# qwen36 CUDA tier, the host pool following the card's residents: raw evidence (2026-10-26)
# Google Colab T4 high-RAM VM: Intel(R) Xeon(R) CPU @ 2.00GHz (4 cores / 8 threads, 38.5 MiB L3) + NVIDIA Tesla T4 (15360 MiB, sm_75, up to PCIe Gen3 x16), 51 GiB; Ubuntu 24.04.4 LTS, Linux 6.6.122+ x86_64 (Google Colab, T4 high-RAM), 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=auto: 6605 of the 10240 experts on the card
#   after 1.79 GB of trunk, the other 3635 page-locked in the host pool).
# v0 = origin/dev 77530b49; v1 = 77530b49 + this change, as measured: a local commit (77061547 in the setup
#   below, c/ git tree 461c4e7556ae20881d8ecda0c52bb77b9fdece02). This commit's c/ differs from it
#   only in the comment of Q36CC_FOLLOW_ROWS, rewritten after the runs to say which VM its numbers come from;
#   the documents and this record came after the runs too. git diff 77061547 HEAD -- c:
#   | diff --git a/c/qwen36_cuda_chain.h b/c/qwen36_cuda_chain.h
#   | index 80536b77..9955a836 100644
#   | --- a/c/qwen36_cuda_chain.h
#   | +++ b/c/qwen36_cuda_chain.h
#   | @@ -1156,10 +1156,13 @@ static void q36cc_kv_down(Q36CChain *ch, Model *m, int i, int n, int pb) {
#   |  /* The host pool follows the residents (qt_chain_host_follow) before a prompt chunk of
#   |   * Q36CC_FOLLOW_ROWS rows or more: the experts that left the card join it, those that came in
#   |   * leave it, or a layer holding one that left goes to the host's step whole. A smaller chunk
#   | - * leaves it behind for the next large one. On a T4 a follow costs 0.3 ms an expert moved
#   | - * (317 + 317 after a 489-token answer: 193 ms), and the layers it gives back to the copies
#   | - * saved 50 ms of a 36-row chunk, 1.1 s of a 348-row one: it pays from about 80 rows, and at
#   | - * 128 for twice those moves. */
#   | + * leaves it behind for the next large one. On the Colab T4 high-RAM VM of the measures
#   | + * (docs/experiments/qwen36-cuda-host-pool-follow-2026-10-26) a follow costs about 0.3 ms an
#   | + * expert moved. On another VM of the same shape an earlier version, which followed before
#   | + * every chunk of more than 32 rows, moved 317 + 317 experts in 193 ms before a 36-row chunk
#   | + * and saved about 50 ms of it, while a 348-row chunk after 50 + 50 moves saved 1.1 s: through
#   | + * those two points the follow pays from about 80 rows, and at 128 the copies save about
#   | + * twice the 193 ms (threshold.txt there). */
#   |  #define Q36CC_FOLLOW_ROWS 128
#   |  /* Promotions (qt_chain_promote): after a decode-sized step, up to this many of its misses move
#   |   * into the buffers of the coldest residents by this conversation's routes. The chain copies
# 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_75 && 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: the 512-token
#   prompt comes after the 64-token one and its 32 steps, in which the CUDA chain's promotions run.
# 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;
#   the steps files (a line a step) by their sha256, not copied;
#   the summary at the end.

===== setup
bench 482e1711d589; engine qwen36; Linux (Intel(R) Xeon(R) CPU @ 2.00GHz, Tesla T4); 2026-10-26 11:42
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/t1/bin-dev/qwen36-bench (commit 77530b49ddd2f4e6ad1b3bb52500491e0c8ed9bb; 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
   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 /content/t1/bin-dev/qwen36-bench --model /content/q36 --text /content/bench/promessi_sposi.txt --lengths 64,512 --gen 32 --teacher --steps /content/t1/misura/v0-runN.steps
v1 pr: /content/t1/bin-pr/qwen36-bench (commit 77061547bd4dca066e9f59094006ae185ad6ffda; 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
   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 /content/t1/bin-pr/qwen36-bench --model /content/q36 --text /content/bench/promessi_sposi.txt --lengths 64,512 --gen 32 --teacher --steps /content/t1/misura/v1-runN.steps

===== warmup.csv (sha256 7ab0ffb4850b2a22043b9c43a88063fe7548bbdf6b956e27c55b9fb2589b149a)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,1.036,61.79,32,65.8,37.583,0,0.00,ffb816333a929871,5cb23346ab6be1d8,2.69296,0.469
512,512,3.295,155.40,32,21.6,45.152,0,0.00,fb5a702b48be64d9,7a8848450b863c49,2.95785,0.438
===== warmup.err (sha256 bcd45f41339b3f8ee25773d259a1857269da64d09d73a6430ddbd1670f4a1394)
[tok] loaded 248044 pieces (max id 248069) from /content/q36/tokenizer.json
[OMP] qwen36: 4 physical-core threads instead of 8 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 13.2s | 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: Tesla T4, 15.6 GB VRAM, sm_75
[qtier] dev 0: 14.5 GB free, allowance 13.4 GB
[place] auto: dev 0 holds 1828.5 MB of trunk (lmhead yes, 30 dnproj layers), 11.64 GB left for experts
[qtier] dev 0: budget 11.64 GB for experts (~6605 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 0.388 ms on CUDA dev 0, 0.612 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: 3635 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), 6605 in VRAM -- 58.2 s
[bench] CUDA ready in 72.76 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: 2048 rows per prompt chunk (286 KiB a row, 1.7 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 19.48 GB
[qtier] host pool: 3635 experts (5.99 GB of the RAM cache) page-locked for the device's copies
----- warmup.steps: sha256 77c8ba814b9687a8674d226da0adcba8ed9707390d5b5ce22acf85deaca0639c

===== v0-run1.csv (sha256 247aa603a739bfa1e11f1729feb0ac68c0d583668a691d1f48372deab7691bab)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,0.867,73.78,32,67.1,39.606,0,0.00,ffb816333a929871,5cb23346ab6be1d8,2.69296,0.469
512,512,2.857,179.19,32,21.6,46.934,0,0.00,a06145f14b7e678d,7a8848450b863c49,2.95784,0.438
===== v0-run1.err (sha256 0f67151f3a75da6ffdec66bd7362708d8cc11f8c20373e41dda5532a342fc853)
[tok] loaded 248044 pieces (max id 248069) from /content/q36/tokenizer.json
[OMP] qwen36: 4 physical-core threads instead of 8 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 13.5s | 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: Tesla T4, 15.6 GB VRAM, sm_75
[qtier] dev 0: 14.5 GB free, allowance 13.4 GB
[place] auto: dev 0 holds 1828.5 MB of trunk (lmhead yes, 30 dnproj layers), 11.64 GB left for experts
[qtier] dev 0: budget 11.64 GB for experts (~6605 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 0.403 ms on CUDA dev 0, 0.617 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
[qwen36] int4 packed weights detected — kept int4, repacked planar for expert_ffn.h
[qtier] host pool: 3635 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), 6605 in VRAM -- 9.3 s
[bench] CUDA ready in 23.98 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: 2048 rows per prompt chunk (286 KiB a row, 1.7 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 19.48 GB
[qtier] host pool: 3635 experts (5.99 GB of the RAM cache) page-locked for the device's copies
----- v0-run1.steps: sha256 d4a01c52bb6127251d125a5395b56f1fec8a54c346c36b0ffa3f358cec793b97

===== v1-run1.csv (sha256 fb9652651176ec304f2198131af486694ace22491f14445f7c901a9571143f38)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,0.822,77.90,32,64.0,40.498,0,0.00,ffb816333a929871,5cb23346ab6be1d8,2.69296,0.469
512,512,1.158,441.96,32,22.3,47.564,0,0.00,4da5addfb0a8021f,7a8848450b863c49,2.93281,0.469
===== v1-run1.err (sha256 30f5319e7c819b4f49818d368cb6327fd96414c52a35cb571b5f76edf62f72bb)
[tok] loaded 248044 pieces (max id 248069) from /content/q36/tokenizer.json
[OMP] qwen36: 4 physical-core threads instead of 8 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 9.7s | 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: Tesla T4, 15.6 GB VRAM, sm_75
[qtier] dev 0: 14.5 GB free, allowance 13.4 GB
[place] auto: dev 0 holds 1828.5 MB of trunk (lmhead yes, 30 dnproj layers), 11.64 GB left for experts
[qtier] dev 0: budget 11.64 GB for experts (~6605 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 0.445 ms on CUDA dev 0, 0.623 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: 3635 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), 6605 in VRAM -- 9.2 s
[bench] CUDA ready in 20.06 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: 2048 rows per prompt chunk (286 KiB a row, 1.7 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 19.47 GB
[qtier] host pool: 3635 experts (5.99 GB of the RAM cache) page-locked for the device's copies
----- v1-run1.steps: sha256 b54b439a268dcab9960bcd03cc46e0a717ea3355d0d4feaffa60c14055f449ea

===== v1-run2.csv (sha256 e03f6880eff9d2d15b33936936f743db39f6db0525e589ad9cd9c11cf69c6d4d)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,0.840,76.20,32,62.4,40.008,0,0.00,ffb816333a929871,5cb23346ab6be1d8,2.69296,0.469
512,512,1.173,436.44,32,22.2,47.543,0,0.00,7d641cca01aa1fc1,7a8848450b863c49,2.93284,0.469
===== v1-run2.err (sha256 a0229b7f6d40ae1a46738dcf0421868d2fb165e01e4061534f8582ddd3ab9116)
[tok] loaded 248044 pieces (max id 248069) from /content/q36/tokenizer.json
[OMP] qwen36: 4 physical-core threads instead of 8 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 9.6s | 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: Tesla T4, 15.6 GB VRAM, sm_75
[qtier] dev 0: 14.5 GB free, allowance 13.4 GB
[place] auto: dev 0 holds 1828.5 MB of trunk (lmhead yes, 30 dnproj layers), 11.64 GB left for experts
[qtier] dev 0: budget 11.64 GB for experts (~6605 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 0.518 ms on CUDA dev 0, 0.550 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: 3635 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), 6605 in VRAM -- 9.0 s
[bench] CUDA ready in 19.79 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: 2048 rows per prompt chunk (286 KiB a row, 1.7 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 19.47 GB
[qtier] host pool: 3635 experts (5.99 GB of the RAM cache) page-locked for the device's copies
----- v1-run2.steps: sha256 5380be8a9556ef669ea91bd42cdcc060798c464ab817b9d16cfe00354bccb6f0

===== v0-run2.csv (sha256 f420218f0be27751833529a052ca93aa9ce0f6df74026e508e76e6d1b0568385)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,0.842,76.02,32,65.1,39.920,0,0.00,ffb816333a929871,5cb23346ab6be1d8,2.69296,0.469
512,512,2.830,180.94,32,22.4,46.811,0,0.00,a06145f14b7e678d,7a8848450b863c49,2.95784,0.438
===== v0-run2.err (sha256 9c7b9fd10cb7907c48408e43f0732e322e5fbc531acc86e38c394ba4e563b081)
[tok] loaded 248044 pieces (max id 248069) from /content/q36/tokenizer.json
[OMP] qwen36: 4 physical-core threads instead of 8 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 9.8s | 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: Tesla T4, 15.6 GB VRAM, sm_75
[qtier] dev 0: 14.5 GB free, allowance 13.4 GB
[place] auto: dev 0 holds 1828.5 MB of trunk (lmhead yes, 30 dnproj layers), 11.64 GB left for experts
[qtier] dev 0: budget 11.64 GB for experts (~6605 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 0.538 ms on CUDA dev 0, 0.564 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: 3635 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), 6605 in VRAM -- 9.1 s
[bench] CUDA ready in 20.10 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: 2048 rows per prompt chunk (286 KiB a row, 1.7 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 19.48 GB
[qtier] host pool: 3635 experts (5.99 GB of the RAM cache) page-locked for the device's copies
----- v0-run2.steps: sha256 e366aeb6797cec711ffcb7945038dbb65bf4b598c2f14621a12bd6fd1f533543

===== v0-run3.csv (sha256 96446c4dc3b11b479f8622bc9a91b1584fc8250e533655ab04808e04d4eae816)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,0.810,79.01,32,63.3,40.086,0,0.00,ffb816333a929871,5cb23346ab6be1d8,2.69296,0.469
512,512,2.841,180.23,32,21.1,47.729,0,0.00,fb5a702b48be64d9,7a8848450b863c49,2.95785,0.438
===== v0-run3.err (sha256 b47ca8ea69ae086a89ff4df4de652b8dc0f5af76d02681904bc4ce120f603dd9)
[tok] loaded 248044 pieces (max id 248069) from /content/q36/tokenizer.json
[OMP] qwen36: 4 physical-core threads instead of 8 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 9.6s | 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: Tesla T4, 15.6 GB VRAM, sm_75
[qtier] dev 0: 14.5 GB free, allowance 13.4 GB
[place] auto: dev 0 holds 1828.5 MB of trunk (lmhead yes, 30 dnproj layers), 11.64 GB left for experts
[qtier] dev 0: budget 11.64 GB for experts (~6605 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 0.437 ms on CUDA dev 0, 0.628 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: 3635 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), 6605 in VRAM -- 9.1 s
[bench] CUDA ready in 19.83 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: 2048 rows per prompt chunk (286 KiB a row, 1.7 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 19.48 GB
[qtier] host pool: 3635 experts (5.99 GB of the RAM cache) page-locked for the device's copies
----- v0-run3.steps: sha256 b9f1dea57a767a3207ff13968891e5610fa917f8801f83216eb64621ea160872

===== v1-run3.csv (sha256 bdc66e87c9f5c0a6092e96070f7add1c0e0bf2b3eb8bedd72f208ac1ddde2e60)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,0.822,77.81,32,63.2,39.992,0,0.00,ffb816333a929871,5cb23346ab6be1d8,2.69296,0.469
512,512,1.164,440.05,32,21.8,47.277,0,0.00,7d641cca01aa1fc1,7a8848450b863c49,2.93284,0.469
===== v1-run3.err (sha256 1cc4d455c32cacd5a150c28a2ebb6d41c887ad132b83c4c8015b0d56c745edf8)
[tok] loaded 248044 pieces (max id 248069) from /content/q36/tokenizer.json
[OMP] qwen36: 4 physical-core threads instead of 8 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 9.7s | 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: Tesla T4, 15.6 GB VRAM, sm_75
[qtier] dev 0: 14.5 GB free, allowance 13.4 GB
[place] auto: dev 0 holds 1828.5 MB of trunk (lmhead yes, 30 dnproj layers), 11.64 GB left for experts
[qtier] dev 0: budget 11.64 GB for experts (~6605 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 0.390 ms on CUDA dev 0, 0.621 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: 3635 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), 6605 in VRAM -- 9.1 s
[bench] CUDA ready in 19.98 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: 2048 rows per prompt chunk (286 KiB a row, 1.7 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 19.47 GB
[qtier] host pool: 3635 experts (5.99 GB of the RAM cache) page-locked for the device's copies
----- v1-run3.steps: sha256 5869840e49db109702ae5844244bbb95cf49ceb6ada3dee267806659d2adbe69

===== v1-run4.csv (sha256 39376f08a919ffebe7e107e67f543e3f9e5240a9d55cab3b2a97377e58fbd0d8)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,0.841,76.09,32,64.1,40.144,0,0.00,ffb816333a929871,5cb23346ab6be1d8,2.69296,0.469
512,512,1.163,440.17,32,21.9,47.431,0,0.00,4da5addfb0a8021f,7a8848450b863c49,2.93281,0.469
===== v1-run4.err (sha256 03f45911fbd18a5daff208cd73b728af4f304d46960c5151e3514fd17ee29e81)
[tok] loaded 248044 pieces (max id 248069) from /content/q36/tokenizer.json
[OMP] qwen36: 4 physical-core threads instead of 8 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 9.8s | 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: Tesla T4, 15.6 GB VRAM, sm_75
[qtier] dev 0: 14.5 GB free, allowance 13.4 GB
[place] auto: dev 0 holds 1828.5 MB of trunk (lmhead yes, 30 dnproj layers), 11.64 GB left for experts
[qtier] dev 0: budget 11.64 GB for experts (~6605 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 0.408 ms on CUDA dev 0, 0.601 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: 3635 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), 6605 in VRAM -- 9.2 s
[bench] CUDA ready in 20.18 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: 2048 rows per prompt chunk (286 KiB a row, 1.7 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 19.47 GB
[qtier] host pool: 3635 experts (5.99 GB of the RAM cache) page-locked for the device's copies
----- v1-run4.steps: sha256 11b133800dc655b6ab860b712596f0b6725a091bb9e166c73476a0914812a717

===== v0-run4.csv (sha256 5f5f118471fa58706710c9cbb32d3e3d855498469f815294a35b5a1d66119daf)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,0.847,75.59,32,68.0,38.850,0,0.00,ffb816333a929871,5cb23346ab6be1d8,2.69296,0.469
512,512,2.843,180.12,32,21.6,47.058,0,0.00,fb5a702b48be64d9,7a8848450b863c49,2.95785,0.438
===== v0-run4.err (sha256 5f81aac4855d3d598a6b00d2f771b9ae569a997e850a45c07089030cb164fac3)
[tok] loaded 248044 pieces (max id 248069) from /content/q36/tokenizer.json
[OMP] qwen36: 4 physical-core threads instead of 8 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 9.7s | 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: Tesla T4, 15.6 GB VRAM, sm_75
[qtier] dev 0: 14.5 GB free, allowance 13.4 GB
[place] auto: dev 0 holds 1828.5 MB of trunk (lmhead yes, 30 dnproj layers), 11.64 GB left for experts
[qtier] dev 0: budget 11.64 GB for experts (~6605 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 0.472 ms on CUDA dev 0, 0.585 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: 3635 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), 6605 in VRAM -- 9.1 s
[bench] CUDA ready in 20.07 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: 2048 rows per prompt chunk (286 KiB a row, 1.7 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 19.48 GB
[qtier] host pool: 3635 experts (5.99 GB of the RAM cache) page-locked for the device's copies
----- v0-run4.steps: sha256 8215205884d4f69246270d10fc8cbd126c28d97a568153692205bf0c978630a0

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

- bench 482e1711d589; engine qwen36; Linux (Intel(R) Xeon(R) CPU @ 2.00GHz, Tesla T4); 2026-10-26 11:42
- 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 | 75.81 (73.78–79.01) | 66 (63–68) | 39.763 (38.850–40.086) | 0 | 0.00 | 2.6930 | 0.469 | ffb816333a929871 |
| 512 | 512 | 180.18 (179.19–180.94) | 22 (21–22) | 46.996 (46.811–47.729) | 0 | 0.00 | 2.9578 | 0.438 | DIFFERENT between runs: ['a06145f14b7e678d', 'fb5a702b48be64d9'] |

peak memory 19948 MiB, model ready in 20.09 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 | 39.763 | 0.000 |
| 512 | 2 to 32 | 0 | 46.996 | 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 | 77.00 (76.09–77.90) | 64 (62–64) | 40.076 (39.992–40.498) | 0 | 0.00 | 2.6930 | 0.469 | ffb816333a929871 |
| 512 | 512 | 440.11 (436.44–441.96) | 22 (22–22) | 47.487 (47.277–47.564) | 0 | 0.00 | 2.9328 | 0.469 | DIFFERENT between runs: ['4da5addfb0a8021f', '7d641cca01aa1fc1'] |

peak memory 19937 MiB, model ready in 20.02 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 | 40.076 | 0.000 |
| 512 | 2 to 32 | 0 | 47.487 | 0.000 |

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

| tokens | prompt | first step | decode | results |
|---|---|---|---|---|
| 64 | +1.6%, ranges overlap | +3.9%, ranges overlap | +0.8%, ranges overlap | same results; nll 2.6930 → 2.6930, top1 0.469 → 0.469 |
| 512 | +144.3%, every run faster | -2.0%, ranges overlap | +1.0%, ranges overlap | same words, logits differ; nll 2.9578 → 2.9328, top1 0.438 → 0.469 |

| tokens | steps | experts read | decode | reads' s |
|---|---|---|---|---|
| 64 | 2 to 32 | 0 → 0 | +0.8%, ranges overlap | 0.000 → 0.000 s |
| 512 | 2 to 32 | 0 → 0 | +1.0%, ranges overlap | 0.000 → 0.000 s |

warm-up (does not count): 64 tokens: prompt 61.79 tok/s, decode 37.583 tok/s; 512 tokens: prompt 155.40 tok/s, decode 45.152 tok/s
runs not cool: none
pages written to swap during the measurement: 0, 0 of them in the runs that count
