[DNM][AMD] agentx-v0.4 rebased from commit chore/agentx-v0.4 commit 823189e

.github/workflows/run-sweep.yml

@@ -416,6 +416,7 @@ jobs:
             decode-additional-settings: ${{ toJson(matrix.config.decode.additional-settings) }}
             conc: ${{ matrix.config.conc }}
             duration: ${{ matrix.config.duration }}
+            offloading: ${{ matrix.config.offloading }}
             run-eval: false
             scenario-type: agentic-coding
 

benchmarks/benchmark_lib.sh

@@ -935,9 +935,9 @@ resolve_trace_source() {
     # pin a specific corpus via WEKA_LOADER_OVERRIDE.
     local default_loader
     if [[ "${MODEL_PREFIX:-}" == dsv4* ]]; then
-        default_loader="semianalysis_cc_traces_weka_with_subagents_060826"
+        default_loader="semianalysis_cc_traces_weka_061526"
     else
-        default_loader="semianalysis_cc_traces_weka_with_subagents_060826_256k"
+        default_loader="semianalysis_cc_traces_weka_061526_256k"
     fi
     local loader="${WEKA_LOADER_OVERRIDE:-$default_loader}"
     local dataset
@@ -966,8 +966,20 @@ resolve_trace_source() {
         semianalysis_cc_traces_weka_with_subagents_060826_256k)
             dataset="semianalysisai/cc-traces-weka-with-subagents-060826-256k"
             ;;
+        semianalysis_cc_traces_weka_061326)
+            dataset="semianalysisai/cc-traces-weka-061326"
+            ;;
+        semianalysis_cc_traces_weka_061326_256k)
+            dataset="semianalysisai/cc-traces-weka-061326-256k"
+            ;;
+        semianalysis_cc_traces_weka_061526)
+            dataset="semianalysisai/cc-traces-weka-061526"
+            ;;
+        semianalysis_cc_traces_weka_061526_256k)
+            dataset="semianalysisai/cc-traces-weka-061526-256k"
+            ;;
         *)
-            echo "Error: unknown WEKA_LOADER_OVERRIDE='$loader'. Allowed: semianalysis_cc_traces_weka_with_subagents, semianalysis_cc_traces_weka_with_subagents_256k, semianalysis_cc_traces_weka_with_subagents_060226, semianalysis_cc_traces_weka_with_subagents_060226_256k, semianalysis_cc_traces_weka_with_subagents_060526, semianalysis_cc_traces_weka_with_subagents_060526_256k, semianalysis_cc_traces_weka_with_subagents_060826, semianalysis_cc_traces_weka_with_subagents_060826_256k" >&2
+            echo "Error: unknown WEKA_LOADER_OVERRIDE='$loader'. Allowed: semianalysis_cc_traces_weka_with_subagents, semianalysis_cc_traces_weka_with_subagents_256k, semianalysis_cc_traces_weka_with_subagents_060226, semianalysis_cc_traces_weka_with_subagents_060226_256k, semianalysis_cc_traces_weka_with_subagents_060526, semianalysis_cc_traces_weka_with_subagents_060526_256k, semianalysis_cc_traces_weka_with_subagents_060826, semianalysis_cc_traces_weka_with_subagents_060826_256k, semianalysis_cc_traces_weka_061326, semianalysis_cc_traces_weka_061326_256k, semianalysis_cc_traces_weka_061526, semianalysis_cc_traces_weka_061526_256k" >&2
             exit 1
             ;;
     esac

benchmarks/multi_node/agentic/dsr1_fp4_mi355x_sglang-disagg.sh

@@ -0,0 +1,176 @@
+#!/usr/bin/env bash
+
+# Agentic trace-replay recipe for a disaggregated SGLang server on MI355X
+# (DeepSeek-R1-0528 MXFP4-v2, 1P1D TP8).
+#
+# CI-style sibling of dsr1_fp8_mi355x_sglang-disagg.sh: driven entirely by
+# environment variables and submits a SLURM job via submit.sh. The agentic /
+# HiCache-offload configuration is ported from local_test_dsr1_agentic_offload.sh
+# and is fully env-overridable so a YAML config can tune it.
+
+source "$(dirname "$0")/../benchmark_lib.sh"
+
+check_env_vars \
+    CONC_LIST \
+    ISL \
+    OSL \
+    IMAGE \
+    SPEC_DECODING \
+    MODEL_PATH \
+    PREFILL_NUM_WORKERS \
+    PREFILL_TP \
+    PREFILL_EP \
+    PREFILL_DP_ATTN \
+    DECODE_NUM_WORKERS \
+    DECODE_TP \
+    DECODE_EP \
+    DECODE_DP_ATTN \
+    PREFILL_NODES \
+    DECODE_NODES \
+    RANDOM_RANGE_RATIO \
+    DURATION \
+    OFFLOADING \
+    IS_AGENTIC \
+    FRAMEWORK
+
+if [[ -n "$SLURM_JOB_ID" ]]; then
+  echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
+fi
+
+set -x
+
+# Use upstreamed multi_node scripts (no external clone needed)
+cd "$GITHUB_WORKSPACE/benchmarks/multi_node/amd_utils" || exit 1
+
+# Set up SGL launch script-specific environment variables
+export TIME_LIMIT="${TIME_LIMIT:-08:00:00}"
+export MODEL_PATH=$MODEL_PATH
+export MODEL_NAME=$MODEL_NAME
+export CONTAINER_IMAGE=$IMAGE
+
+# ── Identity / result naming ──
+export MODEL_PREFIX="${MODEL_PREFIX:-dsr1}"
+export PRECISION="${PRECISION:-fp4}"
+export RESULT_FILENAME="${RESULT_FILENAME:-${RUNNER_NAME:-dsr1-fp4-agentic}}"
+
+# ── Agentic benchmark params ──
+# DURATION threads through submit.sh -> job.slurm -> Docker -> bench.sh.
+# CONC_LIST drives the concurrency sweep (submit.sh splits on 'x').
+export DURATION="${DURATION:-1800}"
+export MAX_MODEL_LEN="${MAX_MODEL_LEN:-163840}"
+
+# ── Aiter fault mitigations (ROCm/ROCm#6023) ──
+export SGLANG_AITER_MLA_PERSIST="${SGLANG_AITER_MLA_PERSIST:-1}"
+# 1 => append --disable-custom-all-reduce to prefill+decode (Aiter fault mitigation).
+export DISABLE_CUSTOM_ALL_REDUCE="${DISABLE_CUSTOM_ALL_REDUCE:-1}"
+
+# # ── Hugging Face cache persistence ──
+# # Persist the HF Hub/datasets cache across runs so traces aren't re-downloaded.
+# export HF_CACHE_HOST_DIR="${HF_CACHE_HOST_DIR:-$HOME/.cache/huggingface}"
+# mkdir -p "${HF_CACHE_HOST_DIR}"
+# export EXTRA_DOCKER_MOUNTS="${EXTRA_DOCKER_MOUNTS:-} -v ${HF_CACHE_HOST_DIR}:/root/.cache/huggingface"
+# # HF auth token: provide via the environment/CI secrets (do NOT hardcode here).
+# export HF_TOKEN="${HF_TOKEN:-}"
+# if [[ -n "${HF_TOKEN:-}" && -z "${HUGGING_FACE_HUB_TOKEN:-}" ]]; then
+#   export HUGGING_FACE_HUB_TOKEN="$HF_TOKEN"
+# fi
+
+# ── In-tree sglang patches ──
+# mori_conn.py targets hybrid-state bugs (GLM-5, Qwen3.5) not present in
+# DSR1-MXFP4-v2 (pure MLA). Skip the auto-apply in job.slurm.
