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 e53578124..8e0c67aef 100755
--- a/benchmarks/single_node/agentic/dsv41flash_fp4_mi355x_vllm_mtp.sh
+++ b/benchmarks/single_node/agentic/dsv41flash_fp4_mi355x_vllm_mtp.sh
@@ -71,22 +71,16 @@ while (( CAPTURE_SIZE < GRAPH_NUM_SEQS * (1 + NUM_SPEC_TOKENS) && CAPTURE_SIZE <
     CAPTURE_SIZE=$((CAPTURE_SIZE * 2))
 done
 
-# The sparse-attention indexer and its companion per-rank buffers scale with
-# --max-num-batched-tokens at roughly 4.4 MiB per token, measured on gfx950, so
-# a smaller prefill chunk buys KV room. TP=2 starts from half the per-rank space
-# and is the arm that runs short: at the upstream 16384 it holds 7.84M KV
-# tokens, 122K per request at c64, where run 35574132719 fell to a 17.6% prefix
-# cache hit rate, 187 s TTFT and 150 tok/s against 94.8%, 1.3 s and 957 tok/s at
-# c32. Every point that held had 232K per request or more, so keep the upstream
-# chunk through c32 (245K at TP=2) and trade it away only above that. B300 runs
-# 8192 at TP=4 and the Blackwell TP=2 arms run 4096 (#3320, #3321).
-if (( CONC >= 128 )); then
-    (( TP == 2 )) && BATCHED_TOKENS=4096 || BATCHED_TOKENS=8192
-elif (( TP == 2 && CONC >= 64 )); then
-    BATCHED_TOKENS=8192
-else
-    BATCHED_TOKENS=16384
-fi
+# ATOM's TP2/TP4 AgentX recipe for this model (ROCm/ATOM
+# dsv41flash_fp4_mi355x_atom_mtp.sh) runs the upstream 16384 chunk unconditionally
+# at every concurrency, including TP=2 c64/c128 where this recipe previously
+# traded chunk size for KV room (see the superseded comment this replaces: the
+# sparse-attention indexer's per-rank buffers scale with --max-num-batched-tokens
+# at ~4.4 MiB/token on gfx950, and run 35574132719 saw TP=2 c64 fall to 17.6%
+# prefix cache hit rate at the smaller 8192 chunk before this change). Matching
+# ATOM's fixed value here is unvalidated at c64/c128 on this SKU -- confirm KV
+# headroom does not regress at those points before taking this out of draft.
+BATCHED_TOKENS=16384
 
 # DSpark verifies 1+5 tokens per sequence, so a decode batch of max_num_seqs
 # needs six times that many token slots. The MI355X API-server default of 1024
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 543c40458..102950f18 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -1409,8 +1409,9 @@ dsv4-fp4-mi355x-sglang-agentic-mtp:
 # Two GPUs per server doubles the servers per node and is the layout that
 # decides whether DSv4.1-Flash is throughput- or interactivity-bound here.
 dsv41flash-fp4-mi355x-vllm-agentic-dspark:
-  # ROCm 10.0 nightly channel, shared with kimik3-fp4-mi355x-vllm-agentic-mtp.
-  image: vllm/vllm-openai-rocm:nightly-rocm100-29468dde8b515031dc6d4d9d06bf0a2fa0442098
+  # TODO: repin once the fixed 16384 prefill chunk above (matching ATOM) has
+  # been validated for KV headroom at TP=2 c64/c128 and TP=4 c128 on this SKU.
+  image: TBD
   model: deepseek-ai/DeepSeek-V4.1-Flash
   model-prefix: dsv41flash
   runner: cluster:mi355x-amds
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 1bc132cd3..7840d4084 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -8950,3 +8950,10 @@
   description:
     - "Update SGLang image from v0.5.19-cu130 to v0.5.20-cu130 and pass --cuda-graph-max-bs-decode after the alias removal."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3334
+
+- config-keys:
+    - dsv41flash-fp4-mi355x-vllm-agentic-dspark
+  description:
+    - "Match ATOM's fixed 16384 --max-num-batched-tokens (prefill chunk) at every concurrency instead of reducing to 4096/8192 at TP=2 c64+/c128+ and TP=4 c128; unpin image to TBD pending KV-headroom validation at those points on this SKU."
+    - "将 --max-num-batched-tokens（预填充块大小）在所有并发下固定为与 ATOM 一致的 16384，不再在 TP=2 c64+/c128+ 与 TP=4 c128 处降为 4096/8192；镜像取消固定为 TBD，待在该 SKU 上验证这些并发点的 KV 显存余量后再恢复固定。"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3451