baalajimaestro/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

Add temporary rms_norm_eps parameter

Browse source
This commit is contained in:
Andrei Betlen 2023-07-24 14:09:24 -04:00
parent 8cd64d4ac3
commit 11dd2bf382
2 changed files with 17 additions and 7 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

14
llama_cpp/llama.py
View file

@ -224,7 +224,7 @@ class Llama:
rope_freq_base: float = 10000.0, rope_freq_base: float = 10000.0,
rope_freq_scale: float = 1.0, rope_freq_scale: float = 1.0,
n_gqa: Optional[int] = None, # (TEMPORARY) must be 8 for llama2 70b n_gqa: Optional[int] = None, # (TEMPORARY) must be 8 for llama2 70b
rms_eps_norm: Optional[float] = None, # (TEMPORARY) rms_norm_eps: Optional[float] = None, # (TEMPORARY)
verbose: bool = True, verbose: bool = True,
): ):
"""Load a llama.cpp model from `model_path`. """Load a llama.cpp model from `model_path`.
@ -287,8 +287,8 @@ class Llama:
if n_gqa is not None: if n_gqa is not None:
self.params.n_gqa = n_gqa self.params.n_gqa = n_gqa
if rms_eps_norm is not None: if rms_norm_eps is not None:
self.params.rms_eps_norm = rms_eps_norm self.params.rms_norm_eps = rms_norm_eps
self.last_n_tokens_size = last_n_tokens_size self.last_n_tokens_size = last_n_tokens_size
self.n_batch = min(n_ctx, n_batch) self.n_batch = min(n_ctx, n_batch)
@ -1533,7 +1533,7 @@ class Llama:
tensor_split=self.tensor_split, tensor_split=self.tensor_split,
### TEMPORARY ### ### TEMPORARY ###
n_gqa=self.params.n_gqa, n_gqa=self.params.n_gqa,
rms_eps_norm=self.params.rms_eps_norm, rms_norm_eps=self.params.rms_norm_eps,
### TEMPORARY ### ### TEMPORARY ###
### DEPRECATED ### ### DEPRECATED ###
n_parts=self.n_parts, n_parts=self.n_parts,
@ -1559,11 +1559,11 @@ class Llama:
lora_base=state["lora_base"], lora_base=state["lora_base"],
lora_path=state["lora_path"], lora_path=state["lora_path"],
tensor_split=state["tensor_split"], tensor_split=state["tensor_split"],
n_gqa=state["n_gqa"],
### TEMPORARY ###
rms_eps_norm=state["rms_eps_norm"],
verbose=state["verbose"], verbose=state["verbose"],
### TEMPORARY ### ### TEMPORARY ###
n_gqa=state["n_gqa"],
rms_norm_eps=state["rms_norm_eps"],
### TEMPORARY ###
### DEPRECATED ### ### DEPRECATED ###
n_parts=state["n_parts"], n_parts=state["n_parts"],
### DEPRECATED ### ### DEPRECATED ###

10
llama_cpp/server/app.py
View file

@ -95,6 +95,14 @@ class Settings(BaseSettings):
default=True, default=True,
description="Whether to interrupt requests when a new request is received.", description="Whether to interrupt requests when a new request is received.",
) )
n_gqa: Optional[int] = Field(
default=None,
description="TEMPORARY: Set to 8 for Llama2 70B",
)
rms_norm_eps: Optional[float] = Field(
default=None,
description="TEMPORARY",
)
class ErrorResponse(TypedDict): class ErrorResponse(TypedDict):
@ -320,6 +328,8 @@ def create_app(settings: Optional[Settings] = None):
last_n_tokens_size=settings.last_n_tokens_size, last_n_tokens_size=settings.last_n_tokens_size,
vocab_only=settings.vocab_only, vocab_only=settings.vocab_only,
verbose=settings.verbose, verbose=settings.verbose,
n_gqa=settings.n_gqa,
rms_norm_eps=settings.rms_norm_eps,
) )
if settings.cache: if settings.cache:
if settings.cache_type == "disk": if settings.cache_type == "disk":