From e1429284278314b70e8a2a1c51e80dea9278b770 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Fri, 24 Apr 2026 17:55:49 -0500
Subject: [PATCH] chore: tag dsv4 b300 benchmark scripts with inference engine

Each model historically used one inference engine, so the b300 launcher
just resolved benchmarks/single_node/${model}_${precision}_b300.sh
regardless of FRAMEWORK. With dsv4 we now want both an sglang and a
vllm script to coexist, so:

- launch_b300-nv.sh prefers an engine-tagged script (e.g.
  dsv4_fp4_b300_sglang.sh) and falls back to the legacy unsuffixed
  name (or the existing _trt suffix) when the tagged variant is
  absent. Existing dsr1/glm5/qwen3.5/kimik2.5/minimaxm2.5 b300
  scripts keep their current names.
- Rename the existing sglang dsv4 b300 script to
  dsv4_fp4_b300_sglang.sh.
- Restore dsv4_fp4_b300_vllm.sh from the abandoned
  origin/claude/add-dsv4-fp4-b300-vllm branch (script only; not
  wired into a config yet).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---
 ...v4_fp4_b300.sh => dsv4_fp4_b300_sglang.sh} |   0
 benchmarks/single_node/dsv4_fp4_b300_vllm.sh  | 104 ++++++++++++++++++
 runners/launch_b300-nv.sh                     |  13 ++-
 3 files changed, 115 insertions(+), 2 deletions(-)
 rename benchmarks/single_node/{dsv4_fp4_b300.sh => dsv4_fp4_b300_sglang.sh} (100%)
 create mode 100755 benchmarks/single_node/dsv4_fp4_b300_vllm.sh

diff --git a/benchmarks/single_node/dsv4_fp4_b300.sh b/benchmarks/single_node/dsv4_fp4_b300_sglang.sh
similarity index 100%
rename from benchmarks/single_node/dsv4_fp4_b300.sh
rename to benchmarks/single_node/dsv4_fp4_b300_sglang.sh
diff --git a/benchmarks/single_node/dsv4_fp4_b300_vllm.sh b/benchmarks/single_node/dsv4_fp4_b300_vllm.sh
new file mode 100755
index 000000000..8ec35beac
--- /dev/null
+++ b/benchmarks/single_node/dsv4_fp4_b300_vllm.sh
@@ -0,0 +1,104 @@
+#!/usr/bin/env bash
+
+# Per https://vllm.ai/blog/deepseek-v4 the DeepSeek-V4-Pro recipe lists
+# 8xB200 and 8xB300 with identical flags, so this script mirrors
+# dsv4_fp4_b200.sh.
+
+source "$(dirname "$0")/../benchmark_lib.sh"
+
+check_env_vars \
+    MODEL \
+    TP \
+    CONC \
+    ISL \
+    OSL \
+    MAX_MODEL_LEN \
+    RANDOM_RANGE_RATIO \
+    RESULT_FILENAME
+
+if [[ -n "$SLURM_JOB_ID" ]]; then
+  echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
+fi
+
+nvidia-smi
+
+SERVER_LOG=/workspace/server.log
+PORT=${PORT:-8888}
+
+# DeepSeek-V4-Pro weights are large and engine startup on B300 can exceed
+# the default 600s. Give it an hour to load.
+export VLLM_ENGINE_READY_TIMEOUT_S=3600
+
+if [ "${EVAL_ONLY}" = "true" ]; then
+    setup_eval_context
+    MAX_MODEL_LEN="$EVAL_MAX_MODEL_LEN"
+fi
+
+# Monkey-patch: bypass persistent_topk unconditionally. It raises "k out of
+# range" during CUDA graph capture when the dummy batch has rows with
+# seq_lens[i] < k (=2048 for DSV4). An attn_metadata.max_seq_len-based gate is
+# not strict enough because dummy batches can have max >= k while individual
+# rows have seq_lens[i] = 1. Fall back to top_k_per_row_decode everywhere so
+# 1k/1k capture completes; 8k/1k already worked without the patch but we trade
+# a small decode-time perf cost there to keep the script single-branch.
+INDEXER_PY=/usr/local/lib/python3.12/dist-packages/vllm/model_executor/layers/sparse_attn_indexer.py
+echo "[monkey-patch] patching $INDEXER_PY"
+sed -i 's/if current_platform.is_cuda() and topk_tokens in (512, 1024, 2048)[^:]*:/if False:  # monkey-patched: bypass persistent_topk (k out of range)/' "$INDEXER_PY"
+if ! grep -Fq 'if False:  # monkey-patched: bypass persistent_topk' "$INDEXER_PY"; then
+    echo "[monkey-patch] FAILED: expected marker not found in $INDEXER_PY" >&2
+    echo "[monkey-patch] current line around persistent_topk dispatch:" >&2
+    grep -n 'topk_tokens in\|persistent_topk' "$INDEXER_PY" >&2 || true
+    exit 1
+fi
+echo "[monkey-patch] applied: $(grep -n 'if False:  # monkey-patched' $INDEXER_PY)"
+
+# Start GPU monitoring (power, temperature, clocks every second)
+start_gpu_monitor
+
+# Per the recipe, run with EP + DP=8 (no --tensor-parallel-size flag). TP
+# from the search space is used only for GPU allocation by the runner and
+# as the DP size.
+set -x
+vllm serve $MODEL --host 0.0.0.0 --port $PORT \
+--trust-remote-code \
+--kv-cache-dtype fp8 \
+--block-size 256 \
+--no-enable-prefix-caching \
+--enable-expert-parallel \
+--data-parallel-size $TP \
+--max-model-len $MAX_MODEL_LEN \
+--compilation-config '{"cudagraph_mode":"FULL_AND_PIECEWISE","custom_ops":["all"]}' \
+--tokenizer-mode deepseek_v4 \
+--tool-call-parser deepseek_v4 \
+--enable-auto-tool-choice \
+--reasoning-parser deepseek_v4 > $SERVER_LOG 2>&1 &
+
+SERVER_PID=$!
+
+# Wait for server to be ready
+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 "$RESULT_FILENAME" \
+    --result-dir /workspace/ \
+    --trust-remote-code
+
+# After throughput, run evaluation only if RUN_EVAL is true
+if [ "${RUN_EVAL}" = "true" ]; then
+    run_eval --framework lm-eval --port "$PORT"
+    append_lm_eval_summary
+fi
+
+# Stop GPU monitoring
+stop_gpu_monitor
+set +x
diff --git a/runners/launch_b300-nv.sh b/runners/launch_b300-nv.sh
index 3daac0167..3c855e805 100644
--- a/runners/launch_b300-nv.sh
+++ b/runners/launch_b300-nv.sh
@@ -259,8 +259,17 @@ else
         export MODEL="$HF_HUB_CACHE_MOUNT/dsv4-pro"
     fi
     SQUASH_FILE="/data/home/sa-shared/gharunners/squash/$(echo "$IMAGE" | sed 's/[\/:@#]/_/g').sqsh"
-    FRAMEWORK_SUFFIX=$([[ "$FRAMEWORK" == "trt" ]] && printf '_trt' || printf '')
     SPEC_SUFFIX=$([[ "$SPEC_DECODING" == "mtp" ]] && printf '_mtp' || printf '')
+    # Prefer a framework-tagged script (e.g. dsv4_fp4_b300_sglang.sh) so models
+    # with multiple inference engines can coexist; fall back to the historical
+    # name without an engine suffix (`_trt` for trt, bare for everyone else)
+    # for scripts that haven't been retagged yet.
+    BENCH_BASE="benchmarks/single_node/${EXP_NAME%%_*}_${PRECISION}_b300"
+    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}${SPEC_SUFFIX}.sh"
+    fi
     LOCK_FILE="${SQUASH_FILE}.lock"
 
     # TODO(Cam): the deepseek-v4 sglang images (lmsysorg/sglang:deepseek-v4-blackwell
@@ -300,6 +309,6 @@ else
         --no-container-mount-home \
         --container-workdir=$CONTAINER_MOUNT_DIR \
         --no-container-entrypoint --export=ALL,PORT=8888 \
-        bash benchmarks/single_node/${EXP_NAME%%_*}_${PRECISION}_b300${FRAMEWORK_SUFFIX}${SPEC_SUFFIX}.sh
+        bash "$BENCH_SCRIPT"
 
 fi