[NV]dsr1-fp4-b200-sglang: add DPA PDL lane

.github/configs/nvidia-master.yaml

@@ -1670,7 +1670,7 @@ dsr1-fp8-b300-dynamo-trt:
           ep: 8
           dp-attn: true
 dsr1-fp4-b200-sglang:
-  image: lmsysorg/sglang:v0.5.12-cu130
+  image: lmsysorg/sglang:v0.5.12.post1
   model: nvidia/DeepSeek-R1-0528-FP4-V2
   model-prefix: dsr1
   runner: b200
@@ -1682,13 +1682,14 @@ dsr1-fp4-b200-sglang:
     - isl: 1024
       osl: 1024
       search-space:
-      - { tp: 4, ep: 4, conc-start: 1, conc-end: 128 }
-      - { tp: 8, ep: 8, conc-start: 1, conc-end: 128 }
+      - { tp: 4, ep: 4, conc-start: 1, conc-end: 256 }
+      - { tp: 8, ep: 8, conc-list: [1] }
     - isl: 8192
       osl: 1024
       search-space:
       - { tp: 4, ep: 4, conc-start: 1, conc-end: 128 }
-      - { tp: 8, ep: 8, conc-start: 1, conc-end: 16 }
+      - { tp: 4, ep: 4, dp-attn: true, conc-start: 64, conc-end: 256 }
+      - { tp: 8, ep: 8, conc-list: [1] }
     # agentic-coding: temporarily disabled — blocked by e2e-tests.yml artifact
     # name mismatch (downloads `agentic_*` but benchmark-tmpl.yml uploads as
     # `bmk_agentic_*`). Re-enable once that workflow is aligned.

benchmarks/single_node/fixed_seq_len/dsr1_fp4_b200.sh

@@ -2,6 +2,8 @@
 
 source "$(dirname "$0")/../../benchmark_lib.sh"
 
+DP_ATTENTION="${DP_ATTENTION:-false}"
+
 check_env_vars \
     MODEL \
     TP \
@@ -10,7 +12,13 @@ check_env_vars \
     OSL \
     RANDOM_RANGE_RATIO \
     RESULT_FILENAME \
-    EP_SIZE
+    EP_SIZE \
+    DP_ATTENTION
+
+if [[ "$DP_ATTENTION" != "true" && "$DP_ATTENTION" != "false" ]]; then
+    echo "DP_ATTENTION must be true or false; got '$DP_ATTENTION'" >&2
+    exit 1
+fi
 
 if [[ -n "$SLURM_JOB_ID" ]]; then
   echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
@@ -22,13 +30,38 @@ nvidia-smi
 
 SERVER_LOG=/workspace/server.log
 
-# Default: recv every ~10 requests; if CONC ≥ 16, relax to ~30 requests between scheduler recv polls.
+# Default: recv every ~10 requests; if CONC >= 16, relax to ~30 requests between scheduler recv polls.
 if [[ $CONC -ge 16 ]]; then
   SCHEDULER_RECV_INTERVAL=30
 else
   SCHEDULER_RECV_INTERVAL=10
 fi
-echo "SCHEDULER_RECV_INTERVAL: $SCHEDULER_RECV_INTERVAL, CONC: $CONC, ISL: $ISL, OSL: $OSL"
+
+CHUNKED_PREFILL_SIZE=16384
+SGLANG_PARALLEL_ARGS=(
+    --tensor-parallel-size="$TP"
+    --data-parallel-size=1
+)
+SGLANG_DPA_ARGS=()
+
+if [[ "$DP_ATTENTION" == "true" ]]; then
+    SCHEDULER_RECV_INTERVAL=1
+    CHUNKED_PREFILL_SIZE=32768
+    SGLANG_PARALLEL_ARGS=(
+        --tensor-parallel-size="$TP"
+        --data-parallel-size="$TP"
+        --enable-dp-attention
+        --enable-dp-attention-local-control-broadcast
+        --enable-dp-lm-head
+    )
+    SGLANG_DPA_ARGS=(
+        --schedule-conservativeness 3.33
+        --enable-prefill-delayer
+    )
+fi
+
+echo "TP: $TP, EP_SIZE: $EP_SIZE, DP_ATTENTION: $DP_ATTENTION, CONC: $CONC, ISL: $ISL, OSL: $OSL"
+echo "SCHEDULER_RECV_INTERVAL: $SCHEDULER_RECV_INTERVAL, CHUNKED_PREFILL_SIZE: $CHUNKED_PREFILL_SIZE"
 
 EVAL_CONTEXT_ARGS=""
 if [ "${EVAL_ONLY}" = "true" ]; then
@@ -39,12 +72,12 @@ fi
 start_gpu_monitor
 
 set -x
-PYTHONNOUSERSITE=1 python3 -m sglang.launch_server --model-path $MODEL --host 0.0.0.0 --port $PORT --trust-remote-code \
---tensor-parallel-size=$TP --data-parallel-size=1 \
+SGLANG_RADIX_FORCE_MISS=1 PYTHONNOUSERSITE=1 python3 -m sglang.launch_server --model-path $MODEL --host 0.0.0.0 --port $PORT --trust-remote-code \
+"${SGLANG_PARALLEL_ARGS[@]}" \
 --cuda-graph-max-bs 256 --max-running-requests 256 --mem-fraction-static 0.85 --kv-cache-dtype fp8_e4m3 \
---chunked-prefill-size 16384 \
+--chunked-prefill-size "$CHUNKED_PREFILL_SIZE" \
 --ep-size $EP_SIZE --quantization modelopt_fp4 --enable-flashinfer-allreduce-fusion --scheduler-recv-interval $SCHEDULER_RECV_INTERVAL \
---enable-symm-mem --disable-radix-cache --attention-backend trtllm_mla --moe-runner-backend flashinfer_trtllm --stream-interval 10 $EVAL_CONTEXT_ARGS > $SERVER_LOG 2>&1 &
+--enable-symm-mem --disable-piecewise-cuda-graph --attention-backend trtllm_mla --moe-runner-backend flashinfer_trtllm --stream-interval 10 "${SGLANG_DPA_ARGS[@]}" $EVAL_CONTEXT_ARGS > $SERVER_LOG 2>&1 &
 
 SERVER_PID=$!
 

perf-changelog.yaml

@@ -4115,3 +4115,12 @@
     - "Use the dedicated ARM64 MiniMax-M3 performance image; benchmark settings unchanged"
     - "Allocate FlashInfer MNNVL workspace for one-shot TP8 all-reduce during CUDA graph capture"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1897
+
+- config-keys:
+    - dsr1-fp4-b200-sglang
+  description:
+    - "Update B200 FP4 SGLang image from lmsysorg/sglang:v0.5.12-cu130 to lmsysorg/sglang:v0.5.12.post1."
+    - "Update search space: 1k/1k TP4/EP4 conc 1-256 and TP8/EP8 conc 1; 8k/1k TP4/EP4 conc 1-128, add TP4/EP4 DP-attention conc 64-256, and keep TP8/EP8 conc 1."
+    - "For DP-attention runs, use TP-sized data parallelism with DP attention local-control broadcast, DP LM head, prefill delayer, scheduler recv interval 1, chunked prefill size 32768, and schedule conservativeness 3.33."
+    - "Set SGLANG_RADIX_FORCE_MISS=1, remove --disable-radix-cache, and explicitly pass --disable-piecewise-cuda-graph."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1792