# qwen36 CUDA chain, the MTP head's adaptive draft list, prose, one draft a step (the default): raw evidence (2026-10-26, UTC)
# 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 = 16d9f670: origin/dev 7bbea341 + the CUDA chain's small copies + the MTP draft picked on the
#   device, the changes this one builds on; v1 = 16d9f670 + this change (the measured build; after it only comments and
#   c/tests/ changed). 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); 2026-10-26 00:04
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 c4: /content/bin-c4/qwen36-bench (commit 16d9f6703aeb716135d5df22092e742be9cc9b5b; 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-c4/qwen36-bench --model /content/q36mtp --text /content/bench/promessi_sposi.txt --lengths 64,512 --gen 32 --teacher --steps /content/c6-prosa-1/v0-runN.steps
v1 c6: /content/bin-c6/qwen36-bench (commit 195e7e5ec793ae8e34c033a1d4513276b095d7da; 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-c6/qwen36-bench --model /content/q36mtp --text /content/bench/promessi_sposi.txt --lengths 64,512 --gen 32 --teacher --steps /content/c6-prosa-1/v1-runN.steps

===== warmup.csv (sha256 d4780b09bbdbb22914eafc5ce999debac85dc1187183c31a780035b43950ba5c)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,0.711,90.07,32,26.3,62.968,0,0.00,1837d0374f5820c3,5cb23346ab6be1d8,2.69849,0.469
512,512,0.659,776.47,32,22.3,67.765,0,0.00,9a9da81aa38cc006,7a8848450b863c49,2.94540,0.469
===== warmup.err (sha256 64dc5d2c608c462a00533e32c4f79522bda83d10f2736e2c122287901a6f0e70)
[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 9.1s | 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 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, 0.413 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 -- 16.0 s
[mtp] Qwen3.6 checkpoint head: 256 experts
[bench] CUDA ready in 27.47 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) 16.9 ms, rows 2:1.36, rho 0.50, acceptance by position 0.43 0.60 0.50 0.50 0.50
[mtp] head rows 22 (3.74 ms each), fed rows 86 (K/V only, 0.06 ms each), drafts 22
[qwen36 spec gate] 64 tokens: T(1) 16.9 ms, rows 2:1.36, 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) 16.2 ms, rows 2:1.30, rho 0.50, acceptance by position 0.47 0.60 0.50 0.50 0.50
[mtp] head rows 43 (3.31 ms each), fed rows 620 (K/V only, 0.03 ms each), drafts 43
[qwen36 spec gate] 512 tokens: T(1) 16.2 ms, rows 2:1.30, 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 01362882646c9254dfa92513aef24c88a76f10955ea178142eb4e8b31541f982)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,0.612,104.52,32,26.1,71.019,0,0.00,1837d0374f5820c3,5cb23346ab6be1d8,2.69849,0.469
512,512,0.310,1651.33,32,20.5,72.627,0,0.00,9a9da81aa38cc006,7a8848450b863c49,2.94540,0.469
===== v0-run1.err (sha256 5f4489993c89f6954af819a3bdbc25c601710d512a38cf1ba11e7278df688d02)
[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 9.1s | 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 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, 0.417 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
[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 -- 14.7 s
[mtp] Qwen3.6 checkpoint head: 256 experts
[bench] CUDA ready in 26.18 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: 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) 14.7 ms, rows 2:1.37, rho 0.50, acceptance by position 0.43 0.60 0.50 0.50 0.50
[mtp] head rows 22 (2.44 ms each), fed rows 86 (K/V only, 0.05 ms each), drafts 22
[qwen36 spec gate] 64 tokens: T(1) 14.7 ms, rows 2:1.37, 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) 14.5 ms, rows 2:1.36, rho 0.50, acceptance by position 0.47 0.60 0.50 0.50 0.50
[mtp] head rows 43 (2.43 ms each), fed rows 620 (K/V only, 0.03 ms each), drafts 43
[qwen36 spec gate] 512 tokens: T(1) 14.5 ms, rows 2:1.36, 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 20.01 GB

===== v1-run1.csv (sha256 f4c42cbe30207135e8f2950d3c3aabbd206406b4e4f73aea3d17492e805cb6ee)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,0.616,103.96,32,24.2,71.233,0,0.00,1837d0374f5820c3,5cb23346ab6be1d8,2.69849,0.469
512,512,0.314,1632.76,32,19.2,78.641,0,0.00,9a9da81aa38cc006,7a8848450b863c49,2.94540,0.469
===== v1-run1.err (sha256 05c201a6868b8157dca614ac25c614f379d1ab405ac9681627c8377362c8923f)
[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 9.1s | 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 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, 0.414 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 -- 14.4 s
[mtp] draft vocabulary: top 16000 + 3/4 rule + conversation, no counts file at /content/q36mtp/.coli_vocab yet
[mtp] Qwen3.6 checkpoint head: 256 experts
[bench] CUDA ready in 25.93 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), drafts score an adaptive vocabulary
[qwen36 mtp] 64 tokens: 1.28 tokens/forward (25 forwards per 32 tokens) | acceptance 29.2% (7/24 drafts) | gate T(1) 14.7 ms, rows 2:1.25, rho 0.50, acceptance by position 0.31 0.60 0.50 0.50 0.50
[mtp] head rows 24 (0.55 ms each), fed rows 88 (K/V only, 0.06 ms each), drafts 24
[mtp] draft list 16004 tokens on average, 68.8% of picks inside, 10 ids appended, 0 rebuilds
[qwen36 spec gate] 64 tokens: T(1) 14.7 ms, rows 2:1.25, rho 0.50, acceptance by position 0.50 0.50 0.50 0.50 0.50
[qwen36 lookup] 64 tokens: 1.28 tokens/forward (25 forwards per 32 tokens) | acceptance 0.0% (0/0 drafts in 24 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) 14.6 ms, rows 2:1.25, rho 0.50, acceptance by position 0.47 0.60 0.50 0.50 0.50
[mtp] head rows 45 (0.54 ms each), fed rows 622 (K/V only, 0.03 ms each), drafts 45
[mtp] draft list 16003 tokens on average, 84.4% of picks inside, 5 ids appended, 0 rebuilds
[qwen36 spec gate] 512 tokens: T(1) 14.6 ms, rows 2:1.25, 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 65fd6b4c1a017d9421de3a6734e488bf6142c5fbc08ca3956e632514d5b0e8dd)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,0.613,104.37,32,24.2,71.289,0,0.00,1837d0374f5820c3,5cb23346ab6be1d8,2.69849,0.469
512,512,0.313,1636.38,32,19.2,78.683,0,0.00,9a9da81aa38cc006,7a8848450b863c49,2.94540,0.469
===== v1-run2.err (sha256 84cc6e9b41c11b1c821b3740d46db6b723b48dae51c216ad7783e37b9a029dcc)
[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 9.2s | 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 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, 0.414 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 -- 14.4 s
[mtp] draft vocabulary: top 16000 + 3/4 rule + conversation, no counts file at /content/q36mtp/.coli_vocab yet
[mtp] Qwen3.6 checkpoint head: 256 experts
[bench] CUDA ready in 25.97 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), drafts score an adaptive vocabulary
[qwen36 mtp] 64 tokens: 1.28 tokens/forward (25 forwards per 32 tokens) | acceptance 29.2% (7/24 drafts) | gate T(1) 14.7 ms, rows 2:1.25, rho 0.50, acceptance by position 0.31 0.60 0.50 0.50 0.50
[mtp] head rows 24 (0.54 ms each), fed rows 88 (K/V only, 0.05 ms each), drafts 24
[mtp] draft list 16004 tokens on average, 68.8% of picks inside, 10 ids appended, 0 rebuilds
[qwen36 spec gate] 64 tokens: T(1) 14.7 ms, rows 2:1.25, rho 0.50, acceptance by position 0.50 0.50 0.50 0.50 0.50
[qwen36 lookup] 64 tokens: 1.28 tokens/forward (25 forwards per 32 tokens) | acceptance 0.0% (0/0 drafts in 24 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) 14.6 ms, rows 2:1.25, rho 0.50, acceptance by position 0.47 0.60 0.50 0.50 0.50
[mtp] head rows 45 (0.54 ms each), fed rows 622 (K/V only, 0.03 ms each), drafts 45
[mtp] draft list 16003 tokens on average, 84.4% of picks inside, 5 ids appended, 0 rebuilds
[qwen36 spec gate] 512 tokens: T(1) 14.6 ms, rows 2:1.25, 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 20.11 GB

===== v0-run2.csv (sha256 a8dd4f4398604bc061976970552608ad63e046282018154cd83da0a318abe383)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,0.605,105.71,32,26.1,70.193,0,0.00,1837d0374f5820c3,5cb23346ab6be1d8,2.69849,0.469
512,512,0.313,1638.25,32,21.0,71.565,0,0.00,9a9da81aa38cc006,7a8848450b863c49,2.94540,0.469
===== v0-run2.err (sha256 197e0163da87d9e6e680d07d155deee409757adab9e44b9055939e6792b6b6c1)
[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 9.1s | 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 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, 0.418 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] warmstart (parallel): all 10240 experts in RAM (planar int4, the only copy), 10240 in VRAM -- 14.5 s
[mtp] Qwen3.6 checkpoint head: 256 experts
[bench] CUDA ready in 25.91 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, 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) 14.7 ms, rows 2:1.38, rho 0.50, acceptance by position 0.43 0.60 0.50 0.50 0.50
[mtp] head rows 22 (2.44 ms each), fed rows 86 (K/V only, 0.05 ms each), drafts 22
[qwen36 spec gate] 64 tokens: T(1) 14.7 ms, rows 2:1.38, 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) 14.6 ms, rows 2:1.38, rho 0.50, acceptance by position 0.47 0.60 0.50 0.50 0.50
[mtp] head rows 43 (2.44 ms each), fed rows 620 (K/V only, 0.03 ms each), drafts 43
[qwen36 spec gate] 512 tokens: T(1) 14.6 ms, rows 2:1.38, 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 bbf8a347e4367fd8fc3a3fe08206c5e111dd5228b13f2aa98f022ad362725dc3)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,0.608,105.25,32,26.0,70.960,0,0.00,1837d0374f5820c3,5cb23346ab6be1d8,2.69849,0.469
512,512,0.310,1649.81,32,21.0,72.450,0,0.00,9a9da81aa38cc006,7a8848450b863c49,2.94540,0.469
===== v0-run3.err (sha256 db355743908d7095ca85fdc2d621816a909ff38219c7b325c88733f86c04c5c9)
[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 9.1s | 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 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, 0.413 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 -- 14.5 s
[mtp] Qwen3.6 checkpoint head: 256 experts
[bench] CUDA ready in 25.98 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: 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 2:1.36, rho 0.50, acceptance by position 0.43 0.60 0.50 0.50 0.50
[mtp] head rows 22 (2.44 ms each), fed rows 86 (K/V only, 0.05 ms each), drafts 22
[qwen36 spec gate] 64 tokens: T(1) 14.8 ms, rows 2:1.36, 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) 14.5 ms, rows 2:1.37, rho 0.50, acceptance by position 0.47 0.60 0.50 0.50 0.50
[mtp] head rows 43 (2.43 ms each), fed rows 620 (K/V only, 0.03 ms each), drafts 43
[qwen36 spec gate] 512 tokens: T(1) 14.5 ms, rows 2:1.37, 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 20.11 GB

===== v1-run3.csv (sha256 43a4f7c8acdbe178bb4a279c2069925508e0de9bbc0be4fb5f4cd4cb87add969)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,0.613,104.48,32,24.2,71.095,0,0.00,1837d0374f5820c3,5cb23346ab6be1d8,2.69849,0.469
512,512,0.312,1640.92,32,19.1,79.226,0,0.00,9a9da81aa38cc006,7a8848450b863c49,2.94540,0.469
===== v1-run3.err (sha256 e0de67e626d6716bb0511a24a7143e037e68f36fdfcdee47bd3e33b9ae2302c5)
[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 9.2s | 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 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, 0.413 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 -- 14.5 s
[mtp] draft vocabulary: top 16000 + 3/4 rule + conversation, no counts file at /content/q36mtp/.coli_vocab yet
[mtp] Qwen3.6 checkpoint head: 256 experts
[bench] CUDA ready in 26.02 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), drafts score an adaptive vocabulary
[qwen36 mtp] 64 tokens: 1.28 tokens/forward (25 forwards per 32 tokens) | acceptance 29.2% (7/24 drafts) | gate T(1) 14.7 ms, rows 2:1.25, rho 0.50, acceptance by position 0.31 0.60 0.50 0.50 0.50
[mtp] head rows 24 (0.55 ms each), fed rows 88 (K/V only, 0.06 ms each), drafts 24
[mtp] draft list 16004 tokens on average, 68.8% of picks inside, 10 ids appended, 0 rebuilds
[qwen36 spec gate] 64 tokens: T(1) 14.7 ms, rows 2:1.25, rho 0.50, acceptance by position 0.50 0.50 0.50 0.50 0.50
[qwen36 lookup] 64 tokens: 1.28 tokens/forward (25 forwards per 32 tokens) | acceptance 0.0% (0/0 drafts in 24 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) 14.5 ms, rows 2:1.24, rho 0.50, acceptance by position 0.47 0.60 0.50 0.50 0.50
[mtp] head rows 45 (0.54 ms each), fed rows 622 (K/V only, 0.03 ms each), drafts 45
[mtp] draft list 16003 tokens on average, 84.4% of picks inside, 5 ids appended, 0 rebuilds
[qwen36 spec gate] 512 tokens: T(1) 14.5 ms, rows 2:1.24, 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 95023a4154293e401a855c250a85a29b323afef0e4475819f1fd75fc11cca85e)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,0.618,103.59,32,24.7,70.705,0,0.00,1837d0374f5820c3,5cb23346ab6be1d8,2.69849,0.469
512,512,0.316,1619.49,32,19.3,78.985,0,0.00,9a9da81aa38cc006,7a8848450b863c49,2.94540,0.469
===== v1-run4.err (sha256 ceb22b398aa365620babbf3d50eb5a7bd8e138174e0031a5bccb72d05d778e16)
[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 9.2s | 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 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, 0.417 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] warmstart (parallel): all 10240 experts in RAM (planar int4, the only copy), 10240 in VRAM -- 14.5 s
[mtp] draft vocabulary: top 16000 + 3/4 rule + conversation, no counts file at /content/q36mtp/.coli_vocab yet
[mtp] Qwen3.6 checkpoint head: 256 experts
[bench] CUDA ready in 26.17 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), drafts score an adaptive vocabulary
[qwen36 mtp] 64 tokens: 1.28 tokens/forward (25 forwards per 32 tokens) | acceptance 29.2% (7/24 drafts) | gate T(1) 14.7 ms, rows 2:1.25, rho 0.50, acceptance by position 0.31 0.60 0.50 0.50 0.50
[mtp] head rows 24 (0.55 ms each), fed rows 88 (K/V only, 0.06 ms each), drafts 24
[mtp] draft list 16004 tokens on average, 68.8% of picks inside, 10 ids appended, 0 rebuilds
[qwen36 spec gate] 64 tokens: T(1) 14.7 ms, rows 2:1.25, rho 0.50, acceptance by position 0.50 0.50 0.50 0.50 0.50
[qwen36 lookup] 64 tokens: 1.28 tokens/forward (25 forwards per 32 tokens) | acceptance 0.0% (0/0 drafts in 24 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) 14.6 ms, rows 2:1.24, rho 0.50, acceptance by position 0.47 0.60 0.50 0.50 0.50
[mtp] head rows 45 (0.55 ms each), fed rows 622 (K/V only, 0.03 ms each), drafts 45
[mtp] draft list 16003 tokens on average, 84.4% of picks inside, 5 ids appended, 0 rebuilds
[qwen36 spec gate] 512 tokens: T(1) 14.6 ms, rows 2:1.24, 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 837771a5ca6a67f239f890b8ad66b9d8efc61a67d16ab037b0d94f2b8b661a5a)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,0.606,105.69,32,26.2,69.793,0,0.00,1837d0374f5820c3,5cb23346ab6be1d8,2.69849,0.469
512,512,0.314,1629.25,32,21.2,71.386,0,0.00,9a9da81aa38cc006,7a8848450b863c49,2.94540,0.469
===== v0-run4.err (sha256 7068b7413f008e795828ae37dc28b13b5c218e316139bb3b63012bfc4a78c739)
[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 9.1s | 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 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, 0.418 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] warmstart (parallel): all 10240 experts in RAM (planar int4, the only copy), 10240 in VRAM -- 14.5 s
[mtp] Qwen3.6 checkpoint head: 256 experts
[bench] CUDA ready in 26.00 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) 14.7 ms, rows 2:1.39, rho 0.50, acceptance by position 0.43 0.60 0.50 0.50 0.50
[mtp] head rows 22 (2.46 ms each), fed rows 86 (K/V only, 0.06 ms each), drafts 22
[qwen36 spec gate] 64 tokens: T(1) 14.7 ms, rows 2:1.39, 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) 14.6 ms, rows 2:1.38, rho 0.50, acceptance by position 0.47 0.60 0.50 0.50 0.50
[mtp] head rows 43 (2.45 ms each), fed rows 620 (K/V only, 0.03 ms each), drafts 43
[qwen36 spec gate] 512 tokens: T(1) 14.6 ms, rows 2:1.38, 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

===== summary (bench's report of the runs above)
# bench: c6-prosa-1

- bench 482e1711d589; engine qwen36; Linux (Intel(R) Xeon(R) CPU @ 2.20GHz, NVIDIA L4); 2026-10-26 00:04
- 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 c4: 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 | 105.47 (104.52–105.71) | 26 (26–26) | 70.576 (69.793–71.019) | 0 | 0.00 | 2.6985 | 0.469 | 1837d0374f5820c3 |
| 512 | 512 | 1644.03 (1629.25–1651.33) | 21 (20–21) | 72.007 (71.386–72.627) | 0 | 0.00 | 2.9454 | 0.469 | 9a9da81aa38cc006 |

peak memory 20582 MiB, model ready in 25.99 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 | 70.577 | 0.000 |
| 512 | 2 to 32 | 0 | 72.007 | 0.000 |

## v1 c6: 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 | 104.16 (103.59–104.48) | 24 (24–25) | 71.164 (70.705–71.289) | 0 | 0.00 | 2.6985 | 0.469 | 1837d0374f5820c3 |
| 512 | 512 | 1634.57 (1619.49–1640.92) | 19 (19–19) | 78.834 (78.641–79.226) | 0 | 0.00 | 2.9454 | 0.469 | 9a9da81aa38cc006 |

peak memory 20582 MiB, model ready in 25.99 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 | 71.164 | 0.000 |
| 512 | 2 to 32 | 0 | 78.834 | 0.000 |

## v1 c6 against v0 c4 (medians; positive = faster)

| tokens | prompt | first step | decode | results |
|---|---|---|---|---|
| 64 | -1.2%, every run slower | +7.9%, every run faster | +0.8%, ranges overlap | same results; nll 2.6985 → 2.6985, top1 0.469 → 0.469 |
| 512 | -0.6%, ranges overlap | +9.4%, every run faster | +9.5%, 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 | +0.8%, ranges overlap | 0.000 → 0.000 s |
| 512 | 2 to 32 | 0 → 0 | +9.5%, every run faster | 0.000 → 0.000 s |

warm-up (does not count): 64 tokens: prompt 90.07 tok/s, decode 62.968 tok/s; 512 tokens: prompt 776.47 tok/s, decode 67.765 tok/s
runs not cool: none
pages written to swap during the measurement: 0, 0 of them in the runs that count
