diff --git a/.gitignore b/.gitignore
index b8e528e..f4641ef 100644
--- a/.gitignore
+++ b/.gitignore
@@ -8,6 +8,7 @@ results.csv
 out/
 data/
 outputs/
+experiments/
 .env
 .venv/
 bench_*/
diff --git a/HARNESS.md b/HARNESS.md
index 3df756c..5342421 100644
--- a/HARNESS.md
+++ b/HARNESS.md
@@ -52,6 +52,18 @@ The config file is the interface for setting instance, vLLM, and benchmark args.
 
 **Note**: Always include `results-csv: /tmp/results.csv` in the benchmark config so results are saved and fetched from the remote instance.
 
+To collect server metrics during the benchmark, add BatchBench's metrics flags
+under `benchmark`. The harness passes them through to the Rust CLI and copies
+the resulting directory back to `results/<model>/<config-hash>/metrics/`.
+
+```yaml
+benchmark:
+  results-csv: /tmp/results.csv
+  metrics-output-dir: /tmp/batchbench-metrics
+  metrics-endpoint: /metrics
+  metrics-interval-ms: 1000
+```
+
 ## Architecture
 
 The harness accepts either a single config file or a directory of configs (all `*.yaml` files in that directory).
diff --git a/README.md b/README.md
index 54d0b10..7f2359e 100644
--- a/README.md
+++ b/README.md
@@ -39,6 +39,7 @@ config = {
         )
     ],
     "dry_run": True,
+    "enable_json_decoding": True,
 }
 
