Qwen3.6 27B MTP weights / lemonade b1295 / native PP131072, full PP4K-128K curve

First-party historical benchmark from daimonionnn, inspected 2026-10-13. Evidence is pinned to source commit 7de1d467c9f2e8282051d9d2332f5ef584f4fa66. No benchmark was rerun or weights downloaded. Hardware: Minisforum MS-S1 MAX, AMD Ryzen AI MAX+ 395, Radeon 8060S (gfx1151), 128 GB-class unified LPDDR5X-8000, Windows 11 Pro. Machine identity and OS are source-reported. Short-pack CSV captures 128 memory units at 8000 MT/s; its PowerShell helper sums Win32_PhysicalMemory Capacity using 1GB (binary GiB), although the column is named ram_total_gb. This is unified capacity, not dedicated VRAM or resident memory. The reported eight 16 GB firmware entries do not establish removable DIMMs. Physical CPU core count and enabled GPU core/CU count are unknown. Source describes 32 CPU threads; actual benchmark worker threads are 16. Participating GPU count = 1, bound to the measured process: every retained native row has gpu_info exactly AMD Radeon(TM) 8060S Graphics, devices auto, main_gpu 0, n_gpu_layers -1, n_cpu_moe 0, layer split, tensor_split 0.00 and no tensor buffer overrides. Both recorded engine revisions' get_gpu_info iterate every GPU/IGPU and join every description without deduplication. Thus this is one enumerated GPU with full layer offload. The older CSV's backends CUDA,ROCm names registered backend libraries, not two GPUs; GPU enumeration contains only the 8060S. Device identity is not borrowed from another run. Timing: native llama-bench synthetic token IDs, not an HTTP request. Every trial clears KV; prompt and generation warmup precede measured trials. Model load and warmup are excluded. The timer starts before test_prompt/test_gen and stops after synchronization. PP counts decoded input tokens; TG counts serial one-token decode calls. avg_ts is the arithmetic mean of per-trial tokens/elapsed time, and stddev_ts is sample SD; it is not tokens divided by avg_ns. The scripts/CSV retain aggregates, not per-trial samples, so these are publisher-reported native statistics, not independently reconstructed trials. No independent TTFT exists; tokens/rate in the source is a prefill-duration proxy and is excluded from the TTFT field. Settings: batch 2048, microbatch 1024, 16 CPU workers, f16 K and V cache, flash attention on, mmap on, direct I/O off, embedding mode off, KV offload on, no CPU MoE, no speculative decoding. Requested context in this benchmark code is n_prompt+n_gen+n_depth; final rounded allocation is not logged and stays unknown. Actual timed tokens and existing context depth are separately recorded below. Model: Qwen3.6-27B MTP weights, UD-Q4_K_XL. Exact measured local filename: Qwen3.6-27B-UD-Q4_K_XL.gguf. Logged tensor parameter count 27320697856; logged model tensor bytes 17898102784. Nominal family size 27B is not a multiplication of architecture labels or measured RAM usage. Artifact provenance: measured local path names unsloth/Qwen3.6-27B-MTP-GGUF/Qwen3.6-27B-UD-Q4_K_XL.gguf. Current HF metadata and the file page were retrieved successfully; current repository revision 5cb35eb3dcbf52dbce5f87dbc64df6aaffadcace is a reference checked today, not a proven tested revision. Historical local weight hash/HF snapshot is absent. HF card lists upstream ['Qwen/Qwen3.6-27B']. Artifact weights were not downloaded. The MTP-pack filename does not imply speculative decoding here: llama-bench runs only ordinary forward passes. Its 27,320,697,856 logged parameters include the loaded pack metadata; do not equate it to the separate 26,895,998,464-parameter non-MTP Qwen quant. Runtime: lemonade ROCm 7 gfx1151 release b1295, logged build 1/049326a. Source states both campaigns predate the July 14 BIOS/IOMMU change, so BIOS 1.06/IOMMU on applies. Exact mode and binary/compiler hash are not logged. This is a PP-only sweep: two measured repetitions per cell, depth 0, n_gen 0. No decode rate is supplied or borrowed from a different file. Sortable headline uses the largest actual input, 131072 tokens. - 2026-07-11T07:38:18Z: PP4096 348.031048 ± 0.445992 t/s; avg_ns 11769074600. - 2026-07-11T07:42:01Z: PP16384 312.373369 ± 0.337756 t/s; avg_ns 52450084450. - 2026-07-11T07:51:12Z: PP32768 271.578058 ± 1.435933 t/s; avg_ns 120659446200. - 2026-07-11T08:12:54Z: PP65536 213.066492 ± 1.742657 t/s; avg_ns 307595020800. - 2026-07-11T09:53:33Z: PP131072 148.790366 ± 0.199385 t/s; avg_ns 880918038850. The scripts first run markdown output and then independently rerun CSV output. README figures (for example 32.3/150.6 at 128K) are not the retained CSV figures (32.145328/148.790366) and are not substituted here. Each CSV row represents the second invocation, after the preceding markdown invocation. These are actual synthetic input counts, not advertised model context capacity. Requested context equals that input count; effective allocated capacity is unknown. Sources: https://github.com/daimonionnn/amd-rocmfpx-for-win/blob/7de1d467c9f2e8282051d9d2332f5ef584f4fa66/llm-inference/results/longctx-prefill-rocm7.csv https://github.com/daimonionnn/amd-rocmfpx-for-win/blob/7de1d467c9f2e8282051d9d2332f5ef584f4fa66/llm-bench/README.md https://github.com/daimonionnn/amd-rocmfpx-for-win/blob/7de1d467c9f2e8282051d9d2332f5ef584f4fa66/llm-bench/results-strix-halo-rocm.md https://github.com/daimonionnn/amd-rocmfpx-for-win/blob/7de1d467c9f2e8282051d9d2332f5ef584f4fa66/llm-bench/Common.ps1 https://github.com/daimonionnn/amd-rocmfpx-for-win/blob/7de1d467c9f2e8282051d9d2332f5ef584f4fa66/llm-bench/Run-RocmTest.ps1 https://github.com/daimonionnn/amd-rocmfpx-for-win/blob/7de1d467c9f2e8282051d9d2332f5ef584f4fa66/llm-inference/README.md https://github.com/ggml-org/llama.cpp/blob/f5525f7e7/tools/llama-bench/llama-bench.cpp https://github.com/lemonade-sdk/llama.cpp/blob/049326a/tools/llama-bench/llama-bench.cpp https://huggingface.co/unsloth/Qwen3.6-27B-MTP-GGUF/blob/5cb35eb3dcbf52dbce5f87dbc64df6aaffadcace/Qwen3.6-27B-UD-Q4_K_XL.gguf https://github.com/daimonionnn/amd-rocmfpx-for-win/blob/7de1d467c9f2e8282051d9d2332f5ef584f4fa66/llm-inference/scripts/longctx-rocm7-ab.ps1 https://github.com/daimonionnn/amd-rocmfpx-for-win/blob/7de1d467c9f2e8282051d9d2332f5ef584f4fa66/llm-inference/scripts/longctx-rocm7-128k.ps1 Structured rate fields are rounded to the form’s 0.1 token/s precision; exact source-reported means, standard deviations and all separate campaign cells are retained above.