server : check that the prompt fits in the slot's context (#10030)

ggml-ci

README.md

@@ -93,6 +93,7 @@ Typically finetunes of the base models below are supported as well.
 - [x] [FalconMamba Models](https://huggingface.co/collections/tiiuae/falconmamba-7b-66b9a580324dd1598b0f6d4a)
 - [x] [Jais](https://huggingface.co/inceptionai/jais-13b-chat)
 - [x] [Bielik-11B-v2.3](https://huggingface.co/collections/speakleash/bielik-11b-v23-66ee813238d9b526a072408a)
+- [x] [RWKV-6](https://github.com/BlinkDL/RWKV-LM)
 
 (instructions for supporting more models: [HOWTO-add-model.md](./docs/development/HOWTO-add-model.md))
 
@@ -173,6 +174,7 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 - [LARS - The LLM & Advanced Referencing Solution](https://github.com/abgulati/LARS) (AGPL)
 - [LLMUnity](https://github.com/undreamai/LLMUnity) (MIT)
 - [Llama Assistant](https://github.com/vietanhdev/llama-assistant) (GPL)
+- [PocketPal AI - An iOS and Android App](https://github.com/a-ghorbani/pocketpal-ai) (MIT)
 
 *(to have a project listed here, it should clearly state that it depends on `llama.cpp`)*
 

ci/run.sh

@@ -53,7 +53,7 @@ if [ ! -z ${GG_BUILD_SYCL} ]; then
         exit 1
     fi
 
-    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_SYCL=1 DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON"
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_SYCL=1 -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON"
 fi
 
 if [ ! -z ${GG_BUILD_VULKAN} ]; then

common/arg.cpp

@@ -1097,7 +1097,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         }
     ).set_examples({LLAMA_EXAMPLE_EMBEDDING, LLAMA_EXAMPLE_RETRIEVAL, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_POOLING"));
     add_opt(common_arg(
-        {"--attention"}, "{causal,non,causal}",
+        {"--attention"}, "{causal,non-causal}",
         "attention type for embeddings, use model default if unspecified",
         [](common_params & params, const std::string & value) {
             /**/ if (value == "causal") { params.attention_type = LLAMA_ATTENTION_TYPE_CAUSAL; }
@@ -1695,7 +1695,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_BENCH}));
     add_opt(common_arg(
         {"--embd-normalize"}, "N",
-        string_format("normalisation for embendings (default: %d) (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)", params.embd_normalize),
+        string_format("normalisation for embeddings (default: %d) (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)", params.embd_normalize),
         [](common_params & params, int value) {
             params.embd_normalize = value;
         }
@@ -1709,7 +1709,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_EMBEDDING}));
     add_opt(common_arg(
         {"--embd-separator"}, "STRING",
-        "separator of embendings (default \\n) for example \"<#sep#>\"",
+        "separator of embeddings (default \\n) for example \"<#sep#>\"",
         [](common_params & params, const std::string & value) {
             params.embd_sep = value;
         }

common/common.cpp

@@ -1035,7 +1035,7 @@ static ggml_type kv_cache_type_from_str(const std::string & s) {
         return GGML_TYPE_Q5_1;
     }
 
-    throw std::runtime_error("Invalid cache type: " + s);
+    throw std::runtime_error("Unsupported cache type: " + s);
 }
 
 struct llama_context_params common_context_params_to_llama(const common_params & params) {
@@ -1047,7 +1047,7 @@ struct llama_context_params common_context_params_to_llama(const common_params &
     cparams.n_ubatch          = params.n_ubatch;
     cparams.n_threads         = params.cpuparams.n_threads;
     cparams.n_threads_batch   = params.cpuparams_batch.n_threads == -1 ?
-                                    params.cpuparams.n_threads : params.cpuparams_batch.n_threads;
+                                params.cpuparams.n_threads : params.cpuparams_batch.n_threads;
     cparams.logits_all        = params.logits_all;
     cparams.embeddings        = params.embedding;
     cparams.rope_scaling_type = params.rope_scaling_type;

common/common.h

@@ -274,9 +274,9 @@ struct common_params {
 
     // embedding
     bool embedding         = false; // get only sentence embedding
-    int32_t embd_normalize = 2;     // normalisation for embendings (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)
+    int32_t embd_normalize = 2;     // normalisation for embeddings (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)
     std::string embd_out   = "";    // empty = default, "array" = [[],[]...], "json" = openai style, "json+" = same "json" + cosine similarity matrix
-    std::string embd_sep   = "\n";  // separator of embendings
+    std::string embd_sep   = "\n";  // separator of embeddings
     bool reranking         = false; // enable reranking support on server
 
     // server params

common/sampling.cpp

@@ -171,60 +171,46 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
                 params.penalize_nl,
                 params.ignore_eos));
 
-    if (params.temp > 0.0f) {
-        if (params.mirostat == 0) {
-            for (const auto & cnstr : params.samplers) {
-                switch (cnstr) {
-                    case COMMON_SAMPLER_TYPE_TOP_K:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_top_k    (params.top_k));
-                        break;
-                    case COMMON_SAMPLER_TYPE_TOP_P:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_top_p    (params.top_p, params.min_keep));
-                        break;
-                    case COMMON_SAMPLER_TYPE_MIN_P:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_min_p    (params.min_p, params.min_keep));
-                        break;
-                    case COMMON_SAMPLER_TYPE_XTC:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_xtc      (params.xtc_probability, params.xtc_threshold, params.min_keep, params.seed));
-                        break;
-                    case COMMON_SAMPLER_TYPE_TFS_Z:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_tail_free(params.tfs_z, params.min_keep));
-                        break;
-                    case COMMON_SAMPLER_TYPE_TYPICAL_P:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_typical  (params.typ_p, params.min_keep));
-                        break;
-                    case COMMON_SAMPLER_TYPE_TEMPERATURE:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_temp_ext (params.temp, params.dynatemp_range, params.dynatemp_exponent));
-                        break;
-                    case COMMON_SAMPLER_TYPE_INFILL:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_infill   (model));
-                        break;
-                    default:
-                        GGML_ASSERT(false && "unknown sampler type");
-                }
+    if (params.mirostat == 0) {
+        for (const auto & cnstr : params.samplers) {
+            switch (cnstr) {
+                case COMMON_SAMPLER_TYPE_TOP_K:
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_top_k    (params.top_k));
+                    break;
+                case COMMON_SAMPLER_TYPE_TOP_P:
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_top_p    (params.top_p, params.min_keep));
+                    break;
+                case COMMON_SAMPLER_TYPE_MIN_P:
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_min_p    (params.min_p, params.min_keep));
+                    break;
+                case COMMON_SAMPLER_TYPE_XTC:
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_xtc      (params.xtc_probability, params.xtc_threshold, params.min_keep, params.seed));
+                    break;
+                case COMMON_SAMPLER_TYPE_TFS_Z:
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_tail_free(params.tfs_z, params.min_keep));
+                    break;
+                case COMMON_SAMPLER_TYPE_TYPICAL_P:
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_typical  (params.typ_p, params.min_keep));
+                    break;
+                case COMMON_SAMPLER_TYPE_TEMPERATURE:
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_temp_ext (params.temp, params.dynatemp_range, params.dynatemp_exponent));
+                    break;
+                case COMMON_SAMPLER_TYPE_INFILL:
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_infill   (model));
+                    break;
+                default:
+                    GGML_ASSERT(false && "unknown sampler type");
             }
-            llama_sampler_chain_add(result->chain, llama_sampler_init_softmax());
-            llama_sampler_chain_add(result->chain, llama_sampler_init_dist(params.seed));
-        } else if (params.mirostat == 1) {
-            llama_sampler_chain_add(result->chain, llama_sampler_init_temp(params.temp));
-            llama_sampler_chain_add(result->chain, llama_sampler_init_mirostat(llama_n_vocab(model), params.seed, params.mirostat_tau, params.mirostat_eta, 100));
-        } else if (params.mirostat == 2) {
-            llama_sampler_chain_add(result->chain, llama_sampler_init_temp(params.temp));
-            llama_sampler_chain_add(result->chain, llama_sampler_init_mirostat_v2(params.seed, params.mirostat_tau, params.mirostat_eta));
-        } else {
-            GGML_ASSERT(false && "unknown mirostat version");
         }
+        llama_sampler_chain_add(result->chain, llama_sampler_init_dist(params.seed));
+    } else if (params.mirostat == 1) {
+        llama_sampler_chain_add(result->chain, llama_sampler_init_temp(params.temp));
+        llama_sampler_chain_add(result->chain, llama_sampler_init_mirostat(llama_n_vocab(model), params.seed, params.mirostat_tau, params.mirostat_eta, 100));
+    } else if (params.mirostat == 2) {
+        llama_sampler_chain_add(result->chain, llama_sampler_init_temp(params.temp));
+        llama_sampler_chain_add(result->chain, llama_sampler_init_mirostat_v2(params.seed, params.mirostat_tau, params.mirostat_eta));
     } else {
-        if (params.n_probs > 0) {
-            // some use cases require to sample greedily, but still obtain the probabilities of the top tokens
-            // ref: https://github.com/ggerganov/llama.cpp/pull/9605
-            //
-            // the following will not produce exactly the same probs as applyging softmax to the full vocabulary, but
-            // it is much faster, since we avoid sorting all tokens and should give a good approximation
-            llama_sampler_chain_add(result->chain, llama_sampler_init_top_k(params.n_probs));
-            llama_sampler_chain_add(result->chain, llama_sampler_init_softmax());
-        }
-        llama_sampler_chain_add(result->chain, llama_sampler_init_greedy());
+        GGML_ASSERT(false && "unknown mirostat version");
     }
 
     return result;

convert_hf_to_gguf.py

@@ -573,6 +573,9 @@ class Model:
         if chkhsh == "0876d13b50744004aa9aeae05e7b0647eac9d801b5ba4668afc01e709c15e19f":
             # ref: https://huggingface.co/BAAI/bge-small-en-v1.5
             res = "bert-bge"
+        if chkhsh == "8e62295832751ca1e8f92f2226f403dea30dc5165e448b5bfa05af5340c64ec7":
+            # ref: https://huggingface.co/BAAI/bge-large-zh-v1.5
+            res = "bert-bge-large"
         if chkhsh == "b6dc8df998e1cfbdc4eac8243701a65afe638679230920b50d6f17d81c098166":
             # ref: https://huggingface.co/mosaicml/mpt-7b
             res = "mpt"
@@ -2864,6 +2867,9 @@ class Rwkv6Model(Model):
         self.gguf_writer.add_token_list(tokens)
         self.gguf_writer.add_token_types(toktypes)
         special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=False)
+        special_vocab.chat_template = "rwkv-world"
+        # hack: Add '\n\n' as the EOT token to make it chat normally
+        special_vocab._set_special_token("eot", 261)
         special_vocab.add_to_gguf(self.gguf_writer)
 
     def set_gguf_parameters(self):

convert_hf_to_gguf_update.py

@@ -72,6 +72,7 @@ models = [
     {"name": "deepseek-coder", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/deepseek-ai/deepseek-coder-6.7b-base", },
     {"name": "falcon",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tiiuae/falcon-7b", },
     {"name": "bert-bge",       "tokt": TOKENIZER_TYPE.WPM, "repo": "https://huggingface.co/BAAI/bge-small-en-v1.5", },
+    {"name": "bert-bge-large", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/BAAI/bge-large-zh-v1.5", },
     {"name": "mpt",            "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mosaicml/mpt-7b", },
     {"name": "starcoder",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/bigcode/starcoder2-3b", },
     {"name": "gpt-2",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/openai-community/gpt2", },

convert_lora_to_gguf.py

@@ -348,6 +348,9 @@ if __name__ == '__main__':
                         if ".base_layer.weight" in name:
                             continue
                         logger.error(f"Unexpected name '{name}': Not a lora_A or lora_B tensor")
+                        if ".embed_tokens.weight" in name or ".lm_head.weight" in name:
+                            logger.error("Embeddings is present in the adapter. This can be due to new tokens added during fine tuning")
+                            logger.error("Hint: if you are using TRL, make sure not to call setup_chat_format()")
                         sys.exit(1)
 
                     if base_name in tensor_map:

examples/llama.swiftui/llama.cpp.swift/LibLlama.swift

@@ -46,7 +46,6 @@ actor LlamaContext {
         let sparams = llama_sampler_chain_default_params()
         self.sampling = llama_sampler_chain_init(sparams)
         llama_sampler_chain_add(self.sampling, llama_sampler_init_temp(0.4))
-        llama_sampler_chain_add(self.sampling, llama_sampler_init_softmax())
         llama_sampler_chain_add(self.sampling, llama_sampler_init_dist(1234))
     }
 

examples/llama.vim

@@ -0,0 +1,783 @@
+" LLM-based text completion using llama.cpp
+"
+" requires:
+"
+"   - neovim or vim
+"   - curl
+"   - llama.cpp server instance
+"   - FIM-compatible model
+"
+" sample config:
+"
+"   - Tab       - accept the current suggestion
+"   - Shift+Tab - accept just the first line of the suggestion
+"   - Ctrl+F    - toggle FIM completion manually
+"
+" make symlink or copy this file to ~/.config/nvim/autoload/llama.vim
+"
+" start the llama.cpp server with a FIM-compatible model. for example:
+"
+"   $ llama-server -m {model.gguf} --port 8012 -ngl 99 -fa -dt 0.1 --ubatch-size 512 --batch-size 1024 --cache-reuse 256
+"
+"   --batch-size [512, model max context]
+"
+"     adjust the batch size to control how much of the provided local context will be used during the inference
+"     lower values will use smaller part of the context around the cursor, which will result in faster processing
+"
+"   --ubatch-size [64, 2048]
+"
+"     chunks the batch into smaller chunks for faster processing
+"     depends on the specific hardware. use llama-bench to profile and determine the best size
+"
+"   --cache-reuse (ge:llama_config.n_predict, 1024]
+"
+"     this should be either 0 (disabled) or strictly larger than g:llama_config.n_predict
+"     using non-zero value enables context reuse on the server side which dramatically improves the performance at
+"     large contexts. a value of 256 should be good for all cases
+"
+" run this once to initialise llama.vim:
+"
+"   :call llama#init()
+"
+" more info: https://github.com/ggerganov/llama.cpp/pull/9787
+"
+
+" colors (adjust to your liking)
+highlight llama_hl_hint guifg=#ff772f ctermfg=202
+highlight llama_hl_info guifg=#77ff2f ctermfg=119
+
+" general parameters:
+"
+"   endpoint:         llama.cpp server endpoint
+"   n_prefix:         number of lines before the cursor location to include in the local prefix
+"   n_suffix:         number of lines after  the cursor location to include in the local suffix
+"   n_predict:        max number of tokens to predict
+"   t_max_prompt_ms:  max alloted time for the prompt processing (TODO: not yet supported)
+"   t_max_predict_ms: max alloted time for the prediction
+"   show_info:        show extra info about the inference (0 - disabled, 1 - statusline, 2 - inline)
+"   auto_fim:         trigger FIM completion automatically on cursor movement
+"   max_line_suffix:  do not auto-trigger FIM completion if there are more than this number of characters to the right of the cursor
+"
+" ring buffer of chunks, accumulated with time upon:
+"
+"  - completion request
+"  - yank
+"  - entering a buffer
+"  - leaving a buffer
+"  - writing a file
+"
+" parameters for the ring-buffer with extra context:
+"
+"   ring_n_chunks:    max number of chunks to pass as extra context to the server (0 to disable)
+"   ring_chunk_size:  max size of the chunks (in number of lines)
+"                     note: adjust these numbers so that you don't overrun your context
+"                           at ring_n_chunks = 64 and ring_chunk_size = 64 you need ~32k context
+"   ring_scope:       the range around the cursor position (in number of lines) for gathering chunks after FIM
+"   ring_update_ms:   how often to process queued chunks in normal mode
+"
+let s:default_config = {
+    \ 'endpoint':         'http://127.0.0.1:8012/infill',
+    \ 'n_prefix':         256,
+    \ 'n_suffix':         64,
+    \ 'n_predict':        128,
+    \ 't_max_prompt_ms':  500,
+    \ 't_max_predict_ms': 3000,
+    \ 'show_info':        2,
+    \ 'auto_fim':         v:true,
+    \ 'max_line_suffix':  8,
+    \ 'ring_n_chunks':    64,
+    \ 'ring_chunk_size':  64,
+    \ 'ring_scope':       1024,
+    \ 'ring_update_ms':   1000,
+    \ }
+
+let g:llama_config = get(g:, 'llama_config', s:default_config)
+
+function! s:get_indent(str)
+    let l:count = 0
+    for i in range(len(a:str))
+        if a:str[i] == "\t"
+            let l:count += &tabstop - 1
+        else
+            break
+        endif
+    endfor
+    return l:count
+endfunction
+
+function! s:rand(i0, i1) abort
+    return a:i0 + rand() % (a:i1 - a:i0 + 1)
+endfunction
+
+function! llama#init()
+    if !executable('curl')
+        echohl WarningMsg
+        echo 'llama.vim requires the "curl" command to be available'
+        echohl None
+        return
+    endif
+
+    let s:pos_x = 0 " cursor position upon start of completion
+    let s:pos_y = 0
+
+    let s:line_cur = ''
+
+    let s:line_cur_prefix = ''
+    let s:line_cur_suffix = ''
+
+    let s:ring_chunks = [] " current set of chunks used as extra context
+    let s:ring_queued = [] " chunks that are queued to be sent for processing
+    let s:ring_n_evict = 0
+
+    let s:hint_shown = v:false
+    let s:pos_y_pick = -9999 " last y where we picked a chunk
+    let s:pos_dx = 0
+    let s:content = []
+    let s:can_accept = v:false
+
+    let s:timer_fim = -1
+    let s:t_fim_start = reltime() " used to measure total FIM time
+    let s:t_last_move = reltime() " last time the cursor moved
+
+    let s:current_job = v:null
+
+    let s:ghost_text_nvim = exists('*nvim_buf_get_mark')
+    let s:ghost_text_vim = has('textprop')
+
+    if s:ghost_text_vim
+        let s:hlgroup_hint = 'llama_hl_hint'
+        let s:hlgroup_info = 'llama_hl_info'
+
+        if empty(prop_type_get(s:hlgroup_hint))
+            call prop_type_add(s:hlgroup_hint, {'highlight': s:hlgroup_hint})
+        endif
+        if empty(prop_type_get(s:hlgroup_info))
+            call prop_type_add(s:hlgroup_info, {'highlight': s:hlgroup_info})
+        endif
+    endif
+
+    augroup llama
+        autocmd!
+        autocmd InsertEnter     * inoremap <expr> <silent> <C-F> llama#fim_inline(v:false)
+        autocmd InsertLeavePre  * call llama#fim_cancel()
+
+        autocmd CursorMoved     * call s:on_move()
+        autocmd CursorMovedI    * call s:on_move()
+        autocmd CompleteChanged * call llama#fim_cancel()
+
+        if g:llama_config.auto_fim
+            autocmd CursorMovedI * call llama#fim(v:true)
+        endif
+
+        " gather chunks upon yanking
+        autocmd TextYankPost    * if v:event.operator ==# 'y' | call s:pick_chunk(v:event.regcontents, v:false, v:true) | endif
+
+        " gather chunks upon entering/leaving a buffer
+        autocmd BufEnter        * call timer_start(100, {-> s:pick_chunk(getline(max([1, line('.') - g:llama_config.ring_chunk_size/2]), min([line('.') + g:llama_config.ring_chunk_size/2, line('$')])), v:true, v:true)})
+        autocmd BufLeave        * call                      s:pick_chunk(getline(max([1, line('.') - g:llama_config.ring_chunk_size/2]), min([line('.') + g:llama_config.ring_chunk_size/2, line('$')])), v:true, v:true)
+
+        " gather chunk upon saving the file
+        autocmd BufWritePost    * call s:pick_chunk(getline(max([1, line('.') - g:llama_config.ring_chunk_size/2]), min([line('.') + g:llama_config.ring_chunk_size/2, line('$')])), v:true, v:true)
+    augroup END
+
+    silent! call llama#fim_cancel()
+
+    " init background update of the ring buffer
+    if g:llama_config.ring_n_chunks > 0
+        call s:ring_update()
+    endif
+endfunction
+
+" compute how similar two chunks of text are
+" 0 - no similarity, 1 - high similarity
+" TODO: figure out something better
+function! s:chunk_sim(c0, c1)
+    let l:lines0 = len(a:c0)
+    let l:lines1 = len(a:c1)
+
+    let l:common = 0
+
+    for l:line0 in a:c0
+        for l:line1 in a:c1
+            if l:line0 == l:line1
+                let l:common += 1
+                break
+            endif
+        endfor
+    endfor
+
+    return 2.0 * l:common / (l:lines0 + l:lines1)
+endfunction
+
+" pick a random chunk of size g:llama_config.ring_chunk_size from the provided text and queue it for processing
+"
+" no_mod   - do not pick chunks from buffers with pending changes
+" do_evict - evict chunks that are very similar to the new one
+"
+function! s:pick_chunk(text, no_mod, do_evict)
+    " do not pick chunks from buffers with pending changes or buffers that are not files
+    if a:no_mod && (getbufvar(bufnr('%'), '&modified') || !buflisted(bufnr('%')) || !filereadable(expand('%')))
+        return
+    endif
+
+    " if the extra context option is disabled - do nothing
+    if g:llama_config.ring_n_chunks <= 0
+        return
+    endif
+
+    " don't pick very small chunks
+    if len(a:text) < 3
+        return
+    endif
+
+    if len(a:text) + 1 < g:llama_config.ring_chunk_size
+        let l:chunk = a:text
+    else
+        let l:l0 = s:rand(0, max([0, len(a:text) - g:llama_config.ring_chunk_size/2]))
+        let l:l1 = min([l:l0 + g:llama_config.ring_chunk_size/2, len(a:text)])
+
+        let l:chunk = a:text[l:l0:l:l1]
+    endif
+
+    let l:chunk_str = join(l:chunk, "\n") . "\n"
+
+    " check if this chunk is already added
+    let l:exist = v:false
+
+    for i in range(len(s:ring_chunks))
+        if s:ring_chunks[i].data == l:chunk
+            let l:exist = v:true
+            break
+        endif
+    endfor
+
+    for i in range(len(s:ring_queued))
+        if s:ring_queued[i].data == l:chunk
+            let l:exist = v:true
+            break
+        endif
+    endfor
+
+    if l:exist
+        return
+    endif
+
+    " evict queued chunks that are very similar to the new one
+    for i in range(len(s:ring_queued) - 1, 0, -1)
+        if s:chunk_sim(s:ring_queued[i].data, l:chunk) > 0.9
+            if a:do_evict
+                call remove(s:ring_queued, i)
+                let s:ring_n_evict += 1
+            else
+                return
+            endif
+        endif
+    endfor
+
+    " also from s:ring_chunks
+    for i in range(len(s:ring_chunks) - 1, 0, -1)
+        if s:chunk_sim(s:ring_chunks[i].data, l:chunk) > 0.9
+            if a:do_evict
+                call remove(s:ring_chunks, i)
+                let s:ring_n_evict += 1
+            else
+                return
+            endif
+        endif
+    endfor
+
+    " TODO: become parameter ?
+    if len(s:ring_queued) == 16
+        call remove(s:ring_queued, 0)
+    endif
+
+    call add(s:ring_queued, {'data': l:chunk, 'str': l:chunk_str, 'time': reltime(), 'filename': expand('%')})
+
+    "let &statusline = 'extra context: ' . len(s:ring_chunks) . ' / ' . len(s:ring_queued)
+endfunction
+
+" picks a queued chunk, sends it for processing and adds it to s:ring_chunks
+" called every g:llama_config.ring_update_ms
+function! s:ring_update()
+    call timer_start(g:llama_config.ring_update_ms, {-> s:ring_update()})
+
+    " update only if in normal mode or if the cursor hasn't moved for a while
+    if mode() !=# 'n' && reltimefloat(reltime(s:t_last_move)) < 3.0
+        return
+    endif
+
+    if len(s:ring_queued) == 0
+        return
+    endif
+
+    " move the first queued chunk to the ring buffer
+    if len(s:ring_chunks) == g:llama_config.ring_n_chunks
+        call remove(s:ring_chunks, 0)
+    endif
+
+    call add(s:ring_chunks, remove(s:ring_queued, 0))
+
+    "let &statusline = 'updated context: ' . len(s:ring_chunks) . ' / ' . len(s:ring_queued)
+
+    " send asynchronous job with the new extra context so that it is ready for the next FIM
+    let l:extra_context = []
+    for l:chunk in s:ring_chunks
+        call add(l:extra_context, {
+            \ 'text':     l:chunk.str,
+            \ 'time':     l:chunk.time,
+            \ 'filename': l:chunk.filename
+            \ })
+    endfor
+
+    " no samplers needed here
+    let l:request = json_encode({
+        \ 'input_prefix':     "",
+        \ 'input_suffix':     "",
+        \ 'input_extra':      l:extra_context,
+        \ 'prompt':           "",
+        \ 'n_predict':        1,
+        \ 'temperature':      0.0,
+        \ 'stream':           v:false,
+        \ 'samplers':         ["temperature"],
+        \ 'cache_prompt':     v:true,
+        \ 't_max_prompt_ms':  1,
+        \ 't_max_predict_ms': 1
+        \ })
+
+    let l:curl_command = [
+        \ "curl",
+        \ "--silent",
+        \ "--no-buffer",
+        \ "--request", "POST",
+        \ "--url", g:llama_config.endpoint,
+        \ "--header", "Content-Type: application/json",
+        \ "--data", l:request
+        \ ]
+
+    " no callbacks because we don't need to process the response
+    if s:ghost_text_nvim
+        call jobstart(l:curl_command, {})
+    elseif s:ghost_text_vim
+        call job_start(l:curl_command, {})
+    endif
+endfunction
+
+" necessary for 'inoremap <expr>'
+function! llama#fim_inline(is_auto) abort
+    call llama#fim(a:is_auto)
+    return ''
+endfunction
+
+" the main FIM call
+" takes local context around the cursor and sends it together with the extra context to the server for completion
+function! llama#fim(is_auto) abort
+    " we already have a suggestion for the current cursor position
+    if s:hint_shown && !a:is_auto
+        call llama#fim_cancel()
+        return
+    endif
+
+    call llama#fim_cancel()
+
+    " avoid sending repeated requests too fast
+    if reltimefloat(reltime(s:t_fim_start)) < 0.6
+        if s:timer_fim != -1
+            call timer_stop(s:timer_fim)
+            let s:timer_fim = -1
+        endif
+
+        let s:t_fim_start = reltime()
+        let s:timer_fim = timer_start(600, {-> llama#fim(v:true)})
+        return
+    endif
+
+    let s:t_fim_start = reltime()
+
+    let s:content = []
+    let s:can_accept = v:false
+
+    let s:pos_x = col('.') - 1
+    let s:pos_y = line('.')
+    let l:max_y = line('$')
+
+    let l:lines_prefix = getline(max([1, s:pos_y - g:llama_config.n_prefix]), s:pos_y - 1)
+    let l:lines_suffix = getline(s:pos_y + 1, min([l:max_y, s:pos_y + g:llama_config.n_suffix]))
+
+    let s:line_cur = getline('.')
+
+    let s:line_cur_prefix = strpart(s:line_cur, 0, s:pos_x)
+    let s:line_cur_suffix = strpart(s:line_cur, s:pos_x)
+
+    if a:is_auto && len(s:line_cur_suffix) > g:llama_config.max_line_suffix
+        return
+    endif
+
+    let l:prefix = ""
+        \ . join(l:lines_prefix, "\n")
+        \ . "\n"
+
+    let l:prompt = ""
+        \ . s:line_cur_prefix
+
+    let l:suffix = ""
+        \ . s:line_cur_suffix
+        \ . "\n"
+        \ . join(l:lines_suffix, "\n")
+        \ . "\n"
+
+    " prepare the extra context data
+    let l:extra_context = []
+    for l:chunk in s:ring_chunks
+        call add(l:extra_context, {
+            \ 'text':     l:chunk.str,
+            \ 'time':     l:chunk.time,
+            \ 'filename': l:chunk.filename
+            \ })
+    endfor
+
+    " the indentation of the current line
+    let l:indent = strlen(matchstr(s:line_cur_prefix, '^\s*'))
+
+    let l:request = json_encode({
+        \ 'input_prefix':     l:prefix,
+        \ 'input_suffix':     l:suffix,
+        \ 'input_extra':      l:extra_context,
+        \ 'prompt':           l:prompt,
+        \ 'n_predict':        g:llama_config.n_predict,
+        \ 'n_indent':         l:indent,
+        \ 'top_k':            40,
+        \ 'top_p':            0.99,
+        \ 'stream':           v:false,
+        \ 'samplers':         ["top_k", "top_p", "infill"],
+        \ 'cache_prompt':     v:true,
+        \ 't_max_prompt_ms':  g:llama_config.t_max_prompt_ms,
+        \ 't_max_predict_ms': g:llama_config.t_max_predict_ms
+        \ })
+
+    let l:curl_command = [
+        \ "curl",
+        \ "--silent",
+        \ "--no-buffer",
+        \ "--request", "POST",
+        \ "--url", g:llama_config.endpoint,
+        \ "--header", "Content-Type: application/json",
+        \ "--data", l:request
+        \ ]
+
+    if s:current_job != v:null
+        if s:ghost_text_nvim
+            call jobstop(s:current_job)
+        elseif s:ghost_text_vim
+            call job_stop(s:current_job)
+        endif
+    endif
+
+    " send the request asynchronously
+    if s:ghost_text_nvim
+        let s:current_job = jobstart(l:curl_command, {
+            \ 'on_stdout': function('s:fim_on_stdout', [s:pos_x, s:pos_y, a:is_auto]),
+            \ 'on_exit':   function('s:fim_on_exit'),
+            \ 'stdout_buffered': v:true
+            \ })
+    elseif s:ghost_text_vim
+        let s:current_job = job_start(l:curl_command, {
+            \ 'out_cb': function('s:fim_on_stdout', [s:pos_x, s:pos_y, a:is_auto]),
+            \ 'exit_cb':   function('s:fim_on_exit')
+            \ })
+    endif
+
+    " TODO: per-file location
+    let l:delta_y = abs(s:pos_y - s:pos_y_pick)
+
+    " gather some extra context nearby and process it in the background
+    " only gather chunks if the cursor has moved a lot
+    " TODO: something more clever? reranking?
+    if a:is_auto && l:delta_y > 32
+        " expand the prefix even further
+        call s:pick_chunk(getline(max([1,       s:pos_y - g:llama_config.ring_scope]), max([1,       s:pos_y - g:llama_config.n_prefix])), v:false, v:false)
+
+        " pick a suffix chunk
+        call s:pick_chunk(getline(min([l:max_y, s:pos_y + g:llama_config.n_suffix]),   min([l:max_y, s:pos_y + g:llama_config.n_suffix + g:llama_config.ring_chunk_size])), v:false, v:false)
+
+        let s:pos_y_pick = s:pos_y
+    endif
+endfunction
+
+" if first_line == v:true accept only the first line of the response
+function! llama#fim_accept(first_line)
+    " insert the suggestion at the cursor location
+    if s:can_accept && len(s:content) > 0
+        call setline(s:pos_y, s:line_cur[:(s:pos_x - 1)] . s:content[0])
+        if len(s:content) > 1
+            if !a:first_line
+                call append(s:pos_y, s:content[1:-1])
+            endif
+        endif
+
+        " move the cursor to the end of the accepted text
+        if !a:first_line && len(s:content) > 1
+            call cursor(s:pos_y + len(s:content) - 1, s:pos_x + s:pos_dx + 1)
+        else
+            call cursor(s:pos_y, s:pos_x + len(s:content[0]))
+        endif
+    endif
+
+    call llama#fim_cancel()
+endfunction
+
+function! llama#fim_cancel()
+    let s:hint_shown = v:false
+
+    " clear the virtual text
+    let l:bufnr = bufnr('%')
+
+    if s:ghost_text_nvim
+        let l:id_vt_fim = nvim_create_namespace('vt_fim')
+        call nvim_buf_clear_namespace(l:bufnr, l:id_vt_fim,  0, -1)
+    elseif s:ghost_text_vim
+        call prop_remove({'type': s:hlgroup_hint, 'all': v:true})
+        call prop_remove({'type': s:hlgroup_info, 'all': v:true})
+    endif
+
+    " remove the mappings
+    silent! iunmap <buffer> <Tab>
+    silent! iunmap <buffer> <S-Tab>
+    silent! iunmap <buffer> <Esc>
+endfunction
+
+function! s:on_move()
+    let s:t_last_move = reltime()
+
+    call llama#fim_cancel()
+endfunction
+
+" callback that processes the FIM result from the server and displays the suggestion
+function! s:fim_on_stdout(pos_x, pos_y, is_auto, job_id, data, event = v:null)
+    if s:ghost_text_nvim
+        let l:raw = join(a:data, "\n")
+    elseif s:ghost_text_vim
+        let l:raw = a:data
+    endif
+
+    if len(l:raw) == 0
+        return
+    endif
+
+    if a:pos_x != col('.') - 1 || a:pos_y != line('.')
+        return
+    endif
+
+    " show the suggestion only in insert mode
+    if mode() !=# 'i'
+        return
+    endif
+
+    let s:pos_x = a:pos_x
+    let s:pos_y = a:pos_y
+
+    let s:can_accept = v:true
+    let l:has_info   = v:false
+
+    if s:can_accept && v:shell_error
+        if !a:is_auto
+            call add(s:content, "<| curl error: is the server on? |>")
+        endif
+        let s:can_accept = v:false
+    endif
+
+    let l:n_prompt    = 0
+    let l:t_prompt_ms = 1.0
+    let l:s_prompt    = 0
+
+    let l:n_predict    = 0
+    let l:t_predict_ms = 1.0
+    let l:s_predict    = 0
+
+    " get the generated suggestion
+    if s:can_accept
+        let l:response = json_decode(l:raw)
+
+        for l:part in split(get(l:response, 'content', ''), "\n", 1)
+            call add(s:content, l:part)
+        endfor
+
+        " remove trailing new lines
+        while len(s:content) > 0 && s:content[-1] == ""
+            call remove(s:content, -1)
+        endwhile
+
+        let l:generation_settings = get(l:response, 'generation_settings', {})
+        let l:n_ctx = get(l:generation_settings, 'n_ctx', 0)
+
+        let l:n_cached  = get(l:response, 'tokens_cached', 0)
+        let l:truncated = get(l:response, 'truncated', v:false)
+
+        " if response.timings is available
+        if len(get(l:response, 'timings', {})) > 0
+            let l:has_info = v:true
+            let l:timings  = get(l:response, 'timings', {})
+
+            let l:n_prompt    = get(l:timings, 'prompt_n', 0)
+            let l:t_prompt_ms = get(l:timings, 'prompt_ms', 1)
+            let l:s_prompt    = get(l:timings, 'prompt_per_second', 0)
+
+            let l:n_predict    = get(l:timings, 'predicted_n', 0)
+            let l:t_predict_ms = get(l:timings, 'predicted_ms', 1)
+            let l:s_predict    = get(l:timings, 'predicted_per_second', 0)
+        endif
+    endif
+
+    if len(s:content) == 0
+        call add(s:content, "")
+        let s:can_accept = v:false
+    endif
+
+    if len(s:content) == 0
+        return
+    endif
+
+    " NOTE: the following is logic for discarding predictions that repeat existing text
+    "       the code is quite ugly and there is very likely a simpler and more canonical way to implement this
+    "
+    "       still, I wonder if there is some better way that avoids having to do these special hacks?
+    "       on one hand, the LLM 'sees' the contents of the file before we start editing, so it is normal that it would
+    "       start generating whatever we have given it via the extra context. but on the other hand, it's not very
+    "       helpful to re-generate the same code that is already there
+
+    " truncate the suggestion if the first line is empty
+    if len(s:content) == 1 && s:content[0] == ""
+        let s:content = [""]
+    endif
+
+    " ... and the next lines are repeated
+    if len(s:content) > 1 && s:content[0] == "" && s:content[1:] == getline(s:pos_y + 1, s:pos_y + len(s:content) - 1)
+        let s:content = [""]
+    endif
+
+    " truncate the suggestion if it repeats the suffix
+    if len(s:content) == 1 && s:content[0] == s:line_cur_suffix
+        let s:content = [""]
+    endif
+
+    " find the first non-empty line (strip whitespace)
+    let l:cmp_y = s:pos_y + 1
+    while l:cmp_y < line('$') && getline(l:cmp_y) =~? '^\s*$'
+        let l:cmp_y += 1
+    endwhile
+
+    if (s:line_cur_prefix . s:content[0]) == getline(l:cmp_y)
+        " truncate the suggestion if it repeats the next line
+        if len(s:content) == 1
+            let s:content = [""]
+        endif
+
+        " ... or if the second line of the suggestion is the prefix of line l:cmp_y + 1
+        if len(s:content) == 2 && s:content[-1] == getline(l:cmp_y + 1)[:len(s:content[-1]) - 1]
+            let s:content = [""]
+        endif
+
+        " ... or if the middle chunk of lines of the suggestion is the same as [l:cmp_y + 1, l:cmp_y + len(s:content) - 1)
+        if len(s:content) > 2 && join(s:content[1:-1], "\n") == join(getline(l:cmp_y + 1, l:cmp_y + len(s:content) - 1), "\n")
+            let s:content = [""]
+        endif
+    endif
+
+    " keep only lines that have the same or larger whitespace prefix as s:line_cur_prefix
+    "let l:indent = strlen(matchstr(s:line_cur_prefix, '^\s*'))
+    "for i in range(1, len(s:content) - 1)
+    "    if strlen(matchstr(s:content[i], '^\s*')) < l:indent
+    "        let s:content = s:content[:i - 1]
+    "        break
+    "    endif
+    "endfor
+
+    let s:pos_dx = len(s:content[-1])
+
+    let s:content[-1] .= s:line_cur_suffix
+
+    call llama#fim_cancel()
+
+    " display virtual text with the suggestion
+    let l:bufnr = bufnr('%')
+
+    if s:ghost_text_nvim
+        let l:id_vt_fim = nvim_create_namespace('vt_fim')
+    endif
+
+    " construct the info message
+    if g:llama_config.show_info > 0 && l:has_info
+        let l:prefix = '   '
+
+        if l:truncated
+            let l:info = printf("%s | WARNING: the context is full: %d / %d, increase the server context size or reduce g:llama_config.ring_n_chunks",
+                \ g:llama_config.show_info == 2 ? l:prefix : 'llama.vim',
+                \ l:n_cached, l:n_ctx
+                \ )
+        else
+            let l:info = printf("%s | c: %d / %d, r: %d / %d, e: %d, q: %d / 16 | p: %d (%.2f ms, %.2f t/s) | g: %d (%.2f ms, %.2f t/s) | t: %.2f ms",
+                \ g:llama_config.show_info == 2 ? l:prefix : 'llama.vim',
+                \ l:n_cached,  l:n_ctx, len(s:ring_chunks), g:llama_config.ring_n_chunks, s:ring_n_evict, len(s:ring_queued),
+                \ l:n_prompt,  l:t_prompt_ms,  l:s_prompt,
+                \ l:n_predict, l:t_predict_ms, l:s_predict,
+                \ 1000.0 * reltimefloat(reltime(s:t_fim_start))
+                \ )
+        endif
+
+        if g:llama_config.show_info == 1
+            " display the info in the statusline
+            let &statusline = l:info
+            let l:info = ''
+        endif
+    endif
+
+    " display the suggestion and append the info to the end of the first line
+    if s:ghost_text_nvim
+        call nvim_buf_set_extmark(l:bufnr, l:id_vt_fim, s:pos_y - 1, s:pos_x - 1, {
+            \ 'virt_text': [[s:content[0], 'llama_hl_hint'], [l:info, 'llama_hl_info']],
+            \ 'virt_text_win_col': virtcol('.') - 1
+            \ })
+
+        call nvim_buf_set_extmark(l:bufnr, l:id_vt_fim, s:pos_y - 1, 0, {
+            \ 'virt_lines': map(s:content[1:], {idx, val -> [[val, 'llama_hl_hint']]}),
+            \ 'virt_text_win_col': virtcol('.')
+            \ })
+    elseif s:ghost_text_vim
+        let l:new_suffix = s:content[0]
+        if !empty(l:new_suffix)
+            call prop_add(s:pos_y, s:pos_x + 1, {
+                        \ 'type': s:hlgroup_hint,
+                        \ 'text': l:new_suffix
+                        \ })
+        endif
+        for line in s:content[1:]
+            call prop_add(s:pos_y, 0, {
+                        \ 'type': s:hlgroup_hint,
+                        \ 'text': line,
+                        \ 'text_padding_left': s:get_indent(line),
+                        \ 'text_align': 'below'
+                        \ })
+        endfor
+        if !empty(l:info)
+            call prop_add(s:pos_y, 0, {
+                        \ 'type': s:hlgroup_info,
+                        \ 'text': l:info,
+                        \ 'text_padding_left': col('$'),
+                        \ 'text_wrap': 'truncate'
+                        \ })
+        endif
+    endif
+
+    " setup accept shortcuts
+    inoremap <buffer> <Tab>   <C-O>:call llama#fim_accept(v:false)<CR>
+    inoremap <buffer> <S-Tab> <C-O>:call llama#fim_accept(v:true)<CR>
+
+    let s:hint_shown = v:true
+endfunction
+
+function! s:fim_on_exit(job_id, exit_code, event = v:null)
+    if a:exit_code != 0
+        echom "Job failed with exit code: " . a:exit_code
+    endif
+
+    let s:current_job = v:null
+endfunction

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

@@ -42,7 +42,6 @@ int main(int argc, char ** argv) {
 
     llama_sampler * smpl = llama_sampler_chain_init(sparams);
 
-    llama_sampler_chain_add(smpl, llama_sampler_init_softmax());
     llama_sampler_chain_add(smpl, llama_sampler_init_dist(params.sparams.seed));
 
     // tokenize prompt
@@ -107,7 +106,6 @@ int main(int argc, char ** argv) {
 
     llama_sampler * smpl2 = llama_sampler_chain_init(sparams);
 
-    llama_sampler_chain_add(smpl2, llama_sampler_init_softmax());
     llama_sampler_chain_add(smpl2, llama_sampler_init_dist(params.sparams.seed));
 
     printf("\nsecond run: %s", params.prompt.c_str());
@@ -171,7 +169,6 @@ int main(int argc, char ** argv) {
 
     llama_sampler * smpl3 = llama_sampler_chain_init(sparams);
 
-    llama_sampler_chain_add(smpl3, llama_sampler_init_softmax());
     llama_sampler_chain_add(smpl3, llama_sampler_init_dist(params.sparams.seed));
 
     printf("\nsingle seq run: %s", params.prompt.c_str());

examples/server/README.md

@@ -319,6 +319,18 @@ node index.js
       - The prompt is a string or an array with the first element given as a string
       - The model's `tokenizer.ggml.add_bos_token` metadata is `true`
 
+    These input shapes and data type are allowed for `prompt`:
+
+      - Single string: `"string"`
+      - Single sequence of tokens: `[12, 34, 56]`
+      - Mixed tokens and strings: `[12, 34, "string", 56, 78]`
+
+    Multiple prompts are also supported. In this case, the completion result will be an array.
+
+      - Only strings: `["string1", "string2"]`
+      - Strings and sequences of tokens: `["string1", [12, 34, 56]]`
+      - Mixed types: `[[12, 34, "string", 56, 78], [12, 34, 56], "string"]`
+
     `temperature`: Adjust the randomness of the generated text. Default: `0.8`
 
     `dynatemp_range`: Dynamic temperature range. The final temperature will be in the range of `[temperature - dynatemp_range; temperature + dynatemp_range]` Default: `0.0`, which is disabled.

examples/server/server.cpp

@@ -43,21 +43,6 @@
 #include <unordered_map>
 #include <unordered_set>
 
-#define SLT_INF(slot, fmt, ...) LOG_INF("slot %12.*s: id %2d | task %d | " fmt, 12, __func__, (slot).id, (slot).id_task, __VA_ARGS__)
-#define SLT_WRN(slot, fmt, ...) LOG_WRN("slot %12.*s: id %2d | task %d | " fmt, 12, __func__, (slot).id, (slot).id_task, __VA_ARGS__)
-#define SLT_ERR(slot, fmt, ...) LOG_ERR("slot %12.*s: id %2d | task %d | " fmt, 12, __func__, (slot).id, (slot).id_task, __VA_ARGS__)
-#define SLT_DBG(slot, fmt, ...) LOG_DBG("slot %12.*s: id %2d | task %d | " fmt, 12, __func__, (slot).id, (slot).id_task, __VA_ARGS__)
-
-#define SRV_INF(fmt, ...) LOG_INF("srv  %12.*s: " fmt, 12, __func__, __VA_ARGS__)
-#define SRV_WRN(fmt, ...) LOG_WRN("srv  %12.*s: " fmt, 12, __func__, __VA_ARGS__)
-#define SRV_ERR(fmt, ...) LOG_ERR("srv  %12.*s: " fmt, 12, __func__, __VA_ARGS__)
-#define SRV_DBG(fmt, ...) LOG_DBG("srv  %12.*s: " fmt, 12, __func__, __VA_ARGS__)
-
-#define QUE_INF(fmt, ...) LOG_INF("que  %12.*s: " fmt, 12, __func__, __VA_ARGS__)
-#define QUE_WRN(fmt, ...) LOG_WRN("que  %12.*s: " fmt, 12, __func__, __VA_ARGS__)
-#define QUE_ERR(fmt, ...) LOG_ERR("que  %12.*s: " fmt, 12, __func__, __VA_ARGS__)
-#define QUE_DBG(fmt, ...) LOG_DBG("que  %12.*s: " fmt, 12, __func__, __VA_ARGS__)
-
 using json = nlohmann::ordered_json;
 
 enum stop_type {
@@ -68,6 +53,7 @@ enum stop_type {
 // state diagram: https://github.com/ggerganov/llama.cpp/pull/9283
 enum slot_state {
     SLOT_STATE_IDLE,
+    SLOT_STATE_STARTED, // TODO: this state is only used for setting up the initial prompt processing; maybe merge it with launch_slot_with_task in the future
     SLOT_STATE_PROCESSING_PROMPT,
     SLOT_STATE_DONE_PROMPT,
     SLOT_STATE_GENERATING,
@@ -79,7 +65,7 @@ enum server_state {
 };
 
 enum server_task_type {
-    SERVER_TASK_TYPE_COMPLETION,
+    SERVER_TASK_TYPE_INFERENCE,
     SERVER_TASK_TYPE_CANCEL,
     SERVER_TASK_TYPE_NEXT_RESPONSE,
     SERVER_TASK_TYPE_METRICS,
@@ -89,21 +75,22 @@ enum server_task_type {
     SERVER_TASK_TYPE_SET_LORA,
 };
 
-enum server_task_cmpl_type {
-    SERVER_TASK_CMPL_TYPE_NORMAL,
-    SERVER_TASK_CMPL_TYPE_EMBEDDING,
-    SERVER_TASK_CMPL_TYPE_RERANK,
-    SERVER_TASK_CMPL_TYPE_INFILL,
+enum server_task_inf_type {
+    SERVER_TASK_INF_TYPE_COMPLETION,
+    SERVER_TASK_INF_TYPE_EMBEDDING,
+    SERVER_TASK_INF_TYPE_RERANK,
+    SERVER_TASK_INF_TYPE_INFILL,
 };
 
 struct server_task {
     int id        = -1; // to be filled by server_queue
     int id_target = -1; // used by SERVER_TASK_TYPE_CANCEL
 
+    llama_tokens prompt_tokens;
     server_task_type type;
     json data;
 
-    server_task_cmpl_type cmpl_type = SERVER_TASK_CMPL_TYPE_NORMAL;
+    server_task_inf_type inf_type = SERVER_TASK_INF_TYPE_COMPLETION;
 
     // utility function
     static std::unordered_set<int> get_list_id(const std::vector<server_task> & tasks) {
@@ -161,26 +148,20 @@ struct server_slot {
     int32_t i_batch     = -1;
     int32_t n_predict   = -1; // TODO: disambiguate from params.n_predict
 
+    // n_prompt_tokens may not be equal to prompt_tokens.size(), because prompt maybe truncated
     int32_t n_prompt_tokens           = 0;
     int32_t n_prompt_tokens_processed = 0;
 
-    json prompt; // can be either a string, array of strings or array of token ids
-
-    json input_prefix;
-    json input_suffix;
-    json input_extra;
-
-    // when a task is submitted, we first tokenize the prompt and store it here
-    std::vector<llama_token> prompt_tokens;
-    std::vector<llama_token> extra_tokens;
+    // input prompt tokens
+    llama_tokens prompt_tokens;
 
     size_t last_nl_pos = 0;
 
     std::string generated_text;
-    std::vector<llama_token> cache_tokens;
+    llama_tokens cache_tokens;
     std::vector<completion_token_output> generated_token_probs;
 
-    server_task_cmpl_type cmpl_type = SERVER_TASK_CMPL_TYPE_NORMAL;
+    server_task_inf_type inf_type = SERVER_TASK_INF_TYPE_COMPLETION;
 
     bool has_next_token = true;
     bool has_new_line   = false;
@@ -229,7 +210,7 @@ struct server_slot {
         n_past             = 0;
         n_sent_text        = 0;
         n_sent_token_probs = 0;
-        cmpl_type          = SERVER_TASK_CMPL_TYPE_NORMAL;
+        inf_type           = SERVER_TASK_INF_TYPE_COMPLETION;
 
         generated_token_probs.clear();
     }
@@ -734,42 +715,6 @@ struct server_context {
         metrics.init();
     }
 
-    std::vector<llama_token> tokenize(const json & json_prompt, bool add_special, bool parse_special) const {
-        // If `add_bos` is true, we only add BOS, when json_prompt is a string,
-        // or the first element of the json_prompt array is a string.
-        std::vector<llama_token> prompt_tokens;
-
-        if (json_prompt.is_array()) {
-            bool first = true;
-            for (const auto & p : json_prompt) {
-                if (p.is_string()) {
-                    auto s = p.template get<std::string>();
-
-                    std::vector<llama_token> p;
-                    if (first) {
-                        p = common_tokenize(ctx, s, add_special, parse_special);
-                        first = false;
-                    } else {
-                        p = common_tokenize(ctx, s, false, parse_special);
-                    }
-
-                    prompt_tokens.insert(prompt_tokens.end(), p.begin(), p.end());
-                } else {
-                    if (first) {
-                        first = false;
-                    }
-
-                    prompt_tokens.push_back(p.template get<llama_token>());
-                }
-            }
-        } else {
-            auto s = json_prompt.template get<std::string>();
-            prompt_tokens = common_tokenize(ctx, s, add_special, parse_special);
-        }
-
-        return prompt_tokens;
-    }
-
     server_slot * get_slot_by_id(int id) {
         for (server_slot & slot : slots) {
             if (slot.id == id) {
@@ -794,22 +739,16 @@ struct server_context {
                     continue;
                 }
 
-                // skip the slot if it does not contains prompt
-                if (!slot.prompt.is_string()) {
+                // skip the slot if it does not contains cached tokens
+                if (slot.prompt_tokens.empty()) {
                     continue;
                 }
 
-                // current slot's prompt
-                std::string slot_prompt = slot.prompt.get<std::string>();
-
-                // length of the current slot's prompt
-                int slot_prompt_len = slot_prompt.size();
-
                 // length of the Longest Common Prefix between the current slot's prompt and the input prompt
-                int lcp_len = longest_common_prefix(slot_prompt, prompt);
+                int lcp_len = longest_common_prefix(slot.cache_tokens, slot.prompt_tokens);
 
                 // fraction of the common substring length compared to the current slot's prompt length
-                similarity = static_cast<float>(lcp_len) / slot_prompt_len;
+                similarity = static_cast<float>(lcp_len) / static_cast<int>(slot.prompt_tokens.size());
 
                 // select the current slot if the criteria match
                 if (lcp_len > max_lcp_len && similarity > slot_prompt_similarity) {
@@ -914,57 +853,6 @@ struct server_context {
             SLT_WRN(slot, "n_predict = %d exceeds server configuration, setting to %d", slot.n_predict, slot.n_predict);
         }
 
-        // infill
-        slot.input_prefix = json_value(data, "input_prefix", json());
-        slot.input_suffix = json_value(data, "input_suffix", json());
-        slot.input_extra  = json_value(data, "input_extra",  json());
-
-        SLT_DBG(slot, "extra_context chunks: %d\n", (int) slot.input_extra.size());
-        for (const auto & chunk : slot.input_extra) {
-            // { "text": string, "filename": string }
-            if (!chunk.contains("text") || !chunk["text"].is_string()) {
-                send_error(task, "extra_context chunk must contain a \"text\" field with a string value", ERROR_TYPE_INVALID_REQUEST);
-                return false;
-            }
-
-            // filename is optional
-            if (chunk.contains("filename") && !chunk["filename"].is_string()) {
-                send_error(task, "extra_context chunk's \"filename\" field must be a string", ERROR_TYPE_INVALID_REQUEST);
-                return false;
-            }
-
-            SLT_DBG(slot, "extra_context chunk in file '%s':\n%s\n", chunk.value("filename", "").c_str(), chunk.value("text", "").c_str());
-        }
-
-        // get prompt
-        {
-            const auto & prompt = data.find("prompt");
-            if (prompt == data.end()) {
-                send_error(task, "\"prompt\" must be provided", ERROR_TYPE_INVALID_REQUEST);
-                return false;
-            }
-
-            if ((prompt->is_string()) ||
-                (prompt->is_array() &&  prompt->size() == 1 && prompt->at(0).is_string()) ||
-                (prompt->is_array() && !prompt->empty()     && prompt->at(0).is_number_integer())) {
-                slot.prompt = *prompt;
-            } else if (prompt->is_array() && prompt->size() == 1 && prompt->at(0).is_array()) {
-                slot.prompt = prompt->at(0);
-            } else if (prompt->is_array() && prompt->size() > 1) {
-                // array of strings
-                for (const auto & el : *prompt) {
-                    if (!el.is_string()) {
-                        send_error(task, "\"prompt\" must be a string, an array of strings or an array of integers", ERROR_TYPE_INVALID_REQUEST);
-                        return false;
-                    }
-                }
-                slot.prompt = *prompt;
-            } else {
-                send_error(task, "\"prompt\" must be a string, an array of strings or an array of integers", ERROR_TYPE_INVALID_REQUEST);
-                return false;
-            }
-        }
-
         {
             slot.sparams.logit_bias.clear();
 
@@ -1044,8 +932,7 @@ struct server_context {
             }
         }
 
-        slot.state = SLOT_STATE_PROCESSING_PROMPT;
-        slot.prompt_tokens.clear();
+        slot.state = SLOT_STATE_STARTED;
 
         SLT_INF(slot, "%s", "processing task\n");
 
@@ -1297,7 +1184,7 @@ struct server_context {
         };
 
         if (slot.sparams.n_probs > 0) {
-            const std::vector<llama_token> to_send_toks = common_tokenize(ctx, tkn.text_to_send, false);
+            const llama_tokens to_send_toks = common_tokenize(ctx, tkn.text_to_send, false);
             const size_t probs_pos      = std::min(slot.n_sent_token_probs,                       slot.generated_token_probs.size());
             const size_t probs_stop_pos = std::min(slot.n_sent_token_probs + to_send_toks.size(), slot.generated_token_probs.size());
 
@@ -1333,7 +1220,7 @@ struct server_context {
             {"tokens_predicted",    slot.n_decoded},
             {"tokens_evaluated",    slot.n_prompt_tokens},
             {"generation_settings", get_formated_generation(slot)},
-            {"prompt",              slot.prompt},
+            {"prompt",              common_detokenize(ctx, slot.prompt_tokens)},
             {"has_new_line",        slot.has_new_line},
             {"truncated",           slot.truncated},
             {"stopped_eos",         slot.stopped_eos},
@@ -1348,7 +1235,7 @@ struct server_context {
         if (slot.sparams.n_probs > 0) {
             std::vector<completion_token_output> probs;
             if (!slot.params.stream && slot.stopped_word) {
-                const std::vector<llama_token> stop_word_toks = common_tokenize(ctx, slot.stopping_word, false);
+                const llama_tokens stop_word_toks = common_tokenize(ctx, slot.stopping_word, false);
 
                 size_t safe_offset = std::min(slot.generated_token_probs.size(), stop_word_toks.size());
                 probs = std::vector<completion_token_output>(
@@ -1457,19 +1344,17 @@ struct server_context {
     // Functions to create new task(s) and receive result(s)
     //
 
-    std::vector<server_task> create_tasks_cmpl(json data, server_task_cmpl_type cmpl_type) {
+    // break the input "prompt" into multiple tasks if needed, then format and tokenize the input prompt(s)
+    std::vector<server_task> create_tasks_inference(json data, server_task_inf_type inf_type) {
         std::vector<server_task> tasks;
-        auto create_task = [&](json & task_data, bool replace_prompt, json prompt) {
+        auto create_task = [&](json & task_data, llama_tokens & prompt_tokens) {
+            SRV_DBG("create task, n_tokens = %d\n", (int) prompt_tokens.size());
             server_task task;
-            task.id        = queue_tasks.get_new_id();
-            task.cmpl_type = cmpl_type;
-            task.type      = SERVER_TASK_TYPE_COMPLETION;
-            if (replace_prompt) {
-                task.data  = task_data;
-                task.data["prompt"] = std::move(prompt);
-            } else {
-                task.data  = std::move(task_data);
-            }
+            task.id            = queue_tasks.get_new_id();
+            task.inf_type      = inf_type;
+            task.type          = SERVER_TASK_TYPE_INFERENCE;
+            task.data          = task_data;
+            task.prompt_tokens = std::move(prompt_tokens);
             tasks.push_back(std::move(task));
         };
 
@@ -1478,41 +1363,49 @@ struct server_context {
             throw std::runtime_error(error_msg);
         }
 
-        json prompt = data.at("prompt");
-
-        // if the prompt is a singleton (i.e. a string or a list of tokens), we only need to create single task
-        if (prompt.is_string() || json_is_array_of_numbers(prompt)) {
-            data["index"] = 0;
-            create_task(data, false, nullptr);
-        } else if (prompt.is_array()) {
-            // otherwise, it's a multiple-prompt task, we break it into smaller tasks
-            std::vector<json> prompts = prompt;
-            if (cmpl_type == SERVER_TASK_CMPL_TYPE_RERANK) {
-                // prompts[0] is the question
-                // the rest are the answers/documents
-                SRV_DBG("creating rerank tasks, n_prompts = %d\n", (int) prompts.size() - 1);
-                for (size_t i = 1; i < prompts.size(); i++) {
-                    json qd;
-                    qd.push_back(prompts[0]);
-                    qd.push_back(prompts[i]);
-                    data["index"] = i - 1;
-                    create_task(data, true, qd);
-                }
-            } else {
-                SRV_DBG("creating multi-prompt tasks, n_prompts = %d\n", (int) prompts.size());
-                for (size_t i = 0; i < prompts.size(); i++) {
-                    const auto & e = prompts[i];
-                    if (e.is_string() || json_is_array_of_numbers(e)) {
+        // because llama_tokenize api is thread-safe, we can tokenize the prompt from HTTP thread
+        bool add_special = inf_type != SERVER_TASK_INF_TYPE_RERANK && inf_type != SERVER_TASK_INF_TYPE_INFILL;
+        std::vector<llama_tokens> tokenized_prompts = tokenize_input_prompts(ctx, data.at("prompt"), add_special, true);
+        switch (inf_type) {
+            case SERVER_TASK_INF_TYPE_RERANK:
+                {
+                    // prompts[0] is the question
+                    // the rest are the answers/documents
+                    GGML_ASSERT(tokenized_prompts.size() > 1);
+                    SRV_DBG("creating rerank tasks, n_prompts = %d\n", (int) tokenized_prompts.size() - 1);
+                    for (size_t i = 1; i < tokenized_prompts.size(); i++) {
+                        data["index"] = i - 1;
+                        auto tokens = format_rerank(model, tokenized_prompts[0], tokenized_prompts[i]);
+                        create_task(data, tokens);
+                    }
+                } break;
+            case SERVER_TASK_INF_TYPE_INFILL:
+                {
+                    SRV_DBG("creating infill tasks, n_prompts = %d\n", (int) tokenized_prompts.size());
+                    for (size_t i = 0; i < tokenized_prompts.size(); i++) {
                         data["index"] = i;
-                        create_task(data, true, e);
-                    } else {
-                        throw std::runtime_error(error_msg);
+                        auto tokens = format_infill(
+                            ctx,
+                            data.at("input_prefix"),
+                            data.at("input_suffix"),
+                            data.at("input_extra"),
+                            params.n_batch,
+                            params.n_predict,
+                            slots[0].n_ctx, // TODO: there should be a better way
+                            params.spm_infill,
+                            tokenized_prompts[i]
+                        );
+                        create_task(data, tokens);
+                    }
+                } break;
+            default:
+                {
+                    SRV_DBG("creating multi-prompt tasks, n_prompts = %d\n", (int) tokenized_prompts.size());
+                    for (size_t i = 0; i < tokenized_prompts.size(); i++) {
+                        data["index"] = i;
+                        create_task(data, tokenized_prompts[i]);
                     }
                 }
-            }
-        } else {
-            // invalid case
-            throw std::runtime_error(error_msg);
         }
 
         return tasks;
@@ -1534,7 +1427,7 @@ struct server_context {
         queue_tasks.post(cancel_tasks, true);
     }
 
-    // receive the results from task(s) created by create_tasks_cmpl
+    // receive the results from task(s) created by create_tasks_inference
     void receive_cmpl_results(
             const std::unordered_set<int> & id_tasks,
             const std::function<void(std::vector<server_task_result>&)> & result_handler,
@@ -1558,7 +1451,7 @@ struct server_context {
         result_handler(results);
     }
 
-    // receive the results from task(s) created by create_tasks_cmpl, in stream mode
+    // receive the results from task(s) created by create_tasks_inference, in stream mode
     void receive_cmpl_results_stream(
             const std::unordered_set<int> & id_tasks, const
             std::function<bool(server_task_result&)> & result_handler, const
@@ -1591,7 +1484,7 @@ struct server_context {
 
     void process_single_task(const server_task & task) {
         switch (task.type) {
-            case SERVER_TASK_TYPE_COMPLETION:
+            case SERVER_TASK_TYPE_INFERENCE:
                 {
                     const int id_slot = json_value(task.data, "id_slot", -1);
 
@@ -1623,9 +1516,10 @@ struct server_context {
 
                     slot->reset();
 
-                    slot->id_task   = task.id;
-                    slot->cmpl_type = task.cmpl_type;
-                    slot->index     = json_value(task.data, "index", 0);
+                    slot->id_task       = task.id;
+                    slot->inf_type      = task.inf_type;
+                    slot->index         = json_value(task.data, "index", 0);
+                    slot->prompt_tokens = std::move(task.prompt_tokens);
 
                     if (!launch_slot_with_task(*slot, task)) {
                         SRV_ERR("failed to launch slot with task, id_task = %d\n", task.id);
@@ -1658,7 +1552,7 @@ struct server_context {
                         slot_data["id"]         = slot.id;
                         slot_data["id_task"]    = slot.id_task;
                         slot_data["state"]      = slot.state;
-                        slot_data["prompt"]     = slot.prompt;
+                        slot_data["prompt"]     = common_detokenize(ctx, slot.prompt_tokens);
                         slot_data["next_token"] = {
                             {"has_next_token", slot.has_next_token},
                             {"has_new_line",   slot.has_new_line},
@@ -1785,9 +1679,6 @@ struct server_context {
                     }
                     slot->cache_tokens.resize(token_count);
 
-                    // TODO: maybe detokenize the slot->cache_tokens instead?
-                    slot->prompt = string_format("[restored %d tokens from file]", (int) token_count);
-
                     const int64_t t_end = ggml_time_us();
                     const double t_restore_ms = (t_end - t_start) / 1000.0;
 
@@ -1954,142 +1845,18 @@ struct server_context {
         if (params.cont_batching || batch.n_tokens == 0) {
             for (auto & slot : slots) {
                 // this slot still has a prompt to be processed
-                if (slot.state == SLOT_STATE_PROCESSING_PROMPT) {
+                if (slot.state == SLOT_STATE_PROCESSING_PROMPT || slot.state == SLOT_STATE_STARTED) {
                     auto & prompt_tokens = slot.prompt_tokens;
 
-                    // we haven't tokenized the prompt yet - do it now:
-                    if (prompt_tokens.empty()) {
-                        SLT_INF(slot, "tokenizing prompt, len = %d\n", (int) slot.prompt.size());
-
+                    // TODO: maybe move branch to outside of this loop in the future
+                    if (slot.state == SLOT_STATE_STARTED) {
                         slot.t_start_process_prompt = ggml_time_us();
                         slot.t_start_generation = 0;
-
-                        switch (slot.cmpl_type) {
-                            case SERVER_TASK_CMPL_TYPE_NORMAL:
-                            case SERVER_TASK_CMPL_TYPE_EMBEDDING:
-                                {
-                                    prompt_tokens = tokenize(slot.prompt, llama_add_bos_token(model), true);
-                                } break;
-                            case SERVER_TASK_CMPL_TYPE_RERANK:
-                                {
-                                    // require slot.prompt to be array of 2 strings
-                                    if (!slot.prompt.is_array() || slot.prompt.size() != 2) {
-                                        SLT_ERR(slot, "%s", "invalid prompt for rerank task\n");
-                                        slot.release();
-                                        send_error(slot, "invalid prompt for rerank task", ERROR_TYPE_INVALID_REQUEST);
-                                        continue;
-                                    }
-
-                                    // prompt: [BOS]query[EOS][SEP]doc[EOS]
-                                    prompt_tokens.clear();
-                                    prompt_tokens.push_back(llama_token_bos(model));
-                                    {
-                                        const auto part = tokenize(slot.prompt[0], false, false);
-                                        prompt_tokens.insert(prompt_tokens.end(), part.begin(), part.end());
-                                    }
-                                    prompt_tokens.push_back(llama_token_eos(model));
-                                    prompt_tokens.push_back(llama_token_sep(model));
-                                    {
-                                        const auto part = tokenize(slot.prompt[1], false, false);
-                                        prompt_tokens.insert(prompt_tokens.end(), part.begin(), part.end());
-                                    }
-                                    prompt_tokens.push_back(llama_token_eos(model));
-                                } break;
-                            case SERVER_TASK_CMPL_TYPE_INFILL:
-                                {
-                                    // TODO: optimize this block by reducing memory allocations and movement
-
-                                    // use FIM repo-level pattern:
-                                    // ref: https://arxiv.org/pdf/2409.12186
-                                    //
-                                    // [FIM_REP]myproject
-                                    // [FIM_SEP]filename0
-                                    // extra chunk 0
-                                    // [FIM_SEP]filename1
-                                    // extra chunk 1
-                                    // ...
-                                    // [FIM_SEP]filename
-                                    // [FIM_PRE]prefix[FIM_SUF]suffix[FIM_MID]prompt
-                                    //
-                                    auto tokens_prefix = tokenize(slot.input_prefix, false, false);
-                                    auto tokens_suffix = tokenize(slot.input_suffix, false, false);
-                                    auto tokens_prompt = tokenize(slot.prompt,       false, false);
-
-                                    slot.extra_tokens.clear();
-                                    if (llama_token_fim_rep(model) != LLAMA_TOKEN_NULL) {
-                                        static const auto k_fim_repo = tokenize("myproject\n", false, false);
-
-                                        slot.extra_tokens.push_back(llama_token_fim_rep(model));
-                                        slot.extra_tokens.insert(slot.extra_tokens.end(), k_fim_repo.begin(), k_fim_repo.end());
-                                    }
-
-                                    for (const auto & chunk : slot.input_extra) {
-                                        // { "text": string, "filename": string }
-                                        const std::string text     = chunk.value("text", "");
-                                        const std::string filename = chunk.value("filename", "tmp");
-
-                                        if (llama_token_fim_sep(model) != LLAMA_TOKEN_NULL) {
-                                            const auto k_fim_file = tokenize(filename + "\n", false, false);
-
-                                            slot.extra_tokens.insert(slot.extra_tokens.end(), llama_token_fim_sep(model));
-                                            slot.extra_tokens.insert(slot.extra_tokens.end(), k_fim_file.begin(), k_fim_file.end());
-                                        } else {
-                                            // chunk separator in binary form to avoid confusing the AI
-                                            static const char k_chunk_prefix_str[] = {0x0a, 0x0a, 0x2d, 0x2d, 0x2d, 0x20, 0x73, 0x6e, 0x69, 0x70, 0x70, 0x65, 0x74, 0x20, 0x2d, 0x2d, 0x2d, 0x0a, 0x0a, 0x00};
-                                            static const auto k_chunk_prefix_tokens = tokenize(k_chunk_prefix_str, false, false);
-
-                                            slot.extra_tokens.insert(slot.extra_tokens.end(), k_chunk_prefix_tokens.begin(), k_chunk_prefix_tokens.end());
-                                        }
-
-                                        const auto chunk_tokens = tokenize(text, false, false);
-                                        slot.extra_tokens.insert(slot.extra_tokens.end(), chunk_tokens.begin(), chunk_tokens.end());
-                                    }
-
-                                    if (llama_token_fim_sep(model) != LLAMA_TOKEN_NULL) {
-                                        // TODO: current filename
-                                        static const auto k_fim_file = tokenize("filename\n", false, false);
-
-                                        slot.extra_tokens.insert(slot.extra_tokens.end(), llama_token_fim_sep(model));
-                                        slot.extra_tokens.insert(slot.extra_tokens.end(), k_fim_file.begin(), k_fim_file.end());
-                                    }
-
-                                    // for now pick FIM context to fit in a batch (ratio prefix:suffix = 3:1, TODO: configurable?)
-                                    const int n_suffix_take = std::min<int>(tokens_suffix.size(),   (n_batch/4));
-                                    const int n_prefix_take = std::min<int>(tokens_prefix.size(), 3*(n_batch/4) - 3);
-
-                                    // fill the rest of the context with extra chunks
-                                    const int n_extra_take = std::min<int>(std::max<int>(0, slot.n_ctx - (n_batch) - 2*slot.n_predict), slot.extra_tokens.size());
-
-                                    tokens_prefix.erase(tokens_prefix.begin(), tokens_prefix.begin() + tokens_prefix.size() - n_prefix_take);
-                                    tokens_suffix.resize(n_suffix_take);
-
-                                    tokens_prefix.insert(tokens_prefix.begin(), llama_token_fim_pre(model));
-                                    tokens_prefix.insert(tokens_prefix.end(),   tokens_prompt.begin(), tokens_prompt.end());
-                                    tokens_suffix.insert(tokens_suffix.begin(), llama_token_fim_suf(model));
-
-                                    auto embd_inp = params.spm_infill ? tokens_suffix : tokens_prefix;
-                                    auto embd_end = params.spm_infill ? tokens_prefix : tokens_suffix;
-
-                                    if (llama_add_bos_token(model)) {
-                                        embd_inp.insert(embd_inp.begin(), llama_token_bos(model));
-                                    }
-
-                                    SLT_DBG(slot, "extra: n_ctx = %d, n_extra_take = %d, n_extra = %d\n", slot.n_ctx, n_extra_take, (int) slot.extra_tokens.size());
-
-                                    // put the extra context before the FIM prefix
-                                    embd_inp.insert(embd_inp.begin(), slot.extra_tokens.end() - n_extra_take, slot.extra_tokens.end());
-
-                                    embd_inp.insert(embd_inp.end(), embd_end.begin(), embd_end.end());
-                                    embd_inp.push_back(llama_token_fim_mid(model));
-
-                                    prompt_tokens = std::move(embd_inp);
-                                } break;
-                        }
-
                         slot.n_past = 0;
                         slot.n_prompt_tokens = prompt_tokens.size();
+                        slot.state = SLOT_STATE_PROCESSING_PROMPT;
 
-                        SLT_INF(slot, "prompt tokenized, n_ctx_slot = %d, n_keep = %d, n_prompt_tokens = %d\n", slot.n_ctx, slot.params.n_keep, slot.n_prompt_tokens);
+                        SLT_INF(slot, "new prompt, n_ctx_slot = %d, n_keep = %d, n_prompt_tokens = %d\n", slot.n_ctx, slot.params.n_keep, slot.n_prompt_tokens);
 
                         // print prompt tokens (for debugging)
                         if (1) {
@@ -2114,13 +1881,18 @@ struct server_context {
                             continue;
                         }
 
-                        if (slot.cmpl_type == SERVER_TASK_CMPL_TYPE_EMBEDDING || slot.cmpl_type == SERVER_TASK_CMPL_TYPE_RERANK) {
-                            // this prompt is too large to process - discard it
+                        if (slot.inf_type == SERVER_TASK_INF_TYPE_EMBEDDING || slot.inf_type == SERVER_TASK_INF_TYPE_RERANK) {
                             if (slot.n_prompt_tokens > n_ubatch) {
                                 slot.release();
                                 send_error(slot, "input is too large to process. increase the physical batch size", ERROR_TYPE_SERVER);
                                 continue;
                             }
+
+                            if (slot.n_prompt_tokens > slot.n_ctx) {
+                                slot.release();
+                                send_error(slot, "input is larger than the max context size. skipping", ERROR_TYPE_SERVER);
+                                continue;
+                            }
                         } else {
                             if (!params.ctx_shift) {
                                 // if context shift is disabled, we make sure prompt size is smaller than KV size
@@ -2144,7 +1916,7 @@ struct server_context {
                                 const int n_block_size = n_left / 2;
                                 const int erased_blocks = (slot.n_prompt_tokens - slot.params.n_keep - n_block_size) / n_block_size;
 
-                                std::vector<llama_token> new_tokens(
+                                llama_tokens new_tokens(
                                         prompt_tokens.begin(),
                                         prompt_tokens.begin() + slot.params.n_keep);
 
@@ -2163,17 +1935,10 @@ struct server_context {
                                 GGML_ASSERT(slot.n_prompt_tokens < slot.n_ctx);
                             }
 
-                            common_sampler_reset(slot.smpl);
-
                             if (slot.params.cache_prompt) {
                                 // reuse any previously computed tokens that are common with the new prompt
                                 slot.n_past = longest_common_prefix(slot.cache_tokens, prompt_tokens);
 
-                                // push the prompt into the sampling context (do not apply grammar)
-                                for (int i = 0; i < slot.n_past; ++i) {
-                                    common_sampler_accept(slot.smpl, slot.cache_tokens[i], false);
-                                }
-
                                 // reuse chunks from the cached prompt by shifting their KV cache in the new position
                                 if (params.n_cache_reuse > 0) {
                                     size_t head_c = slot.n_past; // cache
@@ -2206,8 +1971,6 @@ struct server_context {
                                             for (size_t i = 0; i < n_match; i++) {
                                                 slot.cache_tokens[head_p + i] = slot.cache_tokens[head_c + i];
 
-                                                common_sampler_accept(slot.smpl, slot.cache_tokens[head_p + i], false);
-
                                                 slot.n_past++;
                                             }
 
@@ -2234,7 +1997,7 @@ struct server_context {
                     }
 
                     // non-causal tasks require to fit the entire prompt in the physical batch
-                    if (slot.cmpl_type == SERVER_TASK_CMPL_TYPE_EMBEDDING || slot.cmpl_type == SERVER_TASK_CMPL_TYPE_RERANK) {
+                    if (slot.inf_type == SERVER_TASK_INF_TYPE_EMBEDDING || slot.inf_type == SERVER_TASK_INF_TYPE_RERANK) {
                         // cannot fit the prompt in the current batch - will try next iter
                         if (batch.n_tokens + slot.n_prompt_tokens > n_batch) {
                             continue;
@@ -2243,8 +2006,8 @@ struct server_context {
 
                     // check that we are in the right batch_type, if not defer the slot
                     const bool slot_type =
-                        slot.cmpl_type == SERVER_TASK_CMPL_TYPE_EMBEDDING ||
-                        slot.cmpl_type == SERVER_TASK_CMPL_TYPE_RERANK     ? 1 : 0;
+                        slot.inf_type == SERVER_TASK_INF_TYPE_EMBEDDING ||
+                        slot.inf_type == SERVER_TASK_INF_TYPE_RERANK     ? 1 : 0;
 
                     if (batch_type == -1) {
                         batch_type = slot_type;
@@ -2259,8 +2022,6 @@ struct server_context {
 
                         // there is no common part left
                         slot.n_past = 0;
-
-                        common_sampler_reset(slot.smpl);
                     }
 
                     SLT_INF(slot, "kv cache rm [%d, end)\n", slot.n_past);
@@ -2288,6 +2049,13 @@ struct server_context {
 
                         GGML_ASSERT(batch.n_tokens > 0);
 
+                        common_sampler_reset(slot.smpl);
+
+                        // Process all prompt tokens through sampler system
+                        for (int i = 0; i < slot.n_prompt_tokens; ++i) {
+                            common_sampler_accept(slot.smpl, prompt_tokens[i], false);
+                        }
+
                         // extract the logits only for the last token
                         batch.logits[batch.n_tokens - 1] = true;
 
@@ -2357,7 +2125,7 @@ struct server_context {
                 }
 
                 if (slot.state == SLOT_STATE_DONE_PROMPT) {
-                    if (slot.cmpl_type == SERVER_TASK_CMPL_TYPE_EMBEDDING) {
+                    if (slot.inf_type == SERVER_TASK_INF_TYPE_EMBEDDING) {
                         // prompt evaluated for embedding
                         send_embedding(slot, batch_view);
                         slot.release();
@@ -2365,7 +2133,7 @@ struct server_context {
                         continue; // continue loop of slots
                     }
 
-                    if (slot.cmpl_type == SERVER_TASK_CMPL_TYPE_RERANK) {
+                    if (slot.inf_type == SERVER_TASK_INF_TYPE_RERANK) {
                         send_rerank(slot, batch_view);
                         slot.release();
                         slot.i_batch = -1;
@@ -2919,13 +2687,13 @@ int main(int argc, char ** argv) {
         res_ok(res, {{ "success", true }});
     };
 
-    const auto handle_completions_generic = [&ctx_server, &res_error, &res_ok](server_task_cmpl_type cmpl_type, json & data, httplib::Response & res) {
+    const auto handle_completions_generic = [&ctx_server, &res_error, &res_ok](server_task_inf_type inf_type, json & data, httplib::Response & res) {
         if (ctx_server.params.embedding || ctx_server.params.reranking) {
             res_error(res, format_error_response("This server does not support completions. Start it without `--embeddings` or `--reranking`", ERROR_TYPE_NOT_SUPPORTED));
             return;
         }
 
-        std::vector<server_task> tasks = ctx_server.create_tasks_cmpl(data, cmpl_type);
+        std::vector<server_task> tasks = ctx_server.create_tasks_inference(data, inf_type);
         ctx_server.queue_results.add_waiting_tasks(tasks);
         ctx_server.queue_tasks.post(tasks);
 
@@ -2971,10 +2739,11 @@ int main(int argc, char ** argv) {
 
     const auto handle_completions = [&handle_completions_generic](const httplib::Request & req, httplib::Response & res) {
         json data = json::parse(req.body);
-        return handle_completions_generic(SERVER_TASK_CMPL_TYPE_NORMAL, data, res);
+        return handle_completions_generic(SERVER_TASK_INF_TYPE_COMPLETION, data, res);
     };
 
     const auto handle_infill = [&ctx_server, &res_error, &handle_completions_generic](const httplib::Request & req, httplib::Response & res) {
+        // check model compatibility
         std::string err;
         if (llama_token_fim_pre(ctx_server.model) == LLAMA_TOKEN_NULL) {
             err += "prefix token is missing. ";
@@ -2985,14 +2754,42 @@ int main(int argc, char ** argv) {
         if (llama_token_fim_mid(ctx_server.model) == LLAMA_TOKEN_NULL) {
             err += "middle token is missing. ";
         }
-
         if (!err.empty()) {
             res_error(res, format_error_response(string_format("Infill is not supported by this model: %s", err.c_str()), ERROR_TYPE_NOT_SUPPORTED));
             return;
         }
 
         json data = json::parse(req.body);
-        return handle_completions_generic(SERVER_TASK_CMPL_TYPE_INFILL, data, res);
+
+        // validate input
+        if (!data.contains("input_prefix")) {
+            res_error(res, format_error_response("\"input_prefix\" is required", ERROR_TYPE_INVALID_REQUEST));
+        }
+
+        if (!data.contains("input_suffix")) {
+            res_error(res, format_error_response("\"input_suffix\" is required", ERROR_TYPE_INVALID_REQUEST));
+        }
+
+        if (data.contains("input_extra") && !data.at("input_extra").is_array()) {
+            res_error(res, format_error_response("\"input_extra\" must be an array of {\"filename\": string, \"text\": string}", ERROR_TYPE_INVALID_REQUEST));
+            return;
+        }
+        json input_extra = json_value(data, "input_extra", json::array());
+        for (const auto & chunk : input_extra) {
+            // { "text": string, "filename": string }
+            if (!chunk.contains("text") || !chunk.at("text").is_string()) {
+                res_error(res, format_error_response("extra_context chunk must contain a \"text\" field with a string value", ERROR_TYPE_INVALID_REQUEST));
+                return;
+            }
+            // filename is optional
+            if (chunk.contains("filename") && !chunk.at("filename").is_string()) {
+                res_error(res, format_error_response("extra_context chunk's \"filename\" field must be a string", ERROR_TYPE_INVALID_REQUEST));
+                return;
+            }
+        }
+        data["input_extra"] = input_extra; // default to empty array if it's not exist
+
+        return handle_completions_generic(SERVER_TASK_INF_TYPE_INFILL, data, res);
     };
 
     // TODO: maybe merge this function with "handle_completions_generic"
@@ -3004,7 +2801,7 @@ int main(int argc, char ** argv) {
 
         json data = oaicompat_completion_params_parse(ctx_server.model, json::parse(req.body), params.chat_template);
 
-        std::vector<server_task> tasks = ctx_server.create_tasks_cmpl(data, SERVER_TASK_CMPL_TYPE_NORMAL);
+        std::vector<server_task> tasks = ctx_server.create_tasks_inference(data, SERVER_TASK_INF_TYPE_COMPLETION);
         ctx_server.queue_results.add_waiting_tasks(tasks);
         ctx_server.queue_tasks.post(tasks);
 
@@ -3077,7 +2874,7 @@ int main(int argc, char ** argv) {
             const bool add_special = json_value(body, "add_special", false);
             const bool with_pieces = json_value(body, "with_pieces", false);
 
-            std::vector<llama_token> tokens = ctx_server.tokenize(body.at("content"), add_special, true);
+            llama_tokens tokens = tokenize_mixed(ctx_server.ctx, body.at("content"), add_special, true);
 
             if (with_pieces) {
                 for (const auto& token : tokens) {
@@ -3114,7 +2911,7 @@ int main(int argc, char ** argv) {
 
         std::string content;
         if (body.count("tokens") != 0) {
-            const std::vector<llama_token> tokens = body.at("tokens");
+            const llama_tokens tokens = body.at("tokens");
             content = tokens_to_str(ctx_server.ctx, tokens.cbegin(), tokens.cend());
         }
 
@@ -3148,7 +2945,7 @@ int main(int argc, char ** argv) {
         json responses = json::array();
         bool error = false;
         {
-            std::vector<server_task> tasks = ctx_server.create_tasks_cmpl({{"prompt", prompt}}, SERVER_TASK_CMPL_TYPE_EMBEDDING);
+            std::vector<server_task> tasks = ctx_server.create_tasks_inference({{"prompt", prompt}}, SERVER_TASK_INF_TYPE_EMBEDDING);
             ctx_server.queue_results.add_waiting_tasks(tasks);
             ctx_server.queue_tasks.post(tasks);
 
@@ -3225,7 +3022,7 @@ int main(int argc, char ** argv) {
         json responses = json::array();
         bool error = false;
         {
-            std::vector<server_task> tasks = ctx_server.create_tasks_cmpl({{"prompt", prompt}}, SERVER_TASK_CMPL_TYPE_RERANK);
+            std::vector<server_task> tasks = ctx_server.create_tasks_inference({{"prompt", prompt}}, SERVER_TASK_INF_TYPE_RERANK);
             ctx_server.queue_results.add_waiting_tasks(tasks);
             ctx_server.queue_tasks.post(tasks);
 

examples/server/tests/features/infill.feature

@@ -0,0 +1,36 @@
+@llama.cpp
+@infill
+Feature: llama.cpp server
+
+  # The current model is made by adding FIM tokens to the existing stories260K
+  # We may want to use a better model in the future, maybe something like SmolLM 360M
+
+  Background: Server startup
+    Given a server listening on localhost:8080
+    And   a model file tinyllamas/stories260K-infill.gguf from HF repo ggml-org/models
+    And   a model file test-model-infill.gguf
+    And   a model alias tinyllama-infill
+    And   42 as server seed
+    And   1024 as batch size
+    And   1024 as ubatch size
+    And   2048 KV cache size
+    And   64 max tokens to predict
+    And   0.0 temperature
+    Then  the server is starting
+    Then  the server is healthy
+
+  Scenario: Infill without input_extra
+    Given a prompt "Complete this"
+    And   an infill input extra none none
+    And   an infill input prefix "#include <cstdio>\n#include \"llama.h\"\n\nint main() {\n    int n_threads = llama_"
+    And   an infill input suffix "}\n"
+    And   an infill request with no api error
+    Then  64 tokens are predicted matching One|day|she|saw|big|scary|bird
+
+  Scenario: Infill with input_extra
+    Given a prompt "Complete this"
+    And   an infill input extra "llama.h" "LLAMA_API int32_t llama_n_threads();\n"
+    And   an infill input prefix "#include <cstdio>\n#include \"llama.h\"\n\nint main() {\n    int n_threads = llama_"
+    And   an infill input suffix "}\n"
+    And   an infill request with no api error
+    Then  64 tokens are predicted matching cuts|Jimmy|mom|came|into|the|room"

examples/server/tests/features/steps/steps.py

@@ -80,6 +80,11 @@ def step_server_config(context, server_fqdn: str, server_port: str):
     context.lora_file = None
     context.disable_ctx_shift = False
 
+    # infill
+    context.infill_input_extra = None
+    context.infill_input_suffix = ''
+    context.infill_input_prefix = ''
+
     context.tasks_result = []
     context.concurrent_tasks = []
     context.prompts = []
@@ -291,6 +296,28 @@ async def step_request_completion(context, api_error: Literal['raised'] | str):
         assert completion == api_error_code, f"completion must be an {api_error_code} status code: {completion}"
 
 
+@step('an infill request with {api_error} api error')
+@async_run_until_complete
+async def step_request_completion(context, api_error: Literal['raised'] | str):
+    if api_error != 'no':
+        raise ValueError(f'api_error={api_error} is not yet implemented')
+    payload = {
+        "prompt": context.prompts[0],
+        "input_suffix": context.infill_input_suffix,
+        "input_prefix": context.infill_input_prefix,
+        "n_predict": context.n_predict,
+        "seed": context.seed,
+        "temperature": context.temperature,
+    }
+    if context.infill_input_extra is not None:
+        payload['input_extra'] = context.infill_input_extra
+    async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
+        async with session.post(f'{context.base_url}/infill',
+                                json=payload) as response:
+            assert response.status == 200
+            context.tasks_result = [await response.json()]
+
+
 @step('{predicted_n:d} tokens are predicted matching {re_content}')
 def step_n_tokens_predicted_with_content(context, predicted_n, re_content):
     context.completion = context.tasks_result.pop()
@@ -539,6 +566,25 @@ def step_a_prompt_prompt(context, prompt):
     context.n_prompts = len(context.prompts)
 
 
+# TODO: allow this to be repeated
+@step('an infill input extra {filename} {text}')
+def step_infill_input_extra(context, filename, text):
+    if filename == 'none':
+        context.infill_input_extra = None
+    else:
+        context.infill_input_extra = [{'filename': filename, 'text': text}]
+
+
+@step('an infill input suffix {text}')
+def step_infill_input_suffix(context, text):
+    context.infill_input_suffix = text
+
+
+@step('an infill input prefix {text}')
+def step_infill_input_prefix(context, text):
+    context.infill_input_prefix = text
+
+
 @step('{num_prompts:d} prompts {prompt} with seed {seed:d}')
 def step_many_prompts(context, num_prompts, prompt, seed):
     if context.seed is None:

examples/server/utils.hpp

@@ -24,6 +24,22 @@
 #define DEFAULT_OAICOMPAT_MODEL "gpt-3.5-turbo-0613"
 
 using json = nlohmann::ordered_json;
+using llama_tokens = std::vector<llama_token>;
+
+#define SLT_INF(slot, fmt, ...) LOG_INF("slot %12.*s: id %2d | task %d | " fmt, 12, __func__, (slot).id, (slot).id_task, __VA_ARGS__)
+#define SLT_WRN(slot, fmt, ...) LOG_WRN("slot %12.*s: id %2d | task %d | " fmt, 12, __func__, (slot).id, (slot).id_task, __VA_ARGS__)
+#define SLT_ERR(slot, fmt, ...) LOG_ERR("slot %12.*s: id %2d | task %d | " fmt, 12, __func__, (slot).id, (slot).id_task, __VA_ARGS__)
+#define SLT_DBG(slot, fmt, ...) LOG_DBG("slot %12.*s: id %2d | task %d | " fmt, 12, __func__, (slot).id, (slot).id_task, __VA_ARGS__)
+
+#define SRV_INF(fmt, ...) LOG_INF("srv  %12.*s: " fmt, 12, __func__, __VA_ARGS__)
+#define SRV_WRN(fmt, ...) LOG_WRN("srv  %12.*s: " fmt, 12, __func__, __VA_ARGS__)
+#define SRV_ERR(fmt, ...) LOG_ERR("srv  %12.*s: " fmt, 12, __func__, __VA_ARGS__)
+#define SRV_DBG(fmt, ...) LOG_DBG("srv  %12.*s: " fmt, 12, __func__, __VA_ARGS__)
+
+#define QUE_INF(fmt, ...) LOG_INF("que  %12.*s: " fmt, 12, __func__, __VA_ARGS__)
+#define QUE_WRN(fmt, ...) LOG_WRN("que  %12.*s: " fmt, 12, __func__, __VA_ARGS__)
+#define QUE_ERR(fmt, ...) LOG_ERR("que  %12.*s: " fmt, 12, __func__, __VA_ARGS__)
+#define QUE_DBG(fmt, ...) LOG_DBG("que  %12.*s: " fmt, 12, __func__, __VA_ARGS__)
 
 // https://community.openai.com/t/openai-chat-list-of-error-codes-and-types/357791/11
 enum error_type {
@@ -52,9 +68,235 @@ static T json_value(const json & body, const std::string & key, const T & defaul
 }
 
 //
-// chat template utils
+// tokenizer and input processing utils
 //
 
+static bool json_is_array_of_numbers(const json & data) {
+    if (data.is_array()) {
+        for (const auto & e : data) {
+            if (!e.is_number_integer()) {
+                return false;
+            }
+        }
+        return true;
+    }
+    return false;
+}
+
+// is array having BOTH numbers & strings?
+static bool json_is_array_of_mixed_numbers_strings(const json & data) {
+    bool seen_string = false;
+    bool seen_number = false;
+    if (data.is_array()) {
+        for (const auto & e : data) {
+            seen_string |= e.is_string();
+            seen_number |= e.is_number_integer();
+            if (seen_number && seen_string) {
+                return true;
+            }
+        }
+    }
+    return false;
+}
+
+/**
+ * this handles 2 cases:
+ * - only string, example: "string"
+ * - mixed string and tokens, example: [12, 34, "string", 56, 78]
+ */
+static llama_tokens tokenize_mixed(const llama_context * ctx, const json & json_prompt, bool add_special, bool parse_special) {
+    // If `add_bos` is true, we only add BOS, when json_prompt is a string,
+    // or the first element of the json_prompt array is a string.
+    llama_tokens prompt_tokens;
+
+    if (json_prompt.is_array()) {
+        bool first = true;
+        for (const auto & p : json_prompt) {
+            if (p.is_string()) {
+                auto s = p.template get<std::string>();
+
+                llama_tokens p;
+                if (first) {
+                    p = common_tokenize(ctx, s, add_special, parse_special);
+                    first = false;
+                } else {
+                    p = common_tokenize(ctx, s, false, parse_special);
+                }
+
+                prompt_tokens.insert(prompt_tokens.end(), p.begin(), p.end());
+            } else {
+                if (first) {
+                    first = false;
+                }
+
+                prompt_tokens.push_back(p.template get<llama_token>());
+            }
+        }
+    } else {
+        auto s = json_prompt.template get<std::string>();
+        prompt_tokens = common_tokenize(ctx, s, add_special, parse_special);
+    }
+
+    return prompt_tokens;
+}
+
+/**
+ * break the input "prompt" object into multiple prompt if needed, then tokenize them
+ * this supports these cases:
+ * - "prompt": "string"
+ * - "prompt": [12, 34, 56]
+ * - "prompt": [12, 34, "string", 56, 78]
+ * and multiple prompts (multi-tasks):
+ * - "prompt": ["string1", "string2"]
+ * - "prompt": ["string1", [12, 34, 56]]
+ * - "prompt": [[12, 34, "string", 56, 78], [12, 34, 56]]
+ */
+static std::vector<llama_tokens> tokenize_input_prompts(llama_context * ctx, const json & json_prompt, bool add_special, bool parse_special) {
+    std::vector<llama_tokens> result;
+    if (json_prompt.is_string() || json_is_array_of_mixed_numbers_strings(json_prompt)) {
+        // string or mixed
+        result.push_back(tokenize_mixed(ctx, json_prompt, add_special, parse_special));
+    } else if (json_is_array_of_numbers(json_prompt)) {
+        // array of tokens
+        result.push_back(json_prompt.get<llama_tokens>());
+    } else if (json_prompt.is_array()) {
+        // array of prompts
+        result.reserve(json_prompt.size());
+        for (const auto & p : json_prompt) {
+            if (p.is_string() || json_is_array_of_mixed_numbers_strings(p)) {
+                result.push_back(tokenize_mixed(ctx, p, add_special, parse_special));
+            } else if (json_is_array_of_numbers(p)) {
+                // array of tokens
+                result.push_back(p.get<llama_tokens>());
+            } else {
+                throw std::runtime_error("element of \"prompt\" must be a string, an list of tokens, or a list of mixed strings & tokens");
+            }
+        }
+    } else {
+        throw std::runtime_error("\"prompt\" must be a string, an list of tokens, a list of mixed strings & tokens, or a list of prompts");
+    }
+    return result;
+}
+
+//
+// template utils
+//
+
+// format rerank task: [BOS]query[EOS][SEP]doc[EOS]
+static llama_tokens format_rerank(const struct llama_model * model, const llama_tokens & query, const llama_tokens & doc) {
+    llama_tokens result;
+    result.reserve(doc.size() + query.size() + 4);
+    result.push_back(llama_token_bos(model));
+    result.insert(result.end(), query.begin(), query.end());
+    result.push_back(llama_token_eos(model));
+    result.push_back(llama_token_sep(model));
+    result.insert(result.end(), doc.begin(), doc.end());
+    result.push_back(llama_token_eos(model));
+    return result;
+}
+
+// format infill task
+static llama_tokens format_infill(
+        const llama_context * ctx,
+        const json & input_prefix,
+        const json & input_suffix,
+        const json & input_extra,
+        const int n_batch,
+        const int n_predict,
+        const int n_ctx,
+        const bool spm_infill,
+        const llama_tokens & tokens_prompt
+    ) {
+    // TODO: optimize this block by reducing memory allocations and movement
+
+    // use FIM repo-level pattern:
+    // ref: https://arxiv.org/pdf/2409.12186
+    //
+    // [FIM_REP]myproject
+    // [FIM_SEP]filename0
+    // extra chunk 0
+    // [FIM_SEP]filename1
+    // extra chunk 1
+    // ...
+    // [FIM_SEP]filename
+    // [FIM_PRE]prefix[FIM_SUF]suffix[FIM_MID]prompt
+    //
+    llama_tokens extra_tokens;
+    extra_tokens.reserve(n_ctx);
+
+    auto model = llama_get_model(ctx);
+    auto tokens_prefix = tokenize_mixed(ctx, input_prefix, false, false);
+    auto tokens_suffix = tokenize_mixed(ctx, input_suffix, false, false);
+
+    if (llama_token_fim_rep(model) != LLAMA_TOKEN_NULL) {
+        // TODO: make project name an input
+        static const auto k_fim_repo = common_tokenize(ctx, "myproject\n", false, false);
+
+        extra_tokens.push_back(llama_token_fim_rep(model));
+        extra_tokens.insert(extra_tokens.end(), k_fim_repo.begin(), k_fim_repo.end());
+    }
+    for (const auto & chunk : input_extra) {
+        // { "text": string, "filename": string }
+        const std::string text     = json_value(chunk, "text",     std::string());
+        const std::string filename = json_value(chunk, "filename", std::string("tmp"));
+
+        if (llama_token_fim_sep(model) != LLAMA_TOKEN_NULL) {
+            const auto k_fim_file = common_tokenize(ctx, filename + "\n", false, false);
+
+            extra_tokens.insert(extra_tokens.end(), llama_token_fim_sep(model));
+            extra_tokens.insert(extra_tokens.end(), k_fim_file.begin(), k_fim_file.end());
+        } else {
+            // chunk separator in binary form to avoid confusing the AI
+            static const char k_chunk_prefix_str[] = {0x0a, 0x0a, 0x2d, 0x2d, 0x2d, 0x20, 0x73, 0x6e, 0x69, 0x70, 0x70, 0x65, 0x74, 0x20, 0x2d, 0x2d, 0x2d, 0x0a, 0x0a, 0x00};
+            static const auto k_chunk_prefix_tokens = common_tokenize(ctx, k_chunk_prefix_str, false, false);
+
+            extra_tokens.insert(extra_tokens.end(), k_chunk_prefix_tokens.begin(), k_chunk_prefix_tokens.end());
+        }
+
+        const auto chunk_tokens = common_tokenize(ctx, text, false, false);
+        extra_tokens.insert(extra_tokens.end(), chunk_tokens.begin(), chunk_tokens.end());
+    }
+
+    if (llama_token_fim_sep(model) != LLAMA_TOKEN_NULL) {
+        // TODO: current filename
+        static const auto k_fim_file = common_tokenize(ctx, "filename\n", false, false);
+
+        extra_tokens.insert(extra_tokens.end(), llama_token_fim_sep(model));
+        extra_tokens.insert(extra_tokens.end(), k_fim_file.begin(), k_fim_file.end());
+    }
+
+    // for now pick FIM context to fit in a batch (ratio prefix:suffix = 3:1, TODO: configurable?)
+    const int n_suffix_take = std::min<int>(tokens_suffix.size(),   (n_batch/4));
+    const int n_prefix_take = std::min<int>(tokens_prefix.size(), 3*(n_batch/4) - 3);
+
+    // fill the rest of the context with extra chunks
+    const int n_extra_take = std::min<int>(std::max<int>(0, n_ctx - (n_batch) - 2*n_predict), extra_tokens.size());
+
+    tokens_prefix.erase(tokens_prefix.begin(), tokens_prefix.begin() + tokens_prefix.size() - n_prefix_take);
+    tokens_suffix.resize(n_suffix_take);
+
+    tokens_prefix.insert(tokens_prefix.begin(), llama_token_fim_pre(model));
+    tokens_prefix.insert(tokens_prefix.end(),   tokens_prompt.begin(), tokens_prompt.end());
+    tokens_suffix.insert(tokens_suffix.begin(), llama_token_fim_suf(model));
+
+    auto embd_inp = spm_infill ? tokens_suffix : tokens_prefix;
+    auto embd_end = spm_infill ? tokens_prefix : tokens_suffix;
+
+    if (llama_add_bos_token(model)) {
+        embd_inp.insert(embd_inp.begin(), llama_token_bos(model));
+    }
+
+    SRV_DBG("extra: n_ctx = %d, n_extra_take = %d, n_extra = %d\n", n_ctx, n_extra_take, (int) extra_tokens.size());
+
+    // put the extra context before the FIM prefix
+    embd_inp.insert(embd_inp.begin(), extra_tokens.end() - n_extra_take, extra_tokens.end());
+
+    embd_inp.insert(embd_inp.end(), embd_end.begin(), embd_end.end());
+    embd_inp.push_back(llama_token_fim_mid(model));
+
+    return embd_inp;
+}
+
 // Format given chat. If tmpl is empty, we take the template from model metadata
 inline std::string format_chat(const struct llama_model * model, const std::string & tmpl, const std::vector<json> & messages) {
     std::vector<common_chat_msg> chat;
@@ -229,18 +471,6 @@ static size_t find_partial_stop_string(const std::string &stop, const std::strin
     return std::string::npos;
 }
 
-static bool json_is_array_of_numbers(const json & data) {
-    if (data.is_array()) {
-        for (const auto & e : data) {
-            if (!e.is_number()) {
-                return false;
-            }
-        }
-        return true;
-    }
-    return false;
-}
-
 // TODO: reuse llama_detokenize
 template <class Iter>
 static std::string tokens_to_str(llama_context * ctx, Iter begin, Iter end) {

examples/speculative/speculative.cpp

@@ -39,6 +39,11 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
+    if (params.n_predict < -1) {
+        LOG_ERR("%s: --n-predict must be >= -1\n", __func__);
+        return 1;
+    }
+
     common_init();
 
     if (params.model_draft.empty()) {
@@ -180,8 +185,6 @@ int main(int argc, char ** argv) {
     // target model sampling context (reuse the llama_context's sampling instance)
     struct common_sampler * smpl = common_sampler_init(model_tgt, params.sparams);
 
-    struct llama_sampler * softmax = llama_sampler_init_softmax();
-
     // draft sequence data
     std::vector<seq_draft> drafts(n_seq_dft);
 
@@ -190,8 +193,8 @@ int main(int argc, char ** argv) {
         drafts[s].smpl = common_sampler_init(model_dft, params.sparams);
     }
 
-    llama_batch batch_dft = llama_batch_init(params.n_ctx, 0, 1);
-    llama_batch batch_tgt = llama_batch_init(params.n_ctx, 0, n_seq_dft);
+    llama_batch batch_dft = llama_batch_init(llama_n_batch(ctx_dft), 0, 1);
+    llama_batch batch_tgt = llama_batch_init(llama_n_batch(ctx_tgt), 0, n_seq_dft);
 
     const auto t_dec_start = ggml_time_us();
 
@@ -441,7 +444,7 @@ int main(int argc, char ** argv) {
             ++n_past_dft;
         }
 
-        if (n_predict > params.n_predict || has_eos) {
+        if ((params.n_predict >= 0 && n_predict > params.n_predict) || has_eos) {
             break;
         }
 
@@ -624,7 +627,6 @@ int main(int argc, char ** argv) {
         common_sampler_free(drafts[s].smpl);
     }
 
-    llama_sampler_free(softmax);
     llama_batch_free(batch_dft);
 
     llama_free(ctx_tgt);

flake.lock

@@ -20,11 +20,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1728492678,
-        "narHash": "sha256-9UTxR8eukdg+XZeHgxW5hQA9fIKHsKCdOIUycTryeVw=",
+        "lastModified": 1729256560,
+        "narHash": "sha256-/uilDXvCIEs3C9l73JTACm4quuHUsIHcns1c+cHUJwA=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "5633bcff0c6162b9e4b5f1264264611e950c8ec7",
+        "rev": "4c2fcb090b1f3e5b47eaa7bd33913b574a11e0a0",
         "type": "github"
       },
       "original": {

ggml/include/ggml-cann.h

@@ -34,6 +34,8 @@ extern "C" {
  */
 #define GGML_CANN_MAX_DEVICES 16
 
+GGML_API ggml_backend_reg_t ggml_backend_cann_reg(void);
+
 /**
  * @brief Initializes the CANN backend for a specified device.
  *

ggml/src/ggml-backend.cpp

@@ -561,6 +561,10 @@ void * ggml_backend_reg_get_proc_address(ggml_backend_reg_t reg, const char * na
 #  include "ggml-amx.h"
 #endif
 
+#ifdef GGML_USE_CANN
+#include "ggml-cann.h"
+#endif
+
 struct ggml_backend_registry {
     std::vector<ggml_backend_reg_t> backends;
     std::vector<ggml_backend_dev_t> devices;
@@ -587,8 +591,11 @@ struct ggml_backend_registry {
 #ifdef GGML_USE_AMX
         register_backend(ggml_backend_amx_reg());
 #endif
+#ifdef GGML_USE_CANN
+        register_backend(ggml_backend_cann_reg());
+#endif
 
-        // TODO: kompute, cann
+        // TODO: kompute
 
         register_backend(ggml_backend_cpu_reg());
     }

ggml/src/ggml-cann.cpp

@@ -39,6 +39,8 @@
 
 #include "ggml-common.h"
 
+#define GGML_CANN_NAME "CANN"
+
 /**
  * @brief Handles CANN errors by printing an error message and aborting.
  *
@@ -851,13 +853,6 @@ static void ggml_backend_cann_buffer_set_tensor(
         void *transform_buffer = malloc(size);
         ggml_backend_cann_transform(tensor, data, transform_buffer);
 
-#ifndef NDEBUG
-        void *check_buffer = malloc(size);
-        ggml_backend_cann_transform_back(tensor, transform_buffer,
-                                         check_buffer);
-        GGML_ASSERT(memcmp(data, check_buffer, size) == 0);
-        free(check_buffer);
-#endif
         ACL_CHECK(aclrtMemcpy((char *)tensor->data + offset, size,
                               transform_buffer, size,
                               ACL_MEMCPY_HOST_TO_DEVICE));
@@ -969,7 +964,7 @@ static void ggml_backend_cann_buffer_clear(
  * This structure defines function pointers to operations that can be performed
  * on a CANN buffer within the backend.
  */
-static ggml_backend_buffer_i ggml_backend_cann_buffer_interface = {
+static const ggml_backend_buffer_i ggml_backend_cann_buffer_interface = {
     /* .get_name        = */ ggml_backend_cann_buffer_get_name,
     /* .free_buffer     = */ ggml_backend_cann_buffer_free_buffer,
     /* .get_base        = */ ggml_backend_cann_buffer_get_base,
@@ -1105,19 +1100,25 @@ static size_t ggml_backend_cann_buffer_type_get_alloc_size(
     GGML_UNUSED(buft);
 }
 
+static bool ggml_backend_cann_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
+    return false;
+
+    GGML_UNUSED(buft);
+}
+
 /**
  * @brief Interface for managing CANN buffer types in the GGML backend.
  *
  * Provides function pointers for allocating, querying properties, and managing
  * memory for CANN buffer types in the GGML backend.
  */
-static ggml_backend_buffer_type_i ggml_backend_cann_buffer_type_interface = {
+static const ggml_backend_buffer_type_i ggml_backend_cann_buffer_type_interface = {
     /* .get_name         = */ ggml_backend_cann_buffer_type_name,
     /* .alloc_buffer     = */ ggml_backend_cann_buffer_type_alloc_buffer,
     /* .get_alignment    = */ ggml_backend_cann_buffer_type_get_alignment,
     /* .get_max_size     = */ NULL,  // defaults to SIZE_MAX
     /* .get_alloc_size   = */ ggml_backend_cann_buffer_type_get_alloc_size,
-    /* .is_host          = */ NULL,
+    /* .is_host          = */ ggml_backend_cann_buffer_type_is_host,
 };
 
 /**
@@ -1148,7 +1149,7 @@ ggml_backend_cann_buffer_type(int32_t device) {
         for (int32_t i = 0; i < GGML_CANN_MAX_DEVICES; i++) {
             ggml_backend_cann_buffer_types[i] = {
                 /* .iface    = */ ggml_backend_cann_buffer_type_interface,
-                /* .device    = */ nullptr,
+                /* .device    = */ ggml_backend_reg_dev_get(ggml_backend_cann_reg(), device),
                 /* .context  = */
                  new ggml_backend_cann_buffer_type_context{
                     i, "CANN" + std::to_string(i)},
@@ -1264,7 +1265,7 @@ ggml_backend_buffer_type_t ggml_backend_cann_host_buffer_type() {
             /* .get_alloc_size   = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size,
             /* .is_host          = */ ggml_backend_cpu_buffer_type()->iface.is_host,
         },
-        /* .device   = */ nullptr,
+        /* .device   = */ ggml_backend_reg_dev_get(ggml_backend_cann_reg(), 0),
         /* .context  = */ nullptr,
     };
 
@@ -1511,13 +1512,6 @@ static void ggml_backend_cann_set_tensor_async(ggml_backend_t backend,
         void *transform_buffer = malloc(size);
         ggml_backend_cann_transform(tensor, data, transform_buffer);
 
-#ifndef NDEBUG
-        void *check_buffer = malloc(size);
-        ggml_backend_cann_transform_back(tensor, transform_buffer,
-                                         check_buffer);
-        GGML_ASSERT(memcmp(data, check_buffer, size));
-        free(check_buffer);
-#endif
         ACL_CHECK(aclrtMemcpyAsync(
             (char *)tensor->data + offset, size, transform_buffer, size,
             ACL_MEMCPY_HOST_TO_DEVICE, cann_ctx->stream()));
@@ -1692,7 +1686,7 @@ static enum ggml_status ggml_backend_cann_graph_compute(
  * @return bool Returns true if the operation is supported by the backend,
  *              otherwise false.
  */
-static bool ggml_backend_cann_supports_op(ggml_backend_t backend,
+static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
                                                     const ggml_tensor* op) {
     switch (op->op) {
         case GGML_OP_UNARY:
@@ -1783,7 +1777,7 @@ static bool ggml_backend_cann_supports_op(ggml_backend_t backend,
             return false;
     }
 
-    GGML_UNUSED(backend);
+    GGML_UNUSED(dev);
 }
 
 /**
@@ -1801,31 +1795,6 @@ static bool ggml_backend_buft_is_cann(ggml_backend_buffer_type_t buft) {
     return buft->iface.get_name == ggml_backend_cann_buffer_type_name;
 }
 
-/**
- * @brief Checks if the CANN backend supports a specific backend buffer type.
- *
- * This function determines whether the CANN backend supports the given backend
- * buffer type by comparing the device context of the backend and buffer type.
- * It returns true if the devices are same between the backend context and
- * buffer type context.
- *
- * @param backend Pointer to the CANN backend.
- * @param buft Pointer to the backend buffer type to check.
- * @return bool Returns true if the CANN backend supports the buffer type,
- *              otherwise false.
- */
-static bool ggml_backend_cann_supports_buft(
-    ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
-    if (ggml_backend_buft_is_cann(buft)) {
-        ggml_backend_cann_context * cann_ctx =
-                        (ggml_backend_cann_context *)backend->context;
-        ggml_backend_cann_buffer_type_context * buft_ctx =
-                        (ggml_backend_cann_buffer_type_context *)buft->context;
-        return buft_ctx->device == cann_ctx->device;
-    }
-    return false;
-}
-
 /**
  * @brief Determines if a tensor operation should be offloaded to the CANN
  * backend.
@@ -1840,54 +1809,14 @@ static bool ggml_backend_cann_supports_buft(
  * @return bool Returns true if the operation should be offloaded, otherwise
  * false.
  */
-static bool ggml_backend_cann_offload_op(ggml_backend_t backend,
+static bool ggml_backend_cann_offload_op(ggml_backend_dev_t dev,
                                                    const ggml_tensor* op) {
     const int min_batch_size = 32;
-    GGML_UNUSED(backend);
+    GGML_UNUSED(dev);
 
     return op->ne[1] >= min_batch_size && op->op != GGML_OP_GET_ROWS;
 }
 
-/**
- * @brief Creates a new event for the CANN backend.
- *
- * This function initializes a new event for the CANN backend by setting the
- * device and creating an ACL runtime event. The created event is then wrapped
- * in a ggml_backend_event structure and returned.
- *
- * @param backend Pointer to the CANN backend.
- * @return ggml_backend_event_t Returns a pointer to the new event structure.
- */
-static ggml_backend_event_t ggml_backend_cann_event_new(
-    ggml_backend_t backend) {
-    ggml_backend_cann_context* cann_ctx =
-        (ggml_backend_cann_context*)backend->context;
-
-    ggml_cann_set_device(cann_ctx->device);
-
-    aclrtEvent event;
-    ACL_CHECK(aclrtCreateEvent(&event));
-
-    return new ggml_backend_event{
-        /* .device = */ nullptr,
-        /* .context = */ event,
-    };
-}
-
-/**
- * @brief Frees a CANN backend event.
- *
- * This function destroys the ACL runtime event associated with the given CANN
- * backend event and then deletes the event structure itself.
- *
- * @param event Pointer to the event structure to be freed.
- */
-static void ggml_backend_cann_event_free(ggml_backend_event_t event) {
-    ACL_CHECK(aclrtDestroyEvent((aclrtEvent)event->context));
-
-    delete event;
-}
-
 /**
  * @brief Records an event on the CANN backend stream.
  *
@@ -1924,17 +1853,6 @@ static void ggml_backend_cann_event_wait(ggml_backend_t backend,
     }
 }
 
-/**
- * @brief Synchronizes the given event on the CANN backend.
- *
- * This function waits for the specified event to complete on the ACL runtime.
- *
- * @param event Pointer to the event structure to be synchronized.
- */
-static void ggml_backend_cann_event_synchronize(ggml_backend_event_t event) {
-    ACL_CHECK(aclrtSynchronizeEvent((aclrtEvent)event->context));
-}
-
 /**
  * @brief Structure defining the interface for the CANN backend.
  *
@@ -1942,7 +1860,7 @@ static void ggml_backend_cann_event_synchronize(ggml_backend_event_t event) {
  * supported by the CANN backend, including name retrieval, memory
  * management, tensor operations, synchronization, and event handling.
  */
-static ggml_backend_i ggml_backend_cann_interface = {
+static const ggml_backend_i ggml_backend_cann_interface = {
     /* .get_name                = */ ggml_backend_cann_name,
     /* .free                    = */ ggml_backend_cann_free,
     /* .get_default_buffer_type = */ ggml_backend_cann_get_default_buffer_type,
@@ -1955,9 +1873,9 @@ static ggml_backend_i ggml_backend_cann_interface = {
     /* .graph_plan_update       = */ NULL,
     /* .graph_plan_compute      = */ NULL,
     /* .graph_compute           = */ ggml_backend_cann_graph_compute,
-    /* .supports_op             = */ ggml_backend_cann_supports_op,
-    /* .supports_buft           = */ ggml_backend_cann_supports_buft,
-    /* .offload_op              = */ ggml_backend_cann_offload_op,
+    /* .supports_op             = */ NULL, // moved to device
+    /* .supports_buft           = */ NULL, // moved to device
+    /* .offload_op              = */ NULL, // moved to device
     /* .event_record            = */ ggml_backend_cann_event_record,
     /* .event_wait              = */ ggml_backend_cann_event_wait,
 };
@@ -1976,6 +1894,234 @@ static ggml_guid_t ggml_backend_cann_guid() {
     return &guid;
 }
 
+// backend device
+struct ggml_backend_cann_device_context {
+    int device;
+    std::string name;
+    std::string description;
+};
+
+static const char * ggml_backend_cann_device_get_name(ggml_backend_dev_t dev) {
+    ggml_backend_cann_device_context * ctx = (ggml_backend_cann_device_context *)dev->context;
+    return ctx->name.c_str();
+}
+
+static const char* ggml_backend_cann_device_get_description(ggml_backend_dev_t dev) {
+    ggml_backend_cann_device_context * ctx = (ggml_backend_cann_device_context *)dev->context;
+    return ctx->description.c_str();
+}
+
+static void ggml_backend_cann_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
+    ggml_backend_cann_device_context * ctx = (ggml_backend_cann_device_context *)dev->context;
+    ggml_backend_cann_get_device_memory(ctx->device, free, total);
+}
+
+static enum ggml_backend_dev_type ggml_backend_cann_device_get_type(ggml_backend_dev_t dev) {
+    GGML_UNUSED(dev);
+    return GGML_BACKEND_DEVICE_TYPE_GPU_FULL;
+}
+
+static void ggml_backend_cann_device_get_props(ggml_backend_dev_t dev, ggml_backend_dev_props * props) {
+    props->name        = ggml_backend_cann_device_get_name(dev);
+    props->description = ggml_backend_cann_device_get_description(dev);
+    props->type        = ggml_backend_cann_device_get_type(dev);
+    ggml_backend_cann_device_get_memory(dev, &props->memory_free, &props->memory_total);
+
+    bool host_buffer = getenv("GGML_CANN_NO_PINNED") == nullptr;
+
+    props->caps = {
+        /* .async                 = */ false,
+        /* .host_buffer           = */ host_buffer,
+        /* .buffer_from_host_ptr  = */ false,
+        /* .events                = */ true,
+    };
+}
+
+static ggml_backend_t ggml_backend_cann_device_init(ggml_backend_dev_t dev, const char * params) {
+    GGML_UNUSED(params);
+    ggml_backend_cann_device_context * ctx = (ggml_backend_cann_device_context *)dev->context;
+    return ggml_backend_cann_init(ctx->device);
+}
+
+/**
+ * @brief Checks if the CANN backend supports a specific backend buffer type.
+ *
+ * This function determines whether the CANN backend supports the given backend
+ * buffer type by comparing the device context of the backend and buffer type.
+ * It returns true if the devices are same between the backend context and
+ * buffer type context.
+ *
+ * @param backend Pointer to the CANN backend.
+ * @param buft Pointer to the backend buffer type to check.
+ * @return bool Returns true if the CANN backend supports the buffer type,
+ *              otherwise false.
+ */
+static bool ggml_backend_cann_supports_buft(
+    ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
+    if (ggml_backend_buft_is_cann(buft)) {
+        ggml_backend_cann_device_context * dev_ctx = (ggml_backend_cann_device_context *)dev->context;
+        ggml_backend_cann_buffer_type_context * buft_ctx =
+                        (ggml_backend_cann_buffer_type_context *)buft->context;
+        return buft_ctx->device == dev_ctx->device;
+    }
+    return false;
+}
+
+static ggml_backend_buffer_type_t ggml_backend_cann_device_get_buffer_type(ggml_backend_dev_t dev) {
+    ggml_backend_cann_device_context * ctx = (ggml_backend_cann_device_context *)dev->context;
+    return ggml_backend_cann_buffer_type(ctx->device);
+}
+
+static ggml_backend_buffer_type_t ggml_backend_cann_device_get_host_buffer_type(ggml_backend_dev_t dev) {
+    GGML_UNUSED(dev);
+    return ggml_backend_cann_host_buffer_type();
+}
+
+/**
+ * @brief Creates a new event for the CANN backend device.
+ *
+ * This function initializes a new event for the CANN backend by setting the
+ * device and creating an ACL runtime event. The created event is then wrapped
+ * in a ggml_backend_event structure and returned.
+ *
+ * @param backend Pointer to the CANN backend.
+ * @return ggml_backend_event_t Returns a pointer to the new event structure.
+ */
+static ggml_backend_event_t ggml_backend_cann_device_event_new(
+    ggml_backend_dev_t dev) {
+    ggml_backend_cann_device_context * dev_ctx = (ggml_backend_cann_device_context *)dev->context;
+
+    ggml_cann_set_device(dev_ctx->device);
+
+    aclrtEvent event;
+    ACL_CHECK(aclrtCreateEvent(&event));
+
+    return new ggml_backend_event{
+        /* .device = */ ggml_backend_reg_dev_get(ggml_backend_cann_reg(), dev_ctx->device),
+        /* .context = */ event,
+    };
+}
+
+/**
+ * @brief Frees a CANN backend event.
+ *
+ * This function destroys the ACL runtime event associated with the given CANN
+ * backend event and then deletes the event structure itself.
+ *
+ * @param event Pointer to the event structure to be freed.
+ */
+static void ggml_backend_cann_device_event_free(ggml_backend_dev_t dev, ggml_backend_event_t event) {
+    ACL_CHECK(aclrtDestroyEvent((aclrtEvent)event->context));
+
+    delete event;
+    GGML_UNUSED(dev);
+}
+
+/**
+ * @brief Synchronizes the given event on the CANN backend.
+ *
+ * This function waits for the specified event to complete on the ACL runtime.
+ *
+ * @param event Pointer to the event structure to be synchronized.
+ */
+static void ggml_backend_cann_device_event_synchronize(ggml_backend_dev_t dev, ggml_backend_event_t event) {
+    ACL_CHECK(aclrtSynchronizeEvent((aclrtEvent)event->context));
+
+    GGML_UNUSED(dev);
+}
+
+static const ggml_backend_device_i ggml_backend_cann_device_interface = {
+    /* .get_name                = */ ggml_backend_cann_device_get_name,
+    /* .get_description         = */ ggml_backend_cann_device_get_description,
+    /* .get_memory              = */ ggml_backend_cann_device_get_memory,
+    /* .get_type                = */ ggml_backend_cann_device_get_type,
+    /* .get_props               = */ ggml_backend_cann_device_get_props,
+    /* .init_backend            = */ ggml_backend_cann_device_init,    // called for every card
+    /* .get_buffer_type         = */ ggml_backend_cann_device_get_buffer_type,
+    /* .get_host_buffer_type    = */ ggml_backend_cann_device_get_host_buffer_type,
+    /* .buffer_from_host_ptr    = */ NULL, // not supported for CANN
+    /* .supports_op             = */ ggml_backend_cann_supports_op,
+    /* .supports_buft           = */ ggml_backend_cann_supports_buft,
+    /* .offload_op              = */ ggml_backend_cann_offload_op,
+    /* .event_new               = */ ggml_backend_cann_device_event_new,
+    /* .event_free              = */ ggml_backend_cann_device_event_free,
+    /* .event_synchronize       = */ ggml_backend_cann_device_event_synchronize,
+};
+
+
+// backend reg
+struct ggml_backend_cann_reg_context {
+    std::vector<ggml_backend_dev_t> devices;
+};
+
+static const char * ggml_backend_cann_reg_get_name(ggml_backend_reg_t reg) {
+    GGML_UNUSED(reg);
+    return GGML_CANN_NAME;
+}
+
+static size_t ggml_backend_cann_reg_get_device_count(ggml_backend_reg_t reg) {
+    ggml_backend_cann_reg_context * ctx = (ggml_backend_cann_reg_context *)reg->context;
+    return ctx->devices.size();
+}
+
+static ggml_backend_dev_t ggml_backend_cann_reg_get_device(ggml_backend_reg_t reg, size_t index) {
+    ggml_backend_cann_reg_context * ctx = (ggml_backend_cann_reg_context *)reg->context;
+    GGML_ASSERT(index < ctx->devices.size());
+    return ctx->devices[index];
+}
+
+static void * ggml_backend_cann_reg_get_proc_address(ggml_backend_reg_t reg, const char * name) {
+    GGML_UNUSED(reg);
+    GGML_UNUSED(name);
+    // reserved for future use
+    return nullptr;
+}
+
+static const ggml_backend_reg_i ggml_backend_cann_reg_interface = {
+    /* .get_name          = */ ggml_backend_cann_reg_get_name,
+    /* .get_device_count  = */ ggml_backend_cann_reg_get_device_count,
+    /* .get_device_get    = */ ggml_backend_cann_reg_get_device,
+    /* .get_proc_address  = */ ggml_backend_cann_reg_get_proc_address,
+};
+
+// backend registry, called only once for cann backend
+ggml_backend_reg_t ggml_backend_cann_reg() {
+    static ggml_backend_reg reg;
+    static bool initialized = false;
+
+    {
+        static std::mutex mutex;
+        std::lock_guard<std::mutex> lock(mutex);
+        if (!initialized) {
+            aclInit(nullptr);
+            ggml_backend_cann_reg_context * ctx = new ggml_backend_cann_reg_context;
+
+            for (int i = 0; i < ggml_cann_info().device_count; i++) {
+                ggml_backend_cann_device_context* dev_ctx = new ggml_backend_cann_device_context();
+                dev_ctx->description = aclrtGetSocName();
+                dev_ctx->device = i;
+                dev_ctx->name = GGML_CANN_NAME + std::to_string(i);
+                ggml_cann_set_device(i);
+                ggml_backend_dev_t dev = new ggml_backend_device {
+                    /* .interface = */ ggml_backend_cann_device_interface,
+                    /* .reg       = */ &reg,
+                    /* .context   = */ dev_ctx
+                };
+                ctx->devices.push_back(dev);
+            }
+
+            reg = ggml_backend_reg {
+                /* .interface = */ ggml_backend_cann_reg_interface,
+                /* .context   = */ ctx
+            };
+        }
+
+        initialized = true;
+    }
+
+    return &reg;
+}
+
 ggml_backend_t ggml_backend_cann_init(int32_t device) {
     aclInit(nullptr);
     if (device < 0 || device >= ggml_backend_cann_get_device_count()) {
@@ -1992,7 +2138,7 @@ ggml_backend_t ggml_backend_cann_init(int32_t device) {
     ggml_backend_t cann_backend =
         new ggml_backend{/* .guid      = */ ggml_backend_cann_guid(),
                          /* .interface = */ ggml_backend_cann_interface,
-                         /* .device    = */ nullptr,
+                         /* .device    = */ ggml_backend_reg_dev_get(ggml_backend_cann_reg(), device),
                          /* .context   = */ ctx};
 
     return cann_backend;

ggml/src/ggml-cuda.cu

@@ -1151,8 +1151,8 @@ static cudaError_t ggml_cuda_cpy_tensor_2d(
     void * dst, const struct ggml_tensor * src, int64_t i3, int64_t i2, int64_t i1_low, int64_t i1_high, cudaStream_t stream) {
 
     GGML_ASSERT(ggml_backend_buffer_is_cuda(src->buffer));
-    char * src_ptr = (char *) src->data;
-    char * dst_ptr = (char *) dst;
+    const char * src_ptr = (const char *) src->data;
+    char       * dst_ptr = (char       *) dst;
 
     const int64_t ne0 = src->ne[0];
     const int64_t nb0 = src->nb[0];
@@ -1162,7 +1162,7 @@ static cudaError_t ggml_cuda_cpy_tensor_2d(
     const enum ggml_type type = src->type;
     const int64_t ts = ggml_type_size(type);
     const int64_t bs = ggml_blck_size(type);
-    int64_t i1_diff = i1_high - i1_low;
+    const int64_t i1_diff = i1_high - i1_low;
 
     const char * x = src_ptr + i1_low*nb1 + i2*nb2 + i3*nb3;
     if (nb0 == ts && nb1 == ts*ne0/bs) {
@@ -1479,13 +1479,18 @@ static void ggml_cuda_op_mul_mat(
         if (src0_is_contiguous) {
             dev[id].src0_dd = split ? (char *) src0_extra->data_device[id] : (char *) src0->data;
         } else {
-            dev[id].src0_dd = dev[id].src0_dd_alloc.alloc(ctx.pool(id), ggml_nbytes(src0));
+            // If src0 is not contiguous it will be copied to a temporary buffer.
+            // This buffer needs to be cleared entirely because multiple regions will function as padding.
+            const size_t nbytes_data    = ggml_nbytes(src0);
+            const size_t nbytes_padding = ggml_row_size(src0->type, MATRIX_ROW_PADDING - ne00 % MATRIX_ROW_PADDING);
+            dev[id].src0_dd = dev[id].src0_dd_alloc.alloc(ctx.pool(id), nbytes_data + nbytes_padding);
+            CUDA_CHECK(cudaMemsetAsync(dev[id].src0_dd, 0, nbytes_data + nbytes_padding, stream));
         }
 
-        // If src0 is on a temporary compute buffers (partial offloading) there may be some padding that needs to be cleared:
+        // If src0 is on a temporary compute buffer (partial offloading) there may be some padding that needs to be cleared:
         if (ne00 % MATRIX_ROW_PADDING != 0 && ggml_is_quantized(src0->type) && ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE && src0->view_src == nullptr) {
-            const int64_t nbytes_data    = ggml_row_size(src0->type, (dev[id].row_high - dev[id].row_low)*ne00);
-            const int64_t nbytes_padding = ggml_row_size(src0->type, MATRIX_ROW_PADDING - ne00 % MATRIX_ROW_PADDING);
+            const size_t nbytes_data    = ggml_row_size(src0->type, (dev[id].row_high - dev[id].row_low)*ne00);
+            const size_t nbytes_padding = ggml_row_size(src0->type, MATRIX_ROW_PADDING - ne00 % MATRIX_ROW_PADDING);
             CUDA_CHECK(cudaMemsetAsync(dev[id].src0_dd + nbytes_data , 0, nbytes_padding, stream));
         }
 
@@ -3141,7 +3146,6 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_ROPE:
             return ggml_is_contiguous(op->src[0]);
         case GGML_OP_IM2COL:
-            return op->src[0]->type == GGML_TYPE_F16;
         case GGML_OP_POOL_2D:
         case GGML_OP_SUM:
         case GGML_OP_SUM_ROWS:

ggml/src/ggml-cuda/im2col.cu

@@ -91,9 +91,9 @@ void ggml_cuda_op_im2col(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const int64_t OH = is_2D ? dst->ne[2] : 1;
     const int64_t OW =         dst->ne[1];
 
-    const size_t delta_offset = src1->nb[is_2D ? 2 : 1] / 4; // nb is byte offset, src is type float32
-    const int64_t batch = src1->ne[3];
-    const size_t batch_offset = src1->nb[3] / 4; // nb is byte offset, src is type float32
+    const size_t  delta_offset = src1->nb[is_2D ? 2 : 1] / 4; // nb is byte offset, src is type float32
+    const int64_t batch        = src1->ne[is_2D ? 3 : 2];
+    const size_t  batch_offset = src1->nb[is_2D ? 3 : 2] / 4; // nb is byte offset, src is type float32
 
     if(dst->type == GGML_TYPE_F16) {
         im2col_cuda_f16(src1_d, (half *) dst_d, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, delta_offset, s0, s1, p0, p1, d0, d1, stream);

ggml/src/ggml-cuda/mmq.cu

@@ -8,8 +8,6 @@ void ggml_cuda_op_mul_mat_q(
 
     const int64_t ne00 = src0->ne[0];
 
-    const int64_t nb01 = src0->nb[1];
-
     const int64_t ne10 = src1->ne[0];
     const int64_t ne11 = src1->ne[1];
     GGML_ASSERT(ne10 % QK8_1 == 0);
@@ -17,7 +15,7 @@ void ggml_cuda_op_mul_mat_q(
     const int64_t ne0 = dst->ne[0];
 
     const int64_t row_diff = row_high - row_low;
-    const int64_t stride00 = nb01 / ggml_type_size(src0->type);
+    const int64_t stride00 = ne00 / ggml_blck_size(src0->type);
 
     int id = ggml_cuda_get_device();
     const int compute_capability = ggml_cuda_info().devices[id].cc;

ggml/src/ggml-metal.m

@@ -241,6 +241,8 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_ROPE_NEOX_F16,
     GGML_METAL_KERNEL_TYPE_IM2COL_F16,
     GGML_METAL_KERNEL_TYPE_IM2COL_F32,
+    GGML_METAL_KERNEL_TYPE_IM2COL_EXT_F16,
+    GGML_METAL_KERNEL_TYPE_IM2COL_EXT_F32,
     GGML_METAL_KERNEL_TYPE_UPSCALE_F32,
     GGML_METAL_KERNEL_TYPE_PAD_F32,
     GGML_METAL_KERNEL_TYPE_ARANGE_F32,
@@ -272,6 +274,8 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_SIN,
     GGML_METAL_KERNEL_TYPE_COS,
     GGML_METAL_KERNEL_TYPE_SUM_ROWS,
+    GGML_METAL_KERNEL_TYPE_POOL_2D_AVG_F32,
+    GGML_METAL_KERNEL_TYPE_POOL_2D_MAX_F32,
 
     GGML_METAL_KERNEL_TYPE_COUNT
 };
@@ -685,6 +689,8 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_NEOX_F16,                 rope_neox_f16,                  true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_F16,                    im2col_f16,                     true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_F32,                    im2col_f32,                     true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_EXT_F16,                im2col_ext_f16,                 true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_EXT_F32,                im2col_ext_f32,                 true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_UPSCALE_F32,                   upscale_f32,                    true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_PAD_F32,                       pad_f32,                        true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_TIMESTEP_EMBEDDING_F32,        timestep_embedding_f32,         true);
@@ -716,6 +722,8 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SIN,                           sin,                            true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_COS,                           cos,                            true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUM_ROWS,                      sum_rows,                       true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_POOL_2D_AVG_F32,               pool_2d_avg_f32,                true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_POOL_2D_MAX_F32,               pool_2d_max_f32,                true);
     }
 
     [metal_library release];
@@ -844,8 +852,8 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
         case GGML_OP_IM2COL:
             return op->src[0]->type == GGML_TYPE_F16;
         case GGML_OP_POOL_1D:
-        case GGML_OP_POOL_2D:
             return false;
+        case GGML_OP_POOL_2D:
         case GGML_OP_UPSCALE:
         case GGML_OP_PAD:
         case GGML_OP_ARANGE:
@@ -2545,6 +2553,8 @@ static void ggml_metal_encode_node(
             } break;
         case GGML_OP_IM2COL:
             {
+                GGML_ASSERT(ggml_is_contiguous(src0));
+                GGML_ASSERT(ggml_is_contiguous(src1));
                 GGML_ASSERT(src0->type == GGML_TYPE_F16);
                 GGML_ASSERT(src1->type == GGML_TYPE_F32);
                 GGML_ASSERT( dst->type == GGML_TYPE_F16 || dst->type == GGML_TYPE_F32);
@@ -2574,30 +2584,54 @@ static void ggml_metal_encode_node(
                 const int32_t ofs0 = src1->nb[is_2D ? 3 : 2] / 4;
                 const int32_t ofs1 = src1->nb[is_2D ? 2 : 1] / 4;
 
-                id<MTLComputePipelineState> pipeline = nil;
+                id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_IM2COL_F32].pipeline;
+
+                const bool is_gt_mttpt = ((size_t)(N * KH * KW)) > pipeline.maxTotalThreadsPerThreadgroup;
 
                 switch (dst->type) {
-                    case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_IM2COL_F32].pipeline; break;
-                    case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_IM2COL_F16].pipeline; break;
+                    case GGML_TYPE_F32: {
+                        pipeline = (is_gt_mttpt ?
+                                    ctx->kernels[GGML_METAL_KERNEL_TYPE_IM2COL_EXT_F32].pipeline
+                                    :
+                                    ctx->kernels[GGML_METAL_KERNEL_TYPE_IM2COL_F32].pipeline);
+                    } break;
+                    case GGML_TYPE_F16: {
+                        pipeline = (is_gt_mttpt ?
+                                    ctx->kernels[GGML_METAL_KERNEL_TYPE_IM2COL_EXT_F16].pipeline
+                                    :
+                                    ctx->kernels[GGML_METAL_KERNEL_TYPE_IM2COL_F16].pipeline);
+                    } break;
                     default: GGML_ABORT("fatal error");
                 };
 
