Update ml/backend.go

This commit refactors the Rotary Position Embedding (RoPE) implementation across the codebase to use a structured configuration approach instead of individual parameters.

Key changes:
- Add new RoPEConfig struct with fields for dimension, type, base frequency, and scaling
- Add RopeType enum to formalize different RoPE implementation variants
- Add YarnConfig struct and related configuration for YaRN (Yet Another RoPE extensioN) context extension
- Update RoPE method signature across all tensor interfaces and implementations
- Refactor all model implementations (llama, gemma2, gemma3, mllama) to use the new configuration structure

This change improves code organization, makes the RoPE configuration more explicit, and provides better support for different RoPE variants and context extension methods.

CMakeLists.txt

@@ -86,9 +86,9 @@ if(CMAKE_CUDA_COMPILER)
     )
 endif()
 
-set(WINDOWS_AMDGPU_TARGETS_EXCLUDE_REGEX "^gfx(906|908|90a):xnack[+-]$"
+set(WINDOWS_AMDGPU_TARGETS_EXCLUDE_REGEX "^gfx(906|908|90a|1200|1201):xnack[+-]$"
     CACHE STRING
-    "Regular expression describing AMDGPU_TARGETS not supported on Windows. Override to force building these targets. Default \"^gfx(906|908|90a):xnack[+-]$\"."
+    "Regular expression describing AMDGPU_TARGETS not supported on Windows. Override to force building these targets. Default \"^gfx(906|908|90a|1200|1201):xnack[+-]$\"."
 )
 
 check_language(HIP)
@@ -97,7 +97,7 @@ if(CMAKE_HIP_COMPILER)
 
     find_package(hip REQUIRED)
     if(NOT AMDGPU_TARGETS)
-        list(FILTER AMDGPU_TARGETS INCLUDE REGEX "^gfx(900|94[012]|101[02]|1030|110[012])$")
+        list(FILTER AMDGPU_TARGETS INCLUDE REGEX "^gfx(900|94[012]|101[02]|1030|110[012]|120[01])$")
     elseif(WIN32 AND WINDOWS_AMDGPU_TARGETS_EXCLUDE_REGEX)
         list(FILTER AMDGPU_TARGETS EXCLUDE REGEX ${WINDOWS_AMDGPU_TARGETS_EXCLUDE_REGEX})
     endif()

CMakePresets.json

@@ -56,7 +56,7 @@
       "name": "ROCm 6",
       "inherits": [ "ROCm" ],
       "cacheVariables": {
-        "AMDGPU_TARGETS": "gfx900;gfx940;gfx941;gfx942;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102;gfx1151;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-"
+        "AMDGPU_TARGETS": "gfx900;gfx940;gfx941;gfx942;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102;gfx1151;gfx1200;gfx1201;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-"
       }
     }
   ],

README.md

@@ -285,6 +285,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Bionic GPT](https://github.com/bionic-gpt/bionic-gpt)
 - [HTML UI](https://github.com/rtcfirefly/ollama-ui)
 - [Saddle](https://github.com/jikkuatwork/saddle)
+- [TagSpaces](https://www.tagspaces.org) (A platform for file based apps, [utilizing Ollama](https://docs.tagspaces.org/ai/) for the generation of tags and descriptions)
 - [Chatbot UI](https://github.com/ivanfioravanti/chatbot-ollama)
 - [Chatbot UI v2](https://github.com/mckaywrigley/chatbot-ui)
 - [Typescript UI](https://github.com/ollama-interface/Ollama-Gui?tab=readme-ov-file)
@@ -394,6 +395,8 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Reins](https://github.com/ibrahimcetin/reins) (Easily tweak parameters, customize system prompts per chat, and enhance your AI experiments with reasoning model support.)
 - [Ellama](https://github.com/zeozeozeo/ellama) (Friendly native app to chat with an Ollama instance)
 - [screenpipe](https://github.com/mediar-ai/screenpipe) Build agents powered by your screen history
+- [Ollamb](https://github.com/hengkysteen/ollamb) (Simple yet rich in features, cross-platform built with Flutter and designed for Ollama. Try the [web demo](https://hengkysteen.github.io/demo/ollamb/).)
+- [Writeopia](https://github.com/Writeopia/Writeopia) (Text editor with integration with Ollama)
 
 ### Cloud
 
@@ -433,7 +436,9 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [SwollamaCLI](https://github.com/marcusziade/Swollama) bundled with the Swollama Swift package. [Demo](https://github.com/marcusziade/Swollama?tab=readme-ov-file#cli-usage)
 - [aichat](https://github.com/sigoden/aichat) All-in-one LLM CLI tool featuring Shell Assistant, Chat-REPL, RAG, AI tools & agents, with access to OpenAI, Claude, Gemini, Ollama, Groq, and more.
 - [PowershAI](https://github.com/rrg92/powershai) PowerShell module that brings AI to terminal on Windows, including support for Ollama
+- [DeepShell](https://github.com/Abyss-c0re/deepshell) Your self-hosted AI assistant. Interactive Shell, Files and Folders analysis.
 - [orbiton](https://github.com/xyproto/orbiton) Configuration-free text editor and IDE with support for tab completion with Ollama.
+- [orca-cli](https://github.com/molbal/orca-cli) Ollama Registry CLI Application - Browse, pull and download models from Ollama Registry in your terminal.
 
 ### Apple Vision Pro
 

discover/gpu_linux.go

@@ -111,6 +111,7 @@ func GetCPUDetails() ([]CPU, error) {
 	if err != nil {
 		return nil, err
 	}
+	defer file.Close()
 	return linuxCPUDetails(file)
 }
 
@@ -168,13 +169,11 @@ func linuxCPUDetails(file io.Reader) ([]CPU, error) {
 	for id, s := range socketByID {
 		s.CoreCount = len(coreBySocket[id])
 		s.ThreadCount = 0
-		for _, tc := range threadsByCoreBySocket[id] {
-			s.ThreadCount += tc
-		}
 
 		// This only works if HT is enabled, consider a more reliable model, maybe cache size comparisons?
 		efficiencyCoreCount := 0
 		for _, threads := range threadsByCoreBySocket[id] {
+			s.ThreadCount += threads
 			if threads == 1 {
 				efficiencyCoreCount++
 			}

docs/faq.md

@@ -20,7 +20,13 @@ Please refer to the [GPU docs](./gpu.md).
 
 ## How can I specify the context window size?
 
-By default, Ollama uses a context window size of 2048 tokens. This can be overridden with the `OLLAMA_CONTEXT_LENGTH` environment variable. For example, to set the default context length to 8K, use: `OLLAMA_CONTEXT_LENGTH=8192 ollama serve`.
+By default, Ollama uses a context window size of 2048 tokens. 
+
+This can be overridden with the `OLLAMA_CONTEXT_LENGTH` environment variable. For example, to set the default context window to 8K, use: 
+
+```shell
+OLLAMA_CONTEXT_LENGTH=8192 ollama serve
+```
 
 To change this when using `ollama run`, use `/set parameter`:
 

docs/troubleshooting.md

@@ -9,7 +9,7 @@ cat ~/.ollama/logs/server.log
 On **Linux** systems with systemd, the logs can be found with this command:
 
 ```shell
-journalctl -u ollama --no-pager
+journalctl -u ollama --no-pager --follow --pager-end 
 ```
 
 When you run Ollama in a **container**, the logs go to stdout/stderr in the container:

fs/ggml/ggml.go

@@ -413,7 +413,7 @@ func Decode(rs io.ReadSeeker, maxArraySize int) (*GGML, int64, error) {
 	}, offset, nil
 }
 
-func (f GGML) GraphSize(context, batch uint64, kvCacheType string) (kv, partialOffload, fullOffload uint64) {
+func (f GGML) GraphSize(context, batch uint64, numParallel int, kvCacheType string) (kv []uint64, partialOffload, fullOffload uint64) {
 	embedding := f.KV().EmbeddingLength()
 	heads := f.KV().HeadCount()
 	headsKV := f.KV().HeadCountKV()
@@ -426,7 +426,10 @@ func (f GGML) GraphSize(context, batch uint64, kvCacheType string) (kv, partialO
 	layers := f.Tensors().GroupLayers()
 
 	bytesPerElement := kvCacheBytesPerElement(kvCacheType)
-	kv = uint64(float64(context*f.KV().BlockCount()*(embeddingHeadsK+embeddingHeadsV)*headsKV) * bytesPerElement)
+	kv = make([]uint64, f.KV().BlockCount())
+	for i := range kv {
+		kv[i] = uint64(float64(context*(embeddingHeadsK+embeddingHeadsV)*headsKV) * bytesPerElement)
+	}
 
 	switch f.KV().Architecture() {
 	case "llama":
@@ -460,16 +463,14 @@ func (f GGML) GraphSize(context, batch uint64, kvCacheType string) (kv, partialO
 	case "mllama":
 		var visionTokens, tiles uint64 = 1601, 4
 
-		if crossAttentionLayers, ok := f.KV()["mllama.attention.cross_attention_layers"].(*array); ok {
-			kv = headsKV *
-				(embeddingHeadsK + embeddingHeadsV) * // one for K, one for V
-				(2* // sizeof(float16)
-					(f.KV().BlockCount()-uint64(crossAttentionLayers.size))* // num non-cross attention layers
-					context +
-					4* // sizeof(float32)
-						uint64(crossAttentionLayers.size)* // num cross attention layers
-						visionTokens*
-						tiles)
+		crossAttentionLayers := f.KV().Uints("attention.cross_attention_layers")
+		for i := range kv {
+			if slices.Contains(crossAttentionLayers, uint32(i)) {
+				kv[i] = headsKV * (embeddingHeadsK + embeddingHeadsV) *
+					4 * // sizeof(float32)
+					visionTokens *
+					tiles
+			}
 		}
 
 		fullOffload = max(
@@ -505,6 +506,20 @@ func (f GGML) GraphSize(context, batch uint64, kvCacheType string) (kv, partialO
 				4*embeddingHeadsK*context*8+
 				embedding*embeddingHeadsK*heads*9/16,
 		)
+
+		// Gemma2 also has sliding window attention but we only have an optimized implementation in the Ollama
+		// engine. Gemma3 always uses the Ollama engine.
+		if f.KV().Architecture() == "gemma3" {
+			const gemma3GlobalCacheCount = 6
+			slidingWindow := (uint64(numParallel) * uint64(f.KV().Uint("attention.sliding_window"))) + batch
+			for i := range kv {
+				// Every 6th layer is a global layer, which is the full context size that has already been set. The other
+				// layers are the smaller local (sliding) layers.
+				if (i+1)%gemma3GlobalCacheCount != 0 {
+					kv[i] = uint64(float64(slidingWindow*(embeddingHeadsK+embeddingHeadsV)*headsKV) * bytesPerElement)
+				}
+			}
+		}
 	case "command-r":
 		fullOffload = max(
 			4*batch*(embedding+vocab),

kvcache/causal.go

@@ -119,10 +119,10 @@ func (c *Causal) Init(backend ml.Backend, dtype ml.DType, maxSequences, capacity
 	}
 
 	var cacheSize int
-	if c.windowSize == math.MaxInt32 || capacity < int(c.windowSize)+maxBatch {
+	if c.windowSize == math.MaxInt32 || capacity < int(c.windowSize) {
 		cacheSize = maxSequences * capacity
 	} else {
-		cacheSize = maxSequences * (int(c.windowSize) + maxBatch)
+		cacheSize = (maxSequences * int(c.windowSize)) + maxBatch
 	}
 	cacheSize = roundUp(cacheSize, c.config.CachePadding)
 	c.cells = make([]cacheCell, cacheSize)

kvcache/causal_test.go

@@ -362,7 +362,6 @@ func (c *testContext) FromIntSlice(s []int32, shape ...int) (ml.Tensor, error) {
 }
 
 func (c *testContext) Input() ml.Context    { return c }
-func (c *testContext) Output() ml.Context   { return c }
 func (c *testContext) Layer(int) ml.Context { return c }
 
 func (c *testContext) Forward(...ml.Tensor) ml.Context { return c }
@@ -463,7 +462,7 @@ func (t *testTensor) Conv2D(ctx ml.Context, weight ml.Tensor, s0, s1, p0, p1, d0
 	panic("not implemented")
 }
 
-func (t *testTensor) RoPE(ctx ml.Context, positionIDs, ropeFactors ml.Tensor, dim, ropeType uint32, base, scale float32) ml.Tensor {
+func (t *testTensor) RoPE(ctx ml.Context, positionIDs, ropeFactors ml.Tensor, config ml.RoPEConfig) ml.Tensor {
 	panic("not implemented")
 }
 

llama/llama.go

@@ -166,6 +166,10 @@ func (c *Context) KvCacheDefrag() {
 	C.llama_kv_cache_defrag(c.c)
 }
 
+func (c *Context) KvCacheCanShift() bool {
+	return bool(C.llama_kv_cache_can_shift(c.c))
+}
+
 // Get the embeddings for a sequence id
 func (c *Context) GetEmbeddingsSeq(seqId int) []float32 {
 	e := unsafe.Pointer(C.llama_get_embeddings_seq(c.c, C.int(seqId)))

llama/patches/0022-add-rdna4-support.patch

@@ -0,0 +1,103 @@
+From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
+From: Saman <saman.khatir@amd.com>
+Date: Wed, 19 Mar 2025 14:02:26 -0700
+Subject: [PATCH] add rdna4 support
+
+---
+ ggml/src/ggml-cuda/common.cuh    | 6 ++++--
+ ggml/src/ggml-cuda/mmq.cu        | 2 +-
+ ggml/src/ggml-cuda/mmq.cuh       | 4 ++--
+ ggml/src/ggml-cuda/mmvq.cu       | 4 ++--
+ ggml/src/ggml-cuda/vendors/hip.h | 4 ++++
+ 5 files changed, 13 insertions(+), 7 deletions(-)
+
+diff --git a/ggml/src/ggml-cuda/common.cuh b/ggml/src/ggml-cuda/common.cuh
+index adf0d3ec..b24593fc 100644
+--- a/ggml/src/ggml-cuda/common.cuh
++++ b/ggml/src/ggml-cuda/common.cuh
+@@ -61,11 +61,13 @@
+ #define GGML_CUDA_CC_RDNA1      (GGML_CUDA_CC_OFFSET_AMD + 0x1010) // RX 5000
+ #define GGML_CUDA_CC_RDNA2      (GGML_CUDA_CC_OFFSET_AMD + 0x1030) // RX 6000, minimum for dp4a
+ #define GGML_CUDA_CC_RDNA3      (GGML_CUDA_CC_OFFSET_AMD + 0x1100) // RX 7000, minimum for WMMA
++#define GGML_CUDA_CC_RDNA4      (GGML_CUDA_CC_OFFSET_AMD + 0x1200) // RX 9000
+ 
+ #define GGML_CUDA_CC_IS_RDNA(cc)  (cc >= GGML_CUDA_CC_RDNA1)
+ #define GGML_CUDA_CC_IS_RDNA1(cc) (cc >= GGML_CUDA_CC_RDNA1 && cc < GGML_CUDA_CC_RDNA2)
+ #define GGML_CUDA_CC_IS_RDNA2(cc) (cc >= GGML_CUDA_CC_RDNA2 && cc < GGML_CUDA_CC_RDNA3)
+-#define GGML_CUDA_CC_IS_RDNA3(cc) (cc >= GGML_CUDA_CC_RDNA3)
++#define GGML_CUDA_CC_IS_RDNA3(cc) (cc >= GGML_CUDA_CC_RDNA3 && cc < GGML_CUDA_CC_RDNA4)
++#define GGML_CUDA_CC_IS_RDNA4(cc) (cc >= GGML_CUDA_CC_RDNA4)
+ #define GGML_CUDA_CC_IS_GCN(cc)   (cc > GGML_CUDA_CC_OFFSET_AMD && cc < GGML_CUDA_CC_CDNA)
+ #define GGML_CUDA_CC_IS_CDNA(cc)  (cc >= GGML_CUDA_CC_CDNA && cc < GGML_CUDA_CC_RDNA1)
+ 
+@@ -386,7 +388,7 @@ static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, i
+ #if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+ #if defined(__gfx906__) || defined(__gfx908__) || defined(__gfx90a__) || defined(RDNA2)
+     c = __builtin_amdgcn_sdot4(a, b, c, false);
+-#elif defined(RDNA3)
++#elif defined(RDNA3) || defined(RDNA4)
+     c = __builtin_amdgcn_sudot4( true, a, true, b, c, false);
+ #elif defined(__gfx1010__) || defined(__gfx900__)
+     int tmp1;
+diff --git a/ggml/src/ggml-cuda/mmq.cu b/ggml/src/ggml-cuda/mmq.cu
+index 10f2ebb1..933d945c 100644
+--- a/ggml/src/ggml-cuda/mmq.cu
++++ b/ggml/src/ggml-cuda/mmq.cu
+@@ -149,5 +149,5 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11) {
+         return !fp16_mma_hardware_available(cc) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
+     }
+ 
+-    return (!GGML_CUDA_CC_IS_RDNA3(cc) && !GGML_CUDA_CC_IS_CDNA(cc)) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
++    return (!GGML_CUDA_CC_IS_RDNA4(cc) && !GGML_CUDA_CC_IS_RDNA3(cc) && !GGML_CUDA_CC_IS_CDNA(cc)) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
+ }
+diff --git a/ggml/src/ggml-cuda/mmq.cuh b/ggml/src/ggml-cuda/mmq.cuh
+index 0451c65f..66ce2bc9 100644
+--- a/ggml/src/ggml-cuda/mmq.cuh
++++ b/ggml/src/ggml-cuda/mmq.cuh
+@@ -2577,9 +2577,9 @@ static __device__ void mul_mat_q_process_tile(
+ 
+ template <ggml_type type, int mmq_x, int nwarps, bool need_check>
+ #if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+-#if defined(RDNA3) || defined(RDNA2) || defined(CDNA) || defined(GCN)
++#if defined(RDNA4) || defined(RDNA3) || defined(RDNA2) || defined(CDNA) || defined(GCN)
+     __launch_bounds__(WARP_SIZE*nwarps, 2)
+-#endif // defined(RDNA3) || defined(RDNA2) || defined(CDNA) || defined(GCN)
++#endif // defined(RDNA4) || defined(RDNA3) || defined(RDNA2) || defined(CDNA) || defined(GCN)
+ #else
+ #if __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
+     __launch_bounds__(WARP_SIZE*nwarps, 1)
+diff --git a/ggml/src/ggml-cuda/mmvq.cu b/ggml/src/ggml-cuda/mmvq.cu
+index 4fb466ca..23ae7abc 100644
+--- a/ggml/src/ggml-cuda/mmvq.cu
++++ b/ggml/src/ggml-cuda/mmvq.cu
+@@ -62,13 +62,13 @@ static __global__ void mul_mat_vec_q(
+ 
+     constexpr vec_dot_q_cuda_t vec_dot_q_cuda = get_vec_dot_q_cuda(type);
+ 
+-#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && (defined(RDNA2) || defined(RDNA3))
++#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && (defined(RDNA2) || defined(RDNA3) || defined(RDNA4))
+     constexpr int nwarps              = 1;
+     constexpr int rows_per_cuda_block = 1;
+ #else
+     constexpr int nwarps              = ncols_y <= 4 ? 4 : 2;
+     constexpr int rows_per_cuda_block = ncols_y == 1 ? 1 : 2;
+-#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && !defined(RDNA2) && !defined(RDNA3)
++#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && !defined(RDNA2) && !defined(RDNA3) && !defined(RDNA4)
+ 
+     const     int tid = WARP_SIZE*threadIdx.y + threadIdx.x;
+     const     int row0 = rows_per_cuda_block*blockIdx.x;
+diff --git a/ggml/src/ggml-cuda/vendors/hip.h b/ggml/src/ggml-cuda/vendors/hip.h
+index 81964611..a62544b5 100644
+--- a/ggml/src/ggml-cuda/vendors/hip.h
++++ b/ggml/src/ggml-cuda/vendors/hip.h
+@@ -150,6 +150,10 @@
+ #define CDNA
+ #endif
+ 
++#if defined(__gfx1200__) || defined(__gfx1201__)
++#define RDNA4
++#endif
++
+ #if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx1103__) || \
+     defined(__gfx1150__) || defined(__gfx1151__)
+ #define RDNA3

llm/memory.go

@@ -15,12 +15,12 @@ import (
 )
 
 // This algorithm looks for a complete fit to determine if we need to unload other models
-func PredictServerFit(allGpus discover.GpuInfoList, f *ggml.GGML, adapters, projectors []string, opts api.Options) (bool, uint64) {
+func PredictServerFit(allGpus discover.GpuInfoList, f *ggml.GGML, adapters, projectors []string, opts api.Options, numParallel int) (bool, uint64) {
 	// Split up the GPUs by type and try them
 	var estimatedVRAM uint64
 	for _, gpus := range allGpus.ByLibrary() {
 		var layerCount int
-		estimate := EstimateGPULayers(gpus, f, projectors, opts)
+		estimate := EstimateGPULayers(gpus, f, projectors, opts, numParallel)
 		layerCount, estimatedVRAM = estimate.Layers, estimate.VRAMSize
 		if opts.NumGPU < 0 {
 			if layerCount > 0 && layerCount >= int(f.KV().BlockCount()+1) {
@@ -71,7 +71,7 @@ type MemoryEstimate struct {
 
 // Given a model and one or more GPU targets, predict how many layers and bytes we can load, and the total size
 // The GPUs provided must all be the same Library
-func EstimateGPULayers(gpus []discover.GpuInfo, f *ggml.GGML, projectors []string, opts api.Options) MemoryEstimate {
+func EstimateGPULayers(gpus []discover.GpuInfo, f *ggml.GGML, projectors []string, opts api.Options, numParallel int) MemoryEstimate {
 	// Graph size for a partial offload, applies to all GPUs
 	var graphPartialOffload uint64
 
@@ -137,13 +137,19 @@ func EstimateGPULayers(gpus []discover.GpuInfo, f *ggml.GGML, projectors []strin
 		}
 	}
 
-	kv, graphPartialOffload, graphFullOffload := f.GraphSize(uint64(opts.NumCtx), uint64(min(opts.NumCtx, opts.NumBatch)), kvct)
+	kv, graphPartialOffload, graphFullOffload := f.GraphSize(uint64(opts.NumCtx), uint64(min(opts.NumCtx, opts.NumBatch)), numParallel, kvct)
 
-	// KV is proportional to the number of layers
-	layerSize += kv / f.KV().BlockCount()
+	if len(kv) > 0 {
+		layerSize += kv[0]
+	}
+
+	var kvTotal uint64
+	for _, kvLayer := range kv {
+		kvTotal += kvLayer
+	}
 
 	if graphPartialOffload == 0 {
-		graphPartialOffload = f.KV().GQA() * kv / 6
+		graphPartialOffload = f.KV().GQA() * kvTotal / 6
 	}
 	if graphFullOffload == 0 {
 		graphFullOffload = graphPartialOffload
@@ -217,7 +223,7 @@ func EstimateGPULayers(gpus []discover.GpuInfo, f *ggml.GGML, projectors []strin
 		// Some models have inconsistent layer sizes
 		if blk, ok := layers[fmt.Sprintf("blk.%d", i)]; ok {
 			layerSize = blk.Size()
-			layerSize += kv / f.KV().BlockCount()
+			layerSize += kv[i]
 			memoryWeights += blk.Size()
 		}
 
@@ -315,7 +321,7 @@ func EstimateGPULayers(gpus []discover.GpuInfo, f *ggml.GGML, projectors []strin
 		layersRequested:     opts.NumGPU,
 		layersModel:         int(f.KV().BlockCount()) + 1,
 		availableList:       availableList,
-		kv:                  kv,
+		kv:                  kvTotal,
 		allocationsList:     allocationsList,
 		memoryWeights:       memoryWeights,
 		memoryLayerOutput:   memoryLayerOutput,
@@ -374,7 +380,7 @@ func (m MemoryEstimate) LogValue() slog.Value {
 			slog.Group(
 				"weights",
 				// memory of the weights
-				"total", format.HumanBytes2(m.memoryWeights),
+				"total", format.HumanBytes2(m.memoryWeights+m.memoryLayerOutput),
 				// memory of repeating layers
 				"repeating", format.HumanBytes2(m.memoryWeights),
 				// memory of non-repeating layers

llm/memory_test.go

@@ -61,7 +61,7 @@ func TestEstimateGPULayers(t *testing.T) {
 	projectors := []string{}
 	opts := api.DefaultOptions()
 	t.Run("cpu", func(t *testing.T) {
-		estimate := EstimateGPULayers(gpus, ggml, projectors, opts)
+		estimate := EstimateGPULayers(gpus, ggml, projectors, opts, 1)
 		assert.Equal(t, 0, estimate.Layers)
 		assert.Equal(t, uint64(0), estimate.Graph)
 	})
@@ -112,7 +112,7 @@ func TestEstimateGPULayers(t *testing.T) {
 			gpus[1].FreeMemory += gpuMinimumMemory + layerSize + s.layer1*layerSize + 1
 			gpus[0].FreeMemory += max(graphFullOffload, graphPartialOffload)
 			gpus[1].FreeMemory += max(graphFullOffload, graphPartialOffload)
-			estimate := EstimateGPULayers(gpus, ggml, projectors, opts)
+			estimate := EstimateGPULayers(gpus, ggml, projectors, opts, 1)
 			assert.Equal(t, int(s.expect0+s.expect1), estimate.Layers, "scenario %d: %v", i, s)
 			assert.Equal(t, fmt.Sprintf("%d,%d", s.expect0, s.expect1), estimate.TensorSplit, "scenario %d: %v", i, s)
 			var layerSums uint64

llm/server.go

@@ -109,7 +109,7 @@ func NewLlamaServer(gpus discover.GpuInfoList, modelPath string, f *ggml.GGML, a
 		gpus = discover.GetCPUInfo()
 	}
 
-	estimate := EstimateGPULayers(gpus, f, projectors, opts)
+	estimate := EstimateGPULayers(gpus, f, projectors, opts, numParallel)
 	if len(gpus) > 1 || gpus[0].Library != "cpu" {
 		switch {
 		case gpus[0].Library == "metal" && estimate.VRAMSize > systemTotalMemory:

ml/backend.go

@@ -110,16 +110,61 @@ type Context interface {
 	MaxGraphNodes() int
 	Close()
 
-	// Input returns a context appropriate for creating input tensors
+	// Input returns a context appropriate for creating tensors that are
+	// inputs to the model (which includes things like output locations)
 	Input() Context
 
-	// Output returns a context appropriate for creating output tensors
-	Output() Context
-
 	// Layer returns a context appropriate for creating intermediate tensors
 	Layer(int) Context
 }
 
