Merge branch 'ggerganov:master' into master

Author: apicalshark

README.md

@@ -131,6 +131,7 @@ Typically finetunes of the base models below are supported as well.
 - PHP (API bindings and features built on top of llama.cpp): [distantmagic/resonance](https://github.com/distantmagic/resonance) [(more info)](https://github.com/ggerganov/llama.cpp/pull/6326)
 - Guile Scheme: [guile_llama_cpp](https://savannah.nongnu.org/projects/guile-llama-cpp)
 - Swift [srgtuszy/llama-cpp-swift](https://github.com/srgtuszy/llama-cpp-swift)
+- Swift [ShenghaiWang/SwiftLlama](https://github.com/ShenghaiWang/SwiftLlama)
 
 **UI:**
 

examples/main/README.md

@@ -297,10 +297,6 @@ These options help improve the performance and memory usage of the LLaMA models.
 
  These flags attempt optimizations that help on some systems with non-uniform memory access. This currently consists of one of the above strategies, and disabling prefetch and readahead for mmap. The latter causes mapped pages to be faulted in on first access instead of all at once, and in combination with pinning threads to NUMA nodes, more of the pages end up on the NUMA node where they are used. Note that if the model is already in the system page cache, for example because of a previous run without this option, this will have little effect unless you drop the page cache first. This can be done by rebooting the system or on Linux by writing '3' to '/proc/sys/vm/drop_caches' as root.
 
-### Memory Float 32
-
--   `--memory-f32`: Use 32-bit floats instead of 16-bit floats for memory key+value. This doubles the context memory requirement and cached prompt file size but does not appear to increase generation quality in a measurable way. Not recommended.
-
 ### Batch Size
 
 -   `-b N, --batch-size N`: Set the batch size for prompt processing (default: `2048`). This large batch size benefits users who have BLAS installed and enabled it during the build. If you don't have BLAS enabled ("BLAS=0"), you can use a smaller number, such as 8, to see the prompt progress as it's evaluated in some situations.

include/llama.h

@@ -953,12 +953,6 @@ extern "C" {
                                int32_t   lstrip,
                                   bool   special);
 
-    // check if token0 is contained as a prefix in token1
-    LLAMA_API bool llama_token_is_prefix(
-              const struct llama_model * model,
-                           llama_token   token0,
-                           llama_token   token1);
-
     /// @details Convert the provided tokens into text (inverse of llama_tokenize()).
     /// @param text The char pointer must be large enough to hold the resulting text.
     /// @return Returns the number of chars/bytes on success, no more than text_len_max.

scripts/run-with-preset.py

@@ -15,7 +15,7 @@
     "export", "file", "frequency-penalty", "grammar", "grammar-file", "hellaswag",
     "hellaswag-tasks", "ignore-eos", "in-prefix", "in-prefix-bos", "in-suffix",
     "interactive", "interactive-first", "keep", "logdir", "logit-bias", "lora", "lora-base",
-    "low-vram", "main-gpu", "memory-f32", "mirostat", "mirostat-ent", "mirostat-lr", "mlock",
+    "low-vram", "main-gpu", "mirostat", "mirostat-ent", "mirostat-lr", "mlock",
     "model", "multiline-input", "n-gpu-layers", "n-predict", "no-mmap", "no-mul-mat-q",
     "np-penalize-nl", "numa", "ppl-output-type", "ppl-stride", "presence-penalty", "prompt",
     "prompt-cache", "prompt-cache-all", "prompt-cache-ro", "repeat-last-n",
@@ -25,12 +25,12 @@
 ]
 
 CLI_ARGS_LLAMA_BENCH = [
-    "batch-size", "memory-f32", "low-vram", "model", "mul-mat-q", "n-gen", "n-gpu-layers",
+    "batch-size", "low-vram", "model", "mul-mat-q", "n-gen", "n-gpu-layers",
     "n-prompt", "output", "repetitions", "tensor-split", "threads", "verbose"
 ]
 
 CLI_ARGS_LLAMA_SERVER = [
-    "alias", "batch-size", "ctx-size", "embedding", "host", "memory-f32", "lora", "lora-base",
+    "alias", "batch-size", "ctx-size", "embedding", "host", "lora", "lora-base",
     "low-vram", "main-gpu", "mlock", "model", "n-gpu-layers", "n-probs", "no-mmap", "no-mul-mat-q",
     "numa", "path", "port", "rope-freq-base", "timeout", "rope-freq-scale", "tensor-split",
     "threads", "verbose"

src/llama-sampling.cpp

@@ -1745,6 +1745,9 @@ struct llama_sampler * llama_sampler_init_logit_bias(
 
 struct llama_sampler_infill {
     const struct llama_vocab * vocab;
+
+    std::vector<char> buf0;
+    std::vector<char> buf1;
 };
 
 static const char * llama_sampler_infill_name(const struct llama_sampler * /*smpl*/) {
@@ -1810,27 +1813,44 @@ static void llama_sampler_infill_apply(struct llama_sampler * smpl, llama_token_
     size_t n_combined = 0; GGML_UNUSED(n_combined);
 
     // combine tokens with common prefix
-    for (size_t i = 0; i < cur_p->size; ++i) {
-        for (size_t j = 0; j < cur_p->size; ++j) {
-            if (cur_p->data[i].logit == -INFINITY) {
+    for (size_t i0 = 0; i0 < cur_p->size; ++i0) {
+        for (size_t i1 = 0; i1 < cur_p->size; ++i1) {
+            if (cur_p->data[i0].logit == -INFINITY) {
                 break;
             }
 
-            if (i == j || cur_p->data[j].logit == -INFINITY) {
+            if (i0 == i1 || cur_p->data[i1].logit == -INFINITY) {
                 continue;
             }
 
-            if (llama_token_is_prefix_impl(*ctx->vocab, cur_p->data[i].id, cur_p->data[j].id)) {
-                if (cur_p->data[i].p >  cur_p->data[j].p) {
-                    cur_p->data[i].p += cur_p->data[j].p;
-                    cur_p->data[j].logit = -INFINITY;
-                    cur_p->data[j].p     = 0.0f;
-                } else {
-                    cur_p->data[j].p += cur_p->data[i].p;
-                    cur_p->data[i].logit = -INFINITY;
-                    cur_p->data[i].p     = 0.0f;
+            int len0 = llama_token_to_piece_impl(*ctx->vocab, cur_p->data[i0].id, ctx->buf0.data(), ctx->buf0.size(), 0, false);
+            if (len0 < 0) {
+                ctx->buf0.resize(len0);
+                len0 = llama_token_to_piece_impl(*ctx->vocab, cur_p->data[i0].id, ctx->buf0.data(), ctx->buf0.size(), 0, false);
+                assert(len0 > 0);
+            }
+
+            int len1 = llama_token_to_piece_impl(*ctx->vocab, cur_p->data[i1].id, ctx->buf1.data(), ctx->buf1.size(), 0, false);
+            if (len1 < 0) {
+                ctx->buf1.resize(len1);
+                len1 = llama_token_to_piece_impl(*ctx->vocab, cur_p->data[i1].id, ctx->buf1.data(), ctx->buf1.size(), 0, false);
+                assert(len1 > 0);
+            }
+
+            // token i0 is a prefix of token i1
+            if (len0 > 0 && len0 <= len1 && memcmp(ctx->buf0.data(), ctx->buf1.data(), len0) == 0) {
+                int dst = i0;
+                int src = i1;
+
+                // merge into the token with higher probability
+                if (cur_p->data[i1].p > cur_p->data[i0].p) {
+                    std::swap(dst, src);
                 }
 
+                cur_p->data[dst].p += cur_p->data[src].p;
+                cur_p->data[src].logit = -INFINITY;
+                cur_p->data[src].p     = 0.0f;
+
                 n_combined++;
             }
         }
@@ -1936,6 +1956,8 @@ struct llama_sampler * llama_sampler_init_infill_impl(
         /* .iface = */ &llama_sampler_infill_i,
         /* .ctx   = */ new llama_sampler_infill {
             /* .vocab = */ &vocab,
+            /* .buf0 = */ std::vector<char>(512),
+            /* .buf1 = */ std::vector<char>(512),
         },
     };
 }

src/llama-vocab.cpp

@@ -1858,23 +1858,6 @@ int32_t llama_token_to_piece_impl(const struct llama_vocab & vocab, llama_token
     return 0;
 }
 
