Merge remote-tracking branch 'upstream/main' into vulkanV3

discover/cuda_common.go

@@ -16,7 +16,7 @@ import (
 // Included to drive logic for reducing Ollama-allocated overhead on L4T/Jetson devices.
 var CudaTegra string = os.Getenv("JETSON_JETPACK")
 
-func cudaVariant(gpuInfo CudaGPUInfo) string {
+func cudaVariant(gpuInfos []CudaGPUInfo) string {
 	if runtime.GOARCH == "arm64" && runtime.GOOS == "linux" {
 		if CudaTegra != "" {
 			ver := strings.Split(CudaTegra, ".")
@@ -45,20 +45,19 @@ func cudaVariant(gpuInfo CudaGPUInfo) string {
 		}
 	}
 
-	// Check GPU compute capability FIRST
+	// Check GPU compute capability FIRST, lowest common denominator if multi-gpu
-	isOldGPU := gpuInfo.computeMajor < 7 || (gpuInfo.computeMajor == 7 && gpuInfo.computeMinor < 5)
+	for _, gpuInfo := range gpuInfos {
-	if isOldGPU {
+		if gpuInfo.computeMajor < 7 || (gpuInfo.computeMajor == 7 && gpuInfo.computeMinor < 5) {
 			// GPU is Pascal or older (CC <= 7.4) - use CUDA v12 (supports CC 6.1)
 			return "v12"
 		}
+	}
 
 	// GPU is Turing or newer (CC >= 7.5) - can use newer CUDA
-	if gpuInfo.DriverMajor < 13 {
+	if len(gpuInfos) > 0 && gpuInfos[0].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 !isOldGPU {
+		slog.Warn("old CUDA driver detected - please upgrade to a newer driver for best performance", "version", fmt.Sprintf("%d.%d", gpuInfos[0].DriverMajor, gpuInfos[0].DriverMinor))
-			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 "v13"

discover/gpu.go

@@ -319,18 +319,8 @@ func GetGPUInfo() GpuInfoList {
 				gpuInfo.MinimumMemory = cudaMinimumMemory
 				gpuInfo.DriverMajor = driverMajor
 				gpuInfo.DriverMinor = driverMinor
-				variant := cudaVariant(gpuInfo)
 
-				// Start with our bundled libraries
-				if variant != "" {
-					variantPath := filepath.Join(LibOllamaPath, "cuda_"+variant)
-					if _, err := os.Stat(variantPath); err == nil {
-						// Put the variant directory first in the search path to avoid runtime linking to the wrong library
-						gpuInfo.DependencyPath = append([]string{variantPath}, gpuInfo.DependencyPath...)
-					}
-				}
 				gpuInfo.Name = C.GoString(&memInfo.gpu_name[0])
-				gpuInfo.Variant = variant
 
 				if int(memInfo.major) < cudaComputeMajorMin || (int(memInfo.major) == cudaComputeMajorMin && int(memInfo.minor) < cudaComputeMinorMin) {
 					unsupportedGPUs = append(unsupportedGPUs,
@@ -368,6 +358,24 @@ func GetGPUInfo() GpuInfoList {
 				// TODO potentially sort on our own algorithm instead of what the underlying GPU library does...
 				cudaGPUs = append(cudaGPUs, gpuInfo)
 			}
+			// Second pass on NVIDIA GPUs to set lowest common denominator variant and DependencyPaths
+			variant := cudaVariant(cudaGPUs)
+			var variantPath string
+			// Start with our bundled libraries
+			if variant != "" {
+				variantPath = filepath.Join(LibOllamaPath, "cuda_"+variant)
+				if _, err := os.Stat(variantPath); err != nil {
+					variantPath = ""
+				}
+			}
+
+			for i := range cudaGPUs {
+				cudaGPUs[i].Variant = variant
+				if variantPath != "" {
+					// Put the variant directory first in the search path to avoid runtime linking to the wrong library
+					cudaGPUs[i].DependencyPath = append([]string{variantPath}, cudaGPUs[i].DependencyPath...)
+				}
+			}
 		}
 
 		// Intel

fs/ggml/ggml.go

@@ -243,6 +243,7 @@ func (kv KV) OllamaEngineRequired() bool {
 		"gemma3",
 		"gemma3n",
 		"mistral3",
+		"qwen3",
 		"llama4",
 		"mllama",
 		"qwen25vl",

logutil/logutil.go

@@ -5,6 +5,8 @@ import (
 	"io"
 	"log/slog"
 	"path/filepath"
+	"runtime"
+	"time"
 )
 
 const LevelTrace slog.Level = -8
@@ -29,10 +31,18 @@ func NewLogger(w io.Writer, level slog.Level) *slog.Logger {
 	}))
 }
 
+type key string
+
 func Trace(msg string, args ...any) {
-	slog.Log(context.TODO(), LevelTrace, msg, args...)
+	TraceContext(context.WithValue(context.TODO(), key("skip"), 1), msg, args...)
 }
 
 func TraceContext(ctx context.Context, msg string, args ...any) {
-	slog.Log(ctx, LevelTrace, msg, args...)
+	if logger := slog.Default(); logger.Enabled(ctx, LevelTrace) {
+		skip, _ := ctx.Value(key("skip")).(int)
+		pc, _, _, _ := runtime.Caller(1 + skip)
+		record := slog.NewRecord(time.Now(), LevelTrace, msg, pc)
+		record.Add(args...)
+		logger.Handler().Handle(ctx, record)
+	}
 }

model/models/gemma2/model.go

@@ -63,7 +63,7 @@ func New(c fs.Config) (model.Model, error) {
 			attnValLen:        int(c.Uint("attention.value_length")),
 			eps:               c.Float("attention.layer_norm_rms_epsilon"),
 			ropeBase:          c.Float("rope.freq_base", 10000.0),
-			ropeScale:         c.Float("rope.freq_scale", 1.0),
+			ropeScale:         c.Float("rope.scaling.factor", 1.0),
 			attnLogitSoftcap:  c.Float("attn_logit_softcapping"),
 			finalLogitSoftcap: c.Float("final_logit_softcapping"),
 		},
@@ -88,7 +88,7 @@ func (sa *SelfAttention) Forward(ctx ml.Context, hiddenState, positionIDs ml.Ten
 
 	q := sa.Query.Forward(ctx, hiddenState)
 	q = q.Reshape(ctx, opts.attnKeyLen, opts.numHeads, batchSize)
-	q = fast.RoPE(ctx, q, positionIDs, opts.attnKeyLen, opts.ropeBase, opts.ropeScale, rope.WithTypeNeoX())
+	q = fast.RoPE(ctx, q, positionIDs, opts.attnKeyLen, opts.ropeBase, 1./opts.ropeScale, rope.WithTypeNeoX())
 
 	if opts.largeModelScaling {
 		q = q.Scale(ctx, 1.0/math.Sqrt(float64(opts.hiddenSize/opts.numHeads)))
@@ -98,7 +98,7 @@ func (sa *SelfAttention) Forward(ctx ml.Context, hiddenState, positionIDs ml.Ten
 
 	k := sa.Key.Forward(ctx, hiddenState)
 	k = k.Reshape(ctx, opts.attnKeyLen, opts.numKVHeads, batchSize)
-	k = fast.RoPE(ctx, k, positionIDs, opts.attnKeyLen, opts.ropeBase, opts.ropeScale, rope.WithTypeNeoX())
+	k = fast.RoPE(ctx, k, positionIDs, opts.attnKeyLen, opts.ropeBase, 1./opts.ropeScale, rope.WithTypeNeoX())
 
 	v := sa.Value.Forward(ctx, hiddenState)
 	v = v.Reshape(ctx, opts.attnValLen, opts.numKVHeads, batchSize)

model/models/gemma3/model_text.go

@@ -53,7 +53,7 @@ func newTextModel(c fs.Config) *TextModel {
 			eps:            c.Float("attention.layer_norm_rms_epsilon", 1e-06),
 			ropeLocalBase:  c.Float("rope.local.freq_base", 10000.0),
 			ropeGlobalBase: c.Float("rope.global.freq_base", 1000000.0),
-			ropeScale:      c.Float("rope.freq_scale", 1.0),
+			ropeScale:      c.Float("rope.scaling.factor", 1.0),
 		},
 	}
 
@@ -84,7 +84,7 @@ func (sa *TextSelfAttention) Forward(ctx ml.Context, layer int, hiddenState, pos
 	q := sa.Query.Forward(ctx, hiddenState)
 	q = q.Reshape(ctx, opts.attnKeyLen, opts.numHeads, batchSize)
 	q = sa.QueryNorm.Forward(ctx, q, opts.eps)
-	q = fast.RoPE(ctx, q, positionIDs, opts.attnKeyLen, ropeBase, opts.ropeScale, rope.WithTypeNeoX())
+	q = fast.RoPE(ctx, q, positionIDs, opts.attnKeyLen, ropeBase, 1./opts.ropeScale, rope.WithTypeNeoX())
 
 	if opts.largeModelScaling {
 		q = q.Scale(ctx, 1.0/math.Sqrt(float64(opts.hiddenSize/opts.numHeads)))
@@ -95,7 +95,7 @@ func (sa *TextSelfAttention) Forward(ctx ml.Context, layer int, hiddenState, pos
 	k := sa.Key.Forward(ctx, hiddenState)
 	k = k.Reshape(ctx, opts.attnKeyLen, opts.numKVHeads, batchSize)
 	k = sa.KeyNorm.Forward(ctx, k, opts.eps)
-	k = fast.RoPE(ctx, k, positionIDs, opts.attnKeyLen, ropeBase, opts.ropeScale, rope.WithTypeNeoX())
+	k = fast.RoPE(ctx, k, positionIDs, opts.attnKeyLen, ropeBase, 1./opts.ropeScale, rope.WithTypeNeoX())
 
 	v := sa.Value.Forward(ctx, hiddenState)
 	v = v.Reshape(ctx, opts.attnValLen, opts.numKVHeads, batchSize)

model/models/gemma3n/model_text.go

@@ -95,7 +95,7 @@ func (m *TextModel) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.T
 		ropeBase = m.ropeBaseLocal
 	}
 
-	return fast.RoPE(ctx, key, shift, m.headDim(), ropeBase, m.ropeScale, rope.WithTypeNeoX()), nil
+	return fast.RoPE(ctx, key, shift, m.headDim(), ropeBase, 1./m.ropeScale, rope.WithTypeNeoX()), nil
 }
 
 type TextScaledWordEmbedding struct {
@@ -256,14 +256,14 @@ func (attn TextAttention) Forward(ctx ml.Context, hiddenStates, positions ml.Ten
 	query := attn.Query.Forward(ctx, hiddenStates)
 	query = query.Reshape(ctx, opts.headDim(), opts.numHeads, batchSize)
 	query = attn.QueryNorm.Forward(ctx, query, opts.eps)
-	query = fast.RoPE(ctx, query, positions, opts.headDim(), ropeBase, opts.ropeScale, rope.WithTypeNeoX())
+	query = fast.RoPE(ctx, query, positions, opts.headDim(), ropeBase, 1./opts.ropeScale, rope.WithTypeNeoX())
 
 	var key, value ml.Tensor
 	if !sharedKV {
 		key = attn.Key.Forward(ctx, hiddenStates)
 		key = key.Reshape(ctx, opts.headDim(), opts.numKVHeads, batchSize)
 		key = attn.KeyNorm.Forward(ctx, key, opts.eps)
-		key = fast.RoPE(ctx, key, positions, opts.headDim(), ropeBase, opts.ropeScale, rope.WithTypeNeoX())
+		key = fast.RoPE(ctx, key, positions, opts.headDim(), ropeBase, 1./opts.ropeScale, rope.WithTypeNeoX())
 
 		value = attn.Value.Forward(ctx, hiddenStates)
 		value = value.Reshape(ctx, opts.headDim(), opts.numKVHeads, batchSize)
@@ -349,7 +349,7 @@ func newTextModel(c fs.Config) *TextModel {
 			eps:           c.Float("attention.layer_norm_rms_epsilon", 1e-06),
 			ropeBase:      c.Float("rope.freq_base", 1_000_000),
 			ropeBaseLocal: c.Float("rope.freq_base_local", 10_000),
-			ropeScale:     c.Float("rope.freq_scale", 1.0),
+			ropeScale:     c.Float("rope.scaling.factor", 1.0),
 
 			slidingWindowPattern:    c.Bools("attention.sliding_window_pattern"),
 			activationSparsityScale: c.Floats("activation_sparsity_scale"),

model/models/llama/model.go

@@ -2,7 +2,6 @@ package llama
 
 import (
 	"cmp"
-	"fmt"
 	"math"
 
 	"github.com/ollama/ollama/fs"
@@ -23,30 +22,26 @@ type Options struct {
 
 type Model struct {
 	model.Base
-	model.BytePairEncoding
+	model.TextProcessor
 
 	TokenEmbedding *nn.Embedding `gguf:"token_embd"`
 	Layers         []Layer       `gguf:"blk"`
 	OutputNorm     *nn.RMSNorm   `gguf:"output_norm"`
 	Output         *nn.Linear    `gguf:"output,alt:token_embd"`
 
-	*Options
+	Options
 }
 
 func New(c fs.Config) (model.Model, error) {
-	// This model currently only supports the gpt2 tokenizer
+	if c.Uint("expert_count") > 0 {
-	if c.String("tokenizer.ggml.model") == "llama" {
+		// TODO: support mixtures of experts
-		return nil, fmt.Errorf("unsupported tokenizer: llama")
+		return nil, model.ErrUnsupportedModel
 	}
-	// Best effort detection of library/deepseek-coder model(s) which are incompatible
+
-	if c.String("general.name") == "deepseek-ai" {
+	var processor model.TextProcessor
-		return nil, fmt.Errorf("unsupported model: %s", c.String("general.name"))
+	vocabulary := model.Vocabulary{
-	}
-	m := Model{
-		BytePairEncoding: model.NewBytePairEncoding(
-			c.String("tokenizer.ggml.pretokenizer", `(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\r\n\p{L}\p{N}]?\p{L}+|\p{N}{1,3}| ?[^\s\p{L}\p{N}]+[\r\n]*|\s*[\r\n]+|\s+(?!\S)|\s+`),
-			&model.Vocabulary{
 		Values: c.Strings("tokenizer.ggml.tokens"),
+		Scores: c.Floats("tokenizer.ggml.scores"),
 		Types:  c.Ints("tokenizer.ggml.token_type"),
 		Merges: c.Strings("tokenizer.ggml.merges"),
 		AddBOS: c.Bool("tokenizer.ggml.add_bos_token", true),
@@ -56,18 +51,31 @@ func New(c fs.Config) (model.Model, error) {
 			[]int32{int32(c.Uint("tokenizer.ggml.eos_token_id"))},
 			c.Ints("tokenizer.ggml.eos_token_ids")...,
 		),
-			},
+	}
-		),
+	switch c.String("tokenizer.ggml.model") {
+	case "gpt2":
+		processor = model.NewBytePairEncoding(
+			`(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\r\n\p{L}\p{N}]?\p{L}+|\p{N}{1,3}| ?[^\s\p{L}\p{N}]+[\r\n]*|\s*[\r\n]+|\s+(?!\S)|\s+`,
+			&vocabulary,
+		)
+	case "llama":
+		processor = model.NewSentencePiece(&vocabulary)
+	default:
+		return nil, model.ErrUnsupportedTokenizer
+	}
+
+	m := Model{
+		TextProcessor: processor,
 		Layers:        make([]Layer, c.Uint("block_count")),
-		Options: &Options{
+		Options: Options{
 			hiddenSize: int(c.Uint("embedding_length")),
 			numHeads:   int(c.Uint("attention.head_count")),
 			numKVHeads: int(c.Uint("attention.head_count_kv")),
 			headDim:    int(c.Uint("attention.key_length")),
 			ropeDim:    int(c.Uint("rope.dimension_count")),
 			eps:        c.Float("attention.layer_norm_rms_epsilon"),
-			ropeBase:   c.Float("rope.freq_base"),
+			ropeBase:   c.Float("rope.freq_base", 1e5),
-			ropeScale:  c.Float("rope.freq_scale", 1),
+			ropeScale:  c.Float("rope.scaling.factor", 1),
 		},
 	}
 
@@ -98,8 +106,8 @@ func (sa *SelfAttention) Forward(ctx ml.Context, hiddenState, positions ml.Tenso
 	value := sa.Value.Forward(ctx, hiddenState)
 	value = value.Reshape(ctx, headDim, opts.numKVHeads, batchSize)
 
-	query = fast.RoPE(ctx, query, positions, ropeDim, opts.ropeBase, opts.ropeScale, rope.WithFactors(sa.RopeFactors))
+	query = fast.RoPE(ctx, query, positions, ropeDim, opts.ropeBase, 1./opts.ropeScale, rope.WithFactors(sa.RopeFactors))
-	key = fast.RoPE(ctx, key, positions, ropeDim, opts.ropeBase, opts.ropeScale, rope.WithFactors(sa.RopeFactors))
+	key = fast.RoPE(ctx, key, positions, ropeDim, opts.ropeBase, 1./opts.ropeScale, rope.WithFactors(sa.RopeFactors))
 
 	attention := nn.Attention(ctx, query, key, value, 1.0/math.Sqrt(float64(headDim)), cache)
 	attention = attention.Reshape(ctx, headDim*opts.numHeads, batchSize)
@@ -108,7 +116,7 @@ func (sa *SelfAttention) Forward(ctx ml.Context, hiddenState, positions ml.Tenso
 
