[AMD] Port MI355X DeepSeek V4 fixed-sequence disagg to srt-slurm

benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/mi355x/fixed-seq/disagg-1p1d-dep8-mtp.yaml

@@ -0,0 +1,192 @@
+name: "mi355x-dsv4-disagg-1p1d-dep8-mtp-fixed-seq"
+
+model:
+  path: "hf:deepseek-ai/DeepSeek-V4-Pro"
+  container: "sglang-rocm-dsv4-mi355x"
+  precision: "fp4"
+
+identity:
+  model: {repo: "deepseek-ai/DeepSeek-V4-Pro"}
+  container: {image: "lmsysorg/sglang-rocm:v0.5.15.post1-rocm720-mi35x-20260719"}
+  frameworks:
+    sglang: "0.5.15.post1.dev20260719+ga03ca46a28"
+    sglang-router: "0.3.2"
+    amd-mori: "0.5.15.post1.dev20260719+ga03ca46a28"
+
+slurm: {time_limit: "04:00:00"}
+
+resources:
+  gpu_type: "mi355x"
+  gpus_per_node: 8
+  prefill_nodes: 1
+  decode_nodes: 1
+  prefill_workers: 1
+  decode_workers: 1
+  gpus_per_prefill: 8
+  gpus_per_decode: 8
+
+frontend:
+  type: sglang
+  enable_multiple_frontends: false
+  args:
+    policy: consistent_hashing
+    dp-aware: true
+    cache-threshold: 0.3
+    balance-abs-threshold: 2
+    balance-rel-threshold: 1.1
+    disable-circuit-breaker: true
+    health-failure-threshold: 100
+    health-check-timeout-secs: 600
+    health-check-interval-secs: 30
+
+backend:
+  type: sglang
+  prefill_environment: &common_environment
+    HF_HOME: "/hf_hub_cache"
+    HF_HUB_CACHE: "/hf_hub_cache/hub"
+    HUGGINGFACE_HUB_CACHE: "/hf_hub_cache/hub"
+    PYTHONUNBUFFERED: "1"
+    TORCH_NCCL_BLOCKING_WAIT: "1"
+    NCCL_BLOCKING_WAIT: "1"
+    HSA_NO_SCRATCH_RECLAIM: "0"
+    GPU_MAX_HW_QUEUES: "5"
+    SGLANG_AITER_MLA_PERSIST: "0"
+    SGLANG_ENABLE_UNIFIED_RADIX_TREE: "1"
+    SGLANG_OPT_UNIFIED_CACHE_FREE_OUT_OF_WINDOW_SLOTS: "1"
+    SGLANG_MORI_DISPATCH_DTYPE: "auto"
+    SGLANG_MORI_DISPATCH_INTER_KERNEL_SWITCH_THRESHOLD: "4096"
+    SGLANG_OPT_DEEPGEMM_HC_PRENORM: "false"
+    SGLANG_USE_AITER: "1"
+    SGLANG_USE_ROCM700A: "0"
+    SGLANG_OPT_USE_FUSED_COMPRESS: "true"
+    SGLANG_HACK_FLASHMLA_BACKEND: "unified_kv_triton"
+    SGLANG_OPT_FP8_WO_A_GEMM: "false"
+    SGLANG_OPT_USE_JIT_INDEXER_METADATA: "false"
+    SGLANG_OPT_USE_TOPK_V2: "false"
+    SGLANG_OPT_USE_AITER_INDEXER: "true"
+    SGLANG_OPT_USE_TILELANG_INDEXER: "false"
+    SGLANG_OPT_USE_TILELANG_MHC_PRE: "false"
+    SGLANG_OPT_USE_TILELANG_MHC_POST: "false"
+    SGLANG_FP8_PAGED_MQA_LOGITS_TORCH: "1"
+    SGLANG_OPT_USE_FUSED_COMPRESS_TRITON: "true"
+    SGLANG_OPT_USE_MULTI_STREAM_OVERLAP: "false"
+    SGLANG_ROCM_USE_MULTI_STREAM: "false"
+    AITER_BF16_FP8_MOE_BOUND: "0"
+    SGLANG_EAGER_INPUT_NO_COPY: "true"
+    SGLANG_SHARED_EXPERT_TP1: "1"
+    SGLANG_DP_SHARED_EXPERT_LOCAL: "1"
+    SGLANG_DP_USE_GATHERV: "1"
+    SGLANG_DP_USE_REDUCE_SCATTER: "1"
+    IBDEVICES: "rdma0,rdma1,rdma2,rdma3,rdma4,rdma5,rdma6,rdma7"
+    MORI_RDMA_TC: "104"
+    MORI_IO_SQ_BACKOFF_TIMEOUT_US: "500000"
+    MORI_IO_QP_MAX_SEND_WR: "32767"
+    SGLANG_MORI_COMBINE_DTYPE: "auto"
+    SGLANG_MORI_NUM_MAX_DISPATCH_TOKENS_PER_RANK: "16384"
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: "1800"
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "1800"
+    # The native router may select a different DPA rank from the rank encoded
+    # in bootstrap_room. Use SGLang's supported rank-query fallback so decode
+    # retrieves the actual prefill rank instead of rejecting the KV transfer.
+    SGLANG_DISAGGREGATION_FORCE_QUERY_PREFILL_DP_RANK: "1"
+  decode_environment:
+    <<: *common_environment
+    SGLANG_MORI_NUM_MAX_DISPATCH_TOKENS_PER_RANK: "1024"
+  sglang_config:
+    prefill: &common_server
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      trust-remote-code: true
+      tp-size: 8
+      ep-size: 8
+      dp-size: 8
+      enable-dp-attention: true
+      enable-prefill-delayer: true
+      ep-dispatch-algorithm: fake
+      moe-a2a-backend: mori
+      deepep-mode: normal
+      moe-dense-tp-size: 1
+      enable-dp-lm-head: true
+      watchdog-timeout: 3600
+      load-balance-method: auto
+      kv-cache-dtype: fp8_e4m3
+      attention-backend: dsv4
+      page-size: 256
+      swa-full-tokens-ratio: 0.1
+      disable-shared-experts-fusion: true
+      tool-call-parser: deepseekv4
+      reasoning-parser: deepseek-v4
+      disaggregation-transfer-backend: mori
+      speculative-algorithm: EAGLE
+      speculative-eagle-topk: 1
+      speculative-num-steps: 1
+      speculative-num-draft-tokens: 2
+      log-level: info
+      log-level-http: error
+      # The DSV4 chat encoder adds three framing tokens to the nominal 8192
+      # prompt. Match the workflow's 9472-token admission budget so the full
+      # 1024-token completion fits (8195 + 1024) instead of being rejected.
+      context-length: 9472
+      enable-metrics: true
+      enable-cache-report: true
+      mem-fraction-static: 0.80
+      max-running-requests: 1024
+      chunked-prefill-size: 131072
+      max-total-tokens: 2097152
+      disable-radix-cache: true
+      disable-cuda-graph: true
+    decode:
+      <<: *common_server
+      mem-fraction-static: 0.85
+      max-total-tokens: null
+      disable-radix-cache: false
+      disable-cuda-graph: false
+      prefill-round-robin-balance: true
+      cuda-graph-bs: [1, 2, 4, 8, 16, 32, 64, 128]
+
+sbatch_directives:
+  cpus-per-task: "128"
+  mem: "0"
+
+srun_options:
+  mem: "0"
+  container-writable: ""
+  container-remap-root: ""
+
+health_check: {max_attempts: 720, interval_seconds: 5}
+
+benchmark:
+  type: custom
+  command: |
+    set -euo pipefail
+    result_root="/results/${SLURM_JOB_ID}/fixed-seq"
+    mkdir -p "${result_root}"
+    trap 'tar -C /logs -czf "'"${result_root}"'/runtime-logs.tar.gz" . 2>/dev/null || true' EXIT
+    for concurrency in 256 512; do
+      python3 /infmax-workspace/utils/bench_serving/benchmark_serving.py \
+        --backend openai \
+        --base-url "http://${SRT_FRONTEND_HOST}:${SRT_FRONTEND_PORT}" \
+        --endpoint /v1/completions \
+        --model deepseek-ai/DeepSeek-V4-Pro \
+        --tokenizer deepseek-ai/DeepSeek-V4-Pro \
+        --trust-remote-code \
+        --dataset-name random \
+        --random-input-len 8192 \
+        --random-output-len 1024 \
+        --random-range-ratio 1.0 \
+        --random-num-workers 1 \
+        --use-chat-template \
+        --dsv4 \
+        --num-warmups "$((concurrency * 2))" \
+        --num-prompts "$((concurrency * 10))" \
+        --max-concurrency "${concurrency}" \
+        --request-rate inf \
+        --ignore-eos \
+        --disable-tqdm \
+        --save-result \
+        --result-dir "${result_root}" \
+        --result-filename "dsv4-fp4-disagg-mtp1-dep8-isl8192-osl1024-c${concurrency}.json"
+    done
+  env:
+    HF_HOME: /hf_hub_cache
+    HF_HUB_CACHE: /hf_hub_cache/hub
+    HUGGINGFACE_HUB_CACHE: /hf_hub_cache/hub