# qwen36 CUDA chain, the MTP head's adaptive draft list, code, one draft a step: 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: code.txt beside this file (C: an earlier version of a Metal file of this engine) (sha256
#   ff641dc87569e6e3b56c3e7996fce21627a4b8cea5489d90864d6102232a2e72); 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 00:18
model /content/q36mtp; text codice.txt (sha256 ff641dc87569e6e3); 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/testi/codice.txt --lengths 64,512 --gen 32 --teacher --steps /content/c6-codice-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/testi/codice.txt --lengths 64,512 --gen 32 --teacher --steps /content/c6-codice-1/v1-runN.steps

===== warmup.csv (sha256 777d102b43c6a710466f9bc2c4435882a8688459d353c7a12e40f3e22a648ab6)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,0.650,98.48,32,27.0,64.216,0,0.00,30fd2104818e7459,24b5cd77c9db23fb,2.85121,0.406
512,512,0.414,1235.27,32,21.5,74.356,0,0.00,d8b791a15a3813b4,561100505ac9d4ca,2.56278,0.562
===== warmup.err (sha256 8761a484410c4394b5f8e3fe662719f9cb18ed3ac0a26c83a015a2eeff4966d2)
[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 -- 14.6 s
[mtp] Qwen3.6 checkpoint head: 256 experts
[bench] CUDA ready in 26.20 s; text 7275 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: 1.33 tokens/forward (24 forwards per 32 tokens) | acceptance 36.4% (8/22 drafts) | gate T(1) 16.1 ms, rows 2:1.34, rho 0.50, acceptance by position 0.40 0.60 0.50 0.50 0.50
[mtp] head rows 22 (2.98 ms each), fed rows 87 (K/V only, 0.06 ms each), drafts 22
[qwen36 spec gate] 64 tokens: T(1) 16.1 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: 1.33 tokens/forward (24 forwards per 32 tokens) | acceptance 0.0% (0/0 drafts in 22 verifies) | gate on, 4 declined, 0 probes
[qwen36 mtp] 512 tokens: 1.52 tokens/forward (21 forwards per 32 tokens) | acceptance 57.9% (11/19 drafts) | gate T(1) 14.7 ms, rows 2:1.43, rho 0.50, acceptance by position 0.62 0.60 0.50 0.50 0.50
[mtp] head rows 41 (3.08 ms each), fed rows 619 (K/V only, 0.03 ms each), drafts 41
[qwen36 spec gate] 512 tokens: T(1) 14.7 ms, rows 2:1.43, rho 0.50, acceptance by position 0.50 0.50 0.50 0.50 0.50
[qwen36 lookup] 512 tokens: 1.52 tokens/forward (21 forwards per 32 tokens) | acceptance 0.0% (0/0 drafts in 19 verifies) | gate on, 5 declined, 0 probes
[bench] peak RSS 20.02 GB

===== v0-run1.csv (sha256 9811b0c573e3f0034552117c09cb16169a216f27a724a7b8fae56b91860c58d1)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,0.626,102.29,32,26.5,67.143,0,0.00,30fd2104818e7459,24b5cd77c9db23fb,2.85121,0.406
512,512,0.332,1540.61,32,21.4,78.134,0,0.00,d8b791a15a3813b4,561100505ac9d4ca,2.56278,0.562
===== v0-run1.err (sha256 b08764d66724fb162f4e42da45c6c40a3cee75c05e1047b68ad5d73052541087)
[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, 0.415 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 26.12 s; text 7275 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: 1.33 tokens/forward (24 forwards per 32 tokens) | acceptance 36.4% (8/22 drafts) | gate T(1) 14.8 ms, rows 2:1.39, rho 0.50, acceptance by position 0.40 0.60 0.50 0.50 0.50
[mtp] head rows 22 (2.47 ms each), fed rows 87 (K/V only, 0.06 ms each), drafts 22
[qwen36 spec gate] 64 tokens: T(1) 14.8 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: 1.33 tokens/forward (24 forwards per 32 tokens) | acceptance 0.0% (0/0 drafts in 22 verifies) | gate on, 4 declined, 0 probes
[qwen36 mtp] 512 tokens: 1.52 tokens/forward (21 forwards per 32 tokens) | acceptance 57.9% (11/19 drafts) | gate T(1) 14.5 ms, rows 2:1.39, rho 0.50, acceptance by position 0.62 0.60 0.50 0.50 0.50
[mtp] head rows 41 (2.47 ms each), fed rows 619 (K/V only, 0.03 ms each), drafts 41
[qwen36 spec gate] 512 tokens: T(1) 14.5 ms, rows 2:1.39, rho 0.50, acceptance by position 0.50 0.50 0.50 0.50 0.50
[qwen36 lookup] 512 tokens: 1.52 tokens/forward (21 forwards per 32 tokens) | acceptance 0.0% (0/0 drafts in 19 verifies) | gate on, 5 declined, 0 probes
[bench] peak RSS 20.03 GB

===== v1-run1.csv (sha256 f427385734471a6e45e0af6012510a45e36863cbce7c27e9289e2158295c0234)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,0.628,101.96,32,24.7,72.996,0,0.00,30fd2104818e7459,24b5cd77c9db23fb,2.85121,0.406
512,512,0.327,1567.09,32,19.1,84.236,0,0.00,d8b791a15a3813b4,561100505ac9d4ca,2.56278,0.562
===== v1-run1.err (sha256 e9c4e2573311d791034b0dc17ee3530de60166d8d8e8cd74dab4aa3d321e116a)
[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 9.3s | 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, 0.423 ms on the CPU -> trunk to VRAM
[lmh] lm_head [248320 x 2048] int8 resident on CUDA dev 0 (0.47 GB)
[dnp] 30 DeltaNet-Projektionen auf GPU (0.70 GB VRAM)
[dense] 190 trunk matrices on GPU (30 dnout, 40 attnproj, 120 shexp; 0.61 GB VRAM)
[chain] qwen36: a decode step's routed experts in the frame (the tier's residents by address, format 4, group 64; COLI_CUDA_CHAIN_MOE=0: the host's step)
[chain] qwen36: 40 layers on CUDA device 0 (10 attention), lm_head too; 4.4 MiB of parameters, 35.3 MiB of router, b|a and gate rows, 66 MB of DeltaNet state
[qwen36] int4 packed weights detected — kept int4, repacked planar for expert_ffn.h
[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 26.19 s; text 7275 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: 1.33 tokens/forward (24 forwards per 32 tokens) | acceptance 36.4% (8/22 drafts) | gate T(1) 14.7 ms, rows 2:1.27, rho 0.50, acceptance by position 0.40 0.60 0.50 0.50 0.50
[mtp] head rows 22 (0.56 ms each), fed rows 87 (K/V only, 0.06 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 2:1.27, rho 0.50, acceptance by position 0.50 0.50 0.50 0.50 0.50
[qwen36 lookup] 64 tokens: 1.33 tokens/forward (24 forwards per 32 tokens) | acceptance 0.0% (0/0 drafts in 22 verifies) | gate on, 4 declined, 0 probes
[qwen36 mtp] 512 tokens: 1.52 tokens/forward (21 forwards per 32 tokens) | acceptance 57.9% (11/19 drafts) | gate T(1) 14.7 ms, rows 2:1.27, rho 0.50, acceptance by position 0.62 0.60 0.50 0.50 0.50
[mtp] head rows 41 (0.56 ms each), fed rows 619 (K/V only, 0.03 ms each), drafts 41
[mtp] draft list 16002 tokens on average, 87.5% of picks inside, 4 ids appended, 0 rebuilds
[qwen36 spec gate] 512 tokens: T(1) 14.7 ms, rows 2:1.27, rho 0.50, acceptance by position 0.50 0.50 0.50 0.50 0.50
[qwen36 lookup] 512 tokens: 1.52 tokens/forward (21 forwards per 32 tokens) | acceptance 0.0% (0/0 drafts in 19 verifies) | gate on, 5 declined, 0 probes
[bench] peak RSS 20.02 GB

===== v1-run2.csv (sha256 f7e203c85658d55388cdc5c5d4813f6f3bbcb7de240d7937aab88e320b24287f)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,0.626,102.30,32,25.0,72.857,0,0.00,30fd2104818e7459,24b5cd77c9db23fb,2.85121,0.406
512,512,0.328,1559.52,32,19.4,84.327,0,0.00,d8b791a15a3813b4,561100505ac9d4ca,2.56278,0.562
===== v1-run2.err (sha256 40bbadebcfb1d426e0f3c7942d72f95110251f9c937b6280dc3c80f286a1e86b)
[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 9.3s | 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, 0.418 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 -- 14.6 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 26.33 s; text 7275 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: 1.33 tokens/forward (24 forwards per 32 tokens) | acceptance 36.4% (8/22 drafts) | gate T(1) 14.7 ms, rows 2:1.27, rho 0.50, acceptance by position 0.40 0.60 0.50 0.50 0.50
[mtp] head rows 22 (0.56 ms each), fed rows 87 (K/V only, 0.06 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 2:1.27, rho 0.50, acceptance by position 0.50 0.50 0.50 0.50 0.50
[qwen36 lookup] 64 tokens: 1.33 tokens/forward (24 forwards per 32 tokens) | acceptance 0.0% (0/0 drafts in 22 verifies) | gate on, 4 declined, 0 probes
[qwen36 mtp] 512 tokens: 1.52 tokens/forward (21 forwards per 32 tokens) | acceptance 57.9% (11/19 drafts) | gate T(1) 14.6 ms, rows 2:1.28, rho 0.50, acceptance by position 0.62 0.60 0.50 0.50 0.50
[mtp] head rows 41 (0.56 ms each), fed rows 619 (K/V only, 0.03 ms each), drafts 41
[mtp] draft list 16002 tokens on average, 87.5% of picks inside, 4 ids appended, 0 rebuilds
[qwen36 spec gate] 512 tokens: T(1) 14.6 ms, rows 2:1.28, rho 0.50, acceptance by position 0.50 0.50 0.50 0.50 0.50
[qwen36 lookup] 512 tokens: 1.52 tokens/forward (21 forwards per 32 tokens) | acceptance 0.0% (0/0 drafts in 19 verifies) | gate on, 5 declined, 0 probes
[bench] peak RSS 20.03 GB

===== v0-run2.csv (sha256 fc6a66ed4a308b2a94559b838a63da4c83049f6fc433ff7e66486c587a02502f)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,0.613,104.47,32,26.4,66.499,0,0.00,30fd2104818e7459,24b5cd77c9db23fb,2.85121,0.406
512,512,0.332,1540.55,32,20.9,77.211,0,0.00,d8b791a15a3813b4,561100505ac9d4ca,2.56278,0.562
===== v0-run2.err (sha256 01cb9fe09901c50b93a0f7bea5b3d35c088e3406936645d294cb9f9488aed20b)
[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 0.111 ms on CUDA dev 0, 0.423 ms on the CPU -> trunk to VRAM
[lmh] lm_head [248320 x 2048] int8 resident on CUDA dev 0 (0.47 GB)
[dnp] 30 DeltaNet-Projektionen auf GPU (0.70 GB VRAM)
[dense] 190 trunk matrices on GPU (30 dnout, 40 attnproj, 120 shexp; 0.61 GB VRAM)
[chain] qwen36: a decode step's routed experts in the frame (the tier's residents by address, format 4, group 64; COLI_CUDA_CHAIN_MOE=0: the host's step)
[chain] qwen36: 40 layers on CUDA device 0 (10 attention), lm_head too; 4.4 MiB of parameters, 35.3 MiB of router, b|a and gate rows, 66 MB of DeltaNet state
[qwen36] int4 packed weights detected — kept int4, repacked planar for expert_ffn.h
[qtier] warmstart (parallel): all 10240 experts in RAM (planar int4, the only copy), 10240 in VRAM -- 14.6 s
[mtp] Qwen3.6 checkpoint head: 256 experts
[bench] CUDA ready in 26.30 s; text 7275 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: 1.33 tokens/forward (24 forwards per 32 tokens) | acceptance 36.4% (8/22 drafts) | gate T(1) 14.7 ms, rows 2:1.41, rho 0.50, acceptance by position 0.40 0.60 0.50 0.50 0.50
[mtp] head rows 22 (2.46 ms each), fed rows 87 (K/V only, 0.06 ms each), drafts 22
[qwen36 spec gate] 64 tokens: T(1) 14.7 ms, rows 2:1.41, rho 0.50, acceptance by position 0.50 0.50 0.50 0.50 0.50
[qwen36 lookup] 64 tokens: 1.33 tokens/forward (24 forwards per 32 tokens) | acceptance 0.0% (0/0 drafts in 22 verifies) | gate on, 4 declined, 0 probes
[qwen36 mtp] 512 tokens: 1.52 tokens/forward (21 forwards per 32 tokens) | acceptance 57.9% (11/19 drafts) | gate T(1) 14.6 ms, rows 2:1.41, rho 0.50, acceptance by position 0.62 0.60 0.50 0.50 0.50
[mtp] head rows 41 (2.45 ms each), fed rows 619 (K/V only, 0.03 ms each), drafts 41
[qwen36 spec gate] 512 tokens: T(1) 14.6 ms, rows 2:1.41, rho 0.50, acceptance by position 0.50 0.50 0.50 0.50 0.50
[qwen36 lookup] 512 tokens: 1.52 tokens/forward (21 forwards per 32 tokens) | acceptance 0.0% (0/0 drafts in 19 verifies) | gate on, 5 declined, 0 probes
[bench] peak RSS 20.03 GB

===== v0-run3.csv (sha256 8abd9f96d8ec74d41d4d7debb5f52099622d53932b70e20687bdc68ac54e0748)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,0.612,104.54,32,26.3,67.419,0,0.00,30fd2104818e7459,24b5cd77c9db23fb,2.85121,0.406
512,512,0.328,1560.96,32,21.0,78.371,0,0.00,d8b791a15a3813b4,561100505ac9d4ca,2.56278,0.562
===== v0-run3.err (sha256 871a535eebbc6d6f624219d0ee7ff10f507477c2e9bae3e531819149abeafbf2)
[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 0.116 ms on CUDA dev 0, 0.420 ms on the CPU -> trunk to VRAM
[lmh] lm_head [248320 x 2048] int8 resident on CUDA dev 0 (0.47 GB)
[dnp] 30 DeltaNet-Projektionen auf GPU (0.70 GB VRAM)
[dense] 190 trunk matrices on GPU (30 dnout, 40 attnproj, 120 shexp; 0.61 GB VRAM)
[chain] qwen36: a decode step's routed experts in the frame (the tier's residents by address, format 4, group 64; COLI_CUDA_CHAIN_MOE=0: the host's step)
[chain] qwen36: 40 layers on CUDA device 0 (10 attention), lm_head too; 4.4 MiB of parameters, 35.3 MiB of router, b|a and gate rows, 66 MB of DeltaNet state
[qwen36] int4 packed weights detected — kept int4, repacked planar for expert_ffn.h
[qtier] 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 26.16 s; text 7275 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: 1.33 tokens/forward (24 forwards per 32 tokens) | acceptance 36.4% (8/22 drafts) | gate T(1) 14.7 ms, rows 2:1.38, rho 0.50, acceptance by position 0.40 0.60 0.50 0.50 0.50
[mtp] head rows 22 (2.45 ms each), fed rows 87 (K/V only, 0.06 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: 1.33 tokens/forward (24 forwards per 32 tokens) | acceptance 0.0% (0/0 drafts in 22 verifies) | gate on, 4 declined, 0 probes
[qwen36 mtp] 512 tokens: 1.52 tokens/forward (21 forwards per 32 tokens) | acceptance 57.9% (11/19 drafts) | gate T(1) 14.6 ms, rows 2:1.38, rho 0.50, acceptance by position 0.62 0.60 0.50 0.50 0.50
[mtp] head rows 41 (2.45 ms each), fed rows 619 (K/V only, 0.03 ms each), drafts 41
[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: 1.52 tokens/forward (21 forwards per 32 tokens) | acceptance 0.0% (0/0 drafts in 19 verifies) | gate on, 5 declined, 0 probes
[bench] peak RSS 20.03 GB

===== v1-run3.csv (sha256 663dbbad4a2dfa7c4965f403f1ae70bcc51cea756fe9ca67e1c4647d3afc15df)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,0.631,101.37,32,24.9,72.878,0,0.00,30fd2104818e7459,24b5cd77c9db23fb,2.85121,0.406
512,512,0.318,1607.73,32,19.5,84.118,0,0.00,d8b791a15a3813b4,561100505ac9d4ca,2.56278,0.562
===== v1-run3.err (sha256 2d6cd3cfc2eb9b5d5d613a9eced96808e18ae01980ac3f579bb08a759f0d8c0a)
[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 9.3s | 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
[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 26.18 s; text 7275 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: 1.33 tokens/forward (24 forwards per 32 tokens) | acceptance 36.4% (8/22 drafts) | gate T(1) 14.7 ms, rows 2:1.27, rho 0.50, acceptance by position 0.40 0.60 0.50 0.50 0.50
[mtp] head rows 22 (0.56 ms each), fed rows 87 (K/V only, 0.06 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 2:1.27, rho 0.50, acceptance by position 0.50 0.50 0.50 0.50 0.50
[qwen36 lookup] 64 tokens: 1.33 tokens/forward (24 forwards per 32 tokens) | acceptance 0.0% (0/0 drafts in 22 verifies) | gate on, 4 declined, 0 probes
[qwen36 mtp] 512 tokens: 1.52 tokens/forward (21 forwards per 32 tokens) | acceptance 57.9% (11/19 drafts) | gate T(1) 14.6 ms, rows 2:1.29, rho 0.50, acceptance by position 0.62 0.60 0.50 0.50 0.50
[mtp] head rows 41 (0.56 ms each), fed rows 619 (K/V only, 0.03 ms each), drafts 41
[mtp] draft list 16002 tokens on average, 87.5% of picks inside, 4 ids appended, 0 rebuilds
[qwen36 spec gate] 512 tokens: T(1) 14.6 ms, rows 2:1.29, rho 0.50, acceptance by position 0.50 0.50 0.50 0.50 0.50
[qwen36 lookup] 512 tokens: 1.52 tokens/forward (21 forwards per 32 tokens) | acceptance 0.0% (0/0 drafts in 19 verifies) | gate on, 5 declined, 0 probes
[bench] peak RSS 20.02 GB

===== v1-run4.csv (sha256 66ded596244f18e3df8ba3fe868b847f69094ed128e94f444e09f34e7861faad)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,0.629,101.67,32,24.5,73.115,0,0.00,30fd2104818e7459,24b5cd77c9db23fb,2.85121,0.406
512,512,0.321,1596.74,32,19.4,85.101,0,0.00,d8b791a15a3813b4,561100505ac9d4ca,2.56278,0.562
===== v1-run4.err (sha256 f3ef6152746e92ac5cdde4cc431f9ddd748a322d0c81390061f32be5a38225dd)
[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 9.3s | 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, 0.416 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 -- 14.7 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 26.35 s; text 7275 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: 1.33 tokens/forward (24 forwards per 32 tokens) | acceptance 36.4% (8/22 drafts) | gate T(1) 14.7 ms, rows 2:1.27, rho 0.50, acceptance by position 0.40 0.60 0.50 0.50 0.50
[mtp] head rows 22 (0.56 ms each), fed rows 87 (K/V only, 0.06 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 2:1.27, rho 0.50, acceptance by position 0.50 0.50 0.50 0.50 0.50
[qwen36 lookup] 64 tokens: 1.33 tokens/forward (24 forwards per 32 tokens) | acceptance 0.0% (0/0 drafts in 22 verifies) | gate on, 4 declined, 0 probes
[qwen36 mtp] 512 tokens: 1.52 tokens/forward (21 forwards per 32 tokens) | acceptance 57.9% (11/19 drafts) | gate T(1) 14.6 ms, rows 2:1.27, rho 0.50, acceptance by position 0.62 0.60 0.50 0.50 0.50
[mtp] head rows 41 (0.55 ms each), fed rows 619 (K/V only, 0.03 ms each), drafts 41
[mtp] draft list 16002 tokens on average, 87.5% of picks inside, 4 ids appended, 0 rebuilds
[qwen36 spec gate] 512 tokens: T(1) 14.6 ms, rows 2:1.27, rho 0.50, acceptance by position 0.50 0.50 0.50 0.50 0.50
[qwen36 lookup] 512 tokens: 1.52 tokens/forward (21 forwards per 32 tokens) | acceptance 0.0% (0/0 drafts in 19 verifies) | gate on, 5 declined, 0 probes
[bench] peak RSS 20.04 GB

===== v0-run4.csv (sha256 14eed5ebec9b6a62219cef7a72918d46efb9ea76c68bf51f950f8127cc128be5)
ctx_tokens,prefill_tokens,prefill_s,prefill_tps,gen_tokens,gen_first_ms,gen_steady_tps,expert_miss,disk_s,fingerprint,tokens,nll,top1
64,64,0.615,104.10,32,26.3,66.863,0,0.00,30fd2104818e7459,24b5cd77c9db23fb,2.85121,0.406
512,512,0.330,1553.38,32,20.7,77.612,0,0.00,d8b791a15a3813b4,561100505ac9d4ca,2.56278,0.562
===== v0-run4.err (sha256 65a1a29967be60b7566f15f3eb39b1becc3b5e977a80791ea7f6dc2cbef2b7d1)
[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, 0.420 ms on the CPU -> trunk to VRAM
[lmh] lm_head [248320 x 2048] int8 resident on CUDA dev 0 (0.47 GB)
[dnp] 30 DeltaNet-Projektionen auf GPU (0.70 GB VRAM)
[dense] 190 trunk matrices on GPU (30 dnout, 40 attnproj, 120 shexp; 0.61 GB VRAM)
[chain] qwen36: a decode step's routed experts in the frame (the tier's residents by address, format 4, group 64; COLI_CUDA_CHAIN_MOE=0: the host's step)
[chain] qwen36: 40 layers on CUDA device 0 (10 attention), lm_head too; 4.4 MiB of parameters, 35.3 MiB of router, b|a and gate rows, 66 MB of DeltaNet state
[qwen36] int4 packed weights detected — kept int4, repacked planar for expert_ffn.h
[qtier] warmstart (parallel): all 10240 experts in RAM (planar int4, the only copy), 10240 in VRAM -- 14.6 s
[mtp] Qwen3.6 checkpoint head: 256 experts
[bench] CUDA ready in 26.22 s; text 7275 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: 1.33 tokens/forward (24 forwards per 32 tokens) | acceptance 36.4% (8/22 drafts) | gate T(1) 14.7 ms, rows 2:1.40, rho 0.50, acceptance by position 0.40 0.60 0.50 0.50 0.50
[mtp] head rows 22 (2.45 ms each), fed rows 87 (K/V only, 0.06 ms each), drafts 22
[qwen36 spec gate] 64 tokens: T(1) 14.7 ms, rows 2:1.40, rho 0.50, acceptance by position 0.50 0.50 0.50 0.50 0.50
[qwen36 lookup] 64 tokens: 1.33 tokens/forward (24 forwards per 32 tokens) | acceptance 0.0% (0/0 drafts in 22 verifies) | gate on, 4 declined, 0 probes
[qwen36 mtp] 512 tokens: 1.52 tokens/forward (21 forwards per 32 tokens) | acceptance 57.9% (11/19 drafts) | gate T(1) 14.6 ms, rows 2:1.39, rho 0.50, acceptance by position 0.62 0.60 0.50 0.50 0.50
[mtp] head rows 41 (2.45 ms each), fed rows 619 (K/V only, 0.03 ms each), drafts 41
[qwen36 spec gate] 512 tokens: T(1) 14.6 ms, rows 2:1.39, rho 0.50, acceptance by position 0.50 0.50 0.50 0.50 0.50
[qwen36 lookup] 512 tokens: 1.52 tokens/forward (21 forwards per 32 tokens) | acceptance 0.0% (0/0 drafts in 19 verifies) | gate on, 5 declined, 0 probes
[bench] peak RSS 20.03 GB

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

- bench 482e1711d589; engine qwen36; Linux (Intel(R) Xeon(R) CPU @ 2.20GHz, NVIDIA L4); 2026-10-26 00:18
- model q36mtp; text codice.txt (sha256 ff641dc87569e6e3); 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 | 104.28 (102.29–104.54) | 26 (26–26) | 67.003 (66.499–67.419) | 0 | 0.00 | 2.8512 | 0.406 | 30fd2104818e7459 |
| 512 | 512 | 1546.99 (1540.55–1560.96) | 21 (21–21) | 77.873 (77.211–78.371) | 0 | 0.00 | 2.5628 | 0.562 | d8b791a15a3813b4 |

peak memory 20511 MiB, model ready in 26.19 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 | 67.003 | 0.000 |
| 512 | 2 to 32 | 0 | 77.873 | 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 | 101.81 (101.37–102.30) | 25 (24–25) | 72.937 (72.857–73.115) | 0 | 0.00 | 2.8512 | 0.406 | 30fd2104818e7459 |
| 512 | 512 | 1581.91 (1559.52–1607.73) | 19 (19–20) | 84.281 (84.118–85.101) | 0 | 0.00 | 2.5628 | 0.562 | d8b791a15a3813b4 |

peak memory 20506 MiB, model ready in 26.26 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 | 72.937 | 0.000 |
| 512 | 2 to 32 | 0 | 84.281 | 0.000 |

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

| tokens | prompt | first step | decode | results |
|---|---|---|---|---|
| 64 | -2.4%, ranges overlap | +6.3%, every run faster | +8.9%, every run faster | same results; nll 2.8512 → 2.8512, top1 0.406 → 0.406 |
| 512 | +2.3%, ranges overlap | +8.0%, every run faster | +8.2%, every run faster | same results; nll 2.5628 → 2.5628, top1 0.562 → 0.562 |

| tokens | steps | experts read | decode | reads' s |
|---|---|---|---|---|
| 64 | 2 to 32 | 0 → 0 | +8.9%, every run faster | 0.000 → 0.000 s |
| 512 | 2 to 32 | 0 → 0 | +8.2%, every run faster | 0.000 → 0.000 s |

warm-up (does not count): 64 tokens: prompt 98.48 tok/s, decode 64.216 tok/s; 512 tokens: prompt 1235.27 tok/s, decode 74.356 tok/s
runs not cool: none
pages written to swap during the measurement: 0, 0 of them in the runs that count
