py : add Gemma conversion from HF models (#5647)

* py : add gemma conversion from HF models

* Update convert-hf-to-gguf.py

Co-authored-by: Aarni Koskela <akx@iki.fi>

* Update convert-hf-to-gguf.py

Co-authored-by: Aarni Koskela <akx@iki.fi>

* Update convert-hf-to-gguf.py

Co-authored-by: Jared Van Bortel <jared@nomic.ai>

---------

Co-authored-by: Aarni Koskela <akx@iki.fi>
Co-authored-by: Jared Van Bortel <jared@nomic.ai>

.devops/nix/docker.nix

@@ -0,0 +1,37 @@
+{
+  lib,
+  dockerTools,
+  buildEnv,
+  llama-cpp,
+  interactive ? true,
+  coreutils,
+}:
+
+# A tar that can be fed into `docker load`:
+#
+# $ nix build .#llamaPackages.docker
+# $ docker load < result
+
+# For details and variations cf.
+# - https://nixos.org/manual/nixpkgs/unstable/#ssec-pkgs-dockerTools-buildLayeredImage
+# - https://discourse.nixos.org/t/a-faster-dockertools-buildimage-prototype/16922
+# - https://nixery.dev/
+
+# Approximate (compressed) sizes, at the time of writing, are:
+#
+# .#llamaPackages.docker: 125M;
+# .#llamaPackagesCuda.docker: 537M;
+# .#legacyPackages.aarch64-linux.llamaPackagesXavier.docker: 415M.
+
+dockerTools.buildLayeredImage {
+  name = llama-cpp.pname;
+  tag = "latest";
+
+  contents =
+    [ llama-cpp ]
+    ++ lib.optionals interactive [
+      coreutils
+      dockerTools.binSh
+      dockerTools.caCertificates
+    ];
+}

.devops/nix/scope.nix

@@ -12,5 +12,8 @@ lib.makeScope newScope (
   self: {
     inherit llamaVersion;
     llama-cpp = self.callPackage ./package.nix { };
+    docker = self.callPackage ./docker.nix { };
+    docker-min = self.callPackage ./docker.nix { interactive = false; };
+    sif = self.callPackage ./sif.nix { };
   }
 )

.devops/nix/sif.nix

@@ -0,0 +1,27 @@
+{
+  lib,
+  singularity-tools,
+  llama-cpp,
+  bashInteractive,
+  interactive ? false,
+}:
+
+let
+    optionalInt = cond: x: if cond then x else 0;
+in
+singularity-tools.buildImage rec {
+  inherit (llama-cpp) name;
+  contents = [ llama-cpp ] ++ lib.optionals interactive [ bashInteractive ];
+
+  # These are excessive (but safe) for most variants. Building singularity
+  # images requires superuser privileges, so we build them inside a VM in a
+  # writable image of pre-determined size.
+  #
+  # ROCm is currently affected by https://github.com/NixOS/nixpkgs/issues/276846
+  #
+  # Expected image sizes:
+  # - cpu/blas: 150M,
+  # - cuda, all gencodes: 560M,
+  diskSize = 4096 + optionalInt llama-cpp.useRocm 16384;
+  memSize = diskSize;
+}

.github/workflows/nix-ci-aarch64.yml

@@ -19,7 +19,6 @@ on:
 
 jobs:
   nix-build-aarch64:
-    if: ${{ vars.CACHIX_NAME != '' }}
     runs-on: ubuntu-latest
     steps:
     - name: Checkout repository
@@ -37,8 +36,8 @@ jobs:
         extra-conf: |
           extra-platforms = aarch64-linux
           extra-system-features = nixos-test kvm
-          extra-substituters = https://${{ vars.CACHIX_NAME }}.cachix.org https://cuda-maintainers.cachix.org
-          extra-trusted-public-keys = ${{ vars.CACHIX_PUBLIC_KEY }} cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
+          extra-substituters = https://llama-cpp.cachix.org https://cuda-maintainers.cachix.org
+          extra-trusted-public-keys = llama-cpp.cachix.org-1:H75X+w83wUKTIPSO1KWy9ADUrzThyGs8P5tmAbkWhQc= cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
     - uses: DeterminateSystems/magic-nix-cache-action@v2
       with:
         upstream-cache: https://${{ matrix.cachixName }}.cachix.org
@@ -46,7 +45,7 @@ jobs:
       uses: cachix/cachix-action@v13
       with:
         authToken: '${{ secrets.CACHIX_AUTH_TOKEN }}'
-        name: ${{ vars.CACHIX_NAME }}
+        name: llama-cpp
     - name: Show all output paths
       run: >
           nix run github:nix-community/nix-eval-jobs

.github/workflows/nix-ci.yml

@@ -23,8 +23,8 @@ jobs:
       with:
         github-token: ${{ secrets.GITHUB_TOKEN }}
         extra-conf: |
-          extra-substituters = https://${{ vars.CACHIX_NAME }}.cachix.org https://cuda-maintainers.cachix.org
-          extra-trusted-public-keys = ${{ vars.CACHIX_PUBLIC_KEY }} cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
+          extra-substituters = https://llama-cpp.cachix.org https://cuda-maintainers.cachix.org
+          extra-trusted-public-keys = llama-cpp.cachix.org-1:H75X+w83wUKTIPSO1KWy9ADUrzThyGs8P5tmAbkWhQc= cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
     - uses: DeterminateSystems/magic-nix-cache-action@v2
       with:
         upstream-cache: https://${{ matrix.cachixName }}.cachix.org
@@ -37,7 +37,6 @@ jobs:
           --flake
           ".#packages.$(nix eval --raw --impure --expr builtins.currentSystem)"
   nix-build:
-    if: ${{ vars.CACHIX_NAME != '' }}
     strategy:
       fail-fast: false
       matrix:
@@ -51,8 +50,8 @@ jobs:
       with:
         github-token: ${{ secrets.GITHUB_TOKEN }}
         extra-conf: |
-          extra-substituters = https://${{ vars.CACHIX_NAME }}.cachix.org https://cuda-maintainers.cachix.org
-          extra-trusted-public-keys = ${{ vars.CACHIX_PUBLIC_KEY }} cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
+          extra-substituters = https://llama-cpp.cachix.org https://cuda-maintainers.cachix.org
+          extra-trusted-public-keys = llama-cpp.cachix.org-1:H75X+w83wUKTIPSO1KWy9ADUrzThyGs8P5tmAbkWhQc= cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
     - uses: DeterminateSystems/magic-nix-cache-action@v2
       with:
         upstream-cache: https://${{ matrix.cachixName }}.cachix.org
@@ -60,7 +59,7 @@ jobs:
       uses: cachix/cachix-action@v13
       with:
         authToken: '${{ secrets.CACHIX_AUTH_TOKEN }}'
-        name: ${{ vars.CACHIX_NAME }}
+        name: llama-cpp
     - name: Build
       run: >
           nix run github:Mic92/nix-fast-build

README.md

@@ -10,13 +10,9 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 
 ### Hot topics
 
-- Remove LLAMA_MAX_DEVICES and LLAMA_SUPPORTS_GPU_OFFLOAD: https://github.com/ggerganov/llama.cpp/pull/5240
-- Incoming backends: https://github.com/ggerganov/llama.cpp/discussions/5138
-  - [SYCL backend](README-sycl.md) is ready (1/28/2024), support Linux/Windows in Intel GPUs (iGPU, Arc/Flex/Max series)
-- New SOTA quantized models, including pure 2-bits: https://huggingface.co/ikawrakow
-- Collecting Apple Silicon performance stats:
-  - M-series: https://github.com/ggerganov/llama.cpp/discussions/4167
-  - A-series: https://github.com/ggerganov/llama.cpp/discussions/4508
+- Support for chat templates: [Wiki (contributions welcome)](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template)
+- Support for Gemma models: https://github.com/ggerganov/llama.cpp/pull/5631
+- Non-linear quantization IQ4_NL: https://github.com/ggerganov/llama.cpp/pull/5590
 - Looking for contributions to improve and maintain the `server` example: https://github.com/ggerganov/llama.cpp/issues/4216
 
 ----
@@ -146,6 +142,7 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 - [nat/openplayground](https://github.com/nat/openplayground)
 - [Faraday](https://faraday.dev/) (proprietary)
 - [LMStudio](https://lmstudio.ai/) (proprietary)
+- [LocalAI](https://github.com/mudler/LocalAI) (MIT)
 - [LostRuins/koboldcpp](https://github.com/LostRuins/koboldcpp) (AGPL)
 - [Mozilla-Ocho/llamafile](https://github.com/Mozilla-Ocho/llamafile)
 - [nomic-ai/gpt4all](https://github.com/nomic-ai/gpt4all)

convert-hf-to-gguf.py

@@ -218,6 +218,8 @@ class Model:
             return BertModel
         if model_architecture == "NomicBertModel":
             return NomicBertModel
+        if model_architecture == "GemmaForCausalLM":
+            return GemmaModel
         return Model
 
     def _is_model_safetensors(self) -> bool:
@@ -277,6 +279,8 @@ class Model:
             return gguf.MODEL_ARCH.BERT
         if arch == "NomicBertModel":
             return gguf.MODEL_ARCH.NOMIC_BERT
+        if arch == "GemmaForCausalLM":
+            return gguf.MODEL_ARCH.GEMMA
 
         raise NotImplementedError(f'Architecture "{arch}" not supported!')
 
@@ -655,6 +659,8 @@ class OrionModel(Model):
         self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
         self.gguf_writer.add_head_count(head_count)
         self.gguf_writer.add_head_count_kv(head_count_kv)
+        # note: config provides rms norm but it is actually layer norm
+        # ref:  https://huggingface.co/OrionStarAI/Orion-14B-Chat/blob/276a17221ce42beb45f66fac657a41540e71f4f5/modeling_orion.py#L570-L571
         self.gguf_writer.add_layer_norm_eps(self.hparams["rms_norm_eps"])
 
     def write_tensors(self):
@@ -1031,7 +1037,6 @@ class PersimmonModel(Model):
         self.gguf_writer.add_head_count_kv(head_count_kv)
         self.gguf_writer.add_rope_freq_base(self.hparams["rope_theta"])
         self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_eps"])
-        self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"])
 
     def set_vocab(self):
         self._set_vocab_sentencepiece()
@@ -1785,6 +1790,62 @@ class NomicBertModel(BertModel):
             yield name, data
 
 
+class GemmaModel(Model):
+    def set_vocab(self):
+        self._set_vocab_sentencepiece()
+
+    def set_gguf_parameters(self):
+        hparams = self.hparams
+        block_count = hparams["num_hidden_layers"]
+
+        self.gguf_writer.add_name(self.dir_model.name)
+        self.gguf_writer.add_context_length(hparams["max_position_embeddings"])
+        self.gguf_writer.add_embedding_length(hparams["hidden_size"])
+        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_feed_forward_length(hparams["intermediate_size"])
+        self.gguf_writer.add_head_count(hparams["num_attention_heads"])
+        self.gguf_writer.add_head_count_kv(self.hparams["num_key_value_heads"] if "num_key_value_heads" in hparams else hparams["num_attention_heads"])
+        self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"])
+        self.gguf_writer.add_key_length(hparams["head_dim"])
+        self.gguf_writer.add_value_length(hparams["head_dim"])
+
+    def write_tensors(self):
+        block_count = self.hparams.get("n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer")))
+        tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
+
+        for name, data_torch in self.get_tensors():
+            # ref: https://github.com/huggingface/transformers/blob/fc37f38915372c15992b540dfcbbe00a916d4fc6/src/transformers/models/gemma/modeling_gemma.py#L89
+            if name.endswith("norm.weight"):
+                data_torch = data_torch + 1
+
+            old_dtype = data_torch.dtype
+
+            # convert any unsupported data types to float32
+            if data_torch.dtype not in (torch.float16, torch.float32):
+                data_torch = data_torch.to(torch.float32)
+
+            data = data_torch.squeeze().numpy()
+
+            # map tensor names
+            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
+            if new_name is None:
+                print(f"Can not map tensor {name!r}")
+                sys.exit()
+
+            n_dims = len(data.shape)
+            data_dtype = data.dtype
+
+            data = data.astype(np.float32)
+
+            # if f16 desired, convert any float32 2-dim weight tensors to float16
+            if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
+                data = data.astype(np.float16)
+
+            print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
+
+            self.gguf_writer.add_tensor(new_name, data)
+
+
 ###### CONVERSION LOGIC ######
 
 

examples/llava/README.md

@@ -63,13 +63,12 @@ Now both the LLaMA part and the image encoder is in the `llava-v1.5-7b` director
 ```console
 git clone https://huggingface.co/liuhaotian/llava-v1.6-vicuna-7b
 ```
-2) Backup your pth/safetensor model files as llava-surgery modifies them
-3) Use `llava-surgery-v2.py` which also supports llava-1.5 variants pytorch as well as safetensor models:
+2) Use `llava-surgery-v2.py` which also supports llava-1.5 variants pytorch as well as safetensor models:
 ```console
 python examples/llava/llava-surgery-v2.py -C -m ../llava-v1.6-vicuna-7b/
 ```
 - you will find a llava.projector and a llava.clip file in your model directory
-4) Copy the llava.clip file into a subdirectory (like vit), rename it to pytorch_model.bin and add a fitting vit configuration to the directory:
+3) Copy the llava.clip file into a subdirectory (like vit), rename it to pytorch_model.bin and add a fitting vit configuration to the directory:
 ```console
 mkdir vit
 cp ../llava-v1.6-vicuna-7b/llava.clip vit/pytorch_model.bin
@@ -77,18 +76,18 @@ cp ../llava-v1.6-vicuna-7b/llava.projector vit/
 curl -s -q https://huggingface.co/cmp-nct/llava-1.6-gguf/raw/main/config_vit.json -o vit/config.json
 ```
 
-5) Create the visual gguf model:
+4) Create the visual gguf model:
 ```console
 python ./examples/llava/convert-image-encoder-to-gguf.py -m vit --llava-projector vit/llava.projector --output-dir vit --clip-model-is-vision
 ```
 - This is similar to llava-1.5, the difference is that we tell the encoder that we are working with the pure vision model part of CLIP
 
-6) Then convert the model to gguf format:
+5) Then convert the model to gguf format:
 ```console
-python ./convert.py ../llava-v1.6-vicuna-7b/
+python ./convert.py ../llava-v1.6-vicuna-7b/ --skip-unknown
 ```
 
-7) And finally we can run the llava-cli using the 1.6 model version:
+6) And finally we can run the llava-cli using the 1.6 model version:
 ```console
 ./llava-cli -m ../llava-v1.6-vicuna-7b/ggml-model-f16.gguf --mmproj vit/mmproj-model-f16.gguf --image some-image.jpg -c 4096
 ```

examples/llava/llava-surgery-v2.py

@@ -65,9 +65,7 @@ def clean_vision_tower_from_checkpoint(checkpoint_path):
         for name in clip_tensors:
             del checkpoint[name]
 
-        # Save the updated checkpoint
         checkpoint_path = checkpoint_path
-        save_model(checkpoint, checkpoint_path, file_type)
         return True
     return False
 
@@ -152,16 +150,6 @@ for name in first_mm_tensors:
 if len(projector) > 0:
     save_model(projector, f"{args.model}/llava.projector", 'pytorch')
 
-for name in mm_tensors:
-    del last_checkpoint[name]
-for name in first_mm_tensors:
-    del first_checkpoint[name]
-
-if len(mm_tensors) > 0:
-    save_model(last_checkpoint, projector_checkpoint_path, file_type)
-if len(first_mm_tensors) > 0:
-    save_model(first_checkpoint, newline_checkpoint_path, file_type)
-
 print("Done!")
 print(f"Now you can convert {args.model} to a a regular LLaMA GGUF file.")
 print(f"Also, use {args.model}/llava.projector to prepare a llava-encoder.gguf file.")

examples/main/main.cpp

@@ -334,6 +334,8 @@ int main(int argc, char ** argv) {
     // number of tokens to keep when resetting context
     if (params.n_keep < 0 || params.n_keep > (int) embd_inp.size() || params.instruct || params.chatml) {
         params.n_keep = (int)embd_inp.size();
+    } else {
+        params.n_keep += add_bos; // always keep the BOS token
     }
 
     // prefix & suffix for instruct mode
@@ -383,8 +385,8 @@ int main(int argc, char ** argv) {
             }
         }
 
-        if (params.n_keep > 0) {
-        LOG_TEE("%s: static prompt based on n_keep: '", __func__);
+        if (params.n_keep > add_bos) {
+            LOG_TEE("%s: static prompt based on n_keep: '", __func__);
             for (int i = 0; i < params.n_keep; i++) {
                 LOG_TEE("%s", llama_token_to_piece(ctx, embd_inp[i]).c_str());
             }
@@ -540,14 +542,14 @@ int main(int argc, char ** argv) {
                         break;
                     }
 
-                    const int n_left    = n_past - params.n_keep - 1;
+                    const int n_left    = n_past - params.n_keep;
                     const int n_discard = n_left/2;
 
                     LOG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
                             n_past, n_left, n_ctx, params.n_keep, n_discard);
 
-                    llama_kv_cache_seq_rm   (ctx, 0, params.n_keep + 1            , params.n_keep + n_discard + 1);
-                    llama_kv_cache_seq_shift(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
+                    llama_kv_cache_seq_rm   (ctx, 0, params.n_keep            , params.n_keep + n_discard);
+                    llama_kv_cache_seq_shift(ctx, 0, params.n_keep + n_discard, n_past, -n_discard);
 
                     n_past -= n_discard;
 

examples/server/README.md

@@ -41,6 +41,7 @@ see https://github.com/ggerganov/llama.cpp/issues/1437
 - `--grp-attn-w`: Set the group attention width to extend context size through self-extend(default: 512), used together with group attention factor `--grp-attn-n`
 - `-n, --n-predict`: Set the maximum tokens to predict (default: -1)
 - `--slots-endpoint-disable`: To disable slots state monitoring endpoint. Slots state may contain user data, prompts included.
+- `--chat-template JINJA_TEMPLATE`: Set custom jinja chat template. This parameter accepts a string, not a file name (default: template taken from model's metadata). We only support [some pre-defined templates](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template)
 
 ## Build
 
@@ -140,6 +141,8 @@ node index.js
   - 200 -> `{"status": "no slot available", "slots_idle": 0, "slots_processing": 32}` if no slot are currently available.
   - 503 -> `{"status": "no slot available", "slots_idle": 0, "slots_processing": 32}` if the query parameter `fail_on_no_slot` is provided and no slot are currently available.
 
+  If the query parameter `include_slots` is passed, `slots` field will contain internal slots data except if `--slots-endpoint-disable` is set.
+
 - **POST** `/completion`: Given a `prompt`, it returns the predicted completion.
 
     *Options:*
@@ -148,7 +151,7 @@ node index.js
 
     `temperature`: Adjust the randomness of the generated text (default: 0.8).
 
-    `dynatemp_range`: Dynamic temperature range (default: 0.0, 0.0 = disabled).
+    `dynatemp_range`: Dynamic temperature range. The final temperature will be in the range of `[temperature - dynatemp_range; temperature + dynatemp_range]` (default: 0.0, 0.0 = disabled).
 
     `dynatemp_exponent`: Dynamic temperature exponent (default: 1.0).
 
@@ -206,7 +209,7 @@ node index.js
 
     `slot_id`: Assign the completion task to an specific slot. If is -1 the task will be assigned to a Idle slot (default: -1)
 
-    `cache_prompt`: Save the prompt and generation for avoid reprocess entire prompt if a part of this isn't change (default: false)
+    `cache_prompt`: Re-use previously cached prompt from the last request if possible. This may prevent re-caching the prompt from scratch. (default: false)
 
     `system_prompt`: Change the system prompt (initial prompt of all slots), this is useful for chat applications. [See more](#change-system-prompt-on-runtime)
 
@@ -239,7 +242,7 @@ Notice that each `probs` is an array of length `n_probs`.
 
 - `content`: Completion result as a string (excluding `stopping_word` if any). In case of streaming mode, will contain the next token as a string.
 - `stop`: Boolean for use with `stream` to check whether the generation has stopped (Note: This is not related to stopping words array `stop` from input options)
-- `generation_settings`: The provided options above excluding `prompt` but including `n_ctx`, `model`
+- `generation_settings`: The provided options above excluding `prompt` but including `n_ctx`, `model`. These options may differ from the original ones in some way (e.g. bad values filtered out, strings converted to tokens, etc.).
 - `model`: The path to the model loaded with `-m`
 - `prompt`: The provided `prompt`
 - `stopped_eos`: Indicating whether the completion has stopped because it encountered the EOS token

