A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 17, 04:15 UTC
[Misc] fix line length for entire codebase (#3444)
Simon Mo committed
Mar 16, 2024 at 00:36 UTC
8e67598aa6ea6ce37c4c8cb470412db0ea523573
9 files changed
+174
-128
.github/workflows/ruff.yml
+1
-1
@@ -28,7 +28,7 @@ jobs:
28
pip install ruff==0.1.5 codespell==2.2.6 tomli==2.0.1
29
- name: Analysing the code with ruff
30
run: |
31
- ruff vllm tests
31
+ ruff .
32
- name: Spelling check with codespell
33
run: |
34
codespell --toml pyproject.toml
\ No newline at end of file
benchmarks/backend_request_func.py
+5
-3
@@ -110,7 +110,7 @@ async def async_request_vllm(
110
output.ttft = ttft
111
output.latency = time.perf_counter() - st
112
113
- # When streaming, '\0' is appended to the end of the response.
113
+ # When streaming, '\0' is appended to the end of response.
114
body = data.decode("utf-8").strip("\0")
115
output.generated_text = json.loads(
116
body)["text"][0][len(request_func_input.prompt):]
@@ -192,7 +192,8 @@ async def async_request_deepspeed_mii(
192
output = RequestFuncOutput()
193
output.prompt_len = request_func_input.prompt_len
194
195
- # DeepSpeed-MII doesn't support streaming as of Jan 28 2024, will use 0 as placeholder.
195
+ # DeepSpeed-MII doesn't support streaming as of Jan 28 2024,
196
+ # will use 0 as placeholder.
197
# https://github.com/microsoft/DeepSpeed-MII/pull/311
198
output.ttft = 0
199
@@ -344,7 +345,8 @@ async def async_request_openai_chat_completions(
345
return output
346
347
347
-# Since vllm must support Python 3.8, we can't use str.removeprefix(prefix) introduced in Python 3.9
348
+# Since vllm must support Python 3.8, we can't use str.removeprefix(prefix)
349
+# introduced in Python 3.9
350
def remove_prefix(text: str, prefix: str) -> str:
351
if text.startswith(prefix):
352
return text[len(prefix):]
benchmarks/benchmark_prefix_caching.py
+1
-1
@@ -4,7 +4,7 @@ import time
4
from vllm import LLM
5
from vllm import SamplingParams
6
7
-PROMPT = "You are a helpful assistant in recognizes the content of tables in markdown format. Here is a table as fellows. You need to answer my question about the table.\n# Table\n|Opening|Opening|Sl. No.|Film|Cast|Director|Music Director|Notes|\n|----|----|----|----|----|----|----|----|\n|J A N|9|1|Agni Pushpam|Jayabharathi, Kamalahasan|Jeassy|M. K. Arjunan||\n|J A N|16|2|Priyamvada|Mohan Sharma, Lakshmi, KPAC Lalitha|K. S. Sethumadhavan|V. Dakshinamoorthy||\n|J A N|23|3|Yakshagaanam|Madhu, Sheela|Sheela|M. S. Viswanathan||\n|J A N|30|4|Paalkkadal|Sheela, Sharada|T. K. Prasad|A. T. Ummer||\n|F E B|5|5|Amma|Madhu, Srividya|M. Krishnan Nair|M. K. Arjunan||\n|F E B|13|6|Appooppan|Thikkurissi Sukumaran Nair, Kamal Haasan|P. Bhaskaran|M. S. Baburaj||\n|F E B|20|7|Srishti|Chowalloor Krishnankutty, Ravi Alummoodu|K. T. Muhammad|M. S. Baburaj||\n|F E B|20|8|Vanadevatha|Prem Nazir, Madhubala|Yusufali Kechery|G. Devarajan||\n|F E B|27|9|Samasya|Madhu, Kamalahaasan|K. Thankappan|Shyam||\n|F E B|27|10|Yudhabhoomi|K. P. Ummer, Vidhubala|Crossbelt Mani|R. K. Shekhar||\n|M A R|5|11|Seemantha Puthran|Prem Nazir, Jayabharathi|A. B. Raj|M. K. Arjunan||\n|M A R|12|12|Swapnadanam|Rani Chandra, Dr. Mohandas|K. G. George|Bhaskar Chandavarkar||\n|M A R|19|13|Thulavarsham|Prem Nazir, sreedevi, Sudheer|N. Sankaran Nair|V. Dakshinamoorthy||\n|M A R|20|14|Aruthu|Kaviyoor Ponnamma, Kamalahasan|Ravi|G. Devarajan||\n|M A R|26|15|Swimming Pool|Kamal Haasan, M. G. Soman|J. Sasikumar|M. K. Arjunan||\n\n# Question\nWhat' s the content in the (1,1) cells\n"
7
+PROMPT = "You are a helpful assistant in recognizes the content of tables in markdown format. Here is a table as fellows. You need to answer my question about the table.\n# Table\n|Opening|Opening|Sl. No.|Film|Cast|Director|Music Director|Notes|\n|----|----|----|----|----|----|----|----|\n|J A N|9|1|Agni Pushpam|Jayabharathi, Kamalahasan|Jeassy|M. K. Arjunan||\n|J A N|16|2|Priyamvada|Mohan Sharma, Lakshmi, KPAC Lalitha|K. S. Sethumadhavan|V. Dakshinamoorthy||\n|J A N|23|3|Yakshagaanam|Madhu, Sheela|Sheela|M. S. Viswanathan||\n|J A N|30|4|Paalkkadal|Sheela, Sharada|T. K. Prasad|A. T. Ummer||\n|F E B|5|5|Amma|Madhu, Srividya|M. Krishnan Nair|M. K. Arjunan||\n|F E B|13|6|Appooppan|Thikkurissi Sukumaran Nair, Kamal Haasan|P. Bhaskaran|M. S. Baburaj||\n|F E B|20|7|Srishti|Chowalloor Krishnankutty, Ravi Alummoodu|K. T. Muhammad|M. S. Baburaj||\n|F E B|20|8|Vanadevatha|Prem Nazir, Madhubala|Yusufali Kechery|G. Devarajan||\n|F E B|27|9|Samasya|Madhu, Kamalahaasan|K. Thankappan|Shyam||\n|F E B|27|10|Yudhabhoomi|K. P. Ummer, Vidhubala|Crossbelt Mani|R. K. Shekhar||\n|M A R|5|11|Seemantha Puthran|Prem Nazir, Jayabharathi|A. B. Raj|M. K. Arjunan||\n|M A R|12|12|Swapnadanam|Rani Chandra, Dr. Mohandas|K. G. George|Bhaskar Chandavarkar||\n|M A R|19|13|Thulavarsham|Prem Nazir, sreedevi, Sudheer|N. Sankaran Nair|V. Dakshinamoorthy||\n|M A R|20|14|Aruthu|Kaviyoor Ponnamma, Kamalahasan|Ravi|G. Devarajan||\n|M A R|26|15|Swimming Pool|Kamal Haasan, M. G. Soman|J. Sasikumar|M. K. Arjunan||\n\n# Question\nWhat' s the content in the (1,1) cells\n" # noqa: E501
8
9
10
def test_prefix(llm=None, sampling_params=None, prompts=None):
benchmarks/benchmark_serving.py
+4
-2
@@ -293,7 +293,9 @@ def main(args: argparse.Namespace):
293
294
# Save to file
295
base_model_id = model_id.split("/")[-1]
296
- file_name = f"{backend}-{args.request_rate}qps-{base_model_id}-{current_dt}.json"
296
+ file_name = (
297
+ f"{backend}-{args.request_rate}qps-{base_model_id}-{current_dt}.json"
298
+ )
299
with open(file_name, "w") as outfile:
300
json.dump(result_json, outfile)
301
@@ -341,7 +343,7 @@ if __name__ == "__main__":
343
"--tokenizer",
344
type=str,
345
help=
344
- "Name or path of the tokenizer, if not using the default model tokenizer.",
346
+ "Name or path of the tokenizer, if not using the default tokenizer.",
347
)
348
parser.add_argument(
349
"--best-of",
collect_env.py
+106
-75
@@ -1,3 +1,4 @@
1
+# ruff: noqa
2
# code borrowed from https://github.com/pytorch/pytorch/blob/main/torch/utils/collect_env.py
3
4
# Unlike the rest of the PyTorch this file must be python2 compliant.
@@ -11,7 +12,6 @@ import sys
12
import os
13
from collections import namedtuple
14
14
-
15
try:
16
import torch
17
TORCH_AVAILABLE = True
@@ -19,38 +19,40 @@ except (ImportError, NameError, AttributeError, OSError):
19
TORCH_AVAILABLE = False
20
21
# System Environment Information
22
-SystemEnv = namedtuple('SystemEnv', [
23
- 'torch_version',
24
- 'is_debug_build',
25
- 'cuda_compiled_version',
26
- 'gcc_version',
27
- 'clang_version',
28
- 'cmake_version',
29
- 'os',
30
- 'libc_version',
31
- 'python_version',
32
- 'python_platform',
33
- 'is_cuda_available',
34
- 'cuda_runtime_version',
35
- 'cuda_module_loading',
36
- 'nvidia_driver_version',
37
- 'nvidia_gpu_models',
38
- 'cudnn_version',
39
- 'pip_version', # 'pip' or 'pip3'
40
- 'pip_packages',
41
- 'conda_packages',
42
- 'hip_compiled_version',
43
- 'hip_runtime_version',
44
- 'miopen_runtime_version',
45
- 'caching_allocator_config',
46
- 'is_xnnpack_available',
47
- 'cpu_info',
48
- 'rocm_version', # vllm specific field
49
- 'neuron_sdk_version', # vllm specific field
50
- 'vllm_version', # vllm specific field
51
- 'vllm_build_flags', # vllm specific field
52
- 'gpu_topo', # vllm specific field
53
-])
22
+SystemEnv = namedtuple(
23
+ 'SystemEnv',
24
+ [
25
+ 'torch_version',
26
+ 'is_debug_build',
27
+ 'cuda_compiled_version',
28
+ 'gcc_version',
29
+ 'clang_version',
30
+ 'cmake_version',
31
+ 'os',
32
+ 'libc_version',
33
+ 'python_version',
34
+ 'python_platform',
35
+ 'is_cuda_available',
36
+ 'cuda_runtime_version',
37
+ 'cuda_module_loading',
38
+ 'nvidia_driver_version',
39
+ 'nvidia_gpu_models',
40
+ 'cudnn_version',
41
+ 'pip_version', # 'pip' or 'pip3'
42
+ 'pip_packages',
43
+ 'conda_packages',
44
+ 'hip_compiled_version',
45
+ 'hip_runtime_version',
46
+ 'miopen_runtime_version',
47
+ 'caching_allocator_config',
48
+ 'is_xnnpack_available',
49
+ 'cpu_info',
50
+ 'rocm_version', # vllm specific field
51
+ 'neuron_sdk_version', # vllm specific field
52
+ 'vllm_version', # vllm specific field
53
+ 'vllm_build_flags', # vllm specific field
54
+ 'gpu_topo', # vllm specific field
55
+ ])
56
57
DEFAULT_CONDA_PATTERNS = {
58
"torch",
@@ -77,8 +79,10 @@ DEFAULT_PIP_PATTERNS = {
79
def run(command):
80
"""Return (return-code, stdout, stderr)."""
81
shell = True if type(command) is str else False
80
- p = subprocess.Popen(command, stdout=subprocess.PIPE,
81
- stderr=subprocess.PIPE, shell=shell)
82
+ p = subprocess.Popen(command,
83
+ stdout=subprocess.PIPE,
84
+ stderr=subprocess.PIPE,
85
+ shell=shell)
86
raw_output, raw_err = p.communicate()
87
rc = p.returncode
88
if get_platform() == 'win32':
@@ -108,6 +112,7 @@ def run_and_parse_first_match(run_lambda, command, regex):
112
return None
113
return match.group(1)
114
115
+
116
def run_and_return_first_line(run_lambda, command):
117
"""Run command using run_lambda and returns first line if output is not empty."""
118
rc, out, _ = run_lambda(command)
@@ -124,22 +129,23 @@ def get_conda_packages(run_lambda, patterns=None):
129
if out is None:
130
return out
131
127
- return "\n".join(
128
- line
129
- for line in out.splitlines()
130
- if not line.startswith("#")
131
- and any(name in line for name in patterns)
132
- )
132
+ return "\n".join(line for line in out.splitlines()
133
+ if not line.startswith("#") and any(name in line
134
+ for name in patterns))
135
+
136
137
def get_gcc_version(run_lambda):
138
return run_and_parse_first_match(run_lambda, 'gcc --version', r'gcc (.*)')
139
140
+
141
def get_clang_version(run_lambda):
138
- return run_and_parse_first_match(run_lambda, 'clang --version', r'clang version (.*)')
142
+ return run_and_parse_first_match(run_lambda, 'clang --version',
143
+ r'clang version (.*)')
144
145
146
def get_cmake_version(run_lambda):
142
- return run_and_parse_first_match(run_lambda, 'cmake --version', r'cmake (.*)')
147
+ return run_and_parse_first_match(run_lambda, 'cmake --version',
148
+ r'cmake (.*)')
149
150
151
def get_nvidia_driver_version(run_lambda):
@@ -148,11 +154,13 @@ def get_nvidia_driver_version(run_lambda):
154
return run_and_parse_first_match(run_lambda, cmd,
155
r'com[.]nvidia[.]CUDA [(](.*?)[)]')
156
smi = get_nvidia_smi()
151
- return run_and_parse_first_match(run_lambda, smi, r'Driver Version: (.*?) ')
157
+ return run_and_parse_first_match(run_lambda, smi,
158
+ r'Driver Version: (.*?) ')
159
160
161
def get_gpu_info(run_lambda):
155
- if get_platform() == 'darwin' or (TORCH_AVAILABLE and hasattr(torch.version, 'hip') and torch.version.hip is not None):
162
+ if get_platform() == 'darwin' or (TORCH_AVAILABLE and hasattr(
163
+ torch.version, 'hip') and torch.version.hip is not None):
164
if TORCH_AVAILABLE and torch.cuda.is_available():
165
if torch.version.hip is not None:
166
prop = torch.cuda.get_device_properties(0)
@@ -174,7 +182,8 @@ def get_gpu_info(run_lambda):
182
183
184
def get_running_cuda_version(run_lambda):
177
- return run_and_parse_first_match(run_lambda, 'nvcc --version', r'release .+ V(.*)')
185
+ return run_and_parse_first_match(run_lambda, 'nvcc --version',
186
+ r'release .+ V(.*)')
187
188
189
def get_cudnn_version(run_lambda):
@@ -219,8 +228,10 @@ def get_nvidia_smi():
228
smi = 'nvidia-smi'
229
if get_platform() == 'win32':
230
system_root = os.environ.get('SYSTEMROOT', 'C:\\Windows')
222
- program_files_root = os.environ.get('PROGRAMFILES', 'C:\\Program Files')
223
- legacy_path = os.path.join(program_files_root, 'NVIDIA Corporation', 'NVSMI', smi)
231
+ program_files_root = os.environ.get('PROGRAMFILES',
232
+ 'C:\\Program Files')
233
+ legacy_path = os.path.join(program_files_root, 'NVIDIA Corporation',
234
+ 'NVSMI', smi)
235
new_path = os.path.join(system_root, 'System32', smi)
236
smis = [new_path, legacy_path]
237
for candidate_smi in smis:
@@ -232,7 +243,8 @@ def get_nvidia_smi():
243
244
def get_rocm_version(run_lambda):
245
"""Returns the ROCm version if available, otherwise 'N/A'."""
235
- return run_and_parse_first_match(run_lambda, 'hipcc --version', r'HIP version: (\S+)')
246
+ return run_and_parse_first_match(run_lambda, 'hipcc --version',
247
+ r'HIP version: (\S+)')
248
249
250
def get_neuron_sdk_version(run_lambda):
@@ -342,13 +354,16 @@ def get_gpu_topo(run_lambda):
354
# ProcessorType=3
355
# Revision=27142
356
357
+
358
def get_cpu_info(run_lambda):
359
rc, out, err = 0, '', ''
360
if get_platform() == 'linux':
361
rc, out, err = run_lambda('lscpu')
362
elif get_platform() == 'win32':
350
- rc, out, err = run_lambda('wmic cpu get Name,Manufacturer,Family,Architecture,ProcessorType,DeviceID, \
351
- CurrentClockSpeed,MaxClockSpeed,L2CacheSize,L2CacheSpeed,Revision /VALUE')
363
+ rc, out, err = run_lambda(
364
+ 'wmic cpu get Name,Manufacturer,Family,Architecture,ProcessorType,DeviceID, \
365
+ CurrentClockSpeed,MaxClockSpeed,L2CacheSize,L2CacheSpeed,Revision /VALUE'
366
+ )
367
elif get_platform() == 'darwin':
368
rc, out, err = run_lambda("sysctl -n machdep.cpu.brand_string")
369
cpu_info = 'None'
@@ -373,18 +388,22 @@ def get_platform():
388
389
390
def get_mac_version(run_lambda):
376
- return run_and_parse_first_match(run_lambda, 'sw_vers -productVersion', r'(.*)')
391
+ return run_and_parse_first_match(run_lambda, 'sw_vers -productVersion',
392
+ r'(.*)')
393
394
395
def get_windows_version(run_lambda):
396
system_root = os.environ.get('SYSTEMROOT', 'C:\\Windows')
397
wmic_cmd = os.path.join(system_root, 'System32', 'Wbem', 'wmic')
398
findstr_cmd = os.path.join(system_root, 'System32', 'findstr')
383
- return run_and_read_all(run_lambda, '{} os get Caption | {} /v Caption'.format(wmic_cmd, findstr_cmd))
399
+ return run_and_read_all(
400
+ run_lambda,
401
+ '{} os get Caption | {} /v Caption'.format(wmic_cmd, findstr_cmd))
402
403
404
def get_lsb_version(run_lambda):
387
- return run_and_parse_first_match(run_lambda, 'lsb_release -a', r'Description:\t(.*)')
405
+ return run_and_parse_first_match(run_lambda, 'lsb_release -a',
406
+ r'Description:\t(.*)')
407
408
409
def check_release_file(run_lambda):
@@ -443,11 +462,8 @@ def get_pip_packages(run_lambda, patterns=None):
462
# But here it is invoked as `python -mpip`
463
def run_with_pip(pip):
464
out = run_and_read_all(run_lambda, pip + ["list", "--format=freeze"])
446
- return "\n".join(
447
- line
448
- for line in out.splitlines()
449
- if any(name in line for name in patterns)
450
- )
465
+ return "\n".join(line for line in out.splitlines()
466
+ if any(name in line for name in patterns))
467
468
pip_version = 'pip3' if sys.version[0] == '3' else 'pip'
469
out = run_with_pip([sys.executable, '-mpip'])
@@ -472,10 +488,12 @@ def get_cuda_module_loading_config():
488
def is_xnnpack_available():
489
if TORCH_AVAILABLE:
490
import torch.backends.xnnpack
475
- return str(torch.backends.xnnpack.enabled) # type: ignore[attr-defined]
491
+ return str(
492
+ torch.backends.xnnpack.enabled) # type: ignore[attr-defined]
493
else:
494
return "N/A"
495
496
+
497
def get_env_info():
498
run_lambda = run
499
pip_version, pip_list_output = get_pip_packages(run_lambda)
@@ -485,9 +503,11 @@ def get_env_info():
503
debug_mode_str = str(torch.version.debug)
504
cuda_available_str = str(torch.cuda.is_available())
505
cuda_version_str = torch.version.cuda
488
- if not hasattr(torch.version, 'hip') or torch.version.hip is None: # cuda version
506
+ if not hasattr(torch.version,
507
+ 'hip') or torch.version.hip is None: # cuda version
508
hip_compiled_version = hip_runtime_version = miopen_runtime_version = 'N/A'
509
else: # HIP version
510
+
511
def get_version_or_na(cfg, prefix):
512
_lst = [s.rsplit(None, 1)[-1] for s in cfg if prefix in s]
513
return _lst[0] if _lst else 'N/A'
@@ -514,7 +534,9 @@ def get_env_info():
534
return SystemEnv(
535
torch_version=version_str,
536
is_debug_build=debug_mode_str,
517
- python_version='{} ({}-bit runtime)'.format(sys_version, sys.maxsize.bit_length() + 1),
537
+ python_version='{} ({}-bit runtime)'.format(
538
+ sys_version,
539
+ sys.maxsize.bit_length() + 1),
540
python_platform=get_python_platform(),
541
is_cuda_available=cuda_available_str,
542
cuda_compiled_version=cuda_version_str,
@@ -544,6 +566,7 @@ def get_env_info():
566
gpu_topo=gpu_topo,
567
)
568
569
+
570
env_info_fmt = """
571
PyTorch version: {torch_version}
572
Is debug build: {is_debug_build}
@@ -588,6 +611,7 @@ GPU Topology:
611
612
613
def pretty_str(envinfo):
614
+
615
def replace_nones(dct, replacement='Could not collect'):
616
for key in dct.keys():
617
if dct[key] is not None:
@@ -632,9 +656,10 @@ def pretty_str(envinfo):
656
'nvidia_driver_version',
657
]
658
all_cuda_fields = dynamic_cuda_fields + ['cudnn_version']
635
- all_dynamic_cuda_fields_missing = all(
636
- mutable_dict[field] is None for field in dynamic_cuda_fields)
637
- if TORCH_AVAILABLE and not torch.cuda.is_available() and all_dynamic_cuda_fields_missing:
659
+ all_dynamic_cuda_fields_missing = all(mutable_dict[field] is None
660
+ for field in dynamic_cuda_fields)
661
+ if TORCH_AVAILABLE and not torch.cuda.is_available(
662
+ ) and all_dynamic_cuda_fields_missing:
663
for field in all_cuda_fields:
664
mutable_dict[field] = 'No CUDA'
665
if envinfo.cuda_compiled_version is None:
@@ -647,17 +672,19 @@ def pretty_str(envinfo):
672
mutable_dict = replace_nones(mutable_dict)
673
674
# If either of these are '', replace with 'No relevant packages'
650
- mutable_dict['pip_packages'] = replace_if_empty(mutable_dict['pip_packages'])
651
- mutable_dict['conda_packages'] = replace_if_empty(mutable_dict['conda_packages'])
675
+ mutable_dict['pip_packages'] = replace_if_empty(
676
+ mutable_dict['pip_packages'])
677
+ mutable_dict['conda_packages'] = replace_if_empty(
678
+ mutable_dict['conda_packages'])
679
680
# Tag conda and pip packages with a prefix
681
# If they were previously None, they'll show up as ie '[conda] Could not collect'
682
if mutable_dict['pip_packages']:
656
- mutable_dict['pip_packages'] = prepend(mutable_dict['pip_packages'],
657
- '[{}] '.format(envinfo.pip_version))
683
+ mutable_dict['pip_packages'] = prepend(
684
+ mutable_dict['pip_packages'], '[{}] '.format(envinfo.pip_version))
685
if mutable_dict['conda_packages']:
659
- mutable_dict['conda_packages'] = prepend(mutable_dict['conda_packages'],
660
- '[conda] ')
686
+ mutable_dict['conda_packages'] = prepend(
687
+ mutable_dict['conda_packages'], '[conda] ')
688
mutable_dict['cpu_info'] = envinfo.cpu_info
689
return env_info_fmt.format(**mutable_dict)
690
@@ -671,18 +698,22 @@ def main():
698
output = get_pretty_env_info()
699
print(output)
700
674
- if TORCH_AVAILABLE and hasattr(torch, 'utils') and hasattr(torch.utils, '_crash_handler'):
701
+ if TORCH_AVAILABLE and hasattr(torch, 'utils') and hasattr(
702
+ torch.utils, '_crash_handler'):
703
minidump_dir = torch.utils._crash_handler.DEFAULT_MINIDUMP_DIR
704
if sys.platform == "linux" and os.path.exists(minidump_dir):
677
- dumps = [os.path.join(minidump_dir, dump) for dump in os.listdir(minidump_dir)]
705
+ dumps = [
706
+ os.path.join(minidump_dir, dump)
707
+ for dump in os.listdir(minidump_dir)
708
+ ]
709
latest = max(dumps, key=os.path.getctime)
710
ctime = os.path.getctime(latest)
680
- creation_time = datetime.datetime.fromtimestamp(ctime).strftime('%Y-%m-%d %H:%M:%S')
711
+ creation_time = datetime.datetime.fromtimestamp(ctime).strftime(
712
+ '%Y-%m-%d %H:%M:%S')
713
msg = "\n*** Detected a minidump at {} created on {}, ".format(latest, creation_time) + \
714
"if this is related to your bug please include it when you file a report ***"
715
print(msg, file=sys.stderr)
716
717
686
-
718
if __name__ == '__main__':
719
main()
csrc/punica/bgmv/generator.py
+1
-1
@@ -10,7 +10,7 @@ TEMPLATE = """
10
#include "bgmv_impl.cuh"
11
12
FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, {input_dtype}, {output_dtype}, {weight_dtype})
13
-""".lstrip()
13
+""".lstrip() # noqa: E501
14
15
for input_dtype in DTYPES:
16
for output_dtype in DTYPES:
examples/multilora_inference.py
+37
-32
@@ -1,5 +1,6 @@
1
"""
2
-This example shows how to use the multi-LoRA functionality for offline inference.
2
+This example shows how to use the multi-LoRA functionality
3
+for offline inference.
4
5
Requires HuggingFace credentials for access to Llama2.
6
"""
@@ -16,7 +17,7 @@ def create_test_prompts(
17
lora_path: str
18
) -> List[Tuple[str, SamplingParams, Optional[LoRARequest]]]:
19
"""Create a list of test prompts with their sampling parameters.
19
-
20
+
21
2 requests for base model, 4 requests for the LoRA. We define 2
22
different LoRA adapters (using the same model for demo purposes).
23
Since we also set `max_loras=1`, the expectation is that the requests
@@ -34,36 +35,40 @@ def create_test_prompts(
35
top_k=5,
36
presence_penalty=0.2,
37
max_tokens=128), None),
37
- ("[user] Write a SQL query to answer the question based on the table schema.\n\n context: CREATE TABLE table_name_74 (icao VARCHAR, airport VARCHAR)\n\n question: Name the ICAO for lilongwe international airport [/user] [assistant]",
38
- SamplingParams(temperature=0.0,
39
- logprobs=1,
40
- prompt_logprobs=1,
41
- max_tokens=128,
42
- stop_token_ids=[32003]),
43
- LoRARequest("sql-lora", 1, lora_path)),
44
- ("[user] Write a SQL query to answer the question based on the table schema.\n\n context: CREATE TABLE table_name_11 (nationality VARCHAR, elector VARCHAR)\n\n question: When Anchero Pantaleone was the elector what is under nationality? [/user] [assistant]",
45
- SamplingParams(n=3,
46
- best_of=3,
47
- use_beam_search=True,
48
- temperature=0,
49
- max_tokens=128,
50
- stop_token_ids=[32003]),
51
- LoRARequest("sql-lora", 1, lora_path)),
52
- ("[user] Write a SQL query to answer the question based on the table schema.\n\n context: CREATE TABLE table_name_74 (icao VARCHAR, airport VARCHAR)\n\n question: Name the ICAO for lilongwe international airport [/user] [assistant]",
53
- SamplingParams(temperature=0.0,
54
- logprobs=1,
55
- prompt_logprobs=1,
56
- max_tokens=128,
57
- stop_token_ids=[32003]),
58
- LoRARequest("sql-lora2", 2, lora_path)),
59
- ("[user] Write a SQL query to answer the question based on the table schema.\n\n context: CREATE TABLE table_name_11 (nationality VARCHAR, elector VARCHAR)\n\n question: When Anchero Pantaleone was the elector what is under nationality? [/user] [assistant]",
60
- SamplingParams(n=3,
61
- best_of=3,
62
- use_beam_search=True,
63
- temperature=0,
64
- max_tokens=128,
65
- stop_token_ids=[32003]),
66
- LoRARequest("sql-lora", 1, lora_path)),
38
+ (
39
+ "[user] Write a SQL query to answer the question based on the table schema.\n\n context: CREATE TABLE table_name_74 (icao VARCHAR, airport VARCHAR)\n\n question: Name the ICAO for lilongwe international airport [/user] [assistant]", # noqa: E501
40
+ SamplingParams(temperature=0.0,
41
+ logprobs=1,
42
+ prompt_logprobs=1,
43
+ max_tokens=128,
44
+ stop_token_ids=[32003]),
45
+ LoRARequest("sql-lora", 1, lora_path)),
46
+ (
47
+ "[user] Write a SQL query to answer the question based on the table schema.\n\n context: CREATE TABLE table_name_11 (nationality VARCHAR, elector VARCHAR)\n\n question: When Anchero Pantaleone was the elector what is under nationality? [/user] [assistant]", # noqa: E501
48
+ SamplingParams(n=3,
49
+ best_of=3,
50
+ use_beam_search=True,
51
+ temperature=0,
52
+ max_tokens=128,
53
+ stop_token_ids=[32003]),
54
+ LoRARequest("sql-lora", 1, lora_path)),
55
+ (
56
+ "[user] Write a SQL query to answer the question based on the table schema.\n\n context: CREATE TABLE table_name_74 (icao VARCHAR, airport VARCHAR)\n\n question: Name the ICAO for lilongwe international airport [/user] [assistant]", # noqa: E501
57
+ SamplingParams(temperature=0.0,
58
+ logprobs=1,
59
+ prompt_logprobs=1,
60
+ max_tokens=128,
61
+ stop_token_ids=[32003]),
62
+ LoRARequest("sql-lora2", 2, lora_path)),
63
+ (
64
+ "[user] Write a SQL query to answer the question based on the table schema.\n\n context: CREATE TABLE table_name_11 (nationality VARCHAR, elector VARCHAR)\n\n question: When Anchero Pantaleone was the elector what is under nationality? [/user] [assistant]", # noqa: E501
65
+ SamplingParams(n=3,
66
+ best_of=3,
67
+ use_beam_search=True,
68
+ temperature=0,
69
+ max_tokens=128,
70
+ stop_token_ids=[32003]),
71
+ LoRARequest("sql-lora", 1, lora_path)),
72
]
73
74
examples/offline_inference_with_prefix.py
+4
-3
@@ -37,9 +37,10 @@ for output in outputs:
37
38
print("-" * 80)
39
40
-# The llm.generate call will batch all prompts and send the batch at once if resources allow.
41
-# The prefix will only be cached after the first batch is processed, so we need to call generate once
42
-# to calculate the prefix and cache it.
40
+# The llm.generate call will batch all prompts and send the batch at once
41
+# if resources allow. The prefix will only be cached after the first batch
42
+# is processed, so we need to call generate once to calculate the prefix
43
+# and cache it.
44
outputs = llm.generate(generating_prompts[0], sampling_params)
45
46
# Subsequent batches can leverage the cached prefix
setup.py
+15
-10
@@ -12,7 +12,12 @@ import setuptools
12
import sys
13
import torch
14
import torch.utils.cpp_extension as torch_cpp_ext
15
-from torch.utils.cpp_extension import BuildExtension, CUDAExtension, CUDA_HOME, ROCM_HOME
15
+from torch.utils.cpp_extension import (
16
+ BuildExtension,
17
+ CUDAExtension,
18
+ CUDA_HOME,
19
+ ROCM_HOME,
20
+)
21
22
ROOT_DIR = os.path.dirname(__file__)
23
@@ -57,9 +62,8 @@ NVCC_FLAGS = ["-O2", "-std=c++17"]
62
63
if _is_hip():
64
if ROCM_HOME is None:
60
- raise RuntimeError(
61
- "Cannot find ROCM_HOME. ROCm must be available to build the package."
62
- )
65
+ raise RuntimeError("Cannot find ROCM_HOME. "
66
+ "ROCm must be available to build the package.")
67
NVCC_FLAGS += ["-DUSE_ROCM"]
68
NVCC_FLAGS += ["-U__HIP_NO_HALF_CONVERSIONS__"]
69
NVCC_FLAGS += ["-U__HIP_NO_HALF_OPERATORS__"]
@@ -144,7 +148,8 @@ def get_pytorch_rocm_arch() -> Set[str]:
148
"""
149
env_arch_list = os.environ.get("PYTORCH_ROCM_ARCH", None)
150
147
- # If we don't have PYTORCH_ROCM_ARCH specified pull the list from rocm_agent_enumerator
151
+ # If we don't have PYTORCH_ROCM_ARCH specified pull the list from
152
+ # rocm_agent_enumerator
153
if env_arch_list is None:
154
command = "rocm_agent_enumerator"
155
env_arch_list = (subprocess.check_output(
@@ -255,11 +260,11 @@ if _is_cuda():
260
"CUDA 11.1 or higher is required for compute capability 8.6.")
261
if nvcc_cuda_version < Version("11.8"):
262
if any(cc.startswith("8.9") for cc in compute_capabilities):
258
- # CUDA 11.8 is required to generate the code targeting compute capability 8.9.
259
- # However, GPUs with compute capability 8.9 can also run the code generated by
260
- # the previous versions of CUDA 11 and targeting compute capability 8.0.
261
- # Therefore, if CUDA 11.8 is not available, we target compute capability 8.0
262
- # instead of 8.9.
263
+ # CUDA 11.8 is required to generate the code targeting compute
264
+ # capability 8.9. However, GPUs with compute capability 8.9 can
265
+ # also run the code generated by the previous versions of CUDA 11
266
+ # and targeting compute capability 8.0. Therefore, if CUDA 11.8
267
+ # is not available, we target compute capability 8.0 instead of 8.9.
268
warnings.warn(
269
"CUDA 11.8 or higher is required for compute capability 8.9. "
270
"Targeting compute capability 8.0 instead.",