diff --git a/benchmarks/single_node/fixed_seq_len/dsv4_fp8_mi325x.sh b/benchmarks/single_node/fixed_seq_len/dsv4_fp8_mi325x.sh
new file mode 100755
index 000000000..4dafff650
--- /dev/null
+++ b/benchmarks/single_node/fixed_seq_len/dsv4_fp8_mi325x.sh
@@ -0,0 +1,117 @@
+#!/usr/bin/env bash
+set -eo pipefail
+
+# DeepSeek-V4-Pro FP8 single-node on MI325X (gfx942) via vLLM.
+#
+# EXTRAPOLATED bring-up recipe. The sglang path was abandoned: on gfx942
+# (no native FP4) the dsv4 sglang backend's nvfp4 MoE / TileLang-MLA kernels
+# have no gfx942 equivalents (they exist only for gfx950/MI355X). vLLM instead
+# runs the checkpoint in FP8 via --quantization deepseek_v4_fp8, which
+# dequantizes the FP4 MoE experts to FP8 — the same path the H200 dsv4 vLLM
+# recipe uses (H200 is also a no-FP4 SKU). Derived from:
+#   * same model + framework + AMD family: dsv4_fp4_mi355x_vllm.sh (ROCm vLLM
+#     dsv4 structure: AITER MoE, deepseek_v4 tokenizer/parser, mp executor,
+#     FULL_AND_PIECEWISE compile)
+#   * same model, FP8 path: dsv4_fp8_h200.sh (--quantization deepseek_v4_fp8)
+#   * same SKU, different model: minimaxm3_fp8_mi325x.sh (gfx942 vLLM/AITER)
+#
+# The FP4->FP8 dequant roughly doubles the MoE footprint (~1.05 TB total),
+# which fits 8x256 GB comfortably at TP8, so the sweep is TP8-only.
+#
+# MoE backend is left at auto (NOT --moe-backend aiter) — see dsv4_fp8_mi300x.sh:
+# on gfx942, forcing aiter selects AITER_MXFP4_MXFP4 (W4A4 native-mxfp4) which
+# the gfx942 kernel rejects; auto's ROCm+DeepseekV4 path prefers
+# AITER_MXFP4_BF16 (W4A16, dequant) with a TRITON_UNFUSED fallback.
+
+source "$(dirname "$0")/../../benchmark_lib.sh"
+
+check_env_vars \
+    MODEL \
+    TP \
+    DP_ATTENTION \
+    CONC \
+    ISL \
+    OSL \
+    MAX_MODEL_LEN \
+    RANDOM_RANGE_RATIO \
+    RESULT_FILENAME
+
+if [[ -n "$SLURM_JOB_ID" ]]; then
+  echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
+fi
+
+if [[ "$MODEL" != /* ]]; then hf download "$MODEL"; fi
+
+if [ -n "$ROCR_VISIBLE_DEVICES" ]; then
+    export HIP_VISIBLE_DEVICES="$ROCR_VISIBLE_DEVICES"
+fi
+
+export VLLM_ROCM_USE_AITER=1
+export VLLM_ROCM_USE_AITER_MOE=1
+
+# gsm8k eval at high concurrency (8k1k) OOM-kills the server: hundreds of
+# concurrent 9472-token requests exceed the ~20x KV budget even on 256GB MI325X
+# (c128 fit, so this was originally left at the default, but c512 crashed the
+# EngineCore mid-eval). Cap the eval to a safe in-flight count; only run_eval is
+# affected (throughput jobs use CONC directly). Matches the MI300X script.
+export EVAL_CONCURRENT_REQUESTS=8
+
+SERVER_LOG=/workspace/server.log
+
+if [ "${EVAL_ONLY}" = "true" ]; then
+    setup_eval_context
+    MAX_MODEL_LEN="$EVAL_MAX_MODEL_LEN"
+fi
+
+start_gpu_monitor
+
+PARALLEL_ARGS=(--tensor-parallel-size "$TP" --data-parallel-size 1)
+if [ "${DP_ATTENTION}" = "true" ]; then
+    PARALLEL_ARGS=(--tensor-parallel-size 1 --data-parallel-size "$TP")
+fi
+
+EP_ARGS=()
+if [ "${EP_SIZE:-1}" -gt 1 ]; then
+    EP_ARGS=(--enable-expert-parallel)
+fi
+
+set -x
+vllm serve $MODEL --port $PORT \
+    "${PARALLEL_ARGS[@]}" \
+    "${EP_ARGS[@]}" \
+    --quantization deepseek_v4_fp8 \
+    --async-scheduling \
+    --no-enable-prefix-caching \
+    --distributed-executor-backend mp \
+    --gpu-memory-utilization 0.9 \
+    --max-model-len "$MAX_MODEL_LEN" \
+    --kv-cache-dtype fp8 \
+    --trust-remote-code \
+    --tokenizer-mode deepseek_v4 \
+    --reasoning-parser deepseek_v4 \
+    --compilation-config '{"mode":3,"cudagraph_mode":"FULL_AND_PIECEWISE"}' > $SERVER_LOG 2>&1 &
+
+SERVER_PID=$!
