From d2c554b69db656b0156c97d3bc976980f44c68a2 Mon Sep 17 00:00:00 2001
From: Zuellni <123005779+Zuellni@users.noreply.github.com>
Date: Wed, 25 Oct 2023 19:53:44 +0200
Subject: [PATCH] Add loras, 8bit cache, fix random seed, unloading
---
README.md | 5 +-
exllama.py | 219 +++++++++++++++++++++++++++++++----------------
requirements.txt | 3 +-
3 files changed, 152 insertions(+), 75 deletions(-)
diff --git a/README.md b/README.md
index 419f90b..8e9dd49 100644
--- a/README.md
+++ b/README.md
@@ -13,8 +13,9 @@ If you see any ExLlama-related errors while loading, install it manually followi
## Nodes
Name | Description
:--- | :---
-Loader | Used to load EXL2/GPTQ Llama models. You can find a lot of them on [Hugging Face](https://huggingface.co/TheBloke). Clone the model repository or download all the files in it and place them in an empty directory, then specify the path in `model_dir`. The `model.safetensors` file won't work on its own.
ExLlama allocates memory based on `max_seq_len`. Lowering it is a good way to save on VRAM. It's currently not possible to offload the model to RAM.
-Generator | Generates a `string` based on the given input for use with other nodes. Default values correspond to the `simple-1` preset from [text-generation-webui](https://github.com/oobabooga/text-generation-webui).
ExLlama isn't deterministic, so the outputs may differ even with the same seed.
+Model | Used to load EXL2/GPTQ Llama models. You can find a lot of them on [Hugging Face](https://huggingface.co/TheBloke). Clone the model repository or download all the files in it and place them in an empty directory, then specify the path in `model_dir`. The `model.safetensors` file won't work on its own.
ExLlama allocates memory based on `max_seq_len`. Lowering it is a good way to save on VRAM.
+LoRA | Used to load LoRAs. The directory should contain `adapter_model.bin` or `.safetensors` and `adapter_config.json`. LoRA parameter count has to match the model.
+Generator | Generates a `string` based on the given input for use with other nodes. Default values correspond to the `simple-1` preset from [text-generation-webui](https://github.com/oobabooga/text-generation-webui).
Condition | Checks if the input meets some condition, interrupts processing otherwise.
Format | Replaces variables enclosed in brackets, such as `[a]`, with their values.
Preview | Displays generated outputs in the UI.
diff --git a/exllama.py b/exllama.py
index 93a0e9c..a76ee6b 100644
--- a/exllama.py
+++ b/exllama.py
@@ -1,14 +1,21 @@
-from gc import collect
+import gc
+import random
from time import time
import torch
from comfy.model_management import soft_empty_cache
from comfy.utils import ProgressBar
-from exllamav2 import ExLlamaV2, ExLlamaV2Cache, ExLlamaV2Config, ExLlamaV2Tokenizer
+from exllamav2 import (
+ ExLlamaV2,
+ ExLlamaV2Cache_8bit,
+ ExLlamaV2Config,
+ ExLlamaV2Lora,
+ ExLlamaV2Tokenizer,
+)
from exllamav2.generator import ExLlamaV2Sampler, ExLlamaV2StreamingGenerator
-class Loader:
+class Model:
@classmethod
def INPUT_TYPES(cls):
return {
@@ -18,30 +25,85 @@ class Loader:
},
}
+ CATEGORY = "Zuellni/ExLlama"
+ FUNCTION = "prepare"
+ RETURN_NAMES = ("MODEL",)
+ RETURN_TYPES = ("EXL_MODEL",)
+
+ def __init__(self):
+ self.config = None
+ self.base = None
+ self.cache = None
+ self.tokenizer = None
+ self.generator = None
+
+ def prepare(self, model_dir, max_seq_len):
+ self.unload()
+
+ self.config = ExLlamaV2Config()
