From 8c634513d2f14cd654a5c387d28686a5ec1fa124 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): tune DSv4 MTP frontier on MI355X

---
 .../agentic/dsv4_fp4_mi355x_vllm_mtp.sh         |  8 ++++----
 configs/amd-master.yaml                         | 17 +++++++++--------
 2 files changed, 13 insertions(+), 12 deletions(-)

diff --git a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_vllm_mtp.sh b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_vllm_mtp.sh
index 1b9311eda..40f72f077 100644
--- a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_vllm_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_vllm_mtp.sh
@@ -393,8 +393,6 @@ export VLLM_ROCM_USE_AITER=1
 #export VLLM_ROCM_QUICK_REDUCE_QUANTIZATION=INT4
 export VLLM_ROCM_USE_AITER_MOE=1
 
-sleep 180
-
 { set +x; } 2>/dev/null
 VLLM_CMD=(
     vllm serve "$MODEL_PATH" --served-model-name "$MODEL"
@@ -406,9 +404,11 @@ VLLM_CMD=(
     --kv-cache-dtype fp8
     "${PARALLEL_ARGS[@]}"
     "${EP_ARGS[@]}"
-    --gpu-memory-utilization 0.8
+    --gpu-memory-utilization 0.9
+    --block-size 256
+    --max-num-batched-tokens 8192
     --moe-backend aiter
-    --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 c6580af43..735f9a4ac 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -1227,13 +1227,13 @@ qwen3.5-fp8-mi355x-sglang-agentic-hicache:
       - { tp: 8, ep: 1, kv-offloading: none, conc-list: [1, 2, 4, 8, 16, 32] }
       - { tp: 8, ep: 1, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [16, 32, 48, 64] }
 
-# MTP twin of dsv4-fp4-mi355x-vllm-agentic. The topologies, KV-offload backends,
-# and concurrency points are identical; speculative decoding is enabled on the
-# current immutable ROCm nightly because the base config's retired nightly tag
-# is no longer available from Docker Hub. Throughput uses the three-token golden
-# synthetic acceptance length, while eval-only runs retain real verification.
+# DeepSeek-V4-Pro FP4 MTP refresh for MI355X. Pure TP points densely sample the
+# latency/throughput knee; TP8+EP8 tests expert-sharded MoE without changing the
+# attention topology; DP8+EP8 is the intended high-throughput layout and avoids
+# the old c64 row's full-model replication on every GPU. GPU-resident KV remains
+# the only supported MTP path until vLLM's multi-KV-group recovery fix lands.
 dsv4-fp4-mi355x-vllm-agentic-mtp:
-  image: vllm/vllm-openai-rocm:nightly-821717118fc26667dd474b9b0ab81d29259dfc5c
+  image: vllm/vllm-openai-rocm:v0.26.0
   model: deepseek-ai/DeepSeek-V4-Pro
   model-prefix: dsv4
   runner: cluster:mi355x-amds
@@ -1244,8 +1244,9 @@ dsv4-fp4-mi355x-vllm-agentic-mtp:
     agentic-coding:
     - dram-utilization: 0.60
       search-space:
-      - { tp: 8, ep: 1, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 4, 8, 16, 32, 40, 48] }
-      - { tp: 8, ep: 1, dp-attn: true, kv-offloading: none, spec-decoding: mtp, conc-list: [64], router: { name: vllm-router, version: "0.1.14" } }
+      - { tp: 8, ep: 1, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 4, 8, 12, 16, 20, 24, 28, 32, 36, 40] }
+      - { tp: 8, ep: 8, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [8, 16, 24, 32, 40, 48, 56] }
+      - { tp: 8, ep: 8, dp-attn: true, kv-offloading: none, spec-decoding: mtp, conc-list: [32, 48, 64, 80, 96, 112, 128], router: { name: vllm-router, version: "0.1.14" } }
       # LMCache invalid-block recovery currently assumes one KV-cache group,
       # while MTP creates two. Restore these points after the upstream hybrid
       # KV recovery fix lands: https://github.com/vllm-project/vllm/pull/45497

From 2d5d0c2087d4f5d346778de9c6fb51799a0220e7 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 MI355X DSv4 tuning submission

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

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 471981f60..7cfabeb3e 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-fp4-mi355x-vllm-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Refresh the MI355X DeepSeek-V4-Pro FP4 vLLM AgentX MTP frontier after the prior eight-point run peaked near 66k total tok/s and its only DP-attention point used EP1/full-model replication."
+    - "Use the latest stable ROCm vLLM v0.26.0 release, raise GPU-memory utilization from 0.8 to 0.9, make the official 256-token block and 8192-token scheduler budget explicit, and use FULL_DECODE_ONLY graph capture. Baseline logs showed roughly 44 GiB/GPU of unused fit-safe KV headroom at 0.8."
+    - "Replace the isolated DP8/EP1 c64 point with dense TP8, TP8+EP8 and DP8+EP8 bands through concurrency 128. DP8+EP8 is the intended throughput topology: attention remains data parallel while the FP4 experts are sharded across all eight GPUs."
+    - "Keep MTP and its committed three-token golden synthetic acceptance length of 2.49 on every throughput point; eval-only runs retain real target verification, and AIPerf continues to require a nonempty vllm: server-metrics export."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2540

From 323fc1594980b0484b4565cd511683e0275d157f 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 MI355X

---
 benchmarks/single_node/agentic/dsv4_fp4_mi355x_vllm_mtp.sh | 1 +
 perf-changelog.yaml                                        | 2 +-
 2 files changed, 2 insertions(+), 1 deletion(-)

diff --git a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_vllm_mtp.sh b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_vllm_mtp.sh
index 40f72f077..58d04df7a 100644
--- a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_vllm_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_vllm_mtp.sh
@@ -415,6 +415,7 @@ VLLM_CMD=(
     --reasoning-parser deepseek_v4
     --enable-auto-tool-choice
     --enable-prefix-caching
+    --enable-prompt-tokens-details
     --no-disable-hybrid-kv-cache-manager
     --max-num-seqs "$MAX_NUM_SEQS"
     "${OFFLOAD_ARGS[@]}"
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 7cfabeb3e..a8026fe18 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5718,5 +5718,5 @@
     - "Refresh the MI355X DeepSeek-V4-Pro FP4 vLLM AgentX MTP frontier after the prior eight-point run peaked near 66k total tok/s and its only DP-attention point used EP1/full-model replication."
     - "Use the latest stable ROCm vLLM v0.26.0 release, raise GPU-memory utilization from 0.8 to 0.9, make the official 256-token block and 8192-token scheduler budget explicit, and use FULL_DECODE_ONLY graph capture. Baseline logs showed roughly 44 GiB/GPU of unused fit-safe KV headroom at 0.8."
     - "Replace the isolated DP8/EP1 c64 point with dense TP8, TP8+EP8 and DP8+EP8 bands through concurrency 128. DP8+EP8 is the intended throughput topology: attention remains data parallel while the FP4 experts are sharded across all eight GPUs."
-    - "Keep MTP and its committed three-token golden synthetic acceptance length of 2.49 on every throughput point; eval-only runs retain real target verification, and AIPerf continues to require a nonempty vllm: server-metrics export."
+    - "Keep MTP and its committed three-token golden synthetic acceptance length of 2.49 on every throughput point; eval-only runs retain real target verification. 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/2540

From 4d77e2875bfda653ffa061a1c4bd470a3c5500dc Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Sun, 9 Aug 2026 17:59:51 -0500
Subject: [PATCH 04/18] perf(mi355x): match DSV4 MTP search-space coverage

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

diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 735f9a4ac..7b91d893b 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -1227,11 +1227,11 @@ qwen3.5-fp8-mi355x-sglang-agentic-hicache:
       - { tp: 8, ep: 1, kv-offloading: none, conc-list: [1, 2, 4, 8, 16, 32] }
       - { tp: 8, ep: 1, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [16, 32, 48, 64] }
 
-# DeepSeek-V4-Pro FP4 MTP refresh for MI355X. Pure TP points densely sample the
-# latency/throughput knee; TP8+EP8 tests expert-sharded MoE without changing the
-# attention topology; DP8+EP8 is the intended high-throughput layout and avoids
-# the old c64 row's full-model replication on every GPU. GPU-resident KV remains
-# the only supported MTP path until vLLM's multi-KV-group recovery fix lands.
+# DeepSeek-V4-Pro FP4 MTP refresh for MI355X. The low-concurrency TP8 cells and
+# the DP8+EP8 band mirror the B200/B300 MTP grids, with extra intermediate cells
+# to resolve the MI355X knee. TP8+EP8 isolates expert-sharding without changing
+# the attention topology. GPU-resident KV remains the only supported MTP path
+# until vLLM's multi-KV-group recovery fix lands.
 dsv4-fp4-mi355x-vllm-agentic-mtp:
   image: vllm/vllm-openai-rocm:v0.26.0
   model: deepseek-ai/DeepSeek-V4-Pro
@@ -1244,9 +1244,9 @@ dsv4-fp4-mi355x-vllm-agentic-mtp:
     agentic-coding:
     - dram-utilization: 0.60
       search-space:
-      - { tp: 8, ep: 1, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 4, 8, 12, 16, 20, 24, 28, 32, 36, 40] }
+      - { tp: 8, ep: 1, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 4, 6, 8, 12, 16, 20, 24, 28, 32, 36, 40] }
       - { tp: 8, ep: 8, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [8, 16, 24, 32, 40, 48, 56] }
-      - { tp: 8, ep: 8, dp-attn: true, kv-offloading: none, spec-decoding: mtp, conc-list: [32, 48, 64, 80, 96, 112, 128], router: { name: vllm-router, version: "0.1.14" } }
+      - { tp: 8, ep: 8, dp-attn: true, kv-offloading: none, spec-decoding: mtp, conc-list: [32, 48, 64, 80, 96, 112, 128, 144, 160, 176, 192, 224], router: { name: vllm-router, version: "0.1.14" } }
       # LMCache invalid-block recovery currently assumes one KV-cache group,
       # while MTP creates two. Restore these points after the upstream hybrid
       # KV recovery fix lands: https://github.com/vllm-project/vllm/pull/45497
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index a8026fe18..0ade2b0a9 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5717,6 +5717,6 @@
   description:
     - "Refresh the MI355X DeepSeek-V4-Pro FP4 vLLM AgentX MTP frontier after the prior eight-point run peaked near 66k total tok/s and its only DP-attention point used EP1/full-model replication."
     - "Use the latest stable ROCm vLLM v0.26.0 release, raise GPU-memory utilization from 0.8 to 0.9, make the official 256-token block and 8192-token scheduler budget explicit, and use FULL_DECODE_ONLY graph capture. Baseline logs showed roughly 44 GiB/GPU of unused fit-safe KV headroom at 0.8."
-    - "Replace the isolated DP8/EP1 c64 point with dense TP8, TP8+EP8 and DP8+EP8 bands through concurrency 128. DP8+EP8 is the intended throughput topology: attention remains data parallel while the FP4 experts are sharded across all eight GPUs."
+    - "Replace the isolated DP8/EP1 c64 point with dense TP8, TP8+EP8 and DP8+EP8 bands. Mirror the B200/B300 MTP low-concurrency TP cells and extend DP8+EP8 through concurrency 224, with extra MI355X intermediates to resolve the local knee. DP8+EP8 is the intended throughput topology: attention remains data parallel while the FP4 experts are sharded across all eight GPUs."
     - "Keep MTP and its committed three-token golden synthetic acceptance length of 2.49 on every throughput point; eval-only runs retain real target verification. 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/2540

From 3c2131ee058f530d77e6051a73154d76bcd2677f Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Sun, 9 Aug 2026 18:17:32 -0500
Subject: [PATCH 05/18] perf(mi355x): add fit-validated TP4 MTP band

---
 configs/amd-master.yaml | 13 ++++++++-----
 perf-changelog.yaml     |  3 ++-
 2 files changed, 10 insertions(+), 6 deletions(-)

diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 7b91d893b..ac89eab03 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -1227,11 +1227,11 @@ qwen3.5-fp8-mi355x-sglang-agentic-hicache:
       - { tp: 8, ep: 1, kv-offloading: none, conc-list: [1, 2, 4, 8, 16, 32] }
       - { tp: 8, ep: 1, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [16, 32, 48, 64] }
 
-# DeepSeek-V4-Pro FP4 MTP refresh for MI355X. The low-concurrency TP8 cells and
-# the DP8+EP8 band mirror the B200/B300 MTP grids, with extra intermediate cells
-# to resolve the MI355X knee. TP8+EP8 isolates expert-sharding without changing
-# the attention topology. GPU-resident KV remains the only supported MTP path
-# until vLLM's multi-KV-group recovery fix lands.
+# DeepSeek-V4-Pro FP4 MTP refresh for MI355X. The TP8, fit-validated TP4, and
+# DP8+EP8 bands mirror the B200/B300 MTP grids, with extra intermediate cells to
+# resolve the MI355X knee. TP8+EP8 isolates expert-sharding without changing the
+# attention topology. GPU-resident KV remains the only supported MTP path until
+# vLLM's multi-KV-group recovery fix lands.
 dsv4-fp4-mi355x-vllm-agentic-mtp:
   image: vllm/vllm-openai-rocm:v0.26.0
   model: deepseek-ai/DeepSeek-V4-Pro
@@ -1245,8 +1245,11 @@ dsv4-fp4-mi355x-vllm-agentic-mtp:
     - dram-utilization: 0.60
       search-space:
       - { tp: 8, ep: 1, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 4, 6, 8, 12, 16, 20, 24, 28, 32, 36, 40] }
+      - { tp: 4, ep: 1, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 4, 6, 8, 12, 16, 20] }
       - { tp: 8, ep: 8, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [8, 16, 24, 32, 40, 48, 56] }
       - { tp: 8, ep: 8, dp-attn: true, kv-offloading: none, spec-decoding: mtp, conc-list: [32, 48, 64, 80, 96, 112, 128, 144, 160, 176, 192, 224], router: { name: vllm-router, version: "0.1.14" } }
+      # DP4+EP4 is not fit-safe at the required 1,048,576-token model length:
+      # a direct c56 probe left 12.09 GiB KV memory versus 14.88 GiB required.
       # LMCache invalid-block recovery currently assumes one KV-cache group,
       # while MTP creates two. Restore these points after the upstream hybrid
       # KV recovery fix lands: https://github.com/vllm-project/vllm/pull/45497
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 0ade2b0a9..972808ca4 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5717,6 +5717,7 @@
   description:
     - "Refresh the MI355X DeepSeek-V4-Pro FP4 vLLM AgentX MTP frontier after the prior eight-point run peaked near 66k total tok/s and its only DP-attention point used EP1/full-model replication."
     - "Use the latest stable ROCm vLLM v0.26.0 release, raise GPU-memory utilization from 0.8 to 0.9, make the official 256-token block and 8192-token scheduler budget explicit, and use FULL_DECODE_ONLY graph capture. Baseline logs showed roughly 44 GiB/GPU of unused fit-safe KV headroom at 0.8."
-    - "Replace the isolated DP8/EP1 c64 point with dense TP8, TP8+EP8 and DP8+EP8 bands. Mirror the B200/B300 MTP low-concurrency TP cells and extend DP8+EP8 through concurrency 224, with extra MI355X intermediates to resolve the local knee. DP8+EP8 is the intended throughput topology: attention remains data parallel while the FP4 experts are sharded across all eight GPUs."
+    - "Replace the isolated DP8/EP1 c64 point with 40 points spanning TP8, fit-validated TP4, TP8+EP8 and DP8+EP8. Mirror the B200/B300 MTP low-concurrency cells and extend DP8+EP8 through concurrency 224, with extra MI355X intermediates to resolve the local knee. DP8+EP8 is the intended throughput topology: attention remains data parallel while the FP4 experts are sharded across all eight GPUs."
+    - "Direct fit validation shows TP4 leaves 37.47 GiB/GPU for a 2.82M-token KV cache and serves normally. DP4+EP4 is excluded because it leaves only 12.09 GiB/GPU KV versus 14.88 GiB required for the committed 1,048,576-token model length; the workload is not weakened to force an invalid topology."
     - "Keep MTP and its committed three-token golden synthetic acceptance length of 2.49 on every throughput point; eval-only runs retain real target verification. 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/2540

From d813d647eeee56d97565459794ce7c1e4e9c5638 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Sun, 9 Aug 2026 20:53:14 -0500
Subject: [PATCH 06/18] perf(mi355x): refine fast frontier probes

---
 configs/amd-master.yaml | 17 ++++++++---------
 perf-changelog.yaml     |  1 +
 2 files changed, 9 insertions(+), 9 deletions(-)

diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index ac89eab03..53990f569 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -1227,11 +1227,11 @@ qwen3.5-fp8-mi355x-sglang-agentic-hicache:
       - { tp: 8, ep: 1, kv-offloading: none, conc-list: [1, 2, 4, 8, 16, 32] }
       - { tp: 8, ep: 1, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [16, 32, 48, 64] }
 
-# DeepSeek-V4-Pro FP4 MTP refresh for MI355X. The TP8, fit-validated TP4, and
-# DP8+EP8 bands mirror the B200/B300 MTP grids, with extra intermediate cells to
-# resolve the MI355X knee. TP8+EP8 isolates expert-sharding without changing the
-# attention topology. GPU-resident KV remains the only supported MTP path until
-# vLLM's multi-KV-group recovery fix lands.
+# DeepSeek-V4-Pro FP4 MTP refresh for MI355X. This second agentx-fast pass keeps
+# only the six unresolved cells from the broad B200/B300-shaped discovery grid:
+# low-concurrency TP guards and the high-end TP8+EP8 crossover. The first pass
+# established that DP8+EP8 is dominated and TP4 collapses above c12, so those
+# known-bad tails are intentionally excluded before the full-duration sweep.
 dsv4-fp4-mi355x-vllm-agentic-mtp:
   image: vllm/vllm-openai-rocm:v0.26.0
   model: deepseek-ai/DeepSeek-V4-Pro
@@ -1244,10 +1244,9 @@ dsv4-fp4-mi355x-vllm-agentic-mtp:
     agentic-coding:
     - dram-utilization: 0.60
       search-space:
-      - { tp: 8, ep: 1, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 4, 6, 8, 12, 16, 20, 24, 28, 32, 36, 40] }
-      - { tp: 4, ep: 1, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 4, 6, 8, 12, 16, 20] }
-      - { tp: 8, ep: 8, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [8, 16, 24, 32, 40, 48, 56] }
-      - { tp: 8, ep: 8, dp-attn: true, kv-offloading: none, spec-decoding: mtp, conc-list: [32, 48, 64, 80, 96, 112, 128, 144, 160, 176, 192, 224], router: { name: vllm-router, version: "0.1.14" } }
+      - { tp: 8, ep: 1, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [4, 6, 12] }
+      - { tp: 4, ep: 1, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [4, 6] }
+      - { tp: 8, ep: 8, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [56] }
       # DP4+EP4 is not fit-safe at the required 1,048,576-token model length:
       # a direct c56 probe left 12.09 GiB KV memory versus 14.88 GiB required.
       # LMCache invalid-block recovery currently assumes one KV-cache group,
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 972808ca4..741b7006b 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5720,4 +5720,5 @@
     - "Replace the isolated DP8/EP1 c64 point with 40 points spanning TP8, fit-validated TP4, TP8+EP8 and DP8+EP8. Mirror the B200/B300 MTP low-concurrency cells and extend DP8+EP8 through concurrency 224, with extra MI355X intermediates to resolve the local knee. DP8+EP8 is the intended throughput topology: attention remains data parallel while the FP4 experts are sharded across all eight GPUs."
     - "Direct fit validation shows TP4 leaves 37.47 GiB/GPU for a 2.82M-token KV cache and serves normally. DP4+EP4 is excluded because it leaves only 12.09 GiB/GPU KV versus 14.88 GiB required for the committed 1,048,576-token model length; the workload is not weakened to force an invalid topology."
     - "Keep MTP and its committed three-token golden synthetic acceptance length of 2.49 on every throughput point; eval-only runs retain real target verification. Enable prompt-token details for request-level cache accounting, and require a nonempty vllm: server-metrics export from AIPerf."
+    - "Use agentx-fast discovery before the reusable sweep: 25 completed throughput points show DP8+EP8 is uniformly dominated by TP8+EP8, while TP4 collapses above concurrency 12. Narrow the second fast pass to the six unresolved TP8 c4/c6/c12, TP4 c4/c6, and TP8+EP8 c56 cells; the final full-duration grid will contain only measured frontier points and adjacent guards."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2540

From 8e6d1b205873e34a2fbcfa197ec0ba8643643d73 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Sun, 9 Aug 2026 21:43:58 -0500
Subject: [PATCH 07/18] perf(mi355x): probe upper DSV4 MTP frontier

---
 configs/amd-master.yaml | 15 +++++++--------
 perf-changelog.yaml     |  1 +
 2 files changed, 8 insertions(+), 8 deletions(-)

diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 53990f569..4735a7049 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -1227,11 +1227,12 @@ qwen3.5-fp8-mi355x-sglang-agentic-hicache:
       - { tp: 8, ep: 1, kv-offloading: none, conc-list: [1, 2, 4, 8, 16, 32] }
       - { tp: 8, ep: 1, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [16, 32, 48, 64] }
 
