[bench] base=http://127.0.0.1:8000 model=/models/Qwen3.6-27B-AWQ-INT4
[bench] endpoint=chat thinking=True sampling=canonical(generation_config) max_tokens=256 ignore_eos=False prompts=40 levels=[1]

==== SUMMARY ====
  c=1  median_tok/s= 122.6 mean= 121.5 e2e= 118.2 agg= 116.4  TPOT=8.19ms  accept_len=4.28  out=256  finish={'length': 36, 'stop': 4}  ttft_ms[p50/p95]=78/79  [VALID]

---SAMPLE (clean, first 400 chars)---
To compute $47 \times 89$, we can use the standard vertical multiplication method or break it down using the distributive property. Here, I will demonstrate the standard vertical multiplication method step-by-step, as it is systematic and easy to follow.

### Step 1: Set up the vertical multiplication

$$
\begin{array}{r}
  47 \\
\times 89 \\
\hline
\end{array}
$$

### Step 2: Multiply by the ones
[mtp5_bf16_c1] MTP reduce kernels compiled (the _rw{N} live-diff proof):
     57 _rw1
     12 _rw8
--- full names ---
tqkv_unified_splitk_mtp_g4_v4_k4_v4_m2_v1_smem_rngfull_gms1_tt4_rw1_bp1
tqkv_unified_splitk_mtp_g4_v4_k4_v4_m3_v1_smem_rngfull_gms1_tt4_rw1_bp1
tqkv_unified_splitk_mtp_g4_v4_k4_v4_m4_v1_smem_rngfull_gms1_tt4_rw1_bp1
tqkv_unified_splitk_mtp_g4_v4_k4_v4_m5_v1_smem_rngfull_gms1_tt4_rw1_bp1
tqkv_unified_splitk_mtp_g4_v8_k4_v4_m2_v1_smem_rngfull_fq1_gms1_tt4_rw1
tqkv_unified_splitk_mtp_g4_v8_k4_v4_m2_v1_smem_rngfull_fq1_gms1_tt4_rw1_bp1
tqkv_unified_splitk_mtp_g4_v8_k4_v4_m2_v1_smem_rngfull_fq1_gms1_tt8_rw8_bp1
tqkv_unified_splitk_mtp_g4_v8_k4_v4_m2_v1_smem_rngfull_gms1_tt4_rw1_bp1
tqkv_unified_splitk_mtp_g4_v8_k4_v4_m2_v1_smem_rngfull_lse1_fq1_gms1_tt4_rw1
tqkv_unified_splitk_mtp_g4_v8_k4_v4_m2_v1_smem_rngfull_lse1_fq1_gms1_tt4_rw1_bp1
tqkv_unified_splitk_mtp_g4_v8_k4_v4_m3_v1_smem_rngfull_fq1_gms1_tt4_rw1
tqkv_unified_splitk_mtp_g4_v8_k4_v4_m3_v1_smem_rngfull_gms1_tt4_rw1_bp1
tqkv_unified_splitk_mtp_g4_v8_k4_v4_m4_v1_smem_rngfull_fq1_gms1_tt4_rw1
tqkv_unified_splitk_mtp_g4_v8_k4_v4_m4_v1_smem_rngfull_fq1_gms1_tt4_rw1_bp1
tqkv_unified_splitk_mtp_g4_v8_k4_v4_m4_v1_smem_rngfull_gms1_tt4_rw1_bp1
tqkv_unified_splitk_mtp_g4_v8_k4_v4_m4_v1_smem_rngfull_lse1_fq1_gms1_tt4_rw1
tqkv_unified_splitk_mtp_g4_v8_k4_v4_m4_v1_smem_rngfull_lse1_fq1_gms1_tt4_rw1_bp1
tqkv_unified_splitk_mtp_g4_v8_k4_v4_m5_v1_smem_rngfull_fq1_gms1_tt4_rw1
tqkv_unified_splitk_mtp_g4_v8_k4_v4_m6_v1_smem_rngfull_fq1_gms1_tt4_rw1
tqkv_unified_splitk_mtp_g4_v8_k4_v4_m7_v1_smem_rngfull_fq1_gms1_tt4_rw1
tqkv_unified_splitk_mtp_g4_v8_k4_v4_m8_v1_smem_rngfull_fq1_gms1_tt4_rw1
tqkv_unified_splitk_mtp_g4_v8_k4_v4_m8_v1_smem_rngfull_fq1_gms1_tt4_rw1_bp1
tqkv_unified_splitk_mtp_g4_v8_k4_v4_m8_v1_smem_rngfull_lse1_fq1_gms1_tt4_rw1
tqkv_unified_splitk_mtp_g4_v8_k4_v4_m8_v1_smem_rngfull_lse1_fq1_gms1_tt4_rw1_bp1
tqkv_unified_splitk_mtp_g6_v8_k4_v4_m10_v1_smem_rngfull_fq1_gms1_tt6_rw1_bp1
tqkv_unified_splitk_mtp_g6_v8_k4_v4_m10_v1_smem_rngfull_fq1_gms1_tt6_rw8_bp1
tqkv_unified_splitk_mtp_g6_v8_k4_v4_m11_v1_smem_rngfull_fq1_gms1_tt6_rw1_bp1
tqkv_unified_splitk_mtp_g6_v8_k4_v4_m11_v1_smem_rngfull_fq1_gms1_tt6_rw8_bp1
tqkv_unified_splitk_mtp_g6_v8_k4_v4_m11_v1_smem_rngfull_gms1_tt6_rw1_bp1
tqkv_unified_splitk_mtp_g6_v8_k4_v4_m12_v1_smem_rngfull_fq1_gms1_tt6_rw1_bp1
tqkv_unified_splitk_mtp_g6_v8_k4_v4_m12_v1_smem_rngfull_fq1_gms1_tt6_rw8_bp1
tqkv_unified_splitk_mtp_g6_v8_k4_v4_m2_v1_smem_rngfull_fq1_gms1_tt12_rw1_bp1
tqkv_unified_splitk_mtp_g6_v8_k4_v4_m2_v1_smem_rngfull_fq1_gms1_tt12_rw8_bp1
tqkv_unified_splitk_mtp_g6_v8_k4_v4_m2_v1_smem_rngfull_fq1_gms1_tt6_rw1
tqkv_unified_splitk_mtp_g6_v8_k4_v4_m2_v1_smem_rngfull_fq1_gms1_tt6_rw1_bp1
tqkv_unified_splitk_mtp_g6_v8_k4_v4_m2_v1_smem_rngfull_gms1_tt6_rw1_bp1
tqkv_unified_splitk_mtp_g6_v8_k4_v4_m2_v1_smem_rngfull_lse1_fq1_gms1_tt6_rw1_bp1
tqkv_unified_splitk_mtp_g6_v8_k4_v4_m3_v1_smem_rngfull_fq1_gms1_tt12_rw1_bp1
tqkv_unified_splitk_mtp_g6_v8_k4_v4_m3_v1_smem_rngfull_fq1_gms1_tt12_rw8_bp1
tqkv_unified_splitk_mtp_g6_v8_k4_v4_m3_v1_smem_rngfull_fq1_gms1_tt6_rw1
tqkv_unified_splitk_mtp_g6_v8_k4_v4_m3_v1_smem_rngfull_gms1_tt6_rw1_bp1
tqkv_unified_splitk_mtp_g6_v8_k4_v4_m4_v1_smem_rngfull_fq1_gms1_tt12_rw1_bp1
tqkv_unified_splitk_mtp_g6_v8_k4_v4_m4_v1_smem_rngfull_fq1_gms1_tt12_rw8_bp1
tqkv_unified_splitk_mtp_g6_v8_k4_v4_m4_v1_smem_rngfull_fq1_gms1_tt6_rw1
tqkv_unified_splitk_mtp_g6_v8_k4_v4_m4_v1_smem_rngfull_fq1_gms1_tt6_rw1_bp1
tqkv_unified_splitk_mtp_g6_v8_k4_v4_m4_v1_smem_rngfull_gms1_tt6_rw1_bp1
tqkv_unified_splitk_mtp_g6_v8_k4_v4_m4_v1_smem_rngfull_lse1_fq1_gms1_tt6_rw1
tqkv_unified_splitk_mtp_g6_v8_k4_v4_m4_v1_smem_rngfull_lse1_fq1_gms1_tt6_rw1_bp1
tqkv_unified_splitk_mtp_g6_v8_k4_v4_m5_v1_smem_rngfull_fq1_gms1_tt6_rw1
tqkv_unified_splitk_mtp_g6_v8_k4_v4_m5_v1_smem_rngfull_fq1_gms1_tt9_rw1_bp1
tqkv_unified_splitk_mtp_g6_v8_k4_v4_m5_v1_smem_rngfull_fq1_gms1_tt9_rw8_bp1
tqkv_unified_splitk_mtp_g6_v8_k4_v4_m5_v1_smem_rngfull_gms1_tt6_rw1_bp1
tqkv_unified_splitk_mtp_g6_v8_k4_v4_m6_v1_smem_rngfull_fq1_gms1_tt6_rw1
tqkv_unified_splitk_mtp_g6_v8_k4_v4_m6_v1_smem_rngfull_fq1_gms1_tt8_rw1_bp1
tqkv_unified_splitk_mtp_g6_v8_k4_v4_m6_v1_smem_rngfull_fq1_gms1_tt8_rw8_bp1
tqkv_unified_splitk_mtp_g6_v8_k4_v4_m6_v1_smem_rngfull_gms1_tt6_rw1_bp1
tqkv_unified_splitk_mtp_g6_v8_k4_v4_m7_v1_smem_rngfull_fq1_gms1_tt6_rw1
tqkv_unified_splitk_mtp_g6_v8_k4_v4_m7_v1_smem_rngfull_fq1_gms1_tt6_rw1_bp1
tqkv_unified_splitk_mtp_g6_v8_k4_v4_m7_v1_smem_rngfull_fq1_gms1_tt6_rw8_bp1
tqkv_unified_splitk_mtp_g6_v8_k4_v4_m7_v1_smem_rngfull_gms1_tt6_rw1_bp1
tqkv_unified_splitk_mtp_g6_v8_k4_v4_m8_v1_smem_rngfull_fq1_gms1_tt6_rw1
tqkv_unified_splitk_mtp_g6_v8_k4_v4_m8_v1_smem_rngfull_fq1_gms1_tt6_rw1_bp1
tqkv_unified_splitk_mtp_g6_v8_k4_v4_m8_v1_smem_rngfull_fq1_gms1_tt6_rw8_bp1
tqkv_unified_splitk_mtp_g6_v8_k4_v4_m8_v1_smem_rngfull_gms1_tt6_rw1_bp1
tqkv_unified_splitk_mtp_g6_v8_k4_v4_m8_v1_smem_rngfull_lse1_fq1_gms1_tt6_rw1
tqkv_unified_splitk_mtp_g6_v8_k4_v4_m8_v1_smem_rngfull_lse1_fq1_gms1_tt6_rw1_bp1
tqkv_unified_splitk_mtp_g6_v8_k4_v4_m9_v1_smem_rngfull_fq1_gms1_tt6_rw1_bp1
tqkv_unified_splitk_mtp_g6_v8_k4_v4_m9_v1_smem_rngfull_fq1_gms1_tt6_rw8_bp1
tqkv_unified_splitk_mtp_g6_v8_k4_v4_m9_v1_smem_rngfull_gms1_tt6_rw1_bp1