+ self.config.model_dir = model_dir
+ self.config.prepare()
+
+ if max_seq_len:
+ self.config.max_seq_len = max_seq_len
+
+ self.load()
+
+ return ((self, []),)
+
+ def load(self):
+ if not self.base:
+ self.base = ExLlamaV2(self.config)
+ self.base.load()
+
+ self.cache = ExLlamaV2Cache_8bit(self.base)
+ self.tokenizer = ExLlamaV2Tokenizer(self.config)
+
+ self.generator = ExLlamaV2StreamingGenerator(
+ self.base,
+ self.cache,
+ self.tokenizer,
+ )
+
+ return self.base
+
+ def unload(self):
+ if self.base:
+ self.base.unload()
+
+ del self.base, self.cache, self.tokenizer, self.generator
+ gc.collect()
+ soft_empty_cache()
+
+ self.base = None
+ self.cache = None
+ self.tokenizer = None
+ self.generator = None
+
+
+class Lora:
+ @classmethod
+ def INPUT_TYPES(cls):
+ return {
+ "required": {
+ "model": ("EXL_MODEL",),
+ "lora_dir": ("STRING", {"default": ""}),
+ },
+ }
+
CATEGORY = "Zuellni/ExLlama"
FUNCTION = "load"
RETURN_NAMES = ("MODEL",)
RETURN_TYPES = ("EXL_MODEL",)
- def load(self, model_dir, max_seq_len):
- collect()
- soft_empty_cache()
+ def load(self, model, lora_dir):
+ model, loras = model
- config = ExLlamaV2Config()
- config.model_dir = model_dir
- config.prepare()
+ lora = ExLlamaV2Lora.from_directory(model.load(), lora_dir)
+ loras = loras.copy()
+ loras.append(lora)
- if max_seq_len:
- config.max_seq_len = max_seq_len
-
- model = ExLlamaV2(config)
- model.load()
-
- cache = ExLlamaV2Cache(model)
- tokenizer = ExLlamaV2Tokenizer(config)
- generator = ExLlamaV2StreamingGenerator(model, cache, tokenizer)
-
- return ((tokenizer, generator),)
+ return ((model, loras),)
class Generator:
@@ -57,8 +119,10 @@ class Generator:
"typical_p": ("FLOAT", {"default": 1, "max": 1, "step": 0.01}),
"penalty": ("FLOAT", {"default": 1.15, "min": 1, "max": 2, "step": 0.01}),
"seed": ("INT", {"max": 2**64 - 1}),
+ "unload": ("BOOLEAN", {"default": False}),
"stop_on_newline": ("BOOLEAN", {"default": False}),
- "allowed_strings": ("STRING", {"default": ""}),
+ "allow_strings": ("BOOLEAN", {"default": False}),
+ "strings": ("STRING", {"default": ""}),
"text": ("STRING", {"multiline": True}),
},
"hidden": {
@@ -72,6 +136,34 @@ class Generator:
RETURN_NAMES = ("TEXT",)
RETURN_TYPES = ("STRING",)
+ def format(self, strings):
+ list = []
+
+ for string in strings.split(","):
+ if "-" in string:
+ start, end = string.split("-")
+
+ if start.isdigit() and end.isdigit():
+ start, end = int(start), int(end)
+
+ if start <= end:
+ list.extend(map(str, range(start, end + 1)))
+ else:
+ list.extend(map(str, range(start, end - 1, -1)))
+ elif len(start) == 1 and len(end) == 1:
+ start, end = ord(start), ord(end)
+
+ if start <= end:
+ list.extend(map(chr, range(start, end + 1)))
+ else:
+ list.extend(map(chr, range(start, end + -1, -1)))
+ else:
+ list.append(string)
+ else:
+ list.append(string)
+
+ return list
+
def generate(
self,
model,
@@ -82,26 +174,28 @@ class Generator:
typical_p,
penalty,
seed,
+ unload,
stop_on_newline,
- allowed_strings,
+ allow_strings,
+ strings,
text,
info=None,
id=None,
):
- text = text.strip()
-
if not text:
return ("",)
- tokenizer, generator = model
- text = tokenizer.encode(text)
- stop_conditions = [tokenizer.eos_token_id]
+ model, loras = model
+
+ model.load()
+ text = model.tokenizer.encode(text)
+ stop_conditions = [model.tokenizer.eos_token_id]
if not max_new_tokens:
- max_new_tokens = tokenizer.config.max_seq_len - text.shape[-1]