+export MORI_CONN_PATCH="${MORI_CONN_PATCH:-skip}"
+
+# ── KV cache offloading (HiCache) ──
+# OFFLOADING=hicache (default for this recipe) | none. HICACHE_TIER:
+#   L2 -> GPU + CPU-DRAM host pool only.   L3 -> + Mooncake distributed KV store.
+export OFFLOADING="${OFFLOADING:-hicache}"
+export HICACHE_TIER="${HICACHE_TIER:-L3}"
+export HICACHE_TOTAL_CPU_DRAM_GB="${HICACHE_TOTAL_CPU_DRAM_GB:-64}"
+export HICACHE_HOST_POOL_COUNT="${HICACHE_HOST_POOL_COUNT:-1}"
+export HICACHE_PAGE_SIZE="${HICACHE_PAGE_SIZE:-64}"
+# Per-rank L2 host pool in GB (100GB/rank x TP8 = ~800GB pinned host DRAM/node).
+export HICACHE_SIZE_GB="${HICACHE_SIZE_GB:-100}"
+
+# ── HiCache layout/backend driven by HICACHE_TIER ──
+# Each tier has a canonical (layout, io_backend, write_policy, storage_backend)
+# combo (mirrors server_sglang.sh build_hicache_flags). Any var set explicitly
+# in the environment wins over the tier default.
+#   L3 (Mooncake): page_first + direct + write_through        + storage=mooncake
+#   L2 (CPU DRAM): layer_first + kernel + write_through_selective + storage=none
+if [[ "${HICACHE_TIER^^}" == "L3" ]]; then
+  export HICACHE_MEM_LAYOUT="${HICACHE_MEM_LAYOUT:-page_first}"
+  export HICACHE_IO_BACKEND="${HICACHE_IO_BACKEND:-direct}"
+  export HICACHE_WRITE_POLICY="${HICACHE_WRITE_POLICY:-write_through}"
+  export HICACHE_STORAGE_BACKEND="${HICACHE_STORAGE_BACKEND:-mooncake}"
+else
+  export HICACHE_MEM_LAYOUT="${HICACHE_MEM_LAYOUT:-layer_first}"
+  export HICACHE_IO_BACKEND="${HICACHE_IO_BACKEND:-direct}"
+  # write_through_selective evicts only under GPU memory pressure, giving mori
+  # time to complete RDMA KV transfers before pages are freed. write_through
+  # evicts immediately and races with mori → GPU memory access faults.
+  export HICACHE_WRITE_POLICY="${HICACHE_WRITE_POLICY:-write_through_selective}"
+  export HICACHE_STORAGE_BACKEND="${HICACHE_STORAGE_BACKEND:-}"
+fi
+export HICACHE_DECODE="${HICACHE_DECODE:-0}"
+# Shared nodes: use non-default Mooncake ports to avoid colliding with other users.
+export MC_MASTER_PORT="${MC_MASTER_PORT:-58137}"
+export MC_METRICS_PORT="${MC_METRICS_PORT:-19003}"
+export MC_PATCH_HOSTPOOL="${MC_PATCH_HOSTPOOL:-1}"
+export MC_PROTOCOL="${MC_PROTOCOL:-tcp}"
+export MC_GLOBAL_SEG="${MC_GLOBAL_SEG:-30gb}"
+export MC_DEVICE="${MC_DEVICE:-rdma0}"
+export MC_MASTER_ADDR="${MC_MASTER_ADDR:-}"
+
+# ── MoRIIO RDMA Send Queue tuning (headroom for conc>=8) ──
+export MORI_IO_SQ_BACKOFF_TIMEOUT_US="${MORI_IO_SQ_BACKOFF_TIMEOUT_US:-500000}"
+export MORI_IO_QP_MAX_SEND_WR="${MORI_IO_QP_MAX_SEND_WR:-32768}"
+
+# ── SGLang PD router policy + server metrics ──
+export PREFILL_ROUTER_POLICY="${PREFILL_ROUTER_POLICY:-random}"
+export ENABLE_METRICS="${ENABLE_METRICS:-1}"
+
+# ── MTP ──
+export DECODE_MTP_SIZE="${DECODE_MTP_SIZE:-0}"
+
+# Derive EP/DP enable flags from the topology inputs (same as the fixed-seq recipe).
