{ "recorded_at": "2026-09-07T13:20:00+01:00", "host": { "name": "evox3", "system": "GMKtec EVO-X3", "processor": "AMD Ryzen AI MAX+ 395", "gpu": "Radeon 8060S", "vulkan_device": "RADV STRIX_HALO", "memory_gib": 128 }, "production_control": { "model_name": "Qwen3.8-Flash-Next-Q4KXL-PR27742-Fixed", "llama_cpp_commit": "39817c476489c37747854bebc29d375770618f1c", "llama_server_sha256": "27c7cb0668c6ac053f7958f76a7e1d544788b6b9740d75695379084255f7c04a", "vulkan_bundle": "v0.7.1", "lemonade_version": "11.9.0", "screening": { "prefill_8192_tokens_per_second": 77.27658840477979, "prefill_32768_tokens_per_second": 85.74292847098909, "single_decode_tokens_per_second": 15.698330957583726, "two_stream_aggregate_tokens_per_second": 6.811377411589368, "note": "Sanity control only. The production source predates the row-hoisting machinery required by PR 28501, so PR effects are attributed only from exact upstream base/head pairs." } }, "model": { "repository": "unsloth/Qwen3.8-Flash-Next-GGUF", "quantisation": "UD-Q4_K_XL", "total_bytes": 111334654784, "total_gib": 103.688477, "shards": [ {"file": "Qwen3.8-Flash-Next-UD-Q4_K_XL-00001-of-00004.gguf", "bytes": 10946624}, {"file": "Qwen3.8-Flash-Next-UD-Q4_K_XL-00002-of-00004.gguf", "bytes": 49859583136}, {"file": "Qwen3.8-Flash-Next-UD-Q4_K_XL-00003-of-00004.gguf", "bytes": 49376141504}, {"file": "Qwen3.8-Flash-Next-UD-Q4_K_XL-00004-of-00004.gguf", "bytes": 12087983520} ] }, "profile": { "total_context_tokens": 524288, "parallel_slots": 2, "context_tokens_per_slot": 262144, "gpu_layers": 40, "batch_tokens": 512, "micro_batch_tokens": 128, "flash_attention": true, "kv_key_type": "q8_0", "kv_value_type": "q8_0", "host_prompt_cache_mib": 4096, "prompt_caching": true, "idle_slot_caching": false, "load_mode": "mmap", "per_layer_token_embedding": "CPU", "speculative_decoding": "none" }, "toolchain": { "compiler": "GNU 13.3.0", "build_type": "Release", "ggml_native": false, "llama_curl": false, "glslc": "shaderc v2026.3", "glslc_sha256": "b91ab31aa56c7a6034a242e483f863ca6f9606571ba1cfb329ef6cb79a40234b" }, "pull_requests": { "28489": { "title": "vulkan: make MMVQ path selection independent of batch size", "url": "https://github.com/ggml-org/llama.cpp/pull/28489", "state_at_test": "OPEN", "draft_at_test": false, "merge_state_at_test": "BLOCKED", "base_commit": "73a43d1f69345aee8bb186ef4b3172cef892f2e5", "head_commit": "ef94c0910bc1d6b9bd28507dd7d1f3904ac7cf62", "binary_diff_sha256": "11df07ddeefcf62fdda2e6f66efc09740e44d04cd4f70c3e50bc4412556f8ed7", "changed_files": [ "ggml/src/ggml-vulkan/ggml-vulkan.cpp", "tests/test-backend-ops.cpp" ], "base_llama_server_sha256": "7b24821f66ac9b457c6c3e4766863b69770790a8002b125f172f7397a0d4dbe0", "head_llama_server_sha256": "258a88040ea54f32b50001e1c15d4c21bf399c1b0ca66b34727d798821ce8532", "base_libggml_vulkan_sha256": "182ab58d6562fbae5050e6e3c1d2137a2dce83845a2323dee101bde99cf7e1f3", "head_libggml_vulkan_sha256": "aa37d62890eba6c93e972538d7c1a96163320813b3ffde7112764f306b48c2de" }, "28501": { "title": "vulkan: raise the hoisted row-id limit for mul_mat_id from 256 to 512 experts", "url": "https://github.com/ggml-org/llama.cpp/pull/28501", "state_at_test": "OPEN", "draft_at_test": true, "merge_state_at_test": "BLOCKED", "base_commit": "465e49b9cea78a68b9c244ffb48d0ee24a82873d", "head_commit": "266464c8a331279e157603d9a8728cc79ed5860a", "binary_diff_sha256": "b0c1da46a5630667ee7a73ed8be1d0d2d38a901673d7ab0f55e4d181be5f046e", "changed_files": [ "ggml/src/ggml-vulkan/ggml-vulkan.cpp", "ggml/src/ggml-vulkan/vulkan-shaders/count_experts.comp", "tests/test-backend-ops.cpp" ], "base_llama_server_sha256": "7c5cadd4b0f646024c3a9c4aff9ae4d4a5a2d49514745d3db4bde65721d1b01a", "head_llama_server_sha256": "e5f9ce100a211667febd77af8ab57ac0b564db10314c353730c0f9fab1ed6f3d", "base_libggml_vulkan_sha256": "e00319e20bccf06b64955245a42f71d6da68f97ce2681df18939a2a2e0a76f66", "head_libggml_vulkan_sha256": "dbf889087e07da7529e5031b544c829505f536c9ea9ce2759afd6c93366c5916" } }, "method": { "materiality_threshold_percent": 5, "sampler": {"temperature": 0, "top_k": 1, "top_p": 1, "seed": 2848928501}, "request_cache_prompt": false, "distinct_prefix_per_measured_request": true, "prefill_repetitions_per_pair": 2, "pairing": "base/head compared at the same first or warmed ordinal", "start_gate": {"minimum_mem_available_kib": 115000000, "maximum_gtt_bytes": 1073741824}, "runtime_guard": {"minimum_mem_available_gib": 18, "maximum_gtt_gib": 96} }, "results": { "28489": { "prefill_8192": { "base": [77.73598989288107, 91.68077126460001], "head": [78.36019580191842, 92.94228001188809], "paired_delta_percent": [0.802982, 1.375980], "base_median": 84.70838057874053, "head_median": 85.65123790690325, "median_delta_percent": 1.113063, "outputs_match": true }, "screening": { "single_decode_base": 16.96157110569581, "single_decode_head": 16.972706289408503, "single_decode_delta_percent": 0.065649, "two_stream_aggregate_base": 7.0476923292519835, "two_stream_aggregate_head": 6.920061313312635, "two_stream_delta_percent": -1.810962, "note": "Single screening pass; generation text remained coherent but exact generation hashes changed with the kernel route." }, "decision": "No material production performance benefit" }, "28501": { "prefill_8192": { "base": [77.95161920022113, 92.56029690148088], "head": [77.84748182848666, 93.21662272812604], "paired_delta_percent": [-0.133592, 0.709079], "base_median": 85.25595805085100, "head_median": 85.53205227830635, "median_delta_percent": 0.323842, "outputs_match": true }, "prefill_32768": { "base": [84.54563773783006, 88.17253246094780], "head": [85.03084634026213, 88.13454963421245], "paired_delta_percent": [0.573901, -0.043078], "base_median": 86.35908509938893, "head_median": 86.58269798723728, "median_delta_percent": 0.258934, "outputs_match": true }, "screening": { "single_decode_base": 16.143651357025973, "single_decode_head": 16.567954719779753, "single_decode_delta_percent": 2.628299, "note": "Single screening pass; below the materiality threshold and outside the changed prefill path." }, "vulkan_mul_mat_id_tests": {"executed": 891, "command_exit_code": 0}, "decision": "No material production performance benefit at batch 512 / micro-batch 128" } }, "resources": { "28489": { "base_min_mem_available_gib": 47.855, "head_min_mem_available_gib": 48.295, "base_peak_gtt_gib": 70.718, "head_peak_gtt_gib": 70.718 }, "28501": { "base_min_mem_available_gib": 46.195, "head_min_mem_available_gib": 46.872, "base_peak_gtt_gib": 70.783, "head_peak_gtt_gib": 70.827 }, "guard_stops": 0, "gpu_faults_observed": 0 }, "recovery": { "production_restored": true, "production_model_ready": true, "production_model_pinned": false, "production_auto_evict": false, "qwen35_npu_resident": true, "qwen38_warmer_active": true, "qwen38_residency_guard_active": true, "inhibitor_removed": true }, "boundaries": [ "One host and one Q4_K_XL model", "No 262K prefill repetition", "Speculative decoding disabled, matching production", "Not a perplexity or broad answer-quality evaluation", "PR 28489 focused test filter matched no cases and is excluded from claims", "PR 28501 upstream author used Q5_K with batch 2048; this record does not refute that configuration" ] }