From dad9566b1b904599851ab58c157adda4ad302db3 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Sun, 9 Aug 2026 17:04:44 -0500
Subject: [PATCH 01/18] perf(agentx): add DSv4 MTP frontier on MI325X

---
 .../agentic/dsv4_fp8_mi325x_mtp.sh            | 149 ++++++++++++++++++
 configs/amd-master.yaml                       |  22 +++
 2 files changed, 171 insertions(+)
 create mode 100755 benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh

diff --git a/benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh b/benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh
new file mode 100755
index 000000000..b5188a03c
--- /dev/null
+++ b/benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh
@@ -0,0 +1,149 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+# DeepSeek-V4-Pro FP8 AgentX replay on one 8xMI325X node. The checkpoint is
+# dequantized to FP8 because gfx942 has no native MXFP4 support. Both supported
+# one-node layouts retain MTP: TP8 for the latency frontier, and DP8+EP8 for the
+# high-throughput frontier.
+
+source "$(dirname "$0")/../../benchmark_lib.sh"
+
+check_env_vars MODEL IMAGE TP CONC KV_OFFLOADING RESULT_DIR DURATION EP_SIZE DP_ATTENTION
+
+if [[ "$KV_OFFLOADING" != "none" ]]; then
+    echo "ERROR: DeepSeek-V4 MTP on MI325X currently supports GPU-resident KV only" >&2
+    exit 1
+fi
+
+if [[ -n "${SLURM_JOB_ID:-}" ]]; then
+    echo "JOB $SLURM_JOB_ID running on ${SLURMD_NODENAME:-unknown}"
+fi
+
+if [[ -n "${ROCR_VISIBLE_DEVICES:-}" ]]; then
+    export HIP_VISIBLE_DEVICES="$ROCR_VISIBLE_DEVICES"
+fi
+
+if [[ -n "${MODEL_PATH:-}" ]]; then
+    if [[ ! -d "$MODEL_PATH" || -z "$(ls -A "$MODEL_PATH" 2>/dev/null)" ]]; then
+        hf download "$MODEL" --local-dir "$MODEL_PATH"
+    fi
+else
+    hf download "$MODEL"
+    export MODEL_PATH="$MODEL"
+fi
+
+resolve_trace_source
+install_agentic_deps
+agentic_pip_install --quiet Pillow fastapi uvicorn
+
+export AIPERF_HTTP_TCP_USER_TIMEOUT=900000
+export AIPERF_SERVER_METRICS_URLS="http://localhost:${PORT}/metrics"
+export AIPERF_REQUIRED_SERVER_METRIC_PREFIX="vllm:"
+export VLLM_ENGINE_READY_TIMEOUT_S=10800
+export VLLM_ROCM_USE_AITER=1
+export VLLM_ROCM_USE_AITER_MOE=1
+export PYTHONNOUSERSITE=1
+
+SERVER_LOG="$RESULT_DIR/server.log"
+ROUTER_LOG="$RESULT_DIR/router.log"
+mkdir -p "$RESULT_DIR"
+
+PARALLEL_ARGS=(--tensor-parallel-size "$TP" --data-parallel-size 1)
+if [[ "$DP_ATTENTION" == "true" ]]; then
+    PARALLEL_ARGS=(--tensor-parallel-size 1 --data-parallel-size "$TP")
+fi
+
+EP_ARGS=()
+if (( EP_SIZE > 1 )); then
+    EP_ARGS=(--enable-expert-parallel)
+fi
+
+USE_VLLM_ROUTER=false
+VLLM_BACKEND_PORT="$PORT"
+ROUTER_PID=""
+if [[ "$DP_ATTENTION" == "true" ]]; then
+    if (( EP_SIZE != TP )); then
+        echo "ERROR: MI325X DP-attention requires EP_SIZE == TP so FP8 experts remain sharded" >&2
+        exit 1
+    fi
+    USE_VLLM_ROUTER=true
+    VLLM_BACKEND_PORT=$((PORT + 1))
+    export AIPERF_HTTP_X_SESSION_ID_FROM_CORRELATION_ID=1
+    export AIPERF_SERVER_METRICS_URLS="http://localhost:${VLLM_BACKEND_PORT}/metrics"
+    agentic_pip_install --quiet 'vllm-router==0.1.14'
+fi
+
+MAX_NUM_SEQS=$((2 * CONC))
+NUM_SPEC_TOKENS=3
+SYNTHETIC_ACCEPT_LEN=2.49
+if [[ "${EVAL_ONLY:-false}" == "true" ]]; then
+    SPEC_CONFIG="{\"method\":\"mtp\",\"num_speculative_tokens\":${NUM_SPEC_TOKENS}}"
+else
+    SPEC_CONFIG="{\"method\":\"mtp\",\"num_speculative_tokens\":${NUM_SPEC_TOKENS},\"rejection_sample_method\":\"synthetic\",\"synthetic_acceptance_length\":${SYNTHETIC_ACCEPT_LEN}}"
+fi
+
+cleanup() {
+    local rc=$?
+    trap - EXIT INT TERM
+    set +e
+    stop_background_process_tree "$ROUTER_PID" "vLLM router"
+    stop_background_process_tree "${SERVER_PID:-}" "vLLM server" 60
+    exit "$rc"
+}
+trap cleanup EXIT
+trap 'exit 130' INT
+trap 'exit 143' TERM
+
+VLLM_CMD=(
+    vllm serve "$MODEL_PATH" --served-model-name "$MODEL"
+    --host 0.0.0.0
+    --port "$VLLM_BACKEND_PORT"
+    --trust-remote-code
+    --async-scheduling
+    --distributed-executor-backend mp
+    --quantization deepseek_v4_fp8
+    --kv-cache-dtype fp8
+    "${PARALLEL_ARGS[@]}"
+    "${EP_ARGS[@]}"
+    --gpu-memory-utilization 0.9
+    --block-size 256
+    --max-num-batched-tokens 8192
+    --max-num-seqs "$MAX_NUM_SEQS"
+    --compilation-config '{"mode":3,"cudagraph_mode":"FULL_DECODE_ONLY"}'
+    --speculative-config "$SPEC_CONFIG"
+    --tokenizer-mode deepseek_v4
+    --tool-call-parser deepseek_v4
+    --reasoning-parser deepseek_v4
+    --enable-auto-tool-choice
+    --enable-prefix-caching
+    --no-disable-hybrid-kv-cache-manager
+)
+
+printf '%q ' "${VLLM_CMD[@]}" > "$RESULT_DIR/vllm_command.txt"
+printf '\n' >> "$RESULT_DIR/vllm_command.txt"
+"${VLLM_CMD[@]}" > "$SERVER_LOG" 2>&1 &
+SERVER_PID=$!
+
+wait_for_server_ready --port "$VLLM_BACKEND_PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
+
+if [[ "$USE_VLLM_ROUTER" == "true" ]]; then
+    vllm-router \
+        --worker-urls "http://localhost:$VLLM_BACKEND_PORT" \
+        --policy consistent_hash \
+        --intra-node-data-parallel-size "$TP" \
+        --host 0.0.0.0 \
+        --port "$PORT" \
+        --prometheus-host 127.0.0.1 \
+        --prometheus-port "$((PORT + 10000))" \
+        --request-timeout-secs 14400 \
+        --disable-retries > "$ROUTER_LOG" 2>&1 &
+    ROUTER_PID=$!
+    wait_for_server_ready --port "$PORT" --server-log "$ROUTER_LOG" --server-pid "$ROUTER_PID"
+fi
+
+if [[ "${EVAL_ONLY:-false}" == "true" ]]; then
+    run_eval --port "$PORT"
+else
+    build_replay_cmd "$RESULT_DIR"
+    run_agentic_replay_and_write_outputs "$RESULT_DIR"
+fi
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index c6580af43..971729e8f 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -1471,6 +1471,28 @@ glm5.2-fp8-mi325x-sglang-agentic-mtp:
       search-space:
       - { tp: 8, ep: 1, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 3, 4, 5, 6, 8] }
 
+# DeepSeek-V4-Pro FP8 full-context AgentX frontier on one MI325X node. The
+# dequantized checkpoint is roughly 1.05 TB, so TP8 fits comfortably in the
+# node's 2 TB of HBM while TP4 does not have enough per-rank model headroom.
+# Sample the latency layouts densely around the estimated 20-30 live-tree KV
+# knee, then use DP-attention + EP8 to probe aggregate throughput without
+# replicating the full FP8 expert set on every GPU.
+dsv4-fp8-mi325x-vllm-agentic-mtp:
+  image: vllm/vllm-openai-rocm:v0.26.0
+  model: deepseek-ai/DeepSeek-V4-Pro
+  model-prefix: dsv4
+  runner: cluster:mi325x-amds
+  precision: fp8
+  framework: vllm
+  multinode: false
+  scenarios:
+    agentic-coding:
+    - dram-utilization: 0.90
+      search-space:
+      - { tp: 8, ep: 1, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 28, 32] }
+      - { tp: 8, ep: 8, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [4, 8, 12, 16, 20, 24, 28, 32, 36, 40] }
+      - { tp: 8, ep: 8, dp-attn: true, kv-offloading: none, spec-decoding: mtp, conc-list: [16, 24, 32, 40, 48, 56, 64, 72, 80], router: { name: vllm-router, version: "0.1.14" } }
+
 minimaxm3-fp8-mi325x-vllm-agentic:
   image: vllm/vllm-openai-rocm:nightly-04c2a8deac44fdb1ca3e2b5ec3e6bf16f3f6a914
   model: MiniMaxAI/MiniMax-M3-MXFP8

From 4fb408ee4a75bd6293dad10f0e8450dac951bb05 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Sun, 9 Aug 2026 17:05:26 -0500
Subject: [PATCH 02/18] docs(perf): register MI325X DSv4 AgentX submission

---
 perf-changelog.yaml | 11 +++++++++++
 1 file changed, 11 insertions(+)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 471981f60..3e7e3d369 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5709,3 +5709,14 @@
     - "Enable SGLang metrics on every aggregate, prefill, and decode engine."
     - "Use supported header-based Dynamo session routing with the in-repo AIPerf build."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2520
+
+- config-keys:
+    - dsv4-fp8-mi325x-vllm-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Add the first DeepSeek-V4-Pro FP8 vLLM AgentX submission on MI325X, with MTP enabled for every published point and the committed three-token golden synthetic acceptance length of 2.49; eval-only runs retain real target verification."
+    - "Use one 8xMI325X node: the roughly 1.05 TB FP8-dequantized checkpoint fits in the node's 2 TB HBM at TP8, so multi-node execution would add communication without solving a capacity constraint; TP4 is excluded because its per-rank model footprint is not fit-safe."
+    - "Cover the full one-node frontier with dense TP8, TP8+EP8 and DP8+EP8 bands through concurrency 80. The DP-attention arm requires EP8 so the FP8 expert set remains sharded instead of being replicated on every 256 GB GPU."
+    - "Start from the latest stable ROCm vLLM v0.26.0 release, use FP8 KV, 90% GPU-memory utilization, block size 256, an 8192-token scheduler budget and FULL_DECODE_ONLY graph capture, and require a nonempty vllm: server-metrics export from AIPerf."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2541

From e4150e6803aeb30cb35a732f9a043186cede0e85 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Sun, 9 Aug 2026 17:26:23 -0500
Subject: [PATCH 03/18] bench: report vLLM prompt cache usage on MI325X

---
 benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh | 1 +
 perf-changelog.yaml                                   | 2 +-
 2 files changed, 2 insertions(+), 1 deletion(-)

diff --git a/benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh b/benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh
index b5188a03c..6b7c9fe40 100755
--- a/benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh
@@ -116,6 +116,7 @@ VLLM_CMD=(
     --reasoning-parser deepseek_v4
     --enable-auto-tool-choice
     --enable-prefix-caching
+    --enable-prompt-tokens-details
     --no-disable-hybrid-kv-cache-manager
 )
 
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 3e7e3d369..89da01c14 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5718,5 +5718,5 @@
     - "Add the first DeepSeek-V4-Pro FP8 vLLM AgentX submission on MI325X, with MTP enabled for every published point and the committed three-token golden synthetic acceptance length of 2.49; eval-only runs retain real target verification."
     - "Use one 8xMI325X node: the roughly 1.05 TB FP8-dequantized checkpoint fits in the node's 2 TB HBM at TP8, so multi-node execution would add communication without solving a capacity constraint; TP4 is excluded because its per-rank model footprint is not fit-safe."
     - "Cover the full one-node frontier with dense TP8, TP8+EP8 and DP8+EP8 bands through concurrency 80. The DP-attention arm requires EP8 so the FP8 expert set remains sharded instead of being replicated on every 256 GB GPU."
-    - "Start from the latest stable ROCm vLLM v0.26.0 release, use FP8 KV, 90% GPU-memory utilization, block size 256, an 8192-token scheduler budget and FULL_DECODE_ONLY graph capture, and require a nonempty vllm: server-metrics export from AIPerf."
+    - "Start from the latest stable ROCm vLLM v0.26.0 release, use FP8 KV, 90% GPU-memory utilization, block size 256, an 8192-token scheduler budget and FULL_DECODE_ONLY graph capture. Enable prompt-token details for request-level cache accounting, and require a nonempty vllm: server-metrics export from AIPerf."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2541

From d171f9831a77fd7b029d81c3cf492308ae903b4f Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Sun, 9 Aug 2026 20:00:33 -0500
Subject: [PATCH 04/18] perf(mi325x): screen DSV4 MTP2 correctness

---
 .../single_node/agentic/dsv4_fp8_mi325x_mtp.sh      |  8 ++++++--
 configs/amd-master.yaml                             | 13 +++++++------
 perf-changelog.yaml                                 |  5 +++--
 3 files changed, 16 insertions(+), 10 deletions(-)

diff --git a/benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh b/benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh
index 6b7c9fe40..8136e516a 100755
--- a/benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh
@@ -74,8 +74,12 @@ if [[ "$DP_ATTENTION" == "true" ]]; then
 fi
 
 MAX_NUM_SEQS=$((2 * CONC))
-NUM_SPEC_TOKENS=3
-SYNTHETIC_ACCEPT_LEN=2.49
+# The first AgentX-fast pass used K=3 and exposed empty-content responses on
+# the exploratory gfx942 deepseek_v4_fp8 path.  The existing MI300X/MI325X
+# fixed-sequence recipes use K=2; screen that supported shape independently
+# before changing the stable vLLM image.
+NUM_SPEC_TOKENS=2
+SYNTHETIC_ACCEPT_LEN=2.27
 if [[ "${EVAL_ONLY:-false}" == "true" ]]; then
     SPEC_CONFIG="{\"method\":\"mtp\",\"num_speculative_tokens\":${NUM_SPEC_TOKENS}}"
 else
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 971729e8f..191c811e8 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -1474,9 +1474,10 @@ glm5.2-fp8-mi325x-sglang-agentic-mtp:
 # DeepSeek-V4-Pro FP8 full-context AgentX frontier on one MI325X node. The
 # dequantized checkpoint is roughly 1.05 TB, so TP8 fits comfortably in the
 # node's 2 TB of HBM while TP4 does not have enough per-rank model headroom.
-# Sample the latency layouts densely around the estimated 20-30 live-tree KV
-# knee, then use DP-attention + EP8 to probe aggregate throughput without
-# replicating the full FP8 expert set on every GPU.
+# The broad K=3 AgentX-fast pass established the topology boundaries but also
+# exposed empty-content responses on TEP/DPA.  This K=2 diagnostic pass keeps
+# representative points from each topology so the speculative-depth effect is
+# isolated before a final dense frontier is published.
 dsv4-fp8-mi325x-vllm-agentic-mtp:
   image: vllm/vllm-openai-rocm:v0.26.0
   model: deepseek-ai/DeepSeek-V4-Pro
@@ -1489,9 +1490,9 @@ dsv4-fp8-mi325x-vllm-agentic-mtp:
     agentic-coding:
     - dram-utilization: 0.90
       search-space:
-      - { tp: 8, ep: 1, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 28, 32] }
-      - { tp: 8, ep: 8, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [4, 8, 12, 16, 20, 24, 28, 32, 36, 40] }
-      - { tp: 8, ep: 8, dp-attn: true, kv-offloading: none, spec-decoding: mtp, conc-list: [16, 24, 32, 40, 48, 56, 64, 72, 80], router: { name: vllm-router, version: "0.1.14" } }
+      - { tp: 8, ep: 1, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [8, 12, 16] }
+      - { tp: 8, ep: 8, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [20, 32] }
+      - { tp: 8, ep: 8, dp-attn: true, kv-offloading: none, spec-decoding: mtp, conc-list: [24, 48], router: { name: vllm-router, version: "0.1.14" } }
 
 minimaxm3-fp8-mi325x-vllm-agentic:
   image: vllm/vllm-openai-rocm:nightly-04c2a8deac44fdb1ca3e2b5ec3e6bf16f3f6a914
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 89da01c14..b80df95d0 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5715,8 +5715,9 @@
   scenario-type:
     - agentic-coding
   description:
-    - "Add the first DeepSeek-V4-Pro FP8 vLLM AgentX submission on MI325X, with MTP enabled for every published point and the committed three-token golden synthetic acceptance length of 2.49; eval-only runs retain real target verification."
+    - "Add the first DeepSeek-V4-Pro FP8 vLLM AgentX submission on MI325X, with MTP enabled for every published point; the current diagnostic pass uses two draft tokens and the committed K=2 golden synthetic acceptance length of 2.27, while eval-only runs retain real target verification."
     - "Use one 8xMI325X node: the roughly 1.05 TB FP8-dequantized checkpoint fits in the node's 2 TB HBM at TP8, so multi-node execution would add communication without solving a capacity constraint; TP4 is excluded because its per-rank model footprint is not fit-safe."
-    - "Cover the full one-node frontier with dense TP8, TP8+EP8 and DP8+EP8 bands through concurrency 80. The DP-attention arm requires EP8 so the FP8 expert set remains sharded instead of being replicated on every 256 GB GPU."
+    - "The first broad agentx-fast pass covered dense TP8, TP8+EP8 and DP8+EP8 bands through concurrency 80. It found clean-server duration-coverage failures on slow TP points, repeated empty-content semantic responses on TP8+EP8 and DP8+EP8, and a DP c40 warmup request that did not drain within 1800 seconds."
+    - "Run a second seven-point agentx-fast diagnostic on stable vLLM v0.26.0 at TP8 c8/c12/c16, TP8+EP8 c20/c32, and DP8+EP8 c24/c48. This changes only MTP depth from K=3 to K=2 (and golden AL 2.49 to 2.27) so speculative depth is isolated before considering a nightly image or publishing the final dense frontier."
     - "Start from the latest stable ROCm vLLM v0.26.0 release, use FP8 KV, 90% GPU-memory utilization, block size 256, an 8192-token scheduler budget and FULL_DECODE_ONLY graph capture. Enable prompt-token details for request-level cache accounting, and require a nonempty vllm: server-metrics export from AIPerf."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2541

From 54f86fe5636694c3c05a155a8dade79d3de95626 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Sun, 9 Aug 2026 20:42:29 -0500
Subject: [PATCH 05/18] perf(mi325x): test latest ROCm nightly

---
 configs/amd-master.yaml | 14 +++++++-------
 perf-changelog.yaml     |  4 +++-
 2 files changed, 10 insertions(+), 8 deletions(-)

diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 191c811e8..73f97c05f 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -1474,12 +1474,12 @@ glm5.2-fp8-mi325x-sglang-agentic-mtp:
 # DeepSeek-V4-Pro FP8 full-context AgentX frontier on one MI325X node. The
 # dequantized checkpoint is roughly 1.05 TB, so TP8 fits comfortably in the
 # node's 2 TB of HBM while TP4 does not have enough per-rank model headroom.
-# The broad K=3 AgentX-fast pass established the topology boundaries but also
-# exposed empty-content responses on TEP/DPA.  This K=2 diagnostic pass keeps
-# representative points from each topology so the speculative-depth effect is
-# isolated before a final dense frontier is published.
+# Stable v0.26.0 produced empty-content responses on every expert-parallel
+# topology at both K=3 and K=2.  This third AgentX-fast pass keeps K=2 and moves
+# only the image to the current immutable ROCm nightly, with one TP control and
+# the three minimal TEP/DPA reproductions.
 dsv4-fp8-mi325x-vllm-agentic-mtp:
-  image: vllm/vllm-openai-rocm:v0.26.0
+  image: vllm/vllm-openai-rocm:nightly-f8d03e77416bf90c49acbe50e233275722f02c4b
   model: deepseek-ai/DeepSeek-V4-Pro
   model-prefix: dsv4
   runner: cluster:mi325x-amds
@@ -1490,9 +1490,9 @@ dsv4-fp8-mi325x-vllm-agentic-mtp:
     agentic-coding:
     - dram-utilization: 0.90
       search-space:
-      - { tp: 8, ep: 1, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [8, 12, 16] }
+      - { tp: 8, ep: 1, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [8] }
       - { tp: 8, ep: 8, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [20, 32] }
-      - { tp: 8, ep: 8, dp-attn: true, kv-offloading: none, spec-decoding: mtp, conc-list: [24, 48], router: { name: vllm-router, version: "0.1.14" } }
+      - { tp: 8, ep: 8, dp-attn: true, kv-offloading: none, spec-decoding: mtp, conc-list: [24], router: { name: vllm-router, version: "0.1.14" } }
 
 minimaxm3-fp8-mi325x-vllm-agentic:
   image: vllm/vllm-openai-rocm:nightly-04c2a8deac44fdb1ca3e2b5ec3e6bf16f3f6a914
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index b80df95d0..d9fb88d5e 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5719,5 +5719,7 @@
     - "Use one 8xMI325X node: the roughly 1.05 TB FP8-dequantized checkpoint fits in the node's 2 TB HBM at TP8, so multi-node execution would add communication without solving a capacity constraint; TP4 is excluded because its per-rank model footprint is not fit-safe."
     - "The first broad agentx-fast pass covered dense TP8, TP8+EP8 and DP8+EP8 bands through concurrency 80. It found clean-server duration-coverage failures on slow TP points, repeated empty-content semantic responses on TP8+EP8 and DP8+EP8, and a DP c40 warmup request that did not drain within 1800 seconds."
     - "Run a second seven-point agentx-fast diagnostic on stable vLLM v0.26.0 at TP8 c8/c12/c16, TP8+EP8 c20/c32, and DP8+EP8 c24/c48. This changes only MTP depth from K=3 to K=2 (and golden AL 2.49 to 2.27) so speculative depth is isolated before considering a nightly image or publishing the final dense frontier."
-    - "Start from the latest stable ROCm vLLM v0.26.0 release, use FP8 KV, 90% GPU-memory utilization, block size 256, an 8192-token scheduler budget and FULL_DECODE_ONLY graph capture. Enable prompt-token details for request-level cache accounting, and require a nonempty vllm: server-metrics export from AIPerf."
+    - "The K=2 stable pass reproduced the defect: TEP c20 had 32 empty-content errors in 39 warmup records, TEP c32 had 45 in 57, and DPA c24 aborted after five profiling semantic errors; TP c8 remained semantically clean but produced only 2,993 input and 10.5 output tokens/s. This isolates stable-v0.26 deepseek_v4_fp8 expert parallelism rather than MTP depth."
+    - "Run a four-point third agentx-fast diagnostic on the current immutable ROCm nightly nightly-f8d03e77416bf90c49acbe50e233275722f02c4b: TP c8 control, TEP c20/c32, and DPA c24. Keep K=2/AL2.27 unchanged so the image is the only serving-stack variable."
+    - "Use FP8 KV, 90% GPU-memory utilization, block size 256, an 8192-token scheduler budget and FULL_DECODE_ONLY graph capture. Enable prompt-token details for request-level cache accounting, and require a nonempty vllm: server-metrics export from AIPerf."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2541

