Spaces:
Running
on
Zero
Running
on
Zero
This includes fixes that make checkpointing and reloading work correctly. (#35)
Browse filesIt also batches in a first set of changes for fixing eval code
Summary:
Test Plan:
- apps/main/lingua_train.py +1 -1
- bytelatent/args.py +86 -1
- bytelatent/checkpoint.py +7 -2
- bytelatent/configs/debug.yaml +1 -8
- bytelatent/configs/entropy_model.yaml +1 -8
- bytelatent/data/data_types.py +0 -10
- bytelatent/data/iterators/multiprocess_iterator.py +15 -0
- {apps/main → bytelatent}/eval.py +42 -137
- {apps/main → bytelatent}/generate.py +22 -35
- bytelatent/train.py +46 -35
apps/main/lingua_train.py
CHANGED
@@ -544,7 +544,7 @@ def train(args: TrainArgs):
|
|
544 |
if args.eval is not None and every_n_steps(
|
545 |
train_state, args.checkpoint.eval.every, acc_step=0
|
546 |
):
|
547 |
-
from
|
548 |
|
549 |
eval_args = dataclass_from_dict(EvalArgs, args.eval)
|
550 |
|
|
|
544 |
if args.eval is not None and every_n_steps(
|
545 |
train_state, args.checkpoint.eval.every, acc_step=0
|
546 |
):
|
547 |
+
from bytelatent.eval import EVAL_FOLDER_NAME, EvalArgs, launch_eval
|
548 |
|
549 |
eval_args = dataclass_from_dict(EvalArgs, args.eval)
|
550 |
|
bytelatent/args.py
CHANGED
@@ -5,6 +5,7 @@ from typing import Any
|
|
5 |
|
6 |
import numpy as np
|
7 |
import yaml
|
|
|
8 |
from pydantic import BaseModel, ConfigDict
|
9 |
|
10 |
from bytelatent.checkpoint import CheckpointArgs
|
@@ -39,6 +40,19 @@ def get_rng_state(seed: int, rank: int, world_size: int) -> dict[str, Any]:
|
|
39 |
return np.random.default_rng((seed, rank, world_size)).bit_generator.state
|
40 |
|
41 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
42 |
def distribute_data_to_rank(
|
43 |
*,
|
44 |
dataset_path: str,
|
@@ -71,6 +85,22 @@ def distribute_data_to_rank(
|
|
71 |
return rank_to_arrow_iterator_params[rank]
|
72 |
|
73 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
74 |
class DataloaderArgs(BaseModel):
|
75 |
model_config = ConfigDict(extra="forbid")
|
76 |
s3_profile: str | None = None
|
@@ -168,6 +198,58 @@ class DataloaderArgs(BaseModel):
|
|
168 |
return packing_iterator
|
169 |
|
170 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
171 |
class TrainArgs(BaseModel):
|
172 |
model_config = ConfigDict(extra="forbid")
|
173 |
name: str = "lingua"
|
@@ -186,6 +268,9 @@ class TrainArgs(BaseModel):
|
|
186 |
|
187 |
# Nb optimizer steps to take
|
188 |
steps: int = 1000
|
|
|
|
|
|
|
189 |
|
190 |
data: DataloaderArgs = DataloaderArgs()
|
191 |
optim: OptimArgs = OptimArgs()
|
@@ -203,7 +288,7 @@ class TrainArgs(BaseModel):
|
|
203 |
|
204 |
# If set to None, eval is run locally otherwise it launches a new job with the given number of gpus
|
205 |
async_eval_gpus: int | None = None
|
206 |
-
eval:
|
207 |
eval_on_gpus: int | None = None
|
208 |
|
209 |
def dump_to_yaml_file(
|
|
|
5 |
|
6 |
import numpy as np
|
7 |
import yaml
|
8 |
+
from omegaconf import OmegaConf
|
9 |
from pydantic import BaseModel, ConfigDict
|
10 |
|
11 |
from bytelatent.checkpoint import CheckpointArgs
|
|
|
40 |
return np.random.default_rng((seed, rank, world_size)).bit_generator.state
|
41 |
|
42 |
|
43 |
+
def parse_args(args_cls):
|
44 |
+
cli_args = OmegaConf.from_cli()
|
45 |
+
file_cfg = OmegaConf.load(cli_args.config)
|
46 |
+
# We remove 'config' attribute from config as the underlying DataClass does not have it
|
47 |
+
del cli_args.config
|
48 |
+
|
49 |
+
default_cfg = OmegaConf.create(args_cls().model_dump())
|
50 |
+
cfg = OmegaConf.merge(default_cfg, file_cfg, cli_args)
|
51 |
+
cfg = OmegaConf.to_container(cfg, resolve=True, throw_on_missing=True)
|
52 |
+
pydantic_args = args_cls.model_validate(cfg)
|
53 |
+
return pydantic_args
|
54 |
+
|
55 |
+
|
56 |
def distribute_data_to_rank(
|
57 |
*,
|
58 |
dataset_path: str,
|
|
|
85 |
return rank_to_arrow_iterator_params[rank]
|
86 |
|
87 |
|
88 |
+
class PackedCausalTransformerGeneratorArgs(BaseModel):
|
89 |
+
model_config = ConfigDict(extra="forbid")
|
90 |
+
temperature: float = 0.0
|
91 |
+
top_p: float | None = None
|
92 |
+
top_k: float | None = None
|
93 |
+
max_gen_len: int = 512 # Maximum number of tokens to generate
|
94 |
+
max_tokens: int = 1024 # Maximum number of tokens that can go through the model
|
95 |
+
max_prompt_len: int | None = None
|
96 |
+
until: list[str] = []
|
97 |
+
compile_prefilling: bool = False
|
98 |
+
reduce_generation_overhead: bool = False
|
99 |
+
show_progress: bool = False
|
100 |
+
dtype: str | None = "bf16"
|
101 |
+
device: str | None = "cuda"
|
102 |
+
|
103 |
+
|
104 |
class DataloaderArgs(BaseModel):
|
105 |
model_config = ConfigDict(extra="forbid")
|
106 |
s3_profile: str | None = None
|
|
|
198 |
return packing_iterator
|
199 |
|
200 |
|
201 |
+
class LMHarnessArgs(BaseModel):
|
202 |
+
model_config = ConfigDict(extra="forbid")
|
203 |
+
tasks: list[Any] | None = None
|
204 |
+
num_fewshot: int | None = None
|
205 |
+
device: str | None = None
|
206 |
+
use_cache: str | None = None
|
207 |
+
cache_requests: bool = False
|
208 |
+
rewrite_requests_cache: bool = False
|
209 |
+
delete_requests_cache: bool = False
|
210 |
+
limit: int | float | None = None
|
211 |
+
bootstrap_iters: int = 100000
|
212 |
+
check_integrity: bool = False
|
213 |
+
write_out: bool = False
|
214 |
+
log_samples: bool = True
|
215 |
+
system_instruction: str | None = None
|
216 |
+
apply_chat_template: bool | str = False
|
217 |
+
fewshot_as_multiturn: bool = False
|
218 |
+
gen_kwargs: str | None = None
|
219 |
+
verbosity: str = "INFO"
|
220 |
+
predict_only: bool = False
|
221 |
+
random_seed: int = 0
|
222 |
+
numpy_random_seed: int = 1234
|
223 |
+
torch_random_seed: int = 1234
|
224 |
+
fewshot_random_seed: int = 1234
|
225 |
+
|
226 |
+
|
227 |
+
class ValidationArgs(BaseModel):
|
228 |
+
model_config = ConfigDict(extra="forbid")
|
229 |
+
max_steps: int | None = (
|
230 |
+
None # If None the whole validation file is used -> /!\ This number of steps is gpu dependent (100 max steps on 8 gpus = 800 steps on 1 gpu)
|
231 |
+
)
|
232 |
+
use_val_from_train_src: bool = True # Use the validation set from training sources
|
233 |
+
root_dir: str = ""
|
234 |
+
sources: list[str] = [] # Other sources to eval on
|
235 |
+
|
236 |
+
|
237 |
+
class EvalArgs(BaseModel):
|
238 |
+
model_config = ConfigDict(extra="forbid")
|
239 |
+
dump_dir: str
|
240 |
+
ckpt_dir: str
|
241 |
+
metric_log_dir: str | None = None
|
242 |
+
generator: PackedCausalTransformerGeneratorArgs = (
|
243 |
+
PackedCausalTransformerGeneratorArgs()
|
244 |
+
)
|
245 |
+
|
246 |
+
harness: LMHarnessArgs | None = LMHarnessArgs()
|
247 |
+
validation: ValidationArgs | None = ValidationArgs()
|
248 |
+
|
249 |
+
global_step: int | None = None # for in-training evaluation
|
250 |
+
s3_profile: str | None = None
|
251 |
+
|
252 |
+
|
253 |
class TrainArgs(BaseModel):
|
254 |
model_config = ConfigDict(extra="forbid")
|
255 |
name: str = "lingua"
|
|
|
268 |
|
269 |
# Nb optimizer steps to take
|
270 |
steps: int = 1000
|
271 |
+
# If not None, halt training after this many steps,
|
272 |
+
# useful for debugging
|
273 |
+
max_steps: int | None = None
|
274 |
|
275 |
data: DataloaderArgs = DataloaderArgs()
|
276 |
optim: OptimArgs = OptimArgs()
|
|
|
288 |
|
289 |
# If set to None, eval is run locally otherwise it launches a new job with the given number of gpus
|
290 |
async_eval_gpus: int | None = None
|
291 |
+
eval: EvalArgs | None = None
|
292 |
eval_on_gpus: int | None = None
|
293 |
|
294 |
def dump_to_yaml_file(
|
bytelatent/checkpoint.py
CHANGED
@@ -7,6 +7,7 @@ import re
|
|
7 |
from pathlib import Path
|
8 |
from typing import List, Optional, Tuple
|
9 |
|
|
|
10 |
import torch
|
11 |
import torch.distributed as dist
|
12 |
import torch.distributed.checkpoint as dcp
|
@@ -21,6 +22,7 @@ from torch.distributed.checkpoint.state_dict import (
|
|
21 |
set_state_dict,
|
22 |
)
|
23 |
|
|
|
24 |
from bytelatent.distributed import get_is_master
|
25 |
|
26 |
logger = logging.getLogger("CHECKPOINT")
|
@@ -51,13 +53,14 @@ class CheckpointArgs(BaseModel):
|
|
51 |
path: str | None = None
|
52 |
init_ckpt_path: str | None = None
|
53 |
continue_training_from_init: bool = False
|
|
|
54 |
|
55 |
|
56 |
def _get_key_step(name: str):
|
57 |
return int(re.findall(RE_DIGITS, name)[-1])
|
58 |
|
59 |
|
60 |
-
def consolidate_checkpoints(ckpt_dir: str):
|
61 |
"""
|
62 |
Consolidates all FSDP checkpoints in a directory to a single file
|
63 |
Consolidate checkpoint is saved in a subdirectory of ckpt_dir
|
@@ -102,15 +105,17 @@ def load_from_checkpoint(
|
|
102 |
dcp.load(state_dict, checkpoint_id=ckpt_dir)
|
103 |
|
104 |
|
|
|
105 |
class CheckpointManager:
|
106 |
def __init__(self, args: CheckpointArgs):
|
107 |
self.path = args.path
|
|
|
108 |
self.dump_every = args.dump
|
109 |
self.eval_every = args.eval
|
110 |
self.init_ckpt_path = args.init_ckpt_path
|
111 |
self.continue_training_from_init = args.continue_training_from_init
|
112 |
|
113 |
-
assert
|
114 |
self.path
|
115 |
), f"Path {self.path} does not exist and needs to be created before using CheckpointManager (use instantiate_and_make_dir)"
|
116 |
|
|
|
7 |
from pathlib import Path
|
8 |
from typing import List, Optional, Tuple
|
9 |
|
10 |
+
import fsspec
|
11 |
import torch
|
12 |
import torch.distributed as dist
|
13 |
import torch.distributed.checkpoint as dcp
|
|
|
22 |
set_state_dict,
|
23 |
)
|
24 |
|
25 |
+
from bytelatent.data.file_util import get_fs
|
26 |
from bytelatent.distributed import get_is_master
|
27 |
|
28 |
logger = logging.getLogger("CHECKPOINT")
|
|
|
53 |
path: str | None = None
|
54 |
init_ckpt_path: str | None = None
|
55 |
continue_training_from_init: bool = False
|
56 |
+
s3_profile: str | None = None
|
57 |
|
58 |
|
59 |
def _get_key_step(name: str):
|
60 |
return int(re.findall(RE_DIGITS, name)[-1])
|
61 |
|
62 |
|
63 |
+
def consolidate_checkpoints(fs: fsspec.AbstractFileSystem, ckpt_dir: str):
|
64 |
"""
|
65 |
Consolidates all FSDP checkpoints in a directory to a single file
|
66 |
Consolidate checkpoint is saved in a subdirectory of ckpt_dir
|
|
|
105 |
dcp.load(state_dict, checkpoint_id=ckpt_dir)
|
106 |
|
107 |
|
108 |
+
# TODO: Rewrite the file operations here to use fsspec to enable s3 writing.
|
109 |
class CheckpointManager:
|
110 |
def __init__(self, args: CheckpointArgs):
|
111 |
self.path = args.path
|
112 |
+
self.fs = get_fs(self.path, s3_profile=args.s3_profile)
|
113 |
self.dump_every = args.dump
|
114 |
self.eval_every = args.eval
|
115 |
self.init_ckpt_path = args.init_ckpt_path
|
116 |
self.continue_training_from_init = args.continue_training_from_init
|
117 |
|
118 |
+
assert self.fs.exists(
|
119 |
self.path
|
120 |
), f"Path {self.path} does not exist and needs to be created before using CheckpointManager (use instantiate_and_make_dir)"
|
121 |
|
bytelatent/configs/debug.yaml
CHANGED
@@ -98,11 +98,4 @@ logging:
|
|
98 |
freq: 10
|
99 |
|
100 |
eval_on_gpus: 8
|
101 |
-
eval:
|
102 |
-
dataset_dir: /checkpoint/amaia/codegen/datasets/eval
|
103 |
-
tasks: boolq,hellaswag,nq,piqa,siqa,tqa,winogrande,obqa,arc_easy,arc_challenge,race.middle,race.high,gsm8k,math,bbh,copa,human_eval_plus,mbpp,mmlu
|
104 |
-
generator:
|
105 |
-
max_tokens: 65536
|
106 |
-
dtype: bf16
|
107 |
-
|
108 |
-
mp_size: 1
|
|
|
98 |
freq: 10
|
99 |
|
100 |
eval_on_gpus: 8
|
101 |
+
eval: null
|
|
|
|
|
|
|
|
|
|
|
|
|
|
bytelatent/configs/entropy_model.yaml
CHANGED
@@ -72,11 +72,4 @@ logging:
|
|
72 |
freq: 10
|
73 |
|
74 |
eval_on_gpus: 8
|
75 |
-
eval:
|
76 |
-
dataset_dir: ???
|
77 |
-
tasks: ???
|
78 |
-
generator:
|
79 |
-
max_tokens: 65536
|
80 |
-
dtype: bf16
|
81 |
-
|
82 |
-
mp_size: 1
|
|
|
72 |
freq: 10
|
73 |
|
74 |
eval_on_gpus: 8
|
75 |
+
eval: null
|
|
|
|
|
|
|
|
|
|
|
|
|
|
bytelatent/data/data_types.py
CHANGED
@@ -40,16 +40,6 @@ class BltPackTokensState(BaseModel):
|
|
40 |
n_views: int = 2
|
41 |
|
42 |
|
43 |
-
class DataLoaderState(BaseModel):
|
44 |
-
model_config = ConfigDict(extra="forbid")
|
45 |
-
multi_choice_state: MultiChoiceState
|
46 |
-
pack_tokens_state: BltPackTokensState
|
47 |
-
prefetch_state: PrefetchState
|
48 |
-
|
49 |
-
|
50 |
-
BltIterator = Iterator[tuple[BltExample, DataLoaderState]]
|
51 |
-
|
52 |
-
|
53 |
class BltSequence(BaseModel):
|
54 |
tokens: list[int]
|
55 |
mask: list[bool]
|
|
|
40 |
n_views: int = 2
|
41 |
|
42 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
43 |
class BltSequence(BaseModel):
|
44 |
tokens: list[int]
|
45 |
mask: list[bool]
|
bytelatent/data/iterators/multiprocess_iterator.py
CHANGED
@@ -128,6 +128,13 @@ class MultiprocessIterator(StatefulIterator):
|
|
128 |
self.producer = None
|
129 |
self.stop_iterating_event = None
|
130 |
self.state_dumped_event = None
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
131 |
|
132 |
def get_state(self) -> MultiprocessIteratorState:
|
133 |
"""
|
@@ -135,6 +142,10 @@ class MultiprocessIterator(StatefulIterator):
|
|
135 |
to halt the background process and allow it to write the state to the main loop
|
136 |
in order to not lose data
|
137 |
"""
|
|
|
|
|
|
|
|
|
138 |
if self.producer is None:
|
139 |
serialized_prefetch_buffer = json.dumps(
|
140 |
[b.to_python_dict() for b in self.prefetch_buffer]
|
@@ -187,6 +198,10 @@ class MultiprocessIterator(StatefulIterator):
|
|
187 |
)
|
188 |
|
189 |
def create_iter(self):
|
|
|
|
|
|
|
|
|
190 |
logging.info("Main thread: Creating MP iterator")
|
191 |
# First yield from the stored prefetch buffer.
|
192 |
if self.prefetch_buffer is not None:
|
|
|
128 |
self.producer = None
|
129 |
self.stop_iterating_event = None
|
130 |
self.state_dumped_event = None
|
131 |
+
self.force_shutdown = False
|
132 |
+
|
133 |
+
def shutdown(self):
|
134 |
+
if self.producer is not None:
|
135 |
+
# This properly shuts things down
|
136 |
+
self.producer.kill()
|
137 |
+
self.force_shutdown = True
|
138 |
|
139 |
def get_state(self) -> MultiprocessIteratorState:
|
140 |
"""
|
|
|
142 |
to halt the background process and allow it to write the state to the main loop
|
143 |
in order to not lose data
|
144 |
"""
|
145 |
+
if self.force_shutdown:
|
146 |
+
raise ValueError(
|
147 |
+
"State will be invalid if shutdown was forced before state persisted."
|
148 |
+
)
|
149 |
if self.producer is None:
|
150 |
serialized_prefetch_buffer = json.dumps(
|
151 |
[b.to_python_dict() for b in self.prefetch_buffer]
|
|
|
198 |
)
|
199 |
|
200 |
def create_iter(self):
|
201 |
+
if self.force_shutdown:
|
202 |
+
raise ValueError(
|
203 |
+
"Iterator may be invalid if shutdown was forced before state persisted."
|
204 |
+
)
|
205 |
logging.info("Main thread: Creating MP iterator")
|
206 |
# First yield from the stored prefetch buffer.
|
207 |
if self.prefetch_buffer is not None:
|
{apps/main → bytelatent}/eval.py
RENAMED
@@ -4,20 +4,20 @@ import json
|
|
4 |
import logging
|
5 |
import os
|
6 |
from collections import defaultdict
|
7 |
-
from dataclasses import asdict, dataclass, field
|
8 |
from datetime import datetime
|
9 |
from pathlib import Path
|
10 |
-
from typing import Any
|
11 |
|
12 |
import torch
|
13 |
-
from lingua.args import dump_config
|
14 |
-
from lingua.data import init_choice_state, setup_sources
|
15 |
from lm_eval import simple_evaluate
|
16 |
from lm_eval.api.instance import Instance
|
17 |
from lm_eval.api.model import LM
|
18 |
from omegaconf import OmegaConf
|
|
|
19 |
|
|
|
20 |
from bytelatent.checkpoint import CONSOLIDATE_FOLDER, consolidate_checkpoints
|
|
|
21 |
from bytelatent.distributed import (
|
22 |
DistributedArgs,
|
23 |
dist_mean_dict,
|
@@ -25,72 +25,17 @@ from bytelatent.distributed import (
|
|
25 |
get_world_size,
|
26 |
setup_torch_distributed,
|
27 |
)
|
28 |
-
from bytelatent.
|
29 |
-
|
30 |
-
from apps.main.generate import (
|
31 |
PackedCausalTransformerGenerator,
|
32 |
-
PackedCausalTransformerGeneratorArgs,
|
33 |
load_consolidated_model_and_tokenizer,
|
34 |
)
|
|
|
35 |
|
36 |
EVAL_FOLDER_NAME = "{:010d}"
|
37 |
|
38 |
logger = logging.getLogger()
|
39 |
|
40 |
|
41 |
-
@dataclass
|
42 |
-
class LMHarnessArgs:
|
43 |
-
tasks: Optional[List[Any]] = None
|
44 |
-
num_fewshot: Optional[int] = None
|
45 |
-
device: Optional[str] = None
|
46 |
-
use_cache: Optional[str] = None
|
47 |
-
cache_requests: bool = False
|
48 |
-
rewrite_requests_cache: bool = False
|
49 |
-
delete_requests_cache: bool = False
|
50 |
-
limit: Optional[Union[int, float]] = None
|
51 |
-
bootstrap_iters: int = 100000
|
52 |
-
check_integrity: bool = False
|
53 |
-
write_out: bool = False
|
54 |
-
log_samples: bool = True
|
55 |
-
system_instruction: Optional[str] = None
|
56 |
-
apply_chat_template: Union[bool, str] = False
|
57 |
-
fewshot_as_multiturn: bool = False
|
58 |
-
gen_kwargs: Optional[str] = None
|
59 |
-
verbosity: str = "INFO"
|
60 |
-
predict_only: bool = False
|
61 |
-
random_seed: int = 0
|
62 |
-
numpy_random_seed: int = 1234
|
63 |
-
torch_random_seed: int = 1234
|
64 |
-
fewshot_random_seed: int = 1234
|
65 |
-
|
66 |
-
|
67 |
-
@dataclass
|
68 |
-
class ValidationArgs:
|
69 |
-
max_steps: Optional[int] = (
|
70 |
-
None # If None the whole validation file is used -> /!\ This number of steps is gpu dependent (100 max steps on 8 gpus = 800 steps on 1 gpu)
|
71 |
-
)
|
72 |
-
use_val_from_train_src: bool = True # Use the validation set from training sources
|
73 |
-
root_dir: str = ""
|
74 |
-
sources: List[str] = field(default_factory=list) # Other sources to eval on
|
75 |
-
|
76 |
-
|
77 |
-
@dataclass
|
78 |
-
class EvalArgs:
|
79 |
-
name: str = "evals"
|
80 |
-
dump_dir: Optional[str] = None
|
81 |
-
metric_log_dir: Optional[str] = None
|
82 |
-
ckpt_dir: str = ""
|
83 |
-
generator: PackedCausalTransformerGeneratorArgs = field(
|
84 |
-
default_factory=PackedCausalTransformerGeneratorArgs
|
85 |
-
)
|
86 |
-
harness: Optional[LMHarnessArgs] = field(default_factory=LMHarnessArgs)
|
87 |
-
validation: Optional[ValidationArgs] = field(default_factory=ValidationArgs)
|
88 |
-
|
89 |
-
wandb: Optional[Any] = None
|
90 |
-
|
91 |
-
global_step: Optional[int] = None # for in-training evaluation
|
92 |
-
|
93 |
-
|
94 |
def all_dicts_same(dict_list):
|
95 |
if not dict_list: # Check if the list is empty
|
96 |
return True
|
@@ -120,7 +65,7 @@ class EvalHarnessLM(LM):
|
|
120 |
self._world_size = get_world_size()
|
121 |
self.device = generator.device
|
122 |
|
123 |
-
def generate_until(self, requests:
|
124 |
prompts, gen_args = zip(*[req.args for req in requests])
|
125 |
assert all_dicts_same(gen_args), "Doesn't support different gen args for now"
|
126 |
gen_args = gen_args[0]
|
@@ -141,7 +86,7 @@ class EvalHarnessLM(LM):
|
|
141 |
filtered_gen.append(g)
|
142 |
return filtered_gen
|
143 |
|
144 |
-
def loglikelihood(self, requests:
|
145 |
prompts, continuations = zip(*[req.args for req in requests])
|
146 |
inputs = [req.args[0] + req.args[1] for req in requests]
|
147 |
max_gen_len = self.generator.max_gen_len
|
@@ -158,7 +103,7 @@ class EvalHarnessLM(LM):
|
|
158 |
self.generator.max_gen_len = max_gen_len
|
159 |
return results
|
160 |
|
161 |
-
def loglikelihood_rolling(self, requests:
|
162 |
prompts = [req.args[0] for req in requests]
|
163 |
max_gen_len = self.generator.max_gen_len
|
164 |
# We temporarily lower max gen len
|
@@ -232,68 +177,73 @@ def eval_on_val(generator, val_args: ValidationArgs, train_cfg):
|
|
232 |
return all_val_metrics
|
233 |
|
234 |
|
235 |
-
def launch_eval(
|
236 |
if not torch.distributed.is_initialized():
|
237 |
setup_torch_distributed(DistributedArgs())
|
|
|
|
|
238 |
if (
|
239 |
-
|
240 |
-
and (
|
241 |
-
and
|
242 |
):
|
243 |
-
consolidate_path =
|
244 |
else:
|
245 |
-
consolidate_path =
|
246 |
-
if not
|
247 |
-
consolidate_path = consolidate_checkpoints(
|
248 |
|
249 |
-
|
250 |
-
|
|
|
251 |
|
252 |
-
consolidate_path = str(consolidate_path)
|
253 |
torch.distributed.barrier()
|
254 |
logger.info("Loading model")
|
|
|
|
|
255 |
model, tokenizer, train_cfg = load_consolidated_model_and_tokenizer(
|
256 |
consolidate_path,
|
257 |
-
model_cls=LMTransformer,
|
258 |
-
model_args_cls=LMTransformerArgs,
|
259 |
)
|
260 |
logger.info("Model loaded")
|
261 |
model.eval()
|
262 |
-
generator = PackedCausalTransformerGenerator(
|
263 |
|
264 |
wrap = EvalHarnessLM(generator)
|
265 |
-
|
|
|
266 |
val_results = None
|
267 |
-
if
|
268 |
-
val_results = eval_on_val(generator,
|
269 |
if get_global_rank() == 0:
|
270 |
-
with open(
|
271 |
f.write(json.dumps(results))
|
272 |
logger.info(f"All evaluation results: {results['results']}")
|
273 |
if val_results is not None:
|
274 |
-
with open(
|
275 |
f.write(json.dumps(val_results))
|
276 |
logger.info(f"All validation results: {val_results}")
|
277 |
-
if
|
278 |
-
metric_log_path =
|
279 |
|
280 |
logger.info(f"Writing metric logs to {metric_log_path}")
|
281 |
timestamp = {
|
282 |
"created_at": datetime.utcnow().isoformat(),
|
283 |
}
|
284 |
-
if
|
285 |
-
timestamp["global_step"] =
|
286 |
print(
|
287 |
json.dumps(timestamp | results["results"]),
|
288 |
-
file=open(metric_log_path, mode="a"),
|
289 |
flush=True,
|
290 |
)
|
291 |
|
292 |
-
val_log_path =
|
|
|
|
|
293 |
if val_results is not None:
|
294 |
print(
|
295 |
json.dumps(timestamp | val_results),
|
296 |
-
file=open(val_log_path, mode="a"),
|
297 |
flush=True,
|
298 |
)
|
299 |
|
@@ -301,53 +251,8 @@ def launch_eval(cfg: EvalArgs):
|
|
301 |
|
302 |
|
303 |
def main():
|
304 |
-
|
305 |
-
|
306 |
-
This accepts arguments as a dot list
|
307 |
-
So if the dataclass looks like
|
308 |
-
|
309 |
-
@dataclass
|
310 |
-
class DummyArgs:
|
311 |
-
name: str
|
312 |
-
model: LMTransformerArgsgs
|
313 |
-
|
314 |
-
@dataclass
|
315 |
-
class LMTransformerArgsgs:
|
316 |
-
dim: int
|
317 |
-
|
318 |
-
Then you can pass model.dim=32 to change values in LMTransformerArgsgs
|
319 |
-
or just name=tictac for top level attributes.
|
320 |
-
|
321 |
-
The behavior here is as follows:
|
322 |
-
1. We instantiate EvalArgs with its default values
|
323 |
-
2. We override those default values with the ones in the provided config file
|
324 |
-
3. We override the result with the additional arguments provided through command line
|
325 |
-
|
326 |
-
For example, if the config is the following
|
327 |
-
|
328 |
-
model:
|
329 |
-
dim: 128
|
330 |
-
n_layers: 4
|
331 |
-
|
332 |
-
and you call eval.py with eval.py model.dim=64
|
333 |
-
|
334 |
-
Then the final TrainArgs will have
|
335 |
-
|
336 |
-
model:
|
337 |
-
dim: 64
|
338 |
-
n_layers: 4
|
339 |
-
|
340 |
-
Plus all the default values in EvalArgs dataclass.
|
341 |
-
"""
|
342 |
-
cli_args = OmegaConf.from_cli()
|
343 |
-
file_cfg = OmegaConf.load(cli_args.config)
|
344 |
-
# We remove 'config' attribute from config as the underlying DataClass does not have it
|
345 |
-
del cli_args.config
|
346 |
-
|
347 |
-
default_cfg = OmegaConf.structured(EvalArgs())
|
348 |
-
cfg = OmegaConf.merge(default_cfg, file_cfg, cli_args)
|
349 |
-
cfg = OmegaConf.to_object(cfg)
|
350 |
-
launch_eval(cfg)
|
351 |
|
352 |
|
353 |
if __name__ == "__main__":
|
|
|
4 |
import logging
|
5 |
import os
|
6 |
from collections import defaultdict
|
|
|
7 |
from datetime import datetime
|
8 |
from pathlib import Path
|
9 |
+
from typing import Any
|
10 |
|
11 |
import torch
|
|
|
|
|
12 |
from lm_eval import simple_evaluate
|
13 |
from lm_eval.api.instance import Instance
|
14 |
from lm_eval.api.model import LM
|
15 |
from omegaconf import OmegaConf
|
16 |
+
from pydantic import BaseModel, ConfigDict
|
17 |
|
18 |
+
from bytelatent.args import EvalArgs, ValidationArgs, parse_args
|
19 |
from bytelatent.checkpoint import CONSOLIDATE_FOLDER, consolidate_checkpoints
|
20 |
+
from bytelatent.data.file_util import get_fs
|
21 |
from bytelatent.distributed import (
|
22 |
DistributedArgs,
|
23 |
dist_mean_dict,
|
|
|
25 |
get_world_size,
|
26 |
setup_torch_distributed,
|
27 |
)
|
28 |
+
from bytelatent.generate import (
|
|
|
|
|
29 |
PackedCausalTransformerGenerator,
|
|
|
30 |
load_consolidated_model_and_tokenizer,
|
31 |
)
|
32 |
+
from bytelatent.transformer import LMTransformer, LMTransformerArgs
|
33 |
|
34 |
EVAL_FOLDER_NAME = "{:010d}"
|
35 |
|
36 |
logger = logging.getLogger()
|
37 |
|
38 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
39 |
def all_dicts_same(dict_list):
|
40 |
if not dict_list: # Check if the list is empty
|
41 |
return True
|
|
|
65 |
self._world_size = get_world_size()
|
66 |
self.device = generator.device
|
67 |
|
68 |
+
def generate_until(self, requests: list[Instance]) -> list[str]:
|
69 |
prompts, gen_args = zip(*[req.args for req in requests])
|
70 |
assert all_dicts_same(gen_args), "Doesn't support different gen args for now"
|
71 |
gen_args = gen_args[0]
|
|
|
86 |
filtered_gen.append(g)
|
87 |
return filtered_gen
|
88 |
|
89 |
+
def loglikelihood(self, requests: list[Instance]) -> list[tuple[float, bool]]:
|
90 |
prompts, continuations = zip(*[req.args for req in requests])
|
91 |
inputs = [req.args[0] + req.args[1] for req in requests]
|
92 |
max_gen_len = self.generator.max_gen_len
|
|
|
103 |
self.generator.max_gen_len = max_gen_len
|
104 |
return results
|
105 |
|
106 |
+
def loglikelihood_rolling(self, requests: list[Instance]) -> list[float]:
|
107 |
prompts = [req.args[0] for req in requests]
|
108 |
max_gen_len = self.generator.max_gen_len
|
109 |
# We temporarily lower max gen len
|
|
|
177 |
return all_val_metrics
|
178 |
|
179 |
|
180 |
+
def launch_eval(eval_args: EvalArgs):
|
181 |
if not torch.distributed.is_initialized():
|
182 |
setup_torch_distributed(DistributedArgs())
|
183 |
+
|
184 |
+
fs = get_fs(eval_args.ckpt_dir, s3_profile=eval_args.s3_profile)
|
185 |
if (
|
186 |
+
fs.exists(eval_args.ckpt_dir)
|
187 |
+
and fs.exists(os.path.join(eval_args.ckpt_dir, "params.json"))
|
188 |
+
and len(fs.glob(os.path.join(eval_args.ckpt_dir, "*.pth"))) != 0
|
189 |
):
|
190 |
+
consolidate_path = eval_args.ckpt_dir
|
191 |
else:
|
192 |
+
consolidate_path = os.path.join(eval_args.ckpt_dir, CONSOLIDATE_FOLDER)
|
193 |
+
if not fs.exists(consolidate_path) and get_global_rank() == 0:
|
194 |
+
consolidate_path = consolidate_checkpoints(eval_args.ckpt_dir)
|
195 |
|
196 |
+
fs.mkdirs(eval_args.dump_dir, exist_ok=True)
|
197 |
+
with fs.open(os.path.join(eval_args.dump_dir, "config.yaml"), "w") as f:
|
198 |
+
f.write(eval_args.model_dump_json())
|
199 |
|
|
|
200 |
torch.distributed.barrier()
|
201 |
logger.info("Loading model")
|
202 |
+
# TODO: Make this general so that it works with either
|
203 |
+
# LMTransformer or Blt, similar with args
|
204 |
model, tokenizer, train_cfg = load_consolidated_model_and_tokenizer(
|
205 |
consolidate_path,
|
|
|
|
|
206 |
)
|
207 |
logger.info("Model loaded")
|
208 |
model.eval()
|
209 |
+
generator = PackedCausalTransformerGenerator(eval_args.generator, model, tokenizer)
|
210 |
|
211 |
wrap = EvalHarnessLM(generator)
|
212 |
+
# Redo
|
213 |
+
results = simple_evaluate(wrap, eval_args.harness.model_dump())
|
214 |
val_results = None
|
215 |
+
if eval_args.validation:
|
216 |
+
val_results = eval_on_val(generator, eval_args.validation, train_cfg)
|
217 |
if get_global_rank() == 0:
|
218 |
+
with fs.open(os.path.join(eval_args.dump_dir, "results.json"), "w") as f:
|
219 |
f.write(json.dumps(results))
|
220 |
logger.info(f"All evaluation results: {results['results']}")
|
221 |
if val_results is not None:
|
222 |
+
with fs.open(os.path.join(eval_args.dump_dir, "validation.json"), "w") as f:
|
223 |
f.write(json.dumps(val_results))
|
224 |
logger.info(f"All validation results: {val_results}")
|
225 |
+
if eval_args.metric_log_dir and get_global_rank() == 0:
|
226 |
+
metric_log_path = os.path.join(eval_args.metric_log_dir, "metrics.eval.jsonl")
|
227 |
|
228 |
logger.info(f"Writing metric logs to {metric_log_path}")
|
229 |
timestamp = {
|
230 |
"created_at": datetime.utcnow().isoformat(),
|
231 |
}
|
232 |
+
if eval_args.global_step is not None:
|
233 |
+
timestamp["global_step"] = eval_args.global_step
|
234 |
print(
|
235 |
json.dumps(timestamp | results["results"]),
|
236 |
+
file=fs.open(metric_log_path, mode="a"),
|
237 |
flush=True,
|
238 |
)
|
239 |
|
240 |
+
val_log_path = os.path.join(
|
241 |
+
eval_args.metric_log_dir, "metrics.validation.jsonl"
|
242 |
+
)
|
243 |
if val_results is not None:
|
244 |
print(
|
245 |
json.dumps(timestamp | val_results),
|
246 |
+
file=fs.open(val_log_path, mode="a"),
|
247 |
flush=True,
|
248 |
)
|
249 |
|
|
|
251 |
|
252 |
|
253 |
def main():
|
254 |
+
eval_args = parse_args(EvalArgs)
|
255 |
+
launch_eval(eval_args)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
256 |
|
257 |
|
258 |
if __name__ == "__main__":
|
{apps/main → bytelatent}/generate.py
RENAMED
@@ -1,20 +1,16 @@
|
|
1 |
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
|
|
|
3 |
import time
|
4 |
-
from dataclasses import dataclass, field
|
5 |
-
from pathlib import Path
|
6 |
-
from typing import List, Optional
|
7 |
|
8 |
import torch
|
9 |
-
from lingua.args import dataclass_from_dict
|
10 |
-
from lingua.tokenizers.abstract_tokenizer import Tokenizer
|
11 |
-
from lingua.tokenizers.build_tokenizer import build_tokenizer
|
12 |
from omegaconf import OmegaConf
|
13 |
from torch import nn
|
14 |
from torch.nn import functional as F
|
15 |
from torch.nn.attention.flex_attention import create_block_mask
|
16 |
from tqdm import tqdm
|
17 |
|
|
|
18 |
from bytelatent.base_transformer import (
|
19 |
Attention,
|
20 |
causal_mask,
|
@@ -23,7 +19,10 @@ from bytelatent.base_transformer import (
|
|
23 |
lengths_to_start_ids,
|
24 |
)
|
25 |
from bytelatent.checkpoint import CONSOLIDATE_NAME
|
26 |
-
from bytelatent.
|
|
|
|
|
|
|
27 |
|
28 |
|
29 |
def sample_top_p(probs: torch.Tensor, p: float) -> torch.Tensor:
|
@@ -62,7 +61,7 @@ def sample_tokens(logits, temperature=0.0, top_p=None, top_k=None):
|
|
62 |
return next_token.view(shape[:-1])
|
63 |
|
64 |
|
65 |
-
def pack_prompts(prompts:
|
66 |
res = []
|
67 |
lengths = []
|
68 |
for i, p in enumerate(prompts):
|
@@ -120,22 +119,6 @@ class KVCache(nn.Module):
|
|
120 |
return self.k_cache, self.v_cache
|
121 |
|
122 |
|
123 |
-
@dataclass
|
124 |
-
class PackedCausalTransformerGeneratorArgs:
|
125 |
-
temperature: float = 0.0
|
126 |
-
top_p: Optional[float] = None
|
127 |
-
top_k: Optional[float] = None
|
128 |
-
max_gen_len: int = 512 # Maximum number of tokens to generate
|
129 |
-
max_tokens: int = 1024 # Maximum number of tokens that can go through the model
|
130 |
-
max_prompt_len: Optional[int] = None
|
131 |
-
until: List[str] = field(default_factory=list)
|
132 |
-
compile_prefilling: bool = False
|
133 |
-
reduce_generation_overhead: bool = False
|
134 |
-
show_progress: bool = False
|
135 |
-
dtype: Optional[str] = "bf16"
|
136 |
-
device: Optional[str] = "cuda"
|
137 |
-
|
138 |
-
|
139 |
class PackedCausalTransformerGenerator:
|
140 |
def __init__(
|
141 |
self,
|
@@ -401,25 +384,29 @@ class PackedCausalTransformerGenerator:
|
|
401 |
|
402 |
def load_consolidated_model_and_tokenizer(
|
403 |
consolidated_path,
|
404 |
-
model_cls=LMTransformer,
|
405 |
-
model_args_cls=LMTransformerArgs,
|
406 |
):
|
407 |
-
|
408 |
-
|
409 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
410 |
|
411 |
param_dtype = dict(fp32=torch.float32, fp16=torch.float16, bf16=torch.bfloat16)[
|
412 |
-
|
413 |
]
|
414 |
-
|
415 |
-
|
416 |
-
model = model_cls(model_args)
|
417 |
-
st_dict = torch.load(ckpt_path / CONSOLIDATE_NAME, weights_only=True)
|
418 |
model.load_state_dict(st_dict["model"])
|
419 |
model = model.cuda().eval()
|
420 |
for param in model.parameters():
|
421 |
param.data = param.data.to(dtype=param_dtype)
|
422 |
-
return model, tokenizer,
|
423 |
|
424 |
|
425 |
def main():
|
|
|
1 |
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
|
3 |
+
import os
|
4 |
import time
|
|
|
|
|
|
|
5 |
|
6 |
import torch
|
|
|
|
|
|
|
7 |
from omegaconf import OmegaConf
|
8 |
from torch import nn
|
9 |
from torch.nn import functional as F
|
10 |
from torch.nn.attention.flex_attention import create_block_mask
|
11 |
from tqdm import tqdm
|
12 |
|
13 |
+
from bytelatent.args import PackedCausalTransformerGeneratorArgs, TrainArgs
|
14 |
from bytelatent.base_transformer import (
|
15 |
Attention,
|
16 |
causal_mask,
|
|
|
19 |
lengths_to_start_ids,
|
20 |
)
|
21 |
from bytelatent.checkpoint import CONSOLIDATE_NAME
|
22 |
+
from bytelatent.data.file_util import get_fs
|
23 |
+
from bytelatent.model.blt import ByteLatentTransformer
|
24 |
+
from bytelatent.tokenizers.abstract_tokenizer import Tokenizer
|
25 |
+
from bytelatent.transformer import LMTransformer
|
26 |
|
27 |
|
28 |
def sample_top_p(probs: torch.Tensor, p: float) -> torch.Tensor:
|
|
|
61 |
return next_token.view(shape[:-1])
|
62 |
|
63 |
|
64 |
+
def pack_prompts(prompts: list[int]):
|
65 |
res = []
|
66 |
lengths = []
|
67 |
for i, p in enumerate(prompts):
|
|
|
119 |
return self.k_cache, self.v_cache
|
120 |
|
121 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
122 |
class PackedCausalTransformerGenerator:
|
123 |
def __init__(
|
124 |
self,
|
|
|
384 |
|
385 |
def load_consolidated_model_and_tokenizer(
|
386 |
consolidated_path,
|
|
|
|
|
387 |
):
|
388 |
+
train_args_path = os.path.join(consolidated_path, "params.json")
|
389 |
+
fs = get_fs(train_args_path)
|
390 |
+
with fs.open(train_args_path) as f:
|
391 |
+
train_args = TrainArgs.model_validate_json(f.read())
|
392 |
+
|
393 |
+
if train_args.train_entropy_model:
|
394 |
+
model_args = train_args.entropy_model
|
395 |
+
model = LMTransformer(model_args)
|
396 |
+
else:
|
397 |
+
model_args = train_args.model
|
398 |
+
model = ByteLatentTransformer(model_args)
|
399 |
|
400 |
param_dtype = dict(fp32=torch.float32, fp16=torch.float16, bf16=torch.bfloat16)[
|
401 |
+
train_args.distributed.model_dtype
|
402 |
]
|
403 |
+
tokenizer = train_args.data.tokenizer_args.build()
|
404 |
+
st_dict = torch.load(consolidated_path / CONSOLIDATE_NAME, weights_only=True)
|
|
|
|
|
405 |
model.load_state_dict(st_dict["model"])
|
406 |
model = model.cuda().eval()
|
407 |
for param in model.parameters():
|
408 |
param.data = param.data.to(dtype=param_dtype)
|
409 |
+
return model, tokenizer, train_args
|
410 |
|
411 |
|
412 |
def main():
|
bytelatent/train.py
CHANGED
@@ -10,7 +10,7 @@ from copy import deepcopy
|
|
10 |
from dataclasses import asdict, dataclass
|
11 |
from pathlib import Path
|
12 |
from timeit import default_timer as timer
|
13 |
-
from typing import Any,
|
14 |
|
15 |
import torch
|
16 |
import torch.distributed
|
@@ -23,9 +23,13 @@ from torch.distributed._tensor import DTensor
|
|
23 |
from torch.distributed.checkpoint.stateful import Stateful
|
24 |
from torch.optim import lr_scheduler
|
25 |
|
26 |
-
from bytelatent.args import TrainArgs
|
27 |
from bytelatent.checkpoint import CheckpointManager, load_from_checkpoint
|
28 |
-
from bytelatent.data.
|
|
|
|
|
|
|
|
|
29 |
from bytelatent.distributed import (
|
30 |
check_model_value_range,
|
31 |
clean_env,
|
@@ -39,6 +43,7 @@ from bytelatent.distributed import (
|
|
39 |
setup_env,
|
40 |
setup_torch_distributed,
|
41 |
)
|
|
|
42 |
from bytelatent.logger import init_logger
|
43 |
from bytelatent.metrics import GPUMemoryMonitor, MetricLogger, get_num_params
|
44 |
from bytelatent.model.blt import ByteLatentTransformer
|
@@ -70,36 +75,49 @@ def flatten_dict(d, parent_key="", sep="_"):
|
|
70 |
return dict(items)
|
71 |
|
72 |
|
73 |
-
def
|
74 |
-
|
75 |
-
|
76 |
-
|
77 |
-
|
78 |
-
|
79 |
-
|
80 |
-
return OmegaConf.to_object(OmegaConf.merge(base, override))
|
81 |
|
82 |
|
|
|
|
|
83 |
@dataclass
|
84 |
class TrainState(Stateful):
|
85 |
step: int # Nb of steps taken by the optimizer
|
86 |
acc_step: int # Nb of accumulation steps done since last optimizer step
|
87 |
scheduler: lr_scheduler.LambdaLR
|
88 |
-
data_loader_state:
|
89 |
scale: float = 1.0
|
|
|
90 |
|
91 |
-
def state_dict(self) ->
|
92 |
return {
|
93 |
"step": self.step,
|
94 |
"acc_step": self.acc_step,
|
95 |
-
"data_loader_state": self.data_loader_state.
|
|
|
96 |
"scheduler": self.scheduler.state_dict(),
|
97 |
}
|
98 |
|
99 |
def load_state_dict(self, state_dict):
|
100 |
self.step = state_dict["step"]
|
101 |
self.acc_step = state_dict["acc_step"]
|
102 |
-
self.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
103 |
self.scheduler.load_state_dict(state_dict["scheduler"])
|
104 |
|
105 |
|
@@ -345,7 +363,10 @@ def train(args: TrainArgs):
|
|
345 |
nwords_since_last_log = 0
|
346 |
time_last_log = timer()
|
347 |
gc.collect()
|
348 |
-
|
|
|
|
|
|
|
349 |
# We constrain train_state.acc_step to be in range 0 to args.grad_acc_steps - 1
|
350 |
train_state.acc_step += 1
|
351 |
train_state.acc_step = train_state.acc_step % args.grad_acc_steps
|
@@ -552,7 +573,6 @@ def train(args: TrainArgs):
|
|
552 |
f" pow: {gpu_mem_stats.power_draw/1000} W"
|
553 |
)
|
554 |
|
555 |
-
saved = False
|
556 |
if every_n_steps(
|
557 |
train_state, args.checkpoint.dump.every, acc_step=0
|
558 |
) or every_n_steps(train_state, args.checkpoint.eval.every, acc_step=0):
|
@@ -567,18 +587,14 @@ def train(args: TrainArgs):
|
|
567 |
if args.eval is not None and every_n_steps(
|
568 |
train_state, args.checkpoint.eval.every, acc_step=0
|
569 |
):
|
570 |
-
|
571 |
-
|
572 |
-
eval_args = dataclass_from_dict(EvalArgs, args.eval)
|
573 |
|
574 |
eval_args.global_step = train_state.step
|
575 |
eval_args.ckpt_dir = str(checkpoint.existing_saves[-1])
|
576 |
-
eval_args.dump_dir =
|
577 |
-
|
578 |
-
|
579 |
-
|
580 |
-
EVAL_FOLDER_NAME.format(train_state.step),
|
581 |
-
)
|
582 |
)
|
583 |
eval_args.metric_log_dir = args.dump_dir
|
584 |
if args.async_eval_gpus is None:
|
@@ -619,6 +635,9 @@ def train(args: TrainArgs):
|
|
619 |
args,
|
620 |
device_mesh=world_mesh,
|
621 |
)
|
|
|
|
|
|
|
622 |
gc.collect()
|
623 |
|
624 |
|
@@ -661,15 +680,7 @@ def main():
|
|
661 |
|
662 |
Plus all the default values in TrainArgs dataclass.
|
663 |
"""
|
664 |
-
|
665 |
-
file_cfg = OmegaConf.load(cli_args.config)
|
666 |
-
# We remove 'config' attribute from config as the underlying DataClass does not have it
|
667 |
-
del cli_args.config
|
668 |
-
|
669 |
-
default_cfg = OmegaConf.create(TrainArgs().model_dump())
|
670 |
-
cfg = OmegaConf.merge(default_cfg, file_cfg, cli_args)
|
671 |
-
cfg = OmegaConf.to_container(cfg, resolve=True, throw_on_missing=True)
|
672 |
-
train_args = TrainArgs.model_validate(cfg)
|
673 |
if train_args.debug_dynamo:
|
674 |
import torch._dynamo
|
675 |
|
|
|
10 |
from dataclasses import asdict, dataclass
|
11 |
from pathlib import Path
|
12 |
from timeit import default_timer as timer
|
13 |
+
from typing import Any, TypeVar
|
14 |
|
15 |
import torch
|
16 |
import torch.distributed
|
|
|
23 |
from torch.distributed.checkpoint.stateful import Stateful
|
24 |
from torch.optim import lr_scheduler
|
25 |
|
26 |
+
from bytelatent.args import TrainArgs, parse_args
|
27 |
from bytelatent.checkpoint import CheckpointManager, load_from_checkpoint
|
28 |
+
from bytelatent.data.iterators.multiprocess_iterator import (
|
29 |
+
MultiprocessIterator,
|
30 |
+
MultiprocessIteratorState,
|
31 |
+
)
|
32 |
+
from bytelatent.data.iterators.packing_iterator import PackingIteratorState
|
33 |
from bytelatent.distributed import (
|
34 |
check_model_value_range,
|
35 |
clean_env,
|
|
|
43 |
setup_env,
|
44 |
setup_torch_distributed,
|
45 |
)
|
46 |
+
from bytelatent.eval import EVAL_FOLDER_NAME, launch_eval
|
47 |
from bytelatent.logger import init_logger
|
48 |
from bytelatent.metrics import GPUMemoryMonitor, MetricLogger, get_num_params
|
49 |
from bytelatent.model.blt import ByteLatentTransformer
|
|
|
75 |
return dict(items)
|
76 |
|
77 |
|
78 |
+
def get_iterator_state_name(iterator_state):
|
79 |
+
if isinstance(iterator_state, MultiprocessIteratorState):
|
80 |
+
return "multiprocess"
|
81 |
+
elif isinstance(iterator_state, PackingIteratorState):
|
82 |
+
return "packing"
|
83 |
+
else:
|
84 |
+
raise ValueError(f"Unsupported iterator to get name from: {iterator_state}")
|
|
|
85 |
|
86 |
|
87 |
+
# TODO: Make this pydantic based instead of data class based
|
88 |
+
# TODO: Generalize this to any iterator state
|
89 |
@dataclass
|
90 |
class TrainState(Stateful):
|
91 |
step: int # Nb of steps taken by the optimizer
|
92 |
acc_step: int # Nb of accumulation steps done since last optimizer step
|
93 |
scheduler: lr_scheduler.LambdaLR
|
94 |
+
data_loader_state: MultiprocessIteratorState | PackingIteratorState
|
95 |
scale: float = 1.0
|
96 |
+
data_loader_class: str | None = None
|
97 |
|
98 |
+
def state_dict(self) -> dict[str, Any]:
|
99 |
return {
|
100 |
"step": self.step,
|
101 |
"acc_step": self.acc_step,
|
102 |
+
"data_loader_state": self.data_loader_state.model_dump(),
|
103 |
+
"data_loader_class": get_iterator_state_name(self.data_loader_state),
|
104 |
"scheduler": self.scheduler.state_dict(),
|
105 |
}
|
106 |
|
107 |
def load_state_dict(self, state_dict):
|
108 |
self.step = state_dict["step"]
|
109 |
self.acc_step = state_dict["acc_step"]
|
110 |
+
self.data_loader_class = state_dict["data_loader_class"]
|
111 |
+
if self.data_loader_class == "multiprocess":
|
112 |
+
self.data_loader_state = MultiprocessIteratorState(
|
113 |
+
**state_dict["data_loader_state"]
|
114 |
+
)
|
115 |
+
elif self.data_loader_class == "packing":
|
116 |
+
self.data_loader_state = PackingIteratorState(
|
117 |
+
**state_dict["data_loader_state"]
|
118 |
+
)
|
119 |
+
else:
|
120 |
+
raise ValueError(f"invalid data loader class: {self.data_loader_class}")
|
121 |
self.scheduler.load_state_dict(state_dict["scheduler"])
|
122 |
|
123 |
|
|
|
363 |
nwords_since_last_log = 0
|
364 |
time_last_log = timer()
|
365 |
gc.collect()
|
366 |
+
saved = False
|
367 |
+
while train_state.step < args.steps and (
|
368 |
+
args.max_steps is None or train_state.step < args.max_steps
|
369 |
+
):
|
370 |
# We constrain train_state.acc_step to be in range 0 to args.grad_acc_steps - 1
|
371 |
train_state.acc_step += 1
|
372 |
train_state.acc_step = train_state.acc_step % args.grad_acc_steps
|
|
|
573 |
f" pow: {gpu_mem_stats.power_draw/1000} W"
|
574 |
)
|
575 |
|
|
|
576 |
if every_n_steps(
|
577 |
train_state, args.checkpoint.dump.every, acc_step=0
|
578 |
) or every_n_steps(train_state, args.checkpoint.eval.every, acc_step=0):
|
|
|
587 |
if args.eval is not None and every_n_steps(
|
588 |
train_state, args.checkpoint.eval.every, acc_step=0
|
589 |
):
|
590 |
+
eval_args = args.eval
|
|
|
|
|
591 |
|
592 |
eval_args.global_step = train_state.step
|
593 |
eval_args.ckpt_dir = str(checkpoint.existing_saves[-1])
|
594 |
+
eval_args.dump_dir = os.path.join(
|
595 |
+
args.dump_dir,
|
596 |
+
"evals",
|
597 |
+
EVAL_FOLDER_NAME.format(train_state.step),
|
|
|
|
|
598 |
)
|
599 |
eval_args.metric_log_dir = args.dump_dir
|
600 |
if args.async_eval_gpus is None:
|
|
|
635 |
args,
|
636 |
device_mesh=world_mesh,
|
637 |
)
|
638 |
+
if isinstance(data_loader, MultiprocessIterator):
|
639 |
+
logger.info("Closing MP iterator before exiting")
|
640 |
+
data_loader.shutdown()
|
641 |
gc.collect()
|
642 |
|
643 |
|
|
|
680 |
|
681 |
Plus all the default values in TrainArgs dataclass.
|
682 |
"""
|
683 |
+
train_args = parse_args(TrainArgs)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
684 |
if train_args.debug_dynamo:
|
685 |
import torch._dynamo
|
686 |
|