Update llama.cpp

llama_cpp/llama.py

@@ -83,6 +83,7 @@ class Llama:
         # NOTE: These parameters are likely to change in the future.
         n_ctx: int = 512,
         n_parts: int = -1,
+        n_gpu_layers: int = 0,
         seed: int = 1337,
         f16_kv: bool = True,
         logits_all: bool = False,
@@ -129,6 +130,7 @@ class Llama:
         self.params = llama_cpp.llama_context_default_params()
         self.params.n_ctx = n_ctx
         self.params.n_parts = n_parts
+        self.params.n_gpu_layers = n_gpu_layers
         self.params.seed = seed
         self.params.f16_kv = f16_kv
         self.params.logits_all = logits_all
@@ -1081,6 +1083,7 @@ class Llama:
             model_path=self.model_path,
             n_ctx=self.params.n_ctx,
             n_parts=self.params.n_parts,
+            n_gpu_layers=self.params.n_gpu_layers,
             seed=self.params.seed,
             f16_kv=self.params.f16_kv,
             logits_all=self.params.logits_all,
@@ -1100,6 +1103,7 @@ class Llama:
             model_path=state["model_path"],
             n_ctx=state["n_ctx"],
             n_parts=state["n_parts"],
+            n_gpu_layers=state["n_gpu_layers"],
             seed=state["seed"],
             f16_kv=state["f16_kv"],
             logits_all=state["logits_all"],

llama_cpp/llama_cpp.py

@@ -68,7 +68,7 @@ _lib_base_name = "llama"
 _lib = _load_shared_library(_lib_base_name)
 
 # C types
-LLAMA_FILE_VERSION = c_int(1)
+LLAMA_FILE_VERSION = c_int(2)
 LLAMA_FILE_MAGIC = b"ggjt"
 LLAMA_FILE_MAGIC_UNVERSIONED = b"ggml"
 LLAMA_SESSION_MAGIC = b"ggsn"
@@ -109,6 +109,7 @@ class llama_context_params(Structure):
     _fields_ = [
         ("n_ctx", c_int),  # text context
         ("n_parts", c_int),  # -1 for default
+        ("n_gpu_layers", c_int),  # number of layers to store in VRAM
         ("seed", c_int),  # RNG seed, 0 for random
         ("f16_kv", c_bool),  # use fp16 for KV cache
         (
@@ -135,7 +136,7 @@ LLAMA_FTYPE_MOSTLY_Q4_1 = c_int(3)  # except 1d tensors
 LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = c_int(
     4
 )  # tok_embeddings.weight and output.weight are F16
-LLAMA_FTYPE_MOSTLY_Q4_2 = c_int(5)  # except 1d tensors
+# LLAMA_FTYPE_MOSTLY_Q4_2 = c_int(5)  # except 1d tensors
 # LLAMA_FTYPE_MOSTYL_Q4_3 = c_int(6)  # except 1d tensors
 LLAMA_FTYPE_MOSTLY_Q8_0 = c_int(7)  # except 1d tensors
 LLAMA_FTYPE_MOSTLY_Q5_0 = c_int(8)  # except 1d tensors
@@ -259,9 +260,9 @@ _lib.llama_get_state_size.restype = c_size_t
 # Destination needs to have allocated enough memory.
 # Returns the number of bytes copied
 def llama_copy_state_data(
-    ctx: llama_context_p, dest  # type: Array[c_uint8]
+    ctx: llama_context_p, dst  # type: Array[c_uint8]
 ) -> int:
-    return _lib.llama_copy_state_data(ctx, dest)
+    return _lib.llama_copy_state_data(ctx, dst)
 
 
 _lib.llama_copy_state_data.argtypes = [llama_context_p, c_uint8_p]

llama_cpp/server/app.py

@@ -17,6 +17,11 @@ class Settings(BaseSettings):
         description="The path to the model to use for generating completions."
     )
     n_ctx: int = Field(default=2048, ge=1, description="The context size.")
+    n_gpu_layers: int = Field(
+        default=0,
+        ge=0,
+        description="The number of layers to put on the GPU. The rest will be on the CPU.",
+    )
     n_batch: int = Field(
         default=512, ge=1, description="The batch size to use per eval."
     )
@@ -80,6 +85,7 @@ def create_app(settings: Optional[Settings] = None):
     global llama
     llama = llama_cpp.Llama(
         model_path=settings.model,
+        n_gpu_layers=settings.n_gpu_layers,
         f16_kv=settings.f16_kv,
         use_mlock=settings.use_mlock,
         use_mmap=settings.use_mmap,

vendor/llama.cpp

@@ -1 +1 @@
-Subproject commit b608b55a3ea8e4760c617418538465449175bdb8
+Subproject commit 08737ef720f0510c7ec2aa84d7f70c691073c35d