# qwen36 CUDA chain, a card holding part of the experts (another L4 VM): raw evidence (2026-10-25)
# Google Colab L4 VM: Intel(R) Xeon(R) CPU @ 2.20GHz (6 cores / 12 threads, 38.5 MiB L3) + NVIDIA L4 (23034 MiB, sm_89), 52 GiB; Ubuntu 24.04.4 LTS, Linux 6.6.122+ x86_64 (Google Colab), CUDA 13.0 (nvcc V13.0.88), driver 580.82.07, gcc (Ubuntu 13.3.0-6ubuntu2~24.04.1) 13.3.0.
# Model: Kreuzzelg/qwen36-35b-a3b-colibri-i4-gs64 (CUDA_EXPERT_GB=5, no HEAT_FILE: the warm
#   start fills the card in index order, the same experts in every run; the rest served by the host;
#   the tier and the chain's promotions as by default).
# v0 = origin/dev 7bbea341; v1 = 7bbea341 + this change (the measured build's c/ files outside
#   c/tests/ are this commit's). Harness: bench.c beside this file (sha256 edbe7ddcc146af34d10c3d8a871fb1d3b93404537e374f9bc1504cf2c6886c17),
#   built against each version's c/ with:
#   cd c && make qwen36_tier.o backend_cuda.o cuda_chain.o cu_xb.o cu_tier.o CUDA=1 CUDA_ARCH=sm_89 && gcc -O3 -march=native -fopenmp -pthread -Wall -Wextra -Wno-unused-parameter -Wno-misleading-indentation -Wno-unused-function -DCOLI_CUDA -I. <bench.c> qwen36_tier.o backend_cuda.o cuda_chain.o cu_xb.o cu_tier.o -o qwen36-bench -lm -fopenmp -pthread -L/usr/local/cuda/lib64 -Wl,-rpath,/usr/local/cuda/lib64 -lcudart -lcublasLt -lcublas -lstdc++
#   (Its comments name the driver and scripts it came with; the runs below need only this file.)
# Text: promessi_sposi.txt, speed-bench/promessi_sposi.txt of antirez's ds4 at commit 0aaea5a (sha256
#   f53e0d80cb2d4492d24ebd63c7000c397b16ae70f9bf09b3763e5d8323ec209f); for each length its first N tokens from an empty state, then 32 decode
#   steps fed the text's own next tokens (--teacher). One process a run, the two lengths in order.
# A warm-up run of v0 (not counted), then v0 v1 v1 v0 v0 v1 v1 v0, each once the card was under 70 C
#   with no thermal clock event, then 60 s of rest. Every run's CSV and stderr in full, in the order they ran;
#   decode tok/s = gen_steady_tps (steps 2 to 32), the summary at the end.

===== setup
bench 482e1711d589; engine qwen36; Linux (Intel(R) Xeon(R) CPU @ 2.20GHz, NVIDIA L4); 2026-10-25 14:44
model /content/q36; text promessi_sposi.txt (sha256 f53e0d80cb2d4492); lengths 64,512; gen 32; teacher-forced
runs: a warm-up of v0 (does not count), then 4 per version, the order turning at every round (ABBA); before every run: wait until cool (at most 900 s), then 60 s of rest
v0 dev: /content/bin-dev/qwen36-bench (commit 7bbea341bb4bf9c812fed9d80501cf5fc1615042; harness qwen36_bench.c edbe7ddcc146; CUDA); env: COLI_CUDA=1 COLI_GPUS=0 CUDA_EXPERT_GB=auto OMP_WAIT_POLICY=ACTIVE OMP_PROC_BIND=close CUDA_EXPERT_GB=5
   command: env -i PATH=/usr/bin:/bin HOME=/root LD_LIBRARY_PATH=/usr/lib64-nvidia COLI_CUDA=1 COLI_GPUS=0 CUDA_EXPERT_GB=auto OMP_WAIT_POLICY=ACTIVE OMP_PROC_BIND=close CUDA_EXPERT_GB=5 /content/bin-dev/qwen36-bench --model /content/q36 --text /content/bench/promessi_sposi.txt --lengths 64,512 --gen 32 --teacher --steps /content/misura-hyb/v0-runN.steps
v1 pr: /content/bin-pr/qwen36-bench (commit 1fb2ee782dfd12b243e7a8cd81e2f7748d45823c; harness qwen36_bench.c edbe7ddcc146; CUDA); env: COLI_CUDA=1 COLI_GPUS=0 CUDA_EXPERT_GB=auto OMP_WAIT_POLICY=ACTIVE OMP_PROC_BIND=close CUDA_EXPERT_GB=5
   command: env -i PATH=/usr/bin:/bin HOME=/root LD_LIBRARY_PATH=/usr/lib64-nvidia COLI_CUDA=1 COLI_GPUS=0 CUDA_EXPERT_GB=auto OMP_WAIT_POLICY=ACTIVE OMP_PROC_BIND=close CUDA_EXPERT_GB=5 /content/bin-pr/qwen36-bench --model /content/q36 --text /content/bench/promessi_sposi.txt --lengths 64,512 --gen 32 --teacher --steps /content/misura-hyb/v1-runN.steps

===== warmup.csv (sha256 18836486f0ae7afb95fd963c92a8eaa95ab69fb9a6a0bd89fa3ff9d5ef173267)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,1.433,44.67,32,68.5,29.318,0,0.00,1e40892a3a7ff8df,5cb23346ab6be1d8,2.70028,0.469
512,512,1.654,309.61,32,34.2,31.537,0,0.00,649cee2cbca20368,7a8848450b863c49,2.96052,0.438
===== warmup.err (sha256 bb60d7528efa8302fa84e8273f996f7be3f861bfc6908f18ec468858f1392f01)
[tok] loaded 248044 pieces (max id 248069) from /content/q36/tokenizer.json
[OMP] qwen36: 6 physical-core threads instead of 12 logical CPUs; SMT can halve decode throughput on some CPUs (#718); set OMP_NUM_THREADS=<n> to override
== qwen36 Phase-2 engine | cache=256/layer bits=4 ctx=8192 pilot=0 wide=1 hot=0 smooth=0.30 conf=0.92 ==
[tok] loaded 248044 pieces (max id 248069) from /content/q36/tokenizer.json
[meta] loaded: q_heads=16 kv_heads=2 head_dim=256 q_head_dim=512 k_head_dim=256 v_head_dim=256 o_in=4096 rotary_dim=64 n_experts=256 topk=8 inter=512 shared_inter=512 n_group=1 topk_group=1 attn_output_gate=1 n_active=10
[meta] DeltaNet: vheads=32 kheads=16 kdim=128 vdim=128 convk=4 conv_dim=8192
[qwen36] embedding: rows read in place from the checkpoint, COLI_EMBED_MMAP=0 builds the f32 table
[dense-i8] 291 matrices quantized during load, 7.2 GB f32 freed
[qwen36] group-scaled experts: gs=64
resident weights loaded in 8.9s | RSS after load: 3.03 GB
[qwen36] expert format on disk: int4 packed (tier fmt=4)
[qwen36] expert kernel: planar int4 in RAM (expert_ffn.h), QWEN_EXPERT_KERNEL=0 restores int8 unpack
[CUDA] device 0: NVIDIA L4, 23.7 GB VRAM, sm_89
[qtier] dev 0: 21.8 GB free, allowance 5.0 GB
[place] auto: dev 0 holds 1828.5 MB of trunk (lmhead yes, 30 dnproj layers), 3.21 GB left for experts
[qtier] dev 0: budget 3.21 GB for experts (~1823 experts) after trunk
[qtier] CUDA VRAM expert tier active: 1 device(s), 1.80 MB/expert
[CUDA] mode: routed experts (qwen36 VRAM tier)
[gpu] MoE experts -> CUDA VRAM tier
[place] probe: one [8192 x 2048] int8 GEMV takes 0.112 ms on CUDA dev 0, 0.416 ms on the CPU -> trunk to VRAM
[lmh] lm_head [248320 x 2048] int8 resident on CUDA dev 0 (0.47 GB)
[dnp] 30 DeltaNet-Projektionen auf GPU (0.70 GB VRAM)
[dense] 190 trunk matrices on GPU (30 dnout, 40 attnproj, 120 shexp; 0.61 GB VRAM)
[chain] qwen36: a decode step's routed experts in the frame (the tier's residents by address, format 4, group 64; COLI_CUDA_CHAIN_MOE=0: the host's step)
[chain] qwen36: 40 layers on CUDA device 0 (10 attention), lm_head too; 4.4 MiB of parameters, 35.3 MiB of router, b|a and gate rows, 66 MB of DeltaNet state
[qwen36] int4 packed weights detected — kept int4, repacked planar for expert_ffn.h
[qtier] host pool: 8417 experts of the RAM cache page-locked for the device to copy a prompt chunk's experts from (COLI_CUDA_HOST_POOL=0 off)
[qtier] warmstart (parallel): all 10240 experts in RAM (planar int4, the only copy), 1823 in VRAM -- 7.1 s
[bench] CUDA ready in 17.14 s; text 6182 tokens; lengths 64,512; gen 32; each length from an empty state
[qwen36] KV cache: float16, 0.00 GiB at 96 positions (half of float32; Q36_KV=f32 keeps float32)
[chain] qwen36: 4096 rows per prompt chunk (286 KiB a row, 16.9 GB free on the device; COLI_CUDA_CHAIN_ROWS sets it)
[chain] qwen36: up to 4 experts a step promoted into the card by this conversation's routes, the coldest residents giving way, none while the last 64 steps hit 75 % or more (COLI_CUDA_CHAIN_PROMOTE=0 keeps the warm start's set)
[bench] peak RSS 19.37 GB
[qtier] host pool: 8417 experts (13.87 GB of the RAM cache) page-locked for the device's copies

===== v0-run1.csv (sha256 0143b2324d55e0d8e8654e4c116450da17a14462518d7b38901f9bce7f0b9a45)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,1.439,44.48,32,68.2,29.278,0,0.00,1e40892a3a7ff8df,5cb23346ab6be1d8,2.70028,0.469
512,512,1.668,306.94,32,33.9,31.585,0,0.00,649cee2cbca20368,7a8848450b863c49,2.96052,0.438
===== v0-run1.err (sha256 3b23bb17aa341ab153d5b6de8f8145c125cebf52d305dac0b4469ad8fc226c35)
[tok] loaded 248044 pieces (max id 248069) from /content/q36/tokenizer.json
[OMP] qwen36: 6 physical-core threads instead of 12 logical CPUs; SMT can halve decode throughput on some CPUs (#718); set OMP_NUM_THREADS=<n> to override
== qwen36 Phase-2 engine | cache=256/layer bits=4 ctx=8192 pilot=0 wide=1 hot=0 smooth=0.30 conf=0.92 ==
[tok] loaded 248044 pieces (max id 248069) from /content/q36/tokenizer.json
[meta] loaded: q_heads=16 kv_heads=2 head_dim=256 q_head_dim=512 k_head_dim=256 v_head_dim=256 o_in=4096 rotary_dim=64 n_experts=256 topk=8 inter=512 shared_inter=512 n_group=1 topk_group=1 attn_output_gate=1 n_active=10
[meta] DeltaNet: vheads=32 kheads=16 kdim=128 vdim=128 convk=4 conv_dim=8192
[qwen36] embedding: rows read in place from the checkpoint, COLI_EMBED_MMAP=0 builds the f32 table
[dense-i8] 291 matrices quantized during load, 7.2 GB f32 freed
[qwen36] group-scaled experts: gs=64
resident weights loaded in 9.0s | RSS after load: 3.03 GB
[qwen36] expert format on disk: int4 packed (tier fmt=4)
[qwen36] expert kernel: planar int4 in RAM (expert_ffn.h), QWEN_EXPERT_KERNEL=0 restores int8 unpack
[CUDA] device 0: NVIDIA L4, 23.7 GB VRAM, sm_89
[qtier] dev 0: 21.8 GB free, allowance 5.0 GB
[place] auto: dev 0 holds 1828.5 MB of trunk (lmhead yes, 30 dnproj layers), 3.21 GB left for experts
[qtier] dev 0: budget 3.21 GB for experts (~1823 experts) after trunk
[qtier] CUDA VRAM expert tier active: 1 device(s), 1.80 MB/expert
[CUDA] mode: routed experts (qwen36 VRAM tier)
[gpu] MoE experts -> CUDA VRAM tier
[place] probe: one [8192 x 2048] int8 GEMV takes 0.113 ms on CUDA dev 0, 0.420 ms on the CPU -> trunk to VRAM
[lmh] lm_head [248320 x 2048] int8 resident on CUDA dev 0 (0.47 GB)
[dnp] 30 DeltaNet-Projektionen auf GPU (0.70 GB VRAM)
[dense] 190 trunk matrices on GPU (30 dnout, 40 attnproj, 120 shexp; 0.61 GB VRAM)
[chain] qwen36: a decode step's routed experts in the frame (the tier's residents by address, format 4, group 64; COLI_CUDA_CHAIN_MOE=0: the host's step)
[chain] qwen36: 40 layers on CUDA device 0 (10 attention), lm_head too; 4.4 MiB of parameters, 35.3 MiB of router, b|a and gate rows, 66 MB of DeltaNet state
[qwen36] int4 packed weights detected — kept int4, repacked planar for expert_ffn.h
[qtier] host pool: 8417 experts of the RAM cache page-locked for the device to copy a prompt chunk's experts from (COLI_CUDA_HOST_POOL=0 off)
[qtier] warmstart (parallel): all 10240 experts in RAM (planar int4, the only copy), 1823 in VRAM -- 11.2 s
[bench] CUDA ready in 21.35 s; text 6182 tokens; lengths 64,512; gen 32; each length from an empty state
[qwen36] KV cache: float16, 0.00 GiB at 96 positions (half of float32; Q36_KV=f32 keeps float32)
[chain] qwen36: 4096 rows per prompt chunk (286 KiB a row, 16.9 GB free on the device; COLI_CUDA_CHAIN_ROWS sets it)
[chain] qwen36: up to 4 experts a step promoted into the card by this conversation's routes, the coldest residents giving way, none while the last 64 steps hit 75 % or more (COLI_CUDA_CHAIN_PROMOTE=0 keeps the warm start's set)
[bench] peak RSS 19.37 GB
[qtier] host pool: 8417 experts (13.87 GB of the RAM cache) page-locked for the device's copies

===== v1-run1.csv (sha256 3f1c23907d030c965e4a6b97ed0cf1d599956bb04ec7a6d4c51ab267d4c93541)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,1.431,44.72,32,68.5,29.393,0,0.00,1e40892a3a7ff8df,5cb23346ab6be1d8,2.70028,0.469
512,512,1.669,306.74,32,33.8,31.412,0,0.00,799d16315c29dd05,7a8848450b863c49,2.95938,0.438
===== v1-run1.err (sha256 07143af825476f9b5bed3635045768601d1868c2d4f246bc3dba81ff5202efe1)
[tok] loaded 248044 pieces (max id 248069) from /content/q36/tokenizer.json
[OMP] qwen36: 6 physical-core threads instead of 12 logical CPUs; SMT can halve decode throughput on some CPUs (#718); set OMP_NUM_THREADS=<n> to override
== qwen36 Phase-2 engine | cache=256/layer bits=4 ctx=8192 pilot=0 wide=1 hot=0 smooth=0.30 conf=0.92 ==
[tok] loaded 248044 pieces (max id 248069) from /content/q36/tokenizer.json
[meta] loaded: q_heads=16 kv_heads=2 head_dim=256 q_head_dim=512 k_head_dim=256 v_head_dim=256 o_in=4096 rotary_dim=64 n_experts=256 topk=8 inter=512 shared_inter=512 n_group=1 topk_group=1 attn_output_gate=1 n_active=10
[meta] DeltaNet: vheads=32 kheads=16 kdim=128 vdim=128 convk=4 conv_dim=8192
[qwen36] embedding: rows read in place from the checkpoint, COLI_EMBED_MMAP=0 builds the f32 table
[dense-i8] 291 matrices quantized during load, 7.2 GB f32 freed
[qwen36] group-scaled experts: gs=64
resident weights loaded in 9.1s | RSS after load: 3.03 GB
[qwen36] expert format on disk: int4 packed (tier fmt=4)
[qwen36] expert kernel: planar int4 in RAM (expert_ffn.h), QWEN_EXPERT_KERNEL=0 restores int8 unpack
[CUDA] device 0: NVIDIA L4, 23.7 GB VRAM, sm_89
[qtier] dev 0: 21.8 GB free, allowance 5.0 GB
[place] auto: dev 0 holds 1828.5 MB of trunk (lmhead yes, 30 dnproj layers), 3.21 GB left for experts
[qtier] dev 0: budget 3.21 GB for experts (~1823 experts) after trunk
[qtier] CUDA VRAM expert tier active: 1 device(s), 1.80 MB/expert
[CUDA] mode: routed experts (qwen36 VRAM tier)
[gpu] MoE experts -> CUDA VRAM tier
[place] probe: one [8192 x 2048] int8 GEMV takes 0.113 ms on CUDA dev 0, 0.416 ms on the CPU -> trunk to VRAM
[lmh] lm_head [248320 x 2048] int8 resident on CUDA dev 0 (0.47 GB)
[dnp] 30 DeltaNet-Projektionen auf GPU (0.70 GB VRAM)
[dense] 190 trunk matrices on GPU (30 dnout, 40 attnproj, 120 shexp; 0.61 GB VRAM)
[chain] qwen36: a decode step's routed experts in the frame (the tier's residents by address, format 4, group 64; COLI_CUDA_CHAIN_MOE=0: the host's step)
[chain] qwen36: 40 layers on CUDA device 0 (10 attention), lm_head too; 4.4 MiB of parameters, 35.3 MiB of router, b|a and gate rows, 66 MB of DeltaNet state
[qwen36] int4 packed weights detected — kept int4, repacked planar for expert_ffn.h
[qtier] host pool: 8417 experts of the RAM cache page-locked for the device to copy a prompt chunk's experts from (COLI_CUDA_HOST_POOL=0 off)
[qtier] warmstart (parallel): all 10240 experts in RAM (planar int4, the only copy), 1823 in VRAM -- 10.5 s
[bench] CUDA ready in 20.71 s; text 6182 tokens; lengths 64,512; gen 32; each length from an empty state
[qwen36] KV cache: float16, 0.00 GiB at 96 positions (half of float32; Q36_KV=f32 keeps float32)
[chain] qwen36: 4096 rows per prompt chunk (286 KiB a row, 16.9 GB free on the device; COLI_CUDA_CHAIN_ROWS sets it)
[chain] qwen36: up to 4 experts a step promoted into the card by this conversation's routes, the coldest residents giving way, none while the last 64 steps hit 75 % or more (COLI_CUDA_CHAIN_PROMOTE=0 keeps the warm start's set)
[bench] peak RSS 19.37 GB
[qtier] host pool: 8417 experts (13.87 GB of the RAM cache) page-locked for the device's copies

===== v1-run2.csv (sha256 71091ddd9b50956e572e0ae42c39bcde4dd0497f61f362146353c8ca93415a11)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,1.433,44.65,32,67.6,29.337,0,0.00,1e40892a3a7ff8df,5cb23346ab6be1d8,2.70028,0.469
512,512,1.664,307.67,32,33.6,31.612,0,0.00,95505aeafaa41c33,7a8848450b863c49,2.96304,0.438
===== v1-run2.err (sha256 5aac7bed690a3e40f5642aed0ed4586d6d1951c1046b35c51b4a24e8dd3225e5)
[tok] loaded 248044 pieces (max id 248069) from /content/q36/tokenizer.json
[OMP] qwen36: 6 physical-core threads instead of 12 logical CPUs; SMT can halve decode throughput on some CPUs (#718); set OMP_NUM_THREADS=<n> to override
== qwen36 Phase-2 engine | cache=256/layer bits=4 ctx=8192 pilot=0 wide=1 hot=0 smooth=0.30 conf=0.92 ==
[tok] loaded 248044 pieces (max id 248069) from /content/q36/tokenizer.json
[meta] loaded: q_heads=16 kv_heads=2 head_dim=256 q_head_dim=512 k_head_dim=256 v_head_dim=256 o_in=4096 rotary_dim=64 n_experts=256 topk=8 inter=512 shared_inter=512 n_group=1 topk_group=1 attn_output_gate=1 n_active=10
[meta] DeltaNet: vheads=32 kheads=16 kdim=128 vdim=128 convk=4 conv_dim=8192
[qwen36] embedding: rows read in place from the checkpoint, COLI_EMBED_MMAP=0 builds the f32 table
[dense-i8] 291 matrices quantized during load, 7.2 GB f32 freed
[qwen36] group-scaled experts: gs=64
resident weights loaded in 9.1s | RSS after load: 3.03 GB
[qwen36] expert format on disk: int4 packed (tier fmt=4)
[qwen36] expert kernel: planar int4 in RAM (expert_ffn.h), QWEN_EXPERT_KERNEL=0 restores int8 unpack
[CUDA] device 0: NVIDIA L4, 23.7 GB VRAM, sm_89
[qtier] dev 0: 21.8 GB free, allowance 5.0 GB
[place] auto: dev 0 holds 1828.5 MB of trunk (lmhead yes, 30 dnproj layers), 3.21 GB left for experts
[qtier] dev 0: budget 3.21 GB for experts (~1823 experts) after trunk
[qtier] CUDA VRAM expert tier active: 1 device(s), 1.80 MB/expert
[CUDA] mode: routed experts (qwen36 VRAM tier)
[gpu] MoE experts -> CUDA VRAM tier
[place] probe: one [8192 x 2048] int8 GEMV takes 0.112 ms on CUDA dev 0, 0.423 ms on the CPU -> trunk to VRAM
[lmh] lm_head [248320 x 2048] int8 resident on CUDA dev 0 (0.47 GB)
[dnp] 30 DeltaNet-Projektionen auf GPU (0.70 GB VRAM)
[dense] 190 trunk matrices on GPU (30 dnout, 40 attnproj, 120 shexp; 0.61 GB VRAM)
[chain] qwen36: a decode step's routed experts in the frame (the tier's residents by address, format 4, group 64; COLI_CUDA_CHAIN_MOE=0: the host's step)
[chain] qwen36: 40 layers on CUDA device 0 (10 attention), lm_head too; 4.4 MiB of parameters, 35.3 MiB of router, b|a and gate rows, 66 MB of DeltaNet state
[qwen36] int4 packed weights detected — kept int4, repacked planar for expert_ffn.h
[qtier] host pool: 8417 experts of the RAM cache page-locked for the device to copy a prompt chunk's experts from (COLI_CUDA_HOST_POOL=0 off)
[qtier] warmstart (parallel): all 10240 experts in RAM (planar int4, the only copy), 1823 in VRAM -- 7.6 s
[bench] CUDA ready in 17.80 s; text 6182 tokens; lengths 64,512; gen 32; each length from an empty state
[qwen36] KV cache: float16, 0.00 GiB at 96 positions (half of float32; Q36_KV=f32 keeps float32)
[chain] qwen36: 4096 rows per prompt chunk (286 KiB a row, 16.9 GB free on the device; COLI_CUDA_CHAIN_ROWS sets it)
[chain] qwen36: up to 4 experts a step promoted into the card by this conversation's routes, the coldest residents giving way, none while the last 64 steps hit 75 % or more (COLI_CUDA_CHAIN_PROMOTE=0 keeps the warm start's set)
[bench] peak RSS 19.37 GB
[qtier] host pool: 8417 experts (13.87 GB of the RAM cache) page-locked for the device's copies

===== v0-run2.csv (sha256 37dcde908334c64378e6f1f8ceab1aaa8dd8d46db78a37d586435b086d44953b)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,1.435,44.60,32,68.2,29.203,0,0.00,1e40892a3a7ff8df,5cb23346ab6be1d8,2.70028,0.469
512,512,1.666,307.38,32,34.3,31.540,0,0.00,649cee2cbca20368,7a8848450b863c49,2.96052,0.438
===== v0-run2.err (sha256 71aadc52893ce998ae3d73798e6555c560d6b15aff33c256360b567533b5ec34)
[tok] loaded 248044 pieces (max id 248069) from /content/q36/tokenizer.json
[OMP] qwen36: 6 physical-core threads instead of 12 logical CPUs; SMT can halve decode throughput on some CPUs (#718); set OMP_NUM_THREADS=<n> to override
== qwen36 Phase-2 engine | cache=256/layer bits=4 ctx=8192 pilot=0 wide=1 hot=0 smooth=0.30 conf=0.92 ==
[tok] loaded 248044 pieces (max id 248069) from /content/q36/tokenizer.json
[meta] loaded: q_heads=16 kv_heads=2 head_dim=256 q_head_dim=512 k_head_dim=256 v_head_dim=256 o_in=4096 rotary_dim=64 n_experts=256 topk=8 inter=512 shared_inter=512 n_group=1 topk_group=1 attn_output_gate=1 n_active=10
[meta] DeltaNet: vheads=32 kheads=16 kdim=128 vdim=128 convk=4 conv_dim=8192
[qwen36] embedding: rows read in place from the checkpoint, COLI_EMBED_MMAP=0 builds the f32 table
[dense-i8] 291 matrices quantized during load, 7.2 GB f32 freed
[qwen36] group-scaled experts: gs=64
resident weights loaded in 9.1s | RSS after load: 3.03 GB
[qwen36] expert format on disk: int4 packed (tier fmt=4)
[qwen36] expert kernel: planar int4 in RAM (expert_ffn.h), QWEN_EXPERT_KERNEL=0 restores int8 unpack
[CUDA] device 0: NVIDIA L4, 23.7 GB VRAM, sm_89
[qtier] dev 0: 21.8 GB free, allowance 5.0 GB
[place] auto: dev 0 holds 1828.5 MB of trunk (lmhead yes, 30 dnproj layers), 3.21 GB left for experts
[qtier] dev 0: budget 3.21 GB for experts (~1823 experts) after trunk
[qtier] CUDA VRAM expert tier active: 1 device(s), 1.80 MB/expert
[CUDA] mode: routed experts (qwen36 VRAM tier)
[gpu] MoE experts -> CUDA VRAM tier
[place] probe: one [8192 x 2048] int8 GEMV takes 0.113 ms on CUDA dev 0, 0.418 ms on the CPU -> trunk to VRAM
[lmh] lm_head [248320 x 2048] int8 resident on CUDA dev 0 (0.47 GB)
[dnp] 30 DeltaNet-Projektionen auf GPU (0.70 GB VRAM)
[dense] 190 trunk matrices on GPU (30 dnout, 40 attnproj, 120 shexp; 0.61 GB VRAM)
[chain] qwen36: a decode step's routed experts in the frame (the tier's residents by address, format 4, group 64; COLI_CUDA_CHAIN_MOE=0: the host's step)
[chain] qwen36: 40 layers on CUDA device 0 (10 attention), lm_head too; 4.4 MiB of parameters, 35.3 MiB of router, b|a and gate rows, 66 MB of DeltaNet state
[qwen36] int4 packed weights detected — kept int4, repacked planar for expert_ffn.h
[qtier] host pool: 8417 experts of the RAM cache page-locked for the device to copy a prompt chunk's experts from (COLI_CUDA_HOST_POOL=0 off)
[qtier] warmstart (parallel): all 10240 experts in RAM (planar int4, the only copy), 1823 in VRAM -- 7.6 s
[bench] CUDA ready in 17.75 s; text 6182 tokens; lengths 64,512; gen 32; each length from an empty state
[qwen36] KV cache: float16, 0.00 GiB at 96 positions (half of float32; Q36_KV=f32 keeps float32)
[chain] qwen36: 4096 rows per prompt chunk (286 KiB a row, 16.9 GB free on the device; COLI_CUDA_CHAIN_ROWS sets it)
[chain] qwen36: up to 4 experts a step promoted into the card by this conversation's routes, the coldest residents giving way, none while the last 64 steps hit 75 % or more (COLI_CUDA_CHAIN_PROMOTE=0 keeps the warm start's set)
[bench] peak RSS 19.37 GB
[qtier] host pool: 8417 experts (13.87 GB of the RAM cache) page-locked for the device's copies

===== v0-run3.csv (sha256 d86238f7c2ab03bda1f2e3974a58252050882c0a7d66abd6199401f0615ff14c)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,1.426,44.88,32,68.4,29.160,0,0.00,1e40892a3a7ff8df,5cb23346ab6be1d8,2.70028,0.469
512,512,1.687,303.57,32,34.3,31.361,0,0.00,649cee2cbca20368,7a8848450b863c49,2.96052,0.438
===== v0-run3.err (sha256 cf19cff551139a2af5ce5f92ab53cbcd0895ef35681b9ccf0699b85ea2b929d5)
[tok] loaded 248044 pieces (max id 248069) from /content/q36/tokenizer.json
[OMP] qwen36: 6 physical-core threads instead of 12 logical CPUs; SMT can halve decode throughput on some CPUs (#718); set OMP_NUM_THREADS=<n> to override
== qwen36 Phase-2 engine | cache=256/layer bits=4 ctx=8192 pilot=0 wide=1 hot=0 smooth=0.30 conf=0.92 ==
[tok] loaded 248044 pieces (max id 248069) from /content/q36/tokenizer.json
[meta] loaded: q_heads=16 kv_heads=2 head_dim=256 q_head_dim=512 k_head_dim=256 v_head_dim=256 o_in=4096 rotary_dim=64 n_experts=256 topk=8 inter=512 shared_inter=512 n_group=1 topk_group=1 attn_output_gate=1 n_active=10
[meta] DeltaNet: vheads=32 kheads=16 kdim=128 vdim=128 convk=4 conv_dim=8192
[qwen36] embedding: rows read in place from the checkpoint, COLI_EMBED_MMAP=0 builds the f32 table
[dense-i8] 291 matrices quantized during load, 7.2 GB f32 freed
[qwen36] group-scaled experts: gs=64
resident weights loaded in 9.1s | RSS after load: 3.03 GB
[qwen36] expert format on disk: int4 packed (tier fmt=4)
[qwen36] expert kernel: planar int4 in RAM (expert_ffn.h), QWEN_EXPERT_KERNEL=0 restores int8 unpack
[CUDA] device 0: NVIDIA L4, 23.7 GB VRAM, sm_89
[qtier] dev 0: 21.8 GB free, allowance 5.0 GB
[place] auto: dev 0 holds 1828.5 MB of trunk (lmhead yes, 30 dnproj layers), 3.21 GB left for experts
[qtier] dev 0: budget 3.21 GB for experts (~1823 experts) after trunk
[qtier] CUDA VRAM expert tier active: 1 device(s), 1.80 MB/expert
[CUDA] mode: routed experts (qwen36 VRAM tier)
[gpu] MoE experts -> CUDA VRAM tier
[place] probe: one [8192 x 2048] int8 GEMV takes 0.111 ms on CUDA dev 0, 0.417 ms on the CPU -> trunk to VRAM
[lmh] lm_head [248320 x 2048] int8 resident on CUDA dev 0 (0.47 GB)
[dnp] 30 DeltaNet-Projektionen auf GPU (0.70 GB VRAM)
[dense] 190 trunk matrices on GPU (30 dnout, 40 attnproj, 120 shexp; 0.61 GB VRAM)
[chain] qwen36: a decode step's routed experts in the frame (the tier's residents by address, format 4, group 64; COLI_CUDA_CHAIN_MOE=0: the host's step)
[chain] qwen36: 40 layers on CUDA device 0 (10 attention), lm_head too; 4.4 MiB of parameters, 35.3 MiB of router, b|a and gate rows, 66 MB of DeltaNet state
[qwen36] int4 packed weights detected — kept int4, repacked planar for expert_ffn.h
[qtier] host pool: 8417 experts of the RAM cache page-locked for the device to copy a prompt chunk's experts from (COLI_CUDA_HOST_POOL=0 off)
[qtier] warmstart (parallel): all 10240 experts in RAM (planar int4, the only copy), 1823 in VRAM -- 7.6 s
[bench] CUDA ready in 17.81 s; text 6182 tokens; lengths 64,512; gen 32; each length from an empty state
[qwen36] KV cache: float16, 0.00 GiB at 96 positions (half of float32; Q36_KV=f32 keeps float32)
[chain] qwen36: 4096 rows per prompt chunk (286 KiB a row, 16.9 GB free on the device; COLI_CUDA_CHAIN_ROWS sets it)
[chain] qwen36: up to 4 experts a step promoted into the card by this conversation's routes, the coldest residents giving way, none while the last 64 steps hit 75 % or more (COLI_CUDA_CHAIN_PROMOTE=0 keeps the warm start's set)
[bench] peak RSS 19.37 GB
[qtier] host pool: 8417 experts (13.87 GB of the RAM cache) page-locked for the device's copies

===== v1-run3.csv (sha256 d15569596ec4f4a24d1ea0bd27d79a406b8fb19358ce22d8b7631826ce814d28)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,1.444,44.33,32,68.2,29.331,0,0.00,1e40892a3a7ff8df,5cb23346ab6be1d8,2.70028,0.469
512,512,1.670,306.50,32,33.7,31.595,0,0.00,fc31d7e817064462,7a8848450b863c49,2.97693,0.438
===== v1-run3.err (sha256 ff1e7a495cd2861ab041a1cfbcf81063e23691fc54d1ed681d890757495fcf87)
[tok] loaded 248044 pieces (max id 248069) from /content/q36/tokenizer.json
[OMP] qwen36: 6 physical-core threads instead of 12 logical CPUs; SMT can halve decode throughput on some CPUs (#718); set OMP_NUM_THREADS=<n> to override
== qwen36 Phase-2 engine | cache=256/layer bits=4 ctx=8192 pilot=0 wide=1 hot=0 smooth=0.30 conf=0.92 ==
[tok] loaded 248044 pieces (max id 248069) from /content/q36/tokenizer.json
[meta] loaded: q_heads=16 kv_heads=2 head_dim=256 q_head_dim=512 k_head_dim=256 v_head_dim=256 o_in=4096 rotary_dim=64 n_experts=256 topk=8 inter=512 shared_inter=512 n_group=1 topk_group=1 attn_output_gate=1 n_active=10
[meta] DeltaNet: vheads=32 kheads=16 kdim=128 vdim=128 convk=4 conv_dim=8192
[qwen36] embedding: rows read in place from the checkpoint, COLI_EMBED_MMAP=0 builds the f32 table
[dense-i8] 291 matrices quantized during load, 7.2 GB f32 freed
[qwen36] group-scaled experts: gs=64
resident weights loaded in 9.1s | RSS after load: 3.03 GB
[qwen36] expert format on disk: int4 packed (tier fmt=4)
[qwen36] expert kernel: planar int4 in RAM (expert_ffn.h), QWEN_EXPERT_KERNEL=0 restores int8 unpack
[CUDA] device 0: NVIDIA L4, 23.7 GB VRAM, sm_89
[qtier] dev 0: 21.8 GB free, allowance 5.0 GB
[place] auto: dev 0 holds 1828.5 MB of trunk (lmhead yes, 30 dnproj layers), 3.21 GB left for experts
[qtier] dev 0: budget 3.21 GB for experts (~1823 experts) after trunk
[qtier] CUDA VRAM expert tier active: 1 device(s), 1.80 MB/expert
[CUDA] mode: routed experts (qwen36 VRAM tier)
[gpu] MoE experts -> CUDA VRAM tier
[place] probe: one [8192 x 2048] int8 GEMV takes 0.112 ms on CUDA dev 0, 0.415 ms on the CPU -> trunk to VRAM
[lmh] lm_head [248320 x 2048] int8 resident on CUDA dev 0 (0.47 GB)
[dnp] 30 DeltaNet-Projektionen auf GPU (0.70 GB VRAM)
[dense] 190 trunk matrices on GPU (30 dnout, 40 attnproj, 120 shexp; 0.61 GB VRAM)
[chain] qwen36: a decode step's routed experts in the frame (the tier's residents by address, format 4, group 64; COLI_CUDA_CHAIN_MOE=0: the host's step)
[chain] qwen36: 40 layers on CUDA device 0 (10 attention), lm_head too; 4.4 MiB of parameters, 35.3 MiB of router, b|a and gate rows, 66 MB of DeltaNet state
[qwen36] int4 packed weights detected — kept int4, repacked planar for expert_ffn.h
[qtier] host pool: 8417 experts of the RAM cache page-locked for the device to copy a prompt chunk's experts from (COLI_CUDA_HOST_POOL=0 off)
[qtier] warmstart (parallel): all 10240 experts in RAM (planar int4, the only copy), 1823 in VRAM -- 7.6 s
[bench] CUDA ready in 17.85 s; text 6182 tokens; lengths 64,512; gen 32; each length from an empty state
[qwen36] KV cache: float16, 0.00 GiB at 96 positions (half of float32; Q36_KV=f32 keeps float32)
[chain] qwen36: 4096 rows per prompt chunk (286 KiB a row, 16.9 GB free on the device; COLI_CUDA_CHAIN_ROWS sets it)
[chain] qwen36: up to 4 experts a step promoted into the card by this conversation's routes, the coldest residents giving way, none while the last 64 steps hit 75 % or more (COLI_CUDA_CHAIN_PROMOTE=0 keeps the warm start's set)
[bench] peak RSS 19.37 GB
[qtier] host pool: 8417 experts (13.87 GB of the RAM cache) page-locked for the device's copies

===== v1-run4.csv (sha256 eef457e06c687bed16fa9d7aad8d9cda53820c43985dcb6526060ca4b0b94032)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,1.440,44.43,32,67.7,29.320,0,0.00,1e40892a3a7ff8df,5cb23346ab6be1d8,2.70028,0.469
512,512,1.655,309.29,32,34.3,31.465,0,0.00,649cee2cbca20368,7a8848450b863c49,2.96052,0.438
===== v1-run4.err (sha256 cec48b831a0bc1b71c6b0d738f3349ff7db6c268023fd626f35df3f3b5987d4f)
[tok] loaded 248044 pieces (max id 248069) from /content/q36/tokenizer.json
[OMP] qwen36: 6 physical-core threads instead of 12 logical CPUs; SMT can halve decode throughput on some CPUs (#718); set OMP_NUM_THREADS=<n> to override
== qwen36 Phase-2 engine | cache=256/layer bits=4 ctx=8192 pilot=0 wide=1 hot=0 smooth=0.30 conf=0.92 ==
[tok] loaded 248044 pieces (max id 248069) from /content/q36/tokenizer.json
[meta] loaded: q_heads=16 kv_heads=2 head_dim=256 q_head_dim=512 k_head_dim=256 v_head_dim=256 o_in=4096 rotary_dim=64 n_experts=256 topk=8 inter=512 shared_inter=512 n_group=1 topk_group=1 attn_output_gate=1 n_active=10
[meta] DeltaNet: vheads=32 kheads=16 kdim=128 vdim=128 convk=4 conv_dim=8192
[qwen36] embedding: rows read in place from the checkpoint, COLI_EMBED_MMAP=0 builds the f32 table
[dense-i8] 291 matrices quantized during load, 7.2 GB f32 freed
[qwen36] group-scaled experts: gs=64
resident weights loaded in 9.1s | RSS after load: 3.03 GB
[qwen36] expert format on disk: int4 packed (tier fmt=4)
[qwen36] expert kernel: planar int4 in RAM (expert_ffn.h), QWEN_EXPERT_KERNEL=0 restores int8 unpack
[CUDA] device 0: NVIDIA L4, 23.7 GB VRAM, sm_89
[qtier] dev 0: 21.8 GB free, allowance 5.0 GB
[place] auto: dev 0 holds 1828.5 MB of trunk (lmhead yes, 30 dnproj layers), 3.21 GB left for experts
[qtier] dev 0: budget 3.21 GB for experts (~1823 experts) after trunk
[qtier] CUDA VRAM expert tier active: 1 device(s), 1.80 MB/expert
[CUDA] mode: routed experts (qwen36 VRAM tier)
[gpu] MoE experts -> CUDA VRAM tier
[place] probe: one [8192 x 2048] int8 GEMV takes 0.112 ms on CUDA dev 0, 0.418 ms on the CPU -> trunk to VRAM
[lmh] lm_head [248320 x 2048] int8 resident on CUDA dev 0 (0.47 GB)
[dnp] 30 DeltaNet-Projektionen auf GPU (0.70 GB VRAM)
[dense] 190 trunk matrices on GPU (30 dnout, 40 attnproj, 120 shexp; 0.61 GB VRAM)
[chain] qwen36: a decode step's routed experts in the frame (the tier's residents by address, format 4, group 64; COLI_CUDA_CHAIN_MOE=0: the host's step)
[chain] qwen36: 40 layers on CUDA device 0 (10 attention), lm_head too; 4.4 MiB of parameters, 35.3 MiB of router, b|a and gate rows, 66 MB of DeltaNet state
[qwen36] int4 packed weights detected — kept int4, repacked planar for expert_ffn.h
[qtier] host pool: 8417 experts of the RAM cache page-locked for the device to copy a prompt chunk's experts from (COLI_CUDA_HOST_POOL=0 off)
[qtier] warmstart (parallel): all 10240 experts in RAM (planar int4, the only copy), 1823 in VRAM -- 7.6 s
[bench] CUDA ready in 17.81 s; text 6182 tokens; lengths 64,512; gen 32; each length from an empty state
[qwen36] KV cache: float16, 0.00 GiB at 96 positions (half of float32; Q36_KV=f32 keeps float32)
[chain] qwen36: 4096 rows per prompt chunk (286 KiB a row, 16.9 GB free on the device; COLI_CUDA_CHAIN_ROWS sets it)
[chain] qwen36: up to 4 experts a step promoted into the card by this conversation's routes, the coldest residents giving way, none while the last 64 steps hit 75 % or more (COLI_CUDA_CHAIN_PROMOTE=0 keeps the warm start's set)
[bench] peak RSS 19.37 GB
[qtier] host pool: 8417 experts (13.87 GB of the RAM cache) page-locked for the device's copies

===== v0-run4.csv (sha256 063658dc7939fcd022ff5eba58e62965b1bf1f02264578df980eda665783e8d0)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,1.433,44.67,32,68.4,29.184,0,0.00,1e40892a3a7ff8df,5cb23346ab6be1d8,2.70028,0.469
512,512,1.678,305.14,32,34.3,31.296,0,0.00,95505aeafaa41c33,7a8848450b863c49,2.96304,0.438
===== v0-run4.err (sha256 014956e59c3a796242e14c1950c33a4b65cea23675405c0bd3f506f2529f0803)
[tok] loaded 248044 pieces (max id 248069) from /content/q36/tokenizer.json
[OMP] qwen36: 6 physical-core threads instead of 12 logical CPUs; SMT can halve decode throughput on some CPUs (#718); set OMP_NUM_THREADS=<n> to override
== qwen36 Phase-2 engine | cache=256/layer bits=4 ctx=8192 pilot=0 wide=1 hot=0 smooth=0.30 conf=0.92 ==
[tok] loaded 248044 pieces (max id 248069) from /content/q36/tokenizer.json
[meta] loaded: q_heads=16 kv_heads=2 head_dim=256 q_head_dim=512 k_head_dim=256 v_head_dim=256 o_in=4096 rotary_dim=64 n_experts=256 topk=8 inter=512 shared_inter=512 n_group=1 topk_group=1 attn_output_gate=1 n_active=10
[meta] DeltaNet: vheads=32 kheads=16 kdim=128 vdim=128 convk=4 conv_dim=8192
[qwen36] embedding: rows read in place from the checkpoint, COLI_EMBED_MMAP=0 builds the f32 table
[dense-i8] 291 matrices quantized during load, 7.2 GB f32 freed
[qwen36] group-scaled experts: gs=64
resident weights loaded in 9.2s | RSS after load: 3.03 GB
[qwen36] expert format on disk: int4 packed (tier fmt=4)
[qwen36] expert kernel: planar int4 in RAM (expert_ffn.h), QWEN_EXPERT_KERNEL=0 restores int8 unpack
[CUDA] device 0: NVIDIA L4, 23.7 GB VRAM, sm_89
[qtier] dev 0: 21.8 GB free, allowance 5.0 GB
[place] auto: dev 0 holds 1828.5 MB of trunk (lmhead yes, 30 dnproj layers), 3.21 GB left for experts
[qtier] dev 0: budget 3.21 GB for experts (~1823 experts) after trunk
[qtier] CUDA VRAM expert tier active: 1 device(s), 1.80 MB/expert
[CUDA] mode: routed experts (qwen36 VRAM tier)
[gpu] MoE experts -> CUDA VRAM tier
[place] probe: one [8192 x 2048] int8 GEMV takes 0.112 ms on CUDA dev 0, 0.418 ms on the CPU -> trunk to VRAM
[lmh] lm_head [248320 x 2048] int8 resident on CUDA dev 0 (0.47 GB)
[dnp] 30 DeltaNet-Projektionen auf GPU (0.70 GB VRAM)
[dense] 190 trunk matrices on GPU (30 dnout, 40 attnproj, 120 shexp; 0.61 GB VRAM)
[chain] qwen36: a decode step's routed experts in the frame (the tier's residents by address, format 4, group 64; COLI_CUDA_CHAIN_MOE=0: the host's step)
[chain] qwen36: 40 layers on CUDA device 0 (10 attention), lm_head too; 4.4 MiB of parameters, 35.3 MiB of router, b|a and gate rows, 66 MB of DeltaNet state
[qwen36] int4 packed weights detected — kept int4, repacked planar for expert_ffn.h
[qtier] host pool: 8417 experts of the RAM cache page-locked for the device to copy a prompt chunk's experts from (COLI_CUDA_HOST_POOL=0 off)
[qtier] warmstart (parallel): all 10240 experts in RAM (planar int4, the only copy), 1823 in VRAM -- 7.6 s
[bench] CUDA ready in 17.91 s; text 6182 tokens; lengths 64,512; gen 32; each length from an empty state
[qwen36] KV cache: float16, 0.00 GiB at 96 positions (half of float32; Q36_KV=f32 keeps float32)
[chain] qwen36: 4096 rows per prompt chunk (286 KiB a row, 16.9 GB free on the device; COLI_CUDA_CHAIN_ROWS sets it)
[chain] qwen36: up to 4 experts a step promoted into the card by this conversation's routes, the coldest residents giving way, none while the last 64 steps hit 75 % or more (COLI_CUDA_CHAIN_PROMOTE=0 keeps the warm start's set)
[bench] peak RSS 19.37 GB
[qtier] host pool: 8417 experts (13.87 GB of the RAM cache) page-locked for the device's copies

===== summary (bench's report of the runs above)
# bench: misura-hyb

- bench 482e1711d589; engine qwen36; Linux (Intel(R) Xeon(R) CPU @ 2.20GHz, NVIDIA L4); 2026-10-25 14:44
- model q36; text promessi_sposi.txt (sha256 f53e0d80cb2d4492); lengths 64,512; gen 32; teacher-forced
- runs: a warm-up of v0 (does not count), then 4 per version, the order turning at every round (ABBA); before every run: wait until cool (at most 900 s), then 60 s of rest

## v0 dev: 4 runs (median, min–max)

| tokens | new | prompt tok/s | first step ms | decode tok/s | experts read | reads' s | nll | top1 | fingerprint |
|---|---|---|---|---|---|---|---|---|---|
| 64 | 64 | 44.64 (44.48–44.88) | 68 (68–68) | 29.194 (29.160–29.278) | 0 | 0.00 | 2.7003 | 0.469 | 1e40892a3a7ff8df |
| 512 | 512 | 306.04 (303.57–307.38) | 34 (34–34) | 31.450 (31.296–31.585) | 0 | 0.00 | 2.9605 | 0.438 | DIFFERENT between runs: ['649cee2cbca20368', '95505aeafaa41c33'] |

peak memory 19835 MiB, model ready in 17.86 s (medians)

decode by steps (step 1 gives the second new word; medians of 4 runs):

| tokens | steps | experts read | decode tok/s | reads' s |
|---|---|---|---|---|
| 64 | 2 to 32 | 0 | 29.193 | 0.000 |
| 512 | 2 to 32 | 0 | 31.451 | 0.000 |

## v1 pr: 4 runs (median, min–max)

| tokens | new | prompt tok/s | first step ms | decode tok/s | experts read | reads' s | nll | top1 | fingerprint |
|---|---|---|---|---|---|---|---|---|---|
| 64 | 64 | 44.54 (44.33–44.72) | 68 (68–68) | 29.334 (29.320–29.393) | 0 | 0.00 | 2.7003 | 0.469 | 1e40892a3a7ff8df |
| 512 | 512 | 307.21 (306.50–309.29) | 34 (34–34) | 31.530 (31.412–31.612) | 0 | 0.00 | 2.9618 | 0.438 | DIFFERENT between runs: ['649cee2cbca20368', '799d16315c29dd05', '95505aeafaa41c33', 'fc31d7e817064462'] |

peak memory 19835 MiB, model ready in 17.83 s (medians)

decode by steps (step 1 gives the second new word; medians of 4 runs):

| tokens | steps | experts read | decode tok/s | reads' s |
|---|---|---|---|---|
| 64 | 2 to 32 | 0 | 29.334 | 0.000 |
| 512 | 2 to 32 | 0 | 31.530 | 0.000 |

## v1 pr against v0 dev (medians; positive = faster)

| tokens | prompt | first step | decode | results |
|---|---|---|---|---|
| 64 | -0.2%, ranges overlap | +0.5%, ranges overlap | +0.5%, every run faster | same results; nll 2.7003 → 2.7003, top1 0.469 → 0.469 |
| 512 | +0.4%, ranges overlap | +1.6%, ranges overlap | +0.3%, ranges overlap | same words, logits differ; nll 2.9605 → 2.9618, top1 0.438 → 0.438 |

| tokens | steps | experts read | decode | reads' s |
|---|---|---|---|---|
| 64 | 2 to 32 | 0 → 0 | +0.5%, every run faster | 0.000 → 0.000 s |
| 512 | 2 to 32 | 0 → 0 | +0.3%, ranges overlap | 0.000 → 0.000 s |

warm-up (does not count): 64 tokens: prompt 44.67 tok/s, decode 29.318 tok/s; 512 tokens: prompt 309.61 tok/s, decode 31.537 tok/s
runs not cool: none
pages written to swap during the measurement: 0, 0 of them in the runs that count