 report = batchbench.run_benchmark(config)
@@ -75,9 +76,60 @@ batchbench \
   --output-vary 0
 ```
 
+To benchmark a JSONL dataset instead of generated prompts, pass `--dataset-jsonl`.
+Each non-empty line may be one of:
+
+- `{"text": "prompt text", "input_tokens": 123}` to build a chat completion request.
+- `{"body": {...}, "input_tokens": 123}` to send `body` as the request payload.
+- A full request body object, such as `{"model": "...", "messages": [...]}`.
+
+The CLI `--model` value is injected into every request body at runtime. This means
+dataset rows can omit `model`, and any `model` present in the JSONL is overridden
+by the value passed to `batchbench --model`.
+
+```bash
+batchbench \
+  --dataset-jsonl dataset.jsonl \
+  --model gpt-4o-mini \
+  --users 8 \
+  --requests-per-user 2
+```
+
+The dataset must contain at least `users * requests-per-user` request entries.
+If `--users` is omitted for a dataset run, BatchBench uses as many complete
+request rounds as the dataset can provide for the selected `--requests-per-user`.
+
 Use `--sglang` to apply output token constraints via `min_new_tokens`/`max_new_tokens`
 instead of `min_tokens`/`max_tokens`.
 
+Use `--enable-json-decoding` to request JSON constrained decoding. With `--sglang`,
+BatchBench adds `response_format: {"type": "json_object"}`. Without `--sglang`, it
+adds vLLM `structured_outputs.json_object: true`, equivalent to passing that value
+through the OpenAI SDK's `extra_body`.
+
+Use `--qwen35-disable-thinking` to add
+`chat_template_kwargs: {"enable_thinking": false}` to each request.
+
+Use `--metrics-output-dir` to scrape Prometheus metrics from the benchmarked
+server while the benchmark runs. BatchBench defaults to `--metrics-endpoint
+/metrics`, resolved against `--host`; pass a full URL or another path such as
+`/v1/metrics` if your deployment exposes metrics elsewhere.
+
+```bash
+batchbench \
+  --dataset-jsonl dataset.jsonl \
+  --host http://127.0.0.1:3000 \
+  --model Qwen/Qwen3.6-27B \
+  --users 8 \
+  --requests-per-user 2 \
+  --metrics-output-dir runs/qwen/metrics
+```
+
+Metrics artifacts include raw Prometheus scrape payloads (`raw.promjsonl`),
+parsed long-form samples (`samples.csv`), run metadata (`metadata.json`), and a
+compact derived summary (`summary.json`). SGLang must be launched with
+`--enable-metrics`; vLLM and SGLang both commonly expose metrics at `/metrics`.
+
 Press `Ctrl+C` during a run to cancel active requests and print a partial summary.
 
 ## Rust CLI
diff --git a/bin/batchbench b/bin/batchbench
index ad6e309..af8bfd4 100755
Binary files a/bin/batchbench and b/bin/batchbench differ
diff --git a/configs/k8s/piccolo-k8s-sweep-example-manifest.yaml b/configs/k8s/piccolo-k8s-sweep-example-manifest.yaml
new file mode 100644
index 0000000..f973393
--- /dev/null
+++ b/configs/k8s/piccolo-k8s-sweep-example-manifest.yaml
@@ -0,0 +1,6383 @@
+---
+apiVersion: v1
+kind: ConfigMap
+metadata:
+  name: piccolo-k8s-sweep-example-config
+  namespace: specdec
+  labels:
+    app.kubernetes.io/name: batchbench-sweep
+    app.kubernetes.io/component: benchmark
+    batchbench.doubleword.ai/sweep-name: piccolo-k8s-sweep-example
+data:
+  combinations.json: |-
+    [
+      {
+        "index": 0,
+        "variant_name": "baseline",
+        "variant_extra_args": [
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_16k.jsonl",
+        "dataset_slug": "throughput_16k",
+        "run_slug": "000-baseline-throughput_16k"
+      },
+      {
+        "index": 1,
+        "variant_name": "baseline",
+        "variant_extra_args": [
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_1k.jsonl",
+        "dataset_slug": "throughput_1k",
+        "run_slug": "001-baseline-throughput_1k"
+      },
+      {
+        "index": 2,
+        "variant_name": "baseline",
+        "variant_extra_args": [
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_2k.jsonl",
+        "dataset_slug": "throughput_2k",
+        "run_slug": "002-baseline-throughput_2k"
+      },
+      {
+        "index": 3,
+        "variant_name": "baseline",
+        "variant_extra_args": [
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_32k.jsonl",
+        "dataset_slug": "throughput_32k",
+        "run_slug": "003-baseline-throughput_32k"
+      },
+      {
+        "index": 4,
+        "variant_name": "baseline",
+        "variant_extra_args": [
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_8k.jsonl",
+        "dataset_slug": "throughput_8k",
+        "run_slug": "004-baseline-throughput_8k"
+      },
+      {
+        "index": 5,
+        "variant_name": "baseline",
+        "variant_extra_args": [
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/json_mode_eval/jsonl/train.jsonl",
+        "dataset_slug": "train",
+        "run_slug": "005-baseline-train"
+      },
+      {
+        "index": 6,
+        "variant_name": "mtp-s2-k1-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_16k.jsonl",
+        "dataset_slug": "throughput_16k",
+        "run_slug": "006-mtp-s2-k1-d3-throughput_16k"
+      },
+      {
+        "index": 7,
+        "variant_name": "mtp-s2-k1-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_1k.jsonl",
+        "dataset_slug": "throughput_1k",
+        "run_slug": "007-mtp-s2-k1-d3-throughput_1k"
+      },
+      {
+        "index": 8,
+        "variant_name": "mtp-s2-k1-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_2k.jsonl",
+        "dataset_slug": "throughput_2k",
+        "run_slug": "008-mtp-s2-k1-d3-throughput_2k"
+      },
+      {
+        "index": 9,
+        "variant_name": "mtp-s2-k1-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_32k.jsonl",
+        "dataset_slug": "throughput_32k",
+        "run_slug": "009-mtp-s2-k1-d3-throughput_32k"
+      },
+      {
+        "index": 10,
+        "variant_name": "mtp-s2-k1-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_8k.jsonl",
+        "dataset_slug": "throughput_8k",
+        "run_slug": "010-mtp-s2-k1-d3-throughput_8k"
+      },
+      {
+        "index": 11,
+        "variant_name": "mtp-s2-k1-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/json_mode_eval/jsonl/train.jsonl",
+        "dataset_slug": "train",
+        "run_slug": "011-mtp-s2-k1-d3-train"
+      },
+      {
+        "index": 12,
+        "variant_name": "mtp-s2-k1-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_16k.jsonl",
+        "dataset_slug": "throughput_16k",
+        "run_slug": "012-mtp-s2-k1-d4-throughput_16k"
+      },
+      {
+        "index": 13,
+        "variant_name": "mtp-s2-k1-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_1k.jsonl",
+        "dataset_slug": "throughput_1k",
+        "run_slug": "013-mtp-s2-k1-d4-throughput_1k"
+      },
+      {
+        "index": 14,
+        "variant_name": "mtp-s2-k1-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_2k.jsonl",
+        "dataset_slug": "throughput_2k",
+        "run_slug": "014-mtp-s2-k1-d4-throughput_2k"
+      },
+      {
+        "index": 15,
+        "variant_name": "mtp-s2-k1-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_32k.jsonl",
+        "dataset_slug": "throughput_32k",
+        "run_slug": "015-mtp-s2-k1-d4-throughput_32k"
+      },
+      {
+        "index": 16,
+        "variant_name": "mtp-s2-k1-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_8k.jsonl",
+        "dataset_slug": "throughput_8k",
+        "run_slug": "016-mtp-s2-k1-d4-throughput_8k"
+      },
+      {
+        "index": 17,
+        "variant_name": "mtp-s2-k1-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/json_mode_eval/jsonl/train.jsonl",
+        "dataset_slug": "train",
+        "run_slug": "017-mtp-s2-k1-d4-train"
+      },
+      {
+        "index": 18,
+        "variant_name": "mtp-s2-k1-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_16k.jsonl",
+        "dataset_slug": "throughput_16k",
+        "run_slug": "018-mtp-s2-k1-d5-throughput_16k"
+      },
+      {
+        "index": 19,
+        "variant_name": "mtp-s2-k1-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_1k.jsonl",
+        "dataset_slug": "throughput_1k",
+        "run_slug": "019-mtp-s2-k1-d5-throughput_1k"
+      },
+      {
+        "index": 20,
+        "variant_name": "mtp-s2-k1-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_2k.jsonl",
+        "dataset_slug": "throughput_2k",
+        "run_slug": "020-mtp-s2-k1-d5-throughput_2k"
+      },
+      {
+        "index": 21,
+        "variant_name": "mtp-s2-k1-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_32k.jsonl",
+        "dataset_slug": "throughput_32k",
+        "run_slug": "021-mtp-s2-k1-d5-throughput_32k"
+      },
+      {
+        "index": 22,
+        "variant_name": "mtp-s2-k1-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_8k.jsonl",
+        "dataset_slug": "throughput_8k",
+        "run_slug": "022-mtp-s2-k1-d5-throughput_8k"
+      },
+      {
+        "index": 23,
+        "variant_name": "mtp-s2-k1-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/json_mode_eval/jsonl/train.jsonl",
+        "dataset_slug": "train",
+        "run_slug": "023-mtp-s2-k1-d5-train"
+      },
+      {
+        "index": 24,
+        "variant_name": "mtp-s2-k1-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_16k.jsonl",
+        "dataset_slug": "throughput_16k",
+        "run_slug": "024-mtp-s2-k1-d6-throughput_16k"
+      },
+      {
+        "index": 25,
+        "variant_name": "mtp-s2-k1-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_1k.jsonl",
+        "dataset_slug": "throughput_1k",
+        "run_slug": "025-mtp-s2-k1-d6-throughput_1k"
+      },
+      {
+        "index": 26,
+        "variant_name": "mtp-s2-k1-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_2k.jsonl",
+        "dataset_slug": "throughput_2k",
+        "run_slug": "026-mtp-s2-k1-d6-throughput_2k"
+      },
+      {
+        "index": 27,
+        "variant_name": "mtp-s2-k1-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_32k.jsonl",
+        "dataset_slug": "throughput_32k",
+        "run_slug": "027-mtp-s2-k1-d6-throughput_32k"
+      },
+      {
+        "index": 28,
+        "variant_name": "mtp-s2-k1-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_8k.jsonl",
+        "dataset_slug": "throughput_8k",
+        "run_slug": "028-mtp-s2-k1-d6-throughput_8k"
+      },
+      {
+        "index": 29,
+        "variant_name": "mtp-s2-k1-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/json_mode_eval/jsonl/train.jsonl",
+        "dataset_slug": "train",
+        "run_slug": "029-mtp-s2-k1-d6-train"
+      },
+      {
+        "index": 30,
+        "variant_name": "mtp-s2-k2-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_16k.jsonl",
+        "dataset_slug": "throughput_16k",
+        "run_slug": "030-mtp-s2-k2-d3-throughput_16k"
+      },
+      {
+        "index": 31,
+        "variant_name": "mtp-s2-k2-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_1k.jsonl",
+        "dataset_slug": "throughput_1k",
+        "run_slug": "031-mtp-s2-k2-d3-throughput_1k"
+      },
+      {
+        "index": 32,
+        "variant_name": "mtp-s2-k2-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_2k.jsonl",
+        "dataset_slug": "throughput_2k",
+        "run_slug": "032-mtp-s2-k2-d3-throughput_2k"
+      },
+      {
+        "index": 33,
+        "variant_name": "mtp-s2-k2-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_32k.jsonl",
+        "dataset_slug": "throughput_32k",
+        "run_slug": "033-mtp-s2-k2-d3-throughput_32k"
+      },
+      {
+        "index": 34,
+        "variant_name": "mtp-s2-k2-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_8k.jsonl",
+        "dataset_slug": "throughput_8k",
+        "run_slug": "034-mtp-s2-k2-d3-throughput_8k"
+      },
+      {
+        "index": 35,
+        "variant_name": "mtp-s2-k2-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/json_mode_eval/jsonl/train.jsonl",
+        "dataset_slug": "train",
+        "run_slug": "035-mtp-s2-k2-d3-train"
+      },
+      {
+        "index": 36,
+        "variant_name": "mtp-s2-k2-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_16k.jsonl",
+        "dataset_slug": "throughput_16k",
+        "run_slug": "036-mtp-s2-k2-d4-throughput_16k"
+      },
+      {
+        "index": 37,
+        "variant_name": "mtp-s2-k2-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_1k.jsonl",
+        "dataset_slug": "throughput_1k",
+        "run_slug": "037-mtp-s2-k2-d4-throughput_1k"
+      },
+      {
+        "index": 38,
+        "variant_name": "mtp-s2-k2-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_2k.jsonl",
+        "dataset_slug": "throughput_2k",
+        "run_slug": "038-mtp-s2-k2-d4-throughput_2k"
+      },
+      {
+        "index": 39,
+        "variant_name": "mtp-s2-k2-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_32k.jsonl",
+        "dataset_slug": "throughput_32k",
+        "run_slug": "039-mtp-s2-k2-d4-throughput_32k"
+      },
+      {
+        "index": 40,
+        "variant_name": "mtp-s2-k2-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_8k.jsonl",
+        "dataset_slug": "throughput_8k",
+        "run_slug": "040-mtp-s2-k2-d4-throughput_8k"
+      },
+      {
+        "index": 41,
+        "variant_name": "mtp-s2-k2-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/json_mode_eval/jsonl/train.jsonl",
+        "dataset_slug": "train",
+        "run_slug": "041-mtp-s2-k2-d4-train"
+      },
+      {
+        "index": 42,
+        "variant_name": "mtp-s2-k2-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_16k.jsonl",
+        "dataset_slug": "throughput_16k",
+        "run_slug": "042-mtp-s2-k2-d5-throughput_16k"
+      },
+      {
+        "index": 43,
+        "variant_name": "mtp-s2-k2-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_1k.jsonl",
+        "dataset_slug": "throughput_1k",
+        "run_slug": "043-mtp-s2-k2-d5-throughput_1k"
+      },
+      {
+        "index": 44,
+        "variant_name": "mtp-s2-k2-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_2k.jsonl",
+        "dataset_slug": "throughput_2k",
+        "run_slug": "044-mtp-s2-k2-d5-throughput_2k"
+      },
+      {
+        "index": 45,
+        "variant_name": "mtp-s2-k2-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_32k.jsonl",
+        "dataset_slug": "throughput_32k",
+        "run_slug": "045-mtp-s2-k2-d5-throughput_32k"
+      },
+      {
+        "index": 46,
+        "variant_name": "mtp-s2-k2-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_8k.jsonl",
+        "dataset_slug": "throughput_8k",
+        "run_slug": "046-mtp-s2-k2-d5-throughput_8k"
+      },
+      {
+        "index": 47,
+        "variant_name": "mtp-s2-k2-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/json_mode_eval/jsonl/train.jsonl",
+        "dataset_slug": "train",
+        "run_slug": "047-mtp-s2-k2-d5-train"
+      },
+      {
+        "index": 48,
+        "variant_name": "mtp-s2-k2-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_16k.jsonl",
+        "dataset_slug": "throughput_16k",
+        "run_slug": "048-mtp-s2-k2-d6-throughput_16k"
+      },
+      {
+        "index": 49,
+        "variant_name": "mtp-s2-k2-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_1k.jsonl",
+        "dataset_slug": "throughput_1k",
+        "run_slug": "049-mtp-s2-k2-d6-throughput_1k"
+      },
+      {
+        "index": 50,
+        "variant_name": "mtp-s2-k2-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_2k.jsonl",
+        "dataset_slug": "throughput_2k",
+        "run_slug": "050-mtp-s2-k2-d6-throughput_2k"
+      },
+      {
+        "index": 51,
+        "variant_name": "mtp-s2-k2-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_32k.jsonl",
+        "dataset_slug": "throughput_32k",
+        "run_slug": "051-mtp-s2-k2-d6-throughput_32k"
+      },
+      {
+        "index": 52,
+        "variant_name": "mtp-s2-k2-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_8k.jsonl",
+        "dataset_slug": "throughput_8k",
+        "run_slug": "052-mtp-s2-k2-d6-throughput_8k"
+      },
+      {
+        "index": 53,
+        "variant_name": "mtp-s2-k2-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/json_mode_eval/jsonl/train.jsonl",
+        "dataset_slug": "train",
+        "run_slug": "053-mtp-s2-k2-d6-train"
+      },
+      {
+        "index": 54,
+        "variant_name": "mtp-s2-k3-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_16k.jsonl",
+        "dataset_slug": "throughput_16k",
+        "run_slug": "054-mtp-s2-k3-d3-throughput_16k"
+      },
+      {
+        "index": 55,
+        "variant_name": "mtp-s2-k3-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_1k.jsonl",
+        "dataset_slug": "throughput_1k",
+        "run_slug": "055-mtp-s2-k3-d3-throughput_1k"
+      },
+      {
+        "index": 56,
+        "variant_name": "mtp-s2-k3-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_2k.jsonl",
+        "dataset_slug": "throughput_2k",
+        "run_slug": "056-mtp-s2-k3-d3-throughput_2k"
+      },
+      {
+        "index": 57,
+        "variant_name": "mtp-s2-k3-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_32k.jsonl",
+        "dataset_slug": "throughput_32k",
+        "run_slug": "057-mtp-s2-k3-d3-throughput_32k"
+      },
+      {
+        "index": 58,
+        "variant_name": "mtp-s2-k3-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_8k.jsonl",
+        "dataset_slug": "throughput_8k",
+        "run_slug": "058-mtp-s2-k3-d3-throughput_8k"
+      },
+      {
+        "index": 59,
+        "variant_name": "mtp-s2-k3-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/json_mode_eval/jsonl/train.jsonl",
+        "dataset_slug": "train",
+        "run_slug": "059-mtp-s2-k3-d3-train"
+      },
+      {
+        "index": 60,
+        "variant_name": "mtp-s2-k3-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_16k.jsonl",
+        "dataset_slug": "throughput_16k",
+        "run_slug": "060-mtp-s2-k3-d4-throughput_16k"
+      },
+      {
+        "index": 61,
+        "variant_name": "mtp-s2-k3-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_1k.jsonl",
+        "dataset_slug": "throughput_1k",
+        "run_slug": "061-mtp-s2-k3-d4-throughput_1k"
+      },
+      {
+        "index": 62,
+        "variant_name": "mtp-s2-k3-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_2k.jsonl",
+        "dataset_slug": "throughput_2k",
+        "run_slug": "062-mtp-s2-k3-d4-throughput_2k"
+      },
+      {
+        "index": 63,
+        "variant_name": "mtp-s2-k3-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_32k.jsonl",
+        "dataset_slug": "throughput_32k",
+        "run_slug": "063-mtp-s2-k3-d4-throughput_32k"
+      },
+      {
+        "index": 64,
+        "variant_name": "mtp-s2-k3-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_8k.jsonl",
+        "dataset_slug": "throughput_8k",
+        "run_slug": "064-mtp-s2-k3-d4-throughput_8k"
+      },
+      {
+        "index": 65,
+        "variant_name": "mtp-s2-k3-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/json_mode_eval/jsonl/train.jsonl",
+        "dataset_slug": "train",
+        "run_slug": "065-mtp-s2-k3-d4-train"
+      },
+      {
+        "index": 66,
+        "variant_name": "mtp-s2-k3-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_16k.jsonl",
+        "dataset_slug": "throughput_16k",
+        "run_slug": "066-mtp-s2-k3-d5-throughput_16k"
+      },
+      {
+        "index": 67,
+        "variant_name": "mtp-s2-k3-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_1k.jsonl",
+        "dataset_slug": "throughput_1k",
+        "run_slug": "067-mtp-s2-k3-d5-throughput_1k"
+      },
+      {
+        "index": 68,
+        "variant_name": "mtp-s2-k3-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_2k.jsonl",
+        "dataset_slug": "throughput_2k",
+        "run_slug": "068-mtp-s2-k3-d5-throughput_2k"
+      },
+      {
+        "index": 69,
+        "variant_name": "mtp-s2-k3-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_32k.jsonl",
+        "dataset_slug": "throughput_32k",
+        "run_slug": "069-mtp-s2-k3-d5-throughput_32k"
+      },
+      {
+        "index": 70,
+        "variant_name": "mtp-s2-k3-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_8k.jsonl",
+        "dataset_slug": "throughput_8k",
+        "run_slug": "070-mtp-s2-k3-d5-throughput_8k"
+      },
+      {
+        "index": 71,
+        "variant_name": "mtp-s2-k3-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/json_mode_eval/jsonl/train.jsonl",
+        "dataset_slug": "train",
+        "run_slug": "071-mtp-s2-k3-d5-train"
+      },
+      {
+        "index": 72,
+        "variant_name": "mtp-s2-k3-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_16k.jsonl",
+        "dataset_slug": "throughput_16k",
+        "run_slug": "072-mtp-s2-k3-d6-throughput_16k"
+      },
+      {
+        "index": 73,
+        "variant_name": "mtp-s2-k3-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_1k.jsonl",
+        "dataset_slug": "throughput_1k",
+        "run_slug": "073-mtp-s2-k3-d6-throughput_1k"
+      },
+      {
+        "index": 74,
+        "variant_name": "mtp-s2-k3-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_2k.jsonl",
+        "dataset_slug": "throughput_2k",
+        "run_slug": "074-mtp-s2-k3-d6-throughput_2k"
+      },
+      {
+        "index": 75,
+        "variant_name": "mtp-s2-k3-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_32k.jsonl",
+        "dataset_slug": "throughput_32k",
+        "run_slug": "075-mtp-s2-k3-d6-throughput_32k"
+      },
+      {
+        "index": 76,
+        "variant_name": "mtp-s2-k3-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_8k.jsonl",
+        "dataset_slug": "throughput_8k",
+        "run_slug": "076-mtp-s2-k3-d6-throughput_8k"
+      },
+      {
+        "index": 77,
+        "variant_name": "mtp-s2-k3-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/json_mode_eval/jsonl/train.jsonl",
+        "dataset_slug": "train",
+        "run_slug": "077-mtp-s2-k3-d6-train"
+      },
+      {
+        "index": 78,
+        "variant_name": "mtp-s3-k1-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_16k.jsonl",
+        "dataset_slug": "throughput_16k",
+        "run_slug": "078-mtp-s3-k1-d3-throughput_16k"
+      },
+      {
+        "index": 79,
+        "variant_name": "mtp-s3-k1-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_1k.jsonl",
+        "dataset_slug": "throughput_1k",
+        "run_slug": "079-mtp-s3-k1-d3-throughput_1k"
+      },
+      {
+        "index": 80,
+        "variant_name": "mtp-s3-k1-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_2k.jsonl",
+        "dataset_slug": "throughput_2k",
+        "run_slug": "080-mtp-s3-k1-d3-throughput_2k"
+      },
+      {
+        "index": 81,
+        "variant_name": "mtp-s3-k1-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_32k.jsonl",
+        "dataset_slug": "throughput_32k",
+        "run_slug": "081-mtp-s3-k1-d3-throughput_32k"
+      },
+      {
+        "index": 82,
+        "variant_name": "mtp-s3-k1-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_8k.jsonl",
+        "dataset_slug": "throughput_8k",
+        "run_slug": "082-mtp-s3-k1-d3-throughput_8k"
+      },
+      {
+        "index": 83,
+        "variant_name": "mtp-s3-k1-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/json_mode_eval/jsonl/train.jsonl",
+        "dataset_slug": "train",
+        "run_slug": "083-mtp-s3-k1-d3-train"
+      },
+      {
+        "index": 84,
+        "variant_name": "mtp-s3-k1-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_16k.jsonl",
+        "dataset_slug": "throughput_16k",
+        "run_slug": "084-mtp-s3-k1-d4-throughput_16k"
+      },
+      {
+        "index": 85,
+        "variant_name": "mtp-s3-k1-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_1k.jsonl",
+        "dataset_slug": "throughput_1k",
+        "run_slug": "085-mtp-s3-k1-d4-throughput_1k"
+      },
+      {
+        "index": 86,
+        "variant_name": "mtp-s3-k1-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_2k.jsonl",
+        "dataset_slug": "throughput_2k",
+        "run_slug": "086-mtp-s3-k1-d4-throughput_2k"
+      },
+      {
+        "index": 87,
+        "variant_name": "mtp-s3-k1-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_32k.jsonl",
+        "dataset_slug": "throughput_32k",
+        "run_slug": "087-mtp-s3-k1-d4-throughput_32k"
+      },
+      {
+        "index": 88,
+        "variant_name": "mtp-s3-k1-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_8k.jsonl",
+        "dataset_slug": "throughput_8k",
+        "run_slug": "088-mtp-s3-k1-d4-throughput_8k"
+      },
+      {
+        "index": 89,
+        "variant_name": "mtp-s3-k1-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/json_mode_eval/jsonl/train.jsonl",
+        "dataset_slug": "train",
+        "run_slug": "089-mtp-s3-k1-d4-train"
+      },
+      {
+        "index": 90,
+        "variant_name": "mtp-s3-k1-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_16k.jsonl",
+        "dataset_slug": "throughput_16k",
+        "run_slug": "090-mtp-s3-k1-d5-throughput_16k"
+      },
+      {
+        "index": 91,
+        "variant_name": "mtp-s3-k1-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_1k.jsonl",
+        "dataset_slug": "throughput_1k",
+        "run_slug": "091-mtp-s3-k1-d5-throughput_1k"
+      },
+      {
+        "index": 92,
+        "variant_name": "mtp-s3-k1-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_2k.jsonl",
+        "dataset_slug": "throughput_2k",
+        "run_slug": "092-mtp-s3-k1-d5-throughput_2k"
+      },
+      {
+        "index": 93,
+        "variant_name": "mtp-s3-k1-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_32k.jsonl",
+        "dataset_slug": "throughput_32k",
+        "run_slug": "093-mtp-s3-k1-d5-throughput_32k"
+      },
+      {
+        "index": 94,
+        "variant_name": "mtp-s3-k1-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_8k.jsonl",
+        "dataset_slug": "throughput_8k",
+        "run_slug": "094-mtp-s3-k1-d5-throughput_8k"
+      },
+      {
+        "index": 95,
+        "variant_name": "mtp-s3-k1-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/json_mode_eval/jsonl/train.jsonl",
+        "dataset_slug": "train",
+        "run_slug": "095-mtp-s3-k1-d5-train"
+      },
+      {
+        "index": 96,
+        "variant_name": "mtp-s3-k1-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_16k.jsonl",
+        "dataset_slug": "throughput_16k",
+        "run_slug": "096-mtp-s3-k1-d6-throughput_16k"
+      },
+      {
+        "index": 97,
+        "variant_name": "mtp-s3-k1-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_1k.jsonl",
+        "dataset_slug": "throughput_1k",
+        "run_slug": "097-mtp-s3-k1-d6-throughput_1k"
+      },
+      {
+        "index": 98,
+        "variant_name": "mtp-s3-k1-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_2k.jsonl",
+        "dataset_slug": "throughput_2k",
+        "run_slug": "098-mtp-s3-k1-d6-throughput_2k"
+      },
+      {
+        "index": 99,
+        "variant_name": "mtp-s3-k1-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_32k.jsonl",
+        "dataset_slug": "throughput_32k",
+        "run_slug": "099-mtp-s3-k1-d6-throughput_32k"
+      },
+      {
+        "index": 100,
+        "variant_name": "mtp-s3-k1-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_8k.jsonl",
+        "dataset_slug": "throughput_8k",
+        "run_slug": "100-mtp-s3-k1-d6-throughput_8k"
+      },
+      {
+        "index": 101,
+        "variant_name": "mtp-s3-k1-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/json_mode_eval/jsonl/train.jsonl",
+        "dataset_slug": "train",
+        "run_slug": "101-mtp-s3-k1-d6-train"
+      },
+      {
+        "index": 102,
+        "variant_name": "mtp-s3-k2-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_16k.jsonl",
+        "dataset_slug": "throughput_16k",
+        "run_slug": "102-mtp-s3-k2-d3-throughput_16k"
+      },
+      {
+        "index": 103,
+        "variant_name": "mtp-s3-k2-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_1k.jsonl",
+        "dataset_slug": "throughput_1k",
+        "run_slug": "103-mtp-s3-k2-d3-throughput_1k"
+      },
+      {
+        "index": 104,
+        "variant_name": "mtp-s3-k2-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_2k.jsonl",
+        "dataset_slug": "throughput_2k",
+        "run_slug": "104-mtp-s3-k2-d3-throughput_2k"
+      },
+      {
+        "index": 105,
+        "variant_name": "mtp-s3-k2-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_32k.jsonl",
+        "dataset_slug": "throughput_32k",
+        "run_slug": "105-mtp-s3-k2-d3-throughput_32k"
+      },
+      {
+        "index": 106,
+        "variant_name": "mtp-s3-k2-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_8k.jsonl",
+        "dataset_slug": "throughput_8k",
+        "run_slug": "106-mtp-s3-k2-d3-throughput_8k"
+      },
+      {
+        "index": 107,
+        "variant_name": "mtp-s3-k2-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/json_mode_eval/jsonl/train.jsonl",
+        "dataset_slug": "train",
+        "run_slug": "107-mtp-s3-k2-d3-train"
+      },
+      {
+        "index": 108,
+        "variant_name": "mtp-s3-k2-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_16k.jsonl",
+        "dataset_slug": "throughput_16k",
+        "run_slug": "108-mtp-s3-k2-d4-throughput_16k"
+      },
+      {
+        "index": 109,
+        "variant_name": "mtp-s3-k2-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_1k.jsonl",
+        "dataset_slug": "throughput_1k",
+        "run_slug": "109-mtp-s3-k2-d4-throughput_1k"
+      },
+      {
+        "index": 110,
+        "variant_name": "mtp-s3-k2-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_2k.jsonl",
+        "dataset_slug": "throughput_2k",
+        "run_slug": "110-mtp-s3-k2-d4-throughput_2k"
+      },
+      {
+        "index": 111,
+        "variant_name": "mtp-s3-k2-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_32k.jsonl",
+        "dataset_slug": "throughput_32k",
+        "run_slug": "111-mtp-s3-k2-d4-throughput_32k"
+      },
+      {
+        "index": 112,
+        "variant_name": "mtp-s3-k2-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_8k.jsonl",
+        "dataset_slug": "throughput_8k",
+        "run_slug": "112-mtp-s3-k2-d4-throughput_8k"
+      },
+      {
+        "index": 113,
+        "variant_name": "mtp-s3-k2-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/json_mode_eval/jsonl/train.jsonl",
+        "dataset_slug": "train",
+        "run_slug": "113-mtp-s3-k2-d4-train"
+      },
+      {
+        "index": 114,
+        "variant_name": "mtp-s3-k2-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_16k.jsonl",
+        "dataset_slug": "throughput_16k",
+        "run_slug": "114-mtp-s3-k2-d5-throughput_16k"
+      },
+      {
+        "index": 115,
+        "variant_name": "mtp-s3-k2-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_1k.jsonl",
+        "dataset_slug": "throughput_1k",
+        "run_slug": "115-mtp-s3-k2-d5-throughput_1k"
+      },
+      {
+        "index": 116,
+        "variant_name": "mtp-s3-k2-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_2k.jsonl",
+        "dataset_slug": "throughput_2k",
+        "run_slug": "116-mtp-s3-k2-d5-throughput_2k"
+      },
+      {
+        "index": 117,
+        "variant_name": "mtp-s3-k2-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_32k.jsonl",
+        "dataset_slug": "throughput_32k",
+        "run_slug": "117-mtp-s3-k2-d5-throughput_32k"
+      },
+      {
+        "index": 118,
+        "variant_name": "mtp-s3-k2-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_8k.jsonl",
+        "dataset_slug": "throughput_8k",
+        "run_slug": "118-mtp-s3-k2-d5-throughput_8k"
+      },
+      {
+        "index": 119,
+        "variant_name": "mtp-s3-k2-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/json_mode_eval/jsonl/train.jsonl",
+        "dataset_slug": "train",
+        "run_slug": "119-mtp-s3-k2-d5-train"
+      },
+      {
+        "index": 120,
+        "variant_name": "mtp-s3-k2-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_16k.jsonl",
+        "dataset_slug": "throughput_16k",
+        "run_slug": "120-mtp-s3-k2-d6-throughput_16k"
+      },
+      {
+        "index": 121,
+        "variant_name": "mtp-s3-k2-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_1k.jsonl",
+        "dataset_slug": "throughput_1k",
+        "run_slug": "121-mtp-s3-k2-d6-throughput_1k"
+      },
+      {
+        "index": 122,
+        "variant_name": "mtp-s3-k2-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_2k.jsonl",
+        "dataset_slug": "throughput_2k",
+        "run_slug": "122-mtp-s3-k2-d6-throughput_2k"
+      },
+      {
+        "index": 123,
+        "variant_name": "mtp-s3-k2-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_32k.jsonl",
+        "dataset_slug": "throughput_32k",
+        "run_slug": "123-mtp-s3-k2-d6-throughput_32k"
+      },
+      {
+        "index": 124,
+        "variant_name": "mtp-s3-k2-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_8k.jsonl",
+        "dataset_slug": "throughput_8k",
+        "run_slug": "124-mtp-s3-k2-d6-throughput_8k"
+      },
+      {
+        "index": 125,
+        "variant_name": "mtp-s3-k2-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/json_mode_eval/jsonl/train.jsonl",