From 05beb909c0ff7c326233a26403f4bce9794d07f3 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Sun, 9 Aug 2026 21:32:51 -0500
Subject: [PATCH 06/18] perf(mi325x): narrow DSV4 MTP to valid TP frontier

---
 .../single_node/agentic/dsv4_fp8_mi325x_mtp.sh      | 12 +++++-------
 configs/amd-master.yaml                             | 13 ++++++-------
 perf-changelog.yaml                                 |  3 ++-
 3 files changed, 13 insertions(+), 15 deletions(-)

diff --git a/benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh b/benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh
index 8136e516a..99875d1d7 100755
--- a/benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh
@@ -2,9 +2,9 @@
 set -euo pipefail
 
 # DeepSeek-V4-Pro FP8 AgentX replay on one 8xMI325X node. The checkpoint is
-# dequantized to FP8 because gfx942 has no native MXFP4 support. Both supported
-# one-node layouts retain MTP: TP8 for the latency frontier, and DP8+EP8 for the
-# high-throughput frontier.
+# dequantized to FP8 because gfx942 has no native MXFP4 support. The published
+# path is pure TP8: current stable and nightly vLLM builds both return invalid
+# empty-content responses with expert parallelism on this model/SKU.
 
 source "$(dirname "$0")/../../benchmark_lib.sh"
 
@@ -74,10 +74,8 @@ if [[ "$DP_ATTENTION" == "true" ]]; then
 fi
 
 MAX_NUM_SEQS=$((2 * CONC))
-# The first AgentX-fast pass used K=3 and exposed empty-content responses on
-# the exploratory gfx942 deepseek_v4_fp8 path.  The existing MI300X/MI325X
-# fixed-sequence recipes use K=2; screen that supported shape independently
-# before changing the stable vLLM image.
+# The existing MI300X/MI325X fixed-sequence recipes use K=2. Keep that MTP
+# depth and its measured golden acceptance length for every AgentX point.
 NUM_SPEC_TOKENS=2
 SYNTHETIC_ACCEPT_LEN=2.27
 if [[ "${EVAL_ONLY:-false}" == "true" ]]; then
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 73f97c05f..1fb8106ee 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -1474,10 +1474,11 @@ glm5.2-fp8-mi325x-sglang-agentic-mtp:
 # DeepSeek-V4-Pro FP8 full-context AgentX frontier on one MI325X node. The
 # dequantized checkpoint is roughly 1.05 TB, so TP8 fits comfortably in the
 # node's 2 TB of HBM while TP4 does not have enough per-rank model headroom.
-# Stable v0.26.0 produced empty-content responses on every expert-parallel
-# topology at both K=3 and K=2.  This third AgentX-fast pass keeps K=2 and moves
-# only the image to the current immutable ROCm nightly, with one TP control and
-# the three minimal TEP/DPA reproductions.
+# Stable v0.26.0 and the current immutable nightly both produce empty-content
+# responses on expert-parallel topologies. Upstream also marks MI325X
+# unsupported for this model, while the proposed cross-node TP16 FP8 support is
+# still unmerged. Screen the fit-safe, semantically clean TP8 band densely; do
+# not publish an EP or patched multi-node workaround.
 dsv4-fp8-mi325x-vllm-agentic-mtp:
   image: vllm/vllm-openai-rocm:nightly-f8d03e77416bf90c49acbe50e233275722f02c4b
   model: deepseek-ai/DeepSeek-V4-Pro
@@ -1490,9 +1491,7 @@ dsv4-fp8-mi325x-vllm-agentic-mtp:
     agentic-coding:
     - dram-utilization: 0.90
       search-space:
-      - { tp: 8, ep: 1, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [8] }
-      - { tp: 8, ep: 8, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [20, 32] }
-      - { tp: 8, ep: 8, dp-attn: true, kv-offloading: none, spec-decoding: mtp, conc-list: [24], router: { name: vllm-router, version: "0.1.14" } }
+      - { tp: 8, ep: 1, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 4, 6, 8] }
 
 minimaxm3-fp8-mi325x-vllm-agentic:
   image: vllm/vllm-openai-rocm:nightly-04c2a8deac44fdb1ca3e2b5ec3e6bf16f3f6a914
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index d9fb88d5e..0b41cf4df 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5720,6 +5720,7 @@
     - "The first broad agentx-fast pass covered dense TP8, TP8+EP8 and DP8+EP8 bands through concurrency 80. It found clean-server duration-coverage failures on slow TP points, repeated empty-content semantic responses on TP8+EP8 and DP8+EP8, and a DP c40 warmup request that did not drain within 1800 seconds."
     - "Run a second seven-point agentx-fast diagnostic on stable vLLM v0.26.0 at TP8 c8/c12/c16, TP8+EP8 c20/c32, and DP8+EP8 c24/c48. This changes only MTP depth from K=3 to K=2 (and golden AL 2.49 to 2.27) so speculative depth is isolated before considering a nightly image or publishing the final dense frontier."
     - "The K=2 stable pass reproduced the defect: TEP c20 had 32 empty-content errors in 39 warmup records, TEP c32 had 45 in 57, and DPA c24 aborted after five profiling semantic errors; TP c8 remained semantically clean but produced only 2,993 input and 10.5 output tokens/s. This isolates stable-v0.26 deepseek_v4_fp8 expert parallelism rather than MTP depth."
-    - "Run a four-point third agentx-fast diagnostic on the current immutable ROCm nightly nightly-f8d03e77416bf90c49acbe50e233275722f02c4b: TP c8 control, TEP c20/c32, and DPA c24. Keep K=2/AL2.27 unchanged so the image is the only serving-stack variable."
+    - "The immutable ROCm nightly nightly-f8d03e77416bf90c49acbe50e233275722f02c4b did not make expert parallelism publishable: TEP c20 produced 33 empty-content errors in 41 warmup records and 7 in its first 13 profile records, TEP c32 produced 51 in 63 warmup records, and DPA c24 produced 3 in 49 warmup records plus 1 in its first 11 profile records. Pure TP remained semantically clean."
+    - "Upstream's DeepSeek-V4-Pro recipe explicitly marks MI325X unsupported, and cross-node TP16 FP8 support remains an open, unmerged vLLM PR. Do not pin an unreleased patch: use the current immutable nightly and densely screen the fit-safe single-node pure-TP8 band at concurrency 1, 2, 4, 6, and 8 before the reusable full-duration sweep."
     - "Use FP8 KV, 90% GPU-memory utilization, block size 256, an 8192-token scheduler budget and FULL_DECODE_ONLY graph capture. Enable prompt-token details for request-level cache accounting, and require a nonempty vllm: server-metrics export from AIPerf."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2541

From 2a115199f60a96f0f5596b9e62b547230779d12f Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Sun, 9 Aug 2026 22:11:57 -0500
Subject: [PATCH 07/18] perf(mi325x): refine DSV4 MTP frontier band

---
 benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh | 5 +++++
 configs/amd-master.yaml                               | 2 +-
 perf-changelog.yaml                                   | 2 ++
 3 files changed, 8 insertions(+), 1 deletion(-)

diff --git a/benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh b/benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh
index 99875d1d7..ba87fcaeb 100755
--- a/benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh
@@ -148,5 +148,10 @@ if [[ "${EVAL_ONLY:-false}" == "true" ]]; then
     run_eval --port "$PORT"
 else
     build_replay_cmd "$RESULT_DIR"
+    # Full-context AgentX responses can remain healthy for several minutes
+    # after the admission window closes. Let already-admitted requests drain
+    # so their observed TTFT/ITL enters the strict coverage calculation. This
+    # does not extend admissions, change the workload, or lower the 98% gate.
+    REPLAY_CMD+=" --benchmark-grace-period 1800"
     run_agentic_replay_and_write_outputs "$RESULT_DIR"
 fi
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 1fb8106ee..585c2a1c2 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -1491,7 +1491,7 @@ dsv4-fp8-mi325x-vllm-agentic-mtp:
     agentic-coding:
     - dram-utilization: 0.90
       search-space:
-      - { tp: 8, ep: 1, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 4, 6, 8] }
+      - { tp: 8, ep: 1, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [3, 4, 5] }
 
 minimaxm3-fp8-mi325x-vllm-agentic:
   image: vllm/vllm-openai-rocm:nightly-04c2a8deac44fdb1ca3e2b5ec3e6bf16f3f6a914
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 0b41cf4df..d4821f924 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5722,5 +5722,7 @@
     - "The K=2 stable pass reproduced the defect: TEP c20 had 32 empty-content errors in 39 warmup records, TEP c32 had 45 in 57, and DPA c24 aborted after five profiling semantic errors; TP c8 remained semantically clean but produced only 2,993 input and 10.5 output tokens/s. This isolates stable-v0.26 deepseek_v4_fp8 expert parallelism rather than MTP depth."
     - "The immutable ROCm nightly nightly-f8d03e77416bf90c49acbe50e233275722f02c4b did not make expert parallelism publishable: TEP c20 produced 33 empty-content errors in 41 warmup records and 7 in its first 13 profile records, TEP c32 produced 51 in 63 warmup records, and DPA c24 produced 3 in 49 warmup records plus 1 in its first 11 profile records. Pure TP remained semantically clean."
     - "Upstream's DeepSeek-V4-Pro recipe explicitly marks MI325X unsupported, and cross-node TP16 FP8 support remains an open, unmerged vLLM PR. Do not pin an unreleased patch: use the current immutable nightly and densely screen the fit-safe single-node pure-TP8 band at concurrency 1, 2, 4, 6, and 8 before the reusable full-duration sweep."
+    - "The pure-TP fast pass places the useful band at c1-c4: c1/c2 complete green at 13.49k/31.19 and 7.14k/44.53 input/output tok/s; c4 records 12.02k/67.97 with 0.67s p50 TTFT. c6/c8 collapse to 3.25-4.04k input and 9.70-12.60 output tok/s with 57-85s p50 TTFT, so drop that post-cliff region and probe the missing c3/c5 neighbors around c4."
+    - "Give already-admitted full-context responses up to 1800 seconds to drain after the measurement window, matching the merged MI325X AgentX convention. The c4 server remained healthy at roughly 84 output tok/s through the entire default 30-second grace, but two long responses were cancelled and completed-record TTFT coverage landed at 97.9% versus the unchanged 98% gate. Bounded drain preserves the admission window and strict coverage threshold."
     - "Use FP8 KV, 90% GPU-memory utilization, block size 256, an 8192-token scheduler budget and FULL_DECODE_ONLY graph capture. Enable prompt-token details for request-level cache accounting, and require a nonempty vllm: server-metrics export from AIPerf."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2541

From 7cb3e6e1fceddc1233dccee61e823fa5ab8a1d85 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Sun, 9 Aug 2026 22:51:11 -0500
Subject: [PATCH 08/18] perf(mi325x): publish DSV4 MTP frontier grid

---
 configs/amd-master.yaml | 6 +++---
 perf-changelog.yaml     | 3 ++-
 2 files changed, 5 insertions(+), 4 deletions(-)

diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 585c2a1c2..46c190866 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -1477,8 +1477,8 @@ glm5.2-fp8-mi325x-sglang-agentic-mtp:
 # Stable v0.26.0 and the current immutable nightly both produce empty-content
 # responses on expert-parallel topologies. Upstream also marks MI325X
 # unsupported for this model, while the proposed cross-node TP16 FP8 support is
-# still unmerged. Screen the fit-safe, semantically clean TP8 band densely; do
-# not publish an EP or patched multi-node workaround.
+# still unmerged. Publish the measured, fit-safe, semantically clean TP8 band;
+# do not publish an EP or patched multi-node workaround.
 dsv4-fp8-mi325x-vllm-agentic-mtp:
   image: vllm/vllm-openai-rocm:nightly-f8d03e77416bf90c49acbe50e233275722f02c4b
   model: deepseek-ai/DeepSeek-V4-Pro
@@ -1491,7 +1491,7 @@ dsv4-fp8-mi325x-vllm-agentic-mtp:
     agentic-coding:
     - dram-utilization: 0.90
       search-space:
-      - { tp: 8, ep: 1, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [3, 4, 5] }
+      - { tp: 8, ep: 1, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 3, 4] }
 
 minimaxm3-fp8-mi325x-vllm-agentic:
   image: vllm/vllm-openai-rocm:nightly-04c2a8deac44fdb1ca3e2b5ec3e6bf16f3f6a914
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index d4821f924..2215622dc 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5715,7 +5715,7 @@
   scenario-type:
     - agentic-coding
   description:
-    - "Add the first DeepSeek-V4-Pro FP8 vLLM AgentX submission on MI325X, with MTP enabled for every published point; the current diagnostic pass uses two draft tokens and the committed K=2 golden synthetic acceptance length of 2.27, while eval-only runs retain real target verification."
+    - "Add the first DeepSeek-V4-Pro FP8 vLLM AgentX submission on MI325X, with MTP enabled for every published point; use two draft tokens and the committed K=2 golden synthetic acceptance length of 2.27, while eval-only runs retain real target verification."
     - "Use one 8xMI325X node: the roughly 1.05 TB FP8-dequantized checkpoint fits in the node's 2 TB HBM at TP8, so multi-node execution would add communication without solving a capacity constraint; TP4 is excluded because its per-rank model footprint is not fit-safe."
     - "The first broad agentx-fast pass covered dense TP8, TP8+EP8 and DP8+EP8 bands through concurrency 80. It found clean-server duration-coverage failures on slow TP points, repeated empty-content semantic responses on TP8+EP8 and DP8+EP8, and a DP c40 warmup request that did not drain within 1800 seconds."
     - "Run a second seven-point agentx-fast diagnostic on stable vLLM v0.26.0 at TP8 c8/c12/c16, TP8+EP8 c20/c32, and DP8+EP8 c24/c48. This changes only MTP depth from K=3 to K=2 (and golden AL 2.49 to 2.27) so speculative depth is isolated before considering a nightly image or publishing the final dense frontier."
@@ -5724,5 +5724,6 @@
     - "Upstream's DeepSeek-V4-Pro recipe explicitly marks MI325X unsupported, and cross-node TP16 FP8 support remains an open, unmerged vLLM PR. Do not pin an unreleased patch: use the current immutable nightly and densely screen the fit-safe single-node pure-TP8 band at concurrency 1, 2, 4, 6, and 8 before the reusable full-duration sweep."
     - "The pure-TP fast pass places the useful band at c1-c4: c1/c2 complete green at 13.49k/31.19 and 7.14k/44.53 input/output tok/s; c4 records 12.02k/67.97 with 0.67s p50 TTFT. c6/c8 collapse to 3.25-4.04k input and 9.70-12.60 output tok/s with 57-85s p50 TTFT, so drop that post-cliff region and probe the missing c3/c5 neighbors around c4."
     - "Give already-admitted full-context responses up to 1800 seconds to drain after the measurement window, matching the merged MI325X AgentX convention. The c4 server remained healthy at roughly 84 output tok/s through the entire default 30-second grace, but two long responses were cancelled and completed-record TTFT coverage landed at 97.9% versus the unchanged 98% gate. Bounded drain preserves the admission window and strict coverage threshold."
+    - "The refined c3/c4/c5 fast pass completed all admitted requests with zero request errors and nonempty vllm: metrics exports. c3 reached 6.43k input / 64.35 output tok/s, c4 reached 11.15k / 69.11 with 0.68s p50 TTFT, and c5 fell to 6.30k / 33.04 with 7.8s p50 TTFT. Publish c1-c4 as the final full-duration frontier-and-guard grid and drop the clearly dominated c5-c8 region."
     - "Use FP8 KV, 90% GPU-memory utilization, block size 256, an 8192-token scheduler budget and FULL_DECODE_ONLY graph capture. Enable prompt-token details for request-level cache accounting, and require a nonempty vllm: server-metrics export from AIPerf."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2541

From 138f6c160ad594f865a5a99e99c55b4be7d215c9 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 10 Aug 2026 00:56:53 -0500
Subject: [PATCH 09/18] perf(mi325x): drop underfilled DSV4 c4 point

---
 configs/amd-master.yaml | 2 +-
 perf-changelog.yaml     | 3 ++-
 2 files changed, 3 insertions(+), 2 deletions(-)

diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 46c190866..41a1b89cf 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -1491,7 +1491,7 @@ dsv4-fp8-mi325x-vllm-agentic-mtp:
     agentic-coding:
     - dram-utilization: 0.90
       search-space:
-      - { tp: 8, ep: 1, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 3, 4] }
+      - { tp: 8, ep: 1, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 3] }
 
 minimaxm3-fp8-mi325x-vllm-agentic:
   image: vllm/vllm-openai-rocm:nightly-04c2a8deac44fdb1ca3e2b5ec3e6bf16f3f6a914
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 2215622dc..3091e7930 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5724,6 +5724,7 @@
     - "Upstream's DeepSeek-V4-Pro recipe explicitly marks MI325X unsupported, and cross-node TP16 FP8 support remains an open, unmerged vLLM PR. Do not pin an unreleased patch: use the current immutable nightly and densely screen the fit-safe single-node pure-TP8 band at concurrency 1, 2, 4, 6, and 8 before the reusable full-duration sweep."
     - "The pure-TP fast pass places the useful band at c1-c4: c1/c2 complete green at 13.49k/31.19 and 7.14k/44.53 input/output tok/s; c4 records 12.02k/67.97 with 0.67s p50 TTFT. c6/c8 collapse to 3.25-4.04k input and 9.70-12.60 output tok/s with 57-85s p50 TTFT, so drop that post-cliff region and probe the missing c3/c5 neighbors around c4."
     - "Give already-admitted full-context responses up to 1800 seconds to drain after the measurement window, matching the merged MI325X AgentX convention. The c4 server remained healthy at roughly 84 output tok/s through the entire default 30-second grace, but two long responses were cancelled and completed-record TTFT coverage landed at 97.9% versus the unchanged 98% gate. Bounded drain preserves the admission window and strict coverage threshold."
-    - "The refined c3/c4/c5 fast pass completed all admitted requests with zero request errors and nonempty vllm: metrics exports. c3 reached 6.43k input / 64.35 output tok/s, c4 reached 11.15k / 69.11 with 0.68s p50 TTFT, and c5 fell to 6.30k / 33.04 with 7.8s p50 TTFT. Publish c1-c4 as the final full-duration frontier-and-guard grid and drop the clearly dominated c5-c8 region."
+    - "The refined c3/c4/c5 fast pass completed all admitted requests with zero request errors and nonempty vllm: metrics exports. c3 reached 6.43k input / 64.35 output tok/s, c4 reached 11.15k / 69.11 with 0.68s p50 TTFT, and c5 fell to 6.30k / 33.04 with 7.8s p50 TTFT. Select c1-c4 as the reusable full-duration candidate grid and drop the clearly dominated c5-c8 region."
+    - "The reusable full-duration sweep invalidated c4 as a publishable point despite a healthy server and 295/295 successful requests: the finite five-trajectory corpus had only three distinct root session trees active at the 98% cutoff for configured concurrency four, so TTFT coverage correctly landed at 97.9504% while ITL remained 100%. Keep the strict gate, do not waive or shorten the workload, and publish the fully covered c1-c3 grid instead."
     - "Use FP8 KV, 90% GPU-memory utilization, block size 256, an 8192-token scheduler budget and FULL_DECODE_ONLY graph capture. Enable prompt-token details for request-level cache accounting, and require a nonempty vllm: server-metrics export from AIPerf."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2541

From acf58ba1c3272af9cf1da9347299e32b90c1c75a Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 10 Aug 2026 03:08:40 -0500
Subject: [PATCH 10/18] perf(mi325x): probe larger DSV4 prefill budget

---
 benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh | 2 +-
 perf-changelog.yaml                                   | 3 ++-
 2 files changed, 3 insertions(+), 2 deletions(-)

diff --git a/benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh b/benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh
index ba87fcaeb..fc7a047d8 100755
--- a/benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh
@@ -109,7 +109,7 @@ VLLM_CMD=(
     "${EP_ARGS[@]}"
     --gpu-memory-utilization 0.9
     --block-size 256
-    --max-num-batched-tokens 8192
+    --max-num-batched-tokens 16384
     --max-num-seqs "$MAX_NUM_SEQS"
     --compilation-config '{"mode":3,"cudagraph_mode":"FULL_DECODE_ONLY"}'
     --speculative-config "$SPEC_CONFIG"
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 3091e7930..98e1eb24d 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5726,5 +5726,6 @@
     - "Give already-admitted full-context responses up to 1800 seconds to drain after the measurement window, matching the merged MI325X AgentX convention. The c4 server remained healthy at roughly 84 output tok/s through the entire default 30-second grace, but two long responses were cancelled and completed-record TTFT coverage landed at 97.9% versus the unchanged 98% gate. Bounded drain preserves the admission window and strict coverage threshold."
     - "The refined c3/c4/c5 fast pass completed all admitted requests with zero request errors and nonempty vllm: metrics exports. c3 reached 6.43k input / 64.35 output tok/s, c4 reached 11.15k / 69.11 with 0.68s p50 TTFT, and c5 fell to 6.30k / 33.04 with 7.8s p50 TTFT. Select c1-c4 as the reusable full-duration candidate grid and drop the clearly dominated c5-c8 region."
     - "The reusable full-duration sweep invalidated c4 as a publishable point despite a healthy server and 295/295 successful requests: the finite five-trajectory corpus had only three distinct root session trees active at the 98% cutoff for configured concurrency four, so TTFT coverage correctly landed at 97.9504% while ITL remained 100%. Keep the strict gate, do not waive or shorten the workload, and publish the fully covered c1-c3 grid instead."
-    - "Use FP8 KV, 90% GPU-memory utilization, block size 256, an 8192-token scheduler budget and FULL_DECODE_ONLY graph capture. Enable prompt-token details for request-level cache accounting, and require a nonempty vllm: server-metrics export from AIPerf."
+    - "After the c1-c3 exact-head sweep and c3 eval completed fully green, run one isolated agentx-fast scheduler-budget comparison: increase only --max-num-batched-tokens from the official 8192-token ROCm baseline to 16384 while retaining the exact same c1/c2/c3 topology, image, MTP depth, acceptance length, KV dtype, graph mode, and workload. The completed 8192-token run remains the rollback point."
+    - "Use FP8 KV, 90% GPU-memory utilization, block size 256, the currently tested 16384-token scheduler budget and FULL_DECODE_ONLY graph capture. Enable prompt-token details for request-level cache accounting, and require a nonempty vllm: server-metrics export from AIPerf."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2541

From 11f035c500356497494ca3bbf538d5fb1df4cca7 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 10 Aug 2026 05:09:29 -0500
Subject: [PATCH 11/18] perf(mi325x): probe larger DSV4 sequence capacity

---
 benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh | 8 ++++++--
 configs/amd-master.yaml                               | 3 ++-
 perf-changelog.yaml                                   | 4 +++-
 3 files changed, 11 insertions(+), 4 deletions(-)

diff --git a/benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh b/benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh
index fc7a047d8..ba89850a6 100755
--- a/benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh
@@ -73,7 +73,11 @@ if [[ "$DP_ATTENTION" == "true" ]]; then
     agentic_pip_install --quiet 'vllm-router==0.1.14'
 fi
 
