convert: import support for command-r models from safetensors (#6063)

---------
Co-authored-by: Patrick Devine <patrick@infrahq.com>

convert/convert.go

@@ -191,6 +191,8 @@ func ConvertModel(fsys fs.FS, ws io.WriteSeeker) error {
 		conv = &qwen2Model{}
 	case "BertModel":
 		conv = &bertModel{}
+	case "CohereForCausalLM":
+		conv = &commandrModel{}
 	default:
 		return errors.New("unsupported architecture")
 	}

convert/convert_commandr.go

+package convert
+
+import (
+	"cmp"
+
+	"github.com/ollama/ollama/llm"
+)
+
+type commandrModel 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"`
+	LayerNormEPS          float32 `json:"layer_norm_eps"`
+	RopeTheta             float32 `json:"rope_theta"`
+	UseQKNorm             bool    `json:"use_qk_norm"`
+	MaxLength             uint32  `json:"model_max_length"`
+	LogitScale            float32 `json:"logit_scale"`
+	NCtx                  uint32  `json:"n_ctx"`
+}
+
+var _ ModelConverter = (*commandrModel)(nil)
+
+func (p *commandrModel) KV(t *Tokenizer) llm.KV {
+	kv := p.ModelParameters.KV(t)
+	kv["general.architecture"] = "command-r"
+	kv["general.name"] = "command-r"
+	kv["command-r.context_length"] = cmp.Or(p.MaxLength, p.MaxPositionEmbeddings, p.NCtx)
+	kv["command-r.embedding_length"] = p.HiddenSize
+	kv["command-r.block_count"] = p.HiddenLayers
+	kv["command-r.feed_forward_length"] = p.IntermediateSize
+	kv["command-r.attention.head_count"] = p.NumAttentionHeads
+	kv["command-r.attention.head_count_kv"] = p.NumKeyValueHeads
+	kv["command-r.attention.layer_norm_epsilon"] = p.LayerNormEPS
+	kv["command-r.rope.freq_base"] = p.RopeTheta
+	kv["command-r.max_position_embeddings"] = cmp.Or(p.MaxLength, p.MaxPositionEmbeddings)
+	kv["command-r.logit_scale"] = p.LogitScale
+	kv["command-r.rope.scaling.type"] = "none"
+
+	return kv
+}
+
+func (p *commandrModel) 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 *commandrModel) Replacements() []string {
+	return []string{
+		"self_attn.q_norm", "attn_q_norm",
+		"self_attn.k_norm", "attn_k_norm",
+		"model.layers", "blk",
+		"input_layernorm", "attn_norm",
+		"mlp.down_proj", "ffn_down",
+		"mlp.gate_proj", "ffn_gate",
+		"mlp.up_proj", "ffn_up",
+		"self_attn.k_proj", "attn_k",
+		"self_attn.o_proj", "attn_output",
+		"self_attn.q_proj", "attn_q",
+		"self_attn.v_proj", "attn_v",
+		"model.norm", "output_norm",
+		"model.embed_tokens", "token_embd",
+	}
+}

convert/convert_test.go

@@ -109,6 +109,7 @@ func TestConvertModel(t *testing.T) {
 		"all-MiniLM-L6-v2",
 		"gemma-2-9b-it",
 		"Qwen2.5-0.5B-Instruct",
+		"c4ai-command-r-v01",
 	}
 
