Add HunYuan VL-MoT (hunyuan_vl_mot) VLM model support (#291)

## Summary

Add support for **HunYuan VL-MoT** (HY-Embodied-0.5-X) vision-language
model.

**Model**: tencent/HY-Embodied-0.5-X (3.8B params BF16, ~7.5GB)
**Registry key**: hunyuan_vl_mot
**Task**: Custom HunYuanVLMoTTask (3-model VLM split: decoder +
vision_encoder + embedding)

## Architecture

**Decoder** (32 layers, 2048 hidden, 16Q/4KV heads, head_dim=128):
- MoT dual-pathway: separate Q/K/V/O for text vs vision tokens, merged
before attention
- Shared QK norms, YaRN RoPE with alpha/beta scaling, input_ids for
modality mask

**Vision Encoder** (27-block ViT, 1152 hidden, 16 heads):
- Fused QKV + bias, LayerNorm + bias, patch_size=16, spatial merger (2x2
-> MLP)
- 16384 position embeddings (128x128 grid)

**Embedding**: Token lookup + CumSum-based image feature scatter
(mask_init_id=12)

## Config Extraction
HF config.json has no vision_config sub-object (flat config). Vision
params hardcoded in _configs.py (like phi4mm).

## Test Results

### L1 - Graph builds: PASS
```
4 passed (decoder prefill, decoder decode, vision_encoder, embedding)
```

### L2 - YAML test case: PASS
```
testdata/cases/vision-language/hy-embodied-0_5-x.yaml
222 schema tests passed
```

### CLI build with real weights: PASS
```
mobius build --model tencent/HY-Embodied-0.5-X /tmp/out
Saved decoder to decoder/model.onnx
Saved vision_encoder to vision_encoder/model.onnx
Saved embedding to embedding/model.onnx
```

### Multi-dtype: PASS
```
float32: 3 models built
float16: 3 models built
bfloat16: 3 models built
```

### Weight loading: PASS
All weights loaded. Only 1 unmatched: decoder.model.embed_tokens.weight
(expected - tie_word_embeddings=true).

### L4/L5 Golden tests: SKIPPED
**Reason**: HF repo auto_map references configuration_hunyuan_vl_mot.py
/ modeling_hunyuan_vl_mot.py but those files do not exist in the repo.
The hunyuan_vl_mot model type is not in transformers 5.7.0.
AutoModelForImageTextToText cannot load the model for HF reference
inference. YAML test case has skip_reason documenting this.

## Files Changed

| File | Change |
|------|--------|
| src/mobius/models/hunyuan_vl_mot.py | New model (~550 lines) |
| src/mobius/tasks/_hunyuan_vl_mot.py | Custom task (adds input_ids for
MoT routing) |
| src/mobius/models/__init__.py | Export HunYuanVLMoTModel |
| src/mobius/_registry.py | Register hunyuan_vl_mot |
| src/mobius/_configs.py | Hardcoded vision config + QKV bias fallback |
| tests/_test_configs.py | Tiny config for L1 tests |
| testdata/cases/vision-language/hy-embodied-0_5-x.yaml | L4/L5 test
case (skipped) |

---------

Signed-off-by: Justin Chu <justinchu@microsoft.com>
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Author: justinchuby

src/mobius/_configs.py

@@ -602,6 +602,21 @@ def _extract_vision_config(config, parent_config, model_type: str) -> dict:
             image_token_id=getattr(config, "special_image_token_id", 200010),
         )
 
+    # HunYuan VL-MoT has a flat config.json with no vision_config sub-object.
+    # The vision encoder params are hardcoded (InternViT-style ViT).
+    if model_type == "hunyuan_vl_mot" and not vision_fields.get("hidden_size"):
+        vision_fields.update(
+            hidden_size=1152,
+            intermediate_size=4304,
+            num_hidden_layers=27,
+            num_attention_heads=16,
+            image_size=2048,
+            patch_size=16,
+            norm_eps=1e-6,
+            spatial_merge_size=2,
+            image_token_id=getattr(vision_source, "mask_init_id", 12),
+        )
+
     # InternVL2 doesn't expose image_token_id in its config — default to
     # the Qwen2 <IMG_CONTEXT> token id used by InternVL2-* models.
     parent_model_type = getattr(vision_source, "model_type", None)

src/mobius/_registry.py

@@ -53,6 +53,7 @@
     GraniteMoECausalLMModel,
     HunYuanMoEV1CausalLMModel,
     HunYuanV1DenseCausalLMModel,
+    HunYuanVLMoTModel,
     InternLM2CausalLMModel,
     LayerNormCausalLMModel,
     Llama4CausalLMModel,
@@ -473,6 +474,7 @@ def _detect_fallback_registration(hf_config) -> ModelRegistration | None:
     "glm4v_moe_text": ModelRegistration(Glm4MoECausalLMModel),
     "glm4v_text": ModelRegistration(Glm4CausalLMModel),
     "got_ocr2": ModelRegistration(LLaVAModel, task="vision-language"),
+    "hunyuan_vl_mot": ModelRegistration(HunYuanVLMoTModel, task="hunyuan-vl-mot"),
     "idefics2": ModelRegistration(LLaVAModel, task="vision-language"),
     "idefics3": ModelRegistration(LLaVAModel, task="vision-language"),
     "instructblip": ModelRegistration(LLaVAModel, task="vision-language"),
@@ -835,6 +837,7 @@ def _create_default_registry() -> ModelRegistry:
     "glm4v_moe_text": "THUDM/glm-4v-9b",
     "glm4v_text": "THUDM/glm-4v-9b",
     "got_ocr2": "stepfun-ai/GOT-OCR2_0",
+    "hunyuan_vl_mot": "tencent/HY-Embodied-0.5-X",
     "instructblipvideo": "Salesforce/instructblip-flan-t5-xl",
     "internvl": "OpenGVLab/InternVL2-1B",
     "internvl_chat": "OpenGVLab/InternVL-Chat-V1-5",

src/mobius/models/__init__.py

@@ -55,6 +55,7 @@
     "GraniteMoeHybridCausalLMModel",
     "HunYuanMoEV1CausalLMModel",
     "HunYuanV1DenseCausalLMModel",
+    "HunYuanVLMoTModel",
     "HunyuanDiT2DModel",
     "IPAdapterModel",
     "InternLM2CausalLMModel",
@@ -172,6 +173,7 @@
 from mobius.models.granitemoehybrid import GraniteMoeHybridCausalLMModel
 from mobius.models.hunyuan_dit import HunyuanDiT2DModel
 from mobius.models.hunyuan_v1 import HunYuanV1DenseCausalLMModel
+from mobius.models.hunyuan_vl_mot import HunYuanVLMoTModel
 from mobius.models.internlm import InternLM2CausalLMModel
 from mobius.models.internvl import InternVL2Model
 from mobius.models.jamba import JambaCausalLMModel