Merge branch 'ollama:main' into main

* ollama: add embeddings

README.md

@@ -624,6 +624,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [UnityCodeLama](https://github.com/HardCodeDev777/UnityCodeLama) (Unity Edtior tool to analyze scripts via Ollama)
 - [NativeMind](https://github.com/NativeMindBrowser/NativeMindExtension) (Private, on-device AI Assistant, no cloud dependencies)
 - [GMAI - Gradle Managed AI](https://gmai.premex.se/) (Gradle plugin for automated Ollama lifecycle management during build phases)
+- [NOMYO Router](https://github.com/nomyo-ai/nomyo-router) (A transparent Ollama proxy with model deployment aware routing which auto-manages multiple Ollama instances in a given network)
 
 ### Supported backends
 

harmony/harmonyparser.go

@@ -1,7 +1,6 @@
 package harmony
 
 import (
-	"context"
 	"fmt"
 	"log/slog"
 	"strings"
@@ -292,7 +291,7 @@ func (h *HarmonyMessageHandler) AddContent(content string, toolParser *HarmonyTo
 	for _, event := range events {
 		switch event := event.(type) {
 		case HarmonyEventHeaderComplete:
-			slog.Log(context.TODO(), logutil.LevelTrace, "harmony event header complete", "header", event.Header)
+			logutil.Trace("harmony event header complete", "header", event.Header)
 			switch event.Header.Channel {
 			case "analysis":
 				if event.Header.Recipient != "" {
@@ -315,7 +314,7 @@ func (h *HarmonyMessageHandler) AddContent(content string, toolParser *HarmonyTo
 				h.state = harmonyMessageState_Normal
 			}
 		case HarmonyEventContentEmitted:
-			slog.Log(context.TODO(), logutil.LevelTrace, "harmony event content", "content", event.Content, "state", h.state)
+			logutil.Trace("harmony event content", "content", event.Content, "state", h.state)
 			if h.state == harmonyMessageState_Normal {
 				contentSb.WriteString(event.Content)
 			} else if h.state == harmonyMessageState_Thinking {

llama/patches/0025-harden-uncaught-exception-registration.patch

@@ -0,0 +1,28 @@
+From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
+From: Daniel Hiltgen <daniel@ollama.com>
+Date: Fri, 29 Aug 2025 16:53:08 -0700
+Subject: [PATCH] harden uncaught exception registration
+
+---
+ ggml/src/ggml.cpp | 8 ++++++--
+ 1 file changed, 6 insertions(+), 2 deletions(-)
+
+diff --git a/ggml/src/ggml.cpp b/ggml/src/ggml.cpp
+index 0d388d45..f5bcb446 100644
+--- a/ggml/src/ggml.cpp
++++ b/ggml/src/ggml.cpp
+@@ -19,8 +19,12 @@ static bool ggml_uncaught_exception_init = []{
+         return false;
+     }
+     const auto prev{std::get_terminate()};
+-    GGML_ASSERT(prev != ggml_uncaught_exception);
+-    previous_terminate_handler = prev;
++    // GGML_ASSERT(prev != ggml_uncaught_exception);
++    if (prev != ggml_uncaught_exception) {
++        previous_terminate_handler = prev;
++    } else {
++        GGML_LOG_WARN("%s double registration of ggml_uncaught_exception\n", __func__);
++    }
+     std::set_terminate(ggml_uncaught_exception);
+     return true;
+ }();

llm/server.go

@@ -678,8 +678,12 @@ func (s *ollamaServer) Load(ctx context.Context, gpus discover.GpuInfoList, requ
 
 	if !(len(gpus) == 1 && gpus[0].Library == "cpu") {
 		for _, gpu := range gpus {
+			available := gpu.FreeMemory - envconfig.GpuOverhead() - gpu.MinimumMemory
+			if gpu.FreeMemory < envconfig.GpuOverhead()+gpu.MinimumMemory {
+				available = 0
+			}
 			slog.Info("gpu memory", "id", gpu.ID,
-				"available", format.HumanBytes2(gpu.FreeMemory-envconfig.GpuOverhead()-gpu.MinimumMemory),
+				"available", format.HumanBytes2(available),
 				"free", format.HumanBytes2(gpu.FreeMemory),
 				"minimum", format.HumanBytes2(gpu.MinimumMemory),
 				"overhead", format.HumanBytes2(envconfig.GpuOverhead()))
@@ -861,7 +865,7 @@ func (s *ollamaServer) createLayout(systemInfo discover.SystemInfo, systemGPUs d
 		}
 		layers[i] += memory.CPU.Weights[i].Size
 		layers[i] += memory.CPU.Cache[i].Size
-		slog.Log(context.TODO(), logutil.LevelTrace, "layer to assign", "layer", i, "size", format.HumanBytes2(layers[i]))
+		logutil.Trace("layer to assign", "layer", i, "size", format.HumanBytes2(layers[i]))
 	}
 
