{ "bench_kind": "realistic_final_gate", "bench_path": "/home/steve/qwen36-results-main/data/gemma4-q8-gpu0-strict-vdr2-grouped-reorder-n3-p00475-ub1024-screen-20260627T205542Z/realistic-suite.json", "bench_rc": 0, "bench_run_identity": { "api_mode": "chat", "base_url": "http://127.0.0.1:18390", "created_at_utc": "2026-06-27T20:57:25.152386+00:00", "max_tokens": 512, "model": "gemma4-26b-a4b-q8", "prompt_count": 12, "seed": 1, "suite": { "description": "Fixed cold-response prompt suite for Gemma 4 26B A4B Q8 B70 promotion and LocalMaxxing submissions. Each prompt is sent exactly once. Synthetic/repeated prompts remain diagnostic only.", "metric": "median generated-token throughput for tokens 1-100 after TTFT across prompts", "suite_id": "gemma4-26b-a4b-q8-b70-realistic-v1", "version": 1 }, "suite_path": "/home/steve/qwen36-results-main/repro/gemma4-26b-a4b-q8-b70/realistic-suite-v1.json" }, "bench_summary": { "tok_s_1_100_after_ttft": { "count": 12, "max": 98.43628100359264, "mean": 86.02284187160546, "median": 83.90758854375754, "min": 74.93484368345916, "p10": 78.17221426541892, "stdev": 7.755437144116298 }, "tok_s_after_ttft_full": { "count": 12, "max": 86.80332102696435, "mean": 81.43008546495325, "median": 82.00778999411864, "min": 76.02309048465051, "p10": 77.2358578361824, "stdev": 3.4498591572213386 }, "tok_s_wall_full": { "count": 12, "max": 84.3233523608412, "mean": 78.90075935483047, "median": 79.09301505839233, "min": 74.11337899785764, "p10": 75.00120605131082, "stdev": 3.105617203191548 }, "ttft_ms": { "count": 12, "max": 280.30522004701197, "mean": 189.58008658955805, "median": 179.04035048559308, "min": 167.1963930130005, "p10": 172.63491702033207, "stdev": 30.431726466898773 } }, "canary_pass_all": true, "canary_rows_completed": 32, "fresh_response_validity": { "cached_tokens": [ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ], "cached_tokens_all_zero": true, "classification": "fresh-response", "context_checkpoints_or_prefix_reuse": false, "each_prompt_run_once": true, "history_acceleration": false, "ngram_history_acceleration": false, "note": "Fixed realistic suite; each prompt is sent once as a cold response. Do not average synthetic repeated prompts, warmed continuations, or n-gram/history-accelerated rows into this metric.", "primary_metric_name": "median_tok_s_1_100_after_ttft", "primary_metric_tokens": 100, "prompt_count": 12, "prompts_are_unique": true, "response_reuse": false, "suite_id": "gemma4-26b-a4b-q8-b70-realistic-v1", "suite_version": 1, "token_timing_source": "openai_stream_text_delta", "valid": true }, "headline_eligible_for_gemma_q8": true, "label": "gemma4-q8-gpu0-strict-vdr2-grouped-reorder-n3-p00475-ub1024-screen-20260627T205542Z", "launcher_identity": { "batch_size": "1024", "cache_type_k": null, "cache_type_v": null, "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 3 --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 0.0475 --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": "0", "llama_cpp_commit": "c926ad098", "llama_devices": "SYCL0", "llama_gemma4_fused_final_post_norm_residual": "", "llama_gemma4_moe_fused_branch_post_norm_add": "", "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_fused_router_selected_weights": "", "llama_gemma4_moe_gateup_geglu": "", "llama_gemma4_moe_geglu_down_matmul_epilogue": "", "llama_gemma4_moe_reuse_attn_rms": "1", "llama_gemma4_moe_router_post_scale": "", "llama_gemma4_moe_selected_softmax": "1", "llama_gemma4_moe_selected_softmax_fused": "1", "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_2d": "", "llama_gemma4_moe_weighted_sum_selected_softmax": "", "llama_gemma4_mtp_fused_output_argmax": "1", "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_scores": "", "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-repro-c926/build-sycl-b70-aot-bmg-g31-q8reorder-vdr2/bin/llama-server", "llama_server_spec_profile": "", "llama_spec_adaptive_mtp": "", "llama_spec_adaptive_mtp_alpha": "", "llama_spec_adaptive_mtp_high": "", "llama_spec_adaptive_mtp_low": "", "llama_spec_adaptive_mtp_low_n_max": "", "llama_spec_adaptive_mtp_warmup": "", "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_f16_p021_small_ncols": "", "llama_sycl_mul_mat_id_gate_up_fast_seed_route_cache": "", "llama_sycl_mul_mat_id_gate_up_q8_singleton_direct": "", "llama_sycl_mul_mat_id_multi_token_fast": "1", "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_q8_0_reorder": "1", "llama_sycl_mul_mat_id_route_cache": "1", "llama_sycl_mul_mat_id_route_cache_device_map": "", "llama_sycl_mul_mat_id_route_cache_inplace": "", "llama_sycl_mul_mat_id_route_timing": "", "llama_sycl_mul_mat_id_route_timing_every": "", "llama_sycl_q8_mmvq_small_ncols": "", "llama_tensor_split": "", "oneapi_device_selector": "level_zero:0", "poll": "100", "port": "18390", "realistic_gate": "1", "realistic_metric_tokens": "100", "realistic_suite": null, "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", "realistic_final_gate": { "cached_tokens": [ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ], "cached_tokens_all_zero": true, "chunk_count_matches_completion_tokens_all": false, "chunk_count_matches_completion_tokens_note": "Informational only: llama.cpp usage may count an EOS/final token that is not emitted as a text delta. Promotion requires enough streamed text deltas to measure the first metric window.", "chunk_counts": [ 512, 512, 301, 512, 512, 512, 512, 512, 512, 512, 512, 512 ], "completion_tokens_at_least_metric_window": true, "metric_chunk_events_at_least_window": true, "metric_name": "median_tok_s_1_100_after_ttft", "metric_tokens": 100, "passed": true, "prompts_unique": true, "required_policy": "fixed realistic prompt suite; each prompt once; cached_tokens=0 every row; no repeated/warmed prompt averaging; metric is median tokens 1-100 after TTFT", "token_timing_source": "openai_stream_text_delta" }, "run_dir": "/home/steve/qwen36-results-main/data/gemma4-q8-gpu0-strict-vdr2-grouped-reorder-n3-p00475-ub1024-screen-20260627T205542Z", "server_log": "/mnt/fast-ai/bench-results/gemma4-26b-a4b-q8/servers/gemma4-q8-gpu0-strict-vdr2-grouped-reorder-n3-p00475-ub1024-screen-20260627T205542Z.server.log" }