From 4a91df90fd9d6606d03774f085f7e7ca2dee9732 Mon Sep 17 00:00:00 2001
From: hshrivastava-droid <hshrivastava@nvidia.com>
Date: Fri, 24 Jul 2026 12:01:05 -0700
Subject: [PATCH 01/13] add b300 vllm agentX single-node minimaxm3 fp4
 EAGLE-GQA MTP

---
 .../agentic/minimaxm3_fp4_b300_mtp.sh         | 100 ++++++++++++++++++
 configs/nvidia-master.yaml                    |  16 +++
 perf-changelog.yaml                           |  19 +++-
 3 files changed, 130 insertions(+), 5 deletions(-)
 create mode 100755 benchmarks/single_node/agentic/minimaxm3_fp4_b300_mtp.sh

diff --git a/benchmarks/single_node/agentic/minimaxm3_fp4_b300_mtp.sh b/benchmarks/single_node/agentic/minimaxm3_fp4_b300_mtp.sh
new file mode 100755
index 0000000000..ff43eebc14
--- /dev/null
+++ b/benchmarks/single_node/agentic/minimaxm3_fp4_b300_mtp.sh
@@ -0,0 +1,100 @@
+#!/usr/bin/env bash
+set -euo pipefail
+set -x
+
+# MiniMax-M3 NVFP4 B300 AgentX with EAGLE3-GQA and synthetic acceptance.
+
+source "$(dirname "$0")/../../benchmark_lib.sh"
+
+export EVAL_FRAMEWORK="lm-eval"
+
+check_env_vars MODEL TP CONC KV_OFFLOADING RESULT_DIR DURATION
+
+DRAFT_MODEL="Inferact/MiniMax-M3-EAGLE3-GQA"
+NUM_SPEC_TOKENS=3
+SYNTHETIC_ACCEPT_LEN=2.78
+
+if [[ -n "${SLURM_JOB_ID:-}" ]]; then
+    echo "JOB $SLURM_JOB_ID running on ${SLURMD_NODENAME:-unknown}"
+fi
+
+if [[ -n "${MODEL_PATH:-}" ]]; then
+    if [[ ! -d "$MODEL_PATH" || -z "$(ls -A "$MODEL_PATH" 2>/dev/null)" ]]; then
+        hf download "$MODEL" --local-dir "$MODEL_PATH"
+    fi
+    DRAFT_MODEL_PATH="/data/models/${DRAFT_MODEL##*/}"
+    if [[ ! -d "$DRAFT_MODEL_PATH" || -z "$(ls -A "$DRAFT_MODEL_PATH" 2>/dev/null)" ]]; then
+        hf download "$DRAFT_MODEL" --local-dir "$DRAFT_MODEL_PATH"
+    fi
+else
+    hf download "$MODEL"
+    export MODEL_PATH="$MODEL"
+    hf download "$DRAFT_MODEL"
+    DRAFT_MODEL_PATH="$DRAFT_MODEL"
+fi
+
+nvidia-smi
+resolve_trace_source
+install_agentic_deps
+
+if agentic_kv_offload_enabled; then
+    echo "Error: this submission contains GPU-resident KV points only" >&2
+    exit 1
+fi
+
+export PYTHONNOUSERSITE=1
+export VLLM_ENGINE_READY_TIMEOUT_S=3600
+export VLLM_FLOAT32_MATMUL_PRECISION=high
+export VLLM_FLASHINFER_ALLREDUCE_BACKEND=trtllm
+
+SERVER_LOG="$RESULT_DIR/server.log"
+mkdir -p "$RESULT_DIR"
+
+SERVER_PID=""
+cleanup_agentic_services() {
+    local exit_code=$?
+    trap - EXIT INT TERM
+    set +e
+    stop_background_process_tree "$SERVER_PID" "vLLM server" 60
+    exit "$exit_code"
+}
+trap cleanup_agentic_services EXIT
+trap 'exit 130' INT
+trap 'exit 143' TERM
+
+SPEC_CONFIG=$(printf \
+    '{"method":"eagle3","model":"%s","num_speculative_tokens":%d,"attention_backend":"FLASH_ATTN","rejection_sample_method":"synthetic","synthetic_acceptance_length":%.2f}' \
+    "$DRAFT_MODEL_PATH" "$NUM_SPEC_TOKENS" "$SYNTHETIC_ACCEPT_LEN")
+
+{ set +x; } 2>/dev/null
+VLLM_CMD=(
+    vllm serve "$MODEL_PATH"
+    --served-model-name "$MODEL"
+    --host 0.0.0.0
+    --port "$PORT"
+    --tensor-parallel-size "$TP"
+    --gpu-memory-utilization 0.9
+    --block-size 128
+    --language-model-only
+    --enable-prefix-caching
+    --no-enable-flashinfer-autotune
+    --reasoning-parser minimax_m3
+    --default-chat-template-kwargs '{"thinking_mode":"enabled"}'
+    --attention-config '{"backend":"FLASHINFER","use_trtllm_attention":true,"indexer_kv_dtype":"fp8"}'
+    --all2all-backend flashinfer_nvlink_one_sided
+    --max-cudagraph-capture-size 512
+    --max-num-batched-tokens 16384
+    --stream-interval 20
+    --trust-remote-code
+    --speculative-config "$SPEC_CONFIG"
+)
+printf '%q ' "${VLLM_CMD[@]}" | tee "$RESULT_DIR/vllm_command.txt"
+printf '\n' | tee -a "$RESULT_DIR/vllm_command.txt"
+"${VLLM_CMD[@]}" > "$SERVER_LOG" 2>&1 &
+SERVER_PID=$!
+echo "Server PID: $SERVER_PID"
+set -x
+
+wait_for_server_ready --port "$PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
+build_replay_cmd "$RESULT_DIR"
+run_agentic_replay_and_write_outputs "$RESULT_DIR"
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 8cc99a8a0b..09a42d9901 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -7800,6 +7800,22 @@ minimaxm3-fp8-h200-vllm-agentic:
       - { tp: 8, ep: 8, kv-offloading: none, conc-list: [2, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 16, 18, 20] }
       - { tp: 8, ep: 8, kv-offloading: dram, kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }, conc-list: [5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 16, 18, 20] }
 
+minimaxm3-fp4-b300-vllm-agentic-mtp:
+  image: vllm/vllm-openai:nightly-387189c42997b27e2c04b5d97ef8190ffa2bf909
+  model: nvidia/MiniMax-M3-NVFP4
+  model-prefix: minimaxm3
+  runner: cluster:b300-nv
+  precision: fp4
+  framework: vllm
+  multinode: false
+  scenarios:
+    agentic-coding:
+    - dram-utilization: 0.80
+      search-space:
+      - { tp: 8, spec-decoding: mtp, kv-offloading: none, conc-list: [1] }
+      - { tp: 4, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 2, 16] }
+      - { tp: 2, spec-decoding: mtp, kv-offloading: none, conc-list: [2] }
+
 dsv4-fp4-b200-sglang-agentic-hicache:
   image: lmsysorg/sglang:nightly-dev-cu13-20260707-b4155233
   model: deepseek-ai/DeepSeek-V4-Pro
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 0381848924..ae7443f368 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5061,7 +5061,6 @@
     - "Exclude known-bad nodes mia1-p01-g09,g14 from the disagg node pool"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2301
 
-- config-keys:
     - minimaxm3-fp4-mi355x-vllm-agentic
   description:
     - "Add Minimax-M3 FP4 vLLM Single Node Agentic Support"
@@ -5121,7 +5120,7 @@
   description:
     - "Add GB300 Dynamo-vLLM AgentX MTP3 points: aggregate TP8 c1, aggregate TP4 c4, P/D DEP4/DEP8 c128, and P/D DEP8/DEP8 c384; use upstream vLLM 426e59f, srt-slurm v1.0.36, and random session-affinity routing for P/D."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2341
-  
+
 
 - config-keys:
     - qwen3.5-fp8-mi355x-sglang
@@ -5129,13 +5128,13 @@
   description:
     - "Bump image from lmsysorg/sglang:v0.5.14-rocm720-mi35x to lmsysorg/sglang-rocm:v0.5.16-rocm720-mi35x-20260726"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2349
-  
+
 - config-keys:
     - dsv4-fp4-mi355x-atom-mtp
   description:
     - "Re-sweep dsv4-fp4-mi355x-atom-mtp on latest atom-dev nightly (nightly_202607231538, was atom0.1.3 release). Align ATOM official MTP config (models.json): MTP3 dp-off conc 4-256 + DPA MTP3 (dp-attn) conc 32-1024 (added c32 dp-on; tp4 tried but removed — tp4+MTP+dp-on OOMs, KV budget negative). prefill-only TBO (--enable-tbo, argparse const=prefill so enable_tbo_decode=False) on dp-attn cells at conc>=256 (measured crossover run 30257759947 vs non-TBO 30238071409: TBO -10~14% output tput at c64/c128 but +8~14% at c256+; MTP removes low-conc latency so TBO overlap there is pure overhead); MTP-safe because only decode-TBO (--enable-tbo all) drops spec_decode_metadata. max_num_seqs=conc on dp-on cells and conc>=64 to avoid OOM. bench keeps --dsv4 (InferenceX bench uses encoding_dsv4.py; DSv4-Pro has no jinja chat_template). Fixed benchmark_lib source path (../ -> ../../)."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2345
-  
+
 - config-keys:
     - minimaxm3-fp4-b300-dynamo-vllm-8k1k-tp1
     - minimaxm3-fp4-b300-dynamo-vllm-8k1k-tp1-tep4
@@ -5144,7 +5143,7 @@
     - "Remove VLLM_FLASHINFER_ALLREDUCE_BACKEND=trtllm from the 1P1D TP1-prefill/TP8-decode recipe; retain the existing FP8 indexer KV cache configuration"
     - "Add a one-node 4P1D TP1-prefill/TEP4-decode concurrency-4096 point on the refreshed image while preserving the legacy 4P2D DEP2-prefill/TEP4-decode point unchanged"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2310
-  
+
 
 - config-keys:
     - kimik3-fp4-b300-vllm-agentic
@@ -5173,3 +5172,13 @@
     - "Bring-up validated in run 30326393603: all 12 configs green, zero ServerDisconnectedError after the keep-alive fix. GPU KV resolves to 42.23 GiB / 3,249,215 tokens, i.e. ~3.1 max-length requests, against MAX_NUM_SEQS = 2*CONC."
     - "Measured behaviour: below conc 8 the GPU-resident and DRAM arms are within run-to-run noise (1-5%). At conc 16 and 24 the GPU-resident arm thrashes -- prefix cache hit rate 2.7%, TTFT p50 86s and 191s, 49.6 and 54.9 output tok/s -- because the working set exceeds GPU KV and prefixes are recomputed. The DRAM arm holds TTFT p50 0.85s and 6.2s for 245.0 and 260.6 output tok/s (4-5x), with the CPU tier serving a 62% external prefix cache hit rate at conc 24. The high-conc GPU-resident points are retained deliberately as the honest baseline that makes the offload gain legible."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2386
+
+- config-keys:
+    - minimaxm3-fp4-b300-vllm-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Add five MiniMax-M3 NVFP4 B300 single-node aggregate vLLM AgentX points: TP8 concurrency 1; TP4 concurrency 1, 2, and 16; TP2 concurrency 2."
+    - "Use the Inferact/MiniMax-M3-EAGLE3-GQA draft model with three speculative tokens and synthetic rejection sampling; thinking-on synthetic acceptance length is 2.83 from the canonical MiniMax-M3 EAGLE3 distribution generated by SemiAnalysisAI/InferenceX action 28061204145."
+    - "Use vllm/vllm-openai:nightly-387189c42997b27e2c04b5d97ef8190ffa2bf909 with prefix caching enabled, default KV-cache dtype, FlashInfer TRT-LLM attention with FP8 indexer KV, and the EAGLE3 drafter on FLASH_ATTN."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2328

From 6ad6ee9c91ca344a34a1348f4c9a640de627f2b2 Mon Sep 17 00:00:00 2001
From: hshrivastava-droid <hshrivastava@nvidia.com>
Date: Mon, 27 Jul 2026 20:28:23 -0700
Subject: [PATCH 02/13] update eval logic

---
 benchmarks/single_node/agentic/minimaxm3_fp4_b300_mtp.sh | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)

diff --git a/benchmarks/single_node/agentic/minimaxm3_fp4_b300_mtp.sh b/benchmarks/single_node/agentic/minimaxm3_fp4_b300_mtp.sh
index ff43eebc14..aa3215f5db 100755
--- a/benchmarks/single_node/agentic/minimaxm3_fp4_b300_mtp.sh
+++ b/benchmarks/single_node/agentic/minimaxm3_fp4_b300_mtp.sh
@@ -96,5 +96,9 @@ echo "Server PID: $SERVER_PID"
 set -x
 
 wait_for_server_ready --port "$PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
-build_replay_cmd "$RESULT_DIR"
-run_agentic_replay_and_write_outputs "$RESULT_DIR"
+if [ "${EVAL_ONLY}" = "true" ]; then
+    run_eval --port "$PORT"
+else
+    build_replay_cmd "$RESULT_DIR"
+    run_agentic_replay_and_write_outputs "$RESULT_DIR"
+fi

From 3e84e611f0d95d7ddd8fbcebffa6c02b97884030 Mon Sep 17 00:00:00 2001
From: Xin Li <xinli@nvidia.com>
Date: Tue, 28 Jul 2026 11:48:30 -0400
Subject: [PATCH 03/13] feat: add tool-call-parser + enable-auto-tool-choice;
 drop synthetic AL for eval

Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
---
 .../single_node/agentic/minimaxm3_fp4_b300_mtp.sh  | 14 +++++++++++---
 1 file changed, 11 insertions(+), 3 deletions(-)

diff --git a/benchmarks/single_node/agentic/minimaxm3_fp4_b300_mtp.sh b/benchmarks/single_node/agentic/minimaxm3_fp4_b300_mtp.sh
index aa3215f5db..0f40fdf4e3 100755
--- a/benchmarks/single_node/agentic/minimaxm3_fp4_b300_mtp.sh
+++ b/benchmarks/single_node/agentic/minimaxm3_fp4_b300_mtp.sh
@@ -62,9 +62,15 @@ trap cleanup_agentic_services EXIT
 trap 'exit 130' INT
 trap 'exit 143' TERM
 
-SPEC_CONFIG=$(printf \
-    '{"method":"eagle3","model":"%s","num_speculative_tokens":%d,"attention_backend":"FLASH_ATTN","rejection_sample_method":"synthetic","synthetic_acceptance_length":%.2f}' \
-    "$DRAFT_MODEL_PATH" "$NUM_SPEC_TOKENS" "$SYNTHETIC_ACCEPT_LEN")
+if [ "${EVAL_ONLY:-}" = "true" ]; then
+    SPEC_CONFIG=$(printf \
+        '{"method":"eagle3","model":"%s","num_speculative_tokens":%d,"attention_backend":"FLASH_ATTN"}' \
+        "$DRAFT_MODEL_PATH" "$NUM_SPEC_TOKENS")
+else
+    SPEC_CONFIG=$(printf \
+        '{"method":"eagle3","model":"%s","num_speculative_tokens":%d,"attention_backend":"FLASH_ATTN","rejection_sample_method":"synthetic","synthetic_acceptance_length":%.2f}' \
+        "$DRAFT_MODEL_PATH" "$NUM_SPEC_TOKENS" "$SYNTHETIC_ACCEPT_LEN")
+fi
 
 { set +x; } 2>/dev/null
 VLLM_CMD=(
@@ -79,6 +85,8 @@ VLLM_CMD=(
     --enable-prefix-caching
     --no-enable-flashinfer-autotune
     --reasoning-parser minimax_m3
+    --tool-call-parser minimax_m3
+    --enable-auto-tool-choice
     --default-chat-template-kwargs '{"thinking_mode":"enabled"}'
     --attention-config '{"backend":"FLASHINFER","use_trtllm_attention":true,"indexer_kv_dtype":"fp8"}'
     --all2all-backend flashinfer_nvlink_one_sided

From 3f57bd6646ffcdba377077b5f049ceadaf4e8130 Mon Sep 17 00:00:00 2001
From: Xin Li <xinli@nvidia.com>
Date: Tue, 28 Jul 2026 16:02:51 -0400
Subject: [PATCH 04/13] =?UTF-8?q?perf-changelog:=20add=20trailing=20newlin?=
 =?UTF-8?q?e=20/=20=E6=9C=AB=E5=B0=BE=E6=B7=BB=E5=8A=A0=E7=A9=BA=E8=A1=8C?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
---
 perf-changelog.yaml | 1 +
 1 file changed, 1 insertion(+)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index ae7443f368..ce1d32318c 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5182,3 +5182,4 @@
     - "Use the Inferact/MiniMax-M3-EAGLE3-GQA draft model with three speculative tokens and synthetic rejection sampling; thinking-on synthetic acceptance length is 2.83 from the canonical MiniMax-M3 EAGLE3 distribution generated by SemiAnalysisAI/InferenceX action 28061204145."
     - "Use vllm/vllm-openai:nightly-387189c42997b27e2c04b5d97ef8190ffa2bf909 with prefix caching enabled, default KV-cache dtype, FlashInfer TRT-LLM attention with FP8 indexer KV, and the EAGLE3 drafter on FLASH_ATTN."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2328
+

From 31ec0067b00d8e05aa3f852b144d903782d09f6e Mon Sep 17 00:00:00 2001
From: Xin Li <xinli@nvidia.com>
Date: Tue, 28 Jul 2026 16:05:24 -0400
Subject: [PATCH 05/13] =?UTF-8?q?perf-changelog:=20append=20minimaxm3-fp4-?=
 =?UTF-8?q?b300-vllm-agentic-mtp=20entry=20/=20=E8=BF=BD=E5=8A=A0=20minima?=
 =?UTF-8?q?xm3-fp4-b300-vllm-agentic-mtp=20=E6=9D=A1=E7=9B=AE?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
---
 perf-changelog.yaml | 10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index ce1d32318c..9d69f024e0 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5061,6 +5061,7 @@
     - "Exclude known-bad nodes mia1-p01-g09,g14 from the disagg node pool"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2301
 
+- config-keys:
     - minimaxm3-fp4-mi355x-vllm-agentic
   description:
     - "Add Minimax-M3 FP4 vLLM Single Node Agentic Support"
@@ -5120,7 +5121,7 @@
   description:
     - "Add GB300 Dynamo-vLLM AgentX MTP3 points: aggregate TP8 c1, aggregate TP4 c4, P/D DEP4/DEP8 c128, and P/D DEP8/DEP8 c384; use upstream vLLM 426e59f, srt-slurm v1.0.36, and random session-affinity routing for P/D."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2341
-
+  
 
 - config-keys:
     - qwen3.5-fp8-mi355x-sglang
@@ -5128,13 +5129,13 @@
   description:
     - "Bump image from lmsysorg/sglang:v0.5.14-rocm720-mi35x to lmsysorg/sglang-rocm:v0.5.16-rocm720-mi35x-20260726"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2349
-
+  
 - config-keys:
     - dsv4-fp4-mi355x-atom-mtp
   description:
     - "Re-sweep dsv4-fp4-mi355x-atom-mtp on latest atom-dev nightly (nightly_202607231538, was atom0.1.3 release). Align ATOM official MTP config (models.json): MTP3 dp-off conc 4-256 + DPA MTP3 (dp-attn) conc 32-1024 (added c32 dp-on; tp4 tried but removed — tp4+MTP+dp-on OOMs, KV budget negative). prefill-only TBO (--enable-tbo, argparse const=prefill so enable_tbo_decode=False) on dp-attn cells at conc>=256 (measured crossover run 30257759947 vs non-TBO 30238071409: TBO -10~14% output tput at c64/c128 but +8~14% at c256+; MTP removes low-conc latency so TBO overlap there is pure overhead); MTP-safe because only decode-TBO (--enable-tbo all) drops spec_decode_metadata. max_num_seqs=conc on dp-on cells and conc>=64 to avoid OOM. bench keeps --dsv4 (InferenceX bench uses encoding_dsv4.py; DSv4-Pro has no jinja chat_template). Fixed benchmark_lib source path (../ -> ../../)."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2345
-
+  
 - config-keys:
     - minimaxm3-fp4-b300-dynamo-vllm-8k1k-tp1
     - minimaxm3-fp4-b300-dynamo-vllm-8k1k-tp1-tep4
@@ -5143,7 +5144,7 @@
     - "Remove VLLM_FLASHINFER_ALLREDUCE_BACKEND=trtllm from the 1P1D TP1-prefill/TP8-decode recipe; retain the existing FP8 indexer KV cache configuration"
     - "Add a one-node 4P1D TP1-prefill/TEP4-decode concurrency-4096 point on the refreshed image while preserving the legacy 4P2D DEP2-prefill/TEP4-decode point unchanged"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2310
-
+  
 
 - config-keys:
     - kimik3-fp4-b300-vllm-agentic
@@ -5182,4 +5183,3 @@
     - "Use the Inferact/MiniMax-M3-EAGLE3-GQA draft model with three speculative tokens and synthetic rejection sampling; thinking-on synthetic acceptance length is 2.83 from the canonical MiniMax-M3 EAGLE3 distribution generated by SemiAnalysisAI/InferenceX action 28061204145."
     - "Use vllm/vllm-openai:nightly-387189c42997b27e2c04b5d97ef8190ffa2bf909 with prefix caching enabled, default KV-cache dtype, FlashInfer TRT-LLM attention with FP8 indexer KV, and the EAGLE3 drafter on FLASH_ATTN."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2328
-

From df08354c0dd17baef9fc51bd1546f1f8f418e2f9 Mon Sep 17 00:00:00 2001
From: hshrivastava-droid <hshrivastava@nvidia.com>
Date: Wed, 29 Jul 2026 11:26:34 -0700
Subject: [PATCH 06/13] update config

---
 .../agentic/minimaxm3_fp4_b300_mtp.sh           | 17 ++++++++++++-----
 configs/nvidia-master.yaml                      | 13 ++++++++-----
 perf-changelog.yaml                             |  5 ++---
 3 files changed, 22 insertions(+), 13 deletions(-)

diff --git a/benchmarks/single_node/agentic/minimaxm3_fp4_b300_mtp.sh b/benchmarks/single_node/agentic/minimaxm3_fp4_b300_mtp.sh
index 0f40fdf4e3..18901924d4 100755
--- a/benchmarks/single_node/agentic/minimaxm3_fp4_b300_mtp.sh
+++ b/benchmarks/single_node/agentic/minimaxm3_fp4_b300_mtp.sh
@@ -3,16 +3,17 @@ set -euo pipefail
 set -x
 
 # MiniMax-M3 NVFP4 B300 AgentX with EAGLE3-GQA and synthetic acceptance.
+# DRAM KV offload uses vLLM's SimpleCPUOffloadConnector in lazy mode.
 
 source "$(dirname "$0")/../../benchmark_lib.sh"
 
 export EVAL_FRAMEWORK="lm-eval"
 
-check_env_vars MODEL TP CONC KV_OFFLOADING RESULT_DIR DURATION
+check_env_vars MODEL TP CONC KV_OFFLOADING TOTAL_CPU_DRAM_GB RESULT_DIR DURATION
 
 DRAFT_MODEL="Inferact/MiniMax-M3-EAGLE3-GQA"
 NUM_SPEC_TOKENS=3
-SYNTHETIC_ACCEPT_LEN=2.78
+SYNTHETIC_ACCEPT_LEN=2.83
 
 if [[ -n "${SLURM_JOB_ID:-}" ]]; then
     echo "JOB $SLURM_JOB_ID running on ${SLURMD_NODENAME:-unknown}"
@@ -37,9 +38,14 @@ nvidia-smi
 resolve_trace_source
 install_agentic_deps
 
-if agentic_kv_offload_enabled; then
-    echo "Error: this submission contains GPU-resident KV points only" >&2
-    exit 1
+OFFLOAD_ARGS=()
+if require_agentic_kv_offload_backend vllm-simple; then
+    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}}' \
+        "$CPU_OFFLOAD_BYTES")
+    OFFLOAD_ARGS=(--kv-transfer-config "$OFFLOAD_CONFIG")
 fi
 
 export PYTHONNOUSERSITE=1
@@ -95,6 +101,7 @@ VLLM_CMD=(
     --stream-interval 20
     --trust-remote-code
     --speculative-config "$SPEC_CONFIG"
+    "${OFFLOAD_ARGS[@]}"
 )
 printf '%q ' "${VLLM_CMD[@]}" | tee "$RESULT_DIR/vllm_command.txt"
 printf '\n' | tee -a "$RESULT_DIR/vllm_command.txt"
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 09a42d9901..fca59d082a 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -7810,11 +7810,14 @@ minimaxm3-fp4-b300-vllm-agentic-mtp:
   multinode: false
   scenarios:
     agentic-coding:
-    - dram-utilization: 0.80
-      search-space:
-      - { tp: 8, spec-decoding: mtp, kv-offloading: none, conc-list: [1] }
-      - { tp: 4, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 2, 16] }
-      - { tp: 2, spec-decoding: mtp, kv-offloading: none, conc-list: [2] }
+    # 0.683 maps the B300 TP4 SimpleCPU point to the measured 1,024 GB
+    # engine-level CPU KV budget. GPU-resident points receive a zero budget.
+    - dram-utilization: 0.683
+      search-space:
+      - { tp: 8, spec-decoding: mtp, kv-offloading: none, conc-list: [2] }
+      - { tp: 4, spec-decoding: mtp, kv-offloading: none, conc-list: [2] }
+      - { tp: 2, spec-decoding: mtp, kv-offloading: none, conc-list: [2, 6] }
+      - { tp: 4, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: vllm-simple }, conc-list: [40] }
 
 dsv4-fp4-b200-sglang-agentic-hicache:
   image: lmsysorg/sglang:nightly-dev-cu13-20260707-b4155233
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 9d69f024e0..23329f0cc9 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5179,7 +5179,6 @@
   scenario-type:
     - agentic-coding
   description:
-    - "Add five MiniMax-M3 NVFP4 B300 single-node aggregate vLLM AgentX points: TP8 concurrency 1; TP4 concurrency 1, 2, and 16; TP2 concurrency 2."
-    - "Use the Inferact/MiniMax-M3-EAGLE3-GQA draft model with three speculative tokens and synthetic rejection sampling; thinking-on synthetic acceptance length is 2.83 from the canonical MiniMax-M3 EAGLE3 distribution generated by SemiAnalysisAI/InferenceX action 28061204145."
-    - "Use vllm/vllm-openai:nightly-387189c42997b27e2c04b5d97ef8190ffa2bf909 with prefix caching enabled, default KV-cache dtype, FlashInfer TRT-LLM attention with FP8 indexer KV, and the EAGLE3 drafter on FLASH_ATTN."
+    - "Sample five MiniMax-M3 NVFP4 B300 vLLM AgentX Pareto points on vllm/vllm-openai:nightly-387189c42997b27e2c04b5d97ef8190ffa2bf909: TP8 c2; TP4 c2; TP2 c2 and 6; and TP4 c40 with lazy SimpleCPUOffloadConnector KV offload (1,024 GB engine-level CPU KV budget)."
+    - "Use the Inferact/MiniMax-M3-EAGLE3-GQA drafter with three speculative tokens and synthetic acceptance length 2.83; enable prefix caching, FlashInfer TRT-LLM attention with FP8 indexer KV, and the EAGLE3 drafter on FLASH_ATTN."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2328

From b7556ccd639fe72e4c00603c22a8f34ffc47976c Mon Sep 17 00:00:00 2001
From: Ankur-singh <ankusingh@nvidia.com>
Date: Wed, 29 Jul 2026 16:15:48 -0700
Subject: [PATCH 07/13] =?UTF-8?q?fix(minimaxm3):=20pin=20synthetic=20AL=20?=
 =?UTF-8?q?to=20the=20GQA=20golden=20curve=20(2.83=20->=202.78)=20/=20?=
 =?UTF-8?q?=E4=BF=AE=E5=A4=8D=EF=BC=9A=E5=B0=86=E5=90=88=E6=88=90=E6=8E=A5?=
 =?UTF-8?q?=E5=8F=97=E9=95=BF=E5=BA=A6=E5=AF=B9=E9=BD=90=E5=88=B0=20GQA=20?=
 =?UTF-8?q?=E9=BB=84=E9=87=91=E6=9B=B2=E7=BA=BF=EF=BC=882.83=20=E2=86=92?=
 =?UTF-8?q?=202.78=EF=BC=89?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

This script runs the GQA draft head (Inferact/MiniMax-M3-EAGLE3-GQA) but
injected 2.83, which is minimaxm3_eagle3.yaml thinking_on[3] -- the curve
measured on the NON-GQA head.

#2413 committed golden_al_distribution/minimaxm3_eagle3_gqa.yaml, whose
thinking_on[3] is 2.78 for the head this script actually loads. Since 2.83 is
the higher of the two, the benchmark was injecting a more generous acceptance
than its own golden curve allows, which inflates spec-decode throughput.

Per the AgentX fairness rule a submission may choose any supported draft length
but may not substitute a different acceptance target, so pin 2.78 and note
which curve it comes from -- two curves now exist for this model and the
filenames differ by one suffix.

Note: this changes the benchmarked numbers, so the existing sweep at d0f4fc92
no longer characterises this config; a fresh sweep is required rather than
reusing that run.

中文：本脚本使用 GQA 草稿头，却注入了非 GQA 曲线的 2.83。#2413 已提交
minimaxm3_eagle3_gqa.yaml，其 thinking_on[3] 为 2.78。2.83 更高，会高估投机解码
吞吐，违反 AgentX 不得替换接受率目标的规定。改为 2.78 并注明来源曲线。
该改动会影响基准数值，需要重新跑 sweep，不能复用原有运行。
---
 benchmarks/single_node/agentic/minimaxm3_fp4_b300_mtp.sh | 5 ++++-
 1 file changed, 4 insertions(+), 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 18901924d4..d69f479c94 100755
--- a/benchmarks/single_node/agentic/minimaxm3_fp4_b300_mtp.sh
+++ b/benchmarks/single_node/agentic/minimaxm3_fp4_b300_mtp.sh
@@ -13,7 +13,10 @@ check_env_vars MODEL TP CONC KV_OFFLOADING TOTAL_CPU_DRAM_GB RESULT_DIR DURATION
 
 DRAFT_MODEL="Inferact/MiniMax-M3-EAGLE3-GQA"
 NUM_SPEC_TOKENS=3
-SYNTHETIC_ACCEPT_LEN=2.83
+# Golden AL for the GQA draft head: golden_al_distribution/minimaxm3_eagle3_gqa.yaml
+# minimax-m3.thinking_on[3]. The non-GQA curve (minimaxm3_eagle3.yaml) reads 2.83
+# at the same level -- that head is not what this script runs.
+SYNTHETIC_ACCEPT_LEN=2.78
 
 if [[ -n "${SLURM_JOB_ID:-}" ]]; then
     echo "JOB $SLURM_JOB_ID running on ${SLURMD_NODENAME:-unknown}"

From a52dada798c67263f22c8401f31477f9f4e81cbf Mon Sep 17 00:00:00 2001
From: Ankur-singh <ankusingh@nvidia.com>
Date: Wed, 29 Jul 2026 18:40:17 -0700
Subject: [PATCH 08/13] =?UTF-8?q?fix(changelog):=20correct=20synthetic=20a?=
 =?UTF-8?q?cceptance=20length=202.83=20->=202.78=20/=20=E4=BF=AE=E5=A4=8D?=
 =?UTF-8?q?=EF=BC=9A=E6=9B=B4=E6=AD=A3=E5=90=88=E6=88=90=E6=8E=A5=E5=8F=97?=
 =?UTF-8?q?=E9=95=BF=E5=BA=A6=202.83=20=E2=86=92=202.78?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

The changelog entry still claimed 2.83 while the script was corrected to 2.78
in b7556ccd. 2.78 is the golden value for the GQA drafter this script loads
(golden_al_distribution/minimaxm3_eagle3_gqa.yaml thinking_on[3]); 2.83 is the
non-GQA curve. Prose-only; the executable value was already correct.

中文：changelog 仍写 2.83，而脚本已在 b7556ccd 更正为 2.78。2.78 对应本脚本实际
加载的 GQA 草稿头的黄金值；2.83 为非 GQA 曲线。仅文案更正。
---
 perf-changelog.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index e9649175af..ca7c1305f0 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5280,5 +5280,5 @@
     - agentic-coding
   description:
     - "Sample five MiniMax-M3 NVFP4 B300 vLLM AgentX Pareto points on vllm/vllm-openai:nightly-387189c42997b27e2c04b5d97ef8190ffa2bf909: TP8 c2; TP4 c2; TP2 c2 and 6; and TP4 c40 with lazy SimpleCPUOffloadConnector KV offload (1,024 GB engine-level CPU KV budget)."
-    - "Use the Inferact/MiniMax-M3-EAGLE3-GQA drafter with three speculative tokens and synthetic acceptance length 2.83; enable prefix caching, FlashInfer TRT-LLM attention with FP8 indexer KV, and the EAGLE3 drafter on FLASH_ATTN."
+    - "Use the Inferact/MiniMax-M3-EAGLE3-GQA drafter with three speculative tokens and synthetic acceptance length 2.78; enable prefix caching, FlashInfer TRT-LLM attention with FP8 indexer KV, and the EAGLE3 drafter on FLASH_ATTN."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2328

From 9d0be340f7b1193d1fcbbf5cb238d083c028b91d Mon Sep 17 00:00:00 2001
From: hshrivastava-droid <hshrivastava@nvidia.com>
Date: Thu, 30 Jul 2026 10:27:41 -0700
Subject: [PATCH 09/13] update config

---
 configs/nvidia-master.yaml | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index ff1fb832fa..7c536eec40 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -8456,8 +8456,7 @@ minimaxm3-fp4-b300-vllm-agentic-mtp:
     # engine-level CPU KV budget. GPU-resident points receive a zero budget.
     - dram-utilization: 0.683
       search-space:
-      - { tp: 8, spec-decoding: mtp, kv-offloading: none, conc-list: [2] }
-      - { tp: 4, spec-decoding: mtp, kv-offloading: none, conc-list: [2] }
+      - { tp: 4, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 5] }
       - { tp: 2, spec-decoding: mtp, kv-offloading: none, conc-list: [2, 6] }
       - { tp: 4, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: vllm-simple }, conc-list: [40] }
 

From 01dc27e902f4fbf6bf1dea612c478251406fc43c Mon Sep 17 00:00:00 2001
From: Xin Li <xinli@nvidia.com>
Date: Fri, 31 Jul 2026 17:22:17 -0400
Subject: [PATCH 10/13] chore(agentx): pin MiniMax-M3 sweep to AIPerf ed05782
 (globally anchor profiling handoff)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

中文：将 MiniMax-M3 AgentX 扫描固定到 AIPerf ed05782（全局锚定性能分析切换点）

Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
---
 utils/aiperf | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/utils/aiperf b/utils/aiperf
index deb5421c7f..ed057829b7 160000
--- a/utils/aiperf
+++ b/utils/aiperf
@@ -1 +1 @@
-Subproject commit deb5421c7ff3cd39f78228f469e23e25233ac6c6
+Subproject commit ed057829b78d25d79ce6f3b87763d48fe50363f5

From 2af3fbcd1b2289dafe70d7842d77fe81a8f40a61 Mon Sep 17 00:00:00 2001
From: Xin Li <xinli@nvidia.com>
Date: Sat, 1 Aug 2026 23:06:07 -0400
Subject: [PATCH 11/13] chore(agentx): pin MiniMax-M3 sweep to AIPerf abf55f9
 (keep idle watchdogs active across barriers)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

中文：将 MiniMax-M3 AgentX 扫描固定到 AIPerf abf55f9（跨 barrier 保持空闲看门狗活跃）

Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
---
 utils/aiperf | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/utils/aiperf b/utils/aiperf
index ed057829b7..abf55f902c 160000
--- a/utils/aiperf
+++ b/utils/aiperf
@@ -1 +1 @@
-Subproject commit ed057829b78d25d79ce6f3b87763d48fe50363f5
+Subproject commit abf55f902cde0a3e8389c452a83ecbf5ba15dea9

From 6161fb059d379a560e31e84864d3d52ea86e84a0 Mon Sep 17 00:00:00 2001
From: Xin Li <xinli@nvidia.com>
Date: Sun, 2 Aug 2026 00:52:05 -0400
Subject: [PATCH 12/13] feat(minimaxm3-b300-mtp): expand AgentX search space
 for TP4/TP2/TP4-DRAM arms
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

TP4 GPU-resident: [1,2,5,10,15,20]; TP2 GPU-resident: [1,2,5,10,15,20]; TP4 DRAM offload: [20,30,40,50,60]

中文：扩展 MiniMax-M3 B300 MTP AgentX 扫描空间：TP4 GPU 驻留 [1,2,5,10,15,20]；TP2 GPU 驻留 [1,2,5,10,15,20]；TP4 DRAM 卸载 [20,30,40,50,60]

Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
---
 configs/nvidia-master.yaml | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 05076c2d85..1ba1fed987 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -8600,9 +8600,9 @@ minimaxm3-fp4-b300-vllm-agentic-mtp:
     # engine-level CPU KV budget. GPU-resident points receive a zero budget.
     - dram-utilization: 0.683
       search-space:
-      - { tp: 4, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 5] }
-      - { tp: 2, spec-decoding: mtp, kv-offloading: none, conc-list: [2, 6] }
-      - { tp: 4, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: vllm-simple }, conc-list: [40] }
+      - { 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, 10, 15, 20] }
+      - { tp: 4, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: vllm-simple }, conc-list: [20, 30, 40, 50, 60] }
 
 dsv4-fp4-b200-sglang-agentic-hicache:
   image: lmsysorg/sglang:nightly-dev-cu13-20260707-b4155233

From 9b468247449cc37e2579aaec9fcc8ab286750a28 Mon Sep 17 00:00:00 2001
From: Xin Li <xinli@nvidia.com>
Date: Sun, 2 Aug 2026 10:23:14 -0400
Subject: [PATCH 13/13] =?UTF-8?q?feat(minimaxm3-b300-mtp):=20prune=20searc?=
 =?UTF-8?q?h=20space=20=E2=80=94=20trim=20TP2,=20extend=20DRAM=20offload?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

TP2: [1,2,5] (drop c10,c15,c20); TP4 DRAM: [30,40,50,60,80,90,100] (drop c20, add c80/90/100)

中文：精简扫描空间——TP2 缩减至 [1,2,5]，TP4 DRAM 卸载调整为 [30,40,50,60,80,90,100]

Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
---
 configs/nvidia-master.yaml | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 1ba1fed987..baaf8e4f73 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -8601,8 +8601,8 @@ minimaxm3-fp4-b300-vllm-agentic-mtp:
     - 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, 10, 15, 20] }
-      - { tp: 4, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: vllm-simple }, conc-list: [20, 30, 40, 50, 60] }
+      - { 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, 80, 90, 100] }
 
 dsv4-fp4-b200-sglang-agentic-hicache:
   image: lmsysorg/sglang:nightly-dev-cu13-20260707-b4155233