ViperEkura/AstrAI
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

refactor: dedupe fp8 meta state into per-operand rings

Browse Source 
- collapse FP8TensorMeta's 12 slots + 6 copy-paste methods into three _ScaleRing objects (hist/idx/scale/initialized + update/seed)
- skip meta allocation entirely on the DynamicScaling path (zero rings, scales measured inline)
- drop write-only FP8State._last_device and unused E4M3_MAX alias
This commit is contained in:
ViperEkura
2026-08-24 21:19:45 +08:00
parent cebdd45d3a
commit 998b443aa3
2 changed files with 49 additions and 89 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/extension/test_fp8_mma.py
+5 -5
View File
@@ -334,11 +334,11 @@ def test_fp8_tensor_meta_delayed_update():
"""Meta seeds from data and refreshes the scale from the amax ring."""
meta = FP8TensorMeta(torch.device("cpu"), DelayedScaling(history_len=4, margin=0))
w = torch.randn(8, 8)
meta.init_w(w, "e4m3")
assert meta.w_init
torch.testing.assert_close(meta.w_scale, (w.abs().amax() / 448.0).reshape(1))
meta.update_w(torch.tensor([4.0]), "e4m3")
torch.testing.assert_close(meta.w_scale, torch.tensor(4.0 / 448.0).reshape(1))
meta.w.seed(w, "e4m3")
assert meta.w.initialized
torch.testing.assert_close(meta.w.scale, (w.abs().amax() / 448.0).reshape(1))
meta.w.update(torch.tensor([4.0]), "e4m3")
torch.testing.assert_close(meta.w.scale, torch.tensor(4.0 / 448.0).reshape(1))
def test_quantize_bf16_cpu_fallback():