@Deep research Conduct a fresh technical investigation for new features or improvements that
  could be added to the Eaman llama.cpp patch for this system:

  - RX 6800 16 GiB + RX 6700 XT 12 GiB
  - ROCm/HIP, with Vulkan also relevant for regression compatibility
  - asymmetric fixed multi-GPU placement
  - Qwen3.6-27B with native MTP/speculative decoding
  - priority: greater usable context, lower VRAM use, better MTP efficiency,
    improved HIP performance, and long-session correctness

  Use the attached files as the baseline:

  1. summary_eaman.md is the authoritative current project state.
  2. status is the short current operational state.
  3. improvemets_list.md describes features already present in the Eaman patch.
  4. web_search.md is the earlier research performed on 2026-08-07.
  5. AGENTS.md defines current terminology and project constraints.

  Do not simply repeat proposals already covered in web_search.md. Re-evaluate them
  only if there has been a material upstream development since 2026-08-07.

  Search current sources through today, prioritizing:

  - recent and active ggml-org/llama.cpp pull requests, issues, discussions,
    commits, release notes, and merged changes;
  - ROCm/HIP and AMD GPU repositories or documentation;
  - relevant work in comparable llama.cpp forks only when it contains concrete
    code or measurements that could realistically be ported;
  - MTP/speculative-decoding implementations in other inference engines when
    they expose a transferable memory-management or scheduling technique.

  Investigate these areas particularly:

  - MTP rollback/checkpoint state memory and precision;
  - speculative decoding acceptance, scheduling, batching, and multi-slot use;
  - context fitting and per-device memory accounting;
  - asymmetric multi-GPU tensor placement and pipeline scheduling;
  - HIP flash attention, quantized KV, VEC kernels, and allocation lifetime;
  - ROCm memory pools, VMM, fragmentation, and delayed OOM behavior;
  - GPU sampling operations such as TOP_K;
  - KV-cache formats or compression with realistic HIP support;
  - prompt/decode performance improvements applicable to RDNA2;
  - correctness fixes relevant to long-context MTP sessions.

  For every candidate, determine:

  - source URL and PR/issue/commit number;
  - current state: merged, open, draft, closed/unmerged, or abandoned;
  - date of the latest meaningful activity;
  - what changed technically;
  - whether current llama.cpp mainline already contains it;
  - whether the Eaman patch already contains or supersedes it;
  - likely benefit on this exact dual-RDNA2 configuration;
  - expected context, VRAM, prompt-speed, decode-speed, and correctness effects;
  - likely files/subsystems that would need modification;
  - implementation difficulty and validation requirements;
  - important risks, regressions, or hardware limitations.

  Be skeptical of proposed but unmerged code and unsupported performance claims.
  Distinguish measured facts from your own inference. Cite every time-sensitive
  claim directly to its source. Prefer source code, PR discussions, commits, and
  official documentation over summaries.

  Produce a Markdown report with:

  1. Executive summary
  2. Material changes since the 2026-08-07 search
  3. Ranked candidate table
  4. Detailed analysis of each promising candidate
  5. Ideas rejected or already implemented, with reasons
  6. Recommended top three implementation experiments
  7. A concrete validation matrix for those experiments
  8. A suggested update to improvemets_list.md, clearly separated from the
     research report

  Rank candidates using:
  - expected benefit on this hardware: 0–5
  - implementation feasibility: 0–5
  - evidence quality: 0–5
  - regression risk: low/medium/high
  - overall priority: now/next/watch/reject

  Do not modify or claim authorship of upstream work. Refer to the maintained
  project as the “Eaman patch”; preserve Qwen/Sol names only for historical
  artifacts.

**Prioritize these sites, but allow full-web search**:
 
  github.com/ggml-org/llama.cpp 
  github.com/ROCm 
  rocm.docs.amd.com 
  github.com/ROCm/rocm-systems 

 
  treat summary\_eaman.md, status, and AGENTS.md 
  as higher authority when files disagree. 
improvemets\_list.md contains stale historical statements