Merge branch 'master' into dynamic-temp

.github/workflows/build.yml

@@ -515,6 +515,31 @@ jobs:
       - name: Build Xcode project
         run: xcodebuild -project examples/llama.swiftui/llama.swiftui.xcodeproj -scheme llama.swiftui -sdk iphoneos CODE_SIGNING_REQUIRED=NO CODE_SIGN_IDENTITY= -destination 'generic/platform=iOS' build
 
+  android-build:
+    runs-on: ubuntu-latest
+
+    steps:
+      - name: Clone
+        uses: actions/checkout@v3
+
+      - name: Set up JDK
+        uses: actions/setup-java@v3
+        with:
+          java-version: 17
+          distribution: zulu
+
+      - name: Setup Android SDK
+        uses: android-actions/setup-android@v3
+        with:
+          log-accepted-android-sdk-licenses: false
+
+      - name: Build
+        run: |
+          cd examples/llama.android
+
+          # Skip armeabi-v7a for now (https://github.com/llvm/llvm-project/issues/65820).
+          ./gradlew build --no-daemon -Pskip-armeabi-v7a
+
 #  freeBSD-latest:
 #    runs-on: macos-12
 #    steps:

.gitignore

@@ -105,3 +105,4 @@ poetry.toml
 /tests/test-tokenizer-1-bpe
 /tests/test-rope
 /tests/test-backend-ops
+/tests/test-autorelease

Makefile

@@ -9,7 +9,7 @@ TEST_TARGETS = \
 	tests/test-llama-grammar tests/test-grammar-parser tests/test-double-float tests/test-grad0 tests/test-opt \
 	tests/test-quantize-fns tests/test-quantize-perf tests/test-sampling tests/test-tokenizer-0-llama          \
 	tests/test-tokenizer-0-falcon tests/test-tokenizer-1-llama tests/test-tokenizer-1-bpe tests/test-rope      \
-	tests/test-backend-ops
+	tests/test-backend-ops tests/test-autorelease
 
 # Code coverage output files
 COV_TARGETS = *.gcno tests/*.gcno *.gcda tests/*.gcda *.gcov tests/*.gcov lcov-report gcovr-report
@@ -747,3 +747,6 @@ tests/test-c.o: tests/test-c.c llama.h
 
 tests/test-backend-ops: tests/test-backend-ops.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
+tests/test-autorelease: tests/test-autorelease.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

ci/run.sh

@@ -36,6 +36,10 @@ if [ ! -z ${GG_BUILD_METAL} ]; then
     CMAKE_EXTRA="${CMAKE_EXTRA} -DLLAMA_METAL_SHADER_DEBUG=ON"
 fi
 
+if [ ! -z ${GG_BUILD_CUDA} ]; then
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DLLAMA_CUBLAS=1"
+fi
+
 ## helpers
 
 # download a file if it does not exist or if it is outdated
@@ -160,8 +164,8 @@ function gg_run_open_llama_3b_v2 {
 
     set -e
 
-    (time cmake -DCMAKE_BUILD_TYPE=Release -DLLAMA_QKK_64=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
-    (time make -j                                              ) 2>&1 | tee -a $OUT/${ci}-make.log
+    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} -DLLAMA_QKK_64=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
+    (time make -j                                                             ) 2>&1 | tee -a $OUT/${ci}-make.log
 
     python3 ../convert.py ${path_models}
 
@@ -179,6 +183,8 @@ function gg_run_open_llama_3b_v2 {
 
     wiki_test_60="${path_wiki}/wiki.test-60.raw"
 
+    ./bin/test-autorelease ${model_f16}
+
     ./bin/quantize ${model_f16} ${model_q8_0} q8_0
     ./bin/quantize ${model_f16} ${model_q4_0} q4_0
     ./bin/quantize ${model_f16} ${model_q4_1} q4_1
@@ -214,6 +220,8 @@ function gg_run_open_llama_3b_v2 {
     (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 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
 
+    (time ./bin/imatrix --model ${model_f16} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
+
     (time ./bin/save-load-state --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
 
     function check_ppl {
@@ -241,6 +249,8 @@ 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
 
+    cat $OUT/${ci}-imatrix.log | grep "Final" >> $OUT/${ci}-imatrix-sum.log
+
     # lora
     function compare_ppl {
         qnt="$1"
@@ -282,7 +292,6 @@ function gg_run_open_llama_3b_v2 {
     (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
 }
 
@@ -292,6 +301,7 @@ 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 '- imatrix:\n```\n%s\n```\n' "$(cat $OUT/${ci}-imatrix-sum.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)"
@@ -337,8 +347,8 @@ function gg_run_open_llama_7b_v2 {
 
     set -e
 
-    (time cmake -DCMAKE_BUILD_TYPE=Release -DLLAMA_CUBLAS=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
-    (time make -j                                              ) 2>&1 | tee -a $OUT/${ci}-make.log
+    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} -DLLAMA_CUBLAS=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
+    (time make -j                                                             ) 2>&1 | tee -a $OUT/${ci}-make.log
 
     python3 ../convert.py ${path_models}
 
@@ -391,6 +401,8 @@ function gg_run_open_llama_7b_v2 {
     (time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
     (time ./bin/perplexity --model ${model_q6_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
 
+    (time ./bin/imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
+
     (time ./bin/save-load-state --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
 
     function check_ppl {
@@ -418,6 +430,8 @@ 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
 
+    cat $OUT/${ci}-imatrix.log | grep "Final" >> $OUT/${ci}-imatrix-sum.log
+
     # lora
     function compare_ppl {
         qnt="$1"
@@ -469,6 +483,7 @@ function gg_sum_open_llama_7b_v2 {
     gg_printf 'OpenLLaMA 7B-v2:\n'
     gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
     gg_printf '- perplexity:\n%s\n' "$(cat $OUT/${ci}-ppl.log)"
+    gg_printf '- imatrix:\n```\n%s\n```\n' "$(cat $OUT/${ci}-imatrix-sum.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)"

common/common.cpp

@@ -167,6 +167,24 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
             if (params.n_threads_batch <= 0) {
                 params.n_threads_batch = std::thread::hardware_concurrency();
             }
+        } else if (arg == "-td" || arg == "--threads-draft") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.n_threads_draft = std::stoi(argv[i]);
+            if (params.n_threads_draft <= 0) {
+                params.n_threads_draft = std::thread::hardware_concurrency();
+            }
+        } else if (arg == "-tbd" || arg == "--threads-batch-draft") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.n_threads_batch_draft = std::stoi(argv[i]);
+            if (params.n_threads_batch_draft <= 0) {
+                params.n_threads_batch_draft = std::thread::hardware_concurrency();
+            }
         } else if (arg == "-p" || arg == "--prompt") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -845,6 +863,10 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("  -t N, --threads N     number of threads to use during generation (default: %d)\n", params.n_threads);
     printf("  -tb N, --threads-batch N\n");
     printf("                        number of threads to use during batch and prompt processing (default: same as --threads)\n");
+    printf("  -td N, --threads-draft N");
+    printf("                        number of threads to use during generation (default: same as --threads)");
+    printf("  -tbd N, --threads-batch-draft N\n");
+    printf("                        number of threads to use during batch and prompt processing (default: same as --threads-draft)\n");
     printf("  -p PROMPT, --prompt PROMPT\n");
     printf("                        prompt to start generation with (default: empty)\n");
     printf("  -e, --escape          process prompt escapes sequences (\\n, \\r, \\t, \\', \\\", \\\\)\n");

common/common.h

@@ -46,7 +46,9 @@ struct gpt_params {
     uint32_t seed                           = -1;    // RNG seed
 
     int32_t n_threads                       = get_num_physical_cores();
+    int32_t n_threads_draft                 = -1;
     int32_t n_threads_batch                 = -1;    // number of threads to use for batch processing (-1 = use n_threads)
+    int32_t n_threads_batch_draft           = -1;
     int32_t n_predict                       = -1;    // new tokens to predict
     int32_t n_ctx                           = 512;   // context size
     int32_t n_batch                         = 512;   // batch size for prompt processing (must be >=32 to use BLAS)

common/sampling.h

@@ -17,7 +17,7 @@ typedef struct llama_sampling_params {
     float       min_p                 = 0.05f;    // 0.0 = disabled
     float       tfs_z                 = 1.00f;    // 1.0 = disabled
     float       typical_p             = 1.00f;    // 1.0 = disabled
-    float       temp                  = 0.80f;    // 1.0 = disabled
+    float       temp                  = 0.80f;    // <= 0.0 to sample greedily, 0.0 to not output probabilities
     float       dynatemp_range        = 0.00f;    // 0.0 = disabled
     int32_t     penalty_last_n        = 64;       // last n tokens to penalize (0 = disable penalty, -1 = context size)
     float       penalty_repeat        = 1.10f;    // 1.0 = disabled

convert-hf-to-gguf.py

@@ -266,11 +266,10 @@ def _set_vocab_gpt2(self):
                 toktypes.append(gguf.TokenType.USER_DEFINED)
             elif reverse_vocab[i] in added_vocab:
                 tokens.append(reverse_vocab[i])
-                if hasattr(tokenizer, "added_tokens_decoder"):
-                    if tokenizer.added_tokens_decoder[i].special:
-                        toktypes.append(gguf.TokenType.CONTROL)
-                    else:
-                        toktypes.append(gguf.TokenType.USER_DEFINED)
+                if tokenizer.added_tokens_decoder[i].special:
+                    toktypes.append(gguf.TokenType.CONTROL)
+                else:
+                    toktypes.append(gguf.TokenType.USER_DEFINED)
             else:
                 tokens.append(reverse_vocab[i])
                 toktypes.append(gguf.TokenType.NORMAL)

convert.py

@@ -387,6 +387,7 @@ def __init__(
         self.bpe_tokenizer = json.loads(
             open(str(fname_tokenizer), encoding="utf-8").read()
         )
+        self.vocab = self.bpe_tokenizer["model"]["vocab"]
         added_tokens: dict[str, int]
         if fname_added_tokens is not None:
             # FIXME: Verify that added tokens here _cannot_ overlap with the main vocab.
@@ -405,7 +406,7 @@ def __init__(
                     if item["content"] not in self.bpe_tokenizer
                 )
 