-# DeepSeek-V4-Pro FP4 MTP refresh for MI355X. This second agentx-fast pass keeps
-# only the six unresolved cells from the broad B200/B300-shaped discovery grid:
-# low-concurrency TP guards and the high-end TP8+EP8 crossover. The first pass
-# established that DP8+EP8 is dominated and TP4 collapses above c12, so those
-# known-bad tails are intentionally excluded before the full-duration sweep.
+# DeepSeek-V4-Pro FP4 MTP refresh for MI355X. The broad and guard passes show
+# TP8+EP8 is the throughput topology: it beats DP8+EP8 throughout the overlap,
+# while TP8 weakly dominates TP4 at matched low concurrency. This final fast
+# pass samples the upper TP8+EP8 band every eight sessions. c48/c56 are still
+# scaling, but c56 already has 13.3s p95 TTFT; c96 is the linear KV-capacity
+# estimate and c104/c112 deliberately bracket the expected cache/latency cliff.
 dsv4-fp4-mi355x-vllm-agentic-mtp:
   image: vllm/vllm-openai-rocm:v0.26.0
   model: deepseek-ai/DeepSeek-V4-Pro
@@ -1244,9 +1245,7 @@ dsv4-fp4-mi355x-vllm-agentic-mtp:
     agentic-coding:
     - dram-utilization: 0.60
       search-space:
-      - { tp: 8, ep: 1, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [4, 6, 12] }
-      - { tp: 4, ep: 1, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [4, 6] }
-      - { tp: 8, ep: 8, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [56] }
+      - { tp: 8, ep: 8, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [64, 72, 80, 88, 96, 104, 112] }
       # DP4+EP4 is not fit-safe at the required 1,048,576-token model length:
       # a direct c56 probe left 12.09 GiB KV memory versus 14.88 GiB required.
       # LMCache invalid-block recovery currently assumes one KV-cache group,
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 741b7006b..3386faa6d 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5721,4 +5721,5 @@
     - "Direct fit validation shows TP4 leaves 37.47 GiB/GPU for a 2.82M-token KV cache and serves normally. DP4+EP4 is excluded because it leaves only 12.09 GiB/GPU KV versus 14.88 GiB required for the committed 1,048,576-token model length; the workload is not weakened to force an invalid topology."
     - "Keep MTP and its committed three-token golden synthetic acceptance length of 2.49 on every throughput point; eval-only runs retain real target verification. Enable prompt-token details for request-level cache accounting, and require a nonempty vllm: server-metrics export from AIPerf."
     - "Use agentx-fast discovery before the reusable sweep: 25 completed throughput points show DP8+EP8 is uniformly dominated by TP8+EP8, while TP4 collapses above concurrency 12. Narrow the second fast pass to the six unresolved TP8 c4/c6/c12, TP4 c4/c6, and TP8+EP8 c56 cells; the final full-duration grid will contain only measured frontier points and adjacent guards."
+    - "The second fast pass confirms TP8 weakly dominates TP4 at matched c4/c6. TP8+EP8 c56 reaches 89.22k input and 496.46 output tok/s with 52.9% KV use, but p95 TTFT has risen to 13.29s. Probe the remaining upper band at c64/c72/c80/c88/c96/c104/c112: c96 is the linear KV-capacity estimate from c48, while c104/c112 bracket the expected cache and latency cliff."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2540

From cbbc4bb97d840cac5ef85b61a7cd210236ed275a Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Sun, 9 Aug 2026 23:00:30 -0500
Subject: [PATCH 08/18] perf(mi355x): isolate native DRAM offload at c64

---
 configs/amd-master.yaml | 12 +++++-------
 perf-changelog.yaml     |  2 ++
 2 files changed, 7 insertions(+), 7 deletions(-)

diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 4735a7049..e77315344 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -1227,12 +1227,10 @@ qwen3.5-fp8-mi355x-sglang-agentic-hicache:
       - { tp: 8, ep: 1, kv-offloading: none, conc-list: [1, 2, 4, 8, 16, 32] }
       - { tp: 8, ep: 1, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [16, 32, 48, 64] }
 
-# DeepSeek-V4-Pro FP4 MTP refresh for MI355X. The broad and guard passes show
-# TP8+EP8 is the throughput topology: it beats DP8+EP8 throughout the overlap,
-# while TP8 weakly dominates TP4 at matched low concurrency. This final fast
-# pass samples the upper TP8+EP8 band every eight sessions. c48/c56 are still
-# scaling, but c56 already has 13.3s p95 TTFT; c96 is the linear KV-capacity
-# estimate and c104/c112 deliberately bracket the expected cache/latency cliff.
+# DeepSeek-V4-Pro FP4 MTP refresh for MI355X. GPU-only discovery places the
+# throughput knee at TP8+EP8 c56 and shows a monotonic collapse from c64 onward.
+# Isolate host KV as the next variable: compare native DRAM offload at c64
+# against the preserved GPU-only c64 result before expanding or publishing it.
 dsv4-fp4-mi355x-vllm-agentic-mtp:
   image: vllm/vllm-openai-rocm:v0.26.0
   model: deepseek-ai/DeepSeek-V4-Pro
@@ -1245,7 +1243,7 @@ dsv4-fp4-mi355x-vllm-agentic-mtp:
     agentic-coding:
     - dram-utilization: 0.60
       search-space:
-      - { tp: 8, ep: 8, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [64, 72, 80, 88, 96, 104, 112] }
+      - { tp: 8, ep: 8, dp-attn: false, kv-offloading: dram, kv-offload-backend: { name: vllm-native }, spec-decoding: mtp, conc-list: [64] }
       # DP4+EP4 is not fit-safe at the required 1,048,576-token model length:
       # a direct c56 probe left 12.09 GiB KV memory versus 14.88 GiB required.
       # LMCache invalid-block recovery currently assumes one KV-cache group,
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 3386faa6d..e6a18884b 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5722,4 +5722,6 @@
     - "Keep MTP and its committed three-token golden synthetic acceptance length of 2.49 on every throughput point; eval-only runs retain real target verification. Enable prompt-token details for request-level cache accounting, and require a nonempty vllm: server-metrics export from AIPerf."
     - "Use agentx-fast discovery before the reusable sweep: 25 completed throughput points show DP8+EP8 is uniformly dominated by TP8+EP8, while TP4 collapses above concurrency 12. Narrow the second fast pass to the six unresolved TP8 c4/c6/c12, TP4 c4/c6, and TP8+EP8 c56 cells; the final full-duration grid will contain only measured frontier points and adjacent guards."
     - "The second fast pass confirms TP8 weakly dominates TP4 at matched c4/c6. TP8+EP8 c56 reaches 89.22k input and 496.46 output tok/s with 52.9% KV use, but p95 TTFT has risen to 13.29s. Probe the remaining upper band at c64/c72/c80/c88/c96/c104/c112: c96 is the linear KV-capacity estimate from c48, while c104/c112 bracket the expected cache and latency cliff."
+    - "The GPU-only upper-band pass completes on all seven points and confirms a monotonic post-c56 collapse: c64/c72/c80 reach 79.34k/414.46, 72.17k/306.93, and 59.06k/284.48 input/output tok/s; c88/c96/c104/c112 fall to 34.33k/142.61, 24.39k/117.42, 15.24k/79.54, and 13.28k/70.73. All points have zero request errors; c88/c96/c112 miss only the fast-window completed-record TTFT coverage gate."
+    - "Test one isolated follow-up variable before finalizing: TP8+EP8 c64 with the script's supported vLLM-native DRAM offload, leaving vLLM v0.26.0, MTP3, golden AL 2.49, FP8 KV, topology, concurrency, memory utilization, and AgentX trace unchanged. Keep GPU-only c56/c64 as the preserved baseline and expand offload only if c64 improves."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2540

From 96703c9d3c690df47127885084fdc9f1fa2b0673 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 10 Aug 2026 00:11:12 -0500
Subject: [PATCH 09/18] perf(mi355x): publish measured DSV4 MTP frontier

---
 configs/amd-master.yaml | 13 ++++++++-----
 perf-changelog.yaml     |  2 ++
 2 files changed, 10 insertions(+), 5 deletions(-)

diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index e77315344..5055dd73b 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -1227,10 +1227,12 @@ qwen3.5-fp8-mi355x-sglang-agentic-hicache:
       - { tp: 8, ep: 1, kv-offloading: none, conc-list: [1, 2, 4, 8, 16, 32] }
       - { tp: 8, ep: 1, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [16, 32, 48, 64] }
 
-# DeepSeek-V4-Pro FP4 MTP refresh for MI355X. GPU-only discovery places the
-# throughput knee at TP8+EP8 c56 and shows a monotonic collapse from c64 onward.
-# Isolate host KV as the next variable: compare native DRAM offload at c64
-# against the preserved GPU-only c64 result before expanding or publishing it.
+# DeepSeek-V4-Pro FP4 MTP refresh for MI355X. Fast discovery measured the full
+# B200/B300-shaped topology space one variable at a time: TP8 weakly dominates
+# TP4 at low concurrency, TP8+EP8 dominates DP8+EP8, and native DRAM offload at
+# c64 regresses the GPU-only point without producing an external-cache hit. The
+# reusable sweep therefore follows the measured GPU-resident frontier from TP8
+# c1-c12 into TP8+EP8 c16-c56, with c64 retained as the post-knee guard.
 dsv4-fp4-mi355x-vllm-agentic-mtp:
   image: vllm/vllm-openai-rocm:v0.26.0
   model: deepseek-ai/DeepSeek-V4-Pro
@@ -1243,7 +1245,8 @@ dsv4-fp4-mi355x-vllm-agentic-mtp:
     agentic-coding:
     - dram-utilization: 0.60
       search-space:
-      - { tp: 8, ep: 8, dp-attn: false, kv-offloading: dram, kv-offload-backend: { name: vllm-native }, spec-decoding: mtp, conc-list: [64] }
+      - { tp: 8, ep: 1, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 4, 6, 8, 12] }
+      - { tp: 8, ep: 8, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [16, 24, 32, 40, 48, 56, 64] }
       # DP4+EP4 is not fit-safe at the required 1,048,576-token model length:
       # a direct c56 probe left 12.09 GiB KV memory versus 14.88 GiB required.
       # LMCache invalid-block recovery currently assumes one KV-cache group,
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index e6a18884b..50e9f230a 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5724,4 +5724,6 @@
     - "The second fast pass confirms TP8 weakly dominates TP4 at matched c4/c6. TP8+EP8 c56 reaches 89.22k input and 496.46 output tok/s with 52.9% KV use, but p95 TTFT has risen to 13.29s. Probe the remaining upper band at c64/c72/c80/c88/c96/c104/c112: c96 is the linear KV-capacity estimate from c48, while c104/c112 bracket the expected cache and latency cliff."
     - "The GPU-only upper-band pass completes on all seven points and confirms a monotonic post-c56 collapse: c64/c72/c80 reach 79.34k/414.46, 72.17k/306.93, and 59.06k/284.48 input/output tok/s; c88/c96/c104/c112 fall to 34.33k/142.61, 24.39k/117.42, 15.24k/79.54, and 13.28k/70.73. All points have zero request errors; c88/c96/c112 miss only the fast-window completed-record TTFT coverage gate."
     - "Test one isolated follow-up variable before finalizing: TP8+EP8 c64 with the script's supported vLLM-native DRAM offload, leaving vLLM v0.26.0, MTP3, golden AL 2.49, FP8 KV, topology, concurrency, memory utilization, and AgentX trace unchanged. Keep GPU-only c56/c64 as the preserved baseline and expand offload only if c64 improves."
+    - "Reject native DRAM offload after the isolated c64 test: it completes 719 valid requests with zero request errors but reaches only 73.45k input and 366.17 output tok/s versus the preserved GPU-only c64 result at 79.34k/414.46, while reporting zero external-cache hits despite roughly 20 TB of cumulative host-KV writes. The fast-window TTFT coverage miss (92.8% with ITL at 100%) is consistent with the slower long-response tail and does not change the throughput regression."
+    - "Run the reusable one-hour sweep on the measured GPU-resident Pareto grid only: TP8 c1/c2/c4/c6/c8/c12 for the low-latency band, then TP8+EP8 c16/c24/c32/c40/c48/c56/c64 for the throughput band. This preserves the B200/B300 low-concurrency spacing, switches topology at the measured crossover, includes every sampled rise to the c56 knee, and retains c64 as the adjacent post-knee guard without re-running the dominated TP4, DP8+EP8, offload, or c72+ tails."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2540

