AI-Hypercomputer / AI-Hypercomputer/maxtext
bug: Numeric divergence between the mHC layer and its Pallas kernel
- Vorherrschende Sprache
- Python
- Sterne
- 2.4k
- Forks
- 607
- Ø Merge
- 2 T. 19 Std.
- Gemergte PRs (30 T.)
- 158
Beschreibung
### Bug report
DeepSeek v4. The reference path had its own mapping helper that cast alpha_scale and
beta down to self.dtype before computing sigmoid(alpha_scale * h + beta),
while the Pallas path called common.compute_sigmoid_gate, which keeps the
scale and bias in float32. With dtype: bfloat16 the two branches therefore
gated on different numbers, so toggling use_mhc_pallas_kernel changed the
layer's output beyond accumulation noise.
Fixed in PR:
https://github.com/AI-Hypercomputer/maxtext/pull/5179
### Logs/Output
_No response_
### Environment Information
_No response_
### Additional Context
_No response_
Beitragsleitfaden
Bewertung
Dieses Issue wurde noch nicht bewertet.