 	for i := range cases {

template/command-r.gotmpl

+{{- if or .Tools .System }}<|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|>
+{{- if .Tools }}# Safety Preamble
+The instructions in this section override those in the task description and style guide sections. Don't answer questions that are harmful or immoral.
+
+# System Preamble
+## Basic Rules
+You are a powerful conversational AI trained by Cohere to help people. You are augmented by a number of tools, and your job is to use and consume the output of these tools to best help the user. You will see a conversation history between yourself and a user, ending with an utterance from the user. You will then see a specific instruction instructing you what kind of response to generate. When you answer the user's requests, you cite your sources in your answers, according to those instructions.
+
+{{ if .System }}# User Preamble
+{{ .System }}
+{{- end }}
+
+## Available Tools
+Here is a list of tools that you have available to you:
+{{- range .Tools }}
+
+```python
+def {{ .Function.Name }}(
+{{- range $name, $property := .Function.Parameters.Properties }}{{ $name }}: {{ $property.Type }}, {{ end }}) -> List[Dict]:
+    '''{{ .Function.Description }}
+
+{{- if .Function.Parameters.Properties }}
+
+    Args:
+{{- range $name, $property := .Function.Parameters.Properties }}
+        {{ $name }} ({{ $property.Type }}): {{ $property.Description }}
+{{- end }}
+{{- end }}
+    '''
+    pass
+```
+{{- end }}
+{{- else if .System }}{{ .System }}
+{{- end }}<|END_OF_TURN_TOKEN|>
+{{- end }}
+{{- range .Messages }}
+{{- if eq .Role "system" }}
+{{- continue }}
+{{- end }}<|START_OF_TURN_TOKEN|>
+{{- if eq .Role "user" }}<|USER_TOKEN|>{{ .Content }}
+{{- if $.Tools }}<|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|>Write 'Action:' followed by a json-formatted list of actions that you want to perform in order to produce a good response to the user's last input. You can use any of the supplied tools any number of times, but you should aim to execute the minimum number of necessary actions for the input. You should use the `directly-answer` tool if calling the other tools is unnecessary. The list of actions you want to call should be formatted as a list of json objects, for example:
+```json
+[
+    {
+        "tool_name": title of the tool in the specification,
+        "parameters": a dict of parameters to input into the tool as they are defined in the specs, or {} if it takes no parameters
+    }
+]```
+{{- end }}
+{{- else if eq .Role "assistant" }}<|CHATBOT_TOKEN|>
+{{- if .Content }}{{ .Content }}
+{{- else if .ToolCalls }}
+Action: ```json
+[
+{{- range .ToolCalls }}
+    {
+        "tool_name": "{{ .Function.Name }}",
+        "parameters": {{ .Function.Arguments }}
+    }
+{{- end }}
+]```
+{{- end }}
+{{- else if eq .Role "tool" }}<|SYSTEM_TOKEN|><results>
+console_output: {{ .Content }}
+</results>
+{{- end }}<|END_OF_TURN_TOKEN|>
+{{- end }}<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>

template/command-r.json

+{
+  "stop": [
+    "<|START_OF_TURN_TOKEN|>",
+    "<|END_OF_TURN_TOKEN|>"
+  ]
+}

template/index.json

@@ -138,5 +138,9 @@
   {
     "template": "{% for message in messages %}{% if message['role'] == 'system' %}{% if message['content']%}{{'### System:\n' + message['content']+'\n\n'}}{% endif %}{% elif message['role'] == 'user' %}{{'### User:\n' + message['content']+'\n\n'}}{% elif message['role'] == 'assistant' %}{{'### Assistant:\n'  + message['content']}}{% endif %}{% if loop.last and add_generation_prompt %}{{ '### Assistant:\n' }}{% endif %}{% endfor %}",
     "name": "solar-instruct"
+  },
+  {
+    "template": "{{ bos_token }}{% if messages[0]['role'] == 'system' %}{% set loop_messages = messages[1:] %}{% set system_message = messages[0]['content'] %}{% elif false == true %}{% set loop_messages = messages %}{% set system_message = 'You are Command-R, a brilliant, sophisticated, AI-assistant trained to assist human users by providing thorough responses. You are trained by Cohere.' %}{% else %}{% set loop_messages = messages %}{% set system_message = false %}{% endif %}{% if system_message != false %}{{ '<|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|>' + system_message + '<|END_OF_TURN_TOKEN|>' }}{% endif %}{% for message in loop_messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% set content = message['content'] %}{% if message['role'] == 'user' %}{{ '<|START_OF_TURN_TOKEN|><|USER_TOKEN|>' + content.strip() + '<|END_OF_TURN_TOKEN|>' }}{% elif message['role'] == 'assistant' %}{{ '<|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>'  + content.strip() + '<|END_OF_TURN_TOKEN|>' }}{% endif %}{% endfor %}{% if add_generation_prompt %}{{ '<|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>' }}{% endif %}",
+    "name": "command-r"
   }
 ]

template/testdata/command-r.gotmpl/system-user-assistant-user

+<|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|>You are a helpful assistant.<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|USER_TOKEN|>Hello, how are you?<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>I'm doing great. How can I help you today?<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|USER_TOKEN|>I'd like to show off how chat templating works!<|END_OF_TURN_TOKEN|><|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>

template/testdata/command-r.gotmpl/user

+<|START_OF_TURN_TOKEN|><|USER_TOKEN|>Hello, how are you?<|END_OF_TURN_TOKEN|><|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>

template/testdata/command-r.gotmpl/user-assistant-user

+<|START_OF_TURN_TOKEN|><|USER_TOKEN|>Hello, how are you?<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>I'm doing great. How can I help you today?<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|USER_TOKEN|>I'd like to show off how chat templating works!<|END_OF_TURN_TOKEN|><|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>