metadata.json

{
    "model": "deepseek-ai/DeepSeek-V4-Flash-0731",
    "tp": 2,
    "dtype": "fp8_w8a8",
    "block_shape": [
        128,
        128
    ],
    "device_name": "NVIDIA_GB10",
    "tuned_at": "2026-08-08T06:00:00+00:00",
    "mode": "fp8",
    "fp8_configs": 7,
    "moe_configs": 0,
    "triton_version": "3.6.0",
    "vllm_version": "0.11.0.dev0 (jasl/vllm @ 67f0057da4e3)",
    "driver_version": "580.173.02",
    "cuda_version": "13.0",
    "sm_clock_cap_mhz": 2200,
    "notes": "Complete 7-shape set for DeepSeek-V4-Flash-0731 at TP=2 on GB10. Six shapes originate in jasl/vllm-ds4-sm120-harness (gb10_20260512, MIT). N=4096,K=12288 tuned here: it is the DSpark draft's mtp.0.main_proj. It is absent from the inherited set because that run (2026-05-12) predates every DSpark-equipped checkpoint (the first is DeepSeek-V4-Flash-DSpark, 2026-06-27) - NOT because the shape is unique to 0731. Tuned with SM clocks capped at 2200 MHz (required GB10 thermal mitigation). Measured effect on decode: none."
}