Merge branch 'main' into drifkin/array-head-count-simple

convert/convert_mistral.go

@@ -89,8 +89,8 @@ func (p *mistral3Model) KV(t *Tokenizer) ggml.KV {
 	return kv
 }
 
-func (p *mistral3Model) Tensors(ts []Tensor) []ggml.Tensor {
-	var out []ggml.Tensor
+func (p *mistral3Model) Tensors(ts []Tensor) []*ggml.Tensor {
+	var out []*ggml.Tensor
 
 	for _, t := range ts {
 		if !strings.HasPrefix(t.Name(), "v.") {
@@ -100,7 +100,7 @@ func (p *mistral3Model) Tensors(ts []Tensor) []ggml.Tensor {
 			}
 		}
 
-		out = append(out, ggml.Tensor{
+		out = append(out, &ggml.Tensor{
 			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    t.Shape(),

convert/convert_mixtral.go

@@ -29,7 +29,7 @@ func (p *mixtralModel) KV(t *Tokenizer) ggml.KV {
 	return kv
 }
 
-func (p *mixtralModel) Tensors(ts []Tensor) []ggml.Tensor {
+func (p *mixtralModel) Tensors(ts []Tensor) []*ggml.Tensor {
 	oldnew := []string{
 		"model.layers", "blk",
 		"w1", "ffn_gate_exps",
@@ -56,10 +56,10 @@ func (p *mixtralModel) Tensors(ts []Tensor) []ggml.Tensor {
 		return true
 	})
 
-	var out []ggml.Tensor
+	var out []*ggml.Tensor
 	for n, e := range experts {
 		// TODO(mxyng): sanity check experts
-		out = append(out, ggml.Tensor{
+		out = append(out, &ggml.Tensor{
 			Name:     n,
 			Kind:     e[0].Kind(),
 			Shape:    append([]uint64{uint64(len(e))}, e[0].Shape()...),

convert/convert_phi3.go

@@ -68,19 +68,19 @@ func (p *phi3Model) KV(t *Tokenizer) ggml.KV {
 	return kv
 }
 
-func (p *phi3Model) Tensors(ts []Tensor) []ggml.Tensor {
+func (p *phi3Model) Tensors(ts []Tensor) []*ggml.Tensor {
 	var addRopeFactors sync.Once
 
-	out := make([]ggml.Tensor, 0, len(ts)+2)
+	out := make([]*ggml.Tensor, 0, len(ts)+2)
 	for _, t := range ts {
 		if strings.HasPrefix(t.Name(), "blk.0.") {
 			addRopeFactors.Do(func() {
-				out = append(out, ggml.Tensor{
+				out = append(out, &ggml.Tensor{
 					Name:     "rope_factors_long.weight",
 					Kind:     0,
 					Shape:    []uint64{uint64(len(p.RopeScaling.LongFactor))},
 					WriterTo: p.RopeScaling.LongFactor,
-				}, ggml.Tensor{
+				}, &ggml.Tensor{
 					Name:     "rope_factors_short.weight",
 					Kind:     0,
 					Shape:    []uint64{uint64(len(p.RopeScaling.ShortFactor))},
@@ -89,7 +89,7 @@ func (p *phi3Model) Tensors(ts []Tensor) []ggml.Tensor {
 			})
 		}
 
-		out = append(out, ggml.Tensor{
+		out = append(out, &ggml.Tensor{
 			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    t.Shape(),

convert/convert_qwen2.go

@@ -45,10 +45,10 @@ func (q *qwen2Model) KV(t *Tokenizer) ggml.KV {
 	return kv
 }
 
-func (q *qwen2Model) Tensors(ts []Tensor) []ggml.Tensor {
-	var out []ggml.Tensor
+func (q *qwen2Model) Tensors(ts []Tensor) []*ggml.Tensor {
+	var out []*ggml.Tensor
 	for _, t := range ts {
-		out = append(out, ggml.Tensor{
+		out = append(out, &ggml.Tensor{
 			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    t.Shape(),

convert/convert_test.go

@@ -130,6 +130,7 @@ func TestConvertModel(t *testing.T) {
 			if err != nil {
 				t.Fatal(err)
 			}
+			defer expectFile.Close()
 
 			var expect map[string]string
 			if err := json.NewDecoder(expectFile).Decode(&expect); err != nil {

convert/fs.go

-package convert
-
-import (
-	"archive/zip"
-	"errors"
-	"io"
-	"io/fs"
-	"os"
-	"path/filepath"
-)
-
-type ZipReader struct {
-	r *zip.Reader
-	p string
-
-	// limit is the maximum size of a file that can be read directly
-	// from the zip archive. Files larger than this size will be extracted
-	limit int64
-}
-
-func NewZipReader(r *zip.Reader, p string, limit int64) fs.FS {
-	return &ZipReader{r, p, limit}
-}
-
-func (z *ZipReader) Open(name string) (fs.File, error) {
-	r, err := z.r.Open(name)
-	if err != nil {
-		return nil, err
-	}
-	defer r.Close()
-
-	if fi, err := r.Stat(); err != nil {
-		return nil, err
-	} else if fi.Size() < z.limit {
-		return r, nil
-	}
-
-	if !filepath.IsLocal(name) {
-		return nil, zip.ErrInsecurePath
-	}
-
-	n := filepath.Join(z.p, name)
-	if _, err := os.Stat(n); errors.Is(err, os.ErrNotExist) {
-		w, err := os.Create(n)
-		if err != nil {
-			return nil, err
-		}
-		defer w.Close()
-
-		if _, err := io.Copy(w, r); err != nil {
-			return nil, err
-		}
-	} else if err != nil {
-		return nil, err
-	}
-
-	return os.Open(n)
-}

discover/cuda_common.go

@@ -3,6 +3,7 @@
 package discover
 
 import (
+	"fmt"
 	"log/slog"
 	"os"
 	"regexp"
@@ -59,6 +60,8 @@ func cudaVariant(gpuInfo CudaGPUInfo) string {
 
 	// driver 12.0 has problems with the cuda v12 library, so run v11 on those older drivers
 	if gpuInfo.DriverMajor < 12 || (gpuInfo.DriverMajor == 12 && gpuInfo.DriverMinor == 0) {
+		// The detected driver is older than Feb 2023
+		slog.Warn("old CUDA driver detected - please upgrade to a newer driver", "version", fmt.Sprintf("%d.%d", gpuInfo.DriverMajor, gpuInfo.DriverMinor))
 		return "v11"
 	}
 	return "v12"

discover/gpu_info.h

@@ -27,12 +27,14 @@
 
 #endif
 
+#ifndef LOG
 #define LOG(verbose, ...) \
   do { \
     if (verbose) { \
       fprintf(stderr, __VA_ARGS__); \
     } \
   } while (0)
+#endif
 
 #ifdef __cplusplus
 extern "C" {

discover/gpu_info_cudart.c

 #ifndef __APPLE__  // TODO - maybe consider nvidia support on intel macs?
 
 #include <string.h>
+#include <inttypes.h>
 #include "gpu_info_cudart.h"
 
 void cudart_init(char *cudart_lib_path, cudart_init_resp_t *resp) {
@@ -58,7 +59,7 @@ void cudart_init(char *cudart_lib_path, cudart_init_resp_t *resp) {
     LOG(resp->ch.verbose, "cudaSetDevice err: %d\n", ret);
     UNLOAD_LIBRARY(resp->ch.handle);
     resp->ch.handle = NULL;
-    if (ret == CUDA_ERROR_INSUFFICIENT_DRIVER) {
+    if (ret == CUDART_ERROR_INSUFFICIENT_DRIVER) {
       resp->err = strdup("your nvidia driver is too old or missing.  If you have a CUDA GPU please upgrade to run ollama");
       return;
     }
@@ -168,9 +169,9 @@ void cudart_bootstrap(cudart_handle_t h, int i, mem_info_t *resp) {
   resp->free = memInfo.free;
   resp->used = memInfo.used;
 
