From b998214623d2d608d629b8c1adcabf732d606adc Mon Sep 17 00:00:00 2001 From: Xin Li Date: Fri, 14 Aug 2026 15:18:33 -0400 Subject: [PATCH 1/4] feat(minimaxm3): refresh B300 AgentX Pareto points with EAGLE3-GQA Co-Authored-By: Claude Sonnet 4.6 --- .../single_node/agentic/minimaxm3_fp4_b300_mtp.sh | 2 +- configs/nvidia-master.yaml | 15 +++++++++------ perf-changelog.yaml | 7 +++++++ 3 files changed, 17 insertions(+), 7 deletions(-) diff --git a/benchmarks/single_node/agentic/minimaxm3_fp4_b300_mtp.sh b/benchmarks/single_node/agentic/minimaxm3_fp4_b300_mtp.sh index 3593cac4ce..da608ee847 100755 --- a/benchmarks/single_node/agentic/minimaxm3_fp4_b300_mtp.sh +++ b/benchmarks/single_node/agentic/minimaxm3_fp4_b300_mtp.sh @@ -43,7 +43,7 @@ install_agentic_deps OFFLOAD_ARGS=() if require_agentic_kv_offload_backend vllm-simple; then - CPU_OFFLOAD_BYTES=$((TOTAL_CPU_DRAM_GB * 1024 * 1024 * 1024)) + CPU_OFFLOAD_BYTES=$((TOTAL_CPU_DRAM_GB * 1000 * 1000 * 1000)) export VLLM_USE_SIMPLE_KV_OFFLOAD=1 OFFLOAD_CONFIG=$(printf \ '{"kv_connector":"SimpleCPUOffloadConnector","kv_role":"kv_both","kv_connector_extra_config":{"cpu_bytes_to_use":%d,"lazy_offload":true}}' \ diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml index c7a82fdb1a..d80b0530ad 100644 --- a/configs/nvidia-master.yaml +++ b/configs/nvidia-master.yaml @@ -7395,7 +7395,7 @@ qwen3.5-fp4-gb200-dynamo-sglang-agentic-mtp: decode: { num-worker: 0, tp: 2, ep: 2, dp-attn: false } minimaxm3-fp4-b300-vllm-agentic-mtp: - image: vllm/vllm-openai:nightly-5e35a6f4f9bbc217c599692157ca985c894373f7 + image: vllm/vllm-openai:nightly-ac7509e2b1db40fec2f03dde1ed4e9dfdc2338c9 model: nvidia/MiniMax-M3-NVFP4 model-prefix: minimaxm3 runner: cluster:b300-nv @@ -7404,13 +7404,16 @@ minimaxm3-fp4-b300-vllm-agentic-mtp: multinode: false scenarios: agentic-coding: - # 0.683 maps the B300 TP4 SimpleCPU point to a 1,024 GB - # engine-level CPU KV budget. GPU-resident points receive a zero budget. + # 0.683 maps the B300 TP4 SimpleCPU point to a 1,024 GB engine-level CPU + # KV budget. TP2 uses 1.0 so 2 of 8 GPUs receive the 750 GB policy maximum. - dram-utilization: 0.683 search-space: - - { tp: 4, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 2, 5, 10, 15, 20] } - - { tp: 2, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 2, 5] } - - { tp: 4, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: vllm-simple }, conc-list: [30, 40, 50, 60, 65, 70, 75] } + - { tp: 8, spec-decoding: mtp, kv-offloading: none, conc-list: [1] } + - { tp: 4, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 5, 10, 15, 20] } + - { tp: 4, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: vllm-simple }, conc-list: [30] } + - dram-utilization: 1.0 + search-space: + - { tp: 2, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: vllm-simple }, conc-list: [24] } # GB200 DeepSeek-V4 disaggregated AgentX frontier. The 3P/2D TEP8/TP8 curve # covers the middle/high-interactivity range omitted by the one-decode DEP diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 228b327a4a..0fc60f5ada 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -5971,3 +5971,10 @@ - "Compare the B200 TP8 and DEP8 concurrency range and extend throughput with 1P1D and 2P1D DEP8 disaggregation." - "Use KV-aware Dynamo routing with 4-hour correlation-ID affinity, authoritative vLLM KV events, KV-cache token metrics, and every logical vLLM server-metrics endpoint." pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2567 + +- config-keys: + - minimaxm3-fp4-b300-vllm-agentic-mtp + scenario-type: + - agentic-coding + description: + - "Refresh the MiniMax-M3 B300 AgentX submission as eight EAGLE3-GQA points: no-offload TP8 C1 and TP4 C1/C5/C10/C15/C20, TP4 C30 with 1,024 GB SimpleCPU KV offload, and TP2 C24 with the proportional-policy 750 GB SimpleCPU KV-offload cap." From 15cca5a37d76373fa4094087243282b5309ea210 Mon Sep 17 00:00:00 2001 From: Xin Li Date: Fri, 14 Aug 2026 15:18:59 -0400 Subject: [PATCH 2/4] chore: add PR link to perf-changelog for minimaxm3 B300 refresh Co-Authored-By: Claude Sonnet 4.6 --- perf-changelog.yaml | 1 + 1 file changed, 1 insertion(+) diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 0fc60f5ada..589a3bc799 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -5978,3 +5978,4 @@ - agentic-coding description: - "Refresh the MiniMax-M3 B300 AgentX submission as eight EAGLE3-GQA points: no-offload TP8 C1 and TP4 C1/C5/C10/C15/C20, TP4 C30 with 1,024 GB SimpleCPU KV offload, and TP2 C24 with the proportional-policy 750 GB SimpleCPU KV-offload cap." + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2610 From 8b46c842f4d99652d0d21c8950a5ff2ef675b782 Mon Sep 17 00:00:00 2001 From: Xin Li Date: Fri, 14 Aug 2026 15:21:29 -0400 Subject: [PATCH 3/4] revert: drop benchmark script byte-unit change from B300 PR Co-Authored-By: Claude Sonnet 4.6 --- benchmarks/single_node/agentic/minimaxm3_fp4_b300_mtp.sh | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/benchmarks/single_node/agentic/minimaxm3_fp4_b300_mtp.sh b/benchmarks/single_node/agentic/minimaxm3_fp4_b300_mtp.sh index da608ee847..3593cac4ce 100755 --- a/benchmarks/single_node/agentic/minimaxm3_fp4_b300_mtp.sh +++ b/benchmarks/single_node/agentic/minimaxm3_fp4_b300_mtp.sh @@ -43,7 +43,7 @@ install_agentic_deps OFFLOAD_ARGS=() if require_agentic_kv_offload_backend vllm-simple; then - CPU_OFFLOAD_BYTES=$((TOTAL_CPU_DRAM_GB * 1000 * 1000 * 1000)) + CPU_OFFLOAD_BYTES=$((TOTAL_CPU_DRAM_GB * 1024 * 1024 * 1024)) export VLLM_USE_SIMPLE_KV_OFFLOAD=1 OFFLOAD_CONFIG=$(printf \ '{"kv_connector":"SimpleCPUOffloadConnector","kv_role":"kv_both","kv_connector_extra_config":{"cpu_bytes_to_use":%d,"lazy_offload":true}}' \ From 086298aefe1180e56d5594c39d3308951b96aa91 Mon Sep 17 00:00:00 2001 From: Xin Li Date: Fri, 14 Aug 2026 15:23:48 -0400 Subject: [PATCH 4/4] chore: remove comment from B300 minimaxm3 agentic config Co-Authored-By: Claude Sonnet 4.6 --- configs/nvidia-master.yaml | 2 -- 1 file changed, 2 deletions(-) diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml index d80b0530ad..9993b32175 100644 --- a/configs/nvidia-master.yaml +++ b/configs/nvidia-master.yaml @@ -7404,8 +7404,6 @@ minimaxm3-fp4-b300-vllm-agentic-mtp: multinode: false scenarios: agentic-coding: - # 0.683 maps the B300 TP4 SimpleCPU point to a 1,024 GB engine-level CPU - # KV budget. TP2 uses 1.0 so 2 of 8 GPUs receive the 750 GB policy maximum. - dram-utilization: 0.683 search-space: - { tp: 8, spec-decoding: mtp, kv-offloading: none, conc-list: [1] }