From ad32c4b1602cb3701526c5bcaf7dc770cd9ff34d Mon Sep 17 00:00:00 2001 From: zixi-qi Date: Mon, 27 Jul 2026 16:50:29 -0700 Subject: [PATCH] [NV] kimik3 DSpark: probabilistic draft + block rejection sampling Optimize Kimi-K3 DSpark speculative decoding acceptance rate by setting draft_sample_method=probabilistic and rejection_sample_method=block in the speculative-config. Co-Authored-By: Claude Opus 4.8 (1M context) --- benchmarks/single_node/speedbench/kimik3_fp4_b300_vllm.sh | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/benchmarks/single_node/speedbench/kimik3_fp4_b300_vllm.sh b/benchmarks/single_node/speedbench/kimik3_fp4_b300_vllm.sh index c1ee6c80f3..67514aa18b 100755 --- a/benchmarks/single_node/speedbench/kimik3_fp4_b300_vllm.sh +++ b/benchmarks/single_node/speedbench/kimik3_fp4_b300_vllm.sh @@ -371,7 +371,7 @@ run_cell() { --max-model-len "$MAX_MODEL_LEN" --max-cudagraph-capture-size 256 --attention-config '{"mla_prefill_backend":"FLASHINFER","use_prefill_query_quantization":true}' - --speculative-config "{\"method\": \"dspark\", \"model\": \"$DRAFT_MODEL_PATH\", \"num_speculative_tokens\": $mtp, \"attention_backend\": \"FLASHINFER_MLA\"}" + --speculative-config "{\"method\": \"dspark\", \"model\": \"$DRAFT_MODEL_PATH\", \"num_speculative_tokens\": $mtp, \"attention_backend\": \"FLASHINFER_MLA\", \"draft_sample_method\": \"probabilistic\", \"rejection_sample_method\": \"block\"}" ) local server_log="$RESULTS_DIR/server_${mode}_mtp${mtp}.log"