b70-optimization-lab

Gemma 4 26B A4B LocalMaxxing Targets

Research snapshot: updated 2026-06-30.

This page separates public leaderboard context from this lane’s promoted result rules. The goal is a valid Q8 / INT8-or-better result on Intel Arc Pro B70, not a speed-only lower-precision entry.

Quality/label guardrail: promoted Gemma 26B B70 submissions in this lane must use the literal gemma-4-26B-A4B-it-UD-Q8_K_XL.gguf target/verifier unless the user explicitly accepts a different target quantization. Historical labels like q8target-q40draft are shorthand for a Q8-quality UD-Q8_K_XL target with a Q4_0 MTP draft; they are not claims that the target was gemma-4-26B-A4B-it-Q8_0.gguf. Literal Q8_0.gguf target runs are separate controls and are not LocalMaxxing-promotable under the no-quality-loss rule.

As of 2026-06-27, synthetic/repeated/filled-long benchmark scores are diagnostic only. Do not submit or advertise them as real-world throughput. Promotion and LocalMaxxing submission require the fixed realistic prompt suite, one cold response per prompt, cached_tokens=0 every row, no prompt/KV/context/response reuse or n-gram/history acceleration, verified speculation only, and primary metric median_tok_s_1_100_after_ttft.

Public Target Context

Current public pages are useful as a speed target, but not as direct quality-equivalent comparisons:

Page	Current public top context	Why it is not directly comparable
google/gemma-4-26B-A4B-it	about 87.3 tok/s	Rows include mixed engines, hardware, and quantization such as MXFP4/Q4.
unsloth/gemma-4-26B-A4B-it-GGUF	about 94.3 tok/s	GGUF page, but public top rows are still mixed precision/hardware.
Jackrong/Gemopus-4-26B-A4B-it-GGUF	about 94.5 tok/s	Fine-tune, useful idea source only; not the same checkpoint.

Interpretation for this lane:

• A single-B70 Q8 result near or above 90 tok/s would already be interesting.
• A lower number can still be worth keeping if it is the first validated Q8 B70 baseline.
• Do not compare a Q8/INT8 result against MXFP4/Q4 entries as if the quality lane were identical.

Current policy-compliant LocalMaxxing submission:

• gemma4-q8-gpu0-finalpostnorm-reproexact-full512-20260701T084728Z, llama.cpp c926ad098 on one B70, UD-Q8_K_XL target/verifier with Q4_0 MTP draft, f16 KV, 32K context, FLASH_ATTN=on, GGML_SYCL_ENABLE_VMM=1, reordered-Q8 VDR2, F16 p021 small-ncols, bulk sampled-ID verifier host read, VDR2 selected-down fused weighted-sum, final post-norm residual fusion, n_max=3, n_min=2, p_min=0.0475, UBATCH_SIZE=1024;
• fixed suite: repro/gemma4-26b-a4b-q8-b70/realistic-suite-v1.json, each prompt once, cached_tokens=0 every row;
• primary metric: 124.97714084813418 tok/s median generated-token throughput for tokens 1-100 after TTFT;
• p10 103.83610041293263, mean 122.47435471668817, median TTFT 178.6938319564797 ms, median full-512 after-TTFT 114.87107033590866, median wall full-512 108.58112847853889;
• caveat: valid but high-variance. Same exact reproduction batch support lanes were 121.59076340768573, 119.26425148518223, and 113.63257982764395. A later documentation-pass rerun of the same wrapper measured 120.92334534956485 tok/s with cached_tokens=0 and 512/512 canary rows. Use repeat confirmations for effect-size claims;
• payload: data/localmaxxing-gemma4-26b-a4b-q8-b70-llamacpp-realistic-finalpostnorm-faon-vmm-ctx32768-full512-124tok-20260701.queue.json;
• LocalMaxxing: cmr1u77na01k2ld01kalwzs1e;
• response: data/localmaxxing-responses/gemma4-26b-a4b-q8-b70-llamacpp-realistic-finalpostnorm-faon-vmm-ctx32768-full512-124tok-20260701.submit.log.

Current service/prompt-processing LocalMaxxing submission, separate from the short-decode headline:

• gemma4-26b-a4b-q8-b70-llamacpp-service-32k-smoke-20260703, llama.cpp c926ad098 on one B70, UD-Q8_K_XL target/verifier with Q4_0 MTP draft verified by target, GEMMA4_26B_PROFILE=service;
• one cold request from the long-context suite case lc-24000-late, actual prompt tokens 32571, generated tokens 76, cached_tokens=0, exact JSON retrieval passed, unique prompt, no prompt/KV/context/response reuse and no warmed n-gram/history acceleration;
• metrics: 996.599888707516 approximate prompt/prefill tok/s, 115.17935520385454 decode tok/s after TTFT, 979.1564965957195 prompt+output wall tok/s, TTFT 32682.12285498157 ms;
• LocalMaxxing: cmr47ivql0045nv011pfdjlaa;
• payload: data/localmaxxing-gemma4-26b-a4b-q8-b70-llamacpp-service-32k-20260703.payload.json;
• response: data/localmaxxing-responses/gemma4-26b-a4b-q8-b70-llamacpp-service-32k-20260703.submit.log;
• supporting ladder: data/gemma4-long-context-service-gate-20260702Tservice-ladder-current-rep4.json passed 32/32 long-context rows and 64/64 canary rows across four B70 lanes. Average lane median prefill was 1192.965 tok/s; average lane median long-context decode was 131.786 tok/s.

Previous policy-compliant LocalMaxxing submission, now superseded:

• gemma4-q8-gpu3-q8lmhead-noreorder-control-full512-20260629T224927Z, llama.cpp c926ad098 on one B70, UD-Q8_K_XL target/verifier with Q4_0 MTP draft, f16 KV, 32K context, FLASH_ATTN=on, GGML_SYCL_ENABLE_VMM=1, reordered-Q8 VDR2, F16 p021 small-ncols, bulk sampled-ID verifier host read, VDR2 selected-down fused weighted-sum, n_max=3, n_min=2, p_min=0.0475, UBATCH_SIZE=1024;
• LM-head experiment flags were unset: LLAMA_SYCL_Q8_0_LM_HEAD_1COL_DMMV=0 and LLAMA_SYCL_Q8_0_LM_HEAD_1COL_NO_REORDER=0. The winning row came from the same baseline/control identity, not from the DMMV or no-reorder LM-head experiments;
• fixed suite: repro/gemma4-26b-a4b-q8-b70/realistic-suite-v1.json, each prompt once, cached_tokens=0 every row;
• primary metric: 121.41411987308553 tok/s median generated-token throughput for tokens 1-100 after TTFT;
• p10 107.032, mean 120.136, median TTFT 179.118 ms, median full-512 after-TTFT 110.391, median wall full-512 105.881;
• supporting same-family full512 confirmation: 119.94842631460949 tok/s under the same baseline flags, plus lower variance rows at 113.572, 114.088, and 111.988 tok/s. Treat this as a higher-variance ~120 tok/s baseline lane, not as a source-flag win.
• payload: data/localmaxxing-gemma4-26b-a4b-q8-b70-llamacpp-realistic-vdr2-selecteddown-faon-vmm-ctx32768-full512-121tok-20260629.queue.json;
• LocalMaxxing: cmqztiqdn02vnoe01egox6q3f;
• response: data/localmaxxing-responses/gemma4-26b-a4b-q8-b70-llamacpp-realistic-vdr2-selecteddown-faon-vmm-ctx32768-full512-121tok-20260629.submit.log.

Previous policy-compliant LocalMaxxing submission, now superseded:

• gemma4-q8-gpu3-faon-vmm-ctx32768-full512-20260629T211437Z, llama.cpp c926ad098 on one B70, UD-Q8_K_XL target/verifier with Q4_0 MTP draft, f16 KV, 32K context, FLASH_ATTN=on, GGML_SYCL_ENABLE_VMM=1, reordered-Q8 VDR2, F16 p021 small-ncols, bulk sampled-ID verifier host read, VDR2 selected-down fused weighted-sum, n_max=3, n_min=2, p_min=0.0475, UBATCH_SIZE=1024;
• fixed suite: repro/gemma4-26b-a4b-q8-b70/realistic-suite-v1.json, each prompt once, cached_tokens=0 every row;
• primary metric: 117.91456485086059 tok/s median generated-token throughput for tokens 1-100 after TTFT;
• p10 107.807, mean 118.881, median TTFT 180.169 ms, median full-512 after-TTFT 110.958, median wall full-512 106.807;
• supporting same-identity strict full512 rows: 116.45776605647993, 117.41509141115063, 115.08942949119734, and 117.45737477243767 tok/s; all passed 512/512 canary rows and cached_tokens=0. Treat this as a small confirmed variance-class improvement, not a synthetic/context-only result.
• payload: data/localmaxxing-gemma4-26b-a4b-q8-b70-llamacpp-realistic-vdr2-selecteddown-faon-vmm-ctx32768-full512-20260629.queue.json;
• LocalMaxxing: cmqzq5zu402troe01t774uyox;
• response: data/localmaxxing-responses/gemma4-26b-a4b-q8-b70-llamacpp-realistic-vdr2-selecteddown-faon-vmm-ctx32768-full512-20260629.submit.log.

Previous policy-compliant LocalMaxxing submission, now superseded:

• gemma4-q8-gpu1-selecteddown-bf16retest-control-full512-20260629T051323Z, llama.cpp c926ad098 on one B70, UD-Q8_K_XL target/verifier with Q4_0 MTP draft, f16 KV, 8K context, reordered-Q8 VDR2, F16 p021 small-ncols, bulk sampled-ID verifier host read, VDR2 selected-down fused weighted-sum, n_max=3, n_min=2, p_min=0.0475, UBATCH_SIZE=1024;
• fixed suite: repro/gemma4-26b-a4b-q8-b70/realistic-suite-v1.json, each prompt once, cached_tokens=0 every row;
• primary metric: 115.8466634928202 tok/s median generated-token throughput for tokens 1-100 after TTFT;
• p10 102.573, mean 114.574, median TTFT 181.167 ms, median full-512 after-TTFT 104.661, median wall full-512 100.640;
• supporting same-identity strict full512 rows include the previous 115.72789384447941 record plus 113.47081786263712, 113.81540554086772, 114.8109417270852, and 115.49839246092888 tok/s;
• payload: data/localmaxxing-gemma4-26b-a4b-q8-b70-llamacpp-realistic-vdr2-selecteddown-reordervdr2-full512-repeat-20260629.queue.json;
• LocalMaxxing: cmqyrpox4021dqk01co5o4fcw;
• response: data/localmaxxing-responses/gemma4-26b-a4b-q8-b70-llamacpp-realistic-vdr2-selecteddown-reordervdr2-full512-repeat-20260629.submit.log.

The adjacent BF16-direct retest did not beat controls and is recorded as a negative in ../../patches/gemma4-26b-a4b-q8-b70/20260629-selecteddown-bf16direct-currentstack-negative.md.

Previous policy-compliant LocalMaxxing submission, now superseded:

• gemma4-q8-gpu1-vdr2-selecteddown-reordervdr2-full512-20260629B, same VDR2 selected-down recipe, 115.72789384447941 tok/s median generated-token throughput for tokens 1-100 after TTFT, LocalMaxxing cmqyo0jyt08ippk01vhiobdnm;
• payload: data/localmaxxing-gemma4-26b-a4b-q8-b70-llamacpp-realistic-vdr2-selecteddown-reordervdr2-full512-20260629.queue.json;
• response: data/localmaxxing-responses/gemma4-26b-a4b-q8-b70-llamacpp-realistic-vdr2-selecteddown-reordervdr2-full512-20260629.submit.log.

Earlier policy-compliant LocalMaxxing submission, now superseded:

• gemma4-q8-gpu1-strict-vdr2-f16p021-bulksampled-confirm-B-n3-nmin2-p00475-ub1024-full512-20260628T052158Z, llama.cpp c926ad098 on one B70, UD-Q8_K_XL target/verifier with Q4_0 MTP draft, f16 KV, 8K context, reordered-Q8 VDR2, F16 p021 small-ncols, bulk sampled-ID verifier host read, n_max=3, n_min=2, p_min=0.0475, UBATCH_SIZE=1024;
• fixed suite: repro/gemma4-26b-a4b-q8-b70/realistic-suite-v1.json, each prompt once, cached_tokens=0 every row;
• primary metric: 98.34046474459183 tok/s median generated-token throughput for tokens 1-100 after TTFT;
• p10 85.979, mean 95.953, median TTFT 180.211 ms, median full-512 after-TTFT 91.174, median wall full-512 87.737;
• supporting same-identity strict full512 rows: 96.01452890026427, 95.90275376682132, and 94.94094934974818 tok/s;
• LocalMaxxing: cmqxchyra03xmqr01b963gmi1;
• payload: data/localmaxxing-payloads/gemma4-q8-vdr2-f16p021-bulksampled-confirm-20260628T052158.json;
• response: data/localmaxxing-responses/gemma4-26b-a4b-q8-b70-llamacpp-realistic-vdr2-mtp-n3-nmin2-p00475-ub1024-f16p021-bulksampled-full512-20260628.submit.log.

Previous policy-compliant F16-p021 submission, now superseded:

• gemma4-q8-gpu1-strict-vdr2-f16p021-smallncols-full512-exactconfirm-n3-nmin2-p00475-ub1024-20260628T010121Z;
• primary metric: 95.82453787677183 tok/s median generated-token throughput for tokens 1-100 after TTFT;
• LocalMaxxing: cmqx3687103v4qr01ace1ft3m;
• status: still valid strict evidence, but replaced as the current record by the VDR2 selected-down fused weighted-sum row above.

Previous policy-compliant VDR2 submission, now superseded:

• gemma4-q8-gpu1-strict-vdr2-recordconfirm-n3-nmin2-p00475-ub1024-20260627T221722Z;
• primary metric: 90.98312252660529 tok/s median generated-token throughput for tokens 1-100 after TTFT;
• LocalMaxxing: cmqwxep4a03qiqr010chjn93s;
• status: still valid strict evidence, but replaced as the current record by later rows, most recently the 123.67689864739785 tok/s row above.

Previous policy-compliant VDR2 submission, now superseded:

• gemma4-q8-gpu2-strict-vdr2-n3-p00475-repeat-ub1024-v21-20260627T201757Z;
• primary metric: 90.32179401019857 tok/s median generated-token throughput for tokens 1-100 after TTFT;
• LocalMaxxing: cmqwt1zk803ozqr01hctqss2z;
• status: still valid strict evidence, but replaced as the current record by later rows, most recently the 123.67689864739785 tok/s row above.

Previous policy-compliant VDR2 submission, now superseded:

• gemma4-q8-gpu2-strict-vdr2-n3-p00475-ub1024-v19-20260627T191931Z;
• primary metric: 89.45543282863798 tok/s median generated-token throughput for tokens 1-100 after TTFT;
• LocalMaxxing: cmqwqzayr03o8qr01j6lgx93n;
• status: still valid strict evidence, but replaced as the current record by later rows, most recently the 123.67689864739785 tok/s row above.

Previous policy-compliant VDR4 submission, now superseded:

• gemma4-q8-gpu0-vdr4default-mtp-n3-nmin2-p005-ub1024-realistic-gate-repeat-v8;
• primary metric: 87.61145306230438 tok/s median generated-token throughput for tokens 1-100 after TTFT;
• LocalMaxxing: cmqwnl2ag03lgqr01ch5bxknq;
• status: still valid strict evidence, but replaced as the current record by later rows, most recently the 123.67689864739785 tok/s row above.

Previous realistic-suite local Q8 observation, now superseded:

• gemma4-q8-gpu3-vdr4default-mtp-n3-nmin2-p0075-realistic-gate-v4-20260627T171157Z, llama.cpp c926ad098 on one B70, UD-Q8_K_XL target/verifier with Q4_0 MTP draft, f16 KV, 8K context, default reordered-Q8 VDR4, n_max=3, n_min=2, p_min=0.075, UBATCH_SIZE=720;
• fixed suite: repro/gemma4-26b-a4b-q8-b70/realistic-suite-v1.json, each prompt once, cached_tokens=0 every row;
• realistic_final_gate.passed=true;
• primary metric: 86.47445652599384 tok/s median generated-token throughput for tokens 1-100 after TTFT;
• p10 77.1001287639242, mean 84.8660983117653, median TTFT 182.121 ms, median full-512 after-TTFT 82.05238162525896, median wall full-512 78.41884310237452;
• summary: data/gemma4-q8-gpu3-vdr4default-mtp-n3-nmin2-p0075-realistic-gate-v4-20260627T171157Z/summary.json;
• server log: /mnt/fast-ai/bench-results/gemma4-26b-a4b-q8/servers/gemma4-q8-gpu3-vdr4default-mtp-n3-nmin2-p0075-realistic-gate-v4-20260627T171157Z.server.log;
• status: previously the best valid cold-suite observation, but not representative enough for LocalMaxxing promotion and now superseded by the v8 n3/p0.05/UB1024 strict result. The first exact repeat on GPU0 measured 81.73306503450416 tok/s, and a same-GPU repeat measured 82.89800056264573 tok/s.

Current realistic-suite no-spec control:

• gemma4-q8-gpu0-vdr4default-nospec-realistic-gate-v2-20260627T165335Z;
• same target model/runtime, no speculation, --ctx-checkpoints 0;
• primary metric: 74.29709476830473 tok/s median;
• p10 74.17460514894407, mean 74.26752349723967, median full-512 after-TTFT 72.21419554247626, median wall full-512 70.40573079055511;
• summary: data/gemma4-q8-gpu0-vdr4default-nospec-realistic-gate-v2-20260627T165335Z/summary.json;
• interpretation: this is the clean target-side baseline for future work. Draft-MTP now has a clear median advantage on the fixed suite, while no-spec remains the simplest quality/control reference.

Current local Q8 baseline:

• 20260623T052850Z, llama.cpp SYCL on one B70, UD-Q8_K_XL, f16 KV, 8K context;
• chat canary 128/128 pass;
• p512/o512 chat decode 26.10 tok/s after TTFT, 24.24 tok/s wall;
• status: keep as a control and do not submit as a record unless a baseline-only reference entry is explicitly desired. It is far below the public Gemma 4 family context and should be improved first.

Historical natural-stop local Q8 best:

• gemma4-q8-gpu2-syclopt0-faoff-parallel1-cache0-deep-20260623T0915, llama.cpp SYCL on one B70, UD-Q8_K_XL, f16 KV, 8K context;
• GGML_SYCL_DISABLE_OPT=0, FLASH_ATTN=off, POLL=50, --parallel 1 --cache-ram 0, REASONING=off;
• chat canary 384/384 pass;
• benchmark requested max_tokens=512, but actual completions averaged 146.4 tokens because the model stopped naturally;
• p512/o512 chat decode 42.15 tok/s after TTFT, 36.41 tok/s wall;
• status: submitted to LocalMaxxing and approved as cmqq9nqbh010gqo01a9jnzl6r;
• queue: data/localmaxxing-gemma4-26b-a4b-q8-b70-syclopt0-faoff-parallel1-cache0-20260623.queue.json;
• prior approved result: 41.81 tok/s, LocalMaxxing ID cmqq8phxt0103qo01afcgyjq8.

Historical short-prompt sustained-decode Q8 best:

• gemma4-q8-gpu0-currentbest-longprompt-deep-20260623T0945, llama.cpp SYCL on one B70, same runtime identity as the promoted parallel1-cache0 record;
• BENCH_PROMPT_MODE=long, which is a short instruction designed to prevent early stopping, not a true 512-token prompt fill;
• actual LocalMaxxing packet shape: 75 prompt tokens and 512 output tokens;
• chat canary 384/384 pass;
• decode 42.72 tok/s after TTFT, 41.35 tok/s wall;
• status: submitted to LocalMaxxing and approved as cmqqa6zbx010xqo01cdtfn8e0; this is a separate sustained-decode shape, not a direct replacement for the natural-stop/default-prompt row;
• queue: data/localmaxxing-gemma4-26b-a4b-q8-b70-long512-20260623.queue.json.

Historical short-prompt draft-MTP sustained-decode Q8 best:

• gemma4-q8-gpu0-mtp-n3-aot-repeat-long-deep-20260623T0353, llama.cpp SYCL on one B70, UD-Q8_K_XL main GGUF plus mtp-gemma-4-26B-A4B-it.gguf draft GGUF;
• --spec-type draft-mtp --spec-draft-n-max 3, draft KV f16/f16, AOT BMG build (GGML_SYCL_DEVICE_ARCH=bmg-g31), GGML_SYCL_DISABLE_OPT=0, FLASH_ATTN=off, POLL=50, --parallel 1 --cache-ram 0, REASONING=off;
• actual LocalMaxxing packet shape: 75 prompt tokens and 512 output tokens;
• chat canary 384/384 pass;
• decode 48.35 tok/s after TTFT, 46.60 tok/s wall;
• status: submitted to LocalMaxxing and approved as cmqqctk4w014kqo011gyyks7r;
• queue: data/localmaxxing-gemma4-26b-a4b-q8-b70-mtp-n3-aot-repeat-long512-20260623.queue.json.

Current filled-long draftless ngram-mod warmed/history artifact:

• gemma4-q8-gpu1-ngram-mod-20-32-64-ctx4096ub512-poll100-ctxcp0-filled-long-deep-20260623T1855, llama.cpp SYCL on one B70, UD-Q8_K_XL main GGUF, f16 KV;
• --spec-type ngram-mod --ctx-checkpoints 0 --spec-ngram-mod-n-match 20 --spec-ngram-mod-n-min 32 --spec-ngram-mod-n-max 64, AOT BMG build, llama.cpp c926ad098, GGML_SYCL_DISABLE_OPT=0, FLASH_ATTN=off, POLL=100, --parallel 1 --cache-ram 0, CTX_SIZE=4096, UBATCH_SIZE=512;
• actual LocalMaxxing packet shape: 588 prompt tokens and 512 output tokens (BENCH_PROMPT_MODE=filled-long);
• chat canary 384/384 pass;
• decode 280.64 tok/s after TTFT, 206.24 tok/s warmed wall;
• server log: 3493/3493 accepted/generated n-gram draft tokens, mean accepted length 63.38;
• status: submitted to LocalMaxxing before the fresh/warmed rule clarification and approved as cmqqyby6801dvqo01as3wenz2; retraction-needed if displayed as normal headline throughput. It supersedes prior warmed/history ngram rows only: cmqqxx7bp01dbqo012d2qiiw6 (280.04 tok/s), cmqqxjnif01d0qo01ix4oeixo (255.04 tok/s) and cmqqxbkzx01cxqo01j8p97627 (245.98 tok/s). It does not supersede the historical synthetic diagnostic draft-MTP frontier cmqwkedg303jeqr013z753j62 (176.216 tok/s first no-cache synthetic filled-long row; 176.403 tok/s supporting repeat mean; Q8 target/verifier with Q4_0 MTP draft only);
• queue: data/localmaxxing-gemma4-26b-a4b-q8-b70-ngrammod-20-32-64-poll100-filledlong512-20260623.queue.json;
• response: data/localmaxxing-responses/gemma4-26b-a4b-q8-b70-ngrammod-20-32-64-poll100-filledlong512-20260623.submit.log;
• caveat: this is history-cache acceleration on a repetitive sustained-decode benchmark. It is quality-preserving because every drafted token is verified by the Q8 target model, but it should not be described as unique-prompt no-cache decode throughput or a 32K-context result. API deletion was attempted for all four submitted ngram rows on 2026-06-23, but LocalMaxxing exposes no documented benchmark delete endpoint and DELETE /api/benchmarks/<id> returned 404; see data/localmaxxing-responses/gemma4-ngram-history-accelerated-delete-attempts-20260623.json and data/localmaxxing-responses/localmaxxing-openapi-benchmark-methods-20260623.json.

Historical filled-long draft-MTP Q8-target diagnostic best:

• gemma4-q8-gpu0-q8reorder-vdr2-ub720-nmin3-pmin010-fullconfirm-20260627T155347Z;
• llama.cpp SYCL on one B70, UD-Q8_K_XL main GGUF plus gemma-4-26B-A4B-it-Q4_0-MTP.gguf draft GGUF;
• same c926ad098 n=7 backend-sampling-off route-cache/fused-output/RMS reuse recipe as the previous row, plus LLAMA_SYCL_MUL_MAT_ID_MULTI_TOKEN_FAST=1, LLAMA_SYCL_MUL_MAT_ID_Q8_0_REORDER=1, and reordered Q8_0 MMVQ compile knob GGML_SYCL_REORDER_Q8_0_VDR_MMVQ=2; runtime shape is UBATCH_SIZE=720 with MTP_N_MIN=3 and MTP_P_MIN=0.10, validated on GPU0;
• actual LocalMaxxing packet shape: 588 prompt tokens and 512 output tokens (BENCH_PROMPT_MODE=filled-long);
• chat canary: 1536 repeats / 6144 case rows passed;
• synthetic filled-long row0: first measured no-cache request after TTFT 176.21623213048554 tok/s; supporting repeated-request mean 176.40259133127742 tok/s; first-row wall 139.3169544024847 tok/s; all rows report usage.prompt_tokens_details.cached_tokens=0;
• status: submitted to LocalMaxxing and approved as cmqwkedg303jeqr013z753j62 before the realistic final-gate policy; classify as diagnostic until the fixed suite passes. It is not a promoted real-world >150 tok/s claim.
• queue: data/localmaxxing-gemma4-26b-a4b-q8-b70-llamacpp-mtp-n7-q8target-q40draft-q8reorder-vdr2-ub720-nmin3-pmin010-fresh-20260627.queue.json;
• response: data/localmaxxing-responses/gemma4-26b-a4b-q8-b70-llamacpp-mtp-n7-q8target-q40draft-q8reorder-vdr2-ub720-nmin3-pmin010-fresh-20260627.submit.log.

Superseded Q8 MoE-ID reorder pre-final-gate diagnostic:

• gemma4-q8-gpu0-q8reorder-ub720-nmin3-pmin010-fullconfirm-20260627T144855Z;
• llama.cpp SYCL on one B70, UD-Q8_K_XL main GGUF plus gemma-4-26B-A4B-it-Q4_0-MTP.gguf draft GGUF;
• same c926ad098 n=7 backend-sampling-off route-cache/fused-output/RMS reuse recipe as the previous row, plus LLAMA_SYCL_MUL_MAT_ID_MULTI_TOKEN_FAST=1 and LLAMA_SYCL_MUL_MAT_ID_Q8_0_REORDER=1; runtime shape is UBATCH_SIZE=720 with MTP_N_MIN=3 and MTP_P_MIN=0.10, validated on GPU0;
• actual LocalMaxxing packet shape: 588 prompt tokens and 512 output tokens (BENCH_PROMPT_MODE=filled-long);
• chat canary: 1536 repeats / 6144 case rows passed;
• diagnostic pre-final-gate row0 metric: first measured no-cache request after TTFT 171.1076295077342 tok/s; supporting repeated-request mean 170.12922191012277 tok/s; first-row wall 135.66648146097913 tok/s; all rows report usage.prompt_tokens_details.cached_tokens=0;
• diagnostic pre-final-gate status: submitted to LocalMaxxing and approved as cmqwi45d803gyqr01td3vf9ka; superseded by cmqwkedg303jeqr013z753j62 (176.21623213048554 tok/s);
• queue: data/localmaxxing-gemma4-26b-a4b-q8-b70-llamacpp-mtp-n7-q8target-q40draft-q8reorder-ub720-nmin3-pmin010-fresh-20260627.queue.json;
• response: data/localmaxxing-responses/gemma4-26b-a4b-q8-b70-llamacpp-mtp-n7-q8target-q40draft-q8reorder-ub720-nmin3-pmin010-fresh-20260627.submit.log.

Superseded Q8 MoE-ID reorder pre-final-gate diagnostic:

• gemma4-q8-gpu1-q8reorder-ub704-nmin3-pmin010-fullconfirm-20260627T143126Z;
• 170.11205232778414 tok/s first synthetic / 169.87578310923394 tok/s support mean, LocalMaxxing cmqwhkbzj03guqr01h00c8n04.

Earlier superseded Q8 MoE-ID reorder pre-final-gate diagnostic:

• gemma4-q8-gpu0-mulmatid-fast-q8reorder-ub768-fullconfirm-20260627T142318Z;
• 169.9489959621758 tok/s first synthetic / 169.5501066933547 tok/s support mean, LocalMaxxing cmqwh8du403gfqr01d6ut1ddo.

Superseded UBATCH_SIZE=768, n_min=3, p_min=0.10 pre-final-gate Q8-target diagnostic:

• gemma4-q8-gpu0-ub768-nmin3-pmin010-fullrepeat-20260627T035307Z;
• llama.cpp SYCL on one B70, UD-Q8_K_XL main GGUF plus gemma-4-26B-A4B-it-Q4_0-MTP.gguf draft GGUF;
• same c926ad098 n=7 backend-sampling-off route-cache/fused-output recipe as the previous row; runtime shape is UBATCH_SIZE=768 with MTP_N_MIN=3 and MTP_P_MIN=0.10, validated on GPU0;
• actual LocalMaxxing packet shape: 588 prompt tokens and 512 output tokens (BENCH_PROMPT_MODE=filled-long);
• chat canary: 1536 repeats / 6144 case rows passed;
• diagnostic pre-final-gate row0 metric: first measured no-cache request after TTFT 104.22626983476746 tok/s; supporting repeated-request mean 104.17418893412489 tok/s; first-row wall 90.7413762430611 tok/s; all rows report usage.prompt_tokens_details.cached_tokens=0;
• diagnostic pre-final-gate status: submitted to LocalMaxxing and approved as cmqvv3kop0309qr013ekr8apu; this is a small variance-class micro-record over cmqvmjvzx02qvqr01qh9jikow (104.07050714456982 tok/s). The support mean also improves, but do not treat it as material progress toward >150 tok/s;
• queue: data/localmaxxing-gemma4-26b-a4b-q8-b70-llamacpp-mtp-n7-q8target-q40draft-routecache-mtpfusedoutargmax-selfusedweights-ub768-nmin3-pmin010-fresh-20260627.queue.json;
• response: data/localmaxxing-responses/gemma4-26b-a4b-q8-b70-llamacpp-mtp-n7-q8target-q40draft-routecache-mtpfusedoutargmax-selfusedweights-ub768-nmin3-pmin010-fresh-20260627.submit.log.

Superseded UBATCH_SIZE=768 draft-MTP pre-final-gate Q8-target best:

• gemma4-q8-gpu3-b1024u768-fullrepeat-20260626T235649Z;
• first measured no-cache request after TTFT 104.07050714456982 tok/s; supporting repeated-request mean 103.588578767931 tok/s; first-row wall 90.4869993907642 tok/s; approved as cmqvmjvzx02qvqr01qh9jikow;
• queue: data/localmaxxing-gemma4-26b-a4b-q8-b70-llamacpp-mtp-n7-q8target-q40draft-routecache-mtpfusedoutargmax-selfusedweights-ub768-fresh-20260627.queue.json;
• response: data/localmaxxing-responses/gemma4-26b-a4b-q8-b70-llamacpp-mtp-n7-q8target-q40draft-routecache-mtpfusedoutargmax-selfusedweights-ub768-fresh-20260627.submit.log.

Superseded same-stack filled-long draft-MTP pre-final-gate Q8-target best:

• gemma4-q8-gpu0-currentrecord-control-fullrepeat-20260626T230510Z;
• llama.cpp SYCL on one B70, UD-Q8_K_XL main GGUF plus gemma-4-26B-A4B-it-Q4_0-MTP.gguf draft GGUF;
• same c926ad098 n=7/n-min=2 backend-sampling-off route-cache/fused-output recipe as the previous row; this is a full repeat on GPU0, not a new mechanism;
• actual LocalMaxxing packet shape: 588 prompt tokens and 512 output tokens (BENCH_PROMPT_MODE=filled-long);
• chat canary 1536/1536 pass;
• diagnostic pre-final-gate row0 metric: first measured no-cache request after TTFT 103.9826628154082 tok/s; supporting repeated-request mean 104.09604904731648 tok/s; first-row wall 90.47935762548245 tok/s; all rows report usage.prompt_tokens_details.cached_tokens=0;
• diagnostic pre-final-gate status: submitted to LocalMaxxing and approved as cmqvjupek02pgqr01d46algvg; this is a variance-class micro-record over cmqviful602p0qr01vp27jw5i (103.95374341972274 tok/s) and not material progress toward >150 tok/s; superseded by cmqvmjvzx02qvqr01qh9jikow (104.07050714456982 tok/s), then by cmqvv3kop0309qr013ekr8apu (104.22626983476746 tok/s);
• queue: data/localmaxxing-gemma4-26b-a4b-q8-b70-llamacpp-mtp-n7-q8target-q40draft-routecache-repeat-fresh-20260626.queue.json;
• response: data/localmaxxing-responses/gemma4-26b-a4b-q8-b70-llamacpp-mtp-n7-q8target-q40draft-routecache-repeat-fresh-20260626.submit.log.

Superseded same-stack filled-long draft-MTP pre-final-gate Q8-target best:

• gemma4-q8-gpu2-routecache-mtpfusedoutargmax-selfusedweights-full-20260626T222525Z;
• llama.cpp SYCL on one B70, UD-Q8_K_XL main GGUF plus gemma-4-26B-A4B-it-Q4_0-MTP.gguf draft GGUF;
• submitted from a headless Supermicro AMD Threadripper PRO 5955WX platform with 128 GB DDR4; one Intel Arc Pro B70 32 GB was used for this record;
• same c926ad098 n=7/n-min=2 backend-sampling-off route-cache recipe as the previous row, plus LLAMA_GEMMA4_MTP_FUSED_OUTPUT_ARGMAX=1 and LLAMA_GEMMA4_MOE_SELECTED_SOFTMAX_FUSED=1;
• actual LocalMaxxing packet shape: 588 prompt tokens and 512 output tokens (BENCH_PROMPT_MODE=filled-long);
• chat canary 1536/1536 pass;
• diagnostic pre-final-gate row0 metric: first measured no-cache request after TTFT 103.95374341972274 tok/s; supporting repeated-request mean 104.13506066488091 tok/s; first-row wall 90.68621473793526 tok/s; all rows report usage.prompt_tokens_details.cached_tokens=0;
• diagnostic pre-final-gate status: submitted to LocalMaxxing and approved as cmqviful602p0qr01vp27jw5i; this is a small micro-record over cmqvbq8tf02m1qr010dom0vu1 (103.51547512013657 tok/s) and not material progress toward >150 tok/s; superseded by cmqvjupek02pgqr01d46algvg (103.9826628154082 tok/s), then by cmqvmjvzx02qvqr01qh9jikow (104.07050714456982 tok/s), then by cmqvv3kop0309qr013ekr8apu (104.22626983476746 tok/s);
• queue: data/localmaxxing-gemma4-26b-a4b-q8-b70-llamacpp-mtp-n7-q8target-q40draft-routecache-mtpfusedoutargmax-selfusedweights-fresh-20260626.queue.json;
• response: data/localmaxxing-responses/gemma4-26b-a4b-q8-b70-llamacpp-mtp-n7-q8target-q40draft-routecache-mtpfusedoutargmax-selfusedweights-fresh-20260626.submit.log.

Superseded filled-long draft-MTP route-cache pre-final-gate Q8-target best:

• gemma4-q8-gpu2-routecache-ctx8192-full-20260626T191746Z;
• llama.cpp SYCL on one B70, UD-Q8_K_XL main GGUF plus gemma-4-26B-A4B-it-Q4_0-MTP.gguf draft GGUF;
• validated after a four-GPU CTX screen on GPU2/ctx8192;
• diagnostic pre-final-gate row0 metric: first measured no-cache request after TTFT 103.51547512013657 tok/s; supporting repeated-request mean 103.19340167720759 tok/s; first-row wall 90.22004912439446 tok/s; all rows report usage.prompt_tokens_details.cached_tokens=0;
• LocalMaxxing: cmqvbq8tf02m1qr010dom0vu1;
• superseded by cmqviful602p0qr01vp27jw5i (103.95374341972274 tok/s), then by cmqvjupek02pgqr01d46algvg (103.9826628154082 tok/s), then by cmqvmjvzx02qvqr01qh9jikow (104.07050714456982 tok/s), then by cmqvv3kop0309qr013ekr8apu (104.22626983476746 tok/s);
• queue: data/localmaxxing-gemma4-26b-a4b-q8-b70-llamacpp-mtp-n7-q8target-q40draft-routecache-ctx8192-gpu2-pmin0136-fresh-20260626.queue.json;
• response: data/localmaxxing-responses/gemma4-26b-a4b-q8-b70-llamacpp-mtp-n7-q8target-q40draft-routecache-ctx8192-gpu2-pmin0136-fresh-20260626.submit.log.

Earlier superseded filled-long draft-MTP route-cache pre-final-gate Q8-target best:

• gemma4-q8-gpu0-mulmatid-routecache-full-20260626T184617Z;
• llama.cpp SYCL on one B70, UD-Q8_K_XL main GGUF plus gemma-4-26B-A4B-it-Q4_0-MTP.gguf draft GGUF;
• submitted from a headless Supermicro AMD Threadripper PRO 5955WX platform with 128 GB DDR4; one Intel Arc Pro B70 32 GB was used for this record;
• c926ad098 n=7/n-min=2 backend-sampling-off recipe with MTP_P_MIN=0.136, source-level CPU cleanup, LLAMA_MTP_DRAFT_FAST_ARGMAX=1, direct argmax-ID unroll (LLAMA_MTP_DRAFT_DIRECT_ARGMAX_IDS=1, LLAMA_MTP_DRAFT_DIRECT_ARGMAX_UNROLL=7), Gemma4Assistant q-only attention inputs (LLAMA_GEMMA4_MTP_QONLY_ATTN_INPUTS=1), verifier backend argmax IDs (LLAMA_SPEC_VERIFY_BACKEND_ARGMAX_IDS=1), deferred target h_nextn (LLAMA_MTP_DEFER_TARGET_H_NEXTN=1), LLAMA_GEMMA4_MOE_SELECTED_SOFTMAX=1, LLAMA_GEMMA4_MOE_WEIGHTED_SUM=1, GGML_SYCL_ENABLE_VMM=0, LLAMA_SYCL_MUL_MAT_ID_ROUTE_CACHE=1, UR_L0_USE_IMMEDIATE_COMMANDLISTS=1, BATCH_SIZE=1024, UBATCH_SIZE=1024, THREADS=8, POLL=100, GGML_SYCL_DISABLE_GRAPH=0, --ctx-checkpoints 0, draft threads 32/32;
• actual LocalMaxxing packet shape: 588 prompt tokens and 512 output tokens (BENCH_PROMPT_MODE=filled-long);
• chat canary 1536/1536 pass;
• diagnostic pre-final-gate row0 metric: first measured no-cache request after TTFT 103.30108468098005 tok/s; supporting repeated-request mean 103.06255061691155 tok/s; first-row wall 89.97733776184405 tok/s; all rows report usage.prompt_tokens_details.cached_tokens=0;
• diagnostic pre-final-gate status: submitted to LocalMaxxing and approved as cmqvalync02lhqr01h76rnti3; this was a micro-record over cmqsylo2l011nqr011yydjvne (103.2992004295621 tok/s) and supersedes cmqshlz8j00s0qr01f7lr24oh, cmqsf630x00r1qr01d1usfo2d, cmqsd2jpn00pwqr017fq21akz, cmqs7uyqb00lnqr01u9dtv63r, cmqs56wv100kjqr01de3fdspd, cmqs4jnx100k6qr01d1iy78kl, cmqrsupdk000jqr01af3eu6vu, cmqrjcly601kuqo01rbyub1x6, and earlier fast-argmax/CPU-cleanup result cmqr82niq01hgqo01v42y7ue8;
• superseded by cmqvbq8tf02m1qr010dom0vu1 (103.51547512013657 tok/s), then by cmqviful602p0qr01vp27jw5i (103.95374341972274 tok/s), then by cmqvjupek02pgqr01d46algvg (103.9826628154082 tok/s), then by cmqvmjvzx02qvqr01qh9jikow (104.07050714456982 tok/s), then by cmqvv3kop0309qr013ekr8apu (104.22626983476746 tok/s);
• queue: data/localmaxxing-gemma4-26b-a4b-q8-b70-llamacpp-mtp-n7-q8target-q40draft-routecache-pmin0136-fresh-20260626.queue.json;
• response: data/localmaxxing-responses/gemma4-26b-a4b-q8-b70-llamacpp-mtp-n7-q8target-q40draft-routecache-pmin0136-fresh-20260626.submit.log.

Previous material filled-long draft-MTP pre-final-gate Q8-target best:

• gemma4-q8-gpu0-selectedsoftmax-weightedsum-pmin0136-full-20260625T031510Z, repeated by gemma4-q8-gpu0-selectedsoftmax-weightedsum-pmin0136-full-repeat-20260625T032710Z;
• same Q8 target / Q4_0 MTP draft recipe, without the one-shot route cache;
• diagnostic pre-final-gate row0 metric: 103.2992004295621 tok/s; supporting mean 102.19335537277364 tok/s; first-row wall 89.84890823527608 tok/s;
• LocalMaxxing cmqsylo2l011nqr011yydjvne;
• keep this as the material baseline because the current route-cache record is only +0.001884 tok/s.

Superseded filled-long draft-MTP pre-final-gate Q8-target record:

• gemma4-q8-gpu1-rowargmax-safer-immediatecl1-full-20260624T193222Z, llama.cpp SYCL on one B70, UD-Q8_K_XL main GGUF plus gemma-4-26B-A4B-it-Q4_0-MTP.gguf draft GGUF;
• submitted from a headless Supermicro AMD Threadripper PRO 5955WX platform with 128 GB DDR4; one Intel Arc Pro B70 32 GB was used for this record;
• same c926ad098 n=7/n-min=2 backend-sampling-off recipe, with MTP_P_MIN=0.14, source-level CPU cleanup, LLAMA_MTP_DRAFT_FAST_ARGMAX=1, direct argmax-ID unroll (LLAMA_MTP_DRAFT_DIRECT_ARGMAX_IDS=1, LLAMA_MTP_DRAFT_DIRECT_ARGMAX_UNROLL=7), Gemma4Assistant q-only attention inputs (LLAMA_GEMMA4_MTP_QONLY_ATTN_INPUTS=1), safer verifier row-argmax IDs (LLAMA_SPEC_VERIFY_BACKEND_ARGMAX_IDS=1, with strict sampled-row shape assertions), deferred target h_nextn (LLAMA_MTP_DEFER_TARGET_H_NEXTN=1), GGML_SYCL_ENABLE_VMM=0, UR_L0_USE_IMMEDIATE_COMMANDLISTS=1, BATCH_SIZE=1024, UBATCH_SIZE=1024, THREADS=8, POLL=100, GGML_SYCL_DISABLE_GRAPH=0, --ctx-checkpoints 0, draft threads 32/32;
• actual LocalMaxxing packet shape: 588 prompt tokens and 512 output tokens (BENCH_PROMPT_MODE=filled-long);
• chat canary 1536/1536 pass;
• diagnostic pre-final-gate row0 metric: first measured no-cache request after TTFT 101.60238982389097 tok/s; supporting repeated-request mean 100.83458420322299 tok/s; first-row wall 88.50781195831634 tok/s; all rows report usage.prompt_tokens_details.cached_tokens=0;
• diagnostic pre-final-gate status: submitted to LocalMaxxing and approved as cmqshlz8j00s0qr01f7lr24oh; supersedes the safer row-argmax/defer-H record cmqsf630x00r1qr01d1usfo2d, earlier row-argmax/defer-H record cmqsd2jpn00pwqr017fq21akz, approved direct-unroll/q-only batch/thread/graph pre-final-gate record cmqs7uyqb00lnqr01u9dtv63r, cmqs56wv100kjqr01de3fdspd, cmqs4jnx100k6qr01d1iy78kl, cmqrsupdk000jqr01af3eu6vu, cmqrjcly601kuqo01rbyub1x6, and earlier fast-argmax/CPU-cleanup result cmqr82niq01hgqo01v42y7ue8;
• queue: data/localmaxxing-gemma4-26b-a4b-q8-b70-llamacpp-mtp-n7-q8target-q40draft-rowargmax-safer-deferh-pmin014-immediatecl1-fresh-20260624.queue.json;
• response: data/localmaxxing-responses/gemma4-26b-a4b-q8-b70-llamacpp-mtp-n7-q8target-q40draft-rowargmax-safer-deferh-pmin014-immediatecl1-fresh-20260624.submit.log.

Superseded safer row-argmax/defer-H pre-final-gate Q8-target record:

• gemma4-q8-gpu0-rowargmax-safer-pmin014-full-20260624T183044Z;
• LocalMaxxing cmqsf630x00r1qr01d1usfo2d, 101.4817054635395 tok/s first no-cache request after TTFT, 101.24898926956536 tok/s support mean, 1536/1536 chat canary;
• superseded by the immediate-command-list result above.

Superseded row-argmax/defer-H pre-final-gate Q8-target record:

• gemma4-q8-gpu0-rowargmax-deferh-pmin014-full-20260624T173546Z;
• LocalMaxxing cmqsd2jpn00pwqr017fq21akz, 101.42819815648124 tok/s first no-cache request after TTFT, 100.76942425937877 tok/s support mean, 384/384 chat canary;
• superseded only by the safer verifier sampled-row shape-guard result above.

Previous direct-unroll/q-only batch/thread filled-long draft-MTP Q8-target best:

• gemma4-q8-gpu0-mtp-n7-directunroll7-qonly-b1024u1024-th8-syclgraph0-full-20260624T144749Z, llama.cpp SYCL on one B70, UD-Q8_K_XL main GGUF plus gemma-4-26B-A4B-it-Q4_0-MTP.gguf draft GGUF;
• submitted from a headless Supermicro AMD Threadripper PRO 5955WX platform with 128 GB DDR4; one Intel Arc Pro B70 32 GB was used for this record;
• same c926ad098 n=7/n-min=2/p-min=0.12 backend-sampling-off recipe, plus source-level CPU cleanup, LLAMA_MTP_DRAFT_FAST_ARGMAX=1, direct argmax-ID unroll (LLAMA_MTP_DRAFT_DIRECT_ARGMAX_IDS=1, LLAMA_MTP_DRAFT_DIRECT_ARGMAX_UNROLL=7), Gemma4Assistant q-only attention inputs (LLAMA_GEMMA4_MTP_QONLY_ATTN_INPUTS=1), GGML_SYCL_ENABLE_VMM=0, BATCH_SIZE=1024, UBATCH_SIZE=1024, THREADS=8, POLL=100, GGML_SYCL_DISABLE_GRAPH=0, --ctx-checkpoints 0, draft threads 32/32;
• actual LocalMaxxing packet shape: 588 prompt tokens and 512 output tokens (BENCH_PROMPT_MODE=filled-long);
• chat canary 384/384 pass;
• diagnostic pre-final-gate row0 metric: first measured no-cache request after TTFT 98.61718830251647 tok/s; supporting repeated-request mean 97.95563472401156 tok/s; first-row wall 86.2620078252172 tok/s; all rows cached_tokens=0;
• diagnostic pre-final-gate status: submitted to LocalMaxxing and approved as cmqs7uyqb00lnqr01u9dtv63r; supersedes approved batch/thread tuned pre-final-gate record cmqs56wv100kjqr01de3fdspd, approved direct-unroll/q-only Q8-target/Q4_0-draft pre-final-gate record cmqs4jnx100k6qr01d1iy78kl, approved Q8-target/Q4_0-draft pre-final-gate record cmqrsupdk000jqr01af3eu6vu, approved Q8-draft pre-final-gate record cmqrjcly601kuqo01rbyub1x6, and earlier fast-argmax/CPU-cleanup result cmqr82niq01hgqo01v42y7ue8;
• queue: data/localmaxxing-gemma4-26b-a4b-q8-b70-llamacpp-mtp-n7-q8target-q40draft-directunroll7-qonly-b1024u1024-th8-syclgraph0-fresh-20260624.queue.json;
• response: data/localmaxxing-responses/gemma4-26b-a4b-q8-b70-llamacpp-mtp-n7-q8target-q40draft-directunroll7-qonly-b1024u1024-th8-syclgraph0-fresh-20260624.submit.log.

Previous batch/thread filled-long draft-MTP Q8-target best:

• gemma4-q8-gpu3-mtp-n7-directunroll7-qonly-b1024u1024-th8-full-20260624T135701Z;
• diagnostic pre-final-gate row0 metric: first measured no-cache request after TTFT 98.4913689785019 tok/s; supporting repeated-request mean 97.88614261273774 tok/s; first-row wall 86.19446305286014 tok/s; all rows cached_tokens=0;
• LocalMaxxing cmqs56wv100kjqr01de3fdspd.

Previous direct-unroll/q-only filled-long draft-MTP Q8-target best:

• gemma4-q8-gpu0-mtp-n7-directunroll7-qonly-full-20260624T1432Z;
• diagnostic pre-final-gate row0 metric: first measured no-cache request after TTFT 96.82235022330569 tok/s; supporting repeated-request mean 97.22636780761407 tok/s; first-row wall 82.46162357827212 tok/s; all rows cached_tokens=0;
• LocalMaxxing cmqs4jnx100k6qr01d1iy78kl.

Previous filled-long draft-MTP pre-final-gate Q8-target best:

• gemma4-q8-gpu0-mtp-n7-draftq40-full-20260624T081218Z, llama.cpp SYCL on one B70, UD-Q8_K_XL main GGUF plus gemma-4-26B-A4B-it-Q4_0-MTP.gguf draft GGUF;
• diagnostic pre-final-gate row0 metric: first measured no-cache request after TTFT 95.26352416631231 tok/s; supporting repeated-request mean 95.38558173206405 tok/s; first-row wall 81.28549578539435 tok/s; all rows cached_tokens=0;
• LocalMaxxing cmqrsupdk000jqr01af3eu6vu.

Previous filled-long draft-MTP pre-final-gate Q8-draft diagnostic:

• gemma4-q8-gpu0-mtp-n7-cleanrebuild-control-full-fresh-20260624T032733Z, llama.cpp SYCL on one B70, UD-Q8_K_XL main GGUF plus mtp-gemma-4-26B-A4B-it.gguf draft GGUF;
• diagnostic pre-final-gate row0 metric: first measured no-cache request after TTFT 94.36621149389549 tok/s; supporting repeated-request mean 92.55939821446442 tok/s; first-row wall 80.50542700854518 tok/s; all rows cached_tokens=0;
• LocalMaxxing cmqrjcly601kuqo01rbyub1x6.

Previous filled-long draft-MTP pre-final-gate Q8 diagnostic:

• gemma4-q8-gpu0-mtp-n7-c926-fastargmax-cpucleanup-vmm0-ub512-poll100-full-ctxcp0-nmin2-pmin012-nobs-dthreads32-dtb32-filled-long-20260623T222838Z, llama.cpp SYCL on one B70, UD-Q8_K_XL main GGUF plus mtp-gemma-4-26B-A4B-it.gguf draft GGUF;
• same c926ad098 n=7/n-min=2/p-min=0.12 backend-sampling-off recipe, plus source-level CPU cleanup, LLAMA_MTP_DRAFT_FAST_ARGMAX=1, LLAMA_MTP_DRAFT_FAST_TOPK=0, GGML_SYCL_ENABLE_VMM=0, UBATCH_SIZE=512, and POLL=100;
• actual LocalMaxxing packet shape: 588 prompt tokens and 512 output tokens (BENCH_PROMPT_MODE=filled-long);
• chat canary 384/384 pass;
• conservative diagnostic pre-final-gate row0 metric: first measured request after TTFT 92.397 tok/s; supporting independent repeated-request mean 92.767 tok/s; wall mean 83.289 tok/s; all rows cached_tokens=0;
• diagnostic pre-final-gate status: submitted to LocalMaxxing and approved as cmqr82niq01hgqo01v42y7ue8; supersedes approved CPU-cleanup result cmqr7ni7u01gxqo01wtqsrn3u and fast-top-k result cmqqsecuk01azqo018ahv0i1s;
• queue: data/localmaxxing-gemma4-26b-a4b-q8-b70-mtp-n7-c926ad098-fastargmax-cpucleanup-vmm0-ub512-poll100-filledlong512-20260623.queue.json;
• response: data/localmaxxing-responses/gemma4-26b-a4b-q8-b70-mtp-n7-c926ad098-fastargmax-cpucleanup-vmm0-ub512-poll100-filledlong512-20260623.submit.log.

Previous filled-long draft-MTP pre-final-gate Q8 diagnostic:

• gemma4-q8-gpu0-mtp-n7-c926-fasttopk10-repeat-ctxcp0-nmin2-pmin012-nobs-dthreads32-dtb32-filled-long-deep-20260623T150833Z, llama.cpp SYCL on one B70, UD-Q8_K_XL main GGUF plus mtp-gemma-4-26B-A4B-it.gguf draft GGUF;
• same c926ad098 recipe as the prior result, plus source-level fast top-k MTP draft bypass: LLAMA_MTP_DRAFT_FAST_TOPK=1, LLAMA_MTP_DRAFT_TOP_K=10;
• actual LocalMaxxing packet shape: 588 prompt tokens and 512 output tokens (BENCH_PROMPT_MODE=filled-long);
• chat canary 384/384 pass;
• decode 91.62 tok/s after TTFT, 71.29 tok/s warmed wall;
• diagnostic pre-final-gate status: submitted to LocalMaxxing and approved as cmqqsecuk01azqo018ahv0i1s; superseded by the CPU-cleanup and fast-argmax records above. It superseded approved result cmqqkmbhr017oqo017rdfxqh2 (91.16 tok/s);
• queue: data/localmaxxing-gemma4-26b-a4b-q8-b70-mtp-n7-c926ad098-fasttopk10-filledlong512-20260623.queue.json;
• response: data/localmaxxing-responses/gemma4-26b-a4b-q8-b70-mtp-n7-c926ad098-fasttopk10-filledlong512-20260623.submit.log.

Previous filled-long draft-MTP sustained-decode Q8 best:

• gemma4-q8-gpu0-mtp-n7-latest-c926ad098-ctxcp0-nmin2-pmin012-nobs-dthreads32-dtb32-filled-long-deep-20260623T113058Z, llama.cpp SYCL on one B70, UD-Q8_K_XL main GGUF plus mtp-gemma-4-26B-A4B-it.gguf draft GGUF;
• --spec-type draft-mtp --spec-draft-n-max 7 --spec-draft-n-min 2 --spec-draft-p-min 0.12 --no-spec-draft-backend-sampling --spec-draft-threads 32 --spec-draft-threads-batch 32, --ctx-checkpoints 0, draft KV f16/f16, AOT BMG build (GGML_SYCL_DEVICE_ARCH=bmg-g31), llama.cpp c926ad098, GGML_SYCL_DISABLE_OPT=0, FLASH_ATTN=off, POLL=50, --parallel 1 --cache-ram 0, REASONING=off;
• actual LocalMaxxing packet shape: 588 prompt tokens and 512 output tokens (BENCH_PROMPT_MODE=filled-long);
• chat canary 384/384 pass;
• decode 91.16 tok/s after TTFT, 71.06 tok/s warmed wall;
• status: submitted to LocalMaxxing and approved as cmqqkmbhr017oqo017rdfxqh2; superseded by the fast-top-k result above. It had superseded approved result cmqqi1p2c016jqo01vndau1y9 (91.05 tok/s);
• queue: data/localmaxxing-gemma4-26b-a4b-q8-b70-mtp-n7-c926ad098-ctxcp0-nmin2-pmin012-nobs-dthreads32-dtb32-filledlong512-20260623.queue.json.

Previous draft-MTP approved result:

• gemma4-q8-gpu2-mtp-n7-aot-nmin2-pmin012-nobs-dthreads32-dtb32-filled-long-deep-20260623T101814Z, 91.05 tok/s after TTFT on the filled-long shape, approved as cmqqi1p2c016jqo01vndau1y9.
• gemma4-q8-gpu3-mtp-n7-aot-nmin2-pmin010-nobs-dthreads32-filled-long-deep-20260623T094131Z, 90.42 tok/s after TTFT on the filled-long shape, approved as cmqqgn3cm0163qo010optg91u.
• gemma4-q8-gpu1-mtp-n3-long-deep-20260623T0328, 46.36 tok/s after TTFT, approved as cmqqbyv5w013vqo019pmp161f;
• gemma4-q8-gpu0-mtp-n3-repeat-long-deep-20260623T0337, 47.63 tok/s after TTFT, approved as cmqqc99m2014cqo01s5t61bs6;
• gemma4-q8-gpu3-mtp-n3-aot-bmg-long-deep-20260623T0345, 47.92 tok/s after TTFT, approved as cmqqcje2r014fqo01e8rrgwwr;
• gemma4-q8-gpu1-mtp-n4-long-deep-20260623T1140, 44.50 tok/s after TTFT, approved as cmqqblfw30132qo01jbi1svnu.
• gemma4-q8-gpu3-mtp-n3-aot-filled-long-deep-20260623T085322Z, 68.19 tok/s after TTFT on the filled-long shape, approved as cmqqexo5x0151qo0154xsie7s;
• gemma4-q8-gpu2-mtp-n3-aot-psplit020-filled-long-deep-20260623T085844Z, 68.51 tok/s after TTFT on the filled-long shape, approved as cmqqf759s0154qo01gwqa14uc.
• gemma4-q8-gpu3-mtp-n4-aot-filled-long-deep-20260623T085822Z, 74.39 tok/s after TTFT on the filled-long shape, approved as cmqqf75p70157qo018fsavf0g.
• gemma4-q8-gpu1-mtp-n4-aot-psplit020-filled-long-deep-20260623T090712Z, 74.50 tok/s after TTFT on the filled-long shape, approved as cmqqfe75s015aqo01xr94yxh0.
• gemma4-q8-gpu2-mtp-n6-aot-nmin2-pmin015-filled-long-deep-20260623T091227Z, 83.52 tok/s after TTFT on the filled-long shape, approved as cmqqfnilo015lqo011nm0q2tn.
• gemma4-q8-gpu3-mtp-n7-aot-nmin2-pmin015-filled-long-deep-20260623T091939Z, 87.88 tok/s after TTFT on the filled-long shape, approved as cmqqfv296015sqo0126mym3ko.
• gemma4-q8-gpu1-mtp-n7-aot-nmin2-pmin010-filled-long-deep-20260623T092524Z, 88.35 tok/s after TTFT on the filled-long shape, approved as cmqqg1r0l015xqo01e6d696mx.

Submission Packet

LocalMaxxing requires at minimum:

• hfId;
• hardware;
• engineName;
• quantization;
• tokSOut;
• at least one secondary metric: tokSPrefill, tokSTotal, ttftMs, or peakVramGb.

Useful optional fields for this repo’s records:

• modelRevision;
• engineVersion;
• backend;
• promptTokens;
• outputTokens;
• contextLength;
• batchSize;
• engineFlags;
• notes.

The API supports a dry-run endpoint before writing a real benchmark. The local helper reads the key from LMX_API_KEY or /home/steve/.config/localmaxxing/api_key; never put that key in a payload, note, shell history snippet, or commit.

Gemma 4 Payload Shape

For the primary GGUF lane:

hfId: unsloth/gemma-4-26B-A4B-it-GGUF
modelRevision: 3bb10d594514ef4edb7f3a65d41a7e4eb8c5767a
engineName: llama.cpp
backend: sycl/xpu
quantization: UD-Q8_K_XL
hardware.hwClass: DISCRETE_GPU
hardware.gpuName: Intel Arc Pro B70
hardware.vramGb: 32
hardware.gpuCount: 1 for a single-replica record, 4 only for aggregate service records

Engine flags should include the command snippet and the relevant values from the server log:

• llama.cpp commit;
• CTX_SIZE;
• BATCH_SIZE;
• UBATCH_SIZE;
• CACHE_TYPE_K / CACHE_TYPE_V;
• GGML_SYCL_DISABLE_OPT;
• GGML_SYCL_DISABLE_GRAPH;
• GGML_SYCL_DISABLE_DNN;
• ONEAPI_DEVICE_SELECTOR;
• -fa state;
• benchmark prompt mode and actual prompt/output token counts;
• MTP/spec flags if enabled.
• source-level experimental MTP knobs if enabled, including LLAMA_MTP_DRAFT_FAST_TOPK, LLAMA_MTP_DRAFT_TOP_K, and LLAMA_MTP_DRAFT_LOGIT_GAP_MIN.

Do Not Submit If

• any chat canary fails;
• only raw /v1/completions was tested;
• usage.completion_tokens is missing and output token count was guessed;
• realistic_final_gate.passed is absent or false;
• any fixed-suite row has missing or nonzero cached_tokens;
• engineFlags.primaryMetricName is not median_tok_s_1_100_after_ttft;
• the payload lacks p10, mean, TTFT, wall-clock tok/s, full-512 tok/s, prompt hashes, output hashes, model identity, runtime commit, launch env/flags, and log/result paths;
• the headline uses a synthetic row0, a repeated-prompt average, or a max instead of the fixed-suite median for generated tokens 1-100 after TTFT;
• the speedup is from n-gram/history continuation learned from earlier benchmark requests;
• prefix cache, context checkpoints, or response reuse contributed to the measured request and the result is being labeled fresh-response;
• the result is Q6/Q4/MXFP4/NVFP4 but labeled as the Q8 lane;
• the GPU count mixes independent replicas with a single-session model split;
• a speed win comes from a config family with unresolved nondeterministic failures;
• the model file, runtime commit, or launch identity is incomplete.

Source Links

• LocalMaxxing API docs: https://www.localmaxxing.com/en/api-docs
• Google model page: https://www.localmaxxing.com/en/models/google/gemma-4-26B-A4B-it
• Unsloth GGUF model page: https://localmaxxing.com/en/models/unsloth/gemma-4-26B-A4B-it-GGUF
• LocalMaxxing CLI: https://github.com/LottoLottoLotto/localmaxxing-cli

This site is open source. Improve this page.