOKENIZER_UNK_ID              "tokenizer.ggml.unknown_token_id"
+#define KV_TOKENIZER_SEP_ID              "tokenizer.ggml.seperator_token_id"
+#define KV_TOKENIZER_PAD_ID              "tokenizer.ggml.padding_token_id"
+#define KV_TOKENIZER_HF_JSON             "tokenizer.huggingface.json"
+
+#define KV_CONTEXT_LENGTH                "llama.context_length"
+#define KV_EMBEDDING_LENGTH              "llama.embedding_length"
+#define KV_BLOCK_COUNT                   "llama.block_count"
+#define KV_FEED_FORWARD_LENGTH           "llama.feed_forward_length"
+#define KV_ATTENTION_HEAD_COUNT          "llama.attention.head_count"
+#define KV_ATTENTION_HEAD_COUNT_KV       "llama.attention.head_count_kv"
+#define KV_ATTENTION_LAYERNORM_RMS_EPS   "llama.attention.layer_norm_rms_epsilon"
+#define KV_ROPE_DIMENSION_COUNT          "llama.rope.dimension_count"
+
+#define TN_TOKEN_EMBD  "token_embd.weight"
+#define TN_OUTPUT_NORM "output_norm.weight"
+#define TN_OUTPUT      "output.weight"
+#define TN_ATTN_NORM   "blk.%d.attn_norm.weight"
+#define TN_ATTN_Q      "blk.%d.attn_q.weight"
+#define TN_ATTN_K      "blk.%d.attn_k.weight"
+#define TN_ATTN_V      "blk.%d.attn_v.weight"
+#define TN_ATTN_OUTPUT "blk.%d.attn_output.weight"
+#define TN_FFN_NORM    "blk.%d.ffn_norm.weight"
+#define TN_FFN_GATE    "blk.%d.ffn_gate.weight"
+#define TN_FFN_DOWN    "blk.%d.ffn_down.weight"
+#define TN_FFN_UP      "blk.%d.ffn_up.weight"
+
 #if defined(_MSC_VER)
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
@@ -20,6 +59,11 @@
 #define LLAMA_FILE_MAGIC_GGJT        0x67676a74u // 'ggjt'
 #define LLAMA_FILE_VERSION_GGJT_V3   3
 
+#define TOKENIZER_NAME "llama"
+#define UNKNOWN_TOKEN_ID 0
+#define BOS_TOKEN_ID 1
+#define EOS_TOKEN_ID 2
+
 //////////////////////////////////////// llama2.c model structs and functions to load models, alloc memory etc.
 typedef struct {
     int dim; // transformer dimension
@@ -183,6 +227,7 @@ struct my_llama_hparams {
     uint32_t n_vocab = 32000;
     uint32_t n_ctx   = 512;   // this is provided as user input?
     uint32_t n_embd  = 4096;
+    uint32_t n_ff    = 11008;
     uint32_t n_mult  = 4;
     uint32_t n_head  = 32;
     uint32_t n_layer = 32;
@@ -214,6 +259,8 @@ struct my_llama_layer {
 struct my_llama_model {
     struct ggml_context * ctx = NULL;
 
+    std::string name;
+
     my_llama_hparams hparams;
 
     struct ggml_tensor * tok_embeddings;
@@ -276,18 +323,13 @@ struct train_params {
     int mem_compute1_gb;
 };
 
-uint32_t get_n_ff(const struct my_llama_hparams* hparams) {
-    const uint32_t n_ff = ((2*(4*hparams->n_embd)/3 + hparams->n_mult - 1)/hparams->n_mult)*hparams->n_mult;
-    return n_ff;
-}
-
 void print_params(struct my_llama_hparams * params) {
     printf("%s: n_vocab: %d\n", __func__, params->n_vocab);
     printf("%s: n_ctx:   %d\n", __func__, params->n_ctx);
     printf("%s: n_embd:  %d\n", __func__, params->n_embd);
     printf("%s: n_mult:  %d\n", __func__, params->n_mult);
     printf("%s: n_head:  %d\n", __func__, params->n_head);
-    printf("%s: n_ff:    %d\n", __func__, get_n_ff(params));
+    printf("%s: n_ff:    %d\n", __func__, params->n_ff);
     printf("%s: n_layer: %d\n", __func__, params->n_layer);
     printf("%s: n_rot:   %d\n", __func__, params->n_rot);
 }
@@ -299,7 +341,7 @@ void init_model(struct my_llama_model * model) {
     const uint32_t n_layer = hparams.n_layer;
     const uint32_t n_vocab = hparams.n_vocab;
 
-    const uint32_t n_ff = get_n_ff(&hparams);
+    const uint32_t n_ff = hparams.n_ff;
     struct ggml_context * ctx = model->ctx;
 
     model->train_its = 0;
@@ -481,21 +523,6 @@ struct llama_file {
         return std::string(chars.data(), len);
     }
 
-    void write_raw(const void * ptr, size_t size) {
-        if (size == 0) {
-            return;
-        }
-        errno = 0;
-        size_t ret = std::fwrite(ptr, size, 1, fp);
-        if (ret != 1) {
-            throw std::runtime_error(format("write error: %s", strerror(errno)));
-        }
-    }
-
-    void write_u32(std::uint32_t val) {
-        write_raw(&val, sizeof(val));
-    }
-
     ~llama_file() {
         if (fp) {
             std::fclose(fp);
@@ -503,30 +530,6 @@ struct llama_file {
     }
 };
 
-void write_tensor(struct llama_file * file, struct ggml_tensor * tensor) {
-    if (tensor == NULL) {
-        file->write_u32(0);
-        file->write_u32(0);
-        file->write_u32(GGML_TYPE_F32);
-        file->seek((0-file->tell()) & 31, SEEK_CUR);
-        return;
-    }
-    const char * name = ggml_get_name(tensor);
-    uint32_t name_len = strlen(name);
-    uint32_t nd = tensor->n_dims;
-    uint32_t ne[4] = { (uint32_t)tensor->ne[0],
-                       (uint32_t)tensor->ne[1],
-                       (uint32_t)tensor->ne[2],
-                       (uint32_t)tensor->ne[3] };
-    file->write_u32(nd);
-    file->write_u32(name_len);
-    file->write_u32(tensor->type);
-    file->write_raw(ne, sizeof(ne[0]) * nd);
-    file->write_raw(name, name_len);
-    file->seek((0-file->tell()) & 31, SEEK_CUR);
-    file->write_raw(tensor->data, ggml_nbytes(tensor));
-}
-
 bool is_ggml_file(const char *filename) {
     llama_file file(filename, "rb");
     if (file.size < 4) {
@@ -536,48 +539,96 @@ bool is_ggml_file(const char *filename) {
     return magic == GGUF_MAGIC;
 }
 
+static std::string llama_escape_whitespaces(const std::string& text) {
+    std::ostringstream out;
+    for (char c : text) {
+        if (c == ' ') out << "\xe2\x96\x81";
+        else out << c;
+    }
+    return out.str();
+}
+
 void load_vocab(const char *filename, Config *config, struct llama_vocab *vocab) {
-#pragma message("TODO: implement reading vocabulary using gguf")
+    if (is_ggml_file(filename)) {
-//    // heuristic to infer whether vocab is from ggml or from llama2.c vocabulary
+        struct ggml_context * ctx_data = NULL;
-//    if (is_ggml_file(filename)) {
+
-//
+        struct gguf_init_params params = {
-//        struct llama_context_params llama_params = llama_context_default_params();
+            /*.no_alloc = */ false,
-//        llama_params.vocab_only = true;
+            /*.ctx      = */ &ctx_data,
-//
+        };
-//        struct llama_model * lmodel = llama_load_model_from_file(filename, llama_params);
+
-//        struct llama_context * lctx = llama_new_context_with_model(lmodel, llama_params);
+        struct gguf_context * ctx = gguf_init_from_file(filename, params);
-//
+        GGML_ASSERT(ctx != NULL);
-//        const int n_vocab = llama_n_vocab(lctx);
+
-//        vocab->id_to_token.resize(n_vocab);
+        const int model_idx = gguf_find_key(ctx, KV_TOKENIZER_MODEL);
-//        for (int i=0; i<n_vocab; ++i) {
+        GGML_ASSERT(model_idx >= 0);
-//            vocab->id_to_token[i].text  = llama_token_get_text(lctx, i);
+        std::string tokenizer_name = gguf_get_val_str(ctx, model_idx);
-//            vocab->id_to_token[i].score = llama_token_get_score(lctx, i);
+        GGML_ASSERT(tokenizer_name == TOKENIZER_NAME);
-//            vocab->id_to_token[i].type  = llama_token_get_type(lctx, i);
+
-//            vocab->token_to_id.emplace(vocab->id_to_token[i].text, i);
+        const int token_idx = gguf_find_key(ctx, KV_TOKENIZER_LIST);
-//        }
+        GGML_ASSERT(token_idx >= 0);
-//        llama_free(lctx);
+
-//        llama_free_model(lmodel);
+        const int score_idx = gguf_find_key(ctx, KV_TOKENIZER_SCORES);
-//    } else
+        GGML_ASSERT(score_idx >= 0);
-    { // assume llama2.c vocabulary
+        const float * scores = (const float * ) gguf_get_arr_data(ctx, score_idx);
-        printf("Assuming llama2.c vocabulary since %s is not a ggml file\n", filename);
+
+        const int toktype_idx = gguf_find_key(ctx, KV_TOKENIZER_TOKEN_TYPE);
+        GGML_ASSERT(toktype_idx >= 0);
+        const int * toktypes = (const int * ) gguf_get_arr_data(ctx, toktype_idx);
+
+        const uint32_t n_vocab = gguf_get_arr_n(ctx, token_idx);
+
+        vocab->id_to_token.resize(n_vocab);
+
+        for (uint32_t i = 0; i < n_vocab; i++) {
+            std::string word = gguf_get_arr_str(ctx, token_idx, i);
+
+            vocab->token_to_id[word] = i;
+
+            auto & token_data = vocab->id_to_token[i];
+            token_data.text  = std::move(word);
+            token_data.score = scores[i];
+            token_data.type  = (llama_token_type) toktypes[i];
+        }
+        ggml_free(ctx_data);
+        gguf_free(ctx);
+    } else {
+        // assume llama2.c vocabulary
+        printf("Assuming llama2.c vocabulary since %s is not a gguf file\n", filename);
         llama_file file(filename, "rb");
         const int  n_vocab = config->vocab_size;
         /* uint32_t max_token_length =  */ file.read_u32(); // unused
         vocab->id_to_token.resize(n_vocab);
-        for (int i=0; i<n_vocab; ++i) {
+        for (llama_vocab::id id=0; id<n_vocab; ++id) {
             float_t score = file.read_f32();
             uint32_t len = file.read_u32();
             std::string text = file.read_string(len);
-            // Special-case handling of <0xXX> single byte tokens.
+
-            char byte_val;
+            unsigned char byte_val;
-            if (sscanf(text.c_str(), "<0x%02hhX>", &byte_val) == 1) {
+            llama_vocab::ttype type = LLAMA_TOKEN_TYPE_NORMAL;
-                char cstr[2] = { byte_val, 0 };
+            if (id == UNKNOWN_TOKEN_ID) {
-                text = cstr;
+                text = "<unk>";
+                type = LLAMA_TOKEN_TYPE_UNKNOWN;
+            } else if (id == BOS_TOKEN_ID) {
+                text = "<s>";
+                type = LLAMA_TOKEN_TYPE_CONTROL;
+            } else if (id == EOS_TOKEN_ID) {
+                text = "</s>";
+                type = LLAMA_TOKEN_TYPE_CONTROL;
+            } else if (text.empty()) {
+                type = LLAMA_TOKEN_TYPE_CONTROL;
+            } else if (sscanf(text.c_str(), "<0x%02hhX>", &byte_val) == 1) {
+                // Text of byte tokens is already in the expected format.
+                type = LLAMA_TOKEN_TYPE_BYTE;
+            } else {
+                type = LLAMA_TOKEN_TYPE_NORMAL;
             }
-            vocab->id_to_token[i].text = text;
+            text = llama_escape_whitespaces(text);
-            vocab->id_to_token[i].score = score;
+
-            vocab->id_to_token[i].type = LLAMA_TOKEN_TYPE_UNDEFINED;
+            vocab->id_to_token[id].text = text;
-            vocab->token_to_id.emplace(text, i);
+            vocab->id_to_token[id].score = score;
+            vocab->id_to_token[id].type = type;
+            vocab->token_to_id.emplace(text, id);
         }
     }
 }
@@ -619,33 +670,6 @@ void stuff_karpathy_weights_into_gg(struct ggml_tensor * gg_weights, float * kar
 }
 
 void save_as_llama_model(struct llama_vocab * vocab, struct my_llama_model * model, TransformerWeights* w, const char * filename) {
-    struct llama_file file(filename, "wb");
-    if (file.fp == NULL) {
-        return;
-    }
-
-#pragma message("TODO: implement file saving using gguf")
-    // write_magic
-    file.write_u32(LLAMA_FILE_MAGIC_GGJT);   // magic
-    file.write_u32(LLAMA_FILE_VERSION_GGJT_V3); // version
-    // write_hparams
-    file.write_u32(model->hparams.n_vocab);
-    file.write_u32(model->hparams.n_embd);
-    file.write_u32(model->hparams.n_mult);
-    file.write_u32(model->hparams.n_head);
-    file.write_u32(model->hparams.n_layer);
-    file.write_u32(model->hparams.n_rot);
-    file.write_u32(LLAMA_FTYPE_ALL_F32);
-
-    // write_vocab - for now we are just writing the existing BPE voc. assuming karpathy's vocabulary is the same. idk.
-    uint32_t n_vocab = model->hparams.n_vocab;
-    for (uint32_t i = 0; i < n_vocab; i++) {
-        const auto & token_data = vocab->id_to_token.at(i);
-        file.write_u32((uint32_t) token_data.text.size());
-        file.write_raw(token_data.text.data(), token_data.text.size());
-        file.write_raw(&token_data.score, sizeof(token_data.score));
-    }
-
     // stuff AK weights into GG weights one by one.
     // w->token_embedding_table -> model->tok_embeddings
     // float*                   -> struct ggml_tensor
@@ -658,8 +682,7 @@ void save_as_llama_model(struct llama_vocab * vocab, struct my_llama_model * mod
     // for rms-att-weight
     int row_length = model->hparams.n_embd;
     const auto & hparams = model->hparams;
-    //int n_ff = model->hparams.n_embd;
+    int n_ff = model->hparams.n_ff;
-    int n_ff = get_n_ff(&hparams);
 
     for (uint32_t i = 0; i < model->hparams.n_layer; ++i){
         auto & layer = model->layers[i];
@@ -677,28 +700,91 @@ void save_as_llama_model(struct llama_vocab * vocab, struct my_llama_model * mod
         stuff_karpathy_weights_into_gg(layer.w2            , &w->w2[i*n_ff*row_length]);
         stuff_karpathy_weights_into_gg(layer.w3            , &w->w3[i*row_length*n_ff]);
     }
+
+    struct gguf_context * ctx = gguf_init_empty();
+
+    std::vector<const char*> tokens;
+    std::vector<float> scores;
+    std::vector<llama_token_type> token_types;
+    for (const llama_vocab::token_data & token_data : vocab->id_to_token) {
+        tokens.push_back(token_data.text.c_str());
+        scores.push_back(token_data.score);
+        token_types.push_back(token_data.type);
+    }
+    gguf_set_arr_str(ctx, KV_TOKENIZER_LIST, tokens.data(), tokens.size());
+    gguf_set_arr_data(ctx, KV_TOKENIZER_SCORES, GGUF_TYPE_FLOAT32, scores.data(), scores.size());
+    gguf_set_arr_data(ctx, KV_TOKENIZER_TOKEN_TYPE, GGUF_TYPE_INT32, token_types.data(), token_types.size());
+
+    gguf_set_val_str(ctx, KV_TOKENIZER_MODEL, TOKENIZER_NAME);
+
+    gguf_set_val_str(ctx, KV_GENERAL_ARCHITECTURE, "llama");
+    gguf_set_val_str(ctx, KV_GENERAL_NAME, "llama");
+
+    // special tokens
+    gguf_set_val_u32(ctx, KV_TOKENIZER_UNK_ID, UNKNOWN_TOKEN_ID);
+    gguf_set_val_u32(ctx, KV_TOKENIZER_BOS_ID, BOS_TOKEN_ID);
+    gguf_set_val_u32(ctx, KV_TOKENIZER_EOS_ID, EOS_TOKEN_ID);
+    gguf_set_val_u32(ctx, KV_TOKENIZER_SEP_ID, -1);
+    gguf_set_val_u32(ctx, KV_TOKENIZER_PAD_ID, -1);
+
+    gguf_set_val_u32(ctx, KV_CONTEXT_LENGTH, model->hparams.n_ctx);
+    gguf_set_val_u32(ctx, KV_EMBEDDING_LENGTH, model->hparams.n_embd);
+    gguf_set_val_u32(ctx, KV_FEED_FORWARD_LENGTH, model->hparams.n_ff);
+    gguf_set_val_u32(ctx, KV_ATTENTION_HEAD_COUNT, model->hparams.n_head);
+    // n_head_kv is optional, default to n_head
+    // gguf_set_val_u32(ctx, KV_ATTENTION_HEAD_COUNT_KV, ...);
+    gguf_set_val_u32(ctx, KV_BLOCK_COUNT, model->hparams.n_layer);
+    gguf_set_val_u32(ctx, KV_ROPE_DIMENSION_COUNT, model->hparams.n_rot);
+    gguf_set_val_f32(ctx, KV_ATTENTION_LAYERNORM_RMS_EPS, 1e-5f);
+
     // write tensors
-    write_tensor(&file, model->tok_embeddings);
+    ggml_set_name(model->tok_embeddings, TN_TOKEN_EMBD);
-    write_tensor(&file, model->norm);
+    gguf_add_tensor(ctx, model->tok_embeddings);
-    write_tensor(&file, model->output); // ?
+
+    ggml_set_name(model->norm, TN_OUTPUT_NORM);
+    gguf_add_tensor(ctx, model->norm);
+
+    ggml_set_name(model->output, TN_OUTPUT);
+    gguf_add_tensor(ctx, model->output);
+
     for (uint32_t i = 0; i < model->hparams.n_layer; ++i) {
         auto & layer = model->layers[i];
 
-        write_tensor(&file, layer.attention_norm);
+        ggml_format_name(layer.wq, TN_ATTN_Q, i);
-        write_tensor(&file, layer.wq);
+        gguf_add_tensor(ctx, layer.wq);
-        write_tensor(&file, layer.wk);
+
-        write_tensor(&file, layer.wv);
+        ggml_format_name(layer.wk, TN_ATTN_K, i);
-        write_tensor(&file, layer.wo);
+        gguf_add_tensor(ctx, layer.wk);
-        write_tensor(&file, layer.ffn_norm);
+
-        write_tensor(&file, layer.w1);
+        ggml_format_name(layer.wv, TN_ATTN_V, i);
-        write_tensor(&file, layer.w2);
+        gguf_add_tensor(ctx, layer.wv);
-        write_tensor(&file, layer.w3);
+
+        ggml_format_name(layer.wo, TN_ATTN_OUTPUT, i);
+        gguf_add_tensor(ctx, layer.wo);
+
+        ggml_format_name(layer.attention_norm, TN_ATTN_NORM, i);
+        gguf_add_tensor(ctx, layer.attention_norm);
+
+        ggml_format_name(layer.w1, TN_FFN_GATE, i);
+        gguf_add_tensor(ctx, layer.w1);
+
+        ggml_format_name(layer.w2, TN_FFN_DOWN, i);
+        gguf_add_tensor(ctx, layer.w2);
+
+        ggml_format_name(layer.w3, TN_FFN_UP, i);
+        gguf_add_tensor(ctx, layer.w3);
+
+        ggml_format_name(layer.ffn_norm, TN_FFN_NORM, i);
+        gguf_add_tensor(ctx, layer.ffn_norm);
     }
+
+    gguf_write_to_file(ctx, filename, false);
+    gguf_free(ctx);
 }
 
 struct train_params get_default_train_params() {
     struct train_params params;
-    params.fn_vocab_model    = "tokenizer.bin";
+    params.fn_vocab_model    = "models/7B/ggml-model-f16.gguf";
     params.fn_llama2c_output_model = "ak_llama_model.bin";
     params.fn_train_data     = "shakespeare.txt";
     params.fn_checkpoint_in  = "checkpoint.bin";
@@ -751,7 +837,7 @@ void print_usage(int /*argc*/, char ** argv, const struct train_params * params)
     fprintf(stderr, "\n");
     fprintf(stderr, "options:\n");
     fprintf(stderr, "  -h, --help                       show this help message and exit\n");
-    fprintf(stderr, "  --copy-vocab-from-model FNAME    llama2.c vocabulary or ggmlv3 model path from which to copy vocab (default '%s')\n", params->fn_vocab_model);
+    fprintf(stderr, "  --copy-vocab-from-model FNAME    path of gguf llama model or llama2.c vocabulary from which to copy vocab (default '%s')\n", params->fn_vocab_model);
     fprintf(stderr, "  --llama2c-model FNAME            [REQUIRED] model path from which to load Karpathy's llama2.c model\n");
     fprintf(stderr, "  --llama2c-output-model FNAME     model path to save the converted llama2.c model (default %s')\n", params->fn_llama2c_output_model);
     fprintf(stderr, "\n");
@@ -812,6 +898,14 @@ bool params_parse(int argc, char ** argv, struct train_params * params) {
     return true;
 }
 
