Datasets:
Download data/benchmarks.json from kleinbem/layerstream-orin-nano-benchmarks: direct link, hf CLI and curl.
- Browser
- Download file 317 kB
-
https://huggingface.co/datasets/kleinbem/layerstream-orin-nano-benchmarks/resolve/main/data/benchmarks.json
- Command line
-
hf download hf://datasets/kleinbem/layerstream-orin-nano-benchmarks/data/benchmarks.json
-
curl -L -o benchmarks.json https://huggingface.co/datasets/kleinbem/layerstream-orin-nano-benchmarks/resolve/main/data/benchmarks.json
317 kB
[
{
"round": "airllm",
"source_file": "airllm_14b_4bit_code_edit_warm.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "airllm",
"engine_version": "4.0.0",
"git_rev": null,
"mode": "airllm_4bit_prefetch",
"status": "ok",
"model": "Qwen/Qwen2.5-14B-Instruct",
"draft_model": null,
"workload": "code_edit",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 113,
"completion_tokens": 8,
"load_time_s": 2.55,
"ttft_s": 33.371,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 214.529,
"total_time_s": 247.9,
"decode_tok_per_s": 0.033,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.032,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": null,
"drafted_tokens": null,
"accepted_tokens": null,
"resident_layers": null,
"streamed_layers": null,
"n_layers": null,
"resident_mib": null,
"slot_mib": null,
"expert_slot_mib": null,
"io_mib_read": null,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": null,
"is_unified_memory": null,
"direct_io_active": null,
"peak_rss_mib": 3536.4,
"current_rss_mib": 626.8,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": 1936.2,
"page_cache_delta_mib": -503.0,
"page_cache_delta_since_start_mib": -458.6,
"idle_watts": null,
"avg_watts": 9.3,
"net_watts": null,
"joules_per_tok": 288.21,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": "#[derive(Debug, Clone)]\npub struct"
},
{
"round": "airllm",
"source_file": "airllm_14b_4bit_open_ended_warm.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "airllm",
"engine_version": "4.0.0",
"git_rev": null,
"mode": "airllm_4bit_prefetch",
"status": "ok",
"model": "Qwen/Qwen2.5-14B-Instruct",
"draft_model": null,
"workload": "open_ended",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 31,
"completion_tokens": 8,
"load_time_s": 2.573,
"ttft_s": 32.863,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 213.357,
"total_time_s": 246.219,
"decode_tok_per_s": 0.033,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.032,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": null,
"drafted_tokens": null,
"accepted_tokens": null,
"resident_layers": null,
"streamed_layers": null,
"n_layers": null,
"resident_mib": null,
"slot_mib": null,
"expert_slot_mib": null,
"io_mib_read": null,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": null,
"is_unified_memory": null,
"direct_io_active": null,
"peak_rss_mib": 3531.2,
"current_rss_mib": 617.0,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": 1920.8,
"page_cache_delta_mib": -640.9,
"page_cache_delta_since_start_mib": -591.8,
"idle_watts": null,
"avg_watts": 9.33,
"net_watts": null,
"joules_per_tok": 287.1,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " When discussing the interaction between page-aligned O"
},
{
"round": "airllm",
"source_file": "airllm_14b_4bit_prompt_heavy.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "airllm",
"engine_version": "4.0.0",
"git_rev": null,
"mode": "airllm_4bit_prefetch",
"status": "ok",
"model": "Qwen/Qwen2.5-14B-Instruct",
"draft_model": null,
"workload": "prompt_heavy",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 340,
"completion_tokens": 32,
"load_time_s": 365.529,
"ttft_s": 32.093,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 954.918,
"total_time_s": 987.011,
"decode_tok_per_s": 0.032,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.032,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": null,
"drafted_tokens": null,
"accepted_tokens": null,
"resident_layers": null,
"streamed_layers": null,
"n_layers": null,
"resident_mib": null,
"slot_mib": null,
"expert_slot_mib": null,
"io_mib_read": null,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": null,
"is_unified_memory": null,
"direct_io_active": null,
"peak_rss_mib": 5390.5,
"current_rss_mib": 689.2,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": 1983.3,
"page_cache_delta_mib": 248.0,
"page_cache_delta_since_start_mib": -751.8,
"idle_watts": null,
"avg_watts": null,
"net_watts": null,
"joules_per_tok": null,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " The core memory-management mechanism described is layer-streaming inference, which minimizes memory usage by allocating a small fixed pool of reusable buffers "
},
{
"round": "airllm",
"source_file": "airllm_14b_bf16_open_ended.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "airllm",
"engine_version": "4.0.0",
"git_rev": null,
"mode": "airllm_none_prefetch",
"status": "ok",
"model": "Qwen/Qwen2.5-14B-Instruct",
"draft_model": null,
"workload": "open_ended",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 31,
"completion_tokens": 4,
"load_time_s": 81.81,
"ttft_s": 45.413,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 131.774,
"total_time_s": 177.187,
"decode_tok_per_s": 0.023,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.023,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": null,
"drafted_tokens": null,
"accepted_tokens": null,
"resident_layers": null,
"streamed_layers": null,
"n_layers": null,
"resident_mib": null,
"slot_mib": null,
"expert_slot_mib": null,
"io_mib_read": null,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": null,
"is_unified_memory": null,
"direct_io_active": null,
"peak_rss_mib": 5217.6,
"current_rss_mib": 3489.7,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": 1502.0,
"page_cache_delta_mib": -4310.7,
"page_cache_delta_since_start_mib": -3140.9,
"idle_watts": null,
"avg_watts": 9.31,
"net_watts": null,
"joules_per_tok": 412.35,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " In the context of"
},
{
"round": "competitors",
"source_file": "hf_accelerate_14b_4bit.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "hf-accelerate",
"engine_version": "transformers-5.23.0+bnb-0.50.2",
"git_rev": null,
"mode": "hf_accelerate_14b_4bit_disk",
"status": "timeout_300s",
"model": "Qwen2.5-14B-Instruct",
"draft_model": null,
"workload": "open_ended",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": null,
"completion_tokens": 0,
"load_time_s": 144.0,
"ttft_s": null,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": null,
"total_time_s": 300.0,
"decode_tok_per_s": 0.0,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.0,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": null,
"drafted_tokens": null,
"accepted_tokens": null,
"resident_layers": null,
"streamed_layers": null,
"n_layers": null,
"resident_mib": null,
"slot_mib": null,
"expert_slot_mib": null,
"io_mib_read": null,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": null,
"is_unified_memory": null,
"direct_io_active": null,
"peak_rss_mib": null,
"current_rss_mib": null,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": null,
"page_cache_delta_since_start_mib": null,
"idle_watts": null,
"avg_watts": null,
"net_watts": null,
"joules_per_tok": null,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": "Loaded and disk-offloaded 579 tensors (4-bit NF4, gpu_max_mem=3GiB, cpu_max_mem=1GiB) in 144.0s (2m24s), then hit 300s wall-clock timeout during generation of 16 tokens. PyTorch 2.14.1+cu126 wheel warns Orin CC 8.7 unsupported (same venv as AirLLM 4.0.0).",
"output_preview": null
},
{
"round": "competitors",
"source_file": "llama_cpp_14b_cpu_mmap.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "llama.cpp",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "llama_cpp_14b_cpu_mmap",
"status": "timeout",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": null,
"workload": "open_ended",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": "q8_0",
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": null,
"completion_tokens": 0,
"load_time_s": null,
"ttft_s": null,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 300.458,
"total_time_s": 300.458,
"decode_tok_per_s": 0.0,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.0,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": null,
"drafted_tokens": null,
"accepted_tokens": null,
"resident_layers": null,
"streamed_layers": null,
"n_layers": null,
"resident_mib": null,
"slot_mib": null,
"expert_slot_mib": null,
"io_mib_read": null,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": null,
"is_unified_memory": null,
"direct_io_active": null,
"peak_rss_mib": 6109.6,
"current_rss_mib": null,
"mem_avail_before_mib": 5040.2,
"mem_avail_after_mib": null,
"swap_delta_mib": 362.6,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 5648.3,
"page_cache_delta_since_start_mib": null,
"idle_watts": 7.87,
"avg_watts": 10.61,
"net_watts": null,
"joules_per_tok": null,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": "When streaming transformer layers on an 8 GB edge device, the interaction between page-aligned O_DIRECT NVMe reads and asynchronous ping-pong GPU buffers is cru"
},
{
"round": "competitors",
"source_file": "llama_cpp_14b_ngl0.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "llama.cpp",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "llama_cpp_14b_ngl0",
"status": "error_rc_1",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": null,
"workload": "open_ended",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": "q8_0",
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": null,
"completion_tokens": 0,
"load_time_s": null,
"ttft_s": null,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 22.832,
"total_time_s": 22.832,
"decode_tok_per_s": 0.0,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.0,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": null,
"drafted_tokens": null,
"accepted_tokens": null,
"resident_layers": null,
"streamed_layers": null,
"n_layers": null,
"resident_mib": null,
"slot_mib": null,
"expert_slot_mib": null,
"io_mib_read": null,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": null,
"is_unified_memory": null,
"direct_io_active": null,
"peak_rss_mib": 5991.0,
"current_rss_mib": null,
"mem_avail_before_mib": 4984.7,
"mem_avail_after_mib": null,
"swap_delta_mib": 3683.5,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 135.7,
"page_cache_delta_since_start_mib": null,
"idle_watts": 7.79,
"avg_watts": 9.9,
"net_watts": null,
"joules_per_tok": null,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": null
},
{
"round": "competitors",
"source_file": "llama_cpp_14b_ngl12_b512.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "llama.cpp",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "llama_cpp_14b_ngl12_b512",
"status": "error_rc_1",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": null,
"workload": "prompt_heavy",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": "q8_0",
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": null,
"completion_tokens": 0,
"load_time_s": null,
"ttft_s": null,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 23.562,
"total_time_s": 23.562,
"decode_tok_per_s": 0.0,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.0,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": null,
"drafted_tokens": null,
"accepted_tokens": null,
"resident_layers": null,
"streamed_layers": null,
"n_layers": null,
"resident_mib": null,
"slot_mib": null,
"expert_slot_mib": null,
"io_mib_read": null,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": null,
"is_unified_memory": null,
"direct_io_active": null,
"peak_rss_mib": 5977.9,
"current_rss_mib": null,
"mem_avail_before_mib": 5021.0,
"mem_avail_after_mib": null,
"swap_delta_mib": 3676.6,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 76.5,
"page_cache_delta_since_start_mib": null,
"idle_watts": 7.97,
"avg_watts": 9.84,
"net_watts": null,
"joules_per_tok": null,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": null
},
{
"round": "competitors",
"source_file": "llama_cpp_14b_ngl12_b512.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "llama.cpp",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "llama_cpp_14b_ngl12_b512",
"status": "error_rc_1",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": null,
"workload": "code_edit",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": "q8_0",
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": null,
"completion_tokens": 0,
"load_time_s": null,
"ttft_s": null,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 23.752,
"total_time_s": 23.752,
"decode_tok_per_s": 0.0,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.0,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": null,
"drafted_tokens": null,
"accepted_tokens": null,
"resident_layers": null,
"streamed_layers": null,
"n_layers": null,
"resident_mib": null,
"slot_mib": null,
"expert_slot_mib": null,
"io_mib_read": null,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": null,
"is_unified_memory": null,
"direct_io_active": null,
"peak_rss_mib": 5742.9,
"current_rss_mib": null,
"mem_avail_before_mib": 5163.8,
"mem_avail_after_mib": null,
"swap_delta_mib": 3684.2,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -5.9,
"page_cache_delta_since_start_mib": null,
"idle_watts": 7.97,
"avg_watts": 9.89,
"net_watts": null,
"joules_per_tok": null,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": null
},
{
"round": "competitors",
"source_file": "llama_cpp_14b_ngl12_b512.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "llama.cpp",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "llama_cpp_14b_ngl12_b512",
"status": "error_rc_1",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": null,
"workload": "open_ended",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": "q8_0",
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": null,
"completion_tokens": 0,
"load_time_s": null,
"ttft_s": null,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 23.445,
"total_time_s": 23.445,
"decode_tok_per_s": 0.0,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.0,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": null,
"drafted_tokens": null,
"accepted_tokens": null,
"resident_layers": null,
"streamed_layers": null,
"n_layers": null,
"resident_mib": null,
"slot_mib": null,
"expert_slot_mib": null,
"io_mib_read": null,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": null,
"is_unified_memory": null,
"direct_io_active": null,
"peak_rss_mib": 5723.5,
"current_rss_mib": null,
"mem_avail_before_mib": 5125.5,
"mem_avail_after_mib": null,
"swap_delta_mib": 3710.8,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 53.4,
"page_cache_delta_since_start_mib": null,
"idle_watts": 7.97,
"avg_watts": 9.95,
"net_watts": null,
"joules_per_tok": null,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": null
},
{
"round": "competitors",
"source_file": "llama_cpp_14b_ngl18.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "llama.cpp",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "llama_cpp_14b_ngl18",
"status": "error_rc_1",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": null,
"workload": "prompt_heavy",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": "q8_0",
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": null,
"completion_tokens": 0,
"load_time_s": null,
"ttft_s": null,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 56.129,
"total_time_s": 56.129,
"decode_tok_per_s": 0.0,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.0,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": null,
"drafted_tokens": null,
"accepted_tokens": null,
"resident_layers": null,
"streamed_layers": null,
"n_layers": null,
"resident_mib": null,
"slot_mib": null,
"expert_slot_mib": null,
"io_mib_read": null,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": null,
"is_unified_memory": null,
"direct_io_active": null,
"peak_rss_mib": 5776.3,
"current_rss_mib": null,
"mem_avail_before_mib": 4972.8,
"mem_avail_after_mib": null,
"swap_delta_mib": 3634.6,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -2317.2,
"page_cache_delta_since_start_mib": null,
"idle_watts": 7.92,
"avg_watts": 9.23,
"net_watts": null,
"joules_per_tok": null,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": null
},
{
"round": "competitors",
"source_file": "llama_cpp_14b_ngl18.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "llama.cpp",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "llama_cpp_14b_ngl18",
"status": "error_rc_1",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": null,
"workload": "code_edit",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": "q8_0",
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": null,
"completion_tokens": 0,
"load_time_s": null,
"ttft_s": null,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 22.412,
"total_time_s": 22.412,
"decode_tok_per_s": 0.0,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.0,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": null,
"drafted_tokens": null,
"accepted_tokens": null,
"resident_layers": null,
"streamed_layers": null,
"n_layers": null,
"resident_mib": null,
"slot_mib": null,
"expert_slot_mib": null,
"io_mib_read": null,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": null,
"is_unified_memory": null,
"direct_io_active": null,
"peak_rss_mib": 5736.2,
"current_rss_mib": null,
"mem_avail_before_mib": 5027.8,
"mem_avail_after_mib": null,
"swap_delta_mib": 3626.0,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 114.3,
"page_cache_delta_since_start_mib": null,
"idle_watts": 7.92,
"avg_watts": 9.97,
"net_watts": null,
"joules_per_tok": null,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": null
},
{
"round": "competitors",
"source_file": "llama_cpp_14b_ngl18.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "llama.cpp",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "llama_cpp_14b_ngl18",
"status": "error_rc_1",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": null,
"workload": "open_ended",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": "q8_0",
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": null,
"completion_tokens": 0,
"load_time_s": null,
"ttft_s": null,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 54.08,
"total_time_s": 54.08,
"decode_tok_per_s": 0.0,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.0,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": null,
"drafted_tokens": null,
"accepted_tokens": null,
"resident_layers": null,
"streamed_layers": null,
"n_layers": null,
"resident_mib": null,
"slot_mib": null,
"expert_slot_mib": null,
"io_mib_read": null,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": null,
"is_unified_memory": null,
"direct_io_active": null,
"peak_rss_mib": 5715.3,
"current_rss_mib": null,
"mem_avail_before_mib": 5078.5,
"mem_avail_after_mib": null,
"swap_delta_mib": 3556.4,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -86.6,
"page_cache_delta_since_start_mib": null,
"idle_watts": 7.92,
"avg_watts": 8.73,
"net_watts": null,
"joules_per_tok": null,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": null
},
{
"round": "competitors",
"source_file": "llama_cpp_32b_ngl10.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "llama.cpp",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "llama_cpp_32b_ngl10",
"status": "timeout",
"model": "Qwen2.5-32B-Instruct-Q4_K_M.gguf",
"draft_model": null,
"workload": "open_ended",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": "q8_0",
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": null,
"completion_tokens": 0,
"load_time_s": null,
"ttft_s": null,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 302.828,
"total_time_s": 302.828,
"decode_tok_per_s": 0.0,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.0,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": null,
"drafted_tokens": null,
"accepted_tokens": null,
"resident_layers": null,
"streamed_layers": null,
"n_layers": null,
"resident_mib": null,
"slot_mib": null,
"expert_slot_mib": null,
"io_mib_read": null,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": null,
"is_unified_memory": null,
"direct_io_active": null,
"peak_rss_mib": 5714.7,
"current_rss_mib": null,
"mem_avail_before_mib": 5176.0,
"mem_avail_after_mib": null,
"swap_delta_mib": 17242.8,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -305.4,
"page_cache_delta_since_start_mib": null,
"idle_watts": 7.89,
"avg_watts": 10.29,
"net_watts": null,
"joules_per_tok": null,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": "When streaming transformer layers"
},
{
"round": "competitors",
"source_file": "llama_cpp_qwen3_30b_cpu_mmap.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "llama.cpp",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "llama_cpp_qwen3_30b_cpu_mmap",
"status": "error_rc_1",
"model": "Qwen3-30B-A3B-Q4_K_M.gguf",
"draft_model": null,
"workload": "open_ended",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": "q8_0",
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": null,
"completion_tokens": 0,
"load_time_s": null,
"ttft_s": null,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 66.223,
"total_time_s": 66.223,
"decode_tok_per_s": 0.0,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.0,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": null,
"drafted_tokens": null,
"accepted_tokens": null,
"resident_layers": null,
"streamed_layers": null,
"n_layers": null,
"resident_mib": null,
"slot_mib": null,
"expert_slot_mib": null,
"io_mib_read": null,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": null,
"is_unified_memory": null,
"direct_io_active": null,
"peak_rss_mib": 6044.1,
"current_rss_mib": null,
"mem_avail_before_mib": 5115.9,
"mem_avail_after_mib": null,
"swap_delta_mib": 12324.5,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -62.8,
"page_cache_delta_since_start_mib": null,
"idle_watts": 7.92,
"avg_watts": 10.16,
"net_watts": null,
"joules_per_tok": null,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": null
},
{
"round": "competitors",
"source_file": "llama_cpp_qwen3_30b_cpu_mmap_b512.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "llama.cpp",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "llama_cpp_qwen3_30b_cpu_mmap_b512",
"status": "ok",
"model": "Qwen3-30B-A3B-Q4_K_M.gguf",
"draft_model": null,
"workload": "open_ended",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": "q8_0",
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 31,
"completion_tokens": 23,
"load_time_s": 25.34,
"ttft_s": 17.367,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 18.113,
"total_time_s": 62.445,
"decode_tok_per_s": 1.27,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.648,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": null,
"drafted_tokens": null,
"accepted_tokens": null,
"resident_layers": null,
"streamed_layers": null,
"n_layers": null,
"resident_mib": null,
"slot_mib": null,
"expert_slot_mib": null,
"io_mib_read": null,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": null,
"is_unified_memory": null,
"direct_io_active": null,
"peak_rss_mib": 6070.9,
"current_rss_mib": null,
"mem_avail_before_mib": 6137.6,
"mem_avail_after_mib": null,
"swap_delta_mib": 247.9,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -100.1,
"page_cache_delta_since_start_mib": null,
"idle_watts": 8.07,
"avg_watts": 9.89,
"net_watts": null,
"joules_per_tok": 15.26,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": "the problem is that the GPU is not able to read the data from the host memory as it is being written by the"
},
{
"round": "competitors",
"source_file": "llama_cpp_qwen3_30b_gpu_offload.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "llama.cpp",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "llama_cpp_qwen3_30b_gpu_moe_offload",
"status": "error_rc_1",
"model": "Qwen3-30B-A3B-Q4_K_M.gguf",
"draft_model": null,
"workload": "open_ended",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": "q8_0",
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": null,
"completion_tokens": 0,
"load_time_s": null,
"ttft_s": null,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 54.427,
"total_time_s": 54.427,
"decode_tok_per_s": 0.0,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.0,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": null,
"drafted_tokens": null,
"accepted_tokens": null,
"resident_layers": null,
"streamed_layers": null,
"n_layers": null,
"resident_mib": null,
"slot_mib": null,
"expert_slot_mib": null,
"io_mib_read": null,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": null,
"is_unified_memory": null,
"direct_io_active": null,
"peak_rss_mib": 6014.0,
"current_rss_mib": null,
"mem_avail_before_mib": 6081.8,
"mem_avail_after_mib": null,
"swap_delta_mib": 12342.2,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 81.1,
"page_cache_delta_since_start_mib": null,
"idle_watts": 7.95,
"avg_watts": 10.12,
"net_watts": null,
"joules_per_tok": null,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": null
},
{
"round": "results",
"source_file": "r0_prompt_heavy.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "stream_r0",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": null,
"workload": "prompt_heavy",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 340,
"completion_tokens": 32,
"load_time_s": 2.968,
"ttft_s": 9.257,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 228.946,
"total_time_s": 238.203,
"decode_tok_per_s": 0.135,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.134,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 32,
"drafted_tokens": 0,
"accepted_tokens": 0,
"resident_layers": 0,
"streamed_layers": 48,
"n_layers": 48,
"resident_mib": 1026.8,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 241263.38,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 3225.7,
"current_rss_mib": 3225.7,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 37.9,
"page_cache_delta_since_start_mib": -764.6,
"idle_watts": null,
"avg_watts": null,
"net_watts": null,
"joules_per_tok": null,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " Layer-streaming inference manages memory by allocating a small, fixed pool of buffers for compute, keeping essential layers in memory, and asynchronously strea"
},
{
"round": "results",
"source_file": "r14_draft_prompt_heavy.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "hybrid_r14_draft05b",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "prompt_heavy",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 340,
"completion_tokens": 32,
"load_time_s": 52.833,
"ttft_s": 5.752,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 39.649,
"total_time_s": 45.401,
"decode_tok_per_s": 0.782,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.705,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 15,
"drafted_tokens": 63,
"accepted_tokens": 17,
"resident_layers": 14,
"streamed_layers": 34,
"n_layers": 48,
"resident_mib": 3244.4,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 79826.54,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 5877.1,
"current_rss_mib": 5666.0,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -251.4,
"page_cache_delta_since_start_mib": -2667.7,
"idle_watts": null,
"avg_watts": null,
"net_watts": null,
"joules_per_tok": null,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " Layer-streaming inference manages memory by allocating a small, fixed pool of buffers for active computation, keeping only essential layers in memory, and asyn"
},
{
"round": "round2",
"source_file": "hybrid_r14_draft05b.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "hybrid_r14_draft05b",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "prompt_heavy",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 340,
"completion_tokens": 32,
"load_time_s": 53.386,
"ttft_s": 5.665,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 41.96,
"total_time_s": 47.624,
"decode_tok_per_s": 0.739,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.672,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 15,
"drafted_tokens": 63,
"accepted_tokens": 17,
"resident_layers": 14,
"streamed_layers": 34,
"n_layers": 48,
"resident_mib": 3244.4,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 79826.54,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 6257.4,
"current_rss_mib": 5668.1,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -198.3,
"page_cache_delta_since_start_mib": -2235.9,
"idle_watts": null,
"avg_watts": null,
"net_watts": null,
"joules_per_tok": null,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " Layer-streaming inference manages memory by allocating a small, fixed pool of buffers for active computation, keeping only essential layers in memory, and asyn"
},
{
"round": "round2",
"source_file": "hybrid_r14_draft05b.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "hybrid_r14_draft05b",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "code_edit",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 106,
"completion_tokens": 48,
"load_time_s": 53.386,
"ttft_s": 3.438,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 24.011,
"total_time_s": 27.449,
"decode_tok_per_s": 1.957,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 1.749,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 9,
"drafted_tokens": 40,
"accepted_tokens": 39,
"resident_layers": 14,
"streamed_layers": 34,
"n_layers": 48,
"resident_mib": 3244.4,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 47895.93,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 6257.4,
"current_rss_mib": 5669.1,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -4.3,
"page_cache_delta_since_start_mib": -2240.2,
"idle_watts": null,
"avg_watts": null,
"net_watts": null,
"joules_per_tok": null,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": "#[derive(Debug, Clone)]\npub struct StreamConfig {\npub model_path: PathBuf,\npub n_ctx: u32,\npub n_batch: u32,\npub resident_layers: Option<u32>,\npub vram_budget_b"
},
{
"round": "round2",
"source_file": "hybrid_r14_draft05b.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "hybrid_r14_draft05b",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "open_ended",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 31,
"completion_tokens": 48,
"load_time_s": 53.386,
"ttft_s": 3.155,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 36.017,
"total_time_s": 39.172,
"decode_tok_per_s": 1.305,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 1.225,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 13,
"drafted_tokens": 59,
"accepted_tokens": 32,
"resident_layers": 14,
"streamed_layers": 34,
"n_layers": 48,
"resident_mib": 3244.4,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 69183.0,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 6257.4,
"current_rss_mib": 5669.4,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -14.6,
"page_cache_delta_since_start_mib": -2254.7,
"idle_watts": null,
"avg_watts": null,
"net_watts": null,
"joules_per_tok": null,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " When streaming transformer layers on an 8 GB edge device, the interaction between page-aligned O_DIRECT NVMe reads and asynchronous ping-pong GPU buffers is cr"
},
{
"round": "round2",
"source_file": "hybrid_r16_ngram.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "hybrid_r16_ngram",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": null,
"workload": "prompt_heavy",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 340,
"completion_tokens": 32,
"load_time_s": 75.668,
"ttft_s": 5.074,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 63.663,
"total_time_s": 68.736,
"decode_tok_per_s": 0.487,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.466,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 25,
"drafted_tokens": 26,
"accepted_tokens": 7,
"resident_layers": 16,
"streamed_layers": 32,
"n_layers": 48,
"resident_mib": 3558.5,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 125190.63,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 5982.3,
"current_rss_mib": 5152.0,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 57.9,
"page_cache_delta_since_start_mib": -181.0,
"idle_watts": null,
"avg_watts": null,
"net_watts": null,
"joules_per_tok": null,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " Layer-streaming inference manages memory by allocating a small, fixed pool of buffers for compute, keeping essential layers in memory, and asynchronously strea"
},
{
"round": "round2",
"source_file": "hybrid_r16_ngram.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "hybrid_r16_ngram",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": null,
"workload": "code_edit",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 106,
"completion_tokens": 48,
"load_time_s": 75.668,
"ttft_s": 3.087,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 23.822,
"total_time_s": 26.909,
"decode_tok_per_s": 1.973,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 1.784,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 10,
"drafted_tokens": 45,
"accepted_tokens": 38,
"resident_layers": 16,
"streamed_layers": 32,
"n_layers": 48,
"resident_mib": 3558.5,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 50076.25,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 5982.3,
"current_rss_mib": 5153.0,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 3.7,
"page_cache_delta_since_start_mib": -177.3,
"idle_watts": null,
"avg_watts": null,
"net_watts": null,
"joules_per_tok": null,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": "#[derive(Debug, Clone)]\npub struct StreamConfig {\npub model_path: PathBuf,\npub n_ctx: u32,\npub n_batch: u32,\npub resident_layers: Option<u32>,\npub vram_budget_b"
},
{
"round": "round2",
"source_file": "hybrid_r16_ngram.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "hybrid_r16_ngram",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": null,
"workload": "open_ended",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 31,
"completion_tokens": 48,
"load_time_s": 75.668,
"ttft_s": 2.75,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 79.476,
"total_time_s": 82.225,
"decode_tok_per_s": 0.591,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.584,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 31,
"drafted_tokens": 25,
"accepted_tokens": 17,
"resident_layers": 16,
"streamed_layers": 32,
"n_layers": 48,
"resident_mib": 3558.5,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 155236.38,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 5982.3,
"current_rss_mib": 5155.4,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 2.4,
"page_cache_delta_since_start_mib": -174.9,
"idle_watts": null,
"avg_watts": null,
"net_watts": null,
"joules_per_tok": null,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " When streaming transformer layers on an 8 GB edge device, the interaction between page-aligned O_DIRECT NVMe reads and asynchronous ping-pong GPU buffers is cr"
},
{
"round": "round2",
"source_file": "stream_r0.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "stream_r0",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": null,
"workload": "prompt_heavy",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 340,
"completion_tokens": 32,
"load_time_s": 2.4,
"ttft_s": 9.753,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 247.344,
"total_time_s": 257.097,
"decode_tok_per_s": 0.125,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.124,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 32,
"drafted_tokens": 0,
"accepted_tokens": 0,
"resident_layers": 0,
"streamed_layers": 48,
"n_layers": 48,
"resident_mib": 1026.8,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 241263.38,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 3249.2,
"current_rss_mib": 3249.2,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 2.0,
"page_cache_delta_since_start_mib": 2.3,
"idle_watts": null,
"avg_watts": null,
"net_watts": null,
"joules_per_tok": null,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " Layer-streaming inference manages memory by allocating a small, fixed pool of buffers for compute, keeping essential layers in memory, and asynchronously strea"
},
{
"round": "round2",
"source_file": "stream_r0.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "stream_r0",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": null,
"workload": "code_edit",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 106,
"completion_tokens": 48,
"load_time_s": 2.4,
"ttft_s": 8.395,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 380.266,
"total_time_s": 388.66,
"decode_tok_per_s": 0.124,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.124,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 48,
"drafted_tokens": 0,
"accepted_tokens": 0,
"resident_layers": 0,
"streamed_layers": 48,
"n_layers": 48,
"resident_mib": 1026.8,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 361895.06,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 3249.2,
"current_rss_mib": 3249.2,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 2.1,
"page_cache_delta_since_start_mib": 4.4,
"idle_watts": null,
"avg_watts": null,
"net_watts": null,
"joules_per_tok": null,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": "#[derive(Debug, Clone)]\npub struct StreamConfig {\npub model_path: PathBuf,\npub n_ctx: u32,\npub n_batch: u32,\npub resident_layers: Option<u32>,\npub vram_budget_b"
},
{
"round": "round2",
"source_file": "stream_r0.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "stream_r0",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": null,
"workload": "open_ended",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 31,
"completion_tokens": 48,
"load_time_s": 2.4,
"ttft_s": 8.088,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 385.094,
"total_time_s": 393.183,
"decode_tok_per_s": 0.122,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.122,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 48,
"drafted_tokens": 0,
"accepted_tokens": 0,
"resident_layers": 0,
"streamed_layers": 48,
"n_layers": 48,
"resident_mib": 1026.8,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 361895.06,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 3249.2,
"current_rss_mib": 3249.2,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 9.7,
"page_cache_delta_since_start_mib": 14.1,
"idle_watts": null,
"avg_watts": null,
"net_watts": null,
"joules_per_tok": null,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " When streaming transformer layers on an 8 GB edge device, the interaction between page-aligned O_DIRECT NVMe reads and asynchronous ping-pong GPU buffers is cr"
},
{
"round": "round2",
"source_file": "stream_r0_ngram.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "stream_r0_ngram",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": null,
"workload": "prompt_heavy",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 340,
"completion_tokens": 32,
"load_time_s": 2.102,
"ttft_s": 10.312,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 204.874,
"total_time_s": 215.185,
"decode_tok_per_s": 0.151,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.149,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 25,
"drafted_tokens": 26,
"accepted_tokens": 7,
"resident_layers": 0,
"streamed_layers": 48,
"n_layers": 48,
"resident_mib": 1026.8,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 188487.01,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 3253.6,
"current_rss_mib": 3253.6,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 34.6,
"page_cache_delta_since_start_mib": 46.1,
"idle_watts": null,
"avg_watts": null,
"net_watts": null,
"joules_per_tok": null,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " Layer-streaming inference manages memory by allocating a small, fixed pool of buffers for compute, keeping essential layers in memory, and asynchronously strea"
},
{
"round": "round2",
"source_file": "stream_r0_ngram.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "stream_r0_ngram",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": null,
"workload": "code_edit",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 106,
"completion_tokens": 48,
"load_time_s": 2.102,
"ttft_s": 8.983,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 76.835,
"total_time_s": 85.818,
"decode_tok_per_s": 0.612,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.559,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 10,
"drafted_tokens": 45,
"accepted_tokens": 38,
"resident_layers": 0,
"streamed_layers": 48,
"n_layers": 48,
"resident_mib": 1026.8,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 75394.8,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 3253.6,
"current_rss_mib": 3253.6,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 3.6,
"page_cache_delta_since_start_mib": 49.7,
"idle_watts": null,
"avg_watts": null,
"net_watts": null,
"joules_per_tok": null,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": "#[derive(Debug, Clone)]\npub struct StreamConfig {\npub model_path: PathBuf,\npub n_ctx: u32,\npub n_batch: u32,\npub resident_layers: Option<u32>,\npub vram_budget_b"
},
{
"round": "round2",
"source_file": "stream_r0_ngram.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "stream_r0_ngram",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": null,
"workload": "open_ended",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 31,
"completion_tokens": 48,
"load_time_s": 2.102,
"ttft_s": 8.666,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 256.328,
"total_time_s": 264.994,
"decode_tok_per_s": 0.183,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.181,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 31,
"drafted_tokens": 25,
"accepted_tokens": 17,
"resident_layers": 0,
"streamed_layers": 48,
"n_layers": 48,
"resident_mib": 1026.8,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 233723.89,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 3255.7,
"current_rss_mib": 3255.6,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 39.6,
"page_cache_delta_since_start_mib": 89.3,
"idle_watts": null,
"avg_watts": null,
"net_watts": null,
"joules_per_tok": null,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " When streaming transformer layers on an 8 GB edge device, the interaction between page-aligned O_DIRECT NVMe reads and asynchronous ping-pong GPU buffers is cr"
},
{
"round": "round3",
"source_file": "r3_hybrid_r14_draft05b.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r3_hybrid_r14_draft05b",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "prompt_heavy",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 340,
"completion_tokens": 32,
"load_time_s": 6.886,
"ttft_s": 5.916,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 69.787,
"total_time_s": 75.704,
"decode_tok_per_s": 0.444,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.423,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 15,
"drafted_tokens": 62,
"accepted_tokens": 16,
"resident_layers": 14,
"streamed_layers": 34,
"n_layers": 48,
"resident_mib": 3244.4,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 79826.54,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 6079.0,
"current_rss_mib": 6079.0,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -57.5,
"page_cache_delta_since_start_mib": -1016.1,
"idle_watts": null,
"avg_watts": null,
"net_watts": null,
"joules_per_tok": null,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " Layer-streaming inference manages memory by allocating a small, fixed pool of reusable buffers for compute, keeping essential layers in memory, and asynchronou"
},
{
"round": "round3",
"source_file": "r3_hybrid_r14_draft05b.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r3_hybrid_r14_draft05b",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "code_edit",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 106,
"completion_tokens": 48,
"load_time_s": 6.886,
"ttft_s": 3.49,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 18.261,
"total_time_s": 21.75,
"decode_tok_per_s": 2.574,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 2.207,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 6,
"drafted_tokens": 43,
"accepted_tokens": 36,
"resident_layers": 14,
"streamed_layers": 34,
"n_layers": 48,
"resident_mib": 3244.4,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 31930.62,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 6083.3,
"current_rss_mib": 6083.3,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 0.3,
"page_cache_delta_since_start_mib": -1015.8,
"idle_watts": null,
"avg_watts": null,
"net_watts": null,
"joules_per_tok": null,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": "#[derive(Debug, Clone)]\npub struct StreamConfig {\npub model_path: PathBuf,\npub n_ctx: u32,\npub n_batch: u32,\npub resident_layers: Option<u32>,\npub vram_budget_b"
},
{
"round": "round3",
"source_file": "r3_hybrid_r14_draft05b.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r3_hybrid_r14_draft05b",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "open_ended",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 31,
"completion_tokens": 48,
"load_time_s": 6.886,
"ttft_s": 3.289,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 31.885,
"total_time_s": 35.174,
"decode_tok_per_s": 1.474,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 1.365,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 11,
"drafted_tokens": 64,
"accepted_tokens": 34,
"resident_layers": 14,
"streamed_layers": 34,
"n_layers": 48,
"resident_mib": 3244.4,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 58539.46,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 6083.6,
"current_rss_mib": 6083.6,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 1.2,
"page_cache_delta_since_start_mib": -1014.6,
"idle_watts": null,
"avg_watts": null,
"net_watts": null,
"joules_per_tok": null,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " When streaming transformer layers on an 8 GB edge device, the interaction between page-aligned O_DIRECT NVMe reads and asynchronous ping-pong GPU buffers is cr"
},
{
"round": "round3",
"source_file": "r3_stream_r0.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r3_stream_r0",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": null,
"workload": "prompt_heavy",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 340,
"completion_tokens": 32,
"load_time_s": 1.859,
"ttft_s": 7.419,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 260.939,
"total_time_s": 268.358,
"decode_tok_per_s": 0.119,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.119,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 32,
"drafted_tokens": 0,
"accepted_tokens": 0,
"resident_layers": 0,
"streamed_layers": 48,
"n_layers": 48,
"resident_mib": 1026.8,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 241263.38,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 3168.0,
"current_rss_mib": 3168.0,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 0.6,
"page_cache_delta_since_start_mib": 1.2,
"idle_watts": null,
"avg_watts": null,
"net_watts": null,
"joules_per_tok": null,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " Layer-streaming inference manages memory by allocating a small, fixed pool of buffers for active computation, keeping only essential layers in memory, and asyn"
},
{
"round": "round4",
"source_file": "r4_auto_budget_5800_draft05b.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r4_auto_budget_5800_draft05b",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "prompt_heavy",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 340,
"completion_tokens": 32,
"load_time_s": 8.611,
"ttft_s": 4.823,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 38.613,
"total_time_s": 43.437,
"decode_tok_per_s": 0.803,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.737,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 15,
"drafted_tokens": 62,
"accepted_tokens": 16,
"resident_layers": 19,
"streamed_layers": 29,
"n_layers": 48,
"resident_mib": 4020.4,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 68186.31,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 6482.0,
"current_rss_mib": 5978.3,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 7.5,
"page_cache_delta_since_start_mib": -178.0,
"idle_watts": null,
"avg_watts": 12.65,
"net_watts": null,
"joules_per_tok": 17.17,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " Layer-streaming inference manages memory by allocating a small, fixed pool of reusable buffers for compute, keeping essential layers in memory, and asynchronou"
},
{
"round": "round4",
"source_file": "r4_auto_budget_5800_draft05b.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r4_auto_budget_5800_draft05b",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "code_edit",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 106,
"completion_tokens": 48,
"load_time_s": 8.611,
"ttft_s": 3.137,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 13.633,
"total_time_s": 16.77,
"decode_tok_per_s": 3.447,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 2.862,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 6,
"drafted_tokens": 43,
"accepted_tokens": 36,
"resident_layers": 19,
"streamed_layers": 29,
"n_layers": 48,
"resident_mib": 4020.4,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 27274.52,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 6482.0,
"current_rss_mib": 5971.7,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -1.5,
"page_cache_delta_since_start_mib": -179.5,
"idle_watts": null,
"avg_watts": 12.59,
"net_watts": null,
"joules_per_tok": 4.4,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": "#[derive(Debug, Clone)]\npub struct StreamConfig {\npub model_path: PathBuf,\npub n_ctx: u32,\npub n_batch: u32,\npub resident_layers: Option<u32>,\npub vram_budget_b"
},
{
"round": "round4",
"source_file": "r4_auto_budget_5800_draft05b.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r4_auto_budget_5800_draft05b",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "open_ended",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 31,
"completion_tokens": 48,
"load_time_s": 8.611,
"ttft_s": 2.907,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 27.967,
"total_time_s": 30.874,
"decode_tok_per_s": 1.681,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 1.555,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 11,
"drafted_tokens": 81,
"accepted_tokens": 34,
"resident_layers": 19,
"streamed_layers": 29,
"n_layers": 48,
"resident_mib": 4020.4,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 50003.29,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 6482.0,
"current_rss_mib": 5939.8,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -23.4,
"page_cache_delta_since_start_mib": -202.8,
"idle_watts": null,
"avg_watts": 12.2,
"net_watts": null,
"joules_per_tok": 7.85,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " When streaming transformer layers on an 8 GB edge device, the interaction between page-aligned O_DIRECT NVMe reads and asynchronous ping-pong GPU buffers is cr"
},
{
"round": "round4",
"source_file": "r4_hybrid_r14_draft05b.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r4_hybrid_r14_draft05b",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "prompt_heavy",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 340,
"completion_tokens": 32,
"load_time_s": 7.8,
"ttft_s": 5.093,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 39.028,
"total_time_s": 44.12,
"decode_tok_per_s": 0.794,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.725,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 15,
"drafted_tokens": 62,
"accepted_tokens": 16,
"resident_layers": 14,
"streamed_layers": 34,
"n_layers": 48,
"resident_mib": 3244.4,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 79992.96,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 5708.1,
"current_rss_mib": 5348.1,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -87.8,
"page_cache_delta_since_start_mib": -690.5,
"idle_watts": null,
"avg_watts": 12.85,
"net_watts": null,
"joules_per_tok": 17.72,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " Layer-streaming inference manages memory by allocating a small, fixed pool of reusable buffers for compute, keeping essential layers in memory, and asynchronou"
},
{
"round": "round4",
"source_file": "r4_hybrid_r14_draft05b.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r4_hybrid_r14_draft05b",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "code_edit",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 106,
"completion_tokens": 48,
"load_time_s": 7.8,
"ttft_s": 3.155,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 13.958,
"total_time_s": 17.113,
"decode_tok_per_s": 3.367,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 2.805,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 6,
"drafted_tokens": 43,
"accepted_tokens": 36,
"resident_layers": 14,
"streamed_layers": 34,
"n_layers": 48,
"resident_mib": 3244.4,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 31930.62,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 5708.1,
"current_rss_mib": 5352.6,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -14.7,
"page_cache_delta_since_start_mib": -705.2,
"idle_watts": null,
"avg_watts": 12.36,
"net_watts": null,
"joules_per_tok": 4.41,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": "#[derive(Debug, Clone)]\npub struct StreamConfig {\npub model_path: PathBuf,\npub n_ctx: u32,\npub n_batch: u32,\npub resident_layers: Option<u32>,\npub vram_budget_b"
},
{
"round": "round4",
"source_file": "r4_hybrid_r14_draft05b.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r4_hybrid_r14_draft05b",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "open_ended",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 31,
"completion_tokens": 48,
"load_time_s": 7.8,
"ttft_s": 2.936,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 28.241,
"total_time_s": 31.177,
"decode_tok_per_s": 1.664,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 1.54,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 11,
"drafted_tokens": 81,
"accepted_tokens": 34,
"resident_layers": 14,
"streamed_layers": 34,
"n_layers": 48,
"resident_mib": 3244.4,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 58539.46,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 5708.1,
"current_rss_mib": 5353.0,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 5.8,
"page_cache_delta_since_start_mib": -699.4,
"idle_watts": null,
"avg_watts": 12.27,
"net_watts": null,
"joules_per_tok": 7.97,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " When streaming transformer layers on an 8 GB edge device, the interaction between page-aligned O_DIRECT NVMe reads and asynchronous ping-pong GPU buffers is cr"
},
{
"round": "round4",
"source_file": "r4_hybrid_r16_draft05b.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r4_hybrid_r16_draft05b",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "prompt_heavy",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 340,
"completion_tokens": 32,
"load_time_s": 5.469,
"ttft_s": 4.708,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 38.462,
"total_time_s": 43.17,
"decode_tok_per_s": 0.806,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.741,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 15,
"drafted_tokens": 62,
"accepted_tokens": 16,
"resident_layers": 16,
"streamed_layers": 32,
"n_layers": 48,
"resident_mib": 3558.5,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 75114.38,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 6201.6,
"current_rss_mib": 6201.6,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -54.8,
"page_cache_delta_since_start_mib": -451.7,
"idle_watts": null,
"avg_watts": 12.67,
"net_watts": null,
"joules_per_tok": 17.09,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " Layer-streaming inference manages memory by allocating a small, fixed pool of reusable buffers for compute, keeping essential layers in memory, and asynchronou"
},
{
"round": "round4",
"source_file": "r4_hybrid_r16_draft05b.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r4_hybrid_r16_draft05b",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "code_edit",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 106,
"completion_tokens": 48,
"load_time_s": 5.469,
"ttft_s": 3.142,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 13.715,
"total_time_s": 16.857,
"decode_tok_per_s": 3.427,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 2.848,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 6,
"drafted_tokens": 43,
"accepted_tokens": 36,
"resident_layers": 16,
"streamed_layers": 32,
"n_layers": 48,
"resident_mib": 3558.5,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 30193.48,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 6203.8,
"current_rss_mib": 6198.6,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -11.1,
"page_cache_delta_since_start_mib": -462.8,
"idle_watts": null,
"avg_watts": 12.09,
"net_watts": null,
"joules_per_tok": 4.25,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": "#[derive(Debug, Clone)]\npub struct StreamConfig {\npub model_path: PathBuf,\npub n_ctx: u32,\npub n_batch: u32,\npub resident_layers: Option<u32>,\npub vram_budget_b"
},
{
"round": "round4",
"source_file": "r4_hybrid_r16_draft05b.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r4_hybrid_r16_draft05b",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "open_ended",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 31,
"completion_tokens": 48,
"load_time_s": 5.469,
"ttft_s": 2.91,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 27.794,
"total_time_s": 30.704,
"decode_tok_per_s": 1.691,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 1.563,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 11,
"drafted_tokens": 81,
"accepted_tokens": 34,
"resident_layers": 16,
"streamed_layers": 32,
"n_layers": 48,
"resident_mib": 3558.5,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 55083.88,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 6203.8,
"current_rss_mib": 6179.3,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 7.0,
"page_cache_delta_since_start_mib": -455.8,
"idle_watts": null,
"avg_watts": 12.86,
"net_watts": null,
"joules_per_tok": 8.23,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " When streaming transformer layers on an 8 GB edge device, the interaction between page-aligned O_DIRECT NVMe reads and asynchronous ping-pong GPU buffers is cr"
},
{
"round": "round4",
"source_file": "r4_hybrid_r16_draft05b_t2.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r4_hybrid_r16_draft05b_t2",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "prompt_heavy",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 340,
"completion_tokens": 32,
"load_time_s": 5.178,
"ttft_s": 5.791,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 53.552,
"total_time_s": 59.343,
"decode_tok_per_s": 0.579,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.539,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 15,
"drafted_tokens": 62,
"accepted_tokens": 16,
"resident_layers": 16,
"streamed_layers": 32,
"n_layers": 48,
"resident_mib": 3558.5,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 75262.1,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 6199.6,
"current_rss_mib": 6199.6,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -53.2,
"page_cache_delta_since_start_mib": 4.1,
"idle_watts": null,
"avg_watts": 11.14,
"net_watts": null,
"joules_per_tok": 20.65,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " Layer-streaming inference manages memory by allocating a small, fixed pool of reusable buffers for compute, keeping essential layers in memory, and asynchronou"
},
{
"round": "round4",
"source_file": "r4_hybrid_r16_draft05b_t2.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r4_hybrid_r16_draft05b_t2",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "code_edit",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 106,
"completion_tokens": 48,
"load_time_s": 5.178,
"ttft_s": 3.989,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 18.402,
"total_time_s": 22.391,
"decode_tok_per_s": 2.554,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 2.144,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 6,
"drafted_tokens": 43,
"accepted_tokens": 36,
"resident_layers": 16,
"streamed_layers": 32,
"n_layers": 48,
"resident_mib": 3558.5,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 29898.02,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 6203.1,
"current_rss_mib": 6203.1,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -18.3,
"page_cache_delta_since_start_mib": -14.1,
"idle_watts": null,
"avg_watts": 11.43,
"net_watts": null,
"joules_per_tok": 5.33,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": "#[derive(Debug, Clone)]\npub struct StreamConfig {\npub model_path: PathBuf,\npub n_ctx: u32,\npub n_batch: u32,\npub resident_layers: Option<u32>,\npub vram_budget_b"
},
{
"round": "round4",
"source_file": "r4_hybrid_r16_draft05b_t2.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r4_hybrid_r16_draft05b_t2",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "open_ended",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 31,
"completion_tokens": 48,
"load_time_s": 5.178,
"ttft_s": 3.805,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 38.511,
"total_time_s": 42.316,
"decode_tok_per_s": 1.22,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 1.134,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 11,
"drafted_tokens": 81,
"accepted_tokens": 34,
"resident_layers": 16,
"streamed_layers": 32,
"n_layers": 48,
"resident_mib": 3558.5,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 55083.88,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 6203.1,
"current_rss_mib": 6203.1,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -9.2,
"page_cache_delta_since_start_mib": -23.4,
"idle_watts": null,
"avg_watts": 11.34,
"net_watts": null,
"joules_per_tok": 10.0,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " When streaming transformer layers on an 8 GB edge device, the interaction between page-aligned O_DIRECT NVMe reads and asynchronous ping-pong GPU buffers is cr"
},
{
"round": "round4",
"source_file": "r4_hybrid_r16_draft05b_t4.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r4_hybrid_r16_draft05b_t4",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "prompt_heavy",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 340,
"completion_tokens": 32,
"load_time_s": 5.561,
"ttft_s": 5.592,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 51.504,
"total_time_s": 57.095,
"decode_tok_per_s": 0.602,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.56,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 15,
"drafted_tokens": 62,
"accepted_tokens": 16,
"resident_layers": 16,
"streamed_layers": 32,
"n_layers": 48,
"resident_mib": 3558.5,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 75262.1,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 6203.0,
"current_rss_mib": 6186.4,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -245.6,
"page_cache_delta_since_start_mib": -166.0,
"idle_watts": null,
"avg_watts": 11.47,
"net_watts": null,
"joules_per_tok": 20.47,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " Layer-streaming inference manages memory by allocating a small, fixed pool of reusable buffers for compute, keeping essential layers in memory, and asynchronou"
},
{
"round": "round4",
"source_file": "r4_hybrid_r16_draft05b_t4.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r4_hybrid_r16_draft05b_t4",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "code_edit",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 106,
"completion_tokens": 48,
"load_time_s": 5.561,
"ttft_s": 6.269,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 22.093,
"total_time_s": 28.362,
"decode_tok_per_s": 2.127,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 1.692,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 6,
"drafted_tokens": 43,
"accepted_tokens": 36,
"resident_layers": 16,
"streamed_layers": 32,
"n_layers": 48,
"resident_mib": 3558.5,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 30045.75,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 6203.0,
"current_rss_mib": 6174.4,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 3.3,
"page_cache_delta_since_start_mib": -162.6,
"idle_watts": null,
"avg_watts": 10.5,
"net_watts": null,
"joules_per_tok": 6.21,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": "#[derive(Debug, Clone)]\npub struct StreamConfig {\npub model_path: PathBuf,\npub n_ctx: u32,\npub n_batch: u32,\npub resident_layers: Option<u32>,\npub vram_budget_b"
},
{
"round": "round4",
"source_file": "r4_hybrid_r16_draft05b_t4.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r4_hybrid_r16_draft05b_t4",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "open_ended",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 31,
"completion_tokens": 48,
"load_time_s": 5.561,
"ttft_s": 3.665,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 65.359,
"total_time_s": 69.025,
"decode_tok_per_s": 0.719,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.695,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 11,
"drafted_tokens": 81,
"accepted_tokens": 34,
"resident_layers": 16,
"streamed_layers": 32,
"n_layers": 48,
"resident_mib": 3558.5,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 55083.88,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 6203.0,
"current_rss_mib": 6172.8,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 3.4,
"page_cache_delta_since_start_mib": -159.2,
"idle_watts": null,
"avg_watts": 9.77,
"net_watts": null,
"joules_per_tok": 14.04,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " When streaming transformer layers on an 8 GB edge device, the interaction between page-aligned O_DIRECT NVMe reads and asynchronous ping-pong GPU buffers is cr"
},
{
"round": "round4",
"source_file": "r4_hybrid_r16_ngram.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r4_hybrid_r16_ngram",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": null,
"workload": "prompt_heavy",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 340,
"completion_tokens": 32,
"load_time_s": 4.735,
"ttft_s": 7.575,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 147.639,
"total_time_s": 155.213,
"decode_tok_per_s": 0.21,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.206,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 25,
"drafted_tokens": 29,
"accepted_tokens": 7,
"resident_layers": 16,
"streamed_layers": 32,
"n_layers": 48,
"resident_mib": 3558.5,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 125190.63,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 5596.0,
"current_rss_mib": 5596.0,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 58.6,
"page_cache_delta_since_start_mib": 73.6,
"idle_watts": null,
"avg_watts": 9.86,
"net_watts": null,
"joules_per_tok": 47.84,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " Layer-streaming inference manages memory by allocating a small, fixed pool of buffers for active computation, keeping critical layers in memory, and asynchrono"
},
{
"round": "round4",
"source_file": "r4_hybrid_r16_ngram.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r4_hybrid_r16_ngram",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": null,
"workload": "code_edit",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 106,
"completion_tokens": 48,
"load_time_s": 4.735,
"ttft_s": 6.145,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 28.747,
"total_time_s": 34.892,
"decode_tok_per_s": 1.635,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 1.376,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 6,
"drafted_tokens": 43,
"accepted_tokens": 36,
"resident_layers": 16,
"streamed_layers": 32,
"n_layers": 48,
"resident_mib": 3558.5,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 30045.75,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 5599.5,
"current_rss_mib": 5599.5,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 12.2,
"page_cache_delta_since_start_mib": 85.7,
"idle_watts": null,
"avg_watts": 10.26,
"net_watts": null,
"joules_per_tok": 7.45,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": "#[derive(Debug, Clone)]\npub struct StreamConfig {\npub model_path: PathBuf,\npub n_ctx: u32,\npub n_batch: u32,\npub resident_layers: Option<u32>,\npub vram_budget_b"
},
{
"round": "round4",
"source_file": "r4_hybrid_r16_ngram.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r4_hybrid_r16_ngram",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": null,
"workload": "open_ended",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 31,
"completion_tokens": 48,
"load_time_s": 4.735,
"ttft_s": 5.783,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 196.684,
"total_time_s": 202.467,
"decode_tok_per_s": 0.239,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.237,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 30,
"drafted_tokens": 32,
"accepted_tokens": 18,
"resident_layers": 16,
"streamed_layers": 32,
"n_layers": 48,
"resident_mib": 3558.5,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 150228.75,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 5601.2,
"current_rss_mib": 5164.1,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 166.6,
"page_cache_delta_since_start_mib": 252.4,
"idle_watts": null,
"avg_watts": 9.69,
"net_watts": null,
"joules_per_tok": 40.88,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " When streaming transformer layers on an 8 GB edge device, the interaction between page-aligned O_DIRECT NVMe reads and asynchronous ping-pong GPU buffers is cr"
},
{
"round": "round4",
"source_file": "r4_stream_r0.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r4_stream_r0",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": null,
"workload": "prompt_heavy",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 340,
"completion_tokens": 32,
"load_time_s": 2.595,
"ttft_s": 10.378,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 265.468,
"total_time_s": 275.846,
"decode_tok_per_s": 0.117,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.116,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 32,
"drafted_tokens": 0,
"accepted_tokens": 0,
"resident_layers": 0,
"streamed_layers": 48,
"n_layers": 48,
"resident_mib": 1026.8,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 241429.79,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 3073.8,
"current_rss_mib": 3073.8,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 21.8,
"page_cache_delta_since_start_mib": 35.4,
"idle_watts": null,
"avg_watts": 9.86,
"net_watts": null,
"joules_per_tok": 85.0,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " Layer-streaming inference manages memory by allocating a small, fixed pool of buffers for active computation, keeping only essential layers in memory, and asyn"
},
{
"round": "round5",
"source_file": "r5_zc_auto_budget_5800_draft05b.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r5_zc_auto_budget_5800_draft05b",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "prompt_heavy",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 340,
"completion_tokens": 32,
"load_time_s": 6.523,
"ttft_s": 5.366,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 51.725,
"total_time_s": 57.091,
"decode_tok_per_s": 0.599,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.561,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 16,
"drafted_tokens": 107,
"accepted_tokens": 16,
"resident_layers": 16,
"streamed_layers": 32,
"n_layers": 48,
"resident_mib": 3558.5,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 80269.73,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 5580.8,
"current_rss_mib": 5580.7,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -58.6,
"page_cache_delta_since_start_mib": 369.9,
"idle_watts": null,
"avg_watts": 11.19,
"net_watts": null,
"joules_per_tok": 19.97,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " Layer-streaming inference manages memory by allocating a small, fixed pool of buffers for active computation, keeping only essential layers in memory, and asyn"
},
{
"round": "round5",
"source_file": "r5_zc_auto_budget_5800_draft05b.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r5_zc_auto_budget_5800_draft05b",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "code_edit",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 106,
"completion_tokens": 48,
"load_time_s": 6.523,
"ttft_s": 3.662,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 13.455,
"total_time_s": 17.117,
"decode_tok_per_s": 3.493,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 2.804,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 5,
"drafted_tokens": 44,
"accepted_tokens": 36,
"resident_layers": 16,
"streamed_layers": 32,
"n_layers": 48,
"resident_mib": 3558.5,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 24890.4,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 5584.6,
"current_rss_mib": 5584.4,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -12.0,
"page_cache_delta_since_start_mib": 357.9,
"idle_watts": null,
"avg_watts": 11.84,
"net_watts": null,
"joules_per_tok": 4.22,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": "#[derive(Debug, Clone)]\npub struct StreamConfig {\npub model_path: PathBuf,\npub n_ctx: u32,\npub n_batch: u32,\npub resident_layers: Option<u32>,\npub vram_budget_b"
},
{
"round": "round5",
"source_file": "r5_zc_auto_budget_5800_draft05b.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r5_zc_auto_budget_5800_draft05b",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "open_ended",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 31,
"completion_tokens": 48,
"load_time_s": 6.523,
"ttft_s": 3.534,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 33.908,
"total_time_s": 37.442,
"decode_tok_per_s": 1.386,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 1.282,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 11,
"drafted_tokens": 115,
"accepted_tokens": 34,
"resident_layers": 16,
"streamed_layers": 32,
"n_layers": 48,
"resident_mib": 3558.5,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 55231.6,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 5584.6,
"current_rss_mib": 5584.4,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -12.8,
"page_cache_delta_since_start_mib": 345.1,
"idle_watts": null,
"avg_watts": 11.27,
"net_watts": null,
"joules_per_tok": 8.79,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " When streaming transformer layers on an 8 GB edge device, the interaction between page-aligned O_DIRECT NVMe reads and asynchronous ping-pong GPU buffers is cr"
},
{
"round": "round5",
"source_file": "r5_zc_hybrid_r14_draft15b.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r5_zc_hybrid_r14_draft15b",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-1.5b-instruct-q4_k_m.gguf",
"workload": "prompt_heavy",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 340,
"completion_tokens": 32,
"load_time_s": 7.065,
"ttft_s": 7.763,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 83.656,
"total_time_s": 91.419,
"decode_tok_per_s": 0.371,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.35,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 16,
"drafted_tokens": 103,
"accepted_tokens": 16,
"resident_layers": 14,
"streamed_layers": 34,
"n_layers": 48,
"resident_mib": 3244.4,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 85314.73,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 5865.6,
"current_rss_mib": 5865.6,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -41.6,
"page_cache_delta_since_start_mib": -289.5,
"idle_watts": null,
"avg_watts": 10.33,
"net_watts": null,
"joules_per_tok": 29.51,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " Layer-streaming inference manages memory by allocating a small, fixed pool of buffers for active computation, keeping only essential layers in memory, and asyn"
},
{
"round": "round5",
"source_file": "r5_zc_hybrid_r14_draft15b.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r5_zc_hybrid_r14_draft15b",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-1.5b-instruct-q4_k_m.gguf",
"workload": "code_edit",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 106,
"completion_tokens": 48,
"load_time_s": 7.065,
"ttft_s": 5.808,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 22.176,
"total_time_s": 27.984,
"decode_tok_per_s": 2.119,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 1.715,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 5,
"drafted_tokens": 44,
"accepted_tokens": 36,
"resident_layers": 14,
"streamed_layers": 34,
"n_layers": 48,
"resident_mib": 3244.4,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 26608.85,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 5870.3,
"current_rss_mib": 5870.3,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 4.3,
"page_cache_delta_since_start_mib": -285.2,
"idle_watts": null,
"avg_watts": 9.81,
"net_watts": null,
"joules_per_tok": 5.72,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": "#[derive(Debug, Clone)]\npub struct StreamConfig {\npub model_path: PathBuf,\npub n_ctx: u32,\npub n_batch: u32,\npub resident_layers: Option<u32>,\npub vram_budget_b"
},
{
"round": "round5",
"source_file": "r5_zc_hybrid_r14_draft15b.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r5_zc_hybrid_r14_draft15b",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-1.5b-instruct-q4_k_m.gguf",
"workload": "open_ended",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 31,
"completion_tokens": 48,
"load_time_s": 7.065,
"ttft_s": 5.726,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 50.56,
"total_time_s": 56.285,
"decode_tok_per_s": 0.93,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.853,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 10,
"drafted_tokens": 94,
"accepted_tokens": 35,
"resident_layers": 14,
"streamed_layers": 34,
"n_layers": 48,
"resident_mib": 3244.4,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 53051.28,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 5870.3,
"current_rss_mib": 5870.3,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -14.2,
"page_cache_delta_since_start_mib": -299.4,
"idle_watts": null,
"avg_watts": 10.43,
"net_watts": null,
"joules_per_tok": 12.23,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " When streaming transformer layers on an 8 GB edge device, the interaction between page-aligned O_DIRECT NVMe reads and asynchronous ping-pong GPU buffers is cr"
},
{
"round": "round5",
"source_file": "r5_zc_hybrid_r16_draft05b.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r5_zc_hybrid_r16_draft05b",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "prompt_heavy",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 340,
"completion_tokens": 32,
"load_time_s": 6.533,
"ttft_s": 5.432,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 50.131,
"total_time_s": 55.563,
"decode_tok_per_s": 0.618,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.576,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 16,
"drafted_tokens": 107,
"accepted_tokens": 16,
"resident_layers": 16,
"streamed_layers": 32,
"n_layers": 48,
"resident_mib": 3558.5,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 80269.73,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 5542.5,
"current_rss_mib": 5542.5,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -51.2,
"page_cache_delta_since_start_mib": -1137.8,
"idle_watts": null,
"avg_watts": 11.45,
"net_watts": null,
"joules_per_tok": 19.88,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " Layer-streaming inference manages memory by allocating a small, fixed pool of buffers for active computation, keeping only essential layers in memory, and asyn"
},
{
"round": "round5",
"source_file": "r5_zc_hybrid_r16_draft05b.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r5_zc_hybrid_r16_draft05b",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "code_edit",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 106,
"completion_tokens": 48,
"load_time_s": 6.533,
"ttft_s": 3.664,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 13.386,
"total_time_s": 17.05,
"decode_tok_per_s": 3.511,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 2.815,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 5,
"drafted_tokens": 44,
"accepted_tokens": 36,
"resident_layers": 16,
"streamed_layers": 32,
"n_layers": 48,
"resident_mib": 3558.5,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 25038.13,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 5547.2,
"current_rss_mib": 5547.2,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -13.0,
"page_cache_delta_since_start_mib": -1150.8,
"idle_watts": null,
"avg_watts": 10.73,
"net_watts": null,
"joules_per_tok": 3.81,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": "#[derive(Debug, Clone)]\npub struct StreamConfig {\npub model_path: PathBuf,\npub n_ctx: u32,\npub n_batch: u32,\npub resident_layers: Option<u32>,\npub vram_budget_b"
},
{
"round": "round5",
"source_file": "r5_zc_hybrid_r16_draft05b.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r5_zc_hybrid_r16_draft05b",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "open_ended",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 31,
"completion_tokens": 48,
"load_time_s": 6.533,
"ttft_s": 3.452,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 33.828,
"total_time_s": 37.281,
"decode_tok_per_s": 1.389,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 1.288,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 11,
"drafted_tokens": 115,
"accepted_tokens": 34,
"resident_layers": 16,
"streamed_layers": 32,
"n_layers": 48,
"resident_mib": 3558.5,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 55083.88,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 5547.2,
"current_rss_mib": 5547.2,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 6.8,
"page_cache_delta_since_start_mib": -1144.0,
"idle_watts": null,
"avg_watts": 10.97,
"net_watts": null,
"joules_per_tok": 8.52,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " When streaming transformer layers on an 8 GB edge device, the interaction between page-aligned O_DIRECT NVMe reads and asynchronous ping-pong GPU buffers is cr"
},
{
"round": "round5",
"source_file": "r5_zc_hybrid_r16_draft05b_temp07.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r5_zc_hybrid_r16_draft05b_temp07",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "prompt_heavy",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 340,
"completion_tokens": 32,
"load_time_s": 6.341,
"ttft_s": 7.225,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 52.134,
"total_time_s": 59.359,
"decode_tok_per_s": 0.595,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.539,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 11,
"drafted_tokens": 75,
"accepted_tokens": 21,
"resident_layers": 16,
"streamed_layers": 32,
"n_layers": 48,
"resident_mib": 3558.5,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 55083.88,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 5575.7,
"current_rss_mib": 5575.7,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -24.5,
"page_cache_delta_since_start_mib": 257.5,
"idle_watts": null,
"avg_watts": 10.25,
"net_watts": null,
"joules_per_tok": 19.02,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " The core memory-management mechanism described in the document is layer-streaming inference, which manages to fit large language models into limited memory by "
},
{
"round": "round5",
"source_file": "r5_zc_hybrid_r16_draft05b_temp07.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r5_zc_hybrid_r16_draft05b_temp07",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "code_edit",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 106,
"completion_tokens": 48,
"load_time_s": 6.341,
"ttft_s": 5.607,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 20.974,
"total_time_s": 26.581,
"decode_tok_per_s": 2.241,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 1.806,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 5,
"drafted_tokens": 44,
"accepted_tokens": 36,
"resident_layers": 16,
"streamed_layers": 32,
"n_layers": 48,
"resident_mib": 3558.5,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 25038.13,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 5580.9,
"current_rss_mib": 5580.9,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 0.2,
"page_cache_delta_since_start_mib": 257.7,
"idle_watts": null,
"avg_watts": 10.46,
"net_watts": null,
"joules_per_tok": 5.79,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": "#[derive(Debug, Clone)]\npub struct StreamConfig {\npub model_path: PathBuf,\npub n_ctx: u32,\npub n_batch: u32,\npub resident_layers: Option<u32>,\npub vram_budget_b"
},
{
"round": "round5",
"source_file": "r5_zc_hybrid_r16_draft05b_temp07.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r5_zc_hybrid_r16_draft05b_temp07",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "open_ended",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 31,
"completion_tokens": 48,
"load_time_s": 6.341,
"ttft_s": 5.345,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 52.259,
"total_time_s": 57.603,
"decode_tok_per_s": 0.899,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.833,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 11,
"drafted_tokens": 107,
"accepted_tokens": 32,
"resident_layers": 16,
"streamed_layers": 32,
"n_layers": 48,
"resident_mib": 3558.5,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 55083.88,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 5580.9,
"current_rss_mib": 5580.9,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 0.2,
"page_cache_delta_since_start_mib": 257.9,
"idle_watts": null,
"avg_watts": 9.78,
"net_watts": null,
"joules_per_tok": 11.73,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " The interaction between page-aligned O_DIRECT NVMe reads and asynchronous ping-pong GPU buffers in the context of streaming transformer layers on an 8 GB edge "
},
{
"round": "round5",
"source_file": "r5_zc_hybrid_r16_ngram.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r5_zc_hybrid_r16_ngram",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": null,
"workload": "prompt_heavy",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 340,
"completion_tokens": 32,
"load_time_s": 5.227,
"ttft_s": 6.954,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 125.207,
"total_time_s": 132.161,
"decode_tok_per_s": 0.248,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.242,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 25,
"drafted_tokens": 37,
"accepted_tokens": 7,
"resident_layers": 16,
"streamed_layers": 32,
"n_layers": 48,
"resident_mib": 3558.5,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 125338.35,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 4992.8,
"current_rss_mib": 4992.8,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 0.9,
"page_cache_delta_since_start_mib": -11.4,
"idle_watts": null,
"avg_watts": 9.98,
"net_watts": null,
"joules_per_tok": 41.23,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " Layer-streaming inference manages memory by allocating a small, fixed pool of buffers for active computation, keeping only essential layers in memory, and asyn"
},
{
"round": "round5",
"source_file": "r5_zc_hybrid_r16_ngram.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r5_zc_hybrid_r16_ngram",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": null,
"workload": "code_edit",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 106,
"completion_tokens": 48,
"load_time_s": 5.227,
"ttft_s": 5.484,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 26.205,
"total_time_s": 31.689,
"decode_tok_per_s": 1.794,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 1.515,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 6,
"drafted_tokens": 50,
"accepted_tokens": 31,
"resident_layers": 16,
"streamed_layers": 32,
"n_layers": 48,
"resident_mib": 3558.5,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 29898.02,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 4995.1,
"current_rss_mib": 4995.1,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 0.7,
"page_cache_delta_since_start_mib": -10.7,
"idle_watts": null,
"avg_watts": 9.8,
"net_watts": null,
"joules_per_tok": 6.47,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": "#[derive(Debug, Clone)]\npub struct StreamConfig {\npub model_path: PathBuf,\npub n_ctx: u32,\npub n_batch: u32,\npub resident_layers: Option<u32>,\npub vram_budget_b"
},
{
"round": "round5",
"source_file": "r5_zc_hybrid_r16_ngram.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r5_zc_hybrid_r16_ngram",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": null,
"workload": "open_ended",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 31,
"completion_tokens": 48,
"load_time_s": 5.227,
"ttft_s": 5.319,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 145.983,
"total_time_s": 151.302,
"decode_tok_per_s": 0.322,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.317,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 29,
"drafted_tokens": 30,
"accepted_tokens": 19,
"resident_layers": 16,
"streamed_layers": 32,
"n_layers": 48,
"resident_mib": 3558.5,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 145368.85,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 4997.1,
"current_rss_mib": 4997.1,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 3.4,
"page_cache_delta_since_start_mib": -7.2,
"idle_watts": null,
"avg_watts": 9.8,
"net_watts": null,
"joules_per_tok": 30.88,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " When streaming transformer layers on an 8 GB edge device, the interaction between page-aligned O_DIRECT NVMe reads and asynchronous ping-pong GPU buffers is cr"
},
{
"round": "round5",
"source_file": "r5_zc_hybrid_r19_draft05b.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r5_zc_hybrid_r19_draft05b",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "prompt_heavy",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 340,
"completion_tokens": 32,
"load_time_s": 7.737,
"ttft_s": 4.043,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 31.922,
"total_time_s": 35.965,
"decode_tok_per_s": 0.971,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.89,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 16,
"drafted_tokens": 107,
"accepted_tokens": 16,
"resident_layers": 19,
"streamed_layers": 29,
"n_layers": 48,
"resident_mib": 4020.4,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 72732.06,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 5890.0,
"current_rss_mib": 5890.0,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -77.4,
"page_cache_delta_since_start_mib": -448.5,
"idle_watts": null,
"avg_watts": 12.96,
"net_watts": null,
"joules_per_tok": 14.57,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " Layer-streaming inference manages memory by allocating a small, fixed pool of buffers for active computation, keeping only essential layers in memory, and asyn"
},
{
"round": "round5",
"source_file": "r5_zc_hybrid_r19_draft05b.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r5_zc_hybrid_r19_draft05b",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "code_edit",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 106,
"completion_tokens": 48,
"load_time_s": 7.737,
"ttft_s": 2.45,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 8.27,
"total_time_s": 10.72,
"decode_tok_per_s": 5.684,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 4.478,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 5,
"drafted_tokens": 44,
"accepted_tokens": 36,
"resident_layers": 19,
"streamed_layers": 29,
"n_layers": 48,
"resident_mib": 4020.4,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 22728.77,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 5894.9,
"current_rss_mib": 5894.9,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 0.4,
"page_cache_delta_since_start_mib": -448.1,
"idle_watts": null,
"avg_watts": 12.7,
"net_watts": null,
"joules_per_tok": 2.84,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": "#[derive(Debug, Clone)]\npub struct StreamConfig {\npub model_path: PathBuf,\npub n_ctx: u32,\npub n_batch: u32,\npub resident_layers: Option<u32>,\npub vram_budget_b"
},
{
"round": "round5",
"source_file": "r5_zc_hybrid_r19_draft05b.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r5_zc_hybrid_r19_draft05b",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "open_ended",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 31,
"completion_tokens": 48,
"load_time_s": 7.737,
"ttft_s": 2.262,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 21.636,
"total_time_s": 23.898,
"decode_tok_per_s": 2.172,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 2.009,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 11,
"drafted_tokens": 115,
"accepted_tokens": 34,
"resident_layers": 19,
"streamed_layers": 29,
"n_layers": 48,
"resident_mib": 4020.4,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 50003.29,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 5895.1,
"current_rss_mib": 5895.1,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -33.9,
"page_cache_delta_since_start_mib": -482.0,
"idle_watts": null,
"avg_watts": 12.48,
"net_watts": null,
"joules_per_tok": 6.21,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " When streaming transformer layers on an 8 GB edge device, the interaction between page-aligned O_DIRECT NVMe reads and asynchronous ping-pong GPU buffers is cr"
},
{
"round": "round5",
"source_file": "r5_zc_multiturn_r16_draft05b.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r5_zc_multiturn_r16_draft05b",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "prompt_heavy",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 340,
"completion_tokens": 32,
"load_time_s": 5.897,
"ttft_s": 7.188,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 78.222,
"total_time_s": 85.41,
"decode_tok_per_s": 0.396,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.375,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 16,
"drafted_tokens": 107,
"accepted_tokens": 16,
"resident_layers": 16,
"streamed_layers": 32,
"n_layers": 48,
"resident_mib": 3558.5,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 80122.0,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 5574.5,
"current_rss_mib": 5574.5,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -74.1,
"page_cache_delta_since_start_mib": 16.3,
"idle_watts": null,
"avg_watts": 10.19,
"net_watts": null,
"joules_per_tok": 27.19,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " Layer-streaming inference manages memory by allocating a small, fixed pool of buffers for active computation, keeping only essential layers in memory, and asyn"
},
{
"round": "round5",
"source_file": "r5_zc_multiturn_r16_draft05b.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r5_zc_multiturn_r16_draft05b_multiturn",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": null,
"workload": "multi_turn_prefix_reuse",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": null,
"completion_tokens": null,
"load_time_s": null,
"ttft_s": null,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": null,
"total_time_s": null,
"decode_tok_per_s": null,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": null,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": null,
"drafted_tokens": null,
"accepted_tokens": null,
"resident_layers": null,
"streamed_layers": null,
"n_layers": null,
"resident_mib": null,
"slot_mib": null,
"expert_slot_mib": null,
"io_mib_read": null,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": null,
"is_unified_memory": null,
"direct_io_active": null,
"peak_rss_mib": null,
"current_rss_mib": null,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": null,
"page_cache_delta_since_start_mib": null,
"idle_watts": null,
"avg_watts": null,
"net_watts": null,
"joules_per_tok": null,
"net_joules_per_tok": null,
"turn1_prompt_tokens": 340,
"turn1_ttft_s": 5.174,
"turn2_prompt_tokens": 377,
"turn2_ttft_s": 5.4,
"ttft_speedup_x": 0.96,
"notes": null,
"output_preview": " O_DIRECT prevents page-cache thrashing by bypassing the operating system's page cache"
},
{
"round": "round5",
"source_file": "r5_zc_stream_r0.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r5_zc_stream_r0",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": null,
"workload": "prompt_heavy",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 340,
"completion_tokens": 32,
"load_time_s": 1.99,
"ttft_s": 6.96,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 163.279,
"total_time_s": 170.239,
"decode_tok_per_s": 0.19,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.188,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 32,
"drafted_tokens": 0,
"accepted_tokens": 0,
"resident_layers": 0,
"streamed_layers": 48,
"n_layers": 48,
"resident_mib": 1026.8,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 241429.79,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 2492.4,
"current_rss_mib": 2492.4,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 0.8,
"page_cache_delta_since_start_mib": 1.3,
"idle_watts": null,
"avg_watts": 10.71,
"net_watts": null,
"joules_per_tok": 56.97,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " Layer-streaming inference manages memory by allocating a small, fixed pool of buffers for active computation, keeping only essential layers in memory, and asyn"
},
{
"round": "round5",
"source_file": "r5_zc_stream_r0.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r5_zc_stream_r0",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": null,
"workload": "code_edit",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 106,
"completion_tokens": 48,
"load_time_s": 1.99,
"ttft_s": 5.618,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 247.709,
"total_time_s": 253.327,
"decode_tok_per_s": 0.19,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.189,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 48,
"drafted_tokens": 0,
"accepted_tokens": 0,
"resident_layers": 0,
"streamed_layers": 48,
"n_layers": 48,
"resident_mib": 1026.8,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 361728.64,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 2492.4,
"current_rss_mib": 2492.4,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 36.1,
"page_cache_delta_since_start_mib": 37.4,
"idle_watts": null,
"avg_watts": 10.67,
"net_watts": null,
"joules_per_tok": 56.33,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": "#[derive(Debug, Clone)]\npub struct StreamConfig {\npub model_path: PathBuf,\npub n_ctx: u32,\npub n_batch: u32,\npub resident_layers: Option<u32>,\npub vram_budget_b"
},
{
"round": "round5",
"source_file": "r5_zc_stream_r0.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r5_zc_stream_r0",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": null,
"workload": "open_ended",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 31,
"completion_tokens": 48,
"load_time_s": 1.99,
"ttft_s": 5.368,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 247.761,
"total_time_s": 253.129,
"decode_tok_per_s": 0.19,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.19,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 48,
"drafted_tokens": 0,
"accepted_tokens": 0,
"resident_layers": 0,
"streamed_layers": 48,
"n_layers": 48,
"resident_mib": 1026.8,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 362061.48,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 2492.4,
"current_rss_mib": 2492.4,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 26.8,
"page_cache_delta_since_start_mib": 64.2,
"idle_watts": null,
"avg_watts": 10.7,
"net_watts": null,
"joules_per_tok": 56.41,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " When streaming transformer layers on an 8 GB edge device, the interaction between page-aligned O_DIRECT NVMe reads and asynchronous ping-pong GPU buffers is cr"
},
{
"round": "round5",
"source_file": "r5_zc_stream_r0_ngram.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r5_zc_stream_r0_ngram",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": null,
"workload": "prompt_heavy",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 340,
"completion_tokens": 32,
"load_time_s": 2.011,
"ttft_s": 6.246,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 109.781,
"total_time_s": 116.027,
"decode_tok_per_s": 0.282,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.276,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 25,
"drafted_tokens": 37,
"accepted_tokens": 7,
"resident_layers": 0,
"streamed_layers": 48,
"n_layers": 48,
"resident_mib": 1026.8,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 188653.43,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 2500.8,
"current_rss_mib": 2500.7,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 0.9,
"page_cache_delta_since_start_mib": 1.3,
"idle_watts": null,
"avg_watts": 11.16,
"net_watts": null,
"joules_per_tok": 40.47,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " Layer-streaming inference manages memory by allocating a small, fixed pool of buffers for active computation, keeping only essential layers in memory, and asyn"
},
{
"round": "round5",
"source_file": "r5_zc_stream_r0_ngram.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r5_zc_stream_r0_ngram",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": null,
"workload": "code_edit",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 106,
"completion_tokens": 48,
"load_time_s": 2.011,
"ttft_s": 4.906,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 22.969,
"total_time_s": 27.875,
"decode_tok_per_s": 2.046,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 1.722,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 6,
"drafted_tokens": 50,
"accepted_tokens": 31,
"resident_layers": 0,
"streamed_layers": 48,
"n_layers": 48,
"resident_mib": 1026.8,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 45236.88,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 2503.1,
"current_rss_mib": 2503.1,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 0.2,
"page_cache_delta_since_start_mib": 1.5,
"idle_watts": null,
"avg_watts": 11.03,
"net_watts": null,
"joules_per_tok": 6.41,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": "#[derive(Debug, Clone)]\npub struct StreamConfig {\npub model_path: PathBuf,\npub n_ctx: u32,\npub n_batch: u32,\npub resident_layers: Option<u32>,\npub vram_budget_b"
},
{
"round": "round5",
"source_file": "r5_zc_stream_r0_ngram.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r5_zc_stream_r0_ngram",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": null,
"workload": "open_ended",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 31,
"completion_tokens": 48,
"load_time_s": 2.011,
"ttft_s": 4.612,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 128.056,
"total_time_s": 132.668,
"decode_tok_per_s": 0.367,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.362,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 29,
"drafted_tokens": 30,
"accepted_tokens": 19,
"resident_layers": 0,
"streamed_layers": 48,
"n_layers": 48,
"resident_mib": 1026.8,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 218644.93,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 2505.1,
"current_rss_mib": 2505.1,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 1.5,
"page_cache_delta_since_start_mib": 3.0,
"idle_watts": null,
"avg_watts": 10.97,
"net_watts": null,
"joules_per_tok": 30.31,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " When streaming transformer layers on an 8 GB edge device, the interaction between page-aligned O_DIRECT NVMe reads and asynchronous ping-pong GPU buffers is cr"
},
{
"round": "round6",
"source_file": "r6_auto_budget_6000_draft05b.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r6_auto_budget_6000_draft05b",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "prompt_heavy",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 340,
"completion_tokens": 32,
"load_time_s": 6.615,
"ttft_s": 7.057,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 77.232,
"total_time_s": 84.289,
"decode_tok_per_s": 0.401,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.38,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 16,
"drafted_tokens": 107,
"accepted_tokens": 16,
"resident_layers": 17,
"streamed_layers": 31,
"n_layers": 48,
"resident_mib": 3706.3,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 77758.38,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 5715.0,
"current_rss_mib": 5715.0,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -91.1,
"page_cache_delta_since_start_mib": 151.9,
"idle_watts": null,
"avg_watts": 10.11,
"net_watts": null,
"joules_per_tok": 26.62,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " Layer-streaming inference manages memory by allocating a small, fixed pool of buffers for active computation, keeping only essential layers in memory, and asyn"
},
{
"round": "round6",
"source_file": "r6_auto_budget_6000_draft05b.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r6_auto_budget_6000_draft05b",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "code_edit",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 106,
"completion_tokens": 48,
"load_time_s": 6.615,
"ttft_s": 5.474,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 20.38,
"total_time_s": 25.854,
"decode_tok_per_s": 2.306,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 1.857,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 5,
"drafted_tokens": 44,
"accepted_tokens": 36,
"resident_layers": 17,
"streamed_layers": 31,
"n_layers": 48,
"resident_mib": 3706.3,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 24299.49,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 5719.7,
"current_rss_mib": 5719.7,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 0.4,
"page_cache_delta_since_start_mib": 152.4,
"idle_watts": null,
"avg_watts": 10.14,
"net_watts": null,
"joules_per_tok": 5.46,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": "#[derive(Debug, Clone)]\npub struct StreamConfig {\npub model_path: PathBuf,\npub n_ctx: u32,\npub n_batch: u32,\npub resident_layers: Option<u32>,\npub vram_budget_b"
},
{
"round": "round6",
"source_file": "r6_auto_budget_6000_draft05b.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r6_auto_budget_6000_draft05b",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "open_ended",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 31,
"completion_tokens": 48,
"load_time_s": 6.615,
"ttft_s": 5.291,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 51.813,
"total_time_s": 57.105,
"decode_tok_per_s": 0.907,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.841,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 11,
"drafted_tokens": 115,
"accepted_tokens": 34,
"resident_layers": 17,
"streamed_layers": 31,
"n_layers": 48,
"resident_mib": 3706.3,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 53458.88,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 5719.7,
"current_rss_mib": 5719.7,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 0.1,
"page_cache_delta_since_start_mib": 152.5,
"idle_watts": null,
"avg_watts": 9.78,
"net_watts": null,
"joules_per_tok": 11.64,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " When streaming transformer layers on an 8 GB edge device, the interaction between page-aligned O_DIRECT NVMe reads and asynchronous ping-pong GPU buffers is cr"
},
{
"round": "round6",
"source_file": "r6_hybrid_r12_draft15b.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r6_hybrid_r12_draft15b",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-1.5b-instruct-q4_k_m.gguf",
"workload": "prompt_heavy",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 340,
"completion_tokens": 32,
"load_time_s": 6.343,
"ttft_s": 8.156,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 88.085,
"total_time_s": 96.241,
"decode_tok_per_s": 0.352,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.332,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 16,
"drafted_tokens": 103,
"accepted_tokens": 16,
"resident_layers": 12,
"streamed_layers": 36,
"n_layers": 48,
"resident_mib": 2949.0,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 90023.29,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 5580.4,
"current_rss_mib": 5580.4,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -82.5,
"page_cache_delta_since_start_mib": 391.5,
"idle_watts": null,
"avg_watts": 10.05,
"net_watts": null,
"joules_per_tok": 30.22,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " Layer-streaming inference manages memory by allocating a small, fixed pool of buffers for active computation, keeping only essential layers in memory, and asyn"
},
{
"round": "round6",
"source_file": "r6_hybrid_r12_draft15b.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r6_hybrid_r12_draft15b",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-1.5b-instruct-q4_k_m.gguf",
"workload": "code_edit",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 106,
"completion_tokens": 48,
"load_time_s": 6.343,
"ttft_s": 6.099,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 23.446,
"total_time_s": 29.545,
"decode_tok_per_s": 2.005,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 1.625,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 5,
"drafted_tokens": 44,
"accepted_tokens": 36,
"resident_layers": 12,
"streamed_layers": 36,
"n_layers": 48,
"resident_mib": 2949.0,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 28086.11,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 5585.0,
"current_rss_mib": 5585.0,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -69.6,
"page_cache_delta_since_start_mib": 321.9,
"idle_watts": null,
"avg_watts": 9.65,
"net_watts": null,
"joules_per_tok": 5.94,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": "#[derive(Debug, Clone)]\npub struct StreamConfig {\npub model_path: PathBuf,\npub n_ctx: u32,\npub n_batch: u32,\npub resident_layers: Option<u32>,\npub vram_budget_b"
},
{
"round": "round6",
"source_file": "r6_hybrid_r12_draft15b.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r6_hybrid_r12_draft15b",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-1.5b-instruct-q4_k_m.gguf",
"workload": "open_ended",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 31,
"completion_tokens": 48,
"load_time_s": 6.343,
"ttft_s": 6.014,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 63.741,
"total_time_s": 69.755,
"decode_tok_per_s": 0.737,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.688,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 10,
"drafted_tokens": 94,
"accepted_tokens": 35,
"resident_layers": 12,
"streamed_layers": 36,
"n_layers": 48,
"resident_mib": 2949.0,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 56172.23,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 5585.0,
"current_rss_mib": 5585.0,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 2.4,
"page_cache_delta_since_start_mib": 324.3,
"idle_watts": null,
"avg_watts": 9.78,
"net_watts": null,
"joules_per_tok": 14.22,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " When streaming transformer layers on an 8 GB edge device, the interaction between page-aligned O_DIRECT NVMe reads and asynchronous ping-pong GPU buffers is cr"
},
{
"round": "round6",
"source_file": "r6_hybrid_r19_draft05b_temp07.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r6_hybrid_r19_draft05b_temp07",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "prompt_heavy",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 340,
"completion_tokens": 32,
"load_time_s": 7.145,
"ttft_s": 3.976,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 20.457,
"total_time_s": 24.432,
"decode_tok_per_s": 1.515,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 1.31,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 11,
"drafted_tokens": 75,
"accepted_tokens": 21,
"resident_layers": 19,
"streamed_layers": 29,
"n_layers": 48,
"resident_mib": 4020.4,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 50003.29,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 6037.1,
"current_rss_mib": 6037.1,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -33.1,
"page_cache_delta_since_start_mib": -382.9,
"idle_watts": null,
"avg_watts": 13.05,
"net_watts": null,
"joules_per_tok": 9.97,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " The core memory-management mechanism described in the document is layer-streaming inference, which manages to fit large language models into limited memory by "
},
{
"round": "round6",
"source_file": "r6_hybrid_r19_draft05b_temp07.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r6_hybrid_r19_draft05b_temp07",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "code_edit",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 106,
"completion_tokens": 48,
"load_time_s": 7.145,
"ttft_s": 2.358,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 7.928,
"total_time_s": 10.286,
"decode_tok_per_s": 5.928,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 4.666,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 5,
"drafted_tokens": 44,
"accepted_tokens": 36,
"resident_layers": 19,
"streamed_layers": 29,
"n_layers": 48,
"resident_mib": 4020.4,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 22728.77,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 6042.3,
"current_rss_mib": 6042.3,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -35.3,
"page_cache_delta_since_start_mib": -418.2,
"idle_watts": null,
"avg_watts": 13.81,
"net_watts": null,
"joules_per_tok": 2.96,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": "#[derive(Debug, Clone)]\npub struct StreamConfig {\npub model_path: PathBuf,\npub n_ctx: u32,\npub n_batch: u32,\npub resident_layers: Option<u32>,\npub vram_budget_b"
},
{
"round": "round6",
"source_file": "r6_hybrid_r19_draft05b_temp07.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r6_hybrid_r19_draft05b_temp07",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "open_ended",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 31,
"completion_tokens": 48,
"load_time_s": 7.145,
"ttft_s": 2.168,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 20.461,
"total_time_s": 22.629,
"decode_tok_per_s": 2.297,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 2.121,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 11,
"drafted_tokens": 107,
"accepted_tokens": 32,
"resident_layers": 19,
"streamed_layers": 29,
"n_layers": 48,
"resident_mib": 4020.4,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 50003.29,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 6042.3,
"current_rss_mib": 6042.3,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 4.3,
"page_cache_delta_since_start_mib": -413.9,
"idle_watts": null,
"avg_watts": 13.11,
"net_watts": null,
"joules_per_tok": 6.18,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " The interaction between page-aligned O_DIRECT NVMe reads and asynchronous ping-pong GPU buffers in the context of streaming transformer layers on an 8 GB edge "
},
{
"round": "round6",
"source_file": "r6_hybrid_r19_draft05b_th2.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r6_hybrid_r19_draft05b_th2",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "prompt_heavy",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 340,
"completion_tokens": 32,
"load_time_s": 7.943,
"ttft_s": 3.957,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 31.355,
"total_time_s": 35.312,
"decode_tok_per_s": 0.989,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.906,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 16,
"drafted_tokens": 107,
"accepted_tokens": 16,
"resident_layers": 19,
"streamed_layers": 29,
"n_layers": 48,
"resident_mib": 4020.4,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 72732.06,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 5927.6,
"current_rss_mib": 5927.6,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -76.5,
"page_cache_delta_since_start_mib": -829.4,
"idle_watts": null,
"avg_watts": 12.61,
"net_watts": null,
"joules_per_tok": 13.92,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " Layer-streaming inference manages memory by allocating a small, fixed pool of buffers for active computation, keeping only essential layers in memory, and asyn"
},
{
"round": "round6",
"source_file": "r6_hybrid_r19_draft05b_th2.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r6_hybrid_r19_draft05b_th2",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "code_edit",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 106,
"completion_tokens": 48,
"load_time_s": 7.943,
"ttft_s": 2.355,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 7.874,
"total_time_s": 10.23,
"decode_tok_per_s": 5.969,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 4.692,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 5,
"drafted_tokens": 44,
"accepted_tokens": 36,
"resident_layers": 19,
"streamed_layers": 29,
"n_layers": 48,
"resident_mib": 4020.4,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 22728.77,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 5932.5,
"current_rss_mib": 5932.5,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -11.2,
"page_cache_delta_since_start_mib": -840.6,
"idle_watts": null,
"avg_watts": 13.02,
"net_watts": null,
"joules_per_tok": 2.78,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": "#[derive(Debug, Clone)]\npub struct StreamConfig {\npub model_path: PathBuf,\npub n_ctx: u32,\npub n_batch: u32,\npub resident_layers: Option<u32>,\npub vram_budget_b"
},
{
"round": "round6",
"source_file": "r6_hybrid_r19_draft05b_th2.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r6_hybrid_r19_draft05b_th2",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "open_ended",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 31,
"completion_tokens": 48,
"load_time_s": 7.943,
"ttft_s": 2.189,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 32.682,
"total_time_s": 34.871,
"decode_tok_per_s": 1.438,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 1.377,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 11,
"drafted_tokens": 115,
"accepted_tokens": 34,
"resident_layers": 19,
"streamed_layers": 29,
"n_layers": 48,
"resident_mib": 4020.4,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 50003.29,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 5932.7,
"current_rss_mib": 5932.7,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -2.4,
"page_cache_delta_since_start_mib": -843.0,
"idle_watts": null,
"avg_watts": 11.23,
"net_watts": null,
"joules_per_tok": 8.16,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " When streaming transformer layers on an 8 GB edge device, the interaction between page-aligned O_DIRECT NVMe reads and asynchronous ping-pong GPU buffers is cr"
},
{
"round": "round6",
"source_file": "r6_hybrid_r19_ngram.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r6_hybrid_r19_ngram",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": null,
"workload": "prompt_heavy",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 340,
"completion_tokens": 32,
"load_time_s": 6.029,
"ttft_s": 4.353,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 64.64,
"total_time_s": 68.993,
"decode_tok_per_s": 0.48,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.464,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 25,
"drafted_tokens": 37,
"accepted_tokens": 7,
"resident_layers": 19,
"streamed_layers": 29,
"n_layers": 48,
"resident_mib": 4020.4,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 113643.85,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 5458.6,
"current_rss_mib": 5458.6,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 3.9,
"page_cache_delta_since_start_mib": -45.7,
"idle_watts": null,
"avg_watts": 11.94,
"net_watts": null,
"joules_per_tok": 25.74,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " Layer-streaming inference manages memory by allocating a small, fixed pool of buffers for active computation, keeping only essential layers in memory, and asyn"
},
{
"round": "round6",
"source_file": "r6_hybrid_r19_ngram.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r6_hybrid_r19_ngram",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": null,
"workload": "code_edit",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 106,
"completion_tokens": 48,
"load_time_s": 6.029,
"ttft_s": 3.152,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 13.547,
"total_time_s": 16.699,
"decode_tok_per_s": 3.469,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 2.874,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 6,
"drafted_tokens": 50,
"accepted_tokens": 31,
"resident_layers": 19,
"streamed_layers": 29,
"n_layers": 48,
"resident_mib": 4020.4,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 27274.52,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 5460.9,
"current_rss_mib": 5460.9,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 0.7,
"page_cache_delta_since_start_mib": -45.0,
"idle_watts": null,
"avg_watts": 11.95,
"net_watts": null,
"joules_per_tok": 4.16,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": "#[derive(Debug, Clone)]\npub struct StreamConfig {\npub model_path: PathBuf,\npub n_ctx: u32,\npub n_batch: u32,\npub resident_layers: Option<u32>,\npub vram_budget_b"
},
{
"round": "round6",
"source_file": "r6_hybrid_r19_ngram.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r6_hybrid_r19_ngram",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": null,
"workload": "open_ended",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 31,
"completion_tokens": 48,
"load_time_s": 6.029,
"ttft_s": 2.87,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 75.455,
"total_time_s": 78.325,
"decode_tok_per_s": 0.623,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.613,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 29,
"drafted_tokens": 30,
"accepted_tokens": 19,
"resident_layers": 19,
"streamed_layers": 29,
"n_layers": 48,
"resident_mib": 4020.4,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 131826.86,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 5463.0,
"current_rss_mib": 5462.9,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 22.1,
"page_cache_delta_since_start_mib": -23.0,
"idle_watts": null,
"avg_watts": 11.6,
"net_watts": null,
"joules_per_tok": 18.93,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " When streaming transformer layers on an 8 GB edge device, the interaction between page-aligned O_DIRECT NVMe reads and asynchronous ping-pong GPU buffers is cr"
},
{
"round": "round6",
"source_file": "r6_hybrid_r20_draft05b.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r6_hybrid_r20_draft05b",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "prompt_heavy",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 340,
"completion_tokens": 32,
"load_time_s": 6.749,
"ttft_s": 4.078,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 30.016,
"total_time_s": 34.094,
"decode_tok_per_s": 1.033,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.939,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 16,
"drafted_tokens": 107,
"accepted_tokens": 16,
"resident_layers": 20,
"streamed_layers": 28,
"n_layers": 48,
"resident_mib": 4168.1,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 70534.86,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 6152.5,
"current_rss_mib": 6069.9,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -74.6,
"page_cache_delta_since_start_mib": -171.5,
"idle_watts": null,
"avg_watts": 12.89,
"net_watts": null,
"joules_per_tok": 13.73,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " Layer-streaming inference manages memory by allocating a small, fixed pool of buffers for active computation, keeping only essential layers in memory, and asyn"
},
{
"round": "round6",
"source_file": "r6_hybrid_r20_draft05b.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r6_hybrid_r20_draft05b",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "code_edit",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 106,
"completion_tokens": 48,
"load_time_s": 6.749,
"ttft_s": 2.305,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 7.868,
"total_time_s": 10.173,
"decode_tok_per_s": 5.974,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 4.718,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 5,
"drafted_tokens": 44,
"accepted_tokens": 36,
"resident_layers": 20,
"streamed_layers": 28,
"n_layers": 48,
"resident_mib": 4168.1,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 21990.14,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 6152.5,
"current_rss_mib": 6074.6,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 0.7,
"page_cache_delta_since_start_mib": -170.8,
"idle_watts": null,
"avg_watts": 12.89,
"net_watts": null,
"joules_per_tok": 2.73,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": "#[derive(Debug, Clone)]\npub struct StreamConfig {\npub model_path: PathBuf,\npub n_ctx: u32,\npub n_batch: u32,\npub resident_layers: Option<u32>,\npub vram_budget_b"
},
{
"round": "round6",
"source_file": "r6_hybrid_r20_draft05b.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r6_hybrid_r20_draft05b",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "open_ended",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 31,
"completion_tokens": 48,
"load_time_s": 6.749,
"ttft_s": 2.124,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 22.288,
"total_time_s": 24.412,
"decode_tok_per_s": 2.109,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 1.966,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 11,
"drafted_tokens": 115,
"accepted_tokens": 34,
"resident_layers": 20,
"streamed_layers": 28,
"n_layers": 48,
"resident_mib": 4168.1,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 48378.3,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 6152.5,
"current_rss_mib": 6074.7,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -2.6,
"page_cache_delta_since_start_mib": -173.4,
"idle_watts": null,
"avg_watts": 12.84,
"net_watts": null,
"joules_per_tok": 6.53,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " When streaming transformer layers on an 8 GB edge device, the interaction between page-aligned O_DIRECT NVMe reads and asynchronous ping-pong GPU buffers is cr"
},
{
"round": "round6",
"source_file": "r6_multiturn_long_r19_draft05b.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r6_multiturn_long_r19_draft05b_multiturn",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": null,
"workload": "multi_turn_prefix_reuse",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": null,
"completion_tokens": null,
"load_time_s": null,
"ttft_s": null,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": null,
"total_time_s": null,
"decode_tok_per_s": null,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": null,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": null,
"drafted_tokens": null,
"accepted_tokens": null,
"resident_layers": null,
"streamed_layers": null,
"n_layers": null,
"resident_mib": null,
"slot_mib": null,
"expert_slot_mib": null,
"io_mib_read": null,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": null,
"is_unified_memory": null,
"direct_io_active": null,
"peak_rss_mib": null,
"current_rss_mib": null,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": null,
"page_cache_delta_since_start_mib": null,
"idle_watts": null,
"avg_watts": null,
"net_watts": null,
"joules_per_tok": null,
"net_joules_per_tok": null,
"turn1_prompt_tokens": 1377,
"turn1_ttft_s": 16.705,
"turn2_prompt_tokens": 1414,
"turn2_ttft_s": 3.416,
"ttft_speedup_x": 4.89,
"notes": null,
"output_preview": " O_DIRECT prevents page-cache thrashing by bypassing the operating system's page cache"
},
{
"round": "round6",
"source_file": "r6_stream_r0_th2.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r6_stream_r0_th2",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": null,
"workload": "prompt_heavy",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 340,
"completion_tokens": 32,
"load_time_s": 1.913,
"ttft_s": 7.374,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 174.278,
"total_time_s": 181.652,
"decode_tok_per_s": 0.178,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.176,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 32,
"drafted_tokens": 0,
"accepted_tokens": 0,
"resident_layers": 0,
"streamed_layers": 48,
"n_layers": 48,
"resident_mib": 1026.8,
"slot_mib": 166.4,
"expert_slot_mib": null,
"io_mib_read": 241429.79,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 2484.7,
"current_rss_mib": 2484.7,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 19.2,
"page_cache_delta_since_start_mib": 25.6,
"idle_watts": null,
"avg_watts": 10.58,
"net_watts": null,
"joules_per_tok": 60.05,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " Layer-streaming inference manages memory by allocating a small, fixed pool of buffers for active computation, keeping only essential layers in memory, and asyn"
},
{
"round": "round6_32b",
"source_file": "r6_32b_auto5600_draft05b.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r6_32b_auto5600_draft05b",
"status": "ok",
"model": "Qwen2.5-32B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "prompt_heavy",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 340,
"completion_tokens": 32,
"load_time_s": 6.838,
"ttft_s": 11.69,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 92.127,
"total_time_s": 103.817,
"decode_tok_per_s": 0.336,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.308,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 13,
"drafted_tokens": 87,
"accepted_tokens": 19,
"resident_layers": 6,
"streamed_layers": 58,
"n_layers": 64,
"resident_mib": 2813.1,
"slot_mib": 297.7,
"expert_slot_mib": null,
"io_mib_read": 209768.71,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 5289.1,
"current_rss_mib": 5289.1,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -59.0,
"page_cache_delta_since_start_mib": -4749.6,
"idle_watts": null,
"avg_watts": 12.24,
"net_watts": null,
"joules_per_tok": 39.7,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " The core memory-management mechanism described involves layer-streaming inference, which uses a small fixed pool of reusable compute buffers and asynchronously"
},
{
"round": "round6_32b",
"source_file": "r6_32b_auto5600_draft05b.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r6_32b_auto5600_draft05b",
"status": "ok",
"model": "Qwen2.5-32B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "code_edit",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 106,
"completion_tokens": 48,
"load_time_s": 6.838,
"ttft_s": 8.416,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 49.8,
"total_time_s": 58.216,
"decode_tok_per_s": 0.944,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.825,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 7,
"drafted_tokens": 66,
"accepted_tokens": 41,
"resident_layers": 6,
"streamed_layers": 58,
"n_layers": 64,
"resident_mib": 2813.1,
"slot_mib": 297.7,
"expert_slot_mib": null,
"io_mib_read": 112605.38,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 5293.8,
"current_rss_mib": 5293.8,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -16.0,
"page_cache_delta_since_start_mib": -4765.5,
"idle_watts": null,
"avg_watts": 11.79,
"net_watts": null,
"joules_per_tok": 14.3,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": "#[derive(Debug, Clone)]\npub struct StreamConfig {\n pub model_path: PathBuf,\n pub n_ctx: u32,\n pub n_batch: u32,\n pub resident_layers: Option<u32>,\n "
},
{
"round": "round6_32b",
"source_file": "r6_32b_auto5600_draft05b.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r6_32b_auto5600_draft05b",
"status": "ok",
"model": "Qwen2.5-32B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "open_ended",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 31,
"completion_tokens": 48,
"load_time_s": 6.838,
"ttft_s": 7.916,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 92.905,
"total_time_s": 100.821,
"decode_tok_per_s": 0.506,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.476,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 13,
"drafted_tokens": 127,
"accepted_tokens": 35,
"resident_layers": 6,
"streamed_layers": 58,
"n_layers": 64,
"resident_mib": 2813.1,
"slot_mib": 297.7,
"expert_slot_mib": null,
"io_mib_read": 209657.66,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 5293.8,
"current_rss_mib": 5293.8,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -0.1,
"page_cache_delta_since_start_mib": -4765.6,
"idle_watts": null,
"avg_watts": 11.96,
"net_watts": null,
"joules_per_tok": 25.12,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " When streaming transformer layers on an 8 GB edge device, the interaction between page-aligned O_DIRECT NVMe reads and asynchronous ping-pong GPU buffers can b"
},
{
"round": "round6_32b",
"source_file": "r6_32b_auto5600_draft05b_temp07.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r6_32b_auto5600_draft05b_temp07",
"status": "ok",
"model": "Qwen2.5-32B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "prompt_heavy",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 340,
"completion_tokens": 32,
"load_time_s": 5.459,
"ttft_s": 16.181,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 146.116,
"total_time_s": 162.297,
"decode_tok_per_s": 0.212,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.197,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 13,
"drafted_tokens": 89,
"accepted_tokens": 19,
"resident_layers": 6,
"streamed_layers": 58,
"n_layers": 64,
"resident_mib": 2813.1,
"slot_mib": 297.7,
"expert_slot_mib": null,
"io_mib_read": 209582.03,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 5293.9,
"current_rss_mib": 5293.9,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -95.5,
"page_cache_delta_since_start_mib": 5.7,
"idle_watts": null,
"avg_watts": 10.75,
"net_watts": null,
"joules_per_tok": 54.52,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " The core memory-management mechanism described involves layer-streaming inference, where a fixed pool of reusable compute buffers is allocated for transformer "
},
{
"round": "round6_32b",
"source_file": "r6_32b_auto5600_draft05b_temp07.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r6_32b_auto5600_draft05b_temp07",
"status": "ok",
"model": "Qwen2.5-32B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "code_edit",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 106,
"completion_tokens": 48,
"load_time_s": 5.459,
"ttft_s": 12.943,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 73.28,
"total_time_s": 86.223,
"decode_tok_per_s": 0.641,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.557,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 7,
"drafted_tokens": 66,
"accepted_tokens": 41,
"resident_layers": 6,
"streamed_layers": 58,
"n_layers": 64,
"resident_mib": 2813.1,
"slot_mib": 297.7,
"expert_slot_mib": null,
"io_mib_read": 112792.07,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 5299.2,
"current_rss_mib": 5299.2,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 0.0,
"page_cache_delta_since_start_mib": 5.7,
"idle_watts": null,
"avg_watts": 10.84,
"net_watts": null,
"joules_per_tok": 19.47,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": "#[derive(Debug, Clone)]\npub struct StreamConfig {\n pub model_path: PathBuf,\n pub n_ctx: u32,\n pub n_batch: u32,\n pub resident_layers: Option<u32>,\n "
},
{
"round": "round6_32b",
"source_file": "r6_32b_auto5600_draft05b_temp07.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r6_32b_auto5600_draft05b_temp07",
"status": "ok",
"model": "Qwen2.5-32B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "open_ended",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 31,
"completion_tokens": 48,
"load_time_s": 5.459,
"ttft_s": 12.177,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 121.648,
"total_time_s": 133.825,
"decode_tok_per_s": 0.386,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.359,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 11,
"drafted_tokens": 98,
"accepted_tokens": 37,
"resident_layers": 6,
"streamed_layers": 58,
"n_layers": 64,
"resident_mib": 2813.1,
"slot_mib": 297.7,
"expert_slot_mib": null,
"io_mib_read": 177244.68,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 5299.2,
"current_rss_mib": 5299.2,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -16.2,
"page_cache_delta_since_start_mib": -10.5,
"idle_watts": null,
"avg_watts": 10.53,
"net_watts": null,
"joules_per_tok": 29.35,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " In the context of streaming transformer layers on an 8 GB edge device, understanding how page-aligned O_DIRECT NVMe reads and asynchronous ping-pong GPU buffer"
},
{
"round": "round6_32b",
"source_file": "r6_32b_stream_r0.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r6_32b_stream_r0",
"status": "ok",
"model": "Qwen2.5-32B-Instruct-Q4_K_M.gguf",
"draft_model": null,
"workload": "prompt_heavy",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 340,
"completion_tokens": 32,
"load_time_s": 2.137,
"ttft_s": 14.479,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 330.254,
"total_time_s": 344.733,
"decode_tok_per_s": 0.094,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.093,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": 32,
"drafted_tokens": 0,
"accepted_tokens": 0,
"resident_layers": 0,
"streamed_layers": 64,
"n_layers": 64,
"resident_mib": 1026.8,
"slot_mib": 297.7,
"expert_slot_mib": null,
"io_mib_read": 573082.1,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": false,
"is_unified_memory": null,
"direct_io_active": true,
"peak_rss_mib": 2957.2,
"current_rss_mib": 2957.2,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 6.9,
"page_cache_delta_since_start_mib": 8.2,
"idle_watts": null,
"avg_watts": 11.06,
"net_watts": null,
"joules_per_tok": 119.1,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " The core memory-management mechanism described involves layer-streaming inference, which uses a small fixed pool of reusable compute buffers and asynchronously"
},
{
"round": "round8_v010",
"source_file": "r8_14b_auto_draft05b.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r8_14b_auto_draft05b",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "prompt_heavy",
"chat_template_applied": false,
"warmup": true,
"reps": 3,
"temperature": 0.0,
"seed": 42,
"threads": 4,
"kv_cache_type": "q8_0",
"vram_budget_auto": true,
"vram_budget_mb": 0,
"prompt_tokens": 340,
"completion_tokens": 32,
"load_time_s": 4.936,
"ttft_s": 4.467,
"ttft_s_min": 4.464,
"ttft_s_max": 4.504,
"decode_time_s": 40.794,
"total_time_s": 45.286,
"decode_tok_per_s": 0.76,
"decode_tok_per_s_min": 0.755,
"decode_tok_per_s_max": 0.76,
"total_tok_per_s": 0.707,
"total_tok_per_s_min": 0.703,
"total_tok_per_s_max": 0.707,
"sweeps": 15,
"drafted_tokens": 77,
"accepted_tokens": 18,
"resident_layers": 4,
"streamed_layers": 44,
"n_layers": 48,
"resident_mib": 1692.4,
"slot_mib": 166.4,
"expert_slot_mib": 0.0,
"io_mib_read": 103107.13,
"pread_mib_s": 2796.4,
"effective_io_mib_s": 2276.8,
"expert_cache_hits": 0,
"expert_cache_misses": 0,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": 0.0,
"moe_layers": 0,
"moe_active_streaming": false,
"uma_zero_copy": false,
"is_unified_memory": true,
"direct_io_active": true,
"peak_rss_mib": 3709.2,
"current_rss_mib": 3709.2,
"mem_avail_before_mib": 2397.2,
"mem_avail_after_mib": 2403.7,
"swap_delta_mib": -0.3,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -0.4,
"page_cache_delta_since_start_mib": -1657.4,
"idle_watts": 7.77,
"avg_watts": 13.21,
"net_watts": 5.43,
"joules_per_tok": 18.69,
"net_joules_per_tok": 7.69,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " Layer-streaming inference manages memory by allocating a small, fixed pool of reusable buffers for compute, keeping essential layers in memory, and asynchronou"
},
{
"round": "round8_v010",
"source_file": "r8_14b_auto_draft05b.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r8_14b_auto_draft05b",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "code_edit",
"chat_template_applied": false,
"warmup": true,
"reps": 3,
"temperature": 0.0,
"seed": 42,
"threads": 4,
"kv_cache_type": "q8_0",
"vram_budget_auto": true,
"vram_budget_mb": 0,
"prompt_tokens": 106,
"completion_tokens": 48,
"load_time_s": 4.936,
"ttft_s": 3.258,
"ttft_s_min": 3.236,
"ttft_s_max": 3.281,
"decode_time_s": 14.884,
"total_time_s": 18.158,
"decode_tok_per_s": 3.158,
"decode_tok_per_s_min": 2.244,
"decode_tok_per_s_max": 3.159,
"total_tok_per_s": 2.643,
"total_tok_per_s_min": 1.985,
"total_tok_per_s_max": 2.646,
"sweeps": 6,
"drafted_tokens": 50,
"accepted_tokens": 43,
"resident_layers": 4,
"streamed_layers": 44,
"n_layers": 48,
"resident_mib": 1692.4,
"slot_mib": 166.4,
"expert_slot_mib": 0.0,
"io_mib_read": 41242.85,
"pread_mib_s": 2799.5,
"effective_io_mib_s": 2273.3,
"expert_cache_hits": 0,
"expert_cache_misses": 0,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": 0.0,
"moe_layers": 0,
"moe_active_streaming": false,
"uma_zero_copy": false,
"is_unified_memory": true,
"direct_io_active": true,
"peak_rss_mib": 3713.9,
"current_rss_mib": 3713.9,
"mem_avail_before_mib": 2382.9,
"mem_avail_after_mib": 2385.3,
"swap_delta_mib": -0.3,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 0.2,
"page_cache_delta_since_start_mib": -1654.7,
"idle_watts": 7.77,
"avg_watts": 12.88,
"net_watts": 5.11,
"joules_per_tok": 4.87,
"net_joules_per_tok": 1.93,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": "#[derive(Debug, Clone)]\npub struct StreamConfig {\npub model_path: PathBuf,\npub n_ctx: u32,\npub n_batch: u32,\npub resident_layers: Option<u32>,\npub vram_budget_b"
},
{
"round": "round8_v010",
"source_file": "r8_14b_auto_draft05b.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r8_14b_auto_draft05b",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "open_ended",
"chat_template_applied": false,
"warmup": true,
"reps": 3,
"temperature": 0.0,
"seed": 42,
"threads": 4,
"kv_cache_type": "q8_0",
"vram_budget_auto": true,
"vram_budget_mb": 0,
"prompt_tokens": 31,
"completion_tokens": 48,
"load_time_s": 4.936,
"ttft_s": 2.952,
"ttft_s_min": 2.921,
"ttft_s_max": 2.995,
"decode_time_s": 29.778,
"total_time_s": 32.699,
"decode_tok_per_s": 1.578,
"decode_tok_per_s_min": 1.564,
"decode_tok_per_s_max": 1.582,
"total_tok_per_s": 1.468,
"total_tok_per_s_min": 1.452,
"total_tok_per_s_max": 1.469,
"sweeps": 11,
"drafted_tokens": 97,
"accepted_tokens": 38,
"resident_layers": 4,
"streamed_layers": 44,
"n_layers": 48,
"resident_mib": 1692.4,
"slot_mib": 166.4,
"expert_slot_mib": 0.0,
"io_mib_read": 75611.89,
"pread_mib_s": 2794.3,
"effective_io_mib_s": 2312.4,
"expert_cache_hits": 0,
"expert_cache_misses": 0,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": 0.0,
"moe_layers": 0,
"moe_active_streaming": false,
"uma_zero_copy": false,
"is_unified_memory": true,
"direct_io_active": true,
"peak_rss_mib": 3715.9,
"current_rss_mib": 3715.9,
"mem_avail_before_mib": 2387.5,
"mem_avail_after_mib": 2388.8,
"swap_delta_mib": 0.0,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 0.8,
"page_cache_delta_since_start_mib": -1654.0,
"idle_watts": 7.77,
"avg_watts": 12.84,
"net_watts": 5.06,
"joules_per_tok": 8.75,
"net_joules_per_tok": 3.45,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " When streaming transformer layers on an 8 GB edge device, the interaction between page-aligned O_DIRECT NVMe reads and asynchronous ping-pong GPU buffers is cr"
},
{
"round": "round8_v010",
"source_file": "r8_14b_auto_draft05b.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r8_14b_auto_draft05b_multiturn",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": null,
"workload": "multi_turn_prefix_reuse",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": null,
"completion_tokens": null,
"load_time_s": null,
"ttft_s": null,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": null,
"total_time_s": null,
"decode_tok_per_s": null,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": null,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": null,
"drafted_tokens": null,
"accepted_tokens": null,
"resident_layers": null,
"streamed_layers": null,
"n_layers": null,
"resident_mib": null,
"slot_mib": null,
"expert_slot_mib": null,
"io_mib_read": null,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": null,
"is_unified_memory": null,
"direct_io_active": null,
"peak_rss_mib": null,
"current_rss_mib": null,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": null,
"page_cache_delta_since_start_mib": null,
"idle_watts": null,
"avg_watts": null,
"net_watts": null,
"joules_per_tok": null,
"net_joules_per_tok": null,
"turn1_prompt_tokens": 1377,
"turn1_ttft_s": 14.877,
"turn2_prompt_tokens": 1414,
"turn2_ttft_s": 2.993,
"ttft_speedup_x": 4.97,
"notes": null,
"output_preview": " O_DIRECT prevents page-cache thrashing by bypassing the operating system's page cache"
},
{
"round": "round8_v010",
"source_file": "r8_14b_auto_draft05b_temp07.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r8_14b_auto_draft05b_temp07",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "prompt_heavy",
"chat_template_applied": false,
"warmup": false,
"reps": 3,
"temperature": 0.7,
"seed": 42,
"threads": 4,
"kv_cache_type": "q8_0",
"vram_budget_auto": true,
"vram_budget_mb": 0,
"prompt_tokens": 340,
"completion_tokens": 32,
"load_time_s": 3.799,
"ttft_s": 8.945,
"ttft_s_min": 8.935,
"ttft_s_max": 9.013,
"decode_time_s": 103.065,
"total_time_s": 112.011,
"decode_tok_per_s": 0.301,
"decode_tok_per_s_min": 0.263,
"decode_tok_per_s_max": 0.42,
"total_tok_per_s": 0.286,
"total_tok_per_s_min": 0.252,
"total_tok_per_s_max": 0.386,
"sweeps": 15,
"drafted_tokens": 72,
"accepted_tokens": 18,
"resident_layers": 3,
"streamed_layers": 45,
"n_layers": 48,
"resident_mib": 1526.0,
"slot_mib": 166.4,
"expert_slot_mib": 0.0,
"io_mib_read": 105603.4,
"pread_mib_s": 1031.5,
"effective_io_mib_s": 942.8,
"expert_cache_hits": 0,
"expert_cache_misses": 0,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": 0.0,
"moe_layers": 0,
"moe_active_streaming": false,
"uma_zero_copy": false,
"is_unified_memory": true,
"direct_io_active": true,
"peak_rss_mib": 3542.2,
"current_rss_mib": 3542.2,
"mem_avail_before_mib": 2538.7,
"mem_avail_after_mib": 2522.2,
"swap_delta_mib": -0.5,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 1.1,
"page_cache_delta_since_start_mib": 11.2,
"idle_watts": 7.84,
"avg_watts": 9.95,
"net_watts": 2.11,
"joules_per_tok": 34.83,
"net_joules_per_tok": 7.37,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " Layer-streaming inference manages memory by allocating a small, fixed pool of compute buffers and asynchronously streaming transformer blocks from NVMe storage"
},
{
"round": "round8_v010",
"source_file": "r8_14b_auto_draft05b_temp07.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r8_14b_auto_draft05b_temp07",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "code_edit",
"chat_template_applied": false,
"warmup": false,
"reps": 3,
"temperature": 0.7,
"seed": 42,
"threads": 4,
"kv_cache_type": "q8_0",
"vram_budget_auto": true,
"vram_budget_mb": 0,
"prompt_tokens": 106,
"completion_tokens": 48,
"load_time_s": 3.799,
"ttft_s": 7.714,
"ttft_s_min": 7.712,
"ttft_s_max": 7.719,
"decode_time_s": 37.27,
"total_time_s": 44.989,
"decode_tok_per_s": 1.261,
"decode_tok_per_s_min": 1.261,
"decode_tok_per_s_max": 1.262,
"total_tok_per_s": 1.067,
"total_tok_per_s_min": 1.067,
"total_tok_per_s_max": 1.068,
"sweeps": 6,
"drafted_tokens": 50,
"accepted_tokens": 43,
"resident_layers": 3,
"streamed_layers": 45,
"n_layers": 48,
"resident_mib": 1526.0,
"slot_mib": 166.4,
"expert_slot_mib": 0.0,
"io_mib_read": 42241.36,
"pread_mib_s": 1031.2,
"effective_io_mib_s": 938.9,
"expert_cache_hits": 0,
"expert_cache_misses": 0,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": 0.0,
"moe_layers": 0,
"moe_active_streaming": false,
"uma_zero_copy": false,
"is_unified_memory": true,
"direct_io_active": true,
"peak_rss_mib": 3546.9,
"current_rss_mib": 3546.9,
"mem_avail_before_mib": 2540.0,
"mem_avail_after_mib": 2537.3,
"swap_delta_mib": -0.5,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 0.1,
"page_cache_delta_since_start_mib": 11.9,
"idle_watts": 7.84,
"avg_watts": 9.98,
"net_watts": 2.14,
"joules_per_tok": 9.35,
"net_joules_per_tok": 2.0,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": "#[derive(Debug, Clone)]\npub struct StreamConfig {\npub model_path: PathBuf,\npub n_ctx: u32,\npub n_batch: u32,\npub resident_layers: Option<u32>,\npub vram_budget_b"
},
{
"round": "round8_v010",
"source_file": "r8_14b_auto_draft05b_temp07.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r8_14b_auto_draft05b_temp07",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "open_ended",
"chat_template_applied": false,
"warmup": false,
"reps": 3,
"temperature": 0.7,
"seed": 42,
"threads": 4,
"kv_cache_type": "q8_0",
"vram_budget_auto": true,
"vram_budget_mb": 0,
"prompt_tokens": 31,
"completion_tokens": 48,
"load_time_s": 3.799,
"ttft_s": 7.421,
"ttft_s_min": 7.406,
"ttft_s_max": 7.424,
"decode_time_s": 111.118,
"total_time_s": 118.524,
"decode_tok_per_s": 0.423,
"decode_tok_per_s_min": 0.373,
"decode_tok_per_s_max": 0.454,
"total_tok_per_s": 0.405,
"total_tok_per_s_min": 0.36,
"total_tok_per_s_max": 0.432,
"sweeps": 15,
"drafted_tokens": 112,
"accepted_tokens": 34,
"resident_layers": 3,
"streamed_layers": 45,
"n_layers": 48,
"resident_mib": 1526.0,
"slot_mib": 166.4,
"expert_slot_mib": 0.0,
"io_mib_read": 105603.4,
"pread_mib_s": 1031.8,
"effective_io_mib_s": 950.4,
"expert_cache_hits": 0,
"expert_cache_misses": 0,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": 0.0,
"moe_layers": 0,
"moe_active_streaming": false,
"uma_zero_copy": false,
"is_unified_memory": true,
"direct_io_active": true,
"peak_rss_mib": 3548.9,
"current_rss_mib": 3548.9,
"mem_avail_before_mib": 2527.6,
"mem_avail_after_mib": 2521.9,
"swap_delta_mib": -0.5,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 0.5,
"page_cache_delta_since_start_mib": 34.2,
"idle_watts": 7.84,
"avg_watts": 9.97,
"net_watts": 2.12,
"joules_per_tok": 24.61,
"net_joules_per_tok": 5.24,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " In the context of running transformer layers on an 8 GB edge device, the interaction between page-aligned O_DIRECT NVMe reads and asynchronous ping-pong GPU bu"
},
{
"round": "round8_v010",
"source_file": "r8_14b_b5400_draft05b.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r8_14b_b5400_draft05b",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "prompt_heavy",
"chat_template_applied": false,
"warmup": true,
"reps": 3,
"temperature": 0.0,
"seed": 42,
"threads": 4,
"kv_cache_type": "q8_0",
"vram_budget_auto": false,
"vram_budget_mb": 5400,
"prompt_tokens": 340,
"completion_tokens": 32,
"load_time_s": 4.581,
"ttft_s": 4.238,
"ttft_s_min": 4.23,
"ttft_s_max": 7.883,
"decode_time_s": 37.618,
"total_time_s": 42.173,
"decode_tok_per_s": 0.824,
"decode_tok_per_s_min": 0.817,
"decode_tok_per_s_max": 0.827,
"total_tok_per_s": 0.759,
"total_tok_per_s_min": 0.706,
"total_tok_per_s_max": 0.765,
"sweeps": 15,
"drafted_tokens": 77,
"accepted_tokens": 18,
"resident_layers": 7,
"streamed_layers": 41,
"n_layers": 48,
"resident_mib": 2155.6,
"slot_mib": 166.4,
"expert_slot_mib": 0.0,
"io_mib_read": 96159.67,
"pread_mib_s": 2923.7,
"effective_io_mib_s": 2280.1,
"expert_cache_hits": 0,
"expert_cache_misses": 0,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": 0.0,
"moe_layers": 0,
"moe_active_streaming": false,
"uma_zero_copy": false,
"is_unified_memory": true,
"direct_io_active": true,
"peak_rss_mib": 4171.4,
"current_rss_mib": 4171.1,
"mem_avail_before_mib": 1915.6,
"mem_avail_after_mib": 1914.0,
"swap_delta_mib": 0.0,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -0.8,
"page_cache_delta_since_start_mib": -59.6,
"idle_watts": 7.7,
"avg_watts": 13.29,
"net_watts": 5.59,
"joules_per_tok": 17.51,
"net_joules_per_tok": 7.37,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " Layer-streaming inference manages memory by allocating a small, fixed pool of reusable buffers for compute, keeping essential layers in memory, and asynchronou"
},
{
"round": "round8_v010",
"source_file": "r8_14b_b5400_draft05b.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r8_14b_b5400_draft05b",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "code_edit",
"chat_template_applied": false,
"warmup": true,
"reps": 3,
"temperature": 0.0,
"seed": 42,
"threads": 4,
"kv_cache_type": "q8_0",
"vram_budget_auto": false,
"vram_budget_mb": 5400,
"prompt_tokens": 106,
"completion_tokens": 48,
"load_time_s": 4.581,
"ttft_s": 3.006,
"ttft_s_min": 3.006,
"ttft_s_max": 5.221,
"decode_time_s": 13.803,
"total_time_s": 16.809,
"decode_tok_per_s": 3.405,
"decode_tok_per_s_min": 3.284,
"decode_tok_per_s_max": 3.414,
"total_tok_per_s": 2.856,
"total_tok_per_s_min": 2.457,
"total_tok_per_s_max": 2.862,
"sweeps": 6,
"drafted_tokens": 50,
"accepted_tokens": 43,
"resident_layers": 7,
"streamed_layers": 41,
"n_layers": 48,
"resident_mib": 2155.6,
"slot_mib": 166.4,
"expert_slot_mib": 0.0,
"io_mib_read": 38463.87,
"pread_mib_s": 2961.1,
"effective_io_mib_s": 2288.3,
"expert_cache_hits": 0,
"expert_cache_misses": 0,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": 0.0,
"moe_layers": 0,
"moe_active_streaming": false,
"uma_zero_copy": false,
"is_unified_memory": true,
"direct_io_active": true,
"peak_rss_mib": 4175.9,
"current_rss_mib": 4175.8,
"mem_avail_before_mib": 1909.0,
"mem_avail_after_mib": 1907.4,
"swap_delta_mib": 0.0,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 0.2,
"page_cache_delta_since_start_mib": -66.9,
"idle_watts": 7.7,
"avg_watts": 13.16,
"net_watts": 5.47,
"joules_per_tok": 4.61,
"net_joules_per_tok": 1.91,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": "#[derive(Debug, Clone)]\npub struct StreamConfig {\npub model_path: PathBuf,\npub n_ctx: u32,\npub n_batch: u32,\npub resident_layers: Option<u32>,\npub vram_budget_b"
},
{
"round": "round8_v010",
"source_file": "r8_14b_b5400_draft05b.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r8_14b_b5400_draft05b",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "open_ended",
"chat_template_applied": false,
"warmup": true,
"reps": 3,
"temperature": 0.0,
"seed": 42,
"threads": 4,
"kv_cache_type": "q8_0",
"vram_budget_auto": false,
"vram_budget_mb": 5400,
"prompt_tokens": 31,
"completion_tokens": 48,
"load_time_s": 4.581,
"ttft_s": 2.736,
"ttft_s_min": 2.709,
"ttft_s_max": 2.765,
"decode_time_s": 27.66,
"total_time_s": 30.396,
"decode_tok_per_s": 1.699,
"decode_tok_per_s_min": 1.685,
"decode_tok_per_s_max": 1.713,
"total_tok_per_s": 1.579,
"total_tok_per_s_min": 1.568,
"total_tok_per_s_max": 1.589,
"sweeps": 11,
"drafted_tokens": 97,
"accepted_tokens": 38,
"resident_layers": 7,
"streamed_layers": 41,
"n_layers": 48,
"resident_mib": 2155.6,
"slot_mib": 166.4,
"expert_slot_mib": 0.0,
"io_mib_read": 70517.09,
"pread_mib_s": 2931.5,
"effective_io_mib_s": 2319.9,
"expert_cache_hits": 0,
"expert_cache_misses": 0,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": 0.0,
"moe_layers": 0,
"moe_active_streaming": false,
"uma_zero_copy": false,
"is_unified_memory": true,
"direct_io_active": true,
"peak_rss_mib": 4177.9,
"current_rss_mib": 4177.8,
"mem_avail_before_mib": 1906.0,
"mem_avail_after_mib": 1904.2,
"swap_delta_mib": 0.0,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -0.3,
"page_cache_delta_since_start_mib": -68.6,
"idle_watts": 7.7,
"avg_watts": 13.29,
"net_watts": 5.59,
"joules_per_tok": 8.42,
"net_joules_per_tok": 3.54,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " When streaming transformer layers on an 8 GB edge device, the interaction between page-aligned O_DIRECT NVMe reads and asynchronous ping-pong GPU buffers is cr"
},
{
"round": "round8_v010",
"source_file": "r8_14b_b5400_draft05b_temp07.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r8_14b_b5400_draft05b_temp07",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "prompt_heavy",
"chat_template_applied": false,
"warmup": false,
"reps": 3,
"temperature": 0.7,
"seed": 42,
"threads": 4,
"kv_cache_type": "q8_0",
"vram_budget_auto": false,
"vram_budget_mb": 5400,
"prompt_tokens": 340,
"completion_tokens": 32,
"load_time_s": 4.333,
"ttft_s": 8.337,
"ttft_s_min": 8.329,
"ttft_s_max": 8.556,
"decode_time_s": 95.405,
"total_time_s": 103.742,
"decode_tok_per_s": 0.325,
"decode_tok_per_s_min": 0.284,
"decode_tok_per_s_max": 0.454,
"total_tok_per_s": 0.308,
"total_tok_per_s_min": 0.273,
"total_tok_per_s_max": 0.417,
"sweeps": 15,
"drafted_tokens": 72,
"accepted_tokens": 18,
"resident_layers": 6,
"streamed_layers": 42,
"n_layers": 48,
"resident_mib": 2006.6,
"slot_mib": 166.4,
"expert_slot_mib": 0.0,
"io_mib_read": 98543.92,
"pread_mib_s": 1026.8,
"effective_io_mib_s": 948.4,
"expert_cache_hits": 0,
"expert_cache_misses": 0,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": 0.0,
"moe_layers": 0,
"moe_active_streaming": false,
"uma_zero_copy": false,
"is_unified_memory": true,
"direct_io_active": true,
"peak_rss_mib": 4023.3,
"current_rss_mib": 4023.3,
"mem_avail_before_mib": 2060.9,
"mem_avail_after_mib": 2060.2,
"swap_delta_mib": -0.3,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 0.3,
"page_cache_delta_since_start_mib": 22.2,
"idle_watts": 7.72,
"avg_watts": 10.03,
"net_watts": 2.31,
"joules_per_tok": 32.52,
"net_joules_per_tok": 7.47,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " Layer-streaming inference manages memory by allocating a small, fixed pool of compute buffers and asynchronously streaming transformer blocks from NVMe storage"
},
{
"round": "round8_v010",
"source_file": "r8_14b_b5400_draft05b_temp07.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r8_14b_b5400_draft05b_temp07",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "code_edit",
"chat_template_applied": false,
"warmup": false,
"reps": 3,
"temperature": 0.7,
"seed": 42,
"threads": 4,
"kv_cache_type": "q8_0",
"vram_budget_auto": false,
"vram_budget_mb": 5400,
"prompt_tokens": 106,
"completion_tokens": 48,
"load_time_s": 4.333,
"ttft_s": 7.122,
"ttft_s_min": 7.121,
"ttft_s_max": 7.142,
"decode_time_s": 34.313,
"total_time_s": 41.434,
"decode_tok_per_s": 1.37,
"decode_tok_per_s_min": 1.369,
"decode_tok_per_s_max": 1.372,
"total_tok_per_s": 1.158,
"total_tok_per_s_min": 1.158,
"total_tok_per_s_max": 1.16,
"sweeps": 6,
"drafted_tokens": 50,
"accepted_tokens": 43,
"resident_layers": 6,
"streamed_layers": 42,
"n_layers": 48,
"resident_mib": 2006.6,
"slot_mib": 166.4,
"expert_slot_mib": 0.0,
"io_mib_read": 39208.95,
"pread_mib_s": 1025.5,
"effective_io_mib_s": 946.3,
"expert_cache_hits": 0,
"expert_cache_misses": 0,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": 0.0,
"moe_layers": 0,
"moe_active_streaming": false,
"uma_zero_copy": false,
"is_unified_memory": true,
"direct_io_active": true,
"peak_rss_mib": 4028.0,
"current_rss_mib": 4028.0,
"mem_avail_before_mib": 2052.5,
"mem_avail_after_mib": 2056.1,
"swap_delta_mib": 0.0,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 0.1,
"page_cache_delta_since_start_mib": 22.9,
"idle_watts": 7.72,
"avg_watts": 9.92,
"net_watts": 2.2,
"joules_per_tok": 8.56,
"net_joules_per_tok": 1.9,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": "#[derive(Debug, Clone)]\npub struct StreamConfig {\npub model_path: PathBuf,\npub n_ctx: u32,\npub n_batch: u32,\npub resident_layers: Option<u32>,\npub vram_budget_b"
},
{
"round": "round8_v010",
"source_file": "r8_14b_b5400_draft05b_temp07.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r8_14b_b5400_draft05b_temp07",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "open_ended",
"chat_template_applied": false,
"warmup": false,
"reps": 3,
"temperature": 0.7,
"seed": 42,
"threads": 4,
"kv_cache_type": "q8_0",
"vram_budget_auto": false,
"vram_budget_mb": 5400,
"prompt_tokens": 31,
"completion_tokens": 48,
"load_time_s": 4.333,
"ttft_s": 6.816,
"ttft_s_min": 6.803,
"ttft_s_max": 6.841,
"decode_time_s": 102.588,
"total_time_s": 109.39,
"decode_tok_per_s": 0.458,
"decode_tok_per_s_min": 0.405,
"decode_tok_per_s_max": 0.491,
"total_tok_per_s": 0.439,
"total_tok_per_s_min": 0.391,
"total_tok_per_s_max": 0.468,
"sweeps": 15,
"drafted_tokens": 112,
"accepted_tokens": 34,
"resident_layers": 6,
"streamed_layers": 42,
"n_layers": 48,
"resident_mib": 2006.6,
"slot_mib": 166.4,
"expert_slot_mib": 0.0,
"io_mib_read": 98543.92,
"pread_mib_s": 1027.3,
"effective_io_mib_s": 961.5,
"expert_cache_hits": 0,
"expert_cache_misses": 0,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": 0.0,
"moe_layers": 0,
"moe_active_streaming": false,
"uma_zero_copy": false,
"is_unified_memory": true,
"direct_io_active": true,
"peak_rss_mib": 4030.0,
"current_rss_mib": 4030.0,
"mem_avail_before_mib": 2047.3,
"mem_avail_after_mib": 2046.9,
"swap_delta_mib": 0.0,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 1.2,
"page_cache_delta_since_start_mib": 25.8,
"idle_watts": 7.72,
"avg_watts": 9.77,
"net_watts": 2.05,
"joules_per_tok": 22.27,
"net_joules_per_tok": 4.67,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " In the context of running transformer layers on an 8 GB edge device, the interaction between page-aligned O_DIRECT NVMe reads and asynchronous ping-pong GPU bu"
},
{
"round": "round8_v010",
"source_file": "r8_14b_r0.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r8_14b_r0",
"status": "ok",
"model": "Qwen2.5-14B-Instruct-Q4_K_M.gguf",
"draft_model": null,
"workload": "open_ended",
"chat_template_applied": false,
"warmup": false,
"reps": 2,
"temperature": 0.0,
"seed": 42,
"threads": 4,
"kv_cache_type": "q8_0",
"vram_budget_auto": false,
"vram_budget_mb": 0,
"prompt_tokens": 31,
"completion_tokens": 16,
"load_time_s": 2.993,
"ttft_s": 5.966,
"ttft_s_min": 5.783,
"ttft_s_max": 6.149,
"decode_time_s": 86.265,
"total_time_s": 92.231,
"decode_tok_per_s": 0.174,
"decode_tok_per_s_min": 0.174,
"decode_tok_per_s_max": 0.174,
"total_tok_per_s": 0.173,
"total_tok_per_s_min": 0.173,
"total_tok_per_s_max": 0.174,
"sweeps": 16,
"drafted_tokens": 0,
"accepted_tokens": 0,
"resident_layers": 0,
"streamed_layers": 48,
"n_layers": 48,
"resident_mib": 1026.8,
"slot_mib": 166.4,
"expert_slot_mib": 0.0,
"io_mib_read": 120631.69,
"pread_mib_s": 1426.9,
"effective_io_mib_s": 1308.8,
"expert_cache_hits": 0,
"expert_cache_misses": 0,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": 0.0,
"moe_layers": 0,
"moe_active_streaming": false,
"uma_zero_copy": false,
"is_unified_memory": true,
"direct_io_active": true,
"peak_rss_mib": 2458.0,
"current_rss_mib": 2457.9,
"mem_avail_before_mib": 3656.6,
"mem_avail_after_mib": 3632.7,
"swap_delta_mib": -3.0,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 11.8,
"page_cache_delta_since_start_mib": -677.4,
"idle_watts": 7.71,
"avg_watts": 10.37,
"net_watts": 2.66,
"joules_per_tok": 59.75,
"net_joules_per_tok": 15.33,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " When streaming transformer layers on an 8 GB edge device, the interaction between page"
},
{
"round": "round8_v010",
"source_file": "r8_32b_b5400_draft05b.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r8_32b_b5400_draft05b",
"status": "ok",
"model": "Qwen2.5-32B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "prompt_heavy",
"chat_template_applied": false,
"warmup": false,
"reps": 2,
"temperature": 0.0,
"seed": 42,
"threads": 4,
"kv_cache_type": "q8_0",
"vram_budget_auto": false,
"vram_budget_mb": 5400,
"prompt_tokens": 340,
"completion_tokens": 32,
"load_time_s": 3.764,
"ttft_s": 18.01,
"ttft_s_min": 17.468,
"ttft_s_max": 18.552,
"decode_time_s": 183.023,
"total_time_s": 201.033,
"decode_tok_per_s": 0.171,
"decode_tok_per_s_min": 0.156,
"decode_tok_per_s_max": 0.185,
"total_tok_per_s": 0.16,
"total_tok_per_s_min": 0.148,
"total_tok_per_s_max": 0.172,
"sweeps": 13,
"drafted_tokens": 67,
"accepted_tokens": 19,
"resident_layers": 2,
"streamed_layers": 62,
"n_layers": 64,
"resident_mib": 1622.2,
"slot_mib": 297.7,
"expert_slot_mib": 0.0,
"io_mib_read": 224952.76,
"pread_mib_s": 1229.4,
"effective_io_mib_s": 1125.0,
"expert_cache_hits": 0,
"expert_cache_misses": 0,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": 0.0,
"moe_layers": 0,
"moe_active_streaming": false,
"uma_zero_copy": false,
"is_unified_memory": true,
"direct_io_active": true,
"peak_rss_mib": 4108.7,
"current_rss_mib": 4108.6,
"mem_avail_before_mib": 1961.5,
"mem_avail_after_mib": 1964.1,
"swap_delta_mib": 0.0,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -10.4,
"page_cache_delta_since_start_mib": -21.0,
"idle_watts": 7.94,
"avg_watts": 10.21,
"net_watts": 2.26,
"joules_per_tok": 64.13,
"net_joules_per_tok": 14.21,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " The core memory-management mechanism described involves layer-streaming inference, which uses a small fixed pool of reusable compute buffers and asynchronously"
},
{
"round": "round8_v010",
"source_file": "r8_32b_b5400_draft05b.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r8_32b_b5400_draft05b",
"status": "ok",
"model": "Qwen2.5-32B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "code_edit",
"chat_template_applied": false,
"warmup": false,
"reps": 2,
"temperature": 0.0,
"seed": 42,
"threads": 4,
"kv_cache_type": "q8_0",
"vram_budget_auto": false,
"vram_budget_mb": 5400,
"prompt_tokens": 106,
"completion_tokens": 48,
"load_time_s": 3.764,
"ttft_s": 14.704,
"ttft_s_min": 14.671,
"ttft_s_max": 14.738,
"decode_time_s": 55.967,
"total_time_s": 70.672,
"decode_tok_per_s": 0.84,
"decode_tok_per_s_min": 0.838,
"decode_tok_per_s_max": 0.841,
"total_tok_per_s": 0.679,
"total_tok_per_s_min": 0.678,
"total_tok_per_s_max": 0.681,
"sweeps": 5,
"drafted_tokens": 44,
"accepted_tokens": 44,
"resident_layers": 2,
"streamed_layers": 62,
"n_layers": 64,
"resident_mib": 1622.2,
"slot_mib": 297.7,
"expert_slot_mib": 0.0,
"io_mib_read": 86818.02,
"pread_mib_s": 1337.6,
"effective_io_mib_s": 1224.3,
"expert_cache_hits": 0,
"expert_cache_misses": 0,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": 0.0,
"moe_layers": 0,
"moe_active_streaming": false,
"uma_zero_copy": false,
"is_unified_memory": true,
"direct_io_active": true,
"peak_rss_mib": 4113.2,
"current_rss_mib": 4113.2,
"mem_avail_before_mib": 1958.5,
"mem_avail_after_mib": 1951.6,
"swap_delta_mib": -0.5,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -1.9,
"page_cache_delta_since_start_mib": -24.5,
"idle_watts": 7.94,
"avg_watts": 10.54,
"net_watts": 2.59,
"joules_per_tok": 15.51,
"net_joules_per_tok": 3.82,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": "#[derive(Debug, Clone)]\npub struct StreamConfig {\n pub model_path: PathBuf,\n pub n_ctx: u32,\n pub n_batch: u32,\n pub resident_layers: Option<u32>,\n "
},
{
"round": "round8_v010",
"source_file": "r8_32b_b5400_draft05b.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r8_32b_b5400_draft05b",
"status": "ok",
"model": "Qwen2.5-32B-Instruct-Q4_K_M.gguf",
"draft_model": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
"workload": "open_ended",
"chat_template_applied": false,
"warmup": false,
"reps": 2,
"temperature": 0.0,
"seed": 42,
"threads": 4,
"kv_cache_type": "q8_0",
"vram_budget_auto": false,
"vram_budget_mb": 5400,
"prompt_tokens": 31,
"completion_tokens": 48,
"load_time_s": 3.764,
"ttft_s": 16.558,
"ttft_s_min": 14.005,
"ttft_s_max": 19.111,
"decode_time_s": 154.474,
"total_time_s": 171.031,
"decode_tok_per_s": 0.304,
"decode_tok_per_s_min": 0.304,
"decode_tok_per_s_max": 0.304,
"total_tok_per_s": 0.281,
"total_tok_per_s_min": 0.277,
"total_tok_per_s_max": 0.285,
"sweeps": 12,
"drafted_tokens": 104,
"accepted_tokens": 37,
"resident_layers": 2,
"streamed_layers": 62,
"n_layers": 64,
"resident_mib": 1622.2,
"slot_mib": 297.7,
"expert_slot_mib": 0.0,
"io_mib_read": 207648.7,
"pread_mib_s": 1311.4,
"effective_io_mib_s": 1213.5,
"expert_cache_hits": 0,
"expert_cache_misses": 0,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": 0.0,
"moe_layers": 0,
"moe_active_streaming": false,
"uma_zero_copy": false,
"is_unified_memory": true,
"direct_io_active": true,
"peak_rss_mib": 4115.2,
"current_rss_mib": 4115.2,
"mem_avail_before_mib": 1930.7,
"mem_avail_after_mib": 1955.5,
"swap_delta_mib": 0.0,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -12.6,
"page_cache_delta_since_start_mib": -37.9,
"idle_watts": 7.94,
"avg_watts": 10.27,
"net_watts": 2.33,
"joules_per_tok": 36.61,
"net_joules_per_tok": 8.3,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": " When streaming transformer layers on an 8 GB edge device, the interaction between page-aligned O_DIRECT NVMe reads and asynchronous ping-pong GPU buffers can b"
},
{
"round": "round8_v010",
"source_file": "r8_olmoe_r0_cached.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r8_olmoe_r0_sparse_cached",
"status": "ok",
"model": "OLMoE-1B-7B-0125-Instruct-Q2_K.gguf",
"draft_model": null,
"workload": "prompt_heavy",
"chat_template_applied": false,
"warmup": true,
"reps": 3,
"temperature": 0.0,
"seed": 42,
"threads": 4,
"kv_cache_type": "q8_0",
"vram_budget_auto": false,
"vram_budget_mb": 0,
"prompt_tokens": 367,
"completion_tokens": 32,
"load_time_s": 0.681,
"ttft_s": 2.078,
"ttft_s_min": 2.078,
"ttft_s_max": 2.08,
"decode_time_s": 2.268,
"total_time_s": 4.346,
"decode_tok_per_s": 13.667,
"decode_tok_per_s_min": 13.652,
"decode_tok_per_s_max": 13.922,
"total_tok_per_s": 7.362,
"total_tok_per_s_min": 7.356,
"total_tok_per_s_max": 7.433,
"sweeps": 32,
"drafted_tokens": 0,
"accepted_tokens": 0,
"resident_layers": 0,
"streamed_layers": 16,
"n_layers": 16,
"resident_mib": 218.3,
"slot_mib": 0.0,
"expert_slot_mib": 139.0,
"io_mib_read": 2085.12,
"pread_mib_s": 1404.5,
"effective_io_mib_s": 479.7,
"expert_cache_hits": 11928,
"expert_cache_misses": 0,
"expert_cache_hit_rate": 1.0,
"expert_cache_mib_saved": 8635.38,
"moe_layers": 16,
"moe_active_streaming": true,
"uma_zero_copy": false,
"is_unified_memory": true,
"direct_io_active": true,
"peak_rss_mib": 2921.4,
"current_rss_mib": 2921.4,
"mem_avail_before_mib": 2467.0,
"mem_avail_after_mib": 2465.2,
"swap_delta_mib": 0.0,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 0.4,
"page_cache_delta_since_start_mib": -90.3,
"idle_watts": 8.13,
"avg_watts": 12.32,
"net_watts": 4.2,
"joules_per_tok": 1.67,
"net_joules_per_tok": 0.57,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": "\nThe core memory-management mechanism involves a combination of zero-copy staging from NVMe with page-aligned DMA, layer-streaming inference with reusable"
},
{
"round": "round8_v010",
"source_file": "r8_olmoe_r0_cached.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r8_olmoe_r0_sparse_cached",
"status": "ok",
"model": "OLMoE-1B-7B-0125-Instruct-Q2_K.gguf",
"draft_model": null,
"workload": "code_edit",
"chat_template_applied": false,
"warmup": true,
"reps": 3,
"temperature": 0.0,
"seed": 42,
"threads": 4,
"kv_cache_type": "q8_0",
"vram_budget_auto": false,
"vram_budget_mb": 0,
"prompt_tokens": 133,
"completion_tokens": 48,
"load_time_s": 0.681,
"ttft_s": 1.865,
"ttft_s_min": 1.863,
"ttft_s_max": 1.868,
"decode_time_s": 3.601,
"total_time_s": 5.465,
"decode_tok_per_s": 13.054,
"decode_tok_per_s_min": 12.451,
"decode_tok_per_s_max": 13.24,
"total_tok_per_s": 8.783,
"total_tok_per_s_min": 8.513,
"total_tok_per_s_max": 8.86,
"sweeps": 48,
"drafted_tokens": 0,
"accepted_tokens": 0,
"resident_layers": 0,
"streamed_layers": 16,
"n_layers": 16,
"resident_mib": 218.3,
"slot_mib": 0.0,
"expert_slot_mib": 139.0,
"io_mib_read": 2180.74,
"pread_mib_s": 1269.1,
"effective_io_mib_s": 409.4,
"expert_cache_hits": 17941,
"expert_cache_misses": 131,
"expert_cache_hit_rate": 0.993,
"expert_cache_mib_saved": 12988.27,
"moe_layers": 16,
"moe_active_streaming": true,
"uma_zero_copy": false,
"is_unified_memory": true,
"direct_io_active": true,
"peak_rss_mib": 2953.3,
"current_rss_mib": 2953.3,
"mem_avail_before_mib": 2438.6,
"mem_avail_after_mib": 2437.3,
"swap_delta_mib": 0.0,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 0.1,
"page_cache_delta_since_start_mib": -103.2,
"idle_watts": 8.13,
"avg_watts": 12.61,
"net_watts": 4.49,
"joules_per_tok": 1.44,
"net_joules_per_tok": 0.51,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": "#[derive(Debug, Clone)]\npub struct StreamConfig {\npub model_path: PathBuf,\npub n_ctx: u32,\npub n_batch: u32,\npub resident_layers"
},
{
"round": "round8_v010",
"source_file": "r8_olmoe_r0_cached.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r8_olmoe_r0_sparse_cached",
"status": "ok",
"model": "OLMoE-1B-7B-0125-Instruct-Q2_K.gguf",
"draft_model": null,
"workload": "open_ended",
"chat_template_applied": false,
"warmup": true,
"reps": 3,
"temperature": 0.0,
"seed": 42,
"threads": 4,
"kv_cache_type": "q8_0",
"vram_budget_auto": false,
"vram_budget_mb": 0,
"prompt_tokens": 33,
"completion_tokens": 48,
"load_time_s": 0.681,
"ttft_s": 1.72,
"ttft_s_min": 1.718,
"ttft_s_max": 1.724,
"decode_time_s": 3.901,
"total_time_s": 5.622,
"decode_tok_per_s": 12.047,
"decode_tok_per_s_min": 11.45,
"decode_tok_per_s_max": 12.505,
"total_tok_per_s": 8.538,
"total_tok_per_s_min": 8.235,
"total_tok_per_s_max": 8.764,
"sweeps": 48,
"drafted_tokens": 0,
"accepted_tokens": 0,
"resident_layers": 0,
"streamed_layers": 16,
"n_layers": 16,
"resident_mib": 218.3,
"slot_mib": 0.0,
"expert_slot_mib": 139.0,
"io_mib_read": 2353.59,
"pread_mib_s": 1167.9,
"effective_io_mib_s": 436.5,
"expert_cache_hits": 17705,
"expert_cache_misses": 367,
"expert_cache_hit_rate": 0.98,
"expert_cache_mib_saved": 12816.33,
"moe_layers": 16,
"moe_active_streaming": true,
"uma_zero_copy": false,
"is_unified_memory": true,
"direct_io_active": true,
"peak_rss_mib": 2970.7,
"current_rss_mib": 2970.7,
"mem_avail_before_mib": 2415.1,
"mem_avail_after_mib": 2413.6,
"swap_delta_mib": 0.0,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 0.0,
"page_cache_delta_since_start_mib": -114.5,
"idle_watts": 8.13,
"avg_watts": 11.98,
"net_watts": 3.86,
"joules_per_tok": 1.4,
"net_joules_per_tok": 0.45,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": "\n\n1. **Page-Aligned O_DIRECT NVMe Reads:**\n - **Explanation:** When performing an O_DIRECT read on an NVMe storage device, the read operation is aligned with "
},
{
"round": "round8_v010",
"source_file": "r8_olmoe_r0_full.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r8_olmoe_r0_full",
"status": "ok",
"model": "OLMoE-1B-7B-0125-Instruct-Q2_K.gguf",
"draft_model": null,
"workload": "open_ended",
"chat_template_applied": false,
"warmup": false,
"reps": 2,
"temperature": 0.0,
"seed": 42,
"threads": 4,
"kv_cache_type": "q8_0",
"vram_budget_auto": false,
"vram_budget_mb": 0,
"prompt_tokens": 33,
"completion_tokens": 32,
"load_time_s": 0.507,
"ttft_s": 1.448,
"ttft_s_min": 1.355,
"ttft_s_max": 1.54,
"decode_time_s": 40.615,
"total_time_s": 42.062,
"decode_tok_per_s": 0.763,
"decode_tok_per_s_min": 0.763,
"decode_tok_per_s_max": 0.763,
"total_tok_per_s": 0.761,
"total_tok_per_s_min": 0.759,
"total_tok_per_s_max": 0.762,
"sweeps": 32,
"drafted_tokens": 0,
"accepted_tokens": 0,
"resident_layers": 0,
"streamed_layers": 16,
"n_layers": 16,
"resident_mib": 112.8,
"slot_mib": 145.6,
"expert_slot_mib": 0.0,
"io_mib_read": 74545.94,
"pread_mib_s": 1942.9,
"effective_io_mib_s": 1774.0,
"expert_cache_hits": 0,
"expert_cache_misses": 0,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": 0.0,
"moe_layers": 0,
"moe_active_streaming": false,
"uma_zero_copy": false,
"is_unified_memory": true,
"direct_io_active": true,
"peak_rss_mib": 1312.2,
"current_rss_mib": 1312.2,
"mem_avail_before_mib": 4053.9,
"mem_avail_after_mib": 4054.1,
"swap_delta_mib": -2.0,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": 1.0,
"page_cache_delta_since_start_mib": 1.0,
"idle_watts": 7.86,
"avg_watts": 10.91,
"net_watts": 3.05,
"joules_per_tok": 14.34,
"net_joules_per_tok": 4.01,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": "\n\n1. **Page-Aligned O_DIRECT NVMe Reads:**\n - **Explanation:** When performing an O_DIRECT read on"
},
{
"round": "round8_v010",
"source_file": "r8_qwen3_30b_a3b_draft06b_code.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r8_qwen3_30b_a3b_r0_draft06b",
"status": "ok",
"model": "Qwen3-30B-A3B-Q4_K_M.gguf",
"draft_model": "Qwen3-0.6B-Q4_K_M.gguf",
"workload": "code_edit",
"chat_template_applied": true,
"warmup": false,
"reps": 2,
"temperature": 0.0,
"seed": 42,
"threads": 4,
"kv_cache_type": "q8_0",
"vram_budget_auto": false,
"vram_budget_mb": 0,
"prompt_tokens": 114,
"completion_tokens": 32,
"load_time_s": 33.063,
"ttft_s": 13.596,
"ttft_s_min": 13.512,
"ttft_s_max": 13.681,
"decode_time_s": 331.875,
"total_time_s": 345.471,
"decode_tok_per_s": 0.113,
"decode_tok_per_s_min": 0.066,
"decode_tok_per_s_max": 0.159,
"total_tok_per_s": 0.11,
"total_tok_per_s_min": 0.066,
"total_tok_per_s_max": 0.154,
"sweeps": 11,
"drafted_tokens": 62,
"accepted_tokens": 22,
"resident_layers": 0,
"streamed_layers": 48,
"n_layers": 48,
"resident_mib": 951.3,
"slot_mib": 0.0,
"expert_slot_mib": 373.5,
"io_mib_read": 74126.45,
"pread_mib_s": 267.3,
"effective_io_mib_s": 257.1,
"expert_cache_hits": 4218,
"expert_cache_misses": 22119,
"expert_cache_hit_rate": 0.16,
"expert_cache_mib_saved": 3767.38,
"moe_layers": 48,
"moe_active_streaming": true,
"uma_zero_copy": false,
"is_unified_memory": true,
"direct_io_active": true,
"peak_rss_mib": 3974.3,
"current_rss_mib": 3973.1,
"mem_avail_before_mib": 1549.3,
"mem_avail_after_mib": 1539.5,
"swap_delta_mib": -2.0,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -667.3,
"page_cache_delta_since_start_mib": -1066.2,
"idle_watts": 7.85,
"avg_watts": 8.33,
"net_watts": 0.48,
"joules_per_tok": 89.98,
"net_joules_per_tok": 5.21,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": "<think>\nOkay, let me try to figure out how to do this. The user wants me to take the given Rust struct and add a new field called `"
},
{
"round": "round8_v010",
"source_file": "r8_qwen3_30b_a3b_draft06b_open.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "layerstream-plugin",
"engine_version": "0.1.0",
"git_rev": null,
"mode": "r8_qwen3_30b_a3b_r0_draft06b",
"status": "ok",
"model": "Qwen3-30B-A3B-Q4_K_M.gguf",
"draft_model": "Qwen3-0.6B-Q4_K_M.gguf",
"workload": "open_ended",
"chat_template_applied": true,
"warmup": false,
"reps": 2,
"temperature": 0.0,
"seed": 42,
"threads": 4,
"kv_cache_type": "q8_0",
"vram_budget_auto": false,
"vram_budget_mb": 0,
"prompt_tokens": 39,
"completion_tokens": 32,
"load_time_s": 3.263,
"ttft_s": 11.238,
"ttft_s_min": 11.08,
"ttft_s_max": 11.395,
"decode_time_s": 105.204,
"total_time_s": 116.442,
"decode_tok_per_s": 0.317,
"decode_tok_per_s_min": 0.233,
"decode_tok_per_s_max": 0.4,
"total_tok_per_s": 0.291,
"total_tok_per_s_min": 0.222,
"total_tok_per_s_max": 0.361,
"sweeps": 6,
"drafted_tokens": 48,
"accepted_tokens": 27,
"resident_layers": 0,
"streamed_layers": 48,
"n_layers": 48,
"resident_mib": 951.3,
"slot_mib": 0.0,
"expert_slot_mib": 373.5,
"io_mib_read": 73026.22,
"pread_mib_s": 708.0,
"effective_io_mib_s": 667.7,
"expert_cache_hits": 1826,
"expert_cache_misses": 6019,
"expert_cache_hit_rate": 0.233,
"expert_cache_mib_saved": 1555.77,
"moe_layers": 48,
"moe_active_streaming": true,
"uma_zero_copy": false,
"is_unified_memory": true,
"direct_io_active": true,
"peak_rss_mib": 3969.2,
"current_rss_mib": 3969.2,
"mem_avail_before_mib": 1661.1,
"mem_avail_after_mib": 1629.2,
"swap_delta_mib": 0.0,
"cuda_max_allocated_mib": null,
"page_cache_delta_mib": -24.4,
"page_cache_delta_since_start_mib": -236.8,
"idle_watts": 8.18,
"avg_watts": 9.06,
"net_watts": 0.88,
"joules_per_tok": 32.98,
"net_joules_per_tok": 3.21,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": "<think>\nOkay, so I need to explain how page-aligned O_DIRECT NVMe reads and asynchronous ping-pong GPU buffers interact when streaming transformer layers on an "
}
]
[
{
"round": "airllm",
"source_file": "airllm_14b_4bit_code_edit_warm.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "airllm",
"engine_version": "4.0.0",
"git_rev": null,
"mode": "airllm_4bit_prefetch",
"status": "ok",
"model": "Qwen/Qwen2.5-14B-Instruct",
"draft_model": null,
"workload": "code_edit",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,
"vram_budget_mb": null,
"prompt_tokens": 113,
"completion_tokens": 8,
"load_time_s": 2.55,
"ttft_s": 33.371,
"ttft_s_min": null,
"ttft_s_max": null,
"decode_time_s": 214.529,
"total_time_s": 247.9,
"decode_tok_per_s": 0.033,
"decode_tok_per_s_min": null,
"decode_tok_per_s_max": null,
"total_tok_per_s": 0.032,
"total_tok_per_s_min": null,
"total_tok_per_s_max": null,
"sweeps": null,
"drafted_tokens": null,
"accepted_tokens": null,
"resident_layers": null,
"streamed_layers": null,
"n_layers": null,
"resident_mib": null,
"slot_mib": null,
"expert_slot_mib": null,
"io_mib_read": null,
"pread_mib_s": null,
"effective_io_mib_s": null,
"expert_cache_hits": null,
"expert_cache_misses": null,
"expert_cache_hit_rate": null,
"expert_cache_mib_saved": null,
"moe_layers": null,
"moe_active_streaming": null,
"uma_zero_copy": null,
"is_unified_memory": null,
"direct_io_active": null,
"peak_rss_mib": 3536.4,
"current_rss_mib": 626.8,
"mem_avail_before_mib": null,
"mem_avail_after_mib": null,
"swap_delta_mib": null,
"cuda_max_allocated_mib": 1936.2,
"page_cache_delta_mib": -503.0,
"page_cache_delta_since_start_mib": -458.6,
"idle_watts": null,
"avg_watts": 9.3,
"net_watts": null,
"joules_per_tok": 288.21,
"net_joules_per_tok": null,
"turn1_prompt_tokens": null,
"turn1_ttft_s": null,
"turn2_prompt_tokens": null,
"turn2_ttft_s": null,
"ttft_speedup_x": null,
"notes": null,
"output_preview": "#[derive(Debug, Clone)]\npub struct"
},
{
"round": "airllm",
"source_file": "airllm_14b_4bit_open_ended_warm.json",
"hardware": "NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)",
"hardware_id": "orin-nano-8gb-cuda",
"engine": "airllm",
"engine_version": "4.0.0",
"git_rev": null,
"mode": "airllm_4bit_prefetch",
"status": "ok",
"model": "Qwen/Qwen2.5-14B-Instruct",
"draft_model": null,
"workload": "open_ended",
"chat_template_applied": null,
"warmup": null,
"reps": null,
"temperature": null,
"seed": null,
"threads": null,
"kv_cache_type": null,
"vram_budget_auto": null,