{ "bench_run_identity": { "api_mode": "chat", "base_url": "http://127.0.0.1:18262", "created_at_utc": "2026-06-26T19:26:44.661388+00:00", "max_tokens": 512, "model": "gemma4-26b-a4b-q8", "prompt_chars": 3460, "prompt_mode": "filled-long", "prompt_preview": "You are running a deterministic Gemma B70 decode benchmark. Read the reference context, then produce a long numbered response until the token limit is reached. Do not summarize early.\n\nReference context:\nbenchmark latency memory throughput ", "prompt_sha256": "04f687a68d60254567d3252624e01a696b0c9843dbb02f5a789d75aada5e4da5", "prompt_tokens_requested": 512, "repeats": 8, "seed": 1, "usage_required": true }, "bench_summary": { "completion_tokens": { "cv": 0.0, "max": 512, "mean": 512.0, "median": 512.0, "min": 512, "stdev": 0.0 }, "completion_tokens_total": 4096, "elapsed_s": { "cv": 0.021165081568500527, "max": 5.675013536005281, "mean": 5.403568527755851, "median": 5.351254032022553, "min": 5.349647582974285, "stdev": 0.11436696865093489 }, "prompt_tokens": { "cv": 0.0, "max": 588, "mean": 588.0, "median": 588.0, "min": 588, "stdev": 0.0 }, "requests": 8, "tok_s_after_ttft": { "cv": 0.00652305761553187, "max": 103.51547512013657, "mean": 103.19340167720759, "median": 103.40955604010917, "min": 101.53177845012858, "stdev": 0.6731365046831482 }, "tok_s_wall": { "cv": 0.020377935193904878, "max": 95.70723903934984, "mean": 94.78796818050768, "median": 95.67850857385872, "min": 90.22004912439446, "stdev": 1.931583072744303 }, "ttft_s": { "cv": 0.2625473628142881, "max": 0.7288931669900194, "mean": 0.44182369100599317, "median": 0.40115232349489816, "min": 0.3991685209912248, "stdev": 0.1159996449024984 } }, "canary_pass_all": true, "canary_rows_completed": 1536, "fresh_response_validity": { "all_cached_tokens": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "cached_tokens_all_zero": true, "cached_tokens_reported": true, "headline_cached_tokens": 0, "headline_policy": "Use row0 only as fresh-response headline. Later repeated-prompt rows are support-only unless each request uses an independently fresh prompt with no reusable continuation/history.", "headline_row": 0, "headline_tok_s_after_ttft": 103.51547512013657, "headline_tok_s_wall": 90.22004912439446, "repeated_prompt_rows_support_only": true }, "headline_eligible_for_gemma_q8": true, "label": "gemma4-q8-gpu2-routecache-ctx8192-full-20260626T191746Z", "launcher_identity": { "batch_size": "1024", "cache_type_k": "f16", "cache_type_v": "f16", "ctx_size": "8192", "extra_llama_args": "--parallel 1 --cache-ram 0 --spec-type draft-mtp --spec-draft-model /mnt/fast-ai/llm-models/gemma4-26b-a4b-it-q8-gguf/MTP/gemma-4-26B-A4B-it-Q4_0-MTP.gguf --spec-draft-n-max 7 --spec-draft-device SYCL0 --spec-draft-ngl all --spec-draft-type-k f16 --spec-draft-type-v f16 --spec-draft-n-min 2 --spec-draft-p-min 0.136 --no-spec-draft-backend-sampling --spec-draft-threads 32 --spec-draft-threads-batch 32 --ctx-checkpoints 0", "flash_attn": "off", "ggml_sycl_disable_dnn": "", "ggml_sycl_disable_graph": "0", "ggml_sycl_disable_opt": "0", "ggml_sycl_enable_vmm": "0", "gpu_index": "2", "llama_cpp_commit": "c926ad098", "llama_devices": "SYCL0", "llama_gemma4_moe_fused_down_weighted_sum": "", "llama_gemma4_moe_fused_down_weighted_sum_debug": "", "llama_gemma4_moe_fused_down_weighted_sum_direct_f32_parallel_slots": "", "llama_gemma4_moe_fused_down_weighted_sum_matmul_epilogue": "", "llama_gemma4_moe_fused_down_weighted_sum_parallel_slots": "", "llama_gemma4_moe_fused_geglu_down_weighted_sum": "", "llama_gemma4_moe_geglu_down_matmul_epilogue": "", "llama_gemma4_moe_selected_softmax": "1", "llama_gemma4_moe_selected_softmax_fused": null, "llama_gemma4_moe_selected_softmax_weighted_sum": "", "llama_gemma4_moe_skip_early_weights_expand": "", "llama_gemma4_moe_sorted_top_k": "", "llama_gemma4_moe_top_k": "", "llama_gemma4_moe_weighted_sum": "1", "llama_gemma4_moe_weighted_sum_selected_softmax": "", "llama_gemma4_mtp_fused_output_argmax": "", "llama_gemma4_mtp_qonly_attn_inputs": "1", "llama_main_gpu": "", "llama_mtp_defer_target_h_nextn": "1", "llama_mtp_draft_backend_argmax": "", "llama_mtp_draft_backend_topk": "", "llama_mtp_draft_device_h_handoff": "", "llama_mtp_draft_direct_argmax_ids": "1", "llama_mtp_draft_direct_argmax_unroll": "7", "llama_mtp_draft_fast_argmax": "1", "llama_mtp_draft_fast_topk": "", "llama_mtp_draft_logit_gap_min": "", "llama_mtp_draft_profile": "", "llama_mtp_draft_terminal_logits_only": "", "llama_mtp_draft_top_k": "", "llama_server": "/home/steve/src/llama.cpp-gemma-record-stack/build-sycl-b70-aot-bmg-g31/bin/llama-server", "llama_server_spec_profile": "", "llama_spec_verify_backend_argmax_ids": "1", "llama_spec_verify_fused_output_argmax": "", "llama_spec_verify_greedy_argmax": "", "llama_spec_verify_softcap_argmax": null, "llama_split_mode": "", "llama_sycl_mul_mat_id_multi_token_fast": "", "llama_sycl_mul_mat_id_multi_token_filter": "", "llama_sycl_mul_mat_id_multi_token_grouped_q8_0": "", "llama_sycl_mul_mat_id_multi_token_name_substr": "", "llama_sycl_mul_mat_id_multi_token_no_reorder": "", "llama_sycl_mul_mat_id_multi_token_per_slot_q8_0": "", "llama_sycl_mul_mat_id_route_cache": "1", "llama_tensor_split": "", "oneapi_device_selector": "level_zero:2", "poll": "100", "port": "18262", "reasoning": "off", "threads": "8", "ubatch_size": "1024" }, "model_file_bytes": 27636230944, "model_path": "/mnt/fast-ai/llm-models/gemma4-26b-a4b-it-q8-gguf/gemma-4-26B-A4B-it-UD-Q8_K_XL.gguf", "quality_lane": "q8_target", "run_dir": "/home/steve/qwen36-results-main/data/gemma4-q8-gpu2-routecache-ctx8192-full-20260626T191746Z", "server_log": "/mnt/fast-ai/bench-results/gemma4-26b-a4b-q8/servers/gemma4-q8-gpu2-routecache-ctx8192-full-20260626T191746Z.server.log" }