+// RopeType represents different RoPE (Rotary Position Embedding) implementation types
+type RopeType int
+
+// Available RoPE implementation types
+const (
+	RopeTypeNormal RopeType = iota // Standard RoPE implementation
+	RopeTypeNeox                   // NeoX-style RoPE implementation
+	RopeTypeMRoPE                  // Multimodal RoPE implementation
+	RopeTypeVision                 // Vision-specific RoPE implementation
+)
+
+type YarnConfig struct {
+	YarnCtxTrain   int     // Context size used during training (for YaRN scaling)
+	YarnExtFactor  float32 // Extension factor for YaRN
+	YarnAttnFactor float32 // Attention scaling factor for YaRN
+	YarnBetaFast   float32 // Fast decay parameter for YaRN
+	YarnBetaSlow   float32 // Slow decay parameter for YaRN
+}
+
+// DefaultYarnConfig returns a default configuration for YaRN (Yet Another Rope Extension)
+func DefaultYarnConfig(nCtx int32) *YarnConfig {
+	return &YarnConfig{
+		YarnCtxTrain:   int(nCtx),
+		YarnExtFactor:  0.0,
+		YarnAttnFactor: 1.0,
+		YarnBetaFast:   32.0,
+		YarnBetaSlow:   1.0,
+	}
+}
+
+// RoPEConfig holds configuration for Rotary Position Embedding
+type RoPEConfig struct {
+	// Dim is the dimensionality for applying rotary embeddings
+	Dim uint32
+
+	// Type specifies the RoPE implementation variant
+	Type RopeType
+
+	// Base controls frequency decay for the embeddings
+	Base float32
+
+	// Scale allows scaling the effective context length
+	Scale float32
+
+	*YarnConfig
+}
+
 type Tensor interface {
 	Dim(n int) int
 	Stride(n int) int
@@ -143,7 +188,7 @@ type Tensor interface {
 	AvgPool2D(ctx Context, k, s int, p float32) Tensor
 	Conv2D(ctx Context, weight Tensor, s0, s1, p0, p1, d0, d1 int) Tensor
 
-	RoPE(ctx Context, positionIDs, ropeFactors Tensor, dim, ropeType uint32, base, scale float32) Tensor
+	RoPE(ctx Context, positionIDs, ropeFactors Tensor, config RoPEConfig) Tensor
 
 	Tanh(ctx Context) Tensor
 	GELU(ctx Context) Tensor

ml/backend/ggml/ggml.go

@@ -48,9 +48,6 @@ type Backend struct {
 	// input is the backend used for inputs
 	input *C.struct_ggml_backend_buffer_type
 
-	// output is the backend used for outputs
-	output *C.struct_ggml_backend_buffer_type
-
 	// layers is the backend used for repeating layers
 	layers map[int]*C.struct_ggml_backend_buffer_type
 
@@ -400,8 +397,7 @@ func New(ctx context.Context, r *os.File, params ml.BackendParams) (ml.Backend,
 			C.size_t(maxGraphNodes),
 			C._Bool(len(gpus) > 1 && slices.Contains(gpus, output.d)),
 		),
-		input:  deviceBufferTypes[input.d],
-		output: deviceBufferTypes[output.d],
+		input: deviceBufferTypes[input.d],
 		layers: func() map[int]*C.struct_ggml_backend_buffer_type {
 			m := make(map[int]*C.struct_ggml_backend_buffer_type)
 			for i, layer := range layers {
@@ -482,19 +478,6 @@ func (c Context) Input() ml.Context {
 	return &c
 }
 
-func (c Context) Output() ml.Context {
-	if c.b.output != nil {
-		return &Context{
-			b:             c.b,
-			ctx:           c.ctx,
-			buft:          c.b.output,
-			maxGraphNodes: c.maxGraphNodes,
-		}
-	}
-
-	return &c
-}
-
 func (c Context) Layer(i int) ml.Context {
 	if buft, ok := c.b.layers[i]; ok {
 		return &Context{
@@ -924,6 +907,8 @@ func (t *Tensor) View(ctx ml.Context, offset int, shape ...int) ml.Tensor {
 	}
 }
 
+// GGML RoPE types
+// These are the types used in the C implementation of RoPE
 const (
 	ropeTypeNorm   C.int = 0
 	ropeTypeNeox   C.int = 2
@@ -931,7 +916,8 @@ const (
 	ropeTypeVision C.int = 24
 )
 
-func (t *Tensor) RoPE(ctx ml.Context, positionIDs, ropeFactors ml.Tensor, ropeDim, ropeType uint32, ropeBase, ropeScale float32) ml.Tensor {
+// RoPE applies Rotary Position Embeddings to the tensor
+func (t *Tensor) RoPE(ctx ml.Context, positionIDs, ropeFactors ml.Tensor, config ml.RoPEConfig) ml.Tensor {
 	if ropeFactors == nil {
 		ropeFactors = &Tensor{b: t.b}
 	}
@@ -941,19 +927,41 @@ func (t *Tensor) RoPE(ctx ml.Context, positionIDs, ropeFactors ml.Tensor, ropeDi
 		dequant = C.ggml_cast(ctx.(*Context).ctx, t.t, C.GGML_TYPE_F32)
 	}
 
+	if config.YarnConfig == nil {
+		config.YarnConfig = ml.DefaultYarnConfig(131072) // 131072 is the default for LLaMA, so it is common at the time of writing
+	}
+
+	// Map Go RopeType to C implementation constants
+	var ropeTypeC C.int
+	switch config.Type {
+	case ml.RopeTypeNormal:
+		ropeTypeC = ropeTypeNorm
+	case ml.RopeTypeNeox:
+		ropeTypeC = ropeTypeNeox
+	case ml.RopeTypeMRoPE:
+		ropeTypeC = ropeTypeMrope
+	case ml.RopeTypeVision:
+		ropeTypeC = ropeTypeVision
+	default:
+		ropeTypeC = ropeTypeNorm
+	}
+
 	return &Tensor{
 		b: t.b,
 		t: C.ggml_rope_ext(
-			ctx.(*Context).ctx, dequant, positionIDs.(*Tensor).t, ropeFactors.(*Tensor).t,
-			C.int(ropeDim),
-			C.int(ropeType),
-			131072, // YaRN n_ctx_train
-			C.float(ropeBase),
-			C.float(ropeScale),
-			0.,  // YaRN ext_factor
-			1.,  // YaRN attn_factor
-			32., // YaRN beta_fast
-			1.,  // YaRN beta_slow
+			ctx.(*Context).ctx,
+			dequant,
+			positionIDs.(*Tensor).t,
+			ropeFactors.(*Tensor).t,
+			C.int(config.Dim),
+			ropeTypeC,
+			C.int(config.YarnCtxTrain),
+			C.float(config.Base),
+			C.float(config.Scale),
+			C.float(config.YarnExtFactor),
+			C.float(config.YarnAttnFactor),
+			C.float(config.YarnBetaFast),
+			C.float(config.YarnBetaSlow),
 		),
 	}
 }

ml/backend/ggml/ggml/src/ggml-cuda/common.cuh

@@ -61,11 +61,13 @@
 #define GGML_CUDA_CC_RDNA1      (GGML_CUDA_CC_OFFSET_AMD + 0x1010) // RX 5000
 #define GGML_CUDA_CC_RDNA2      (GGML_CUDA_CC_OFFSET_AMD + 0x1030) // RX 6000, minimum for dp4a
 #define GGML_CUDA_CC_RDNA3      (GGML_CUDA_CC_OFFSET_AMD + 0x1100) // RX 7000, minimum for WMMA
+#define GGML_CUDA_CC_RDNA4      (GGML_CUDA_CC_OFFSET_AMD + 0x1200) // RX 9000
 
 #define GGML_CUDA_CC_IS_RDNA(cc)  (cc >= GGML_CUDA_CC_RDNA1)
 #define GGML_CUDA_CC_IS_RDNA1(cc) (cc >= GGML_CUDA_CC_RDNA1 && cc < GGML_CUDA_CC_RDNA2)
 #define GGML_CUDA_CC_IS_RDNA2(cc) (cc >= GGML_CUDA_CC_RDNA2 && cc < GGML_CUDA_CC_RDNA3)
-#define GGML_CUDA_CC_IS_RDNA3(cc) (cc >= GGML_CUDA_CC_RDNA3)
+#define GGML_CUDA_CC_IS_RDNA3(cc) (cc >= GGML_CUDA_CC_RDNA3 && cc < GGML_CUDA_CC_RDNA4)
+#define GGML_CUDA_CC_IS_RDNA4(cc) (cc >= GGML_CUDA_CC_RDNA4)
 #define GGML_CUDA_CC_IS_GCN(cc)   (cc > GGML_CUDA_CC_OFFSET_AMD && cc < GGML_CUDA_CC_CDNA)
 #define GGML_CUDA_CC_IS_CDNA(cc)  (cc >= GGML_CUDA_CC_CDNA && cc < GGML_CUDA_CC_RDNA1)
 
@@ -386,7 +388,7 @@ static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, i
 #if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 #if defined(__gfx906__) || defined(__gfx908__) || defined(__gfx90a__) || defined(RDNA2)
     c = __builtin_amdgcn_sdot4(a, b, c, false);
-#elif defined(RDNA3)
+#elif defined(RDNA3) || defined(RDNA4)
     c = __builtin_amdgcn_sudot4( true, a, true, b, c, false);
 #elif defined(__gfx1010__) || defined(__gfx900__)
     int tmp1;

ml/backend/ggml/ggml/src/ggml-cuda/mmq.cu

@@ -149,5 +149,5 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11) {
         return !fp16_mma_hardware_available(cc) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
     }
 
-    return (!GGML_CUDA_CC_IS_RDNA3(cc) && !GGML_CUDA_CC_IS_CDNA(cc)) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
+    return (!GGML_CUDA_CC_IS_RDNA4(cc) && !GGML_CUDA_CC_IS_RDNA3(cc) && !GGML_CUDA_CC_IS_CDNA(cc)) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
 }

ml/backend/ggml/ggml/src/ggml-cuda/mmq.cuh

@@ -2577,9 +2577,9 @@ static __device__ void mul_mat_q_process_tile(
 
 template <ggml_type type, int mmq_x, int nwarps, bool need_check>
 #if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
-#if defined(RDNA3) || defined(RDNA2) || defined(CDNA) || defined(GCN)
+#if defined(RDNA4) || defined(RDNA3) || defined(RDNA2) || defined(CDNA) || defined(GCN)
     __launch_bounds__(WARP_SIZE*nwarps, 2)
-#endif // defined(RDNA3) || defined(RDNA2) || defined(CDNA) || defined(GCN)
+#endif // defined(RDNA4) || defined(RDNA3) || defined(RDNA2) || defined(CDNA) || defined(GCN)
 #else
 #if __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
     __launch_bounds__(WARP_SIZE*nwarps, 1)

ml/backend/ggml/ggml/src/ggml-cuda/mmvq.cu

@@ -62,13 +62,13 @@ static __global__ void mul_mat_vec_q(
 
     constexpr vec_dot_q_cuda_t vec_dot_q_cuda = get_vec_dot_q_cuda(type);
 
-#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && (defined(RDNA2) || defined(RDNA3))
+#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && (defined(RDNA2) || defined(RDNA3) || defined(RDNA4))
     constexpr int nwarps              = 1;
     constexpr int rows_per_cuda_block = 1;
 #else
     constexpr int nwarps              = ncols_y <= 4 ? 4 : 2;
     constexpr int rows_per_cuda_block = ncols_y == 1 ? 1 : 2;
-#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && !defined(RDNA2) && !defined(RDNA3)
+#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && !defined(RDNA2) && !defined(RDNA3) && !defined(RDNA4)
 
     const     int tid = WARP_SIZE*threadIdx.y + threadIdx.x;
     const     int row0 = rows_per_cuda_block*blockIdx.x;

ml/backend/ggml/ggml/src/ggml-cuda/vendors/hip.h

@@ -150,6 +150,10 @@
 #define CDNA
 #endif
 
+#if defined(__gfx1200__) || defined(__gfx1201__)
+#define RDNA4
+#endif
+
 #if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx1103__) || \
     defined(__gfx1150__) || defined(__gfx1151__)
 #define RDNA3

model/models/gemma2/model.go

@@ -13,10 +13,11 @@ import (
 type Options struct {
 	hiddenSize, numHeads, numKVHeads int
 	attnKeyLen, attnValLen           int
-	eps, ropeBase, ropeScale         float32
+	eps                              float32
 	attnLogitSoftcap                 float32
 	finalLogitSoftcap                float32
 	largeModelScaling                bool
+	ropeConfig                       ml.RoPEConfig
 }
 
 type Model struct {
@@ -55,10 +56,15 @@ func New(c ml.Config) (model.Model, error) {
 			attnKeyLen:        int(c.Uint("attention.key_length")),
 			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),
 			attnLogitSoftcap:  c.Float("attn_logit_softcapping"),
 			finalLogitSoftcap: c.Float("final_logit_softcapping"),
+			ropeConfig: ml.RoPEConfig{
+				Base:       c.Float("rope.freq_base", 10000.0),
+				Scale:      c.Float("rope.freq_scale", 1.0),
+				Dim:        c.Uint("attention.key_length"),
+				Type:       ml.RopeTypeNormal,
+				YarnConfig: ml.DefaultYarnConfig(int32(c.Uint("context_length", 131072))),
+			},
 		},
 	}
 
