# qwen36 CUDA chain, the MTP head's adaptive draft list, prose, one draft a step, warm counts: 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 (this commit's code).
#   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.)
# Counts: both versions were given the same counts file, coli_vocab beside this file (sha256
#   9f32c38c5b516c41fc61f66c1b4c9cbadb2b86b847a0ce733cb5ade1a93843c1; the same before and after the runs), with
#   COLI_VOCAB_USAGE=<it> USAGE_SAVE=0; v0 does not read it. This change's engine wrote it on the same VM
#   before the runs, as a user's earlier runs would (counts.txt beside 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 02:32
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 COLI_VOCAB_USAGE=/content/warm/coli_vocab USAGE_SAVE=0
   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 COLI_VOCAB_USAGE=/content/warm/coli_vocab USAGE_SAVE=0 /content/bin-c4/qwen36-bench --model /content/q36mtp --text /content/bench/promessi_sposi.txt --lengths 64,512 --gen 32 --teacher --steps /content/c6caldo-prosa-1/v0-runN.steps
v1 c6: /content/bin-c6/qwen36-bench (commit 8aaf4d788308ddac1e0e9e63ea507e152fda9cc5; 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 COLI_VOCAB_USAGE=/content/warm/coli_vocab USAGE_SAVE=0
   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 COLI_VOCAB_USAGE=/content/warm/coli_vocab USAGE_SAVE=0 /content/bin-c6/qwen36-bench --model /content/q36mtp --text /content/bench/promessi_sposi.txt --lengths 64,512 --gen 32 --teacher --steps /content/c6caldo-prosa-1/v1-runN.steps

===== warmup.csv (sha256 c6b88b758ed8b9ef2461b5a50dd6c1ca19f79887a08d4a8058683d0094b6f350)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,0.607,105.41,32,25.8,71.869,0,0.00,1837d0374f5820c3,5cb23346ab6be1d8,2.69849,0.469
512,512,0.325,1577.74,32,20.5,73.216,0,0.00,9a9da81aa38cc006,7a8848450b863c49,2.94540,0.469
===== warmup.err (sha256 b44f08ab3b227a470d097b0903aae5f0d6e4d4debe9c96c077cda488aeaf6bd5)
[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 0.116 ms on CUDA dev 0, 0.416 ms on the CPU -> trunk to VRAM
[lmh] lm_head [248320 x 2048] int8 resident on CUDA dev 0 (0.47 GB)
[dnp] 30 DeltaNet-Projektionen auf GPU (0.70 GB VRAM)
[dense] 190 trunk matrices on GPU (30 dnout, 40 attnproj, 120 shexp; 0.61 GB VRAM)
[chain] qwen36: a decode step's routed experts in the frame (the tier's residents by address, format 4, group 64; COLI_CUDA_CHAIN_MOE=0: the host's step)
[chain] qwen36: 40 layers on CUDA device 0 (10 attention), lm_head too; 4.4 MiB of parameters, 35.3 MiB of router, b|a and gate rows, 66 MB of DeltaNet state
[qwen36] int4 packed weights detected — kept int4, repacked planar for expert_ffn.h
[qtier] warmstart (parallel): all 10240 experts in RAM (planar int4, the only copy), 10240 in VRAM -- 14.4 s
[mtp] Qwen3.6 checkpoint head: 256 experts
[bench] CUDA ready in 25.85 s; text 6182 tokens; lengths 64,512; gen 32; each length from an empty state
[qwen36] KV cache: float16, 0.00 GiB at 96 positions (half of float32; Q36_KV=f32 keeps float32)
[chain] qwen36: 4096 rows per prompt chunk (286 KiB a row, 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 2:1.36, 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) 14.7 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) 14.8 ms, rows 2:1.33, 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) 14.8 ms, rows 2:1.33, 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

