The none case previously passed --no-enable-prefix-caching, which disables the
on-GPU prefix cache entirely, not just DRAM offload. That crippled the no-offload
baseline (no reuse even within HBM) and made lmcache look artificially good. Leave
prefix caching at vLLM's default (ON) so none is an honest GPU-only-KV baseline,
apples-to-apples vs lmcache. Matches the kimik2.5 / dsv4 agentic recipes.

Set the kimik2.7-fp4-mi355x-vllm-agentic sweep to conc [1,4,8,16,32,48] for
both none and dram/LMCache at TP4, dram-utilization 0.80. Add the corresponding
perf-changelog entry (pr-link placeholder to fill on PR open).

Previous commit accidentally dropped perf-changelog.yaml (blob too large for
inline arg). Restore it with the kimik2.7-fp4-mi355x-vllm-agentic entry appended.

- Frontier-only sweep: none [1,4,8] (low conc, fits GPU HBM) + lmcache [16,32]
  (above the KV cliff). Drops dominated/cliff points per review.
- Fix recipe header: 'Variant of kimik2.5_fp4_mi355x.sh' (was self-referential)
- perf-changelog: correct runner (cluster:mi355x-amds) and conc lists

… smoke

At TP4 the conc16/32 cells were GPU-KV-capacity bound: the ~121 GiB KV
pool held only ~12 resident ~90k-token contexts, so throughput plateaued
(~160 tok/s) and requests preempted/thrashed from conc16 up (0 preemptions
at c8, 24 at c16, 35 at c32). LMCache was healthy (82% external prefix-hit)
so offload quality was not the limiter -- raw resident-KV room was.

Move the agentic-coding sweep to TP8 (doubles the KV pool and per-token
decode compute) at conc [16,32,48] for both KV-offloading none and
dram/LMCache, as a 900s trace-replay smoke.

Adds an optional per-scenario `duration` override to the agentic-coding
config schema (defaults to 3600s when unset) so smoke sweeps can run
shorter without changing the global default.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Add a third KV tier to the TP8 agentic comparison alongside none and
LMCache: dram/Mooncake via the embedded MooncakeStoreConnector store
(built from source for ROCm, tcp protocol), modeled on the MI355X DSv4
vLLM recipe. Wire it as an additive `mooncake)` case so the existing
none/lmcache paths are unchanged.

configs/amd-master.yaml now carries all three backends at conc [16,32,48].
configs/smoke-kimik2.7-mooncake.yaml is a temporary overlay to dispatch
only the Mooncake cells, so the in-flight none/lmcache sweep is not rerun.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…32,48]; fix Mooncake worker import

Reshape the TP8 agentic sweep for the real 3600s comparison run:
- none (GPU-only) swept only at low conc [4,8] (it collapses under load)
- LMCache and Mooncake carry the high-conc [16,32,48] range
- duration 3600s (the prior 900s smoke was warmup-dominated for the
  ~90k-token agentic inputs, so per-request metrics were unreliable)

Fix Mooncake so the vLLM worker subprocesses can import it: the previous
bare `make install` only laid down C++ libs, so the launcher imported
`mooncake` but the workers hit "Please install mooncake ..." at KV-cache
init. Build the wheel via Mooncake's scripts/build_wheel.sh (auditwheel
bundles every .so) and pip install it into site-packages, and build with
-DWITH_STORE=ON so store.so (MooncakeStoreConnector) is present.

Drop the temporary configs/smoke-kimik2.7-mooncake.yaml overlay; all three
backends now run from one canonical config.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…weep

- Set VLLM_ROCM_USE_AITER_MLA=0 explicitly (64 KV heads incompatible with the
  ROCm AITER-MLA kernel; was only working via implicit fallback).
- Add kimik2.7-fp4-mi355x-vllm-agentic-tune config key: TP8+LMCache, EP1 vs EP8
  at conc[16,32], 900s. Probes whether expert parallelism improves perf (current
  record sweep runs EP=1 on a large MoE).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…sweep

Wire DP_ATTENTION into the vLLM recipe, mirroring dsv4_fp4_b300_vllm.sh:
- DP_ATTENTION=true -> TP1 x DP=$TP with --enable-expert-parallel (experts
  EP-sharded across DP ranks), fronted by vllm-router on $PORT (consistent_hash
  + X-session affinity from aiperf's X-Correlation-ID). DP engine binds $PORT+1.
- DP_ATTENTION=false -> unchanged pure-TP path on $PORT.
- Add DP_ATTENTION to check_env_vars; clean up ROUTER_PID on exit.

Retarget the -tune config key from the (non-working) EP1-vs-EP8 probe to a
scaling experiment: pure-TP8 vs DEP, each with LMCache/Mooncake, conc[16,32,48],
900s. Aims for a throughput curve that grows with concurrency instead of the
GPU-KV-capacity collapse the pure-TP path hits at conc>=16.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…_MLA)

With VLLM_ROCM_USE_AITER_MLA=0 the only ROCm MLA backend is TRITON_MLA, which
rejects block_size=1 ("No valid attention backend found ... TRITON_MLA:
block_size not supported"). --block-size=1 only ever existed to force the
now-disabled AITER-MLA path. Drop it and let vLLM pick a TRITON_MLA-valid
default, matching AMD's reference serve command (AITER_MLA=0, no --block-size).
Fixes bringup failure for every cell (pure-TP and DEP alike).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…st isolation)

The DEP Mooncake c16 cell fails fast at bringup; GH matrix fail-fast then
cancels the still-running LMCache/pure-TP cells before they finish. Split the
LMCache scaling experiment (-tune) from Mooncake (-tune-mc) so a Mooncake
bringup failure can't poison the scaling run, and so Mooncake runs alone and
uploads its own server_logs for diagnosis.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…ache[16]

Adopt the best-performing config from the sweep analysis:
- Attention: keep AITER-MLA ON (VLLM_ROCM_USE_AITER_MLA=1) + --block-size=1.
  AITER-MLA is ~3x faster than TRITON_MLA on the agentic trace; deliberately
  chosen for throughput despite AMD's 16/128-head guidance (revisit if accuracy
  validation shows the 64-head path is wrong).
- Record sweep narrowed to the best cell per regime: none [4,8] (GPU-only viable
  at low conc, peaks ~26.7 tok/s/GPU at c8), LMCache [16] (crossover where DRAM
  offload overtakes none). Dropped Mooncake and lmcache c32/48 (no headroom over
  this envelope) and the experimental -tune/-tune-mc keys.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>