Add Command R chat template (#6650)

* Add chat template for command-r model series

* Fix indentation

* Add chat template test for command-r models and update the implementation to trim whitespaces

* Remove debug print

.github/workflows/build.yml

@@ -52,7 +52,7 @@ jobs:
         id: cmake_test
         run: |
           cd build
-          ctest -L main --verbose --timeout 900
+          ctest -L 'main|curl' --verbose --timeout 900
 
       - name: Determine tag name
         id: tag
@@ -101,7 +101,9 @@ jobs:
           sysctl -a
           mkdir build
           cd build
-          cmake -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_METAL_EMBED_LIBRARY=ON -DLLAMA_CURL=ON ..
+          # Metal is disabled due to intermittent failures with Github runners not having a GPU:
+          # https://github.com/ggerganov/llama.cpp/actions/runs/8635935781/job/23674807267#step:5:2313
+          cmake -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_METAL=OFF -DLLAMA_CURL=ON ..
           cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
 
       - name: Test
@@ -209,21 +211,21 @@ jobs:
         id: depends
         run: |
           sudo apt-get update
-          sudo apt-get install build-essential
+          sudo apt-get install build-essential libcurl4-openssl-dev
 
       - name: Build
         id: cmake_build
         run: |
           mkdir build
           cd build
-          cmake .. -DLLAMA_FATAL_WARNINGS=ON
+          cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=ON
           cmake --build . --config Release -j $(nproc)
 
       - name: Test
         id: cmake_test
         run: |
           cd build
-          ctest -L main --verbose --timeout 900
+          ctest -L 'main|curl' --verbose --timeout 900
 
       - name: Test llama2c conversion
         id: llama2c_test
@@ -938,6 +940,12 @@ jobs:
       - name: Download artifacts
         id: download-artifact
         uses: actions/download-artifact@v4
+        with:
+          path: ./artifact
+
+      - name: Move artifacts
+        id: move_artifacts
+        run: mkdir -p ./artifact/release && mv ./artifact/*/*.zip ./artifact/release
 
       - name: Create release
         id: create_release
@@ -956,7 +964,7 @@ jobs:
             const path = require('path');
             const fs = require('fs');
             const release_id = '${{ steps.create_release.outputs.id }}';
-            for (let file of await fs.readdirSync('./artifact')) {
+            for (let file of await fs.readdirSync('./artifact/release')) {
               if (path.extname(file) === '.zip') {
                 console.log('uploadReleaseAsset', file);
                 await github.repos.uploadReleaseAsset({
@@ -964,7 +972,7 @@ jobs:
                   repo: context.repo.repo,
                   release_id: release_id,
                   name: file,
-                  data: await fs.readFileSync(`./artifact/${file}`)
+                  data: await fs.readFileSync(`./artifact/release/${file}`)
                 });
               }
             }

.github/workflows/docker.yml

@@ -91,6 +91,12 @@ jobs:
             echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
           fi
 
+      - name: Downcase github.repository_owner
+        run: |
+          echo "repository_owner_lowercase=${GITHUB_REPOSITORY_OWNER@L}" >> $GITHUB_ENV
+        env:
+          GITHUB_REPOSITORY_OWNER: '${{ github.repository_owner }}'
+
       - name: Build and push Docker image (versioned)
         if: github.event_name == 'push'
         uses: docker/build-push-action@v4
@@ -98,7 +104,7 @@ jobs:
           context: .
           push: true
           platforms: ${{ matrix.config.platforms }}
-          tags: "ghcr.io/${{ github.repository_owner }}/llama.cpp:${{ matrix.config.tag }}-${{ env.COMMIT_SHA }}"
+          tags: "ghcr.io/${{ env.repository_owner_lowercase }}/llama.cpp:${{ matrix.config.tag }}-${{ env.COMMIT_SHA }}"
           file: ${{ matrix.config.dockerfile }}
 
       - name: Build and push Docker image (tagged)
@@ -107,5 +113,5 @@ jobs:
           context: .
           push: ${{ github.event_name == 'push' }}
           platforms: ${{ matrix.config.platforms }}
-          tags: "ghcr.io/${{ github.repository_owner }}/llama.cpp:${{ matrix.config.tag }},ghcr.io/${{ github.repository_owner }}/llama.cpp:${{ matrix.config.tag }}-${{ steps.tag.outputs.name }}"
+          tags: "ghcr.io/${{ env.repository_owner_lowercase }}/llama.cpp:${{ matrix.config.tag }},ghcr.io/${{ env.repository_owner_lowercase }}/llama.cpp:${{ matrix.config.tag }}-${{ steps.tag.outputs.name }}"
           file: ${{ matrix.config.dockerfile }}

.gitignore

@@ -48,6 +48,7 @@ models-mnt
 /convert-llama2c-to-ggml
 /embd-input-test
 /embedding
+/eval-callback
 /gguf
 /gguf-llama-simple
 /gguf-split

Makefile

@@ -1,7 +1,7 @@
 # Define the default target now so that it is always the first target
 BUILD_TARGETS = \
 	main quantize quantize-stats perplexity imatrix embedding vdot q8dot train-text-from-scratch convert-llama2c-to-ggml \
-	simple batched batched-bench save-load-state server gguf gguf-split llama-bench libllava.a llava-cli baby-llama beam-search  \
+	simple batched batched-bench save-load-state server gguf gguf-split eval-callback llama-bench libllava.a llava-cli baby-llama beam-search  \
 	retrieval speculative infill tokenize benchmark-matmult parallel finetune export-lora lookahead lookup passkey gritlm tests/test-c.o
 
 # Binaries only useful for tests
@@ -646,7 +646,7 @@ CUDA_VERSION := $(shell $(NVCC) --version | grep -oP 'release (\K[0-9]+\.[0-9])'
 ifeq ($(shell awk -v "v=$(CUDA_VERSION)" 'BEGIN { print (v < 11.7) }'),1)
 ifndef CUDA_DOCKER_ARCH
 ifndef CUDA_POWER_ARCH
-$(error I ERROR: For CUDA versions < 11.7 a target CUDA architecture must be explicitly provided via CUDA_DOCKER_ARCH)
+$(error I ERROR: For CUDA versions < 11.7 a target CUDA architecture must be explicitly provided via environment variable CUDA_DOCKER_ARCH, e.g. by running "export CUDA_DOCKER_ARCH=compute_XX" on Unix-like systems, where XX is the minimum compute capability that the code needs to run on. A list with compute capabilities can be found here: https://developer.nvidia.com/cuda-gpus )
 endif # CUDA_POWER_ARCH
 endif # CUDA_DOCKER_ARCH
 endif # eq ($(shell echo "$(CUDA_VERSION) < 11.7" | bc),1)
@@ -800,6 +800,10 @@ gguf-split: examples/gguf-split/gguf-split.cpp ggml.o llama.o $(COMMON_DEPS) $(O
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
+eval-callback: examples/eval-callback/eval-callback.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+
 train-text-from-scratch: examples/train-text-from-scratch/train-text-from-scratch.cpp ggml.o llama.o $(COMMON_DEPS) train.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)

README-sycl.md

@@ -8,9 +8,9 @@
 - [Linux](#linux)
 - [Windows](#windows)
 - [Environment Variable](#environment-variable)
-- [Known Issue](#known-issue)
-- [Q&A](#q&a)
-- [Todo](#todo)
+- [Known Issue](#known-issues)
+- [Q&A](#qa)
+- [TODO](#todo)
 
 ## Background
 
@@ -54,10 +54,10 @@ It has the similar design of other llama.cpp BLAS-based paths such as *OpenBLAS,
 
 ## OS
 
-|OS|Status|Verified|
-|-|-|-|
-|Linux|Support|Ubuntu 22.04, Fedora Silverblue 39|
-|Windows|Support|Windows 11|
+| OS      | Status  | Verified                           |
+|---------|---------|------------------------------------|
+| Linux   | Support | Ubuntu 22.04, Fedora Silverblue 39 |
+| Windows | Support | Windows 11                         |
 
 
 ## Hardware
@@ -66,13 +66,13 @@ It has the similar design of other llama.cpp BLAS-based paths such as *OpenBLAS,
 
 **Verified devices**
 
-|Intel GPU| Status | Verified Model|
-|-|-|-|
-|Intel Data Center Max Series| Support| Max 1550|
-|Intel Data Center Flex Series| Support| Flex 170|
-|Intel Arc Series| Support| Arc 770, 730M|
-|Intel built-in Arc GPU| Support| built-in Arc GPU in Meteor Lake|
-|Intel iGPU| Support| iGPU in i5-1250P, i7-1260P, i7-1165G7|
+| Intel GPU                     | Status  | Verified Model                        |
+|-------------------------------|---------|---------------------------------------|
+| Intel Data Center Max Series  | Support | Max 1550, 1100                        |
+| Intel Data Center Flex Series | Support | Flex 170                              |
+| Intel Arc Series              | Support | Arc 770, 730M                         |
+| Intel built-in Arc GPU        | Support | built-in Arc GPU in Meteor Lake       |
+| Intel iGPU                    | Support | iGPU in i5-1250P, i7-1260P, i7-1165G7 |
 
 *Notes:*
 
@@ -84,24 +84,18 @@ It has the similar design of other llama.cpp BLAS-based paths such as *OpenBLAS,
 - **Execution Unit (EU)**
   - If the iGPU has less than 80 EUs, the inference speed will likely be too slow for practical use.
 
-### Nvidia GPU
-The BLAS acceleration on Nvidia GPU through oneAPI can be obtained using the Nvidia plugins for oneAPI and the cuBLAS backend of the upstream oneMKL library. Details and instructions on how to setup the runtime and library can be found in [this section](#i-setup-environment)
+### Other Vendor GPU
 
 **Verified devices**
 
-|Nvidia GPU| Status | Verified Model|
-|-|-|-|
-|Ampere Series| Support| A100, A4000|
-|Ampere Series *(Mobile)*| Support| RTX 40 Series|
-
-*Notes:*
-  - Support for Nvidia targets through oneAPI is currently limited to Linux platforms.
-
-  - Please make sure the native oneAPI MKL *(dedicated to intel CPUs and GPUs)* is not "visible" at this stage to properly setup and use the built-from-source oneMKL with cuBLAS backend in llama.cpp for Nvidia GPUs.
-
+| Nvidia GPU               | Status  | Verified Model |
+|--------------------------|---------|----------------|
+| Ampere Series            | Support | A100, A4000    |
+| Ampere Series *(Mobile)* | Support | RTX 40 Series  |
 
 ## Docker
 The docker build option is currently limited to *intel GPU* targets.
+
 ### Build image
 ```sh
 # Using FP16
@@ -167,30 +161,11 @@ Platform #0: Intel(R) OpenCL HD Graphics
 
 - **Nvidia GPU**
 
-In order to target Nvidia GPUs through SYCL, please make sure the CUDA/CUBLAS native requirements *-found [here](README.md#cublas)-* are installed.
-Installation can be verified by running the following:
-```sh
-nvidia-smi
-```
-Please make sure at least one CUDA device is available, which can be displayed like this *(here an A100-40GB Nvidia GPU)*:
-```
-+---------------------------------------------------------------------------------------+
-| NVIDIA-SMI 535.54.03              Driver Version: 535.54.03    CUDA Version: 12.2     |
-|-----------------------------------------+----------------------+----------------------+
-| GPU  Name                 Persistence-M | Bus-Id        Disp.A | Volatile Uncorr. ECC |
-| Fan  Temp   Perf          Pwr:Usage/Cap |         Memory-Usage | GPU-Util  Compute M. |
-|                                         |                      |               MIG M. |
-|=========================================+======================+======================|
-|   0  NVIDIA A100-PCIE-40GB          On  | 00000000:8D:00.0 Off |                    0 |
-| N/A   36C    P0              57W / 250W |      4MiB / 40960MiB |      0%      Default |
-|                                         |                      |             Disabled |
-+-----------------------------------------+----------------------+----------------------+
-```
-
+In order to target Nvidia GPUs through SYCL, please make sure the CUDA/CUBLAS native requirements *-found [here](README.md#cuda)-* are installed.
 
 2. **Install Intel® oneAPI Base toolkit**
 
-- **Base installation**
+- **For Intel GPU**
 
 The base toolkit can be obtained from the official [Intel® oneAPI Base Toolkit](https://www.intel.com/content/www/us/en/developer/tools/oneapi/base-toolkit.html) page.
 
@@ -202,10 +177,10 @@ Upon a successful installation, SYCL is enabled for the available intel devices,
 
 - **Adding support to Nvidia GPUs**
 
-**oneAPI**: In order to enable SYCL support on Nvidia GPUs, please install the [Codeplay oneAPI Plugin for Nvidia GPUs](https://developer.codeplay.com/products/oneapi/nvidia/download). User should also make sure the plugin version matches the installed base toolkit one *(previous step)* for a seamless "oneAPI on Nvidia GPU" setup.
+**oneAPI Plugin**: In order to enable SYCL support on Nvidia GPUs, please install the [Codeplay oneAPI Plugin for Nvidia GPUs](https://developer.codeplay.com/products/oneapi/nvidia/download). User should also make sure the plugin version matches the installed base toolkit one *(previous step)* for a seamless "oneAPI on Nvidia GPU" setup.
 
 
-**oneMKL**: The current oneMKL releases *(shipped with the oneAPI base-toolkit)* do not contain the cuBLAS backend. A build from source of the upstream [oneMKL](https://github.com/oneapi-src/oneMKL) with the *cuBLAS* backend enabled is thus required to run it on Nvidia GPUs.
+**oneMKL for cuBlas**: The current oneMKL releases *(shipped with the oneAPI base-toolkit)* do not contain the cuBLAS backend. A build from source of the upstream [oneMKL](https://github.com/oneapi-src/oneMKL) with the *cuBLAS* backend enabled is thus required to run it on Nvidia GPUs.
 
 ```sh
 git clone https://github.com/oneapi-src/oneMKL
@@ -237,7 +212,7 @@ When targeting an intel GPU, the user should expect one or more level-zero devic
 
 - **Nvidia GPU**
 
-Similarly, user targetting Nvidia GPUs should expect at least one SYCL-CUDA device [`ext_oneapi_cuda:gpu`] as bellow:
+Similarly, user targeting Nvidia GPUs should expect at least one SYCL-CUDA device [`ext_oneapi_cuda:gpu`] as bellow:
 ```
 [opencl:acc:0] Intel(R) FPGA Emulation Platform for OpenCL(TM), Intel(R) FPGA Emulation Device OpenCL 1.2  [2023.16.12.0.12_195853.xmain-hotfix]
 [opencl:cpu:1] Intel(R) OpenCL, Intel(R) Xeon(R) Gold 6326 CPU @ 2.90GHz OpenCL 3.0 (Build 0) [2023.16.12.0.12_195853.xmain-hotfix]
@@ -260,6 +235,9 @@ cmake --build .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icp
 
 # Option 2: Use FP32 by default
 cmake --build .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+
+#build all binary
+cmake --build . --config Release -j -v
 ```
 
 #### Nvidia GPU
@@ -278,6 +256,10 @@ cmake --build .. -DLLAMA_SYCL=ON -DLLAMA_SYCL_TARGET=NVIDIA -DCMAKE_C_COMPILER=i
 
 # Option 2: Use FP32 by default
 cmake --build .. -DLLAMA_SYCL=ON -DLLAMA_SYCL_TARGET=NVIDIA -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+
+#build all binary
+cmake --build . --config Release -j -v
+
 ```
 
 ### III. Run the inference
@@ -313,10 +295,10 @@ found 6 SYCL devices:
 | 5|    [opencl:acc:0]|               Intel(R) FPGA Emulation Device|       1.2|         24|67108864|     64|    67064815616|
 ```
 
-|Attribute|Note|
-|-|-|
-|compute capability 1.3|Level-zero driver/runtime, recommended |
-|compute capability 3.0|OpenCL driver/runtime, slower than level-zero in most cases|
+| Attribute              | Note                                                        |
+|------------------------|-------------------------------------------------------------|
+| compute capability 1.3 | Level-zero driver/runtime, recommended                      |
+| compute capability 3.0 | OpenCL driver/runtime, slower than level-zero in most cases |
 
 4. Launch inference
 
@@ -325,10 +307,10 @@ There are two device selection modes:
 - Single device: Use one device target specified by the user.
 - Multiple devices: Automatically select the devices with the same largest Max compute-units.
 
-|Device selection|Parameter|
-|-|-|
-|Single device|--split-mode none --main-gpu DEVICE_ID |
-|Multiple devices|--split-mode layer (default)|
+| Device selection | Parameter                              |
+|------------------|----------------------------------------|
+| Single device    | --split-mode none --main-gpu DEVICE_ID |
+| Multiple devices | --split-mode layer (default)           |
 
 Examples:
 
@@ -357,7 +339,6 @@ Otherwise, you can run the script:
 
 *Notes:*
 
-- By default, `mmap` is used to read the model file. In some cases, it causes runtime hang issues. Please disable it by passing `--no-mmap` to the `/bin/main` if faced with the issue.
 - Upon execution, verify the selected device(s) ID(s) in the output log, which can for instance be displayed as follow:
 
 ```sh
@@ -438,7 +419,7 @@ cd build
 
 cmake -G "MinGW Makefiles" ..  -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx  -DCMAKE_BUILD_TYPE=Release -DLLAMA_SYCL_F16=ON
 
-make
+make -j
 ```
 
 Otherwise, run the `win-build-sycl.bat` wrapper which encapsulates the former instructions:
@@ -486,10 +467,10 @@ found 6 SYCL devices:
 
 ```
 
-|Attribute|Note|
-|-|-|
-|compute capability 1.3|Level-zero running time, recommended |
-|compute capability 3.0|OpenCL running time, slower than level-zero in most cases|
+| Attribute              | Note                                                      |
+|------------------------|-----------------------------------------------------------|
+| compute capability 1.3 | Level-zero running time, recommended                      |
+| compute capability 3.0 | OpenCL running time, slower than level-zero in most cases |
 
 
 4. Launch inference
@@ -499,10 +480,10 @@ There are two device selection modes:
 - Single device: Use one device assigned by user.
 - Multiple devices: Automatically choose the devices with the same biggest Max compute units.
 
-|Device selection|Parameter|
-|-|-|
-|Single device|--split-mode none --main-gpu DEVICE_ID |
-|Multiple devices|--split-mode layer (default)|
+| Device selection | Parameter                              |
+|------------------|----------------------------------------|
+| Single device    | --split-mode none --main-gpu DEVICE_ID |
+| Multiple devices | --split-mode layer (default)           |
 
 Examples:
 
@@ -525,7 +506,6 @@ Otherwise, run the following wrapper script:
 
 Note:
 
-- By default, `mmap` is used to read the model file. In some cases, it causes runtime hang issues. Please disable it by passing `--no-mmap` to the `main.exe` if faced with the issue.
 - Upon execution, verify the selected device(s) ID(s) in the output log, which can for instance be displayed as follow:
 
 ```sh
@@ -540,29 +520,23 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 
 #### Build
 
-|Name|Value|Function|
-|-|-|-|
-|LLAMA_SYCL|ON (mandatory)|Enable build with SYCL code path.|
-|LLAMA_SYCL_TARGET | INTEL *(default)* \| NVIDIA|Set the SYCL target device type.|
-|LLAMA_SYCL_F16|OFF *(default)* \|ON *(optional)*|Enable FP16 build with SYCL code path.|
-|CMAKE_C_COMPILER|icx|Set *icx* compiler for SYCL code path.|
-|CMAKE_CXX_COMPILER|icpx *(Linux)*, icx *(Windows)*|Set `icpx/icx` compiler for SYCL code path.|
+| Name               | Value                             | Function                                    |
+|--------------------|-----------------------------------|---------------------------------------------|
+| LLAMA_SYCL         | ON (mandatory)                    | Enable build with SYCL code path.           |
+| LLAMA_SYCL_TARGET  | INTEL *(default)* \| NVIDIA       | Set the SYCL target device type.            |
+| LLAMA_SYCL_F16     | OFF *(default)* \|ON *(optional)* | Enable FP16 build with SYCL code path.      |
+| CMAKE_C_COMPILER   | icx                               | Set *icx* compiler for SYCL code path.      |
+| CMAKE_CXX_COMPILER | icpx *(Linux)*, icx *(Windows)*   | Set `icpx/icx` compiler for SYCL code path. |
 
 #### Runtime
 
-|Name|Value|Function|
-|-|-|-|
-|GGML_SYCL_DEBUG|0 (default) or 1|Enable log function by macro: GGML_SYCL_DEBUG|
-|ZES_ENABLE_SYSMAN| 0 (default) or 1|Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory.<br>Recommended to use when --split-mode = layer|
+| Name              | Value            | Function                                                                                                                  |
+|-------------------|------------------|---------------------------------------------------------------------------------------------------------------------------|
+| GGML_SYCL_DEBUG   | 0 (default) or 1 | Enable log function by macro: GGML_SYCL_DEBUG                                                                             |
+| ZES_ENABLE_SYSMAN | 0 (default) or 1 | Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory.<br>Recommended to use when --split-mode = layer |
 
 ## Known Issues
 
-- Hanging during startup
-
-  llama.cpp uses *mmap* as the default mode for reading the model file and copying it to the GPU. In some systems, `memcpy` might behave abnormally and therefore hang.
-
-  - **Solution**: add `--no-mmap` or `--mmap 0` flag to the `main` executable.
-
 - `Split-mode:[row]` is not supported.
 
 ## Q&A
@@ -574,7 +548,7 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 
 - General compiler error:
 
-  - Remove build folder or try a clean-build.
+  - Remove **build** folder or try a clean-build.
 
 - I can **not** see `[ext_oneapi_level_zero:gpu]` afer installing the GPU driver on Linux.
 
@@ -591,6 +565,6 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 ### **GitHub contribution**:
 Please add the **[SYCL]** prefix/tag in issues/PRs titles to help the SYCL-team check/address them without delay.
 
-## Todo
+## TODO
 
 - Support row layer split for multiple card runs.

README.md

@@ -94,6 +94,7 @@ Typically finetunes of the base models below are supported as well.
 - [x] LLaMA 2 🦙🦙
 - [X] [Mistral 7B](https://huggingface.co/mistralai/Mistral-7B-v0.1)
 - [x] [Mixtral MoE](https://huggingface.co/models?search=mistral-ai/Mixtral)
+- [x] [DBRX](https://huggingface.co/databricks/dbrx-instruct)
 - [X] Falcon
 - [X] [Chinese LLaMA / Alpaca](https://github.com/ymcui/Chinese-LLaMA-Alpaca) and [Chinese LLaMA-2 / Alpaca-2](https://github.com/ymcui/Chinese-LLaMA-Alpaca-2)
 - [X] [Vigogne (French)](https://github.com/bofenghuang/vigogne)
@@ -485,14 +486,14 @@ Building the program with BLAS support may lead to some performance improvements
 
   The environment variable [`CUDA_VISIBLE_DEVICES`](https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#env-vars) can be used to specify which GPU(s) will be used. The following compilation options are also available to tweak performance:
 
-  | Option                         | Legal values           | Default | Description |
-  |--------------------------------|------------------------|---------|-------------|
-  | LLAMA_CUDA_FORCE_DMMV          | Boolean                |   false | Force the use of dequantization + matrix vector multiplication kernels instead of using kernels that do matrix vector multiplication on quantized data. By default the decision is made based on compute capability (MMVQ for 6.1/Pascal/GTX 1000 or higher). Does not affect k-quants. |
-  | LLAMA_CUDA_DMMV_X              | Positive integer >= 32 |      32 | Number of values in x direction processed by the CUDA dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants. |
-  | LLAMA_CUDA_MMV_Y               | Positive integer       |       1 | Block size in y direction for the CUDA mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended. |
-  | LLAMA_CUDA_F16                 | Boolean                |   false | If enabled, use half-precision floating point arithmetic for the CUDA dequantization + mul mat vec kernels and for the q4_1 and q5_1 matrix matrix multiplication kernels. Can improve performance on relatively recent GPUs. |
-  | LLAMA_CUDA_KQUANTS_ITER        | 1 or 2                 |       2 | Number of values processed per iteration and per CUDA thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs. |
-  | LLAMA_CUDA_PEER_MAX_BATCH_SIZE | Positive integer       |     128 | Maximum batch size for which to enable peer access between multiple GPUs. Peer access requires either Linux or NVLink. When using NVLink enabling peer access for larger batch sizes is potentially beneficial. |
+  | Option                         | Legal values           | Default | Description                                                                                                                                                                                                                                                                             |
+  |--------------------------------|------------------------|---------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
+  | LLAMA_CUDA_FORCE_DMMV          | Boolean                | false   | Force the use of dequantization + matrix vector multiplication kernels instead of using kernels that do matrix vector multiplication on quantized data. By default the decision is made based on compute capability (MMVQ for 6.1/Pascal/GTX 1000 or higher). Does not affect k-quants. |
+  | LLAMA_CUDA_DMMV_X              | Positive integer >= 32 | 32      | Number of values in x direction processed by the CUDA dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants.                                         |
+  | LLAMA_CUDA_MMV_Y               | Positive integer       | 1       | Block size in y direction for the CUDA mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended.                                                                                                                                         |
+  | LLAMA_CUDA_F16                 | Boolean                | false   | If enabled, use half-precision floating point arithmetic for the CUDA dequantization + mul mat vec kernels and for the q4_1 and q5_1 matrix matrix multiplication kernels. Can improve performance on relatively recent GPUs.                                                           |
+  | LLAMA_CUDA_KQUANTS_ITER        | 1 or 2                 | 2       | Number of values processed per iteration and per CUDA thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs.                                                                                                                     |
+  | LLAMA_CUDA_PEER_MAX_BATCH_SIZE | Positive integer       | 128     | Maximum batch size for which to enable peer access between multiple GPUs. Peer access requires either Linux or NVLink. When using NVLink enabling peer access for larger batch sizes is potentially beneficial.                                                                         |
 
 - #### hipBLAS
 
@@ -534,11 +535,11 @@ Building the program with BLAS support may lead to some performance improvements
   If your GPU is not officially supported you can use the environment variable [`HSA_OVERRIDE_GFX_VERSION`] set to a similar GPU, for example 10.3.0 on RDNA2 (e.g. gfx1030, gfx1031, or gfx1035) or 11.0.0 on RDNA3.
   The following compilation options are also available to tweak performance (yes, they refer to CUDA, not HIP, because it uses the same code as the cuBLAS version above):
 
-  | Option                  | Legal values           | Default | Description |
-  |-------------------------|------------------------|---------|-------------|
-  | LLAMA_CUDA_DMMV_X       | Positive integer >= 32 |      32 | Number of values in x direction processed by the HIP dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants. |
-  | LLAMA_CUDA_MMV_Y        | Positive integer       |       1 | Block size in y direction for the HIP mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended. Does not affect k-quants. |
-  | LLAMA_CUDA_KQUANTS_ITER | 1 or 2                 |       2 | Number of values processed per iteration and per HIP thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs. |
+  | Option                  | Legal values           | Default | Description                                                                                                                                                                                                                                    |
+  |-------------------------|------------------------|---------|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
+  | LLAMA_CUDA_DMMV_X       | Positive integer >= 32 | 32      | Number of values in x direction processed by the HIP dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants. |
+  | LLAMA_CUDA_MMV_Y        | Positive integer       | 1       | Block size in y direction for the HIP mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended. Does not affect k-quants.                                                                       |
+  | LLAMA_CUDA_KQUANTS_ITER | 1 or 2                 | 2       | Number of values processed per iteration and per HIP thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs.                                                                             |
 
 - #### CLBlast
 
@@ -746,11 +747,11 @@ From the unzipped folder, open a terminal/cmd window here and place a pre-conver
 As the models are currently fully loaded into memory, you will need adequate disk space to save them and sufficient RAM to load them. At the moment, memory and disk requirements are the same.
 
 | Model | Original size | Quantized size (Q4_0) |
-|------:|--------------:|-----------------------:|
-|    7B |         13 GB |                 3.9 GB |
-|   13B |         24 GB |                 7.8 GB |
-|   30B |         60 GB |                19.5 GB |
-|   65B |        120 GB |                38.5 GB |
+|------:|--------------:|----------------------:|
+|    7B |         13 GB |                3.9 GB |
+|   13B |         24 GB |                7.8 GB |
+|   30B |         60 GB |               19.5 GB |
+|   65B |        120 GB |               38.5 GB |
 
 ### Quantization
 
@@ -758,7 +759,7 @@ Several quantization methods are supported. They differ in the resulting model d
 
 *(outdated)*
 
-| Model | Measure      | F16    | Q4_0   | Q4_1   | Q5_0   | Q5_1   | Q8_0   |
+| Model | Measure      |    F16 |   Q4_0 |   Q4_1 |   Q5_0 |   Q5_1 |   Q8_0 |
 |------:|--------------|-------:|-------:|-------:|-------:|-------:|-------:|
 |    7B | perplexity   | 5.9066 | 6.1565 | 6.0912 | 5.9862 | 5.9481 | 5.9070 |
 |    7B | file size    |  13.0G |   3.5G |   3.9G |   4.3G |   4.7G |   6.7G |

SECURITY.md

@@ -49,11 +49,11 @@ If you intend to run multiple models in parallel with shared memory, it is your
 
 1. Tenant Isolation: Models should run separately with strong isolation methods to prevent unwanted data access. Separating networks is crucial for isolation, as it prevents unauthorized access to data or models and malicious users from sending graphs to execute under another tenant's identity.
 
-1. Resource Allocation: A denial of service caused by one model can impact the overall system health. Implement safeguards like rate limits, access controls, and health monitoring.
+2. Resource Allocation: A denial of service caused by one model can impact the overall system health. Implement safeguards like rate limits, access controls, and health monitoring.
 
-1. Model Sharing: In a multitenant model sharing design, tenants and users must understand the security risks of running code provided by others. Since there are no reliable methods to detect malicious models, sandboxing the model execution is the recommended approach to mitigate the risk.
+3. Model Sharing: In a multitenant model sharing design, tenants and users must understand the security risks of running code provided by others. Since there are no reliable methods to detect malicious models, sandboxing the model execution is the recommended approach to mitigate the risk.
 
-1. Hardware Attacks: GPUs or TPUs can also be attacked. [Researches](https://scholar.google.com/scholar?q=gpu+side+channel) has shown that side channel attacks on GPUs are possible, which can make data leak from other models or processes running on the same system at the same time.
+4. Hardware Attacks: GPUs or TPUs can also be attacked. [Researches](https://scholar.google.com/scholar?q=gpu+side+channel) has shown that side channel attacks on GPUs are possible, which can make data leak from other models or processes running on the same system at the same time.
 
 ## Reporting a vulnerability
 

ci/run.sh

@@ -153,6 +153,52 @@ function gg_sum_ctest_release {
     gg_printf '```\n'
 }
 
+# test_scripts_debug
+
+function gg_run_test_scripts_debug {
+    cd ${SRC}
+
+    set -e
+
+    (cd ./examples/gguf-split && time bash tests.sh "$SRC/build-ci-debug/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log
+
+    set +e
+}
+
+function gg_sum_test_scripts_debug {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'Runs test scripts in debug mode\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '```\n'
+    gg_printf '%s\n' "$(cat $OUT/${ci}-scripts.log)"
+    gg_printf '```\n'
+    gg_printf '\n'
+}
+
+# test_scripts_release
+
+function gg_run_test_scripts_release {
+    cd ${SRC}
+
+    set -e
+
+    (cd ./examples/gguf-split && time bash tests.sh "$SRC/build-ci-release/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log
+
+    set +e
+}
+
+function gg_sum_test_scripts_release {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'Runs test scripts in release mode\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '```\n'
+    gg_printf '%s\n' "$(cat $OUT/${ci}-scripts.log)"
+    gg_printf '```\n'
+    gg_printf '\n'
+}
+
 function gg_get_model {
     local gguf_3b="$MNT/models/open-llama/3B-v2/ggml-model-f16.gguf"
     local gguf_7b="$MNT/models/open-llama/7B-v2/ggml-model-f16.gguf"
@@ -642,6 +688,9 @@ test $ret -eq 0 && gg_run ctest_release
 if [ -z ${GG_BUILD_LOW_PERF} ]; then
     test $ret -eq 0 && gg_run embd_bge_small
 
+    test $ret -eq 0 && gg_run test_scripts_debug
+    test $ret -eq 0 && gg_run test_scripts_release
+
     if [ -z ${GG_BUILD_VRAM_GB} ] || [ ${GG_BUILD_VRAM_GB} -ge 8 ]; then
         if [ -z ${GG_BUILD_CUDA} ]; then
             test $ret -eq 0 && gg_run open_llama_3b_v2

common/common.cpp

@@ -1745,6 +1745,8 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
     cparams.yarn_orig_ctx     = params.yarn_orig_ctx;
     cparams.pooling_type      = params.pooling_type;
     cparams.defrag_thold      = params.defrag_thold;
+    cparams.cb_eval           = params.cb_eval;
+    cparams.cb_eval_user_data = params.cb_eval_user_data;
     cparams.offload_kqv       = !params.no_kv_offload;
 
     cparams.type_k = kv_cache_type_from_str(params.cache_type_k);
@@ -2192,7 +2194,7 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
         params.sparams.logit_bias[llama_token_eos(model)] = -INFINITY;
     }
 
-    {
+    if (params.warmup) {
         LOG("warming up the model with an empty run\n");
 
         std::vector<llama_token> tmp = { llama_token_bos(model), llama_token_eos(model), };

common/common.h

@@ -80,6 +80,9 @@ struct gpt_params {
     int32_t yarn_orig_ctx         = 0;     // YaRN original context length
     float   defrag_thold          = -1.0f; // KV cache defragmentation threshold
 
+    ggml_backend_sched_eval_callback cb_eval = nullptr;
+    void * cb_eval_user_data                 = nullptr;
+
     ggml_numa_strategy numa = GGML_NUMA_STRATEGY_DISABLED;
 
     llama_rope_scaling_type rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED;
@@ -156,6 +159,7 @@ struct gpt_params {
     bool infill            = false; // use infill mode
     bool dump_kv_cache     = false; // dump the KV cache contents for debugging purposes
     bool no_kv_offload     = false; // disable KV offloading
+    bool warmup            = true;  // warmup run
 
     std::string cache_type_k = "f16"; // KV cache data type for the K
     std::string cache_type_v = "f16"; // KV cache data type for the V

common/json-schema-to-grammar.cpp

@@ -11,35 +11,101 @@
 
 using json = nlohmann::ordered_json;
 
+template <typename Iterator>
+static std::string join(Iterator begin, Iterator end, const std::string & separator);
+
+static std::string repeat(const std::string & str, size_t n);
+
+static std::string build_repetition(const std::string & item_rule, int min_items, int max_items, const std::string & separator_rule = "", bool item_rule_is_literal = false) {
+    if (separator_rule.empty()) {
+        if (min_items == 0 && max_items == 1) {
+            return item_rule + "?";
+        } else if (min_items == 1 && max_items == std::numeric_limits<int>::max()) {
+            return item_rule + "+";
+        }
+    }
+
+    std::string result;
+    if (min_items > 0) {
+        if (item_rule_is_literal && separator_rule.empty()) {
+            result = "\"" + repeat(std::string(item_rule.begin() + 1, item_rule.end() - 1), min_items) + "\"";
+        } else {
+            std::vector<std::string> items(min_items, item_rule);
+            result = join(items.begin(), items.end(), separator_rule.empty() ? " " : " " + separator_rule + " ");
+        }
+    }
+
+    std::function<std::string(int, bool)> opt_repetitions = [&](int up_to_n, bool prefix_with_sep) -> std::string {
+        auto content = prefix_with_sep && !separator_rule.empty() ? separator_rule + " " + item_rule : item_rule;
+
+        if (up_to_n == 0) {
+            return "";
+        } else if (up_to_n == 1) {
+            return "(" + content + ")?";
+        } else if (!separator_rule.empty() && !prefix_with_sep) {
+            return "(" + content + " " + opt_repetitions(up_to_n - 1, true) + ")?";
+        } else {
+            std::string res = repeat("(" + content + " ", up_to_n);
+            // strip trailing space
+            res = res.substr(0, res.length() - 1);
+            res += repeat(")?", up_to_n);
+            return res;
+        }
+    };
+
+    if (min_items > 0 && max_items != min_items) {
+        result += " ";
+    }
+
+    if (max_items != std::numeric_limits<int>::max()) {
+        result += opt_repetitions(max_items - min_items, min_items > 0);
+    } else {
+        std::string item_operator = "(" + (separator_rule.empty() ? "" : separator_rule + " ") + item_rule + ")";
+        if (min_items == 0 && !separator_rule.empty()) {
+            result = "(" + item_rule + " " + item_operator + "*)?";
+        } else {
+            result += item_operator + "*";
+        }
+    }
+
+    return result;
+}
+
 const std::string SPACE_RULE = "\" \"?";
 
-std::unordered_map<std::string, std::string> PRIMITIVE_RULES = {
-    {"boolean", "(\"true\" | \"false\") space"},
-    {"number", "(\"-\"? ([0-9] | [1-9] [0-9]*)) (\".\" [0-9]+)? ([eE] [-+]? [0-9]+)? space"},
-    {"integer", "(\"-\"? ([0-9] | [1-9] [0-9]*)) space"},
-    {"value", "object | array | string | number | boolean"},
-    {"object", "\"{\" space ( string \":\" space value (\",\" space string \":\" space value)* )? \"}\" space"},
-    {"array", "\"[\" space ( value (\",\" space value)* )? \"]\" space"},
-    {"uuid", "\"\\\"\" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] "
-                "\"-\" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] "
-                "\"-\" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] "
-                "\"-\" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] "
-                "\"-\" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] \"\\\"\" space"},
-    {"string", " \"\\\"\" (\n"
-               "        [^\"\\\\] |\n"
-               "        \"\\\\\" ([\"\\\\/bfnrt] | \"u\" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])\n"
-               "      )* \"\\\"\" space"},
-    {"null", "\"null\" space"}
+struct BuiltinRule {
+    std::string content;
+    std::vector<std::string> deps;
 };
-std::vector<std::string> OBJECT_RULE_NAMES = {"object", "array", "string", "number", "boolean", "null", "value"};
 
-std::unordered_map<std::string, std::string> DATE_RULES = {
-    {"date", "[0-9] [0-9] [0-9] [0-9] \"-\" ( \"0\" [1-9] | \"1\" [0-2] ) \"-\" ( \"0\" [1-9] | [1-2] [0-9] | \"3\" [0-1] )"},
-    {"time", "([01] [0-9] | \"2\" [0-3]) \":\" [0-5] [0-9] \":\" [0-5] [0-9] ( \".\" [0-9] [0-9] [0-9] )? ( \"Z\" | ( \"+\" | \"-\" ) ( [01] [0-9] | \"2\" [0-3] ) \":\" [0-5] [0-9] )"},
-    {"date-time", "date \"T\" time"},
-    {"date-string", "\"\\\"\" date \"\\\"\" space"},
-    {"time-string", "\"\\\"\" time \"\\\"\" space"},
-    {"date-time-string", "\"\\\"\" date-time \"\\\"\" space"}
+const std::string _up_to_15_digits = build_repetition("[0-9]", 0, 15);
+
+std::unordered_map<std::string, BuiltinRule> PRIMITIVE_RULES = {
+    {"boolean", {"(\"true\" | \"false\") space", {}}},
+    {"decimal-part", {"[0-9] " + _up_to_15_digits, {}}},
+    {"integral-part", {"[0-9] | [1-9] " + _up_to_15_digits, {}}},
+    {"number", {"(\"-\"? integral-part) (\".\" decimal-part)? ([eE] [-+]? integral-part)? space", {"integral-part", "decimal-part"}}},
+    {"integer", {"(\"-\"? integral-part) space", {"integral-part"}}},
+    {"value", {"object | array | string | number | boolean | null", {"object", "array", "string", "number", "boolean", "null"}}},
+    {"object", {"\"{\" space ( string \":\" space value (\",\" space string \":\" space value)* )? \"}\" space", {"string", "value"}}},
+    {"array", {"\"[\" space ( value (\",\" space value)* )? \"]\" space", {"value"}}},
+    {"uuid", {"\"\\\"\" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] "
+                "\"-\" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] "
+                "\"-\" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] "
+                "\"-\" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] "
+                "\"-\" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] \"\\\"\" space", {}}},
+    {"char",   {"[^\"\\\\] | \"\\\\\" ([\"\\\\/bfnrt] | \"u\" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])", {}}},
+    {"string", {"\"\\\"\" char* \"\\\"\" space", {"char"}}},
+    {"null", {"\"null\" space", {}}},
+};
+
+std::unordered_map<std::string, BuiltinRule> STRING_FORMAT_RULES = {
+    {"date", {"[0-9] [0-9] [0-9] [0-9] \"-\" ( \"0\" [1-9] | \"1\" [0-2] ) \"-\" ( \"0\" [1-9] | [1-2] [0-9] | \"3\" [0-1] )", {}}},
+    {"time", {"([01] [0-9] | \"2\" [0-3]) \":\" [0-5] [0-9] \":\" [0-5] [0-9] ( \".\" [0-9] [0-9] [0-9] )? ( \"Z\" | ( \"+\" | \"-\" ) ( [01] [0-9] | \"2\" [0-3] ) \":\" [0-5] [0-9] )", {}}},
+    {"date-time", {"date \"T\" time", {"date", "time"}}},
+    {"date-string", {"\"\\\"\" date \"\\\"\" space", {"date"}}},
+    {"time-string", {"\"\\\"\" time \"\\\"\" space", {"time"}}},
+    {"date-time-string", {"\"\\\"\" date-time \"\\\"\" space", {"date-time"}}}
 };
 
 static bool is_reserved_name(const std::string & name) {
@@ -47,7 +113,7 @@ static bool is_reserved_name(const std::string & name) {
     if (RESERVED_NAMES.empty()) {
         RESERVED_NAMES.insert("root");
         for (const auto &p : PRIMITIVE_RULES) RESERVED_NAMES.insert(p.first);
-        for (const auto &p : DATE_RULES) RESERVED_NAMES.insert(p.first);
+        for (const auto &p : STRING_FORMAT_RULES) RESERVED_NAMES.insert(p.first);
     }
     return RESERVED_NAMES.find(name) != RESERVED_NAMES.end();
 }
@@ -192,7 +258,7 @@ private:
                 if (_dotall) {
                     rule = "[\\U00000000-\\U0010FFFF]";
                 } else {
-                    rule = "[\\U00000000-\\x09\\x0B\\x0C\\x0E-\\U0010FFFF]";
+                    rule = "[^\\x0A\\x0D]";
                 }
                 return _add_rule("dot", rule);
             };
@@ -308,47 +374,21 @@ private:
                     auto &sub = last.first;
                     auto sub_is_literal = last.second;
 
-                    if (min_times == 0 && max_times == std::numeric_limits<int>::max()) {
-                        sub += "*";
-                    } else if (min_times == 0 && max_times == 1) {
-                        sub += "?";
-                    } else if (min_times == 1 && max_times == std::numeric_limits<int>::max()) {
-                        sub += "+";
-                    } else {
-                        if (!sub_is_literal) {
-                            std::string & sub_id = sub_rule_ids[sub];
-                            if (sub_id.empty()) {
-                                sub_id = _add_rule(name + "-" + std::to_string(sub_rule_ids.size()), sub);
-                            }
-                            sub = sub_id;
+                    if (!sub_is_literal) {
+                        std::string & sub_id = sub_rule_ids[sub];
+                        if (sub_id.empty()) {
+                            sub_id = _add_rule(name + "-" + std::to_string(sub_rule_ids.size()), sub);
                         }
-                        std::string result;
-                        if (sub_is_literal && min_times > 0) {
-                            result = "\"" + repeat(sub.substr(1, sub.length() - 2), min_times) + "\"";
-                        } else {
-                            for (int j = 0; j < min_times; j++) {
-                                if (j > 0) {
-                                    result += " ";
-                                }
-                                result += sub;
-                            }
-                        }
-                        if (min_times > 0 && min_times < max_times) {
-                            result += " ";
-                        }
-                        if (max_times == std::numeric_limits<int>::max()) {
-                            result += sub + "*";
-                        } else {
-                            for (int j = min_times; j < max_times; j++) {
-                                if (j > min_times) {
-                                    result += " ";
-                                }
-                                result += sub + "?";
-                            }
-                        }
-                        seq.back().first = result;
-                        seq.back().second = false;
+                        sub = sub_id;
                     }
+                    seq.back().first = build_repetition(
+                        sub_is_literal ? "\"" + sub + "\"" : sub,
+                        min_times,
+                        max_times,
+                        "",
+                        sub_is_literal
+                    );
+                    seq.back().second = false;
                 } else {
                     std::string literal;
                     auto is_non_literal = [&](char c) {
@@ -424,7 +464,7 @@ private:
         if (additional_properties.is_object() || (additional_properties.is_boolean() && additional_properties.get<bool>())) {
             std::string sub_name = name + (name.empty() ? "" : "-") + "additional";
             std::string value_rule = visit(additional_properties.is_object() ? additional_properties : json::object(), sub_name + "-value");
-            std::string kv_rule = _add_rule(sub_name + "-kv", _add_rule("string", PRIMITIVE_RULES.at("string")) + " \":\" space " + value_rule);
+            std::string kv_rule = _add_rule(sub_name + "-kv", _add_primitive("string", PRIMITIVE_RULES.at("string")) + " \":\" space " + value_rule);
             prop_kv_rule_names["*"] = kv_rule;
             optional_props.push_back("*");
         }
@@ -486,6 +526,25 @@ private:
         return rule;
     }
 
+    std::string _add_primitive(const std::string & name, const BuiltinRule & rule) {
+        auto n = _add_rule(name, rule.content);
+        for (const auto & dep : rule.deps) {
+            BuiltinRule dep_rule;
+            auto it = PRIMITIVE_RULES.find(dep);
+            if (it == PRIMITIVE_RULES.end()) {
+                it = STRING_FORMAT_RULES.find(dep);
+                if (it == STRING_FORMAT_RULES.end()) {
+                    _errors.push_back("Rule " + dep + " not known");
+                    continue;
+                }
+            }
+            if (_rules.find(dep) == _rules.end()) {
+                _add_primitive(dep, it->second);
+            }
+        }
+        return n;
+    }
+
 public:
     SchemaConverter(
         const std::function<json(const std::string &)> & fetch_json,
@@ -647,49 +706,33 @@ public:
                 return _add_rule(rule_name, rule);
             } else {
                 std::string item_rule_name = visit(items, name + (name.empty() ? "" : "-") + "item");
-                std::string list_item_operator = "( \",\" space " + item_rule_name + " )";
-                std::string successive_items;
                 int min_items = schema.contains("minItems") ? schema["minItems"].get<int>() : 0;
                 json max_items_json = schema.contains("maxItems") ? schema["maxItems"] : json();
-                int max_items = max_items_json.is_number_integer() ? max_items_json.get<int>() : -1;
-                if (min_items > 0) {
-                    successive_items += repeat(list_item_operator, min_items - 1);
-                    min_items--;
-                }
-                if (max_items >= 0 && max_items > min_items) {
-                    successive_items += repeat(list_item_operator + "?", max_items - min_items - 1);
-                } else {
-                    successive_items += list_item_operator + "*";
-                }
-                std::string rule;
-                if (min_items == 0) {
-                    rule =  "\"[\" space ( " + item_rule_name + " " + successive_items + " )? \"]\" space";
-                } else {
-                    rule =  "\"[\" space " + item_rule_name + " " + successive_items + " \"]\" space";
-                }
-                return _add_rule(rule_name, rule);
+                int max_items = max_items_json.is_number_integer() ? max_items_json.get<int>() : std::numeric_limits<int>::max();
+
+                return _add_rule(rule_name, "\"[\" space " + build_repetition(item_rule_name, min_items, max_items, "\",\" space") + " \"]\" space");
             }
         } else if ((schema_type.is_null() || schema_type == "string") && schema.contains("pattern")) {
             return _visit_pattern(schema["pattern"], rule_name);
         } else if ((schema_type.is_null() || schema_type == "string") && std::regex_match(schema_format, std::regex("^uuid[1-5]?$"))) {
-            return _add_rule(rule_name == "root" ? "root" : schema_format, PRIMITIVE_RULES.at("uuid"));
-        } else if ((schema_type.is_null() || schema_type == "string") && DATE_RULES.find(schema_format) != DATE_RULES.end()) {
-            for (const auto & kv : DATE_RULES) {
-                _add_rule(kv.first, kv.second);
-            }
-            return schema_format + "-string";
+            return _add_primitive(rule_name == "root" ? "root" : schema_format, PRIMITIVE_RULES.at("uuid"));
+        } else if ((schema_type.is_null() || schema_type == "string") && STRING_FORMAT_RULES.find(schema_format + "-string") != STRING_FORMAT_RULES.end()) {
+            auto prim_name = schema_format + "-string";
+            return _add_rule(rule_name, _add_primitive(prim_name, STRING_FORMAT_RULES.at(prim_name)));
+        } else if (schema_type == "string" && (schema.contains("minLength") || schema.contains("maxLength"))) {
+            std::string char_rule = _add_primitive("char", PRIMITIVE_RULES.at("char"));
+            int min_len = schema.contains("minLength") ? schema["minLength"].get<int>() : 0;
+            int max_len = schema.contains("maxLength") ? schema["maxLength"].get<int>() : std::numeric_limits<int>::max();
+            return _add_rule(rule_name, "\"\\\"\" " + build_repetition(char_rule, min_len, max_len) + " \"\\\"\" space");
         } else if (schema.empty() || schema_type == "object") {
-            for (const auto & n : OBJECT_RULE_NAMES) {
-                _add_rule(n, PRIMITIVE_RULES.at(n));
-            }
-            return _add_rule(rule_name, "object");
+            return _add_rule(rule_name, _add_primitive("object", PRIMITIVE_RULES.at("object")));
         } else {
             if (!schema_type.is_string() || PRIMITIVE_RULES.find(schema_type.get<std::string>()) == PRIMITIVE_RULES.end()) {
                 _errors.push_back("Unrecognized schema: " + schema.dump());
                 return "";
             }
             // TODO: support minimum, maximum, exclusiveMinimum, exclusiveMaximum at least for zero
-            return _add_rule(rule_name == "root" ? "root" : schema_type.get<std::string>(), PRIMITIVE_RULES.at(schema_type.get<std::string>()));
+            return _add_primitive(rule_name == "root" ? "root" : schema_type.get<std::string>(), PRIMITIVE_RULES.at(schema_type.get<std::string>()));
         }
     }
 

convert-hf-to-gguf.py

@@ -43,17 +43,18 @@ AnyModel = TypeVar("AnyModel", bound="type[Model]")
 class Model(ABC):
     _model_classes: dict[str, type[Model]] = {}
 
