From a48ffcb298a85054d051d6a36d2c16322ec2d2bc Mon Sep 17 00:00:00 2001
From: yhyang201 <yhyang201@gmail.com>
Date: Tue, 28 Apr 2026 23:50:59 +0800
Subject: [PATCH 1/9] sglang dpskv4 hopper

---
 .github/configs/nvidia-master.yaml            | 19 +++++
 .../single_node/dsv4_fp4_h200_sglang.sh       | 73 +++++++++++++++++++
 2 files changed, 92 insertions(+)
 create mode 100644 benchmarks/single_node/dsv4_fp4_h200_sglang.sh

diff --git a/.github/configs/nvidia-master.yaml b/.github/configs/nvidia-master.yaml
index 38d1101f3..e8a0faab6 100644
--- a/.github/configs/nvidia-master.yaml
+++ b/.github/configs/nvidia-master.yaml
@@ -2604,6 +2604,25 @@ dsv4-fp8-h200-vllm:
       search-space:
       - { tp: 8, ep: 8, dp-attn: true, conc-start: 4, conc-end: 64 }
 
+# DeepSeek-V4-Pro H200 single-node with SGLang (Marlin FP4, TP-only).
+dsv4-fp4-h200-sglang:
+  image: lmsysorg/sglang:deepseek-v4-hopper
+  model: deepseek-ai/DeepSeek-V4-Pro
+  model-prefix: dsv4
+  runner: h200
+  precision: fp4
+  framework: sglang
+  multinode: false
+  seq-len-configs:
+  - isl: 1024
+    osl: 1024
+    search-space:
+    - { tp: 8, ep: 1, conc-start: 1, conc-end: 32 }
+  - isl: 8192
+    osl: 1024
+    search-space:
+    - { tp: 8, ep: 1, conc-start: 1, conc-end: 32 }
+
   # DeepSeek-V4-Pro B300 single-node aggregate recipe from the submitted B300
   # pareto sweep. The single-node schema has no explicit data-parallel-size
   # field, so dp-attn=true is used as the existing vLLM script switch for DP4
diff --git a/benchmarks/single_node/dsv4_fp4_h200_sglang.sh b/benchmarks/single_node/dsv4_fp4_h200_sglang.sh
new file mode 100644
index 000000000..a7e822596
--- /dev/null
+++ b/benchmarks/single_node/dsv4_fp4_h200_sglang.sh
@@ -0,0 +1,73 @@
+#!/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 $SLURMD_NODENAME"
+fi
+
+hf download "$MODEL"
+
+nvidia-smi
+
+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"
+
+set -x
+PYTHONNOUSERSITE=1 sglang serve \
+    --model-path $MODEL \
+    --host 0.0.0.0 \
+    --port $PORT \
+    --trust-remote-code \
+    --tp $TP \
+    --moe-runner-backend marlin \
+    --chunked-prefill-size 4096 \
+    --disable-flashinfer-autotune \
+    --mem-fraction-static 0.88 \
+    --max-running-requests "$(( CONC * 3 / 2 > 8 ? CONC * 3 / 2 : 8 ))" \
+    $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

From a479062bf5baa474e5e36ed142122f32506b60da Mon Sep 17 00:00:00 2001
From: yhyang201 <yhyang201@gmail.com>
Date: Wed, 29 Apr 2026 00:32:50 +0800
Subject: [PATCH 2/9] h200 runner: support framework-tagged script names

---
 runners/launch_h200-dgxc-slurm.sh | 13 ++++++++++++-
 1 file changed, 12 insertions(+), 1 deletion(-)

diff --git a/runners/launch_h200-dgxc-slurm.sh b/runners/launch_h200-dgxc-slurm.sh
index 199f5e4ae..7fbbf1cb6 100755
--- a/runners/launch_h200-dgxc-slurm.sh
+++ b/runners/launch_h200-dgxc-slurm.sh
@@ -292,13 +292,24 @@ else
         fi
     "
 
+    # Prefer a framework-tagged script (e.g. dsv4_fp4_h200_sglang.sh) so models
+    # with multiple inference engines can coexist; fall back to the historical
+    # name without an engine suffix for scripts that haven't been retagged yet.
+    SPEC_SUFFIX=$([[ "$SPEC_DECODING" == "mtp" ]] && printf '_mtp' || printf '')
+    BENCH_BASE="benchmarks/single_node/${SCENARIO_SUBDIR}${EXP_NAME%%_*}_${PRECISION}_h200"
+    BENCH_SCRIPT="${BENCH_BASE}_${FRAMEWORK}${SPEC_SUFFIX}.sh"
+    if [[ ! -f "$BENCH_SCRIPT" ]]; then
+        LEGACY_FW_SUFFIX=$([[ "$FRAMEWORK" == "trt" ]] && printf '_trt' || printf '')
+        BENCH_SCRIPT="${BENCH_BASE}${LEGACY_FW_SUFFIX}${SPEC_SUFFIX}.sh"
+    fi
+
     srun --jobid=$JOB_ID \
         --container-image=$SQUASH_FILE \
         --container-mounts=$GITHUB_WORKSPACE:/workspace/,$HF_HUB_CACHE_MOUNT:$HF_HUB_CACHE \
         --no-container-mount-home \
         --container-workdir=/workspace/ \
         --no-container-entrypoint --export=ALL,PORT=8888 \
-        bash benchmarks/single_node/${SCENARIO_SUBDIR}${EXP_NAME%%_*}_${PRECISION}_h200$([[ "$FRAMEWORK" == "trt" ]] && printf '_trt')$([[ "$SPEC_DECODING" == "mtp" ]] && printf '_mtp').sh
+        bash $BENCH_SCRIPT
 
     scancel $JOB_ID
 

From 31da83572b7ab4c03701eb9ad43dbbf960a00365 Mon Sep 17 00:00:00 2001
From: yhyang201 <yhyang201@gmail.com>
Date: Wed, 29 Apr 2026 00:40:52 +0800
Subject: [PATCH 3/9] h200 runners: fix script path and /workspace mount
 conflict

---
 runners/launch_h200-cw.sh         | 19 +++++++++++++++----
 runners/launch_h200-dgxc-slurm.sh | 13 ++++++++-----
 runners/launch_h200-nb.sh         | 20 ++++++++++++++++----
 3 files changed, 39 insertions(+), 13 deletions(-)

diff --git a/runners/launch_h200-cw.sh b/runners/launch_h200-cw.sh
index 1486c4fa6..4be548c87 100644
--- a/runners/launch_h200-cw.sh
+++ b/runners/launch_h200-cw.sh
@@ -4,8 +4,13 @@ export HF_HUB_CACHE_MOUNT="/mnt/vast/gharunner/hf-hub-cache"
 export PORT=8888
 
 MODEL_CODE="${EXP_NAME%%_*}"
-FRAMEWORK_SUFFIX=$([[ "$FRAMEWORK" == "trt" ]] && printf '_trt' || printf '')
 SPEC_SUFFIX=$([[ "$SPEC_DECODING" == "mtp" ]] && printf '_mtp' || printf '')
+BENCH_BASE="benchmarks/single_node/${SCENARIO_SUBDIR}${MODEL_CODE}_${PRECISION}_h200"
+BENCH_SCRIPT="${BENCH_BASE}_${FRAMEWORK}${SPEC_SUFFIX}.sh"
+if [[ ! -f "$BENCH_SCRIPT" ]]; then
+    LEGACY_FW_SUFFIX=$([[ "$FRAMEWORK" == "trt" ]] && printf '_trt' || printf '')
+    BENCH_SCRIPT="${BENCH_BASE}${LEGACY_FW_SUFFIX}${SPEC_SUFFIX}.sh"
+fi
 
 PARTITION="h200"
 SQUASH_FILE="/mnt/vast/gharunner/squash/$(echo "$IMAGE" | sed 's/[\/:@#]/_/g').sqsh"
@@ -38,13 +43,19 @@ else
     CONTAINER_IMAGE=$(realpath $SQUASH_FILE)
 fi
 
+if [[ "$IMAGE" == *deepseek-v4-hopper* ]]; then
+    CONTAINER_MOUNT_DIR=/ix
+else
+    CONTAINER_MOUNT_DIR=/workspace
+fi
+
 srun --jobid=$JOB_ID \
 --container-image=$CONTAINER_IMAGE \
---container-mounts=$GITHUB_WORKSPACE:/workspace/,$HF_HUB_CACHE_MOUNT:$HF_HUB_CACHE \
+--container-mounts=$GITHUB_WORKSPACE:$CONTAINER_MOUNT_DIR/,$HF_HUB_CACHE_MOUNT:$HF_HUB_CACHE \
 --container-mount-home \
---container-workdir=/workspace/ \
+--container-workdir=$CONTAINER_MOUNT_DIR/ \
 --no-container-entrypoint --export=ALL \
-bash benchmarks/single_node/${SCENARIO_SUBDIR}${MODEL_CODE}_${PRECISION}_h200${FRAMEWORK_SUFFIX}${SPEC_SUFFIX}.sh
+bash $BENCH_SCRIPT
 
 rmdir $SAGEMAKER_SHM_PATH
 scancel $JOB_ID
diff --git a/runners/launch_h200-dgxc-slurm.sh b/runners/launch_h200-dgxc-slurm.sh
index 7fbbf1cb6..7dd04a6d7 100755
--- a/runners/launch_h200-dgxc-slurm.sh
+++ b/runners/launch_h200-dgxc-slurm.sh
@@ -292,9 +292,6 @@ else
         fi
     "
 
-    # Prefer a framework-tagged script (e.g. dsv4_fp4_h200_sglang.sh) so models
-    # with multiple inference engines can coexist; fall back to the historical
-    # name without an engine suffix for scripts that haven't been retagged yet.
     SPEC_SUFFIX=$([[ "$SPEC_DECODING" == "mtp" ]] && printf '_mtp' || printf '')
     BENCH_BASE="benchmarks/single_node/${SCENARIO_SUBDIR}${EXP_NAME%%_*}_${PRECISION}_h200"
     BENCH_SCRIPT="${BENCH_BASE}_${FRAMEWORK}${SPEC_SUFFIX}.sh"
@@ -303,11 +300,17 @@ else
         BENCH_SCRIPT="${BENCH_BASE}${LEGACY_FW_SUFFIX}${SPEC_SUFFIX}.sh"
     fi
 
+    if [[ "$IMAGE" == *deepseek-v4-hopper* ]]; then
+        CONTAINER_MOUNT_DIR=/ix
+    else
+        CONTAINER_MOUNT_DIR=/workspace
+    fi
+
     srun --jobid=$JOB_ID \
         --container-image=$SQUASH_FILE \
-        --container-mounts=$GITHUB_WORKSPACE:/workspace/,$HF_HUB_CACHE_MOUNT:$HF_HUB_CACHE \
+        --container-mounts=$GITHUB_WORKSPACE:$CONTAINER_MOUNT_DIR/,$HF_HUB_CACHE_MOUNT:$HF_HUB_CACHE \
         --no-container-mount-home \
-        --container-workdir=/workspace/ \
+        --container-workdir=$CONTAINER_MOUNT_DIR/ \
         --no-container-entrypoint --export=ALL,PORT=8888 \
         bash $BENCH_SCRIPT
 
diff --git a/runners/launch_h200-nb.sh b/runners/launch_h200-nb.sh
index 158c30792..de2505158 100644
--- a/runners/launch_h200-nb.sh
+++ b/runners/launch_h200-nb.sh
@@ -4,19 +4,31 @@ export HF_HUB_CACHE_MOUNT="/mnt/data/gharunners/hf-hub-cache/"
 export PORT=8888
 
 MODEL_CODE="${EXP_NAME%%_*}"
-FRAMEWORK_SUFFIX=$([[ "$FRAMEWORK" == "trt" ]] && printf '_trt' || printf '')
 SPEC_SUFFIX=$([[ "$SPEC_DECODING" == "mtp" ]] && printf '_mtp' || printf '')
 
+BENCH_BASE="benchmarks/single_node/${SCENARIO_SUBDIR}${MODEL_CODE}_${PRECISION}_h200"
+BENCH_SCRIPT="${BENCH_BASE}_${FRAMEWORK}${SPEC_SUFFIX}.sh"
+if [[ ! -f "$BENCH_SCRIPT" ]]; then
+    LEGACY_FW_SUFFIX=$([[ "$FRAMEWORK" == "trt" ]] && printf '_trt' || printf '')
+    BENCH_SCRIPT="${BENCH_BASE}${LEGACY_FW_SUFFIX}${SPEC_SUFFIX}.sh"
+fi
+
+if [[ "$IMAGE" == *deepseek-v4-hopper* ]]; then
+    CONTAINER_MOUNT_DIR=/ix
+else
+    CONTAINER_MOUNT_DIR=/workspace
+fi
+
 PARTITION="main"
 
 set -x
 srun --partition=$PARTITION --gres=gpu:$TP --exclusive --job-name="$RUNNER_NAME" \
 --container-image=$IMAGE \
 --container-name=$(echo "$IMAGE" | sed 's/[\/:@#]/_/g')-${USER} \
---container-mounts=$GITHUB_WORKSPACE:/workspace/,$HF_HUB_CACHE_MOUNT:$HF_HUB_CACHE \
+--container-mounts=$GITHUB_WORKSPACE:$CONTAINER_MOUNT_DIR/,$HF_HUB_CACHE_MOUNT:$HF_HUB_CACHE \
 --container-remap-root \
 --container-writable \
 --container-mount-home \
---container-workdir=/workspace/ \
+--container-workdir=$CONTAINER_MOUNT_DIR/ \
 --no-container-entrypoint --export=ALL \
-bash benchmarks/single_node/${SCENARIO_SUBDIR}${MODEL_CODE}_${PRECISION}_h200${FRAMEWORK_SUFFIX}${SPEC_SUFFIX}.sh
+bash $BENCH_SCRIPT

From 49661cef02f65325699946cafb619a1f077653f0 Mon Sep 17 00:00:00 2001
From: yhyang201 <yhyang201@gmail.com>
Date: Wed, 29 Apr 2026 00:52:17 +0800
Subject: [PATCH 4/9] pin deepseek-v4-hopper image digest

---
 .github/configs/nvidia-master.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/.github/configs/nvidia-master.yaml b/.github/configs/nvidia-master.yaml
index e8a0faab6..b7993a0b0 100644
--- a/.github/configs/nvidia-master.yaml
+++ b/.github/configs/nvidia-master.yaml
@@ -2606,7 +2606,7 @@ dsv4-fp8-h200-vllm:
 
 # DeepSeek-V4-Pro H200 single-node with SGLang (Marlin FP4, TP-only).
 dsv4-fp4-h200-sglang:
-  image: lmsysorg/sglang:deepseek-v4-hopper
+  image: lmsysorg/sglang:deepseek-v4-hopper@sha256:7f19c6dc092e47a10fac2e41f47eab78970280d06648b8e50d312a82f0ae722f
   model: deepseek-ai/DeepSeek-V4-Pro
   model-prefix: dsv4
   runner: h200

From a5f403e553e44754f93acecd040c9771d57b2294 Mon Sep 17 00:00:00 2001
From: yhyang201 <yhyang201@gmail.com>
Date: Wed, 29 Apr 2026 12:18:13 +0800
Subject: [PATCH 5/9] fp4 -> fp8

---
 .github/configs/nvidia-master.yaml                          | 6 +++---
 .../{dsv4_fp4_h200_sglang.sh => dsv4_fp8_h200_sglang.sh}    | 0
 2 files changed, 3 insertions(+), 3 deletions(-)
 rename benchmarks/single_node/{dsv4_fp4_h200_sglang.sh => dsv4_fp8_h200_sglang.sh} (100%)

diff --git a/.github/configs/nvidia-master.yaml b/.github/configs/nvidia-master.yaml
index b7993a0b0..1d793f0f0 100644
--- a/.github/configs/nvidia-master.yaml
+++ b/.github/configs/nvidia-master.yaml
@@ -2604,13 +2604,13 @@ dsv4-fp8-h200-vllm:
       search-space:
       - { tp: 8, ep: 8, dp-attn: true, conc-start: 4, conc-end: 64 }
 
-# DeepSeek-V4-Pro H200 single-node with SGLang (Marlin FP4, TP-only).
-dsv4-fp4-h200-sglang:
+# DeepSeek-V4-Pro H200 single-node with SGLang (Marlin FP8, TP-only).
+dsv4-fp8-h200-sglang:
   image: lmsysorg/sglang:deepseek-v4-hopper@sha256:7f19c6dc092e47a10fac2e41f47eab78970280d06648b8e50d312a82f0ae722f
   model: deepseek-ai/DeepSeek-V4-Pro
   model-prefix: dsv4
   runner: h200
-  precision: fp4
+  precision: fp8
   framework: sglang
   multinode: false
   seq-len-configs:
diff --git a/benchmarks/single_node/dsv4_fp4_h200_sglang.sh b/benchmarks/single_node/dsv4_fp8_h200_sglang.sh
similarity index 100%
rename from benchmarks/single_node/dsv4_fp4_h200_sglang.sh
rename to benchmarks/single_node/dsv4_fp8_h200_sglang.sh

From d58ca0682ef78d8dd80700ef279638df85303622 Mon Sep 17 00:00:00 2001
From: yhyang201 <yhyang201@gmail.com>
Date: Wed, 29 Apr 2026 12:19:26 +0800
Subject: [PATCH 6/9] conc: drop 2, add 64

---
 .github/configs/nvidia-master.yaml | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

diff --git a/.github/configs/nvidia-master.yaml b/.github/configs/nvidia-master.yaml
index 1d793f0f0..9f44c2d64 100644
--- a/.github/configs/nvidia-master.yaml
+++ b/.github/configs/nvidia-master.yaml
@@ -2617,11 +2617,13 @@ dsv4-fp8-h200-sglang:
   - isl: 1024
     osl: 1024
     search-space:
-    - { tp: 8, ep: 1, conc-start: 1, conc-end: 32 }
+    - { tp: 8, ep: 1, conc-start: 1, conc-end: 1 }
+    - { tp: 8, ep: 1, conc-start: 4, conc-end: 64 }
   - isl: 8192
     osl: 1024
     search-space:
-    - { tp: 8, ep: 1, conc-start: 1, conc-end: 32 }
+    - { tp: 8, ep: 1, conc-start: 1, conc-end: 1 }
+    - { tp: 8, ep: 1, conc-start: 4, conc-end: 64 }
 
   # DeepSeek-V4-Pro B300 single-node aggregate recipe from the submitted B300
   # pareto sweep. The single-node schema has no explicit data-parallel-size

From 189a71d8e780076a647b5a13f006b1f306b19fe6 Mon Sep 17 00:00:00 2001
From: functionstackx <47992694+functionstackx@users.noreply.github.com>
Date: Sat, 2 May 2026 16:10:18 -0400
Subject: [PATCH 7/9] dsv4-fp8-h200-sglang: disable radix cache and migrate to
 scenarios.fixed-seq-len schema

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---
 .github/configs/nvidia-master.yaml            | 23 ++++++++++---------
 .../single_node/dsv4_fp8_h200_sglang.sh       |  1 +
 2 files changed, 13 insertions(+), 11 deletions(-)

diff --git a/.github/configs/nvidia-master.yaml b/.github/configs/nvidia-master.yaml
index 9f44c2d64..8c9fc0c7a 100644
--- a/.github/configs/nvidia-master.yaml
+++ b/.github/configs/nvidia-master.yaml
@@ -2613,17 +2613,18 @@ dsv4-fp8-h200-sglang:
   precision: fp8
   framework: sglang
   multinode: false
-  seq-len-configs:
-  - isl: 1024
-    osl: 1024
-    search-space:
-    - { tp: 8, ep: 1, conc-start: 1, conc-end: 1 }
-    - { tp: 8, ep: 1, conc-start: 4, conc-end: 64 }
-  - isl: 8192
-    osl: 1024
-    search-space:
-    - { tp: 8, ep: 1, conc-start: 1, conc-end: 1 }
-    - { tp: 8, ep: 1, conc-start: 4, conc-end: 64 }
+  scenarios:
+    fixed-seq-len:
+    - isl: 1024
+      osl: 1024
+      search-space:
+      - { tp: 8, ep: 1, conc-start: 1, conc-end: 1 }
+      - { tp: 8, ep: 1, conc-start: 4, conc-end: 64 }
+    - isl: 8192
+      osl: 1024
+      search-space:
+      - { tp: 8, ep: 1, conc-start: 1, conc-end: 1 }
+      - { tp: 8, ep: 1, conc-start: 4, conc-end: 64 }
 
   # DeepSeek-V4-Pro B300 single-node aggregate recipe from the submitted B300
   # pareto sweep. The single-node schema has no explicit data-parallel-size
diff --git a/benchmarks/single_node/dsv4_fp8_h200_sglang.sh b/benchmarks/single_node/dsv4_fp8_h200_sglang.sh
index a7e822596..f2fc8541e 100644
--- a/benchmarks/single_node/dsv4_fp8_h200_sglang.sh
+++ b/benchmarks/single_node/dsv4_fp8_h200_sglang.sh
@@ -42,6 +42,7 @@ PYTHONNOUSERSITE=1 sglang serve \
     --moe-runner-backend marlin \
     --chunked-prefill-size 4096 \
     --disable-flashinfer-autotune \
+    --disable-radix-cache \
     --mem-fraction-static 0.88 \
     --max-running-requests "$(( CONC * 3 / 2 > 8 ? CONC * 3 / 2 : 8 ))" \
     $EVAL_CONTEXT_ARGS >> $SERVER_LOG 2>&1 &

From 0fa05d309529ee12f68732b08f1be9a3e2c4ba85 Mon Sep 17 00:00:00 2001
From: functionstackx <47992694+functionstackx@users.noreply.github.com>
Date: Sat, 2 May 2026 16:24:17 -0400
Subject: [PATCH 8/9] perf-changelog: add dsv4-fp8-h200-sglang entry

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---
 perf-changelog.yaml | 10 ++++++++++
 1 file changed, 10 insertions(+)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 089ce36f2..6a4f08be5 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -2103,3 +2103,13 @@
     - "Sweep DP attention on/off and EP=8 via DP_ATTENTION and EP_SIZE matrix env vars"
     - "Ship a DeepSeek-V4 thinking-mode chat template so eval /v1/chat/completions works (the canonical checkpoint ships no chat_template)"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1244
+
+- config-keys:
+    - dsv4-fp8-h200-sglang
+  description:
+    - "Add DeepSeek-V4-Pro FP8 H200 single-node SGLang recipe (Marlin MoE backend, TP=8, EP=1)"
+    - "Image: lmsysorg/sglang:deepseek-v4-hopper pinned by digest"
+    - "Server flags: --moe-runner-backend marlin, --chunked-prefill-size 4096, --disable-flashinfer-autotune, --disable-radix-cache, --mem-fraction-static 0.88"
+    - "Search space: TP=8 EP=1, conc 1 and 4-64 for both 1k1k and 8k1k"
+    - "Pinned to the h200-dgxc runner pool (new runners.yaml group); launch_h200-dgxc-slurm.sh extended to support framework-tagged script names and mount /ix instead of /workspace for the deepseek-v4-hopper image"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1264

From 66a2446442057fd394207dfe59e72cfca3a08e94 Mon Sep 17 00:00:00 2001
From: functionstackx <47992694+functionstackx@users.noreply.github.com>
Date: Sat, 2 May 2026 17:18:46 -0400
Subject: [PATCH 9/9] dsv4-fp8-h200-sglang: pin to h200-dgxc runner pool

Revert the cw/nb launcher generalizations from the previous PR-1212 import
since the deepseek-v4-hopper image only needs the /ix mount layout from
launch_h200-dgxc-slurm.sh. Restrict the recipe to that runner pool by
adding a new h200-dgxc group in runners.yaml (parallel to the existing
gb300-cw / h200-multinode precedents) and setting runner: h200-dgxc.

All 14 h200-dgxc-slurm_* runners already carry the h200-dgxc label in
their GH Actions config, so runs-on: h200-dgxc dispatches cleanly. The
runner.name prefix is still h200-dgxc-slurm, which matches the existing
launch_h200-dgxc-slurm.sh.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---
 .github/configs/nvidia-master.yaml |  4 +++-
 .github/configs/runners.yaml       | 15 +++++++++++++++
 runners/launch_h200-cw.sh          | 19 ++++---------------
 runners/launch_h200-nb.sh          | 20 ++++----------------
 4 files changed, 26 insertions(+), 32 deletions(-)

diff --git a/.github/configs/nvidia-master.yaml b/.github/configs/nvidia-master.yaml
index 8c9fc0c7a..4ce7b823e 100644
--- a/.github/configs/nvidia-master.yaml
+++ b/.github/configs/nvidia-master.yaml
@@ -2605,11 +2605,13 @@ dsv4-fp8-h200-vllm:
       - { tp: 8, ep: 8, dp-attn: true, conc-start: 4, conc-end: 64 }
 
 # DeepSeek-V4-Pro H200 single-node with SGLang (Marlin FP8, TP-only).
+# Pinned to the h200-dgxc-slurm runner pool because the deepseek-v4-hopper
+# image needs the /ix mount layout that only launch_h200-dgxc-slurm.sh sets up.
 dsv4-fp8-h200-sglang:
   image: lmsysorg/sglang:deepseek-v4-hopper@sha256:7f19c6dc092e47a10fac2e41f47eab78970280d06648b8e50d312a82f0ae722f
   model: deepseek-ai/DeepSeek-V4-Pro
   model-prefix: dsv4
-  runner: h200
+  runner: h200-dgxc
   precision: fp8
   framework: sglang
   multinode: false
diff --git a/.github/configs/runners.yaml b/.github/configs/runners.yaml
index f574c629c..48a7173d4 100644
--- a/.github/configs/runners.yaml
+++ b/.github/configs/runners.yaml
@@ -49,6 +49,21 @@ h200-multinode:
 - 'h200-dgxc-slurm_11'
 - 'h200-dgxc-slurm_12'
 - 'h200-dgxc-slurm_13'
+h200-dgxc:
+- 'h200-dgxc-slurm_0'
+- 'h200-dgxc-slurm_1'
+- 'h200-dgxc-slurm_2'
+- 'h200-dgxc-slurm_3'
+- 'h200-dgxc-slurm_4'
+- 'h200-dgxc-slurm_5'
+- 'h200-dgxc-slurm_6'
+- 'h200-dgxc-slurm_7'
+- 'h200-dgxc-slurm_8'
+- 'h200-dgxc-slurm_9'
+- 'h200-dgxc-slurm_10'
+- 'h200-dgxc-slurm_11'
+- 'h200-dgxc-slurm_12'
+- 'h200-dgxc-slurm_13'
 b200:
 - 'b200-cw_00'
 - 'b200-cw_01'
diff --git a/runners/launch_h200-cw.sh b/runners/launch_h200-cw.sh
index 4be548c87..1486c4fa6 100644
--- a/runners/launch_h200-cw.sh
+++ b/runners/launch_h200-cw.sh
@@ -4,13 +4,8 @@ export HF_HUB_CACHE_MOUNT="/mnt/vast/gharunner/hf-hub-cache"
 export PORT=8888
 
 MODEL_CODE="${EXP_NAME%%_*}"
+FRAMEWORK_SUFFIX=$([[ "$FRAMEWORK" == "trt" ]] && printf '_trt' || printf '')
 SPEC_SUFFIX=$([[ "$SPEC_DECODING" == "mtp" ]] && printf '_mtp' || printf '')
-BENCH_BASE="benchmarks/single_node/${SCENARIO_SUBDIR}${MODEL_CODE}_${PRECISION}_h200"
-BENCH_SCRIPT="${BENCH_BASE}_${FRAMEWORK}${SPEC_SUFFIX}.sh"
-if [[ ! -f "$BENCH_SCRIPT" ]]; then
-    LEGACY_FW_SUFFIX=$([[ "$FRAMEWORK" == "trt" ]] && printf '_trt' || printf '')
-    BENCH_SCRIPT="${BENCH_BASE}${LEGACY_FW_SUFFIX}${SPEC_SUFFIX}.sh"
-fi
 
 PARTITION="h200"
 SQUASH_FILE="/mnt/vast/gharunner/squash/$(echo "$IMAGE" | sed 's/[\/:@#]/_/g').sqsh"
@@ -43,19 +38,13 @@ else
     CONTAINER_IMAGE=$(realpath $SQUASH_FILE)
 fi
 
-if [[ "$IMAGE" == *deepseek-v4-hopper* ]]; then
-    CONTAINER_MOUNT_DIR=/ix
-else
-    CONTAINER_MOUNT_DIR=/workspace
-fi
-
 srun --jobid=$JOB_ID \
 --container-image=$CONTAINER_IMAGE \
---container-mounts=$GITHUB_WORKSPACE:$CONTAINER_MOUNT_DIR/,$HF_HUB_CACHE_MOUNT:$HF_HUB_CACHE \
+--container-mounts=$GITHUB_WORKSPACE:/workspace/,$HF_HUB_CACHE_MOUNT:$HF_HUB_CACHE \
 --container-mount-home \
---container-workdir=$CONTAINER_MOUNT_DIR/ \
+--container-workdir=/workspace/ \
 --no-container-entrypoint --export=ALL \
-bash $BENCH_SCRIPT
+bash benchmarks/single_node/${SCENARIO_SUBDIR}${MODEL_CODE}_${PRECISION}_h200${FRAMEWORK_SUFFIX}${SPEC_SUFFIX}.sh
 
 rmdir $SAGEMAKER_SHM_PATH
 scancel $JOB_ID
diff --git a/runners/launch_h200-nb.sh b/runners/launch_h200-nb.sh
index de2505158..158c30792 100644
--- a/runners/launch_h200-nb.sh
+++ b/runners/launch_h200-nb.sh
@@ -4,31 +4,19 @@ export HF_HUB_CACHE_MOUNT="/mnt/data/gharunners/hf-hub-cache/"
 export PORT=8888
 
 MODEL_CODE="${EXP_NAME%%_*}"
+FRAMEWORK_SUFFIX=$([[ "$FRAMEWORK" == "trt" ]] && printf '_trt' || printf '')
 SPEC_SUFFIX=$([[ "$SPEC_DECODING" == "mtp" ]] && printf '_mtp' || printf '')
 
-BENCH_BASE="benchmarks/single_node/${SCENARIO_SUBDIR}${MODEL_CODE}_${PRECISION}_h200"
-BENCH_SCRIPT="${BENCH_BASE}_${FRAMEWORK}${SPEC_SUFFIX}.sh"
-if [[ ! -f "$BENCH_SCRIPT" ]]; then
-    LEGACY_FW_SUFFIX=$([[ "$FRAMEWORK" == "trt" ]] && printf '_trt' || printf '')
-    BENCH_SCRIPT="${BENCH_BASE}${LEGACY_FW_SUFFIX}${SPEC_SUFFIX}.sh"
-fi
-
-if [[ "$IMAGE" == *deepseek-v4-hopper* ]]; then
-    CONTAINER_MOUNT_DIR=/ix
-else
-    CONTAINER_MOUNT_DIR=/workspace
-fi
-
 PARTITION="main"
 
 set -x
 srun --partition=$PARTITION --gres=gpu:$TP --exclusive --job-name="$RUNNER_NAME" \
 --container-image=$IMAGE \
 --container-name=$(echo "$IMAGE" | sed 's/[\/:@#]/_/g')-${USER} \
---container-mounts=$GITHUB_WORKSPACE:$CONTAINER_MOUNT_DIR/,$HF_HUB_CACHE_MOUNT:$HF_HUB_CACHE \
+--container-mounts=$GITHUB_WORKSPACE:/workspace/,$HF_HUB_CACHE_MOUNT:$HF_HUB_CACHE \
 --container-remap-root \
 --container-writable \
 --container-mount-home \
---container-workdir=$CONTAINER_MOUNT_DIR/ \
+--container-workdir=/workspace/ \
 --no-container-entrypoint --export=ALL \
-bash $BENCH_SCRIPT
+bash benchmarks/single_node/${SCENARIO_SUBDIR}${MODEL_CODE}_${PRECISION}_h200${FRAMEWORK_SUFFIX}${SPEC_SUFFIX}.sh