fix(agentic-mtp): revert mem_fraction_static override, narrow sweep to c48

Disable the 0.75 mem_fraction_static override that caused a 6x KV cache
reduction (5.1M → 873K tokens), falling back to the default 0.90.
Narrow the sweep to DPA c48 only for a focused retest.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: seungrokj

benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh

@@ -82,7 +82,8 @@ if [ "$DP_ATTENTION" = "true" ]; then
     USE_SGLANG_ROUTER=true
     # DPA + MTP needs additional runtime headroom for speculative decode and
     # communication buffers beyond SGLang's static KV pool.
-    MEM_FRACTION_STATIC=0.75
+    # 0719
+    #MEM_FRACTION_STATIC=0.75
     export AIPERF_HTTP_X_SMG_ROUTING_KEY_FROM_CORRELATION_ID=true
     SGLANG_BACKEND_PORT=$((PORT + 1))
     SGLANG_ROUTER_METRICS_PORT=$((PORT + 10000))

configs/amd-master.yaml

@@ -1713,8 +1713,9 @@ dsv4-fp4-mi355x-sglang-agentic-mtp:
     agentic-coding:
     - dram-utilization: 0.80
       search-space:
-      - { tp: 8, kv-offloading: none, conc-list: [4, 8], spec-decoding: mtp }
-      - { tp: 8, dp-attn: true, kv-offloading: none, conc-list: [16, 32, 48, 52], spec-decoding: mtp }
+      #- { tp: 8, kv-offloading: none, conc-list: [4, 8], spec-decoding: mtp }
+      #- { tp: 8, dp-attn: true, kv-offloading: none, conc-list: [16, 32, 48, 52], spec-decoding: mtp }
+      - { tp: 8, dp-attn: true, kv-offloading: none, conc-list: [48], spec-decoding: mtp }
 
 # MiniMax-M3 MXFP8 MI355X recipe:
 # https://github.com/vllm-project/recipes/commit/2a3728ed9892debfd767a72a58ebc90b33f186e5