-MAX_NUM_SEQS=$((2 * CONC))
+# The preceding fast run showed the 16K prefill budget is exactly neutral at
+# c1. Restore the official 8K budget and isolate scheduler slot capacity at c3,
+# where the baseline reached its 2*CONC running-request ceiling with queued
+# branch work despite low KV utilization.
+MAX_NUM_SEQS=$((4 * CONC))
 # The existing MI300X/MI325X fixed-sequence recipes use K=2. Keep that MTP
 # depth and its measured golden acceptance length for every AgentX point.
 NUM_SPEC_TOKENS=2
@@ -109,7 +113,7 @@ VLLM_CMD=(
     "${EP_ARGS[@]}"
     --gpu-memory-utilization 0.9
     --block-size 256
-    --max-num-batched-tokens 16384
+    --max-num-batched-tokens 8192
     --max-num-seqs "$MAX_NUM_SEQS"
     --compilation-config '{"mode":3,"cudagraph_mode":"FULL_DECODE_ONLY"}'
     --speculative-config "$SPEC_CONFIG"
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 41a1b89cf..db0354f8e 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -1491,7 +1491,8 @@ dsv4-fp8-mi325x-vllm-agentic-mtp:
     agentic-coding:
     - dram-utilization: 0.90
       search-space:
-      - { tp: 8, ep: 1, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 3] }
+      # Isolated agentx-fast scheduler-slot probe at the retained upper point.
+      - { tp: 8, ep: 1, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [3] }
 
 minimaxm3-fp8-mi325x-vllm-agentic:
   image: vllm/vllm-openai-rocm:nightly-04c2a8deac44fdb1ca3e2b5ec3e6bf16f3f6a914
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 98e1eb24d..0b78b30e2 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5727,5 +5727,7 @@
     - "The refined c3/c4/c5 fast pass completed all admitted requests with zero request errors and nonempty vllm: metrics exports. c3 reached 6.43k input / 64.35 output tok/s, c4 reached 11.15k / 69.11 with 0.68s p50 TTFT, and c5 fell to 6.30k / 33.04 with 7.8s p50 TTFT. Select c1-c4 as the reusable full-duration candidate grid and drop the clearly dominated c5-c8 region."
     - "The reusable full-duration sweep invalidated c4 as a publishable point despite a healthy server and 295/295 successful requests: the finite five-trajectory corpus had only three distinct root session trees active at the 98% cutoff for configured concurrency four, so TTFT coverage correctly landed at 97.9504% while ITL remained 100%. Keep the strict gate, do not waive or shorten the workload, and publish the fully covered c1-c3 grid instead."
     - "After the c1-c3 exact-head sweep and c3 eval completed fully green, run one isolated agentx-fast scheduler-budget comparison: increase only --max-num-batched-tokens from the official 8192-token ROCm baseline to 16384 while retaining the exact same c1/c2/c3 topology, image, MTP depth, acceptance length, KV dtype, graph mode, and workload. The completed 8192-token run remains the rollback point."
-    - "Use FP8 KV, 90% GPU-memory utilization, block size 256, the currently tested 16384-token scheduler budget and FULL_DECODE_ONLY graph capture. Enable prompt-token details for request-level cache accounting, and require a nonempty vllm: server-metrics export from AIPerf."
+    - "The isolated 16384-token c1 fast probe completed 48/48 requests with zero errors, TTFT 99.7%, ITL 100%, and 13,488.30 input / 31.19 output tok/s: exactly neutral versus the retained 8192-token result. Its c3 eval also completed 1,319/1,319 HTTP 200 responses. Reject the larger budget and restore 8192."
+    - "Next isolate only scheduler slot capacity at c3: raise --max-num-seqs from 2*CONC to 4*CONC while keeping the restored 8192-token budget and every topology, model, MTP, KV, graph, memory, and workload setting fixed. Baseline logs reached the 2*CONC running-request ceiling with queued branch work while KV use stayed below 18%, making c3 the informative upper retained point."
+    - "Use FP8 KV, 90% GPU-memory utilization, block size 256, the official 8192-token scheduler budget, the currently tested 4*CONC sequence cap, and FULL_DECODE_ONLY graph capture. Enable prompt-token details for request-level cache accounting, and require a nonempty vllm: server-metrics export from AIPerf."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2541

From 74f8b9f9ae0dcb66c6508992c3b55693ea7af0a9 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 10 Aug 2026 07:11:37 -0500
Subject: [PATCH 12/18] perf(mi325x): probe piecewise DSV4 graphs

---
 benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh | 10 ++++------
 perf-changelog.yaml                                   |  3 ++-
 2 files changed, 6 insertions(+), 7 deletions(-)

diff --git a/benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh b/benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh
index ba89850a6..1ecdd5511 100755
--- a/benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh
@@ -73,11 +73,9 @@ if [[ "$DP_ATTENTION" == "true" ]]; then
     agentic_pip_install --quiet 'vllm-router==0.1.14'
 fi
 
-# The preceding fast run showed the 16K prefill budget is exactly neutral at
-# c1. Restore the official 8K budget and isolate scheduler slot capacity at c3,
-# where the baseline reached its 2*CONC running-request ceiling with queued
-# branch work despite low KV utilization.
-MAX_NUM_SEQS=$((4 * CONC))
+# The 16K prefill budget and 4*CONC sequence-cap probes were both neutral.
+# Restore the official 8K/2*CONC baseline before isolating graph capture mode.
+MAX_NUM_SEQS=$((2 * CONC))
 # The existing MI300X/MI325X fixed-sequence recipes use K=2. Keep that MTP
 # depth and its measured golden acceptance length for every AgentX point.
 NUM_SPEC_TOKENS=2
@@ -115,7 +113,7 @@ VLLM_CMD=(
     --block-size 256
     --max-num-batched-tokens 8192
     --max-num-seqs "$MAX_NUM_SEQS"
-    --compilation-config '{"mode":3,"cudagraph_mode":"FULL_DECODE_ONLY"}'
+    --compilation-config '{"mode":3,"cudagraph_mode":"FULL_AND_PIECEWISE"}'
     --speculative-config "$SPEC_CONFIG"
     --tokenizer-mode deepseek_v4
     --tool-call-parser deepseek_v4
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 0b78b30e2..5bd65bd30 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5729,5 +5729,6 @@
     - "After the c1-c3 exact-head sweep and c3 eval completed fully green, run one isolated agentx-fast scheduler-budget comparison: increase only --max-num-batched-tokens from the official 8192-token ROCm baseline to 16384 while retaining the exact same c1/c2/c3 topology, image, MTP depth, acceptance length, KV dtype, graph mode, and workload. The completed 8192-token run remains the rollback point."
     - "The isolated 16384-token c1 fast probe completed 48/48 requests with zero errors, TTFT 99.7%, ITL 100%, and 13,488.30 input / 31.19 output tok/s: exactly neutral versus the retained 8192-token result. Its c3 eval also completed 1,319/1,319 HTTP 200 responses. Reject the larger budget and restore 8192."
     - "Next isolate only scheduler slot capacity at c3: raise --max-num-seqs from 2*CONC to 4*CONC while keeping the restored 8192-token budget and every topology, model, MTP, KV, graph, memory, and workload setting fixed. Baseline logs reached the 2*CONC running-request ceiling with queued branch work while KV use stayed below 18%, making c3 the informative upper retained point."
-    - "Use FP8 KV, 90% GPU-memory utilization, block size 256, the official 8192-token scheduler budget, the currently tested 4*CONC sequence cap, and FULL_DECODE_ONLY graph capture. Enable prompt-token details for request-level cache accounting, and require a nonempty vllm: server-metrics export from AIPerf."
+    - "Reject the 4*CONC sequence cap after its c3 fast run completed fully green at 6,440.50 input / 64.42 output tok/s over 1,392.11 seconds, versus 6,433.20 / 64.35 over 1,393.73 seconds for the matched 2*CONC control: both token rates changed by only 0.11%. Its c3 eval also completed 1,319/1,319 HTTP 200 responses and collect-evals passed. Restore 2*CONC before any further experiment."
+    - "Next isolate only graph capture mode at c3: switch FULL_DECODE_ONLY to FULL_AND_PIECEWISE while keeping K=2/golden AL 2.27, the official 8192-token scheduler budget, max-num-seqs=2*CONC, FP8 KV, 90% GPU-memory utilization, block size 256, topology, image, and AgentX trace unchanged. Continue requiring nonempty vllm: server metrics from the explicit AIPerf endpoint."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2541

From 7d028235bb2316fda6139fba2b62677a4fdac16b Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 10 Aug 2026 07:49:56 -0500
Subject: [PATCH 13/18] perf(mi325x): probe INT4 quick reduce

---
 benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh | 8 +++++---
 configs/amd-master.yaml                               | 2 +-
 perf-changelog.yaml                                   | 2 ++
 3 files changed, 8 insertions(+), 4 deletions(-)

diff --git a/benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh b/benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh
index 1ecdd5511..2dae4c517 100755
--- a/benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh
@@ -42,6 +42,7 @@ export AIPERF_REQUIRED_SERVER_METRIC_PREFIX="vllm:"
 export VLLM_ENGINE_READY_TIMEOUT_S=10800
 export VLLM_ROCM_USE_AITER=1
 export VLLM_ROCM_USE_AITER_MOE=1
+export VLLM_ROCM_QUICK_REDUCE_QUANTIZATION=INT4
 export PYTHONNOUSERSITE=1
 
 SERVER_LOG="$RESULT_DIR/server.log"
@@ -73,8 +74,9 @@ if [[ "$DP_ATTENTION" == "true" ]]; then
     agentic_pip_install --quiet 'vllm-router==0.1.14'
 fi
 
-# The 16K prefill budget and 4*CONC sequence-cap probes were both neutral.
-# Restore the official 8K/2*CONC baseline before isolating graph capture mode.
+# The 16K prefill budget and 4*CONC sequence-cap probes were neutral, while
+# piecewise graphs regressed. Restore the official 8K/2*CONC/FULL_DECODE_ONLY
+# baseline before isolating INT4 Quick Reduce.
 MAX_NUM_SEQS=$((2 * CONC))
 # The existing MI300X/MI325X fixed-sequence recipes use K=2. Keep that MTP
 # depth and its measured golden acceptance length for every AgentX point.
@@ -113,7 +115,7 @@ VLLM_CMD=(
     --block-size 256
     --max-num-batched-tokens 8192
     --max-num-seqs "$MAX_NUM_SEQS"
-    --compilation-config '{"mode":3,"cudagraph_mode":"FULL_AND_PIECEWISE"}'
+    --compilation-config '{"mode":3,"cudagraph_mode":"FULL_DECODE_ONLY"}'
     --speculative-config "$SPEC_CONFIG"
     --tokenizer-mode deepseek_v4
     --tool-call-parser deepseek_v4
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index db0354f8e..c1ac1339b 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -1491,7 +1491,7 @@ dsv4-fp8-mi325x-vllm-agentic-mtp:
     agentic-coding:
     - dram-utilization: 0.90
       search-space:
-      # Isolated agentx-fast scheduler-slot probe at the retained upper point.
+      # Isolated agentx-fast INT4 Quick Reduce probe at the retained upper point.
       - { tp: 8, ep: 1, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [3] }
 
 minimaxm3-fp8-mi325x-vllm-agentic:
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 5bd65bd30..18a0497ef 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5731,4 +5731,6 @@
     - "Next isolate only scheduler slot capacity at c3: raise --max-num-seqs from 2*CONC to 4*CONC while keeping the restored 8192-token budget and every topology, model, MTP, KV, graph, memory, and workload setting fixed. Baseline logs reached the 2*CONC running-request ceiling with queued branch work while KV use stayed below 18%, making c3 the informative upper retained point."
     - "Reject the 4*CONC sequence cap after its c3 fast run completed fully green at 6,440.50 input / 64.42 output tok/s over 1,392.11 seconds, versus 6,433.20 / 64.35 over 1,393.73 seconds for the matched 2*CONC control: both token rates changed by only 0.11%. Its c3 eval also completed 1,319/1,319 HTTP 200 responses and collect-evals passed. Restore 2*CONC before any further experiment."
     - "Next isolate only graph capture mode at c3: switch FULL_DECODE_ONLY to FULL_AND_PIECEWISE while keeping K=2/golden AL 2.27, the official 8192-token scheduler budget, max-num-seqs=2*CONC, FP8 KV, 90% GPU-memory utilization, block size 256, topology, image, and AgentX trace unchanged. Continue requiring nonempty vllm: server metrics from the explicit AIPerf endpoint."
+    - "Reject FULL_AND_PIECEWISE after the exact c3 fast comparison processed the same 76 requests with zero errors but reached 6,358.88 input / 63.60 output tok/s over 1,410.00 seconds, versus 6,433.20 / 64.35 over 1,393.73 seconds for the duration-matched FULL_DECODE_ONLY control: both rates regressed about 1.16%. TTFT coverage was 99.8%, ITL coverage was 100%, and server metrics exports were nonempty, so the longer completion time is a measured graph-mode regression rather than infrastructure noise."
+    - "Next isolate only INT4 Quick Reduce at the retained c3 baseline: enable VLLM_ROCM_QUICK_REDUCE_QUANTIZATION=INT4 while restoring FULL_DECODE_ONLY and keeping the immutable nightly, TP8, K=2/golden AL 2.27, the 8192-token scheduler budget, max-num-seqs=2*CONC, FP8 KV, 90% GPU-memory utilization, block size 256, and the exact AgentX trace unchanged. The paired eval guards accuracy."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2541

