From 45142771e2894e1eb3f01ce83693416cfc249536 Mon Sep 17 00:00:00 2001 From: jasonlizhengjian Date: Mon, 15 Jun 2026 12:03:00 -0700 Subject: [PATCH 1/8] Update MiniMax M3 B300 vLLM serving settings --- .github/configs/nvidia-master.yaml | 3 +++ .../single_node/fixed_seq_len/minimaxm3_fp8_b300.sh | 9 ++------- perf-changelog.yaml | 7 +++++++ 3 files changed, 12 insertions(+), 7 deletions(-) diff --git a/.github/configs/nvidia-master.yaml b/.github/configs/nvidia-master.yaml index 5293fcb1df..3d6c6f0005 100644 --- a/.github/configs/nvidia-master.yaml +++ b/.github/configs/nvidia-master.yaml @@ -12618,6 +12618,7 @@ minimaxm3-fp8-b300-vllm: - { tp: 8, ep: 8, conc-start: 1, conc-end: 512 } - { tp: 4, conc-start: 1, conc-end: 64 } - { tp: 4, ep: 4, conc-start: 64, conc-end: 512 } + - { tp: 4, ep: 4, dp-attn: true, conc-start: 128, conc-end: 512 } # tp2 fits MXFP8 weights (~222 GB/GPU of 288) but KV headroom is thin; # 1k1k only, drop if it OOMs at the high end. - { tp: 2, ep: 2, conc-start: 16, conc-end: 128 } @@ -12628,6 +12629,8 @@ minimaxm3-fp8-b300-vllm: - { tp: 8, conc-start: 1, conc-end: 64 } - { tp: 8, ep: 8, conc-start: 1, conc-end: 512 } - { tp: 4, conc-start: 1, conc-end: 128 } + - { tp: 4, ep: 4, conc-start: 64, conc-end: 256 } + - { tp: 4, ep: 4, dp-attn: true, conc-start: 64, conc-end: 128 } - { tp: 8, ep: 8, dp-attn: true, conc-start: 128, conc-end: 512 } # MiniMax-M3 day-zero (https://recipes.vllm.ai/MiniMaxAI/MiniMax-M3). diff --git a/benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_b300.sh b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_b300.sh index 9d19d37248..53dd5a5123 100755 --- a/benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_b300.sh +++ b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_b300.sh @@ -46,6 +46,7 @@ SERVER_LOG=/workspace/server.log # 444 GB of MXFP8 weights off shared FS; engine startup can exceed the # default 600s readiness window. export VLLM_ENGINE_READY_TIMEOUT_S=3600 +export VLLM_FLOAT32_MATMUL_PRECISION=high if [ "${DP_ATTENTION}" = "true" ]; then PARALLEL_ARGS="--tensor-parallel-size=1 --data-parallel-size=$TP --enable-expert-parallel" @@ -55,12 +56,6 @@ else PARALLEL_ARGS="--tensor-parallel-size=$TP" fi -# Fixed-seq-len runs don't need graphs past the request concurrency: capture -# up to the next power of two >= CONC, capped at vLLM's 2048 ceiling. -CAPTURE_SIZE=4 -while (( CAPTURE_SIZE < CONC )); do CAPTURE_SIZE=$((CAPTURE_SIZE * 2)); done -(( CAPTURE_SIZE > 2048 )) && CAPTURE_SIZE=2048 - if [ "${EVAL_ONLY}" = "true" ]; then setup_eval_context MAX_MODEL_LEN="$EVAL_MAX_MODEL_LEN" @@ -75,7 +70,7 @@ $PARALLEL_ARGS \ --max-model-len $MAX_MODEL_LEN \ --block-size 128 \ --language-model-only \ ---max-cudagraph-capture-size $CAPTURE_SIZE \ +--max-cudagraph-capture-size 2048 \ --max-num-batched-tokens "$((ISL * 2 ))" \ --stream-interval 20 --no-enable-prefix-caching \ --trust-remote-code > $SERVER_LOG 2>&1 & diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 56daa76c0e..103a56bd56 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -3849,3 +3849,10 @@ - "Align MiniMax-M3 B200 vLLM fixed-sequence serving with MiniMax-M2.5 FP8 B200 settings by setting VLLM_FLOAT32_MATMUL_PRECISION=high and restoring max cudagraph capture size 2048." - "Add TP4+EP4 coverage for MiniMax-M3 B200: DP-attention rows for 1k1k/8k1k and the missing non-DP-attention row for 8k1k." pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1779 + +- config-keys: + - minimaxm3-fp8-b300-vllm + description: + - "Align MiniMax-M3 B300 vLLM fixed-sequence serving with MiniMax-M2.5 FP8 settings by setting VLLM_FLOAT32_MATMUL_PRECISION=high and restoring max cudagraph capture size 2048." + - "Add TP4+EP4 coverage for MiniMax-M3 B300: DP-attention rows for 1k1k/8k1k and the missing non-DP-attention row for 8k1k." + pr-link: TODO From 1546aebd07bc1806f80e1e85a512c3c568a6cf76 Mon Sep 17 00:00:00 2001 From: jasonlizhengjian Date: Mon, 15 Jun 2026 12:03:34 -0700 Subject: [PATCH 2/8] Update MiniMax M3 B300 changelog PR link --- perf-changelog.yaml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 103a56bd56..f67f061cd6 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -3855,4 +3855,4 @@ description: - "Align MiniMax-M3 B300 vLLM fixed-sequence serving with MiniMax-M2.5 FP8 settings by setting VLLM_FLOAT32_MATMUL_PRECISION=high and restoring max cudagraph capture size 2048." - "Add TP4+EP4 coverage for MiniMax-M3 B300: DP-attention rows for 1k1k/8k1k and the missing non-DP-attention row for 8k1k." - pr-link: TODO + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1781 From 16ccdff51724f7582b45b0710f661854adf06705 Mon Sep 17 00:00:00 2001 From: jasonlizhengjian Date: Mon, 15 Jun 2026 12:31:27 -0700 Subject: [PATCH 3/8] Update MiniMax M3 B200 B300 MTP settings --- .github/configs/nvidia-master.yaml | 6 ++++++ .../fixed_seq_len/minimaxm3_fp8_b200_mtp.sh | 11 ++--------- .../fixed_seq_len/minimaxm3_fp8_b300_mtp.sh | 11 ++--------- perf-changelog.yaml | 8 ++++++++ 4 files changed, 18 insertions(+), 18 deletions(-) diff --git a/.github/configs/nvidia-master.yaml b/.github/configs/nvidia-master.yaml index 5293fcb1df..8e84531f20 100644 --- a/.github/configs/nvidia-master.yaml +++ b/.github/configs/nvidia-master.yaml @@ -12693,6 +12693,7 @@ minimaxm3-fp8-b200-vllm-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 } - isl: 8192 osl: 1024 @@ -12700,6 +12701,8 @@ minimaxm3-fp8-b200-vllm-mtp: - { 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 } # EAGLE3 speculative-decoding (spec-decoding: mtp) variant of @@ -12727,6 +12730,7 @@ minimaxm3-fp8-b300-vllm-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 } - isl: 8192 osl: 1024 @@ -12734,6 +12738,8 @@ minimaxm3-fp8-b300-vllm-mtp: - { 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 } # EAGLE3 speculative-decoding (spec-decoding: mtp) variant of diff --git a/benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_b200_mtp.sh b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_b200_mtp.sh index 42e3b9e254..dde0e69c0f 100644 --- a/benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_b200_mtp.sh +++ b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_b200_mtp.sh @@ -61,6 +61,7 @@ SERVER_LOG=/workspace/server.log # 444 GB of MXFP8 weights off shared FS; engine startup can exceed the # default 600s readiness window. export VLLM_ENGINE_READY_TIMEOUT_S=3600 +export VLLM_FLOAT32_MATMUL_PRECISION=high if [ "${DP_ATTENTION}" = "true" ]; then PARALLEL_ARGS="--tensor-parallel-size=1 --data-parallel-size=$TP --enable-expert-parallel" @@ -73,14 +74,6 @@ fi # use 3 speculative tokens for all configs for now NUM_SPEC_TOKENS=3 -# Fixed-seq-len runs don't need graphs past the decode step's token count: -# with spec decoding every running request contributes 1 + NUM_SPEC_TOKENS -# tokens per step, so capture up to the next power of two >= -# CONC * (1 + NUM_SPEC_TOKENS), capped at vLLM's 2048 ceiling. -CAPTURE_SIZE=4 -while (( CAPTURE_SIZE < CONC * (1 + NUM_SPEC_TOKENS) )); do CAPTURE_SIZE=$((CAPTURE_SIZE * 2)); done -(( CAPTURE_SIZE > 2048 )) && CAPTURE_SIZE=2048 - if [ "${EVAL_ONLY}" = "true" ]; then setup_eval_context MAX_MODEL_LEN="$EVAL_MAX_MODEL_LEN" @@ -95,7 +88,7 @@ $PARALLEL_ARGS \ --max-model-len $MAX_MODEL_LEN \ --block-size 128 \ --language-model-only \ ---max-cudagraph-capture-size $CAPTURE_SIZE \ +--max-cudagraph-capture-size 2048 \ --max-num-batched-tokens "$((ISL * 2 ))" \ --speculative-config "{\"method\": \"eagle3\", \"model\": \"$DRAFT_MODEL_PATH\", \"num_speculative_tokens\": $NUM_SPEC_TOKENS, \"attention_backend\": \"FLASH_ATTN\"}" \ --stream-interval 20 --no-enable-prefix-caching \ diff --git a/benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_b300_mtp.sh b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_b300_mtp.sh index b832f1f390..a93da87b0f 100644 --- a/benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_b300_mtp.sh +++ b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_b300_mtp.sh @@ -62,6 +62,7 @@ SERVER_LOG=/workspace/server.log # 444 GB of MXFP8 weights off shared FS; engine startup can exceed the # default 600s readiness window. export VLLM_ENGINE_READY_TIMEOUT_S=3600 +export VLLM_FLOAT32_MATMUL_PRECISION=high if [ "${DP_ATTENTION}" = "true" ]; then PARALLEL_ARGS="--tensor-parallel-size=1 --data-parallel-size=$TP --enable-expert-parallel" @@ -74,14 +75,6 @@ fi # use 3 speculative tokens for all configs for now NUM_SPEC_TOKENS=3 -# Fixed-seq-len runs don't need graphs past the decode step's token count: -# with spec decoding every running request contributes 1 + NUM_SPEC_TOKENS -# tokens per step, so capture up to the next power of two >= -# CONC * (1 + NUM_SPEC_TOKENS), capped at vLLM's 2048 ceiling. -CAPTURE_SIZE=4 -while (( CAPTURE_SIZE < CONC * (1 + NUM_SPEC_TOKENS) )); do CAPTURE_SIZE=$((CAPTURE_SIZE * 2)); done -(( CAPTURE_SIZE > 2048 )) && CAPTURE_SIZE=2048 - if [ "${EVAL_ONLY}" = "true" ]; then setup_eval_context MAX_MODEL_LEN="$EVAL_MAX_MODEL_LEN" @@ -96,7 +89,7 @@ $PARALLEL_ARGS \ --max-model-len $MAX_MODEL_LEN \ --block-size 128 \ --language-model-only \ ---max-cudagraph-capture-size $CAPTURE_SIZE \ +--max-cudagraph-capture-size 2048 \ --max-num-batched-tokens "$((ISL * 2 ))" \ --speculative-config "{\"method\": \"eagle3\", \"model\": \"$DRAFT_MODEL_PATH\", \"num_speculative_tokens\": $NUM_SPEC_TOKENS, \"attention_backend\": \"FLASH_ATTN\"}" \ --stream-interval 20 --no-enable-prefix-caching \ diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 56daa76c0e..88e454a0a5 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -3849,3 +3849,11 @@ - "Align MiniMax-M3 B200 vLLM fixed-sequence serving with MiniMax-M2.5 FP8 B200 settings by setting VLLM_FLOAT32_MATMUL_PRECISION=high and restoring max cudagraph capture size 2048." - "Add TP4+EP4 coverage for MiniMax-M3 B200: DP-attention rows for 1k1k/8k1k and the missing non-DP-attention row for 8k1k." pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1779 + +- config-keys: + - minimaxm3-fp8-b200-vllm-mtp + - minimaxm3-fp8-b300-vllm-mtp + description: + - "Align MiniMax-M3 B200/B300 EAGLE3 MTP serving with the MiniMax-M2.5 FP8 serving settings by setting VLLM_FLOAT32_MATMUL_PRECISION=high and using max cudagraph capture size 2048." + - "Add TP4+EP4 MTP coverage: DP-attention rows for 1k1k/8k1k and the missing non-DP-attention row for 8k1k." + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXXX From 389037e714792a4236054b1180b9188d5f0b46f5 Mon Sep 17 00:00:00 2001 From: jasonlizhengjian Date: Mon, 15 Jun 2026 12:32:22 -0700 Subject: [PATCH 4/8] Update MiniMax M3 MTP changelog link --- perf-changelog.yaml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 88e454a0a5..59c7c8729d 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -3856,4 +3856,4 @@ description: - "Align MiniMax-M3 B200/B300 EAGLE3 MTP serving with the MiniMax-M2.5 FP8 serving settings by setting VLLM_FLOAT32_MATMUL_PRECISION=high and using max cudagraph capture size 2048." - "Add TP4+EP4 MTP coverage: DP-attention rows for 1k1k/8k1k and the missing non-DP-attention row for 8k1k." - pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXXX + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1784 From 161adc023c95f177060725c3088893c100c0a4e0 Mon Sep 17 00:00:00 2001 From: jasonlizhengjian Date: Tue, 16 Jun 2026 08:40:12 -0700 Subject: [PATCH 5/8] Use Marlin for MiniMax M3 TP-only configs --- .../single_node/fixed_seq_len/minimaxm3_fp8_b200.sh | 2 +- .../single_node/fixed_seq_len/minimaxm3_fp8_b200_mtp.sh | 2 +- .../single_node/fixed_seq_len/minimaxm3_fp8_b300.sh | 2 +- .../single_node/fixed_seq_len/minimaxm3_fp8_b300_mtp.sh | 2 +- perf-changelog.yaml | 9 +++++++++ 5 files changed, 13 insertions(+), 4 deletions(-) diff --git a/benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_b200.sh b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_b200.sh index 49ab8df755..16041a2ea4 100755 --- a/benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_b200.sh +++ b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_b200.sh @@ -45,7 +45,7 @@ if [ "${DP_ATTENTION}" = "true" ]; then elif [ "$EP_SIZE" -gt 1 ]; then PARALLEL_ARGS="--tensor-parallel-size=$TP --enable-expert-parallel" else - PARALLEL_ARGS="--tensor-parallel-size=$TP" + PARALLEL_ARGS="--tensor-parallel-size=$TP --moe-backend marlin" fi if [ "${EVAL_ONLY}" = "true" ]; then diff --git a/benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_b200_mtp.sh b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_b200_mtp.sh index dde0e69c0f..54aabd3422 100644 --- a/benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_b200_mtp.sh +++ b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_b200_mtp.sh @@ -68,7 +68,7 @@ if [ "${DP_ATTENTION}" = "true" ]; then elif [ "$EP_SIZE" -gt 1 ]; then PARALLEL_ARGS="--tensor-parallel-size=$TP --enable-expert-parallel" else - PARALLEL_ARGS="--tensor-parallel-size=$TP" + PARALLEL_ARGS="--tensor-parallel-size=$TP --moe-backend marlin" fi # use 3 speculative tokens for all configs for now diff --git a/benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_b300.sh b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_b300.sh index 53dd5a5123..1f4187c044 100755 --- a/benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_b300.sh +++ b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_b300.sh @@ -53,7 +53,7 @@ if [ "${DP_ATTENTION}" = "true" ]; then elif [ "$EP_SIZE" -gt 1 ]; then PARALLEL_ARGS="--tensor-parallel-size=$TP --enable-expert-parallel" else - PARALLEL_ARGS="--tensor-parallel-size=$TP" + PARALLEL_ARGS="--tensor-parallel-size=$TP --moe-backend marlin" fi if [ "${EVAL_ONLY}" = "true" ]; then diff --git a/benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_b300_mtp.sh b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_b300_mtp.sh index a93da87b0f..742933deb1 100644 --- a/benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_b300_mtp.sh +++ b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_b300_mtp.sh @@ -69,7 +69,7 @@ if [ "${DP_ATTENTION}" = "true" ]; then elif [ "$EP_SIZE" -gt 1 ]; then PARALLEL_ARGS="--tensor-parallel-size=$TP --enable-expert-parallel" else - PARALLEL_ARGS="--tensor-parallel-size=$TP" + PARALLEL_ARGS="--tensor-parallel-size=$TP --moe-backend marlin" fi # use 3 speculative tokens for all configs for now diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 5e5bebde80..39f829c1e2 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -3879,3 +3879,12 @@ - "Align MiniMax-M3 B200/B300 EAGLE3 MTP serving with the MiniMax-M2.5 FP8 serving settings by setting VLLM_FLOAT32_MATMUL_PRECISION=high and using max cudagraph capture size 2048." - "Add TP4+EP4 MTP coverage: DP-attention rows for 1k1k/8k1k and the missing non-DP-attention row for 8k1k." pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1784 + +- config-keys: + - minimaxm3-fp8-b200-vllm + - minimaxm3-fp8-b300-vllm + - minimaxm3-fp8-b200-vllm-mtp + - minimaxm3-fp8-b300-vllm-mtp + description: + - "Use the Marlin MoE backend for MiniMax-M3 B200/B300 TP-only vLLM configurations by adding --moe-backend marlin when expert parallelism is disabled." + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXXX From 02f6f476a3b8363fafffa2d912df411e73ad7891 Mon Sep 17 00:00:00 2001 From: jasonlizhengjian Date: Tue, 16 Jun 2026 08:41:04 -0700 Subject: [PATCH 6/8] Update MiniMax M3 Marlin changelog link --- perf-changelog.yaml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 39f829c1e2..3a8bae8375 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -3887,4 +3887,4 @@ - minimaxm3-fp8-b300-vllm-mtp description: - "Use the Marlin MoE backend for MiniMax-M3 B200/B300 TP-only vLLM configurations by adding --moe-backend marlin when expert parallelism is disabled." - pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXXX + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1807 From 32c5fe0411fc7fb0ec37de2c72a4e8741b0e37f3 Mon Sep 17 00:00:00 2001 From: jasonlizhengjian Date: Tue, 16 Jun 2026 08:51:45 -0700 Subject: [PATCH 7/8] Update MiniMax M3 Marlin changelog link --- perf-changelog.yaml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 3a8bae8375..4d57724167 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -3887,4 +3887,4 @@ - minimaxm3-fp8-b300-vllm-mtp description: - "Use the Marlin MoE backend for MiniMax-M3 B200/B300 TP-only vLLM configurations by adding --moe-backend marlin when expert parallelism is disabled." - pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1807 + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1809 From a580725795bada4a43687331e02658c25c9cbf57 Mon Sep 17 00:00:00 2001 From: jasonlizhengjian Date: Tue, 16 Jun 2026 08:55:57 -0700 Subject: [PATCH 8/8] Preserve append-only changelog --- perf-changelog.yaml | 7 +++++++ 1 file changed, 7 insertions(+) diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 4d57724167..39e63a8a1d 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -3857,6 +3857,13 @@ - "Add TP4+EP4 coverage for MiniMax-M3 B200: DP-attention rows for 1k1k/8k1k and the missing non-DP-attention row for 8k1k." pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1779 +- config-keys: + - minimaxm3-fp8-b200-vllm + description: + - "Align MiniMax-M3 B200 vLLM fixed-sequence serving with MiniMax-M2.5 FP8 B200 settings by setting VLLM_FLOAT32_MATMUL_PRECISION=high and restoring max cudagraph capture size 2048." + - "Add TP4+EP4 coverage for MiniMax-M3 B200: DP-attention rows for 1k1k/8k1k and the missing non-DP-attention row for 8k1k." + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1779 + - config-keys: - kimik2.5-int4-mi355x-vllm description: