File size: 16,528 Bytes
14e4843
 
 
 
2d754ab
14e4843
 
 
 
 
 
 
 
d936aea
14e4843
 
 
 
 
 
 
 
 
 
 
 
 
a4a186c
 
 
 
 
 
 
 
 
 
 
 
d10adef
a4a186c
d10adef
14e4843
 
 
 
 
d6d7ec6
 
 
14e4843
2d754ab
 
14e4843
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d6d7ec6
 
 
 
 
 
14e4843
 
 
d6d7ec6
14e4843
 
 
d6d7ec6
 
 
 
 
 
 
 
 
 
 
 
 
14e4843
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d936aea
faf48c3
14e4843
d6d7ec6
 
 
 
 
 
 
d936aea
d6d7ec6
14e4843
 
 
d6d7ec6
 
 
 
 
 
 
d936aea
d6d7ec6
14e4843
 
 
88d1c0e
14e4843
d6d7ec6
88d1c0e
14e4843
d6d7ec6
 
 
14e4843
 
 
 
d6d7ec6
 
 
 
 
 
 
 
 
14e4843
 
 
 
 
 
 
 
 
d6d7ec6
 
 
14e4843
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d6d7ec6
 
 
14e4843
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d6d7ec6
 
 
 
 
 
 
 
 
 
 
 
 
 
14e4843
 
 
d6d7ec6
 
 
 
 
 
 
 
 
 
 
 
 
 
14e4843
 
 
 
 
 
 
 
 
 
 
 
d6d7ec6
 
 
14e4843
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d6d7ec6
 
 
14e4843
 
 
 
 
 
 
 
 
 
 
 
d6d7ec6
 
 
 
 
 
 
14e4843
 
d6d7ec6
 
 
 
 
 
 
 
 
 
 
 
 
 
14e4843
 
 
d6d7ec6
 
 
 
 
 
 
 
 
 
 
 
 
 
14e4843
 
 
 
 
 
 
 
 
 
 
 
d6d7ec6
 
 
14e4843
 
 
 
 
 
 
 
 
 
 
 
 
d6d7ec6
 
 
 
 
 
 
 
 
 
14e4843
 
 
 
 
 
 
d6d7ec6
 
 
 
 
 
 
 
 
 
14e4843
 
 
 
2d754ab
d6d7ec6
 
d936aea
 
 
 
 
 
2d754ab
 
 
14e4843
2d754ab
 
d6d7ec6
14e4843
d936aea
 
14e4843
 
 
 
 
 
d6d7ec6
d936aea
 
 
 
 
 
d6d7ec6
d936aea
b3946ef
 
 
14e4843
b3946ef
 
 
 
14e4843
b3946ef
 
 
 
 
14e4843
b3946ef
14e4843
b3946ef
 
 
 
 
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
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
#!/usr/bin/env python

import os
import json
import argparse

import socket
import random
from datetime import datetime

from src.backend.run_eval_suite import run_evaluation
from src.backend.manage_requests import check_completed_evals, get_eval_requests, set_eval_request
from src.backend.sort_queue import sort_models_by_priority
from src.backend.envs import Tasks, EVAL_REQUESTS_PATH_BACKEND, EVAL_RESULTS_PATH_BACKEND, DEVICE, Task
from src.backend.manage_requests import EvalRequest
from src.leaderboard.read_evals import EvalResult

from src.envs import QUEUE_REPO, RESULTS_REPO, API
from src.utils import my_snapshot_download

from src.leaderboard.read_evals import get_raw_eval_results

from typing import Optional

import time

import pprint
import logging


# Configure the root logger
logging.basicConfig(
    format="%(asctime)s,%(msecs)03d %(levelname)-8s [%(filename)s:%(lineno)d] %(message)s",
    datefmt="%Y-%m-%d:%H:%M:%S",
    level=logging.WARNING,
)

# Get the 'lm-eval' logger from the third-party library
eval_logger = logging.getLogger("lm-eval")

# Explicitly set the level for 'lm-eval' logger to WARNING
eval_logger.setLevel(logging.WARNING)


def my_set_eval_request(api, eval_request, set_to_status, hf_repo, local_dir):
    for i in range(10):
        try:
            set_eval_request(
                api=api, eval_request=eval_request, set_to_status=set_to_status, hf_repo=hf_repo, local_dir=local_dir
            )
            return
        except Exception as e:
            print(f"Error setting eval request to {set_to_status}: {e}. Retrying in 60 seconds")
            time.sleep(60)
    return


logging.getLogger("openai").setLevel(logging.WARNING)

logging.basicConfig(level=logging.ERROR)
pp = pprint.PrettyPrinter(width=80)

PENDING_STATUS = "PENDING"
RUNNING_STATUS = "RUNNING"
FINISHED_STATUS = "FINISHED"
FAILED_STATUS = "FAILED"

TASKS_HARNESS = [task.value for task in Tasks]


my_snapshot_download(
    repo_id=RESULTS_REPO, revision="main", local_dir=EVAL_RESULTS_PATH_BACKEND, repo_type="dataset", max_workers=60
)
my_snapshot_download(
    repo_id=QUEUE_REPO, revision="main", local_dir=EVAL_REQUESTS_PATH_BACKEND, repo_type="dataset", max_workers=60
)


def sanity_checks():
    print(f"Device: {DEVICE}")

    # pull the eval dataset from the hub and parse any eval requests
    # check completed evals and set them to finished
    my_snapshot_download(
        repo_id=QUEUE_REPO, revision="main", local_dir=EVAL_REQUESTS_PATH_BACKEND, repo_type="dataset", max_workers=60
    )
    check_completed_evals(
        api=API,
        checked_status=RUNNING_STATUS,
        completed_status=FINISHED_STATUS,
        failed_status=FAILED_STATUS,
        hf_repo=QUEUE_REPO,
        local_dir=EVAL_REQUESTS_PATH_BACKEND,
        hf_repo_results=RESULTS_REPO,
        local_dir_results=EVAL_RESULTS_PATH_BACKEND,
    )
    return


def request_to_result_name(request: EvalRequest) -> str:
    # Request: EvalRequest(model='meta-llama/Llama-2-13b-hf', private=False, status='FINISHED',
    # json_filepath='./eval-queue-bk/meta-llama/Llama-2-13b-hf_eval_request_False_False_False.json',
    # weight_type='Original', model_type='pretrained', precision='float32', base_model='', revision='main',
    # submitted_time='2023-09-09T10:52:17Z', likes=389, params=13.016, license='?')
    #
    # EvalResult(eval_name='meta-llama_Llama-2-13b-hf_float32', full_model='meta-llama/Llama-2-13b-hf',
    # org='meta-llama', model='Llama-2-13b-hf', revision='main',
    # results={'nq_open': 33.739612188365655, 'triviaqa': 74.12505572893447},
    # precision=<Precision.float32: ModelDetails(name='float32', symbol='')>,
    # model_type=<ModelType.PT: ModelDetails(name='pretrained', symbol='🟢')>,
    # weight_type=<WeightType.Original: ModelDetails(name='Original', symbol='')>,
    # architecture='LlamaForCausalLM', license='?', likes=389, num_params=13.016, date='2023-09-09T10:52:17Z', still_on_hub=True)
    #
    org_and_model = request.model.split("/", 1)
    if len(org_and_model) == 1:
        model = org_and_model[0]
        res = f"{model}_{request.precision}"
    else:
        org = org_and_model[0]
        model = org_and_model[1]
        res = f"{org}_{model}_{request.precision}"
    return res


def process_evaluation(task: Task, eval_request: EvalRequest, limit: Optional[int] = None) -> dict:
    batch_size = 4
    try:
        results = run_evaluation(
            eval_request=eval_request,
            task_names=[task.benchmark],
            num_fewshot=task.num_fewshot,
            batch_size=batch_size,
            device=DEVICE,
            use_cache=None,
            limit=limit,
        )
    except RuntimeError as e:
        if "No executable batch size found" in str(e):
            batch_size = 1
            results = run_evaluation(
                eval_request=eval_request,
                task_names=[task.benchmark],
                num_fewshot=task.num_fewshot,
                batch_size=batch_size,
                device=DEVICE,
                use_cache=None,
                limit=limit,
            )
        else:
            raise

    # print("RESULTS", results)

    dumped = json.dumps(results, indent=2, default=lambda o: "<not serializable>")
    # print(dumped)

    output_path = os.path.join(
        EVAL_RESULTS_PATH_BACKEND, *eval_request.model.split("/"), f"results_{datetime.now()}.json"
    )
    os.makedirs(os.path.dirname(output_path), exist_ok=True)
    with open(output_path, "w") as f:
        f.write(dumped)

    my_snapshot_download(
        repo_id=RESULTS_REPO, revision="main", local_dir=EVAL_RESULTS_PATH_BACKEND, repo_type="dataset", max_workers=60
    )
    API.upload_file(
        path_or_fileobj=output_path,
        path_in_repo=f"{eval_request.model}/results_{datetime.now()}.json",
        repo_id=RESULTS_REPO,
        repo_type="dataset",
    )
    return results


def process_finished_requests(thr: int, hard_task_lst: Optional[list[str]] = None) -> bool:
    sanity_checks()

    current_finished_status = [FINISHED_STATUS, FAILED_STATUS]

    # Get all eval request that are FINISHED, if you want to run other evals, change this parameter
    eval_requests: list[EvalRequest] = get_eval_requests(
        job_status=current_finished_status, hf_repo=QUEUE_REPO, local_dir=EVAL_REQUESTS_PATH_BACKEND
    )
    # Sort the evals by priority (first submitted, first run)
    eval_requests: list[EvalRequest] = sort_models_by_priority(api=API, models=eval_requests)

    random.shuffle(eval_requests)

    eval_results: list[EvalResult] = get_raw_eval_results(EVAL_RESULTS_PATH_BACKEND, EVAL_REQUESTS_PATH_BACKEND)

    result_name_to_request = {request_to_result_name(r): r for r in eval_requests}
    result_name_to_result = {r.eval_name: r for r in eval_results}

    for eval_request in eval_requests:
        if eval_request.likes >= thr:
            result_name: str = request_to_result_name(eval_request)

            # Check the corresponding result
            eval_result: Optional[EvalResult] = (
                result_name_to_result[result_name] if result_name in result_name_to_result else None
            )

            # breakpoint()

            task_lst = TASKS_HARNESS.copy()
            random.shuffle(task_lst)

            # Iterate over tasks and, if we do not have results for a task, run the relevant evaluations
            for task in task_lst:
                task_name = task.benchmark

                do_run_task = False
                if hard_task_lst is None or any(ss in task_name for ss in hard_task_lst):
                    do_run_task = True

                if (eval_result is None or task_name not in eval_result.results) and do_run_task:
                    eval_request: EvalRequest = result_name_to_request[result_name]

                    my_snapshot_download(
                        repo_id=QUEUE_REPO,
                        revision="main",
                        local_dir=EVAL_REQUESTS_PATH_BACKEND,
                        repo_type="dataset",
                        max_workers=60,
                    )
                    my_set_eval_request(
                        api=API,
                        eval_request=eval_request,
                        set_to_status=RUNNING_STATUS,
                        hf_repo=QUEUE_REPO,
                        local_dir=EVAL_REQUESTS_PATH_BACKEND,
                    )

                    results = process_evaluation(task, eval_request)

                    my_snapshot_download(
                        repo_id=QUEUE_REPO,
                        revision="main",
                        local_dir=EVAL_REQUESTS_PATH_BACKEND,
                        repo_type="dataset",
                        max_workers=60,
                    )
                    my_set_eval_request(
                        api=API,
                        eval_request=eval_request,
                        set_to_status=FINISHED_STATUS,
                        hf_repo=QUEUE_REPO,
                        local_dir=EVAL_REQUESTS_PATH_BACKEND,
                    )

                    return True

    return False


def maybe_refresh_results(thr: int, hard_task_lst: Optional[list[str]] = None) -> bool:
    sanity_checks()

    current_finished_status = [PENDING_STATUS, FINISHED_STATUS, FAILED_STATUS]

    # Get all eval request that are FINISHED, if you want to run other evals, change this parameter
    eval_requests: list[EvalRequest] = get_eval_requests(
        job_status=current_finished_status, hf_repo=QUEUE_REPO, local_dir=EVAL_REQUESTS_PATH_BACKEND
    )
    # Sort the evals by priority (first submitted, first run)
    eval_requests: list[EvalRequest] = sort_models_by_priority(api=API, models=eval_requests)

    random.shuffle(eval_requests)

    eval_results: list[EvalResult] = get_raw_eval_results(EVAL_RESULTS_PATH_BACKEND, EVAL_REQUESTS_PATH_BACKEND)

    result_name_to_request = {request_to_result_name(r): r for r in eval_requests}
    result_name_to_result = {r.eval_name: r for r in eval_results}

    for eval_request in eval_requests:
        if eval_request.likes >= thr:
            result_name: str = request_to_result_name(eval_request)

            # Check the corresponding result
            eval_result: Optional[EvalResult] = (
                result_name_to_result[result_name] if result_name in result_name_to_result else None
            )

            task_lst = TASKS_HARNESS.copy()
            random.shuffle(task_lst)

            # Iterate over tasks and, if we do not have results for a task, run the relevant evaluations
            for task in task_lst:
                task_name = task.benchmark

                do_run_task = False
                if hard_task_lst is None or any(ss in task_name for ss in hard_task_lst):
                    do_run_task = True

                task_lst = ["nq", "trivia", "tqa", "self"]
                if (
                    eval_result is None
                    or do_run_task
                    or task_name not in eval_result.results
                    or any(ss in task_name for ss in task_lst)
                ):
                    eval_request: EvalRequest = result_name_to_request[result_name]

                    my_snapshot_download(
                        repo_id=QUEUE_REPO,
                        revision="main",
                        local_dir=EVAL_REQUESTS_PATH_BACKEND,
                        repo_type="dataset",
                        max_workers=60,
                    )
                    my_set_eval_request(
                        api=API,
                        eval_request=eval_request,
                        set_to_status=RUNNING_STATUS,
                        hf_repo=QUEUE_REPO,
                        local_dir=EVAL_REQUESTS_PATH_BACKEND,
                    )

                    results = process_evaluation(task, eval_request)

                    my_snapshot_download(
                        repo_id=QUEUE_REPO,
                        revision="main",
                        local_dir=EVAL_REQUESTS_PATH_BACKEND,
                        repo_type="dataset",
                        max_workers=60,
                    )
                    my_set_eval_request(
                        api=API,
                        eval_request=eval_request,
                        set_to_status=FINISHED_STATUS,
                        hf_repo=QUEUE_REPO,
                        local_dir=EVAL_REQUESTS_PATH_BACKEND,
                    )

                    return True

    return False


def process_pending_requests() -> bool:
    sanity_checks()

    current_pending_status = [PENDING_STATUS]

    # Get all eval request that are PENDING, if you want to run other evals, change this parameter
    eval_requests = get_eval_requests(
        job_status=current_pending_status, hf_repo=QUEUE_REPO, local_dir=EVAL_REQUESTS_PATH_BACKEND
    )
    # Sort the evals by priority (first submitted, first run)
    eval_requests = sort_models_by_priority(api=API, models=eval_requests)

    random.shuffle(eval_requests)

    print(f"Found {len(eval_requests)} {','.join(current_pending_status)} eval requests")

    if len(eval_requests) == 0:
        return False

    eval_request = eval_requests[0]
    pp.pprint(eval_request)

    my_snapshot_download(
        repo_id=QUEUE_REPO, revision="main", local_dir=EVAL_REQUESTS_PATH_BACKEND, repo_type="dataset", max_workers=60
    )
    my_set_eval_request(
        api=API,
        eval_request=eval_request,
        set_to_status=RUNNING_STATUS,
        hf_repo=QUEUE_REPO,
        local_dir=EVAL_REQUESTS_PATH_BACKEND,
    )

    task_lst = TASKS_HARNESS.copy()
    random.shuffle(task_lst)

    for task in task_lst:
        results = process_evaluation(task, eval_request)

    my_snapshot_download(
        repo_id=QUEUE_REPO, revision="main", local_dir=EVAL_REQUESTS_PATH_BACKEND, repo_type="dataset", max_workers=60
    )
    my_set_eval_request(
        api=API,
        eval_request=eval_request,
        set_to_status=FINISHED_STATUS,
        hf_repo=QUEUE_REPO,
        local_dir=EVAL_REQUESTS_PATH_BACKEND,
    )

    return True


def get_args():
    parser = argparse.ArgumentParser(description="Run the backend")
    parser.add_argument("--debug", action="store_true", help="Run in debug mode")
    # debug parameters
    parser.add_argument("--task", type=str, default="selfcheckgpt", help="Task to debug")
    parser.add_argument("--model", type=str, default="facebook/opt-1.3b", help="Model to debug")
    parser.add_argument("--precision", type=str, default="float16", help="Precision to debug")
    parser.add_argument("--inference-framework", type=str, default="hf-chat", help="Inference framework to debug")
    parser.add_argument("--limit", type=int, default=None, help="Limit for the number of samples")
    return parser.parse_args()


if __name__ == "__main__":
    args = get_args()
    local_debug = args.debug
    # debug specific task by ping
    if local_debug:
        debug_model_names = [args.model]  # Use model from arguments
        debug_task_name = args.task  # Use task from arguments
        task_lst = TASKS_HARNESS.copy()
        for task in task_lst:
            for debug_model_name in debug_model_names:
                task_name = task.benchmark
                if task_name != debug_task_name:
                    continue
                eval_request = EvalRequest(
                    model=debug_model_name, 
                    private=False, 
                    status="", 
                    json_filepath="", 
                    precision=args.precision,  # Use precision from arguments
                    inference_framework=args.inference_framework  # Use inference framework from arguments
                )
                results = process_evaluation(task, eval_request, limit=args.limit)
    else:
        while True:
            res = False

            # if random.randint(0, 10) == 0:
            res = process_pending_requests()
            print(f"waiting for 60 seconds")
            time.sleep(60)

            # if res is False:
            #     if random.randint(0, 5) == 0:
            #         res = maybe_refresh_results(100)
            #     else:
            #         res = process_finished_requests(100)

            # time.sleep(60)

            # if res is False:
            #     if random.randint(0, 5) == 0:
            #         res = maybe_refresh_results(0)
            #     else:
            #         res = process_finished_requests(0)