spec: keep DFlash vision positions contiguous

[Sphinx-8]

Problem

DFlash2 uses a one-dimensional draft KV cache, while multimodal target batches can carry multi-dimensional M-RoPE positions with repeated or jumping values. Copying those positions into the draft batch makes the draft sequence non-consecutive and vision requests fail during llama_decode.

A second mismatch appears after multimodal prefill: dp.n_past is measured on the target sequence, but the text-only draft cache advances on its own position scale.

Change

• Assign injected target features strictly increasing positions per sequence, beginning at the current draft-cache maximum plus one.
• Start each speculative draft block at the actual next draft-cache position instead of the target token count.

This keeps target positions untouched. It is intentionally narrow and is offered for comparison with the zero-fill approach in #1. The initial dense-position direction was also discussed in ggml-org#27342.

Validation

Base: z-lab dflash2 at 1deefcc.

Hardware/runtime: DGX Spark GB10, CUDA; Qwen3.8 27B target BF16, DFlash2 draft BF16, F16 vision projector, speculative depth 4.

Measured vision requests, each cell after three warmups:

Prompt/context	Parallelism	Result	Mean TTFT	Mean decode	Acceptance
Native image, about 1k prompt tokens	1	10/10	2.50 s	12.32 tok/s	62.66%
Native image, about 1k prompt tokens	2	10/10	4.32 s	11.61 tok/s	62.34%
Image plus about 7k text tokens	1	10/10	11.02 s	11.36 tok/s	60.23%
Image plus about 7k text tokens	2	10/10	19.85 s	8.19 tok/s	60.27%
Image plus about 129.5k text tokens	1	10/10	237.08 s	8.72 tok/s	64.85%

No request errors occurred across the 50 measured requests. A separate two-slot, 262144-context-per-slot smoke test passed 2/2 concurrent vision requests with active DFlash2.

The patched current branch also builds llama-server successfully on aarch64 with GCC 13.3.0. Patch SHA-256 from the validated deployment: ad8e244316fb51c088258d532f271c32e15289045605cd2c5eb093f13f8ef91f.

Disclosure

Implementation and PR text were prepared with OpenAI Codex under Sphinx-8 direction. All reported measurements came from actual runs on the stated hardware; no results are inferred.