 func (m *Model) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) {
 	ropeDim := cmp.Or(m.ropeDim, m.hiddenSize/m.numHeads)
-	return fast.RoPE(ctx, key, shift, ropeDim, m.ropeBase, m.ropeScale, rope.WithFactors(m.Layers[layer].SelfAttention.RopeFactors)), nil
+	return fast.RoPE(ctx, key, shift, ropeDim, m.ropeBase, 1./m.ropeScale, rope.WithFactors(m.Layers[layer].SelfAttention.RopeFactors)), nil
 }
 
 type MLP struct {
@@ -163,7 +171,7 @@ func (m *Model) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
 			outputs = batch.Outputs
 		}
 
-		hiddenState = layer.Forward(ctx, hiddenState, positions, outputs, m.Cache, m.Options)
+		hiddenState = layer.Forward(ctx, hiddenState, positions, outputs, m.Cache, &m.Options)
 	}
 
 	hiddenState = m.OutputNorm.Forward(ctx, hiddenState, m.eps)

model/models/llama4/model_text.go

@@ -33,8 +33,8 @@ func (sa *TextAttention) Forward(ctx ml.Context, hiddenStates, positions, attent
 	value = value.Reshape(ctx, headDim, opts.numKVHeads, batchSize)
 
 	if useRope {
-		query = fast.RoPE(ctx, query, positions, opts.ropeDim, opts.ropeBase, opts.ropeScale, rope.WithFactors(sa.RopeFactors))
+		query = fast.RoPE(ctx, query, positions, opts.ropeDim, opts.ropeBase, 1./opts.ropeScale, rope.WithFactors(sa.RopeFactors))
-		key = fast.RoPE(ctx, key, positions, opts.ropeDim, opts.ropeBase, opts.ropeScale, rope.WithFactors(sa.RopeFactors))
+		key = fast.RoPE(ctx, key, positions, opts.ropeDim, opts.ropeBase, 1./opts.ropeScale, rope.WithFactors(sa.RopeFactors))
 	}
 
 	if opts.useQKNorm {
@@ -196,7 +196,7 @@ func newTextModel(c fs.Config) *TextModel {
 			numExpertsUsed:             int(c.Uint("expert_used_count")),
 			ropeDim:                    int(c.Uint("rope.dimension_count")),
 			ropeBase:                   c.Float("rope.freq_base"),
-			ropeScale:                  c.Float("rope.freq_scale", 1),
+			ropeScale:                  c.Float("rope.scaling.factor", 1),
 			eps:                        c.Float("attention.layer_norm_rms_epsilon"),
 			interleaveLayerStep:        int(c.Uint("interleave_moe_layer_step", 1)),
 			noRopeInterval:             int(c.Uint("no_rope_interval", 4)),
@@ -248,5 +248,5 @@ func (m *TextModel) Forward(ctx ml.Context, inputs, positions, outputs ml.Tensor
 }
 
 func (m *TextModel) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) {
-	return fast.RoPE(ctx, key, shift, m.ropeDim, m.ropeBase, m.ropeScale, rope.WithFactors(m.Layers[layer].Attention.RopeFactors)), nil
+	return fast.RoPE(ctx, key, shift, m.ropeDim, m.ropeBase, 1./m.ropeScale, rope.WithFactors(m.Layers[layer].Attention.RopeFactors)), nil
 }

model/models/mistral3/model_text.go

@@ -40,11 +40,11 @@ func (sa *SelfAttention) Forward(ctx ml.Context, hiddenState, positionIDs ml.Ten
 
 	q := sa.Query.Forward(ctx, hiddenState)
 	q = q.Reshape(ctx, headDim, opts.numHeads, batchSize)
-	q = fast.RoPE(ctx, q, positionIDs, opts.ropeDim, opts.ropeBase, opts.ropeScale)
+	q = fast.RoPE(ctx, q, positionIDs, opts.ropeDim, opts.ropeBase, 1./opts.ropeScale)
 
 	k := sa.Key.Forward(ctx, hiddenState)
 	k = k.Reshape(ctx, headDim, opts.numKVHeads, batchSize)
-	k = fast.RoPE(ctx, k, positionIDs, opts.ropeDim, opts.ropeBase, opts.ropeScale)
+	k = fast.RoPE(ctx, k, positionIDs, opts.ropeDim, opts.ropeBase, 1./opts.ropeScale)
 
 	v := sa.Value.Forward(ctx, hiddenState)
 	v = v.Reshape(ctx, headDim, opts.numKVHeads, batchSize)
@@ -55,7 +55,7 @@ func (sa *SelfAttention) Forward(ctx ml.Context, hiddenState, positionIDs ml.Ten
 }
 
 func (m *TextModel) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) {
-	return fast.RoPE(ctx, key, shift, m.ropeDim, m.ropeBase, m.ropeScale), nil
+	return fast.RoPE(ctx, key, shift, m.ropeDim, m.ropeBase, 1./m.ropeScale), nil
 }
 
 type MLP struct {
@@ -132,7 +132,7 @@ func newTextModel(c fs.Config) *TextModel {
 			ropeDim:    int(c.Uint("rope.dimension_count")),
 			eps:        c.Float("attention.layer_norm_rms_epsilon"),
 			ropeBase:   c.Float("rope.freq_base"),
-			ropeScale:  c.Float("rope.freq_scale", 1),
+			ropeScale:  c.Float("rope.scaling.factor", 1),
 		},
 	}
 }

model/models/mllama/model_text.go

@@ -26,11 +26,11 @@ func (sa *TextSelfAttention) Forward(ctx ml.Context, hiddenState, positions ml.T
 
 	query := sa.Query.Forward(ctx, hiddenState)
 	query = query.Reshape(ctx, headDim, opts.numHeads, batchSize)
-	query = fast.RoPE(ctx, query, positions, opts.ropeDim, opts.ropeBase, opts.ropeScale, rope.WithFactors(sa.RopeFactors))
+	query = fast.RoPE(ctx, query, positions, opts.ropeDim, opts.ropeBase, 1./opts.ropeScale, rope.WithFactors(sa.RopeFactors))
 
 	key := sa.Key.Forward(ctx, hiddenState)
 	key = key.Reshape(ctx, headDim, opts.numKVHeads, batchSize)
-	key = fast.RoPE(ctx, key, positions, opts.ropeDim, opts.ropeBase, opts.ropeScale, rope.WithFactors(sa.RopeFactors))
+	key = fast.RoPE(ctx, key, positions, opts.ropeDim, opts.ropeBase, 1./opts.ropeScale, rope.WithFactors(sa.RopeFactors))
 
 	value := sa.Value.Forward(ctx, hiddenState)
 	value = value.Reshape(ctx, headDim, opts.numKVHeads, batchSize)
@@ -45,7 +45,7 @@ func (sa *TextSelfAttention) Forward(ctx ml.Context, hiddenState, positions ml.T
 func (m *TextModel) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) {
 	// This will only get called for layers in the cache, which are just the self attention layers
 	if sa, ok := m.Transformer.Layers[layer].(*TextSelfAttentionDecoderLayer); ok {
-		return fast.RoPE(ctx, key, shift, m.ropeDim, m.ropeBase, m.ropeScale, rope.WithFactors(sa.SelfAttention.RopeFactors)), nil
+		return fast.RoPE(ctx, key, shift, m.ropeDim, m.ropeBase, 1./m.ropeScale, rope.WithFactors(sa.SelfAttention.RopeFactors)), nil
 	}
 
 	return key, nil
@@ -244,7 +244,7 @@ func newTextModel(c fs.Config) *TextModel {
 			ropeDim:              int(c.Uint("rope.dimension_count")),
 			eps:                  c.Float("attention.layer_norm_rms_epsilon"),
 			ropeBase:             c.Float("rope.freq_base"),
-			ropeScale:            c.Float("rope.freq_scale", 1),
+			ropeScale:            c.Float("rope.scaling.factor", 1),
 			crossAttentionLayers: c.Ints("attention.cross_attention_layers"),
 		},
 	}

model/models/qwen2/model.go

@@ -43,8 +43,8 @@ func (attn Attention) Forward(ctx ml.Context, hiddenStates, positions ml.Tensor,
 	value := attn.Value.Forward(ctx, hiddenStates)
 	value = value.Reshape(ctx, headDim, opts.numKVHeads, batchSize)
 
