diff --git a/.github/configs/amd-master.yaml b/.github/configs/amd-master.yaml
index 2528798b1..f18b3f94e 100644
--- a/.github/configs/amd-master.yaml
+++ b/.github/configs/amd-master.yaml
@@ -2871,3 +2871,33 @@ minimaxm3-fp8-mi300x-vllm:
       search-space:
       - { tp: 8, conc-start: 4, conc-end: 64 }
       - { tp: 8, ep: 8, conc-start: 128, conc-end: 256 }
+
+# EAGLE3 speculative-decoding (spec-decoding: mtp) variant of
+# minimaxm3-fp8-mi300x-vllm, pairing MiniMaxAI/MiniMax-M3-MXFP8 with the
+# Inferact/MiniMax-M3-EAGLE3 draft head (3 speculative tokens). Same TP8-only
+# search space as the non-MTP MI300X entry (gfx942 192 GB is memory-tight, like
+# H100), with the TP8 latency rows started at conc 1 to capture single-request
+# latency — matching the H100/MI355X MTP recipes. The shipped ROCm image lacks
+# SupportsEagle3 on the AMD MiniMax-M3 model, so the recipe applies that fix
+# in-place at runtime (functionstackx/vllm#1, upstream vllm-project/vllm#45546;
+# validated green on MI355X) before serving.
+minimaxm3-fp8-mi300x-vllm-mtp:
+  image: vllm/vllm-openai-rocm:minimax-m3
+  model: MiniMaxAI/MiniMax-M3-MXFP8
+  model-prefix: minimaxm3
+  runner: mi300x
+  precision: fp8
+  framework: vllm
+  multinode: false
+  scenarios:
+    fixed-seq-len:
+    - isl: 1024
+      osl: 1024
+      search-space:
+      - { tp: 8, conc-start: 1, conc-end: 128, spec-decoding: mtp }
+      - { tp: 8, ep: 8, conc-start: 256, conc-end: 256, 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: 128, conc-end: 256, spec-decoding: mtp }
diff --git a/.github/configs/nvidia-master.yaml b/.github/configs/nvidia-master.yaml
index 2b2848714..46772c123 100644
--- a/.github/configs/nvidia-master.yaml
+++ b/.github/configs/nvidia-master.yaml
@@ -11773,6 +11773,101 @@ minimaxm2.5-fp8-gb300-dynamo-vllm:
           ep: 4
           dp-attn: true
 
+# MiniMax-M3 GB300 full sweep — safetensors-load-strategy removed from all
+# GB300 recipes (host-memory OOM with prefetch on CW Grace Blackwell nodes).
+# srun_options mem=0 required (DefMemPerCPU=4096 cgroup limit).
+minimaxm3-fp8-gb300-dynamo-vllm:
+  image: vllm/vllm-openai:minimax-m3
+  model: MiniMaxAI/MiniMax-M3-MXFP8
+  model-prefix: minimaxm3
+  runner: gb300-cw
+  precision: fp8
+  framework: dynamo-vllm
+  multinode: true
+  disagg: true
+  scenarios:
+    fixed-seq-len:
+    - isl: 1024
+      osl: 1024
+      search-space:
+      # Disagg-only. Every recipe enables NixlConnector KV transfer over
+      # multi-node NVLink (UCX_CUDA_IPC_ENABLE_MNNVL=y + --enable-cumem-allocator),
+      # which moved the cross-node prefill->decode KV handoff off the RDMA/TCP
+      # fallback (~268 MB/s) onto the NVL fabric (~1.4-1.7 GB/s): +17/+23/+49%
+      # out tok/s/gpu @ conc 64/128/256 (1P1D). GB300-only win — B300's 8-GPU
+      # IB islands cannot move KV over multi-node NVLink. GB300-cw is 4 GPU/node,
+      # so prefill and decode each take whole 4-GPU nodes (no 8-GPU collocation).
+      # 1P+1D TP4+EP4 split, 2 nodes (8 GPU). Best per-GPU efficiency:
+      # peaks at ~1646 out/s/gpu @ conc 256 (1k1k). Covers latency-to-mid.
+      - conc-list: [8, 16, 32, 64, 128, 256, 512]
+        prefill:
+          num-worker: 1
+          tp: 4
+          ep: 4
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb300-fp8/1k1k/disagg-gb300-1p1d-tp4ep4-2n.yaml"
+        decode:
+          num-worker: 1
+          tp: 4
+          ep: 4
+          dp-attn: false
+
+      # Rack-saturating max throughput: 5 prefill + 12 decode, TP4 ep1 (17 nodes).
+      # Balanced prefill:decode ratio (single prefill starved the decode pool)
+      # + TP-only decode (Qwen3.5-A17B pattern; M3 wide-EP all-to-all was slower).
+      - conc-list: [2048, 4096, 8192]
+        prefill:
+          num-worker: 5
+          tp: 4
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb300-fp8/1k1k/disagg-gb300-5p12d-tp4ep1-17n.yaml"
+        decode:
+          num-worker: 12
+          tp: 4
+          ep: 1
+          dp-attn: false
+
+    - isl: 8192
+      osl: 1024
+      search-space:
+      # 8k1k long context: prefill is heavier and KV is larger, so concurrency
+      # is lower than 1k1k. Same disagg shapes + multi-node-NVLink KV transfer.
+      # 1P+1D TP4+EP4 split, 2 nodes (8 GPU). Best per-GPU efficiency:
+      # peaks at ~1209 out/s/gpu @ conc 256 (8k1k). Covers latency-to-mid.
+      - conc-list: [8, 16, 32, 64, 128, 256]
+        prefill:
+          num-worker: 1
+          tp: 4
+          ep: 4
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb300-fp8/8k1k/disagg-gb300-1p1d-tp4ep4-2n.yaml"
+        decode:
+          num-worker: 1
+          tp: 4
+          ep: 4
+          dp-attn: false
+
+      # Prefill-heavy rack-saturating: 10P + 7D, TP4 ep1 (17 nodes).
+      # At 8k context, prefill is 8x heavier — 5P:12D showed 616-req prefill
+      # backlog. DSR1/DSV4 GB300 patterns use 6-10 prefill workers for 8k1k.
+      - conc-list: [1024, 2048, 4096]
+        prefill:
+          num-worker: 10
+          tp: 4
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb300-fp8/8k1k/disagg-gb300-10p7d-tp4ep1-17n.yaml"
+        decode:
+          num-worker: 7
+          tp: 4
+          ep: 1
+          dp-attn: false
+
 # MiniMax-M3 day-zero (https://recipes.vllm.ai/MiniMaxAI/MiniMax-M3).
 # 427B total / 26B active MoE with MSA sparse attention; MXFP8 checkpoint
 # (MiniMaxAI/MiniMax-M3-MXFP8, ~444 GB) quantized by NVIDIA — native MX tensor
diff --git a/.github/workflows/benchmark-multinode-tmpl.yml b/.github/workflows/benchmark-multinode-tmpl.yml
index 81727ef39..85b399e6c 100644
--- a/.github/workflows/benchmark-multinode-tmpl.yml
+++ b/.github/workflows/benchmark-multinode-tmpl.yml
@@ -123,6 +123,11 @@ on:
 
 env:
   RANDOM_RANGE_RATIO: 0.8
+  # Day-zero models resolved via hf: ids download from the Hub inside the
+  # slurm job (srtctl pre-download + dynamo hub fetch). Anonymous requests
+  # get 429-rate-limited when several workers pull a 444 GB snapshot at
+  # once; sbatch/srun inherit this env so the token reaches the workers.
+  HF_TOKEN: ${{ secrets.INFERENCEX_OFFICIAL_RO_HF_TOKEN }}
   EXP_NAME: ${{ inputs.exp-name }}
   IMAGE: ${{ inputs.image }}
   MODEL_PREFIX: ${{ inputs.model-prefix }}
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/disagg-gb300-1p1d-tp4ep4-2n.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/disagg-gb300-1p1d-tp4ep4-2n.yaml
new file mode 100644
index 000000000..8e13b522f
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/disagg-gb300-1p1d-tp4ep4-2n.yaml
@@ -0,0 +1,126 @@
+name: "minimax-m3-vllm-disagg-gb300-1p1d-tp4ep4-2n-1k1k"
+
+# MiniMax-M3 disaggregated 1P+1D recipe for GB300, 2 nodes.
+# 1 node prefill (4 GPUs, TP4+EP4) → NixlConnector → 1 node decode (4 GPUs, TP4+EP4).
+# --block-size 128 is mandatory (MSA sparse/index cache alignment).
+# safetensors-load-strategy omitted — prefetch OOMs on CW GB300 host memory.
+
+model:
+  path: "hf:MiniMaxAI/MiniMax-M3-MXFP8"
+  container: "vllm/vllm-openai:minimax-m3"
+  precision: "fp8"
+
+dynamo:
+  install: true
+  wheel: "1.2.0.dev20260526"
+
+# CW gb300 has no per-GPU Slurm defaults: without sbatch-level mem=0 the
+# job allocation is ntasks x DefMemPerCPU = 4 GB/node and worker cgroups
+# OOM-kill during engine init; srun_options.mem=0 alone only grants a
+# step what the job already holds. cpus-per-task=72 (one NUMA socket)
+# keeps the gpu-less infra step (etcd/nats) off the 1-CPU default.
+# Full rationale: vllm/deepseek-v4/agentic/disagg-gb300-1p6d-dep4-tp4-agentic.yaml.
+sbatch_directives:
+  mem: "0"
+  cpus-per-task: "72"
+srun_options:
+  mem: "0"
+
+slurm:
+  time_limit: "8:00:00"
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+
+extra_mount:
+  - "__M3_HF_HOME__:__M3_HF_HOME__"
+
+resources:
+  gpu_type: "gb300"
+  gpus_per_node: 4
+  prefill_nodes: 1
+  decode_nodes: 1
+  prefill_workers: 1
+  decode_workers: 1
+  gpus_per_prefill: 4
+  gpus_per_decode: 4
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+
+backend:
+  type: vllm
+  connector: null
+
+  prefill_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_FLASHINFER_ALLREDUCE_BACKEND: "mnnvl"
+    HF_HOME: "__M3_HF_HOME__"
+    # Cache-only at runtime: the launcher pre-stages the full snapshot and
+    # verifies it offline before submit, so workers must NOT re-fetch. Without
+    # this, dynamo fetch_model takes a per-blob .lock on the shared NFS cache;
+    # co-fetching workers (e.g. both nodes of a TP8-2n job) collide and the
+    # loser dies with "Lock acquisition failed" (no retry like Python hf_hub).
+    HF_HUB_OFFLINE: "1"
+    # Multi-node-NVLink KV transfer: keep cross-node prefill->decode KV on the
+    # NVL fabric (5-6x vs RDMA/TCP fallback, +17-49% tok/s/gpu). Needs VMM-
+    # registered KV (enable-cumem-allocator below).
+    UCX_CUDA_IPC_ENABLE_MNNVL: "y"
+
+  decode_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_FLASHINFER_ALLREDUCE_BACKEND: "mnnvl"
+    HF_HOME: "__M3_HF_HOME__"
+    # Cache-only at runtime: the launcher pre-stages the full snapshot and
+    # verifies it offline before submit, so workers must NOT re-fetch. Without
+    # this, dynamo fetch_model takes a per-blob .lock on the shared NFS cache;
+    # co-fetching workers (e.g. both nodes of a TP8-2n job) collide and the
+    # loser dies with "Lock acquisition failed" (no retry like Python hf_hub).
+    HF_HUB_OFFLINE: "1"
+    # Multi-node-NVLink KV transfer: keep cross-node prefill->decode KV on the
+    # NVL fabric (5-6x vs RDMA/TCP fallback, +17-49% tok/s/gpu). Needs VMM-
+    # registered KV (enable-cumem-allocator below).
+    UCX_CUDA_IPC_ENABLE_MNNVL: "y"
+
+  vllm_config:
+    prefill:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      tensor-parallel-size: 4
+      pipeline-parallel-size: 1
+      enable-expert-parallel: true
+      enable-cumem-allocator: true
+      enforce-eager: true
+      max-model-len: 2304
+      max-num-seqs: 16
+      max-num-batched-tokens: 16384
+      block-size: 128
+      language-model-only: true
+      gpu-memory-utilization: 0.9
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      stream-interval: 32
+
+    decode:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      tensor-parallel-size: 4
+      pipeline-parallel-size: 1
+      enable-expert-parallel: true
+      enable-cumem-allocator: true
+      max-model-len: 2304
+      max-num-seqs: 256
+      max-num-batched-tokens: 256
+      max-cudagraph-capture-size: 512
+      block-size: 128
+      language-model-only: true
+      gpu-memory-utilization: 0.9
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      stream-interval: 32
+
+benchmark:
+  type: "sa-bench"
+  isl: 1024
+  osl: 1024
+  concurrencies: "64x128x256x512"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/disagg-gb300-5p12d-tp4ep1-17n.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/disagg-gb300-5p12d-tp4ep1-17n.yaml
new file mode 100644
index 000000000..eb39f71f9
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/disagg-gb300-5p12d-tp4ep1-17n.yaml
@@ -0,0 +1,102 @@
+name: "minimax-m3-vllm-disagg-gb300-5p12d-tp4ep1-1k1k"
+
+# MiniMax-M3 GB300 max-throughput disagg, rack-saturating (17 nodes / 68 GPU).
+# Lessons baked in (validated on gb300-cw): (1) TP-only decode (ep1, NO expert
+# parallelism) per Qwen3.5-397B-A17B recipes — M3's MoE all-to-all overhead
+# made wide-EP/DP-attention slower; (2) PREFILL is the bottleneck, not decode
+# or KV transfer — a single prefill starved 64 decode GPUs (3967 reqs queued),
+# so this balances 5 prefill : 12 decode TP4 workers (prefill backlog -> 0,
+# +57% tok/s/gpu vs the 1-prefill ratio at conc 2048). NixlConnector KV
+# transfer stays on multi-node NVLink (MNNVL + cumem). --block-size 128
+# mandatory (MSA cache); text-only -> language-model-only.
+
+model:
+  path: "hf:MiniMaxAI/MiniMax-M3-MXFP8"
+  container: "vllm/vllm-openai:minimax-m3"
+  precision: "fp8"
+dynamo:
+  install: true
+  wheel: "1.2.0.dev20260526"
+sbatch_directives:
+  mem: "0"
+  cpus-per-task: "72"
+srun_options:
+  mem: "0"
+slurm:
+  time_limit: "8:00:00"
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+extra_mount:
+  - "__M3_HF_HOME__:__M3_HF_HOME__"
+
+resources:
+  gpu_type: "gb300"
+  gpus_per_node: 4
+  prefill_nodes: 5
+  decode_nodes: 12
+  prefill_workers: 5
+  decode_workers: 12
+  gpus_per_prefill: 4
+  gpus_per_decode: 4
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+
+backend:
+  type: vllm
+  connector: null
+
+  prefill_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_FLASHINFER_ALLREDUCE_BACKEND: "mnnvl"
+    HF_HOME: "__M3_HF_HOME__"
+    HF_HUB_OFFLINE: "1"
+    UCX_CUDA_IPC_ENABLE_MNNVL: "y"
+
+  decode_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_FLASHINFER_ALLREDUCE_BACKEND: "mnnvl"
+    HF_HOME: "__M3_HF_HOME__"
+    HF_HUB_OFFLINE: "1"
+    UCX_CUDA_IPC_ENABLE_MNNVL: "y"
+
+  vllm_config:
+    prefill:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      tensor-parallel-size: 4
+      pipeline-parallel-size: 1
+      enable-cumem-allocator: true
+      enforce-eager: true
+      max-num-seqs: 16
+      max-num-batched-tokens: 16384
+      max-model-len: 2304
+      block-size: 128
+      language-model-only: true
+      gpu-memory-utilization: 0.9
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      stream-interval: 32
+
+    decode:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      tensor-parallel-size: 4
+      pipeline-parallel-size: 1
+      enable-cumem-allocator: true
+      max-num-seqs: 256
+      max-num-batched-tokens: 256
+      max-cudagraph-capture-size: 512
+      max-model-len: 2304
+      block-size: 128
+      language-model-only: true
+      gpu-memory-utilization: 0.9
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      stream-interval: 32
+
+benchmark:
+  type: "sa-bench"
+  isl: 1024
+  osl: 1024
+  concurrencies: "2048x4096x8192"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/disagg-gb300-10p7d-tp4ep1-17n.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/disagg-gb300-10p7d-tp4ep1-17n.yaml
new file mode 100644
index 000000000..f005aa6dc
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/disagg-gb300-10p7d-tp4ep1-17n.yaml
@@ -0,0 +1,99 @@
+name: "minimax-m3-vllm-disagg-gb300-10p7d-tp4ep1-8k1k"
+
+# MiniMax-M3 GB300 prefill-heavy max-throughput disagg, 17 nodes / 68 GPU.
+# At 8k context, prefill is 8x heavier per request: the 5P:12D ratio (tuned
+# for light 1k1k prefill) is prefill-bound — 616 reqs queued behind 5 workers.
+# This config rebalances to 10P:7D per DSR1/DSV4 GB300 patterns (DSR1 8k1k
+# uses 6P:1D; DSV4 uses 10P:8D at 18 nodes). TP4 ep1 decode (no wide EP —
+# M3's MoE all-to-all overhead makes it slower). MNNVL KV + cumem.
+
+model:
+  path: "hf:MiniMaxAI/MiniMax-M3-MXFP8"
+  container: "vllm/vllm-openai:minimax-m3"
+  precision: "fp8"
+dynamo:
+  install: true
+  wheel: "1.2.0.dev20260526"
+sbatch_directives:
+  mem: "0"
+  cpus-per-task: "72"
+srun_options:
+  mem: "0"
+slurm:
+  time_limit: "8:00:00"
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+extra_mount:
+  - "__M3_HF_HOME__:__M3_HF_HOME__"
+
+resources:
+  gpu_type: "gb300"
+  gpus_per_node: 4
+  prefill_nodes: 10
+  decode_nodes: 7
+  prefill_workers: 10
+  decode_workers: 7
+  gpus_per_prefill: 4
+  gpus_per_decode: 4
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+
+backend:
+  type: vllm
+  connector: null
+
+  prefill_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_FLASHINFER_ALLREDUCE_BACKEND: "mnnvl"
+    HF_HOME: "__M3_HF_HOME__"
+    HF_HUB_OFFLINE: "1"
+    UCX_CUDA_IPC_ENABLE_MNNVL: "y"
+
+  decode_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_FLASHINFER_ALLREDUCE_BACKEND: "mnnvl"
+    HF_HOME: "__M3_HF_HOME__"
+    HF_HUB_OFFLINE: "1"
+    UCX_CUDA_IPC_ENABLE_MNNVL: "y"
+
+  vllm_config:
+    prefill:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      tensor-parallel-size: 4
+      pipeline-parallel-size: 1
+      enable-cumem-allocator: true
+      enforce-eager: true
+      max-num-seqs: 16
+      max-num-batched-tokens: 16384
+      max-model-len: 9472
+      block-size: 128
+      language-model-only: true
+      gpu-memory-utilization: 0.9
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      stream-interval: 32
+
+    decode:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      tensor-parallel-size: 4
+      pipeline-parallel-size: 1
+      enable-cumem-allocator: true
+      max-num-seqs: 256
+      max-num-batched-tokens: 256
+      max-cudagraph-capture-size: 512
+      max-model-len: 9472
+      block-size: 128
+      language-model-only: true
+      gpu-memory-utilization: 0.9
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      stream-interval: 32
+
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  concurrencies: "1024x2048x4096"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/disagg-gb300-1p1d-tp4ep4-2n.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/disagg-gb300-1p1d-tp4ep4-2n.yaml
new file mode 100644
index 000000000..1e0d67e3c
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/disagg-gb300-1p1d-tp4ep4-2n.yaml
@@ -0,0 +1,126 @@
+name: "minimax-m3-vllm-disagg-gb300-1p1d-tp4ep4-2n-8k1k"
+
+# MiniMax-M3 disaggregated 1P+1D recipe for GB300, 2 nodes.
+# 1 node prefill (4 GPUs, TP4+EP4) → NixlConnector → 1 node decode (4 GPUs, TP4+EP4).
+# --block-size 128 is mandatory (MSA sparse/index cache alignment).
+# safetensors-load-strategy omitted — prefetch OOMs on CW GB300 host memory.
+
+model:
+  path: "hf:MiniMaxAI/MiniMax-M3-MXFP8"
+  container: "vllm/vllm-openai:minimax-m3"
+  precision: "fp8"
+
+dynamo:
+  install: true
+  wheel: "1.2.0.dev20260526"
+
+# CW gb300 has no per-GPU Slurm defaults: without sbatch-level mem=0 the
+# job allocation is ntasks x DefMemPerCPU = 4 GB/node and worker cgroups
+# OOM-kill during engine init; srun_options.mem=0 alone only grants a
+# step what the job already holds. cpus-per-task=72 (one NUMA socket)
+# keeps the gpu-less infra step (etcd/nats) off the 1-CPU default.
+# Full rationale: vllm/deepseek-v4/agentic/disagg-gb300-1p6d-dep4-tp4-agentic.yaml.
+sbatch_directives:
+  mem: "0"
+  cpus-per-task: "72"
+srun_options:
+  mem: "0"
+
+slurm:
+  time_limit: "8:00:00"
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+
+extra_mount:
+  - "__M3_HF_HOME__:__M3_HF_HOME__"
+
+resources:
+  gpu_type: "gb300"
+  gpus_per_node: 4
+  prefill_nodes: 1
+  decode_nodes: 1
+  prefill_workers: 1
+  decode_workers: 1
+  gpus_per_prefill: 4
+  gpus_per_decode: 4
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+
+backend:
+  type: vllm
+  connector: null
+
+  prefill_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_FLASHINFER_ALLREDUCE_BACKEND: "mnnvl"
+    HF_HOME: "__M3_HF_HOME__"
+    # Cache-only at runtime: the launcher pre-stages the full snapshot and
+    # verifies it offline before submit, so workers must NOT re-fetch. Without
+    # this, dynamo fetch_model takes a per-blob .lock on the shared NFS cache;
+    # co-fetching workers (e.g. both nodes of a TP8-2n job) collide and the
+    # loser dies with "Lock acquisition failed" (no retry like Python hf_hub).
+    HF_HUB_OFFLINE: "1"
+    # Multi-node-NVLink KV transfer: keep cross-node prefill->decode KV on the
+    # NVL fabric (5-6x vs RDMA/TCP fallback, +17-49% tok/s/gpu). Needs VMM-
+    # registered KV (enable-cumem-allocator below).
+    UCX_CUDA_IPC_ENABLE_MNNVL: "y"
+
+  decode_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_FLASHINFER_ALLREDUCE_BACKEND: "mnnvl"
+    HF_HOME: "__M3_HF_HOME__"
+    # Cache-only at runtime: the launcher pre-stages the full snapshot and
+    # verifies it offline before submit, so workers must NOT re-fetch. Without
+    # this, dynamo fetch_model takes a per-blob .lock on the shared NFS cache;
+    # co-fetching workers (e.g. both nodes of a TP8-2n job) collide and the
+    # loser dies with "Lock acquisition failed" (no retry like Python hf_hub).
+    HF_HUB_OFFLINE: "1"
+    # Multi-node-NVLink KV transfer: keep cross-node prefill->decode KV on the
+    # NVL fabric (5-6x vs RDMA/TCP fallback, +17-49% tok/s/gpu). Needs VMM-
+    # registered KV (enable-cumem-allocator below).
+    UCX_CUDA_IPC_ENABLE_MNNVL: "y"
+
+  vllm_config:
+    prefill:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      tensor-parallel-size: 4
+      pipeline-parallel-size: 1
+      enable-expert-parallel: true
+      enable-cumem-allocator: true
+      enforce-eager: true
+      max-model-len: 9472
+      max-num-seqs: 16
+      max-num-batched-tokens: 16384
+      block-size: 128
+      language-model-only: true
+      gpu-memory-utilization: 0.9
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      stream-interval: 32
+
+    decode:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      tensor-parallel-size: 4
+      pipeline-parallel-size: 1
+      enable-expert-parallel: true
+      enable-cumem-allocator: true
+      max-model-len: 9472
+      max-num-seqs: 256
+      max-num-batched-tokens: 256
+      max-cudagraph-capture-size: 512
+      block-size: 128
+      language-model-only: true
+      gpu-memory-utilization: 0.9
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      stream-interval: 32
+
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  concurrencies: "16x32x64x128x256"
diff --git a/benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_mi300x.sh b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_mi300x.sh
index e3522e00a..f2cdaf284 100755
--- a/benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_mi300x.sh
+++ b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_mi300x.sh
@@ -32,6 +32,7 @@ fi
 
 SERVER_LOG=/workspace/server.log
 export VLLM_ENGINE_READY_TIMEOUT_S=3600
+export VLLM_USE_BREAKABLE_CUDAGRAPH=0
 
 if [ "${EVAL_ONLY}" = "true" ]; then
     setup_eval_context
@@ -58,7 +59,6 @@ vllm serve "$MODEL" --port "$PORT" \
     --language-model-only \
     --max-model-len "$MAX_MODEL_LEN" \
     --attention-backend TRITON_ATTN \
-    --enforce-eager \
     --tool-call-parser minimax_m3 \
     --reasoning-parser minimax_m3 \
     --enable-auto-tool-choice > "$SERVER_LOG" 2>&1 &
diff --git a/benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_mi300x_mtp.sh b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_mi300x_mtp.sh
new file mode 100644
index 000000000..40fbab536
--- /dev/null
+++ b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_mi300x_mtp.sh
@@ -0,0 +1,214 @@
+#!/usr/bin/env bash
+
+# MiniMax-M3 MXFP8 MI300X (gfx942) single-node vLLM recipe with EAGLE3
+# speculative decoding — the spec-decoding=mtp variant of
+# minimaxm3_fp8_mi300x.sh. Adds the Inferact/MiniMax-M3-EAGLE3 draft head via
+# --speculative-config with 3 speculative tokens. Everything else mirrors the
+# non-MTP MI300X recipe: mandatory --block-size 128, --language-model-only for
+# the text-only benchmark, --attention-backend TRITON_ATTN, and
+# --no-enable-prefix-caching. Runs with CUDA graphs (no --enforce-eager);
+# VLLM_USE_BREAKABLE_CUDAGRAPH=0 avoids the M3-decode breakable-cudagraph path.
+# The default BF16 KV cache is retained (unlike
+# the MI355X recipe's FP8 KV cache): gfx942 has no calibrated q/prob scales for
+# ROCm FP8 attention and vLLM's fallback scale of 1.0 corrupts accuracy.
+#
+# Unlike the CUDA recipes, the drafter needs no attention_backend override:
+# the FlashInfer "page size 128 requires GQA/MQA" limitation that forced
+# FLASH_ATTN for the EAGLE3 MHA head on Blackwell is FlashInfer/CUDA-specific.
+# Here the whole server runs on TRITON_ATTN (set globally below), which serves
+# the MHA draft fine.
+#
+# [AI generated draft test] The shipped vllm/vllm-openai-rocm:minimax-m3 image
+# does NOT implement SupportsEagle3 on the AMD MiniMax-M3 model, so EAGLE3
+# engine init fails with "Model does not support EAGLE3 interface but
+# aux_hidden_state_outputs was requested". This recipe applies that fix
+# (functionstackx/vllm#1 — ported from nvidia/model.py, upstreamed as
+# vllm-project/vllm#45546) in-place to the installed vllm before serving, so we
+# can validate EAGLE3 on real MI300X hardware ahead of an image rebuild. The
+# same patch is validated green on MI355X. It is idempotent and fails the job
+# loudly if the installed amd/model.py has drifted from the expected base.
+
+source "$(dirname "$0")/../../benchmark_lib.sh"
+
+check_env_vars \
+    MODEL \
+    TP \
+    EP_SIZE \
+    DP_ATTENTION \
+    CONC \
+    ISL \
+    OSL \
+    MAX_MODEL_LEN \
+    RANDOM_RANGE_RATIO \
+    RESULT_FILENAME
+
+DRAFT_MODEL="Inferact/MiniMax-M3-EAGLE3"
+
+if [[ -n "$SLURM_JOB_ID" ]]; then
+  echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
+fi
+
+# MODEL is a bare HF id on the mi300x single-node runner (a fast cache hit when
+# pre-staged). The EAGLE3 draft is not staged; fetch it into the same cache.
+if [[ "$MODEL" != /* ]]; then
+  hf download "$MODEL"
+  hf download "$DRAFT_MODEL"
+fi
+
+if [ -n "$ROCR_VISIBLE_DEVICES" ]; then
+    export HIP_VISIBLE_DEVICES="$ROCR_VISIBLE_DEVICES"
+fi
+
+SERVER_LOG=/workspace/server.log
+export VLLM_ENGINE_READY_TIMEOUT_S=3600
+export VLLM_USE_BREAKABLE_CUDAGRAPH=0
+
+if [ "${EVAL_ONLY}" = "true" ]; then
+    setup_eval_context
+fi
+
+PARALLEL_ARGS=(--tensor-parallel-size "$TP")
+if [ "${DP_ATTENTION}" = "true" ]; then
+    PARALLEL_ARGS=(
+        --tensor-parallel-size 1
+        --data-parallel-size "$TP"
+        --enable-expert-parallel
+    )
+elif [ "$EP_SIZE" -gt 1 ]; then
+    PARALLEL_ARGS+=(--enable-expert-parallel)
+fi
+
+# use 3 speculative tokens for all configs for now
+NUM_SPEC_TOKENS=3
+
+# [AI generated draft test] Patch the installed AMD MiniMax-M3 model to add the
+# SupportsEagle3 interface (functionstackx/vllm#1, upstream vllm-project/vllm#45546).
+# Mirrors nvidia/model.py: adds EagleModelMixin to the inner model +
+# aux-hidden-state emission, and SupportsEagle3 to the two outer classes.
+# Idempotent; hard-fails if the installed file has drifted from the expected
+# base (so we never silently run unpatched and mislabel the result).
+python3 - <<'PYEOF' || { echo "EAGLE3 in-place patch failed" >&2; exit 1; }
+import ast, importlib.util, pathlib, sys
+
+spec = importlib.util.find_spec("vllm")
+root = pathlib.Path(spec.submodule_search_locations[0])
+target = root / "models" / "minimax_m3" / "amd" / "model.py"
+src = target.read_text()
+
+if "EagleModelMixin" in src and "class MiniMaxM3Model(nn.Module, EagleModelMixin):" in src:
+    print(f"[eagle3-patch] already applied: {target}")
+    sys.exit(0)
+
+edits = [
+    (
+        "from vllm.model_executor.models.interfaces import (\n"
+        "    MultiModalEmbeddings,\n"
+        "    SupportsMultiModal,\n"
+        ")",
+        "from vllm.model_executor.models.interfaces import (\n"
+        "    EagleModelMixin,\n"
+        "    MultiModalEmbeddings,\n"
+        "    SupportsEagle3,\n"
+        "    SupportsMultiModal,\n"
+        ")",
+    ),
+    (
+        "class MiniMaxM3Model(nn.Module):",
+        "class MiniMaxM3Model(nn.Module, EagleModelMixin):",
+    ),
+    (
+        "        inputs_embeds: torch.Tensor | None = None,\n"
+        "    ) -> torch.Tensor:\n"
+        "        if inputs_embeds is not None:",
+        "        inputs_embeds: torch.Tensor | None = None,\n"
+        "    ) -> torch.Tensor | tuple[torch.Tensor, list[torch.Tensor]]:\n"
+        "        if inputs_embeds is not None:",
+    ),
+    (
+        "        residual = None\n\n"
+        "        for layer in self.layers[self.start_layer : self.end_layer]:\n"
+        "            hidden_states, residual = layer(positions, hidden_states, residual)\n\n"
+        "        hidden_states, _ = self.norm(hidden_states, residual)\n"
+        "        return hidden_states",
+        "        residual = None\n\n"
+        "        # EAGLE3 is not yet compatible with pipeline parallel\n"
+        "        aux_hidden_states = self._maybe_add_hidden_state([], 0, hidden_states, residual)\n"
+        "        for idx, layer in enumerate(self.layers[self.start_layer : self.end_layer]):\n"
+        "            hidden_states, residual = layer(positions, hidden_states, residual)\n"
+        "            self._maybe_add_hidden_state(\n"
+        "                aux_hidden_states, idx + 1, hidden_states, residual\n"
+        "            )\n\n"
+        "        hidden_states, _ = self.norm(hidden_states, residual)\n\n"
+        "        if len(aux_hidden_states) > 0:\n"
+        "            return hidden_states, aux_hidden_states\n"
+        "        return hidden_states",
+    ),
+    (
+        "class MiniMaxM3SparseForCausalLM(nn.Module):",
+        "class MiniMaxM3SparseForCausalLM(nn.Module, SupportsEagle3):",
+    ),
+    (
+        "class MiniMaxM3SparseForConditionalGeneration(nn.Module, SupportsMultiModal):",
+        "class MiniMaxM3SparseForConditionalGeneration(\n"
+        "    nn.Module, SupportsMultiModal, SupportsEagle3\n"
+        "):",
+    ),
+]
+
+for old, new in edits:
+    count = src.count(old)
+    if count != 1:
+        sys.exit(
+            f"[eagle3-patch] anchor matched {count} times (expected 1); "
+            f"installed {target} has drifted from the expected base — aborting"
+        )
+    src = src.replace(old, new)
+
+ast.parse(src)
+target.write_text(src)
+print(f"[eagle3-patch] applied EAGLE3 support to {target}")
+PYEOF
+
+start_gpu_monitor
+
+set -x
+vllm serve "$MODEL" --port "$PORT" \
+    "${PARALLEL_ARGS[@]}" \
+    --block-size 128 \
+    --no-enable-prefix-caching \
+    --language-model-only \
+    --max-model-len "$MAX_MODEL_LEN" \
+    --attention-backend TRITON_ATTN \
+    --speculative-config "{\"method\": \"eagle3\", \"model\": \"$DRAFT_MODEL\", \"num_speculative_tokens\": $NUM_SPEC_TOKENS}" \
+    --tool-call-parser minimax_m3 \
+    --reasoning-parser minimax_m3 \
+    --enable-auto-tool-choice > "$SERVER_LOG" 2>&1 &
+
+SERVER_PID=$!
+wait_for_server_ready --port "$PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
+
+pip install -q datasets pandas
+
+# Spec-decode acceptance rate degrades on raw random tokens; route prompts
+# through the chat template as the other MTP recipes do.
+run_benchmark_serving \
+    --model "$MODEL" \
+    --port "$PORT" \
+    --backend vllm \
+    --input-len "$ISL" \
+    --output-len "$OSL" \
+    --random-range-ratio "$RANDOM_RANGE_RATIO" \
+    --num-prompts "$((CONC * 10))" \
+    --max-concurrency "$CONC" \
+    --result-filename "$RESULT_FILENAME" \
+    --result-dir /workspace/ \
+    --trust-remote-code \
+    --use-chat-template
+
+if [ "${RUN_EVAL}" = "true" ]; then
+    run_eval --framework lm-eval --port "$PORT"
+    append_lm_eval_summary
+fi
+
+stop_gpu_monitor
+set +x
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 6b16a41c5..0b9067114 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -3649,6 +3649,19 @@
     - "Serves from the launch_b300-nv.sh MODEL/MODEL_PATH split (model not in the SRE-staged /scratch/models list -> writable /data/models)"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1724
 
+- config-keys:
+    - minimaxm3-fp8-gb300-dynamo-vllm
+  description:
+    - "Initial submission: MiniMax-M3 MXFP8 day-zero vLLM sweep for GB300 via Dynamo"
+    - "Model: MiniMaxAI/MiniMax-M3-MXFP8 (427B total / 26B active MoE, MSA sparse attention, ~444 GB MXFP8 checkpoint)"
+    - "Image: vllm/vllm-openai:minimax-m3 (multi-arch arm64+amd64 from m3_release branch, vllm-project/vllm#45381)"
+    - "Dynamo orchestration with NixlConnector for disaggregated prefill/decode"
+    - "7 topologies: TP4 (1n), TP8 (2n), TP4+EP4 (1n), 1P+1D disagg TP4+EP4 (2n), 1P+1D disagg collocated (1n), DEP4 (1n), DEP8 (2n)"
+    - "Concurrency sweep: TP 4-64, TEP 128-512, disagg 64-512, DEP4 256-1024, DEP8 512-2048"
+    - "GB300 recipes require srun_options mem=0 (CW DefMemPerCPU=4096 causes host-RAM OOM) and omit safetensors-load-strategy prefetch (~600 GB host-memory limit)"
+    - "--block-size 128 mandatory everywhere (MSA sparse/index cache alignment); --language-model-only for text-only benchmarks"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1735
+
 - config-keys:
     - minimaxm3-fp8-b200-vllm
   description:
@@ -3745,3 +3758,26 @@
     - "H100-aligned layouts and concurrency ranges: TP8 and TP8+EP8 across 1k1k and 8k1k"
     - "Fix launch_mi300x-amds.sh node exclusion to use the current short Slurm node name"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1746
+
+- config-keys:
+    - minimaxm3-fp8-mi300x-vllm-mtp
+  description:
+    - "Initial submission: MiniMax-M3 MXFP8 MI300X (gfx942) vLLM benchmark with EAGLE3 speculative decoding (target: MiniMaxAI/MiniMax-M3-MXFP8, draft: Inferact/MiniMax-M3-EAGLE3, 3 speculative tokens) — spec-decoding=mtp variant of the MI300X day-zero recipe"
+    - "Image: vllm/vllm-openai-rocm:minimax-m3 (same day-zero ROCm build as the non-MTP entry)"
+    - "Serve shape follows minimaxm3-fp8-mi300x-vllm: --block-size 128, --no-enable-prefix-caching, --language-model-only, --attention-backend TRITON_ATTN, --enforce-eager, minimax_m3 parsers, and the default BF16 KV cache (gfx942 lacks calibrated ROCm FP8 attention scales); prompts routed through the chat template for realistic acceptance"
+    - "TP8-only search space (gfx942 192 GB is memory-tight, like H100): TP8 latency rows started at conc 1, TP8+EP8 (TEP) at high concurrency, across 1k1k and 8k1k"
+    - "[AI generated draft test] The shipped ROCm image's AMD MiniMax-M3 model lacks SupportsEagle3, so the recipe patches it in-place at runtime (functionstackx/vllm#1, upstream vllm-project/vllm#45546; validated green on MI355X) before serving — validates EAGLE3 on MI300X ahead of an image rebuild"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1749
+
+- config-keys:
+    - minimaxm3-fp8-mi300x-vllm
+  description:
+    - "Enable CUDA graphs for MiniMax-M3 MXFP8 on MI300X and set VLLM_USE_BREAKABLE_CUDAGRAPH=0 per AMD guidance"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1750
+
+- config-keys:
+    - minimaxm3-fp8-mi300x-vllm-mtp
+  description:
+    - "Run the MiniMax-M3 MXFP8 MI300X EAGLE3 MTP recipe with CUDA graphs instead of --enforce-eager"
+    - "Drop --enforce-eager and set VLLM_USE_BREAKABLE_CUDAGRAPH=0 (matching the non-MTP MI300X recipe, #1750), which avoids the M3-decode breakable-cudagraph path that previously forced eager execution"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1756
diff --git a/runners/launch_gb300-cw.sh b/runners/launch_gb300-cw.sh
index 6a5c50e38..50dae4465 100644
--- a/runners/launch_gb300-cw.sh
+++ b/runners/launch_gb300-cw.sh
@@ -59,8 +59,24 @@ elif [[ $MODEL_PREFIX == "glm5" && $PRECISION == "fp8" ]]; then
         echo "Unsupported framework on gb300-cw for glm5/fp8: $FRAMEWORK. Currently supported: dynamo-sglang"
         exit 1
     fi
+elif [[ $MODEL_PREFIX == "minimaxm3" && $PRECISION == "fp8" ]]; then
+    # Day-zero: MiniMax-M3-MXFP8 is not staged on this cluster. The recipes
+    # carry an hf: model id directly, so srtctl pre-downloads the snapshot
+    # into the shared HF_HOME sed-injected below; MODEL_PATH only feeds the
+    # (unreferenced) model_paths aliases in srtslurm.yaml.
+    export MODEL_PATH="hf:MiniMaxAI/MiniMax-M3-MXFP8"
+
+    if [[ $FRAMEWORK == "dynamo-vllm" ]]; then
+        SRT_SLURM_RECIPES_REPO="https://github.com/NVIDIA/srt-slurm.git"
+        SRT_SLURM_RECIPES_REF="main"
+        SRT_RECIPE_SRC="$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8"
+        SRT_RECIPE_DST="recipes/vllm/minimax-m3-gb300-fp8"
+    else
+        echo "Unsupported framework on gb300-cw for minimaxm3/fp8: $FRAMEWORK. Currently supported: dynamo-vllm"
+        exit 1
+    fi
 else
-    echo "Unsupported model prefix/precision combination on gb300-cw: $MODEL_PREFIX/$PRECISION. Currently supported: dsv4/fp4, glm5/fp8"
+    echo "Unsupported model prefix/precision combination on gb300-cw: $MODEL_PREFIX/$PRECISION. Currently supported: dsv4/fp4, glm5/fp8, minimaxm3/fp8"
     exit 1
 fi
 
@@ -276,6 +292,39 @@ else
     mv "$TMP_CONFIG_FILE" "$CONFIG_FILE"
 fi
 
+# MiniMax-M3 day-zero: the recipes use an hf: model id and need a shared-FS
+# HF_HOME visible (and writable) on compute nodes for srtctl's one-time
+# pre-download of the 444 GB snapshot. /mnt/vast is the shared NFS that
+# already hosts models and squash files on this cluster.
+if [[ $MODEL_PREFIX == "minimaxm3" ]]; then
+    M3_HF_HOME="/mnt/vast/hf-home"
+    mkdir -p "$M3_HF_HOME"
+    sed -i "s|__M3_HF_HOME__|${M3_HF_HOME}|g" "$CONFIG_FILE"
+    # Dynamo's rust fetch_model dies instantly on ANY held .lock file
+    # ("Lock acquisition failed") — it doesn't retry like Python's
+    # huggingface_hub.  Concurrent GHA jobs race on the 444 GB download
+    # and create fresh locks that survive the old "mmin +10" cleanup.
+    # Fix: nuke ALL locks (safe — HF uses atomic rename from .incomplete),
+    # then force-download with the Python client (which DOES wait for
+    # locks) so srtctl's pre-download is a no-op and dynamo sees a fully
+    # cached snapshot with zero lock files.
+    find "$M3_HF_HOME" -name '*.lock' -delete 2>/dev/null || true
+    export HF_HOME="$M3_HF_HOME"
+    DL_CMD="huggingface-cli download"
+    command -v huggingface-cli >/dev/null 2>&1 || DL_CMD="hf download"
+    for _attempt in 1 2 3; do
+        if HF_HUB_OFFLINE=1 $DL_CMD MiniMaxAI/MiniMax-M3-MXFP8 --quiet 2>/dev/null; then
+            echo "MiniMax-M3-MXFP8 fully cached (verified offline)"
+            break
+        fi
+        echo "MiniMax-M3 cache incomplete, downloading (attempt $_attempt)..."
+        find "$M3_HF_HOME" -name '*.lock' -delete 2>/dev/null || true
+        $DL_CMD MiniMaxAI/MiniMax-M3-MXFP8 --quiet 2>&1 | tail -5 || true
+        sleep 5
+    done
+    find "$M3_HF_HOME" -name '*.lock' -delete 2>/dev/null || true
+fi
+
 SRTCTL_OUTPUT=$(srtctl apply -f "$CONFIG_FILE" --tags "gb300,${MODEL_PREFIX},${PRECISION},${ISL}x${OSL},infmax-$(date +%Y%m%d)" 2>&1)
 echo "$SRTCTL_OUTPUT"
 
diff --git a/runners/launch_mi300x-amds.sh b/runners/launch_mi300x-amds.sh
index ce04ceadd..b0c1e22c8 100644
--- a/runners/launch_mi300x-amds.sh
+++ b/runners/launch_mi300x-amds.sh
@@ -7,6 +7,10 @@ PARTITION="compute"
 SQUASH_FILE="/home/gharunner/gharunners/squash/$(echo "$IMAGE" | sed 's/[\/:@#]/_/g').sqsh"
 LOCK_FILE="${SQUASH_FILE}.lock"
 
+# Route spec-decoding=mtp configs to the _mtp benchmark script (parity with
+# the h200 launchers, which have carried SPEC_SUFFIX since #392).
+SPEC_SUFFIX=$([[ "$SPEC_DECODING" == "mtp" ]] && printf '_mtp' || printf '')
+
 set -x
 
 # Exclude known-bad nodes; let Slurm pick from anything else:
@@ -37,6 +41,6 @@ srun --jobid=$JOB_ID \
 --container-remap-root \
 --container-workdir=/workspace/ \
 --no-container-entrypoint --export=ALL \
-bash benchmarks/single_node/${SCENARIO_SUBDIR}${EXP_NAME%%_*}_${PRECISION}_mi300x.sh
+bash benchmarks/single_node/${SCENARIO_SUBDIR}${EXP_NAME%%_*}_${PRECISION}_mi300x${SPEC_SUFFIX}.sh
 
 scancel $JOB_ID
diff --git a/utils/aiperf b/utils/aiperf
index 062a5de92..ff2b646c0 160000
--- a/utils/aiperf
+++ b/utils/aiperf
@@ -1 +1 @@
-Subproject commit 062a5de92c8ac8a0a6dd5d2a7fb9a539a147f3d9
+Subproject commit ff2b646c0425aff9307a0e73161b23d77003a357