[AMD] Add dsv4-fp4-mi355x-atom-disagg-mtp config

configs/amd-master.yaml

@@ -1603,6 +1603,75 @@ dsv4-fp4-mi355x-atom-disagg:
           additional-settings:
           - "DECODE_NODES=1"
       # 1P1D TP8
+
+# MTP variant of dsv4-fp4-mi355x-atom-disagg. Identical topology; enables
+# DeepSeek-V4-Pro built-in MTP via spec-decoding=mtp + DECODE_MTP_SIZE>0 on the
+# decode workers. server_atom.sh then appends models_atom.yaml's mtp_flags
+# ("--method mtp --num-speculative-tokens <N>"). Reuses the same launch script
+# (dsv4_fp4_mi355x_atom-disagg.sh) — the multinode launcher adds no _mtp suffix.
+# Results are disambiguated from the base config by the spec-mtp RESULT_FILENAME
+# segment, so no filename collision. Image bumped to nightly_202607201550
+# (2026-07-20): verified on-node it carries the dsv4 mtp path
+# (deepseek_v4 -> deepseek_v4_mtp -> DeepseekV4MTPModel) and the atomesh router.
+# Draft depth (DECODE_MTP_SIZE -> --num-speculative-tokens) is 3 on both arms,
+# matching the single-node dsv4_fp4_mi355x_atom_mtp.sh recipe. The Atom MTP head
+# runs autoregressively N times, so N=3 is valid even though dsv4 ships 1 nextn
+# layer. At high concurrency (>=512) a shallower depth may be worth revisiting
+# once the curve is measured (cf. the concurrency taper in
+# dsr1-fp4-mi355x-sglang-disagg-mtp), but we start with a uniform 3.
+dsv4-fp4-mi355x-atom-disagg-mtp:
+  image: rocm/atom-dev:nightly_202607201550
+  model: deepseek-ai/DeepSeek-V4-Pro
+  model-prefix: dsv4
+  runner: mi355x
+  precision: fp4
+  framework: atom-disagg
+  router: { name: atomesh, version: "087b82d9c1f630e79149ba37e6213257ec9a76f8" }
+  kv-p2p-transfer: mooncake
+  multinode: true
+  disagg: true
+  scenarios:
+    fixed-seq-len:
+    - isl: 8192
+      osl: 1024
+      search-space:
+      # 2P1D DPA+TP8
+      - spec-decoding: "mtp"
+        conc-list: [ 256, 512, 768, 1024, 2048 ]
+        prefill:
+          num-worker: 2
+          tp: 8
+          ep: 1
+          dp-attn: true
+          additional-settings:
+          - "PREFILL_NODES=2"
+        decode:
+          num-worker: 1
+          tp: 8
+          ep: 1
+          dp-attn: true
+          additional-settings:
+          - "DECODE_NODES=1"
+          - "DECODE_MTP_SIZE=3"
+      # 1P1D TP8
+      - spec-decoding: "mtp"
+        conc-list: [ 4, 8, 16, 32, 64, 128 ]
+        prefill:
+          num-worker: 1
+          tp: 8
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "PREFILL_NODES=1"
+        decode:
+          num-worker: 1
+          tp: 8
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "DECODE_NODES=1"
+          - "DECODE_MTP_SIZE=3"
+
 minimaxm3-fp8-mi355x-vllm:
   image: vllm/vllm-openai-rocm:nightly-09663abde0f50944a8d5ea30120666024b503faa
   model: MiniMaxAI/MiniMax-M3-MXFP8

perf-changelog.yaml

@@ -4982,3 +4982,12 @@
     - "DEP arm runs conc-list [48] only (was [48, 64, 96, 128, 192, 256, 512]): conc 48 is the frontier peak (163 tok/s/gpu out, ttft 2.8s); conc >=64 overflows the HBM+host radix cache (GPU hit 0.93->0.57 at 48->64) and thrashes on re-prefill, so it is strictly dominated"
     - "Low-latency TP8 arm (conc [1, 2, 4, 8, 16, 32], fp8 KV + cutedsl bf16 GEMM) unchanged; DEP mirrors the proven dsv4-fp4-b300-sglang tp8/ep8/dp-attn DEP arm"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2281
+
+- config-keys:
+    - dsv4-fp4-mi355x-atom-disagg-mtp
+  description:
+    - "Add DeepSeek-V4-Pro FP4 MI355X ATOM disaggregated MTP benchmark (spec-decoding=mtp, num-speculative-tokens=3 via DECODE_MTP_SIZE)"
+    - "Config-only: mirrors dsv4-fp4-mi355x-atom-disagg topology (2P1D DPA+TP8 conc 256-2048, 1P1D TP8 conc 4-128) and reuses benchmarks/multi_node/dsv4_fp4_mi355x_atom-disagg.sh; MTP is driven by models_atom.yaml mtp_flags + DECODE_MTP_SIZE in the decode additional-settings"
+    - "Draft depth 3 on both arms (matches single-node dsv4_fp4_mi355x_atom_mtp.sh); high-conc depth may be reduced later per the dsr1-fp4-mi355x-sglang-disagg-mtp concurrency taper once results land"
+    - "Image rocm/atom-dev:nightly_202607201550 (bumped from the base config's stale 2026-06-10 nightly): verified on-node it carries the dsv4 mtp path (deepseek_v4 -> deepseek_v4_mtp -> DeepseekV4MTPModel) and the atomesh disagg router"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2294