ENH: skip 4-bit quantization for non-linux or non-cuda local deployme…

…nt (xorbitsai#264)

xinference/core/api.py

@@ -66,6 +66,11 @@ async def get_model(self, model_uid: str) -> xo.ActorRefType["ModelActor"]:
         supervisor_ref = await self._get_supervisor_ref()
         return await supervisor_ref.get_model(model_uid)
 
+    async def is_local_deployment(self) -> bool:
+        # TODO: temporary.
+        supervisor_ref = await self._get_supervisor_ref()
+        return await supervisor_ref.is_local_deployment()
+
 
 class SyncSupervisorAPI:
     def __init__(self, supervisor_address: str):
@@ -124,3 +129,11 @@ async def _get_model():
             return await supervisor_ref.get_model(model_uid)
 
         return self._isolation.call(_get_model())
+
+    def is_local_deployment(self) -> bool:
+        # TODO: temporary.
+        async def _is_local_deployment():
+            supervisor_ref = await self._get_supervisor_ref()
+            return await supervisor_ref.is_local_deployment()
+
+        return self._isolation.call(_is_local_deployment())

xinference/core/gradio.py

@@ -301,7 +301,11 @@ def select_model(
             progress=gr.Progress(),
         ):
             match_result = match_llm(
-                _model_name, _model_format, int(_model_size_in_billions), _quantization
+                _model_name,
+                _model_format,
+                int(_model_size_in_billions),
+                _quantization,
+                self._api.is_local_deployment(),
             )
             if not match_result:
                 raise ValueError(

xinference/core/service.py

@@ -182,6 +182,13 @@ async def list_models(self) -> Dict[str, Dict[str, Any]]:
             ret.update(await worker.list_models())
         return ret
 
+    def is_local_deployment(self) -> bool:
+        # TODO: temporary.
+        return (
+            len(self._worker_address_to_worker) == 1
+            and list(self._worker_address_to_worker)[0] == self.address
+        )
+
     @log
     async def add_worker(self, worker_address: str):
         assert worker_address not in self._worker_address_to_worker
@@ -290,8 +297,13 @@ async def launch_builtin_model(
 
         from ..model.llm import match_llm, match_llm_cls
 
+        assert self._supervisor_ref is not None
         match_result = match_llm(
-            model_name, model_format, model_size_in_billions, quantization
+            model_name,
+            model_format,
+            model_size_in_billions,
+            quantization,
+            await self._supervisor_ref.is_local_deployment(),
         )
         if not match_result:
             raise ValueError(

xinference/model/llm/__init__.py

@@ -13,7 +13,9 @@
 # limitations under the License.
 
 import json
+import logging
 import os
+import platform
 from typing import List, Optional, Tuple, Type
 
 from .core import LLM
@@ -29,12 +31,29 @@
 
 LLM_FAMILIES: List["LLMFamilyV1"] = []
 
+logger = logging.getLogger(__name__)
+
+
+def _is_linux():
+    return platform.system() == "Linux"
+
+
+def _has_cuda_device():
+    cuda_visible_devices = os.environ.get("CUDA_VISIBLE_DEVICES")
+    if cuda_visible_devices:
+        return True
+    else:
+        from xorbits._mars.resource import cuda_count
+
+        return cuda_count() > 0
+
 
 def match_llm(
     model_name: str,
     model_format: Optional[str] = None,
     model_size_in_billions: Optional[int] = None,
     quantization: Optional[str] = None,
+    is_local_deployment: bool = False,
 ) -> Optional[Tuple[LLMFamilyV1, LLMSpecV1, str]]:
     """
     Find an LLM family, spec, and quantization that satisfy given criteria.
@@ -52,8 +71,25 @@ def match_llm(
                 and quantization not in spec.quantizations
             ):
                 continue
-            # by default, choose the most coarse-grained quantization.
-            return family, spec, quantization or spec.quantizations[0]
+            if quantization:
+                return family, spec, quantization
+            else:
+                # by default, choose the most coarse-grained quantization.
+                # TODO: too hacky.
+                quantizations = spec.quantizations
+                quantizations.sort()
+                for q in quantizations:
+                    if (
+                        is_local_deployment
+                        and not (_is_linux() and _has_cuda_device())
+                        and q == "4-bit"
+                    ):
+                        logger.warning(
+                            "Skipping %s for non-linux or non-cuda local deployment .",
+                            q,
+                        )
+                        continue
+                    return family, spec, q
     return None