diff --git a/convert/convert.go b/convert/convert.go
index e9c2ef2d..e71a0ff3 100644
--- a/convert/convert.go
+++ b/convert/convert.go
@@ -18,6 +18,16 @@ import (
 	"github.com/ollama/ollama/llm"
 )
 
+const (
+	_ int32 = iota
+	tokenTypeNormal
+	tokenTypeUnknown
+	tokenTypeControl
+	tokenTypeUserDefined
+	tokenTypeUnused
+	tokenTypeByte
+)
+
 type Params struct {
 	Architectures     []string `json:"architectures"`
 	VocabSize         int      `json:"vocab_size"`
@@ -172,7 +182,7 @@ func LoadSentencePieceTokens(dirpath string, params *Params) (*Vocab, error) {
 		}
 		v.Tokens = append(v.Tokens, t.key)
 		v.Scores = append(v.Scores, -1000.0)
-		v.Types = append(v.Types, int32(llm.GGUFTokenUserDefined))
+		v.Types = append(v.Types, tokenTypeUserDefined)
 	}
 	slog.Info(fmt.Sprintf("vocab size w/ extra tokens: %d", len(v.Tokens)))
 
@@ -182,7 +192,7 @@ func LoadSentencePieceTokens(dirpath string, params *Params) (*Vocab, error) {
 		for cnt := 0; cnt < missingTokens; cnt++ {
 			v.Tokens = append(v.Tokens, fmt.Sprintf("<dummy%05d>", cnt+1))
 			v.Scores = append(v.Scores, -1)
-			v.Types = append(v.Types, int32(llm.GGUFTokenUserDefined))
+			v.Types = append(v.Types, tokenTypeUserDefined)
 		}
 	}
 
diff --git a/convert/gemma.go b/convert/gemma.go
index 88abe646..e24b8ec5 100644
--- a/convert/gemma.go
+++ b/convert/gemma.go
@@ -71,8 +71,6 @@ func (m *GemmaModel) GetTensors() error {
 	}
 
 	slog.Debug(fmt.Sprintf("Total tensors: %d", len(t)))
-
-	m.Tensors = []llm.Tensor{}
 	for _, l := range t {
 		if strings.HasSuffix(l.Name, "norm.weight") {
 			wt := l.WriterTo.(safetensorWriterTo)
diff --git a/convert/llama.go b/convert/llama.go
index 83d942cb..a10670e6 100644
--- a/convert/llama.go
+++ b/convert/llama.go
@@ -105,8 +105,6 @@ func (m *LlamaModel) GetTensors() error {
 		return err
 	}
 
-	m.Tensors = []llm.Tensor{}
-
 	pattern := `^blk\.[0-9]+\.attn_(?P<layer>q|k)\.weight$`
 	re, err := regexp.Compile(pattern)
 	if err != nil {
@@ -133,30 +131,22 @@ func (m *LlamaModel) GetTensors() error {
 	return nil
 }
 
-func (m *LlamaModel) LoadVocab() error {
-	v := &Vocab{}
-
-	tokpath := filepath.Join(m.Path, "tokenizer.json")
-	pre, ts, merges, err := parseTokens(tokpath)
+func (m *LlamaModel) LoadVocab() (err error) {
+	pre, ts, merges, err := parseTokens(filepath.Join(m.Path, "tokenizer.json"))
 	if errors.Is(err, os.ErrNotExist) {
-		v, err = LoadSentencePieceTokens(m.Path, m.Params)
-		if err != nil {
-			return err
-		}
+		return nil
 	} else if err != nil {
 		return err
-	} else {
-		for _, t := range ts {
-			v.Tokens = append(v.Tokens, t.Content)
-			v.Types = append(v.Types, t.Type())
-		}
-
-		m.Params.PreTokenizer = pre
-		v.Merges = merges
 	}
 
-	m.Vocab = v
+	m.Vocab = &Vocab{}
+	for _, t := range ts {
+		m.Vocab.Tokens = append(m.Vocab.Tokens, t.Content)
+		m.Vocab.Types = append(m.Vocab.Types, t.Type())
+	}
 
+	m.Vocab.Merges = merges
+	m.Params.PreTokenizer = pre
 	return nil
 }
 
@@ -174,7 +164,7 @@ func (m *LlamaModel) WriteGGUF(ws io.WriteSeeker) error {
 		"llama.attention.head_count":             uint32(m.Params.AttentionHeads),
 		"llama.attention.head_count_kv":          uint32(m.Params.KeyValHeads),
 		"llama.attention.layer_norm_rms_epsilon": float32(m.Params.NormEPS),
-		"general.file_type":                      uint32(2),
+		"general.file_type":                      uint32(1),
 		"tokenizer.ggml.model":                   "gpt2",
 
 		"tokenizer.ggml.pre":        m.Params.PreTokenizer,
diff --git a/convert/mistral.go b/convert/mistral.go
index f88de12b..89d2e084 100644
--- a/convert/mistral.go
+++ b/convert/mistral.go
@@ -102,8 +102,6 @@ func (m *MistralModel) GetTensors() error {
 		return err
 	}
 
-	m.Tensors = []llm.Tensor{}
-
 	pattern := `^blk\.[0-9]+\.attn_(?P<layer>q|k)\.weight$`
 	re, err := regexp.Compile(pattern)
 	if err != nil {
diff --git a/convert/mixtral.go b/convert/mixtral.go
index 940df55d..66546fd7 100644
--- a/convert/mixtral.go
+++ b/convert/mixtral.go
@@ -17,8 +17,6 @@ func (m *MixtralModel) GetTensors() error {
 		return err
 	}
 
-	m.Tensors = []llm.Tensor{}
-
 	pattern := `^blk\.[0-9]+\.attn_(?P<layer>q|k)\.weight$`
 	re, err := regexp.Compile(pattern)
 	if err != nil {
diff --git a/convert/safetensors.go b/convert/safetensors.go
index b52a048d..2107ae81 100644
--- a/convert/safetensors.go
+++ b/convert/safetensors.go
@@ -11,6 +11,7 @@ import (
 	"path/filepath"
 	"regexp"
 	"slices"
+	"strings"
 
 	"github.com/d4l3k/go-bfloat16"
 	"github.com/mitchellh/mapstructure"
@@ -97,6 +98,10 @@ func (m *SafetensorFormat) readTensors(fn string, offset uint64, params *Params)
 
 	var tensors []llm.Tensor
 	for _, k := range keys {
+		if strings.HasSuffix(k, "self_attn.rotary_emb.inv_freq") {
+			continue
+		}
+
 		vals := parsed[k].(map[string]interface{})
 		var data tensorMetaData
 		if err = mapstructure.Decode(vals, &data); err != nil {
diff --git a/convert/tokenizer.go b/convert/tokenizer.go
index a847a84c..e0fe0bb7 100644
--- a/convert/tokenizer.go
+++ b/convert/tokenizer.go
@@ -44,11 +44,11 @@ type Token struct {
 func (t *Token) Type() int32 {
 	switch {
 	case t.Special:
-		return 3
+		return tokenTypeControl
 	case t.UserDefined:
-		return 4
+		return tokenTypeUserDefined
 	default:
-		return 1
+		return tokenTypeNormal
 	}
 }
 
diff --git a/convert/torch.go b/convert/torch.go
index 803827ba..cb8d74b0 100644
--- a/convert/torch.go
+++ b/convert/torch.go
@@ -34,18 +34,13 @@ func (tf *TorchFormat) GetTensors(dirpath string, params *Params) ([]llm.Tensor,
 	slog.Debug("getting torch tensors")
 
 	var files []string
-	var err error
-	files, err = filepath.Glob(filepath.Join(dirpath, "consolidated.*.pth"))
-	if err != nil {
-		files, err = filepath.Glob(filepath.Join(dirpath, "pytorch_model-*.bin"))
-		if err != nil {
-			slog.Error("didn't find any torch files")
-			return nil, err
-		}
+	if pt, _ := filepath.Glob(filepath.Join(dirpath, "consolidated*.pth")); len(pt) > 0 {
+		files = append(files, pt...)
+	} else if pt, _ := filepath.Glob(filepath.Join(dirpath, "pytorch_model*.pth")); len(pt) > 0 {
+		files = append(files, pt...)
 	}
 
 	var offset uint64
-
 	var tensors []llm.Tensor
 	for _, fn := range files {
 		m, err := pytorch.Load(fn)
diff --git a/llm/gguf.go b/llm/gguf.go
index 179b3255..eb7d7b75 100644
--- a/llm/gguf.go
+++ b/llm/gguf.go
@@ -62,16 +62,6 @@ func (c *containerGGUF) Decode(rs io.ReadSeeker) (model, error) {
 	return model, nil
 }
 
-const (
-	_ uint32 = iota
-	GGUFTokenNormal
-	GGUFTokenUnknown
-	GGUFTokenControl
-	GGUFTokenUserDefined
-	GGUFTokenUnused
-	GGUFTokenByte
-)
-
 const (
 	ggufTypeUint8 uint32 = iota
 	ggufTypeInt8