[AMD][AgentX] Port MI355X DeepSeek V4 MTP disagg to srt-slurm

.github/AGENT_OPERATIONS.md

@@ -92,7 +92,7 @@ For multi-node configurations, `--all-evals` creates one eval job per engine top
 
 Single-node fixed-sequence results may include `power_valid`, `avg_power_w`, `avg_total_gpu_power_w`, `total_gpu_energy_j`, and joules per query/input/output/total token. Invalid telemetry records `power_valid: 0` without energy metrics and fails only with `REQUIRE_POWER=1`.
 
-Multinode disaggregated results add `prefill_gpu_energy_j`, `decode_gpu_energy_j`, `prefill_joules_per_input_token`, and `decode_joules_per_output_token`. Role energy covers the full formal benchmark window, not kernel-level phases.
+Multinode disaggregated results add `prefill_gpu_energy_j`, `decode_gpu_energy_j`, `prefill_avg_power_w`, `decode_avg_power_w`, `prefill_joules_per_input_token`, and `decode_joules_per_output_token`. Role energy covers the full formal benchmark window, not kernel-level phases, and the role watts are that energy divided by the same window and by the role's GPU count.
 
 For srt-slurm recipes, `telemetry: {provider: dcgm-power}` enables official energy collection. `runners/launch_gb200-nv.sh` and `runners/launch_gb300-nv.sh` are the source of truth for `POWER_SRT_SLURM_PIN`. CI derives `POWER_PRODUCER_SHA` from the launcher stamp. `utils/test_gb200_power_official_contract.py` and `utils/test_gb300_power_official_contract.py` enforce the recipe/launcher contract. Only `PRECISION=fp8` dcgm-power lanes are validated.
 

.github/workflows/e2e-tests.yml

@@ -261,6 +261,7 @@ jobs:
                   MULTI_AGENTIC_EVAL=$(echo "$CONFIG_JSON" | python3 -c "import sys,json; d=json.load(sys.stdin); print(json.dumps([x for x in d if x.get('scenario-type') == 'agentic-coding' and 'prefill' in x and x.get('run-eval', False)]))" | score_matrix multi-agentic-eval)
                   SINGLE=$(echo "$CONFIG_JSON" | python3 -c "import sys,json; d=json.load(sys.stdin); print(json.dumps([x for x in d if 'prefill' not in x and x.get('scenario-type') != 'agentic-coding' and not x.get('eval-only', False)]))" | score_matrix single)
                   MULTI=$(echo "$CONFIG_JSON" | python3 -c "import sys,json; d=json.load(sys.stdin); print(json.dumps([x for x in d if 'prefill' in x and x.get('scenario-type') != 'agentic-coding' and not x.get('eval-only', False)]))" | score_matrix multi)
+                  EVALS=$(echo "$CONFIG_JSON" | python3 -c "import sys,json; d=json.load(sys.stdin); print(json.dumps([x for x in d if 'prefill' not in x and x.get('scenario-type') != 'agentic-coding' and x.get('run-eval', False)]))" | score_matrix eval)
                   MULTI_EVAL=$(echo "$CONFIG_JSON" | python3 -c "import sys,json; d=json.load(sys.stdin); print(json.dumps([x for x in d if 'prefill' in x and x.get('scenario-type') != 'agentic-coding' and x.get('run-eval', False)]))" | score_matrix multi-eval)
                   {
                     echo "agentic-config=$AGENTIC"

