From a54d9ba74d69887eceed775f3387597feb3f51fb Mon Sep 17 00:00:00 2001 From: wzhao18 Date: Fri, 3 Jul 2026 11:13:05 -0700 Subject: [PATCH 1/6] update --- .../fixed_seq_len/minimaxm3_fp4_b300_mtp.sh | 36 +++++++++++++++++-- configs/nvidia-master.yaml | 34 +++++++++++------- 2 files changed, 54 insertions(+), 16 deletions(-) diff --git a/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_b300_mtp.sh b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_b300_mtp.sh index 74cbcd0202..786f89e0c9 100755 --- a/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_b300_mtp.sh +++ b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_b300_mtp.sh @@ -52,6 +52,7 @@ SERVER_LOG=/workspace/server.log export VLLM_ENGINE_READY_TIMEOUT_S=3600 export VLLM_FLOAT32_MATMUL_PRECISION=high +export VLLM_FLASHINFER_ALLREDUCE_BACKEND=trtllm if [ "${DP_ATTENTION}" = "true" ]; then PARALLEL_ARGS="--tensor-parallel-size=1 --data-parallel-size=$TP --enable-expert-parallel" @@ -61,8 +62,37 @@ else PARALLEL_ARGS="--tensor-parallel-size=$TP" fi -# use 3 speculative tokens for all configs for now -NUM_SPEC_TOKENS=3 +# Speculative-token count is picked per operating point to trace the +# Total-TPS/GPU vs median-interactivity Pareto frontier of the EAGLE3 offline +# sweep (bench_results.json). The best num_speculative_tokens is not constant: +# fewer tokens win at the throughput end (high concurrency), more tokens win at +# the latency end (low concurrency). The frontier only ever selects 1..4 tokens, +# so 3 is the default and the table below overrides the non-3 operating points. +# Key: ISL:TP:EP_SIZE:DP_ATTENTION:CONC (exactly the env vars the launcher sets; +# for dp-attn configs TP holds the data-parallel size, as in PARALLEL_ARGS below). +declare -A NUM_SPEC_TOKENS_MAP=( + # --- ISL=1024 / OSL=1024 --- + [1024:2:1:false:512]=1 + [1024:2:2:false:512]=1 + [1024:2:2:false:1]=4 + [1024:4:4:false:2]=4 + [1024:8:1:false:2]=2 + [1024:8:1:false:1]=4 + # --- ISL=8192 / OSL=1024 --- + [8192:2:1:false:2]=2 + [8192:2:1:false:4]=4 + [8192:2:1:false:32]=2 + [8192:2:1:false:256]=2 + [8192:2:2:false:32]=4 + [8192:2:2:false:128]=2 + [8192:2:2:true:128]=4 + [8192:4:1:false:1]=4 + [8192:4:1:false:2]=4 + [8192:4:4:false:1]=4 + [8192:8:1:false:1]=4 +) +NUM_SPEC_TOKENS="${NUM_SPEC_TOKENS_MAP[${ISL}:${TP}:${EP_SIZE}:${DP_ATTENTION}:${CONC}]:-3}" +echo "Selected NUM_SPEC_TOKENS=$NUM_SPEC_TOKENS for ISL=$ISL TP=$TP EP_SIZE=$EP_SIZE DP_ATTENTION=$DP_ATTENTION CONC=$CONC" if [ "${EVAL_ONLY}" = "true" ]; then setup_eval_context @@ -73,7 +103,7 @@ start_gpu_monitor set -x vllm serve "$MODEL_PATH" --served-model-name "$MODEL" --host 0.0.0.0 --port $PORT \ $PARALLEL_ARGS \ ---gpu-memory-utilization 0.90 \ +--gpu-memory-utilization 0.95 \ --max-model-len $MAX_MODEL_LEN \ --block-size 128 \ --language-model-only \ diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml index bdb6a46029..9a50bf21b3 100644 --- a/configs/nvidia-master.yaml +++ b/configs/nvidia-master.yaml @@ -12631,7 +12631,7 @@ minimaxm3-fp4-b300-vllm: # /scratch/models/MiniMax-M3-NVFP4 (added to the STAGED_MODELS allow-list in # launch_b300-nv.sh); the EAGLE3 draft is downloaded to the writable models dir. minimaxm3-fp4-b300-vllm-mtp: - image: vllm/vllm-openai:vllm-minimax-m3-perf-x86_64-13.0.1-8b00f41 + image: vllm/vllm-openai:nightly-93d8f834dd8acf33eb0e2a75b2711b628cb6e226 model: nvidia/MiniMax-M3-NVFP4 model-prefix: minimaxm3 runner: b300 @@ -12640,24 +12640,32 @@ minimaxm3-fp4-b300-vllm-mtp: multinode: false scenarios: fixed-seq-len: + # Search space trimmed to the (parallelism, concurrency) points that lie on + # the Total-TPS/GPU vs median-interactivity Pareto frontier of the EAGLE3 + # offline sweep (bench_results.json). The best num_speculative_tokens varies + # along the frontier (1 at the throughput end, up to 4 at the latency end); + # the recipe shell (minimaxm3_fp4_b300_mtp.sh) selects it per point from the + # same ISL:TP:EP_SIZE:DP_ATTENTION:CONC key. TP2 (plain / EP / DP-attention) + # carries the throughput-to-mid front; TP4/TP8 only survive at the + # low-concurrency, high-interactivity tail. - isl: 1024 osl: 1024 search-space: - - { tp: 8, conc-start: 1, conc-end: 64, spec-decoding: mtp } - - { tp: 8, ep: 8, conc-start: 1, conc-end: 256, spec-decoding: mtp } - - { tp: 4, conc-start: 1, conc-end: 64, spec-decoding: mtp } - - { tp: 4, ep: 4, conc-start: 64, conc-end: 256, spec-decoding: mtp } - - { tp: 4, ep: 4, dp-attn: true, conc-start: 128, conc-end: 512, spec-decoding: mtp } - - { tp: 8, ep: 8, dp-attn: true, conc-start: 256, conc-end: 512, spec-decoding: mtp } + - { tp: 2, conc-list: [2, 32, 64, 512], spec-decoding: mtp } + - { tp: 2, ep: 2, conc-list: [1, 128, 512], spec-decoding: mtp } + - { tp: 2, ep: 2, dp-attn: true, conc-list: [256, 512], spec-decoding: mtp } + - { tp: 4, conc-list: [8, 16, 32], spec-decoding: mtp } + - { tp: 4, ep: 4, conc-list: [2, 32], spec-decoding: mtp } + - { tp: 8, conc-list: [1, 2], spec-decoding: mtp } - isl: 8192 osl: 1024 search-space: - - { tp: 8, conc-start: 1, conc-end: 64, spec-decoding: mtp } - - { tp: 8, ep: 8, conc-start: 1, conc-end: 256, spec-decoding: mtp } - - { tp: 4, conc-start: 1, conc-end: 64, spec-decoding: mtp } - - { tp: 4, ep: 4, conc-start: 64, conc-end: 256, spec-decoding: mtp } - - { tp: 4, ep: 4, dp-attn: true, conc-start: 64, conc-end: 128, spec-decoding: mtp } - - { tp: 8, ep: 8, dp-attn: true, conc-start: 128, conc-end: 256, spec-decoding: mtp } + - { tp: 2, conc-list: [2, 4, 8, 16, 32, 128, 256], spec-decoding: mtp } + - { tp: 2, ep: 2, conc-list: [1, 2, 32, 64, 128], spec-decoding: mtp } + - { tp: 2, ep: 2, dp-attn: true, conc-list: [128], spec-decoding: mtp } + - { tp: 4, conc-list: [1, 2], spec-decoding: mtp } + - { tp: 4, ep: 4, conc-list: [1], spec-decoding: mtp } + - { tp: 8, conc-list: [1], spec-decoding: mtp } # MiniMax-M3 day-zero (https://recipes.vllm.ai/MiniMaxAI/MiniMax-M3). # 427B total / 26B active MoE with MSA sparse attention; MXFP8 checkpoint From 305ea25777c5a1aa5f4b72573766ca71c77da142 Mon Sep 17 00:00:00 2001 From: wzhao18 Date: Fri, 3 Jul 2026 11:16:20 -0700 Subject: [PATCH 2/6] update --- perf-changelog.yaml | 7 +++++++ 1 file changed, 7 insertions(+) diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 3944e67c5f..aea651062d 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -4433,3 +4433,10 @@ - "Add --online_quant_config with ptpc_fp8 and MoE layer exclusions (*block_sparse_moe) to all scripts." - "Replace deprecated AITER_QUICK_REDUCE_CAST_BF16_TO_FP16=0 and ATOM_M3_SPARSE_USE_ASM_PA=1 with ATOM_FORCE_ATTN_TRITON=1." pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2001 + +- config-keys: + - minimaxm3-fp4-b300-vllm-mtp + description: + - "Update image tag to nightly" + - "Update B300 MTP search space" + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2006 \ No newline at end of file From bab8412861383d1eaa432d19211a8621a6ccd1f5 Mon Sep 17 00:00:00 2001 From: Wei Zhao <51183510+wzhao18@users.noreply.github.com> Date: Fri, 3 Jul 2026 14:19:12 -0400 Subject: [PATCH 3/6] Update perf-changelog.yaml --- perf-changelog.yaml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/perf-changelog.yaml b/perf-changelog.yaml index aea651062d..8936854538 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -4439,4 +4439,4 @@ description: - "Update image tag to nightly" - "Update B300 MTP search space" - pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2006 \ No newline at end of file + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2006 From 938a1349c91ad22f525b4ef8d651828908321da0 Mon Sep 17 00:00:00 2001 From: wzhao18 Date: Fri, 3 Jul 2026 13:36:27 -0700 Subject: [PATCH 4/6] fixup --- benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_b300_mtp.sh | 1 - configs/nvidia-master.yaml | 2 +- 2 files changed, 1 insertion(+), 2 deletions(-) diff --git a/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_b300_mtp.sh b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_b300_mtp.sh index 786f89e0c9..e4fdc2f7d9 100755 --- a/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_b300_mtp.sh +++ b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_b300_mtp.sh @@ -76,7 +76,6 @@ declare -A NUM_SPEC_TOKENS_MAP=( [1024:2:2:false:512]=1 [1024:2:2:false:1]=4 [1024:4:4:false:2]=4 - [1024:8:1:false:2]=2 [1024:8:1:false:1]=4 # --- ISL=8192 / OSL=1024 --- [8192:2:1:false:2]=2 diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml index 9a50bf21b3..91fc7fcd90 100644 --- a/configs/nvidia-master.yaml +++ b/configs/nvidia-master.yaml @@ -12656,7 +12656,7 @@ minimaxm3-fp4-b300-vllm-mtp: - { tp: 2, ep: 2, dp-attn: true, conc-list: [256, 512], spec-decoding: mtp } - { tp: 4, conc-list: [8, 16, 32], spec-decoding: mtp } - { tp: 4, ep: 4, conc-list: [2, 32], spec-decoding: mtp } - - { tp: 8, conc-list: [1, 2], spec-decoding: mtp } + - { tp: 8, conc-list: [1], spec-decoding: mtp } - isl: 8192 osl: 1024 search-space: From 47bd6ca76ab4631b504b77264914234791309044 Mon Sep 17 00:00:00 2001 From: Wei Zhao <51183510+wzhao18@users.noreply.github.com> Date: Fri, 3 Jul 2026 19:09:51 -0400 Subject: [PATCH 5/6] Update nvidia-master.yaml --- configs/nvidia-master.yaml | 20 ++++++++++---------- 1 file changed, 10 insertions(+), 10 deletions(-) diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml index 91fc7fcd90..cb11596b1d 100644 --- a/configs/nvidia-master.yaml +++ b/configs/nvidia-master.yaml @@ -12651,21 +12651,21 @@ minimaxm3-fp4-b300-vllm-mtp: - isl: 1024 osl: 1024 search-space: - - { tp: 2, conc-list: [2, 32, 64, 512], spec-decoding: mtp } - - { tp: 2, ep: 2, conc-list: [1, 128, 512], spec-decoding: mtp } + - { tp: 2, conc-list: [1, 2, 4, 8, 16, 32, 64, 512], spec-decoding: mtp } + - { tp: 2, ep: 2, conc-list: [1, 2, 4, 8, 16, 128, 512], spec-decoding: mtp } - { tp: 2, ep: 2, dp-attn: true, conc-list: [256, 512], spec-decoding: mtp } - - { tp: 4, conc-list: [8, 16, 32], spec-decoding: mtp } - - { tp: 4, ep: 4, conc-list: [2, 32], spec-decoding: mtp } - - { tp: 8, conc-list: [1], spec-decoding: mtp } + - { tp: 4, conc-list: [1, 2, 4, 8, 16, 32], spec-decoding: mtp } + - { tp: 4, ep: 4, conc-list: [1, 2, 32], spec-decoding: mtp } + - { tp: 8, conc-list: [1, 2, 4, 8], spec-decoding: mtp } - isl: 8192 osl: 1024 search-space: - - { tp: 2, conc-list: [2, 4, 8, 16, 32, 128, 256], spec-decoding: mtp } - - { tp: 2, ep: 2, conc-list: [1, 2, 32, 64, 128], spec-decoding: mtp } + - { tp: 2, conc-list: [1, 2, 4, 8, 16, 32, 128, 256], spec-decoding: mtp } + - { tp: 2, ep: 2, conc-list: [1, 2, 4, 8, 32, 64, 128], spec-decoding: mtp } - { tp: 2, ep: 2, dp-attn: true, conc-list: [128], spec-decoding: mtp } - - { tp: 4, conc-list: [1, 2], spec-decoding: mtp } - - { tp: 4, ep: 4, conc-list: [1], spec-decoding: mtp } - - { tp: 8, conc-list: [1], spec-decoding: mtp } + - { tp: 4, conc-list: [1, 2, 4, 8], spec-decoding: mtp } + - { tp: 4, ep: 4, conc-list: [1, 2, 4, 8], spec-decoding: mtp } + - { tp: 8, conc-list: [1, 2, 4, 8], spec-decoding: mtp } # MiniMax-M3 day-zero (https://recipes.vllm.ai/MiniMaxAI/MiniMax-M3). # 427B total / 26B active MoE with MSA sparse attention; MXFP8 checkpoint From 0523519b93f68fbf4bd76d640ffdc3f880875c47 Mon Sep 17 00:00:00 2001 From: Wei Zhao <51183510+wzhao18@users.noreply.github.com> Date: Fri, 3 Jul 2026 20:46:42 -0400 Subject: [PATCH 6/6] Adjust GPU memory utilization to 0.9 --- benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_b300_mtp.sh | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_b300_mtp.sh b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_b300_mtp.sh index e4fdc2f7d9..4ed72bc022 100755 --- a/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_b300_mtp.sh +++ b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_b300_mtp.sh @@ -102,7 +102,7 @@ start_gpu_monitor set -x vllm serve "$MODEL_PATH" --served-model-name "$MODEL" --host 0.0.0.0 --port $PORT \ $PARALLEL_ARGS \ ---gpu-memory-utilization 0.95 \ +--gpu-memory-utilization 0.9 \ --max-model-len $MAX_MODEL_LEN \ --block-size 128 \ --language-model-only \