Hugging Face's logo

flashrt
/
flashrt-flex-attention-train

Model card Files
xet
 Community
flashrt-flex-attention-train / VALIDATION.md
liangsu9988's picture
liangsu9988
Add package source and torch-universal build variant
b22e03e verified
preview code
|
Raw
History Blame Contribute Delete
1.72 kB

Validation

Correctness smoke:

python flashrt-flex-attention-train/tests/test_flashrt_flex_attention_train.py --backend source --mode smoke
python flashrt-flex-attention-train/tests/test_flashrt_flex_attention_train.py --backend source --mode full

Installed artifact smoke:

python flashrt-flex-attention-train/tests/test_flashrt_flex_attention_train.py --backend installed --mode full
python scripts/prebuild_check.py --package flashrt-flex-attention-train --check-config

Microbenchmark and acceptance gates:

python flashrt-flex-attention-train/benchmarks/benchmark.py --device cuda --dtype bf16 --mode all --output /tmp/flex_attention_a100.json
python flashrt-flex-attention-train/benchmarks/benchmark.py --device cuda --dtype bf16 --mode all --require-gates

Shape/tile matrix:

python flashrt-flex-attention-train/benchmarks/shape_matrix.py --presets a100 --output /tmp/a100_flex_matrix.jsonl
python flashrt-flex-attention-train/benchmarks/shape_matrix.py --presets consumer --output /tmp/consumer_flex_matrix.jsonl

Minimum gate for connecting native kernels:

  • forward time <= 0.95 * SDPA
  • forward+backward time <= 0.95 * SDPA
  • peak memory <= 1.03 * SDPA
  • forward max abs diff <= 2e-3
  • gradient norm relative diff <= 1e-2

Higher gate for publishing native kernels:

  • full PI052 step speedup >= 1.08x on A100 and RTX 5090
  • isolated attention forward+backward speedup >= 1.25x
  • no text/flow step regression
  • unsupported shapes automatically fall back to SDPA

The current package provides the reference/fallback path and therefore is not expected to pass the performance gates until native CUDA fwd/bwd kernels are added behind the same API.