diff options
| author | Yuren Hao <yurenh2@illinois.edu> | 2026-07-06 09:17:36 -0500 |
|---|---|---|
| committer | Yuren Hao <yurenh2@illinois.edu> | 2026-07-06 09:17:36 -0500 |
| commit | 9a8b2796ca12e4d4c24717485a635a301aa6d07f (patch) | |
| tree | 550c7a60705c6f02976da27b86184faaac70b891 /ep_run/agg_bench.log | |
| parent | 488c50e1bdbf8f420b2ad5b4021a7f950d121835 (diff) | |
aggregate speed bench: speed tier hf+sd 1.47x; accuracy tier hf+sd+t80+avg (0.91x, cos 0.94, avg free); compile demoted
Full-ep_step wall times on quiet A6000 (warm s2000, B24), res parity across
all 8 configs. compile only 1.12x at this shape (historical 1.46x was a
different workload split); FULL cmp_sdpa saves 4% over eager at t80 — not
worth the guard complexity. tforce_sdpa added (flash baked into compiled
graph, flag-free so grad paths never see SDPA). bp_lm --tie probe in flight.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
Diffstat (limited to 'ep_run/agg_bench.log')
| -rw-r--r-- | ep_run/agg_bench.log | 13 |
1 files changed, 13 insertions, 0 deletions
diff --git a/ep_run/agg_bench.log b/ep_run/agg_bench.log new file mode 100644 index 0000000..35cb799 --- /dev/null +++ b/ep_run/agg_bench.log @@ -0,0 +1,13 @@ +/home/yurenh2/miniconda3/lib/python3.13/site-packages/torch/autograd/graph.py:865: UserWarning: Attempting to run cuBLAS, but there was no current CUDA context! Attempting to set the primary context... (Triggered internally at /pytorch/aten/src/ATen/cuda/CublasHandlePool.cpp:330.) + return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass + base: 4.50s/step (res 8.2e-03) + hf: 3.49s/step (res 8.2e-03) + sd: 4.07s/step (res 8.2e-03) + hf+sd: 3.06s/step (res 8.2e-03) + hf+sd+t80: 4.97s/step (res 8.2e-03) + hf+sd+t80+avg: 4.97s/step (res 8.2e-03) + cmp: 4.03s/step (res 8.2e-03) +/home/yurenh2/miniconda3/lib/python3.13/site-packages/torch/_inductor/compile_fx.py:321: UserWarning: TensorFloat32 tensor cores for float32 matrix multiplication available but not enabled. Consider setting `torch.set_float32_matmul_precision('high')` for better performance. + warnings.warn( + FULL(cmp_sdpa): 4.77s/step (res 8.2e-03) +AGG_BENCH_DONE |