From b5a6eaaa3576c59650b5d9468edc2f08d56c58fb Mon Sep 17 00:00:00 2001 From: functionstackx <47992694+functionstackx@users.noreply.github.com> Date: Thu, 23 Apr 2026 23:54:29 -0400 Subject: [PATCH 1/3] Add B300 config: dsv4-fp4-vllm (DeepSeek-V4-Pro) Mirrors the B200 dsv4-fp4-vllm config for B300. The recipe at https://vllm.ai/blog/deepseek-v4 lists 8xB200 and 8xB300 with identical flags for DeepSeek-V4-Pro, so the launch script is a straight port of dsv4_fp4_b200.sh with a note at the top. Launch: EP + DP=$TP (no --tensor-parallel-size), FP8 KV cache, block size 256, FP4 indexer cache, prefix caching disabled, deepseek_v4 tokenizer/tool-call/reasoning parsers. Configs: 1k1k conc 4-1024, 8k1k conc 4-512. Co-Authored-By: Claude Opus 4.7 (1M context) --- .github/configs/nvidia-master.yaml | 22 +++++++ benchmarks/single_node/dsv4_fp4_b300.sh | 85 +++++++++++++++++++++++++ perf-changelog.yaml | 18 +++++- 3 files changed, 124 insertions(+), 1 deletion(-) create mode 100644 benchmarks/single_node/dsv4_fp4_b300.sh diff --git a/.github/configs/nvidia-master.yaml b/.github/configs/nvidia-master.yaml index a64803497..0984b5061 100644 --- a/.github/configs/nvidia-master.yaml +++ b/.github/configs/nvidia-master.yaml @@ -1728,6 +1728,28 @@ dsr1-fp4-b300-sglang: - { tp: 4, ep: 4, conc-start: 4, conc-end: 128 } - { tp: 8, ep: 8, conc-start: 4, conc-end: 16 } +# DeepSeek-V4-Pro recipe from https://vllm.ai/blog/deepseek-v4 +# The Pro recipe lists 8xB200 and 8xB300 with identical flags. Runs with +# DP=8 + expert parallelism (no --tensor-parallel-size flag), FP8 KV cache, +# block size 256, and an FP4 indexer cache. +dsv4-fp4-b300-vllm: + image: vllm/vllm-openai:deepseekv4-cu130 + model: deepseek-ai/DeepSeek-V4-Pro + model-prefix: dsv4 + runner: b300 + precision: fp4 + 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: 1024 } + - isl: 8192 + osl: 1024 + search-space: + - { tp: 8, ep: 8, dp-attn: true, conc-start: 4, conc-end: 512 } + dsr1-fp4-b200-trt: image: nvcr.io#nvidia/tensorrt-llm/release:1.2.0rc6.post2 model: nvidia/DeepSeek-R1-0528-FP4-V2 diff --git a/benchmarks/single_node/dsv4_fp4_b300.sh b/benchmarks/single_node/dsv4_fp4_b300.sh new file mode 100644 index 000000000..1cbfbbcd9 --- /dev/null +++ b/benchmarks/single_node/dsv4_fp4_b300.sh @@ -0,0 +1,85 @@ +#!/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 $SLURMD_NODENAME" +fi + +nvidia-smi + +hf download "$MODEL" + +SERVER_LOG=/workspace/server.log +PORT=${PORT:-8888} + +if [ "${EVAL_ONLY}" = "true" ]; then + setup_eval_context + MAX_MODEL_LEN="$EVAL_MAX_MODEL_LEN" +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 $MAX_MODEL_LEN \ +--compilation-config '{"cudagraph_mode":"FULL_AND_PIECEWISE","custom_ops":["all"]}' \ +--attention_config.use_fp4_indexer_cache=True \ +--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 "$PORT" + append_lm_eval_summary +fi + +# Stop GPU monitoring +stop_gpu_monitor +set +x diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 528a12d02..07119324a 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -1755,7 +1755,7 @@ - "VLLM_ENGINE_READY_TIMEOUT_S=3600 to accommodate large weight loading" - "Configs: 1k1k conc 4-64, 8k1k conc 4-64" pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1130 - + - config-keys: - dsv4-fp4-b300-sglang description: @@ -1775,3 +1775,19 @@ - "Model: sgl-project/DeepSeek-V4-Pro-FP8" - "https://github.com/sgl-project/sglang/pull/23608#issuecomment-4311952977" pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1134 + +- config-keys: + - dsv4-fp4-b300-vllm + description: + - "Add DeepSeek-V4-Pro vLLM B300 benchmark per https://vllm.ai/blog/deepseek-v4" + - "Image: vllm/vllm-openai:deepseekv4-cu130" + - "Model: deepseek-ai/DeepSeek-V4-Pro" + - "Recipe lists 8xB200 and 8xB300 with identical flags; this mirrors the B200 config" + - "EP + DP=8 (no --tensor-parallel-size), FP8 KV cache, block size 256, FP4 indexer cache" + - "Flags: --trust-remote-code, --kv-cache-dtype fp8, --block-size 256, --no-enable-prefix-caching," + - " --enable-expert-parallel, --data-parallel-size=$TP," + - " --compilation-config cudagraph_mode=FULL_AND_PIECEWISE custom_ops=all," + - " --attention_config.use_fp4_indexer_cache=True, --tokenizer-mode deepseek_v4," + - " --tool-call-parser deepseek_v4, --enable-auto-tool-choice, --reasoning-parser deepseek_v4" + - "Configs: 1k1k conc 4-1024, 8k1k conc 4-512" + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXXX From eb1375650192714c24e428b1f880605542a60df3 Mon Sep 17 00:00:00 2001 From: functionstackx <47992694+functionstackx@users.noreply.github.com> Date: Fri, 24 Apr 2026 00:15:41 -0400 Subject: [PATCH 2/3] dsv4-fp4-b300-vllm: extend engine-ready timeout, cap conc at 64 - Set VLLM_ENGINE_READY_TIMEOUT_S=1200 (20 min) in the launch script. DeepSeek-V4-Pro weights are large and engine startup on B300 was exceeding the default 600s, surfacing as: TimeoutError: Timed out waiting for engine core processes to start - Cap both 1k1k and 8k1k search-space concurrency at 64 (conc-start 4, conc-end 64) to keep the initial sweep bounded while the config stabilizes. Higher-conc points can be re-enabled once the runner is consistently coming up green. Co-Authored-By: Claude Opus 4.7 (1M context) --- .github/configs/nvidia-master.yaml | 4 ++-- benchmarks/single_node/dsv4_fp4_b300.sh | 4 ++++ 2 files changed, 6 insertions(+), 2 deletions(-) diff --git a/.github/configs/nvidia-master.yaml b/.github/configs/nvidia-master.yaml index 0984b5061..c649fc213 100644 --- a/.github/configs/nvidia-master.yaml +++ b/.github/configs/nvidia-master.yaml @@ -1744,11 +1744,11 @@ dsv4-fp4-b300-vllm: - isl: 1024 osl: 1024 search-space: - - { tp: 8, ep: 8, dp-attn: true, conc-start: 4, conc-end: 1024 } + - { 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: 512 } + - { tp: 8, ep: 8, dp-attn: true, conc-start: 4, conc-end: 64 } dsr1-fp4-b200-trt: image: nvcr.io#nvidia/tensorrt-llm/release:1.2.0rc6.post2 diff --git a/benchmarks/single_node/dsv4_fp4_b300.sh b/benchmarks/single_node/dsv4_fp4_b300.sh index 1cbfbbcd9..6c4847fe3 100644 --- a/benchmarks/single_node/dsv4_fp4_b300.sh +++ b/benchmarks/single_node/dsv4_fp4_b300.sh @@ -27,6 +27,10 @@ hf download "$MODEL" SERVER_LOG=/workspace/server.log PORT=${PORT:-8888} +# DeepSeek-V4-Pro weights are large and engine startup on B300 can exceed +# the default 600s. Give it 20 minutes to load. +export VLLM_ENGINE_READY_TIMEOUT_S=1200 + if [ "${EVAL_ONLY}" = "true" ]; then setup_eval_context MAX_MODEL_LEN="$EVAL_MAX_MODEL_LEN" From 9434fca3ede989c6273dab3a5a6e3b6b452b5966 Mon Sep 17 00:00:00 2001 From: functionstackx <47992694+functionstackx@users.noreply.github.com> Date: Fri, 24 Apr 2026 01:40:20 -0400 Subject: [PATCH 3/3] dsv4 B300: retag precision fp4fp8, bump startup timeout to 1h MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit - Rename config key dsv4-fp4-b300-vllm -> dsv4-fp4fp8-b300-vllm and precision fp4 -> fp4fp8, reflecting that the model runs fp4 weights with an fp8 KV cache. - Rename benchmarks/single_node/dsv4_fp4_b300.sh -> dsv4_fp4fp8_b300.sh so the runner (${EXP_NAME%%_*}_${PRECISION}_b300.sh) resolves to the new script. User wrote "fp4p8" in the script name; I'm using "fp4fp8" to match the PRECISION field — otherwise the launcher would look for a script that doesn't exist. - Bump VLLM_ENGINE_READY_TIMEOUT_S from 1200 (20m) to 3600 (1h) so slow weight loading doesn't trip the engine-ready gate. - Update perf-changelog entry accordingly. Co-Authored-By: Claude Opus 4.7 (1M context) --- .github/configs/nvidia-master.yaml | 6 ++++-- .../{dsv4_fp4_b300.sh => dsv4_fp4fp8_b300.sh} | 4 ++-- perf-changelog.yaml | 12 +++++++----- 3 files changed, 13 insertions(+), 9 deletions(-) rename benchmarks/single_node/{dsv4_fp4_b300.sh => dsv4_fp4fp8_b300.sh} (96%) diff --git a/.github/configs/nvidia-master.yaml b/.github/configs/nvidia-master.yaml index c649fc213..86fade6c1 100644 --- a/.github/configs/nvidia-master.yaml +++ b/.github/configs/nvidia-master.yaml @@ -1732,12 +1732,14 @@ dsr1-fp4-b300-sglang: # The Pro recipe lists 8xB200 and 8xB300 with identical flags. Runs with # DP=8 + expert parallelism (no --tensor-parallel-size flag), FP8 KV cache, # block size 256, and an FP4 indexer cache. -dsv4-fp4-b300-vllm: +# precision is tagged fp4fp8 because the model runs fp4 weights with an +# fp8 KV cache. +dsv4-fp4fp8-b300-vllm: image: vllm/vllm-openai:deepseekv4-cu130 model: deepseek-ai/DeepSeek-V4-Pro model-prefix: dsv4 runner: b300 - precision: fp4 + precision: fp4fp8 framework: vllm multinode: false seq-len-configs: diff --git a/benchmarks/single_node/dsv4_fp4_b300.sh b/benchmarks/single_node/dsv4_fp4fp8_b300.sh similarity index 96% rename from benchmarks/single_node/dsv4_fp4_b300.sh rename to benchmarks/single_node/dsv4_fp4fp8_b300.sh index 6c4847fe3..e4f8507c8 100644 --- a/benchmarks/single_node/dsv4_fp4_b300.sh +++ b/benchmarks/single_node/dsv4_fp4fp8_b300.sh @@ -28,8 +28,8 @@ SERVER_LOG=/workspace/server.log PORT=${PORT:-8888} # DeepSeek-V4-Pro weights are large and engine startup on B300 can exceed -# the default 600s. Give it 20 minutes to load. -export VLLM_ENGINE_READY_TIMEOUT_S=1200 +# the default 600s. Give it an hour to load. +export VLLM_ENGINE_READY_TIMEOUT_S=3600 if [ "${EVAL_ONLY}" = "true" ]; then setup_eval_context diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 07119324a..196297d4b 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -1755,7 +1755,7 @@ - "VLLM_ENGINE_READY_TIMEOUT_S=3600 to accommodate large weight loading" - "Configs: 1k1k conc 4-64, 8k1k conc 4-64" pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1130 - + - config-keys: - dsv4-fp4-b300-sglang description: @@ -1777,17 +1777,19 @@ pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1134 - config-keys: - - dsv4-fp4-b300-vllm + - dsv4-fp4fp8-b300-vllm description: - "Add DeepSeek-V4-Pro vLLM B300 benchmark per https://vllm.ai/blog/deepseek-v4" - "Image: vllm/vllm-openai:deepseekv4-cu130" - "Model: deepseek-ai/DeepSeek-V4-Pro" - "Recipe lists 8xB200 and 8xB300 with identical flags; this mirrors the B200 config" - - "EP + DP=8 (no --tensor-parallel-size), FP8 KV cache, block size 256, FP4 indexer cache" + - "EP + DP=8 (no --tensor-parallel-size), FP4 weights + FP8 KV cache, block size 256, FP4 indexer cache" + - "Precision tag fp4fp8 reflects fp4 weights with fp8 KV cache" + - "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," - " --compilation-config cudagraph_mode=FULL_AND_PIECEWISE custom_ops=all," - " --attention_config.use_fp4_indexer_cache=True, --tokenizer-mode deepseek_v4," - " --tool-call-parser deepseek_v4, --enable-auto-tool-choice, --reasoning-parser deepseek_v4" - - "Configs: 1k1k conc 4-1024, 8k1k conc 4-512" - pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXXX + - "Configs: 1k1k conc 4-64, 8k1k conc 4-64" + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1128