b70-optimization-lab

DeepSeek V4 Flash uniform-K160 on 4x B70

Status: paused/closed frontier; best verified record preserved.

The lane closed on 2026-07-21 after the exact target-verified DSpark7 path reached a strict high of 80.820052 tok/s for one active generation. No later verified endpoint result exceeded it. The three independent strict suite medians were 80.820052, 76.900178, and 78.287226 tok/s, making 78.287226 tok/s the median-of-medians.

Use these entry points

• Standalone record recipe: repro/deepseek-v4-flash-k160-b70-80tps-20260718/
• Record note: 2026-07-18-sharded-target-argmax-record.md
• Closeout and reopen conditions: 2026-07-21-deepseek-v4-flash-frontier-closeout.md
• Validity rules: validity-gates.md
• Rejected paths and lessons: bugs-failed-paths.md
• Full experiment history: experiment-ledger.md

Record identity

• Model: 0xSero/DeepSeek-V4-Flash-180B, revision 7c360e1cd4a5168099dbc54d16d929bf6df04990.
• Hardware: four Intel Arc Pro B70 32 GB GPUs, TP4+EP, concurrency one.
• Runtime: vLLM 264c7f2f7df21ddeeab32ecca0353133344f1ac9, XPU kernels 31315673737d95da0f79179c8f755260ef02c1d6, oneCCL 48fda4f0e074db005596d6899d5227d3f0316c12.
• Graphs: target PIECEWISE and private breakable draft PIECEWISE.
• Speculation: DSpark7, exact M=7 draft query capture, unchanged target verifies accepted tokens at M=8.
• Mechanism: persistent Markov with W1-only replication, exact M8 compressors, selective M8 W8A16, MXFP4 N128, native M8 router normalization, and guarded sharded greedy target argmax/native target-token rejection.
• Benchmark: 12 unique cold prompts, 128 output tokens, median throughput for generated tokens 1-100 after TTFT.
• Validity: all 36 realistic requests cache-zero; four ordered six-case exact suites pass 24/24.
• LocalMaxxing: cmrquta9905w3lg013m5vxoqx.

The public uniform-K160 checkpoint is a useful experimental performance artifact, not a quality-certified official REAP construction: its hash layers are pruned and its calibration/ranking provenance is unavailable. Keep that caveat attached to every result.

Final decision

The endpoint record is intact and all reusable source, patches, harnesses, and negative results are preserved. Reopen only for a 10-20M-token EAGLE capture and hybrid-draft training effort, or a genuinely new mechanism that removes device execution time. Do not restart configuration sweeps, submission-only fusion, or already rejected M=1 occupancy experiments.

This site is open source. Improve this page.