From a2cd2ebe45b907bd32eafdf07c7444263f953702 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 10 Aug 2026 12:50:09 -0500
Subject: [PATCH 14/18] perf(mi325x): expand DSV4 collection grid

---
 .../agentic/dsv4_fp8_mi325x_mtp.sh            | 125 +++++++++++++++++-
 configs/amd-master.yaml                       |  15 ++-
 perf-changelog.yaml                           |   1 +
 3 files changed, 127 insertions(+), 14 deletions(-)

diff --git a/benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh b/benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh
index 2dae4c517..91de14369 100755
--- a/benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh
@@ -8,12 +8,7 @@ set -euo pipefail
 
 source "$(dirname "$0")/../../benchmark_lib.sh"
 
-check_env_vars MODEL IMAGE TP CONC KV_OFFLOADING RESULT_DIR DURATION EP_SIZE DP_ATTENTION
-
-if [[ "$KV_OFFLOADING" != "none" ]]; then
-    echo "ERROR: DeepSeek-V4 MTP on MI325X currently supports GPU-resident KV only" >&2
-    exit 1
-fi
+check_env_vars MODEL IMAGE TP CONC KV_OFFLOADING TOTAL_CPU_DRAM_GB RESULT_DIR DURATION EP_SIZE DP_ATTENTION
 
 if [[ -n "${SLURM_JOB_ID:-}" ]]; then
     echo "JOB $SLURM_JOB_ID running on ${SLURMD_NODENAME:-unknown}"
@@ -47,8 +42,123 @@ export PYTHONNOUSERSITE=1
 
 SERVER_LOG="$RESULT_DIR/server.log"
 ROUTER_LOG="$RESULT_DIR/router.log"
+MOONCAKE_MASTER_LOG="$RESULT_DIR/mooncake_master.log"
 mkdir -p "$RESULT_DIR"
 
+SERVER_PID=""
+ROUTER_PID=""
+MOONCAKE_MASTER_PID=""
+
+# Mooncake does not publish a ROCm wheel. Reuse the established MI300X/MI325X
+# recipe convention: build the pinned release once per immutable ROCm/Python
+# combination, cache the staged install on shared storage, and verify that the
+# loaded transfer engine is actually linked against HIP.
+install_mooncake_rocm() {
+    local mooncake_tag="v0.3.11.post1"
+    local mooncake_src="/tmp/Mooncake-$mooncake_tag"
+    local mooncake_stage="/tmp/mooncake-stage-$mooncake_tag"
+    local build_jobs cache_root cache_key cache_archive cache_tmp
+    local engine_path os_version python_abi rocm_version
+
+    build_jobs=$(nproc)
+    if ((build_jobs > 32)); then
+        build_jobs=32
+    fi
+
+    os_version=$(. /etc/os-release && printf '%s-%s' "$ID" "$VERSION_ID")
+    python_abi=$(python3 -c 'import sys; print(f"cp{sys.version_info.major}{sys.version_info.minor}")')
+    rocm_version=$(sed -n '1p' /opt/rocm/.info/version 2>/dev/null || true)
+    if [[ -z "$rocm_version" ]]; then
+        rocm_version=$(hipconfig --version)
+    fi
+    rocm_version=${rocm_version//[^[:alnum:]._-]/_}
+    cache_root="${HF_HUB_CACHE:?HF_HUB_CACHE must be set}/inferencex/mooncake"
+    cache_key="${mooncake_tag}-${os_version}-${python_abi}-${rocm_version}-$(uname -m)-hip"
+    cache_archive="$cache_root/$cache_key.tar.gz"
+    mkdir -p "$cache_root"
+
+    apt-get update
+    DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends \
+        build-essential cmake git libasio-dev libboost-dev libcurl4-openssl-dev \
+        libgflags-dev libgoogle-glog-dev libibverbs-dev libjsoncpp-dev \
+        libnuma-dev libpython3-dev libssl-dev libunwind-dev liburing-dev \
+        libxxhash-dev libyaml-cpp-dev libzstd-dev ninja-build pybind11-dev
+
+    exec 9>"$cache_archive.lock"
+    flock -w 1800 9
+    if [[ -f "$cache_archive" ]] && ! tar -tzf "$cache_archive" >/dev/null 2>&1; then
+        rm -f "$cache_archive"
+    fi
+    if [[ ! -f "$cache_archive" ]]; then
+        echo "Building HIP Mooncake cache artifact: $cache_archive"
+        rm -rf "$mooncake_src" "$mooncake_stage"
+        git clone --depth 1 --branch "$mooncake_tag" --recurse-submodules \
+            --shallow-submodules https://github.com/kvcache-ai/Mooncake.git "$mooncake_src"
+        cmake -S "$mooncake_src/extern/yalantinglibs" \
+            -B "$mooncake_src/extern/yalantinglibs/build" \
+            -DBUILD_EXAMPLES=OFF -DBUILD_BENCHMARK=OFF -DBUILD_UNIT_TESTS=OFF
+        cmake --build "$mooncake_src/extern/yalantinglibs/build" -j "$build_jobs"
+        cmake --install "$mooncake_src/extern/yalantinglibs/build"
+        cmake -S "$mooncake_src" -B "$mooncake_src/build" -G Ninja \
+            -DCMAKE_BUILD_TYPE=Release -DUSE_CUDA=OFF -DUSE_HIP=ON \
+            -DWITH_EP=OFF -DWITH_STORE=ON -DWITH_STORE_RUST=OFF \
+            -DWITH_RUST_EXAMPLE=OFF -DBUILD_EXAMPLES=OFF -DBUILD_UNIT_TESTS=OFF
+        cmake --build "$mooncake_src/build" -j "$build_jobs"
+        mkdir -p "$mooncake_stage"
+        DESTDIR="$mooncake_stage" cmake --install "$mooncake_src/build"
+        cache_tmp=$(mktemp "$cache_root/$cache_key.tmp.XXXXXX")
+        tar -C "$mooncake_stage" -czf "$cache_tmp" .
+        mv -f "$cache_tmp" "$cache_archive"
+    else
+        echo "Using HIP Mooncake cache artifact: $cache_archive"
+    fi
+    tar -C / -xzf "$cache_archive"
+    engine_path=$(python3 -c 'import mooncake.engine; print(mooncake.engine.__file__)')
+    ldd "$engine_path" | grep -q 'libamdhip64.so'
+    exec 9>&-
+}
+
+OFFLOAD_ARGS=()
+if agentic_kv_offload_enabled; then
+    require_agentic_kv_offload_backend mooncake
+    # TOTAL_CPU_DRAM_GB is the generator-capped aggregate node budget.
+    # Embedded Mooncake contributes one segment per rank, so divide it here.
+    PER_RANK_GB=$((TOTAL_CPU_DRAM_GB / TP))
+    if ! python3 -c "from mooncake.store import MooncakeDistributedStore" >/dev/null 2>&1; then
+        install_mooncake_rocm
+    fi
+    python3 -c "from mooncake.store import MooncakeDistributedStore" >/dev/null
+    MOONCAKE_MASTER_PORT=$((PORT + 12000))
+    MOONCAKE_CONFIG_PATH="$RESULT_DIR/mooncake_config.json"
+    cat > "$MOONCAKE_CONFIG_PATH" <<EOF
+{
+  "mode": "embedded",
+  "metadata_server": "P2PHANDSHAKE",
+  "master_server_address": "127.0.0.1:$MOONCAKE_MASTER_PORT",
+  "global_segment_size": "${PER_RANK_GB}GB",
+  "local_buffer_size": "4GB",
+  "protocol": "tcp",
+  "device_name": "",
+  "enable_offload": false
+}
+EOF
+    export MOONCAKE_CONFIG_PATH PYTHONHASHSEED=0 MC_SLICE_SIZE=1048576 MC_WORKERS_PER_CTX=4
+    export MC_TCP_ENABLE_CONNECTION_POOL=1
+    mooncake_master --port "$MOONCAKE_MASTER_PORT" \
+        --default_kv_lease_ttl=120s \
+        --eviction_high_watermark_ratio=0.80 \
+        --eviction_ratio=0.10 > "$MOONCAKE_MASTER_LOG" 2>&1 &
+    MOONCAKE_MASTER_PID=$!
+    sleep 2
+    kill -0 "$MOONCAKE_MASTER_PID"
+    OFFLOAD_ARGS=(
+        --kv-transfer-config
+        '{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_both","kv_connector_extra_config":{"load_async":true}}'
+    )
+else
+    require_agentic_kv_offload_none
+fi
+
 PARALLEL_ARGS=(--tensor-parallel-size "$TP" --data-parallel-size 1)
 if [[ "$DP_ATTENTION" == "true" ]]; then
     PARALLEL_ARGS=(--tensor-parallel-size 1 --data-parallel-size "$TP")
@@ -61,7 +171,6 @@ fi
 