                 [encoder setComputePipelineState:pipeline];
-                [encoder setBuffer:id_src1 offset:offs_src1        atIndex:0];
-                [encoder setBuffer:id_dst  offset:offs_dst         atIndex:1];
-                [encoder setBytes:&ofs0    length:sizeof( int32_t) atIndex:2];
-                [encoder setBytes:&ofs1    length:sizeof( int32_t) atIndex:3];
-                [encoder setBytes:&IW      length:sizeof( int32_t) atIndex:4];
-                [encoder setBytes:&IH      length:sizeof( int32_t) atIndex:5];
-                [encoder setBytes:&CHW     length:sizeof( int32_t) atIndex:6];
-                [encoder setBytes:&s0      length:sizeof( int32_t) atIndex:7];
-                [encoder setBytes:&s1      length:sizeof( int32_t) atIndex:8];
-                [encoder setBytes:&p0      length:sizeof( int32_t) atIndex:9];
-                [encoder setBytes:&p1      length:sizeof( int32_t) atIndex:10];
-                [encoder setBytes:&d0      length:sizeof( int32_t) atIndex:11];
-                [encoder setBytes:&d1      length:sizeof( int32_t) atIndex:12];
+                [encoder setBuffer:id_src1 offset:offs_src1       atIndex:0];
+                [encoder setBuffer:id_dst  offset:offs_dst        atIndex:1];
+                [encoder setBytes:&ofs0    length:sizeof(int32_t) atIndex:2];
+                [encoder setBytes:&ofs1    length:sizeof(int32_t) atIndex:3];
+                [encoder setBytes:&IW      length:sizeof(int32_t) atIndex:4];
+                [encoder setBytes:&IH      length:sizeof(int32_t) atIndex:5];
+                [encoder setBytes:&CHW     length:sizeof(int32_t) atIndex:6];
+                [encoder setBytes:&s0      length:sizeof(int32_t) atIndex:7];
+                [encoder setBytes:&s1      length:sizeof(int32_t) atIndex:8];
+                [encoder setBytes:&p0      length:sizeof(int32_t) atIndex:9];
+                [encoder setBytes:&p1      length:sizeof(int32_t) atIndex:10];
+                [encoder setBytes:&d0      length:sizeof(int32_t) atIndex:11];
+                [encoder setBytes:&d1      length:sizeof(int32_t) atIndex:12];
 
-                [encoder dispatchThreadgroups:MTLSizeMake(IC, OH, OW) threadsPerThreadgroup:MTLSizeMake(N, KH, KW)];
+                if (is_gt_mttpt) {
+                    [encoder setBytes:&N   length:sizeof(int32_t) atIndex:13];
+                    [encoder setBytes:&KH  length:sizeof(int32_t) atIndex:14];
+                    [encoder setBytes:&KW  length:sizeof(int32_t) atIndex:15];
+
+                    const uint64_t n_threads = MIN(pipeline.maxTotalThreadsPerThreadgroup, (uint64_t)N);
+
+                    const int64_t  quotient  = N / n_threads + (N % n_threads > 0 ? 1 : 0);
+
+                    [encoder dispatchThreadgroups:MTLSizeMake(quotient * CHW, OH, OW) threadsPerThreadgroup:MTLSizeMake(n_threads, 1, 1)];
+                } else {
+                    [encoder dispatchThreadgroups:MTLSizeMake(IC, OH, OW) threadsPerThreadgroup:MTLSizeMake(N, KH, KW)];
+                }
             } break;
         case GGML_OP_UPSCALE:
             {
@@ -3001,6 +3035,64 @@ static void ggml_metal_encode_node(
 
                 [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
             } break;
+        case GGML_OP_POOL_2D:
+            {
+                GGML_ASSERT(ggml_is_contiguous(src0));
+                GGML_ASSERT(src0t == GGML_TYPE_F32 && src0t == dstt);
+
+                const int32_t * opts = dst->op_params;
+                enum ggml_op_pool op = opts[0];
+
+                id<MTLComputePipelineState> pipeline = nil;
+                switch (src0t) {
+                    case GGML_TYPE_F32: {
+                        switch(op) {
+                            case GGML_OP_POOL_AVG:
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_POOL_2D_AVG_F32].pipeline; break;
+                            case GGML_OP_POOL_MAX:
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_POOL_2D_MAX_F32].pipeline; break;
+                            default: GGML_ASSERT(false && "not implemented");
+                        }
+                    } break;
+                    default: GGML_ASSERT(false && "not implemented");
+                }
+
+                const int32_t k0 = opts[1];
+                const int32_t k1 = opts[2];
+                const int32_t s0 = opts[3];
+                const int32_t s1 = opts[4];
+                const int32_t p0 = opts[5];
+                const int32_t p1 = opts[6];
+
+                const int64_t IH = src0->ne[1];
+                const int64_t IW = src0->ne[0];
+
+                const int64_t N  = dst->ne[3];
+                const int64_t OC = dst->ne[2];
+                const int64_t OH = dst->ne[1];
+                const int64_t OW = dst->ne[0];
+
+                const int64_t parallel_elements = N * OC * OH * OW;
+                const int64_t n_threads = MIN((int64_t)[pipeline maxTotalThreadsPerThreadgroup], parallel_elements);
+                const int64_t n_tg = (parallel_elements + n_threads - 1) / n_threads;
+
+                [encoder setComputePipelineState:pipeline];
+                [encoder setBuffer:id_src0 offset:offs_src0       atIndex:0];
+                [encoder setBuffer:id_dst  offset:offs_dst        atIndex:1];
+                [encoder setBytes:&k0      length:sizeof(int32_t) atIndex:2];
+                [encoder setBytes:&k1      length:sizeof(int32_t) atIndex:3];
+                [encoder setBytes:&s0      length:sizeof(int32_t) atIndex:4];
+                [encoder setBytes:&s1      length:sizeof(int32_t) atIndex:5];
+                [encoder setBytes:&p0      length:sizeof(int32_t) atIndex:6];
+                [encoder setBytes:&p1      length:sizeof(int32_t) atIndex:7];
+                [encoder setBytes:&IH      length:sizeof(int64_t) atIndex:8];
+                [encoder setBytes:&IW      length:sizeof(int64_t) atIndex:9];
+                [encoder setBytes:&OH      length:sizeof(int64_t) atIndex:10];
+                [encoder setBytes:&OW      length:sizeof(int64_t) atIndex:11];
+                [encoder setBytes:&parallel_elements length:sizeof(int64_t) atIndex:12];
+
+                [encoder dispatchThreadgroups:MTLSizeMake(n_tg, 1, 1) threadsPerThreadgroup:MTLSizeMake(n_threads, 1, 1)];
+            } break;
        default:
             {
                 GGML_LOG_ERROR("%s: error: node %3d, op = %8s not implemented\n", __func__, idx, ggml_op_name(dst->op));

ggml/src/ggml-metal.metal

@@ -1933,6 +1933,85 @@ kernel void kernel_im2col(
 template [[host_name("kernel_im2col_f32")]] kernel im2col_t kernel_im2col<float>;
 template [[host_name("kernel_im2col_f16")]] kernel im2col_t kernel_im2col<half>;
 
+typedef void (im2col_ext_t)(
+        device const float * x,
+        device        char * dst,
+        constant   int32_t & ofs0,
+        constant   int32_t & ofs1,
+        constant   int32_t & IW,
+        constant   int32_t & IH,
+        constant   int32_t & CHW,
+        constant   int32_t & s0,
+        constant   int32_t & s1,
+        constant   int32_t & p0,
+        constant   int32_t & p1,
+        constant   int32_t & d0,
+        constant   int32_t & d1,
+        constant   int32_t & N,
+        constant   int32_t & KH,
+        constant   int32_t & KW,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3  tgpg[[threadgroups_per_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]);
+
+template <typename T>
+kernel void kernel_im2col_ext(
+        device const float * x,
+        device        char * dst,
+        constant   int32_t & ofs0,
+        constant   int32_t & ofs1,
+        constant   int32_t & IW,
+        constant   int32_t & IH,
+        constant   int32_t & CHW,
+        constant   int32_t & s0,
+        constant   int32_t & s1,
+        constant   int32_t & p0,
+        constant   int32_t & p1,
+        constant   int32_t & d0,
+        constant   int32_t & d1,
+        constant   int32_t & N,
+        constant   int32_t & KH,
+        constant   int32_t & KW,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3  tgpg[[threadgroups_per_grid]],      // tgpg[0] = D x IC x KH x KW, CHW = IC x KH x KW
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {  // [M, 1, 1]
+    const int32_t KHW = KH * KW;             // KHW == ntg[1] * ntg[2], KW == ntg[2]
+
+    const int32_t d = tgpig[0] / CHW;
+    const int32_t chw = tgpig[0] % CHW;
+    const int32_t tgpig_0 = chw / KHW;  // 0 ~ (IC - 1)
+    const int32_t HW = tgpig[0] % KHW;
+
+    const int32_t tpitg_0 = (d * ntg[0]) + tpitg[0];
+    if (tpitg_0 >= N) {
+        return;
+    }
+
+    const int32_t tpitg_1 = HW / KW;
+    const int32_t tpitg_2 = HW % KW;
+
+    const int32_t iiw = tgpig[2] * s0 + tpitg_2 * d0 - p0;
+    const int32_t iih = tgpig[1] * s1 + tpitg_1 * d1 - p1;
+
+    const int32_t offset_dst =
+        (tpitg_0 * tgpg[1] * tgpg[2] + tgpig[1] * tgpg[2] + tgpig[2]) * CHW +
+        (tgpig_0 * KHW + tpitg_1 * KW + tpitg_2);
+
+    device T * pdst = (device T *) (dst);
+
+    if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) {
+        pdst[offset_dst] = 0.0f;
+    } else {
+        const int32_t offset_src = tpitg_0 * ofs0 + tgpig_0 * ofs1;
+        pdst[offset_dst] = x[offset_src + iih * IW + iiw];
+    }
+}
+
+template [[host_name("kernel_im2col_ext_f32")]] kernel im2col_ext_t kernel_im2col_ext<float>;
+template [[host_name("kernel_im2col_ext_f16")]] kernel im2col_ext_t kernel_im2col_ext<half>;
+
 kernel void kernel_upscale_f32(
     device  const char * src0,
     device        char * dst,
@@ -6372,3 +6451,102 @@ template [[host_name("kernel_mul_mv_id_iq3_s_f32")]]   kernel kernel_mul_mv_id_t
 template [[host_name("kernel_mul_mv_id_iq2_s_f32")]]   kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq2_s_f32_impl>>;
 template [[host_name("kernel_mul_mv_id_iq4_nl_f32")]]  kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq4_nl_f32_impl>>;
 template [[host_name("kernel_mul_mv_id_iq4_xs_f32")]]  kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq4_xs_f32_impl>>;
+
+kernel void kernel_pool_2d_max_f32(
+        device  const float * src0,
+        device        float * dst,
+        constant    int32_t & k0,
+        constant    int32_t & k1,
+        constant    int32_t & s0,
+        constant    int32_t & s1,
+        constant    int32_t & p0,
+        constant    int32_t & p1,
+        constant    int64_t & IH,
+        constant    int64_t & IW,
+        constant    int64_t & OH,
+        constant    int64_t & OW,
+        constant    int64_t & parallel_elements,
+        uint        gid[[thread_position_in_grid]]) {
+
+    if (gid >= parallel_elements) {
+        return;
+    }
+
+    const int idx = gid;
+    const int I_HW = IH * IW;
+    const int O_HW = OH * OW;
+    const int nc = idx / O_HW;
+    const int cur_oh = idx % O_HW / OW;
+    const int cur_ow = idx % O_HW % OW;
+
+    device const float * i_ptr = src0 + nc * I_HW;
+    device       float * o_ptr = dst  + nc * O_HW;
+
+    const int start_h = cur_oh * s1 - p1;
+    const int bh = MAX(0,  start_h);
+    const int eh = MIN(IH, start_h + k1);
+    const int start_w = cur_ow * s0 - p0;
+    const int bw = MAX(0,  start_w);
+    const int ew = MIN(IW, start_w + k0);
+
+    float res = -INFINITY;
+
+    for (int i = bh; i < eh; i += 1) {
+        for (int j = bw; j < ew; j += 1) {
+            res = MAX(res, i_ptr[i * IW + j]);
+        }
+    }
+
+    o_ptr[cur_oh * OW + cur_ow] = res;
+}
+
+kernel void kernel_pool_2d_avg_f32(
+        device  const float * src0,
+        device        float * dst,
+        constant    int32_t & k0,
+        constant    int32_t & k1,
+        constant    int32_t & s0,
+        constant    int32_t & s1,
+        constant    int32_t & p0,
+        constant    int32_t & p1,
+        constant    int64_t & IH,
+        constant    int64_t & IW,
+        constant    int64_t & OH,
+        constant    int64_t & OW,
+        constant    int64_t & parallel_elements,
+        uint        gid[[thread_position_in_grid]]) {
+
+    if (gid >= parallel_elements) {
+        return;
+    }
+
+    const int idx = gid;
+    const int I_HW = IH * IW;
+    const int O_HW = OH * OW;
+    const int nc = idx / O_HW;
+    const int cur_oh = idx % O_HW / OW;
+    const int cur_ow = idx % O_HW % OW;
+
+    device const float * i_ptr = src0 + nc * I_HW;
+    device       float * o_ptr = dst  + nc * O_HW;
+
+    const int start_h = cur_oh * s1 - p1;
+    const int bh = MAX(0,  start_h);
+    const int eh = MIN(IH, start_h + k1);
+    const int start_w = cur_ow * s0 - p0;
+    const int bw = MAX(0,  start_w);
+    const int ew = MIN(IW, start_w + k0);
+    // const float scale = 1. / ((eh - bh) * (ew - bw));
+    const float scale = 1. / (k0 * k1);
+
+    float res = 0;
+
+    for (int i = bh; i < eh; i += 1) {
+        for (int j = bw; j < ew; j += 1) {
+            float cur = i_ptr[i * IW + j];
+            res += cur * scale;
+        }
+    }
+
+    o_ptr[cur_oh * OW + cur_ow] = res;
+}

ggml/src/ggml-rpc.cpp

@@ -57,8 +57,9 @@ struct socket_t {
     }
 };
 
+// all RPC structures must be packed
+#pragma pack(push, 1)
 // ggml_tensor is serialized into rpc_tensor
-#pragma pack(1)
 struct rpc_tensor {
     uint64_t id;
     uint32_t type;
@@ -95,76 +96,64 @@ enum rpc_cmd {
     RPC_CMD_COUNT,
 };
 
-#pragma pack(1)
 struct rpc_msg_alloc_buffer_req {
     uint64_t size;
 };
 
-#pragma pack(1)
 struct rpc_msg_alloc_buffer_rsp {
     uint64_t remote_ptr;
     uint64_t remote_size;
 };
 
-#pragma pack(1)
 struct rpc_msg_get_alignment_rsp {
     uint64_t alignment;
 };
 
-#pragma pack(1)
 struct rpc_msg_get_max_size_rsp {
     uint64_t max_size;
 };
 
-#pragma pack(1)
 struct rpc_msg_buffer_get_base_req {
     uint64_t remote_ptr;
 };
 
-#pragma pack(1)
 struct rpc_msg_buffer_get_base_rsp {
     uint64_t base_ptr;
 };
 
-#pragma pack(1)
 struct rpc_msg_free_buffer_req {
     uint64_t remote_ptr;
 };
 
-#pragma pack(1)
 struct rpc_msg_buffer_clear_req {
     uint64_t remote_ptr;
     uint8_t value;
 };
 
-#pragma pack(1)
 struct rpc_msg_get_tensor_req {
     rpc_tensor tensor;
     uint64_t offset;
     uint64_t size;
 };
 
-#pragma pack(1)
 struct rpc_msg_copy_tensor_req {
     rpc_tensor src;
     rpc_tensor dst;
 };
 
-#pragma pack(1)
 struct rpc_msg_copy_tensor_rsp {
     uint8_t result;
 };
 
-#pragma pack(1)
 struct rpc_msg_graph_compute_rsp {
     uint8_t result;
 };
 
-#pragma pack(1)
 struct rpc_msg_get_device_memory_rsp {
     uint64_t free_mem;
     uint64_t total_mem;
 };
+#pragma pack(pop)
 
 // RPC data structures
 

ggml/src/ggml.c

@@ -324,8 +324,9 @@ struct ggml_logger_state {
 static struct ggml_logger_state g_logger_state = {ggml_log_callback_default, NULL};
 
 static void ggml_log_internal_v(enum ggml_log_level level, const char * format, va_list args) {
-    if (format == NULL)
+    if (format == NULL) {
         return;
+    }
     va_list args_copy;
     va_copy(args_copy, args);
     char buffer[128];
@@ -3463,7 +3464,7 @@ int64_t ggml_nrows(const struct ggml_tensor * tensor) {
 
 size_t ggml_nbytes(const struct ggml_tensor * tensor) {
     size_t nbytes;
-    size_t blck_size = ggml_blck_size(tensor->type);
+    const size_t blck_size = ggml_blck_size(tensor->type);
     if (blck_size == 1) {
         nbytes = ggml_type_size(tensor->type);
         for (int i = 0; i < GGML_MAX_DIMS; ++i) {
@@ -3851,10 +3852,6 @@ struct ggml_context * ggml_init(struct ggml_init_params params) {
                 },
             };
 
-            for (int i = 0; i < GGML_MAX_CONTEXTS; ++i) {
-                g_state.contexts[i].used = false;
-            }
-
             const uint64_t t_end = ggml_time_us(); UNUSED(t_end);
 
             GGML_PRINT_DEBUG("%s: g_state initialized in %f ms\n", __func__, (t_end - t_start)/1000.0f);
@@ -15723,6 +15720,9 @@ static void ggml_compute_forward_flash_attn_ext_f16(
     ggml_vec_dot_t    const kq_vec_dot     = type_traits[k->type].vec_dot;
     ggml_to_float_t   const v_to_float     = type_traits[v->type].to_float;
 
+    GGML_ASSERT(q_to_vec_dot && "fattn: unsupported K-type");
+    GGML_ASSERT(v_to_float   && "fattn: unsupported V-type");
+
     // loop over n_batch and n_head
     for (int ir = ir0; ir < ir1; ++ir) {
         // q indices

include/llama.h

@@ -217,6 +217,7 @@ extern "C" {
 
     typedef struct llama_token_data_array {
         // TODO: consider SoA
+        // NOTE: this pointer can be modified by the samplers
         llama_token_data * data;
         size_t size;
         int64_t selected; // this is the index in the data array (i.e. not the token id)
@@ -1069,12 +1070,13 @@ extern "C" {
 
     // available samplers:
 
-    LLAMA_API struct llama_sampler * llama_sampler_init_greedy     (void);
-    LLAMA_API struct llama_sampler * llama_sampler_init_dist       (uint32_t seed);
+    LLAMA_API struct llama_sampler * llama_sampler_init_greedy(void);
+    LLAMA_API struct llama_sampler * llama_sampler_init_dist  (uint32_t seed);
 
     /// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits.
     /// NOTE: Avoid using on the full vocabulary as the sorting can become slow. For example, apply top-k or top-p sampling first.
-    LLAMA_API struct llama_sampler * llama_sampler_init_softmax    (void);
+    DEPRECATED(LLAMA_API struct llama_sampler * llama_sampler_init_softmax    (void),
+        "will be removed in the future (see https://github.com/ggerganov/llama.cpp/pull/9896#discussion_r1800920915)");
 
     /// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
     LLAMA_API struct llama_sampler * llama_sampler_init_top_k      (int32_t k);
@@ -1090,6 +1092,8 @@ extern "C" {
 
     /// @details Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666.
     LLAMA_API struct llama_sampler * llama_sampler_init_typical    (float   p, size_t min_keep);
+
+    /// #details Updates the logits l_i` = l_i/t. When t <= 0.0f, the maximum logit is kept at it's original value, the rest are set to -inf
     LLAMA_API struct llama_sampler * llama_sampler_init_temp       (float   t);
 
     /// @details Dynamic temperature implementation (a.k.a. entropy) described in the paper https://arxiv.org/abs/2309.02772.

scripts/sync-ggml.last

@@ -1 +1 @@
-2327bda7a55ac6b72614ac5ebd5c5a5e02553b9b
+6dccc647264f5429df2624f36138f601e7ce23e5

src/llama-sampling.cpp

@@ -63,6 +63,30 @@ static void llama_log_softmax(float * array, size_t size) {
 }
 */
 
+static void llama_sampler_temp_impl(llama_token_data_array * cur_p, float temp) {
+    if (temp <= 0.0f) {
+        // find the token with the highest logit and set the rest to -inf
+        size_t max_i = 0;
+        float  max_l = cur_p->data[0].logit;
+
+        for (size_t i = 1; i < cur_p->size; ++i) {
+            if (cur_p->data[i    ].logit > max_l) {
+                cur_p->data[max_i].logit = -INFINITY;
+                max_i = i;
+                max_l = cur_p->data[i].logit;
+            } else {
+                cur_p->data[i].logit = -INFINITY;
+            }
+        }
+
+        return;
+    }
+
+    for (size_t i = 0; i < cur_p->size; ++i) {
+        cur_p->data[i].logit /= temp;
+    }
+}
+
 static void llama_sampler_softmax_impl(llama_token_data_array * cur_p) {
     GGML_ASSERT(cur_p->size > 0);
 
@@ -427,6 +451,9 @@ static const char * llama_sampler_dist_name(const struct llama_sampler * /*smpl*
 
 static void llama_sampler_dist_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
     auto * ctx = (llama_sampler_dist *) smpl->ctx;
+
+    llama_sampler_softmax_impl(cur_p);
+
     cur_p->selected = llama_sample_dist(cur_p, ctx->rng);
 }
 
@@ -912,9 +939,8 @@ static const char * llama_sampler_temp_name(const struct llama_sampler * /*smpl*
 
 static void llama_sampler_temp_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
     const auto * ctx = (llama_sampler_temp *) smpl->ctx;
-    for (size_t i = 0; i < cur_p->size; ++i) {
-        cur_p->data[i].logit /= ctx->temp;
-    }
+
+    llama_sampler_temp_impl(cur_p, ctx->temp);
 }
 
 static struct llama_sampler * llama_sampler_temp_clone(const struct llama_sampler * smpl) {
@@ -961,6 +987,7 @@ static void llama_sampler_temp_ext_apply(struct llama_sampler * smpl, llama_toke
     if (ctx->delta > 0) {
         const float min_temp = std::max(0.0f, ctx->temp - ctx->delta);
         const float max_temp = ctx->temp + ctx->delta;
+
         float exponent_val = ctx->exponent;
 
         // no need to do anything if there is only one (or zero) candidates
@@ -998,9 +1025,7 @@ static void llama_sampler_temp_ext_apply(struct llama_sampler * smpl, llama_toke
     #endif
 
         // Apply the dynamically calculated temperature scaling
-        for (size_t i = 0; i < cur_p->size; ++i) {
-            cur_p->data[i].logit /= dyn_temp;
-        }
+        llama_sampler_temp_impl(cur_p, dyn_temp);
 
         // Re-compute softmax probabilities after scaling logits with dynamic temperature
         const double max_l_double = cur_p->data[0].logit;
@@ -1024,9 +1049,7 @@ static void llama_sampler_temp_ext_apply(struct llama_sampler * smpl, llama_toke
         }
     #endif
     } else {
-        for (size_t i = 0; i < cur_p->size; ++i) {
-            cur_p->data[i].logit /= ctx->temp;
-        }
+        llama_sampler_temp_impl(cur_p, ctx->temp);
     }
 }
 

src/llama.cpp

@@ -10,8 +10,6 @@
 
 #if defined(GGML_USE_KOMPUTE)
 #   include "ggml-kompute.h"
-#elif defined(GGML_USE_CANN)
-#   include "ggml-cann.h"
 #endif
 
 #ifndef __AMX_INT8__
@@ -3399,10 +3397,6 @@ static int llama_get_device_count(const llama_model & model) {
     count += (int) model.rpc_servers.size();
 #endif
 
-#if defined(GGML_USE_CANN)
-    count += ggml_backend_cann_get_device_count();
-#endif
-
     return count;
 
     GGML_UNUSED(model);
@@ -3420,11 +3414,7 @@ static ggml_backend_buffer_type_t llama_default_buffer_type_cpu(const llama_mode
         }
     }
 
-#if defined(GGML_USE_CANN)
-    if (host_buffer) {
-        buft = ggml_backend_cann_host_buffer_type();
-    }
-#elif defined(GGML_USE_CPU_HBM)
+#if defined(GGML_USE_CPU_HBM)
     buft = ggml_backend_cpu_hbm_buffer_type();
 #endif
 
@@ -3446,8 +3436,6 @@ static ggml_backend_buffer_type_t llama_default_buffer_type_offload(const llama_
 
 #if defined(GGML_USE_KOMPUTE)
     buft = ggml_backend_kompute_buffer_type(device);
-#elif defined(GGML_USE_CANN)
-    buft = ggml_backend_cann_buffer_type(device);
 #endif
 
     if (buft == nullptr) {
@@ -3491,14 +3479,13 @@ static size_t llama_get_device_memory(const llama_model & model, int device) {
         return free;
     }
 
-#if defined(GGML_USE_CANN)
-    size_t total;
-    size_t free;
-    ggml_backend_cann_get_device_memory(device, &free, &total);
-    return free;
-#else
+    if (model.devices.size() > 0) {
+        ggml_backend_reg_t reg = ggml_backend_dev_backend_reg(model.devices[0]);
+        LLAMA_LOG_WARN("%s: failed to get free memmory of device:%d of backend:%s, for device id is out of range.\n", __func__, device, ggml_backend_reg_name(reg));
+    } else {
+        LLAMA_LOG_WARN("%s: failed to get free memmory of device, no devices in inputted model.\n", __func__);
+    }
     return 1;
-#endif
 
     GGML_UNUSED(model);
     GGML_UNUSED(device);
@@ -5190,6 +5177,57 @@ struct llama_model_loader {
     }
 };
 
+// temporary allocate memory for the input batch if needed
+static const llama_seq_id batch_default_seq_id = 0;
+struct llama_batch_allocr {
+    std::array<llama_seq_id, 1> seq_id_0 = {batch_default_seq_id};
+    std::vector<llama_pos>      pos;
+    std::vector<int32_t>        n_seq_id;
+    std::vector<llama_seq_id *> seq_id;
+    std::vector<int8_t>         logits;
+    struct llama_batch          batch;
+    // optionally fulfill the batch returned by llama_batch_get_one
+    llama_batch_allocr(llama_context & ctx, struct llama_batch in_batch) {
+        batch = in_batch;
+        GGML_ASSERT(batch.n_tokens > 0);
+        if (!batch.pos) {
+            // determine the last position in KV cache
+            llama_pos last_pos = -1;
+            for (const auto & cell : ctx.kv_self.cells) {
+                if (cell.has_seq_id(batch_default_seq_id)) {
+                    last_pos = std::max(last_pos, cell.pos);
+                }
+            }
+            last_pos++; // next position
+            pos.resize(batch.n_tokens);
+            for (int32_t i = 0; i < batch.n_tokens; i++) {
+                pos[i] = i+last_pos;
+            }
+            batch.pos = pos.data();
+        }
+        if (!batch.n_seq_id) {
+            n_seq_id.resize(batch.n_tokens);
+            for (int32_t i = 0; i < batch.n_tokens; i++) {
+                n_seq_id[i] = seq_id_0.size();
+            }
+            batch.n_seq_id = n_seq_id.data();
+        }
+        if (!batch.seq_id) {
+            seq_id.resize(batch.n_tokens + 1);
+            seq_id[batch.n_tokens] = NULL;
+            for (int32_t i = 0; i < batch.n_tokens; i++) {
+                seq_id[i] = seq_id_0.data();
+            }
+            batch.seq_id = seq_id.data();
+        }
+        if (!batch.logits) {
+            logits.resize(batch.n_tokens);
+            logits[logits.size() - 1] = true;
+            batch.logits = logits.data();
+        }
+    }
+};
+
 template<>
 bool llama_model_loader::get_key(const enum llm_kv kid, enum llama_pooling_type & result, const bool required) {
     uint32_t tmp;
@@ -10030,7 +10068,7 @@ struct llm_build_context {
           llama_context  & lctx;
     const llama_hparams  & hparams;
     const llama_cparams  & cparams;
-    const llama_ubatch   & batch;
+    const llama_ubatch   & ubatch;
     const llama_kv_cache & kv_self;
 
     const int64_t n_embd;
@@ -10076,14 +10114,14 @@ struct llm_build_context {
     // TODO: consider making the entire interface noexcept
     llm_build_context(
         llama_context  & lctx,
-    const llama_ubatch & batch,
+    const llama_ubatch & ubatch,
     const llm_build_cb & cb,
                   bool   worst_case) :
         model            (lctx.model),
         lctx             (lctx),
         hparams          (model.hparams),
         cparams          (lctx.cparams),
-        batch            (batch),
+        ubatch           (ubatch),
         kv_self          (lctx.kv_self),
         n_embd           (hparams.n_embd),
         n_layer          (hparams.n_layer),
@@ -10105,7 +10143,7 @@ struct llm_build_context {
         beta_slow        (cparams.yarn_beta_slow),
         norm_eps         (hparams.f_norm_eps),
         norm_rms_eps     (hparams.f_norm_rms_eps),
-        n_tokens         (batch.n_tokens),
+        n_tokens         (ubatch.n_tokens),
         n_kv             (worst_case ? kv_self.size : kv_self.n),
         n_outputs        (worst_case ? n_tokens : lctx.n_outputs),
         n_outputs_enc    (worst_case ? n_tokens : lctx.embd_enc.size() / hparams.n_embd),
@@ -10474,7 +10512,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -10634,7 +10672,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = model.type == MODEL_7B ? build_inp_pos() : nullptr;
@@ -10749,7 +10787,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -10853,7 +10891,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -10975,7 +11013,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // multiply by embedding_multiplier_scale of 78.38367176906169
         inpL = ggml_scale(ctx0, inpL, 78.38367176906169f);
@@ -11133,7 +11171,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -11255,7 +11293,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -11358,7 +11396,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
         struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
@@ -11460,7 +11498,7 @@ struct llm_build_context {
         }
 
         // construct input embeddings (token, type, position)
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // token types are hardcoded to zero ("Sentence A")
         struct ggml_tensor * type_row0 = ggml_view_1d(ctx0, model.type_embd, n_embd, 0);
@@ -11647,7 +11685,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
         struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
@@ -11749,7 +11787,7 @@ struct llm_build_context {
         struct ggml_tensor * pos;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
         struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
@@ -11887,7 +11925,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -12037,7 +12075,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -12150,7 +12188,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -12265,7 +12303,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -12410,7 +12448,7 @@ struct llm_build_context {
         struct ggml_tensor * ffn_output;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -12529,7 +12567,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -12657,7 +12695,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -12762,7 +12800,7 @@ struct llm_build_context {
         struct ggml_tensor * pos;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -12867,7 +12905,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -12977,7 +13015,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -13095,7 +13133,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -13222,7 +13260,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // scale the input embeddings
         inpL = ggml_scale(ctx0, inpL, scale_embd);
@@ -13366,7 +13404,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // scale the input embeddings
         inpL = ggml_scale(ctx0, inpL, scale_embd);
@@ -13567,7 +13605,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
         cb(inpL, "inp_scaled", -1);
@@ -13675,7 +13713,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
         cb(inpL, "inp_scaled", -1);
@@ -13813,7 +13851,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -13929,7 +13967,7 @@ struct llm_build_context {
         struct ggml_tensor * inpL;
 
         // {n_embd, n_tokens}
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         struct ggml_tensor * state_copy = build_inp_s_copy();
         struct ggml_tensor * state_mask = build_inp_s_mask();
@@ -13941,7 +13979,7 @@ struct llm_build_context {
                     LLM_NORM_RMS, cb, il);
             cb(cur, "attn_norm", il);
 
-            cur = llm_build_mamba(ctx0, lctx, batch, gf, cur,
+            cur = llm_build_mamba(ctx0, lctx, ubatch, gf, cur,
                     state_copy, state_mask,
                     kv_head, n_kv, cb, il);
 
@@ -13987,7 +14025,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -14144,7 +14182,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -14272,7 +14310,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -14391,7 +14429,7 @@ struct llm_build_context {
 
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -14518,7 +14556,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -14663,7 +14701,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -14804,7 +14842,7 @@ struct llm_build_context {
         struct ggml_tensor * inpL;
 
         // {n_embd, n_tokens}
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -15019,7 +15057,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -15173,7 +15211,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         GGML_ASSERT(lctx.is_encoding);
         struct ggml_tensor * pos_bucket_enc = llm_build_pos_bucket(false);
@@ -15305,7 +15343,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         GGML_ASSERT(!lctx.is_encoding);
         GGML_ASSERT(n_outputs_enc > 0 && "call llama_encode() first");
@@ -15507,7 +15545,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
         struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
@@ -15599,7 +15637,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -15713,7 +15751,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -15837,7 +15875,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -15957,11 +15995,11 @@ struct llm_build_context {
         // Token shift state dimensions should be 2 * n_emb
         GGML_ASSERT(n_embd == hparams.n_embd_k_s() / 2);
 
-        const int64_t n_seqs = batch.n_seqs;
-        const int64_t n_seq_tokens = batch.n_seq_tokens;
-        const int64_t n_tokens = batch.n_tokens;
+        const int64_t n_seqs = ubatch.n_seqs;
+        const int64_t n_seq_tokens = ubatch.n_seq_tokens;
+        const int64_t n_tokens = ubatch.n_tokens;
         GGML_ASSERT(n_seqs != 0);
-        GGML_ASSERT(batch.equal_seqs);
+        GGML_ASSERT(ubatch.equal_seqs);
         GGML_ASSERT(n_tokens == n_seq_tokens * n_seqs);
 
         struct ggml_tensor * cur;
@@ -15969,7 +16007,7 @@ struct llm_build_context {
         struct ggml_tensor * state_copy = build_inp_s_copy();
         struct ggml_tensor * state_mask = build_inp_s_mask();
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
         inpL = llm_build_norm(ctx0, inpL, hparams, model.tok_norm, model.tok_norm_b, LLM_NORM, cb, -1);
 
         for (int il = 0; il < n_layer; ++il) {
@@ -16083,7 +16121,7 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
         struct ggml_tensor * inp_pos = build_inp_pos();
@@ -16279,7 +16317,7 @@ static struct ggml_cgraph * llama_build_graph_k_shift(llama_context & lctx) {
 
 static struct ggml_cgraph * llama_build_graph(
          llama_context & lctx,
-    const llama_ubatch & batch,
+    const llama_ubatch & ubatch,
                   bool   worst_case) {
     const auto & model = lctx.model;
 
@@ -16301,7 +16339,7 @@ static struct ggml_cgraph * llama_build_graph(
         // norm may be automatically assigned to the backend of the previous layer, increasing data transfer between backends
         // FIXME: fix in ggml_backend_sched
         const bool full_offload = lctx.model.n_gpu_layers > (int)lctx.model.hparams.n_layer;
-        if (batch.n_tokens < 32 || full_offload) {
+        if (ubatch.n_tokens < 32 || full_offload) {
             if (il != -1 && strcmp(name, "norm") == 0) {
                 for (auto * backend : lctx.backends) {
                     if (ggml_backend_supports_buft(backend, lctx.model.buft_layer[il].buft) &&
@@ -16316,7 +16354,7 @@ static struct ggml_cgraph * llama_build_graph(
 
     struct ggml_cgraph * result = NULL;
 
-    struct llm_build_context llm(lctx, batch, cb, worst_case);
+    struct llm_build_context llm(lctx, ubatch, cb, worst_case);
 
     llm.init();
 
@@ -16567,7 +16605,7 @@ static int32_t llama_relative_position_bucket(llama_pos x, llama_pos y, uint64_t
     return relative_bucket;
 }
 
-static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
+static void llama_set_inputs(llama_context & lctx, const llama_ubatch & ubatch) {
     //
     // set input data
     //
@@ -16576,28 +16614,28 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
     const auto & cparams = lctx.cparams;
     const auto & kv_self = lctx.kv_self;
 
-    if (batch.token) {
-        const int64_t n_tokens = batch.n_tokens;
+    if (ubatch.token) {
+        const int64_t n_tokens = ubatch.n_tokens;
 
-        ggml_backend_tensor_set(lctx.inp_tokens, batch.token, 0, n_tokens*ggml_element_size(lctx.inp_tokens));
+        ggml_backend_tensor_set(lctx.inp_tokens, ubatch.token, 0, n_tokens*ggml_element_size(lctx.inp_tokens));
     }
 
-    if (batch.embd) {
+    if (ubatch.embd) {
         const int64_t n_embd   = hparams.n_embd;
-        const int64_t n_tokens = batch.n_tokens;
+        const int64_t n_tokens = ubatch.n_tokens;
 
-        ggml_backend_tensor_set(lctx.inp_embd, batch.embd, 0, n_tokens*n_embd*ggml_element_size(lctx.inp_embd));
+        ggml_backend_tensor_set(lctx.inp_embd, ubatch.embd, 0, n_tokens*n_embd*ggml_element_size(lctx.inp_embd));
     }
 
-    if (batch.pos && lctx.inp_pos) {
-        const int64_t n_tokens = batch.n_tokens;
+    if (ubatch.pos && lctx.inp_pos) {
+        const int64_t n_tokens = ubatch.n_tokens;
 
-        ggml_backend_tensor_set(lctx.inp_pos, batch.pos, 0, n_tokens*ggml_element_size(lctx.inp_pos));
+        ggml_backend_tensor_set(lctx.inp_pos, ubatch.pos, 0, n_tokens*ggml_element_size(lctx.inp_pos));
     }
 
     if (hparams.causal_attn || cparams.pooling_type == LLAMA_POOLING_TYPE_NONE) {
         GGML_ASSERT(lctx.inp_out_ids && "every model that can must skip unused outputs");
-        const int64_t n_tokens = batch.n_tokens;
+        const int64_t n_tokens = ubatch.n_tokens;
 
         GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_out_ids->buffer));
         int32_t * data = (int32_t *) lctx.inp_out_ids->data;
@@ -16606,10 +16644,10 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
             for (int i = 0; i < n_tokens; ++i) {
                 data[i] = i;
             }
-        } else if (batch.output) {
+        } else if (ubatch.output) {
             int32_t n_outputs = 0;
             for (int i = 0; i < n_tokens; ++i) {
-                if (batch.output[i]) {
+                if (ubatch.output[i]) {
                     data[n_outputs++] = i;
                 }
             }
@@ -16634,9 +16672,9 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
         // NOTE: hparams.causal_attn indicates the model is capable of generation and uses the kv cache.
         if (cparams.causal_attn && !lctx.is_encoding) {
             const int64_t n_kv         = kv_self.n;
-            const int64_t n_tokens     = batch.n_tokens;
-            const int64_t n_seq_tokens = batch.n_seq_tokens;
-            const int64_t n_seqs       = batch.n_seqs;
+            const int64_t n_tokens     = ubatch.n_tokens;
+            const int64_t n_seq_tokens = ubatch.n_seq_tokens;
+            const int64_t n_seqs       = ubatch.n_seqs;
 
 
             float * data     = nullptr;
@@ -16653,14 +16691,14 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
             }
 
             // For causal attention, use only the previous KV cells
-            // of the correct sequence for each token of the batch.
+            // of the correct sequence for each token of the ubatch.
             // It's assumed that if a token in the batch has multiple sequences, they are equivalent.
             for (int h = 0; h < 1; ++h) {
                 for (int s = 0; s < n_seqs; ++s) {
-                    const llama_seq_id seq_id = batch.seq_id[s][0];
+                    const llama_seq_id seq_id = ubatch.seq_id[s][0];
 
                     for (int j = 0; j < n_seq_tokens; ++j) {
-                        const llama_pos pos = batch.pos[s*n_seq_tokens + j];
+                        const llama_pos pos = ubatch.pos[s*n_seq_tokens + j];
 
                         for (int i = 0; i < n_kv; ++i) {
                             float f;
@@ -16706,9 +16744,9 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
                 }
             }
         } else {
-            const int64_t n_tokens     = batch.n_tokens;
-            const int64_t n_seq_tokens = batch.n_seq_tokens;
-            const int64_t n_seqs       = batch.n_seqs;
+            const int64_t n_tokens     = ubatch.n_tokens;
+            const int64_t n_seq_tokens = ubatch.n_seq_tokens;
+            const int64_t n_seqs       = ubatch.n_seqs;
             // when using kv cache, the mask needs to match the kv cache size
             const int64_t n_stride = hparams.causal_attn && !lctx.is_encoding ? kv_self.n : n_tokens;
 
@@ -16718,7 +16756,7 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
 
             for (int h = 0; h < 1; ++h) {
                 for (int s1 = 0; s1 < n_seqs; ++s1) {
-                    const llama_seq_id seq_id = batch.seq_id[s1][0];
+                    const llama_seq_id seq_id = ubatch.seq_id[s1][0];
 
                     for (int j = 0; j < n_seq_tokens; ++j) {
                         const int32_t tj = s1*n_seq_tokens + j;
@@ -16728,10 +16766,10 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
                                 const int32_t ti = s0*n_seq_tokens + i;
                                 float f = -INFINITY;
 
-                                for (int s = 0; s < batch.n_seq_id[s0]; ++s) {
-                                    if (batch.seq_id[s0][s] == seq_id) {
+                                for (int s = 0; s < ubatch.n_seq_id[s0]; ++s) {
+                                    if (ubatch.seq_id[s0][s] == seq_id) {
                                         if (hparams.use_alibi) {
-                                            f = -std::abs(batch.pos[ti] - batch.pos[tj]);
+                                            f = -std::abs(ubatch.pos[ti] - ubatch.pos[tj]);
                                         } else {
                                             f = 0.0f;
                                         }
@@ -16753,9 +16791,9 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
     }
 
     if (cparams.embeddings && cparams.pooling_type == LLAMA_POOLING_TYPE_MEAN) {
-        const int64_t n_tokens     = batch.n_tokens;
-        const int64_t n_seq_tokens = batch.n_seq_tokens;
-        const int64_t n_seqs       = batch.n_seqs;
+        const int64_t n_tokens     = ubatch.n_tokens;
+        const int64_t n_seq_tokens = ubatch.n_seq_tokens;
+        const int64_t n_seqs       = ubatch.n_seqs;
 
         GGML_ASSERT(lctx.inp_mean);
         GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_mean->buffer));
@@ -16766,12 +16804,12 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
         std::vector<uint64_t> sum(n_tokens, 0);
 
         for (int s = 0; s < n_seqs; ++s) {
-            const llama_seq_id seq_id = batch.seq_id[s][0];
+            const llama_seq_id seq_id = ubatch.seq_id[s][0];
 
-            // TODO: adapt limits to n_seqs when batch.equal_seqs is true
+            // TODO: adapt limits to n_seqs when ubatch.equal_seqs is true
             GGML_ASSERT(seq_id < n_tokens && "seq_id cannot be larger than n_tokens with pooling_type == MEAN");
 
-            sum[seq_id] += batch.n_seq_tokens;
+            sum[seq_id] += ubatch.n_seq_tokens;
         }
 
         std::vector<float> div(n_tokens, 0.0f);
@@ -16783,7 +16821,7 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
         }
 
         for (int s = 0; s < n_seqs; ++s) {
-            const llama_seq_id seq_id = batch.seq_id[s][0];
+            const llama_seq_id seq_id = ubatch.seq_id[s][0];
 
             for (int i = 0; i < n_seq_tokens; ++i) {
                 data[seq_id*n_tokens + s*n_seq_tokens + i] = div[seq_id];
@@ -16794,9 +16832,9 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
     if (cparams.embeddings && (
                 cparams.pooling_type == LLAMA_POOLING_TYPE_CLS ||
                 cparams.pooling_type == LLAMA_POOLING_TYPE_RANK)) {
-        const int64_t n_tokens     = batch.n_tokens;
-        const int64_t n_seq_tokens = batch.n_seq_tokens;
-        const int64_t n_seqs       = batch.n_seqs;
+        const int64_t n_tokens     = ubatch.n_tokens;
+        const int64_t n_seq_tokens = ubatch.n_seq_tokens;
+        const int64_t n_seqs       = ubatch.n_seqs;
 
         GGML_ASSERT(lctx.inp_cls);
         GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_cls->buffer));
@@ -16805,13 +16843,13 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
         memset(lctx.inp_cls->data, 0, n_tokens * ggml_element_size(lctx.inp_cls));
 
         for (int s = 0; s < n_seqs; ++s) {
-            const llama_seq_id seq_id = batch.seq_id[s][0];
+            const llama_seq_id seq_id = ubatch.seq_id[s][0];
 
-            // TODO: adapt limits to n_seqs when batch.equal_seqs is true
+            // TODO: adapt limits to n_seqs when ubatch.equal_seqs is true
             GGML_ASSERT(seq_id < n_tokens && "seq_id cannot be larger than n_tokens with pooling_type == CLS or RANK");
 
             for (int i = 0; i < n_seq_tokens; ++i) {
-                const llama_pos pos = batch.pos[s*n_seq_tokens + i];
+                const llama_pos pos = ubatch.pos[s*n_seq_tokens + i];
 
                 if (pos == 0) {
                     data[seq_id] = s*n_seq_tokens + i;
@@ -16821,9 +16859,9 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
     }
 
     if (cparams.embeddings && cparams.pooling_type == LLAMA_POOLING_TYPE_LAST) {
-        const int64_t n_tokens     = batch.n_tokens;
-        const int64_t n_seq_tokens = batch.n_seq_tokens;
-        const int64_t n_seqs       = batch.n_seqs;
+        const int64_t n_tokens     = ubatch.n_tokens;
+        const int64_t n_seq_tokens = ubatch.n_seq_tokens;
+        const int64_t n_seqs       = ubatch.n_seqs;
 
         GGML_ASSERT(lctx.inp_cls);
         GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_cls->buffer));
@@ -16835,13 +16873,13 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
         std::vector<int> last_row(n_tokens, -1);
 
         for (int s = 0; s < n_seqs; ++s) {
-            const llama_seq_id seq_id = batch.seq_id[s][0];
+            const llama_seq_id seq_id = ubatch.seq_id[s][0];
 
-            // TODO: adapt limits to n_seqs when batch.equal_seqs is true
+            // TODO: adapt limits to n_seqs when ubatch.equal_seqs is true
             GGML_ASSERT(seq_id < n_tokens && "seq_id cannot be larger than n_tokens with pooling_type == LAST");
 
             for (int i = 0; i < n_seq_tokens; ++i) {
-                const llama_pos pos = batch.pos[s*n_seq_tokens + i];
+                const llama_pos pos = ubatch.pos[s*n_seq_tokens + i];
 
                 if (pos >= last_pos[seq_id]) {
                     last_pos[seq_id] = pos;
@@ -16903,10 +16941,10 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
     }
 
     if (lctx.inp_pos_bucket) {
-        const int64_t n_tokens = batch.n_tokens;
+        const int64_t n_tokens = ubatch.n_tokens;
 
         GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_pos_bucket->buffer));
-        GGML_ASSERT(!batch.equal_seqs); // TODO: use batch.n_seqs instead of failing
+        GGML_ASSERT(!ubatch.equal_seqs); // TODO: use ubatch.n_seqs instead of failing
 
         int32_t * data = (int32_t *) lctx.inp_pos_bucket->data;
 
@@ -16915,7 +16953,7 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
             for (int h = 0; h < 1; ++h) {
                 for (int j = 0; j < n_tokens; ++j) {
                     for (int i = 0; i < n_kv; ++i) {
-                        data[h*(n_kv*n_tokens) + j*n_kv + i] = llama_relative_position_bucket(lctx.kv_self.cells[i].pos, batch.pos[j], hparams.n_rel_attn_bkts, lctx.is_encoding);
+                        data[h*(n_kv*n_tokens) + j*n_kv + i] = llama_relative_position_bucket(lctx.kv_self.cells[i].pos, ubatch.pos[j], hparams.n_rel_attn_bkts, lctx.is_encoding);
                     }
                 }
             }
@@ -16923,7 +16961,7 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
             for (int h = 0; h < 1; ++h) {
                 for (int j = 0; j < n_tokens; ++j) {
                     for (int i = 0; i < n_tokens; ++i) {
-                        data[h*(n_tokens*n_tokens) + j*n_tokens + i] = llama_relative_position_bucket(batch.pos[i], batch.pos[j], hparams.n_rel_attn_bkts, lctx.is_encoding);
+                        data[h*(n_tokens*n_tokens) + j*n_tokens + i] = llama_relative_position_bucket(ubatch.pos[i], ubatch.pos[j], hparams.n_rel_attn_bkts, lctx.is_encoding);
                     }
                 }
             }
@@ -16939,10 +16977,10 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
 
     if (!lctx.is_encoding && lctx.inp_KQ_mask_cross) {
         const int64_t n_output_enc = lctx.embd_enc.size() / hparams.n_embd;
-        const int64_t n_tokens = batch.n_tokens;
+        const int64_t n_tokens = ubatch.n_tokens;
 
         GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_KQ_mask_cross->buffer));
-        GGML_ASSERT(!batch.equal_seqs); // TODO: use batch.n_seqs instead of failing
+        GGML_ASSERT(!ubatch.equal_seqs); // TODO: use ubatch.n_seqs instead of failing
 
         float * data = (float *) lctx.inp_KQ_mask_cross->data;
 
@@ -16950,8 +16988,8 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
             for (int j = 0; j < n_tokens; ++j) {
                 for (int i = 0; i < n_output_enc; ++i) {
                     float f = -INFINITY;
-                    for (int s = 0; s < batch.n_seq_id[j]; ++s) {
-                        const llama_seq_id seq_id = batch.seq_id[j][s];
+                    for (int s = 0; s < ubatch.n_seq_id[j]; ++s) {
+                        const llama_seq_id seq_id = ubatch.seq_id[j][s];
                         if (lctx.seq_ids_enc[i].find(seq_id) != lctx.seq_ids_enc[i].end()) {
                             f = 0.0f;
                         }
@@ -17108,16 +17146,20 @@ static void llama_graph_compute(
 //
 static int llama_decode_internal(
          llama_context & lctx,
-           llama_batch   batch) {
+           llama_batch   inp_batch) {
 
     lctx.is_encoding = false;
-    const uint32_t n_tokens_all = batch.n_tokens;
 
-    if (n_tokens_all == 0) {
+    if (inp_batch.n_tokens == 0) {
         LLAMA_LOG_ERROR("%s: n_tokens == 0\n", __func__);
         return -1;
     }
 
+    // temporary allocate memory for the input batch if needed
+    llama_batch_allocr batch_allocr(lctx, inp_batch);
+    const llama_batch & batch = batch_allocr.batch;
+    const uint32_t n_tokens_all = batch.n_tokens;
+
     const auto & model   = lctx.model;
     const auto & hparams = model.hparams;
     const auto & cparams = lctx.cparams;
@@ -17422,17 +17464,20 @@ static int llama_decode_internal(
 //
 static int llama_encode_internal(
          llama_context & lctx,
-           llama_batch   batch) {
+           llama_batch   inp_batch) {
 
     lctx.is_encoding = true;
 
-    const uint32_t n_tokens = batch.n_tokens;
-
-    if (n_tokens == 0) {
+    if (inp_batch.n_tokens == 0) {
         LLAMA_LOG_ERROR("%s: n_tokens == 0\n", __func__);
         return -1;
     }
 
+    // temporary allocate memory for the input batch if needed
+    llama_batch_allocr batch_allocr(lctx, inp_batch);
+    const llama_batch & batch = batch_allocr.batch;
+    const uint32_t n_tokens = batch.n_tokens;
+
     const auto & model   = lctx.model;
     const auto & hparams = model.hparams;
     const auto & cparams = lctx.cparams;
@@ -19243,7 +19288,7 @@ struct llama_context * llama_new_context_with_model(
         params.flash_attn = false;
     }
 
-    if (params.type_v != GGML_TYPE_F16 && !params.flash_attn) {
+    if (ggml_is_quantized(params.type_v) && !params.flash_attn) {
         LLAMA_LOG_ERROR("%s: V cache quantization requires flash_attn\n", __func__);
         return nullptr;
     }
@@ -19396,30 +19441,6 @@ struct llama_context * llama_new_context_with_model(
             }
             ctx->backends.push_back(backend);
         }
-#elif defined(GGML_USE_CANN)
-        // with split_mode LLAMA_SPLIT_MODE_NONE or LLAMA_SPLIT_MODE_ROW, only the main GPU backend is used
-        // TODO: ggml_backend_cann is not support split tensor now, just leave code here.
-        if (model->split_mode == LLAMA_SPLIT_MODE_NONE || model->split_mode == LLAMA_SPLIT_MODE_ROW) {
-            ggml_backend_t backend = ggml_backend_cann_init(main_gpu);
-            if (backend == nullptr) {
-                LLAMA_LOG_ERROR("%s: failed to initialize CANN%d backend\n", __func__, main_gpu);
-                llama_free(ctx);
-                return nullptr;
-            }
-            ctx->backends.push_back(backend);
-        } else {
-            // LLAMA_SPLIT_MODE_LAYER requires a backend for each GPU
-            // TODO: currently, CANN can't use multi-gpus, just leave code here for further cann version.
-            for (int32_t device = 0; device < ggml_backend_cann_get_device_count(); ++device) {
-                ggml_backend_t backend = ggml_backend_cann_init(device);
-                if (backend == nullptr) {
-                    LLAMA_LOG_ERROR("%s: failed to initialize CANN%d backend\n", __func__, device);
-                    llama_free(ctx);
-                    return nullptr;
-                }
-                ctx->backends.push_back(backend);
-            }
-        }
 #endif
 
         // add other backends (such as BLAS)
@@ -21127,61 +21148,10 @@ void llama_batch_free(struct llama_batch batch) {
     if (batch.logits)   free(batch.logits);
 }
 
-// temporary allocate memory for the input batch if needed
-static const llama_seq_id batch_default_seq_id = 0;
-struct llama_batch_allocr {
-    std::array<llama_seq_id, 1> seq_id_0 = {batch_default_seq_id};
-    std::vector<llama_pos>      pos;
-    std::vector<int32_t>        n_seq_id;
-    std::vector<llama_seq_id *> seq_id;
-    std::vector<int8_t>         logits;
-    struct llama_batch          batch;
-    // optionally fulfill the batch returned by llama_batch_get_one
-    llama_batch_allocr(struct llama_context * ctx, struct llama_batch in_batch) {
-        batch = in_batch;
-        if (!batch.pos) {
-            // determine the last position in KV cache
-            llama_pos last_pos = -1;
-            for (const auto & cell : ctx->kv_self.cells) {
-                if (cell.has_seq_id(batch_default_seq_id)) {
-                    last_pos = std::max(last_pos, cell.pos);
-                }
-            }
-            last_pos++; // next position
-            pos.resize(batch.n_tokens);
-            for (int32_t i = 0; i < batch.n_tokens; i++) {
-                pos[i] = i+last_pos;
-            }
-            batch.pos = pos.data();
-        }
-        if (!batch.n_seq_id) {
-            n_seq_id.resize(batch.n_tokens);
-            for (int32_t i = 0; i < batch.n_tokens; i++) {
-                n_seq_id[i] = seq_id_0.size();
-            }
-            batch.n_seq_id = n_seq_id.data();
-        }
-        if (!batch.seq_id) {
-            seq_id.resize(batch.n_tokens + 1);
-            seq_id[batch.n_tokens] = NULL;
-            for (int32_t i = 0; i < batch.n_tokens; i++) {
-                seq_id[i] = seq_id_0.data();
-            }
-            batch.seq_id = seq_id.data();
-        }
-        if (!batch.logits) {
-            logits.resize(batch.n_tokens);
-            logits[logits.size() - 1] = true;
-            batch.logits = logits.data();
-        }
-    }
-};
-
 int32_t llama_encode(
         struct llama_context * ctx,
           struct llama_batch   batch) {
-    llama_batch_allocr batch_allocr(ctx, batch);
-    const int ret = llama_encode_internal(*ctx, batch_allocr.batch);
+    const int ret = llama_encode_internal(*ctx, batch);
     if (ret != 0) {
         LLAMA_LOG_ERROR("%s: failed to encode, ret = %d\n", __func__, ret);
     }
@@ -21192,8 +21162,7 @@ int32_t llama_encode(
 int32_t llama_decode(
         struct llama_context * ctx,
           struct llama_batch   batch) {
-    llama_batch_allocr batch_allocr(ctx, batch);
-    const int ret = llama_decode_internal(*ctx, batch_allocr.batch);
+    const int ret = llama_decode_internal(*ctx, batch);
     if (ret != 0) {
         LLAMA_LOG_ERROR("%s: failed to decode, ret = %d\n", __func__, ret);
     }
@@ -21734,6 +21703,16 @@ static int32_t llama_chat_apply_template_internal(
         if (add_ass) {
             ss << "[|assistant|]";
         }
+    } else if (tmpl == "rwkv-world" || tmpl_contains("rwkv-world")) {
+        // this template requires the model to have "\n\n" as EOT token
+        for (auto message : chat) {
+            std::string role(message->role);
+            if (role == "user") {
+                ss << "User: " << message->content << "\n\nAssistant:";
+            } else {
+                ss << message->content << "\n\n";
+            }
+        }
     } else {
         // template not supported
         return -1;

tests/test-backend-ops.cpp

@@ -1650,11 +1650,12 @@ struct test_mul_mat : public test_case {
     const int64_t m;
     const int64_t n;
     const int64_t k;
-    const std::array<int64_t, 2> bs; // dims 3 and 4
-    const std::array<int64_t, 2> nr; // repeat in dims 3 and 4
+    const std::array<int64_t, 2> bs;  // dims 3 and 4
+    const std::array<int64_t, 2> nr;  // repeat in dims 3 and 4
+    const std::array<int64_t, 4> per; // permutation of dimensions
 
     std::string vars() override {
-        return VARS_TO_STR7(type_a, type_b, m, n, k, bs, nr);
+        return VARS_TO_STR8(type_a, type_b, m, n, k, bs, nr, per);
     }
 
     double max_nmse_err() override {
@@ -1669,17 +1670,44 @@ struct test_mul_mat : public test_case {
     test_mul_mat(ggml_type type_a = GGML_TYPE_F32, ggml_type type_b = GGML_TYPE_F32,
             int64_t m = 32, int64_t n = 32, int64_t k = 32,
             std::array<int64_t, 2> bs = {10, 10},
-            std::array<int64_t, 2> nr = {2, 2})
-        : type_a(type_a), type_b(type_b), m(m), n(n), k(k), bs(bs), nr(nr) {}
+            std::array<int64_t, 2> nr = {2, 2},
+            std::array<int64_t, 4> per = {0, 1, 2, 3})
+        : type_a(type_a), type_b(type_b), m(m), n(n), k(k), bs(bs), nr(nr), per(per) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         // C^T = A * B^T: (k, m) * (k, n) => (m, n)
-        ggml_tensor * a = ggml_new_tensor_4d(ctx, type_a, k, m, bs[0]      , bs[1]);
-        ggml_tensor * b = ggml_new_tensor_4d(ctx, type_b, k, n, bs[0]*nr[0], bs[1]*nr[1]);
-        ggml_set_param(ctx, a);
-        ggml_set_param(ctx, b);
-        ggml_set_name(a, "a");
-        ggml_set_name(b, "b");
+        ggml_tensor * a;
+        ggml_tensor * b;
+
+        const int npermuted = (per[0] != 0) + (per[1] != 1) + (per[2] != 2) + (per[3] != 3);
+        if (npermuted > 0) {
+            GGML_ASSERT(npermuted == 2);
+            GGML_ASSERT(!ggml_is_quantized(type_a) || per[0] == 0);
+            GGML_ASSERT(!ggml_is_quantized(type_b) || per[0] == 0);
+
+            // Create tensors with the permuted dimensions, then permute them back to the dimensions given by m,n,k.
+            const int64_t ne_a[4] = {k, m, bs[0],       bs[1]};
+            const int64_t ne_b[4] = {k, n, bs[0]*nr[0], bs[1]*nr[1]};
+
+            a = ggml_new_tensor_4d(ctx, type_a, ne_a[per[0]], ne_a[per[1]], ne_a[per[2]], ne_a[per[3]]);
+            b = ggml_new_tensor_4d(ctx, type_b, ne_b[per[0]], ne_b[per[1]], ne_b[per[2]], ne_b[per[3]]);
+            ggml_set_param(ctx, a);
+            ggml_set_param(ctx, b);
+            ggml_set_name(a, "a");
+            ggml_set_name(b, "b");
+
+            a = ggml_permute(ctx, a, per[0], per[1], per[2], per[3]);
+            b = ggml_permute(ctx, b, per[0], per[1], per[2], per[3]);
+            ggml_set_name(a, "a_permuted");
+            ggml_set_name(b, "b_permuted");
+        } else {
+            a = ggml_new_tensor_4d(ctx, type_a, k, m, bs[0],       bs[1]);
+            b = ggml_new_tensor_4d(ctx, type_b, k, n, bs[0]*nr[0], bs[1]*nr[1]);
+            ggml_set_param(ctx, a);
+            ggml_set_param(ctx, b);
+            ggml_set_name(a, "a");
+            ggml_set_name(b, "b");
+        }
 
         ggml_tensor * out = ggml_mul_mat(ctx, a, b);
         ggml_set_name(out, "out");
@@ -3308,13 +3336,49 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
         }
     }
 
-    test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F32, GGML_TYPE_F32));
-    test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F32));
-    test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16));
-    // test cases for 1D im2col
+    // im2col 1D
     test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F32, GGML_TYPE_F32, {3000, 128, 1, 1}, {3, 128, 1280, 1}, 1, 0, 1, 0, 1, 0, false));
     test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F32, {3000, 128, 1, 1}, {3, 128, 1280, 1}, 1, 0, 1, 0, 1, 0, false));
     test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {3000, 128, 1, 1}, {3, 128, 1280, 1}, 1, 0, 1, 0, 1, 0, false));
+    for (int s0 : {1, 3}) {
+        for (int p0 : {0, 3}) {
+            for (int d0 : {1, 3}) {
+                test_cases.emplace_back(new test_im2col(
+                    GGML_TYPE_F32, GGML_TYPE_F32, GGML_TYPE_F32, {20, 2, 2, 1}, {3, 2, 2, 1},
+                    s0, 0, p0, 0, d0, 0, false));
+            }
+        }
+    }
+
+    // im2col 2D
+    test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F32, GGML_TYPE_F32));
+    test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F32));
+    test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16));
+    for (int s0 : {1, 3}) {
+        for (int s1 : {1, 3}) {
+            for (int p0 : {0, 3}) {
+                for (int p1 : {0, 3}) {
+                    for (int d0 : {1, 3}) {
+                        for (int d1 : {1, 3}) {
+                            test_cases.emplace_back(new test_im2col(
+                                GGML_TYPE_F32, GGML_TYPE_F32, GGML_TYPE_F32, {20, 20, 2, 2}, {3, 3, 2, 2},
+                                s0, s1, p0, p1, d0, d1, true));
+                        }
+                    }
+                }
+            }
+        }
+    }
+
+    // extra tests for im2col 2D
+    test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {12, 12, 1, 32}, {3, 3, 1, 32}, 1, 1, 1, 1, 1, 1, true));
+    test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {12, 12, 2, 32}, {3, 3, 2, 32}, 1, 1, 1, 1, 1, 1, true));
+    test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {12, 12, 1, 1024}, {3, 3, 1, 1024}, 1, 1, 1, 1, 1, 1, true));
+    test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {12, 12, 2, 1024}, {3, 3, 2, 1024}, 1, 1, 1, 1, 1, 1, true));
+    test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {12, 12, 1, 2048}, {3, 3, 1, 2048}, 1, 1, 1, 1, 1, 1, true));
+    test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {12, 12, 2, 2048}, {3, 3, 2, 2048}, 1, 1, 1, 1, 1, 1, true));
+    test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {12, 12, 1, 2560}, {3, 3, 1, 2560}, 1, 1, 1, 1, 1, 1, true));
+    test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {12, 12, 2, 2560}, {3, 3, 2, 2560}, 1, 1, 1, 1, 1, 1, true));
 
     // sycl backend will limit task global_range < MAX_INT
     // test cases for 2D im2col with large input W and H (occurs in stable-diffusion)
@@ -3442,13 +3506,14 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
 #if 1
     for (ggml_type type_a : base_types) {
         for (ggml_type type_b : {GGML_TYPE_F32, GGML_TYPE_F16}) {
-            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, { 1,  1}, {1, 1}));
-            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {10,  1}, {1, 1}));
-            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {10,  1}, {2, 1}));
-            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {10, 10}, {1, 1}));
-            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {10, 10}, {2, 1}));
-            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {10, 10}, {1, 2}));
-            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {10, 10}, {2, 2}));
+            // test cases without permutation
+            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  1, 256, { 1,  1}, {1, 1}));
+            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  1, 256, {10,  1}, {1, 1}));
+            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  1, 256, {10,  1}, {2, 1}));
+            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  1, 256, {10, 10}, {1, 1}));
+            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  1, 256, {10, 10}, {2, 1}));
+            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  1, 256, {10, 10}, {1, 2}));
+            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  1, 256, {10, 10}, {2, 2}));
 
             test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, { 1,  1}, {1, 1}));
             test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {10,  1}, {1, 1}));
@@ -3457,6 +3522,19 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
             test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {10, 10}, {2, 1}));
             test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {10, 10}, {1, 2}));
             test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {10, 10}, {2, 2}));
+
+            // test cases with permutation
+            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  1, 256, {2, 3}, {1, 1}, {0, 2, 1, 3}));
+            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  1, 256, {2, 3}, {1, 1}, {0, 1, 3, 2}));
+            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  1, 256, {2, 3}, {1, 1}, {0, 3, 2, 1}));
+
+            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  8, 256, {2, 3}, {1, 1}, {0, 2, 1, 3}));
+            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  8, 256, {2, 3}, {1, 1}, {0, 1, 3, 2}));
+            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  8, 256, {2, 3}, {1, 1}, {0, 3, 2, 1}));
+
+            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {2, 3}, {1, 1}, {0, 2, 1, 3}));
+            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {2, 3}, {1, 1}, {0, 1, 3, 2}));
+            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {2, 3}, {1, 1}, {0, 3, 2, 1}));
         }
     }
     for (ggml_type type_a : other_types) {

tests/test-sampling.cpp

@@ -18,203 +18,176 @@ static void dump(const llama_token_data_array * cur_p) {
 
 #define DUMP(__cur_p) do { printf("%s:%d (%s)\n", __FILE__, __LINE__, __func__); dump((__cur_p)); printf("-\n"); } while(0)
 
-#define APPLY(__cnstr, __cur_p) do { \
-    auto * cnstr = (__cnstr); \
-    llama_sampler_apply(cnstr, (__cur_p)); \
-    llama_sampler_free(cnstr); \
-} while(0)
+struct sampler_tester {
+    sampler_tester(size_t n_vocab) {
+        cur.reserve(n_vocab);
+        for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) {
+            const float logit = logf(token_id);
+            cur.emplace_back(llama_token_data{token_id, logit, 0.0f});
+        }
 
-static void test_top_k(const std::vector<float> & probs, const std::vector<float> & expected_probs, int k) {
-    const size_t n_vocab = probs.size();
+        cur_p = llama_token_data_array { cur.data(), cur.size(), -1, false };
+    }
+
+    sampler_tester(const std::vector<float> & probs, const std::vector<float> & probs_expected) : probs_expected(probs_expected) {
+        cur.reserve(probs.size());
+        for (llama_token token_id = 0; token_id < (llama_token)probs.size(); token_id++) {
+            const float logit = logf(probs[token_id]);
+            cur.emplace_back(llama_token_data{token_id, logit, probs[token_id]});
+        }
+
+        cur_p = llama_token_data_array { cur.data(), cur.size(), -1, false };
+    }
+
+    void apply(llama_sampler * sampler) {
+        llama_sampler_apply(sampler, &cur_p);
+        llama_sampler_free(sampler);
+    }
+
+    void check() {
+        GGML_ASSERT(cur_p.size == probs_expected.size());
+        for (size_t i = 0; i < cur_p.size; i++) {
+            GGML_ASSERT(fabs(cur_p.data[i].p - probs_expected[i]) < 1e-5);
+        }
+    }
+
+    llama_token_data_array cur_p;
+
+private:
+    const std::vector<float> probs_expected;
 
     std::vector<llama_token_data> cur;
-    cur.reserve(n_vocab);
-    for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) {
-        const float logit = logf(probs[token_id]);
-        cur.emplace_back(llama_token_data{token_id, logit, 0.0f});
-    }
+};
 
-    llama_token_data_array cur_p = { cur.data(), cur.size(), -1, false };
-    APPLY(llama_sampler_init_softmax(), &cur_p);
-    DUMP(&cur_p);
-    APPLY(llama_sampler_init_top_k(k), &cur_p);
-    DUMP(&cur_p);
+static void test_temp(const std::vector<float> & probs, const std::vector<float> & probs_expected, float temp) {
+    sampler_tester tester(probs, probs_expected);
 
-    GGML_ASSERT(cur_p.size == expected_probs.size());
-    for (size_t i = 0; i < cur_p.size; i++) {
-        GGML_ASSERT(fabs(cur_p.data[i].p - expected_probs[i]) < 1e-5);
-    }
+    DUMP(&tester.cur_p);
+    tester.apply(llama_sampler_init_temp(temp));
+    tester.apply(llama_sampler_init_dist(0));
+    DUMP(&tester.cur_p);
+
+    tester.check();
 }
 
-static void test_top_p(const std::vector<float> & probs, const std::vector<float> & expected_probs, float p) {
-    const size_t n_vocab = probs.size();
+static void test_temp_ext(const std::vector<float> & probs, const std::vector<float> & probs_expected, float temp, float delta, float exponent) {
+    sampler_tester tester(probs, probs_expected);
 
-    std::vector<llama_token_data> cur;
-    cur.reserve(n_vocab);
-    for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) {
-        const float logit = logf(probs[token_id]);
-        cur.emplace_back(llama_token_data{token_id, logit, 0.0f});
-    }
+    DUMP(&tester.cur_p);
+    tester.apply(llama_sampler_init_temp_ext(temp, delta, exponent));
+    tester.apply(llama_sampler_init_dist (0));
+    DUMP(&tester.cur_p);
 
-    llama_token_data_array cur_p = { cur.data(), cur.size(), -1, false };
-    APPLY(llama_sampler_init_softmax(), &cur_p);
-    DUMP(&cur_p);
-    APPLY(llama_sampler_init_top_p(p, 1), &cur_p);
-    DUMP(&cur_p);
-
-    GGML_ASSERT(cur_p.size == expected_probs.size());
-    for (size_t i = 0; i < cur_p.size; i++) {
-        GGML_ASSERT(fabs(cur_p.data[i].p - expected_probs[i]) < 1e-3);
-    }
+    tester.check();
 }
 
-static void test_tfs(const std::vector<float> & probs, const std::vector<float> & expected_probs, float z) {
-    const size_t n_vocab = probs.size();
+static void test_top_k(const std::vector<float> & probs, const std::vector<float> & probs_expected, int k) {
+    sampler_tester tester(probs, probs_expected);
 
-    std::vector<llama_token_data> cur;
-    cur.reserve(n_vocab);
-    for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) {
-        const float logit = logf(probs[token_id]);
-        cur.emplace_back(llama_token_data{token_id, logit, 0.0f});
-    }
+    DUMP(&tester.cur_p);
+    tester.apply(llama_sampler_init_top_k(k));
+    tester.apply(llama_sampler_init_dist (0));
+    DUMP(&tester.cur_p);
 
-    llama_token_data_array cur_p = { cur.data(), cur.size(), -1, false };
-    DUMP(&cur_p);
-    APPLY(llama_sampler_init_tail_free(z, 1), &cur_p);
-    DUMP(&cur_p);
-
-    GGML_ASSERT(cur_p.size == expected_probs.size());
-    for (size_t i = 0; i < cur_p.size; i++) {
-        GGML_ASSERT(fabs(cur_p.data[i].p - expected_probs[i]) < 1e-3);
-    }
+    tester.check();
 }
 
-static void test_min_p(const std::vector<float> & probs, const std::vector<float> & expected_probs, float p) {
-    const size_t n_vocab = probs.size();
+static void test_top_p(const std::vector<float> & probs, const std::vector<float> & probs_expected, float p) {
+    sampler_tester tester(probs, probs_expected);
 
-    std::vector<llama_token_data> cur;
-    cur.reserve(n_vocab);
-    for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) {
-        const float logit = logf(probs[token_id]);
-        cur.emplace_back(llama_token_data{token_id, logit, 0.0f});
-    }
+    DUMP(&tester.cur_p);
+    tester.apply(llama_sampler_init_top_p(p, 1));
+    tester.apply(llama_sampler_init_dist (0));
+    DUMP(&tester.cur_p);
 
-    llama_token_data_array cur_p = { cur.data(), cur.size(), -1, false };
-    DUMP(&cur_p);
-    APPLY(llama_sampler_init_min_p(p, 1), &cur_p);
-    DUMP(&cur_p);
-    APPLY(llama_sampler_init_softmax(), &cur_p);
-
-    GGML_ASSERT(cur_p.size == expected_probs.size());
-    for (size_t i = 0; i < cur_p.size; i++) {
-        GGML_ASSERT(fabs(cur_p.data[i].p - expected_probs[i]) < 1e-3);
-    }
+    tester.check();
 }
 
-static void test_xtc(const std::vector<float> & probs, const std::vector<float> & expected_probs, float p, float t) {
-    const size_t n_vocab = probs.size();
+static void test_tfs(const std::vector<float> & probs, const std::vector<float> & probs_expected, float z) {
+    sampler_tester tester(probs, probs_expected);
 
-    std::vector<llama_token_data> cur;
-    cur.reserve(n_vocab);
-    for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) {
-        const float logit = logf(probs[token_id]);
-        cur.emplace_back(llama_token_data{token_id, logit, 0.0f});
-    }
+    DUMP(&tester.cur_p);
+    tester.apply(llama_sampler_init_tail_free(z, 1));
+    DUMP(&tester.cur_p);
 
-    llama_token_data_array cur_p = { cur.data(), cur.size(), -1, false };
-    APPLY(llama_sampler_init_softmax(), &cur_p);
-    DUMP(&cur_p);
-    APPLY(llama_sampler_init_xtc(p, t, 0, 0), &cur_p);
-    DUMP(&cur_p);
-
-    GGML_ASSERT(cur_p.size == expected_probs.size());
-    for (size_t i = 0; i < cur_p.size; i++) {
-        GGML_ASSERT(fabs(cur_p.data[i].p - expected_probs[i]) < 1e-5);
-    }
+    tester.check();
 }
 
-static void test_typical(const std::vector<float> & probs, const std::vector<float> & expected_probs, float p) {
-    const size_t n_vocab = probs.size();
+static void test_min_p(const std::vector<float> & probs, const std::vector<float> & probs_expected, float p) {
+    sampler_tester tester(probs, probs_expected);
 
-    std::vector<llama_token_data> cur;
-    cur.reserve(n_vocab);
-    for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) {
-        const float logit = logf(probs[token_id]);
-        cur.emplace_back(llama_token_data{token_id, logit, 0.0f});
-    }
+    DUMP(&tester.cur_p);
+    tester.apply(llama_sampler_init_min_p(p, 1));
+    tester.apply(llama_sampler_init_dist (0));
+    DUMP(&tester.cur_p);
 
-    llama_token_data_array cur_p = { cur.data(), cur.size(), -1, false };
-    DUMP(&cur_p);
-    APPLY(llama_sampler_init_typical(p, 1), &cur_p);
-    DUMP(&cur_p);
+    tester.check();
+}
 
-    GGML_ASSERT(cur_p.size == expected_probs.size());
-    for (size_t i = 0; i < cur_p.size; i++) {
-        GGML_ASSERT(fabs(cur_p.data[i].p - expected_probs[i]) < 1e-3);
-    }
+static void test_xtc(const std::vector<float> & probs, const std::vector<float> & probs_expected, float p, float t) {
+    sampler_tester tester(probs, probs_expected);
+
+    DUMP(&tester.cur_p);
+    tester.apply(llama_sampler_init_xtc(p, t, 0, 0));
+    DUMP(&tester.cur_p);
+
+    tester.check();
+}
+
+static void test_typical(const std::vector<float> & probs, const std::vector<float> & probs_expected, float p) {
+    sampler_tester tester(probs, probs_expected);
+
+    DUMP(&tester.cur_p);
+    tester.apply(llama_sampler_init_typical(p, 1));
+    DUMP(&tester.cur_p);
+
+    tester.check();
 }
 
 static void test_penalties(
     const std::vector<float> & probs, const std::vector<llama_token> & last_tokens,
-    const std::vector<float> & expected_probs, float repeat_penalty, float alpha_frequency, float alpha_presence
+    const std::vector<float> & probs_expected, float repeat_penalty, float alpha_frequency, float alpha_presence
 ) {
-    GGML_ASSERT(probs.size() == expected_probs.size());
+    GGML_ASSERT(probs.size() == probs_expected.size());
+
+    sampler_tester tester(probs, probs_expected);
 
     const size_t n_vocab = probs.size();
-
-    std::vector<llama_token_data> cur;
-    cur.reserve(n_vocab);
-    for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) {
-        const float logit = logf(probs[token_id]);
-        cur.emplace_back(llama_token_data{token_id, logit, 0.0f});
-    }
-
-    llama_token_data_array cur_p = { cur.data(), cur.size(), -1, false };
-
     auto * sampler = llama_sampler_init_penalties(n_vocab, LLAMA_TOKEN_NULL, LLAMA_TOKEN_NULL, last_tokens.size(), repeat_penalty, alpha_frequency, alpha_presence, false, false);
 
     for (size_t i = 0; i < last_tokens.size(); i++) {
         llama_sampler_accept(sampler, last_tokens[i]);
     }
 
-    APPLY(llama_sampler_init_softmax(), &cur_p);
-    DUMP(&cur_p);
-    APPLY(sampler, &cur_p);
-    APPLY(llama_sampler_init_softmax(), &cur_p);
-    DUMP(&cur_p);
+    DUMP(&tester.cur_p);
+    tester.apply(sampler);
+    tester.apply(llama_sampler_init_dist(0));
+    DUMP(&tester.cur_p);
 
-    GGML_ASSERT(cur_p.size == expected_probs.size());
-    for (size_t i = 0; i < cur_p.size; i++) {
-        GGML_ASSERT(fabs(cur_p.data[i].p - expected_probs[i]) < 1e-3);
-    }
+    tester.check();
 }
 
 static void test_sampler_queue(const size_t n_vocab, const std::string & samplers_sequence, const int top_k, const float top_p, const float min_p
 ) {
-    std::vector<llama_token_data> cur;
-    cur.reserve(n_vocab);
-    for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) {
-        const float logit = logf(token_id);
-        cur.emplace_back(llama_token_data{token_id, logit, 0.0f});
-    }
-
-    llama_token_data_array cur_p = { cur.data(), cur.size(), -1, false };
+    sampler_tester tester(n_vocab);
 
           llama_token min_token_id = 0;
     const llama_token max_token_id = n_vocab-1;
 
     for (auto s : samplers_sequence) {
         switch (s){
-            case 'k': APPLY(llama_sampler_init_top_k(top_k), &cur_p); break;
+            case 'k': tester.apply(llama_sampler_init_top_k(top_k)); break;
             case 'f': GGML_ABORT("tail_free test not implemented");
             case 'y': GGML_ABORT("typical test not implemented");
-            case 'p': APPLY(llama_sampler_init_top_p(top_p, 1), &cur_p); break;
-            case 'm': APPLY(llama_sampler_init_min_p(min_p, 1), &cur_p); break;
+            case 'p': tester.apply(llama_sampler_init_top_p(top_p, 1)); break;
+            case 'm': tester.apply(llama_sampler_init_min_p(min_p, 1)); break;
             case 't': GGML_ABORT("temperature test not implemented");
             default : GGML_ABORT("Unknown sampler");
         }
 
-        APPLY(llama_sampler_init_softmax(), &cur_p); // make sure tokens are sorted for tests
+        tester.apply(llama_sampler_init_dist(0));
+
+        auto & cur_p = tester.cur_p;
 
         const int size = cur_p.size;
 
@@ -307,21 +280,26 @@ static void test_perf() {
     BENCH(llama_sampler_init_tail_free(0.5f, 1),                data, 32);
     BENCH(llama_sampler_init_typical  (0.5f, 1),                data, 32);
     BENCH(llama_sampler_init_xtc      (1.0f, 0.1f, 1, 1),       data, 32);
-    BENCH(llama_sampler_init_softmax  (),                       data, 32);
 }
 
 int main(void) {
     ggml_time_init();
 
-    test_top_k({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f}, 1);
-    test_top_k({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f}, 3);
+    test_temp({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f, 0.1f}, 1.0f);
+    test_temp({0.1f, 0.2f, 0.3f, 0.4f}, {1.0f, 0.0f, 0.0f, 0.0f}, 0.0f);
+
+    test_temp_ext({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f, 0.1f}, 1.0f, 0.0f, 1.0f);
+    test_temp_ext({0.1f, 0.2f, 0.3f, 0.4f}, {1.0f, 0.0f, 0.0f, 0.0f}, 0.0f, 0.0f, 1.0f);
+
+    test_top_k({0.1f, 0.2f, 0.3f, 0.4f}, {1.0f}, 1);
+    test_top_k({0.1f, 0.2f, 0.3f, 0.4f}, {0.44444f, 0.33333f, 0.22222f}, 3);
     test_top_k({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f, 0.1f}, 4);
     test_top_k({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f, 0.1f}, 0);
 
-    test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f}, 0);
-    test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f}, 0.7f);
-    test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f}, 0.8f);
-    test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f, 0.1f}, 1);
+    test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {1.0f}, 0);
+    test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.571429f, 0.428571f}, 0.7f);
+    test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.44444f, 0.33333f, 0.22222f}, 0.8f);
+    test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f, 0.1f}, 1.0f);
 
     test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/1.0f, 0.3f/1.0f, 0.2f/1.0f, 0.1f/1.0f}, 0.00f);
     test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/1.0f, 0.3f/1.0f, 0.2f/1.0f, 0.1f/1.0f}, 0.24f);