@@ -1,5 +1,87 @@
 # SlimServe Optimization Status
 
+## 2026-10-15 - PRs #55/#56 (Qwen GGUF GDN head layout, embedded MTP heads) validated on one MI300X; the ROCm launch_pdl boot break (#89)
+
+- Scope: the `qwen38-q2kxl-1` mi300x record (TP1, the registered DFlash2
+  k=3 drafter through `--spec`, Qwen3.8-27B-UD-Q2_K_XL.gguf) on one card of
+  the 8x MI300X box, native ROCm build of main's csrc (the PRs touch no
+  native code). Heads validated: #55 at 0bfc84736f (main 74b5cb8b61 merged
+  in, plus the CPU test fix below), #56 at 6887994ab2 (#55 merged in). The
+  arms ran on local branches `val-main` = main af8026706 + the #89 fix and
+  `val-pr56` = #56 + the #89 fix.
+- Blocker first, on main rather than the PRs: since #29 every GDN/KDA
+  profile died at boot on ROCm with `KeyError: 'Keyword argument launch_pdl
+  was specified but unrecognised'`. The KDA / causal_conv1d / gate-projection
+  Triton kernels passed `launch_pdl=_PDL` as a launch option on every
+  backend; CUDA accepts the keyword (the A100 validation never saw it), the
+  ROCm backend rejects it even when False. Fix: `PDL_LAUNCH =
+  {"launch_pdl": True} if PDL else {}` in `vllm/triton_utils/pdl.py`,
+  splatted at the seven launch sites, the pattern the minimax index_topk
+  launches already used. PR #89, merged ad61e06c2.
+- Tests: the two PR suites on the box 25 passed / 1 skipped after pinning the
+  torch GEMM in the CPU module check (on a ROCm host the unquantized linear
+  dispatches CPU tensors to `vllm::rocm_unquantized_gemm`, which has no CPU
+  kernel; commit 0bfc84736f); tests/slimserve 307 passed. After merging main
+  0eb1fc4d4d into #55 (only overlap: tests/kernels/mamba/test_gdn_fused_mtp.py,
+  merged clean) the GDN layout and fused-MTP suites pass on a Mac CPU
+  (18 passed, 1 skipped).
+- Arm M (main + #89) and arm P (#56): text, reasoning, image and tool
+  canaries pass on both; model load 20.79 GiB and KV pool 1,202,879 tokens
+  on both (the PR dequantizes nothing); boot about 30 s longer on the PR
+  (load-time untiling of the GGUF tensors).
+- Parity: prompt_logprobs over four 3000-character slices of
+  benchmarks/sonnet.txt (3128 tokens). Main vs PR: mean |dlogprob| 0.041
+  nats, max 0.70, argmax agreement 3104/3128, ppl 3.2199 vs 3.2097. Noise
+  floor from two boots of the same PR code (arm P vs arm E): 0.038 nats,
+  max 0.68, 3092/3128, ppl 3.2097 vs 3.2196. Main and the PR are equivalent
+  within run-to-run noise.
+- First exact pass (seed 42, 1000 in / 2000 out, temperature 1.0 / top-p
+  0.95 / top-k 20): main c1 84.40 / c8 220.15 tok/s, PR c1 73.64 / c8
+  186.17 (-12.7 % / -15.4 %), which read as a regression. Attribution from
+  the harness JSON's spec counters (steps = `spec_decode_drafts`): c1 main
+  642 steps at 36.91 ms, PR 754 steps at 36.02 ms; acceptance 70.5 % vs
+  55.0 %. The engine's interval metrics show the same drafted throughput on
+  both arms (223.2 tok/s at c8) and acceptance climbing to 99-100 % as the
+  ignore_eos generation falls into repetition, at different points on the
+  two trajectories. A torch-profiler capture of a 256-token decode on each
+  arm: kernel total 476.3 vs 468.1 ms with the PR's graph replays slightly
+  shorter; the PR's only new kernel is the per-layer index_select (48 per
+  step, 0.09 ms replayed / 0.70 ms eager at T=1, measured separately).
+  Nothing in the step got slower.
+- Multi-seed exact pass, same boots, warmed, six seeds at c1 and three at c8
+  (`seeds-val-main/`, `seeds-val-pr56/`), every run exact:
+
+  | arm | c1 tok/s per seed (1..6) | c1 mean | c1 ms/step | c1 accept | c8 tok/s (42 / 7 / 99) | c8 mean | c8 ms/step | c8 accept |
+  |---|---|---|---|---|---|---|---|---|
+  | main + #89 | 87.07 / 91.46 / 90.62 / 70.92 / 94.03 / 74.61 | 84.8 | 37.13 | 71.6 % | 199.85 / 209.63 / 217.26 | 208.9 | 14.45 | 67.3 % |
+  | #56 | 89.91 / 88.34 / 84.01 / 84.68 / 88.29 / 83.82 | 86.5 | 36.17 | 70.9 % | 210.50 / 212.34 / 200.34 | 207.7 | 14.60 | 67.8 % |
+
+  (c8 ms/step is wall / total drafts, i.e. per request-step; multiply by 8
+  for the batch step.) Decision: no regression. Per-step cost is equal (PR
+  2.6 % faster at c1, 1 % slower at c8, both inside the seed spread); the
+  single-seed gap was draft acceptance on seed 42's trajectory. The
+  recorded V2 DFlash2 baseline of 2026-08-25 (c1 ~75.5 / c8 194-197.5) sits
+  inside the same spread.
+- Methodology note: with ignore_eos and 2000 output tokens a sampled
+  generation collapses into repetition and DFlash acceptance goes to 100 %,
+  so one seeded exact number on this profile carries about +-12 % of
+  trajectory noise (70.9 to 94.0 tok/s across six seeds on main). A/B
+  comparisons on this record need several seeds, or the per-step time from
+  the spec counters; the registered c1/c8 numbers should be read with that
+  spread.
+- Arm E (#56's feature: the GGUF's embedded NextN head served as an MTP
+  drafter, k=2, through a validation-only registry edit reverted afterwards;
+  the GGUF carries `qwen35.nextn_predict_layers=1` and four
+  `blk.64.nextn.*` tensors): boots with method=mtp, canaries pass, pool
+  1,426,371 tokens, seed-42 c1 77.06 / c8 201.51 (32.60 / 11.96 ms per
+  step, acceptance 75.6 % / 70.5 %). Single seed, so not ranked against the
+  DFlash2 drafter; the registered drafter stays.
+- Raw (on the MI300X box): `perf/results/2026-10-15/pr5556-mi300x/`:
+  armM/, armP/, armE/ (exact JSON, warmups), `*-server.log`, `*-parity.json`,
+  `profile-val-main/`, `profile-val-pr56/` (traces + profiler tables),
+  `seeds-val-main/`, `seeds-val-pr56/`; driver scripts and logs in
+  `~/.local/scratch/pr5556/`.
+
 ## 2026-10-14 - PR #86 (LTX-2.5 on Apple Silicon) render and serve check on a Mac Studio M4 Max
 
 - Box: Mac Studio M4 Max (40-core GPU, 128 GiB, macOS 27.0.1), Xcode 27.0