From 11edf7a53ac844b8815902ff54d909826debb36b Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 10 Aug 2026 04:29:31 -0500
Subject: [PATCH 10/18] perf(mi355x): probe larger DSV4 prefill budget

---
 .../single_node/agentic/dsv4_fp4_mi355x_vllm_mtp.sh      | 2 +-
 configs/amd-master.yaml                                  | 9 +++++----
 perf-changelog.yaml                                      | 1 +
 3 files changed, 7 insertions(+), 5 deletions(-)

diff --git a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_vllm_mtp.sh b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_vllm_mtp.sh
index 58d04df7a..3b5596b2f 100644
--- a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_vllm_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_vllm_mtp.sh
@@ -406,7 +406,7 @@ VLLM_CMD=(
     "${EP_ARGS[@]}"
     --gpu-memory-utilization 0.9
     --block-size 256
-    --max-num-batched-tokens 8192
+    --max-num-batched-tokens 16384
     --moe-backend aiter
     --compilation-config '{"mode":3,"cudagraph_mode":"FULL_DECODE_ONLY"}'
     --speculative-config "$SPEC_CONFIG"
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 5055dd73b..eb43766b4 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -1231,8 +1231,10 @@ qwen3.5-fp8-mi355x-sglang-agentic-hicache:
 # B200/B300-shaped topology space one variable at a time: TP8 weakly dominates
 # TP4 at low concurrency, TP8+EP8 dominates DP8+EP8, and native DRAM offload at
 # c64 regresses the GPU-only point without producing an external-cache hit. The
-# reusable sweep therefore follows the measured GPU-resident frontier from TP8
-# c1-c12 into TP8+EP8 c16-c56, with c64 retained as the post-knee guard.
+# reusable sweep followed the measured GPU-resident frontier from TP8 c1-c12
+# into TP8+EP8 c16-c64. The current fast follow-up isolates only c48 while
+# testing a larger prefill token budget; the all-green full grid remains the
+# rollback baseline in run 31357800541.
 dsv4-fp4-mi355x-vllm-agentic-mtp:
   image: vllm/vllm-openai-rocm:v0.26.0
   model: deepseek-ai/DeepSeek-V4-Pro
@@ -1245,8 +1247,7 @@ dsv4-fp4-mi355x-vllm-agentic-mtp:
     agentic-coding:
     - dram-utilization: 0.60
       search-space:
-      - { tp: 8, ep: 1, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 4, 6, 8, 12] }
-      - { tp: 8, ep: 8, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [16, 24, 32, 40, 48, 56, 64] }
+      - { tp: 8, ep: 8, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [48] }
       # DP4+EP4 is not fit-safe at the required 1,048,576-token model length:
       # a direct c56 probe left 12.09 GiB KV memory versus 14.88 GiB required.
       # LMCache invalid-block recovery currently assumes one KV-cache group,
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 50e9f230a..1cb731717 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5726,4 +5726,5 @@
     - "Test one isolated follow-up variable before finalizing: TP8+EP8 c64 with the script's supported vLLM-native DRAM offload, leaving vLLM v0.26.0, MTP3, golden AL 2.49, FP8 KV, topology, concurrency, memory utilization, and AgentX trace unchanged. Keep GPU-only c56/c64 as the preserved baseline and expand offload only if c64 improves."
     - "Reject native DRAM offload after the isolated c64 test: it completes 719 valid requests with zero request errors but reaches only 73.45k input and 366.17 output tok/s versus the preserved GPU-only c64 result at 79.34k/414.46, while reporting zero external-cache hits despite roughly 20 TB of cumulative host-KV writes. The fast-window TTFT coverage miss (92.8% with ITL at 100%) is consistent with the slower long-response tail and does not change the throughput regression."
     - "Run the reusable one-hour sweep on the measured GPU-resident Pareto grid only: TP8 c1/c2/c4/c6/c8/c12 for the low-latency band, then TP8+EP8 c16/c24/c32/c40/c48/c56/c64 for the throughput band. This preserves the B200/B300 low-concurrency spacing, switches topology at the measured crossover, includes every sampled rise to the c56 knee, and retains c64 as the adjacent post-knee guard without re-running the dominated TP4, DP8+EP8, offload, or c72+ tails."
+    - "The reusable GPU-only grid completed green in run 31357800541. Isolate one follow-up variable at its input-throughput peak, TP8+EP8 c48: raise max-num-batched-tokens from 8192 to 16384 while keeping vLLM v0.26.0, topology, concurrency, MTP3/golden AL 2.49, FP8 KV, GPU memory utilization, and AgentX trace unchanged. Use agentx-fast and preserve the full-grid run as the rollback baseline."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2540

From 86773f811263eccf45aa9600520d9e6bc97efae7 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 10 Aug 2026 05:43:07 -0500
Subject: [PATCH 11/18] perf(mi355x): probe shallower DSV4 MTP

---
 .../agentic/dsv4_fp4_mi355x_vllm_mtp.sh           | 15 +++++++++------
 perf-changelog.yaml                               |  2 ++
 2 files changed, 11 insertions(+), 6 deletions(-)

diff --git a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_vllm_mtp.sh b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_vllm_mtp.sh
index 3b5596b2f..168ee6223 100644
--- a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_vllm_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_vllm_mtp.sh
@@ -376,11 +376,14 @@ fi
 # leave 2x headroom rather than clipping those bursts at the scheduler.
 MAX_NUM_SEQS=$((2 * CONC))
 
-# DeepSeek-V4-Pro ships a native MTP head. AgentX throughput pins its
-# three-token draft to the committed thinking-on golden acceptance length;
-# eval-only runs use real target verification so accuracy remains meaningful.
-NUM_SPEC_TOKENS=3
-SYNTHETIC_ACCEPT_LEN=2.49
+# DeepSeek-V4-Pro ships a native MTP head. AgentX throughput pins each tested
+# draft depth to its committed thinking-on golden acceptance length; eval-only
+# runs use real target verification so accuracy remains meaningful.
+# The isolated 16K scheduler-budget probe regressed the matched c48 control, so
+# restore 8K and isolate the K=2 golden MTP operating point. Draft depth and its
+# measured golden acceptance length are one coupled speculative configuration.
+NUM_SPEC_TOKENS=2
+SYNTHETIC_ACCEPT_LEN=2.27
 if [ "${EVAL_ONLY:-false}" = "true" ]; then
     SPEC_CONFIG="{\"method\": \"mtp\", \"num_speculative_tokens\": $NUM_SPEC_TOKENS}"
 else
@@ -406,7 +409,7 @@ VLLM_CMD=(
     "${EP_ARGS[@]}"
     --gpu-memory-utilization 0.9
     --block-size 256
-    --max-num-batched-tokens 16384
+    --max-num-batched-tokens 8192
     --moe-backend aiter
     --compilation-config '{"mode":3,"cudagraph_mode":"FULL_DECODE_ONLY"}'
     --speculative-config "$SPEC_CONFIG"
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 1cb731717..be2103861 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5727,4 +5727,6 @@
     - "Reject native DRAM offload after the isolated c64 test: it completes 719 valid requests with zero request errors but reaches only 73.45k input and 366.17 output tok/s versus the preserved GPU-only c64 result at 79.34k/414.46, while reporting zero external-cache hits despite roughly 20 TB of cumulative host-KV writes. The fast-window TTFT coverage miss (92.8% with ITL at 100%) is consistent with the slower long-response tail and does not change the throughput regression."
     - "Run the reusable one-hour sweep on the measured GPU-resident Pareto grid only: TP8 c1/c2/c4/c6/c8/c12 for the low-latency band, then TP8+EP8 c16/c24/c32/c40/c48/c56/c64 for the throughput band. This preserves the B200/B300 low-concurrency spacing, switches topology at the measured crossover, includes every sampled rise to the c56 knee, and retains c64 as the adjacent post-knee guard without re-running the dominated TP4, DP8+EP8, offload, or c72+ tails."
     - "The reusable GPU-only grid completed green in run 31357800541. Isolate one follow-up variable at its input-throughput peak, TP8+EP8 c48: raise max-num-batched-tokens from 8192 to 16384 while keeping vLLM v0.26.0, topology, concurrency, MTP3/golden AL 2.49, FP8 KV, GPU memory utilization, and AgentX trace unchanged. Use agentx-fast and preserve the full-grid run as the rollback baseline."
+    - "Reject the 16384-token scheduler budget after its exact c48 fast comparison completed fully green but regressed to 80,316.53 input / 412.36 output tok/s over 1,228.83 seconds, versus 86,334.85 / 471.67 over 1,229.32 seconds for the duration-matched 8192-token control from run 31341602956. Restore 8192 before any further experiment."
+    - "Next isolate only the golden MTP operating point at c48: switch from K=3 with its measured synthetic acceptance length 2.49 to K=2 with its measured synthetic acceptance length 2.27, while retaining vLLM v0.26.0, TP8+EP8, max-num-seqs 96, FP8 KV, 0.9 memory utilization, FULL_DECODE_ONLY graphs, and the exact AgentX trace. Draft depth and its matching golden acceptance length are one coupled speculative configuration."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2540

From 9e245f4c1b0fe723f035f71210ccb0b45ff2410b Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 10 Aug 2026 06:41:47 -0500
Subject: [PATCH 12/18] perf(mi355x): probe piecewise DSV4 graphs

---
 .../agentic/dsv4_fp4_mi355x_vllm_mtp.sh         | 17 ++++++++---------
 perf-changelog.yaml                             |  2 ++
 2 files changed, 10 insertions(+), 9 deletions(-)

diff --git a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_vllm_mtp.sh b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_vllm_mtp.sh
index 168ee6223..a8aabe847 100644
--- a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_vllm_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_vllm_mtp.sh
@@ -376,14 +376,13 @@ fi
 # leave 2x headroom rather than clipping those bursts at the scheduler.
 MAX_NUM_SEQS=$((2 * CONC))
 
-# DeepSeek-V4-Pro ships a native MTP head. AgentX throughput pins each tested
-# draft depth to its committed thinking-on golden acceptance length; eval-only
-# runs use real target verification so accuracy remains meaningful.
-# The isolated 16K scheduler-budget probe regressed the matched c48 control, so
-# restore 8K and isolate the K=2 golden MTP operating point. Draft depth and its
-# measured golden acceptance length are one coupled speculative configuration.
-NUM_SPEC_TOKENS=2
-SYNTHETIC_ACCEPT_LEN=2.27
+# DeepSeek-V4-Pro ships a native MTP head. AgentX throughput pins its
+# three-token draft to the committed thinking-on golden acceptance length;
+# eval-only runs use real target verification so accuracy remains meaningful.
+# The isolated 16K scheduler-budget and K=2 probes both regressed their matched
+# c48 controls. Restore the K=3/8K baseline before isolating graph capture mode.
+NUM_SPEC_TOKENS=3
+SYNTHETIC_ACCEPT_LEN=2.49
 if [ "${EVAL_ONLY:-false}" = "true" ]; then
     SPEC_CONFIG="{\"method\": \"mtp\", \"num_speculative_tokens\": $NUM_SPEC_TOKENS}"
 else
@@ -411,7 +410,7 @@ VLLM_CMD=(
     --block-size 256
     --max-num-batched-tokens 8192
     --moe-backend aiter
-    --compilation-config '{"mode":3,"cudagraph_mode":"FULL_DECODE_ONLY"}'
+    --compilation-config '{"mode":3,"cudagraph_mode":"FULL_AND_PIECEWISE"}'
     --speculative-config "$SPEC_CONFIG"
     --tokenizer-mode deepseek_v4
     --tool-call-parser deepseek_v4
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index be2103861..e40ed8615 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5729,4 +5729,6 @@
     - "The reusable GPU-only grid completed green in run 31357800541. Isolate one follow-up variable at its input-throughput peak, TP8+EP8 c48: raise max-num-batched-tokens from 8192 to 16384 while keeping vLLM v0.26.0, topology, concurrency, MTP3/golden AL 2.49, FP8 KV, GPU memory utilization, and AgentX trace unchanged. Use agentx-fast and preserve the full-grid run as the rollback baseline."
     - "Reject the 16384-token scheduler budget after its exact c48 fast comparison completed fully green but regressed to 80,316.53 input / 412.36 output tok/s over 1,228.83 seconds, versus 86,334.85 / 471.67 over 1,229.32 seconds for the duration-matched 8192-token control from run 31341602956. Restore 8192 before any further experiment."
     - "Next isolate only the golden MTP operating point at c48: switch from K=3 with its measured synthetic acceptance length 2.49 to K=2 with its measured synthetic acceptance length 2.27, while retaining vLLM v0.26.0, TP8+EP8, max-num-seqs 96, FP8 KV, 0.9 memory utilization, FULL_DECODE_ONLY graphs, and the exact AgentX trace. Draft depth and its matching golden acceptance length are one coupled speculative configuration."
+    - "Reject K=2 after the exact c48 fast run completed fully green at 83,283.76 input / 448.52 output tok/s over 1,226.65 seconds, versus 86,334.85 / 471.67 over 1,229.32 seconds for the duration-matched K=3 control: input regressed 3.53% and output regressed 4.91%. Restore K=3 with golden acceptance length 2.49 before any further experiment."
+    - "Next isolate only graph capture mode at c48: switch FULL_DECODE_ONLY to FULL_AND_PIECEWISE while keeping vLLM v0.26.0, TP8+EP8, K=3/golden AL 2.49, max-num-batched-tokens 8192, max-num-seqs 96, FP8 KV, 0.9 memory utilization, and the exact AgentX trace unchanged."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2540

From 1de2b07945a9b5173721da33d57800d2dddb5956 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 10 Aug 2026 07:39:39 -0500
Subject: [PATCH 13/18] perf(mi355x): probe INT4 quick reduce

---
 .../single_node/agentic/dsv4_fp4_mi355x_vllm_mtp.sh      | 9 +++++----
 configs/amd-master.yaml                                  | 6 +++---
 perf-changelog.yaml                                      | 2 ++
 3 files changed, 10 insertions(+), 7 deletions(-)

diff --git a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_vllm_mtp.sh b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_vllm_mtp.sh
index a8aabe847..2db1c452c 100644
--- a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_vllm_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_vllm_mtp.sh
@@ -379,8 +379,9 @@ MAX_NUM_SEQS=$((2 * CONC))
 # DeepSeek-V4-Pro ships a native MTP head. AgentX throughput pins its
 # three-token draft to the committed thinking-on golden acceptance length;
 # eval-only runs use real target verification so accuracy remains meaningful.
-# The isolated 16K scheduler-budget and K=2 probes both regressed their matched
-# c48 controls. Restore the K=3/8K baseline before isolating graph capture mode.
+# The isolated 16K scheduler-budget, K=2, and piecewise-graph probes all
+# regressed their matched c48 controls. Restore the K=3/8K/FULL_DECODE_ONLY
+# baseline before isolating INT4 Quick Reduce.
 NUM_SPEC_TOKENS=3
 SYNTHETIC_ACCEPT_LEN=2.49
 if [ "${EVAL_ONLY:-false}" = "true" ]; then
@@ -392,7 +393,7 @@ fi
 echo "Starting vllm server..."
 set -x
 export VLLM_ROCM_USE_AITER=1
-#export VLLM_ROCM_QUICK_REDUCE_QUANTIZATION=INT4
+export VLLM_ROCM_QUICK_REDUCE_QUANTIZATION=INT4
 export VLLM_ROCM_USE_AITER_MOE=1
 
 { set +x; } 2>/dev/null
@@ -410,7 +411,7 @@ VLLM_CMD=(
     --block-size 256
     --max-num-batched-tokens 8192
     --moe-backend aiter
-    --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 eb43766b4..97e52a249 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -1232,9 +1232,9 @@ qwen3.5-fp8-mi355x-sglang-agentic-hicache:
 # TP4 at low concurrency, TP8+EP8 dominates DP8+EP8, and native DRAM offload at
 # c64 regresses the GPU-only point without producing an external-cache hit. The
 # reusable sweep followed the measured GPU-resident frontier from TP8 c1-c12
-# into TP8+EP8 c16-c64. The current fast follow-up isolates only c48 while
-# testing a larger prefill token budget; the all-green full grid remains the
-# rollback baseline in run 31357800541.
+# into TP8+EP8 c16-c64. The current fast follow-up isolates INT4 Quick Reduce
+# at c48; the all-green full grid remains the rollback baseline in run
+# 31357800541.
 dsv4-fp4-mi355x-vllm-agentic-mtp:
   image: vllm/vllm-openai-rocm:v0.26.0
   model: deepseek-ai/DeepSeek-V4-Pro
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index e40ed8615..32378d16a 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5731,4 +5731,6 @@
     - "Next isolate only the golden MTP operating point at c48: switch from K=3 with its measured synthetic acceptance length 2.49 to K=2 with its measured synthetic acceptance length 2.27, while retaining vLLM v0.26.0, TP8+EP8, max-num-seqs 96, FP8 KV, 0.9 memory utilization, FULL_DECODE_ONLY graphs, and the exact AgentX trace. Draft depth and its matching golden acceptance length are one coupled speculative configuration."
     - "Reject K=2 after the exact c48 fast run completed fully green at 83,283.76 input / 448.52 output tok/s over 1,226.65 seconds, versus 86,334.85 / 471.67 over 1,229.32 seconds for the duration-matched K=3 control: input regressed 3.53% and output regressed 4.91%. Restore K=3 with golden acceptance length 2.49 before any further experiment."
     - "Next isolate only graph capture mode at c48: switch FULL_DECODE_ONLY to FULL_AND_PIECEWISE while keeping vLLM v0.26.0, TP8+EP8, K=3/golden AL 2.49, max-num-batched-tokens 8192, max-num-seqs 96, FP8 KV, 0.9 memory utilization, and the exact AgentX trace unchanged."
+    - "Reject FULL_AND_PIECEWISE after the exact c48 fast run completed fully green at 83,955.53 input / 455.90 output tok/s over 1,228.87 seconds, versus 86,334.85 / 471.67 over 1,229.32 seconds for the duration-matched FULL_DECODE_ONLY control: input regressed 2.76% and output regressed 3.34%. The run processed 795 valid requests with zero request errors, passed TTFT/ITL coverage, and exported nonempty vllm: metrics, so this is a measured graph-mode regression rather than infrastructure noise."
+    - "Next isolate only INT4 Quick Reduce at the retained c48 baseline: enable VLLM_ROCM_QUICK_REDUCE_QUANTIZATION=INT4 while restoring FULL_DECODE_ONLY and keeping vLLM v0.26.0, TP8+EP8, K=3/golden AL 2.49, max-num-batched-tokens 8192, max-num-seqs 96, FP8 KV, 0.9 memory utilization, and the exact AgentX trace unchanged. This is a standard ROCm vLLM communication optimization already used by other MI355X recipes; the paired eval guards accuracy."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2540

From 4c52de6c8aacf4a89d6d4c7a99cee79d6ade1c4d Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 10 Aug 2026 08:31:17 -0500
Subject: [PATCH 14/18] perf(mi355x): test compiled DSV4 graph path

---
 .../single_node/agentic/dsv4_fp4_mi355x_vllm_mtp.sh       | 8 ++++----
 perf-changelog.yaml                                       | 2 ++
 2 files changed, 6 insertions(+), 4 deletions(-)

diff --git a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_vllm_mtp.sh b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_vllm_mtp.sh
index 2db1c452c..7beb30101 100644
--- a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_vllm_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_vllm_mtp.sh
@@ -379,9 +379,9 @@ MAX_NUM_SEQS=$((2 * CONC))
 # DeepSeek-V4-Pro ships a native MTP head. AgentX throughput pins its
 # three-token draft to the committed thinking-on golden acceptance length;
 # eval-only runs use real target verification so accuracy remains meaningful.
-# The isolated 16K scheduler-budget, K=2, and piecewise-graph probes all
-# regressed their matched c48 controls. Restore the K=3/8K/FULL_DECODE_ONLY
-# baseline before isolating INT4 Quick Reduce.
+# The isolated 16K scheduler-budget, K=2, piecewise-graph, and INT4 Quick
+# Reduce probes did not improve their matched c48 controls. Restore the
+# K=3/8K/FULL_DECODE_ONLY baseline before isolating breakable CUDA graphs.
 NUM_SPEC_TOKENS=3
 SYNTHETIC_ACCEPT_LEN=2.49
 if [ "${EVAL_ONLY:-false}" = "true" ]; then
@@ -393,7 +393,7 @@ fi
 echo "Starting vllm server..."
 set -x
 export VLLM_ROCM_USE_AITER=1
-export VLLM_ROCM_QUICK_REDUCE_QUANTIZATION=INT4
+export VLLM_USE_BREAKABLE_CUDAGRAPH=0
 export VLLM_ROCM_USE_AITER_MOE=1
 
 { set +x; } 2>/dev/null
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 32378d16a..34f6c8951 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5733,4 +5733,6 @@
     - "Next isolate only graph capture mode at c48: switch FULL_DECODE_ONLY to FULL_AND_PIECEWISE while keeping vLLM v0.26.0, TP8+EP8, K=3/golden AL 2.49, max-num-batched-tokens 8192, max-num-seqs 96, FP8 KV, 0.9 memory utilization, and the exact AgentX trace unchanged."
     - "Reject FULL_AND_PIECEWISE after the exact c48 fast run completed fully green at 83,955.53 input / 455.90 output tok/s over 1,228.87 seconds, versus 86,334.85 / 471.67 over 1,229.32 seconds for the duration-matched FULL_DECODE_ONLY control: input regressed 2.76% and output regressed 3.34%. The run processed 795 valid requests with zero request errors, passed TTFT/ITL coverage, and exported nonempty vllm: metrics, so this is a measured graph-mode regression rather than infrastructure noise."
     - "Next isolate only INT4 Quick Reduce at the retained c48 baseline: enable VLLM_ROCM_QUICK_REDUCE_QUANTIZATION=INT4 while restoring FULL_DECODE_ONLY and keeping vLLM v0.26.0, TP8+EP8, K=3/golden AL 2.49, max-num-batched-tokens 8192, max-num-seqs 96, FP8 KV, 0.9 memory utilization, and the exact AgentX trace unchanged. This is a standard ROCm vLLM communication optimization already used by other MI355X recipes; the paired eval guards accuracy."
+    - "Reject INT4 Quick Reduce after the exact c48 fast run completed successfully at 87,176.57 input / 470.37 output tok/s over 1,228.05 seconds, versus 86,334.85 / 471.67 over 1,229.32 seconds for the duration-matched control: input improved only 0.98%, output regressed 0.28%, and the candidate introduced one empty-content response while the control was error-free. The 0.12% error rate remained below the harness threshold, TTFT/ITL coverage was 100%, GSM8K passed, and vllm: metrics exported, but this is not a clean Pareto improvement, so restore unquantized collectives."
+    - "Next isolate only vLLM's graph execution path at the retained c48 baseline: set VLLM_USE_BREAKABLE_CUDAGRAPH=0 while keeping FULL_DECODE_ONLY, vLLM v0.26.0, TP8+EP8, K=3/golden AL 2.49, max-num-batched-tokens 8192, max-num-seqs 96, FP8 KV, 0.9 memory utilization, and the exact AgentX trace unchanged. Current logs show auto-enabled breakable CUDA graphs force CompilationMode.NONE despite compilation mode 3; opting out tests the actual torch.compile path without changing graph capture mode or any workload setting."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2540

From 428e6b740455022fce112f7da376047da6e238e1 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 10 Aug 2026 12:50:09 -0500
Subject: [PATCH 15/18] perf(mi355x): add offload and DEP collection points

---
 .../agentic/dsv4_fp4_mi355x_vllm_mtp.sh       | 144 +++++++++++-------
 configs/amd-master.yaml                       |  25 +--
 perf-changelog.yaml                           |   1 +
 3 files changed, 103 insertions(+), 67 deletions(-)

diff --git a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_vllm_mtp.sh b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_vllm_mtp.sh
index 7beb30101..187f003c8 100644
--- a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_vllm_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_vllm_mtp.sh
@@ -107,6 +107,87 @@ mkdir -p "$RESULT_DIR"
 SERVER_PID=""
 ROUTER_PID=""
 MOONCAKE_MASTER_PID=""
+LMCACHE_PID=""
+
+cleanup_agentic_services() {
+    local exit_code=$?
+    trap - EXIT INT TERM
+    set +e
+    stop_background_process_tree "$ROUTER_PID" "vLLM router"
+    stop_background_process_tree "$SERVER_PID" "vLLM server" 60
+    stop_background_process_tree "$MOONCAKE_MASTER_PID" "Mooncake master"
+    stop_background_process_tree "$LMCACHE_PID" "LMCache server"
+    exit "$exit_code"
+}
+trap cleanup_agentic_services EXIT
+trap 'exit 130' INT
+trap 'exit 143' TERM
+
+# Mooncake does not publish a ROCm wheel. Build the pinned release once for
+# each immutable OS/Python/ROCm tuple and reuse the shared cache on later jobs.
+install_mooncake_rocm() {
+    local mooncake_tag="v0.3.11.post1"
+    local mooncake_src="/tmp/Mooncake-$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=()
 
@@ -139,30 +220,13 @@ case "${KV_OFFLOAD_BACKEND:-}" in
     ;;
   mooncake)
     require_agentic_kv_offload_backend mooncake
-    # ---- Mooncake config ----------------------------------------------------------
         # Embedded mode contributes one segment per GPU rank to a shared
-        # distributed store, so pre-divide the aggregate host-memory budget.
+        # distributed store, so pre-divide the generated aggregate budget.
         PER_RANK_GB=$((TOTAL_CPU_DRAM_GB / TP))
-
-        #MOONCAKE_VERSION=0.3.11.post1
-        #apt-get update && apt-get install -y libcurl4 libibverbs1 rdma-core librdmacm1 libnuma1 liburing2
-        #agentic_pip_install --quiet --no-cache-dir --no-deps \
-        #    --force-reinstall "mooncake-transfer-engine-non-cuda==$MOONCAKE_VERSION"
-
-        git clone https://github.com/kvcache-ai/Mooncake.git
-        cd Mooncake
-        bash dependencies.sh
-        mkdir build
-        cd build
-        cmake ..
-        make -j
-        sudo make install # optional, make it ready to be used by vLLM/SGLang
-        cd ..
-        cd ..
-
+        if ! python3 -c "from mooncake.store import MooncakeDistributedStore" >/dev/null 2>&1; then
+            install_mooncake_rocm
+        fi
         python3 -c "from mooncake.store import MooncakeDistributedStore" >/dev/null
-        export INFERENCEX_MOONCAKE_MAX_TRANSFER_BATCH_KEYS=32
-        python3 "$(dirname "$0")/patch_vllm_mooncake_transfer_batches.py"
 
         MOONCAKE_MASTER_PORT=$((PORT + 12000))
         MOONCAKE_CONFIG_PATH="$RESULT_DIR/mooncake_config.json"
@@ -178,22 +242,15 @@ case "${KV_OFFLOAD_BACKEND:-}" in
   "enable_offload": false
 }
 EOF
-# (srok)
-  #"protocol": "rdma",
-  #"device_name": "mlx5_0",
-  #"local_buffer_size": "4GB",
         export MOONCAKE_CONFIG_PATH
         export MC_ENABLE_DEST_DEVICE_AFFINITY=1
         export PYTHONHASHSEED=0
         export MC_SLICE_SIZE=1048576
-        # (srok)
-        #export MC_WORKERS_PER_CTX=4
         export MC_WORKERS_PER_CTX=8
 
         MOONCAKE_EVICTION_HIGH_WATERMARK_RATIO=0.80
         MOONCAKE_EVICTION_RATIO=0.10
         MOONCAKE_KV_LEASE_TTL=60s
-        #MOONCAKE_KV_LEASE_TTL=3600s
 
         echo "Starting Mooncake master on port $MOONCAKE_MASTER_PORT..."
         mooncake_master --port "$MOONCAKE_MASTER_PORT" \
@@ -219,30 +276,6 @@ EOF
   lmcache)
     require_agentic_kv_offload_backend lmcache
     # ---- Lmcache config ----------------------------------------------------------
-    LMCACHE_PID=""
-
-    cleanup_lmcache_server() {
-        if [[ -n "$LMCACHE_PID" ]] && kill -0 "$LMCACHE_PID" 2>/dev/null; then
-            kill "$LMCACHE_PID" 2>/dev/null || true
-            wait "$LMCACHE_PID" 2>/dev/null || true
-        fi
-    }
-
-    trap cleanup_lmcache_server EXIT
-
-    cleanup_agentic_services() {
-        local exit_code=$?
-        trap - EXIT INT TERM
-        set +e
-        stop_background_process_tree "$ROUTER_PID" "vLLM router"
-        stop_background_process_tree "$SERVER_PID" "vLLM server" 60
-        stop_background_process_tree "$MOONCAKE_MASTER_PID" "Mooncake master"
-        exit "$exit_code"
-    }
-    trap cleanup_agentic_services EXIT
-    trap 'exit 130' INT
-    trap 'exit 143' TERM
-
     wait_for_lmcache_ready() {
         { set +x; } 2>/dev/null
         local attempts="${LMCACHE_READY_ATTEMPTS:-120}"
@@ -379,9 +412,9 @@ MAX_NUM_SEQS=$((2 * CONC))
 # DeepSeek-V4-Pro ships a native MTP head. AgentX throughput pins its
 # three-token draft to the committed thinking-on golden acceptance length;
 # eval-only runs use real target verification so accuracy remains meaningful.
-# The isolated 16K scheduler-budget, K=2, piecewise-graph, and INT4 Quick
-# Reduce probes did not improve their matched c48 controls. Restore the
-# K=3/8K/FULL_DECODE_ONLY baseline before isolating breakable CUDA graphs.
+# The isolated scheduler, MTP-depth, graph-mode, collective, and breakable-
+# graph probes did not produce a clean improvement. Keep the verified
+# K=3/8K/FULL_DECODE_ONLY baseline while expanding topology coverage.
 NUM_SPEC_TOKENS=3
 SYNTHETIC_ACCEPT_LEN=2.49
 if [ "${EVAL_ONLY:-false}" = "true" ]; then
@@ -393,7 +426,6 @@ fi
 echo "Starting vllm server..."
 set -x
 export VLLM_ROCM_USE_AITER=1
-export VLLM_USE_BREAKABLE_CUDAGRAPH=0
 export VLLM_ROCM_USE_AITER_MOE=1
 
 { set +x; } 2>/dev/null
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 97e52a249..72ac84d81 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -1227,14 +1227,11 @@ qwen3.5-fp8-mi355x-sglang-agentic-hicache:
       - { tp: 8, ep: 1, kv-offloading: none, conc-list: [1, 2, 4, 8, 16, 32] }
       - { tp: 8, ep: 1, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [16, 32, 48, 64] }
 
-# DeepSeek-V4-Pro FP4 MTP refresh for MI355X. Fast discovery measured the full
-# B200/B300-shaped topology space one variable at a time: TP8 weakly dominates
-# TP4 at low concurrency, TP8+EP8 dominates DP8+EP8, and native DRAM offload at
-# c64 regresses the GPU-only point without producing an external-cache hit. The
-# reusable sweep followed the measured GPU-resident frontier from TP8 c1-c12
-# into TP8+EP8 c16-c64. The current fast follow-up isolates INT4 Quick Reduce
-# at c48; the all-green full grid remains the rollback baseline in run
-# 31357800541.
+# DeepSeek-V4-Pro FP4 MTP collection sweep for MI355X. Keep the verified
+# GPU-resident frontier, add denser c52/c60/c68 TEP guards, collect the same
+# points with pinned Mooncake DRAM offload, and retain representative DEP8 and
+# DEP4 cells even when dominated. The fast pass is a functionality gate;
+# Pareto optimality is deliberately not required for inclusion.
 dsv4-fp4-mi355x-vllm-agentic-mtp:
   image: vllm/vllm-openai-rocm:v0.26.0
   model: deepseek-ai/DeepSeek-V4-Pro
@@ -1247,9 +1244,15 @@ dsv4-fp4-mi355x-vllm-agentic-mtp:
     agentic-coding:
     - dram-utilization: 0.60
       search-space:
-      - { tp: 8, ep: 8, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [48] }
-      # DP4+EP4 is not fit-safe at the required 1,048,576-token model length:
-      # a direct c56 probe left 12.09 GiB KV memory versus 14.88 GiB required.
+      - { tp: 8, ep: 1, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 4, 6, 8, 12] }
+      - { tp: 8, ep: 8, dp-attn: false, kv-offloading: none, spec-decoding: mtp, conc-list: [16, 24, 32, 40, 48, 52, 56, 60, 64, 68] }
+      - { tp: 8, ep: 8, dp-attn: false, kv-offloading: dram, kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }, spec-decoding: mtp, conc-list: [52, 60, 68] }
+      - { tp: 8, ep: 8, dp-attn: true, kv-offloading: none, spec-decoding: mtp, conc-list: [64, 128, 192, 224], router: { name: vllm-router, version: "0.1.14" } }
+      - { tp: 4, ep: 4, dp-attn: true, kv-offloading: none, spec-decoding: mtp, conc-list: [4, 8], router: { name: vllm-router, version: "0.1.14" } }
+      # DEP4 previously had only 12.09 GiB/GPU available for the native
+      # 1,048,576-token context versus a 14.88 GiB minimum. Retain two cells in
+      # the fast gate to reconfirm whether the current base can now initialize;
+      # do not weaken the AgentX context limit to make them pass.
       # LMCache invalid-block recovery currently assumes one KV-cache group,
       # while MTP creates two. Restore these points after the upstream hybrid
       # KV recovery fix lands: https://github.com/vllm-project/vllm/pull/45497
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 34f6c8951..7ae48e4da 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5735,4 +5735,5 @@
     - "Next isolate only INT4 Quick Reduce at the retained c48 baseline: enable VLLM_ROCM_QUICK_REDUCE_QUANTIZATION=INT4 while restoring FULL_DECODE_ONLY and keeping vLLM v0.26.0, TP8+EP8, K=3/golden AL 2.49, max-num-batched-tokens 8192, max-num-seqs 96, FP8 KV, 0.9 memory utilization, and the exact AgentX trace unchanged. This is a standard ROCm vLLM communication optimization already used by other MI355X recipes; the paired eval guards accuracy."
     - "Reject INT4 Quick Reduce after the exact c48 fast run completed successfully at 87,176.57 input / 470.37 output tok/s over 1,228.05 seconds, versus 86,334.85 / 471.67 over 1,229.32 seconds for the duration-matched control: input improved only 0.98%, output regressed 0.28%, and the candidate introduced one empty-content response while the control was error-free. The 0.12% error rate remained below the harness threshold, TTFT/ITL coverage was 100%, GSM8K passed, and vllm: metrics exported, but this is not a clean Pareto improvement, so restore unquantized collectives."
     - "Next isolate only vLLM's graph execution path at the retained c48 baseline: set VLLM_USE_BREAKABLE_CUDAGRAPH=0 while keeping FULL_DECODE_ONLY, vLLM v0.26.0, TP8+EP8, K=3/golden AL 2.49, max-num-batched-tokens 8192, max-num-seqs 96, FP8 KV, 0.9 memory utilization, and the exact AgentX trace unchanged. Current logs show auto-enabled breakable CUDA graphs force CompilationMode.NONE despite compilation mode 3; opting out tests the actual torch.compile path without changing graph capture mode or any workload setting."
+    - "Expand the final collection matrix rather than limiting it to Pareto points: keep the verified TP8 and TEP8 grid, add non-offload TEP8 c52/c60/c68 around the measured c56 knee, pair those three cells with pinned Mooncake 0.3.11.post1 DRAM offload under the generated aggregate host-memory cap, restore representative DEP8 c64/c128/c192/c224, and probe DEP4 c4/c8 without reducing the native AgentX context limit. Run agentx-fast first as a functionality gate, then collect every functional cell in the reusable full-duration sweep even when dominated."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2540

From f894cc8fb2b7efef4b1155fae7a2d922d58f93ca Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 10 Aug 2026 13:02:31 -0500
Subject: [PATCH 16/18] fix(mi355x): bound Mooncake transfer logging

---
 benchmarks/single_node/agentic/dsv4_fp4_mi355x_vllm_mtp.sh | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_vllm_mtp.sh b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_vllm_mtp.sh
index 187f003c8..707c67114 100644
--- a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_vllm_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_vllm_mtp.sh
@@ -242,6 +242,9 @@ case "${KV_OFFLOAD_BACKEND:-}" in
   "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
         export MC_ENABLE_DEST_DEVICE_AFFINITY=1
         export PYTHONHASHSEED=0

From 1b48a1a1436eaa4040b3c651978cb6037d889add Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 10 Aug 2026 13:38:07 -0500
Subject: [PATCH 17/18] fix(mi355x): fit native context on DEP4

---
 .../single_node/agentic/dsv4_fp4_mi355x_vllm_mtp.sh  | 12 +++++++++++-
 perf-changelog.yaml                                  |  1 +
 2 files changed, 12 insertions(+), 1 deletion(-)

diff --git a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_vllm_mtp.sh b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_vllm_mtp.sh
index 707c67114..c5f0a95e7 100644
--- a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_vllm_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_vllm_mtp.sh
@@ -426,6 +426,16 @@ else
     SPEC_CONFIG="{\"method\": \"mtp\", \"num_speculative_tokens\": $NUM_SPEC_TOKENS, \"rejection_sample_method\": \"synthetic\", \"synthetic_acceptance_length\": $SYNTHETIC_ACCEPT_LEN}"
 fi
 
+# DEP4 replicates the attention weights across four DP ranks. At the common
+# 0.90 setting vLLM leaves only 12.06--12.50 GiB for KV on MI355X, below the
+# 14.88 GiB required to admit one native 1,048,576-token request. Give only
+# this topology the extra headroom it needs; established TP/TEP/DEP8 results
+# retain their original memory fraction.
+GPU_MEM_UTIL=0.9
+if [ "$DP_ATTENTION" = "true" ] && [ "$TP" -eq 4 ]; then
+    GPU_MEM_UTIL=0.92
+fi
+
 echo "Starting vllm server..."
 set -x
 export VLLM_ROCM_USE_AITER=1
@@ -442,7 +452,7 @@ VLLM_CMD=(
     --kv-cache-dtype fp8
     "${PARALLEL_ARGS[@]}"
     "${EP_ARGS[@]}"
-    --gpu-memory-utilization 0.9
+    --gpu-memory-utilization "$GPU_MEM_UTIL"
     --block-size 256
     --max-num-batched-tokens 8192
     --moe-backend aiter
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 7ae48e4da..46dde7793 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5736,4 +5736,5 @@
     - "Reject INT4 Quick Reduce after the exact c48 fast run completed successfully at 87,176.57 input / 470.37 output tok/s over 1,228.05 seconds, versus 86,334.85 / 471.67 over 1,229.32 seconds for the duration-matched control: input improved only 0.98%, output regressed 0.28%, and the candidate introduced one empty-content response while the control was error-free. The 0.12% error rate remained below the harness threshold, TTFT/ITL coverage was 100%, GSM8K passed, and vllm: metrics exported, but this is not a clean Pareto improvement, so restore unquantized collectives."
     - "Next isolate only vLLM's graph execution path at the retained c48 baseline: set VLLM_USE_BREAKABLE_CUDAGRAPH=0 while keeping FULL_DECODE_ONLY, vLLM v0.26.0, TP8+EP8, K=3/golden AL 2.49, max-num-batched-tokens 8192, max-num-seqs 96, FP8 KV, 0.9 memory utilization, and the exact AgentX trace unchanged. Current logs show auto-enabled breakable CUDA graphs force CompilationMode.NONE despite compilation mode 3; opting out tests the actual torch.compile path without changing graph capture mode or any workload setting."
     - "Expand the final collection matrix rather than limiting it to Pareto points: keep the verified TP8 and TEP8 grid, add non-offload TEP8 c52/c60/c68 around the measured c56 knee, pair those three cells with pinned Mooncake 0.3.11.post1 DRAM offload under the generated aggregate host-memory cap, restore representative DEP8 c64/c128/c192/c224, and probe DEP4 c4/c8 without reducing the native AgentX context limit. Run agentx-fast first as a functionality gate, then collect every functional cell in the reusable full-duration sweep even when dominated."
+    - "Use 0.92 GPU-memory utilization only for DEP4: the isolated native-context c4/c8 probe measured 12.06--12.50 GiB of free KV at 0.90 versus the 14.88 GiB required for one 1,048,576-token request. Keep 0.90 unchanged for established TP8, TEP8, and DEP8 cells."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2540

