Spaces:
Sleeping
Sleeping
File size: 9,725 Bytes
5caedb4 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 |
import gc
import logging
import os
import numpy as np
import torch
from accelerate import dispatch_model, infer_auto_device_map
from accelerate.utils import get_balanced_memory
from h2o_wave import Q
from h2o_wave import data as chat_data
from h2o_wave import ui
from llm_studio.app_utils.utils import get_experiments, get_ui_elements_for_cfg, set_env
from llm_studio.python_configs.base import DefaultConfigProblemBase
from llm_studio.src.datasets.text_utils import get_texts, get_tokenizer
from llm_studio.src.utils.config_utils import (
NON_GENERATION_PROBLEM_TYPES,
load_config_yaml,
)
from llm_studio.src.utils.export_utils import get_prediction_dataframe
from llm_studio.src.utils.modeling_utils import load_checkpoint
logger = logging.getLogger(__name__)
async def chat_tab(q: Q, load_model=True):
if not await should_start_chat(q):
return
if load_model:
q.page["experiment/display/chat"] = ui.form_card(
box="first",
items=[ui.progress(label="Loading the model...")],
)
q.client["experiment/display/chat/messages"] = []
q.client.delete_cards.add("experiment/display/chat")
q.page["experiment/display/chat/settings"] = ui.form_card(
box="second",
items=[
ui.expander(
name="chat_settings",
label="Chat Settings",
items=[ui.progress(label="Loading model configuration...")],
expanded=True,
)
],
)
q.client.delete_cards.add("experiment/display/chat/settings")
await q.page.save()
logger.info(torch.cuda.memory_allocated())
if load_model:
with set_env(HF_TOKEN=q.client["default_huggingface_api_token"]):
gpu_id = q.client["gpu_used_for_chat"] - 1
cfg, model, tokenizer = load_cfg_model_tokenizer(
q.client["experiment/display/experiment_path"], device=f"cuda:{gpu_id}"
)
q.client["experiment/display/chat/cfg"] = cfg
q.client["experiment/display/chat/model"] = model
q.client["experiment/display/chat/tokenizer"] = tokenizer
initial_message = "Model successfully loaded, how can I help you?"
else:
cfg = q.client["experiment/display/chat/cfg"]
assert q.client["experiment/display/chat/model"] is not None
assert q.client["experiment/display/chat/tokenizer"] is not None
initial_message = "Chat History cleaned. How can I help you?"
# Load validation dataframe and texts
validation_dataframe = get_prediction_dataframe(cfg.output_directory)
if cfg.dataset.parent_id_column != "None":
# sample and parent ids can have any dtype, such as str, int, float, etc.
# id column can be int, while parent_id column can be float
# (as some values are NaN) so we cast id to the same dtype
sample_ids = (
validation_dataframe["id"]
.astype(validation_dataframe[cfg.dataset.parent_id_column].dtype)
.tolist()
)
parent_ids = validation_dataframe[cfg.dataset.parent_id_column].tolist()
sample_ids_set = set(sample_ids)
is_seed_prompt = [
False if idx in sample_ids_set else True for idx in parent_ids
]
validation_dataframe["is_seed_prompt"] = is_seed_prompt
validation_dataframe = validation_dataframe.loc[
validation_dataframe["is_seed_prompt"]
]
validation_texts = get_texts(validation_dataframe, cfg)
# Hide fields that are should not be visible in the UI
cfg.prediction._visibility["metric"] = -1
cfg.prediction._visibility["batch_size_inference"] = -1
cfg.prediction._visibility["min_length_inference"] = -1
cfg.prediction._visibility["stop_tokens"] = -1
logger.info(torch.cuda.memory_allocated())
q.page["experiment/display/chat"] = ui.chatbot_card(
box="first",
data=chat_data(fields="content from_user", t="list"), # type: ignore
name="experiment/display/chat/chatbot",
events=["stop", "suggestion"],
suggestions=[
ui.chat_suggestion(
"Write a poem about H2O LLM Studio",
label="Write a poem",
caption="about H2O LLM Studio",
icon="Edit",
),
ui.chat_suggestion(
"Plan a trip to Europe",
label="Plan a trip",
caption="to Europe",
icon="Airplane",
),
ui.chat_suggestion(
"Give me ideas for a new project",
label="Give me ideas",
caption="for a new project",
icon="Lightbulb",
),
ui.chat_suggestion(
np.random.choice(validation_texts),
label="Random sample from validation set",
icon="Chat",
),
],
)
q.page["experiment/display/chat"].data += [initial_message, False]
option_items = get_ui_elements_for_cfg(
cfg=q.client["experiment/display/chat/cfg"].prediction,
q=q,
pre="chat/cfg_predictions",
)
q.page["experiment/display/chat/settings"] = ui.form_card(
box="second",
items=[
ui.buttons(
[
ui.button(
name="experiment/display/chat/clear_history",
label="Clear History",
primary=True,
),
ui.button(
name="experiment/display/chat/copy_chat",
label="Copy to clipboard",
primary=True,
),
]
),
ui.expander(
name="chat_settings",
label="Chat Settings",
items=option_items,
expanded=True,
),
],
)
async def should_start_chat(q: Q):
cfg: DefaultConfigProblemBase = load_config_yaml(
os.path.join(q.client["experiment/display/experiment_path"], "cfg.yaml")
)
if cfg.problem_type in NON_GENERATION_PROBLEM_TYPES:
q.page["experiment/display/chat"] = ui.form_card(
box="first",
items=[
ui.text(
"Chatbot is not available for this problem type. "
"Please select a text generation problem."
)
],
title="",
)
q.client.delete_cards.add("experiment/display/chat")
return False
# gpu id in UI is offset by 1 to be in sync with experiment UI
gpu_id = q.client["gpu_used_for_chat"] - 1
if gpu_is_blocked(q, gpu_id):
q.page["experiment/display/chat"] = ui.form_card(
box="first",
items=[
ui.text(
f"""Chatbot is not available when GPU{q.client["gpu_used_for_chat"]}
is blocked by another experiment.
You can change "Gpu used for Chat" in the settings tab
to use another GPU for the chatbot. """
)
],
title="",
)
q.client.delete_cards.add("experiment/display/chat")
return False
return True
def gpu_is_blocked(q, gpu_id):
experiments = get_experiments(q=q)
running_experiments = experiments[experiments.status.isin(["running"])]
gpu_blocked = any(
[
str(gpu_id) in gpu_list
for gpu_list in running_experiments["gpu_list"]
.apply(lambda x: x.split(","))
.to_list()
]
)
return gpu_blocked
def load_cfg_model_tokenizer(
experiment_path: str, merge: bool = False, device: str = "cuda:0"
):
"""Loads the model, tokenizer and configuration from the experiment path."""
cfg = load_config_yaml(os.path.join(experiment_path, "cfg.yaml"))
cfg.architecture.pretrained = False
cfg.architecture.gradient_checkpointing = False
cfg.environment._device = device.replace("_shard", "")
cfg.environment._local_rank = 0
cfg.prediction._visibility["num_history"] = 1
tokenizer = get_tokenizer(cfg)
gc.collect()
torch.cuda.empty_cache()
if (
merge
and cfg.training.lora
and cfg.architecture.backbone_dtype in ("int4", "int8")
):
# Force to float16 for merging LORA weights.
# TODO: Could be configurable in the future to allow bfloat16.
logger.info("Loading backbone in float16 for merging LORA weights.")
cfg.architecture.backbone_dtype = "float16"
cfg.architecture.pretrained = True
# if "cpu" in device:
# cfg.architecture.backbone_dtype = "float32"
with torch.device(cfg.environment._device):
model = cfg.architecture.model_class(cfg)
cfg.architecture.pretrained_weights = os.path.join(
experiment_path, "checkpoint.pth"
)
load_checkpoint(cfg, model, strict=False)
if device == "cpu_shard":
max_memory = get_balanced_memory(
model,
)
device_map = infer_auto_device_map(model, max_memory=max_memory)
model = dispatch_model(
model,
device_map=device_map,
)
if merge and cfg.training.lora:
# merges the LoRa layers into the base model.
# This is needed if one wants to use the base model as a standalone model.
logger.info("Merging LORA layers with base model.")
model.backbone = model.backbone.merge_and_unload()
model = model.eval()
model.backbone.use_cache = True
return cfg, model, tokenizer
|