D-Flash2
A speculative decoding method that serves as a new approach to accelerating large-model inference and is offered as an optional feature in Strata.
Timeline (last 90 days)
Strata 0.2.4 disables speculative decoding (D-Flash2) by default, because the official team believes no graphics card has yet been proven to make that path comprehensively faster, and it can only be enabled manually in the terminal.
D-Flash2 mode achieves a code throughput of 50.9 tokens/s, a long-text processing throughput of 25.4 tokens/s, and 66.6 tokens/s in thinking-type short-chat scenarios.
Under 16-way concurrent load, D-Flash2 achieves a total throughput of 227.6 tokens/s, an average of 28.2 tokens/s per channel, with a time-to-first-token latency of 4.18 seconds.
D-Spark is recommended for pure code high-frequency scenarios, while D-Flash2 is recommended for long documents, reasoning, multi-turn dialogue, and multi-user concurrency.
All Facts (2)
Strata 0.2.4 disables speculative decoding (D-Flash2) by default, because the official team believes no graphics card has yet been proven to make that path comprehensively faster, and it can only be enabled manually in the terminal.
50%UnverifiedD-Spark is recommended for pure code high-frequency scenarios, while D-Flash2 is recommended for long documents, reasoning, multi-turn dialogue, and multi-user concurrency.
50%