+std::string basename(const std::string &path) {
+    size_t pos = path.find_last_of("/");
+    if (pos == std::string::npos) {
+        return path;
+    }
+    return path.substr(pos + 1);
+}
+
 int main(int argc, char ** argv) {
     struct train_params params = get_default_train_params();
     if (!params_parse(argc, argv, &params)) {
@@ -840,6 +934,7 @@ int main(int argc, char ** argv) {
     model.hparams.n_vocab = config.vocab_size; //llama_n_vocab(lctx);
     model.hparams.n_ctx   = params.n_ctx;
     model.hparams.n_embd  = config.dim; //params.n_embd;
+    model.hparams.n_ff    = config.hidden_dim;
     model.hparams.n_mult  = 32;//params.n_mult;
     model.hparams.n_head  = config.n_heads; //params.n_head;
     model.hparams.n_layer = config.n_layers; //params.n_layer;
@@ -853,6 +948,7 @@ int main(int argc, char ** argv) {
     model.ctx = ggml_init(lcparams);
 
     init_model(&model);
+    model.name = basename(params.fn_llama2c_model);
     save_as_llama_model(&vocab, &model, &weights, params.fn_llama2c_output_model);
 
     printf("Saving llama.c model file %s in ggml format at %s\n", params.fn_llama2c_model, params.fn_llama2c_output_model);

examples/embd-input/embd-input-lib.cpp

@@ -214,7 +214,7 @@ const char * sampling(struct MyModel * mymodel) {
     if (id == llama_token_eos(ctx)) {
         ret = "</s>";
     } else {
-        ret = llama_token_to_str(ctx, id);
+        ret = llama_token_to_piece(ctx, id);
     }
     eval_id(mymodel, id);
     return ret.c_str();

examples/embedding/embedding.cpp

@@ -56,9 +56,6 @@ int main(int argc, char ** argv) {
 
     int n_past = 0;
 
-    // Add a space in front of the first character to match OG llama tokenizer behavior
-    params.prompt.insert(0, 1, ' ');
-
     // tokenize the prompt
     auto embd_inp = ::llama_tokenize(ctx, params.prompt, true);
 
@@ -67,7 +64,7 @@ int main(int argc, char ** argv) {
         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());
         for (int i = 0; i < (int) embd_inp.size(); i++) {
-            fprintf(stderr, "%6d -> '%s'\n", embd_inp[i], llama_token_to_str(ctx, embd_inp[i]).c_str());
+            fprintf(stderr, "%6d -> '%s'\n", embd_inp[i], llama_token_to_piece(ctx, embd_inp[i]).c_str());
         }
         fprintf(stderr, "\n");
     }

examples/gguf/gguf.cpp

@@ -30,6 +30,9 @@ bool gguf_ex_write(const std::string & fname) {
     gguf_set_val_u32 (ctx, "some.parameter.uint32",   0x12345678);
     gguf_set_val_i32 (ctx, "some.parameter.int32",   -0x12345679);
     gguf_set_val_f32 (ctx, "some.parameter.float32",  0.123456789f);
+    gguf_set_val_u64 (ctx, "some.parameter.uint64",   0x123456789abcdef0ull);
+    gguf_set_val_i64 (ctx, "some.parameter.int64",   -0x123456789abcdef1ll);
+    gguf_set_val_f64 (ctx, "some.parameter.float64",  0.1234567890123456789);
     gguf_set_val_bool(ctx, "some.parameter.bool",     true);
     gguf_set_val_str (ctx, "some.parameter.string",   "hello world");
 

examples/llama-bench/llama-bench.cpp

@@ -441,6 +441,8 @@ struct test {
     static const std::string gpu_info;
     std::string model_filename;
     std::string model_type;
+    uint64_t model_size;
+    uint64_t model_n_params;
     int n_batch;
     int n_threads;
     bool f32_kv;
@@ -457,8 +459,10 @@ struct test {
     test(const cmd_params_instance & inst, const llama_model * lmodel, const llama_context * ctx) {
         model_filename = inst.model;
         char buf[128];
-        llama_model_type(lmodel, buf, sizeof(buf));
+        llama_model_desc(lmodel, buf, sizeof(buf));
         model_type = buf;
+        model_size = llama_model_size(lmodel);
+        model_n_params = llama_model_n_params(lmodel);
         n_batch = inst.n_batch;
         n_threads = inst.n_threads;
         f32_kv = inst.f32_kv;
@@ -524,7 +528,7 @@ struct test {
             "build_commit", "build_number",
             "cuda", "opencl", "metal", "gpu_blas", "blas",
             "cpu_info", "gpu_info",
-            "model_filename", "model_type",
+            "model_filename", "model_type", "model_size", "model_n_params",
             "n_batch", "n_threads", "f16_kv",
             "n_gpu_layers", "main_gpu", "mul_mat_q", "low_vram", "tensor_split",
             "n_prompt", "n_gen", "test_time",
@@ -538,6 +542,7 @@ struct test {
 
     static field_type get_field_type(const std::string & field) {
         if (field == "build_number" || field == "n_batch" || field == "n_threads" ||
+            field == "model_size" || field == "model_n_params" ||
             field == "n_gpu_layers" || field == "main_gpu" ||
             field == "n_prompt" || field == "n_gen" ||
             field == "avg_ns" || field == "stddev_ns") {
@@ -573,7 +578,7 @@ struct test {
             build_commit, std::to_string(build_number),
             std::to_string(cuda), std::to_string(opencl), std::to_string(metal), std::to_string(gpu_blas), std::to_string(blas),
             cpu_info, gpu_info,
-            model_filename, model_type,
+            model_filename, model_type, std::to_string(model_size), std::to_string(model_n_params),
             std::to_string(n_batch), std::to_string(n_threads), std::to_string(!f32_kv),
             std::to_string(n_gpu_layers), std::to_string(main_gpu), std::to_string(mul_mat_q), std::to_string(low_vram), tensor_split_str,
             std::to_string(n_prompt), std::to_string(n_gen), test_time,
@@ -709,8 +714,15 @@ struct markdown_printer : public printer {
             return -30;
         }
         if (field == "t/s") {
-            return 15;
+            return 16;
         }
+        if (field == "size" || field == "params") {
+            return 10;
+        }
+        if (field == "n_gpu_layers") {
+            return 3;
+        }
+
         int width = std::max((int)field.length(), 10);
 
         if (test::get_field_type(field) == test::STRING) {
@@ -719,9 +731,28 @@ struct markdown_printer : public printer {
         return width;
     }
 
+    static std::string get_field_display_name(const std::string & field) {
+        if (field == "n_gpu_layers") {
+            return "ngl";
+        }
+        if (field == "n_threads") {
+            return "threads";
+        }
+        if (field == "mul_mat_q") {
+            return "mmq";
+        }
+        if (field == "tensor_split") {
+            return "ts";
+        }
+        return field;
+    }
+
     void print_header(const cmd_params & params) override {
         // select fields to print
-        fields = { "model", "backend" };
+        fields.push_back("model");
+        fields.push_back("size");
+        fields.push_back("params");
+        fields.push_back("backend");
         bool is_cpu_backend = test::get_backend() == "CPU" || test::get_backend() == "BLAS";
         if (!is_cpu_backend) {
             fields.push_back("n_gpu_layers");
@@ -752,7 +783,7 @@ struct markdown_printer : public printer {
 
         fprintf(fout, "|");
         for (const auto & field : fields) {
-            fprintf(fout, " %*s |", get_field_width(field), field.c_str());
+            fprintf(fout, " %*s |", get_field_width(field), get_field_display_name(field).c_str());
         }
         fprintf(fout, "\n");
         fprintf(fout, "|");
@@ -769,12 +800,26 @@ struct markdown_printer : public printer {
         fprintf(fout, "|");
         for (const auto & field : fields) {
             std::string value;
+            char buf[128];
             if (field == "model") {
                 value = t.model_type;
+            } else if (field == "size") {
+                if (t.model_size < 1024*1024*1024) {
+                    snprintf(buf, sizeof(buf), "%.2f MiB", t.model_size / 1024.0 / 1024.0);
+                } else {
+                    snprintf(buf, sizeof(buf), "%.2f GiB", t.model_size / 1024.0 / 1024.0 / 1024.0);
+                }
+                value = buf;
+            } else if (field == "params") {
+                if (t.model_n_params < 1000*1000*1000) {
+                    snprintf(buf, sizeof(buf), "%.2f M", t.model_n_params / 1e6);
+                } else {
+                    snprintf(buf, sizeof(buf), "%.2f B", t.model_n_params / 1e9);
+                }
+                value = buf;
             } else if (field == "backend") {
                 value = test::get_backend();
             } else if (field == "test") {
-                char buf[128];
                 if (t.n_prompt > 0 && t.n_gen == 0) {
                     snprintf(buf, sizeof(buf), "pp %d", t.n_prompt);
                 } else if (t.n_gen > 0 && t.n_prompt == 0) {
@@ -785,7 +830,6 @@ struct markdown_printer : public printer {
                 }
                 value = buf;
             } else if (field == "t/s") {
-                char buf[128];
                 snprintf(buf, sizeof(buf), "%.2f ± %.2f", t.avg_ts(), t.stdev_ts());
                 value = buf;
             } else if (vmap.find(field) != vmap.end()) {

examples/main/main.cpp

@@ -189,12 +189,14 @@ int main(int argc, char ** argv) {
         }
     }
 
-    const bool is_spm = llama_vocab_type(ctx) == LLAMA_VOCAB_TYPE_SPM;
+    // Add BOS if SPM tokenizer
+    const bool add_bos = llama_vocab_type(ctx) == LLAMA_VOCAB_TYPE_SPM;
 
     // tokenize the prompt
     std::vector<llama_token> embd_inp;
+
     if (params.interactive_first || params.instruct || !params.prompt.empty() || session_tokens.empty()) {
-        embd_inp = ::llama_tokenize(ctx, params.prompt, is_spm);
+        embd_inp = ::llama_tokenize(ctx, params.prompt, add_bos);
     } else {
         embd_inp = session_tokens;
     }
@@ -209,10 +211,9 @@ int main(int argc, char ** argv) {
     int guidance_offset = 0;
     int original_prompt_len = 0;
     if (ctx_guidance) {
-        params.cfg_negative_prompt.insert(0, 1, ' ');
+        guidance_inp = ::llama_tokenize(ctx_guidance, params.cfg_negative_prompt, add_bos);
-        guidance_inp = ::llama_tokenize(ctx_guidance, params.cfg_negative_prompt, is_spm);
 
-        std::vector<llama_token> original_inp = ::llama_tokenize(ctx, params.prompt, is_spm);
+        std::vector<llama_token> original_inp = ::llama_tokenize(ctx, params.prompt, add_bos);
         original_prompt_len = original_inp.size();
         guidance_offset = (int)guidance_inp.size() - original_prompt_len;
     }
@@ -259,7 +260,7 @@ int main(int argc, char ** argv) {
     }
 
     // prefix & suffix for instruct mode
-    const auto inp_pfx = ::llama_tokenize(ctx, "\n\n### Instruction:\n\n", is_spm);
+    const auto inp_pfx = ::llama_tokenize(ctx, "\n\n### Instruction:\n\n", add_bos);
     const auto inp_sfx = ::llama_tokenize(ctx, "\n\n### Response:\n\n",    false);
 
     // in instruct mode, we inject a prefix and a suffix to each input by the user
@@ -278,7 +279,7 @@ int main(int argc, char ** argv) {
         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());
         for (int i = 0; i < (int) embd_inp.size(); i++) {
-            fprintf(stderr, "%6d -> '%s'\n", embd_inp[i], llama_token_to_str(ctx, embd_inp[i]).c_str());
+            fprintf(stderr, "%6d -> '%s'\n", embd_inp[i], llama_token_to_piece(ctx, embd_inp[i]).c_str());
         }
 
         if (ctx_guidance) {
@@ -286,14 +287,14 @@ int main(int argc, char ** argv) {
             fprintf(stderr, "%s: negative prompt: '%s'\n", __func__, params.cfg_negative_prompt.c_str());
             fprintf(stderr, "%s: number of tokens in negative prompt = %zu\n", __func__, guidance_inp.size());
             for (int i = 0; i < (int) guidance_inp.size(); i++) {
-                fprintf(stderr, "%6d -> '%s'\n", guidance_inp[i], llama_token_to_str(ctx, guidance_inp[i]).c_str());
+                fprintf(stderr, "%6d -> '%s'\n", guidance_inp[i], llama_token_to_piece(ctx, guidance_inp[i]).c_str());
             }
         }
 
         if (params.n_keep > 0) {
         fprintf(stderr, "%s: static prompt based on n_keep: '", __func__);
             for (int i = 0; i < params.n_keep; i++) {
-                fprintf(stderr, "%s", llama_token_to_str(ctx, embd_inp[i]).c_str());
+                fprintf(stderr, "%s", llama_token_to_piece(ctx, embd_inp[i]).c_str());
             }
             fprintf(stderr, "'\n");
         }
@@ -449,7 +450,7 @@ int main(int argc, char ** argv) {
                 //printf("\n---\n");
                 //printf("resetting: '");
                 //for (int i = 0; i < (int) embd.size(); i++) {
-                //    printf("%s", llama_token_to_str(ctx, embd[i]));
+                //    printf("%s", llama_token_to_piece(ctx, embd[i]));
                 //}
                 //printf("'\n");
                 //printf("\n---\n");
@@ -502,7 +503,7 @@ int main(int argc, char ** argv) {
                     input_size = embd_guidance.size();
                     //fprintf(stderr, "\n---------------------\n");
                     //for (int i = 0; i < (int) embd_guidance.size(); i++) {
-                        //fprintf(stderr, "%s", llama_token_to_str(ctx, embd_guidance[i]));
+                        //fprintf(stderr, "%s", llama_token_to_piece(ctx, embd_guidance[i]));
                     //}
                     //fprintf(stderr, "\n---------------------\n");
                 } else {
@@ -597,7 +598,12 @@ int main(int argc, char ** argv) {
                     last_n_tokens.data() + last_n_tokens.size() - last_n_repeat,
                     last_n_repeat, alpha_frequency, alpha_presence);
                 if (!penalize_nl) {
-                    logits[llama_token_nl(ctx)] = nl_logit;
+                    for (size_t idx = 0; idx < candidates_p.size; idx++) {
+                        if (candidates_p.data[idx].id == llama_token_nl(ctx)) {
+                            candidates_p.data[idx].logit = nl_logit;
+                            break;
+                        }
+                    }
                 }
 
                 if (grammar != NULL) {
@@ -661,7 +667,7 @@ int main(int argc, char ** argv) {
         // display text
         if (input_echo) {
             for (auto id : embd) {
-                printf("%s", llama_token_to_str(ctx, id).c_str());
+                printf("%s", llama_token_to_piece(ctx, id).c_str());
             }
             fflush(stdout);
         }
@@ -677,7 +683,7 @@ int main(int argc, char ** argv) {
             if (params.antiprompt.size()) {
                 std::string last_output;
                 for (auto id : last_n_tokens) {
-                    last_output += llama_token_to_str(ctx, id);
+                    last_output += llama_token_to_piece(ctx, id);
                 }
 
                 is_antiprompt = false;

examples/perplexity/perplexity.cpp

@@ -6,6 +6,8 @@
 #include <ctime>
 #include <sstream>
 #include <cstring>
+#include <thread>
+#include <mutex>
 
 #if defined(_MSC_VER)
 #pragma warning(disable: 4244 4267) // possible loss of data
@@ -27,6 +29,40 @@ std::vector<float> softmax(const std::vector<float>& logits) {
     return probs;
 }
 
+float log_softmax(int n_vocab, const float * logits, int tok) {
+    float max_logit = logits[0];
+    for (int i = 1; i < n_vocab; ++i) max_logit = std::max(max_logit, logits[i]);
+    double sum_exp = 0.0;
+    for (int i = 0; i < n_vocab; ++i) sum_exp += expf(logits[i] - max_logit);
+    return logits[tok] - max_logit - log(sum_exp);
+}
+
+void process_logits(int n_vocab, const float * logits, const int * tokens, int n_token, std::vector<std::thread>& workers,
+        double& nll, double& nll2) {
+
+    std::mutex mutex;
+    int counter = 0;
+    auto compute = [&mutex, &counter, &nll, &nll2, n_vocab, logits, tokens, n_token] () {
+        double local_nll = 0, local_nll2 = 0;
+        while (true) {
+            std::unique_lock<std::mutex> lock(mutex);
+            int i = counter++;
+            if (i >= n_token) {
+                nll += local_nll; nll2 += local_nll2;
+                break;
+            }
+            lock.unlock();
+            double v = -log_softmax(n_vocab, logits + i*n_vocab, tokens[i+1]);
+            local_nll += v;
+            local_nll2 += v*v;
+        }
+    };
+    for (auto& w : workers) w = std::thread(compute);
+    compute();
+    for (auto& w : workers) w.join();
+
+}
+
 void perplexity_v2(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`
@@ -154,10 +190,14 @@ void perplexity(llama_context * ctx, const gpt_params & params) {
     const bool is_spm = llama_vocab_type(ctx) == LLAMA_VOCAB_TYPE_SPM;
     const bool add_bos = is_spm;
 
+    auto tim1 = std::chrono::high_resolution_clock::now();
     fprintf(stderr, "%s: tokenizing the input ..\n", __func__);
 
     auto tokens = ::llama_tokenize(ctx, params.prompt, add_bos);
 
+    auto tim2 = std::chrono::high_resolution_clock::now();
+    fprintf(stderr, "%s: tokenization took %g ms\n",__func__,1e-3*std::chrono::duration_cast<std::chrono::microseconds>(tim2-tim1).count());
+
     const int n_chunk_max = tokens.size() / params.n_ctx;
 
     const int n_chunk = params.n_chunks < 0 ? n_chunk_max : std::min(params.n_chunks, n_chunk_max);
@@ -166,9 +206,12 @@ void perplexity(llama_context * ctx, const gpt_params & params) {
 
     int count = 0;
     double nll = 0.0;
+    double nll2 = 0.0;
 
     fprintf(stderr, "%s: calculating perplexity over %d chunks, batch_size=%d\n", __func__, n_chunk, n_batch);
 
+    std::vector<std::thread> workers(std::thread::hardware_concurrency() - 1);
+
     for (int i = 0; i < n_chunk; ++i) {
         const int start =     i * params.n_ctx;
         const int end   = start + params.n_ctx;
@@ -228,26 +271,32 @@ void perplexity(llama_context * ctx, const gpt_params & params) {
         // Example, we have a context window of 512, we will compute perplexity for each of the
         // last 256 tokens.  Then, we split the input up into context window size chunks to
         // process the entire prompt.
-        for (int j = std::min(512, params.n_ctx / 2); j < params.n_ctx - 1; ++j) {
+        const int first = std::min(512, params.n_ctx/2);
-            // Calculate probability of next token, given the previous ones.
+        process_logits(n_vocab, logits.data() + first*n_vocab, tokens.data() + start + first, params.n_ctx - 1 - first, workers, nll, nll2);
-            const std::vector<float> tok_logits(
+        count += params.n_ctx - first - 1;
-                logits.begin() + (j + 0) * n_vocab,
-                logits.begin() + (j + 1) * n_vocab);
 
-            const float prob = softmax(tok_logits)[tokens[start + j + 1]];
-
-            nll += -std::log(prob);
-            ++count;
-        }
         // perplexity is e^(average negative log-likelihood)
         if (params.ppl_output_type == 0) {
             printf("[%d]%.4lf,", i + 1, std::exp(nll / count));
         } else {
-            printf("%8d  %.4lf\n", i*params.n_ctx, std::exp(nll / count));
+            double av = nll/count;
+            double av2 = nll2/count - av*av;
+            if (av2 > 0) av2 = sqrt(av2/(count-1));
+            printf("%8d  %.4lf  %4lf  %4lf\n", i*params.n_ctx, std::exp(nll / count), av, av2);
         }
         fflush(stdout);
     }
     printf("\n");
+    nll2 /= count;
+    nll /= count;
+    nll2 -= nll * nll;
+    if (nll2 > 0) {
+        nll2 = sqrt(nll2/(count-1));
+        double ppl = exp(nll);
+        printf("Final estimate: PPL = %.4lf +/- %.5lf\n", ppl, nll2*ppl);
+    } else {
+        printf("Unexpected negative standard deviation of log(prob)\n");
+    }
 }
 
 std::vector<float> hellaswag_evaluate_tokens(llama_context * ctx, const std::vector<int>& tokens, int n_past, int n_batch,
@@ -306,6 +355,7 @@ void hellaswag_score(llama_context * ctx, const gpt_params & params) {
     fprintf(stderr, "%s : loaded %zu tasks from prompt.\n", __func__, hs_task_count);
 
     const bool is_spm = llama_vocab_type(ctx) == LLAMA_VOCAB_TYPE_SPM;
+    fprintf(stderr, "================================= is_spm = %d\n", is_spm);
 
     // This is needed as usual for LLaMA models
     const bool add_bos = is_spm;
@@ -346,7 +396,7 @@ void hellaswag_score(llama_context * ctx, const gpt_params & params) {
         hs_data[i].context = prompt_lines[idx*6];
         hs_data[i].gold_ending_idx = std::stoi( prompt_lines[idx*6+1] );
         for (size_t j=0; j < 4; j++) {
-            hs_data[i].ending[j] = " " + prompt_lines[idx*6+2+j];
+            hs_data[i].ending[j] = prompt_lines[idx*6+2+j];
         }
 
         // Delete the selected random example from the prompt
@@ -361,6 +411,8 @@ void hellaswag_score(llama_context * ctx, const gpt_params & params) {
     double acc = 0.0f;
     const int n_vocab = llama_n_vocab(ctx);
 
+    std::vector<std::vector<int>> ending_tokens(4);
+
     std::vector<float> tok_logits(n_vocab);
 
     for (size_t task_idx = 0; task_idx < hs_task_count; task_idx++) {
@@ -368,11 +420,21 @@ void hellaswag_score(llama_context * ctx, const gpt_params & params) {
         std::vector<int> context_embd = ::llama_tokenize(ctx, hs_data[task_idx].context, add_bos);
         size_t context_size = context_embd.size();
 
+        for (int i = 0; i < 4; ++i) {
+            ending_tokens[i] = ::llama_tokenize(ctx, hs_data[task_idx].context + " " + hs_data[task_idx].ending[i], add_bos);
+            for (int k = 0; k < int(context_size); ++k) {
+                if (ending_tokens[i][k] != context_embd[k]) {
+                    fprintf(stderr, "Oops: ending %d of task %d differs from context at position %d\n",i,int(task_idx),k);
+                    break;
+                }
+            }
+        }
+
         // Do the 1st ending
         // In this case we include the context when evaluating
-        auto query_embd = ::llama_tokenize(ctx, hs_data[task_idx].context + hs_data[task_idx].ending[0], add_bos);
+        //auto query_embd = ::llama_tokenize(ctx, hs_data[task_idx].context + hs_data[task_idx].ending[0], add_bos);
+        auto query_embd = ending_tokens[0];
         auto query_size = query_embd.size();
-        //printf("First query: %d\n",(int)query_size);
 
         // Stop if query wont fit the ctx window
         if (query_size > (size_t)params.n_ctx) {
@@ -417,7 +479,8 @@ void hellaswag_score(llama_context * ctx, const gpt_params & params) {
         for (size_t ending_idx = 1; ending_idx < 4; ending_idx++) {
 
             // Tokenize the query
-            query_embd = ::llama_tokenize(ctx, hs_data[task_idx].ending[ending_idx], false);
+            query_embd.resize(ending_tokens[ending_idx].size() - context_size);
+            std::memcpy(query_embd.data(), ending_tokens[ending_idx].data() + context_size, query_embd.size()*sizeof(int));
             query_size = query_embd.size();
 
             // Stop if query wont fit the ctx window

examples/save-load-state/save-load-state.cpp

@@ -87,7 +87,7 @@ int main(int argc, char ** argv) {
         }
         llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
         auto next_token = llama_sample_token(ctx, &candidates_p);
-        auto next_token_str = llama_token_to_str(ctx, next_token);
+        auto next_token_str = llama_token_to_piece(ctx, next_token);
         last_n_tokens_data.push_back(next_token);
 
         printf("%s", next_token_str.c_str());
@@ -147,7 +147,7 @@ int main(int argc, char ** argv) {
         }
         llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
         auto next_token = llama_sample_token(ctx2, &candidates_p);
-        auto next_token_str = llama_token_to_str(ctx2, next_token);
+        auto next_token_str = llama_token_to_piece(ctx2, next_token);
         last_n_tokens_data.push_back(next_token);
 
         printf("%s", next_token_str.c_str());

examples/server/README.md

@@ -77,34 +77,31 @@ You need to have [Node.js](https://nodejs.org/en) installed.
 ```bash
 mkdir llama-client
 cd llama-client
-npm init
-npm install axios
 ```
 
 Create a index.js file and put inside this:
 
 ```javascript
-const axios = require("axios");
-
 const prompt = `Building a website can be done in 10 simple steps:`;
 
 async function Test() {
-    let result = await axios.post("http://127.0.0.1:8080/completion", {
+    let response = await fetch("http://127.0.0.1:8080/completion", {
-        prompt,
+        method: 'POST',
-        n_predict: 512,
+        body: JSON.stringify({
-    });
+            prompt,
-
+            n_predict: 512,
-    // the response is received until completion finish
+        })
-    console.log(result.data.content);
+    })
+    console.log((await response.json()).content)
 }
 
-Test();
+Test()
 ```
 
 And run it:
 
 ```bash
-node .
+node index.js
 ```
 
 ## API Endpoints
@@ -167,6 +164,12 @@ node .
 
     Note that the special `BOS` token is not added in front of the text and also a space character is not inserted automatically as it is for `/completion`.
 
+-   **POST** `/detokenize`: Convert tokens to text.
+
+    *Options:*
+
+    `tokens`: Set the tokens to detokenize.
+
 -   **POST** `/embedding`: Generate embedding of a given text just as [the embedding example](../embedding) does.
 
     *Options:*

examples/server/server.cpp

@@ -94,7 +94,7 @@ static std::string tokens_to_str(llama_context *ctx, Iter begin, Iter end)
     std::string ret;
     for (; begin != end; ++begin)
     {
-        ret += llama_token_to_str(ctx, *begin);
+        ret += llama_token_to_piece(ctx, *begin);
     }
     return ret;
 }
@@ -123,7 +123,7 @@ static void server_log(const char *level, const char *function, int line,
 // format incomplete utf-8 multibyte character for output
 static std::string tokens_to_output_formatted_string(const llama_context *ctx, const llama_token token)
 {
-    std::string out = token == -1 ? "" : llama_token_to_str(ctx, token);
+    std::string out = token == -1 ? "" : llama_token_to_piece(ctx, token);
     // if the size is 1 and first bit is 1, meaning it's a partial character
     //   (size > 1 meaning it's already a known token)
     if (out.size() == 1 && (out[0] & 0x80) == 0x80)
@@ -286,7 +286,6 @@ struct llama_server_context
                     std::vector<llama_token> p;
                     if (first)
                     {
-                        s.insert(0, 1, ' '); // add a space if it's the first
                         p = ::llama_tokenize(ctx, s, add_bos);
                         first = false;
                     }
@@ -309,7 +308,6 @@ struct llama_server_context
         else
         {
             auto s = json_prompt.template get<std::string>();
-            s.insert(0, 1, ' '); // always add a first space
             prompt_tokens = ::llama_tokenize(ctx, s, add_bos);
         }
 
@@ -566,7 +564,7 @@ struct llama_server_context
 
         if (!embd.empty() && embd.back() == llama_token_eos(ctx))
         {
-            // stopping_word = llama_token_to_str(ctx, embd.back());
+            // stopping_word = llama_token_to_piece(ctx, embd.back());
             has_next_token = false;
             stopped_eos = true;
             LOG_VERBOSE("eos token found", {});
@@ -613,7 +611,7 @@ struct llama_server_context
     {
         const completion_token_output token_with_probs = nextToken();
 
-        const std::string token_text = token_with_probs.tok == -1 ? "" : llama_token_to_str(ctx, token_with_probs.tok);
+        const std::string token_text = token_with_probs.tok == -1 ? "" : llama_token_to_piece(ctx, token_with_probs.tok);
         generated_text += token_text;
 
         if (params.n_probs > 0)
@@ -1104,6 +1102,12 @@ static json format_tokenizer_response(const std::vector<llama_token> &tokens)
         {"tokens", tokens}};
 }
 
+static json format_detokenized_response(std::string content)
+{
+    return json{
+        {"content", content}};
+}
+
 template <typename T>
 static T json_value(const json &body, const std::string &key, const T &default_value)
 {
@@ -1209,6 +1213,62 @@ static void log_server_request(const Request &req, const Response &res)
                            });
 }
 
+bool is_at_eob(llama_server_context & server_context, const llama_token * tokens, const size_t n_tokens) {
+    return n_tokens && tokens[n_tokens-1] == llama_token_eos(server_context.ctx);
+}
+
+// Function matching type llama_beam_search_callback_fn_t.
+// Custom callback example is called each time the beams lengths increase:
+//  * Show progress by printing ',' following by number of convergent beam tokens if any.
+//  * When all beams converge to a common prefix, they are made available in beams_state.beams[0].
+//    This is also called when the stop condition is met.
+//    Collect tokens into std::vector<llama_token> response which is pointed to by callback_data.
+void beam_search_callback(void * callback_data, llama_beams_state beams_state) {
+    auto & llama = *static_cast<llama_server_context*>(callback_data);
+    // Mark beams as EOS as needed.
+    for (size_t i = 0 ; i < beams_state.n_beams ; ++i) {
+        llama_beam_view& beam_view = beams_state.beam_views[i];
+        if (!beam_view.eob && is_at_eob(llama, beam_view.tokens, beam_view.n_tokens)) {
+            beam_view.eob = true;
+        }
+    }
+    printf(",");  // Show progress
+    if (const size_t n = beams_state.common_prefix_length) {
+        llama.generated_token_probs.resize(llama.generated_token_probs.size() + n);
+        assert(0u < beams_state.n_beams);
+        const llama_token * tokens = beams_state.beam_views[0].tokens;
+        const auto map = [](llama_token tok) { return completion_token_output{{},tok}; };
+        std::transform(tokens, tokens + n, llama.generated_token_probs.end() - n, map);
+        printf("%lu", n);
+    }
+    fflush(stdout);
+#if 0 // DEBUG: print current beams for this iteration
+    std::cout << "\n\nCurrent beams:\n";
+    for (size_t i=0 ; i < beams_state.n_beams ; ++i) {
+        std::cout << "beams["<<i<<"]: " << ostream_beam_view{state.ctx,beams_state.beam_views[i]} << std::endl;
+    }
+#endif
+}
+
+struct token_translator {
+    llama_context * ctx;
+    std::string operator()(llama_token tok) const { return llama_token_to_piece(ctx, tok); }
+    std::string operator()(completion_token_output cto) const { return (*this)(cto.tok); }
+};
+
+void append_to_generated_text_from_generated_token_probs(llama_server_context & llama) {
+    auto & gtps = llama.generated_token_probs;
+    auto translator = token_translator{llama.ctx};
+    auto add_strlen = [=](size_t sum, const completion_token_output & cto) { return sum + translator(cto).size(); };
+    const size_t len = std::accumulate(gtps.begin(), gtps.end(), size_t(0), add_strlen);
+    if (llama.generated_text.capacity() < llama.generated_text.size() + len) {
+        llama.generated_text.reserve(llama.generated_text.size() + len);
+    }
+    for (const completion_token_output & cto : gtps) {
+        llama.generated_text += translator(cto);
+    }
+}
+
 int main(int argc, char **argv)
 {
     // own arguments required by this example
@@ -1291,22 +1351,30 @@ int main(int argc, char **argv)
         llama.beginCompletion();
 
         if (!llama.stream) {
-            size_t stop_pos = std::string::npos;
+            if (llama.params.n_beams) {
+                // Fill llama.generated_token_probs vector with final beam.
+                llama_beam_search(llama.ctx, beam_search_callback, &llama, llama.params.n_beams,
+                                  llama.n_past, llama.n_remain, llama.params.n_threads);
+                // Translate llama.generated_token_probs to llama.generated_text.
+                append_to_generated_text_from_generated_token_probs(llama);
+            } else {
+                size_t stop_pos = std::string::npos;
 
-            while (llama.has_next_token) {
+                while (llama.has_next_token) {
-                const completion_token_output token_with_probs = llama.doCompletion();
+                    const completion_token_output token_with_probs = llama.doCompletion();
-                const std::string token_text = token_with_probs.tok == -1 ? "" : llama_token_to_str(llama.ctx, token_with_probs.tok);
+                    const std::string token_text = token_with_probs.tok == -1 ? "" : llama_token_to_piece(llama.ctx, token_with_probs.tok);
 
-                stop_pos = llama.findStoppingStrings(llama.generated_text,
+                    stop_pos = llama.findStoppingStrings(llama.generated_text,
-                    token_text.size(), STOP_FULL);
+                        token_text.size(), STOP_FULL);
-            }
+                }
 
-            if (stop_pos == std::string::npos) {
+                if (stop_pos == std::string::npos) {
-                stop_pos = llama.findStoppingStrings(llama.generated_text, 0, STOP_PARTIAL);
+                    stop_pos = llama.findStoppingStrings(llama.generated_text, 0, STOP_PARTIAL);
-            }
+                }
-            if (stop_pos != std::string::npos) {
+                if (stop_pos != std::string::npos) {
-                llama.generated_text.erase(llama.generated_text.begin() + stop_pos,
+                    llama.generated_text.erase(llama.generated_text.begin() + stop_pos,
-                    llama.generated_text.end());
+                        llama.generated_text.end());
+                }
             }
 
             const json data = format_final_response(llama, llama.generated_text, llama.generated_token_probs);
@@ -1325,7 +1393,7 @@ int main(int argc, char **argv)
                     if (token_with_probs.tok == -1 || llama.multibyte_pending > 0) {
                         continue;
                     }
-                    const std::string token_text = llama_token_to_str(llama.ctx, token_with_probs.tok);
+                    const std::string token_text = llama_token_to_piece(llama.ctx, token_with_probs.tok);
 
                     size_t pos = std::min(sent_count, llama.generated_text.size());
 
@@ -1437,6 +1505,21 @@ int main(int argc, char **argv)
         const json data = format_tokenizer_response(tokens);
         return res.set_content(data.dump(), "application/json"); });
 
+    svr.Post("/detokenize", [&llama](const Request &req, Response &res)
+             {
+        auto lock = llama.lock();
+
+        const json body = json::parse(req.body);
+        std::string content;
+        if (body.count("tokens") != 0)
+        {
+            const std::vector<llama_token> tokens = body["tokens"];
+            content = tokens_to_str(llama.ctx, tokens.cbegin(), tokens.cend());
+        }
+
+        const json data = format_detokenized_response(content);
+        return res.set_content(data.dump(), "application/json"); });
+
     svr.Post("/embedding", [&llama](const Request &req, Response &res)
              {
         auto lock = llama.lock();

examples/simple/simple.cpp

@@ -63,7 +63,7 @@ int main(int argc, char ** argv) {
     fprintf(stderr, "\n\n");
 
     for (auto id : tokens_list) {
-        fprintf(stderr, "%s", llama_token_to_str(ctx, id).c_str());
+        fprintf(stderr, "%s", llama_token_to_piece(ctx, id).c_str());
     }
 
     fflush(stderr);
@@ -112,7 +112,7 @@ int main(int argc, char ** argv) {
         }
 
         // print the new token :
-        printf("%s", llama_token_to_str(ctx, new_token_id).c_str());
+        printf("%s", llama_token_to_piece(ctx, new_token_id).c_str());
         fflush(stdout);
 
         // push this new token for next evaluation

examples/train-text-from-scratch/train-text-from-scratch.cpp

@@ -1964,7 +1964,7 @@ void print_matrix(struct ggml_tensor * probs) {
 
 
 void print_token(struct llama_context * ctx, llama_token token) {
-    printf("%s", llama_token_to_str(ctx, token).c_str());
+    printf("%s", llama_token_to_piece(ctx, token).c_str());
 }
 
 void print_tokens(struct llama_context* ctx, struct ggml_tensor * tokens) {
@@ -2202,7 +2202,7 @@ int tokenize_file(struct llama_context * lctx, const char * filename, std::vecto
         const char * in  = buf.data();
         const char * end = buf.data() + buf.size();
         for (int i = 0; i < (int) out.size(); ++i) {
-            std::string s = llama_token_to_str(lctx, out[i]);
+            std::string s = llama_token_to_piece(lctx, out[i]);
             int len = s.length();
             if (in >= end) {
                 printf("%s: unexpected end of original text.\n", __func__);

flake.lock

@@ -5,11 +5,11 @@
         "systems": "systems"
       },
       "locked": {
-        "lastModified": 1685518550,
+        "lastModified": 1692799911,
-        "narHash": "sha256-o2d0KcvaXzTrPRIo0kOLV0/QXHhDQ5DTi+OxcjO8xqY=",
+        "narHash": "sha256-3eihraek4qL744EvQXsK1Ha6C3CR7nnT8X2qWap4RNk=",
         "owner": "numtide",
         "repo": "flake-utils",
-        "rev": "a1720a10a6cfe8234c0e93907ffe81be440f4cef",
+        "rev": "f9e7cf818399d17d347f847525c5a5a8032e4e44",
         "type": "github"
       },
       "original": {
@@ -20,11 +20,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1685931219,
+        "lastModified": 1692913444,
-        "narHash": "sha256-8EWeOZ6LKQfgAjB/USffUSELPRjw88A+xTcXnOUvO5M=",
+        "narHash": "sha256-1SvMQm2DwofNxXVtNWWtIcTh7GctEVrS/Xel/mdc6iY=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "7409480d5c8584a1a83c422530419efe4afb0d19",
+        "rev": "18324978d632ffc55ef1d928e81630c620f4f447",
         "type": "github"
       },
       "original": {

flake.nix

@@ -6,6 +6,9 @@
   outputs = { self, nixpkgs, flake-utils }:
     flake-utils.lib.eachDefaultSystem (system:
       let
+        name = "llama.cpp";
+        src = ./.;
+        meta.mainProgram = "llama";
         inherit (pkgs.stdenv) isAarch32 isAarch64 isDarwin;
         buildInputs = with pkgs; [ openmpi ];
         osSpecific = with pkgs; buildInputs ++
@@ -21,11 +24,17 @@
               CoreGraphics
               CoreVideo
             ]
+          else if isDarwin then
+            with pkgs.darwin.apple_sdk.frameworks; [
+              Accelerate
+              CoreGraphics
+              CoreVideo
+            ]
           else
             with pkgs; [ openblas ]
         );
         pkgs = import nixpkgs { inherit system; };
-        nativeBuildInputs = with pkgs; [ cmake pkgconfig ];
+        nativeBuildInputs = with pkgs; [ cmake ninja pkgconfig ];
         llama-python =
           pkgs.python3.withPackages (ps: with ps; [ numpy sentencepiece ]);
         postPatch = ''
@@ -38,35 +47,35 @@
           mv $out/bin/server $out/bin/llama-server
         '';
         cmakeFlags = [ "-DLLAMA_BUILD_SERVER=ON" "-DLLAMA_MPI=ON" "-DBUILD_SHARED_LIBS=ON" "-DCMAKE_SKIP_BUILD_RPATH=ON" ];
-      in {
+      in
+      {
         packages.default = pkgs.stdenv.mkDerivation {
-          name = "llama.cpp";
+          inherit name src meta postPatch nativeBuildInputs buildInputs postInstall;
-          src = ./.;
-          postPatch = postPatch;
-          nativeBuildInputs = nativeBuildInputs;
-          buildInputs = osSpecific;
           cmakeFlags = cmakeFlags
             ++ (if isAarch64 && isDarwin then [
-              "-DCMAKE_C_FLAGS=-D__ARM_FEATURE_DOTPROD=1"
+            "-DCMAKE_C_FLAGS=-D__ARM_FEATURE_DOTPROD=1"
-              "-DLLAMA_METAL=ON"
+            "-DLLAMA_METAL=ON"
-            ] else [
+          ] else [
-              "-DLLAMA_BLAS=ON"
+            "-DLLAMA_BLAS=ON"
-              "-DLLAMA_BLAS_VENDOR=OpenBLAS"
+            "-DLLAMA_BLAS_VENDOR=OpenBLAS"
           ]);
-          postInstall = postInstall;
-          meta.mainProgram = "llama";
         };
         packages.opencl = pkgs.stdenv.mkDerivation {
-          name = "llama.cpp";
+          inherit name src meta postPatch nativeBuildInputs postInstall;
-          src = ./.;
-          postPatch = postPatch;
-          nativeBuildInputs = nativeBuildInputs;
           buildInputs = with pkgs; buildInputs ++ [ clblast ];
           cmakeFlags = cmakeFlags ++ [
             "-DLLAMA_CLBLAST=ON"
           ];
-          postInstall = postInstall;
+        };
-          meta.mainProgram = "llama";
+        packages.rocm = pkgs.stdenv.mkDerivation {
+          inherit name src meta postPatch nativeBuildInputs postInstall;
+          buildInputs = with pkgs; buildInputs ++ [ hip hipblas rocblas ];
+          cmakeFlags = cmakeFlags ++ [
+            "-DLLAMA_HIPBLAS=1"
+            "-DCMAKE_C_COMPILER=hipcc"
+            "-DCMAKE_CXX_COMPILER=hipcc"
+            "-DCMAKE_POSITION_INDEPENDENT_CODE=ON"
+          ];
         };
         apps.llama-server = {
           type = "app";
@@ -80,8 +89,13 @@
           type = "app";
           program = "${self.packages.${system}.default}/bin/llama";
         };
+        apps.quantize = {
+          type = "app";
+          program = "${self.packages.${system}.default}/bin/quantize";
+        };
         apps.default = self.apps.${system}.llama;
         devShells.default = pkgs.mkShell {
+          buildInputs = [ llama-python ];
           packages = nativeBuildInputs ++ osSpecific;
         };
       });

ggml-cuda.cu

@@ -306,11 +306,11 @@ typedef struct {
 #define QI4_K (QK_K / (4*QR4_K))
 #ifdef GGML_QKK_64
 typedef struct {
-    half    d[2];              // super-block scales/mins
+    half    dm[2];             // super-block scales/mins
     uint8_t scales[2];         // 4-bit block scales/mins
     uint8_t qs[QK_K/2];        // 4--bit quants
 } block_q4_K;
-static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_fp16_t) + QK_K/2 + 2, "wrong q4_K block size/padding");
+static_assert(sizeof(block_q4_K) == sizeof(half2) + QK_K/2 + 2, "wrong q4_K block size/padding");
 #else
 typedef struct {
     half2 dm;                  // super-block scale for quantized scales/mins
@@ -737,8 +737,8 @@ static __global__ void dequantize_block_q4_K(const void * __restrict__ vx, float
     const int tid = threadIdx.x;
     const uint8_t * q = x[i].qs;
     float * y = yy + i*QK_K;
-    const float d = (float)x[i].d[0];
+    const float d = (float)x[i].dm[0];
-    const float m = (float)x[i].d[1];
+    const float m = (float)x[i].dm[1];
     y[tid+ 0] = d * (x[i].scales[0] & 0xF) * (q[tid] & 0xF) - m * (x[i].scales[0] >> 4);
     y[tid+32] = d * (x[i].scales[1] & 0xF) * (q[tid] >>  4) - m * (x[i].scales[1] >> 4);
 #endif
@@ -1155,8 +1155,8 @@ static __global__ void dequantize_mul_mat_vec_q4_k(const void * __restrict__ vx,
         const uint16_t * a = (const uint16_t *)x[i].scales;
         aux16[0] = a[0] & 0x0f0f;
         aux16[1] = (a[0] >> 4) & 0x0f0f;
-        const float d = (float)x[i].d[0];
+        const float d = (float)x[i].dm[0];
-        const float m = (float)x[i].d[1];
+        const float m = (float)x[i].dm[1];
         float sum = 0.f;
         for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) {
             sum += y[j+ 0] * (d * s[0] * (q[j+ 0] & 0xF) - m * s[2])
@@ -2845,8 +2845,8 @@ static __device__ __forceinline__ float vec_dot_q4_K_q8_1(
     aux16[0] = a[0] & 0x0f0f;
     aux16[1] = (a[0] >> 4) & 0x0f0f;
 
-    const float dall = bq4_K->d[0];
+    const float dall = bq4_K->dm[0];
-    const float dmin = bq4_K->d[1];
+    const float dmin = bq4_K->dm[1];
 
     const float d8_1 = __low2float(bq8_1[0].ds);
     const float d8_2 = __low2float(bq8_1[1].ds);
@@ -2929,7 +2929,11 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 
         const block_q4_K * bxi = bx0 + i*blocks_per_row + kbxd;
 
+#if QK_K == 256
         x_dm[i * (WARP_SIZE/QI4_K) + i / QI4_K + kbxd] = bxi->dm;
+#else
+        x_dm[i * (WARP_SIZE/QI4_K) + i / QI4_K + kbxd] = {bxi->dm[0], bxi->dm[1]};
+#endif
     }
 
 #pragma unroll
@@ -3119,7 +3123,9 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 
         const block_q5_K * bxi = bx0 + i*blocks_per_row + kbxd;
 
+#if QK_K == 256
         x_dm[i * (WARP_SIZE/QI5_K) + i / QI5_K + kbxd] = bxi->dm;
+#endif
     }
 
 #pragma unroll
@@ -4709,6 +4715,8 @@ static void ggml_mul_mat_q3_K_q8_1_cuda(
     const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x,
     const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) {
 
+#if QK_K == 256
+
     int id;
     CUDA_CHECK(cudaGetDevice(&id));
     const int compute_capability = g_compute_capabilities[id];
@@ -4740,6 +4748,7 @@ static void ggml_mul_mat_q3_K_q8_1_cuda(
         mul_mat_q3_K<need_check><<<block_nums, block_dims, 0, stream>>>
             (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
     }
+#endif
 }
 
 static void ggml_mul_mat_q4_K_q8_1_cuda(
@@ -6328,9 +6337,11 @@ void ggml_cuda_soft_max(const ggml_tensor * src0, const ggml_tensor * src1, ggml
 
 void ggml_cuda_rope(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     GGML_ASSERT(src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32);
+    GGML_ASSERT(ggml_is_contiguous(src0)); // TODO: this restriction is temporary until non-cont support is implemented
 
     const int mode = ((int32_t *) dst->op_params)[2];
     const bool is_glm = mode & 4;
+
     ggml_cuda_op(src0, src1, dst, ggml_cuda_op_rope, true, !is_glm); // flatten support not implemented for glm
 }
 

ggml.c

@@ -19394,7 +19394,7 @@ size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, i
 ////////////////////////////////////////////////////////////////////////////////
 
 struct gguf_str {
-    uint32_t n;
+    uint64_t n;  // GGUFv2
     char * data;
 };
 
@@ -19408,9 +19408,12 @@ static const size_t GGUF_TYPE_SIZE[GGUF_TYPE_COUNT] = {
     [GGUF_TYPE_FLOAT32] = sizeof(float),
     [GGUF_TYPE_BOOL]    = sizeof(bool),
     [GGUF_TYPE_STRING]  = sizeof(struct gguf_str),
+    [GGUF_TYPE_UINT64]  = sizeof(uint64_t),
+    [GGUF_TYPE_INT64]   = sizeof(int64_t),
+    [GGUF_TYPE_FLOAT64] = sizeof(double),
     [GGUF_TYPE_ARRAY]   = 0, // undefined
 };
-static_assert(GGUF_TYPE_COUNT == 10, "GGUF_TYPE_COUNT != 10");
+static_assert(GGUF_TYPE_COUNT == 13, "GGUF_TYPE_COUNT != 13");
 
 static const char * GGUF_TYPE_NAME[GGUF_TYPE_COUNT] = {
     [GGUF_TYPE_UINT8]   = "u8",
@@ -19423,8 +19426,11 @@ static const char * GGUF_TYPE_NAME[GGUF_TYPE_COUNT] = {
     [GGUF_TYPE_BOOL]    = "bool",
     [GGUF_TYPE_STRING]  = "str",
     [GGUF_TYPE_ARRAY]   = "arr",
+    [GGUF_TYPE_UINT64]  = "u64",
+    [GGUF_TYPE_INT64]   = "i64",
+    [GGUF_TYPE_FLOAT64] = "f64",
 };
-static_assert(GGUF_TYPE_COUNT == 10, "GGUF_TYPE_COUNT != 10");
+static_assert(GGUF_TYPE_COUNT == 13, "GGUF_TYPE_COUNT != 13");
 
 union gguf_value {
     uint8_t  uint8;
@@ -19434,6 +19440,9 @@ union gguf_value {
     uint32_t uint32;
     int32_t  int32;
     float    float32;
+    uint64_t uint64;
+    int64_t  int64;
+    double   float64;
     bool     bool_;
 
     struct gguf_str str;
@@ -19441,7 +19450,7 @@ union gguf_value {
     struct {
         enum gguf_type type;
 
-        uint32_t n;
+        uint64_t n;  // GGUFv2
         void * data;
     } arr;
 };
@@ -19449,8 +19458,6 @@ union gguf_value {
 struct gguf_kv {
     struct gguf_str key;
 
-    uint32_t n_bytes; // TODO: is this actually needed?
-
     enum  gguf_type  type;
     union gguf_value value;
 };
@@ -19458,15 +19465,15 @@ struct gguf_kv {
 struct gguf_header {
     uint32_t magic;
     uint32_t version;
-    uint32_t n_tensors;
+    uint64_t n_tensors; // GGUFv2
-    uint32_t n_kv;
+    uint64_t n_kv;      // GGUFv2
 };
 
 struct gguf_tensor_info {
     struct gguf_str name;
 
     uint32_t n_dims;
-    uint32_t ne[GGML_MAX_DIMS];
+    uint64_t ne[GGML_MAX_DIMS];
 
     enum ggml_type type;
 
@@ -19497,19 +19504,32 @@ static bool gguf_fread_el(FILE * file, void * dst, size_t size, size_t * offset)
     return n == size;
 }
 
-static bool gguf_fread_str(FILE * file, struct gguf_str * p, size_t * offset) {
+// NOTE: temporary handling of GGUFv1 >> remove after Oct 2023
+static bool gguf_fread_str_cur(FILE * file, struct gguf_str * p, size_t * offset) {
     p->n    = 0;
     p->data = NULL;
 
     bool ok = true;
 
-    // TODO: how to avoid mallocs for strings?
     ok = ok && gguf_fread_el(file, &p->n,    sizeof(p->n), offset); p->data = calloc(p->n + 1, 1);
     ok = ok && gguf_fread_el(file,  p->data, p->n,         offset);
 
     return ok;
 }
 
+static bool gguf_fread_str_v1(FILE * file, struct gguf_str * p, size_t * offset) {
+    p->n    = 0;
+    p->data = NULL;
+
+    bool ok = true;
+
+    uint32_t n = 0;
+    ok = ok && gguf_fread_el(file, &n,       sizeof(n),   offset); p->data = calloc(n + 1, 1); p->n = n;
+    ok = ok && gguf_fread_el(file,  p->data, p->n,         offset);
+
+    return ok;
+}
+
 struct gguf_context * gguf_init_empty(void) {
     struct gguf_context * ctx = GGML_ALIGNED_MALLOC(sizeof(struct gguf_context));
 
@@ -19565,8 +19585,21 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
         ctx->data  = NULL;
 
         ok = ok && gguf_fread_el(file, &ctx->header.version,   sizeof(ctx->header.version),   &offset);
-        ok = ok && gguf_fread_el(file, &ctx->header.n_tensors, sizeof(ctx->header.n_tensors), &offset);
+
-        ok = ok && gguf_fread_el(file, &ctx->header.n_kv,      sizeof(ctx->header.n_kv),      &offset);
+        if (ctx->header.version == 1) {
+            // NOTE: temporary handling of GGUFv1 >> remove after Oct 2023
+            uint32_t n_tensors = 0;
+            uint32_t n_kv      = 0;
+
+            ok = ok && gguf_fread_el(file, &n_tensors, sizeof(n_tensors), &offset);
+            ok = ok && gguf_fread_el(file, &n_kv,      sizeof(n_kv),      &offset);
+
+            ctx->header.n_tensors = n_tensors;
+            ctx->header.n_kv      = n_kv;
+        } else {
+            ok = ok && gguf_fread_el(file, &ctx->header.n_tensors, sizeof(ctx->header.n_tensors), &offset);
+            ok = ok && gguf_fread_el(file, &ctx->header.n_kv,      sizeof(ctx->header.n_kv),      &offset);
+        }
 
         if (!ok) {
             fprintf(stderr, "%s: failed to read header\n", __func__);
@@ -19576,6 +19609,12 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
         }
     }
 
+    // NOTE: temporary handling of GGUFv1 >> remove after Oct 2023
+    bool (* gguf_fread_str)(FILE *, struct gguf_str *, size_t *) = gguf_fread_str_cur;
+    if (ctx->header.version == 1) {
+        gguf_fread_str = gguf_fread_str_v1;
+    }
+
     // read the kv pairs
     {
         ctx->kv = GGML_ALIGNED_MALLOC(ctx->header.n_kv * sizeof(struct gguf_kv));
@@ -19585,9 +19624,8 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
 
             //fprintf(stderr, "%s: reading kv %d\n", __func__, i);
 
-            ok = ok && gguf_fread_str(file, &kv->key,                          &offset);
+            ok = ok && gguf_fread_str(file, &kv->key,                    &offset);
-          //ok = ok && gguf_fread_el (file, &kv->n_bytes, sizeof(kv->n_bytes), &offset);
+            ok = ok && gguf_fread_el (file, &kv->type, sizeof(kv->type), &offset);
-            ok = ok && gguf_fread_el (file, &kv->type,    sizeof(kv->type),    &offset);
 
             //fprintf(stderr, "%s: reading kv with key %s\n", __func__, kv->key.data);
 
@@ -19599,12 +19637,23 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
                 case GGUF_TYPE_UINT32:  ok = ok && gguf_fread_el (file, &kv->value.uint32,  sizeof(kv->value.uint32),  &offset); break;
                 case GGUF_TYPE_INT32:   ok = ok && gguf_fread_el (file, &kv->value.int32,   sizeof(kv->value.int32),   &offset); break;
                 case GGUF_TYPE_FLOAT32: ok = ok && gguf_fread_el (file, &kv->value.float32, sizeof(kv->value.float32), &offset); break;
+                case GGUF_TYPE_UINT64:  ok = ok && gguf_fread_el (file, &kv->value.uint64,  sizeof(kv->value.uint64),  &offset); break;
+                case GGUF_TYPE_INT64:   ok = ok && gguf_fread_el (file, &kv->value.int64,   sizeof(kv->value.int64),   &offset); break;
+                case GGUF_TYPE_FLOAT64: ok = ok && gguf_fread_el (file, &kv->value.float64, sizeof(kv->value.float64), &offset); break;
                 case GGUF_TYPE_BOOL:    ok = ok && gguf_fread_el (file, &kv->value.bool_,   sizeof(kv->value.bool_),   &offset); break;
                 case GGUF_TYPE_STRING:  ok = ok && gguf_fread_str(file, &kv->value.str,                                &offset); break;
                 case GGUF_TYPE_ARRAY:
                     {
                         ok = ok && gguf_fread_el(file, &kv->value.arr.type, sizeof(kv->value.arr.type), &offset);
-                        ok = ok && gguf_fread_el(file, &kv->value.arr.n,    sizeof(kv->value.arr.n),    &offset);
+
+                        if (ctx->header.version == 1) {
+                            // NOTE: temporary handling of GGUFv1 >> remove after Oct 2023
+                            uint32_t n = 0;
+                            ok = ok && gguf_fread_el(file, &n, sizeof(n), &offset);
+                            kv->value.arr.n = n;
+                        } else {
+                            ok = ok && gguf_fread_el(file, &kv->value.arr.n, sizeof(kv->value.arr.n), &offset);
+                        }
 
                         switch (kv->value.arr.type) {
                             case GGUF_TYPE_UINT8:
@@ -19614,6 +19663,9 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
                             case GGUF_TYPE_UINT32:
                             case GGUF_TYPE_INT32:
                             case GGUF_TYPE_FLOAT32:
+                            case GGUF_TYPE_UINT64:
+                            case GGUF_TYPE_INT64:
+                            case GGUF_TYPE_FLOAT64:
                             case GGUF_TYPE_BOOL:
                                 {
                                     kv->value.arr.data = malloc(kv->value.arr.n * GGUF_TYPE_SIZE[kv->value.arr.type]);
@@ -19660,7 +19712,14 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
             ok = ok && gguf_fread_str(file, &info->name,                          &offset);
             ok = ok && gguf_fread_el (file, &info->n_dims, sizeof(info->n_dims),  &offset);
             for (uint32_t j = 0; j < info->n_dims; ++j) {
-                ok = ok && gguf_fread_el(file, &info->ne[j], sizeof(info->ne[j]), &offset);
+                if (ctx->header.version == 1) {
+                    // NOTE: temporary handling of GGUFv1 >> remove after Oct 2023
+                    uint32_t t = 0;
+                    ok = ok && gguf_fread_el(file, &t, sizeof(t), &offset);
+                    info->ne[j] = t;
+                } else {
+                    ok = ok && gguf_fread_el(file, &info->ne[j], sizeof(info->ne[j]), &offset);
+                }
             }
             ok = ok && gguf_fread_el (file, &info->type,   sizeof(info->type),    &offset);
             ok = ok && gguf_fread_el (file, &info->offset, sizeof(info->offset),  &offset);
@@ -19954,6 +20013,18 @@ float gguf_get_val_f32(struct gguf_context * ctx, int i) {
     return ctx->kv[i].value.float32;
 }
 
+uint64_t gguf_get_val_u64(struct gguf_context * ctx, int i) {
+    return ctx->kv[i].value.uint64;
+}
+
+int64_t gguf_get_val_i64(struct gguf_context * ctx, int i) {
+    return ctx->kv[i].value.int64;
+}
+
+double gguf_get_val_f64(struct gguf_context * ctx, int i) {
+    return ctx->kv[i].value.float64;
+}
+
 bool gguf_get_val_bool(struct gguf_context * ctx, int i) {
     return ctx->kv[i].value.bool_;
 }
@@ -20056,6 +20127,27 @@ void gguf_set_val_f32(struct gguf_context * ctx, const char * key, float val) {
     ctx->kv[idx].value.float32 = val;
 }
 
+void gguf_set_val_u64(struct gguf_context * ctx, const char * key, uint64_t val) {
+    const int idx = gguf_get_or_add_key(ctx, key);
+
+    ctx->kv[idx].type         = GGUF_TYPE_UINT64;
+    ctx->kv[idx].value.uint64 = val;
+}
+
+void gguf_set_val_i64(struct gguf_context * ctx, const char * key, int64_t val) {
+    const int idx = gguf_get_or_add_key(ctx, key);
+
+    ctx->kv[idx].type        = GGUF_TYPE_INT64;
+    ctx->kv[idx].value.int64 = val;
+}
+
+void gguf_set_val_f64(struct gguf_context * ctx, const char * key, double val) {
+    const int idx = gguf_get_or_add_key(ctx, key);
+
+    ctx->kv[idx].type          = GGUF_TYPE_FLOAT64;
+    ctx->kv[idx].value.float64 = val;
+}
+
 void gguf_set_val_bool(struct gguf_context * ctx, const char * key, bool val) {
     const int idx = gguf_get_or_add_key(ctx, key);
 
@@ -20106,6 +20198,9 @@ void gguf_set_kv(struct gguf_context * ctx, struct gguf_context * src) {
             case GGUF_TYPE_UINT32:  gguf_set_val_u32 (ctx, src->kv[i].key.data, src->kv[i].value.uint32);   break;
             case GGUF_TYPE_INT32:   gguf_set_val_i32 (ctx, src->kv[i].key.data, src->kv[i].value.int32);    break;
             case GGUF_TYPE_FLOAT32: gguf_set_val_f32 (ctx, src->kv[i].key.data, src->kv[i].value.float32);  break;
+            case GGUF_TYPE_UINT64:  gguf_set_val_u64 (ctx, src->kv[i].key.data, src->kv[i].value.uint64);   break;
+            case GGUF_TYPE_INT64:   gguf_set_val_i64 (ctx, src->kv[i].key.data, src->kv[i].value.int64);    break;
+            case GGUF_TYPE_FLOAT64: gguf_set_val_f64 (ctx, src->kv[i].key.data, src->kv[i].value.float64);  break;
             case GGUF_TYPE_BOOL:    gguf_set_val_bool(ctx, src->kv[i].key.data, src->kv[i].value.bool_);    break;
             case GGUF_TYPE_STRING:  gguf_set_val_str (ctx, src->kv[i].key.data, src->kv[i].value.str.data); break;
             case GGUF_TYPE_ARRAY:
@@ -20267,6 +20362,9 @@ static void gguf_write_to_buf(struct gguf_context * ctx, struct gguf_buf * buf,
             case GGUF_TYPE_UINT32:  gguf_bwrite_el (buf, &kv->value.uint32,  sizeof(kv->value.uint32) ); break;
             case GGUF_TYPE_INT32:   gguf_bwrite_el (buf, &kv->value.int32,   sizeof(kv->value.int32)  ); break;
             case GGUF_TYPE_FLOAT32: gguf_bwrite_el (buf, &kv->value.float32, sizeof(kv->value.float32)); break;
+            case GGUF_TYPE_UINT64:  gguf_bwrite_el (buf, &kv->value.uint64,  sizeof(kv->value.uint64) ); break;
+            case GGUF_TYPE_INT64:   gguf_bwrite_el (buf, &kv->value.int64,   sizeof(kv->value.int64)  ); break;
+            case GGUF_TYPE_FLOAT64: gguf_bwrite_el (buf, &kv->value.float64, sizeof(kv->value.float64)); break;
             case GGUF_TYPE_BOOL:    gguf_bwrite_el (buf, &kv->value.bool_,   sizeof(kv->value.bool_)  ); break;
             case GGUF_TYPE_STRING:  gguf_bwrite_str(buf, &kv->value.str                               ); break;
             case GGUF_TYPE_ARRAY:
@@ -20282,6 +20380,9 @@ static void gguf_write_to_buf(struct gguf_context * ctx, struct gguf_buf * buf,
                         case GGUF_TYPE_UINT32:
                         case GGUF_TYPE_INT32:
                         case GGUF_TYPE_FLOAT32:
+                        case GGUF_TYPE_UINT64:
+                        case GGUF_TYPE_INT64:
+                        case GGUF_TYPE_FLOAT64:
                         case GGUF_TYPE_BOOL:
                             {
                                 gguf_bwrite_el(buf, kv->value.arr.data, kv->value.arr.n * GGUF_TYPE_SIZE[kv->value.arr.type]);
@@ -20516,6 +20617,14 @@ int ggml_cpu_has_sse3(void) {
 #endif
 }
 
+int ggml_cpu_has_ssse3(void) {
+#if defined(__SSSE3__)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
 int ggml_cpu_has_vsx(void) {
 #if defined(__POWER9_VECTOR__)
     return 1;

ggml.h

@@ -216,7 +216,7 @@
 #define GGML_EXIT_ABORTED 1
 
 #define GGUF_MAGIC   0x46554747 // "GGUF"
-#define GGUF_VERSION 1
+#define GGUF_VERSION 2
 
 #define GGUF_DEFAULT_ALIGNMENT 32
 
@@ -1827,6 +1827,9 @@ extern "C" {
         GGUF_TYPE_BOOL    = 7,
         GGUF_TYPE_STRING  = 8,
         GGUF_TYPE_ARRAY   = 9,
+        GGUF_TYPE_UINT64  = 10,
+        GGUF_TYPE_INT64   = 11,
+        GGUF_TYPE_FLOAT64 = 12,
         GGUF_TYPE_COUNT,       // marks the end of the enum
     };
 
@@ -1867,6 +1870,9 @@ extern "C" {
     GGML_API uint32_t     gguf_get_val_u32 (struct gguf_context * ctx, int i);
     GGML_API int32_t      gguf_get_val_i32 (struct gguf_context * ctx, int i);
     GGML_API float        gguf_get_val_f32 (struct gguf_context * ctx, int i);
+    GGML_API uint64_t     gguf_get_val_u64 (struct gguf_context * ctx, int i);
+    GGML_API int64_t      gguf_get_val_i64 (struct gguf_context * ctx, int i);
+    GGML_API double       gguf_get_val_f64 (struct gguf_context * ctx, int i);
     GGML_API bool         gguf_get_val_bool(struct gguf_context * ctx, int i);
     GGML_API const char * gguf_get_val_str (struct gguf_context * ctx, int i);
     GGML_API int          gguf_get_arr_n   (struct gguf_context * ctx, int i);
@@ -1886,6 +1892,9 @@ extern "C" {
     GGML_API void gguf_set_val_u32 (struct gguf_context * ctx, const char * key, uint32_t val);
     GGML_API void gguf_set_val_i32 (struct gguf_context * ctx, const char * key, int32_t  val);
     GGML_API void gguf_set_val_f32 (struct gguf_context * ctx, const char * key, float    val);
+    GGML_API void gguf_set_val_u64 (struct gguf_context * ctx, const char * key, uint64_t val);
+    GGML_API void gguf_set_val_i64 (struct gguf_context * ctx, const char * key, int64_t  val);
+    GGML_API void gguf_set_val_f64 (struct gguf_context * ctx, const char * key, double   val);
     GGML_API void gguf_set_val_bool(struct gguf_context * ctx, const char * key, bool     val);
     GGML_API void gguf_set_val_str (struct gguf_context * ctx, const char * key, const char * val);
     GGML_API void gguf_set_arr_data(struct gguf_context * ctx, const char * key, enum gguf_type type, const void * data, int n);
@@ -1944,6 +1953,7 @@ extern "C" {
     GGML_API int ggml_cpu_has_clblast    (void);
     GGML_API int ggml_cpu_has_gpublas    (void);
     GGML_API int ggml_cpu_has_sse3       (void);
+    GGML_API int ggml_cpu_has_ssse3      (void);
     GGML_API int ggml_cpu_has_vsx        (void);
 
     //

gguf-py/gguf/gguf.py

@@ -13,7 +13,7 @@ from typing import Any, IO, List, Optional
 #
 
 GGUF_MAGIC             = 0x46554747
-GGUF_VERSION           = 1
+GGUF_VERSION           = 2
 GGUF_DEFAULT_ALIGNMENT = 32
 
 # general
@@ -365,6 +365,9 @@ class GGUFValueType(IntEnum):
     BOOL    = 7
     STRING  = 8
     ARRAY   = 9
+    UINT64  = 10
+    INT64   = 11
+    FLOAT64 = 12
 
     @staticmethod
     def get_type(val):
@@ -378,6 +381,7 @@ class GGUFValueType(IntEnum):
             return GGUFValueType.BOOL
         elif isinstance(val, int):
             return GGUFValueType.INT32
+        # TODO: need help with 64-bit types in Python
         else:
             print("Unknown type: "+str(type(val)))
             sys.exit()
@@ -400,8 +404,8 @@ class GGUFWriter:
     def write_header_to_file(self):
         self.fout.write(struct.pack("<I", GGUF_MAGIC))
         self.fout.write(struct.pack("<I", GGUF_VERSION))
-        self.fout.write(struct.pack("<I", self.ti_data_count))
+        self.fout.write(struct.pack("<Q", self.ti_data_count))
-        self.fout.write(struct.pack("<I", self.kv_data_count))
+        self.fout.write(struct.pack("<Q", self.kv_data_count))
         self.flush()
 #        print("tensors " + str(self.ti_data_count) + " kv " + str(self.kv_data_count))
 
@@ -444,6 +448,18 @@ class GGUFWriter:
         self.add_key(key)
         self.add_val(val, GGUFValueType.FLOAT32)
 
+    def add_uint64(self, key: str, val: int):
+        self.add_key(key)
+        self.add_val(val, GGUFValueType.UINT64)
+
+    def add_int64(self, key: str, val: int):
+        self.add_key(key)
+        self.add_val(val, GGUFValueType.INT64)
+
+    def add_float64(self, key: str, val: float):
+        self.add_key(key)
+        self.add_val(val, GGUFValueType.FLOAT64)
+
     def add_bool(self, key: str, val: bool):
         self.add_key(key)
         self.add_val(val, GGUFValueType.BOOL)
@@ -483,17 +499,23 @@ class GGUFWriter:
             self.kv_data += struct.pack("<i", val)
         elif vtype == GGUFValueType.FLOAT32:
             self.kv_data += struct.pack("<f", val)
+        elif vtype == GGUFValueType.UINT64:
+            self.kv_data += struct.pack("<Q", val)
+        elif vtype == GGUFValueType.INT64:
+            self.kv_data += struct.pack("<q", val)
+        elif vtype == GGUFValueType.FLOAT64:
+            self.kv_data += struct.pack("<d", val)
         elif vtype == GGUFValueType.BOOL:
             self.kv_data += struct.pack("?", val)
         elif vtype == GGUFValueType.STRING:
             encoded_val = val.encode("utf8") if isinstance(val, str) else val
-            self.kv_data += struct.pack("<I", len(encoded_val))
+            self.kv_data += struct.pack("<Q", len(encoded_val))
             self.kv_data += encoded_val
         elif vtype == GGUFValueType.ARRAY:
             ltype = set([GGUFValueType.get_type(item) for item in val])
             assert len(ltype) == 1, "All items in a GGUF array should be of the same type"
             self.kv_data += struct.pack("<I", list(ltype)[0])
-            self.kv_data += struct.pack("<I", len(val))
+            self.kv_data += struct.pack("<Q", len(val))
             for item in val:
                 self.add_val(item, add_vtype=False)
         else:
@@ -507,12 +529,12 @@ class GGUFWriter:
         assert raw_dtype is not None or tensor_dtype in (np.float32, np.float16), "Only F32 and F16 tensors are supported for now"
 
         encoded_name = name.encode("utf8")
-        self.ti_data += struct.pack("<I", len(encoded_name))
+        self.ti_data += struct.pack("<Q", len(encoded_name))
         self.ti_data += encoded_name
         n_dims = len(tensor_shape)
         self.ti_data += struct.pack("<I", n_dims)
         for i in range(n_dims):
-            self.ti_data += struct.pack("<I", tensor_shape[n_dims - 1 - i])
+            self.ti_data += struct.pack("<Q", tensor_shape[n_dims - 1 - i])
         if raw_dtype is None:
             dtype = GGMLQuantizationType.F32 if tensor_dtype == np.float32 else GGMLQuantizationType.F16
         else:

llama.cpp

@@ -1,9 +1,6 @@
 // Defines fileno on msys:
 #ifndef _GNU_SOURCE
 #define _GNU_SOURCE
-#include <cstddef>
-#include <cstdint>
-#include <cstdio>
 #endif
 
 #include "llama.h"
@@ -62,6 +59,9 @@
 #include <cinttypes>
 #include <climits>
 #include <cstdarg>
+#include <cstddef>
+#include <cstdint>
+#include <cstdio>
 #include <cstring>
 #include <ctime>
 #include <fstream>
@@ -114,12 +114,17 @@ static size_t utf8_len(char src) {
 }
 
 void replace_all(std::string & s, const std::string & search, const std::string & replace) {
-    for (size_t pos = 0; ; pos += replace.length()) {
+    std::string result;
-        pos = s.find(search, pos);
+    for (size_t pos = 0; ; pos += search.length()) {
-        if (pos == std::string::npos) break;
+        auto new_pos = s.find(search, pos);
-        s.erase(pos, search.length());
+        if (new_pos == std::string::npos) {
-        s.insert(pos, replace);
+            result += s.substr(pos, s.size() - pos);
+            break;
+        }
+        result += s.substr(pos, new_pos - pos) + replace;
+        pos = new_pos;
     }
+    s = std::move(result);
 }
 
 static void zeros(std::ofstream & file, size_t n) {
@@ -796,12 +801,12 @@ static void llama_nop(struct ggml_tensor * tensor) { // don't offload by default
     (void) tensor;
 }
 
-static std::string llama_token_to_text(const struct llama_context * ctx, llama_token token) {
+static std::string llama_token_to_str(const struct llama_context * ctx, llama_token token) {
     std::vector<char> result(8, 0);
-    const int n_tokens = llama_token_to_str(ctx, token, result.data(), result.size());
+    const int n_tokens = llama_token_to_piece(ctx, token, result.data(), result.size());
     if (n_tokens < 0) {
         result.resize(-n_tokens);
-        int check = llama_token_to_str(ctx, token, result.data(), result.size());
+        int check = llama_token_to_piece(ctx, token, result.data(), result.size());
         GGML_ASSERT(check == -n_tokens);
     } else {
         result.resize(n_tokens);
@@ -955,10 +960,10 @@ struct llama_vocab {
     id linefeed_id = 13;
 
     int find_bpe_rank(std::string token_left, std::string token_right) const {
-        replace_all(token_left,  " ",  "Ġ");
+        replace_all(token_left,  " ",  "\u0120");
-        replace_all(token_left,  "\n", "Ċ");
+        replace_all(token_left,  "\n", "\u010A");
-        replace_all(token_right, " ",  "Ġ");
+        replace_all(token_right, " ",  "\u0120");
-        replace_all(token_right, "\n", "Ċ");
+        replace_all(token_right, "\n", "\u010A");
 
         auto it = bpe_ranks.find(std::make_pair(token_left, token_right));
         if (it == bpe_ranks.end()) {
@@ -1144,11 +1149,13 @@ static bool llama_kv_cache_init(
 
 enum llama_fver {
     GGUF_FILE_VERSION_V1 = 1,
+    GGUF_FILE_VERSION_V2 = 2,
 };
 
 static const char * llama_file_version_name(llama_fver version) {
     switch (version) {
-        case GGUF_FILE_VERSION_V1: return "GGUF V1 (latest)";
+        case GGUF_FILE_VERSION_V1: return "GGUF V1 (support until nov 2023)";
+        case GGUF_FILE_VERSION_V2: return "GGUF V2 (latest)";
     }
 
     return "unknown";
@@ -1635,7 +1642,8 @@ static void llm_load_hparams(
 }
 
 // TODO: This should probably be in llama.h
-static std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab & vocab, const std::string & raw_text, bool bos, bool escape);
+static std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab & vocab, std::string raw_text, bool bos);
+static llama_token llama_byte_to_token(const llama_vocab & vocab, uint8_t ch);
 
 static void llm_load_vocab(
         llama_model_loader & ml,
@@ -1737,7 +1745,11 @@ static void llm_load_vocab(
     }
 
     // determine the newline token: LLaMA "<0x0A>" == 10 == '\n', Falcon 193 == '\n'
-    vocab.linefeed_id = llama_tokenize_internal(vocab, "\n", false, false)[0];
+    if (vocab.type == LLAMA_VOCAB_TYPE_SPM) {
+        vocab.linefeed_id = llama_byte_to_token(vocab, '\n');
+    } else {
+        vocab.linefeed_id = llama_tokenize_internal(vocab, "\n", false)[0];
+    }
 
     // special tokens
     GGUF_GET_KEY(ctx, vocab.special_bos_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_BOS_ID));
@@ -2635,18 +2647,20 @@ static struct ggml_cgraph * llm_build_falcon(
 
             const size_t wsize = ggml_type_size(cur->type);
 
-            struct ggml_tensor * tmpq = ggml_view_3d(
+            // TODO: these 2 ggml_conts are technically not needed, but we add them until CUDA support for
+            //       non-contiguous views is added for the rope operator
+            struct ggml_tensor * tmpq = ggml_cont(ctx0, ggml_view_3d(
                 ctx0, cur, n_embd_head, n_head, N,
                 wsize * n_embd_head,
                 wsize * n_embd_head * (n_head + 2 * n_head_kv),
-                0);
+                0));
             offload_func_kq(tmpq);
 
-            struct ggml_tensor * tmpk = ggml_view_3d(
+            struct ggml_tensor * tmpk = ggml_cont(ctx0, ggml_view_3d(
                 ctx0, cur, n_embd_head, n_head_kv, N,
                 wsize * n_embd_head,
                 wsize * n_embd_head * (n_head + 2 * n_head_kv),
-                wsize * n_embd_head *  n_head);
+                wsize * n_embd_head *  n_head));
             offload_func_kq(tmpk);
 
             struct ggml_tensor * tmpv = ggml_view_3d(
@@ -2831,7 +2845,6 @@ static bool llama_eval_internal(
 
     GGML_ASSERT(n_tokens > 0);
     GGML_ASSERT(n_past >= 0);
-    GGML_ASSERT(n_threads > 0);
     // TODO: keep the values of n_batch and n_ctx
     // GGML_ASSERT(n_tokens <= n_batch);
     // GGML_ASSERT(n_past + n_tokens <= n_ctx);
@@ -2842,6 +2855,8 @@ static bool llama_eval_internal(
     ggml_mpi_eval_init(lctx.ctx_mpi, &n_tokens, &n_past, &n_threads);
 #endif
 
+    GGML_ASSERT(n_threads > 0);
+
     const int N = n_tokens;
 
     const auto & model   = lctx.model;
@@ -3026,16 +3041,8 @@ static llama_token llama_byte_to_token(const llama_vocab & vocab, uint8_t ch) {
     return vocab.token_to_id.at(buf);
 }
 
-static std::string llama_escape_whitespace(const std::string& text) {
+static void llama_escape_whitespace(std::string & text) {
-    std::string result = "\xe2\x96\x81";
+    replace_all(text, " ", "\xe2\x96\x81");
-    for (size_t offs = 0; offs < text.length(); ++offs) {
-        if (text[offs] == ' ') {
-            result += "\xe2\x96\x81";
-        } else {
-            result += text[offs];
-        }
-    }
-    return result;
 }
 
 static void llama_unescape_whitespace(std::string & word) {
@@ -3219,7 +3226,7 @@ struct llm_bigram_bpe {
 };
 
 struct llm_tokenizer_bpe {
-    llm_tokenizer_bpe(const llama_vocab & vocab, bool g2ws): vocab(vocab) { flag_g2ws = g2ws; }
+    llm_tokenizer_bpe(const llama_vocab & vocab): vocab(vocab) {}
 
     void tokenize(const std::string & text, std::vector<llama_vocab::id> & output) {
         int final_prev_index = -1;
@@ -3371,8 +3378,6 @@ private:
         return words;
     }
 
-    bool flag_g2ws = false;
-
     const llama_vocab & vocab;
 
     std::vector<llm_symbol> symbols;
@@ -3381,9 +3386,18 @@ private:
     llm_bigram_bpe::queue work_queue;
 };
 
-static std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab & vocab, const std::string & raw_text, bool bos, bool escape) {
+static std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab & vocab, std::string raw_text, bool bos) {
     std::vector<llama_vocab::id> output;
 
+    // OG tokenizer behavior:
+    //
+    // tokenizer.encode('', add_bos=True)  returns [1]
+    // tokenizer.encode('', add_bos=False) returns []
+
+    if (bos && vocab.special_bos_id != -1) {
+        output.push_back(vocab.special_bos_id);
+    }
+
     if (raw_text.empty()) {
         return output;
     }
@@ -3391,29 +3405,16 @@ static std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab &
     switch (vocab.type) {
         case LLAMA_VOCAB_TYPE_SPM:
             {
+                // without adding this leading whitespace, we do not get the same results as the original tokenizer
+                raw_text = " " + raw_text;
+
                 llm_tokenizer_spm tokenizer(vocab);
-
+                llama_escape_whitespace(raw_text);
-                if (bos) {
+                tokenizer.tokenize(raw_text, output);
-                    output.push_back(vocab.special_bos_id);
-                }
-
-                std::string text;
-                if (escape) {
-                    text = llama_escape_whitespace(raw_text);
-                } else {
-                    text = raw_text;
-                }
-
-                tokenizer.tokenize(text, output);
             } break;
         case LLAMA_VOCAB_TYPE_BPE:
             {
-                llm_tokenizer_bpe tokenizer(vocab, escape);
+                llm_tokenizer_bpe tokenizer(vocab);
-
-                if (bos && vocab.special_bos_id != -1) {
-                    output.push_back(vocab.special_bos_id);
-                }
-
                 tokenizer.tokenize(raw_text, output);
             } break;
     };
@@ -3908,7 +3909,7 @@ void llama_sample_tail_free(struct llama_context * ctx, llama_token_data_array *
 
     // Calculate absolute value of second derivatives
     for (size_t i = 0; i < second_derivatives.size(); ++i) {
-        second_derivatives[i] = abs(second_derivatives[i]);
+        second_derivatives[i] = std::abs(second_derivatives[i]);
     }
 
     // Normalize the second derivatives
@@ -4099,16 +4100,16 @@ void llama_sample_grammar(struct llama_context * ctx, llama_token_data_array * c
     std::vector<llama_grammar_candidate>                              candidates_grammar;
 
     for (size_t i = 0; i < candidates->size; ++i) {
-        const llama_token id   = candidates->data[i].id;
+        const llama_token id    = candidates->data[i].id;
-        const std::string text = llama_token_to_text(ctx, id);
+        const std::string piece = llama_token_to_str(ctx, id);
         if (id == eos) {
             if (!allow_eos) {
                 candidates->data[i].logit = -INFINITY;
             }
-        } else if (text.empty() || text[0] == 0) {
+        } else if (piece.empty() || piece[0] == 0) {
             candidates->data[i].logit = -INFINITY;
         } else {
-            candidates_decoded.push_back(decode_utf8(text.c_str(), grammar->partial_utf8));
+            candidates_decoded.push_back(decode_utf8(piece.c_str(), grammar->partial_utf8));
             candidates_grammar.push_back({ i, candidates_decoded.back().first.data(), candidates_decoded.back().second });
         }
     }
@@ -4312,10 +4313,10 @@ void llama_grammar_accept_token(struct llama_context * ctx, struct llama_grammar
         GGML_ASSERT(false);
     }
 
-    const std::string text = llama_token_to_text(ctx, token);
+    const std::string piece = llama_token_to_str(ctx, token);
 
     // Note terminating 0 in decoded string
-    const auto   decoded     = decode_utf8(text.c_str(), grammar->partial_utf8);
+    const auto   decoded     = decode_utf8(piece.c_str(), grammar->partial_utf8);
     const auto & code_points = decoded.first;
     for (auto it = code_points.begin(), end = code_points.end() - 1; it != end; ++it) {
         grammar->stacks = llama_grammar_accept(grammar->rules, grammar->stacks, *it);
@@ -4326,6 +4327,257 @@ void llama_grammar_accept_token(struct llama_context * ctx, struct llama_grammar
     ctx->t_sample_us += ggml_time_us() - t_start_sample_us;
 }
 
+//
+// Beam search
+//
+
+struct llama_beam {
+    std::vector<llama_token> tokens;
+    float p;  // Cumulative beam probability (renormalized relative to all beams)
+    bool eob; // Initialize end-of-beam to false. Callback sets this to true.
+    // Sort beams by probability. In case of ties, prefer beams at eob.
+    bool operator<(const llama_beam & rhs) const {
+        return std::make_pair(p, eob) < std::make_pair(rhs.p, rhs.eob);
+    }
+    // Shift off first n tokens and discard them.
+    void shift_tokens(const size_t n) {
+        if (n) {
+            std::copy(tokens.begin() + n, tokens.end(), tokens.begin());
+            tokens.resize(tokens.size() - n);
+        }
+    }
+    llama_beam_view view() const { return {tokens.data(), tokens.size(), p, eob}; }
+};
+
+// A struct for calculating logit-related info.
+struct llama_logit_info {
+    const float * const logits;
+    const int n_vocab;
+    const float max_l;
+    const float normalizer;
+    struct sum_exp {
+        float max_l;
+        float operator()(float sum, float l) const { return sum + std::exp(l - max_l); }
+    };
+    llama_logit_info(llama_context * ctx)
+      : logits(llama_get_logits(ctx))
+      , n_vocab(llama_n_vocab(ctx))
+      , max_l(*std::max_element(logits, logits + n_vocab))
+      , normalizer(1.0f / std::accumulate(logits, logits + n_vocab, 0.0f, sum_exp{max_l}))
+      { }
+    llama_token_data get_token_data(const llama_token token_id) const {
+        constexpr auto p = std::numeric_limits<float>::quiet_NaN();  // never used
+        return {token_id, logits[token_id], p};
+    }
+    // Return top k token_data by logit.
+    std::vector<llama_token_data> top_k(size_t k) {
+        std::vector<llama_token_data> min_heap;  // min-heap by logit
+        const llama_token k_min = std::min(static_cast<llama_token>(k), n_vocab);
+        min_heap.reserve(k_min);
+        for (llama_token token_id = 0 ; token_id < k_min ; ++token_id) {
+            min_heap.push_back(get_token_data(token_id));
+        }
+        auto comp = [](const llama_token_data & a, const llama_token_data & b) { return a.logit > b.logit; };
+        std::make_heap(min_heap.begin(), min_heap.end(), comp);
+        for (llama_token token_id = k_min ; token_id < n_vocab ; ++token_id) {
+            if (min_heap.front().logit < logits[token_id]) {
+                std::pop_heap(min_heap.begin(), min_heap.end(), comp);
+                min_heap.back().id = token_id;
+                min_heap.back().logit = logits[token_id];
+                std::push_heap(min_heap.begin(), min_heap.end(), comp);
+            }
+        }
+        return min_heap;
+    }
+    float probability_from_logit(float logit) {
+        return normalizer * std::exp(logit - max_l);
+    }
+};
+
+struct llama_beam_search_data {
+    llama_context * ctx;
+    size_t n_beams;
+    int n_past;
+    int n_predict;
+    int n_threads;
+    std::vector<llama_beam> beams;
+    std::vector<llama_beam> next_beams;
+
+    // Re-calculated on each loop iteration
+    size_t common_prefix_length;
+
+    // Used to communicate to/from callback on beams state.
+    std::vector<llama_beam_view> beam_views;
+
+    llama_beam_search_data(llama_context * ctx, size_t n_beams, int n_past, int n_predict, int n_threads)
+      : ctx(ctx)
+      , n_beams(n_beams)
+      , n_past(n_past)
+      , n_predict(n_predict)
+      , n_threads(n_threads)
+      , beam_views(n_beams) {
+        beams.reserve(n_beams);
+        next_beams.reserve(n_beams);
+    }
+
+    // Collapse beams to a single beam given by index.
+    void collapse_beams(const size_t beam_idx) {
+        if (0u < beam_idx) {
+            std::swap(beams[0], beams[beam_idx]);
+        }
+        beams.resize(1);
+    }
+
+    // Min-heaps are used to efficiently collect the top-k elements (k=n_beams).
+    // The repetative patterns below reflect the 2 stages of heaps:
+    //  * Gather elements until the vector is full, then call std::make_heap() on it.
+    //  * If the heap is full and a new element is found that should be included, pop the
+    //    least element to the back(), replace it with the new, then push it into the heap.
+    void fill_next_beams_by_top_probabilities(llama_beam & beam) {
+        // Min-heaps use a greater-than comparator.
+        const auto comp = [](const llama_beam & a, const llama_beam & b) { return a.p > b.p; };
+        if (beam.eob) {
+            // beam is at end-of-sentence, so just copy it to next_beams if its probability is high enough.
+            if (next_beams.size() < n_beams) {
+                next_beams.push_back(std::move(beam));
+                if (next_beams.size() == n_beams) {
+                    std::make_heap(next_beams.begin(), next_beams.end(), comp);
+                }
+            } else if (next_beams.front().p < beam.p) {
+                std::pop_heap(next_beams.begin(), next_beams.end(), comp);
+                next_beams.back() = std::move(beam);
+                std::push_heap(next_beams.begin(), next_beams.end(), comp);
+            }
+        } else {
+            // beam is not at end-of-sentence, so branch with next top_k tokens.
+            if (!beam.tokens.empty()) {
+                llama_eval(ctx, beam.tokens.data(), beam.tokens.size(), n_past, n_threads);
+            }
+            llama_logit_info logit_info(ctx);
+            std::vector<llama_token_data> next_tokens = logit_info.top_k(n_beams);
+            size_t i=0;
+            if (next_beams.size() < n_beams) {
+                for (; next_beams.size() < n_beams ; ++i) {
+                    llama_beam next_beam = beam;
+                    next_beam.tokens.push_back(next_tokens[i].id);
+                    next_beam.p *= logit_info.probability_from_logit(next_tokens[i].logit);
+                    next_beams.push_back(std::move(next_beam));
+                }
+                std::make_heap(next_beams.begin(), next_beams.end(), comp);
+            } else {
+                for (; next_beams.front().p == 0.0f ; ++i) {
+                    std::pop_heap(next_beams.begin(), next_beams.end(), comp);
+                    next_beams.back() = beam;
+                    next_beams.back().tokens.push_back(next_tokens[i].id);
+                    next_beams.back().p *= logit_info.probability_from_logit(next_tokens[i].logit);
+                    std::push_heap(next_beams.begin(), next_beams.end(), comp);
+                }
+            }
+            for (; i < n_beams ; ++i) {
+                const float next_p = beam.p * logit_info.probability_from_logit(next_tokens[i].logit);
+                if (next_beams.front().p < next_p) {
+                    std::pop_heap(next_beams.begin(), next_beams.end(), comp);
+                    next_beams.back() = beam;
+                    next_beams.back().tokens.push_back(next_tokens[i].id);
+                    next_beams.back().p = next_p;
+                    std::push_heap(next_beams.begin(), next_beams.end(), comp);
+                }
+            }
+        }
+    }
+
+    // Find common_prefix_length based on beams.
+    // Requires beams is not empty.
+    size_t find_common_prefix_length() {
+        size_t common_prefix_length = beams[0].tokens.size();
+        for (size_t i = 1 ; i < beams.size() ; ++i) {
+            common_prefix_length = std::min(common_prefix_length, beams[i].tokens.size());
+            for (size_t j = 0 ; j < common_prefix_length ; ++j) {
+                if (beams[0].tokens[j] != beams[i].tokens[j]) {
+                    common_prefix_length = j;
+                    break;
+                }
+            }
+        }
+        return common_prefix_length;
+    }
+
+    // Construct beams_state to send back to caller via the callback function.
+    // Side effect: set common_prefix_length = find_common_prefix_length();
+    llama_beams_state get_beams_state(const bool last_call) {
+        for (size_t i = 0 ; i < beams.size() ; ++i) {
+            beam_views[i] = beams[i].view();
+        }
+        common_prefix_length = find_common_prefix_length();
+        return {beam_views.data(), beams.size(), common_prefix_length, last_call};
+    }
+
+    // Loop:
+    //  * while i < n_predict, AND
+    //  * any of the beams have not yet reached end-of-beam (eob), AND
+    //  * the highest probability beam(s) (plural in case of ties) are not at end-of-sentence
+    //    (since all other beam probabilities can only decrease)
+    void loop(const llama_beam_search_callback_fn_t callback, void * const callback_data) {
+        beams.push_back({{}, 1.0f, false});  // Start with one empty beam w/ probability = 1.0 and !eob.
+        const auto not_eob = [](const llama_beam & beam) { return !beam.eob; };
+        for (int i = 0 ; i < n_predict && std::any_of(beams.begin(),beams.end(),not_eob) &&
+                       !beams[top_beam_index()].eob ; ++i) {
+            callback(callback_data, get_beams_state(false));  // Sets common_prefix_length
+            update_beams_from_beam_views();   // Update values (p,eob) that callback may have changed.
+            if (common_prefix_length) {
+                llama_eval(ctx, beams[0].tokens.data(), common_prefix_length, n_past, n_threads);
+                n_past += common_prefix_length;
+            }
+            // Zero-out next_beam probabilities to place them last in following min-heap.
+            std::for_each(next_beams.begin(), next_beams.end(), [](llama_beam & beam) { beam.p = 0.0f; });
+            for (llama_beam & beam : beams) {
+                beam.shift_tokens(common_prefix_length);
+                fill_next_beams_by_top_probabilities(beam);
+            }
+            // next_beams become the beams of next/final iteration. Swap them to re-use memory.
+            beams.swap(next_beams);
+            renormalize_beam_probabilities(beams);
+        }
+        collapse_beams(top_beam_index());
+        callback(callback_data, get_beams_state(true));
+    }
+
+    // As beams grow, the cumulative probabilities decrease.
+    // Renormalize them to avoid floating point underflow.
+    static void renormalize_beam_probabilities(std::vector<llama_beam> & beams) {
+        const auto sum_p = [](float sum, llama_beam & beam) { return sum + beam.p; };
+        const float inv_sum = 1.0f / std::accumulate(beams.begin(), beams.end(), 0.0f, sum_p);
+        std::for_each(beams.begin(), beams.end(), [=](llama_beam & beam) { beam.p *= inv_sum; });
+    }
+
+    // Assumes beams is non-empty.  Uses llama_beam::operator<() for ordering.
+    size_t top_beam_index() {
+        return std::max_element(beams.begin(), beams.end()) - beams.begin();
+    }
+
+    // Copy (p,eob) for each beam which may have been changed by the callback.
+    void update_beams_from_beam_views() {
+        for (size_t i = 0 ; i < beams.size() ; ++i) {
+            beams[i].p = beam_views[i].p;
+            beams[i].eob = beam_views[i].eob;
+        }
+    }
+};
+
+void llama_beam_search(llama_context * ctx,
+                       llama_beam_search_callback_fn_t callback, void * callback_data,
+                       size_t n_beams, int n_past, int n_predict, int n_threads) {
+    assert(ctx);
+    const int64_t t_start_sample_us = ggml_time_us();
+
+    llama_beam_search_data beam_search_data(ctx, n_beams, n_past, n_predict, n_threads);
+
+    beam_search_data.loop(callback, callback_data);
+
+    ctx->t_sample_us += ggml_time_us() - t_start_sample_us;
+    ctx->n_sample++;
+}
+
 //
 // quantization
 //
@@ -4423,6 +4675,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
 
     std::unique_ptr<llama_model_loader> ml(new llama_model_loader(fname_inp, /*use_mmap*/ false));
 
+    llama_model model;
+    llm_load_arch(*ml, model);
+    llm_load_hparams(*ml, model, 0, 0, 0);
+
     const size_t align = GGUF_DEFAULT_ALIGNMENT;
     struct gguf_context * ctx_out = gguf_init_empty();
 
@@ -4448,6 +4704,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
             ++n_feed_forward_w2;
         }
     }
+    if (n_attention_wv != n_feed_forward_w2 || (uint32_t)n_attention_wv != model.hparams.n_layer) {
+        LLAMA_LOG_WARN("%s ============ Strange model: n_attention_wv = %d, n_feed_forward_w2 = %d, hparams.n_layer = %d\n",
+                __func__, n_attention_wv, n_feed_forward_w2, model.hparams.n_layer);
+    }
 
     int i_attention_wv = 0;
     int i_feed_forward_w2 = 0;
@@ -4524,8 +4784,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
 
             if (name == tn(LLM_TENSOR_OUTPUT, "weight")) {
                 int nx = tensor->ne[0];
-                int ny = tensor->ne[1];
+                if (model.arch == LLM_ARCH_FALCON || nx % QK_K != 0) {
-                if (nx % QK_K == 0 && ny % QK_K == 0) {
+                    new_type = GGML_TYPE_Q8_0;
+                }
+                else if (new_type != GGML_TYPE_Q8_0) {
                     new_type = GGML_TYPE_Q6_K;
                 }
             } else if (name.find("attn_v.weight") != std::string::npos) {
@@ -4539,21 +4801,49 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
                 else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && i_attention_wv < 4) new_type = GGML_TYPE_Q5_K;
                 else if (QK_K == 64 && (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_S) &&
                         (i_attention_wv < n_attention_wv/8 || i_attention_wv >= 7*n_attention_wv/8)) new_type = GGML_TYPE_Q6_K;
+                if (model.type == MODEL_70B) {
+                    // In the 70B model we have 8 heads sharing the same attn_v weights. As a result, the attn_v.weight tensor is
+                    // 8x smaller compared to attn_q.weight. Hence, we can get a nice boost in quantization accuracy with
+                    // nearly negligible increase in model size by quantizing this tensor with more bits:
+                    if (new_type == GGML_TYPE_Q3_K || new_type == GGML_TYPE_Q4_K) new_type = GGML_TYPE_Q5_K;
+                }
                 ++i_attention_wv;
             } else if (name.find("ffn_down.weight") != std::string::npos) {
                 if      (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
                 else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) {
-                    new_type = i_feed_forward_w2 < 2 ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K;
+                    new_type = i_feed_forward_w2 < 2 ? GGML_TYPE_Q5_K
+                             : model.arch != LLM_ARCH_FALCON || use_more_bits(i_feed_forward_w2, n_feed_forward_w2) ? GGML_TYPE_Q4_K
+                             : GGML_TYPE_Q3_K;
+                }
+                else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) {
+                    new_type = model.arch == LLM_ARCH_FALCON ? GGML_TYPE_Q4_K : GGML_TYPE_Q5_K;
+                }
+                else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) {
+                    if (model.arch == LLM_ARCH_FALCON) {
+                        new_type = i_feed_forward_w2 < 2 ? GGML_TYPE_Q6_K :
+                                   use_more_bits(i_feed_forward_w2, n_feed_forward_w2) ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K;
+                    } else {
+                        if (use_more_bits(i_feed_forward_w2, n_feed_forward_w2)) new_type = GGML_TYPE_Q6_K;
+                    }
+                }
+                else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M && use_more_bits(i_feed_forward_w2, n_feed_forward_w2)) new_type = GGML_TYPE_Q6_K;
+                else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && model.arch != LLM_ARCH_FALCON && i_feed_forward_w2 < 4) {
+                    new_type = GGML_TYPE_Q5_K;
                 }
-                else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K;
-                else if ((ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) &&
-                         use_more_bits(i_feed_forward_w2, n_feed_forward_w2)) new_type = GGML_TYPE_Q6_K;
-                else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && i_feed_forward_w2 < 4) new_type = GGML_TYPE_Q5_K;
                 ++i_feed_forward_w2;
             } else if (name.find("attn_output.weight") != std::string::npos) {
-                if      (ftype == LLAMA_FTYPE_MOSTLY_Q2_K  ) new_type = GGML_TYPE_Q3_K;
+                if (model.arch != LLM_ARCH_FALCON) {
-                else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) new_type = GGML_TYPE_Q4_K;
+                    if      (ftype == LLAMA_FTYPE_MOSTLY_Q2_K  ) new_type = GGML_TYPE_Q3_K;
-                else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K;
+                    else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) new_type = GGML_TYPE_Q4_K;
+                    else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K;
+                } else {
+                    if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q4_K;
+                }
+            }
+            else if (name.find("attn_qkv.weight") != std::string::npos) {
+                if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q4_K;
+                else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) new_type = GGML_TYPE_Q5_K;
+                else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) new_type = GGML_TYPE_Q6_K;
             }
             else if (name.find("ffn_gate.weight") != std::string::npos || name.find("ffn_up.weight") != std::string::npos) {
                 if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
@@ -4568,8 +4858,8 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
                 new_type == GGML_TYPE_Q5_K || new_type == GGML_TYPE_Q6_K) {
                 int nx = tensor->ne[0];
                 int ny = tensor->ne[1];
-                if (nx % QK_K != 0 || ny % QK_K != 0) {
+                if (nx % QK_K != 0) {
-                    LLAMA_LOG_INFO("\n\nTensor sizes %d x %d are not divisible by %d, required for k-quants.\n",nx,ny,QK_K);
+                    LLAMA_LOG_WARN("\n\n%s : tensor cols %d x %d are not divisible by %d, required for k-quants\n", __func__, nx, ny, QK_K);
                     convert_incompatible_tensor = true;
                 }
             }
@@ -5297,13 +5587,29 @@ int llama_model_n_embd(const struct llama_model * model) {
     return model->hparams.n_embd;
 }
 
-int llama_model_type(const struct llama_model * model, char * buf, size_t buf_size) {
+int llama_model_desc(const struct llama_model * model, char * buf, size_t buf_size) {
     return snprintf(buf, buf_size, "%s %s %s",
             model->name.c_str(),
             llama_model_type_name(model->type),
             llama_model_ftype_name(model->ftype).c_str());
 }
 
+uint64_t llama_model_size(const struct llama_model * model) {
+    uint64_t size = 0;
+    for (const auto & it : model->tensors_by_name) {
+        size += ggml_nbytes(it.second);
+    }
+    return size;
+}
+
+uint64_t llama_model_n_params(const struct llama_model * model) {
+    uint64_t nparams = 0;
+    for (const auto & it : model->tensors_by_name) {
+        nparams += ggml_nelements(it.second);
+    }
+    return nparams;
+}
+
 int llama_model_quantize(
         const char * fname_inp,
         const char * fname_out,
@@ -5828,8 +6134,7 @@ int llama_tokenize_with_model(
                  llama_token * tokens,
                          int   n_max_tokens,
                         bool   add_bos) {
-    auto escape = llama_vocab_get_type(model->vocab) == LLAMA_VOCAB_TYPE_SPM;
+    auto res = llama_tokenize_internal(model->vocab, text, add_bos);
-    auto res = llama_tokenize_internal(model->vocab, text, add_bos, escape);
 
     if (n_max_tokens < (int) res.size()) {
         LLAMA_LOG_ERROR("%s: too many tokens\n", __func__);
@@ -5843,12 +6148,12 @@ int llama_tokenize_with_model(
     return res.size();
 }
 
-int llama_token_to_str(const struct llama_context * ctx, llama_token token, char * buf, int length) {
+int llama_token_to_piece(const struct llama_context * ctx, llama_token token, char * buf, int length) {
-    return llama_token_to_str_with_model(&ctx->model, token, buf, length);
+    return llama_token_to_piece_with_model(&ctx->model, token, buf, length);
 }
 
-// does not write null-terminator to str
+// does not write null-terminator to buf
-int llama_token_to_str_with_model(const struct llama_model * model, llama_token token, char * buf, int length) {
+int llama_token_to_piece_with_model(const struct llama_model * model, llama_token token, char * buf, int length) {
     if (0 <= token && token < llama_model_n_vocab(model)) {
         if (llama_is_normal_token(model->vocab, token)) {
             std::string result = model->vocab.id_to_token[token].text;
@@ -5936,6 +6241,7 @@ const char * llama_print_system_info(void) {
     s += "WASM_SIMD = "   + std::to_string(ggml_cpu_has_wasm_simd())   + " | ";
     s += "BLAS = "        + std::to_string(ggml_cpu_has_blas())        + " | ";
     s += "SSE3 = "        + std::to_string(ggml_cpu_has_sse3())        + " | ";
+    s += "SSSE3 = "       + std::to_string(ggml_cpu_has_ssse3())       + " | ";
     s += "VSX = "         + std::to_string(ggml_cpu_has_vsx())         + " | ";
 
     return s.c_str();

llama.h

@@ -254,7 +254,11 @@ extern "C" {
     LLAMA_API int llama_model_n_embd (const struct llama_model * model);
 
     // Get a string describing the model type
-    LLAMA_API int llama_model_type(const struct llama_model * model, char * buf, size_t buf_size);
+    LLAMA_API int llama_model_desc(const struct llama_model * model, char * buf, size_t buf_size);
+    // Returns the total size of all the tensors in the model in bytes
+    LLAMA_API uint64_t llama_model_size(const struct llama_model * model);
+    // Returns the total number of parameters in the model
+    LLAMA_API uint64_t llama_model_n_params(const struct llama_model * model);
 
     // Returns 0 on success
     LLAMA_API int llama_model_quantize(
@@ -377,15 +381,17 @@ extern "C" {
                              int   n_max_tokens,
                             bool   add_bos);
 
-    // Token Id -> String. Uses the vocabulary in the provided context
+    // Token Id -> Piece.
-    // Does not write null terminator to the buffer
+    // Uses the vocabulary in the provided context.
-    LLAMA_API int llama_token_to_str(
+    // Does not write null terminator to the buffer.
+    // User code is responsible to remove the leading whitespace of the first non-BOS token when decoding multiple tokens.
+    LLAMA_API int llama_token_to_piece(
             const struct llama_context * ctx,
                            llama_token   token,
                                   char * buf,
                                   int    length);
 
-    LLAMA_API int llama_token_to_str_with_model(
+    LLAMA_API int llama_token_to_piece_with_model(
               const struct llama_model * model,
                            llama_token   token,
                                   char * buf,
@@ -465,6 +471,43 @@ extern "C" {
     /// @details Accepts the sampled token into the grammar
     LLAMA_API void llama_grammar_accept_token(struct llama_context * ctx, struct llama_grammar * grammar, llama_token token);
 
+    //
+    // Beam search
+    //
+
+    struct llama_beam_view {
+        const llama_token * tokens;
+        size_t n_tokens;
+        float p;   // Cumulative beam probability (renormalized relative to all beams)
+        bool eob;  // Callback should set this to true when a beam is at end-of-beam.
+    };
+
+    // Passed to beam_search_callback function.
+    // Whenever 0 < common_prefix_length, this number of tokens should be copied from any of the beams
+    // (e.g. beams[0]) as they will be removed (shifted) from all beams in all subsequent callbacks.
+    // These pointers are valid only during the synchronous callback, so should not be saved.
+    struct llama_beams_state {
+        struct llama_beam_view * beam_views;
+        size_t n_beams;               // Number of elements in beam_views[].
+        size_t common_prefix_length;  // Current max length of prefix tokens shared by all beams.
+        bool last_call;               // True iff this is the last callback invocation.
+    };
+
+    // Type of pointer to the beam_search_callback function.
+    // void* callback_data is any custom data passed to llama_beam_search, that is subsequently
+    // passed back to beam_search_callback. This avoids having to use global variables in the callback.
+    typedef void (*llama_beam_search_callback_fn_t)(void * callback_data, llama_beams_state);
+
+    /// @details Deterministically returns entire sentence constructed by a beam search.
+    /// @param ctx Pointer to the llama_context.
+    /// @param callback Invoked for each iteration of the beam_search loop, passing in beams_state.
+    /// @param callback_data A pointer that is simply passed back to callback.
+    /// @param n_beams Number of beams to use.
+    /// @param n_past Number of tokens already evaluated.
+    /// @param n_predict Maximum number of tokens to predict. EOS may occur earlier.
+    /// @param n_threads Number of threads as passed to llama_eval().
+    LLAMA_API void llama_beam_search(struct llama_context * ctx, llama_beam_search_callback_fn_t callback, void * callback_data, size_t n_beams, int n_past, int n_predict, int n_threads);
+
     // Performance information
     LLAMA_API struct llama_timings llama_get_timings(struct llama_context * ctx);
     LLAMA_API void llama_print_timings(struct llama_context * ctx);

scripts/convert-gg.sh

@@ -0,0 +1,26 @@
+#!/bin/bash
+
+set -e
+
+# LLaMA v1
+python3 convert.py ../llama1/7B  --outfile models/llama-7b/ggml-model-f16.gguf  --outtype f16
+python3 convert.py ../llama1/13B --outfile models/llama-13b/ggml-model-f16.gguf --outtype f16
+python3 convert.py ../llama1/30B --outfile models/llama-30b/ggml-model-f16.gguf --outtype f16
+python3 convert.py ../llama1/65B --outfile models/llama-65b/ggml-model-f16.gguf --outtype f16
+
+# LLaMA v2
+python3 convert.py ../llama2/llama-2-7b  --outfile models/llama-7b-v2/ggml-model-f16.gguf  --outtype f16
+python3 convert.py ../llama2/llama-2-13b --outfile models/llama-13b-v2/ggml-model-f16.gguf --outtype f16
+python3 convert.py ../llama2/llama-2-70b --outfile models/llama-70b-v2/ggml-model-f16.gguf --outtype f16
+
+# Code Llama
+python3 convert.py ../codellama/CodeLlama-7b/  --outfile models/codellama-7b/ggml-model-f16.gguf  --outtype f16
+python3 convert.py ../codellama/CodeLlama-13b/ --outfile models/codellama-13b/ggml-model-f16.gguf --outtype f16
+python3 convert.py ../codellama/CodeLlama-34b/ --outfile models/codellama-34b/ggml-model-f16.gguf --outtype f16
+
+# Falcon
+python3 convert-falcon-hf-to-gguf.py ../falcon/falcon-7b  1
+mv -v ../falcon/falcon-7b/ggml-model-f16.gguf models/falcon-7b/ggml-model-f16.gguf
+
+python3 convert-falcon-hf-to-gguf.py ../falcon/falcon-40b 1
+mv -v ../falcon/falcon-40b/ggml-model-f16.gguf models/falcon-40b/ggml-model-f16.gguf

scripts/qnt-all.sh

@@ -20,6 +20,8 @@ fi
 model="$1"
 out="../tmp/results-${model}"
 
+set -e
+
 mkdir -p ${out}
 
 for q in ${qnt[@]}; do

scripts/run-all-perf.sh

@@ -20,6 +20,8 @@ fi
 model="$1"
 out="../tmp/results-${model}"
 
+set -e
+
 mkdir -p ${out}
 
 mstr=""

scripts/run-all-ppl.sh

@@ -17,6 +17,8 @@ if [ ! -z "$3" ]; then
     args="$3"
 fi
 
+set -e
+
 model="$1"
 out="../tmp/results-${model}"
 

tests/CMakeLists.txt

@@ -25,8 +25,10 @@ endfunction()
 llama_build_and_test_executable(test-quantize-fns.cpp)
 llama_build_and_test_executable(test-quantize-perf.cpp)
 llama_build_and_test_executable(test-sampling.cpp)
-llama_build_executable(test-tokenizer-0.cpp)
+llama_build_executable(test-tokenizer-0-llama.cpp)
-llama_test_executable (test-tokenizer-0.llama test-tokenizer-0.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama.gguf)
+llama_test_executable (test-tokenizer-0-llama test-tokenizer-0-llama.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama.gguf)
+llama_build_executable(test-tokenizer-0-falcon.cpp)
+#llama_test_executable (test-tokenizer-0-falcon test-tokenizer-0-falcon.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-falcon.gguf)
 llama_build_executable(test-tokenizer-1.cpp)
 # test-tokenizer-1 requires a BPE vocab. re-enable when we have one.
 #llama_test_executable (test-tokenizer-1.llama test-tokenizer-1.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-falcon.gguf)

tests/test-tokenizer-0-falcon.cpp

@@ -0,0 +1,178 @@
+#include "llama.h"
+#include "common.h"
+
+#include <cstdio>
+#include <string>
+#include <map>
+#include <vector>
+#include <fstream>
+
+// generate using test-tokenizer-0-falcon.py
+static const std::map<std::string, std::vector<llama_token>> & k_tests() {
+    static std::map<std::string, std::vector<llama_token>> _k_tests = {
+        { ""                      , {  }, },
+        { " "                     , {     204, }, },
+        { "  "                    , {     258, }, },
+        { "   "                   , {     466, }, },
+        { "\t"                    , {     192, }, },
+        { "\n"                    , {     193, }, },
+        { "\t\n"                  , {   19125, }, },
+        { "Hello world"           , {    9856,   1079, }, },
+        { " Hello world"          , {   23090,   1079, }, },
+        { "Hello World"           , {    9856,   2889, }, },
+        { " Hello World"          , {   23090,   2889, }, },
+        { " Hello World!"         , {   23090,   2889,     12, }, },
+        { "Hello, world!"         , {    9856,     23,   1079,     12, }, },
+        { " Hello, world!"        , {   23090,     23,   1079,     12, }, },
+        { " this is 🦙.cpp"        , {     414,    304,   3346,    111,    231,     25,  29247, }, },
+        { "w048 7tuijk dsdfhu"    , {      98,  55866,    204,     34,  16682,   7149,  36190,   6869,  11481, }, },
+        { "нещо на Български"     , {     150,    133,   6207,    151,    215,    150,    134,   5052,    133,   6279,   5052,    223,    151,    216,  49679,    123,  53110,  47043,   7795, }, },
+        { "កាន់តែពិសេសអាចខលចេញ"   , {   38154,    206,  38154,    126,  38154,    225,    167,    237,    217,  38154,    221,    167,    237,    208,  38154,    228,  38154,    127,  38154,    237,    167,    237,    207,  38154,    237,  38154,    107,  38154,    126,  38154,    211,  38154,    207,  38154,    233,  38154,    211,    167,    237,    207,  38154,    215, }, },
+        { "🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ (only emoji that has its own token)", {    2571,    232,    206,    204,     19,  11003,     20,   8196,    126,    283,    219,  48778,    116,  13392,    204,     19,  51831,    732,  63209,   1741,   7955,    522,     20,  22438,    211,    204,     19,   7927,  53360,    325,    504,    701,    946,  10930,     20, }, },
+        { "Hello"                 , {    9856, }, },
+        { " Hello"                , {   23090, }, },
+        { "  Hello"               , {     204,  23090, }, },
+        { "   Hello"              , {     258,  23090, }, },
+        { "    Hello"             , {     466,  23090, }, },
+        { "    Hello\n    Hello"  , {     466,  23090,    742,  23090, }, },
+    };
+
+    return _k_tests;
+}
+
+int main(int argc, char **argv) {
+    if (argc < 2) {
+        fprintf(stderr, "Usage: %s vocab-file [text-file]\n", argv[0]);
+        return 1;
+    }
+
+    const std::string fname = argv[1];
+
+    std::string fname_text;
+    if (argc > 2) {
+        fname_text = argv[2];
+    }
+
+    fprintf(stderr, "%s : reading vocab from: '%s'\n", __func__, fname.c_str());
+
+    llama_model * model;
+    llama_context * ctx;
+
+    llama_backend_init(false);
+
+    // load the vocab
+    {
+        auto lparams = llama_context_default_params();
+
+        lparams.vocab_only = true;
+
+        model = llama_load_model_from_file(fname.c_str(), lparams);
+
+        if (model == NULL) {
+            fprintf(stderr, "%s: error: failed to load vocab '%s'\n", __func__, fname.c_str());
+            return 1;
+        }
+
+        ctx = llama_new_context_with_model(model, lparams);
+
+        if (ctx == NULL) {
+            fprintf(stderr, "%s: error: failed to load vocab '%s'\n", __func__, fname.c_str());
+            llama_free_model(model);
+            return 1;
+        }
+    }
+
+    if (llama_vocab_type(ctx) != LLAMA_VOCAB_TYPE_BPE) {
+        fprintf(stderr, "%s : error: vocab type is not SPM\n", __func__);
+        llama_free_model(model);
+        llama_free(ctx);
+        return 2;
+    }
+
+    bool success = true;
+
+    for (const auto & test_kv : k_tests()) {
+        const std::vector<llama_token> res = llama_tokenize(ctx, test_kv.first, false);
+
+        printf("\n");
+        printf("src: '%s'\n", test_kv.first.c_str());
+        printf("res: '%s'\n", llama_detokenize_bpe(ctx, res).c_str());
+        printf("tok: ");
+        for (const auto & tok : res) {
+            printf("%d ", tok);
+        }
+        printf("\n");
+
+        bool correct = res.size() == test_kv.second.size();
+
+        for (int i = 0; i < (int) res.size() && correct; ++i) {
+            if (test_kv.second[i] != res[i]) {
+                correct = false;
+            }
+        }
+
+        if (!correct) {
+            fprintf(stderr, "%s : failed test:    '%s'\n", __func__, test_kv.first.c_str());
+            fprintf(stderr, "%s : detokenized to: '%s' instead of '%s'\n", __func__,
+                llama_detokenize_bpe(ctx, res).c_str(),
+                llama_detokenize_bpe(ctx, test_kv.second).c_str());
+            fprintf(stderr, "%s : expected tokens: ", __func__);
+            for (const auto & t : test_kv.second) {
+                fprintf(stderr, "%6d, ", t);
+            }
+            fprintf(stderr, "\n");
+            fprintf(stderr, "%s : got tokens:      ", __func__);
+            for (const auto & t : res) {
+                fprintf(stderr, "%6d, ", t);
+            }
+            fprintf(stderr, "\n");
+
+            success = false;
+        }
+    }
+
+    if (!fname_text.empty()) {
+        fprintf(stderr, "%s : tokenizing: '%s'\n", __func__, fname_text.c_str());
+
+        std::string text;
+        {
+            std::ifstream ifs(fname_text);
+            if (!ifs) {
+                fprintf(stderr, "%s : error: could not open file '%s'\n", __func__, fname_text.c_str());
+                return 1;
+            }
+            text = std::string(std::istreambuf_iterator<char>(ifs), std::istreambuf_iterator<char>());
+        }
+
+        fprintf(stderr, "%s : text size: %zu\n", __func__, text.size());
+
+        const std::vector<llama_token> res = llama_tokenize(ctx, text, true);
+
+        fprintf(stderr, "%s : tokens: %zu\n", __func__, res.size());
+
+        {
+            const std::string fname_out = fname_text + ".tokcpp";
+
+            std::ofstream ofs(fname_out);
+            if (!ofs) {
+                fprintf(stderr, "%s : error: could not open file '%s'\n", __func__, fname_out.c_str());
+                return 1;
+            }
+
+            for (const auto & tok : res) {
+                ofs << tok << " ";
+            }
+
+            ofs << "\n";
+        }
+
+        fprintf(stderr, "%s : tokens written to '%s'\n", __func__, (fname_text + ".tokcpp").c_str());
+    }
+
+    llama_free_model(model);
+    llama_free(ctx);
+
+    llama_backend_free();
+
+    return success ? 0 : 3;
+}

tests/test-tokenizer-0-falcon.py

@@ -0,0 +1,83 @@
+# tests with BPE tokenizer
+
+import os
+import sys
+import argparse
+
+from transformers import AutoTokenizer
+
+parser = argparse.ArgumentParser()
+parser.add_argument("dir_tokenizer", help="directory containing 'tokenizer.model' file")
+parser.add_argument("--fname-tok",   help="path to a text file to tokenize")
+args = parser.parse_args()
+
+dir_tokenizer = args.dir_tokenizer
+
+tokenizer = AutoTokenizer.from_pretrained(dir_tokenizer)
+
+tests = [
+        "",
+        " ",
+        "  ",
+        "   ",
+        "\t",
+        "\n",
+        "\t\n",
+        "Hello world",
+        " Hello world",
+        "Hello World",
+        " Hello World",
+        " Hello World!",
+        "Hello, world!",
+        " Hello, world!",
+        " this is 🦙.cpp",
+        "w048 7tuijk dsdfhu",
+        "нещо на Български",
+        "កាន់តែពិសេសអាចខលចេញ",
+        "🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ (only emoji that has its own token)",
+        "Hello",
+        " Hello",
+        "  Hello",
+        "   Hello",
+        "    Hello",
+        "    Hello\n    Hello",
+    ]
+
+for text in tests:
+    print('text: ', text)
+    print(tokenizer.encode(text))
+    print(tokenizer.decode(tokenizer.encode(text)))
+
+print("\n\ntests for C++:\n")
+for text in tests:
+    res = tokenizer.encode(text)
+
+    k = text.replace('\n', '\\n')
+    k = k.replace('\t', '\\t')
+    k = '"' + k + '"'
+    print("{ %-24s, { " % k, end='')
+    for x in res:
+        print("%7d," % x, end='')
+    print(" }, },")
+
+print(tokenizer.encode('hello'))
+print(tokenizer.encode('world'))
+print(tokenizer.encode(' world'))
+print(tokenizer.encode('hello world'))
+
+fname_tok = args.fname_tok
+if fname_tok:
+    print('tokenizing file: ', fname_tok)
+    fname_out = fname_tok + '.tok'
+    with open(fname_tok, 'r') as f:
+        lines = f.readlines()
+        s = ''.join(lines)
+        res = tokenizer.encode(s)
+        # write to file
+        with open(fname_out, 'w') as f:
+            for x in res:
+                f.write(str(x) + ' ')
+            f.write('\n')
+        print('len(res): ', len(res))
+        print('len(lines): ', len(lines))
+    print('results written to: ', fname_out)

tests/test-tokenizer-0-llama.cpp

@@ -0,0 +1,182 @@
+#include "llama.h"
+#include "common.h"
+
+#include <cstdio>
+#include <string>
+#include <map>
+#include <vector>
+#include <fstream>
+
+// generate using test-tokenizer-0-llama.py
+static const std::map<std::string, std::vector<llama_token>> & k_tests() {
+    static std::map<std::string, std::vector<llama_token>> _k_tests = {
+        { ""                      , {  }, },
+        { " "                     , {     259, }, },
+        { "  "                    , {    1678, }, },
+        { "   "                   , {     268, }, },
+        { "\t"                    , {   29871,     12, }, },
+        { "\n"                    , {   29871,     13, }, },
+        { "\t\n"                  , {   29871,     12,     13, }, },
+        { "Hello world"           , {   15043,   3186, }, },
+        { " Hello world"          , {   29871,  15043,   3186, }, },
+        { "Hello World"           , {   15043,   2787, }, },
+        { " Hello World"          , {   29871,  15043,   2787, }, },
+        { " Hello World!"         , {   29871,  15043,   2787,  29991, }, },
+        { "Hello, world!"         , {   15043,  29892,   3186,  29991, }, },
+        { " Hello, world!"        , {   29871,  15043,  29892,   3186,  29991, }, },
+        { " this is 🦙.cpp"        , {   29871,    445,    338,  29871,    243,    162,    169,    156,  29889,   8223, }, },
+        { "w048 7tuijk dsdfhu"    , {     281,  29900,  29946,  29947,  29871,  29955,   9161,  13535,  18031,   2176,   6905, }, },
+        { "нещо на Български"     , {    1538,   4851,    665,   1386,  29713,   1305, }, },
+        { "កាន់តែពិសេសអាចខលចេញ"   , {   29871,  31849,  31324,  31934,    228,    162,    142,    228,    161,    146,    228,    162,    133,    228,    161,    153,    228,    161,    186,  31708,    228,    162,    132,  31708,    228,    161,    165,  31324,    228,    161,    136,    228,    161,    132,    228,    161,    158,    228,    161,    136,    228,    162,    132,    228,    161,    140, }, },
+        { "🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ (only emoji that has its own token)", {   29871,    243,    162,    157,    131,    313,   8945,  29897,  29871,    243,    162,    155,    185,  30722,    243,    162,    143,    174,  30598,    313,  20787,    953,   3848,    275,  16125,    630,  29897,  29871,  31681,    313,   6194,    953,  29877,   2397,    393,    756,    967,   1914,   5993,  29897, }, },
+        { "Hello"                 , {   15043, }, },
+        { " Hello"                , {   29871,  15043, }, },
+        { "  Hello"               , {     259,  15043, }, },
+        { "   Hello"              , {    1678,  15043, }, },
+        { "    Hello"             , {     268,  15043, }, },
+        { "    Hello\n    Hello"  , {     268,  15043,     13,   1678,  15043, }, },
+    };
+
+    return _k_tests;
+}
+
+int main(int argc, char **argv) {
+    if (argc < 2) {
+        fprintf(stderr, "Usage: %s vocab-file [text-file]\n", argv[0]);
+        return 1;
+    }
+
+    const std::string fname = argv[1];
+
+    std::string fname_text;
+    if (argc > 2) {
+        fname_text = argv[2];
+    }
+
+    fprintf(stderr, "%s : reading vocab from: '%s'\n", __func__, fname.c_str());
+
+    llama_model * model;
+    llama_context * ctx;
+
+    llama_backend_init(false);
+
+    // load the vocab
+    {
+        auto lparams = llama_context_default_params();
+
+        lparams.vocab_only = true;
+
+        model = llama_load_model_from_file(fname.c_str(), lparams);
+
+        if (model == NULL) {
+            fprintf(stderr, "%s: error: failed to load vocab '%s'\n", __func__, fname.c_str());
+            return 1;
+        }
+
+        ctx = llama_new_context_with_model(model, lparams);
+
+        if (ctx == NULL) {
+            fprintf(stderr, "%s: error: failed to load vocab '%s'\n", __func__, fname.c_str());
+            llama_free_model(model);
+            return 1;
+        }
+    }
+
+    if (llama_vocab_type(ctx) != LLAMA_VOCAB_TYPE_SPM) {
+        fprintf(stderr, "%s : error: vocab type is not SPM\n", __func__);
+        llama_free_model(model);
+        llama_free(ctx);
+        return 2;
+    }
+
+    bool success = true;
+
+    for (const auto & test_kv : k_tests()) {
+        const std::vector<llama_token> res_bos   = llama_tokenize(ctx, test_kv.first, true);
+        const std::vector<llama_token> res_nobos = llama_tokenize(ctx, test_kv.first, false);
+
+        printf("\n");
+        printf("src: '%s'\n", test_kv.first.c_str());
+        printf("res: '%s'\n", llama_detokenize_spm(ctx, res_bos).c_str());
+        printf("tok: ");
+        for (const auto & tok : res_bos) {
+            printf("%d ", tok);
+        }
+        printf("\n");
+
+        bool correct = res_nobos.size() == test_kv.second.size() && res_bos.size() == res_nobos.size() + 1 && res_bos[0] == 1;
+
+        for (int i = 0; i < (int) res_nobos.size() && correct; ++i) {
+            if (test_kv.second[i] != res_bos[i + 1]) {
+                correct = false;
+            }
+            if (test_kv.second[i] != res_nobos[i]) {
+                correct = false;
+            }
+        }
+
+        if (!correct) {
+            fprintf(stderr, "%s : failed test:    '%s'\n", __func__, test_kv.first.c_str());
+            fprintf(stderr, "%s : detokenized to: '%s' instead of '%s'\n", __func__,
+                llama_detokenize_spm(ctx, res_nobos).c_str(),
+                llama_detokenize_spm(ctx, test_kv.second).c_str());
+            fprintf(stderr, "%s : expected tokens: ", __func__);
+            for (const auto & t : test_kv.second) {
+                fprintf(stderr, "%6d, ", t);
+            }
+            fprintf(stderr, "\n");
+            fprintf(stderr, "%s : got tokens:      ", __func__);
+            for (const auto & t : res_nobos) {
+                fprintf(stderr, "%6d, ", t);
+            }
+            fprintf(stderr, "\n");
+
+            success = false;
+        }
+    }
+
+    if (!fname_text.empty()) {
+        fprintf(stderr, "%s : tokenizing: '%s'\n", __func__, fname_text.c_str());
+
+        std::string text;
+        {
+            std::ifstream ifs(fname_text);
+            if (!ifs) {
+                fprintf(stderr, "%s : error: could not open file '%s'\n", __func__, fname_text.c_str());
+                return 1;
+            }
+            text = std::string(std::istreambuf_iterator<char>(ifs), std::istreambuf_iterator<char>());
+        }
+
+        fprintf(stderr, "%s : text size: %zu\n", __func__, text.size());
+
+        const std::vector<llama_token> res = llama_tokenize(ctx, text, true);
+
+        fprintf(stderr, "%s : tokens: %zu\n", __func__, res.size());
+
+        {
+            const std::string fname_out = fname_text + ".tokcpp";
+
+            std::ofstream ofs(fname_out);
+            if (!ofs) {
+                fprintf(stderr, "%s : error: could not open file '%s'\n", __func__, fname_out.c_str());
+                return 1;
+            }
+
+            for (const auto & tok : res) {
+                ofs << tok << " ";
+            }
+
+            ofs << "\n";
+        }
+
+        fprintf(stderr, "%s : tokens written to '%s'\n", __func__, (fname_text + ".tokcpp").c_str());
+    }
+
+    llama_free_model(model);
+    llama_free(ctx);
+
+    llama_backend_free();
+
+    return success ? 0 : 3;
+}

tests/test-tokenizer-0-llama.py

@@ -0,0 +1,95 @@
+# tests with SPM tokenizer
+
+import os
+import sys
+import argparse
+
+from sentencepiece import SentencePieceProcessor
+
+parser = argparse.ArgumentParser()
+parser.add_argument("dir_tokenizer", help="directory containing 'tokenizer.model' file")
+parser.add_argument("--fname-tok",   help="path to a text file to tokenize")
+args = parser.parse_args()
+
+dir_tokenizer = args.dir_tokenizer
+
+tokenizer = SentencePieceProcessor(dir_tokenizer + '/tokenizer.model')
+
+tests = [
+        "",
+        " ",
+        "  ",
+        "   ",
+        "\t",
+        "\n",
+        "\t\n",
+        "Hello world",
+        " Hello world",
+        "Hello World",
+        " Hello World",
+        " Hello World!",
+        "Hello, world!",
+        " Hello, world!",
+        " this is 🦙.cpp",
+        "w048 7tuijk dsdfhu",
+        "нещо на Български",
+        "កាន់តែពិសេសអាចខលចេញ",
+        "🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ (only emoji that has its own token)",
+        "Hello",
+        " Hello",
+        "  Hello",
+        "   Hello",
+        "    Hello",
+        "    Hello\n    Hello",
+    ]
+
+
+for text in tests:
+    print('text: ', text)
+    print('\nwith bos:')
+    print(tokenizer.encode(text, add_bos=True))
+    print(tokenizer.decode(tokenizer.encode(text, add_bos=True)))
+    print('\nwithout bos:')
+    print(tokenizer.encode(text, add_bos=False))
+    print(tokenizer.decode(tokenizer.encode(text, add_bos=False)))
+
+print("'" + tokenizer.id_to_piece(15043) + "'") # '_Hello'
+print("'" + tokenizer.id_to_piece(29871) + "'") # '_'
+print("'" + tokenizer.decode([15043]) + "'")        # 'Hello'
+print("'" + tokenizer.decode([15043, 15043]) + "'") # 'Hello Hello'
+print("'" + tokenizer.decode([29871, 15043]) + "'")               # ' Hello'
+print("'" + tokenizer.decode([29871, 15043, 29871, 15043]) + "'") # ' Hello  Hello'
+
+print("\n\ntests for C++:\n")
+for text in tests:
+    res = tokenizer.encode(text, add_bos=False)
+
+    k = text.replace('\n', '\\n')
+    k = k.replace('\t', '\\t')
+    k = '"' + k + '"'
+    print("{ %-24s, { " % k, end='')
+    for x in res:
+        print("%7d," % x, end='')
+    print(" }, },")
+
+print(tokenizer.encode('hello'))
+print(tokenizer.encode('world'))
+print(tokenizer.encode(' world'))
+print(tokenizer.encode('hello world'))
+
+fname_tok = args.fname_tok
+if fname_tok:
+    print('tokenizing file: ', fname_tok)
+    fname_out = fname_tok + '.tok'
+    with open(fname_tok, 'r') as f:
+        lines = f.readlines()
+        s = ''.join(lines)
+        res = tokenizer.encode(s, add_bos=True)
+        # write to file
+        with open(fname_out, 'w') as f:
+            for x in res:
+                f.write(str(x) + ' ')
+            f.write('\n')
+        print('len(res): ', len(res))
+        print('len(lines): ', len(lines))
+    print('results written to: ', fname_out)

tests/test-tokenizer-0.cpp

@@ -1,140 +0,0 @@
-#include "llama.h"
-#include "common.h"
-
-#include <cstdio>
-#include <string>
-#include <map>
-#include <vector>
-
-static std::string unescape_whitespace(llama_context* ctx, const std::vector<llama_token>& tokens) {
-    std::string result;
-    for (size_t i = 0; i < tokens.size(); ++i) {
-        result += llama_token_to_str(ctx, tokens[i]);
-    }
-    return result;
-}
-
-static const std::map<std::string, std::vector<llama_token>> & k_tests() {
-    static std::map<std::string, std::vector<llama_token>> _k_tests = {
-        { " ",                      {1,    259, }, },
-        { "  ",                     { 1,    1678, }, },
-        { "   ",                    { 1,     268, }, },
-        { "\t",                     { 1,    29871,   12, }, },
-        { "\n",                     { 1,    29871,   13, }, },
-        { "\t\n",                   { 1,    29871,   12,     13, }, },
-        { "Hello world",            { 1,  15043,   3186, }, },
-        { " Hello world",           { 1,  29871,  15043,   3186, }, },
-        { "Hello World",            { 1,  15043,   2787, }, },
-        { " Hello World",           { 1,  29871,  15043,   2787, }, },
-        { " Hello World!",          { 1,  29871,  15043,   2787,  29991, }, },
-        { " this is 🦙.cpp",        { 1,  29871,    445,    338,  29871,    243,    162,    169,    156,  29889,   8223, }, },
-        { "w048 7tuijk dsdfhu",     { 1,    281,  29900,  29946,  29947,  29871,  29955,   9161,  13535,  18031,   2176,   6905, }, },
-        { "нещо на Български",      { 1,   1538,   4851,    665,   1386,  29713,   1305, }, },
-        { "កាន់តែពិសេសអាចខលចេញ",   { 1,  29871,  31849,  31324,  31934,    228,    162,    142,    228,    161,
-                                     146,    228,    162,    133,    228,    161,    153,    228,    161,    186,
-                                     31708,    228,    162,    132,  31708,    228,    161,    165,  31324,    228,
-                                     161,    136,    228,    161,    132,    228,    161,    158,    228,    161,
-                                     136,    228,    162,    132,    228,    161,    140, }, },
-        { "🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ (only emoji that has its own token)",
-            { 1,  29871,    243,    162,    157,    131,    313,   8945,  29897,  29871,
-                243,    162,    155,    185,  30722,    243,    162,    143,    174,  30598,
-                313,  20787,    953,   3848,    275,  16125,    630,  29897,  29871,  31681,
-                313,   6194,    953,  29877,   2397,    393,    756,    967,   1914,   5993,  29897, }, },
-        { "Hello",                  { 1,    15043 }, },
-        { " Hello",                 { 1,    29871,  15043 }, },
-        { "  Hello",                { 1,    259,    15043 }, },
-        { "   Hello",               { 1,    1678,   15043 }, },
-        { "    Hello",              { 1,    268,    15043 }, },
-        { "    Hello\n    Hello",   { 1,    268,    15043,  13,     1678,   15043 }, },
-    };
-
-    return _k_tests;
-}
-
-int main(int argc, char **argv) {
-    if (argc < 2) {
-        fprintf(stderr, "Usage: %s <vocab-file>\n", argv[0]);
-        return 1;
-    }
-
-    const std::string fname = argv[1];
-
-    fprintf(stderr, "%s : reading vocab from: '%s'\n", __func__, fname.c_str());
-
-    llama_model * model;
-    llama_context * ctx;
-
-    llama_backend_init(false);
-
-    // load the vocab
-    {
-        auto lparams = llama_context_default_params();
-
-        lparams.vocab_only = true;
-
-        model = llama_load_model_from_file(fname.c_str(), lparams);
-
-        if (model == NULL) {
-            fprintf(stderr, "%s: error: failed to load vocab '%s'\n", __func__, fname.c_str());
-            return 1;
-        }
-
-        ctx = llama_new_context_with_model(model, lparams);
-
-        if (ctx == NULL) {
-            fprintf(stderr, "%s: error: failed to load vocab '%s'\n", __func__, fname.c_str());
-            llama_free_model(model);
-            return 1;
-        }
-    }
-
-    const int n_vocab = llama_n_vocab(ctx);
-
-    if (n_vocab != 32000) {
-        fprintf(stderr, "%s : expected 32000 tokens, got %d\n", __func__, n_vocab);
-        llama_free_model(model);
-        llama_free(ctx);
-        return 2;
-    }
-
-    bool success = true;
-
-    for (const auto & test_kv : k_tests()) {
-        std::vector<llama_token> res = llama_tokenize(ctx, test_kv.first, true);
-        fprintf(stderr, "%s : '%s' tokenized to '%s'\n",
-            __func__, test_kv.first.c_str(), unescape_whitespace(ctx, res).c_str());
-
-        bool correct = res.size() == test_kv.second.size();
-
-        for (int i = 0; i < (int) res.size() && correct; ++i) {
-            if (res[i] != test_kv.second[i]) {
-                correct = false;
-            }
-        }
-
-        if (!correct) {
-            fprintf(stderr, "%s : failed test:    '%s'\n", __func__, test_kv.first.c_str());
-            fprintf(stderr, "%s : detokenized to: '%s' instead of '%s'\n", __func__,
-                unescape_whitespace(ctx, res).c_str(), unescape_whitespace(ctx, test_kv.second).c_str());
-            fprintf(stderr, "%s : expected tokens: ", __func__);
-            for (const auto & t : test_kv.second) {
-                fprintf(stderr, "%6d, ", t);
-            }
-            fprintf(stderr, "\n");
-            fprintf(stderr, "%s : got tokens:      ", __func__);
-            for (const auto & t : res) {
-                fprintf(stderr, "%6d, ", t);
-            }
-            fprintf(stderr, "\n");
-
-            success = false;
-        }
-    }
-
-    llama_free_model(model);
-    llama_free(ctx);
-
-    llama_backend_free();
-
-    return success ? 0 : 3;
-}

tests/test-tokenizer-1.cpp

@@ -22,14 +22,6 @@ static std::string escape_whitespace(const std::string& text) {
     return result;
 }
 
-static std::string unescape_whitespace(llama_context * ctx, const std::vector<llama_token> & tokens) {
-    std::string result;
-    for (size_t i = 0; i < tokens.size(); ++i) {
-        result += llama_token_to_str(ctx, tokens[i]);
-    }
-    return result;
-}
-
 int main(int argc, char **argv) {
     if (argc < 2) {
         fprintf(stderr, "Usage: %s <vocab-file>\n", argv[0]);
@@ -72,13 +64,13 @@ int main(int argc, char **argv) {
     const int n_vocab = llama_n_vocab(ctx);
 
     for (int i = 0; i < n_vocab; ++i) {
-        std::string forward = llama_token_to_str(ctx, i);
+        std::string forward = llama_token_to_piece(ctx, i);
         std::vector<llama_token> tokens = llama_tokenize(ctx, forward, false);
         if (tokens.size() == 1) {
             if (i != tokens[0]) {
-                std::string backward = llama_token_to_str(ctx, tokens[0]);
+                std::string backward = llama_token_to_piece(ctx, tokens[0]);
                 fprintf(stderr, "%s : error: token %d is string %s but bpe returns token %d %s\n",
-                    __func__, i, llama_token_to_str(ctx, i).c_str(), tokens[0], backward.c_str());
+                    __func__, i, llama_token_to_piece(ctx, i).c_str(), tokens[0], backward.c_str());
                 return 2;
             }
         }