# qwen36 CUDA chain, accelerated mode (the MTP head's drafts): 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=auto: all 10240 experts in VRAM);
#   with its MTP head (mtp.* of Qwen/Qwen3.6-35B-A3B, convert_qwen36.py --mtp-only), Q36_MTP=1 Q36_MTP_DRAFTS=1.
# 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 12:52
model /content/q36mtp; 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 Q36_MTP=1 Q36_MTP_DRAFTS=1
   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 Q36_MTP=1 Q36_MTP_DRAFTS=1 /content/bin-dev/qwen36-bench --model /content/q36mtp --text /content/bench/promessi_sposi.txt --lengths 64,512 --gen 32 --teacher --steps /content/misura-acc/v0-runN.steps
v1 pr: /content/bin-pr/qwen36-bench (commit 56152a7e392cffab45b2d0e1dbbc4973584e48e6; 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 Q36_MTP=1 Q36_MTP_DRAFTS=1
   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 Q36_MTP=1 Q36_MTP_DRAFTS=1 /content/bin-pr/qwen36-bench --model /content/q36mtp --text /content/bench/promessi_sposi.txt --lengths 64,512 --gen 32 --teacher --steps /content/misura-acc/v1-runN.steps

===== warmup.csv (sha256 13eb4773d944fb22fcea78bfd3e10f7b4e3abeb67a0aa0494d0ad3c9384c8081)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,0.605,105.75,32,30.2,60.179,0,0.00,1837d0374f5820c3,5cb23346ab6be1d8,2.69849,0.469
512,512,0.313,1636.97,32,26.1,60.794,0,0.00,9a9da81aa38cc006,7a8848450b863c49,2.94540,0.469
===== warmup.err (sha256 dd820ccfaca45ce3f4fd08bae58d0452dc1277720d7737926c5445c844006724)
[tok] loaded 248044 pieces (max id 248069) from /content/q36mtp/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/q36mtp/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 9.4s | 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 20.8 GB
[place] auto: dev 0 holds 1828.5 MB of trunk (lmhead yes, 30 dnproj layers), 19.02 GB left for experts
[qtier] dev 0: budget 19.02 GB for experts (~10794 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, 0.431 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] warmstart (parallel): all 10240 experts in RAM (planar int4, the only copy), 10240 in VRAM -- 15.0 s
[mtp] Qwen3.6 checkpoint head: 256 experts
[bench] CUDA ready in 26.94 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, 3.1 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)
[mtp] the head runs on the CUDA chain's device, its experts in the frame (format 4)
[qwen36 mtp] 64 tokens: 1.39 tokens/forward (23 forwards per 32 tokens) | acceptance 40.9% (9/22 drafts) | gate T(1) 15.2 ms, rows 2:1.56, rho 0.50, acceptance by position 0.43 0.60 0.50 0.50 0.50
[mtp] head rows 22 (2.57 ms each), fed rows 86 (K/V only, 0.08 ms each), drafts 22
[qwen36 spec gate] 64 tokens: T(1) 15.2 ms, rows 2:1.56, rho 0.50, acceptance by position 0.50 0.50 0.50 0.50 0.50
[qwen36 lookup] 64 tokens: 1.39 tokens/forward (23 forwards per 32 tokens) | acceptance 0.0% (0/0 drafts in 22 verifies) | gate on, 1 declined, 0 probes
[qwen36 mtp] 512 tokens: 1.39 tokens/forward (23 forwards per 32 tokens) | acceptance 42.9% (9/21 drafts) | gate T(1) 15.2 ms, rows 2:1.57, rho 0.50, acceptance by position 0.47 0.60 0.50 0.50 0.50
[mtp] head rows 43 (2.58 ms each), fed rows 620 (K/V only, 0.04 ms each), drafts 43
[qwen36 spec gate] 512 tokens: T(1) 15.2 ms, rows 2:1.57, rho 0.50, acceptance by position 0.50 0.50 0.50 0.50 0.50
[qwen36 lookup] 512 tokens: 1.39 tokens/forward (23 forwards per 32 tokens) | acceptance 0.0% (0/0 drafts in 21 verifies) | gate on, 3 declined, 0 probes
[bench] peak RSS 20.09 GB

===== v0-run1.csv (sha256 0201f2c00d37451e024906e64152963d1b469bde69d89810833efca5ae76e655)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,0.607,105.46,32,30.1,60.829,0,0.00,1837d0374f5820c3,5cb23346ab6be1d8,2.69849,0.469
512,512,0.315,1627.14,32,24.0,62.332,0,0.00,9a9da81aa38cc006,7a8848450b863c49,2.94540,0.469
===== v0-run1.err (sha256 2362bf344e6ae4f3ccb8fe663ee1e4f8813b988c93834c088a1c416cb0350f5b)
[tok] loaded 248044 pieces (max id 248069) from /content/q36mtp/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/q36mtp/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 9.4s | 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 20.8 GB
[place] auto: dev 0 holds 1828.5 MB of trunk (lmhead yes, 30 dnproj layers), 19.02 GB left for experts
[qtier] dev 0: budget 19.02 GB for experts (~10794 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 0.114 ms on CUDA dev 0, 0.425 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] warmstart (parallel): all 10240 experts in RAM (planar int4, the only copy), 10240 in VRAM -- 15.0 s
[mtp] Qwen3.6 checkpoint head: 256 experts
[bench] CUDA ready in 26.99 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, 3.1 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)
[mtp] the head runs on the CUDA chain's device, its experts in the frame (format 4)
[qwen36 mtp] 64 tokens: 1.39 tokens/forward (23 forwards per 32 tokens) | acceptance 40.9% (9/22 drafts) | gate T(1) 15.3 ms, rows 2:1.54, rho 0.50, acceptance by position 0.43 0.60 0.50 0.50 0.50
[mtp] head rows 22 (2.58 ms each), fed rows 86 (K/V only, 0.08 ms each), drafts 22
[qwen36 spec gate] 64 tokens: T(1) 15.3 ms, rows 2:1.54, rho 0.50, acceptance by position 0.50 0.50 0.50 0.50 0.50
[qwen36 lookup] 64 tokens: 1.39 tokens/forward (23 forwards per 32 tokens) | acceptance 0.0% (0/0 drafts in 22 verifies) | gate on, 1 declined, 0 probes
[qwen36 mtp] 512 tokens: 1.39 tokens/forward (23 forwards per 32 tokens) | acceptance 42.9% (9/21 drafts) | gate T(1) 15.1 ms, rows 2:1.54, rho 0.50, acceptance by position 0.47 0.60 0.50 0.50 0.50
[mtp] head rows 43 (2.58 ms each), fed rows 620 (K/V only, 0.04 ms each), drafts 43
[qwen36 spec gate] 512 tokens: T(1) 15.1 ms, rows 2:1.54, rho 0.50, acceptance by position 0.50 0.50 0.50 0.50 0.50
[qwen36 lookup] 512 tokens: 1.39 tokens/forward (23 forwards per 32 tokens) | acceptance 0.0% (0/0 drafts in 21 verifies) | gate on, 3 declined, 0 probes
[bench] peak RSS 20.10 GB

===== v1-run1.csv (sha256 f4d7d1a8110d548112b272f18bf1679371086622bb391a31f98176bb5cd8e170)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,0.603,106.06,32,29.7,62.273,0,0.00,1837d0374f5820c3,5cb23346ab6be1d8,2.69849,0.469
512,512,0.309,1657.67,32,23.7,63.546,0,0.00,9a9da81aa38cc006,7a8848450b863c49,2.94540,0.469
===== v1-run1.err (sha256 7d757c1b0818ad631af7548d61e5e53815efa38e448cd510051e1c0ebd21722f)
[tok] loaded 248044 pieces (max id 248069) from /content/q36mtp/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/q36mtp/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 9.4s | 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 20.8 GB
[place] auto: dev 0 holds 1828.5 MB of trunk (lmhead yes, 30 dnproj layers), 19.02 GB left for experts
[qtier] dev 0: budget 19.02 GB for experts (~10794 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, 0.424 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] warmstart (parallel): all 10240 experts in RAM (planar int4, the only copy), 10240 in VRAM -- 15.1 s
[mtp] Qwen3.6 checkpoint head: 256 experts
[bench] CUDA ready in 27.73 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, 3.1 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)
[mtp] the head runs on the CUDA chain's device, its experts in the frame (format 4)
[qwen36 mtp] 64 tokens: 1.39 tokens/forward (23 forwards per 32 tokens) | acceptance 40.9% (9/22 drafts) | gate T(1) 14.8 ms, rows 2:1.55, rho 0.50, acceptance by position 0.43 0.60 0.50 0.50 0.50
[mtp] head rows 22 (2.53 ms each), fed rows 86 (K/V only, 0.07 ms each), drafts 22
[qwen36 spec gate] 64 tokens: T(1) 14.8 ms, rows 2:1.55, rho 0.50, acceptance by position 0.50 0.50 0.50 0.50 0.50
[qwen36 lookup] 64 tokens: 1.39 tokens/forward (23 forwards per 32 tokens) | acceptance 0.0% (0/0 drafts in 22 verifies) | gate on, 1 declined, 0 probes
[qwen36 mtp] 512 tokens: 1.39 tokens/forward (23 forwards per 32 tokens) | acceptance 42.9% (9/21 drafts) | gate T(1) 14.7 ms, rows 2:1.55, rho 0.50, acceptance by position 0.47 0.60 0.50 0.50 0.50
[mtp] head rows 43 (2.54 ms each), fed rows 620 (K/V only, 0.03 ms each), drafts 43
[qwen36 spec gate] 512 tokens: T(1) 14.7 ms, rows 2:1.55, rho 0.50, acceptance by position 0.50 0.50 0.50 0.50 0.50
[qwen36 lookup] 512 tokens: 1.39 tokens/forward (23 forwards per 32 tokens) | acceptance 0.0% (0/0 drafts in 21 verifies) | gate on, 3 declined, 0 probes
[bench] peak RSS 20.10 GB

===== v1-run2.csv (sha256 b00b6a5d7de4416e300de4fb5bee503029f585ffdb5d8957be7863b36f0afaff)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,0.609,105.02,32,29.3,62.345,0,0.00,1837d0374f5820c3,5cb23346ab6be1d8,2.69849,0.469
512,512,0.315,1625.66,32,23.7,63.625,0,0.00,9a9da81aa38cc006,7a8848450b863c49,2.94540,0.469
===== v1-run2.err (sha256 34ed47ee901cf0afef82c89dca6876fe0cfae196ca13312e7ad097c511ab6016)
[tok] loaded 248044 pieces (max id 248069) from /content/q36mtp/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/q36mtp/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 9.4s | 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 20.8 GB
[place] auto: dev 0 holds 1828.5 MB of trunk (lmhead yes, 30 dnproj layers), 19.02 GB left for experts
[qtier] dev 0: budget 19.02 GB for experts (~10794 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, 0.422 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] warmstart (parallel): all 10240 experts in RAM (planar int4, the only copy), 10240 in VRAM -- 14.9 s
[mtp] Qwen3.6 checkpoint head: 256 experts
[bench] CUDA ready in 27.57 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, 3.1 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)
[mtp] the head runs on the CUDA chain's device, its experts in the frame (format 4)
[qwen36 mtp] 64 tokens: 1.39 tokens/forward (23 forwards per 32 tokens) | acceptance 40.9% (9/22 drafts) | gate T(1) 14.8 ms, rows 2:1.54, rho 0.50, acceptance by position 0.43 0.60 0.50 0.50 0.50
[mtp] head rows 22 (2.53 ms each), fed rows 86 (K/V only, 0.07 ms each), drafts 22
[qwen36 spec gate] 64 tokens: T(1) 14.8 ms, rows 2:1.54, rho 0.50, acceptance by position 0.50 0.50 0.50 0.50 0.50
[qwen36 lookup] 64 tokens: 1.39 tokens/forward (23 forwards per 32 tokens) | acceptance 0.0% (0/0 drafts in 22 verifies) | gate on, 1 declined, 0 probes
[qwen36 mtp] 512 tokens: 1.39 tokens/forward (23 forwards per 32 tokens) | acceptance 42.9% (9/21 drafts) | gate T(1) 14.7 ms, rows 2:1.55, rho 0.50, acceptance by position 0.47 0.60 0.50 0.50 0.50
[mtp] head rows 43 (2.54 ms each), fed rows 620 (K/V only, 0.03 ms each), drafts 43
[qwen36 spec gate] 512 tokens: T(1) 14.7 ms, rows 2:1.55, rho 0.50, acceptance by position 0.50 0.50 0.50 0.50 0.50
[qwen36 lookup] 512 tokens: 1.39 tokens/forward (23 forwards per 32 tokens) | acceptance 0.0% (0/0 drafts in 21 verifies) | gate on, 3 declined, 0 probes
[bench] peak RSS 20.10 GB

===== v0-run2.csv (sha256 f1fd4f627cde8cf53db572bfb0e5e68d0fdfc13894c922a77944dfbaae23712b)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,0.607,105.43,32,30.3,60.827,0,0.00,1837d0374f5820c3,5cb23346ab6be1d8,2.69849,0.469
512,512,0.315,1627.17,32,24.4,62.434,0,0.00,9a9da81aa38cc006,7a8848450b863c49,2.94540,0.469
===== v0-run2.err (sha256 fd84fe8d795f7f4ab461527660fe839960aa5e7ae2b9b07e0234a2b923d82cbb)
[tok] loaded 248044 pieces (max id 248069) from /content/q36mtp/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/q36mtp/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 9.4s | 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 20.8 GB
[place] auto: dev 0 holds 1828.5 MB of trunk (lmhead yes, 30 dnproj layers), 19.02 GB left for experts
[qtier] dev 0: budget 19.02 GB for experts (~10794 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 0.114 ms on CUDA dev 0, 0.427 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] warmstart (parallel): all 10240 experts in RAM (planar int4, the only copy), 10240 in VRAM -- 15.0 s
[mtp] Qwen3.6 checkpoint head: 256 experts
[bench] CUDA ready in 26.96 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, 3.1 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)
[mtp] the head runs on the CUDA chain's device, its experts in the frame (format 4)
[qwen36 mtp] 64 tokens: 1.39 tokens/forward (23 forwards per 32 tokens) | acceptance 40.9% (9/22 drafts) | gate T(1) 15.3 ms, rows 2:1.54, rho 0.50, acceptance by position 0.43 0.60 0.50 0.50 0.50
[mtp] head rows 22 (2.57 ms each), fed rows 86 (K/V only, 0.08 ms each), drafts 22
[qwen36 spec gate] 64 tokens: T(1) 15.3 ms, rows 2:1.54, rho 0.50, acceptance by position 0.50 0.50 0.50 0.50 0.50
[qwen36 lookup] 64 tokens: 1.39 tokens/forward (23 forwards per 32 tokens) | acceptance 0.0% (0/0 drafts in 22 verifies) | gate on, 1 declined, 0 probes
[qwen36 mtp] 512 tokens: 1.39 tokens/forward (23 forwards per 32 tokens) | acceptance 42.9% (9/21 drafts) | gate T(1) 15.2 ms, rows 2:1.52, rho 0.50, acceptance by position 0.47 0.60 0.50 0.50 0.50
[mtp] head rows 43 (2.58 ms each), fed rows 620 (K/V only, 0.04 ms each), drafts 43
[qwen36 spec gate] 512 tokens: T(1) 15.2 ms, rows 2:1.52, rho 0.50, acceptance by position 0.50 0.50 0.50 0.50 0.50
[qwen36 lookup] 512 tokens: 1.39 tokens/forward (23 forwards per 32 tokens) | acceptance 0.0% (0/0 drafts in 21 verifies) | gate on, 3 declined, 0 probes
[bench] peak RSS 20.10 GB

===== v0-run3.csv (sha256 3ca3daab6b10c214e05cd5fba32828d0815455330c20fd3e843092958db1dd4d)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,0.607,105.42,32,30.2,60.788,0,0.00,1837d0374f5820c3,5cb23346ab6be1d8,2.69849,0.469
512,512,0.315,1627.58,32,24.4,62.438,0,0.00,9a9da81aa38cc006,7a8848450b863c49,2.94540,0.469
===== v0-run3.err (sha256 adf3b8ad2280ec5337683c375c41ba93dbb148b92b597b2fafc44493a13c4a8a)
[tok] loaded 248044 pieces (max id 248069) from /content/q36mtp/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/q36mtp/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 9.4s | 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 20.8 GB
[place] auto: dev 0 holds 1828.5 MB of trunk (lmhead yes, 30 dnproj layers), 19.02 GB left for experts
[qtier] dev 0: budget 19.02 GB for experts (~10794 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, 0.425 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] warmstart (parallel): all 10240 experts in RAM (planar int4, the only copy), 10240 in VRAM -- 15.1 s
[mtp] Qwen3.6 checkpoint head: 256 experts
[bench] CUDA ready in 27.00 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, 3.1 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)
[mtp] the head runs on the CUDA chain's device, its experts in the frame (format 4)
[qwen36 mtp] 64 tokens: 1.39 tokens/forward (23 forwards per 32 tokens) | acceptance 40.9% (9/22 drafts) | gate T(1) 15.3 ms, rows 2:1.54, rho 0.50, acceptance by position 0.43 0.60 0.50 0.50 0.50
[mtp] head rows 22 (2.57 ms each), fed rows 86 (K/V only, 0.08 ms each), drafts 22
[qwen36 spec gate] 64 tokens: T(1) 15.3 ms, rows 2:1.54, rho 0.50, acceptance by position 0.50 0.50 0.50 0.50 0.50
[qwen36 lookup] 64 tokens: 1.39 tokens/forward (23 forwards per 32 tokens) | acceptance 0.0% (0/0 drafts in 22 verifies) | gate on, 1 declined, 0 probes
[qwen36 mtp] 512 tokens: 1.39 tokens/forward (23 forwards per 32 tokens) | acceptance 42.9% (9/21 drafts) | gate T(1) 15.2 ms, rows 2:1.53, rho 0.50, acceptance by position 0.47 0.60 0.50 0.50 0.50
[mtp] head rows 43 (2.58 ms each), fed rows 620 (K/V only, 0.04 ms each), drafts 43
[qwen36 spec gate] 512 tokens: T(1) 15.2 ms, rows 2:1.53, rho 0.50, acceptance by position 0.50 0.50 0.50 0.50 0.50
[qwen36 lookup] 512 tokens: 1.39 tokens/forward (23 forwards per 32 tokens) | acceptance 0.0% (0/0 drafts in 21 verifies) | gate on, 3 declined, 0 probes
[bench] peak RSS 20.08 GB

===== v1-run3.csv (sha256 883a500f39fb9932f6e483394ed102992a7fd9d18222f834a3d299ef920124cf)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,0.612,104.50,32,29.5,62.093,0,0.00,1837d0374f5820c3,5cb23346ab6be1d8,2.69849,0.469
512,512,0.315,1627.61,32,23.9,63.518,0,0.00,9a9da81aa38cc006,7a8848450b863c49,2.94540,0.469
===== v1-run3.err (sha256 65d7277a5bef60c974de4bc3dbf7cdd85503610805a546768efec7f42f53aca4)
[tok] loaded 248044 pieces (max id 248069) from /content/q36mtp/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/q36mtp/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 9.4s | 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 20.8 GB
[place] auto: dev 0 holds 1828.5 MB of trunk (lmhead yes, 30 dnproj layers), 19.02 GB left for experts
[qtier] dev 0: budget 19.02 GB for experts (~10794 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] warmstart (parallel): all 10240 experts in RAM (planar int4, the only copy), 10240 in VRAM -- 15.0 s
[mtp] Qwen3.6 checkpoint head: 256 experts
[bench] CUDA ready in 27.57 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, 3.1 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)
[mtp] the head runs on the CUDA chain's device, its experts in the frame (format 4)
[qwen36 mtp] 64 tokens: 1.39 tokens/forward (23 forwards per 32 tokens) | acceptance 40.9% (9/22 drafts) | gate T(1) 14.8 ms, rows 2:1.55, rho 0.50, acceptance by position 0.43 0.60 0.50 0.50 0.50
[mtp] head rows 22 (2.54 ms each), fed rows 86 (K/V only, 0.07 ms each), drafts 22
[qwen36 spec gate] 64 tokens: T(1) 14.8 ms, rows 2:1.55, rho 0.50, acceptance by position 0.50 0.50 0.50 0.50 0.50
[qwen36 lookup] 64 tokens: 1.39 tokens/forward (23 forwards per 32 tokens) | acceptance 0.0% (0/0 drafts in 22 verifies) | gate on, 1 declined, 0 probes
[qwen36 mtp] 512 tokens: 1.39 tokens/forward (23 forwards per 32 tokens) | acceptance 42.9% (9/21 drafts) | gate T(1) 14.7 ms, rows 2:1.55, rho 0.50, acceptance by position 0.47 0.60 0.50 0.50 0.50
[mtp] head rows 43 (2.55 ms each), fed rows 620 (K/V only, 0.03 ms each), drafts 43
[qwen36 spec gate] 512 tokens: T(1) 14.7 ms, rows 2:1.55, rho 0.50, acceptance by position 0.50 0.50 0.50 0.50 0.50
[qwen36 lookup] 512 tokens: 1.39 tokens/forward (23 forwards per 32 tokens) | acceptance 0.0% (0/0 drafts in 21 verifies) | gate on, 3 declined, 0 probes
[bench] peak RSS 20.09 GB

===== v1-run4.csv (sha256 1ee4a5bb68645adb365e11a88532efbfe5888f1d960a42af713ff7f663aacc09)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,0.617,103.71,32,29.6,61.908,0,0.00,1837d0374f5820c3,5cb23346ab6be1d8,2.69849,0.469
512,512,0.317,1613.96,32,23.9,63.354,0,0.00,9a9da81aa38cc006,7a8848450b863c49,2.94540,0.469
===== v1-run4.err (sha256 3d802547cc6c45c522a99a409905512c3d50e77ddb5ca63d3408cab8ddf72e26)
[tok] loaded 248044 pieces (max id 248069) from /content/q36mtp/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/q36mtp/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 9.4s | 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 20.8 GB
[place] auto: dev 0 holds 1828.5 MB of trunk (lmhead yes, 30 dnproj layers), 19.02 GB left for experts
[qtier] dev 0: budget 19.02 GB for experts (~10794 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 0.114 ms on CUDA dev 0, 0.426 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] warmstart (parallel): all 10240 experts in RAM (planar int4, the only copy), 10240 in VRAM -- 15.2 s
[mtp] Qwen3.6 checkpoint head: 256 experts
[bench] CUDA ready in 27.78 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, 3.1 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)
[mtp] the head runs on the CUDA chain's device, its experts in the frame (format 4)
[qwen36 mtp] 64 tokens: 1.39 tokens/forward (23 forwards per 32 tokens) | acceptance 40.9% (9/22 drafts) | gate T(1) 14.9 ms, rows 2:1.56, rho 0.50, acceptance by position 0.43 0.60 0.50 0.50 0.50
[mtp] head rows 22 (2.53 ms each), fed rows 86 (K/V only, 0.07 ms each), drafts 22
[qwen36 spec gate] 64 tokens: T(1) 14.9 ms, rows 2:1.56, rho 0.50, acceptance by position 0.50 0.50 0.50 0.50 0.50
[qwen36 lookup] 64 tokens: 1.39 tokens/forward (23 forwards per 32 tokens) | acceptance 0.0% (0/0 drafts in 22 verifies) | gate on, 1 declined, 0 probes
[qwen36 mtp] 512 tokens: 1.39 tokens/forward (23 forwards per 32 tokens) | acceptance 42.9% (9/21 drafts) | gate T(1) 14.7 ms, rows 2:1.55, rho 0.50, acceptance by position 0.47 0.60 0.50 0.50 0.50
[mtp] head rows 43 (2.55 ms each), fed rows 620 (K/V only, 0.03 ms each), drafts 43
[qwen36 spec gate] 512 tokens: T(1) 14.7 ms, rows 2:1.55, rho 0.50, acceptance by position 0.50 0.50 0.50 0.50 0.50
[qwen36 lookup] 512 tokens: 1.39 tokens/forward (23 forwards per 32 tokens) | acceptance 0.0% (0/0 drafts in 21 verifies) | gate on, 3 declined, 0 probes
[bench] peak RSS 20.10 GB

===== v0-run4.csv (sha256 c6c6a1a2cadf3f4ee8ea5bc5ea292b0e41a6c6ab269fdf67484bcbf737c172fd)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,0.606,105.69,32,29.9,61.171,0,0.00,1837d0374f5820c3,5cb23346ab6be1d8,2.69849,0.469
512,512,0.316,1620.67,32,23.9,62.686,0,0.00,9a9da81aa38cc006,7a8848450b863c49,2.94540,0.469
===== v0-run4.err (sha256 538ada1f3f8b1f4ab8232fe862ca126ba8c83e28701501663589c489c43111ce)
[tok] loaded 248044 pieces (max id 248069) from /content/q36mtp/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/q36mtp/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 9.4s | 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 20.8 GB
[place] auto: dev 0 holds 1828.5 MB of trunk (lmhead yes, 30 dnproj layers), 19.02 GB left for experts
[qtier] dev 0: budget 19.02 GB for experts (~10794 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 0.114 ms on CUDA dev 0, 0.422 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] warmstart (parallel): all 10240 experts in RAM (planar int4, the only copy), 10240 in VRAM -- 15.1 s
[mtp] Qwen3.6 checkpoint head: 256 experts
[bench] CUDA ready in 27.04 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, 3.1 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)
[mtp] the head runs on the CUDA chain's device, its experts in the frame (format 4)
[qwen36 mtp] 64 tokens: 1.39 tokens/forward (23 forwards per 32 tokens) | acceptance 40.9% (9/22 drafts) | gate T(1) 15.3 ms, rows 2:1.53, rho 0.50, acceptance by position 0.43 0.60 0.50 0.50 0.50
[mtp] head rows 22 (2.56 ms each), fed rows 86 (K/V only, 0.08 ms each), drafts 22
[qwen36 spec gate] 64 tokens: T(1) 15.3 ms, rows 2:1.53, rho 0.50, acceptance by position 0.50 0.50 0.50 0.50 0.50
[qwen36 lookup] 64 tokens: 1.39 tokens/forward (23 forwards per 32 tokens) | acceptance 0.0% (0/0 drafts in 22 verifies) | gate on, 1 declined, 0 probes
[qwen36 mtp] 512 tokens: 1.39 tokens/forward (23 forwards per 32 tokens) | acceptance 42.9% (9/21 drafts) | gate T(1) 15.1 ms, rows 2:1.52, rho 0.50, acceptance by position 0.47 0.60 0.50 0.50 0.50
[mtp] head rows 43 (2.57 ms each), fed rows 620 (K/V only, 0.04 ms each), drafts 43
[qwen36 spec gate] 512 tokens: T(1) 15.1 ms, rows 2:1.52, rho 0.50, acceptance by position 0.50 0.50 0.50 0.50 0.50
[qwen36 lookup] 512 tokens: 1.39 tokens/forward (23 forwards per 32 tokens) | acceptance 0.0% (0/0 drafts in 21 verifies) | gate on, 3 declined, 0 probes
[bench] peak RSS 20.09 GB

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

- bench 482e1711d589; engine qwen36; Linux (Intel(R) Xeon(R) CPU @ 2.20GHz, NVIDIA L4); 2026-10-25 12:52
- model q36mtp; 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 | 105.44 (105.42–105.69) | 30 (30–30) | 60.828 (60.788–61.171) | 0 | 0.00 | 2.6985 | 0.469 | 1837d0374f5820c3 |
| 512 | 512 | 1627.16 (1620.67–1627.58) | 24 (24–24) | 62.436 (62.332–62.686) | 0 | 0.00 | 2.9454 | 0.469 | 9a9da81aa38cc006 |

peak memory 20577 MiB, model ready in 26.99 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 | 60.828 | 0.000 |
| 512 | 2 to 32 | 0 | 62.436 | 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 | 104.76 (103.71–106.06) | 30 (29–30) | 62.183 (61.908–62.345) | 0 | 0.00 | 2.6985 | 0.469 | 1837d0374f5820c3 |
| 512 | 512 | 1626.63 (1613.96–1657.67) | 24 (24–24) | 63.532 (63.354–63.625) | 0 | 0.00 | 2.9454 | 0.469 | 9a9da81aa38cc006 |

peak memory 20582 MiB, model ready in 27.65 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 | 62.183 | 0.000 |
| 512 | 2 to 32 | 0 | 63.532 | 0.000 |

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

| tokens | prompt | first step | decode | results |
|---|---|---|---|---|
| 64 | -0.6%, ranges overlap | +2.0%, every run faster | +2.2%, every run faster | same results; nll 2.6985 → 2.6985, top1 0.469 → 0.469 |
| 512 | -0.0%, ranges overlap | +1.7%, ranges overlap | +1.8%, every run faster | same results; nll 2.9454 → 2.9454, top1 0.469 → 0.469 |

| tokens | steps | experts read | decode | reads' s |
|---|---|---|---|---|
| 64 | 2 to 32 | 0 → 0 | +2.2%, every run faster | 0.000 → 0.000 s |
| 512 | 2 to 32 | 0 → 0 | +1.8%, every run faster | 0.000 → 0.000 s |

warm-up (does not count): 64 tokens: prompt 105.75 tok/s, decode 60.179 tok/s; 512 tokens: prompt 1636.97 tok/s, decode 60.794 tok/s
runs not cool: none
pages written to swap during the measurement: 0, 0 of them in the runs that count
