hexagon: apply repl optimization in flash attn softmax as #22993 (#23455)

* requirements: relax torch~=2.6.0 to torch>=2.6.0 for convert_hf_to_gguf

The ~=2.6.0 operator resolves to >=2.6.0, <2.7.0, which fails on
PyPI for platform/CPython combinations where 2.6.x is not present.
The accompanying comment already says 'PyTorch 2.6.0 or later', so
the looser >=2.6.0 matches the documented intent and unblocks
pip install -r requirements/requirements-convert_hf_to_gguf.txt.

Fixes #23408

* requirements: bump torch floor to 2.11.0 per maintainer

* requirements: pin torch to ==2.11.0 per project policy

* requirements: pin mtmd torch and torchvision to 2.11.0/0.26.0 per project policy

* requirements: suppress check_requirements pin warning on mtmd

The check_requirements script flags '==' on lines in files matched by
*/**/requirements*.txt. Append the documented suppression comment to the
pinned torch and torchvision lines (and to the s390x platform marker lines)
so the check passes while keeping the pins required by project policy.

* ty: silence Tensor/Module union check on model[0].auto_model

With torch 2.11.0 stubs, nn.Sequential.__getitem__ now returns
Tensor | Module rather than Module, so model[0].auto_model fails ty
on the SentenceTransformer code path. The runtime behavior is
unchanged because SentenceTransformer always wraps a Module at
index 0. Adding a targeted unresolved-attribute ignore keeps the
type-check green without altering behavior. A follow-up issue
tracks typing the variable explicitly.

.devops/cann.Dockerfile

@@ -5,6 +5,9 @@
 # Define the CANN base image for easier version updates later
 ARG CHIP_TYPE=910b
 ARG CANN_BASE_IMAGE=quay.io/ascend/cann:8.5.0-${CHIP_TYPE}-openeuler24.03-py3.11
+ARG BUILD_DATE=N/A
+ARG APP_VERSION=N/A
+ARG APP_REVISION=N/A
 
 # ==============================================================================
 # BUILD STAGE
@@ -55,6 +58,7 @@ RUN mkdir -p /app/lib && \
 RUN mkdir -p /app/full && \
     cp build/bin/* /app/full/ && \
     cp *.py /app/full/ && \
+    cp -r conversion /app/full/ && \
     cp -r gguf-py /app/full/ && \
     cp -r requirements /app/full/ && \
     cp requirements.txt /app/full/
@@ -67,6 +71,19 @@ RUN mkdir -p /app/full && \
 # ==============================================================================
 FROM ${CANN_BASE_IMAGE} AS base
 
+ARG BUILD_DATE=N/A
+ARG APP_VERSION=N/A
+ARG APP_REVISION=N/A
+ARG IMAGE_URL=https://github.com/ggml-org/llama.cpp
+ARG IMAGE_SOURCE=https://github.com/ggml-org/llama.cpp
+LABEL org.opencontainers.image.created=$BUILD_DATE \
+      org.opencontainers.image.version=$APP_VERSION \
+      org.opencontainers.image.revision=$APP_REVISION \
+      org.opencontainers.image.title="llama.cpp" \
+      org.opencontainers.image.description="LLM inference in C/C++" \
+      org.opencontainers.image.url=$IMAGE_URL \
+      org.opencontainers.image.source=$IMAGE_SOURCE
+
 # -- Install runtime dependencies --
 RUN yum install -y libgomp curl && \
     yum clean all && \

.devops/cpu.Dockerfile

@@ -1,4 +1,7 @@
 ARG UBUNTU_VERSION=24.04
+ARG BUILD_DATE=N/A
+ARG APP_VERSION=N/A
+ARG APP_REVISION=N/A
 
 FROM ubuntu:$UBUNTU_VERSION AS build
 
@@ -27,6 +30,7 @@ RUN mkdir -p /app/lib && \
 RUN mkdir -p /app/full \
     && cp build/bin/* /app/full \
     && cp *.py /app/full \
+    && cp -r conversion /app/full \
     && cp -r gguf-py /app/full \
     && cp -r requirements /app/full \
     && cp requirements.txt /app/full \
@@ -35,6 +39,19 @@ RUN mkdir -p /app/full \
 ## Base image
 FROM ubuntu:$UBUNTU_VERSION AS base
 
+ARG BUILD_DATE=N/A
+ARG APP_VERSION=N/A
+ARG APP_REVISION=N/A
+ARG IMAGE_URL=https://github.com/ggml-org/llama.cpp
+ARG IMAGE_SOURCE=https://github.com/ggml-org/llama.cpp
+LABEL org.opencontainers.image.created=$BUILD_DATE \
+      org.opencontainers.image.version=$APP_VERSION \
+      org.opencontainers.image.revision=$APP_REVISION \
+      org.opencontainers.image.title="llama.cpp" \
+      org.opencontainers.image.description="LLM inference in C/C++" \
+      org.opencontainers.image.url=$IMAGE_URL \
+      org.opencontainers.image.source=$IMAGE_SOURCE
+
 RUN apt-get update \
     && apt-get install -y libgomp1 curl \
     && apt autoremove -y \

.devops/cuda.Dockerfile

@@ -6,6 +6,10 @@ ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VER
 
 ARG BASE_CUDA_RUN_CONTAINER=nvidia/cuda:${CUDA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}
 
+ARG BUILD_DATE=N/A
+ARG APP_VERSION=N/A
+ARG APP_REVISION=N/A
+
 FROM ${BASE_CUDA_DEV_CONTAINER} AS build
 
 # CUDA architecture to build for (defaults to all supported archs)
@@ -32,6 +36,7 @@ RUN mkdir -p /app/lib && \
 RUN mkdir -p /app/full \
     && cp build/bin/* /app/full \
     && cp *.py /app/full \
+    && cp -r conversion /app/full \
     && cp -r gguf-py /app/full \
     && cp -r requirements /app/full \
     && cp requirements.txt /app/full \
@@ -40,6 +45,19 @@ RUN mkdir -p /app/full \
 ## Base image
 FROM ${BASE_CUDA_RUN_CONTAINER} AS base
 
+ARG BUILD_DATE=N/A
+ARG APP_VERSION=N/A
+ARG APP_REVISION=N/A
+ARG IMAGE_URL=https://github.com/ggml-org/llama.cpp
+ARG IMAGE_SOURCE=https://github.com/ggml-org/llama.cpp
+LABEL org.opencontainers.image.created=$BUILD_DATE \
+      org.opencontainers.image.version=$APP_VERSION \
+      org.opencontainers.image.revision=$APP_REVISION \
+      org.opencontainers.image.title="llama.cpp" \
+      org.opencontainers.image.description="LLM inference in C/C++" \
+      org.opencontainers.image.url=$IMAGE_URL \
+      org.opencontainers.image.source=$IMAGE_SOURCE
+
 RUN apt-get update \
     && apt-get install -y libgomp1 curl \
     && apt autoremove -y \

.devops/intel.Dockerfile

@@ -1,4 +1,7 @@
 ARG ONEAPI_VERSION=2025.3.3-0-devel-ubuntu24.04
+ARG BUILD_DATE=N/A
+ARG APP_VERSION=N/A
+ARG APP_REVISION=N/A
 
 ## Build Image
 
@@ -33,6 +36,7 @@ RUN mkdir -p /app/lib && \
 RUN mkdir -p /app/full \
     && cp build/bin/* /app/full \
     && cp *.py /app/full \
+    && cp -r conversion /app/full \
     && cp -r gguf-py /app/full \
     && cp -r requirements /app/full \
     && cp requirements.txt /app/full \
@@ -40,6 +44,19 @@ RUN mkdir -p /app/full \
 
 FROM intel/deep-learning-essentials:$ONEAPI_VERSION AS base
 
+ARG BUILD_DATE=N/A
+ARG APP_VERSION=N/A
+ARG APP_REVISION=N/A
+ARG IMAGE_URL=https://github.com/ggml-org/llama.cpp
+ARG IMAGE_SOURCE=https://github.com/ggml-org/llama.cpp
+LABEL org.opencontainers.image.created=$BUILD_DATE \
+      org.opencontainers.image.version=$APP_VERSION \
+      org.opencontainers.image.revision=$APP_REVISION \
+      org.opencontainers.image.title="llama.cpp" \
+      org.opencontainers.image.description="LLM inference in C/C++" \
+      org.opencontainers.image.url=$IMAGE_URL \
+      org.opencontainers.image.source=$IMAGE_SOURCE
+
 ARG IGC_VERSION=v2.20.5
 ARG IGC_VERSION_FULL=2_2.20.5+19972
 ARG COMPUTE_RUNTIME_VERSION=25.40.35563.10

.devops/llama-cli-cann.Dockerfile

@@ -1,4 +1,7 @@
 ARG ASCEND_VERSION=8.5.0-910b-openeuler22.03-py3.10
+ARG BUILD_DATE=N/A
+ARG APP_VERSION=N/A
+ARG APP_REVISION=N/A
 
 FROM ascendai/cann:$ASCEND_VERSION AS build
 
@@ -28,6 +31,20 @@ RUN echo "Building with static libs" && \
 
 # TODO: use image with NNRT
 FROM ascendai/cann:$ASCEND_VERSION AS runtime
+
+ARG BUILD_DATE=N/A
+ARG APP_VERSION=N/A
+ARG APP_REVISION=N/A
+ARG IMAGE_URL=https://github.com/ggml-org/llama.cpp
+ARG IMAGE_SOURCE=https://github.com/ggml-org/llama.cpp
+LABEL org.opencontainers.image.created=$BUILD_DATE \
+      org.opencontainers.image.version=$APP_VERSION \
+      org.opencontainers.image.revision=$APP_REVISION \
+      org.opencontainers.image.title="llama.cpp" \
+      org.opencontainers.image.description="LLM inference in C/C++" \
+      org.opencontainers.image.url=$IMAGE_URL \
+      org.opencontainers.image.source=$IMAGE_SOURCE
+
 COPY --from=build /app/build/bin/llama-cli /app/build/bin/llama-completion /
 
 ENV LC_ALL=C.utf8

.devops/musa.Dockerfile

@@ -6,6 +6,10 @@ ARG BASE_MUSA_DEV_CONTAINER=mthreads/musa:${MUSA_VERSION}-devel-ubuntu${UBUNTU_V
 
 ARG BASE_MUSA_RUN_CONTAINER=mthreads/musa:${MUSA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}-amd64
 
+ARG BUILD_DATE=N/A
+ARG APP_VERSION=N/A
+ARG APP_REVISION=N/A
+
 FROM ${BASE_MUSA_DEV_CONTAINER} AS build
 
 # MUSA architecture to build for (defaults to all supported archs)
@@ -37,6 +41,7 @@ RUN mkdir -p /app/lib && \
 RUN mkdir -p /app/full \
     && cp build/bin/* /app/full \
     && cp *.py /app/full \
+    && cp -r conversion /app/full \
     && cp -r gguf-py /app/full \
     && cp -r requirements /app/full \
     && cp requirements.txt /app/full \
@@ -45,6 +50,19 @@ RUN mkdir -p /app/full \
 ## Base image
 FROM ${BASE_MUSA_RUN_CONTAINER} AS base
 
+ARG BUILD_DATE=N/A
+ARG APP_VERSION=N/A
+ARG APP_REVISION=N/A
+ARG IMAGE_URL=https://github.com/ggml-org/llama.cpp
+ARG IMAGE_SOURCE=https://github.com/ggml-org/llama.cpp
+LABEL org.opencontainers.image.created=$BUILD_DATE \
+      org.opencontainers.image.version=$APP_VERSION \
+      org.opencontainers.image.revision=$APP_REVISION \
+      org.opencontainers.image.title="llama.cpp" \
+      org.opencontainers.image.description="LLM inference in C/C++" \
+      org.opencontainers.image.url=$IMAGE_URL \
+      org.opencontainers.image.source=$IMAGE_SOURCE
+
 RUN apt-get update \
     && apt-get install -y libgomp1 curl \
     && apt autoremove -y \

.devops/openvino.Dockerfile

@@ -18,6 +18,10 @@ ARG LIBZE1_VERSION=1.27.0-1~24.04~ppa2
 ARG http_proxy=
 ARG https_proxy=
 
+ARG BUILD_DATE=N/A
+ARG APP_VERSION=N/A
+ARG APP_REVISION=N/A
+
 ## Build Image
 FROM ubuntu:${UBUNTU_VERSION} AS build
 
@@ -77,6 +81,7 @@ RUN mkdir -p /app/lib && \
 RUN mkdir -p /app/full \
     && cp build/ReleaseOV/bin/* /app/full/ \
     && cp *.py /app/full \
+    && cp -r conversion /app/full \
     && cp -r gguf-py /app/full \
     && cp -r requirements /app/full \
     && cp requirements.txt /app/full \
@@ -88,6 +93,18 @@ FROM ubuntu:${UBUNTU_VERSION} AS base
 # Pass proxy args to runtime stage
 ARG http_proxy
 ARG https_proxy
+ARG BUILD_DATE=N/A
+ARG APP_VERSION=N/A
+ARG APP_REVISION=N/A
+ARG IMAGE_URL=https://github.com/ggml-org/llama.cpp
+ARG IMAGE_SOURCE=https://github.com/ggml-org/llama.cpp
+LABEL org.opencontainers.image.created=$BUILD_DATE \
+      org.opencontainers.image.version=$APP_VERSION \
+      org.opencontainers.image.revision=$APP_REVISION \
+      org.opencontainers.image.title="llama.cpp" \
+      org.opencontainers.image.description="LLM inference in C/C++" \
+      org.opencontainers.image.url=$IMAGE_URL \
+      org.opencontainers.image.source=$IMAGE_SOURCE
 
 RUN apt-get update \
     && apt-get install -y libgomp1 libtbb12 curl wget ocl-icd-libopencl1 \

.devops/rocm.Dockerfile

@@ -7,6 +7,10 @@ ARG AMDGPU_VERSION=7.2.1
 # Target the ROCm build image
 ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete
 
+ARG BUILD_DATE=N/A
+ARG APP_VERSION=N/A
+ARG APP_REVISION=N/A
+
 ### Build image
 FROM ${BASE_ROCM_DEV_CONTAINER} AS build
 
@@ -49,6 +53,7 @@ RUN mkdir -p /app/lib \
 RUN mkdir -p /app/full \
     && cp build/bin/* /app/full \
     && cp *.py /app/full \
+    && cp -r conversion /app/full \
     && cp -r gguf-py /app/full \
     && cp -r requirements /app/full \
     && cp requirements.txt /app/full \
@@ -57,6 +62,19 @@ RUN mkdir -p /app/full \
 ## Base image
 FROM ${BASE_ROCM_DEV_CONTAINER} AS base
 
+ARG BUILD_DATE=N/A
+ARG APP_VERSION=N/A
+ARG APP_REVISION=N/A
+ARG IMAGE_URL=https://github.com/ggml-org/llama.cpp
+ARG IMAGE_SOURCE=https://github.com/ggml-org/llama.cpp
+LABEL org.opencontainers.image.created=$BUILD_DATE \
+      org.opencontainers.image.version=$APP_VERSION \
+      org.opencontainers.image.revision=$APP_REVISION \
+      org.opencontainers.image.title="llama.cpp" \
+      org.opencontainers.image.description="LLM inference in C/C++" \
+      org.opencontainers.image.url=$IMAGE_URL \
+      org.opencontainers.image.source=$IMAGE_SOURCE
+
 RUN apt-get update \
     && apt-get install -y libgomp1 curl \
     && apt autoremove -y \

.devops/s390x.Dockerfile

@@ -1,5 +1,8 @@
 ARG GCC_VERSION=15.2.0
 ARG UBUNTU_VERSION=24.04
+ARG BUILD_DATE=N/A
+ARG APP_VERSION=N/A
+ARG APP_REVISION=N/A
 
 ### Build Llama.cpp stage
 FROM gcc:${GCC_VERSION} AS build
@@ -34,6 +37,7 @@ RUN --mount=type=cache,target=/root/.ccache \
 
 COPY *.py             /opt/llama.cpp/bin
 COPY .devops/tools.sh /opt/llama.cpp/bin
+COPY conversion       /opt/llama.cpp/conversion
 
 COPY gguf-py          /opt/llama.cpp/gguf-py
 COPY requirements.txt /opt/llama.cpp/gguf-py
@@ -44,14 +48,28 @@ COPY requirements     /opt/llama.cpp/gguf-py/requirements
 FROM scratch AS collector
 
 # Copy llama.cpp binaries and libraries
-COPY --from=build /opt/llama.cpp/bin     /llama.cpp/bin
-COPY --from=build /opt/llama.cpp/lib     /llama.cpp/lib
-COPY --from=build /opt/llama.cpp/gguf-py /llama.cpp/gguf-py
+COPY --from=build /opt/llama.cpp/bin        /llama.cpp/bin
+COPY --from=build /opt/llama.cpp/lib        /llama.cpp/lib
+COPY --from=build /opt/llama.cpp/gguf-py    /llama.cpp/gguf-py
+COPY --from=build /opt/llama.cpp/conversion /llama.cpp/conversion
 
 
 ### Base image
 FROM ubuntu:${UBUNTU_VERSION} AS base
 
+ARG BUILD_DATE=N/A
+ARG APP_VERSION=N/A
+ARG APP_REVISION=N/A
+ARG IMAGE_URL=https://github.com/ggml-org/llama.cpp
+ARG IMAGE_SOURCE=https://github.com/ggml-org/llama.cpp
+LABEL org.opencontainers.image.created=$BUILD_DATE \
+      org.opencontainers.image.version=$APP_VERSION \
+      org.opencontainers.image.revision=$APP_REVISION \
+      org.opencontainers.image.title="llama.cpp" \
+      org.opencontainers.image.description="LLM inference in C/C++" \
+      org.opencontainers.image.url=$IMAGE_URL \
+      org.opencontainers.image.source=$IMAGE_SOURCE
+
 RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
     --mount=type=cache,target=/var/lib/apt/lists,sharing=locked \
     apt update -y && \
@@ -91,6 +109,7 @@ RUN curl https://sh.rustup.rs -sSf | bash -s -- -y
 
 COPY --from=collector /llama.cpp/bin /app
 COPY --from=collector /llama.cpp/gguf-py /app/gguf-py
+COPY --from=collector /llama.cpp/conversion /app/conversion
 
 RUN pip install --no-cache-dir --break-system-packages \
         -r /app/gguf-py/requirements.txt

.devops/vulkan.Dockerfile

@@ -1,4 +1,7 @@
 ARG UBUNTU_VERSION=26.04
+ARG BUILD_DATE=N/A
+ARG APP_VERSION=N/A
+ARG APP_REVISION=N/A
 
 FROM ubuntu:$UBUNTU_VERSION AS build
 
@@ -23,6 +26,7 @@ RUN mkdir -p /app/lib && \
 RUN mkdir -p /app/full \
     && cp build/bin/* /app/full \
     && cp *.py /app/full \
+    && cp -r conversion /app/full \
     && cp -r gguf-py /app/full \
     && cp -r requirements /app/full \
     && cp requirements.txt /app/full \
@@ -31,6 +35,19 @@ RUN mkdir -p /app/full \
 ## Base image
 FROM ubuntu:$UBUNTU_VERSION AS base
 
+ARG BUILD_DATE=N/A
+ARG APP_VERSION=N/A
+ARG APP_REVISION=N/A
+ARG IMAGE_URL=https://github.com/ggml-org/llama.cpp
+ARG IMAGE_SOURCE=https://github.com/ggml-org/llama.cpp
+LABEL org.opencontainers.image.created=$BUILD_DATE \
+      org.opencontainers.image.version=$APP_VERSION \
+      org.opencontainers.image.revision=$APP_REVISION \
+      org.opencontainers.image.title="llama.cpp" \
+      org.opencontainers.image.description="LLM inference in C/C++" \
+      org.opencontainers.image.url=$IMAGE_URL \
+      org.opencontainers.image.source=$IMAGE_SOURCE
+
 RUN apt-get update \
     && apt-get install -y libgomp1 curl libvulkan1 mesa-vulkan-drivers \
     libglvnd0 libgl1 libglx0 libegl1 libgles2 \

.github/ISSUE_TEMPLATE/011-bug-results.yml

@@ -100,8 +100,8 @@ body:
       label: Relevant log output
       description: >
           Please copy and paste any relevant log output, including the command that you entered and any generated text.
-          For very long logs (thousands of lines), preferably upload them as files instead.
-          On Linux you can redirect console output into a file by appending ` > llama.log 2>&1` to your command.
+          For very long logs (thousands of lines), please upload them as files instead; the `--log-file` CLI argument can be used for this purpose.
+          On Linux you can alternatively redirect the console output of any command into a file by appending ` > llama.log 2>&1` to your command.
       value: |
         <details>
         <summary>Logs</summary>

.github/ISSUE_TEMPLATE/019-bug-misc.yml

@@ -88,8 +88,8 @@ body:
       description: >
           If applicable, please copy and paste any relevant log output, including any generated text.
           If you are encountering problems specifically with the `llama_params_fit` module, always upload `--verbose` logs as well.
-          For very long logs (thousands of lines), please upload them as files instead.
-          On Linux you can redirect console output into a file by appending ` > llama.log 2>&1` to your command.
+          For very long logs (thousands of lines), please upload them as files instead; the `--log-file` CLI argument can be used for this purpose.
+          On Linux you can alternatively redirect the console output of any command into a file by appending ` > llama.log 2>&1` to your command.
       value: |
         <details>
         <summary>Logs</summary>

.github/workflows/build-and-test-snapdragon.yml

@@ -31,7 +31,7 @@ jobs:
   android-ndk-snapdragon:
     runs-on: ubuntu-latest
     container:
-      image: 'ghcr.io/snapdragon-toolchain/arm64-android:v0.3'
+      image: 'ghcr.io/snapdragon-toolchain/arm64-android:v0.6'
     defaults:
       run:
         shell: bash
@@ -61,7 +61,7 @@ jobs:
   linux-iot-snapdragon:
     runs-on: ubuntu-latest
     container:
-      image: 'ghcr.io/snapdragon-toolchain/arm64-linux:v0.1'
+      image: 'ghcr.io/snapdragon-toolchain/arm64-linux:v0.6'
     defaults:
       run:
         shell: bash

.github/workflows/build-apple.yml

@@ -59,6 +59,7 @@ jobs:
           cmake -B build -G Xcode \
             -DGGML_METAL_USE_BF16=ON \
             -DGGML_METAL_EMBED_LIBRARY=ON \
+            -DLLAMA_BUILD_APP=OFF \
             -DLLAMA_BUILD_COMMON=OFF \
             -DLLAMA_BUILD_EXAMPLES=OFF \
             -DLLAMA_BUILD_TOOLS=OFF \
@@ -89,6 +90,7 @@ jobs:
             -DGGML_METAL_USE_BF16=ON \
             -DGGML_METAL_EMBED_LIBRARY=ON \
             -DLLAMA_OPENSSL=OFF \
+            -DLLAMA_BUILD_APP=OFF \
             -DLLAMA_BUILD_EXAMPLES=OFF \
             -DLLAMA_BUILD_TOOLS=OFF \
             -DLLAMA_BUILD_TESTS=OFF \
@@ -138,6 +140,7 @@ jobs:
             -DGGML_METAL_USE_BF16=ON \
             -DGGML_METAL_EMBED_LIBRARY=ON \
             -DLLAMA_BUILD_COMMON=OFF \
+            -DLLAMA_BUILD_APP=OFF \
             -DLLAMA_BUILD_EXAMPLES=OFF \
             -DLLAMA_BUILD_TOOLS=OFF \
             -DLLAMA_BUILD_TESTS=OFF \
@@ -163,6 +166,7 @@ jobs:
             -DGGML_METAL_USE_BF16=ON \
             -DGGML_METAL_EMBED_LIBRARY=ON \
             -DLLAMA_BUILD_COMMON=OFF \
+            -DLLAMA_BUILD_APP=OFF \
             -DLLAMA_BUILD_EXAMPLES=OFF \
             -DLLAMA_BUILD_TOOLS=OFF \
             -DLLAMA_BUILD_TESTS=OFF \
@@ -206,6 +210,7 @@ jobs:
             -DGGML_METAL_USE_BF16=ON \
             -DGGML_METAL_EMBED_LIBRARY=ON \
             -DLLAMA_OPENSSL=OFF \
+            -DLLAMA_BUILD_APP=OFF \
             -DLLAMA_BUILD_EXAMPLES=OFF \
             -DLLAMA_BUILD_TOOLS=OFF \
             -DLLAMA_BUILD_TESTS=OFF \

.github/workflows/build-cmake-pkg.yml

@@ -19,9 +19,14 @@ jobs:
       - name: Build
         run: |
           PREFIX="$(pwd)"/inst
-          cmake -S . -B build -DCMAKE_PREFIX_PATH="$PREFIX" \
-                -DLLAMA_OPENSSL=OFF -DLLAMA_BUILD_TESTS=OFF -DLLAMA_BUILD_TOOLS=OFF \
-                -DLLAMA_BUILD_EXAMPLES=OFF -DCMAKE_BUILD_TYPE=Release
+          cmake -S . -B build \
+                -DCMAKE_PREFIX_PATH="$PREFIX" \
+                -DLLAMA_OPENSSL=OFF \
+                -DLLAMA_BUILD_TESTS=OFF \
+                -DLLAMA_BUILD_TOOLS=OFF \
+                -DLLAMA_BUILD_EXAMPLES=OFF \
+                -DLLAMA_BUILD_APP=OFF \
+                -DCMAKE_BUILD_TYPE=Release
           cmake --build build --config Release
           cmake --install build --prefix "$PREFIX" --config Release
 

.github/workflows/docker.yml

@@ -11,6 +11,11 @@ name: Publish Docker image
 
 on:
   workflow_dispatch: # allows manual triggering
+    inputs:
+      skip_s390x:
+        description: "Skip the s390x build target (useful for fast test runs that do not need the IBM Z runner)"
+        type: boolean
+        default: false
   schedule:
     # Rebuild daily rather than on every push because it is expensive
     - cron: '12 4 * * *'
@@ -64,6 +69,8 @@ jobs:
       - name: Generate build and merge matrices
         id: matrices
         shell: bash
+        env:
+          SKIP_S390X: ${{ inputs.skip_s390x || 'false' }}
         run: |
           set -euo pipefail
 
@@ -86,6 +93,11 @@ jobs:
           ]
           JSON
 
+          if [ "${SKIP_S390X}" = "true" ]; then
+            jq 'map(select(.platforms != "linux/s390x"))' build-matrix.json > build-matrix.json.tmp
+            mv build-matrix.json.tmp build-matrix.json
+          fi
+
           BUILD_MATRIX="$(jq -c . build-matrix.json)"
           MERGE_MATRIX="$(jq -c '
             reduce .[] as $entry ({}; .[$entry.tag] |= (
@@ -132,6 +144,7 @@ jobs:
         config: ${{ fromJSON(needs.prepare_matrices.outputs.build_matrix) }}
     steps:
       - name: Check out the repo
+        id: checkout
         uses: actions/checkout@v6
         with:
           fetch-depth: 0
@@ -187,6 +200,10 @@ jobs:
         env:
           GITHUB_REPOSITORY_OWNER: '${{ github.repository_owner }}'
 
+      - name: Get build date
+        id: build_date
+        run: echo "date=$(date -u +"%Y-%m-%dT%H:%M:%SZ")" >> $GITHUB_OUTPUT
+
       - name: Free Disk Space (Ubuntu)
         if: ${{ matrix.config.free_disk_space == true }}
         uses: ggml-org/free-disk-space@v1.3.1
@@ -211,13 +228,26 @@ jobs:
         with:
           context: .
           platforms: ${{ matrix.config.platforms }}
-          outputs: type=image,name=${{ steps.meta.outputs.image_repo }},push-by-digest=true,name-canonical=true,push=true
+          outputs: type=image,name=${{ steps.meta.outputs.image_repo }},push-by-digest=true,name-canonical=true,push=true,oci-mediatypes=true
           file: ${{ matrix.config.dockerfile }}
           target: full
           provenance: false
           build-args: |
+            BUILD_DATE=${{ steps.build_date.outputs.date }}
+            APP_VERSION=${{ needs.create_tag.outputs.source_tag }}
+            APP_REVISION=${{ steps.checkout.outputs.commit }}
+            IMAGE_URL=${{ github.server_url }}/${{ github.repository }}
+            IMAGE_SOURCE=${{ github.server_url }}/${{ github.repository }}
             ${{ matrix.config.ubuntu_version && format('UBUNTU_VERSION={0}', matrix.config.ubuntu_version) || '' }}
             ${{ matrix.config.cuda_version && format('CUDA_VERSION={0}', matrix.config.cuda_version) || '' }}
+          annotations: |
+            manifest:org.opencontainers.image.created=${{ steps.build_date.outputs.date }}
+            manifest:org.opencontainers.image.version=${{ needs.create_tag.outputs.source_tag }}
+            manifest:org.opencontainers.image.revision=${{ steps.checkout.outputs.commit }}
+            manifest:org.opencontainers.image.title=llama.cpp
+            manifest:org.opencontainers.image.description=LLM inference in C/C++
+            manifest:org.opencontainers.image.url=${{ github.server_url }}/${{ github.repository }}
+            manifest:org.opencontainers.image.source=${{ github.server_url }}/${{ github.repository }}
           # using github experimental cache
           #cache-from: type=gha
           #cache-to: type=gha,mode=max
@@ -235,13 +265,26 @@ jobs:
         with:
           context: .
           platforms: ${{ matrix.config.platforms }}
-          outputs: type=image,name=${{ steps.meta.outputs.image_repo }},push-by-digest=true,name-canonical=true,push=true
+          outputs: type=image,name=${{ steps.meta.outputs.image_repo }},push-by-digest=true,name-canonical=true,push=true,oci-mediatypes=true
           file: ${{ matrix.config.dockerfile }}
           target: light
           provenance: false
           build-args: |
+            BUILD_DATE=${{ steps.build_date.outputs.date }}
+            APP_VERSION=${{ needs.create_tag.outputs.source_tag }}
+            APP_REVISION=${{ steps.checkout.outputs.commit }}
+            IMAGE_URL=${{ github.server_url }}/${{ github.repository }}
+            IMAGE_SOURCE=${{ github.server_url }}/${{ github.repository }}
             ${{ matrix.config.ubuntu_version && format('UBUNTU_VERSION={0}', matrix.config.ubuntu_version) || '' }}
             ${{ matrix.config.cuda_version && format('CUDA_VERSION={0}', matrix.config.cuda_version) || '' }}
+          annotations: |
+            manifest:org.opencontainers.image.created=${{ steps.build_date.outputs.date }}
+            manifest:org.opencontainers.image.version=${{ needs.create_tag.outputs.source_tag }}
+            manifest:org.opencontainers.image.revision=${{ steps.checkout.outputs.commit }}
+            manifest:org.opencontainers.image.title=llama.cpp
+            manifest:org.opencontainers.image.description=LLM inference in C/C++
+            manifest:org.opencontainers.image.url=${{ github.server_url }}/${{ github.repository }}
+            manifest:org.opencontainers.image.source=${{ github.server_url }}/${{ github.repository }}
           # using github experimental cache
           #cache-from: type=gha
           #cache-to: type=gha,mode=max
@@ -259,13 +302,26 @@ jobs:
         with:
           context: .
           platforms: ${{ matrix.config.platforms }}
-          outputs: type=image,name=${{ steps.meta.outputs.image_repo }},push-by-digest=true,name-canonical=true,push=true
+          outputs: type=image,name=${{ steps.meta.outputs.image_repo }},push-by-digest=true,name-canonical=true,push=true,oci-mediatypes=true
           file: ${{ matrix.config.dockerfile }}
           target: server
           provenance: false
           build-args: |
+            BUILD_DATE=${{ steps.build_date.outputs.date }}
+            APP_VERSION=${{ needs.create_tag.outputs.source_tag }}
+            APP_REVISION=${{ steps.checkout.outputs.commit }}
+            IMAGE_URL=${{ github.server_url }}/${{ github.repository }}
+            IMAGE_SOURCE=${{ github.server_url }}/${{ github.repository }}
             ${{ matrix.config.ubuntu_version && format('UBUNTU_VERSION={0}', matrix.config.ubuntu_version) || '' }}
             ${{ matrix.config.cuda_version && format('CUDA_VERSION={0}', matrix.config.cuda_version) || '' }}
+          annotations: |
+            manifest:org.opencontainers.image.created=${{ steps.build_date.outputs.date }}
+            manifest:org.opencontainers.image.version=${{ needs.create_tag.outputs.source_tag }}
+            manifest:org.opencontainers.image.revision=${{ steps.checkout.outputs.commit }}
+            manifest:org.opencontainers.image.title=llama.cpp
+            manifest:org.opencontainers.image.description=LLM inference in C/C++
+            manifest:org.opencontainers.image.url=${{ github.server_url }}/${{ github.repository }}
+            manifest:org.opencontainers.image.source=${{ github.server_url }}/${{ github.repository }}
           # using github experimental cache
           #cache-from: type=gha
           #cache-to: type=gha,mode=max
@@ -330,10 +386,15 @@ jobs:
 
     steps:
       - name: Check out the repo
+        id: checkout
         uses: actions/checkout@v6
         with:
           fetch-depth: 0
 
+      - name: Get build date
+        id: build_date
+        run: echo "date=$(date -u +"%Y-%m-%dT%H:%M:%SZ")" >> $GITHUB_OUTPUT
+
       - name: Download digest metadata
         uses: actions/download-artifact@3e5f45b2cfb9172054b4087a40e8e0b5a5461e7c # v8
         with:
@@ -361,6 +422,8 @@ jobs:
           IMAGE_REPO="ghcr.io/${REPO_OWNER}/${REPO_NAME}"
           PREFIX="${IMAGE_REPO}:"
           SRC_TAG="${{ needs.create_tag.outputs.source_tag }}"
+          BUILD_DATE="${{ steps.build_date.outputs.date }}"
+          COMMIT_SHA="${{ steps.checkout.outputs.commit }}"
           TAGS="${{ matrix.config.tag }}"
           ARCHES="${{ matrix.config.arches }}"
           DIGEST_GLOB="/tmp/digests/*.tsv"
@@ -412,11 +475,21 @@ jobs:
                   refs+=("${IMAGE_REPO}@${digest}")
               done
 
+              local annotations=(
+                  --annotation "index:org.opencontainers.image.created=${BUILD_DATE}"
+                  --annotation "index:org.opencontainers.image.version=${SRC_TAG}"
+                  --annotation "index:org.opencontainers.image.revision=${COMMIT_SHA}"
+                  --annotation "index:org.opencontainers.image.title=llama.cpp"
+                  --annotation "index:org.opencontainers.image.description=LLM inference in C/C++"
+                  --annotation "index:org.opencontainers.image.url=${{ github.server_url }}/${{ github.repository }}"
+                  --annotation "index:org.opencontainers.image.source=${{ github.server_url }}/${{ github.repository }}"
+              )
+
               echo "Creating ${merged_tag} from ${refs[*]}"
-              docker buildx imagetools create --tag "${merged_tag}" "${refs[@]}"
+              docker buildx imagetools create "${annotations[@]}" --tag "${merged_tag}" "${refs[@]}"
 
               echo "Creating ${merged_versioned_tag} from ${refs[*]}"
-              docker buildx imagetools create --tag "${merged_versioned_tag}" "${refs[@]}"
+              docker buildx imagetools create "${annotations[@]}" --tag "${merged_versioned_tag}" "${refs[@]}"
           }
 
           for tag in $TAGS; do

.github/workflows/release.yml

@@ -1108,6 +1108,7 @@ jobs:
             -DGGML_METAL_USE_BF16=ON \
             -DGGML_METAL_EMBED_LIBRARY=ON \
             -DLLAMA_OPENSSL=OFF \
+            -DLLAMA_BUILD_APP=OFF \
             -DLLAMA_BUILD_EXAMPLES=OFF \
             -DLLAMA_BUILD_TOOLS=OFF \
             -DLLAMA_BUILD_TESTS=OFF \
@@ -1233,6 +1234,9 @@ jobs:
           path: llama-${{ steps.tag.outputs.name }}-bin-${{ matrix.chip_type }}-openEuler-${{ matrix.arch }}${{ matrix.use_acl_graph == 'on' && '-aclgraph' || '' }}.tar.gz
           name: llama-bin-${{ matrix.chip_type }}-openEuler-${{ matrix.arch }}${{ matrix.use_acl_graph == 'on' && '-aclgraph' || '' }}.tar.gz
 
+  ui-build:
+    uses: ./.github/workflows/ui-build.yml
+
   release:
     if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
 
@@ -1258,6 +1262,7 @@ jobs:
       - macOS-cpu
       - ios-xcode-build
       - openEuler-cann
+      - ui-build
 
     outputs:
       tag_name: ${{ steps.tag.outputs.name }}
@@ -1317,6 +1322,18 @@ jobs:
           mv -v artifact/*.zip release
           mv -v artifact/*.tar.gz release
 
+      - name: Download UI build
+        id: download_ui
+        uses: actions/download-artifact@v7
+        with:
+          name: ui-build
+          path: ./ui-dist
+
+      - name: Package UI
+        id: package_ui
+        run: |
+          tar -czvf release/llama-${{ steps.tag.outputs.name }}-ui.tar.gz --transform "s,^\.,llama-${{ steps.tag.outputs.name }}," -C ./ui-dist .
+
       - name: Create release
         id: create_release
         uses: ggml-org/action-create-release@v1
@@ -1366,6 +1383,9 @@ jobs:
             - [openEuler aarch64 (310p)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-310p-openEuler-aarch64.tar.gz)
             - [openEuler aarch64 (910b, ACL Graph)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-910b-openEuler-aarch64-aclgraph.tar.gz)
 
+            **UI:**
+            - [UI](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-ui.tar.gz)
+
       - name: Upload release
         id: upload_release
         uses: actions/github-script@v8

.github/workflows/server-self-hosted.yml

@@ -130,3 +130,68 @@ jobs:
   #          pip install -r requirements.txt
   #          export ${{ matrix.extra_args }}
   #          pytest -v -x -m "not slow"
+
+  server-kleidiai:
+    runs-on: ah-ubuntu_22_04-c8g_8x
+
+    name: server-kleidiai (${{ matrix.wf_name }})
+    strategy:
+      matrix:
+        include:
+          - build_type: Release
+            extra_build_flags: "-DGGML_CPU_KLEIDIAI=ON"
+            extra_args: ""
+            wf_name:    "CPUx1, kleidiai"
+      fail-fast: false
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+        with:
+          fetch-depth: 0
+          ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
+
+      - name: Dependencies
+        id: depends
+        run: |
+          set -euxo pipefail
+          sudo apt-get update
+          sudo DEBIAN_FRONTEND=noninteractive NEEDRESTART_MODE=a \
+          apt-get install -y \
+           build-essential \
+           libssl-dev \
+           python3-venv \
+           gpg \
+           wget \
+           time \
+           git-lfs
+
+          git lfs install
+
+          # install the latest cmake
+          sudo install -d /usr/share/keyrings
+          wget -O - https://apt.kitware.com/keys/kitware-archive-latest.asc \
+           | gpg --dearmor \
+           | sudo tee /usr/share/keyrings/kitware-archive-keyring.gpg >/dev/null
+          echo 'deb [signed-by=/usr/share/keyrings/kitware-archive-keyring.gpg] https://apt.kitware.com/ubuntu/ jammy main' \
+           | sudo tee /etc/apt/sources.list.d/kitware.list
+          sudo apt-get update
+          sudo apt-get install -y cmake
+
+      - name: Build
+        id: cmake_build
+        run: |
+          cmake -B build -DGGML_SCHED_NO_REALLOC=ON ${{ matrix.extra_build_flags }}
+          cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server
+
+      - name: Tests
+        id: server_integration_tests
+        if: ${{ (!matrix.disabled_on_pr || !github.event.pull_request) }}
+        run: |
+          cd tools/server/tests
+          python3 -m venv venv
+          source venv/bin/activate
+          pip install -r requirements.txt
+          export ${{ matrix.extra_args }}
+          pytest -v -x -m "not slow"

.github/workflows/server.yml

@@ -54,8 +54,13 @@ concurrency:
   cancel-in-progress: true
 
 jobs:
+  ui-build:
+    name: Build Web UI
+    uses: ./.github/workflows/ui-build.yml
+
   server:
     runs-on: ubuntu-latest
+    needs: ui-build
 
     name: server (${{ matrix.wf_name }})
     strategy:
@@ -93,12 +98,11 @@ jobs:
           fetch-depth: 0
           ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
 
-      - name: Setup Node.js
-        uses: actions/setup-node@v6
+      - name: Download built UI
+        uses: actions/download-artifact@v7
         with:
-          node-version: "24"
-          cache: "npm"
-          cache-dependency-path: "tools/ui/package-lock.json"
+          name: ui-build
+          path: tools/ui/dist
 
       - name: Build
         id: cmake_build

.github/workflows/ui-build.yml

@@ -31,7 +31,7 @@ jobs:
 
       - name: Generate checksums
         run: |
-          cd build/tools/ui/dist
+          cd tools/ui/dist
           for f in *; do
             sha256sum "$f" | awk '{print $1, $2}' >> checksums.txt
           done
@@ -40,5 +40,5 @@ jobs:
         uses: actions/upload-artifact@v6
         with:
           name: ui-build
-          path: build/tools/ui/dist/
+          path: tools/ui/dist/
           retention-days: 1

.github/workflows/ui-ci.yml

@@ -41,7 +41,7 @@ jobs:
   ui-checks:
     name: UI Checks
     needs: ui-build
-    runs-on: ubuntu-24.04-arm
+    runs-on: ubuntu-latest
     continue-on-error: true
     steps:
       - name: Checkout code
@@ -93,7 +93,7 @@ jobs:
   e2e-tests:
     name: E2E Tests
     needs: ui-build
-    runs-on: ubuntu-24.04-arm
+    runs-on: ubuntu-latest
     steps:
       - name: Checkout code
         uses: actions/checkout@v6

.github/workflows/ui-publish.yml

@@ -38,7 +38,7 @@ jobs:
         uses: actions/download-artifact@v7
         with:
           name: ui-build
-          path: build/tools/ui/dist/
+          path: tools/ui/dist/
 
       - name: Install Hugging Face Hub CLI
         run: pip install -U huggingface_hub
@@ -49,12 +49,12 @@ jobs:
       - name: Sync built files to Hugging Face bucket (version tag)
         run: |
           # Upload the built files to the Hugging Face bucket under the release version
-          hf buckets sync build/tools/ui/dist hf://buckets/ggml-org/${{ env.HF_BUCKET_NAME }}/${{ inputs.version_tag }} --delete --quiet
+          hf buckets sync tools/ui/dist hf://buckets/ggml-org/${{ env.HF_BUCKET_NAME }}/${{ inputs.version_tag }} --delete --quiet
 
       - name: Sync built files to Hugging Face bucket (latest)
         run: |
           # Also upload to the 'latest' directory for fallback downloads
-          hf buckets sync build/tools/ui/dist hf://buckets/ggml-org/${{ env.HF_BUCKET_NAME }}/latest --delete --quiet
+          hf buckets sync tools/ui/dist hf://buckets/ggml-org/${{ env.HF_BUCKET_NAME }}/latest --delete --quiet
 
       - name: Verify upload
         run: |

.pi/gg/SYSTEM.md

@@ -1,7 +1,7 @@
 You are a coding agent. Here are some very important rules that you must follow:
 
 General:
-- By very precise and concise when writing code, comments, explanations, etc.
+- Be very precise and concise when writing code, comments, explanations, etc.
 - PR and commit titles format: `<module> : <title>`. Lookup recents for examples
 - Don't try to build or run the code unless you are explicitly asked to do so
 - Use the `gh` CLI tool when querying PRs, issues, or other GitHub resources
@@ -16,12 +16,15 @@ Pull requests (PRs):
 - New branch names are prefixed with "gg/"
 - Before opening a pull request, ask the user to confirm the description
 - When creating a pull request, look for the repository's PR template and follow it
-- For the AI usage disclosure section, write "YES. llama.cpp + pi"
+- For the AI usage disclosure section, write "YES. llama.cpp + pi + [MODEL]"
+- Ask the user to tell you what model was used and write it in place of [MODEL]
 - Always create the pull requests in draft mode
 
 Commits:
 - On every commit that you make, include a "Assisted-by: llama.cpp:local pi" tag
 - Do not explicitly set the git author in commits - rely on the default git config
+- Always use `--no-gpg-sign` when committing
+- Never `git push` without explicit confirmation from the user
 
 Resources (read on demand):
 - [CONTRIBUTING.md](CONTRIBUTING.md)

CMakeLists.txt

@@ -104,29 +104,16 @@ option(LLAMA_SANITIZE_UNDEFINED "llama: enable undefined sanitizer" OFF)
 option(LLAMA_BUILD_COMMON "llama: build common utils library" ${LLAMA_STANDALONE})
 
 # extra artifacts
-option(LLAMA_BUILD_TESTS            "llama: build tests"                                                                            ${LLAMA_STANDALONE})
-option(LLAMA_BUILD_TOOLS            "llama: build tools"                                                                            ${LLAMA_STANDALONE})
-option(LLAMA_BUILD_EXAMPLES         "llama: build examples"                                                                         ${LLAMA_STANDALONE})
-option(LLAMA_BUILD_SERVER           "llama: build server example"                                                                   ${LLAMA_STANDALONE})
-# Deprecated: use LLAMA_BUILD_UI instead (kept for backward compat)
-option(LLAMA_BUILD_WEBUI            "llama: build the embedded Web UI for server (deprecated: use LLAMA_BUILD_UI)"                   ON)
-option(LLAMA_USE_PREBUILT_WEBUI     "llama: use prebuilt WebUI from HF Bucket when available (deprecated: use LLAMA_USE_PREBUILT_UI)" ON)
+option(LLAMA_BUILD_TESTS     "llama: build tests"                                                                ${LLAMA_STANDALONE})
+option(LLAMA_BUILD_TOOLS     "llama: build tools"                                                                ${LLAMA_STANDALONE})
+option(LLAMA_BUILD_EXAMPLES  "llama: build examples"                                                             ${LLAMA_STANDALONE})
+option(LLAMA_BUILD_SERVER    "llama: build server example"                                                       ${LLAMA_STANDALONE})
+option(LLAMA_BUILD_APP       "llama: build the unified binary"                                                   ${LLAMA_STANDALONE})
+option(LLAMA_BUILD_UI        "llama: build the embedded Web UI for server"                                       ON)
+option(LLAMA_USE_PREBUILT_UI "llama: use prebuilt UI from HF Bucket when available (requires LLAMA_BUILD_UI=ON)" ON)
 
-# New option names
-option(LLAMA_BUILD_UI                "llama: build the embedded Web UI for server"                                                   ON)
-option(LLAMA_USE_PREBUILT_UI         "llama: use prebuilt UI from HF Bucket when available (requires LLAMA_BUILD_UI=ON)"             ON)
-
-# Backward compat: when old var is set but new one isn't, forward the value
-if(DEFINED LLAMA_BUILD_WEBUI AND NOT DEFINED LLAMA_BUILD_UI)
-    set(LLAMA_BUILD_UI ${LLAMA_BUILD_WEBUI})
-    message(DEPRECATION "LLAMA_BUILD_WEBUI is deprecated, use LLAMA_BUILD_UI instead")
-endif()
-if(DEFINED LLAMA_USE_PREBUILT_WEBUI AND NOT DEFINED LLAMA_USE_PREBUILT_UI)
-    set(LLAMA_USE_PREBUILT_UI ${LLAMA_USE_PREBUILT_WEBUI})
-    message(DEPRECATION "LLAMA_USE_PREBUILT_WEBUI is deprecated, use LLAMA_USE_PREBUILT_UI instead")
-endif()
-option(LLAMA_TOOLS_INSTALL          "llama: install tools"                                                                          ${LLAMA_TOOLS_INSTALL_DEFAULT})
-option(LLAMA_TESTS_INSTALL          "llama: install tests"                                                                          ON)
+option(LLAMA_TOOLS_INSTALL "llama: install tools" ${LLAMA_TOOLS_INSTALL_DEFAULT})
+option(LLAMA_TESTS_INSTALL "llama: install tests" ON)
 
 # 3rd party libs
 option(LLAMA_OPENSSL    "llama: use openssl to support HTTPS" ON)
@@ -231,6 +218,10 @@ if (LLAMA_BUILD_COMMON AND LLAMA_BUILD_TOOLS)
     add_subdirectory(tools)
 endif()
 
+if (LLAMA_BUILD_APP)
+    add_subdirectory(app)
+endif()
+
 # Automatically add all files from the 'licenses' directory
 file(GLOB EXTRA_LICENSES "${CMAKE_SOURCE_DIR}/licenses/LICENSE-*")
 
@@ -286,18 +277,6 @@ install(FILES ${CMAKE_CURRENT_BINARY_DIR}/llama-config.cmake
               ${CMAKE_CURRENT_BINARY_DIR}/llama-version.cmake
         DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/llama)
 
-install(
-    FILES convert_hf_to_gguf.py
-    PERMISSIONS
-        OWNER_READ
-        OWNER_WRITE
-        OWNER_EXECUTE
-        GROUP_READ
-        GROUP_EXECUTE
-        WORLD_READ
-        WORLD_EXECUTE
-    DESTINATION ${CMAKE_INSTALL_BINDIR})
-
 configure_file(cmake/llama.pc.in
         "${CMAKE_CURRENT_BINARY_DIR}/llama.pc"
         @ONLY)

CODEOWNERS

@@ -26,6 +26,7 @@
 /common/fit.*                           @JohannesGaessler
 /common/jinja/                          @CISC
 /common/ngram-map.*                     @srogmann
+/conversion/                            @CISC
 /convert_*.py                           @CISC
 /docs/backend/snapdragon/               @ggml-org/ggml-hexagon
 /examples/batched.swift/                @ggerganov
@@ -48,7 +49,6 @@
 /examples/parallel/                     @ggerganov
 /examples/passkey/                      @ggerganov
 /examples/retrieval/                    @ggerganov
-/examples/save-load-state/              @ggerganov
 /examples/speculative-simple/           @ggerganov
 /examples/speculative/                  @ggerganov
 /ggml/cmake/                            @ggerganov

README.md

@@ -27,6 +27,7 @@ LLM inference in C/C++
 - Vim/Neovim plugin for FIM completions: https://github.com/ggml-org/llama.vim
 - Hugging Face Inference Endpoints now support GGUF out of the box! https://github.com/ggml-org/llama.cpp/discussions/9669
 - Hugging Face GGUF editor: [discussion](https://github.com/ggml-org/llama.cpp/discussions/9268) | [tool](https://huggingface.co/spaces/CISCai/gguf-editor)
+- WebGPU support is now available in the browser, see a blog/demo introducing it [here](https://reeselevine.github.io/llamas-on-the-web/).
 
 ----
 
@@ -280,7 +281,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 | [Metal](docs/build.md#metal-build) | Apple Silicon |
 | [BLAS](docs/build.md#blas-build) | All |
 | [BLIS](docs/backend/BLIS.md) | All |
-| [SYCL](docs/backend/SYCL.md) | Intel and Nvidia GPU |
+| [SYCL](docs/backend/SYCL.md) | Intel GPU |
 | [OpenVINO [In Progress]](docs/backend/OPENVINO.md) | Intel CPUs, GPUs, and NPUs |
 | [MUSA](docs/build.md#musa) | Moore Threads GPU |
 | [CUDA](docs/build.md#cuda) | Nvidia GPU |
@@ -290,7 +291,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 | [CANN](docs/build.md#cann) | Ascend NPU |
 | [OpenCL](docs/backend/OPENCL.md) | Adreno GPU |
 | [IBM zDNN](docs/backend/zDNN.md) | IBM Z & LinuxONE |
-| [WebGPU [In Progress]](docs/build.md#webgpu) | All |
+| [WebGPU](docs/build.md#webgpu) | All |
 | [RPC](https://github.com/ggml-org/llama.cpp/tree/master/tools/rpc) | All |
 | [Hexagon [In Progress]](docs/backend/snapdragon/README.md) | Snapdragon |
 | [VirtGPU](docs/backend/VirtGPU.md) | VirtGPU APIR |

app/CMakeLists.txt

@@ -0,0 +1,20 @@
+set(TARGET llama-app)
+
+add_executable(${TARGET} llama.cpp)
+set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama)
+
+target_link_libraries(${TARGET} PRIVATE
+    llama-server-impl
+    llama-cli-impl
+    llama-completion-impl
+    llama-bench-impl
+    llama-batched-bench-impl
+    llama-fit-params-impl
+    llama-quantize-impl
+    llama-perplexity-impl
+)
+target_compile_features(${TARGET} PRIVATE cxx_std_17)
+
+if(LLAMA_TOOLS_INSTALL)
+    install(TARGETS ${TARGET} RUNTIME)
+endif()

app/llama.cpp

@@ -0,0 +1,95 @@
+#include "build-info.h"
+
+#include <cstdio>
+#include <cstdlib>
+#include <string>
+#include <vector>
+
+// visible
+int llama_server(int argc, char ** argv);
+int llama_cli(int argc, char ** argv);
+
+// hidden
+int llama_completion(int argc, char ** argv);
+int llama_bench(int argc, char ** argv);
+int llama_batched_bench(int argc, char ** argv);
+int llama_fit_params(int argc, char ** argv);
+int llama_quantize(int argc, char ** argv);
+int llama_perplexity(int argc, char ** argv);
+
+static int help(int argc, char ** argv);
+static int version(int argc, char ** argv);
+
+struct command {
+    const char * name;
+    const char * desc;
+    std::vector<std::string> aliases;
+    bool hidden;
+    int (*func)(int, char **);
+};
+
+static const command cmds[] = {
+    {"serve",         "HTTP API server",                                    {"server"},   false, llama_server       },
+    {"cli",           "Command-line interactive interface",                 {"client"},   false, llama_cli          },
+    {"completion",    "Text completion",                                    {"complete"}, true,  llama_completion   },
+    {"bench",         "Benchmark prompt processing and text generation",    {},           true,  llama_bench        },
+    {"batched-bench", "Benchmark batched decoding performance",             {},           true,  llama_batched_bench},
+    {"fit-params",    "Compute parameters to fit a model in device memory", {},           true,  llama_fit_params   },
+    {"quantize",      "Quantize a model",                                   {},           true,  llama_quantize     },
+    {"perplexity",    "Compute model perplexity and KL divergence",         {},           true,  llama_perplexity   },
+    {"version",       "Show version",                                       {},           true,  version            },
+    {"help",          "Show available commands",                            {},           true,  help               },
+};
+
+static int version(int argc, char ** argv) {
+    printf("%s\n", llama_build_info());
+    return 0;
+}
+
+static int help(int argc, char ** argv) {
+    const bool show_all = argc >= 2 && std::string(argv[1]) == "all";
+
+    printf("Usage: llama <command> [options]\n\nAvailable commands:\n");
+
+    for (const auto & cmd : cmds) {
+        if (show_all || !cmd.hidden) {
+            printf("  %-15s %s\n", cmd.name, cmd.desc);
+        }
+    }
+    printf("\nRun 'llama <command> --help' for command-specific usage.\n");
+
+    return 0;
+}
+
+static bool matches(const std::string & arg, const command & cmd) {
+    if (arg == cmd.name) {
+        return true;
+    }
+    for (const auto & alias : cmd.aliases) {
+        if (arg == alias) {
+            return true;
+        }
+    }
+    return false;
+}
+
+int main(int argc, char ** argv) {
+    const std::string arg = argc >= 2 ? argv[1] : "help";
+
+    for (const auto & cmd : cmds) {
+        if (matches(arg, cmd)) {
+
+            // router spawns children through this same binary, it needs the
+            // subcommand to relaunch as 'llama serve' and not bare options
+#ifdef _WIN32
+            _putenv_s("LLAMA_APP_CMD", cmd.name);
+#else
+            setenv("LLAMA_APP_CMD", cmd.name, 1);
+#endif
+            return cmd.func(argc - 1, argv + 1);
+        }
+    }
+
+    fprintf(stderr, "error: unknown command '%s'\n", arg.c_str());
+    return 1;
+}

build-xcframework.sh

@@ -7,6 +7,7 @@ VISIONOS_MIN_OS_VERSION=1.0
 TVOS_MIN_OS_VERSION=16.4
 
 BUILD_SHARED_LIBS=OFF
+LLAMA_BUILD_APP=OFF
 LLAMA_BUILD_EXAMPLES=OFF
 LLAMA_BUILD_TOOLS=OFF
 LLAMA_BUILD_TESTS=OFF
@@ -31,6 +32,7 @@ COMMON_CMAKE_ARGS=(
     -DCMAKE_XCODE_ATTRIBUTE_STRIP_INSTALLED_PRODUCT=NO
     -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
     -DBUILD_SHARED_LIBS=${BUILD_SHARED_LIBS}
+    -DLLAMA_BUILD_APP=${LLAMA_BUILD_APP}
     -DLLAMA_BUILD_EXAMPLES=${LLAMA_BUILD_EXAMPLES}
     -DLLAMA_BUILD_TOOLS=${LLAMA_BUILD_TOOLS}
     -DLLAMA_BUILD_TESTS=${LLAMA_BUILD_TESTS}

ci/run.sh

@@ -117,6 +117,12 @@ if [ ! -z ${GG_BUILD_VULKAN} ]; then
     # if on Mac, disable METAL
     if [[ "$OSTYPE" == "darwin"* ]]; then
         CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_METAL=OFF -DGGML_BLAS=OFF"
+
+        MACOS_RUNNER_CUSTOM_VULKAN_CMAKE_LOCATION="/usr/local/lib/cmake/vulkan"
+        MACOS_RUNNER_CUSTOM_SPIRV_HEADERS_LOCATION="${MACOS_RUNNER_CUSTOM_VULKAN_CMAKE_LOCATION}/SPIRV-Headers/SPIRV-HeadersConfig.cmake"
+        if [[ -f "${MACOS_RUNNER_CUSTOM_SPIRV_HEADERS_LOCATION}" || -h "${MACOS_RUNNER_CUSTOM_SPIRV_HEADERS_LOCATION}" ]]; then
+            CMAKE_EXTRA="${CMAKE_EXTRA} -DSPIRV-Headers_DIR=${MACOS_RUNNER_CUSTOM_VULKAN_CMAKE_LOCATION}/SPIRV-Headers"
+        fi
     fi
 
     # Build shared libs on Windows
@@ -455,10 +461,10 @@ function gg_run_qwen3_0_6b {
 
     (time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
 
-    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 1024 -fa off --no-op-offload) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 1024 -fa on  --no-op-offload) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 1024 -fa off                ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 1024 -fa on                 ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/test-save-load-state --model ${model_q4_0} -ngl 10 -c 1024 -fa off --no-op-offload) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/test-save-load-state --model ${model_q4_0} -ngl 10 -c 1024 -fa on  --no-op-offload) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/test-save-load-state --model ${model_q4_0} -ngl 99 -c 1024 -fa off                ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/test-save-load-state --model ${model_q4_0} -ngl 99 -c 1024 -fa on                 ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
 
     function check_ppl {
         qnt="$1"

cmake/llama-config.cmake.in

@@ -7,7 +7,7 @@ set(LLAMA_SHARED_LIB   @BUILD_SHARED_LIBS@)
 
 set_and_check(LLAMA_INCLUDE_DIR "@PACKAGE_LLAMA_INCLUDE_INSTALL_DIR@")
 set_and_check(LLAMA_LIB_DIR     "@PACKAGE_LLAMA_LIB_INSTALL_DIR@")
-set_and_check(LLAMA_BIN_DIR     "@PACKAGE_LLAMA_BIN_INSTALL_DIR@")
+set(LLAMA_BIN_DIR "@PACKAGE_LLAMA_BIN_INSTALL_DIR@")
 
 find_package(ggml REQUIRED HINTS ${LLAMA_LIB_DIR}/cmake)
 

common/arg.cpp

@@ -4,7 +4,6 @@
 #include "chat.h"
 #include "common.h"
 #include "download.h"
-#include "hf-cache.h"
 #include "json-schema-to-grammar.h"
 #include "log.h"
 #include "sampling.h"
@@ -537,7 +536,11 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
                 throw std::invalid_argument(string_format("error: invalid argument: %s", arg.c_str()));
             }
             if (!seen_args.insert(arg).second) {
-                LOG_WRN("DEPRECATED: argument '%s' specified multiple times, use comma-separated values instead (only last value will be used)\n", arg.c_str());
+                const bool skip = (arg == "--spec-type");
+
+                if (!skip) {
+                    LOG_WRN("DEPRECATED: argument '%s' specified multiple times, use comma-separated values instead (only last value will be used)\n", arg.c_str());
+                }
             }
             auto & tmp = arg_to_options[arg];
             auto opt = *tmp.first;
@@ -586,12 +589,6 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
     // parse the first time to get -hf option (used for remote preset)
     parse_cli_args();
 
-    // TODO: Remove later
-    try {
-        hf_cache::migrate_old_cache_to_hf_cache(params.hf_token, params.offline);
-    } catch (const std::exception & e) {
-        LOG_WRN("HF cache migration failed: %s\n", e.what());
-    }
     // export_graph_ops loads only metadata
     const bool skip_model_download = ctx_arg.ex == LLAMA_EXAMPLE_EXPORT_GRAPH_OPS;
 
@@ -900,7 +897,11 @@ bool common_params_to_map(int argc, char ** argv, llama_example ex, std::map<com
             throw std::invalid_argument(string_format("error: invalid argument: %s", arg.c_str()));
         }
         if (!seen_args.insert(arg).second) {
-            LOG_WRN("DEPRECATED: argument '%s' specified multiple times, use comma-separated values instead (only last value will be used)\n", arg.c_str());
+            const bool skip = (arg == "--spec-type");
+
+            if (!skip) {
+                LOG_WRN("DEPRECATED: argument '%s' specified multiple times, use comma-separated values instead (only last value will be used)\n", arg.c_str());
+            }
         }
         auto opt = *arg_to_options[arg];
         std::string val;
@@ -2808,7 +2809,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, int value) {
             params.embd_normalize = value;
         }
-    ).set_examples({LLAMA_EXAMPLE_EMBEDDING, LLAMA_EXAMPLE_DEBUG}));
+    ).set_examples({LLAMA_EXAMPLE_EMBEDDING, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_DEBUG}));
     add_opt(common_arg(
         {"--embd-output-format"}, "FORMAT",
         "empty = default, \"array\" = [[],[]...], \"json\" = openai style, \"json+\" = same \"json\" + cosine similarity matrix, \"raw\" = plain whitespace-delimited output (one embedding per line)",
@@ -3363,7 +3364,8 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             " - 1: error\n"
             " - 2: warning\n"
             " - 3: info\n"
-            " - 4: debug\n"
+            " - 4: trace (more info)\n"
+            " - 5: debug\n"
             "(default: %d)\n", params.verbosity),
         [](common_params & params, int value) {
             params.verbosity = value;
@@ -3589,6 +3591,15 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.speculative.draft.p_min = std::stof(value);
         }
     ).set_spec().set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_SPEC_DRAFT_P_MIN"));
+    add_opt(common_arg(
+        {"--spec-draft-backend-sampling"},
+        {"--no-spec-draft-backend-sampling"},
+        string_format("offload draft sampling to the backend (default: %s)",
+                      params.speculative.draft.backend_sampling ? "enabled" : "disabled"),
+        [](common_params & params, bool value) {
+            params.speculative.draft.backend_sampling = value;
+        }
+    ).set_spec().set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_SPEC_DRAFT_BACKEND_SAMPLING"));
     add_opt(common_arg(
         {"--spec-draft-device", "-devd", "--device-draft"}, "<dev1,dev2,..>",
         "comma-separated list of devices to use for offloading the draft model (none = don't offload)\n"
@@ -4124,6 +4135,12 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.speculative.ngram_mod.n_match = 24;
             params.speculative.ngram_mod.n_min = 48;
             params.speculative.ngram_mod.n_max = 64;
+
+            // TODO: not sure if this is a good config - explore more settings and potentially enable it
+            //params.speculative.types.push_back(COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V);
+            //params.speculative.ngram_map_k4v.size_n = 8;
+            //params.speculative.ngram_map_k4v.size_m = 24;
+            //params.speculative.ngram_map_k4v.min_hits = 2;
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}));
 

common/chat-auto-parser-generator.cpp

@@ -43,11 +43,33 @@ common_chat_params peg_generator::generate_parser(const common_chat_template &
                                                   const autoparser &              autoparser) {
     // Create the result structure
     common_chat_params data;
-    data.prompt           = common_chat_template_direct_apply(tmpl, inputs);
-    data.format           = COMMON_CHAT_FORMAT_PEG_NATIVE;
-    data.preserved_tokens = autoparser.preserved_tokens;
+    data.prompt            = common_chat_template_direct_apply(tmpl, inputs);
+    data.generation_prompt = common_chat_template_generation_prompt(tmpl, inputs);
+    data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
+    data.preserved_tokens  = autoparser.preserved_tokens;
 
-    auto parser = autoparser.build_parser(inputs);
+    std::string parser_generation_prompt = data.generation_prompt;
+
+    if (inputs.continue_final_message != COMMON_CHAT_CONTINUATION_NONE && !inputs.continue_msg.empty()) {
+        // Build up generation prompt manually
+        const auto & msg = inputs.continue_msg;
+
+        if (!autoparser.reasoning.start.empty()) {
+            data.generation_prompt = data.generation_prompt.substr(0, data.generation_prompt.find(autoparser.reasoning.start));
+            data.generation_prompt += autoparser.reasoning.start + msg.reasoning_content;
+            if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
+                data.generation_prompt += autoparser.reasoning.end;
+            }
+        }
+
+        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
+            data.generation_prompt += msg.render_content();
+        }
+
+        data.prompt += data.generation_prompt;
+    }
+
+    auto parser = autoparser.build_parser(inputs, parser_generation_prompt);
     data.parser = parser.save();
 
     // Build grammar if tools are present
@@ -87,7 +109,7 @@ common_chat_params peg_generator::generate_parser(const common_chat_template &
     return data;
 }
 
-common_peg_arena autoparser::build_parser(const generation_params & inputs) const {
+common_peg_arena autoparser::build_parser(const generation_params & inputs, const std::string & generation_prompt) const {
     if (!analysis_complete) {
         throw std::invalid_argument("Cannot call build_parser on autoparser without performing analysis first, call analyze_template(...)");
     }
@@ -121,7 +143,7 @@ common_peg_arena autoparser::build_parser(const generation_params & inputs) cons
         } else {
             parser = content.build_parser(ctx);
         }
-        return pure_content ? p.prefix(inputs.generation_prompt, reasoning.start) + parser : p.prefix(inputs.generation_prompt, reasoning.start) << parser;
+        return pure_content ? p.prefix(generation_prompt, reasoning.start) + parser : p.prefix(generation_prompt, reasoning.start) << parser;
     });
 }
 

common/chat-auto-parser.h

@@ -60,16 +60,21 @@ struct generation_params {
     common_reasoning_format               reasoning_format    = COMMON_REASONING_FORMAT_AUTO;
     bool                                  stream              = true;
     std::string                           grammar;
-    bool                                  add_generation_prompt = false;
-    bool                                  enable_thinking       = true;
-    std::chrono::system_clock::time_point now                   = std::chrono::system_clock::now();
-    std::string                           generation_prompt;
+    bool                                  add_generation_prompt  = false;
+    common_chat_continuation              continue_final_message = COMMON_CHAT_CONTINUATION_NONE;
+    common_chat_msg                       continue_msg;
+    bool                                  enable_thinking        = true;
+    std::chrono::system_clock::time_point now                    = std::chrono::system_clock::now();
     json                                  extra_context;
     bool                                  add_bos       = false;
     bool                                  add_eos       = false;
     bool                                  is_inference  = true;
     bool                                  add_inference = false;
     bool                                  mark_input    = true;  // whether to mark input strings in the jinja context
+
+    bool has_continuation() const {
+        return continue_final_message != COMMON_CHAT_CONTINUATION_NONE && !continue_msg.empty();
+    }
 };
 
 // ============================================================================
@@ -386,7 +391,7 @@ struct autoparser {
     void analyze_template(const common_chat_template & tmpl);
 
     // Build the PEG parser for this template
-    common_peg_arena build_parser(const generation_params & inputs) const;
+    common_peg_arena build_parser(const generation_params & inputs, const std::string & generation_prompt) const;
 
   private:
     // Collect tokens from entire analysis to preserve

common/chat-peg-parser.cpp

@@ -358,35 +358,7 @@ void common_chat_peg_mapper::map(const common_peg_ast_node & node) {
             if (is_potential_container) {
                 value_content = normalize_container_value(value_content);
             }
-
-            // Try to parse as JSON value (number, bool, null, object, array)
-            try {
-                ordered_json parsed = ordered_json::parse(value_content);
-                if (parsed.is_string()) {
-                    // Don't add closing quote yet (added by arg_close) for monotonic streaming
-                    std::string escaped = parsed.dump();
-                    if (!escaped.empty() && escaped.back() == '"') {
-                        escaped.pop_back();
-                    }
-                    value_to_add          = escaped;
-                    closing_quote_pending = true;
-                } else {
-                    // Non-string values: use raw content to preserve whitespace for monotonicity
-                    value_to_add = value_content;
-                }
-            } catch (...) {
-                if (node.is_partial && is_potential_container) {
-                    // Partial container: pass through the already-normalized content
-                    value_to_add = value_content;
-                } else {
-                    // Not valid JSON - treat as string value
-                    if (!closing_quote_pending) {
-                        value_to_add          = "\"";
-                        closing_quote_pending = true;
-                    }
-                    value_to_add += escape_json_string_inner(value_content);
-                }
-            }
+            value_to_add += value_content;
         }
 
         args_target() += value_to_add;
@@ -813,7 +785,7 @@ common_peg_parser common_chat_peg_builder::prefix(const std::string & s, const s
     if (delimiter.empty()) {
         return literal(s);
     }
-    return literal(s.substr(0, s.rfind(delimiter)));
+    return literal(s.substr(0, s.find(delimiter)));
 }
 
 common_peg_parser common_chat_peg_builder::optspace(const std::string & tag) {

common/chat-peg-parser.h

@@ -90,7 +90,7 @@ class common_chat_peg_builder : public common_peg_parser_builder {
 
     // Use for schema-declared string types - won't be treated as potential JSON container
     common_peg_parser tool_arg_string_value(const common_peg_parser & p) { return tag(TOOL_ARG_STRING_VALUE, p); }
-    common_peg_parser tool_arg_json_value(const common_peg_parser & p) { return atomic(tag(TOOL_ARG_VALUE, p)); }
+    common_peg_parser tool_arg_json_value(const common_peg_parser & p) { return tag(TOOL_ARG_VALUE, p); }
 
 
     // Return a parser that parses the prefix of a string, up to a given delimiter.

common/chat.cpp

@@ -70,6 +70,26 @@ static bool has_content_or_tool_calls(const common_chat_msg & msg) {
     return !msg.content.empty() || !msg.tool_calls.empty();
 }
 
+std::string common_chat_msg::render_content(const std::string & delimiter) const {
+    if (!content.empty() && !content_parts.empty()) {
+        throw std::runtime_error("Cannot specify both content and content_parts");
+    }
+    if (!content.empty()) {
+        return content;
+    }
+
+    std::string text;
+    for (const auto & part : content_parts) {
+        if (part.type == "text") {
+            if (!text.empty()) {
+                text += delimiter;
+            }
+            text += part.text;
+        }
+    }
+    return text;
+}
+
 json common_chat_msg::to_json_oaicompat(bool concat_typed_text) const {
     if (!content.empty() && !content_parts.empty()) {
         throw std::runtime_error("Cannot specify both content and content_parts");
@@ -451,6 +471,22 @@ std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const json & too
     return result;
 }
 
+common_chat_continuation common_chat_continuation_parse(const nlohmann::ordered_json & value) {
+    if (value.is_boolean() && value.get<bool>()) {
+        return COMMON_CHAT_CONTINUATION_AUTO;
+    }
+    if (value.is_string()) {
+        auto value_str = value.get<std::string>();
+        if (value_str == "reasoning_content") {
+            return COMMON_CHAT_CONTINUATION_REASONING;
+        }
+        if (value_str == "content") {
+            return COMMON_CHAT_CONTINUATION_CONTENT;
+        }
+    }
+    return COMMON_CHAT_CONTINUATION_NONE;
+}
+
 bool common_chat_verify_template(const std::string & tmpl, bool use_jinja) {
     if (use_jinja) {
         try {
@@ -811,6 +847,36 @@ std::string common_chat_template_direct_apply(
     return common_chat_template_direct_apply_impl(tmpl, inputs, std::nullopt, std::nullopt, std::nullopt);
 }
 
+static std::string common_chat_template_generation_prompt_impl(
+    const common_chat_template & tmpl,
+    const autoparser::generation_params & inputs,
+    const std::optional<json> & messages_override = std::nullopt,
+    const std::optional<json> & tools_override = std::nullopt,
+    const std::optional<json> & additional_context = std::nullopt) {
+
+    auto adjusted_messages = messages_override ? *messages_override : inputs.messages;
+
+    autoparser::generation_params params = inputs;
+    params.add_generation_prompt = false;
+    params.continue_final_message = COMMON_CHAT_CONTINUATION_NONE;
+    std::string no_gen_prompt    = common_chat_template_direct_apply_impl(tmpl, params, adjusted_messages, tools_override, additional_context);
+    params.add_generation_prompt = true;
+    std::string gen_prompt       = common_chat_template_direct_apply_impl(tmpl, params, adjusted_messages, tools_override, additional_context);
+
+    size_t prefix_len = 0;
+    size_t min_size = std::min(no_gen_prompt.size(), gen_prompt.size());
+    while (prefix_len < min_size && no_gen_prompt[prefix_len] == gen_prompt[prefix_len]) {
+        prefix_len++;
+    }
+    return gen_prompt.substr(prefix_len);
+}
+
+std::string common_chat_template_generation_prompt(
+    const common_chat_template & tmpl,
+    const autoparser::generation_params & inputs) {
+    return common_chat_template_generation_prompt_impl(tmpl, inputs, std::nullopt, std::nullopt, std::nullopt);
+}
+
 static common_chat_params common_chat_params_init_ministral_3(const common_chat_template &    tmpl,
                                                               const autoparser::generation_params & inputs) {
     common_chat_params data;
@@ -863,6 +929,7 @@ static common_chat_params common_chat_params_init_ministral_3(const common_chat_
     data.thinking_start_tag = "[THINK]";
     data.thinking_end_tag   = "[/THINK]";
     data.prompt            = common_chat_template_direct_apply_impl(tmpl, inputs, /* messages_override = */ adjusted_messages);
+    data.generation_prompt = common_chat_template_generation_prompt_impl(tmpl, inputs, /* messages_override = */ adjusted_messages);
     data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
     data.preserved_tokens  = {
         "[THINK]",
@@ -871,8 +938,19 @@ static common_chat_params common_chat_params_init_ministral_3(const common_chat_
         "[ARGS]",
     };
 
+    if (inputs.has_continuation()) {
+        const auto & msg = inputs.continue_msg;
+
+        data.generation_prompt = "[THINK]" + msg.reasoning_content;
+        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
+            data.generation_prompt += "[/THINK]" + msg.render_content();
+        }
+
+        data.prompt += data.generation_prompt;
+    }
+
     auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
-        auto generation_prompt = p.prefix(inputs.generation_prompt, "[THINK]");
+        auto generation_prompt = p.eps();
         auto reasoning =
             extract_reasoning ? p.optional("[THINK]" + p.reasoning(p.until("[/THINK]")) + "[/THINK]") : p.eps();
 
@@ -963,6 +1041,7 @@ static common_chat_params common_chat_params_init_gpt_oss(const common_chat_temp
     }
 
     data.prompt            = prompt;
+    data.generation_prompt = common_chat_template_generation_prompt_impl(tmpl, inputs, /* messages_override= */ adjusted_messages);
     data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
     data.supports_thinking = true;
 
@@ -972,6 +1051,18 @@ static common_chat_params common_chat_params_init_gpt_oss(const common_chat_temp
         "<|channel|>", "<|constrain|>", "<|message|>", "<|start|>", "<|end|>",
     };
 
+    // Adjust prompt for continuation
+    if (inputs.has_continuation()) {
+        const auto & msg = inputs.continue_msg;
+
+        data.generation_prompt = "<|start|>assistant<|channel|>analysis<|message|>" + msg.reasoning_content;
+        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
+            data.generation_prompt += "<|end|><|start|>assistant<|channel|>final<|message|>" + msg.render_content();
+        }
+
+        data.prompt += data.generation_prompt;
+    }
+
     auto has_tools           = inputs.tools.is_array() && !inputs.tools.empty();
     auto has_response_format = !inputs.json_schema.is_null() && inputs.json_schema.is_object();
     auto include_grammar     = has_response_format || (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE);
@@ -1080,12 +1171,14 @@ static common_chat_params common_chat_params_init_gemma4(const common_chat_templ
     common_chat_params data;
 
     data.prompt            = common_chat_template_direct_apply_impl(tmpl, inputs);
+    data.generation_prompt = common_chat_template_generation_prompt_impl(tmpl, inputs);
 
     if (inputs.add_generation_prompt && string_ends_with(data.prompt, "<turn|>\n")) {
         // This may happen if the model generates content + tool_call, the
         // template does not add the model's next turn and confuses the model
         // from emitting its proper reasoning token sequence.
-        data.prompt += "<|turn>model\n";
+        data.generation_prompt = "<|turn>model\n";
+        data.prompt += data.generation_prompt;
     }
 
     data.format            = COMMON_CHAT_FORMAT_PEG_GEMMA4;
@@ -1101,13 +1194,25 @@ static common_chat_params common_chat_params_init_gemma4(const common_chat_templ
         "<|turn>",
     };
 
+    if (inputs.has_continuation()) {
+        const auto & msg = inputs.continue_msg;
+
+        data.generation_prompt = string_ends_with(data.prompt, "<turn|>\n") ? "<|turn>model\n" : "";
+        data.generation_prompt += "<|channel>thought\n" + msg.reasoning_content;
+        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
+            data.generation_prompt += "<channel|>" + msg.render_content();
+        }
+
+        data.prompt += data.generation_prompt;
+    }
+
     auto has_tools           = inputs.tools.is_array() && !inputs.tools.empty();
     auto has_response_format = !inputs.json_schema.is_null() && inputs.json_schema.is_object();
     auto include_grammar     = has_response_format || (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE);
     auto extract_reasoning   = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
 
     auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
-        auto start = p.rule("start", p.prefix(inputs.generation_prompt, "<|channel>"));
+        auto start = p.rule("start", p.optional(p.literal("<|turn>model\n")));
 
         if (extract_reasoning) {
             p.rule("thought", p.literal("<|channel>thought") + p.space() + p.reasoning(p.until("<channel|>")) + p.literal("<channel|>"));
@@ -1224,15 +1329,22 @@ static common_chat_params common_chat_params_init_functionary_v3_2(const common_
                                                                    const autoparser::generation_params & inputs) {
     common_chat_params data;
 
-    data.prompt           = common_chat_template_direct_apply_impl(tmpl, inputs);
-    data.format           = COMMON_CHAT_FORMAT_PEG_NATIVE;
-    data.preserved_tokens = {
+    data.prompt            = common_chat_template_direct_apply_impl(tmpl, inputs);
+    data.generation_prompt = common_chat_template_generation_prompt_impl(tmpl, inputs);
+    data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
+    data.preserved_tokens  = {
         ">>>all",
     };
 
     auto has_tools         = inputs.tools.is_array() && !inputs.tools.empty();
     auto include_grammar   = has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE;
 
+    if (inputs.has_continuation()) {
+        const auto & msg = inputs.continue_msg;
+        data.generation_prompt = "<|start_header_id|>assistant<|end_header_id|>\n\n>>>all\n" + msg.render_content();
+        data.prompt += data.generation_prompt;
+    }
+
     auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
         // Functionary v3.2 format:
         // - Normal content: >>>all\n{content}
@@ -1244,7 +1356,7 @@ static common_chat_params common_chat_params_init_functionary_v3_2(const common_
         // When no tools, content goes until end
         auto content_until_tool = p.literal("all\n") + p.content(p.until(">>>"));
         auto content_until_end  = p.literal("all\n") + p.content(p.rest());
-        auto generation_prompt  = p.literal(inputs.generation_prompt);
+        auto generation_prompt  = p.literal("<|start_header_id|>assistant<|end_header_id|>\n\n>>>");
 
         // If no tools or tool_choice is NONE, just parse content
         if (!has_tools || inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_NONE) {
@@ -1318,9 +1430,10 @@ static common_chat_params common_chat_params_init_kimi_k2(const common_chat_temp
                                                           const autoparser::generation_params & inputs) {
     common_chat_params data;
 
-    data.prompt             = common_chat_template_direct_apply_impl(tmpl, inputs);
-    data.format             = COMMON_CHAT_FORMAT_PEG_NATIVE;
-    data.supports_thinking  = true;
+    data.prompt            = common_chat_template_direct_apply_impl(tmpl, inputs);
+    data.generation_prompt = common_chat_template_generation_prompt_impl(tmpl, inputs);
+    data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
+    data.supports_thinking = true;
     data.preserved_tokens  = {
         "<|tool_calls_section_begin|>",
         "<|tool_calls_section_end|>",
@@ -1343,10 +1456,22 @@ static common_chat_params common_chat_params_init_kimi_k2(const common_chat_temp
 
     const std::string THINK_START = "<think>";
     const std::string THINK_END   = "</think>";
+    const std::string GEN_PROMPT  = "<|im_assistant|>assistant<|im_middle|>";
 
     data.thinking_start_tag = THINK_START;
     data.thinking_end_tag   = THINK_END;
 
+    if (inputs.has_continuation()) {
+        const auto & msg = inputs.continue_msg;
+
+        data.generation_prompt = GEN_PROMPT + THINK_START + msg.reasoning_content;
+        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
+            data.generation_prompt += THINK_END + msg.render_content();
+        }
+
+        data.prompt += data.generation_prompt;
+    }
+
     auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
         // Kimi K2 Thinking format:
         // - Reasoning: <think>{reasoning}</think>
@@ -1366,7 +1491,7 @@ static common_chat_params common_chat_params_init_kimi_k2(const common_chat_temp
         auto reasoning = extract_reasoning ? p.optional(THINK_START + p.reasoning(
             p.until_one_of({ THINK_END, "<|tool_calls_section_begin|>", "<|tool_call_begin|>" })) +
             p.optional(p.literal(THINK_END))) : p.eps();
-        auto generation_prompt = p.prefix(inputs.generation_prompt, THINK_START);
+        auto generation_prompt = p.literal(GEN_PROMPT);
 
 
         // Content only parser (no tools)
@@ -1442,6 +1567,7 @@ static common_chat_params common_chat_params_init_lfm2(const common_chat_templat
     common_chat_params data;
 
     data.prompt            = common_chat_template_direct_apply_impl(tmpl, inputs);
+    data.generation_prompt = common_chat_template_generation_prompt_impl(tmpl, inputs);
     data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
     data.supports_thinking = true;
     data.preserved_tokens  = {
@@ -1461,12 +1587,24 @@ static common_chat_params common_chat_params_init_lfm2(const common_chat_templat
     const std::string TOOL_CALL_END   = "<|tool_call_end|>";
     const std::string THINK_START     = "<think>";
     const std::string THINK_END       = "</think>";
+    const std::string GEN_PROMPT      = "<|im_start|>assistant\n";
 
     data.thinking_start_tag = THINK_START;
     data.thinking_end_tag   = THINK_END;
 
+    if (inputs.has_continuation()) {
+        const auto & msg = inputs.continue_msg;
+
+        data.generation_prompt = GEN_PROMPT + THINK_START + msg.reasoning_content;
+        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
+            data.generation_prompt += THINK_END + msg.render_content();
+        }
+
+        data.prompt += data.generation_prompt;
+    }
+
     auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
-        auto generation_prompt = p.prefix(inputs.generation_prompt, THINK_START);
+        auto generation_prompt = p.literal(GEN_PROMPT);
         auto end = p.end();
 
         auto reasoning = p.eps();
@@ -1521,6 +1659,7 @@ static common_chat_params common_chat_params_init_lfm2_5(const common_chat_templ
     common_chat_params data;
 
     data.prompt            = common_chat_template_direct_apply_impl(tmpl, inputs);
+    data.generation_prompt = common_chat_template_generation_prompt_impl(tmpl, inputs);
     data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
     data.supports_thinking = true;
     data.preserved_tokens  = {
@@ -1536,12 +1675,24 @@ static common_chat_params common_chat_params_init_lfm2_5(const common_chat_templ
 
     const std::string THINK_START     = "<think>";
     const std::string THINK_END       = "</think>";
+    const std::string GEN_PROMPT      = "<|im_start|>assistant\n";
 
     data.thinking_start_tag = THINK_START;
     data.thinking_end_tag   = THINK_END;
 
+    if (inputs.has_continuation()) {
+        const auto & msg = inputs.continue_msg;
+
+        data.generation_prompt = GEN_PROMPT + THINK_START + msg.reasoning_content;
+        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
+            data.generation_prompt += THINK_END + msg.render_content();
+        }
+
+        data.prompt += data.generation_prompt;
+    }
+
     auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
-        auto generation_prompt = p.prefix(inputs.generation_prompt, THINK_START);
+        auto generation_prompt = p.literal(GEN_PROMPT);
         auto end = p.end();
 
         auto reasoning = p.eps();
@@ -1592,6 +1743,7 @@ static common_chat_params common_chat_params_init_gigachat_v3(
     common_chat_params data;
 
     data.prompt            = common_chat_template_direct_apply_impl(tmpl, inputs);
+    data.generation_prompt = common_chat_template_generation_prompt_impl(tmpl, inputs);
     data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
     data.supports_thinking = false;
     data.preserved_tokens  = {
@@ -1599,6 +1751,12 @@ static common_chat_params common_chat_params_init_gigachat_v3(
         "<|role_sep|>\n",
     };
 
+    if (inputs.has_continuation()) {
+        const auto & msg = inputs.continue_msg;
+        data.generation_prompt = "assistant<|role_sep|>\n" + msg.render_content();
+        data.prompt += data.generation_prompt;
+    }
+
     auto has_tools         = inputs.tools.is_array() && !inputs.tools.empty();
     auto include_grammar   = has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE;
     const auto *tool_call_start_prefix = "<|message_sep|>\n\nfunction call<|role_sep|>\n";
@@ -1634,7 +1792,7 @@ static common_chat_params common_chat_params_init_gigachat_v3(
             ret = p.content(p.rest());
         }
 
-        return p.literal(inputs.generation_prompt) + ret;
+        return p.literal("assistant<|role_sep|>\n") + ret;
     });
 
     data.parser = parser.save();
@@ -1662,12 +1820,13 @@ static common_chat_params common_chat_params_init_deepseek_v3_2(const common_cha
                                                                  const autoparser::generation_params & inputs) {
     common_chat_params data;
 
-    data.prompt            = common_chat_template_direct_apply_impl(tmpl, inputs);
-    data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
-    data.supports_thinking = true;
+    data.prompt             = common_chat_template_direct_apply_impl(tmpl, inputs);
+    data.generation_prompt  = common_chat_template_generation_prompt_impl(tmpl, inputs);
+    data.format             = COMMON_CHAT_FORMAT_PEG_NATIVE;
+    data.supports_thinking  = true;
     data.thinking_start_tag = "<think>";
     data.thinking_end_tag   = "</think>";
-    data.preserved_tokens  = {
+    data.preserved_tokens   = {
         "｜DSML｜",
         "<think>",
         "</think>",
@@ -1687,9 +1846,21 @@ static common_chat_params common_chat_params_init_deepseek_v3_2(const common_cha
     const std::string INVOKE_END   = "</" + DSML + "invoke>";
     const std::string PARAM_START  = "<" + DSML + "parameter";
     const std::string PARAM_END    = "</" + DSML + "parameter>";
+    const std::string GEN_PROMPT   = "<｜Assistant｜>";
+
+    if (inputs.has_continuation()) {
+        const auto & msg = inputs.continue_msg;
+
+        data.generation_prompt = GEN_PROMPT + THINK_START + msg.reasoning_content;
+        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
+            data.generation_prompt += THINK_END + msg.render_content();
+        }
+
+        data.prompt += data.generation_prompt;
+    }
 
     auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
-        auto generation_prompt = p.prefix(inputs.generation_prompt, THINK_START);
+        auto generation_prompt = p.literal(GEN_PROMPT);
         auto end = p.end();
 
         auto reasoning = p.eps();
@@ -2116,21 +2287,6 @@ std::optional<common_chat_params> common_chat_try_specialized_template(
     return std::nullopt;
 }
 
-static std::string common_chat_templates_generation_prompt(const common_chat_template & tmpl, const autoparser::generation_params & inputs) {
-    autoparser::generation_params params = inputs;
-    params.add_generation_prompt = false;
-    std::string no_gen_prompt    = common_chat_template_direct_apply_impl(tmpl, params);
-    params.add_generation_prompt = true;
-    std::string gen_prompt       = common_chat_template_direct_apply_impl(tmpl, params);
-
-    size_t prefix_len = 0;
-    size_t min_size = std::min(no_gen_prompt.size(), gen_prompt.size());
-    while (prefix_len < min_size && no_gen_prompt[prefix_len] == gen_prompt[prefix_len]) {
-        prefix_len++;
-    }
-    return gen_prompt.substr(prefix_len);
-}
-
 static common_chat_params common_chat_templates_apply_jinja(const struct common_chat_templates *        tmpls,
                                                             const struct common_chat_templates_inputs & inputs) {
     autoparser::generation_params params;
@@ -2149,6 +2305,27 @@ static common_chat_params common_chat_templates_apply_jinja(const struct common_
     params.add_bos               = tmpls->add_bos;
     params.add_eos               = tmpls->add_eos;
 
+    params.continue_final_message = inputs.continue_final_message;
+    if (params.continue_final_message != COMMON_CHAT_CONTINUATION_NONE) {
+        params.add_generation_prompt = false;
+
+        if (!inputs.messages.empty()) {
+            // Render messages[:-1] and store continuation message separately
+            params.continue_msg = inputs.messages.back();
+            params.messages.erase(params.messages.size() - 1);
+        }
+
+        if (params.continue_final_message == COMMON_CHAT_CONTINUATION_AUTO && !inputs.messages.empty()) {
+            // Resolve based on message content
+            params.continue_final_message = COMMON_CHAT_CONTINUATION_CONTENT;
+            if (!params.continue_msg.reasoning_content.empty() &&
+                params.continue_msg.content.empty() &&
+                params.continue_msg.content_parts.empty()) {
+                params.continue_final_message = COMMON_CHAT_CONTINUATION_REASONING;
+            }
+        }
+    }
+
     if (src.find("<|channel|>") == std::string::npos) {
         // map developer to system for all models except for GPT-OSS
         workaround::map_developer_role_to_system(params.messages);
@@ -2169,8 +2346,6 @@ static common_chat_params common_chat_templates_apply_jinja(const struct common_
         workaround::func_args_not_string(params.messages);
     }
 
-    params.generation_prompt = common_chat_templates_generation_prompt(tmpl, params);
-
     params.extra_context = common_chat_extra_context();
     for (auto el : inputs.chat_template_kwargs) {
         params.extra_context[el.first] = json::parse(el.second);
@@ -2200,17 +2375,16 @@ static common_chat_params common_chat_templates_apply_jinja(const struct common_
         auto params_copy               = params;
         params_copy.reasoning_format   = COMMON_REASONING_FORMAT_NONE;
         data.prompt                    = common_chat_template_direct_apply_impl(tmpl, params_copy);
+        data.generation_prompt         = common_chat_template_generation_prompt_impl(tmpl, params);
         data.format                    = COMMON_CHAT_FORMAT_PEG_NATIVE;
-        data.generation_prompt         = params.generation_prompt;
-        auto parser                    = build_chat_peg_parser([&params](common_chat_peg_builder &p) {
-            return p.prefix(params.generation_prompt) << p.content(p.rest());
+        auto parser                    = build_chat_peg_parser([&data](common_chat_peg_builder &p) {
+            return p.literal(data.generation_prompt) << p.content(p.rest());
         });
         data.parser                    = parser.save();
         return data;
     }
 
     if (auto result = common_chat_try_specialized_template(tmpl, src, params)) {
-        result->generation_prompt = params.generation_prompt;
         return *result;
     }
 
@@ -2224,7 +2398,6 @@ static common_chat_params common_chat_templates_apply_jinja(const struct common_
             auto_params.thinking_start_tag = trim_whitespace(autoparser.reasoning.start);
             auto_params.thinking_end_tag   = trim_whitespace(autoparser.reasoning.end);
         }
-        auto_params.generation_prompt = params.generation_prompt;
         common_peg_arena arena;
         arena.load(auto_params.parser);
         LOG_DBG("%s: generated parser:\n%s\n\nparser generation prompt: %s\n", __func__, arena.dump(arena.root()).c_str(), auto_params.generation_prompt.c_str());

common/chat.h

@@ -89,6 +89,8 @@ struct common_chat_msg {
 
     nlohmann::ordered_json to_json_oaicompat(bool concat_typed_text = false) const;
 
+    std::string render_content(const std::string & delimiter = "\n\n") const;
+
     bool empty() const {
         return content.empty() && content_parts.empty() && tool_calls.empty() && reasoning_content.empty() &&
                tool_name.empty() && tool_call_id.empty();
@@ -164,12 +166,22 @@ enum common_chat_format {
     COMMON_CHAT_FORMAT_COUNT,  // Not a format, just the # formats
 };
 
+
+// Continuation method provided via `continue_final_message`
+enum common_chat_continuation {
+    COMMON_CHAT_CONTINUATION_NONE,
+    COMMON_CHAT_CONTINUATION_AUTO,
+    COMMON_CHAT_CONTINUATION_REASONING,
+    COMMON_CHAT_CONTINUATION_CONTENT,
+};
+
 struct common_chat_templates_inputs {
     std::vector<common_chat_msg>          messages;
     std::string                           grammar;
     std::string                           json_schema;
-    bool                                  add_generation_prompt = true;
-    bool                                  use_jinja             = true;
+    bool                                  add_generation_prompt  = true;
+    common_chat_continuation              continue_final_message = COMMON_CHAT_CONTINUATION_NONE;
+    bool                                  use_jinja              = true;
     // Parameters below only supported when use_jinja is true
     std::vector<common_chat_tool>         tools;
     common_chat_tool_choice               tool_choice         = COMMON_CHAT_TOOL_CHOICE_AUTO;
@@ -207,6 +219,8 @@ struct common_chat_parser_params {
     bool                    reasoning_in_content = false;
     std::string             generation_prompt;
     bool                    parse_tool_calls     = true;
+    bool                    is_continuation      = false;
+    bool                    echo                 = false;  // Include assistant prefilled msg in output
     bool                    debug                = false;  // Enable debug output for PEG parser
     common_peg_arena        parser               = {};
     common_chat_parser_params() = default;
@@ -267,6 +281,8 @@ std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const nlohmann::or
 
 std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const nlohmann::ordered_json & tools);
 
+common_chat_continuation common_chat_continuation_parse(const nlohmann::ordered_json & value);
+
 // DEPRECATED: only used in tests
 nlohmann::ordered_json common_chat_msgs_to_json_oaicompat(const std::vector<common_chat_msg> & msgs, bool concat_typed_text = false);
 
@@ -279,6 +295,10 @@ std::string common_chat_template_direct_apply(
     const common_chat_template & tmpl,
     const autoparser::generation_params & inputs);
 
+std::string common_chat_template_generation_prompt(
+    const common_chat_template &          tmpl,
+    const autoparser::generation_params & inputs);
+
 std::optional<common_chat_params> common_chat_try_specialized_template(
         const common_chat_template &          tmpl,
         const std::string &                   src,

common/common.cpp

@@ -1160,7 +1160,7 @@ struct common_init_result::impl {
     std::vector<llama_sampler_seq_config> samplers_seq_config;
 };
 
-common_init_result::common_init_result(common_params & params) :
+common_init_result::common_init_result(common_params & params, bool model_only) :
     pimpl(new impl{}) {
     auto mparams = common_model_params_to_llama(params);
     auto cparams = common_context_params_to_llama(params);
@@ -1173,7 +1173,7 @@ common_init_result::common_init_result(common_params & params) :
             params.tensor_buft_overrides.data(),
             params.fit_params_target.data(),
             params.fit_params_min_ctx,
-            params.verbosity >= 4 ? GGML_LOG_LEVEL_DEBUG : GGML_LOG_LEVEL_ERROR);
+            params.verbosity >= LOG_LEVEL_DEBUG ? GGML_LOG_LEVEL_DEBUG : GGML_LOG_LEVEL_ERROR);
     }
 
     llama_model * model = llama_model_load_from_file(params.model.path.c_str(), mparams);
@@ -1183,6 +1183,10 @@ common_init_result::common_init_result(common_params & params) :
 
     pimpl->model.reset(model);
 
+    if (model_only) {
+        return;
+    }
+
     const llama_vocab * vocab = llama_model_get_vocab(model);
 
     // load and optionally apply lora adapters
@@ -1252,29 +1256,6 @@ common_init_result::common_init_result(common_params & params) :
         cparams.n_samplers = pimpl->samplers_seq_config.size();
     }
 
-    // [TAG_RS_STATE_ROLLBACK_SUPPORT]
-    // TODO: ngram speculative methods require checkpointing in addition to partial RS rollback
-    //       currently this is not supported. so we disable the partial rollback
-    if (cparams.n_rs_seq > 0 && (llama_model_is_recurrent(model) || llama_model_is_hybrid(model))) {
-        auto & types = params.speculative.types;
-
-        for (int i = 0; i < (int) types.size(); i++) {
-            if (types[i] == COMMON_SPECULATIVE_TYPE_NONE) {
-                continue;
-            }
-            if (types[i] == COMMON_SPECULATIVE_TYPE_DRAFT_MTP) {
-                continue;
-            }
-
-            cparams.n_rs_seq = 0;
-
-            LOG_WRN("%s: recurrent state rollback is not compatible with '%s' - disabling rollback support\n", __func__,
-                    common_speculative_type_to_str(types[i]).c_str());
-
-            break;
-        }
-    }
-
     llama_context * lctx = llama_init_from_model(model, cparams);
     if (lctx == NULL) {
         LOG_ERR("%s: failed to create context with model '%s'\n", __func__, params.model.path.c_str());
@@ -1309,8 +1290,8 @@ std::vector<llama_adapter_lora_ptr> & common_init_result::lora() {
     return pimpl->lora;
 }
 
-common_init_result_ptr common_init_from_params(common_params & params) {
-    common_init_result_ptr res(new common_init_result(params));
+common_init_result_ptr common_init_from_params(common_params & params, bool model_only) {
+    common_init_result_ptr res(new common_init_result(params, model_only));
 
     llama_model * model = res->model();
     if (model == NULL) {
@@ -1318,6 +1299,10 @@ common_init_result_ptr common_init_from_params(common_params & params) {
         return res;
     }
 
+    if (model_only) {
+        return res;
+    }
+
     llama_context * lctx = res->context();
     if (lctx == NULL) {
         LOG_ERR("%s: failed to create context with model '%s'\n", __func__, params.model.path.c_str());
@@ -1381,7 +1366,7 @@ common_init_result_ptr common_init_from_params(common_params & params) {
     }
 
     if (params.warmup) {
-        LOG_WRN("%s: warming up the model with an empty run - please wait ... (--no-warmup to disable)\n", __func__);
+        LOG_INF("%s: warming up the model with an empty run - please wait ... (--no-warmup to disable)\n", __func__);
 
         llama_set_warmup(lctx, true);
 

common/common.h

@@ -299,11 +299,13 @@ struct common_params_model {
 
 // draft-model-based speculative decoding parameters
 struct common_params_speculative_draft {
-    int32_t n_max = 16; // maximum number of tokens to draft during speculative decoding
-    int32_t n_min = 0;  // minimum number of draft tokens to use for speculative decoding
+    int32_t n_max = 3; // maximum number of tokens to draft during speculative decoding
+    int32_t n_min = 0; // minimum number of draft tokens to use for speculative decoding
 
-    float p_split = 0.1f;  // speculative decoding split probability
-    float p_min   = 0.75f; // minimum speculative decoding probability (greedy) // TODO: change default to 0.0f
+    float p_split = 0.1f; // speculative decoding split probability
+    float p_min   = 0.0f; // minimum speculative decoding probability (greedy)
+
+    bool backend_sampling = true; // offload draft sampling to the backend (default: on)
 
     common_params_model mparams;
 
@@ -615,13 +617,7 @@ struct common_params {
     std::map<std::string, std::string> default_template_kwargs;
 
     // UI configs
-#ifdef LLAMA_UI_DEFAULT_ENABLED
-    bool ui = LLAMA_UI_DEFAULT_ENABLED != 0;
-#elif defined(LLAMA_WEBUI_DEFAULT_ENABLED)
-    bool ui = LLAMA_WEBUI_DEFAULT_ENABLED != 0;
-#else
-    bool ui = true; // default to enabled when not set
-#endif
+    bool ui = true;
 
     // Deprecated: use ui, ui_mcp_proxy, ui_config_json instead
     bool webui = ui;
@@ -859,7 +855,7 @@ struct common_sampler;
 
 // note: defines the model, context, samplers, ets. lifetimes
 struct common_init_result {
-    common_init_result(common_params & params);
+    common_init_result(common_params & params, bool model_only = false);
     ~common_init_result();
 
     llama_model * model();
@@ -877,7 +873,7 @@ private:
 
 using common_init_result_ptr = std::unique_ptr<common_init_result>;
 
-common_init_result_ptr common_init_from_params(common_params & params);
+common_init_result_ptr common_init_from_params(common_params & params, bool model_only = false);
 
 struct llama_model_params     common_model_params_to_llama  (      common_params & params);
 struct llama_context_params   common_context_params_to_llama(const common_params & params);

common/hf-cache.cpp

@@ -11,7 +11,6 @@
 #include <filesystem>
 #include <fstream>
 #include <atomic>
-#include <regex> // migration only
 #include <string>
 #include <string_view>
 #include <stdexcept>
@@ -336,15 +335,9 @@ hf_files get_repo_files(const std::string & repo_id,
                 if (item["lfs"].contains("oid") && item["lfs"]["oid"].is_string()) {
                     file.oid = item["lfs"]["oid"].get<std::string>();
                 }
-                if (item["lfs"].contains("size") && item["lfs"]["size"].is_number()) {
-                    file.size = item["lfs"]["size"].get<size_t>();
-                }
             } else if (item.contains("oid") && item["oid"].is_string()) {
                 file.oid = item["oid"].get<std::string>();
             }
-            if (file.size == 0 && item.contains("size") && item["size"].is_number()) {
-                file.size = item["size"].get<size_t>();
-            }
 
             if (!file.oid.empty() && !is_valid_oid(file.oid)) {
                 LOG_WRN("%s: skip invalid oid: %s\n", __func__, file.oid.c_str());
@@ -502,271 +495,4 @@ std::string finalize_file(const hf_file & file) {
     return file.final_path;
 }
 
-// delete everything after this line, one day
-
-// copied from download.cpp without the tag part
-struct gguf_split_info {
-    std::string prefix; // tag included
-    int index;
-    int count;
-};
-
-static gguf_split_info get_gguf_split_info(const std::string & path) {
-    static const std::regex re_split("^(.+)-([0-9]{5})-of-([0-9]{5})$", std::regex::icase);
-    std::smatch m;
-
-    std::string prefix = path;
-    if (!string_remove_suffix(prefix, ".gguf")) {
-        return {};
-    }
-
-    int index = 1;
-    int count = 1;
-
-    if (std::regex_match(prefix, m, re_split)) {
-        index = std::stoi(m[2].str());
-        count = std::stoi(m[3].str());
-        prefix = m[1].str();
-    }
-
-    return {std::move(prefix), index, count};
-}
-
-static std::pair<std::string, std::string> parse_manifest_name(std::string & filename) {
-    static const std::regex re(R"(^manifest=([^=]+)=([^=]+)=.*\.json$)");
-    std::smatch match;
-    if (std::regex_match(filename, match, re)) {
-        return {match[1].str(), match[2].str()};
-    }
-    return {};
-}
-
-static std::string make_old_cache_filename(const std::string & owner,
-                                           const std::string & repo,
-                                           const std::string & filename) {
-    auto result = owner + "_" + repo + "_" + filename;
-    string_replace_all(result, "/", "_");
-    return result;
-}
-
-struct migrate_file {
-    std::string path;
-    std::string sha256;
-    size_t size;
-    fs::path old_path;
-    fs::path etag_path;
-    const hf_file * file;
-};
-
-using migrate_files = std::vector<migrate_file>;
-
-static bool collect_file(const fs::path    & old_cache,
-                         const std::string & owner,
-                         const std::string & repo,
-                         const std::string & path,
-                         const std::string & sha256,
-                         const hf_files    & files,
-                         migrate_files     & to_migrate) {
-
-    const hf_file * file = nullptr;
-
-    for (const auto & f : files) {
-        if (f.path == path) {
-            file = &f;
-            break;
-        }
-    }
-
-    std::string old_filename = make_old_cache_filename(owner, repo, path);
-    fs::path old_path = old_cache / old_filename;
-    fs::path etag_path = old_path.string() + ".etag";
-
-    if (!fs::exists(old_path)) {
-        if (file && fs::exists(file->final_path)) {
-            return true;
-        }
-        LOG_WRN("%s: %s not found in old cache or HF cache\n", __func__, old_filename.c_str());
-        return false;
-    }
-
-    if (!file) {
-        LOG_WRN("%s: %s not found in current repo\n", __func__, old_filename.c_str());
-        return false;
-    }
-
-    if (!sha256.empty() && !file->oid.empty() && sha256 != file->oid) {
-        LOG_WRN("%s: %s is not up to date (sha256 mismatch)\n", __func__, old_filename.c_str());
-        return false;
-    }
-
-    if (file->size > 0) {
-        size_t size = fs::file_size(old_path);
-        if (size != file->size) {
-            LOG_WRN("%s: %s has wrong size %zu (expected %zu)\n", __func__, old_filename.c_str(), size, file->size);
-            return false;
-        }
-    }
-
-    to_migrate.push_back({path, sha256, file->size, old_path, etag_path, file});
-    return true;
-}
-
-static bool collect_files(const fs::path    & old_cache,
-                          const std::string & owner,
-                          const std::string & repo,
-                          const nl::json    & node,
-                          const hf_files    & files,
-                          migrate_files     & to_migrate) {
-
-    if (!node.contains("rfilename") ||
-        !node.contains("lfs")       ||
-        !node["lfs"].contains("sha256")) {
-        return true;
-    }
-
-    std::string path = node["rfilename"];
-    std::string sha256 = node["lfs"]["sha256"];
-
-    auto split = get_gguf_split_info(path);
-
-    if (split.count <= 1) {
-        return collect_file(old_cache, owner, repo, path, sha256, files, to_migrate);
-    }
-
-    std::vector<std::pair<std::string, std::string>> splits;
-
-    for (const auto & f : files) {
-        auto split_f = get_gguf_split_info(f.path);
-        if (split_f.count == split.count && split_f.prefix == split.prefix) {
-            // sadly the manifest only provides the sha256 of the first file (index == 1)
-            // the rest will be verified using the size...
-            std::string f_sha256 = (split_f.index == 1) ? sha256 : "";
-            splits.emplace_back(f.path, f_sha256);
-        }
-    }
-
-    if ((int)splits.size() != split.count) {
-        LOG_WRN("%s: expected %d split files but found %d in repo\n", __func__, split.count, (int)splits.size());
-        return false;
-    }
-
-    for (const auto & [f_path, f_sha256] : splits) {
-        if (!collect_file(old_cache, owner, repo, f_path, f_sha256, files, to_migrate)) {
-            return false;
-        }
-    }
-
-    return true;
-}
-
-static bool migrate_file(const migrate_file & file) {
-    std::error_code ec;
-
-    fs::path new_path(file.file->local_path);
-    fs::create_directories(new_path.parent_path(), ec);
-
-    if (!fs::exists(new_path, ec)) {
-        fs::rename(file.old_path, new_path, ec);
-        if (ec) {
-            fs::copy_file(file.old_path, new_path, ec);
-            if (ec) {
-                LOG_ERR("%s: failed to move/copy %s: %s\n", __func__, file.old_path.string().c_str(), ec.message().c_str());
-                return false;
-            }
-        }
-        fs::remove(file.old_path, ec);
-    }
-    fs::remove(file.etag_path, ec);
-
-    std::string filename = finalize_file(*file.file);
-    LOG_INF("%s: migrated %s -> %s\n", __func__, file.old_path.filename().string().c_str(), filename.c_str());
-    return true;
-}
-
-void migrate_old_cache_to_hf_cache(const std::string & token, bool offline) {
-    fs::path old_cache = fs_get_cache_directory();
-    if (!fs::exists(old_cache)) {
-        return;
-    }
-
-    if (offline) {
-        LOG_WRN("%s: skipping migration in offline mode (will run when online)\n", __func__);
-        return; // -hf is not going to work
-    }
-
-    bool warned = false;
-
-    for (const auto & entry : fs::directory_iterator(old_cache)) {
-        if (!entry.is_regular_file()) {
-            continue;
-        }
-        auto filename = entry.path().filename().string();
-        auto [owner, repo] = parse_manifest_name(filename);
-
-        if (owner.empty() || repo.empty()) {
-            continue;
-        }
-
-        if (!warned) {
-            warned = true;
-            LOG_WRN("================================================================================\n"
-                    "WARNING: Migrating cache to HuggingFace cache directory\n"
-                    "  Old cache: %s\n"
-                    "  New cache: %s\n"
-                    "This one-time migration moves models previously downloaded with -hf\n"
-                    "from the legacy llama.cpp cache to the standard HuggingFace cache.\n"
-                    "Models downloaded with --model-url are not affected.\n"
-                    "================================================================================\n",
-                    old_cache.string().c_str(), get_cache_directory().string().c_str());
-        }
-
-        auto repo_id = owner + "/" + repo;
-        auto files = get_repo_files(repo_id, token);
-
-        if (files.empty()) {
-            LOG_WRN("%s: could not get repo files for %s, skipping\n", __func__, repo_id.c_str());
-            continue;
-        }
-
-        migrate_files to_migrate;
-        bool ok = true;
-
-        try {
-            std::ifstream manifest(entry.path());
-            auto json = nl::json::parse(manifest);
-            for (const char * key : {"ggufFile", "mmprojFile"}) {
-                if (json.contains(key)) {
-                    if (!collect_files(old_cache, owner, repo, json[key], files, to_migrate)) {
-                        ok = false;
-                        break;
-                    }
-                }
-            }
-        } catch (const std::exception & e) {
-            LOG_WRN("%s: failed to parse manifest %s: %s\n", __func__, filename.c_str(), e.what());
-            continue;
-        }
-
-        if (!ok) {
-            LOG_WRN("%s: migration skipped: one or more files failed validation\n", __func__);
-            continue;
-        }
-
-        for (const auto & file : to_migrate) {
-            if (!migrate_file(file)) {
-                ok = false;
-                break;
-            }
-        }
-
-        if (!ok) {
-            LOG_WRN("%s: migration failed: could not migrate all files\n", __func__);
-            continue;
-        }
-
-        LOG_INF("%s: migration complete, deleting manifest: %s\n", __func__, entry.path().string().c_str());
-        fs::remove(entry.path());
-    }
-}
-
 } // namespace hf_cache

common/hf-cache.h

@@ -14,7 +14,6 @@ struct hf_file {
     std::string final_path;
     std::string oid;
     std::string repo_id;
-    size_t size = 0; // only for the migration
 };
 
 using hf_files = std::vector<hf_file>;
@@ -30,7 +29,4 @@ hf_files get_cached_files(const std::string & repo_id = {});
 // Create snapshot path (link or move/copy) and return it
 std::string finalize_file(const hf_file & file);
 
-// TODO: Remove later
-void migrate_old_cache_to_hf_cache(const std::string & token, bool offline = false);
-
 } // namespace hf_cache

common/ngram-map.cpp

@@ -471,7 +471,7 @@ void common_ngram_map_draft(common_ngram_map & map,
         sum_occur += curr_occur;
     }
 
-    LOG_INF("%s: key_offset = %zu, max_occur = %d, sum_occur = %d, slot_max = %d [%zu/%d, %zu/%d, %zu/%d, %zu/%d]\n", __func__,
+    LOG_DBG("%s: key_offset = %zu, max_occur = %d, sum_occur = %d, slot_max = %d [%zu/%d, %zu/%d, %zu/%d, %zu/%d]\n", __func__,
             key_offset,
             max_occur, sum_occur, slot_max,
             curr_key.values[0].value_idx, curr_key.values[0].value_num,
@@ -482,7 +482,7 @@ void common_ngram_map_draft(common_ngram_map & map,
     // Print the tokens of the four values (if idx != 0), use LOG_INF
     for (int v = 0; v < COMMON_NGRAM_MAX_VALUES; ++v) {
         if (curr_key.values[v].value_idx != 0) {
-            LOG_INF("%s: value[%d] = %s\n", __func__, v, common_tokens_to_str(inp, curr_key.values[v].value_idx, m).c_str());
+            LOG_DBG("%s: value[%d] = %s\n", __func__, v, common_tokens_to_str(inp, curr_key.values[v].value_idx, m).c_str());
         }
     }
 
@@ -500,7 +500,7 @@ void common_ngram_map_draft(common_ngram_map & map,
         draft.push_back(inp[match_pos + n + i]);
     }
 
-    LOG_INF("%s: key_offset = %zu, slot_max = %d, key_num = %d, draft.size = %zu\n", __func__,
+    LOG_DBG("%s: key_offset = %zu, slot_max = %d, key_num = %d, draft.size = %zu\n", __func__,
             key_offset, slot_max,
             curr_key.key_num, draft.size());
 

common/speculative.cpp

@@ -32,6 +32,18 @@ const std::map<std::string, common_speculative_type> common_speculative_type_fro
     {"ngram-cache",   COMMON_SPECULATIVE_TYPE_NGRAM_CACHE}
 };
 
+static std::string common_speculative_get_devices_str(const std::vector<ggml_backend_dev_t> & devices) {
+    std::string result;
+    for (size_t i = 0; i < devices.size(); i++) {
+        if (devices[i] == nullptr) {
+            continue;
+        }
+        if (!result.empty()) result += ", ";
+        result += ggml_backend_dev_name(devices[i]);
+    }
+    return result.empty() ? "default" : result;
+}
+
 struct common_speculative_config {
     common_speculative_type type;
     common_params_speculative params;
@@ -144,10 +156,13 @@ struct common_speculative_impl {
 
     virtual void draft(common_speculative_draft_params_vec & dparams) = 0;
 
-    virtual void accept(llama_seq_id seq_id, uint16_t n_accepted) = 0;
+    virtual void accept(llama_seq_id seq_id, uint16_t n_accepted, bool is_other) = 0;
 
-    // true if this implementation requires the target context to extract embeddings
+    // true if this implementation requires the target context to extract post-norm embeddings
     virtual bool need_embd() const = 0;
+
+    // true if this implementation requires the target context to extract pre-norm embeddings
+    virtual bool need_embd_pre_norm() const { return false; }
 };
 
 struct common_speculative_impl_draft_simple : public common_speculative_impl {
@@ -164,6 +179,16 @@ struct common_speculative_impl_draft_simple : public common_speculative_impl {
         auto * ctx_dft = this->params.ctx_dft;
         auto * ctx_tgt = this->params.ctx_tgt;
 
+        LOG_INF("%s: adding speculative implementation 'draft-simple'\n", __func__);
+        LOG_INF("%s: - n_max=%d, n_min=%d, p_min=%f\n", __func__, this->params.n_max, this->params.n_min, this->params.p_min);
+        LOG_INF("%s: - gpu_layers=%d, cache_k=%s, cache_v=%s, ctx_tgt=%s, ctx_dft=%s, devices=[%s]\n", __func__,
+                this->params.n_gpu_layers,
+                ggml_type_name(this->params.cache_type_k),
+                ggml_type_name(this->params.cache_type_v),
+                ctx_tgt ? "yes" : "no",
+                ctx_dft ? "yes" : "no",
+                common_speculative_get_devices_str(this->params.devices).c_str());
+
         batch = llama_batch_init(llama_n_batch(ctx_dft), 0, 1);
 
         // TODO: optimize or pass from outside?
@@ -340,7 +365,7 @@ struct common_speculative_impl_draft_simple : public common_speculative_impl {
         }
     }
 
-    void accept(llama_seq_id /*seq_id*/, uint16_t /*n_accepted*/) override {
+    void accept(llama_seq_id /*seq_id*/, uint16_t /*n_accepted*/, bool /*is_other*/) override {
         // noop
     }
 
@@ -352,8 +377,12 @@ struct common_speculative_impl_draft_simple : public common_speculative_impl {
 struct common_speculative_impl_draft_eagle3 : public common_speculative_impl {
     //common_params_speculative_eagle3 params;
 
-    common_speculative_impl_draft_eagle3(const common_params_speculative & /*params*/, uint32_t n_seq)
-        : common_speculative_impl(COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3, n_seq) {}
+    common_speculative_impl_draft_eagle3(const common_params_speculative & params, uint32_t n_seq)
+        : common_speculative_impl(COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3, n_seq)
+    {
+        LOG_INF("%s: adding speculative implementation 'draft-eagle3'\n", __func__);
+        LOG_INF("%s: - n_max=%d, n_min=%d, p_min=%f\n", __func__, params.draft.n_max, params.draft.n_min, params.draft.p_min);
+    }
 
     void begin(llama_seq_id /*seq_id*/, const llama_tokens & /*prompt*/) override {
         // noop
@@ -368,7 +397,7 @@ struct common_speculative_impl_draft_eagle3 : public common_speculative_impl {
         // TODO: implement
     }
 
-    void accept(llama_seq_id /*seq_id*/, uint16_t /*n_accepted*/) override {
+    void accept(llama_seq_id /*seq_id*/, uint16_t /*n_accepted*/, bool /*is_other*/) override {
         // noop
     }
 
@@ -377,13 +406,16 @@ struct common_speculative_impl_draft_eagle3 : public common_speculative_impl {
     }
 };
 
-struct common_speculative_state_draft_mtp : public common_speculative_impl {
+struct common_speculative_impl_draft_mtp : public common_speculative_impl {
     common_params_speculative_draft params; // reuses the draft-model params slot (ctx_tgt/ctx_dft)
 
     llama_batch batch;
 
     std::vector<common_sampler_ptr> smpls;
 
+    // backend sampler chain per seq, attached to ctx_dft
+    std::vector<llama_sampler *> backend_chains;
+
     int32_t n_embd = 0;
 
     // Per-sequence cross-batch carryover: pair (h_p, x_{p+1}) at MTP pos p+1.
@@ -404,7 +436,7 @@ struct common_speculative_state_draft_mtp : public common_speculative_impl {
     // pre-advancement before process() mirrored the verify batch.
     std::vector<uint16_t> last_n_drafted;
 
-    common_speculative_state_draft_mtp(const common_params_speculative & params, uint32_t n_seq)
+    common_speculative_impl_draft_mtp(const common_params_speculative & params, uint32_t n_seq)
         : common_speculative_impl(COMMON_SPECULATIVE_TYPE_DRAFT_MTP, n_seq)
         , params(params.draft)
     {
@@ -414,6 +446,16 @@ struct common_speculative_state_draft_mtp : public common_speculative_impl {
 
         n_embd = llama_model_n_embd(llama_get_model(ctx_dft));
 
+        LOG_INF("%s: adding speculative implementation 'draft-mtp'\n", __func__);
+        LOG_INF("%s: - n_max=%d, n_min=%d, p_min=%.2f, n_embd=%d, backend_sampling=%d\n", __func__, this->params.n_max, this->params.n_min, this->params.p_min, n_embd, (int) this->params.backend_sampling);
+        LOG_INF("%s: - gpu_layers=%d, cache_k=%s, cache_v=%s, ctx_tgt=%s, ctx_dft=%s, devices=[%s]\n", __func__,
+                this->params.n_gpu_layers,
+                ggml_type_name(this->params.cache_type_k),
+                ggml_type_name(this->params.cache_type_v),
+                ctx_tgt ? "yes" : "no",
+                ctx_dft ? "yes" : "no",
+                common_speculative_get_devices_str(this->params.devices).c_str());
+
         const int32_t n_b = (int32_t) llama_n_batch(ctx_dft);
         batch = llama_batch_init(/*n_tokens=*/ n_b, /*embd=*/ n_embd, /*n_seq_max=*/ 1);
         // llama_batch_init allocates only one of token/embd; MTP needs both.
@@ -424,13 +466,29 @@ struct common_speculative_state_draft_mtp : public common_speculative_impl {
         for (auto & s : smpls) {
             common_params_sampling sparams;
             sparams.no_perf  = false;
-            sparams.top_k    = 1; // TODO: re-enable top_k == 10 and utilize `p_min` spec param
+            sparams.top_k    = 10;
             sparams.samplers = { COMMON_SAMPLER_TYPE_TOP_K };
             s.reset(common_sampler_init(llama_get_model(ctx_dft), sparams));
         }
 
-        llama_set_embeddings_pre_norm(ctx_tgt, true);
-        llama_set_embeddings_pre_norm(ctx_dft, true);
+        // offload draft sampling to the backend
+        backend_chains.assign(n_seq, nullptr);
+        if (this->params.backend_sampling) {
+            for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) n_seq; ++seq_id) {
+                llama_sampler * chain = llama_sampler_chain_init(llama_sampler_chain_default_params());
+                llama_sampler_chain_add(chain, llama_sampler_init_top_k(10));
+
+                if (!llama_set_sampler(ctx_dft, seq_id, chain)) {
+                    LOG_WRN("%s: backend offload failed for seq_id=%d; using CPU sampler\n", __func__, (int) seq_id);
+                    llama_sampler_free(chain);
+                    chain = nullptr;
+                }
+                backend_chains[seq_id] = chain;
+            }
+        }
+
+        llama_set_embeddings_pre_norm(ctx_tgt, true, /*masked*/ false);
+        llama_set_embeddings_pre_norm(ctx_dft, true, /*masked*/ true);
 
         pending_h.assign(n_seq, std::vector<float>(n_embd, 0.0f));
 
@@ -443,7 +501,19 @@ struct common_speculative_state_draft_mtp : public common_speculative_impl {
         last_n_drafted.assign(n_seq, 0);
     }
 
-    ~common_speculative_state_draft_mtp() override {
+    ~common_speculative_impl_draft_mtp() override {
+        auto * ctx_dft = this->params.ctx_dft;
+        for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) backend_chains.size(); ++seq_id) {
+            if (backend_chains[seq_id] == nullptr) {
+                continue;
+            }
+            if (ctx_dft) {
+                llama_set_sampler(ctx_dft, seq_id, nullptr);
+            }
+            llama_sampler_free(backend_chains[seq_id]);
+        }
+        backend_chains.clear();
+
         if (batch.token != nullptr) {
             free(batch.token);
             batch.token = nullptr;
@@ -459,7 +529,7 @@ struct common_speculative_state_draft_mtp : public common_speculative_impl {
         auto * ctx_dft = this->params.ctx_dft;
         const llama_pos pos_max = llama_memory_seq_pos_max(llama_get_memory(ctx_dft), seq_id);
         if (pos_max < N - 1) {
-            LOG_WRN("%s: ctx_dft pos_max=%d < N-1=%d — "
+            LOG_WRN("%s: ctx_dft pos_max=%d < N-1=%d - "
                     "process() hook may not have run on every prefill ubatch "
                     "(need_embd / logits=1 on every prompt position?). "
                     "Drafts may degrade.\n",
@@ -630,6 +700,14 @@ struct common_speculative_state_draft_mtp : public common_speculative_impl {
                 // add drafted token for each sequence
                 const llama_token id = cur_p->data[0].id;
 
+                // only collect very high-confidence draft tokens
+                if (cur_p->data[0].p < params.p_min) {
+                    drafting[seq_id] = false;
+                    n_drafting--;
+
+                    continue;
+                }
+
                 common_sampler_accept(smpl, id, true);
 
                 auto & dp = dparams.at(seq_id);
@@ -675,7 +753,7 @@ struct common_speculative_state_draft_mtp : public common_speculative_impl {
         }
     }
 
-    void accept(llama_seq_id seq_id, uint16_t n_accepted) override {
+    void accept(llama_seq_id seq_id, uint16_t n_accepted, bool /*is_other*/) override {
         if (seq_id < 0 || seq_id >= (llama_seq_id) n_seq) {
             return;
         }
@@ -691,6 +769,10 @@ struct common_speculative_state_draft_mtp : public common_speculative_impl {
     }
 
     bool need_embd() const override {
+        return false;
+    }
+
+    bool need_embd_pre_norm() const override {
         return true;
     }
 };
@@ -707,7 +789,12 @@ struct common_speculative_impl_ngram_simple : public common_speculative_impl {
             common_ngram_simple_config config)
         : common_speculative_impl(COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE, n_seq)
         , params(params.ngram_simple)
-        , config(config) {}
+        , config(config)
+    {
+        LOG_INF("%s: adding speculative implementation 'ngram-simple'\n", __func__);
+        LOG_INF("%s: - size_n=%d, size_m=%d, min_hits=%d\n", __func__,
+                this->params.size_n, this->params.size_m, this->params.min_hits);
+    }
 
     void begin(llama_seq_id /*seq_id*/, const llama_tokens & /*prompt*/) override {
         // noop
@@ -731,7 +818,7 @@ struct common_speculative_impl_ngram_simple : public common_speculative_impl {
         }
     }
 
-    void accept(llama_seq_id /*seq_id*/, uint16_t /*n_accepted*/) override {
+    void accept(llama_seq_id /*seq_id*/, uint16_t /*n_accepted*/, bool /*is_other*/) override {
         // noop
     }
 
@@ -741,20 +828,21 @@ struct common_speculative_impl_ngram_simple : public common_speculative_impl {
 };
 
 struct common_speculative_impl_ngram_map_k : public common_speculative_impl {
-    common_params_speculative_ngram_map params;
-
     // n_seq configs
     std::vector<common_ngram_map> config;
 
     common_speculative_impl_ngram_map_k(
-            const common_params_speculative & params,
             const common_ngram_map & config,
             uint32_t n_seq)
         : common_speculative_impl(COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K, n_seq)
-        , params(params.ngram_map_k) {
+    {
         for (uint32_t i = 0; i < n_seq; i++) {
             this->config.push_back(config);
         }
+
+        LOG_INF("%s: adding speculative implementation '%s'\n", __func__, common_speculative_type_to_str(this->type).c_str());
+        LOG_INF("%s: - size_key=%d, size_value=%d, key_only=%d, min_hits=%d\n", __func__,
+                config.size_key, config.size_value, config.key_only, config.min_hits);
     }
 
     void begin(llama_seq_id seq_id, const llama_tokens & prompt) override {
@@ -781,9 +869,13 @@ struct common_speculative_impl_ngram_map_k : public common_speculative_impl {
         }
     }
 
-    void accept(llama_seq_id seq_id, uint16_t n_accepted) override {
+    void accept(llama_seq_id seq_id, uint16_t n_accepted, bool is_other) override {
         GGML_ASSERT((seq_id < (llama_seq_id) config.size()));
 
+        if (is_other) {
+            return;
+        }
+
         common_ngram_map_accept(config[seq_id], n_accepted);
     }
 
@@ -805,7 +897,7 @@ struct common_speculative_impl_ngram_mod : public common_speculative_impl {
         // the last position in the prompt that was added to the ngram container
         size_t i_last = 0;
 
-        // length of the last drafted n‑gram (number of tokens returned by draft)
+        // length of the last drafted n-gram (number of tokens returned by draft)
         size_t n_draft_last = 0;
 
         // consecutive accept rounds with low acceptance fraction (< 0.5)
@@ -823,8 +915,11 @@ struct common_speculative_impl_ngram_mod : public common_speculative_impl {
         , verbose(std::getenv("LLAMA_TRACE") != nullptr) {
         static_assert(sizeof(llama_token) == sizeof(common_ngram_mod::entry_t));
 
-        LOG_INF("%s: initialized ngram_mod with n_match=%d, size=%zu (%.3f MB)\n", __func__,
-                this->params.n_match, mod.size(), (float)(mod.size_bytes())/1024/1024);
+        LOG_INF("%s: adding speculative implementation 'ngram-mod'\n", __func__);
+        LOG_INF("%s: - n_match=%d, n_max=%d, n_min=%d\n", __func__,
+                this->params.n_match, this->params.n_max, this->params.n_min);
+        LOG_INF("%s: - mod size=%zu (%.3f MB)\n", __func__,
+                mod.size(), (float)(mod.size_bytes())/1024/1024);
 
         if (this->params.n_match < 16) {
             LOG_WRN("%s: ngram_mod n_match=%d is too small - poor quality is possible, "
@@ -914,7 +1009,7 @@ struct common_speculative_impl_ngram_mod : public common_speculative_impl {
         }
         result.resize(result.size() - n);
 
-        // store length of drafted n‑gram for later acceptance analysis
+        // store length of drafted n-gram for later acceptance analysis
         sinfo.n_draft_last = result.size();
     }
 
@@ -936,17 +1031,21 @@ struct common_speculative_impl_ngram_mod : public common_speculative_impl {
         }
     }
 
-    void accept(llama_seq_id seq_id, uint16_t n_accepted) override {
+    void accept(llama_seq_id seq_id, uint16_t n_accepted, bool is_other) override {
+        if (is_other) {
+            return;
+        }
+
         auto & sinfo = sinfos[seq_id];
 
         // compute acceptance fraction if we have a recorded draft length
         if (sinfo.n_draft_last > 0) {
             const double f_acc = (double)n_accepted / (double)sinfo.n_draft_last;
-            if (f_acc < 0.5) {
+            if (f_acc < 0.25) {
                 sinfo.n_low++;
-                if (sinfo.n_low >= 3) {
+                if (sinfo.n_low >= 5) {
                     if (verbose) {
-                        LOG_WRN("%s: low acceptance streak (%d) – resetting ngram_mod\n", __func__, sinfo.n_low);
+                        LOG_WRN("%s: low acceptance streak (%d) - resetting ngram_mod\n", __func__, sinfo.n_low);
                     }
 
                     mod.reset();
@@ -996,6 +1095,12 @@ struct common_speculative_impl_ngram_cache : public common_speculative_impl {
         , save_dynamic(save_dynamic)
         , save_static(save_static)
     {
+        LOG_INF("%s: adding speculative implementation 'ngram-cache'\n", __func__);
+        LOG_INF("%s: - n_draft=%d, cache_static=%s, cache_dynamic=%s\n", __func__,
+                n_draft,
+                path_static.empty() ? "none" : path_static.c_str(),
+                path_dynamic.empty() ? "none" : path_dynamic.c_str());
+
         sinfos.resize(n_seq);
 
         if (!path_static.empty()) {
@@ -1092,7 +1197,7 @@ struct common_speculative_impl_ngram_cache : public common_speculative_impl {
         }
     }
 
-    void accept(llama_seq_id /*seq_id*/, uint16_t /*n_accepted*/) override {
+    void accept(llama_seq_id /*seq_id*/, uint16_t /*n_accepted*/, bool /*is_other*/) override {
         // noop
     }
 
@@ -1278,7 +1383,6 @@ common_speculative * common_speculative_init(common_params_speculative & params,
     std::vector<std::unique_ptr<common_speculative_impl>> impls = {};
 
     for (const common_speculative_config & config : configs) {
-        LOG_INF("%s: adding speculative implementation '%s'\n", __func__, common_speculative_type_to_str(config.type).c_str());
         switch (config.type) {
             case COMMON_SPECULATIVE_TYPE_NONE:
                 break;
@@ -1291,7 +1395,7 @@ common_speculative * common_speculative_init(common_params_speculative & params,
                 break;
             }
             case COMMON_SPECULATIVE_TYPE_DRAFT_MTP: {
-                impls.push_back(std::make_unique<common_speculative_state_draft_mtp>(config.params, n_seq));
+                impls.push_back(std::make_unique<common_speculative_impl_draft_mtp>(config.params, n_seq));
                 break;
             }
             case COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE: {
@@ -1312,11 +1416,16 @@ common_speculative * common_speculative_init(common_params_speculative & params,
                 impls.push_back(std::move(state));
                 break;
             }
-            case COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K:
+            case COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K: {
+                impls.push_back(
+                        std::make_unique<common_speculative_impl_ngram_map_k>(
+                            get_common_ngram_map(config.type, config.params.ngram_map_k), n_seq));
+                break;
+            }
             case COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V: {
                 impls.push_back(
                         std::make_unique<common_speculative_impl_ngram_map_k>(
-                            config.params, get_common_ngram_map(config.type, config.params.ngram_map_k), n_seq));
+                            get_common_ngram_map(config.type, config.params.ngram_map_k4v), n_seq));
                 break;
             }
             case COMMON_SPECULATIVE_TYPE_NGRAM_MOD: {
@@ -1408,6 +1517,20 @@ bool common_speculative_need_embd(common_speculative * spec) {
     return false;
 }
 
+bool common_speculative_need_embd_pre_norm(common_speculative * spec) {
+    if (spec == nullptr) {
+        return false;
+    }
+
+    for (auto & impl : spec->impls) {
+        if (impl->need_embd_pre_norm()) {
+            return true;
+        }
+    }
+
+    return false;
+}
+
 void common_speculative_draft(common_speculative * spec) {
     if (spec == nullptr) {
         return;
@@ -1494,11 +1617,6 @@ void common_speculative_accept(common_speculative * spec, llama_seq_id seq_id, u
 
     GGML_ASSERT(impl);
 
-    // TODO: currently only the implementation that generated the draft is used to accept it
-    //       however, some implementations (such as MTP) need to also "see" the accepted tokens
-    //       extend `common_speculative_impl::accept()` with an extra argument `bool is_other` to
-    //       inform the implementation if the accepted tokens are from another implementation and
-    //       pass the accepted tokens to all remaining implementations using `is_other == true`
     {
         common_time_meas tm(impl->t_accept_us, !impl->gen_perf);
         if (n_accepted > 0) {
@@ -1506,9 +1624,16 @@ void common_speculative_accept(common_speculative * spec, llama_seq_id seq_id, u
             impl->n_acc_tokens += n_accepted;
         }
 
-        impl->accept(seq_id, n_accepted);
+        impl->accept(seq_id, n_accepted, false);
         impl->n_call_accept++;
     }
+
+    // accept with the rest of the implementations, using is_other == true
+    for (auto & impl_other : spec->impls) {
+        if (impl_other.get() != impl) {
+            impl_other->accept(seq_id, n_accepted, true);
+        }
+    }
 }
 
 void common_speculative_print_stats(const common_speculative * spec) {
@@ -1528,7 +1653,7 @@ void common_speculative_print_stats(const common_speculative * spec) {
             str_perf = "";
         }
 
-        LOG_INF("statistics %s: #calls(b,g,a) = %zu %zu %zu, #gen drafts = %zu, #acc drafts = %zu, #gen tokens = %zu, #acc tokens = %zu%s\n",
+        LOG_INF("statistics %16s: #calls(b,g,a) = %4zu %6zu %6zu, #gen drafts = %6zu, #acc drafts = %5zu, #gen tokens = %6zu, #acc tokens = %5zu%s\n",
                 common_speculative_type_to_str(impl->type).c_str(),
                 impl->n_call_begin, impl->n_call_draft, impl->n_call_accept,
                 impl->n_gen_drafts,

common/speculative.h

@@ -53,9 +53,12 @@ void common_speculative_begin(common_speculative * spec, llama_seq_id seq_id, co
 // process the batch and update the internal state of the speculative context
 bool common_speculative_process(common_speculative * spec, const llama_batch & batch);
 
-// true if any implementation requires target embeddings to be extracted
+// true if any implementation requires target post-norm embeddings to be extracted
 bool common_speculative_need_embd(common_speculative * spec);
 
+// true if any implementation requires target pre-norm embeddings to be extracted
+bool common_speculative_need_embd_pre_norm(common_speculative * spec);
+
 // generate drafts for the sequences specified with `common_speculative_get_draft_params`
 void common_speculative_draft(common_speculative * spec);
 

conversion/base.py

@@ -1610,6 +1610,47 @@ class TextModel(ModelBase):
         special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
         special_vocab.add_to_gguf(self.gguf_writer)
 
+    def _set_vocab_hybriddna(self):
+        from transformers import AutoTokenizer
+        tokenizer = AutoTokenizer.from_pretrained(self.dir_model, trust_remote_code=True)
+        vocab_size = self.hparams.get("vocab_size", len(tokenizer.vocab))  # ty: ignore[unresolved-attribute]
+        assert max(tokenizer.vocab.values()) < vocab_size  # ty: ignore[unresolved-attribute]
+
+        reverse_vocab = {id_: encoded_tok for encoded_tok, id_ in tokenizer.vocab.items()}  # ty: ignore[unresolved-attribute]
+        # k-mers can share text with a base-vocab BPE token (e.g. CCCCCC) and get
+        # dropped by get_vocab(); a reserved marker suffix (U+E000) keeps each
+        # k-mer's own id (llama.cpp strips it on detokenization)
+        for kmer in tokenizer.kmers:  # ty: ignore[unresolved-attribute]
+            reverse_vocab[tokenizer.dna_token_to_id[kmer]] = kmer + "\ue000"  # ty: ignore[unresolved-attribute]
+        added_vocab = tokenizer.get_added_vocab()  # ty: ignore[unresolved-attribute]
+        added_tokens_decoder = tokenizer.added_tokens_decoder  # ty: ignore[unresolved-attribute]
+
+        tokens: list[str] = []
+        toktypes: list[int] = []
+        for i in range(vocab_size):
+            if i not in reverse_vocab:
+                tokens.append(f"[PAD{i}]")
+                toktypes.append(gguf.TokenType.UNUSED)
+            else:
+                token: str = reverse_vocab[i]
+                if token in added_vocab:
+                    if added_tokens_decoder[i].special or self.does_token_look_special(token):
+                        toktypes.append(gguf.TokenType.CONTROL)
+                    else:
+                        toktypes.append(gguf.TokenType.USER_DEFINED)
+                else:
+                    toktypes.append(gguf.TokenType.NORMAL)
+                tokens.append(token)
+
+        tokpre = self.get_vocab_base_pre(tokenizer)
+        self.gguf_writer.add_tokenizer_model("hybriddna")
+        self.gguf_writer.add_tokenizer_pre(tokpre)
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_types(toktypes)
+
+        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
+        special_vocab.add_to_gguf(self.gguf_writer)
+
     def _set_vocab_qwen(self):
         from .qwen import QwenModel
 

conversion/hunyuan.py

@@ -189,7 +189,8 @@ class HunYuanModel(TextModel):
             self.gguf_writer.add_token_list(tokens)
             self.gguf_writer.add_token_types(toktypes)
 
-            # HunyuanOCR has pad_token_id=-1 in config.json; exclude pad from SpecialVocab
+            # Some HunYuanVL variants (e.g. OCR-style configs) have pad_token_id=-1;
+            # guard SpecialVocab so it doesn't try to emit an invalid pad id.
             token_types = None
             if (self.hparams.get("pad_token_id") or 0) < 0:
                 token_types = ('bos', 'eos', 'unk', 'sep', 'cls', 'mask')
@@ -250,7 +251,8 @@ class HunYuanModel(TextModel):
             self._fix_special_tokens()
 
     def set_gguf_parameters(self):
-        # HunyuanOCR has num_experts=1 which is not MoE, prevent parent from writing it
+        # Some HunYuanVL variants set num_experts=1 (not real MoE);
+        # prevent the parent class from emitting expert_count metadata in that case.
         saved_num_experts = self.hparams.pop("num_experts", None)
         super().set_gguf_parameters()
         if saved_num_experts is not None and saved_num_experts > 1:
@@ -288,51 +290,21 @@ class HunYuanModel(TextModel):
 
 @ModelBase.register("HunYuanVLForConditionalGeneration")
 class HunyuanVLVisionModel(MmprojModel):
-    # Handles both HunyuanOCR and HunyuanVL, which share the HF architecture name
-    # "HunYuanVLForConditionalGeneration" and the `vit.perceive.*` vision layout.
-    # Each variant maps to a different projector type in clip.cpp so image
-    # preprocessing follows the correct code path.
-
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
         assert self.hparams_vision is not None
-        # HunyuanOCR / HunyuanVL uses max_image_size instead of image_size
+        # HunyuanVL uses max_image_size instead of image_size
         if "image_size" not in self.hparams_vision:
             self.hparams_vision["image_size"] = self.hparams_vision.get("max_image_size", 2048)
 
-    @staticmethod
-    def is_ocr_variant(hparams: dict) -> bool:
-        """Return True for HunyuanOCR, False for HunyuanVL.
-
-        The projector's output dim must equal the text model's hidden_size by
-        construction (that's what "projector" means). HunyuanOCR pairs a 1B text
-        backbone (hidden=1024); HunyuanVL pairs a 4B one (hidden=3072). So the
-        ViT -> LLM projection dim is a hard architectural signature, not a
-        magic number.
-        """
-        vision_out = int((hparams.get("vision_config") or {}).get("out_hidden_size", 0))
-        return vision_out == 1024
-
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
         assert self.hparams_vision is not None
         vcfg = self.hparams_vision
-
-        if self.is_ocr_variant(self.global_config):
-            # --- HunyuanOCR ---
-            self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.HUNYUANOCR)
-            self.gguf_writer.add_vision_use_gelu(True)
-            self.gguf_writer.add_vision_attention_layernorm_eps(vcfg.get("rms_norm_eps", 1e-5))
-            self.gguf_writer.add_vision_spatial_merge_size(vcfg.get("spatial_merge_size", 2))
-            self.gguf_writer.add_vision_min_pixels(self.preprocessor_config["min_pixels"])
-            self.gguf_writer.add_vision_max_pixels(self.preprocessor_config["max_pixels"])
-            return
-
-        # --- HunyuanVL ---
         self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.HUNYUANVL)
-        self.gguf_writer.add_vision_use_gelu(str(vcfg["hidden_act"]).lower() == "gelu")
-        self.gguf_writer.add_vision_attention_layernorm_eps(float(vcfg["rms_norm_eps"]))
-        self.gguf_writer.add_vision_spatial_merge_size(int(vcfg["spatial_merge_size"]))
+        self.gguf_writer.add_vision_use_gelu(True)
+        self.gguf_writer.add_vision_attention_layernorm_eps(vcfg.get("rms_norm_eps", 1e-5))
+        self.gguf_writer.add_vision_spatial_merge_size(vcfg.get("spatial_merge_size", 2))
         self.gguf_writer.add_vision_min_pixels(int(self.preprocessor_config["min_pixels"]))
         self.gguf_writer.add_vision_max_pixels(int(self.preprocessor_config["max_pixels"]))
 
@@ -353,7 +325,7 @@ class HunyuanVLVisionModel(MmprojModel):
 
     def tensor_force_quant(self, name, new_name, bid, n_dims):
         # force conv weights to F32 or F16 to avoid BF16 IM2COL issues on Metal
-        # Both HunyuanOCR and HunyuanVL emit the ViT -> LLM projection as mm.0/mm.2.
+        # HunyuanVL emit the ViT -> LLM projection as mm.0/mm.2.
         if ("mm.0." in new_name or "mm.2." in new_name) and new_name.endswith(".weight"):
             return gguf.GGMLQuantizationType.F16 if self.ftype == gguf.LlamaFileType.MOSTLY_F16 else gguf.GGMLQuantizationType.F32
         return super().tensor_force_quant(name, new_name, bid, n_dims)
@@ -361,40 +333,18 @@ class HunyuanVLVisionModel(MmprojModel):
 
 @ModelBase.register("HunYuanVLForConditionalGeneration")
 class HunyuanVLTextModel(HunYuanModel):
-    # The "HunYuanVLForConditionalGeneration" HF architecture covers both HunyuanOCR
-    # and HunyuanVL. HunyuanOCR reuses the HunYuan-Dense text backbone (standard RoPE),
-    # while HunyuanVL introduces a new LLM arch with XD-RoPE. Detect the variant from
-    # the config and pick the matching GGUF architecture.
     model_arch = gguf.MODEL_ARCH.HUNYUAN_VL
 
-    @staticmethod
-    def _is_ocr_config(hparams: dict) -> bool:
-        # OCR pairs a 1B text backbone (hidden=1024) with a ViT projector that
-        # outputs 1024-d; HunyuanVL uses 3072-d. Keep in sync with
-        # HunyuanVLVisionModel.is_ocr_variant.
-        return int((hparams.get("vision_config") or {}).get("out_hidden_size", 0)) == 1024
-
     def __init__(self, dir_model: Path, *args, **kwargs):
-        raw_hparams = kwargs.get("hparams") or ModelBase.load_hparams(dir_model, is_mistral_format=False)
-        if self._is_ocr_config(raw_hparams):
-            self.model_arch = gguf.MODEL_ARCH.HUNYUAN_DENSE
-        else:
-            self.model_arch = gguf.MODEL_ARCH.HUNYUAN_VL
         super().__init__(dir_model, *args, **kwargs)
 
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
 
-        # Only emit XD-RoPE metadata for the HunyuanVL backbone; HunyuanOCR uses
-        # the HunYuan-Dense arch which already handles standard rope in super().
-        if self.model_arch != gguf.MODEL_ARCH.HUNYUAN_VL:
-            return
-
+        # XD-RoPE metadata for the HunyuanVL;
         if self.rope_parameters.get("rope_type") != "xdrope":
             return
 
-        # defaults for HunyuanVL. The C++ side later computes:
-        #   freq_base = rope_theta * alpha ** (head_dim / (head_dim - 2))
         self.gguf_writer.add_rope_freq_base(float(self.rope_parameters["rope_theta"]))
         self.gguf_writer.add_rope_scaling_alpha(float(self.rope_parameters["alpha"]))
         self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.NONE)

conversion/llama.py

@@ -51,6 +51,15 @@ class LlamaModel(TextModel):
         if path_tekken_json.is_file() and not path_tokenizer_json.is_file():
             self._set_vocab_mistral()
 
+        tokenizer_config_file = self.dir_model / 'tokenizer_config.json'
+        if tokenizer_config_file.is_file():
+            with open(tokenizer_config_file, "r", encoding="utf-8") as f:
+                tokenizer_config_json = json.load(f)
+                if (add_prefix_space := tokenizer_config_json.get("add_prefix_space")) is not None:
+                    self.gguf_writer.add_add_space_prefix(add_prefix_space)
+                if tokenizer_config_json.get("tokenizer_class") == "HybridDNATokenizer":
+                    return self._set_vocab_hybriddna()
+
         try:
             self._set_vocab_sentencepiece()
         except FileNotFoundError:
@@ -72,13 +81,6 @@ class LlamaModel(TextModel):
             special_vocab._set_special_token("eot",    32010)
             special_vocab.add_to_gguf(self.gguf_writer)
 
-        tokenizer_config_file = self.dir_model / 'tokenizer_config.json'
-        if tokenizer_config_file.is_file():
-            with open(tokenizer_config_file, "r", encoding="utf-8") as f:
-                tokenizer_config_json = json.load(f)
-                if "add_prefix_space" in tokenizer_config_json:
-                    self.gguf_writer.add_add_space_prefix(tokenizer_config_json["add_prefix_space"])
-
         # Apply to granite small models only
         if self.hparams.get("vocab_size", 32000) == 49152:
             self.gguf_writer.add_add_bos_token(False)

conversion/qwen.py

@@ -600,6 +600,7 @@ class _Qwen35MtpMixin:
             if name.find("layers.") != -1:
                 assert bid is not None
                 name = name.replace(f"mtp.layers.{bid}", f"model.layers.{bid + n_layer}")
+                bid = bid + n_layer
             else:
                 remapper = {
                     "mtp.fc":                    "model.layers.{bid}.eh_proj",

convert_hf_to_gguf.py

@@ -115,15 +115,15 @@ def parse_args() -> argparse.Namespace:
     )
     parser.add_argument(
         "--mmproj", action="store_true",
-        help="(Experimental) Export multimodal projector (mmproj) for vision models. This will only work on some vision models. A prefix 'mmproj-' will be added to the output file name.",
+        help="Export multimodal projector (mmproj) for vision models. This will only work on some vision models. An 'mmproj-' prefix will be added to the output file name.",
     )
     parser.add_argument(
         "--mtp", action="store_true",
-        help="(Experimental) Export only the multi-token prediction (MTP) head as a separate GGUF, suitable for use as a speculative draft. Output file name will get a '-MTP' suffix.",
+        help="Export only the multi-token prediction (MTP) head as a separate GGUF, suitable for use as a speculative draft. An 'mtp-' prefix will be added to the output file name.",
     )
     parser.add_argument(
         "--no-mtp", action="store_true",
-        help="(Experimental) Exclude the multi-token prediction (MTP) head from the converted GGUF. Pair with --mtp on a second run to publish trunk and MTP as two files. Note: the split form duplicates embeddings, so the bundled default is more space-efficient overall.",
+        help="Exclude the multi-token prediction (MTP) head from the converted GGUF. Pair with --mtp on a second run to publish trunk and MTP as two files. Note: the split form duplicates embeddings, but even though the bundled default is more space-efficient overall, this allows differing quantization which may be more performant.",
     )
     parser.add_argument(
         "--mistral-format", action="store_true",

convert_lora_to_gguf.py

@@ -445,6 +445,11 @@ if __name__ == '__main__':
                     if self.lazy:
                         tensor = LazyTorchTensor.from_eager(tensor)
                     base_name = get_base_tensor_name(name)
+                    # filter base name, ignore tensor transformations for now
+                    data_gen = lambda g=tensor: g  # noqa: E731
+                    if (titem := self.filter_tensors((base_name, data_gen))) is None:
+                        continue
+                    base_name, _ = titem
                     # note: mergekit-extract-lora also adds token embeddings to the adapter
                     is_lora_a = ".lora_A.weight" in name or ".lora_embedding_A" in name
                     is_lora_b = ".lora_B.weight" in name or ".lora_embedding_B" in name

docs/autoparser.md

@@ -489,6 +489,7 @@ The following templates have active tests in `tests/test-chat.cpp`:
 | Qwen-QwQ-32B | Reasoning | Forced-open thinking |
 | NousResearch Hermes 2 Pro | JSON_NATIVE | `<tool_call>` wrapper |
 | IBM Granite 3.3 | JSON_NATIVE | `<think></think>` + `<response></response>` |
+| IBM Granite 4.0 | JSON_NATIVE | `<tool_call>` wrapper (same template used by 4.1) |
 | ByteDance Seed-OSS | TAG_WITH_TAGGED | Custom `<seed:think>` and `<seed:tool_call>` tags |
 | Qwen3-Coder | TAG_WITH_TAGGED | XML-style tool format |
 | DeepSeek V3.1 | JSON_NATIVE | Forced thinking mode |

docs/backend/SYCL.md

@@ -5,6 +5,7 @@
 - [News](#news)
 - [OS](#os)
 - [Hardware](#hardware)
+- [Performance Reference](#performance-reference)
 - [Docker](#docker)
 - [Linux](#linux)
 - [Windows](#windows)
@@ -51,9 +52,8 @@ The packages for FP32 and FP16 would have different accuracy and performance on
 
 ## News
 
-- 2026.04
-
-  - Optimize mul_mat by reorder feature for data type: Q4_K, Q5_K, Q_K, Q8_0.
+- 2026.04-05
+  - Optimize mul_mat by reorder feature for data type: Q4_K, Q5_K, Q6_K, Q8_0.
   - Fused MoE.
   - Upgrate CI and built package for oneAPI 2025.3.3, support Ubuntu 24.04 built package.
 
@@ -150,6 +150,13 @@ On older Intel GPUs, you may try [OpenCL](/docs/backend/OPENCL.md) although the
 
 NA
 
+## Performance Reference
+
+
+To get the supported LLMs, GPUs, and performance reference, please check [Performance of llama.cpp on Intel GPU with SYCL backend](https://github.com/ggml-org/llama.cpp/discussions/23313).
+
+You could update your test result in it directly.
+
 ## Docker
 
 The docker build option is currently limited to *Intel GPU* targets.

docs/backend/snapdragon/CMakeUserPresets.json

@@ -10,8 +10,8 @@
             "ANDROID_ABI":      "arm64-v8a",
             "ANDROID_PLATFORM": "android-31",
             "CMAKE_TOOLCHAIN_FILE": "$env{ANDROID_NDK_ROOT}/build/cmake/android.toolchain.cmake",
-            "CMAKE_C_FLAGS":   "-march=armv8.7a+fp16 -fvectorize -ffp-model=fast -fno-finite-math-only -flto -D_GNU_SOURCE",
-            "CMAKE_CXX_FLAGS": "-march=armv8.7a+fp16 -fvectorize -ffp-model=fast -fno-finite-math-only -flto -D_GNU_SOURCE",
+            "CMAKE_C_FLAGS":   "-march=armv8.7a+fp16+dotprod+i8mm -fvectorize -ffp-model=fast -fno-finite-math-only -flto -D_GNU_SOURCE",
+            "CMAKE_CXX_FLAGS": "-march=armv8.7a+fp16+dotprod+i8mm -fvectorize -ffp-model=fast -fno-finite-math-only -flto -D_GNU_SOURCE",
             "CMAKE_C_FLAGS_RELEASE":          "-O3 -DNDEBUG",
             "CMAKE_CXX_FLAGS_RELEASE":        "-O3 -DNDEBUG",
             "CMAKE_C_FLAGS_RELWITHDEBINFO":   "-O3 -DNDEBUG -g",
@@ -33,8 +33,8 @@
         "name": "arm64-windows-snapdragon",
         "inherits": [ "base", "arm64-windows-llvm" ],
         "cacheVariables": {
-            "CMAKE_C_FLAGS":   "-march=armv8.7a+fp16 -fvectorize -ffp-model=fast -flto -D_GNU_SOURCE",
-            "CMAKE_CXX_FLAGS": "-march=armv8.7a+fp16 -fvectorize -ffp-model=fast -flto -D_GNU_SOURCE",
+            "CMAKE_C_FLAGS":   "-march=armv8.7a+fp16+dotprod+i8mm -fvectorize -ffp-model=fast -flto -D_GNU_SOURCE",
+            "CMAKE_CXX_FLAGS": "-march=armv8.7a+fp16+dotprod+i8mm -fvectorize -ffp-model=fast -flto -D_GNU_SOURCE",
             "CMAKE_C_FLAGS_RELEASE":          "-O3 -DNDEBUG",
             "CMAKE_CXX_FLAGS_RELEASE":        "-O3 -DNDEBUG",
             "CMAKE_C_FLAGS_RELWITHDEBINFO":   "-O3 -DNDEBUG -g",
@@ -59,8 +59,8 @@
         "toolset":      { "value": "host=x86_64", "strategy": "external" },
         "cacheVariables": {
             "CMAKE_TOOLCHAIN_FILE": "cmake/arm64-linux-clang.cmake",
-            "CMAKE_C_FLAGS":   "-march=armv8 -fno-finite-math-only -flto -D_GNU_SOURCE",
-            "CMAKE_CXX_FLAGS": "-march=armv8 -fno-finite-math-only -flto -D_GNU_SOURCE",
+            "CMAKE_C_FLAGS":   "-march=armv8.2a+fp16+dotprod -fvectorize -fno-finite-math-only -flto -D_GNU_SOURCE",
+            "CMAKE_CXX_FLAGS": "-march=armv8.2a+fp16+dotprod -fvectorize -fno-finite-math-only -flto -D_GNU_SOURCE",
             "CMAKE_C_FLAGS_RELEASE":          "-O3 -DNDEBUG",
             "CMAKE_CXX_FLAGS_RELEASE":        "-O3 -DNDEBUG",
             "CMAKE_C_FLAGS_RELWITHDEBINFO":   "-O3 -DNDEBUG -g",

docs/backend/snapdragon/README.md

@@ -10,7 +10,7 @@ This image includes Android NDK, OpenCL SDK, Hexagon SDK, CMake, etc.
 This method works on Linux, macOS, and Windows. macOS and Windows users should install Docker Desktop.
 
 ```
-~/src/llama.cpp$ docker run -it -u $(id -u):$(id -g) --volume $(pwd):/workspace --platform linux/amd64 ghcr.io/snapdragon-toolchain/arm64-android:v0.3
+~/src/llama.cpp$ docker run -it -u $(id -u):$(id -g) --volume $(pwd):/workspace --platform linux/amd64 ghcr.io/snapdragon-toolchain/arm64-android:v0.6
 [d]/> cd /workspace
 ```
 
@@ -24,7 +24,7 @@ Native Windows 11 arm64 builds has the following tools dependencies:
   - UCRT and Driver Kit
 - LLVM core libraries and Clang compiler (winget)
 - CMake, Git, Python (winget)
-- Hexagon SDK Community Edition 6.4 or later (see windows.md)
+- Hexagon SDK Community Edition 6.6 or later (see windows.md)
 - OpenCL SDK 2.3 or later (see windows.md)
 
 Note: The rest of the **Windows** build process assumes that you're running natively in Powershell.
@@ -45,7 +45,7 @@ Preset CMake variables:
   GGML_HEXAGON="ON"
   GGML_OPENCL="ON"
   GGML_OPENMP="OFF"
-  HEXAGON_SDK_ROOT="/opt/hexagon/6.4.0.2"
+  HEXAGON_SDK_ROOT="/opt/hexagon/6.6.0.0"
 ...
 -- Including OpenCL backend
 -- Including Hexagon backend

docs/backend/snapdragon/windows.md

@@ -28,15 +28,15 @@ c:\Qualcomm\OpenCL_SDK\2.3.2
 
 Either use the trimmed down version (optimized for CI) from
 
-    https://github.com/snapdragon-toolchain/hexagon-sdk/releases/download/v6.4.0.2/hexagon-sdk-v6.4.0.2-arm64-wos.tar.xz
+    https://github.com/snapdragon-toolchain/hexagon-sdk/releases/download/v6.6.0.0/hexagon-sdk-v6.6.0.0-arm64-wos.tar.xz
 
 Or download the complete official version from
 
-    https://softwarecenter.qualcomm.com/catalog/item/Hexagon_SDK?version=6.4.0.2
+    https://softwarecenter.qualcomm.com/catalog/item/Hexagon_SDK?version=6.6.0.0
 
 Unzip/untar the archive into
 ```
-c:\Qualcomm\Hexagon_SDK\6.4.0.2
+c:\Qualcomm\Hexagon_SDK\6.6.0.0
 ```
 
 ## Install the latest Adreno GPU driver
@@ -123,10 +123,10 @@ The overall Hexagon backend build procedure for Windows on Snapdragon is the sam
 However, additional settings are required for generating and signing HTP Ops libraries.
 ```
 > $env:OPENCL_SDK_ROOT="C:\Qualcomm\OpenCL_SDK\2.3.2"
-> $env:HEXAGON_SDK_ROOT="C:\Qualcomm\Hexagon_SDK\6.4.0.2"
-> $env:HEXAGON_TOOLS_ROOT="C:\Qualcomm\Hexagon_SDK\6.4.0.2\tools\HEXAGON_Tools\19.0.04"
+> $env:HEXAGON_SDK_ROOT="C:\Qualcomm\Hexagon_SDK\6.6.0.0"
+> $env:HEXAGON_TOOLS_ROOT="C:\Qualcomm\Hexagon_SDK\6.6.0.0\tools\HEXAGON_Tools\19.0.07"
 > $env:HEXAGON_HTP_CERT="c:\Users\MyUsers\Certs\ggml-htp-v1.pfx"
-> $env:WINDOWS_SDK_BIN="C:\Program Files (x86)\Windows Kits\10\bin\10.0.26100.0\arm64"
+> $env:WINDOWS_SDK_BIN="C:\Program Files (x86)\Windows Kits\10\bin\10.0.26100.0"
 
 > cmake --preset arm64-windows-snapdragon-release -B build-wos
 ...

docs/build.md

@@ -735,7 +735,7 @@ ninja
 
 To read documentation for how to build on Android, [click here](./android.md)
 
-## WebGPU [In Progress]
+## WebGPU
 
 The WebGPU backend relies on [Dawn](https://dawn.googlesource.com/dawn). Follow the instructions [here](https://dawn.googlesource.com/dawn/+/refs/heads/main/docs/quickstart-cmake.md) to install Dawn locally so that llama.cpp can find it using CMake. The current implementation is up-to-date with Dawn commit `18eb229`.
 

docs/function-calling.md

@@ -291,6 +291,7 @@ Here are some models known to work (w/ chat template override when needed):
 llama-server --jinja -fa -hf bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M
 llama-server --jinja -fa -hf bartowski/Mistral-Nemo-Instruct-2407-GGUF:Q6_K_L
 llama-server --jinja -fa -hf bartowski/Llama-3.3-70B-Instruct-GGUF:Q4_K_M
+llama-server --jinja -fa -hf ibm-granite/granite-4.1-3b-GGUF:Q4_K_M
 
 # Native support for DeepSeek R1 works best w/ our template override (official template is buggy, although we do work around it)
 

docs/speculative.md

@@ -108,11 +108,12 @@ If a draft model is combined with a draftless decoding the draftless decoding ha
 ### General Speculative Parameters
 
 ```
---spec-type [none|ngram-cache|ngram-simple|ngram-map-k|ngram-map-k4v|ngram-mod]
-                                        type of speculative decoding to use when no draft model is provided
+--spec-type [none|draft-simple|draft-mtp|ngram-cache|ngram-simple|ngram-map-k|ngram-map-k4v|ngram-mod]
+                                        comma-separated list of types of speculative decoding to use
                                         (default: none)
                                         (env: LLAMA_ARG_SPEC_TYPE)
---spec-default                          use default speculative decoding
+--spec-default                          use default speculative decoding config
+                                        (enables ngram-mod)
 ```
 
 ### Draft Model Parameters
@@ -123,8 +124,9 @@ If a draft model is combined with a draftless decoding the draftless decoding ha
                                         (env: LLAMA_ARG_SPEC_DRAFT_MODEL)
 --spec-draft-hf, -hfd, -hfrd, --hf-repo-draft  <user>/<model>[:quant]
                                         HuggingFace repository for the draft model
+                                        (env: LLAMA_ARG_SPEC_DRAFT_HF_REPO)
 --spec-draft-n-max                      N
-                                        number of tokens to draft for speculative decoding (default: 16)
+                                        number of tokens to draft for speculative decoding (default: 3)
                                         (env: LLAMA_ARG_SPEC_DRAFT_N_MAX)
 --spec-draft-n-min                      N
                                         minimum number of draft tokens to use for speculative decoding (default: 0)
@@ -133,18 +135,64 @@ If a draft model is combined with a draftless decoding the draftless decoding ha
                                         speculative decoding split probability (default: 0.10)
                                         (env: LLAMA_ARG_SPEC_DRAFT_P_SPLIT)
 --spec-draft-p-min, --draft-p-min       P
-                                        minimum speculative decoding probability (greedy) (default: 0.75)
+                                        minimum speculative decoding probability (greedy) (default: 0.00)
                                         (env: LLAMA_ARG_SPEC_DRAFT_P_MIN)
---spec-draft-ctx-size, -cd, --ctx-size-draft  N
-                                        size of the prompt context for the draft model (default: 0, 0 = loaded from model)
-                                        (env: LLAMA_ARG_SPEC_DRAFT_CTX_SIZE)
 --spec-draft-ngl, -ngld, --gpu-layers-draft, --n-gpu-layers-draft  N
                                         max. number of draft model layers to store in VRAM, either an exact number, 'auto', or 'all' (default: auto)
                                         (env: LLAMA_ARG_N_GPU_LAYERS_DRAFT)
 --spec-draft-device, -devd, --device-draft  <dev1,dev2,..>
                                         comma-separated list of devices to use for offloading the draft model
---spec-draft-replace, --spec-replace    TARGET  DRAFT
-                                        translate the string in TARGET into DRAFT if the draft model and main model are not compatible
+                                        (use --list-devices to see available devices)
+```
+
+### Draft Model CPU Scheduling Parameters
+
+```
+--spec-draft-threads, -td, --threads-draft  N
+                                        number of CPU threads to use during generation
+--spec-draft-threads-batch, -tbd, --threads-batch-draft  N
+                                        number of threads to use during batch and prompt processing (default: same as --threads-draft)
+--spec-draft-cpu-mask, -Cd, --cpu-mask-draft  M
+                                        Draft model CPU affinity mask. Complements cpu-range-draft
+--spec-draft-cpu-range, -Crd, --cpu-range-draft  lo-hi
+                                        Ranges of CPUs for affinity. Complements --cpu-mask-draft
+--spec-draft-cpu-strict, --cpu-strict-draft  <0|1>
+                                        Use strict CPU placement for draft model (default: same as --cpu-strict)
+--spec-draft-prio, --prio-draft  N
+                                        set draft process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime
+--spec-draft-poll, --poll-draft  <0|1>
+                                        Use polling to wait for draft model work (default: same as --poll)
+--spec-draft-cpu-mask-batch, -Cbd, --cpu-mask-batch-draft  M
+                                        Draft model CPU affinity mask for batch. Complements cpu-range-batch-draft
+--spec-draft-cpu-range-batch, -Crbd, --cpu-range-batch-draft  lo-hi
+                                        Ranges of CPUs for affinity for batch. Complements --cpu-mask-batch-draft
+--spec-draft-cpu-strict-batch, --cpu-strict-batch-draft  <0|1>
+                                        Use strict CPU placement for draft model batch (default: --cpu-strict-draft)
+--spec-draft-prio-batch, --prio-batch-draft  N
+                                        set draft process/thread priority for batch : 0-normal, 1-medium, 2-high, 3-realtime
+--spec-draft-poll-batch, --poll-batch-draft  <0|1>
+                                        Use polling to wait for draft model work for batch (default: --poll-draft)
+```
+
+### Draft Model KV Cache and Tensor Override Parameters
+
+```
+--spec-draft-type-k, -ctkd, --cache-type-k-draft  TYPE
+                                        KV cache data type for K for the draft model
+                                        allowed values: f32, f16, bf16, q8_0, q4_0, q4_1, iq4_nl, q5_0, q5_1
+                                        (env: LLAMA_ARG_SPEC_DRAFT_CACHE_TYPE_K)
+--spec-draft-type-v, -ctvd, --cache-type-v-draft  TYPE
+                                        KV cache data type for V for the draft model
+                                        allowed values: f32, f16, bf16, q8_0, q4_0, q4_1, iq4_nl, q5_0, q5_1
+                                        (env: LLAMA_ARG_SPEC_DRAFT_CACHE_TYPE_V)
+--spec-draft-override-tensor, -otd, --override-tensor-draft  <tensor name pattern>=<buffer type>,...
+                                        override tensor buffer type for draft model
+--spec-draft-cpu-moe, -cmoed, --cpu-moe-draft
+                                        keep all Mixture of Experts (MoE) weights in the CPU for the draft model
+                                        (env: LLAMA_ARG_SPEC_DRAFT_CPU_MOE)
+--spec-draft-n-cpu-moe, --spec-draft-ncmoe, -ncmoed, --n-cpu-moe-draft  N
+                                        keep the MoE weights of the first N layers in the CPU for the draft model
+                                        (env: LLAMA_ARG_SPEC_DRAFT_N_CPU_MOE)
 ```
 
 ### n-gram Mod Parameters
@@ -193,11 +241,13 @@ If a draft model is combined with a draftless decoding the draftless decoding ha
 
 ### `--spec-type TYPE`
 
-Specifies a type of speculative decoding without draft model.
+Specifies a comma-separated list of speculative decoding types to use.
 
 | Type | Description |
 |------|-------------|
 | `none` | No speculative decoding (default) |
+| `draft-simple` | Use a simple draft model for speculation |
+| `draft-mtp` | Use Multi Token Prediction (MTP) heads from the main model |
 | `ngram-cache` | Use n-gram cache lookup |
 | `ngram-simple` | Use simple n-gram pattern matching |
 | `ngram-map-k` | Use n-gram pattern matching with n-gram-keys |
@@ -209,6 +259,11 @@ Specifies a type of speculative decoding without draft model.
 ./llama-server [...] --spec-type ngram-simple
 ```
 
+**Example:** Multiple speculative implementations.
+```bash
+./llama-server [...] --spec-type ngram-mod,ngram-map-k4v
+```
+
 ### `--spec-ngram-*-size-n N`
 
 Sets the size N of the lookup n-gram for n-gram map based speculative decoding.

examples/CMakeLists.txt

@@ -27,7 +27,6 @@ else()
     add_subdirectory(parallel)
     add_subdirectory(passkey)
     add_subdirectory(retrieval)
-    add_subdirectory(save-load-state)
     add_subdirectory(simple)
     add_subdirectory(simple-chat)
     add_subdirectory(speculative)

examples/llama-eval/llama-eval.py

@@ -149,6 +149,8 @@ class TaskState:
     t_gen_ms: Optional[float] = None
     reasoning_content: Optional[str] = None
     server_name: Optional[str] = None
+    chunk_idx: int = 0
+    problem_idx: int = 0
 
 
 class EvalState:
@@ -233,7 +235,9 @@ class EvalState:
         tps_gen: Optional[float] = None,
         t_gen_ms: Optional[float] = None,
         reasoning_content: Optional[str] = None,
-        server_name: Optional[str] = None
+        server_name: Optional[str] = None,
+        chunk_idx: int = 0,
+        problem_idx: int = 0,
     ):
         with self._lock:
             if "cases" not in self.task_states:
@@ -252,7 +256,9 @@ class EvalState:
                 "tps_gen": tps_gen,
                 "t_gen_ms": t_gen_ms,
                 "reasoning_content": reasoning_content,
-                "server_name": server_name
+                "server_name": server_name,
+                "chunk_idx": chunk_idx,
+                "problem_idx": problem_idx,
             }
 
             self.correct = sum(1 for c in self.task_states.get("cases", {}).values() if c.get("correct", False))
@@ -289,6 +295,9 @@ class EvalState:
             all_cases = {}
             for i, task_id in tasks_to_save:
                 question_text, prompt, expected = self.get_case(i)
+                # Extract chunk_idx from task_id for pending cases
+                _parts = task_id.rsplit("_", 2)
+                _chunk_idx = int(_parts[-2]) if len(_parts) >= 3 else 0
                 if task_id in self.task_states.get("cases", {}):
                     all_cases[task_id] = self.task_states["cases"][task_id]
                 else:
@@ -306,7 +315,9 @@ class EvalState:
                         "tps_gen": None,
                         "t_gen_ms": None,
                         "reasoning_content": None,
-                        "server_name": None
+                        "server_name": None,
+                        "chunk_idx": _chunk_idx,
+                        "problem_idx": i,
                     }
 
             ci_lower, ci_upper = self.accuracy_ci()
@@ -382,11 +393,12 @@ class EvalState:
             grader_log_str = self._escape_html(json.dumps(grader_log, indent=2))
             escaped_server = self._escape_html(server_name)
 
+            answer_class = status_class if status == "ok" else ""
             rows.append(f"""<tr class="task-row" onclick="toggleDetails('{task_id}')">
                 <td>{task_id}</td>
                 <td class="{status_class}">{status_text}</td>
                 <td>{self._escape_html(expected)}</td>
-                <td>{self._escape_html(answer)}</td>
+                <td class="{answer_class}">{self._escape_html(answer)}</td>
                 <td>{tokens_str}</td>
                 <td>{tps_str}</td>
                 <td>{t_gen_str}</td>
@@ -405,6 +417,53 @@ class EvalState:
 
         rows_html = "\n".join(rows)
 
+        # ---- per-problem summary table ----
+        problem_groups: Dict[int, List[Dict[str, Any]]] = {}
+        for _tid, _case in cases.items():
+            if _case.get("status") != "ok":
+                continue
+            _pidx = _case.get("problem_idx")
+            if _pidx is None:
+                _p_parts = _tid.rsplit("_", 2)
+                _pidx = int(_p_parts[-1]) if len(_p_parts) >= 3 else 0
+            problem_groups.setdefault(_pidx, []).append(_case)
+
+        summary_rows_html = ""
+        if problem_groups:
+            def _stat(v, fmt=".1f", avg_fmt=None):
+                if not v:
+                    return ("–", "–", "–")
+                af = fmt if avg_fmt is None else avg_fmt
+                return (f"{min(v):{fmt}}", f"{sum(v)/len(v):{af}}", f"{max(v):{fmt}}")
+
+            summary_data = []
+            for pidx, g in problem_groups.items():
+                runs = len(g)
+                n_ok = sum(1 for c in g if c.get("correct", False))
+                toks = [c["tokens"] for c in g if c.get("tokens") is not None]
+                tps = [c["tps_gen"] for c in g if c.get("tps_gen") is not None]
+                tg = [c["t_gen_ms"] / 1000 for c in g if c.get("t_gen_ms") is not None]
+                summary_data.append((
+                    pidx, runs, n_ok,
+                    _stat(toks, "d", ".0f"),
+                    _stat(tps),
+                    _stat(tg),
+                ))
+
+            summary_data.sort(key=lambda r: r[0])  # sort by problem index ascending
+
+            summary_rows_html = "\n".join(
+                f"""<tr class="summary-row">
+                    <td>{p:03d}</td>
+                    <td>{r}</td>
+                    <td>{n}/{r}</td>
+                    <td>{tk[0]}</td><td>{tk[1]}</td><td>{tk[2]}</td>
+                    <td>{tp[0]}</td><td>{tp[1]}</td><td>{tp[2]}</td>
+                    <td>{tg[0]}</td><td>{tg[1]}</td><td>{tg[2]}</td>
+                </tr>"""
+                for p, r, n, tk, tp, tg in summary_data
+            )
+
         html_content = f"""<!DOCTYPE html>
 <html>
 <head>
@@ -412,10 +471,10 @@ class EvalState:
 <title>{self.dataset_type.upper()} Eval</title>
 <style>
         body {{ font-family: system-ui, sans-serif; margin: 0; padding: 16px; background: #fff; color: #222; }}
-        .bar {{ padding: 8px 0; font-size: 14px; color: #555; }}
-        .bar span {{ margin-right: 20px; }}
-        .bar b {{ color: #222; }}
-        table {{ width: 100%; border-collapse: collapse; font-size: 13px; }}
+        .bar {{ padding: 8px 0; font-size: 13px; color: #555; font-family: 'SF Mono', 'Menlo', 'Consolas', monospace; display: grid; grid-template-columns: auto 1fr auto 1fr; gap: 2px 12px; align-items: baseline; }}
+        .bar .label {{ color: #888; }}
+        .bar .value {{ color: #222; }}
+        table {{ width: 100%; border-collapse: collapse; font-size: 13px; font-family: 'SF Mono', 'Menlo', 'Consolas', monospace; }}
         th {{ text-align: left; padding: 6px 8px; border-bottom: 2px solid #ccc; font-weight: 600; }}
         td {{ padding: 4px 8px; border-bottom: 1px solid #eee; vertical-align: top; }}
         .task-row {{ cursor: pointer; }}
@@ -429,37 +488,88 @@ class EvalState:
         .details-content {{ padding: 8px 16px; background: #f6f8fa; font-size: 12px; }}
         .details-content b {{ color: #555; }}
         .details-content pre {{ background: #fff; border: 1px solid #e1e4e8; padding: 8px; overflow-x: auto; white-space: pre-wrap; word-wrap: break-word; margin: 4px 0 8px; }}
+        .summary-table {{ margin-bottom: 16px; font-size: 13px; width: 100%; }}
+        .summary-row {{ background: #fafbfc; }}
+        .summary-row:hover {{ background: #f5f5f5; }}
+        .summary-table th {{ text-align: right; font-weight: 600; }}
+        .summary-table th:first-child {{ text-align: left; }}
+        .summary-table th[colspan] {{ text-align: center; }}
+        .summary-table td {{ text-align: right; }}
+        .summary-table td:first-child {{ text-align: left; }}
+        .tabs {{ display: flex; border-bottom: 2px solid #ddd; margin: 12px 0 0; }}
+        .tab-btn {{ padding: 6px 16px; border: none; background: none; font-size: 13px; cursor: pointer; color: #555; border-bottom: 2px solid transparent; margin-bottom: -2px; font-weight: 500; }}
+        .tab-btn:hover {{ color: #222; }}
+        .tab-btn.active {{ color: #222; border-bottom-color: #222; font-weight: 600; }}
+        .tab-content {{ display: none; }}
+        .tab-content.active {{ display: block; }}
 </style>
 </head>
 <body>
     <div class="bar">
-        <span><b>{self.dataset_type.upper()}</b></span>
-        <span>Model: {self.model_name or 'N/A'}</span>
-        <span>Accuracy: <b>{accuracy:.1f}%</b> [{ci_lower*100:.1f}%, {ci_upper*100:.1f}%]</span>
-        <span>Correct: <span class="correct">{n_correct}</span> / {len(completed)}</span>
-        <span>Pending: {n_pending}</span>
-        <span>Time: {self.total_time:.1f}s</span>
-        <span>Sampling: {sampling_str}</span>
+        <div class="label">Dataset</div><div class="value"><b>{self.dataset_type.upper()}</b></div>
+        <div class="label">Model</div><div class="value"><b>{self.model_name or 'N/A'}</b></div>
+        <div class="label">Accuracy</div><div class="value"><b>{accuracy:.1f}%</b> [{ci_lower*100:.1f}%, {ci_upper*100:.1f}%]</div>
+        <div class="label">Correct</div><div class="value"><span class="correct">{n_correct}</span> / {len(completed)}</div>
+        <div class="label">Pending</div><div class="value">{n_pending}</div>
+        <div class="label">Time</div><div class="value">{self.total_time:.1f}s</div>
+        <div class="label">Sampling</div><div class="value">{sampling_str}</div>
+    </div>
+    <div class="tabs">
+        <button class="tab-btn active" data-tab="detailed" onclick="switchTab(this)">Detailed</button>
+        <button class="tab-btn" data-tab="summary" onclick="switchTab(this)">Summary</button>
+    </div>
+    <div id="tab-detailed" class="tab-content active">
+        <table>
+            <thead>
+                <tr>
+                    <th>ID</th>
+                    <th></th>
+                    <th>Gold</th>
+                    <th>Answer</th>
+                    <th>Tokens</th>
+                    <th>T/s</th>
+                    <th>Gen s</th>
+                    <th>Server</th>
+                </tr>
+            </thead>
+            <tbody>
+                {rows_html}
+            </tbody>
+        </table>
+    </div>
+    <div id="tab-summary" class="tab-content">
+        <table class="summary-table">
+            <thead>
+                <tr>
+                    <th>Problem</th>
+                    <th>Runs</th>
+                    <th>Correct</th>
+                    <th colspan="3">Tokens</th>
+                    <th colspan="3">T/s</th>
+                    <th colspan="3">Gen s</th>
+                </tr>
+                <tr>
+                    <th></th>
+                    <th></th>
+                    <th></th>
+                    <th>min</th><th>avg</th><th>max</th>
+                    <th>min</th><th>avg</th><th>max</th>
+                    <th>min</th><th>avg</th><th>max</th>
+                </tr>
+            </thead>
+            <tbody>
+                {summary_rows_html}
+            </tbody>
+        </table>
     </div>
-    <table>
-        <thead>
-            <tr>
-                <th>ID</th>
-                <th></th>
-                <th>Gold</th>
-                <th>Answer</th>
-                <th>Tokens</th>
-                <th>T/s</th>
-                <th>Gen s</th>
-                <th>Server</th>
-            </tr>
-        </thead>
-        <tbody>
-            {rows_html}
-        </tbody>
-    </table>
     <script>
         function toggleDetails(id) {{ document.getElementById('details-'+id).classList.toggle('open'); }}
+        function switchTab(btn) {{
+            document.querySelectorAll('.tab-btn').forEach(b => b.classList.remove('active'));
+            document.querySelectorAll('.tab-content').forEach(c => c.classList.remove('active'));
+            btn.classList.add('active');
+            document.getElementById('tab-'+btn.dataset.tab).classList.add('active');
+        }}
     </script>
 </body>
 </html>"""
@@ -1062,12 +1172,19 @@ class Processor:
     ) -> TaskState:
         question_text, prompt, expected = eval_state.get_case(i)
 
+        # Extract chunk_idx from task_id: "{dataset_type}_{chunk_idx:03d}_{index:03d}"
+        _parts = task_id.rsplit("_", 2)
+        chunk_idx = int(_parts[-2]) if len(_parts) >= 3 else 0
+        problem_idx = i
+
         task_state = TaskState(
             task_id=task_id,
             prompt=prompt,
             expected=expected,
             question_text=question_text,
-            server_name=server_config.name
+            server_name=server_config.name,
+            chunk_idx=chunk_idx,
+            problem_idx=problem_idx,
         )
 
         try:
@@ -1085,7 +1202,8 @@ class Processor:
                 eval_state.add_result(
                     task_id, prompt, expected, result, None,
                     {"finish_reason": finish_reason}, False, task_state.status,
-                    tokens, tps_gen, t_gen_ms, reasoning_content, server_config.name
+                    tokens, tps_gen, t_gen_ms, reasoning_content, server_config.name,
+                    chunk_idx, problem_idx,
                 )
                 eval_state.dump()
                 return task_state
@@ -1108,7 +1226,8 @@ class Processor:
             eval_state.add_result(
                 task_id, prompt, expected, result, answer,
                 grader_log, is_correct, "ok",
-                tokens, tps_gen, t_gen_ms, reasoning_content, server_config.name
+                tokens, tps_gen, t_gen_ms, reasoning_content, server_config.name,
+                chunk_idx, problem_idx,
             )
 
             eval_state.dump()

examples/llama-eval/llama-server-simulator.py

@@ -65,34 +65,70 @@ def normalize_number(s: str) -> Optional[int]:
     return int(match.group(0))
 
 class AimeDataset:
-    def __init__(self, split: str = "train"):
+    def __init__(self, split: str = "train", dataset_type: str = "aime"):
         self.split = split
+        self.dataset_type = dataset_type
         self.questions: List[Dict] = []
         self._load_dataset()
 
-    def _load_dataset(self):
-        print(f"Loading AIME dataset (split: {self.split})...")
+    def _get_question_text(self, question: Dict) -> str:
+        """Get question text, handling different dataset field names."""
+        return question.get("problem", question.get("question", ""))
 
-        cache_path = Path.home() / ".cache" / "huggingface" / "datasets" / "AI-MO___aimo-validation-aime" / "default" / "0.0.0"
-        if cache_path.exists():
-            print(f"Using cached dataset from {cache_path}")
-            ds = datasets.load_dataset("AI-MO/aimo-validation-aime", split=self.split, cache_dir=str(cache_path))
+    def _load_dataset(self):
+        if self.dataset_type == "aime":
+            print(f"Loading AIME dataset (split: {self.split})...")
+            cache_path = Path.home() / ".cache" / "huggingface" / "datasets" / "AI-MO___aimo-validation-aime" / "default" / "0.0.0"
+            if cache_path.exists():
+                print(f"Using cached dataset from {cache_path}")
+                ds = datasets.load_dataset("AI-MO/aimo-validation-aime", split=self.split, cache_dir=str(cache_path))
+            else:
+                ds = datasets.load_dataset("AI-MO/aimo-validation-aime", split=self.split)
+        elif self.dataset_type == "aime2025":
+            print(f"Loading AIME2025 dataset...")
+            ds_list = []
+            for config_name in ["AIME2025-I", "AIME2025-II"]:
+                cache_path = Path.home() / ".cache" / "huggingface" / "datasets" / "opencompass___AIME2025" / "default" / "0.0.0"
+                if cache_path.exists():
+                    print(f"Using cached dataset from {cache_path}")
+                    ds = datasets.load_dataset("opencompass/AIME2025", config_name, split="test", cache_dir=str(cache_path))
+                else:
+                    ds = datasets.load_dataset("opencompass/AIME2025", config_name, split="test")
+                ds_list.extend(ds)
+            ds = ds_list
         else:
-            ds = datasets.load_dataset("AI-MO/aimo-validation-aime", split=self.split)
+            raise ValueError(f"Unknown dataset type: {self.dataset_type}")
 
         self.questions = list(ds)
-        print(f"AIME dataset loaded: {len(self.questions)} questions")
+        print(f"{self.dataset_type} dataset loaded: {len(self.questions)} questions")
 
     def find_question(self, request_text: str) -> Optional[Dict]:
+        # Strip common template prefixes to get the actual question text
+        # Templates include things like "Solve the following math problem step by step..."
+        # The actual question usually follows a blank line or after the template instruction
+        cleaned = request_text
+        # Split on double newline and take the part that looks like the problem
+        parts = cleaned.split('\n\n')
+        if len(parts) > 1:
+            # Find the part that's longest (likely the actual problem text)
+            problem_parts = [p for p in parts if len(p.strip()) > 100]
+            if problem_parts:
+                cleaned = max(problem_parts, key=lambda x: len(x))
+
         best_match = None
         best_distance = -1
         best_index = -1
 
         for i, question in enumerate(self.questions):
-            question_text = question["problem"]
-            request_lower = request_text.lower()
+            question_text = self._get_question_text(question)
+            request_lower = cleaned.lower()
             question_lower = question_text.lower()
 
+            # Check if question text is contained in the cleaned request
+            if question_lower in request_lower or request_lower in question_lower:
+                debug_log(f"DEBUG: Found substring match at index {i}")
+                return question
+
             # Exact match
             if question_lower == request_lower:
                 debug_log(f"DEBUG: Found exact match at index {i}")
@@ -118,7 +154,7 @@ class AimeDataset:
             debug_log(f"DEBUG: Found best partial match at index {best_index} with distance {best_distance:.3f}")
             return best_match
 
-        debug_log(f"DEBUG: No matching question found for: {request_text[:100]}...")
+        debug_log(f"DEBUG: No matching question found for cleaned: {cleaned[:100]}...")
         return None
 
     def get_answer(self, question: Dict) -> str:
@@ -134,15 +170,16 @@ class Simulator:
         port: int = 8033,
         host: str = "localhost",
         success_rate: float = 0.8,
-        dataset_split: str = "train"
+        dataset_split: str = "train",
+        dataset_type: str = "aime"
     ):
         self.port = port
         self.host = host
         self.success_rate = success_rate
-        self.dataset = AimeDataset(dataset_split)
+        self.dataset = AimeDataset(dataset_split, dataset_type)
         self.eval_state = EvalState(
-            id="aime-2025",
-            tasks=["aime"],
+            id=dataset_type,
+            tasks=[dataset_type],
             task_states={},
             sampling_config={"temperature": 0, "max_tokens": 2048}
         )
@@ -159,6 +196,10 @@ class Simulator:
         else:
             response_text = self._generate_wrong_answer(question)
 
+        comp_tokens = random.randint(10000, 60000)
+        tps_gen = random.uniform(90.0, 110.0)
+        t_gen_ms = comp_tokens / tps_gen * 1000
+
         return {
             "id": f"chatcmpl-{int(time.time())}",
             "object": "chat.completion",
@@ -176,8 +217,12 @@ class Simulator:
             ],
             "usage": {
                 "prompt_tokens": 100,
-                "completion_tokens": 50,
-                "total_tokens": 150
+                "completion_tokens": comp_tokens,
+                "total_tokens": 100 + comp_tokens
+            },
+            "timings": {
+                "predicted_ms": t_gen_ms,
+                "predicted_per_second": tps_gen
             }
         }
 
@@ -218,6 +263,12 @@ class Simulator:
         return response
 
 class RequestHandler(BaseHTTPRequestHandler):
+    def do_GET(self):
+        if self.path == "/v1/models":
+            self._send_json({"data": [{"id": "llama", "object": "model"}]}, 200)
+            return
+        self._send_json({"error": "Not found"}, 404)
+
     def do_POST(self):
         if self.path != "/v1/chat/completions":
             self._send_json({"error": "Not found"}, 404)
@@ -280,6 +331,13 @@ def main():
         default=0.8,
         help="Success rate 0-1 (default: 0.8)"
     )
+    parser.add_argument(
+        "--dataset",
+        type=str,
+        default="aime",
+        choices=["aime", "aime2025"],
+        help="Dataset type (default: aime)"
+    )
     parser.add_argument(
         "--dataset-split",
         type=str,
@@ -294,7 +352,8 @@ def main():
         port=args.port,
         host=args.host,
         success_rate=args.success_rate,
-        dataset_split=args.dataset_split
+        dataset_split=args.dataset_split,
+        dataset_type=args.dataset
     )
 
     server = HTTPServer((args.host, args.port), RequestHandler)
@@ -304,7 +363,7 @@ def main():
     print("\n=== llama-server-simulator ===")
     print(f"Server running on http://{args.host}:{args.port}")
     print(f"Success rate: {args.success_rate}")
-    print(f"AIME dataset loaded: {len(simulator.dataset.questions)} questions")
+    print(f"{args.dataset} dataset loaded: {len(simulator.dataset.questions)} questions")
     print("\nPress Ctrl+C to stop\n")
 
     try:

examples/llama.android/lib/build.gradle.kts

@@ -25,6 +25,7 @@ android {
                 arguments += "-DCMAKE_VERBOSE_MAKEFILE=ON"
 
                 arguments += "-DBUILD_SHARED_LIBS=ON"
+                arguments += "-DLLAMA_BUILD_APP=OFF"
                 arguments += "-DLLAMA_BUILD_COMMON=ON"
                 arguments += "-DLLAMA_OPENSSL=OFF"
 

examples/model-conversion/scripts/embedding/run-original-model.py

@@ -64,7 +64,7 @@ def load_model_and_tokenizer(model_path, use_sentence_transformers=False, device
         print("Using SentenceTransformer to apply all numbered layers")
         model = SentenceTransformer(model_path)
         tokenizer = model.tokenizer
-        config = model[0].auto_model.config
+        config = model[0].auto_model.config  # ty: ignore[unresolved-attribute]
     else:
         tokenizer = AutoTokenizer.from_pretrained(model_path)
         config = AutoConfig.from_pretrained(model_path, trust_remote_code=True)

examples/save-load-state/CMakeLists.txt

@@ -1,5 +0,0 @@
-set(TARGET llama-save-load-state)
-add_executable(${TARGET} save-load-state.cpp)
-install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE llama-common llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_17)

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

@@ -1,320 +0,0 @@
-#include "arg.h"
-#include "common.h"
-#include "llama.h"
-
-#include <clocale>
-#include <vector>
-#include <cstdio>
-
-
-int main(int argc, char ** argv) {
-    std::setlocale(LC_NUMERIC, "C");
-
-    common_params params;
-
-    params.prompt = "The quick brown fox";
-    params.sampling.seed = 1234;
-
-    const std::string_view state_file = "dump_state.bin";
-
-    common_init();
-
-    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) {
-        return 1;
-    }
-
-    if (params.n_parallel == 1) {
-        // the example uses 2 sequences, so when n_parallel == 1, we need to enable unified kv cache
-        printf("%s: n_parallel == 1, enabling unified kv cache\n", __func__);
-        params.kv_unified = true;
-    }
-
-    if (params.n_predict < 0) {
-        params.n_predict = 16;
-    }
-
-    auto n_past = 0;
-
-    std::string result0;
-    std::string result1;
-    std::string result2;
-    std::string result3;
-
-    // init
-
-    ggml_backend_load_all();
-
-    auto llama_init = common_init_from_params(params);
-
-    auto * model = llama_init->model();
-    auto * ctx   = llama_init->context();
-
-    if (model == nullptr || ctx == nullptr) {
-        fprintf(stderr, "%s : failed to init\n", __func__);
-        return 1;
-    }
-
-    auto sparams = llama_sampler_chain_default_params();
-
-    llama_sampler * smpl = llama_sampler_chain_init(sparams);
-
-    llama_sampler_chain_add(smpl, llama_sampler_init_dist(params.sampling.seed));
-
-    // tokenize prompt
-    auto tokens = common_tokenize(ctx, params.prompt, true);
-
-    const bool save_state = true;
-    if (!common_prompt_batch_decode(ctx, tokens, n_past, params.n_batch, state_file, save_state)) {
-        return 1;
-    }
-
-    // first run
-    printf("\nfirst run: %s", params.prompt.c_str());
-
-    llama_batch batch = llama_batch_init(1, 0, 1);
-
-    for (auto i = 0; i < params.n_predict; i++) {
-        auto next_token     = llama_sampler_sample(smpl, ctx, -1);
-        auto next_token_str = common_token_to_piece(ctx, next_token);
-
-        printf("%s", next_token_str.c_str());
-        result0 += next_token_str;
-
-        common_batch_clear(batch);
-        common_batch_add(batch, next_token, n_past, {0}, true);
-
-        if (llama_decode(ctx, batch)) {
-            fprintf(stderr, "\n%s : failed to evaluate\n", __func__);
-            llama_batch_free(batch);
-            return 1;
-        }
-        n_past += 1;
-    }
-
-    printf("\n\n");
-
-    // make new context
-    llama_context * ctx2 = llama_init_from_model(model, common_context_params_to_llama(params));
-
-    llama_sampler * smpl2 = llama_sampler_chain_init(sparams);
-
-    llama_sampler_chain_add(smpl2, llama_sampler_init_dist(params.sampling.seed));
-
-    printf("\nsecond run: %s", params.prompt.c_str());
-
-    // load state from file
-    std::vector<llama_token> unused_sts(tokens.size()); // unused session tokens.
-    size_t n_token_count_out = 0;
-
-    if (!llama_state_load_file(ctx2, state_file.data(), unused_sts.data(), unused_sts.size(), &n_token_count_out)) {
-        fprintf(stderr, "\n%s : failed to load state\n", __func__);
-        return 1;
-    }
-
-    fprintf(stderr, "%s : loaded state with %zu tokens\n", __func__, n_token_count_out);
-
-    // restore state (last tokens)
-    n_past = n_token_count_out;
-    if (!common_replay_last_token(ctx2, tokens.back(), n_past)) {
-        return 1;
-    }
-    ++n_past;
-
-    // second run
-    for (auto i = 0; i < params.n_predict; i++) {
-        auto next_token     = llama_sampler_sample(smpl2, ctx2, -1);
-        auto next_token_str = common_token_to_piece(ctx2, next_token);
-
-        printf("%s", next_token_str.c_str());
-        result1 += next_token_str;
-
-        common_batch_clear(batch);
-        common_batch_add(batch, next_token, n_past, {0}, true);
-
-        if (llama_decode(ctx2, batch)) {
-            fprintf(stderr, "\n%s : failed to evaluate\n", __func__);
-            llama_batch_free(batch);
-            return 1;
-        }
-        n_past += 1;
-    }
-
-    printf("\n\n");
-
-    if (result0 != result1) {
-        fprintf(stderr, "\n%s : error : the 2 generations are different\n", __func__);
-        return 1;
-    }
-
-    // make new context
-    auto params_ctx3 = common_context_params_to_llama(params);
-    params_ctx3.n_seq_max = 2;
-    llama_context * ctx3 = llama_init_from_model(model, params_ctx3);
-
-    llama_sampler * smpl3 = llama_sampler_chain_init(sparams);
-
-    llama_sampler_chain_add(smpl3, llama_sampler_init_dist(params.sampling.seed));
-
-    printf("\nsingle seq run: %s", params.prompt.c_str());
-
-    // load state (rng, logits, embedding and kv_cache) from file
-    n_token_count_out = 0;
-
-    if (!llama_state_load_file(ctx3, state_file.data(), unused_sts.data(), unused_sts.size(), &n_token_count_out)) {
-        fprintf(stderr, "\n%s : failed to load state\n", __func__);
-        return 1;
-    }
-
-    fprintf(stderr, "%s : loaded state with %zu tokens\n", __func__, n_token_count_out);
-
-    // restore state (last tokens)
-    n_past = n_token_count_out;
-    if (!common_replay_last_token(ctx3, tokens.back(), n_past)) {
-        return 1;
-    }
-    ++n_past;
-
-    // save seq 0 and load into seq 1
-    {
-        // save kv of seq 0
-        std::vector<uint8_t> seq_store(llama_state_seq_get_size(ctx3, 0));
-        const size_t ncopy = llama_state_seq_get_data(ctx3, seq_store.data(), seq_store.size(), 0);
-        if (ncopy != seq_store.size()) {
-            fprintf(stderr, "\n%s : seq copy data length %zd does not match expected length %zd\n", __func__, ncopy, seq_store.size());
-            return 1;
-        }
-        fprintf(stderr, "%s : seq 0 copied, %zd bytes\n", __func__, ncopy);
-
-        // erase whole kv
-        llama_memory_clear(llama_get_memory(ctx3), true);
-        fprintf(stderr, "%s : kv cache cleared\n", __func__);
-
-        // restore kv into seq 1
-        const size_t nset = llama_state_seq_set_data(ctx3, seq_store.data(), seq_store.size(), 1);
-        if (nset != seq_store.size()) {
-            fprintf(stderr, "\n%s : seq set data length %zd does not match expected length %zd\n", __func__, nset, seq_store.size());
-            return 1;
-        }
-        fprintf(stderr, "%s : seq 1 restored, %zd bytes\n", __func__, nset);
-    }
-
-    // third run with seq 1 instead of 0
-    for (auto i = 0; i < params.n_predict; i++) {
-        auto next_token     = llama_sampler_sample(smpl3, ctx3, -1);
-        auto next_token_str = common_token_to_piece(ctx3, next_token);
-
-        printf("%s", next_token_str.c_str());
-        result2 += next_token_str;
-
-        common_batch_clear(batch);
-        common_batch_add(batch, next_token, n_past, {1}, true);
-
-        if (llama_decode(ctx3, batch)) {
-            fprintf(stderr, "\n%s : failed to evaluate\n", __func__);
-            llama_batch_free(batch);
-            return 1;
-        }
-        n_past += 1;
-    }
-
-    // test on-device state save/load
-    auto params_ctx4 = common_context_params_to_llama(params);
-    params_ctx4.n_seq_max = 2;
-    llama_context * ctx4 = llama_init_from_model(model, params_ctx4);
-
-    llama_sampler * smpl4 = llama_sampler_chain_init(sparams);
-
-    llama_sampler_chain_add(smpl4, llama_sampler_init_dist(params.sampling.seed));
-
-    printf("\nsingle seq run: %s", params.prompt.c_str());
-
-    // load state (rng, logits, embedding and kv_cache) from file
-    n_token_count_out = 0;
-
-    if (!llama_state_load_file(ctx4, state_file.data(), unused_sts.data(), unused_sts.size(), &n_token_count_out)) {
-        fprintf(stderr, "\n%s : failed to load state\n", __func__);
-        return 1;
-    }
-
-    fprintf(stderr, "%s : loaded state with %zu tokens\n", __func__, n_token_count_out);
-
-    // restore state (last tokens)
-    n_past = n_token_count_out;
-    if (!common_replay_last_token(ctx4, tokens.back(), n_past)) {
-        return 1;
-    }
-    ++n_past;
-
-    // save seq 0 and load into seq 1
-    {
-        // save kv of seq 0
-        std::vector<uint8_t> seq_store(llama_state_seq_get_size_ext(ctx4, 0, LLAMA_STATE_SEQ_FLAGS_ON_DEVICE));
-        const size_t ncopy = llama_state_seq_get_data_ext(ctx4, seq_store.data(), seq_store.size(), 0, LLAMA_STATE_SEQ_FLAGS_ON_DEVICE);
-        if (ncopy != seq_store.size()) {
-            fprintf(stderr, "\n%s : seq copy data length %zd does not match expected length %zd\n", __func__, ncopy, seq_store.size());
-            return 1;
-        }
-        fprintf(stderr, "%s : seq 0 copied, %zd bytes\n", __func__, ncopy);
-
-        // erase whole kv
-        llama_memory_clear(llama_get_memory(ctx4), true);
-        fprintf(stderr, "%s : kv cache cleared\n", __func__);
-
-        // restore kv into seq 0
-        const size_t nset = llama_state_seq_set_data_ext(ctx4, seq_store.data(), seq_store.size(), 1, LLAMA_STATE_SEQ_FLAGS_ON_DEVICE);
-        if (nset != seq_store.size()) {
-            fprintf(stderr, "\n%s : seq set data length %zd does not match expected length %zd\n", __func__, nset, seq_store.size());
-            return 1;
-        }
-        fprintf(stderr, "%s : seq 1 restored, %zd bytes\n", __func__, nset);
-    }
-
-    // forth run
-    for (auto i = 0; i < params.n_predict; i++) {
-        auto next_token     = llama_sampler_sample(smpl4, ctx4, -1);
-        auto next_token_str = common_token_to_piece(ctx4, next_token);
-
-        printf("%s", next_token_str.c_str());
-        result3 += next_token_str;
-
-        common_batch_clear(batch);
-        common_batch_add(batch, next_token, n_past, {1}, true);
-
-        if (llama_decode(ctx4, batch)) {
-            fprintf(stderr, "\n%s : failed to evaluate\n", __func__);
-            llama_batch_free(batch);
-            return 1;
-        }
-        n_past += 1;
-    }
-
-    printf("\n");
-
-    llama_sampler_free(smpl);
-    llama_sampler_free(smpl2);
-    llama_sampler_free(smpl3);
-    llama_sampler_free(smpl4);
-
-    llama_batch_free(batch);
-
-    // this one is managed by common_init_result
-    //llama_free(ctx);
-
-    llama_free(ctx2);
-    llama_free(ctx3);
-    llama_free(ctx4);
-
-    if (result0 != result2) {
-        fprintf(stderr, "\n%s : error : the seq restore generation is different\n", __func__);
-        return 1;
-    }
-
-    if (result0 != result3) {
-        fprintf(stderr, "\n%s : error : the seq restore generation is different\n", __func__);
-        return 1;
-    }
-
-    fprintf(stderr, "\n%s : success\n", __func__);
-
-    return 0;
-}

examples/sycl/start-svr.sh

@@ -111,7 +111,6 @@ if [ $GGML_SYCL_DEVICE -ne -1 ]; then
     echo "Use $GGML_SYCL_DEVICE as main GPU"
     #use signle GPU only
     GPUS_SETTING="-mg $GGML_SYCL_DEVICE -sm ${SPLIT_MODE}"
-    export ONEAPI_DEVICE_SELECTOR="level_zero:${GGML_SYCL_DEVICE}"
     echo "ONEAPI_DEVICE_SELECTOR=${ONEAPI_DEVICE_SELECTOR}"
 else
     echo "Use all Intel GPUs, including iGPU & dGPU"

examples/sycl/test.sh

@@ -119,7 +119,6 @@ if [ $GGML_SYCL_DEVICE -ne -1 ]; then
     echo "Use $GGML_SYCL_DEVICE as main GPU"
     #use signle GPU only
     GPUS_SETTING="-mg $GGML_SYCL_DEVICE -sm ${SPLIT_MODE}"
-    export ONEAPI_DEVICE_SELECTOR="level_zero:${GGML_SYCL_DEVICE}"
     echo "ONEAPI_DEVICE_SELECTOR=${ONEAPI_DEVICE_SELECTOR}"
 else
     echo "Use all Intel GPUs, including iGPU & dGPU"

examples/sycl/win-start-svr.bat

@@ -164,7 +164,6 @@ if not "%GGML_SYCL_DEVICE%"=="-1" (
   echo Use %GGML_SYCL_DEVICE% as main GPU
   REM Use single GPU only.
   set "GPUS_SETTING=-mg %GGML_SYCL_DEVICE% -sm %SPLIT_MODE%"
-  set "ONEAPI_DEVICE_SELECTOR=level_zero:%GGML_SYCL_DEVICE%"
   echo ONEAPI_DEVICE_SELECTOR=%ONEAPI_DEVICE_SELECTOR%
 ) else (
   echo Use all Intel GPUs, including iGPU ^& dGPU

examples/sycl/win-test.bat

@@ -186,7 +186,6 @@ if not "%GGML_SYCL_DEVICE%"=="-1" (
   echo Use %GGML_SYCL_DEVICE% as main GPU
   REM Use single GPU only.
   set "GPUS_SETTING=-mg %GGML_SYCL_DEVICE% -sm %SPLIT_MODE%"
-  set "ONEAPI_DEVICE_SELECTOR=level_zero:%GGML_SYCL_DEVICE%"
   echo ONEAPI_DEVICE_SELECTOR=%ONEAPI_DEVICE_SELECTOR%
 ) else (
   echo Use all Intel GPUs, including iGPU ^& dGPU

ggml/src/ggml-backend.cpp

@@ -306,7 +306,7 @@ void ggml_backend_tensor_get_2d_async(ggml_backend_t backend, const struct ggml_
     GGML_ASSERT(tensor);
     GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
 
-    if (n_copies <= 1 || backend->iface.set_tensor_2d_async == NULL) {
+    if (n_copies <= 1 || backend->iface.get_tensor_2d_async == NULL) {
         for (size_t i = 0; i < n_copies; i++) {
             ggml_backend_tensor_get_async(backend, tensor, (char *) data + i*stride_data, offset + i*stride_tensor, size);
         }
@@ -317,7 +317,7 @@ void ggml_backend_tensor_get_2d_async(ggml_backend_t backend, const struct ggml_
     }
 
     GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
-    GGML_ASSERT(offset + (n_copies-1)*stride_tensor + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
+    GGML_ASSERT(offset + (n_copies-1)*stride_tensor + size <= ggml_nbytes(tensor) && "tensor read out of bounds");
     backend->iface.get_tensor_2d_async(backend, tensor, data, offset, size, n_copies, stride_tensor, stride_data);
 }
 
@@ -379,7 +379,7 @@ void ggml_backend_tensor_get_2d(const struct ggml_tensor * tensor, void * data,
     ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
     GGML_ASSERT(buf != NULL && "tensor buffer not set");
 
-    if (n_copies <= 1 || buf->iface.set_tensor_2d == NULL) {
+    if (n_copies <= 1 || buf->iface.get_tensor_2d == NULL) {
         for (size_t i = 0; i < n_copies; i++) {
             ggml_backend_tensor_get(tensor, (char *) data + i*stride_data, offset + i*stride_tensor, size);
         }

ggml/src/ggml-cuda/CMakeLists.txt

@@ -15,6 +15,7 @@ if (CUDAToolkit_FOUND)
         # 80     == Ampere, asynchronous data loading, faster tensor core instructions
         # 86     == RTX 3000, needs CUDA v11.1
         # 89     == RTX 4000, needs CUDA v11.8
+        # 90     == Hopper H100/200, needs CUDA v11.8
         # 120    == Blackwell, needs CUDA v12.8, FP4 tensor cores
         #
         # XX-virtual == compile CUDA code as PTX, do JIT compilation to binary code on first run
@@ -33,7 +34,7 @@ if (CUDAToolkit_FOUND)
             list(APPEND CMAKE_CUDA_ARCHITECTURES 75-virtual 80-virtual 86-real)
 
             if (CUDAToolkit_VERSION VERSION_GREATER_EQUAL "11.8")
-                list(APPEND CMAKE_CUDA_ARCHITECTURES 89-real)
+                list(APPEND CMAKE_CUDA_ARCHITECTURES 89-real 90-virtual)
             endif()
 
             if (CUDAToolkit_VERSION VERSION_GREATER_EQUAL "12.8")

ggml/src/ggml-cuda/binbcast.cu

@@ -2,6 +2,9 @@
 #include <cstdint>
 #include <utility>
 
+template<typename T, size_t>
+using type_for_index = T;
+
 static __device__ __forceinline__ float op_repeat(const float a, const float b) {
     return b;
     GGML_UNUSED(a);
@@ -52,6 +55,7 @@ static __global__ void k_bin_bcast(const src0_t *         src0,
                                    const int              s12,
                                    const int              s13,
                                    src1_ptrs... src1s) {
+    ggml_cuda_pdl_lc();
     const uint32_t i0s = blockDim.x * blockIdx.x + threadIdx.x;
     const uint32_t i1  = (blockDim.y * blockIdx.y + threadIdx.y);
     const uint32_t i2  = fastdiv((blockDim.z * blockIdx.z + threadIdx.z), ne3);
@@ -72,6 +76,7 @@ static __global__ void k_bin_bcast(const src0_t *         src0,
     const src0_t * src0_row = src0 ? (src0 + i_src0) : nullptr;
     dst_t * dst_row = dst + i_dst;
 
+    ggml_cuda_pdl_sync();
     for (int i0 = i0s; i0 < ne0; i0 += blockDim.x * gridDim.x) {
         const uint32_t i10 = fastmodulo(i0, ne10);
 
@@ -141,6 +146,7 @@ static __global__ void k_bin_bcast_unravel(const src0_t *         src0,
 
     const int i10 = fastmodulo(i0, ne10);
 
+    ggml_cuda_pdl_sync();
     float result = src0_row ? (float) src0_row[i0*s00] : 0.0f;
     if constexpr (sizeof...(src1_ptrs) > 0) {
         result = (..., (result = bin_op(result, (float)src1s[i_src1 + i10*s10])));
@@ -282,35 +288,24 @@ static void launch_bin_bcast_pack(const ggml_tensor * src0, const ggml_tensor *
             const uint3 ne1_fastdiv = init_fastdiv_values((uint32_t) ne1);
             const uint3 ne2_fastdiv = init_fastdiv_values((uint32_t) ne2);
 
-            if constexpr (sizeof...(I) > 0) {
-                k_bin_bcast_unravel<bin_op, src0_t, src1_t, dst_t><<<block_num, block_size, 0, stream>>>(
+            {
+                const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params((dim3)block_num, block_size, 0, stream);
+                ggml_cuda_kernel_launch(k_bin_bcast_unravel<bin_op, src0_t, src1_t, dst_t, type_for_index<const src1_t *, I>...>, launch_params,
                     src0_dd, src1_dd, dst_dd, ne0_fastdiv, ne1_fastdiv, ne2_fastdiv, ne3, prod_012, prod_01, ne10, ne11,
                     ne12, ne13,
                   /*s0,*/ s1,  s2,  s3,
                     s00, s01, s02, s03,
                     s10, s11, s12, s13, (const src1_t *) dst->src[I + 1]->data...);
-            } else {
-                k_bin_bcast_unravel<bin_op, src0_t, src1_t, dst_t>
-                    <<<block_num, block_size, 0, stream>>>(src0_dd, src1_dd, dst_dd, ne0_fastdiv, ne1_fastdiv,
-                                                           ne2_fastdiv, ne3, prod_012, prod_01, ne10, ne11, ne12, ne13,
-                                                         /*s0,*/ s1,  s2,  s3,
-                                                           s00, s01, s02, s03,
-                                                           s10, s11, s12, s13);
             }
         } else {
             const uint3 ne3_fastdiv = init_fastdiv_values((uint32_t) ne3);
-            if constexpr (sizeof...(I) > 0) {
-                k_bin_bcast<bin_op, src0_t, src1_t, dst_t><<<block_nums, block_dims, 0, stream>>>(
+            {
+                const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(block_nums, block_dims, 0, stream);
+                ggml_cuda_kernel_launch(k_bin_bcast<bin_op, src0_t, src1_t, dst_t, type_for_index<const src1_t *, I>...>, launch_params,
                     src0_dd, src1_dd, dst_dd, ne0, ne1, ne2, ne3_fastdiv, ne10, ne11, ne12, ne13,
                   /*s0,*/ s1, s2,  s3,
-                    s00 ,s01, s02, s03,
-                    s10, s11, s12, s13, (const src1_t *) dst->src[I + 1]->data...);
-            } else {
-                k_bin_bcast<bin_op, src0_t, src1_t, dst_t><<<block_nums, block_dims, 0, stream>>>(
-                    src0_dd, src1_dd, dst_dd, ne0, ne1, ne2, ne3_fastdiv, ne10, ne11, ne12, ne13,
-                  /*s0,*/ s1,  s2,  s3,
                     s00, s01, s02, s03,
-                    s10, s11, s12, s13);
+                    s10, s11, s12, s13, (const src1_t *) dst->src[I + 1]->data...);
             }
         }
     }
@@ -333,6 +328,7 @@ static __global__ void k_repeat_back(
     }
 
     T sum = 0;
+    ggml_cuda_pdl_sync();
     for (int64_t i3 = tid3; i3 < ne03; i3 += ne3) {
         for (int64_t i2 = tid2; i2 < ne02; i2 += ne2) {
             for (int64_t i1 = tid1; i1 < ne01; i1 += ne1) {

ggml/src/ggml-cuda/common.cuh

@@ -5,6 +5,7 @@
 #include "ggml-cuda.h"
 
 #include <cstdint>
+#include <cstdlib>
 #include <memory>
 
 #if defined(GGML_USE_HIP)
@@ -27,6 +28,7 @@
 #include <cstdio>
 #include <string>
 #include <unordered_map>
+#include <utility>
 #include <vector>
 
 #if defined(GGML_USE_HIP)
@@ -50,6 +52,7 @@
 #define GGML_CUDA_CC_TURING          750
 #define GGML_CUDA_CC_AMPERE          800
 #define GGML_CUDA_CC_ADA_LOVELACE    890
+#define GGML_CUDA_CC_HOPPER          900
 // While BW spans CC 1000, 1100 & 1200, we are integrating Tensor Core instructions available to 1200 family, see
 // https://docs.nvidia.com/cutlass/media/docs/cpp/blackwell_functionality.html#blackwell-sm120-gemms
 #define GGML_CUDA_CC_BLACKWELL       1200
@@ -107,6 +110,24 @@
 #    define GGML_CUDA_USE_CUB
 #endif  // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11070
 
+// PDL host-side support (cudaLaunchKernelEx) requires CUDART >= 11.8 and excludes HIP/MUSA.
+// __CUDA_ARCH__  is undefined in host passes; GPU arch check happens in device-side code.
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11080
+#    define GGML_CUDA_USE_PDL
+#endif  // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11080
+
+static __device__ __forceinline__ void ggml_cuda_pdl_sync() {
+#if defined(GGML_CUDA_USE_PDL) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= GGML_CUDA_CC_HOPPER
+    cudaGridDependencySynchronize();
+#endif // defined(GGML_CUDA_USE_PDL) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= GGML_CUDA_CC_HOPPER
+}
+
+static __device__ __forceinline__ void ggml_cuda_pdl_lc() {
+#if defined(GGML_CUDA_USE_PDL) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= GGML_CUDA_CC_HOPPER
+    cudaTriggerProgrammaticLaunchCompletion();
+#endif // defined(GGML_CUDA_USE_PDL) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= GGML_CUDA_CC_HOPPER
+}
+
 #ifdef __CUDA_ARCH_LIST__
 constexpr bool ggml_cuda_has_arch_impl(int) {
     return false;
@@ -165,6 +186,7 @@ void ggml_cuda_error(const char * stmt, const char * func, const char * file, in
 
 #define CUDA_CHECK(err) CUDA_CHECK_GEN(err, cudaSuccess, cudaGetErrorString)
 
+
 #if CUDART_VERSION >= 12000 || defined(GGML_USE_MUSA)
     static const char * cublas_get_error_str(const cublasStatus_t err) {
         return cublasGetStatusString(err);
@@ -1487,3 +1509,68 @@ struct ggml_cuda_mm_fusion_args_device {
     const void * gate_bias = nullptr;
     ggml_glu_op glu_op;
 };
+
+struct ggml_cuda_kernel_launch_params {
+    dim3 block_nums;
+    dim3 block_dims;
+    size_t shmem;
+    cudaStream_t stream;
+
+    // size_t shmem
+    ggml_cuda_kernel_launch_params(const dim3& block_nums_, const dim3& block_dims_, const size_t shmem_, const cudaStream_t stream_)
+        : block_nums(block_nums_), block_dims(block_dims_), shmem(shmem_), stream(stream_) {}
+
+    // Some call sites pass ints instead of the required size_t. This 2nd constructor casts int->size_t to avoid these -Wnarrowing warnings.
+    ggml_cuda_kernel_launch_params(const dim3& block_nums_, const dim3& block_dims_, const int shmem_, const cudaStream_t stream_)
+        : block_nums(block_nums_), block_dims(block_dims_), shmem((size_t)shmem_), stream(stream_) {}
+};
+
+#if defined(GGML_CUDA_USE_PDL)
+struct ggml_cuda_pdl_config {
+    cudaLaunchAttribute attr;
+    cudaLaunchConfig_t  cfg;
+
+    ggml_cuda_pdl_config(const ggml_cuda_kernel_launch_params & params) {
+        attr.id = cudaLaunchAttributeProgrammaticStreamSerialization;
+        attr.val.programmaticStreamSerializationAllowed = 1;
+
+        cfg = {};
+        cfg.gridDim          = params.block_nums;
+        cfg.blockDim         = params.block_dims;
+        cfg.dynamicSmemBytes = params.shmem;
+        cfg.stream           = params.stream;
+        cfg.attrs            = &attr;
+        cfg.numAttrs         = 1;
+    }
+
+    // Delete due to &attr
+    ggml_cuda_pdl_config(const ggml_cuda_pdl_config&) = delete;
+    ggml_cuda_pdl_config& operator=(const ggml_cuda_pdl_config&) = delete;
+    ggml_cuda_pdl_config& operator=(ggml_cuda_pdl_config&&) = delete;
+
+};
+#endif //defined(GGML_CUDA_USE_PDL)
+
+
+template<typename Kernel, typename... Args>
+static __inline__ void ggml_cuda_kernel_launch(Kernel kernel, const ggml_cuda_kernel_launch_params & launch_params, Args&&... args) {
+#if defined(GGML_CUDA_USE_PDL)
+
+    static const bool env_pdl_enabled = []() {
+        const char * env = getenv("GGML_CUDA_PDL");
+        return env == nullptr || std::atoi(env) != 0;
+    }();
+
+    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
+    if (env_pdl_enabled && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_HOPPER) {
+        auto pdl_cfg = ggml_cuda_pdl_config(launch_params);
+
+        CUDA_CHECK(cudaLaunchKernelEx(&pdl_cfg.cfg, kernel, std::forward<Args>(args)... ));
+        return;
+    }
+#endif //defined(GGML_CUDA_USE_PDL)
+
+    kernel<<<launch_params.block_nums, launch_params.block_dims, launch_params.shmem, launch_params.stream>>>(std::forward<Args>(args)... );
+    CUDA_CHECK(cudaGetLastError());
+}
+

ggml/src/ggml-cuda/concat.cu

@@ -15,6 +15,7 @@ static __global__ void __launch_bounds__(CUDA_CONCAT_BLOCK_SIZE) concat_f32_cont
 
     const int64_t n = ne0 * ne1 * ne2;
 
+    ggml_cuda_pdl_sync();
     for (int64_t i = (int64_t) blockIdx.x * blockDim.x + threadIdx.x; i < n; i += (int64_t) blockDim.x * gridDim.x) {
         if constexpr (dim == 0) {
             const int64_t row = i / ne0;
@@ -64,8 +65,8 @@ static void concat_f32_cuda(const float * x,
     const int     num_blocks = (n + CUDA_CONCAT_BLOCK_SIZE - 1) / CUDA_CONCAT_BLOCK_SIZE;
 
     if (dim == 0) {
-        concat_f32_cont<0>
-            <<<num_blocks, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(x, y, dst, ne00, ne01, ne02, ne0, ne1, ne2);
+        const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(num_blocks, CUDA_CONCAT_BLOCK_SIZE, 0, stream);
+        ggml_cuda_kernel_launch(concat_f32_cont<0>, launch_params,x, y, dst, ne00, ne01, ne02, ne0, ne1, ne2);
         return;
     }
     if (dim == 1) {

ggml/src/ggml-cuda/cpy.cu

@@ -16,6 +16,7 @@ static __global__ void cpy_scalar(const char * cx, char * cdst, const int64_t ne
                                   const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t nb00, const int64_t nb01, const int64_t nb02,
                                   const int64_t nb03, const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t nb10, const int64_t nb11,
                                   const int64_t nb12, const int64_t nb13) {
+    ggml_cuda_pdl_lc();
     const int64_t i = (int64_t)blockDim.x*blockIdx.x + threadIdx.x;
 
     if (i >= ne) {
@@ -36,6 +37,7 @@ static __global__ void cpy_scalar(const char * cx, char * cdst, const int64_t ne
     const int64_t i10 = i - i13*ne10*ne11*ne12 - i12*ne10*ne11 - i11*ne10;
     const int64_t dst_offset = i10*nb10 + i11*nb11 + i12*nb12 + i13 * nb13;
 
+    ggml_cuda_pdl_sync();
     cpy_1(cx + x_offset, cdst + dst_offset);
 }
 
@@ -59,6 +61,7 @@ static __global__ void cpy_scalar_transpose(const char * cx, char * cdst, const
     __shared__ float tile[2][CUDA_CPY_TILE_DIM_2D][CUDA_CPY_TILE_DIM_2D+1];
     int cur_tile_buf = 0;
 
+    ggml_cuda_pdl_sync();
 #pragma unroll
     for (int i = 0; i < CUDA_CPY_BLOCK_NM; ++i) {
 
@@ -142,6 +145,7 @@ static __global__ void cpy_f32_q(const char * cx, char * cdst, const int64_t ne,
     const int64_t i10 = i - i13*ne10*ne11*ne12 - i12*ne10*ne11 - i11*ne10;
     const int64_t dst_offset = (i10/qk)*nb10 + i11*nb11 + i12*nb12 + i13*nb13;
 
+    ggml_cuda_pdl_sync();
     cpy_blck(cx + x_offset, cdst + dst_offset);
 }
 
@@ -168,6 +172,7 @@ static __global__ void cpy_q_f32(const char * cx, char * cdst, const int64_t ne,
     const int64_t i10 = i - i13*ne10*ne11*ne12 - i12*ne10*ne11 - i11*ne10;
     const int64_t dst_offset = i10*nb10 + i11*nb11 + i12*nb12 + i13*nb13;
 
+    ggml_cuda_pdl_sync();
     cpy_blck(cx + x_offset, cdst + dst_offset);
 }
 
@@ -182,6 +187,7 @@ static __global__ void cpy_scalar_contiguous(const char * cx, char * cdst, const
     const src_t * x = (const src_t *) cx;
     dst_t *     dst = (dst_t *) cdst;
 
+    ggml_cuda_pdl_sync();
     dst[i] = ggml_cuda_cast<dst_t>(x[i]);
 }
 
@@ -192,8 +198,8 @@ cudaStream_t stream) {
 
     const int64_t num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
     GGML_ASSERT(num_blocks < UINT_MAX);
-    cpy_scalar_contiguous<src_t, dst_t><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
-        (cx, cdst, ne);
+    const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params((dim3)num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream);
+    ggml_cuda_kernel_launch(cpy_scalar_contiguous<src_t, dst_t>, launch_params, cx, cdst, ne);
 }
 
 template<typename src_t, typename dst_t, bool transposed = false>
@@ -223,13 +229,15 @@ static void ggml_cpy_scalar_cuda(
         GGML_ASSERT(grid_z < USHRT_MAX);
         dim3 dimGrid(grid_x, grid_y, grid_z);
         dim3 dimBlock(CUDA_CPY_TILE_DIM_2D, CUDA_CPY_BLOCK_ROWS, 1);
-        cpy_scalar_transpose<dst_t><<<dimGrid, dimBlock, 0, stream>>>
-            (cx, cdst, ne, ne00n, ne01n, ne02n, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+        const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(dimGrid, dimBlock, 0, stream);
+        ggml_cuda_kernel_launch(cpy_scalar_transpose<dst_t>, launch_params,
+            cx, cdst, ne, ne00n, ne01n, ne02n, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
     } else {
         const int64_t num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
         GGML_ASSERT(num_blocks < UINT_MAX);
-        cpy_scalar<cpy_1_scalar<src_t, dst_t>><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
-            (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+        const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params((dim3)num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream);
+        ggml_cuda_kernel_launch(cpy_scalar<cpy_1_scalar<src_t, dst_t>>, launch_params,
+            cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
     }
 }
 

ggml/src/ggml-cuda/fattn-common.cuh

@@ -636,6 +636,7 @@ static __global__ void flash_attn_mask_to_KV_max(
     if (tid < WARP_SIZE) {
         buf_iw[tid] = 1;
     }
+    ggml_cuda_pdl_sync();
     __syncthreads();
 
     int KV_max_sj = (ne30 - 1) * FATTN_KQ_STRIDE;
@@ -687,6 +688,7 @@ static __global__ void flash_attn_stream_k_fixup_uniform(
         const uint3 fd_iter_j_z,
         const uint3 fd_iter_j) {
     constexpr int ncols = ncols1*ncols2;
+    ggml_cuda_pdl_lc();
 
     const int tile_idx = blockIdx.x; // One block per output tile.
     const int j        = blockIdx.y;
@@ -718,6 +720,7 @@ static __global__ void flash_attn_stream_k_fixup_uniform(
 
     dst += sequence*ne02*ne01*D + jt*ne02*(ncols1*D) + zt_Q*D + (j*ne02 + c)*D + tid;
 
+    ggml_cuda_pdl_sync();
     // Load the partial result that needs a fixup
     float dst_val = *dst;
     float max_val;
@@ -809,6 +812,7 @@ static __global__ void flash_attn_stream_k_fixup_general(
     float dst_val = 0.0f;
     float max_val = 0.0f;
     float rowsum  = 0.0f;
+    ggml_cuda_pdl_sync();
     {
         dst_val = *dst;
 
@@ -867,6 +871,7 @@ static __global__ void flash_attn_combine_results(
         const float2 * __restrict__ VKQ_meta,
         float * __restrict__ dst,
         const int parallel_blocks) {
+    ggml_cuda_pdl_lc();
     // Dimension 0: threadIdx.x
     // Dimension 1: blockIdx.x
     // Dimension 2: blockIdx.y
@@ -890,6 +895,7 @@ static __global__ void flash_attn_combine_results(
     __builtin_assume(tid < D);
 
     extern __shared__ float2 meta[];
+    ggml_cuda_pdl_sync();
     for (int i = tid; i < 2*parallel_blocks; i += D) {
         ((float *) meta)[i] = ((const float *)VKQ_meta) [i];
     }
@@ -1146,7 +1152,9 @@ void launch_fattn(
     const uint3 ne01 = init_fastdiv_values(Q->ne[1]);
 
     GGML_ASSERT(block_dim.x % warp_size == 0);
-    fattn_kernel<<<blocks_num, block_dim, nbytes_shared, main_stream>>>(
+
+    const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(blocks_num, block_dim, nbytes_shared, main_stream);
+    ggml_cuda_kernel_launch(fattn_kernel, launch_params,
         (const char *) Q->data,
         K_data,
         V_data,
@@ -1176,9 +1184,9 @@ void launch_fattn(
             const dim3 block_dim_combine(DV, 1, 1);
             const dim3 blocks_num_combine = {(unsigned)ntiles_dst, ncols1, ncols2};
 
-            flash_attn_stream_k_fixup_uniform<DV, ncols1, ncols2>
-                <<<blocks_num_combine, block_dim_combine, 0, main_stream>>>
-                ((float *) KQV->data, dst_tmp_meta.ptr,
+            const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(blocks_num_combine, block_dim_combine, 0, main_stream);
+            ggml_cuda_kernel_launch(flash_attn_stream_k_fixup_uniform<DV, ncols1, ncols2>, launch_params,
+                (float *) KQV->data, dst_tmp_meta.ptr,
                  Q->ne[1], Q->ne[2], K->ne[2], nblocks_sk,
                  gqa_ratio, bpt, fd0, fd1, fd2);
         } else if (ntiles_dst % blocks_num.x != 0) {
@@ -1193,9 +1201,9 @@ void launch_fattn(
             const dim3 block_dim_combine(DV, 1, 1);
             const dim3 blocks_num_combine = {blocks_num.x, ncols1, ncols2};
 
-            flash_attn_stream_k_fixup_general<DV, ncols1, ncols2>
-                <<<blocks_num_combine, block_dim_combine, 0, main_stream>>>
-                ((float *) KQV->data, dst_tmp_meta.ptr,
+            const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(blocks_num_combine, block_dim_combine, 0, main_stream);
+            ggml_cuda_kernel_launch(flash_attn_stream_k_fixup_general<DV, ncols1, ncols2>, launch_params,
+                (float *) KQV->data, dst_tmp_meta.ptr,
                  Q->ne[1], Q->ne[2], gqa_ratio, total_work,
                  fd_k_j_z_ne12, fd_k_j_z, fd_k_j, fd_k);
         }
@@ -1204,9 +1212,9 @@ void launch_fattn(
         const dim3 blocks_num_combine(Q->ne[1], Q->ne[2], Q->ne[3]);
         const size_t nbytes_shared_combine = parallel_blocks*sizeof(float2);
 
-        flash_attn_combine_results<DV>
-            <<<blocks_num_combine, block_dim_combine, nbytes_shared_combine, main_stream>>>
-            (dst_tmp.ptr, dst_tmp_meta.ptr, (float *) KQV->data, parallel_blocks);
+        const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(blocks_num_combine, block_dim_combine, nbytes_shared_combine, main_stream);
+        ggml_cuda_kernel_launch(flash_attn_combine_results<DV>, launch_params,
+            dst_tmp.ptr, dst_tmp_meta.ptr, (float *) KQV->data, parallel_blocks);
     }
     CUDA_CHECK(cudaGetLastError());
 }

ggml/src/ggml-cuda/fattn-mma-f16.cuh

@@ -1724,6 +1724,7 @@ static __global__ void flash_attn_ext_f16(
                             const int32_t nb21, const int32_t nb22, const int64_t nb23,
                             const int32_t ne31, const int32_t ne32, const int32_t ne33,
                             const int32_t nb31, const int32_t nb32, const int64_t nb33) {
+    ggml_cuda_pdl_sync(); // TODO optimize placement
 #if defined(FLASH_ATTN_AVAILABLE) && (defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) || defined(AMD_MFMA_AVAILABLE))
 
     // Skip unused kernel variants for faster compilation:

ggml/src/ggml-cuda/fattn-tile.cuh

@@ -894,6 +894,8 @@ static __global__ void flash_attn_tile(
     }
     float KQ_sum[cpw] = {0.0f};
 
+    ggml_cuda_pdl_sync();
+
     // Load Q data, convert to FP16 if fast:
 #pragma unroll
     for (int jc0 = 0; jc0 < cpw; ++jc0) {

ggml/src/ggml-cuda/fattn-vec.cuh

@@ -40,6 +40,7 @@ static __global__ void flash_attn_ext_vec(
                             const int32_t nb21, const int32_t nb22, const int64_t nb23,
                             const int32_t ne31, const int32_t ne32, const int32_t ne33,
                             const int32_t nb31, const int32_t nb32, const int64_t nb33) {
+    ggml_cuda_pdl_lc();
 #ifdef FLASH_ATTN_AVAILABLE
 
     // Skip unused kernel variants for faster compilation:
@@ -136,6 +137,8 @@ static __global__ void flash_attn_ext_vec(
 #endif // V_DOT2_F32_F16_AVAILABLE
     int    Q_i32[ncols][1 > D/(sizeof(int)*nthreads_KQ) ? 1 : D/(sizeof(int)*nthreads_KQ)];
     float2  Q_ds[ncols][1 > D/(sizeof(int)*nthreads_KQ) ? 1 : D/(sizeof(int)*nthreads_KQ)];
+
+    ggml_cuda_pdl_sync();
     if constexpr (Q_q8_1) {
 #pragma unroll
         for (int j0 = 0; j0 < ncols; j0 += nwarps) {

ggml/src/ggml-cuda/fattn-wmma-f16.cu

@@ -86,6 +86,7 @@ static __global__ void flash_attn_ext_f16(
     constexpr int kqs_padded = FATTN_KQ_STRIDE + 8;
     constexpr int kqar = sizeof(KQ_acc_t)/sizeof(half);
 
+    ggml_cuda_pdl_sync();
     const int sequence = blockIdx.z / ne02;
     const int head = blockIdx.z - sequence*ne02;
     const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.

ggml/src/ggml-cuda/gated_delta_net.cu

@@ -1,4 +1,5 @@
 #include "gated_delta_net.cuh"
+#include "ggml-cuda/common.cuh"
 
 template <int S_v, bool KDA, bool keep_rs_t>
 __global__ void __launch_bounds__((ggml_cuda_get_physical_warp_size() < S_v ? ggml_cuda_get_physical_warp_size() : S_v) * 4, 2)
@@ -53,6 +54,7 @@ gated_delta_net_cuda(const float * q,
     float         s_shard[rows_per_lane];
     // state is stored transposed: M[col][i] = S[i][col], row col is contiguous
 
+    ggml_cuda_pdl_sync();
 #pragma unroll
     for (int r = 0; r < rows_per_lane; r++) {
         const int i = r * warp_size + lane;
@@ -189,28 +191,29 @@ static void launch_gated_delta_net(
 
     int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
 
+    const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(grid_dims, block_dims, 0, stream);
     switch (S_v) {
         case 16:
-            gated_delta_net_cuda<16, KDA, keep_rs_t><<<grid_dims, block_dims, 0, stream>>>(
+            ggml_cuda_kernel_launch(gated_delta_net_cuda<16, KDA, keep_rs_t>, launch_params,
                 q_d, k_d, v_d, g_d, b_d, s_d, dst_d, H,
                 n_tokens, n_seqs, sq1, sq2, sq3, sv1, sv2, sv3,
                 sb1, sb2, sb3, neqk1_magic, rq3_magic, scale, K);
             break;
         case 32:
-            gated_delta_net_cuda<32, KDA, keep_rs_t><<<grid_dims, block_dims, 0, stream>>>(
+            ggml_cuda_kernel_launch(gated_delta_net_cuda<32, KDA, keep_rs_t>, launch_params,
                 q_d, k_d, v_d, g_d, b_d, s_d, dst_d, H,
                 n_tokens, n_seqs, sq1, sq2, sq3, sv1, sv2, sv3,
                 sb1, sb2, sb3, neqk1_magic, rq3_magic, scale, K);
             break;
         case 64: {
-            gated_delta_net_cuda<64, KDA, keep_rs_t><<<grid_dims, block_dims, 0, stream>>>(
+            ggml_cuda_kernel_launch(gated_delta_net_cuda<64, KDA, keep_rs_t>, launch_params,
                 q_d, k_d, v_d, g_d, b_d, s_d, dst_d, H,
                 n_tokens, n_seqs, sq1, sq2, sq3, sv1, sv2, sv3,
                 sb1, sb2, sb3, neqk1_magic, rq3_magic, scale, K);
             break;
         }
         case 128: {
-            gated_delta_net_cuda<128, KDA, keep_rs_t><<<grid_dims, block_dims, 0, stream>>>(
+            ggml_cuda_kernel_launch(gated_delta_net_cuda<128, KDA, keep_rs_t>, launch_params,
                 q_d, k_d, v_d, g_d, b_d, s_d, dst_d, H,
                 n_tokens, n_seqs, sq1, sq2, sq3, sv1, sv2, sv3,
                 sb1, sb2, sb3, neqk1_magic, rq3_magic, scale, K);

ggml/src/ggml-cuda/getrows.cu

@@ -11,6 +11,7 @@ static __global__ void k_get_rows(
         /*const size_t nb00,*/ const size_t nb01, const size_t nb02, const size_t nb03,
         const size_t s10, const size_t s11, const size_t s12/*, const size_t s13*/) {
 
+    ggml_cuda_pdl_sync();
     for (int64_t z = blockIdx.z; z < ne11*(int64_t)ne12_fdv.z; z += gridDim.z) {
         for (int64_t i00 = 2*(blockIdx.y*blockDim.x + threadIdx.x); i00 < ne00; i00 += gridDim.y*blockDim.x) {
             // The x and y dimensions of the grid are swapped because the maximum allowed grid size for x is higher.
@@ -48,6 +49,8 @@ static __global__ void k_get_rows_float(
         /*const size_t nb00,*/ const size_t nb01, const size_t nb02, const size_t nb03,
         const size_t s10, const size_t s11, const size_t s12/*, const size_t s13*/) {
 
+    ggml_cuda_pdl_lc();
+    ggml_cuda_pdl_sync();
     for (int64_t z = blockIdx.z; z < ne11*(int64_t)ne12_fdv.z; z += gridDim.z) {
         for (int64_t i00 = blockIdx.y*blockDim.x + threadIdx.x; i00 < ne00; i00 += gridDim.y*blockDim.x) {
             // The x and y dimensions of the grid are swapped because the maximum allowed grid size for x is higher.
@@ -83,6 +86,7 @@ static __global__ void k_get_rows_back_float(
 
     float sum = 0.0f;
 
+    ggml_cuda_pdl_sync();
     for (int64_t i = 0; i < nrows_grad; ++i) {
         if (rows[i] != dst_row) {
             continue;
@@ -156,7 +160,8 @@ static void get_rows_cuda_float(
     GGML_ASSERT(ne11 <= std::numeric_limits<uint32_t>::max() / ne12);
     const uint3 ne12_fdv = init_fastdiv_values(ne12);
 
-    k_get_rows_float<<<block_nums, block_dims, 0, stream>>>(
+    const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params{block_nums, block_dims, 0, stream};
+    ggml_cuda_kernel_launch(k_get_rows_float<src0_t, dst_t>, launch_params,
         src0_d, src1_d, dst_d,
         ne00, /*ne01, ne02, ne03,*/
         /*ne10,*/ ne11, ne12_fdv, /*ne13,*/

ggml/src/ggml-cuda/mean.cu

@@ -67,9 +67,11 @@ void ggml_cuda_op_mean(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     // See discussion in: https://github.com/ggml-org/llama.cpp/pull/15132
     if ((nrows / nsm) < 2) {
         const dim3 block_dims(512, 1, 1);
-        reduce_rows_f32</*norm=*/true><<<block_nums, block_dims, 0, stream>>>(src0_d, dst_d, ncols);
+        const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(block_nums, block_dims, 0, stream);
+        ggml_cuda_kernel_launch(reduce_rows_f32</*norm=*/true>, launch_params, src0_d, dst_d, ncols);
     } else {
         const dim3 block_dims(ncols < 1024 ? 32 : 128, 1, 1);
-        reduce_rows_f32</*norm=*/true><<<block_nums, block_dims, 0, stream>>>(src0_d, dst_d, ncols);
+        const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(block_nums, block_dims, 0, stream);
+        ggml_cuda_kernel_launch(reduce_rows_f32</*norm=*/true>, launch_params, src0_d, dst_d, ncols);
     }
 }

ggml/src/ggml-cuda/mmvf.cu

@@ -21,6 +21,7 @@ static __global__ void mul_mat_vec_f(
     int channel_y;
     int sample_dst;
 
+    ggml_cuda_pdl_sync();
     if constexpr (is_multi_token_id) {
         // Multi-token MUL_MAT_ID path, adding these in the normal path causes a perf regression for n_tokens=1 case
         token_idx  = blockIdx.z;
@@ -298,6 +299,7 @@ static __global__ void mul_mat_vec_f(
         static_assert(std::is_same_v<T, void>, "unsupported type");
     }
 
+    ggml_cuda_pdl_lc();
 #pragma unroll
     for (int j = 0; j < ncols_dst; ++j) {
         sumf[j] = warp_reduce_sum<warp_size>(sumf[j]);
@@ -382,11 +384,13 @@ static void mul_mat_vec_f_switch_fusion(
         const uint3 sample_ratio, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst,
         const dim3 & block_dims, const dim3 & block_nums, const int nbytes_shared, const int ids_stride, const cudaStream_t stream) {
 
+    const ggml_cuda_kernel_launch_params launch_params = {block_nums, block_dims, nbytes_shared, stream};
+
     const bool has_fusion = fusion.gate != nullptr || fusion.x_bias != nullptr || fusion.gate_bias != nullptr;
     if constexpr (ncols_dst == 1) {
         if (has_fusion) {
-            mul_mat_vec_f<T, type_acc, ncols_dst, block_size, true, is_multi_token_id><<<block_nums, block_dims, nbytes_shared, stream>>>
-                (x, y, ids, fusion, dst, ncols, nchannels_y, stride_row, stride_col_y, stride_col_dst,
+            ggml_cuda_kernel_launch(mul_mat_vec_f<T, type_acc, ncols_dst, block_size, true, is_multi_token_id>, launch_params,
+                x, y, ids, fusion, dst, ncols, nchannels_y, stride_row, stride_col_y, stride_col_dst,
                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride);
             return;
@@ -395,8 +399,8 @@ static void mul_mat_vec_f_switch_fusion(
 
     GGML_ASSERT(!has_fusion && "fusion only supported for ncols_dst=1");
 
-    mul_mat_vec_f<T, type_acc, ncols_dst, block_size, false, is_multi_token_id><<<block_nums, block_dims, nbytes_shared, stream>>>
-        (x, y, ids, fusion, dst, ncols, nchannels_y, stride_row, stride_col_y, stride_col_dst,
+    ggml_cuda_kernel_launch(mul_mat_vec_f<T, type_acc, ncols_dst, block_size, false, is_multi_token_id>, launch_params,
+        x, y, ids, fusion, dst, ncols, nchannels_y, stride_row, stride_col_y, stride_col_dst,
         channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
         sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride);
 

ggml/src/ggml-cuda/mmvq.cu

@@ -359,7 +359,9 @@ static constexpr __host__ __device__ int calc_nwarps(ggml_type type, int ncols_d
                 case GGML_TYPE_Q5_1:
                 case GGML_TYPE_Q8_0:
                 case GGML_TYPE_Q4_K:
+                    return 8;
                 case GGML_TYPE_Q6_K:
+                    return 2;
                 case GGML_TYPE_IQ4_NL:
                     return 8;
                 default:
@@ -422,6 +424,7 @@ static __global__ void mul_mat_vec_q(
     uint32_t channel_y;
     uint32_t sample_dst;
 
+    ggml_cuda_pdl_sync();
     channel_x  = ncols_dst == 1 && ids ? ids[channel_dst]                     : fastdiv(channel_dst, channel_ratio);
     channel_y  = ncols_dst == 1 && ids ? fastmodulo(channel_dst, nchannels_y) : channel_dst;
     sample_dst = blockIdx.z;
@@ -681,8 +684,9 @@ static void mul_mat_vec_q_switch_fusion(
     const bool has_fusion = fusion.gate != nullptr || fusion.x_bias != nullptr || fusion.gate_bias != nullptr;
     if constexpr (c_ncols_dst == 1) {
         if (has_fusion) {
-            mul_mat_vec_q<type, c_ncols_dst, true, small_k><<<block_nums, block_dims, nbytes_shared, stream>>>
-                (vx, vy, ids, fusion, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
+            const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(block_nums, block_dims, nbytes_shared, stream);
+            ggml_cuda_kernel_launch(mul_mat_vec_q<type, c_ncols_dst, true, small_k>, launch_params,
+                 vx, vy, ids, fusion, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
                  sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride);
             return;
@@ -691,8 +695,9 @@ static void mul_mat_vec_q_switch_fusion(
 
     GGML_ASSERT(!has_fusion && "fusion only supported for ncols_dst=1");
 
-    mul_mat_vec_q<type, c_ncols_dst, false, small_k><<<block_nums, block_dims, nbytes_shared, stream>>>
-        (vx, vy, ids, fusion, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
+    const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(block_nums, block_dims, nbytes_shared, stream);
+    ggml_cuda_kernel_launch(mul_mat_vec_q<type, c_ncols_dst, false, small_k>, launch_params,
+        vx, vy, ids, fusion, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
         channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
         sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride);
 }

ggml/src/ggml-cuda/norm.cu

@@ -18,6 +18,7 @@ static __global__ void norm_f32(
 
     float2 mean_var = make_float2(0.0f, 0.0f);
 
+    ggml_cuda_pdl_sync();
     for (int col = tid; col < ncols; col += block_size) {
         const float xi = x[col];
         mean_var.x += xi;
@@ -46,6 +47,7 @@ static __global__ void group_norm_f32(const float * x, float * dst, const int gr
 
     float tmp = 0.0f; // partial sum for thread in warp
 
+    ggml_cuda_pdl_sync();
     for (int j = start; j < end; j += block_size) {
         tmp += x[j];
     }
@@ -95,6 +97,7 @@ static __global__ void rms_norm_f32(const float * x,
                                     const uint3   add_nrows_packed     = make_uint3(0, 0, 0),
                                     const uint3   add_nchannels_packed = make_uint3(0, 0, 0),
                                     const uint3   add_nsamples_packed  = make_uint3(0, 0, 0)) {
+    ggml_cuda_pdl_lc();
     const int nrows     = gridDim.x;
     const int nchannels = gridDim.y;
 
@@ -124,6 +127,7 @@ static __global__ void rms_norm_f32(const float * x,
 
     float tmp = 0.0f; // partial sum for thread in warp
 
+    ggml_cuda_pdl_sync();
     for (int col = tid; col < ncols; col += block_size) {
         const float xi = x[col];
         tmp += xi * xi;
@@ -163,6 +167,7 @@ static __global__ void rms_norm_back_f32(
     float sum_xx = 0.0f; // sum for squares of x, equivalent to forward pass
     float sum_xg = 0.0f; // sum for x * gradient, needed because RMS norm mixes inputs
 
+    ggml_cuda_pdl_sync();
     for (int col = tid; col < ncols; col += block_size) {
         const float xfi = xf[col];
         sum_xx += xfi * xfi;
@@ -253,6 +258,7 @@ static __global__ void l2_norm_f32(
 
     float tmp = 0.0f; // partial sum for thread in warp
 
+    ggml_cuda_pdl_sync();
     for (int col = tid; col < ncols; col += block_size) {
         const float xi = x[col];
         tmp += xi * xi;
@@ -261,6 +267,7 @@ static __global__ void l2_norm_f32(
     // sum up partial sums
     extern __shared__ float s_sum[];
     tmp = block_reduce<block_reduce_method::SUM, block_size>(tmp, s_sum);
+    ggml_cuda_pdl_lc();
 
     // from https://pytorch.org/docs/stable/generated/torch.nn.functional.normalize.html
     const float scale = rsqrtf(fmaxf(tmp, eps * eps));
@@ -300,10 +307,19 @@ static void rms_norm_f32_cuda(
     const dim3 blocks_num(nrows, nchannels, nsamples);
     if (ncols < 1024) {
         const dim3 block_dims(256, 1, 1);
-        rms_norm_f32<256, false><<<blocks_num, block_dims, block_dims.x > WARP_SIZE ? 32 * sizeof(float): 0, stream>>>(x, dst, ncols, stride_row, stride_channel, stride_sample, eps);
+        const ggml_cuda_kernel_launch_params launch_params = {blocks_num, block_dims, block_dims.x > WARP_SIZE ? 32 * sizeof(float): 0, stream};
+        ggml_cuda_kernel_launch(rms_norm_f32<256, false>, launch_params,
+            x, dst, ncols, stride_row, stride_channel, stride_sample, eps,
+        // underlying cudaLaunchKernelEx does not support default params
+        nullptr, 0, 0, 0, make_uint3(0, 0, 0), make_uint3(0, 0, 0), make_uint3(0, 0, 0), make_uint3(0, 0, 0),
+        nullptr, 0, 0, 0, make_uint3(0, 0, 0), make_uint3(0, 0, 0), make_uint3(0, 0, 0), make_uint3(0, 0, 0));
     } else {
         const dim3 block_dims(1024, 1, 1);
-        rms_norm_f32<1024, false><<<blocks_num, block_dims, block_dims.x > WARP_SIZE ? 32 * sizeof(float): 0, stream>>>(x, dst, ncols, stride_row, stride_channel, stride_sample, eps);
+        const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params{blocks_num, block_dims, block_dims.x > WARP_SIZE ? 32 * sizeof(float): 0, stream};
+        ggml_cuda_kernel_launch(rms_norm_f32<1024, false>, launch_params, x, dst, ncols, stride_row, stride_channel, stride_sample, eps,
+        // underlying cudaLaunchKernelEx does not support default params
+        nullptr, 0, 0, 0, make_uint3(0, 0, 0), make_uint3(0, 0, 0), make_uint3(0, 0, 0), make_uint3(0, 0, 0),
+        nullptr, 0, 0, 0, make_uint3(0, 0, 0), make_uint3(0, 0, 0), make_uint3(0, 0, 0), make_uint3(0, 0, 0));
     }
 }
 
@@ -346,14 +362,20 @@ static void rms_norm_mul_f32_cuda(const float *  x,
         const uint3 mul_nsamples_packed  = init_fastdiv_values(mul_nsamples);
         if (ncols < 1024) {
             const dim3 block_dims(256, 1, 1);
-            rms_norm_f32<256, true><<<blocks_num, block_dims, block_dims.x > WARP_SIZE ? 32 * sizeof(float): 0, stream>>>(
+            const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params{blocks_num, block_dims, block_dims.x > WARP_SIZE ? 32 * sizeof(float): 0, stream};
+            ggml_cuda_kernel_launch(rms_norm_f32<256, true>, launch_params,
                 x, dst, ncols, stride_row, stride_channel, stride_sample, eps, mul, mul_stride_row, mul_stride_channel,
-                mul_stride_sample, mul_ncols_packed, mul_nrows_packed, mul_nchannels_packed, mul_nsamples_packed);
+                mul_stride_sample, mul_ncols_packed, mul_nrows_packed, mul_nchannels_packed, mul_nsamples_packed,
+                // underlying cudaLaunchKernelEx does not support default params
+            nullptr, 0, 0, 0, make_uint3(0, 0, 0), make_uint3(0, 0, 0), make_uint3(0, 0, 0), make_uint3(0, 0, 0));
         } else {
             const dim3 block_dims(1024, 1, 1);
-            rms_norm_f32<1024, true><<<blocks_num, block_dims, block_dims.x > WARP_SIZE ? 32 * sizeof(float): 0, stream>>>(
+            const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params{blocks_num, block_dims, block_dims.x > WARP_SIZE ? 32 * sizeof(float): 0, stream};
+            ggml_cuda_kernel_launch(rms_norm_f32<1024, true>, launch_params,
                 x, dst, ncols, stride_row, stride_channel, stride_sample, eps, mul, mul_stride_row, mul_stride_channel,
-                mul_stride_sample, mul_ncols_packed, mul_nrows_packed, mul_nchannels_packed, mul_nsamples_packed);
+                mul_stride_sample, mul_ncols_packed, mul_nrows_packed, mul_nchannels_packed, mul_nsamples_packed,
+                // underlying cudaLaunchKernelEx does not support default params
+            nullptr, 0, 0, 0, make_uint3(0, 0, 0), make_uint3(0, 0, 0), make_uint3(0, 0, 0), make_uint3(0, 0, 0));
         }
     } else {
         const uint3 mul_ncols_packed     = init_fastdiv_values(mul_ncols);
@@ -367,14 +389,16 @@ static void rms_norm_mul_f32_cuda(const float *  x,
         const uint3 add_nsamples_packed  = init_fastdiv_values(add_nsamples);
         if (ncols < 1024) {
             const dim3 block_dims(256, 1, 1);
-            rms_norm_f32<256, true, true><<<blocks_num, block_dims, block_dims.x > WARP_SIZE ? 32 * sizeof(float): 0, stream>>>(
+            const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params{blocks_num, block_dims,block_dims.x > WARP_SIZE ? 32 * sizeof(float): 0, stream};
+            ggml_cuda_kernel_launch(rms_norm_f32<256, true, true>, launch_params,
                 x, dst, ncols, stride_row, stride_channel, stride_sample, eps, mul, mul_stride_row, mul_stride_channel,
                 mul_stride_sample, mul_ncols_packed, mul_nrows_packed, mul_nchannels_packed, mul_nsamples_packed, add,
                 add_stride_row, add_stride_channel, add_stride_sample, add_ncols_packed, add_nrows_packed,
                 add_nchannels_packed, add_nsamples_packed);
         } else {
             const dim3 block_dims(1024, 1, 1);
-            rms_norm_f32<1024, true, true><<<blocks_num, block_dims, block_dims.x > WARP_SIZE ? 32 * sizeof(float): 0, stream>>>(
+            const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params{blocks_num, block_dims, block_dims.x > WARP_SIZE ? 32 * sizeof(float): 0, stream};
+            ggml_cuda_kernel_launch(rms_norm_f32<1024, true, true>, launch_params,
                 x, dst, ncols, stride_row, stride_channel, stride_sample, eps, mul, mul_stride_row, mul_stride_channel,
                 mul_stride_sample, mul_ncols_packed, mul_nrows_packed, mul_nchannels_packed, mul_nsamples_packed, add,
                 add_stride_row, add_stride_channel, add_stride_sample, add_ncols_packed, add_nrows_packed,
@@ -399,10 +423,12 @@ static void l2_norm_f32_cuda(
     const dim3 blocks_num(nrows, nchannels, nsamples);
     if (ncols < 1024) {
         const dim3 block_dims(WARP_SIZE, 1, 1);
-        l2_norm_f32<WARP_SIZE><<<blocks_num, block_dims, 0, stream>>>(x, dst, ncols, stride_row, stride_channel, stride_sample, eps);
+        const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params{blocks_num, block_dims, 0, stream};
+        ggml_cuda_kernel_launch(l2_norm_f32<WARP_SIZE>, launch_params, x, dst, ncols, stride_row, stride_channel, stride_sample, eps);
     } else {
         const dim3 block_dims(1024, 1, 1);
-        l2_norm_f32<1024><<<blocks_num, block_dims, block_dims.x > WARP_SIZE ? 32 * sizeof(float): 0, stream>>>(x, dst, ncols, stride_row, stride_channel, stride_sample, eps);
+        const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params{blocks_num, block_dims, block_dims.x > WARP_SIZE ? 32 * sizeof(float): 0, stream};
+        ggml_cuda_kernel_launch(l2_norm_f32<1024>, launch_params, x, dst, ncols, stride_row, stride_channel, stride_sample, eps);
     }
 }
 

ggml/src/ggml-cuda/quantize.cu

@@ -6,6 +6,7 @@ static __global__ void quantize_q8_1(
         const float * __restrict__ x, void * __restrict__ vy,
         const int64_t ne00, const int64_t s01, const int64_t s02, const int64_t s03,
         const int64_t ne0, const uint32_t ne1, const uint3 ne2) {
+    ggml_cuda_pdl_lc();
     const int64_t i0 = (int64_t)blockDim.x*blockIdx.x + threadIdx.x;
 
     if (i0 >= ne0) {
@@ -28,6 +29,7 @@ static __global__ void quantize_q8_1(
     const int64_t ib  = i_cont / QK8_1; // block index
     const int64_t iqs = i_cont % QK8_1; // quant index
 
+    ggml_cuda_pdl_sync();
     const float xi = i0 < ne00 ? x[i03*s03 + i02*s02 + i01*s01 + i00] : 0.0f;
     float amax = fabsf(xi);
     float sum = xi;
@@ -196,6 +198,7 @@ static __global__ void quantize_mmq_mxfp4(const float * __restrict__ x,
     const int64_t i2 = blockIdx.z % ne2;
     const int64_t i3 = blockIdx.z / ne2;
 
+    ggml_cuda_pdl_sync();
     const int64_t i01 = ids ? ids[i1] : i1;
     const int64_t i02 = i2;
     const int64_t i03 = i3;
@@ -288,6 +291,7 @@ static __global__ void quantize_mmq_q8_1(
     const int64_t i3 = blockIdx.z / ne2;
 
     const int64_t i00 = i0;
+    ggml_cuda_pdl_sync();
     const int64_t i01 = ids ? ids[i1] : i1;
     const int64_t i02 = i2;
     const int64_t i03 = i3;
@@ -378,7 +382,8 @@ void quantize_row_q8_1_cuda(
     const int64_t block_num_x = (ne0 + CUDA_QUANTIZE_BLOCK_SIZE - 1) / CUDA_QUANTIZE_BLOCK_SIZE;
     const dim3 num_blocks(block_num_x, ne1, ne2*ne3);
     const dim3 block_size(CUDA_QUANTIZE_BLOCK_SIZE, 1, 1);
-    quantize_q8_1<<<num_blocks, block_size, 0, stream>>>(x, vy, ne00, s01, s02, s03, ne0, ne1, ne2_fastdiv);
+    const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(num_blocks, block_size, 0, stream);
+    ggml_cuda_kernel_launch(quantize_q8_1, launch_params, x, vy, ne00, s01, s02, s03, ne0, ne1, ne2_fastdiv);
     GGML_UNUSED(type_src0);
 }
 

ggml/src/ggml-cuda/reduce_rows.cuh

@@ -10,6 +10,8 @@ static __global__ void reduce_rows_f32(const float * __restrict__ x, float * __r
     const int num_unroll = 8;
     float     temp[num_unroll];
     float     sum_temp[num_unroll] = { 0.0f };
+
+    ggml_cuda_pdl_sync();
     for (int i = col; i < ncols;) {
         for (int j = 0; j < num_unroll; ++j) {
             if (i < ncols) {

ggml/src/ggml-cuda/rope.cu

@@ -134,6 +134,7 @@ static __global__ void rope_neox(const T *            x,
                                  const float *        freq_factors,
                                  const int64_t *      row_indices,
                                  const int            set_rows_stride) {
+    ggml_cuda_pdl_lc();
     const int i0 = 2*(blockDim.y*blockIdx.y + threadIdx.y);
 
     if (i0 >= ne00) {
@@ -148,6 +149,7 @@ static __global__ void rope_neox(const T *            x,
 
     int       idst = i0 / 2 + i1 * s1  + i2 * s2  + i3 * s3;
     const int ix   = i0 / 2 + i1 * s01 + i2 * s02 + i3 * s03;
+    ggml_cuda_pdl_sync();
 
     // Fusion optimization: ROPE + VIEW + SET_ROWS.
     // The rope output is viewed as a 1D tensor and offset based on a row index in row_indices.
@@ -216,6 +218,7 @@ static __global__ void rope_multi(const T *            x,
     int       idst = i0 / 2 + i1 * s1  + i2 * s2  + i3 * s3;
     const int ix   = i0 / 2 + i1 * s01 + i2 * s02 + i3 * s03;
 
+    ggml_cuda_pdl_sync();
     if (i0 >= n_dims) {
         dst[idst + i0/2 + 0] = x[ix + i0/2 + 0];
         dst[idst + i0/2 + 1] = x[ix + i0/2 + 1];
@@ -300,6 +303,7 @@ static __global__ void rope_vision(const T *            x,
     int       idst = i0 / 2 + i1 * s1  + i2 * s2  + i3 * s3;
     const int ix   = i0 / 2 + i1 * s01 + i2 * s02 + i3 * s03;
 
+    ggml_cuda_pdl_sync();
     const int sect_dims = sections.v[0] + sections.v[1];
     const int sec_w     = sections.v[1] + sections.v[0];
     const int sector    = (i0 / 2) % sect_dims;
@@ -399,13 +403,14 @@ static void rope_neox_cuda(const T *            x,
     const dim3 block_nums(nr, n_blocks_x, 1);
 
     const float theta_scale = powf(freq_base, -2.0f / n_dims);
+    const ggml_cuda_kernel_launch_params launch_params = {block_nums, block_dims, 0, stream};
 
     if (freq_factors == nullptr) {
-        rope_neox<forward, false><<<block_nums, block_dims, 0, stream>>>(
+        ggml_cuda_kernel_launch(rope_neox<forward, false, T, D>, launch_params,
             x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, pos, freq_scale, ext_factor,
             attn_factor, corr_dims, theta_scale, freq_factors, row_indices, set_rows_stride);
     } else {
-        rope_neox<forward, true><<<block_nums, block_dims, 0, stream>>>(
+        ggml_cuda_kernel_launch(rope_neox<forward, true, T, D>, launch_params,
             x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, pos, freq_scale, ext_factor,
             attn_factor, corr_dims, theta_scale, freq_factors, row_indices, set_rows_stride);
     }
@@ -443,11 +448,13 @@ static void rope_multi_cuda(const T *            x,
     const float theta_scale = powf(freq_base, -2.0f / n_dims);
 
     if (freq_factors == nullptr) {
-        rope_multi<forward, false, T><<<block_nums, block_dims, 0, stream>>>(
+        const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(block_nums, block_dims, 0, stream);
+        ggml_cuda_kernel_launch(rope_multi<forward, false, T>, launch_params,
             x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, pos, freq_scale, ext_factor,
             attn_factor, corr_dims, theta_scale, freq_factors, sections, is_imrope);
     } else {
-        rope_multi<forward, true, T><<<block_nums, block_dims, 0, stream>>>(
+        const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(block_nums, block_dims, 0, stream);
+        ggml_cuda_kernel_launch(rope_multi<forward, true, T>, launch_params,
             x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, pos, freq_scale, ext_factor,
             attn_factor, corr_dims, theta_scale, freq_factors, sections, is_imrope);
     }

ggml/src/ggml-cuda/scale.cu

@@ -3,9 +3,11 @@
 #define MAX_GRIDDIM_X 0x7FFFFFFF
 
 static __global__ void scale_f32(const float * x, float * dst, const float scale, const float bias, const int64_t nelements) {
+    ggml_cuda_pdl_lc();
     int64_t tid = (int64_t)blockIdx.x * (int64_t)blockDim.x + (int64_t)threadIdx.x;
     int64_t stride = (int64_t)blockDim.x * (int64_t)gridDim.x;
 
+    ggml_cuda_pdl_sync();
     for (int64_t i = tid; i < nelements; i += stride) {
         dst[i] = scale * x[i] + bias;
     }
@@ -13,7 +15,8 @@ static __global__ void scale_f32(const float * x, float * dst, const float scale
 
 static void scale_f32_cuda(const float * x, float * dst, const float scale, const float bias, const int64_t nelements, cudaStream_t stream) {
     const int64_t num_blocks = (nelements + CUDA_SCALE_BLOCK_SIZE - 1) / CUDA_SCALE_BLOCK_SIZE;
-    scale_f32<<<MIN(MAX_GRIDDIM_X, num_blocks), CUDA_SCALE_BLOCK_SIZE, 0, stream>>>(x, dst, scale, bias, nelements);
+    const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(MIN(MAX_GRIDDIM_X, num_blocks), CUDA_SCALE_BLOCK_SIZE, 0, stream);
+    ggml_cuda_kernel_launch(scale_f32, launch_params, x, dst, scale, bias, nelements);
 }
 
 void ggml_cuda_op_scale(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {

ggml/src/ggml-cuda/set-rows.cu

@@ -53,6 +53,7 @@ static __global__ void k_set_rows_quant(const float * __restrict__ src0,
     const int64_t i11 = fastmodulo((uint32_t) i02, ne11_fd);
     const int64_t i10 = i01;
 
+    ggml_cuda_pdl_sync();
     const int64_t dst_row = *(src1 + i10*s10 + i11*s11 + i12*s12);
 
     const float * src0_row = src0 + i01*s01 + i02*s02 + i03*s03;
@@ -157,7 +158,9 @@ static __global__ void k_set_rows(const src_t * __restrict__ src0,
     const int64_t i11 = fastmodulo((uint32_t) i02, ne11_fd);
     const int64_t i10 = i01;
 
+    ggml_cuda_pdl_sync();
     const int64_t dst_row = *(src1 + i10*s10 + i11*s11 + i12*s12);
+    ggml_cuda_pdl_lc();
 
     const src_t * src0_row = src0 + i01*s01 + i02*s02 + i03*s03;
     dst_t * dst_row_ptr    = dst + dst_row*s1 + i02*s2 + i03*s3;
@@ -203,9 +206,11 @@ static void set_rows_cuda(
         const uint3 ne11_fd = init_fastdiv_values((uint32_t) ne11);
         const uint3 ne12_fd = init_fastdiv_values((uint32_t) ne12);
 
-        k_set_rows<<<grid_size, block_size, 0, stream>>>(src0_d, src1_d, dst_d, ne_total, ne10, ne11, ne12, ne13, s01,
-                                                         s02, s03, s10, s11, s12, s1, s2, s3, ne00_fd, ne01_fd, ne02_fd,
-                                                         ne11_fd, ne12_fd);
+        const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(grid_size, block_size, 0, stream);
+        ggml_cuda_kernel_launch(k_set_rows<src_t, idx_t, dst_t>, launch_params,
+            src0_d, src1_d, dst_d, ne_total, ne10, ne11, ne12, ne13, s01,
+            s02, s03, s10, s11, s12, s1, s2, s3, ne00_fd, ne01_fd, ne02_fd,
+            ne11_fd, ne12_fd);
     }
 }
 

ggml/src/ggml-cuda/softcap.cu

@@ -1,18 +1,21 @@
 #include "softcap.cuh"
 
 static __global__ void softcap_f32(const float * x, float * dst, const float scale, const float softcap, const int k) {
+    ggml_cuda_pdl_lc();
     const int i = blockDim.x*blockIdx.x + threadIdx.x;
 
     if (i >= k) {
         return;
     }
 
+    ggml_cuda_pdl_sync();
     dst[i] = tanhf(scale * x[i]) * softcap;
 }
 
 static void softcap_f32_cuda(const float * x, float * dst, const float scale, const float softcap, const int k, cudaStream_t stream) {
     const int num_blocks = (k + CUDA_SOFTCAP_BLOCK_SIZE - 1) / CUDA_SOFTCAP_BLOCK_SIZE;
-    softcap_f32<<<num_blocks, CUDA_SOFTCAP_BLOCK_SIZE, 0, stream>>>(x, dst, scale, softcap, k);
+    const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(num_blocks, CUDA_SOFTCAP_BLOCK_SIZE, 0, stream);
+    ggml_cuda_kernel_launch(softcap_f32, launch_params, x, dst, scale, softcap, k);
 }
 
 // fused GGML_OP_SCALE + GGML_UNARY_OP_TANH + GGML_OP_SCALE

ggml/src/ggml-cuda/ssm-conv.cu

@@ -1,3 +1,4 @@
+#include "common.cuh"
 #include "ssm-conv.cuh"
 #include "unary.cuh"
 
@@ -7,6 +8,7 @@ static __global__ void ssm_conv_f32(const float * __restrict__ src0, const float
                                     const int src0_nb0, const int src0_nb1, const int src0_nb2, const int src1_nb1,
                                     float * __restrict__ dst, const int dst_nb0, const int dst_nb1, const int dst_nb2,
                                     const int64_t n_t) {
+    ggml_cuda_pdl_lc();
     GGML_UNUSED(src0_nb0);
     const int tid  = threadIdx.x;
     const int bidx = blockIdx.x;
@@ -23,6 +25,7 @@ static __global__ void ssm_conv_f32(const float * __restrict__ src0, const float
     float x[d_conv] = { 0.0f };
     float w[d_conv] = { 0.0f };
 
+    ggml_cuda_pdl_sync();
 #pragma unroll
     for (size_t j = 0; j < d_conv; j++) {
         w[j] = w_block[tid * stride_w + j];
@@ -128,8 +131,9 @@ static void ssm_conv_f32_cuda(const float * src0, const float * src1, const floa
         constexpr int kNC = decltype(NC)::value;
         if (n_t <= 32) {
             const dim3 blocks(n_s, (nr + threads - 1) / threads, 1);
-            ssm_conv_f32<apply_silu, threads, kNC><<<blocks, threads, 0, stream>>>(src0, src1, bias, src0_nb0, src0_nb1, src0_nb2, src1_nb1,
-                                                                       dst, dst_nb0, dst_nb1, dst_nb2, n_t);
+            const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(blocks, threads, 0, stream);
+            ggml_cuda_kernel_launch(ssm_conv_f32<apply_silu, threads, kNC>, launch_params, src0, src1, bias, src0_nb0, src0_nb1,
+                                                                        src0_nb2, src1_nb1, dst, dst_nb0, dst_nb1, dst_nb2, n_t);
         } else {
             const int64_t split_n_t = 32;
             dim3          blocks(n_s, (nr + threads - 1) / threads, (n_t + split_n_t - 1) / split_n_t);
@@ -140,11 +144,12 @@ static void ssm_conv_f32_cuda(const float * src0, const float * src1, const floa
     };
 
     switch (nc) {
-        case 3: launch_kernel(std::integral_constant<int, 3>{}); break;
-        case 4: launch_kernel(std::integral_constant<int, 4>{}); break;
-        case 5: launch_kernel(std::integral_constant<int, 5>{}); break;
-        case 9: launch_kernel(std::integral_constant<int, 9>{}); break;
-        default: GGML_ABORT("Only support kernel sizes 3, 4, 5, 9 right now.");
+        case 3:  launch_kernel(std::integral_constant<int, 3 >{}); break;
+        case 4:  launch_kernel(std::integral_constant<int, 4 >{}); break;
+        case 5:  launch_kernel(std::integral_constant<int, 5 >{}); break;
+        case 9:  launch_kernel(std::integral_constant<int, 9 >{}); break;
+        case 15: launch_kernel(std::integral_constant<int, 15>{}); break;
+        default: GGML_ABORT("Only support kernel sizes 3, 4, 5, 9, 15 right now.");
     }
 }
 

ggml/src/ggml-cuda/ssm-scan.cu

@@ -26,6 +26,7 @@ __global__ void __launch_bounds__(splitD, 1)
                  const int64_t s_off, const int64_t d_inner, const int64_t L_param)
 {
     const size_t L = L_template == 0 ? L_param : L_template;
+    ggml_cuda_pdl_sync();
     const float *s0_block = (const float *)((const char *)src0 + src6[blockIdx.x] * src0_nb3 + blockIdx.y * splitD * src0_nb2);
     const float *x_block = (const float *)((const char *)src1 + (blockIdx.x * src1_nb3) + blockIdx.y * splitD * sizeof(float));
     const float *dt_block = (const float *)((const char *)src2 + (blockIdx.x * src2_nb2) + blockIdx.y * splitD * sizeof(float));
@@ -135,6 +136,7 @@ __global__ void __launch_bounds__(d_state, 1)
 
     const int group_off = (head_idx / (n_head / n_group)) * d_state * sizeof(float);
 
+    ggml_cuda_pdl_sync();
     // TODO: refactor strides to be in elements/floats instead of bytes to be cleaner and consistent with the rest of the codebase
     const float * s0_warp = (const float *) ((const char *) src0 + src6[seq_idx] * src0_nb3 + head_idx * src0_nb2 + head_off * d_state);
     const float * x_warp  = (const float *) ((const char *) src1 + (seq_idx * src1_nb3) + (warp_idx * sizeof(float)));
@@ -206,7 +208,8 @@ static void ssm_scan_f32_cuda(const float * src0, const float * src1, const floa
             constexpr int num_warps = threads/WARP_SIZE;
 
             const dim3 blocks((n_head * head_dim + (num_warps - 1)) / num_warps, n_seq, 1);
-            ssm_scan_f32_group<128/WARP_SIZE, 128><<<blocks, threads, 0, stream>>>(
+            const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(blocks, threads, 0, stream);
+            ggml_cuda_kernel_launch(ssm_scan_f32_group<128/WARP_SIZE, 128>, launch_params,
                     src0, src1, src2, src3, src4, src5, src6, dst,
                     src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2, src3_nb1,
                     src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, head_dim, n_group, n_tok);
@@ -215,7 +218,8 @@ static void ssm_scan_f32_cuda(const float * src0, const float * src1, const floa
             constexpr int num_warps = threads/WARP_SIZE;
 
             const dim3 blocks((n_head * head_dim + (num_warps - 1)) / num_warps, n_seq, 1);
-            ssm_scan_f32_group<256/WARP_SIZE, 256><<<blocks, threads, 0, stream>>>(
+            const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(blocks, threads, 0, stream);
+            ggml_cuda_kernel_launch(ssm_scan_f32_group<256/WARP_SIZE, 256>, launch_params,
                     src0, src1, src2, src3, src4, src5, src6, dst,
                     src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2, src3_nb1,
                     src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, head_dim, n_group, n_tok);
@@ -231,58 +235,59 @@ static void ssm_scan_f32_cuda(const float * src0, const float * src1, const floa
         const dim3 blocks(n_seq, (n_head + threads - 1) / threads, 1);
         const int  smem_size = (threads * (d_state + 1) * 2) * sizeof(float);
         if (d_state == 16) {
+            const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(blocks, threads, smem_size, stream);
             switch (n_tok)
             {
             case 1:
-                ssm_scan_f32<threads, 16, 1><<<blocks, threads, smem_size, stream>>>(
+                ggml_cuda_kernel_launch(ssm_scan_f32<threads, 16, 1>, launch_params,
                     src0, src1, src2, src3, src4, src5, src6, dst,
                 src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2,
                 src3_nb1, src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, n_tok);
                 break;
             case 2:
-                ssm_scan_f32<threads, 16, 2><<<blocks, threads, smem_size, stream>>>(
+                ggml_cuda_kernel_launch(ssm_scan_f32<threads, 16, 2>, launch_params,
                     src0, src1, src2, src3, src4, src5, src6, dst,
                 src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2,
                 src3_nb1, src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, n_tok);
                 break;
             case 3:
-                ssm_scan_f32<threads, 16, 3><<<blocks, threads, smem_size, stream>>>(
+                ggml_cuda_kernel_launch(ssm_scan_f32<threads, 16, 3>, launch_params,
                     src0, src1, src2, src3, src4, src5, src6, dst,
                 src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2,
                 src3_nb1, src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, n_tok);
                 break;
             case 4:
-                ssm_scan_f32<threads, 16, 4><<<blocks, threads, smem_size, stream>>>(
+                ggml_cuda_kernel_launch(ssm_scan_f32<threads, 16, 4>, launch_params,
                     src0, src1, src2, src3, src4, src5, src6, dst,
                 src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2,
                 src3_nb1, src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, n_tok);
                 break;
             case 5:
-                ssm_scan_f32<threads, 16, 5><<<blocks, threads, smem_size, stream>>>(
+                ggml_cuda_kernel_launch(ssm_scan_f32<threads, 16, 5>, launch_params,
                     src0, src1, src2, src3, src4, src5, src6, dst,
                 src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2,
                 src3_nb1, src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, n_tok);
                 break;
             case 6:
-                ssm_scan_f32<threads, 16, 6><<<blocks, threads, smem_size, stream>>>(
+                ggml_cuda_kernel_launch(ssm_scan_f32<threads, 16, 6>, launch_params,
                     src0, src1, src2, src3, src4, src5, src6, dst,
                 src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2,
                 src3_nb1, src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, n_tok);
                 break;
             case 7:
-                ssm_scan_f32<threads, 16, 7><<<blocks, threads, smem_size, stream>>>(
+                ggml_cuda_kernel_launch(ssm_scan_f32<threads, 16, 7>, launch_params,
                     src0, src1, src2, src3, src4, src5, src6, dst,
                 src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2,
                 src3_nb1, src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, n_tok);
                 break;
             case 8:
-                ssm_scan_f32<threads, 16, 8><<<blocks, threads, smem_size, stream>>>(
+                ggml_cuda_kernel_launch(ssm_scan_f32<threads, 16, 8>, launch_params,
                     src0, src1, src2, src3, src4, src5, src6, dst,
                 src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2,
                 src3_nb1, src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, n_tok);
                 break;
             default:
-                ssm_scan_f32<threads, 16, 0><<<blocks, threads, smem_size, stream>>>(
+                ggml_cuda_kernel_launch(ssm_scan_f32<threads, 16, 0>, launch_params,
                     src0, src1, src2, src3, src4, src5, src6, dst,
                 src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2,
                 src3_nb1, src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, n_tok);

ggml/src/ggml-cuda/sumrows.cu

@@ -7,10 +7,12 @@ void sum_rows_f32_cuda(const float * x, float * dst, const int ncols, const int
     const dim3 block_nums(nrows, 1, 1);
     if ((nrows / nsm) < 2) {
         const dim3 block_dims(512, 1, 1);
-        reduce_rows_f32</*norm=*/false><<<block_nums, block_dims, 0, stream>>>(x, dst, ncols);
+        const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(block_nums, block_dims, 0, stream);
+        ggml_cuda_kernel_launch(reduce_rows_f32</*norm=*/false>, launch_params, x, dst, ncols);
     } else {
         const dim3 block_dims(ncols < 1024 ? 32 : 128, 1, 1);
-        reduce_rows_f32</*norm=*/false><<<block_nums, block_dims, 0, stream>>>(x, dst, ncols);
+        const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(block_nums, block_dims, 0, stream);
+        ggml_cuda_kernel_launch(reduce_rows_f32</*norm=*/false>, launch_params, x, dst, ncols);
     }
 }
 
@@ -34,10 +36,12 @@ void ggml_cuda_op_sum_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     if ((nrows / nsm) < 2) {
         // Increase num threads to 512 for small nrows to better hide the latency
         const dim3 block_dims(512, 1, 1);
-        reduce_rows_f32</*norm=*/false><<<block_nums, block_dims, 0, stream>>>(src0_d, dst_d, ncols);
+        const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(block_nums, block_dims, 0, stream);
+        ggml_cuda_kernel_launch(reduce_rows_f32</*norm=*/false>, launch_params, src0_d, dst_d, ncols);
     } else {
         // Enough active SMs to hide latency, use smaller blocks to allow better scheduling
         const dim3 block_dims(ncols < 1024 ? 32 : 128, 1, 1);
-        reduce_rows_f32</*norm=*/false><<<block_nums, block_dims, 0, stream>>>(src0_d, dst_d, ncols);
+        const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(block_nums, block_dims, 0, stream);
+        ggml_cuda_kernel_launch(reduce_rows_f32</*norm=*/false>, launch_params, src0_d, dst_d, ncols);
     }
 }

ggml/src/ggml-cuda/top-k.cu

@@ -5,6 +5,7 @@
 #    include <cub/cub.cuh>
 #    if (CCCL_MAJOR_VERSION >= 3 && CCCL_MINOR_VERSION >= 2)
 #        define CUB_TOP_K_AVAILABLE
+#        include <cuda/iterator>
 using namespace cub;
 #    endif  // CCCL_MAJOR_VERSION >= 3 && CCCL_MINOR_VERSION >= 2
 #endif      // GGML_CUDA_USE_CUB

ggml/src/ggml-cuda/topk-moe.cu

@@ -105,6 +105,7 @@ __launch_bounds__(4 * WARP_SIZE, 1) __global__ void topk_moe_cuda(const float *
         wt[i] = -INFINITY;
     }
 
+    ggml_cuda_pdl_sync();
 #pragma unroll
     for (int i = 0; i < n_experts; i += WARP_SIZE) {
         const int expert  = i + threadIdx.x;
@@ -161,6 +162,7 @@ __launch_bounds__(4 * WARP_SIZE, 1) __global__ void topk_moe_cuda(const float *
         output_weights[i] = 0.f;
     }
 
+    ggml_cuda_pdl_lc();
     for (int k = 0; k < n_expert_used; k++) {
         float max_val    = wt[0];
         int   max_expert = threadIdx.x;
@@ -271,51 +273,52 @@ static void launch_topk_moe_cuda(ggml_backend_cuda_context & ctx,
     dim3         grid_dims((n_rows + rows_per_block - 1) / rows_per_block, 1, 1);
     dim3         block_dims(WARP_SIZE, rows_per_block, 1);
     cudaStream_t stream = ctx.stream();
+    const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(grid_dims, block_dims, 0, stream);
 
     switch (n_expert) {
         case 1:
-            topk_moe_cuda<1, has_bias><<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, bias, n_rows, n_expert_used,
-                                                                   clamp_val, scale_val, config);
+            ggml_cuda_kernel_launch(topk_moe_cuda<1, has_bias>, launch_params,
+                logits, weights, ids, bias, n_rows, n_expert_used, clamp_val, scale_val, config);
             break;
         case 2:
-            topk_moe_cuda<2, has_bias><<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, bias, n_rows, n_expert_used,
-                                                                   clamp_val, scale_val, config);
+            ggml_cuda_kernel_launch(topk_moe_cuda<2, has_bias>, launch_params,
+                logits, weights, ids, bias, n_rows, n_expert_used, clamp_val, scale_val, config);
             break;
         case 4:
-            topk_moe_cuda<4, has_bias><<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, bias, n_rows, n_expert_used,
-                                                                   clamp_val, scale_val, config);
+            ggml_cuda_kernel_launch(topk_moe_cuda<4, has_bias>, launch_params,
+                logits, weights, ids, bias, n_rows, n_expert_used, clamp_val, scale_val, config);
             break;
         case 8:
-            topk_moe_cuda<8, has_bias><<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, bias, n_rows, n_expert_used,
-                                                                   clamp_val, scale_val, config);
+            ggml_cuda_kernel_launch(topk_moe_cuda<8, has_bias>, launch_params,
+                logits, weights, ids, bias, n_rows, n_expert_used, clamp_val, scale_val, config);
             break;
         case 16:
-            topk_moe_cuda<16, has_bias><<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, bias, n_rows, n_expert_used,
-                                                                    clamp_val, scale_val, config);
+            ggml_cuda_kernel_launch(topk_moe_cuda<16, has_bias>, launch_params,
+                logits, weights, ids, bias, n_rows, n_expert_used, clamp_val, scale_val, config);
             break;
         case 32:
-            topk_moe_cuda<32, has_bias><<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, bias, n_rows, n_expert_used,
-                                                                    clamp_val, scale_val, config);
+            ggml_cuda_kernel_launch(topk_moe_cuda<32, has_bias>, launch_params,
+                logits, weights, ids, bias, n_rows, n_expert_used, clamp_val, scale_val, config);
             break;
         case 64:
-            topk_moe_cuda<64, has_bias><<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, bias, n_rows, n_expert_used,
-                                                                    clamp_val, scale_val, config);
+            ggml_cuda_kernel_launch(topk_moe_cuda<64, has_bias>, launch_params,
+                logits, weights, ids, bias, n_rows, n_expert_used, clamp_val, scale_val, config);
             break;
         case 128:
-            topk_moe_cuda<128, has_bias><<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, bias, n_rows, n_expert_used,
-                                                                     clamp_val, scale_val, config);
+            ggml_cuda_kernel_launch(topk_moe_cuda<128, has_bias>, launch_params,
+                logits, weights, ids, bias, n_rows, n_expert_used, clamp_val, scale_val, config);
             break;
         case 256:
-            topk_moe_cuda<256, has_bias><<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, bias, n_rows, n_expert_used,
-                                                                     clamp_val, scale_val, config);
+            ggml_cuda_kernel_launch(topk_moe_cuda<256, has_bias>, launch_params,
+                logits, weights, ids, bias, n_rows, n_expert_used, clamp_val, scale_val, config);
             break;
         case 512:
-            topk_moe_cuda<512, has_bias><<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, bias, n_rows, n_expert_used,
-                                                                     clamp_val, scale_val, config);
+            ggml_cuda_kernel_launch(topk_moe_cuda<512, has_bias>, launch_params,
+                logits, weights, ids, bias, n_rows, n_expert_used, clamp_val, scale_val, config);
             break;
         case 576:
-            topk_moe_cuda<576, has_bias><<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, bias, n_rows, n_expert_used,
-                                                                     clamp_val, scale_val, config);
+            ggml_cuda_kernel_launch(topk_moe_cuda<576, has_bias>, launch_params,
+                logits, weights, ids, bias, n_rows, n_expert_used, clamp_val, scale_val, config);
             break;
         default:
             GGML_ASSERT(false && "fatal error");

ggml/src/ggml-cuda/unary.cu

@@ -116,19 +116,22 @@ static __device__ __forceinline__ float op_trunc(float x) {
 
 template <float (*op)(float), typename T>
 static __global__ void unary_op_kernel(const T * x, T * dst, const int k) {
+    ggml_cuda_pdl_lc();
     const int i = blockDim.x*blockIdx.x + threadIdx.x;
 
     if (i >= k) {
         return;
     }
 
+    ggml_cuda_pdl_sync();
     dst[i] = (T)op((float)x[i]);
 }
 
 template <float (*op)(float), typename T>
 static void unary_cuda(const T * x, T * dst, const int k, cudaStream_t stream) {
     const int num_blocks = (k + CUDA_NEG_BLOCK_SIZE - 1) / CUDA_NEG_BLOCK_SIZE;
-    unary_op_kernel<op><<<num_blocks, CUDA_NEG_BLOCK_SIZE, 0, stream>>>(x, dst, k);
+    const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params((dim3)num_blocks, CUDA_NEG_BLOCK_SIZE, 0, stream);
+    ggml_cuda_kernel_launch(unary_op_kernel<op, T>, launch_params, x, dst, k);
 }
 
 template <float (*op)(float)>
@@ -258,6 +261,7 @@ void ggml_cuda_op_softplus(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
 
 template <float (*op)(float), typename T>
 static __global__ void unary_gated_op_kernel(const T * x, const T * g, T * dst, const int64_t k, const int64_t n, const int64_t o0, const int64_t o1) {
+    ggml_cuda_pdl_lc();
     const int64_t i = int64_t(blockDim.x)*blockIdx.x + threadIdx.x;
 
     if (i >= k) {
@@ -268,13 +272,15 @@ static __global__ void unary_gated_op_kernel(const T * x, const T * g, T * dst,
     const int64_t j0 = (i / n) * o0 + (i % n);
     const int64_t j1 = o0 == o1 ? j0 : (i / n) * o1 + (i % n);
 
+    ggml_cuda_pdl_sync();
     dst[i] = (T)(op((float)x[j0]) * (float)g[j1]);
 }
 
 template <float (*op)(float), typename T>
 static void unary_gated_cuda(const T * x, const T * g, T * dst, const int64_t k, const int64_t n, const int64_t o0, const int64_t o1, cudaStream_t stream) {
     const int64_t num_blocks = (k + CUDA_GLU_BLOCK_SIZE - 1) / CUDA_GLU_BLOCK_SIZE;
-    unary_gated_op_kernel<op><<<num_blocks, CUDA_GLU_BLOCK_SIZE, 0, stream>>>(x, g, dst, k, n, o0, o1);
+    const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params((dim3)num_blocks, CUDA_GLU_BLOCK_SIZE, 0, stream);
+    ggml_cuda_kernel_launch(unary_gated_op_kernel<op, T>, launch_params, x, g, dst, k, n, o0, o1);
 }
 
 template <float (*op)(float)>

ggml/src/ggml-hexagon/ggml-hexagon.cpp

@@ -2661,7 +2661,7 @@ static bool ggml_hexagon_supported_rope(const struct ggml_hexagon_session * sess
 
     int mode = op_params[2];
 
-    if ((mode & GGML_ROPE_TYPE_MROPE) || (mode & GGML_ROPE_TYPE_VISION)) {
+    if (mode == GGML_ROPE_TYPE_VISION) {
         return false;
     }
     if (mode & 1) {
@@ -2735,15 +2735,28 @@ static bool ggml_hexagon_supported_ssm_conv(const struct ggml_hexagon_session *
     if (dst->ne[0] != d_inner || dst->ne[1] != n_t || dst->ne[2] != n_s) {
         return false;
     }
-
-    // TODO: add support for non-contiguous tensors
-    if (!ggml_is_contiguous(src0) || !ggml_is_contiguous(src1) || !ggml_is_contiguous(dst)) {
+    if (src0->nb[0] != sizeof(float) || src1->nb[0] != sizeof(float) || dst->nb[0] != sizeof(float)) {
+        return false;
+    }
+    if (src0->nb[1] != src0->ne[0] * sizeof(float) || src1->nb[1] != src1->ne[0] * sizeof(float)) {
         return false;
     }
 
     return true;
 }
 
+static bool ggml_hexagon_supported_pad(const struct ggml_hexagon_session * sess, const struct ggml_tensor * op) {
+    const struct ggml_tensor * src0 = op->src[0];
+    const struct ggml_tensor * dst  = op;
+
+    if (src0->type != GGML_TYPE_F32 || dst->type != GGML_TYPE_F32) {
+        return false;
+    }
+
+    GGML_UNUSED(sess);
+    return true;
+}
+
 static bool ggml_hexagon_supported_cumsum(const struct ggml_hexagon_session * sess, const struct ggml_tensor * op) {
     const struct ggml_tensor * src0 = op->src[0];
     const struct ggml_tensor * dst  = op;
@@ -2816,6 +2829,21 @@ static bool ggml_hexagon_supported_solve_tri(const struct ggml_hexagon_session *
     return true;
 }
 
+static bool ggml_hexagon_supported_tri(const struct ggml_hexagon_session * sess, const struct ggml_tensor * op) {
+
+    const struct ggml_tensor * src0 = op->src[0];
+    const struct ggml_tensor * dst  = op;
+
+    if (src0->type != GGML_TYPE_F32) { return false; }
+    if (dst->type  != GGML_TYPE_F32) { return false; }
+    if (!ggml_are_same_shape(src0, dst)) { return false; }
+    if (!ggml_is_contiguous(src0) || !ggml_is_contiguous(dst)) { return false; }
+
+    return true;
+
+    GGML_UNUSED(sess);
+}
+
 static const char * ggml_backend_hexagon_name(ggml_backend_t backend) {
     auto sess = static_cast<ggml_hexagon_session *>(backend->context);
     return sess->c_name();
@@ -2843,6 +2871,7 @@ static htp_op_code op_remap_to_htp(const ggml_tensor * t) {
         case GGML_OP_SET_ROWS:        return HTP_OP_SET_ROWS;
         case GGML_OP_SUM_ROWS:        return HTP_OP_SUM_ROWS;
         case GGML_OP_ARGSORT:         return HTP_OP_ARGSORT;
+        case GGML_OP_NORM:            return HTP_OP_NORM;
         case GGML_OP_L2_NORM:         return HTP_OP_L2_NORM;
         case GGML_OP_RMS_NORM:        return HTP_OP_RMS_NORM;
         case GGML_OP_SCALE:           return HTP_OP_SCALE;
@@ -2857,6 +2886,9 @@ static htp_op_code op_remap_to_htp(const ggml_tensor * t) {
         case GGML_OP_FILL:            return HTP_OP_FILL;
         case GGML_OP_DIAG:            return HTP_OP_DIAG;
         case GGML_OP_SOLVE_TRI:       return HTP_OP_SOLVE_TRI;
+        case GGML_OP_TRI:             return HTP_OP_TRI;
+        case GGML_OP_PAD:             return HTP_OP_PAD;
+
         case GGML_OP_UNARY:
             switch (ggml_get_unary_op(t)) {
                 case GGML_UNARY_OP_SILU:     return HTP_OP_UNARY_SILU;
@@ -3308,10 +3340,8 @@ static bool ggml_backend_hexagon_device_supports_op(ggml_backend_dev_t dev, cons
             supp = ggml_hexagon_supported_add_id(sess, op);
             break;
 
+        case GGML_OP_NORM:
         case GGML_OP_L2_NORM:
-            supp = ggml_hexagon_supported_unary(sess, op);
-            break;
-
         case GGML_OP_RMS_NORM:
         case GGML_OP_SCALE:
             supp = ggml_hexagon_supported_unary(sess, op);
@@ -3416,6 +3446,14 @@ static bool ggml_backend_hexagon_device_supports_op(ggml_backend_dev_t dev, cons
             supp = ggml_hexagon_supported_solve_tri(sess, op);
             break;
 
+        case GGML_OP_TRI:
+            supp = ggml_hexagon_supported_tri(sess, op);
+            break;
+
+        case GGML_OP_PAD:
+            supp = ggml_hexagon_supported_pad(sess, op);
+            break;
+
         default:
             break;
     }