-  LOG(h.verbose, "[%s] CUDA totalMem %lu\n", resp->gpu_id, resp->total);
-  LOG(h.verbose, "[%s] CUDA freeMem %lu\n", resp->gpu_id, resp->free);
-  LOG(h.verbose, "[%s] CUDA usedMem %lu\n", resp->gpu_id, resp->used);
+  LOG(h.verbose, "[%s] CUDA totalMem %" PRId64 "\n", resp->gpu_id, resp->total);
+  LOG(h.verbose, "[%s] CUDA freeMem %" PRId64 "\n", resp->gpu_id, resp->free);
+  LOG(h.verbose, "[%s] CUDA usedMem %" PRId64 "\n", resp->gpu_id, resp->used);
   LOG(h.verbose, "[%s] Compute Capability %d.%d\n", resp->gpu_id, resp->major, resp->minor);
 }
 
@@ -180,4 +181,4 @@ void cudart_release(cudart_handle_t h) {
   h.handle = NULL;
 }
 
-#endif  // __APPLE__
\ No newline at end of file
+#endif  // __APPLE__

discover/gpu_info_nvcuda.c

 #ifndef __APPLE__  // TODO - maybe consider nvidia support on intel macs?
 
 #include <string.h>
+#include <inttypes.h>
 #include "gpu_info_nvcuda.h"
 
 void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp) {
@@ -193,8 +194,8 @@ void nvcuda_bootstrap(nvcuda_handle_t h, int i, mem_info_t *resp) {
   resp->total = memInfo.total;
   resp->free = memInfo.free;
 
-  LOG(h.verbose, "[%s] CUDA totalMem %lu mb\n", resp->gpu_id, resp->total / 1024 / 1024);
-  LOG(h.verbose, "[%s] CUDA freeMem %lu mb\n", resp->gpu_id, resp->free / 1024 / 1024);
+  LOG(h.verbose, "[%s] CUDA totalMem %" PRId64 "mb\n", resp->gpu_id, resp->total / 1024 / 1024);
+  LOG(h.verbose, "[%s] CUDA freeMem %" PRId64 "mb\n", resp->gpu_id, resp->free / 1024 / 1024);
   LOG(h.verbose, "[%s] Compute Capability %d.%d\n", resp->gpu_id, resp->major, resp->minor);
 
   
@@ -247,4 +248,4 @@ void nvcuda_release(nvcuda_handle_t h) {
   h.handle = NULL;
 }
 
-#endif  // __APPLE__
\ No newline at end of file
+#endif  // __APPLE__

discover/path.go

@@ -12,7 +12,7 @@ import (
 // '../lib/ollama' on Linux and the executable's directory on macOS
 // note: distribution builds, additional GPU-specific libraries are
 // found in subdirectories of the returned path, such as
-// 'cuda_v11', 'cuda_v12', 'rocm', etc.
+// 'cuda_v12', 'rocm', etc.
 var LibOllamaPath string = func() string {
 	exe, err := os.Executable()
 	if err != nil {

docs/api.md

@@ -394,9 +394,6 @@ curl http://localhost:11434/api/generate -d '{
     "repeat_penalty": 1.2,
     "presence_penalty": 1.5,
     "frequency_penalty": 1.0,
-    "mirostat": 1,
-    "mirostat_tau": 0.8,
-    "mirostat_eta": 0.6,
     "penalize_newline": true,
     "stop": ["\n", "user:"],
     "numa": false,
@@ -404,10 +401,7 @@ curl http://localhost:11434/api/generate -d '{
     "num_batch": 2,
     "num_gpu": 1,
     "main_gpu": 0,
-    "low_vram": false,
-    "vocab_only": false,
     "use_mmap": true,
-    "use_mlock": false,
     "num_thread": 8
   }
 }'

docs/faq.md

@@ -20,7 +20,7 @@ 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 4096 tokens, unless you have a single GPU with <= 4 GB of VRAM, in which case it will default to 2048 tokens. 
+By default, Ollama uses a context window size of 4096 tokens. 
 
 This can be overridden with the `OLLAMA_CONTEXT_LENGTH` environment variable. For example, to set the default context window to 8K, use: 
 
@@ -31,7 +31,7 @@ OLLAMA_CONTEXT_LENGTH=8192 ollama serve
 To change this when using `ollama run`, use `/set parameter`:
 
 ```shell
-/set parameter num_ctx 8192
+/set parameter num_ctx 4096
 ```
 
 When using the API, specify the `num_ctx` parameter:
@@ -41,7 +41,7 @@ curl http://localhost:11434/api/generate -d '{
   "model": "llama3.2",
   "prompt": "Why is the sky blue?",
   "options": {
-    "num_ctx": 8192
+    "num_ctx": 4096
   }
 }'
 ```

docs/gpu.md

 # GPU
 ## Nvidia
-Ollama supports Nvidia GPUs with compute capability 5.0+.
+Ollama supports Nvidia GPUs with compute capability 5.0+ and driver version 531 and newer.
 
 Check your compute compatibility to see if your card is supported:
 [https://developer.nvidia.com/cuda-gpus](https://developer.nvidia.com/cuda-gpus)

docs/modelfile.md

@@ -150,9 +150,6 @@ PARAMETER <parameter> <parametervalue>
 
 | Parameter      | Description                                                                                                                                                                                                                                             | Value Type | Example Usage        |
 | -------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------- | -------------------- |
-| mirostat       | Enable Mirostat sampling for controlling perplexity. (default: 0, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0)                                                                                                                                         | int        | mirostat 0           |
-| mirostat_eta   | Influences how quickly the algorithm responds to feedback from the generated text. A lower learning rate will result in slower adjustments, while a higher learning rate will make the algorithm more responsive. (Default: 0.1)                        | float      | mirostat_eta 0.1     |
-| mirostat_tau   | Controls the balance between coherence and diversity of the output. A lower value will result in more focused and coherent text. (Default: 5.0)                                                                                                         | float      | mirostat_tau 5.0     |
 | num_ctx        | Sets the size of the context window used to generate the next token. (Default: 2048)                                                                                                                                                                    | int        | num_ctx 4096         |
 | repeat_last_n  | Sets how far back for the model to look back to prevent repetition. (Default: 64, 0 = disabled, -1 = num_ctx)                                                                                                                                           | int        | repeat_last_n 64     |
 | repeat_penalty | Sets how strongly to penalize repetitions. A higher value (e.g., 1.5) will penalize repetitions more strongly, while a lower value (e.g., 0.9) will be more lenient. (Default: 1.1)                                                                     | float      | repeat_penalty 1.1   |

docs/troubleshooting.md

@@ -43,7 +43,7 @@ Ollama includes multiple LLM libraries compiled for different GPUs and CPU vecto
 In the server log, you will see a message that looks something like this (varies from release to release):
 
 ```
-Dynamic LLM libraries [rocm_v6 cpu cpu_avx cpu_avx2 cuda_v11 rocm_v5]
+Dynamic LLM libraries [rocm_v6 cpu cpu_avx cpu_avx2 cuda_v12 rocm_v5]
 ```
 
 **Experimental LLM Library Override**

envconfig/config.go

@@ -169,7 +169,7 @@ var (
 	// Enable the new Ollama engine
 	NewEngine = Bool("OLLAMA_NEW_ENGINE")
 	// ContextLength sets the default context length
-	ContextLength = Int64("OLLAMA_CONTEXT_LENGTH", -1)
+	ContextLength = Uint("OLLAMA_CONTEXT_LENGTH", 4096)
 )
 
 func String(s string) func() string {
@@ -227,20 +227,6 @@ func Uint64(key string, defaultValue uint64) func() uint64 {
 	}
 }
 
-func Int64(key string, defaultValue int64) func() int64 {
-	return func() int64 {
-		if s := Var(key); s != "" {
-			if n, err := strconv.ParseInt(s, 10, 64); err != nil {
-				slog.Warn("invalid environment variable, using default", "key", key, "value", s, "default", defaultValue)
-			} else {
-				return n
-			}
-		}
-
-		return defaultValue
-	}
-}
-
 // Set aside VRAM per GPU
 var GpuOverhead = Uint64("OLLAMA_GPU_OVERHEAD", 0)
 
@@ -269,7 +255,7 @@ func AsMap() map[string]EnvVar {
 		"OLLAMA_ORIGINS":           {"OLLAMA_ORIGINS", AllowedOrigins(), "A comma separated list of allowed origins"},
 		"OLLAMA_SCHED_SPREAD":      {"OLLAMA_SCHED_SPREAD", SchedSpread(), "Always schedule model across all GPUs"},
 		"OLLAMA_MULTIUSER_CACHE":   {"OLLAMA_MULTIUSER_CACHE", MultiUserCache(), "Optimize prompt caching for multi-user scenarios"},
-		"OLLAMA_CONTEXT_LENGTH":    {"OLLAMA_CONTEXT_LENGTH", ContextLength(), "Context length to use unless otherwise specified (default 4096 or 2048 with low VRAM)"},
+		"OLLAMA_CONTEXT_LENGTH":    {"OLLAMA_CONTEXT_LENGTH", ContextLength(), "Context length to use unless otherwise specified (default: 4096)"},
 		"OLLAMA_NEW_ENGINE":        {"OLLAMA_NEW_ENGINE", NewEngine(), "Enable the new Ollama engine"},
 
 		// Informational

envconfig/config_test.go

@@ -278,9 +278,9 @@ func TestVar(t *testing.T) {
 }
 
 func TestContextLength(t *testing.T) {
-	cases := map[string]int64{
-		"":     -1,
-		"4096": 4096,
+	cases := map[string]uint{
+		"":     4096,
+		"2048": 2048,
 	}
 
 	for k, v := range cases {

fs/ggml/ggml.go

@@ -37,12 +37,12 @@ func (kv KV) ParameterCount() uint64 {
 	return val
 }
 
-func (kv KV) FileType() fileType {
+func (kv KV) FileType() FileType {
 	if t := kv.Uint("general.file_type"); t > 0 {
-		return fileType(t)
+		return FileType(t)
 	}
 
-	return fileTypeUnknown
+	return FileTypeUnknown
 }
 
 func (kv KV) BlockCount() uint64 {
@@ -194,7 +194,7 @@ func keyValue[T valueTypes | arrayValueTypes](kv KV, key string, defaultValue ..
 		return val, true
 	}
 
-	slog.Warn("key with type not found", "key", key, "default", defaultValue[0])
+	slog.Debug("key with type not found", "key", key, "default", defaultValue[0])
 	return defaultValue[0], false
 }
 
@@ -271,7 +271,11 @@ func (t Tensor) block() (n int) {
 }
 
 func (t Tensor) blockSize() uint64 {
-	switch t.Kind {
+	return (TensorType)(t.Kind).BlockSize()
+}
+
+func (t TensorType) BlockSize() uint64 {
+	switch t {
 	case
 		0,  // F32
