[FEATURE REQUEST] Optional mask & bias inputs with adaptive computation skipping

[LoserCheems]

Is your feature request related to a problem?
Current kernels always assume both attn_mask and attn_bias are active. This causes:

• Unnecessary global memory loads when only one is needed.
• needless dbias computation when no bias is conceptually required.
• For bias-only usage we still fake a full mask or disable skipping logic inefficiently.

Describe the solution you'd like
Make both attn_mask and attn_bias optional with 4 explicit modes:

Case	attn_mask	attn_bias	Behavior
A	None	None	Dense path, no block skip, no bias load/add, fastest
B	Tensor	None	Block skip using mask, no bias add/dbias
C	None	Tensor	No block skip (all blocks active), add bias + compute dbias
D	Tensor	Tensor	Current behavior (mask skip + bias add + dbias)

Rules:

• Only load tensors that are not None.
• Only compute / return dbias if bias provided.
• Block skip only when mask is present (never infer from bias alone).

Describe alternatives you've considered

• Forcing dummy all-ones mask (bandwidth waste).
• Sentinel zero-sized tensors (confusing, still conditional branches).
• Mode enum instead of optional tensors (less ergonomic for users).

Implementation details
Kernel / Launch:

• Add flags: use_mask, use_bias.
• If use_mask: load mask → build skip map (OR reduction per block).
• If !use_mask: all blocks active (skip map implicit true).
• If use_bias: load bias (subject to skip only if mask present).
  Forward:
• Guard bias addition with if (use_bias).
  Backward:
• If use_bias: compute/store dbias; else skip.
• Reuse existing skip logic only when use_mask.
  Performance:
• Case A removes both memory streams + skip logic overhead.
• Case B removes bias path (saves reads/writes + math).
• Case C removes mask loads / OR reductions (simpler control flow).
  Autograd API:
• Gradient for absent tensors must be None (not zeros).
  Python wrapper:
• Accept attn_mask: Optional[Tensor] = None, attn_bias: Optional[Tensor] = None.
  Testing:

1. Case A ≈ Case D with all-ones mask & zero bias.
2. Case B ≈ Case D with zero bias (and no dbias returned).
3. Case C ≈ Case D with all-ones mask (no skip) same bias.
4. Case D unchanged baseline regression tests.
5. Gradient checks: dbias only when bias passed.
   Docs:

• Add mode table + guidance on when to omit tensors.
• Note: omitting mask disables block skipping even if sparsity exists.

Describe alternatives you've considered (API level)

• Separate functions (explodes surface area).
• Single enum “mode” (harder to mix future extensions).

Use case

• Dense long-context causal inference (no mask/bias).
• Relative position / ALiBi style bias-only (no sparsity).
• Structured sparsity pruning (mask-only).
• Combined sparse + bias training (both).

Related work

• FlashAttention dense fast path patterns.
• Sparse block skipping in dynamic mask attention.
• Relative position bias (bias-only adopters).

Why it’s valuable
Eliminates wasted DRAM traffic and unnecessary gradient work; improves throughput and reduces latency for large sequence lengths and specialized training/inference regimes.