+        "dataset_slug": "train",
+        "run_slug": "125-mtp-s3-k2-d6-train"
+      },
+      {
+        "index": 126,
+        "variant_name": "mtp-s3-k3-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_16k.jsonl",
+        "dataset_slug": "throughput_16k",
+        "run_slug": "126-mtp-s3-k3-d3-throughput_16k"
+      },
+      {
+        "index": 127,
+        "variant_name": "mtp-s3-k3-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_1k.jsonl",
+        "dataset_slug": "throughput_1k",
+        "run_slug": "127-mtp-s3-k3-d3-throughput_1k"
+      },
+      {
+        "index": 128,
+        "variant_name": "mtp-s3-k3-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_2k.jsonl",
+        "dataset_slug": "throughput_2k",
+        "run_slug": "128-mtp-s3-k3-d3-throughput_2k"
+      },
+      {
+        "index": 129,
+        "variant_name": "mtp-s3-k3-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_32k.jsonl",
+        "dataset_slug": "throughput_32k",
+        "run_slug": "129-mtp-s3-k3-d3-throughput_32k"
+      },
+      {
+        "index": 130,
+        "variant_name": "mtp-s3-k3-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_8k.jsonl",
+        "dataset_slug": "throughput_8k",
+        "run_slug": "130-mtp-s3-k3-d3-throughput_8k"
+      },
+      {
+        "index": 131,
+        "variant_name": "mtp-s3-k3-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/json_mode_eval/jsonl/train.jsonl",
+        "dataset_slug": "train",
+        "run_slug": "131-mtp-s3-k3-d3-train"
+      },
+      {
+        "index": 132,
+        "variant_name": "mtp-s3-k3-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_16k.jsonl",
+        "dataset_slug": "throughput_16k",
+        "run_slug": "132-mtp-s3-k3-d4-throughput_16k"
+      },
+      {
+        "index": 133,
+        "variant_name": "mtp-s3-k3-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_1k.jsonl",
+        "dataset_slug": "throughput_1k",
+        "run_slug": "133-mtp-s3-k3-d4-throughput_1k"
+      },
+      {
+        "index": 134,
+        "variant_name": "mtp-s3-k3-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_2k.jsonl",
+        "dataset_slug": "throughput_2k",
+        "run_slug": "134-mtp-s3-k3-d4-throughput_2k"
+      },
+      {
+        "index": 135,
+        "variant_name": "mtp-s3-k3-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_32k.jsonl",
+        "dataset_slug": "throughput_32k",
+        "run_slug": "135-mtp-s3-k3-d4-throughput_32k"
+      },
+      {
+        "index": 136,
+        "variant_name": "mtp-s3-k3-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_8k.jsonl",
+        "dataset_slug": "throughput_8k",
+        "run_slug": "136-mtp-s3-k3-d4-throughput_8k"
+      },
+      {
+        "index": 137,
+        "variant_name": "mtp-s3-k3-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/json_mode_eval/jsonl/train.jsonl",
+        "dataset_slug": "train",
+        "run_slug": "137-mtp-s3-k3-d4-train"
+      },
+      {
+        "index": 138,
+        "variant_name": "mtp-s3-k3-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_16k.jsonl",
+        "dataset_slug": "throughput_16k",
+        "run_slug": "138-mtp-s3-k3-d5-throughput_16k"
+      },
+      {
+        "index": 139,
+        "variant_name": "mtp-s3-k3-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_1k.jsonl",
+        "dataset_slug": "throughput_1k",
+        "run_slug": "139-mtp-s3-k3-d5-throughput_1k"
+      },
+      {
+        "index": 140,
+        "variant_name": "mtp-s3-k3-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_2k.jsonl",
+        "dataset_slug": "throughput_2k",
+        "run_slug": "140-mtp-s3-k3-d5-throughput_2k"
+      },
+      {
+        "index": 141,
+        "variant_name": "mtp-s3-k3-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_32k.jsonl",
+        "dataset_slug": "throughput_32k",
+        "run_slug": "141-mtp-s3-k3-d5-throughput_32k"
+      },
+      {
+        "index": 142,
+        "variant_name": "mtp-s3-k3-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_8k.jsonl",
+        "dataset_slug": "throughput_8k",
+        "run_slug": "142-mtp-s3-k3-d5-throughput_8k"
+      },
+      {
+        "index": 143,
+        "variant_name": "mtp-s3-k3-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/json_mode_eval/jsonl/train.jsonl",
+        "dataset_slug": "train",
+        "run_slug": "143-mtp-s3-k3-d5-train"
+      },
+      {
+        "index": 144,
+        "variant_name": "mtp-s3-k3-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_16k.jsonl",
+        "dataset_slug": "throughput_16k",
+        "run_slug": "144-mtp-s3-k3-d6-throughput_16k"
+      },
+      {
+        "index": 145,
+        "variant_name": "mtp-s3-k3-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_1k.jsonl",
+        "dataset_slug": "throughput_1k",
+        "run_slug": "145-mtp-s3-k3-d6-throughput_1k"
+      },
+      {
+        "index": 146,
+        "variant_name": "mtp-s3-k3-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_2k.jsonl",
+        "dataset_slug": "throughput_2k",
+        "run_slug": "146-mtp-s3-k3-d6-throughput_2k"
+      },
+      {
+        "index": 147,
+        "variant_name": "mtp-s3-k3-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_32k.jsonl",
+        "dataset_slug": "throughput_32k",
+        "run_slug": "147-mtp-s3-k3-d6-throughput_32k"
+      },
+      {
+        "index": 148,
+        "variant_name": "mtp-s3-k3-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_8k.jsonl",
+        "dataset_slug": "throughput_8k",
+        "run_slug": "148-mtp-s3-k3-d6-throughput_8k"
+      },
+      {
+        "index": 149,
+        "variant_name": "mtp-s3-k3-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/json_mode_eval/jsonl/train.jsonl",
+        "dataset_slug": "train",
+        "run_slug": "149-mtp-s3-k3-d6-train"
+      },
+      {
+        "index": 150,
+        "variant_name": "mtp-s4-k1-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_16k.jsonl",
+        "dataset_slug": "throughput_16k",
+        "run_slug": "150-mtp-s4-k1-d3-throughput_16k"
+      },
+      {
+        "index": 151,
+        "variant_name": "mtp-s4-k1-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_1k.jsonl",
+        "dataset_slug": "throughput_1k",
+        "run_slug": "151-mtp-s4-k1-d3-throughput_1k"
+      },
+      {
+        "index": 152,
+        "variant_name": "mtp-s4-k1-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_2k.jsonl",
+        "dataset_slug": "throughput_2k",
+        "run_slug": "152-mtp-s4-k1-d3-throughput_2k"
+      },
+      {
+        "index": 153,
+        "variant_name": "mtp-s4-k1-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_32k.jsonl",
+        "dataset_slug": "throughput_32k",
+        "run_slug": "153-mtp-s4-k1-d3-throughput_32k"
+      },
+      {
+        "index": 154,
+        "variant_name": "mtp-s4-k1-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_8k.jsonl",
+        "dataset_slug": "throughput_8k",
+        "run_slug": "154-mtp-s4-k1-d3-throughput_8k"
+      },
+      {
+        "index": 155,
+        "variant_name": "mtp-s4-k1-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/json_mode_eval/jsonl/train.jsonl",
+        "dataset_slug": "train",
+        "run_slug": "155-mtp-s4-k1-d3-train"
+      },
+      {
+        "index": 156,
+        "variant_name": "mtp-s4-k1-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_16k.jsonl",
+        "dataset_slug": "throughput_16k",
+        "run_slug": "156-mtp-s4-k1-d4-throughput_16k"
+      },
+      {
+        "index": 157,
+        "variant_name": "mtp-s4-k1-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_1k.jsonl",
+        "dataset_slug": "throughput_1k",
+        "run_slug": "157-mtp-s4-k1-d4-throughput_1k"
+      },
+      {
+        "index": 158,
+        "variant_name": "mtp-s4-k1-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_2k.jsonl",
+        "dataset_slug": "throughput_2k",
+        "run_slug": "158-mtp-s4-k1-d4-throughput_2k"
+      },
+      {
+        "index": 159,
+        "variant_name": "mtp-s4-k1-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_32k.jsonl",
+        "dataset_slug": "throughput_32k",
+        "run_slug": "159-mtp-s4-k1-d4-throughput_32k"
+      },
+      {
+        "index": 160,
+        "variant_name": "mtp-s4-k1-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_8k.jsonl",
+        "dataset_slug": "throughput_8k",
+        "run_slug": "160-mtp-s4-k1-d4-throughput_8k"
+      },
+      {
+        "index": 161,
+        "variant_name": "mtp-s4-k1-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/json_mode_eval/jsonl/train.jsonl",
+        "dataset_slug": "train",
+        "run_slug": "161-mtp-s4-k1-d4-train"
+      },
+      {
+        "index": 162,
+        "variant_name": "mtp-s4-k1-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_16k.jsonl",
+        "dataset_slug": "throughput_16k",
+        "run_slug": "162-mtp-s4-k1-d5-throughput_16k"
+      },
+      {
+        "index": 163,
+        "variant_name": "mtp-s4-k1-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_1k.jsonl",
+        "dataset_slug": "throughput_1k",
+        "run_slug": "163-mtp-s4-k1-d5-throughput_1k"
+      },
+      {
+        "index": 164,
+        "variant_name": "mtp-s4-k1-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_2k.jsonl",
+        "dataset_slug": "throughput_2k",
+        "run_slug": "164-mtp-s4-k1-d5-throughput_2k"
+      },
+      {
+        "index": 165,
+        "variant_name": "mtp-s4-k1-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_32k.jsonl",
+        "dataset_slug": "throughput_32k",
+        "run_slug": "165-mtp-s4-k1-d5-throughput_32k"
+      },
+      {
+        "index": 166,
+        "variant_name": "mtp-s4-k1-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_8k.jsonl",
+        "dataset_slug": "throughput_8k",
+        "run_slug": "166-mtp-s4-k1-d5-throughput_8k"
+      },
+      {
+        "index": 167,
+        "variant_name": "mtp-s4-k1-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/json_mode_eval/jsonl/train.jsonl",
+        "dataset_slug": "train",
+        "run_slug": "167-mtp-s4-k1-d5-train"
+      },
+      {
+        "index": 168,
+        "variant_name": "mtp-s4-k1-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_16k.jsonl",
+        "dataset_slug": "throughput_16k",
+        "run_slug": "168-mtp-s4-k1-d6-throughput_16k"
+      },
+      {
+        "index": 169,
+        "variant_name": "mtp-s4-k1-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_1k.jsonl",
+        "dataset_slug": "throughput_1k",
+        "run_slug": "169-mtp-s4-k1-d6-throughput_1k"
+      },
+      {
+        "index": 170,
+        "variant_name": "mtp-s4-k1-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_2k.jsonl",
+        "dataset_slug": "throughput_2k",
+        "run_slug": "170-mtp-s4-k1-d6-throughput_2k"
+      },
+      {
+        "index": 171,
+        "variant_name": "mtp-s4-k1-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_32k.jsonl",
+        "dataset_slug": "throughput_32k",
+        "run_slug": "171-mtp-s4-k1-d6-throughput_32k"
+      },
+      {
+        "index": 172,
+        "variant_name": "mtp-s4-k1-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_8k.jsonl",
+        "dataset_slug": "throughput_8k",
+        "run_slug": "172-mtp-s4-k1-d6-throughput_8k"
+      },
+      {
+        "index": 173,
+        "variant_name": "mtp-s4-k1-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/json_mode_eval/jsonl/train.jsonl",
+        "dataset_slug": "train",
+        "run_slug": "173-mtp-s4-k1-d6-train"
+      },
+      {
+        "index": 174,
+        "variant_name": "mtp-s4-k2-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_16k.jsonl",
+        "dataset_slug": "throughput_16k",
+        "run_slug": "174-mtp-s4-k2-d3-throughput_16k"
+      },
+      {
+        "index": 175,
+        "variant_name": "mtp-s4-k2-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_1k.jsonl",
+        "dataset_slug": "throughput_1k",
+        "run_slug": "175-mtp-s4-k2-d3-throughput_1k"
+      },
+      {
+        "index": 176,
+        "variant_name": "mtp-s4-k2-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_2k.jsonl",
+        "dataset_slug": "throughput_2k",
+        "run_slug": "176-mtp-s4-k2-d3-throughput_2k"
+      },
+      {
+        "index": 177,
+        "variant_name": "mtp-s4-k2-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_32k.jsonl",
+        "dataset_slug": "throughput_32k",
+        "run_slug": "177-mtp-s4-k2-d3-throughput_32k"
+      },
+      {
+        "index": 178,
+        "variant_name": "mtp-s4-k2-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_8k.jsonl",
+        "dataset_slug": "throughput_8k",
+        "run_slug": "178-mtp-s4-k2-d3-throughput_8k"
+      },
+      {
+        "index": 179,
+        "variant_name": "mtp-s4-k2-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/json_mode_eval/jsonl/train.jsonl",
+        "dataset_slug": "train",
+        "run_slug": "179-mtp-s4-k2-d3-train"
+      },
+      {
+        "index": 180,
+        "variant_name": "mtp-s4-k2-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_16k.jsonl",
+        "dataset_slug": "throughput_16k",
+        "run_slug": "180-mtp-s4-k2-d4-throughput_16k"
+      },
+      {
+        "index": 181,
+        "variant_name": "mtp-s4-k2-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_1k.jsonl",
+        "dataset_slug": "throughput_1k",
+        "run_slug": "181-mtp-s4-k2-d4-throughput_1k"
+      },
+      {
+        "index": 182,
+        "variant_name": "mtp-s4-k2-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_2k.jsonl",
+        "dataset_slug": "throughput_2k",
+        "run_slug": "182-mtp-s4-k2-d4-throughput_2k"
+      },
+      {
+        "index": 183,
+        "variant_name": "mtp-s4-k2-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_32k.jsonl",
+        "dataset_slug": "throughput_32k",
+        "run_slug": "183-mtp-s4-k2-d4-throughput_32k"
+      },
+      {
+        "index": 184,
+        "variant_name": "mtp-s4-k2-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_8k.jsonl",
+        "dataset_slug": "throughput_8k",
+        "run_slug": "184-mtp-s4-k2-d4-throughput_8k"
+      },
+      {
+        "index": 185,
+        "variant_name": "mtp-s4-k2-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/json_mode_eval/jsonl/train.jsonl",
+        "dataset_slug": "train",
+        "run_slug": "185-mtp-s4-k2-d4-train"
+      },
+      {
+        "index": 186,
+        "variant_name": "mtp-s4-k2-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_16k.jsonl",
+        "dataset_slug": "throughput_16k",
+        "run_slug": "186-mtp-s4-k2-d5-throughput_16k"
+      },
+      {
+        "index": 187,
+        "variant_name": "mtp-s4-k2-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_1k.jsonl",
+        "dataset_slug": "throughput_1k",
+        "run_slug": "187-mtp-s4-k2-d5-throughput_1k"
+      },
+      {
+        "index": 188,
+        "variant_name": "mtp-s4-k2-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_2k.jsonl",
+        "dataset_slug": "throughput_2k",
+        "run_slug": "188-mtp-s4-k2-d5-throughput_2k"
+      },
+      {
+        "index": 189,
+        "variant_name": "mtp-s4-k2-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_32k.jsonl",
+        "dataset_slug": "throughput_32k",
+        "run_slug": "189-mtp-s4-k2-d5-throughput_32k"
+      },
+      {
+        "index": 190,
+        "variant_name": "mtp-s4-k2-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_8k.jsonl",
+        "dataset_slug": "throughput_8k",
+        "run_slug": "190-mtp-s4-k2-d5-throughput_8k"
+      },
+      {
+        "index": 191,
+        "variant_name": "mtp-s4-k2-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/json_mode_eval/jsonl/train.jsonl",
+        "dataset_slug": "train",
+        "run_slug": "191-mtp-s4-k2-d5-train"
+      },
+      {
+        "index": 192,
+        "variant_name": "mtp-s4-k2-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_16k.jsonl",
+        "dataset_slug": "throughput_16k",
+        "run_slug": "192-mtp-s4-k2-d6-throughput_16k"
+      },
+      {
+        "index": 193,
+        "variant_name": "mtp-s4-k2-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_1k.jsonl",
+        "dataset_slug": "throughput_1k",
+        "run_slug": "193-mtp-s4-k2-d6-throughput_1k"
+      },
+      {
+        "index": 194,
+        "variant_name": "mtp-s4-k2-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_2k.jsonl",
+        "dataset_slug": "throughput_2k",
+        "run_slug": "194-mtp-s4-k2-d6-throughput_2k"
+      },
+      {
+        "index": 195,
+        "variant_name": "mtp-s4-k2-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_32k.jsonl",
+        "dataset_slug": "throughput_32k",
+        "run_slug": "195-mtp-s4-k2-d6-throughput_32k"
+      },
+      {
+        "index": 196,
+        "variant_name": "mtp-s4-k2-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_8k.jsonl",
+        "dataset_slug": "throughput_8k",
+        "run_slug": "196-mtp-s4-k2-d6-throughput_8k"
+      },
+      {
+        "index": 197,
+        "variant_name": "mtp-s4-k2-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/json_mode_eval/jsonl/train.jsonl",
+        "dataset_slug": "train",
+        "run_slug": "197-mtp-s4-k2-d6-train"
+      },
+      {
+        "index": 198,
+        "variant_name": "mtp-s4-k3-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_16k.jsonl",
+        "dataset_slug": "throughput_16k",
+        "run_slug": "198-mtp-s4-k3-d3-throughput_16k"
+      },
+      {
+        "index": 199,
+        "variant_name": "mtp-s4-k3-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_1k.jsonl",
+        "dataset_slug": "throughput_1k",
+        "run_slug": "199-mtp-s4-k3-d3-throughput_1k"
+      },
+      {
+        "index": 200,
+        "variant_name": "mtp-s4-k3-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_2k.jsonl",
+        "dataset_slug": "throughput_2k",
+        "run_slug": "200-mtp-s4-k3-d3-throughput_2k"
+      },
+      {
+        "index": 201,
+        "variant_name": "mtp-s4-k3-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_32k.jsonl",
+        "dataset_slug": "throughput_32k",
+        "run_slug": "201-mtp-s4-k3-d3-throughput_32k"
+      },
+      {
+        "index": 202,
+        "variant_name": "mtp-s4-k3-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_8k.jsonl",
+        "dataset_slug": "throughput_8k",
+        "run_slug": "202-mtp-s4-k3-d3-throughput_8k"
+      },
+      {
+        "index": 203,
+        "variant_name": "mtp-s4-k3-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/json_mode_eval/jsonl/train.jsonl",
+        "dataset_slug": "train",
+        "run_slug": "203-mtp-s4-k3-d3-train"
+      },
+      {
+        "index": 204,
+        "variant_name": "mtp-s4-k3-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_16k.jsonl",
+        "dataset_slug": "throughput_16k",
+        "run_slug": "204-mtp-s4-k3-d4-throughput_16k"
+      },
+      {
+        "index": 205,
+        "variant_name": "mtp-s4-k3-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_1k.jsonl",
+        "dataset_slug": "throughput_1k",
+        "run_slug": "205-mtp-s4-k3-d4-throughput_1k"
+      },
+      {
+        "index": 206,
+        "variant_name": "mtp-s4-k3-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_2k.jsonl",
+        "dataset_slug": "throughput_2k",
+        "run_slug": "206-mtp-s4-k3-d4-throughput_2k"
+      },
+      {
+        "index": 207,
+        "variant_name": "mtp-s4-k3-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_32k.jsonl",
+        "dataset_slug": "throughput_32k",
+        "run_slug": "207-mtp-s4-k3-d4-throughput_32k"
+      },
+      {
+        "index": 208,
+        "variant_name": "mtp-s4-k3-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_8k.jsonl",
+        "dataset_slug": "throughput_8k",
+        "run_slug": "208-mtp-s4-k3-d4-throughput_8k"
+      },
+      {
+        "index": 209,
+        "variant_name": "mtp-s4-k3-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/json_mode_eval/jsonl/train.jsonl",
+        "dataset_slug": "train",
+        "run_slug": "209-mtp-s4-k3-d4-train"
+      },
+      {
+        "index": 210,
+        "variant_name": "mtp-s4-k3-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_16k.jsonl",
+        "dataset_slug": "throughput_16k",
+        "run_slug": "210-mtp-s4-k3-d5-throughput_16k"
+      },
+      {
+        "index": 211,
+        "variant_name": "mtp-s4-k3-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_1k.jsonl",
+        "dataset_slug": "throughput_1k",
+        "run_slug": "211-mtp-s4-k3-d5-throughput_1k"
+      },
+      {
+        "index": 212,
+        "variant_name": "mtp-s4-k3-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_2k.jsonl",
+        "dataset_slug": "throughput_2k",
+        "run_slug": "212-mtp-s4-k3-d5-throughput_2k"
+      },
+      {
+        "index": 213,
+        "variant_name": "mtp-s4-k3-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_32k.jsonl",
+        "dataset_slug": "throughput_32k",
+        "run_slug": "213-mtp-s4-k3-d5-throughput_32k"
+      },
+      {
+        "index": 214,
+        "variant_name": "mtp-s4-k3-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_8k.jsonl",
+        "dataset_slug": "throughput_8k",
+        "run_slug": "214-mtp-s4-k3-d5-throughput_8k"
+      },
+      {
+        "index": 215,
+        "variant_name": "mtp-s4-k3-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/json_mode_eval/jsonl/train.jsonl",
+        "dataset_slug": "train",
+        "run_slug": "215-mtp-s4-k3-d5-train"
+      },
+      {
+        "index": 216,
+        "variant_name": "mtp-s4-k3-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_16k.jsonl",
+        "dataset_slug": "throughput_16k",
+        "run_slug": "216-mtp-s4-k3-d6-throughput_16k"
+      },
+      {
+        "index": 217,
+        "variant_name": "mtp-s4-k3-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_1k.jsonl",
+        "dataset_slug": "throughput_1k",
+        "run_slug": "217-mtp-s4-k3-d6-throughput_1k"
+      },
+      {
+        "index": 218,
+        "variant_name": "mtp-s4-k3-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_2k.jsonl",
+        "dataset_slug": "throughput_2k",
+        "run_slug": "218-mtp-s4-k3-d6-throughput_2k"
+      },
+      {
+        "index": 219,
+        "variant_name": "mtp-s4-k3-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_32k.jsonl",
+        "dataset_slug": "throughput_32k",
+        "run_slug": "219-mtp-s4-k3-d6-throughput_32k"
+      },
+      {
+        "index": 220,
+        "variant_name": "mtp-s4-k3-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_8k.jsonl",
+        "dataset_slug": "throughput_8k",
+        "run_slug": "220-mtp-s4-k3-d6-throughput_8k"
+      },
+      {
+        "index": 221,
+        "variant_name": "mtp-s4-k3-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/json_mode_eval/jsonl/train.jsonl",
+        "dataset_slug": "train",
+        "run_slug": "221-mtp-s4-k3-d6-train"
+      },
+      {
+        "index": 222,
+        "variant_name": "mtp-s5-k1-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_16k.jsonl",
+        "dataset_slug": "throughput_16k",
+        "run_slug": "222-mtp-s5-k1-d3-throughput_16k"
+      },
+      {
+        "index": 223,
+        "variant_name": "mtp-s5-k1-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_1k.jsonl",
+        "dataset_slug": "throughput_1k",
+        "run_slug": "223-mtp-s5-k1-d3-throughput_1k"
+      },
+      {
+        "index": 224,
+        "variant_name": "mtp-s5-k1-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_2k.jsonl",
+        "dataset_slug": "throughput_2k",
+        "run_slug": "224-mtp-s5-k1-d3-throughput_2k"
+      },
+      {
+        "index": 225,
+        "variant_name": "mtp-s5-k1-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_32k.jsonl",
+        "dataset_slug": "throughput_32k",
+        "run_slug": "225-mtp-s5-k1-d3-throughput_32k"
+      },
+      {
+        "index": 226,
+        "variant_name": "mtp-s5-k1-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_8k.jsonl",
+        "dataset_slug": "throughput_8k",
+        "run_slug": "226-mtp-s5-k1-d3-throughput_8k"
+      },
+      {
+        "index": 227,
+        "variant_name": "mtp-s5-k1-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/json_mode_eval/jsonl/train.jsonl",
+        "dataset_slug": "train",
+        "run_slug": "227-mtp-s5-k1-d3-train"
+      },
+      {
+        "index": 228,
+        "variant_name": "mtp-s5-k1-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_16k.jsonl",
+        "dataset_slug": "throughput_16k",
+        "run_slug": "228-mtp-s5-k1-d4-throughput_16k"
+      },
+      {
+        "index": 229,
+        "variant_name": "mtp-s5-k1-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_1k.jsonl",
+        "dataset_slug": "throughput_1k",
+        "run_slug": "229-mtp-s5-k1-d4-throughput_1k"
+      },
+      {
+        "index": 230,
+        "variant_name": "mtp-s5-k1-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_2k.jsonl",
+        "dataset_slug": "throughput_2k",
+        "run_slug": "230-mtp-s5-k1-d4-throughput_2k"
+      },
+      {
+        "index": 231,
+        "variant_name": "mtp-s5-k1-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_32k.jsonl",
+        "dataset_slug": "throughput_32k",
+        "run_slug": "231-mtp-s5-k1-d4-throughput_32k"
+      },
+      {
+        "index": 232,
+        "variant_name": "mtp-s5-k1-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_8k.jsonl",
+        "dataset_slug": "throughput_8k",
+        "run_slug": "232-mtp-s5-k1-d4-throughput_8k"
+      },
+      {
+        "index": 233,
+        "variant_name": "mtp-s5-k1-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/json_mode_eval/jsonl/train.jsonl",
+        "dataset_slug": "train",
+        "run_slug": "233-mtp-s5-k1-d4-train"
+      },
+      {
+        "index": 234,
+        "variant_name": "mtp-s5-k1-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_16k.jsonl",
+        "dataset_slug": "throughput_16k",
+        "run_slug": "234-mtp-s5-k1-d5-throughput_16k"
+      },
+      {
+        "index": 235,
+        "variant_name": "mtp-s5-k1-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_1k.jsonl",
+        "dataset_slug": "throughput_1k",
+        "run_slug": "235-mtp-s5-k1-d5-throughput_1k"
+      },
+      {
+        "index": 236,
+        "variant_name": "mtp-s5-k1-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_2k.jsonl",
+        "dataset_slug": "throughput_2k",
+        "run_slug": "236-mtp-s5-k1-d5-throughput_2k"
+      },
+      {
+        "index": 237,
+        "variant_name": "mtp-s5-k1-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_32k.jsonl",
+        "dataset_slug": "throughput_32k",
+        "run_slug": "237-mtp-s5-k1-d5-throughput_32k"
+      },
+      {
+        "index": 238,
+        "variant_name": "mtp-s5-k1-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_8k.jsonl",
+        "dataset_slug": "throughput_8k",
+        "run_slug": "238-mtp-s5-k1-d5-throughput_8k"
+      },
+      {
+        "index": 239,
+        "variant_name": "mtp-s5-k1-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/json_mode_eval/jsonl/train.jsonl",
+        "dataset_slug": "train",
+        "run_slug": "239-mtp-s5-k1-d5-train"
+      },
+      {
+        "index": 240,
+        "variant_name": "mtp-s5-k1-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_16k.jsonl",
+        "dataset_slug": "throughput_16k",
+        "run_slug": "240-mtp-s5-k1-d6-throughput_16k"
+      },
+      {
+        "index": 241,
+        "variant_name": "mtp-s5-k1-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_1k.jsonl",
+        "dataset_slug": "throughput_1k",
+        "run_slug": "241-mtp-s5-k1-d6-throughput_1k"
+      },
+      {
+        "index": 242,
+        "variant_name": "mtp-s5-k1-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_2k.jsonl",
+        "dataset_slug": "throughput_2k",
+        "run_slug": "242-mtp-s5-k1-d6-throughput_2k"
+      },
+      {
+        "index": 243,
+        "variant_name": "mtp-s5-k1-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_32k.jsonl",
+        "dataset_slug": "throughput_32k",
+        "run_slug": "243-mtp-s5-k1-d6-throughput_32k"
+      },
+      {
+        "index": 244,
+        "variant_name": "mtp-s5-k1-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_8k.jsonl",
+        "dataset_slug": "throughput_8k",
+        "run_slug": "244-mtp-s5-k1-d6-throughput_8k"
+      },
+      {
+        "index": 245,
+        "variant_name": "mtp-s5-k1-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/json_mode_eval/jsonl/train.jsonl",
+        "dataset_slug": "train",
+        "run_slug": "245-mtp-s5-k1-d6-train"
+      },
+      {
+        "index": 246,
+        "variant_name": "mtp-s5-k2-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_16k.jsonl",
+        "dataset_slug": "throughput_16k",
+        "run_slug": "246-mtp-s5-k2-d3-throughput_16k"
+      },
+      {
+        "index": 247,
+        "variant_name": "mtp-s5-k2-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_1k.jsonl",
+        "dataset_slug": "throughput_1k",
+        "run_slug": "247-mtp-s5-k2-d3-throughput_1k"
+      },
+      {
+        "index": 248,
+        "variant_name": "mtp-s5-k2-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_2k.jsonl",
+        "dataset_slug": "throughput_2k",
+        "run_slug": "248-mtp-s5-k2-d3-throughput_2k"
+      },
+      {
+        "index": 249,
+        "variant_name": "mtp-s5-k2-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_32k.jsonl",
+        "dataset_slug": "throughput_32k",
+        "run_slug": "249-mtp-s5-k2-d3-throughput_32k"
+      },
+      {
+        "index": 250,
+        "variant_name": "mtp-s5-k2-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_8k.jsonl",
+        "dataset_slug": "throughput_8k",
+        "run_slug": "250-mtp-s5-k2-d3-throughput_8k"
+      },
+      {
+        "index": 251,
+        "variant_name": "mtp-s5-k2-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/json_mode_eval/jsonl/train.jsonl",
+        "dataset_slug": "train",
+        "run_slug": "251-mtp-s5-k2-d3-train"
+      },
+      {
+        "index": 252,
+        "variant_name": "mtp-s5-k2-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_16k.jsonl",
+        "dataset_slug": "throughput_16k",
+        "run_slug": "252-mtp-s5-k2-d4-throughput_16k"
+      },
+      {
+        "index": 253,
+        "variant_name": "mtp-s5-k2-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_1k.jsonl",
+        "dataset_slug": "throughput_1k",
+        "run_slug": "253-mtp-s5-k2-d4-throughput_1k"
+      },
+      {
+        "index": 254,
+        "variant_name": "mtp-s5-k2-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_2k.jsonl",
+        "dataset_slug": "throughput_2k",
+        "run_slug": "254-mtp-s5-k2-d4-throughput_2k"
+      },
+      {
+        "index": 255,
+        "variant_name": "mtp-s5-k2-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_32k.jsonl",
+        "dataset_slug": "throughput_32k",
+        "run_slug": "255-mtp-s5-k2-d4-throughput_32k"
+      },
+      {
+        "index": 256,
+        "variant_name": "mtp-s5-k2-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_8k.jsonl",
+        "dataset_slug": "throughput_8k",
+        "run_slug": "256-mtp-s5-k2-d4-throughput_8k"
+      },
+      {
+        "index": 257,
+        "variant_name": "mtp-s5-k2-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/json_mode_eval/jsonl/train.jsonl",
+        "dataset_slug": "train",
+        "run_slug": "257-mtp-s5-k2-d4-train"
+      },
+      {
+        "index": 258,
+        "variant_name": "mtp-s5-k2-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_16k.jsonl",
+        "dataset_slug": "throughput_16k",
+        "run_slug": "258-mtp-s5-k2-d5-throughput_16k"
+      },
+      {
+        "index": 259,
+        "variant_name": "mtp-s5-k2-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_1k.jsonl",
+        "dataset_slug": "throughput_1k",
+        "run_slug": "259-mtp-s5-k2-d5-throughput_1k"
+      },
+      {
+        "index": 260,
+        "variant_name": "mtp-s5-k2-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_2k.jsonl",
+        "dataset_slug": "throughput_2k",
+        "run_slug": "260-mtp-s5-k2-d5-throughput_2k"
+      },
+      {
+        "index": 261,
+        "variant_name": "mtp-s5-k2-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_32k.jsonl",
+        "dataset_slug": "throughput_32k",
+        "run_slug": "261-mtp-s5-k2-d5-throughput_32k"
+      },
+      {
+        "index": 262,
+        "variant_name": "mtp-s5-k2-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_8k.jsonl",
+        "dataset_slug": "throughput_8k",
+        "run_slug": "262-mtp-s5-k2-d5-throughput_8k"
+      },
+      {
+        "index": 263,
+        "variant_name": "mtp-s5-k2-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/json_mode_eval/jsonl/train.jsonl",
+        "dataset_slug": "train",
+        "run_slug": "263-mtp-s5-k2-d5-train"
+      },
+      {
+        "index": 264,
+        "variant_name": "mtp-s5-k2-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_16k.jsonl",
+        "dataset_slug": "throughput_16k",
+        "run_slug": "264-mtp-s5-k2-d6-throughput_16k"
+      },
+      {
+        "index": 265,
+        "variant_name": "mtp-s5-k2-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_1k.jsonl",
+        "dataset_slug": "throughput_1k",
+        "run_slug": "265-mtp-s5-k2-d6-throughput_1k"
+      },
+      {
+        "index": 266,
+        "variant_name": "mtp-s5-k2-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_2k.jsonl",
+        "dataset_slug": "throughput_2k",
+        "run_slug": "266-mtp-s5-k2-d6-throughput_2k"
+      },
+      {
+        "index": 267,
+        "variant_name": "mtp-s5-k2-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_32k.jsonl",
+        "dataset_slug": "throughput_32k",
+        "run_slug": "267-mtp-s5-k2-d6-throughput_32k"
+      },
+      {
+        "index": 268,
+        "variant_name": "mtp-s5-k2-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_8k.jsonl",
+        "dataset_slug": "throughput_8k",
+        "run_slug": "268-mtp-s5-k2-d6-throughput_8k"
+      },
+      {
+        "index": 269,
+        "variant_name": "mtp-s5-k2-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/json_mode_eval/jsonl/train.jsonl",
+        "dataset_slug": "train",
+        "run_slug": "269-mtp-s5-k2-d6-train"
+      },
+      {
+        "index": 270,
+        "variant_name": "mtp-s5-k3-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_16k.jsonl",
+        "dataset_slug": "throughput_16k",
+        "run_slug": "270-mtp-s5-k3-d3-throughput_16k"
+      },
+      {
+        "index": 271,
+        "variant_name": "mtp-s5-k3-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_1k.jsonl",
+        "dataset_slug": "throughput_1k",
+        "run_slug": "271-mtp-s5-k3-d3-throughput_1k"
+      },
+      {
+        "index": 272,
+        "variant_name": "mtp-s5-k3-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_2k.jsonl",
+        "dataset_slug": "throughput_2k",
+        "run_slug": "272-mtp-s5-k3-d3-throughput_2k"
+      },
+      {
+        "index": 273,
+        "variant_name": "mtp-s5-k3-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_32k.jsonl",
+        "dataset_slug": "throughput_32k",
+        "run_slug": "273-mtp-s5-k3-d3-throughput_32k"
+      },
+      {
+        "index": 274,
+        "variant_name": "mtp-s5-k3-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_8k.jsonl",
+        "dataset_slug": "throughput_8k",
+        "run_slug": "274-mtp-s5-k3-d3-throughput_8k"
+      },
+      {
+        "index": 275,
+        "variant_name": "mtp-s5-k3-d3",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/json_mode_eval/jsonl/train.jsonl",
+        "dataset_slug": "train",
+        "run_slug": "275-mtp-s5-k3-d3-train"
+      },
+      {
+        "index": 276,
+        "variant_name": "mtp-s5-k3-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_16k.jsonl",
+        "dataset_slug": "throughput_16k",
+        "run_slug": "276-mtp-s5-k3-d4-throughput_16k"
+      },
+      {
+        "index": 277,
+        "variant_name": "mtp-s5-k3-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_1k.jsonl",
+        "dataset_slug": "throughput_1k",
+        "run_slug": "277-mtp-s5-k3-d4-throughput_1k"
+      },
+      {
+        "index": 278,
+        "variant_name": "mtp-s5-k3-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_2k.jsonl",
+        "dataset_slug": "throughput_2k",
+        "run_slug": "278-mtp-s5-k3-d4-throughput_2k"
+      },
+      {
+        "index": 279,
+        "variant_name": "mtp-s5-k3-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_32k.jsonl",
+        "dataset_slug": "throughput_32k",
+        "run_slug": "279-mtp-s5-k3-d4-throughput_32k"
+      },
+      {
+        "index": 280,
+        "variant_name": "mtp-s5-k3-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_8k.jsonl",
+        "dataset_slug": "throughput_8k",
+        "run_slug": "280-mtp-s5-k3-d4-throughput_8k"
+      },
+      {
+        "index": 281,
+        "variant_name": "mtp-s5-k3-d4",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/json_mode_eval/jsonl/train.jsonl",
+        "dataset_slug": "train",
+        "run_slug": "281-mtp-s5-k3-d4-train"
+      },
+      {
+        "index": 282,
+        "variant_name": "mtp-s5-k3-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_16k.jsonl",
+        "dataset_slug": "throughput_16k",
+        "run_slug": "282-mtp-s5-k3-d5-throughput_16k"
+      },
+      {
+        "index": 283,
+        "variant_name": "mtp-s5-k3-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_1k.jsonl",
+        "dataset_slug": "throughput_1k",
+        "run_slug": "283-mtp-s5-k3-d5-throughput_1k"
+      },
+      {
+        "index": 284,
+        "variant_name": "mtp-s5-k3-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_2k.jsonl",
+        "dataset_slug": "throughput_2k",
+        "run_slug": "284-mtp-s5-k3-d5-throughput_2k"
+      },
+      {
+        "index": 285,
+        "variant_name": "mtp-s5-k3-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_32k.jsonl",
+        "dataset_slug": "throughput_32k",
+        "run_slug": "285-mtp-s5-k3-d5-throughput_32k"
+      },
+      {
+        "index": 286,
+        "variant_name": "mtp-s5-k3-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_8k.jsonl",
+        "dataset_slug": "throughput_8k",
+        "run_slug": "286-mtp-s5-k3-d5-throughput_8k"
+      },
+      {
+        "index": 287,
+        "variant_name": "mtp-s5-k3-d5",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/json_mode_eval/jsonl/train.jsonl",
+        "dataset_slug": "train",
+        "run_slug": "287-mtp-s5-k3-d5-train"
+      },
+      {
+        "index": 288,
+        "variant_name": "mtp-s5-k3-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_16k.jsonl",
+        "dataset_slug": "throughput_16k",
+        "run_slug": "288-mtp-s5-k3-d6-throughput_16k"
+      },
+      {
+        "index": 289,
+        "variant_name": "mtp-s5-k3-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_1k.jsonl",
+        "dataset_slug": "throughput_1k",
+        "run_slug": "289-mtp-s5-k3-d6-throughput_1k"
+      },
+      {
+        "index": 290,
+        "variant_name": "mtp-s5-k3-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_2k.jsonl",
+        "dataset_slug": "throughput_2k",
+        "run_slug": "290-mtp-s5-k3-d6-throughput_2k"
+      },
+      {
+        "index": 291,
+        "variant_name": "mtp-s5-k3-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_32k.jsonl",
+        "dataset_slug": "throughput_32k",
+        "run_slug": "291-mtp-s5-k3-d6-throughput_32k"
+      },
+      {
+        "index": 292,
+        "variant_name": "mtp-s5-k3-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/speed_bench/jsonl/throughput_8k.jsonl",
+        "dataset_slug": "throughput_8k",
+        "run_slug": "292-mtp-s5-k3-d6-throughput_8k"
+      },
+      {
+        "index": 293,
+        "variant_name": "mtp-s5-k3-d6",
+        "variant_extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ],
+        "dataset_path": "examples/json_mode_eval/jsonl/train.jsonl",
+        "dataset_slug": "train",
+        "run_slug": "293-mtp-s5-k3-d6-train"
+      }
+    ]
+  variants.json: |-
+    [
+      {
+        "name": "baseline",
+        "extra_args": [
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ]
+      },
+      {
+        "name": "mtp-s2-k1-d3",
+        "extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ]
+      },
+      {
+        "name": "mtp-s2-k1-d4",
+        "extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ]
+      },
+      {
+        "name": "mtp-s2-k1-d5",
+        "extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ]
+      },
+      {
+        "name": "mtp-s2-k1-d6",
+        "extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ]
+      },
+      {
+        "name": "mtp-s2-k2-d3",
+        "extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ]
+      },
+      {
+        "name": "mtp-s2-k2-d4",
+        "extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ]
+      },
+      {
+        "name": "mtp-s2-k2-d5",
+        "extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ]
+      },
+      {
+        "name": "mtp-s2-k2-d6",
+        "extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ]
+      },
+      {
+        "name": "mtp-s2-k3-d3",
+        "extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ]
+      },
+      {
+        "name": "mtp-s2-k3-d4",
+        "extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ]
+      },
+      {
+        "name": "mtp-s2-k3-d5",
+        "extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ]
+      },
+      {
+        "name": "mtp-s2-k3-d6",
+        "extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "2",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ]
+      },
+      {
+        "name": "mtp-s3-k1-d3",
+        "extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ]
+      },
+      {
+        "name": "mtp-s3-k1-d4",
+        "extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ]
+      },
+      {
+        "name": "mtp-s3-k1-d5",
+        "extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ]
+      },
+      {
+        "name": "mtp-s3-k1-d6",
+        "extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ]
+      },
+      {
+        "name": "mtp-s3-k2-d3",
+        "extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ]
+      },
+      {
+        "name": "mtp-s3-k2-d4",
+        "extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ]
+      },
+      {
+        "name": "mtp-s3-k2-d5",
+        "extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ]
+      },
+      {
+        "name": "mtp-s3-k2-d6",
+        "extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ]
+      },
+      {
+        "name": "mtp-s3-k3-d3",
+        "extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ]
+      },
+      {
+        "name": "mtp-s3-k3-d4",
+        "extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ]
+      },
+      {
+        "name": "mtp-s3-k3-d5",
+        "extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ]
+      },
+      {
+        "name": "mtp-s3-k3-d6",
+        "extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "3",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ]
+      },
+      {
+        "name": "mtp-s4-k1-d3",
+        "extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ]
+      },
+      {
+        "name": "mtp-s4-k1-d4",
+        "extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ]
+      },
+      {
+        "name": "mtp-s4-k1-d5",
+        "extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ]
+      },
+      {
+        "name": "mtp-s4-k1-d6",
+        "extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ]
+      },
+      {
+        "name": "mtp-s4-k2-d3",
+        "extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ]
+      },
+      {
+        "name": "mtp-s4-k2-d4",
+        "extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ]
+      },
+      {
+        "name": "mtp-s4-k2-d5",
+        "extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ]
+      },
+      {
+        "name": "mtp-s4-k2-d6",
+        "extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ]
+      },
+      {
+        "name": "mtp-s4-k3-d3",
+        "extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ]
+      },
+      {
+        "name": "mtp-s4-k3-d4",
+        "extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ]
+      },
+      {
+        "name": "mtp-s4-k3-d5",
+        "extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ]
+      },
+      {
+        "name": "mtp-s4-k3-d6",
+        "extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "4",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ]
+      },
+      {
+        "name": "mtp-s5-k1-d3",
+        "extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ]
+      },
+      {
+        "name": "mtp-s5-k1-d4",
+        "extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ]
+      },
+      {
+        "name": "mtp-s5-k1-d5",
+        "extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ]
+      },
+      {
+        "name": "mtp-s5-k1-d6",
+        "extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "1",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ]
+      },
+      {
+        "name": "mtp-s5-k2-d3",
+        "extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ]
+      },
+      {
+        "name": "mtp-s5-k2-d4",
+        "extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ]
+      },
+      {
+        "name": "mtp-s5-k2-d5",
+        "extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ]
+      },
+      {
+        "name": "mtp-s5-k2-d6",
+        "extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "2",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ]
+      },
+      {
+        "name": "mtp-s5-k3-d3",
+        "extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "3",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ]
+      },
+      {
+        "name": "mtp-s5-k3-d4",
+        "extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "4",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ]
+      },
+      {
+        "name": "mtp-s5-k3-d5",
+        "extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "5",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ]
+      },
+      {
+        "name": "mtp-s5-k3-d6",
+        "extra_args": [
+          "--speculative-algorithm",
+          "EAGLE",
+          "--speculative-num-steps",
+          "5",
+          "--speculative-eagle-topk",
+          "3",
+          "--speculative-num-draft-tokens",
+          "6",
+          "--mamba-scheduler-strategy",
+          "extra_buffer"
+        ]
+      }
+    ]
+---
+apiVersion: batch/v1
+kind: Job
+metadata:
+  name: piccolo-k8s-sweep-example-job
+  namespace: specdec
+  labels:
+    app.kubernetes.io/name: batchbench-sweep
+    app.kubernetes.io/component: benchmark
+    batchbench.doubleword.ai/sweep-name: piccolo-k8s-sweep-example
+  annotations:
+    batchbench.doubleword.ai/image: tytn/batchbench-sglang-sweep:example
+    batchbench.doubleword.ai/created-at: '2026-04-27T16:56:23+00:00'
+    batchbench.doubleword.ai/git-sha: 6c8524e0b4ea
+spec:
+  completionMode: Indexed
+  completions: 294
+  parallelism: 1
+  backoffLimit: 1
+  template:
+    metadata:
+      labels:
+        app.kubernetes.io/name: batchbench-sweep
+        app.kubernetes.io/component: benchmark
+        batchbench.doubleword.ai/sweep-name: piccolo-k8s-sweep-example
+    spec:
+      restartPolicy: Never
+      runtimeClassName: nvidia
+      priorityClassName: research
+      containers:
+      - name: benchmark
+        image: tytn/batchbench-sglang-sweep:example
+        imagePullPolicy: IfNotPresent
+        env:
+        - name: SWEEP_NAME
+          value: piccolo-k8s-sweep-example
+        - name: OUTPUT_ROOT
+          value: /outputs/piccolo-k8s-sweep-example
+        - name: COMBINATIONS_JSON
+          value: /config/combinations.json
+        - name: SHARED_RESULTS_CSV
+          value: /outputs/piccolo-k8s-sweep-example/results/summary.csv
+        - name: CSV_LOCK_PATH
+          value: /outputs/piccolo-k8s-sweep-example/results/summary.csv.lock
+        - name: REQUEST_TIMEOUT_SECS
+          value: '6000'
+        - name: REQUESTS_PER_USER
+          value: '1'
+        - name: IMAGE_REF
+          value: tytn/batchbench-sglang-sweep:example
+        - name: JOB_COMPLETION_INDEX
+          valueFrom:
+            fieldRef:
+              fieldPath: metadata.annotations['batch.kubernetes.io/job-completion-index']
+        - name: POD_NAME
+          valueFrom:
+            fieldRef:
+              fieldPath: metadata.name
+        - name: NODE_NAME
+          valueFrom:
+            fieldRef:
+              fieldPath: spec.nodeName
+        - name: SGLANG_USE_CUDA_IPC_TRANSPORT
+          value: '0'
+        - name: SGLANG_ENABLE_SPEC_V2
+          value: '1'
+        resources:
+          limits:
+            nvidia.com/gpu: '4'
+            cpu: '64'
+            memory: 512Gi
+          requests:
+            nvidia.com/gpu: '4'
+            cpu: '64'
+            memory: 512Gi
+        volumeMounts:
+        - name: config
+          mountPath: /config
+          readOnly: true
+        - name: outputs
+          mountPath: /outputs
+        - name: cache
+          mountPath: /root/.cache
+        - name: dshm
+          mountPath: /dev/shm
+      volumes:
+      - name: config
+        configMap:
+          name: piccolo-k8s-sweep-example-config
+      - name: outputs
+        hostPath:
+          path: /mnt/raid/batchbench
+          type: DirectoryOrCreate
+      - name: cache
+        hostPath:
+          path: /mnt/raid/cache
+          type: DirectoryOrCreate
+      - name: dshm
+        emptyDir:
+          medium: Memory
+          sizeLimit: 16Gi
diff --git a/docker/Dockerfile.cu126 b/docker/Dockerfile.cu126
deleted file mode 100644
index 422c371..0000000
--- a/docker/Dockerfile.cu126
+++ /dev/null
@@ -1,105 +0,0 @@
-# ============================================================
-# Stage 1: NVSHMEM Builder
-# ============================================================
-FROM pytorch/pytorch:2.9.0-cuda12.6-cudnn9-devel AS deepep-builder
-
-# Add rdma libraries and build dependencies
-RUN apt-get update && apt-get install -y --no-install-recommends \
-    git build-essential ninja-build cmake pkg-config wget \
-    librdmacm-dev rdma-core libfabric-dev libibverbs-dev \
-    devscripts debhelper fakeroot \
- && rm -rf /var/lib/apt/lists/*
-
-# Build GDRCopy
-RUN cd /tmp && \
-    wget https://github.com/NVIDIA/gdrcopy/archive/refs/tags/v2.5.1.tar.gz && \
-    tar -xf v2.5.1.tar.gz && \
-    cd gdrcopy-2.5.1/packages/ && \
-    CUDA=/usr/local/cuda ./build-deb-packages.sh -t -k
-
-RUN apt-get update && apt-get install -y --no-install-recommends git \
- && dpkg -i /tmp/gdrcopy-2.5.1/packages/libgdrapi_*.deb \
- && rm -rf /var/lib/apt/lists/*
-
-# Build NVSHMEM - Need GDRCOPY and IBGDA support for DeepEP all2all kernels
-RUN set -eux; \
-    cd /tmp; \
-    wget https://developer.nvidia.com/downloads/assets/secure/nvshmem/nvshmem_src_3.2.5-1.txz; \
-    mkdir -p nvshmem_src_3.2.5-1; \
-    tar xf nvshmem_src_3.2.5-1.txz -C nvshmem_src_3.2.5-1; \
-    cd nvshmem_src_3.2.5-1/nvshmem_src; \
-    mkdir -p build; \
-    cd build; \
-    cmake \
-        -DNVSHMEM_PREFIX=/opt/nvshmem-3.2.5 \
-        -DCMAKE_CUDA_ARCHITECTURES=90a \
-        -DNVSHMEM_MPI_SUPPORT=0 \
-        -DNVSHMEM_PMIX_SUPPORT=0 \
-        -DNVSHMEM_USE_GDRCOPY=1 \
-        -DNVSHMEM_IBGDA_SUPPORT=1 \
-        -DNVSHMEM_BUILD_TESTS=1 \
-        -DNVSHMEM_BUILD_EXAMPLES=1 \
-        -DNVSHMEM_BUILD_HYDRA_LAUNCHER=1 \
-        -DNVSHMEM_BUILD_TXZ_PACKAGE=1 \
-        -G Ninja \
-        ..; \
-        ninja; \
-        ninja install; \
-        rm -rf /tmp/nvshmem_src_3.2.5-1.txz
-
-
-
-
-# ============================================================
-# Stage 2: vLLM/Python Dependencies (slow, cache separately)
-# ============================================================
-FROM pytorch/pytorch:2.9.0-cuda12.6-cudnn9-devel AS python-deps
-
-RUN apt-get update && apt-get install -y --no-install-recommends curl \
- && rm -rf /var/lib/apt/lists/*
-
-RUN curl -LsSf https://astral.sh/uv/install.sh | sh \
- && ln -s /root/.local/bin/uv /usr/local/bin/uv
-
-# flashinfer-cubin installs pre-compiled CUDA kernels for vLLM
-# Needed for FP8-MOE models
-RUN uv venv /opt/batchbench/.venv && \
-    . /opt/batchbench/.venv/bin/activate && \
-    uv pip install vllm --torch-backend=auto && \
-    uv pip install flashinfer-cubin flashinfer-python
-# ============================================================
-# Stage 3: Final Runtime Image
-# ============================================================
-FROM pytorch/pytorch:2.9.0-cuda12.6-cudnn9-devel AS runtime
-
-# Copy built artifacts from builder stages
-COPY --from=deepep-builder /opt/nvshmem-3.2.5 /opt/nvshmem-3.2.5
-COPY --from=deepep-builder /tmp/gdrcopy-2.5.1/packages/*.deb /tmp/gdrcopy/
-COPY --from=deepep-builder /opt/conda/lib/python3.11/site-packages/deep_ep* /opt/conda/lib/python3.11/site-packages/
-COPY --from=python-deps /opt/batchbench/.venv /opt/batchbench/.venv
-COPY --from=python-deps /root/.local/bin/uv /usr/local/bin/uv
-
-# Install minimal runtime dependencies
-RUN apt-get update && apt-get install -y --no-install-recommends \
-    librdmacm-dev rdma-core libfabric-dev libibverbs-dev \
-    git curl openssh-server librdmacm1 libibverbs1 libfabric1 \
- && dpkg -i /tmp/gdrcopy/libgdrapi_*.deb \
- && rm -rf /tmp/gdrcopy /var/lib/apt/lists/*
-
-# NVSHMEM Environment
-ENV NVSHMEM_HOME=/opt/nvshmem-3.2.5
-ENV LD_LIBRARY_PATH=${NVSHMEM_HOME}/lib:/usr/local/cuda/lib64:${LD_LIBRARY_PATH}
-ENV PATH="/opt/batchbench/.venv/bin:/opt/hydra/bin:/usr/local/cuda/bin:${PATH}"
-ENV VLLM_HAS_FLASHINFER_CUBIN=1
-
-# DeepEP installation - TODO: pull into the deepep-builder stage
-RUN git clone https://github.com/deepseek-ai/DeepEP /deepep && \
-    cd /deepep && \
-    NVSHMEM_DIR=/opt/nvshmem-3.2.5 python setup.py install
-
-# SSH configuration
-RUN mkdir -p /var/run/sshd /root/.ssh && \
-    chmod 700 /root/.ssh && \
-    sed -i 's/#PermitRootLogin prohibit-password/PermitRootLogin prohibit-password/' /etc/ssh/sshd_config && \
-    sed -i 's@session\s*required\s*pam_loginuid.so@session optional pam_loginuid.so@g' /etc/pam.d/sshd && \
-    ssh-keygen -A
diff --git a/docker/Dockerfile.cu129 b/docker/Dockerfile.cu129
deleted file mode 100644
index ed56297..0000000
--- a/docker/Dockerfile.cu129
+++ /dev/null
@@ -1,116 +0,0 @@
-# syntax=docker/dockerfile:1
-# ============================================================
-# Stage 1: NVSHMEM Builder
-# ============================================================
-FROM pytorch/pytorch:2.8.0-cuda12.9-cudnn9-devel AS deepep-builder
-
-# Add rdma libraries and build dependencies
-RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
-    --mount=type=cache,target=/var/lib/apt,sharing=locked \
-    apt-get update && apt-get install -y --no-install-recommends \
-    git build-essential ninja-build cmake pkg-config wget \
-    librdmacm-dev rdma-core libfabric-dev libibverbs-dev \
-    devscripts debhelper fakeroot \
- && rm -rf /var/lib/apt/lists/*
-
-# Build GDRCopy
-RUN cd /tmp && \
-    wget https://github.com/NVIDIA/gdrcopy/archive/refs/tags/v2.5.1.tar.gz && \
-    tar -xf v2.5.1.tar.gz && \
-    cd gdrcopy-2.5.1/packages/ && \
-    CUDA=/usr/local/cuda ./build-deb-packages.sh -t -k
-
-RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
-    --mount=type=cache,target=/var/lib/apt,sharing=locked \
-    apt-get update && apt-get install -y --no-install-recommends git \
- && dpkg -i /tmp/gdrcopy-2.5.1/packages/libgdrapi_*.deb
-
-# Build NVSHMEM - Need GDRCOPY and IBGDA support for DeepEP all2all kernels
-RUN set -eux; \
-    cd /tmp; \
-    wget https://developer.nvidia.com/downloads/assets/secure/nvshmem/nvshmem_src_3.2.5-1.txz; \
-    mkdir -p nvshmem_src_3.2.5-1; \
-    tar xf nvshmem_src_3.2.5-1.txz -C nvshmem_src_3.2.5-1; \
-    cd nvshmem_src_3.2.5-1/nvshmem_src; \
-    mkdir -p build; \
-    cd build; \
-    cmake \
-        -DNVSHMEM_PREFIX=/opt/nvshmem-3.2.5 \
-        -DCMAKE_CUDA_ARCHITECTURES=90a \
-        -DNVSHMEM_MPI_SUPPORT=0 \
-        -DNVSHMEM_PMIX_SUPPORT=0 \
-        -DNVSHMEM_USE_GDRCOPY=1 \
-        -DNVSHMEM_IBGDA_SUPPORT=1 \
-        -DNVSHMEM_BUILD_TESTS=0 \
-        -DNVSHMEM_BUILD_EXAMPLES=0 \
-        -DNVSHMEM_BUILD_HYDRA_LAUNCHER=1 \
-        -DNVSHMEM_BUILD_TXZ_PACKAGE=1 \
-        -G Ninja \
-        ..; \
-        ninja; \
-        ninja install; \
-        rm -rf /tmp/nvshmem_src_3.2.5-1.txz
-
-
-
-
-# ============================================================
-# Stage 2: vLLM/Python Dependencies (slow, cache separately)
-# ============================================================
-FROM pytorch/pytorch:2.8.0-cuda12.9-cudnn9-devel AS python-deps
-
-RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
-    --mount=type=cache,target=/var/lib/apt,sharing=locked \
-    apt-get update && apt-get install -y --no-install-recommends curl
-
-RUN curl -LsSf https://astral.sh/uv/install.sh | sh \
- && ln -s /root/.local/bin/uv /usr/local/bin/uv
-
-# flashinfer-cubin installs pre-compiled CUDA kernels for vLLM
-# Needed for FP8-MOE models
-RUN uv venv /opt/batchbench/.venv && \
-    . /opt/batchbench/.venv/bin/activate && \
-    uv pip install vllm --torch-backend=auto && \
-    uv pip install flashinfer-cubin flashinfer-python && \
-    uv pip install flashinfer-jit-cache --index-url https://flashinfer.ai/whl/cu129
-
-
-# ============================================================
-# Stage 3: Final Runtime Image
-# ============================================================
-FROM pytorch/pytorch:2.8.0-cuda12.9-cudnn9-devel AS runtime
-
-# Copy built artifacts from builder stages
-COPY --from=deepep-builder /opt/nvshmem-3.2.5 /opt/nvshmem-3.2.5
-COPY --from=deepep-builder /tmp/gdrcopy-2.5.1/packages/*.deb /tmp/gdrcopy/
-COPY --from=deepep-builder /opt/conda/lib/python3.11/site-packages/deep_ep* /opt/conda/lib/python3.11/site-packages/
-COPY --from=python-deps /opt/batchbench/.venv /opt/batchbench/.venv
-COPY --from=python-deps /root/.local/bin/uv /usr/local/bin/uv
-
-# Install minimal runtime dependencies
-RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
-    --mount=type=cache,target=/var/lib/apt,sharing=locked \
-    apt-get update && apt-get install -y --no-install-recommends \
-    librdmacm-dev rdma-core libfabric-dev libibverbs-dev \
-    git curl openssh-server librdmacm1 libibverbs1 libfabric1 \
- && dpkg -i /tmp/gdrcopy/libgdrapi_*.deb \
- && rm -rf /tmp/gdrcopy
-
-# NVSHMEM Environment
-ENV NVSHMEM_HOME=/opt/nvshmem-3.2.5
-ENV LD_LIBRARY_PATH=${NVSHMEM_HOME}/lib:/usr/local/cuda/lib64:${LD_LIBRARY_PATH}
-ENV PATH="/opt/batchbench/.venv/bin:/opt/hydra/bin:/usr/local/cuda/bin:${PATH}"
-ENV VLLM_HAS_FLASHINFER_CUBIN=1
-
-# DeepEP installation - TODO: pull into the deepep-builder stage
-RUN git clone https://github.com/deepseek-ai/DeepEP /deepep && \
-    cd /deepep && \
-    NVSHMEM_DIR=/opt/nvshmem-3.2.5 python setup.py install
-
-# SSH configuration
-RUN mkdir -p /var/run/sshd /root/.ssh && \
-    chmod 700 /root/.ssh && \
-    sed -i 's/#PermitRootLogin prohibit-password/PermitRootLogin prohibit-password/' /etc/ssh/sshd_config && \
-    sed -i 's@session\s*required\s*pam_loginuid.so@session optional pam_loginuid.so@g' /etc/pam.d/sshd && \
-    ssh-keygen -A
-
diff --git a/docker/Dockerfile.cu130 b/docker/Dockerfile.cu130
deleted file mode 100644
index c0f952f..0000000
--- a/docker/Dockerfile.cu130
+++ /dev/null
@@ -1,114 +0,0 @@
-# ============================================================
-# Stage 1: NVSHMEM Builder
-# ============================================================
-FROM pytorch/pytorch:2.9.0-cuda13.0-cudnn9-devel AS deepep-builder
-
-# Add rdma libraries and build dependencies
-RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
-    --mount=type=cache,target=/var/lib/apt,sharing=locked \
-    apt-get update && apt-get install -y --no-install-recommends \
-    git build-essential ninja-build cmake pkg-config wget \
-    librdmacm-dev rdma-core libfabric-dev libibverbs-dev \
-    devscripts debhelper fakeroot \
- && rm -rf /var/lib/apt/lists/*
-
-# Build GDRCopy
-RUN cd /tmp && \
-    wget https://github.com/NVIDIA/gdrcopy/archive/refs/tags/v2.5.1.tar.gz && \
-    tar -xf v2.5.1.tar.gz && \
-    cd gdrcopy-2.5.1/packages/ && \
-    CUDA=/usr/local/cuda ./build-deb-packages.sh -t -k
-
-RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
-    --mount=type=cache,target=/var/lib/apt,sharing=locked \
-    apt-get update && apt-get install -y --no-install-recommends git \
- && dpkg -i /tmp/gdrcopy-2.5.1/packages/libgdrapi_*.deb
-
-# Build NVSHMEM - Need GDRCOPY and IBGDA support for DeepEP all2all kernels
-RUN set -eux; \
-    cd /tmp; \
-    wget https://developer.nvidia.com/downloads/assets/secure/nvshmem/nvshmem_src_3.2.5-1.txz; \
-    mkdir -p nvshmem_src_3.2.5-1; \
-    tar xf nvshmem_src_3.2.5-1.txz -C nvshmem_src_3.2.5-1; \
-    cd nvshmem_src_3.2.5-1/nvshmem_src; \
-    mkdir -p build; \
-    cd build; \
-    cmake \
-        -DNVSHMEM_PREFIX=/opt/nvshmem-3.2.5 \
-        -DCMAKE_CUDA_ARCHITECTURES=90a \
-        -DNVSHMEM_MPI_SUPPORT=0 \
-        -DNVSHMEM_PMIX_SUPPORT=0 \
-        -DNVSHMEM_USE_GDRCOPY=1 \
-        -DNVSHMEM_IBGDA_SUPPORT=1 \
-        -DNVSHMEM_BUILD_TESTS=0 \
-        -DNVSHMEM_BUILD_EXAMPLES=0 \
-        -DNVSHMEM_BUILD_HYDRA_LAUNCHER=1 \
-        -DNVSHMEM_BUILD_TXZ_PACKAGE=1 \
-        -G Ninja \
-        ..; \
-        ninja; \
-        ninja install; \
-        rm -rf /tmp/nvshmem_src_3.2.5-1.txz
-
-
-
-
-# ============================================================
-# Stage 2: vLLM/Python Dependencies (slow, cache separately)
-# ============================================================
-FROM pytorch/pytorch:2.9.0-cuda13.0-cudnn9-devel AS python-deps
-
-RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
-    --mount=type=cache,target=/var/lib/apt,sharing=locked \
-    apt-get update && apt-get install -y --no-install-recommends curl
-
-RUN curl -LsSf https://astral.sh/uv/install.sh | sh \
- && ln -s /root/.local/bin/uv /usr/local/bin/uv
-
-# flashinfer-cubin installs pre-compiled CUDA kernels for vLLM
-# Needed for FP8-MOE models
-RUN uv venv /opt/batchbench/.venv && \
-    . /opt/batchbench/.venv/bin/activate && \
-    uv pip install vllm --torch-backend=auto && \
-    uv pip install flashinfer-cubin 
-
-# ============================================================
-# Stage 3: Final Runtime Image
-# ============================================================
-FROM pytorch/pytorch:2.9.0-cuda13.0-cudnn9-devel AS runtime
-
-# Copy built artifacts from builder stages
-COPY --from=deepep-builder /opt/nvshmem-3.2.5 /opt/nvshmem-3.2.5
-COPY --from=deepep-builder /tmp/gdrcopy-2.5.1/packages/*.deb /tmp/gdrcopy/
-COPY --from=deepep-builder /opt/conda/lib/python3.11/site-packages/deep_ep* /opt/conda/lib/python3.11/site-packages/
-COPY --from=python-deps /opt/batchbench/.venv /opt/batchbench/.venv
-COPY --from=python-deps /root/.local/bin/uv /usr/local/bin/uv
-
-# Install minimal runtime dependencies
-RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
-    --mount=type=cache,target=/var/lib/apt,sharing=locked \
-    apt-get update && apt-get install -y --no-install-recommends \
-    librdmacm-dev rdma-core libfabric-dev libibverbs-dev \
-    git curl openssh-server librdmacm1 libibverbs1 libfabric1 \
- && dpkg -i /tmp/gdrcopy/libgdrapi_*.deb \
- && rm -rf /tmp/gdrcopy
-
-# Environment
-ENV NVSHMEM_HOME=/opt/nvshmem-3.2.5
-ENV LD_LIBRARY_PATH=${NVSHMEM_HOME}/lib:${LD_LIBRARY_PATH}
-ENV PATH="/opt/batchbench/.venv/bin:/opt/hydra/bin:${PATH}"
-ENV VLLM_HAS_FLASHINFER_CUBIN=1
-
-# DeepEP installation - TODO: pull into the deepep-builder stage
-RUN git clone https://github.com/deepseek-ai/DeepEP /deepep && \
-    cd /deepep && \
-    NVSHMEM_DIR=/opt/nvshmem-3.2.5 python setup.py install
-
-# SSH configuration
-RUN mkdir -p /var/run/sshd /root/.ssh && \
-    chmod 700 /root/.ssh && \
-    sed -i 's/#PermitRootLogin prohibit-password/PermitRootLogin prohibit-password/' /etc/ssh/sshd_config && \
-    sed -i 's@session\s*required\s*pam_loginuid.so@session optional pam_loginuid.so@g' /etc/pam.d/sshd && \
-    ssh-keygen -A
-
-    
\ No newline at end of file
diff --git a/docker/run_vllm.sh b/docker/run_vllm.sh
deleted file mode 100644
index 859196c..0000000
--- a/docker/run_vllm.sh
+++ /dev/null
@@ -1,56 +0,0 @@
-#!/bin/bash
-
-export HF_HOME="/data"
-# uv pip list
-
-# Start vLLM server
-vllm serve \
-    --model Qwen/Qwen3-30B-A3B-Instruct-2507-FP8 \
-    --max-model-len 131072 \
-    --gpu-memory-utilization 0.9 \
-    --trust-remote-code \
-    --host 0.0.0.0 \
-    --port 8000
-
-# VLLM_HAS_FLASHINFER_CUBIN=1 \
-# VLLM_USE_FLASHINFER_MOE_FP8=1 \
-# VLLM_FLASHINFER_MOE_BACKEND=throughput \
-# vllm serve Qwen/Qwen3-30B-A3B-Instruct-2507-FP8 \
-#   --gpu-memory-utilization 0.9 \
-#   --trust-remote-code \
-#   --host 0.0.0.0 \
-#   --port 8000 \
-#   --enable-expert-parallel \
-#   --tensor-parallel-size 1 \
-#   --data-parallel-size 2
-
-# # Most verbose logging
-# VLLM_LOGGING_LEVEL=DEBUG \
-# VLLM_TRACE_FUNCTION=1 \
-# VLLM_HAS_FLASHINFER_CUBIN=1 \
-# VLLM_USE_FLASHINFER_MOE_FP8=1 \
-# VLLM_FLASHINFER_MOE_BACKEND=throughput \
-# vllm serve Qwen/Qwen3-30B-A3B-Instruct-2507-FP8 \
-#   --gpu-memory-utilization 0.9 \
-#   --trust-remote-code \
-#   --host 0.0.0.0 \
-#   --port 8000 \
-#   --enable-expert-parallel \
-#   --tensor-parallel-size 1 \
-#   --data-parallel-size 2 \
-# #   --all2all-backend deepep_high_throughput \
-#   --compilation-mode NONE \
-#   --cudagraph-mode FULL
-
-# VLLM_LOGGING_LEVEL=DEBUG \
-# VLLM_TRACE_FUNCTION=1 \
-# vllm serve Qwen/Qwen3-30B-A3B-Instruct-2507-FP8 \
-#   --gpu-memory-utilization 0.9 \
-#   --trust-remote-code \
-#   --host 0.0.0.0 \
-#   --port 8000 \
-#   --enable-expert-parallel \
-#   --tensor-parallel-size 1 \
-#   --data-parallel-size 2 \
-#   --compilation-config '{"mode": 0, "cudagraph_mode": "full"}' \
-#   --all2all-backend deepep_high_throughput
\ No newline at end of file
diff --git a/rust/README.md b/rust/README.md
index 4b4a762..74a063e 100644
--- a/rust/README.md
+++ b/rust/README.md
@@ -10,7 +10,7 @@ The repository includes a ready-to-go CLI that reads prompts from a JSONL file (
 
 ```
 cargo run --bin batchbench -- \
-  --jsonl data/requests.jsonl \
+  --dataset-jsonl data/requests.jsonl \
   --users 32 \
   --model gpt-4o-mini \
   --host https://api.openai.com \
@@ -21,9 +21,9 @@ By default every user issues a single request, so make sure the JSONL file conta
 
 The CLI reuses the library entry points under the hood and prints a `BenchmarkReport` when the run completes.
 
-- `--jsonl <path>` Path to the JSONL input whose objects expose a `text` field (required).
-- `--users <num>` Number of concurrent workers; defaults to the JSONL record count.
-- `--model <name>` Model identifier injected into every request; default `gpt-4o-mini`.
+- `--dataset-jsonl <path>` Path to the JSONL input. Each non-empty line may contain a `text` field, a `body` object, or a full request body object. The legacy alias `--jsonl` is also accepted.
+- `--users <num>` Number of concurrent workers. For generated prompts this defaults to `1`; for JSONL datasets this defaults to as many complete request rounds as the dataset can provide for the selected `--requests-per-user`.
+- `--model <name>` Model identifier injected into every request, overriding any `model` field in JSONL request bodies; default `gpt-4o-mini`.
 - `--host <url>` Base host for the API, including scheme; default `https://api.openai.com`.
 - `--endpoint <path-or-url>` Endpoint path or full URL; default `/v1/chat/completions`.
 - `--requests-per-user <num>` Iterations each worker performs before stopping; default `1`.
@@ -35,6 +35,8 @@ The CLI reuses the library entry points under the hood and prints a `BenchmarkRe
 - `--output-tokens <num>` Force completions to emit exactly `<num>` tokens (sets `max_completion_tokens` and enables `nvext.ignore_eos`).
 - `--output-vary <num>` When used with `--output-tokens`, add a per-request uniform variation in `[-num, +num]` tokens (lower bounded at 1).
 - `--sglang` Use `min_new_tokens` and `max_new_tokens` for output token constraints (default uses `min_tokens` and `max_tokens`).
+- `--enable-json-decoding` Enable JSON constrained decoding. With `--sglang`, requests include `response_format: {"type": "json_object"}`. Without `--sglang`, requests include vLLM `structured_outputs.json_object: true`, equivalent to passing that value through the OpenAI SDK's `extra_body`.
+- `--metrics-output-dir <path>` Enable Prometheus scraping from the benchmarked server and write raw scrapes, parsed samples, metadata, and a summary to `<path>`. The default metrics endpoint is `/metrics`; override it with `--metrics-endpoint`.
 
 ### Library Usage
 
@@ -102,6 +104,7 @@ The CLI reuses the library entry points under the hood and prints a `BenchmarkRe
 - Derived throughput metrics, such as tokens-per-second and requests-per-second.
 - Latency percentiles (p50/p90/p99) across all successful requests.
 - A list of `FailureRecord` items capturing the last error seen per failed request.
+- Optional `MetricsRunReport` scrape counts and artifact location when metrics scraping is enabled.
 
 Use these fields to build dashboards, generate CSVs, or trigger alerts after a run.
 
diff --git a/rust/src/cli.rs b/rust/src/cli.rs
index 887a171..3837996 100644
--- a/rust/src/cli.rs
+++ b/rust/src/cli.rs
@@ -1,11 +1,12 @@
 use std::ffi::OsString;
 use std::fs;
+use std::io::{BufRead, BufReader};
 use std::path::Path;
 use std::time::Duration;
 
 use crate::{
-    generate_requests, run_benchmark, BenchmarkConfig, BenchmarkReport, DistMode, GenerateOptions,
-    RequestEntry, RunMode,
+    generate_requests, resolve_metrics_endpoint, run_benchmark, BenchmarkConfig, BenchmarkReport,
+    DistMode, GenerateOptions, MetricsConfig, RequestEntry, RunMode,
 };
 use anyhow::{anyhow, Context, Result};
 use clap::Parser;
@@ -41,11 +42,19 @@ struct CsvResult {
     output_lognorm_sigma: Option<f64>,
     output_lognorm_max: Option<usize>,
     sglang: bool,
+    enable_json_decoding: bool,
+    qwen35_disable_thinking: bool,
     request_timeout_secs: u64,
     max_retries: usize,
     retry_delay_ms: u64,
     host: String,
     endpoint: String,
+    metrics_enabled: bool,
+    metrics_output_dir: Option<String>,
+    metrics_scrape_count: Option<u64>,
+    metrics_scrape_error_count: Option<u64>,
+    metrics_first_timestamp: Option<String>,
+    metrics_last_timestamp: Option<String>,
 }
 
 #[derive(Parser, Debug)]
@@ -98,6 +107,11 @@ struct Args {
     #[arg(long = "gen-dist-mode", hide = true)]
     legacy_gen_dist_mode: Option<String>,
 
+    /// Path to a JSONL dataset. Each line may contain {"text": "..."},
+    /// {"body": {...}}, or a full request body object.
+    #[arg(long, alias = "jsonl")]
+    dataset_jsonl: Option<PathBuf>,
+
     /// Number of concurrent users to spawn (default: 1)
     #[arg(long)]
     users: Option<usize>,
@@ -166,6 +180,14 @@ struct Args {
     #[arg(long)]
     sglang: bool,
 
+    /// Enable JSON constrained decoding using provider-specific request fields
+    #[arg(long)]
+    enable_json_decoding: bool,
+
+    /// Add chat_template_kwargs.enable_thinking=false to each request
+    #[arg(long)]
+    qwen35_disable_thinking: bool,
+
     /// Enable verbose mode to print request/response details
     #[arg(long, short)]
     verbose: bool,
@@ -179,6 +201,26 @@ struct Args {
     #[arg(long)]
     results_csv: Option<PathBuf>,
 
+    /// Directory for Prometheus metrics artifacts; enables scraping when set
+    #[arg(long)]
+    metrics_output_dir: Option<PathBuf>,
+
+    /// Metrics endpoint path or full URL (default: /metrics)
+    #[arg(long, default_value = "/metrics")]
+    metrics_endpoint: String,
+
+    /// Metrics scrape interval in milliseconds
+    #[arg(long, default_value_t = 1000)]
+    metrics_interval_ms: u64,
+
+    /// Metrics scrape request timeout in milliseconds
+    #[arg(long, default_value_t = 2000)]
+    metrics_timeout_ms: u64,
+
+    /// Fail the benchmark if any metrics scrape fails
+    #[arg(long)]
+    metrics_fail_on_error: bool,
+
     /// Random seed for reproducible benchmarking (default: None)
     #[arg(long)]
     seed: Option<u64>,
@@ -245,6 +287,18 @@ async fn run(args: Args) -> Result<()> {
         || args.input_lognorm_median.is_some()
         || args.input_lognorm_sigma.is_some()
         || args.input_lognorm_max.is_some();
+    let dataset_jsonl = args.dataset_jsonl.clone();
+    let input_generation_requested = args.input_tokens > 0
+        || args.input_vary > 0
+        || input_lognorm_requested
+        || args.input_prefix_overlap != 0.0
+        || legacy_input_dist_mode.is_some();
+
+    if dataset_jsonl.is_some() && input_generation_requested {
+        return Err(anyhow!(
+            "--dataset-jsonl cannot be combined with input generation flags"
+        ));
+    }
 
     if args.input_tokens == 0 && args.input_vary > 0 {
         return Err(anyhow!("input-vary requires --input-tokens to be set"));
@@ -380,6 +434,15 @@ async fn run(args: Args) -> Result<()> {
         }
     };
 
+    if args.metrics_output_dir.is_some() {
+        if args.metrics_interval_ms == 0 {
+            return Err(anyhow!("metrics-interval-ms must be greater than zero"));
+        }
+        if args.metrics_timeout_ms == 0 {
+            return Err(anyhow!("metrics-timeout-ms must be greater than zero"));
+        }
+    }
+
     let api_key = args
         .api_key
         .or_else(|| std::env::var(&args.api_key_env).ok());
@@ -390,8 +453,60 @@ async fn run(args: Args) -> Result<()> {
         return Err(anyhow!("requests_per_user must be greater than zero"));
     }
 
-    // Always generate inline: size to users * requests_per_user
-    let user_count = args.users.unwrap_or(1);
+    let (mut request_bodies, dataset_label, dataset_is_generated) =
+        if let Some(dataset_path) = dataset_jsonl.as_ref() {
+            let requests = load_dataset_jsonl(dataset_path, &args.model)?;
+            let label = dataset_path.display().to_string();
+            (requests, label, false)
+        } else {
+            let user_count = args.users.unwrap_or(1);
+            if user_count == 0 {
+                return Err(anyhow!("users must be greater than zero"));
+            }
+            let total_requests = user_count
+                .checked_mul(requests_per_user)
+                .ok_or_else(|| anyhow!("users * requests_per_user overflowed"))?;
+
+            let target_tokens = if args.input_tokens > 0 {
+                Some(args.input_tokens)
+            } else {
+                None
+            };
+
+            let gen_opts = GenerateOptions {
+                count: total_requests,
+                prefix_overlap: args.input_prefix_overlap,
+                target_tokens,
+                token_tolerance: if args.input_tokens > 0 {
+                    Some(args.input_vary)
+                } else {
+                    None
+                },
+                tokenizer_model: args.model.clone(),
+                dist_mode: input_dist_mode,
+                dist_mu: input_dist_mu,
+                dist_median: input_dist_median,
+                dist_sigma: input_dist_sigma,
+                dist_max: input_dist_max,
+                seed: args.seed,
+            };
+
+            let requests = generate_requests(&gen_opts, &args.model)?;
+            let label = format!(
+                "generated (count={}, tokenizer={})",
+                total_requests, args.model
+            );
+            (requests, label, true)
+        };
+
+    let dataset_size = request_bodies.len();
+    let user_count = if dataset_is_generated {
+        args.users.unwrap_or(1)
+    } else if let Some(users) = args.users {
+        users
+    } else {
+        dataset_size / requests_per_user
+    };
     if user_count == 0 {
         return Err(anyhow!("users must be greater than zero"));
     }
@@ -399,32 +514,16 @@ async fn run(args: Args) -> Result<()> {
     let total_requests = user_count
         .checked_mul(requests_per_user)
         .ok_or_else(|| anyhow!("users * requests_per_user overflowed"))?;
+    if total_requests > dataset_size {
+        return Err(anyhow!(
+            "dataset contains {} request entries but benchmark needs {} (users {} * requests-per-user {})",
+            dataset_size,
+            total_requests,
+            user_count,
+            requests_per_user
+        ));
+    }
 
-    let target_tokens = if args.input_tokens > 0 {
-        Some(args.input_tokens)
-    } else {
-        None
-    };
-
-    let gen_opts = GenerateOptions {
-        count: total_requests,
-        prefix_overlap: args.input_prefix_overlap,
-        target_tokens,
-        token_tolerance: if args.input_tokens > 0 {
-            Some(args.input_vary)
-        } else {
-            None
-        },
-        tokenizer_model: args.model.clone(),
-        dist_mode: input_dist_mode,
-        dist_mu: input_dist_mu,
-        dist_median: input_dist_median,
-        dist_sigma: input_dist_sigma,
-        dist_max: input_dist_max,
-        seed: args.seed,
-    };
-
-    let mut request_bodies = generate_requests(&gen_opts, &args.model)?;
     let output_vary = if args.output_tokens.is_some() {
         Some(args.output_vary)
     } else {
@@ -437,11 +536,6 @@ async fn run(args: Args) -> Result<()> {
         args.seed,
         args.sglang,
     )?;
-    let dataset_label = format!(
-        "generated (count={}, tokenizer={})",
-        total_requests, args.model
-    );
-    let dataset_size = request_bodies.len();
 
     // Print input token histogram
     let input_tokens: Vec<usize> = request_bodies.iter().map(|r| r.input_tokens).collect();
@@ -459,7 +553,14 @@ async fn run(args: Args) -> Result<()> {
     println!("Dataset: {}", dataset_label);
     println!("Dataset size: {}", dataset_size);
     println!("Users: {}", user_count);
-    println!("Mode: Deterministic mapping m*N+n into generated dataset");
+    println!(
+        "Mode: Deterministic mapping m*N+n into {}",
+        if dataset_is_generated {
+            "generated dataset"
+        } else {
+            "dataset"
+        }
+    );
     println!("Requests per user: {}", requests_per_user);
     println!("Total requests: {}", total_requests);
     if let Some(tokens) = args.output_tokens {
@@ -495,8 +596,36 @@ async fn run(args: Args) -> Result<()> {
             "default (min_tokens/max_tokens)"
         }
     );
+    println!(
+        "JSON decoding: {}",
+        if args.enable_json_decoding {
+            if args.sglang {
+                "enabled (SGLang response_format)"
+            } else {
+                "enabled (vLLM structured_outputs)"
+            }
+        } else {
+            "disabled"
+        }
+    );
+    println!(
+        "Qwen 3.5 thinking: {}",
+        if args.qwen35_disable_thinking {
+            "disabled (chat_template_kwargs.enable_thinking=false)"
+        } else {
+            "unchanged"
+        }
+    );
     println!("Max retries: {}", args.max_retries);
     println!("Retry delay: {}ms", args.retry_delay_ms);
+    if let Some(metrics_output_dir) = args.metrics_output_dir.as_ref() {
+        let metrics_endpoint = resolve_metrics_endpoint(&args.host, &args.metrics_endpoint)?;
+        println!("Metrics endpoint: {}", metrics_endpoint);
+        println!("Metrics output: {}", metrics_output_dir.display());
+        println!("Metrics interval: {}ms", args.metrics_interval_ms);
+    } else {
+        println!("Metrics scraping: disabled");
+    }
     println!("===============================\n");
 
     let mode = RunMode::Finite { requests_per_user };
@@ -515,7 +644,9 @@ async fn run(args: Args) -> Result<()> {
     .with_retry(args.max_retries, Duration::from_millis(args.retry_delay_ms))
     .with_verbose(args.verbose)
     .with_dry_run(args.dry_run)
-    .with_sglang(args.sglang);
+    .with_sglang(args.sglang)
+    .with_json_decoding(args.enable_json_decoding)
+    .with_qwen35_disable_thinking(args.qwen35_disable_thinking);
 
     if let Some(seed) = args.seed {
         config = config.with_seed(seed);
@@ -527,6 +658,30 @@ async fn run(args: Args) -> Result<()> {
         config = config.with_output_lognorm(mu, sigma, max);
     }
 
+    if let Some(metrics_output_dir) = args.metrics_output_dir.as_ref() {
+        let metrics_endpoint = resolve_metrics_endpoint(&args.host, &args.metrics_endpoint)?;
+        config = config.with_metrics(MetricsConfig {
+            endpoint: metrics_endpoint,
+            output_dir: metrics_output_dir.clone(),
+            interval: Duration::from_millis(args.metrics_interval_ms),
+            timeout: Duration::from_millis(args.metrics_timeout_ms),
+            fail_on_error: args.metrics_fail_on_error,
+            metadata: json!({
+                "model": args.model.clone(),
+                "dataset_path": dataset_label.clone(),
+                "dataset_size": dataset_size,
+                "users": user_count,
+                "requests_per_user": requests_per_user,
+                "total_requests": total_requests,
+                "host": args.host.clone(),
+                "endpoint": endpoint_for_config.clone(),
+                "sglang": args.sglang,
+                "enable_json_decoding": args.enable_json_decoding,
+                "qwen35_disable_thinking": args.qwen35_disable_thinking,
+            }),
+        });
+    }
+
     let start_time = chrono::Utc::now();
     let report = run_benchmark(config).await?;
 
@@ -562,11 +717,31 @@ async fn run(args: Args) -> Result<()> {
             output_lognorm_sigma: args.output_lognorm_sigma,
             output_lognorm_max: args.output_lognorm_max,
             sglang: args.sglang,
+            enable_json_decoding: args.enable_json_decoding,
+            qwen35_disable_thinking: args.qwen35_disable_thinking,
             request_timeout_secs: args.request_timeout_secs,
             max_retries: args.max_retries,
             retry_delay_ms: args.retry_delay_ms,
             host: args.host.clone(),
             endpoint: endpoint_for_config.clone(),
+            metrics_enabled: report.metrics.is_some(),
+            metrics_output_dir: report
+                .metrics
+                .as_ref()
+                .map(|metrics| metrics.output_dir.clone()),
+            metrics_scrape_count: report.metrics.as_ref().map(|metrics| metrics.scrape_count),
+            metrics_scrape_error_count: report
+                .metrics
+                .as_ref()
+                .map(|metrics| metrics.scrape_error_count),
+            metrics_first_timestamp: report
+                .metrics
+                .as_ref()
+                .and_then(|metrics| metrics.first_timestamp.clone()),
+            metrics_last_timestamp: report
+                .metrics
+                .as_ref()
+                .and_then(|metrics| metrics.last_timestamp.clone()),
         };
 
         match write_results_csv(csv_path.as_path(), &record) {
@@ -593,6 +768,112 @@ fn parse_dist_mode(mode: &str) -> Result<DistMode> {
     }
 }
 
+fn load_dataset_jsonl(path: &Path, model: &str) -> Result<Vec<RequestEntry>> {
+    let file = fs::File::open(path)
+        .with_context(|| format!("failed to open dataset JSONL {}", path.display()))?;
+    let reader = BufReader::new(file);
+    let mut requests = Vec::new();
+
+    for (line_number, line) in reader.lines().enumerate() {
+        let line_number = line_number + 1;
+        let line = line.with_context(|| {
+            format!(
+                "failed to read line {} from dataset JSONL {}",
+                line_number,
+                path.display()
+            )
+        })?;
+        let trimmed = line.trim();
+        if trimmed.is_empty() {
+            continue;
+        }
+
+        let value: serde_json::Value = serde_json::from_str(trimmed).with_context(|| {
+            format!(
+                "failed to parse dataset JSONL {} line {} as JSON",
+                path.display(),
+                line_number
+            )
+        })?;
+        requests.push(dataset_value_to_request_entry(
+            value,
+            requests.len(),
+            line_number,
+            model,
+        )?);
+    }
+
+    if requests.is_empty() {
+        return Err(anyhow!("dataset JSONL {} is empty", path.display()));
+    }
+
+    Ok(requests)
+}
+
+fn dataset_value_to_request_entry(
+    value: serde_json::Value,
+    line_idx: usize,
+    line_number: usize,
+    model: &str,
+) -> Result<RequestEntry> {
+    let input_tokens = value
+        .get("input_tokens")
+        .and_then(|v| v.as_u64())
+        .map(|v| v as usize)
+        .unwrap_or(0);
+
+    if let Some(mut body) = value.get("body").cloned() {
+        if !body.is_object() {
+            return Err(anyhow!(
+                "dataset line {} has a body field, but body must be a JSON object",
+                line_number
+            ));
+        }
+        apply_request_model(&mut body, model);
+        return Ok(RequestEntry {
+            body,
+            line_idx,
+            input_tokens,
+        });
+    }
+
+    if let Some(text) = value.get("text").and_then(|v| v.as_str()) {
+        return Ok(RequestEntry {
+            body: json!({
+                "messages": [
+                    {"role": "user", "content": text}
+                ],
+                "model": model,
+            }),
+            line_idx,
+            input_tokens,
+        });
+    }
+
+    let mut body = value;
+    let Some(map) = body.as_object_mut() else {
+        return Err(anyhow!(
+            "dataset line {} must be a JSON object, or contain a body object or text string",
+            line_number
+        ));
+    };
+    map.remove("input_tokens");
+    map.remove("line_idx");
+    map.insert("model".to_string(), json!(model));
+
+    Ok(RequestEntry {
+        body,
+        line_idx,
+        input_tokens,
+    })
+}
+
+fn apply_request_model(body: &mut serde_json::Value, model: &str) {
+    if let Some(map) = body.as_object_mut() {
+        map.insert("model".to_string(), json!(model));
+    }
+}
+
 /// Apply output token settings to generated request bodies, honoring optional variation and seed.
 fn apply_output_tokens(
     bodies: &mut [RequestEntry],
@@ -694,6 +975,13 @@ fn print_summary(report: &BenchmarkReport) -> Result<()> {
         }
     }
 
+    if let Some(metrics) = report.metrics.as_ref() {
+        println!(
+            "Metrics: {} scrape(s), {} error(s), artifacts at {}",
+            metrics.scrape_count, metrics.scrape_error_count, metrics.output_dir
+        );
+    }
+
     Ok(())
 }
 
@@ -763,6 +1051,85 @@ fn print_histogram(label: &str, data: &[usize], bins: usize, bar_width: usize) {
     }
 }
 
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn dataset_text_line_becomes_chat_request() {
+        let entry =
+            dataset_value_to_request_entry(json!({"text": "hello", "input_tokens": 3}), 0, 1, "m")
+                .unwrap();
+
+        assert_eq!(entry.line_idx, 0);
+        assert_eq!(entry.input_tokens, 3);
+        assert_eq!(entry.body["model"], "m");
+        assert_eq!(entry.body["messages"][0]["content"], "hello");
+    }
+
+    #[test]
+    fn dataset_body_line_uses_body_object_and_overrides_model() {
+        let entry = dataset_value_to_request_entry(
+            json!({"body": {"model": "x", "prompt": "hello"}, "input_tokens": 5}),
+            4,
+            5,
+            "runtime-model",
+        )
+        .unwrap();
+
+        assert_eq!(entry.line_idx, 4);
+        assert_eq!(entry.input_tokens, 5);
+        assert_eq!(
+            entry.body,
+            json!({"model": "runtime-model", "prompt": "hello"})
+        );
+    }
+
+    #[test]
+    fn dataset_body_line_without_model_gets_runtime_model() {
+        let entry = dataset_value_to_request_entry(
+            json!({"body": {"messages": []}, "input_tokens": 5}),
+            4,
+            5,
+            "runtime-model",
+        )
+        .unwrap();
+
+        assert_eq!(
+            entry.body,
+            json!({"model": "runtime-model", "messages": []})
+        );
+    }
+
+    #[test]
+    fn dataset_full_request_line_drops_metadata_fields_and_overrides_model() {
+        let entry = dataset_value_to_request_entry(
+            json!({"model": "x", "messages": [], "input_tokens": 7, "line_idx": 99}),
+            1,
+            2,
+            "runtime-model",
+        )
+        .unwrap();
+
+        assert_eq!(entry.input_tokens, 7);
+        assert_eq!(
+            entry.body,
+            json!({"model": "runtime-model", "messages": []})
+        );
+    }
+
+    #[test]
+    fn parse_args_accepts_qwen35_disable_thinking_flag() {
+        let parsed = parse_args(["batchbench", "--qwen35-disable-thinking"]).unwrap();
+
+        let ParsedArgs::Ready(args) = parsed else {
+            panic!("expected parsed args");
+        };
+
+        assert!(args.qwen35_disable_thinking);
+    }
+}
+
 fn format_latency(latency: Option<Duration>) -> String {
     match latency {
         Some(value) => format!("{:.2}", value.as_secs_f64() * 1000.0),
diff --git a/rust/src/config.rs b/rust/src/config.rs
index 2a89f67..1002c10 100644
--- a/rust/src/config.rs
+++ b/rust/src/config.rs
@@ -5,6 +5,8 @@ use reqwest::header::{HeaderMap, HeaderName, HeaderValue, AUTHORIZATION, CONTENT
 use reqwest::Url;
 use serde_json::Value;
 
+use crate::metrics::MetricsConfig;
+
 #[derive(Clone, Debug)]
 pub struct RequestEntry {
     pub body: Value,
@@ -40,6 +42,12 @@ pub struct BenchmarkConfig {
     pub dry_run: bool,
     /// Use SGLang sampling parameter names for output-token constraints
     pub sglang: bool,
+    /// Add provider-specific JSON constrained decoding fields to each request
+    pub enable_json_decoding: bool,
+    /// Disable Qwen 3.5 thinking via chat_template_kwargs.enable_thinking=false
+    pub qwen35_disable_thinking: bool,
+    /// Optional Prometheus metrics scraping configuration
+    pub metrics: Option<MetricsConfig>,
 }
 
 impl BenchmarkConfig {
@@ -100,6 +108,9 @@ impl BenchmarkConfig {
             seed: None,
             dry_run: false,
             sglang: false,
+            enable_json_decoding: false,
+            qwen35_disable_thinking: false,
+            metrics: None,
         })
     }
 
@@ -143,6 +154,21 @@ impl BenchmarkConfig {
         self
     }
 
+    pub fn with_json_decoding(mut self, enable_json_decoding: bool) -> Self {
+        self.enable_json_decoding = enable_json_decoding;
+        self
+    }
+
+    pub fn with_qwen35_disable_thinking(mut self, qwen35_disable_thinking: bool) -> Self {
+        self.qwen35_disable_thinking = qwen35_disable_thinking;
+        self
+    }
+
+    pub fn with_metrics(mut self, metrics: MetricsConfig) -> Self {
+        self.metrics = Some(metrics);
+        self
+    }
+
     pub fn add_header(mut self, name: HeaderName, value: HeaderValue) -> Self {
         self.headers.insert(name, value);
         self
diff --git a/rust/src/lib.rs b/rust/src/lib.rs
index ea1911a..7b3e14f 100644
--- a/rust/src/lib.rs
+++ b/rust/src/lib.rs
@@ -1,6 +1,7 @@
 mod cli;
 mod config;
 mod generator;
+mod metrics;
 #[cfg(feature = "python")]
 mod py_bindings;
 mod report;
@@ -9,6 +10,7 @@ mod runner;
 pub use cli::{run_from_argv, run_from_env};
 pub use config::{BenchmarkConfig, RequestEntry, RunMode};
 pub use generator::{generate_requests, DistMode, GenerateOptions};
+pub use metrics::{resolve_metrics_endpoint, MetricsConfig, MetricsRunReport};
 pub use report::{BenchmarkReport, FailureRecord};
 pub use reqwest::header::{HeaderMap, HeaderName, HeaderValue};
 pub use runner::run_benchmark;
diff --git a/rust/src/metrics.rs b/rust/src/metrics.rs
new file mode 100644
index 0000000..5a729d8
--- /dev/null
+++ b/rust/src/metrics.rs
@@ -0,0 +1,902 @@
+use std::collections::{BTreeMap, HashMap};
+use std::fs::{self, File};
+use std::io::{BufWriter, Write};
+use std::path::PathBuf;
+use std::time::{Duration, Instant};
+
+use anyhow::{anyhow, Context, Result};
+use chrono::{DateTime, Utc};
+use reqwest::{Client, StatusCode, Url};
+use serde::{Deserialize, Serialize};
+use serde_json::{json, Value};
+use tokio::sync::oneshot;
+use tokio::task::JoinHandle;
+
+#[derive(Clone, Debug)]
+pub struct MetricsConfig {
+    pub endpoint: Url,
+    pub output_dir: PathBuf,
+    pub interval: Duration,
+    pub timeout: Duration,
+    pub fail_on_error: bool,
+    pub metadata: Value,
+}
+
+#[derive(Clone, Debug, Default, Serialize, Deserialize)]
+pub struct MetricsRunReport {
+    pub output_dir: String,
+    pub scrape_count: u64,
+    pub scrape_error_count: u64,
+    pub first_timestamp: Option<String>,
+    pub last_timestamp: Option<String>,
+}
+
+pub struct MetricsCollectorHandle {
+    stop_tx: Option<oneshot::Sender<()>>,
+    join_handle: JoinHandle<Result<MetricsRunReport>>,
+}
+
+impl MetricsCollectorHandle {
+    pub async fn stop(mut self) -> Result<MetricsRunReport> {
+        if let Some(stop_tx) = self.stop_tx.take() {
+            let _ = stop_tx.send(());
+        }
+        self.join_handle
+            .await
+            .map_err(|err| anyhow!("metrics collector task failed: {}", err))?
+    }
+}
+
+pub async fn spawn_metrics_collector(config: MetricsConfig) -> Result<MetricsCollectorHandle> {
+    let (stop_tx, stop_rx) = oneshot::channel();
+    let (ready_tx, ready_rx) = oneshot::channel();
+    let join_handle = tokio::spawn(async move { collect_metrics(config, stop_rx, ready_tx).await });
+    ready_rx
+        .await
+        .map_err(|_| anyhow!("metrics collector exited before initial scrape"))??;
+    Ok(MetricsCollectorHandle {
+        stop_tx: Some(stop_tx),
+        join_handle,
+    })
+}
+
+async fn collect_metrics(
+    config: MetricsConfig,
+    mut stop_rx: oneshot::Receiver<()>,
+    ready_tx: oneshot::Sender<Result<()>>,
+) -> Result<MetricsRunReport> {
+    fs::create_dir_all(&config.output_dir).with_context(|| {
+        format!(
+            "failed to create metrics output directory {}",
+            config.output_dir.display()
+        )
+    })?;
+
+    let raw_path = config.output_dir.join("raw.promjsonl");
+    let samples_path = config.output_dir.join("samples.csv");
+    let metadata_path = config.output_dir.join("metadata.json");
+    let summary_path = config.output_dir.join("summary.json");
+
+    let mut raw_writer = BufWriter::new(
+        File::create(&raw_path)
+            .with_context(|| format!("failed to create metrics raw log {}", raw_path.display()))?,
+    );
+    let mut sample_writer = csv::Writer::from_path(&samples_path).with_context(|| {
+        format!(
+            "failed to create parsed metrics samples CSV {}",
+            samples_path.display()
+        )
+    })?;
+
+    let client = Client::builder()
+        .timeout(config.timeout)
+        .build()
+        .context("failed to construct metrics HTTP client")?;
+    let started_at = Utc::now();
+    let start = Instant::now();
+    let mut state = MetricsSummaryBuilder::new(config.output_dir.display().to_string());
+    let mut seq = 0u64;
+
+    if let Err(err) = scrape_once(
+        &client,
+        &config,
+        &mut raw_writer,
+        &mut sample_writer,
+        &mut state,
+        seq,
+        start,
+    )
+    .await
+    {
+        let message = err.to_string();
+        let _ = ready_tx.send(Err(anyhow!(message)));
+        return Err(err);
+    }
+    let _ = ready_tx.send(Ok(()));
+    seq += 1;
+
+    loop {
+        tokio::select! {
+            _ = tokio::time::sleep(config.interval) => {
+                scrape_once(
+                    &client,
+                    &config,
+                    &mut raw_writer,
+                    &mut sample_writer,
+                    &mut state,
+                    seq,
+                    start,
+                )
+                .await?;
+                seq += 1;
+            }
+            _ = &mut stop_rx => {
+                break;
+            }
+        }
+    }
+
+    scrape_once(
+        &client,
+        &config,
+        &mut raw_writer,
+        &mut sample_writer,
+        &mut state,
+        seq,
+        start,
+    )
+    .await?;
+
+    raw_writer.flush()?;
+    sample_writer.flush()?;
+
+    let ended_at = Utc::now();
+    let report = state.report();
+    let summary = state.summary();
+    let metadata = json!({
+        "schema_version": 1,
+        "benchmark_start_utc": started_at.to_rfc3339(),
+        "benchmark_end_utc": ended_at.to_rfc3339(),
+        "metrics_url": config.endpoint.as_str(),
+        "interval_ms": config.interval.as_millis(),
+        "timeout_ms": config.timeout.as_millis(),
+        "fail_on_error": config.fail_on_error,
+        "run": config.metadata,
+    });
+    fs::write(&metadata_path, serde_json::to_vec_pretty(&metadata)?)
+        .with_context(|| format!("failed to write {}", metadata_path.display()))?;
+    fs::write(&summary_path, serde_json::to_vec_pretty(&summary)?)
+        .with_context(|| format!("failed to write {}", summary_path.display()))?;
+
+    if config.fail_on_error && report.scrape_error_count > 0 {
+        return Err(anyhow!(
+            "metrics scraping recorded {} error(s); see {}",
+            report.scrape_error_count,
+            raw_path.display()
+        ));
+    }
+
+    Ok(report)
+}
+
+async fn scrape_once(
+    client: &Client,
+    config: &MetricsConfig,
+    raw_writer: &mut BufWriter<File>,
+    sample_writer: &mut csv::Writer<File>,
+    state: &mut MetricsSummaryBuilder,
+    seq: u64,
+    start: Instant,
+) -> Result<()> {
+    let timestamp = Utc::now();
+    let elapsed_ms = start.elapsed().as_millis() as u64;
+    let scrape_start = Instant::now();
+    let result = client.get(config.endpoint.clone()).send().await;
+    let scrape_duration_ms = scrape_start.elapsed().as_millis() as u64;
+
+    match result {
+        Ok(response) => {
+            let status = response.status();
+            let body = response.text().await.unwrap_or_else(|err| err.to_string());
+            if status.is_success() {
+                match parse_prometheus_text(&body) {
+                    Ok(samples) => {
+                        write_raw_record(
+                            raw_writer,
+                            RawScrapeRecord::success(
+                                seq,
+                                timestamp,
+                                elapsed_ms,
+                                scrape_duration_ms,
+                                status,
+                                &body,
+                            ),
+                        )?;
+                        state.record_success(timestamp, &samples);
+                        for sample in samples {
+                            sample_writer.serialize(SampleCsvRecord {
+                                seq,
+                                timestamp_utc: timestamp.to_rfc3339(),
+                                elapsed_ms,
+                                name: sample.name,
+                                metric_type: sample.metric_type,
+                                labels_json: serde_json::to_string(&sample.labels)?,
+                                value: sample.value.to_string(),
+                            })?;
+                        }
+                    }
+                    Err(err) => {
+                        state.record_error();
+                        write_raw_record(
+                            raw_writer,
+                            RawScrapeRecord::parse_error(
+                                seq,
+                                timestamp,
+                                elapsed_ms,
+                                scrape_duration_ms,
+                                status,
+                                body,
+                                err.to_string(),
+                            ),
+                        )?;
+                    }
+                }
+            } else {
+                state.record_error();
+                write_raw_record(
+                    raw_writer,
+                    RawScrapeRecord::http_error(
+                        seq,
+                        timestamp,
+                        elapsed_ms,
+                        scrape_duration_ms,
+                        status,
+                        body,
+                    ),
+                )?;
+            }
+        }
+        Err(err) => {
+            state.record_error();
+            write_raw_record(
+                raw_writer,
+                RawScrapeRecord::request_error(
+                    seq,
+                    timestamp,
+                    elapsed_ms,
+                    scrape_duration_ms,
+                    err.to_string(),
+                ),
+            )?;
+        }
+    }
+
+    raw_writer.flush()?;
+    sample_writer.flush()?;
+    Ok(())
+}
+
+fn write_raw_record(writer: &mut BufWriter<File>, record: RawScrapeRecord) -> Result<()> {
+    serde_json::to_writer(&mut *writer, &record)?;
+    writer.write_all(b"\n")?;
+    Ok(())
+}
+
+#[derive(Serialize)]
+struct RawScrapeRecord {
+    seq: u64,
+    timestamp_utc: String,
+    elapsed_ms: u64,
+    status: String,
+    http_status: Option<u16>,
+    scrape_duration_ms: u64,
+    body: Option<String>,
+    error: Option<String>,
+}
+
+impl RawScrapeRecord {
+    fn success(
+        seq: u64,
+        timestamp: DateTime<Utc>,
+        elapsed_ms: u64,
+        scrape_duration_ms: u64,
+        status: StatusCode,
+        body: &str,
+    ) -> Self {
+        Self {
+            seq,
+            timestamp_utc: timestamp.to_rfc3339(),
+            elapsed_ms,
+            status: "ok".to_string(),
+            http_status: Some(status.as_u16()),
+            scrape_duration_ms,
+            body: Some(body.to_string()),
+            error: None,
+        }
+    }
+
+    fn parse_error(
+        seq: u64,
+        timestamp: DateTime<Utc>,
+        elapsed_ms: u64,
+        scrape_duration_ms: u64,
+        status: StatusCode,
+        body: String,
+        error: String,
+    ) -> Self {
+        Self {
+            seq,
+            timestamp_utc: timestamp.to_rfc3339(),
+            elapsed_ms,
+            status: "parse_error".to_string(),
+            http_status: Some(status.as_u16()),
+            scrape_duration_ms,
+            body: Some(body),
+            error: Some(error),
+        }
+    }
+
+    fn http_error(
+        seq: u64,
+        timestamp: DateTime<Utc>,
+        elapsed_ms: u64,
+        scrape_duration_ms: u64,
+        status: StatusCode,
+        body: String,
+    ) -> Self {
+        Self {
+            seq,
+            timestamp_utc: timestamp.to_rfc3339(),
+            elapsed_ms,
+            status: "http_error".to_string(),
+            http_status: Some(status.as_u16()),
+            scrape_duration_ms,
+            body: Some(body),
+            error: None,
+        }
+    }
+
+    fn request_error(
+        seq: u64,
+        timestamp: DateTime<Utc>,
+        elapsed_ms: u64,
+        scrape_duration_ms: u64,
+        error: String,
+    ) -> Self {
+        Self {
+            seq,
+            timestamp_utc: timestamp.to_rfc3339(),
+            elapsed_ms,
+            status: "request_error".to_string(),
+            http_status: None,
+            scrape_duration_ms,
+            body: None,
+            error: Some(error),
+        }
+    }
+}
+
+#[derive(Serialize)]
+struct SampleCsvRecord {
+    seq: u64,
+    timestamp_utc: String,
+    elapsed_ms: u64,
+    name: String,
+    metric_type: String,
+    labels_json: String,
+    value: String,
+}
+
+#[derive(Clone, Debug)]
+struct ParsedSample {
+    name: String,
+    metric_type: String,
+    labels: BTreeMap<String, String>,
+    value: f64,
+}
+
+fn parse_prometheus_text(text: &str) -> Result<Vec<ParsedSample>> {
+    let mut metric_types = HashMap::new();
+    let mut samples = Vec::new();
+
+    for line in text.lines() {
+        let line = line.trim();
+        if line.is_empty() {
+            continue;
+        }
+        if let Some(rest) = line.strip_prefix("# TYPE ") {
+            let mut parts = rest.split_whitespace();
+            if let (Some(name), Some(metric_type)) = (parts.next(), parts.next()) {
+                metric_types.insert(name.to_string(), metric_type.to_string());
+            }
+            continue;
+        }
+        if line.starts_with('#') {
+            continue;
+        }
+
+        let sample = parse_sample_line(line, &metric_types)
+            .with_context(|| format!("failed to parse Prometheus sample line: {}", line))?;
+        samples.push(sample);
+    }
+
+    Ok(samples)
+}
+
+fn parse_sample_line(line: &str, metric_types: &HashMap<String, String>) -> Result<ParsedSample> {
+    let (name_and_labels, rest) =
+        split_sample_name_and_rest(line).ok_or_else(|| anyhow!("sample line missing value"))?;
+    let value_text = rest
+        .split_whitespace()
+        .next()
+        .ok_or_else(|| anyhow!("sample line missing value"))?;
+    let value = parse_prometheus_float(value_text)?;
+
+    let (name, labels) = if let Some(label_start) = name_and_labels.find('{') {
+        let label_end = name_and_labels
+            .rfind('}')
+            .ok_or_else(|| anyhow!("metric labels missing closing brace"))?;
+        let name = &name_and_labels[..label_start];
+        let labels_text = &name_and_labels[label_start + 1..label_end];
+        (name.to_string(), parse_labels(labels_text)?)
+    } else {
+        (name_and_labels.to_string(), BTreeMap::new())
+    };
+
+    let metric_type = infer_metric_type(&name, metric_types);
+    Ok(ParsedSample {
+        name,
+        metric_type,
+        labels,
+        value,
+    })
+}
+
+fn split_sample_name_and_rest(line: &str) -> Option<(&str, &str)> {
+    let mut in_labels = false;
+    let mut in_quotes = false;
+    let mut escaped = false;
+
+    for (idx, ch) in line.char_indices() {
+        if escaped {
+            escaped = false;
+            continue;
+        }
+        if in_quotes && ch == '\\' {
+            escaped = true;
+            continue;
+        }
+        if in_labels && ch == '"' {
+            in_quotes = !in_quotes;
+            continue;
+        }
+        if !in_quotes {
+            if ch == '{' {
+                in_labels = true;
+                continue;
+            }
+            if ch == '}' {
+                in_labels = false;
+                continue;
+            }
+            if ch.is_whitespace() {
+                return Some((&line[..idx], line[idx..].trim_start()));
+            }
+        }
+    }
+
+    None
+}
+
+fn infer_metric_type(name: &str, metric_types: &HashMap<String, String>) -> String {
+    if let Some(metric_type) = metric_types.get(name) {
+        return metric_type.clone();
+    }
+    for suffix in ["_bucket", "_sum", "_count"] {
+        if let Some(base) = name.strip_suffix(suffix) {
+            if let Some(metric_type) = metric_types.get(base) {
+                return metric_type.clone();
+            }
+        }
+    }
+    "untyped".to_string()
+}
+
+fn parse_labels(labels_text: &str) -> Result<BTreeMap<String, String>> {
+    let mut labels = BTreeMap::new();
+    let mut key = String::new();
+    let mut value = String::new();
+    let mut in_key = true;
+    let mut in_quotes = false;
+    let mut escaped = false;
+
+    for ch in labels_text.chars().chain(std::iter::once(',')) {
+        if escaped {
+            value.push(match ch {
+                'n' => '\n',
+                't' => '\t',
+                '\\' => '\\',
+                '"' => '"',
+                other => other,
+            });
+            escaped = false;
+            continue;
+        }
+        if in_quotes && ch == '\\' {
+            escaped = true;
+            continue;
+        }
+        if ch == '"' {
+            in_quotes = !in_quotes;
+            continue;
+        }
+        if !in_quotes && in_key && ch == '=' {
+            in_key = false;
+            continue;
+        }
+        if !in_quotes && ch == ',' {
+            let label_key = key.trim();
+            if !label_key.is_empty() {
+                labels.insert(label_key.to_string(), value.clone());
+            }
+            key.clear();
+            value.clear();
+            in_key = true;
+            continue;
+        }
+        if in_key {
+            key.push(ch);
+        } else {
+            value.push(ch);
+        }
+    }
+
+    if in_quotes {
+        return Err(anyhow!("unterminated label quote"));
+    }
+
+    Ok(labels)
+}
+
+fn parse_prometheus_float(value: &str) -> Result<f64> {
+    match value {
+        "NaN" | "nan" => Ok(f64::NAN),
+        "Inf" | "+Inf" | "inf" | "+inf" => Ok(f64::INFINITY),
+        "-Inf" | "-inf" => Ok(f64::NEG_INFINITY),
+        other => other
+            .parse::<f64>()
+            .with_context(|| format!("invalid Prometheus float: {}", value)),
+    }
+}
+
+#[derive(Default)]
+struct MetricsSummaryBuilder {
+    output_dir: String,
+    scrape_count: u64,
+    scrape_error_count: u64,
+    first_timestamp: Option<DateTime<Utc>>,
+    last_timestamp: Option<DateTime<Utc>>,
+    scalar_stats: BTreeMap<String, ScalarMetricStats>,
+    histogram_buckets: BTreeMap<String, BTreeMap<String, BucketStats>>,
+}
+
+impl MetricsSummaryBuilder {
+    fn new(output_dir: String) -> Self {
+        Self {
+            output_dir,
+            ..Self::default()
+        }
+    }
+
+    fn record_success(&mut self, timestamp: DateTime<Utc>, samples: &[ParsedSample]) {
+        self.scrape_count += 1;
+        self.first_timestamp.get_or_insert(timestamp);
+        self.last_timestamp = Some(timestamp);
+
+        let mut aggregated = BTreeMap::<(String, String), f64>::new();
+        let mut buckets = BTreeMap::<(String, String), f64>::new();
+
+        for sample in samples {
+            if !sample.value.is_finite() {
+                continue;
+            }
+            if sample.name.ends_with("_bucket") {
+                if let Some(le) = sample.labels.get("le") {
+                    let base = sample.name.trim_end_matches("_bucket").to_string();
+                    *buckets.entry((base, le.clone())).or_default() += sample.value;
+                }
+                continue;
+            }
+            if sample.name.ends_with("_sum") || sample.name.ends_with("_count") {
+                continue;
+            }
+            *aggregated
+                .entry((sample.name.clone(), sample.metric_type.clone()))
+                .or_default() += sample.value;
+        }
+
+        for ((name, metric_type), value) in aggregated {
+            self.scalar_stats
+                .entry(name)
+                .or_insert_with(|| ScalarMetricStats::new(metric_type))
+                .record(value);
+        }
+        for ((name, le), value) in buckets {
+            self.histogram_buckets
+                .entry(name)
+                .or_default()
+                .entry(le)
+                .or_default()
+                .record(value);
+        }
+    }
+
+    fn record_error(&mut self) {
+        self.scrape_error_count += 1;
+    }
+
+    fn report(&self) -> MetricsRunReport {
+        MetricsRunReport {
+            output_dir: self.output_dir.clone(),
+            scrape_count: self.scrape_count,
+            scrape_error_count: self.scrape_error_count,
+            first_timestamp: self.first_timestamp.map(|ts| ts.to_rfc3339()),
+            last_timestamp: self.last_timestamp.map(|ts| ts.to_rfc3339()),
+        }
+    }
+
+    fn summary(&self) -> Value {
+        let mut counters = serde_json::Map::new();
+        let mut gauges = serde_json::Map::new();
+        for (name, stats) in &self.scalar_stats {
+            match stats.metric_type.as_str() {
+                "counter" => {
+                    counters.insert(name.clone(), stats.counter_json());
+                }
+                "gauge" => {
+                    gauges.insert(name.clone(), stats.gauge_json());
+                }
+                _ => {}
+            }
+        }
+
+        let mut histograms = serde_json::Map::new();
+        for (name, buckets) in &self.histogram_buckets {
+            histograms.insert(name.clone(), histogram_json(buckets));
+        }
+
+        json!({
+            "schema_version": 1,
+            "scrape_count": self.scrape_count,
+            "scrape_error_count": self.scrape_error_count,
+            "first_timestamp": self.first_timestamp.map(|ts| ts.to_rfc3339()),
+            "last_timestamp": self.last_timestamp.map(|ts| ts.to_rfc3339()),
+            "counters": counters,
+            "gauges": gauges,
+            "histograms": histograms,
+        })
+    }
+}
+
+#[derive(Default)]
+struct ScalarMetricStats {
+    metric_type: String,
+    first: Option<f64>,
+    last: Option<f64>,
+    min: f64,
+    max: f64,
+    sum: f64,
+    count: u64,
+}
+
+impl ScalarMetricStats {
+    fn new(metric_type: String) -> Self {
+        Self {
+            metric_type,
+            ..Self::default()
+        }
+    }
+
+    fn record(&mut self, value: f64) {
+        self.first.get_or_insert(value);
+        self.last = Some(value);
+        if self.count == 0 {
+            self.min = value;
+            self.max = value;
+        } else {
+            self.min = self.min.min(value);
+            self.max = self.max.max(value);
+        }
+        self.sum += value;
+        self.count += 1;
+    }
+
+    fn counter_json(&self) -> Value {
+        let first = self.first.unwrap_or(0.0);
+        let last = self.last.unwrap_or(first);
+        json!({
+            "first": first,
+            "last": last,
+            "delta": if last >= first { last - first } else { last },
+        })
+    }
+
+    fn gauge_json(&self) -> Value {
+        json!({
+            "first": self.first,
+            "last": self.last,
+            "min": self.min,
+            "max": self.max,
+            "mean": if self.count > 0 { Some(self.sum / self.count as f64) } else { None },
+        })
+    }
+}
+
+#[derive(Default)]
+struct BucketStats {
+    first: Option<f64>,
+    last: Option<f64>,
+}
+
+impl BucketStats {
+    fn record(&mut self, value: f64) {
+        self.first.get_or_insert(value);
+        self.last = Some(value);
+    }
+
+    fn delta(&self) -> f64 {
+        let first = self.first.unwrap_or(0.0);
+        let last = self.last.unwrap_or(first);
+        if last >= first {
+            last - first
+        } else {
+            last
+        }
+    }
+}
+
+fn histogram_json(buckets: &BTreeMap<String, BucketStats>) -> Value {
+    let mut bucket_values = Vec::new();
+    for (le, stats) in buckets {
+        bucket_values.push(json!({
+            "le": le,
+            "delta": stats.delta(),
+        }));
+    }
+
+    json!({
+        "buckets": bucket_values,
+        "p50": approximate_histogram_quantile(buckets, 0.50),
+        "p90": approximate_histogram_quantile(buckets, 0.90),
+        "p99": approximate_histogram_quantile(buckets, 0.99),
+    })
+}
+
+fn approximate_histogram_quantile(
+    buckets: &BTreeMap<String, BucketStats>,
+    quantile: f64,
+) -> Option<f64> {
+    let mut parsed = buckets
+        .iter()
+        .filter_map(|(le, stats)| parse_bucket_bound(le).map(|bound| (bound, stats.delta())))
+        .collect::<Vec<_>>();
+    parsed.sort_by(|a, b| a.0.total_cmp(&b.0));
+
+    let total = parsed
+        .iter()
+        .find(|(bound, _)| bound.is_infinite())
+        .map(|(_, count)| *count)
+        .or_else(|| parsed.last().map(|(_, count)| *count))?;
+    if total <= 0.0 {
+        return None;
+    }
+
+    let target = total * quantile;
+    let mut previous_bound = 0.0;
+    let mut previous_count = 0.0;
+    for (bound, cumulative_count) in parsed {
+        if cumulative_count >= target {
+            if bound.is_infinite() {
+                return Some(previous_bound);
+            }
+            let bucket_count = cumulative_count - previous_count;
+            if bucket_count <= 0.0 {
+                return Some(bound);
+            }
+            let within_bucket = (target - previous_count) / bucket_count;
+            return Some(previous_bound + (bound - previous_bound) * within_bucket);
+        }
+        previous_bound = bound;
+        previous_count = cumulative_count;
+    }
+    None
+}
+
+fn parse_bucket_bound(value: &str) -> Option<f64> {
+    match value {
+        "+Inf" | "Inf" | "inf" | "+inf" => Some(f64::INFINITY),
+        other => other.parse::<f64>().ok(),
+    }
+}
+
+pub fn resolve_metrics_endpoint(host: &str, endpoint: &str) -> Result<Url> {
+    if endpoint.starts_with("http://") || endpoint.starts_with("https://") {
+        return Url::parse(endpoint)
+            .with_context(|| format!("invalid metrics endpoint: {}", endpoint));
+    }
+
+    let normalized_host = if host.starts_with("http://") || host.starts_with("https://") {
+        host.trim_end_matches('/').to_string()
+    } else {
+        format!("https://{}", host.trim_end_matches('/'))
+    };
+    let resolved = format!("{}/{}", normalized_host, endpoint.trim_start_matches('/'));
+    Url::parse(&resolved).with_context(|| format!("invalid metrics endpoint: {}", resolved))
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn resolves_relative_metrics_endpoint_against_host() {
+        let url = resolve_metrics_endpoint("http://127.0.0.1:3000/", "/metrics").unwrap();
+        assert_eq!(url.as_str(), "http://127.0.0.1:3000/metrics");
+    }
+
+    #[test]
+    fn parses_colon_metric_names_and_labels() {
+        let parsed = parse_prometheus_text(
+            r#"
+# HELP vllm:prompt_tokens_total Number of prefill tokens processed.
+# TYPE vllm:prompt_tokens_total counter
+vllm:prompt_tokens_total{model_name="Qwen/Qwen3"} 12
+# TYPE sglang:token_usage gauge
+sglang:token_usage{model_name="Qwen/Qwen3"} 0.25
+"#,
+        )
+        .unwrap();
+
+        assert_eq!(parsed[0].name, "vllm:prompt_tokens_total");
+        assert_eq!(parsed[0].metric_type, "counter");
+        assert_eq!(parsed[0].labels.get("model_name").unwrap(), "Qwen/Qwen3");
+        assert_eq!(parsed[1].name, "sglang:token_usage");
+        assert_eq!(parsed[1].metric_type, "gauge");
+    }
+
+    #[test]
+    fn summarizes_counter_deltas_and_gauge_stats() {
+        let mut summary = MetricsSummaryBuilder::new("metrics".to_string());
+        let first = parse_prometheus_text(
+            r#"
+# TYPE vllm:prompt_tokens_total counter
+vllm:prompt_tokens_total 10
+# TYPE vllm:num_requests_running gauge
+vllm:num_requests_running 1
+"#,
+        )
+        .unwrap();
+        let second = parse_prometheus_text(
+            r#"
+# TYPE vllm:prompt_tokens_total counter
+vllm:prompt_tokens_total 25
+# TYPE vllm:num_requests_running gauge
+vllm:num_requests_running 3
+"#,
+        )
+        .unwrap();
+
+        summary.record_success(Utc::now(), &first);
+        summary.record_success(Utc::now(), &second);
+        let payload = summary.summary();
+
+        assert_eq!(
+            payload["counters"]["vllm:prompt_tokens_total"]["delta"],
+            json!(15.0)
+        );
+        assert_eq!(
+            payload["gauges"]["vllm:num_requests_running"]["max"],
+            json!(3.0)
+        );
+    }
+}
diff --git a/rust/src/py_bindings.rs b/rust/src/py_bindings.rs
index 61b415e..464b020 100644
--- a/rust/src/py_bindings.rs
+++ b/rust/src/py_bindings.rs
@@ -1,4 +1,5 @@
 use std::collections::HashMap;
+use std::path::PathBuf;
 use std::time::Duration;
 
 use anyhow::{anyhow, Context, Result};
@@ -10,8 +11,9 @@ use serde::{Deserialize, Serialize};
 use serde_json::Value;
 
 use crate::{
-    run_benchmark, run_from_argv, BenchmarkConfig, BenchmarkReport, DistMode, FailureRecord,
-    GenerateOptions, RequestEntry, RunMode,
+    resolve_metrics_endpoint, run_benchmark, run_from_argv, BenchmarkConfig, BenchmarkReport,
+    DistMode, FailureRecord, GenerateOptions, MetricsConfig, MetricsRunReport, RequestEntry,
+    RunMode,
 };
 
 #[derive(Debug, Clone, Deserialize, Serialize)]
@@ -173,9 +175,25 @@ struct PyBenchmarkConfig {
     #[serde(default)]
     sglang: bool,
     #[serde(default)]
+    enable_json_decoding: bool,
+    #[serde(default)]
+    qwen35_disable_thinking: bool,
+    #[serde(default)]
     seed: Option<u64>,
     #[serde(default)]
     dry_run: bool,
+    #[serde(default)]
+    metrics_host: Option<String>,
+    #[serde(default)]
+    metrics_endpoint: Option<String>,
+    #[serde(default)]
+    metrics_output_dir: Option<PathBuf>,
+    #[serde(default)]
+    metrics_interval_ms: Option<u64>,
+    #[serde(default)]
+    metrics_timeout_ms: Option<u64>,
+    #[serde(default)]
+    metrics_fail_on_error: bool,
 }
 
 fn to_benchmark_config(value: PyBenchmarkConfig) -> Result<BenchmarkConfig> {
@@ -220,6 +238,14 @@ fn to_benchmark_config(value: PyBenchmarkConfig) -> Result<BenchmarkConfig> {
         config = config.with_sglang(true);
     }
 
+    if value.enable_json_decoding {
+        config = config.with_json_decoding(true);
+    }
+
+    if value.qwen35_disable_thinking {
+        config = config.with_qwen35_disable_thinking(true);
+    }
+
     if let Some(seed) = value.seed {
         config = config.with_seed(seed);
     }
@@ -228,6 +254,31 @@ fn to_benchmark_config(value: PyBenchmarkConfig) -> Result<BenchmarkConfig> {
         config = config.with_dry_run(true);
     }
 
+    if let Some(metrics_output_dir) = value.metrics_output_dir {
+        let metrics_interval_ms = value.metrics_interval_ms.unwrap_or(1000);
+        if metrics_interval_ms == 0 {
+            return Err(anyhow!("metrics_interval_ms must be greater than zero"));
+        }
+        let metrics_timeout_ms = value.metrics_timeout_ms.unwrap_or(2000);
+        if metrics_timeout_ms == 0 {
+            return Err(anyhow!("metrics_timeout_ms must be greater than zero"));
+        }
+        let metrics_host = value
+            .metrics_host
+            .as_deref()
+            .unwrap_or(config.endpoint.as_str());
+        let metrics_endpoint = value.metrics_endpoint.as_deref().unwrap_or("/metrics");
+        let endpoint = resolve_metrics_endpoint(metrics_host, metrics_endpoint)?;
+        config = config.with_metrics(MetricsConfig {
+            endpoint,
+            output_dir: metrics_output_dir,
+            interval: Duration::from_millis(metrics_interval_ms),
+            timeout: Duration::from_millis(metrics_timeout_ms),
+            fail_on_error: value.metrics_fail_on_error,
+            metadata: serde_json::json!({}),
+        });
+    }
+
     for (name, value) in value.headers {
         let header_name = HeaderName::from_bytes(name.as_bytes())
             .with_context(|| format!("invalid header name: {}", name))?;
@@ -275,6 +326,7 @@ struct PyBenchmarkReport {
     latency_p90_ms: Option<f64>,
     latency_p99_ms: Option<f64>,
     failures: Vec<PyFailureRecord>,
+    metrics: Option<MetricsRunReport>,
 }
 
 impl From<BenchmarkReport> for PyBenchmarkReport {
@@ -304,6 +356,7 @@ impl From<BenchmarkReport> for PyBenchmarkReport {
                 .into_iter()
                 .map(PyFailureRecord::from)
                 .collect(),
+            metrics: value.metrics,
         }
     }
 }
diff --git a/rust/src/report.rs b/rust/src/report.rs
index b0c185a..451bd4d 100644
--- a/rust/src/report.rs
+++ b/rust/src/report.rs
@@ -1,5 +1,7 @@
 use std::time::Duration;
 
+use crate::metrics::MetricsRunReport;
+
 #[derive(Debug, Clone)]
 pub struct FailureRecord {
     pub user_id: usize,
@@ -21,6 +23,7 @@ pub struct BenchmarkReport {
     pub latency_p90: Option<Duration>,
     pub latency_p99: Option<Duration>,
     pub failures: Vec<FailureRecord>,
+    pub metrics: Option<MetricsRunReport>,
 }
 
 impl BenchmarkReport {
diff --git a/rust/src/runner.rs b/rust/src/runner.rs
index 30ab348..81873a2 100644
--- a/rust/src/runner.rs
+++ b/rust/src/runner.rs
@@ -14,6 +14,7 @@ use tokio::sync::mpsc;
 use tokio::task::JoinSet;
 
 use crate::config::{BenchmarkConfig, RequestEntry, RunMode};
+use crate::metrics::spawn_metrics_collector;
 use crate::report::{BenchmarkReport, FailureRecord};
 use std::cmp;
 
@@ -33,6 +34,50 @@ fn output_token_field_names(use_sglang: bool) -> (&'static str, &'static str) {
     }
 }
 
+fn apply_json_decoding(body: &mut Value, use_sglang: bool) {
+    let Some(map) = body.as_object_mut() else {
+        return;
+    };
+
+    if use_sglang {
+        map.insert(
+            "response_format".to_string(),
+            serde_json::json!({"type": "json_object"}),
+        );
+        return;
+    }
+
+    let structured_outputs = map
+        .entry("structured_outputs".to_string())
+        .or_insert_with(|| serde_json::json!({}));
+    if !structured_outputs.is_object() {
+        *structured_outputs = serde_json::json!({});
+    }
+
+    structured_outputs
+        .as_object_mut()
+        .expect("structured_outputs was normalized to an object")
+        .insert("json_object".to_string(), serde_json::json!(true));
+}
+
+fn apply_qwen35_disable_thinking(body: &mut Value) {
+    let Some(map) = body.as_object_mut() else {
+        return;
+    };
+
+    let chat_template_kwargs = map
+        .entry("chat_template_kwargs".to_string())
+        .or_insert_with(|| serde_json::json!({}));
+    if !chat_template_kwargs.is_object() {
+        *chat_template_kwargs = serde_json::json!({});
+    }
+
+    chat_template_kwargs
+        .as_object_mut()
+        .expect("chat_template_kwargs was normalized to an object")
+        .insert("enable_thinking".to_string(), serde_json::json!(false));
+}
+
 pub async fn run_benchmark(config: BenchmarkConfig) -> Result<BenchmarkReport> {
     let start = Instant::now();
     let client = Client::builder()
@@ -48,11 +93,20 @@ pub async fn run_benchmark(config: BenchmarkConfig) -> Result<BenchmarkReport> {
 
     let config = Arc::new(config);
 
+    let scrape_handle = if !config.dry_run {
+        match config.metrics.clone() {
+            Some(metrics_config) => Some(spawn_metrics_collector(metrics_config).await?),
+            None => None,
+        }
+    } else {
+        None
+    };
+
     let (event_tx, event_rx) = mpsc::unbounded_channel();
     let (status_tx, status_rx) = mpsc::unbounded_channel();
 
     let metrics_status_tx = status_tx.clone();
-    let metrics_handle = tokio::spawn(async move {
+    let aggregator_handle = tokio::spawn(async move {
         collect_metrics(event_rx, planned_total_requests, metrics_status_tx).await
     });
 
@@ -70,6 +124,7 @@ pub async fn run_benchmark(config: BenchmarkConfig) -> Result<BenchmarkReport> {
     drop(status_tx);
 
     let mut interrupted = false;
+    let mut fatal_error: Option<anyhow::Error> = None;
     let mut ctrl_c = pin!(tokio::signal::ctrl_c());
     while !join_set.is_empty() {
         tokio::select! {
@@ -94,7 +149,8 @@ pub async fn run_benchmark(config: BenchmarkConfig) -> Result<BenchmarkReport> {
                                 if interrupted {
                                     eprintln!("worker exited with error during shutdown: {}", err);
                                 } else {
-                                    return Err(err);
+                                    fatal_error = Some(err);
+                                    join_set.abort_all();
                                 }
                             }
                         }
@@ -103,7 +159,8 @@ pub async fn run_benchmark(config: BenchmarkConfig) -> Result<BenchmarkReport> {
                         if interrupted && err.is_cancelled() {
                             continue;
                         }
-                        return Err(anyhow!("worker task failed: {}", err));
+                        fatal_error = Some(anyhow!("worker task failed: {}", err));
+                        join_set.abort_all();
                     }
                     None => break,
                 }
@@ -111,14 +168,23 @@ pub async fn run_benchmark(config: BenchmarkConfig) -> Result<BenchmarkReport> {
         }
     }
 
-    let aggregator = metrics_handle.await??;
+    let aggregator = aggregator_handle.await??;
     let total_duration = start.elapsed();
 
     tracker_handle
         .await
         .map_err(|err| anyhow!("status tracker task failed: {}", err))?;
 
-    let (report, dry_run_events) = aggregator.finalize(total_duration);
+    let metrics_report = match scrape_handle {
+        Some(handle) => Some(handle.stop().await?),
+        None => None,
+    };
+
+    if let Some(err) = fatal_error {
+        return Err(err);
+    }
+
+    let (report, dry_run_events) = aggregator.finalize(total_duration, metrics_report);
 
     if config.dry_run && !dry_run_events.is_empty() {
         print_sorted_dry_run_events(&dry_run_events);
@@ -204,6 +270,14 @@ async fn dispatch_request(
 
     let mut request_body = request_entry.body.clone();
 
+    if config.enable_json_decoding {
+        apply_json_decoding(&mut request_body, config.sglang);
+    }
+
+    if config.qwen35_disable_thinking {
+        apply_qwen35_disable_thinking(&mut request_body);
+    }
+
     // If lognormal output sampling is configured, sample and inject tokens
     let mut lognorm_tokens: Option<usize> = None;
     if let Some((mu, sigma, max)) = config.output_lognorm {
@@ -508,7 +582,11 @@ impl MetricsAggregator {
         Ok(())
     }
 
-    fn finalize(self, total_duration: Duration) -> (BenchmarkReport, Vec<DryRunRecord>) {
+    fn finalize(
+        self,
+        total_duration: Duration,
+        metrics: Option<crate::metrics::MetricsRunReport>,
+    ) -> (BenchmarkReport, Vec<DryRunRecord>) {
         let total_requests = self.successful_requests + self.failed_requests;
         let duration_secs = total_duration.as_secs_f64();
         let prompt_tokens_per_second = if duration_secs > 0.0 {
@@ -547,6 +625,7 @@ impl MetricsAggregator {
             latency_p90,
             latency_p99,
             failures: self.failures,
+            metrics,
         };
 
         (report, self.dry_run_events)
@@ -753,3 +832,85 @@ fn print_histogram(label: &str, data: &[usize], bins: usize, bar_width: usize) {
         );
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use serde_json::json;
+
+    #[test]
+    fn json_decoding_for_sglang_adds_response_format() {
+        let mut body = json!({
+            "model": "m",
+            "messages": [{"role": "user", "content": "ping"}]
+        });
+
+        apply_json_decoding(&mut body, true);
+
+        assert_eq!(body["response_format"], json!({"type": "json_object"}));
+        assert!(body.get("extra_body").is_none());
+    }
+
+    #[test]
+    fn json_decoding_for_vllm_adds_structured_outputs() {
+        let mut body = json!({
+            "model": "m",
+            "messages": [{"role": "user", "content": "ping"}]
+        });
+
+        apply_json_decoding(&mut body, false);
+
+        assert_eq!(body["structured_outputs"]["json_object"], json!(true));
+        assert!(body.get("response_format").is_none());
+    }
+
+    #[test]
+    fn json_decoding_for_vllm_preserves_existing_structured_outputs_fields() {
+        let mut body = json!({
+            "model": "m",
+            "priority": 1,
+            "structured_outputs": {
+                "regex": "[0-9]+"
+            }
+        });
+
+        apply_json_decoding(&mut body, false);
+
+        assert_eq!(body["priority"], json!(1));
+        assert_eq!(body["structured_outputs"]["regex"], json!("[0-9]+"));
+        assert_eq!(body["structured_outputs"]["json_object"], json!(true));
+    }
+
+    #[test]
+    fn qwen35_disable_thinking_adds_chat_template_kwargs() {
+        let mut body = json!({
+            "model": "m",
+            "messages": [{"role": "user", "content": "ping"}]
+        });
+
+        apply_qwen35_disable_thinking(&mut body);
+
+        assert_eq!(
+            body["chat_template_kwargs"]["enable_thinking"],
+            json!(false)
+        );
+    }
+
+    #[test]
+    fn qwen35_disable_thinking_preserves_existing_chat_template_kwargs_fields() {
+        let mut body = json!({
+            "model": "m",
+            "chat_template_kwargs": {
+                "foo": "bar"
+            }
+        });
+
+        apply_qwen35_disable_thinking(&mut body);
+
+        assert_eq!(body["chat_template_kwargs"]["foo"], json!("bar"));
+        assert_eq!(
+            body["chat_template_kwargs"]["enable_thinking"],
+            json!(false)
+        );
+    }
+}
diff --git a/src/batchbench/harness.py b/src/batchbench/harness.py
index ebc24df..4e591c5 100644
--- a/src/batchbench/harness.py
+++ b/src/batchbench/harness.py
@@ -11,11 +11,15 @@
 """
 
 import argparse
+import base64
 import hashlib
+import io
 import json
 import os
 import re
+import shlex
 import sys
+import tarfile
 import time
 from concurrent.futures import ThreadPoolExecutor, as_completed
 from dataclasses import dataclass
@@ -449,9 +453,35 @@ def run_benchmark(env: RemoteEnvironment, config: dict, run_dir: Path) -> None:
                 f.write(stdout)
             print(f"Results saved to: {local_results}")
 
+    remote_metrics = bench_cfg.get("metrics-output-dir")
+    if remote_metrics:
+        fetch_metrics_artifacts(env, remote_metrics, run_dir / "metrics")
+
     print("\nBenchmark complete!")
 
 
+def fetch_metrics_artifacts(env: RemoteEnvironment, remote_dir: str, local_dir: Path) -> None:
+    """Save BatchBench metrics artifacts from container to local run directory."""
+    print("\nSaving metrics artifacts...")
+    quoted_remote_dir = shlex.quote(remote_dir)
+    cmd = f"test -d {quoted_remote_dir} && tar -C {quoted_remote_dir} -czf - . | base64 -w0"
+    stdout, _, rc = env.exec(cmd, timeout=120)
+    if rc != 0 or not stdout.strip():
+        print("Warning: Could not retrieve metrics artifacts")
+        return
+
+    local_dir.mkdir(parents=True, exist_ok=True)
+    archive = base64.b64decode(stdout.strip())
+    local_root = local_dir.resolve()
+    with tarfile.open(fileobj=io.BytesIO(archive), mode="r:gz") as tar:
+        for member in tar.getmembers():
+            target = (local_dir / member.name).resolve()
+            if not target.is_relative_to(local_root):
+                raise PipelineError(f"Unsafe metrics archive path: {member.name}")
+        tar.extractall(local_dir)
+    print(f"Metrics artifacts saved to: {local_dir}")
+
+
 def wait_for_vllm_ready(env: RemoteEnvironment, port: int, timeout: int) -> None:
     print(f"\nWaiting for vLLM server to be ready (timeout: {timeout}s)...")
     print("-" * 60)