From a6f84240dd1685ad3b8dd4b8f3b445bf8f895b6f Mon Sep 17 00:00:00 2001 From: functionstackx <47992694+functionstackx@users.noreply.github.com> Date: Fri, 24 Apr 2026 01:30:17 -0400 Subject: [PATCH 1/4] Add H200 config: dsv4-fp8-vllm (DeepSeek-V4-Pro) Port the DeepSeek-V4-Pro vLLM recipe to H200 per https://vllm.ai/blog/deepseek-v4. Uses the cu129 image and omits the FP4 indexer cache flag (H200 has no FP4 path). Max-model-len is pinned at 800k per the recipe. Prefix caching is disabled (matches the B200/B300 configs and the user's note) and VLLM_ENGINE_READY_TIMEOUT_S is bumped to 1200s to tolerate slow weight loading. Launch: EP + DP=$TP (no --tensor-parallel-size), FP8 KV cache, block size 256, max-model-len 800000, prefix caching disabled, deepseek_v4 tokenizer/tool-call/reasoning parsers. Configs: 1k1k conc 4-64, 8k1k conc 4-64. Co-Authored-By: Claude Opus 4.7 (1M context) --- .github/configs/nvidia-master.yaml | 21 ++++++ benchmarks/single_node/dsv4_fp8_h200.sh | 89 +++++++++++++++++++++++++ perf-changelog.yaml | 17 +++++ 3 files changed, 127 insertions(+) create mode 100644 benchmarks/single_node/dsv4_fp8_h200.sh diff --git a/.github/configs/nvidia-master.yaml b/.github/configs/nvidia-master.yaml index ec9cbc11e..96273444f 100644 --- a/.github/configs/nvidia-master.yaml +++ b/.github/configs/nvidia-master.yaml @@ -2363,6 +2363,27 @@ dsr1-fp8-h200-sglang: search-space: - { tp: 8, conc-start: 4, conc-end: 64 } +# DeepSeek-V4-Pro H200 recipe from https://vllm.ai/blog/deepseek-v4 +# Uses the cu129 image. H200 has no FP4 path, so the FP4 indexer cache +# flag is omitted. Max-model-len is pinned at 800k per the recipe. +dsv4-fp8-h200-vllm: + image: vllm/vllm-openai:deepseekv4-cu129 + model: deepseek-ai/DeepSeek-V4-Pro + model-prefix: dsv4 + runner: h200 + precision: fp8 + framework: vllm + multinode: false + seq-len-configs: + - isl: 1024 + osl: 1024 + search-space: + - { tp: 8, ep: 8, dp-attn: true, conc-start: 4, conc-end: 64 } + - isl: 8192 + osl: 1024 + search-space: + - { tp: 8, ep: 8, dp-attn: true, conc-start: 4, conc-end: 64 } + qwen3.5-fp8-h200-sglang: image: lmsysorg/sglang:v0.5.9-cu129-amd64 model: Qwen/Qwen3.5-397B-A17B-FP8 diff --git a/benchmarks/single_node/dsv4_fp8_h200.sh b/benchmarks/single_node/dsv4_fp8_h200.sh new file mode 100644 index 000000000..329b204f2 --- /dev/null +++ b/benchmarks/single_node/dsv4_fp8_h200.sh @@ -0,0 +1,89 @@ +#!/usr/bin/env bash + +# Per https://vllm.ai/blog/deepseek-v4 the DeepSeek-V4-Pro H200 recipe uses +# the cu129 image and omits the FP4 indexer cache flag (H200 has no FP4 +# path). Max-model-len is pinned at 800k per the recipe. + +source "$(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 $SLURMD_NODENAME" +fi + +nvidia-smi + +hf download "$MODEL" + +SERVER_LOG=/workspace/server.log +PORT=${PORT:-8888} + +# DeepSeek-V4-Pro weights are large; engine startup can exceed the default +# 600s. Give it 20 minutes to load. +export VLLM_ENGINE_READY_TIMEOUT_S=1200 + +if [ "${EVAL_ONLY}" = "true" ]; then + setup_eval_context + MAX_MODEL_LEN_ARG="--max-model-len $EVAL_MAX_MODEL_LEN" +else + MAX_MODEL_LEN_ARG="--max-model-len 800000" +fi + +# 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_ARG \ +--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 "$PORT" + append_lm_eval_summary +fi + +# Stop GPU monitoring +stop_gpu_monitor +set +x diff --git a/perf-changelog.yaml b/perf-changelog.yaml index ddc6409c2..5e8b000db 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -1733,3 +1733,20 @@ - "TP=2 and TP=4, concurrency 4-256 for 1k1k and 8k1k sequence lengths" - "Add --max-num-seqs and --gpu-memory-utilization 0.9 to server launch" pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1043 + +- config-keys: + - dsv4-fp8-h200-vllm + description: + - "Add DeepSeek-V4-Pro vLLM H200 benchmark per https://vllm.ai/blog/deepseek-v4" + - "Image: vllm/vllm-openai:deepseekv4-cu129" + - "Model: deepseek-ai/DeepSeek-V4-Pro" + - "EP + DP=8, FP8 KV cache, block size 256, max-model-len 800000, prefix caching disabled" + - "H200 has no FP4 path, so --attention_config.use_fp4_indexer_cache is omitted" + - "VLLM_ENGINE_READY_TIMEOUT_S=1200 to accommodate large weight loading" + - "Flags: --trust-remote-code, --kv-cache-dtype fp8, --block-size 256, --no-enable-prefix-caching," + - " --enable-expert-parallel, --data-parallel-size=$TP, --max-model-len 800000," + - " --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" + - "Configs: 1k1k conc 4-64, 8k1k conc 4-64" + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXXX From 7a0812c84c84636a49f3da946c4e082f0750e138 Mon Sep 17 00:00:00 2001 From: functionstackx <47992694+functionstackx@users.noreply.github.com> Date: Fri, 24 Apr 2026 02:09:13 -0400 Subject: [PATCH 2/4] dsv4-fp8-h200-vllm: bump engine-ready timeout to 1h VLLM_ENGINE_READY_TIMEOUT_S 1200 -> 3600. Matches the B300 config; DeepSeek-V4-Pro weight loading was tripping the 20-min gate during sweeps. Also update the changelog entry text. Co-Authored-By: Claude Opus 4.7 (1M context) --- benchmarks/single_node/dsv4_fp8_h200.sh | 4 ++-- perf-changelog.yaml | 2 +- 2 files changed, 3 insertions(+), 3 deletions(-) diff --git a/benchmarks/single_node/dsv4_fp8_h200.sh b/benchmarks/single_node/dsv4_fp8_h200.sh index 329b204f2..d5eb0687f 100644 --- a/benchmarks/single_node/dsv4_fp8_h200.sh +++ b/benchmarks/single_node/dsv4_fp8_h200.sh @@ -27,8 +27,8 @@ SERVER_LOG=/workspace/server.log PORT=${PORT:-8888} # DeepSeek-V4-Pro weights are large; engine startup can exceed the default -# 600s. Give it 20 minutes to load. -export VLLM_ENGINE_READY_TIMEOUT_S=1200 +# 600s. Give it an hour to load. +export VLLM_ENGINE_READY_TIMEOUT_S=3600 if [ "${EVAL_ONLY}" = "true" ]; then setup_eval_context diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 5e8b000db..beda30e18 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -1742,7 +1742,7 @@ - "Model: deepseek-ai/DeepSeek-V4-Pro" - "EP + DP=8, FP8 KV cache, block size 256, max-model-len 800000, prefix caching disabled" - "H200 has no FP4 path, so --attention_config.use_fp4_indexer_cache is omitted" - - "VLLM_ENGINE_READY_TIMEOUT_S=1200 to accommodate large weight loading" + - "VLLM_ENGINE_READY_TIMEOUT_S=3600 to accommodate large weight loading" - "Flags: --trust-remote-code, --kv-cache-dtype fp8, --block-size 256, --no-enable-prefix-caching," - " --enable-expert-parallel, --data-parallel-size=$TP, --max-model-len 800000," - " --compilation-config cudagraph_mode=FULL_AND_PIECEWISE custom_ops=all," From 620afcdbd7e4c5ebc0ce88b93c48165e870e93ed Mon Sep 17 00:00:00 2001 From: Oseltamivir Date: Fri, 24 Apr 2026 01:10:02 -0700 Subject: [PATCH 3/4] update to pr 380 --- benchmarks/single_node/dsv4_fp8_h200.sh | 6 +++++- perf-changelog.yaml | 4 +++- 2 files changed, 8 insertions(+), 2 deletions(-) diff --git a/benchmarks/single_node/dsv4_fp8_h200.sh b/benchmarks/single_node/dsv4_fp8_h200.sh index d5eb0687f..167a50a57 100644 --- a/benchmarks/single_node/dsv4_fp8_h200.sh +++ b/benchmarks/single_node/dsv4_fp8_h200.sh @@ -52,7 +52,11 @@ vllm serve $MODEL --host 0.0.0.0 --port $PORT \ --enable-expert-parallel \ --data-parallel-size $TP \ $MAX_MODEL_LEN_ARG \ ---compilation-config '{"cudagraph_mode":"FULL_AND_PIECEWISE","custom_ops":["all"]}' \ +--gpu-memory-utilization 0.95 \ +--max-num-seqs 512 \ +--max-num-batched-tokens 512 \ +--no-enable-flashinfer-autotune \ +--compilation-config '{"mode":0,"cudagraph_mode":"FULL_DECODE_ONLY"}' \ --tokenizer-mode deepseek_v4 \ --tool-call-parser deepseek_v4 \ --enable-auto-tool-choice \ diff --git a/perf-changelog.yaml b/perf-changelog.yaml index beda30e18..798470dd2 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -1745,7 +1745,9 @@ - "VLLM_ENGINE_READY_TIMEOUT_S=3600 to accommodate large weight loading" - "Flags: --trust-remote-code, --kv-cache-dtype fp8, --block-size 256, --no-enable-prefix-caching," - " --enable-expert-parallel, --data-parallel-size=$TP, --max-model-len 800000," - - " --compilation-config cudagraph_mode=FULL_AND_PIECEWISE custom_ops=all," + - " --gpu-memory-utilization 0.95, --max-num-seqs 512, --max-num-batched-tokens 512," + - " --no-enable-flashinfer-autotune," + - " --compilation-config mode=0 cudagraph_mode=FULL_DECODE_ONLY," - " --tokenizer-mode deepseek_v4, --tool-call-parser deepseek_v4," - " --enable-auto-tool-choice, --reasoning-parser deepseek_v4" - "Configs: 1k1k conc 4-64, 8k1k conc 4-64" From 55a6b107448b306863a1ccb19d3ad83ef4a6081c Mon Sep 17 00:00:00 2001 From: Cameron Quilici Date: Fri, 24 Apr 2026 10:09:32 -0500 Subject: [PATCH 4/4] Update perf-changelog.yaml --- perf-changelog.yaml | 9 +-------- 1 file changed, 1 insertion(+), 8 deletions(-) diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 798470dd2..2b2e138c8 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -1743,12 +1743,5 @@ - "EP + DP=8, FP8 KV cache, block size 256, max-model-len 800000, prefix caching disabled" - "H200 has no FP4 path, so --attention_config.use_fp4_indexer_cache is omitted" - "VLLM_ENGINE_READY_TIMEOUT_S=3600 to accommodate large weight loading" - - "Flags: --trust-remote-code, --kv-cache-dtype fp8, --block-size 256, --no-enable-prefix-caching," - - " --enable-expert-parallel, --data-parallel-size=$TP, --max-model-len 800000," - - " --gpu-memory-utilization 0.95, --max-num-seqs 512, --max-num-batched-tokens 512," - - " --no-enable-flashinfer-autotune," - - " --compilation-config mode=0 cudagraph_mode=FULL_DECODE_ONLY," - - " --tokenizer-mode deepseek_v4, --tool-call-parser deepseek_v4," - - " --enable-auto-tool-choice, --reasoning-parser deepseek_v4" - "Configs: 1k1k conc 4-64, 8k1k conc 4-64" - pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXXX + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1130