Dependency conflict: causal student net (atten_backend="i4") requires natten≥0.21.5, but pinned install (torch 2.7 + natten 0.21.0) cannot satisfy it

[zihenghe04]

Summary

The release appears to be internally inconsistent. The distilled causal student net (with atten_backend="i4") hard-depends on NATTEN, and the code requires natten>=0.21.5 for the causal varlen API. However, the official install instructions pin torch 2.7 + natten 0.21.0, which does not provide that API. The versions of NATTEN that do ship the API (0.21.5+) only publish prebuilt wheels for torch>=2.8. As a result, there is no combination that satisfies the constraints out of the box.

Details

• The causal student net architecture (atten_backend="i4") is tied to NATTEN.
• The code path uses the causal varlen API, which requires natten>=0.21.5.
• The official installation instructions pin torch 2.7 + natten 0.21.0.
• natten 0.21.0 does not include the causal varlen API, so the pinned stack fails at runtime/import.
• NATTEN 0.21.5+ (which contains the required API) only provides prebuilt wheels for torch>=2.8, so it is incompatible with the pinned torch 2.7.

The contradiction

Requirement	Source	Conflict
natten>=0.21.5 (causal varlen API)	code / atten_backend="i4"	Not met by pinned natten 0.21.0
torch 2.7 + natten 0.21.0	official install instructions	Lacks the required API
natten>=0.21.5 prebuilt wheels	NATTEN release channel	Only built for torch>=2.8, conflicts with pinned torch 2.7

Questions / Requested fix

1. What is the intended, working version combination for running the causal student net with atten_backend="i4"?
2. Should the official install instructions be updated to torch>=2.8 together with natten>=0.21.5?
3. Alternatively, is there a supported way to build natten>=0.21.5 from source against torch 2.7, or a fallback backend that avoids the causal varlen API requirement?

Could you please clarify the correct, self-consistent dependency set? Thanks!