[Chazyl]

Overview

Fixes #27408. Follow-up to #28587.

With --spec-type draft-dflash and an mmproj model, every image request returns HTTP 500
(failed to process speculative batch). #28587 skips M-RoPE-pinned mtmd embedding batches in
the DFlash process() and assumes "the draft can jump over the gap" -- but that gap is not
harmless: the next token batch fails the draft KV cache's consecutive-position check. This
patch detects the per-sequence gap at the top of process() and seeds it with zero-feature
injections. Drafted tokens are still verified by the target, so output stays distribution-exact;
only draft acceptance immediately after an image is degraded.

Additional information

• speculative: fix failed to decode mtmd chunk with DFlash #28587 (merged) added the pos_pinned skip; this is the missing half that fills the resulting hole.
• The approach is due to @dagnarf (dflash: zero-fill draft-cache holes left by mtmd chunks and reused prefixes z-lab/llama.cpp-fork#1); this is the minimal port to upstream.
• Known remaining limitation noted in spec: mtmd image chunks leave positional holes in the draft KV cache -> llama_decode(ctx_dft) rc=-1 -> HTTP 500 with draft-dflash (CUDA); zero-fill patch crash-eliminating but no speculative benefit on images #27408: cross-request position gaps (non-consecutive
  positions spanning slots/requests) are not covered by this patch.
• Tested on 2x RTX 5060 Ti (CUDA, build b11434) with Qwen3.6-35B-A3B-Q4_K_M + mmproj-Q8_0 + a
  DFlash sidecar: image requests go from HTTP 500 to HTTP 200 with no failed to process speculative batch in the log; text-only n=6 quality is unchanged (paired bootstrap P = 0.505)
  and speed is within noise.
• No CI regression test is included: it would require a multimodal target + draft model in CI.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES

[ggml-gh-bot]

Hi @Chazyl, thanks for your contribution!

Per our contribution guidelines, the automated PR checker found the following issue(s) that need your attention:

• PR Template not respected: Please respect the template when creating a new pull request. Make sure to fill out all required sections.

---

Please note that maintainers reserve the right to make final decisions on PRs. If you believe there is a mistake, please comment below.