[Qubitium]

Follow-up to #3545 and #3549 (both merged). Two defects found in review of the mirrored HF patch.

1. Preallocated caches broke causal attention

StaticCache layers return their full preallocated capacity from update(), so k.shape[2] is capacity, not occupied context. A 6-token prefill into a 16-slot cache was therefore treated as a 10-token prefix: the end-anchored verify mask gave query positions 10-15, letting each query attend slots of a different absolute position. Changing token 6 changed token 1's output (measured max diff 0.4830), which is impossible under causal attention.

Fix: read the layer's real length via Cache.get_seq_length(layer_idx), trim the unwritten tail of k/v, build the mask from cache_position and the absolute key positions (trailing window for evicting caches), and align an incoming mask that is longer than the keys actually present.

2. labels + logits_to_keep=1 returned NaN loss

The patch trimmed hidden states and labels to one position, then the causal shift produced empty tensors and cross_entropy returned NaN.

Fix: when labels are present, keep the full sequence so the shifted loss is defined, and trim only the returned logits. Unlabeled calls still slice the LM-head input, so the [batch, seq, vocab] projection is never materialised during generation.

Validation

• tests/test_ricursive_support.py gains test_static_cache_uses_actual_positions_not_capacity and test_labels_with_logits_to_keep_keeps_finite_loss; 7 focused tests pass on CPU.
• Tiny-fixture GPU run (original upstream modeling file, so these patches are exercised): StaticCache causality diff 0.0 (was 0.4830), StaticCache warm block 0.0, labels+logits_to_keep=1 loss 5.562519 (full 5.562519, finite), logits_to_keep=1 shape (1,1,256) and max diff 8.9e-08, cached == uncached generation, bf16 causality 0.0, peak test allocation 9.4 MiB.

[Qubitium]

Full-model verification of this branch: Ripple-65B-Preview GPTQ W4A16 (Marlin), single RTX 4090, bf16, 40.2 GiB peak, using the native modeling_ricursive.py plus these load-time patches.

Scenario	Result
StaticCache causality: 6-token prefill into a 16-slot cache, change token 6	max diff on tokens 1-5 = 0.0
Warm verify block, DynamicCache, N=48 + M=8 vs one 56-token forward	max abs 0.25 (rel 1.4e-2)
Warm verify block, StaticCache 64 slots	max abs 0.25
Control: single-token decode step vs same full-forward row (no block mask involved)	max abs 0.156 (bf16/Marlin shape-noise floor)
Pre-fix mask emulation (top-left SDPA causal)	max abs 32.31 (rel 1.76), argmax correct 0/8
labels + logits_to_keep=1	loss 16.5 = full 16.5, logits (1,1,100352)
logits_to_keep=1	(1,1,100352), argmax matches full last row
cached vs uncached greedy generate	identical
Chunked MLP, RIPPLE_MLP_CHUNK=64 vs 0	torch.equal=True
DFlash2 dflash_generate, block_size=1 vs AR	0 divergences
DFlash2 default block vs AR	first divergence at token 1 is an exact bf16 tie (top-2 both 18.0, gap 0.0); accept_mean 2.56

The block-path error (0.25) is ~1.4x the no-mask control (0.156) and ~130x smaller than the pre-fix 32.31, i.e. the verify mask is correct and the residual is bf16 kernel shape noise. The DFlash token mismatch is a genuine tie at bf16 precision, not a masking error.

Re-ran the same suite with apply_ricursive_generate_patch disabled (patches all report False), so only the native HF file is active: all pass with the same numbers.