DeepSeek V4 Flash K160 record validity
The 80.820052 tok/s row is valid only under all of these conditions:
- one active generation on four B70s; never aggregate concurrent throughput;
- exact model, draft, runtime commits, topology, graph modes, and selector
identity from the standalone repro;
- unchanged K160 target verifies every accepted DSpark token at M=8;
- plain temperature-zero greedy sampling for the sharded-target-argmax path;
unsupported grammar, logprob, penalty, bias, bad-word, draft-logit, or
non-greedy configurations must fall back to the canonical path;
- each realistic prompt is unique and sent once cold;
cached_tokens=0 for every request and no prefix cache, response reuse,
n-gram/history acceleration, or context checkpoints;- primary metric is the median generated-token throughput for tokens 1-100
after TTFT across the fixed 12-prompt suite;
- exact canaries pass before, between, and after the three strict suites.
The public K160 checkpoint is hash-pruned and its calibration is not
reproducible. The gate supports a result for this exact experimental artifact;
it does not certify the artifact as official DeepSeek V4 Flash or true REAP.
Evidence is the record note,
the compact JSON,
and LocalMaxxing approval cmrquta9905w3lg013m5vxoqx.