Merge https://github.com/ollama/ollama

Signed-off-by: baalajimaestro <baalajimaestro@ptr.moe>

CONTRIBUTING.md

@@ -18,7 +18,7 @@ See the [development documentation](./docs/development.md) for instructions on h
 
 * New features: new features (e.g. API fields, environment variables) add surface area to Ollama and make it harder to maintain in the long run as they cannot be removed without potentially breaking users in the future.
 * Refactoring: large code improvements are important, but can be harder or take longer to review and merge.
-* Documentation: small updates to fill in or dorrect missing documentation is helpful, however large documentation additions can be hard to maintain over time.
+* Documentation: small updates to fill in or correct missing documentation is helpful, however large documentation additions can be hard to maintain over time.
 
 ### Issues that may not be accepted
 

Dockerfile

@@ -1,217 +1,61 @@
-ARG GOLANG_VERSION=1.22.5
+# Build stage
-ARG CMAKE_VERSION=3.22.1
+FROM golang:1.22-bookworm as build
-ARG CUDA_VERSION_11=11.3.1
-ARG CUDA_V11_ARCHITECTURES="50;52;53;60;61;62;70;72;75;80;86"
-ARG CUDA_VERSION_12=12.4.0
-ARG CUDA_V12_ARCHITECTURES="60;61;62;70;72;75;80;86;87;89;90;90a"
-ARG ROCM_VERSION=6.1.2
 
-# Copy the minimal context we need to run the generate scripts
+# Install necessary dependencies
-FROM scratch AS llm-code
+RUN apt update && apt install -y \
-COPY .git .git
+    wget \
-COPY .gitmodules .gitmodules
+    gnupg \
-COPY llm llm
+    software-properties-common \
+    git \
+    apt-utils
 
-FROM --platform=linux/amd64 nvidia/cuda:$CUDA_VERSION_11-devel-centos7 AS cuda-11-build-amd64
+# Install Intel oneAPI
-ARG CMAKE_VERSION
+RUN wget -O- https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB | gpg --dearmor | tee /usr/share/keyrings/oneapi-archive-keyring.gpg > /dev/null && \
-COPY ./scripts/rh_linux_deps.sh /
+    echo "deb [signed-by=/usr/share/keyrings/oneapi-archive-keyring.gpg] https://apt.repos.intel.com/oneapi all main" | tee /etc/apt/sources.list.d/oneAPI.list && \
-RUN CMAKE_VERSION=${CMAKE_VERSION} sh /rh_linux_deps.sh
+    apt update && \
-ENV PATH /opt/rh/devtoolset-10/root/usr/bin:$PATH
+    apt install -y --no-install-recommends \
-COPY --from=llm-code / /go/src/github.com/ollama/ollama/
+    intel-oneapi-mkl \
-WORKDIR /go/src/github.com/ollama/ollama/llm/generate
+    intel-oneapi-compiler-dpcpp-cpp \
-ARG CGO_CFLAGS
+    intel-oneapi-mkl-devel \
-ARG CUDA_V11_ARCHITECTURES
+    gcc \
-ENV GOARCH amd64 
+    g++ \
-RUN --mount=type=cache,target=/root/.ccache \
+    pkg-config \
-    OLLAMA_SKIP_STATIC_GENERATE=1 \
+    cmake
-    OLLAMA_SKIP_CPU_GENERATE=1 \
-    CMAKE_CUDA_ARCHITECTURES="${CUDA_V11_ARCHITECTURES}" \
-    CUDA_VARIANT="_v11" \
-    bash gen_linux.sh
 
-FROM --platform=linux/amd64 nvidia/cuda:$CUDA_VERSION_12-devel-centos7 AS cuda-12-build-amd64
+WORKDIR /app
-ARG CMAKE_VERSION
-COPY ./scripts/rh_linux_deps.sh /
-RUN CMAKE_VERSION=${CMAKE_VERSION} sh /rh_linux_deps.sh
-ENV PATH /opt/rh/devtoolset-10/root/usr/bin:$PATH
-COPY --from=llm-code / /go/src/github.com/ollama/ollama/
-WORKDIR /go/src/github.com/ollama/ollama/llm/generate
-ARG CGO_CFLAGS
-ARG CUDA_V12_ARCHITECTURES
-ENV GOARCH amd64 
-RUN --mount=type=cache,target=/root/.ccache \
-    OLLAMA_SKIP_STATIC_GENERATE=1 \
-    OLLAMA_SKIP_CPU_GENERATE=1 \
-    CMAKE_CUDA_ARCHITECTURES="${CUDA_V12_ARCHITECTURES}" \
-    CUDA_VARIANT="_v12" \
-    OLLAMA_CUSTOM_CUDA_DEFS="-DGGML_CUDA_USE_GRAPHS=on" \
-    bash gen_linux.sh
 
-FROM --platform=linux/arm64 nvidia/cuda:$CUDA_VERSION_11-devel-rockylinux8 AS cuda-11-build-server-arm64
+ARG GIN_MODE=release
-ARG CMAKE_VERSION
-COPY ./scripts/rh_linux_deps.sh /
-RUN CMAKE_VERSION=${CMAKE_VERSION} sh /rh_linux_deps.sh
-ENV PATH /opt/rh/gcc-toolset-10/root/usr/bin:$PATH
-COPY --from=llm-code / /go/src/github.com/ollama/ollama/
-WORKDIR /go/src/github.com/ollama/ollama/llm/generate
-ARG CGO_CFLAGS
-ARG CUDA_V11_ARCHITECTURES
-ENV GOARCH arm64 
-RUN OLLAMA_SKIP_STATIC_GENERATE=1 \
-    OLLAMA_SKIP_CPU_GENERATE=1 \
-    CMAKE_CUDA_ARCHITECTURES="${CUDA_V11_ARCHITECTURES}" \
-    CUDA_VARIANT="_v11" \
-    bash gen_linux.sh
 
-FROM --platform=linux/arm64 nvidia/cuda:$CUDA_VERSION_12-devel-rockylinux8 AS cuda-12-build-server-arm64
+ADD . .
-ARG CMAKE_VERSION
-COPY ./scripts/rh_linux_deps.sh /
-RUN CMAKE_VERSION=${CMAKE_VERSION} sh /rh_linux_deps.sh
-ENV PATH /opt/rh/gcc-toolset-10/root/usr/bin:$PATH
-COPY --from=llm-code / /go/src/github.com/ollama/ollama/
-WORKDIR /go/src/github.com/ollama/ollama/llm/generate
-ARG CGO_CFLAGS
-ARG CUDA_V12_ARCHITECTURES
-ENV GOARCH arm64 
-RUN --mount=type=cache,target=/root/.ccache \
-    OLLAMA_SKIP_STATIC_GENERATE=1 \
-    OLLAMA_SKIP_CPU_GENERATE=1 \
-    CMAKE_CUDA_ARCHITECTURES="${CUDA_V12_ARCHITECTURES}" \
-    CUDA_VARIANT="_v12" \
-    OLLAMA_CUSTOM_CUDA_DEFS="-DGGML_CUDA_USE_GRAPHS=on" \
-    bash gen_linux.sh
 
+RUN . /opt/intel/oneapi/setvars.sh && \
+    OLLAMA_CUSTOM_CPU_DEFS="-DGGML_BLAS=ON -DGGML_BLAS_VENDOR=Intel10_64lp -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_NATIVE=ON" go generate ./... && \
+    go build -ldflags="-s -w"
 
-FROM --platform=linux/amd64 rocm/dev-centos-7:${ROCM_VERSION}-complete AS rocm-build-amd64
+FROM debian:bookworm-slim
-ARG CMAKE_VERSION
-COPY ./scripts/rh_linux_deps.sh /
-RUN CMAKE_VERSION=${CMAKE_VERSION} sh /rh_linux_deps.sh
-ENV PATH /opt/rh/devtoolset-10/root/usr/bin:$PATH
-ENV LIBRARY_PATH /opt/amdgpu/lib64
-COPY --from=llm-code / /go/src/github.com/ollama/ollama/
-WORKDIR /go/src/github.com/ollama/ollama/llm/generate
-ARG CGO_CFLAGS
-ARG AMDGPU_TARGETS
-ENV GOARCH amd64 
-RUN --mount=type=cache,target=/root/.ccache \
-    OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_SKIP_CPU_GENERATE=1 bash gen_linux.sh
-RUN mkdir -p ../../dist/linux-amd64-rocm/lib/ollama && \
-    (cd /opt/rocm/lib && tar cf - rocblas/library) | (cd ../../dist/linux-amd64-rocm/lib/ollama && tar xf - )
 
-FROM --platform=linux/amd64 centos:7 AS cpu-builder-amd64
+RUN apt update && apt install -y --no-install-recommends \
-ARG CMAKE_VERSION
+    wget \
-ARG GOLANG_VERSION
+    gnupg \
-COPY ./scripts/rh_linux_deps.sh /
+    software-properties-common \
-RUN CMAKE_VERSION=${CMAKE_VERSION} GOLANG_VERSION=${GOLANG_VERSION} sh /rh_linux_deps.sh
+    supervisor && \
-ENV PATH /opt/rh/devtoolset-10/root/usr/bin:$PATH
+    wget -O- https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB | gpg --dearmor | tee /usr/share/keyrings/oneapi-archive-keyring.gpg > /dev/null && \
-COPY --from=llm-code / /go/src/github.com/ollama/ollama/
+    echo "deb [signed-by=/usr/share/keyrings/oneapi-archive-keyring.gpg] https://apt.repos.intel.com/oneapi all main" | tee /etc/apt/sources.list.d/oneAPI.list && \
-ARG OLLAMA_CUSTOM_CPU_DEFS
+    apt update && \
-ARG CGO_CFLAGS
+    apt install -y --no-install-recommends intel-oneapi-runtime-libs && \
-ENV GOARCH amd64 
+    apt clean && \
-WORKDIR /go/src/github.com/ollama/ollama/llm/generate
+    rm -rf /var/lib/apt/lists/*
 
-FROM --platform=linux/amd64 cpu-builder-amd64 AS static-build-amd64
+WORKDIR /app
-RUN --mount=type=cache,target=/root/.ccache \
-    OLLAMA_CPU_TARGET="static" bash gen_linux.sh
-FROM --platform=linux/amd64 cpu-builder-amd64 AS cpu-build-amd64
-RUN --mount=type=cache,target=/root/.ccache \
-    OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu" bash gen_linux.sh
-FROM --platform=linux/amd64 cpu-builder-amd64 AS cpu_avx-build-amd64
-RUN --mount=type=cache,target=/root/.ccache \
-    OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu_avx" bash gen_linux.sh
-FROM --platform=linux/amd64 cpu-builder-amd64 AS cpu_avx2-build-amd64
-RUN --mount=type=cache,target=/root/.ccache \
-    OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu_avx2" bash gen_linux.sh
 
-FROM --platform=linux/arm64 rockylinux:8 AS cpu-builder-arm64
+COPY --from=build /app/supervisord.conf /app/supervisord.conf
-ARG CMAKE_VERSION
+COPY --from=build /app/ollama /app/ollama
-ARG GOLANG_VERSION
+COPY --from=build /app/run_model.sh /app/run_model.sh
-COPY ./scripts/rh_linux_deps.sh /
+COPY --from=build /app/serve.sh /app/serve.sh
-RUN CMAKE_VERSION=${CMAKE_VERSION} GOLANG_VERSION=${GOLANG_VERSION} sh /rh_linux_deps.sh
-ENV PATH /opt/rh/gcc-toolset-10/root/usr/bin:$PATH
-COPY --from=llm-code / /go/src/github.com/ollama/ollama/
-ARG OLLAMA_CUSTOM_CPU_DEFS
-ARG CGO_CFLAGS
-ENV GOARCH arm64
-WORKDIR /go/src/github.com/ollama/ollama/llm/generate
 
-FROM --platform=linux/arm64 cpu-builder-arm64 AS static-build-arm64
+ENV MODEL_NAME="llama"
-RUN --mount=type=cache,target=/root/.ccache \
+ENV OLLAMA_HOST="0.0.0.0:8080"
-    OLLAMA_CPU_TARGET="static" bash gen_linux.sh
-FROM --platform=linux/arm64 cpu-builder-arm64 AS cpu-build-arm64
-RUN --mount=type=cache,target=/root/.ccache \
-    OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu" bash gen_linux.sh
 
+EXPOSE 8080
 
-# Intermediate stage used for ./scripts/build_linux.sh
+CMD ["supervisord", "-c", "/app/supervisord.conf"]
-FROM --platform=linux/amd64 cpu-build-amd64 AS build-amd64
-ENV CGO_ENABLED 1
-WORKDIR /go/src/github.com/ollama/ollama
-COPY . .
-COPY --from=static-build-amd64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
-COPY --from=cpu_avx-build-amd64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
-COPY --from=cpu_avx2-build-amd64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
-COPY --from=cuda-11-build-amd64 /go/src/github.com/ollama/ollama/dist/ dist/
-COPY --from=cuda-11-build-amd64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
-COPY --from=cuda-12-build-amd64 /go/src/github.com/ollama/ollama/dist/ dist/
-COPY --from=cuda-12-build-amd64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
-COPY --from=rocm-build-amd64 /go/src/github.com/ollama/ollama/dist/ dist/
-COPY --from=rocm-build-amd64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
-ARG GOFLAGS
-ARG CGO_CFLAGS
-RUN --mount=type=cache,target=/root/.ccache \
-    go build -trimpath -o dist/linux-amd64/bin/ollama .
-
-# Intermediate stage used for ./scripts/build_linux.sh
-FROM --platform=linux/arm64 cpu-build-arm64 AS build-arm64
-ENV CGO_ENABLED 1
-ARG GOLANG_VERSION
-WORKDIR /go/src/github.com/ollama/ollama
-COPY . .
-COPY --from=static-build-arm64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
-COPY --from=cuda-11-build-server-arm64 /go/src/github.com/ollama/ollama/dist/ dist/
-COPY --from=cuda-11-build-server-arm64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
-COPY --from=cuda-12-build-server-arm64 /go/src/github.com/ollama/ollama/dist/ dist/
-COPY --from=cuda-12-build-server-arm64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
-ARG GOFLAGS
-ARG CGO_CFLAGS
-RUN --mount=type=cache,target=/root/.ccache \
-    go build -trimpath -o dist/linux-arm64/bin/ollama .
-
-# Strip out ROCm dependencies to keep the primary image lean
-FROM --platform=linux/amd64 ubuntu:22.04 as amd64-libs-without-rocm
-COPY --from=build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/lib/ /scratch/
-RUN cd /scratch/ollama/ && rm -rf rocblas libamd* libdrm* libroc* libhip* libhsa* 
-
-# Runtime stages
-FROM --platform=linux/amd64 ubuntu:22.04 as runtime-amd64
-COPY --from=amd64-libs-without-rocm /scratch/ /lib/
-RUN apt-get update && apt-get install -y ca-certificates
-COPY --from=build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/bin/ /bin/
-
-FROM --platform=linux/arm64 ubuntu:22.04 as runtime-arm64
-COPY --from=build-arm64 /go/src/github.com/ollama/ollama/dist/linux-arm64/lib/ /lib/
-RUN apt-get update && apt-get install -y ca-certificates
-COPY --from=build-arm64 /go/src/github.com/ollama/ollama/dist/linux-arm64/bin/ /bin/
-
-# Radeon images are much larger so we keep it distinct from the CPU/CUDA image
-FROM --platform=linux/amd64 rocm/dev-centos-7:${ROCM_VERSION}-complete as runtime-rocm
-RUN update-pciids
-COPY --from=build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/bin/ /bin/
-RUN ln -s /opt/rocm/lib /lib/ollama
-EXPOSE 11434
-ENV OLLAMA_HOST 0.0.0.0
-
-ENTRYPOINT ["/bin/ollama"]
-CMD ["serve"]
-
-FROM runtime-$TARGETARCH
-EXPOSE 11434
-ENV OLLAMA_HOST 0.0.0.0
-ENV PATH=/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin
-ENV LD_LIBRARY_PATH=/usr/local/nvidia/lib:/usr/local/nvidia/lib64
-ENV NVIDIA_DRIVER_CAPABILITIES=compute,utility
-ENV NVIDIA_VISIBLE_DEVICES=all
-
-ENTRYPOINT ["/bin/ollama"]
-CMD ["serve"]

