diff --git a/.github/ISSUE_TEMPLATE/custom.md b/.github/ISSUE_TEMPLATE/bug.md
similarity index 96%
rename from .github/ISSUE_TEMPLATE/custom.md
rename to .github/ISSUE_TEMPLATE/bug.md
index 8fd955356..d7879b232 100644
--- a/.github/ISSUE_TEMPLATE/custom.md
+++ b/.github/ISSUE_TEMPLATE/bug.md
@@ -1,8 +1,7 @@
 ---
-name: Issue and enhancement template
-about: Used to report issues and request enhancements for llama.cpp
-title: "[User] Insert summary of your issue or enhancement.."
-labels: ''
+name: Bug template
+about: Used to report bugs in llama.cpp
+labels: ["bug"]
 assignees: ''
 
 ---
@@ -46,7 +45,7 @@ $ g++ --version
 
 # Failure Information (for bugs)
 
-Please help provide information about the failure if this is a bug. If it is not a bug, please remove the rest of this template.
+Please help provide information about the failure / bug.
 
 # Steps to Reproduce
 
diff --git a/.github/ISSUE_TEMPLATE/enhancement.md b/.github/ISSUE_TEMPLATE/enhancement.md
new file mode 100644
index 000000000..dcffda750
--- /dev/null
+++ b/.github/ISSUE_TEMPLATE/enhancement.md
@@ -0,0 +1,28 @@
+---
+name: Enhancement template
+about: Used to request enhancements for llama.cpp
+labels: ["enhancement"]
+assignees: ''
+
+---
+
+# Prerequisites
+
+Please answer the following questions for yourself before submitting an issue.
+
+- [ ] I am running the latest code. Development is very rapid so there are no tagged versions as of now.
+- [ ] I carefully followed the [README.md](https://github.com/ggerganov/llama.cpp/blob/master/README.md).
+- [ ] I [searched using keywords relevant to my issue](https://docs.github.com/en/issues/tracking-your-work-with-issues/filtering-and-searching-issues-and-pull-requests) to make sure that I am creating a new issue that is not already open (or closed).
+- [ ] I reviewed the [Discussions](https://github.com/ggerganov/llama.cpp/discussions), and have a new bug or useful enhancement to share.
+
+# Feature Description
+
+Please provide a detailed written description of what you were trying to do, and what you expected `llama.cpp` to do as an enhancement.
+
+# Motivation
+
+Please provide a detailed written description of reasons why this feature is necessary and how it is useful to `llama.cpp` users.
+
+# Possible Implementation
+
+If you have an idea as to how it can be implemented, please write a detailed description. Feel free to give links to external sources or share visuals that might be helpful to understand the details better.
diff --git a/CMakeLists.txt b/CMakeLists.txt
index 6af42a6c2..202f26049 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -331,6 +331,7 @@ if (LLAMA_CUBLAS)
             set(CMAKE_CUDA_ARCHITECTURES "60;61;70") # needed for f16 CUDA intrinsics
         else()
             set(CMAKE_CUDA_ARCHITECTURES "52;61;70") # lowest CUDA 12 standard + lowest for integer intrinsics
+            #set(CMAKE_CUDA_ARCHITECTURES "") # use this to compile much faster, but only F16 models work
         endif()
     endif()
     message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")
diff --git a/Makefile b/Makefile
index 705c4acb4..80179631f 100644
--- a/Makefile
+++ b/Makefile
@@ -391,12 +391,9 @@ else
 endif #LLAMA_CUDA_NVCC
 ifdef CUDA_DOCKER_ARCH
 	NVCCFLAGS += -Wno-deprecated-gpu-targets -arch=$(CUDA_DOCKER_ARCH)
-endif # CUDA_DOCKER_ARCH
-ifdef CUDA_NATIVE_ARCH
-	NVCCFLAGS += -arch=$(CUDA_NATIVE_ARCH)
 else
 	NVCCFLAGS += -arch=native
-endif # CUDA_NATIVE_ARCH
+endif # CUDA_DOCKER_ARCH
 ifdef LLAMA_CUDA_FORCE_DMMV
 	NVCCFLAGS += -DGGML_CUDA_FORCE_DMMV
 endif # LLAMA_CUDA_FORCE_DMMV
diff --git a/common/common.cpp b/common/common.cpp
index ac406ff72..73c2d4d42 100644
--- a/common/common.cpp
+++ b/common/common.cpp
@@ -933,13 +933,13 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
     }
 
     if (params.ignore_eos) {
-        params.sparams.logit_bias[llama_token_eos(lctx)] = -INFINITY;
+        params.sparams.logit_bias[llama_token_eos(model)] = -INFINITY;
     }
 
     {
         LOG("warming up the model with an empty run\n");
 
-        std::vector<llama_token> tmp = { llama_token_bos(lctx), llama_token_eos(lctx), };
+        std::vector<llama_token> tmp = { llama_token_bos(model), llama_token_eos(model), };
         llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch), 0, 0));
         llama_kv_cache_tokens_rm(lctx, -1, -1);
         llama_reset_timings(lctx);
@@ -994,7 +994,7 @@ std::string llama_token_to_piece(const struct llama_context * ctx, llama_token t
 }
 
 std::string llama_detokenize_spm(llama_context * ctx, const std::vector<llama_token> & tokens) {
-    const llama_token bos_id = llama_token_bos(ctx);
+    const llama_token bos_id = llama_token_bos(llama_get_model(ctx));
 
     std::string piece;
     std::string result;
@@ -1239,7 +1239,7 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     fprintf(stream, "hellaswag: %s # default: false\n", params.hellaswag ? "true" : "false");
     fprintf(stream, "hellaswag_tasks: %zu # default: 400\n", params.hellaswag_tasks);
 
-    const auto logit_bias_eos = sparams.logit_bias.find(llama_token_eos(lctx));
+    const auto logit_bias_eos = sparams.logit_bias.find(llama_token_eos(llama_get_model(lctx)));
     const bool ignore_eos = logit_bias_eos != sparams.logit_bias.end() && logit_bias_eos->second == -INFINITY;
     fprintf(stream, "ignore_eos: %s # default: false\n", ignore_eos ? "true" : "false");
 
diff --git a/common/sampling.cpp b/common/sampling.cpp
index 6f0af3c4a..5258d4e82 100644
--- a/common/sampling.cpp
+++ b/common/sampling.cpp
@@ -147,7 +147,7 @@ llama_token llama_sampling_sample(
 
     // apply penalties
     if (!prev.empty()) {
-        const float nl_logit = logits[llama_token_nl(ctx_main)];
+        const float nl_logit = logits[llama_token_nl(llama_get_model(ctx_main))];
 
         llama_sample_repetition_penalties(ctx_main, &cur_p,
                 prev.data() + prev.size() - penalty_last_n,
@@ -155,7 +155,7 @@ llama_token llama_sampling_sample(
 
         if (!penalize_nl) {
             for (size_t idx = 0; idx < cur_p.size; idx++) {
-                if (cur_p.data[idx].id == llama_token_nl(ctx_main)) {
+                if (cur_p.data[idx].id == llama_token_nl(llama_get_model(ctx_main))) {
                     cur_p.data[idx].logit = nl_logit;
                     break;
                 }
diff --git a/common/train.cpp b/common/train.cpp
index 154ca56e5..3cce5da26 100644
--- a/common/train.cpp
+++ b/common/train.cpp
@@ -236,8 +236,8 @@ int64_t get_example_targets_batch(
     int64_t used_samples = 0;
 
     ggml_set_f32(target_probs, 0.0f);
-    llama_token bos = llama_token_bos(lctx);
-    llama_token eos = llama_token_eos(lctx);
+    llama_token bos = llama_token_bos(llama_get_model(lctx));
+    llama_token eos = llama_token_eos(llama_get_model(lctx));
     // printf("%s: example_id=%d n_batch=%d n_train_samples=%zu\n", __func__, example_id, n_batch, n_train_samples);
     for (int k=0; k<n_batch; ++k) {
         // printf("%s: batch %d\n", __func__, k);
@@ -924,7 +924,7 @@ size_t tokenize_file(
         for (llama_token token=0; token < n_vocab; ++token) {
             max_token_text_size = std::max(
                 max_token_text_size,
-                strlen(llama_token_get_text(lctx, token)));
+                strlen(llama_token_get_text(llama_get_model(lctx), token)));
         }
 
         // upper bound of context byte length.
diff --git a/convert-bloom-hf-to-gguf.py b/convert-bloom-hf-to-gguf.py
index 14dbd793c..6e866d943 100755
--- a/convert-bloom-hf-to-gguf.py
+++ b/convert-bloom-hf-to-gguf.py
@@ -118,15 +118,24 @@ tokenizer = AutoTokenizer.from_pretrained(dir_model)
 vocab_size = hparams.get("vocab_size", len(tokenizer.vocab))
 assert max(tokenizer.vocab.values()) < vocab_size
 
+added_vocab = tokenizer.get_added_vocab()
 reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
 
 for i in range(vocab_size):
-    tokens.append(reverse_vocab[i] if i in reverse_vocab else f"[PAD{i}]")
-    scores.append(0.0)  # dummy
-    toktypes.append(gguf.TokenType.NORMAL)
+    if i not in reverse_vocab:
+        tokens.append(f"[PAD{i}]")
+        toktypes.append(gguf.TokenType.USER_DEFINED)
+    elif reverse_vocab[i] in added_vocab:
+        tokens.append(reverse_vocab[i])
+        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)
 
 gguf_writer.add_token_list(tokens)
-gguf_writer.add_token_scores(scores)
 gguf_writer.add_token_types(toktypes)
 
 special_vocab = gguf.SpecialVocab(dir_model, load_merges=True, n_vocab = len(tokens))
diff --git a/convert-gptneox-hf-to-gguf.py b/convert-gptneox-hf-to-gguf.py
index f1599b0c4..02d1fdf16 100755
--- a/convert-gptneox-hf-to-gguf.py
+++ b/convert-gptneox-hf-to-gguf.py
@@ -123,15 +123,24 @@ tokenizer = AutoTokenizer.from_pretrained(dir_model)
 vocab_size = hparams.get("vocab_size", len(tokenizer.vocab))
 assert max(tokenizer.vocab.values()) < vocab_size
 
+added_vocab = tokenizer.get_added_vocab()
 reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
 
 for i in range(vocab_size):
-    tokens.append(reverse_vocab[i] if i in reverse_vocab else f"[PAD{i}]")
-    scores.append(0.0) # dummy
-    toktypes.append(gguf.TokenType.NORMAL)
+    if i not in reverse_vocab:
+        tokens.append(f"[PAD{i}]")
+        toktypes.append(gguf.TokenType.USER_DEFINED)
+    elif reverse_vocab[i] in added_vocab:
+        tokens.append(reverse_vocab[i])
+        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)
 
 gguf_writer.add_token_list(tokens)
-gguf_writer.add_token_scores(scores)
 gguf_writer.add_token_types(toktypes)
 
 special_vocab = gguf.SpecialVocab(dir_model, load_merges = True, n_vocab = len(tokens))
diff --git a/convert-mpt-hf-to-gguf.py b/convert-mpt-hf-to-gguf.py
index 2d2fa2329..70d154b3f 100755
--- a/convert-mpt-hf-to-gguf.py
+++ b/convert-mpt-hf-to-gguf.py
@@ -136,9 +136,11 @@ for i in range(vocab_size):
         tokens.append(f"[PAD{i}]")
         toktypes.append(gguf.TokenType.USER_DEFINED)
     elif reverse_vocab[i] in added_vocab:
-        # NOTE: wouldn't we like to distinguish CONTROL tokens here?
         tokens.append(reverse_vocab[i])
-        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)
diff --git a/convert-refact-hf-to-gguf.py b/convert-refact-hf-to-gguf.py
index 934f3852b..f0cfe84d8 100755
--- a/convert-refact-hf-to-gguf.py
+++ b/convert-refact-hf-to-gguf.py
@@ -139,15 +139,24 @@ tokenizer = AutoTokenizer.from_pretrained(dir_model)
 vocab_size = hparams.get("vocab_size", len(tokenizer.vocab))
 assert max(tokenizer.vocab.values()) < vocab_size
 
+added_vocab = tokenizer.get_added_vocab()
 reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
 
 for i in range(vocab_size):
-    tokens.append(reverse_vocab[i] if i in reverse_vocab else f"[PAD{i}]")
-    scores.append(0.0) # dummy
-    toktypes.append(gguf.TokenType.NORMAL)
+    if i not in reverse_vocab:
+        tokens.append(f"[PAD{i}]")
+        toktypes.append(gguf.TokenType.USER_DEFINED)
+    elif reverse_vocab[i] in added_vocab:
+        tokens.append(reverse_vocab[i])
+        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)
 
 gguf_writer.add_token_list(tokens)
-gguf_writer.add_token_scores(scores)
 gguf_writer.add_token_types(toktypes)
 
 special_vocab = gguf.SpecialVocab(dir_model, load_merges=True, n_vocab = len(tokens))
diff --git a/convert-starcoder-hf-to-gguf.py b/convert-starcoder-hf-to-gguf.py
index fe8815cbf..a9bfed85e 100755
--- a/convert-starcoder-hf-to-gguf.py
+++ b/convert-starcoder-hf-to-gguf.py
@@ -111,17 +111,25 @@ tokenizer = AutoTokenizer.from_pretrained(dir_model)
 vocab_size = hparams.get("vocab_size", len(tokenizer.vocab))
 assert max(tokenizer.vocab.values()) < vocab_size
 
+added_vocab = tokenizer.get_added_vocab()
 reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
 
 for i in range(vocab_size):
-    tokens.append(reverse_vocab[i] if i in reverse_vocab else f"[PAD{i}]")
-    scores.append(0.0) # dummy
-    toktypes.append(gguf.TokenType.NORMAL)
+    if i not in reverse_vocab:
+        tokens.append(f"[PAD{i}]")
+        toktypes.append(gguf.TokenType.USER_DEFINED)
+    elif reverse_vocab[i] in added_vocab:
+        tokens.append(reverse_vocab[i])
+        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)
 
 gguf_writer.add_token_list(tokens)
-gguf_writer.add_token_scores(scores)
 gguf_writer.add_token_types(toktypes)
-
 special_vocab = gguf.SpecialVocab(dir_model, load_merges = True, n_vocab = len(tokens))
 special_vocab.add_to_gguf(gguf_writer)
 
diff --git a/examples/batched/batched.cpp b/examples/batched/batched.cpp
index 2797329b4..22a4265df 100644
--- a/examples/batched/batched.cpp
+++ b/examples/batched/batched.cpp
@@ -11,7 +11,7 @@ int main(int argc, char ** argv) {
     gpt_params params;
 
     if (argc == 1 || argv[1][0] == '-') {
-        printf("usage: %s MODEL_PATH [PROMPT] [PARALLEL] [LEN]\n" , argv[0]);
+        printf("usage: %s MODEL_PATH [PROMPT] [PARALLEL] [LEN] [NGL]\n" , argv[0]);
         return 1 ;
     }
 
@@ -21,6 +21,9 @@ int main(int argc, char ** argv) {
     // total length of the sequences including the prompt
     int n_len = 32;
 
+    // number of layers to offload to the GPU
+    int n_gpu_layers = 0;
+
     if (argc >= 2) {
         params.model = argv[1];
     }
@@ -37,6 +40,10 @@ int main(int argc, char ** argv) {
         n_len = std::atoi(argv[4]);
     }
 
+    if (argc >= 6) {
+        n_gpu_layers = std::atoi(argv[5]);
+    }
+
     if (params.prompt.empty()) {
         params.prompt = "Hello my name is";
     }
@@ -49,7 +56,7 @@ int main(int argc, char ** argv) {
 
     llama_model_params model_params = llama_model_default_params();
 
-    // model_params.n_gpu_layers = 99; // offload all layers to the GPU
+    model_params.n_gpu_layers = n_gpu_layers;
 
     llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
 
@@ -180,7 +187,7 @@ int main(int argc, char ** argv) {
             //const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p);
 
             // is it an end of stream? -> mark the stream as finished
-            if (new_token_id == llama_token_eos(ctx) || n_cur == n_len) {
+            if (new_token_id == llama_token_eos(model) || n_cur == n_len) {
                 i_batch[i] = -1;
                 LOG_TEE("\n");
                 if (n_parallel > 1) {
diff --git a/examples/beam-search/beam-search.cpp b/examples/beam-search/beam-search.cpp
index f078ab8a8..679b382e1 100644
--- a/examples/beam-search/beam-search.cpp
+++ b/examples/beam-search/beam-search.cpp
@@ -47,7 +47,7 @@ struct beam_search_callback_data {
 // 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.
 static bool is_at_eob(const beam_search_callback_data & callback_data, const llama_token * tokens, size_t n_tokens) {
-    return n_tokens && tokens[n_tokens-1] == llama_token_eos(callback_data.ctx);
+    return n_tokens && tokens[n_tokens-1] == llama_token_eos(llama_get_model(callback_data.ctx));
 }
 
 // Function matching type llama_beam_search_callback_fn_t.
diff --git a/examples/infill/infill.cpp b/examples/infill/infill.cpp
index 6331335e3..9c52b7bba 100644
--- a/examples/infill/infill.cpp
+++ b/examples/infill/infill.cpp
@@ -246,14 +246,14 @@ int main(int argc, char ** argv) {
     if (suff_rm_leading_spc && inp_sfx[0] == space_token) {
         inp_sfx.erase(inp_sfx.begin());
     }
-    inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(ctx));
+    inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(model));
     if (add_bos) {
-        inp_pfx.insert(inp_pfx.begin(), llama_token_bos(ctx));
+        inp_pfx.insert(inp_pfx.begin(), llama_token_bos(model));
     }
-    inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(ctx));
+    inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(model));
     embd_inp = inp_pfx;
     embd_inp.insert(embd_inp.end(), inp_sfx.begin(), inp_sfx.end());
-    embd_inp.push_back(llama_token_middle(ctx));
+    embd_inp.push_back(llama_token_middle(model));
 
     LOG("prefix: \"%s\"\n", log_tostr(params.input_prefix));
     LOG("suffix: \"%s\"\n", log_tostr(params.input_suffix));
@@ -261,7 +261,7 @@ int main(int argc, char ** argv) {
 
     // Should not run without any tokens
     if (embd_inp.empty()) {
-        embd_inp.push_back(llama_token_bos(ctx));
+        embd_inp.push_back(llama_token_bos(model));
         LOG("embd_inp was considered empty and bos was added: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp).c_str());
     }
 
@@ -577,10 +577,10 @@ int main(int argc, char ** argv) {
         if ((int) embd_inp.size() <= n_consumed) {
 
             // deal with eot token in infill mode
-            if ((llama_sampling_last(ctx_sampling) == llama_token_eot(ctx) || is_interacting) && params.interactive){
+            if ((llama_sampling_last(ctx_sampling) == llama_token_eot(model) || is_interacting) && params.interactive){
                 if(is_interacting && !params.interactive_first) {
                     // print an eot token
-                    printf("%s", llama_token_to_piece(ctx, llama_token_eot(ctx)).c_str());
+                    printf("%s", llama_token_to_piece(ctx, llama_token_eot(model)).c_str());
                 }
                 fflush(stdout);
                 printf("\n");
@@ -627,14 +627,14 @@ int main(int argc, char ** argv) {
                 if (suff_rm_leading_spc && inp_sfx[0] == space_token) {
                     inp_sfx.erase(inp_sfx.begin());
                 }
-                inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(ctx));
+                inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(model));
                 if (add_bos) {
-                    inp_pfx.insert(inp_pfx.begin(), llama_token_bos(ctx));
+                    inp_pfx.insert(inp_pfx.begin(), llama_token_bos(model));
                 }
-                inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(ctx));
+                inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(model));
                 embd_inp = inp_pfx;
                 embd_inp.insert(embd_inp.end(), inp_sfx.begin(), inp_sfx.end());
-                embd_inp.push_back(llama_token_middle(ctx));
+                embd_inp.push_back(llama_token_middle(model));
                 embd.clear();
                 embd_guidance.clear();
                 n_remain = params.n_predict;
@@ -644,7 +644,7 @@ int main(int argc, char ** argv) {
                 is_interacting = false;
             }
             // deal with end of text token in interactive mode
-            else if (llama_sampling_last(ctx_sampling) == llama_token_eos(ctx)) {
+            else if (llama_sampling_last(ctx_sampling) == llama_token_eos(model)) {
                 LOG("found EOS token\n");
 
                 if (params.interactive) {
@@ -661,7 +661,7 @@ int main(int argc, char ** argv) {
 
                 if (params.input_prefix_bos) {
                     LOG("adding input prefix BOS token\n");
-                    embd_inp.push_back(llama_token_bos(ctx));
+                    embd_inp.push_back(llama_token_bos(model));
                 }
 
                 std::string buffer;
@@ -724,7 +724,7 @@ int main(int argc, char ** argv) {
         }
 
         // end of text token
-        if (!embd.empty() && embd.back() == llama_token_eos(ctx) && !params.interactive) {
+        if (!embd.empty() && embd.back() == llama_token_eos(model) && !params.interactive) {
             break;
         }
 
@@ -736,7 +736,7 @@ int main(int argc, char ** argv) {
         }
     }
     if (!params.interactive && n_remain <= 0) {
-        printf("%s", llama_token_to_piece(ctx, llama_token_eot(ctx)).c_str());
+        printf("%s", llama_token_to_piece(ctx, llama_token_eot(model)).c_str());
         fflush(stdout);
     }
 
diff --git a/examples/llama-bench/llama-bench.cpp b/examples/llama-bench/llama-bench.cpp
index a04115c96..20767d555 100644
--- a/examples/llama-bench/llama-bench.cpp
+++ b/examples/llama-bench/llama-bench.cpp
@@ -933,7 +933,7 @@ struct sql_printer : public printer {
 };
 
 static void test_prompt(llama_context * ctx, int n_prompt, int n_past, int n_batch, int n_threads) {
-    std::vector<llama_token> tokens(n_batch, llama_token_bos(ctx));
+    std::vector<llama_token> tokens(n_batch, llama_token_bos(llama_get_model(ctx)));
     int n_processed = 0;
 
     llama_set_n_threads(ctx, n_threads, n_threads);
@@ -946,7 +946,7 @@ static void test_prompt(llama_context * ctx, int n_prompt, int n_past, int n_bat
 }
 
 static void test_gen(llama_context * ctx, int n_gen, int n_past, int n_threads) {
-    llama_token token = llama_token_bos(ctx);
+    llama_token token = llama_token_bos(llama_get_model(ctx));
 
     llama_set_n_threads(ctx, n_threads, n_threads);
 
diff --git a/examples/llava/llava-utils.h b/examples/llava/llava-utils.h
index 45b2b1ad3..320c71967 100644
--- a/examples/llava/llava-utils.h
+++ b/examples/llava/llava-utils.h
@@ -137,7 +137,7 @@ inline llama_token sample_id(llama_context * ctx_llama, gpt_params & params) {
 inline const char * sample(struct llama_context * ctx_llama, gpt_params & params, int * n_past) {
     int id = sample_id(ctx_llama, params);
     static std::string ret;
-    if (id == llama_token_eos(ctx_llama)) {
+    if (id == llama_token_eos(llama_get_model(ctx_llama))) {
         ret = "</s>";
     } else {
         ret = llama_token_to_piece(ctx_llama, id);
diff --git a/examples/main-cmake-pkg/CMakeLists.txt b/examples/main-cmake-pkg/CMakeLists.txt
index 908131884..cb00edbbb 100644
--- a/examples/main-cmake-pkg/CMakeLists.txt
+++ b/examples/main-cmake-pkg/CMakeLists.txt
@@ -16,6 +16,8 @@ add_library(common OBJECT
     ${_common_path}/console.cpp
     ${_common_path}/grammar-parser.h
     ${_common_path}/grammar-parser.cpp
+    ${_common_path}/sampling.h
+    ${_common_path}/sampling.cpp
     )
 
 # WARNING: because build-info.h is auto-generated, it will only
diff --git a/examples/main/main.cpp b/examples/main/main.cpp
index 2621bd539..3d9f670b9 100644
--- a/examples/main/main.cpp
+++ b/examples/main/main.cpp
@@ -248,7 +248,7 @@ int main(int argc, char ** argv) {
 
     // Should not run without any tokens
     if (embd_inp.empty()) {
-        embd_inp.push_back(llama_token_bos(ctx));
+        embd_inp.push_back(llama_token_bos(model));
         LOG("embd_inp was considered empty and bos was added: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp).c_str());
     }
 
@@ -693,7 +693,7 @@ int main(int argc, char ** argv) {
             }
 
             // deal with end of text token in interactive mode
-            if (llama_sampling_last(ctx_sampling) == llama_token_eos(ctx)) {
+            if (llama_sampling_last(ctx_sampling) == llama_token_eos(model)) {
                 LOG("found EOS token\n");
 
                 if (params.interactive) {
@@ -720,7 +720,7 @@ int main(int argc, char ** argv) {
 
                 if (params.input_prefix_bos) {
                     LOG("adding input prefix BOS token\n");
-                    embd_inp.push_back(llama_token_bos(ctx));
+                    embd_inp.push_back(llama_token_bos(model));
                 }
 
                 std::string buffer;
@@ -804,7 +804,7 @@ int main(int argc, char ** argv) {
         }
 
         // end of text token
-        if (!embd.empty() && embd.back() == llama_token_eos(ctx) && !(params.instruct || params.interactive)) {
+        if (!embd.empty() && embd.back() == llama_token_eos(model) && !(params.instruct || params.interactive)) {
             LOG_TEE(" [end of text]\n");
             break;
         }
diff --git a/examples/parallel/parallel.cpp b/examples/parallel/parallel.cpp
index eb64adef8..9a0b9c183 100644
--- a/examples/parallel/parallel.cpp
+++ b/examples/parallel/parallel.cpp
@@ -347,7 +347,7 @@ int main(int argc, char ** argv) {
                 //        client.id, client.seq_id, id, client.n_decoded, client.i_batch, token_str.c_str());
 
                 if (client.n_decoded > 2 &&
-                        (id == llama_token_eos(ctx) ||
+                        (id == llama_token_eos(model) ||
                          (params.n_predict > 0 && client.n_decoded + client.n_prompt >= params.n_predict) ||
                          client.response.find("User:") != std::string::npos ||
                          client.response.find('\n') != std::string::npos)) {
diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp
index 7d0038bd4..3c2542e8c 100644
--- a/examples/perplexity/perplexity.cpp
+++ b/examples/perplexity/perplexity.cpp
@@ -227,7 +227,7 @@ static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params &
 
             // add BOS token for the first batch of each chunk
             if (add_bos && j == 0) {
-                tokens[batch_start] = llama_token_bos(ctx);
+                tokens[batch_start] = llama_token_bos(llama_get_model(ctx));
             }
 
             const auto batch_logits = llama_get_logits(ctx);
@@ -350,7 +350,7 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
 
             // add BOS token for the first batch of each chunk
             if (add_bos && j == 0) {
-                tokens[batch_start] = llama_token_bos(ctx);
+                tokens[batch_start] = llama_token_bos(llama_get_model(ctx));
             }
 
             if (llama_decode(ctx, llama_batch_get_one(tokens.data() + batch_start, batch_size, j * n_batch, 0))) {
diff --git a/examples/server/server.cpp b/examples/server/server.cpp
index 210c41736..460a6baa0 100644
--- a/examples/server/server.cpp
+++ b/examples/server/server.cpp
@@ -454,7 +454,7 @@ struct llama_client_slot
     }
 
     void release() {
-        if (state == PROCESSING)
+        if (state == IDLE || state == PROCESSING)
         {
             t_token_generation = (ggml_time_us() - t_start_genereration) / 1e3;
             command = RELEASE;
@@ -726,7 +726,7 @@ struct llama_server_context
 
         if (json_value(data, "ignore_eos", false))
         {
-            slot->sparams.logit_bias[llama_token_eos(ctx)] = -INFINITY;
+            slot->sparams.logit_bias[llama_token_eos(model)] = -INFINITY;
         }
 
         const auto &logit_bias = data.find("logit_bias");
@@ -754,6 +754,7 @@ struct llama_server_context
         }
 
         slot->params.antiprompt.clear();
+
         const auto &stop = data.find("stop");
         if (stop != data.end() && stop->is_array())
         {
@@ -867,7 +868,7 @@ struct llama_server_context
 
         kv_cache_clear();
 
-        for (int32_t i = 0; i < batch.n_tokens; ++i)
+        for (int i = 0; i < (int) system_tokens.size(); ++i)
         {
             llama_batch_add(batch, system_tokens[i], i, { 0 }, false);
         }
@@ -894,16 +895,8 @@ struct llama_server_context
         {
             slot.release();
         }
-        wait_all_are_idle();
-        all_slots_are_idle = true;
 
-        // wait until system prompt load
         system_need_update = true;
-        while (system_need_update)
-        {
-            std::this_thread::sleep_for(std::chrono::milliseconds(5));
-        }
-        // system prompt loaded, continue
     }
 
     void process_system_prompt_data(const json &sys_props) {
@@ -915,26 +908,6 @@ struct llama_server_context
         {
             notify_system_prompt_changed();
         }
-        else
-        {
-            system_need_update = true;
-        }
-    }
-
-    void wait_all_are_idle() {
-        bool wait = true;
-        while (wait)
-        {
-            wait = false;
-            for (auto &slot : slots)
-            {
-                if (!slot.available())
-                {
-                    wait = true;
-                    break;
-                }
-            }
-        }
     }
 
     static size_t find_stopping_strings(const std::string &text, const size_t last_token_size,
@@ -965,7 +938,6 @@ struct llama_server_context
                     slot.has_next_token = false;
                 }
                 stop_pos = pos;
-
             }
         }
 
@@ -1056,7 +1028,7 @@ struct llama_server_context
             slot.has_next_token = false;
         }
 
-        if (!slot.cache_tokens.empty() && result.tok == llama_token_eos(ctx))
+        if (!slot.cache_tokens.empty() && result.tok == llama_token_eos(model))
         {
             slot.stopped_eos = true;
             slot.has_next_token = false;
@@ -1130,7 +1102,7 @@ struct llama_server_context
 
     json get_formated_generation(llama_client_slot &slot)
     {
-        const auto eos_bias = slot.sparams.logit_bias.find(llama_token_eos(ctx));
+        const auto eos_bias = slot.sparams.logit_bias.find(llama_token_eos(model));
         const bool ignore_eos = eos_bias != slot.sparams.logit_bias.end() &&
                                 eos_bias->second < 0.0f && std::isinf(eos_bias->second);
         return json {
@@ -1444,7 +1416,7 @@ struct llama_server_context
         process_tasks();
 
         // update the system prompt wait until all slots are idle state
-        if (system_need_update)
+        if (system_need_update && all_slots_are_idle)
         {
             LOG_TEE("updating system prompt\n");
             update_system_prompt();
@@ -1498,7 +1470,7 @@ struct llama_server_context
         for (auto & slot : slots)
         {
             // release the slot
-            if (slot.state == PROCESSING && slot.command == RELEASE)
+            if (slot.command == RELEASE)
             {
                 slot.state = IDLE;
                 slot.command = NONE;
@@ -1509,7 +1481,7 @@ struct llama_server_context
                 continue;
             }
 
-            if (slot.state == IDLE || slot.command == RELEASE)
+            if (slot.state == IDLE)
             {
                 continue;
             }
@@ -1530,6 +1502,17 @@ struct llama_server_context
         {
             for (auto & slot : slots)
             {
+                const bool has_prompt = slot.prompt.is_array() || (slot.prompt.is_string() && !slot.prompt.get<std::string>().empty());
+
+                // empty prompt passed -> release the slot and send empty response
+                if (slot.state == IDLE && slot.command == LOAD_PROMPT && !has_prompt)
+                {
+                    slot.release();
+                    slot.print_timings();
+                    send_final_response(slot);
+                    continue;
+                }
+
                 // need process the prompt
                 if (slot.state == IDLE && slot.command == LOAD_PROMPT)
                 {
@@ -1555,11 +1538,11 @@ struct llama_server_context
                             suffix_tokens.erase(suffix_tokens.begin());
                         }
 
-                        prefix_tokens.insert(prefix_tokens.begin(), llama_token_prefix(ctx));
-                        prefix_tokens.insert(prefix_tokens.begin(), llama_token_bos(ctx)); // always add BOS
-                        prefix_tokens.insert(prefix_tokens.end(), llama_token_suffix(ctx));
+                        prefix_tokens.insert(prefix_tokens.begin(), llama_token_prefix(model));
+                        prefix_tokens.insert(prefix_tokens.begin(), llama_token_bos(model)); // always add BOS
+                        prefix_tokens.insert(prefix_tokens.end(), llama_token_suffix(model));
                         prefix_tokens.insert(prefix_tokens.end(), suffix_tokens.begin(), suffix_tokens.end());
-                        prefix_tokens.push_back(llama_token_middle(ctx));
+                        prefix_tokens.push_back(llama_token_middle(model));
                         prompt_tokens = prefix_tokens;
                     }
                     else
@@ -1749,8 +1732,8 @@ struct llama_server_context
                 if (!process_token(result, slot))
                 {
                     slot.release();
-                    send_final_response(slot);
                     slot.print_timings();
+                    send_final_response(slot);
                 }
 
                 slot.i_batch = -1;
@@ -1766,21 +1749,22 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
     printf("usage: %s [options]\n", argv0);
     printf("\n");
     printf("options:\n");
-    printf("  -h, --help            show this help message and exit\n");
-    printf("  -v, --verbose         verbose output (default: %s)\n", server_verbose ? "enabled" : "disabled");
-    printf("  -t N, --threads N     number of threads to use during computation (default: %d)\n", params.n_threads);
-    printf("  -c N, --ctx-size N    size of the prompt context (default: %d)\n", params.n_ctx);
+    printf("  -h, --help                show this help message and exit\n");
+    printf("  -v, --verbose             verbose output (default: %s)\n", server_verbose ? "enabled" : "disabled");
+    printf("  -t N, --threads N         number of threads to use during computation (default: %d)\n", params.n_threads);
+    printf("  -tb N, --threads-batch N  number of threads to use during batch and prompt processing (default: same as --threads)\n");
+    printf("  -c N, --ctx-size N        size of the prompt context (default: %d)\n", params.n_ctx);
     printf("  --rope-scaling {none,linear,yarn}\n");
-    printf("                        RoPE frequency scaling method, defaults to linear unless specified by the model\n");
-    printf("  --rope-freq-base N    RoPE base frequency (default: loaded from model)\n");
-    printf("  --rope-freq-scale N   RoPE frequency scaling factor, expands context by a factor of 1/N\n");
-    printf("  --yarn-ext-factor N   YaRN: extrapolation mix factor (default: 1.0, 0.0 = full interpolation)\n");
-    printf("  --yarn-attn-factor N  YaRN: scale sqrt(t) or attention magnitude (default: 1.0)\n");
-    printf("  --yarn-beta-slow N    YaRN: high correction dim or alpha (default: %.1f)\n", params.yarn_beta_slow);
-    printf("  --yarn-beta-fast N    YaRN: low correction dim or beta (default: %.1f)\n", params.yarn_beta_fast);
-    printf("  -b N, --batch-size N  batch size for prompt processing (default: %d)\n", params.n_batch);
-    printf("  --memory-f32          use f32 instead of f16 for memory key+value (default: disabled)\n");
-    printf("                        not recommended: doubles context memory required and no measurable increase in quality\n");
+    printf("                            RoPE frequency scaling method, defaults to linear unless specified by the model\n");
+    printf("  --rope-freq-base N        RoPE base frequency (default: loaded from model)\n");
+    printf("  --rope-freq-scale N       RoPE frequency scaling factor, expands context by a factor of 1/N\n");
+    printf("  --yarn-ext-factor N       YaRN: extrapolation mix factor (default: 1.0, 0.0 = full interpolation)\n");
+    printf("  --yarn-attn-factor N      YaRN: scale sqrt(t) or attention magnitude (default: 1.0)\n");
+    printf("  --yarn-beta-slow N        YaRN: high correction dim or alpha (default: %.1f)\n", params.yarn_beta_slow);
+    printf("  --yarn-beta-fast N        YaRN: low correction dim or beta (default: %.1f)\n", params.yarn_beta_fast);
+    printf("  -b N, --batch-size N      batch size for prompt processing (default: %d)\n", params.n_batch);
+    printf("  --memory-f32              use f32 instead of f16 for memory key+value (default: disabled)\n");
+    printf("                            not recommended: doubles context memory required and no measurable increase in quality\n");
     if (llama_mlock_supported())
     {
         printf("  --mlock               force system to keep model in RAM rather than swapping or compressing\n");
@@ -1975,6 +1959,15 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
             }
             params.n_threads = std::stoi(argv[i]);
         }
+        else if (arg == "--threads-batch" || arg == "-tb")
+        {
+            if (++i >= argc)
+            {
+                invalid_param = true;
+                break;
+            }
+            params.n_threads_batch = std::stoi(argv[i]);
+        }
         else if (arg == "-b" || arg == "--batch-size")
         {
             if (++i >= argc)
@@ -2336,7 +2329,7 @@ int main(int argc, char **argv)
                 if (!json_value(data, "stream", false)) {
                     std::string completion_text;
                     task_result result = llama.next_result(task_id);
-                    if(!result.error && result.stop) {
+                    if (!result.error && result.stop) {
                         res.set_content(result.result_json.dump(-1, ' ', false, json::error_handler_t::replace), "application/json");
                     }
                     else
@@ -2363,7 +2356,7 @@ int main(int argc, char **argv)
                                 {
                                     return false;
                                 }
-                                if(result.stop) {
+                                if (result.stop) {
                                     break;
                                 }
                             } else {
diff --git a/examples/simple/simple.cpp b/examples/simple/simple.cpp
index 55385f566..f376c0509 100644
--- a/examples/simple/simple.cpp
+++ b/examples/simple/simple.cpp
@@ -138,7 +138,7 @@ int main(int argc, char ** argv) {
             const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p);
 
             // is it an end of stream?
-            if (new_token_id == llama_token_eos(ctx) || n_cur == n_len) {
+            if (new_token_id == llama_token_eos(model) || n_cur == n_len) {
                 LOG_TEE("\n");
 
                 break;
diff --git a/examples/speculative/speculative.cpp b/examples/speculative/speculative.cpp
index 894321ce9..92ad27e8e 100644
--- a/examples/speculative/speculative.cpp
+++ b/examples/speculative/speculative.cpp
@@ -163,7 +163,7 @@ int main(int argc, char ** argv) {
             printf("%s", token_str.c_str());
             fflush(stdout);
 
-            if (id == llama_token_eos(ctx_tgt)) {
+            if (id == llama_token_eos(model_tgt)) {
                 has_eos = true;
             }
 
diff --git a/ggml-cuda.cu b/ggml-cuda.cu
index 4c6a36ca1..a43a5873d 100644
--- a/ggml-cuda.cu
+++ b/ggml-cuda.cu
@@ -29,6 +29,8 @@
 #define __shfl_xor_sync(mask, var, laneMask, width) __shfl_xor(var, laneMask, width)
 #define cublasCreate hipblasCreate
 #define cublasGemmEx hipblasGemmEx
+#define cublasGemmBatchedEx hipblasGemmBatchedEx
+#define cublasGemmStridedBatchedEx hipblasGemmStridedBatchedEx
 #define cublasHandle_t hipblasHandle_t
 #define cublasSetMathMode(handle, mode) CUBLAS_STATUS_SUCCESS
 #define cublasSetStream hipblasSetStream
@@ -4326,13 +4328,13 @@ static __global__ void mul_mat_vec_nc_f16_f32( // nc == non-contiguous
 
     const half * x = (const half *) vx;
 
-    const int row_x = blockDim.y*blockIdx.y + threadIdx.y;
-    const int channel = blockDim.z*blockIdx.z + threadIdx.z;
+    const int row_x     = blockDim.y*blockIdx.y + threadIdx.y;
+    const int channel   = blockDim.z*blockIdx.z + threadIdx.z;
     const int channel_x = channel / channel_x_divisor;
 
-    const int nrows_y = ncols_x;
+    const int nrows_y   = ncols_x;
     const int nrows_dst = nrows_x;
-    const int row_dst = row_x;
+    const int row_dst   = row_x;
 
     const int idst = channel*nrows_dst + row_dst;
 
@@ -4345,13 +4347,13 @@ static __global__ void mul_mat_vec_nc_f16_f32( // nc == non-contiguous
             break;
         }
 
-        const int ix = channel_x*channel_stride_x + row_x*row_stride_x + col_x;
-        const float xi = __half2float(x[ix]);
-
         const int row_y = col_x;
 
+        const int ix = channel_x*channel_stride_x + row_x*row_stride_x + col_x;
         const int iy = channel*nrows_y + row_y;
 
+        const float xi = __half2float(x[ix]);
+
         tmp += xi * y[iy];
     }
 
@@ -5698,10 +5700,10 @@ void ggml_init_cublas() {
         GGML_ASSERT(g_device_count <= GGML_CUDA_MAX_DEVICES);
         int64_t total_vram = 0;
         fprintf(stderr, "%s: found %d " GGML_CUDA_NAME " devices:\n", __func__, g_device_count);
-        for (int64_t id = 0; id < g_device_count; ++id) {
+        for (int id = 0; id < g_device_count; ++id) {
             cudaDeviceProp prop;
             CUDA_CHECK(cudaGetDeviceProperties(&prop, id));
-            fprintf(stderr, "  Device %ld: %s, compute capability %d.%d\n", id, prop.name, prop.major, prop.minor);
+            fprintf(stderr, "  Device %d: %s, compute capability %d.%d\n", id, prop.name, prop.major, prop.minor);
 
             g_tensor_split[id] = total_vram;
             total_vram += prop.totalGlobalMem;
@@ -5711,15 +5713,15 @@ void ggml_init_cublas() {
             g_compute_capabilities[id] = 100*prop.major + 10*prop.minor;
 #endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
         }
-        for (int64_t id = 0; id < g_device_count; ++id) {
+        for (int id = 0; id < g_device_count; ++id) {
             g_tensor_split[id] /= total_vram;
         }
 
-        for (int64_t id = 0; id < g_device_count; ++id) {
+        for (int id = 0; id < g_device_count; ++id) {
             CUDA_CHECK(ggml_cuda_set_device(id));
 
             // create cuda streams
-            for (int64_t is = 0; is < MAX_STREAMS; ++is) {
+            for (int is = 0; is < MAX_STREAMS; ++is) {
                 CUDA_CHECK(cudaStreamCreateWithFlags(&g_cudaStreams[id][is], cudaStreamNonBlocking));
             }
 
@@ -7063,7 +7065,8 @@ static void ggml_cuda_mul_mat_vec_p021(const ggml_tensor * src0, const ggml_tens
 }
 
 static void ggml_cuda_mul_mat_vec_nc(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst){
-    GGML_ASSERT(!ggml_is_contiguous(src0) && ggml_is_contiguous(src1));
+    GGML_ASSERT(!ggml_is_transposed(src0));
+    GGML_ASSERT(!ggml_is_transposed(src1));
     GGML_ASSERT(!ggml_is_permuted(src0));
     GGML_ASSERT(src0->backend != GGML_BACKEND_GPU_SPLIT);
     GGML_ASSERT(src0->type == GGML_TYPE_F16);
@@ -7073,11 +7076,11 @@ static void ggml_cuda_mul_mat_vec_nc(const ggml_tensor * src0, const ggml_tensor
     const int64_t ne01 = src0->ne[1];
     const int64_t ne02 = src0->ne[2];
 
-    const int64_t ne12 = src1->ne[2];
-
     const int64_t nb01 = src0->nb[1];
     const int64_t nb02 = src0->nb[2];
 
+    const int64_t ne12 = src1->ne[2];
+
     CUDA_CHECK(ggml_cuda_set_device(g_main_device));
     cudaStream_t main_stream = g_cudaStreams[g_main_device][0];
 
@@ -7096,6 +7099,159 @@ static void ggml_cuda_mul_mat_vec_nc(const ggml_tensor * src0, const ggml_tensor
     ggml_mul_mat_vec_nc_f16_f32_cuda(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, row_stride_x, ne02, ne12, channel_stride_x, main_stream);
 }
 
+static void ggml_cuda_mul_mat_mat_batched_cublas(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst){
+    GGML_ASSERT(!ggml_is_transposed(src0));
+    GGML_ASSERT(!ggml_is_transposed(src1));
+    GGML_ASSERT(src0->backend != GGML_BACKEND_GPU_SPLIT);
+    GGML_ASSERT(src0->type == GGML_TYPE_F16);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+
+    const int64_t ne00 = src0->ne[0]; GGML_UNUSED(ne00);
+    const int64_t ne01 = src0->ne[1];
+    const int64_t ne02 = src0->ne[2];
+    const int64_t ne03 = src0->ne[3];
+
+    const int64_t nb01 = src0->nb[1];
+    const int64_t nb02 = src0->nb[2]; GGML_UNUSED(nb02);
+    const int64_t nb03 = src0->nb[3]; GGML_UNUSED(nb03);
+
+    const int64_t ne10 = src1->ne[0];
+    const int64_t ne11 = src1->ne[1];
+    const int64_t ne12 = src1->ne[2];
+    const int64_t ne13 = src1->ne[3];
+
+    const int64_t nb11 = src1->nb[1];
+    const int64_t nb12 = src1->nb[2]; GGML_UNUSED(nb12);
+    const int64_t nb13 = src1->nb[3]; GGML_UNUSED(nb13);
+
+    const int64_t ne1 = ggml_nelements(src1);
+    const int64_t ne  = ggml_nelements(dst);
+
+    CUDA_CHECK(ggml_cuda_set_device(g_main_device));
+    cudaStream_t main_stream = g_cudaStreams[g_main_device][0];
+
+    int id;
+    CUDA_CHECK(cudaGetDevice(&id));
+    CUBLAS_CHECK(cublasSetStream(g_cublas_handles[id], main_stream));
+
+    ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra;
+    void * src0_ddq = src0_extra->data_device[g_main_device];
+    half * src0_as_f16 = (half *) src0_ddq;
+
+    ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra;
+    float * src1_ddf = (float *) src1_extra->data_device[g_main_device];
+
+    ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra;
+    float * dst_ddf = (float *) dst_extra->data_device[g_main_device];
+
+    // convert src1 to fp16
+    const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(src1->type);
+    GGML_ASSERT(to_fp16_cuda != nullptr);
+
+    size_t src1_as = 0;
+    half * src1_as_f16 = (half *) ggml_cuda_pool_malloc(ne1 * sizeof(half), &src1_as);
+    to_fp16_cuda(src1_ddf, src1_as_f16, ne1, main_stream);
+
+    size_t dst_as = 0;
+    half * dst_f16 = (half *) ggml_cuda_pool_malloc(ne * sizeof(half), &dst_as);
+
+    GGML_ASSERT(ne12 % ne02 == 0);
+    GGML_ASSERT(ne13 % ne03 == 0);
+
+    // broadcast factors
+    const int64_t r2 = ne12/ne02;
+    const int64_t r3 = ne13/ne03;
+
+    const half alpha_f16 = 1.0f;
+    const half beta_f16  = 0.0f;
+
+#if 0
+    // use cublasGemmEx
+    {
+        for (int i13 = 0; i13 < ne13; ++i13) {
+            for (int i12 = 0; i12 < ne12; ++i12) {
+                int i03 = i13 / r3;
+                int i02 = i12 / r2;
+
+                CUBLAS_CHECK(
+                        cublasGemmEx(g_cublas_handles[id], CUBLAS_OP_T, CUBLAS_OP_N,
+                            ne01, ne11, ne10,
+                            &alpha_f16, (const char *) src0_as_f16 + i02*src0->nb[2]   + i03*src0->nb[3]  , CUDA_R_16F, nb01/sizeof(half),
+                                        (const char *) src1_as_f16 + i12*src1->nb[2]/2 + i13*src1->nb[3]/2, CUDA_R_16F, nb11/sizeof(float),
+                            &beta_f16,  (      char *)     dst_f16 + i12* dst->nb[2]/2 + i13* dst->nb[3]/2, CUDA_R_16F, ne01,
+                            CUBLAS_COMPUTE_16F,
+                            CUBLAS_GEMM_DEFAULT_TENSOR_OP));
+            }
+        }
+    }
+#else
+    if (r2 == 1 && r3 == 1 && src0->nb[2]*src0->ne[2] == src0->nb[3] && src1->nb[2]*src1->ne[2] == src1->nb[3]) {
+        // there is no broadcast and src0, src1 are contiguous across dims 2, 3
+        // use cublasGemmStridedBatchedEx
+        CUBLAS_CHECK(
+        cublasGemmStridedBatchedEx(g_cublas_handles[id], CUBLAS_OP_T, CUBLAS_OP_N,
+                ne01, ne11, ne10,
+                &alpha_f16, (const char *) src0_as_f16, CUDA_R_16F, nb01/sizeof(half),  src0->nb[2]/sizeof(half),  // strideA
+                            (const char *) src1_as_f16, CUDA_R_16F, nb11/sizeof(float), src1->nb[2]/sizeof(float), // strideB
+                &beta_f16,  (      char *)     dst_f16, CUDA_R_16F, ne01,                dst->nb[2]/sizeof(float), // strideC
+                ne12*ne13,
+                CUBLAS_COMPUTE_16F,
+                CUBLAS_GEMM_DEFAULT_TENSOR_OP));
+    } else {
+        // use cublasGemmBatchedEx
+        // TODO: https://github.com/ggerganov/llama.cpp/pull/3749#discussion_r1369997000
+        const int ne23 = ne12*ne13;
+
+        // TODO: avoid this alloc
+        void ** ptrs = (void **) malloc(3*ne23*sizeof(void *));
+
+        for (int i13 = 0; i13 < ne13; ++i13) {
+            for (int i12 = 0; i12 < ne12; ++i12) {
+                int i03 = i13 / r3;
+                int i02 = i12 / r2;
+
+                ptrs[0*ne23 + i12 + i13*ne12] = (char *) src0_as_f16 + i02*src0->nb[2]   + i03*src0->nb[3];
+                ptrs[1*ne23 + i12 + i13*ne12] = (char *) src1_as_f16 + i12*src1->nb[2]/2 + i13*src1->nb[3]/2;
+                ptrs[2*ne23 + i12 + i13*ne12] = (char *)     dst_f16 + i12* dst->nb[2]/2 + i13* dst->nb[3]/2;
+            }
+        }
+
+        // allocate device memory for pointers
+        void ** ptrs_as = nullptr;
+        CUDA_CHECK(cudaMalloc(&ptrs_as, 3*ne23*sizeof(void *)));
+
+        // TODO: this does not work for some reason -- not sure why?
+        //size_t ptrs_s = 0;
+        //ptrs_as = (void **) ggml_cuda_pool_malloc(3*ne23*sizeof(void *), &ptrs_s);
+
+        // copy pointers to device
+        CUDA_CHECK(cudaMemcpy(ptrs_as, ptrs, 3*ne23*sizeof(void *), cudaMemcpyHostToDevice));
+
+        free(ptrs);
+
+        CUBLAS_CHECK(
+        cublasGemmBatchedEx(g_cublas_handles[id], CUBLAS_OP_T, CUBLAS_OP_N,
+                ne01, ne11, ne10,
+                &alpha_f16, (const void **) (ptrs_as + 0*ne23), CUDA_R_16F, nb01/sizeof(half),
+                            (const void **) (ptrs_as + 1*ne23), CUDA_R_16F, nb11/sizeof(float),
+                &beta_f16,  (      void **) (ptrs_as + 2*ne23), CUDA_R_16F, ne01,
+                ne23,
+                CUBLAS_COMPUTE_16F,
+                CUBLAS_GEMM_DEFAULT_TENSOR_OP));
+
+        // free device memory for pointers
+        CUDA_CHECK(cudaFree(ptrs_as));
+        //ggml_cuda_pool_free(ptrs_as, ptrs_s);
+    }
+#endif
+
+    const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(GGML_TYPE_F16);
+    to_fp32_cuda(dst_f16, dst_ddf, ne, main_stream);
+
+    ggml_cuda_pool_free(src1_as_f16, src1_as);
+    ggml_cuda_pool_free(dst_f16, dst_as);
+}
+
 static void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     bool all_on_device = (src0->backend == GGML_BACKEND_GPU || src0->backend == GGML_BACKEND_GPU_SPLIT) &&
         src1->backend == GGML_BACKEND_GPU && dst->backend == GGML_BACKEND_GPU;
@@ -7108,10 +7264,22 @@ static void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1
         }
     }
 
+    // debug helpers
+    //printf("src0: %8d %8d %8d %8d\n", src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3]);
+    //printf("      %8d %8d %8d %8d\n", src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3]);
+    //printf("src1: %8d %8d %8d %8d\n", src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3]);
+    //printf("      %8d %8d %8d %8d\n", src1->nb[0], src1->nb[1], src1->nb[2], src1->nb[3]);
+    //printf("src0 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src0), ggml_is_transposed(src0), ggml_type_name(src0->type), src0->name);
+    //printf("src1 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src1), ggml_is_transposed(src1), ggml_type_name(src1->type), src1->name);
+
     if (all_on_device && src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && src1->ne[1] == 1) {
+        // KQ
         ggml_cuda_mul_mat_vec_p021(src0, src1, dst);
-    } else if (all_on_device && !ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && src1->ne[1] == 1) {
+    } else if (all_on_device && src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && src1->ne[1] == 1) {
+        // KQV
         ggml_cuda_mul_mat_vec_nc(src0, src1, dst);
+    } else if (all_on_device && src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32 && !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
+        ggml_cuda_mul_mat_mat_batched_cublas(src0, src1, dst);
     } else if (src0->type == GGML_TYPE_F32) {
         ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_mul_mat_cublas, false);
     } else if (ggml_is_quantized(src0->type) || src0->type == GGML_TYPE_F16) {
diff --git a/ggml-metal.m b/ggml-metal.m
index d3165a218..2cc7aad3e 100644
--- a/ggml-metal.m
+++ b/ggml-metal.m
@@ -62,6 +62,7 @@ struct ggml_metal_context {
     GGML_METAL_DECL_KERNEL(mul);
     GGML_METAL_DECL_KERNEL(mul_row); // TODO: avoid this extra kernel, instead extend the "mul" kernel to support broadcast
     GGML_METAL_DECL_KERNEL(scale);
+    GGML_METAL_DECL_KERNEL(scale_4);
     GGML_METAL_DECL_KERNEL(silu);
     GGML_METAL_DECL_KERNEL(relu);
     GGML_METAL_DECL_KERNEL(gelu);
@@ -249,6 +250,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(mul);
         GGML_METAL_ADD_KERNEL(mul_row);
         GGML_METAL_ADD_KERNEL(scale);
+        GGML_METAL_ADD_KERNEL(scale_4);
         GGML_METAL_ADD_KERNEL(silu);
         GGML_METAL_ADD_KERNEL(relu);
         GGML_METAL_ADD_KERNEL(gelu);
@@ -347,6 +349,7 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
     GGML_METAL_DEL_KERNEL(mul);
     GGML_METAL_DEL_KERNEL(mul_row);
     GGML_METAL_DEL_KERNEL(scale);
+    GGML_METAL_DEL_KERNEL(scale_4);
     GGML_METAL_DEL_KERNEL(silu);
     GGML_METAL_DEL_KERNEL(relu);
     GGML_METAL_DEL_KERNEL(gelu);
@@ -923,15 +926,20 @@ void ggml_metal_graph_compute(
 
                             const float scale = *(const float *) src1->data;
 
-                            [encoder setComputePipelineState:ctx->pipeline_scale];
+                            int64_t n = ggml_nelements(dst);
+
+                            if (n % 4 == 0) {
+                                n /= 4;
+                                [encoder setComputePipelineState:ctx->pipeline_scale_4];
+                            } else {
+                                [encoder setComputePipelineState:ctx->pipeline_scale];
+                            }
+
                             [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                             [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
                             [encoder setBytes:&scale length:sizeof(scale) atIndex:2];
 
-                            const int64_t n = ggml_nelements(dst);
-                            GGML_ASSERT(n % 4 == 0);
-
-                            [encoder dispatchThreadgroups:MTLSizeMake(n/4, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                            [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                         } break;
                     case GGML_OP_UNARY:
                         switch (ggml_get_unary_op(gf->nodes[i])) {
diff --git a/ggml-metal.metal b/ggml-metal.metal
index fcb5327e8..ddfe37813 100644
--- a/ggml-metal.metal
+++ b/ggml-metal.metal
@@ -125,9 +125,17 @@ kernel void kernel_mul_row(
 }
 
 kernel void kernel_scale(
+        device const float * src0,
+        device       float * dst,
+        constant     float & scale,
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = src0[tpig] * scale;
+}
+
+kernel void kernel_scale_4(
         device const float4 * src0,
         device       float4 * dst,
-        constant     float & scale,
+        constant     float  & scale,
         uint tpig[[thread_position_in_grid]]) {
     dst[tpig] = src0[tpig] * scale;
 }
diff --git a/ggml.c b/ggml.c
index 5403ae1d1..0034067d5 100644
--- a/ggml.c
+++ b/ggml.c
@@ -572,7 +572,6 @@ int64_t ggml_cycles_per_ms(void) {
 #define ggml_perf_cycles_per_ms() 0
 #endif
 
-
 //
 // cache line
 //
@@ -1829,7 +1828,6 @@ ggml_type_traits_t ggml_internal_get_type_traits(enum ggml_type type) {
     return type_traits[type];
 }
 
-
 //
 // simd mappings
 //
@@ -4058,16 +4056,17 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "ALIBI",
     "CLAMP",
     "CONV_1D",
+    "CONV_1D_STAGE_0",
+    "CONV_1D_STAGE_1",
     "CONV_TRANSPOSE_1D",
     "CONV_2D",
+    "CONV_2D_STAGE_0",
+    "CONV_2D_STAGE_1",
     "CONV_TRANSPOSE_2D",
     "POOL_1D",
     "POOL_2D",
     "UPSCALE",
 
-    "CONV_1D_STAGE_0",
-    "CONV_1D_STAGE_1",
-
     "FLASH_ATTN",
     "FLASH_FF",
     "FLASH_ATTN_BACK",
@@ -4093,7 +4092,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "CROSS_ENTROPY_LOSS_BACK",
 };
 
-static_assert(GGML_OP_COUNT == 71, "GGML_OP_COUNT != 71");
+static_assert(GGML_OP_COUNT == 73, "GGML_OP_COUNT != 73");
 
 static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "none",
@@ -4144,16 +4143,17 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "alibi(x)",
     "clamp(x)",
     "conv_1d(x)",
+    "conv_1d_stage_0(x)",
+    "conv_1d_stage_1(x)",
     "conv_transpose_1d(x)",
     "conv_2d(x)",
+    "conv_2d_stage_0(x)",
+    "conv_2d_stage_1(x)",
     "conv_transpose_2d(x)",
     "pool_1d(x)",
     "pool_2d(x)",
     "upscale(x)",
 
-    "conv_1d_stage_0(x)",
-    "conv_1d_stage_1(x)",
-
     "flash_attn(x)",
     "flash_ff(x)",
     "flash_attn_back(x)",
@@ -4179,7 +4179,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "cross_entropy_loss_back(x,y)",
 };
 
-static_assert(GGML_OP_COUNT == 71, "GGML_OP_COUNT != 71");
+static_assert(GGML_OP_COUNT == 73, "GGML_OP_COUNT != 73");
 
 static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2");
 
@@ -4210,8 +4210,10 @@ static void ggml_setup_op_has_task_pass(void) {
         p[GGML_OP_CONV_1D                ] = true;
         p[GGML_OP_CONV_1D_STAGE_0        ] = true;
         p[GGML_OP_CONV_1D_STAGE_1        ] = true;
-        p[GGML_OP_CONV_2D                ] = true;
         p[GGML_OP_CONV_TRANSPOSE_1D      ] = true;
+        p[GGML_OP_CONV_2D                ] = true;
+        p[GGML_OP_CONV_2D_STAGE_0        ] = true;
+        p[GGML_OP_CONV_2D_STAGE_1        ] = true;
         p[GGML_OP_CONV_TRANSPOSE_2D      ] = true;
         p[GGML_OP_FLASH_ATTN_BACK        ] = true;
         p[GGML_OP_CROSS_ENTROPY_LOSS     ] = true;
@@ -5955,7 +5957,6 @@ struct ggml_tensor * ggml_sqrt_inplace(
     return ggml_sqrt_impl(ctx, a, true);
 }
 
-
 // ggml_log
 
 static struct ggml_tensor * ggml_log_impl(
@@ -6009,7 +6010,6 @@ struct ggml_tensor * ggml_sum(
     return result;
 }
 
-
 // ggml_sum_rows
 
 struct ggml_tensor * ggml_sum_rows(
@@ -6641,7 +6641,6 @@ struct ggml_tensor * ggml_set_2d_inplace(
     return ggml_set_impl(ctx, a, b, nb1, a->nb[2], a->nb[3], offset, false);
 }
 
-
 // ggml_cpy
 
 static struct ggml_tensor * ggml_cpy_impl(
@@ -6721,7 +6720,6 @@ struct ggml_tensor * ggml_cont_inplace(
     return ggml_cont_impl(ctx, a, true);
 }
 
-
 // make contiguous, with new shape
 GGML_API struct ggml_tensor * ggml_cont_1d(
         struct ggml_context * ctx,
@@ -7174,7 +7172,6 @@ struct ggml_tensor * ggml_diag(
     return result;
 }
 
-
 // ggml_diag_mask_inf
 
 static struct ggml_tensor * ggml_diag_mask_inf_impl(
@@ -7286,7 +7283,6 @@ struct ggml_tensor * ggml_soft_max_inplace(
     return ggml_soft_max_impl(ctx, a, true);
 }
 
-
 // ggml_soft_max_back
 
 static struct ggml_tensor * ggml_soft_max_back_impl(
@@ -7732,7 +7728,11 @@ GGML_API struct ggml_tensor * ggml_conv_transpose_1d(
 
 // ggml_conv_2d
 
-struct ggml_tensor * ggml_conv_2d(
+// im2col: [N, IC, IH, IW] => [N, OH, OW, IC*KH*KW]
+// a: [OC，IC, KH, KW]
+// b: [N, IC, IH, IW]
+// result: [N, OH, OW, IC*KH*KW]
+static struct ggml_tensor * ggml_conv_2d_stage_0(
     struct ggml_context * ctx,
     struct ggml_tensor  * a,
     struct ggml_tensor  * b,
@@ -7751,17 +7751,21 @@ struct ggml_tensor * ggml_conv_2d(
         is_node = true;
     }
 
