From e4711d2bc27a60e80f38f0487c3d6957584add52 Mon Sep 17 00:00:00 2001
From: wzhao18 <wzhao18.sz@gmail.com>
Date: Fri, 3 Jul 2026 11:28:12 -0700
Subject: [PATCH 1/7] update

---
 .../fixed_seq_len/minimaxm3_fp4_b200_mtp.sh   | 35 +++++++++++++++++--
 configs/nvidia-master.yaml                    | 26 +++++++++-----
 2 files changed, 50 insertions(+), 11 deletions(-)

diff --git a/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_b200_mtp.sh b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_b200_mtp.sh
index 94ee390834..2608263863 100755
--- a/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_b200_mtp.sh
+++ b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_b200_mtp.sh
@@ -47,6 +47,7 @@ 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"
@@ -56,8 +57,36 @@ 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_b200.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). 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 ---
+  [1024:4:4:false:4]=4
+  [1024:8:1:false:2]=6
+  [1024:8:1:false:4]=4
+  # --- ISL=8192 / OSL=1024 ---
+  [8192:2:1:false:1]=4
+  [8192:2:1:false:256]=4
+  [8192:2:2:false:1]=4
+  [8192:2:2:false:16]=4
+  [8192:2:2:false:32]=2
+  [8192:2:2:false:64]=4
+  [8192:2:2:false:128]=4
+  [8192:2:2:false:256]=4
+  [8192:2:2:false:512]=2
+  [8192:4:1:false:256]=2
+  [8192:4:4:false:256]=4
+  [8192:8:1:false:1]=4
+  [8192:8:1:false:2]=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
@@ -68,7 +97,7 @@ start_gpu_monitor
 set -x
 vllm serve $MODEL --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..9e83a2c3c7 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -12784,20 +12784,30 @@ minimaxm3-fp4-b200-vllm-mtp:
   multinode: false
   scenarios:
     fixed-seq-len:
+    # Search space trimmed to the (parallelism, concurrency) points on the
+    # Total-TPS/GPU vs median-interactivity Pareto frontier of the EAGLE3 offline
+    # sweep (bench_results_b200.json). The best num_speculative_tokens varies
+    # along the frontier (fewer at the throughput end, up to 6 at the latency
+    # tail); the recipe shell (minimaxm3_fp4_b200_mtp.sh) selects it per point
+    # from the same ISL:TP:EP_SIZE:DP_ATTENTION:CONC key. TP2 (plain / EP) carries
+    # the throughput-to-mid front; TP4/TP8 hold the low-concurrency tail. DP-attn
+    # never reached the frontier on B200, so it is dropped.
     - 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: 2, conc-list: [64, 128, 256, 512], spec-decoding: mtp }
+      - { tp: 2, ep: 2, conc-list: [8, 512], spec-decoding: mtp }
+      - { tp: 4, conc-list: [2, 4, 16, 32, 64], spec-decoding: mtp }
+      - { tp: 4, ep: 4, conc-list: [4, 8], spec-decoding: mtp }
+      - { tp: 8, conc-list: [2, 4], 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: 2, conc-list: [1, 8, 16, 32, 256], spec-decoding: mtp }
+      - { tp: 2, ep: 2, conc-list: [1, 4, 16, 32, 64, 128, 256, 512], spec-decoding: mtp }
+      - { tp: 4, conc-list: [2, 8, 256], spec-decoding: mtp }
+      - { tp: 4, ep: 4, conc-list: [2, 4, 256], spec-decoding: mtp }
+      - { tp: 8, conc-list: [1, 2], spec-decoding: mtp }
 
 # EAGLE3 speculative-decoding (spec-decoding: mtp) variant of
 # minimaxm3-fp8-b300-vllm, pairing MiniMaxAI/MiniMax-M3-MXFP8 with the

From 6eb837d75c509c7512e3ec45220c7d12f9bb3000 Mon Sep 17 00:00:00 2001
From: wzhao18 <wzhao18.sz@gmail.com>
Date: Fri, 3 Jul 2026 11:29:19 -0700
Subject: [PATCH 2/7] update

---
 perf-changelog.yaml | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 3944e67c5f..fca225fcb5 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-b200-vllm-mtp
+  description:
+    - "Update image tag to nightly"
+    - "Update B200 MTP search space"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2007
\ No newline at end of file

From 11fcaa1e8c5a696a04f46c4c43d32e62954439aa Mon Sep 17 00:00:00 2001
From: Wei Zhao <51183510+wzhao18@users.noreply.github.com>
Date: Fri, 3 Jul 2026 14:30:44 -0400
Subject: [PATCH 3/7] 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 fca225fcb5..b7529a7a93 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -4439,4 +4439,4 @@
   description:
     - "Update image tag to nightly"
     - "Update B200 MTP search space"
-  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2007
\ No newline at end of file
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2007

From c50b6bcb59f2a651489632dc5587b859f6ace173 Mon Sep 17 00:00:00 2001
From: Wei Zhao <51183510+wzhao18@users.noreply.github.com>
Date: Fri, 3 Jul 2026 14:52:30 -0400
Subject: [PATCH 4/7] Update minimaxm3_fp4_b200_mtp.sh

---
 benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_b200_mtp.sh | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_b200_mtp.sh b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_b200_mtp.sh
index 2608263863..3d00f4e75d 100755
--- a/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_b200_mtp.sh
+++ b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_b200_mtp.sh
@@ -97,7 +97,7 @@ start_gpu_monitor
 set -x
 vllm serve $MODEL --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 \

From 26a532634f34b5477a589590db0082399c508cd7 Mon Sep 17 00:00:00 2001
From: Wei Zhao <51183510+wzhao18@users.noreply.github.com>
Date: Fri, 3 Jul 2026 16:40:19 -0400
Subject: [PATCH 5/7] Update nvidia-master.yaml

---
 configs/nvidia-master.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 9e83a2c3c7..786d08a3ae 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -12799,7 +12799,7 @@ minimaxm3-fp4-b200-vllm-mtp:
       - { tp: 2, ep: 2, conc-list: [8, 512], spec-decoding: mtp }
       - { tp: 4, conc-list: [2, 4, 16, 32, 64], spec-decoding: mtp }
       - { tp: 4, ep: 4, conc-list: [4, 8], spec-decoding: mtp }
-      - { tp: 8, conc-list: [2, 4], spec-decoding: mtp }
+      - { tp: 8, conc-list: [1, 2, 4], spec-decoding: mtp }
     - isl: 8192
       osl: 1024
       search-space:

From 989199014efc0201297fe30289eb10206fe7c88b Mon Sep 17 00:00:00 2001
From: Wei Zhao <51183510+wzhao18@users.noreply.github.com>
Date: Fri, 3 Jul 2026 16:40:41 -0400
Subject: [PATCH 6/7] Update nvidia-master.yaml

---
 configs/nvidia-master.yaml | 8 --------
 1 file changed, 8 deletions(-)

diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 786d08a3ae..f7686924ff 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -12784,14 +12784,6 @@ minimaxm3-fp4-b200-vllm-mtp:
   multinode: false
   scenarios:
     fixed-seq-len:
-    # Search space trimmed to the (parallelism, concurrency) points on the
-    # Total-TPS/GPU vs median-interactivity Pareto frontier of the EAGLE3 offline
-    # sweep (bench_results_b200.json). The best num_speculative_tokens varies
-    # along the frontier (fewer at the throughput end, up to 6 at the latency
-    # tail); the recipe shell (minimaxm3_fp4_b200_mtp.sh) selects it per point
-    # from the same ISL:TP:EP_SIZE:DP_ATTENTION:CONC key. TP2 (plain / EP) carries
-    # the throughput-to-mid front; TP4/TP8 hold the low-concurrency tail. DP-attn
-    # never reached the frontier on B200, so it is dropped.
     - isl: 1024
       osl: 1024
       search-space:

From 396eded8ce172386506051fc950f30b07836b9da Mon Sep 17 00:00:00 2001
From: Wei Zhao <51183510+wzhao18@users.noreply.github.com>
Date: Fri, 3 Jul 2026 21:12:45 -0400
Subject: [PATCH 7/7] Expand search space parameters in nvidia-master.yaml

---
 configs/nvidia-master.yaml | 16 ++++++++--------
 1 file changed, 8 insertions(+), 8 deletions(-)

diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index f7686924ff..00f4372515 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -12787,19 +12787,19 @@ minimaxm3-fp4-b200-vllm-mtp:
     - isl: 1024
       osl: 1024
       search-space:
-      - { tp: 2, conc-list: [64, 128, 256, 512], spec-decoding: mtp }
+      - { tp: 2, conc-list: [1, 2, 4, 8, 64, 128, 256, 512], spec-decoding: mtp }
       - { tp: 2, ep: 2, conc-list: [8, 512], spec-decoding: mtp }
-      - { tp: 4, conc-list: [2, 4, 16, 32, 64], spec-decoding: mtp }
-      - { tp: 4, ep: 4, conc-list: [4, 8], spec-decoding: mtp }
-      - { tp: 8, conc-list: [1, 2, 4], spec-decoding: mtp }
+      - { tp: 4, conc-list: [1, 2, 4, 16, 32, 64], spec-decoding: mtp }
+      - { tp: 4, ep: 4, conc-list: [1, 2, 4, 8], spec-decoding: mtp }
+      - { tp: 8, conc-list: [1, 2, 4, 8, 16], spec-decoding: mtp }
     - isl: 8192
       osl: 1024
       search-space:
       - { tp: 2, conc-list: [1, 8, 16, 32, 256], spec-decoding: mtp }
-      - { tp: 2, ep: 2, conc-list: [1, 4, 16, 32, 64, 128, 256, 512], spec-decoding: mtp }
-      - { tp: 4, conc-list: [2, 8, 256], spec-decoding: mtp }
-      - { tp: 4, ep: 4, conc-list: [2, 4, 256], spec-decoding: mtp }
-      - { tp: 8, conc-list: [1, 2], spec-decoding: mtp }
+      - { tp: 2, ep: 2, conc-list: [1, 2, 4, 16, 32, 64, 128, 256, 512], spec-decoding: mtp }
+      - { tp: 4, conc-list: [1, 2, 4, 8, 16, 256], spec-decoding: mtp }
+      - { tp: 4, ep: 4, conc-list: [1, 2, 4, 8, 256], spec-decoding: mtp }
+      - { tp: 8, conc-list: [1, 2, 4, 8], spec-decoding: mtp }
 
 # EAGLE3 speculative-decoding (spec-decoding: mtp) variant of
 # minimaxm3-fp8-b300-vllm, pairing MiniMaxAI/MiniMax-M3-MXFP8 with the