benchmarks/multi_node/amd_utils/job.slurm

@@ -458,7 +458,6 @@ DOCKER_ENV_COMMON=(
     -e PREFILL_DP_ATTN=\${PREFILL_DP_ATTN:-false}
     -e PREFILL_CONTEXT_LENGTH=\${PREFILL_CONTEXT_LENGTH:-}
     -e PREFILL_CHUNKED_PREFILL_SIZE=\${PREFILL_CHUNKED_PREFILL_SIZE:-}
-    -e DISABLE_CUSTOM_ALL_REDUCE=\${DISABLE_CUSTOM_ALL_REDUCE:-}
     -e MAX_MODEL_LEN=\${MAX_MODEL_LEN:-}
     -e DURATION=\${DURATION:-1800}
     -e IS_AGENTIC=\${IS_AGENTIC:-0}
@@ -467,8 +466,6 @@ DOCKER_ENV_COMMON=(
     -e KV_OFFLOAD_BACKEND_METADATA=\"\${KV_OFFLOAD_BACKEND_METADATA:-}\"
     -e TOTAL_CPU_DRAM_GB=\${TOTAL_CPU_DRAM_GB:-}
     -e ENABLE_METRICS=\${ENABLE_METRICS:-0}
-    -e PREFILL_ROUTER_POLICY=\${PREFILL_ROUTER_POLICY:-random}
-    -e DECODE_ROUTER_POLICY=\${DECODE_ROUTER_POLICY:-random}
     -e MORI_IO_SQ_BACKOFF_TIMEOUT_US=\${MORI_IO_SQ_BACKOFF_TIMEOUT_US:-}
     -e MORI_IO_QP_MAX_SEND_WR=\${MORI_IO_QP_MAX_SEND_WR:-}
     -e DECODE_TP_SIZE=\$DECODE_TP_SIZE
@@ -538,7 +535,6 @@ fi
 HICACHE_MC_CONFIG="${BENCHMARK_LOGS_DIR}/hicache_mc_${SLURM_JOB_ID}.env"
 cat > "$HICACHE_MC_CONFIG" <<EOF
 HICACHE_RATIO=${HICACHE_RATIO:-}
-FORCE_HICACHE_RATIO=${FORCE_HICACHE_RATIO:-}
 HICACHE_HOST_POOL_COUNT=${HICACHE_HOST_POOL_COUNT:-}
 HICACHE_PAGE_SIZE=${HICACHE_PAGE_SIZE:-}
 HICACHE_IO_BACKEND=${HICACHE_IO_BACKEND:-}

benchmarks/multi_node/amd_utils/models.yaml

@@ -363,41 +363,6 @@ DeepSeek-R1-0528-MXFP4-v2:
       chunked_prefill_size: 262144
       cuda_graph_bs_range: "1-128"
 
-DeepSeek-V4-Pro-AgentX:
-  base_flags: "--watchdog-timeout 3600 --load-balance-method round_robin --kv-cache-dtype fp8_e4m3 --attention-backend dsv4 --page-size 256 --swa-full-tokens-ratio 0.1 --disable-shared-experts-fusion --tool-call-parser deepseekv4 --reasoning-parser deepseek-v4 --disaggregation-transfer-backend mori --log-level info --log-level-http error"
-  dp_flags: "--enable-dp-attention --enable-prefill-delayer"
-  ep_flags: "--ep-dispatch-algorithm fake --moe-a2a-backend mori --deepep-mode normal"
-  mtp_flags: "--speculative-algorithm EAGLE --speculative-eagle-topk 1"
-  prefill:
-    mem_fraction_static: 0.72
-    disable_radix_cache: false
-    disable_cuda_graph: true
-    dp:
-      max_running_requests: 1024
-      chunked_prefill_size: "MORI_MAX_DISPATCH_TOKENS_PREFILL * PREFILL_TP_SIZE"  # dsv4 compressor kernel uint16 token cap (255*256)
-      context_length: 1048576
-      # max_total_tokens: 1048576
-    no_dp:
-      max_running_requests: 64
-      # Small prefill chunks interleave long-context agentic prefills across
-      # requests instead of letting one ~100K-token prefill monopolize the
-      # engine (the conc>=16 queue-saturation / decode-stall failure mode).
-      # Mirrors the single-node DSv4 agentic recipe (dsv4_fp4_mi355x.sh=8192).
-      # Was 65280 (255*256, the dsv4 compressor kernel uint16 token cap); 8192
-      # (32*256) stays a page-size multiple well under that cap.
-      chunked_prefill_size: 8192
-      context_length: 1048576
-  decode:
-    mem_fraction_static: 0.85
-    prefill_round_robin_balance: true
-    disagg_decode_enable_radix_cache: false
-    dp:
-      max_running_requests: 1024
-      cuda_graph_bs_range: "1-128"
-    no_dp:
-      max_running_requests: 128
-      cuda_graph_bs_range: "1-128"
-
 DeepSeek-V4-Pro-DI:
   base_flags: "--decode-log-interval 100 --log-level info --watchdog-timeout 3600 --load-balance-method round_robin --kv-cache-dtype fp8_e4m3 --attention-backend dsv4 --page-size 256 --swa-full-tokens-ratio 0.1 --disable-shared-experts-fusion --tool-call-parser deepseekv4 --reasoning-parser deepseek-v4 --disaggregation-transfer-backend mori"
   # DeepSeek-V4 built-in MTP: load via the EAGLE spec path with eagle-topk 1 (a single

