add model registration and other boilerplate

Author: nihalgeorge01

python/mlc_llm/conversation_template/qwen2.py

@@ -18,3 +18,17 @@
         stop_token_ids=[151643, 151645],
     )
 )
+
+ConvTemplateRegistry.register_conv_template(
+    Conversation(
+        name="qwen2-vl",
+        system_template=f"<|im_start|>system\n{MessagePlaceholders.SYSTEM.value}<|im_end|>\n",
+        system_message="You are a helpful assistant.",
+        roles={"user": "<|im_start|>user", "assistant": "<|im_start|>assistant"},
+        seps=["<|im_end|>\n"],
+        role_content_sep="\n",
+        role_empty_sep="\n",
+        stop_str=["<|endoftext|>", "<|im_end|>"],
+        stop_token_ids=[151643, 151645],
+    )
+)

python/mlc_llm/interface/gen_config.py

@@ -267,6 +267,7 @@ def gen_config(  # pylint: disable=too-many-locals,too-many-arguments,too-many-b
     "chatml",
     "chatml_nosystem",
     "qwen2",
+    "qwen2-vl",
     "open_hermes_mistral",
     "neural_hermes_mistral",
     "llama_default",

python/mlc_llm/model/model.py

@@ -38,6 +38,7 @@
 from .qwen import qwen_loader, qwen_model, qwen_quantization
 from .qwen2 import qwen2_loader, qwen2_model, qwen2_quantization
 from .qwen2_moe import qwen2_moe_loader, qwen2_moe_model, qwen2_moe_quantization
+from .qwen2_vl import qwen2_vl_loader, qwen2_vl_model, qwen2_vl_quantization
 from .rwkv5 import rwkv5_loader, rwkv5_model, rwkv5_quantization
 from .rwkv6 import rwkv6_loader, rwkv6_model, rwkv6_quantization
 from .stable_lm import stablelm_loader, stablelm_model, stablelm_quantization
@@ -299,6 +300,20 @@ class Model:
             "ft-quant": qwen2_moe_quantization.ft_quant,
         },
     ),
+    "qwen2_vl": Model(
+        name="qwen2_vl",
+        model=qwen2_vl_model.QWen2VLLMHeadModel,
+        config=qwen2_vl_model.QWen2VLConfig,
+        source={
+            "huggingface-torch": qwen2_vl_loader.huggingface,
+            "huggingface-safetensor": qwen2_vl_loader.huggingface,
+        },
+        quantize={
+            "no-quant": qwen2_vl_quantization.no_quant,
+            "group-quant": qwen2_vl_quantization.group_quant,
+            "ft-quant": qwen2_vl_quantization.ft_quant,
+        },
+    ),
     "deepseek_v2": Model(
         name="deepseek_v2",
         model=deepseek_v2_model.DeepseekV2ForCausalLM,

python/mlc_llm/model/qwen2_vl/__init__.py

@@ -84,7 +84,7 @@ def get_output_image_size(image: Tensor):
                 assert False, "not supported resize parameter"
 
         (new_h, new_w) = get_output_image_size(image)
-        out = op.interpolate(image, (new_h, new_w), data_layout="NCHW", mode="linear")
+        out = op.interpolate(image, (new_h, new_w), data_layout="NCHW", mode="bicubic")
         return out
 
     # pylint: disable=too-many-arguments,too-many-locals