From 207b090dd2cc7260c76e064015d187660ea86665 Mon Sep 17 00:00:00 2001
From: Fangzhou Ai <31551580+Fangzhou-Ai@users.noreply.github.com>
Date: Sat, 12 Sep 2026 01:49:47 +0000
Subject: [PATCH 1/7] config: switch MI355X DSv4.1 Flash to the ROCm nightly
 image

The deepseekv41-flash-0909 tag predates vllm-project/vllm#56503, which
moves the mHC delayed pre block off the eager Torch reference and onto
AITER. That block is 85% of the decoder's kernel launches for this model,
so the release tag leaves most of the decode cost unaddressed.

Tag is nightly-TBD until the first nightly containing that commit is
published; it is then pinned to nightly-<sha> like the other ROCm entries.
---
 .../single_node/agentic/dsv41flash_fp4_mi355x_vllm_mtp.sh   | 2 +-
 configs/amd-master.yaml                                     | 6 ++++--
 docs/configuration-procedures.md                            | 2 +-
 docs/configuration-procedures_zh.md                         | 2 +-
 4 files changed, 7 insertions(+), 5 deletions(-)

diff --git a/benchmarks/single_node/agentic/dsv41flash_fp4_mi355x_vllm_mtp.sh b/benchmarks/single_node/agentic/dsv41flash_fp4_mi355x_vllm_mtp.sh
index 40ec8d58a..0a9fc85d0 100755
--- a/benchmarks/single_node/agentic/dsv41flash_fp4_mi355x_vllm_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv41flash_fp4_mi355x_vllm_mtp.sh
@@ -3,7 +3,7 @@ set -eo pipefail
 
 # DeepSeek-V4.1-Flash on MI355X: native DSpark and GPU-resident KV.
 # Follow upstream AMD defaults for Engram; storage behavior needs verification.
-# Image: vllm/vllm-openai-rocm:deepseekv41-flash-0909 (configured in amd-master.yaml); GPU validation is pending.
+# Image: vllm/vllm-openai-rocm:nightly-TBD (configured in amd-master.yaml); GPU validation is pending.
 # https://github.com/vllm-project/recipes/blob/main/models/deepseek-ai/DeepSeek-V4.1-Flash.yaml
 source "$(dirname "$0")/../../benchmark_lib.sh"
 check_env_vars MODEL TP CONC KV_OFFLOADING TOTAL_CPU_DRAM_GB RESULT_DIR DURATION
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 3d9a3dcca..3d429a7aa 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -1753,9 +1753,11 @@ dsv4-fp4-mi355x-sglang-agentic-mtp:
       - { tp: 8, ep: 1, dp-attn: false, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [32, 48], spec-decoding: mtp }
       - { tp: 8, ep: 1, dp-attn: true, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [128, 256], spec-decoding: mtp }
 
-# Official upstream ROCm image from the vLLM recipe; MI355X runtime validation is pending.
+# Upstream ROCm nightly rather than the deepseekv41-flash-0909 release tag: that
+# tag predates vllm-project/vllm#56503, which moves the mHC delayed pre block off
+# the eager Torch reference and onto AITER. MI355X runtime validation is pending.
 dsv41flash-fp4-mi355x-vllm-agentic-dspark:
-  image: vllm/vllm-openai-rocm:deepseekv41-flash-0909
+  image: vllm/vllm-openai-rocm:nightly-TBD
   model: deepseek-ai/DeepSeek-V4.1-Flash
   model-prefix: dsv41flash
   runner: cluster:mi355x-amds
diff --git a/docs/configuration-procedures.md b/docs/configuration-procedures.md
index d853f6493..977a02ebc 100644
--- a/docs/configuration-procedures.md
+++ b/docs/configuration-procedures.md
@@ -403,4 +403,4 @@ The draft `dsv41flash-fp4-mi355x-vllm-agentic-dspark` recipe extends [#2958](htt
 
 Follow the AMD overrides in [upstream recipe #946](https://github.com/vllm-project/recipes/pull/946): `VLLM_ROCM_USE_AITER=1`, `VLLM_ROCM_USE_AITER_MOE=1`, and `--moe-backend aiter_triton_mxfp4_bf16`. The recipe pins `semianalysis_cc_traces_weka_062126` (the unfiltered corpus) via `WEKA_LOADER_OVERRIDE`. KV stays GPU-resident; Engram follows upstream AMD defaults. Do not copy the NVIDIA `--engram-config` option: upstream currently rejects it on ROCm. The MI355X launcher uses the shared HF cache and mounts this model's repository at `/ix`, and exports `INFMAX_CONTAINER_WORKSPACE=/ix` so AgentX dependencies and outputs resolve inside that mount.
 
-**GPU validation is pending:** The recipe uses `vllm/vllm-openai-rocm:deepseekv41-flash-0909`, the image referenced by the [upstream vLLM recipe](https://recipes.vllm.ai/deepseek-ai/DeepSeek-V4.1-Flash?hardware=mi355x). The tag was published to Docker Hub on 2026-09-11 after AMD verification; it returned HTTP 404 in run 34466680355 before publication. Engram behavior and serving compatibility still require GPU validation. Follow the [AgentX procedure](./eval-agentx-procedures.md#7-run-agentx-fast-feedback-versus-canonical-evidence) for runtime evidence; local generation and registry metadata are not GPU proof.
+**GPU validation is pending:** The recipe uses `vllm/vllm-openai-rocm:nightly-TBD`. The `deepseekv41-flash-0909` tag referenced by the [upstream vLLM recipe](https://recipes.vllm.ai/deepseek-ai/DeepSeek-V4.1-Flash?hardware=mi355x) predates [vllm-project/vllm#56503](https://github.com/vllm-project/vllm/pull/56503), which moves the mHC delayed pre block off the eager Torch reference and onto AITER; that block is 85% of the decoder's kernel launches on this model, so the release tag leaves most of the decode cost on the table. The nightly tag is a placeholder until the first nightly containing that commit is published, at which point it is pinned to `nightly-<sha>` in line with the other ROCm entries. Engram behavior and serving compatibility still require GPU validation. Follow the [AgentX procedure](./eval-agentx-procedures.md#7-run-agentx-fast-feedback-versus-canonical-evidence) for runtime evidence; local generation and registry metadata are not GPU proof.
diff --git a/docs/configuration-procedures_zh.md b/docs/configuration-procedures_zh.md
index 883f1f8e2..fdac9ccad 100644
--- a/docs/configuration-procedures_zh.md
+++ b/docs/configuration-procedures_zh.md
@@ -381,4 +381,4 @@ python -m pytest utils/matrix_logic/ -v
 
 遵循[上游配方 #946](https://github.com/vllm-project/recipes/pull/946) 的 AMD 设置：`VLLM_ROCM_USE_AITER=1`、`VLLM_ROCM_USE_AITER_MOE=1` 和 `--moe-backend aiter_triton_mxfp4_bf16`。配方通过 `WEKA_LOADER_OVERRIDE` 固定使用完整语料 `semianalysis_cc_traces_weka_062126`。KV 驻留 GPU；Engram 沿用上游 AMD 默认设置。不要复制 NVIDIA 的 `--engram-config` 选项：上游目前在 ROCm 上拒绝该选项。MI355X launcher 使用共享 HF 缓存，并将此模型的仓库挂载至 `/ix`，同时导出 `INFMAX_CONTAINER_WORKSPACE=/ix`，确保 AgentX 依赖与输出路径位于该挂载中。
 
-**GPU 验证尚待完成：** 配方使用 `vllm/vllm-openai-rocm:deepseekv41-flash-0909`，即[上游 vLLM 配方](https://recipes.vllm.ai/deepseek-ai/DeepSeek-V4.1-Flash?hardware=mi355x)引用的镜像。该标签经 AMD 验证后于 2026-09-11 发布到 Docker Hub；发布前在运行 34466680355 中返回 HTTP 404。Engram 行为与服务兼容性仍需 GPU 验证。请按 [AgentX 流程](./eval-agentx-procedures_zh.md) 获取运行时证据；本地矩阵生成和镜像元数据不等于 GPU 验证。
+**GPU 验证尚待完成：** 配方使用 `vllm/vllm-openai-rocm:nightly-TBD`。[上游 vLLM 配方](https://recipes.vllm.ai/deepseek-ai/DeepSeek-V4.1-Flash?hardware=mi355x)引用的 `deepseekv41-flash-0909` 标签早于 [vllm-project/vllm#56503](https://github.com/vllm-project/vllm/pull/56503)，该 PR 将 mHC delayed pre 块从 eager Torch 参考实现切换到 AITER；该块占此模型解码器内核启动数的 85%，因此发布标签会浪费大部分解码开销。nightly 标签目前为占位符，待包含该提交的首个 nightly 发布后，将按其他 ROCm 条目的惯例固定为 `nightly-<sha>`。Engram 行为与服务兼容性仍需 GPU 验证。请按 [AgentX 流程](./eval-agentx-procedures_zh.md) 获取运行时证据；本地矩阵生成和镜像元数据不等于 GPU 验证。

From 68064a0565aeb5ed0f35c394805fdbc1390dbb98 Mon Sep 17 00:00:00 2001
From: Fangzhou Ai <31551580+Fangzhou-Ai@users.noreply.github.com>
Date: Sat, 12 Sep 2026 01:50:41 +0000
Subject: [PATCH 2/7] changelog: record the MI355X nightly image switch

---
 perf-changelog.yaml | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 1f90deb80..b0908bed7 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -7408,3 +7408,10 @@
     - "Enable required power for the seven existing B200 Kimi-K3 recipes, use the pinned AgentX producer and shared collector, fix the exporter import URI, and record DCP8 with offload disabled to match the serving commands."
     - "启用 Kimi-K3 B200 实测功耗并修正 DCP 与 Offload 元数据。"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3043
+
+- config-keys:
+    - dsv41flash-fp4-mi355x-vllm-agentic-dspark
+  description:
+    - "Switch to the upstream ROCm nightly image. The deepseekv41-flash-0909 tag predates vllm-project/vllm#56503, which moves the mHC delayed pre block off the eager Torch reference onto AITER; that block is 85% of the decoder's kernel launches, measured 15.9-20.9x faster under cudagraph with gsm8k unchanged at 0.9719. Tag is nightly-TBD until the first nightly containing the commit is published."
+    - "切换到上游 ROCm nightly 镜像。deepseekv41-flash-0909 标签早于 vllm-project/vllm#56503，该 PR 将 mHC delayed pre 块从 eager Torch 参考实现切换到 AITER；该块占解码器内核启动数的 85%，在 cudagraph 下实测加速 15.9-20.9 倍，gsm8k 保持 0.9719 不变。标签暂为 nightly-TBD，待包含该提交的首个 nightly 发布后固定。"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3035

From 35dc4c98d64a1fe8ea2148617bd0fed3feedeff3 Mon Sep 17 00:00:00 2001
From: Fangzhou Ai <31551580+Fangzhou-Ai@users.noreply.github.com>
Date: Sat, 12 Sep 2026 03:32:19 +0000
Subject: [PATCH 3/7] config: silence the AITER Gluon-fallback warning on
 MI355X DSv4.1 Flash

AITER's Triton MoE GEMM logs "GLUON backend not available. Using TRITON
backend!!!" on every call. Gluon supports only gfx1250, so on gfx950 the
message reports a fixed property of the hardware rather than a condition,
and it dominates the server log: 410,984 of 417,504 lines in a gsm8k run,
29.8 MiB of 32 MiB.

Set AITER_TRITON_LOG_LEVEL=ERROR, the knob aiter.ops.triton.utils.logger
reads. Every warning that module emits is about Gluon availability -- the
two Gluon GEMM depth fallbacks are inside gfx1250-only paths, and
fused_mxfp4_quant warns only when Gluon is requested explicitly -- so
raising the threshold hides nothing actionable on this architecture.

Log hygiene rather than performance: the emits run at 61/s per worker and
4.33 us each, or 0.026% of wall time.

Signed-off-by: Fangzhou Ai <fangzhou@semianalysis.ai>
Co-authored-by: Cursor <cursoragent@cursor.com>
---
 .../single_node/agentic/dsv41flash_fp4_mi355x_vllm_mtp.sh | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/benchmarks/single_node/agentic/dsv41flash_fp4_mi355x_vllm_mtp.sh b/benchmarks/single_node/agentic/dsv41flash_fp4_mi355x_vllm_mtp.sh
index 0a9fc85d0..ad4297d41 100755
--- a/benchmarks/single_node/agentic/dsv41flash_fp4_mi355x_vllm_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv41flash_fp4_mi355x_vllm_mtp.sh
@@ -23,6 +23,14 @@ if [[ -n "${ROCR_VISIBLE_DEVICES:-}" ]]; then
 fi
 export VLLM_ROCM_USE_AITER=1
 export VLLM_ROCM_USE_AITER_MOE=1
+# AITER's Triton MoE GEMM warns on every call that Gluon is unavailable and it
+# is falling back to Triton. Gluon supports only gfx1250, so on gfx950 that is
+# a fixed property rather than a condition worth reporting, and it was 98% of
+# the lines in a gsm8k server log (411k of 417k, 30 MiB of 32 MiB). Every
+# warning aiter.ops.triton emits is about Gluon availability, so raising the
+# threshold loses nothing actionable here. Log hygiene only: the emits cost
+# 0.03% of wall time per worker, so this is not a throughput change.
+export AITER_TRITON_LOG_LEVEL=ERROR
 # DeepseekV41ForCausalLM is not torch-compiled upstream, so the default
 # cudagraph_mode=FULL_AND_PIECEWISE aborts at engine init with "piecewise CUDA
 # graphs unavailable" (run 34566727564). The model is built for the breakable

From 737210a4eee1e102fad859817c316f24be55a12a Mon Sep 17 00:00:00 2001
From: Fangzhou Ai <31551580+Fangzhou-Ai@users.noreply.github.com>
Date: Sat, 12 Sep 2026 04:43:23 +0000
Subject: [PATCH 4/7] config: raise MI355X max_num_seqs to the upstream default

The 2*CONC cap sat below AgentX's subagent fan-out. At CONC=1 the engine
admitted 2 requests and queued the rest on scheduling capacity, with vLLM
reporting queue times up to 16.2 s. Removing the cap takes queue time to
zero at identical output.
---
 .../single_node/agentic/dsv41flash_fp4_mi355x_vllm_mtp.sh  | 7 +++++--
 perf-changelog.yaml                                        | 7 +++++++
 2 files changed, 12 insertions(+), 2 deletions(-)

diff --git a/benchmarks/single_node/agentic/dsv41flash_fp4_mi355x_vllm_mtp.sh b/benchmarks/single_node/agentic/dsv41flash_fp4_mi355x_vllm_mtp.sh
index ad4297d41..8ae683c01 100755
--- a/benchmarks/single_node/agentic/dsv41flash_fp4_mi355x_vllm_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv41flash_fp4_mi355x_vllm_mtp.sh
@@ -47,8 +47,11 @@ export VLLM_ENGINE_READY_TIMEOUT_S=3600
 export VLLM_USE_RUST_FRONTEND=1
 export PYTHONUNBUFFERED=1
 
-# Match the sibling's scheduler headroom for AgentX subagent fan-out.
-MAX_NUM_SEQS=$((2 * CONC))
+# Upstream default. The previous 2*CONC cap sat below AgentX's subagent
+# fan-out, so at CONC=1 the engine admitted 2 requests and left the rest
+# queued on scheduling capacity. Pinned rather than inherited so CAPTURE_SIZE
+# below stays consistent with it.
+MAX_NUM_SEQS=128
 NUM_SPEC_TOKENS=5
 CAPTURE_SIZE=1
 while (( CAPTURE_SIZE < MAX_NUM_SEQS * (1 + NUM_SPEC_TOKENS) && CAPTURE_SIZE < 2048 )); do
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index b0908bed7..a3b258351 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -7415,3 +7415,10 @@
     - "Switch to the upstream ROCm nightly image. The deepseekv41-flash-0909 tag predates vllm-project/vllm#56503, which moves the mHC delayed pre block off the eager Torch reference onto AITER; that block is 85% of the decoder's kernel launches, measured 15.9-20.9x faster under cudagraph with gsm8k unchanged at 0.9719. Tag is nightly-TBD until the first nightly containing the commit is published."
     - "切换到上游 ROCm nightly 镜像。deepseekv41-flash-0909 标签早于 vllm-project/vllm#56503，该 PR 将 mHC delayed pre 块从 eager Torch 参考实现切换到 AITER；该块占解码器内核启动数的 85%，在 cudagraph 下实测加速 15.9-20.9 倍，gsm8k 保持 0.9719 不变。标签暂为 nightly-TBD，待包含该提交的首个 nightly 发布后固定。"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3035
+
+- config-keys:
+    - dsv41flash-fp4-mi355x-vllm-agentic-dspark
+  description:
+    - "Raise max_num_seqs to the upstream default of 128. The 2*CONC cap sat below AgentX's subagent fan-out, so requests queued on scheduling capacity; vLLM reported queue times up to 16.2 s at CONC=1, and removing the cap took them to zero."
+    - "将 max_num_seqs 提升到上游默认值 128。此前的 2*CONC 上限低于 AgentX 的子代理扇出，导致请求因调度容量排队；vLLM 在 CONC=1 时报告排队时间最高达 16.2 秒，取消该上限后归零。"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3035

From 3af2d0a75c689100dd9a81bc01888342afd9bc2a Mon Sep 17 00:00:00 2001
From: Fangzhou Ai <31551580+Fangzhou-Ai@users.noreply.github.com>
Date: Sat, 12 Sep 2026 05:50:38 +0000
Subject: [PATCH 5/7] config: pin the MI355X nightly image to the published tag

The nightly containing vllm-project/vllm#56503 is out:
nightly-eed1f3d0c6043bd494424a22443ee198dd56f657
(sha256:960228cf..., published 2026-09-12). Replace the nightly-TBD
placeholder in the config, recipe header, and both docs, and record the
upstream recipe PR that moves vllm-project/recipes to the same image.

Co-authored-by: Cursor <cursoragent@cursor.com>
---
 .../single_node/agentic/dsv41flash_fp4_mi355x_vllm_mtp.sh     | 3 ++-
 configs/amd-master.yaml                                       | 2 +-
 docs/configuration-procedures.md                              | 2 +-
 docs/configuration-procedures_zh.md                           | 2 +-
 perf-changelog.yaml                                           | 4 ++--
 5 files changed, 7 insertions(+), 6 deletions(-)

diff --git a/benchmarks/single_node/agentic/dsv41flash_fp4_mi355x_vllm_mtp.sh b/benchmarks/single_node/agentic/dsv41flash_fp4_mi355x_vllm_mtp.sh
index 8ae683c01..fa2ded18a 100755
--- a/benchmarks/single_node/agentic/dsv41flash_fp4_mi355x_vllm_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv41flash_fp4_mi355x_vllm_mtp.sh
@@ -3,7 +3,8 @@ set -eo pipefail
 
 # DeepSeek-V4.1-Flash on MI355X: native DSpark and GPU-resident KV.
 # Follow upstream AMD defaults for Engram; storage behavior needs verification.
-# Image: vllm/vllm-openai-rocm:nightly-TBD (configured in amd-master.yaml); GPU validation is pending.
+# Image: vllm/vllm-openai-rocm:nightly-eed1f3d0c6043bd494424a22443ee198dd56f657
+# (configured in amd-master.yaml); GPU validation is pending.
 # https://github.com/vllm-project/recipes/blob/main/models/deepseek-ai/DeepSeek-V4.1-Flash.yaml
 source "$(dirname "$0")/../../benchmark_lib.sh"
 check_env_vars MODEL TP CONC KV_OFFLOADING TOTAL_CPU_DRAM_GB RESULT_DIR DURATION
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 3d429a7aa..2e061d32f 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -1757,7 +1757,7 @@ dsv4-fp4-mi355x-sglang-agentic-mtp:
 # tag predates vllm-project/vllm#56503, which moves the mHC delayed pre block off
 # the eager Torch reference and onto AITER. MI355X runtime validation is pending.
 dsv41flash-fp4-mi355x-vllm-agentic-dspark:
-  image: vllm/vllm-openai-rocm:nightly-TBD
+  image: vllm/vllm-openai-rocm:nightly-eed1f3d0c6043bd494424a22443ee198dd56f657
   model: deepseek-ai/DeepSeek-V4.1-Flash
   model-prefix: dsv41flash
   runner: cluster:mi355x-amds
diff --git a/docs/configuration-procedures.md b/docs/configuration-procedures.md
index 977a02ebc..98e265a58 100644
--- a/docs/configuration-procedures.md
+++ b/docs/configuration-procedures.md
@@ -403,4 +403,4 @@ The draft `dsv41flash-fp4-mi355x-vllm-agentic-dspark` recipe extends [#2958](htt
 
 Follow the AMD overrides in [upstream recipe #946](https://github.com/vllm-project/recipes/pull/946): `VLLM_ROCM_USE_AITER=1`, `VLLM_ROCM_USE_AITER_MOE=1`, and `--moe-backend aiter_triton_mxfp4_bf16`. The recipe pins `semianalysis_cc_traces_weka_062126` (the unfiltered corpus) via `WEKA_LOADER_OVERRIDE`. KV stays GPU-resident; Engram follows upstream AMD defaults. Do not copy the NVIDIA `--engram-config` option: upstream currently rejects it on ROCm. The MI355X launcher uses the shared HF cache and mounts this model's repository at `/ix`, and exports `INFMAX_CONTAINER_WORKSPACE=/ix` so AgentX dependencies and outputs resolve inside that mount.
 
-**GPU validation is pending:** The recipe uses `vllm/vllm-openai-rocm:nightly-TBD`. The `deepseekv41-flash-0909` tag referenced by the [upstream vLLM recipe](https://recipes.vllm.ai/deepseek-ai/DeepSeek-V4.1-Flash?hardware=mi355x) predates [vllm-project/vllm#56503](https://github.com/vllm-project/vllm/pull/56503), which moves the mHC delayed pre block off the eager Torch reference and onto AITER; that block is 85% of the decoder's kernel launches on this model, so the release tag leaves most of the decode cost on the table. The nightly tag is a placeholder until the first nightly containing that commit is published, at which point it is pinned to `nightly-<sha>` in line with the other ROCm entries. Engram behavior and serving compatibility still require GPU validation. Follow the [AgentX procedure](./eval-agentx-procedures.md#7-run-agentx-fast-feedback-versus-canonical-evidence) for runtime evidence; local generation and registry metadata are not GPU proof.
+**GPU validation is pending:** The recipe uses `vllm/vllm-openai-rocm:nightly-eed1f3d0c6043bd494424a22443ee198dd56f657` (digest `sha256:960228cf…`, published 2026-09-12). The `deepseekv41-flash-0909` tag referenced by the [upstream vLLM recipe](https://recipes.vllm.ai/deepseek-ai/DeepSeek-V4.1-Flash?hardware=mi355x) predates [vllm-project/vllm#56503](https://github.com/vllm-project/vllm/pull/56503), which moves the mHC delayed pre block off the eager Torch reference and onto AITER; that block is 85% of the decoder's kernel launches on this model, so the release tag leaves most of the decode cost on the table. The upstream recipe is being moved to the same nightly in [vllm-project/recipes#962](https://github.com/vllm-project/recipes/pull/962). Engram behavior and serving compatibility still require GPU validation. Follow the [AgentX procedure](./eval-agentx-procedures.md#7-run-agentx-fast-feedback-versus-canonical-evidence) for runtime evidence; local generation and registry metadata are not GPU proof.
diff --git a/docs/configuration-procedures_zh.md b/docs/configuration-procedures_zh.md
index fdac9ccad..8046b94fc 100644
--- a/docs/configuration-procedures_zh.md
+++ b/docs/configuration-procedures_zh.md
@@ -381,4 +381,4 @@ python -m pytest utils/matrix_logic/ -v
 
 遵循[上游配方 #946](https://github.com/vllm-project/recipes/pull/946) 的 AMD 设置：`VLLM_ROCM_USE_AITER=1`、`VLLM_ROCM_USE_AITER_MOE=1` 和 `--moe-backend aiter_triton_mxfp4_bf16`。配方通过 `WEKA_LOADER_OVERRIDE` 固定使用完整语料 `semianalysis_cc_traces_weka_062126`。KV 驻留 GPU；Engram 沿用上游 AMD 默认设置。不要复制 NVIDIA 的 `--engram-config` 选项：上游目前在 ROCm 上拒绝该选项。MI355X launcher 使用共享 HF 缓存，并将此模型的仓库挂载至 `/ix`，同时导出 `INFMAX_CONTAINER_WORKSPACE=/ix`，确保 AgentX 依赖与输出路径位于该挂载中。
 
-**GPU 验证尚待完成：** 配方使用 `vllm/vllm-openai-rocm:nightly-TBD`。[上游 vLLM 配方](https://recipes.vllm.ai/deepseek-ai/DeepSeek-V4.1-Flash?hardware=mi355x)引用的 `deepseekv41-flash-0909` 标签早于 [vllm-project/vllm#56503](https://github.com/vllm-project/vllm/pull/56503)，该 PR 将 mHC delayed pre 块从 eager Torch 参考实现切换到 AITER；该块占此模型解码器内核启动数的 85%，因此发布标签会浪费大部分解码开销。nightly 标签目前为占位符，待包含该提交的首个 nightly 发布后，将按其他 ROCm 条目的惯例固定为 `nightly-<sha>`。Engram 行为与服务兼容性仍需 GPU 验证。请按 [AgentX 流程](./eval-agentx-procedures_zh.md) 获取运行时证据；本地矩阵生成和镜像元数据不等于 GPU 验证。
+**GPU 验证尚待完成：** 配方使用 `vllm/vllm-openai-rocm:nightly-eed1f3d0c6043bd494424a22443ee198dd56f657`（摘要 `sha256:960228cf…`，发布于 2026-09-12）。[上游 vLLM 配方](https://recipes.vllm.ai/deepseek-ai/DeepSeek-V4.1-Flash?hardware=mi355x)引用的 `deepseekv41-flash-0909` 标签早于 [vllm-project/vllm#56503](https://github.com/vllm-project/vllm/pull/56503)，该 PR 将 mHC delayed pre 块从 eager Torch 参考实现切换到 AITER；该块占此模型解码器内核启动数的 85%，因此发布标签会浪费大部分解码开销。上游配方正在 [vllm-project/recipes#962](https://github.com/vllm-project/recipes/pull/962) 中切换到同一 nightly。Engram 行为与服务兼容性仍需 GPU 验证。请按 [AgentX 流程](./eval-agentx-procedures_zh.md) 获取运行时证据；本地矩阵生成和镜像元数据不等于 GPU 验证。
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index a3b258351..ea5505840 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -7412,8 +7412,8 @@
 - config-keys:
     - dsv41flash-fp4-mi355x-vllm-agentic-dspark
   description:
-    - "Switch to the upstream ROCm nightly image. The deepseekv41-flash-0909 tag predates vllm-project/vllm#56503, which moves the mHC delayed pre block off the eager Torch reference onto AITER; that block is 85% of the decoder's kernel launches, measured 15.9-20.9x faster under cudagraph with gsm8k unchanged at 0.9719. Tag is nightly-TBD until the first nightly containing the commit is published."
-    - "切换到上游 ROCm nightly 镜像。deepseekv41-flash-0909 标签早于 vllm-project/vllm#56503，该 PR 将 mHC delayed pre 块从 eager Torch 参考实现切换到 AITER；该块占解码器内核启动数的 85%，在 cudagraph 下实测加速 15.9-20.9 倍，gsm8k 保持 0.9719 不变。标签暂为 nightly-TBD，待包含该提交的首个 nightly 发布后固定。"
+    - "Switch to the upstream ROCm nightly image nightly-eed1f3d0c6043bd494424a22443ee198dd56f657. The deepseekv41-flash-0909 tag predates vllm-project/vllm#56503, which moves the mHC delayed pre block off the eager Torch reference onto AITER; that block is 85% of the decoder's kernel launches, measured 15.9-20.9x faster under cudagraph with gsm8k unchanged at 0.9719. The upstream recipe moves to the same nightly in vllm-project/recipes#962."
+    - "切换到上游 ROCm nightly 镜像 nightly-eed1f3d0c6043bd494424a22443ee198dd56f657。deepseekv41-flash-0909 标签早于 vllm-project/vllm#56503，该 PR 将 mHC delayed pre 块从 eager Torch 参考实现切换到 AITER；该块占解码器内核启动数的 85%，在 cudagraph 下实测加速 15.9-20.9 倍，gsm8k 保持 0.9719 不变。上游配方在 vllm-project/recipes#962 中切换到同一 nightly。"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3035
 
 - config-keys:

From ab76c1313fe8a870cf2049bff8d93d7d3b3c02de Mon Sep 17 00:00:00 2001
From: Fangzhou Ai <31551580+Fangzhou-Ai@users.noreply.github.com>
Date: Sat, 12 Sep 2026 15:10:44 +0000
Subject: [PATCH 6/7] perf: select the CK a8w4 MoE experts on MI355X DSv4.1
 Flash

The recipe pinned --moe-backend aiter_triton_mxfp4_bf16, which names a
single W4A16 entry and dispatched _moe_gemm_a16w4. Passing plain aiter
opens vLLM's priority list, whose head is the Composable Kernel backend,
and CK quantizes activations to FP8 internally: the decode trace
switches to mfma_moe1_silu_mul_afp8_wfp4_bf16 and
mfma_moe2_afp8_wfp4_bf16, the same a8w4 expert family the DSV4-Pro
MI355X recipe already runs. The BF16 in the backend name and this
checkpoint's activation_scheme=dynamic both suggest otherwise, hence the
comment at the flag.

Measured on one MI355X half, TP4, 131k context, CUDA graphs on, both
arms captured on the same four GPUs under identical profiling overhead:

  conc   MoE GEMM ms/step      whole step ms
   1     2.72 -> 2.30          19.11 -> 18.31   (4.2%)
   4     4.91 -> 4.27          19.42 -> 18.40   (5.3%)
  16     9.06 -> 7.18          28.38 -> 25.92   (8.7%)

AgentX replay at conc 1, each arm solo on a quiet node, improves mean
ITL from 4.01 to 3.91 ms and mean TTFT from 681 to 639 ms. gsm8k is
unchanged: 0.9719 +/- 0.0045 strict-match, 0.9712 +/- 0.0046 flexible,
identical to the Triton arm.

Signed-off-by: Fangzhou Ai <31551580+Fangzhou-Ai@users.noreply.github.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
---
 .../agentic/dsv41flash_fp4_mi355x_vllm_mtp.sh            | 9 ++++++++-
 docs/configuration-procedures.md                         | 2 +-
 docs/configuration-procedures_zh.md                      | 2 +-
 perf-changelog.yaml                                      | 7 +++++++
 4 files changed, 17 insertions(+), 3 deletions(-)

diff --git a/benchmarks/single_node/agentic/dsv41flash_fp4_mi355x_vllm_mtp.sh b/benchmarks/single_node/agentic/dsv41flash_fp4_mi355x_vllm_mtp.sh
index fa2ded18a..4045fea9b 100755
--- a/benchmarks/single_node/agentic/dsv41flash_fp4_mi355x_vllm_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv41flash_fp4_mi355x_vllm_mtp.sh
@@ -82,7 +82,14 @@ VLLM_CMD=(
     --tokenizer-mode deepseek_v41
     --tool-call-parser deepseek_v41 --enable-auto-tool-choice
     --reasoning-parser deepseek_v41
-    --moe-backend aiter_triton_mxfp4_bf16
+    # aiter, not aiter_triton_mxfp4_bf16: the plain name opens vLLM's full
+    # priority list and the CK kernel at its head wins. Despite the BF16
+    # backend name and this checkpoint's activation_scheme=dynamic, CK
+    # quantizes activations to FP8 internally and dispatches the a8w4 experts
+    # (mfma_moe1_silu_mul_afp8_wfp4_bf16 / mfma_moe2_afp8_wfp4_bf16) that the
+    # DSV4-Pro MI355X recipe already gets. Pinning the Triton name instead
+    # forced the W4A16 _moe_gemm_a16w4 kernel.
+    --moe-backend aiter
     --gpu-memory-utilization 0.9
     --speculative-config "$SPEC_CONFIG"
     --max-model-len 1048576
diff --git a/docs/configuration-procedures.md b/docs/configuration-procedures.md
index 98e265a58..a7930e969 100644
--- a/docs/configuration-procedures.md
+++ b/docs/configuration-procedures.md
@@ -401,6 +401,6 @@ A configuration is ready for sweep only when the executable files agree, the exa
 
 The draft `dsv41flash-fp4-mi355x-vllm-agentic-dspark` recipe extends [#2958](https://github.com/SemiAnalysisAI/InferenceX/pull/2958) to MI355X AgentX: TP4, concurrency 1–32, native five-token DSpark. Throughput uses the [committed golden AL](../golden_al_distribution/dsv41flash_dspark.yaml) of 3.51 for thinking on and five draft tokens, with synthetic rejection sampling and adaptive verification disabled. Accuracy evals retain real block rejection but, unlike the CUDA arms, also keep adaptive verification disabled: it trims verification requests on device, which the ROCm `DeepseekV4IndexerBackend` does not support, and the engine refused to start with it enabled ([run 34651830283](https://github.com/SemiAnalysisAI/InferenceX/actions/runs/34651830283)). FP4 describes the MXFP4 experts; the checkpoint also contains MXFP8 weights.
 
-Follow the AMD overrides in [upstream recipe #946](https://github.com/vllm-project/recipes/pull/946): `VLLM_ROCM_USE_AITER=1`, `VLLM_ROCM_USE_AITER_MOE=1`, and `--moe-backend aiter_triton_mxfp4_bf16`. The recipe pins `semianalysis_cc_traces_weka_062126` (the unfiltered corpus) via `WEKA_LOADER_OVERRIDE`. KV stays GPU-resident; Engram follows upstream AMD defaults. Do not copy the NVIDIA `--engram-config` option: upstream currently rejects it on ROCm. The MI355X launcher uses the shared HF cache and mounts this model's repository at `/ix`, and exports `INFMAX_CONTAINER_WORKSPACE=/ix` so AgentX dependencies and outputs resolve inside that mount.
+Follow the AMD overrides in [upstream recipe #946](https://github.com/vllm-project/recipes/pull/946): `VLLM_ROCM_USE_AITER=1`, `VLLM_ROCM_USE_AITER_MOE=1`, and `--moe-backend aiter` (the upstream recipe names the Triton backend explicitly; `aiter` lets vLLM pick the CK a8w4 experts instead, matching the DSV4-Pro MI355X recipe). The recipe pins `semianalysis_cc_traces_weka_062126` (the unfiltered corpus) via `WEKA_LOADER_OVERRIDE`. KV stays GPU-resident; Engram follows upstream AMD defaults. Do not copy the NVIDIA `--engram-config` option: upstream currently rejects it on ROCm. The MI355X launcher uses the shared HF cache and mounts this model's repository at `/ix`, and exports `INFMAX_CONTAINER_WORKSPACE=/ix` so AgentX dependencies and outputs resolve inside that mount.
 
 **GPU validation is pending:** The recipe uses `vllm/vllm-openai-rocm:nightly-eed1f3d0c6043bd494424a22443ee198dd56f657` (digest `sha256:960228cf…`, published 2026-09-12). The `deepseekv41-flash-0909` tag referenced by the [upstream vLLM recipe](https://recipes.vllm.ai/deepseek-ai/DeepSeek-V4.1-Flash?hardware=mi355x) predates [vllm-project/vllm#56503](https://github.com/vllm-project/vllm/pull/56503), which moves the mHC delayed pre block off the eager Torch reference and onto AITER; that block is 85% of the decoder's kernel launches on this model, so the release tag leaves most of the decode cost on the table. The upstream recipe is being moved to the same nightly in [vllm-project/recipes#962](https://github.com/vllm-project/recipes/pull/962). Engram behavior and serving compatibility still require GPU validation. Follow the [AgentX procedure](./eval-agentx-procedures.md#7-run-agentx-fast-feedback-versus-canonical-evidence) for runtime evidence; local generation and registry metadata are not GPU proof.
diff --git a/docs/configuration-procedures_zh.md b/docs/configuration-procedures_zh.md
index 8046b94fc..0fb3daeff 100644
--- a/docs/configuration-procedures_zh.md
+++ b/docs/configuration-procedures_zh.md
@@ -379,6 +379,6 @@ python -m pytest utils/matrix_logic/ -v
 
 草案配方 `dsv41flash-fp4-mi355x-vllm-agentic-dspark` 将 [#2958](https://github.com/SemiAnalysisAI/InferenceX/pull/2958) 扩展至 MI355X AgentX：TP4、并发 1–32、原生五 token DSpark。吞吐测试使用[已提交的黄金 AL](../golden_al_distribution/dsv41flash_dspark.yaml)：thinking 开启、五个草稿 token 对应 3.51，采用合成拒绝采样并关闭自适应验证。准确率 eval 保留真实块拒绝采样，但与 CUDA 分支不同，同样关闭自适应验证：它会在设备端裁剪验证请求，而 ROCm 的 `DeepseekV4IndexerBackend` 不支持该操作，启用后引擎拒绝启动（[运行 34651830283](https://github.com/SemiAnalysisAI/InferenceX/actions/runs/34651830283)）。FP4 表示 MXFP4 专家权重；检查点还包含 MXFP8 权重。
 
-遵循[上游配方 #946](https://github.com/vllm-project/recipes/pull/946) 的 AMD 设置：`VLLM_ROCM_USE_AITER=1`、`VLLM_ROCM_USE_AITER_MOE=1` 和 `--moe-backend aiter_triton_mxfp4_bf16`。配方通过 `WEKA_LOADER_OVERRIDE` 固定使用完整语料 `semianalysis_cc_traces_weka_062126`。KV 驻留 GPU；Engram 沿用上游 AMD 默认设置。不要复制 NVIDIA 的 `--engram-config` 选项：上游目前在 ROCm 上拒绝该选项。MI355X launcher 使用共享 HF 缓存，并将此模型的仓库挂载至 `/ix`，同时导出 `INFMAX_CONTAINER_WORKSPACE=/ix`，确保 AgentX 依赖与输出路径位于该挂载中。
+遵循[上游配方 #946](https://github.com/vllm-project/recipes/pull/946) 的 AMD 设置：`VLLM_ROCM_USE_AITER=1`、`VLLM_ROCM_USE_AITER_MOE=1` 和 `--moe-backend aiter`（上游配方显式指定 Triton 后端；使用 `aiter` 可让 vLLM 选择 CK a8w4 专家内核，与 DSV4-Pro MI355X 配方一致）。配方通过 `WEKA_LOADER_OVERRIDE` 固定使用完整语料 `semianalysis_cc_traces_weka_062126`。KV 驻留 GPU；Engram 沿用上游 AMD 默认设置。不要复制 NVIDIA 的 `--engram-config` 选项：上游目前在 ROCm 上拒绝该选项。MI355X launcher 使用共享 HF 缓存，并将此模型的仓库挂载至 `/ix`，同时导出 `INFMAX_CONTAINER_WORKSPACE=/ix`，确保 AgentX 依赖与输出路径位于该挂载中。
 
 **GPU 验证尚待完成：** 配方使用 `vllm/vllm-openai-rocm:nightly-eed1f3d0c6043bd494424a22443ee198dd56f657`（摘要 `sha256:960228cf…`，发布于 2026-09-12）。[上游 vLLM 配方](https://recipes.vllm.ai/deepseek-ai/DeepSeek-V4.1-Flash?hardware=mi355x)引用的 `deepseekv41-flash-0909` 标签早于 [vllm-project/vllm#56503](https://github.com/vllm-project/vllm/pull/56503)，该 PR 将 mHC delayed pre 块从 eager Torch 参考实现切换到 AITER；该块占此模型解码器内核启动数的 85%，因此发布标签会浪费大部分解码开销。上游配方正在 [vllm-project/recipes#962](https://github.com/vllm-project/recipes/pull/962) 中切换到同一 nightly。Engram 行为与服务兼容性仍需 GPU 验证。请按 [AgentX 流程](./eval-agentx-procedures_zh.md) 获取运行时证据；本地矩阵生成和镜像元数据不等于 GPU 验证。
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index ea5505840..c1ea8ea33 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -7422,3 +7422,10 @@
     - "Raise max_num_seqs to the upstream default of 128. The 2*CONC cap sat below AgentX's subagent fan-out, so requests queued on scheduling capacity; vLLM reported queue times up to 16.2 s at CONC=1, and removing the cap took them to zero."
     - "将 max_num_seqs 提升到上游默认值 128。此前的 2*CONC 上限低于 AgentX 的子代理扇出，导致请求因调度容量排队；vLLM 在 CONC=1 时报告排队时间最高达 16.2 秒，取消该上限后归零。"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3035
+
+- config-keys:
+    - dsv41flash-fp4-mi355x-vllm-agentic-dspark
+  description:
+    - "Switch --moe-backend from aiter_triton_mxfp4_bf16 to aiter, which selects the CK a8w4 experts (mfma_moe1_silu_mul_afp8_wfp4_bf16 / mfma_moe2_afp8_wfp4_bf16) instead of the W4A16 _moe_gemm_a16w4 Triton kernel, matching the DSV4-Pro MI355X recipe. Decode traces at 131k context show MoE GEMM falling from 2.72 to 2.30 ms/step at conc 1, 4.91 to 4.27 at conc 4 and 9.06 to 7.18 at conc 16, for whole-step gains of 4.2%, 5.3% and 8.7%. AgentX conc 1 improves mean ITL 4.01 to 3.91 ms and mean TTFT 681 to 639 ms; gsm8k is unchanged at 0.9719 strict-match."
+    - "将 --moe-backend 从 aiter_triton_mxfp4_bf16 改为 aiter，从而选择 CK a8w4 专家内核（mfma_moe1_silu_mul_afp8_wfp4_bf16 / mfma_moe2_afp8_wfp4_bf16），取代 W4A16 的 _moe_gemm_a16w4 Triton 内核，与 DSV4-Pro MI355X 配方保持一致。131k 上下文的解码 trace 显示 MoE GEMM 在并发 1 时从 2.72 降至 2.30 ms/step，并发 4 从 4.91 降至 4.27，并发 16 从 9.06 降至 7.18，整步分别加速 4.2%、5.3% 和 8.7%。AgentX 并发 1 的平均 ITL 从 4.01 降至 3.91 ms，平均 TTFT 从 681 降至 639 ms；gsm8k strict-match 保持 0.9719 不变。"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3058

From 15a373823a113a7da6f9c0b39092c6ef27d356dc Mon Sep 17 00:00:00 2001
From: Chun Fang <chun.fang@amd.com>
Date: Sun, 13 Sep 2026 01:49:43 +0000
Subject: [PATCH 7/7] Fix per-changelog

---
 perf-changelog.yaml | 14 --------------
 1 file changed, 14 deletions(-)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index bc9cfee29..6511841ca 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -7429,20 +7429,6 @@
     - "增加固定序列请求诊断与多节点并发文件名处理的代表性 AMD 覆盖；共享处理器变更也影响未列出的固定序列配置。"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3026
 
-- config-keys:
-    - dsv41flash-fp4-mi355x-vllm-agentic-dspark
-  description:
-    - "Switch to the upstream ROCm nightly image nightly-eed1f3d0c6043bd494424a22443ee198dd56f657. The deepseekv41-flash-0909 tag predates vllm-project/vllm#56503, which moves the mHC delayed pre block off the eager Torch reference onto AITER; that block is 85% of the decoder's kernel launches, measured 15.9-20.9x faster under cudagraph with gsm8k unchanged at 0.9719. The upstream recipe moves to the same nightly in vllm-project/recipes#962."
-    - "切换到上游 ROCm nightly 镜像 nightly-eed1f3d0c6043bd494424a22443ee198dd56f657。deepseekv41-flash-0909 标签早于 vllm-project/vllm#56503，该 PR 将 mHC delayed pre 块从 eager Torch 参考实现切换到 AITER；该块占解码器内核启动数的 85%，在 cudagraph 下实测加速 15.9-20.9 倍，gsm8k 保持 0.9719 不变。上游配方在 vllm-project/recipes#962 中切换到同一 nightly。"
-  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3035
-
-- config-keys:
-    - dsv41flash-fp4-mi355x-vllm-agentic-dspark
-  description:
-    - "Raise max_num_seqs to the upstream default of 128. The 2*CONC cap sat below AgentX's subagent fan-out, so requests queued on scheduling capacity; vLLM reported queue times up to 16.2 s at CONC=1, and removing the cap took them to zero."
-    - "将 max_num_seqs 提升到上游默认值 128。此前的 2*CONC 上限低于 AgentX 的子代理扇出，导致请求因调度容量排队；vLLM 在 CONC=1 时报告排队时间最高达 16.2 秒，取消该上限后归零。"
-  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3035
-
 - config-keys:
     - dsv41flash-fp4-mi355x-vllm-agentic-dspark
   description: