riaz.somc/transformers
1
0
Fork 0
mirror of https://github.com/huggingface/transformers.git synced 2025-08-03 03:31:05 +06:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

feat: run benchmarks on A100 (#34287)

Browse Source
This commit is contained in:
Luc Georges 2024-10-28 19:33:17 +01:00 committed by GitHub
parent d21dbd1520
commit 6cc4a67b3d
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
3 changed files with 943 additions and 779 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

9
.github/workflows/benchmark.yml vendored
View File

@ -16,8 +16,11 @@ env:
jobs:
benchmark:
name: Benchmark
strategy:
matrix:
group: [aws-g5-4xlarge-cache, aws-p4d-24xlarge-plus]
runs-on:
group: aws-g5-4xlarge-cache
group: ${{ matrix.group }}
if: |
(github.event_name == 'pull_request' && contains( github.event.pull_request.labels.*.name, 'run-benchmark') )||
(github.event_name == 'push' && github.ref == 'refs/heads/main')
@ -60,9 +63,13 @@ jobs:
commit_id=$GITHUB_SHA
fi
commit_msg=$(git show -s --format=%s | cut -c1-70)
df -h
python3 benchmark/llama.py "${{ github.head_ref || github.ref_name }}" "$commit_id" "$commit_msg"
env:
HF_TOKEN: ${{ secrets.HF_HUB_READ_TOKEN }}
# Enable this to see debug logs
# HF_HUB_VERBOSITY: debug
# TRANSFORMERS_VERBOSITY: debug
PGHOST: ${{ secrets.TRANSFORMERS_BENCHMARKS_PGHOST }}
PGUSER: transformers_benchmarks
PGPASSWORD: ${{ secrets.TRANSFORMERS_BENCHMARKS_PGPASSWORD }}

1709
benchmark/grafana_dashboard.json
View File

File diff suppressed because it is too large Load Diff

4
benchmark/llama.py
View File

@ -96,17 +96,21 @@ def run_benchmark(branch: str, commit_id: str, commit_msg: str, num_tokens_to_ge
)
conn.commit()
benchmark_id = cur.fetchone()[0]
logger.info(f"running benchmark #{benchmark_id} on {gpu_name}")
metrics_thread = Thread(target=collect_metrics, args=[benchmark_id, continue_metric_collection])
metrics_thread.start()
logger.info("started background thread to fetch device metrics")
os.environ["TOKENIZERS_PARALLELISM"] = "false" # silence warnings when compiling
device = "cuda"
ckpt = "meta-llama/Llama-2-7b-hf"
logger.info("downloading weights")
# This is to avoid counting download in model load time measurement
model = AutoModelForCausalLM.from_pretrained(ckpt, torch_dtype=torch.float16)
gen_config = GenerationConfig(do_sample=False, top_p=1, temperature=1)
logger.info("loading model")
start = perf_counter()
model = AutoModelForCausalLM.from_pretrained(
ckpt, torch_dtype=torch.float16, generation_config=gen_config