janus: image generation

Author: foldl

README.md

@@ -31,6 +31,7 @@ LittleAcademia[<a href="https://github.com/foldl/little-academia"   style="text-
 
 **What's New:**
 
+* 2025-10-10: [I can draw](./docs/multimodal.md): Janus-Pro
 * 2025-09-23: Qwen2.5-VL
 * 2025-09-15: Ling/Ring-mini-2.0
 * 2025-09-08: GroveMoE

convert.py

@@ -7730,6 +7730,17 @@ def dump_config(f, config, ggml_type):
         assert config.vision_config['params']['model_name'] == 'siglip_large_patch16_384'
         JanusConverter.lang_config = AttributeDict(config.language_config)
 
+        DEF = {
+            "hidden_size": 4096,
+            "intermediate_size": 11008,
+            "num_attention_heads": 32,
+            "num_hidden_layers": 32,
+        }
+
+        for k in DEF:
+            if k not in JanusConverter.lang_config:
+                JanusConverter.lang_config[k] = DEF[k]
+
         JanusConverter.lang_config.hidden_act = 'silu'
 
         DeepSeekConverter.dump_config(f, JanusConverter.lang_config, ggml_type)
@@ -8138,6 +8149,11 @@ def main():
         config = AttributeDict({})
 
     if arch == '':
+        if config.architectures is None:
+            if "model_type" in config:
+                if config.model_type == "multi_modality":
+                    config.architectures = ['MultiModalityCausalLM']
+
         if len(config.architectures) != 1:
             raise Exception(f'unknown architectures: {config.architectures}')
         arch = config.architectures[0]

docs/models.md

@@ -353,6 +353,11 @@ Please use `--format completion` for these models.
 
     Note: Only download `tokenizer.model` and DO NOT download `tokenizer.json` when converting. Use `--set do-pan-and-scan 1` to enable _Pan and Scan_.
 
+* Janus (`MultiModalityCausalLM`)
+    * [x] Pro: [1B](https://huggingface.co/deepseek-ai/Janus-Pro-1B/tree/960ab33191f61342a4c60ae74d8dc356a39fafcb), [7B](https://huggingface.co/deepseek-ai/Janus-Pro-7B/tree/5c3eb3fb2a3b61094328465ba61fcd4272090d67)
+
+    Note: Use `--set parallel-size N` to generate `N` images in a single run (default: 2); `--set gen-head-temperature T` to set temperature of `gen-head` (default: 1.0). Add prefix "/gen " to prompts to generate images.
+
 * Kimi (`KimiVLForConditionalGeneration`)
     * [x] VL: [A3B-Instruct](https://huggingface.co/moonshotai/Kimi-VL-A3B-Instruct/tree/7a3c132a7b0f1f1677f5a72f258bd3afded7d357), [A3B-Thinking](https://huggingface.co/moonshotai/Kimi-VL-A3B-Thinking/commit/16681d8ac24e505088698e4e34ea494dd6e24400), [A3B-Thinking-2506](https://huggingface.co/moonshotai/Kimi-VL-A3B-Thinking-2506/tree/f124f44fb6ab5778cfac5117e3902ef03e860ad4)