-        vocab_size: int = len(self.bpe_tokenizer)
+        vocab_size: int = len(self.vocab)
         expected_ids = list(range(vocab_size, vocab_size + len(added_tokens)))
         actual_ids = sorted(added_tokens.values())
         if expected_ids != actual_ids:
@@ -415,17 +416,17 @@ def __init__(
             )
 
         items = sorted(added_tokens.items(), key=lambda text_idx: text_idx[1])
+        self.added_tokens_dict = added_tokens
         self.added_tokens_list = [text for (text, idx) in items]
         self.vocab_size_base: int = vocab_size
         self.vocab_size: int = self.vocab_size_base + len(self.added_tokens_list)
         self.fname_tokenizer = fname_tokenizer
         self.fname_added_tokens = fname_added_tokens
 
     def bpe_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
-        tokenizer = self.bpe_tokenizer
-        reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.items()}
+        reverse_vocab = {id: encoded_tok for encoded_tok, id in self.vocab.items()}
 
-        for i, _ in enumerate(tokenizer):
+        for i, _ in enumerate(self.vocab):
             yield reverse_vocab[i], 0.0, gguf.TokenType.NORMAL
 
     def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
@@ -466,6 +467,7 @@ def __init__(
             )
 
         # Token pieces that were added to the base vocabulary.
+        self.added_tokens_dict = added_tokens
         self.added_tokens_list = [new_tokens[id] for id in actual_new_ids]
         self.vocab_size_base = vocab_size
         self.vocab_size = self.vocab_size_base + len(self.added_tokens_list)
@@ -1006,6 +1008,7 @@ def check_vocab_size(params: Params, vocab: Vocab, pad_vocab: bool = False) -> N
         )
         for i in range(1, pad_count + 1):
             vocab.added_tokens_dict[f"<dummy{i:05}>"] = -1
+            vocab.added_tokens_list.append(f"<dummy{i:05}>")
         vocab.vocab_size = params.n_vocab
         return
 
@@ -1097,6 +1100,8 @@ def extract_vocabulary_from_model(self, vocab: Vocab) -> Tuple[list, list, list]
             scores.append(score)
             toktypes.append(toktype)
 
+        assert len(tokens) == vocab.vocab_size
+
         return tokens, scores, toktypes
 
     def add_meta_vocab(self, vocab: Vocab) -> None:
@@ -1373,15 +1378,14 @@ def _detect_files(self):
                 self.files[file] = file_path
             elif parent_file_path.exists():
                 self.files[file] = parent_file_path
+        print(f"Found vocab files: {self.files}")
 
     def _select_file(self, vocabtype: Optional[str]) -> Path:
         if vocabtype in ["spm", "bpe"]:
-            # For SentencePiece and BPE, return specific files as before
-            file_key = "tokenizer.model" if vocabtype == "spm" else "vocab.json"
-            if self.files[file_key]:
-                return self.files[file_key]
-            else:
-                raise FileNotFoundError(f"{vocabtype} {file_key} not found.")
+            for file_key in self.files.keys():
+                if self.files[file_key]:
+                    return self.files[file_key]
+            raise FileNotFoundError(f"{vocabtype} vocab not found.")
         elif vocabtype == "hfft":
             # For Hugging Face Fast Tokenizer, return the directory path instead of a specific file
             return self.path

examples/finetune/finetune.cpp

@@ -1138,9 +1138,8 @@ static void save_as_llama_lora(const char * filename, struct my_llama_lora * lor
         return tn_buf.data();
     };
 
-    uint32_t LLAMA_FILE_MAGIC_LORA = 0x67676C61; // 'ggla'
     // write_magic
-    file.write_u32(LLAMA_FILE_MAGIC_LORA);   // magic
+    file.write_u32(LLAMA_FILE_MAGIC_GGLA);   // magic
     file.write_u32(1); // version
     // write_hparams
     file.write_u32(lora->hparams.lora_r);
@@ -1800,7 +1799,7 @@ int main(int argc, char ** argv) {
     std::vector<llama_token> train_tokens;
     std::vector<size_t> train_samples_begin;
     std::vector<size_t> train_samples_size;
-    printf("%s: tokenize training data\n", __func__);
+    printf("%s: tokenize training data from %s\n", __func__, params.common.fn_train_data);
     tokenize_file(lctx,
             params.common.fn_train_data,
             params.common.sample_start,