Feature: k2_horizon_uno decode is about 20% slower than needed (no CUDA graph; LoRA branch computed on rows whose mask is 0) #2054

Request type: Improvement to existing behavior

Problem and use case

families/k2_horizon_uno runs speculative decoding with one dynamic engine that takes 1 to 8 rows per call and a per-row LoRA mask (draft rows mask 1; prompt and verify rows mask 0). Two costs on main, found during the FP8 study recorded in families/k2_horizon_uno/README.md (#2039, section 7):

1. The Uno native runtime never enables CUDA graphs. A bf16 8-row call launches about 830 kernels; the host loop spends about 6 ms per 8-row call against 5.3 to 5.7 ms when the same call is replayed as a CUDA graph.
2. The LoRA branch is computed for every row of every call, including prompt and verify calls whose mask is all 0. The branch (rank 128 on all 7 projections, in bf16) is 19% to 23% of a call: the bf16 engine without the branch is 1.21x to 1.24x faster.

Both are larger than what FP8 gives on this family (FP8 was measured and not shipped, see the README).

Proposed outcome

1. Replay the decode calls as CUDA graphs in the Uno native runtime (per row count used).
2. Skip the LoRA branch when the mask of a call is all 0, for example with a second engine profile (or engine) without LoRA for prompt and verify calls.

Model and target details

IFM/K2-Horizon-7B-Uno @ ec92bbd768f4a404319625204544782e3377bcd7 (the k2-horizon-7b-uno manifest, linear_spec_lora mode) on the K2-Horizon-7B base; bf16; TensorRT 11.4; B200 / B300; max_sequence_length 1024.

Alternatives considered

FP8 of the base projections (measured, not shipped): the 8-row call is 0.96x (B200) and 0.88x (B300) of bf16 without a CUDA graph, because the bf16 LoRA branch and the kernel count stay.

Additional context

The same study found that speculative decode equals the engine's own greedy autoregressive output on 13 of 14 prompts only (bf16): 8-row and 1-row calls differ in the last bits, which changes the argmax at near ties.