-    def __init__(self, dir_model: Path, ftype: int, fname_out: Path, is_big_endian: bool):
+    def __init__(self, dir_model: Path, ftype: int, fname_out: Path, is_big_endian: bool, use_temp_file: bool):
         self.dir_model = dir_model
         self.ftype = ftype
         self.fname_out = fname_out
         self.is_big_endian = is_big_endian
         self.endianess = gguf.GGUFEndian.BIG if is_big_endian else gguf.GGUFEndian.LITTLE
+        self.use_temp_file = use_temp_file
         self.is_safetensors = self._is_model_safetensors()
         self.num_parts = Model.count_model_parts(self.dir_model, ".safetensors" if self.is_safetensors else ".bin")
         self.part_names = self._get_part_names()
         self.hparams = Model.load_hparams(self.dir_model)
-        self.gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[self.model_arch], endianess=self.endianess, use_temp_file=False)
+        self.gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[self.model_arch], endianess=self.endianess, use_temp_file=self.use_temp_file)
         self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer"])
 
     @property
@@ -1427,6 +1428,102 @@ class GrokModel(Model):
             self.gguf_writer.add_tensor(new_name, data)
 
 
+@Model.register("DbrxForCausalLM")
+class DbrxModel(Model):
+    model_arch = gguf.MODEL_ARCH.DBRX
+
+    def set_gguf_parameters(self):
+        ffn_config = self.hparams["ffn_config"]
+        attn_config = self.hparams["attn_config"]
+        self.gguf_writer.add_name(self.hparams["model_type"])
+        self.gguf_writer.add_block_count(self.hparams["n_layers"])
+
+        self.gguf_writer.add_context_length(self.hparams["max_seq_len"])
+        self.gguf_writer.add_embedding_length(self.hparams["d_model"])
+        self.gguf_writer.add_feed_forward_length(ffn_config["ffn_hidden_size"])
+
+        self.gguf_writer.add_head_count(self.hparams["n_heads"])
+        self.gguf_writer.add_head_count_kv(attn_config["kv_n_heads"])
+
+        self.gguf_writer.add_rope_freq_base(attn_config["rope_theta"])
+
+        self.gguf_writer.add_clamp_kqv(attn_config["clip_qkv"])
+        self.gguf_writer.add_file_type(self.ftype)
+
+        self.gguf_writer.add_expert_count(ffn_config["moe_num_experts"])
+        self.gguf_writer.add_expert_used_count(ffn_config["moe_top_k"])
+
+        self.gguf_writer.add_layer_norm_eps(1e-5)
+
+        self.gguf_writer.add_file_type(self.ftype)
+        print(f"gguf: file type = {self.ftype}")
+
+    def write_tensors(self):
+        block_count = self.hparams.get("n_layers")
+        tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
+        for name, data_torch in self.get_tensors():
+            n_expert = self.hparams["ffn_config"]["moe_num_experts"]
+            n_ff = self.hparams["ffn_config"]["ffn_hidden_size"]
+            n_embd = self.hparams["d_model"]
+
+            # Specific behavior for experts tensors: suffix .weight, view as 3D and transpose
+            # original implementation expects (n_expert, n_ff, n_embd) for all experts weights
+            # But llama.cpp moe graph works differently
+            # AND the dimensions in ggml are typically in the reverse order of the pytorch dimensions
+            # so (n_expert, n_ff, n_embd) in pytorch is {n_embd, n_ff, n_expert} in ggml_tensor
+            exp_tensor_names = {"ffn.experts.mlp.w1": None,       # LLM_TENSOR_FFN_GATE_EXPS ggml_tensor->ne{n_embd, n_ff,   n_expert}
+                                "ffn.experts.mlp.w2": (0, 2, 1),  # LLM_TENSOR_FFN_DOWN_EXPS ggml_tensor->ne{n_ff,   n_embd, n_expert}
+                                "ffn.experts.mlp.v1": None}       # LLM_TENSOR_FFN_UP_EXPS   ggml_tensor->ne{n_embd, n_ff,   n_expert}
+            experts = False
+            for exp_tensor_name in exp_tensor_names.keys():
+                if name.find(exp_tensor_name) != -1 and name.find(".weight") == -1:
+                    experts = True
+                    data_torch = data_torch.view(n_expert, n_ff, n_embd)
+                    if (permute_tensor := exp_tensor_names[exp_tensor_name]) is not None:
+                        data_torch = data_torch.permute(*permute_tensor)
+                    break
+
+            old_dtype = data_torch.dtype
+
+            # convert any unsupported data types to float32
+            if data_torch.dtype not in (torch.float16, torch.float32):
+                data_torch = data_torch.to(torch.float32)
+
+            data = data_torch.squeeze().numpy()
+
+            # map tensor names
+            # In MoE models the ffn tensors are typically most of the model weights,
+            # and need to be quantizable. Quantize expects tensor names to be suffixed by .weight.
+            # Every other model has the weight names ending in .weight,
+            # let's assume that is the convention which is not the case for dbrx:
+            # https://huggingface.co/databricks/dbrx-instruct/blob/main/model.safetensors.index.json#L15
+            new_name = tensor_map.get_name(name if not experts else name + ".weight", try_suffixes=(".weight",))
+            if new_name is None:
+                print(f"Can not map tensor {name!r}")
+                sys.exit()
+
+            n_dims = len(data.shape)
+            data_dtype = data.dtype
+
+            # Most of the codebase that takes in 1D tensors only handles F32 tensors
+            # and most of the outputs tensors are F32.
+            if data_dtype != np.float32 and n_dims == 1:
+                print(f"Can not map tensor {name!r}: all 1D tensors must be F32")
+                sys.exit()
+
+            # if f32 desired, convert any float16 to float32
+            if self.ftype == 0 and data_dtype == np.float16:
+                data = data.astype(np.float32)
+
+            # if f16 desired, convert any float32 2-dim weight tensors to float16
+            if self.ftype == 1 and data_dtype == np.float32 and n_dims > 1:
+                data = data.astype(np.float16)
+
+            print(f"{new_name}, n_dims = {n_dims}, shape = {data.shape}, {old_dtype} --> {data.dtype}")
+
+            self.gguf_writer.add_tensor(new_name, data)
+
+
 @Model.register("MiniCPMForCausalLM")
 class MiniCPMModel(Model):
     model_arch = gguf.MODEL_ARCH.MINICPM
@@ -2363,6 +2460,7 @@ def parse_args() -> argparse.Namespace:
         "model", type=Path,
         help="directory containing model file",
     )
+    parser.add_argument("--use-temp-file", action="store_true", help="use the tempfile library while processing (helpful when running out of memory, process killed)")
 
     return parser.parse_args()
 
@@ -2406,7 +2504,7 @@ def main() -> None:
 
     with torch.inference_mode():
         model_class = Model.from_model_architecture(hparams["architectures"][0])
-        model_instance = model_class(dir_model, ftype_map[args.outtype], fname_out, args.bigendian)
+        model_instance = model_class(dir_model, ftype_map[args.outtype], fname_out, args.bigendian, args.use_temp_file)
 
         print("Set model parameters")
         model_instance.set_gguf_parameters()

docs/HOWTO-add-model.md

@@ -100,6 +100,8 @@ Have a look to existing implementation like `build_llama`, `build_dbrx` or `buil
 
 When implementing a new graph, please note that the underlying `ggml` backends might not support them all, support of missing backend operations can be added in another PR.
 
+Note: to debug the inference graph: you can use [eval-callback](../examples/eval-callback).
+
 ## GGUF specification
 
 https://github.com/ggerganov/ggml/blob/master/docs/gguf.md

examples/CMakeLists.txt

@@ -19,6 +19,7 @@ else()
     add_subdirectory(benchmark)
     add_subdirectory(convert-llama2c-to-ggml)
     add_subdirectory(embedding)
+    add_subdirectory(eval-callback)
     add_subdirectory(finetune)
     add_subdirectory(gritlm)
     add_subdirectory(gguf-split)

examples/eval-callback/CMakeLists.txt

@@ -0,0 +1,9 @@
+set(TARGET eval-callback)
+add_executable(${TARGET} eval-callback.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)
+
+set(TEST_TARGET test-eval-callback)
+add_test(NAME ${TEST_TARGET} COMMAND eval-callback --hf-repo ggml-org/models --hf-file tinyllamas/stories260K.gguf --model stories260K.gguf --prompt hello --seed 42 -ngl 0)
+set_property(TEST ${TEST_TARGET} PROPERTY LABELS eval-callback curl)

examples/eval-callback/README.md

@@ -0,0 +1,95 @@
+# llama.cpp/examples/eval-callback
+
+A simple example which demonstrates how to use callback during the inference.
+It simply prints to the console all operations and tensor data.
+
+Usage:
+
+```shell
+eval-callback \
+  --hf-repo ggml-org/models \
+  --hf-file phi-2/ggml-model-q4_0.gguf \
+  --model phi-2-q4_0.gguf \
+  --prompt hello \
+  --seed 42 \
+  -ngl 33
+```
+
+Will print:
+
+```shell
+llm_load_tensors: offloaded 33/33 layers to GPU
+...
+llama_new_context_with_model: n_ctx      = 512
+...
+llama_new_context_with_model:      CUDA0 compute buffer size =   105.00 MiB
+llama_new_context_with_model:  CUDA_Host compute buffer size =     6.01 MiB
+llama_new_context_with_model: graph nodes  = 1225
+llama_new_context_with_model: graph splits = 2
+ggml_debug:                 inp_embd = (f32)   GET_ROWS(token_embd.weight{2560, 51200, 1, 1}, inp_tokens{1, 1, 1, 1}}) = {2560, 1, 1, 1}
+                                     [
+                                      [
+                                       [ -0.0181,   0.0272,   0.0272, ...],
+                                      ],
+                                     ]
+ggml_debug:                   norm-0 = (f32)       NORM(CUDA0#inp_embd#0{2560, 1, 1, 1}, }) = {2560, 1, 1, 1}
+                                     [
+                                      [
+                                       [ -0.6989,   1.0636,   1.0636, ...],
+                                      ],
+                                     ]
+ggml_debug:                 norm_w-0 = (f32)        MUL(norm-0{2560, 1, 1, 1}, blk.0.attn_norm.weight{2560, 1, 1, 1}}) = {2560, 1, 1, 1}
+                                     [
+                                      [
+                                       [ -0.1800,   0.2817,   0.2632, ...],
+                                      ],
+                                     ]
+ggml_debug:              attn_norm-0 = (f32)        ADD(norm_w-0{2560, 1, 1, 1}, blk.0.attn_norm.bias{2560, 1, 1, 1}}) = {2560, 1, 1, 1}
+                                     [
+                                      [
+                                       [ -0.1863,   0.2970,   0.2604, ...],
+                                      ],
+                                     ]
+ggml_debug:                   wqkv-0 = (f32)    MUL_MAT(blk.0.attn_qkv.weight{2560, 7680, 1, 1}, attn_norm-0{2560, 1, 1, 1}}) = {7680, 1, 1, 1}
+                                     [
+                                      [
+                                       [ -1.1238,   1.2876,  -1.8086, ...],
+                                      ],
+                                     ]
+ggml_debug:                   bqkv-0 = (f32)        ADD(wqkv-0{7680, 1, 1, 1}, blk.0.attn_qkv.bias{7680, 1, 1, 1}}) = {7680, 1, 1, 1}
+                                     [
+                                      [
+                                       [ -1.1135,   1.4604,  -1.9226, ...],
+                                      ],
+                                     ]
+ggml_debug:            bqkv-0 (view) = (f32)       VIEW(bqkv-0{7680, 1, 1, 1}, }) = {2560, 1, 1, 1}
+                                     [
+                                      [
+                                       [ -1.1135,   1.4604,  -1.9226, ...],
+                                      ],
+                                     ]
+ggml_debug:                   Qcur-0 = (f32)       CONT(bqkv-0 (view){2560, 1, 1, 1}, }) = {2560, 1, 1, 1}
+                                     [
+                                      [
+                                       [ -1.1135,   1.4604,  -1.9226, ...],
+                                      ],
+                                     ]
+ggml_debug:        Qcur-0 (reshaped) = (f32)    RESHAPE(Qcur-0{2560, 1, 1, 1}, }) = {80, 32, 1, 1}
+                                     [
+                                      [
+                                       [ -1.1135,   1.4604,  -1.9226, ...],
+                                       [ -0.3608,   0.5076,  -1.8866, ...],
+                                       [  1.7643,   0.0273,  -2.1065, ...],
+                                       ...
+                                      ],
+                                     ]
+ggml_debug:                   Qcur-0 = (f32)       ROPE(Qcur-0 (reshaped){80, 32, 1, 1}, CUDA0#inp_pos#0{1, 1, 1, 1}}) = {80, 32, 1, 1}
+                                     [
+                                      [
+                                       [ -1.1135,   1.4604,  -1.9226, ...],
+                                       [ -0.3608,   0.5076,  -1.8866, ...],
+                                       [  1.7643,   0.0273,  -2.1065, ...],
+                                       ...
+                                      ],
+                                     ]
+```

examples/eval-callback/eval-callback.cpp

@@ -0,0 +1,195 @@
+#include "common.h"
+#include "llama.h"
+#include "ggml.h"
+
+#include <cstdio>
+#include <random>
+#include <string>
+#include <tuple>
+#include <vector>
+
+/**
+ * This the arbitrary data which will be passed to each callback.
+ * Later on we can for example add operation or tensor name filter from the CLI arg, or a file descriptor to dump the tensor.
+ */
+struct callback_data {
+    std::vector<uint8_t> data;
+};
+
+static std::string ggml_ne_string(const ggml_tensor * t) {
+    std::string str;
+    for (int i = 0; i < GGML_MAX_DIMS; ++i) {
+        str += std::to_string(t->ne[i]);
+        if (i + 1 < GGML_MAX_DIMS) {
+            str += ", ";
+        }
+    }
+    return str;
+}
+
+static void ggml_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne, const size_t * nb, int64_t n) {
+    GGML_ASSERT(n > 0);
+    float sum = 0;
+    for (int64_t i3 = 0; i3 < ne[3]; i3++) {
+        printf("                                     [\n");
+        for (int64_t i2 = 0; i2 < ne[2]; i2++) {
+            if (i2 == n && ne[2] > 2*n) {
+                printf("                                      ..., \n");
+                i2 = ne[2] - n;
+            }
+            printf("                                      [\n");
+            for (int64_t i1 = 0; i1 < ne[1]; i1++) {
+                if (i1 == n && ne[1] > 2*n) {
+                    printf("                                       ..., \n");
+                    i1 = ne[1] - n;
+                }
+                printf("                                       [");
+                for (int64_t i0 = 0; i0 < ne[0]; i0++) {
+                    if (i0 == n && ne[0] > 2*n) {
+                        printf("..., ");
+                        i0 = ne[0] - n;
+                    }
+                    size_t i = i3 * nb[3] + i2 * nb[2] + i1 * nb[1] + i0 * nb[0];
+                    float v;
+                    if (type == GGML_TYPE_F16) {
+                        v = ggml_fp16_to_fp32(*(ggml_fp16_t *) data + i);
+                    } else if (type == GGML_TYPE_F32) {
+                        v = *(float *) data + i;
+                    } else if (type == GGML_TYPE_I32) {
+                        v = (float) *(int32_t *) data + i;
+                    } else if (type == GGML_TYPE_I16) {
+                        v = (float) *(int16_t *) data + i;
+                    } else if (type == GGML_TYPE_I8) {
+                        v = (float) *(int8_t *) data + i;
+                    } else {
+                        GGML_ASSERT(false);
+                    }
+                    printf("%12.4f", v);
+                    sum += v;
+                    if (i0 < ne[0] - 1) printf(", ");
+                }
+                printf("],\n");
+            }
+            printf("                                      ],\n");
+        }
+        printf("                                     ]\n");
+        printf("                                     sum = %f\n", sum);
+    }
+}
+
+/**
+ * GGML operations callback during the graph execution.
+ *
+ * @param t current tensor
+ * @param ask when ask is true, the scheduler wants to know if we are interested in data from this tensor
+ *            if we return true, a follow-up call will be made with ask=false in which we can do the actual collection.
+ *            see ggml_backend_sched_eval_callback
+ * @param user_data user data to pass at each call back
+ * @return true to receive data or continue the graph, false otherwise
+ */
+static bool ggml_debug(struct ggml_tensor * t, bool ask, void * user_data) {
+    auto * cb_data = (callback_data *) user_data;
+
+    const struct ggml_tensor * src0 = t->src[0];
+    const struct ggml_tensor * src1 = t->src[1];
+
+    if (ask) {
+        return true; // Always retrieve data
+    }
+
+    char src1_str[128] = {0};
+    if (src1) {
+        sprintf(src1_str, "%s{%s}", src1->name, ggml_ne_string(src1).c_str());
+    }
+
+    printf("%s: %24s = (%s) %10s(%s{%s}, %s}) = {%s}\n", __func__,
+           t->name, ggml_type_name(t->type), ggml_op_desc(t),
+           src0->name, ggml_ne_string(src0).c_str(),
+           src1 ? src1_str : "",
+           ggml_ne_string(t).c_str());
+
+
+    // copy the data from the GPU memory if needed
+    const bool is_host = ggml_backend_buffer_is_host(t->buffer);
+
+    if (!is_host) {
+        auto n_bytes = ggml_nbytes(t);
+        cb_data->data.resize(n_bytes);
+        ggml_backend_tensor_get(t, cb_data->data.data(), 0, n_bytes);
+    }
+
+    if (!ggml_is_quantized(t->type)) {
+        uint8_t * data = is_host ? (uint8_t *) t->data : cb_data->data.data();
+        ggml_print_tensor(data, t->type, t->ne, t->nb, 3);
+    }
+
+    return true;
+}
+
+static bool run(llama_context * ctx, const gpt_params & params) {
+    const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
+
+    std::vector<llama_token> tokens = ::llama_tokenize(ctx, params.prompt, add_bos);
+
+    if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size(), 0, 0))) {
+        fprintf(stderr, "%s : failed to eval\n", __func__);
+        return false;
+    }
+
+    return true;
+}
+
+int main(int argc, char ** argv) {
+
+    callback_data cb_data;
+
+    gpt_params params;
+    if (!gpt_params_parse(argc, argv, params)) {
+        return 1;
+    }
+
+    print_build_info();
+
+    std::mt19937 rng(params.seed);
+    if (params.random_prompt) {
+        params.prompt = gpt_random_prompt(rng);
+    }
+
+    llama_backend_init();
+    llama_numa_init(params.numa);
+
+    // pass the callback to the backend scheduler
+    // it will be executed for each node during the graph computation
+    params.cb_eval = ggml_debug;
+    params.cb_eval_user_data = &cb_data;
+    params.warmup = false;
+
+    // init
+    llama_model * model;
+    llama_context * ctx;
+    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    if (model == nullptr || ctx == nullptr) {
+        fprintf(stderr, "%s : failed to init\n", __func__);
+        return 1;
+    }
+
+    // print system information
+    {
+        fprintf(stderr, "\n");
+        fprintf(stderr, "%s\n", get_system_info(params).c_str());
+    }
+
+    bool OK = run(ctx, params);
+    if (!OK) {
+        return 1;
+    }
+
+    llama_print_timings(ctx);
+
+    llama_free(ctx);
+    llama_free_model(model);
+
+    llama_backend_free();
+
+    return 0;
+}

examples/gbnf-validator/gbnf-validator.cpp

@@ -17,7 +17,7 @@ static bool llama_sample_grammar_string(struct llama_grammar * grammar, const st
     size_t pos = 0;
     for (auto it = code_points.begin(), end = code_points.end() - 1; it != end; ++it) {
         auto prev_stacks = grammar->stacks;
-        grammar->stacks = llama_grammar_accept(grammar->rules, grammar->stacks, *it);
+        llama_grammar_accept(grammar->rules, prev_stacks, *it, grammar->stacks);
         if (grammar->stacks.empty()) {
             error_pos = pos;
             error_msg = "Unexpected character '" + unicode_cpt_to_utf8(*it) + "'";

examples/gguf-split/README.md

@@ -5,5 +5,6 @@ CLI to split / merge GGUF files.
 **Command line options:**
 
 - `--split`: split GGUF to multiple GGUF, default operation.
+- `--split-max-size`: max size per split in `M` or `G`, f.ex. `500M` or `2G`.
 - `--split-max-tensors`: maximum tensors in each split: default(128)
 - `--merge`: merge multiple GGUF to a single GGUF.

examples/gguf-split/gguf-split.cpp

@@ -59,10 +59,10 @@ static size_t split_str_to_n_bytes(std::string str) {
     int n;
     if (str.back() == 'M') {
         sscanf(str.c_str(), "%d", &n);
-        n_bytes = n * 1024 * 1024; // megabytes
+        n_bytes = (size_t)n * 1024 * 1024; // megabytes
     } else if (str.back() == 'G') {
         sscanf(str.c_str(), "%d", &n);
-        n_bytes = n * 1024 * 1024 * 1024; // gigabytes
+        n_bytes = (size_t)n * 1024 * 1024 * 1024; // gigabytes
     } else {
         throw std::invalid_argument("error: supported units are M (megabytes) or G (gigabytes), but got: " + std::string(1, str.back()));
     }

examples/gguf-split/tests.sh

@@ -0,0 +1,89 @@
+#!/bin/bash
+
+set -eu
+
+if [ $# -lt 1 ]
+then
+  echo "usage:   $0 path_to_build_binary [path_to_temp_folder]"
+  echo "example: $0 ../../build/bin ../../tmp"
+  exit 1
+fi
+
+if [ $# -gt 1 ]
+then
+  TMP_DIR=$2
+else
+  TMP_DIR=/tmp
+fi
+
+set -x
+
+SPLIT=$1/gguf-split
+MAIN=$1/main
+WORK_PATH=$TMP_DIR/gguf-split
+CUR_DIR=$(pwd)
+
+mkdir -p "$WORK_PATH"
+
+# Clean up in case of previously failed test
+rm -f $WORK_PATH/ggml-model-split*.gguf $WORK_PATH/ggml-model-merge*.gguf
+
+# 1. Get a model
+(
+  cd $WORK_PATH
+  "$CUR_DIR"/../../scripts/hf.sh --repo ggml-org/gemma-1.1-2b-it-Q8_0-GGUF --file gemma-1.1-2b-it.Q8_0.gguf
+)
+echo PASS
+
+# 2. Split with max tensors strategy
+$SPLIT --split-max-tensors 28  $WORK_PATH/gemma-1.1-2b-it.Q8_0.gguf $WORK_PATH/ggml-model-split
+echo PASS
+echo
+
+# 2b. Test the sharded model is loading properly
+$MAIN --model $WORK_PATH/ggml-model-split-00001-of-00006.gguf --random-prompt --n-predict 32
+echo PASS
+echo
+
+# 3. Merge
+$SPLIT --merge $WORK_PATH/ggml-model-split-00001-of-00006.gguf $WORK_PATH/ggml-model-merge.gguf
+echo PASS
+echo
+
+# 3b. Test the merged model is loading properly
+$MAIN --model $WORK_PATH/ggml-model-merge.gguf --random-prompt --n-predict 32
+echo PASS
+echo
+
+# 4. Split with no tensor in metadata
+#$SPLIT --split-max-tensors 32 --no-tensor-in-metadata $WORK_PATH/ggml-model-merge.gguf $WORK_PATH/ggml-model-split-32-tensors
+#echo PASS
+#echo
+
+# 4b. Test the sharded model is loading properly
+#$MAIN --model $WORK_PATH/ggml-model-split-32-tensors-00001-of-00006.gguf --random-prompt --n-predict 32
+#echo PASS
+#echo
+
+# 5. Merge
+#$SPLIT --merge $WORK_PATH/ggml-model-split-32-tensors-00001-of-00006.gguf $WORK_PATH/ggml-model-merge-2.gguf
+#echo PASS
+#echo
+
+# 5b. Test the merged model is loading properly
+#$MAIN --model $WORK_PATH/ggml-model-merge-2.gguf --random-prompt --n-predict 32
+#echo PASS
+#echo
+
+# 6. Split with size strategy
+$SPLIT --split-max-size 2G $WORK_PATH/ggml-model-merge.gguf $WORK_PATH/ggml-model-split-2G
+echo PASS
+echo
+
+# 6b. Test the sharded model is loading properly
+$MAIN --model $WORK_PATH/ggml-model-split-2G-00001-of-00002.gguf --random-prompt --n-predict 32
+echo PASS
+echo
+
+# Clean up
+rm -f $WORK_PATH/ggml-model-split*.gguf $WORK_PATH/ggml-model-merge*.gguf

examples/gguf/gguf.cpp

@@ -142,7 +142,7 @@ static bool gguf_ex_read_0(const std::string & fname) {
 }
 
 // read and create ggml_context containing the tensors and their data
-static bool gguf_ex_read_1(const std::string & fname) {
+static bool gguf_ex_read_1(const std::string & fname, bool check_data) {
     struct ggml_context * ctx_data = NULL;
 
     struct gguf_init_params params = {
@@ -206,7 +206,7 @@ static bool gguf_ex_read_1(const std::string & fname) {
             printf("\n\n");
 
             // check data
-            {
+            if (check_data) {
                 const float * data = (const float *) cur->data;
                 for (int j = 0; j < ggml_nelements(cur); ++j) {
                     if (data[j] != 100 + i) {
@@ -229,9 +229,16 @@ static bool gguf_ex_read_1(const std::string & fname) {
 
 int main(int argc, char ** argv) {
     if (argc < 3) {
-        printf("usage: %s data.gguf r|w\n", argv[0]);
+        printf("usage: %s data.gguf r|w [n]\n", argv[0]);
+        printf("r: read data.gguf file\n");
+        printf("w: write data.gguf file\n");
+        printf("n: no check of tensor data\n");
         return -1;
     }
+    bool check_data = true;
+    if (argc == 4) {
+        check_data = false;
+    }
 
     const std::string fname(argv[1]);
     const std::string mode (argv[2]);
@@ -242,7 +249,7 @@ int main(int argc, char ** argv) {
         GGML_ASSERT(gguf_ex_write(fname) && "failed to write gguf file");
     } else if (mode == "r") {
         GGML_ASSERT(gguf_ex_read_0(fname) && "failed to read gguf file");
-        GGML_ASSERT(gguf_ex_read_1(fname) && "failed to read gguf file");
+        GGML_ASSERT(gguf_ex_read_1(fname, check_data) && "failed to read gguf file");
     }
 
     return 0;

examples/imatrix/imatrix.cpp

@@ -107,9 +107,7 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
 
         // the top-k selected expert ids are stored in the ids tensor
         // for simplicity, always copy ids to host, because it is small
-        // take into account that ids is not contiguous!
         GGML_ASSERT(ids->ne[1] == src1->ne[1]);
-        GGML_ASSERT(n_as*ggml_nrows(ids)*sizeof(int) == GGML_PAD(ggml_nbytes(ids), n_as*sizeof(int)));
         m_ids.resize(ggml_nbytes(ids)/sizeof(int));
         ggml_backend_tensor_get(ids, m_ids.data(), 0, ggml_nbytes(ids));
 
@@ -597,24 +595,18 @@ int main(int argc, char ** argv) {
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    llama_model_params mparams = llama_model_params_from_gpt_params(params);
-
-    llama_model * model = llama_load_model_from_file(params.model.c_str(), mparams);
-    if (model == NULL) {
-        fprintf(stderr, "%s: error: unable to load model\n", __func__);
-        return 1;
-    }
-
-    llama_context_params cparams = llama_context_params_from_gpt_params(params);
-
     // pass the callback to the backend scheduler
     // it will be executed for each node during the graph computation
-    cparams.cb_eval = ik_collect_imatrix;
-    cparams.cb_eval_user_data = NULL;
+    params.cb_eval = ik_collect_imatrix;
+    params.cb_eval_user_data = NULL;
+    params.warmup = false;
 
-    llama_context * ctx = llama_new_context_with_model(model, cparams);
-    if (ctx == NULL) {
-        fprintf(stderr, "%s: error: unable to create context\n", __func__);
+    // init
+    llama_model * model;
+    llama_context * ctx;
+    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    if (model == nullptr || ctx == nullptr) {
+        fprintf(stderr, "%s : failed to init\n", __func__);
         return 1;
     }
 

examples/infill/README.md

@@ -36,6 +36,11 @@ The `infill` program offers a seamless way to interact with LLaMA models, allowi
 
 ### Example
 
+Download a model that supports infill, for example CodeLlama:
+```console
+scripts/hf.sh --repo TheBloke/CodeLlama-13B-GGUF --file codellama-13b.Q5_K_S.gguf --outdir models
+```
+
 ```bash
 ./infill -t 10 -ngl 0 -m models/codellama-13b.Q5_K_S.gguf -c 4096 --temp 0.7 --repeat_penalty 1.1 -n 20 --in-prefix "def helloworld():\n    print(\"hell" --in-suffix "\n   print(\"goodbye world\")\n    "
 ```

examples/json_schema_to_grammar.py

@@ -6,37 +6,94 @@ import re
 import sys
 from typing import Any, Dict, List, Set, Tuple, Union
 
+def _build_repetition(item_rule, min_items, max_items, separator_rule=None, item_rule_is_literal=False):
+    if not separator_rule:
+        if min_items == 0 and max_items == 1:
+            return f'{item_rule}?'
+        elif min_items == 1 and max_items is None:
+            return f'{item_rule}+'
+
+    result = ''
+
+    if min_items > 0:
+        if item_rule_is_literal and separator_rule is None:
+            result = '"' + (item_rule[1:-1] * min_items) + '"'
+        else:
+            result = (f' {separator_rule} ' if separator_rule else ' ').join([item_rule] * min_items)
+
+    def opt_repetitions(up_to_n, prefix_with_sep=False):
+        '''
+            - n=4, no sep:             '(a (a (a (a)?)?)?)?'
+            - n=4, sep=',', prefix:    '("," a ("," a ("," a ("," a)?)?)?)?'
+            - n=4, sep=',', no prefix: '(a ("," a ("," a ("," a)?)?)?)?'
+        '''
+
+        content = f'{separator_rule} {item_rule}' if prefix_with_sep and separator_rule else item_rule
+        if up_to_n == 0:
+            return ''
+        elif up_to_n == 1:
+            return f'({content})?'
+        elif separator_rule and not prefix_with_sep:
+            return f'({content} {opt_repetitions(up_to_n - 1, prefix_with_sep=True)})?'
+        else:
+            return (f'({content} ' * up_to_n).rstrip() + (')?' * up_to_n)
+
+    if min_items > 0 and max_items != min_items:
+        result += ' '
+
+    if max_items is not None:
+        result += opt_repetitions(max_items - min_items, prefix_with_sep=min_items > 0)
+    else:
+        item_operator = f'({separator_rule + " " if separator_rule else ""}{item_rule})'
+
+        if min_items == 0 and separator_rule:
+            result = f'({item_rule} {item_operator}*)?'
+        else:
+            result += f'{item_operator}*'
+
+    return result
+
+
+class BuiltinRule:
+    def __init__(self, content: str, deps: list = None):
+        self.content = content
+        self.deps = deps or []
+
+_up_to_15_digits = _build_repetition('[0-9]', 0, 15)
+
 # whitespace is constrained to a single space char to prevent model "running away" in
 # whitespace. Also maybe improves generation quality?
 SPACE_RULE = '" "?'
 
 PRIMITIVE_RULES = {
-    'boolean': '("true" | "false") space',
-    'number': '("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space',
-    'integer': '("-"? ([0-9] | [1-9] [0-9]*)) space',
-    'value'  : 'object | array | string | number | boolean',
-    'object' : '"{" space ( string ":" space value ("," space string ":" space value)* )? "}" space',
-    'array'  : '"[" space ( value ("," space value)* )? "]" space',
-    'uuid'   : '"\\"" ' + ' "-" '.join('[0-9a-fA-F]' * n for n in [8, 4, 4, 4, 12]) + ' "\\"" space',
-    'string': r''' "\"" (
-        [^"\\] |
-        "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
-      )* "\"" space''',
-    'null': '"null" space',
+    'boolean'      : BuiltinRule('("true" | "false") space', []),
+    'decimal-part' : BuiltinRule('[0-9] ' + _up_to_15_digits, []),
+    'integral-part': BuiltinRule('[0-9] | [1-9] ' + _up_to_15_digits, []),
+    'number'       : BuiltinRule('("-"? integral-part) ("." decimal-part)? ([eE] [-+]? integral-part)? space', ['integral-part', 'decimal-part']),
+    'integer'      : BuiltinRule('("-"? integral-part) space', ['integral-part']),
+    'value'        : BuiltinRule('object | array | string | number | boolean | null', ['object', 'array', 'string', 'number', 'boolean', 'null']),
+    'object'       : BuiltinRule('"{" space ( string ":" space value ("," space string ":" space value)* )? "}" space', ['string', 'value']),
+    'array'        : BuiltinRule('"[" space ( value ("," space value)* )? "]" space', ['value']),
+    'uuid'         : BuiltinRule(r'"\"" ' + ' "-" '.join('[0-9a-fA-F]' * n for n in [8, 4, 4, 4, 12]) + r' "\"" space', []),
+    'char'         : BuiltinRule(r'[^"\\] | "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])', []),
+    'string'       : BuiltinRule(r'"\"" char* "\"" space', ['char']),
+    'null'         : BuiltinRule('"null" space', []),
 }
-OBJECT_RULE_NAMES = ['object', 'array', 'string', 'number', 'boolean', 'null', 'value']
 
 # TODO: support "uri", "email" string formats
-DATE_RULES = {
-    'date'   : '[0-9] [0-9] [0-9] [0-9] "-" ( "0" [1-9] | "1" [0-2] ) "-" ( \"0\" [1-9] | [1-2] [0-9] | "3" [0-1] )',
-    'time'   : '([01] [0-9] | "2" [0-3]) ":" [0-5] [0-9] ":" [0-5] [0-9] ( "." [0-9] [0-9] [0-9] )? ( "Z" | ( "+" | "-" ) ( [01] [0-9] | "2" [0-3] ) ":" [0-5] [0-9] )',
-    'date-time': 'date "T" time',
-    'date-string': '"\\"" date "\\"" space',
-    'time-string': '"\\"" time "\\"" space',
-    'date-time-string': '"\\"" date-time "\\"" space',
+STRING_FORMAT_RULES = {
+    'date'            : BuiltinRule('[0-9] [0-9] [0-9] [0-9] "-" ( "0" [1-9] | "1" [0-2] ) "-" ( \"0\" [1-9] | [1-2] [0-9] | "3" [0-1] )', []),
+    'time'            : BuiltinRule('([01] [0-9] | "2" [0-3]) ":" [0-5] [0-9] ":" [0-5] [0-9] ( "." [0-9] [0-9] [0-9] )? ( "Z" | ( "+" | "-" ) ( [01] [0-9] | "2" [0-3] ) ":" [0-5] [0-9] )', []),
+    'date-time'       : BuiltinRule('date "T" time', ['date', 'time']),
+    'date-string'     : BuiltinRule('"\\"" date "\\"" space', ['date']),
+    'time-string'     : BuiltinRule('"\\"" time "\\"" space', ['time']),
+    'date-time-string': BuiltinRule('"\\"" date-time "\\"" space', ['date-time']),
 }
 
-RESERVED_NAMES = set(["root", *PRIMITIVE_RULES.keys(), *DATE_RULES.keys()])
+DOTALL = '[\\U00000000-\\U0010FFFF]'
+DOT = '[^\\x0A\\x0D]'
+
+RESERVED_NAMES = set(["root", "dot", *PRIMITIVE_RULES.keys(), *STRING_FORMAT_RULES.keys()])
 
 INVALID_RULE_CHARS_RE = re.compile(r'[^a-zA-Z0-9-]+')
 GRAMMAR_LITERAL_ESCAPE_RE = re.compile(r'[\r\n"]')
@@ -46,8 +103,6 @@ GRAMMAR_LITERAL_ESCAPES = {'\r': '\\r', '\n': '\\n', '"': '\\"', '-': '\\-', ']'
 NON_LITERAL_SET = set('|.()[]{}*+?')
 ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS = set('[]()|{}*+?')
 
-DATE_PATTERN = '[0-9]{4}-(0[1-9]|1[0-2])-([0-2][0-9]|3[0-1])'
-TIME_PATTERN = '([01][0-9]|2[0-3])(:[0-5][0-9]){2}(\\.[0-9]{1,3})?(Z|[+-](([01][0-9]|2[0-3]):[0-5][0-9]))' # Cap millisecond precision w/ 3 digits
 
 class SchemaConverter:
     def __init__(self, *, prop_order, allow_fetch, dotall, raw_pattern):
@@ -55,7 +110,9 @@ class SchemaConverter:
         self._allow_fetch = allow_fetch
         self._dotall = dotall
         self._raw_pattern = raw_pattern
-        self._rules = {'space': SPACE_RULE}
+        self._rules = {
+            'space': SPACE_RULE,
+        }
         self._refs = {}
         self._refs_being_resolved = set()
 
@@ -65,6 +122,29 @@ class SchemaConverter:
         )
         return f'"{escaped}"'
 
+    def not_literal(self, literal: str, dotall: bool = True, maybe_escaped_underscores = False) -> str:
+        '''
+            not_literal('a') -> '[^a]'
+            not_literal('abc') -> '([^a] | "a" ([^b] | "b" ([^c])?)?)?'
+        '''
+        assert len(literal) > 0, 'Empty literal not supported'
+        def recurse(i: int):
+            c = literal[i]
+            if maybe_escaped_underscores and c == '_':
+                yield f'[^{c}\\\\]'
+                yield ' | '
+                yield f'"\\\\"? "{c}"'
+            else:
+                yield f'[^{c}]'
+            if i < len(literal) - 1:
+                yield ' | '
+                yield self._format_literal(c)
+                yield ' ('
+                yield from recurse(i + 1)
+                yield ')?'
+
+        return ''.join(('(', *recurse(0), ')'))
+
     def _add_rule(self, name, rule):
         esc_name = INVALID_RULE_CHARS_RE.sub('-', name)
         if esc_name not in self._rules or self._rules[esc_name] == rule:
@@ -169,10 +249,10 @@ class SchemaConverter:
 
             def get_dot():
                 if self._dotall:
-                    rule = '[\\U00000000-\\U0010FFFF]'
+                    rule = DOTALL
                 else:
                     # Accept any character... except \n and \r line break chars (\x0A and \xOD)
-                    rule = '[\\U00000000-\\x09\\x0B\\x0C\\x0E-\\U0010FFFF]'
+                    rule = DOT
                 return self._add_rule(f'dot', rule)
 
             def join_seq():
@@ -246,26 +326,14 @@ class SchemaConverter:
 
                     (sub, sub_is_literal) = seq[-1]
 
-                    if min_times == 0 and max_times is None:
-                        seq[-1] = (f'{sub}*', False)
-                    elif min_times == 0 and max_times == 1:
-                        seq[-1] = (f'{sub}?', False)
-                    elif min_times == 1 and max_times is None:
-                        seq[-1] = (f'{sub}+', False)
-                    else:
-                        if not sub_is_literal:
-                            id = sub_rule_ids.get(sub)
-                            if id is None:
-                                id = self._add_rule(f'{name}-{len(sub_rule_ids) + 1}', sub)
-                                sub_rule_ids[sub] = id
-                            sub = id
+                    if not sub_is_literal:
+                        id = sub_rule_ids.get(sub)
+                        if id is None:
+                            id = self._add_rule(f'{name}-{len(sub_rule_ids) + 1}', sub)
+                            sub_rule_ids[sub] = id
+                        sub = id
 
-                        seq[-1] = (
-                            ' '.join(
-                                ([f'"{sub[1:-1] * min_times}"'] if sub_is_literal else [sub] * min_times) +
-                                ([f'{sub}?'] * (max_times - min_times) if max_times is not None else [f'{sub}*'])),
-                            False
-                        )
+                    seq[-1] = (_build_repetition(f'"{sub}"' if sub_is_literal else sub, min_times, max_times, item_rule_is_literal=sub_is_literal), False)
                 else:
                     literal = ''
                     while i < length:
@@ -373,49 +441,47 @@ class SchemaConverter:
                     ' "]" space')
             else:
                 item_rule_name = self.visit(items, f'{name}{"-" if name else ""}item')
-                list_item_operator = f'( "," space {item_rule_name} )'
-                successive_items = ""
                 min_items = schema.get("minItems", 0)
                 max_items = schema.get("maxItems")
-                if min_items > 0:
-                    successive_items = list_item_operator * (min_items - 1)
-                    min_items -= 1
-                if max_items is not None and max_items > min_items:
-                    successive_items += (list_item_operator + "?") * (max_items - min_items - 1)
-                else:
-                    successive_items += list_item_operator + "*"
-                if min_items == 0:
-                    rule = f'"[" space ( {item_rule_name} {successive_items} )? "]" space'
-                else:
-                    rule = f'"[" space {item_rule_name} {successive_items} "]" space'
-                return self._add_rule(rule_name, rule)
+                return self._add_rule(rule_name, '"[" space ' + _build_repetition(item_rule_name, min_items, max_items, separator_rule='"," space') + ' "]" space')
 
         elif schema_type in (None, 'string') and 'pattern' in schema:
             return self._visit_pattern(schema['pattern'], rule_name)
 
         elif schema_type in (None, 'string') and re.match(r'^uuid[1-5]?$', schema_format or ''):
-            return self._add_rule(
+            return self._add_primitive(
                 'root' if rule_name == 'root' else schema_format,
                 PRIMITIVE_RULES['uuid']
             )
 
-        elif schema_type in (None, 'string') and schema_format in DATE_RULES:
-            for t, r in DATE_RULES.items():
-                self._add_rule(t, r)
-            return schema_format + '-string'
+        elif schema_type in (None, 'string') and f'{schema_format}-string' in STRING_FORMAT_RULES:
+            prim_name = f'{schema_format}-string'
+            return self._add_rule(rule_name, self._add_primitive(prim_name, STRING_FORMAT_RULES[prim_name]))
+
+        elif schema_type == 'string' and ('minLength' in schema or 'maxLength' in schema):
+            char_rule = self._add_primitive('char', PRIMITIVE_RULES['char'])
+            min_len = schema.get('minLength', 0)
+            max_len = schema.get('maxLength')
+
+            return self._add_rule(rule_name, r'"\"" ' + _build_repetition(char_rule, min_len, max_len) + r' "\"" space')
 
         elif (schema_type == 'object') or (len(schema) == 0):
-            for n in OBJECT_RULE_NAMES:
-                self._add_rule(n, PRIMITIVE_RULES[n])
-            return self._add_rule(rule_name, 'object')
+            return self._add_rule(rule_name, self._add_primitive('object', PRIMITIVE_RULES['object']))
 
         else:
             assert schema_type in PRIMITIVE_RULES, f'Unrecognized schema: {schema}'
             # TODO: support minimum, maximum, exclusiveMinimum, exclusiveMaximum at least for zero
-            return self._add_rule(
-                'root' if rule_name == 'root' else schema_type,
-                PRIMITIVE_RULES[schema_type]
-            )
+            return self._add_primitive('root' if rule_name == 'root' else schema_type, PRIMITIVE_RULES[schema_type])
+
+    def _add_primitive(self, name: str, rule: BuiltinRule):
+        n = self._add_rule(name, rule.content)
+
+        for dep in rule.deps:
+            dep_rule = PRIMITIVE_RULES.get(dep) or STRING_FORMAT_RULES.get(dep)
+            assert dep_rule, f'Rule {dep} not known'
+            if dep not in self._rules:
+                self._add_primitive(dep, dep_rule)
+        return n
 
     def _build_object_rule(self, properties: List[Tuple[str, Any]], required: Set[str], name: str, additional_properties: Union[bool, Any]):
         prop_order = self._prop_order
@@ -437,7 +503,7 @@ class SchemaConverter:
             value_rule = self.visit({} if additional_properties == True else additional_properties, f'{sub_name}-value')
             prop_kv_rule_names["*"] = self._add_rule(
                 f'{sub_name}-kv',
-                self._add_rule('string', PRIMITIVE_RULES['string']) + f' ":" space {value_rule}'
+                self._add_primitive('string', PRIMITIVE_RULES['string']) + f' ":" space {value_rule}'
             )
             optional_props.append("*")
 

examples/llava/MobileVLM-README.md

@@ -22,7 +22,7 @@ After building, run: `./llava-cli` to see the usage. For example:
 
 ## Model conversion
 
-- Clone `mobileVLM-1.7B` and `clip-vit-large-patch14-336` locally:
+1. Clone `mobileVLM-1.7B` and `clip-vit-large-patch14-336` locally:
 
 ```sh
 git clone https://huggingface.co/mtgv/MobileVLM-1.7B

examples/llava/README.md

@@ -24,7 +24,7 @@ After building, run: `./llava-cli` to see the usage. For example:
 
 ## LLaVA 1.5
 
-- Clone a LLaVA and a CLIP model ([available options](https://github.com/haotian-liu/LLaVA/blob/main/docs/MODEL_ZOO.md)). For example:
+1. Clone a LLaVA and a CLIP model ([available options](https://github.com/haotian-liu/LLaVA/blob/main/docs/MODEL_ZOO.md)). For example:
 
 ```sh
 git clone https://huggingface.co/liuhaotian/llava-v1.5-7b

examples/main/README.md

@@ -310,7 +310,7 @@ These options help improve the performance and memory usage of the LLaMA models.
 
 ### Quantization
 
-For information about 4-bit quantization, which can significantly improve performance and reduce memory usage, please refer to llama.cpp's primary [README](../../README.md#prepare-data--run).
+For information about 4-bit quantization, which can significantly improve performance and reduce memory usage, please refer to llama.cpp's primary [README](../../README.md#prepare-and-quantize).
 
 ## Additional Options
 

examples/perplexity/README.md

@@ -3,19 +3,18 @@
 TODO
 
 ## Llama 2 70B Scorechart
-Quantization | Model size (GiB) | Perplexity | Delta to fp16
--- | -- | -- | --
-Q4_0 | 36.20 | 3.5550 | 3.61%
-Q4_1 | 40.20 | 3.5125 | 2.37%
-Q5_0 | 44.20 | 3.4744 | 1.26%
-Q2_K | 27.27 | 3.7339 | 8.82%
-Q3_K_S | 27.86 | 3.7019 | 7.89%
-Q3_K_M | 30.83 | 3.5932 | 4.72%
-Q3_K_L | 33.67 | 3.5617 | 3.80%
-Q4_K_S | 36.39 | 3.4852 | 1.57%
-Q4_K_M | 38.54 | 3.4725 | 1.20%
-Q5_K_S | 44.20 | 3.4483 | 0.50%
-Q5_K_M | 45.41 | 3.4451 | 0.40%
-Q6_K | 52.70 | 3.4367 | 0.16%
-fp16 | 128.5 | 3.4313 | -
-
+| Quantization | Model size (GiB) | Perplexity | Delta to fp16 |
+|--------------|------------------|------------|---------------|
+| Q4_0         | 36.20            | 3.5550     | 3.61%         |
+| Q4_1         | 40.20            | 3.5125     | 2.37%         |
+| Q5_0         | 44.20            | 3.4744     | 1.26%         |
+| Q2_K         | 27.27            | 3.7339     | 8.82%         |
+| Q3_K_S       | 27.86            | 3.7019     | 7.89%         |
+| Q3_K_M       | 30.83            | 3.5932     | 4.72%         |
+| Q3_K_L       | 33.67            | 3.5617     | 3.80%         |
+| Q4_K_S       | 36.39            | 3.4852     | 1.57%         |
+| Q4_K_M       | 38.54            | 3.4725     | 1.20%         |
+| Q5_K_S       | 44.20            | 3.4483     | 0.50%         |
+| Q5_K_M       | 45.41            | 3.4451     | 0.40%         |
+| Q6_K         | 52.70            | 3.4367     | 0.16%         |
+| fp16         | 128.5            | 3.4313     | -             |

examples/quantize/README.md

@@ -4,17 +4,17 @@ TODO
 
 ## Llama 2 7B
 
-Quantization | Bits per Weight (BPW)
--- | --
-Q2_K | 3.35
-Q3_K_S | 3.50
-Q3_K_M | 3.91
-Q3_K_L | 4.27
-Q4_K_S | 4.58
-Q4_K_M | 4.84
-Q5_K_S | 5.52
-Q5_K_M | 5.68
-Q6_K | 6.56
+| Quantization | Bits per Weight (BPW) |
+|--------------|-----------------------|
+| Q2_K         | 3.35                  |
+| Q3_K_S       | 3.50                  |
+| Q3_K_M       | 3.91                  |
+| Q3_K_L       | 4.27                  |
+| Q4_K_S       | 4.58                  |
+| Q4_K_M       | 4.84                  |
+| Q5_K_S       | 5.52                  |
+| Q5_K_M       | 5.68                  |
+| Q6_K         | 6.56                  |
 
 ## Llama 2 13B
 Quantization | Bits per Weight (BPW)

examples/regex-to-grammar.py

@@ -8,7 +8,7 @@ print(subprocess.check_output(
         "python",
         os.path.join(
         os.path.dirname(os.path.realpath(__file__)),
-        "json-schema-to-grammar.py"),
+        "json_schema_to_grammar.py"),
         *rest,
         "-",
         "--raw-pattern",

examples/server/README.md

@@ -11,6 +11,7 @@ Set of LLM REST APIs and a simple web front end to interact with llama.cpp.
  * Continuous batching
  * Multimodal (wip)
  * Monitoring endpoints
+ * Schema-constrained JSON response format
 
 The project is under active development, and we are [looking for feedback and contributors](https://github.com/ggerganov/llama.cpp/issues/4216).
 
@@ -250,6 +251,8 @@ node index.js
 
     `grammar`: Set grammar for grammar-based sampling.  Default: no grammar
 
+    `json_schema`: Set a JSON schema for grammar-based sampling (e.g. `{"items": {"type": "string"}, "minItems": 10, "maxItems": 100}` of a list of strings, or `{}` for any JSON). See [tests](../../tests/test-json-schema-to-grammar.cpp) for supported features.  Default: no JSON schema.
+
     `seed`: Set the random number generator (RNG) seed.  Default: `-1`, which is a random seed.
 
     `ignore_eos`: Ignore end of stream token and continue generating.  Default: `false`
@@ -365,6 +368,8 @@ Notice that each `probs` is an array of length `n_probs`.
 
     See [OpenAI Chat Completions API documentation](https://platform.openai.com/docs/api-reference/chat). While some OpenAI-specific features such as function calling aren't supported, llama.cpp `/completion`-specific features such as `mirostat` are supported.
 
+    The `response_format` parameter supports both plain JSON output (e.g. `{"type": "json_object"}`) and schema-constrained JSON (e.g. `{"type": "json_object", "schema": {"type": "string", "minLength": 10, "maxLength": 100}}`), similar to other OpenAI-inspired API providers.
+
     *Examples:*
 
     You can use either Python `openai` library with appropriate checkpoints:

examples/server/public/json-schema-to-grammar.mjs

@@ -1,33 +1,95 @@
-// WARNING: This file was ported from json-schema-to-grammar.py, please fix bugs / add features there first.
+// WARNING: This file was ported from json_schema_to_grammar.py, please fix bugs / add features there first.
 const SPACE_RULE = '" "?';
 
+function _buildRepetition(itemRule, minItems, maxItems, opts={}) {
+  const separatorRule = opts.separatorRule ?? '';
+  const itemRuleIsLiteral = opts.itemRuleIsLiteral ?? false
+
+  if (separatorRule === '') {
+    if (minItems === 0 && maxItems === 1) {
+      return `${itemRule}?`;
+    } else if (minItems === 1 && maxItems === undefined) {
+      return `${itemRule}+`;
+    }
+  }
+
+  let result = '';
+  if (minItems > 0) {
+    if (itemRuleIsLiteral && separatorRule === '') {
+      result = `"${itemRule.slice(1, -1).repeat(minItems)}"`;
+    } else {
+      result = Array.from({ length: minItems }, () => itemRule)
+        .join(separatorRule !== '' ? ` ${separatorRule} ` : ' ');
+    }
+  }
+
+  const optRepetitions = (upToN, prefixWithSep=false) => {
+    const content = separatorRule !== '' && prefixWithSep ? `${separatorRule} ${itemRule}` : itemRule;
+    if (upToN === 0) {
+      return '';
+    } else if (upToN === 1) {
+      return `(${content})?`;
+    } else if (separatorRule !== '' && !prefixWithSep) {
+      return `(${content} ${optRepetitions(upToN - 1, true)})?`;
+    } else {
+      return Array.from({ length: upToN }, () => `(${content}`).join(' ').trim() + Array.from({ length: upToN }, () => ')?').join('');
+    }
+  };
+
+  if (minItems > 0 && maxItems !== minItems) {
+    result += ' ';
+  }
+
+  if (maxItems !== undefined) {
+    result += optRepetitions(maxItems - minItems, minItems > 0);
+  } else {
+    const itemOperator = `(${separatorRule !== '' ? separatorRule + ' ' : ''}${itemRule})`;
+
+    if (minItems === 0 && separatorRule !== '') {
+      result = `(${itemRule} ${itemOperator}*)?`;
+    } else {
+      result += `${itemOperator}*`;
+    }
+  }
+
+  return result;
+}
+
+class BuiltinRule {
+  constructor(content, deps) {
+    this.content = content;
+    this.deps = deps || [];
+  }
+}
+
+const UP_TO_15_DIGITS = _buildRepetition('[0-9]', 0, 15);
+
 const PRIMITIVE_RULES = {
-  boolean: '("true" | "false") space',
-  number: '("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space',
-  integer: '("-"? ([0-9] | [1-9] [0-9]*)) space',
-  value: 'object | array | string | number | boolean',
-  object: '"{" space ( string ":" space value ("," space string ":" space value)* )? "}" space',
-  array: '"[" space ( value ("," space value)* )? "]" space',
-  uuid: '"\\"" ' + [8, 4, 4, 4, 12].map(n => [...new Array(n)].map(_ => '[0-9a-fA-F]').join('')).join(' "-" ') + ' "\\"" space',
-  string: ` "\\"" (
-        [^"\\\\] |
-        "\\\\" (["\\\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
-      )* "\\"" space`,
-  null: '"null" space',
+  boolean        : new BuiltinRule('("true" | "false") space', []),
+  'decimal-part' : new BuiltinRule('[0-9] ' + UP_TO_15_DIGITS, []),
+  'integral-part': new BuiltinRule('[0-9] | [1-9] ' + UP_TO_15_DIGITS, []),
+  number         : new BuiltinRule('("-"? integral-part) ("." decimal-part)? ([eE] [-+]? integral-part)? space', ['integral-part', 'decimal-part']),
+  integer        : new BuiltinRule('("-"? integral-part) space', ['integral-part']),
+  value          : new BuiltinRule('object | array | string | number | boolean | null', ['object', 'array', 'string', 'number', 'boolean', 'null']),
+  object         : new BuiltinRule('"{" space ( string ":" space value ("," space string ":" space value)* )? "}" space', ['string', 'value']),
+  array          : new BuiltinRule('"[" space ( value ("," space value)* )? "]" space', ['value']),
+  uuid           : new BuiltinRule('"\\"" ' + [8, 4, 4, 4, 12].map(n => [...new Array(n)].map(_ => '[0-9a-fA-F]').join('')).join(' "-" ') + ' "\\"" space', []),
+  char           : new BuiltinRule(`[^"\\\\] | "\\\\" (["\\\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])`, []),
+  string         : new BuiltinRule(`"\\"" char* "\\"" space`, ['char']),
+  null           : new BuiltinRule('"null" space', []),
 };
-const OBJECT_RULE_NAMES = ['object', 'array', 'string', 'number', 'boolean', 'null', 'value'];
 
 // TODO: support "uri", "email" string formats
-const DATE_RULES = {
-    'date'   : '[0-9] [0-9] [0-9] [0-9] "-" ( "0" [1-9] | "1" [0-2] ) "-" ( \"0\" [1-9] | [1-2] [0-9] | "3" [0-1] )',
-    'time'   : '([01] [0-9] | "2" [0-3]) ":" [0-5] [0-9] ":" [0-5] [0-9] ( "." [0-9] [0-9] [0-9] )? ( "Z" | ( "+" | "-" ) ( [01] [0-9] | "2" [0-3] ) ":" [0-5] [0-9] )',
-    'date-time': 'date "T" time',
-    'date-string': '"\\"" date "\\"" space',
-    'time-string': '"\\"" time "\\"" space',
-    'date-time-string': '"\\"" date-time "\\"" space',
-};
+const STRING_FORMAT_RULES = {
+  'date'            : new BuiltinRule('[0-9] [0-9] [0-9] [0-9] "-" ( "0" [1-9] | "1" [0-2] ) "-" ( \"0\" [1-9] | [1-2] [0-9] | "3" [0-1] )', []),
+  'time'            : new BuiltinRule('([01] [0-9] | "2" [0-3]) ":" [0-5] [0-9] ":" [0-5] [0-9] ( "." [0-9] [0-9] [0-9] )? ( "Z" | ( "+" | "-" ) ( [01] [0-9] | "2" [0-3] ) ":" [0-5] [0-9] )', []),
+  'date-time'       : new BuiltinRule('date "T" time', ['date', 'time']),
+  'date-string'     : new BuiltinRule('"\\"" date "\\"" space', ['date']),
+  'time-string'     : new BuiltinRule('"\\"" time "\\"" space', ['time']),
+  'date-time-string': new BuiltinRule('"\\"" date-time "\\"" space', ['date-time']),
+}
 
-const RESERVED_NAMES = {'root': true, ...PRIMITIVE_RULES, ...DATE_RULES};
+const RESERVED_NAMES = {'root': true, ...PRIMITIVE_RULES, ...STRING_FORMAT_RULES};
 
 const INVALID_RULE_CHARS_RE = /[^\dA-Za-z-]+/g;
 const GRAMMAR_LITERAL_ESCAPE_RE = /[\n\r"]/g;
@@ -158,7 +220,7 @@ export class SchemaConverter {
         rule = '[\\U00000000-\\U0010FFFF]';
       } else {
         // Accept any character... except \n and \r line break chars (\x0A and \xOD)
-        rule = '[\\U00000000-\\x09\\x0B\\x0C\\x0E-\\U0010FFFF]';
+        rule = '[^\\x0A\\x0D]';
       }
       return this._addRule('dot', rule);
     };
@@ -259,26 +321,19 @@ export class SchemaConverter {
 
           let [sub, subIsLiteral] = seq[seq.length - 1];
 
-          if (minTimes === 0 && maxTimes === Infinity) {
-            seq[seq.length - 1] = [`${sub}*`, false];
-          } else if (minTimes === 0 && maxTimes === 1) {
-            seq[seq.length - 1] = [`${sub}?`, false];
-          } else if (minTimes === 1 && maxTimes === Infinity) {
-            seq[seq.length - 1] = [`${sub}+`, false];
-          } else {
-            if (!subIsLiteral) {
-              let id = subRuleIds[sub];
-              if (id === undefined) {
-                id = this._addRule(`${name}-${Object.keys(subRuleIds).length + 1}`, sub);
-                subRuleIds[sub] = id;
-              }
-              sub = id;
+          if (!subIsLiteral) {
+            let id = subRuleIds[sub];
+            if (id === undefined) {
+              id = this._addRule(`${name}-${Object.keys(subRuleIds).length + 1}`, sub);
+              subRuleIds[sub] = id;
             }
-
-            const repeatedSub = Array.from({ length: minTimes }, () => subIsLiteral ? `"${sub.slice(1, -1).repeat(minTimes)}"` : sub);
-            const optionalSub = maxTimes !== undefined ? Array.from({ length: maxTimes - minTimes }, () => `${sub}?`) : [`${sub}*`];
-            seq[seq.length - 1] = [repeatedSub.concat(optionalSub).join(' '), false];
+            sub = id;
           }
+
+          seq[seq.length - 1] = [
+            _buildRepetition(subIsLiteral ? `"${sub}"` : sub, minTimes, maxTimes, {itemRuleIsLiteral: subIsLiteral}),
+            false
+          ];
         } else {
           let literal = '';
           while (i < length) {
@@ -394,49 +449,50 @@ export class SchemaConverter {
         );
       } else {
         const itemRuleName = this.visit(items, `${name ?? ''}${name ? '-' : ''}item`);
-        const listItemOperator = `( "," space ${itemRuleName} )`;
-        let successiveItems = '';
-        let minItems = schema.minItems || 0;
+        const minItems = schema.minItems || 0;
         const maxItems = schema.maxItems;
-        if (minItems > 0) {
-          successiveItems = listItemOperator.repeat(minItems - 1);
-          minItems--;
-        }
-        if (maxItems !== undefined && maxItems > minItems) {
-          successiveItems += `${listItemOperator}?`.repeat(maxItems - minItems - 1);
-        } else {
-          successiveItems += `${listItemOperator}*`;
-        }
-        const rule = minItems === 0
-          ? `"[" space ( ${itemRuleName} ${successiveItems} )? "]" space`
-          : `"[" space ${itemRuleName} ${successiveItems} "]" space`;
-        return this._addRule(ruleName, rule);
+        return this._addRule(ruleName, '"[" space ' + _buildRepetition(itemRuleName, minItems, maxItems, {separatorRule: '"," space'}) + ' "]" space');
       }
     } else if ((schemaType === undefined || schemaType === 'string') && 'pattern' in schema) {
       return this._visitPattern(schema.pattern, ruleName);
     } else if ((schemaType === undefined || schemaType === 'string') && /^uuid[1-5]?$/.test(schema.format || '')) {
-      return this._addRule(
-          ruleName === 'root' ? 'root' : schemaFormat,
-          PRIMITIVE_RULES['uuid'])
-    } else if ((schemaType === undefined || schemaType === 'string') && schema.format in DATE_RULES) {
-      for (const [t, r] of Object.entries(DATE_RULES)) {
-        this._addRule(t, r);
-      }
-      return schemaFormat + '-string';
+      return this._addPrimitive(
+        ruleName === 'root' ? 'root' : schemaFormat,
+        PRIMITIVE_RULES['uuid']
+      );
+    } else if ((schemaType === undefined || schemaType === 'string') && `${schema.format}-string` in STRING_FORMAT_RULES) {
+      const primName = `${schema.format}-string`
+      return this._addRule(ruleName, this._addPrimitive(primName, STRING_FORMAT_RULES[primName]));
+    } else if (schemaType === 'string' && ('minLength' in schema || 'maxLength' in schema)) {
+      const charRuleName = this._addPrimitive('char', PRIMITIVE_RULES['char']);
+      const minLen = schema.minLength || 0;
+      const maxLen = schema.maxLength;
+      return this._addRule(ruleName, '"\\\"" ' + _buildRepetition(charRuleName, minLen, maxLen) + ' "\\\"" space');
     } else if ((schemaType === 'object') || (Object.keys(schema).length === 0)) {
-      for (const n of OBJECT_RULE_NAMES) {
-        this._addRule(n, PRIMITIVE_RULES[n]);
-      }
-      return this._addRule(ruleName, 'object');
+      return this._addRule(ruleName, this._addPrimitive('object', PRIMITIVE_RULES['object']));
     } else {
       if (!(schemaType in PRIMITIVE_RULES)) {
         throw new Error(`Unrecognized schema: ${JSON.stringify(schema)}`);
       }
       // TODO: support minimum, maximum, exclusiveMinimum, exclusiveMaximum at least for zero
-      return this._addRule(ruleName === 'root' ? 'root' : schemaType, PRIMITIVE_RULES[schemaType]);
+      return this._addPrimitive(ruleName === 'root' ? 'root' : schemaType, PRIMITIVE_RULES[schemaType]);
     }
   }
 
+  _addPrimitive(name, rule) {
+    let n = this._addRule(name, rule.content);
+    for (const dep of rule.deps) {
+      const depRule = PRIMITIVE_RULES[dep] || STRING_FORMAT_RULES[dep];
+      if (!depRule) {
+        throw new Error(`Rule ${dep} not known`);
+      }
+      if (!(dep in this._rules)) {
+        this._addPrimitive(dep, depRule);
+      }
+    }
+    return n;
+  }
+
   _buildObjectRule(properties, required, name, additionalProperties) {
     const propOrder = this._propOrder;
     // sort by position in prop_order (if specified) then by original order
@@ -462,7 +518,7 @@ export class SchemaConverter {
       const valueRule = this.visit(additionalProperties === true ? {} : additionalProperties, `${subName}-value`);
       propKvRuleNames['*'] = this._addRule(
         `${subName}-kv`,
-        `${this._addRule('string', PRIMITIVE_RULES['string'])} ":" space ${valueRule}`);
+        `${this._addPrimitive('string', PRIMITIVE_RULES['string'])} ":" space ${valueRule}`);
       optionalProps.push('*');
     }
 

examples/server/server.cpp

@@ -859,7 +859,7 @@ struct server_context {
         slot.sparams.min_keep          = json_value(data, "min_keep",          default_sparams.min_keep);
 
         // process "json_schema" and "grammar"
-        if (data.contains("json_schema") && data.contains("grammar")) {
+        if (data.contains("json_schema") && !data["json_schema"].is_null() && data.contains("grammar") && !data["grammar"].is_null()) {
             send_error(task, "Either \"json_schema\" or \"grammar\" can be specified, but not both", ERROR_TYPE_INVALID_REQUEST);
             return false;
         } else if (data.contains("json_schema") && !data.contains("grammar")) {
@@ -1083,7 +1083,7 @@ struct server_context {
                 };
 
                 if (llama_decode(ctx, batch_view) != 0) {
-                    LOG_TEE("%s: llama_decode() failed\n", __func__);
+                    LOG_ERROR("llama_decode() failed", {});
                     return;
                 }
             }
@@ -1281,7 +1281,11 @@ struct server_context {
     }
 
     void send_error(const int id_task, const int id_multi, const std::string & error, const enum error_type type = ERROR_TYPE_SERVER) {
-        LOG_TEE("task %i - error: %s\n", id_task, error.c_str());
+        LOG_ERROR("task error", {
+            {"id_multi", id_multi},
+            {"id_task", id_task},
+            {"error", error},
+        });
 
         server_task_result res;
         res.id       = id_task;
@@ -2186,7 +2190,11 @@ struct server_context {
             if (ret != 0) {
                 if (n_batch == 1 || ret < 0) {
                     // if you get here, it means the KV cache is full - try increasing it via the context size
-                    LOG_TEE("%s : failed to decode the batch, n_batch = %d, ret = %d\n", __func__, n_batch, ret);
+                    LOG_ERROR("failed to decode the batch: KV cache is full - try increasing it via the context size", {
+                        {"i",   i},
+                        {"n_batch",  ret},
+                        {"ret",   ret},
+                    });
                     for (auto & slot : slots) {
                         slot.state = SLOT_STATE_PROCESSING;
                         slot.command = SLOT_COMMAND_NONE;
@@ -2196,12 +2204,16 @@ struct server_context {
                     break; // break loop of n_batch
                 }
 
-                LOG_TEE("%s : failed to find free space in the KV cache, retrying with smaller n_batch = %d\n", __func__, n_batch / 2);
-
                 // retry with half the batch size to try to find a free slot in the KV cache
                 n_batch /= 2;
                 i -= n_batch;
 
+                LOG_WARNING("failed to find free space in the KV cache, retrying with smaller batch size - try increasing it via the context size or enable defragmentation", {
+                    {"i",   i},
+                    {"n_batch",  n_batch},
+                    {"ret",   ret},
+                });
+
                 continue; // continue loop of n_batch
             }
 

examples/sycl/build.sh

@@ -20,4 +20,4 @@ cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
 #cmake --build . --config Release --target llama-bench
 
 #build all binary
-cmake --build . --config Release -v
+cmake --build . --config Release -j -v

examples/sycl/run-llama2.sh

@@ -12,6 +12,7 @@ if [ $# -gt 0 ]; then
     GGML_SYCL_SINGLE_GPU=1
 else
     GGML_SYCL_DEVICE=0
+    GGML_SYCL_SINGLE_GPU=0
 fi
 
 #export GGML_SYCL_DEBUG=1

examples/ts-type-to-grammar.sh

@@ -1,7 +1,7 @@
 #!/bin/bash
 #
 # ./examples/ts-type-to-grammar.sh "{a:string,b:string,c?:string}"
-# python examples/json-schema-to-grammar.py https://json.schemastore.org/tsconfig.json
+# python examples/json_schema_to_grammar.py https://json.schemastore.org/tsconfig.json
 #
 set -euo pipefail
 
@@ -25,4 +25,4 @@ npx ts-json-schema-generator --unstable --no-top-ref --path "$DTS_FILE" --type M
 # https://github.com/YousefED/typescript-json-schema
 # npx typescript-json-schema --defaultProps --required "$DTS_FILE" MyType | tee "$SCHEMA_FILE" >&2
 
-./examples/json-schema-to-grammar.py "$SCHEMA_FILE"
+./examples/json_schema_to_grammar.py "$SCHEMA_FILE"

flake.lock

@@ -20,11 +20,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1712163089,
-        "narHash": "sha256-Um+8kTIrC19vD4/lUCN9/cU9kcOsD1O1m+axJqQPyMM=",
+        "lastModified": 1712791164,
+        "narHash": "sha256-3sbWO1mbpWsLepZGbWaMovSO7ndZeFqDSdX0hZ9nVyw=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "fd281bd6b7d3e32ddfa399853946f782553163b5",
+        "rev": "1042fd8b148a9105f3c0aca3a6177fd1d9360ba5",
         "type": "github"
       },
       "original": {

ggml-cuda.cu

@@ -1946,7 +1946,7 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
     } else if (!split && !fp16_performance_good && src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && src1->ne[1] == 1) {
         // KQV single-batch
         ggml_cuda_mul_mat_vec_nc(ctx, src0, src1, dst);
-    } else if (!split && fp16_performance_good && src0->type == GGML_TYPE_F16 && !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
+    } else if (!split && src0->type == GGML_TYPE_F16 && (src1->type == GGML_TYPE_F16 || fp16_performance_good) && !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
         // KQ + KQV multi-batch
         ggml_cuda_mul_mat_batched_cublas(ctx, src0, src1, dst);
     } else if (use_dequantize_mul_mat_vec) {

ggml-metal.m

@@ -37,6 +37,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_DIV_ROW,
     GGML_METAL_KERNEL_TYPE_SCALE,
     GGML_METAL_KERNEL_TYPE_SCALE_4,
+    GGML_METAL_KERNEL_TYPE_CLAMP,
     GGML_METAL_KERNEL_TYPE_TANH,
     GGML_METAL_KERNEL_TYPE_RELU,
     GGML_METAL_KERNEL_TYPE_GELU,
@@ -468,6 +469,7 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV_ROW,                   div_row,                true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SCALE,                     scale,                  true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SCALE_4,                   scale_4,                true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CLAMP,                     clamp,                  true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_TANH,                      tanh,                   true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RELU,                      relu,                   true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU,                      gelu,                   true);
@@ -713,6 +715,7 @@ static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const
         case GGML_OP_MUL:
         case GGML_OP_DIV:
         case GGML_OP_SCALE:
+        case GGML_OP_CLAMP:
         case GGML_OP_SQR:
         case GGML_OP_SUM_ROWS:
             return true;
@@ -1154,6 +1157,25 @@ static enum ggml_status ggml_metal_graph_compute(
 
                         [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                     } break;
+                case GGML_OP_CLAMP:
+                {
+                    id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CLAMP].pipeline;
+
+                    float min;
+                    float max;
+                    memcpy(&min, ((int32_t *) dst->op_params) + 0, sizeof(float));
+                    memcpy(&max, ((int32_t *) dst->op_params) + 1, sizeof(float));
+
+                    [encoder setComputePipelineState:pipeline];
+                    [encoder setBuffer:id_src0   offset:offs_src0 atIndex:0];
+                    [encoder setBuffer:id_dst    offset:offs_dst  atIndex:1];
+                    [encoder setBytes:&min length:sizeof(min) atIndex:2];
+                    [encoder setBytes:&max length:sizeof(max) atIndex:3];
+
+                    const int64_t n = ggml_nelements(dst);
+
+                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                } break;
                 case GGML_OP_UNARY:
                     switch (ggml_get_unary_op(gf->nodes[i])) {
                         case GGML_UNARY_OP_TANH:
@@ -1926,7 +1948,12 @@ static enum ggml_status ggml_metal_graph_compute(
                                     {
                                         nth0 = 4;
                                         nth1 = 16;
+                                    #if QK_K == 64
+                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_NL_F32].pipeline;
+                                    #else
                                         pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_XS_F32].pipeline;
+                                    #endif
+
                                     } break;
                                 default:
                                     {

ggml-sycl.cpp

@@ -3154,7 +3154,6 @@ typedef float (*vec_dot_q_mul_mat_sycl_t)(
 #define SYCL_SCALE_BLOCK_SIZE 256
 #define SYCL_CLAMP_BLOCK_SIZE 256
 #define SYCL_ROPE_BLOCK_SIZE 256
-#define SYCL_SOFT_MAX_BLOCK_SIZE 1024
 #define SYCL_ALIBI_BLOCK_SIZE 32
 #define SYCL_DIAG_MASK_INF_BLOCK_SIZE 32
 #define SYCL_QUANTIZE_BLOCK_SIZE 256
@@ -13080,11 +13079,13 @@ static void soft_max_f32_sycl(const float * x, const float * mask, const float *
                               const int nrows_y, const float scale, const float max_bias,
                               dpct::queue_ptr stream) {
     int nth = WARP_SIZE;
-    while (nth < ncols_x && nth < SYCL_SOFT_MAX_BLOCK_SIZE) nth *= 2;
+    int max_block_size = g_work_group_size;
+    while (nth < ncols_x && nth < max_block_size) nth *= 2;
+    if (nth>max_block_size) nth = max_block_size;
+
     const sycl::range<3> block_dims(1, 1, nth);
     const sycl::range<3> block_nums(1, 1, nrows_x);
     const size_t n_local_scratch = (GGML_PAD(ncols_x, WARP_SIZE) + WARP_SIZE);
-    static_assert(SYCL_SOFT_MAX_BLOCK_SIZE == 1024, "These values need to be adjusted.");
 
     const uint32_t n_head_kv   = nrows_x/nrows_y;
     const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head_kv));
@@ -13094,6 +13095,12 @@ static void soft_max_f32_sycl(const float * x, const float * mask, const float *
 
     const size_t local_mem_size = stream->get_device().get_info<sycl::info::device::local_mem_size>();
     if (n_local_scratch*sizeof(float) < local_mem_size) {
+        if (ncols_x > max_block_size) {
+            soft_max_f32_submitter<true, 0, 0>(x, mask, pos, dst, ncols_x, nrows_y, scale,
+                                               max_bias, m0, m1, n_head_log2, block_nums,
+                                               block_dims, n_local_scratch, stream);
+            return;
+        }
         switch (ncols_x) {
             case 32:
                 soft_max_f32_submitter<true, 32, 32>(x, mask, pos, dst, ncols_x, nrows_y, scale,
@@ -16814,11 +16821,13 @@ static void ggml_backend_sycl_buffer_set_tensor(ggml_backend_buffer_t buffer,
     const dpct::queue_ptr stream = g_syclStreams[ctx->device][0];
     SYCL_CHECK(
         CHECK_TRY_ERROR(dpct::dev_mgr::instance().get_device(ctx->device).queues_wait_and_throw()));
-
+    char* host_buf = (char*)malloc(size);
+    memcpy(host_buf, data, size);
     SYCL_CHECK(
         CHECK_TRY_ERROR((*stream)
-                             .memcpy((char *)tensor->data + offset, data, size)
+                             .memcpy((char *)tensor->data + offset, host_buf, size)
                              .wait()));
+    free(host_buf);
 }
 catch (sycl::exception const &exc) {
   std::cerr << exc.what() << "Exception caught at file:" << __FILE__

ggml.c

@@ -11012,7 +11012,6 @@ static void ggml_compute_forward_mul_mat_id(
         }
 
         // initialize matrix_row_counts
-        GGML_ASSERT(wdata == wdata_src1_end);
         memset(matrix_row_counts, 0, n_as*sizeof(int64_t));
 
         // group rows by src0 matrix
@@ -20550,6 +20549,32 @@ static bool gguf_fread_str(FILE * file, struct gguf_str * p, size_t * offset) {
     return ok;
 }
 
+static void gguf_free_kv(struct gguf_kv * kv) {
+    if (kv->key.data) {
+        GGML_FREE(kv->key.data);
+    }
+
+    if (kv->type == GGUF_TYPE_STRING) {
+        if (kv->value.str.data) {
+            GGML_FREE(kv->value.str.data);
+        }
+    }
+
+    if (kv->type == GGUF_TYPE_ARRAY) {
+        if (kv->value.arr.data) {
+            if (kv->value.arr.type == GGUF_TYPE_STRING) {
+                for (uint64_t j = 0; j < kv->value.arr.n; ++j) {
+                    struct gguf_str * str = &((struct gguf_str *) kv->value.arr.data)[j];
+                    if (str->data) {
+                        GGML_FREE(str->data);
+                    }
+                }
+            }
+            GGML_FREE(kv->value.arr.data);
+        }
+    }
+}
+
 struct gguf_context * gguf_init_empty(void) {
     struct gguf_context * ctx = GGML_ALIGNED_MALLOC(sizeof(struct gguf_context));
 
@@ -20899,31 +20924,7 @@ void gguf_free(struct gguf_context * ctx) {
     if (ctx->kv) {
         // free string memory - not great..
         for (uint64_t i = 0; i < ctx->header.n_kv; ++i) {
-            struct gguf_kv * kv = &ctx->kv[i];
-
-            if (kv->key.data) {
-                GGML_FREE(kv->key.data);
-            }
-
-            if (kv->type == GGUF_TYPE_STRING) {
-                if (kv->value.str.data) {
-                    GGML_FREE(kv->value.str.data);
-                }
-            }
-
-            if (kv->type == GGUF_TYPE_ARRAY) {
-                if (kv->value.arr.data) {
-                    if (kv->value.arr.type == GGUF_TYPE_STRING) {
-                        for (uint64_t j = 0; j < kv->value.arr.n; ++j) {
-                            struct gguf_str * str = &((struct gguf_str *) kv->value.arr.data)[j];
-                            if (str->data) {
-                                GGML_FREE(str->data);
-                            }
-                        }
-                    }
-                    GGML_FREE(kv->value.arr.data);
-                }
-            }
+            gguf_free_kv(&ctx->kv[i]);
         }
 
         GGML_FREE(ctx->kv);
@@ -21148,6 +21149,19 @@ static int gguf_get_or_add_key(struct gguf_context * ctx, const char * key) {
     return n_kv;
 }
 
+void gguf_remove_key(struct gguf_context * ctx, const char * key) {
+    const int idx = gguf_find_key(ctx, key);
+    if (idx >= 0) {
+        const int n_kv = gguf_get_n_kv(ctx);
+        gguf_free_kv(&ctx->kv[idx]);
+        for (int i = idx; i < n_kv-1; ++i) {
+            ctx->kv[i] = ctx->kv[i+1];
+        }
+        ctx->kv = realloc(ctx->kv, (n_kv - 1) * sizeof(struct gguf_kv));
+        ctx->header.n_kv--;
+    }
+}
+
 void gguf_set_val_u8(struct gguf_context * ctx, const char * key, uint8_t val) {
     const int idx = gguf_get_or_add_key(ctx, key);
 

ggml.h

@@ -2289,6 +2289,9 @@ extern "C" {
     GGML_API char *         gguf_get_tensor_name  (const struct gguf_context * ctx, int i);
     GGML_API enum ggml_type gguf_get_tensor_type  (const struct gguf_context * ctx, int i);
 
+    // removes key if it exists
+    GGML_API void gguf_remove_key(struct gguf_context * ctx, const char * key);
+
     // overrides existing values or adds a new one
     GGML_API void gguf_set_val_u8  (struct gguf_context * ctx, const char * key, uint8_t  val);
     GGML_API void gguf_set_val_i8  (struct gguf_context * ctx, const char * key, int8_t   val);

gguf-py/gguf/constants.py

@@ -126,6 +126,7 @@ class MODEL_ARCH(IntEnum):
     MAMBA      = auto()
     XVERSE     = auto()
     COMMAND_R  = auto()
+    DBRX       = auto()
 
 
 class MODEL_TENSOR(IntEnum):
@@ -195,6 +196,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.MAMBA:          "mamba",
     MODEL_ARCH.XVERSE:         "xverse",
     MODEL_ARCH.COMMAND_R:      "command-r",
+    MODEL_ARCH.DBRX:           "dbrx",
 }
 
 TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
@@ -642,6 +644,19 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.ATTN_K_NORM,
         MODEL_TENSOR.ATTN_Q_NORM,
     ],
+    MODEL_ARCH.DBRX: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.ATTN_OUT_NORM,
+        MODEL_TENSOR.FFN_GATE_INP,
+        MODEL_TENSOR.FFN_GATE_EXP,
+        MODEL_TENSOR.FFN_DOWN_EXP,
+        MODEL_TENSOR.FFN_UP_EXP,
+    ],
     # TODO
 }
 

gguf-py/gguf/tensor_mapping.py

@@ -10,7 +10,7 @@ class TensorNameMap:
         # Token embeddings
         MODEL_TENSOR.TOKEN_EMBD: (
             "gpt_neox.embed_in",                         # gptneox
-            "transformer.wte",                           # gpt2 gpt-j mpt refact qwen
+            "transformer.wte",                           # gpt2 gpt-j mpt refact qwen dbrx
             "transformer.word_embeddings",               # falcon
             "word_embeddings",                           # bloom
             "model.embed_tokens",                        # llama-hf
@@ -48,7 +48,7 @@ class TensorNameMap:
         # Output
         MODEL_TENSOR.OUTPUT: (
             "embed_out",                 # gptneox
-            "lm_head",                   # gpt2 mpt falcon llama-hf baichuan qwen mamba
+            "lm_head",                   # gpt2 mpt falcon llama-hf baichuan qwen mamba dbrx
             "output",                    # llama-pth bloom internlm2
             "word_embeddings_for_head",  # persimmon
             "lm_head.linear",            # phi2
@@ -60,7 +60,7 @@ class TensorNameMap:
             "transformer.ln_f",                        # gpt2 gpt-j falcon
             "model.norm",                              # llama-hf baichuan internlm2
             "norm",                                    # llama-pth
-            "transformer.norm_f",                      # mpt
+            "transformer.norm_f",                      # mpt dbrx
             "ln_f",                                    # refact bloom qwen gpt2
             "language_model.encoder.final_layernorm",  # persimmon
             "model.final_layernorm",                   # persimmon
@@ -96,6 +96,7 @@ class TensorNameMap:
             "model.layers.{bid}.norm",                              # mamba-qbert
             "backbone.layers.{bid}.norm",                           # mamba
             "transformer.decoder_layer.{bid}.rms_norm",             # Grok
+            "transformer.blocks.{bid}.norm_attn_norm.norm_1",       # dbrx
         ),
 
         # Attention norm 2
@@ -108,6 +109,7 @@ class TensorNameMap:
             "gpt_neox.layers.{bid}.attention.query_key_value",                     # gptneox
             "transformer.h.{bid}.attn.c_attn",                                     # gpt2 qwen
             "transformer.blocks.{bid}.attn.Wqkv",                                  # mpt
+            "transformer.blocks.{bid}.norm_attn_norm.attn.Wqkv",                   # dbrx
             "transformer.h.{bid}.self_attention.query_key_value",                  # falcon
             "h.{bid}.self_attention.query_key_value",                              # bloom
             "language_model.encoder.layers.{bid}.self_attention.query_key_value",  # persimmon
@@ -152,23 +154,24 @@ class TensorNameMap:
 
         # Attention output
         MODEL_TENSOR.ATTN_OUT: (
-            "gpt_neox.layers.{bid}.attention.dense",                     # gptneox
-            "transformer.h.{bid}.attn.c_proj",                           # gpt2 refact qwen
-            "transformer.blocks.{bid}.attn.out_proj",                    # mpt
-            "transformer.h.{bid}.self_attention.dense",                  # falcon
-            "h.{bid}.self_attention.dense",                              # bloom
-            "model.layers.{bid}.self_attn.o_proj",                       # llama-hf
-            "layers.{bid}.attention.wo",                                 # llama-pth
-            "encoder.layer.{bid}.attention.output.dense",                # bert
-            "transformer.h.{bid}.attn.out_proj",                         # gpt-j
-            "language_model.encoder.layers.{bid}.self_attention.dense",  # persimmon
-            "model.layers.{bid}.self_attn.dense",                        # persimmon
-            "h.{bid}.attn.c_proj",                                       # gpt2
-            "transformer.h.{bid}.mixer.out_proj",                        # phi2
-            "model.layers.layers.{bid}.self_attn.o_proj",                # plamo
-            "model.layers.{bid}.attention.wo",                           # internlm2
-            "encoder.layers.{bid}.attn.out_proj",                        # nomic-bert
-            "transformer.decoder_layer.{bid}.multi_head_attention.linear"# Grok
+            "gpt_neox.layers.{bid}.attention.dense",                        # gptneox
+            "transformer.h.{bid}.attn.c_proj",                              # gpt2 refact qwen
+            "transformer.blocks.{bid}.attn.out_proj",                       # mpt
+            "transformer.h.{bid}.self_attention.dense",                     # falcon
+            "h.{bid}.self_attention.dense",                                 # bloom
+            "model.layers.{bid}.self_attn.o_proj",                          # llama-hf
+            "layers.{bid}.attention.wo",                                    # llama-pth
+            "encoder.layer.{bid}.attention.output.dense",                   # bert
+            "transformer.h.{bid}.attn.out_proj",                            # gpt-j
+            "language_model.encoder.layers.{bid}.self_attention.dense",     # persimmon
+            "model.layers.{bid}.self_attn.dense",                           # persimmon
+            "h.{bid}.attn.c_proj",                                          # gpt2
+            "transformer.h.{bid}.mixer.out_proj",                           # phi2
+            "model.layers.layers.{bid}.self_attn.o_proj",                   # plamo
+            "model.layers.{bid}.attention.wo",                              # internlm2
+            "encoder.layers.{bid}.attn.out_proj",                           # nomic-bert
+            "transformer.decoder_layer.{bid}.multi_head_attention.linear",  # Grok
+            "transformer.blocks.{bid}.norm_attn_norm.attn.out_proj",        # dbrx
         ),
 
         # Attention output norm
@@ -176,6 +179,7 @@ class TensorNameMap:
             "encoder.layer.{bid}.attention.output.LayerNorm",  # bert
             "encoder.layers.{bid}.norm1",                      # nomic-bert
             "transformer.decoder_layer.{bid}.rms_norm_1",      # Grok
+            "transformer.blocks.{bid}.norm_attn_norm.norm_2",  # dbrx
         ),
 
         # Rotary embeddings
@@ -202,9 +206,10 @@ class TensorNameMap:
         ),
 
         MODEL_TENSOR.FFN_GATE_INP: (
-            "layers.{bid}.feed_forward.gate",           # mixtral
-            "model.layers.{bid}.block_sparse_moe.gate", # mixtral
-            "transformer.decoder_layer.{bid}.router"    # Grok
+            "layers.{bid}.feed_forward.gate",             # mixtral
+            "model.layers.{bid}.block_sparse_moe.gate",   # mixtral
+            "transformer.decoder_layer.{bid}.router",     # Grok
+            "transformer.blocks.{bid}.ffn.router.layer",  # dbrx
         ),
 
         # Feed-forward up
@@ -233,6 +238,7 @@ class TensorNameMap:
         MODEL_TENSOR.FFN_UP_EXP: (
             "layers.{bid}.feed_forward.experts.w3",                 # mixtral (merged)
             "transformer.decoder_layer.{bid}.moe.linear_v",         # Grok (merged)
+            "transformer.blocks.{bid}.ffn.experts.mlp.v1",          # dbrx
         ),
 
         # AWQ-activation gate
@@ -251,8 +257,9 @@ class TensorNameMap:
         ),
 
         MODEL_TENSOR.FFN_GATE_EXP: (
-            "layers.{bid}.feed_forward.experts.w1",                 # mixtral (merged)
-            "transformer.decoder_layer.{bid}.moe.linear"            # Grok (merged)
+            "layers.{bid}.feed_forward.experts.w1",         # mixtral (merged)
+            "transformer.decoder_layer.{bid}.moe.linear",   # Grok (merged)
+            "transformer.blocks.{bid}.ffn.experts.mlp.w1",  # dbrx
         ),
 
         # Feed-forward down
@@ -280,6 +287,7 @@ class TensorNameMap:
         MODEL_TENSOR.FFN_DOWN_EXP: (
             "layers.{bid}.feed_forward.experts.w2",                 # mixtral (merged)
             "transformer.decoder_layer.{bid}.moe.linear_1",         # Grok (merged)
+            "transformer.blocks.{bid}.ffn.experts.mlp.w2",          # dbrx
         ),
 
         MODEL_TENSOR.ATTN_Q_NORM: (

grammars/README.md

@@ -89,3 +89,13 @@ This guide provides a brief overview. Check out the GBNF files in this directory
 ```
 ./main -m <model> --grammar-file grammars/some-grammar.gbnf -p 'Some prompt'
 ```
+
+## Troubleshooting
+
+Grammars currently have performance gotchas (see https://github.com/ggerganov/llama.cpp/issues/4218).
+
+### Efficient optional repetitions
+
+A common pattern is to allow repetitions of a pattern `x` up to N times.
+
+While semantically correct, the syntax `x? x? x?.... x?` (with N repetitions) will result in extremely slow inference. Instead, you can write `(x (x (x ... (x)?...)?)?)?` (w/ N-deep nesting)

llama.cpp

@@ -105,7 +105,7 @@
 #endif
 
 #define LLAMA_MAX_NODES   8192
-#define LLAMA_MAX_EXPERTS 8
+#define LLAMA_MAX_EXPERTS 16
 
 
 //
@@ -220,6 +220,7 @@ enum llm_arch {
     LLM_ARCH_MAMBA,
     LLM_ARCH_XVERSE,
     LLM_ARCH_COMMAND_R,
+    LLM_ARCH_DBRX,
     LLM_ARCH_UNKNOWN,
 };
 
@@ -252,6 +253,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_MAMBA,           "mamba"      },
     { LLM_ARCH_XVERSE,          "xverse"     },
     { LLM_ARCH_COMMAND_R,       "command-r"  },
+    { LLM_ARCH_DBRX,            "dbrx"       },
     { LLM_ARCH_UNKNOWN,         "(unknown)"  },
 };
 
@@ -934,6 +936,22 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
             { LLM_TENSOR_ATTN_K_NORM,     "blk.%d.attn_k_norm" },
         },
     },
+    {
+        LLM_ARCH_DBRX,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_QKV,        "blk.%d.attn_qkv" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_ATTN_OUT_NORM,   "blk.%d.attn_output_norm" },
+            { LLM_TENSOR_FFN_GATE_INP,    "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_GATE_EXPS,   "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,   "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,     "blk.%d.ffn_up_exps" },
+        },
+    },
     {
         LLM_ARCH_UNKNOWN,
         {
@@ -1638,17 +1656,17 @@ static size_t llama_get_device_memory(int device) {
 #if defined(GGML_USE_CUDA)
     size_t total;
     size_t free;
-    ggml_backend_cuda_get_device_memory(device, &total, &free);
+    ggml_backend_cuda_get_device_memory(device, &free, &total);
     return free;
 #elif defined(GGML_USE_SYCL)
     size_t total;
     size_t free;
-    ggml_backend_sycl_get_device_memory(device, &total, &free);
+    ggml_backend_sycl_get_device_memory(device, &free, &total);
     return free;
 #elif defined(GGML_USE_VULKAN)
     size_t total;
     size_t free;
-    ggml_backend_vk_get_device_memory(device, &total, &free);
+    ggml_backend_vk_get_device_memory(device, &free, &total);
     return free;
 #else
     return 1;
@@ -1707,6 +1725,7 @@ enum e_model {
     MODEL_XL,
     MODEL_8x7B,
     MODEL_8x22B,
+    MODEL_16x12B,
 };
 
 static const size_t kiB = 1024;
@@ -3562,6 +3581,7 @@ static const char * llama_model_type_name(e_model type) {
         case MODEL_XL:     return "1.5B";
         case MODEL_8x7B:   return "8x7B";
         case MODEL_8x22B:  return "8x22B";
+        case MODEL_16x12B: return "16x12B";
         default:           return "?B";
     }
 }
@@ -3983,6 +4003,16 @@ static void llm_load_hparams(
                     default: model.type = e_model::MODEL_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_DBRX:
+        {
+            ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS,  hparams.f_norm_eps);
+            ml.get_key(LLM_KV_ATTENTION_CLAMP_KQV,      hparams.f_clamp_kqv);
+
+            switch (hparams.n_layer) {
+                case 40: model.type = e_model::MODEL_16x12B; break;
+                default: model.type = e_model::MODEL_UNKNOWN;
+            }
+        } break;
         default: (void)0;
     }
 
@@ -4671,6 +4701,39 @@ static bool llm_load_tensors(
                         layer.layer_out_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_LAYER_OUT_NORM, "weight", i), {n_embd});
                     }
                 } break;
+            case LLM_ARCH_DBRX:
+            {
+                if (n_expert == 0) {
+                    throw std::runtime_error("DBRX model cannot have zero experts");
+                }
+
+                model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+
+                // output
+                {
+                    model.output_norm = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                    model.output      = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab});
+                }
+
+                for (int i = 0; i < n_layer; ++i) {
+                    ggml_context * ctx_layer = ctx_for_layer(i);
+                    ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                    auto & layer = model.layers[i];
+
+                    layer.attn_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM,  "weight", i), {n_embd});
+
+                    layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa});
+                    layer.wo   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
+
+                    layer.attn_out_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT_NORM, "weight", i), {n_embd});
+
+                    layer.ffn_gate_inp  = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), {n_embd, n_expert});
+                    layer.ffn_gate_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff,   n_expert});
+                    layer.ffn_down_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff,   n_embd, n_expert});
+                    layer.ffn_up_exps   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd, n_ff,   n_expert});
+                }
+            } break;
             case LLM_ARCH_BAICHUAN:
                 {
                     model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
@@ -6433,62 +6496,7 @@ struct llm_build_context {
                         LLM_NORM_RMS, cb, il);
                 cb(cur, "ffn_norm", il);
 
