fix: stop silently swallowing unknown Llama kwargs; --rpc_servers is a no-op (#2210)

llama_cpp/llama.py

@@ -198,6 +198,21 @@ def __init__(
         Returns:
             A Llama instance.
         """
+        if "embeddings" in kwargs:
+            # `embeddings` (plural) is the spelling used by llama.cpp's context
+            # params and by `Llama.__getstate__`, so it is a natural thing to
+            # pass here. Accept it as an alias instead of dropping it.
+            embedding = bool(kwargs.pop("embeddings"))
+
+        if kwargs:
+            warnings.warn(
+                "Llama.__init__ got unexpected keyword argument(s): "
+                f"{', '.join(sorted(kwargs))}. They are ignored. Pass only "
+                "arguments that appear in the Llama.__init__ signature.",
+                UserWarning,
+                stacklevel=2,
+            )
+
         self.verbose = verbose
         self._stack = contextlib.ExitStack()
 

llama_cpp/server/model.py

@@ -1,6 +1,7 @@
 from __future__ import annotations
 
 import json
+import warnings
 
 from typing import Dict, Optional, Union, List
 
@@ -252,6 +253,24 @@ def load_llama_from_model_settings(settings: ModelSettings) -> llama_cpp.Llama:
 
         import functools
 
+        # `rpc_servers` and `mul_mat_q` are not parameters of Llama.__init__:
+        # llama_model_params no longer carries rpc_servers and mul_mat_q is long
+        # gone upstream. Passing them here only fed Llama's **kwargs sink, so a
+        # user who set them got a silent no-op. Say so instead.
+        if settings.rpc_servers:
+            warnings.warn(
+                "The `rpc_servers` server setting is no longer supported by "
+                "llama.cpp's model params and has no effect.",
+                UserWarning,
+                stacklevel=2,
+            )
+        if not settings.mul_mat_q:
+            warnings.warn(
+                "The `mul_mat_q` server setting is obsolete and has no effect.",
+                UserWarning,
+                stacklevel=2,
+            )
+
         kwargs = {}
 
         if settings.hf_model_repo_id is not None:
@@ -275,7 +294,6 @@ def load_llama_from_model_settings(settings: ModelSettings) -> llama_cpp.Llama:
             use_mmap=settings.use_mmap,
             use_mlock=settings.use_mlock,
             kv_overrides=kv_overrides,
-            rpc_servers=settings.rpc_servers,
             # Context Params
             seed=settings.seed,
             n_ctx=settings.n_ctx,
@@ -291,7 +309,6 @@ def load_llama_from_model_settings(settings: ModelSettings) -> llama_cpp.Llama:
             yarn_beta_fast=settings.yarn_beta_fast,
             yarn_beta_slow=settings.yarn_beta_slow,
             yarn_orig_ctx=settings.yarn_orig_ctx,
-            mul_mat_q=settings.mul_mat_q,
             logits_all=settings.logits_all,
             embedding=settings.embedding,
             offload_kqv=settings.offload_kqv,

llama_cpp/server/settings.py

@@ -60,7 +60,7 @@ class ModelSettings(BaseSettings):
     )
     rpc_servers: Optional[str] = Field(
         default=None,
-        description="comma separated list of rpc servers for offloading",
+        description="Deprecated and ignored: llama.cpp's model params no longer accept a list of rpc servers.",
     )
     # Context Params
     seed: int = Field(
@@ -96,7 +96,8 @@ class ModelSettings(BaseSettings):
     yarn_beta_slow: float = Field(default=1.0)
     yarn_orig_ctx: int = Field(default=0)
     mul_mat_q: bool = Field(
-        default=True, description="if true, use experimental mul_mat_q kernels"
+        default=True,
+        description="Deprecated and ignored: the experimental mul_mat_q kernels were removed upstream.",
     )
     logits_all: bool = Field(default=True, description="Whether to return logits.")
     embedding: bool = Field(default=False, description="Whether to use embeddings.")