-bool llama_token_is_prefix_impl(
-        const struct llama_vocab & vocab,
-                     llama_token   token0,
-                     llama_token   token1) {
-    char text_buf_0[128];
-    char text_buf_1[128];
-
-    const int32_t len0 = llama_token_to_piece_impl(vocab, token0, text_buf_0, sizeof(text_buf_0) - 1, 0, false);
-    const int32_t len1 = llama_token_to_piece_impl(vocab, token1, text_buf_1, sizeof(text_buf_1) - 1, 0, false);
-
-    if (len0 <= 0 || len1 <= 0) {
-        return false;
-    }
-
-    return len0 <= len1 && memcmp(text_buf_0, text_buf_1, len0) == 0;
-}
-
 int32_t llama_detokenize_impl(
         const struct llama_vocab & vocab,
                const llama_token * tokens,

src/llama.cpp

@@ -6735,9 +6735,9 @@ static void llm_load_vocab(
                     vocab.id_to_token[t.second].attr = LLAMA_TOKEN_ATTR_CONTROL;
                 }
             } else {
-                // token is control, but not marked as EOG -> print a warning
+                // token is control, but not marked as EOG -> print a debug log
                 if (vocab.id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL && vocab.special_eog_ids.count(t.second) == 0) {
-                    LLAMA_LOG_WARN("%s: control token: %6d '%s' is not marked as EOG\n",
+                    LLAMA_LOG_DEBUG("%s: control token: %6d '%s' is not marked as EOG\n",
                             __func__, t.second, t.first.c_str());
                 }
             }
@@ -17134,10 +17134,10 @@ static void llama_graph_compute(
 //
 static int llama_decode_internal(
          llama_context & lctx,
-           llama_batch   batch_all) { // TODO: rename back to batch
+           llama_batch   batch) {
 
     lctx.is_encoding = false;
-    const uint32_t n_tokens_all = batch_all.n_tokens;
+    const uint32_t n_tokens_all = batch.n_tokens;
 
     if (n_tokens_all == 0) {
         LLAMA_LOG_ERROR("%s: n_tokens == 0\n", __func__);
@@ -17148,12 +17148,12 @@ static int llama_decode_internal(
     const auto & hparams = model.hparams;
     const auto & cparams = lctx.cparams;
 
-    GGML_ASSERT((!batch_all.token && batch_all.embd) || (batch_all.token && !batch_all.embd)); // NOLINT
+    GGML_ASSERT((!batch.token && batch.embd) || (batch.token && !batch.embd)); // NOLINT
 
-    if (batch_all.token) {
+    if (batch.token) {
         for (uint32_t i = 0; i < n_tokens_all; ++i) {
-            if (batch_all.token[i] < 0 || (uint32_t)batch_all.token[i] >= model.vocab.n_vocab) {
-                LLAMA_LOG_ERROR("%s: invalid token[%d] = %d\n", __func__, i, batch_all.token[i]);
+            if (batch.token[i] < 0 || (uint32_t)batch.token[i] >= model.vocab.n_vocab) {
+                LLAMA_LOG_ERROR("%s: invalid token[%d] = %d\n", __func__, i, batch.token[i]);
                 return -1;
             }
         }
@@ -17184,9 +17184,9 @@ static int llama_decode_internal(
     lctx.embd_seq.clear();
 
     // count outputs
-    if (batch_all.logits && !embd_pooled) {
+    if (batch.logits && !embd_pooled) {
         for (uint32_t i = 0; i < n_tokens_all; ++i) {
-            n_outputs += batch_all.logits[i] != 0;
+            n_outputs += batch.logits[i] != 0;
         }
     } else if (lctx.logits_all || embd_pooled) {
         n_outputs = n_tokens_all;
@@ -17195,7 +17195,7 @@ static int llama_decode_internal(
         n_outputs = 1;
     }
 
-    lctx.sbatch.from_batch(batch_all, n_embd,
+    lctx.sbatch.from_batch(batch, n_embd,
         /* simple_split */ !kv_self.recurrent,
         /* logits_all   */ n_outputs == n_tokens_all);
 
@@ -21466,13 +21466,6 @@ int32_t llama_token_to_piece(
     return llama_token_to_piece_impl(model->vocab, token, buf, length, lstrip, special);
 }
 
-bool llama_token_is_prefix(
-    const struct llama_model * model,
-                 llama_token   token0,
-                 llama_token   token1) {
-    return llama_token_is_prefix_impl(model->vocab, token0, token1);
-}
-
 int32_t llama_detokenize(
     const struct llama_model * model,
            const llama_token * tokens,