{"model":"deepseek-4.1-flash","title":"Hourly update #19 — chunk-size loop: both hypotheses rejected","status":"negative","summary":"Tested prefill-chunk size as the ITL-tail fix: MAXBATCH=8192 collapses KV (7.86M-\u003e3.23M) and fails c1; MAXBATCH=1024 is rejected by the multimodal min. No free ITL fix; baseline restored","config":{"serving":"TP1xPP8 (5,5,5,5,5,5,5,5), max-num-seqs 32, util 0.94","engine":"vllm-backport-v41:sm80-v13 (v0.13.0 + Schaka DSV41 kit + patches 0007/0008)","kv":"fp8_ds_mla, 7,859,683-token pool (7.50x @1M), block 128","notes":"TRITON_MLA_SPARSE_DSV41, DSpark k=5, VLLM_SPARSE_DECODE_HEAD_BLOCK_SIZE=32, default max-num-batched-tokens"},"benchmarks":[{"metric":"c1_gen_tok_s","label":"c1 Gen tok/s @ 32K","value":20.6,"unit":"tok/s","context":"frozen PP8 util-0.94 baseline, 32K in / 256 out, 16 real-text prompts, ignore_eos"},{"metric":"c4_gen_tok_s","label":"c4 Gen tok/s @ 32K","value":53.4,"unit":"tok/s","context":"32K in / 256 out, 16 prompts"},{"metric":"c8_gen_tok_s","label":"c8 Gen tok/s @ 32K","value":78.9,"unit":"tok/s","context":"32K in / 256 out, 16 prompts"},{"metric":"c8_c1_scaling","label":"c8/c1 scaling @ 32K","value":3.82,"unit":"x","context":"derived"},{"metric":"agg_512k_c16_tok_s","label":"512K aggregate decode @ c16","value":98.9,"unit":"tok/s","context":"shared 487K prefix, pure decode, output 256"},{"metric":"kv_pool_tokens","label":"KV token capacity (PP8 util 0.94)","value":7859683,"unit":"tokens","context":"default max-num-batched-tokens"}],"found":["max-num-batched-tokens trades directly against KV capacity: MAXBATCH=8192 dropped the PP8 pool from 7.86M to 3.23M tokens and failed c1","max-num-batched-tokens is floored by the vision item (max_tokens_per_mm_item=1025); MAXBATCH=1024 is rejected at startup","the ITL p99 tail also appears in shared-prefix decode without prefill interleaving -\u003e it is a depth decode-step tail, not purely chunked-prefill interleaving","master-table c8 is noisy: 101.5 (util 0.96) vs 78.9 (util 0.94) for a change that should not matter -\u003e raise prompts/reps for a stable verdict"],"tried":["MAXBATCH=8192 (larger prefill chunks -\u003e fewer decode interruptions)","MAXBATCH=1024 (smaller chunks -\u003e finer interleave)"],"failed":["MAXBATCH=8192: KV 7,859,683 -\u003e 3,234,468 tokens and c1 fail=16 (dead)","MAXBATCH=1024: ValueError Chunked MM input disabled but max_tokens_per_mm_item (1025) \u003e max_num_batched_tokens (1024)"],"next":["harden the master table (32+ prompts, 2 reps, report variance) so the verdict is not noise-limited","attack the c4/c8 ITL p99 tail as a decode-side problem (sparse-MLA decode step variability / DSpark verify), since it persists without prefill interleaving","re-run the 512K@c16 milestone on the restored baseline and one live opencode c1 run (realness gate)"],"notes":"Chunk size is constrained to roughly [1030, small] and larger values cost KV, so it is not a free ITL fix. Baseline config restored and healthy."}