+ max_new_tokens = model.config.max_seq_len - text.shape[-1]
if stop_on_newline:
- stop_conditions.append(tokenizer.newline_token_id)
+ stop_conditions.append(model.tokenizer.newline_token_id)
settings = ExLlamaV2Sampler.Settings()
settings.temperature = temperature
@@ -110,47 +204,23 @@ class Generator:
settings.typical = typical_p
settings.token_repetition_penalty = penalty
- if allowed_strings:
- strings = []
-
- for string in allowed_strings.split(","):
- string = string.strip()
-
- if "-" in string:
- start, end = string.split("-")
-
- if start.isdigit() and end.isdigit():
- start, end = int(start), int(end)
-
- if start <= end:
- strings.extend(map(str, range(start, end + 1)))
- else:
- strings.extend(map(str, range(start, end - 1, -1)))
- elif len(start) == 1 and len(end) == 1:
- start, end = ord(start), ord(end)
-
- if start <= end:
- strings.extend(map(chr, range(start, end + 1)))
- else:
- strings.extend(map(chr, range(start, end + -1, -1)))
- else:
- strings.append(string)
- else:
- strings.append(string)
-
- allowed_strings = strings
- allowed_tokens = tokenizer.encode(allowed_strings)
- max_new_tokens = allowed_tokens.shape[-1]
-
- vocab_size = tokenizer.config.vocab_size
+ if strings:
+ strings = self.format(strings)
+ tokens = model.tokenizer.encode(strings)
+ vocab_size = model.config.vocab_size
padding = vocab_size + (-vocab_size % 32)
- settings.token_bias = torch.full((padding,), float("-inf"))
- settings.token_bias[allowed_tokens] = 0
+ if allow_strings:
+ settings.token_bias = torch.full((padding,), float("-inf"))
+ settings.token_bias[tokens] = 0
+ max_new_tokens = tokens.shape[-1]
+ else:
+ settings.token_bias = torch.zeros((padding,))
+ settings.token_bias[tokens] = float("-inf")
- torch.manual_seed(seed)
- generator.set_stop_conditions(stop_conditions)
- generator.begin_stream(text, settings)
+ random.seed(seed)
+ model.generator.set_stop_conditions(stop_conditions)
+ model.generator.begin_stream(text, settings, loras=loras)
progress = ProgressBar(max_new_tokens)
start = time()
eos = False
@@ -158,12 +228,12 @@ class Generator:
tokens = 0
while not eos and tokens < max_new_tokens:
- chunk, eos, _ = generator.stream()
+ chunk, eos, _ = model.generator.stream()
- if allowed_strings:
+ if strings and allow_strings:
c = (output + chunk).strip()
- if not any(c in s for s in allowed_strings):
+ if not any(c in s for s in strings):
break
progress.update(1)
@@ -175,6 +245,9 @@ class Generator:
speed = round(tokens / total, 2)
print(f"Output generated in {total} seconds ({tokens} tokens, {speed} tokens/s)")
+ if unload:
+ model.unload()
+
if id and info and "workflow" in info:
nodes = info["workflow"]["nodes"]
node = next((n for n in nodes if str(n["id"]) == id), None)
@@ -186,11 +259,13 @@ class Generator:
NODE_CLASS_MAPPINGS = {
- "ZuellniExLlamaLoader": Loader,
+ "ZuellniExLlamaModel": Model,
+ "ZuellniExLlamaLora": Lora,
"ZuellniExLlamaGenerator": Generator,
}
NODE_DISPLAY_NAME_MAPPINGS = {
- "ZuellniExLlamaLoader": "Loader",
+ "ZuellniExLlamaModel": "Model",
+ "ZuellniExLlamaLora": "LoRA",
"ZuellniExLlamaGenerator": "Generator",
}
diff --git a/requirements.txt b/requirements.txt
index 7aea264..3815689 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -1 +1,2 @@
-exllamav2
+exllamav2>=0.0.7; platform_system == "Linux"
+https://github.com/turboderp/exllamav2/releases/download/v0.0.7/exllamav2-0.0.7+cu121-cp311-cp311-win_amd64.whl; platform_system == "Windows"