examples/server/server.cpp

@@ -400,6 +400,16 @@ struct llama_server_context
         return true;
     }
 
+    void validate_model_chat_template(server_params & sparams) {
+        llama_chat_message chat[] = {{"user", "test"}};
+        std::vector<char> buf(1);
+        int res = llama_chat_apply_template(model, nullptr, chat, 1, true, buf.data(), buf.size());
+        if (res < 0) {
+            LOG_ERROR("The chat template comes with this model is not yet supported, falling back to chatml. This may cause the model to output suboptimal responses", {});
+            sparams.chat_template = "<|im_start|>"; // llama_chat_apply_template only checks if <|im_start|> exist in the template
+        }
+    }
+
     void initialize() {
         // create slots
         all_slots_are_idle = true;
@@ -1394,6 +1404,46 @@ struct llama_server_context
             case TASK_TYPE_NEXT_RESPONSE: {
                 // do nothing
             } break;
+            case TASK_TYPE_SLOTS_DATA: {
+                json slots_data        = json::array();
+                int n_idle_slots       = 0;
+                int n_processing_slots = 0;
+
+                for (llama_client_slot &slot: slots) {
+                    if (slot.available()) {
+                        n_idle_slots++;
+                    } else {
+                        n_processing_slots++;
+                    }
+                    json slot_data = get_formated_generation(slot);
+                    slot_data["id"] = slot.id;
+                    slot_data["task_id"] = slot.task_id;
+                    slot_data["state"] = slot.state;
+                    slot_data["prompt"] = slot.prompt;
+                    slot_data["next_token"] = {
+                            {"has_next_token", slot.has_next_token},
+                            {"n_remain", slot.n_remaining},
+                            {"num_tokens_predicted", slot.n_decoded},
+                            {"stopped_eos", slot.stopped_eos},
+                            {"stopped_word", slot.stopped_word},
+                            {"stopped_limit", slot.stopped_limit},
+                            {"stopping_word", slot.stopping_word},
+                    };
+                    slots_data.push_back(slot_data);
+                }
+                LOG_TEE("task %i - slots data: idle=%i processing=%i\n", task.id, n_idle_slots, n_processing_slots);
+                task_result res;
+                res.id = task.id;
+                res.multitask_id = task.multitask_id;
+                res.stop = true;
+                res.error = false;
+                res.result_json = {
+                        { "idle",       n_idle_slots       },
+                        { "processing", n_processing_slots },
+                        { "slots",      slots_data         }
+                };
+                queue_results.send(res);
+            } break;
         }
     }
 
@@ -1447,14 +1497,15 @@ struct llama_server_context
                 if (slot.is_processing() && system_tokens.size() + slot.cache_tokens.size() >= (size_t) slot.n_ctx)
                 {
                     // Shift context
-                    const int n_left    = system_tokens.size() + slot.n_past - slot.params.n_keep - 1;
+                    const int n_keep    = slot.params.n_keep + add_bos_token;
+                    const int n_left    = system_tokens.size() + slot.n_past - n_keep;
                     const int n_discard = n_left / 2;
 
-                    LOG_TEE("slot %d: context shift - n_keep = %d, n_left = %d, n_discard = %d\n", slot.id, slot.params.n_keep, n_left, n_discard);
-                    llama_kv_cache_seq_rm   (ctx, slot.id, slot.params.n_keep + 1            , slot.params.n_keep + n_discard + 1);
-                    llama_kv_cache_seq_shift(ctx, slot.id, slot.params.n_keep + 1 + n_discard, system_tokens.size() + slot.n_past, -n_discard);
+                    LOG_TEE("slot %d: context shift - n_keep = %d, n_left = %d, n_discard = %d\n", slot.id, n_keep, n_left, n_discard);
+                    llama_kv_cache_seq_rm   (ctx, slot.id, n_keep            , n_keep + n_discard);
+                    llama_kv_cache_seq_shift(ctx, slot.id, n_keep + n_discard, system_tokens.size() + slot.n_past, -n_discard);
 
-                    for (size_t i = slot.params.n_keep + 1 + n_discard; i < slot.cache_tokens.size(); i++)
+                    for (size_t i = n_keep + n_discard; i < slot.cache_tokens.size(); i++)
                     {
                         slot.cache_tokens[i - n_discard] = slot.cache_tokens[i];
                     }
@@ -1467,7 +1518,7 @@ struct llama_server_context
 
                     LOG_VERBOSE("context shift", {
                         { "n_ctx", n_ctx },
-                        { "n_keep", params.n_keep },
+                        { "n_keep", n_keep },
                         { "n_left", n_left },
                     });
                 }
@@ -2557,34 +2608,38 @@ int main(int argc, char **argv)
         server_state current_state = state.load();
         switch(current_state) {
             case SERVER_STATE_READY: {
-                int available_slots  = 0;
-                int processing_slots = 0;
-                for (llama_client_slot &slot: llama.slots) {
-                    if (slot.available()) {
-                        available_slots++;
-                    } else {
-                        processing_slots++;
-                    }
+                // request slots data using task queue
+                task_server task;
+                task.id   = llama.queue_tasks.get_new_id();
+                task.type = TASK_TYPE_SLOTS_DATA;
+                task.target_id = -1;
+
+                llama.queue_results.add_waiting_task_id(task.id);
+                llama.queue_tasks.post(task);
+
+                // get the result
+                task_result result = llama.queue_results.recv(task.id);
+                llama.queue_results.remove_waiting_task_id(task.id);
+
+                int n_idle_slots       = result.result_json["idle"];
+                int n_processing_slots = result.result_json["processing"];
+
+                json health = {
+                        {"status",           "ok"},
+                        {"slots_idle",       n_idle_slots},
+                        {"slots_processing", n_processing_slots}};
+                res.status = 200; // HTTP OK
+                if (sparams.slots_endpoint && req.has_param("include_slots")) {
+                    health["slots"] = result.result_json["slots"];
                 }
-                if (available_slots > 0) {
-                    json health = {
-                            {"status",           "ok"},
-                            {"slots_idle",       available_slots},
-                            {"slots_processing", processing_slots}};
-                    res.set_content(health.dump(), "application/json");
-                    res.status = 200; // HTTP OK
-                } else {
-                    json health = {
-                            {"status",           "no slot available"},
-                            {"slots_idle",       available_slots},
-                            {"slots_processing", processing_slots}};
-                    res.set_content(health.dump(), "application/json");
+
+                if (n_idle_slots == 0) {
+                    health["status"] = "no slot available";
                     if (req.has_param("fail_on_no_slot")) {
                         res.status = 503; // HTTP Service Unavailable
-                    } else {
-                        res.status = 200; // HTTP OK
                     }
                 }
+                res.set_content(health.dump(), "application/json");
                 break;
             }
             case SERVER_STATE_LOADING_MODEL:
@@ -2600,26 +2655,20 @@ int main(int argc, char **argv)
 
     if (sparams.slots_endpoint) {
         svr.Get("/slots", [&](const httplib::Request&, httplib::Response& res) {
-            json slots;
-            for (llama_client_slot & slot : llama.slots) {
-                json slot_data = llama.get_formated_generation(slot);
-                slot_data["id"] = slot.id;
-                slot_data["task_id"] = slot.task_id;
-                slot_data["state"] = slot.state;
-                slot_data["prompt"] = slot.prompt;
-                slot_data["next_token"] = {
-                        {"has_next_token", slot.has_next_token},
-                        {"n_remain", slot.n_remaining},
-                        {"num_tokens_predicted", slot.n_decoded},
-                        {"stopped_eos", slot.stopped_eos},
-                        {"stopped_word", slot.stopped_word},
-                        {"stopped_limit", slot.stopped_limit},
-                        {"stopping_word", slot.stopping_word},
-                };
+            // request slots data using task queue
+            task_server task;
+            task.id = llama.queue_tasks.get_new_id();
+            task.type = TASK_TYPE_SLOTS_DATA;
+            task.target_id = -1;
 
-                slots.push_back(slot_data);
-            }
-            res.set_content(slots.dump(), "application/json");
+            llama.queue_results.add_waiting_task_id(task.id);
+            llama.queue_tasks.post(task);
+
+            // get the result
+            task_result result = llama.queue_results.recv(task.id);
+            llama.queue_results.remove_waiting_task_id(task.id);
+
+            res.set_content(result.result_json["slots"].dump(), "application/json");
             res.status = 200; // HTTP OK
         });
     }
@@ -2713,6 +2762,11 @@ int main(int argc, char **argv)
         LOG_INFO("model loaded", {});
     }
 
+    if (sparams.chat_template.empty()) { // custom chat template is not supplied
+        // check if the template comes with the model is supported by us
+        llama.validate_model_chat_template(sparams);
+    }
+
     // Middleware for API key validation
     auto validate_api_key = [&sparams](const httplib::Request &req, httplib::Response &res) -> bool {
         // If API key is not set, skip validation

examples/server/utils.hpp

@@ -49,7 +49,8 @@ enum server_state {
 enum task_type {
     TASK_TYPE_COMPLETION,
     TASK_TYPE_CANCEL,
-    TASK_TYPE_NEXT_RESPONSE
+    TASK_TYPE_NEXT_RESPONSE,
+    TASK_TYPE_SLOTS_DATA
 };
 
 struct task_server {

ggml-cuda.cu

@@ -1,3 +1,7 @@
+#include "ggml-cuda.h"
+#include "ggml.h"
+#include "ggml-backend-impl.h"
+
 #include <algorithm>
 #include <assert.h>
 #include <atomic>
@@ -121,11 +125,6 @@
 
 #endif // defined(GGML_USE_HIPBLAS)
 
-// ggml-cuda need half type so keep ggml headers include at last
-#include "ggml-cuda.h"
-#include "ggml.h"
-#include "ggml-backend-impl.h"
-
 #define CUDART_HMAX     11070 // CUDA 11.7, min. ver. for which __hmax and __hmax2 are known to work (may be higher than needed)
 
 #define CC_PASCAL     600

ggml-impl.h

@@ -53,11 +53,23 @@ extern "C" {
 //
 #include <arm_neon.h>
 
-#define GGML_COMPUTE_FP16_TO_FP32(x) ((float) (x))
-#define GGML_COMPUTE_FP32_TO_FP16(x) (x)
+#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
+#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
 
-#define GGML_FP16_TO_FP32(x) ((float) (x))
-#define GGML_FP32_TO_FP16(x) (x)
+#define GGML_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
+
+static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
+    __fp16 tmp;
+    memcpy(&tmp, &h, sizeof(ggml_fp16_t));
+    return (float)tmp;
+}
+
+static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
+    ggml_fp16_t res;
+    __fp16 tmp = f;
+    memcpy(&res, &tmp, sizeof(ggml_fp16_t));
+    return res;
+}
 
 #else
 
@@ -214,8 +226,7 @@ extern float ggml_table_f32_f16[1 << 16];
 // On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
 // so we define GGML_FP16_TO_FP32 and GGML_FP32_TO_FP16 elsewhere for NEON.
 // This is also true for POWER9.
-#if !defined(GGML_FP16_TO_FP32) || !defined(GGML_FP32_TO_FP16)
-
+#if !defined(GGML_FP16_TO_FP32)
 inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
     uint16_t s;
     memcpy(&s, &f, sizeof(uint16_t));
@@ -223,8 +234,10 @@ inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
 }
 
 #define GGML_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x)
-#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
+#endif
 
+#if !defined(GGML_FP32_TO_FP16)
+#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
 #endif
 
 #define GGML_HASHTABLE_FULL ((size_t)-1)

ggml-quants.c

@@ -438,6 +438,30 @@ inline static ggml_int8x16x4_t ggml_vld1q_s8_x4(const int8_t * ptr) {
     return res;
 }
 
+// NOTE: not tested
+inline static int8x16_t ggml_vqtbl1q_s8(int8x16_t a, uint8x16_t b) {
+    int8x16_t res;
+
+    res[ 0] = a[b[ 0]];
+    res[ 1] = a[b[ 1]];
+    res[ 2] = a[b[ 2]];
+    res[ 3] = a[b[ 3]];
+    res[ 4] = a[b[ 4]];
+    res[ 5] = a[b[ 5]];
+    res[ 6] = a[b[ 6]];
+    res[ 7] = a[b[ 7]];
+    res[ 8] = a[b[ 8]];
+    res[ 9] = a[b[ 9]];
+    res[10] = a[b[10]];
+    res[11] = a[b[11]];
+    res[12] = a[b[12]];
+    res[13] = a[b[13]];
+    res[14] = a[b[14]];
+    res[15] = a[b[15]];
+
+    return res;
+}
+
 #else
 
 #define ggml_int16x8x2_t  int16x8x2_t
@@ -451,6 +475,7 @@ inline static ggml_int8x16x4_t ggml_vld1q_s8_x4(const int8_t * ptr) {
 #define ggml_vld1q_u8_x4  vld1q_u8_x4
 #define ggml_vld1q_s8_x2  vld1q_s8_x2
 #define ggml_vld1q_s8_x4  vld1q_s8_x4
+#define ggml_vqtbl1q_s8   vqtbl1q_s8
 
 #endif
 
@@ -5629,8 +5654,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 
     for (int i = 0; i < nb; ++i) {
 
-        const float d = y[i].d * (float)x[i].d;
-        const float dmin = -y[i].d * (float)x[i].dmin;
+        const float d    =  y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
 
         const uint8_t * restrict q2 = x[i].qs;
         const int8_t  * restrict q8 = y[i].qs;
@@ -5779,8 +5804,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 
     for (int i = 0; i < nb; ++i) {
 
-        const float d = y[i].d * (float)x[i].d;
-        const float dmin = -y[i].d * (float)x[i].dmin;
+        const float d    =  y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
 
         const uint8_t * restrict q2 = x[i].qs;
         const int8_t  * restrict q8 = y[i].qs;
@@ -6433,7 +6458,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 
         int32_t isum = -4*(scales[0] * y[i].bsums[0] + scales[2] * y[i].bsums[1] + scales[1] * y[i].bsums[2] + scales[3] * y[i].bsums[3]);
 
-        const float d = y[i].d * (float)x[i].d;
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
 
         const uint8x16_t htmp = vcombine_u8(hbits, vshr_n_u8(hbits, 1));
         q3h.val[0] = vandq_u8(mh, vshlq_n_u8(htmp, 2));
@@ -6635,7 +6660,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 
         int32_t isum = -4*(scales[0] * y[i].bsums[0] + scales[2] * y[i].bsums[1] + scales[1] * y[i].bsums[2] + scales[3] * y[i].bsums[3]);
 
-        const float d = y[i].d * (float)x[i].d;
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
 
         vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1);
 
@@ -7138,9 +7163,9 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r
         aux16[1] = (a[0] >> 4) & 0x0f0f;
 
         const int32_t summi = scales[2] * (y[i].bsums[0] + y[i].bsums[1]) + scales[3] * (y[i].bsums[2] + y[i].bsums[3]);
-        sum_mins += y[i].d * (float)x[i].d[1] * summi;
+        sum_mins += y[i].d * GGML_FP16_TO_FP32(x[i].d[1]) * summi;
 
-        const float d = y[i].d * (float)x[i].d[0];
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d[0]);
 
         const ggml_uint8x16x2_t q4bits = ggml_vld1q_u8_x2(q4);
 
@@ -7798,7 +7823,7 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 
     for (int i = 0; i < nb; ++i) {
 
-        const float d = y[i].d * (float)x[i].d;
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
         const int8_t * sc = x[i].scales;
 
         const uint8_t * restrict q5 = x[i].qs;
@@ -7940,7 +7965,7 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 
     for (int i = 0; i < nb; ++i) {
 
-        const float d = y[i].d * (float)x[i].d;
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
         const int8_t * sc = x[i].scales;
 
         const uint8_t * restrict q5 = x[i].qs;
@@ -8508,7 +8533,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 
     for (int i = 0; i < nb; ++i) {
 
-        const float d_all = (float)x[i].d;
+        const float d_all = GGML_FP16_TO_FP32(x[i].d);
 
         const uint8_t * restrict q6 = x[i].ql;
         const uint8_t * restrict qh = x[i].qh;
@@ -8679,7 +8704,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 
     for (int i = 0; i < nb; ++i) {
 
-        const float d_all = (float)x[i].d;
+        const float d_all = GGML_FP16_TO_FP32(x[i].d);
 
         const uint8_t * restrict q6 = x[i].ql;
         const uint8_t * restrict qh = x[i].qh;
@@ -9333,7 +9358,7 @@ void ggml_vec_dot_iq1_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const
     uint16_t gindex[8];
     uint16x8x2_t vindex;
     int8x16x4_t q1b;
-    int8x16x4_t q8b;
+    ggml_int8x16x4_t q8b;
     uint16x8x4_t scales;
     int32x4x2_t sumi;
     int32x4x2_t dotq;
@@ -9498,7 +9523,6 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * restrict s, size_t bs, const void *
     float sumf = 0;
 
     for (int ib = 0; ib < nb; ib += 2) {
-
         q4bits.val[0] = vld1q_u8(x[ib+0].qs);
         q4bits.val[1] = vld1q_u8(x[ib+1].qs);
         q8b.val[0]    = vld1q_s8(y[ib+0].qs);
@@ -9506,16 +9530,17 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * restrict s, size_t bs, const void *
         q8b.val[2]    = vld1q_s8(y[ib+1].qs);
         q8b.val[3]    = vld1q_s8(y[ib+1].qs + 16);
 
-        q4b.val[0] = vqtbl1q_s8(values, vandq_u8(q4bits.val[0], m4b));
-        q4b.val[1] = vqtbl1q_s8(values, vshrq_n_u8(q4bits.val[0], 4));
-        q4b.val[2] = vqtbl1q_s8(values, vandq_u8(q4bits.val[1], m4b));
-        q4b.val[3] = vqtbl1q_s8(values, vshrq_n_u8(q4bits.val[1], 4));
+        q4b.val[0] = ggml_vqtbl1q_s8(values, vandq_u8  (q4bits.val[0], m4b));
+        q4b.val[1] = ggml_vqtbl1q_s8(values, vshrq_n_u8(q4bits.val[0], 4));
+        q4b.val[2] = ggml_vqtbl1q_s8(values, vandq_u8  (q4bits.val[1], m4b));
+        q4b.val[3] = ggml_vqtbl1q_s8(values, vshrq_n_u8(q4bits.val[1], 4));
 
         prod_1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q4b.val[0], q8b.val[0]), q4b.val[1], q8b.val[1]);
         prod_2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q4b.val[2], q8b.val[2]), q4b.val[3], q8b.val[3]);
 
-        sumf += (float)x[ib+0].d * (float)y[ib+0].d * vaddvq_s32(prod_1) + (float)x[ib+1].d * (float)y[ib+1].d * vaddvq_s32(prod_2);
-
+        sumf +=
+            GGML_FP16_TO_FP32(x[ib+0].d) * GGML_FP16_TO_FP32(y[ib+0].d) * vaddvq_s32(prod_1) +
+            GGML_FP16_TO_FP32(x[ib+1].d) * GGML_FP16_TO_FP32(y[ib+1].d) * vaddvq_s32(prod_2);
     }
 
     *s = sumf;

ggml.h

@@ -315,13 +315,7 @@
 extern "C" {
 #endif
 
-#if defined(__ARM_NEON) && defined(__CUDACC__)
-    typedef half ggml_fp16_t;
-#elif defined(__ARM_NEON) && !defined(_MSC_VER)
-    typedef __fp16 ggml_fp16_t;
-#else
     typedef uint16_t ggml_fp16_t;
-#endif
 
     // convert FP16 <-> FP32
     GGML_API float       ggml_fp16_to_fp32(ggml_fp16_t x);

llama.cpp

@@ -2791,13 +2791,7 @@ struct llama_model_loader {
 
         std::vector<no_init<uint8_t>> read_buf;
 
-        for (int i = 0; i < gguf_get_n_tensors(ctx_gguf); i++) {
-            struct ggml_tensor * cur = ggml_get_tensor(ctx, gguf_get_tensor_name(ctx_gguf, i));
-            if (!cur) {
-                // some tensors may be allocated in a different context
-                continue;
-            }
-
+        for (struct ggml_tensor * cur = ggml_get_first_tensor(ctx); cur != NULL; cur = ggml_get_next_tensor(ctx, cur)) {
             if (progress_callback) {
                 if (!progress_callback((float) size_done / size_data, progress_callback_user_data)) {
                     return false;
@@ -3722,7 +3716,7 @@ static bool llm_load_tensors(
     }
 
     // create one context per buffer type
-    size_t ctx_size = ggml_tensor_overhead()*ml.n_tensors;
+    size_t ctx_size = ggml_tensor_overhead()*(ml.n_tensors + 1); // +1 for models where tok_embd is duplicated as output
     std::map<ggml_backend_buffer_type_t, ggml_context *> ctx_map;
     for (auto & it : buft_layer_count) {
         struct ggml_init_params params = {
@@ -3860,6 +3854,7 @@ static bool llm_load_tensors(
                         } else {
                             model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); // needs to be on GPU
                             ml.n_created--; // artificial tensor
+                            ml.size_data += ggml_nbytes(model.output);
                         }
                     }
 
@@ -4059,6 +4054,8 @@ static bool llm_load_tensors(
                     // output
                     {
                         model.output_norm   = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                        model.output_norm_b = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd}, false);
+
                         model.output        = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab});
                     }
 
@@ -4068,14 +4065,23 @@ static bool llm_load_tensors(
 
                         auto & layer = model.layers[i];
 
-                        layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+                        layer.attn_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+                        layer.attn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "bias", i),   {n_embd}, false);
 
                         layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa});
-                        layer.wo   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
+                        layer.bqkv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd + 2*n_embd_gqa}, false);
 
-                        layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
-                        layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd});
-                        layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
+                        layer.wo   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
+                        layer.bo   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT, "bias", i),   {n_embd}, false);
+
+                        layer.ffn_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
+                        layer.ffn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "bias", i),   {n_embd}, false);
+
+                        layer.ffn_down   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd});
+                        layer.ffn_down_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd}, false);
+
+                        layer.ffn_up     = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
+                        layer.ffn_up_b   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP,   "bias", i),   {n_ff}, false);
 
                         // AWQ ScaleActivation layer
                         layer.ffn_act = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_ACT, "scales", i), {n_ff}, false);
@@ -4394,6 +4400,9 @@ static bool llm_load_tensors(
 
                     // output
                     model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                    model.output      = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}); // same as tok_embd, duplicated to allow offloading
+                    ml.n_created--; // artificial tensor
+                    ml.size_data += ggml_nbytes(model.output);
 
                     const int64_t n_ff          = hparams.n_ff;
                     const int64_t n_embd_head_k = hparams.n_embd_head_k;
@@ -6173,7 +6182,7 @@ struct llm_build_context {
 
             attn_norm = llm_build_norm(ctx0, inpL, hparams,
                     model.layers[il].attn_norm,
-                    NULL,
+                    model.layers[il].attn_norm_b,
                     LLM_NORM, cb, il);
             cb(attn_norm, "attn_norm", il);
 
@@ -6184,6 +6193,11 @@ struct llm_build_context {
                 cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, cur);
                 cb(cur, "wqkv", il);
 
+                if (model.layers[il].bqkv){
+                    cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+                    cb(cur, "bqkv", il);
+                }
+
                 if (hparams.f_clamp_kqv > 0.0f) {
                     cur = ggml_clamp(ctx0, cur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv);
                     cb(cur, "wqkv_clamped", il);
@@ -6200,7 +6214,7 @@ struct llm_build_context {
                 Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
 
                 cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
-                        model.layers[il].wo, NULL,
+                        model.layers[il].wo, model.layers[il].bo,
                         Kcur, Vcur, Qcur, KQ_mask, KQ_pos, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                 cb(cur, "kqv_out", il);
             }
@@ -6213,13 +6227,13 @@ struct llm_build_context {
             {
                 cur = llm_build_norm(ctx0, ffn_inp, hparams,
                         model.layers[il].ffn_norm,
-                        NULL,
+                        model.layers[il].ffn_norm_b,
                         LLM_NORM, cb, il);
                 cb(cur, "ffn_norm", il);
                 cur = llm_build_ffn(ctx0, cur,
-                        model.layers[il].ffn_up,   NULL,
+                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,
                         NULL,                      NULL,
-                        model.layers[il].ffn_down, NULL,
+                        model.layers[il].ffn_down, model.layers[il].ffn_down_b,
                         model.layers[il].ffn_act,
                         LLM_FFN_GELU, LLM_FFN_SEQ, cb, il);
                 cb(cur, "ffn_out", il);
@@ -6236,7 +6250,7 @@ struct llm_build_context {
 
         cur = llm_build_norm(ctx0, cur, hparams,
                 model.output_norm,
-                NULL,
+                model.output_norm_b,
                 LLM_NORM, cb, -1);
         cb(cur, "result_norm", -1);
 
@@ -7436,6 +7450,7 @@ struct llm_build_context {
 
         inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb);
         cb(inpL, "inp_embd", -1);
+
         inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
         cb(inpL, "inp_scaled", -1);
 
@@ -7477,6 +7492,7 @@ struct llm_build_context {
                         n_embd_head_k, 2, 0, n_orig_ctx, freq_base, freq_scale,
                         ext_factor, attn_factor, beta_fast, beta_slow);
                 cb(Qcur, "Qcur", il);
+
                 Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head_k)));
                 cb(Qcur, "Qcur_scaled", il);
 
@@ -7491,6 +7507,7 @@ struct llm_build_context {
                         Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f, cb, il);
                 cb(cur, "kqv_out", il);
             }
+
             struct ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL);
             cb(sa_out, "sa_out", il);
 
@@ -7525,7 +7542,7 @@ struct llm_build_context {
         cb(cur, "result_norm", -1);
 
         // lm_head
-        cur = ggml_mul_mat(ctx0, model.tok_embd, cur);
+        cur = ggml_mul_mat(ctx0, model.output, cur);
         cb(cur, "result_output", -1);
 
         ggml_build_forward_expand(gf, cur);
@@ -12174,18 +12191,19 @@ static void llama_copy_state_data_internal(struct llama_context * ctx, llama_dat
         data_ctx->write(&kv_used,     sizeof(kv_used));
 
         if (kv_buf_size) {
-            const size_t elt_size = ggml_element_size(kv_self.k_l[0]);
-
             std::vector<uint8_t> tmp_buf;
             for (int il = 0; il < (int) n_layer; ++il) {
-                tmp_buf.resize(elt_size*n_embd_k_gqa*kv_head);
+                size_t k_size = ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa*kv_head);
+                tmp_buf.resize(k_size);
                 ggml_backend_tensor_get(kv_self.k_l[il], tmp_buf.data(), 0, tmp_buf.size());
                 data_ctx->write(tmp_buf.data(), tmp_buf.size());
 
                 // v is not contiguous, copy row by row
-                tmp_buf.resize(elt_size*kv_head);
+                size_t v_row_size = ggml_row_size(kv_self.v_l[il]->type, kv_head);
+                size_t v_row_stride = ggml_row_size(kv_self.v_l[il]->type, n_ctx);
+                tmp_buf.resize(v_row_size);
                 for (int ir = 0; ir < (int) n_embd_v_gqa; ++ir) {
-                    ggml_backend_tensor_get(kv_self.v_l[il], tmp_buf.data(), ir*elt_size*n_ctx, tmp_buf.size());
+                    ggml_backend_tensor_get(kv_self.v_l[il], tmp_buf.data(), ir*v_row_stride, tmp_buf.size());
                     data_ctx->write(tmp_buf.data(), tmp_buf.size());
                 }
             }
@@ -12287,17 +12305,16 @@ size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src) {
         if (kv_buf_size) {
             GGML_ASSERT(kv_self.total_size() == kv_buf_size);
 
-            const size_t elt_size = ggml_element_size(kv_self.k_l[0]);
-
             for (int il = 0; il < (int) n_layer; ++il) {
-                size_t k_size = elt_size*n_embd_k_gqa*kv_head;
+                size_t k_size = ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa*kv_head);
                 ggml_backend_tensor_set(kv_self.k_l[il], inp, 0, k_size);
                 inp += k_size;
 
                 // v is not contiguous, copy row by row
-                size_t v_row_size = elt_size*kv_head;
+                size_t v_row_size = ggml_row_size(kv_self.v_l[il]->type, kv_head);
+                size_t v_row_stride = ggml_row_size(kv_self.v_l[il]->type, n_ctx);
                 for (int ir = 0; ir < (int) n_embd_v_gqa; ++ir) {
-                    ggml_backend_tensor_set(kv_self.v_l[il], inp, ir*elt_size*n_ctx, v_row_size);
+                    ggml_backend_tensor_set(kv_self.v_l[il], inp, ir*v_row_stride, v_row_size);
                     inp += v_row_size;
                 }
             }
@@ -12759,6 +12776,37 @@ static int32_t llama_chat_apply_template_internal(
         if (add_ass) {
             ss << "<|assistant|>\n";
         }
+    } else if (tmpl.find("bos_token + message['role']") != std::string::npos) {
+        // mlabonne/AlphaMonarch-7B template (the <s> is included inside history)
+        for (auto message : chat) {
+            std::string bos = (message == chat.front()) ? "" : "<s>"; // skip BOS for first message
+            ss << bos << message->role << "\n" << message->content << "</s>\n";
+        }
+        if (add_ass) {
+            ss << "<s>assistant\n";
+        }
+    } else if (tmpl.find("<start_of_turn>") != std::string::npos) {
+        // google/gemma-7b-it
+        std::string system_prompt = "";
+        for (auto message : chat) {
+            std::string role(message->role);
+            if (role == "system") {
+                // there is no system message for gemma, but we will merge it with user prompt, so nothing is broken
+                system_prompt = trim(message->content);
+                continue;
+            }
+            // in gemma, "assistant" is "model"
+            role = role == "assistant" ? "model" : message->role;
+            ss << "<start_of_turn>" << role << "\n";
+            if (!system_prompt.empty() && role != "model") {
+                ss << system_prompt << "\n\n";
+                system_prompt = "";
+            }
+            ss << trim(message->content) << "<end_of_turn>\n";
+        }
+        if (add_ass) {
+            ss << "<start_of_turn>model\n";
+        }
     } else {
         // template not supported
         return -1;

llama.h

@@ -708,7 +708,7 @@ extern "C" {
 
     /// Apply chat template. Inspired by hf apply_chat_template() on python.
     /// Both "model" and "custom_template" are optional, but at least one is required. "custom_template" has higher precedence than "model"
-    /// NOTE: This function only support some known jinja templates. It is not a jinja parser.
+    /// NOTE: This function does not use a jinja parser. It only support a pre-defined list of template. See more: https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template
     /// @param tmpl A Jinja template to use for this chat. If this is nullptr, the model’s default chat template will be used instead.
     /// @param chat Pointer to a list of multiple llama_chat_message
     /// @param n_msg Number of llama_chat_message in this chat

scripts/sync-ggml.last

@@ -1 +1 @@
-818eeb8a3be99125746a90ec63af8f51516a2ec6
+8cdf783f288a98eddf521b0ab1b4d405be9e18ba

tests/test-chat-template.cpp

@@ -27,12 +27,24 @@ int main(void) {
         "{%- for idx in range(0, messages|length) -%}\\n{%- if messages[idx]['role'] == 'user' -%}\\n{%- if idx > 1 -%}\\n{{- bos_token + '[INST] ' + messages[idx]['content'] + ' [/INST]' -}}\\n{%- else -%}\\n{{- messages[idx]['content'] + ' [/INST]' -}}\\n{%- endif -%}\\n{% elif messages[idx]['role'] == 'system' %}\\n{{- '[INST] <<SYS>>\\\\n' + messages[idx]['content'] + '\\\\n<</SYS>>\\\\n\\\\n' -}}\\n{%- elif messages[idx]['role'] == 'assistant' -%}\\n{{- ' '  + messages[idx]['content'] + ' ' + eos_token -}}\\n{% endif %}\\n{% endfor %}",
         // bofenghuang/vigogne-2-70b-chat
         "{{ bos_token }}{% if messages[0]['role'] == 'system' %}{% set loop_messages = messages[1:] %}{% set system_message = messages[0]['content'] %}{% elif true == true and not '<<SYS>>' in messages[0]['content'] %}{% set loop_messages = messages %}{% set system_message = 'Vous êtes Vigogne, un assistant IA créé par Zaion Lab. Vous suivez extrêmement bien les instructions. Aidez autant que vous le pouvez.' %}{% else %}{% set loop_messages = messages %}{% set system_message = false %}{% endif %}{% for message in loop_messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if loop.index0 == 0 and system_message != false %}{% set content = '<<SYS>>\\\\n' + system_message + '\\\\n<</SYS>>\\\\n\\\\n' + message['content'] %}{% else %}{% set content = message['content'] %}{% endif %}{% if message['role'] == 'user' %}{{ '[INST] ' + content.strip() + ' [/INST]' }}{% elif message['role'] == 'system' %}{{ '<<SYS>>\\\\n' + content.strip() + '\\\\n<</SYS>>\\\\n\\\\n' }}{% elif message['role'] == 'assistant' %}{{ ' '  + content.strip() + ' ' + eos_token }}{% endif %}{% endfor %}",
+        // mlabonne/AlphaMonarch-7B
+        "{% for message in messages %}{{bos_token + message['role'] + '\\n' + message['content'] + eos_token + '\\n'}}{% endfor %}{% if add_generation_prompt %}{{ bos_token + 'assistant\\n' }}{% endif %}",
+        // google/gemma-7b-it
+        "{% if messages[0]['role'] == 'system' %}{{ raise_exception('System role not supported') }}{% endif %}{% for message in messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if (message['role'] == 'assistant') %}{% set role = 'model' %}{% else %}{% set role = message['role'] %}{% endif %}{{ '<start_of_turn>' + role + '\\n' + message['content'] | trim + '<end_of_turn>\\n' }}{% endfor %}{% if add_generation_prompt %}{{'<start_of_turn>model\\n'}}{% endif %}",
     };
-    std::vector<std::string> expected_substr = {
-        "<|im_start|>assistant\n   I am an assistant   <|im_end|>\n<|im_start|>user\nAnother question<|im_end|>\n<|im_start|>assistant",
-        "[/INST]Hi there</s>[INST] Who are you [/INST]   I am an assistant   </s>[INST] Another question [/INST]",
-        "</s><s>[INST] Who are you [/INST]    I am an assistant    </s><s>[INST] Another question [/INST]",
-        "[/INST] Hi there </s>[INST] Who are you [/INST] I am an assistant </s>[INST] Another question [/INST]",
+    std::vector<std::string> expected_output = {
+        // teknium/OpenHermes-2.5-Mistral-7B
+        "<|im_start|>system\nYou are a helpful assistant<|im_end|>\n<|im_start|>user\nHello<|im_end|>\n<|im_start|>assistant\nHi there<|im_end|>\n<|im_start|>user\nWho are you<|im_end|>\n<|im_start|>assistant\n   I am an assistant   <|im_end|>\n<|im_start|>user\nAnother question<|im_end|>\n<|im_start|>assistant\n",
+        // mistralai/Mistral-7B-Instruct-v0.2
+        "[INST] You are a helpful assistant\nHello [/INST]Hi there</s>[INST] Who are you [/INST]   I am an assistant   </s>[INST] Another question [/INST]",
+        // TheBloke/FusionNet_34Bx2_MoE-AWQ
+        "[INST] <<SYS>>\nYou are a helpful assistant\n<</SYS>>\n\nHello [/INST] Hi there </s><s>[INST] Who are you [/INST]    I am an assistant    </s><s>[INST] Another question [/INST]",
+        // bofenghuang/vigogne-2-70b-chat
+        "[INST] <<SYS>>\nYou are a helpful assistant\n<</SYS>>\n\nHello [/INST] Hi there </s>[INST] Who are you [/INST] I am an assistant </s>[INST] Another question [/INST]",
+        // mlabonne/AlphaMonarch-7B
+        "system\nYou are a helpful assistant</s>\n<s>user\nHello</s>\n<s>assistant\nHi there</s>\n<s>user\nWho are you</s>\n<s>assistant\n   I am an assistant   </s>\n<s>user\nAnother question</s>\n<s>assistant\n",
+        // google/gemma-7b-it
+        "<start_of_turn>user\nYou are a helpful assistant\n\nHello<end_of_turn>\n<start_of_turn>model\nHi there<end_of_turn>\n<start_of_turn>user\nWho are you<end_of_turn>\n<start_of_turn>model\nI am an assistant<end_of_turn>\n<start_of_turn>user\nAnother question<end_of_turn>\n<start_of_turn>model\n",
     };
     std::vector<char> formatted_chat(1024);
     int32_t res;
@@ -43,7 +55,7 @@ int main(void) {
 
     for (size_t i = 0; i < templates.size(); i++) {
         std::string custom_template = templates[i];
-        std::string substr = expected_substr[i];
+        std::string expected = expected_output[i];
         formatted_chat.resize(1024);
         res = llama_chat_apply_template(
             nullptr,
@@ -57,8 +69,7 @@ int main(void) {
         formatted_chat.resize(res);
         std::string output(formatted_chat.data(), formatted_chat.size());
         std::cout << output << "\n-------------------------\n";
-        // expect the "formatted_chat" to contain pre-defined strings
-        assert(output.find(substr) != std::string::npos);
+        assert(output == expected);
     }
     return 0;
 }