Skip to content

Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦

C

docs(vulkan): attach repeated MTP measurements and reproducible inputs

JustVugg1d758d7ee27e
diff --git a/docs/experiments/vulkan-qwen38-mtp-2026-10-05/baseline.json b/docs/experiments/vulkan-qwen38-mtp-2026-10-05/baseline.json
new file mode 100644
index 00000000..7dc738dc
--- /dev/null
+++ b/docs/experiments/vulkan-qwen38-mtp-2026-10-05/baseline.json
@@ -0,0 +1,166 @@
+{
+ "runs": [
+ {
+ "tag": "q38-i4g64-mtp-baseline-1",
+ "family": "qwen38",
+ "model": "/root/colibri-test/models/qwen38-flash-next-fp8",
+ "prompt": "/root/colibri-test/p4mtp/prompts/code.txt",
+ "new": 128,
+ "cap": 170,
+ "cold": true,
+ "env": {
+ "COLI_VULKAN": "1",
+ "COLI_VK_CHAIN": "0",
+ "COLI_VK_DENSE": "0",
+ "Q38_EXPERT_INT4": "1",
+ "Q38_MTP": "0",
+ "Q38_MTP_DRAFTS": "3"
+ },
+ "kind": "generate",
+ "prompt_sha256": "b840c508d4d473840929f53cc370614e1438c6d0fd5f9afc67176ce1d99dfc2f",
+ "history_sha256": "d7d86f69d2c83d900c7d869b7a04e34be19a4f466eb40ec7c52deb70d50fc108",
+ "cmd": [
+ "/root/colibri-test/colibri-pr-validation-20261005/src/c/qwen38",
+ "170",
+ "8",
+ "/root/colibri-test/p4mtp/prompts/code.txt"
+ ],
+ "settings": {
+ "OMP_NUM_THREADS": "8",
+ "COLI_NO_OMP_TUNE": "1",
+ "COLI_TIMERS": "1",
+ "VK_ICD_FILENAMES": "/usr/share/vulkan/icd.d/radeon_icd.json",
+ "COLI_LOOKUP": "0",
+ "Q38_MTP": "0",
+ "COLI_VULKAN": "1",
+ "COLI_VK_CHAIN": "0",
+ "COLI_VK_DENSE": "0",
+ "Q38_EXPERT_INT4": "1",
+ "Q38_MTP_DRAFTS": "3",
+ "SNAP": "/root/colibri-test/models/qwen38-flash-next-fp8",
+ "COLI_USAGE": "/root/colibri-test/colibri-pr-validation-20261005/results/q38-i4g64-mtp-baseline-1/run.usage",
+ "N_NEW": "128"
+ },
+ "load_start": [
+ 0.09716796875,
+ 0.1455078125,
+ 0.76318359375
+ ],
+ "source": "2f33f211ea6e28186d62cf9c00d511df826590a0",
+ "rc": 0,
+ "wall_s": 54.302,
+ "peak_rss_kib": 29819624,
+ "other_engines": [],
+ "stdout": "```c\nstatic int count_words(const char *s) {\n int total = 0, in_word = 0;\n for (const char *p = s; *p; p++) {\n if (*p == ' ' || *p == '\\n' || *p == '\\t') {\n in_word = 0;\n } else if (!in_word) {\n in_word = 1;\n total++;\n }\n }\n return total;\n}\n\nstatic void print_counts(const char **lines, int n) {\n int total",
+ "stderr": "== Qwen3.8-Flash-Next native text engine | cache=170/layer | CPU ==\n[tok] loaded 248044 base pieces + 33 added tokens (max id 248076) from /root/colibri-test/models/qwen38-flash-next-fp8/tokenizer.json\n[enc] prompt tokens: 241 | generating 128 new tokens\n[qwen38] vision tower: 27 blocks, hidden 1152, 16 heads, patch 16, merge 2\n[qwen38] native text weights: prefix=model.language_model, 48 layers, PLE=1, cache=170/layer, FP8=native, BF16=native (resident matrices 10.88 GiB)\n[qwen38] trunk: 553 matrices int8 on the CPU (3.97 GiB, 7.93 GiB of BF16 released) in 1.9s; Q38_TRUNK_CPU_INT8=0 keeps BF16\n[qwen38] routed experts: int4-g64 from experts-int4g64/, 2764800 bytes (2.64 MiB) per expert (the snapshot's native FP8: 4915200 bytes); cache 170/layer x 48 layers = 21.01 GiB full; 56.1 GB available now holds 422/layer\n[VK] VRAM pressure-proofing: memory_priority on, memory_budget on\n[VK] ready: AMD Radeon 780M Graphics (RADV PHOENIX), compute qfam 0, memtype 3, fused gate+up, absorb attention, tiled GEMM from S=2 and S*O>=4096 (cooperative matrix)\n[VK] qwen38: device ready, dense matrices on the CPU (COLI_VK_DENSE=0)\n[VK] qwen38: dense chain off (COLI_VK_CHAIN=0)\n[VK] qwen38: dense weights on the device and in host RAM (the dense part runs on the CPU)\n[USAGE] expert history: 158400 selections (/root/colibri-test/colibri-pr-validation-20261005/results/q38-i4g64-mtp-baseline-1/run.usage)\n[VK] tier qwen38: on, AMD Radeon 780M Graphics (RADV PHOENIX), budget 7.31 GiB = 2770 experts of 2.6 MiB (fmt 4 gs 64, down fmt 4 gs 64), shared RAM: a quarter of what the expert cache leaves, own queue, up to 16 promotions per token, balanced against the CPU, streaming cold experts of prompt steps through 64 slots\n[VK] tier qwen38: warm start, 2770 experts from the history in 1.9s\nresident weights loaded in 4.5s | RSS after load: 11.61 GB\nGenerated (128 new tokens):\nText : [VK] tier qwen38 stream: a step of 241 rows; cold experts with 16 rows or more go to the device (the GEMM's 16 rows: the upload and the CPU's rows not measured yet); layer 0: 61 resident, 31 streamed, 289 kept on the CPU (1475 rows)\n\nTTFT: 16.69 s (time to first token)\n[qwen38 timers] total: 128 forwards\n[qwen38 timers] expert-read 8.288 s 64.749 ms/forward\n[qwen38 timers] fp8-expand 0.000 s 0.000 ms/forward\n[qwen38 timers] routed-expert 5.547 s 43.333 ms/forward\n[qwen38 timers] shared-expert 1.054 s 8.234 ms/forward\n[qwen38 timers] resident-mm 11.637 s 90.913 ms/forward\n[qwen38 timers] deltanet 10.243 s 80.022 ms/forward\n[qwen38 timers] qsa-index 0.110 s 0.863 ms/forward\n[qwen38 timers] qsa-attn 2.083 s 16.274 ms/forward\n[qwen38 timers] ple 1.031 s 8.057 ms/forward\n[qwen38 timers] lm-head 1.359 s 10.614 ms/forward\n[qwen38 timers] architecture phases overlap resident-mm; expert-read is disk service and fp8-expand is synchronous miss work\n[qwen38 timers] prefill: 1 forwards\n[qwen38 timers] expert-read 3.906 s 3906.416 ms/forward\n[qwen38 timers] fp8-expand 0.000 s 0.000 ms/forward\n[qwen38 timers] routed-expert 1.933 s 1933.317 ms/forward\n[qwen38 timers] shared-expert 0.247 s 246.866 ms/forward\n[qwen38 timers] resident-mm 2.866 s 2865.993 ms/forward\n[qwen38 timers] deltanet 3.752 s 3751.728 ms/forward\n[qwen38 timers] qsa-index 0.011 s 10.927 ms/forward\n[qwen38 timers] qsa-attn 0.224 s 224.093 ms/forward\n[qwen38 timers] ple 0.650 s 649.645 ms/forward\n[qwen38 timers] lm-head 0.012 s 12.328 ms/forward\n[qwen38 timers] architecture phases overlap resident-mm; expert-read is disk service and fp8-expand is synchronous miss work\n[qwen38 timers] decode: 127 forwards\n[qwen38 timers] expert-read 4.381 s 34.500 ms/forward\n[qwen38 timers] fp8-expand 0.000 s 0.000 ms/forward\n[qwen38 timers] routed-expert 3.613 s 28.451 ms/forward\n[qwen38 timers] shared-expert 0.807 s 6.355 ms/forward\n[qwen38 timers] resident-mm 8.771 s 69.062 ms/forward\n[qwen38 timers] deltanet 6.491 s 51.111 ms/forward\n[qwen38 timers] qsa-index 0.100 s 0.784 ms/forward\n[qwen38 timers] qsa-attn 1.859 s 14.638 ms/forward\n[qwen38 timers] ple 0.382 s 3.005 ms/forward\n[qwen38 timers] lm-head 1.346 s 10.601 ms/forward\n[qwen38 timers] architecture phases overlap resident-mm; expert-read is disk service and fp8-expand is synchronous miss work\n[qwen38 expert I/O] weight-ranges=18554 scale-ranges=0 coalesced-gate-up=0 prefetched=16815 parallel-batches=2146 resident-scales=0.00 MiB\n\nPEAK RSS: 28.44 GB\nExpert cache hit rate: 69.0% (hit=41260 miss=18554)\nSpeed: 2.93 tok/s (43.7s for 128 tokens)\n[VK] qwen38: 0 matmuls on the GPU (0 matrices resident, 0.0 MiB; placed int8 0, bf16 0, f32 0)\n[VK] tier qwen38 run: device 89124 of 176640 routed experts (50.5%; this run 89124 of 176640) | CPU RAM hits 41260, disk loads 18554 | resident 2770 (budget 2770, 7.30 GiB of 7.31 GiB, 30 blocks, frag 0.42) | uploads 4866 (11.14 GiB, 2770 warm), evictions 2096, skipped 145 queue + 27546 rate, failed 0 | device 7178.6 ms, CPU share 15258.8 ms, waited 1275.3 ms (82% of device time hidden) | balance: device share 0.29, 4868 rows handed to the CPU | stream: 48 steps, 1739 cold experts (4.48 GiB, 10.42 GB/s) and 61912 rows streamed in 85 sub-batches, prefetched 0 (0 used), 9443 cold experts (38884 rows) kept on the CPU | cooperative matmuls: 612\n[STATS] 335040 selections across 20176 distinct experts -> /root/colibri-test/colibri-pr-validation-20261005/results/q38-i4g64-mtp-baseline-1/run.usage\n"
+ },
+ {
+ "tag": "q38-i4g64-mtp-baseline-2",
+ "family": "qwen38",
+ "model": "/root/colibri-test/models/qwen38-flash-next-fp8",
+ "prompt": "/root/colibri-test/p4mtp/prompts/code.txt",
+ "new": 128,
+ "cap": 170,
+ "cold": true,
+ "env": {
+ "COLI_VULKAN": "1",
+ "COLI_VK_CHAIN": "0",
+ "COLI_VK_DENSE": "0",
+ "Q38_EXPERT_INT4": "1",
+ "Q38_MTP": "0",
+ "Q38_MTP_DRAFTS": "3"
+ },
+ "kind": "generate",
+ "prompt_sha256": "b840c508d4d473840929f53cc370614e1438c6d0fd5f9afc67176ce1d99dfc2f",
+ "history_sha256": "d7d86f69d2c83d900c7d869b7a04e34be19a4f466eb40ec7c52deb70d50fc108",
+ "cmd": [
+ "/root/colibri-test/colibri-pr-validation-20261005/src/c/qwen38",
+ "170",
+ "8",
+ "/root/colibri-test/p4mtp/prompts/code.txt"
+ ],
+ "settings": {
+ "OMP_NUM_THREADS": "8",
+ "COLI_NO_OMP_TUNE": "1",
+ "COLI_TIMERS": "1",
+ "VK_ICD_FILENAMES": "/usr/share/vulkan/icd.d/radeon_icd.json",
+ "COLI_LOOKUP": "0",
+ "Q38_MTP": "0",
+ "COLI_VULKAN": "1",
+ "COLI_VK_CHAIN": "0",
+ "COLI_VK_DENSE": "0",
+ "Q38_EXPERT_INT4": "1",
+ "Q38_MTP_DRAFTS": "3",
+ "SNAP": "/root/colibri-test/models/qwen38-flash-next-fp8",
+ "COLI_USAGE": "/root/colibri-test/colibri-pr-validation-20261005/results/q38-i4g64-mtp-baseline-2/run.usage",
+ "N_NEW": "128"
+ },
+ "load_start": [
+ 1.62109375,
+ 1.81298828125,
+ 1.4033203125
+ ],
+ "source": "2f33f211ea6e28186d62cf9c00d511df826590a0",
+ "rc": 0,
+ "wall_s": 54.304,
+ "peak_rss_kib": 29819676,
+ "other_engines": [],
+ "stdout": "```c\nstatic int count_words(const char *s) {\n int total = 0, in_word = 0;\n for (const char *p = s; *p; p++) {\n if (*p == ' ' || *p == '\\n' || *p == '\\t') {\n in_word = 0;\n } else if (!in_word) {\n in_word = 1;\n total++;\n }\n }\n return total;\n}\n\nstatic void print_counts(const char **lines, int n) {\n int total",
+ "stderr": "== Qwen3.8-Flash-Next native text engine | cache=170/layer | CPU ==\n[tok] loaded 248044 base pieces + 33 added tokens (max id 248076) from /root/colibri-test/models/qwen38-flash-next-fp8/tokenizer.json\n[enc] prompt tokens: 241 | generating 128 new tokens\n[qwen38] vision tower: 27 blocks, hidden 1152, 16 heads, patch 16, merge 2\n[qwen38] native text weights: prefix=model.language_model, 48 layers, PLE=1, cache=170/layer, FP8=native, BF16=native (resident matrices 10.88 GiB)\n[qwen38] trunk: 553 matrices int8 on the CPU (3.97 GiB, 7.93 GiB of BF16 released) in 1.9s; Q38_TRUNK_CPU_INT8=0 keeps BF16\n[qwen38] routed experts: int4-g64 from experts-int4g64/, 2764800 bytes (2.64 MiB) per expert (the snapshot's native FP8: 4915200 bytes); cache 170/layer x 48 layers = 21.01 GiB full; 56.1 GB available now holds 422/layer\n[VK] VRAM pressure-proofing: memory_priority on, memory_budget on\n[VK] ready: AMD Radeon 780M Graphics (RADV PHOENIX), compute qfam 0, memtype 3, fused gate+up, absorb attention, tiled GEMM from S=2 and S*O>=4096 (cooperative matrix)\n[VK] qwen38: device ready, dense matrices on the CPU (COLI_VK_DENSE=0)\n[VK] qwen38: dense chain off (COLI_VK_CHAIN=0)\n[VK] qwen38: dense weights on the device and in host RAM (the dense part runs on the CPU)\n[USAGE] expert history: 158400 selections (/root/colibri-test/colibri-pr-validation-20261005/results/q38-i4g64-mtp-baseline-2/run.usage)\n[VK] tier qwen38: on, AMD Radeon 780M Graphics (RADV PHOENIX), budget 7.31 GiB = 2773 experts of 2.6 MiB (fmt 4 gs 64, down fmt 4 gs 64), shared RAM: a quarter of what the expert cache leaves, own queue, up to 16 promotions per token, balanced against the CPU, streaming cold experts of prompt steps through 64 slots\n[VK] tier qwen38: warm start, 2773 experts from the history in 2.0s\nresident weights loaded in 4.5s | RSS after load: 11.61 GB\nGenerated (128 new tokens):\nText : [VK] tier qwen38 stream: a step of 241 rows; cold experts with 16 rows or more go to the device (the GEMM's 16 rows: the upload and the CPU's rows not measured yet); layer 0: 61 resident, 31 streamed, 289 kept on the CPU (1475 rows)\n\nTTFT: 16.75 s (time to first token)\n[qwen38 timers] total: 128 forwards\n[qwen38 timers] expert-read 8.274 s 64.640 ms/forward\n[qwen38 timers] fp8-expand 0.000 s 0.000 ms/forward\n[qwen38 timers] routed-expert 5.566 s 43.483 ms/forward\n[qwen38 timers] shared-expert 1.046 s 8.173 ms/forward\n[qwen38 timers] resident-mm 11.654 s 91.049 ms/forward\n[qwen38 timers] deltanet 10.215 s 79.802 ms/forward\n[qwen38 timers] qsa-index 0.110 s 0.857 ms/forward\n[qwen38 timers] qsa-attn 2.075 s 16.211 ms/forward\n[qwen38 timers] ple 1.044 s 8.160 ms/forward\n[qwen38 timers] lm-head 1.359 s 10.616 ms/forward\n[qwen38 timers] architecture phases overlap resident-mm; expert-read is disk service and fp8-expand is synchronous miss work\n[qwen38 timers] prefill: 1 forwards\n[qwen38 timers] expert-read 3.952 s 3952.437 ms/forward\n[qwen38 timers] fp8-expand 0.000 s 0.000 ms/forward\n[qwen38 timers] routed-expert 1.935 s 1934.785 ms/forward\n[qwen38 timers] shared-expert 0.238 s 238.284 ms/forward\n[qwen38 timers] resident-mm 2.890 s 2889.701 ms/forward\n[qwen38 timers] deltanet 3.746 s 3745.537 ms/forward\n[qwen38 timers] qsa-index 0.011 s 11.000 ms/forward\n[qwen38 timers] qsa-attn 0.225 s 224.504 ms/forward\n[qwen38 timers] ple 0.656 s 655.910 ms/forward\n[qwen38 timers] lm-head 0.012 s 12.029 ms/forward\n[qwen38 timers] architecture phases overlap resident-mm; expert-read is disk service and fp8-expand is synchronous miss work\n[qwen38 timers] decode: 127 forwards\n[qwen38 timers] expert-read 4.321 s 34.027 ms/forward\n[qwen38 timers] fp8-expand 0.000 s 0.000 ms/forward\n[qwen38 timers] routed-expert 3.631 s 28.591 ms/forward\n[qwen38 timers] shared-expert 0.808 s 6.362 ms/forward\n[qwen38 timers] resident-mm 8.765 s 69.012 ms/forward\n[qwen38 timers] deltanet 6.469 s 50.937 ms/forward\n[qwen38 timers] qsa-index 0.099 s 0.777 ms/forward\n[qwen38 timers] qsa-attn 1.851 s 14.571 ms/forward\n[qwen38 timers] ple 0.389 s 3.060 ms/forward\n[qwen38 timers] lm-head 1.347 s 10.605 ms/forward\n[qwen38 timers] architecture phases overlap resident-mm; expert-read is disk service and fp8-expand is synchronous miss work\n[qwen38 expert I/O] weight-ranges=18409 scale-ranges=0 coalesced-gate-up=0 prefetched=16662 parallel-batches=2080 resident-scales=0.00 MiB\n\nPEAK RSS: 28.44 GB\nExpert cache hit rate: 69.3% (hit=41476 miss=18409)\nSpeed: 2.93 tok/s (43.8s for 128 tokens)\n[VK] qwen38: 0 matmuls on the GPU (0 matrices resident, 0.0 MiB; placed int8 0, bf16 0, f32 0)\n[VK] tier qwen38 run: device 89191 of 176640 routed experts (50.5%; this run 89191 of 176640) | CPU RAM hits 41476, disk loads 18409 | resident 2773 (budget 2773, 7.31 GiB of 7.31 GiB, 30 blocks, frag 0.47) | uploads 4869 (11.14 GiB, 2773 warm), evictions 2096, skipped 146 queue + 27747 rate, failed 0 | device 7181.5 ms, CPU share 15241.2 ms, waited 1284.8 ms (82% of device time hidden) | balance: device share 0.17, 4962 rows handed to the CPU | stream: 48 steps, 1747 cold experts (4.50 GiB, 10.37 GB/s) and 62038 rows streamed in 84 sub-batches, prefetched 0 (0 used), 9438 cold experts (38749 rows) kept on the CPU | cooperative matmuls: 615\n[STATS] 335040 selections across 20115 distinct experts -> /root/colibri-test/colibri-pr-validation-20261005/results/q38-i4g64-mtp-baseline-2/run.usage\n"
+ },
+ {
+ "tag": "q38-i4g64-mtp-baseline-3",
+ "family": "qwen38",
+ "model": "/root/colibri-test/models/qwen38-flash-next-fp8",
+ "prompt": "/root/colibri-test/p4mtp/prompts/code.txt",
+ "new": 128,
+ "cap": 170,
+ "cold": true,
+ "env": {
+ "COLI_VULKAN": "1",
+ "COLI_VK_CHAIN": "0",
+ "COLI_VK_DENSE": "0",
+ "Q38_EXPERT_INT4": "1",
+ "Q38_MTP": "0",
+ "Q38_MTP_DRAFTS": "3"
+ },
+ "kind": "generate",
+ "prompt_sha256": "b840c508d4d473840929f53cc370614e1438c6d0fd5f9afc67176ce1d99dfc2f",
+ "history_sha256": "d7d86f69d2c83d900c7d869b7a04e34be19a4f466eb40ec7c52deb70d50fc108",
+ "cmd": [
+ "/root/colibri-test/colibri-pr-validation-20261005/src/c/qwen38",
+ "170",
+ "8",
+ "/root/colibri-test/p4mtp/prompts/code.txt"
+ ],
+ "settings": {
+ "OMP_NUM_THREADS": "8",
+ "COLI_NO_OMP_TUNE": "1",
+ "COLI_TIMERS": "1",
+ "VK_ICD_FILENAMES": "/usr/share/vulkan/icd.d/radeon_icd.json",
+ "COLI_LOOKUP": "0",
+ "Q38_MTP": "0",
+ "COLI_VULKAN": "1",
+ "COLI_VK_CHAIN": "0",
+ "COLI_VK_DENSE": "0",
+ "Q38_EXPERT_INT4": "1",
+ "Q38_MTP_DRAFTS": "3",
+ "SNAP": "/root/colibri-test/models/qwen38-flash-next-fp8",
+ "COLI_USAGE": "/root/colibri-test/colibri-pr-validation-20261005/results/q38-i4g64-mtp-baseline-3/run.usage",
+ "N_NEW": "128"
+ },
+ "load_start": [
+ 1.56982421875,
+ 2.05029296875,
+ 1.5576171875
+ ],
+ "source": "2f33f211ea6e28186d62cf9c00d511df826590a0",
+ "rc": 0,
+ "wall_s": 55.305,
+ "peak_rss_kib": 29822384,
+ "other_engines": [],
+ "stdout": "```c\nstatic int count_words(const char *s) {\n int total = 0, in_word = 0;\n for (const char *p = s; *p; p++) {\n if (*p == ' ' || *p == '\\n' || *p == '\\t') {\n in_word = 0;\n } else if (!in_word) {\n in_word = 1;\n total++;\n }\n }\n return total;\n}\n\nstatic void print_counts(const char **lines, int n) {\n int total",
+ "stderr": "== Qwen3.8-Flash-Next native text engine | cache=170/layer | CPU ==\n[tok] loaded 248044 base pieces + 33 added tokens (max id 248076) from /root/colibri-test/models/qwen38-flash-next-fp8/tokenizer.json\n[enc] prompt tokens: 241 | generating 128 new tokens\n[qwen38] vision tower: 27 blocks, hidden 1152, 16 heads, patch 16, merge 2\n[qwen38] native text weights: prefix=model.language_model, 48 layers, PLE=1, cache=170/layer, FP8=native, BF16=native (resident matrices 10.88 GiB)\n[qwen38] trunk: 553 matrices int8 on the CPU (3.97 GiB, 7.93 GiB of BF16 released) in 1.9s; Q38_TRUNK_CPU_INT8=0 keeps BF16\n[qwen38] routed experts: int4-g64 from experts-int4g64/, 2764800 bytes (2.64 MiB) per expert (the snapshot's native FP8: 4915200 bytes); cache 170/layer x 48 layers = 21.01 GiB full; 56.1 GB available now holds 422/layer\n[VK] VRAM pressure-proofing: memory_priority on, memory_budget on\n[VK] ready: AMD Radeon 780M Graphics (RADV PHOENIX), compute qfam 0, memtype 3, fused gate+up, absorb attention, tiled GEMM from S=2 and S*O>=4096 (cooperative matrix)\n[VK] qwen38: device ready, dense matrices on the CPU (COLI_VK_DENSE=0)\n[VK] qwen38: dense chain off (COLI_VK_CHAIN=0)\n[VK] qwen38: dense weights on the device and in host RAM (the dense part runs on the CPU)\n[USAGE] expert history: 158400 selections (/root/colibri-test/colibri-pr-validation-20261005/results/q38-i4g64-mtp-baseline-3/run.usage)\n[VK] tier qwen38: on, AMD Radeon 780M Graphics (RADV PHOENIX), budget 7.30 GiB = 2769 experts of 2.6 MiB (fmt 4 gs 64, down fmt 4 gs 64), shared RAM: a quarter of what the expert cache leaves, own queue, up to 16 promotions per token, balanced against the CPU, streaming cold experts of prompt steps through 64 slots\n[VK] tier qwen38: warm start, 2769 experts from the history in 1.9s\nresident weights loaded in 4.5s | RSS after load: 11.61 GB\nGenerated (128 new tokens):\nText : [VK] tier qwen38 stream: a step of 241 rows; cold experts with 16 rows or more go to the device (the GEMM's 16 rows: the upload and the CPU's rows not measured yet); layer 0: 61 resident, 31 streamed, 289 kept on the CPU (1475 rows)\n\nTTFT: 17.01 s (time to first token)\n[qwen38 timers] total: 128 forwards\n[qwen38 timers] expert-read 8.387 s 65.527 ms/forward\n[qwen38 timers] fp8-expand 0.000 s 0.000 ms/forward\n[qwen38 timers] routed-expert 5.675 s 44.336 ms/forward\n[qwen38 timers] shared-expert 1.057 s 8.259 ms/forward\n[qwen38 timers] resident-mm 11.740 s 91.722 ms/forward\n[qwen38 timers] deltanet 10.318 s 80.606 ms/forward\n[qwen38 timers] qsa-index 0.113 s 0.880 ms/forward\n[qwen38 timers] qsa-attn 2.071 s 16.178 ms/forward\n[qwen38 timers] ple 1.038 s 8.109 ms/forward\n[qwen38 timers] lm-head 1.362 s 10.642 ms/forward\n[qwen38 timers] architecture phases overlap resident-mm; expert-read is disk service and fp8-expand is synchronous miss work\n[qwen38 timers] prefill: 1 forwards\n[qwen38 timers] expert-read 4.001 s 4001.120 ms/forward\n[qwen38 timers] fp8-expand 0.000 s 0.000 ms/forward\n[qwen38 timers] routed-expert 2.032 s 2031.749 ms/forward\n[qwen38 timers] shared-expert 0.255 s 255.384 ms/forward\n[qwen38 timers] resident-mm 2.932 s 2931.647 ms/forward\n[qwen38 timers] deltanet 3.777 s 3777.196 ms/forward\n[qwen38 timers] qsa-index 0.011 s 10.846 ms/forward\n[qwen38 timers] qsa-attn 0.224 s 223.894 ms/forward\n[qwen38 timers] ple 0.654 s 653.562 ms/forward\n[qwen38 timers] lm-head 0.012 s 12.109 ms/forward\n[qwen38 timers] architecture phases overlap resident-mm; expert-read is disk service and fp8-expand is synchronous miss work\n[qwen38 timers] decode: 127 forwards\n[qwen38 timers] expert-read 4.386 s 34.538 ms/forward\n[qwen38 timers] fp8-expand 0.000 s 0.000 ms/forward\n[qwen38 timers] routed-expert 3.643 s 28.687 ms/forward\n[qwen38 timers] shared-expert 0.802 s 6.313 ms/forward\n[qwen38 timers] resident-mm 8.809 s 69.360 ms/forward\n[qwen38 timers] deltanet 6.540 s 51.499 ms/forward\n[qwen38 timers] qsa-index 0.102 s 0.802 ms/forward\n[qwen38 timers] qsa-attn 1.847 s 14.542 ms/forward\n[qwen38 timers] ple 0.384 s 3.027 ms/forward\n[qwen38 timers] lm-head 1.350 s 10.631 ms/forward\n[qwen38 timers] architecture phases overlap resident-mm; expert-read is disk service and fp8-expand is synchronous miss work\n[qwen38 expert I/O] weight-ranges=18473 scale-ranges=0 coalesced-gate-up=0 prefetched=16734 parallel-batches=2112 resident-scales=0.00 MiB\n\nPEAK RSS: 28.44 GB\nExpert cache hit rate: 69.1% (hit=41329 miss=18473)\nSpeed: 2.90 tok/s (44.2s for 128 tokens)\n[VK] qwen38: 0 matmuls on the GPU (0 matrices resident, 0.0 MiB; placed int8 0, bf16 0, f32 0)\n[VK] tier qwen38 run: device 89136 of 176640 routed experts (50.5%; this run 89136 of 176640) | CPU RAM hits 41329, disk loads 18473 | resident 2769 (budget 2769, 7.29 GiB of 7.30 GiB, 30 blocks, frag 0.70) | uploads 4865 (11.14 GiB, 2769 warm), evictions 2096, skipped 145 queue + 27474 rate, failed 0 | device 7192.9 ms, CPU share 15451.7 ms, waited 1292.4 ms (82% of device time hidden) | balance: device share 0.24, 4963 rows handed to the CPU | stream: 48 steps, 1739 cold experts (4.48 GiB, 10.34 GB/s) and 61912 rows streamed in 85 sub-batches, prefetched 0 (0 used), 9443 cold experts (38884 rows) kept on the CPU | cooperative matmuls: 612\n[STATS] 335040 selections across 20154 distinct experts -> /root/colibri-test/colibri-pr-validation-20261005/results/q38-i4g64-mtp-baseline-3/run.usage\n"
+ }
+ ]
+}
diff --git a/docs/experiments/vulkan-qwen38-mtp-2026-10-05/history.usage.gz b/docs/experiments/vulkan-qwen38-mtp-2026-10-05/history.usage.gz
new file mode 100644
index 00000000..c98294a1
Binary files /dev/null and b/docs/experiments/vulkan-qwen38-mtp-2026-10-05/history.usage.gz differ
diff --git a/docs/experiments/vulkan-qwen38-mtp-2026-10-05/manifest.json b/docs/experiments/vulkan-qwen38-mtp-2026-10-05/manifest.json
new file mode 100644
index 00000000..99013ca6
--- /dev/null
+++ b/docs/experiments/vulkan-qwen38-mtp-2026-10-05/manifest.json
@@ -0,0 +1,92 @@
+{
+ "version": 1,
+ "hypothesis": "On the existing int4-g64 Qwen3.8 checkpoint, enabling three MTP drafts reduces prompt-plus-generation time for a code-edit request with an unchanged expert-tier configuration.",
+ "commit": "2f33f211ea6e28186d62cf9c00d511df826590a0",
+ "model": "Qwen3.8-Flash-Next native snapshot with existing experts-int4g64 sidecar; MTP experts remain native FP8; default CPU int8 trunk.",
+ "prompt_hash": "b840c508d4d473840929f53cc370614e1438c6d0fd5f9afc67176ce1d99dfc2f",
+ "history_hash": "d7d86f69d2c83d900c7d869b7a04e34be19a4f466eb40ec7c52deb70d50fc108",
+ "hardware": {
+ "cpu": "AMD Ryzen 7 PRO 8700GE w/ Radeon 780M Graphics",
+ "ram": "MemTotal: 64897288 kB",
+ "storage": "NAME MODEL SIZE ROTA\nnvme0n1 SAMSUNG MZVL2512HCJQ-00B07 476.9G 0\nnvme1n1 SAMSUNG MZVL2512HCJQ-00B07 476.9G 0\next4 /dev/md2 rw,relatime,stripe=256\n",
+ "os": "Linux 7.0.0-22-generic; gcc (Ubuntu 15.2.0-16ubuntu1) 15.2.0; PRETTY_NAME=\"Ubuntu 26.04 LTS\"\nNAME=\"Ubuntu\"\nVERSION_ID=\"26.04\"\nVERSION=\"26.04 LTS (Resolute Raccoon)\"\nVERSION_CODENAME=resolute\nID=ubuntu\nID_LIKE=debian\nHOME_URL=\"https://www.ubuntu.com/\"\nSUPPORT_URL=\"https://help.ubuntu.com/\"\nBUG_REPORT_URL=\"https://bugs.launchpad.net/ubuntu/\"\nPRIVACY_POLICY_URL=\"https://www.ubuntu.com/legal/terms-and-policies/privacy-policy\"\nUBUNTU_CODENAME=resolute\nLOGO=ubuntu-logo\n\nGPU: AMD Radeon 780M Graphics (RADV PHOENIX), Vulkan; VK_ICD_FILENAMES=/usr/share/vulkan/icd.d/radeon_icd.json"
+ },
+ "warmup_runs": 0,
+ "changed_variables": [
+ "Q38_MTP"
+ ],
+ "metric": "Engine-reported tok/s: 128 generated tokens divided by prompt (241 tokens) plus generation time, excluding model loading. Not steady decode throughput.",
+ "protocol": "Three runs per arm, alternating order A/B, B/A, A/B; eight threads and cap 170. Before each run decompress the attached history.usage.gz to COLI_USAGE, advise checkpoint file pages out using POSIX_FADV_DONTNEED, and wait for three idle checks five seconds apart (load average below 2, no other inference). Advisory eviction does not guarantee physically cold storage. No global page-cache flush.",
+ "command": "make -j4 -C c qwen38 VK=1\nenv OMP_NUM_THREADS=8 COLI_NO_OMP_TUNE=1 COLI_TIMERS=1 VK_ICD_FILENAMES=/usr/share/vulkan/icd.d/radeon_icd.json COLI_LOOKUP=0 Q38_MTP=0 COLI_VULKAN=1 COLI_VK_CHAIN=0 COLI_VK_DENSE=0 Q38_EXPERT_INT4=1 Q38_MTP_DRAFTS=3 SNAP=/root/colibri-test/models/qwen38-flash-next-fp8 COLI_USAGE=/root/colibri-test/colibri-pr-validation-20261005/results/q38-i4g64-mtp-baseline-1/run.usage N_NEW=128 /root/colibri-test/colibri-pr-validation-20261005/src/c/qwen38 170 8 /root/colibri-test/p4mtp/prompts/code.txt\nenv OMP_NUM_THREADS=8 COLI_NO_OMP_TUNE=1 COLI_TIMERS=1 VK_ICD_FILENAMES=/usr/share/vulkan/icd.d/radeon_icd.json COLI_LOOKUP=0 Q38_MTP=0 COLI_VULKAN=1 COLI_VK_CHAIN=0 COLI_VK_DENSE=0 Q38_EXPERT_INT4=1 Q38_MTP_DRAFTS=3 SNAP=/root/colibri-test/models/qwen38-flash-next-fp8 COLI_USAGE=/root/colibri-test/colibri-pr-validation-20261005/results/q38-i4g64-mtp-baseline-2/run.usage N_NEW=128 /root/colibri-test/colibri-pr-validation-20261005/src/c/qwen38 170 8 /root/colibri-test/p4mtp/prompts/code.txt\nenv OMP_NUM_THREADS=8 COLI_NO_OMP_TUNE=1 COLI_TIMERS=1 VK_ICD_FILENAMES=/usr/share/vulkan/icd.d/radeon_icd.json COLI_LOOKUP=0 Q38_MTP=0 COLI_VULKAN=1 COLI_VK_CHAIN=0 COLI_VK_DENSE=0 Q38_EXPERT_INT4=1 Q38_MTP_DRAFTS=3 SNAP=/root/colibri-test/models/qwen38-flash-next-fp8 COLI_USAGE=/root/colibri-test/colibri-pr-validation-20261005/results/q38-i4g64-mtp-baseline-3/run.usage N_NEW=128 /root/colibri-test/colibri-pr-validation-20261005/src/c/qwen38 170 8 /root/colibri-test/p4mtp/prompts/code.txt\nenv OMP_NUM_THREADS=8 COLI_NO_OMP_TUNE=1 COLI_TIMERS=1 VK_ICD_FILENAMES=/usr/share/vulkan/icd.d/radeon_icd.json COLI_LOOKUP=0 Q38_MTP=1 COLI_VULKAN=1 COLI_VK_CHAIN=0 COLI_VK_DENSE=0 Q38_EXPERT_INT4=1 Q38_MTP_DRAFTS=3 SNAP=/root/colibri-test/models/qwen38-flash-next-fp8 COLI_USAGE=/root/colibri-test/colibri-pr-validation-20261005/results/q38-i4g64-mtp-trial-1/run.usage N_NEW=128 /root/colibri-test/colibri-pr-validation-20261005/src/c/qwen38 170 8 /root/colibri-test/p4mtp/prompts/code.txt\nenv OMP_NUM_THREADS=8 COLI_NO_OMP_TUNE=1 COLI_TIMERS=1 VK_ICD_FILENAMES=/usr/share/vulkan/icd.d/radeon_icd.json COLI_LOOKUP=0 Q38_MTP=1 COLI_VULKAN=1 COLI_VK_CHAIN=0 COLI_VK_DENSE=0 Q38_EXPERT_INT4=1 Q38_MTP_DRAFTS=3 SNAP=/root/colibri-test/models/qwen38-flash-next-fp8 COLI_USAGE=/root/colibri-test/colibri-pr-validation-20261005/results/q38-i4g64-mtp-trial-2/run.usage N_NEW=128 /root/colibri-test/colibri-pr-validation-20261005/src/c/qwen38 170 8 /root/colibri-test/p4mtp/prompts/code.txt\nenv OMP_NUM_THREADS=8 COLI_NO_OMP_TUNE=1 COLI_TIMERS=1 VK_ICD_FILENAMES=/usr/share/vulkan/icd.d/radeon_icd.json COLI_LOOKUP=0 Q38_MTP=1 COLI_VULKAN=1 COLI_VK_CHAIN=0 COLI_VK_DENSE=0 Q38_EXPERT_INT4=1 Q38_MTP_DRAFTS=3 SNAP=/root/colibri-test/models/qwen38-flash-next-fp8 COLI_USAGE=/root/colibri-test/colibri-pr-validation-20261005/results/q38-i4g64-mtp-trial-3/run.usage N_NEW=128 /root/colibri-test/colibri-pr-validation-20261005/src/c/qwen38 170 8 /root/colibri-test/p4mtp/prompts/code.txt",
+ "baseline": {
+ "config": {
+ "OMP_NUM_THREADS": "8",
+ "COLI_NO_OMP_TUNE": "1",
+ "COLI_TIMERS": "1",
+ "VK_ICD_FILENAMES": "/usr/share/vulkan/icd.d/radeon_icd.json",
+ "COLI_LOOKUP": "0",
+ "Q38_MTP": "0",
+ "COLI_VULKAN": "1",
+ "COLI_VK_CHAIN": "0",
+ "COLI_VK_DENSE": "0",
+ "Q38_EXPERT_INT4": "1",
+ "Q38_MTP_DRAFTS": "3",
+ "SNAP": "/root/colibri-test/models/qwen38-flash-next-fp8",
+ "COLI_USAGE": "fresh decompressed copy of attached history.usage.gz",
+ "N_NEW": "128"
+ },
+ "samples": {
+ "tok_s": [
+ 2.93,
+ 2.93,
+ 2.9
+ ]
+ },
+ "median_tok_s": 2.93,
+ "quality": {
+ "method": "All six generated texts are byte-identical; SHA-256 3ff36e598e3e0651a0c9349193df0c47804fbb8bbc54826ea935ec090bb52b9e",
+ "passed": true
+ },
+ "evidence": {
+ "uri": "docs/experiments/vulkan-qwen38-mtp-2026-10-05/baseline.json",
+ "sha256": "9cb9ab1dd599dee0540ad2a07aa31b7e9109524e96f49d9ec88eaf2fa9a8ab80"
+ }
+ },
+ "trial": {
+ "config": {
+ "OMP_NUM_THREADS": "8",
+ "COLI_NO_OMP_TUNE": "1",
+ "COLI_TIMERS": "1",
+ "VK_ICD_FILENAMES": "/usr/share/vulkan/icd.d/radeon_icd.json",
+ "COLI_LOOKUP": "0",
+ "Q38_MTP": "1",
+ "COLI_VULKAN": "1",
+ "COLI_VK_CHAIN": "0",
+ "COLI_VK_DENSE": "0",
+ "Q38_EXPERT_INT4": "1",
+ "Q38_MTP_DRAFTS": "3",
+ "SNAP": "/root/colibri-test/models/qwen38-flash-next-fp8",
+ "COLI_USAGE": "fresh decompressed copy of attached history.usage.gz",
+ "N_NEW": "128"
+ },
+ "samples": {
+ "tok_s": [
+ 3.49,
+ 3.5,
+ 3.47
+ ]
+ },
+ "median_tok_s": 3.49,
+ "quality": {
+ "method": "All six generated texts are byte-identical; SHA-256 3ff36e598e3e0651a0c9349193df0c47804fbb8bbc54826ea935ec090bb52b9e",
+ "passed": true
+ },
+ "evidence": {
+ "uri": "docs/experiments/vulkan-qwen38-mtp-2026-10-05/trial.json",
+ "sha256": "496040ccc2209b2645cd75922a3069ffdc6e717bd5f7ef330b608041e6615aa6"
+ }
+ },
+ "outcome": "improvement",
+ "history_compression": "gzip; history_hash is SHA-256 of the decompressed bytes",
+ "prompt_encoding": "prompt.json text field, encoded as UTF-8; preserves the original trailing newlines"
+}
diff --git a/docs/experiments/vulkan-qwen38-mtp-2026-10-05/prompt.json b/docs/experiments/vulkan-qwen38-mtp-2026-10-05/prompt.json
new file mode 100644
index 00000000..1ee8505f
--- /dev/null
+++ b/docs/experiments/vulkan-qwen38-mtp-2026-10-05/prompt.json
@@ -0,0 +1,4 @@
+{
+ "text": "<|im_start|>user\nRename the variable `count` to `total` in this C code and print the whole code again, with no explanation.\n\n```c\nstatic int count_words(const char *s) {\n int count = 0, in_word = 0;\n for (const char *p = s; *p; p++) {\n if (*p == ' ' || *p == '\\n' || *p == '\\t') {\n in_word = 0;\n } else if (!in_word) {\n in_word = 1;\n count++;\n }\n }\n return count;\n}\n\nstatic void print_counts(const char **lines, int n) {\n int count = 0;\n for (int i = 0; i < n; i++) {\n int words = count_words(lines[i]);\n printf(\"line %d: %d words\\n\", i, words);\n count += words;\n }\n printf(\"all lines: %d words\\n\", count);\n}\n```<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n",
+ "sha256": "b840c508d4d473840929f53cc370614e1438c6d0fd5f9afc67176ce1d99dfc2f"
+}
diff --git a/docs/experiments/vulkan-qwen38-mtp-2026-10-05/trial.json b/docs/experiments/vulkan-qwen38-mtp-2026-10-05/trial.json
new file mode 100644
index 00000000..0a2363b5
--- /dev/null
+++ b/docs/experiments/vulkan-qwen38-mtp-2026-10-05/trial.json
@@ -0,0 +1,166 @@
+{
+ "runs": [
+ {
+ "tag": "q38-i4g64-mtp-trial-1",
+ "family": "qwen38",
+ "model": "/root/colibri-test/models/qwen38-flash-next-fp8",
+ "prompt": "/root/colibri-test/p4mtp/prompts/code.txt",
+ "new": 128,
+ "cap": 170,
+ "cold": true,
+ "env": {
+ "COLI_VULKAN": "1",
+ "COLI_VK_CHAIN": "0",
+ "COLI_VK_DENSE": "0",
+ "Q38_EXPERT_INT4": "1",
+ "Q38_MTP": "1",
+ "Q38_MTP_DRAFTS": "3"
+ },
+ "kind": "generate",
+ "prompt_sha256": "b840c508d4d473840929f53cc370614e1438c6d0fd5f9afc67176ce1d99dfc2f",
+ "history_sha256": "d7d86f69d2c83d900c7d869b7a04e34be19a4f466eb40ec7c52deb70d50fc108",
+ "cmd": [
+ "/root/colibri-test/colibri-pr-validation-20261005/src/c/qwen38",
+ "170",
+ "8",
+ "/root/colibri-test/p4mtp/prompts/code.txt"
+ ],
+ "settings": {
+ "OMP_NUM_THREADS": "8",
+ "COLI_NO_OMP_TUNE": "1",
+ "COLI_TIMERS": "1",
+ "VK_ICD_FILENAMES": "/usr/share/vulkan/icd.d/radeon_icd.json",
+ "COLI_LOOKUP": "0",
+ "Q38_MTP": "1",
+ "COLI_VULKAN": "1",
+ "COLI_VK_CHAIN": "0",
+ "COLI_VK_DENSE": "0",
+ "Q38_EXPERT_INT4": "1",
+ "Q38_MTP_DRAFTS": "3",
+ "SNAP": "/root/colibri-test/models/qwen38-flash-next-fp8",
+ "COLI_USAGE": "/root/colibri-test/colibri-pr-validation-20261005/results/q38-i4g64-mtp-trial-1/run.usage",
+ "N_NEW": "128"
+ },
+ "load_start": [
+ 1.64990234375,
+ 1.15234375,
+ 1.08544921875
+ ],
+ "source": "2f33f211ea6e28186d62cf9c00d511df826590a0",
+ "rc": 0,
+ "wall_s": 47.26,
+ "peak_rss_kib": 31081908,
+ "other_engines": [],
+ "stdout": "```c\nstatic int count_words(const char *s) {\n int total = 0, in_word = 0;\n for (const char *p = s; *p; p++) {\n if (*p == ' ' || *p == '\\n' || *p == '\\t') {\n in_word = 0;\n } else if (!in_word) {\n in_word = 1;\n total++;\n }\n }\n return total;\n}\n\nstatic void print_counts(const char **lines, int n) {\n int total",
+ "stderr": "== Qwen3.8-Flash-Next native text engine | cache=170/layer | CPU ==\n[tok] loaded 248044 base pieces + 33 added tokens (max id 248076) from /root/colibri-test/models/qwen38-flash-next-fp8/tokenizer.json\n[enc] prompt tokens: 241 | generating 128 new tokens\n[qwen38] vision tower: 27 blocks, hidden 1152, 16 heads, patch 16, merge 2\n[qwen38] native text weights: prefix=model.language_model, 48 layers, PLE=1, cache=170/layer, FP8=native, BF16=native (resident matrices 10.88 GiB)\n[qwen38] MTP head mtp.*: wiring b (a norm per stream, fc_hidden per stream), RoPE base 1e+07, 172.66 MiB resident; routed experts F8_E4M3 from the snapshot (the int4-g64 sidecar covers the model's layers), cache 170 (Q38_MTP_CAP) x 4.69 MiB = 0.78 GiB full\n[qwen38] MTP norm weights, mean: embedding -0.7642, hidden -0.3284, mixer 3.7881 (q38_rms0 scales by 1+w; a plain norm's weights would sit near 1)\n[qwen38] trunk: 570 matrices int8 on the CPU (4.06 GiB, 8.10 GiB of BF16 released) in 1.9s; Q38_TRUNK_CPU_INT8=0 keeps BF16\n[qwen38] routed experts: int4-g64 from experts-int4g64/, 2764800 bytes (2.64 MiB) per expert (the snapshot's native FP8: 4915200 bytes); cache 170/layer x 48 layers = 21.01 GiB full; 55.7 GB available now holds 419/layer\n[VK] VRAM pressure-proofing: memory_priority on, memory_budget on\n[VK] ready: AMD Radeon 780M Graphics (RADV PHOENIX), compute qfam 0, memtype 3, fused gate+up, absorb attention, tiled GEMM from S=2 and S*O>=4096 (cooperative matrix)\n[VK] qwen38: device ready, dense matrices on the CPU (COLI_VK_DENSE=0)\n[VK] qwen38: dense chain off (COLI_VK_CHAIN=0)\n[VK] qwen38: dense weights on the device and in host RAM (the dense part runs on the CPU)\n[USAGE] expert history: 158400 selections (/root/colibri-test/colibri-pr-validation-20261005/results/q38-i4g64-mtp-trial-1/run.usage)\n[VK] tier qwen38: on, AMD Radeon 780M Graphics (RADV PHOENIX), budget 7.21 GiB = 2731 experts of 2.6 MiB (fmt 4 gs 64, down fmt 4 gs 64), shared RAM: a quarter of what the expert cache leaves, own queue, up to 16 promotions per token, balanced against the CPU, streaming cold experts of prompt steps through 64 slots\n[VK] tier qwen38: warm start, 2731 experts from the history in 1.9s\nresident weights loaded in 4.5s | RSS after load: 11.77 GB\nGenerated (128 new tokens):\nText : [VK] tier qwen38 stream: a step of 241 rows; cold experts with 16 rows or more go to the device (the GEMM's 16 rows: the upload and the CPU's rows not measured yet); layer 0: 61 resident, 31 streamed, 289 kept on the CPU (1475 rows)\n\nTTFT: 17.52 s (time to first token)\n[qwen38 MTP] run: 3.97 tokens/forward (32 forwards per 127 tokens) | acceptance 100.0% (95/95 drafts) | wiring b | fixed depth, by position: 1 100.0% (32/32) 2 100.0% (32/32) 3 100.0% (31/31)\n[qwen38 timers] total: 33 forwards\n[qwen38 timers] expert-read 7.018 s 212.673 ms/forward\n[qwen38 timers] fp8-expand 0.000 s 0.000 ms/forward\n[qwen38 timers] routed-expert 5.342 s 161.884 ms/forward\n[qwen38 timers] shared-expert 0.508 s 15.402 ms/forward\n[qwen38 timers] resident-mm 6.120 s 185.459 ms/forward\n[qwen38 timers] deltanet 7.210 s 218.479 ms/forward\n[qwen38 timers] qsa-index 0.118 s 3.585 ms/forward\n[qwen38 timers] qsa-attn 2.318 s 70.249 ms/forward\n[qwen38 timers] ple 1.019 s 30.868 ms/forward\n[qwen38 timers] lm-head 1.399 s 42.405 ms/forward\n[qwen38 timers] architecture phases overlap resident-mm; expert-read is disk service and fp8-expand is synchronous miss work\n[qwen38 timers] prefill: 1 forwards\n[qwen38 timers] expert-read 4.150 s 4149.825 ms/forward\n[qwen38 timers] fp8-expand 0.000 s 0.000 ms/forward\n[qwen38 timers] routed-expert 2.115 s 2115.136 ms/forward\n[qwen38 timers] shared-expert 0.250 s 249.827 ms/forward\n[qwen38 timers] resident-mm 2.980 s 2979.585 ms/forward\n[qwen38 timers] deltanet 3.749 s 3748.699 ms/forward\n[qwen38 timers] qsa-index 0.013 s 13.140 ms/forward\n[qwen38 timers] qsa-attn 0.257 s 256.694 ms/forward\n[qwen38 timers] ple 0.652 s 651.836 ms/forward\n[qwen38 timers] lm-head 0.012 s 12.204 ms/forward\n[qwen38 timers] architecture phases overlap resident-mm; expert-read is disk service and fp8-expand is synchronous miss work\n[qwen38 timers] decode: 32 forwards\n[qwen38 timers] expert-read 2.868 s 89.637 ms/forward\n[qwen38 timers] fp8-expand 0.000 s 0.000 ms/forward\n[qwen38 timers] routed-expert 3.227 s 100.845 ms/forward\n[qwen38 timers] shared-expert 0.258 s 8.076 ms/forward\n[qwen38 timers] resident-mm 3.141 s 98.143 ms/forward\n[qwen38 timers] deltanet 3.461 s 108.160 ms/forward\n[qwen38 timers] qsa-index 0.105 s 3.286 ms/forward\n[qwen38 timers] qsa-attn 2.062 s 64.423 ms/forward\n[qwen38 timers] ple 0.367 s 11.462 ms/forward\n[qwen38 timers] lm-head 1.387 s 43.349 ms/forward\n[qwen38 timers] architecture phases overlap resident-mm; expert-read is disk service and fp8-expand is synchronous miss work\n[qwen38 expert I/O] weight-ranges=19173 scale-ranges=2 coalesced-gate-up=437 prefetched=17412 parallel-batches=1507 resident-scales=0.59 MiB\n\nPEAK RSS: 29.65 GB\nExpert cache hit rate: 62.9% (hit=31728 miss=18736)\nSpeed: 3.49 tok/s (36.7s for 128 tokens)\n[VK] qwen38: 0 matmuls on the GPU (0 matrices resident, 0.0 MiB; placed int8 0, bf16 0, f32 0)\n[VK] tier qwen38 run: device 89648 of 176640 routed experts (50.8%; this run 89648 of 176640) | CPU RAM hits 31728, disk loads 18736 | resident 2731 (budget 2731, 7.20 GiB of 7.21 GiB, 29 blocks, frag 0.80) | uploads 4827 (11.05 GiB, 2731 warm), evictions 2096, skipped 149 queue + 20140 rate, failed 0 | device 6536.3 ms, CPU share 15246.1 ms, waited 604.6 ms (91% of device time hidden) | balance: device share 0.41, 3467 rows handed to the CPU | stream: 48 steps, 1761 cold experts (4.53 GiB, 10.53 GB/s) and 62210 rows streamed in 88 sub-batches, prefetched 0 (0 used), 9432 cold experts (38689 rows) kept on the CPU | cooperative matmuls: 612\n[STATS] 339320 selections across 20430 distinct experts -> /root/colibri-test/colibri-pr-validation-20261005/results/q38-i4g64-mtp-trial-1/run.usage\n"
+ },
+ {
+ "tag": "q38-i4g64-mtp-trial-2",
+ "family": "qwen38",
+ "model": "/root/colibri-test/models/qwen38-flash-next-fp8",
+ "prompt": "/root/colibri-test/p4mtp/prompts/code.txt",
+ "new": 128,
+ "cap": 170,
+ "cold": true,
+ "env": {
+ "COLI_VULKAN": "1",
+ "COLI_VK_CHAIN": "0",
+ "COLI_VK_DENSE": "0",
+ "Q38_EXPERT_INT4": "1",
+ "Q38_MTP": "1",
+ "Q38_MTP_DRAFTS": "3"
+ },
+ "kind": "generate",
+ "prompt_sha256": "b840c508d4d473840929f53cc370614e1438c6d0fd5f9afc67176ce1d99dfc2f",
+ "history_sha256": "d7d86f69d2c83d900c7d869b7a04e34be19a4f466eb40ec7c52deb70d50fc108",
+ "cmd": [
+ "/root/colibri-test/colibri-pr-validation-20261005/src/c/qwen38",
+ "170",
+ "8",
+ "/root/colibri-test/p4mtp/prompts/code.txt"
+ ],
+ "settings": {
+ "OMP_NUM_THREADS": "8",
+ "COLI_NO_OMP_TUNE": "1",
+ "COLI_TIMERS": "1",
+ "VK_ICD_FILENAMES": "/usr/share/vulkan/icd.d/radeon_icd.json",
+ "COLI_LOOKUP": "0",
+ "Q38_MTP": "1",
+ "COLI_VULKAN": "1",
+ "COLI_VK_CHAIN": "0",
+ "COLI_VK_DENSE": "0",
+ "Q38_EXPERT_INT4": "1",
+ "Q38_MTP_DRAFTS": "3",
+ "SNAP": "/root/colibri-test/models/qwen38-flash-next-fp8",
+ "COLI_USAGE": "/root/colibri-test/colibri-pr-validation-20261005/results/q38-i4g64-mtp-trial-2/run.usage",
+ "N_NEW": "128"
+ },
+ "load_start": [
+ 1.6201171875,
+ 1.71484375,
+ 1.328125
+ ],
+ "source": "2f33f211ea6e28186d62cf9c00d511df826590a0",
+ "rc": 0,
+ "wall_s": 47.243,
+ "peak_rss_kib": 31081284,
+ "other_engines": [],
+ "stdout": "```c\nstatic int count_words(const char *s) {\n int total = 0, in_word = 0;\n for (const char *p = s; *p; p++) {\n if (*p == ' ' || *p == '\\n' || *p == '\\t') {\n in_word = 0;\n } else if (!in_word) {\n in_word = 1;\n total++;\n }\n }\n return total;\n}\n\nstatic void print_counts(const char **lines, int n) {\n int total",
+ "stderr": "== Qwen3.8-Flash-Next native text engine | cache=170/layer | CPU ==\n[tok] loaded 248044 base pieces + 33 added tokens (max id 248076) from /root/colibri-test/models/qwen38-flash-next-fp8/tokenizer.json\n[enc] prompt tokens: 241 | generating 128 new tokens\n[qwen38] vision tower: 27 blocks, hidden 1152, 16 heads, patch 16, merge 2\n[qwen38] native text weights: prefix=model.language_model, 48 layers, PLE=1, cache=170/layer, FP8=native, BF16=native (resident matrices 10.88 GiB)\n[qwen38] MTP head mtp.*: wiring b (a norm per stream, fc_hidden per stream), RoPE base 1e+07, 172.66 MiB resident; routed experts F8_E4M3 from the snapshot (the int4-g64 sidecar covers the model's layers), cache 170 (Q38_MTP_CAP) x 4.69 MiB = 0.78 GiB full\n[qwen38] MTP norm weights, mean: embedding -0.7642, hidden -0.3284, mixer 3.7881 (q38_rms0 scales by 1+w; a plain norm's weights would sit near 1)\n[qwen38] trunk: 570 matrices int8 on the CPU (4.06 GiB, 8.10 GiB of BF16 released) in 1.9s; Q38_TRUNK_CPU_INT8=0 keeps BF16\n[qwen38] routed experts: int4-g64 from experts-int4g64/, 2764800 bytes (2.64 MiB) per expert (the snapshot's native FP8: 4915200 bytes); cache 170/layer x 48 layers = 21.01 GiB full; 55.7 GB available now holds 419/layer\n[VK] VRAM pressure-proofing: memory_priority on, memory_budget on\n[VK] ready: AMD Radeon 780M Graphics (RADV PHOENIX), compute qfam 0, memtype 3, fused gate+up, absorb attention, tiled GEMM from S=2 and S*O>=4096 (cooperative matrix)\n[VK] qwen38: device ready, dense matrices on the CPU (COLI_VK_DENSE=0)\n[VK] qwen38: dense chain off (COLI_VK_CHAIN=0)\n[VK] qwen38: dense weights on the device and in host RAM (the dense part runs on the CPU)\n[USAGE] expert history: 158400 selections (/root/colibri-test/colibri-pr-validation-20261005/results/q38-i4g64-mtp-trial-2/run.usage)\n[VK] tier qwen38: on, AMD Radeon 780M Graphics (RADV PHOENIX), budget 7.21 GiB = 2732 experts of 2.6 MiB (fmt 4 gs 64, down fmt 4 gs 64), shared RAM: a quarter of what the expert cache leaves, own queue, up to 16 promotions per token, balanced against the CPU, streaming cold experts of prompt steps through 64 slots\n[VK] tier qwen38: warm start, 2732 experts from the history in 1.9s\nresident weights loaded in 4.5s | RSS after load: 11.77 GB\nGenerated (128 new tokens):\nText : [VK] tier qwen38 stream: a step of 241 rows; cold experts with 16 rows or more go to the device (the GEMM's 16 rows: the upload and the CPU's rows not measured yet); layer 0: 61 resident, 31 streamed, 289 kept on the CPU (1475 rows)\n\nTTFT: 17.46 s (time to first token)\n[qwen38 MTP] run: 3.97 tokens/forward (32 forwards per 127 tokens) | acceptance 100.0% (95/95 drafts) | wiring b | fixed depth, by position: 1 100.0% (32/32) 2 100.0% (32/32) 3 100.0% (31/31)\n[qwen38 timers] total: 33 forwards\n[qwen38 timers] expert-read 7.004 s 212.231 ms/forward\n[qwen38 timers] fp8-expand 0.000 s 0.000 ms/forward\n[qwen38 timers] routed-expert 5.258 s 159.341 ms/forward\n[qwen38 timers] shared-expert 0.511 s 15.485 ms/forward\n[qwen38 timers] resident-mm 6.124 s 185.562 ms/forward\n[qwen38 timers] deltanet 7.227 s 218.989 ms/forward\n[qwen38 timers] qsa-index 0.118 s 3.577 ms/forward\n[qwen38 timers] qsa-attn 2.316 s 70.187 ms/forward\n[qwen38 timers] ple 1.025 s 31.053 ms/forward\n[qwen38 timers] lm-head 1.387 s 42.040 ms/forward\n[qwen38 timers] architecture phases overlap resident-mm; expert-read is disk service and fp8-expand is synchronous miss work\n[qwen38 timers] prefill: 1 forwards\n[qwen38 timers] expert-read 4.104 s 4103.720 ms/forward\n[qwen38 timers] fp8-expand 0.000 s 0.000 ms/forward\n[qwen38 timers] routed-expert 2.059 s 2058.949 ms/forward\n[qwen38 timers] shared-expert 0.254 s 253.672 ms/forward\n[qwen38 timers] resident-mm 2.999 s 2998.691 ms/forward\n[qwen38 timers] deltanet 3.788 s 3787.908 ms/forward\n[qwen38 timers] qsa-index 0.013 s 12.840 ms/forward\n[qwen38 timers] qsa-attn 0.255 s 255.385 ms/forward\n[qwen38 timers] ple 0.657 s 657.305 ms/forward\n[qwen38 timers] lm-head 0.012 s 12.167 ms/forward\n[qwen38 timers] architecture phases overlap resident-mm; expert-read is disk service and fp8-expand is synchronous miss work\n[qwen38 timers] decode: 32 forwards\n[qwen38 timers] expert-read 2.900 s 90.622 ms/forward\n[qwen38 timers] fp8-expand 0.000 s 0.000 ms/forward\n[qwen38 timers] routed-expert 3.199 s 99.978 ms/forward\n[qwen38 timers] shared-expert 0.257 s 8.042 ms/forward\n[qwen38 timers] resident-mm 3.125 s 97.651 ms/forward\n[qwen38 timers] deltanet 3.439 s 107.460 ms/forward\n[qwen38 timers] qsa-index 0.105 s 3.287 ms/forward\n[qwen38 timers] qsa-attn 2.061 s 64.400 ms/forward\n[qwen38 timers] ple 0.367 s 11.483 ms/forward\n[qwen38 timers] lm-head 1.375 s 42.974 ms/forward\n[qwen38 timers] architecture phases overlap resident-mm; expert-read is disk service and fp8-expand is synchronous miss work\n[qwen38 expert I/O] weight-ranges=19216 scale-ranges=2 coalesced-gate-up=435 prefetched=17460 parallel-batches=1531 resident-scales=0.59 MiB\n\nPEAK RSS: 29.65 GB\nExpert cache hit rate: 62.8% (hit=31734 miss=18781)\nSpeed: 3.50 tok/s (36.6s for 128 tokens)\n[VK] qwen38: 0 matmuls on the GPU (0 matrices resident, 0.0 MiB; placed int8 0, bf16 0, f32 0)\n[VK] tier qwen38 run: device 89785 of 176640 routed experts (50.8%; this run 89785 of 176640) | CPU RAM hits 31734, disk loads 18781 | resident 2732 (budget 2732, 7.20 GiB of 7.21 GiB, 29 blocks, frag 0.52) | uploads 4828 (11.05 GiB, 2732 warm), evictions 2096, skipped 154 queue + 20102 rate, failed 0 | device 6574.7 ms, CPU share 15116.0 ms, waited 617.7 ms (91% of device time hidden) | balance: device share 0.41, 3200 rows handed to the CPU | stream: 48 steps, 1756 cold experts (4.52 GiB, 10.76 GB/s) and 62207 rows streamed in 88 sub-batches, prefetched 0 (0 used), 9441 cold experts (38734 rows) kept on the CPU | cooperative matmuls: 612\n[STATS] 339320 selections across 20449 distinct experts -> /root/colibri-test/colibri-pr-validation-20261005/results/q38-i4g64-mtp-trial-2/run.usage\n"
+ },
+ {
+ "tag": "q38-i4g64-mtp-trial-3",
+ "family": "qwen38",
+ "model": "/root/colibri-test/models/qwen38-flash-next-fp8",
+ "prompt": "/root/colibri-test/p4mtp/prompts/code.txt",
+ "new": 128,
+ "cap": 170,
+ "cold": true,
+ "env": {
+ "COLI_VULKAN": "1",
+ "COLI_VK_CHAIN": "0",
+ "COLI_VK_DENSE": "0",
+ "Q38_EXPERT_INT4": "1",
+ "Q38_MTP": "1",
+ "Q38_MTP_DRAFTS": "3"
+ },
+ "kind": "generate",
+ "prompt_sha256": "b840c508d4d473840929f53cc370614e1438c6d0fd5f9afc67176ce1d99dfc2f",
+ "history_sha256": "d7d86f69d2c83d900c7d869b7a04e34be19a4f466eb40ec7c52deb70d50fc108",
+ "cmd": [
+ "/root/colibri-test/colibri-pr-validation-20261005/src/c/qwen38",
+ "170",
+ "8",
+ "/root/colibri-test/p4mtp/prompts/code.txt"
+ ],
+ "settings": {
+ "OMP_NUM_THREADS": "8",
+ "COLI_NO_OMP_TUNE": "1",
+ "COLI_TIMERS": "1",
+ "VK_ICD_FILENAMES": "/usr/share/vulkan/icd.d/radeon_icd.json",
+ "COLI_LOOKUP": "0",
+ "Q38_MTP": "1",
+ "COLI_VULKAN": "1",
+ "COLI_VK_CHAIN": "0",
+ "COLI_VK_DENSE": "0",
+ "Q38_EXPERT_INT4": "1",
+ "Q38_MTP_DRAFTS": "3",
+ "SNAP": "/root/colibri-test/models/qwen38-flash-next-fp8",
+ "COLI_USAGE": "/root/colibri-test/colibri-pr-validation-20261005/results/q38-i4g64-mtp-trial-3/run.usage",
+ "N_NEW": "128"
+ },
+ "load_start": [
+ 1.56494140625,
+ 2.2109375,
+ 1.6943359375
+ ],
+ "source": "2f33f211ea6e28186d62cf9c00d511df826590a0",
+ "rc": 0,
+ "wall_s": 48.253,
+ "peak_rss_kib": 31081916,
+ "other_engines": [],
+ "stdout": "```c\nstatic int count_words(const char *s) {\n int total = 0, in_word = 0;\n for (const char *p = s; *p; p++) {\n if (*p == ' ' || *p == '\\n' || *p == '\\t') {\n in_word = 0;\n } else if (!in_word) {\n in_word = 1;\n total++;\n }\n }\n return total;\n}\n\nstatic void print_counts(const char **lines, int n) {\n int total",
+ "stderr": "== Qwen3.8-Flash-Next native text engine | cache=170/layer | CPU ==\n[tok] loaded 248044 base pieces + 33 added tokens (max id 248076) from /root/colibri-test/models/qwen38-flash-next-fp8/tokenizer.json\n[enc] prompt tokens: 241 | generating 128 new tokens\n[qwen38] vision tower: 27 blocks, hidden 1152, 16 heads, patch 16, merge 2\n[qwen38] native text weights: prefix=model.language_model, 48 layers, PLE=1, cache=170/layer, FP8=native, BF16=native (resident matrices 10.88 GiB)\n[qwen38] MTP head mtp.*: wiring b (a norm per stream, fc_hidden per stream), RoPE base 1e+07, 172.66 MiB resident; routed experts F8_E4M3 from the snapshot (the int4-g64 sidecar covers the model's layers), cache 170 (Q38_MTP_CAP) x 4.69 MiB = 0.78 GiB full\n[qwen38] MTP norm weights, mean: embedding -0.7642, hidden -0.3284, mixer 3.7881 (q38_rms0 scales by 1+w; a plain norm's weights would sit near 1)\n[qwen38] trunk: 570 matrices int8 on the CPU (4.06 GiB, 8.10 GiB of BF16 released) in 1.9s; Q38_TRUNK_CPU_INT8=0 keeps BF16\n[qwen38] routed experts: int4-g64 from experts-int4g64/, 2764800 bytes (2.64 MiB) per expert (the snapshot's native FP8: 4915200 bytes); cache 170/layer x 48 layers = 21.01 GiB full; 56.0 GB available now holds 421/layer\n[VK] VRAM pressure-proofing: memory_priority on, memory_budget on\n[VK] ready: AMD Radeon 780M Graphics (RADV PHOENIX), compute qfam 0, memtype 3, fused gate+up, absorb attention, tiled GEMM from S=2 and S*O>=4096 (cooperative matrix)\n[VK] qwen38: device ready, dense matrices on the CPU (COLI_VK_DENSE=0)\n[VK] qwen38: dense chain off (COLI_VK_CHAIN=0)\n[VK] qwen38: dense weights on the device and in host RAM (the dense part runs on the CPU)\n[USAGE] expert history: 158400 selections (/root/colibri-test/colibri-pr-validation-20261005/results/q38-i4g64-mtp-trial-3/run.usage)\n[VK] tier qwen38: on, AMD Radeon 780M Graphics (RADV PHOENIX), budget 7.28 GiB = 2759 experts of 2.6 MiB (fmt 4 gs 64, down fmt 4 gs 64), shared RAM: a quarter of what the expert cache leaves, own queue, up to 16 promotions per token, balanced against the CPU, streaming cold experts of prompt steps through 64 slots\n[VK] tier qwen38: warm start, 2759 experts from the history in 1.9s\nresident weights loaded in 4.5s | RSS after load: 11.77 GB\nGenerated (128 new tokens):\nText : [VK] tier qwen38 stream: a step of 241 rows; cold experts with 16 rows or more go to the device (the GEMM's 16 rows: the upload and the CPU's rows not measured yet); layer 0: 61 resident, 31 streamed, 289 kept on the CPU (1475 rows)\n\nTTFT: 17.65 s (time to first token)\n[qwen38 MTP] run: 3.97 tokens/forward (32 forwards per 127 tokens) | acceptance 100.0% (95/95 drafts) | wiring b | fixed depth, by position: 1 100.0% (32/32) 2 100.0% (32/32) 3 100.0% (31/31)\n[qwen38 timers] total: 33 forwards\n[qwen38 timers] expert-read 7.029 s 213.013 ms/forward\n[qwen38 timers] fp8-expand 0.000 s 0.000 ms/forward\n[qwen38 timers] routed-expert 5.311 s 160.935 ms/forward\n[qwen38 timers] shared-expert 0.514 s 15.583 ms/forward\n[qwen38 timers] resident-mm 6.182 s 187.319 ms/forward\n[qwen38 timers] deltanet 7.254 s 219.821 ms/forward\n[qwen38 timers] qsa-index 0.119 s 3.610 ms/forward\n[qwen38 timers] qsa-attn 2.329 s 70.563 ms/forward\n[qwen38 timers] ple 1.023 s 30.988 ms/forward\n[qwen38 timers] lm-head 1.392 s 42.179 ms/forward\n[qwen38 timers] architecture phases overlap resident-mm; expert-read is disk service and fp8-expand is synchronous miss work\n[qwen38 timers] prefill: 1 forwards\n[qwen38 timers] expert-read 4.123 s 4122.917 ms/forward\n[qwen38 timers] fp8-expand 0.000 s 0.000 ms/forward\n[qwen38 timers] routed-expert 2.122 s 2121.847 ms/forward\n[qwen38 timers] shared-expert 0.257 s 256.837 ms/forward\n[qwen38 timers] resident-mm 3.049 s 3049.040 ms/forward\n[qwen38 timers] deltanet 3.807 s 3807.400 ms/forward\n[qwen38 timers] qsa-index 0.013 s 12.882 ms/forward\n[qwen38 timers] qsa-attn 0.253 s 253.302 ms/forward\n[qwen38 timers] ple 0.654 s 653.609 ms/forward\n[qwen38 timers] lm-head 0.012 s 12.160 ms/forward\n[qwen38 timers] architecture phases overlap resident-mm; expert-read is disk service and fp8-expand is synchronous miss work\n[qwen38 timers] decode: 32 forwards\n[qwen38 timers] expert-read 2.907 s 90.829 ms/forward\n[qwen38 timers] fp8-expand 0.000 s 0.000 ms/forward\n[qwen38 timers] routed-expert 3.189 s 99.656 ms/forward\n[qwen38 timers] shared-expert 0.257 s 8.044 ms/forward\n[qwen38 timers] resident-mm 3.132 s 97.891 ms/forward\n[qwen38 timers] deltanet 3.447 s 107.709 ms/forward\n[qwen38 timers] qsa-index 0.106 s 3.320 ms/forward\n[qwen38 timers] qsa-attn 2.075 s 64.853 ms/forward\n[qwen38 timers] ple 0.369 s 11.531 ms/forward\n[qwen38 timers] lm-head 1.380 s 43.118 ms/forward\n[qwen38 timers] architecture phases overlap resident-mm; expert-read is disk service and fp8-expand is synchronous miss work\n[qwen38 expert I/O] weight-ranges=19252 scale-ranges=2 coalesced-gate-up=429 prefetched=17501 parallel-batches=1525 resident-scales=0.59 MiB\n\nPEAK RSS: 29.64 GB\nExpert cache hit rate: 62.7% (hit=31588 miss=18823)\nSpeed: 3.47 tok/s (36.9s for 128 tokens)\n[VK] qwen38: 0 matmuls on the GPU (0 matrices resident, 0.0 MiB; placed int8 0, bf16 0, f32 0)\n[VK] tier qwen38 run: device 89834 of 176640 routed experts (50.9%; this run 89834 of 176640) | CPU RAM hits 31588, disk loads 18823 | resident 2759 (budget 2759, 7.27 GiB of 7.28 GiB, 30 blocks, frag 0.68) | uploads 4855 (11.11 GiB, 2759 warm), evictions 2096, skipped 156 queue + 20329 rate, failed 0 | device 6636.2 ms, CPU share 15247.3 ms, waited 641.3 ms (90% of device time hidden) | balance: device share 0.45, 3263 rows handed to the CPU | stream: 48 steps, 1751 cold experts (4.51 GiB, 10.24 GB/s) and 62119 rows streamed in 87 sub-batches, prefetched 0 (0 used), 9429 cold experts (38693 rows) kept on the CPU | cooperative matmuls: 615\n[STATS] 339320 selections across 20454 distinct experts -> /root/colibri-test/colibri-pr-validation-20261005/results/q38-i4g64-mtp-trial-3/run.usage\n"
+ }
+ ]
+}
diff --git a/docs/speculative.md b/docs/speculative.md
index cebcefbb..ee968c3c 100644
--- a/docs/speculative.md
+++ b/docs/speculative.md
@@ -228,6 +228,16 @@ enabled, every verify used MTP and lookup supplied no drafts. Their gains theref
do not add. The small differences between MTP alone and the combined mode are not
evidence of a lookup benefit.
+A subsequent three-run comparison on commit
+`2f33f211ea6e28186d62cf9c00d511df826590a0` isolated `Q38_MTP=0` versus `1`, with
+`Q38_MTP_DRAFTS=3` fixed and the expert-tier configuration above. The samples were
+2.93, 2.93 and 2.90 tok/s without MTP, and 3.49, 3.50 and 3.47 with it: median
+2.93 to 3.49 tok/s (+19.1%, including prefill). All six outputs were byte-identical.
+The [validated experiment manifest](experiments/vulkan-qwen38-mtp-2026-10-05/manifest.json)
+links the raw logs and records the command, hardware, cache policy, prompt and starting
+history. It measures MTP on this workload, not the combined effect of every Vulkan
+change.
+
The tier-only run took 44.2 seconds without drafts and 36.7 seconds with MTP 3;
time to first token was 16.99 and 17.44 seconds respectively. Subtracting those rounded
times gives approximately 4.67 and 6.59 tokens/s after the first token. The dense