+    const int64_t OH = ggml_calc_conv_output_size(b->ne[1], a->ne[1], s1, p1, d1);
+    const int64_t OW = ggml_calc_conv_output_size(b->ne[0], a->ne[0], s0, p0, d0);
+
     const int64_t ne[4] = {
-        ggml_calc_conv_output_size(b->ne[0], a->ne[0], s0, p0, d0),
-        ggml_calc_conv_output_size(b->ne[1], a->ne[1], s1, p1, d1),
-        a->ne[3], b->ne[3],
+        a->ne[2] * a->ne[1] * a->ne[0],
+        OW,
+        OH,
+        b->ne[3],
     };
-    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
+    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F16, 4, ne);
 
     int32_t params[] = { s0, s1, p0, p1, d0, d1 };
     ggml_set_op_params(result, params, sizeof(params));
 
-    result->op = GGML_OP_CONV_2D;
+    result->op = GGML_OP_CONV_2D_STAGE_0;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
     result->src[0] = a;
     result->src[1] = b;
@@ -7770,8 +7774,61 @@ struct ggml_tensor * ggml_conv_2d(
 
 }
 
-// ggml_conv_2d_sk_p0
+// gemm: [N, OC, OH, OW] = [OC, IC * KH * KW] x [N*OH*OW, IC * KH * KW]
+// a: [OC, IC, KH, KW]
+// b: [N, OH, OW, IC * KH * KW]
+// result: [N, OC, OH, OW]
+static struct ggml_tensor * ggml_conv_2d_stage_1(
+    struct ggml_context * ctx,
+    struct ggml_tensor  * a,
+    struct ggml_tensor  * b) {
 
+    bool is_node = false;
+
+    if (a->grad || b->grad) {
+        GGML_ASSERT(false); // TODO: implement backward
+        is_node = true;
+    }
+
+    const int64_t ne[4] = {
+        b->ne[1],
+        b->ne[2],
+        a->ne[3],
+        b->ne[3],
+    };
+    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
+
+    result->op = GGML_OP_CONV_2D_STAGE_1;
+    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->src[0] = a;
+    result->src[1] = b;
+
+    return result;
+
+}
+
+// a: [OC，IC, KH, KW]
+// b: [N, IC, IH, IW]
+// result: [N, OC, OH, OW]
+struct ggml_tensor * ggml_conv_2d(
+    struct ggml_context * ctx,
+    struct ggml_tensor  * a,
+    struct ggml_tensor  * b,
+    int                  s0,
+    int                  s1,
+    int                  p0,
+    int                  p1,
+    int                  d0,
+    int                  d1) {
+
+    struct ggml_tensor * result = ggml_conv_2d_stage_0(ctx, a, b, s0, s1, p0, p1, d0, d1); // [N, OH, OW, IC * KH * KW]
+    result = ggml_conv_2d_stage_1(ctx, a, result);
+
+    return result;
+
+}
+
+// ggml_conv_2d_sk_p0
 struct ggml_tensor * ggml_conv_2d_sk_p0(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
@@ -8210,7 +8267,6 @@ static struct ggml_tensor * ggml_add_rel_pos_impl(
     return result;
 }
 
-
 struct ggml_tensor * ggml_add_rel_pos(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
@@ -8655,8 +8711,6 @@ struct ggml_tensor * ggml_map_custom3_inplace(
     return ggml_map_custom3_impl(ctx, a, b, c, fun, n_tasks, userdata, true);
 }
 
-
-
 // ggml_cross_entropy_loss
 
 struct ggml_tensor * ggml_cross_entropy_loss(
@@ -9858,7 +9912,6 @@ static void ggml_compute_forward_add1(
     }
 }
 
-
 // ggml_compute_forward_acc
 
 static void ggml_compute_forward_acc_f32(
@@ -9998,7 +10051,6 @@ static void ggml_compute_forward_sub_f32(
             const int i2 = (ir - i3*ne2*ne1)/ne1;
             const int i1 = (ir - i3*ne2*ne1 - i2*ne1);
 
-
 #ifdef GGML_USE_ACCELERATE
             vDSP_vsub(
                     (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11), 1,
@@ -10179,7 +10231,6 @@ static void ggml_compute_forward_div_f32(
             const int i2 = (ir - i3*ne2*ne1)/ne1;
             const int i1 = (ir - i3*ne2*ne1 - i2*ne1);
 
-
 #ifdef GGML_USE_ACCELERATE
             UNUSED(ggml_vec_div_f32);
 
@@ -10317,7 +10368,6 @@ static void ggml_compute_forward_sqrt(
     }
 }
 
-
 // ggml_compute_forward_log
 
 static void ggml_compute_forward_log_f32(
@@ -12150,7 +12200,6 @@ static void ggml_compute_forward_out_prod_f32(
         }
     }
 
-
     //int64_t t1 = ggml_perf_time_us();
     //static int64_t acc = 0;
     //acc += t1 - t0;
@@ -12346,7 +12395,6 @@ static void ggml_compute_forward_scale_f32(
 
     const size_t nb1 = dst->nb[1];
 
-
     for (int i1 = ir0; i1 < ir1; i1++) {
         if (dst->data != src0->data) {
             // src0 is same shape as dst => same indices
@@ -12744,7 +12792,6 @@ static void ggml_compute_forward_get_rows_back_f32(
     }
 }
 
-
 static void ggml_compute_forward_get_rows_back(
         const struct ggml_compute_params * params,
         const struct ggml_tensor * src0,
@@ -14087,6 +14134,7 @@ static void ggml_compute_forward_conv_1d_f32(
     }
 }
 
+// TODO: reuse ggml_mul_mat or implement ggml_im2col and remove stage_0 and stage_1
 static void gemm_f16_out_f32(int64_t m, int64_t n, int64_t k,
                              ggml_fp16_t * A,
                              ggml_fp16_t * B,
@@ -14388,6 +14436,9 @@ static void ggml_compute_forward_conv_transpose_1d_f16_f32(
             }
         }
 
+        // need to zero dst since we are accumulating into it
+        memset(dst->data, 0, ggml_nbytes(dst));
+
         return;
     }
 
@@ -14460,7 +14511,7 @@ static void ggml_compute_forward_conv_transpose_1d_f32(
                     const float * const src = (float *)((char *) src0->data + i02*nb02 + i01*nb01);
                     float * dst_data = wdata + i01*ne00*ne02;
                     for (int64_t i00 = 0; i00 < ne00; i00++) {
-                        dst_data[i01*ne00*ne02 + i00*ne02 + i02] = src[i00];
+                        dst_data[i00*ne02 + i02] = src[i00];
                     }
                 }
             }
@@ -14479,6 +14530,9 @@ static void ggml_compute_forward_conv_transpose_1d_f32(
             }
         }
 
+        // need to zero dst since we are accumulating into it
+        memset(dst->data, 0, ggml_nbytes(dst));
+
         return;
     }
 
@@ -14540,6 +14594,144 @@ static void ggml_compute_forward_conv_transpose_1d(
 
 // ggml_compute_forward_conv_2d
 
+// src0: kernel [OC, IC, KH, KW]
+// src1: image [N, IC, IH, IW]
+// dst:  result [N, OH, OW, IC*KH*KW]
+static void ggml_compute_forward_conv_2d_stage_0_f32(
+        const struct ggml_compute_params * params,
+        const struct ggml_tensor * src0,
+        const struct ggml_tensor * src1,
+              struct ggml_tensor * dst) {
+    GGML_ASSERT(src0->type == GGML_TYPE_F16);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F16);
+
+    int64_t t0 = ggml_perf_time_us();
+    UNUSED(t0);
+
+    GGML_TENSOR_BINARY_OP_LOCALS;
+
+    const int64_t N = ne13;
+    const int64_t IC = ne12;
+    const int64_t IH = ne11;
+    const int64_t IW = ne10;
+
+    // const int64_t OC = ne03;
+    // const int64_t IC = ne02;
+    const int64_t KH = ne01;
+    const int64_t KW = ne00;
+
+    const int64_t OH = ne2;
+    const int64_t OW = ne1;
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int32_t s0 = ((const int32_t*)(dst->op_params))[0];
+    const int32_t s1 = ((const int32_t*)(dst->op_params))[1];
+    const int32_t p0 = ((const int32_t*)(dst->op_params))[2];
+    const int32_t p1 = ((const int32_t*)(dst->op_params))[3];
+    const int32_t d0 = ((const int32_t*)(dst->op_params))[4];
+    const int32_t d1 = ((const int32_t*)(dst->op_params))[5];
+
+    GGML_ASSERT(nb00 == sizeof(ggml_fp16_t));
+    GGML_ASSERT(nb10 == sizeof(float));
+
+    if (params->type == GGML_TASK_INIT) {
+        memset(dst->data, 0, ggml_nbytes(dst));
+        return;
+    }
+
+    if (params->type == GGML_TASK_FINALIZE) {
+        return;
+    }
+
+    // im2col: [N, IC, IH, IW] => [N, OH, OW, IC*KH*KW]
+    {
+        ggml_fp16_t * const wdata = (ggml_fp16_t *) dst->data;
+
+        for (int64_t in = 0; in < N; in++) {
+            for (int64_t ioh = 0; ioh < OH; ioh++) {
+                for (int64_t iow = 0; iow < OW; iow++) {
+                    for (int64_t iic = ith; iic < IC; iic+=nth) {
+
+                        // micro kernel
+                        ggml_fp16_t * dst_data = wdata + (in*OH*OW + ioh*OW + iow)*(IC*KH*KW); // [IC, KH, KW]
+                        const float * const src_data = (float *)((char *) src1->data + in*nb13 + iic*nb12); // [IH, IW]
+
+                        for (int64_t ikh = 0; ikh < KH; ikh++) {
+                            for (int64_t ikw = 0; ikw < KW; ikw++) {
+                                const int64_t iiw = iow*s0 + ikw*d0 - p0;
+                                const int64_t iih = ioh*s1 + ikh*d1 - p1;
+
+                                if (!(iih < 0 || iih >= IH || iiw < 0 || iiw >= IW)) {
+                                    dst_data[iic*(KH*KW) + ikh*KW + ikw] = GGML_FP32_TO_FP16(src_data[iih*IW + iiw]);
+                                }
+                            }
+                        }
+                    }
+                }
+            }
+        }
+    }
+}
+
+// gemm: [N, OC, OH, OW] = [OC, IC * KH * KW] x [N*OH*OW, IC * KH * KW]
+// src0: [OC, IC, KH, KW]
+// src1: [N, OH, OW, IC * KH * KW]
+// result: [N, OC, OH, OW]
+static void ggml_compute_forward_conv_2d_stage_1_f16(
+        const struct ggml_compute_params * params,
+        const struct ggml_tensor * src0,
+        const struct ggml_tensor * src1,
+              struct ggml_tensor * dst) {
+    GGML_ASSERT(src0->type == GGML_TYPE_F16);
+    GGML_ASSERT(src1->type == GGML_TYPE_F16);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    int64_t t0 = ggml_perf_time_us();
+    UNUSED(t0);
+
+    if (params->type == GGML_TASK_INIT) {
+        return;
+    }
+
+    if (params->type == GGML_TASK_FINALIZE) {
+        return;
+    }
+
+    GGML_TENSOR_BINARY_OP_LOCALS;
+
+    GGML_ASSERT(nb00 == sizeof(ggml_fp16_t));
+    GGML_ASSERT(nb10 == sizeof(ggml_fp16_t));
+    GGML_ASSERT(nb0  == sizeof(float));
+
+    const int N = ne13;
+    const int OH = ne12;
+    const int OW = ne11;
+
+    const int OC = ne03;
+    const int IC = ne02;
+    const int KH = ne01;
+    const int KW = ne00;
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    int64_t m = OC;
+    int64_t n = OH * OW;
+    int64_t k = IC * KH * KW;
+
+    // [N, OC, OH, OW] = [OC, IC * KH * KW] x [N*OH*OW, IC * KH * KW]
+    for (int i = 0; i < N; i++) {
+        ggml_fp16_t * A = (ggml_fp16_t *)src0->data; // [m, k]
+        ggml_fp16_t * B = (ggml_fp16_t *)src1->data + i * m * k; // [n, k]
+        float * C = (float *)dst->data + i * m * n; // [m, n]
+
+        gemm_f16_out_f32(m, n, k, A, B, C, ith, nth);
+    }
+}
+
 static void ggml_compute_forward_conv_2d_f16_f32(
         const struct ggml_compute_params * params,
         const struct ggml_tensor * src0,
@@ -14552,16 +14744,40 @@ static void ggml_compute_forward_conv_2d_f16_f32(
     int64_t t0 = ggml_perf_time_us();
     UNUSED(t0);
 
-    GGML_TENSOR_BINARY_OP_LOCALS;
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    // src1: image [N, IC, IH, IW]
+    // src0: kernel [OC, IC, KH, KW]
+    // dst:  result [N, OC, OH, OW]
+    // ne12: IC
+    // ne0: OW
+    // ne1: OH
+    // nk0: KW
+    // nk1: KH
+    // ne13: N
+
+    const int N = ne13;
+    const int IC = ne12;
+    const int IH = ne11;
+    const int IW = ne10;
+
+    const int OC = ne03;
+    // const int IC = ne02;
+    const int KH = ne01;
+    const int KW = ne00;
+
+    const int OH = ne1;
+    const int OW = ne0;
 
     const int ith = params->ith;
     const int nth = params->nth;
 
-    const int nk0 = ne00;
-    const int nk1 = ne01;
+    // const int nk0 = ne00;
+    // const int nk1 = ne01;
 
     // size of the convolution row - the kernel size unrolled across all channels
-    const int ew0 = nk0*nk1*ne02;
+    // const int ew0 = nk0*nk1*ne02;
+    // ew0: IC*KH*KW
 
     const int32_t s0 = ((const int32_t*)(dst->op_params))[0];
     const int32_t s1 = ((const int32_t*)(dst->op_params))[1];
@@ -14577,24 +14793,27 @@ static void ggml_compute_forward_conv_2d_f16_f32(
         memset(params->wdata, 0, params->wsize);
 
         // prepare source data (src1)
+        // im2col: [N, IC, IH, IW] => [N*OH*OW, IC*KH*KW]
+
         {
             ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + 0;
 
-            for (int i13 = 0; i13 < ne13; i13++) {
-                for (int i12 = 0; i12 < ne12; i12++) {
-                    const float * const src = (float *)((char *) src1->data + i13*nb13 + i12*nb12);
-                    ggml_fp16_t * dst_data = wdata + i13*(ne1*ne0*ew0);
+            for (int in = 0; in < N; in++) {
+                for (int iic = 0; iic < IC; iic++) {
+                    for (int ioh = 0; ioh < OH; ioh++) {
+                        for (int iow = 0; iow < OW; iow++) {
 
-                    for (int i1 = 0; i1 < ne1; i1++) {
-                        for (int i0 = 0; i0 < ne0; i0++) {
-                            for (int ik1 = 0; ik1 < nk1; ik1++) {
-                                for (int ik0 = 0; ik0 < nk0; ik0++) {
-                                    const int idx0 = i0*s0 + ik0*d0 - p0;
-                                    const int idx1 = i1*s1 + ik1*d1 - p1;
+                            // micro kernel
+                            ggml_fp16_t * dst_data = wdata + (in*OH*OW + ioh*OW + iow)*(IC*KH*KW); // [IC, KH, KW]
+                            const float * const src_data = (float *)((char *) src1->data + in*nb13 + iic*nb12); // [IH, IW]
 
-                                    if (!(idx1 < 0 || idx1 >= ne11 || idx0 < 0 || idx0 >= ne10)) {
-                                        dst_data[(i1*ne0 + i0)*ew0 + i12*(nk0*nk1) + ik1*nk0 + ik0] =
-                                            GGML_FP32_TO_FP16(src[idx1*ne10 + idx0]);
+                            for (int ikh = 0; ikh < KH; ikh++) {
+                                for (int ikw = 0; ikw < KW; ikw++) {
+                                    const int iiw = iow*s0 + ikw*d0 - p0;
+                                    const int iih = ioh*s1 + ikh*d1 - p1;
+
+                                    if (!(iih < 0 || iih >= IH || iiw < 0 || iiw >= IW)) {
+                                        dst_data[iic*(KH*KW) + ikh*KW + ikw] = GGML_FP32_TO_FP16(src_data[iih*IW + iiw]);
                                     }
                                 }
                             }
@@ -14611,30 +14830,22 @@ static void ggml_compute_forward_conv_2d_f16_f32(
         return;
     }
 
-    // total patches in dst
-    const int np = ne2;
-
-    // patches per thread
-    const int dp = (np + nth - 1)/nth;
-
-    // patch range for this thread
-    const int ip0 = dp*ith;
-    const int ip1 = MIN(ip0 + dp, np);
-
     ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + 0;
+    // wdata: [N*OH*OW, IC*KH*KW]
+    // dst: result [N, OC, OH, OW]
+    // src0: kernel [OC, IC, KH, KW]
 
-    for (int i3 = 0; i3 < ne3; i3++) {
-        for (int i2 = ip0; i2 < ip1; i2++) {
-            float * dst_data = (float *)((char *) dst->data + i3*nb3 + i2*nb2);
+    int64_t m = OC;
+    int64_t n = OH * OW;
+    int64_t k = IC * KH * KW;
 
-            for (int i1 = 0; i1 < ne1; ++i1) {
-                for (int i0 = 0; i0 < ne0; ++i0) {
-                    ggml_vec_dot_f16(ew0, dst_data + i1*ne0 + i0,
-                            (ggml_fp16_t *) ((char *) src0->data + i2*nb03),
-                            (ggml_fp16_t *)                wdata + i3*nb3 + (i1*ne0 + i0)*ew0);
-                }
-            }
-        }
+    // [N, OC, OH, OW] = [OC, IC * KH * KW] x [N*OH*OW, IC * KH * KW]
+    for (int i = 0; i < N; i++) {
+        ggml_fp16_t * A = (ggml_fp16_t *)src0->data; // [m, k]
+        ggml_fp16_t * B = (ggml_fp16_t *)wdata + i * m * k; // [n, k]
+        float * C = (float *)dst->data + i * m * n; // [m * k]
+
+        gemm_f16_out_f32(m, n, k, A, B, C, ith, nth);
     }
 }
 
@@ -14660,6 +14871,48 @@ static void ggml_compute_forward_conv_2d(
     }
 }
 
+static void ggml_compute_forward_conv_2d_stage_0(
+        const struct ggml_compute_params * params,
+        const struct ggml_tensor * src0,
+        const struct ggml_tensor * src1,
+              struct ggml_tensor * dst) {
+    switch (src0->type) {
+        case GGML_TYPE_F16:
+            {
+                ggml_compute_forward_conv_2d_stage_0_f32(params, src0, src1, dst);
+            } break;
+        case GGML_TYPE_F32:
+            {
+                GGML_ASSERT(false);
+            } break;
+        default:
+            {
+                GGML_ASSERT(false);
+            } break;
+    }
+}
+
+static void ggml_compute_forward_conv_2d_stage_1(
+        const struct ggml_compute_params * params,
+        const struct ggml_tensor * src0,
+        const struct ggml_tensor * src1,
+              struct ggml_tensor * dst) {
+    switch (src0->type) {
+        case GGML_TYPE_F16:
+            {
+                ggml_compute_forward_conv_2d_stage_1_f16(params, src0, src1, dst);
+            } break;
+        case GGML_TYPE_F32:
+            {
+                GGML_ASSERT(false);
+            } break;
+        default:
+            {
+                GGML_ASSERT(false);
+            } break;
+    }
+}
+
 // ggml_compute_forward_conv_transpose_2d
 
 static void ggml_compute_forward_conv_transpose_2d(
@@ -14718,6 +14971,8 @@ static void ggml_compute_forward_conv_transpose_2d(
             }
         }
 
+        memset(dst->data, 0, ggml_nbytes(dst));
+
         return;
     }
 
@@ -16216,7 +16471,6 @@ static void ggml_compute_forward_add_rel_pos_f32(
     const int ip0 = dp*ith;
     const int ip1 = MIN(ip0 + dp, np);
 
-
     for (int64_t i13 = ip0; i13 < ip1; ++i13) {
         for (int64_t i12 = 0; i12 < ne12; ++i12) {
             for (int64_t i11 = 0; i11 < ne11; ++i11) {
@@ -16283,7 +16537,6 @@ static void ggml_compute_forward_map_unary_f32(
     }
 }
 
-
 static void ggml_compute_forward_map_unary(
         const struct ggml_compute_params * params,
         const struct ggml_tensor * src0,
@@ -16331,7 +16584,6 @@ static void ggml_compute_forward_map_binary_f32(
     }
 }
 
-
 static void ggml_compute_forward_map_binary(
         const struct ggml_compute_params * params,
         const struct ggml_tensor * src0,
@@ -16383,7 +16635,6 @@ static void ggml_compute_forward_map_custom2_f32(
     fun(dst, a, b);
 }
 
-
 // ggml_compute_forward_map_custom3
 
 static void ggml_compute_forward_map_custom3_f32(
@@ -16658,7 +16909,6 @@ static void ggml_compute_forward_cross_entropy_loss_back_f32(
         ggml_vec_sub_f32(nc, ds0, ds0, s1);
         ggml_vec_scale_f32(nc, ds0, d[0] / (float) nr);
 
-
 #ifndef NDEBUG
         for (int i = 0; i < nc; ++i) {
             assert(!isnan(ds0[i]));
@@ -16686,12 +16936,15 @@ static void ggml_compute_forward_cross_entropy_loss_back(
     }
 }
 
-
 /////////////////////////////////
 
 static void ggml_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor) {
     GGML_ASSERT(params);
 
+    if (tensor->op == GGML_OP_NONE) {
+        return;
+    }
+
 #ifdef GGML_USE_CUBLAS
     bool skip_cpu = ggml_cuda_compute_forward(params, tensor);
     if (skip_cpu) {
@@ -16894,6 +17147,14 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_conv_2d(params, tensor->src[0], tensor->src[1], tensor);
             } break;
+        case GGML_OP_CONV_2D_STAGE_0:
+            {
+                ggml_compute_forward_conv_2d_stage_0(params, tensor->src[0], tensor->src[1], tensor);
+            } break;
+        case GGML_OP_CONV_2D_STAGE_1:
+            {
+                ggml_compute_forward_conv_2d_stage_1(params, tensor->src[0], tensor->src[1], tensor);
+            } break;
         case GGML_OP_CONV_TRANSPOSE_2D:
             {
                 ggml_compute_forward_conv_transpose_2d(params, tensor->src[0], tensor->src[1], tensor);
@@ -17828,11 +18089,19 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
             {
                 GGML_ASSERT(false); // TODO: not implemented
             } break;
+        case GGML_OP_CONV_TRANSPOSE_1D:
+            {
+                GGML_ASSERT(false); // TODO: not implemented
+            } break;
         case GGML_OP_CONV_2D:
             {
                 GGML_ASSERT(false); // TODO: not implemented
             } break;
-        case GGML_OP_CONV_TRANSPOSE_1D:
+        case GGML_OP_CONV_2D_STAGE_0:
+            {
+                GGML_ASSERT(false); // TODO: not implemented
+            } break;
+        case GGML_OP_CONV_2D_STAGE_1:
             {
                 GGML_ASSERT(false); // TODO: not implemented
             } break;
@@ -18761,6 +19030,7 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
                     const int64_t ne0 = node->ne[0];
                     const int64_t ne1 = node->ne[1];
                     const int64_t ne2 = node->ne[2];
+                    const int64_t ne3 = node->ne[3];
                     const int64_t nk = ne00*ne01;
                     const int64_t ew0 = nk * ne02;
 
@@ -18771,7 +19041,8 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
 
                     if (node->src[0]->type == GGML_TYPE_F16 &&
                         node->src[1]->type == GGML_TYPE_F32) {
-                        cur = sizeof(ggml_fp16_t)*(ne0*ne1*ew0);
+                        // im2col: [N*OH*OW, IC*KH*KW]
+                        cur = sizeof(ggml_fp16_t)*(ne3*ne0*ne1*ew0);
                     } else if (node->src[0]->type == GGML_TYPE_F32 &&
                                node->src[1]->type == GGML_TYPE_F32) {
                         cur = sizeof(float)*      (ne10*ne11*ne12);
@@ -18781,6 +19052,14 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
 
                     work_size = MAX(work_size, cur);
                 } break;
+            case GGML_OP_CONV_2D_STAGE_0:
+                {
+                    n_tasks = n_threads;
+                } break;
+            case GGML_OP_CONV_2D_STAGE_1:
+                {
+                    n_tasks = n_threads;
+                } break;
             case GGML_OP_CONV_TRANSPOSE_2D:
                 {
                     n_tasks = n_threads;
@@ -19969,7 +20248,6 @@ static enum ggml_opt_result ggml_opt_adam(
 
         opt->loss_after = fx;
 
-
         // check convergence
         if (fabsf(fx - fx_prev[0])/fx < params.adam.eps_f) {
             GGML_PRINT_DEBUG("converged\n");
diff --git a/ggml.h b/ggml.h
index 3a0b8153c..ffebad35d 100644
--- a/ggml.h
+++ b/ggml.h
@@ -401,15 +401,16 @@ extern "C" {
         GGML_OP_ALIBI,
         GGML_OP_CLAMP,
         GGML_OP_CONV_1D,
-        GGML_OP_CONV_2D,
+        GGML_OP_CONV_1D_STAGE_0,  // internal
+        GGML_OP_CONV_1D_STAGE_1,  // internal
         GGML_OP_CONV_TRANSPOSE_1D,
+        GGML_OP_CONV_2D,
+        GGML_OP_CONV_2D_STAGE_0, // internal
+        GGML_OP_CONV_2D_STAGE_1, // internal
         GGML_OP_CONV_TRANSPOSE_2D,
         GGML_OP_POOL_1D,
         GGML_OP_POOL_2D,
 
-        GGML_OP_CONV_1D_STAGE_0,  // internal
-        GGML_OP_CONV_1D_STAGE_1,  // internal
-
         GGML_OP_UPSCALE, // nearest interpolate
 
         GGML_OP_FLASH_ATTN,
@@ -1020,9 +1021,9 @@ extern "C" {
             struct ggml_tensor  * b,
             float                 eps);
 
-    // A: n columns, m rows
-    // B: n columns, p rows  (i.e. we transpose it internally)
-    // result is m columns, p rows
+    // A: k columns, n rows => [ne03, ne02, n, k]
+    // B: k columns, m rows  (i.e. we transpose it internally) => [ne03 * x, ne02 * y, m, k]
+    // result is n columns, m rows => [ne03 * x, ne02 * y, m, n]
     GGML_API struct ggml_tensor * ggml_mul_mat(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
diff --git a/llama.cpp b/llama.cpp
index ff713c14e..e0794f90c 100644
--- a/llama.cpp
+++ b/llama.cpp
@@ -7593,7 +7593,7 @@ void llama_sample_grammar(struct llama_context * ctx, llama_token_data_array * c
         }
     }
 
-    const llama_token eos = llama_token_eos(ctx);
+    const llama_token eos = llama_token_eos(&ctx->model);
 
     std::vector<std::pair<std::vector<uint32_t>, llama_partial_utf8>> candidates_decoded;
     std::vector<llama_grammar_candidate>                              candidates_grammar;
@@ -7803,7 +7803,7 @@ llama_token llama_sample_token(struct llama_context * ctx, llama_token_data_arra
 void llama_grammar_accept_token(struct llama_context * ctx, struct llama_grammar * grammar, llama_token token) {
     const int64_t t_start_sample_us = ggml_time_us();
 
-    if (token == llama_token_eos(ctx)) {
+    if (token == llama_token_eos(&ctx->model)) {
         for (const auto & stack : grammar->stacks) {
             if (stack.empty()) {
                 return;
@@ -9033,7 +9033,7 @@ struct llama_context * llama_new_context_with_model(
             // build worst-case graph
             int n_tokens = (int)std::min(cparams.n_ctx, cparams.n_batch);
             int n_past = cparams.n_ctx - n_tokens;
-            llama_token token = llama_token_bos(ctx); // not actually used by llama_build_graph, but required to choose between token and embedding inputs graph
+            llama_token token = llama_token_bos(&ctx->model); // not actually used by llama_build_graph, but required to choose between token and embedding inputs graph
             ggml_cgraph * gf = llama_build_graph(*ctx, llama_batch_get_one(&token, n_tokens, n_past, 0));
 
 #ifdef GGML_USE_METAL
@@ -9794,43 +9794,44 @@ float * llama_get_embeddings(struct llama_context * ctx) {
     return ctx->embedding.data();
 }
 
-const char * llama_token_get_text(const struct llama_context * ctx, llama_token token) {
-    return ctx->model.vocab.id_to_token[token].text.c_str();
+const char * llama_token_get_text(const struct llama_model * model, llama_token token) {
+    return model->vocab.id_to_token[token].text.c_str();
 }
 
-float llama_token_get_score(const struct llama_context * ctx, llama_token token) {
-    return ctx->model.vocab.id_to_token[token].score;
+float llama_token_get_score(const struct llama_model * model, llama_token token) {
+    return model->vocab.id_to_token[token].score;
 }
 
-llama_token_type llama_token_get_type(const struct llama_context * ctx, llama_token token) {
-    return ctx->model.vocab.id_to_token[token].type;
+llama_token_type llama_token_get_type(const struct llama_model * model, llama_token token) {
+    return model->vocab.id_to_token[token].type;
 }
 
-llama_token llama_token_bos(const struct llama_context * ctx) {
-    return ctx->model.vocab.special_bos_id;
+llama_token llama_token_bos(const struct llama_model * model) {
+    return model->vocab.special_bos_id;
 }
 
-llama_token llama_token_eos(const struct llama_context * ctx) {
-    return ctx->model.vocab.special_eos_id;
+llama_token llama_token_eos(const struct llama_model * model) {
+    return model->vocab.special_eos_id;
 }
 
-llama_token llama_token_nl(const struct llama_context * ctx) {
-    return ctx->model.vocab.linefeed_id;
-}
-llama_token llama_token_prefix(const struct llama_context * ctx) {
-    return ctx->model.vocab.special_prefix_id;
+llama_token llama_token_nl(const struct llama_model * model) {
+    return model->vocab.linefeed_id;
 }
 
-llama_token llama_token_middle(const struct llama_context * ctx) {
-    return ctx->model.vocab.special_middle_id;
+llama_token llama_token_prefix(const struct llama_model * model) {
+    return model->vocab.special_prefix_id;
 }
 
-llama_token llama_token_suffix(const struct llama_context * ctx) {
-    return ctx->model.vocab.special_suffix_id;
+llama_token llama_token_middle(const struct llama_model * model) {
+    return model->vocab.special_middle_id;
 }
 
-llama_token llama_token_eot(const struct llama_context * ctx) {
-    return ctx->model.vocab.special_eot_id;
+llama_token llama_token_suffix(const struct llama_model * model) {
+    return model->vocab.special_suffix_id;
+}
+
+llama_token llama_token_eot(const struct llama_model * model) {
+    return model->vocab.special_eot_id;
 }
 
 int llama_tokenize(
diff --git a/llama.h b/llama.h
index 7c0f9ab1f..2048272bf 100644
--- a/llama.h
+++ b/llama.h
@@ -508,21 +508,22 @@ extern "C" {
     // Vocab
     //
 
-    LLAMA_API const char * llama_token_get_text(const struct llama_context * ctx, llama_token token);
+    LLAMA_API const char * llama_token_get_text(const struct llama_model * model, llama_token token);
 
-    LLAMA_API float llama_token_get_score(const struct llama_context * ctx, llama_token token);
+    LLAMA_API float llama_token_get_score(const struct llama_model * model, llama_token token);
 
-    LLAMA_API enum llama_token_type llama_token_get_type(const struct llama_context * ctx, llama_token token);
+    LLAMA_API enum llama_token_type llama_token_get_type(const struct llama_model * model, llama_token token);
 
     // Special tokens
-    LLAMA_API llama_token llama_token_bos(const struct llama_context * ctx);  // beginning-of-sentence
-    LLAMA_API llama_token llama_token_eos(const struct llama_context * ctx);  // end-of-sentence
-    LLAMA_API llama_token llama_token_nl (const struct llama_context * ctx);  // next-line
+    LLAMA_API llama_token llama_token_bos(const struct llama_model * model); // beginning-of-sentence
+    LLAMA_API llama_token llama_token_eos(const struct llama_model * model); // end-of-sentence
+    LLAMA_API llama_token llama_token_nl (const struct llama_model * model); // next-line
+
     // codellama infill tokens
-    LLAMA_API llama_token llama_token_prefix(const struct llama_context * ctx); // Beginning of infill prefix
-    LLAMA_API llama_token llama_token_middle(const struct llama_context * ctx); // Beginning of infill middle
-    LLAMA_API llama_token llama_token_suffix(const struct llama_context * ctx); // Beginning of infill suffix
-    LLAMA_API llama_token llama_token_eot   (const struct llama_context * ctx); // End of infill middle
+    LLAMA_API llama_token llama_token_prefix(const struct llama_model * model); // Beginning of infill prefix
+    LLAMA_API llama_token llama_token_middle(const struct llama_model * model); // Beginning of infill middle
+    LLAMA_API llama_token llama_token_suffix(const struct llama_model * model); // Beginning of infill suffix
+    LLAMA_API llama_token llama_token_eot   (const struct llama_model * model); // End of infill middle
 
     //
     // Tokenization
diff --git a/models/ggml-vocab-baichuan.gguf b/models/ggml-vocab-baichuan.gguf
new file mode 100644
index 000000000..7caaf8239
Binary files /dev/null and b/models/ggml-vocab-baichuan.gguf differ
diff --git a/models/ggml-vocab-gpt-neox.gguf b/models/ggml-vocab-gpt-neox.gguf
new file mode 100644
index 000000000..b9af16845
Binary files /dev/null and b/models/ggml-vocab-gpt-neox.gguf differ
diff --git a/models/ggml-vocab-refact.gguf b/models/ggml-vocab-refact.gguf
new file mode 100644
index 000000000..8f26cfb76
Binary files /dev/null and b/models/ggml-vocab-refact.gguf differ
diff --git a/models/ggml-vocab-starcoder.gguf b/models/ggml-vocab-starcoder.gguf
new file mode 100644
index 000000000..a52983fdb
Binary files /dev/null and b/models/ggml-vocab-starcoder.gguf differ
diff --git a/tests/CMakeLists.txt b/tests/CMakeLists.txt
index 1c73de0a3..6757ad1cc 100644
--- a/tests/CMakeLists.txt
+++ b/tests/CMakeLists.txt
@@ -28,10 +28,14 @@ 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-llama.cpp)
 llama_test_executable (test-tokenizer-1-llama test-tokenizer-1-llama.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama.gguf)
+llama_test_executable(test-tokenizer-1-baichuan test-tokenizer-1-llama.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-baichuan.gguf)
 llama_build_executable(test-tokenizer-1-bpe.cpp)
 llama_test_executable (test-tokenizer-1-falcon test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-falcon.gguf)
 llama_test_executable(test-tokenizer-1-aquila test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-aquila.gguf)
 llama_test_executable(test-tokenizer-1-mpt test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-mpt.gguf)
+llama_test_executable(test-tokenizer-1-gpt-neox test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-gpt-neox.gguf)
+llama_test_executable(test-tokenizer-1-refact test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-refact.gguf)
+llama_test_executable(test-tokenizer-1-starcoder test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-starcoder.gguf)
 llama_build_and_test_executable(test-grammar-parser.cpp)
 llama_build_and_test_executable(test-llama-grammar.cpp)
 llama_build_and_test_executable(test-grad0.cpp) # SLOW
diff --git a/tests/test-tokenizer-1-bpe.cpp b/tests/test-tokenizer-1-bpe.cpp
index 85a59a14d..386530f23 100644
--- a/tests/test-tokenizer-1-bpe.cpp
+++ b/tests/test-tokenizer-1-bpe.cpp
@@ -91,9 +91,19 @@ int main(int argc, char **argv) {
             }
         }
     }
-    // TODO: why doesn't this work for the full range of Unicodes?
+    // Restrict to assigned unicode planes
     // for (uint32_t cp = 0x10000; cp < 0x0010ffff; ++cp) {
-    for (uint32_t cp = 0x10000; cp < 0x00080000; ++cp) {
+    for (uint32_t cp = 0x10000; cp < 0x00040000; ++cp) {
+        std::string str = codepoint_to_utf8(cp);
+        std::vector<llama_token> tokens = llama_tokenize(ctx, str, false);
+        std::string check = llama_detokenize_bpe(ctx, tokens);
+        if (str != check) {
+            fprintf(stderr, "%s : error: codepoint %x detokenizes to '%s'(%zu) instead of '%s'(%zu)\n",
+                __func__, cp, check.c_str(), check.length(), str.c_str(), str.length());
+            return 4;
+        }
+    }
+    for (uint32_t cp = 0x000e0000; cp < 0x0010ffff; ++cp) {
         std::string str = codepoint_to_utf8(cp);
         std::vector<llama_token> tokens = llama_tokenize(ctx, str, false);
         std::string check = llama_detokenize_bpe(ctx, tokens);
@@ -103,7 +113,6 @@ int main(int argc, char **argv) {
             return 4;
         }
     }
-
     llama_free_model(model);
     llama_free(ctx);