xinference 1.6.0.post1__py3-none-any.whl → 1.6.1__py3-none-any.whl

xinference/model/llm/transformers/utils.py

@@ -191,42 +191,13 @@ def _get_pad_param(seq_len_idx: int, pad_len: int) -> Tuple:
     return tuple(dimensions)
 
 
-def _merge_kv_cache(
-    xinf_model_obj: "PytorchModel",
-    past_cache: DynamicCache,
-    new_cache: DynamicCache,
-) -> DynamicCache:
-    from torch.nn.functional import pad
-
-    _, seq_len_idx = xinf_model_obj.get_batch_size_and_seq_len_indexes_from_kv()
-    past_seq_len = past_cache[0][0].shape[seq_len_idx]
-    new_seq_len = new_cache[0][0].shape[seq_len_idx]
-    if past_seq_len != new_seq_len:
-        padding_target = new_cache if past_seq_len > new_seq_len else past_cache
-        padding_len = abs(past_seq_len - new_seq_len)
-        pad_param = _get_pad_param(seq_len_idx, padding_len)
-        for idx in range(len(padding_target)):
-            k = padding_target.key_cache[idx]
-            v = padding_target.value_cache[idx]
-            _k = pad(k, pad_param)
-            _v = pad(v, pad_param)
-            padding_target.key_cache[idx] = _k
-            padding_target.value_cache[idx] = _v
-
-    ret_kv = DynamicCache()
-    for idx in range(len(past_cache)):
-        k1, k2 = new_cache.key_cache[idx], past_cache.key_cache[idx]
-        v1, v2 = new_cache.value_cache[idx], past_cache.value_cache[idx]
-        ret_kv.update(
-            torch.cat((k1, k2), 0).contiguous(),
-            torch.cat((v1, v2), 0).contiguous(),
-            idx,
-        )
-    return ret_kv
-
-
 def get_batch_size_and_seq_len_from_kv_cache(kv, xinf_model_obj: "PytorchModel"):
+    from transformers import HybridCache
+
     bs_idx, seq_len_idx = xinf_model_obj.get_batch_size_and_seq_len_indexes_from_kv()
+
+    if isinstance(kv, HybridCache):
+        return kv.key_cache[0].shape[bs_idx], kv.get_seq_length()
     return kv[0][0].shape[bs_idx], kv[0][0].shape[seq_len_idx] + 1
 
 
@@ -304,9 +275,7 @@ def _batch_inference_one_step_internal(
         if decode_reqs:
             decode_kv = decode_reqs[0].kv_cache
             # prefill and decode kv cache need to be merged at `batch_size` and `seq_len` dimensions.
-            merged_kv_cache = _merge_kv_cache(
-                xinf_model_obj, decode_kv, past_key_values
-            )
+            merged_kv_cache = xinf_model_obj.merge_kv_cache(decode_kv, past_key_values)
             for r in valid_req_list:
                 r.kv_cache = merged_kv_cache
             empty_cache()

xinference/model/llm/vllm/core.py

@@ -199,7 +199,11 @@ if VLLM_INSTALLED and vllm.__version__ >= "0.5.1":
     VLLM_SUPPORTED_CHAT_MODELS.append("deepseek-v2-chat-0628")
     VLLM_SUPPORTED_CHAT_MODELS.append("deepseek-v2.5")
     VLLM_SUPPORTED_CHAT_MODELS.append("deepseek-v3")
+    VLLM_SUPPORTED_CHAT_MODELS.append("deepseek-v3-0324")
     VLLM_SUPPORTED_CHAT_MODELS.append("deepseek-r1")
+    VLLM_SUPPORTED_CHAT_MODELS.append("deepseek-r1-0528")
+    VLLM_SUPPORTED_CHAT_MODELS.append("deepseek-prover-v2")
+    VLLM_SUPPORTED_CHAT_MODELS.append("deepseek-r1-0528-qwen3")
 
 if VLLM_INSTALLED and vllm.__version__ >= "0.5.3":
     VLLM_SUPPORTED_CHAT_MODELS.append("gemma-2-it")

xinference/model/rerank/core.py

@@ -265,7 +265,13 @@ class RerankModel:
         if max_chunks_per_doc is not None:
             raise ValueError("rerank hasn't support `max_chunks_per_doc` parameter.")
         logger.info("Rerank with kwargs: %s, model: %s", kwargs, self._model)
-        sentence_combinations = [[query, doc] for doc in documents]
+
+        from .utils import preprocess_sentence
+
+        pre_query = preprocess_sentence(
+            query, kwargs.get("instruction", None), self._model_spec.model_name
+        )
+        sentence_combinations = [[pre_query, doc] for doc in documents]
         # reset n tokens
         self._model.model.n_tokens = 0
         if self._model_spec.type == "normal":

xinference/model/rerank/utils.py

@@ -11,8 +11,25 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+from typing import Any
+
 from .core import RerankModelSpec
 
 
 def get_model_version(rerank_model: RerankModelSpec) -> str:
     return rerank_model.model_name
+
+
+instruction_cfg = {
+    "minicpm-reranker": "Query: ",
+}
+
+
+def preprocess_sentence(query: str, instruction: Any, model_name: str) -> str:
+    if instruction and isinstance(instruction, str):
+        return f"{instruction}{query}"
+    if instruction is None:
+        for k, v in instruction_cfg.items():
+            if k.lower() in model_name.lower():
+                return f"{v}{query}"
+    return query

xinference/web/ui/build/asset-manifest.json

@@ -1,14 +1,14 @@
 {
   "files": {
     "main.css": "./static/css/main.337afe76.css",
-    "main.js": "./static/js/main.ae579a97.js",
+    "main.js": "./static/js/main.ddf9eaee.js",
     "static/media/icon.webp": "./static/media/icon.4603d52c63041e5dfbfd.webp",
     "index.html": "./index.html",
     "main.337afe76.css.map": "./static/css/main.337afe76.css.map",
-    "main.ae579a97.js.map": "./static/js/main.ae579a97.js.map"
+    "main.ddf9eaee.js.map": "./static/js/main.ddf9eaee.js.map"
   },
   "entrypoints": [
     "static/css/main.337afe76.css",
-    "static/js/main.ae579a97.js"
+    "static/js/main.ddf9eaee.js"
   ]
 }

xinference/web/ui/build/index.html

@@ -1 +1 @@
-<!doctype html><html lang="en"><head><meta charset="utf-8"/><link rel="icon" href="./favicon.svg"/><meta name="viewport" content="width=device-width,initial-scale=1"/><meta name="theme-color" content="#000000"/><meta name="description" content="Web site created using create-react-app"/><link rel="apple-touch-icon" href="./logo192.png"/><link rel="manifest" href="./manifest.json"/><title>Xinference</title><script defer="defer" src="./static/js/main.ae579a97.js"></script><link href="./static/css/main.337afe76.css" rel="stylesheet"></head><body><noscript>You need to enable JavaScript to run this app.</noscript><div id="root"></div></body></html>
+<!doctype html><html lang="en"><head><meta charset="utf-8"/><link rel="icon" href="./favicon.svg"/><meta name="viewport" content="width=device-width,initial-scale=1"/><meta name="theme-color" content="#000000"/><meta name="description" content="Web site created using create-react-app"/><link rel="apple-touch-icon" href="./logo192.png"/><link rel="manifest" href="./manifest.json"/><title>Xinference</title><script defer="defer" src="./static/js/main.ddf9eaee.js"></script><link href="./static/css/main.337afe76.css" rel="stylesheet"></head><body><noscript>You need to enable JavaScript to run this app.</noscript><div id="root"></div></body></html>