+
+wait_for_server_ready --port "$PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
+
+run_benchmark_serving \
+    --model "$MODEL" \
+    --port "$PORT" \
+    --backend vllm \
+    --input-len "$ISL" \
+    --output-len "$OSL" \
+    --random-range-ratio "$RANDOM_RANGE_RATIO" \
+    --num-prompts "$((CONC * 10))" \
+    --max-concurrency "$CONC" \
+    --result-filename "$RESULT_FILENAME" \
+    --result-dir /workspace/ \
+    --trust-remote-code
+
+if [ "${RUN_EVAL}" = "true" ]; then
+    run_eval --framework lm-eval --port "$PORT"
+    append_lm_eval_summary
+fi
+
+stop_gpu_monitor
+set +x
diff --git a/benchmarks/single_node/fixed_seq_len/dsv4_fp8_mi325x_mtp.sh b/benchmarks/single_node/fixed_seq_len/dsv4_fp8_mi325x_mtp.sh
new file mode 100755
index 000000000..7233da9dd
--- /dev/null
+++ b/benchmarks/single_node/fixed_seq_len/dsv4_fp8_mi325x_mtp.sh
@@ -0,0 +1,126 @@
+#!/usr/bin/env bash
+set -eo pipefail
+
+# DeepSeek-V4-Pro FP8 single-node on MI325X (gfx942) via vLLM, MTP variant.
+#
+# MTP sibling of dsv4_fp8_mi325x.sh: adds --speculative-config
+# '{"method":"mtp","num_speculative_tokens":2}' (DeepSeek-V4 built-in MTP)
+# and --dsv4 chat-template encoding for run_benchmark_serving.
+#
+# EXTRAPOLATED bring-up recipe. The sglang path was abandoned: on gfx942
+# (no native FP4) the dsv4 sglang backend's nvfp4 MoE / TileLang-MLA kernels
+# have no gfx942 equivalents (they exist only for gfx950/MI355X). vLLM instead
+# runs the checkpoint in FP8 via --quantization deepseek_v4_fp8, which
+# dequantizes the FP4 MoE experts to FP8 — the same path the H200 dsv4 vLLM
+# recipe uses (H200 is also a no-FP4 SKU). Derived from:
+#   * same model + framework + AMD family: dsv4_fp4_mi355x_vllm.sh (ROCm vLLM
+#     dsv4 structure: AITER MoE, deepseek_v4 tokenizer/parser, mp executor,
+#     FULL_AND_PIECEWISE compile)
+#   * same model, FP8 path: dsv4_fp8_h200.sh (--quantization deepseek_v4_fp8)
+#   * same SKU, different model: minimaxm3_fp8_mi325x.sh (gfx942 vLLM/AITER)
+#
+# The FP4->FP8 dequant roughly doubles the MoE footprint (~1.05 TB total),
+# which fits 8x256 GB comfortably at TP8, so the sweep is TP8-only.
+#
+# MoE backend is left at auto (NOT --moe-backend aiter) — see dsv4_fp8_mi300x.sh:
+# on gfx942, forcing aiter selects AITER_MXFP4_MXFP4 (W4A4 native-mxfp4) which
+# the gfx942 kernel rejects; auto's ROCm+DeepseekV4 path prefers
+# AITER_MXFP4_BF16 (W4A16, dequant) with a TRITON_UNFUSED fallback.
