diff --git a/.github/configs/nvidia-master.yaml b/.github/configs/nvidia-master.yaml
index 39e299cb0..5ac7a1074 100644
--- a/.github/configs/nvidia-master.yaml
+++ b/.github/configs/nvidia-master.yaml
@@ -7714,3 +7714,122 @@ dsv4-fp4-gb200-dynamo-vllm:
         tp: 16
         ep: 16
         dp-attn: true
+
+dsv4-fp4-gb200-dynamo-sglang:
+  image: lmsysorg/sglang:deepseek-v4-grace-blackwell
+  model: deepseek-ai/DeepSeek-V4-Pro
+  model-prefix: dsv4
+  runner: gb200
+  precision: fp4
+  framework: dynamo-sglang
+  multinode: true
+  disagg: true
+  seq-len-configs:
+  # 1k/1k — TP=8 (2 GB200 nodes per worker) with DP-attention but no
+  # DeepEP. The lmsysorg/sglang:deepseek-v4-grace-blackwell image's
+  # sglang fork has a fork-only mxfp4_deepseek kernel that crashes any
+  # DeepEP forward path (both DeepEPLLDispatchOutput and
+  # DeepEPNormalDispatchOutput lack the `topk_output` field the kernel
+  # reads). At TP=8 the shared-experts gate_up_proj would also fail
+  # FP8 block-quant divisibility (1536/8=192, not divisible by 128)
+  # unless `moe-dense-tp-size: 1` runs the dense MLP layers replicated
+  # — and that flag is gated on `enable_dp_attention=True` in sglang
+  # dp_attention.py. So: DP-attention on; `moe-a2a-backend` left at
+  # its default `"none"` — sglang `forward_normal` path runs (verified
+  # in deepseek_v2.py: `_enable_a2a_moe` is False unless backend is
+  # deepep|mooncake|nixl|mori|ascend_fuseep|flashinfer). Filenames keep
+  # the historical 'dep8'/'dep16' tag for symmetry with the dsv4-fp4-
+  # gb200-dynamo-vllm sibling; the actual recipe is TP=8 + DP=8 with
+  # all-reduce/all-gather MoE dispatch.
+  - isl: 1024
+    osl: 1024
+    search-space:
+    # Low-concurrency: 1 prefill (TP=8) + 1 decode (TP=8). 4 nodes.
+    - conc-list: [1, 4, 8, 16, 32, 64]
+      prefill:
+        num-worker: 1
+        tp: 8
+        ep: 1
+        dp-attn: true
+        additional-settings:
+        - "CONFIG_FILE=recipes/sglang/deepseek-v4/1k1k/disagg-gb200-1p1d-dep8-tep8.yaml"
+      decode:
+        num-worker: 1
+        tp: 8
+        ep: 1
+        dp-attn: true
+    # Mid throughput: 1 prefill (TP=8) + 1 decode (TP=8). 4 nodes.
+    - conc-list: [128, 256, 1024, 2048, 4096]
+      prefill:
+        num-worker: 1
+        tp: 8
+        ep: 1
+        dp-attn: true
+        additional-settings:
+        - "CONFIG_FILE=recipes/sglang/deepseek-v4/1k1k/disagg-gb200-1p1d-dep8-dep16.yaml"
+      decode:
+        num-worker: 1
+        tp: 8
+        ep: 1
+        dp-attn: true
+    # High throughput: 3 prefills (TP=8) + 1 decode (TP=8). 8 nodes.
+    # 4096 overlap with the 1p1d block gives a prefill-scaling A/B.
+    - conc-list: [4096, 8192]
+      prefill:
+        num-worker: 3
+        tp: 8
+        ep: 1
+        dp-attn: true
+        additional-settings:
+        - "CONFIG_FILE=recipes/sglang/deepseek-v4/1k1k/disagg-gb200-3p1d-dep8-dep16.yaml"
+      decode:
+        num-worker: 1
+        tp: 8
+        ep: 1
+        dp-attn: true
+
+  - isl: 8192
+    osl: 1024
+    search-space:
+    # Low-concurrency: 1 prefill (TP=8) + 1 decode (TP=8). 4 nodes.
+    - conc-list: [1, 4, 8, 16, 32, 64]
+      prefill:
+        num-worker: 1
+        tp: 8
+        ep: 1
+        dp-attn: true
+        additional-settings:
+        - "CONFIG_FILE=recipes/sglang/deepseek-v4/8k1k/disagg-gb200-1p1d-dep8-tep8.yaml"
+      decode:
+        num-worker: 1
+        tp: 8
+        ep: 1
+        dp-attn: true
+    # Mid: 3 prefills (TP=8) + 1 decode (TP=8). 8 nodes.
+    - conc-list: [512, 1024]
+      prefill:
+        num-worker: 3
+        tp: 8
+        ep: 1
+        dp-attn: true
+        additional-settings:
+        - "CONFIG_FILE=recipes/sglang/deepseek-v4/8k1k/disagg-gb200-3p1d-dep8-dep16.yaml"
+      decode:
+        num-worker: 1
+        tp: 8
+        ep: 1
+        dp-attn: true
+    # Max throughput: 7 prefills (TP=8) + 1 decode (TP=8). 16 nodes.
+    - conc-list: [4096, 8192]
+      prefill:
+        num-worker: 7
+        tp: 8
+        ep: 1
+        dp-attn: true
+        additional-settings:
+        - "CONFIG_FILE=recipes/sglang/deepseek-v4/8k1k/disagg-gb200-7p1d-dep8-dep16.yaml"
+      decode:
+        num-worker: 1
+        tp: 8
+        ep: 1
+        dp-attn: true
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/1k1k/disagg-gb200-1p1d-dep8-dep16.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/1k1k/disagg-gb200-1p1d-dep8-dep16.yaml
new file mode 100644
index 000000000..d309562a1
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/1k1k/disagg-gb200-1p1d-dep8-dep16.yaml
@@ -0,0 +1,113 @@
+name: "dsv4-sglang-disagg-gb200-1p1d-dep8-dep16"
+
+# Hand-rolled — see ./disagg-gb200-1p1d-dep8-tep8.yaml header for the
+# upstream-reference list (PR #69 GB200 agg, PR #75 GB300 disagg).
+# Topology mirrors the dsv4-fp4-gb200-dynamo-vllm sibling.
+#
+# Topology: 1 prefill (DP=8 EP=8) + 1 decode (DP=16 EP=16). 6 nodes.
+# Single prefill is enough for 1k prompts up to ~conc 4096 (per-rank
+# prefill TFlops at 1k ISL is high; matches the vLLM sibling sizing).
+
+model:
+  path: "deepseek-v4-pro"
+  container: "lmsysorg/sglang:deepseek-v4-grace-blackwell"
+  precision: "fp4"
+
+# See ./disagg-gb200-1p1d-dep8-tep8.yaml for the dynamo pin rationale.
+dynamo:
+  hash: 21f135f5edf40e12e6ff5db2b462d862a6d6ab9b
+  install: true
+
+slurm:
+  time_limit: "8:00:00"
+
+health_check:
+  max_attempts: 1440
+  interval_seconds: 10
+
+resources:
+  gpu_type: "gb200"
+  gpus_per_node: 4
+  prefill_nodes: 2
+  decode_nodes: 2
+  prefill_workers: 1
+  decode_workers: 1
+  gpus_per_prefill: 8
+  gpus_per_decode: 8
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+
+backend:
+  type: sglang
+
+  prefill_environment:
+    PYTHONUNBUFFERED: "1"
+    SGLANG_JIT_DEEPGEMM_PRECOMPILE: "0"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: "100000"
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: "100000"
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "100000"
+    SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK: "1024"
+
+  decode_environment:
+    PYTHONUNBUFFERED: "1"
+    SGLANG_JIT_DEEPGEMM_PRECOMPILE: "0"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: "100000"
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: "100000"
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "100000"
+    SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK: "1024"
+
+  sglang_config:
+    prefill:
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      trust-remote-code: true
+      tensor-parallel-size: 8
+      moe-dense-tp-size: 1
+      enable-dp-attention: true
+      dp-size: 8
+      moe-runner-backend: "flashinfer_mxfp4"
+      chunked-prefill-size: 4096
+      disable-flashinfer-autotune: true
+      disable-radix-cache: true
+      mem-fraction-static: 0.82
+      context-length: 3072
+      max-running-requests: 16
+      stream-interval: 50
+      decode-log-interval: 1000
+      disaggregation-mode: "prefill"
+      disaggregation-bootstrap-port: 30001
+      disaggregation-transfer-backend: nixl
+
+    decode:
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      trust-remote-code: true
+      tensor-parallel-size: 8
+      moe-dense-tp-size: 1
+      enable-dp-attention: true
+      dp-size: 8
+      moe-runner-backend: "flashinfer_mxfp4"
+      chunked-prefill-size: 4096
+      disable-flashinfer-autotune: true
+      disable-radix-cache: true
+      mem-fraction-static: 0.82
+      context-length: 3072
+      max-running-requests: 512
+      cuda-graph-max-bs: 512
+      stream-interval: 50
+      decode-log-interval: 1000
+      disaggregation-mode: "decode"
+      disaggregation-bootstrap-port: 30001
+      disaggregation-transfer-backend: nixl
+
+benchmark:
+  type: "sa-bench"
+  isl: 1024
+  osl: 1024
+  concurrencies: "128x256x1024x2048x4096"
+  req_rate: "inf"
+  use_chat_template: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/1k1k/disagg-gb200-1p1d-dep8-tep8.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/1k1k/disagg-gb200-1p1d-dep8-tep8.yaml
new file mode 100644
index 000000000..e20c9c0a2
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/1k1k/disagg-gb200-1p1d-dep8-tep8.yaml
@@ -0,0 +1,153 @@
+name: "dsv4-sglang-disagg-gb200-1p1d-dep8-tep8"
+
+# Hand-rolled — no GB200 DSV4 sglang disagg recipe exists upstream. The
+# closest references on NVIDIA/srt-slurm are:
+#   * PR #69 (recipes/gb200-fp4/1k1k-dsv4/agg-2n-low-latency.yaml) —
+#     GB200 DSV4 sglang AGGREGATED: per-worker flag set + env vars.
+#   * PR #75 (recipes/gb300-fp4/1k1k-dsv4/disagg-1p1d-tp4-mxfp4.yaml) —
+#     GB300 DSV4 sglang DISAGG: confirms nixl + flashinfer_mxfp4 +
+#     chunked-prefill-size=4096 + disable-flashinfer-autotune.
+# Topology mirrors the dsv4-fp4-gb200-dynamo-vllm sibling so cross-
+# framework numbers stay directly comparable.
+#
+# Topology: 1 prefill (TP=8 / DP=8) + 1 decode (TP=8 / DP=8). 4 nodes.
+# Targets very low concurrency (1-64).
+#
+# Why TP=8 + DP-attention but NO `moe-a2a-backend` (default "none"):
+#   1. DSV4-Pro at MXFP4 is too large for TP=4 single-node — OOM.
+#      TP=8 across 2 GB200 nodes (8 GPUs * 96 GB = 768 GB) fits.
+#   2. The lmsysorg/sglang:deepseek-v4-grace-blackwell sglang fork
+#      ships a fork-only quant kernel `mxfp4_deepseek.py` that reads
+#      `dispatch_output.topk_output`. Neither `DeepEPLLDispatchOutput`
+#      nor `DeepEPNormalDispatchOutput` exposes that field in this
+#      fork, so `forward_deepep` always crashes the prefill scheduler.
+#      We must stay off the DeepEP path.
+#   3. At TP=8 the shared-experts gate_up_proj fails FP8 block-quant
+#      divisibility (1536/8=192, not divisible by block_n=128).
+#      `moe-dense-tp-size: 1` runs the dense MLP layers replicated
+#      (TP=1) so the divisibility check passes — but that flag is
+#      gated on `enable_dp_attention=True` in sglang
+#      `python/sglang/srt/layers/dp_attention.py`
+#      (`compute_dp_attention_local_info` returns the full `tp_size`
+#      and ignores `moe_dense_tp_size` when DP-attn is off).
+# So: `enable-dp-attention: true` + `dp-size: 8` (DP-attn active so
+# `moe-dense-tp-size: 1` takes effect) AND no `moe-a2a-backend` set.
+# The default `"none"` lands the MoE on `forward_normal` instead of
+# `forward_deepep` — verified in deepseek_v2.py:
+#   `_enable_a2a_moe = is_deepep|is_mooncake|is_nixl|is_mori|
+#    is_ascend_fuseep|is_flashinfer`  → False with default.
+
+model:
+  path: "deepseek-v4-pro"
+  container: "lmsysorg/sglang:deepseek-v4-grace-blackwell"
+  precision: "fp4"
+
+# Pin dynamo to the v1.2.0-sglang-deepseek-v4-dev.1 tag. The PyPI
+# 0.8.0/0.8.1 releases (srtctl's default) reference `sgl.Engine` in
+# `dynamo.sglang.health_check` *eagerly* (no `from __future__ import
+# annotations`), and the lmsysorg/sglang:deepseek-v4-grace-blackwell
+# image's sglang fork does not expose `sgl.Engine`, so they crash at
+# import with `AttributeError: module 'sglang' has no attribute
+# 'Engine'`. The DSV4-targeted tag adds `from __future__ import
+# annotations` (commit cdb7218a, ai-dynamo PR #7255), making the
+# annotation lazy so the module imports cleanly.
+dynamo:
+  hash: 21f135f5edf40e12e6ff5db2b462d862a6d6ab9b
+  install: true
+
+slurm:
+  time_limit: "8:00:00"
+
+health_check:
+  max_attempts: 1440
+  interval_seconds: 10
+
+resources:
+  gpu_type: "gb200"
+  gpus_per_node: 4
+  prefill_nodes: 2
+  decode_nodes: 2
+  prefill_workers: 1
+  decode_workers: 1
+  gpus_per_prefill: 8
+  gpus_per_decode: 8
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+
+backend:
+  type: sglang
+
+  # Env var set mirrored from PR #69 (the GB200 DSV4 aggregated baseline
+  # that's actually been run upstream) plus the disaggregation timeout
+  # triple — heartbeat 100k matches the DSR1 sglang disagg convention.
+  prefill_environment:
+    PYTHONUNBUFFERED: "1"
+    SGLANG_JIT_DEEPGEMM_PRECOMPILE: "0"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: "100000"
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: "100000"
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "100000"
+    SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK: "1024"
+
+  decode_environment:
+    PYTHONUNBUFFERED: "1"
+    SGLANG_JIT_DEEPGEMM_PRECOMPILE: "0"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: "100000"
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: "100000"
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "100000"
+    SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK: "1024"
+
+  sglang_config:
+    prefill:
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      trust-remote-code: true
+      tensor-parallel-size: 8
+      moe-dense-tp-size: 1
+      enable-dp-attention: true
+      dp-size: 8
+      moe-runner-backend: "flashinfer_mxfp4"
+      chunked-prefill-size: 4096
+      disable-flashinfer-autotune: true
+      disable-radix-cache: true
+      mem-fraction-static: 0.82
+      context-length: 3072
+      max-running-requests: 16
+      stream-interval: 50
+      decode-log-interval: 1000
+      disaggregation-mode: "prefill"
+      disaggregation-bootstrap-port: 30001
+      disaggregation-transfer-backend: nixl
+
+    decode:
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      trust-remote-code: true
+      tensor-parallel-size: 8
+      moe-dense-tp-size: 1
+      enable-dp-attention: true
+      dp-size: 8
+      moe-runner-backend: "flashinfer_mxfp4"
+      chunked-prefill-size: 4096
+      disable-flashinfer-autotune: true
+      disable-radix-cache: true
+      mem-fraction-static: 0.82
+      context-length: 3072
+      max-running-requests: 64
+      cuda-graph-max-bs: 64
+      stream-interval: 50
+      decode-log-interval: 1000
+      disaggregation-mode: "decode"
+      disaggregation-bootstrap-port: 30001
+      disaggregation-transfer-backend: nixl
+
+benchmark:
+  type: "sa-bench"
+  isl: 1024
+  osl: 1024
+  concurrencies: "1x4x8x16x32x64"
+  req_rate: "inf"
+  use_chat_template: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/1k1k/disagg-gb200-3p1d-dep8-dep16.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/1k1k/disagg-gb200-3p1d-dep8-dep16.yaml
new file mode 100644
index 000000000..a8a161798
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/1k1k/disagg-gb200-3p1d-dep8-dep16.yaml
@@ -0,0 +1,123 @@
+name: "dsv4-sglang-disagg-gb200-3p1d-dep8-dep16"
+
+# High-concurrency 4096/8192 topology. Same TP=8 + DP-attn + no-EP
+# shape as the 1p1d siblings — see ./disagg-gb200-1p1d-dep8-tep8.yaml
+# header for the full constraint chain.
+#
+# Both EP backends available upstream (deepep, flashinfer) are dead on
+# this image:
+#   * deepep — mxfp4_deepseek.py:347 reads dispatch_output.topk_output;
+#     neither DeepEPNormalDispatchOutput nor DeepEPLLDispatchOutput
+#     exposes that field in this fork.
+#   * flashinfer — `_handle_a2a_moe` in server_args.py asserts
+#     "Flashinfer MoE A2A is only supported with flashinfer_cutlass
+#     moe runner backend", and flashinfer_cutlass is FP8-only — won't
+#     load DSV4-Pro's MXFP4 weights.
+# Adds prefill capacity (3 workers vs 1) for the high-conc tail —
+# single prefill saturates around conc 4096 at 1k prompts.
+#
+# Topology: 3 prefill (TP=8 / DP=8) + 1 decode (TP=8 / DP=8). 8 nodes.
+
+model:
+  path: "deepseek-v4-pro"
+  container: "lmsysorg/sglang:deepseek-v4-grace-blackwell"
+  precision: "fp4"
+
+# See ./disagg-gb200-1p1d-dep8-tep8.yaml for the dynamo pin rationale.
+dynamo:
+  hash: 21f135f5edf40e12e6ff5db2b462d862a6d6ab9b
+  install: true
+
+slurm:
+  time_limit: "8:00:00"
+
+health_check:
+  max_attempts: 1440
+  interval_seconds: 10
+
+resources:
+  gpu_type: "gb200"
+  gpus_per_node: 4
+  prefill_nodes: 6
+  decode_nodes: 2
+  prefill_workers: 3
+  decode_workers: 1
+  gpus_per_prefill: 8
+  gpus_per_decode: 8
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+
+backend:
+  type: sglang
+
+  prefill_environment:
+    PYTHONUNBUFFERED: "1"
+    SGLANG_JIT_DEEPGEMM_PRECOMPILE: "0"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: "100000"
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: "100000"
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "100000"
+    SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK: "1024"
+
+  decode_environment:
+    PYTHONUNBUFFERED: "1"
+    SGLANG_JIT_DEEPGEMM_PRECOMPILE: "0"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: "100000"
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: "100000"
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "100000"
+    SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK: "1024"
+
+  sglang_config:
+    prefill:
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      trust-remote-code: true
+      tensor-parallel-size: 8
+      moe-dense-tp-size: 1
+      enable-dp-attention: true
+      dp-size: 8
+      moe-runner-backend: "flashinfer_mxfp4"
+      chunked-prefill-size: 4096
+      disable-flashinfer-autotune: true
+      disable-radix-cache: true
+      mem-fraction-static: 0.82
+      context-length: 3072
+      max-running-requests: 16
+      stream-interval: 50
+      decode-log-interval: 1000
+      disaggregation-mode: "prefill"
+      disaggregation-bootstrap-port: 30001
+      disaggregation-transfer-backend: nixl
+
+    decode:
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      trust-remote-code: true
+      tensor-parallel-size: 8
+      moe-dense-tp-size: 1
+      enable-dp-attention: true
+      dp-size: 8
+      moe-runner-backend: "flashinfer_mxfp4"
+      chunked-prefill-size: 4096
+      disable-flashinfer-autotune: true
+      disable-radix-cache: true
+      mem-fraction-static: 0.82
+      context-length: 3072
+      max-running-requests: 1024
+      cuda-graph-max-bs: 1024
+      stream-interval: 50
+      decode-log-interval: 1000
+      disaggregation-mode: "decode"
+      disaggregation-bootstrap-port: 30001
+      disaggregation-transfer-backend: nixl
+
+benchmark:
+  type: "sa-bench"
+  isl: 1024
+  osl: 1024
+  concurrencies: "4096x8192"
+  req_rate: "inf"
+  use_chat_template: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-1p1d-dep8-tep8.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-1p1d-dep8-tep8.yaml
new file mode 100644
index 000000000..218ad01f6
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-1p1d-dep8-tep8.yaml
@@ -0,0 +1,113 @@
+name: "dsv4-sglang-disagg-gb200-1p1d-dep8-tep8"
+
+# 8k/1k variant of the 1k/1k 1p1d-dep8-tep8 recipe. Same topology and
+# tuning; only context-length grows from 3072 (1k+1k+pad) to 9280
+# (8k+1k+pad), and prefill max-running-requests halves to keep the per-
+# rank prefill working set inside the GPU memory budget.
+#
+# See ../1k1k/disagg-gb200-1p1d-dep8-tep8.yaml for the full upstream-
+# reference list (PR #69 GB200 agg, PR #75 GB300 disagg).
+
+model:
+  path: "deepseek-v4-pro"
+  container: "lmsysorg/sglang:deepseek-v4-grace-blackwell"
+  precision: "fp4"
+
+# See ../1k1k/disagg-gb200-1p1d-dep8-tep8.yaml for the dynamo pin rationale.
+dynamo:
+  hash: 21f135f5edf40e12e6ff5db2b462d862a6d6ab9b
+  install: true
+
+slurm:
+  time_limit: "8:00:00"
+
+health_check:
+  max_attempts: 1440
+  interval_seconds: 10
+
+resources:
+  gpu_type: "gb200"
+  gpus_per_node: 4
+  prefill_nodes: 2
+  decode_nodes: 2
+  prefill_workers: 1
+  decode_workers: 1
+  gpus_per_prefill: 8
+  gpus_per_decode: 8
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+
+backend:
+  type: sglang
+
+  prefill_environment:
+    PYTHONUNBUFFERED: "1"
+    SGLANG_JIT_DEEPGEMM_PRECOMPILE: "0"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: "100000"
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: "100000"
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "100000"
+    SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK: "1024"
+
+  decode_environment:
+    PYTHONUNBUFFERED: "1"
+    SGLANG_JIT_DEEPGEMM_PRECOMPILE: "0"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: "100000"
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: "100000"
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "100000"
+    SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK: "1024"
+
+  sglang_config:
+    prefill:
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      trust-remote-code: true
+      tensor-parallel-size: 8
+      moe-dense-tp-size: 1
+      enable-dp-attention: true
+      dp-size: 8
+      moe-runner-backend: "flashinfer_mxfp4"
+      chunked-prefill-size: 4096
+      disable-flashinfer-autotune: true
+      disable-radix-cache: true
+      mem-fraction-static: 0.82
+      context-length: 9280
+      max-running-requests: 8
+      stream-interval: 50
+      decode-log-interval: 1000
+      disaggregation-mode: "prefill"
+      disaggregation-bootstrap-port: 30001
+      disaggregation-transfer-backend: nixl
+
+    decode:
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      trust-remote-code: true
+      tensor-parallel-size: 8
+      moe-dense-tp-size: 1
+      enable-dp-attention: true
+      dp-size: 8
+      moe-runner-backend: "flashinfer_mxfp4"
+      chunked-prefill-size: 4096
+      disable-flashinfer-autotune: true
+      disable-radix-cache: true
+      mem-fraction-static: 0.82
+      context-length: 9280
+      max-running-requests: 64
+      cuda-graph-max-bs: 64
+      stream-interval: 50
+      decode-log-interval: 1000
+      disaggregation-mode: "decode"
+      disaggregation-bootstrap-port: 30001
+      disaggregation-transfer-backend: nixl
+
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  concurrencies: "1x4x8x16x32x64"
+  req_rate: "inf"
+  use_chat_template: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-3p1d-dep8-dep16.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-3p1d-dep8-dep16.yaml
new file mode 100644
index 000000000..a1fd14571
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-3p1d-dep8-dep16.yaml
@@ -0,0 +1,112 @@
+name: "dsv4-sglang-disagg-gb200-3p1d-dep8-dep16"
+
+# 8k/1k mid-throughput topology: 3 prefill (DP=8 EP=8) + 1 wide decode
+# (DP=16 EP=16). 10 nodes. Targets conc 512-1024 — 8k prompts saturate
+# a single prefill worker below conc=512.
+#
+# See ../1k1k/disagg-gb200-1p1d-dep8-tep8.yaml for the upstream-reference
+# list. Topology mirrors the dsv4-fp4-gb200-dynamo-vllm sibling.
+
+model:
+  path: "deepseek-v4-pro"
+  container: "lmsysorg/sglang:deepseek-v4-grace-blackwell"
+  precision: "fp4"
+
+# See ../1k1k/disagg-gb200-1p1d-dep8-tep8.yaml for the dynamo pin rationale.
+dynamo:
+  hash: 21f135f5edf40e12e6ff5db2b462d862a6d6ab9b
+  install: true
+
+slurm:
+  time_limit: "8:00:00"
+
+health_check:
+  max_attempts: 1440
+  interval_seconds: 10
+
+resources:
+  gpu_type: "gb200"
+  gpus_per_node: 4
+  prefill_nodes: 6
+  decode_nodes: 2
+  prefill_workers: 3
+  decode_workers: 1
+  gpus_per_prefill: 8
+  gpus_per_decode: 8
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+
+backend:
+  type: sglang
+
+  prefill_environment:
+    PYTHONUNBUFFERED: "1"
+    SGLANG_JIT_DEEPGEMM_PRECOMPILE: "0"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: "100000"
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: "100000"
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "100000"
+    SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK: "1024"
+
+  decode_environment:
+    PYTHONUNBUFFERED: "1"
+    SGLANG_JIT_DEEPGEMM_PRECOMPILE: "0"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: "100000"
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: "100000"
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "100000"
+    SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK: "1024"
+
+  sglang_config:
+    prefill:
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      trust-remote-code: true
+      tensor-parallel-size: 8
+      moe-dense-tp-size: 1
+      enable-dp-attention: true
+      dp-size: 8
+      moe-runner-backend: "flashinfer_mxfp4"
+      chunked-prefill-size: 4096
+      disable-flashinfer-autotune: true
+      disable-radix-cache: true
+      mem-fraction-static: 0.82
+      context-length: 9280
+      max-running-requests: 8
+      stream-interval: 50
+      decode-log-interval: 1000
+      disaggregation-mode: "prefill"
+      disaggregation-bootstrap-port: 30001
+      disaggregation-transfer-backend: nixl
+
+    decode:
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      trust-remote-code: true
+      tensor-parallel-size: 8
+      moe-dense-tp-size: 1
+      enable-dp-attention: true
+      dp-size: 8
+      moe-runner-backend: "flashinfer_mxfp4"
+      chunked-prefill-size: 4096
+      disable-flashinfer-autotune: true
+      disable-radix-cache: true
+      mem-fraction-static: 0.82
+      context-length: 9280
+      max-running-requests: 256
+      cuda-graph-max-bs: 256
+      stream-interval: 50
+      decode-log-interval: 1000
+      disaggregation-mode: "decode"
+      disaggregation-bootstrap-port: 30001
+      disaggregation-transfer-backend: nixl
+
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  concurrencies: "512x1024"
+  req_rate: "inf"
+  use_chat_template: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-7p1d-dep8-dep16.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-7p1d-dep8-dep16.yaml
new file mode 100644
index 000000000..4eb0f2716
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-7p1d-dep8-dep16.yaml
@@ -0,0 +1,113 @@
+name: "dsv4-sglang-disagg-gb200-7p1d-dep8-dep16"
+
+# 8k/1k max-throughput topology: 7 prefill (DP=8 EP=8) + 1 wide decode
+# (DP=16 EP=16). 18 nodes — full GB200 cluster. Targets conc 4096-8192.
+# Per-worker tunings identical to the 3p1d sibling; only prefill_workers
+# and prefill_nodes scale up.
+#
+# See ../1k1k/disagg-gb200-1p1d-dep8-tep8.yaml for the upstream-reference
+# list. Topology mirrors the dsv4-fp4-gb200-dynamo-vllm sibling.
+
+model:
+  path: "deepseek-v4-pro"
+  container: "lmsysorg/sglang:deepseek-v4-grace-blackwell"
+  precision: "fp4"
+
+# See ../1k1k/disagg-gb200-1p1d-dep8-tep8.yaml for the dynamo pin rationale.
+dynamo:
+  hash: 21f135f5edf40e12e6ff5db2b462d862a6d6ab9b
+  install: true
+
+slurm:
+  time_limit: "8:00:00"
+
+health_check:
+  max_attempts: 1440
+  interval_seconds: 10
+
+resources:
+  gpu_type: "gb200"
+  gpus_per_node: 4
+  prefill_nodes: 14
+  decode_nodes: 2
+  prefill_workers: 7
+  decode_workers: 1
+  gpus_per_prefill: 8
+  gpus_per_decode: 8
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+
+backend:
+  type: sglang
+
+  prefill_environment:
+    PYTHONUNBUFFERED: "1"
+    SGLANG_JIT_DEEPGEMM_PRECOMPILE: "0"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: "100000"
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: "100000"
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "100000"
+    SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK: "1024"
+
+  decode_environment:
+    PYTHONUNBUFFERED: "1"
+    SGLANG_JIT_DEEPGEMM_PRECOMPILE: "0"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: "100000"
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: "100000"
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "100000"
+    SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK: "1024"
+
+  sglang_config:
+    prefill:
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      trust-remote-code: true
+      tensor-parallel-size: 8
+      moe-dense-tp-size: 1
+      enable-dp-attention: true
+      dp-size: 8
+      moe-runner-backend: "flashinfer_mxfp4"
+      chunked-prefill-size: 4096
+      disable-flashinfer-autotune: true
+      disable-radix-cache: true
+      mem-fraction-static: 0.82
+      context-length: 9280
+      max-running-requests: 8
+      stream-interval: 50
+      decode-log-interval: 1000
+      disaggregation-mode: "prefill"
+      disaggregation-bootstrap-port: 30001
+      disaggregation-transfer-backend: nixl
+
+    decode:
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      trust-remote-code: true
+      tensor-parallel-size: 8
+      moe-dense-tp-size: 1
+      enable-dp-attention: true
+      dp-size: 8
+      moe-runner-backend: "flashinfer_mxfp4"
+      chunked-prefill-size: 4096
+      disable-flashinfer-autotune: true
+      disable-radix-cache: true
+      mem-fraction-static: 0.82
+      context-length: 9280
+      max-running-requests: 256
+      cuda-graph-max-bs: 256
+      stream-interval: 50
+      decode-log-interval: 1000
+      disaggregation-mode: "decode"
+      disaggregation-bootstrap-port: 30001
+      disaggregation-transfer-backend: nixl
+
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  concurrencies: "4096x8192"
+  req_rate: "inf"
+  use_chat_template: false
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index a29c278f2..6772f90bd 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -1918,3 +1918,12 @@
     - "Three CONC bands: A=TP8 (1-8), B=TP4 (16-128), C=DP4 dp-attn (64-512); B/C overlap at conc 64,128"
     - "Configs: 1k1k and 8k1k, no validation.py / launcher / yaml-field changes (knob-free)"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1180
+
+- config-keys:
+    - dsv4-fp4-gb200-dynamo-sglang
+  description:
+    - "Add DeepSeek-V4-Pro FP4 GB200 disaggregated SGLang benchmarks via Dynamo (1k/1k sweep; 8k/1k recipes shipped but commented out)"
+    - "Container: lmsysorg/sglang:deepseek-v4-grace-blackwell (linux/arm64); model from /mnt/numa1/models/deepseek-v4-pro/ (compute-node-local NVMe)"
+    - "Topologies mirror the dsv4-fp4-gb200-dynamo-vllm sibling: low-conc 1p1d-dep8-tep8 (4 nodes), mid 1p1d-dep8-dep16 (6 nodes), high 3p1d-dep8-dep16 (10 nodes). 4096 overlap between mid and high gives a topology-crossover A/B"
+    - "No upstream GB200 DSV4 sglang disagg recipe exists. Per-worker sglang_config (env vars + flashinfer_mxfp4 + chunked-prefill-size 4096 + disable-flashinfer-autotune + mem-fraction-static 0.82) is mirrored from NVIDIA/srt-slurm PR #69 (recipes/gb200-fp4/1k1k-dsv4/agg-2n-low-latency.yaml — GB200 DSV4 SGLang aggregated). Disagg flag set (nixl transfer backend, enable-dp-attention + moe-a2a-backend deepep) cross-checked against PR #75 (recipes/gb300-fp4/1k1k-dsv4/disagg-1p1d-tp4-mxfp4.yaml — GB300 DSV4 SGLang disagg) and the SGLang DeepSeek-V4 cookbook. Stored under benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/ and overlaid onto the upstream srt-slurm checkout at runtime"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1157
diff --git a/runners/launch_gb200-nv.sh b/runners/launch_gb200-nv.sh
index 224c3a928..08897874e 100755
--- a/runners/launch_gb200-nv.sh
+++ b/runners/launch_gb200-nv.sh
@@ -15,6 +15,12 @@ if [[ $FRAMEWORK == "dynamo-sglang" ]]; then
     elif [[ $MODEL_PREFIX == "dsr1" && $PRECISION == "fp4" ]]; then
         export MODEL_PATH="/mnt/lustre01/models/deepseek-r1-0528-fp4-v2/"
         export SRT_SLURM_MODEL_PREFIX="dsr1-fp4"
+    elif [[ $MODEL_PREFIX == "dsv4" && $PRECISION == "fp4" ]]; then
+        # Same compute-node-local NVMe path as the dynamo-vllm dsv4
+        # branch — see that branch for rationale. SRT_SLURM_MODEL_PREFIX
+        # matches the model.path alias in our DSV4 sglang recipes.
+        export MODEL_PATH="/mnt/numa1/models/deepseek-v4-pro/"
+        export SRT_SLURM_MODEL_PREFIX="deepseek-v4-pro"
     else
         export MODEL_PATH=$MODEL
     fi
@@ -150,6 +156,16 @@ if [[ $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "dsv4" ]]; then
     # `recipes/vllm/deepseek-v4/deepseek-v4/...` in that case).
     mkdir -p recipes/vllm/deepseek-v4
     cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4" recipes/vllm/deepseek-v4
+elif [[ $FRAMEWORK == "dynamo-sglang" && $MODEL_PREFIX == "dsv4" ]]; then
+    # Mirrors the dynamo-vllm dsv4 branch above: pin to the q2-2026
+    # NVIDIA srt-slurm (newer srtctl + dynamo-sglang container alias)
+    # and overlay our hand-rolled DSV4 sglang recipes. NVIDIA/srt-slurm
+    # has no upstream sglang DSV4 disagg recipes yet, hence the overlay.
+    git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
+    cd "$SRT_REPO_DIR"
+    git checkout sa-submission-q2-2026
+    mkdir -p recipes/sglang/deepseek-v4
+    cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4" recipes/sglang/deepseek-v4
 elif [[ $FRAMEWORK == "dynamo-vllm" ]]; then
     git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
     cd "$SRT_REPO_DIR"