 		1,  // F16
@@ -297,73 +301,77 @@ func (t Tensor) blockSize() uint64 {
 }
 
 func (t Tensor) typeSize() uint64 {
-	blockSize := t.blockSize()
+	return TensorType(t.Kind).TypeSize()
+}
+
+func (t TensorType) TypeSize() uint64 {
+	blockSize := t.BlockSize()
 
-	switch t.Kind {
-	case 0: // FP32
+	switch t {
+	case TensorTypeF32:
 		return 4
-	case 1: // FP16
+	case TensorTypeF16:
 		return 2
-	case 2: // Q4_0
+	case TensorTypeQ4_0:
 		return 2 + blockSize/2
-	case 3: // Q4_1
+	case TensorTypeQ4_1:
 		return 2 + 2 + blockSize/2
-	case 6: // Q5_0
+	case TensorTypeQ5_0:
 		return 2 + 4 + blockSize/2
-	case 7: // Q5_1
+	case TensorTypeQ5_1:
 		return 2 + 2 + 4 + blockSize/2
-	case 8: // Q8_0
+	case TensorTypeQ8_0:
 		return 2 + blockSize
-	case 9: // Q8_1
+	case TensorTypeQ8_1:
 		return 2 + 2 + blockSize
-	case 10: // Q2_K
+	case TensorTypeQ2_K:
 		return blockSize/16 + blockSize/4 + 2 + 2
-	case 11: // Q3_K
+	case TensorTypeQ3_K:
 		return blockSize/8 + blockSize/4 + 12 + 2
-	case 12: // Q4_K
+	case TensorTypeQ4_K:
 		return 2 + 2 + 12 + blockSize/2
-	case 13: // Q5_K
+	case TensorTypeQ5_K:
 		return 2 + 2 + 12 + blockSize/8 + blockSize/2
-	case 14: // Q6_K
+	case TensorTypeQ6_K:
 		return blockSize/2 + blockSize/4 + blockSize/16 + 2
-	case 15: // Q8_K
+	case TensorTypeQ8_K:
 		return 4 + blockSize + 2*blockSize/16
-	case 16: // IQ2_XXS
+	case tensorTypeIQ2_XXS:
 		return 2 + 2*blockSize/8
-	case 17: // IQ2_XS
+	case tensorTypeIQ2_XS:
 		return 2 + 2*blockSize/8 + blockSize/32
-	case 18: // IQ3_XXS
+	case tensorTypeIQ3_XXS:
 		return 2 + blockSize/4 + blockSize/8
-	case 19: // IQ1_S
+	case tensorTypeIQ1_S:
 		return 2 + blockSize/8 + blockSize/16
-	case 20: // IQ4_NL
+	case tensorTypeIQ4_NL:
 		return 2 + blockSize/2
-	case 21: // IQ3_S
+	case tensorTypeIQ3_S:
 		return 2 + blockSize/4 + blockSize/8 + blockSize/32 + 4
-	case 22: // IQ2_S
+	case tensorTypeIQ2_S:
 		return 2 + blockSize/4 + blockSize/16
-	case 23: // IQ4_XS
+	case tensorTypeIQ4_XS:
 		return 2 + 2 + blockSize/2 + blockSize/64
-	case 24: // I8
+	case TensorTypeI8:
 		return 1
-	case 25: // I16
+	case TensorTypeI16:
 		return 2
-	case 26: // I32
+	case TensorTypeI32:
 		return 4
-	case 27: // I64
+	case TensorTypeI64:
 		return 8
-	case 28: // F64
+	case TensorTypeF64:
 		return 8
-	case 29: // IQ1_M
+	case tensorTypeIQ1_M:
 		return blockSize/8 + blockSize/16 + blockSize/32
-	case 30: // BF16
+	case TensorTypeBF16:
 		return 2
 	default:
 		return 0
 	}
 }
 
-func (t Tensor) parameters() uint64 {
+func (t Tensor) Elements() uint64 {
 	var count uint64 = 1
 	for _, n := range t.Shape {
 		count *= n
@@ -372,11 +380,11 @@ func (t Tensor) parameters() uint64 {
 }
 
 func (t Tensor) Size() uint64 {
-	return t.parameters() * t.typeSize() / t.blockSize()
+	return t.Elements() * t.typeSize() / t.blockSize()
 }
 
 func (t Tensor) Type() string {
-	return fileType(t.Kind).String()
+	return TensorType(t.Kind).String()
 }
 
 type container interface {
@@ -525,7 +533,7 @@ func (f GGML) GraphSize(context, batch uint64, numParallel int, kvCacheType stri
 		var ropeFreqsCount uint64
 		if ropeFreqs, ok := f.Tensors().GroupLayers()["rope_freqs"]; ok {
 			if ropeFreqsWeights, ok := ropeFreqs["weights"]; ok {
-				ropeFreqsCount = ropeFreqsWeights.parameters()
+				ropeFreqsCount = ropeFreqsWeights.Elements()
 			}
 		}
 

fs/ggml/gguf.go

@@ -9,8 +9,12 @@ import (
 	"io"
 	"log/slog"
 	"maps"
+	"os"
+	"runtime"
 	"slices"
 	"strings"
+
+	"golang.org/x/sync/errgroup"
 )
 
 type containerGGUF struct {
@@ -225,7 +229,7 @@ func (llm *gguf) Decode(rs io.ReadSeeker) error {
 		}
 
 		llm.tensors = append(llm.tensors, &tensor)
-		llm.parameters += tensor.parameters()
+		llm.parameters += tensor.Elements()
 	}
 
 	// patch KV with parameter count
@@ -488,25 +492,38 @@ func writeGGUFArray[S ~[]E, E any](w io.Writer, t uint32, s S) error {
 		return err
 	}
 
+	if t == ggufTypeString {
+		for _, e := range any(s).([]string) {
+			if err := binary.Write(w, binary.LittleEndian, uint64(len(e))); err != nil {
+				return err
+			}
+
+			if err := binary.Write(w, binary.LittleEndian, []byte(e)); err != nil {
+				return err
+			}
+		}
+		return nil
+	}
+
 	return binary.Write(w, binary.LittleEndian, s)
 }
 
-func WriteGGUF(ws io.WriteSeeker, kv KV, ts []Tensor) error {
+func WriteGGUF(f *os.File, kv KV, ts []*Tensor) error {
 	alignment := kv.Uint("general.alignment", 32)
 
-	if err := binary.Write(ws, binary.LittleEndian, []byte("GGUF")); err != nil {
+	if err := binary.Write(f, binary.LittleEndian, []byte("GGUF")); err != nil {
 		return err
 	}
 
-	if err := binary.Write(ws, binary.LittleEndian, uint32(3)); err != nil {
+	if err := binary.Write(f, binary.LittleEndian, uint32(3)); err != nil {
 		return err
 	}
 
-	if err := binary.Write(ws, binary.LittleEndian, uint64(len(ts))); err != nil {
+	if err := binary.Write(f, binary.LittleEndian, uint64(len(ts))); err != nil {
 		return err
 	}
 
-	if err := binary.Write(ws, binary.LittleEndian, uint64(len(kv))); err != nil {
+	if err := binary.Write(f, binary.LittleEndian, uint64(len(kv))); err != nil {
 		return err
 	}
 
@@ -514,12 +531,12 @@ func WriteGGUF(ws io.WriteSeeker, kv KV, ts []Tensor) error {
 	slices.Sort(keys)
 
 	for _, key := range keys {
-		if err := ggufWriteKV(ws, key, kv[key]); err != nil {
+		if err := ggufWriteKV(f, key, kv[key]); err != nil {
 			return err
 		}
 	}
 
-	slices.SortStableFunc(ts, func(a, b Tensor) int {
+	slices.SortStableFunc(ts, func(a, b *Tensor) int {
 		if i, j := a.block(), b.block(); i < 0 && j > 0 {
 			return 1
 		} else if i > 0 && j < 0 {
@@ -530,21 +547,34 @@ func WriteGGUF(ws io.WriteSeeker, kv KV, ts []Tensor) error {
 	})
 
 	var s uint64
-	for _, t := range ts {
-		t.Offset = s + uint64(ggufPadding(int64(s), int64(alignment)))
-		if err := ggufWriteTensorInfo(ws, t); err != nil {
+	for i := range ts {
+		ts[i].Offset = s
+		if err := ggufWriteTensorInfo(f, ts[i]); err != nil {
 			return err
 		}
-		s += t.Size()
+		s += ts[i].Size()
+		s += uint64(ggufPadding(int64(s), int64(alignment)))
 	}
 
+	offset, err := f.Seek(0, io.SeekCurrent)
+	if err != nil {
+		return err
+	}
+	offset += ggufPadding(offset, int64(alignment))
+
+	var g errgroup.Group
+	g.SetLimit(runtime.GOMAXPROCS(0))
+	// TODO consider reducing if tensors size * gomaxprocs is larger than free memory
 	for _, t := range ts {
-		if err := ggufWriteTensor(ws, t, int64(alignment)); err != nil {
+		t := t
+		w := io.NewOffsetWriter(f, offset+int64(t.Offset))
+		g.Go(func() error {
+			_, err := t.WriteTo(w)
 			return err
-		}
+		})
 	}
 
-	return nil
+	return g.Wait()
 }
 
 func ggufWriteKV(ws io.WriteSeeker, k string, v any) error {
@@ -559,8 +589,10 @@ func ggufWriteKV(ws io.WriteSeeker, k string, v any) error {
 
 	var err error
 	switch v := v.(type) {
-	case uint32:
+	case uint32, FileType:
 		err = writeGGUF(ws, ggufTypeUint32, v)
+	case uint64:
+		err = writeGGUF(ws, ggufTypeUint64, v)
 	case float32:
 		err = writeGGUF(ws, ggufTypeFloat32, v)
 	case bool:
@@ -569,32 +601,20 @@ func ggufWriteKV(ws io.WriteSeeker, k string, v any) error {
 		err = writeGGUFString(ws, v)
 	case []int32:
 		err = writeGGUFArray(ws, ggufTypeInt32, v)
+	case *array[int32]:
+		err = writeGGUFArray(ws, ggufTypeInt32, v.values)
 	case []uint32:
 		err = writeGGUFArray(ws, ggufTypeUint32, v)
+	case *array[uint32]:
+		err = writeGGUFArray(ws, ggufTypeUint32, v.values)
 	case []float32:
 		err = writeGGUFArray(ws, ggufTypeFloat32, v)
+	case *array[float32]:
+		err = writeGGUFArray(ws, ggufTypeFloat32, v.values)
 	case []string:
-		if err := binary.Write(ws, binary.LittleEndian, ggufTypeArray); err != nil {
-			return err
-		}
-
-		if err := binary.Write(ws, binary.LittleEndian, ggufTypeString); err != nil {
-			return err
-		}
-
-		if err := binary.Write(ws, binary.LittleEndian, uint64(len(v))); err != nil {
-			return err
-		}
-
-		for _, e := range v {
-			if err := binary.Write(ws, binary.LittleEndian, uint64(len(e))); err != nil {
-				return err
-			}
-
-			if err := binary.Write(ws, binary.LittleEndian, []byte(e)); err != nil {
-				return err
-			}
-		}
+		err = writeGGUFArray(ws, ggufTypeString, v)
+	case *array[string]:
+		err = writeGGUFArray(ws, ggufTypeString, v.values)
 	default:
 		return fmt.Errorf("improper type for '%s'", k)
 	}
@@ -602,7 +622,7 @@ func ggufWriteKV(ws io.WriteSeeker, k string, v any) error {
 	return err
 }
 
-func ggufWriteTensorInfo(ws io.WriteSeeker, t Tensor) error {
+func ggufWriteTensorInfo(ws io.WriteSeeker, t *Tensor) error {
 	slog.Debug(t.Name, "kind", t.Kind, "shape", t.Shape, "offset", t.Offset)
 	if err := binary.Write(ws, binary.LittleEndian, uint64(len(t.Name))); err != nil {
 		return err
@@ -629,20 +649,6 @@ func ggufWriteTensorInfo(ws io.WriteSeeker, t Tensor) error {
 	return binary.Write(ws, binary.LittleEndian, t.Offset)
 }
 
-func ggufWriteTensor(ws io.WriteSeeker, t Tensor, alignment int64) error {
-	offset, err := ws.Seek(0, io.SeekCurrent)
-	if err != nil {
-		return err
-	}
-
-	if err := binary.Write(ws, binary.LittleEndian, bytes.Repeat([]byte{0}, int(ggufPadding(offset, alignment)))); err != nil {
-		return err
-	}
-
-	_, err = t.WriteTo(ws)
-	return err
-}
-
 func ggufPadding(offset, align int64) int64 {
 	return (align - offset%align) % align
 }