+
+source "$(dirname "$0")/../../benchmark_lib.sh"
+
+check_env_vars \
+    MODEL \
+    TP \
+    DP_ATTENTION \
+    CONC \
+    ISL \
+    OSL \
+    MAX_MODEL_LEN \
+    RANDOM_RANGE_RATIO \
+    RESULT_FILENAME
+
+if [[ -n "$SLURM_JOB_ID" ]]; then
+  echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
+fi
+
+if [[ "$MODEL" != /* ]]; then hf download "$MODEL"; fi
+
+if [ -n "$ROCR_VISIBLE_DEVICES" ]; then
+    export HIP_VISIBLE_DEVICES="$ROCR_VISIBLE_DEVICES"
+fi
+
+export VLLM_ROCM_USE_AITER=1
+export VLLM_ROCM_USE_AITER_MOE=1
+
+# gsm8k eval at high concurrency (8k1k) OOM-kills the server: hundreds of
+# concurrent 9472-token requests exceed the ~20x KV budget even on 256GB MI325X
+# (c128 fit, so this was originally left at the default, but c512 crashed the
+# EngineCore mid-eval). Cap the eval to a safe in-flight count; only run_eval is
+# affected (throughput jobs use CONC directly). Matches the MI300X script.
+export EVAL_CONCURRENT_REQUESTS=8
+
+SERVER_LOG=/workspace/server.log
+
+if [ "${EVAL_ONLY}" = "true" ]; then
+    setup_eval_context
+    MAX_MODEL_LEN="$EVAL_MAX_MODEL_LEN"
+fi
+
+start_gpu_monitor
+
+PARALLEL_ARGS=(--tensor-parallel-size "$TP" --data-parallel-size 1)
+if [ "${DP_ATTENTION}" = "true" ]; then
+    PARALLEL_ARGS=(--tensor-parallel-size 1 --data-parallel-size "$TP")
+fi
+
+EP_ARGS=()
+if [ "${EP_SIZE:-1}" -gt 1 ]; then
+    EP_ARGS=(--enable-expert-parallel)
+fi
+
+# Use 2 speculative tokens (matches dsv4_fp4_mi355x_vllm_mtp.sh).
+NUM_SPEC_TOKENS=2
+
+set -x
+vllm serve $MODEL --port $PORT \
+    "${PARALLEL_ARGS[@]}" \
+    "${EP_ARGS[@]}" \
+    --quantization deepseek_v4_fp8 \
+    --async-scheduling \
+    --no-enable-prefix-caching \
+    --distributed-executor-backend mp \
+    --gpu-memory-utilization 0.9 \
+    --max-model-len "$MAX_MODEL_LEN" \
+    --kv-cache-dtype fp8 \
+    --trust-remote-code \
+    --tokenizer-mode deepseek_v4 \
+    --reasoning-parser deepseek_v4 \
+    --speculative-config "{\"method\": \"mtp\", \"num_speculative_tokens\": $NUM_SPEC_TOKENS}" \
+    --compilation-config '{"mode":3,"cudagraph_mode":"FULL_AND_PIECEWISE"}' > $SERVER_LOG 2>&1 &
+
+SERVER_PID=$!
+
+wait_for_server_ready --port "$PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
+
+run_benchmark_serving \
+    --model "$MODEL" \
+    --port "$PORT" \
+    --backend vllm \
+    --input-len "$ISL" \
+    --output-len "$OSL" \
+    --random-range-ratio "$RANDOM_RANGE_RATIO" \
+    --num-prompts "$((CONC * 10))" \
+    --max-concurrency "$CONC" \
+    --result-filename "$RESULT_FILENAME" \
+    --result-dir /workspace/ \
+    --trust-remote-code \
+    --dsv4
+
+if [ "${RUN_EVAL}" = "true" ]; then
+    run_eval --framework lm-eval --port "$PORT"
+    append_lm_eval_summary
+fi
+
+stop_gpu_monitor
+set +x
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 4b1ea4aed..56ff8da58 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -119,6 +119,64 @@ dsr1-fp8-mi325x-sglang:
       search-space:
       - { tp: 8, conc-start: 4, conc-end: 64 }
 
+# DeepSeek-V4-Pro FP8 single-node on MI325X (gfx942) via vLLM.
+# EXTRAPOLATED bring-up. Same rationale as dsv4-fp8-mi300x-vllm: sglang has no
+# gfx942 build of the dsv4 nvfp4 MoE / TileLang-MLA kernels, so vLLM runs the
+# checkpoint in FP8 via --quantization deepseek_v4_fp8 (dequant FP4 MoE -> FP8),
+# the H200 dsv4 vLLM path. Config mirrors the same-model dsv4-fp4-mi355x-vllm
+# (TP8, conc 4-512); 8x256GB (2TB) has ample headroom for the ~1.05TB FP8
+# footprint. Launch script dsv4_fp8_mi325x.sh carries the deepseek_v4 + gfx942
+# AITER flags.
+dsv4-fp8-mi325x-vllm:
+  image: vllm/vllm-openai-rocm:nightly-09663abde0f50944a8d5ea30120666024b503faa
+  model: deepseek-ai/DeepSeek-V4-Pro
+  model-prefix: dsv4
+  # mi325x (mi325x-amds pool) runners are down; use the mi325x-tw cluster.
+  runner: cluster:mi325x-tw
+  precision: fp8
+  framework: vllm
+  multinode: false
+  scenarios:
+    fixed-seq-len:
+    - isl: 1024
+      osl: 1024
+      search-space:
+      - { tp: 8, conc-start: 4, conc-end: 512 }
+    - isl: 8192
+      osl: 1024
+      search-space:
+      # 8k1k KV fits only ~20x concurrency (9472-token requests), so conc512 is
+      # ~25x oversubscribed -> request timeouts. conc256 is proven green; cap here.
+      - { tp: 8, conc-start: 4, conc-end: 256 }
+
+# MTP variant of dsv4-fp8-mi325x-vllm. Mirrors the base recipe and adds
+# DeepSeek-V4 built-in MTP via --speculative-config (num_speculative_tokens=2),
+# routing to dsv4_fp8_mi325x_mtp.sh; benchmark uses --dsv4 chat-template
+# encoding (required for meaningful MTP acceptance).
+dsv4-fp8-mi325x-vllm-mtp:
+  image: vllm/vllm-openai-rocm:nightly-09663abde0f50944a8d5ea30120666024b503faa
+  model: deepseek-ai/DeepSeek-V4-Pro
+  model-prefix: dsv4
+  # mi325x (mi325x-amds pool) runners are down; use the mi325x-tw cluster.
+  runner: cluster:mi325x-tw
+  precision: fp8
+  framework: vllm
+  multinode: false
+  scenarios:
+    fixed-seq-len:
+    - isl: 1024
+      osl: 1024
+      search-space:
+      - { tp: 8, conc-start: 4, conc-end: 512, spec-decoding: mtp }
+    - isl: 8192
+      osl: 1024
+      search-space:
+      # 8k1k MTP KV is tighter than normal (draft model): conc256 is borderline
+      # (MI300X 19.2% req failures, flaked pass->fail across runs) and conc512
+      # is 55.8%. conc128 passed cleanly on the memory-tightest SKU (MI300X);
+      # cap 8k1k MTP at 128 (normal holds 256, 1k1k holds 512).
+      - { tp: 8, conc-start: 4, conc-end: 128, spec-decoding: mtp }
+
 dsr1-fp8-mi355x-sglang:
   image: lmsysorg/sglang:v0.5.12-rocm700-mi35x
   model: deepseek-ai/DeepSeek-R1-0528
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index d77c35f1f..eec5dde49 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -4781,3 +4781,21 @@
     - "Bump image to lmsysorg/sglang-rocm:v0.5.14-rocm720-mi35x-20260708"
     - "Clean the export envs"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2198
+
+- config-keys:
+    - dsv4-fp8-mi325x-vllm
+  description:
+    - "Add DeepSeek-V4-Pro FP8 single-node MI325X vLLM benchmark (new SKU, previously no dsv4 data on gfx942)"
+    - "vLLM, not sglang: same rationale as dsv4-fp8-mi300x-vllm — sglang has no gfx942 build of the dsv4 nvfp4 MoE / TileLang-MLA kernels (gfx950/MI355X only). vLLM runs the checkpoint in FP8 via --quantization deepseek_v4_fp8 (dequant FP4 MoE -> FP8), the no-FP4 H200 dsv4 vLLM path"
+    - "Recipe mirrors the same-model dsv4-fp4-mi355x-vllm (ROCm vLLM dsv4: deepseek_v4 tokenizer/reasoning-parser, mp executor, FULL_AND_PIECEWISE compile) plus gfx942 AITER infra from minimaxm3-fp8-mi325x-vllm; TP8 conc 4-512, ample headroom on 8x256GB"
+    - "MoE backend left at auto (unlike MI355X's --moe-backend aiter): forcing aiter on gfx942 selects AITER_MXFP4_MXFP4 (W4A4 native-mxfp4) which the gfx942 kernel rejects; auto's ROCm+DeepseekV4 path prefers AITER_MXFP4_BF16 (W4A16, dequant) with TRITON_UNFUSED fallback (see dsv4-fp8-mi300x-vllm)"
+    - "Image vllm/vllm-openai-rocm:nightly-09663abde0f50944a8d5ea30120666024b503faa (the DSv4-validated ROCm vLLM build)"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2195
+
+- config-keys:
+    - dsv4-fp8-mi325x-vllm-mtp
+  description:
+    - "Add DeepSeek-V4-Pro FP8 MI325X vLLM MTP (speculative-decoding) variant, mirroring dsv4-fp8-mi325x-vllm plus --speculative-config {\"method\":\"mtp\",\"num_speculative_tokens\":2} (DeepSeek-V4 built-in MTP)"
+    - "run_benchmark_serving uses --dsv4 chat-template encoding, required for meaningful MTP acceptance rate; routes to dsv4_fp8_mi325x_mtp.sh via the launcher's spec-decoding=mtp suffix"
+    - "256GB MI325X keeps the default eval concurrency (no KV cap needed)"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2195
diff --git a/runners/launch_mi325x-tw.sh b/runners/launch_mi325x-tw.sh
new file mode 100644
index 000000000..0ed4be196
--- /dev/null
+++ b/runners/launch_mi325x-tw.sh
@@ -0,0 +1,35 @@
+#!/usr/bin/bash
+set -euo pipefail
+
+# mi325x-tw is a NON-SLURM cluster: the runner executes on the GPU node itself
+# (docker + ROCm, no salloc/srun). So run the container directly on the node,
+# like launch_h100-cr.sh but AMD/ROCm (--device=/dev/kfd,/dev/dri instead of
+# --runtime=nvidia --gpus). dsv4 MI325X is single-node (TP8), so one node's
+# 8 GPUs suffice. Runs the same _mi325x.sh benchmark as the amds launcher.
+
+HF_HUB_CACHE_MOUNT="${HF_HUB_CACHE_MOUNT:-/home/gharunner/hf_hub_cache/}"
+mkdir -p "$HF_HUB_CACHE_MOUNT"
+export HF_HUB_CACHE="${HF_HUB_CACHE:-/hf_hub_cache}"
+PORT=8888
+server_name="bmk-server-${RUNNER_NAME:-mi325x-tw}"
+
+# Route spec-decoding=mtp configs to the _mtp benchmark script.
+SPEC_SUFFIX=$([[ "${SPEC_DECODING:-}" == "mtp" ]] && printf '_mtp' || printf '')
+
+export GPU_COUNT="${GPU_COUNT:-${TP:?TP must be set}}"
+
+set -x
+docker rm -f "$server_name" >/dev/null 2>&1 || true
+docker run --rm --network=host --name="$server_name" \
+--device=/dev/kfd --device=/dev/dri --group-add video --group-add render \
+--ipc=host --privileged --shm-size=32g --ulimit memlock=-1 --ulimit stack=67108864 \
+--security-opt seccomp=unconfined --cap-add=SYS_PTRACE \
+-v "$HF_HUB_CACHE_MOUNT:$HF_HUB_CACHE" \
+-v "$GITHUB_WORKSPACE:/workspace/" -w /workspace/ \
+-e HF_TOKEN -e HF_HUB_CACHE -e MODEL -e TP -e PP_SIZE -e DCP_SIZE -e PCP_SIZE -e GPU_COUNT -e CONC -e MAX_MODEL_LEN -e ISL -e OSL -e RUN_EVAL -e EVAL_ONLY -e RUNNER_TYPE -e RESULT_FILENAME -e RANDOM_RANGE_RATIO -e PORT="$PORT" \
+-e DP_ATTENTION -e EP_SIZE -e DP_SIZE -e EVAL_MAX_MODEL_LEN -e SPEC_DECODING -e NUM_SPEC_TOKENS \
+-e PROFILE -e SGLANG_TORCH_PROFILER_DIR -e VLLM_TORCH_PROFILER_DIR -e VLLM_RPC_TIMEOUT \
+-e PYTHONPYCACHEPREFIX=/tmp/pycache/ -e CUDA_DEVICE_ORDER=PCI_BUS_ID \
+--entrypoint=/bin/bash \
+"$IMAGE" \
+benchmarks/single_node/${SCENARIO_SUBDIR}"${EXP_NAME%%_*}_${PRECISION}_mi325x${SPEC_SUFFIX}.sh"