benchmarks/multi_node/amd_utils/server_sglang.sh

@@ -537,22 +537,14 @@ if [[ "$KV_OFFLOADING" != "none" && "$KV_OFFLOAD_BACKEND" == "hicache" ]]; then
     # per-node DRAM budget computed by the sweep generator (enforcement); fall
     # back to --hicache-ratio (relative to the GPU KV pool) when no budget is
     # provided, keeping configs that predate the budget unchanged.
-    # FORCE_HICACHE_RATIO lets a recipe opt into ratio-based sizing without
-    # unsetting TOTAL_CPU_DRAM_GB — that var is also the shared client-side
-    # gate (benchmark_lib.sh requires it whenever KV_OFFLOADING=dram) and is
-    # forwarded verbatim into client.env below, so unsetting it here would
-    # make the aiperf client container fail its own env validation before
-    # ever sending a request.
     HICACHE_RATIO="${HICACHE_RATIO:-5}"
     HICACHE_SIZING_FLAGS="--hicache-ratio ${HICACHE_RATIO}"
     # DeepSeek V4's hybrid HiCache pool rejects --hicache-size (requires
     # --hicache-ratio), so the absolute per-node budget cannot be applied to it.
     # See sglang _deepseek_v4_num_host_pages() (raises ValueError when
     # server_args.hicache_size > 0):
     # https://github.com/sgl-project/sglang/blob/9dd57ef8c48e2cd82292d849f01e2130c5203e67/python/sglang/srt/mem_cache/hybrid_cache/hybrid_pool_assembler.py#L262-L266
-    # FORCE_HICACHE_RATIO additionally lets a recipe opt into ratio-based sizing
-    # for any other model without unsetting TOTAL_CPU_DRAM_GB (see comment above).
-    if [[ "${FORCE_HICACHE_RATIO:-0}" != "1" && -n "${TOTAL_CPU_DRAM_GB:-}" && "${TOTAL_CPU_DRAM_GB}" -gt 0 && "${MODEL_NAME}" != *DeepSeek-V4* ]]; then
+    if [[ -n "${TOTAL_CPU_DRAM_GB:-}" && "${TOTAL_CPU_DRAM_GB}" -gt 0 && "${MODEL_NAME}" != *DeepSeek-V4* ]]; then
         # TOTAL_CPU_DRAM_GB is the prefill worker's per-node budget (only prefill
         # offloads KV to CPU DRAM today); --hicache-size is per rank per host
         # pool. A prefill server may span nodes (PREFILL_TP_SIZE is its total
@@ -791,13 +783,7 @@ if [ "$NODE_RANK" -eq 0 ]; then
         # across the agentic trace; round_robin decode keeps the single decode worker
         # fed evenly. Override via ROUTER_RESILIENCE_FLAGS / ROUTER_POLICY_FLAGS.
         ROUTER_RESILIENCE_FLAGS="${ROUTER_RESILIENCE_FLAGS:---disable-circuit-breaker --health-failure-threshold 100 --health-check-timeout-secs 600 --health-check-interval-secs 30}"
-        # server_sglang.sh previously read ROUTER_PREFILL_POLICY, but the recipe
-        # scripts export PREFILL_ROUTER_POLICY, so the recipe's policy override was
-        # silently ignored and the router always fell back to this hardcoded
-        # default. Also comment out ROUTER_DECODE_POLICY for now (superseded by
-        # --dp-aware below).
-        ROUTER_PREFILL_POLICY="${PREFILL_ROUTER_POLICY:-consistent_hashing}"
-        # ROUTER_DECODE_POLICY="${ROUTER_DECODE_POLICY:-round_robin}"
+        ROUTER_PREFILL_POLICY="consistent_hashing"
         ROUTER_CACHE_THRESHOLD="${ROUTER_CACHE_THRESHOLD:-0.3}"
         ROUTER_BALANCE_ABS_THRESHOLD="${ROUTER_BALANCE_ABS_THRESHOLD:-2}"
         ROUTER_BALANCE_REL_THRESHOLD="${ROUTER_BALANCE_REL_THRESHOLD:-1.1}"
@@ -1275,35 +1261,7 @@ else
     fi
     set +x
 
