BUG: fix custom vision model (#1280)

xorbitsai · Apr 11, 2024 · e3a947e · e3a947e
1 parent a7f0c3b
commit e3a947e
Show file tree

Hide file tree

Showing 3 changed files with 51 additions and 28 deletions.
diff --git a/xinference/model/llm/llm_family.py b/xinference/model/llm/llm_family.py
@@ -199,6 +199,21 @@ def parse_raw(
                 )
             llm_spec.prompt_style = BUILTIN_LLM_PROMPT_STYLE[prompt_style_name]
 
+        # check model ability, registering LLM only provides generate and chat
+        # but for vision models, we add back the abilities so that
+        # gradio chat interface can be generated properly
+        if (
+            llm_spec.model_family != "other"
+            and llm_spec.model_family
+            in {
+                family.model_name
+                for family in BUILTIN_LLM_FAMILIES
+                if "vision" in family.model_ability
+            }
+            and "vision" not in llm_spec.model_ability
+        ):
+            llm_spec.model_ability.append("vision")
+
         return llm_spec
 
 

diff --git a/xinference/model/llm/pytorch/core.py b/xinference/model/llm/pytorch/core.py
@@ -42,6 +42,25 @@
 
 logger = logging.getLogger(__name__)
 
+NON_DEFAULT_MODEL_LIST: List[str] = [
+    "baichuan-chat",
+    "baichuan-2-chat",
+    "vicuna-v1.3",
+    "falcon",
+    "falcon-instruct",
+    "chatglm",
+    "chatglm2",
+    "chatglm2-32k",
+    "chatglm2-128k",
+    "llama-2",
+    "llama-2-chat",
+    "internlm2-chat",
+    "qwen-vl-chat",
+    "OmniLMM",
+    "yi-vl-chat",
+    "deepseek-vl-chat",
+]
+
 
 class PytorchModel(LLM):
     def __init__(
@@ -233,17 +252,7 @@ def match(
         if llm_spec.model_format not in ["pytorch", "gptq", "awq"]:
             return False
         model_family = llm_family.model_family or llm_family.model_name
-        if model_family in [
-            "baichuan-chat",
-            "vicuna-v1.3",
-            "falcon",
-            "falcon-instruct",
-            "chatglm",
-            "chatglm2",
-            "chatglm2-32k",
-            "llama-2",
-            "llama-2-chat",
-        ]:
+        if model_family in NON_DEFAULT_MODEL_LIST:
             return False
         if "generate" not in llm_family.model_ability:
             return False
@@ -452,23 +461,8 @@ def match(
     ) -> bool:
         if llm_spec.model_format not in ["pytorch", "gptq", "awq"]:
             return False
-        if llm_family.model_name in [
-            "baichuan-chat",
-            "baichuan-2-chat",
-            "vicuna-v1.3",
-            "falcon",
-            "falcon-instruct",
-            "chatglm",
-            "chatglm2",
-            "chatglm2-32k",
-            "llama-2",
-            "llama-2-chat",
-            "internlm2-chat",
-            "qwen-vl-chat",
-            "OmniLMM",
-            "yi-vl-chat",
-            "deepseek-vl-chat",
-        ]:
+        model_family = llm_family.model_family or llm_family.model_name
+        if model_family in NON_DEFAULT_MODEL_LIST:
             return False
         if "chat" not in llm_family.model_ability:
             return False

diff --git a/xinference/model/llm/tests/test_llm_family.py b/xinference/model/llm/tests/test_llm_family.py
@@ -992,6 +992,20 @@ def test_parse_prompt_style():
     model_spec = CustomLLMFamilyV1.parse_raw(bytes(llm_family.json(), "utf8"))
     assert model_spec.model_name == llm_family.model_name
 
+    # test vision
+    llm_family = CustomLLMFamilyV1(
+        version=1,
+        model_type="LLM",
+        model_name="test_LLM",
+        model_lang=["en"],
+        model_ability=["chat", "generate"],
+        model_specs=[hf_spec, ms_spec],
+        model_family="qwen-vl-chat",
+        prompt_style="qwen-vl-chat",
+    )
+    model_spec = CustomLLMFamilyV1.parse_raw(bytes(llm_family.json(), "utf-8"))
+    assert "vision" in model_spec.model_ability
+
     # error: missing model_family
     llm_family = CustomLLMFamilyV1(
         version=1,