 	gpuLayers := ml.GPULayersList{}

logutil/logutil.go

@@ -1,6 +1,7 @@
 package logutil
 
 import (
+	"context"
 	"io"
 	"log/slog"
 	"path/filepath"
@@ -27,3 +28,11 @@ func NewLogger(w io.Writer, level slog.Level) *slog.Logger {
 		},
 	}))
 }
+
+func Trace(msg string, args ...any) {
+	slog.Log(context.TODO(), LevelTrace, msg, args...)
+}
+
+func TraceContext(ctx context.Context, msg string, args ...any) {
+	slog.Log(ctx, LevelTrace, msg, args...)
+}

ml/backend.go

@@ -266,7 +266,7 @@ func (m DeviceMemory) LogValue() slog.Value {
 // allocation is guaranteed to be provided so that if it failed, the caller can
 // accommodate that to make forward progress.
 type BackendMemory struct {
-	// InputsWeights are always located on the CPU and cannot be moved
+	// InputWeights are always located on the CPU and cannot be moved
 	InputWeights Memory
 
 	// CPU model components are located in system memory. This does not

ml/backend/ggml/ggml.go

@@ -271,7 +271,7 @@ func New(modelPath string, params ml.BackendParams) (ml.Backend, error) {
 			tt := C.ggml_new_tensor(ctxs[bt], kind, C.int(len(t.source.Shape)), (*C.int64_t)(unsafe.Pointer(&t.source.Shape[0])))
 			C.ggml_set_name(tt, cname)
 
-			slog.Log(context.TODO(), logutil.LevelTrace, "created tensor", "name", name, "shape", t.source.Shape, "dtype", t.source.Kind, "buffer_type", C.GoString(C.ggml_backend_buft_name(bt)))
+			logutil.Trace("created tensor", "name", name, "shape", t.source.Shape, "dtype", t.source.Kind, "buffer_type", C.GoString(C.ggml_backend_buft_name(bt)))
 
 			size := pad(C.ggml_backend_buft_get_alloc_size(bt, tt), C.ggml_backend_buft_get_alignment(bt))
 			if layer == -1 {
@@ -378,7 +378,7 @@ func New(modelPath string, params ml.BackendParams) (ml.Backend, error) {
 	}
 
 	for bs := range maps.Values(bbs) {
-		slog.Log(context.TODO(), logutil.LevelTrace, "model weights", "buffer", C.GoString(C.ggml_backend_buffer_name(bs)),
+		logutil.Trace("model weights", "buffer", C.GoString(C.ggml_backend_buffer_name(bs)),
 			"size", format.HumanBytes2(uint64(C.ggml_backend_buffer_get_size(bs))))
 	}
 
@@ -811,7 +811,7 @@ func (c *Context) Reserve() {
 			}
 		}
 
-		slog.Log(context.TODO(), logutil.LevelTrace, "compute graph", "backend", C.GoString(C.ggml_backend_name(c.b.schedBackends[i])),
+		logutil.Trace("compute graph", "backend", C.GoString(C.ggml_backend_name(c.b.schedBackends[i])),
 			"buffer_type", C.GoString(C.ggml_backend_buft_name(c.b.schedBufts[i])), "size", format.HumanBytes2(uint64(bufferStatus.size)))
 	}
 

