Merge branch 'ollama:main' into main

2025-12-21 22:33:56 +00:00 · 2025-09-13 10:45:51 +08:00
parent 501cb38b8c 92b96d54ef
commit 9f3f80891d
28 changed files with 520 additions and 129 deletions
--- a/.github/workflows/release.yaml
+++ b/.github/workflows/release.yaml
@@ -65,14 +65,36 @@ jobs:
            arch: amd64
            preset: 'CUDA 12'
            install: https://developer.download.nvidia.com/compute/cuda/12.8.0/local_installers/cuda_12.8.0_571.96_windows.exe
+            cuda-components:
+              - '"cudart"'
+              - '"nvcc"'
+              - '"cublas"'
+              - '"cublas_dev"'
            cuda-version: '12.8'
            flags: ''
+            runner_dir: 'cuda_v12'
+          - os: windows
+            arch: amd64
+            preset: 'CUDA 13'
+            install: https://developer.download.nvidia.com/compute/cuda/13.0.0/local_installers/cuda_13.0.0_windows.exe
+            cuda-components:
+              - '"cudart"'
+              - '"nvcc"'
+              - '"cublas"'
+              - '"cublas_dev"'
+              - '"crt"'
+              - '"nvvm"'
+              - '"nvptxcompiler"'
+            cuda-version: '13.0'
+            flags: ''
+            runner_dir: 'cuda_v13'
          - os: windows
            arch: amd64
            preset: 'ROCm 6'
            install: https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q4-WinSvr2022-For-HIP.exe
            rocm-version: '6.2'
            flags: '-DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_C_FLAGS="-parallel-jobs=4 -Wno-ignored-attributes -Wno-deprecated-pragma" -DCMAKE_CXX_FLAGS="-parallel-jobs=4 -Wno-ignored-attributes -Wno-deprecated-pragma"'
+            runner_dir: ''
    runs-on: ${{ matrix.arch == 'arm64' && format('{0}-{1}', matrix.os, matrix.arch) || matrix.os }}
    environment: release
    env:
@@ -96,7 +118,7 @@ jobs:
          $ErrorActionPreference = "Stop"
          if ("${{ steps.cache-install.outputs.cache-hit }}" -ne 'true') {
            Invoke-WebRequest -Uri "${{ matrix.install }}" -OutFile "install.exe"
-            $subpackages = @("cudart", "nvcc", "cublas", "cublas_dev") | Foreach-Object {"${_}_${{ matrix.cuda-version }}"}
+            $subpackages = @(${{ join(matrix.cuda-components, ', ') }}) | Foreach-Object {"${_}_${{ matrix.cuda-version }}"}
            Start-Process -FilePath .\install.exe -ArgumentList (@("-s") + $subpackages) -NoNewWindow -Wait
          }

@@ -138,7 +160,7 @@ jobs:
        run: |
          Import-Module 'C:\Program Files\Microsoft Visual Studio\2022\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
          Enter-VsDevShell -VsInstallPath 'C:\Program Files\Microsoft Visual Studio\2022\Enterprise' -SkipAutomaticLocation  -DevCmdArguments '-arch=x64 -no_logo'
-          cmake --preset "${{ matrix.preset }}" ${{ matrix.flags }}
+          cmake --preset "${{ matrix.preset }}" ${{ matrix.flags }} -DOLLAMA_RUNNER_DIR="${{ matrix.runner_dir }}"
          cmake --build --parallel --preset "${{ matrix.preset }}"
          cmake --install build --component "${{ startsWith(matrix.preset, 'CUDA ') && 'CUDA' || startsWith(matrix.preset, 'ROCm ') && 'HIP' || 'CPU' }}" --strip --parallel 8
        env:
@@ -232,7 +254,7 @@ jobs:
            case "$COMPONENT" in
              bin/ollama)                echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
              lib/ollama/*.so*)          echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
-              lib/ollama/cuda_sbsa)      echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
+              lib/ollama/cuda_v*)        echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
              lib/ollama/cuda_jetpack5)  echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}-jetpack5.tar.in ;;
              lib/ollama/cuda_jetpack6)  echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}-jetpack6.tar.in ;;
              lib/ollama/rocm)           echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}-rocm.tar.in ;;
--- a/.github/workflows/test.yaml
+++ b/.github/workflows/test.yaml
@@ -46,7 +46,7 @@ jobs:
        include:
          - preset: CPU
          - preset: CUDA
-            container: nvidia/cuda:12.8.1-devel-ubuntu22.04
+            container: nvidia/cuda:13.0.0-devel-ubuntu22.04
            flags: '-DCMAKE_CUDA_ARCHITECTURES=87'
          - preset: ROCm
            container: rocm/dev-ubuntu-22.04:6.1.2
@@ -78,8 +78,17 @@ jobs:
        include:
          - preset: CPU
          - preset: CUDA
-            install: https://developer.download.nvidia.com/compute/cuda/12.8.0/local_installers/cuda_12.8.0_571.96_windows.exe
+            install: https://developer.download.nvidia.com/compute/cuda/13.0.0/local_installers/cuda_13.0.0_windows.exe
            flags: '-DCMAKE_CUDA_ARCHITECTURES=80'
+            cuda-components:
+              - '"cudart"'
+              - '"nvcc"'
+              - '"cublas"'
+              - '"cublas_dev"'
+              - '"crt"'
+              - '"nvvm"'
+              - '"nvptxcompiler"'
+            cuda-version: '13.0'
          - preset: ROCm
            install: https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q4-WinSvr2022-For-HIP.exe
            flags: '-DAMDGPU_TARGETS=gfx1010 -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_C_FLAGS="-parallel-jobs=4 -Wno-ignored-attributes -Wno-deprecated-pragma" -DCMAKE_CXX_FLAGS="-parallel-jobs=4 -Wno-ignored-attributes -Wno-deprecated-pragma"'
@@ -102,7 +111,8 @@ jobs:
          $ErrorActionPreference = "Stop"
          if ("${{ steps.cache-install.outputs.cache-hit }}" -ne 'true') {
            Invoke-WebRequest -Uri "${{ matrix.install }}" -OutFile "install.exe"
-            Start-Process -FilePath .\install.exe -ArgumentList (@("-s", "cudart_12.8", "nvcc_12.8", "cublas_12.8", "cublas_dev_12.8")) -NoNewWindow -Wait
+            $subpackages = @(${{ join(matrix.cuda-components, ', ') }}) | Foreach-Object {"${_}_${{ matrix.cuda-version }}"}
+            Start-Process -FilePath .\install.exe -ArgumentList (@("-s") + $subpackages) -NoNewWindow -Wait
          }

          $cudaPath = (Resolve-Path "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\*").path
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -38,7 +38,7 @@ if (CMAKE_OSX_ARCHITECTURES MATCHES "x86_64")
 endif()

 set(OLLAMA_BUILD_DIR ${CMAKE_BINARY_DIR}/lib/ollama)
-set(OLLAMA_INSTALL_DIR ${CMAKE_INSTALL_PREFIX}/lib/ollama)
+set(OLLAMA_INSTALL_DIR ${CMAKE_INSTALL_PREFIX}/lib/ollama/${OLLAMA_RUNNER_DIR})

 set(CMAKE_RUNTIME_OUTPUT_DIRECTORY         ${OLLAMA_BUILD_DIR})
 set(CMAKE_RUNTIME_OUTPUT_DIRECTORY_DEBUG   ${OLLAMA_BUILD_DIR})
@@ -81,7 +81,7 @@ if(CMAKE_CUDA_COMPILER)
    add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-cuda)
    install(TARGETS ggml-cuda
        RUNTIME_DEPENDENCIES
-            DIRECTORIES ${CUDAToolkit_BIN_DIR} ${CUDAToolkit_LIBRARY_DIR}
+            DIRECTORIES ${CUDAToolkit_BIN_DIR} ${CUDAToolkit_BIN_DIR}/x64 ${CUDAToolkit_LIBRARY_DIR}
            PRE_INCLUDE_REGEXES cublas cublasLt cudart
            PRE_EXCLUDE_REGEXES ".*"
        RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT CUDA
--- a/CMakePresets.json
+++ b/CMakePresets.json
@@ -18,6 +18,14 @@
      "name": "CUDA",
      "inherits": [ "Default" ]
    },
+    {
+      "name": "CUDA 11",
+      "inherits": [ "CUDA" ],
+      "cacheVariables": {
+        "CMAKE_CUDA_ARCHITECTURES": "50-virtual;60-virtual;61-virtual;70-virtual;75-virtual;80-virtual;86-virtual;87-virtual;89-virtual;90-virtual",
+        "CMAKE_CUDA_FLAGS": "-Wno-deprecated-gpu-targets -t 2"
+      }
+    },
    {
      "name": "CUDA 12",
      "inherits": [ "CUDA" ],
@@ -26,6 +34,14 @@
        "CMAKE_CUDA_FLAGS": "-Wno-deprecated-gpu-targets -t 2"
      }
    },
+    {
+      "name": "CUDA 13",
+      "inherits": [ "CUDA" ],
+      "cacheVariables": {
+        "CMAKE_CUDA_ARCHITECTURES": "75-virtual;80-virtual;86-virtual;87-virtual;89-virtual;90-virtual;90a-virtual;100-virtual;110-virtual;120-virtual;121-virtual",
+        "CMAKE_CUDA_FLAGS": "-t 2"
+      }
+    },
    {
      "name": "JetPack 5",
      "inherits": [ "CUDA" ],
@@ -72,11 +88,21 @@
      "configurePreset": "CUDA",
      "targets": [ "ggml-cuda" ]
    },
+    {
+      "name": "CUDA 11",
+      "inherits": [ "CUDA" ],
+      "configurePreset": "CUDA 11"
+    },
    {
      "name": "CUDA 12",
      "inherits": [ "CUDA" ],
      "configurePreset": "CUDA 12"
    },
+    {
+      "name": "CUDA 13",
+      "inherits": [ "CUDA" ],
+      "configurePreset": "CUDA 13"
+    },
    {
      "name": "JetPack 5",
      "inherits": [ "CUDA" ],
--- a/30
+++ b/30
@@ -39,15 +39,35 @@ RUN --mount=type=cache,target=/root/.ccache \
        && cmake --build --parallel --preset 'CPU' \
        && cmake --install build --component CPU --strip --parallel 8

+FROM base AS cuda-11
+ARG CUDA11VERSION=11.8
+RUN dnf install -y cuda-toolkit-${CUDA11VERSION//./-}
+ENV PATH=/usr/local/cuda-11/bin:$PATH
+RUN --mount=type=cache,target=/root/.ccache \
+    cmake --preset 'CUDA 11' -DOLLAMA_RUNNER_DIR="cuda_v11" \
+        && cmake --build --parallel --preset 'CUDA 11' \
+        && cmake --install build --component CUDA --strip --parallel 8
+
 FROM base AS cuda-12
 ARG CUDA12VERSION=12.8
 RUN dnf install -y cuda-toolkit-${CUDA12VERSION//./-}
 ENV PATH=/usr/local/cuda-12/bin:$PATH
 RUN --mount=type=cache,target=/root/.ccache \
-    cmake --preset 'CUDA 12' \
+    cmake --preset 'CUDA 12' -DOLLAMA_RUNNER_DIR="cuda_v12"\
        && cmake --build --parallel --preset 'CUDA 12' \
        && cmake --install build --component CUDA --strip --parallel 8

+
+FROM base AS cuda-13
+ARG CUDA13VERSION=13.0
+RUN dnf install -y cuda-toolkit-${CUDA13VERSION//./-}
+ENV PATH=/usr/local/cuda-13/bin:$PATH
+RUN --mount=type=cache,target=/root/.ccache \
+    cmake --preset 'CUDA 13' -DOLLAMA_RUNNER_DIR="cuda_v13" \
+        && cmake --build --parallel --preset 'CUDA 13' \
+        && cmake --install build --component CUDA --strip --parallel 8
+
+
 FROM base AS rocm-6
 ENV PATH=/opt/rocm/hcc/bin:/opt/rocm/hip/bin:/opt/rocm/bin:/opt/rocm/hcc/bin:$PATH
 RUN --mount=type=cache,target=/root/.ccache \
@@ -92,10 +112,14 @@ RUN --mount=type=cache,target=/root/.cache/go-build \
    go build -trimpath -buildmode=pie -o /bin/ollama .

 FROM --platform=linux/amd64 scratch AS amd64
-COPY --from=cuda-12 dist/lib/ollama /lib/ollama
+# COPY --from=cuda-11 dist/lib/ollama/ /lib/ollama/
+COPY --from=cuda-12 dist/lib/ollama /lib/ollama/
+COPY --from=cuda-13 dist/lib/ollama/ /lib/ollama/

 FROM --platform=linux/arm64 scratch AS arm64
-COPY --from=cuda-12 dist/lib/ollama /lib/ollama/cuda_sbsa
+# COPY --from=cuda-11 dist/lib/ollama/ /lib/ollama/
+COPY --from=cuda-12 dist/lib/ollama /lib/ollama/
+COPY --from=cuda-13 dist/lib/ollama/ /lib/ollama/
 COPY --from=jetpack-5 dist/lib/ollama /lib/ollama/cuda_jetpack5
 COPY --from=jetpack-6 dist/lib/ollama /lib/ollama/cuda_jetpack6

--- a/README.md
+++ b/README.md
@@ -435,6 +435,8 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Mayan EDMS](https://gitlab.com/mayan-edms/mayan-edms) (Open source document management system to organize, tag, search, and automate your files with powerful Ollama driven workflows.)
 - [Serene Pub](https://github.com/doolijb/serene-pub) (Beginner friendly, open source AI Roleplaying App for Windows, Mac OS and Linux. Search, download and use models with Ollama all inside the app.)
 - [Andes](https://github.com/aqerd/andes) (A Visual Studio Code extension that provides a local UI interface for Ollama models)
+- [Clueless](https://github.com/KashyapTan/clueless) (Open Source & Local Cluely: A desktop application LLM assistant to help you talk to anything on your screen using locally served Ollama models. Also undetectable to screenshare)
+- [ollama-co2](https://github.com/carbonatedWaterOrg/ollama-co2) (FastAPI web interface for monitoring and managing local and remote Ollama servers with real-time model monitoring and concurrent downloads)

 ### Cloud

--- a/api/types.go
+++ b/api/types.go
@@ -388,8 +388,12 @@ type EmbedRequest struct {
 	// this request.
 	KeepAlive *Duration `json:"keep_alive,omitempty"`

+	// Truncate truncates the input to fit the model's max sequence length.
 	Truncate *bool `json:"truncate,omitempty"`

+	// Dimensions truncates the output embedding to the specified dimension.
+	Dimensions int `json:"dimensions,omitempty"`
+
 	// Options lists model-specific options.
 	Options map[string]any `json:"options"`
 }
--- a/cmd/cmd.go
+++ b/cmd/cmd.go
@@ -56,10 +56,8 @@ func ensureThinkingSupport(ctx context.Context, client *api.Client, name string)
 	if err != nil {
 		return
 	}
-	for _, cap := range resp.Capabilities {
-		if cap == model.CapabilityThinking {
-			return
-		}
+	if slices.Contains(resp.Capabilities, model.CapabilityThinking) {
+		return
 	}
 	fmt.Fprintf(os.Stderr, "warning: model %q does not support thinking output\n", name)
 }
--- a/discover/cuda_common.go
+++ b/discover/cuda_common.go
@@ -43,14 +43,15 @@ func cudaVariant(gpuInfo CudaGPUInfo) string {
 				}
 			}
 		}
-		return "sbsa"
 	}

-	// driver 12.0 has problems with the cuda v12 library, so run v11 on those older drivers
-	if gpuInfo.DriverMajor < 12 || (gpuInfo.DriverMajor == 12 && gpuInfo.DriverMinor == 0) {
-		// The detected driver is older than Feb 2023
-		slog.Warn("old CUDA driver detected - please upgrade to a newer driver", "version", fmt.Sprintf("%d.%d", gpuInfo.DriverMajor, gpuInfo.DriverMinor))
-		return "v11"
+	if gpuInfo.DriverMajor < 13 {
+		// The detected driver is older than 580 (Aug 2025)
+		// Warn if their CC is compatible with v13 and they should upgrade their driver to get better performance
+		if gpuInfo.computeMajor > 7 || (gpuInfo.computeMajor == 7 && gpuInfo.computeMinor >= 5) {
+			slog.Warn("old CUDA driver detected - please upgrade to a newer driver for best performance", "version", fmt.Sprintf("%d.%d", gpuInfo.DriverMajor, gpuInfo.DriverMinor))
+		}
+		return "v12"
 	}
-	return "v12"
+	return "v13"
 }
--- a/docs/api.md
+++ b/docs/api.md
@@ -1708,6 +1708,7 @@ Advanced parameters:
 - `truncate`: truncates the end of each input to fit within context length. Returns error if `false` and context length is exceeded. Defaults to `true`
 - `options`: additional model parameters listed in the documentation for the [Modelfile](./modelfile.md#valid-parameters-and-values) such as `temperature`
 - `keep_alive`: controls how long the model will stay loaded into memory following the request (default: `5m`)
+- `dimensions`: number of dimensions for the embedding

 ### Examples

--- a/docs/linux.md
+++ b/docs/linux.md
@@ -11,12 +11,13 @@ curl -fsSL https://ollama.com/install.sh | sh
 ## Manual install

 > [!NOTE]
-> If you are upgrading from a prior version, you should remove the old libraries with `sudo rm -rf /usr/lib/ollama` first.
+> If you are upgrading from a prior version, you **MUST** remove the old libraries with `sudo rm -rf /usr/lib/ollama` first.

 Download and extract the package:

 ```shell
 curl -LO https://ollama.com/download/ollama-linux-amd64.tgz
+sudo rm -rf /usr/lib/ollama
 sudo tar -C /usr -xzf ollama-linux-amd64.tgz
 ```

--- a/docs/troubleshooting.md
+++ b/docs/troubleshooting.md
@@ -92,6 +92,9 @@ If none of those resolve the problem, gather additional information and file an
 - Set `CUDA_ERROR_LEVEL=50` and try again to get more diagnostic logs
 - Check dmesg for any errors `sudo dmesg | grep -i nvrm` and `sudo dmesg | grep -i nvidia`

+You may get more details for initialization failures by enabling debug prints in the uvm driver.  You should only use this temporarily while troubleshooting
+- `sudo rmmod nvidia_uvm` then `sudo modprobe nvidia_uvm uvm_debug_prints=1`
+

 ## AMD GPU Discovery

--- a/envconfig/config.go
+++ b/envconfig/config.go
@@ -185,8 +185,6 @@ var (
 	ContextLength = Uint("OLLAMA_CONTEXT_LENGTH", 4096)
 	// Auth enables authentication between the Ollama client and server
 	UseAuth = Bool("OLLAMA_AUTH")
-	// Enable the new memory estimation logic
-	NewMemoryEstimates = Bool("OLLAMA_NEW_ESTIMATES")
 )

 func String(s string) func() string {
@@ -272,7 +270,6 @@ func AsMap() map[string]EnvVar {
 		"OLLAMA_MULTIUSER_CACHE":   {"OLLAMA_MULTIUSER_CACHE", MultiUserCache(), "Optimize prompt caching for multi-user scenarios"},
 		"OLLAMA_CONTEXT_LENGTH":    {"OLLAMA_CONTEXT_LENGTH", ContextLength(), "Context length to use unless otherwise specified (default: 4096)"},
 		"OLLAMA_NEW_ENGINE":        {"OLLAMA_NEW_ENGINE", NewEngine(), "Enable the new Ollama engine"},
-		"OLLAMA_NEW_ESTIMATES":     {"OLLAMA_NEW_ESTIMATES", NewMemoryEstimates(), "Enable the new memory estimation logic"},

 		// Informational
 		"HTTP_PROXY":  {"HTTP_PROXY", String("HTTP_PROXY")(), "HTTP proxy"},
--- a/fs/ggml/ggml.go
+++ b/fs/ggml/ggml.go
@@ -57,10 +57,28 @@ func (kv KV) EmbeddingLength() uint64 {
 	return uint64(kv.Uint("embedding_length"))
 }

+func (kv KV) HeadCount() []uint64 {
+	headCountDefault := uint32(1)
+	headCount := kv.UintOrArrayValueAsArray("attention.head_count", headCountDefault)
+	if len(headCount) == 1 {
+		headCountDefault = headCount[0]
+	}
+	nLayers := int(kv.BlockCount())
+	if len(headCount) > nLayers {
+		slog.Warn("got more elements of attention.head_count than layers", "len(headCount)", len(headCount), "layers", nLayers)
+	}
+	out := make([]uint64, nLayers)
+	for i := range nLayers {
+		if i >= len(headCount) {
+			out[i] = uint64(headCountDefault)
+		} else {
+			out[i] = uint64(headCount[i])
+		}
+	}
+	return out
+}
+
 func (kv KV) HeadCountMax() uint64 {
-	// TODO(drifkin): using the max value can cause an overestimation. In the
-	// future if array values become more popular, we can adapt the more invasive
-	// <https://github.com/ollama/ollama/pull/10225>
 	return uint64(kv.UintOrMaxArrayValue("attention.head_count", 1))
 }

@@ -68,6 +86,27 @@ func (kv KV) HeadCountMin() uint64 {
 	return uint64(kv.UintOrMinArrayValue("attention.head_count", 1))
 }

+func (kv KV) HeadCountKV() []uint64 {
+	headCountKVDefault := uint32(1)
+	headCountKV := kv.UintOrArrayValueAsArray("attention.head_count_kv", headCountKVDefault)
+	if len(headCountKV) == 1 {
+		headCountKVDefault = headCountKV[0]
+	}
+	nLayers := int(kv.BlockCount())
+	if len(headCountKV) > nLayers {
+		slog.Warn("got more elements of attention.head_count than layers", "len(headCountKV)", len(headCountKV), "layers", nLayers)
+	}
+	out := make([]uint64, nLayers)
+	for i := range nLayers {
+		if i >= len(headCountKV) {
+			out[i] = uint64(headCountKVDefault)
+		} else {
+			out[i] = uint64(headCountKV[i])
+		}
+	}
+	return out
+}
+
 func (kv KV) HeadCountKVMax() uint64 {
 	return uint64(kv.UintOrMaxArrayValue("attention.head_count_kv", 1))
 }
@@ -100,6 +139,26 @@ func (kv KV) ChatTemplate() string {
 	return kv.String("tokenizer.chat_template")
 }

+// ssm architecture parameters
+
+func (kv KV) SSMConvKernel() uint64 {
+	return uint64(kv.Uint("ssm.conv_kernel"))
+}
+
+func (kv KV) SSMInnerSize() uint64 {
+	return uint64(kv.Uint("ssm.inner_size"))
+}
+
+func (kv KV) SSMStateSize() uint64 {
+	return uint64(kv.Uint("ssm.state_size"))
+}
+
+func (kv KV) SSMGroupCount() uint64 {
+	return uint64(kv.Uint("ssm.group_count"))
+}
+
+// general types
+
 func (kv KV) String(key string, defaultValue ...string) string {
 	val, _ := keyValue(kv, key, append(defaultValue, "")...)
 	return val
@@ -131,22 +190,27 @@ func (kv KV) UintOrMinArrayValue(key string, defaultValue uint32) uint32 {
 }

 func (kv KV) UintOrArrayValue(key string, defaultValue uint32) (uint32, uint32) {
+	arrVal := kv.UintOrArrayValueAsArray(key, defaultValue)
+	return slices.Min(arrVal), slices.Max(arrVal)
+}
+
+func (kv KV) UintOrArrayValueAsArray(key string, defaultValue uint32) []uint32 {
 	if u32, ok := keyValue(kv, key, uint32(0)); ok {
-		return u32, u32
+		return []uint32{u32}
 	} else if u32s, ok := keyValue(kv, key, &array[uint32]{}); ok {
-		min := slices.Min(u32s.values)
-		max := slices.Max(u32s.values)
-		return min, max
+		return u32s.values
 	} else if i32s, ok := keyValue(kv, key, &array[int32]{}); ok {
-		min := slices.Min(i32s.values)
-		max := slices.Max(i32s.values)
-		if min < 0 || max < 0 {
-			slog.Warn("array values are unexpectedly negative", "key", key, "min", min, "max", max)
+		dst := make([]uint32, len(i32s.values))
+		for i, v := range i32s.values {
+			if v < 0 {
+				slog.Warn("array values are unexpectedly negative", "key", key, "i", i, "v", v)
+			}
+			dst[i] = uint32(v)
 		}
-		return uint32(min), uint32(max)
+		return dst
 	}

-	return defaultValue, defaultValue
+	return []uint32{defaultValue}
 }

 func (kv KV) Strings(key string, defaultValue ...[]string) []string {
@@ -486,7 +550,9 @@ func (f GGML) GraphSize(context, batch uint64, numParallel int, kvCacheType stri

 	embedding := f.KV().EmbeddingLength()
 	heads := f.KV().HeadCountMax()
+	headsArr := f.KV().HeadCount()
 	headsKV := f.KV().HeadCountKVMax()
+	headsKVArr := f.KV().HeadCountKV()
 	vocab := uint64(f.KV()["tokenizer.ggml.tokens"].(*array[string]).size)

 	embeddingHeads := f.KV().EmbeddingHeadCountMax()
@@ -496,12 +562,51 @@ func (f GGML) GraphSize(context, batch uint64, numParallel int, kvCacheType stri
 	layers := f.Tensors().GroupLayers()

 	bytesPerElement := kvCacheBytesPerElement(kvCacheType)
+
+	// Default for models unless special-cased below. These defaults mirror the
+	// cache usage in llama.cpp under the assumption that models without special
+	// cases below will use the llamarunner and caching will be handled by the
+	// llama.cpp layer.
+	//
+	// This also assumes that a layer without heads or headsKV set is recurrent
+	// which is usually the case. Some models (eg nemotronh) use "blocks" in
+	// place of layers where some are MLP blocks that don't have any cache.
+	// Models like this will need a special case below to be accurately
+	// estimated.
 	var kvTotal uint64
 	kv = make([]uint64, f.KV().BlockCount())
+	kvSizeAttn := uint64(0)
+	kvSizeRecurrent := uint64(0)
 	for i := range kv {
-		kv[i] = uint64(float64(context*(embeddingHeadsK+embeddingHeadsV)*headsKV) * bytesPerElement)
+		headsL := headsArr[i]
+		headsKVL := headsKVArr[i]
+		if headsL > 0 && headsKVL > 0 {
+			// full attention layer
+			// NOTE: Assumes uniform values for all attn layers
+			kv[i] = uint64(float64(context*(embeddingHeadsK+embeddingHeadsV)*headsKVL) * bytesPerElement)
+			kvSizeAttn += kv[i]
+		} else {
+			// recurrent layer
+			ssmDConv := f.KV().SSMConvKernel()
+			ssmDState := f.KV().SSMStateSize()
+			ssmDInner := f.KV().SSMInnerSize()
+			ssmNGroups := f.KV().SSMGroupCount()
+			nEmbdR := uint64(0)
+			if ssmDConv > 0 {
+				nEmbdR = (ssmDConv - 1) * (ssmDInner + 2*ssmNGroups*ssmDState)
+			}
+			nEmbdS := ssmDState * ssmDInner
+
+			// recurrent always uses F32 in llama.cpp backend
+			// https://github.com/ggml-org/llama.cpp/blob/master/src/llama-model.cpp#L18644
+			bytesPerElementRecurrent := kvCacheBytesPerElement("f32")
+
+			kv[i] = (nEmbdR + nEmbdS) * uint64(bytesPerElementRecurrent)
+			kvSizeRecurrent += kv[i]
+		}
 		kvTotal += kv[i]
 	}
+	slog.Debug("default cache size estimate", "attention MiB", float32(kvSizeAttn)/(1024.*1024.), "attention bytes", kvSizeAttn, "recurrent MiB", float32(kvSizeRecurrent)/(1024.*1024.), "recurrent bytes", kvSizeRecurrent)

 	switch f.KV().Architecture() {
 	case "llama", "llama4":
@@ -759,12 +864,16 @@ func (llm GGML) VisionGraphSize() (weights, graphSize uint64) {

 // SupportsKVCacheType checks if the requested cache type is supported
 func (f GGML) SupportsKVCacheType(cacheType string) bool {
+	if cacheType == "" || cacheType == "f16" {
+		return true
+	}
+
 	if arch := f.KV().Architecture(); slices.Contains([]string{"gptoss", "gpt-oss"}, arch) {
 		// gpt-oss uses attention with sinks which does not support quantized cache types
-		slog.Warn("model only supports non-quantized cache types ", "mode", arch)
-		return cacheType == "f16"
+		slog.Warn("model only supports non-quantized cache types", "model", arch)
+		return false
 	}
-	return slices.Contains([]string{"f16", "q8_0", "q4_0"}, cacheType)
+	return slices.Contains([]string{"q8_0", "q4_0"}, cacheType)
 }

 // SupportsFlashAttention checks if the model supports flash attention
@@ -774,6 +883,10 @@ func (f GGML) SupportsFlashAttention() bool {
 		return false
 	}

+	if arch := f.KV().Architecture(); slices.Contains([]string{"gemma2"}, arch) {
+		return false
+	}
+
 	// Check head counts match and are non-zero
 	headCountK := f.KV().EmbeddingHeadCountK()
 	headCountV := f.KV().EmbeddingHeadCountV()
@@ -794,6 +907,8 @@ func kvCacheBytesPerElement(cacheType string) float64 {
 		return 1 // 1/2 of fp16
 	case "q4_0":
 		return 0.5 // 1/4 of fp16
+	case "f32":
+		return 4 // f32 (default for recurrent)
 	default:
 		return 2 // f16 (default)
 	}
--- a/integration/api_test.go
+++ b/integration/api_test.go
@@ -410,3 +410,99 @@ func TestAPIEmbeddings(t *testing.T) {
 		t.Errorf("zero length embedding response")
 	}
 }
+
+func TestAPIToolCalling(t *testing.T) {
+	initialTimeout := 60 * time.Second
+	streamTimeout := 30 * time.Second
+	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute)
+	defer cancel()
+
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
+
+	modelName := "qwen3:0.6b"
+	if err := PullIfMissing(ctx, client, modelName); err != nil {
+		t.Fatalf("pull failed %s", err)
+	}
+
+	tools := []api.Tool{
+		{
+			Type: "function",
+			Function: api.ToolFunction{
+				Name:        "get_weather",
+				Description: "Get the current weather in a given location",
+				Parameters: api.ToolFunctionParameters{
+					Type:     "object",
+					Required: []string{"location"},
+					Properties: map[string]api.ToolProperty{
+						"location": {
+							Type:        api.PropertyType{"string"},
+							Description: "The city and state, e.g. San Francisco, CA",
+						},
+					},
+				},
+			},
+		},
+	}
+
+	req := api.ChatRequest{
+		Model: modelName,
+		Messages: []api.Message{
+			{
+				Role:    "user",
+				Content: "Call get_weather with location set to San Francisco.",
+			},
+		},
+		Tools: tools,
+		Options: map[string]any{
+			"temperature": 0,
+		},
+	}
+
+	stallTimer := time.NewTimer(initialTimeout)
+	var gotToolCall bool
+	var lastToolCall api.ToolCall
+
+	fn := func(response api.ChatResponse) error {
+		if len(response.Message.ToolCalls) > 0 {
+			gotToolCall = true
+			lastToolCall = response.Message.ToolCalls[len(response.Message.ToolCalls)-1]
+		}
+		if !stallTimer.Reset(streamTimeout) {
+			return fmt.Errorf("stall was detected while streaming response, aborting")
+		}
+		return nil
+	}
+
+	stream := true
+	req.Stream = &stream
+	done := make(chan int)
+	var genErr error
+	go func() {
+		genErr = client.Chat(ctx, &req, fn)
+		done <- 0
+	}()
+
+	select {
+	case <-stallTimer.C:
+		t.Errorf("tool-calling chat never started. Timed out after: %s", initialTimeout.String())
+	case <-done:
+		if genErr != nil {
+			t.Fatalf("chat failed: %v", genErr)
+		}
+
+		if !gotToolCall {
+			t.Fatalf("expected at least one tool call, got none")
+		}
+
+		if lastToolCall.Function.Name != "get_weather" {
+			t.Errorf("unexpected tool called: got %q want %q", lastToolCall.Function.Name, "get_weather")
+		}
+
+		if _, ok := lastToolCall.Function.Arguments["location"]; !ok {
+			t.Errorf("expected tool arguments to include 'location', got: %s", lastToolCall.Function.Arguments.String())
+		}
+	case <-ctx.Done():
+		t.Error("outer test context done while waiting for tool-calling chat")
+	}
+}
--- a/integration/concurrency_test.go
+++ b/integration/concurrency_test.go
@@ -121,6 +121,7 @@ func TestMultiModelStress(t *testing.T) {
 	// The intent is to go 1 over what can fit so we force the scheduler to thrash
 	targetLoadCount := 0
 	slog.Info("Loading models to find how many can fit in VRAM before overflowing")
+chooseModels:
 	for i, model := range chosenModels {
 		req := &api.GenerateRequest{Model: model}
 		slog.Info("loading", "model", model)
@@ -142,6 +143,13 @@ func TestMultiModelStress(t *testing.T) {
 				slog.Info("found model load capacity", "target", targetLoadCount, "current", loaded, "chosen", chosenModels[:targetLoadCount])
 				break
 			}
+			// Effectively limit model count to 2 on CPU only systems to avoid thrashing and timeouts
+			for _, m := range models.Models {
+				if m.SizeVRAM == 0 {
+					slog.Info("model running on CPU", "name", m.Name, "target", targetLoadCount, "chosen", chosenModels[:targetLoadCount])
+					break chooseModels
+				}
+			}
 		}
 	}
 	if targetLoadCount == len(chosenModels) {
--- a/integration/context_test.go
+++ b/integration/context_test.go
@@ -36,7 +36,7 @@ func TestLongInputContext(t *testing.T) {
 	if err := PullIfMissing(ctx, client, req.Model); err != nil {
 		t.Fatalf("PullIfMissing failed: %v", err)
 	}
-	DoGenerate(ctx, t, client, req, []string{"russia", "germany", "france", "england", "austria", "prussia", "individuals", "coalition", "conflict"}, 120*time.Second, 10*time.Second)
+	DoGenerate(ctx, t, client, req, []string{"russia", "germany", "france", "england", "austria", "prussia", "europe", "individuals", "coalition", "conflict"}, 120*time.Second, 10*time.Second)
 }

 func TestContextExhaustion(t *testing.T) {
@@ -50,7 +50,7 @@ func TestContextExhaustion(t *testing.T) {
 	// Set up the test data
 	req := api.GenerateRequest{
 		Model:  smol,
-		Prompt: "Write me a story with a ton of emojis?",
+		Prompt: "Write me a story in english with a lot of emojis",
 		Stream: &stream,
 		Options: map[string]any{
 			"temperature": 0,
--- a/integration/embed_test.go
+++ b/integration/embed_test.go
@@ -38,8 +38,9 @@ func TestAllMiniLMEmbeddings(t *testing.T) {
 	defer cleanup()

 	req := api.EmbeddingRequest{
-		Model:  "all-minilm",
-		Prompt: "why is the sky blue?",
+		Model:     "all-minilm",
+		Prompt:    "why is the sky blue?",
+		KeepAlive: &api.Duration{Duration: 10 * time.Second},
 	}

 	res, err := embeddingTestHelper(ctx, client, t, req)
--- a/integration/utils_test.go
+++ b/integration/utils_test.go
@@ -502,6 +502,22 @@ func DoGenerate(ctx context.Context, t *testing.T, client *api.Client, genReq ap
 		done <- 0
 	}()

+	var response string
+	verify := func() {
+		// Verify the response contains the expected data
+		response = buf.String()
+		atLeastOne := false
+		for _, resp := range anyResp {
+			if strings.Contains(strings.ToLower(response), resp) {
+				atLeastOne = true
+				break
+			}
+		}
+		if !atLeastOne {
+			t.Fatalf("%s: none of %v found in %s", genReq.Model, anyResp, response)
+		}
+	}
+
 	select {
 	case <-stallTimer.C:
 		if buf.Len() == 0 {
@@ -517,21 +533,14 @@ func DoGenerate(ctx context.Context, t *testing.T, client *api.Client, genReq ap
 		if genErr != nil {
 			t.Fatalf("%s failed with %s request prompt %s", genErr, genReq.Model, genReq.Prompt)
 		}
-		// Verify the response contains the expected data
-		response := buf.String()
-		atLeastOne := false
-		for _, resp := range anyResp {
-			if strings.Contains(strings.ToLower(response), resp) {
-				atLeastOne = true
-				break
-			}
-		}
-		if !atLeastOne {
-			t.Fatalf("%s: none of %v found in %s", genReq.Model, anyResp, response)
-		}
+		verify()
 		slog.Info("test pass", "model", genReq.Model, "prompt", genReq.Prompt, "contains", anyResp, "response", response)
 	case <-ctx.Done():
-		t.Error("outer test context done while waiting for generate")
+		// On slow systems, we might timeout before some models finish rambling, so check what we have so far to see
+		// if it's considered a pass - the stallTimer will detect hangs, but we want to consider slow systems a pass
+		// if they are still generating valid responses
+		slog.Warn("outer test context done while waiting for generate")
+		verify()
 	}
 	return context
 }
@@ -552,7 +561,7 @@ func GenerateRequests() ([]api.GenerateRequest, [][]string) {
 				KeepAlive: &api.Duration{Duration: 10 * time.Second},
 			}, {
 				Model:     smol,
-				Prompt:    "what is the origin of the US thanksgiving holiday? Be brief but factual in your reply",
+				Prompt:    "how do rainbows form? Be brief but factual in your reply",
 				Stream:    &stream,
 				KeepAlive: &api.Duration{Duration: 10 * time.Second},
 			}, {
@@ -570,9 +579,9 @@ func GenerateRequests() ([]api.GenerateRequest, [][]string) {
 		[][]string{
 			{"sunlight", "scattering", "interact", "color", "surface", "depth", "red", "orange", "yellow", "absorbs", "wavelength"},
 			{"soil", "organic", "earth", "black", "tan", "chemical", "processes", "pigments", "particles", "iron oxide", "rust", "air", "water", "mixture", "mixing"},
-			{"england", "english", "massachusetts", "pilgrims", "colonists", "independence", "british", "feast", "family", "gatherings", "traditions", "turkey", "colonial", "period", "harvest", "agricultural", "european settlers", "american revolution", "civil war", "16th century", "17th century", "native american", "united states", "cultural", "hardship", "autumn", "festival"},
+			{"water", "droplet", "refracted", "reflect", "color", "spectrum"},
 			{"fourth", "july", "declaration", "independence"},
-			{"nitrogen", "oxygen", "carbon", "dioxide"},
+			{"nitrogen", "oxygen", "carbon", "dioxide", "water", "vapor"},
 		}
 }

@@ -599,6 +608,22 @@ func DoChat(ctx context.Context, t *testing.T, client *api.Client, req api.ChatR
 		done <- 0
 	}()

+	var response string
+	verify := func() {
+		// Verify the response contains the expected data
+		response = buf.String()
+		atLeastOne := false
+		for _, resp := range anyResp {
+			if strings.Contains(strings.ToLower(response), resp) {
+				atLeastOne = true
+				break
+			}
+		}
+		if !atLeastOne {
+			t.Fatalf("%s: none of %v found in \"%s\" -- request was:%v", req.Model, anyResp, response, req.Messages)
+		}
+	}
+
 	select {
 	case <-stallTimer.C:
 		if buf.Len() == 0 {
@@ -614,23 +639,14 @@ func DoChat(ctx context.Context, t *testing.T, client *api.Client, req api.ChatR
 		if genErr != nil {
 			t.Fatalf("%s failed with %s request prompt %v", genErr, req.Model, req.Messages)
 		}
-
-		// Verify the response contains the expected data
-		response := buf.String()
-		atLeastOne := false
-		for _, resp := range anyResp {
-			if strings.Contains(strings.ToLower(response), resp) {
-				atLeastOne = true
-				break
-			}
-		}
-		if !atLeastOne {
-			t.Fatalf("%s: none of %v found in \"%s\" -- request was:%v", req.Model, anyResp, response, req.Messages)
-		}
-
+		verify()
 		slog.Info("test pass", "model", req.Model, "messages", req.Messages, "contains", anyResp, "response", response)
 	case <-ctx.Done():
-		t.Error("outer test context done while waiting for generate")
+		// On slow systems, we might timeout before some models finish rambling, so check what we have so far to see
+		// if it's considered a pass - the stallTimer will detect hangs, but we want to consider slow systems a pass
+		// if they are still generating valid responses
+		slog.Warn("outer test context done while waiting for chat")
+		verify()
 	}
 	return &api.Message{Role: role, Content: buf.String()}
 }
--- a/llama/llama.go
+++ b/llama/llama.go
@@ -515,33 +515,34 @@ func (c *MtmdContext) NewEmbed(llamaContext *Context, data []byte) ([][]float32,
 	}
 	nChunks := C.mtmd_input_chunks_size(ic)
 	numEmbed := llamaContext.Model().NEmbd()
-	lastChunkSize := 0
+	embed := make([][]float32, 0)
 	for i := range int(nChunks) {
 		chunk := C.mtmd_input_chunks_get(ic, C.size_t(i))
 		numTokens := int(C.mtmd_input_chunk_get_n_tokens(chunk))
-		lastChunkSize = numTokens
+		slog.Debug("chunk tokens", "index", i, "numTokens", numTokens)

 		// Encode the chunk
 		if C.int32_t(0) != C.mtmd_encode_chunk(c.c, chunk) {
 			return nil, errors.New("unable to encode mtmd image chunk")
 		}
-	}

-	// Get the embeddings
-	embed := make([][]float32, lastChunkSize)
-	embd := C.mtmd_get_output_embd(c.c)
-	if nil == embd {
-		return nil, errors.New("failed to get image embedding")
-	}
+		// Get the embeddings for this chunk
+		chunkEmbed := make([][]float32, numTokens)
+		chunkEmbd := C.mtmd_get_output_embd(c.c)
+		if nil == chunkEmbd {
+			continue
+		}

-	// Extend the embedding array for each token
-	s := unsafe.Slice((*float32)(embd), numEmbed*lastChunkSize)
-	rows := make([]float32, len(s))
-	copy(rows, s)
-	for i := range lastChunkSize {
-		embed[i] = rows[i*numEmbed : (i+1)*numEmbed]
+		// Extend the embedding array for each token
+		s := unsafe.Slice((*float32)(chunkEmbd), numTokens*numEmbed)
+		rows := make([]float32, len(s))
+		copy(rows, s)
+		for i := range numTokens {
+			chunkEmbed[i] = rows[i*numEmbed : (i+1)*numEmbed]
+		}
+		embed = append(embed, chunkEmbed...)
 	}
-
+	slog.Debug("image embeddings", "totalEmbeddings", len(embed))
 	return embed, nil
 }

--- a/llm/memory.go
+++ b/llm/memory.go
@@ -202,7 +202,7 @@ func estimateGPULayers(gpus []discover.GpuInfo, f *ggml.GGML, projectors []strin
 	var kvct string
 	if useFlashAttention {
 		requested := strings.ToLower(envconfig.KvCacheType())
-		if requested != "" && f.SupportsKVCacheType(requested) {
+		if f.SupportsKVCacheType(requested) {
 			kvct = requested
 		}
 	}
--- a/llm/server.go
+++ b/llm/server.go
@@ -148,7 +148,11 @@ func NewLlamaServer(gpus discover.GpuInfoList, modelPath string, f *ggml.GGML, a
 	var textProcessor model.TextProcessor
 	var err error
 	if envconfig.NewEngine() || f.KV().OllamaEngineRequired() {
-		textProcessor, err = model.NewTextProcessor(modelPath)
+		if len(projectors) == 0 {
+			textProcessor, err = model.NewTextProcessor(modelPath)
+		} else {
+			err = errors.New("split vision models aren't supported")
+		}
 		if err != nil {
 			// To prepare for opt-out mode, instead of treating this as an error, we fallback to the old runner
 			slog.Debug("model not yet supported by Ollama engine, switching to compatibility mode", "model", modelPath, "error", err)
@@ -161,11 +165,6 @@ func NewLlamaServer(gpus discover.GpuInfoList, modelPath string, f *ggml.GGML, a
 		}
 	}

-	newEstimates := textProcessor != nil && envconfig.NewMemoryEstimates()
-	if newEstimates {
-		slog.Info("enabling new memory estimates")
-	}
-
 	// Verify the requested context size is <= the model training size
 	trainCtx := f.KV().ContextLength()
 	if opts.NumCtx > int(trainCtx) && trainCtx > 0 {
@@ -173,6 +172,8 @@ func NewLlamaServer(gpus discover.GpuInfoList, modelPath string, f *ggml.GGML, a
 		opts.NumCtx = int(trainCtx)
 	}

+	opts.NumBatch = min(opts.NumBatch, opts.NumCtx)
+
 	loadRequest := LoadRequest{LoraPath: adapters, KvSize: opts.NumCtx * numParallel, BatchSize: opts.NumBatch, Parallel: numParallel, MultiUserCache: envconfig.MultiUserCache()}

 	defaultThreads := discover.GetSystemInfo().GetOptimalThreadCount()
@@ -218,7 +219,7 @@ func NewLlamaServer(gpus discover.GpuInfoList, modelPath string, f *ggml.GGML, a

 		// Flash Attention also supports kv cache quantization
 		// Enable if the requested and kv cache type is supported by the model
-		if kvct != "" && f.SupportsKVCacheType(kvct) {
+		if f.SupportsKVCacheType(kvct) {
 			loadRequest.KvCacheType = kvct
 		} else {
 			slog.Warn("kv cache type not supported by model", "type", kvct)
@@ -431,7 +432,7 @@ func NewLlamaServer(gpus discover.GpuInfoList, modelPath string, f *ggml.GGML, a
 			}
 		}()

-		if newEstimates {
+		if textProcessor != nil {
 			return &ollamaServer{llmServer: s}, nil
 		} else {
 			return &llamaServer{llmServer: s, ggml: f}, nil
--- a/openai/openai.go
+++ b/openai/openai.go
@@ -76,8 +76,9 @@ type JsonSchema struct {
 }

 type EmbedRequest struct {
-	Input any    `json:"input"`
-	Model string `json:"model"`
+	Input      any    `json:"input"`
+	Model      string `json:"model"`
+	Dimensions int    `json:"dimensions,omitempty"`
 }

 type StreamOptions struct {
@@ -1005,7 +1006,7 @@ func EmbeddingsMiddleware() gin.HandlerFunc {
 		}

 		var b bytes.Buffer
-		if err := json.NewEncoder(&b).Encode(api.EmbedRequest{Model: req.Model, Input: req.Input}); err != nil {
+		if err := json.NewEncoder(&b).Encode(api.EmbedRequest{Model: req.Model, Input: req.Input, Dimensions: req.Dimensions}); err != nil {
 			c.AbortWithStatusJSON(http.StatusInternalServerError, NewError(http.StatusInternalServerError, err.Error()))
 			return
 		}
--- a/parser/parser.go
+++ b/parser/parser.go
@@ -246,7 +246,7 @@ func filesForModel(path string) ([]string, error) {
 		for _, match := range matches {
 			if ct, err := detectContentType(match); err != nil {
 				return nil, err
-			} else if ct != contentType {
+			} else if len(contentType) > 0 && ct != contentType {
 				return nil, fmt.Errorf("invalid content type: expected %s for %s", ct, match)
 			}
 		}
@@ -255,7 +255,8 @@ func filesForModel(path string) ([]string, error) {
 	}

 	var files []string
-	if st, _ := glob(filepath.Join(path, "*.safetensors"), "application/octet-stream"); len(st) > 0 {
+	// some safetensors files do not properly match "application/octet-stream", so skip checking their contentType
+	if st, _ := glob(filepath.Join(path, "*.safetensors"), ""); len(st) > 0 {
 		// safetensors files might be unresolved git lfs references; skip if they are
 		// covers model-x-of-y.safetensors, model.fp32-x-of-y.safetensors, model.safetensors
 		files = append(files, st...)
--- a/runner/ollamarunner/cache.go
+++ b/runner/ollamarunner/cache.go
@@ -34,8 +34,8 @@ type InputCache struct {
 func NewInputCache(model model.Model, kvCacheType string, kvSize int32, numSlots int, batchSize int, multiUserCache bool) (*InputCache, error) {
 	numCtx := kvSize / int32(numSlots)

-	if numCtx < 1 {
-		return nil, fmt.Errorf("must have at least one kv cache entry per parallel sequence (kv: %v parallel: %v)", kvSize, numSlots)
+	if int(numCtx) < batchSize {
+		return nil, fmt.Errorf("kv size must be at least as large as batch size * parallel (kv: %v batch: %v parallel: %v)", kvSize, batchSize, numSlots)
 	}

 	slots := make([]InputCacheSlot, numSlots)
@@ -70,11 +70,9 @@ func kvCacheTypeFromStr(s string) ml.DType {
 }

 func (c *InputCache) Close() {
-	if c == nil {
-		return
+	if c != nil && c.cache != nil {
+		c.cache.Close()
 	}
-
-	c.cache.Close()
 }

 // Locking: Operations on InputCacheSlot (including finding one
--- a/runner/ollamarunner/runner.go
+++ b/runner/ollamarunner/runner.go
@@ -18,7 +18,6 @@ import (
 	"reflect"
 	"regexp"
 	"runtime"
-	"runtime/debug"
 	"strconv"
 	"strings"
 	"sync"
@@ -1086,9 +1085,13 @@ func (s *Server) allocModel(
 	// Convert memory allocation panics to errors
 	defer func() {
 		if r := recover(); r != nil {
-			debug.PrintStack()
 			if err, ok := r.(error); ok {
-				panicErr = err
+				var noMem ml.ErrNoMem
+				if errors.As(err, &noMem) {
+					panicErr = noMem
+				} else {
+					panic(r)
+				}
 			} else {
 				panic(r)
 			}
--- a/scripts/build_windows.ps1
+++ b/scripts/build_windows.ps1
@@ -78,7 +78,7 @@ function checkEnv() {
 }


-function buildOllama() {
+function buildCPU() {
    mkdir -Force -path "${script:DIST_DIR}\"
    if ($script:ARCH -ne "arm64") {
        Remove-Item -ea 0 -recurse -force -path "${script:SRC_DIR}\dist\windows-${script:ARCH}"
@@ -90,20 +90,72 @@ function buildOllama() {
        if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
        & cmake --install build --component CPU --strip
        if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
+    }
+}

+function buildCUDA11() {
+    # CUDA v11 claims to be compatible with MSVC 2022, but the latest updates are no longer compatible
+    # 19.40 is the last compiler version that works, but recent udpates are 19.43
+    # So this pins to MSVC 2019 for best compatibility
+    mkdir -Force -path "${script:DIST_DIR}\"
+    if ($script:ARCH -ne "arm64") {
        $hashEnv = @{}
        Get-ChildItem env: | foreach { $hashEnv[$_.Name] = $_.Value }
-        if ("$script:CUDA_DIRS".Contains("v12")) {
-            $hashEnv.Keys | foreach { if ($_.Contains("CUDA_PATH_V12")) { $v12="$_" }}
-            $env:CUDAToolkit_ROOT=$hashEnv[$v12]
-            write-host "Building CUDA v12 backend libraries"
-            & cmake --fresh --preset "CUDA 12" --install-prefix $script:DIST_DIR
+        if ("$script:CUDA_DIRS".Contains("v11")) {
+            $hashEnv.Keys | foreach { if ($_.Contains("CUDA_PATH_V11")) { $x=$hashEnv[$_]; if (test-path -literalpath "$x\bin\nvcc.exe" ) { $cuda=$x}  }}
+            write-host "Building CUDA v11 backend libraries $cuda"
+            $env:CUDAToolkit_ROOT=$cuda
+            & cmake --fresh --preset "CUDA 11" -T cuda="$cuda" -DCMAKE_CUDA_COMPILER="$cuda\bin\nvcc.exe" -G "Visual Studio 16 2019" --install-prefix $script:DIST_DIR -DOLLAMA_RUNNER_DIR="cuda_v11"
+            if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
+            & cmake --build --preset "CUDA 11"  --config Release --parallel $script:JOBS
+            if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
+            & cmake --install build --component "CUDA" --strip
+            if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
+        }
+    }
+}
+
+function buildCUDA12() {
+    mkdir -Force -path "${script:DIST_DIR}\"
+    if ($script:ARCH -ne "arm64") {
+        $hashEnv = @{}
+        Get-ChildItem env: | foreach { $hashEnv[$_.Name] = $_.Value }
+        if ("$script:CUDA_DIRS".Contains("v12.8")) {
+            $hashEnv.Keys | foreach { if ($_.Contains("CUDA_PATH_V12_8")) { $x=$hashEnv[$_]; if (test-path -literalpath "$x\bin\nvcc.exe" ) { $cuda=$x}  }}
+            write-host "Building CUDA v12 backend libraries $cuda"
+            $env:CUDAToolkit_ROOT=$cuda
+            & cmake --fresh --preset "CUDA 12" -T cuda="$cuda" --install-prefix $script:DIST_DIR -DOLLAMA_RUNNER_DIR="cuda_v12"
            if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
            & cmake --build --preset "CUDA 12"  --config Release --parallel $script:JOBS
            if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
            & cmake --install build --component "CUDA" --strip
            if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
        }
+    }
+}
+
+function buildCUDA13() {
+    mkdir -Force -path "${script:DIST_DIR}\"
+    if ($script:ARCH -ne "arm64") {
+        $hashEnv = @{}
+        Get-ChildItem env: | foreach { $hashEnv[$_.Name] = $_.Value }
+        if ("$script:CUDA_DIRS".Contains("v13")) {
+            $hashEnv.Keys | foreach { if ($_.Contains("CUDA_PATH_V13")) { $x=$hashEnv[$_]; if (test-path -literalpath "$x\bin\nvcc.exe" ) { $cuda=$x}  }}
+            $env:CUDAToolkit_ROOT=$cuda
+            write-host "Building CUDA v13 backend libraries $cuda"
+            & cmake --fresh --preset "CUDA 13" -T cuda="$cuda" --install-prefix $script:DIST_DIR -DOLLAMA_RUNNER_DIR="cuda_v13"
+            if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
+            & cmake --build --preset "CUDA 13"  --config Release --parallel $script:JOBS
+            if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
+            & cmake --install build --component "CUDA" --strip
+            if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
+        }
+    }
+}
+
+function buildROCm() {
+    mkdir -Force -path "${script:DIST_DIR}\"
+    if ($script:ARCH -ne "arm64") {
        if ($env:HIP_PATH) {
            write-host "Building ROCm backend libraries"
            if (-Not (get-command -ErrorAction silent ninja)) {
@@ -129,6 +181,10 @@ function buildOllama() {
            if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
        }
    }
+}
+
+function buildOllama() {
+    mkdir -Force -path "${script:DIST_DIR}\"
    write-host "Building ollama CLI"
    & go build -trimpath -ldflags "-s -w -X=github.com/ollama/ollama/version.Version=$script:VERSION -X=github.com/ollama/ollama/server.mode=release" .
    if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
@@ -236,6 +292,10 @@ function distZip() {
 checkEnv
 try {
    if ($($args.count) -eq 0) {
+        buildCPU
+        buildCUDA12
+        buildCUDA13
+        buildROCm
        buildOllama
        buildApp
        gatherDependencies
--- a/server/routes.go
+++ b/server/routes.go
@@ -576,7 +576,12 @@ func (s *Server) EmbedHandler(c *gin.Context) {
 			if err != nil {
 				return err
 			}
-			embeddings[i] = normalize(embedding)
+			// TODO: this first normalization should be done by the model
+			embedding = normalize(embedding)
+			if req.Dimensions > 0 && req.Dimensions < len(embedding) {
+				embedding = normalize(embedding[:req.Dimensions])
+			}
+			embeddings[i] = embedding
 			return nil
 		})
 	}
@@ -602,11 +607,7 @@ func normalize(vec []float32) []float32 {
 		sum += v * v
 	}

-	norm := float32(0.0)
-	if sum > 0 {
-		norm = float32(1.0 / math.Sqrt(float64(sum)))
-	}
-
+	norm := float32(1.0 / max(math.Sqrt(float64(sum)), 1e-12))
 	for i := range vec {
 		vec[i] *= norm
 	}