diff --git a/.github/configs/nvidia-master.yaml b/.github/configs/nvidia-master.yaml
index 1d467308f..e24408017 100644
--- a/.github/configs/nvidia-master.yaml
+++ b/.github/configs/nvidia-master.yaml
@@ -1832,9 +1832,10 @@ dsr1-fp8-b300-sglang:
     - { tp: 8, ep: 1, conc-start: 4, conc-end: 4 }
     - { tp: 4, ep: 1, conc-start: 4, conc-end: 32 }
 
-# NOTE: Low-latency fallback (TP=8, EP=1, no DP-attn, no DeepEP) while
-# the DeepEP FP8 weight-postprocess path is broken for DeepSeek-V4-Pro
-# on B300. Re-introduce balanced/max-throughput rows once fixed upstream.
+# NOTE: https://docs.sglang.io/cookbook/autoregressive/DeepSeek/DeepSeek-V4
+# lists B200 (not B300) as the Blackwell target; we reuse the B200 Pro FP4
+# recipes on B300 until a B300-specific recipe ships. Prefix caching is
+# disabled. Parallelisms mirror dsv4-fp4-b200-sglang.
 dsv4-fp4-b300-sglang:
   image: lmsysorg/sglang:deepseek-v4-b300
   model: deepseek-ai/DeepSeek-V4-Pro
@@ -1843,22 +1844,63 @@ dsv4-fp4-b300-sglang:
   precision: fp4
   framework: sglang
   multinode: false
-  # TODO(Cam): low-latency recipe only (TP-only, no DP-attn, no DeepEP)
-  # while the DeepEP FP8 weight-postprocess path is broken for this
-  # checkpoint on B300 (RuntimeError: Recipe must be a list/tuple of 3
-  # integers. raised from sglang.srt.layers.quantization.fp8
-  # .process_weights_after_loading_block_quant). Full concurrency sweep
-  # retained; revert to the recipe-per-CONC split on chore/dsv4-sgl-b300
-  # once sglang can load the checkpoint under --moe-a2a-backend deepep.
+  # Three recipes from https://docs.sglang.io/cookbook/autoregressive/DeepSeek/DeepSeek-V4
+  # are selected inside benchmarks/single_node/dsv4_fp4_b300.sh by CONC:
+  #   low-latency    (CONC <= 32):       TP-only
+  #   balanced       (32 < CONC <= 128): + DP-attn
+  #   max-throughput (CONC > 128):       + DP-attn
+  # Split so result filenames (ep=, dpa=) accurately reflect the recipe.
+  # ep is implicit in sglang: --moe-a2a-backend deepep forces ep_size=tp_size,
+  # while low-latency leaves ep_size at the default of 1.
+  seq-len-configs:
+  - isl: 1024
+    osl: 1024
+    search-space:
+    # low-latency
+    - { tp: 8, ep: 1, conc-start: 4, conc-end: 32 }
+    # balanced
+    - { tp: 8, ep: 8, dp-attn: true, conc-start: 64, conc-end: 128 }
+    # max-throughput
+    - { tp: 8, ep: 8, dp-attn: true, conc-start: 256, conc-end: 1024 }
+  - isl: 8192
+    osl: 1024
+    search-space:
+    # low-latency
+    - { tp: 8, ep: 1, conc-start: 4, conc-end: 32 }
+    # balanced
+    - { tp: 8, ep: 8, dp-attn: true, conc-start: 64, conc-end: 128 }
+    # max-throughput
+    - { tp: 8, ep: 8, dp-attn: true, conc-start: 256, conc-end: 512 }
+
+dsv4-fp4-b300-sglang-mtp:
+  image: lmsysorg/sglang:deepseek-v4-b300
+  model: deepseek-ai/DeepSeek-V4-Pro
+  model-prefix: dsv4
+  runner: b300
+  precision: fp4
+  framework: sglang
+  multinode: false
+  # Mirrors dsv4-fp4-b300-sglang's low-latency and balanced rows with EAGLE
+  # MTP enabled per https://docs.sglang.io/cookbook/autoregressive/DeepSeek/DeepSeek-V4:
+  #   low-latency  (CONC <= 32):       EAGLE 3 steps / 4 draft tokens
+  #   balanced     (32 < CONC <= 128): EAGLE 1 step  / 2 draft tokens
+  # Max-throughput is intentionally omitted -- the cookbook says MTP off
+  # at saturation because the verify step costs more than it saves.
   seq-len-configs:
   - isl: 1024
     osl: 1024
     search-space:
-    - { tp: 8, ep: 1, conc-start: 4, conc-end: 1024 }
+    # low-latency
+    - { tp: 8, ep: 1, conc-start: 4, conc-end: 32, spec-decoding: mtp }
+    # balanced
+    - { tp: 8, ep: 8, dp-attn: true, conc-start: 64, conc-end: 128, spec-decoding: mtp }
   - isl: 8192
     osl: 1024
     search-space:
-    - { tp: 8, ep: 1, conc-start: 4, conc-end: 512 }
+    # low-latency
+    - { tp: 8, ep: 1, conc-start: 4, conc-end: 32, spec-decoding: mtp }
+    # balanced
+    - { tp: 8, ep: 8, dp-attn: true, conc-start: 64, conc-end: 128, spec-decoding: mtp }
 
 qwen3.5-bf16-b200-sglang:
   image: lmsysorg/sglang:nightly-dev-20260216-d3bae71e
diff --git a/benchmarks/single_node/dsv4_fp4_b300.sh b/benchmarks/single_node/dsv4_fp4_b300.sh
new file mode 100755
index 000000000..faa946174
--- /dev/null
+++ b/benchmarks/single_node/dsv4_fp4_b300.sh
@@ -0,0 +1,129 @@
+#!/usr/bin/env bash
+
+source "$(dirname "$0")/../benchmark_lib.sh"
+
+check_env_vars \
+    MODEL \
+    TP \
+    CONC \
+    ISL \
+    OSL \
+    RANDOM_RANGE_RATIO \
+    RESULT_FILENAME
+
+if [[ -n "$SLURM_JOB_ID" ]]; then
+  echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
+fi
+
+# The B300 runner overrides MODEL to a pre-staged /data/models path, so skip
+# `hf download`. Only fetch when MODEL looks like a HF repo ID.
+if [[ "$MODEL" != /* ]]; then
+    hf download "$MODEL"
+fi
+
+nvidia-smi
+
+export SGLANG_JIT_DEEPGEMM_PRECOMPILE=0
+
+# The deepseek-v4 sglang images (lmsysorg/sglang:deepseek-v4-blackwell and its
+# B300 forks) bake CUDA_VISIBLE_DEVICES=4,5,6,7 into their ENV, which masks half
+# of the 8 GPUs Slurm allocates us. Clear it so TP=8 can bind to all ranks.
+unset CUDA_VISIBLE_DEVICES
+
+# TODO(Cam): the deepseek-v4 sglang images install sglang editable at
+# /workspace/sglang/python; prior sglang tags used /sgl-workspace/sglang.
+# The runner mounts our repo at a non-/workspace path for these images so the
+# editable install stays visible. Paths in this script are $PWD-relative for
+# that reason. Drop the runner conditional once lmsys moves sglang back out of
+# /workspace.
+
+SERVER_LOG="$PWD/server.log"
+PORT=${PORT:-8888}
+
+echo "TP: $TP, CONC: $CONC, ISL: $ISL, OSL: $OSL"
+
+EVAL_CONTEXT_ARGS=""
+if [ "${EVAL_ONLY}" = "true" ]; then
+    setup_eval_context
+    EVAL_CONTEXT_ARGS="--context-length $EVAL_MAX_MODEL_LEN"
+fi
+
+start_gpu_monitor --output "$PWD/gpu_metrics.csv"
+
+# Three recipes from https://docs.sglang.io/cookbook/autoregressive/DeepSeek/DeepSeek-V4
+# (spec-decoding / MTP and prefix-caching flags dropped for the baseline):
+#   - low-latency    (CONC <= 32):        TP-only, chunked-prefill, disable autotune
+#   - balanced       (32 < CONC <= 128):  + DP-attn, max-running-requests=128
+#   - max-throughput (CONC > 128):        + DP-attn, max-running-requests=256
+DEEPEP_CONFIG='{"normal_dispatch":{"num_sms":96},"normal_combine":{"num_sms":96}}'
+
+if [[ $CONC -le 32 ]]; then
+    RECIPE=low-latency
+    RECIPE_FLAGS=(
+        --moe-runner-backend flashinfer_mxfp4
+        --chunked-prefill-size 4096
+        --disable-flashinfer-autotune
+        --mem-fraction-static 0.82
+    )
+elif [[ $CONC -le 128 ]]; then
+    RECIPE=balanced
+    export SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=256
+    RECIPE_FLAGS=(
+        --dp-size "$TP"
+        --enable-dp-attention
+        --moe-a2a-backend deepep
+        --deepep-config "$DEEPEP_CONFIG"
+        --mem-fraction-static 0.82
+        --cuda-graph-max-bs 64
+        --max-running-requests 128
+    )
+else
+    RECIPE=max-throughput
+    export SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=256
+    RECIPE_FLAGS=(
+        --dp-size "$TP"
+        --enable-dp-attention
+        --moe-a2a-backend deepep
+        --deepep-config "$DEEPEP_CONFIG"
+        --mem-fraction-static 0.82
+        --cuda-graph-max-bs 64
+        --max-running-requests 256
+    )
+fi
+echo "Recipe: $RECIPE (CONC=$CONC)"
+
+set -x
+PYTHONNOUSERSITE=1 sglang serve \
+    --model-path $MODEL \
+    --host 0.0.0.0 \
+    --port $PORT \
+    --trust-remote-code \
+    --tp $TP \
+    --disable-radix-cache \
+    "${RECIPE_FLAGS[@]}" $EVAL_CONTEXT_ARGS > $SERVER_LOG 2>&1 &
+
+SERVER_PID=$!
+
+wait_for_server_ready --port "$PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
+
+pip install -q datasets pandas
+
+run_benchmark_serving \
+    --model "$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 "$RESULT_FILENAME" \
+    --result-dir "$PWD/"
+
+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/dsv4_fp4_b300_mtp.sh b/benchmarks/single_node/dsv4_fp4_b300_mtp.sh
new file mode 100755
index 000000000..2e383ead2
--- /dev/null
+++ b/benchmarks/single_node/dsv4_fp4_b300_mtp.sh
@@ -0,0 +1,130 @@
+#!/usr/bin/env bash
+
+source "$(dirname "$0")/../benchmark_lib.sh"
+
+check_env_vars \
+    MODEL \
+    TP \
+    CONC \
+    ISL \
+    OSL \
+    RANDOM_RANGE_RATIO \
+    RESULT_FILENAME
+
+if [[ -n "$SLURM_JOB_ID" ]]; then
+  echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
+fi
+
+# The B300 runner overrides MODEL to a pre-staged /data/models path, so skip
+# `hf download`. Only fetch when MODEL looks like a HF repo ID.
+if [[ "$MODEL" != /* ]]; then
+    hf download "$MODEL"
+fi
+
+nvidia-smi
+
+export SGLANG_JIT_DEEPGEMM_PRECOMPILE=0
+# Cookbook note: "MTP currently requires SGLANG_ENABLE_SPEC_V2=1."
+export SGLANG_ENABLE_SPEC_V2=1
+
+# The deepseek-v4 sglang images (lmsysorg/sglang:deepseek-v4-blackwell and its
+# B300 forks) bake CUDA_VISIBLE_DEVICES=4,5,6,7 into their ENV, which masks half
+# of the 8 GPUs Slurm allocates us. Clear it so TP=8 can bind to all ranks.
+unset CUDA_VISIBLE_DEVICES
+
+# TODO(Cam): the deepseek-v4 sglang images install sglang editable at
+# /workspace/sglang/python; prior sglang tags used /sgl-workspace/sglang.
+# The runner mounts our repo at a non-/workspace path for these images so the
+# editable install stays visible. Paths in this script are $PWD-relative for
+# that reason. Drop the runner conditional once lmsys moves sglang back out of
+# /workspace.
+
+SERVER_LOG="$PWD/server.log"
+PORT=${PORT:-8888}
+
+echo "TP: $TP, CONC: $CONC, ISL: $ISL, OSL: $OSL"
+
+EVAL_CONTEXT_ARGS=""
+if [ "${EVAL_ONLY}" = "true" ]; then
+    setup_eval_context
+    EVAL_CONTEXT_ARGS="--context-length $EVAL_MAX_MODEL_LEN"
+fi
+
+start_gpu_monitor --output "$PWD/gpu_metrics.csv"
+
+# Two recipes from https://docs.sglang.io/cookbook/autoregressive/DeepSeek/DeepSeek-V4
+# with EAGLE / MTP enabled per the cookbook (prefix-caching dropped):
+#   - low-latency (CONC <= 32):       TP-only + EAGLE 3 steps / 4 draft tokens
+#   - balanced    (32 < CONC <= 128): + DP-attn + EAGLE 1 step / 2 draft tokens
+# Max-throughput is intentionally not handled here -- the cookbook says
+# MTP off at saturation because the verify step costs more than it saves.
+# dsv4-fp4-b300-sglang-mtp's search-space caps CONC at 128 to match.
+DEEPEP_CONFIG='{"normal_dispatch":{"num_sms":96},"normal_combine":{"num_sms":96}}'
+
+if [[ $CONC -le 32 ]]; then
+    RECIPE=low-latency
+    RECIPE_FLAGS=(
+        --moe-runner-backend flashinfer_mxfp4
+        --chunked-prefill-size 4096
+        --disable-flashinfer-autotune
+        --mem-fraction-static 0.82
+        --speculative-algo EAGLE
+        --speculative-num-steps 3
+        --speculative-eagle-topk 1
+        --speculative-num-draft-tokens 4
+    )
+else
+    RECIPE=balanced
+    export SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=256
+    RECIPE_FLAGS=(
+        --dp-size "$TP"
+        --enable-dp-attention
+        --moe-a2a-backend deepep
+        --deepep-config "$DEEPEP_CONFIG"
+        --mem-fraction-static 0.82
+        --cuda-graph-max-bs 64
+        --max-running-requests 128
+        --speculative-algo EAGLE
+        --speculative-num-steps 1
+        --speculative-eagle-topk 1
+        --speculative-num-draft-tokens 2
+    )
+fi
+echo "Recipe: $RECIPE (CONC=$CONC)"
+
+set -x
+PYTHONNOUSERSITE=1 sglang serve \
+    --model-path $MODEL \
+    --host 0.0.0.0 \
+    --port $PORT \
+    --trust-remote-code \
+    --tp $TP \
+    --disable-radix-cache \
+    "${RECIPE_FLAGS[@]}" $EVAL_CONTEXT_ARGS > $SERVER_LOG 2>&1 &
+
+SERVER_PID=$!
+
+wait_for_server_ready --port "$PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
+
+pip install -q datasets pandas
+
+run_benchmark_serving \
+    --model "$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 "$RESULT_FILENAME" \
+    --result-dir "$PWD/" \
+    --use-chat-template
+
+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/perf-changelog.yaml b/perf-changelog.yaml
index a6c811748..64ff305f8 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -1812,3 +1812,26 @@
     - "Topologies: low-conc 1p1d-dep8-tep8 (4 nodes, mirrored from NVIDIA srt-slurm PR #71 with offload kept and numa-bind dropped); mid 1p1d-dep8-dep16 (6 nodes) and high 3p1d-dep8-dep16 (10 nodes) hand-rolled, structurally derived from the kimi-k2.5 1k/1k pattern"
     - "Recipes stored under benchmarks/multi_node/srt-slurm-recipes/ and overlaid onto the upstream srt-slurm checkout at runtime"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1129
+  
+- config-keys:
+    - dsv4-fp4-b300-sglang
+  description:
+    - "Add DeepSeek-V4-Pro FP4 B300 SGLang benchmark"
+    - "Image: lmsysorg/sglang:deepseek-v4-blackwell"
+    - "Model: deepseek-ai/DeepSeek-V4-Pro (FP4 MoE experts + FP8 attention/dense)"
+    - "Reuses the B200 Pro Max-Throughput recipe from https://docs.sglang.io/cookbook/autoregressive/DeepSeek/DeepSeek-V4 on B300 until a B300-specific recipe ships"
+    - "DP=8 + DeepEP, prefix caching disabled, no speculative decoding"
+    - "Parallelism (TP=8/EP=8/dp-attn=true) and concurrency ranges (4-1024 for 1k1k, 4-512 for 8k1k) mirror dsv4-fp4-b200-vllm"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1132
+
+- config-keys:
+    - dsv4-fp4-b300-sglang-mtp
+  description:
+    - "Add DeepSeek-V4-Pro FP4 B300 SGLang MTP benchmark"
+    - "Image: lmsysorg/sglang:deepseek-v4-b300"
+    - "Model: deepseek-ai/DeepSeek-V4-Pro (FP4 MoE experts + FP8 attention/dense)"
+    - "Mirrors dsv4-fp4-b300-sglang's low-latency and balanced rows with EAGLE MTP enabled per https://docs.sglang.io/cookbook/autoregressive/DeepSeek/DeepSeek-V4"
+    - "EAGLE 3 steps / 4 draft tokens on low-latency, 1 step / 2 draft tokens on balanced"
+    - "Max-throughput intentionally omitted: cookbook says MTP off at saturation because the verify step costs more than it saves"
+    - "SGLANG_ENABLE_SPEC_V2=1 required for MTP; --use-chat-template passed to bench_serving"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1151