-                ggml_tensor * logits = ggml_mul_mat(ctx0, model.layers[il].ffn_gate_inp, cur); // [n_tokens, num_experts]
-                cb(logits, "ffn_moe_logits", il);
-
-                ggml_tensor * probs = ggml_soft_max(ctx0, logits); // [n_tokens, num_experts]
-                cb(probs, "ffn_moe_probs", il);
-
-                // select experts
-                ggml_tensor * selected_experts = ggml_top_k(ctx0, probs, n_expert_used); // [n_tokens, num_experts_per_tok]
-                cb(selected_experts->src[0], "ffn_moe_argsort", il);
-
-                ggml_tensor * weights = ggml_get_rows(ctx0,
-                        ggml_reshape_3d(ctx0, probs, 1, n_expert, n_tokens), selected_experts);
-                cb(weights, "ffn_moe_weights", il);
-
-                weights = ggml_reshape_2d(ctx0, weights, n_expert_used, n_tokens); // [n_tokens, num_experts_per_tok]
-
-                ggml_tensor * weights_sum = ggml_sum_rows(ctx0, weights);
-                cb(weights_sum, "ffn_moe_weights_sum", il);
-
-                weights = ggml_div(ctx0, weights, weights_sum); // [n_tokens, num_experts_per_tok]
-                cb(weights, "ffn_moe_weights_norm", il);
-
-                // compute expert outputs
-                ggml_tensor * moe_out = nullptr;
-
-                for (int i = 0; i < n_expert_used; ++i) {
-                    ggml_tensor * cur_expert;
-
-                    ggml_tensor * cur_up = ggml_mul_mat_id(ctx0, model.layers[il].ffn_up_exps, selected_experts, i, cur);
-                    cb(cur_up, "ffn_moe_up", il);
-
-                    ggml_tensor * cur_gate = ggml_mul_mat_id(ctx0, model.layers[il].ffn_gate_exps, selected_experts, i, cur);
-                    cb(cur_gate, "ffn_moe_gate", il);
-
-                    cur_gate = ggml_silu(ctx0, cur_gate);
-                    cb(cur_gate, "ffn_moe_silu", il);
-
-                    cur_expert = ggml_mul(ctx0, cur_up, cur_gate);
-                    cb(cur_expert, "ffn_moe_gate_par", il);
-
-                    cur_expert = ggml_mul_mat_id(ctx0, model.layers[il].ffn_down_exps, selected_experts, i, cur_expert); // [n_tokens, n_embd]
-                    cb(cur_expert, "ffn_moe_down", il);
-
-                    cur_expert = ggml_mul(ctx0, cur_expert,
-                            ggml_view_2d(ctx0, weights, 1, n_tokens, weights->nb[1], i*weights->nb[0]));
-                    cb(cur_expert, "ffn_moe_weighted", il);
-
-                    if (i == 0) {
-                        moe_out = cur_expert;
-                    } else {
-                        moe_out = ggml_add(ctx0, moe_out, cur_expert);
-                        cb(moe_out, "ffn_moe_out", il);
-                    }
-                }
-
-                cur = moe_out;
+                cur = build_moe_ffn(cur, n_tokens, LLM_FFN_SILU, il);
             }
 
             cur = ggml_add(ctx0, cur, ffn_inp);
@@ -6520,6 +6528,78 @@ struct llm_build_context {
         return gf;
     }
 
+    // REVIEW: will be replaced by https://github.com/ggerganov/llama.cpp/pull/6505
+    ggml_tensor * build_moe_ffn(ggml_tensor * cur, int32_t n_tokens, llm_ffn_op_type type_op, int il) {
+        ggml_tensor * logits = ggml_mul_mat(ctx0, model.layers[il].ffn_gate_inp, cur); // [n_tokens, num_experts]
+        cb(logits, "ffn_moe_logits", il);
+
+        ggml_tensor * probs = ggml_soft_max(ctx0, logits); // [n_tokens, num_experts]
+        cb(probs, "ffn_moe_probs", il);
+
+        // select experts
+        ggml_tensor * selected_experts = ggml_top_k(ctx0, probs, n_expert_used); // [n_tokens, num_experts_per_tok]
+        cb(selected_experts->src[0], "ffn_moe_argsort", il);
+
+        ggml_tensor * weights = ggml_get_rows(ctx0,
+                                              ggml_reshape_3d(ctx0, probs, 1, n_expert, n_tokens), selected_experts);
+        cb(weights, "ffn_moe_weights", il);
+
+        weights = ggml_reshape_2d(ctx0, weights, n_expert_used, n_tokens); // [n_tokens, num_experts_per_tok]
+
+        ggml_tensor * weights_sum = ggml_sum_rows(ctx0, weights);
+        cb(weights_sum, "ffn_moe_weights_sum", il);
+
+        weights = ggml_div(ctx0, weights, weights_sum); // [n_tokens, num_experts_per_tok]
+        cb(weights, "ffn_moe_weights_norm", il);
+
+        // compute expert outputs
+        ggml_tensor * moe_out = nullptr;
+
+        for (int i = 0; i < n_expert_used; ++i) {
+            ggml_tensor * cur_expert;
+
+            ggml_tensor * cur_up = ggml_mul_mat_id(ctx0, model.layers[il].ffn_up_exps, selected_experts, i, cur);
+            cb(cur_up, "ffn_moe_up", il);
+
+            ggml_tensor * gate = ggml_mul_mat_id(ctx0, model.layers[il].ffn_gate_exps, selected_experts, i, cur);
+            cb(gate, "ffn_moe_gate", il);
+
+            switch (type_op) {
+                case LLM_FFN_SILU:
+                {
+                    gate = ggml_silu(ctx0, gate);
+                    cb(gate, "ffn_moe_silu", il);
+                } break;
+                case LLM_FFN_GELU:
+                {
+                    gate = ggml_gelu(ctx0, gate);
+                    cb(gate, "ffn_moe_gelu", il);
+                } break;
+                default:
+                    GGML_ASSERT(false);
+            }
+
+            cur_expert = ggml_mul(ctx0, cur_up, gate);
+            cb(cur_expert, "ffn_moe_gate_par", il);
+
+            cur_expert = ggml_mul_mat_id(ctx0, model.layers[il].ffn_down_exps, selected_experts, i, cur_expert); // [n_tokens, n_embd]
+            cb(cur_expert, "ffn_moe_down", il);
+
+            cur_expert = ggml_mul(ctx0, cur_expert,
+                                  ggml_view_2d(ctx0, weights, 1, n_tokens, weights->nb[1], i*weights->nb[0]));
+            cb(cur_expert, "ffn_moe_weighted", il);
+
+            if (i == 0) {
+                moe_out = cur_expert;
+            } else {
+                moe_out = ggml_add(ctx0, moe_out, cur_expert);
+                cb(moe_out, "ffn_moe_out", il);
+            }
+        }
+
+        return moe_out;
+    }
+
     struct ggml_cgraph * build_baichuan() {
         struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
 
@@ -6967,63 +7047,7 @@ struct llm_build_context {
                     LLM_NORM_RMS, cb, il);
             cb(cur, "ffn_norm", il);
 
-            ggml_tensor * logits = ggml_mul_mat(ctx0, model.layers[il].ffn_gate_inp, cur); // [n_tokens, num_experts]
-            cb(logits, "ffn_moe_logits", il);
-
-            ggml_tensor * probs = ggml_soft_max(ctx0, logits); // [n_tokens, num_experts]
-            cb(probs, "ffn_moe_probs", il);
-
-            // select experts
-            ggml_tensor * selected_experts = ggml_top_k(ctx0, probs, n_expert_used); // [n_tokens, num_experts_per_tok]
-            cb(selected_experts->src[0], "ffn_moe_argsort", il);
-
-            ggml_tensor * weights = ggml_get_rows(ctx0,
-                    ggml_reshape_3d(ctx0, probs, 1, n_expert, n_tokens), selected_experts);
-            cb(weights, "ffn_moe_weights", il);
-
-            weights = ggml_reshape_2d(ctx0, weights, n_expert_used, n_tokens); // [n_tokens, num_experts_per_tok]
-
-            ggml_tensor * weights_sum = ggml_sum_rows(ctx0, weights);
-            cb(weights_sum, "ffn_moe_weights_sum", il);
-
-            weights = ggml_div(ctx0, weights, weights_sum); // [n_tokens, num_experts_per_tok]
-            cb(weights, "ffn_moe_weights_norm", il);
-
-            // compute expert outputs
-            ggml_tensor * moe_out = nullptr;
-
-            for (int i = 0; i < n_expert_used; ++i) {
-                ggml_tensor * cur_expert;
-
-                ggml_tensor * cur_up = ggml_mul_mat_id(ctx0, model.layers[il].ffn_up_exps, selected_experts, i, cur);
-                cb(cur_up, "ffn_moe_up", il);
-
-                ggml_tensor * cur_gate = ggml_mul_mat_id(ctx0, model.layers[il].ffn_gate_exps, selected_experts, i, cur);
-                cb(cur_gate, "ffn_moe_gate", il);
-
-                //GeLU
-                cur_gate = ggml_gelu(ctx0, cur_gate);
-                cb(cur_gate, "ffn_moe_gelu", il);
-
-                cur_expert = ggml_mul(ctx0, cur_up, cur_gate);
-                cb(cur_expert, "ffn_moe_gate_par", il);
-
-                cur_expert = ggml_mul_mat_id(ctx0, model.layers[il].ffn_down_exps, selected_experts, i, cur_expert); // [n_tokens, n_embd]
-                cb(cur_expert, "ffn_moe_down", il);
-
-                cur_expert = ggml_mul(ctx0, cur_expert,
-                        ggml_view_2d(ctx0, weights, 1, n_tokens, weights->nb[1], i*weights->nb[0]));
-                cb(cur_expert, "ffn_moe_weighted", il);
-
-                if (i == 0) {
-                    moe_out = cur_expert;
-                } else {
-                    moe_out = ggml_add(ctx0, moe_out, cur_expert);
-                    cb(moe_out, "ffn_moe_out", il);
-                }
-            }
-
-            cur = moe_out;
+            cur = build_moe_ffn(cur, n_tokens, LLM_FFN_GELU, il);
 
             // Grok
             // if layer_out_norm is present then apply it before adding the input
@@ -7071,6 +7095,126 @@ struct llm_build_context {
         return gf;
     }
 
+    struct ggml_cgraph * build_dbrx() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+        // mutable variable, needed during the last layer of the computation to skip unused tokens
+        int32_t n_tokens = this->n_tokens;
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * inpSA = inpL;
+
+            // norm
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                                 model.layers[il].attn_norm, NULL,
+                                 LLM_NORM, cb, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                struct ggml_tensor * Qcur = nullptr;
+                struct ggml_tensor * Kcur = nullptr;
+                struct ggml_tensor * Vcur = nullptr;
+
+                cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, cur);
+                cb(cur, "wqkv", il);
+
+                cur = ggml_clamp(ctx0, cur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv);
+                cb(cur, "wqkv_clamped", il);
+
+                Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
+                Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
+                Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
+
+                cb(Qcur, "Qcur", il);
+                cb(Kcur, "Kcur", il);
+                cb(Vcur, "Vcur", il);
+
+                Qcur = ggml_rope_custom(
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos,
+                    n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Qcur, "Qcur", il);
+
+                Kcur = ggml_rope_custom(
+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos,
+                    n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur", il);
+
+                cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
+                                   model.layers[il].wo, NULL,
+                                   Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                n_tokens = n_outputs;
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // feed-forward network
+            // MoE branch
+            cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                                 model.layers[il].attn_out_norm, NULL,
+                                 LLM_NORM, cb, il);
+            cb(cur, "attn_out_norm", il);
+
+            cur = build_moe_ffn(cur, n_tokens, LLM_FFN_SILU, il);
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cb(cur, "ffn_out", il);
+
+            ggml_tensor * layer_dir = lctx.cvec.tensor_for(il);
+            if (layer_dir != nullptr) {
+                cur = ggml_add(ctx0, cur, layer_dir);
+            }
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = llm_build_norm(ctx0, cur, hparams,
+                             model.output_norm, NULL,
+                             LLM_NORM, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        // lm_head
+        cur = ggml_mul_mat(ctx0, model.output, cur);
+
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
     struct ggml_cgraph * build_starcoder() {
         struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
 
@@ -9785,6 +9929,10 @@ static struct ggml_cgraph * llama_build_graph(
             {
                 result = llm.build_command_r();
             } break;
+        case LLM_ARCH_DBRX:
+            {
+                result = llm.build_dbrx();
+            } break;
         default:
             GGML_ASSERT(false);
     }
@@ -11121,7 +11269,7 @@ struct llm_tokenizer_bpe {
                 add_new_bigram(bigram.left, left_symbol.next);  // right side of current symbol
             }
 
-            // add the fnished tokens to the final list keeping correct order for next and prev
+            // add the finished tokens to the final list keeping correct order for next and prev
             for (auto & sym : symbols) {
                 if (sym.n > 0) {
                     sym.prev = final_prev_index;
@@ -11861,7 +12009,9 @@ static void llama_grammar_advance_stack(
         std::vector<std::vector<const llama_grammar_element *>> & new_stacks) {
 
     if (stack.empty()) {
-        new_stacks.emplace_back(stack);
+        if (std::find(new_stacks.begin(), new_stacks.end(), stack) == new_stacks.end()) {
+            new_stacks.emplace_back(stack);
+        }
         return;
     }
 
@@ -11898,7 +12048,10 @@ static void llama_grammar_advance_stack(
         }
         case LLAMA_GRETYPE_CHAR:
         case LLAMA_GRETYPE_CHAR_NOT:
-            new_stacks.emplace_back(stack);
+            if (std::find(new_stacks.begin(), new_stacks.end(), stack) == new_stacks.end()) {
+                // only add the stack if it's not a duplicate of one we already have
+                new_stacks.emplace_back(stack);
+            }
             break;
         default:
             // end of alternate (LLAMA_GRETYPE_END, LLAMA_GRETYPE_ALT) or middle of char range
@@ -11912,12 +12065,13 @@ static void llama_grammar_advance_stack(
 // be positioned at a character range (see `llama_grammar_advance_stack`), and
 // produces the N possible stacks if the given char is accepted at those
 // positions
-std::vector<std::vector<const llama_grammar_element *>> llama_grammar_accept(
+void llama_grammar_accept(
         const std::vector<std::vector<llama_grammar_element>>         & rules,
         const std::vector<std::vector<const llama_grammar_element *>> & stacks,
-        const uint32_t                                                  chr) {
+        const uint32_t                                                  chr,
+        std::vector<std::vector<const llama_grammar_element *>>       & new_stacks) {
 
-    std::vector<std::vector<const llama_grammar_element *>> new_stacks;
+    new_stacks.clear();
 
     for (const auto & stack : stacks) {
         if (stack.empty()) {
@@ -11936,8 +12090,6 @@ std::vector<std::vector<const llama_grammar_element *>> llama_grammar_accept(
             llama_grammar_advance_stack(rules, new_stack, new_stacks);
         }
     }
-
-    return new_stacks;
 }
 
 static std::vector<llama_grammar_candidate> llama_grammar_reject_candidates(
@@ -11951,6 +12103,7 @@ static std::vector<llama_grammar_candidate> llama_grammar_reject_candidates_for_
         const std::vector<llama_grammar_candidate>            & candidates) {
 
     std::vector<llama_grammar_candidate> rejects;
+    rejects.reserve(candidates.size());
 
     if (stack.empty()) {
         for (const auto & tok : candidates) {
@@ -11964,6 +12117,8 @@ static std::vector<llama_grammar_candidate> llama_grammar_reject_candidates_for_
     const llama_grammar_element * stack_pos = stack.back();
 
     std::vector<llama_grammar_candidate> next_candidates;
+    next_candidates.reserve(candidates.size());
+
     for (const auto & tok : candidates) {
         if (*tok.code_points == 0) {
             // reached end of full codepoints in token, reject iff it ended in a partial sequence
@@ -12771,8 +12926,10 @@ void llama_grammar_accept_token(struct llama_context * ctx, struct llama_grammar
     // Note terminating 0 in decoded string
     const auto   decoded     = decode_utf8(piece, grammar->partial_utf8);
     const auto & code_points = decoded.first;
+    std::vector<std::vector<const llama_grammar_element *>> tmp_new_stacks;
     for (auto it = code_points.begin(), end = code_points.end() - 1; it != end; ++it) {
-        grammar->stacks = llama_grammar_accept(grammar->rules, grammar->stacks, *it);
+        llama_grammar_accept(grammar->rules, grammar->stacks, *it, tmp_new_stacks);
+        grammar->stacks = tmp_new_stacks;
     }
     grammar->partial_utf8 = decoded.second;
     GGML_ASSERT(!grammar->stacks.empty());
@@ -13526,6 +13683,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
     gguf_set_kv     (ctx_out, ml.meta);
     gguf_set_val_u32(ctx_out, "general.quantization_version", GGML_QNT_VERSION);
     gguf_set_val_u32(ctx_out, "general.file_type", ftype);
+    // Remove split metadata
+    gguf_remove_key(ctx_out, ml.llm_kv(LLM_KV_SPLIT_NO).c_str());
+    gguf_remove_key(ctx_out, ml.llm_kv(LLM_KV_SPLIT_COUNT).c_str());
+    gguf_remove_key(ctx_out, ml.llm_kv(LLM_KV_SPLIT_TENSORS_COUNT).c_str());
 
     if (params->kv_overrides) {
         const std::vector<llama_model_kv_override> & overrides = *(const std::vector<llama_model_kv_override> *)params->kv_overrides;
@@ -14625,6 +14786,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
         // the pairs of head values are offset by n_rot/2
         case LLM_ARCH_FALCON:
         case LLM_ARCH_GROK:
+        case LLM_ARCH_DBRX:
         case LLM_ARCH_PERSIMMON:
         case LLM_ARCH_BERT:
         case LLM_ARCH_NOMIC_BERT:
@@ -16463,6 +16625,21 @@ static int32_t llama_chat_apply_template_internal(
         if (add_ass) {
             ss << "### Response:\n";
         }
+    } else if (tmpl == "command-r" || (tmpl.find("<|START_OF_TURN_TOKEN|>") != std::string::npos && tmpl.find("<|USER_TOKEN|>") != std::string::npos)) {
+        // CohereForAI/c4ai-command-r-plus
+        for (auto message : chat) {
+            std::string role(message->role);
+            if (role == "system") {
+                ss << "<|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|>" << trim(message->content) << "<|END_OF_TURN_TOKEN|>";
+            } else if (role == "user") {
+                ss << "<|START_OF_TURN_TOKEN|><|USER_TOKEN|>" << trim(message->content) << "<|END_OF_TURN_TOKEN|>";
+            } else if (role == "assistant") {
+                ss << "<|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>" << trim(message->content) << "<|END_OF_TURN_TOKEN|>";
+            }
+        }
+        if (add_ass) {
+            ss << "<|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>";
+        }
     } else {
         // template not supported
         return -1;

llama.h

@@ -1097,10 +1097,11 @@ const std::vector<std::pair<std::string, struct ggml_tensor *>> & llama_internal
     struct llama_context * ctx
 );
 
-std::vector<std::vector<const llama_grammar_element *>> llama_grammar_accept(
+void llama_grammar_accept(
         const std::vector<std::vector<llama_grammar_element>>         & rules,
         const std::vector<std::vector<const llama_grammar_element *>> & stacks,
-        const uint32_t                                                  chr);
+        const uint32_t                                                  chr,
+        std::vector<std::vector<const llama_grammar_element *>>       & new_stacks);
 
 std::pair<std::vector<uint32_t>, llama_partial_utf8> decode_utf8(
         const std::string & src,

scripts/get-wikitext-2.sh

@@ -1,10 +1,11 @@
 #!/bin/bash
 
 wget https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
+unzip wikitext-2-raw-v1.zip
 
 echo "Usage:"
 echo ""
-echo "  ./perplexity -m model.gguf -f wiki.test.raw [other params]"
+echo "  ./perplexity -m model.gguf -f wikitext-2-raw/wiki.test.raw [other params]"
 echo ""
 
 exit 0

scripts/hf.sh

@@ -3,9 +3,9 @@
 # Shortcut for downloading HF models
 #
 # Usage:
-#   ./main -m $(./examples/hf.sh https://huggingface.co/TheBloke/Mixtral-8x7B-v0.1-GGUF/resolve/main/mixtral-8x7b-v0.1.Q4_K_M.gguf)
-#   ./main -m $(./examples/hf.sh --url https://huggingface.co/TheBloke/Mixtral-8x7B-v0.1-GGUF/blob/main/mixtral-8x7b-v0.1.Q4_K_M.gguf)
-#   ./main -m $(./examples/hf.sh --repo TheBloke/Mixtral-8x7B-v0.1-GGUF --file mixtral-8x7b-v0.1.Q4_K_M.gguf)
+#   ./main -m $(./scripts/hf.sh https://huggingface.co/TheBloke/Mixtral-8x7B-v0.1-GGUF/resolve/main/mixtral-8x7b-v0.1.Q4_K_M.gguf)
+#   ./main -m $(./scripts/hf.sh --url https://huggingface.co/TheBloke/Mixtral-8x7B-v0.1-GGUF/blob/main/mixtral-8x7b-v0.1.Q4_K_M.gguf)
+#   ./main -m $(./scripts/hf.sh --repo TheBloke/Mixtral-8x7B-v0.1-GGUF --file mixtral-8x7b-v0.1.Q4_K_M.gguf)
 #
 
 # all logs go to stderr
@@ -14,7 +14,7 @@ function log {
 }
 
 function usage {
-    log "Usage: $0 [[--url] <url>] [--repo <repo>] [--file <file>] [-h|--help]"
+    log "Usage: $0 [[--url] <url>] [--repo <repo>] [--file <file>] [--outdir <dir> [-h|--help]"
     exit 1
 }
 
@@ -26,9 +26,9 @@ function has_cmd {
 }
 
 if has_cmd wget; then
-    cmd="wget -q --show-progress -c -O %s %s"
+    cmd="wget -q --show-progress -c -O %s/%s %s"
 elif has_cmd curl; then
-    cmd="curl -C - -f -o %s -L %s"
+    cmd="curl -C - -f --output-dir %s -o %s -L %s"
 else
     log "[E] curl or wget not found"
     exit 1
@@ -37,6 +37,7 @@ fi
 url=""
 repo=""
 file=""
+outdir="."
 
