diff --git a/benchmarks/single_node/fixed_seq_len/dsv4_fp8_mi300x.sh b/benchmarks/single_node/fixed_seq_len/dsv4_fp8_mi300x.sh
new file mode 100755
index 000000000..b921b6054
--- /dev/null
+++ b/benchmarks/single_node/fixed_seq_len/dsv4_fp8_mi300x.sh
@@ -0,0 +1,123 @@
+#!/usr/bin/env bash
+set -eo pipefail
+
+# DeepSeek-V4-Pro FP8 single-node on MI300X (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_mi300x.sh (gfx942 vLLM/AITER)
+#
+# The FP4->FP8 dequant roughly doubles the MoE footprint (~1.05 TB total),
+# which fits 8x192 GB only at TP8, so the sweep is TP8-only.
+#
+# MoE backend is left at auto (NOT --moe-backend aiter). --quantization
+# deepseek_v4_fp8 only handles the dense/attention weights; the MoE experts
+# stay mxfp4 and go through vLLM's mxfp4 MoE selector. On gfx942, forcing
+# aiter selects AITER_MXFP4_MXFP4 (W4A4, native mxfp4) which the gfx942 kernel
+# rejects ("Mxfp4 MoE backend 'AITER_MXFP4_MXFP4' does not support ... QuantKey
+# (u8 ... col=32)"). With auto, vLLM's select_deepseek_v4_mxfp4_moe_backend
+# takes its ROCm+DeepseekV4 branch and prefers AITER_MXFP4_BF16 (W4A16 CK,
+# dequantizes weights — no native FP4), falling back to TRITON_UNFUSED. MI355X
+# keeps --moe-backend aiter because gfx950 supports the W4A4 kernel.
+
+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="$ROCR_VISIBLE_DEVICES"
+fi
+
+export VLLM_ROCM_USE_AITER=1
+export VLLM_ROCM_USE_AITER_MOE=1
+
+# Cap eval concurrency for gfx942's tight KV. FP8 weights (~131GB/GPU) leave
+# only ~71k tokens of KV on the 192GB MI300X ("Maximum concurrency ... 7.52x"
+# for a 9472-token request). The eval defaults to CONC (128) concurrent
+# requests, which OOM-kills the server mid-gsm8k. Cap to the KV budget; this
+# only affects run_eval (throughput jobs use CONC directly). MI325X (256GB)
+# has the headroom and keeps the default.
+export EVAL_CONCURRENT_REQUESTS=8
+
+SERVER_LOG=/workspace/server.log
+
+if [ "${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_mi300x_mtp.sh b/benchmarks/single_node/fixed_seq_len/dsv4_fp8_mi300x_mtp.sh
new file mode 100755
index 000000000..0f71361f6
--- /dev/null
+++ b/benchmarks/single_node/fixed_seq_len/dsv4_fp8_mi300x_mtp.sh
@@ -0,0 +1,133 @@
+#!/usr/bin/env bash
+set -eo pipefail
+
+# DeepSeek-V4-Pro FP8 single-node on MI300X (gfx942) via vLLM, MTP variant.
+#
+# MTP sibling of dsv4_fp8_mi300x.sh: adds --speculative-config
+# '{"method":"mtp","num_speculative_tokens":2}' (DeepSeek-V4 built-in MTP)
+# and --dsv4 chat-template encoding for run_benchmark_serving (EAGLE/MTP is
+# trained on chat-formatted inputs).
+#
+# 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_mi300x.sh (gfx942 vLLM/AITER)
+#
+# The FP4->FP8 dequant roughly doubles the MoE footprint (~1.05 TB total),
+# which fits 8x192 GB only at TP8, so the sweep is TP8-only.
+#
+# MoE backend is left at auto (NOT --moe-backend aiter). --quantization
+# deepseek_v4_fp8 only handles the dense/attention weights; the MoE experts
+# stay mxfp4 and go through vLLM's mxfp4 MoE selector. On gfx942, forcing
+# aiter selects AITER_MXFP4_MXFP4 (W4A4, native mxfp4) which the gfx942 kernel
+# rejects ("Mxfp4 MoE backend 'AITER_MXFP4_MXFP4' does not support ... QuantKey
+# (u8 ... col=32)"). With auto, vLLM's select_deepseek_v4_mxfp4_moe_backend
+# takes its ROCm+DeepseekV4 branch and prefers AITER_MXFP4_BF16 (W4A16 CK,
+# dequantizes weights — no native FP4), falling back to TRITON_UNFUSED. MI355X
+# keeps --moe-backend aiter because gfx950 supports the W4A4 kernel.
+
+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="$ROCR_VISIBLE_DEVICES"
+fi
+
+export VLLM_ROCM_USE_AITER=1
+export VLLM_ROCM_USE_AITER_MOE=1
+
+# Cap eval concurrency for gfx942's tight KV. FP8 weights (~131GB/GPU) leave
+# only ~71k tokens of KV on the 192GB MI300X ("Maximum concurrency ... 7.52x"
+# for a 9472-token request). The eval defaults to CONC (128) concurrent
+# requests, which OOM-kills the server mid-gsm8k. Cap to the KV budget; this
+# only affects run_eval (throughput jobs use CONC directly). MI325X (256GB)
+# has the headroom and keeps the default.
+export EVAL_CONCURRENT_REQUESTS=8
+
+SERVER_LOG=/workspace/server.log
+
+if [ "${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..0d6577d59 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -119,6 +119,61 @@ dsr1-fp8-mi325x-sglang:
       search-space:
       - { tp: 8, conc-start: 4, conc-end: 64 }
 
+# DeepSeek-V4-Pro FP8 single-node on MI300X (gfx942) via vLLM.
+# EXTRAPOLATED bring-up. sglang was abandoned: on gfx942 (no native FP4) the
+# dsv4 sglang backend's nvfp4 MoE / TileLang-MLA kernels have no gfx942 build
+# (gfx950/MI355X only). vLLM runs the checkpoint in FP8 via --quantization
+# deepseek_v4_fp8 (dequant FP4 MoE -> FP8), the same path the H200 dsv4 vLLM
+# recipe uses. Config mirrors the same-model dsv4-fp4-mi355x-vllm (TP8, conc
+# 4-512); the FP4->FP8 dequant (~1.05TB) fits 8x192GB only at TP8. Launch
+# script dsv4_fp8_mi300x.sh carries the deepseek_v4 + gfx942 AITER flags.
+dsv4-fp8-mi300x-vllm:
+  image: vllm/vllm-openai-rocm:nightly-09663abde0f50944a8d5ea30120666024b503faa
+  model: deepseek-ai/DeepSeek-V4-Pro
+  model-prefix: dsv4
+  runner: mi300x
+  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-mi300x-vllm. Mirrors the base recipe and adds
+# DeepSeek-V4 built-in MTP via --speculative-config (num_speculative_tokens=2),
+# routing to dsv4_fp8_mi300x_mtp.sh; benchmark uses --dsv4 chat-template
+# encoding (required for meaningful MTP acceptance).
+dsv4-fp8-mi300x-vllm-mtp:
+  image: vllm/vllm-openai-rocm:nightly-09663abde0f50944a8d5ea30120666024b503faa
+  model: deepseek-ai/DeepSeek-V4-Pro
+  model-prefix: dsv4
+  runner: mi300x
+  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 failed here
+      # (19.2% req failures) and conc512 was 55.8%; conc128 passed cleanly.
+      # 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 e05c2ca1c..55228a17c 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -4751,6 +4751,22 @@
     - "6 topologies across 1k/1k and 8k/1k: 1P1D TP4 STP + wide-EP (DEP4 prefill / DEP16 decode) from 1P1D up to 8P1D, recipes under benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb300-fp8/"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2137
 
+- config-keys:
+    - dsv4-fp8-h200-vllm
+    - dsv4-fp8-h200-vllm-mtp
+  description:
+    - "Bump vLLM image from v0.21.0 to v0.25.0 for DeepSeek-V4-Pro FP8 on H200, matching the B200/B300 dsv4 vLLM bump (#2169)"
+    - "Refresh stale H200 dsv4 submissions (last run 2026-05-21)"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2191
+
+- config-keys:
+    - dsv4-fp8-h200-sglang
+    - dsv4-fp8-h200-sglang-mtp
+  description:
+    - "Bump the pinned lmsysorg/sglang:deepseek-v4-hopper digest from the 2026-05-02 push (7f19c6dc) to the current 2026-05-13 push (1bf5d508)"
+    - "Refresh stale H200 dsv4 SGLang submissions (last run 2026-05-04)"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2191
+
 - config-keys:
     - dsv4-fp4-mi355x-sglang
   description:
@@ -4765,3 +4781,22 @@
     - "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-mi300x-vllm
+  description:
+    - "Add DeepSeek-V4-Pro FP8 single-node MI300X vLLM benchmark (new SKU, previously no dsv4 data on gfx942)"
+    - "vLLM, not sglang: on gfx942 (no native FP4) the dsv4 sglang backend's nvfp4 MoE and TileLang-MLA kernels have no gfx942 build (gfx950/MI355X only). vLLM runs the checkpoint in FP8 via --quantization deepseek_v4_fp8 (dequant FP4 MoE -> FP8), the same path the no-FP4 H200 dsv4 vLLM recipe uses"
+    - "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-mi300x-vllm; TP8 conc 4-512 (the FP4->FP8 dequant ~1.05TB fits 8x192GB only at TP8)"
+    - "MoE backend left at auto (unlike MI355X's --moe-backend aiter): deepseek_v4_fp8 only dequantizes dense/attention, the MoE experts stay mxfp4; forcing aiter on gfx942 selects AITER_MXFP4_MXFP4 (W4A4 native-mxfp4) which the gfx942 kernel rejects, while auto's ROCm+DeepseekV4 path prefers AITER_MXFP4_BF16 (W4A16, dequant) with TRITON_UNFUSED fallback"
+    - "EVAL_CONCURRENT_REQUESTS=8: FP8 weights (~131GB/GPU) leave only ~71k tokens of KV on 8x192GB (max 7.52x concurrency for a 9472-token request); the eval default of CONC (128) concurrent requests OOM-kills the server mid-gsm8k, so eval concurrency is capped to the KV budget (throughput unaffected; MI325X's 256GB keeps the default)"
+    - "Image vllm/vllm-openai-rocm:nightly-09663abde0f50944a8d5ea30120666024b503faa (the DSv4-validated ROCm vLLM build)"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2194
+
+- config-keys:
+    - dsv4-fp8-mi300x-vllm-mtp
+  description:
+    - "Add DeepSeek-V4-Pro FP8 MI300X vLLM MTP (speculative-decoding) variant, mirroring dsv4-fp8-mi300x-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 (EAGLE/MTP is trained on chat-formatted inputs)"
+    - "Same gfx942 auto-MoE (AITER_MXFP4_BF16) + EVAL_CONCURRENT_REQUESTS=8 KV cap as the base recipe; routes to dsv4_fp8_mi300x_mtp.sh via the launcher's spec-decoding=mtp suffix"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2194