[AMD][AgentX] kimik2.7: add EAGLE3 MTP + SWE-bench accuracy path

- New recipe kimik2.7_fp4_mi355x_mtp.sh: EAGLE3 spec decoding with the
  K2.7-Code-matched draft lightseekorg/kimi-k2.7-coder-eagle3.1-mla
  (pre-staged; num_speculative_tokens 3, env-overridable). spec-decoding: mtp
  routes here via the runner's _mtp suffix.
- Accuracy path (EVAL_ONLY -> run_eval SWE-bench agentic) + Kimi eval parser
  flags (--tool-call-parser/--reasoning-parser kimi_k2 --enable-auto-tool-choice)
  added to BOTH base and MTP recipes.
- Align server flags with recipes.vllm.ai Kimi-K2.7-Code MI355X recipe:
  --moe-backend flydsl + --compilation-config {fuse_allreduce_rms:false}.
- Harden LMCache MP server (flags verified at pinned aaf7c0d3): split worker
  pools --max-gpu-workers 1/--max-cpu-workers 8 (was --max-workers TP*2),
  --chunk-size 1024, --l1-align-bytes 16384, --eviction-trigger-watermark 0.85,
  --eviction-ratio 0.10, --supported-transfer-mode lmcache_driven.
- Config: kimik2.7-fp4-mi355x-vllm-agentic-mtp (record) +
  kimik2.7-fp4-mi355x-vllm-agentic-mtp-smoke (TP4/TP8 x none c4 / lmcache c16, 900s).

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

Author: hyukjlee

benchmarks/single_node/agentic/kimik2.7_fp4_mi355x.sh

@@ -180,8 +180,20 @@ case "$OFFLOAD_MODE" in
         fi
         LMCACHE_L1_INIT_SIZE_GB="${LMCACHE_L1_INIT_SIZE_GB:-20}"
         LMCACHE_L1_READ_TTL_SECONDS="${LMCACHE_L1_READ_TTL_SECONDS:-7200}"
-        LMCACHE_CHUNK_SIZE="${LMCACHE_CHUNK_SIZE:-256}"
-        LMCACHE_MAX_WORKERS="${LMCACHE_MAX_WORKERS:-$((TP * 2))}"
+        # Server flags below track the MI355X Kimi-K2.7 LMCache runbook.
+        # chunk-size 1024: fewer server round-trips than the 256 default; the MP
+        # connector does not pin a matching value, so this is self-contained.
+        LMCACHE_CHUNK_SIZE="${LMCACHE_CHUNK_SIZE:-1024}"
+        LMCACHE_L1_ALIGN_BYTES="${LMCACHE_L1_ALIGN_BYTES:-16384}"
+        # Split worker pools: cap the GPU-affinity pool (STORE/RETRIEVE) at 1 to
+        # avoid concurrent-GPU-transfer stalls under heavy async-load pressure
+        # (the conc64 LMCache stall class); keep 8 CPU-side workers for LOOKUP.
+        # Supersedes the old --max-workers (which set both pools equal).
+        LMCACHE_MAX_GPU_WORKERS="${LMCACHE_MAX_GPU_WORKERS:-1}"
+        LMCACHE_MAX_CPU_WORKERS="${LMCACHE_MAX_CPU_WORKERS:-8}"
+        LMCACHE_EVICTION_TRIGGER_WATERMARK="${LMCACHE_EVICTION_TRIGGER_WATERMARK:-0.85}"
+        LMCACHE_EVICTION_RATIO="${LMCACHE_EVICTION_RATIO:-0.10}"
+        LMCACHE_TRANSFER_MODE="${LMCACHE_TRANSFER_MODE:-lmcache_driven}"
         export PYTHONHASHSEED="${PYTHONHASHSEED:-0}"
         export LMCACHE_BLOCKING_TIMEOUT_SECS=60
 
@@ -196,8 +208,13 @@ case "$OFFLOAD_MODE" in
             --l1-init-size-gb "$LMCACHE_L1_INIT_SIZE_GB"
             --l1-read-ttl-seconds "$LMCACHE_L1_READ_TTL_SECONDS"
             --chunk-size "$LMCACHE_CHUNK_SIZE"
-            --max-workers "$LMCACHE_MAX_WORKERS"
+            --l1-align-bytes "$LMCACHE_L1_ALIGN_BYTES"
+            --max-gpu-workers "$LMCACHE_MAX_GPU_WORKERS"
+            --max-cpu-workers "$LMCACHE_MAX_CPU_WORKERS"
+            --eviction-trigger-watermark "$LMCACHE_EVICTION_TRIGGER_WATERMARK"
+            --eviction-ratio "$LMCACHE_EVICTION_RATIO"
             --eviction-policy LRU
+            --supported-transfer-mode "$LMCACHE_TRANSFER_MODE"
         )
         printf '%q ' "${LMCACHE_CMD[@]}" > "$RESULT_DIR/lmcache_command.txt"
         printf '\n' >> "$RESULT_DIR/lmcache_command.txt"
@@ -348,6 +365,18 @@ VLLM_CMD=(
     --trust-remote-code
     --max-num-seqs "$CONC"
     --mm-encoder-tp-mode data
+    # MI355X (gfx950) recipe defaults from recipes.vllm.ai/moonshotai/Kimi-K2.7-Code:
+    # FlyDSL W4A16 MoE kernel + disable the fuse_allreduce_rms compile pass
+    # (required on MI350X/MI355X).
+    --moe-backend flydsl
+    --compilation-config '{"pass_config": {"fuse_allreduce_rms": false}}'
+    # Structured tool-calls + reasoning for the SWE-bench agentic eval
+    # (EVAL_ONLY path). Kimi-K2 emits the kimi_k2 tool-call/reasoning format;
+    # without these the agent harness can't parse message.tool_calls. Harmless
+    # for the throughput replay path.
+    --tool-call-parser kimi_k2
+    --enable-auto-tool-choice
+    --reasoning-parser kimi_k2
     "${PREFIX_CACHE_ARGS[@]}"
     "${OFFLOAD_ARGS[@]}"
 )
@@ -380,7 +409,13 @@ if [ "$USE_VLLM_ROUTER" = "true" ]; then
     wait_for_server_ready --port "$PORT" --server-log "$ROUTER_LOG" --server-pid "$ROUTER_PID"
 fi
 
-# ---- Run benchmark ----------------------------------------------------------
-build_replay_cmd "$RESULT_DIR"
-
-run_agentic_replay_and_write_outputs "$RESULT_DIR"
+# ---- Run benchmark or accuracy eval -----------------------------------------
+# EVAL_ONLY=true (workflow eval-only input) runs the SWE-bench agentic accuracy
+# eval against the live server instead of the throughput replay; run_eval
+# dispatches to the swebench harness for agentic-coding scenarios.
+if [ "${EVAL_ONLY:-false}" = "true" ]; then
+    run_eval --port "$PORT"
+else
+    build_replay_cmd "$RESULT_DIR"
+    run_agentic_replay_and_write_outputs "$RESULT_DIR"
+fi