Day 0 DeepSeek V4 Pro FP4 GB300 disaggregated SGLang benchmarks

.github/configs/nvidia-master.yaml

@@ -7748,3 +7748,101 @@ dsv4-fp4-gb200-dynamo-vllm:
         tp: 8
         ep: 8
         dp-attn: true
+
+dsv4-fp4-gb300-dynamo-sglang:
+  image: lmsysorg/sglang:deepseek-v4-grace-blackwell
+  model: deepseek-ai/DeepSeek-V4-Pro
+  model-prefix: dsv4
+  runner: gb300-cw
+  precision: fp4
+  framework: dynamo-sglang
+  multinode: true
+  disagg: true
+  seq-len-configs:
+  - isl: 8192
+    osl: 1024
+    search-space:
+    # WideEP TP=16 decode: 1p1d-dep4-dep16. 5 nodes (4P + 16D = 20 GPUs).
+    - conc-list: [512]
+      prefill:
+        num-worker: 1
+        tp: 4
+        ep: 4
+        dp-attn: true
+        additional-settings:
+        - "CONFIG_FILE=recipes/sglang/deepseek-v4/8k1k/conc512-20.yaml"
+      decode:
+        num-worker: 1
+        tp: 16
+        ep: 16
+        dp-attn: true
+    # DP-attn wideep: 1p1d-dep4-dep8. 3 nodes.
+    - conc-list: [512]
+      prefill:
+        num-worker: 1
+        tp: 4
+        ep: 4
+        dp-attn: true
+        additional-settings:
+        - "CONFIG_FILE=recipes/sglang/deepseek-v4/8k1k/conc512.yaml"
+      decode:
+        num-worker: 1
+        tp: 8
+        ep: 8
+        dp-attn: true
+    # DP-attn wideep: 2p1d-dep4-dep8. 4 nodes.
+    - conc-list: [1024]
+      prefill:
+        num-worker: 2
+        tp: 4
+        ep: 4
+        dp-attn: true
+        additional-settings:
+        - "CONFIG_FILE=recipes/sglang/deepseek-v4/8k1k/conc1024.yaml"
+      decode:
+        num-worker: 1
+        tp: 8
+        ep: 8
+        dp-attn: true
+    # Low concurrency
+    - conc-list: [1]
+      prefill:
+        num-worker: 1
+        tp: 4
+        ep: 1
+        dp-attn: false
+        additional-settings:
+        - "CONFIG_FILE=recipes/sglang/deepseek-v4/8k1k/conc1.yaml"
+      decode:
+        num-worker: 1
+        tp: 4
+        ep: 1
+        dp-attn: false
+    # Mid concurrency
+    - conc-list: [2048]
+      prefill:
+        num-worker: 4
+        tp: 4
+        ep: 4
+        dp-attn: true
+        additional-settings:
+        - "CONFIG_FILE=recipes/sglang/deepseek-v4/8k1k/conc2048.yaml"
+      decode:
+        num-worker: 1
+        tp: 8
+        ep: 8
+        dp-attn: true
+    # Max concurrency
+    - conc-list: [16384]
+      prefill:
+        num-worker: 14
+        tp: 4
+        ep: 4
+        dp-attn: true
+        additional-settings:
+        - "CONFIG_FILE=recipes/sglang/deepseek-v4/8k1k/conc16384.yaml"
+      decode:
+        num-worker: 1
+        tp: 16
+        ep: 16
+        dp-attn: true

.github/configs/runners.yaml

@@ -139,3 +139,8 @@ gb300:
 - 'gb300-nv_0'
 - 'gb300-nv_1'
 - 'gb300-nv_2'
+gb300-cw:
+- 'gb300-cw_0'
+- 'gb300-cw_1'
+- 'gb300-cw_2'
+- 'gb300-cw_3'

benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/conc1.yaml

@@ -0,0 +1,167 @@
+name: "conc1"
+
+# 8k/1k high-throughput topology for the wideep DSV4-Pro setup.
+#
+# Schema/values come from PR #1213 (513cbef) — that PR introduced the
+# `dsv4-pro-gb300-fp4` upstream-style recipe with two `zip_override`
+# variants (wideep [0] / narrow_ep [1]) and `backend.benchmark`. Our
+# pinned srtctl (NVIDIA/srt-slurm @ sa-submission-q2-2026) doesn't
+# support either: `zip_override_*_hightpt` rejects with `Unknown field`
+# and `benchmark` only validates at top level. So this file inlines the
+# wideep [0] override and lifts `benchmark` back out — same operational
+# values, schema the pinned srtctl will accept.
+#
+# Other adjustments back to the InferenceX cluster shape: container &
+# model.path restored to the aliases mapped in launch_gb300.sh's
+# srtslurm.yaml (`lmsysorg/sglang:deepseek-v4-grace-blackwell` and
+# `deepseek-v4-pro`); `dynamo.install: true` added so the container
+# (which has no dynamo baked in) installs from the pinned hash.
+#
+# Cluster-specific items NOT inlined (require InferenceX-side equivalents):
+#   - slurm.partition (yangminl's gb300-cw uses `hpc-mid`)
+#   - frontend.nginx_container (yangminl's `nginx-1.27.4.sqsh` path)
+#   - extra_mount: yangminl/sglang-patched/sglang. Earlier diff analysis
+#     showed only `expert_location_dispatch.py` topk_ids int32 cast is an
+#     active runtime diff vs container sglang; other patched files are
+#     env-gated dead code under the same SGLANG_OPT_* flags this yaml
+#     already sets.
+#
+# DG-related env intentionally diverged (DG cache path is host-specific):
+#   - SGLANG_DG_CACHE_DIR=/configs/deepgemm_cache (yangminl host)
+#   - SGLANG_JIT_DEEPGEMM_PRECOMPILE=0 (yangminl uses prebuilt cache)
+#   This yaml uses SGLANG_JIT_DEEPGEMM_FAST_WARMUP=1 instead.
+
+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. Hash bumped from PR #1213 to track the dynamo-sglang dsv4
+# dev branch.
+dynamo:
+  hash: "9d3c913d300eb368cda28b3f98a23a5762621e0d"
+  install: true
+
+slurm:
+  time_limit: "03:00:00"
+
+# Match yangminl's working all-dynamo.yaml on the same gb300-cw cluster:
+#   cpus-per-task=144  — without this slurm hands out 1 CPU/task, which
+#     turns the dynamo `hash:` cold source build (~500 rust crates,
+#     ravif/exr/zip/pyo3 stack) into a 30+ min serial compile. With 144
+#     cargo finishes in ~5 min.
+#   mem=0              — slurm's "give the whole node's memory"; needed
+#     for sglang loading 671B FP4 weights + dynamo build at the same
+#     time without OOM.
+sbatch_directives:
+  cpus-per-task: "144"
+  mem: "0"
+
+# Topology: 7 prefill (TP=4 / DP=4 / EP=4 / 1 node each) + 1 decode
+# (TP=8 / DP=8 / EP=8 / 2 nodes). 9 nodes total.
+resources:
+  gpu_type: "gb300"
+  gpus_per_node: 4
+  prefill_nodes: 1
+  prefill_workers: 1
+  gpus_per_prefill: 4
+  decode_nodes: 1
+  decode_workers: 1
+  gpus_per_decode: 4
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: true
+  num_additional_frontends: 8
+
+backend:
+  type: sglang
+
+  prefill_environment:
+    PYTHONUNBUFFERED: "1"
+    SGLANG_JIT_DEEPGEMM_FAST_WARMUP: "1"
+    SGLANG_ENABLE_THINKING: "1"
+    SGLANG_REASONING_EFFORT: "max"
+    SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: "1"
+    SGLANG_OPT_USE_JIT_NORM: "1"
+    SGLANG_OPT_USE_JIT_INDEXER_METADATA: "1"
+    SGLANG_OPT_USE_TOPK_V2: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: "True"
+    MC_FORCE_MNNVL: "1"
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: "100000"
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "100000"
+    SGLANG_OPT_SWA_RELEASE_LEAF_LOCK_AFTER_WINDOW: "1"
+
+  decode_environment:
+    PYTHONUNBUFFERED: "1"
+    SGLANG_JIT_DEEPGEMM_FAST_WARMUP: "1"
+    SGLANG_ENABLE_THINKING: "1"
+    SGLANG_REASONING_EFFORT: "max"
+    SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: "1"
+    SGLANG_OPT_USE_JIT_NORM: "1"
+    SGLANG_OPT_USE_JIT_INDEXER_METADATA: "1"
+    SGLANG_OPT_USE_TOPK_V2: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: "True"
+    MC_FORCE_MNNVL: "1"
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: "100000"
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "100000"
+    SGLANG_OPT_SWA_RELEASE_LEAF_LOCK_AFTER_WINDOW: "1"
+    # SGLANG_OPT_USE_CUSTOM_ALL_REDUCE_V2 intentionally NOT set: CAR_V2
+    # is single-node only and corrupts results in 2-node decode setups.
+
+  sglang_config:
+    prefill:
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      model-path: "/model/"
+      trust-remote-code: true
+      disable-radix-cache: true
+
+      disaggregation-mode: "prefill"
+      disaggregation-transfer-backend: mooncake
+
+      tensor-parallel-size: 4
+      data-parallel-size:   1
+      expert-parallel-size: 1
+
+      moe-runner-backend: "flashinfer_mxfp4"
+      disable-flashinfer-autotune: true
+
+      mem-fraction-static: 0.90
+      max-running-requests: 512
+      cuda-graph-max-bs: 512
+      chunked-prefill-size: 32768
+
+    decode:
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      model-path: "/model/"
+      trust-remote-code: true
+      disable-radix-cache: true
+
+      disaggregation-mode: "decode"
+      disaggregation-transfer-backend: mooncake
+
+      tensor-parallel-size: 4
+      data-parallel-size:   1
+      expert-parallel-size: 1
+
+      moe-runner-backend: "flashinfer_mxfp4"
+      disable-flashinfer-autotune: true
+
+      mem-fraction-static: 0.9
+      max-running-requests: 1024
+      cuda-graph-max-bs:    512
+      swa-full-tokens-ratio: 0.1
+      context-length: 16384
+
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  concurrencies: "1"
+  req_rate: "inf"
+  use_chat_template: false