diff --git a/.github/configs/amd-master.yaml b/.github/configs/amd-master.yaml
index 418ad5ab9..41d12d251 100644
--- a/.github/configs/amd-master.yaml
+++ b/.github/configs/amd-master.yaml
@@ -589,6 +589,98 @@ kimik2.5-int4-mi355x-vllm:
       search-space:
       - { tp: 8, conc-start: 4, conc-end: 64 }
 
+kimik2.5-mxfp4-mi355x-vllm-eagle3:
+  image: vllm/vllm-openai-rocm:v0.21.0
+  model: amd/Kimi-K2.5-MXFP4
+  model-prefix: kimik2.5
+  runner: mi355x
+  precision: fp4
+  framework: vllm
+  multinode: false
+  scenarios:
+    fixed-seq-len:
+    - isl: 1024
+      osl: 1024
+      search-space:
+      - { tp: 8, conc-start: 4, conc-end: 64, spec-decoding: eagle3 }
+    - isl: 8192
+      osl: 1024
+      search-space:
+      - { tp: 8, conc-start: 4, conc-end: 64, spec-decoding: eagle3 }
+
+kimik2.5-int4-mi355x-vllm-eagle3:
+  image: vllm/vllm-openai-rocm:v0.21.0
+  model: moonshotai/Kimi-K2.5
+  model-prefix: kimik2.5
+  runner: mi355x
+  precision: int4
+  framework: vllm
+  multinode: false
+  scenarios:
+    fixed-seq-len:
+    - isl: 1024
+      osl: 1024
+      search-space:
+      - { tp: 8, conc-start: 4, conc-end: 64, spec-decoding: eagle3 }
+    - isl: 8192
+      osl: 1024
+      search-space:
+      - { tp: 8, conc-start: 4, conc-end: 64, spec-decoding: eagle3 }
+
+kimik2.5-int4-mi355x-vllm-fixed-ar-mtp:
+  image: vllm/vllm-openai-rocm:v0.21.0
+  model: moonshotai/Kimi-K2.5
+  model-prefix: kimik2.5
+  runner: mi355x
+  precision: int4
+  framework: vllm
+  multinode: false
+  scenarios:
+    fixed-ar-mtp:
+    - isl: 1024
+      osl: 1024
+      draft-model: nvidia/Kimi-K2.5-Thinking-Eagle3
+      num-speculative-tokens: 3
+      rejection-sample-method: synthetic
+      synthetic-acceptance-rates: [0.778774, 0.57543, 0.412793]
+      search-space:
+      - { tp: 8, conc-start: 4, conc-end: 64, spec-decoding: eagle3 }
+    - isl: 8192
+      osl: 1024
+      draft-model: nvidia/Kimi-K2.5-Thinking-Eagle3
+      num-speculative-tokens: 3
+      rejection-sample-method: synthetic
+      synthetic-acceptance-rates: [0.778774, 0.57543, 0.412793]
+      search-space:
+      - { tp: 8, conc-start: 4, conc-end: 64, spec-decoding: eagle3 }
+
+kimik2.5-fp4-mi355x-vllm-fixed-ar-mtp:
+  image: vllm/vllm-openai-rocm:v0.21.0
+  model: amd/Kimi-K2.5-MXFP4
+  model-prefix: kimik2.5
+  runner: mi355x
+  precision: fp4
+  framework: vllm
+  multinode: false
+  scenarios:
+    fixed-ar-mtp:
+    - isl: 1024
+      osl: 1024
+      draft-model: lightseekorg/kimi-k2.5-eagle3
+      num-speculative-tokens: 3
+      rejection-sample-method: synthetic
+      synthetic-acceptance-rates: [0.778774, 0.57543, 0.412793]
+      search-space:
+      - { tp: 8, conc-start: 4, conc-end: 64, spec-decoding: eagle3 }
+    - isl: 8192
+      osl: 1024
+      draft-model: lightseekorg/kimi-k2.5-eagle3
+      num-speculative-tokens: 3
+      rejection-sample-method: synthetic
+      synthetic-acceptance-rates: [0.778774, 0.57543, 0.412793]
+      search-space:
+      - { tp: 8, conc-start: 4, conc-end: 64, spec-decoding: eagle3 }
+
 kimik2.5-int4-mi325x-vllm:
   image: vllm/vllm-openai-rocm:v0.21.0
   model: moonshotai/Kimi-K2.5
@@ -724,6 +816,25 @@ minimaxm2.5-fp8-mi355x-vllm:
       - { tp: 4, ep: 4, conc-start: 4, conc-end: 512 }
       - { tp: 8, ep: 8, conc-start: 2, conc-end: 2 }
 
+minimaxm2.5-fp8-mi355x-vllm-eagle3:
+  image: vllm/vllm-openai-rocm:v0.21.0
+  model: MiniMaxAI/MiniMax-M2.5
+  model-prefix: minimaxm2.5
+  runner: mi355x
+  precision: fp8
+  framework: vllm
+  multinode: false
+  scenarios:
+    fixed-seq-len:
+    - isl: 1024
+      osl: 1024
+      search-space:
+      - { tp: 8, conc-start: 4, conc-end: 64, spec-decoding: eagle3 }
+    - isl: 8192
+      osl: 1024
+      search-space:
+      - { tp: 8, conc-start: 4, conc-end: 64, spec-decoding: eagle3 }
+
 # Diverged from minimaxm2.5-fp8-mi355x-vllm (agentic-coding sibling). Reasons below;
 # the original minimaxm2.5-fp8-mi355x-vllm entry is left identical to origin/main so
 # its fixed-seq-len sweep is unaffected.
diff --git a/.github/workflows/benchmark-tmpl.yml b/.github/workflows/benchmark-tmpl.yml
index cca6031c3..82153570e 100644
--- a/.github/workflows/benchmark-tmpl.yml
+++ b/.github/workflows/benchmark-tmpl.yml
@@ -53,7 +53,6 @@ on:
       run-eval:
         type: boolean
         required: true
-        default: false
       eval-only:
         description: "Run only evals (skip throughput benchmark)"
         type: boolean
@@ -68,10 +67,30 @@ on:
         required: false
         type: string
       scenario-type:
-        description: "Scenario type (fixed-seq-len or agentic-coding)"
+        description: "Scenario type (fixed-seq-len, agentic-coding, or fixed-ar-mtp)"
         required: false
         type: string
         default: 'fixed-seq-len'
+      draft-model:
+        description: "Draft model for fixed-AR MTP scenarios"
+        required: false
+        type: string
+        default: ''
+      num-speculative-tokens:
+        description: "Number of speculative tokens for fixed-AR MTP scenarios"
+        required: false
+        type: string
+        default: ''
+      rejection-sample-method:
+        description: "Speculative rejection sampling method"
+        required: false
+        type: string
+        default: ''
+      synthetic-acceptance-rates:
+        description: "JSON array of synthetic acceptance rates for fixed-AR MTP scenarios"
+        required: false
+        type: string
+        default: ''
       offloading:
         description: "KV offload backend for agentic scenarios (none/cpu/ssd)"
         required: false
@@ -111,6 +130,10 @@ env:
   SCENARIO_TYPE: ${{ inputs.scenario-type }}
   SCENARIO_SUBDIR: ${{ inputs.scenario-type == 'agentic-coding' && 'agentic/' || '' }}
   IS_AGENTIC: ${{ inputs.scenario-type == 'agentic-coding' && '1' || '0' }}
+  DRAFT_MODEL: ${{ inputs.draft-model }}
+  NUM_SPECULATIVE_TOKENS: ${{ inputs.num-speculative-tokens }}
+  REJECTION_SAMPLE_METHOD: ${{ inputs.rejection-sample-method }}
+  SYNTHETIC_ACCEPTANCE_RATES: ${{ inputs.synthetic-acceptance-rates }}
   OFFLOADING: ${{ inputs.offloading }}
   TOTAL_CPU_DRAM_GB: ${{ inputs.total-cpu-dram-gb }}
   DURATION: ${{ inputs.duration }}
diff --git a/.github/workflows/e2e-tests.yml b/.github/workflows/e2e-tests.yml
index fea89fcae..290c503ba 100644
--- a/.github/workflows/e2e-tests.yml
+++ b/.github/workflows/e2e-tests.yml
@@ -51,6 +51,7 @@ jobs:
             multi-node-eval-config: ${{ steps.get-jobs.outputs.multi-node-eval-config }}
             agentic-config: ${{ steps.get-jobs.outputs.agentic-config }}
             multi-node-agentic-config: ${{ steps.get-jobs.outputs.multi-node-agentic-config }}
+            fixed-ar-mtp-config: ${{ steps.get-jobs.outputs.fixed-ar-mtp-config }}
         steps:
             - name: Checkout code (ref)
               if: ${{ inputs.ref && inputs.ref != '' }}
@@ -71,12 +72,14 @@ jobs:
                     ${{ inputs.generate-cli-command || github.event.inputs.generate-cli-command }})
                   AGENTIC=$(echo "$CONFIG_JSON" | python3 -c "import sys,json; d=json.load(sys.stdin); print(json.dumps([x for x in d if x.get('scenario-type') == 'agentic-coding' and 'prefill' not in x]))")
                   MULTI_AGENTIC=$(echo "$CONFIG_JSON" | python3 -c "import sys,json; d=json.load(sys.stdin); print(json.dumps([x for x in d if x.get('scenario-type') == 'agentic-coding' and 'prefill' in x]))")
-                  SINGLE=$(echo "$CONFIG_JSON" | python3 -c "import sys,json; d=json.load(sys.stdin); print(json.dumps([x for x in d if 'prefill' not in x and x.get('scenario-type') != 'agentic-coding' and not x.get('eval-only', False)]))")
+                  FIXED_AR_MTP=$(echo "$CONFIG_JSON" | python3 -c "import sys,json; d=json.load(sys.stdin); print(json.dumps([x for x in d if x.get('scenario-type') == 'fixed-ar-mtp']))")
+                  SINGLE=$(echo "$CONFIG_JSON" | python3 -c "import sys,json; d=json.load(sys.stdin); print(json.dumps([x for x in d if 'prefill' not in x and x.get('scenario-type') not in ('agentic-coding', 'fixed-ar-mtp') and not x.get('eval-only', False)]))")
                   MULTI=$(echo "$CONFIG_JSON" | python3 -c "import sys,json; d=json.load(sys.stdin); print(json.dumps([x for x in d if 'prefill' in x and x.get('scenario-type') != 'agentic-coding' and not x.get('eval-only', False)]))")
                   EVALS=$(echo "$CONFIG_JSON" | python3 -c "import sys,json; d=json.load(sys.stdin); print(json.dumps([x for x in d if 'prefill' not in x and x.get('scenario-type') != 'agentic-coding' and x.get('run-eval', False)]))")
                   MULTI_EVAL=$(echo "$CONFIG_JSON" | python3 -c "import sys,json; d=json.load(sys.stdin); print(json.dumps([x for x in d if 'prefill' in x and x.get('run-eval', False)]))")
                   echo "agentic-config=$AGENTIC" >> $GITHUB_OUTPUT
                   echo "multi-node-agentic-config=$MULTI_AGENTIC" >> $GITHUB_OUTPUT
+                  echo "fixed-ar-mtp-config=$FIXED_AR_MTP" >> $GITHUB_OUTPUT
                   echo "single-node-config=$SINGLE" >> $GITHUB_OUTPUT
                   echo "multi-node-config=$MULTI" >> $GITHUB_OUTPUT
                   echo "eval-config=$EVALS" >> $GITHUB_OUTPUT
@@ -190,7 +193,7 @@ jobs:
             osl: '0'
             max-model-len: '0'
             spec-decoding: 'none'
-            disagg: 'false'
+            disagg: ${{ 'false' }}
             run-eval: false
             scenario-type: agentic-coding
             ref: ${{ inputs.ref }}
@@ -235,6 +238,41 @@ jobs:
             scenario-type: agentic-coding
             ref: ${{ inputs.ref }}
 