From 42fc55445ba368aa661f0c88ccf48abdc83a70e8 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 10 Aug 2026 23:18:17 -0500
Subject: [PATCH 18/18] Simplify MI355X performance changelog

---
 perf-changelog.yaml | 27 ++++-----------------------
 1 file changed, 4 insertions(+), 23 deletions(-)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index ca5307664..ea1562333 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5731,29 +5731,10 @@
   scenario-type:
     - agentic-coding
   description:
-    - "Refresh the MI355X DeepSeek-V4-Pro FP4 vLLM AgentX MTP frontier after the prior eight-point run peaked near 66k total tok/s and its only DP-attention point used EP1/full-model replication."
-    - "Use the latest stable ROCm vLLM v0.26.0 release, raise GPU-memory utilization from 0.8 to 0.9, make the official 256-token block and 8192-token scheduler budget explicit, and use FULL_DECODE_ONLY graph capture. Baseline logs showed roughly 44 GiB/GPU of unused fit-safe KV headroom at 0.8."
-    - "Replace the isolated DP8/EP1 c64 point with 40 points spanning TP8, fit-validated TP4, TP8+EP8 and DP8+EP8. Mirror the B200/B300 MTP low-concurrency cells and extend DP8+EP8 through concurrency 224, with extra MI355X intermediates to resolve the local knee. DP8+EP8 is the intended throughput topology: attention remains data parallel while the FP4 experts are sharded across all eight GPUs."
-    - "Direct fit validation shows TP4 leaves 37.47 GiB/GPU for a 2.82M-token KV cache and serves normally. DP4+EP4 is excluded because it leaves only 12.09 GiB/GPU KV versus 14.88 GiB required for the committed 1,048,576-token model length; the workload is not weakened to force an invalid topology."
-    - "Keep MTP and its committed three-token golden synthetic acceptance length of 2.49 on every throughput point; eval-only runs retain real target verification. Enable prompt-token details for request-level cache accounting, and require a nonempty vllm: server-metrics export from AIPerf."
-    - "Use agentx-fast discovery before the reusable sweep: 25 completed throughput points show DP8+EP8 is uniformly dominated by TP8+EP8, while TP4 collapses above concurrency 12. Narrow the second fast pass to the six unresolved TP8 c4/c6/c12, TP4 c4/c6, and TP8+EP8 c56 cells; the final full-duration grid will contain only measured frontier points and adjacent guards."
-    - "The second fast pass confirms TP8 weakly dominates TP4 at matched c4/c6. TP8+EP8 c56 reaches 89.22k input and 496.46 output tok/s with 52.9% KV use, but p95 TTFT has risen to 13.29s. Probe the remaining upper band at c64/c72/c80/c88/c96/c104/c112: c96 is the linear KV-capacity estimate from c48, while c104/c112 bracket the expected cache and latency cliff."
-    - "The GPU-only upper-band pass completes on all seven points and confirms a monotonic post-c56 collapse: c64/c72/c80 reach 79.34k/414.46, 72.17k/306.93, and 59.06k/284.48 input/output tok/s; c88/c96/c104/c112 fall to 34.33k/142.61, 24.39k/117.42, 15.24k/79.54, and 13.28k/70.73. All points have zero request errors; c88/c96/c112 miss only the fast-window completed-record TTFT coverage gate."
-    - "Test one isolated follow-up variable before finalizing: TP8+EP8 c64 with the script's supported vLLM-native DRAM offload, leaving vLLM v0.26.0, MTP3, golden AL 2.49, FP8 KV, topology, concurrency, memory utilization, and AgentX trace unchanged. Keep GPU-only c56/c64 as the preserved baseline and expand offload only if c64 improves."
-    - "Reject native DRAM offload after the isolated c64 test: it completes 719 valid requests with zero request errors but reaches only 73.45k input and 366.17 output tok/s versus the preserved GPU-only c64 result at 79.34k/414.46, while reporting zero external-cache hits despite roughly 20 TB of cumulative host-KV writes. The fast-window TTFT coverage miss (92.8% with ITL at 100%) is consistent with the slower long-response tail and does not change the throughput regression."
-    - "Run the reusable one-hour sweep on the measured GPU-resident Pareto grid only: TP8 c1/c2/c4/c6/c8/c12 for the low-latency band, then TP8+EP8 c16/c24/c32/c40/c48/c56/c64 for the throughput band. This preserves the B200/B300 low-concurrency spacing, switches topology at the measured crossover, includes every sampled rise to the c56 knee, and retains c64 as the adjacent post-knee guard without re-running the dominated TP4, DP8+EP8, offload, or c72+ tails."
-    - "The reusable GPU-only grid completed green in run 31357800541. Isolate one follow-up variable at its input-throughput peak, TP8+EP8 c48: raise max-num-batched-tokens from 8192 to 16384 while keeping vLLM v0.26.0, topology, concurrency, MTP3/golden AL 2.49, FP8 KV, GPU memory utilization, and AgentX trace unchanged. Use agentx-fast and preserve the full-grid run as the rollback baseline."
-    - "Reject the 16384-token scheduler budget after its exact c48 fast comparison completed fully green but regressed to 80,316.53 input / 412.36 output tok/s over 1,228.83 seconds, versus 86,334.85 / 471.67 over 1,229.32 seconds for the duration-matched 8192-token control from run 31341602956. Restore 8192 before any further experiment."
-    - "Next isolate only the golden MTP operating point at c48: switch from K=3 with its measured synthetic acceptance length 2.49 to K=2 with its measured synthetic acceptance length 2.27, while retaining vLLM v0.26.0, TP8+EP8, max-num-seqs 96, FP8 KV, 0.9 memory utilization, FULL_DECODE_ONLY graphs, and the exact AgentX trace. Draft depth and its matching golden acceptance length are one coupled speculative configuration."
-    - "Reject K=2 after the exact c48 fast run completed fully green at 83,283.76 input / 448.52 output tok/s over 1,226.65 seconds, versus 86,334.85 / 471.67 over 1,229.32 seconds for the duration-matched K=3 control: input regressed 3.53% and output regressed 4.91%. Restore K=3 with golden acceptance length 2.49 before any further experiment."
-    - "Next isolate only graph capture mode at c48: switch FULL_DECODE_ONLY to FULL_AND_PIECEWISE while keeping vLLM v0.26.0, TP8+EP8, K=3/golden AL 2.49, max-num-batched-tokens 8192, max-num-seqs 96, FP8 KV, 0.9 memory utilization, and the exact AgentX trace unchanged."
-    - "Reject FULL_AND_PIECEWISE after the exact c48 fast run completed fully green at 83,955.53 input / 455.90 output tok/s over 1,228.87 seconds, versus 86,334.85 / 471.67 over 1,229.32 seconds for the duration-matched FULL_DECODE_ONLY control: input regressed 2.76% and output regressed 3.34%. The run processed 795 valid requests with zero request errors, passed TTFT/ITL coverage, and exported nonempty vllm: metrics, so this is a measured graph-mode regression rather than infrastructure noise."
-    - "Next isolate only INT4 Quick Reduce at the retained c48 baseline: enable VLLM_ROCM_QUICK_REDUCE_QUANTIZATION=INT4 while restoring FULL_DECODE_ONLY and keeping vLLM v0.26.0, TP8+EP8, K=3/golden AL 2.49, max-num-batched-tokens 8192, max-num-seqs 96, FP8 KV, 0.9 memory utilization, and the exact AgentX trace unchanged. This is a standard ROCm vLLM communication optimization already used by other MI355X recipes; the paired eval guards accuracy."
-    - "Reject INT4 Quick Reduce after the exact c48 fast run completed successfully at 87,176.57 input / 470.37 output tok/s over 1,228.05 seconds, versus 86,334.85 / 471.67 over 1,229.32 seconds for the duration-matched control: input improved only 0.98%, output regressed 0.28%, and the candidate introduced one empty-content response while the control was error-free. The 0.12% error rate remained below the harness threshold, TTFT/ITL coverage was 100%, GSM8K passed, and vllm: metrics exported, but this is not a clean Pareto improvement, so restore unquantized collectives."
-    - "Next isolate only vLLM's graph execution path at the retained c48 baseline: set VLLM_USE_BREAKABLE_CUDAGRAPH=0 while keeping FULL_DECODE_ONLY, vLLM v0.26.0, TP8+EP8, K=3/golden AL 2.49, max-num-batched-tokens 8192, max-num-seqs 96, FP8 KV, 0.9 memory utilization, and the exact AgentX trace unchanged. Current logs show auto-enabled breakable CUDA graphs force CompilationMode.NONE despite compilation mode 3; opting out tests the actual torch.compile path without changing graph capture mode or any workload setting."
-    - "Expand the final collection matrix rather than limiting it to Pareto points: keep the verified TP8 and TEP8 grid, add non-offload TEP8 c52/c60/c68 around the measured c56 knee, pair those three cells with pinned Mooncake 0.3.11.post1 DRAM offload under the generated aggregate host-memory cap, restore representative DEP8 c64/c128/c192/c224, and probe DEP4 c4/c8 without reducing the native AgentX context limit. Run agentx-fast first as a functionality gate, then collect every functional cell in the reusable full-duration sweep even when dominated."
-    - "Use 0.92 GPU-memory utilization only for DEP4: the isolated native-context c4/c8 probe measured 12.06--12.50 GiB of free KV at 0.90 versus the 14.88 GiB required for one 1,048,576-token request. Keep 0.90 unchanged for established TP8, TEP8, and DEP8 cells."
-    - "Upgrade the final 26-point collection to the latest stable vLLM ROCm v0.27.0 release after the isolated Mooncake c52 functionality gate completed 575/575 warmup requests and 248/248 measured requests with zero errors, live external-cache telemetry, and nonempty vllm: metrics."
+    - "Refresh the MI355X DeepSeek-V4-Pro MTP submission on stable vLLM ROCm v0.27.0; raise GPU-memory utilization from 0.80 to 0.90 and use 256-token blocks, an 8192-token scheduler budget, and FULL_DECODE_ONLY graphs."
+    - "Keep the measured MTP operating point at three draft tokens with golden synthetic acceptance length 2.49."
+    - "Expand collection to TP8 c1-c12, TP8/EP8 c16-c68, Mooncake DRAM offload c52/c60/c68, DEP8 c64-c224, and DEP4 c4/c8, including dominated points requested for comparison."
+    - "Use 0.92 GPU-memory utilization only for DEP4 so the native 1,048,576-token context fits; all other points retain 0.90."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2540
 
 - config-keys: