Merge branch 'master' into metal-memory-reduction

.devops/server-cuda.Dockerfile

@@ -0,0 +1,32 @@
+ARG UBUNTU_VERSION=22.04
+# This needs to generally match the container host's environment.
+ARG CUDA_VERSION=11.7.1
+# Target the CUDA build image
+ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
+# Target the CUDA runtime image
+ARG BASE_CUDA_RUN_CONTAINER=nvidia/cuda:${CUDA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}
+
+FROM ${BASE_CUDA_DEV_CONTAINER} as build
+
+# Unless otherwise specified, we make a fat build.
+ARG CUDA_DOCKER_ARCH=all
+
+RUN apt-get update && \
+    apt-get install -y build-essential git
+
+WORKDIR /app
+
+COPY . .
+
+# Set nvcc architecture
+ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
+# Enable cuBLAS
+ENV LLAMA_CUBLAS=1
+
+RUN make
+
+FROM ${BASE_CUDA_RUN_CONTAINER} as runtime
+
+COPY --from=build /app/server /server
+
+ENTRYPOINT [ "/server" ]

.devops/server-intel.Dockerfile

@@ -0,0 +1,25 @@
+ARG ONEAPI_VERSION=2024.0.1-devel-ubuntu22.04
+ARG UBUNTU_VERSION=22.04
+
+FROM intel/hpckit:$ONEAPI_VERSION as build
+
+RUN apt-get update && \
+    apt-get install -y git
+
+WORKDIR /app
+
+COPY . .
+
+# for some reasons, "-DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=Intel10_64lp -DLLAMA_NATIVE=ON" give worse performance
+RUN mkdir build && \
+    cd build && \
+    cmake .. -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx && \
+    cmake --build . --config Release --target main server
+
+FROM ubuntu:$UBUNTU_VERSION as runtime
+
+COPY --from=build /app/build/bin/server /server
+
+ENV LC_ALL=C.utf8
+
+ENTRYPOINT [ "/server" ]

.devops/server-rocm.Dockerfile

@@ -0,0 +1,45 @@
+ARG UBUNTU_VERSION=22.04
+
+# This needs to generally match the container host's environment.
+ARG ROCM_VERSION=5.6
+
+# Target the CUDA build image
+ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete
+
+FROM ${BASE_ROCM_DEV_CONTAINER} as build
+
+# Unless otherwise specified, we make a fat build.
+# List from https://github.com/ggerganov/llama.cpp/pull/1087#issuecomment-1682807878
+# This is mostly tied to rocBLAS supported archs.
+ARG ROCM_DOCKER_ARCH=\
+    gfx803 \
+    gfx900 \
+    gfx906 \
+    gfx908 \
+    gfx90a \
+    gfx1010 \
+    gfx1030 \
+    gfx1100 \
+    gfx1101 \
+    gfx1102
+
+COPY requirements.txt   requirements.txt
+COPY requirements       requirements
+
+RUN pip install --upgrade pip setuptools wheel \
+    && pip install -r requirements.txt
+
+WORKDIR /app
+
+COPY . .
+
+# Set nvcc architecture
+ENV GPU_TARGETS=${ROCM_DOCKER_ARCH}
+# Enable ROCm
+ENV LLAMA_HIPBLAS=1
+ENV CC=/opt/rocm/llvm/bin/clang
+ENV CXX=/opt/rocm/llvm/bin/clang++
+
+RUN make
+
+ENTRYPOINT [ "/app/server" ]

.devops/server.Dockerfile

@@ -0,0 +1,20 @@
+ARG UBUNTU_VERSION=22.04
+
+FROM ubuntu:$UBUNTU_VERSION as build
+
+RUN apt-get update && \
+    apt-get install -y build-essential git
+
+WORKDIR /app
+
+COPY . .
+
+RUN make
+
+FROM ubuntu:$UBUNTU_VERSION as runtime
+
+COPY --from=build /app/server /server
+
+ENV LC_ALL=C.utf8
+
+ENTRYPOINT [ "/server" ]

.github/workflows/docker.yml

@@ -28,14 +28,18 @@ jobs:
         config:
           - { tag: "light", dockerfile: ".devops/main.Dockerfile", platforms: "linux/amd64,linux/arm64" }
           - { tag: "full", dockerfile: ".devops/full.Dockerfile", platforms: "linux/amd64,linux/arm64" }
+          - { tag: "server", dockerfile: ".devops/server.Dockerfile", platforms: "linux/amd64,linux/arm64" }
           # NOTE(canardletter): The CUDA builds on arm64 are very slow, so I
           #                     have disabled them for now until the reason why
           #                     is understood.
           - { tag: "light-cuda", dockerfile: ".devops/main-cuda.Dockerfile", platforms: "linux/amd64" }
           - { tag: "full-cuda", dockerfile: ".devops/full-cuda.Dockerfile", platforms: "linux/amd64" }
+          - { tag: "server-cuda", dockerfile: ".devops/server-cuda.Dockerfile", platforms: "linux/amd64" }
           - { tag: "light-rocm", dockerfile: ".devops/main-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
           - { tag: "full-rocm", dockerfile: ".devops/full-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
+          - { tag: "server-rocm", dockerfile: ".devops/server-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
           - { tag: "light-intel", dockerfile: ".devops/main-intel.Dockerfile", platforms: "linux/amd64" }
+          - { tag: "server-intel", dockerfile: ".devops/server-intel.Dockerfile", platforms: "linux/amd64" }
     steps:
       - name: Check out the repo
         uses: actions/checkout@v3

README.md

@@ -122,7 +122,8 @@ as the main playground for developing new features for the [ggml](https://github
 - Node.js: [withcatai/node-llama-cpp](https://github.com/withcatai/node-llama-cpp)
 - JS/TS (llama.cpp server client): [lgrammel/modelfusion](https://modelfusion.dev/integration/model-provider/llamacpp)
 - Ruby: [yoshoku/llama_cpp.rb](https://github.com/yoshoku/llama_cpp.rb)
-- Rust: [mdrokz/rust-llama.cpp](https://github.com/mdrokz/rust-llama.cpp)
+- Rust (nicer API): [mdrokz/rust-llama.cpp](https://github.com/mdrokz/rust-llama.cpp)
+- Rust (more direct bindings): [utilityai/llama-cpp-rs](https://github.com/utilityai/llama-cpp-rs)
 - C#/.NET: [SciSharp/LLamaSharp](https://github.com/SciSharp/LLamaSharp)
 - Scala 3: [donderom/llm4s](https://github.com/donderom/llm4s)
 - Clojure: [phronmophobic/llama.clj](https://github.com/phronmophobic/llama.clj)
@@ -931,17 +932,20 @@ Place your desired model into the `~/llama.cpp/models/` directory and execute th
 * Create a folder to store big models & intermediate files (ex. /llama/models)
 
 #### Images
-We have two Docker images available for this project:
+We have three Docker images available for this project:
 
 1. `ghcr.io/ggerganov/llama.cpp:full`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization. (platforms: `linux/amd64`, `linux/arm64`)
 2. `ghcr.io/ggerganov/llama.cpp:light`: This image only includes the main executable file. (platforms: `linux/amd64`, `linux/arm64`)
+3. `ghcr.io/ggerganov/llama.cpp:server`: This image only includes the server executabhle file. (platforms: `linux/amd64`, `linux/arm64`)
 
 Additionally, there the following images, similar to the above:
 
 - `ghcr.io/ggerganov/llama.cpp:full-cuda`: Same as `full` but compiled with CUDA support. (platforms: `linux/amd64`)
 - `ghcr.io/ggerganov/llama.cpp:light-cuda`: Same as `light` but compiled with CUDA support. (platforms: `linux/amd64`)
+- `ghcr.io/ggerganov/llama.cpp:server-cuda`: Same as `server` but compiled with CUDA support. (platforms: `linux/amd64`)
 - `ghcr.io/ggerganov/llama.cpp:full-rocm`: Same as `full` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
 - `ghcr.io/ggerganov/llama.cpp:light-rocm`: Same as `light` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
+- `ghcr.io/ggerganov/llama.cpp:server-rocm`: Same as `server` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
 
 The GPU enabled images are not currently tested by CI beyond being built. They are not built with any variation from the ones in the Dockerfiles defined in [.devops/](.devops/) and the GitHub Action defined in [.github/workflows/docker.yml](.github/workflows/docker.yml). If you need different settings (for example, a different CUDA or ROCm library, you'll need to build the images locally for now).
 
@@ -967,6 +971,12 @@ or with a light image:
 docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:light -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512
 ```
 
+or with a server image:
+
+```bash
+docker run -v /path/to/models:/models -p 8000:8000 ghcr.io/ggerganov/llama.cpp:server -m /models/7B/ggml-model-q4_0.gguf --port 8000 --host 0.0.0.0 -n 512
+```
+
 ### Docker With CUDA
 
 Assuming one has the [nvidia-container-toolkit](https://github.com/NVIDIA/nvidia-container-toolkit) properly installed on Linux, or is using a GPU enabled cloud, `cuBLAS` should be accessible inside the container.
@@ -976,6 +986,7 @@ Assuming one has the [nvidia-container-toolkit](https://github.com/NVIDIA/nvidia
 ```bash
 docker build -t local/llama.cpp:full-cuda -f .devops/full-cuda.Dockerfile .
 docker build -t local/llama.cpp:light-cuda -f .devops/main-cuda.Dockerfile .
+docker build -t local/llama.cpp:server-cuda -f .devops/server-cuda.Dockerfile .
 ```
 
 You may want to pass in some different `ARGS`, depending on the CUDA environment supported by your container host, as well as the GPU architecture.
@@ -989,6 +1000,7 @@ The resulting images, are essentially the same as the non-CUDA images:
 
 1. `local/llama.cpp:full-cuda`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization.
 2. `local/llama.cpp:light-cuda`: This image only includes the main executable file.
+3. `local/llama.cpp:server-cuda`: This image only includes the server executable file.
 
 #### Usage
 
@@ -997,6 +1009,7 @@ After building locally, Usage is similar to the non-CUDA examples, but you'll ne
 ```bash
 docker run --gpus all -v /path/to/models:/models local/llama.cpp:full-cuda --run -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512 --n-gpu-layers 1
 docker run --gpus all -v /path/to/models:/models local/llama.cpp:light-cuda -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512 --n-gpu-layers 1
+docker run --gpus all -v /path/to/models:/models local/llama.cpp:server-cuda -m /models/7B/ggml-model-q4_0.gguf --port 8000 --host 0.0.0.0 -n 512 --n-gpu-layers 1
 ```
 
 ### Contributing

convert-hf-to-gguf.py

@@ -201,6 +201,8 @@ class Model:
             return PlamoModel
         if model_architecture == "CodeShellForCausalLM":
             return CodeShellModel
+        if model_architecture == "OrionForCausalLM":
+            return OrionModel
         return Model
 
     def _is_model_safetensors(self) -> bool:
@@ -250,6 +252,8 @@ class Model:
             return gguf.MODEL_ARCH.PLAMO
         if arch == "CodeShellForCausalLM":
             return gguf.MODEL_ARCH.CODESHELL
+        if arch == "OrionForCausalLM":
+            return gguf.MODEL_ARCH.ORION
 
         raise NotImplementedError(f'Architecture "{arch}" not supported!')
 
@@ -572,6 +576,83 @@ class MPTModel(Model):
                 self.gguf_writer.add_tensor("output.weight", data)
 
 
+class OrionModel(Model):
+    def set_vocab(self):
+        self._set_vocab_sentencepiece()
+
+    def set_gguf_parameters(self):
+        block_count = self.hparams["num_hidden_layers"]
+        head_count = self.hparams["num_attention_heads"]
+        head_count_kv = self.hparams.get("num_key_value_heads", head_count)
+        hf_repo = self.hparams.get("_name_or_path", "")
+
+        ctx_length = 0
+        if "max_sequence_length" in self.hparams:
+            ctx_length = self.hparams["max_sequence_length"]
+        elif "max_position_embeddings" in self.hparams:
+            ctx_length = self.hparams["max_position_embeddings"]
+        elif "model_max_length" in self.hparams:
+            ctx_length = self.hparams["model_max_length"]
+        else:
+            print("gguf: can not find ctx length parameter.")
+            sys.exit()
+
+        self.gguf_writer.add_file_type(self.ftype)
+        self.gguf_writer.add_name(self.dir_model.name)
+        self.gguf_writer.add_source_hf_repo(hf_repo)
+        self.gguf_writer.add_tensor_data_layout("Meta AI original pth")
+        self.gguf_writer.add_context_length(ctx_length)
+        self.gguf_writer.add_embedding_length(self.hparams["hidden_size"])
+        self.gguf_writer.add_block_count(block_count)
+        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)
+        self.gguf_writer.add_layer_norm_eps(self.hparams["rms_norm_eps"])
+
+    def write_tensors(self):
+        # Collect tensors from generator object
+        model_kv = dict(self.get_tensors())
+        block_count = self.hparams["num_hidden_layers"]
+        tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
+
+        for name, data_torch in model_kv.items():
+            # we don't need these
+            if name.endswith(".rotary_emb.inv_freq"):
+                continue
+
+            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
+
+            # if f32 desired, convert any float16 to float32
+            if self.ftype == 0 and data_dtype == np.float16:
+                data = data.astype(np.float32)
+
+            # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
+            if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
+                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"{name} -> {new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
+            self.gguf_writer.add_tensor(new_name, data)
+
+
 class BaichuanModel(Model):
     def set_vocab(self):
         self._set_vocab_sentencepiece()

examples/server/README.md

@@ -66,6 +66,14 @@ server.exe -m models\7B\ggml-model.gguf -c 2048
 The above command will start a server that by default listens on `127.0.0.1:8080`.
 You can consume the endpoints with Postman or NodeJS with axios library. You can visit the web front end at the same url.
 
+### Docker:
+```bash
+docker run -p 8080:8080 -v /path/to/models:/models ggerganov/llama.cpp:server -m models/7B/ggml-model.gguf -c 512 --host 0.0.0.0 --port 8080
+
+# or, with CUDA:
+docker run -p 8080:8080 -v /path/to/models:/models --gpus all ggerganov/llama.cpp:server-cuda -m models/7B/ggml-model.gguf -c 512 --host 0.0.0.0 --port 8080 --n-gpu-layers 99
+```
+
 ## Testing with CURL
 
 Using [curl](https://curl.se/). On Windows `curl.exe` should be available in the base OS.

gguf-py/gguf/constants.py

@@ -101,6 +101,7 @@ class MODEL_ARCH(IntEnum):
     PHI2      = auto()
     PLAMO     = auto()
     CODESHELL = auto()
+    ORION     = auto()
 
 
 class MODEL_TENSOR(IntEnum):
@@ -151,6 +152,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.PHI2:           "phi2",
     MODEL_ARCH.PLAMO:          "plamo",
     MODEL_ARCH.CODESHELL:      "codeshell",
+    MODEL_ARCH.ORION:          "orion",
 }
 
 TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
@@ -427,7 +429,23 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_NORM,
         MODEL_TENSOR.FFN_DOWN,
         MODEL_TENSOR.FFN_UP,
-    ]
+    ],
+    MODEL_ARCH.ORION: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.ATTN_ROT_EMBD,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
     # TODO
 }
 
@@ -452,6 +470,10 @@ MODEL_TENSOR_SKIP: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.ROPE_FREQS,
         MODEL_TENSOR.ATTN_ROT_EMBD,
     ],
+    MODEL_ARCH.ORION: [
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.ATTN_ROT_EMBD,
+    ],
 }
 
 #

llama.cpp

@@ -52,6 +52,7 @@
 #include <algorithm>
 #include <array>
 #include <cassert>
+#include <cfloat>
 #include <cinttypes>
 #include <climits>
 #include <cmath>
@@ -196,6 +197,7 @@ enum llm_arch {
     LLM_ARCH_PHI2,
     LLM_ARCH_PLAMO,
     LLM_ARCH_CODESHELL,
+    LLM_ARCH_ORION,
     LLM_ARCH_UNKNOWN,
 };
 
@@ -217,6 +219,7 @@ static std::map<llm_arch, std::string> LLM_ARCH_NAMES = {
     { LLM_ARCH_PHI2,            "phi2"      },
     { LLM_ARCH_PLAMO,           "plamo"     },
     { LLM_ARCH_CODESHELL,       "codeshell" },
+    { LLM_ARCH_ORION,           "orion"     },
 };
 
 enum llm_kv {
@@ -641,6 +644,25 @@ static std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NAMES =
             { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
         },
     },
+    {
+        LLM_ARCH_ORION,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ROPE_FREQS,      "rope_freqs" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_ATTN_ROT_EMBD,   "blk.%d.attn_rot_embd" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+        },
+    },
 
     {
         LLM_ARCH_UNKNOWN,
@@ -1332,6 +1354,7 @@ enum e_model {
     MODEL_7B,
     MODEL_8B,
     MODEL_13B,
+    MODEL_14B,
     MODEL_15B,
     MODEL_30B,
     MODEL_34B,
@@ -2683,6 +2706,7 @@ static const char * llama_model_type_name(e_model type) {
         case MODEL_7B:     return "7B";
         case MODEL_8B:     return "8B";
         case MODEL_13B:    return "13B";
+        case MODEL_14B:    return "14B";
         case MODEL_15B:    return "15B";
         case MODEL_30B:    return "30B";
         case MODEL_34B:    return "34B";
@@ -2950,7 +2974,15 @@ static void llm_load_hparams(
                     default: model.type = e_model::MODEL_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_ORION:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
 
+                switch (hparams.n_layer) {
+                    case 40: model.type = e_model::MODEL_14B; break;
+                    default: model.type = e_model::MODEL_UNKNOWN;
+                }
+            } break;
         default: (void)0;
     }
 
@@ -3933,6 +3965,38 @@ static bool llm_load_tensors(
                         layer.ffn_up_b   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP, "bias", i),     {n_ff});
                     }
                 } break;
+            case LLM_ARCH_ORION:
+                {
+                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+                    {
+                        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});
+                        model.output        = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab});
+                    }
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        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_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "bias", i),   {n_embd});
+
+                        layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd});
+                        layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_gqa});
+                        layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_gqa});
+                        layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
+
+                        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});
+
+                        layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff});
+                        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});
+                    }
+                } break;
+
+
             default:
                 throw std::runtime_error("unknown architecture");
         }
@@ -4563,6 +4627,126 @@ struct llm_build_context {
             ctx0 = nullptr;
         }
     }
+    struct ggml_cgraph * build_orion() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb);
+        cb(inpL, "inp_embd", -1);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0);
+        cb(inp_pos, "inp_pos", -1);
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0);
+        cb(KQ_mask, "KQ_mask", -1);
+
+        // shift the entire K-cache if needed
+        if (do_rope_shift) {
+            llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, lctx.inp_K_shift, LLM_ROPE, n_ctx, freq_base, freq_scale, cb);
+        }
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * inpSA = inpL;
+
+            // norm
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm, model.layers[il].attn_norm_b,
+                    LLM_NORM, cb, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                // compute Q and K and RoPE them
+                struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+                // if (model.layers[il].bq) {
+                //     Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                //     cb(Qcur, "Qcur", il);
+                // }
+
+                struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+                // if (model.layers[il].bk) {
+                //     Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                //     cb(Kcur, "Kcur", il);
+                // }
+
+                struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+                // if (model.layers[il].bv) {
+                //     Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                //     cb(Vcur, "Vcur", il);
+                // }
+
+                Qcur = ggml_rope_custom(
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens), inp_pos,
+                    hparams.n_rot, 2, 0, n_orig_ctx, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Qcur, "Qcur", il);
+
+                Kcur = ggml_rope_custom(
+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos,
+                    hparams.n_rot, 2, 0, n_orig_ctx, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur", il);
+
+                cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
+                        model.layers[il].wo, NULL,
+                        Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                cb(cur, "kqv_out", il);
+            }
+
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // feed-forward network
+            cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                    model.layers[il].ffn_norm, 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_gate, NULL,
+                    model.layers[il].ffn_down, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
+            cb(cur, "ffn_out", il);
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = llm_build_norm(ctx0, cur, hparams,
+                model.output_norm, model.output_norm_b,
+                LLM_NORM, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        // lm_head
+        cur = ggml_mul_mat(ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
+
 
     struct ggml_cgraph * build_llama() {
         struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
@@ -6520,6 +6704,10 @@ static struct ggml_cgraph * llama_build_graph(
             {
                 result = llm.build_codeshell();
             } break;
+        case LLM_ARCH_ORION:
+            {
+                result = llm.build_orion();
+            } break;
         default:
             GGML_ASSERT(false);
     }
@@ -7946,6 +8134,11 @@ void llama_sample_softmax(struct llama_context * ctx, llama_token_data_array * c
 }
 
 void llama_sample_top_k(struct llama_context * ctx, llama_token_data_array * candidates, int32_t k, size_t min_keep) {
+    // TODO: move bucket sort to separate function so that top_p/tail_free/typical/softmax first is equally fast
+    // if (k >= (int32_t)candidates->size) {
+    //     return;
+    // }
+
     const int64_t t_start_sample_us = ggml_time_us();
 
     k = std::max(k, (int) min_keep);
@@ -8054,21 +8247,56 @@ void llama_sample_min_p(struct llama_context * ctx, llama_token_data_array * can
         return;
     }
 
-    llama_sample_softmax(ctx, candidates);
-
     const int64_t t_start_sample_us = ggml_time_us();
 
-    float scale = candidates->data[0].p; // scale by max prob
+    bool min_p_applied = false;
+
+    // if the candidates aren't sorted, try the unsorted implementation first
+    if (!candidates->sorted) {
+        std::vector<llama_token_data> filtered_tokens;
+
+        float max_logit = -FLT_MAX;
+        for (size_t i = 0; i < candidates->size; ++i) {
+            max_logit = std::max(max_logit, candidates->data[i].logit);
+        }
+        const float min_logit = max_logit + logf(p); // min logit for p_i >= p * p_max
+
+        for (size_t i = 0; i < candidates->size; ++i) {
+            if (candidates->data[i].logit >= min_logit) {
+                filtered_tokens.push_back(candidates->data[i]);
+            }
+        }
+
+        // if we have enough values the operation was a success
+        if (filtered_tokens.size() >= min_keep) {
+            memcpy(candidates->data, filtered_tokens.data(), filtered_tokens.size()*sizeof(llama_token_data));
+            candidates->size = filtered_tokens.size();
+            min_p_applied = true;
+        }
+    }
+
+    // if the candidates are sorted or the unsorted implementation failed, use this implementation
+    if (!min_p_applied) {
+        // Sort the logits in descending order
+        if (!candidates->sorted) {
+            std::sort(candidates->data, candidates->data + candidates->size, [](const llama_token_data & a, const llama_token_data & b) {
+                return a.logit > b.logit;
+            });
+            candidates->sorted = true;
+        }
+
+        const float min_logit = candidates->data[0].logit + logf(p); // min logit for p_i >= p * p_max
         size_t i = 1; // first token always matches
 
         for (; i < candidates->size; ++i) {
-        if (candidates->data[i].p < p * scale && i >= min_keep) {
+            if (candidates->data[i].logit < min_logit && i >= min_keep) {
                 break; // prob too small
             }
         }
 
         // Resize the output vector to keep only the matching tokens
         candidates->size = i;
+    }
 
     if (ctx) {
         ctx->t_sample_us += ggml_time_us() - t_start_sample_us;

tests/test-sampling.cpp

@@ -5,11 +5,10 @@
 #undef NDEBUG
 #endif
 
-#include <cmath>
-#include <numeric>
-#include <cassert>
-#include <vector>
 #include <algorithm>
+#include <cmath>
+#include <string>
+#include <vector>
 
 static void dump(const llama_token_data_array * candidates) {
     for (size_t i = 0; i < candidates->size; i++) {
@@ -20,11 +19,11 @@ static void dump(const llama_token_data_array * candidates) {
 #define DUMP(__candidates) do { printf("%s:%d (%s)\n", __FILE__, __LINE__, __func__); dump((__candidates)); printf("-\n"); } while(0)
 
 static void test_top_k(const std::vector<float> & probs, const std::vector<float> & expected_probs, int k) {
-    size_t n_vocab = probs.size();
+    const size_t n_vocab = probs.size();
     std::vector<llama_token_data> candidates;
     candidates.reserve(n_vocab);
     for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) {
-        float logit = log(probs[token_id]);
+        const float logit = logf(probs[token_id]);
         candidates.emplace_back(llama_token_data{token_id, logit, 0.0f});
     }
 
@@ -41,11 +40,11 @@ static void test_top_k(const std::vector<float> & probs, const std::vector<float
 }
 
 static void test_top_p(const std::vector<float> & probs, const std::vector<float> & expected_probs, float p) {
-    size_t n_vocab = probs.size();
+    const size_t n_vocab = probs.size();
     std::vector<llama_token_data> candidates;
     candidates.reserve(n_vocab);
     for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) {
-        float logit = log(probs[token_id]);
+        const float logit = logf(probs[token_id]);
         candidates.emplace_back(llama_token_data{token_id, logit, 0.0f});
     }
 
@@ -62,11 +61,11 @@ static void test_top_p(const std::vector<float> & probs, const std::vector<float
 }
 
 static void test_tfs(const std::vector<float> & probs, const std::vector<float> & expected_probs, float z) {
-    size_t n_vocab = probs.size();
+    const size_t n_vocab = probs.size();
     std::vector<llama_token_data> candidates;
     candidates.reserve(n_vocab);
     for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) {
-        float logit = log(probs[token_id]);
+        const float logit = logf(probs[token_id]);
         candidates.emplace_back(llama_token_data{token_id, logit, 0.0f});
     }
 
@@ -81,12 +80,33 @@ static void test_tfs(const std::vector<float> & probs, const std::vector<float>
     }
 }
 
-static void test_typical(const std::vector<float> & probs, const std::vector<float> & expected_probs, float p) {
+static void test_min_p(const std::vector<float> & probs, const std::vector<float> & expected_probs, float p) {
-    size_t n_vocab = probs.size();
+    const size_t n_vocab = probs.size();
     std::vector<llama_token_data> candidates;
     candidates.reserve(n_vocab);
     for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) {
-        float logit = log(probs[token_id]);
+        const float logit = logf(probs[token_id]);
+        candidates.emplace_back(llama_token_data{token_id, logit, 0.0f});
+    }
+
+    llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
+    DUMP(&candidates_p);
+    llama_sample_min_p(nullptr, &candidates_p, p, 1);
+    DUMP(&candidates_p);
+    llama_sample_softmax(nullptr, &candidates_p);
+
+    GGML_ASSERT(candidates_p.size == expected_probs.size());
+    for (size_t i = 0; i < candidates_p.size; i++) {
+        GGML_ASSERT(fabs(candidates_p.data[i].p - expected_probs[i]) < 1e-3);
+    }
+}
+
+static void test_typical(const std::vector<float> & probs, const std::vector<float> & expected_probs, float p) {
+    const size_t n_vocab = probs.size();
+    std::vector<llama_token_data> candidates;
+    candidates.reserve(n_vocab);
+    for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) {
+        const float logit = logf(probs[token_id]);
         candidates.emplace_back(llama_token_data{token_id, logit, 0.0f});
     }
 
@@ -107,11 +127,11 @@ static void test_repetition_penalties(
 ) {
     GGML_ASSERT(probs.size() == expected_probs.size());
 
-    size_t n_vocab = probs.size();
+    const size_t n_vocab = probs.size();
     std::vector<llama_token_data> candidates;
     candidates.reserve(n_vocab);
     for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) {
-        float logit = log(probs[token_id]);
+        const float logit = logf(probs[token_id]);
         candidates.emplace_back(llama_token_data{token_id, logit, 0.0f});
     }
 
@@ -128,6 +148,88 @@ static void test_repetition_penalties(
     }
 }
 
+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> candidates;
+    candidates.reserve(n_vocab);
+    for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) {
+        const float logit = logf(token_id);
+        candidates.emplace_back(llama_token_data{token_id, logit, 0.0f});
+    }
+
+    llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
+
+          llama_token min_token_id = 0;
+    const llama_token max_token_id = n_vocab-1;
+
+    for (auto s : samplers_sequence) {
+        switch (s){
+            case 'k': llama_sample_top_k    (nullptr, &candidates_p, top_k, 1); break;
+            case 'f': GGML_ASSERT(false && "tail_free test not implemented");   break;
+            case 'y': GGML_ASSERT(false && "typical test not implemented");     break;
+            case 'p': llama_sample_top_p    (nullptr, &candidates_p, top_p, 1); break;
+            case 'm': llama_sample_min_p    (nullptr, &candidates_p, min_p, 1); break;
+            case 't': GGML_ASSERT(false && "temperature test not implemented"); break;
+            default : GGML_ASSERT(false && "Unknown sampler");                  break;
+        }
+
+        llama_sample_softmax(nullptr, &candidates_p); // make sure tokens are sorted for tests
+
+        const int size = candidates_p.size;
+
+        if (s == 'k') {
+            const int expected_size = std::min(size, top_k);
+            min_token_id = std::max(min_token_id, (llama_token)(n_vocab - top_k));
+
+            GGML_ASSERT(size == expected_size);
+            GGML_ASSERT(candidates_p.data[0].id == max_token_id);
+            GGML_ASSERT(candidates_p.data[expected_size-1].id == min_token_id);
+        } else if (s == 'p') {
+            const int softmax_divisor = n_vocab * (n_vocab-1) / 2 - min_token_id * (min_token_id-1) / 2;
+            const int softmax_numerator_target = ceilf(top_p * softmax_divisor);
+
+                min_token_id  = n_vocab;
+            int expected_size = 0;
+            int cumsum        = 0;
+            do { // do-while because always at least one token is sampled
+                min_token_id--;
+                expected_size++;
+
+                cumsum += min_token_id;
+            } while (cumsum < softmax_numerator_target);
+
+            // token 0 has p == 0, need special consideration for cumsum because top_p immediately returns
+            if (min_token_id == 1) {
+                min_token_id--;
+                expected_size += 1;
+            }
+
+            GGML_ASSERT(size == expected_size);
+            GGML_ASSERT(candidates_p.data[0].id == max_token_id);
+            GGML_ASSERT(candidates_p.data[expected_size-1].id == min_token_id);
+        } else if (s == 'm') {
+            int expected_size = ceilf((1.0f-min_p) * n_vocab);
+            expected_size = std::max(expected_size, 1);
+            expected_size = std::min(expected_size, size);
+
+            min_token_id = floorf(min_p * n_vocab);
+            min_token_id = std::max(min_token_id, 1);
+            min_token_id = std::max(min_token_id, (llama_token)(n_vocab - size));
+            min_token_id = std::min(min_token_id, (llama_token)(n_vocab - 1));
+
+            GGML_ASSERT(size == expected_size);
+            GGML_ASSERT(candidates_p.data[0].id == max_token_id);
+            GGML_ASSERT(candidates_p.data[expected_size-1].id == min_token_id);
+        } else {
+            GGML_ASSERT(false);
+        }
+    }
+
+    printf("Sampler queue %3s OK with n_vocab=%05ld top_k=%05d top_p=%f min_p=%f\n",
+           samplers_sequence.c_str(), n_vocab, top_k, top_p, min_p);
+}
+
 int main(void) {
     ggml_time_init();
 
@@ -139,6 +241,15 @@ int main(void) {
     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_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);
+    test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/0.9f, 0.3f/0.9f, 0.2f/0.9f},            0.26f);
+    test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/0.9f, 0.3f/0.9f, 0.2f/0.9f},            0.49f);
+    test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/0.7f, 0.3f/0.7f},                       0.51f);
+    test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/0.7f, 0.3f/0.7f},                       0.74f);
+    test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/0.4f},                                  0.76f);
+    test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/0.4f},                                  1.00f);
+
     test_tfs({0.1f, 0.15f, 0.2f, 0.25f, 0.3f}, {0.3f}, 0.25f);
     test_tfs({0.1f, 0.15f, 0.2f, 0.25f, 0.3f}, {0.3f, 0.25f}, 0.75f);
     test_tfs({0.1f, 0.15f, 0.2f, 0.25f, 0.3f}, {0.3f, 0.25f}, 0.99f);
@@ -154,6 +265,34 @@ int main(void) {
     test_repetition_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2},       {0.499966f, 0.499966f, 0.000023f, 0.000023f, 0.000023f}, 1.0f, 5.0f, 5.0f);
     test_repetition_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2, 0, 0}, {0.499977f, 0.499977f, 0.000023f, 0.000023f, 0.000000f}, 1.0f, 5.0f, 5.0f);
 
+    test_sampler_queue(10000, "k", 10000, 1.0f, 1.0f);
+    test_sampler_queue(10000, "k",     1, 1.0f, 1.0f);
+    test_sampler_queue(10000, "p", 10000, 1.0f, 1.0f);
+    test_sampler_queue(10000, "p", 10000, 0.0f, 1.0f);
+    test_sampler_queue(10000, "m", 10000, 1.0f, 1.0f);
+    test_sampler_queue(10000, "m", 10000, 1.0f, 1e-12);
+
+    test_sampler_queue(10000, "k",   100, 1.0000f, 1.0f);
+    test_sampler_queue(10000, "p", 10000, 0.0002f, 1.0f);
+    test_sampler_queue(10000, "p", 10000, 0.8000f, 1.0f);
+    test_sampler_queue(10000, "m", 10000, 1.0000f, 9997.9f/9999.0f);
+    test_sampler_queue(10000, "m", 10000, 1.0000f, 0.1f);
+
+    test_sampler_queue(10000, "kp", 100, 0.8f, 0.1f);
+    test_sampler_queue(10000, "km", 100, 0.8f, 0.1f);
+    test_sampler_queue(10000, "pk", 100, 0.8f, 0.1f);
+    test_sampler_queue(10000, "pm", 100, 0.8f, 0.1f);
+    test_sampler_queue(10000, "mk", 100, 0.8f, 0.1f);
+    test_sampler_queue(10000, "mp", 100, 0.8f, 9997.9f/9999.0f);
+    test_sampler_queue(10000, "mp", 100, 0.8f, 0.1f);
+
+    test_sampler_queue(10000, "kpm", 100, 0.8f, 0.1f);
+    test_sampler_queue(10000, "kmp", 100, 0.8f, 0.1f);
+    test_sampler_queue(10000, "pkm", 100, 0.8f, 0.1f);
+    test_sampler_queue(10000, "pmk", 100, 0.8f, 0.1f);
+    test_sampler_queue(10000, "mkp", 100, 0.8f, 0.1f);
+    test_sampler_queue(10000, "mpk", 100, 0.8f, 0.1f);
+
     printf("OK\n");
 
     return 0;