Current ROCm MTP status

Production patches (base llama.cpp `e86f5fd1e`, 2026-10-27 refresh):

  latest_rocm_improvement_e86f5fd1e.patch
  cumulative `eaman` through `7318b9325` (merge of e86f5fd1e plus the
  ROCm MTP graph/pool fit-headroom follow-up; upstream
  common_fit_apply/common_fit_log_scope adapted while keeping the MTP
  pipeline-disabled fit and checkpoint reservation, adaptive draft +
  is_lilicorr, upstream draft-state slot save), tested build 2618
  (incremental rebuild after the 2617 full HIP/Vulkan recompile;
  llama-server relinked), tree-validated to `2d18012d`,
  SHA-256 f077b2fa677f5340c46f4fe569c0b64eda90ab02224f85fb7f049dac6d17cc26

  mtp_compact_rollback_c811cb8f.patch
  standalone MTP CR `ec77263a` on c811cb8f mainline, tested build 2159
  (CPU/default build, server + RS-rollback + arg-parser targets;
  arg-parser passed, model-dependent rollback tests skipped - no local
  GGUF), tree-validated to `c0cd7f9d`,
  SHA-256 3470c08dc27e4e190548e4a161b9162192752115faf2a204539f6d76baefb1a6

  kv_prefix_checkpoint_sidecar_f0c41e01.patch
  standalone sidecar `9d27fed0` on pure f0c41e01 mainline (sidecar +
  storage-mode fields + prompt-cache commits), tested build 2090
  (CPU/default build), tree-validated to `fceee33f`,
  SHA-256 07fc4ddafb06b960d386520e35b29d46fad6ad841b8b80f759e690b9ebe7cd17

  Superseded exports moved to patches/:
  latest_rocm_improvement_ff5a23376.patch,
  latest_rocm_improvement_37d6d653.patch,
  latest_rocm_improvement_781dbc5a.patch,
  latest_rocm_improvement_c811cb8f.patch,
  latest_rocm_improvement_f0c41e01.patch,
  mtp_compact_rollback_f0c41e01.patch,
  mtp_compact_rollback_c6824a9e.patch, kv_prefix_checkpoint_sidecar_9adc7f42.patch

The cumulative patch carries the ROCm fitting/pipeline work, MTP Compact
Rollback, ordered MTP multi-ubatch execution, HIP VEC control, the ported
rdna-boosts blocks 09/10, and the prompt-cache best-match + restore-park
server changes (the KV prefix checkpoint sidecar was dropped in b8531ac0;
disk persistence is the native #26004 appendix, no patch needed). Validated
F16/off configuration:

  --pipeline-parallel off
  --cache-type-k-draft f16
  --cache-type-v-draft f16

Key controls: `--spec-mtp-cr-depth` (MTP CR, standalone + cumulative),
`--hip-fa-force-vec` (default `off`, cumulative only, HIP-specific),
`--spec-draft-adaptive` (opt-in; keep static `--spec-draft-p-min 0.75`).

Current mainline `e86f5fd1e` (216 commits past `ff5a23376`) includes the
upstream fit refactor (`common_fit_apply`, `common_fit_log_scope`), the
`test-llama-archs` `kv_writer` refactor, and the speculative
rejection-sampling test. Carried features remain: server default port 9931,
`--moe-cache-mib N`, and the native slot checkpoint appendix (#26004) as the
sole checkpoint persistence path (sidecar dropped).

Fitter validation 2026-10-27: exact `one_shot.shlong_single` launcher
arguments (ThinkingCap-Qwen3.8-27B IQ4_XS MIX, `--fit-target 30`, rocm0,
VEC off) fit 116,736 ctx on build 2618 and completed a 610-token prompt
with no OOM. The ROCm-only 64 MiB MTP graph/pool fit headroom covers the
transient MMQ pool allocation that the fitter does not model; without it
the 118,784-ctx fit OOMs on the first prompt. The earlier startup OOM
cause was upstream moving the fit block into `common_fit_apply()` and
dropping the MTP pipeline-disabled setting; the merge restores it. The
64 MiB headroom is a temporary upstream-fit workaround (~2,048 ctx);
re-test removal on each refresh with a >=459-token prompt.

Source branches (`/home/eaman/llama/llama.cpp`):

  eaman                              7318b9325 (tip; merge of e86f5fd1e + MTP fit headroom)
  eaman-pre-latest-e86f5fd1e         98a2b3cd8 (recovery)
  eaman-kv-prefix-checkpoint-sidecar 9d27fed0 (standalone: pure mainline f0c41e01 + sidecar + prompt-cache)
  eaman-mtp-compact-rollback         ec77263a (merge on c811cb8f)
  eaman-mtp-compact-rollback-pre-c811cb8f     8a589204 (recovery)
  eaman-kv-prefix-checkpoint-sidecar-pre-f0c41e01 c9a5e217 (recovery)
  eaman-rdna3-port                   branch only (worktree dir removed)
  latest                             e86f5fd1e

Deployment: `/home/eaman/llama/bin_vulkan` now holds build 2618
(code 7318b9325), matching the canonical `build/bin`. The deployed build
passed the 610-token `--fit-target 30` prompt. `/home/eaman/llama/bin_vulkan_previous`
is a stale copy that segfaults; do not reuse it.

Detailed history: summary_eaman.md. Feature inventory:
improvements_list.md. Dual-GPU placement procedure:
dual_gpu_context_balancing_guide.md. Logs: logs/. Archived material: old/.

Author and maintainer: eaman.