-    # Agentic trace replay doesn't reproduce real token-by-token traffic, so
-    # measured MTP/EAGLE acceptance there isn't representative (PR #2309
-    # review: https://github.com/SemiAnalysisAI/InferenceX/pull/2309#pullrequestreview-4778348624).
-    # Per the AgentX fairness guidelines (golden_al_distribution/README.md),
-    # agentic throughput benchmarks simulate acceptance at the model's
-    # committed golden AL instead of measuring real (non-representative)
-    # acceptance. Eval runs (RUN_EVAL / EVAL_ONLY) need real acceptance so
-    # GSM8K scores reflect actual MTP behavior. Golden curve source:
-    # golden_al_distribution/dsv4_mtp.yaml (thinking_on).
-    DECODE_SIM_ACC_ENV=""
-    if [[ "$DECODE_MTP_SIZE" -gt 0 ]] && { [[ "${IS_AGENTIC:-0}" == "1" ]] || [[ "${IS_AGENTIC:-}" == "true" ]]; }; then
-        if [[ "${EVAL_ONLY:-false}" == "true" ]] || [[ "${RUN_EVAL:-false}" == "true" ]]; then
-            echo "[INFO] Eval mode: synthetic MTP disabled (using real acceptance)"
-        else
-            DSV4_GOLDEN_AL=""
-            case "${MODEL_NAME}:${DECODE_MTP_SIZE}" in
-                *DeepSeek-V4*:1) DSV4_GOLDEN_AL=1.79 ;;
-                *DeepSeek-V4*:2) DSV4_GOLDEN_AL=2.27 ;;
-                *DeepSeek-V4*:3) DSV4_GOLDEN_AL=2.49 ;;
-            esac
-            if [[ -n "$DSV4_GOLDEN_AL" ]]; then
-                DECODE_SIM_ACC_ENV="SGLANG_SIMULATE_ACC_LEN=${DSV4_GOLDEN_AL} SGLANG_SIMULATE_ACC_METHOD=match-expected SGLANG_SIMULATE_ACC_TOKEN_MODE=real-draft-token"
-            else
-                echo "WARNING: agentic MTP run (model=${MODEL_NAME}, DECODE_MTP_SIZE=${DECODE_MTP_SIZE}) has no golden AL wired in server_sglang.sh -- falling back to real (unsimulated, non-representative) acceptance. Add a case in server_sglang.sh and golden_al_distribution/ before shipping this arm. See golden_al_distribution/README.md." >&2
-            fi
-        fi
-    fi
-
-    DECODE_CMD="SGLANG_MORI_COMBINE_DTYPE=${MORI_COMBINE_DTYPE_DECODE} ${DECODE_MORI_MOE_ENV} SGLANG_MORI_NUM_MAX_DISPATCH_TOKENS_PER_RANK=${MORI_NUM_MAX_DISPATCH_TOKENS_PER_RANK_DECODE:-${MORI_MAX_DISPATCH_TOKENS_DECODE}} MORI_IO_SQ_BACKOFF_TIMEOUT_US=${MORI_IO_SQ_BACKOFF_TIMEOUT_US} MORI_IO_QP_MAX_SEND_WR=${MORI_IO_QP_MAX_SEND_WR} ${DECODE_SIM_ACC_ENV} ${LAUNCH_PREFIX:-} python3 -m sglang.launch_server \
+    DECODE_CMD="SGLANG_MORI_COMBINE_DTYPE=${MORI_COMBINE_DTYPE_DECODE} ${DECODE_MORI_MOE_ENV} SGLANG_MORI_NUM_MAX_DISPATCH_TOKENS_PER_RANK=${MORI_NUM_MAX_DISPATCH_TOKENS_PER_RANK_DECODE:-${MORI_MAX_DISPATCH_TOKENS_DECODE}} MORI_IO_SQ_BACKOFF_TIMEOUT_US=${MORI_IO_SQ_BACKOFF_TIMEOUT_US} MORI_IO_QP_MAX_SEND_WR=${MORI_IO_QP_MAX_SEND_WR} ${LAUNCH_PREFIX:-} python3 -m sglang.launch_server \
         --model-path ${MODEL_DIR}/${MODEL_NAME} \
         --disaggregation-mode decode \
         --disaggregation-ib-device ${IBDEVICES} \