 # parse args
 while [[ $# -gt 0 ]]; do
@@ -53,6 +54,10 @@ while [[ $# -gt 0 ]]; do
             file="$2"
             shift 2
             ;;
+        --outdir)
+            outdir="$2"
+            shift 2
+            ;;
         -h|--help)
             usage
             ;;
@@ -94,10 +99,10 @@ basename=$(basename $url)
 log "[+] attempting to download $basename"
 
 if [ -n "$cmd" ]; then
-    cmd=$(printf "$cmd" "$basename" "$url")
+    cmd=$(printf "$cmd" "$outdir" "$basename" "$url")
     log "[+] $cmd"
     if $cmd; then
-        echo $basename
+        echo $outdir/$basename
         exit 0
     fi
 fi

tests/test-backend-ops.cpp

@@ -2014,6 +2014,7 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
             for (int n_mats : {2, 4, 8}) {
                 for (int id = 0; id < n_mats; id++) {
                     for (bool v : {false, true}) {
+                        test_cases.emplace_back(new test_mul_mat_id(type_a, type_b, n_mats, id, 16,  1, 256, v));
                         test_cases.emplace_back(new test_mul_mat_id(type_a, type_b, n_mats, id, 16, 16, 256, v));
                     }
                 }

tests/test-chat-template.cpp

@@ -45,6 +45,8 @@ int main(void) {
         // Orca-Vicuna
         // No template included in tokenizer_config.json, so this template likely needs to be manually set.
         "{%- for message in messages %}{%- if message['role'] == 'system' -%}{{-'SYSTEM: ' + message['content'] + '\n' -}}{%- else -%}{%- if message['role'] == 'user' -%}{{-'USER: ' + message['content'] + '\n'-}}{%- else -%}{{-'ASSISTANT: ' + message['content'] + '</s>\n' -}}{%- endif -%}{%- endif -%}{%- endfor -%}{%- if add_generation_prompt -%}{{-'ASSISTANT:'-}}{%- endif -%}",
+        // CohereForAI/c4ai-command-r-plus
+        "{{ bos_token }}{% if messages[0]['role'] == 'system' %}{% set loop_messages = messages[1:] %}{% set system_message = messages[0]['content'] %}{% elif false == true %}{% set loop_messages = messages %}{% set system_message = 'You are Command-R, a brilliant, sophisticated, AI-assistant trained to assist human users by providing thorough responses. You are trained by Cohere.' %}{% else %}{% set loop_messages = messages %}{% set system_message = false %}{% endif %}{% if system_message != false %}{{ '<|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|>' + system_message + '<|END_OF_TURN_TOKEN|>' }}{% endif %}{% for message in loop_messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% set content = message['content'] %}{% if message['role'] == 'user' %}{{ '<|START_OF_TURN_TOKEN|><|USER_TOKEN|>' + content.strip() + '<|END_OF_TURN_TOKEN|>' }}{% elif message['role'] == 'assistant' %}{{ '<|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>'  + content.strip() + '<|END_OF_TURN_TOKEN|>' }}{% endif %}{% endfor %}{% if add_generation_prompt %}{{ '<|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>' }}{% endif %}"
     };
     std::vector<std::string> expected_output = {
         // teknium/OpenHermes-2.5-Mistral-7B
@@ -69,6 +71,8 @@ int main(void) {
         "You are a helpful assistant\n\nUSER: Hello\nASSISTANT: Hi there</s>\nUSER: Who are you\nASSISTANT:    I am an assistant   </s>\nUSER: Another question\nASSISTANT:",
         // Orca-Vicuna
         "SYSTEM: You are a helpful assistant\nUSER: Hello\nASSISTANT: Hi there</s>\nUSER: Who are you\nASSISTANT:    I am an assistant   </s>\nUSER: Another question\nASSISTANT:",
+        // CohereForAI/c4ai-command-r-plus
+        "<|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|>You are a helpful assistant<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|USER_TOKEN|>Hello<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>Hi there<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|USER_TOKEN|>Who are you<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>I am an assistant<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|USER_TOKEN|>Another question<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>",
     };
     std::vector<char> formatted_chat(1024);
     int32_t res;

tests/test-grammar-integration.cpp

@@ -38,7 +38,7 @@ number ::= [0-9]+)""";
 
     for (auto it = code_points.begin(), end = code_points.end() - 1; it != end; ++it) {
         auto prev_stacks = grammar->stacks;
-        grammar->stacks = llama_grammar_accept(grammar->rules, grammar->stacks, *it);
+        llama_grammar_accept(grammar->rules, prev_stacks, *it, grammar->stacks);
         assert(!grammar->stacks.empty());
     }
 
@@ -138,7 +138,7 @@ ws ::= [ \t\n\r]?)""";
         for (auto it = code_points.begin(), end = code_points.end() - 1; it != end; ++it) {
             ++pos;
             auto prev_stacks = grammar->stacks;
-            grammar->stacks = llama_grammar_accept(grammar->rules, grammar->stacks, *it);
+            llama_grammar_accept(grammar->rules, prev_stacks, *it, grammar->stacks);
 
             // Expect that each code point will not cause the grammar to fail
             if (grammar->stacks.empty()) {
@@ -173,7 +173,7 @@ ws ::= [ \t\n\r]?)""";
 
         for (auto it = code_points.begin(), end = code_points.end() - 1; it != end; ++it) {
             auto prev_stacks = grammar->stacks;
-            grammar->stacks = llama_grammar_accept(grammar->rules, grammar->stacks, *it);
+            llama_grammar_accept(grammar->rules, prev_stacks, *it, grammar->stacks);
             if (grammar->stacks.empty()) {
                 parse_failed = true;
                 break;

tests/test-json-schema-to-grammar.cpp

@@ -104,16 +104,16 @@ static void test_all(const std::string & lang, std::function<void(const TestCase
         R"""(
             array ::= "[" space ( value ("," space value)* )? "]" space
             boolean ::= ("true" | "false") space
+            char ::= [^"\\] | "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
+            decimal-part ::= [0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9])?)?)?)?)?)?)?)?)?)?)?)?)?)?)?
+            integral-part ::= [0-9] | [1-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9])?)?)?)?)?)?)?)?)?)?)?)?)?)?)?
             null ::= "null" space
-            number ::= ("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space
+            number ::= ("-"? integral-part) ("." decimal-part)? ([eE] [-+]? integral-part)? space
             object ::= "{" space ( string ":" space value ("," space string ":" space value)* )? "}" space
             root ::= object
             space ::= " "?
-            string ::=  "\"" (
-                    [^"\\] |
-                    "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
-                    )* "\"" space
-            value ::= object | array | string | number | boolean
+            string ::= "\"" char* "\"" space
+            value ::= object | array | string | number | boolean | null
         )"""
     });
 
@@ -133,10 +133,13 @@ static void test_all(const std::string & lang, std::function<void(const TestCase
             date-string ::= "\"" date "\"" space
             date-time ::= date "T" time
             date-time-string ::= "\"" date-time "\"" space
-            root ::= "[" space date-string "," space uuid "," space time-string "," space date-time-string "]" space
+            root ::= "[" space tuple-0 "," space uuid "," space tuple-2 "," space tuple-3 "]" space
             space ::= " "?
             time ::= ([01] [0-9] | "2" [0-3]) ":" [0-5] [0-9] ":" [0-5] [0-9] ( "." [0-9] [0-9] [0-9] )? ( "Z" | ( "+" | "-" ) ( [01] [0-9] | "2" [0-3] ) ":" [0-5] [0-9] )
             time-string ::= "\"" time "\"" space
+            tuple-0 ::= date-string
+            tuple-2 ::= time-string
+            tuple-3 ::= date-time-string
             uuid ::= "\"" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] "-" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] "-" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] "-" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] "-" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] "\"" space
         )"""
     });
@@ -148,10 +151,65 @@ static void test_all(const std::string & lang, std::function<void(const TestCase
             "type": "string"
         })""",
         R"""(
-            root ::=  "\"" (
-                    [^"\\] |
-                    "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
-                    )* "\"" space
+            char ::= [^"\\] | "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
+            root ::= "\"" char* "\"" space
+            space ::= " "?
+        )"""
+    });
+
+    test({
+        SUCCESS,
+        "string w/ min length 1",
+        R"""({
+            "type": "string",
+            "minLength": 1
+        })""",
+        R"""(
+            char ::= [^"\\] | "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
+            root ::= "\"" char+ "\"" space
+            space ::= " "?
+        )"""
+    });
+
+    test({
+        SUCCESS,
+        "string w/ min length 3",
+        R"""({
+            "type": "string",
+            "minLength": 3
+        })""",
+        R"""(
+            char ::= [^"\\] | "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
+            root ::= "\"" char char char (char)* "\"" space
+            space ::= " "?
+        )"""
+    });
+
+    test({
+        SUCCESS,
+        "string w/ max length",
+        R"""({
+            "type": "string",
+            "maxLength": 3
+        })""",
+        R"""(
+            char ::= [^"\\] | "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
+            root ::= "\"" (char (char (char)?)?)? "\"" space
+            space ::= " "?
+        )"""
+    });
+
+    test({
+        SUCCESS,
+        "string w/ min & max length",
+        R"""({
+            "type": "string",
+            "minLength": 1,
+            "maxLength": 4
+        })""",
+        R"""(
+            char ::= [^"\\] | "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
+            root ::= "\"" char (char (char (char)?)?)? "\"" space
             space ::= " "?
         )"""
     });
@@ -175,7 +233,8 @@ static void test_all(const std::string & lang, std::function<void(const TestCase
             "type": "integer"
         })""",
         R"""(
-            root ::= ("-"? ([0-9] | [1-9] [0-9]*)) space
+            integral-part ::= [0-9] | [1-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9])?)?)?)?)?)?)?)?)?)?)?)?)?)?)?
+            root ::= ("-"? integral-part) space
             space ::= " "?
         )"""
     });
@@ -223,12 +282,10 @@ static void test_all(const std::string & lang, std::function<void(const TestCase
             "prefixItems": [{ "type": "string" }]
         })""",
         R"""(
+            char ::= [^"\\] | "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
             root ::= "[" space string "]" space
             space ::= " "?
-            string ::=  "\"" (
-                    [^"\\] |
-                    "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
-                    )* "\"" space
+            string ::= "\"" char* "\"" space
         )"""
     });
 
@@ -239,13 +296,13 @@ static void test_all(const std::string & lang, std::function<void(const TestCase
             "prefixItems": [{ "type": "string" }, { "type": "number" }]
         })""",
         R"""(
-            number ::= ("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space
+            char ::= [^"\\] | "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
+            decimal-part ::= [0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9])?)?)?)?)?)?)?)?)?)?)?)?)?)?)?
+            integral-part ::= [0-9] | [1-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9])?)?)?)?)?)?)?)?)?)?)?)?)?)?)?
+            number ::= ("-"? integral-part) ("." decimal-part)? ([eE] [-+]? integral-part)? space
             root ::= "[" space string "," space number "]" space
             space ::= " "?
-            string ::=  "\"" (
-                    [^"\\] |
-                    "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
-                    )* "\"" space
+            string ::= "\"" char* "\"" space
         )"""
     });
 
@@ -256,7 +313,9 @@ static void test_all(const std::string & lang, std::function<void(const TestCase
             "type": "number"
         })""",
         R"""(
-            root ::= ("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space
+            decimal-part ::= [0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9])?)?)?)?)?)?)?)?)?)?)?)?)?)?)?
+            integral-part ::= [0-9] | [1-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9])?)?)?)?)?)?)?)?)?)?)?)?)?)?)?
+            root ::= ("-"? integral-part) ("." decimal-part)? ([eE] [-+]? integral-part)? space
             space ::= " "?
         )"""
     });
@@ -272,7 +331,7 @@ static void test_all(const std::string & lang, std::function<void(const TestCase
         })""",
         R"""(
             boolean ::= ("true" | "false") space
-            root ::= "[" space boolean ( "," space boolean )( "," space boolean )* "]" space
+            root ::= "[" space boolean "," space boolean ("," space boolean)* "]" space
             space ::= " "?
         )"""
     });
@@ -288,7 +347,7 @@ static void test_all(const std::string & lang, std::function<void(const TestCase
         })""",
         R"""(
             boolean ::= ("true" | "false") space
-            root ::= "[" space ( boolean  )? "]" space
+            root ::= "[" space (boolean)? "]" space
             space ::= " "?
         )"""
     });
@@ -304,7 +363,7 @@ static void test_all(const std::string & lang, std::function<void(const TestCase
         })""",
         R"""(
             boolean ::= ("true" | "false") space
-            root ::= "[" space ( boolean ( "," space boolean )? )? "]" space
+            root ::= "[" space (boolean ("," space boolean)?)? "]" space
             space ::= " "?
         )"""
     });
@@ -320,10 +379,12 @@ static void test_all(const std::string & lang, std::function<void(const TestCase
             "maxItems": 5
         })""",
         R"""(
-            integer ::= ("-"? ([0-9] | [1-9] [0-9]*)) space
+            decimal-part ::= [0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9])?)?)?)?)?)?)?)?)?)?)?)?)?)?)?
+            integer ::= ("-"? integral-part) space
+            integral-part ::= [0-9] | [1-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9])?)?)?)?)?)?)?)?)?)?)?)?)?)?)?
             item ::= number | integer
-            number ::= ("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space
-            root ::= "[" space item ( "," space item )( "," space item )( "," space item )?( "," space item )? "]" space
+            number ::= ("-"? integral-part) ("." decimal-part)? ([eE] [-+]? integral-part)? space
+            root ::= "[" space item "," space item "," space item ("," space item ("," space item)?)? "]" space
             space ::= " "?
         )"""
     });
@@ -372,11 +433,11 @@ static void test_all(const std::string & lang, std::function<void(const TestCase
         "regexp",
         R"""({
             "type": "string",
-            "pattern": "^(\\([0-9]{1,3}\\))?[0-9]{3}-[0-9]{4} and...$"
+            "pattern": "^(\\([0-9]{1,3}\\))?[0-9]{3}-[0-9]{4} a{3,5}nd...$"
         })""",
         R"""(
-            dot ::= [\U00000000-\x09\x0B\x0C\x0E-\U0010FFFF]
-            root ::= "\"" ("(" root-1 root-1? root-1? ")")? root-1 root-1 root-1 "-" root-1 root-1 root-1 root-1 " and" dot dot dot "\"" space
+            dot ::= [^\x0A\x0D]
+            root ::= "\"" ("(" root-1 (root-1 (root-1)?)? ")")? root-1 root-1 root-1 "-" root-1 root-1 root-1 root-1 " " "aaa" ("a" ("a")?)? "nd" dot dot dot "\"" space
             root-1 ::= [0-9]
             space ::= " "?
         )"""
@@ -404,12 +465,10 @@ static void test_all(const std::string & lang, std::function<void(const TestCase
             a-kv ::= "\"a\"" space ":" space string
             b-kv ::= "\"b\"" space ":" space string
             c-kv ::= "\"c\"" space ":" space string
+            char ::= [^"\\] | "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
             root ::= "{" space b-kv "," space c-kv "," space a-kv "}" space
             space ::= " "?
-            string ::=  "\"" (
-                    [^"\\] |
-                    "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
-                    )* "\"" space
+            string ::= "\"" char* "\"" space
         )"""
     });
 
@@ -426,12 +485,10 @@ static void test_all(const std::string & lang, std::function<void(const TestCase
         })""",
         R"""(
             a-kv ::= "\"a\"" space ":" space string
+            char ::= [^"\\] | "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
             root ::= "{" space  (a-kv )? "}" space
             space ::= " "?
-            string ::=  "\"" (
-                    [^"\\] |
-                    "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
-                    )* "\"" space
+            string ::= "\"" char* "\"" space
         )"""
     });
 
@@ -452,12 +509,10 @@ static void test_all(const std::string & lang, std::function<void(const TestCase
             b-kv ::= "\"b\"" space ":" space string
             b-rest ::= ( "," space c-kv )?
             c-kv ::= "\"c\"" space ":" space string
+            char ::= [^"\\] | "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
             root ::= "{" space  (a-kv a-rest | b-kv b-rest | c-kv )? "}" space
             space ::= " "?
-            string ::=  "\"" (
-                    [^"\\] |
-                    "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
-                    )* "\"" space
+            string ::= "\"" char* "\"" space
         )"""
     });
 
@@ -478,14 +533,12 @@ static void test_all(const std::string & lang, std::function<void(const TestCase
             a-kv ::= "\"a\"" space ":" space string
             b-kv ::= "\"b\"" space ":" space string
             c-kv ::= "\"c\"" space ":" space string
+            char ::= [^"\\] | "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
             d-kv ::= "\"d\"" space ":" space string
             d-rest ::= ( "," space c-kv )?
             root ::= "{" space b-kv "," space a-kv ( "," space ( d-kv d-rest | c-kv ) )? "}" space
             space ::= " "?
-            string ::=  "\"" (
-                    [^"\\] |
-                    "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
-                )* "\"" space
+            string ::= "\"" char* "\"" space
         )"""
     });
 
@@ -499,14 +552,14 @@ static void test_all(const std::string & lang, std::function<void(const TestCase
         R"""(
             additional-kv ::= string ":" space additional-value
             additional-kvs ::= additional-kv ( "," space additional-kv )*
-            additional-value ::= "[" space ( number ( "," space number )* )? "]" space
-            number ::= ("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space
+            additional-value ::= "[" space (number ("," space number)*)? "]" space
+            char ::= [^"\\] | "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
+            decimal-part ::= [0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9])?)?)?)?)?)?)?)?)?)?)?)?)?)?)?
+            integral-part ::= [0-9] | [1-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9])?)?)?)?)?)?)?)?)?)?)?)?)?)?)?
+            number ::= ("-"? integral-part) ("." decimal-part)? ([eE] [-+]? integral-part)? space
             root ::= "{" space  (additional-kvs )? "}" space
             space ::= " "?
-            string ::=  "\"" (
-                    [^"\\] |
-                    "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
-                    )* "\"" space
+            string ::= "\"" char* "\"" space
         )"""
     });
 
@@ -520,16 +573,16 @@ static void test_all(const std::string & lang, std::function<void(const TestCase
         R"""(
             array ::= "[" space ( value ("," space value)* )? "]" space
             boolean ::= ("true" | "false") space
+            char ::= [^"\\] | "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
+            decimal-part ::= [0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9])?)?)?)?)?)?)?)?)?)?)?)?)?)?)?
+            integral-part ::= [0-9] | [1-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9])?)?)?)?)?)?)?)?)?)?)?)?)?)?)?
             null ::= "null" space
-            number ::= ("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space
+            number ::= ("-"? integral-part) ("." decimal-part)? ([eE] [-+]? integral-part)? space
             object ::= "{" space ( string ":" space value ("," space string ":" space value)* )? "}" space
             root ::= object
             space ::= " "?
-            string ::=  "\"" (
-                    [^"\\] |
-                    "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
-                    )* "\"" space
-            value ::= object | array | string | number | boolean
+            string ::= "\"" char* "\"" space
+            value ::= object | array | string | number | boolean | null
         )"""
     });
 
@@ -542,16 +595,16 @@ static void test_all(const std::string & lang, std::function<void(const TestCase
         R"""(
             array ::= "[" space ( value ("," space value)* )? "]" space
             boolean ::= ("true" | "false") space
+            char ::= [^"\\] | "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
+            decimal-part ::= [0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9])?)?)?)?)?)?)?)?)?)?)?)?)?)?)?
+            integral-part ::= [0-9] | [1-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9])?)?)?)?)?)?)?)?)?)?)?)?)?)?)?
             null ::= "null" space
-            number ::= ("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space
+            number ::= ("-"? integral-part) ("." decimal-part)? ([eE] [-+]? integral-part)? space
             object ::= "{" space ( string ":" space value ("," space string ":" space value)* )? "}" space
             root ::= object
             space ::= " "?
-            string ::=  "\"" (
-                    [^"\\] |
-                    "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
-                    )* "\"" space
-            value ::= object | array | string | number | boolean
+            string ::= "\"" char* "\"" space
+            value ::= object | array | string | number | boolean | null
         )"""
     });
 
@@ -583,13 +636,13 @@ static void test_all(const std::string & lang, std::function<void(const TestCase
             a-kv ::= "\"a\"" space ":" space number
             additional-kv ::= string ":" space string
             additional-kvs ::= additional-kv ( "," space additional-kv )*
-            number ::= ("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space
+            char ::= [^"\\] | "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
+            decimal-part ::= [0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9])?)?)?)?)?)?)?)?)?)?)?)?)?)?)?
+            integral-part ::= [0-9] | [1-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9])?)?)?)?)?)?)?)?)?)?)?)?)?)?)?
+            number ::= ("-"? integral-part) ("." decimal-part)? ([eE] [-+]? integral-part)? space
             root ::= "{" space a-kv ( "," space ( additional-kvs ) )? "}" space
             space ::= " "?
-            string ::=  "\"" (
-                    [^"\\] |
-                    "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
-                    )* "\"" space
+            string ::= "\"" char* "\"" space
         )"""
     });
 
@@ -608,13 +661,13 @@ static void test_all(const std::string & lang, std::function<void(const TestCase
             a-rest ::= additional-kvs
             additional-kv ::= string ":" space number
             additional-kvs ::= additional-kv ( "," space additional-kv )*
-            number ::= ("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space
+            char ::= [^"\\] | "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
+            decimal-part ::= [0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9])?)?)?)?)?)?)?)?)?)?)?)?)?)?)?
+            integral-part ::= [0-9] | [1-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9])?)?)?)?)?)?)?)?)?)?)?)?)?)?)?
+            number ::= ("-"? integral-part) ("." decimal-part)? ([eE] [-+]? integral-part)? space
             root ::= "{" space  (a-kv a-rest | additional-kvs )? "}" space
             space ::= " "?
-            string ::=  "\"" (
-                    [^"\\] |
-                    "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
-                    )* "\"" space
+            string ::= "\"" char* "\"" space
         )"""
     });
 
@@ -636,13 +689,13 @@ static void test_all(const std::string & lang, std::function<void(const TestCase
             additional-kvs ::= additional-kv ( "," space additional-kv )*
             b-kv ::= "\"b\"" space ":" space number
             b-rest ::= additional-kvs
-            number ::= ("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space
+            char ::= [^"\\] | "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
+            decimal-part ::= [0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9])?)?)?)?)?)?)?)?)?)?)?)?)?)?)?
+            integral-part ::= [0-9] | [1-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9])?)?)?)?)?)?)?)?)?)?)?)?)?)?)?
+            number ::= ("-"? integral-part) ("." decimal-part)? ([eE] [-+]? integral-part)? space
             root ::= "{" space a-kv ( "," space ( b-kv b-rest | additional-kvs ) )? "}" space
             space ::= " "?
-            string ::=  "\"" (
-                    [^"\\] |
-                    "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
-                    )* "\"" space
+            string ::= "\"" char* "\"" space
         )"""
     });
 
@@ -650,9 +703,9 @@ static void test_all(const std::string & lang, std::function<void(const TestCase
         SUCCESS,
         "top-level $ref",
         R"""({
-            "$ref": "#/definitions/MyType",
+            "$ref": "#/definitions/foo",
             "definitions": {
-                "MyType": {
+                "foo": {
                     "type": "object",
                     "properties": {
                         "a": {
@@ -667,14 +720,12 @@ static void test_all(const std::string & lang, std::function<void(const TestCase
             }
         })""",
         R"""(
-            MyType ::= "{" space MyType-a-kv "}" space
-            MyType-a-kv ::= "\"a\"" space ":" space string
-            root ::= MyType
+            char ::= [^"\\] | "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
+            foo ::= "{" space foo-a-kv "}" space
+            foo-a-kv ::= "\"a\"" space ":" space string
+            root ::= foo
             space ::= " "?
-            string ::=  "\"" (
-                    [^"\\] |
-                    "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
-                    )* "\"" space
+            string ::= "\"" char* "\"" space
         )"""
     });
 
@@ -701,9 +752,11 @@ static void test_all(const std::string & lang, std::function<void(const TestCase
             alternative-1 ::= bar
             bar ::= "{" space  (bar-b-kv )? "}" space
             bar-b-kv ::= "\"b\"" space ":" space number
+            decimal-part ::= [0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9])?)?)?)?)?)?)?)?)?)?)?)?)?)?)?
             foo ::= "{" space  (foo-a-kv )? "}" space
             foo-a-kv ::= "\"a\"" space ":" space number
-            number ::= ("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space
+            integral-part ::= [0-9] | [1-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9])?)?)?)?)?)?)?)?)?)?)?)?)?)?)?
+            number ::= ("-"? integral-part) ("." decimal-part)? ([eE] [-+]? integral-part)? space
             root ::= alternative-0 | alternative-1
             space ::= " "?
         )"""
@@ -745,7 +798,9 @@ static void test_all(const std::string & lang, std::function<void(const TestCase
             c-kv ::= "\"c\"" space ":" space number
             d-kv ::= "\"d\"" space ":" space number
             d-rest ::= ( "," space c-kv )?
-            number ::= ("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space
+            decimal-part ::= [0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9])?)?)?)?)?)?)?)?)?)?)?)?)?)?)?
+            integral-part ::= [0-9] | [1-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9])?)?)?)?)?)?)?)?)?)?)?)?)?)?)?
+            number ::= ("-"? integral-part) ("." decimal-part)? ([eE] [-+]? integral-part)? space
             root ::= "{" space a-kv "," space b-kv ( "," space ( d-kv d-rest | c-kv ) )? "}" space
             space ::= " "?
         )"""
@@ -786,7 +841,9 @@ static void test_all(const std::string & lang, std::function<void(const TestCase
             "definitions": {}
         })""",
         R"""(
-            number ::= ("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space
+            decimal-part ::= [0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9])?)?)?)?)?)?)?)?)?)?)?)?)?)?)?
+            integral-part ::= [0-9] | [1-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9] ([0-9])?)?)?)?)?)?)?)?)?)?)?)?)?)?)?
+            number ::= ("-"? integral-part) ("." decimal-part)? ([eE] [-+]? integral-part)? space
             number- ::= "{" space number-number-kv "}" space
             number-kv ::= "\"number\"" space ":" space number-
             number-number ::= "{" space number-number-root-kv "}" space
@@ -816,7 +873,7 @@ int main() {
         test_all("Python", [](const TestCase & tc) {
             write("test-json-schema-input.tmp", tc.schema);
             tc.verify_status(std::system(
-                "python ./examples/json-schema-to-grammar.py test-json-schema-input.tmp > test-grammar-output.tmp") == 0 ? SUCCESS : FAILURE);
+                "python ./examples/json_schema_to_grammar.py test-json-schema-input.tmp > test-grammar-output.tmp") == 0 ? SUCCESS : FAILURE);
             tc.verify(read("test-grammar-output.tmp"));
         });
     } else {