-
Notifications
You must be signed in to change notification settings - Fork 0
/
llm_benchmark_ipu.yaml
315 lines (301 loc) · 13.8 KB
/
llm_benchmark_ipu.yaml
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
name: llm-ipu
outpath: llm_benchmark_ipu_run
comment: llm benchmark jube script for IPU-GC200
# GPT-2 benchmark for
# JURECA-M2000 Graphcore Node:
# The IPU-POD4 consists of two parts:
# AMD EPYC based access server on which user applications are launched with
# CPU: AMD EPYC 7413 (Milan); 2 sockets, 24 cores per socket, SMT-2 (total: 2×24×2 = 96 threads) in NPS-4 1 configuration (details for AMD EPYC 7413 on WikiChip)
# Memory: 512 GiB DDR4-3200 RAM (of which at least 20 GB is taken by the system software stack, including the file system); 256 GB per socket; 8 memory channels per socket (2 channels per NUMA domain)
# Network: 1 × Mellanox EDR InfiniBand ConnectX 5 (100 Gbit/s) to connect to other compute nodes and 1 × Mellanox 100 GigE ConnectX 5 to connect to the IPU-M2000
# Graphcore IPU-M2000 which is connected directly to the access server with
# IPUs: 4 × GC200 IPUs
# Available: 1 node
parameterset:
- name: systemInfo
parameter:
- {name: system_name, type: str, _: "GC200"}
- {name: system_version, mode: shell, _: "echo 2024.01"}
- name: modelParameter
parameter:
- {name: llm_model_size, type: str, tag: "117M", _: "117M"}
- {name: modelidx, type: int, mode: python, _: "{'117M': '0'}['$llm_model_size']"}
- {name: model_type, type: str, _: "GPT"}
- {name: model, type: str, _: "gpt2"}
- {name: nhidden, type: int, mode: python, _: "[768][$modelidx]"}
- {name: nlayers, type: int, mode: python, _: "[12][$modelidx]"}
- {name: nheads, type: int, mode: python, _: "[12][$modelidx]"}
- {name: dataset, type: str, tag: "!synthetic", _: "OSCAR"}
- {name: dataset, type: str, tag: "synthetic", _: "Synthetic"}
- {name: dataformat, type: str, tag: "!synthetic", _: "'mmap'"}
- {name: dataformat, type: str, tag: "synthetic", _: "'generated'"}
- {name: datadir, type: str, tag: "!synthetic", _: "$jube_benchmark_home/llm_data/oscar_text_document"}
- {name: replication_fact, type: int, seperator: ",", _: "1"}
- {name: ipus_per_replica, type: int, seperator: ",", _: "4"}
- {name: layers_per_ipu, type: str, seperator: ",", _: "0 4 4 4"}
- {name: matmul_prop, type: str, seperator: ",", _: "0.15 0.15 0.15 0.15"}
- {name: device_iter, type: int, seperator: ",", _: "4"}
- {name: num_instances, type: int, seperator: ",", _: "1"}
- {name: embedding_serial_fact, type: int, seperator: ",", _: "4"}
- {name: total_devices, type: int, mode: python, _: "${replication_fact}*${ipus_per_replica}*${num_instances}"}
- {name: batch_size, type: int, seperator: ",", _: "1"}
# gas: 16,32,64,128,256,512,1024,2048,4096,8192,16384
- {name: gas, type: int, seperator: ",", _: "128"}
- {name: global_batch_size, type: int, mode: python,
_: "int($num_instances*$replication_fact*$batch_size*$gas*$device_iter)"}
- {name: sequence_length, type: int, _: 1024}
- name: optimizerargs
mode: text
_: >
--optimizer AdamW
--learning-rate 0.00015
--lr-schedule cosine
--lr-warmup 0.01
- name: parallelargs
_: >
--replication-factor ${replication_fact}
--ipus-per-replica ${ipus_per_replica}
--layers-per-ipu ${layers_per_ipu}
--matmul-proportion ${matmul_prop}
--device-iterations ${device_iter}
--embedding-serialization-factor ${embedding_serial_fact}
--replicated-tensor-sharding True
--recompute-checkpoint-every-layer True
--remap-logit True
--enable-sequence-serialized True
- name: dataargs
tag: "synthetic"
_: >
--dataset ${dataformat}
- name: dataargs
tag: "!synthetic"
_: >
--dataset ${dataformat}
--input-files ${datadir}
- name: modelargs
_: >
--model ${model}
--epochs 1
$dataargs
--batch-size ${batch_size}
--gradient-accumulation ${gas}
--max-len ${sequence_length}
--enable-half-partials True
$parallelargs
$optimizerargs
- name: args_executable
mode: text
tag: "!container"
separator: |
_: $modelargs
- name: systemParameter
init_with: platform.xml
parameter:
- {name: account, _: "zam"}
- {name: queue, _: "dc-ipu"}
- {name: nodes, type: int, _: 1}
- {name: gpus_per_node, type: int, tag: "!container", _: 4}
- {name: gpus_per_node, type: int, tag: "container", _: 1}
- {name: taskspernode, type: int, _: 1}
- {name: tasks, type: int, mode: python, _: "$taskspernode*$nodes"}
- {name: threadspertask, type: int, mode: python, tag: "!container" ,_: 12}
- {name: threadspertask, type: int, mode: python, tag: "container", _: 4}
- {name: timelimit, tag: "container", _: "01:30:00"}
- {name: timelimit, tag: "!container", _: "00:40:00"}
- {name: hint, _: "nomultithread"}
- {name: outlogfile, _: "job.out"}
- {name: outerrfile, _: "job.err"}
- {name: ready_file, _: "ready"}
- {name: error_file, _: "error"}
- {name: root_dir, type: str, _: "$jube_benchmark_home/.."}
# https://hub.docker.com/layers/graphcore/pytorch/3.3.0-ubuntu-20.04-20230703/images/sha256-7f65b5ff5bdc2dad3c112e45e380dc2549113d3eec181d4cf04df6a006cd42a4?context=explore
# pytorch:3.3.0-ubuntu-20.04-20230703
- {name: singularity_file, type: str,
_: "$root_dir/containers/ipu_pytorch20_poplar33_py38.sif"}
- name: variable_export
mode: shell
separator: "!"
_: |
echo "export BENCH_DIR=${jube_benchmark_home}"
echo "export ACCELERATOR=${system_name}"
echo "export ENERGY_PATH=${system_name}_power.csv"
- {name: measurement, _: $variable_export}
- name: container_env
mode: text
separator: "!"
_: |
POPLAR_ROOT=\$(cd \"\$( dirname \"\${BASH_SOURCE[0]}\" )\" && pwd)
export CMAKE_PREFIX_PATH=\${POPLAR_ROOT}\${CMAKE_PREFIX_PATH:+:\${CMAKE_PREFIX_PATH}}
export PATH=\${POPLAR_ROOT}/bin\${PATH:+:\${PATH}}
export CPATH=\${POPLAR_ROOT}/include\${CPATH:+:\${CPATH}}
export LIBRARY_PATH=\${POPLAR_ROOT}/lib\${LIBRARY_PATH:+:\${LIBRARY_PATH}}
export LD_LIBRARY_PATH=\${POPLAR_ROOT}/lib\${LD_LIBRARY_PATH:+:\${LD_LIBRARY_PATH}}
export OPAL_PREFIX=\${POPLAR_ROOT}
export PYTHONPATH=\${POPLAR_ROOT}/python:\${POPLAR_ROOT}/lib/python\${PYTHONPATH:+:\${PYTHONPATH}}
export POPLAR_SDK_ENABLED=\${POPLAR_ROOT}
POPART_ROOT=\$(cd \"\$( dirname \"\${BASH_SOURCE[0]}\" )\" && pwd)
export CMAKE_PREFIX_PATH=\${POPART_ROOT}\${CMAKE_PREFIX_PATH:+:\${CMAKE_PREFIX_PATH}}
export CPATH=\${POPART_ROOT}/include\${CPATH:+:\${CPATH}}
export LIBRARY_PATH=\${POPART_ROOT}/lib\${LIBRARY_PATH:+:\${LIBRARY_PATH}}
export LD_LIBRARY_PATH=\${POPART_ROOT}/lib\${LD_LIBRARY_PATH:+:\${LD_LIBRARY_PATH}}
export PYTHONPATH=\${POPART_ROOT}/python:\$PYTHONPATH
- {name: packages, _: $jube_benchmark_home/ipu_torch_packages/lib/python3.8/site-packages}
- name: executable
separator: ";"
tag: "!container"
_:
python3 ${folder}/train_gpt2.py $args_executable
- name: executable
separator: "!"
tag: "container"
_:
bash ${jube_benchmark_home}/get_pytorch_container.sh
- name: preprocess
mode: text
tag: "!container"
update_mode: step
separator: |
_: |
echo "START TIME: $(date)"
echo "Submitted batch job $${SLURM_JOBID}"
export SRUN_CPUS_PER_TASK=$${SLURM_CPUS_PER_TASK}
export BENCH_DIR=$jube_benchmark_home
mkdir -p $jube_benchmark_home/ipu_gc200_cache
if [ "x$$BENCH_DIR" = "x" ]; then
echo "BENCH_DIR is not set. Please set it to the llm_training directory of benchmark" >&2
exit 1
fi
export POPLAR_LOG_LEVEL=INFO
export POPART_LOG_LEVEL=INFO
export TF_POPLAR_FLAGS='--executable_cache_path=$jube_benchmark_home/ipu_gc200_cache'
- name: executeset
init_with: platform.xml
parameter:
- name: folder
_: $jube_benchmark_home/$jube_wp_relpath/graphcore_benchmarks/nlp/gpt2/pytorch
- name: args_starter
mode: text
tag: "!container"
separator: |
_: |
--cpu_bind=none,v --mpi=pspmix env PMIX_SECURITY_MODE=native apptainer exec --bind=${jube_benchmark_home} ${singularity_file} ${jube_benchmark_home}/ipu_torch_wrap.sh
fileset:
- name: files
prepare: git clone -b jpwr_integration --recursive https://github.com/chelseajohn/examples.git graphcore_benchmarks
patternset:
- name: perf_patterns
pattern:
- {name: jobid, type: int, _: "Submitted batch job $jube_pat_int" }
# https://github.com/graphcore/examples/blob/3cc8f81f13728dd8c17d0d3c8d2fc1549b159485/nlp/gpt2/pytorch/train_gpt2.py#L271
- {name: step_time, type: float, _: "step_time: ${jube_pat_fp}"}
- {name: step_throughput, type: float, _: "throughput: ${jube_pat_fp}"}
- {name: tokens_per_second, type: float, mode: python, _: "($global_batch_size/$step_time_last)"}
- {name: throughput_in_time, type: float, mode: python, _: "(20000000/$tokens_per_second)"}
# - {name: energy_average, type: float, _: "Average-Energy-per-GPU(Wh): ${jube_pat_fp}"}
# - {name: energy_list_reimann,type: str, _: "Energy-per-GPU-list integrated\\(Wh\\): (\\[.*?\\])"}
- name: efile_patterns
pattern:
- {name: energy_file_path, type: str, _: "Writing combined energy DataFrame to (.*)"}
analyser:
- name: analyse
reduce: false
analyse:
- step: combine_energy
file:
{use: efile_patterns, _: stdout}
- step: benchmark
file:
{use: perf_patterns, _: job.err}
result:
use: analyse
table:
name: result
style: csv
sort: iter_pat
column:
- {title: "JobID", _: jobid}
- {title: "System", _: system_name}
- {title: "Version", _: system_version}
- {title: "Queue", _: queue}
- {title: "JobTime", _: timelimit}
- {title: "Model", _: model_type}
- {title: "ModelSize", _: llm_model_size}
- {title: "Dataset", _: dataset}
- {title: "Nodes", _: nodes}
- {title: "Devices", _: total_devices}
- {title: "DataParallel", _: replication_fact}
- {title: "IPU/replica", _: ipus_per_replica}
- {title: "GlobalBatchSize", _: global_batch_size}
- {title: "Time/Iteration(s)", format: ".2f", _: step_time_last}
# - {title: "StepThroughput(tokens/s)",format: ".2f", _: step_throughput_last}
# - {title: "Tokens/second",format: ".2f", _: tokens_per_second}
- {title: "Tokens/second",format: ".2f", _: step_throughput_last}
# - {title: "time_to_report_in_seconds",format: ".2f", _: throughput_in_time}
# - {title: "Energy/Device(Wh)", format: ".2f", _: energy_average}
# - {title: "Energy/Device(Wh)", _: energy_list_reimann}
- {title: "EnergyFile", _: energy_file_path}
step:
- name: wrap
tag: "!container"
use:
- systemInfo
- systemParameter
- modelParameter
do:
mode: shell
_: |
printf "%s\n" '#!/bin/bash' "${container_env}" "export PYTHONPATH=${packages}:\$PYTHONPATH" "\$*" > ${jube_benchmark_home}/ipu_torch_wrap.sh
chmod u+rwx ${jube_benchmark_home}/ipu_torch_wrap.sh
- name: wrap
tag: container
use:
- systemInfo
- systemParameter
do:
mode: shell
_: |
printf "%s\n" '#!/bin/bash' "${container_env}" "\$*" > ${jube_benchmark_home}/ipu_torch_wrap.sh
chmod u+rwx ${jube_benchmark_home}/ipu_torch_wrap.sh
- name: get_container
tag: "container"
depend: wrap
use:
- systemInfo
- systemParameter
- executeset
- from: platform.xml
_: jobfiles
- from: platform.xml
_: executesub
do:
done_file: $ready_file
error_file: $error_file
_: $submit $submit_script
- name: benchmark
tag: "!container"
depend: wrap
use:
- files
- systemInfo
- systemParameter
- modelParameter
- executeset
- from: platform.xml
_: jobfiles
- from: platform.xml
_: executesub
do:
done_file: $ready_file
error_file: $error_file
_: $submit $submit_script
- name: combine_energy
tag: "!container"
iterations: 1
depend: benchmark
use:
- systemInfo
- systemParameter
- modelParameter
do:
- module load GCC SciPy-bundle
- python3 ${jube_benchmark_home}/aux/combine_energy.py --output ${jube_benchmark_home}/${jube_wp_relpath}/combined_energy.csv $$(printf "${jube_benchmark_home}/${jube_wp_relpath}/benchmark/${system_name}_power.%d.energy.csv " $$(seq 0 $$((${tasks}-1))))