ml/backend/ggml/ggml/src/ggml.cpp

@@ -19,8 +19,12 @@ static bool ggml_uncaught_exception_init = []{
         return false;
     }
     const auto prev{std::get_terminate()};
-    GGML_ASSERT(prev != ggml_uncaught_exception);
+    // GGML_ASSERT(prev != ggml_uncaught_exception);
-    previous_terminate_handler = prev;
+    if (prev != ggml_uncaught_exception) {
+        previous_terminate_handler = prev;
+    } else {
+        GGML_LOG_WARN("%s double registration of ggml_uncaught_exception\n", __func__);
+    }
     std::set_terminate(ggml_uncaught_exception);
     return true;
 }();

model/bytepairencoding.go

@@ -2,7 +2,6 @@ package model
 
 import (
 	"cmp"
-	"context"
 	"fmt"
 	"iter"
 	"log/slog"
@@ -202,12 +201,11 @@ func (bpe BytePairEncoding) Encode(s string, addSpecial bool) ([]int32, error) {
 		}
 	}
 
-	slog.Log(context.TODO(), logutil.LevelTrace, "encoded", "string", s, "ids", ids)
-
 	if addSpecial && len(ids) > 0 {
 		ids = bpe.vocab.addSpecials(ids)
 	}
 
+	logutil.Trace("encoded", "string", s, "ids", ids)
 	return ids, nil
 }
 
@@ -243,6 +241,6 @@ func (bpe BytePairEncoding) Decode(ids []int32) (string, error) {
 		}
 	}
 
-	slog.Log(context.TODO(), logutil.LevelTrace, "decoded", "string", sb.String(), "from", lazyIdsString{ids: ids})
+	logutil.Trace("decoded", "string", sb.String(), "from", lazyIdsString{ids: ids})
 	return sb.String(), nil
 }

model/model.go

@@ -1,12 +1,11 @@
 package model
 
 import (
-	"context"
 	"errors"
 	"fmt"
 	_ "image/jpeg"
 	_ "image/png"
-	"log/slog"
+	"math"
 	"os"
 	"reflect"
 	"strconv"
@@ -105,6 +104,10 @@ func New(modelPath string, params ml.BackendParams) (Model, error) {
 	}
 
 	arch := b.Config().Architecture()
+	if b.Config().Uint("pooling_type", math.MaxUint32) != math.MaxUint32 {
+		arch = arch + "_embed"
+	}
+
 	f, ok := models[arch]
 	if !ok {
 		return nil, fmt.Errorf("unsupported model architecture %q", arch)
@@ -198,7 +201,7 @@ func populateFields(base Base, v reflect.Value, tags ...Tag) reflect.Value {
 				names := fn(tagsCopy)
 				for _, name := range names {
 					if tensor := base.Backend().Get(strings.Join(name, ".")); tensor != nil {
-						slog.Log(context.TODO(), logutil.LevelTrace, "found tensor", "", tensor)
+						logutil.Trace("found tensor", "", tensor)
 						vv.Set(reflect.ValueOf(tensor))
 						break
 					}

model/models/gemma3/embed.go

@@ -0,0 +1,73 @@
+package gemma3
+
+import (
+	"errors"
+
+	"github.com/ollama/ollama/fs"
+	"github.com/ollama/ollama/kvcache"
+	"github.com/ollama/ollama/ml"
+	"github.com/ollama/ollama/ml/nn"
+	"github.com/ollama/ollama/model"
+	"github.com/ollama/ollama/model/input"
+)
+
+type embedModel struct {
+	model.Base
+	model.SentencePieceModel
+
+	*TextModel
+	PoolingType uint32
+
+	Dense [2]*nn.Linear `gguf:"dense"`
+}
+
+func (m *embedModel) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
+	batch.Outputs = batch.Positions // return all positions
+	hiddenStates := m.TextModel.Forward(ctx, batch, m.Cache)
+
+	switch m.PoolingType {
+	case 0: // None
+	case 1: // Mean
+		hiddenStates = hiddenStates.Permute(ctx, 1, 0, 2, 3).Contiguous(ctx).Mean(ctx)
+		hiddenStates = hiddenStates.Permute(ctx, 1, 0, 2, 3).Contiguous(ctx)
+	default:
+		return nil, errors.New("unsupported pooling type")
+	}
+
+	for _, dense := range m.Dense {
+		hiddenStates = dense.Forward(ctx, hiddenStates)
+	}
+
+	return hiddenStates, nil
+}
+
+func newEmbedModel(c fs.Config) (model.Model, error) {
+	m := &embedModel{
+		SentencePieceModel: model.NewSentencePieceModel(
+			&model.Vocabulary{
+				Values: c.Strings("tokenizer.ggml.tokens"),
+				Scores: c.Floats("tokenizer.ggml.scores"),
+				Types:  c.Ints("tokenizer.ggml.token_type"),
+				AddBOS: c.Bool("tokenizer.ggml.add_bos_token", true),
+				BOS:    []int32{int32(c.Uint("tokenizer.ggml.bos_token_id"))},
+				AddEOS: c.Bool("tokenizer.ggml.add_eos_token", false),
+				EOS: append(
+					[]int32{
+						int32(c.Uint("tokenizer.ggml.eos_token_id")),
+						int32(c.Uint("tokenizer.ggml.eot_token_id", 106)),
+					},
+					c.Ints("tokenizer.ggml.eos_token_ids")...,
+				),
+			},
+		),
+		TextModel:   newTextModel(c),
+		PoolingType: c.Uint("pooling_type", 0),
+	}
+
+	m.Cache = kvcache.NewWrapperCache(
+		kvcache.NewSWACache(int32(c.Uint("attention.sliding_window")), m.Shift),
+		kvcache.NewCausalCache(m.Shift),
+	)
+
+	return m, nil
+}

model/models/gemma3/model.go

@@ -141,12 +141,11 @@ func (m *Model) PostTokenize(inputs []*input.Input) ([]*input.Input, error) {
 }
 
 func (m *Model) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
-	positions := ctx.Input().FromIntSlice(batch.Positions, len(batch.Positions))
+	hiddenStates := m.TextModel.Forward(ctx, batch, m.Cache)
-	outputs := ctx.Input().FromIntSlice(batch.Outputs, len(batch.Outputs))
+	return m.Output.Forward(ctx, hiddenStates), nil
-
-	return m.TextModel.Forward(ctx, batch.Inputs, positions, outputs, batch, m.Cache), nil
 }
 
 func init() {
 	model.Register("gemma3", New)
+	model.Register("gemma3_embed", newEmbedModel)
 }

model/models/gemma3/model_text.go

@@ -159,8 +159,11 @@ func (l *TextLayer) Forward(ctx ml.Context, layer int, hiddenState, positionIDs,
 	return hiddenState.Add(ctx, residual)
 }
 
-func (m *TextModel) Forward(ctx ml.Context, inputs, positions, outputs ml.Tensor, batch input.Batch, cache kvcache.Cache) ml.Tensor {
+func (m *TextModel) Forward(ctx ml.Context, batch input.Batch, cache kvcache.Cache) ml.Tensor {
-	hiddenState := m.TokenEmbedding.Forward(ctx, inputs)
+	positions := ctx.Input().FromIntSlice(batch.Positions, len(batch.Positions))
+	outputs := ctx.Input().FromIntSlice(batch.Outputs, len(batch.Outputs))
+
+	hiddenState := m.TokenEmbedding.Forward(ctx, batch.Inputs)
 	hiddenState = hiddenState.Scale(ctx, math.Sqrt(float64(m.TextConfig.hiddenSize)))
 
 	// set image embeddings
@@ -198,5 +201,5 @@ func (m *TextModel) Forward(ctx ml.Context, inputs, positions, outputs ml.Tensor
 	}
 
 	hiddenState = m.OutputNorm.Forward(ctx, hiddenState, m.eps)
-	return m.Output.Forward(ctx, hiddenState)
+	return hiddenState
 }

model/sentencepiece.go

@@ -2,7 +2,6 @@ package model
 
 import (
 	"container/heap"
-	"context"
 	"fmt"
 	"log/slog"
 	"strconv"
@@ -25,7 +24,7 @@ func (spm SentencePieceModel) Vocabulary() *Vocabulary {
 }
 
 func NewSentencePieceModel(vocab *Vocabulary) SentencePieceModel {
-	slog.Log(context.TODO(), logutil.LevelTrace, "Tokens", "num tokens", len(vocab.Values), "vals", vocab.Values[:5], "scores", vocab.Scores[:5], "types", vocab.Types[:5])
+	logutil.Trace("Tokens", "num tokens", len(vocab.Values), "vals", vocab.Values[:5], "scores", vocab.Scores[:5], "types", vocab.Types[:5])
 
 	counter := map[int]int{}
 	var maxTokenLen int
@@ -39,7 +38,7 @@ func NewSentencePieceModel(vocab *Vocabulary) SentencePieceModel {
 		}
 	}
 
-	slog.Log(context.TODO(), logutil.LevelTrace, "Token counts", "normal", counter[TOKEN_TYPE_NORMAL], "unknown", counter[TOKEN_TYPE_UNKNOWN], "control", counter[TOKEN_TYPE_CONTROL],
+	logutil.Trace("Token counts", "normal", counter[TOKEN_TYPE_NORMAL], "unknown", counter[TOKEN_TYPE_UNKNOWN], "control", counter[TOKEN_TYPE_CONTROL],
 		"user defined", counter[TOKEN_TYPE_USER_DEFINED], "unused", counter[TOKEN_TYPE_UNUSED], "byte", counter[TOKEN_TYPE_BYTE],
 		"max token len", maxTokenLen)
 
@@ -182,12 +181,11 @@ func (spm SentencePieceModel) Encode(s string, addSpecial bool) ([]int32, error)
 		}
 	}
 
-	slog.Log(context.TODO(), logutil.LevelTrace, "encoded", "string", s, "ids", ids)
-
 	if addSpecial && len(ids) > 0 {
 		ids = spm.vocab.addSpecials(ids)
 	}
 
+	logutil.Trace("encoded", "string", s, "ids", ids)
 	return ids, nil
 }
 
@@ -246,6 +244,6 @@ func (spm SentencePieceModel) Decode(ids []int32) (string, error) {
 		}
 	}
 
-	slog.Log(context.TODO(), logutil.LevelTrace, "decoded", "ids", ids, "string", sb.String())
+	logutil.Trace("decoded", "ids", ids, "string", sb.String())
 	return sb.String(), nil
 }

model/vocabulary.go

@@ -49,7 +49,7 @@ func (v *Vocabulary) addSpecials(ids []int32) []int32 {
 			slog.Warn("adding bos token to prompt which already has it", "id", v.BOS)
 		}
 
-		slog.Debug("adding bos token to prompt", "id", v.BOS)
+		slog.Debug("adding bos token to prompt", "id", v.BOS[0])
 		ids = append([]int32{v.BOS[0]}, ids...)
 	}
 
@@ -58,7 +58,7 @@ func (v *Vocabulary) addSpecials(ids []int32) []int32 {
 			slog.Warn("adding eos token to prompt which already has it", "id", v.EOS)
 		}
 
-		slog.Debug("adding eos token to prompt", "id", v.EOS)
+		slog.Debug("adding eos token to prompt", "id", v.EOS[0])
 		ids = append(ids, v.EOS[0])
 	}
 

runner/ollamarunner/cache.go

@@ -95,7 +95,7 @@ type InputCacheSlot struct {
 	lastUsed time.Time
 }
 
-func (c *InputCache) LoadCacheSlot(prompt []*input.Input) (*InputCacheSlot, []*input.Input, error) {
+func (c *InputCache) LoadCacheSlot(prompt []*input.Input, cachePrompt bool) (*InputCacheSlot, []*input.Input, error) {
 	var slot *InputCacheSlot
 	var numPast int32
 	var err error
@@ -113,6 +113,10 @@ func (c *InputCache) LoadCacheSlot(prompt []*input.Input) (*InputCacheSlot, []*i
 		return nil, nil, err
 	}
 
+	if !cachePrompt {
+		numPast = 0
+	}
+
 	slot.InUse = true
 	slot.lastUsed = time.Now()
 

runner/ollamarunner/cache_test.go

@@ -393,7 +393,7 @@ func TestLoadCacheSlot(t *testing.T) {
 
 	for _, tt := range tests {
 		t.Run(tt.name, func(t *testing.T) {
-			slot, remainingPrompt, err := tt.cache.LoadCacheSlot(tt.prompt)
+			slot, remainingPrompt, err := tt.cache.LoadCacheSlot(tt.prompt, true)
 
 			// Check error state
 			if (err != nil) != tt.wantErr {

runner/ollamarunner/runner.go

@@ -11,6 +11,7 @@ import (
 	"image"
 	"log"
 	"log/slog"
+	"math"
 	"net"
 	"net/http"
 	"os"
@@ -405,6 +406,8 @@ func (s *Server) removeSequence(seqIndex int, reason llm.DoneReason) {
 func (s *Server) run(ctx context.Context) {
 	s.ready.Wait()
 
+	supportsAsync := s.model.Backend().Config().Uint("pooling_type", math.MaxUint32) == math.MaxUint32
+
 	var activeBatch batchState
 	for {
 		select {
@@ -418,7 +421,12 @@ func (s *Server) run(ctx context.Context) {
 			if err != nil {
 				panic(err)
 			}
-			go s.computeBatch(activeBatch)
+
+			if supportsAsync {
+				go s.computeBatch(activeBatch)
+			} else {
+				s.computeBatch(activeBatch)
+			}
 		}
 	}
 }
@@ -429,12 +437,12 @@ func (s *Server) forwardBatch(pendingBatch batchState) (nextBatch batchState, er
 	// before setting up the next batch so the seqs inputs are ready to receive their
 	// token values and we get the correct input pointers for the batchInputs
 	if pendingBatch.ctx != nil {
-		slog.Log(context.TODO(), logutil.LevelTrace, "forwardBatch waiting for compute to start", "pendingBatch.id", pendingBatch.id)
+		logutil.Trace("forwardBatch waiting for compute to start", "pendingBatch.id", pendingBatch.id)
 		<-pendingBatch.computeStartedCh
-		slog.Log(context.TODO(), logutil.LevelTrace, "forwardBatch compute started, setting up next batch", "pendingBatch.id", pendingBatch.id, "id", s.batchID)
+		logutil.Trace("forwardBatch compute started, setting up next batch", "pendingBatch.id", pendingBatch.id, "id", s.batchID)
 		nextBatch.inputsReadyCh = pendingBatch.outputsReadyCh // Chain the ouputs from the pending batch to the next inputs batch
 	} else {
-		slog.Log(context.TODO(), logutil.LevelTrace, "forwardBatch no pending batch detected", "batchID", s.batchID)
+		logutil.Trace("forwardBatch no pending batch detected", "batchID", s.batchID)
 		// No pendingBatch, so the inputs will be ready in the seqs immediately
 		nextBatch.inputsReadyCh = make(chan struct{}, 1)
 		nextBatch.inputsReadyCh <- struct{}{}
@@ -546,7 +554,7 @@ func (s *Server) forwardBatch(pendingBatch batchState) (nextBatch batchState, er
 			if i+1 == len(seq.inputs) {
 				batch.Outputs = append(batch.Outputs, int32(len(batchInputs)-1))
 			}
-			slog.Log(context.TODO(), logutil.LevelTrace, "forwardBatch iBatch", "batchID", s.batchID, "seqIdx", seqIdx, "seq.iBatch", seq.iBatch, "i+1", i+1, "len(seq.inputs)", len(seq.inputs))
+			logutil.Trace("forwardBatch iBatch", "batchID", s.batchID, "seqIdx", seqIdx, "seq.iBatch", seq.iBatch, "i+1", i+1, "len(seq.inputs)", len(seq.inputs))
 			seq.pendingInputs = append(seq.pendingInputs, inp)
 		}
 
@@ -560,7 +568,7 @@ func (s *Server) forwardBatch(pendingBatch batchState) (nextBatch batchState, er
 	}
 
 	if len(batchInputs) == 0 {
-		slog.Log(context.TODO(), logutil.LevelTrace, "forwardBatch no batchInputs, going idle", "batchID", s.batchID)
+		logutil.Trace("forwardBatch no batchInputs, going idle", "batchID", s.batchID)
 		nextBatch.ctx.Close()
 		nextBatch.ctx = nil
 		return
@@ -589,14 +597,14 @@ func (s *Server) computeBatch(activeBatch batchState) {
 	defer activeBatch.ctx.Close()
 
 	// Wait until inputs are ready
-	slog.Log(context.TODO(), logutil.LevelTrace, "computeBatch: waiting for inputs to be ready", "batchID", activeBatch.id)
+	logutil.Trace("computeBatch: waiting for inputs to be ready", "batchID", activeBatch.id)
 	<-activeBatch.inputsReadyCh
-	slog.Log(context.TODO(), logutil.LevelTrace, "computeBatch: inputs are ready", "batchID", activeBatch.id)
+	logutil.Trace("computeBatch: inputs are ready", "batchID", activeBatch.id)
 
 	// Once we complete, signal the next batch of inputs are ready
 	// This will unblock the next computeBatch, or forwardBatch if new seqs come in
 	defer func() {
-		slog.Log(context.TODO(), logutil.LevelTrace, "computeBatch: outputs are ready", "batchID", activeBatch.id)
+		logutil.Trace("computeBatch: outputs are ready", "batchID", activeBatch.id)
 		activeBatch.outputsReadyCh <- struct{}{}
 	}()
 
@@ -626,7 +634,7 @@ func (s *Server) computeBatch(activeBatch batchState) {
 		// Detect if the sequence we're processing has already been completed and replaced
 		// with a new sequence
 		if seq != activeBatch.seqs[i] {
-			slog.Log(context.TODO(), logutil.LevelTrace, "computeBatch: sequence replaced, discarding its results", "batchID", activeBatch.id, "seqIdx", i)
+			logutil.Trace("computeBatch: sequence replaced, discarding its results", "batchID", activeBatch.id, "seqIdx", i)
 			continue
 		}
 
@@ -666,18 +674,19 @@ func (s *Server) computeBatch(activeBatch batchState) {
 	activeBatch.batch.Inputs.SetValueFromIntSlice(batchInputs)
 	activeBatch.ctx.ComputeWithNotify(
 		func() {
-			slog.Log(context.TODO(), logutil.LevelTrace, "computeBatch: signaling computeStartedCh", "batchID", activeBatch.id)
+			logutil.Trace("computeBatch: signaling computeStartedCh", "batchID", activeBatch.id)
 			activeBatch.computeStartedCh <- struct{}{}
 		},
 		activeBatch.modelOutput)
-	logits := activeBatch.modelOutput.Floats()
 
-	slog.Log(context.TODO(), logutil.LevelTrace, "computeBatch: logits ready", "batchID", activeBatch.id)
+	outputs := activeBatch.modelOutput.Floats()
+
+	logutil.Trace("computeBatch: logits ready", "batchID", activeBatch.id)
 
 	s.mu.Lock()
 	defer s.mu.Unlock()
 
-	slog.Log(context.TODO(), logutil.LevelTrace, "computeBatch: decoding", "batchID", activeBatch.id)
+	logutil.Trace("computeBatch: decoding", "batchID", activeBatch.id)
 	for i, seq := range s.seqs {
 		if seq == nil || nextBatchTokens[i] == nil {
 			continue
@@ -689,16 +698,15 @@ func (s *Server) computeBatch(activeBatch batchState) {
 
 		// if done processing the prompt, generate an embedding and return
 		if seq.embeddingOnly {
-			// TODO(jessegross): Embedding support
+			seq.embedding <- outputs
-			slog.Warn("generation of embedding outputs not yet supported", "id", activeBatch.id, "seqIdx", i)
 			s.removeSequence(i, llm.DoneReasonStop)
 			continue
 		}
 
 		// sample a token
-		vocabSize := len(logits) / len(activeBatch.batch.Outputs)
+		vocabSize := len(outputs) / len(activeBatch.batch.Outputs)
-		slog.Log(context.TODO(), logutil.LevelTrace, "computeBatch: vocab details", "batchID", activeBatch.id, "seqIdx", i, "len(logits)", len(logits), "len(activeBatch.batch.Outputs)", len(activeBatch.batch.Outputs), "vocabSize", vocabSize, "iBatches", iBatches)
+		logutil.Trace("computeBatch: vocab details", "batchID", activeBatch.id, "seqIdx", i, "len(logits)", len(outputs), "len(activeBatch.batch.Outputs)", len(activeBatch.batch.Outputs), "vocabSize", vocabSize, "iBatches", iBatches)
-		token, err := seq.sampler.Sample(logits[iBatches[i]*vocabSize : (iBatches[i]+1)*vocabSize])
+		token, err := seq.sampler.Sample(outputs[iBatches[i]*vocabSize : (iBatches[i]+1)*vocabSize])
 		if err != nil {
 			s.hardErrCh <- fmt.Errorf("failed to sample token: %w", err)
 			return
@@ -711,7 +719,7 @@ func (s *Server) computeBatch(activeBatch batchState) {
 			// TODO (jmorganca): we should send this back
 			// as it's important for the /api/generate context
 			// seq.responses <- piece
-			slog.Log(context.TODO(), logutil.LevelTrace, "computeBatch: EOS", "batchID", activeBatch.id, "seqIdx", i)
+			logutil.Trace("computeBatch: EOS", "batchID", activeBatch.id, "seqIdx", i)
 			s.removeSequence(i, llm.DoneReasonStop)
 			continue
 		}
@@ -834,7 +842,7 @@ func (s *Server) completion(w http.ResponseWriter, r *http.Request) {
 	found := false
 	for i, sq := range s.seqs {
 		if sq == nil {
-			seq.cache, seq.inputs, err = s.cache.LoadCacheSlot(seq.inputs)
+			seq.cache, seq.inputs, err = s.cache.LoadCacheSlot(seq.inputs, true)
 			if err != nil {
 				s.mu.Unlock()
 				s.seqsSem.Release(1)
@@ -890,6 +898,67 @@ func (s *Server) completion(w http.ResponseWriter, r *http.Request) {
 	}
 }
 
+func (s *Server) embeddings(w http.ResponseWriter, r *http.Request) {
+	if s.model.Backend().Config().Uint("pooling_type", math.MaxUint32) == math.MaxUint32 {
+		http.Error(w, "this model does not support embeddings", http.StatusNotImplemented)
+		return
+	}
+
+	var req llm.EmbeddingRequest
+	if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
+		http.Error(w, fmt.Sprintf("bad request: %s", err), http.StatusBadRequest)
+		return
+	}
+
+	w.Header().Set("Content-Type", "application/json")
+	seq, err := s.NewSequence(req.Content, nil, NewSequenceParams{embedding: true})
+	if err != nil {
+		http.Error(w, fmt.Sprintf("failed to create new sequence: %v", err), http.StatusInternalServerError)
+		return
+	}
+
+	if err := s.seqsSem.Acquire(r.Context(), 1); err != nil {
+		if errors.Is(err, context.Canceled) {
+			slog.Info("aborting embedding request due to client closing the connection")
+		} else {
+			http.Error(w, fmt.Sprintf("failed to acquire semaphore: %v", err), http.StatusInternalServerError)
+		}
+		return
+	}
+
+	s.mu.Lock()
+	found := false
+	for i, sq := range s.seqs {
+		if sq == nil {
+			seq.cache, seq.inputs, err = s.cache.LoadCacheSlot(seq.inputs, false)
+			if err != nil {
+				s.mu.Unlock()
+				s.seqsSem.Release(1)
+				http.Error(w, fmt.Sprintf("failed to load cache: %v", err), http.StatusInternalServerError)
+				return
+			}
+
+			s.seqs[i] = seq
+			s.cond.Signal()
+			found = true
+			break
+		}
+	}
+	s.mu.Unlock()
+
+	if !found {
+		s.seqsSem.Release(1)
+		http.Error(w, "could not find an available sequence", http.StatusInternalServerError)
+		return
+	}
+
+	if err := json.NewEncoder(w).Encode(&llm.EmbeddingResponse{
+		Embedding: <-seq.embedding,
+	}); err != nil {
+		http.Error(w, fmt.Sprintf("failed to encode response: %v", err), http.StatusInternalServerError)
+	}
+}
+
 func (s *Server) health(w http.ResponseWriter, r *http.Request) {
 	w.Header().Set("Content-Type", "application/json")
 	if err := json.NewEncoder(w).Encode(&llm.ServerStatusResponse{
@@ -1206,10 +1275,7 @@ func Execute(args []string) error {
 	mux := http.NewServeMux()
 	// TODO: support embeddings
 	mux.HandleFunc("POST /load", server.load)
-	mux.HandleFunc("POST /embedding", func(w http.ResponseWriter, r *http.Request) {
+	mux.HandleFunc("POST /embedding", server.embeddings)
-		http.Error(w, "this model does not support embeddings", http.StatusNotImplemented)
-	})
-
 	mux.HandleFunc("POST /completion", server.completion)
 	mux.HandleFunc("GET /health", server.health)