+    test-sweep-fixed-ar-mtp:
+        needs: get-jobs
+        if: ${{ needs.get-jobs.outputs.fixed-ar-mtp-config != '[]' }}
+        uses: ./.github/workflows/benchmark-tmpl.yml
+        name: Fixed-AR-MTP throughput /
+        strategy:
+            fail-fast: false
+            matrix:
+                config: ${{ fromJson(needs.get-jobs.outputs.fixed-ar-mtp-config) }}
+        secrets: inherit
+        with:
+            exp-name: ${{ matrix.config.exp-name }}
+            isl: ${{ matrix.config.isl }}
+            osl: ${{ matrix.config.osl }}
+            max-model-len: ${{ matrix.config.max-model-len }}
+            runner: ${{ matrix.config.runner }}
+            image: ${{ matrix.config.image }}
+            model: ${{ matrix.config.model }}
+            model-prefix: ${{ matrix.config.model-prefix }}
+            framework: ${{ matrix.config.framework }}
+            precision: ${{ matrix.config.precision }}
+            tp: ${{ matrix.config.tp }}
+            ep: ${{ matrix.config.ep }}
+            dp-attn: ${{ matrix.config.dp-attn }}
+            conc: ${{ matrix.config.conc }}
+            spec-decoding: ${{ matrix.config.spec-decoding }}
+            disagg: ${{ matrix.config.disagg }}
+            run-eval: false
+            scenario-type: fixed-ar-mtp
+            draft-model: ${{ matrix.config.draft-model }}
+            num-speculative-tokens: ${{ matrix.config.num-speculative-tokens }}
+            rejection-sample-method: ${{ matrix.config.rejection-sample-method }}
+            synthetic-acceptance-rates: ${{ toJson(matrix.config.synthetic-acceptance-rates) }}
+            ref: ${{ inputs.ref }}
+
     test-sweep-single-node:
         needs: get-jobs
         if: ${{ needs.get-jobs.outputs.single-node-config != '[]' }}
@@ -297,8 +335,8 @@ jobs:
             ref: ${{ inputs.ref }}
 
     collect-results:
-        needs: [test-sweep-multi-node, test-sweep-single-node, test-sweep-agentic, test-sweep-multi-node-agentic]
-        if: ${{ always() && (needs.test-sweep-multi-node.result != 'skipped' || needs.test-sweep-single-node.result != 'skipped' || needs.test-sweep-agentic.result != 'skipped' || needs.test-sweep-multi-node-agentic.result != 'skipped') }}
+        needs: [test-sweep-multi-node, test-sweep-single-node, test-sweep-fixed-ar-mtp, test-sweep-agentic, test-sweep-multi-node-agentic]
+        if: ${{ always() && (needs.test-sweep-multi-node.result != 'skipped' || needs.test-sweep-single-node.result != 'skipped' || needs.test-sweep-fixed-ar-mtp.result != 'skipped' || needs.test-sweep-agentic.result != 'skipped' || needs.test-sweep-multi-node-agentic.result != 'skipped') }}
         uses: ./.github/workflows/collect-results.yml
         secrets: inherit
         with:
diff --git a/.github/workflows/mtp-fixed-ar-amd.yml b/.github/workflows/mtp-fixed-ar-amd.yml
new file mode 100644
index 000000000..3f6d0a391
--- /dev/null
+++ b/.github/workflows/mtp-fixed-ar-amd.yml
@@ -0,0 +1,619 @@
+name: MTP benchmark with fixed AR
+run-name: MTP fixed AR - ${{ inputs.test-name || inputs.model }}
+
+on:
+  workflow_dispatch:
+    inputs:
+      test-name:
+        description: "Display name for this run"
+        required: false
+        type: string
+        default: "lcb-v6-kimik2.5-MXFP4-mtp-fixed-ar"
+      runner:
+        description: "MI355X self-hosted runner to use"
+        required: true
+        type: string
+        default: "mi355x-amds_0"
+      image:
+        description: "ROCm vLLM Docker image"
+        required: true
+        type: string
+        default: "vllm/vllm-openai-rocm:v0.21.0"
+      model:
+        description: "Target model served by vLLM"
+        required: true
+        type: string
+        default: "amd/Kimi-K2.5-MXFP4"
+      draft-model:
+        description: "Draft model used by Eagle3 speculative decoding"
+        required: true
+        type: string
+        default: "lightseekorg/kimi-k2.5-eagle3"
+      dataset:
+        description: "Hugging Face dataset used for acceptance-rate sampling"
+        required: true
+        type: string
+        default: "livecodebench/code_generation_lite"
+      dataset-config:
+        description: "Optional Hugging Face dataset config/subset name"
+        required: false
+        type: string
+        default: ""
+      dataset-split:
+        description: "Dataset split to load. Empty means test if present, otherwise first split."
+        required: false
+        type: string
+        default: "test6"
+      dataset-prompt-fields:
+        description: "Comma-separated candidate prompt fields, supports dotted paths"
+        required: false
+        type: string
+        default: "question_content"
+      prompt-turns:
+        description: "For list-based prompts such as MT-Bench turns: first or all"
+        required: true
+        type: choice
+        options:
+          - first
+          - all
+        default: "first"
+      num-prompts:
+        description: "Number of prompts to sample from the dataset"
+        required: true
+        type: string
+        default: "200"
+      max-concurrency:
+        description: "vLLM bench serve max concurrency. Supports a single value or comma-separated values, e.g. 32,64"
+        required: true
+        type: string
+        default: "32,64"
+      port:
+        description: "vLLM server port"
+        required: true
+        type: string
+        default: "8000"
+      num-speculative-tokens:
+        description: "Number of speculative tokens"
+        required: true
+        type: string
+        default: "3"
+      tensor-parallel-size:
+        description: "Tensor parallel size for vLLM serve"
+        required: true
+        type: string
+        default: "8"
+      draft-tensor-parallel-size:
+        description: "Draft model tensor parallel size for Eagle3 speculative decoding"
+        required: true
+        type: string
+        default: "1"
+      vllm-env-json:
+        description: "Optional JSON object of extra environment variables for vLLM serve"
+        required: false
+        type: string
+        default: '{"VLLM_ROCM_USE_AITER":"1","VLLM_ROCM_QUICK_REDUCE_QUANTIZATION":"INT4","VLLM_ROCM_USE_AITER_RMSNORM":"0"}'
+      vllm-extra-args:
+        description: "Optional extra arguments appended to vLLM serve"
+        required: false
+        type: string
+        default: "--enable-expert-parallel --long-prefill-token-threshold 8192 --max-num-batched-tokens 16384"
+      bench-extra-args:
+        description: "Optional extra arguments appended to vLLM bench serve"
+        required: false
+        type: string
+        default: ""
+
+permissions:
+  contents: read
+
+jobs:
+  generate-synthetic-acceptance-rates:
+    runs-on: ${{ inputs.runner }}
+    timeout-minutes: 500
+    outputs:
+      synthetic-acceptance-rates: ${{ steps.compute-rates.outputs.synthetic_acceptance_rates }}
+      speculative-config: ${{ steps.compute-rates.outputs.speculative_config }}
+
+    env:
+      CONTAINER_NAME: mtp-fixed-ar-${{ github.run_id }}-${{ github.run_attempt }}
+      HOST_WORK_DIR: /tmp/mtp-fixed-ar-${{ github.run_id }}-${{ github.run_attempt }}
+      CONTAINER_WORK_DIR: /work/tmp/mtp-fixed-ar-${{ github.run_id }}-${{ github.run_attempt }}
+      DATASET_JSONL: dataset_prompts.jsonl
+      HF_TOKEN: ${{ secrets.HF_TOKEN }}
+      HUGGING_FACE_HUB_TOKEN: ${{ secrets.HF_TOKEN }}
+
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+
+      - name: Prepare workspace
+        run: |
+          set -euo pipefail
+          mkdir -p "$HOST_WORK_DIR"
+          docker rm -f "$CONTAINER_NAME" >/dev/null 2>&1 || true
+
+      - name: Launch ROCm vLLM container
+        run: |
+          set -euo pipefail
+          docker run -d -it \
+            --ipc=host \
+            --network=host \
+            --privileged \
+            --cap-add=CAP_SYS_ADMIN \
+            --device=/dev/kfd \
+            --device=/dev/dri \
+            --device=/dev/mem \
+            --group-add video \
+            --cap-add=SYS_PTRACE \
+            --security-opt seccomp=unconfined \
+            --shm-size 32G \
+            -e HF_TOKEN \
+            -e HUGGING_FACE_HUB_TOKEN \
+            -v ~/.cache/huggingface:/root/.cache/huggingface \
+            -v /:/work \
+            --entrypoint "/bin/bash" \
+            --name "$CONTAINER_NAME" \
+            "${{ inputs.image }}"
+
+      - name: Download dataset and convert to vLLM JSONL
+        env:
+          DATASET: ${{ inputs.dataset }}
+          DATASET_CONFIG: ${{ inputs.dataset-config }}
+          DATASET_SPLIT: ${{ inputs.dataset-split }}
+          DATASET_PROMPT_FIELDS: ${{ inputs.dataset-prompt-fields }}
+          PROMPT_TURNS: ${{ inputs.prompt-turns }}
+          NUM_PROMPTS: ${{ inputs.num-prompts }}
+        run: |
+          set -euo pipefail
+          docker exec \
+            -e HF_TOKEN \
+            -e HUGGING_FACE_HUB_TOKEN \
+            -e DATASET \
+            -e DATASET_CONFIG \
+            -e DATASET_SPLIT \
+            -e DATASET_PROMPT_FIELDS \
+            -e PROMPT_TURNS \
+            -e NUM_PROMPTS \
+            -e CONTAINER_WORK_DIR \
+            -e DATASET_JSONL \
+            "$CONTAINER_NAME" \
+            bash -lc '
+              set -euo pipefail
+              mkdir -p "$CONTAINER_WORK_DIR"
+              python3 -m pip install -q --no-cache-dir datasets huggingface_hub || \
+                python3 -m pip install -q --no-cache-dir --break-system-packages datasets huggingface_hub
+              hf download "$DATASET" --repo-type=dataset >/dev/null
+              python3 - <<'"'"'PY'"'"'
+          import json
+          import os
+          from pathlib import Path
+          from datasets import get_dataset_split_names, load_dataset
+          from huggingface_hub import snapshot_download
+
+          dataset_name = os.environ["DATASET"]
+          dataset_config = os.environ.get("DATASET_CONFIG") or None
+          requested_split = os.environ.get("DATASET_SPLIT") or None
+          prompt_fields = [
+              field.strip()
+              for field in os.environ.get("DATASET_PROMPT_FIELDS", "").split(",")
+              if field.strip()
+          ]
+          prompt_turns = os.environ.get("PROMPT_TURNS", "first")
+          num_prompts = int(os.environ["NUM_PROMPTS"])
+          out_path = os.path.join(os.environ["CONTAINER_WORK_DIR"], os.environ["DATASET_JSONL"])
+
+          split_kwargs = {"path": dataset_name}
+          load_kwargs = {"path": dataset_name}
+          if dataset_config:
+              split_kwargs["name"] = dataset_config
+              load_kwargs["name"] = dataset_config
+
+          def get_path(row, path):
+              value = row
+              for part in path.split("."):
+                  if isinstance(value, dict) and part in value:
+                      value = value[part]
+                  else:
+                      return None
+              return value
+
+          def message_text(message):
+              if isinstance(message, str):
+                  return message
+              if isinstance(message, dict):
+                  content = message.get("content") or message.get("value") or message.get("text")
+                  if isinstance(content, str):
+                      return content
+              return None
+
+          def normalize_prompt(value):
+              if isinstance(value, str) and value.strip():
+                  return value
+              if isinstance(value, list):
+                  texts = []
+                  for item in value:
+                      text = message_text(item)
+                      if text and text.strip():
+                          texts.append(text)
+                  if texts:
+                      return "\n\n".join(texts) if prompt_turns == "all" else texts[0]
+              if isinstance(value, dict):
+                  for key in ("prompt", "content", "value", "text"):
+                      prompt = normalize_prompt(value.get(key))
+                      if prompt:
+                          return prompt
+              return None
+
+          def load_rows_from_downloaded_files():
+              snapshot = Path(snapshot_download(dataset_name, repo_type="dataset"))
+              split_name = requested_split or "test"
+              candidates = [snapshot / f"{split_name}.jsonl", snapshot / f"{split_name}.json"]
+              if not requested_split:
+                  candidates.extend(sorted(snapshot.glob("*.jsonl")))
+                  candidates.extend(sorted(snapshot.glob("*.json")))
+
+              data_file = next((path for path in candidates if path.is_file()), None)
+              if data_file is None:
+                  raise FileNotFoundError(
+                      f"No JSON/JSONL data file found in downloaded dataset snapshot: {snapshot}"
+                  )
+
+              rows = []
+              with data_file.open(encoding="utf-8") as fh:
+                  if data_file.suffix == ".jsonl":
+                      for line in fh:
+                          if line.strip():
+                              rows.append(json.loads(line))
+                              if len(rows) >= num_prompts:
+                                  break
+                  else:
+                      data = json.load(fh)
+                      if isinstance(data, dict):
+                          for key in ("data", "rows", "examples"):
+                              if isinstance(data.get(key), list):
+                                  data = data[key]
+                                  break
+                      if not isinstance(data, list):
+                          raise ValueError(f"Unsupported JSON dataset shape in {data_file}")
+                      rows = data[:num_prompts]
+              return rows, data_file.name
+
+          try:
+              splits = get_dataset_split_names(**split_kwargs)
+              split = requested_split or ("test" if "test" in splits else splits[0])
+              dataset = load_dataset(**load_kwargs, split=split)
+              rows = dataset.select(range(min(num_prompts, len(dataset))))
+              source = f"split={split}"
+          except RuntimeError as exc:
+              if "Dataset scripts are no longer supported" not in str(exc):
+                  raise
+              rows, source = load_rows_from_downloaded_files()
+
+          rows = list(rows)
+          if rows and len(rows) < num_prompts:
+              rows = [rows[i % len(rows)] for i in range(num_prompts)]
+
+          # vLLM's custom dataset mode consumes JSONL rows with a prompt field.
+          # This replaces a fixed path like /work/.../livecodebench_v6_prompts.jsonl.
+          with open(out_path, "w", encoding="utf-8") as out:
+              for row in rows:
+                  prompt = None
+                  for field in prompt_fields:
+                      prompt = normalize_prompt(get_path(row, field))
+                      if prompt:
+                          break
+                  if prompt is None:
+                      prompt = json.dumps(row, ensure_ascii=False)
+                  out.write(json.dumps({"prompt": prompt}, ensure_ascii=False) + "\n")
+
+          print(f"Wrote prompts from {dataset_name} {source} to {out_path}")
+          PY
+            '
+
+      - name: Launch vLLM server with standard speculative decoding
+        env:
+          MODEL: ${{ inputs.model }}
+          DRAFT_MODEL: ${{ inputs.draft-model }}
+          PORT: ${{ inputs.port }}
+          NUM_SPECULATIVE_TOKENS: ${{ inputs.num-speculative-tokens }}
+          TENSOR_PARALLEL_SIZE: ${{ inputs.tensor-parallel-size }}
+          DRAFT_TENSOR_PARALLEL_SIZE: ${{ inputs.draft-tensor-parallel-size }}
+          VLLM_ENV_JSON: ${{ inputs.vllm-env-json }}
+          VLLM_EXTRA_ARGS: ${{ inputs.vllm-extra-args }}
+        run: |
+          set -euo pipefail
+          docker exec -d \
+            -e MODEL \
+            -e DRAFT_MODEL \
+            -e PORT \
+            -e NUM_SPECULATIVE_TOKENS \
+            -e TENSOR_PARALLEL_SIZE \
+            -e DRAFT_TENSOR_PARALLEL_SIZE \
+            -e VLLM_ENV_JSON \
+            -e VLLM_EXTRA_ARGS \
+            -e CONTAINER_WORK_DIR \
+            "$CONTAINER_NAME" \
+            bash -lc '
+              set -euo pipefail
+              DEFAULT_DRAFT_MODEL="lightseekorg/kimi-k2.5-eagle3"
+              DEFAULT_VLLM_ENV_JSON='"'"'{"VLLM_ROCM_USE_AITER":"1","VLLM_ROCM_QUICK_REDUCE_QUANTIZATION":"INT4","VLLM_ROCM_USE_AITER_RMSNORM":"0"}'"'"'
+              DEFAULT_VLLM_EXTRA_ARGS="--enable-expert-parallel --long-prefill-token-threshold 8192 --max-num-batched-tokens 16384"
+              if [[ "$MODEL" == "moonshotai/Kimi-K2.5" || "$MODEL" == "amd/Kimi-K2.5-MXFP4" ]]; then
+                if [[ -z "${VLLM_EXTRA_ARGS:-}" ]]; then
+                  VLLM_EXTRA_ARGS="$DEFAULT_VLLM_EXTRA_ARGS"
+                fi
+              fi
+              if [[ "$MODEL" == "MiniMaxAI/MiniMax-M2.5" ]]; then
+                if [[ "$DRAFT_MODEL" == "$DEFAULT_DRAFT_MODEL" ]]; then
+                  DRAFT_MODEL="thoughtworks/MiniMax-M2.5-Eagle3"
+                fi
+                if [[ "${VLLM_ENV_JSON:-}" == "$DEFAULT_VLLM_ENV_JSON" ]]; then
+                  VLLM_ENV_JSON='"'"'{"VLLM_ROCM_USE_AITER":"1","VLLM_ROCM_QUICK_REDUCE_QUANTIZATION":"INT4"}'"'"'
+                fi
+                if [[ -z "${VLLM_EXTRA_ARGS:-}" || "${VLLM_EXTRA_ARGS:-}" == "$DEFAULT_VLLM_EXTRA_ARGS" ]]; then
+                  VLLM_EXTRA_ARGS="--enable-expert-parallel --kv-cache-dtype fp8 --block-size=32 --attention-backend ROCM_AITER_FA"
+                fi
+              fi
+              python3 -m pip install -q --no-cache-dir amd-quark || \
+                python3 -m pip install -q --no-cache-dir --break-system-packages amd-quark
+              SPECULATIVE_CONFIG=$(python3 - <<'"'"'PY'"'"'
+          import json
+          import os
+
+          print(json.dumps({
+              "model": os.environ["DRAFT_MODEL"],
+              "method": "eagle3",
+              "num_speculative_tokens": int(os.environ["NUM_SPECULATIVE_TOKENS"]),
+              "rejection_sample_method": "standard",
+              "draft_tensor_parallel_size": int(os.environ["DRAFT_TENSOR_PARALLEL_SIZE"]),
+          }))
+          PY
+          )
+              while IFS= read -r export_line; do
+                export "$export_line"
+              done < <(python3 - <<'"'"'PY'"'"'
+          import json
+          import os
+
+          env = json.loads(os.environ.get("VLLM_ENV_JSON") or "{}")
+          if not isinstance(env, dict):
+              raise SystemExit("vllm-env-json must be a JSON object")
+          for key, value in env.items():
+              print(f"{key}={value}")
+          PY
+          )
+              vllm serve "$MODEL" \
+                --host 0.0.0.0 \
+                --port "$PORT" \
+                -tp "$TENSOR_PARALLEL_SIZE" \
+                --mm-encoder-tp-mode data \
+                --no-enable-prefix-caching \
+                --speculative-config "$SPECULATIVE_CONFIG" \
+                --trust-remote-code \
+                $VLLM_EXTRA_ARGS \
+                > "$CONTAINER_WORK_DIR/vllm_server.log" 2>&1
+            '
+
+      - name: Wait for vLLM server
+        env:
+          PORT: ${{ inputs.port }}
+        run: |
+          set -euo pipefail
+          for i in {1..180}; do
+            if curl --silent --fail "http://127.0.0.1:${PORT}/health" >/dev/null; then
+              exit 0
+            fi
+            docker exec "$CONTAINER_NAME" bash -lc "tail -n 80 '$CONTAINER_WORK_DIR/vllm_server.log' 2>/dev/null || true"
+            sleep 10
+          done
+          echo "vLLM server did not become healthy" >&2
+          exit 1
+
+      - name: Run LiveCodeBench sampling benchmark
+        env:
+          MODEL: ${{ inputs.model }}
+          PORT: ${{ inputs.port }}
+          NUM_PROMPTS: ${{ inputs.num-prompts }}
+          MAX_CONCURRENCY: ${{ inputs.max-concurrency }}
+          BENCH_EXTRA_ARGS: ${{ inputs.bench-extra-args }}
+        run: |
+          set -euo pipefail
+          docker exec \
+            -e MODEL \
+            -e PORT \
+            -e NUM_PROMPTS \
+            -e MAX_CONCURRENCY \
+            -e BENCH_EXTRA_ARGS \
+            -e CONTAINER_WORK_DIR \
+            -e DATASET_JSONL \
+            "$CONTAINER_NAME" \
+            bash -lc '
+              set -euo pipefail
+              combined_log="$CONTAINER_WORK_DIR/vllm_bench_serve.log"
+              : > "$combined_log"
+
+              IFS="," read -ra concurrency_values <<< "$MAX_CONCURRENCY"
+              ran_any=false
+              for raw_concurrency in "${concurrency_values[@]}"; do
+                concurrency="${raw_concurrency//[[:space:]]/}"
+                if [[ -z "$concurrency" ]]; then
+                  continue
+                fi
+                if [[ ! "$concurrency" =~ ^[0-9]+$ ]]; then
+                  echo "Invalid max-concurrency value: $raw_concurrency" >&2
+                  exit 1
+                fi
+
+                log_file="$CONTAINER_WORK_DIR/vllm_bench_serve_conc${concurrency}.log"
+                ran_any=true
+                {
+                  echo "===== max-concurrency=${concurrency} ====="
+                  echo "dataset=$CONTAINER_WORK_DIR/$DATASET_JSONL"
+                  echo "num-prompts=$NUM_PROMPTS"
+                  echo ""
+                } > "$log_file"
+
+                if ! vllm bench serve \
+                  --backend openai \
+                  --base-url "http://127.0.0.1:${PORT}" \
+                  --endpoint /v1/completions \
+                  --model "$MODEL" \
+                  --dataset-name custom \
+                  --dataset-path "$CONTAINER_WORK_DIR/$DATASET_JSONL" \
+                  --num-prompts "$NUM_PROMPTS" \
+                  --max-concurrency "$concurrency" \
+                  --request-rate inf \
+                  --temperature 0.0 \
+                  --ignore-eos \
+                  --trust-remote-code \
+                  $BENCH_EXTRA_ARGS \
+                  >> "$log_file" 2>&1; then
+                  cat "$log_file" >> "$combined_log"
+                  exit 1
+                fi
+
+                cat "$log_file" >> "$combined_log"
+                echo "" >> "$combined_log"
+              done
+
+              if [[ "$ran_any" != "true" ]]; then
+                echo "No max-concurrency values were provided." >&2
+                exit 1
+              fi
+            '
+
+      - name: Capture vLLM metrics
+        env:
+          PORT: ${{ inputs.port }}
+        run: |
+          set -euo pipefail
+          curl --silent --fail "http://127.0.0.1:${PORT}/metrics" -o "$HOST_WORK_DIR/livecodebench_metrics.txt"
+
+      - name: Compute synthetic acceptance rates
+        id: compute-rates
+        env:
+          MODEL: ${{ inputs.model }}
+          DRAFT_MODEL: ${{ inputs.draft-model }}
+          NUM_SPECULATIVE_TOKENS: ${{ inputs.num-speculative-tokens }}
+        run: |
+          set -euo pipefail
+          python3 - <<'PY' | tee "$HOST_WORK_DIR/synthetic_acceptance_rates.log"
+          import json
+          import os
+          import re
+          from collections import defaultdict
+          from pathlib import Path
+
+          path = Path(os.environ["HOST_WORK_DIR"]) / "livecodebench_metrics.txt"
+
+          drafts_by_metric = defaultdict(float)
+          accepted_by_metric = defaultdict(lambda: defaultdict(float))
+
+          for line in path.read_text(encoding="utf-8").splitlines():
+              line = line.strip()
+              if not line or line.startswith("#"):
+                  continue
+
+              parts = line.split()
+              if len(parts) < 2:
+                  continue
+
+              series, value = parts[0], float(parts[1])
+              metric = series.split("{", 1)[0]
+
+              if metric in ("vllm:spec_decode_num_drafts_total", "vllm:spec_decode_num_drafts"):
+                  drafts_by_metric[metric] += value
+
+              if metric in (
+                  "vllm:spec_decode_num_accepted_tokens_per_pos_total",
+                  "vllm:spec_decode_num_accepted_tokens_per_pos",
+              ):
+                  match = re.search(r'position="(\d+)"', series)
+                  if match:
+                      accepted_by_metric[metric][int(match.group(1))] += value
+
+          draft_metric = (
+              "vllm:spec_decode_num_drafts_total"
+              if drafts_by_metric["vllm:spec_decode_num_drafts_total"] > 0
+              else "vllm:spec_decode_num_drafts"
+          )
+          accepted_metric = (
+              "vllm:spec_decode_num_accepted_tokens_per_pos_total"
+              if accepted_by_metric["vllm:spec_decode_num_accepted_tokens_per_pos_total"]
+              else "vllm:spec_decode_num_accepted_tokens_per_pos"
+          )
+
+          num_drafts = drafts_by_metric[draft_metric]
+          accepted = accepted_by_metric[accepted_metric]
+          num_speculative_tokens = int(os.environ["NUM_SPECULATIVE_TOKENS"])
+
+          if num_drafts <= 0:
+              raise SystemExit("No spec decode drafts found. Did you run standard speculative decoding first?")
+
+          rates = [accepted[i] / num_drafts for i in range(num_speculative_tokens)]
+          rounded_rates = [round(rate, 6) for rate in rates]
+          draft_model = os.environ["DRAFT_MODEL"]
+          if (
+              os.environ["MODEL"] == "MiniMaxAI/MiniMax-M2.5"
+              and draft_model == "lightseekorg/kimi-k2.5-eagle3"
+          ):
+              draft_model = "thoughtworks/MiniMax-M2.5-Eagle3"
+
+          speculative_config = {
+              "method": "eagle3",
+              "model": draft_model,
+              "num_speculative_tokens": num_speculative_tokens,
+              "rejection_sample_method": "synthetic",
+              "synthetic_acceptance_rates": rounded_rates,
+          }
+
+          out_dir = Path(os.environ["HOST_WORK_DIR"])
+          (out_dir / "synthetic_acceptance_rates.json").write_text(
+              json.dumps({
+                  "num_drafts": int(num_drafts),
+                  "accepted_per_position": {str(i): int(accepted[i]) for i in range(num_speculative_tokens)},
+                  "synthetic_acceptance_rates": rounded_rates,
+                  "speculative_config": speculative_config,
+              }, indent=2) + "\n",
+              encoding="utf-8",
+          )
+
+          print("num_drafts =", int(num_drafts))
+          print("accepted_per_position =", {i: int(accepted[i]) for i in range(num_speculative_tokens)})
+          print("synthetic_acceptance_rates =", json.dumps(rounded_rates))
+          print("speculative_config =", json.dumps(speculative_config))
+
+          with open(os.environ["GITHUB_OUTPUT"], "a", encoding="utf-8") as fh:
+              fh.write(f"synthetic_acceptance_rates={json.dumps(rounded_rates)}\n")
+              fh.write(f"speculative_config={json.dumps(speculative_config)}\n")
+          PY
+
+      - name: Add workflow summary
+        run: |
+          set -euo pipefail
+          {
+            echo "## Synthetic Acceptance Rates"
+            echo ""
+            echo '```text'
+            cat "$HOST_WORK_DIR/synthetic_acceptance_rates.log"
+            echo '```'
+          } >> "$GITHUB_STEP_SUMMARY"
+
+      - name: Upload acceptance-rate artifacts
+        if: always()
+        uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7.0.1
+        with:
+          name: mtp-fixed-ar-artifacts
+          path: |
+            ${{ env.HOST_WORK_DIR }}/${{ env.DATASET_JSONL }}
+            ${{ env.HOST_WORK_DIR }}/livecodebench_metrics.txt
+            ${{ env.HOST_WORK_DIR }}/synthetic_acceptance_rates.json
+            ${{ env.HOST_WORK_DIR }}/synthetic_acceptance_rates.log
+            ${{ env.HOST_WORK_DIR }}/vllm_server.log
+            ${{ env.HOST_WORK_DIR }}/vllm_bench_serve*.log
+          if-no-files-found: ignore
+
+      - name: Cleanup container
+        if: always()
+        run: |
+          docker rm -f "$CONTAINER_NAME" >/dev/null 2>&1 || true
diff --git a/benchmarks/single_node/kimik2.5_fp4_mi355x_vllm_eagle3.sh b/benchmarks/single_node/kimik2.5_fp4_mi355x_vllm_eagle3.sh
new file mode 100755
index 000000000..e5bf04e24
--- /dev/null
+++ b/benchmarks/single_node/kimik2.5_fp4_mi355x_vllm_eagle3.sh
@@ -0,0 +1,100 @@
+#!/usr/bin/env bash
+
+source "$(dirname "$0")/../benchmark_lib.sh"
+
+check_env_vars \
+    MODEL \
+    TP \
+    CONC \
+    ISL \
+    OSL \
+    MAX_MODEL_LEN \
+    RANDOM_RANGE_RATIO \
+    RESULT_FILENAME
+
+if [[ -n "$SLURM_JOB_ID" ]]; then
+  echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
+fi
+
+if [[ "$MODEL" != /* ]]; then hf download "$MODEL"; fi
+
+# Install amd-quark for MXFP4 quantization support.
+pip install amd-quark
+
+# Set HIP_VISIBLE_DEVICES to match ROCR_VISIBLE_DEVICES for Ray compatibility in vLLM 0.14+
+if [ -n "$ROCR_VISIBLE_DEVICES" ]; then
+    export HIP_VISIBLE_DEVICES="$ROCR_VISIBLE_DEVICES"
+fi
+
+SERVER_LOG=/workspace/server.log
+PORT=${PORT:-8888}
+
+if [ "${EVAL_ONLY}" = "true" ]; then
+    setup_eval_context
+    MAX_MODEL_LEN="$EVAL_MAX_MODEL_LEN"
+fi
+
+SPECULATIVE_CONFIG=$(python3 - <<'PY'
+import json
+import os
+
+config = {
+    "model": os.environ.get("DRAFT_MODEL", "lightseekorg/kimi-k2.5-eagle3"),
+    "method": "eagle3",
+    "num_speculative_tokens": int(os.environ.get("NUM_SPECULATIVE_TOKENS", "3")),
+    "rejection_sample_method": "standard",
+    "draft_tensor_parallel_size": int(os.environ.get("DRAFT_TENSOR_PARALLEL_SIZE", "1")),
+}
+print(json.dumps(config))
+PY
+)
+
+# Start GPU monitoring (power, temperature, clocks every second)
+start_gpu_monitor
+
+set -x
+export VLLM_ROCM_USE_AITER=1
+export VLLM_ROCM_QUICK_REDUCE_QUANTIZATION=INT4
+export VLLM_ROCM_USE_AITER_RMSNORM=0
+
+vllm serve "$MODEL" --port "$PORT" \
+--tensor-parallel-size="$TP" \
+--enable-expert-parallel \
+--gpu-memory-utilization "${GPU_MEMORY_UTILIZATION:-0.90}" \
+--max-model-len "$MAX_MODEL_LEN" \
+--max-num-seqs "$CONC" \
+--long-prefill-token-threshold 8192 \
+--max-num-batched-tokens 16384 \
+--mm-encoder-tp-mode data \
+--no-enable-prefix-caching \
+--speculative-config "$SPECULATIVE_CONFIG" \
+--trust-remote-code > "$SERVER_LOG" 2>&1 &
+
+SERVER_PID=$!
+
+# Wait for server to be ready
+wait_for_server_ready --port "$PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
+
+run_benchmark_serving \
+    --model "$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 "$RESULT_FILENAME" \
+    --result-dir /workspace/ \
+    --trust-remote-code \
+    --server-pid "$SERVER_PID"
+
+# After throughput, run evaluation only if RUN_EVAL is true
+if [ "${RUN_EVAL}" = "true" ]; then
+    run_eval --framework lm-eval --port "$PORT"
+    append_lm_eval_summary
+fi
+
+# Stop GPU monitoring
+stop_gpu_monitor
+set +x
diff --git a/benchmarks/single_node/kimik2.5_fp4_mi355x_vllm_mtp_fixed_AR.sh b/benchmarks/single_node/kimik2.5_fp4_mi355x_vllm_mtp_fixed_AR.sh
new file mode 100755
index 000000000..7472e2cd5
--- /dev/null
+++ b/benchmarks/single_node/kimik2.5_fp4_mi355x_vllm_mtp_fixed_AR.sh
@@ -0,0 +1,105 @@
+#!/usr/bin/env bash
+
+source "$(dirname "$0")/../benchmark_lib.sh"
+
+check_env_vars \
+    MODEL \
+    DRAFT_MODEL \
+    NUM_SPECULATIVE_TOKENS \
+    REJECTION_SAMPLE_METHOD \
+    SYNTHETIC_ACCEPTANCE_RATES \
+    TP \
+    CONC \
+    ISL \
+    OSL \
+    MAX_MODEL_LEN \
+    RANDOM_RANGE_RATIO \
+    RESULT_FILENAME
+
+if [[ -n "$SLURM_JOB_ID" ]]; then
+  echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
+fi
+
+if [[ "$MODEL" != /* ]]; then hf download "$MODEL"; fi
+
+# Install amd-quark for MXFP4 quantization support.
+pip install amd-quark
+
+# Set HIP_VISIBLE_DEVICES to match ROCR_VISIBLE_DEVICES for Ray compatibility in vLLM 0.14+
+if [ -n "$ROCR_VISIBLE_DEVICES" ]; then
+    export HIP_VISIBLE_DEVICES="$ROCR_VISIBLE_DEVICES"
+fi
+
+SERVER_LOG=/workspace/server.log
+PORT=${PORT:-8888}
+
+if [ "${EVAL_ONLY}" = "true" ]; then
+    setup_eval_context
+    MAX_MODEL_LEN="$EVAL_MAX_MODEL_LEN"
+fi
+
+SPECULATIVE_CONFIG=$(python3 - <<'PY'
+import json
+import os
+
+print(json.dumps({
+    "method": "eagle3",
+    "model": os.environ["DRAFT_MODEL"],
+    "num_speculative_tokens": int(os.environ["NUM_SPECULATIVE_TOKENS"]),
+    # Matrix/config fields use kebab-case; vLLM expects snake_case JSON.
+    "rejection_sample_method": os.environ["REJECTION_SAMPLE_METHOD"],
+    "synthetic_acceptance_rates": json.loads(os.environ["SYNTHETIC_ACCEPTANCE_RATES"]),
+    "draft_tensor_parallel_size": int(os.environ.get("DRAFT_TENSOR_PARALLEL_SIZE", "1")),
+}))
+PY
+)
+
+# Start GPU monitoring (power, temperature, clocks every second)
+start_gpu_monitor
+
+set -x
+export VLLM_ROCM_USE_AITER=1
+export VLLM_ROCM_QUICK_REDUCE_QUANTIZATION=INT4
+export VLLM_ROCM_USE_AITER_RMSNORM=0
+
+vllm serve "$MODEL" --port "$PORT" \
+--tensor-parallel-size="$TP" \
+--enable-expert-parallel \
+--gpu-memory-utilization "${GPU_MEMORY_UTILIZATION:-0.90}" \
+--max-model-len "$MAX_MODEL_LEN" \
+--max-num-seqs "$CONC" \
+--long-prefill-token-threshold 8192 \
+--max-num-batched-tokens 16384 \
+--mm-encoder-tp-mode data \
+--no-enable-prefix-caching \
+--speculative-config "$SPECULATIVE_CONFIG" \
+--trust-remote-code > "$SERVER_LOG" 2>&1 &
+
+SERVER_PID=$!
+
+# Wait for server to be ready
+wait_for_server_ready --port "$PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
+
+run_benchmark_serving \
+    --model "$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 "$RESULT_FILENAME" \
+    --result-dir /workspace/ \
+    --trust-remote-code \
+    --server-pid "$SERVER_PID"
+
+# After throughput, run evaluation only if RUN_EVAL is true
+if [ "${RUN_EVAL}" = "true" ]; then
+    run_eval --framework lm-eval --port "$PORT"
+    append_lm_eval_summary
+fi
+
+# Stop GPU monitoring
+stop_gpu_monitor
+set +x
diff --git a/benchmarks/single_node/kimik2.5_int4_mi355x_vllm_eagle3.sh b/benchmarks/single_node/kimik2.5_int4_mi355x_vllm_eagle3.sh
new file mode 100755
index 000000000..68345dca2
--- /dev/null
+++ b/benchmarks/single_node/kimik2.5_int4_mi355x_vllm_eagle3.sh
@@ -0,0 +1,97 @@
+#!/usr/bin/env bash
+
+source "$(dirname "$0")/../benchmark_lib.sh"
+
+check_env_vars \
+    MODEL \
+    TP \
+    CONC \
+    ISL \
+    OSL \
+    MAX_MODEL_LEN \
+    RANDOM_RANGE_RATIO \
+    RESULT_FILENAME
+
+if [[ -n "$SLURM_JOB_ID" ]]; then
+  echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
+fi
+
+if [[ "$MODEL" != /* ]]; then hf download "$MODEL"; fi
+
+# Set HIP_VISIBLE_DEVICES to match ROCR_VISIBLE_DEVICES for Ray compatibility in vLLM 0.14+
+if [ -n "$ROCR_VISIBLE_DEVICES" ]; then
+    export HIP_VISIBLE_DEVICES="$ROCR_VISIBLE_DEVICES"
+fi
+
+SERVER_LOG=/workspace/server.log
+PORT=${PORT:-8888}
+
+if [ "${EVAL_ONLY}" = "true" ]; then
+    setup_eval_context
+    MAX_MODEL_LEN="$EVAL_MAX_MODEL_LEN"
+fi
+
+SPECULATIVE_CONFIG=$(python3 - <<'PY'
+import json
+import os
+
+config = {
+    "model": os.environ.get("DRAFT_MODEL", "lightseekorg/kimi-k2.5-eagle3"),
+    "method": "eagle3",
+    "num_speculative_tokens": int(os.environ.get("NUM_SPECULATIVE_TOKENS", "3")),
+    "rejection_sample_method": "standard",
+    "draft_tensor_parallel_size": int(os.environ.get("DRAFT_TENSOR_PARALLEL_SIZE", "1")),
+}
+print(json.dumps(config))
+PY
+)
+
+# Start GPU monitoring (power, temperature, clocks every second)
+start_gpu_monitor
+
+set -x
+export VLLM_ROCM_USE_AITER=1
+export VLLM_ROCM_QUICK_REDUCE_QUANTIZATION=INT4
+export VLLM_ROCM_USE_AITER_RMSNORM=0
+
+vllm serve "$MODEL" --port "$PORT" \
+--tensor-parallel-size="$TP" \
+--enable-expert-parallel \
+--gpu-memory-utilization "${GPU_MEMORY_UTILIZATION:-0.90}" \
+--max-model-len "$MAX_MODEL_LEN" \
+--max-num-seqs "$CONC" \
+--long-prefill-token-threshold 8192 \
+--max-num-batched-tokens 16384 \
+--mm-encoder-tp-mode data \
+--no-enable-prefix-caching \
+--speculative-config "$SPECULATIVE_CONFIG" \
+--trust-remote-code > "$SERVER_LOG" 2>&1 &
+
+SERVER_PID=$!
+
+# Wait for server to be ready
+wait_for_server_ready --port "$PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
+
+run_benchmark_serving \
+    --model "$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 "$RESULT_FILENAME" \
+    --result-dir /workspace/ \
+    --trust-remote-code \
+    --server-pid "$SERVER_PID"
+
+# After throughput, run evaluation only if RUN_EVAL is true
+if [ "${RUN_EVAL}" = "true" ]; then
+    run_eval --framework lm-eval --port "$PORT"
+    append_lm_eval_summary
+fi
+
+# Stop GPU monitoring
+stop_gpu_monitor
+set +x
diff --git a/benchmarks/single_node/kimik2.5_int4_mi355x_vllm_mtp_fixed_AR.sh b/benchmarks/single_node/kimik2.5_int4_mi355x_vllm_mtp_fixed_AR.sh
new file mode 100755
index 000000000..8a782427d
--- /dev/null
+++ b/benchmarks/single_node/kimik2.5_int4_mi355x_vllm_mtp_fixed_AR.sh
@@ -0,0 +1,98 @@
+#!/usr/bin/env bash
+
+source "$(dirname "$0")/../benchmark_lib.sh"
+
+check_env_vars \
+    MODEL \
+    DRAFT_MODEL \
+    NUM_SPECULATIVE_TOKENS \
+    REJECTION_SAMPLE_METHOD \
+    SYNTHETIC_ACCEPTANCE_RATES \
+    TP \
+    CONC \
+    ISL \
+    OSL \
+    MAX_MODEL_LEN \
+    RANDOM_RANGE_RATIO \
+    RESULT_FILENAME
+
+if [[ -n "$SLURM_JOB_ID" ]]; then
+  echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
+fi
+
+if [[ "$MODEL" != /* ]]; then hf download "$MODEL"; fi
+
+# Set HIP_VISIBLE_DEVICES to match ROCR_VISIBLE_DEVICES for Ray compatibility in vLLM 0.14+
+if [ -n "$ROCR_VISIBLE_DEVICES" ]; then
+    export HIP_VISIBLE_DEVICES="$ROCR_VISIBLE_DEVICES"
+fi
+
+SERVER_LOG=/workspace/server.log
+PORT=${PORT:-8888}
+
+if [ "${EVAL_ONLY}" = "true" ]; then
+    setup_eval_context
+    MAX_MODEL_LEN="$EVAL_MAX_MODEL_LEN"
+fi
+
+SPECULATIVE_CONFIG=$(python3 - <<'PY'
+import json
+import os
+
+print(json.dumps({
+    "method": "eagle3",
+    "model": os.environ["DRAFT_MODEL"],
+    "num_speculative_tokens": int(os.environ["NUM_SPECULATIVE_TOKENS"]),
+    # Matrix/config fields use kebab-case; vLLM expects snake_case JSON.
+    "rejection_sample_method": os.environ["REJECTION_SAMPLE_METHOD"],
+    "synthetic_acceptance_rates": json.loads(os.environ["SYNTHETIC_ACCEPTANCE_RATES"]),
+}))
+PY
+)
+
+# Start GPU monitoring (power, temperature, clocks every second)
+start_gpu_monitor
+
+set -x
+export VLLM_ROCM_USE_AITER=1
+export VLLM_ROCM_QUICK_REDUCE_QUANTIZATION=INT4
+export VLLM_ROCM_USE_AITER_RMSNORM=0
+
+vllm serve "$MODEL" --port "$PORT" \
+--tensor-parallel-size="$TP" \
+--gpu-memory-utilization "${GPU_MEMORY_UTILIZATION:-0.90}" \
+--max-model-len "$MAX_MODEL_LEN" \
+--trust-remote-code \
+--no-enable-prefix-caching \
+--max-num-seqs "$CONC" \
+--mm-encoder-tp-mode data \
+--speculative-config "$SPECULATIVE_CONFIG" > "$SERVER_LOG" 2>&1 &
+
+SERVER_PID=$!
+
+# Wait for server to be ready
+wait_for_server_ready --port "$PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
+
+run_benchmark_serving \
+    --model "$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 "$RESULT_FILENAME" \
+    --result-dir /workspace/ \
+    --trust-remote-code \
+    --server-pid "$SERVER_PID"
+
+# After throughput, run evaluation only if RUN_EVAL is true
+if [ "${RUN_EVAL}" = "true" ]; then
+    run_eval --framework lm-eval --port "$PORT"
+    append_lm_eval_summary
+fi
+
+# Stop GPU monitoring
+stop_gpu_monitor
+set +x
diff --git a/benchmarks/single_node/minimaxm2.5_fp8_mi355x_vllm_eagle3.sh b/benchmarks/single_node/minimaxm2.5_fp8_mi355x_vllm_eagle3.sh
new file mode 100755
index 000000000..9c5c6fb51
--- /dev/null
+++ b/benchmarks/single_node/minimaxm2.5_fp8_mi355x_vllm_eagle3.sh
@@ -0,0 +1,109 @@
+#!/usr/bin/env bash
+
+source "$(dirname "$0")/../benchmark_lib.sh"
+
+check_env_vars \
+    MODEL \
+    TP \
+    CONC \
+    ISL \
+    OSL \
+    MAX_MODEL_LEN \
+    RANDOM_RANGE_RATIO \
+    RESULT_FILENAME
+
+if [[ -n "$SLURM_JOB_ID" ]]; then
+  echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
+fi
+
+if [[ "$MODEL" != /* ]]; then hf download "$MODEL"; fi
+
+# Set HIP_VISIBLE_DEVICES to match ROCR_VISIBLE_DEVICES for Ray compatibility in vLLM 0.14+
+if [ -n "$ROCR_VISIBLE_DEVICES" ]; then
+    export HIP_VISIBLE_DEVICES="$ROCR_VISIBLE_DEVICES"
+fi
+
+SERVER_LOG=/workspace/server.log
+PORT=${PORT:-8888}
+ASYNC_SCHEDULING_ARGS=""
+VLLM_BLOCK_SIZE=32
+
+if [ "${EVAL_ONLY}" = "true" ]; then
+    setup_eval_context
+    MAX_MODEL_LEN="$EVAL_MAX_MODEL_LEN"
+fi
+
+SPECULATIVE_CONFIG=$(python3 - <<'PY'
+import json
+import os
+
+config = {
+    "model": os.environ.get("DRAFT_MODEL", "thoughtworks/MiniMax-M2.5-Eagle3"),
+    "method": "eagle3",
+    "num_speculative_tokens": int(os.environ.get("NUM_SPECULATIVE_TOKENS", "3")),
+    "rejection_sample_method": "standard",
+    "draft_tensor_parallel_size": int(os.environ.get("DRAFT_TENSOR_PARALLEL_SIZE", "1")),
+}
+print(json.dumps(config))
+PY
+)
+
+# Start GPU monitoring (power, temperature, clocks every second)
+start_gpu_monitor
+
+set -x
+export VLLM_ROCM_USE_AITER=1
+export VLLM_ROCM_QUICK_REDUCE_QUANTIZATION=INT4
+
+if [[ "$ISL" == "8192" && "$OSL" == "1024" && "$TP" == "8" ]]; then
+    export VLLM_ROCM_USE_AITER_MOE=0
+    ASYNC_SCHEDULING_ARGS="--no-async-scheduling"
+    echo "8k1k TP8: disabling AITER MoE and async scheduling to avoid ROCm resource exhaustion."
+    if (( CONC >= 64 )); then
+        VLLM_BLOCK_SIZE=16
+        echo "8k1k TP8 c${CONC}: using block size 16."
+    fi
+fi
+
+vllm serve "$MODEL" --port "$PORT" \
+--tensor-parallel-size="$TP" \
+--enable-expert-parallel \
+--gpu-memory-utilization "${GPU_MEMORY_UTILIZATION:-0.90}" \
+--max-model-len "$MAX_MODEL_LEN" \
+--max-num-seqs "$CONC" \
+--kv-cache-dtype fp8 \
+--block-size="$VLLM_BLOCK_SIZE" \
+--no-enable-prefix-caching \
+--attention-backend "ROCM_AITER_FA" \
+$ASYNC_SCHEDULING_ARGS \
+--speculative-config "$SPECULATIVE_CONFIG" \
+--trust-remote-code > "$SERVER_LOG" 2>&1 &
+
+SERVER_PID=$!
+
+# Wait for server to be ready
+wait_for_server_ready --port "$PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
+
+run_benchmark_serving \
+    --model "$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 "$RESULT_FILENAME" \
+    --result-dir /workspace/ \
+    --trust-remote-code \
+    --server-pid "$SERVER_PID"
+
+# After throughput, run evaluation only if RUN_EVAL is true
+if [ "${RUN_EVAL}" = "true" ]; then
+    run_eval --framework lm-eval --port "$PORT"
+    append_lm_eval_summary
+fi
+
+# Stop GPU monitoring
+stop_gpu_monitor
+set +x
diff --git a/benchmarks/single_node/minimaxm2.5_fp8_mi355x_vllm_eagle3_fixed_AR.sh b/benchmarks/single_node/minimaxm2.5_fp8_mi355x_vllm_eagle3_fixed_AR.sh
new file mode 100755
index 000000000..ed585f239
--- /dev/null
+++ b/benchmarks/single_node/minimaxm2.5_fp8_mi355x_vllm_eagle3_fixed_AR.sh
@@ -0,0 +1,114 @@
+#!/usr/bin/env bash
+
+source "$(dirname "$0")/../benchmark_lib.sh"
+
+check_env_vars \
+    MODEL \
+    DRAFT_MODEL \
+    NUM_SPECULATIVE_TOKENS \
+    REJECTION_SAMPLE_METHOD \
+    SYNTHETIC_ACCEPTANCE_RATES \
+    TP \
+    CONC \
+    ISL \
+    OSL \
+    MAX_MODEL_LEN \
+    RANDOM_RANGE_RATIO \
+    RESULT_FILENAME
+
+if [[ -n "$SLURM_JOB_ID" ]]; then
+  echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
+fi
+
+if [[ "$MODEL" != /* ]]; then hf download "$MODEL"; fi
+
+# Set HIP_VISIBLE_DEVICES to match ROCR_VISIBLE_DEVICES for Ray compatibility in vLLM 0.14+
+if [ -n "$ROCR_VISIBLE_DEVICES" ]; then
+    export HIP_VISIBLE_DEVICES="$ROCR_VISIBLE_DEVICES"
+fi
+
+SERVER_LOG=/workspace/server.log
+PORT=${PORT:-8888}
+ASYNC_SCHEDULING_ARGS=""
+VLLM_BLOCK_SIZE=32
+
+if [ "${EVAL_ONLY}" = "true" ]; then
+    setup_eval_context
+    MAX_MODEL_LEN="$EVAL_MAX_MODEL_LEN"
+fi
+
+SPECULATIVE_CONFIG=$(python3 - <<'PY'
+import json
+import os
+
+print(json.dumps({
+    "model": os.environ["DRAFT_MODEL"],
+    "method": "eagle3",
+    "num_speculative_tokens": int(os.environ["NUM_SPECULATIVE_TOKENS"]),
+    # Matrix/config fields use kebab-case; vLLM expects snake_case JSON.
+    "rejection_sample_method": os.environ["REJECTION_SAMPLE_METHOD"],
+    "synthetic_acceptance_rates": json.loads(os.environ["SYNTHETIC_ACCEPTANCE_RATES"]),
+    "draft_tensor_parallel_size": int(os.environ.get("DRAFT_TENSOR_PARALLEL_SIZE", "1")),
+}))
+PY
+)
+
+# Start GPU monitoring (power, temperature, clocks every second)
+start_gpu_monitor
+
+set -x
+export VLLM_ROCM_USE_AITER=1
+export VLLM_ROCM_QUICK_REDUCE_QUANTIZATION=INT4
+
+if [[ "$ISL" == "8192" && "$OSL" == "1024" && "$TP" == "8" ]]; then
+    export VLLM_ROCM_USE_AITER_MOE=0
+    ASYNC_SCHEDULING_ARGS="--no-async-scheduling"
+    echo "8k1k TP8: disabling AITER MoE and async scheduling to avoid ROCm resource exhaustion."
+    if (( CONC >= 64 )); then
+        VLLM_BLOCK_SIZE=16
+        echo "8k1k TP8 c${CONC}: using block size 16."
+    fi
+fi
+
+vllm serve "$MODEL" --port "$PORT" \
+--tensor-parallel-size="$TP" \
+--enable-expert-parallel \
+--gpu-memory-utilization "${GPU_MEMORY_UTILIZATION:-0.90}" \
+--max-model-len "$MAX_MODEL_LEN" \
+--max-num-seqs "$CONC" \
+--kv-cache-dtype fp8 \
+--block-size="$VLLM_BLOCK_SIZE" \
+--no-enable-prefix-caching \
+--attention-backend "ROCM_AITER_FA" \
+$ASYNC_SCHEDULING_ARGS \
+--speculative-config "$SPECULATIVE_CONFIG" \
+--trust-remote-code > "$SERVER_LOG" 2>&1 &
+
+SERVER_PID=$!
+
+# Wait for server to be ready
+wait_for_server_ready --port "$PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
+
+run_benchmark_serving \
+    --model "$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 "$RESULT_FILENAME" \
+    --result-dir /workspace/ \
+    --trust-remote-code \
+    --server-pid "$SERVER_PID"
+
+# After throughput, run evaluation only if RUN_EVAL is true
+if [ "${RUN_EVAL}" = "true" ]; then
+    run_eval --framework lm-eval --port "$PORT"
+    append_lm_eval_summary
+fi
+
+# Stop GPU monitoring
+stop_gpu_monitor
+set +x
diff --git a/runners/launch_mi355x-amds.sh b/runners/launch_mi355x-amds.sh
index 789e363cb..9dbbc50c6 100644
--- a/runners/launch_mi355x-amds.sh
+++ b/runners/launch_mi355x-amds.sh
@@ -180,7 +180,17 @@ else
     export PORT_OFFSET=${RUNNER_NAME: -1}
     export PORT=$(( 8888 + ${PORT_OFFSET} ))
     FRAMEWORK_SUFFIX=$([[ "$FRAMEWORK" == "atom" ]] && printf '_atom' || printf '')
-    SPEC_SUFFIX=$([[ "$SPEC_DECODING" == "mtp" ]] && printf '_mtp' || printf '')
+    case "$SPEC_DECODING" in
+        mtp)
+            SPEC_SUFFIX="_mtp"
+            ;;
+        eagle3)
+            SPEC_SUFFIX="_eagle3"
+            ;;
+        *)
+            SPEC_SUFFIX=""
+            ;;
+    esac
 
     PARTITION="compute"
     SQUASH_FILE="/var/lib/squash/$(echo "$IMAGE" | sed 's/[\/:@#]/_/g').sqsh"
@@ -224,8 +234,14 @@ else
 
     SCRIPT_BASE="${EXP_NAME%%_*}_${PRECISION}_mi355x"
     SCRIPT_FW="benchmarks/single_node/${SCENARIO_SUBDIR:-}${SCRIPT_BASE}_${FRAMEWORK}${SPEC_SUFFIX}.sh"
+    SCRIPT_FIXED_AR="benchmarks/single_node/${SCENARIO_SUBDIR:-}${SCRIPT_BASE}_${FRAMEWORK}${SPEC_SUFFIX}_fixed_AR.sh"
+    SCRIPT_FIXED_AR_MTP="benchmarks/single_node/${SCENARIO_SUBDIR:-}${SCRIPT_BASE}_${FRAMEWORK}_mtp_fixed_AR.sh"
     SCRIPT_FALLBACK="benchmarks/single_node/${SCENARIO_SUBDIR:-}${SCRIPT_BASE}${FRAMEWORK_SUFFIX}${SPEC_SUFFIX}.sh"
-    if [[ -f "$SCRIPT_FW" ]]; then
+    if [[ "$SCENARIO_TYPE" == "fixed-ar-mtp" && -f "$SCRIPT_FIXED_AR" ]]; then
+        BENCHMARK_SCRIPT="$SCRIPT_FIXED_AR"
+    elif [[ "$SCENARIO_TYPE" == "fixed-ar-mtp" && -f "$SCRIPT_FIXED_AR_MTP" ]]; then
+        BENCHMARK_SCRIPT="$SCRIPT_FIXED_AR_MTP"
+    elif [[ -f "$SCRIPT_FW" ]]; then
         BENCHMARK_SCRIPT="$SCRIPT_FW"
     else
         BENCHMARK_SCRIPT="$SCRIPT_FALLBACK"
diff --git a/utils/matrix_logic/generate_sweep_configs.py b/utils/matrix_logic/generate_sweep_configs.py
index 9f38292f4..a43d2dc1a 100644
--- a/utils/matrix_logic/generate_sweep_configs.py
+++ b/utils/matrix_logic/generate_sweep_configs.py
@@ -26,6 +26,24 @@
 seq_len_itos = {v: k for k, v in seq_len_stoi.items()}
 
 
+def parse_concurrency_filter(values):
+    if values is None:
+        return None
+
+    conc_values = []
+    for value in values:
+        for part in str(value).split(","):
+            part = part.strip()
+            if not part:
+                continue
+            try:
+                conc_values.append(int(part))
+            except ValueError as exc:
+                raise argparse.ArgumentTypeError(
+                    f"invalid concurrency value: {part!r}") from exc
+    return conc_values
+
+
 def seq_len_to_str(isl: int, osl: int) -> str:
     """Convert sequence lengths to short string representation.
 
@@ -124,7 +142,7 @@ def _max_eval_conc(ie):
 
     # Mark the selected entries (skip agentic entries which don't support evals)
     for i, entry in enumerate(matrix_values):
-        if entry.get(Fields.SCENARIO_TYPE.value) == 'agentic-coding':
+        if entry.get(Fields.SCENARIO_TYPE.value) in ('agentic-coding', 'fixed-ar-mtp'):
             continue
         entry[Fields.RUN_EVAL.value] = i in eval_indices
         if i in mn_eval_conc:
@@ -187,6 +205,7 @@ def generate_full_sweep(args, all_config_data, runner_data):
         scenarios = val[Fields.SCENARIOS.value]
         scenario_filter = set(args.scenario_type) if getattr(args, 'scenario_type', None) else None
         seq_len_configs = scenarios.get(Fields.FIXED_SEQ_LEN.value, []) if (scenario_filter is None or 'fixed-seq-len' in scenario_filter) else []
+        fixed_ar_mtp_configs = scenarios.get(Fields.FIXED_AR_MTP.value, []) if (scenario_filter is None or 'fixed-ar-mtp' in scenario_filter) else []
         image = val[Fields.IMAGE.value]
         model = val[Fields.MODEL.value]
         precision = val[Fields.PRECISION.value]
@@ -378,6 +397,97 @@ def generate_full_sweep(args, all_config_data, runner_data):
                         if conc > conc_end:
                             conc = conc_end
 
+        # ---- Fixed-AR MTP throughput scenarios ----
+        if not is_multinode:
+            for fixed_ar_config in fixed_ar_mtp_configs:
+                isl = fixed_ar_config[Fields.ISL.value]
+                osl = fixed_ar_config[Fields.OSL.value]
+
+                if seq_lens_filter and (isl, osl) not in seq_lens_filter:
+                    continue
+
+                bmk_space = fixed_ar_config[Fields.SEARCH_SPACE.value]
+                seq_len_str = seq_len_to_str(isl, osl)
+                runners_for_entry = runner_nodes_to_use if runner_nodes_to_use else [runner]
+
+                for bmk in bmk_space:
+                    tp = bmk[Fields.TP.value]
+                    ep = bmk.get(Fields.EP.value)
+                    dp_attn = bmk.get(Fields.DP_ATTN.value)
+                    spec_decoding = bmk.get(Fields.SPEC_DECODING.value, "mtp")
+
+                    if Fields.CONC_LIST.value in bmk:
+                        conc_values = bmk[Fields.CONC_LIST.value]
+                    else:
+                        conc_start = bmk[Fields.CONC_START.value]
+                        conc_end = bmk[Fields.CONC_END.value]
+                        conc_values = []
+                        conc = conc_start
+                        while conc <= conc_end:
+                            conc_values.append(conc)
+                            if conc == conc_end:
+                                break
+                            conc *= args.step_size
+                            if conc > conc_end:
+                                conc = conc_end
+
+                    if args.max_tp is not None:
+                        if args.max_tp <= 0:
+                            continue
+                        if tp > args.max_tp:
+                            continue
+
+                    if args.max_ep is not None:
+                        if args.max_ep <= 0:
+                            continue
+                        if ep is not None and ep > args.max_ep:
+                            ep = args.max_ep
+
+                    if args.min_conc is not None:
+                        if args.min_conc <= 0:
+                            continue
+                        conc_values = [c for c in conc_values if c >= args.min_conc]
+                        if not conc_values:
+                            continue
+
+                    if args.max_conc is not None:
+                        if args.max_conc <= 0:
+                            continue
+                        filtered_conc = [c for c in conc_values if c <= args.max_conc]
+                        if not filtered_conc:
+                            conc_values = [args.max_conc]
+                        else:
+                            conc_values = filtered_conc
+
+                    for conc in conc_values:
+                        for runner_value in runners_for_entry:
+                            entry = {
+                                Fields.IMAGE.value: image,
+                                Fields.MODEL.value: model,
+                                Fields.MODEL_PREFIX.value: model_code,
+                                Fields.PRECISION.value: precision,
+                                Fields.FRAMEWORK.value: framework,
+                                Fields.RUNNER.value: runner_value,
+                                Fields.ISL.value: isl,
+                                Fields.OSL.value: osl,
+                                Fields.TP.value: tp,
+                                Fields.CONC.value: conc,
+                                Fields.MAX_MODEL_LEN.value: isl + osl + 256,
+                                Fields.EP.value: ep if ep is not None else 1,
+                                Fields.DP_ATTN.value: dp_attn if dp_attn is not None else False,
+                                Fields.SPEC_DECODING.value: spec_decoding,
+                                Fields.EXP_NAME.value: f"{model_code}_{seq_len_str}",
+                                Fields.DISAGG.value: disagg,
+                                Fields.RUN_EVAL.value: False,
+                                Fields.SCENARIO_TYPE.value: "fixed-ar-mtp",
+                                Fields.DRAFT_MODEL.value: fixed_ar_config[Fields.DRAFT_MODEL.value],
+                                Fields.NUM_SPECULATIVE_TOKENS.value: fixed_ar_config[Fields.NUM_SPECULATIVE_TOKENS.value],
+                                Fields.REJECTION_SAMPLE_METHOD.value: fixed_ar_config[Fields.REJECTION_SAMPLE_METHOD.value],
+                                Fields.SYNTHETIC_ACCEPTANCE_RATES.value: fixed_ar_config[Fields.SYNTHETIC_ACCEPTANCE_RATES.value],
+                            }
+                            validate_matrix_entry(entry, is_multinode=False)
+                            matrix_values.append(entry)
+
         # ---- Agentic-coding scenarios ----
         agentic_configs = scenarios.get(Fields.AGENTIC_CODING.value, []) if (scenario_filter is None or 'agentic-coding' in scenario_filter) else []
 
@@ -682,6 +792,7 @@ def generate_test_config_sweep(args, all_config_data, runner_data=None):
 
         scenario_filter = set(args.scenario_type) if getattr(args, 'scenario_type', None) else None
         fixed_configs = val[Fields.SCENARIOS.value].get(Fields.FIXED_SEQ_LEN.value, []) if (scenario_filter is None or 'fixed-seq-len' in scenario_filter) else []
+        fixed_ar_mtp_configs = val[Fields.SCENARIOS.value].get(Fields.FIXED_AR_MTP.value, []) if (scenario_filter is None or 'fixed-ar-mtp' in scenario_filter) else []
         for seq_len_config in fixed_configs:
             isl = seq_len_config[Fields.ISL.value]
             osl = seq_len_config[Fields.OSL.value]
@@ -794,6 +905,71 @@ def generate_test_config_sweep(args, all_config_data, runner_data=None):
                             }
                             matrix_values.append(validate_matrix_entry(entry, is_multinode=False))
 
+        # ---- Fixed-AR MTP throughput scenarios ----
+        if not is_multinode:
+            for fixed_ar_config in fixed_ar_mtp_configs:
+                isl = fixed_ar_config[Fields.ISL.value]
+                osl = fixed_ar_config[Fields.OSL.value]
+
+                if seq_lens_filter and (isl, osl) not in seq_lens_filter:
+                    continue
+
+                seq_len_str = seq_len_to_str(isl, osl)
+
+                for bmk in fixed_ar_config[Fields.SEARCH_SPACE.value]:
+                    tp = bmk[Fields.TP.value]
+                    ep = bmk.get(Fields.EP.value)
+                    dp_attn = bmk.get(Fields.DP_ATTN.value)
+                    spec_decoding = bmk.get(Fields.SPEC_DECODING.value, "mtp")
+
+                    if Fields.CONC_LIST.value in bmk:
+                        conc_values = bmk[Fields.CONC_LIST.value]
+                    else:
+                        conc_start = bmk[Fields.CONC_START.value]
+                        conc_end = bmk[Fields.CONC_END.value]
+                        conc_values = []
+                        conc = conc_start
+                        while conc <= conc_end:
+                            conc_values.append(conc)
+                            if conc == conc_end:
+                                break
+                            conc *= 2
+                            if conc > conc_end:
+                                conc = conc_end
+
+                    if getattr(args, 'conc', None):
+                        conc_values = [c for c in conc_values if c in args.conc]
+                        if not conc_values:
+                            continue
+
+                    for conc in conc_values:
+                        for runner_value in runners_for_entry:
+                            entry = {
+                                Fields.IMAGE.value: image,
+                                Fields.MODEL.value: model,
+                                Fields.MODEL_PREFIX.value: model_code,
+                                Fields.PRECISION.value: precision,
+                                Fields.FRAMEWORK.value: framework,
+                                Fields.RUNNER.value: runner_value,
+                                Fields.ISL.value: isl,
+                                Fields.OSL.value: osl,
+                                Fields.TP.value: tp,
+                                Fields.CONC.value: conc,
+                                Fields.MAX_MODEL_LEN.value: isl + osl + 256,
+                                Fields.EP.value: ep if ep is not None else 1,
+                                Fields.DP_ATTN.value: dp_attn if dp_attn is not None else False,
+                                Fields.SPEC_DECODING.value: spec_decoding,
+                                Fields.EXP_NAME.value: f"{model_code}_{seq_len_str}",
+                                Fields.DISAGG.value: disagg,
+                                Fields.RUN_EVAL.value: False,
+                                Fields.SCENARIO_TYPE.value: "fixed-ar-mtp",
+                                Fields.DRAFT_MODEL.value: fixed_ar_config[Fields.DRAFT_MODEL.value],
+                                Fields.NUM_SPECULATIVE_TOKENS.value: fixed_ar_config[Fields.NUM_SPECULATIVE_TOKENS.value],
+                                Fields.REJECTION_SAMPLE_METHOD.value: fixed_ar_config[Fields.REJECTION_SAMPLE_METHOD.value],
+                                Fields.SYNTHETIC_ACCEPTANCE_RATES.value: fixed_ar_config[Fields.SYNTHETIC_ACCEPTANCE_RATES.value],
+                            }
+                            matrix_values.append(validate_matrix_entry(entry, is_multinode=False))
+
         # ---- Agentic-coding scenarios ----
         agentic_configs = val[Fields.SCENARIOS.value].get(Fields.AGENTIC_CODING.value, []) if (scenario_filter is None or 'agentic-coding' in scenario_filter) else []
         for agentic_config in agentic_configs:
@@ -947,7 +1123,7 @@ def main():
     parent_parser.add_argument(
         '--scenario-type',
         nargs='+',
-        choices=['fixed-seq-len', 'agentic-coding'],
+        choices=['fixed-seq-len', 'agentic-coding', 'fixed-ar-mtp'],
         required=False,
         help='Scenario type(s) to include. If not specified, all scenario types are generated.'
     )
@@ -1116,9 +1292,8 @@ def main():
     test_config_keys_parser.add_argument(
         '--conc',
         nargs='+',
-        type=int,
         required=False,
-        help='Only include these concurrency values. Values must exist in the config conc-range/list.'
+        help='Only include these concurrency values. Supports space-separated or comma-separated values, e.g. 32 64 or 32,64. Values must exist in the config conc-range/list.'
     )
     test_config_keys_parser.add_argument(
         '--seq-lens',
@@ -1134,6 +1309,8 @@ def main():
     )
 
     args = parser.parse_args()
+    if args.command == 'test-config':
+        args.conc = parse_concurrency_filter(args.conc)
     apply_node_type_defaults(args)
 
     # Load and validate configuration files (validation happens by default in load functions)
diff --git a/utils/matrix_logic/validation.py b/utils/matrix_logic/validation.py
index dd245aec7..53f30e2fb 100644
--- a/utils/matrix_logic/validation.py
+++ b/utils/matrix_logic/validation.py
@@ -26,6 +26,7 @@ class Fields(Enum):
     # Scenario type keys
     FIXED_SEQ_LEN = 'fixed-seq-len'
     AGENTIC_CODING = 'agentic-coding'
+    FIXED_AR_MTP = 'fixed-ar-mtp'
 
     # Seq-len-config fields
     ISL = 'isl'
@@ -53,6 +54,12 @@ class Fields(Enum):
     OFFLOADING = 'offloading'
     DURATION = 'duration'
 
+    # Fixed acceptance-rate MTP fields
+    DRAFT_MODEL = 'draft-model'
+    NUM_SPECULATIVE_TOKENS = 'num-speculative-tokens'
+    REJECTION_SAMPLE_METHOD = 'rejection-sample-method'
+    SYNTHETIC_ACCEPTANCE_RATES = 'synthetic-acceptance-rates'
+
     # Matrix entry fields
     CONC = 'conc'
     MAX_MODEL_LEN = 'max-model-len'
@@ -86,7 +93,7 @@ class SingleNodeMatrixEntry(BaseModel):
     model_prefix: str = Field(alias=Fields.MODEL_PREFIX.value)
     precision: str
     framework: str
-    spec_decoding: Literal["mtp", "draft_model", "none"] = Field(
+    spec_decoding: Literal["mtp", "eagle3", "draft_model", "none"] = Field(
         alias=Fields.SPEC_DECODING.value
     )
     runner: str
@@ -103,6 +110,28 @@ class SingleNodeMatrixEntry(BaseModel):
     eval_only: bool = Field(alias=Fields.EVAL_ONLY.value, default=False)
 
 
+class SingleNodeFixedArMtpMatrixEntry(SingleNodeMatrixEntry):
+    """Single-node throughput entry with synthetic fixed acceptance rates for MTP."""
+    model_config = ConfigDict(extra='forbid', populate_by_name=True)
+
+    scenario_type: Literal["fixed-ar-mtp"] = Field(alias=Fields.SCENARIO_TYPE.value)
+    draft_model: str = Field(alias=Fields.DRAFT_MODEL.value)
+    num_speculative_tokens: int = Field(alias=Fields.NUM_SPECULATIVE_TOKENS.value)
+    rejection_sample_method: Literal["synthetic"] = Field(alias=Fields.REJECTION_SAMPLE_METHOD.value)
+    synthetic_acceptance_rates: List[float] = Field(alias=Fields.SYNTHETIC_ACCEPTANCE_RATES.value)
+
+    @model_validator(mode='after')
+    def validate_synthetic_acceptance_rates(self):
+        if self.spec_decoding != "eagle3":
+            raise ValueError("fixed-ar-mtp entries must use spec-decoding: eagle3")
+        if len(self.synthetic_acceptance_rates) != self.num_speculative_tokens:
+            raise ValueError(
+                f"'{Fields.SYNTHETIC_ACCEPTANCE_RATES.value}' length must match "
+                f"'{Fields.NUM_SPECULATIVE_TOKENS.value}'"
+            )
+        return self
+
+
 class WorkerConfig(BaseModel):
     """Pydantic model for validating worker configuration in multinode entries."""
     model_config = ConfigDict(extra='forbid', populate_by_name=True)
@@ -125,7 +154,7 @@ class MultiNodeMatrixEntry(BaseModel):
     model_prefix: str = Field(alias=Fields.MODEL_PREFIX.value)
     precision: str
     framework: str
-    spec_decoding: Literal["mtp", "draft_model", "none"] = Field(
+    spec_decoding: Literal["mtp", "eagle3", "draft_model", "none"] = Field(
         alias=Fields.SPEC_DECODING.value
     )
     runner: str
@@ -171,7 +200,7 @@ class MultiNodeAgenticMatrixEntry(BaseModel):
     model_prefix: str = Field(alias=Fields.MODEL_PREFIX.value)
     precision: str
     framework: str
-    spec_decoding: Literal["mtp", "draft_model", "none"] = Field(
+    spec_decoding: Literal["mtp", "eagle3", "draft_model", "none"] = Field(
         alias=Fields.SPEC_DECODING.value
     )
     runner: str
@@ -207,7 +236,11 @@ def validate_matrix_entry(entry: dict, is_multinode: bool) -> dict:
     Returns the original list if all entries are valid.
     """
     try:
-        if is_multinode:
+        if entry.get(Fields.SCENARIO_TYPE.value) == Fields.FIXED_AR_MTP.value:
+            if is_multinode:
+                raise ValueError("fixed-ar-mtp is only supported for single-node entries")
+            SingleNodeFixedArMtpMatrixEntry(**entry)
+        elif is_multinode:
             MultiNodeMatrixEntry(**entry)
         else:
             SingleNodeMatrixEntry(**entry)
@@ -271,7 +304,7 @@ class SingleNodeSearchSpaceEntry(BaseModel):
 
     tp: int
     ep: Optional[int] = None
-    spec_decoding: Literal["mtp", "draft_model", "none"] = Field(
+    spec_decoding: Literal["mtp", "eagle3", "draft_model", "none"] = Field(
         default="none", alias=Fields.SPEC_DECODING.value)
     dp_attn: Optional[bool] = Field(
         default=None, alias=Fields.DP_ATTN.value)
@@ -291,7 +324,7 @@ class MultiNodeSearchSpaceEntry(BaseModel):
     """Multinode search space configuration."""
     model_config = ConfigDict(extra='forbid', populate_by_name=True)
 
-    spec_decoding: Literal["mtp", "draft_model", "none"] = Field(
+    spec_decoding: Literal["mtp", "eagle3", "draft_model", "none"] = Field(
         default="none", alias=Fields.SPEC_DECODING.value)
     prefill: WorkerConfig
     decode: WorkerConfig
@@ -317,6 +350,28 @@ class SingleNodeSeqLenConfig(BaseModel):
         alias=Fields.SEARCH_SPACE.value)
 
 
+class FixedArMtpSeqLenConfig(BaseModel):
+    """Single-node fixed-AR MTP scenario configuration."""
+    model_config = ConfigDict(extra='forbid', populate_by_name=True)
+
+    isl: int
+    osl: int
+    draft_model: str = Field(alias=Fields.DRAFT_MODEL.value)
+    num_speculative_tokens: int = Field(alias=Fields.NUM_SPECULATIVE_TOKENS.value)
+    rejection_sample_method: Literal["synthetic"] = Field(alias=Fields.REJECTION_SAMPLE_METHOD.value)
+    synthetic_acceptance_rates: List[float] = Field(alias=Fields.SYNTHETIC_ACCEPTANCE_RATES.value)
+    search_space: List[SingleNodeSearchSpaceEntry] = Field(alias=Fields.SEARCH_SPACE.value)
+
+    @model_validator(mode='after')
+    def validate_synthetic_acceptance_rates(self):
+        if len(self.synthetic_acceptance_rates) != self.num_speculative_tokens:
+            raise ValueError(
+                f"'{Fields.SYNTHETIC_ACCEPTANCE_RATES.value}' length must match "
+                f"'{Fields.NUM_SPECULATIVE_TOKENS.value}'"
+            )
+        return self
+
+
 class MultiNodeSeqLenConfig(BaseModel):
     """Multinode sequence length configuration."""
     model_config = ConfigDict(extra='forbid', populate_by_name=True)
@@ -334,7 +389,7 @@ class AgenticCodingSearchSpaceEntry(BaseModel):
     tp: Optional[int] = None
     ep: Optional[int] = None
     dp_attn: Optional[bool] = Field(default=None, alias=Fields.DP_ATTN.value)
-    spec_decoding: Literal["mtp", "draft_model", "none"] = Field(
+    spec_decoding: Literal["mtp", "eagle3", "draft_model", "none"] = Field(
         default="none", alias=Fields.SPEC_DECODING.value)
     prefill: Optional[WorkerConfig] = None
     decode: Optional[WorkerConfig] = None
@@ -375,12 +430,14 @@ class SingleNodeScenarios(BaseModel):
 
     fixed_seq_len: Optional[List[SingleNodeSeqLenConfig]] = Field(
         default=None, alias=Fields.FIXED_SEQ_LEN.value)
+    fixed_ar_mtp: Optional[List[FixedArMtpSeqLenConfig]] = Field(
+        default=None, alias=Fields.FIXED_AR_MTP.value)
     agentic_coding: Optional[List[AgenticCodingConfig]] = Field(
         default=None, alias=Fields.AGENTIC_CODING.value)
 
     @model_validator(mode='after')
     def at_least_one_scenario(self):
-        if not self.fixed_seq_len and not self.agentic_coding:
+        if not self.fixed_seq_len and not self.fixed_ar_mtp and not self.agentic_coding:
             raise ValueError("At least one scenario type must be specified")
         return self
 
@@ -504,7 +561,7 @@ class ChangelogMatrixEntry(BaseModel):
     """
     model_config = ConfigDict(extra="forbid", populate_by_name=True)
 
-    single_node: dict[str, list[Union[SingleNodeMatrixEntry, SingleNodeAgenticMatrixEntry]]
+    single_node: dict[str, list[Union[SingleNodeMatrixEntry, SingleNodeFixedArMtpMatrixEntry, SingleNodeAgenticMatrixEntry]]
                       ] = Field(default_factory=dict)
     multi_node: dict[str, list[Union[MultiNodeMatrixEntry, MultiNodeAgenticMatrixEntry]]
                      ] = Field(default_factory=dict)