convert: qwen2 from safetensors (#8408)

Add native support for converting Qwen2 family models (including Qwen2.5)
from safetensors to gguf format so we can run it.

convert/convert.go

@@ -187,6 +187,8 @@ func ConvertModel(fsys fs.FS, ws io.WriteSeeker) error {
 		conv = &gemma2Model{}
 	case "Phi3ForCausalLM":
 		conv = &phi3Model{}
+	case "Qwen2ForCausalLM":
+		conv = &qwen2Model{}
 	case "BertModel":
 		conv = &bertModel{}
 	default:

convert/convert_qwen2.go

+package convert
+
+import "github.com/ollama/ollama/llm"
+
+type qwen2Model struct {
+	ModelParameters
+	MaxPositionEmbeddings uint32  `json:"max_position_embeddings"`
+	HiddenSize            uint32  `json:"hidden_size"`
+	HiddenLayers          uint32  `json:"num_hidden_layers"`
+	IntermediateSize      uint32  `json:"intermediate_size"`
+	NumAttentionHeads     uint32  `json:"num_attention_heads"`
+	NumKeyValueHeads      uint32  `json:"num_key_value_heads"`
+	RopeTheta             float32 `json:"rope_theta"`
+	RopeScaling           struct {
+		Type                          string     `json:"type"`
+		Factor                        ropeFactor `json:"factor"`
+		OriginalMaxPositionEmbeddings uint32     `json:"original_max_position_embeddings"`
+	} `json:"rope_scaling"`
+	RMSNormEPS float32 `json:"rms_norm_eps"`
+}
+
+var _ ModelConverter = (*qwen2Model)(nil)
+
+func (q *qwen2Model) KV(t *Tokenizer) llm.KV {
+	kv := q.ModelParameters.KV(t)
+	kv["general.architecture"] = "qwen2"
+	kv["qwen2.block_count"] = q.HiddenLayers
+	kv["qwen2.context_length"] = q.MaxPositionEmbeddings
+	kv["qwen2.embedding_length"] = q.HiddenSize
+	kv["qwen2.feed_forward_length"] = q.IntermediateSize
+	kv["qwen2.attention.head_count"] = q.NumAttentionHeads
+	kv["qwen2.attention.head_count_kv"] = q.NumKeyValueHeads
+	kv["qwen2.rope.freq_base"] = q.RopeTheta
+	kv["qwen2.attention.layer_norm_rms_epsilon"] = q.RMSNormEPS
+
+	switch q.RopeScaling.Type {
+	case "":
+		// no scaling
+	case "yarn":
+		kv["qwen2.rope.scaling.type"] = q.RopeScaling.Type
+		kv["qwen2.rope.scaling.factor"] = q.RopeScaling.Factor
+	default:
+		panic("unknown rope scaling type")
+	}
+	return kv
+}
+
+func (q *qwen2Model) Tensors(ts []Tensor) []llm.Tensor {
+	var out []llm.Tensor
+	for _, t := range ts {
+		out = append(out, llm.Tensor{
+			Name:     t.Name(),
+			Kind:     t.Kind(),
+			Shape:    t.Shape(),
+			WriterTo: t,
+		})
+	}
+
+	return out
+}
+
+func (p *qwen2Model) Replacements() []string {
+	return []string{
+		"lm_head", "output",
+		"model.embed_tokens", "token_embd",
+		"model.layers", "blk",
+		"input_layernorm", "attn_norm",
+		"self_attn.k_proj", "attn_k",
+		"self_attn.v_proj", "attn_v",
+		"self_attn.q_proj", "attn_q",
+		"self_attn.o_proj", "attn_output",
+		"mlp.down_proj", "ffn_down",
+		"mlp.gate_proj", "ffn_gate",
+		"mlp.up_proj", "ffn_up",
+		"post_attention_layernorm", "ffn_norm",
+		"model.norm", "output_norm",
+	}
+}

convert/convert_test.go

@@ -108,6 +108,7 @@ func TestConvertModel(t *testing.T) {
 		"Phi-3-mini-128k-instruct",
 		"all-MiniLM-L6-v2",
 		"gemma-2-9b-it",
+		"Qwen2.5-0.5B-Instruct",
 	}
 
 	for i := range cases {

convert/tokenizer.go

@@ -100,6 +100,8 @@ func parseTokenizer(fsys fs.FS, specialTokenTypes []string) (*Tokenizer, error)
 			t.Pre = "deepseek-llm"
 		case "21cde974d587f0d54dc8d56b183cc1e6239600172035c68fbd6d4b9f8da0576e":
 			t.Pre = "deepseek-coder"
+		case "1ff7f41064896984db5d1bb6ff64fa4bc29007d08c1b439e505b7392777a319e":
+			t.Pre = "qwen2"
 		case "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855":
 			// noop, empty pretokenizer
 		default: