Per-token routed experts of accepted generated tokens (135,144 tokens each; EPLB off, so no popularity file). Provenance has host names and absolute paths removed.
10 days ago
Record each request's acceptance in six speculative captures' traces
Every speculative capture whose run recorded per-request decode progress: glm52_nvfp4 vllm_mtp_k5 diverse_100 20260902 and 20260922, glm52_nvfp4 vllm_mtp_k5 quadrant_c48 20260923, glm53_nvfp4 vllm_dflash2_k7 diverse_100 20260922, glm53_flash_fp8 vllm_mtp_k5 diverse_100 20260922 and deepseek_v4_flash vllm_dspark_k7 diverse_100 20260922. trace.csv gains the speculative tag's accept_rate column, each request's conditional acceptance chain; the pooled rates equal the server's spec-decode counters of the same pass. Other columns are unchanged. provenance.json names the pass and whether it is the routing's own, and the README describes the column. glm52 vllm_mtp_k5 diverse_100 20260830 and glm53 vllm_dflash2_k7 diverse_100 20260920 recorded no per-request progress and are unchanged.
11 days ago
Add 6 MoE routing captures (DeepSeek-V4.1-Flash, GLM-5.3-Flash MTP5, DeepSeek-V4-Flash DSpark7, GLM-5.2 no-spec and MTP5 corpora, GLM-5.3 DFlash2 replicate); point 6 provenance files at trace/ specs