+if [[ "${PREFILL_EP:-1}" -eq 1 ]]; then
+export PREFILL_ENABLE_EP=false
+else
+export PREFILL_ENABLE_EP=true
+fi
+
+if [[ "$PREFILL_DP_ATTN" == "true" ]]; then
+export PREFILL_ENABLE_DP=true
+else
+export PREFILL_ENABLE_DP=false
+fi
+
+if [[ "${DECODE_EP:-1}" -eq 1 ]]; then
+export DECODE_ENABLE_EP=false
+else
+export DECODE_ENABLE_EP=true
+fi
+
+if [[ "$DECODE_DP_ATTN" == "true" ]]; then
+export DECODE_ENABLE_DP=true
+else
+export DECODE_ENABLE_DP=false
+fi
+
+# Launch the job. CONC_LIST is space-delimited in YAML; submit.sh wants 'x'.
+JOB_ID=$(bash ./submit.sh $PREFILL_NODES \
+    $PREFILL_NUM_WORKERS \
+    $DECODE_NODES \
+    $DECODE_NUM_WORKERS \
+    $ISL $OSL "${CONC_LIST// /x}" inf \
+    ${PREFILL_ENABLE_EP} ${PREFILL_ENABLE_DP} \
+    ${DECODE_ENABLE_EP} ${DECODE_ENABLE_DP} \
+    ${PREFILL_TP} ${DECODE_TP} \
+    ${RANDOM_RANGE_RATIO})
+
+if [[ $? -ne 0 ]]; then
+    echo "Failed to submit job" >&2
+    exit 1
+fi
+
+echo "$JOB_ID"

benchmarks/multi_node/agentic/dsv4_fp4_mi355x_sglang-disagg.sh

@@ -0,0 +1,166 @@
+#!/usr/bin/env bash
+
+# Agentic trace-replay recipe for a disaggregated SGLang server on MI355X
+# (DeepSeek-V4-Pro FP4, 1P1D TP8).
+#
+# CI-style sibling of dsr1_fp4_mi355x_sglang-disagg.sh: driven entirely by
+# environment variables and submits a SLURM job via submit.sh. The agentic /
+# HiCache-offload configuration mirrors the DSR1 recipe but uses DSV4-Pro
+# specific flags (dsv4 attention backend, page-size 256, SWA settings).
+
+source "$(dirname "$0")/../benchmark_lib.sh"
+
+check_env_vars \
+    CONC_LIST \
+    ISL \
+    OSL \
+    IMAGE \
+    SPEC_DECODING \
+    MODEL_PATH \
+    PREFILL_NUM_WORKERS \
+    PREFILL_TP \
+    PREFILL_EP \
+    PREFILL_DP_ATTN \
+    DECODE_NUM_WORKERS \
+    DECODE_TP \
+    DECODE_EP \
+    DECODE_DP_ATTN \
+    PREFILL_NODES \
+    DECODE_NODES \
+    RANDOM_RANGE_RATIO \
+    DURATION \
+    OFFLOADING \
+    IS_AGENTIC \
+    FRAMEWORK
+
+if [[ -n "$SLURM_JOB_ID" ]]; then
+  echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
+fi
+
+set -x
+
+# Use upstreamed multi_node scripts (no external clone needed)
+cd "$GITHUB_WORKSPACE/benchmarks/multi_node/amd_utils" || exit 1
+
+# Set up SGL launch script-specific environment variables
+export TIME_LIMIT="${TIME_LIMIT:-08:00:00}"
+export MODEL_PATH=$MODEL_PATH
+export MODEL_NAME=$MODEL_NAME
+export CONTAINER_IMAGE=$IMAGE
+
+# ── Identity / result naming ──
+export MODEL_PREFIX="${MODEL_PREFIX:-dsv4}"
+export PRECISION="${PRECISION:-fp4}"
+export RESULT_FILENAME="${RESULT_FILENAME:-${RUNNER_NAME:-dsv4-fp4-agentic}}"
+
+# ── Agentic benchmark params ──
+export DURATION="${DURATION:-1800}"
+# DSV4-Pro max model len for agentic traces (matches single-node recipe).
+export MAX_MODEL_LEN="${MAX_MODEL_LEN:-1000000}"
+
+# ── In-tree sglang patches ──
+# mori_conn.py targets hybrid-state bugs (GLM-5, Qwen3.5). DSV4-Pro uses a
+# pure MoE/DSA architecture without hybrid state; skip to avoid interference.
+export MORI_CONN_PATCH="${MORI_CONN_PATCH:-skip}"
+
+# ── Aiter fault mitigation ──
+# --disable-custom-all-reduce avoids a known aiter fault on MI355X.
+export DISABLE_CUSTOM_ALL_REDUCE="${DISABLE_CUSTOM_ALL_REDUCE:-0}"
+
+# ── KV cache offloading (HiCache) ──
+# OFFLOADING=hicache | none (passed from YAML; default none for disagg).
+# HICACHE_TIER: L2 -> GPU + CPU-DRAM host pool. L3 -> + Mooncake store.
+export OFFLOADING="${OFFLOADING:-none}"
+# HiCache/Mooncake tunables only matter when KV offloading is enabled.
+if [[ "$OFFLOADING" == "hicache" ]]; then
+  export HICACHE_TIER="${HICACHE_TIER:-L2}"
+  export HICACHE_HOST_POOL_COUNT="${HICACHE_HOST_POOL_COUNT:-1}"
+  # DSV4 uses page-size 256 (set in models.yaml); HiCache must match.
+  export HICACHE_PAGE_SIZE="${HICACHE_PAGE_SIZE:-256}"
+  # HiCache ratio (host pool = ratio * GPU KV pool). Default derived in server_sglang.sh.
+  export HICACHE_RATIO="${HICACHE_RATIO:-}"
+
+  # ── HiCache layout/backend by tier ──
+  #   L3 (Mooncake): page_first + direct + write_through     + storage=mooncake
+  #   L2 (CPU DRAM): layer_first + direct + write_through_selective + storage=none
+  # NOTE: write_through_selective evicts only under GPU memory pressure, avoiding
+  # the mori RDMA race that causes GPU memory access faults with write_through.
+  if [[ "${HICACHE_TIER^^}" == "L3" ]]; then
+    export HICACHE_MEM_LAYOUT="${HICACHE_MEM_LAYOUT:-page_first}"
+    export HICACHE_IO_BACKEND="${HICACHE_IO_BACKEND:-direct}"
+    export HICACHE_WRITE_POLICY="${HICACHE_WRITE_POLICY:-write_through}"
+    export HICACHE_STORAGE_BACKEND="${HICACHE_STORAGE_BACKEND:-mooncake}"
+  else
+    export HICACHE_MEM_LAYOUT="${HICACHE_MEM_LAYOUT:-layer_first}"
+    export HICACHE_IO_BACKEND="${HICACHE_IO_BACKEND:-direct}"
+    export HICACHE_WRITE_POLICY="${HICACHE_WRITE_POLICY:-write_through_selective}"
+    export HICACHE_STORAGE_BACKEND="${HICACHE_STORAGE_BACKEND:-}"
+  fi
+  export HICACHE_PREFETCH_POLICY="${HICACHE_PREFETCH_POLICY:-wait_complete}"
+  # Shared nodes: use non-default Mooncake ports to avoid collisions.
+  export MC_MASTER_PORT="${MC_MASTER_PORT:-58137}"
+  export MC_METADATA_PORT="${MC_METADATA_PORT:-8080}"
+  export MC_METRICS_PORT="${MC_METRICS_PORT:-19003}"
+  export MC_MASTER_THREADS="${MC_MASTER_THREADS:-64}"
+  export MC_EVICTION_HIGH_WATERMARK="${MC_EVICTION_HIGH_WATERMARK:-0.95}"
+  export MC_PATCH_HOSTPOOL="${MC_PATCH_HOSTPOOL:-1}"
+  export MC_PROTOCOL="${MC_PROTOCOL:-tcp}"
+  export MC_GLOBAL_SEG="${MC_GLOBAL_SEG:-64gb}"
+  export MC_DEVICE="${MC_DEVICE:-}"
+  export MC_MASTER_ADDR="${MC_MASTER_ADDR:-}"
+  export MC_METADATA_SERVER="${MC_METADATA_SERVER:-}"
+fi
+
+# ── MoRIIO RDMA Send Queue tuning ──
+export MORI_IO_SQ_BACKOFF_TIMEOUT_US="${MORI_IO_SQ_BACKOFF_TIMEOUT_US:-500000}"
+export MORI_IO_QP_MAX_SEND_WR="${MORI_IO_QP_MAX_SEND_WR:-32768}"
+
+# ── SGLang PD router policy + server metrics ──
+export PREFILL_ROUTER_POLICY="${PREFILL_ROUTER_POLICY:-random}"
+export ENABLE_METRICS="${ENABLE_METRICS:-1}"
+
+# ── MTP ──
+export DECODE_MTP_SIZE="${DECODE_MTP_SIZE:-0}"
+
+# Derive EP/DP enable flags from the topology inputs.
+if [[ "${PREFILL_EP:-1}" -eq 1 ]]; then
+export PREFILL_ENABLE_EP=false
+else
+export PREFILL_ENABLE_EP=true
+fi
+
+if [[ "$PREFILL_DP_ATTN" == "true" ]]; then
+export PREFILL_ENABLE_DP=true
+else
+export PREFILL_ENABLE_DP=false
+fi
+
+if [[ "${DECODE_EP:-1}" -eq 1 ]]; then
+export DECODE_ENABLE_EP=false
+else
+export DECODE_ENABLE_EP=true
+fi
+
+if [[ "$DECODE_DP_ATTN" == "true" ]]; then
+export DECODE_ENABLE_DP=true
+else
+export DECODE_ENABLE_DP=false
+fi
+
+# Launch the job. CONC_LIST is space-delimited in YAML; submit.sh wants 'x'.
+JOB_ID=$(bash ./submit.sh $PREFILL_NODES \
+    $PREFILL_NUM_WORKERS \
+    $DECODE_NODES \
+    $DECODE_NUM_WORKERS \
+    $ISL $OSL "${CONC_LIST// /x}" inf \
+    ${PREFILL_ENABLE_EP} ${PREFILL_ENABLE_DP} \
+    ${DECODE_ENABLE_EP} ${DECODE_ENABLE_DP} \
+    ${PREFILL_TP} ${DECODE_TP} \
+    ${RANDOM_RANGE_RATIO})
+
+if [[ $? -ne 0 ]]; then
+    echo "Failed to submit job" >&2
+    exit 1
+fi
+
+echo "$JOB_ID"

benchmarks/multi_node/amd_utils/bench.sh

@@ -79,7 +79,12 @@ for max_concurrency in "${chosen_concurrencies[@]}"; do
     if [[ "$ENGINE" == "vllm-disagg" ]]; then
         extra_flags="--trust-remote-code --tokenizer $MODEL_PATH"
     else
-        if [ "$IS_MTP" = "true" ]; then
+        # DeepSeek-V4-Pro ships no jinja chat_template, so --use-chat-template crashes;
+        # --dsv4 applies the DSv4 <bos><User>...<Assistant><think> framing instead
+        # (chat-formatted inputs are required for correct EAGLE/MTP acceptance too).
+        if [[ "$model_name" == "DeepSeek-V4-Pro" ]]; then
+            extra_flags="--dsv4"
+        elif [ "$IS_MTP" = "true" ]; then
             extra_flags="--use-chat-template"
         fi
     fi