@@ -78,11 +84,10 @@ type SelfAttention struct {
 
 func (sa *SelfAttention) Forward(ctx ml.Context, hiddenState, positionIDs ml.Tensor, cache kvcache.Cache, opts *Options) ml.Tensor {
 	batchSize := hiddenState.Dim(1)
-	ropeType := uint32(2)
 
 	q := sa.Query.Forward(ctx, hiddenState)
 	q = q.Reshape(ctx, opts.attnKeyLen, opts.numHeads, batchSize)
-	q = q.RoPE(ctx, positionIDs, nil, uint32(opts.attnKeyLen), ropeType, opts.ropeBase, opts.ropeScale)
+	q = q.RoPE(ctx, positionIDs, nil, opts.ropeConfig)
 
 	if opts.largeModelScaling {
 		q = q.Scale(ctx, 1.0/math.Sqrt(float64(opts.hiddenSize/opts.numHeads)))
@@ -92,7 +97,7 @@ 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 = k.RoPE(ctx, positionIDs, nil, uint32(opts.attnKeyLen), ropeType, opts.ropeBase, opts.ropeScale)
+	k = k.RoPE(ctx, positionIDs, nil, opts.ropeConfig)
 
 	v := sa.Value.Forward(ctx, hiddenState)
 	v = v.Reshape(ctx, opts.attnValLen, opts.numKVHeads, batchSize)
@@ -122,7 +127,7 @@ func (sa *SelfAttention) Forward(ctx ml.Context, hiddenState, positionIDs ml.Ten
 }
 
 func (m *Model) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) {
-	return key.RoPE(ctx, shift, nil, uint32(m.Options.attnKeyLen), uint32(2), m.Options.ropeBase, m.Options.ropeScale), nil
+	return key.RoPE(ctx, shift, nil, m.ropeConfig), nil
 }
 
 type MLP struct {

model/models/gemma3/model_text.go

@@ -13,9 +13,11 @@ import (
 type TextOptions struct {
 	hiddenSize, numHeads, numKVHeads int
 	attnKeyLen, attnValLen           int
-	eps, ropeScale                   float32
-	ropeLocalBase, ropeGlobalBase    float32
+	eps                              float32
 	largeModelScaling                bool
+
+	ropeLocalConfig  ml.RoPEConfig
+	ropeGlobalConfig ml.RoPEConfig
 }
 
 type TextModel struct {
@@ -56,15 +58,27 @@ func newTextModel(c ml.Config) *TextModel {
 		),
 		Layers: make([]TextLayer, numBlocks),
 		TextOptions: &TextOptions{
-			hiddenSize:     int(c.Uint("embedding_length")),
-			numHeads:       int(c.Uint("attention.head_count")),
-			numKVHeads:     int(c.Uint("attention.head_count_kv")),
-			attnKeyLen:     int(c.Uint("attention.key_length", 256)),
-			attnValLen:     int(c.Uint("attention.value_length", 256)),
-			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),
+			hiddenSize: int(c.Uint("embedding_length")),
+			numHeads:   int(c.Uint("attention.head_count")),
+			numKVHeads: int(c.Uint("attention.head_count_kv")),
+			attnKeyLen: int(c.Uint("attention.key_length", 256)),
+			attnValLen: int(c.Uint("attention.value_length", 256)),
+			eps:        c.Float("attention.layer_norm_rms_epsilon", 1e-06),
+
+			ropeLocalConfig: ml.RoPEConfig{
+				Base:       c.Float("rope.local.freq_base", 10000.0),
+				Scale:      c.Float("rope.freq_scale", 1.0),
+				Dim:        c.Uint("attention.key_length", 256),
+				Type:       ml.RopeTypeNeox,
+				YarnConfig: ml.DefaultYarnConfig(int32(c.Uint("context_length", 131072))),
+			},
+			ropeGlobalConfig: ml.RoPEConfig{
+				Base:       c.Float("rope.global.freq_base", 1000000.0),
+				Scale:      c.Float("rope.freq_scale", 1.0),
+				Dim:        c.Uint("attention.key_length", 256),
+				Type:       ml.RopeTypeNeox,
+				YarnConfig: ml.DefaultYarnConfig(int32(c.Uint("context_length", 131072))),
+			},
 		},
 	}
 
@@ -86,17 +100,16 @@ type TextSelfAttention struct {
 
 func (sa *TextSelfAttention) Forward(ctx ml.Context, layer int, hiddenState, positionIDs ml.Tensor, cache kvcache.Cache, opts *TextOptions) ml.Tensor {
 	batchSize := hiddenState.Dim(1)
-	ropeType := uint32(2)
 
-	ropeBase := opts.ropeLocalBase
+	ropeConfig := opts.ropeLocalConfig
 	if (layer+1)%gemmaGlobalCacheCount == 0 {
-		ropeBase = opts.ropeGlobalBase
+		ropeConfig = opts.ropeGlobalConfig
 	}
 
 	q := sa.Query.Forward(ctx, hiddenState)
 	q = q.Reshape(ctx, opts.attnKeyLen, opts.numHeads, batchSize)
 	q = sa.QueryNorm.Forward(ctx, q, opts.eps)
-	q = q.RoPE(ctx, positionIDs, nil, uint32(opts.attnKeyLen), ropeType, ropeBase, opts.ropeScale)
+	q = q.RoPE(ctx, positionIDs, nil, ropeConfig)
 
 	if opts.largeModelScaling {
 		q = q.Scale(ctx, 1.0/math.Sqrt(float64(opts.hiddenSize/opts.numHeads)))
@@ -107,7 +120,7 @@ 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 = k.RoPE(ctx, positionIDs, nil, uint32(opts.attnKeyLen), ropeType, ropeBase, opts.ropeScale)
+	k = k.RoPE(ctx, positionIDs, nil, ropeConfig)
 
 	v := sa.Value.Forward(ctx, hiddenState)
 	v = v.Reshape(ctx, opts.attnValLen, opts.numKVHeads, batchSize)
@@ -120,12 +133,12 @@ func (sa *TextSelfAttention) Forward(ctx ml.Context, layer int, hiddenState, pos
 }
 
 func (m *TextModel) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) {
-	ropeBase := m.TextOptions.ropeLocalBase
+	ropeConfig := m.ropeLocalConfig
 	if (layer+1)%gemmaGlobalCacheCount == 0 {
-		ropeBase = m.TextOptions.ropeGlobalBase
+		ropeConfig = m.ropeGlobalConfig
 	}
 
-	return key.RoPE(ctx, shift, nil, uint32(m.TextOptions.attnKeyLen), uint32(2), ropeBase, m.TextOptions.ropeScale), nil
+	return key.RoPE(ctx, shift, nil, ropeConfig), nil
 }
 
 type TextMLP struct {

model/models/llama/model.go

@@ -14,8 +14,8 @@ import (
 
 type Options struct {
 	hiddenSize, numHeads, numKVHeads int
-	eps, ropeBase, ropeScale         float32
-	ropeDim                          uint32
+	eps                              float32
+	ropeConfig                       ml.RoPEConfig
 }
 
 type Model struct {
@@ -54,9 +54,13 @@ func New(c ml.Config) (model.Model, error) {
 			numHeads:   int(c.Uint("attention.head_count")),
 			numKVHeads: int(c.Uint("attention.head_count_kv")),
 			eps:        c.Float("attention.layer_norm_rms_epsilon"),
-			ropeBase:   c.Float("rope.freq_base"),
-			ropeScale:  c.Float("rope.freq_scale", 1),
-			ropeDim:    c.Uint("rope.dimension_count"),
+			ropeConfig: ml.RoPEConfig{
+				Base:       c.Float("rope.freq_base"),
+				Scale:      c.Float("rope.freq_scale", 1),
+				Dim:        c.Uint("rope.dimension_count"),
+				Type:       ml.RopeTypeNormal,
+				YarnConfig: ml.DefaultYarnConfig(int32(c.Uint("context_length", 131072))),
+			},
 		},
 	}
 
@@ -76,15 +80,14 @@ type SelfAttention struct {
 func (sa *SelfAttention) Forward(ctx ml.Context, hiddenState, positionIDs ml.Tensor, cache kvcache.Cache, opts *Options) ml.Tensor {
 	batchSize := hiddenState.Dim(1)
 	headDim := opts.hiddenSize / opts.numHeads
-	ropeType := uint32(0)
 
 	q := sa.Query.Forward(ctx, hiddenState)
 	q = q.Reshape(ctx, headDim, opts.numHeads, batchSize)
-	q = q.RoPE(ctx, positionIDs, sa.RopeFactors, opts.ropeDim, ropeType, opts.ropeBase, opts.ropeScale)
+	q = q.RoPE(ctx, positionIDs, sa.RopeFactors, opts.ropeConfig)
 
 	k := sa.Key.Forward(ctx, hiddenState)
 	k = k.Reshape(ctx, headDim, opts.numKVHeads, batchSize)
-	k = k.RoPE(ctx, positionIDs, sa.RopeFactors, opts.ropeDim, ropeType, opts.ropeBase, opts.ropeScale)
+	k = k.RoPE(ctx, positionIDs, sa.RopeFactors, opts.ropeConfig)
 
 	v := sa.Value.Forward(ctx, hiddenState)
 	v = v.Reshape(ctx, headDim, opts.numKVHeads, batchSize)
@@ -97,7 +100,7 @@ func (sa *SelfAttention) Forward(ctx ml.Context, hiddenState, positionIDs ml.Ten
 }
 
 func (m *Model) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) {
-	return key.RoPE(ctx, shift, m.Layers[layer].SelfAttention.RopeFactors, uint32(0), m.ropeDim, m.ropeBase, m.ropeScale), nil
+	return key.RoPE(ctx, shift, m.Layers[layer].SelfAttention.RopeFactors, m.ropeConfig), nil
 }
 
 type MLP struct {

model/models/mllama/model_text.go

@@ -20,15 +20,14 @@ type TextSelfAttention struct {
 func (sa *TextSelfAttention) Forward(ctx ml.Context, hiddenState, positions, _ ml.Tensor, cache *kvcache.WrapperCache, opts *TextModelOptions) ml.Tensor {
 	batchSize := hiddenState.Dim(1)
 	headDim := opts.hiddenSize / opts.numHeads
-	ropeType := uint32(0)
 
 	query := sa.Query.Forward(ctx, hiddenState)
 	query = query.Reshape(ctx, headDim, opts.numHeads, batchSize)
-	query = query.RoPE(ctx, positions, sa.RopeFactors, opts.ropeDim, ropeType, opts.ropeBase, opts.ropeScale)
+	query = query.RoPE(ctx, positions, sa.RopeFactors, opts.ropeConfig)
 
 	key := sa.Key.Forward(ctx, hiddenState)
 	key = key.Reshape(ctx, headDim, opts.numKVHeads, batchSize)
-	key = key.RoPE(ctx, positions, sa.RopeFactors, opts.ropeDim, ropeType, opts.ropeBase, opts.ropeScale)
+	key = key.RoPE(ctx, positions, sa.RopeFactors, opts.ropeConfig)
 
 	value := sa.Value.Forward(ctx, hiddenState)
 	value = value.Reshape(ctx, headDim, opts.numKVHeads, batchSize)
@@ -43,7 +42,7 @@ func (sa *TextSelfAttention) Forward(ctx ml.Context, hiddenState, positions, _ m
 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 key.RoPE(ctx, shift, sa.SelfAttention.RopeFactors, m.ropeDim, uint32(0), m.ropeBase, m.ropeScale), nil
+		return key.RoPE(ctx, shift, sa.SelfAttention.RopeFactors, m.ropeConfig), nil
 	}
 
 	return key, nil
@@ -198,8 +197,8 @@ func (d *TextDecoder) Forward(ctx ml.Context, hiddenState, positionIDs, outputs,
 
 type TextModelOptions struct {
 	hiddenSize, numHeads, numKVHeads int
-	eps, ropeBase, ropeScale         float32
-	ropeDim                          uint32
+	eps                              float32
+	ropeConfig                       ml.RoPEConfig
 
 	crossAttentionLayers []uint32
 }
@@ -240,10 +239,14 @@ func newTextModel(c ml.Config) *TextModel {
 			numHeads:             int(c.Uint("attention.head_count")),
 			numKVHeads:           int(c.Uint("attention.head_count_kv")),
 			eps:                  c.Float("attention.layer_norm_rms_epsilon"),
-			ropeBase:             c.Float("rope.freq_base"),
-			ropeScale:            c.Float("rope.freq_scale", 1),
-			ropeDim:              c.Uint("rope.dimension_count"),
 			crossAttentionLayers: c.Uints("attention.cross_attention_layers"),
+			ropeConfig: ml.RoPEConfig{
+				Base:       c.Float("rope.freq_base"),
+				Scale:      c.Float("rope.freq_scale", 1),
+				Dim:        c.Uint("rope.dimension_count"),
+				Type:       ml.RopeTypeNormal,
+				YarnConfig: ml.DefaultYarnConfig(int32(c.Uint("context_length", 131072))),
+			},
 		},
 	}
 }

runner/llamarunner/cache.go

@@ -213,8 +213,16 @@ func (c *InputCache) ShiftDiscard(inputLen int, numKeep int) int {
 	return discard
 }
 
-// Frees up space in the KV cache by deleting the oldest half of history and shifting
-// the newest half into that space (saving numKeep inputs at the beginning).
+type ErrReprocessInputs struct {
+	Inputs []input
+}
+
+func (e *ErrReprocessInputs) Error() string {
+	return fmt.Sprintf("kv cache shift not supported, inputs need reprocessing (input count: %v)", len(e.Inputs))
+}
+
+// ShiftCacheSlot frees up space in the KV cache by deleting the oldest half of history
+// and shifting the newest half into that space (saving numKeep inputs at the beginning).
 //
 // Assumes that at least 1 entry can be freed up by shifting (i.e. numKeep < numCtx)
 func (c *InputCache) ShiftCacheSlot(slot *InputCacheSlot, numKeep int) error {
@@ -222,7 +230,8 @@ func (c *InputCache) ShiftCacheSlot(slot *InputCacheSlot, numKeep int) error {
 		return fmt.Errorf("unable to shift context - keep exceeds context (keep: %v context: %v)", numKeep, c.numCtx)
 	}
 
-	discard := c.ShiftDiscard(len(slot.Inputs), numKeep)
+	inputLen := len(slot.Inputs)
+	discard := c.ShiftDiscard(inputLen, numKeep)
 
 	if discard <= 0 {
 		return nil
@@ -231,16 +240,42 @@ func (c *InputCache) ShiftCacheSlot(slot *InputCacheSlot, numKeep int) error {
 	slog.Debug("context limit hit - shifting", "id", slot.Id, "limit", c.numCtx, "input", len(slot.Inputs),
 		"keep", numKeep, "discard", discard)
 
-	// TODO (jessegross): KV cache removal can fail for certain types of models
-	if !c.lc.KvCacheSeqRm(slot.Id, numKeep, numKeep+discard) {
-		return fmt.Errorf("unable to remove old kv cache entries (id: %v, keep: %v discard: %v)", slot.Id, numKeep, discard)
-	}
-	c.lc.KvCacheSeqAdd(slot.Id, numKeep+discard, len(slot.Inputs), -discard)
+	var shiftFailed bool
 
-	for i := numKeep + discard; i < len(slot.Inputs); i++ {
+	if c.lc.KvCacheCanShift() {
+		// For models that support shifting, attempt to shift the KV cache
+		if !c.lc.KvCacheSeqRm(slot.Id, numKeep, numKeep+discard) {
+			shiftFailed = true
+			slog.Debug("kv cache removal not supported, clearing cache and returning inputs for reprocessing", "id", slot.Id)
+		} else {
+			c.lc.KvCacheSeqAdd(slot.Id, numKeep+discard, inputLen, -discard)
+		}
+	} else {
+		// For models that don't support shifting
+		shiftFailed = true
+		slog.Debug("kv cache cannot shift, clearing cache and returning inputs for reprocessing", "id", slot.Id)
+	}
+
+	if shiftFailed {
+		// Create new input slice with preserved tokens (numKeep + remaining tokens after discard)
+		newInputs := make([]input, numKeep+inputLen-(numKeep+discard))
+		copy(newInputs[:numKeep], slot.Inputs[:numKeep])
+		copy(newInputs[numKeep:], slot.Inputs[numKeep+discard:])
+
+		// Clear the entire KV cache
+		_ = c.lc.KvCacheSeqRm(slot.Id, 0, -1)
+		// Reset the slot inputs since we've cleared the cache
+		slot.Inputs = []input{}
+
+		// Return error with inputs that need to be reprocessed
+		return &ErrReprocessInputs{Inputs: newInputs}
+	}
+
+	// Standard shift succeeded - update input array
+	for i := numKeep + discard; i < inputLen; i++ {
 		slot.Inputs[i-discard] = slot.Inputs[i]
 	}
-	slot.Inputs = slot.Inputs[:len(slot.Inputs)-discard]
+	slot.Inputs = slot.Inputs[:inputLen-discard]
 
 	return nil
 }

runner/llamarunner/runner.go

@@ -389,7 +389,15 @@ func (s *Server) processBatch(tokenBatch *llama.Batch, embedBatch *llama.Batch)
 				if len(seq.pendingInputs) == 0 {
 					err := s.cache.ShiftCacheSlot(seq.cache, seq.numKeep)
 					if err != nil {
-						return err
+						var reprocess *ErrReprocessInputs
+						if errors.As(err, &reprocess) {
+							// Prepend these inputs to the sequence's inputs queue for reprocessing
+							seq.inputs = append(reprocess.Inputs, seq.inputs...)
+							// Continue processing as normal
+							continue
+						} else {
+							return err
+						}
 					}
 				} else {
 					break
@@ -599,7 +607,7 @@ func (s *Server) completion(w http.ResponseWriter, r *http.Request) {
 		if errors.Is(err, context.Canceled) {
 			slog.Info("aborting completion request due to client closing the connection")
 		} else {
-			slog.Error("Failed to acquire semaphore", "error", err)
+			http.Error(w, fmt.Sprintf("Failed to acquire semaphore: %v", err), http.StatusInternalServerError)
 		}
 		return
 	}
@@ -611,6 +619,7 @@ func (s *Server) completion(w http.ResponseWriter, r *http.Request) {
 			seq.cache, seq.inputs, err = s.cache.LoadCacheSlot(seq.inputs, true)
 			if err != nil {
 				s.mu.Unlock()
+				s.seqsSem.Release(1)
 				http.Error(w, fmt.Sprintf("Failed to load cache: %v", err), http.StatusInternalServerError)
 				return
 			}
@@ -626,6 +635,7 @@ func (s *Server) completion(w http.ResponseWriter, r *http.Request) {
 	s.mu.Unlock()
 
 	if !found {
+		s.seqsSem.Release(1)
 		http.Error(w, "could not find an available sequence", http.StatusInternalServerError)
 		return
 	}
@@ -691,7 +701,7 @@ func (s *Server) embeddings(w http.ResponseWriter, r *http.Request) {
 		if errors.Is(err, context.Canceled) {
 			slog.Info("aborting embeddings request due to client closing the connection")
 		} else {
-			slog.Error("Failed to acquire semaphore", "error", err)
+			http.Error(w, fmt.Sprintf("Failed to acquire semaphore: %v", err), http.StatusInternalServerError)
 		}
 		return
 	}
@@ -703,6 +713,7 @@ func (s *Server) embeddings(w http.ResponseWriter, r *http.Request) {
 			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
 			}
@@ -715,6 +726,7 @@ func (s *Server) embeddings(w http.ResponseWriter, r *http.Request) {
 	s.mu.Unlock()
 
 	if !found {
+		s.seqsSem.Release(1)
 		http.Error(w, "could not find an available sequence", http.StatusInternalServerError)
 		return
 	}

runner/ollamarunner/cache.go

@@ -239,6 +239,14 @@ func (c *InputCache) ShiftDiscard(inputLen int32, numKeep int32) int32 {
 	return discard
 }
 
+type ErrReprocessInputs struct {
+	Inputs []input.Input
+}
+
+func (e *ErrReprocessInputs) Error() string {
+	return fmt.Sprintf("kv cache shift not supported, inputs need reprocessing (input count: %v)", len(e.Inputs))
+}
+
 // Frees up space in the KV cache by deleting the oldest half of history and shifting
 // the newest half into that space (saving numKeep inputs at the beginning).
 //
@@ -258,11 +266,23 @@ func (c *InputCache) ShiftCacheSlot(slot *InputCacheSlot, numKeep int32) error {
 	slog.Debug("context limit hit - shifting", "id", slot.Id, "limit", c.numCtx, "input", len(slot.Inputs),
 		"keep", numKeep, "discard", discard)
 
-	// TODO (jessegross): KV cache removal can fail for certain types of models
 	if c.cache != nil {
 		err := c.cache.Remove(slot.Id, numKeep, numKeep+discard)
 		if err != nil {
-			return fmt.Errorf("unable to remove old kv cache entries (id: %v, keep: %v discard: %v): %w", slot.Id, numKeep, discard, err)
+			slog.Debug("kv cache removal unsupported, clearing cache and returning inputs for reprocessing",
+				"id", slot.Id, "error", err)
+
+			// Create new input slice with preserved tokens (numKeep + remaining tokens after discard)
+			newInputs := make([]input.Input, numKeep+inputLen-(numKeep+discard))
+			copy(newInputs[:numKeep], slot.Inputs[:numKeep])
+			copy(newInputs[numKeep:], slot.Inputs[numKeep+discard:])
+
+			// Reset the cache
+			_ = c.cache.Remove(slot.Id, 0, -1)
+			slot.Inputs = []input.Input{}
+
+			// Return error with inputs that need to be reprocessed
+			return &ErrReprocessInputs{Inputs: newInputs}
 		}
 	}
 

runner/ollamarunner/cache_test.go

@@ -1,10 +1,13 @@
 package ollamarunner
 
 import (
+	"errors"
+	"fmt"
 	"image"
 	"testing"
 	"time"
 
+	"github.com/ollama/ollama/ml"
 	"github.com/ollama/ollama/model/input"
 )
 
@@ -425,3 +428,91 @@ func TestLoadCacheSlot(t *testing.T) {
 		})
 	}
 }
+
+// Mock implementation of the Cache interface
+type mockCache struct {
+	shouldFail bool
+}
+
+// Implement only the methods needed for the test
+func (m *mockCache) Remove(seq int, beginIndex, endIndex int32) error {
+	if m.shouldFail {
+		return fmt.Errorf("mock cache removal error")
+	}
+	return nil
+}
+
+// Stub implementations for other interface methods
+func (m *mockCache) SetLayer(layer int)                                                            {}
+func (m *mockCache) Get(ctx ml.Context) (ml.Tensor, ml.Tensor, ml.Tensor)                          { return nil, nil, nil }
+func (m *mockCache) Put(ctx ml.Context, key, value ml.Tensor)                                      {}
+func (m *mockCache) Init(backend ml.Backend, dtype ml.DType, maxSequences, capacity, maxBatch int) {}
+func (m *mockCache) Close()                                                                        {}
+func (m *mockCache) StartForward(ctx ml.Context, batch input.Batch) error                          { return nil }
+func (m *mockCache) CopyPrefix(srcSeq, dstSeq int, len int32)                                      {}
+func (m *mockCache) SetConfig(ml.CacheConfig)                                                      {}
+
+func TestShiftCacheSlot(t *testing.T) {
+	tests := []struct {
+		name          string
+		numCtx        int32
+		inputs        []input.Input
+		numKeep       int32
+		cacheErr      bool
+		wantErr       any
+		wantInputsLen int
+	}{
+		{
+			name:          "Normal shift",
+			numCtx:        10,
+			inputs:        []input.Input{{Token: 1}, {Token: 2}, {Token: 3}, {Token: 4}, {Token: 5}, {Token: 6}, {Token: 7}, {Token: 8}, {Token: 9}, {Token: 10}},
+			numKeep:       2,
+			cacheErr:      false, // No error
+			wantErr:       nil,
+			wantInputsLen: 6, // After discarding 4 tokens
+		},
+		{
+			name:          "Cache removal fails",
+			numCtx:        10,
+			inputs:        []input.Input{{Token: 1}, {Token: 2}, {Token: 3}, {Token: 4}, {Token: 5}, {Token: 6}, {Token: 7}, {Token: 8}, {Token: 9}, {Token: 10}},
+			numKeep:       2,
+			cacheErr:      true,
+			wantErr:       &ErrReprocessInputs{},
+			wantInputsLen: 0, // Original inputs should be cleared
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			mock := &mockCache{shouldFail: tt.cacheErr}
+			c := InputCache{
+				numCtx: tt.numCtx,
+				cache:  mock,
+			}
+			slot := &InputCacheSlot{
+				Id:     123,
+				Inputs: make([]input.Input, len(tt.inputs)),
+			}
+			copy(slot.Inputs, tt.inputs)
+
+			err := c.ShiftCacheSlot(slot, tt.numKeep)
+
+			if tt.wantErr != nil {
+				if err == nil {
+					t.Errorf("Expected error but got nil")
+					return
+				}
+
+				if !errors.As(err, &tt.wantErr) {
+					t.Errorf("Expected error of type %T but got %T: %v", tt.wantErr, err, err)
+				}
+			} else if err != nil {
+				t.Errorf("Unexpected error: %v", err)
+			}
+
+			if len(slot.Inputs) != tt.wantInputsLen {
+				t.Errorf("Slot inputs length after operation: got %v, want %v", len(slot.Inputs), tt.wantInputsLen)
+			}
+		})
+	}
+}

runner/ollamarunner/runner.go

@@ -267,6 +267,9 @@ type Server struct {
 	// KV cache
 	cache *InputCache
 
+	// next sequence for prompt processing to avoid starvation
+	nextSeq int
+
 	// multimodalHash generates hashes for comparing equality
 	// of non-text data
 	multimodalHash maphash.Hash
@@ -351,14 +354,19 @@ func (s *Server) processBatch() error {
 	var batchInputs []int32
 	var batch input.Batch
 
-	for i, seq := range s.seqs {
+	resumeSeq := -1
+	seqIdx := s.nextSeq - 1
+	for range s.seqs {
+		seqIdx = (seqIdx + 1) % len(s.seqs)
+		seq := s.seqs[seqIdx]
+
 		if seq == nil {
 			continue
 		}
 
 		// if past the num predict limit
 		if seq.numPredict > 0 && seq.numPredicted >= seq.numPredict {
-			s.removeSequence(i, "limit")
+			s.removeSequence(seqIdx, "limit")
 			continue
 		}
 
@@ -369,16 +377,23 @@ func (s *Server) processBatch() error {
 
 		batchSize := s.batchSize
 
-		for j, inp := range seq.inputs {
+		for i, inp := range seq.inputs {
 			// If we are required to put following inputs into a single batch then extend the
 			// batch size. Since we are only extending the size the minimum amount possible, this
-			// will cause a break if we have pending inputs.
+			// will cause a break if we have existing inputs.
 			minBatch := 1 + inp.SameBatch
 			if minBatch > batchSize {
 				batchSize = minBatch
 			}
 
-			if len(seq.pendingInputs)+minBatch > batchSize {
+			// Stop if the required batch would put us over the total batch size (including tokens
+			// added by other sequences). If we haven't been able to add anything yet then pick up
+			// here again for the next batch to avoid starvation, though we can opportunistically
+			// check if other sequences can still squeeze something in.
+			if len(batchInputs)+minBatch > batchSize {
+				if len(seq.pendingInputs) == 0 && resumeSeq == -1 {
+					resumeSeq = seqIdx
+				}
 				break
 			}
 
@@ -392,7 +407,15 @@ func (s *Server) processBatch() error {
 
 				err := s.cache.ShiftCacheSlot(seq.cache, seq.numKeep)
 				if err != nil {
-					return err
+					var reprocess *ErrReprocessInputs
+					if errors.As(err, &reprocess) {
+						// Prepend these inputs to the sequence's inputs queue for reprocessing
+						seq.inputs = append(reprocess.Inputs, seq.inputs...)
+						// Skip this sequence but continue processing the rest
+						continue
+					} else {
+						return err
+					}
 				}
 			}
 
@@ -405,7 +428,7 @@ func (s *Server) processBatch() error {
 			batch.Sequences = append(batch.Sequences, seq.cache.Id)
 
 			seq.iBatch = len(batch.Outputs)
-			if j+1 == len(seq.inputs) {
+			if i+1 == len(seq.inputs) {
 				batch.Outputs = append(batch.Outputs, int32(len(batchInputs)-1))
 			}
 			seq.pendingInputs = append(seq.pendingInputs, inp)
@@ -414,6 +437,12 @@ func (s *Server) processBatch() error {
 		seq.inputs = seq.inputs[len(seq.pendingInputs):]
 	}
 
+	if resumeSeq != -1 {
+		s.nextSeq = resumeSeq
+	} else {
+		s.nextSeq = seqIdx + 1
+	}
+
 	if len(batchInputs) == 0 {
 		return nil
 	}
@@ -588,7 +617,7 @@ func (s *Server) completion(w http.ResponseWriter, r *http.Request) {
 		if errors.Is(err, context.Canceled) {
 			slog.Info("aborting completion request due to client closing the connection")
 		} else {
-			slog.Error("Failed to acquire semaphore", "error", err)
+			http.Error(w, fmt.Sprintf("Failed to acquire semaphore: %v", err), http.StatusInternalServerError)
 		}
 		return
 	}
@@ -600,6 +629,7 @@ func (s *Server) completion(w http.ResponseWriter, r *http.Request) {
 			seq.cache, seq.inputs, err = s.cache.LoadCacheSlot(seq.inputs)
 			if err != nil {
 				s.mu.Unlock()
+				s.seqsSem.Release(1)
 				http.Error(w, fmt.Sprintf("Failed to load cache: %v", err), http.StatusInternalServerError)
 				return
 			}
@@ -613,6 +643,7 @@ func (s *Server) completion(w http.ResponseWriter, r *http.Request) {
 	s.mu.Unlock()
 
 	if !found {
+		s.seqsSem.Release(1)
 		http.Error(w, "could not find an available sequence", http.StatusInternalServerError)
 		return
 	}

server/download.go

@@ -29,8 +29,9 @@ import (
 const maxRetries = 6
 
 var (
-	errMaxRetriesExceeded = errors.New("max retries exceeded")
-	errPartStalled        = errors.New("part stalled")
+	errMaxRetriesExceeded   = errors.New("max retries exceeded")
+	errPartStalled          = errors.New("part stalled")
+	errMaxRedirectsExceeded = errors.New("maximum redirects exceeded (10) for directURL")
 )
 
 var blobDownloadManager sync.Map
@@ -236,7 +237,7 @@ func (b *blobDownload) run(ctx context.Context, requestURL *url.URL, opts *regis
 
 			newOpts.CheckRedirect = func(req *http.Request, via []*http.Request) error {
 				if len(via) > 10 {
-					return errors.New("maximum redirects exceeded (10) for directURL")
+					return errMaxRedirectsExceeded
 				}
 
 				// if the hostname is the same, allow the redirect

server/images.go

@@ -35,6 +35,7 @@ var (
 	errCapabilityCompletion = errors.New("completion")
 	errCapabilityTools      = errors.New("tools")
 	errCapabilityInsert     = errors.New("insert")
+	errInsecureProtocol     = errors.New("insecure protocol http")
 )
 
 type Capability string
@@ -479,7 +480,7 @@ func PushModel(ctx context.Context, name string, regOpts *registryOptions, fn fu
 	fn(api.ProgressResponse{Status: "retrieving manifest"})
 
 	if mp.ProtocolScheme == "http" && !regOpts.Insecure {
-		return errors.New("insecure protocol http")
+		return errInsecureProtocol
 	}
 
 	manifest, _, err := GetManifest(mp)
@@ -543,7 +544,7 @@ func PullModel(ctx context.Context, name string, regOpts *registryOptions, fn fu
 	}
 
 	if mp.ProtocolScheme == "http" && !regOpts.Insecure {
-		return errors.New("insecure protocol http")
+		return errInsecureProtocol
 	}
 
 	fn(api.ProgressResponse{Status: "pulling manifest"})

server/internal/client/ollama/registry.go

@@ -421,14 +421,6 @@ func (r *Registry) Push(ctx context.Context, name string, p *PushParams) error {
 	return err
 }
 
-func canRetry(err error) bool {
-	var re *Error
-	if !errors.As(err, &re) {
-		return false
-	}
-	return re.Status >= 500
-}
-
 // trackingReader is an io.Reader that tracks the number of bytes read and
 // calls the update function with the layer, the number of bytes read.
 //
@@ -514,13 +506,40 @@ func (r *Registry) Pull(ctx context.Context, name string) error {
 				break
 			}
 
+			cacheKey := fmt.Sprintf(
+				"v1 pull chunksum %s %s %d-%d",
+				l.Digest,
+				cs.Digest,
+				cs.Chunk.Start,
+				cs.Chunk.End,
+			)
+			cacheKeyDigest := blob.DigestFromBytes(cacheKey)
+			_, err := c.Get(cacheKeyDigest)
+			if err == nil {
+				received.Add(cs.Chunk.Size())
+				t.update(l, cs.Chunk.Size(), ErrCached)
+				continue
+			}
+
 			wg.Add(1)
 			g.Go(func() (err error) {
 				defer func() {
 					if err == nil {
+						// Ignore cache key write errors for now. We've already
+						// reported to trace that the chunk is complete.
+						//
+						// Ideally, we should only report completion to trace
+						// after successful cache commit. This current approach
+						// works but could trigger unnecessary redownloads if
+						// the checkpoint key is missing on next pull.
+						//
+						// Not incorrect, just suboptimal - fix this in a
+						// future update.
+						_ = blob.PutBytes(c, cacheKeyDigest, cacheKey)
+
 						received.Add(cs.Chunk.Size())
 					} else {
-						err = fmt.Errorf("error downloading %s: %w", cs.Digest.Short(), err)
+						t.update(l, 0, err)
 					}
 					wg.Done()
 				}()
@@ -563,7 +582,7 @@ func (r *Registry) Pull(ctx context.Context, name string) error {
 		return err
 	}
 	if received.Load() != expected {
-		return fmt.Errorf("%w: received %d/%d", ErrIncomplete, received.Load(), expected)
+		return fmt.Errorf("%w: received %d/%d bytes", ErrIncomplete, received.Load(), expected)
 	}
 
 	md := blob.DigestFromBytes(m.Data)
@@ -608,6 +627,30 @@ func (m *Manifest) Layer(d blob.Digest) *Layer {
 	return nil
 }
 
+func (m *Manifest) All() iter.Seq[*Layer] {
+	return func(yield func(*Layer) bool) {
+		if !yield(m.Config) {
+			return
+		}
+		for _, l := range m.Layers {
+			if !yield(l) {
+				return
+			}
+		}
+	}
+}
+
+func (m *Manifest) Size() int64 {
+	var size int64
+	if m.Config != nil {
+		size += m.Config.Size
+	}
+	for _, l := range m.Layers {
+		size += l.Size
+	}
+	return size
+}
+
 // MarshalJSON implements json.Marshaler.
 //
 // NOTE: It adds an empty config object to the manifest, which is required by
@@ -750,20 +793,32 @@ func (r *Registry) chunksums(ctx context.Context, name string, l *Layer) iter.Se
 			return
 		}
 
-		// A chunksums response is a sequence of chunksums in a
-		// simple, easy to parse line-oriented format.
+		// The response is a sequence of chunksums.
 		//
-		// Example:
+		// Chunksums are chunks of a larger blob that can be
+		// downloaded and verified independently.
 		//
-		//     >> GET /v2/<namespace>/<model>/chunksums/<digest>
+		// The chunksums endpoint is a GET request that returns a
+		// sequence of chunksums in the following format:
 		//
-		//     << HTTP/1.1 200 OK
-		//     << Content-Location: <blobURL>
-		//     <<
-		//     << <digest> <start>-<end>
-		//     << ...
+		//     > GET /v2/<namespace>/<model>/chunksums/<digest>
 		//
-		// The blobURL is the URL to download the chunks from.
+		//     < HTTP/1.1 200 OK
+		//     < Content-Location: <blobURL>
+		//     <
+		//     < <digest> <start>-<end>
+		//     < ...
+		//
+		// The <blobURL> is the URL to download the chunks from and
+		// each <digest> is the digest of the chunk, and <start>-<end>
+		// is the range the chunk in the blob.
+		//
+		// Ranges may be used directly in Range headers like
+		// "bytes=<start>-<end>".
+		//
+		// The chunksums returned are guaranteed to be contiguous and
+		// include all bytes of the layer. If the stream is cut short,
+		// clients should retry.
 
 		chunksumsURL := fmt.Sprintf("%s://%s/v2/%s/%s/chunksums/%s",
 			scheme,

server/internal/client/ollama/registry_test.go

@@ -9,17 +9,14 @@ import (
 	"fmt"
 	"io"
 	"io/fs"
-	"math/rand/v2"
+	"net"
 	"net/http"
 	"net/http/httptest"
 	"os"
-	"path"
 	"reflect"
-	"slices"
 	"strings"
-	"sync"
+	"sync/atomic"
 	"testing"
-	"time"
 
 	"github.com/ollama/ollama/server/internal/cache/blob"
 	"github.com/ollama/ollama/server/internal/testutil"
@@ -338,15 +335,8 @@ func TestPushCommitRoundtripError(t *testing.T) {
 	}
 }
 
-func checkNotExist(t *testing.T, err error) {
-	t.Helper()
-	if !errors.Is(err, fs.ErrNotExist) {
-		t.Fatalf("err = %v; want fs.ErrNotExist", err)
-	}
-}
-
 func TestRegistryPullInvalidName(t *testing.T) {
-	rc, _ := newClient(t, nil)
+	rc, _ := newRegistryClient(t, nil)
 	err := rc.Pull(t.Context(), "://")
 	if !errors.Is(err, ErrNameInvalid) {
 		t.Errorf("err = %v; want %v", err, ErrNameInvalid)
@@ -362,197 +352,16 @@ func TestRegistryPullInvalidManifest(t *testing.T) {
 	}
 
 	for _, resp := range cases {
-		rc, _ := newClient(t, func(w http.ResponseWriter, r *http.Request) {
+		rc, _ := newRegistryClient(t, func(w http.ResponseWriter, r *http.Request) {
 			io.WriteString(w, resp)
 		})
-		err := rc.Pull(t.Context(), "x")
+		err := rc.Pull(t.Context(), "http://example.com/a/b")
 		if !errors.Is(err, ErrManifestInvalid) {
 			t.Errorf("err = %v; want invalid manifest", err)
 		}
 	}
 }
 
-func TestRegistryPullNotCached(t *testing.T) {
-	check := testutil.Checker(t)
-
-	var c *blob.DiskCache
-	var rc *Registry
-
-	d := blob.DigestFromBytes("some data")
-	rc, c = newClient(t, func(w http.ResponseWriter, r *http.Request) {
-		if strings.Contains(r.URL.Path, "/blobs/") {
-			io.WriteString(w, "some data")
-			return
-		}
-		fmt.Fprintf(w, `{"layers":[{"digest":%q,"size":9}]}`, d)
-	})
-
-	// Confirm that the layer does not exist locally
-	_, err := rc.ResolveLocal("model")
-	checkNotExist(t, err)
-
-	_, err = c.Get(d)
-	checkNotExist(t, err)
-
-	err = rc.Pull(t.Context(), "model")
-	check(err)
-
-	mw, err := rc.Resolve(t.Context(), "model")
-	check(err)
-	mg, err := rc.ResolveLocal("model")
-	check(err)
-	if !reflect.DeepEqual(mw, mg) {
-		t.Errorf("mw = %v; mg = %v", mw, mg)
-	}
-
-	// Confirm successful download
-	info, err := c.Get(d)
-	check(err)
-	if info.Digest != d {
-		t.Errorf("info.Digest = %v; want %v", info.Digest, d)
-	}
-	if info.Size != 9 {
-		t.Errorf("info.Size = %v; want %v", info.Size, 9)
-	}
-
-	data, err := os.ReadFile(c.GetFile(d))
-	check(err)
-	if string(data) != "some data" {
-		t.Errorf("data = %q; want %q", data, "exists")
-	}
-}
-
-func TestRegistryPullCached(t *testing.T) {
-	cached := blob.DigestFromBytes("exists")
-	rc, _ := newClient(t, func(w http.ResponseWriter, r *http.Request) {
-		if strings.Contains(r.URL.Path, "/blobs/") {
-			w.WriteHeader(499) // should not be called
-			return
-		}
-		if strings.Contains(r.URL.Path, "/manifests/") {
-			fmt.Fprintf(w, `{"layers":[{"digest":%q,"size":6}]}`, cached)
-		}
-	})
-
-	var errs []error
-	var reads []int64
-	ctx := WithTrace(t.Context(), &Trace{
-		Update: func(d *Layer, n int64, err error) {
-			t.Logf("update %v %d %v", d, n, err)
-			reads = append(reads, n)
-			errs = append(errs, err)
-		},
-	})
-
-	ctx, cancel := context.WithTimeout(ctx, 3*time.Second)
-	defer cancel()
-
-	err := rc.Pull(ctx, "single")
-	testutil.Check(t, err)
-
-	want := []int64{0, 6}
-	if !errors.Is(errors.Join(errs...), ErrCached) {
-		t.Errorf("errs = %v; want %v", errs, ErrCached)
-	}
-	if !slices.Equal(reads, want) {
-		t.Errorf("pairs = %v; want %v", reads, want)
-	}
-}
-
-func TestRegistryPullManifestNotFound(t *testing.T) {
-	rc, _ := newClient(t, func(w http.ResponseWriter, r *http.Request) {
-		w.WriteHeader(http.StatusNotFound)
-	})
-	err := rc.Pull(t.Context(), "notfound")
-	checkErrCode(t, err, 404, "")
-}
-
-func TestRegistryPullResolveRemoteError(t *testing.T) {
-	rc, _ := newClient(t, func(w http.ResponseWriter, r *http.Request) {
-		w.WriteHeader(http.StatusInternalServerError)
-		io.WriteString(w, `{"errors":[{"code":"an_error"}]}`)
-	})
-	err := rc.Pull(t.Context(), "single")
-	checkErrCode(t, err, 500, "an_error")
-}
-
-func TestRegistryPullResolveRoundtripError(t *testing.T) {
-	rc, _ := newClient(t, func(w http.ResponseWriter, r *http.Request) {
-		if strings.Contains(r.URL.Path, "/manifests/") {
-			w.WriteHeader(499) // force RoundTrip error
-			return
-		}
-	})
-	err := rc.Pull(t.Context(), "single")
-	if !errors.Is(err, errRoundTrip) {
-		t.Errorf("err = %v; want %v", err, errRoundTrip)
-	}
-}
-
-// TestRegistryPullMixedCachedNotCached tests that cached layers do not
-// interfere with pulling layers that are not cached
-func TestRegistryPullMixedCachedNotCached(t *testing.T) {
-	x := blob.DigestFromBytes("xxxxxx")
-	e := blob.DigestFromBytes("exists")
-	y := blob.DigestFromBytes("yyyyyy")
-
-	for i := range 10 {
-		t.Logf("iteration %d", i)
-
-		digests := []blob.Digest{x, e, y}
-
-		rand.Shuffle(len(digests), func(i, j int) {
-			digests[i], digests[j] = digests[j], digests[i]
-		})
-
-		manifest := fmt.Sprintf(`{
-			"layers": [
-				{"digest":"%s","size":6},
-				{"digest":"%s","size":6},
-				{"digest":"%s","size":6}
-			]
-		}`, digests[0], digests[1], digests[2])
-
-		rc, c := newClient(t, func(w http.ResponseWriter, r *http.Request) {
-			switch path.Base(r.URL.Path) {
-			case "latest":
-				io.WriteString(w, manifest)
-			case x.String():
-				io.WriteString(w, "xxxxxx")
-			case e.String():
-				io.WriteString(w, "exists")
-			case y.String():
-				io.WriteString(w, "yyyyyy")
-			default:
-				panic(fmt.Sprintf("unexpected request: %v", r))
-			}
-		})
-
-		ctx := WithTrace(t.Context(), &Trace{
-			Update: func(l *Layer, n int64, err error) {
-				t.Logf("update %v %d %v", l, n, err)
-			},
-		})
-
-		// Check that we pull all layers that we can.
-
-		err := rc.Pull(ctx, "mixed")
-		if err != nil {
-			t.Fatal(err)
-		}
-
-		for _, d := range digests {
-			info, err := c.Get(d)
-			if err != nil {
-				t.Fatalf("Get(%v): %v", d, err)
-			}
-			if info.Size != 6 {
-				t.Errorf("info.Size = %v; want %v", info.Size, 6)
-			}
-		}
-	}
-}
-
 func TestRegistryResolveByDigest(t *testing.T) {
 	check := testutil.Checker(t)
 
@@ -590,26 +399,6 @@ func TestInsecureSkipVerify(t *testing.T) {
 	testutil.Check(t, err)
 }
 
-func TestCanRetry(t *testing.T) {
-	cases := []struct {
-		err  error
-		want bool
-	}{
-		{nil, false},
-		{errors.New("x"), false},
-		{ErrCached, false},
-		{ErrManifestInvalid, false},
-		{ErrNameInvalid, false},
-		{&Error{Status: 100}, false},
-		{&Error{Status: 500}, true},
-	}
-	for _, tt := range cases {
-		if got := canRetry(tt.err); got != tt.want {
-			t.Errorf("CanRetry(%v) = %v; want %v", tt.err, got, tt.want)
-		}
-	}
-}
-
 func TestErrorUnmarshal(t *testing.T) {
 	cases := []struct {
 		name    string
@@ -761,17 +550,23 @@ func TestParseNameExtended(t *testing.T) {
 
 func TestUnlink(t *testing.T) {
 	t.Run("found by name", func(t *testing.T) {
-		rc, _ := newClient(t, nil)
+		check := testutil.Checker(t)
+
+		rc, _ := newRegistryClient(t, nil)
+		// make a blob and link it
+		d := blob.DigestFromBytes("{}")
+		err := blob.PutBytes(rc.Cache, d, "{}")
+		check(err)
+		err = rc.Cache.Link("registry.ollama.ai/library/single:latest", d)
+		check(err)
 
 		// confirm linked
-		_, err := rc.ResolveLocal("single")
-		if err != nil {
-			t.Errorf("unexpected error: %v", err)
-		}
+		_, err = rc.ResolveLocal("single")
+		check(err)
 
 		// unlink
 		_, err = rc.Unlink("single")
-		testutil.Check(t, err)
+		check(err)
 
 		// confirm unlinked
 		_, err = rc.ResolveLocal("single")
@@ -780,7 +575,7 @@ func TestUnlink(t *testing.T) {
 		}
 	})
 	t.Run("not found by name", func(t *testing.T) {
-		rc, _ := newClient(t, nil)
+		rc, _ := newRegistryClient(t, nil)
 		ok, err := rc.Unlink("manifestNotFound")
 		if err != nil {
 			t.Fatal(err)
@@ -791,78 +586,368 @@ func TestUnlink(t *testing.T) {
 	})
 }
 
-func TestPullChunksums(t *testing.T) {
-	check := testutil.Checker(t)
+// Many tests from here out, in this file are based on a single blob, "abc",
+// with the checksum of its sha256 hash. The checksum is:
+//
+//	"abc" -> sha256:ba7816bf8f01cfea414140de5dae2223b00361a396177a9cb410ff61f20015ad
+//
+// Using the literal value instead of a constant with fmt.Xprintf calls proved
+// to be the most readable and maintainable approach. The sum is consistently
+// used in the tests and unique so searches do not yield false positives.
 
-	content := "hello"
-	var chunksums string
-	contentDigest := func() blob.Digest {
-		return blob.DigestFromBytes(content)
+func checkRequest(t *testing.T, req *http.Request, method, path string) {
+	t.Helper()
+	if got := req.URL.Path; got != path {
+		t.Errorf("URL = %q, want %q", got, path)
 	}
-	rc, c := newClient(t, func(w http.ResponseWriter, r *http.Request) {
-		switch {
-		case strings.Contains(r.URL.Path, "/manifests/latest"):
-			fmt.Fprintf(w, `{"layers":[{"digest":%q,"size":%d}]}`, contentDigest(), len(content))
-		case strings.HasSuffix(r.URL.Path, "/chunksums/"+contentDigest().String()):
-			loc := fmt.Sprintf("http://blob.store/v2/library/test/blobs/%s", contentDigest())
-			w.Header().Set("Content-Location", loc)
-			io.WriteString(w, chunksums)
-		case strings.Contains(r.URL.Path, "/blobs/"+contentDigest().String()):
-			http.ServeContent(w, r, contentDigest().String(), time.Time{}, strings.NewReader(content))
-		default:
-			t.Errorf("unexpected request: %v", r)
-			http.NotFound(w, r)
-		}
-	})
+	if req.Method != method {
+		t.Errorf("Method = %q, want %q", req.Method, method)
+	}
+}
 
-	rc.MaxStreams = 1        // prevent concurrent chunk downloads
-	rc.ChunkingThreshold = 1 // for all blobs to be chunked
+func newRegistryClient(t *testing.T, h http.HandlerFunc) (*Registry, context.Context) {
+	s := httptest.NewServer(h)
+	t.Cleanup(s.Close)
+	cache, err := blob.Open(t.TempDir())
+	if err != nil {
+		t.Fatal(err)
+	}
 
-	var mu sync.Mutex
-	var reads []int64
 	ctx := WithTrace(t.Context(), &Trace{
 		Update: func(l *Layer, n int64, err error) {
-			t.Logf("Update: %v %d %v", l, n, err)
-			mu.Lock()
-			reads = append(reads, n)
-			mu.Unlock()
+			t.Log("trace:", l.Digest.Short(), n, err)
 		},
 	})
 
-	chunksums = fmt.Sprintf("%s 0-2\n%s 3-4\n",
-		blob.DigestFromBytes("hel"),
-		blob.DigestFromBytes("lo"),
-	)
-	err := rc.Pull(ctx, "test")
-	check(err)
-	wantReads := []int64{
-		0, // initial signaling of layer pull starting
-		3, // first chunk read
-		2, // second chunk read
-	}
-	if !slices.Equal(reads, wantReads) {
-		t.Errorf("reads = %v; want %v", reads, wantReads)
+	rc := &Registry{
+		Cache: cache,
+		HTTPClient: &http.Client{Transport: &http.Transport{
+			Dial: func(network, addr string) (net.Conn, error) {
+				return net.Dial(network, s.Listener.Addr().String())
+			},
+		}},
 	}
+	return rc, ctx
+}
 
-	mw, err := rc.Resolve(t.Context(), "test")
-	check(err)
-	mg, err := rc.ResolveLocal("test")
-	check(err)
-	if !reflect.DeepEqual(mw, mg) {
-		t.Errorf("mw = %v; mg = %v", mw, mg)
-	}
-	for i := range mg.Layers {
-		_, err = c.Get(mg.Layers[i].Digest)
-		if err != nil {
-			t.Errorf("Get(%v): %v", mg.Layers[i].Digest, err)
+func TestPullChunked(t *testing.T) {
+	var steps atomic.Int64
+	c, ctx := newRegistryClient(t, func(w http.ResponseWriter, r *http.Request) {
+		switch steps.Add(1) {
+		case 1:
+			checkRequest(t, r, "GET", "/v2/library/abc/manifests/latest")
+			io.WriteString(w, `{"layers":[{"size":3,"digest":"sha256:ba7816bf8f01cfea414140de5dae2223b00361a396177a9cb410ff61f20015ad"}]}`)
+		case 2:
+			checkRequest(t, r, "GET", "/v2/library/abc/chunksums/sha256:ba7816bf8f01cfea414140de5dae2223b00361a396177a9cb410ff61f20015ad")
+			w.Header().Set("Content-Location", "http://blob.store/v2/library/abc/blobs/sha256:ba7816bf8f01cfea414140de5dae2223b00361a396177a9cb410ff61f20015ad")
+			fmt.Fprintf(w, "%s 0-1\n", blob.DigestFromBytes("ab"))
+			fmt.Fprintf(w, "%s 2-2\n", blob.DigestFromBytes("c"))
+		case 3, 4:
+			checkRequest(t, r, "GET", "/v2/library/abc/blobs/sha256:ba7816bf8f01cfea414140de5dae2223b00361a396177a9cb410ff61f20015ad")
+			switch rng := r.Header.Get("Range"); rng {
+			case "bytes=0-1":
+				io.WriteString(w, "ab")
+			case "bytes=2-2":
+				t.Logf("writing c")
+				io.WriteString(w, "c")
+			default:
+				t.Errorf("unexpected range %q", rng)
+			}
+		default:
+			t.Errorf("unexpected steps %d: %v", steps.Load(), r)
+			http.Error(w, "unexpected steps", http.StatusInternalServerError)
 		}
-	}
+	})
 
-	// missing chunks
-	content = "llama"
-	chunksums = fmt.Sprintf("%s 0-1\n", blob.DigestFromBytes("ll"))
-	err = rc.Pull(ctx, "missingchunks")
-	if err == nil {
-		t.Error("expected error because of missing chunks")
+	c.ChunkingThreshold = 1 // force chunking
+
+	err := c.Pull(ctx, "http://o.com/library/abc")
+	testutil.Check(t, err)
+
+	_, err = c.Cache.Resolve("o.com/library/abc:latest")
+	testutil.Check(t, err)
+
+	if g := steps.Load(); g != 4 {
+		t.Fatalf("got %d steps, want 4", g)
+	}
+}
+
+func TestPullCached(t *testing.T) {
+	c, ctx := newRegistryClient(t, func(w http.ResponseWriter, r *http.Request) {
+		checkRequest(t, r, "GET", "/v2/library/abc/manifests/latest")
+		io.WriteString(w, `{"layers":[{"size":3,"digest":"sha256:ba7816bf8f01cfea414140de5dae2223b00361a396177a9cb410ff61f20015ad"}]}`)
+	})
+
+	check := testutil.Checker(t)
+
+	// Premeptively cache the blob
+	d, err := blob.ParseDigest("sha256:ba7816bf8f01cfea414140de5dae2223b00361a396177a9cb410ff61f20015ad")
+	check(err)
+	err = blob.PutBytes(c.Cache, d, []byte("abc"))
+	check(err)
+
+	// Pull only the manifest, which should be enough to resolve the cached blob
+	err = c.Pull(ctx, "http://o.com/library/abc")
+	check(err)
+}
+
+func TestPullManifestError(t *testing.T) {
+	c, ctx := newRegistryClient(t, func(w http.ResponseWriter, r *http.Request) {
+		checkRequest(t, r, "GET", "/v2/library/abc/manifests/latest")
+		w.WriteHeader(http.StatusNotFound)
+		io.WriteString(w, `{"errors":[{"code":"MANIFEST_UNKNOWN"}]}`)
+	})
+
+	err := c.Pull(ctx, "http://o.com/library/abc")
+	if err == nil {
+		t.Fatalf("expected error")
+	}
+	var got *Error
+	if !errors.Is(err, ErrModelNotFound) {
+		t.Fatalf("err = %v, want %v", got, ErrModelNotFound)
+	}
+}
+
+func TestPullLayerError(t *testing.T) {
+	c, ctx := newRegistryClient(t, func(w http.ResponseWriter, r *http.Request) {
+		checkRequest(t, r, "GET", "/v2/library/abc/manifests/latest")
+		io.WriteString(w, `!`)
+	})
+
+	err := c.Pull(ctx, "http://o.com/library/abc")
+	if err == nil {
+		t.Fatalf("expected error")
+	}
+	var want *json.SyntaxError
+	if !errors.As(err, &want) {
+		t.Fatalf("err = %T, want %T", err, want)
+	}
+}
+
+func TestPullLayerChecksumError(t *testing.T) {
+	var step atomic.Int64
+	c, _ := newRegistryClient(t, func(w http.ResponseWriter, r *http.Request) {
+		switch step.Add(1) {
+		case 1:
+			checkRequest(t, r, "GET", "/v2/library/abc/manifests/latest")
+			io.WriteString(w, `{"layers":[{"size":3,"digest":"sha256:ba7816bf8f01cfea414140de5dae2223b00361a396177a9cb410ff61f20015ad"}]}`)
+		case 2:
+			checkRequest(t, r, "GET", "/v2/library/abc/chunksums/sha256:ba7816bf8f01cfea414140de5dae2223b00361a396177a9cb410ff61f20015ad")
+			w.Header().Set("Content-Location", "http://blob.store/v2/library/abc/blobs/sha256:ba7816bf8f01cfea414140de5dae2223b00361a396177a9cb410ff61f20015ad")
+			fmt.Fprintf(w, "%s 0-1\n", blob.DigestFromBytes("ab"))
+			fmt.Fprintf(w, "%s 2-2\n", blob.DigestFromBytes("c"))
+		case 3:
+			w.WriteHeader(http.StatusNotFound)
+			io.WriteString(w, `{"errors":[{"code":"BLOB_UNKNOWN"}]}`)
+		case 4:
+			io.WriteString(w, "c")
+		default:
+			t.Errorf("unexpected steps %d: %v", step.Load(), r)
+			http.Error(w, "unexpected steps", http.StatusInternalServerError)
+		}
+	})
+
+	c.MaxStreams = 1
+	c.ChunkingThreshold = 1 // force chunking
+
+	var written atomic.Int64
+	ctx := WithTrace(t.Context(), &Trace{
+		Update: func(l *Layer, n int64, err error) {
+			t.Log("trace:", l.Digest.Short(), n, err)
+			written.Add(n)
+		},
+	})
+
+	err := c.Pull(ctx, "http://o.com/library/abc")
+	var got *Error
+	if !errors.As(err, &got) || got.Code != "BLOB_UNKNOWN" {
+		t.Fatalf("err = %v, want %v", err, got)
+	}
+
+	if g := written.Load(); g != 1 {
+		t.Fatalf("wrote %d bytes, want 1", g)
+	}
+}
+
+func TestPullChunksumStreamError(t *testing.T) {
+	var step atomic.Int64
+	c, ctx := newRegistryClient(t, func(w http.ResponseWriter, r *http.Request) {
+		switch step.Add(1) {
+		case 1:
+			checkRequest(t, r, "GET", "/v2/library/abc/manifests/latest")
+			io.WriteString(w, `{"layers":[{"size":3,"digest":"sha256:ba7816bf8f01cfea414140de5dae2223b00361a396177a9cb410ff61f20015ad"}]}`)
+		case 2:
+			w.Header().Set("Content-Location", "http://blob.store/v2/library/abc/blobs/sha256:ba7816bf8f01cfea414140de5dae2223b00361a396177a9cb410ff61f20015ad")
+
+			// Write one valid chunksum and one invalid chunksum
+			fmt.Fprintf(w, "%s 0-1\n", blob.DigestFromBytes("ab")) // valid
+			fmt.Fprint(w, "sha256:!")                              // invalid
+		case 3:
+			io.WriteString(w, "ab")
+		default:
+			t.Errorf("unexpected steps %d: %v", step.Load(), r)
+			http.Error(w, "unexpected steps", http.StatusInternalServerError)
+		}
+	})
+
+	c.ChunkingThreshold = 1 // force chunking
+
+	got := c.Pull(ctx, "http://o.com/library/abc")
+	if !errors.Is(got, ErrIncomplete) {
+		t.Fatalf("err = %v, want %v", got, ErrIncomplete)
+	}
+}
+
+type flushAfterWriter struct {
+	w io.Writer
+}
+
+func (f *flushAfterWriter) Write(p []byte) (n int, err error) {
+	n, err = f.w.Write(p)
+	f.w.(http.Flusher).Flush() // panic if not a flusher
+	return
+}
+
+func TestPullChunksumStreaming(t *testing.T) {
+	csr, csw := io.Pipe()
+	defer csw.Close()
+
+	var step atomic.Int64
+	c, _ := newRegistryClient(t, func(w http.ResponseWriter, r *http.Request) {
+		switch step.Add(1) {
+		case 1:
+			checkRequest(t, r, "GET", "/v2/library/abc/manifests/latest")
+			io.WriteString(w, `{"layers":[{"size":3,"digest":"sha256:ba7816bf8f01cfea414140de5dae2223b00361a396177a9cb410ff61f20015ad"}]}`)
+		case 2:
+			w.Header().Set("Content-Location", "http://blob.store/v2/library/abc/blobs/sha256:ba7816bf8f01cfea414140de5dae2223b00361a396177a9cb410ff61f20015ad")
+			fw := &flushAfterWriter{w} // ensure client gets data as it arrives by aggressively flushing
+			_, err := io.Copy(fw, csr)
+			if err != nil {
+				t.Errorf("copy: %v", err)
+			}
+		case 3:
+			io.WriteString(w, "ab")
+		case 4:
+			io.WriteString(w, "c")
+		default:
+			t.Errorf("unexpected steps %d: %v", step.Load(), r)
+			http.Error(w, "unexpected steps", http.StatusInternalServerError)
+		}
+	})
+
+	c.ChunkingThreshold = 1 // force chunking
+
+	update := make(chan int64, 1)
+	ctx := WithTrace(t.Context(), &Trace{
+		Update: func(l *Layer, n int64, err error) {
+			t.Log("trace:", l.Digest.Short(), n, err)
+			if n > 0 {
+				update <- n
+			}
+		},
+	})
+
+	errc := make(chan error, 1)
+	go func() {
+		errc <- c.Pull(ctx, "http://o.com/library/abc")
+	}()
+
+	// Send first chunksum and ensure it kicks off work immediately
+	fmt.Fprintf(csw, "%s 0-1\n", blob.DigestFromBytes("ab"))
+	if g := <-update; g != 2 {
+		t.Fatalf("got %d, want 2", g)
+	}
+
+	// now send the second chunksum and ensure it kicks off work immediately
+	fmt.Fprintf(csw, "%s 2-2\n", blob.DigestFromBytes("c"))
+	if g := <-update; g != 1 {
+		t.Fatalf("got %d, want 1", g)
+	}
+	csw.Close()
+	testutil.Check(t, <-errc)
+}
+
+func TestPullChunksumsCached(t *testing.T) {
+	var step atomic.Int64
+	c, _ := newRegistryClient(t, func(w http.ResponseWriter, r *http.Request) {
+		switch step.Add(1) {
+		case 1:
+			checkRequest(t, r, "GET", "/v2/library/abc/manifests/latest")
+			io.WriteString(w, `{"layers":[{"size":3,"digest":"sha256:ba7816bf8f01cfea414140de5dae2223b00361a396177a9cb410ff61f20015ad"}]}`)
+		case 2:
+			w.Header().Set("Content-Location", "http://blob.store/v2/library/abc/blobs/sha256:ba7816bf8f01cfea414140de5dae2223b00361a396177a9cb410ff61f20015ad")
+			fmt.Fprintf(w, "%s 0-1\n", blob.DigestFromBytes("ab"))
+			fmt.Fprintf(w, "%s 2-2\n", blob.DigestFromBytes("c"))
+		case 3, 4:
+			switch rng := r.Header.Get("Range"); rng {
+			case "bytes=0-1":
+				io.WriteString(w, "ab")
+			case "bytes=2-2":
+				io.WriteString(w, "c")
+			default:
+				t.Errorf("unexpected range %q", rng)
+			}
+		default:
+			t.Errorf("unexpected steps %d: %v", step.Load(), r)
+			http.Error(w, "unexpected steps", http.StatusInternalServerError)
+		}
+	})
+
+	c.MaxStreams = 1        // force serial processing of chunksums
+	c.ChunkingThreshold = 1 // force chunking
+
+	ctx, cancel := context.WithCancel(t.Context())
+	defer cancel()
+
+	// Cancel the pull after the first chunksum is processed, but before
+	// the second chunksum is processed (which is waiting because
+	// MaxStreams=1). This should cause the second chunksum to error out
+	// leaving the blob incomplete.
+	ctx = WithTrace(ctx, &Trace{
+		Update: func(l *Layer, n int64, err error) {
+			if n > 0 {
+				cancel()
+			}
+		},
+	})
+	err := c.Pull(ctx, "http://o.com/library/abc")
+	if !errors.Is(err, context.Canceled) {
+		t.Fatalf("err = %v, want %v", err, context.Canceled)
+	}
+
+	_, err = c.Cache.Resolve("o.com/library/abc:latest")
+	if !errors.Is(err, fs.ErrNotExist) {
+		t.Fatalf("err = %v, want nil", err)
+	}
+
+	// Reset state and pull again to ensure the blob chunks that should
+	// have been cached are, and the remaining chunk was downloaded, making
+	// the blob complete.
+	step.Store(0)
+	var written atomic.Int64
+	var cached atomic.Int64
+	ctx = WithTrace(t.Context(), &Trace{
+		Update: func(l *Layer, n int64, err error) {
+			t.Log("trace:", l.Digest.Short(), n, err)
+			if errors.Is(err, ErrCached) {
+				cached.Add(n)
+			}
+			written.Add(n)
+		},
+	})
+
+	check := testutil.Checker(t)
+
+	err = c.Pull(ctx, "http://o.com/library/abc")
+	check(err)
+
+	_, err = c.Cache.Resolve("o.com/library/abc:latest")
+	check(err)
+
+	if g := written.Load(); g != 3 {
+		t.Fatalf("wrote %d bytes, want 3", g)
+	}
+	if g := cached.Load(); g != 2 { // "ab" should have been cached
+		t.Fatalf("cached %d bytes, want 3", g)
 	}
 }

server/modelpath.go

@@ -31,9 +31,10 @@ const (
 
 var (
 	ErrInvalidImageFormat  = errors.New("invalid image format")
+	ErrInvalidDigestFormat = errors.New("invalid digest format")
 	ErrInvalidProtocol     = errors.New("invalid protocol scheme")
 	ErrInsecureProtocol    = errors.New("insecure protocol http")
-	ErrInvalidDigestFormat = errors.New("invalid digest format")
+	ErrModelPathInvalid    = errors.New("invalid model path")
 )
 
 func ParseModelPath(name string) ModelPath {
@@ -73,8 +74,6 @@ func ParseModelPath(name string) ModelPath {
 	return mp
 }
 
-var errModelPathInvalid = errors.New("invalid model path")
-
 func (mp ModelPath) GetNamespaceRepository() string {
 	return fmt.Sprintf("%s/%s", mp.Namespace, mp.Repository)
 }

server/routes.go

@@ -777,7 +777,7 @@ func (s *Server) ShowHandler(c *gin.Context) {
 func GetModelInfo(req api.ShowRequest) (*api.ShowResponse, error) {
 	name := model.ParseName(req.Model)
 	if !name.IsValid() {
-		return nil, errModelPathInvalid
+		return nil, ErrModelPathInvalid
 	}
 	name, err := getExistingName(name)
 	if err != nil {

server/sched.go

@@ -711,7 +711,7 @@ func pickBestFullFitByLibrary(req *LlmRequest, f *ggml.GGML, gpus discover.GpuIn
 			req.opts.NumCtx = req.origNumCtx * p
 			if !envconfig.SchedSpread() {
 				for _, g := range sgl {
-					if ok, estimatedVRAM = llm.PredictServerFit([]discover.GpuInfo{g}, f, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts); ok {
+					if ok, estimatedVRAM = llm.PredictServerFit([]discover.GpuInfo{g}, f, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts, p); ok {
 						slog.Info("new model will fit in available VRAM in single GPU, loading", "model", req.model.ModelPath, "gpu", g.ID, "parallel", p, "available", g.FreeMemory, "required", format.HumanBytes2(estimatedVRAM))
 						*numParallel = p
 						return []discover.GpuInfo{g}
@@ -727,7 +727,7 @@ func pickBestFullFitByLibrary(req *LlmRequest, f *ggml.GGML, gpus discover.GpuIn
 		// Now try all the GPUs
 		for _, p := range numParallelToTry {
 			req.opts.NumCtx = req.origNumCtx * p
-			if ok, estimatedVRAM = llm.PredictServerFit(sgl, f, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts); ok {
+			if ok, estimatedVRAM = llm.PredictServerFit(sgl, f, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts, p); ok {
 				slog.Info("new model will fit in available VRAM, loading", "model", req.model.ModelPath, "library", sgl[0].Library, "parallel", p, "required", format.HumanBytes2(estimatedVRAM))
 				*numParallel = p
 				return sgl
@@ -750,7 +750,7 @@ func pickBestPartialFitByLibrary(req *LlmRequest, f *ggml.GGML, gpus discover.Gp
 	var bestEstimate uint64
 	var bestFit int
 	for i, gl := range byLibrary {
-		_, estimatedVRAM := llm.PredictServerFit(gl, f, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts)
+		_, estimatedVRAM := llm.PredictServerFit(gl, f, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts, *numParallel)
 		if estimatedVRAM > bestEstimate {
 			bestEstimate = estimatedVRAM
 			bestFit = i
@@ -825,7 +825,7 @@ func (s *Scheduler) expireRunner(model *Model) {
 // If not, pick a runner to unload, else return nil and the request can be loaded
 func (s *Scheduler) maybeFindCPURunnerToUnload(req *LlmRequest, f *ggml.GGML, gpus discover.GpuInfoList) *runnerRef {
 	slog.Debug("evaluating if CPU model load will fit in available system memory")
-	estimate := llm.EstimateGPULayers(gpus, f, req.model.ProjectorPaths, req.opts)
+	estimate := llm.EstimateGPULayers(gpus, f, req.model.ProjectorPaths, req.opts, req.opts.NumCtx/req.origNumCtx)
 	if estimate.TotalSize <= gpus[0].FreeMemory {
 		slog.Debug("cpu inference mode, model fits in available system memory", "model", format.HumanBytes2(estimate.TotalSize), "available", format.HumanBytes2(gpus[0].FreeMemory))
 		return nil