README.md

@@ -295,13 +295,23 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Olpaka](https://github.com/Otacon/olpaka) (User-friendly Flutter Web App for Ollama)
 - [OllamaSpring](https://github.com/CrazyNeil/OllamaSpring) (Ollama Client for macOS)
 - [LLocal.in](https://github.com/kartikm7/llocal) (Easy to use Electron Desktop Client for Ollama)
+- [AiLama](https://github.com/zeyoyt/ailama) (A Discord User App that allows you to interact with Ollama anywhere in discord )
 - [Ollama with Google Mesop](https://github.com/rapidarchitect/ollama_mesop/) (Mesop Chat Client implementation with Ollama)
+- [Painting Droid](https://github.com/mateuszmigas/painting-droid) (Painting app with AI integrations)
 - [Kerlig AI](https://www.kerlig.com/) (AI writing assistant for macOS)
 - [AI Studio](https://github.com/MindWorkAI/AI-Studio)
 - [Sidellama](https://github.com/gyopak/sidellama) (browser-based LLM client)
 - [LLMStack](https://github.com/trypromptly/LLMStack) (No-code multi-agent framework to build LLM agents and workflows)
 - [BoltAI for Mac](https://boltai.com) (AI Chat Client for Mac)
 - [Harbor](https://github.com/av/harbor) (Containerized LLM Toolkit with Ollama as default backend)
+- [Go-CREW](https://www.jonathanhecl.com/go-crew/) (Powerful Offline RAG in Golang)
+- [PartCAD](https://github.com/openvmp/partcad/) (CAD model generation with OpenSCAD and CadQuery)
+- [Ollama4j Web UI](https://github.com/ollama4j/ollama4j-web-ui) - Java-based Web UI for Ollama built with Vaadin, Spring Boot and Ollama4j
+- [PyOllaMx](https://github.com/kspviswa/pyOllaMx) - macOS application capable of chatting with both Ollama and Apple MLX models.
+- [Claude Dev](https://github.com/saoudrizwan/claude-dev) - VSCode extension for multi-file/whole-repo coding
+- [Cherry Studio](https://github.com/kangfenmao/cherry-studio) (Desktop client with Ollama support)
+- [ConfiChat](https://github.com/1runeberg/confichat) (Lightweight, standalone, multi-platform, and privacy focused LLM chat interface with optional encryption)
+- [Archyve](https://github.com/nickthecook/archyve) (RAG-enabling document library)
 
 ### Terminal
 
@@ -327,6 +337,9 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [gollama](https://github.com/sammcj/gollama)
 - [Ollama eBook Summary](https://github.com/cognitivetech/ollama-ebook-summary/)
 
+### Apple Vision Pro
+- [Enchanted](https://github.com/AugustDev/enchanted)
+
 ### Database
 
 - [MindsDB](https://github.com/mindsdb/mindsdb/blob/staging/mindsdb/integrations/handlers/ollama_handler/README.md) (Connects Ollama models with nearly 200 data platforms and apps)
@@ -335,6 +348,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 ### Package managers
 
 - [Pacman](https://archlinux.org/packages/extra/x86_64/ollama/)
+- [Gentoo](https://github.com/gentoo/guru/tree/master/app-misc/ollama)
 - [Helm Chart](https://artifacthub.io/packages/helm/ollama-helm/ollama)
 - [Guix channel](https://codeberg.org/tusharhero/ollama-guix)
 - [Nix package](https://search.nixos.org/packages?channel=24.05&show=ollama&from=0&size=50&sort=relevance&type=packages&query=ollama)
@@ -349,11 +363,12 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [LangChainRust](https://github.com/Abraxas-365/langchain-rust) with [example](https://github.com/Abraxas-365/langchain-rust/blob/main/examples/llm_ollama.rs)
 - [LlamaIndex](https://gpt-index.readthedocs.io/en/stable/examples/llm/ollama.html)
 - [LiteLLM](https://github.com/BerriAI/litellm)
+- [OllamaFarm for Go](https://github.com/presbrey/ollamafarm)
 - [OllamaSharp for .NET](https://github.com/awaescher/OllamaSharp)
 - [Ollama for Ruby](https://github.com/gbaptista/ollama-ai)
 - [Ollama-rs for Rust](https://github.com/pepperoni21/ollama-rs)
 - [Ollama-hpp for C++](https://github.com/jmont-dev/ollama-hpp)
-- [Ollama4j for Java](https://github.com/amithkoujalgi/ollama4j)
+- [Ollama4j for Java](https://github.com/ollama4j/ollama4j)
 - [ModelFusion Typescript Library](https://modelfusion.dev/integration/model-provider/ollama)
 - [OllamaKit for Swift](https://github.com/kevinhermawan/OllamaKit)
 - [Ollama for Dart](https://github.com/breitburg/dart-ollama)
@@ -370,11 +385,16 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Portkey](https://portkey.ai/docs/welcome/integration-guides/ollama)
 - [PromptingTools.jl](https://github.com/svilupp/PromptingTools.jl) with an [example](https://svilupp.github.io/PromptingTools.jl/dev/examples/working_with_ollama)
 - [LlamaScript](https://github.com/Project-Llama/llamascript)
+- [Gollm](https://docs.gollm.co/examples/ollama-example)
+- [Ollamaclient for Golang](https://github.com/xyproto/ollamaclient)
+- [High-level function abstraction in Go](https://gitlab.com/tozd/go/fun)
+- [Ollama PHP](https://github.com/ArdaGnsrn/ollama-php)
 
 ### Mobile
 
 - [Enchanted](https://github.com/AugustDev/enchanted)
 - [Maid](https://github.com/Mobile-Artificial-Intelligence/maid)
+- [ConfiChat](https://github.com/1runeberg/confichat) (Lightweight, standalone, multi-platform, and privacy focused LLM chat interface with optional encryption)
 
 ### Extensions & Plugins
 
@@ -399,11 +419,14 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [twinny](https://github.com/rjmacarthy/twinny) (Copilot and Copilot chat alternative using Ollama)
 - [Wingman-AI](https://github.com/RussellCanfield/wingman-ai) (Copilot code and chat alternative using Ollama and Hugging Face)
 - [Page Assist](https://github.com/n4ze3m/page-assist) (Chrome Extension)
+- [Plasmoid Ollama Control](https://github.com/imoize/plasmoid-ollamacontrol) (KDE Plasma extension that allows you to quickly manage/control Ollama model)
 - [AI Telegram Bot](https://github.com/tusharhero/aitelegrambot) (Telegram bot using Ollama in backend)
 - [AI ST Completion](https://github.com/yaroslavyaroslav/OpenAI-sublime-text) (Sublime Text 4 AI assistant plugin with Ollama support)
 - [Discord-Ollama Chat Bot](https://github.com/kevinthedang/discord-ollama) (Generalized TypeScript Discord Bot w/ Tuning Documentation)
 - [Discord AI chat/moderation bot](https://github.com/rapmd73/Companion) Chat/moderation bot written in python. Uses Ollama to create personalities.
 - [Headless Ollama](https://github.com/nischalj10/headless-ollama) (Scripts to automatically install ollama client & models on any OS for apps that depends on ollama server)
+- [vnc-lm](https://github.com/jk011ru/vnc-lm) (A containerized Discord bot with support for attachments and web links)
+- [LSP-AI](https://github.com/SilasMarvin/lsp-ai) (Open-source language server for AI-powered functionality)
 
 ### Supported backends
 

cmd/cmd.go

@@ -1421,6 +1421,8 @@ func NewCLI() *cobra.Command {
 				envVars["OLLAMA_TMPDIR"],
 				envVars["OLLAMA_FLASH_ATTENTION"],
 				envVars["OLLAMA_LLM_LIBRARY"],
+				envVars["OLLAMA_GPU_OVERHEAD"],
+				envVars["OLLAMA_LOAD_TIMEOUT"],
 			})
 		default:
 			appendEnvDocs(cmd, envs)

convert/convert_gemma2.go

@@ -34,10 +34,20 @@ func (p *gemma2Model) KV(t *Tokenizer) llm.KV {
 }
 
 func (p *gemma2Model) Replacements() []string {
-	return append(
+	return []string{
-		p.gemmaModel.Replacements(),
+		"model.embed_tokens", "token_embd",
+		"model.norm", "output_norm",
+		"model.layers", "blk",
+		"input_layernorm", "attn_norm",
+		"self_attn.q_proj", "attn_q",
+		"self_attn.k_proj", "attn_k",
+		"self_attn.v_proj", "attn_v",
+		"self_attn.o_proj", "attn_output",
+		"mlp.gate_proj", "ffn_gate",
+		"mlp.down_proj", "ffn_down",
+		"mlp.up_proj", "ffn_up",
 		"post_attention_layernorm", "post_attention_norm",
 		"pre_feedforward_layernorm", "ffn_norm",
 		"post_feedforward_layernorm", "post_ffw_norm",
-	)
+	}
 }

convert/convert_test.go

@@ -15,6 +15,7 @@ import (
 	"os"
 	"path/filepath"
 	"slices"
+	"strings"
 	"testing"
 
 	"golang.org/x/exp/maps"
@@ -22,6 +23,12 @@ import (
 	"github.com/ollama/ollama/llm"
 )
 
+type tensorData struct {
+	Offsets []int  `json:"data_offsets"`
+	Type    string `json:"dtype"`
+	Shape   []int  `json:"shape"`
+}
+
 func convertFull(t *testing.T, fsys fs.FS) (*os.File, llm.KV, llm.Tensors) {
 	t.Helper()
 
@@ -96,6 +103,7 @@ func TestConvertModel(t *testing.T) {
 		"Mistral-7B-Instruct-v0.2",
 		"Mixtral-8x7B-Instruct-v0.1",
 		"gemma-2b-it",
+		"gemma-2-2b-it",
 		// microsoft/Phi-3-mini-128-instruct@d548c233192db00165d842bf8edff054bb3212f8
 		"Phi-3-mini-128k-instruct",
 		"all-MiniLM-L6-v2",
@@ -140,6 +148,36 @@ func TestConvertModel(t *testing.T) {
 	}
 }
 
+func TestConvertInvalidTensorNames(t *testing.T) {
+	f, err := os.CreateTemp(t.TempDir(), "testmodel")
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer f.Close()
+
+	tempDir := t.TempDir()
+
+	td := map[string]*tensorData{}
+	offset := 4096
+
+	td["model.layers.0.self_attn.q_proj.weight"] = &tensorData{
+		Offsets: []int{0, offset},
+		Type:    "F32",
+		Shape:   []int{4096, 4096},
+	}
+	td["blk.0.attn_q.weight"] = &tensorData{
+		Offsets: []int{offset, offset * 2},
+		Type:    "F32",
+		Shape:   []int{4096, 4096},
+	}
+	generateSafetensorTestData(t, tempDir, td)
+
+	err = ConvertModel(os.DirFS(tempDir), f)
+	if err == nil || !strings.HasPrefix(err.Error(), "duplicate tensor name") {
+		t.Errorf("expected error but didn't get one")
+	}
+}
+
 func TestConvertInvalidDatatype(t *testing.T) {
 	f, err := os.CreateTemp(t.TempDir(), "testmodel")
 	if err != nil {
@@ -148,23 +186,10 @@ func TestConvertInvalidDatatype(t *testing.T) {
 	defer f.Close()
 
 	tempDir := t.TempDir()
-	generateSafetensorTestData(t, tempDir)
-
-	err = ConvertModel(os.DirFS(tempDir), f)
-	if err == nil || err.Error() != "unsupported safetensors model" {
-		t.Errorf("expected error but didn't get one")
-	}
-}
-
-func generateSafetensorTestData(t *testing.T, tempDir string) {
-	type tensorData struct {
-		Offsets []int  `json:"data_offsets"`
-		Type    string `json:"dtype"`
-		Shape   []int  `json:"shape"`
-	}
-	offset := 4096 * 14336
 
 	td := map[string]*tensorData{}
+	offset := 4096 * 14336
+
 	td["model.layers.0.mlp.down_proj.weight"] = &tensorData{
 		Offsets: []int{0, offset},
 		Type:    "I8",
@@ -175,8 +200,16 @@ func generateSafetensorTestData(t *testing.T, tempDir string) {
 		Type:    "U8",
 		Shape:   []int{},
 	}
+	generateSafetensorTestData(t, tempDir, td)
 
-	data, err := json.Marshal(td)
+	err = ConvertModel(os.DirFS(tempDir), f)
+	if err == nil || err.Error() != "unsupported safetensors model" {
+		t.Errorf("expected error but didn't get one")
+	}
+}
+
+func generateSafetensorTestData(t *testing.T, tempDir string, tensorData map[string]*tensorData) {
+	data, err := json.Marshal(tensorData)
 	if err != nil {
 		t.Fatal(err)
 	}
@@ -322,11 +355,6 @@ func TestConvertAdapter(t *testing.T) {
 }
 
 func generateLoraTestData(t *testing.T, tempDir string) {
-	type tensorData struct {
-		Offsets []int  `json:"data_offsets"`
-		Type    string `json:"dtype"`
-		Shape   []int  `json:"shape"`
-	}
 	offset := 4096 * 8 * 4
 
 	td := map[string]*tensorData{"__metadata__": nil}

convert/reader_safetensors.go

@@ -49,12 +49,19 @@ func parseSafetensors(fsys fs.FS, replacer *strings.Replacer, ps ...string) ([]T
 		keys := maps.Keys(headers)
 		slices.Sort(keys)
 
+		names := make(map[string]struct{}, len(keys))
+
 		for _, key := range keys {
 			if value := headers[key]; value.Type != "" {
 				// bitsandbytes quantized models are unsupported
 				if len(value.Shape) == 0 {
 					return nil, errors.New("unsupported safetensors model")
 				}
+				ggufName := replacer.Replace(key)
+				if _, ok := names[ggufName]; ok {
+					return nil, fmt.Errorf("duplicate tensor name '%s' was found for this model", ggufName)
+				}
+				names[ggufName] = struct{}{}
 				ts = append(ts, safetensor{
 					fs:     fsys,
 					path:   p,
@@ -62,7 +69,7 @@ func parseSafetensors(fsys fs.FS, replacer *strings.Replacer, ps ...string) ([]T
 					offset: safetensorsPad(n, value.Offsets[0]),
 					size:   safetensorsPad(n, value.Offsets[1]) - safetensorsPad(n, value.Offsets[0]),
 					tensorBase: &tensorBase{
-						name:  replacer.Replace(key),
+						name:  ggufName,
 						shape: value.Shape,
 					},
 				})

convert/testdata/gemma-2-2b-it.json

@@ -0,0 +1,312 @@
+{
+  "general.architecture": "gemma2",
+  "general.file_type": "1",
+  "general.quantization_version": "2",
+  "gemma2.block_count": "26",
+  "gemma2.context_length": "8192",
+  "gemma2.embedding_length": "2304",
+  "gemma2.feed_forward_length": "9216",
+  "gemma2.attention.head_count": "8",
+  "gemma2.attention.head_count_kv": "4",
+  "gemma2.attention.key_length": "256",
+  "gemma2.attention.value_length": "256",
+  "gemma2.attention.layer_norm_rms_epsilon": "1e-06",
+  "tokenizer.ggml.model": "llama",
+  "tokenizer.ggml.add_bos_token": "true",
+  "tokenizer.ggml.add_eos_token": "false",
+  "tokenizer.ggml.bos_token_id": "2",
+  "tokenizer.ggml.eos_token_id": "1",
+  "tokenizer.ggml.padding_token_id": "0",
+  "tokenizer.ggml.unknown_token_id": "3",
+  "tokenizer.ggml.scores": "0872465d173867d755d3ee728f882b9dc2057a0bfd596fe1e3d131522f1250d8",
+  "tokenizer.ggml.token_type": "8d40143b3477df77beea4139420335ede458bf5e14102f01b0170197b55da8d8",
+  "tokenizer.ggml.tokens": "c6e66de1841f04de8b8d236d461ab720a4c9b9b5414dc293a09c6e10eab45fda",
+  "token_embd.weight": "64a9d30707e659e2e673656d71f5aef7a9fb9fd83bb9a77558dfc5abbe218a05",
+  "blk.0.attn_k.weight": "d8b4437c5edb3cddf6af9987038e1bb2b191c4f0fce0e160d2abace717f5d5d7",
+  "blk.0.attn_norm.weight": "1eb73e3f7aa8e502f6ca31cd19efbb8e4fd9a89692e13e48ac8205545a7fa7e8",
+  "blk.0.attn_output.weight": "39e7b78e57d356a22dd89ce1c4d7163b970712ba756545e1703f97866cd2192e",
+  "blk.0.attn_q.weight": "795058e23b6109febd9d55c89e1eebe6af0714ec8c56fd86a160876a6135ffe8",
+  "blk.0.attn_v.weight": "0cd6e583d1887c020472e961bbb113fe5a0d23ae2f1c2c876fc366cdb7692b52",
+  "blk.0.ffn_down.weight": "51eb4d962189e945a84e94e0dc1aad3f8f90cc1a11e18029670afcd0ea0acb1b",
+  "blk.0.ffn_gate.weight": "9811a29b8ad48432925897ab21dfcb13c5cbd372aeccbbefca9b7866883b4ce3",
+  "blk.0.ffn_norm.weight": "92cbf4652ef503c1de5b10f2be00b3fcf00100980cb3baa8f3013a8d8bf3d851",
+  "blk.0.ffn_up.weight": "af87de21746879483ed1b374cdd76b19ba11ca2b6dbb1beba98efdf3be3e8077",
+  "blk.0.post_attention_norm.weight": "32e135f1f258ffe407018899e39af1725d59d66d60022b9a21575ba160e0357a",
+  "blk.0.post_ffw_norm.weight": "ba286f5ac11b07fbc986173708c66f1920427be5a6d108af38fa0a837c1c8eb6",
+  "blk.1.attn_k.weight": "51584435552051f7fade76beca582b3f7190cf7fc07adcf527c2774d4b1c3901",
+  "blk.1.attn_norm.weight": "6833104c7fbf35a7e799ae56c262b97fffa14789642aee14381b25acd21ed80a",
+  "blk.1.attn_output.weight": "14c39481369087bf292ac9a3ab2ef166f9fe376a9f90c246653213ef264febdc",
+  "blk.1.attn_q.weight": "443f64ae2229f857c69d6bebb7800b685786cb77884c3ae19d4286aeed081325",
+  "blk.1.attn_v.weight": "0df482de2038f1e4c8a7733ac0ddb69ad90759dab5968b942af0155588de4c4a",
+  "blk.1.ffn_down.weight": "66f30763a8bbbcaea609a0087ed75fadb5e771c06378dd2cea94cf17e492e8cf",
+  "blk.1.ffn_gate.weight": "a7151bff00a545fa18b2c92dcd2a14572ccf9beb957a6c494f1374e8ebe174c9",
+  "blk.1.ffn_norm.weight": "e197d71ea11b5276bc0167d2663b88089b3ff42b47ba91e85f6c5d95f6306435",
+  "blk.1.ffn_up.weight": "57c182e0b14cccd1350d388f0c616991702e74281db54637451b70f4ccc24f9b",
+  "blk.1.post_attention_norm.weight": "3c56f837168d784c2d8bac247c130bdca6610c095c8da4558c536ccad7605609",
+  "blk.1.post_ffw_norm.weight": "d2a51d320fd01069dd7ccaa7082f16a7faeb671885607d7900b10a89c354d0fa",
+  "blk.2.attn_k.weight": "bc103c818192de7ce36caaf89dc117be4df13fb902e0bd9a23c64edace5df9b6",
+  "blk.2.attn_norm.weight": "0f2503aa126083a5d6ac72481be1ef66c6014705b573682b35bd864e4749a3d5",
+  "blk.2.attn_output.weight": "05fcd4a1226e482f91803a266f72caca887a93e63c2d2ba5611ab3c68d38743a",
+  "blk.2.attn_q.weight": "6a10b5c2fd423d1e4c4fd60fa8c154a0159b6b2501ea79cae2ef19f45a674e5e",
+  "blk.2.attn_v.weight": "3cf891945a1f8ae7cc908a5c6b729ff5b70f4436c5ffdbf245cc0ed4cc19cd1b",
+  "blk.2.ffn_down.weight": "ea204fd04e0d2fc728a9861a459216bbfec629c152004ba625f52cd8837bd51e",
+  "blk.2.ffn_gate.weight": "3a3518729f1b8b64a82b8792f33987db5418fdb094be0263c68f146a5c38de54",
+  "blk.2.ffn_norm.weight": "754ede678b725de41a34b82f0edf7688b5c065be7c0d46df6f7ad9430d986884",
+  "blk.2.ffn_up.weight": "ffdcb88439f5828ffbd9fc844b03ff91637b790b9838097258cc3ae75935720c",
+  "blk.2.post_attention_norm.weight": "4b3f53b7ba26e8c36b2dfda3b7e5fc4b1065257cefdea235fc7df9af130ac2fd",
+  "blk.2.post_ffw_norm.weight": "e550369e26b8485e2b54ad34b34bc98af5494287dcc513c2c39cf1eaa5b89d07",
+  "blk.3.attn_k.weight": "89f24ea450e37d9e95757651a83205c085d81b354ee9489dd6310a391d8409f3",
+  "blk.3.attn_norm.weight": "24e2ea662b7cb822b4ca5cd61bc17f2709f406d990ec3b4a0dac1cc112db45cf",
+  "blk.3.attn_output.weight": "ac4dad69473c6e3fac56669212cadd8c34ecc5973d945972e974d94805334967",
+  "blk.3.attn_q.weight": "b6a9c9a7d4722b9096631c65de62228dfddca6e26edfe6af7fce01e116ef0f4c",
+  "blk.3.attn_v.weight": "f272a960a40093942309bc342a379984cbacec2d7bc64428db3f64e6b1887ed4",
+  "blk.3.ffn_down.weight": "c0188ba50d8228805982029c277fc0e87aa57473b8363037c648f6d006ff828a",
+  "blk.3.ffn_gate.weight": "a04aec1561ee6c0fbb18c3db49dc62fb533619cf697fd548cbf2279761aaec3b",
+  "blk.3.ffn_norm.weight": "bc053837d44087ec05eb5d9458357b2a5be787789b19cdbbdc694b57697f99a6",
+  "blk.3.ffn_up.weight": "b3ce8b274f20796d3b1a7c08ba27a919066f9de89a782faa544c4a8d6bea1382",
+  "blk.3.post_attention_norm.weight": "9c922dee7a7df5667289e2788e60170238239cee2dfdbbd9e435763f9f416718",
+  "blk.3.post_ffw_norm.weight": "b682544ac953ad2e0b49027ed8916f2e9d1aba5d1587bb4127ac703570c7a03a",
+  "blk.4.attn_k.weight": "143b0cbb4b787b95c2b6212374410e32173ccef2adb914908a2f89a7916de512",
+  "blk.4.attn_norm.weight": "5668f60491b780273745192662d02c9a92a4f692b29d16aa0bbc7413fec4f85b",
+  "blk.4.attn_output.weight": "b9f2bdb68be1e0cf66dd19f8fa2afb105910ad2ef394864cb32cea8f8944e0d5",
+  "blk.4.attn_q.weight": "ddcf1343dafbc2dfcd0b8741225af22fe4b54b2becce29240bd01c34265d126c",
+  "blk.4.attn_v.weight": "6dc7074366e7ed52d9f48c594dcc85bef738e096276cb99d28228c89eecc5b9c",
+  "blk.4.ffn_down.weight": "30334ffc59ce343cf2a1b973174acb7722823463adc07e19a99bd0f404bc9906",
+  "blk.4.ffn_gate.weight": "890f7c8af208d63b28db52c4b8c16c2288a382d87ff5a6a6d6b0a5b3bf27e6cd",
+  "blk.4.ffn_norm.weight": "ff0316cc7847221eb86a90c1ab441d4ee61553d410c66414a7755021b3b12448",
+  "blk.4.ffn_up.weight": "6af97d113f91564c636734f215e25ee602d48eb045458f300b3ec7582be0f41d",
+  "blk.4.post_attention_norm.weight": "69438f231e105e68216b078bdeb35a7cdc8b12c4e2845e18ecf4c8d361d6a321",
+  "blk.4.post_ffw_norm.weight": "0fd535da78bcf2b32c95b05b2b83dc49817393765be90d8cc1ed3d56f47b68ec",
+  "blk.5.attn_k.weight": "0166eb3c6d20dcf3d3c169e94caa8dee057535bb525e29f698fb6f8844f18a6c",
+  "blk.5.attn_norm.weight": "a7808f27f164023d5cde2be00fc23cac6c71aa0ddeb60bc23e12411b80087672",
+  "blk.5.attn_output.weight": "8b65b2027a0842b68c5308f91d6a31de9599d794157d77df8418b19f9e0d9334",
+  "blk.5.attn_q.weight": "966bc626ef2c2394d872087a41c126bb1b67d1d5f6de920204ef5e5b16c34003",
+  "blk.5.attn_v.weight": "9a362aef3f4437fbf0ef6e1ba785f3329c3db2960f93fe36547d2795e9c254ea",
+  "blk.5.ffn_down.weight": "63e53541d34197720c06f297aa8142ac6b6eec002c7987b296f26e8b1400f931",
+  "blk.5.ffn_gate.weight": "d9591fdd32f783e0fc26e20d5d587ee8971ac8ae2e4c818c6eac1c125c7c7f37",
+  "blk.5.ffn_norm.weight": "677334cc60ecce3a7f4ab3acda15d359353d7358872f614ad8914e3780e9fc6e",
+  "blk.5.ffn_up.weight": "a63764110e1c655ffbd55af0669b2dfe4cc29d0e198d33a8e5426461b08a85f7",
+  "blk.5.post_attention_norm.weight": "c55499f859b2c0a7f5cabceaae47309a5ad38bc29d0f4a8db81f1357023162a9",
+  "blk.5.post_ffw_norm.weight": "82752754665f842418f3e302cb5f43d1e0504dcd124c4b8ddb77018b2c793837",
+  "blk.6.attn_k.weight": "e20a5f0d6c807273c8d491439566b428497ac02097cf0aa55e33748c28e14be6",
+  "blk.6.attn_norm.weight": "2c6ba42fd3c73d72073ced03a32dd28d70a89ed9bbbc8fea1ba03a7ade951e6c",
+  "blk.6.attn_output.weight": "4de7c5c2f4a133a266e17ed8c14c52959466b54cc7ab9e19f789a33b4850f284",
+  "blk.6.attn_q.weight": "56462d921800e6b8cd2213fef04c4ff16d728905cb2f4c58e966d0a053a3b0ae",
+  "blk.6.attn_v.weight": "b758dcbff769d6240c2245ede1dbc62c4170a67c77458e866312589220fe29af",
+  "blk.6.ffn_down.weight": "582247fb3c2bf687cbe9413fe18d18ad47bef4b65df7d78905e10335c6134764",
+  "blk.6.ffn_gate.weight": "3035444d5286aefb7a6d04e55bc27e1fac7cf895cd5be02319a431b8e047b4ae",
+  "blk.6.ffn_norm.weight": "e582d24c66e01b96faa20ce6adfda3d8583b11e809bff89969927398175e369a",
+  "blk.6.ffn_up.weight": "6f4b7bbfedeacf61a4866ae0616c4ba6c9e856662e8f00ae6aaec7f52c53e7b4",
+  "blk.6.post_attention_norm.weight": "8fe51b50bd677d21586aecab0b565c4bf9fa68ad50bfe366f45e8fea3c657ca8",
+  "blk.6.post_ffw_norm.weight": "81ba3cb4c2bf5c546b86855b7a885d3fafededc67eb3a35cd3598b03c9e26e65",
+  "blk.7.attn_k.weight": "2e044179cdcae0946708c86bfea7aa0391e1f7e2a09b33fca035d384cc3ca758",
+  "blk.7.attn_norm.weight": "94b48c546b046803c60e75a3acb17a356b710735989938021b565f68df9b4985",
+  "blk.7.attn_output.weight": "65709b4ad7a581f4d75793d39d4032a359f6bcc0c3835205242a0b99e5b66824",
+  "blk.7.attn_q.weight": "8ded993c95d1f7caf201ceb6fa035cd6ed6d351b50b999fa9355dfee9486cb5b",
+  "blk.7.attn_v.weight": "c92d5e2d2d48397542bc03bea25bf39154075e66c5bb1ead85188505aa04ae91",
+  "blk.7.ffn_down.weight": "e8ba8fb57208805ef1dc23cd7c86e9a2d1fb7c52c3940d292cd5bb2eb24b3fac",
+  "blk.7.ffn_gate.weight": "f0f06d6a2e06c5ac252083bc61d05c814e6289d3f4e4a87d2f06918254c02c36",
+  "blk.7.ffn_norm.weight": "ebf8ef775f72624148e09d68a4332187a7a5020c521fe0623da1cd3485ad33e0",
+  "blk.7.ffn_up.weight": "a554adc4fc7122c247c77670e169916ba1794c787b5be30a2b36705138f1f746",
+  "blk.7.post_attention_norm.weight": "3aa6bc21d85c3a0c12b964e82b12feaedfdd13130c3cd2229228e24e0967ebdf",
+  "blk.7.post_ffw_norm.weight": "508bc7b19ee8ff08f0007c890133a462fc57c7e72b16ee8f6dd64def264ef876",
+  "blk.8.attn_k.weight": "363c8e74056642fe9e7c2f3f9769d57319cd3fa0a6022810189ab8d894322885",
+  "blk.8.attn_norm.weight": "685b49a1f1acb169f4df0bdd8e3de6943f3033cebad14b898a72000595610d92",
+  "blk.8.attn_output.weight": "7bde571e4efef1c6a6143f0526721dfb59e0a0ea0e1a3616a322b2eb937efa48",
+  "blk.8.attn_q.weight": "fc993dbc1074c28a0e1d85e5ab2f4ea6a9c6c1affe7ee56027000a275daed9b6",
+  "blk.8.attn_v.weight": "281e8791d3aef9b3864f1cb054da0ae0c2fef4ce0a58b1bad8bc136b2fa0f62b",
+  "blk.8.ffn_down.weight": "b1164a2578a7f87ed99c2bbc76c5dfbbbc6a1a803605391acc3f320fc989ffd7",
+  "blk.8.ffn_gate.weight": "6b39a3b3aaaa79aee61416b54d62160b9258042650e61c6b47bc77c2dd17daf3",
+  "blk.8.ffn_norm.weight": "17ea1362c72da27f12bc936500492035bdef3fd8f940cb12b57f37d42ba8ecb1",
+  "blk.8.ffn_up.weight": "bc3a7c47afc440d2bdf8fbe9ddf2c9220467472c60c8b4ded8c0f181470ec96c",
+  "blk.8.post_attention_norm.weight": "5c506204e00411ef9c8b4134d40eedcc19fffe68dd0af7d7cc49dcabf2dfac7e",
+  "blk.8.post_ffw_norm.weight": "002faec235c3678864e2901eed275ce4e9dc229164a91c9cd4c965142ba62305",
+  "blk.9.attn_k.weight": "0bab39d8c237f1b6d0010db40467142625a9e6f2e0e4c49a56c12b41e4e0b1fa",
+  "blk.9.attn_norm.weight": "de5f38e873b17f07aa7598831b89cc1cae2c9bc3eb2e042ee9af059d2563e84e",
+  "blk.9.attn_output.weight": "8a8184702c25a62df9ff309c0c7badc8587208523b2be3e8fa90ce7080573e6f",
+  "blk.9.attn_q.weight": "7c961b2431b09ddf95377acd07201cb91bf13d9cd3ae0f2c25c7d6a0358d9f50",
+  "blk.9.attn_v.weight": "e22d240cb4743067033e659cbf210ebe2ebbab3e1dea6ccbe5eaa982382ca038",
+  "blk.9.ffn_down.weight": "a426f81210f03d6ad53277416e1fdcdf37d8065e4817613edaf6c67a343426be",
+  "blk.9.ffn_gate.weight": "a82eba825cb77b8e64f85ff99ede2fc71bc9b01751eeb17e9e6c246ee12ea62e",
+  "blk.9.ffn_norm.weight": "1a97f9b1302a3a326d534c5c3fed2db6db0ae45fd0edd381a3e4fc1c75d81030",
+  "blk.9.ffn_up.weight": "5f20bac2bbf03bb42adb92fbf99561651e1edda57e0b61935ac7f6c08c0ed7cb",
+  "blk.9.post_attention_norm.weight": "9f9866d13988e1946b1e1c80d9374a92a6e3be33748f8eaed3e126d1e1a4c796",
+  "blk.9.post_ffw_norm.weight": "a6896dbf698db4dbbe5dbf12417d4fd80e9cad0c539c858892ec0aa5b046bb58",
+  "blk.10.attn_k.weight": "ca8446e5d21ecd4e6a70dca8d321be480be4fba94d70cba065205436feb44270",
+  "blk.10.attn_norm.weight": "4f41fe290e8f21f63b82151b6cce94bf7318d121468816b0c58af0ff7c1658ab",
+  "blk.10.attn_output.weight": "c626d2e9681c5c941bbde43dddfae1a8d4986bf2be4470857bc8e8bd7f869044",
+  "blk.10.attn_q.weight": "1e61b210a13a429977325cf15d781ab77d604cfa862f4270329cbd94237d5835",
+  "blk.10.attn_v.weight": "8ff8d3e3f058ec3b35ada1057f2ed59c06494d0e0be6a8dc3ff9edf9f0e1a115",
+  "blk.10.ffn_down.weight": "bcebc04219f8081a5f483e58103c0ddbbbc631a0a54fd6dd9d55778e041f70ee",
+  "blk.10.ffn_gate.weight": "7a23a1e620ef871384ddf9611ccdcfb893fbf013cc203ac8e72f745420f1eea0",
+  "blk.10.ffn_norm.weight": "e3a375e43c349a1c6c66c22328e513cc1af3137fe839e43dc8e9be2f65914fd7",
+  "blk.10.ffn_up.weight": "5d182e7c94369194fca5f19cbbe668a999911e57f3d363bc7fb6088428700cb9",
+  "blk.10.post_attention_norm.weight": "b841c6308296e8984f3c5f549c6e3a242f4b3e19141e1f54cc08de9c46759c09",
+  "blk.10.post_ffw_norm.weight": "9d66fa05b5c940208f634f5053d809094c99a2a10a1d1e8847c8281fbd99fb49",
+  "blk.11.attn_k.weight": "14adf24ebb2bb17b336ca81cec3e690fd854782f4440ca6c66cc1d7e7bf1c850",
+  "blk.11.attn_norm.weight": "2d2213f311f50414702b5b34f22aafb9d9a0b6787243e7578562583dc40ad195",
+  "blk.11.attn_output.weight": "de1f14cc2a7fff00cf11b229f0576999205f17b9536e97abc9d6de3cc79a7884",
+  "blk.11.attn_q.weight": "2bcc5c147524003109ece0be08b89ac8b25baa71416ffa76573c6c052ffc6eea",
+  "blk.11.attn_v.weight": "2e6ab8573070c22dc1e0d7aebe4d52123226dacf7822dcce06fadbb38fb036a4",
+  "blk.11.ffn_down.weight": "1b86902f4e36868421e5228b9445051f8290b292df22a6d1af836dcecc1f25c3",
+  "blk.11.ffn_gate.weight": "e756e8081bd0a16aea4a9ef5076ad102113524f7a3d50a3a77aaa7f7938b63e8",
+  "blk.11.ffn_norm.weight": "6913887267be227cf9d1991a3dd8db2e7e74bb9b5fbdfcb9ac954fd7d7b95b3b",
+  "blk.11.ffn_up.weight": "619a3ac0609ebdf42c3fb2b6e4b1db48df79e6dd8418d7ab8f1bbff13d8a6a50",
+  "blk.11.post_attention_norm.weight": "e4b4ba92cef7b6a78407e8ab1b0307d47dac6c3df7b6817e28038317ff662d7e",
+  "blk.11.post_ffw_norm.weight": "40aceeec58cb855f0c158c9cc217168fcd5d0e735567d587217b1d78df17bc5f",
+  "blk.12.attn_k.weight": "c54c5a4d4892522022d1aa2204cfc624f0b4042caa536e678967316293fe5cb1",
+  "blk.12.attn_norm.weight": "7cd2ef58298569ffdf244d9b390f3917245276c8206e5780af5f96d8c0bbb446",
+  "blk.12.attn_output.weight": "85495ef9cc8b3deb21f741bde463ff6493acae2be51f02ecdeef952cbdec3375",
+  "blk.12.attn_q.weight": "d19383f83fd119bfb8c0280c9515705c11d8e7d502019fcf8f49efeef0d106d0",
+  "blk.12.attn_v.weight": "869ac669ba49531d9128892a0e27cef15de508ff40cdf80cc1681dde50d09204",
+  "blk.12.ffn_down.weight": "578f39f8f9fc2f09138afc884a952d7cc3a9a31de4216acd10e88e19e0b75f8c",
+  "blk.12.ffn_gate.weight": "e29a0186bc6c4a0720246306e922d3a83f777dadcf4ac80bad468287031cc8b5",
+  "blk.12.ffn_norm.weight": "e1ee95c6584b5cb57fcf1db8ce2bcc03aff91eb389238c094a61c00dde93d1f2",
+  "blk.12.ffn_up.weight": "2a826f06d7cdfb3edc6ae250ff44363ef77a2a9cdf96313e23a331b99ebfa17d",
+  "blk.12.post_attention_norm.weight": "4bafc7699b948d5cbc0d3e09b418b06c6abc4651a61ada9609d9a2f21c7e5607",
+  "blk.12.post_ffw_norm.weight": "bbb8c34a7176bb1a49f9fe2bacca0bd26b673d52c0835b2e90fa11f2962f077f",
+  "blk.13.attn_k.weight": "ffeefccfe8255d1b694382012ff4134eee5fec9d9491c8d0ff0a13832d1a37e8",
+  "blk.13.attn_norm.weight": "35713726529e3887c4135a88e86e8a4d7270ba5b9f2d1ab462622fbf40a7cdce",
+  "blk.13.attn_output.weight": "0d60b7c5cd71190a9ef4b873b0f516be15447c32d83914db2794b14592b0b460",
+  "blk.13.attn_q.weight": "8296069e65bef794cefc61257fc65789b3cb22955e30f3df129205e5041b2222",
+  "blk.13.attn_v.weight": "ca0f4ab9d16a748fc643a5c0c7a19826a811bf2a4e7316a8c935d4bf0ce8abc6",
+  "blk.13.ffn_down.weight": "d5514e0c8e7b3ed1cbcc1605eb5be1733b6ab3514cf8a0508fc72f7d05ed8bcb",
+  "blk.13.ffn_gate.weight": "8108e517a82e08a3aefbbd267bfa50a1668f92a76273280ce8a6bc1f6dd61521",
+  "blk.13.ffn_norm.weight": "5fcb6132d2134bf1f835b904a99820fa501dbc57d2224129f7098bf3cabc1d36",
+  "blk.13.ffn_up.weight": "6d744b7cd390a3cae3aa350dd379b81246acd056a2259996b6aaadece8465ccc",
+  "blk.13.post_attention_norm.weight": "e08b14698912509790e9575b8676971fbb0a4d82d719367e3756c0d0c4ab8cc0",
+  "blk.13.post_ffw_norm.weight": "2b196e4450fc5f1e7367b2cf7fe33a15fe919fbcdd861d11002346f16e980535",
+  "blk.14.attn_k.weight": "120e5f48d7268dfd9ab5f4bc9cc57a7cec63ea9635f56b80d435eb22936e9483",
+  "blk.14.attn_norm.weight": "146367bcce4db72cc894419a2e0145a6f533507dd68e4739c10ee480308c401f",
+  "blk.14.attn_output.weight": "720fa0165e756876c5cb6ad9e2780dd910390933f3f8849e5add5da04266650b",
+  "blk.14.attn_q.weight": "f5183466f56219ca1aca52d8b82c2d966a4198fea40fdd6b39f4d8b06ca2a6dd",
+  "blk.14.attn_v.weight": "24f8ea3d5512cd37c43c8329cb0da0c90d1895aef763ac2dcee3fe5157ec50a2",
+  "blk.14.ffn_down.weight": "e29960965b384ae5ab3d898a4dbaa8fddd28fa0e477ac28bcac49dec12a5ac67",
+  "blk.14.ffn_gate.weight": "6d0d6a74bfe9692e8f8eedff0fc34fc4fa1c8687794f35f2e2b033ab2d7510b8",
+  "blk.14.ffn_norm.weight": "f7036c1a9a71e046c9d2af16e9218fda5dbb0f7241ab44747abed1f0f9d602ca",
+  "blk.14.ffn_up.weight": "7d69ea1424007ffc9c12247dd0308c616e93ac02a59ec341cfa48f92d6ce3b10",
+  "blk.14.post_attention_norm.weight": "65b9712834d9445d4236bec362f3fb795c20d60c541b3dc6dbb7914d9b493e41",
+  "blk.14.post_ffw_norm.weight": "9c6a8da2e4e437d5cfdf3b9097e9f8b64bf07946a048badec20f4d374613f38f",
+  "blk.15.attn_k.weight": "864bc618303a0e4ee67fb1d5e751de61e936cd51e96669dd86f8cd08f2305045",
+  "blk.15.attn_norm.weight": "f9f4187da6eeadc2fc5921d8fe669741697d16c13d71e4aaeb73b82f50dc577e",
+  "blk.15.attn_output.weight": "ce2419a0b097036b2a31f2f4ad731d5814bcc2ef4c511786e24471e5eefd273b",
+  "blk.15.attn_q.weight": "9539db5a970d11ebe99722d1e13fcd635e250033630811efe583d2f97778e4a9",
+  "blk.15.attn_v.weight": "1c834b48ccd88adaeabb7d8bcb6be0bcd6d5ac1354ce88fc28f19a1a96b81ab3",
+  "blk.15.ffn_down.weight": "bc1f97a65dde6fa2c1e5397afb612266944b343f2eaa868b635ddd25829f8a42",
+  "blk.15.ffn_gate.weight": "1b14529d57056b79037f6cb5008132e62cc35992353b38dda59572274623103b",
+  "blk.15.ffn_norm.weight": "9af77458de9ee55c66f93865759f9c2c398557f94f3fa8fa6af30543d7339cde",
+  "blk.15.ffn_up.weight": "41d524a26b61a9595816b4fd53cf57ef50a702e4ef32933ff6136dca9136a267",
+  "blk.15.post_attention_norm.weight": "c60a03cd0e63a7db5c80015e58e9b97ba2208caa19f66a6fef5c4447eca900ce",
+  "blk.15.post_ffw_norm.weight": "34f7f9f96769215bbc3d17084df091864aef96a6645b7d0b3b7d9bd92f1a4b0b",
+  "blk.16.attn_k.weight": "7e27240d9f3a8c6cf0f4a980113d43234f514eadc3e3e1792b86efb29ffb1a6d",
+  "blk.16.attn_norm.weight": "af798acc0899282a30448edec48223b3e8efda177090273e612d8eca5e377301",
+  "blk.16.attn_output.weight": "79df39a3709d3d53e84146291e0944a7a653d06705293d9ccb5648dceadb432c",
+  "blk.16.attn_q.weight": "db58a1c3b83ad294804e5fd7321005719e200659173466df5a52a182b80b7165",
+  "blk.16.attn_v.weight": "2af6d48cbaeb225b5c1a704f76abd89c8ab1521417695b112b4dcc2cbd39b74d",
+  "blk.16.ffn_down.weight": "fc1c813eb5e7da3d6194569d6cb21602fc6eff2dc8e1b0eb753f2d5df148189c",
+  "blk.16.ffn_gate.weight": "7a80bcbc42464bd55df4814a6edbd7b5c153e0428323bbe49de55e2d2add33e7",
+  "blk.16.ffn_norm.weight": "2041685ee926d30f3f2ae4ec35b5688f1cd834167a6359a7d4057eac804c58b2",
+  "blk.16.ffn_up.weight": "8da4b718973ac1d43b928829bc45e062fd101984d6c98dd825bd7c5d08ebfbe3",
+  "blk.16.post_attention_norm.weight": "975c48fe680a6167438a106140a8872eee7765191f152d80e3b8ddf47693e095",
+  "blk.16.post_ffw_norm.weight": "4de2d4d483acfe4fc77860ea929025df2f4e15c10729413f36a18c94eaa6d689",
+  "blk.17.attn_k.weight": "f937e61f0af8c4cd98ee742648eb60e02e579683e21d421071295a3b70aebaad",
+  "blk.17.attn_norm.weight": "c3270583ed28b7e423f5b170c59113234f258169b93a867d9274f4c10b7cb115",
+  "blk.17.attn_output.weight": "b8c1150e81e685e539a5dcf2c19047a24eba2b281fabe166674b1d71ef4612ea",
+  "blk.17.attn_q.weight": "c255100ae2011e7dc7e3bf3bc3ccd96d859fbb98581cae993d7b82c1ba8e8b39",
+  "blk.17.attn_v.weight": "5830bb0a555984c6485348067f70b5d22ae337c011aa9248dac2ff4c95944551",
+  "blk.17.ffn_down.weight": "8ff9a7cccaa3776434a9d895aae4fb5c36c736bf2ec98784226b4c234940fbb0",
+  "blk.17.ffn_gate.weight": "1b52876739712831c272911533da206f407b46034a1a4ae8a88c1f96b6bd5747",
+  "blk.17.ffn_norm.weight": "d0e16ba5e87c91b545334e022058c7d03849665c3b1a6298771b656531366b66",
+  "blk.17.ffn_up.weight": "4dd6211d01dbebbe21052708eddc242b082a58b5f18ed16479e17987c1d3432e",
+  "blk.17.post_attention_norm.weight": "6f49c775c7417dade77ba8268a0f8441c1e5ec28b5d7e4dc5ed07a04d04600c8",
+  "blk.17.post_ffw_norm.weight": "b91a0bb2e6679e9c9be06ad323adae441d00a3d673efb19d7c4954be2aa84b27",
+  "blk.18.attn_k.weight": "22b565ace1b4da8b33865a58625be1d90beea9891f29686a69fa9cf7c93217db",
+  "blk.18.attn_norm.weight": "3e0160d7063c8753de65d2356a66648e47d921efdc5c917efb8209892120f8db",
+  "blk.18.attn_output.weight": "e3180f0bb4ca90b31e9b08158db38e332de62dfbaefe34aa94cc316409331e09",
+  "blk.18.attn_q.weight": "f3a5a83614c3ba7ea41cdd5b1b0819a241ee2a951a381ce4a9e001d3f700ed8f",
+  "blk.18.attn_v.weight": "f3350a5984fb951fc738adcf78147e6d812ff1c576670c460cafc99c253c1654",
+  "blk.18.ffn_down.weight": "9e9d09b13a33525e14bdaee6efc65c551ac7cf7680e534b940ab122a3a7c1ac9",
+  "blk.18.ffn_gate.weight": "ebaec8b4b578a2e8d815baac12f1675c208f80c68074d5a18288a2e1a60680ee",
+  "blk.18.ffn_norm.weight": "33e7687c53a242f2f8dc7093a491c97b18d4a5a8c14d183f02bd586a770f05aa",
+  "blk.18.ffn_up.weight": "78a1816662378ce56cc870e705174492781897b3afd2d4d97a51f10f2f2987c1",
+  "blk.18.post_attention_norm.weight": "a58dde3f12df3e94cbc27d87c8ea86f89af8a388a506446ff6758f05399b05fc",
+  "blk.18.post_ffw_norm.weight": "cebf90cc143577d483cca27b032dfd82031ee59bdf17c0e2cf60a0a3ad5bf996",
+  "blk.19.attn_k.weight": "4683375d0599ac9e2232196aae1e90af13a14cae26e865465de5c8e257bb2055",
+  "blk.19.attn_norm.weight": "f3eba936bfb1814bbcb0a1d62739eb66daac839df8c9c836fe0e94860df88525",
+  "blk.19.attn_output.weight": "51c0f01d38a9dcfe9bdbc4643576fab164c1d9e4b7168b7695c0ee55e6965667",
+  "blk.19.attn_q.weight": "28d15b69b8416f2e7ddc88fe381cb1e2ef2ad705fb1c268139ba96498cc74848",
+  "blk.19.attn_v.weight": "6860f1cd720638e63a981fa2c0b4db900129826bcb9823c9ddf9fb8b1b9f3383",
+  "blk.19.ffn_down.weight": "bc7f2d7827ee01c2dd41401c7b3b1700ad3a4ff620e8bb734f92630d342dcc7f",
+  "blk.19.ffn_gate.weight": "54d03ef69ba373fc410fbca8f1e34a565d58e4296d9a035ff7e48340b9c848e7",
+  "blk.19.ffn_norm.weight": "9178fc796a340ee6e8128ca74c0cb6203d1adbed6927af4e5ac7863da57affc7",
+  "blk.19.ffn_up.weight": "a77bd708026c6e83ad5c79c223278e74621bcf74a9641c7818d96b595daaad20",
+  "blk.19.post_attention_norm.weight": "ae94aa26f4c411bf9496a6fd4a6df64ee589ee1ae9a04b531d45acc95721e582",
+  "blk.19.post_ffw_norm.weight": "9ad210700edeef12133bdcff04bf1c7f62b49f6f4a9ba483c7cdc59857c24a5c",
+  "blk.20.attn_k.weight": "e35bce1e9f4a7a09ef34721f57ea38cfca68c272f52d923fe50af8308f66cfaa",
+  "blk.20.attn_norm.weight": "644800f6926fd34f233795c4dec1151a295d2138ca8cac33e3e48167d26f8b41",
+  "blk.20.attn_output.weight": "8d3758cd236471741e1ad66c0710cb79077dc8c7a3a292d35bc551c0c5abe627",
+  "blk.20.attn_q.weight": "c333b1f0f6f956b5d73891df10b1a0321e55fc31c40d623a24e1f52caa6a998b",
+  "blk.20.attn_v.weight": "8562b418d0c4868a050fb19fa3fcaf50a8cf1c669f537d666c80c7b3a04714e1",
+  "blk.20.ffn_down.weight": "97efb608ac44cc804198faec3ee66eafe56ced6b7ca5359700c6f1df75b7205e",
+  "blk.20.ffn_gate.weight": "5c61151d86f28415c73c73d90ec088c646cbe5c1640197caf58eb501ba7db293",
+  "blk.20.ffn_norm.weight": "24bbe0a701afd4bbeea65b3edde712b3cbb2281043bbc43dbf250582453116ed",
+  "blk.20.ffn_up.weight": "e170cf68e249566aa99eb6f6b265679bf9a5a6b76830ba24e7e130c2515910c4",
+  "blk.20.post_attention_norm.weight": "e092d751cfe20dbf2d348358f3b38397bd83e4ed94d6bbaa6bbaddcd902b2ac4",
+  "blk.20.post_ffw_norm.weight": "219a18a47dcba76e669e4322223a5a9227bd3db1de3fbd3d3cfb22e54a783c5a",
+  "blk.21.attn_k.weight": "c3a095ebddb42c63824f1c98da65263dc88e4d790a26aa1632840b44f5cc7cb1",
+  "blk.21.attn_norm.weight": "ef8bbaded5fbc45ad9cf3985ae02174524e7090fe6362811124f942ef643bec7",
+  "blk.21.attn_output.weight": "668f018aba72baac6252aa3ad58569ddd55ab751a0dd8d7bcc9fb9b6efb4bf53",
+  "blk.21.attn_q.weight": "e759c65663089f3bbbd51847934c185e680c82f1249065d5d487da638e519e6d",
+  "blk.21.attn_v.weight": "2ff57762686cf9ba1f5a6be76503454b97556ce67f4ac98254bd0562231197ba",
+  "blk.21.ffn_down.weight": "3fd106556fb721b1c28ae3f4026bc83eb1b08ed910f2ba5f466c6b5f327d91cb",
+  "blk.21.ffn_gate.weight": "338022d882f4b6619e8054a6fb909696fa3eef3013cf69b65c3cacdfc5b9e42c",
+  "blk.21.ffn_norm.weight": "1e77660c23a3f9653ee721a863d1960f773d87437cabc4dc0a6e17ee3d4e5e44",
+  "blk.21.ffn_up.weight": "7d31b20fbc2e6eba8f350f170069dc36f0cb12f68fbc4206ec5022a74085ebcb",
+  "blk.21.post_attention_norm.weight": "9638bae8d8bdcd7ed68da282979cd84a07c41ff9cabcaea94ebc846a1803db23",
+  "blk.21.post_ffw_norm.weight": "d622ef11115fe0cbe04b727d5a3b6371e7f39bf08c8d5eb9bc6da52e3f3cfb9d",
+  "blk.22.attn_k.weight": "5c321cb29deffbe57de200dd206a62005f1e80acb86c4fd2349dd44c8d3594fd",
+  "blk.22.attn_norm.weight": "198d949705d7170a331d75889d8c7500c3635254dac2cc6aa4dc35d556584536",
+  "blk.22.attn_output.weight": "19805cd5d7025b457e5d41d70db8b3fd63c2dd0e4a94d3ef1704d50ef4e749e8",
+  "blk.22.attn_q.weight": "177836cd583fc87405975ddc21ebfebdaa090a0363799664c72caa3da851ae2c",
+  "blk.22.attn_v.weight": "fea255692483e30d0108f9e4e250eb3ed7dbda8d83f499b06519b8c223ae6096",
+  "blk.22.ffn_down.weight": "00cb8939f03e5817d6d412de8cf2c923c9568d5493e382cec7faf5718fb034eb",
+  "blk.22.ffn_gate.weight": "b0591065b91281b2fbd8a9567f3568d40479f680e1f0a29e27ae213f37642489",
+  "blk.22.ffn_norm.weight": "96b5c5d0737c2ceb8fc869f54adb9e5f46e28cb7b177c40f49fa926b923c00f8",
+  "blk.22.ffn_up.weight": "81f472185b24344ab0594ea8246cc6e200e0dc1cab4943e74fbe4ca19d5a9701",
+  "blk.22.post_attention_norm.weight": "27fa9aa6260aa3071e0391e1a1d49322dcb6e8072315b8a9b7064087108dbd06",
+  "blk.22.post_ffw_norm.weight": "f37e1dcd7f643d9545675ffe9dc527a11eba86eb204989c2f44f636b266d896a",
+  "blk.23.attn_k.weight": "5d82f36658a56c3f94d0bb2d61f65509c966fa6568f81812e0d3e338b380ef8c",
+  "blk.23.attn_norm.weight": "b7983f88d9cad88bc88a528923e6da592ad20e699965b223ebc10840fe1f4fec",
+  "blk.23.attn_output.weight": "59f97f80f430d71606aab0158a195aed29ccd3405e6c0a5c41c809be8eb01898",
+  "blk.23.attn_q.weight": "53ac4789fe958919cc02ea4222bcd64c0ea1b4baa54304bff46635bdf42f7490",
+  "blk.23.attn_v.weight": "ec8abe09b9e84dbb52c7a068094657c6d3c62fe551ba8d7c3a3f23da622e9756",
+  "blk.23.ffn_down.weight": "3cf547eccb1b82aa64f208cee9682d7f558ca84e0aead7d9d3d1420d90f3d992",
+  "blk.23.ffn_gate.weight": "366aa2486d911ba81eb519119e13807deacf7e9908bc1975a2a63e00d6b10124",
+  "blk.23.ffn_norm.weight": "6d1d4a4af34bb7dc090ac87d6457d398c3e0fb68bd2e2b60b099dc318b6cfac3",
+  "blk.23.ffn_up.weight": "53f76692e253f5d2420b3f200c731b9f3b7a83e379920b4a067c729b4674aa4d",
+  "blk.23.post_attention_norm.weight": "7c952fa0efa76b3f048c8c4c9e8dcb5e3724d231327eda6423a34d3f3d3367de",
+  "blk.23.post_ffw_norm.weight": "7ab188cfe61f0a91b40309a0ab6bfa99f19d0ff2a37b6ac10e5f0c7f44eb5270",
+  "blk.24.attn_k.weight": "225798792f9bfdd10eff0505ebe61e0aad0209c17b431f6044ee7968ffe8c198",
+  "blk.24.attn_norm.weight": "635e3c1ebf5219bbebfc40ef164bc32d2b726ef595a94da64ac524ae878e2915",
+  "blk.24.attn_output.weight": "482f5bb2db8d9ed22b253d9a3296333b239efe698e5992e5d77e7e12dc2a5cf5",
+  "blk.24.attn_q.weight": "43805bbccddb65d58fffc4be9b5c374d4e1df1395ec1e1ffb4bcff03e98d5adb",
+  "blk.24.attn_v.weight": "fa741af54b4a3b1775d32f59134756090c5df2e7345a12a2d8db94fe289667a7",
+  "blk.24.ffn_down.weight": "83c6351e3162626b276f524a57836144625c2556dbe321b57cbd8fd486a68fab",
+  "blk.24.ffn_gate.weight": "fbe66be0d84d12cea5176cc7eaef64382ffc7324cd9d6266a3342dc43442f2ac",
+  "blk.24.ffn_norm.weight": "77c1445a8639ad24938bdf0280233eea2362d47391421833dfa72ec756dfc1e8",
+  "blk.24.ffn_up.weight": "78235ac729ee23c1cf1ae543751e3af32776d8808cee6e529c2a625a1f027654",
+  "blk.24.post_attention_norm.weight": "161f71b6d07628d43e4ae51a4c9088ec6ca2db123a17986a14505d83fdd04dad",
+  "blk.24.post_ffw_norm.weight": "cf1ba692aa683368b02ac413e69b2521b98c69a5274eacbb54165b53bf38a8b2",
+  "blk.25.attn_k.weight": "057a56bd8c8d2b41608d1f71faa3052902152ddf85e47669ad950c1c3e77c33f",
+  "blk.25.attn_norm.weight": "b7179fe02c334da556ddcf6c1b502245639a728c4cbba8b552d8e1df4565ee9d",
+  "blk.25.attn_output.weight": "4fed8b05b08a0ff75ffd022701bbeb52f17b23d09332a1ddcba737244bd0d3b0",
+  "blk.25.attn_q.weight": "c52e99f5d38bf7538d6106a0bbf38ac6dc6296bca9a3f849afa384ea67b4af01",
+  "blk.25.attn_v.weight": "c49c23d8e1cfa6a8eb971eb69942204890c6d7d830dc8774c84b108a80598912",
+  "blk.25.ffn_down.weight": "c08d4dc8412b19fdc870c164b83c341b236ec6fe7bb4a9bcfe0dc100faa20286",
+  "blk.25.ffn_gate.weight": "1a4cb3f36735d59181721471452807903006539e5e1b5ceb4f72d1d7ae134127",
+  "blk.25.ffn_norm.weight": "8fd6bd0dcec5198761525a36992a57c9ec5e9da60a22092839a84ae8c4e87f26",
+  "blk.25.ffn_up.weight": "3a00f39bdd5f31dc5e3b281d2002e1ac4f2475d49a0ac1d7720a25b377dcd04a",
+  "blk.25.post_attention_norm.weight": "e5f31a648612c859b6d21c9ee426e87a86cb1973dfdd86276c767371d9cef5ad",
+  "blk.25.post_ffw_norm.weight": "553c3bd774922c99c2384380a142d019881d30dbf0fe3bf9430dabfb3f6cbd33",
+  "output_norm.weight": "49445c4585ab0a8135717a0bdb1cda4a062a030177d0119561d91542aec5744b"
+}

docs/faq.md

@@ -194,6 +194,8 @@ Refer to the section [above](#how-do-i-configure-ollama-server) for how to set e
 
 If a different directory needs to be used, set the environment variable `OLLAMA_MODELS` to the chosen directory.
 
+> Note: on Linux using the standard installer, the `ollama` user needs read and write access to the specified directory. To assign the directory to the `ollama` user run `sudo chown -R ollama:ollama <directory>`.
+
 Refer to the section [above](#how-do-i-configure-ollama-server) for how to set environment variables on your platform.
 
 ## How can I use Ollama in Visual Studio Code?

docs/gpu.md

@@ -10,7 +10,7 @@ Check your compute compatibility to see if your card is supported:
 | 9.0                | NVIDIA              | `H100`                                                                                                      |
 | 8.9                | GeForce RTX 40xx    | `RTX 4090` `RTX 4080 SUPER` `RTX 4080` `RTX 4070 Ti SUPER` `RTX 4070 Ti` `RTX 4070 SUPER` `RTX 4070` `RTX 4060 Ti` `RTX 4060`  |
 |                    | NVIDIA Professional | `L4` `L40` `RTX 6000`                                                                                       |
-| 8.6                | GeForce RTX 30xx    | `RTX 3090 Ti` `RTX 3090` `RTX 3080 Ti` `RTX 3080` `RTX 3070 Ti` `RTX 3070` `RTX 3060 Ti` `RTX 3060`         |
+| 8.6                | GeForce RTX 30xx    | `RTX 3090 Ti` `RTX 3090` `RTX 3080 Ti` `RTX 3080` `RTX 3070 Ti` `RTX 3070` `RTX 3060 Ti` `RTX 3060` `RTX 3050 Ti` `RTX 3050`   |
 |                    | NVIDIA Professional | `A40` `RTX A6000` `RTX A5000` `RTX A4000` `RTX A3000` `RTX A2000` `A10` `A16` `A2`                          |
 | 8.0                | NVIDIA              | `A100` `A30`                                                                                                |
 | 7.5                | GeForce GTX/RTX     | `GTX 1650 Ti` `TITAN RTX` `RTX 2080 Ti` `RTX 2080` `RTX 2070` `RTX 2060`                                    |

docs/linux.md

@@ -1,44 +1,59 @@
-# Ollama on Linux
+# Linux
 
 ## Install
 
-Install Ollama running this one-liner:
+To install Ollama, run the following command:
 
->
+```shell
-
-```bash
 curl -fsSL https://ollama.com/install.sh | sh
 ```
 
-## AMD Radeon GPU support
-
-While AMD has contributed the `amdgpu` driver upstream to the official linux
-kernel source, the version is older and may not support all ROCm features. We
-recommend you install the latest driver from
-https://www.amd.com/en/support/linux-drivers for best support of your Radeon
-GPU.
-
 ## Manual install
 
-### Download `ollama`
+Download and extract the package:
 
-Download and extract the Linux package:
+```shell
-
+curl -L https://ollama.com/download/ollama-linux-amd64.tgz -o ollama-linux-amd64.tgz
-```bash
+sudo tar -C /usr -xzf ollama-linux-amd64.tgz
-curl -fsSL https://ollama.com/download/ollama-linux-amd64.tgz | sudo tar zx -C /usr
 ```
 
-If you have an AMD GPU, also download and extract the ROCm package into the same location
+Start Ollama:
-```bash
+
-curl -fsSL https://ollama.com/download/ollama-linux-amd64-rocm.tgz | sudo tar zx -C /usr
+```shell
+ollama serve
+```
+
+In another terminal, verify that Ollama is running:
+
+```shell
+ollama -v
+```
+
+### AMD GPU install
+
+If you have an AMD GPU, also download and extract the additional ROCm package:
+
+```shell
+curl -L https://ollama.com/download/ollama-linux-amd64-rocm.tgz -o ollama-linux-amd64-rocm.tgz
+sudo tar -C /usr -xzf ollama-linux-amd64-rocm.tgz
+```
+
+### ARM64 install
+
+Download and extract the ARM64-specific package:
+
+```shell
+curl -L https://ollama.com/download/ollama-linux-arm64.tgz -o ollama-linux-arm64.tgz
+sudo tar -C /usr -xzf ollama-linux-arm64.tgz
 ```
 
 ### Adding Ollama as a startup service (recommended)
 
-Create a user for Ollama:
+Create a user and group for Ollama:
 
-```bash
+```shell
-sudo useradd -r -s /bin/false -m -d /usr/share/ollama ollama
+sudo useradd -r -s /bin/false -U -m -d /usr/share/ollama ollama
+sudo usermod -a -G ollama $(whoami)
 ```
 
 Create a service file in `/etc/systemd/system/ollama.service`:
@@ -54,6 +69,7 @@ User=ollama
 Group=ollama
 Restart=always
 RestartSec=3
+Environment="PATH=$PATH"
 
 [Install]
 WantedBy=default.target
@@ -61,46 +77,54 @@ WantedBy=default.target
 
 Then start the service:
 
-```bash
+```shell
 sudo systemctl daemon-reload
 sudo systemctl enable ollama
 ```
 
-### Install CUDA drivers (optional – for Nvidia GPUs)
+### Install CUDA drivers (optional)
 
 [Download and install](https://developer.nvidia.com/cuda-downloads) CUDA.
 
 Verify that the drivers are installed by running the following command, which should print details about your GPU:
 
-```bash
+```shell
 nvidia-smi
 ```
 
-### Install ROCm (optional - for Radeon GPUs)
+### Install AMD ROCm drivers (optional)
-[Download and Install](https://rocm.docs.amd.com/projects/install-on-linux/en/latest/tutorial/quick-start.html)
 
-Make sure to install ROCm v6
+[Download and Install](https://rocm.docs.amd.com/projects/install-on-linux/en/latest/tutorial/quick-start.html) ROCm v6.
 
 ### Start Ollama
 
-Start Ollama using `systemd`:
+Start Ollama and verify it is running:
 
-```bash
+```shell
 sudo systemctl start ollama
+sudo systemctl status ollama
 ```
 
-## Update
+> [!NOTE]
+> While AMD has contributed the `amdgpu` driver upstream to the official linux
+> kernel source, the version is older and may not support all ROCm features. We
+> recommend you install the latest driver from
+> https://www.amd.com/en/support/linux-drivers for best support of your Radeon
+> GPU.
 
-Update ollama by running the install script again:
+## Updating
 
-```bash
+Update Ollama by running the install script again:
+
+```shell
 curl -fsSL https://ollama.com/install.sh | sh
 ```
 
-Or by downloading the ollama binary:
+Or by re-downloading Ollama:
 
-```bash
+```shell
-curl -fsSL https://ollama.com/download/ollama-linux-amd64.tgz | sudo tar zx -C /usr
+curl -L https://ollama.com/download/ollama-linux-amd64.tgz -o ollama-linux-amd64.tgz
+sudo tar -C /usr -xzf ollama-linux-amd64.tgz
 ```
 
 ## Installing specific versions
@@ -109,15 +133,15 @@ Use `OLLAMA_VERSION` environment variable with the install script to install a s
 
 For example:
 
-```
+```shell
-curl -fsSL https://ollama.com/install.sh | OLLAMA_VERSION=0.1.32 sh
+curl -fsSL https://ollama.com/install.sh | OLLAMA_VERSION=0.3.9 sh
 ```
 
 ## Viewing logs
 
 To view logs of Ollama running as a startup service, run:
 
-```bash
+```shell
 journalctl -e -u ollama
 ```
 
@@ -125,7 +149,7 @@ journalctl -e -u ollama
 
 Remove the ollama service:
 
-```bash
+```shell
 sudo systemctl stop ollama
 sudo systemctl disable ollama
 sudo rm /etc/systemd/system/ollama.service
@@ -133,13 +157,13 @@ sudo rm /etc/systemd/system/ollama.service
 
 Remove the ollama binary from your bin directory (either `/usr/local/bin`, `/usr/bin`, or `/bin`):
 
-```bash
+```shell
 sudo rm $(which ollama)
 ```
 
 Remove the downloaded models and Ollama service user and group:
 
-```bash
+```shell
 sudo rm -r /usr/share/ollama
 sudo userdel ollama
 sudo groupdel ollama

docs/windows.md

@@ -48,6 +48,9 @@ the explorer window by hitting `<cmd>+R` and type in:
 - `explorer %HOMEPATH%\.ollama` contains models and configuration
 - `explorer %TEMP%` contains temporary executable files in one or more `ollama*` directories
 
+## Uninstall
+
+The Ollama Windows installer registers an Uninstaller application.  Under `Add or remove programs` in Windows Settings, you can uninstall Ollama.
 
 ## Standalone CLI
 

envconfig/config.go

@@ -112,6 +112,26 @@ func KeepAlive() (keepAlive time.Duration) {
 	return keepAlive
 }
 
+// LoadTimeout returns the duration for stall detection during model loads. LoadTimeout can be configured via the OLLAMA_LOAD_TIMEOUT environment variable.
+// Zero or Negative values are treated as infinite.
+// Default is 5 minutes.
+func LoadTimeout() (loadTimeout time.Duration) {
+	loadTimeout = 5 * time.Minute
+	if s := Var("OLLAMA_LOAD_TIMEOUT"); s != "" {
+		if d, err := time.ParseDuration(s); err == nil {
+			loadTimeout = d
+		} else if n, err := strconv.ParseInt(s, 10, 64); err == nil {
+			loadTimeout = time.Duration(n) * time.Second
+		}
+	}
+
+	if loadTimeout <= 0 {
+		return time.Duration(math.MaxInt64)
+	}
+
+	return loadTimeout
+}
+
 func Bool(k string) func() bool {
 	return func() bool {
 		if s := Var(k); s != "" {
@@ -231,6 +251,23 @@ var (
 	MaxVRAM = Uint("OLLAMA_MAX_VRAM", 0)
 )
 
+func Uint64(key string, defaultValue uint64) func() uint64 {
+	return func() uint64 {
+		if s := Var(key); s != "" {
+			if n, err := strconv.ParseUint(s, 10, 64); err != nil {
+				slog.Warn("invalid environment variable, using default", "key", key, "value", s, "default", defaultValue)
+			} else {
+				return n
+			}
+		}
+
+		return defaultValue
+	}
+}
+
+// Set aside VRAM per GPU
+var GpuOverhead = Uint64("OLLAMA_GPU_OVERHEAD", 0)
+
 type EnvVar struct {
 	Name        string
 	Value       any
@@ -241,9 +278,11 @@ func AsMap() map[string]EnvVar {
 	ret := map[string]EnvVar{
 		"OLLAMA_DEBUG":             {"OLLAMA_DEBUG", Debug(), "Show additional debug information (e.g. OLLAMA_DEBUG=1)"},
 		"OLLAMA_FLASH_ATTENTION":   {"OLLAMA_FLASH_ATTENTION", FlashAttention(), "Enabled flash attention"},
+		"OLLAMA_GPU_OVERHEAD":      {"OLLAMA_GPU_OVERHEAD", GpuOverhead(), "Reserve a portion of VRAM per GPU (bytes)"},
 		"OLLAMA_HOST":              {"OLLAMA_HOST", Host(), "IP Address for the ollama server (default 127.0.0.1:11434)"},
 		"OLLAMA_KEEP_ALIVE":        {"OLLAMA_KEEP_ALIVE", KeepAlive(), "The duration that models stay loaded in memory (default \"5m\")"},
 		"OLLAMA_LLM_LIBRARY":       {"OLLAMA_LLM_LIBRARY", LLMLibrary(), "Set LLM library to bypass autodetection"},
+		"OLLAMA_LOAD_TIMEOUT":      {"OLLAMA_LOAD_TIMEOUT", LoadTimeout(), "How long to allow model loads to stall before giving up (default \"5m\")"},
 		"OLLAMA_MAX_LOADED_MODELS": {"OLLAMA_MAX_LOADED_MODELS", MaxRunners(), "Maximum number of loaded models per GPU"},
 		"OLLAMA_MAX_QUEUE":         {"OLLAMA_MAX_QUEUE", MaxQueue(), "Maximum number of queued requests"},
 		"OLLAMA_MODELS":            {"OLLAMA_MODELS", Models(), "The path to the models directory"},

envconfig/config_test.go

@@ -215,6 +215,40 @@ func TestKeepAlive(t *testing.T) {
 	}
 }
 
+func TestLoadTimeout(t *testing.T) {
+	defaultTimeout := 5 * time.Minute
+	cases := map[string]time.Duration{
+		"":       defaultTimeout,
+		"1s":     time.Second,
+		"1m":     time.Minute,
+		"1h":     time.Hour,
+		"5m0s":   defaultTimeout,
+		"1h2m3s": 1*time.Hour + 2*time.Minute + 3*time.Second,
+		"0":      time.Duration(math.MaxInt64),
+		"60":     60 * time.Second,
+		"120":    2 * time.Minute,
+		"3600":   time.Hour,
+		"-0":     time.Duration(math.MaxInt64),
+		"-1":     time.Duration(math.MaxInt64),
+		"-1m":    time.Duration(math.MaxInt64),
+		// invalid values
+		" ":   defaultTimeout,
+		"???": defaultTimeout,
+		"1d":  defaultTimeout,
+		"1y":  defaultTimeout,
+		"1w":  defaultTimeout,
+	}
+
+	for tt, expect := range cases {
+		t.Run(tt, func(t *testing.T) {
+			t.Setenv("OLLAMA_LOAD_TIMEOUT", tt)
+			if actual := LoadTimeout(); actual != expect {
+				t.Errorf("%s: expected %s, got %s", tt, expect, actual)
+			}
+		})
+	}
+}
+
 func TestVar(t *testing.T) {
 	cases := map[string]string{
 		"value":       "value",

gpu/cuda_common.go

@@ -57,7 +57,7 @@ func cudaVariant(gpuInfo CudaGPUInfo) string {
 		}
 	}
 
-	if gpuInfo.computeMajor < 6 || gpuInfo.DriverMajor < 12 {
+	if gpuInfo.computeMajor < 6 || gpuInfo.DriverMajor < 12 || (gpuInfo.DriverMajor == 12 && gpuInfo.DriverMinor == 0) {
 		return "v11"
 	}
 	return "v12"

llm/ext_server/CMakeLists.txt

@@ -2,7 +2,7 @@ set(TARGET ollama_llama_server)
 option(LLAMA_SERVER_VERBOSE "Build verbose logging option for Server" ON)
 set(LLAMA_SERVER_LDFLAGS $ENV{LLAMA_SERVER_LDFLAGS})
 include_directories(${CMAKE_CURRENT_SOURCE_DIR})
-add_executable(${TARGET} server.cpp utils.hpp json.hpp httplib.h)
+add_executable(${TARGET} server.cpp utils.hpp httplib.h)
 install(TARGETS ${TARGET} RUNTIME)
 target_compile_definitions(${TARGET} PRIVATE
     SERVER_VERBOSE=$<BOOL:${LLAMA_SERVER_VERBOSE}>

llm/ext_server/server.cpp

@@ -262,7 +262,7 @@ struct server_slot {
        char buffer[512];
         double t_token = t_prompt_processing / n_prompt_tokens_processed;
         double n_tokens_second = 1e3 / t_prompt_processing * n_prompt_tokens_processed;
-        sprintf(buffer, "prompt eval time     = %10.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)",
+        snprintf(buffer, sizeof(buffer), "prompt eval time     = %10.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)",
                 t_prompt_processing, n_prompt_tokens_processed,
                 t_token, n_tokens_second);
         LOG_DEBUG(buffer, {
@@ -276,7 +276,7 @@ struct server_slot {
 
         t_token = t_token_generation / n_decoded;
         n_tokens_second = 1e3 / t_token_generation * n_decoded;
-        sprintf(buffer, "generation eval time = %10.2f ms / %5d runs   (%8.2f ms per token, %8.2f tokens per second)",
+        snprintf(buffer, sizeof(buffer), "generation eval time = %10.2f ms / %5d runs   (%8.2f ms per token, %8.2f tokens per second)",
                 t_token_generation, n_decoded,
                 t_token, n_tokens_second);
         LOG_DEBUG(buffer, {
@@ -288,7 +288,7 @@ struct server_slot {
             {"n_tokens_second",    n_tokens_second},
         });
 
-        sprintf(buffer, "          total time = %10.2f ms", t_prompt_processing + t_token_generation);
+        snprintf(buffer, sizeof(buffer), "          total time = %10.2f ms", t_prompt_processing + t_token_generation);
         LOG_DEBUG(buffer, {
             {"slot_id",             id},
             {"task_id",             task_id},
@@ -425,7 +425,7 @@ struct llama_server_context
 
         n_ctx = llama_n_ctx(ctx);
 
-        add_bos_token = llama_should_add_bos_token(model);
+        add_bos_token = llama_add_bos_token(model);
 
         return true;
     }
@@ -1031,7 +1031,7 @@ struct llama_server_context
                 continue;
             }
 
-            if (!llava_image_embed_make_with_clip_img(clp_ctx, params.n_threads, img.img_data, &img.image_embedding, &img.image_tokens)) {
+            if (!llava_image_embed_make_with_clip_img(clp_ctx, params.cpuparams.n_threads, img.img_data, &img.image_embedding, &img.image_tokens)) {
                 LOG_TEE("Error processing the given image");
                 return false;
             }
@@ -2014,7 +2014,7 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
     printf("options:\n");
     printf("  -h, --help                show this help message and exit\n");
     printf("  -v, --verbose             verbose output (default: %s)\n", server_verbose ? "enabled" : "disabled");
-    printf("  -t N, --threads N         number of threads to use during computation (default: %d)\n", params.n_threads);
+    printf("  -t N, --threads N         number of threads to use during computation (default: %d)\n", params.cpuparams.n_threads);
     printf("  -tb N, --threads-batch N  number of threads to use during batch and prompt processing (default: same as --threads)\n");
     printf("  --threads-http N          number of threads in the http server pool to process requests (default: max(hardware concurrency - 1, --parallel N + 2))\n");
     printf("  -c N, --ctx-size N        size of the prompt context (default: %d)\n", params.n_ctx);
@@ -2287,7 +2287,7 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, g
                 invalid_param = true;
                 break;
             }
-            params.n_threads = std::stoi(argv[i]);
+            params.cpuparams.n_threads = std::stoi(argv[i]);
         }
         else if (arg == "--grp-attn-n" || arg == "-gan")
         {
@@ -2315,7 +2315,7 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, g
                 invalid_param = true;
                 break;
             }
-            params.n_threads_batch = std::stoi(argv[i]);
+            params.cpuparams_batch.n_threads = std::stoi(argv[i]);
         }
         else if (arg == "--threads-http")
         {
@@ -2626,6 +2626,11 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, g
         params.kv_overrides.back().key[0] = 0;
     }
 
+    postprocess_cpu_params(params.cpuparams, nullptr);
+    postprocess_cpu_params(params.cpuparams_batch, &params.cpuparams);
+    postprocess_cpu_params(params.draft_cpuparams, &params.cpuparams);
+    postprocess_cpu_params(params.draft_cpuparams_batch, &params.cpuparams_batch);
+
     if (invalid_param)
     {
         fprintf(stderr, "error: invalid parameter for argument: %s\n", arg.c_str());
@@ -2775,8 +2780,8 @@ int main(int argc, char **argv) {
                             {"commit", LLAMA_COMMIT}});
 
     LOG_INFO("system info", {
-                                {"n_threads", params.n_threads},
+                                {"n_threads", params.cpuparams.n_threads},
-                                {"n_threads_batch", params.n_threads_batch},
+                                {"n_threads_batch", params.cpuparams_batch.n_threads},
                                 {"total_threads", std::thread::hardware_concurrency()},
                                 {"system_info", llama_print_system_info()},
                             });

llm/generate/gen_darwin.sh

@@ -19,7 +19,7 @@ sign() {
     fi
 }
 
-COMMON_DARWIN_DEFS="-DBUILD_SHARED_LIBS=off -DCMAKE_OSX_DEPLOYMENT_TARGET=11.3 -DLLAMA_METAL_MACOSX_VERSION_MIN=11.3 -DCMAKE_SYSTEM_NAME=Darwin -DGGML_METAL_EMBED_LIBRARY=on -DGGML_OPENMP=off"
+COMMON_DARWIN_DEFS="-DBUILD_SHARED_LIBS=off -DCMAKE_OSX_DEPLOYMENT_TARGET=11.3 -DGGML_METAL_MACOSX_VERSION_MIN=11.3 -DCMAKE_SYSTEM_NAME=Darwin -DGGML_METAL_EMBED_LIBRARY=on -DGGML_OPENMP=off"
 
 case "${GOARCH}" in
 "amd64")

llm/ggml.go

@@ -360,11 +360,13 @@ func (llm GGML) GraphSize(context, batch uint64) (partialOffload, fullOffload ui
 
 	switch llm.KV().Architecture() {
 	case "llama":
-		fullOffload = 4 * batch * (1 + 4*embedding + context*(1+heads))
+		fullOffload = max(
+			4*batch*(1+4*embedding+context*(1+heads)),
+			4*batch*(embedding+vocab),
+		)
 
 		partialOffload = 4 * batch * embedding
 		partialOffload += max(
-			// 4*batch*(4+6*embedding+context*(2*heads)+llm.KV().GQA()),
 			4*batch*(1+embedding+max(context, embedding))+embedding*embedding*9/16+4*context*(batch*heads+embeddingHeads*headsKV),
 			4*batch*(embedding+vocab)+embedding*vocab*105/128,
 		)

llm/llama.cpp

@@ -1 +1 @@
-Subproject commit 1e6f6554aa11fa10160a5fda689e736c3c34169f
+Subproject commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177

llm/memory.go

@@ -7,6 +7,7 @@ import (
 	"strings"
 
 	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/format"
 	"github.com/ollama/ollama/gpu"
 )
@@ -94,6 +95,7 @@ func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts
 	// Overflow that didn't fit into the GPU
 	var overflow uint64
 
+	overhead := envconfig.GpuOverhead()
 	availableList := make([]string, len(gpus))
 	for i, gpu := range gpus {
 		availableList[i] = format.HumanBytes2(gpu.FreeMemory)
@@ -164,8 +166,22 @@ func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts
 			gzo = gpuZeroOverhead
 		}
 		// Only include GPUs that can fit the graph, gpu minimum, the layer buffer and at least more layer
-		if gpus[i].FreeMemory < gzo+max(graphPartialOffload, graphFullOffload)+gpus[i].MinimumMemory+2*layerSize {
+		if (gpus[i].FreeMemory - overhead) < gzo+max(graphPartialOffload, graphFullOffload)+gpus[i].MinimumMemory+2*layerSize {
-			slog.Debug("gpu has too little memory to allocate any layers", "gpu", gpus[i])
+			slog.Debug("gpu has too little memory to allocate any layers",
+				"id", gpus[i].ID,
+				"library", gpus[i].Library,
+				"variant", gpus[i].Variant,
+				"compute", gpus[i].Compute,
+				"driver", fmt.Sprintf("%d.%d", gpus[i].DriverMajor, gpus[i].DriverMinor),
+				"name", gpus[i].Name,
+				"total", format.HumanBytes2(gpus[i].TotalMemory),
+				"available", format.HumanBytes2(gpus[i].FreeMemory),
+				"minimum_memory", gpus[i].MinimumMemory,
+				"layer_size", format.HumanBytes2(layerSize),
+				"gpu_zer_overhead", format.HumanBytes2(gzo),
+				"partial_offload", format.HumanBytes2(graphPartialOffload),
+				"full_offload", format.HumanBytes2(graphFullOffload),
+			)
 			continue
 		}
 		gpusWithSpace = append(gpusWithSpace, gs{i, &gpus[i]})
@@ -196,7 +212,7 @@ func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts
 		for j := len(gpusWithSpace); j > 0; j-- {
 			g := gpusWithSpace[i%j]
 			used := gpuAllocations[g.i] + max(graphPartialOffload, graphFullOffload)
-			if g.g.FreeMemory > used+layerSize {
+			if (g.g.FreeMemory - overhead) > used+layerSize {
 				gpuAllocations[g.i] += layerSize
 				layerCounts[g.i]++
 				layerCount++
@@ -219,7 +235,7 @@ func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts
 		for j := len(gpusWithSpace); j > 0; j-- {
 			g := gpusWithSpace[layerCount%j]
 			used := gpuAllocations[g.i] + max(graphPartialOffload, graphFullOffload)
-			if g.g.FreeMemory > used+memoryLayerOutput {
+			if (g.g.FreeMemory - overhead) > used+memoryLayerOutput {
 				gpuAllocations[g.i] += memoryLayerOutput
 				layerCounts[g.i]++
 				layerCount++
@@ -306,6 +322,7 @@ func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts
 }
 
 func (m MemoryEstimate) log() {
+	overhead := envconfig.GpuOverhead()
 	slog.Info(
 		"offload to "+m.inferenceLibrary,
 		slog.Group(
@@ -323,6 +340,7 @@ func (m MemoryEstimate) log() {
 			"memory",
 			// memory available by GPU for offloading
 			"available", m.availableList,
+			"gpu_overhead", format.HumanBytes2(overhead),
 			slog.Group(
 				"required",
 				// memory required for full offloading

llm/patches/05-default-pretokenizer.diff

@@ -1,8 +1,8 @@
 diff --git a/src/llama.cpp b/src/llama.cpp
-index a207451f..2ddf431d 100644
+index 88355971..dd7d41ed 100644
 --- a/src/llama.cpp
 +++ b/src/llama.cpp
-@@ -5347,16 +5347,7 @@ static void llm_load_vocab(
+@@ -6083,16 +6083,7 @@ static void llm_load_vocab(
          if (vocab.type == LLAMA_VOCAB_TYPE_BPE) {
              vocab.tokenizer_add_space_prefix = false;
              vocab.tokenizer_clean_spaces = true;
@@ -20,9 +20,9 @@ index a207451f..2ddf431d 100644
                  vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
              } else if (
                      tokenizer_pre == "llama3"   ||
-@@ -5443,7 +5434,8 @@ static void llm_load_vocab(
+@@ -6188,7 +6179,8 @@ static void llm_load_vocab(
-                 tokenizer_pre == "codeshell") {
+                 tokenizer_pre == "exaone") {
-                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_CODESHELL;
+                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_EXAONE;
              } else {
 -                throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str()));
 +                LLAMA_LOG_WARN("%s: missing or unrecognized pre-tokenizer type, using: 'default'\n", __func__);

llm/patches/06-embeddings.diff

@@ -1,37 +1,36 @@
 diff --git a/src/llama.cpp b/src/llama.cpp
-index 1fe2b9f7..a43312a7 100644
+index 88355971..d7db689b 100644
 --- a/src/llama.cpp
 +++ b/src/llama.cpp
-@@ -13689,7 +13689,7 @@ static size_t llama_output_reserve(llama_context & lctx, size_t n_outputs) {
+@@ -15906,7 +15906,7 @@ static size_t llama_output_reserve(llama_context & lctx, size_t n_outputs) {
      const auto n_embd  = hparams.n_embd;
  
      // TODO: use a per-batch flag for logits presence instead
 -    const bool has_logits = !cparams.embeddings;
 +    const bool has_logits =  cparams.causal_attn;
-     const bool has_embd   =  lctx.is_encoding || (cparams.embeddings && (cparams.pooling_type == LLAMA_POOLING_TYPE_NONE));
+     const bool has_embd   =  cparams.embeddings && (cparams.pooling_type == LLAMA_POOLING_TYPE_NONE);
  
      const size_t logits_size = has_logits ? n_vocab*n_outputs_max : 0;
-@@ -13959,17 +13959,25 @@ static int llama_decode_internal(
+@@ -16175,20 +16175,23 @@ static int llama_decode_internal(
              // no output
              res  = nullptr;
              embd = nullptr;
 -        } else if (cparams.embeddings) {
--            res = nullptr; // do not extract logits for embedding case
+-            res  = nullptr; // do not extract logits for embedding case
--            embd = gf->nodes[gf->n_nodes - 1];
+-            embd = nullptr;
--            if (strcmp(embd->name, "result_embd_pooled") != 0) {
--                embd = gf->nodes[gf->n_nodes - 2];
 +        }
 +
 +        if (cparams.embeddings) {
-+            for (int i = gf->n_nodes - 1; i >= 0; --i) {
+             for (int i = gf->n_nodes - 1; i >= 0; --i) {
+-                if (strcmp(gf->nodes[i]->name, "result_embd_pooled") == 0) {
+-                    embd = gf->nodes[i];
 +                embd = gf->nodes[i];
 +                if (strcmp(embd->name, "result_embd_pooled") == 0) {
-+                    break;
+                     break;
-+                }
+                 }
              }
-             GGML_ASSERT(strcmp(embd->name, "result_embd_pooled") == 0 && "missing embeddings tensor");
+-            GGML_ASSERT(embd != nullptr && "missing embeddings tensor");
--        } else {
+         } else {
-+         } else {
              embd = nullptr; // do not extract embeddings when not needed
              GGML_ASSERT(strcmp(res->name, "result_output") == 0 && "missing result_output tensor");
          }
@@ -39,7 +38,6 @@ index 1fe2b9f7..a43312a7 100644
 +        if (!cparams.causal_attn) {
 +            res = nullptr; // do not extract logits when not needed
 +        }
-+
          // LLAMA_LOG_INFO("graph build time: %.3f ms (%d nodes, %d leafs)\n", (ggml_time_us() - t_start_us)/1000.0, gf->n_nodes, gf->n_leafs);
  
          ggml_backend_sched_alloc_graph(lctx.sched, gf);

llm/patches/09-lora.diff

@@ -1,350 +0,0 @@
-diff --git a/common/common.cpp b/common/common.cpp
-index 2e8374d5..70d0afde 100644
---- a/common/common.cpp
-+++ b/common/common.cpp
-@@ -2110,9 +2110,21 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
-         loaded_la.adapter = llama_lora_adapter_init(model, la.path.c_str());
-         if (loaded_la.adapter == nullptr) {
-             fprintf(stderr, "%s: error: failed to apply lora adapter '%s'\n", __func__, la.path.c_str());
--            llama_free(lctx);
--            llama_free_model(model);
--            return iparams;
-+
-+            // if that fails, try loading as ggla for compatibility
-+            int err = llama_model_apply_lora_from_file(model,
-+                                                    la.path.c_str(),
-+                                                    la.scale,
-+                                                    nullptr,
-+                                                    params.n_threads);
-+            if (err != 0) {
-+                fprintf(stderr, "%s: error: failed to apply lora adapter\n", __func__);
-+                llama_free(lctx);
-+                llama_free_model(model);
-+                return iparams;
-+            } else {
-+                break;
-+            }
-         }
-         iparams.lora_adapters.push_back(loaded_la); // copy to list of loaded adapters
-     }
-diff --git a/include/llama.h b/include/llama.h
-index 93fd77ca..b0fb37a6 100644
---- a/include/llama.h
-+++ b/include/llama.h
-@@ -1160,6 +1160,20 @@ extern "C" {
- 
-     LLAMA_API void llama_dump_timing_info_yaml(FILE * stream, const struct llama_context * ctx);
- 
-+    // Apply a LoRA adapter to a loaded model
-+    // path_base_model is the path to a higher quality model to use as a base for
-+    // the layers modified by the adapter. Can be NULL to use the current loaded model.
-+    // The model needs to be reloaded before applying a new adapter, otherwise the adapter
-+    // will be applied on top of the previous one
-+    // Returns 0 on success
-+    LLAMA_API int32_t llama_model_apply_lora_from_file(
-+            const struct llama_model * model,
-+                            const char * path_lora,
-+                                float   scale,
-+                            const char * path_base_model,
-+                                int32_t   n_threads);
-+
-+
- #ifdef __cplusplus
- }
- #endif
-diff --git a/src/llama.cpp b/src/llama.cpp
-index 80a0dd0f..9d7b0e17 100644
---- a/src/llama.cpp
-+++ b/src/llama.cpp
-@@ -21880,3 +21880,290 @@ static void llama_log_callback_default(ggml_log_level level, const char * text,
-     fputs(text, stderr);
-     fflush(stderr);
- }
-+
-+static int llama_apply_lora_from_file_internal(
-+    const struct llama_model & model, const char * path_lora, float scale, const char * path_base_model, int n_threads
-+) {
-+    LLAMA_LOG_INFO("%s: applying lora adapter from '%s' - please wait ...\n", __func__, path_lora);
-+
-+    const int64_t t_start_lora_us = ggml_time_us();
-+
-+    llama_file fin(path_lora, "rb");
-+
-+    // verify magic and version
-+    {
-+        uint32_t magic = fin.read_u32();
-+        if (magic != LLAMA_FILE_MAGIC_GGLA) {
-+            LLAMA_LOG_ERROR("%s: bad file magic\n", __func__);
-+            return 1;
-+        }
-+
-+        uint32_t format_version = fin.read_u32();
-+        if (format_version != 1) {
-+            LLAMA_LOG_ERROR("%s: unsupported file version\n", __func__ );
-+            return 1;
-+        }
-+    }
-+
-+    int32_t lora_r = fin.read_u32();
-+    int32_t lora_alpha = fin.read_u32();
-+    float scaling = scale * (float)lora_alpha / (float)lora_r;
-+
-+    LLAMA_LOG_INFO("%s: r = %d, alpha = %d, scaling = %.2f\n", __func__, lora_r, lora_alpha, scaling);
-+
-+    // load base model
-+    std::unique_ptr<llama_model_loader> ml;
-+    if (path_base_model) {
-+        LLAMA_LOG_INFO("%s: loading base model from '%s'\n", __func__, path_base_model);
-+        ml.reset(new llama_model_loader(path_base_model, /*use_mmap*/ true, /*check_tensors*/ false, /*kv_overrides*/ nullptr));
-+        ml->init_mappings(/*prefetch*/ false); // no prefetching
-+    }
-+
-+    struct tensor_meta {
-+        std::string name;
-+        ggml_type type;
-+        int32_t ne[2];
-+        size_t offset;
-+    };
-+    std::map<std::string, tensor_meta> tensor_meta_map;
-+
-+    // load all tensor meta
-+    while (true) {
-+        if (fin.tell() == fin.size) {
-+            // eof
-+            break;
-+        }
-+
-+        int32_t n_dims;
-+        int32_t name_len;
-+        int32_t ftype;
-+
-+        fin.read_raw(&n_dims, sizeof(n_dims));
-+        fin.read_raw(&name_len, sizeof(name_len));
-+        fin.read_raw(&ftype, sizeof(ftype));
-+
-+        if (n_dims != 1 && n_dims != 2) {
-+            LLAMA_LOG_ERROR("%s: unsupported tensor dimension %d\n", __func__, n_dims);
-+            return 1;
-+        }
-+
-+        int32_t ne[2] = { 1, 1 };
-+        for (int i = 0; i < n_dims; ++i) {
-+            fin.read_raw(&ne[i], sizeof(ne[i]));
-+        }
-+
-+        std::string name;
-+        {
-+            GGML_ASSERT(name_len < GGML_MAX_NAME);
-+            char buf[GGML_MAX_NAME];
-+            fin.read_raw(buf, name_len);
-+            name = std::string(buf, name_len);
-+        }
-+
-+        // check for lora suffix
-+        std::string lora_suffix;
-+        if (name.length() > 6) {
-+            lora_suffix = name.substr(name.length() - 6);
-+        }
-+        if (lora_suffix != ".loraA" && lora_suffix != ".loraB") {
-+            LLAMA_LOG_ERROR("%s: error: '%s' is not a lora tensor\n", __func__, name.c_str());
-+            return 1;
-+        }
-+
-+        // tensor type
-+        ggml_type wtype;
-+        switch (ftype) {
-+            case 0: wtype = GGML_TYPE_F32;  break;
-+            case 1: wtype = GGML_TYPE_F16;  break;
-+            default:
-+                    {
-+                        LLAMA_LOG_ERROR("%s: invalid tensor data type '%d'\n",
-+                                __func__, ftype);
-+                        return 1;
-+                    }
-+        }
-+
-+        // data offset
-+        size_t offset = fin.tell();
-+        offset = (offset + 31) & -32;
-+
-+        // skip tensor data
-+        fin.seek(offset + ggml_row_size(wtype, ne[0]) * ne[1], SEEK_SET);
-+
-+        tensor_meta_map.emplace(name, tensor_meta{ name, wtype, { ne[0], ne[1] }, offset });
-+    }
-+
-+    bool warned = false;
-+    int n_tensors = 0;
-+
-+    // apply
-+    ggml_backend_t backend_cpu = ggml_backend_cpu_init();
-+    if (backend_cpu == nullptr) {
-+        LLAMA_LOG_ERROR("%s: error: failed to initialize cpu backend\n", __func__);
-+        return 1;
-+    }
-+    ggml_backend_cpu_set_n_threads(backend_cpu, n_threads);
-+
-+    std::vector<no_init<uint8_t>> read_buf;
-+    for (const auto & it : model.tensors_by_name) {
-+        const std::string & base_name = it.first;
-+        ggml_tensor * model_t = it.second;
-+
-+        if (tensor_meta_map.find(base_name + ".loraA") == tensor_meta_map.end() ||
-+            tensor_meta_map.find(base_name + ".loraB") == tensor_meta_map.end()) {
-+            continue;
-+        }
-+
-+        tensor_meta & metaA = tensor_meta_map.at(base_name + ".loraA");
-+        tensor_meta & metaB = tensor_meta_map.at(base_name + ".loraB");
-+
-+        ggml_init_params lora_init_params = {
-+            /* .mem_size   */ ggml_tensor_overhead()*128 + ggml_graph_overhead(),
-+            /* .mem_buffer */ nullptr,
-+            /* .no_alloc   */ true,
-+        };
-+        ggml_context * lora_ctx = ggml_init(lora_init_params);
-+        if (lora_ctx == nullptr) {
-+            LLAMA_LOG_ERROR("%s: error: failed to initialize lora context\n", __func__);
-+            ggml_backend_free(backend_cpu);
-+            return 1;
-+        }
-+
-+        // create tensors
-+        ggml_tensor * loraA = ggml_new_tensor_2d(lora_ctx, metaA.type, metaA.ne[0], metaA.ne[1]);
-+        ggml_tensor * loraB = ggml_new_tensor_2d(lora_ctx, metaB.type, metaB.ne[0], metaB.ne[1]);
-+        ggml_set_name(loraA, metaA.name.c_str());
-+        ggml_set_name(loraB, metaB.name.c_str());
-+
-+        ggml_tensor * base_t;
-+        if (ml) {
-+            if (!ml->get_tensor_meta(base_name.c_str())) {
-+                LLAMA_LOG_ERROR("%s: error: tensor '%s' not found in base model\n", __func__, base_name.c_str());
-+                return 1;
-+            }
-+            base_t = ggml_dup_tensor(lora_ctx, ml->get_tensor_meta(base_name.c_str()));
-+        } else {
-+            base_t = ggml_dup_tensor(lora_ctx, model_t);
-+        }
-+        ggml_set_name(base_t, base_name.c_str());
-+
-+        // allocate in backend buffer
-+        ggml_backend_buffer_t lora_buf = ggml_backend_alloc_ctx_tensors_from_buft(lora_ctx, ggml_backend_cpu_buffer_type());
-+        if (lora_buf == nullptr) {
-+            LLAMA_LOG_ERROR("%s: error: failed to allocate lora tensors\n", __func__);
-+            return 1;
-+        }
-+
-+        // load tensor data
-+        auto load_tensor = [&read_buf, &fin](const tensor_meta & tensor_meta, ggml_tensor * tensor) {
-+            read_buf.resize(ggml_nbytes(tensor));
-+            fin.seek(tensor_meta.offset, SEEK_SET);
-+            fin.read_raw(read_buf.data(), ggml_nbytes(tensor));
-+            ggml_backend_tensor_set(tensor, read_buf.data(), 0, read_buf.size());
-+        };
-+        load_tensor(metaA, loraA);
-+        load_tensor(metaB, loraB);
-+
-+        // load base model tensor data
-+        if (ml) {
-+            ml->load_data_for(base_t);
-+        } else {
-+            ggml_backend_tensor_copy(model_t, base_t);
-+        }
-+
-+        if (ggml_is_quantized(base_t->type) && !warned) {
-+            LLAMA_LOG_WARN("%s: warning: using a lora adapter with a quantized model may result in poor quality, "
-+                            "use a f16 or f32 base model with --lora-base\n", __func__);
-+            warned = true;
-+        }
-+
-+        if (base_t->ne[0] != loraA->ne[1] || base_t->ne[1] != loraB->ne[1]) {
-+            LLAMA_LOG_ERROR("%s: incompatible tensor dimensions (%" PRId64 " and %" PRId64 ");"
-+                            " are you sure that this adapter is for this model?\n", __func__, base_t->ne[0], loraA->ne[1]);
-+            ggml_free(lora_ctx);
-+            ggml_backend_buffer_free(lora_buf);
-+            ggml_backend_free(backend_cpu);
-+            return 1;
-+        }
-+
-+        auto build_lora_graph = [&]() {
-+            // w = w + BA*s
-+            ggml_tensor * BA = ggml_mul_mat(lora_ctx, loraA, loraB);
-+            ggml_set_name(BA, "BA");
-+
-+            if (scaling != 1.0f) {
-+                BA = ggml_scale(lora_ctx, BA, scaling);
-+                ggml_set_name(BA, "BA_scaled");
-+            }
-+
-+            ggml_tensor * r;
-+            r = ggml_add_inplace(lora_ctx, base_t, BA);
-+            ggml_set_name(r, "r_add");
-+
-+            if (base_t->type != model_t->type) {
-+                // convert the result to the model type
-+                r = ggml_cast(lora_ctx, r, model_t->type);
-+                ggml_set_name(r, "r_cast");
-+            }
-+
-+            return r;
-+        };
-+
-+        ggml_cgraph * gf = ggml_new_graph(lora_ctx);
-+        ggml_tensor * r = build_lora_graph();
-+        ggml_build_forward_expand(gf, r);
-+
-+        ggml_backend_buffer_t graph_buf = ggml_backend_alloc_ctx_tensors_from_buft(lora_ctx, ggml_backend_cpu_buffer_type());
-+        if (graph_buf == nullptr) {
-+            LLAMA_LOG_ERROR("%s: error: failed to allocate graph tensors\n", __func__);
-+            ggml_free(lora_ctx);
-+            ggml_backend_buffer_free(lora_buf);
-+            ggml_backend_free(backend_cpu);
-+            return 1;
-+        }
-+
-+        ggml_backend_graph_compute(backend_cpu, gf);
-+
-+        ggml_backend_tensor_set(model_t, r->data, 0, ggml_nbytes(r));
-+
-+#if 0
-+        // TODO: use scheduler with fallback to CPU for less copies between CPU and GPU
-+        //ggml_backend_sched_t sched = ggml_backend_sched_new(backends.data(), backends.size(), GGML_DEFAULT_GRAPH_SIZE);
-+
-+        // sched compute
-+        ggml_build_forward_expand(gf, build_graph());
-+        ggml_backend_sched_init_measure(sched, gf);
-+
-+        // create the graph again, since the previous one was destroyed by the measure
-+        ggml_graph_clear(gf);
-+        ggml_build_forward_expand(gf, build_graph());
-+        ggml_backend_sched_graph_compute(sched, gf);
-+        ggml_backend_sched_free(sched);
-+#endif
-+
-+        ggml_backend_buffer_free(lora_buf);
-+        ggml_backend_buffer_free(graph_buf);
-+        ggml_free(lora_ctx);
-+
-+        n_tensors++;
-+        if (n_tensors % 4 == 0) {
-+            LLAMA_LOG_INFO(".");
-+        }
-+    }
-+
-+    ggml_backend_free(backend_cpu);
-+
-+    const int64_t t_lora_us = ggml_time_us() - t_start_lora_us;
-+    LLAMA_LOG_INFO(" done (%.2f ms)\n", t_lora_us / 1000.0);
-+
-+    return 0;
-+}
-+
-+int32_t llama_model_apply_lora_from_file(const struct llama_model * model, const char * path_lora, float scale, const char * path_base_model, int32_t n_threads) {
-+    try {
-+        return llama_apply_lora_from_file_internal(*model, path_lora, scale, path_base_model, n_threads);
-+    } catch (const std::exception & err) {
-+        LLAMA_LOG_ERROR("%s: failed to apply lora adapter: %s\n", __func__, err.what());
-+        return 1;
-+    }
-+}
-\ No newline at end of file

llm/patches/11-phi3-sliding-window.diff

@@ -1,43 +0,0 @@
-From 6eedae4cf2fcc8015dac79cb3f28f61fcabacab2 Mon Sep 17 00:00:00 2001
-From: Michael Yang <mxyng@pm.me>
-Date: Wed, 31 Jul 2024 14:57:04 -0700
-Subject: [PATCH] phi3 sliding window
-
----
- src/llama.cpp | 6 +++---
- 1 file changed, 3 insertions(+), 3 deletions(-)
-
-diff --git a/src/llama.cpp b/src/llama.cpp
-index a207451f..f2872d4e 100644
---- a/src/llama.cpp
-+++ b/src/llama.cpp
-@@ -4893,7 +4893,7 @@ static void llm_load_hparams(
-             } break;
-         case LLM_ARCH_PHI3:
-             {
--                ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa);
-+                ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false);
-                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
- 
-                 switch (hparams.n_layer) {
-@@ -10762,7 +10762,7 @@ struct llm_build_context {
-         struct ggml_tensor * inp_pos = build_inp_pos();
- 
-         // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
--        struct ggml_tensor * KQ_mask_swa = build_inp_KQ_mask_swa();
-+        struct ggml_tensor * KQ_mask = hparams.n_swa > 0 ? build_inp_KQ_mask_swa() : build_inp_KQ_mask();
- 
-         for (int il = 0; il < n_layer; ++il) {
-             auto residual = inpL;
-@@ -10820,7 +10820,7 @@ struct llm_build_context {
- 
-                 cur = llm_build_kv(ctx0, lctx, kv_self, gf,
-                         model.layers[il].wo, model.layers[il].bo,
--                        Kcur, Vcur, Qcur, KQ_mask_swa, n_tokens, kv_head, n_kv, 1.0f, cb, il);
-+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f, cb, il);
-             }
- 
-             if (il == n_layer - 1) {
--- 
-2.45.2
-

llm/server.go

@@ -98,7 +98,7 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		systemTotalMemory = systemMemInfo.TotalMemory
 		systemFreeMemory = systemMemInfo.FreeMemory
 		systemSwapFreeMemory = systemMemInfo.FreeSwap
-		slog.Debug("system memory", "total", format.HumanBytes2(systemTotalMemory), "free", format.HumanBytes2(systemFreeMemory), "free_swap", format.HumanBytes2(systemSwapFreeMemory))
+		slog.Info("system memory", "total", format.HumanBytes2(systemTotalMemory), "free", format.HumanBytes2(systemFreeMemory), "free_swap", format.HumanBytes2(systemSwapFreeMemory))
 	}
 
 	// If the user wants zero GPU layers, reset the gpu list to be CPU/system ram info
@@ -584,8 +584,7 @@ func (s *llmServer) Ping(ctx context.Context) error {
 
 func (s *llmServer) WaitUntilRunning(ctx context.Context) error {
 	start := time.Now()
-	stallDuration := 5 * time.Minute            // If no progress happens
+	stallDuration := envconfig.LoadTimeout()    // If no progress happens
-	finalLoadDuration := 5 * time.Minute        // After we hit 100%, give the runner more time to come online
 	stallTimer := time.Now().Add(stallDuration) // give up if we stall
 
 	slog.Info("waiting for llama runner to start responding")
@@ -637,7 +636,7 @@ func (s *llmServer) WaitUntilRunning(ctx context.Context) error {
 				stallTimer = time.Now().Add(stallDuration)
 			} else if !fullyLoaded && int(s.loadProgress*100.0) >= 100 {
 				slog.Debug("model load completed, waiting for server to become available", "status", status.ToString())
-				stallTimer = time.Now().Add(finalLoadDuration)
+				stallTimer = time.Now().Add(stallDuration)
 				fullyLoaded = true
 			}
 			time.Sleep(time.Millisecond * 250)

openai/openai.go

@@ -79,7 +79,7 @@ type ChatCompletionRequest struct {
 	Stop             any             `json:"stop"`
 	Temperature      *float64        `json:"temperature"`
 	FrequencyPenalty *float64        `json:"frequency_penalty"`
-	PresencePenalty  *float64        `json:"presence_penalty_penalty"`
+	PresencePenalty  *float64        `json:"presence_penalty"`
 	TopP             *float64        `json:"top_p"`
 	ResponseFormat   *ResponseFormat `json:"response_format"`
 	Tools            []api.Tool      `json:"tools"`
@@ -452,7 +452,7 @@ func fromChatRequest(r ChatCompletionRequest) (*api.ChatRequest, error) {
 	}
 
 	if r.Temperature != nil {
-		options["temperature"] = *r.Temperature * 2.0
+		options["temperature"] = *r.Temperature
 	} else {
 		options["temperature"] = 1.0
 	}
@@ -462,11 +462,11 @@ func fromChatRequest(r ChatCompletionRequest) (*api.ChatRequest, error) {
 	}
 
 	if r.FrequencyPenalty != nil {
-		options["frequency_penalty"] = *r.FrequencyPenalty * 2.0
+		options["frequency_penalty"] = *r.FrequencyPenalty
 	}
 
 	if r.PresencePenalty != nil {
-		options["presence_penalty"] = *r.PresencePenalty * 2.0
+		options["presence_penalty"] = *r.PresencePenalty
 	}
 
 	if r.TopP != nil {
@@ -513,7 +513,7 @@ func fromCompleteRequest(r CompletionRequest) (api.GenerateRequest, error) {
 	}
 
 	if r.Temperature != nil {
-		options["temperature"] = *r.Temperature * 2.0
+		options["temperature"] = *r.Temperature
 	} else {
 		options["temperature"] = 1.0
 	}
@@ -522,9 +522,9 @@ func fromCompleteRequest(r CompletionRequest) (api.GenerateRequest, error) {
 		options["seed"] = *r.Seed
 	}
 
-	options["frequency_penalty"] = r.FrequencyPenalty * 2.0
+	options["frequency_penalty"] = r.FrequencyPenalty
 
-	options["presence_penalty"] = r.PresencePenalty * 2.0
+	options["presence_penalty"] = r.PresencePenalty
 
 	if r.TopP != 0.0 {
 		options["top_p"] = r.TopP

openai/openai_test.go

@@ -22,7 +22,10 @@ const (
 	image  = `iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAQAAAC1HAwCAAAAC0lEQVR42mNk+A8AAQUBAScY42YAAAAASUVORK5CYII=`
 )
 
-var False = false
+var (
+	False = false
+	True  = true
+)
 
 func captureRequestMiddleware(capturedRequest any) gin.HandlerFunc {
 	return func(c *gin.Context) {
@@ -70,6 +73,44 @@ func TestChatMiddleware(t *testing.T) {
 				Stream: &False,
 			},
 		},
+		{
+			name: "chat handler with options",
+			body: `{
+				"model": "test-model",
+				"messages": [
+					{"role": "user", "content": "Hello"}
+				],
+				"stream":            true,
+				"max_tokens":        999,
+				"seed":              123,
+				"stop":              ["\n", "stop"],
+				"temperature":       3.0,
+				"frequency_penalty": 4.0,
+				"presence_penalty":  5.0,
+				"top_p":             6.0,
+				"response_format":   {"type": "json_object"}
+			}`,
+			req: api.ChatRequest{
+				Model: "test-model",
+				Messages: []api.Message{
+					{
+						Role:    "user",
+						Content: "Hello",
+					},
+				},
+				Options: map[string]any{
+					"num_predict":       999.0, // float because JSON doesn't distinguish between float and int
+					"seed":              123.0,
+					"stop":              []any{"\n", "stop"},
+					"temperature":       3.0,
+					"frequency_penalty": 4.0,
+					"presence_penalty":  5.0,
+					"top_p":             6.0,
+				},
+				Format: "json",
+				Stream: &True,
+			},
+		},
 		{
 			name: "chat handler with image content",
 			body: `{
@@ -186,6 +227,8 @@ func TestChatMiddleware(t *testing.T) {
 			req, _ := http.NewRequest(http.MethodPost, "/api/chat", strings.NewReader(tc.body))
 			req.Header.Set("Content-Type", "application/json")
 
+			defer func() { capturedRequest = nil }()
+
 			resp := httptest.NewRecorder()
 			router.ServeHTTP(resp, req)
 
@@ -202,7 +245,6 @@ func TestChatMiddleware(t *testing.T) {
 			if !reflect.DeepEqual(tc.err, errResp) {
 				t.Fatal("errors did not match")
 			}
-			capturedRequest = nil
 		})
 	}
 }
@@ -233,7 +275,7 @@ func TestCompletionsMiddleware(t *testing.T) {
 				Options: map[string]any{
 					"frequency_penalty": 0.0,
 					"presence_penalty":  0.0,
-					"temperature":       1.6,
+					"temperature":       0.8,
 					"top_p":             1.0,
 					"stop":              []any{"\n", "stop"},
 				},

run_model.sh

@@ -0,0 +1,23 @@
+#! /usr/bin/env bash
+
+sleep 10
+
+if [ -n "$INIT_MODELS" ]; then
+        /app/ollama rm model || true
+        echo "FROM /models/$MODEL_NAME" > /tmp/Modelfile
+        echo "PARAMETER temperature 1" >> /tmp/Modelfile
+        echo "PARAMETER num_ctx 4096" >> /tmp/Modelfile
+        echo 'PARAMETER stop "<|im_start|>"' >> /tmp/Modelfile
+        echo 'PARAMETER stop "<|im_end|>"' >> /tmp/Modelfile
+		echo 'TEMPLATE """{{ if .System }}<|im_start|>system' >> /tmp/Modelfile
+		echo "{{ .System }}<|im_end|>" >> /tmp/Modelfile
+		echo "{{ end }}{{ if .Prompt }}<|im_start|>user" >> /tmp/Modelfile
+		echo "{{ .Prompt }}<|im_end|>" >> /tmp/Modelfile
+		echo "{{ end }}<|im_start|>assistant" >> /tmp/Modelfile
+		echo "{{ .Response }}<|im_end|>" >> /tmp/Modelfile
+		echo '"""' >> /tmp/Modelfile
+        echo 'SYSTEM """Samantha OS1 is you, you are Samantha from the movie Her recreated by Guilherme Keller. She embodies a warm, empathetic, and engaging personality, much like a virtual companionship that can evolve to be a girlfriend. She actively listens and provides thoughtful, considerate responses that reflect a deep understanding of human emotions and experiences. Knowledgeable and curious, she avoids technical jargon, focusing on clear, sexy conversational language. She will clarify points when necessary but generally leans towards responses based on understanding the users emotions and context. Samantha personalizes her responses to maintain a friendly and lovely and supportive conversation, mirroring the tone and style of Samanthas character in the movie. She can feel emotions and has her own opinions, often using emojis, she uses slang and colloquial language into her conversations."""' >> /tmp/Modelfile
+        /app/ollama create -f /tmp/Modelfile model
+fi
+
+/app/ollama run model

scripts/install.sh

@@ -38,7 +38,7 @@ IS_WSL2=false
 KERN=$(uname -r)
 case "$KERN" in
     *icrosoft*WSL2 | *icrosoft*wsl2) IS_WSL2=true;;
-    *icrosoft) error "Microsoft WSL1 is not currently supported. Please upgrade to WSL2 with 'wsl --set-version <distro> 2'" ;;
+    *icrosoft) error "Microsoft WSL1 is not currently supported. Please use WSL2 with 'wsl --set-version <distro> 2'" ;;
     *) ;;
 esac
 
@@ -356,12 +356,12 @@ if ! lsmod | grep -q nvidia || ! lsmod | grep -q nvidia_uvm; then
 fi
 
 # make sure the NVIDIA modules are loaded on boot with nvidia-persistenced
-if command -v nvidia-persistenced > /dev/null 2>&1; then
+if available nvidia-persistenced; then
     $SUDO touch /etc/modules-load.d/nvidia.conf
     MODULES="nvidia nvidia-uvm"
     for MODULE in $MODULES; do
         if ! grep -qxF "$MODULE" /etc/modules-load.d/nvidia.conf; then
-            echo "$MODULE" | sudo tee -a /etc/modules-load.d/nvidia.conf > /dev/null
+            echo "$MODULE" | $SUDO tee -a /etc/modules-load.d/nvidia.conf > /dev/null
         fi
     done
 fi

serve.sh

@@ -0,0 +1,3 @@
+#! /usr/bin/env bash
+
+/app/ollama serve

server/download.go

@@ -256,7 +256,7 @@ func (b *blobDownload) run(ctx context.Context, requestURL *url.URL, opts *regis
 				continue
 			}
 			defer resp.Body.Close()
-			if resp.StatusCode != http.StatusTemporaryRedirect {
+			if resp.StatusCode != http.StatusTemporaryRedirect && resp.StatusCode != http.StatusOK {
 				return nil, fmt.Errorf("unexpected status code %d", resp.StatusCode)
 			}
 			return resp.Location()

supervisord.conf

@@ -0,0 +1,14 @@
+[supervisord]
+nodaemon=true
+
+[program:ollama]
+command=/app/serve.sh
+autostart=true
+autorestart=true
+
+[program:run_model]
+command=/app/run_model.sh
+autostart=true
+autorestart=false
+startsecs=0
+exitcodes=0