 USE_VLLM_ROUTER=false
 VLLM_BACKEND_PORT="$PORT"
-ROUTER_PID=""
 if [[ "$DP_ATTENTION" == "true" ]]; then
     if (( EP_SIZE != TP )); then
         echo "ERROR: MI325X DP-attention requires EP_SIZE == TP so FP8 experts remain sharded" >&2
@@ -94,6 +203,7 @@ cleanup() {
     set +e
     stop_background_process_tree "$ROUTER_PID" "vLLM router"
     stop_background_process_tree "${SERVER_PID:-}" "vLLM server" 60
+    stop_background_process_tree "$MOONCAKE_MASTER_PID" "Mooncake master"
     exit "$rc"
 }
 trap cleanup EXIT
@@ -124,6 +234,7 @@ VLLM_CMD=(
     --enable-prefix-caching
     --enable-prompt-tokens-details
     --no-disable-hybrid-kv-cache-manager
+    "${OFFLOAD_ARGS[@]}"
 )
 
 printf '%q ' "${VLLM_CMD[@]}" > "$RESULT_DIR/vllm_command.txt"
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index c1ac1339b..15522c056 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -1474,11 +1474,11 @@ glm5.2-fp8-mi325x-sglang-agentic-mtp:
 # DeepSeek-V4-Pro FP8 full-context AgentX frontier on one MI325X node. The
 # dequantized checkpoint is roughly 1.05 TB, so TP8 fits comfortably in the
 # node's 2 TB of HBM while TP4 does not have enough per-rank model headroom.
-# Stable v0.26.0 and the current immutable nightly both produce empty-content
-# responses on expert-parallel topologies. Upstream also marks MI325X
-# unsupported for this model, while the proposed cross-node TP16 FP8 support is
-# still unmerged. Publish the measured, fit-safe, semantically clean TP8 band;
-# do not publish an EP or patched multi-node workaround.
+# Stable v0.26.0 and the current immutable nightly both produced empty-content
+# responses on earlier expert-parallel probes. Retain c4-c6 TEP8 explicitly in
+# this collection sweep so that behavior is represented beside the pure-TP
+# points. Mooncake uses the standard generated aggregate host-DRAM cap and the
+# pinned ROCm build convention already exercised by MI325X AgentX recipes.
 dsv4-fp8-mi325x-vllm-agentic-mtp:
   image: vllm/vllm-openai-rocm:nightly-f8d03e77416bf90c49acbe50e233275722f02c4b
   model: deepseek-ai/DeepSeek-V4-Pro
@@ -1491,8 +1491,9 @@ dsv4-fp8-mi325x-vllm-agentic-mtp:
     agentic-coding:
     - dram-utilization: 0.90
       search-space:
-      # Isolated agentx-fast INT4 Quick Reduce probe at the retained upper point.
-      - { tp: 8, ep: 1, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [3] }
+      - { tp: 8, ep: 1, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 3, 4, 5, 6] }
+      - { tp: 8, ep: 1, dp-attn: false, kv-offloading: dram, kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }, spec-decoding: mtp, conc-list: [4, 5, 6] }
+      - { tp: 8, ep: 8, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [4, 5, 6] }
 
 minimaxm3-fp8-mi325x-vllm-agentic:
   image: vllm/vllm-openai-rocm:nightly-04c2a8deac44fdb1ca3e2b5ec3e6bf16f3f6a914
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 18a0497ef..52cc12d7c 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5733,4 +5733,5 @@
     - "Next isolate only graph capture mode at c3: switch FULL_DECODE_ONLY to FULL_AND_PIECEWISE while keeping K=2/golden AL 2.27, the official 8192-token scheduler budget, max-num-seqs=2*CONC, FP8 KV, 90% GPU-memory utilization, block size 256, topology, image, and AgentX trace unchanged. Continue requiring nonempty vllm: server metrics from the explicit AIPerf endpoint."
     - "Reject FULL_AND_PIECEWISE after the exact c3 fast comparison processed the same 76 requests with zero errors but reached 6,358.88 input / 63.60 output tok/s over 1,410.00 seconds, versus 6,433.20 / 64.35 over 1,393.73 seconds for the duration-matched FULL_DECODE_ONLY control: both rates regressed about 1.16%. TTFT coverage was 99.8%, ITL coverage was 100%, and server metrics exports were nonempty, so the longer completion time is a measured graph-mode regression rather than infrastructure noise."
     - "Next isolate only INT4 Quick Reduce at the retained c3 baseline: enable VLLM_ROCM_QUICK_REDUCE_QUANTIZATION=INT4 while restoring FULL_DECODE_ONLY and keeping the immutable nightly, TP8, K=2/golden AL 2.27, the 8192-token scheduler budget, max-num-seqs=2*CONC, FP8 KV, 90% GPU-memory utilization, block size 256, and the exact AgentX trace unchanged. The paired eval guards accuracy."
+    - "Expand the collection grid beyond the measured Pareto candidates: retain pure TP8 c1-c6, add Mooncake DRAM-offload TP8 c4/c5/c6 using the generated aggregate host-memory cap and pinned ROCm Mooncake 0.3.11.post1 build, and add TEP8 c4/c5/c6 even if those points remain dominated or reproduce the earlier empty-content behavior. Functionality is the acceptance criterion for inclusion, not Pareto optimality."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2541

From 2fa98a2231d34424d9ec2bd1fddd613d6c0474aa Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 10 Aug 2026 13:02:28 -0500
Subject: [PATCH 15/18] fix(mi325x): bound Mooncake transfer logging

---
 benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh b/benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh
index 91de14369..80ae8be31 100755
--- a/benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh
@@ -142,6 +142,9 @@ if agentic_kv_offload_enabled; then
   "enable_offload": false
 }
 EOF
+    # Mooncake v0.3.11.post1 emits its transfer polling loop at VLOG(1).
+    # Keep normal INFO diagnostics while suppressing that unbounded hot-loop output.
+    export GLOG_v=0
     export MOONCAKE_CONFIG_PATH PYTHONHASHSEED=0 MC_SLICE_SIZE=1048576 MC_WORKERS_PER_CTX=4
     export MC_TCP_ENABLE_CONNECTION_POOL=1
     mooncake_master --port "$MOONCAKE_MASTER_PORT" \

From 221772aff5ab1e6c233a50f15135e103c7d71a3d Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 10 Aug 2026 16:17:26 -0500
Subject: [PATCH 16/18] fix MI325X TEP sampling and Mooncake transport

---
 .../agentic/dsv4_fp8_mi325x_mtp.sh            | 26 ++++++++++++++++---
 perf-changelog.yaml                           |  2 ++
 2 files changed, 25 insertions(+), 3 deletions(-)

diff --git a/benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh b/benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh
index 80ae8be31..820bd9bc2 100755
--- a/benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh
@@ -145,10 +145,22 @@ EOF
     # Mooncake v0.3.11.post1 emits its transfer polling loop at VLOG(1).
     # Keep normal INFO diagnostics while suppressing that unbounded hot-loop output.
     export GLOG_v=0
-    export MOONCAKE_CONFIG_PATH PYTHONHASHSEED=0 MC_SLICE_SIZE=1048576 MC_WORKERS_PER_CTX=4
-    export MC_TCP_ENABLE_CONNECTION_POOL=1
+    export MOONCAKE_CONFIG_PATH PYTHONHASHSEED=0 MC_SLICE_SIZE=1048576
+    # Match the proven MI355X TCP-store convention. Mooncake 0.3.11's pooled
+    # TCP path retained thousands of simultaneous sockets on this workload and
+    # timed out 60-second transfers; the unpooled path plus eight workers keeps
+    # transfer concurrency bounded by the store workers instead.
+    export MC_ENABLE_DEST_DEVICE_AFFINITY=1
+    export MC_WORKERS_PER_CTX=8
+    unset MC_TCP_ENABLE_CONNECTION_POOL
+    MOONCAKE_KV_LEASE_TTL=120s
+    # A full-context transfer can starve a rank's client heartbeat beyond the
+    # master's 10-second default even while that rank remains healthy. Keep the
+    # client registered long enough for the bounded transfer timeout to resolve.
+    MOONCAKE_CLIENT_TTL=120
     mooncake_master --port "$MOONCAKE_MASTER_PORT" \
-        --default_kv_lease_ttl=120s \
+        --default_kv_lease_ttl="$MOONCAKE_KV_LEASE_TTL" \
+        --client_ttl="$MOONCAKE_CLIENT_TTL" \
         --eviction_high_watermark_ratio=0.80 \
         --eviction_ratio=0.10 > "$MOONCAKE_MASTER_LOG" 2>&1 &
     MOONCAKE_MASTER_PID=$!
@@ -186,6 +198,14 @@ if [[ "$DP_ATTENTION" == "true" ]]; then
     agentic_pip_install --quiet 'vllm-router==0.1.14'
 fi
 
+if (( EP_SIZE > 1 )) && [[ "$DP_ATTENTION" != "true" ]]; then
+    # TEP's correlated low-concurrency trajectories contain a few deterministic
+    # metadata-only turns. Do not let three early turns abort a one-hour run
+    # before the sample is representative; the unchanged strict 10% post-run
+    # validator remains authoritative for the completed result.
+    AIPERF_LIVE_FAILED_REQUEST_THRESHOLD=0.50
+fi
+
 # The 16K prefill budget and 4*CONC sequence-cap probes were neutral, while
 # piecewise graphs regressed. Restore the official 8K/2*CONC/FULL_DECODE_ONLY
 # baseline before isolating INT4 Quick Reduce.
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 52cc12d7c..f49d11e27 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5734,4 +5734,6 @@
     - "Reject FULL_AND_PIECEWISE after the exact c3 fast comparison processed the same 76 requests with zero errors but reached 6,358.88 input / 63.60 output tok/s over 1,410.00 seconds, versus 6,433.20 / 64.35 over 1,393.73 seconds for the duration-matched FULL_DECODE_ONLY control: both rates regressed about 1.16%. TTFT coverage was 99.8%, ITL coverage was 100%, and server metrics exports were nonempty, so the longer completion time is a measured graph-mode regression rather than infrastructure noise."
     - "Next isolate only INT4 Quick Reduce at the retained c3 baseline: enable VLLM_ROCM_QUICK_REDUCE_QUANTIZATION=INT4 while restoring FULL_DECODE_ONLY and keeping the immutable nightly, TP8, K=2/golden AL 2.27, the 8192-token scheduler budget, max-num-seqs=2*CONC, FP8 KV, 90% GPU-memory utilization, block size 256, and the exact AgentX trace unchanged. The paired eval guards accuracy."
     - "Expand the collection grid beyond the measured Pareto candidates: retain pure TP8 c1-c6, add Mooncake DRAM-offload TP8 c4/c5/c6 using the generated aggregate host-memory cap and pinned ROCm Mooncake 0.3.11.post1 build, and add TEP8 c4/c5/c6 even if those points remain dominated or reproduce the earlier empty-content behavior. Functionality is the acceptance criterion for inclusion, not Pareto optimality."
+    - "Keep TEP c4-c6 running long enough to collect a representative sample when a few early correlated metadata-only turns return no content: raise only the live fail-fast threshold to 50%, while retaining the unchanged strict 10% completed-run request-error validator."
+    - "Set Mooncake's independent client-heartbeat TTL to 120 seconds for the full-context offload points. The first c4-c6 run kept the 120-second object lease but used Mooncake's 10-second client default; a healthy rank was unmounted after missing that heartbeat window. A full-primer c4 isolation then kept all eight clients registered but exposed a second issue: the 0.3.11 TCP connection pool retained about 7,000 simultaneous sockets and 60-second batch transfers timed out. Match the proven MI355X convention by disabling that pool, enabling destination affinity, and using eight transfer workers."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2541

From c875b370e5fcf7109c60ee1e917a5281fa09e988 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 10 Aug 2026 23:28:56 -0500
Subject: [PATCH 17/18] refresh MI325X MTP collection on vLLM v0.27.0

---
 .../agentic/dsv4_fp8_mi325x_mtp.sh            | 37 +++++++++++++------
 configs/amd-master.yaml                       | 12 +++---
 perf-changelog.yaml                           |  1 +
 3 files changed, 32 insertions(+), 18 deletions(-)

diff --git a/benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh b/benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh
index 820bd9bc2..f2e1617b3 100755
--- a/benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv4_fp8_mi325x_mtp.sh
@@ -2,9 +2,9 @@
 set -euo pipefail
 
 # DeepSeek-V4-Pro FP8 AgentX replay on one 8xMI325X node. The checkpoint is
-# dequantized to FP8 because gfx942 has no native MXFP4 support. The published
-# path is pure TP8: current stable and nightly vLLM builds both return invalid
-# empty-content responses with expert parallelism on this model/SKU.
+# dequantized to FP8 because gfx942 has no native MXFP4 support. The collection
+# includes pure TP8, TP8/EP8, and Mooncake DRAM-offload points explicitly, even
+# where the latter two do not improve the Pareto frontier.
 
 source "$(dirname "$0")/../../benchmark_lib.sh"
 
@@ -31,6 +31,26 @@ resolve_trace_source
 install_agentic_deps
 agentic_pip_install --quiet Pillow fastapi uvicorn
 
+install_vllm_hybrid_kv_recovery() {
+    # vLLM PR #45497 fixes hybrid-KV connector load failures by recomputing the
+    # complete request when any KV group is unavailable. Keep v0.27.0's ROCm
+    # binaries and install only the pinned Python scheduler source revision.
+    local recovery_commit="cb1917efe34f16423de2ceb7f210fd015d53459a"
+    local recovery_src="/tmp/vllm-hybrid-kv-recovery"
+    rm -rf "$recovery_src"
+    git clone --depth 1 --branch test/v027-hybrid-kv-recovery \
+        https://github.com/cquil11/vllm.git "$recovery_src"
+    if [[ "$(git -C "$recovery_src" rev-parse HEAD)" != "$recovery_commit" ]]; then
+        echo "Unexpected vLLM hybrid-KV recovery revision" >&2
+        exit 1
+    fi
+    VLLM_USE_PRECOMPILED=1 \
+    VLLM_ROCM_WHEEL_INDEX=https://wheels.vllm.ai/rocm/0.27.0/rocm723 \
+        uv pip install --system --no-deps --no-build-isolation --editable \
+        "$recovery_src"
+    python3 -c 'import vllm; print("vLLM source:", vllm.__file__)'
+}
+
 export AIPERF_HTTP_TCP_USER_TIMEOUT=900000
 export AIPERF_SERVER_METRICS_URLS="http://localhost:${PORT}/metrics"
 export AIPERF_REQUIRED_SERVER_METRIC_PREFIX="vllm:"
@@ -121,6 +141,7 @@ install_mooncake_rocm() {
 OFFLOAD_ARGS=()
 if agentic_kv_offload_enabled; then
     require_agentic_kv_offload_backend mooncake
+    install_vllm_hybrid_kv_recovery
     # TOTAL_CPU_DRAM_GB is the generator-capped aggregate node budget.
     # Embedded Mooncake contributes one segment per rank, so divide it here.
     PER_RANK_GB=$((TOTAL_CPU_DRAM_GB / TP))
@@ -168,7 +189,7 @@ EOF
     kill -0 "$MOONCAKE_MASTER_PID"
     OFFLOAD_ARGS=(
         --kv-transfer-config
-        '{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_both","kv_connector_extra_config":{"load_async":true}}'
+        '{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_both","kv_load_failure_policy":"recompute","kv_connector_extra_config":{"load_async":true}}'
     )
 else
     require_agentic_kv_offload_none
@@ -198,14 +219,6 @@ if [[ "$DP_ATTENTION" == "true" ]]; then
     agentic_pip_install --quiet 'vllm-router==0.1.14'
 fi
 
-if (( EP_SIZE > 1 )) && [[ "$DP_ATTENTION" != "true" ]]; then
-    # TEP's correlated low-concurrency trajectories contain a few deterministic
-    # metadata-only turns. Do not let three early turns abort a one-hour run
-    # before the sample is representative; the unchanged strict 10% post-run
-    # validator remains authoritative for the completed result.
-    AIPERF_LIVE_FAILED_REQUEST_THRESHOLD=0.50
-fi
-
 # The 16K prefill budget and 4*CONC sequence-cap probes were neutral, while
 # piecewise graphs regressed. Restore the official 8K/2*CONC/FULL_DECODE_ONLY
 # baseline before isolating INT4 Quick Reduce.
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 4645dbacd..176bede87 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -1474,13 +1474,13 @@ glm5.2-fp8-mi325x-sglang-agentic-mtp:
 # DeepSeek-V4-Pro FP8 full-context AgentX frontier on one MI325X node. The
 # dequantized checkpoint is roughly 1.05 TB, so TP8 fits comfortably in the
 # node's 2 TB of HBM while TP4 does not have enough per-rank model headroom.
-# Stable v0.26.0 and the current immutable nightly both produced empty-content
-# responses on earlier expert-parallel probes. Retain c4-c6 TEP8 explicitly in
-# this collection sweep so that behavior is represented beside the pure-TP
-# points. Mooncake uses the standard generated aggregate host-DRAM cap and the
-# pinned ROCm build convention already exercised by MI325X AgentX recipes.
+# Stable v0.26.0 and v0.27.0 both produced empty-content responses on earlier
+# expert-parallel probes. Retain c4-c6 TEP8 explicitly in this collection sweep
+# so that behavior is represented beside the pure-TP points. Mooncake uses the
+# generated aggregate host-DRAM cap, the pinned ROCm build convention, and the
+# upstream hybrid-KV load-recovery scheduler fix validated against v0.27.0.
 dsv4-fp8-mi325x-vllm-agentic-mtp:
-  image: vllm/vllm-openai-rocm:nightly-f8d03e77416bf90c49acbe50e233275722f02c4b
+  image: vllm/vllm-openai-rocm:v0.27.0
   model: deepseek-ai/DeepSeek-V4-Pro
   model-prefix: dsv4
   runner: cluster:mi325x-amds
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index ac6aaacda..fc7878059 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5752,6 +5752,7 @@
     - "Expand the collection grid beyond the measured Pareto candidates: retain pure TP8 c1-c6, add Mooncake DRAM-offload TP8 c4/c5/c6 using the generated aggregate host-memory cap and pinned ROCm Mooncake 0.3.11.post1 build, and add TEP8 c4/c5/c6 even if those points remain dominated or reproduce the earlier empty-content behavior. Functionality is the acceptance criterion for inclusion, not Pareto optimality."
     - "Keep TEP c4-c6 running long enough to collect a representative sample when a few early correlated metadata-only turns return no content: raise only the live fail-fast threshold to 50%, while retaining the unchanged strict 10% completed-run request-error validator."
     - "Set Mooncake's independent client-heartbeat TTL to 120 seconds for the full-context offload points. The first c4-c6 run kept the 120-second object lease but used Mooncake's 10-second client default; a healthy rank was unmounted after missing that heartbeat window. A full-primer c4 isolation then kept all eight clients registered but exposed a second issue: the 0.3.11 TCP connection pool retained about 7,000 simultaneous sockets and 60-second batch transfers timed out. Match the proven MI355X convention by disabling that pool, enabling destination affinity, and using eight transfer workers."
+    - "Use the stable vLLM v0.27.0 ROCm image. For Mooncake points only, pin the Python-only scheduler recovery from upstream vLLM PR #45497 and set kv_load_failure_policy=recompute; the isolated c6 gate completed 65/65 warmup and 31/31 measured requests with zero errors while safely recomputing 22 partial hybrid-KV misses."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2541
 
 - config-keys:

From 29af29b5d662e1b870ba1f72ab0fcc005c284261 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 10 Aug 2026 23:57:05 -0500
Subject: [PATCH 18/18] omit invalid MI325X TEP points

---
 configs/amd-master.yaml | 11 +++++------
 perf-changelog.yaml     |  4 ++--
 2 files changed, 7 insertions(+), 8 deletions(-)

diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 176bede87..28d818e0d 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -1474,11 +1474,11 @@ glm5.2-fp8-mi325x-sglang-agentic-mtp:
 # DeepSeek-V4-Pro FP8 full-context AgentX frontier on one MI325X node. The
 # dequantized checkpoint is roughly 1.05 TB, so TP8 fits comfortably in the
 # node's 2 TB of HBM while TP4 does not have enough per-rank model headroom.
-# Stable v0.26.0 and v0.27.0 both produced empty-content responses on earlier
-# expert-parallel probes. Retain c4-c6 TEP8 explicitly in this collection sweep
-# so that behavior is represented beside the pure-TP points. Mooncake uses the
-# generated aggregate host-DRAM cap, the pinned ROCm build convention, and the
-# upstream hybrid-KV load-recovery scheduler fix validated against v0.27.0.
+# Stable v0.26.0 and v0.27.0 both produced invalid empty-content responses with
+# expert parallelism, including strict K=1 and K=2 gates. Publish only the
+# functional TP8 and Mooncake points. Mooncake uses the generated aggregate
+# host-DRAM cap, the pinned ROCm build convention, and the upstream hybrid-KV
+# load-recovery scheduler fix validated against v0.27.0.
 dsv4-fp8-mi325x-vllm-agentic-mtp:
   image: vllm/vllm-openai-rocm:v0.27.0
   model: deepseek-ai/DeepSeek-V4-Pro
@@ -1493,7 +1493,6 @@ dsv4-fp8-mi325x-vllm-agentic-mtp:
       search-space:
       - { tp: 8, ep: 1, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 3, 4, 5, 6] }
       - { tp: 8, ep: 1, dp-attn: false, kv-offloading: dram, kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }, spec-decoding: mtp, conc-list: [4, 5, 6] }
-      - { tp: 8, ep: 8, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [4, 5, 6] }
 
 minimaxm3-fp8-mi325x-vllm-agentic:
   image: vllm/vllm-openai-rocm:nightly-04c2a8deac44fdb1ca3e2b5ec3e6bf16f3f6a914
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index fc7878059..053ba4aba 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5749,8 +5749,8 @@
     - "Next isolate only graph capture mode at c3: switch FULL_DECODE_ONLY to FULL_AND_PIECEWISE while keeping K=2/golden AL 2.27, the official 8192-token scheduler budget, max-num-seqs=2*CONC, FP8 KV, 90% GPU-memory utilization, block size 256, topology, image, and AgentX trace unchanged. Continue requiring nonempty vllm: server metrics from the explicit AIPerf endpoint."
     - "Reject FULL_AND_PIECEWISE after the exact c3 fast comparison processed the same 76 requests with zero errors but reached 6,358.88 input / 63.60 output tok/s over 1,410.00 seconds, versus 6,433.20 / 64.35 over 1,393.73 seconds for the duration-matched FULL_DECODE_ONLY control: both rates regressed about 1.16%. TTFT coverage was 99.8%, ITL coverage was 100%, and server metrics exports were nonempty, so the longer completion time is a measured graph-mode regression rather than infrastructure noise."
     - "Next isolate only INT4 Quick Reduce at the retained c3 baseline: enable VLLM_ROCM_QUICK_REDUCE_QUANTIZATION=INT4 while restoring FULL_DECODE_ONLY and keeping the immutable nightly, TP8, K=2/golden AL 2.27, the 8192-token scheduler budget, max-num-seqs=2*CONC, FP8 KV, 90% GPU-memory utilization, block size 256, and the exact AgentX trace unchanged. The paired eval guards accuracy."
-    - "Expand the collection grid beyond the measured Pareto candidates: retain pure TP8 c1-c6, add Mooncake DRAM-offload TP8 c4/c5/c6 using the generated aggregate host-memory cap and pinned ROCm Mooncake 0.3.11.post1 build, and add TEP8 c4/c5/c6 even if those points remain dominated or reproduce the earlier empty-content behavior. Functionality is the acceptance criterion for inclusion, not Pareto optimality."
-    - "Keep TEP c4-c6 running long enough to collect a representative sample when a few early correlated metadata-only turns return no content: raise only the live fail-fast threshold to 50%, while retaining the unchanged strict 10% completed-run request-error validator."
+    - "Expand the collection grid beyond the measured Pareto candidates: retain pure TP8 c1-c6 and add Mooncake DRAM-offload TP8 c4/c5/c6 using the generated aggregate host-memory cap and pinned ROCm Mooncake 0.3.11.post1 build."
+    - "Omit TEP8 c4-c6 after strict v0.27.0 gates reproduced invalid empty-content responses with both real MTP K=2 (3/10 profiling requests failed) and K=1 (2/18 failed). Keep the standard 10% request-error gate with no topology-specific exception."
     - "Set Mooncake's independent client-heartbeat TTL to 120 seconds for the full-context offload points. The first c4-c6 run kept the 120-second object lease but used Mooncake's 10-second client default; a healthy rank was unmounted after missing that heartbeat window. A full-primer c4 isolation then kept all eight clients registered but exposed a second issue: the 0.3.11 TCP connection pool retained about 7,000 simultaneous sockets and 60-second batch transfers timed out. Match the proven MI355X convention by disabling that pool, enabling destination affinity, and using eight transfer workers."
     - "Use the stable vLLM v0.27.0 ROCm image. For Mooncake points only, pin the Python-only scheduler recovery from upstream vLLM PR #45497 and set kv_load_failure_policy=recompute; the isolated c6 gate completed 65/65 warmup and 31/31 measured requests with zero errors while safely recomputing 22 partial hybrid-KV misses."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2541