Publish RTX PRO 4000 benchmark scaffold

CONTRIBUTING.md

@@ -0,0 +1,41 @@
+# Contributing Results
+
+Benchmarks and configs are welcome. Support requests are out of scope.
+
+## Result Rules
+
+- Rows must be owner-measured on hardware the submitter operates.
+- Cloud, vendor, review-site, or copied rows are rejected.
+- `variant` is mandatory.
+- `bandwidth_gbps`, `tdp_w`, and `pcie_lanes` are mandatory and must match the
+  declared desktop/SFF variant.
+- `vram_gb`, `memory_type`, and `bus_width_bit` are required metadata, but they
+  do not identify desktop versus SFF. Desktop and SFF entries are both 24 GB
+  GDDR7 ECC on a 192-bit bus.
+- Desktop, SFF, and laptop rows are accepted but segregated by variant and never
+  merged into desktop tables.
+- Include enough metadata for another owner to repeat the run.
+- Tag related Hugging Face model cards with `rtx-pro-4000`.
+
+## Required Metadata
+
+Use `templates/result-row.json` as the starting point.
+
+Keep notes short but useful: backend flags, power method, whether the run was
+served or local, and what else was stopped to isolate the GPU.
+
+## Validation
+
+Before submitting:
+
+```bash
+python validate.py data/results.jsonl
+```
+
+To validate a proposed row separately:
+
+```bash
+python validate.py path/to/proposed.jsonl
+```
+
+Rows with `source` other than `owner-measured` fail validation.

METHODOLOGY.md

@@ -0,0 +1,65 @@
+# Methodology
+
+This methodology is canonical for the benchmark dataset.
+
+## Scope
+
+Rows must come from owner-operated hardware. Cloud instances, vendor marketing
+numbers, and third-party reposts are excluded because the value of this dataset
+is reproducible local workstation behavior.
+
+## Throughput Mode
+
+Throughput rows should run with model "thinking" disabled when the backend and
+model expose that control. Thinking can improve quality, but it changes token
+budget use and makes raw tok/s comparisons noisy. Quality belongs in a separate
+axis from speed.
+
+If thinking cannot be disabled, set `thinking=n/a` and explain why in `notes`.
+
+## Served vs Local
+
+Record whether the model was served through `llama.cpp`, `vllm`,
+`tensorrt-llm`, or another backend. Served HTTP measurements include scheduler
+and loopback overhead; local CLI measurements may not. Do not merge those
+without keeping the `backend` and `backend_version` fields visible.
+
+## Clean Runs
+
+GPU bandwidth contention invalidates results. Do not run benchmarks while another
+model is using the same GPU's memory bus.
+
+The seed rows come from llm-lab card-bench, which stops the production stack for
+clean card-bench windows. Current P620 placement from DC-BRIEFING is:
+
+- GPU 0: `gemma-aeon`, single-card production model.
+- GPU 1: utility/lab card, with embed/rerank and free VRAM for suitable
+  single-GPU lab work.
+
+Practical rule for this rig: a single-GPU run that fits on the lab card does not
+need a production teardown. A TP=2 or all-GPU run requires stopping
+`gemma-aeon`, entering maintenance mode, running the benchmark, then restarting
+and LAN-verifying the production endpoint.
+
+For other owners: report what else was running on each GPU and how you isolated
+the measurement window.
+
+## Required Reporting
+
+Every row must report:
+
+- Variant: desktop, sff, or laptop.
+- Bandwidth, TDP, and PCIe lanes. These separate desktop from SFF.
+- VRAM, memory type, and memory bus width. These are required metadata but do
+  not separate desktop from SFF because both are 24 GB GDDR7 ECC on a 192-bit
+  bus.
+- Driver version.
+- Backend and backend version.
+- Single-GPU vs multi-GPU.
+- Context length and concurrency.
+- Prefill tok/s and decode tok/s.
+- Peak VRAM used.
+- Power measurement method in `notes`.
+
+Power should be measured from board telemetry where possible, not a wall plug
+unless that is explicitly stated.

README.md

@@ -0,0 +1,67 @@
+---
+license: cc-by-4.0
+task_categories:
+- text-generation
+language:
+- en
+tags:
+- rtx-pro-4000
+- blackwell
+- llm-inference
+- owner-measured
+pretty_name: RTX PRO 4000 Blackwell Owner Benchmarks
+---
+
+# RTX PRO 4000 Blackwell Owner Benchmarks
+
+This dataset records owner-operated inference benchmark rows for the NVIDIA RTX
+PRO 4000 Blackwell family, anchored on the desktop 24 GB card.
+
+The point of the dataset is measured workstation behavior. Cloud benchmark rows,
+vendor marketing rows, and unverified screenshots are out of scope.
+
+## Files
+
+- `data/results.jsonl` - benchmark rows, one JSON object per line.
+- `schema.json` - machine-readable JSON Schema.
+- `SCHEMA.md` - human-readable field contract and variant rules.
+- `METHODOLOGY.md` - how runs should be measured and compared.
+- `CONTRIBUTING.md` - contribution rules and submission template.
+- `validate.py` - local validator used before accepting rows.
+- `tests/` - validator fixtures; not benchmark data.
+
+## Variant Rule
+
+Every row must set `variant` to one of:
+
+- `desktop`
+- `sff`
+- `laptop`
+
+The validator cross-checks `variant` against `bandwidth_gbps`, `tdp_w`, and
+`pcie_lanes` for desktop/SFF rows. `vram_gb`, `memory_type`, and
+`bus_width_bit` are required metadata, but they are not desktop/SFF
+discriminators: desktop and SFF rows are both 24 GB GDDR7 ECC on a 192-bit bus.
+Results for SFF and laptop hardware are welcome, but they remain segregated by
+variant and are never merged into desktop tables.
+
+## Seed Rows
+
+The initial seed rows are Andrew's owner-measured llm-lab card-bench results on
+dual desktop RTX PRO 4000 Blackwell cards. They are marked:
+
+- `variant=desktop`
+- `memory_type=gddr7`
+- `bus_width_bit=192`
+- `tdp_w=140`
+- `pcie_lanes=16`
+- `source=owner-measured`
+- `submitter=Andrew`
+
+The seed rows are representative, not a full export of every local artifact.
+Their notes record the llm-lab run label and provenance.
+
+## License
+
+Dataset rows are released under CC-BY-4.0 unless a future row states a stricter
+compatible license in its notes.

SCHEMA.md

@@ -0,0 +1,88 @@
+# Benchmark Schema
+
+Each line in `data/results.jsonl` is one benchmark measurement row.
+
+JSONL files do not include a header row. The field contract below is the header.
+Run `python validate.py data/results.jsonl` before submitting changes.
+
+## Required Fields
+
+| Field | Type | Rule |
+|---|---|---|
+| `variant` | string | enum: `desktop`, `sff`, `laptop` |
+| `vram_gb` | number | required metadata; never identifies desktop vs SFF by itself |
+| `memory_type` | string | enum: `gddr7`, `gddr6` |
+| `bus_width_bit` | integer | memory bus width in bits |
+| `bandwidth_gbps` | number | primary desktop/SFF variant discriminator |
+| `tdp_w` | number | per-card canonical/reference total board power for the variant |
+| `pcie_lanes` | integer | per-card PCIe link width for the variant, e.g. `16` for x16 |
+| `gpu_count` | integer | one or more GPUs used for the run |
+| `model` | string | model or run label |
+| `params_b` | number | parameter count in billions |
+| `quant` | string | enum below |
+| `backend` | string | enum below |
+| `backend_version` | string | commit, version, or digest |
+| `driver_version` | string | NVIDIA driver version |
+| `os` | string | OS and version |
+| `context_len` | integer | prompt/context target for the run |
+| `concurrency` | integer | simultaneous streams represented by the row |
+| `tok_s_prefill` | number | prefill or prompt-processing tokens per second |
+| `tok_s_decode` | number | decode/generation tokens per second |
+| `vram_used_gb` | number | peak VRAM used by the measured run |
+| `power_w` | number | board power during measured window |
+| `thinking` | string | enum: `on`, `off`, `n/a` |
+| `source` | string | must be `owner-measured` |
+| `submitter` | string | person or organization submitting |
+| `date` | string | ISO date, `YYYY-MM-DD` |
+| `notes` | string | provenance and measurement caveats |
+
+## Enums
+
+`memory_type`:
+
+- `gddr7`
+- `gddr6`
+
+`quant`:
+
+- `gguf-q4_k_m`
+- `gguf-q8_0`
+- `nvfp4`
+- `fp8`
+- `bf16`
+- `other`
+
+`backend`:
+
+- `llama.cpp`
+- `vllm`
+- `tensorrt-llm`
+- `other`
+
+`source`:
+
+- `owner-measured`
+
+## Variant Sanity Checks
+
+The validator identifies the physical desktop/SFF variant from bandwidth, TDP,
+and PCIe lanes. VRAM, memory type, and bus width are recorded metadata, but they
+do not separate desktop from SFF because both are 24 GB GDDR7 ECC on a 192-bit
+bus.
+
+The validator enforces these current public specs:
+
+| Variant | `vram_gb` | `memory_type` | `bus_width_bit` | `bandwidth_gbps` | `tdp_w` | `pcie_lanes` |
+|---|---:|---|---:|---:|---:|---:|
+| desktop | 24 | `gddr7` | 192 | 672 | 140 | 16 |
+| sff | 24 | `gddr7` | 192 | 432 | 70 | 8 |
+| laptop | 16 | `gddr7` | 256 | 896 | OEM-dependent | OEM-dependent |
+
+For laptop rows, `tdp_w` and `pcie_lanes` still must be reported from the owner
+or OEM configuration, but they are not hard-coded by the validator because mobile
+implementations vary.
+
+SFF source: NVIDIA RTX PRO 4000 Blackwell SFF Edition datasheet,
+`workstation-datasheet-blackwell-rtx-pro-4000-sff-nvidia-us-4016700.pdf`.
+It lists 24 GB GDDR7 ECC, 192-bit, 432 GB/s, PCIe 5.0 x8, 70 W, and dual-slot
+half-height form factor.

data/results.jsonl

@@ -0,0 +1,16 @@
+{"backend":"llama.cpp","backend_version":"94a220cd6745e6e3f8de62870b66fd5b9bc92700","bandwidth_gbps":672,"concurrency":1,"context_len":2048,"date":"2026-06-06","driver_version":"610.43.02","gpu_count":1,"model":"gemma4-12b-ggml-q4km","notes":"Seeded from llm-lab card-bench gemma4-12b-ggml-q4km-1gpu; one row per context; median of kept runs; production stack stopped for clean run.","os":"Ubuntu 24.04.4 LTS","params_b":12,"power_w":150.082,"quant":"gguf-q4_k_m","source":"owner-measured","submitter":"Andrew","thinking":"off","tok_s_decode":58.09,"tok_s_prefill":54.84,"variant":"desktop","vram_gb":24,"vram_used_gb":8.014,"memory_type":"gddr7","bus_width_bit":192,"tdp_w":140,"pcie_lanes":16}
+{"backend":"llama.cpp","backend_version":"94a220cd6745e6e3f8de62870b66fd5b9bc92700","bandwidth_gbps":672,"concurrency":1,"context_len":32768,"date":"2026-06-06","driver_version":"610.43.02","gpu_count":1,"model":"gemma4-12b-ggml-q4km","notes":"Seeded from llm-lab card-bench gemma4-12b-ggml-q4km-1gpu; one row per context; median of kept runs; production stack stopped for clean run.","os":"Ubuntu 24.04.4 LTS","params_b":12,"power_w":145.876,"quant":"gguf-q4_k_m","source":"owner-measured","submitter":"Andrew","thinking":"off","tok_s_decode":54.19,"tok_s_prefill":44.18,"variant":"desktop","vram_gb":24,"vram_used_gb":8.291,"memory_type":"gddr7","bus_width_bit":192,"tdp_w":140,"pcie_lanes":16}
+{"backend":"llama.cpp","backend_version":"94a220cd6745e6e3f8de62870b66fd5b9bc92700","bandwidth_gbps":672,"concurrency":1,"context_len":2048,"date":"2026-06-06","driver_version":"610.43.02","gpu_count":2,"model":"gemma4-12b-ggml-q4km","notes":"Seeded from llm-lab card-bench gemma4-12b-ggml-q4km-2gpu; one row per context; median of kept runs; production stack stopped for clean run.","os":"Ubuntu 24.04.4 LTS","params_b":12,"power_w":188.319,"quant":"gguf-q4_k_m","source":"owner-measured","submitter":"Andrew","thinking":"off","tok_s_decode":57.31,"tok_s_prefill":54.81,"variant":"desktop","vram_gb":24,"vram_used_gb":8.67,"memory_type":"gddr7","bus_width_bit":192,"tdp_w":140,"pcie_lanes":16}
+{"backend":"llama.cpp","backend_version":"94a220cd6745e6e3f8de62870b66fd5b9bc92700","bandwidth_gbps":672,"concurrency":1,"context_len":32768,"date":"2026-06-06","driver_version":"610.43.02","gpu_count":2,"model":"gemma4-12b-ggml-q4km","notes":"Seeded from llm-lab card-bench gemma4-12b-ggml-q4km-2gpu; one row per context; median of kept runs; production stack stopped for clean run.","os":"Ubuntu 24.04.4 LTS","params_b":12,"power_w":188.536,"quant":"gguf-q4_k_m","source":"owner-measured","submitter":"Andrew","thinking":"off","tok_s_decode":53.7,"tok_s_prefill":44.74,"variant":"desktop","vram_gb":24,"vram_used_gb":9.15,"memory_type":"gddr7","bus_width_bit":192,"tdp_w":140,"pcie_lanes":16}
+{"backend":"llama.cpp","backend_version":"94a220cd6745e6e3f8de62870b66fd5b9bc92700","bandwidth_gbps":672,"concurrency":1,"context_len":2048,"date":"2026-06-10","driver_version":"610.43.02","gpu_count":1,"model":"gemma4-26b-a4b-q4-k-m","notes":"Seeded from llm-lab card-bench gemma4-26b-a4b-q4-k-m-1gpu; one row per context; median of kept runs; production stack stopped for clean run.","os":"Ubuntu 24.04.4 LTS","params_b":26,"power_w":130.124,"quant":"gguf-q4_k_m","source":"owner-measured","submitter":"Andrew","thinking":"off","tok_s_decode":121.76,"tok_s_prefill":94.87,"variant":"desktop","vram_gb":24,"vram_used_gb":16.613,"memory_type":"gddr7","bus_width_bit":192,"tdp_w":140,"pcie_lanes":16}
+{"backend":"llama.cpp","backend_version":"94a220cd6745e6e3f8de62870b66fd5b9bc92700","bandwidth_gbps":672,"concurrency":1,"context_len":32768,"date":"2026-06-10","driver_version":"610.43.02","gpu_count":1,"model":"gemma4-26b-a4b-q4-k-m","notes":"Seeded from llm-lab card-bench gemma4-26b-a4b-q4-k-m-1gpu; one row per context; median of kept runs; production stack stopped for clean run.","os":"Ubuntu 24.04.4 LTS","params_b":26,"power_w":127.894,"quant":"gguf-q4_k_m","source":"owner-measured","submitter":"Andrew","thinking":"off","tok_s_decode":99.04,"tok_s_prefill":66.11,"variant":"desktop","vram_gb":24,"vram_used_gb":17.039,"memory_type":"gddr7","bus_width_bit":192,"tdp_w":140,"pcie_lanes":16}
+{"backend":"llama.cpp","backend_version":"94a220cd6745e6e3f8de62870b66fd5b9bc92700","bandwidth_gbps":672,"concurrency":1,"context_len":2048,"date":"2026-06-10","driver_version":"610.43.02","gpu_count":2,"model":"gemma4-26b-a4b-q4-k-m","notes":"Seeded from llm-lab card-bench gemma4-26b-a4b-q4-k-m-2gpu; one row per context; median of kept runs; production stack stopped for clean run.","os":"Ubuntu 24.04.4 LTS","params_b":26,"power_w":173.613,"quant":"gguf-q4_k_m","source":"owner-measured","submitter":"Andrew","thinking":"off","tok_s_decode":124.78,"tok_s_prefill":96.44,"variant":"desktop","vram_gb":24,"vram_used_gb":17.166,"memory_type":"gddr7","bus_width_bit":192,"tdp_w":140,"pcie_lanes":16}
+{"backend":"llama.cpp","backend_version":"94a220cd6745e6e3f8de62870b66fd5b9bc92700","bandwidth_gbps":672,"concurrency":1,"context_len":32768,"date":"2026-06-10","driver_version":"610.43.02","gpu_count":2,"model":"gemma4-26b-a4b-q4-k-m","notes":"Seeded from llm-lab card-bench gemma4-26b-a4b-q4-k-m-2gpu; one row per context; median of kept runs; production stack stopped for clean run.","os":"Ubuntu 24.04.4 LTS","params_b":26,"power_w":171.012,"quant":"gguf-q4_k_m","source":"owner-measured","submitter":"Andrew","thinking":"off","tok_s_decode":102.48,"tok_s_prefill":68.78,"variant":"desktop","vram_gb":24,"vram_used_gb":17.893,"memory_type":"gddr7","bus_width_bit":192,"tdp_w":140,"pcie_lanes":16}
+{"backend":"vllm","backend_version":"0.22.1rc1.dev245+g9c7f7741d","bandwidth_gbps":672,"concurrency":1,"context_len":2048,"date":"2026-06-10","driver_version":"610.43.02","gpu_count":1,"model":"gemma4-26b-a4b-aeon-nvfp4","notes":"Seeded from llm-lab card-bench gemma4-26b-a4b-aeon-nvfp4-1gpu; one row per context; median of kept runs; production stack stopped for clean run.","os":"Ubuntu 24.04.4 LTS","params_b":26,"power_w":140.535,"quant":"nvfp4","source":"owner-measured","submitter":"Andrew","thinking":"off","tok_s_decode":120.63,"tok_s_prefill":13064.82,"variant":"desktop","vram_gb":24,"vram_used_gb":21.262,"memory_type":"gddr7","bus_width_bit":192,"tdp_w":140,"pcie_lanes":16}
+{"backend":"vllm","backend_version":"0.22.1rc1.dev245+g9c7f7741d","bandwidth_gbps":672,"concurrency":1,"context_len":8192,"date":"2026-06-10","driver_version":"610.43.02","gpu_count":1,"model":"gemma4-26b-a4b-aeon-nvfp4","notes":"Seeded from llm-lab card-bench gemma4-26b-a4b-aeon-nvfp4-1gpu; one row per context; median of kept runs; production stack stopped for clean run.","os":"Ubuntu 24.04.4 LTS","params_b":26,"power_w":146.558,"quant":"nvfp4","source":"owner-measured","submitter":"Andrew","thinking":"off","tok_s_decode":118.49,"tok_s_prefill":12605.53,"variant":"desktop","vram_gb":24,"vram_used_gb":21.934,"memory_type":"gddr7","bus_width_bit":192,"tdp_w":140,"pcie_lanes":16}
+{"backend":"vllm","backend_version":"0.22.1rc1.dev245+g9c7f7741d","bandwidth_gbps":672,"concurrency":1,"context_len":32768,"date":"2026-06-10","driver_version":"610.43.02","gpu_count":1,"model":"gemma4-26b-a4b-aeon-nvfp4","notes":"Seeded from llm-lab card-bench gemma4-26b-a4b-aeon-nvfp4-1gpu; one row per context; median of kept runs; production stack stopped for clean run.","os":"Ubuntu 24.04.4 LTS","params_b":26,"power_w":149.598,"quant":"nvfp4","source":"owner-measured","submitter":"Andrew","thinking":"off","tok_s_decode":110.79,"tok_s_prefill":9334.83,"variant":"desktop","vram_gb":24,"vram_used_gb":22.355,"memory_type":"gddr7","bus_width_bit":192,"tdp_w":140,"pcie_lanes":16}
+{"backend":"vllm","backend_version":"0.22.1rc1.dev245+g9c7f7741d","bandwidth_gbps":672,"concurrency":1,"context_len":131072,"date":"2026-06-10","driver_version":"610.43.02","gpu_count":1,"model":"gemma4-26b-a4b-aeon-nvfp4","notes":"Seeded from llm-lab card-bench gemma4-26b-a4b-aeon-nvfp4-1gpu; one row per context; median of kept runs; production stack stopped for clean run.","os":"Ubuntu 24.04.4 LTS","params_b":26,"power_w":149.786,"quant":"nvfp4","source":"owner-measured","submitter":"Andrew","thinking":"off","tok_s_decode":103.3,"tok_s_prefill":4470.77,"variant":"desktop","vram_gb":24,"vram_used_gb":22.357,"memory_type":"gddr7","bus_width_bit":192,"tdp_w":140,"pcie_lanes":16}
+{"backend":"vllm","backend_version":"0.22.1rc1.dev245+g9c7f7741d","bandwidth_gbps":672,"concurrency":1,"context_len":2048,"date":"2026-06-10","driver_version":"610.43.02","gpu_count":2,"model":"gemma4-26b-a4b-aeon-nvfp4","notes":"Seeded from llm-lab card-bench gemma4-26b-a4b-aeon-nvfp4-2gpu; one row per context; median of kept runs; production stack stopped for clean run.","os":"Ubuntu 24.04.4 LTS","params_b":26,"power_w":169.035,"quant":"nvfp4","source":"owner-measured","submitter":"Andrew","thinking":"off","tok_s_decode":182.58,"tok_s_prefill":11943.84,"variant":"desktop","vram_gb":24,"vram_used_gb":43.984,"memory_type":"gddr7","bus_width_bit":192,"tdp_w":140,"pcie_lanes":16}
+{"backend":"vllm","backend_version":"0.22.1rc1.dev245+g9c7f7741d","bandwidth_gbps":672,"concurrency":1,"context_len":8192,"date":"2026-06-10","driver_version":"610.43.02","gpu_count":2,"model":"gemma4-26b-a4b-aeon-nvfp4","notes":"Seeded from llm-lab card-bench gemma4-26b-a4b-aeon-nvfp4-2gpu; one row per context; median of kept runs; production stack stopped for clean run.","os":"Ubuntu 24.04.4 LTS","params_b":26,"power_w":206.528,"quant":"nvfp4","source":"owner-measured","submitter":"Andrew","thinking":"off","tok_s_decode":174.03,"tok_s_prefill":12379.46,"variant":"desktop","vram_gb":24,"vram_used_gb":44.34,"memory_type":"gddr7","bus_width_bit":192,"tdp_w":140,"pcie_lanes":16}
+{"backend":"vllm","backend_version":"0.22.1rc1.dev245+g9c7f7741d","bandwidth_gbps":672,"concurrency":1,"context_len":32768,"date":"2026-06-10","driver_version":"610.43.02","gpu_count":2,"model":"gemma4-26b-a4b-aeon-nvfp4","notes":"Seeded from llm-lab card-bench gemma4-26b-a4b-aeon-nvfp4-2gpu; one row per context; median of kept runs; production stack stopped for clean run.","os":"Ubuntu 24.04.4 LTS","params_b":26,"power_w":246.071,"quant":"nvfp4","source":"owner-measured","submitter":"Andrew","thinking":"off","tok_s_decode":154.39,"tok_s_prefill":10850.93,"variant":"desktop","vram_gb":24,"vram_used_gb":44.598,"memory_type":"gddr7","bus_width_bit":192,"tdp_w":140,"pcie_lanes":16}
+{"backend":"vllm","backend_version":"0.22.1rc1.dev245+g9c7f7741d","bandwidth_gbps":672,"concurrency":1,"context_len":131072,"date":"2026-06-10","driver_version":"610.43.02","gpu_count":2,"model":"gemma4-26b-a4b-aeon-nvfp4","notes":"Seeded from llm-lab card-bench gemma4-26b-a4b-aeon-nvfp4-2gpu; one row per context; median of kept runs; production stack stopped for clean run.","os":"Ubuntu 24.04.4 LTS","params_b":26,"power_w":280.518,"quant":"nvfp4","source":"owner-measured","submitter":"Andrew","thinking":"off","tok_s_decode":149.28,"tok_s_prefill":6685.85,"variant":"desktop","vram_gb":24,"vram_used_gb":44.531,"memory_type":"gddr7","bus_width_bit":192,"tdp_w":140,"pcie_lanes":16}

schema.json

@@ -0,0 +1,69 @@
+{
+  "$schema": "https://json-schema.org/draft/2020-12/schema",
+  "$id": "https://huggingface.co/datasets/rtx-pro-4000/benchmarks/schema.json",
+  "title": "RTX PRO 4000 Blackwell benchmark row",
+  "type": "object",
+  "additionalProperties": false,
+  "required": [
+    "variant",
+    "vram_gb",
+    "memory_type",
+    "bus_width_bit",
+    "bandwidth_gbps",
+    "tdp_w",
+    "pcie_lanes",
+    "gpu_count",
+    "model",
+    "params_b",
+    "quant",
+    "backend",
+    "backend_version",
+    "driver_version",
+    "os",
+    "context_len",
+    "concurrency",
+    "tok_s_prefill",
+    "tok_s_decode",
+    "vram_used_gb",
+    "power_w",
+    "thinking",
+    "source",
+    "submitter",
+    "date",
+    "notes"
+  ],
+  "properties": {
+    "variant": { "type": "string", "enum": ["desktop", "sff", "laptop"] },
+    "vram_gb": { "type": "number", "exclusiveMinimum": 0 },
+    "memory_type": { "type": "string", "enum": ["gddr7", "gddr6"] },
+    "bus_width_bit": { "type": "integer", "minimum": 1 },
+    "bandwidth_gbps": { "type": "number", "exclusiveMinimum": 0 },
+    "tdp_w": { "type": "number", "exclusiveMinimum": 0 },
+    "pcie_lanes": { "type": "integer", "minimum": 1 },
+    "gpu_count": { "type": "integer", "minimum": 1 },
+    "model": { "type": "string", "minLength": 1 },
+    "params_b": { "type": "number", "exclusiveMinimum": 0 },
+    "quant": {
+      "type": "string",
+      "enum": ["gguf-q4_k_m", "gguf-q8_0", "nvfp4", "fp8", "bf16", "other"]
+    },
+    "backend": {
+      "type": "string",
+      "enum": ["llama.cpp", "vllm", "tensorrt-llm", "other"]
+    },
+    "backend_version": { "type": "string", "minLength": 1 },
+    "driver_version": { "type": "string", "minLength": 1 },
+    "os": { "type": "string", "minLength": 1 },
+    "context_len": { "type": "integer", "minimum": 1 },
+    "concurrency": { "type": "integer", "minimum": 1 },
+    "tok_s_prefill": { "type": "number", "minimum": 0 },
+    "tok_s_decode": { "type": "number", "minimum": 0 },
+    "vram_used_gb": { "type": "number", "minimum": 0 },
+    "power_w": { "type": "number", "minimum": 0 },
+    "thinking": { "type": "string", "enum": ["on", "off", "n/a"] },
+    "source": { "type": "string", "enum": ["owner-measured"] },
+    "submitter": { "type": "string", "minLength": 1 },
+    "date": { "type": "string", "pattern": "^\\d{4}-\\d{2}-\\d{2}$" },
+    "notes": { "type": "string", "minLength": 1 }
+  }
+}

templates/result-row.json

@@ -0,0 +1,28 @@
+{
+  "variant": "desktop",
+  "vram_gb": 24,
+  "memory_type": "gddr7",
+  "bus_width_bit": 192,
+  "bandwidth_gbps": 672,
+  "tdp_w": 140,
+  "pcie_lanes": 16,
+  "gpu_count": 1,
+  "model": "example-model-q4",
+  "params_b": 12,
+  "quant": "gguf-q4_k_m",
+  "backend": "llama.cpp",
+  "backend_version": "commit-or-version",
+  "driver_version": "610.43.02",
+  "os": "Ubuntu 24.04 LTS",
+  "context_len": 2048,
+  "concurrency": 1,
+  "tok_s_prefill": 0,
+  "tok_s_decode": 0,
+  "vram_used_gb": 0,
+  "power_w": 0,
+  "thinking": "off",
+  "source": "owner-measured",
+  "submitter": "name-or-handle",
+  "date": "2026-06-14",
+  "notes": "Describe measurement method, backend flags, and isolation."
+}

tests/README.md

@@ -0,0 +1,7 @@
+Validator fixtures only. These JSONL files are not benchmark data.
+
+Expected failures:
+
+- `invalid-desktop-sff-bandwidth.jsonl`: desktop row with SFF bandwidth.
+- `invalid-sff-desktop-bandwidth.jsonl`: SFF row with desktop bandwidth.
+- `invalid-laptop-vram.jsonl`: laptop row with desktop/SFF VRAM.

tests/invalid-desktop-sff-bandwidth.jsonl

@@ -0,0 +1 @@
+{"variant":"desktop","vram_gb":24,"memory_type":"gddr7","bus_width_bit":192,"bandwidth_gbps":432,"tdp_w":140,"pcie_lanes":16,"gpu_count":1,"model":"validator-fixture","params_b":1,"quant":"other","backend":"other","backend_version":"fixture","driver_version":"fixture","os":"fixture","context_len":1,"concurrency":1,"tok_s_prefill":0,"tok_s_decode":0,"vram_used_gb":0,"power_w":0,"thinking":"n/a","source":"owner-measured","submitter":"validator","date":"2026-06-14","notes":"Invalid fixture: desktop variant with SFF bandwidth."}

tests/invalid-laptop-vram.jsonl

@@ -0,0 +1 @@
+{"variant":"laptop","vram_gb":24,"memory_type":"gddr7","bus_width_bit":256,"bandwidth_gbps":896,"tdp_w":120,"pcie_lanes":16,"gpu_count":1,"model":"validator-fixture","params_b":1,"quant":"other","backend":"other","backend_version":"fixture","driver_version":"fixture","os":"fixture","context_len":1,"concurrency":1,"tok_s_prefill":0,"tok_s_decode":0,"vram_used_gb":0,"power_w":0,"thinking":"n/a","source":"owner-measured","submitter":"validator","date":"2026-06-14","notes":"Invalid fixture: laptop variant with desktop/SFF VRAM."}

tests/invalid-sff-desktop-bandwidth.jsonl

@@ -0,0 +1 @@
+{"variant":"sff","vram_gb":24,"memory_type":"gddr7","bus_width_bit":192,"bandwidth_gbps":672,"tdp_w":70,"pcie_lanes":8,"gpu_count":1,"model":"validator-fixture","params_b":1,"quant":"other","backend":"other","backend_version":"fixture","driver_version":"fixture","os":"fixture","context_len":1,"concurrency":1,"tok_s_prefill":0,"tok_s_decode":0,"vram_used_gb":0,"power_w":0,"thinking":"n/a","source":"owner-measured","submitter":"validator","date":"2026-06-14","notes":"Invalid fixture: SFF variant with desktop bandwidth."}

validate.py

@@ -0,0 +1,208 @@
+#!/usr/bin/env python3
+"""Validate RTX PRO 4000 benchmark JSONL rows."""
+
+from __future__ import annotations
+
+import argparse
+import json
+import re
+import sys
+from pathlib import Path
+from typing import Any
+
+
+REQUIRED = [
+    "variant",
+    "vram_gb",
+    "memory_type",
+    "bus_width_bit",
+    "bandwidth_gbps",
+    "tdp_w",
+    "pcie_lanes",
+    "gpu_count",
+    "model",
+    "params_b",
+    "quant",
+    "backend",
+    "backend_version",
+    "driver_version",
+    "os",
+    "context_len",
+    "concurrency",
+    "tok_s_prefill",
+    "tok_s_decode",
+    "vram_used_gb",
+    "power_w",
+    "thinking",
+    "source",
+    "submitter",
+    "date",
+    "notes",
+]
+
+ENUMS = {
+    "variant": {"desktop", "sff", "laptop"},
+    "memory_type": {"gddr7", "gddr6"},
+    "quant": {"gguf-q4_k_m", "gguf-q8_0", "nvfp4", "fp8", "bf16", "other"},
+    "backend": {"llama.cpp", "vllm", "tensorrt-llm", "other"},
+    "thinking": {"on", "off", "n/a"},
+    "source": {"owner-measured"},
+}
+
+VARIANT_DISCRIMINATORS = {
+    "desktop": {
+        "bandwidth_gbps": 672,
+        "tdp_w": 140,
+        "pcie_lanes": 16,
+    },
+    "sff": {
+        "bandwidth_gbps": 432,
+        "tdp_w": 70,
+        "pcie_lanes": 8,
+    },
+    "laptop": {
+        "vram_gb": 16,
+        "bandwidth_gbps": 896,
+    },
+}
+
+VARIANT_METADATA = {
+    "desktop": {
+        "vram_gb": 24,
+        "memory_type": "gddr7",
+        "bus_width_bit": 192,
+    },
+    "sff": {
+        "vram_gb": 24,
+        "memory_type": "gddr7",
+        "bus_width_bit": 192,
+    },
+    "laptop": {
+        "vram_gb": 16,
+        "memory_type": "gddr7",
+        "bus_width_bit": 256,
+    },
+}
+
+NUMERIC = {
+    "vram_gb",
+    "bandwidth_gbps",
+    "tdp_w",
+    "params_b",
+    "tok_s_prefill",
+    "tok_s_decode",
+    "vram_used_gb",
+    "power_w",
+}
+
+INTEGER = {"gpu_count", "bus_width_bit", "pcie_lanes", "context_len", "concurrency"}
+
+
+def row_errors(row: dict[str, Any], line_no: int) -> list[str]:
+    errors: list[str] = []
+
+    for field in REQUIRED:
+        if field not in row:
+            errors.append(f"line {line_no}: missing required field {field}")
+
+    for field in row:
+        if field not in REQUIRED:
+            errors.append(f"line {line_no}: unexpected field {field}")
+
+    for field, allowed in ENUMS.items():
+        value = row.get(field)
+        if value not in allowed:
+            errors.append(f"line {line_no}: {field}={value!r} not in {sorted(allowed)}")
+
+    for field in NUMERIC:
+        value = row.get(field)
+        if not isinstance(value, (int, float)) or isinstance(value, bool):
+            errors.append(f"line {line_no}: {field} must be numeric")
+        elif value < 0:
+            errors.append(f"line {line_no}: {field} must be >= 0")
+
+    for field in INTEGER:
+        value = row.get(field)
+        if not isinstance(value, int) or isinstance(value, bool):
+            errors.append(f"line {line_no}: {field} must be an integer")
+        elif value < 1:
+            errors.append(f"line {line_no}: {field} must be >= 1")
+
+    for field in ("model", "backend_version", "driver_version", "os", "submitter", "notes"):
+        value = row.get(field)
+        if not isinstance(value, str) or not value.strip():
+            errors.append(f"line {line_no}: {field} must be a non-empty string")
+
+    date_value = row.get("date")
+    if not isinstance(date_value, str) or not re.fullmatch(r"\d{4}-\d{2}-\d{2}", date_value):
+        errors.append(f"line {line_no}: date must be YYYY-MM-DD")
+
+    variant = row.get("variant")
+    discriminator_expected = VARIANT_DISCRIMINATORS.get(variant)
+    if discriminator_expected:
+        for field, expected_value in discriminator_expected.items():
+            actual = row.get(field)
+            if actual != expected_value:
+                errors.append(
+                    f"line {line_no}: {variant} discriminator requires {field}={expected_value}, got {actual!r}"
+                )
+
+    metadata_expected = VARIANT_METADATA.get(variant)
+    if metadata_expected:
+        for field, expected_value in metadata_expected.items():
+            actual = row.get(field)
+            if actual != expected_value:
+                errors.append(
+                    f"line {line_no}: {variant} metadata expects {field}={expected_value}, got {actual!r}"
+                )
+
+    if row.get("source") != "owner-measured":
+        errors.append(f"line {line_no}: source must be owner-measured")
+
+    return errors
+
+
+def validate(path: Path) -> tuple[int, list[str]]:
+    errors: list[str] = []
+    count = 0
+
+    try:
+        lines = path.read_text(encoding="utf-8").splitlines()
+    except OSError as exc:
+        return 0, [f"{path}: {exc}"]
+
+    for line_no, line in enumerate(lines, start=1):
+        if not line.strip():
+            continue
+        try:
+            row = json.loads(line)
+        except json.JSONDecodeError as exc:
+            errors.append(f"line {line_no}: invalid JSON: {exc}")
+            continue
+        if not isinstance(row, dict):
+            errors.append(f"line {line_no}: row must be a JSON object")
+            continue
+        count += 1
+        errors.extend(row_errors(row, line_no))
+
+    if count == 0:
+        errors.append(f"{path}: no rows found")
+    return count, errors
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description="Validate benchmark JSONL rows")
+    parser.add_argument("path", nargs="?", default="data/results.jsonl")
+    args = parser.parse_args()
+
+    count, errors = validate(Path(args.path))
+    if errors:
+        for error in errors:
+            print(error, file=sys.stderr)
+        return 1
+    print(f"OK: {count} rows valid")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())