[AMD] Qwen3.5-FP8 MI355X SGLang disagg MTP: EAGLE speculative decoding / Qwen3.5-FP8 MI355X SGLang disagg MTP：新增EAGLE投机解码配置

benchmarks/multi_node/amd_utils/models.yaml

@@ -202,7 +202,10 @@ Qwen3.5-397B-A17B-MXFP4:
 
 Qwen3.5-397B-A17B-FP8:
   base_flags: "--decode-log-interval 1000 --log-level warning --watchdog-timeout 3600 --load-balance-method round_robin --kv-cache-dtype fp8_e4m3 --attention-backend aiter --disaggregation-transfer-backend mori --moe-dense-tp-size 1"
-  mtp_flags: ""
+  # Qwen3.5 built-in MTP head uses EAGLE (in-checkpoint draft, no draft-model-path);
+  # build_server_config appends --speculative-num-steps=DECODE_MTP_SIZE and
+  # --speculative-num-draft-tokens=DECODE_MTP_SIZE+1.
+  mtp_flags: "--speculative-algorithm EAGLE --speculative-eagle-topk 1"
   dp_flags: "--enable-dp-attention --enable-dp-lm-head"
   ep_flags: "--moe-a2a-backend mori"
   prefill:

benchmarks/multi_node/amd_utils/submit.sh

@@ -186,7 +186,7 @@ fi
 # Optional: exclude specific nodes (e.g. nodes with broken Docker sockets).
 # Set SLURM_EXCLUDE_NODES env var to a comma-separated list of hostnames.
 EXCLUDE_OPT=()
-SLURM_EXCLUDE_NODES="${SLURM_EXCLUDE_NODES:-mia1-p01-g09,mia1-p01-g10,mia1-p01-g11,mia1-p01-g12}"
+SLURM_EXCLUDE_NODES="${SLURM_EXCLUDE_NODES:-mia1-p01-g09,mia1-p01-g10,mia1-p01-g11,mia1-p01-g12,mia1-p01-g14,mia1-p01-g16,mia1-p01-g17,mia1-p01-g19,mia1-p01-g31,mia1-p01-g37}"
 if [[ -n "${SLURM_EXCLUDE_NODES:-}" ]]; then
     EXCLUDE_OPT=(--exclude "$SLURM_EXCLUDE_NODES")
 fi

configs/amd-master.yaml

@@ -312,6 +312,63 @@ qwen3.5-fp8-mi355x-sglang-disagg:
           - "DECODE_NODES=1"
           - "DECODE_MTP_SIZE=0"
 
+qwen3.5-fp8-mi355x-sglang-disagg-mtp:
+  image: lmsysorg/sglang:v0.5.16-rocm720-mi35x
+  model: Qwen/Qwen3.5-397B-A17B-FP8
+  model-prefix: qwen3.5
+  runner: mi355x-disagg
+  precision: fp8
+  framework: sglang-disagg
+  router: { name: sglang-router, version: "0.3.2" }
+  kv-p2p-transfer: mori
+  multinode: true
+  disagg: true
+  scenarios:
+    fixed-seq-len:
+    - isl: 8192
+      osl: 1024
+      search-space:
+      # 1P+1D TP4P+TP8D/EP1, EAGLE MTP depth 3 — conc 8-64.
+      - spec-decoding: "mtp"
+        conc-list: [ 8, 16, 32, 64 ]
+        prefill:
+          num-worker: 1
+          tp: 4
+          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"
+      # 1P+1D TP4P+TP4D/EP1, EAGLE MTP depth 3 — conc 128.
+      # At conc=128, TP8 decode bottlenecks MTP (draft verification
+      # saturates 8 GPUs, causing -38% tput/GPU and TTFT explosion).
+      # Switching to TP4D resolves the collapse: +83% tput/GPU, 7.5x
+      # better TTFT vs TP8D, while interactivity converges (39.5 vs 49.5).
+      - spec-decoding: "mtp"
+        conc-list: [ 128 ]
+        prefill:
+          num-worker: 1
+          tp: 4
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "PREFILL_NODES=1"
+        decode:
+          num-worker: 1
+          tp: 4
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "DECODE_NODES=1"
+          - "DECODE_MTP_SIZE=3"
+
 qwen3.5-fp4-mi355x-sglang:
   image: lmsysorg/sglang-rocm:v0.5.16-rocm720-mi35x-20260730
   model: amd/Qwen3.5-397B-A17B-MXFP4

perf-changelog.yaml

@@ -6100,7 +6100,6 @@
     - "Add GB200 DeepSeek-V4-Pro FP4 Dynamo-vLLM AgentX mirroring the GB300 PR #2571 MTP tuning, with every GB300 4-GPU worker sized to 8 GPUs on GB200."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2636
 
-
 - config-keys:
     - qwen3.5-fp8-h200-sglang-agentic-hicache-mtp
   scenario-type:
@@ -6195,3 +6194,11 @@
     - "Use AITER INT4 quick-reduce, ptpc_fp8 online quantization excluding embeddings, lm_head, gates, and experts, an FP8 KV cache, and three-token MTP with golden synthetic acceptance length 2.99 for benchmark runs."
     - "Set max-num-seqs to twice the concurrency, select CUDA graph capture sizes by concurrency, and cap max-num-batched-tokens at 16384."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2576
+
+- config-keys:
+    - qwen3.5-fp8-mi355x-sglang-disagg-mtp
+  description:
+    - "Add new qwen3.5-fp8-mi355x-sglang-disagg-mtp config with native EAGLE MTP (depth 3, top-k 1) on image v0.5.16."
+    - "Two search-space arms: conc 8-64 with TP4P+TP8D, and conc 128 with TP4P+TP4D (avoids TP8D draft-verification bottleneck at high concurrency: +83% tput/GPU, 7.5x TTFT improvement)."
+    - "models.yaml: add mtp_flags (--speculative-algorithm EAGLE --speculative-eagle-topk 1) to Qwen3.5-397B-A17B-FP8."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2607