-	query = fast.RoPE(ctx, query, positions, ropeDim, opts.ropeBase, opts.ropeScale, rope.WithTypeNeoX())
+	query = fast.RoPE(ctx, query, positions, ropeDim, opts.ropeBase, 1./opts.ropeScale, rope.WithTypeNeoX())
-	key = fast.RoPE(ctx, key, positions, ropeDim, opts.ropeBase, opts.ropeScale, rope.WithTypeNeoX())
+	key = fast.RoPE(ctx, key, positions, ropeDim, opts.ropeBase, 1./opts.ropeScale, rope.WithTypeNeoX())
 
 	attention := nn.Attention(ctx, query, key, value, 1.0/math.Sqrt(float64(headDim)), cache)
 	attention = attention.Reshape(ctx, headDim*opts.numHeads, batchSize)
@@ -124,7 +124,7 @@ func (m Model) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
 
 func (m Model) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) {
 	ropeDim := cmp.Or(m.ropeDim, m.hiddenSize/m.numHeads)
-	return fast.RoPE(ctx, key, shift, ropeDim, m.ropeBase, m.ropeScale, rope.WithTypeNeoX()), nil
+	return fast.RoPE(ctx, key, shift, ropeDim, m.ropeBase, 1./m.ropeScale, rope.WithTypeNeoX()), nil
 }
 
 func New(c fs.Config) (model.Model, error) {
@@ -160,7 +160,7 @@ func New(c fs.Config) (model.Model, error) {
 			headDim:    int(c.Uint("attention.key_length")),
 			ropeDim:    int(c.Uint("rope.dimension_count")),
 			ropeBase:   c.Float("rope.freq_base"),
-			ropeScale:  c.Float("rope.freq_scale", 1),
+			ropeScale:  c.Float("rope.scaling.factor", 1),
 			eps:        c.Float("attention.layer_norm_rms_epsilon"),
 		},
 	}

model/models/qwen25vl/model_text.go

@@ -38,7 +38,7 @@ func NewTextModel(c fs.Config) *TextModel {
 			originalContextLength: int(c.Uint("context_length", 128000)),
 			eps:                   c.Float("attention.layer_norm_rms_epsilon"),
 			ropeBase:              c.Float("rope.freq_base"),
-			ropeScale:             c.Float("rope.freq_scale", 1),
+			ropeScale:             c.Float("rope.scaling.factor", 1),
 		},
 	}
 
@@ -60,11 +60,11 @@ func (sa *SelfAttention) Forward(ctx ml.Context, hiddenState, positionIDs ml.Ten
 
 	q := sa.Query.Forward(ctx, hiddenState)
 	q = q.Reshape(ctx, headDim, opts.numHeads, batchSize)
-	q = fast.RoPE(ctx, q, positionIDs, opts.ropeDim, opts.ropeBase, opts.ropeScale, rope.WithOriginalContextLength(opts.originalContextLength), rope.WithTypeNeoX())
+	q = fast.RoPE(ctx, q, positionIDs, opts.ropeDim, opts.ropeBase, 1./opts.ropeScale, rope.WithOriginalContextLength(opts.originalContextLength), rope.WithTypeNeoX())
 
 	k := sa.Key.Forward(ctx, hiddenState)
 	k = k.Reshape(ctx, headDim, opts.numKVHeads, batchSize)
-	k = fast.RoPE(ctx, k, positionIDs, opts.ropeDim, opts.ropeBase, opts.ropeScale, rope.WithOriginalContextLength(opts.originalContextLength), rope.WithTypeNeoX())
+	k = fast.RoPE(ctx, k, positionIDs, opts.ropeDim, opts.ropeBase, 1./opts.ropeScale, rope.WithOriginalContextLength(opts.originalContextLength), rope.WithTypeNeoX())
 
 	v := sa.Value.Forward(ctx, hiddenState)
 	v = v.Reshape(ctx, headDim, opts.numKVHeads, batchSize)
@@ -78,7 +78,7 @@ func (sa *SelfAttention) Forward(ctx ml.Context, hiddenState, positionIDs ml.Ten
 
 // Shift applies rotary position embeddings to the key tensor for causal attention caching
 func (m *TextModel) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) {
-	return fast.RoPE(ctx, key, shift, m.ropeDim, m.ropeBase, m.ropeScale, rope.WithOriginalContextLength(m.originalContextLength), rope.WithTypeNeoX()), nil
+	return fast.RoPE(ctx, key, shift, m.ropeDim, m.ropeBase, 1./m.ropeScale, rope.WithOriginalContextLength(m.originalContextLength), rope.WithTypeNeoX()), nil
 }
 
 // MLP implements the feed-forward network component with SwiGLU activation

model/models/qwen3/model.go

@@ -52,8 +52,8 @@ func (sa *Attention) Forward(ctx ml.Context, hiddenStates, positions ml.Tensor,
 	query = sa.QueryNorm.Forward(ctx, query, opts.eps)
 	key = sa.KeyNorm.Forward(ctx, key, opts.eps)
 
-	query = fast.RoPE(ctx, query, positions, opts.headDim(), opts.ropeBase, opts.ropeScale, rope.WithTypeNeoX())
+	query = fast.RoPE(ctx, query, positions, opts.headDim(), opts.ropeBase, 1./opts.ropeScale, rope.WithTypeNeoX())
-	key = fast.RoPE(ctx, key, positions, opts.headDim(), opts.ropeBase, opts.ropeScale, rope.WithTypeNeoX())
+	key = fast.RoPE(ctx, key, positions, opts.headDim(), opts.ropeBase, 1./opts.ropeScale, rope.WithTypeNeoX())
 
 	attention := nn.Attention(ctx, query, key, value, 1./math.Sqrt(float64(opts.headDim())), cache)
 	attention = attention.Reshape(ctx, attention.Dim(0)*attention.Dim(1), batchSize)
@@ -173,7 +173,7 @@ func (m *Model) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
 }
 
 func (m *Model) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) {
-	return fast.RoPE(ctx, key, shift, m.headDim(), m.ropeBase, m.ropeScale, rope.WithTypeNeoX()), nil
+	return fast.RoPE(ctx, key, shift, m.headDim(), m.ropeBase, 1./m.ropeScale, rope.WithTypeNeoX()), nil
 }
 
 var _ model.Model = (*Model)(nil)
@@ -213,7 +213,7 @@ func New(c fs.Config) (model.Model, error) {
 			valueLength:    int(c.Uint("attention.value_length")),
 			eps:            c.Float("attention.layer_norm_rms_epsilon"),
 			ropeBase:       c.Float("rope.freq_base"),
-			ropeScale:      c.Float("rope.freq_scale", 1),
+			ropeScale:      c.Float("rope.scaling.factor", 1),
 			numExperts:     int(c.Uint("expert_count")),
 			numExpertsUsed: int(c.Uint("expert_used_count")),
 			normTopKProb:   c.Bool("norm_top_k_prob", true),

runner/llamarunner/cache.go

@@ -204,13 +204,8 @@ func (c *InputCache) ShiftDiscard(inputLen int, numKeep int) int {
 	targetFree = max(targetFree, 1)
 
 	currentFree := c.numCtx - inputLen
-	discard := targetFree - currentFree
 
-	if discard < 0 {
+	return max(targetFree-currentFree, 0)
-		discard = 0
-	}
-
-	return discard
 }
 
 type ErrReprocessInputs struct {

runner/ollamarunner/cache.go

@@ -242,13 +242,8 @@ func (c *InputCache) ShiftDiscard(inputLen int32, numKeep int32) int32 {
 	targetFree = max(targetFree, 1)
 
 	currentFree := c.numCtx - inputLen
-	discard := targetFree - currentFree
 
-	if discard < 0 {
+	return max(targetFree-currentFree, 0)
-		discard = 0
-	}
-
-	return discard
 }
 
 type ErrReprocessInputs struct {

server/internal/internal/backoff/backoff.go

@@ -25,10 +25,7 @@ func Loop(ctx context.Context, maxBackoff time.Duration) iter.Seq2[int, error] {
 
 			// n^2 backoff timer is a little smoother than the
 			// common choice of 2^n.
-			d := time.Duration(n*n) * 10 * time.Millisecond
+			d := min(time.Duration(n*n)*10*time.Millisecond, maxBackoff)
-			if d > maxBackoff {
-				d = maxBackoff
-			}
 			// Randomize the delay between 0.5-1.5 x msec, in order
 			// to prevent accidental "thundering herd" problems.
 			d = time.Duration(float64(d) * (rand.Float64() + 0.5))

server/sched.go

@@ -382,10 +382,7 @@ func (pending *LlmRequest) useLoadedRunner(runner *runnerRef, finished chan *Llm
 // load creates a new model based on req and loads it. If requireFull is true then the model must be loaded fully onto GPUs
 // (if any). Returns whether the scheduler needs to evict a model to make this one fit.
 func (s *Scheduler) load(req *LlmRequest, f *ggml.GGML, gpus discover.GpuInfoList, requireFull bool) bool {
-	numParallel := int(envconfig.NumParallel())
+	numParallel := max(int(envconfig.NumParallel()), 1)
-	if numParallel < 1 {
-		numParallel = 1
-	}
 
 	// Embedding models should always be loaded with parallel=1
 	if req.model.CheckCapabilities(model.CapabilityCompletion) != nil {