===== v0-run1.csv (sha256 ee7f76e151fda07d1967174b691e72b1528b5e9060e8e160f8dfdc829a9628b2)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,0.602,106.27,32,25.9,72.085,0,0.00,1837d0374f5820c3,5cb23346ab6be1d8,2.69849,0.469
512,512,0.327,1568.06,32,20.4,73.588,0,0.00,9a9da81aa38cc006,7a8848450b863c49,2.94540,0.469
===== v0-run1.err (sha256 a8735953a37a34a040570e38881625be1545f8c6d1a0673a2be5f74b300fd85e)
[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 0.114 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] Qwen3.6 checkpoint head: 256 experts
[bench] CUDA ready in 25.84 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) 14.6 ms, rows 2:1.35, 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) 14.6 ms, rows 2:1.35, 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 2:1.35, 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 2:1.35, 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 76b24554f79fd1dce1644ce113e2e411ed1959f9c9e2a10af7603c24bbafe02b)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,0.614,104.25,32,24.0,79.499,0,0.00,1837d0374f5820c3,5cb23346ab6be1d8,2.69849,0.469
512,512,0.320,1602.48,32,18.7,80.786,0,0.00,9a9da81aa38cc006,7a8848450b863c49,2.94540,0.469
===== v1-run1.err (sha256 6a40c4c867fd5706510641146930abc7c973bf116ee43e91c6ead22ec80b1ccd)
[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, 0.415 ms on the CPU -> trunk to VRAM
[lmh] lm_head [248320 x 2048] int8 resident on CUDA dev 0 (0.47 GB)
[dnp] 30 DeltaNet-Projektionen auf GPU (0.70 GB VRAM)
[dense] 190 trunk matrices on GPU (30 dnout, 40 attnproj, 120 shexp; 0.61 GB VRAM)
[chain] qwen36: a decode step's routed experts in the frame (the tier's residents by address, format 4, group 64; COLI_CUDA_CHAIN_MOE=0: the host's step)
[chain] qwen36: 40 layers on CUDA device 0 (10 attention), lm_head too; 4.4 MiB of parameters, 35.3 MiB of router, b|a and gate rows, 66 MB of DeltaNet state
[qwen36] int4 packed weights detected — kept int4, repacked planar for expert_ffn.h
[qtier] 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, counts from /content/warm/coli_vocab (1588 tokens, 5346 distinct)
[mtp] Qwen3.6 checkpoint head: 256 experts
[bench] CUDA ready in 25.86 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: 1.39 tokens/forward (23 forwards per 32 tokens) | acceptance 40.9% (9/22 drafts) | gate T(1) 14.7 ms, rows 2:1.22, rho 0.50, acceptance by position 0.43 0.60 0.50 0.50 0.50
[mtp] head rows 22 (0.54 ms each), fed rows 86 (K/V only, 0.05 ms each), drafts 22
[mtp] draft list 16002 tokens on average, 90.6% of picks inside, 3 ids appended, 0 rebuilds
[qwen36 spec gate] 64 tokens: T(1) 14.7 ms, rows 2:1.22, 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 2:1.21, rho 0.50, acceptance by position 0.47 0.60 0.50 0.50 0.50
[mtp] head rows 43 (0.54 ms each), fed rows 620 (K/V only, 0.03 ms each), drafts 43
[mtp] draft list 16001 tokens on average, 93.8% of picks inside, 2 ids appended, 0 rebuilds
[qwen36 spec gate] 512 tokens: T(1) 14.6 ms, rows 2:1.21, 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 85a88a8827a34e04bbb01a5446c4c5bb76e4d0816c06d364ed9315125ed87360)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,0.619,103.38,32,24.3,79.736,0,0.00,1837d0374f5820c3,5cb23346ab6be1d8,2.69849,0.469
512,512,0.325,1576.16,32,18.5,80.987,0,0.00,9a9da81aa38cc006,7a8848450b863c49,2.94540,0.469
===== v1-run2.err (sha256 dbf64e866ac7a48933741021f64aaa014d60bf11b847038de75ea7303216c5a2)
[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, 0.416 ms on the CPU -> trunk to VRAM
[lmh] lm_head [248320 x 2048] int8 resident on CUDA dev 0 (0.47 GB)
[dnp] 30 DeltaNet-Projektionen auf GPU (0.70 GB VRAM)
[dense] 190 trunk matrices on GPU (30 dnout, 40 attnproj, 120 shexp; 0.61 GB VRAM)
[chain] qwen36: a decode step's routed experts in the frame (the tier's residents by address, format 4, group 64; COLI_CUDA_CHAIN_MOE=0: the host's step)
[chain] qwen36: 40 layers on CUDA device 0 (10 attention), lm_head too; 4.4 MiB of parameters, 35.3 MiB of router, b|a and gate rows, 66 MB of DeltaNet state
[qwen36] int4 packed weights detected — kept int4, repacked planar for expert_ffn.h
[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 -- 14.3 s
[mtp] draft vocabulary: top 16000 + 3/4 rule + conversation, counts from /content/warm/coli_vocab (1588 tokens, 5346 distinct)
[mtp] Qwen3.6 checkpoint head: 256 experts
[bench] CUDA ready in 25.76 s; text 6182 tokens; lengths 64,512; gen 32; each length from an empty state
[qwen36] KV cache: float16, 0.00 GiB at 96 positions (half of float32; Q36_KV=f32 keeps float32)
[chain] qwen36: 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: 1.39 tokens/forward (23 forwards per 32 tokens) | acceptance 40.9% (9/22 drafts) | gate T(1) 14.9 ms, rows 2:1.20, rho 0.50, acceptance by position 0.43 0.60 0.50 0.50 0.50
[mtp] head rows 22 (0.53 ms each), fed rows 86 (K/V only, 0.05 ms each), drafts 22
[mtp] draft list 16002 tokens on average, 90.6% of picks inside, 3 ids appended, 0 rebuilds
[qwen36 spec gate] 64 tokens: T(1) 14.9 ms, rows 2:1.20, 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 2:1.21, rho 0.50, acceptance by position 0.47 0.60 0.50 0.50 0.50
[mtp] head rows 43 (0.53 ms each), fed rows 620 (K/V only, 0.03 ms each), drafts 43
[mtp] draft list 16001 tokens on average, 93.8% of picks inside, 2 ids appended, 0 rebuilds
[qwen36 spec gate] 512 tokens: T(1) 14.5 ms, rows 2:1.21, 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 ed56299604e19fd12a5be500df299369c16c738ee632b60f4856cb595390c635)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,0.604,105.94,32,26.0,71.992,0,0.00,1837d0374f5820c3,5cb23346ab6be1d8,2.69849,0.469
512,512,0.325,1577.21,32,20.4,73.399,0,0.00,9a9da81aa38cc006,7a8848450b863c49,2.94540,0.469
===== v0-run2.err (sha256 9f1f53a7b4b607adb52f365241b918efc128517140e1d522ab1ee57784093a9f)
[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, 0.414 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 -- 14.5 s
[mtp] Qwen3.6 checkpoint head: 256 experts
[bench] CUDA ready in 25.86 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 2:1.34, 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 2:1.34, 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 2:1.35, 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 2:1.35, 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 68edd501094da50ee40cdbc5d9810a1a530c8b6a1b151c279a4ac5b9dd600c83)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,0.601,106.44,32,25.8,72.027,0,0.00,1837d0374f5820c3,5cb23346ab6be1d8,2.69849,0.469
512,512,0.323,1584.31,32,20.5,73.405,0,0.00,9a9da81aa38cc006,7a8848450b863c49,2.94540,0.469
===== v0-run3.err (sha256 3d21d9fd023269d03c51ea8a03fc2e93269826749435a9d567707eb56b9d1016)
[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 0.111 ms on CUDA dev 0, 0.415 ms on the CPU -> trunk to VRAM
[lmh] lm_head [248320 x 2048] int8 resident on CUDA dev 0 (0.47 GB)
[dnp] 30 DeltaNet-Projektionen auf GPU (0.70 GB VRAM)
[dense] 190 trunk matrices on GPU (30 dnout, 40 attnproj, 120 shexp; 0.61 GB VRAM)
[chain] qwen36: a decode step's routed experts in the frame (the tier's residents by address, format 4, group 64; COLI_CUDA_CHAIN_MOE=0: the host's step)
[chain] qwen36: 40 layers on CUDA device 0 (10 attention), lm_head too; 4.4 MiB of parameters, 35.3 MiB of router, b|a and gate rows, 66 MB of DeltaNet state
[qwen36] int4 packed weights detected — kept int4, repacked planar for expert_ffn.h
[qtier] 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 25.87 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 2:1.34, 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) 14.9 ms, rows 2:1.34, 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 2:1.35, 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 2:1.35, 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-run3.csv (sha256 5fd18f32cb752654c0be005ada804d9f4db285c8de2912d2be34f85b928b2d5c)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,0.613,104.34,32,23.9,79.701,0,0.00,1837d0374f5820c3,5cb23346ab6be1d8,2.69849,0.469
512,512,0.321,1592.65,32,18.6,81.232,0,0.00,9a9da81aa38cc006,7a8848450b863c49,2.94540,0.469
===== v1-run3.err (sha256 c722d63e35fb334a00565e6c3c665a8f741fef8e18c7ae7cb095e499e25b1e1f)
[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, 0.413 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] draft vocabulary: top 16000 + 3/4 rule + conversation, counts from /content/warm/coli_vocab (1588 tokens, 5346 distinct)
[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: 1.39 tokens/forward (23 forwards per 32 tokens) | acceptance 40.9% (9/22 drafts) | gate T(1) 14.6 ms, rows 2:1.22, rho 0.50, acceptance by position 0.43 0.60 0.50 0.50 0.50
[mtp] head rows 22 (0.54 ms each), fed rows 86 (K/V only, 0.05 ms each), drafts 22
[mtp] draft list 16002 tokens on average, 90.6% of picks inside, 3 ids appended, 0 rebuilds
[qwen36 spec gate] 64 tokens: T(1) 14.6 ms, rows 2:1.22, 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 2:1.21, rho 0.50, acceptance by position 0.47 0.60 0.50 0.50 0.50
[mtp] head rows 43 (0.53 ms each), fed rows 620 (K/V only, 0.03 ms each), drafts 43
[mtp] draft list 16001 tokens on average, 93.8% of picks inside, 2 ids appended, 0 rebuilds
[qwen36 spec gate] 512 tokens: T(1) 14.5 ms, rows 2:1.21, 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-run4.csv (sha256 5f92360cf9e1992630d74bb62306c4937ef60144801c93ae122a72cef421e986)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,0.613,104.32,32,23.8,79.534,0,0.00,1837d0374f5820c3,5cb23346ab6be1d8,2.69849,0.469
512,512,0.326,1572.11,32,18.7,81.179,0,0.00,9a9da81aa38cc006,7a8848450b863c49,2.94540,0.469
===== v1-run4.err (sha256 5ab4da83e5570d33f7cad8ee42b10c51474bd3c227e9e1f7fa780d273b317c43)
[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 0.114 ms on CUDA dev 0, 0.415 ms on the CPU -> trunk to VRAM
[lmh] lm_head [248320 x 2048] int8 resident on CUDA dev 0 (0.47 GB)
[dnp] 30 DeltaNet-Projektionen auf GPU (0.70 GB VRAM)
[dense] 190 trunk matrices on GPU (30 dnout, 40 attnproj, 120 shexp; 0.61 GB VRAM)
[chain] qwen36: a decode step's routed experts in the frame (the tier's residents by address, format 4, group 64; COLI_CUDA_CHAIN_MOE=0: the host's step)
[chain] qwen36: 40 layers on CUDA device 0 (10 attention), lm_head too; 4.4 MiB of parameters, 35.3 MiB of router, b|a and gate rows, 66 MB of DeltaNet state
[qwen36] int4 packed weights detected — kept int4, repacked planar for expert_ffn.h
[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, counts from /content/warm/coli_vocab (1588 tokens, 5346 distinct)
[mtp] Qwen3.6 checkpoint head: 256 experts
[bench] CUDA ready in 25.87 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: 1.39 tokens/forward (23 forwards per 32 tokens) | acceptance 40.9% (9/22 drafts) | gate T(1) 15.0 ms, rows 2:1.19, rho 0.50, acceptance by position 0.43 0.60 0.50 0.50 0.50
[mtp] head rows 22 (0.54 ms each), fed rows 86 (K/V only, 0.05 ms each), drafts 22
[mtp] draft list 16002 tokens on average, 90.6% of picks inside, 3 ids appended, 0 rebuilds
[qwen36 spec gate] 64 tokens: T(1) 15.0 ms, rows 2:1.19, 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 2:1.21, rho 0.50, acceptance by position 0.47 0.60 0.50 0.50 0.50
[mtp] head rows 43 (0.53 ms each), fed rows 620 (K/V only, 0.03 ms each), drafts 43
[mtp] draft list 16001 tokens on average, 93.8% of picks inside, 2 ids appended, 0 rebuilds
[qwen36 spec gate] 512 tokens: T(1) 14.5 ms, rows 2:1.21, 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-run4.csv (sha256 28f725856f4fbe9a41602447898f1775223d364b633152291f59afb27fe0218e)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,0.603,106.16,32,25.6,71.903,0,0.00,1837d0374f5820c3,5cb23346ab6be1d8,2.69849,0.469
512,512,0.319,1605.90,32,20.4,73.496,0,0.00,9a9da81aa38cc006,7a8848450b863c49,2.94540,0.469
===== v0-run4.err (sha256 33a1a587cf31bed004ab0a22da7599412bb5a9f29cfc7aed599bbedf1de3c320)
[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, 0.415 ms on the CPU -> trunk to VRAM
[lmh] lm_head [248320 x 2048] int8 resident on CUDA dev 0 (0.47 GB)
[dnp] 30 DeltaNet-Projektionen auf GPU (0.70 GB VRAM)
[dense] 190 trunk matrices on GPU (30 dnout, 40 attnproj, 120 shexp; 0.61 GB VRAM)
[chain] qwen36: a decode step's routed experts in the frame (the tier's residents by address, format 4, group 64; COLI_CUDA_CHAIN_MOE=0: the host's step)
[chain] qwen36: 40 layers on CUDA device 0 (10 attention), lm_head too; 4.4 MiB of parameters, 35.3 MiB of router, b|a and gate rows, 66 MB of DeltaNet state
[qwen36] int4 packed weights detected — kept int4, repacked planar for expert_ffn.h
[qtier] warmstart (parallel): all 10240 experts in RAM (planar int4, the only copy), 10240 in VRAM -- 14.4 s
[mtp] Qwen3.6 checkpoint head: 256 experts
[bench] CUDA ready in 25.84 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 2:1.35, 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 2:1.35, 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 2:1.35, 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) 14.6 ms, rows 2:1.35, 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

===== summary (bench's report of the runs above)
# bench: c6caldo-prosa-1

- bench 482e1711d589; engine qwen36; Linux (Intel(R) Xeon(R) CPU @ 2.20GHz, NVIDIA L4); 2026-10-26 02:32
- 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 | 106.22 (105.94–106.44) | 26 (26–26) | 72.010 (71.903–72.085) | 0 | 0.00 | 2.6985 | 0.469 | 1837d0374f5820c3 |
| 512 | 512 | 1580.76 (1568.06–1605.90) | 20 (20–20) | 73.451 (73.399–73.588) | 0 | 0.00 | 2.9454 | 0.469 | 9a9da81aa38cc006 |

peak memory 20582 MiB, model ready in 25.85 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 | 72.009 | 0.000 |
| 512 | 2 to 32 | 0 | 73.451 | 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 | 104.28 (103.38–104.34) | 24 (24–24) | 79.618 (79.499–79.736) | 0 | 0.00 | 2.6985 | 0.469 | 1837d0374f5820c3 |
| 512 | 512 | 1584.41 (1572.11–1602.48) | 19 (18–19) | 81.083 (80.786–81.232) | 0 | 0.00 | 2.9454 | 0.469 | 9a9da81aa38cc006 |

peak memory 20582 MiB, model ready in 25.87 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 | 79.618 | 0.000 |
| 512 | 2 to 32 | 0 | 81.083 | 0.000 |

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

| tokens | prompt | first step | decode | results |
|---|---|---|---|---|
| 64 | -1.8%, every run slower | +7.9%, every run faster | +10.6%, every run faster | same results; nll 2.6985 → 2.6985, top1 0.469 → 0.469 |
| 512 | +0.2%, ranges overlap | +9.4%, every run faster | +10.4%, 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 | +10.6%, every run faster | 0.000 → 0.000 s |
| 512 | 2 to 32 | 0 → 0 | +10.4%, every run faster | 0.000 → 0.000 s |

warm-up (does not count): 64 tokens: prompt 105.41 tok/s, decode 71.869 tok/s; 512 tokens: prompt 1577.74 tok/s, decode 73.216 tok/s
runs not cool: none
pages written to swap during the measurement: 0, 0 of them in the runs that count
