Dataset Viewer
Auto-converted to Parquet Duplicate
round
string
source_file
string
hardware
string
hardware_id
string
engine
string
engine_version
string
git_rev
null
mode
string
status
string
model
string
draft_model
string
workload
string
chat_template_applied
bool
warmup
bool
reps
int64
temperature
float64
seed
int64
threads
int64
kv_cache_type
string
vram_budget_auto
bool
vram_budget_mb
int64
prompt_tokens
int64
completion_tokens
int64
load_time_s
float64
ttft_s
float64
ttft_s_min
float64
ttft_s_max
float64
decode_time_s
float64
total_time_s
float64
decode_tok_per_s
float64
decode_tok_per_s_min
float64
decode_tok_per_s_max
float64
total_tok_per_s
float64
total_tok_per_s_min
float64
total_tok_per_s_max
float64
sweeps
int64
drafted_tokens
int64
accepted_tokens
int64
resident_layers
int64
streamed_layers
int64
n_layers
int64
resident_mib
float64
slot_mib
float64
expert_slot_mib
float64
io_mib_read
float64
pread_mib_s
float64
effective_io_mib_s
float64
expert_cache_hits
int64
expert_cache_misses
int64
expert_cache_hit_rate
float64
expert_cache_mib_saved
float64
moe_layers
int64
moe_active_streaming
bool
uma_zero_copy
bool
is_unified_memory
bool
direct_io_active
bool
peak_rss_mib
float64
current_rss_mib
float64
mem_avail_before_mib
float64
mem_avail_after_mib
float64
swap_delta_mib
float64
cuda_max_allocated_mib
float64
page_cache_delta_mib
float64
page_cache_delta_since_start_mib
float64
idle_watts
float64
avg_watts
float64
net_watts
float64
joules_per_tok
float64
net_joules_per_tok
float64
turn1_prompt_tokens
int64
turn1_ttft_s
float64
turn2_prompt_tokens
int64
turn2_ttft_s
float64
ttft_speedup_x
float64
notes
string
output_preview
string
airllm
airllm_14b_4bit_code_edit_warm.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
airllm
4.0.0
null
airllm_4bit_prefetch
ok
Qwen/Qwen2.5-14B-Instruct
null
code_edit
null
null
null
null
null
null
null
null
null
113
8
2.55
33.371
null
null
214.529
247.9
0.033
null
null
0.032
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
3,536.4
626.8
null
null
null
1,936.2
-503
-458.6
null
9.3
null
288.21
null
null
null
null
null
null
null
#[derive(Debug, Clone)] pub struct
airllm
airllm_14b_4bit_open_ended_warm.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
airllm
4.0.0
null
airllm_4bit_prefetch
ok
Qwen/Qwen2.5-14B-Instruct
null
open_ended
null
null
null
null
null
null
null
null
null
31
8
2.573
32.863
null
null
213.357
246.219
0.033
null
null
0.032
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
3,531.2
617
null
null
null
1,920.8
-640.9
-591.8
null
9.33
null
287.1
null
null
null
null
null
null
null
When discussing the interaction between page-aligned O
airllm
airllm_14b_4bit_prompt_heavy.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
airllm
4.0.0
null
airllm_4bit_prefetch
ok
Qwen/Qwen2.5-14B-Instruct
null
prompt_heavy
null
null
null
null
null
null
null
null
null
340
32
365.529
32.093
null
null
954.918
987.011
0.032
null
null
0.032
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
5,390.5
689.2
null
null
null
1,983.3
248
-751.8
null
null
null
null
null
null
null
null
null
null
null
The core memory-management mechanism described is layer-streaming inference, which minimizes memory usage by allocating a small fixed pool of reusable buffers
airllm
airllm_14b_bf16_open_ended.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
airllm
4.0.0
null
airllm_none_prefetch
ok
Qwen/Qwen2.5-14B-Instruct
null
open_ended
null
null
null
null
null
null
null
null
null
31
4
81.81
45.413
null
null
131.774
177.187
0.023
null
null
0.023
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
5,217.6
3,489.7
null
null
null
1,502
-4,310.7
-3,140.9
null
9.31
null
412.35
null
null
null
null
null
null
null
In the context of
competitors
hf_accelerate_14b_4bit.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
hf-accelerate
transformers-5.23.0+bnb-0.50.2
null
hf_accelerate_14b_4bit_disk
timeout_300s
Qwen2.5-14B-Instruct
null
open_ended
null
null
null
null
null
null
null
null
null
null
0
144
null
null
null
null
300
0
null
null
0
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
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).
null
competitors
llama_cpp_14b_cpu_mmap.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
llama.cpp
0.1.0
null
llama_cpp_14b_cpu_mmap
timeout
Qwen2.5-14B-Instruct-Q4_K_M.gguf
null
open_ended
null
null
null
null
null
null
q8_0
null
null
null
0
null
null
null
null
300.458
300.458
0
null
null
0
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
6,109.6
null
5,040.2
null
362.6
null
5,648.3
null
7.87
10.61
null
null
null
null
null
null
null
null
null
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
competitors
llama_cpp_14b_ngl0.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
llama.cpp
0.1.0
null
llama_cpp_14b_ngl0
error_rc_1
Qwen2.5-14B-Instruct-Q4_K_M.gguf
null
open_ended
null
null
null
null
null
null
q8_0
null
null
null
0
null
null
null
null
22.832
22.832
0
null
null
0
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
5,991
null
4,984.7
null
3,683.5
null
135.7
null
7.79
9.9
null
null
null
null
null
null
null
null
null
null
competitors
llama_cpp_14b_ngl12_b512.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
llama.cpp
0.1.0
null
llama_cpp_14b_ngl12_b512
error_rc_1
Qwen2.5-14B-Instruct-Q4_K_M.gguf
null
prompt_heavy
null
null
null
null
null
null
q8_0
null
null
null
0
null
null
null
null
23.562
23.562
0
null
null
0
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
5,977.9
null
5,021
null
3,676.6
null
76.5
null
7.97
9.84
null
null
null
null
null
null
null
null
null
null
competitors
llama_cpp_14b_ngl12_b512.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
llama.cpp
0.1.0
null
llama_cpp_14b_ngl12_b512
error_rc_1
Qwen2.5-14B-Instruct-Q4_K_M.gguf
null
code_edit
null
null
null
null
null
null
q8_0
null
null
null
0
null
null
null
null
23.752
23.752
0
null
null
0
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
5,742.9
null
5,163.8
null
3,684.2
null
-5.9
null
7.97
9.89
null
null
null
null
null
null
null
null
null
null
competitors
llama_cpp_14b_ngl12_b512.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
llama.cpp
0.1.0
null
llama_cpp_14b_ngl12_b512
error_rc_1
Qwen2.5-14B-Instruct-Q4_K_M.gguf
null
open_ended
null
null
null
null
null
null
q8_0
null
null
null
0
null
null
null
null
23.445
23.445
0
null
null
0
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
5,723.5
null
5,125.5
null
3,710.8
null
53.4
null
7.97
9.95
null
null
null
null
null
null
null
null
null
null
competitors
llama_cpp_14b_ngl18.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
llama.cpp
0.1.0
null
llama_cpp_14b_ngl18
error_rc_1
Qwen2.5-14B-Instruct-Q4_K_M.gguf
null
prompt_heavy
null
null
null
null
null
null
q8_0
null
null
null
0
null
null
null
null
56.129
56.129
0
null
null
0
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
5,776.3
null
4,972.8
null
3,634.6
null
-2,317.2
null
7.92
9.23
null
null
null
null
null
null
null
null
null
null
competitors
llama_cpp_14b_ngl18.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
llama.cpp
0.1.0
null
llama_cpp_14b_ngl18
error_rc_1
Qwen2.5-14B-Instruct-Q4_K_M.gguf
null
code_edit
null
null
null
null
null
null
q8_0
null
null
null
0
null
null
null
null
22.412
22.412
0
null
null
0
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
5,736.2
null
5,027.8
null
3,626
null
114.3
null
7.92
9.97
null
null
null
null
null
null
null
null
null
null
competitors
llama_cpp_14b_ngl18.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
llama.cpp
0.1.0
null
llama_cpp_14b_ngl18
error_rc_1
Qwen2.5-14B-Instruct-Q4_K_M.gguf
null
open_ended
null
null
null
null
null
null
q8_0
null
null
null
0
null
null
null
null
54.08
54.08
0
null
null
0
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
5,715.3
null
5,078.5
null
3,556.4
null
-86.6
null
7.92
8.73
null
null
null
null
null
null
null
null
null
null
competitors
llama_cpp_32b_ngl10.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
llama.cpp
0.1.0
null
llama_cpp_32b_ngl10
timeout
Qwen2.5-32B-Instruct-Q4_K_M.gguf
null
open_ended
null
null
null
null
null
null
q8_0
null
null
null
0
null
null
null
null
302.828
302.828
0
null
null
0
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
5,714.7
null
5,176
null
17,242.8
null
-305.4
null
7.89
10.29
null
null
null
null
null
null
null
null
null
When streaming transformer layers
competitors
llama_cpp_qwen3_30b_cpu_mmap.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
llama.cpp
0.1.0
null
llama_cpp_qwen3_30b_cpu_mmap
error_rc_1
Qwen3-30B-A3B-Q4_K_M.gguf
null
open_ended
null
null
null
null
null
null
q8_0
null
null
null
0
null
null
null
null
66.223
66.223
0
null
null
0
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
6,044.1
null
5,115.9
null
12,324.5
null
-62.8
null
7.92
10.16
null
null
null
null
null
null
null
null
null
null
competitors
llama_cpp_qwen3_30b_cpu_mmap_b512.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
llama.cpp
0.1.0
null
llama_cpp_qwen3_30b_cpu_mmap_b512
ok
Qwen3-30B-A3B-Q4_K_M.gguf
null
open_ended
null
null
null
null
null
null
q8_0
null
null
31
23
25.34
17.367
null
null
18.113
62.445
1.27
null
null
0.648
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
6,070.9
null
6,137.6
null
247.9
null
-100.1
null
8.07
9.89
null
15.26
null
null
null
null
null
null
null
the problem is that the GPU is not able to read the data from the host memory as it is being written by the
competitors
llama_cpp_qwen3_30b_gpu_offload.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
llama.cpp
0.1.0
null
llama_cpp_qwen3_30b_gpu_moe_offload
error_rc_1
Qwen3-30B-A3B-Q4_K_M.gguf
null
open_ended
null
null
null
null
null
null
q8_0
null
null
null
0
null
null
null
null
54.427
54.427
0
null
null
0
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
6,014
null
6,081.8
null
12,342.2
null
81.1
null
7.95
10.12
null
null
null
null
null
null
null
null
null
null
results
r0_prompt_heavy.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
stream_r0
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
null
prompt_heavy
null
null
null
null
null
null
null
null
null
340
32
2.968
9.257
null
null
228.946
238.203
0.135
null
null
0.134
null
null
32
0
0
0
48
48
1,026.8
166.4
null
241,263.38
null
null
null
null
null
null
null
null
false
null
true
3,225.7
3,225.7
null
null
null
null
37.9
-764.6
null
null
null
null
null
null
null
null
null
null
null
Layer-streaming inference manages memory by allocating a small, fixed pool of buffers for compute, keeping essential layers in memory, and asynchronously strea
results
r14_draft_prompt_heavy.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
hybrid_r14_draft05b
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
qwen2.5-0.5b-instruct-q4_k_m.gguf
prompt_heavy
null
null
null
null
null
null
null
null
null
340
32
52.833
5.752
null
null
39.649
45.401
0.782
null
null
0.705
null
null
15
63
17
14
34
48
3,244.4
166.4
null
79,826.54
null
null
null
null
null
null
null
null
false
null
true
5,877.1
5,666
null
null
null
null
-251.4
-2,667.7
null
null
null
null
null
null
null
null
null
null
null
Layer-streaming inference manages memory by allocating a small, fixed pool of buffers for active computation, keeping only essential layers in memory, and asyn
round2
hybrid_r14_draft05b.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
hybrid_r14_draft05b
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
qwen2.5-0.5b-instruct-q4_k_m.gguf
prompt_heavy
null
null
null
null
null
null
null
null
null
340
32
53.386
5.665
null
null
41.96
47.624
0.739
null
null
0.672
null
null
15
63
17
14
34
48
3,244.4
166.4
null
79,826.54
null
null
null
null
null
null
null
null
false
null
true
6,257.4
5,668.1
null
null
null
null
-198.3
-2,235.9
null
null
null
null
null
null
null
null
null
null
null
Layer-streaming inference manages memory by allocating a small, fixed pool of buffers for active computation, keeping only essential layers in memory, and asyn
round2
hybrid_r14_draft05b.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
hybrid_r14_draft05b
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
qwen2.5-0.5b-instruct-q4_k_m.gguf
code_edit
null
null
null
null
null
null
null
null
null
106
48
53.386
3.438
null
null
24.011
27.449
1.957
null
null
1.749
null
null
9
40
39
14
34
48
3,244.4
166.4
null
47,895.93
null
null
null
null
null
null
null
null
false
null
true
6,257.4
5,669.1
null
null
null
null
-4.3
-2,240.2
null
null
null
null
null
null
null
null
null
null
null
#[derive(Debug, Clone)] pub struct StreamConfig { pub model_path: PathBuf, pub n_ctx: u32, pub n_batch: u32, pub resident_layers: Option<u32>, pub vram_budget_b
round2
hybrid_r14_draft05b.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
hybrid_r14_draft05b
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
qwen2.5-0.5b-instruct-q4_k_m.gguf
open_ended
null
null
null
null
null
null
null
null
null
31
48
53.386
3.155
null
null
36.017
39.172
1.305
null
null
1.225
null
null
13
59
32
14
34
48
3,244.4
166.4
null
69,183
null
null
null
null
null
null
null
null
false
null
true
6,257.4
5,669.4
null
null
null
null
-14.6
-2,254.7
null
null
null
null
null
null
null
null
null
null
null
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
round2
hybrid_r16_ngram.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
hybrid_r16_ngram
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
null
prompt_heavy
null
null
null
null
null
null
null
null
null
340
32
75.668
5.074
null
null
63.663
68.736
0.487
null
null
0.466
null
null
25
26
7
16
32
48
3,558.5
166.4
null
125,190.63
null
null
null
null
null
null
null
null
false
null
true
5,982.3
5,152
null
null
null
null
57.9
-181
null
null
null
null
null
null
null
null
null
null
null
Layer-streaming inference manages memory by allocating a small, fixed pool of buffers for compute, keeping essential layers in memory, and asynchronously strea
round2
hybrid_r16_ngram.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
hybrid_r16_ngram
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
null
code_edit
null
null
null
null
null
null
null
null
null
106
48
75.668
3.087
null
null
23.822
26.909
1.973
null
null
1.784
null
null
10
45
38
16
32
48
3,558.5
166.4
null
50,076.25
null
null
null
null
null
null
null
null
false
null
true
5,982.3
5,153
null
null
null
null
3.7
-177.3
null
null
null
null
null
null
null
null
null
null
null
#[derive(Debug, Clone)] pub struct StreamConfig { pub model_path: PathBuf, pub n_ctx: u32, pub n_batch: u32, pub resident_layers: Option<u32>, pub vram_budget_b
round2
hybrid_r16_ngram.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
hybrid_r16_ngram
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
null
open_ended
null
null
null
null
null
null
null
null
null
31
48
75.668
2.75
null
null
79.476
82.225
0.591
null
null
0.584
null
null
31
25
17
16
32
48
3,558.5
166.4
null
155,236.38
null
null
null
null
null
null
null
null
false
null
true
5,982.3
5,155.4
null
null
null
null
2.4
-174.9
null
null
null
null
null
null
null
null
null
null
null
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
round2
stream_r0.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
stream_r0
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
null
prompt_heavy
null
null
null
null
null
null
null
null
null
340
32
2.4
9.753
null
null
247.344
257.097
0.125
null
null
0.124
null
null
32
0
0
0
48
48
1,026.8
166.4
null
241,263.38
null
null
null
null
null
null
null
null
false
null
true
3,249.2
3,249.2
null
null
null
null
2
2.3
null
null
null
null
null
null
null
null
null
null
null
Layer-streaming inference manages memory by allocating a small, fixed pool of buffers for compute, keeping essential layers in memory, and asynchronously strea
round2
stream_r0.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
stream_r0
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
null
code_edit
null
null
null
null
null
null
null
null
null
106
48
2.4
8.395
null
null
380.266
388.66
0.124
null
null
0.124
null
null
48
0
0
0
48
48
1,026.8
166.4
null
361,895.06
null
null
null
null
null
null
null
null
false
null
true
3,249.2
3,249.2
null
null
null
null
2.1
4.4
null
null
null
null
null
null
null
null
null
null
null
#[derive(Debug, Clone)] pub struct StreamConfig { pub model_path: PathBuf, pub n_ctx: u32, pub n_batch: u32, pub resident_layers: Option<u32>, pub vram_budget_b
round2
stream_r0.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
stream_r0
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
null
open_ended
null
null
null
null
null
null
null
null
null
31
48
2.4
8.088
null
null
385.094
393.183
0.122
null
null
0.122
null
null
48
0
0
0
48
48
1,026.8
166.4
null
361,895.06
null
null
null
null
null
null
null
null
false
null
true
3,249.2
3,249.2
null
null
null
null
9.7
14.1
null
null
null
null
null
null
null
null
null
null
null
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
round2
stream_r0_ngram.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
stream_r0_ngram
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
null
prompt_heavy
null
null
null
null
null
null
null
null
null
340
32
2.102
10.312
null
null
204.874
215.185
0.151
null
null
0.149
null
null
25
26
7
0
48
48
1,026.8
166.4
null
188,487.01
null
null
null
null
null
null
null
null
false
null
true
3,253.6
3,253.6
null
null
null
null
34.6
46.1
null
null
null
null
null
null
null
null
null
null
null
Layer-streaming inference manages memory by allocating a small, fixed pool of buffers for compute, keeping essential layers in memory, and asynchronously strea
round2
stream_r0_ngram.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
stream_r0_ngram
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
null
code_edit
null
null
null
null
null
null
null
null
null
106
48
2.102
8.983
null
null
76.835
85.818
0.612
null
null
0.559
null
null
10
45
38
0
48
48
1,026.8
166.4
null
75,394.8
null
null
null
null
null
null
null
null
false
null
true
3,253.6
3,253.6
null
null
null
null
3.6
49.7
null
null
null
null
null
null
null
null
null
null
null
#[derive(Debug, Clone)] pub struct StreamConfig { pub model_path: PathBuf, pub n_ctx: u32, pub n_batch: u32, pub resident_layers: Option<u32>, pub vram_budget_b
round2
stream_r0_ngram.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
stream_r0_ngram
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
null
open_ended
null
null
null
null
null
null
null
null
null
31
48
2.102
8.666
null
null
256.328
264.994
0.183
null
null
0.181
null
null
31
25
17
0
48
48
1,026.8
166.4
null
233,723.89
null
null
null
null
null
null
null
null
false
null
true
3,255.7
3,255.6
null
null
null
null
39.6
89.3
null
null
null
null
null
null
null
null
null
null
null
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
round3
r3_hybrid_r14_draft05b.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r3_hybrid_r14_draft05b
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
qwen2.5-0.5b-instruct-q4_k_m.gguf
prompt_heavy
null
null
null
null
null
null
null
null
null
340
32
6.886
5.916
null
null
69.787
75.704
0.444
null
null
0.423
null
null
15
62
16
14
34
48
3,244.4
166.4
null
79,826.54
null
null
null
null
null
null
null
null
false
null
true
6,079
6,079
null
null
null
null
-57.5
-1,016.1
null
null
null
null
null
null
null
null
null
null
null
Layer-streaming inference manages memory by allocating a small, fixed pool of reusable buffers for compute, keeping essential layers in memory, and asynchronou
round3
r3_hybrid_r14_draft05b.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r3_hybrid_r14_draft05b
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
qwen2.5-0.5b-instruct-q4_k_m.gguf
code_edit
null
null
null
null
null
null
null
null
null
106
48
6.886
3.49
null
null
18.261
21.75
2.574
null
null
2.207
null
null
6
43
36
14
34
48
3,244.4
166.4
null
31,930.62
null
null
null
null
null
null
null
null
false
null
true
6,083.3
6,083.3
null
null
null
null
0.3
-1,015.8
null
null
null
null
null
null
null
null
null
null
null
#[derive(Debug, Clone)] pub struct StreamConfig { pub model_path: PathBuf, pub n_ctx: u32, pub n_batch: u32, pub resident_layers: Option<u32>, pub vram_budget_b
round3
r3_hybrid_r14_draft05b.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r3_hybrid_r14_draft05b
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
qwen2.5-0.5b-instruct-q4_k_m.gguf
open_ended
null
null
null
null
null
null
null
null
null
31
48
6.886
3.289
null
null
31.885
35.174
1.474
null
null
1.365
null
null
11
64
34
14
34
48
3,244.4
166.4
null
58,539.46
null
null
null
null
null
null
null
null
false
null
true
6,083.6
6,083.6
null
null
null
null
1.2
-1,014.6
null
null
null
null
null
null
null
null
null
null
null
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
round3
r3_stream_r0.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r3_stream_r0
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
null
prompt_heavy
null
null
null
null
null
null
null
null
null
340
32
1.859
7.419
null
null
260.939
268.358
0.119
null
null
0.119
null
null
32
0
0
0
48
48
1,026.8
166.4
null
241,263.38
null
null
null
null
null
null
null
null
false
null
true
3,168
3,168
null
null
null
null
0.6
1.2
null
null
null
null
null
null
null
null
null
null
null
Layer-streaming inference manages memory by allocating a small, fixed pool of buffers for active computation, keeping only essential layers in memory, and asyn
round4
r4_auto_budget_5800_draft05b.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r4_auto_budget_5800_draft05b
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
qwen2.5-0.5b-instruct-q4_k_m.gguf
prompt_heavy
null
null
null
null
null
null
null
null
null
340
32
8.611
4.823
null
null
38.613
43.437
0.803
null
null
0.737
null
null
15
62
16
19
29
48
4,020.4
166.4
null
68,186.31
null
null
null
null
null
null
null
null
false
null
true
6,482
5,978.3
null
null
null
null
7.5
-178
null
12.65
null
17.17
null
null
null
null
null
null
null
Layer-streaming inference manages memory by allocating a small, fixed pool of reusable buffers for compute, keeping essential layers in memory, and asynchronou
round4
r4_auto_budget_5800_draft05b.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r4_auto_budget_5800_draft05b
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
qwen2.5-0.5b-instruct-q4_k_m.gguf
code_edit
null
null
null
null
null
null
null
null
null
106
48
8.611
3.137
null
null
13.633
16.77
3.447
null
null
2.862
null
null
6
43
36
19
29
48
4,020.4
166.4
null
27,274.52
null
null
null
null
null
null
null
null
false
null
true
6,482
5,971.7
null
null
null
null
-1.5
-179.5
null
12.59
null
4.4
null
null
null
null
null
null
null
#[derive(Debug, Clone)] pub struct StreamConfig { pub model_path: PathBuf, pub n_ctx: u32, pub n_batch: u32, pub resident_layers: Option<u32>, pub vram_budget_b
round4
r4_auto_budget_5800_draft05b.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r4_auto_budget_5800_draft05b
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
qwen2.5-0.5b-instruct-q4_k_m.gguf
open_ended
null
null
null
null
null
null
null
null
null
31
48
8.611
2.907
null
null
27.967
30.874
1.681
null
null
1.555
null
null
11
81
34
19
29
48
4,020.4
166.4
null
50,003.29
null
null
null
null
null
null
null
null
false
null
true
6,482
5,939.8
null
null
null
null
-23.4
-202.8
null
12.2
null
7.85
null
null
null
null
null
null
null
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
round4
r4_hybrid_r14_draft05b.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r4_hybrid_r14_draft05b
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
qwen2.5-0.5b-instruct-q4_k_m.gguf
prompt_heavy
null
null
null
null
null
null
null
null
null
340
32
7.8
5.093
null
null
39.028
44.12
0.794
null
null
0.725
null
null
15
62
16
14
34
48
3,244.4
166.4
null
79,992.96
null
null
null
null
null
null
null
null
false
null
true
5,708.1
5,348.1
null
null
null
null
-87.8
-690.5
null
12.85
null
17.72
null
null
null
null
null
null
null
Layer-streaming inference manages memory by allocating a small, fixed pool of reusable buffers for compute, keeping essential layers in memory, and asynchronou
round4
r4_hybrid_r14_draft05b.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r4_hybrid_r14_draft05b
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
qwen2.5-0.5b-instruct-q4_k_m.gguf
code_edit
null
null
null
null
null
null
null
null
null
106
48
7.8
3.155
null
null
13.958
17.113
3.367
null
null
2.805
null
null
6
43
36
14
34
48
3,244.4
166.4
null
31,930.62
null
null
null
null
null
null
null
null
false
null
true
5,708.1
5,352.6
null
null
null
null
-14.7
-705.2
null
12.36
null
4.41
null
null
null
null
null
null
null
#[derive(Debug, Clone)] pub struct StreamConfig { pub model_path: PathBuf, pub n_ctx: u32, pub n_batch: u32, pub resident_layers: Option<u32>, pub vram_budget_b
round4
r4_hybrid_r14_draft05b.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r4_hybrid_r14_draft05b
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
qwen2.5-0.5b-instruct-q4_k_m.gguf
open_ended
null
null
null
null
null
null
null
null
null
31
48
7.8
2.936
null
null
28.241
31.177
1.664
null
null
1.54
null
null
11
81
34
14
34
48
3,244.4
166.4
null
58,539.46
null
null
null
null
null
null
null
null
false
null
true
5,708.1
5,353
null
null
null
null
5.8
-699.4
null
12.27
null
7.97
null
null
null
null
null
null
null
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
round4
r4_hybrid_r16_draft05b.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r4_hybrid_r16_draft05b
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
qwen2.5-0.5b-instruct-q4_k_m.gguf
prompt_heavy
null
null
null
null
null
null
null
null
null
340
32
5.469
4.708
null
null
38.462
43.17
0.806
null
null
0.741
null
null
15
62
16
16
32
48
3,558.5
166.4
null
75,114.38
null
null
null
null
null
null
null
null
false
null
true
6,201.6
6,201.6
null
null
null
null
-54.8
-451.7
null
12.67
null
17.09
null
null
null
null
null
null
null
Layer-streaming inference manages memory by allocating a small, fixed pool of reusable buffers for compute, keeping essential layers in memory, and asynchronou
round4
r4_hybrid_r16_draft05b.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r4_hybrid_r16_draft05b
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
qwen2.5-0.5b-instruct-q4_k_m.gguf
code_edit
null
null
null
null
null
null
null
null
null
106
48
5.469
3.142
null
null
13.715
16.857
3.427
null
null
2.848
null
null
6
43
36
16
32
48
3,558.5
166.4
null
30,193.48
null
null
null
null
null
null
null
null
false
null
true
6,203.8
6,198.6
null
null
null
null
-11.1
-462.8
null
12.09
null
4.25
null
null
null
null
null
null
null
#[derive(Debug, Clone)] pub struct StreamConfig { pub model_path: PathBuf, pub n_ctx: u32, pub n_batch: u32, pub resident_layers: Option<u32>, pub vram_budget_b
round4
r4_hybrid_r16_draft05b.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r4_hybrid_r16_draft05b
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
qwen2.5-0.5b-instruct-q4_k_m.gguf
open_ended
null
null
null
null
null
null
null
null
null
31
48
5.469
2.91
null
null
27.794
30.704
1.691
null
null
1.563
null
null
11
81
34
16
32
48
3,558.5
166.4
null
55,083.88
null
null
null
null
null
null
null
null
false
null
true
6,203.8
6,179.3
null
null
null
null
7
-455.8
null
12.86
null
8.23
null
null
null
null
null
null
null
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
round4
r4_hybrid_r16_draft05b_t2.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r4_hybrid_r16_draft05b_t2
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
qwen2.5-0.5b-instruct-q4_k_m.gguf
prompt_heavy
null
null
null
null
null
null
null
null
null
340
32
5.178
5.791
null
null
53.552
59.343
0.579
null
null
0.539
null
null
15
62
16
16
32
48
3,558.5
166.4
null
75,262.1
null
null
null
null
null
null
null
null
false
null
true
6,199.6
6,199.6
null
null
null
null
-53.2
4.1
null
11.14
null
20.65
null
null
null
null
null
null
null
Layer-streaming inference manages memory by allocating a small, fixed pool of reusable buffers for compute, keeping essential layers in memory, and asynchronou
round4
r4_hybrid_r16_draft05b_t2.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r4_hybrid_r16_draft05b_t2
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
qwen2.5-0.5b-instruct-q4_k_m.gguf
code_edit
null
null
null
null
null
null
null
null
null
106
48
5.178
3.989
null
null
18.402
22.391
2.554
null
null
2.144
null
null
6
43
36
16
32
48
3,558.5
166.4
null
29,898.02
null
null
null
null
null
null
null
null
false
null
true
6,203.1
6,203.1
null
null
null
null
-18.3
-14.1
null
11.43
null
5.33
null
null
null
null
null
null
null
#[derive(Debug, Clone)] pub struct StreamConfig { pub model_path: PathBuf, pub n_ctx: u32, pub n_batch: u32, pub resident_layers: Option<u32>, pub vram_budget_b
round4
r4_hybrid_r16_draft05b_t2.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r4_hybrid_r16_draft05b_t2
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
qwen2.5-0.5b-instruct-q4_k_m.gguf
open_ended
null
null
null
null
null
null
null
null
null
31
48
5.178
3.805
null
null
38.511
42.316
1.22
null
null
1.134
null
null
11
81
34
16
32
48
3,558.5
166.4
null
55,083.88
null
null
null
null
null
null
null
null
false
null
true
6,203.1
6,203.1
null
null
null
null
-9.2
-23.4
null
11.34
null
10
null
null
null
null
null
null
null
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
round4
r4_hybrid_r16_draft05b_t4.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r4_hybrid_r16_draft05b_t4
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
qwen2.5-0.5b-instruct-q4_k_m.gguf
prompt_heavy
null
null
null
null
null
null
null
null
null
340
32
5.561
5.592
null
null
51.504
57.095
0.602
null
null
0.56
null
null
15
62
16
16
32
48
3,558.5
166.4
null
75,262.1
null
null
null
null
null
null
null
null
false
null
true
6,203
6,186.4
null
null
null
null
-245.6
-166
null
11.47
null
20.47
null
null
null
null
null
null
null
Layer-streaming inference manages memory by allocating a small, fixed pool of reusable buffers for compute, keeping essential layers in memory, and asynchronou
round4
r4_hybrid_r16_draft05b_t4.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r4_hybrid_r16_draft05b_t4
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
qwen2.5-0.5b-instruct-q4_k_m.gguf
code_edit
null
null
null
null
null
null
null
null
null
106
48
5.561
6.269
null
null
22.093
28.362
2.127
null
null
1.692
null
null
6
43
36
16
32
48
3,558.5
166.4
null
30,045.75
null
null
null
null
null
null
null
null
false
null
true
6,203
6,174.4
null
null
null
null
3.3
-162.6
null
10.5
null
6.21
null
null
null
null
null
null
null
#[derive(Debug, Clone)] pub struct StreamConfig { pub model_path: PathBuf, pub n_ctx: u32, pub n_batch: u32, pub resident_layers: Option<u32>, pub vram_budget_b
round4
r4_hybrid_r16_draft05b_t4.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r4_hybrid_r16_draft05b_t4
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
qwen2.5-0.5b-instruct-q4_k_m.gguf
open_ended
null
null
null
null
null
null
null
null
null
31
48
5.561
3.665
null
null
65.359
69.025
0.719
null
null
0.695
null
null
11
81
34
16
32
48
3,558.5
166.4
null
55,083.88
null
null
null
null
null
null
null
null
false
null
true
6,203
6,172.8
null
null
null
null
3.4
-159.2
null
9.77
null
14.04
null
null
null
null
null
null
null
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
round4
r4_hybrid_r16_ngram.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r4_hybrid_r16_ngram
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
null
prompt_heavy
null
null
null
null
null
null
null
null
null
340
32
4.735
7.575
null
null
147.639
155.213
0.21
null
null
0.206
null
null
25
29
7
16
32
48
3,558.5
166.4
null
125,190.63
null
null
null
null
null
null
null
null
false
null
true
5,596
5,596
null
null
null
null
58.6
73.6
null
9.86
null
47.84
null
null
null
null
null
null
null
Layer-streaming inference manages memory by allocating a small, fixed pool of buffers for active computation, keeping critical layers in memory, and asynchrono
round4
r4_hybrid_r16_ngram.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r4_hybrid_r16_ngram
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
null
code_edit
null
null
null
null
null
null
null
null
null
106
48
4.735
6.145
null
null
28.747
34.892
1.635
null
null
1.376
null
null
6
43
36
16
32
48
3,558.5
166.4
null
30,045.75
null
null
null
null
null
null
null
null
false
null
true
5,599.5
5,599.5
null
null
null
null
12.2
85.7
null
10.26
null
7.45
null
null
null
null
null
null
null
#[derive(Debug, Clone)] pub struct StreamConfig { pub model_path: PathBuf, pub n_ctx: u32, pub n_batch: u32, pub resident_layers: Option<u32>, pub vram_budget_b
round4
r4_hybrid_r16_ngram.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r4_hybrid_r16_ngram
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
null
open_ended
null
null
null
null
null
null
null
null
null
31
48
4.735
5.783
null
null
196.684
202.467
0.239
null
null
0.237
null
null
30
32
18
16
32
48
3,558.5
166.4
null
150,228.75
null
null
null
null
null
null
null
null
false
null
true
5,601.2
5,164.1
null
null
null
null
166.6
252.4
null
9.69
null
40.88
null
null
null
null
null
null
null
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
round4
r4_stream_r0.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r4_stream_r0
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
null
prompt_heavy
null
null
null
null
null
null
null
null
null
340
32
2.595
10.378
null
null
265.468
275.846
0.117
null
null
0.116
null
null
32
0
0
0
48
48
1,026.8
166.4
null
241,429.79
null
null
null
null
null
null
null
null
false
null
true
3,073.8
3,073.8
null
null
null
null
21.8
35.4
null
9.86
null
85
null
null
null
null
null
null
null
Layer-streaming inference manages memory by allocating a small, fixed pool of buffers for active computation, keeping only essential layers in memory, and asyn
round5
r5_zc_auto_budget_5800_draft05b.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r5_zc_auto_budget_5800_draft05b
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
qwen2.5-0.5b-instruct-q4_k_m.gguf
prompt_heavy
null
null
null
null
null
null
null
null
null
340
32
6.523
5.366
null
null
51.725
57.091
0.599
null
null
0.561
null
null
16
107
16
16
32
48
3,558.5
166.4
null
80,269.73
null
null
null
null
null
null
null
null
false
null
true
5,580.8
5,580.7
null
null
null
null
-58.6
369.9
null
11.19
null
19.97
null
null
null
null
null
null
null
Layer-streaming inference manages memory by allocating a small, fixed pool of buffers for active computation, keeping only essential layers in memory, and asyn
round5
r5_zc_auto_budget_5800_draft05b.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r5_zc_auto_budget_5800_draft05b
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
qwen2.5-0.5b-instruct-q4_k_m.gguf
code_edit
null
null
null
null
null
null
null
null
null
106
48
6.523
3.662
null
null
13.455
17.117
3.493
null
null
2.804
null
null
5
44
36
16
32
48
3,558.5
166.4
null
24,890.4
null
null
null
null
null
null
null
null
false
null
true
5,584.6
5,584.4
null
null
null
null
-12
357.9
null
11.84
null
4.22
null
null
null
null
null
null
null
#[derive(Debug, Clone)] pub struct StreamConfig { pub model_path: PathBuf, pub n_ctx: u32, pub n_batch: u32, pub resident_layers: Option<u32>, pub vram_budget_b
round5
r5_zc_auto_budget_5800_draft05b.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r5_zc_auto_budget_5800_draft05b
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
qwen2.5-0.5b-instruct-q4_k_m.gguf
open_ended
null
null
null
null
null
null
null
null
null
31
48
6.523
3.534
null
null
33.908
37.442
1.386
null
null
1.282
null
null
11
115
34
16
32
48
3,558.5
166.4
null
55,231.6
null
null
null
null
null
null
null
null
false
null
true
5,584.6
5,584.4
null
null
null
null
-12.8
345.1
null
11.27
null
8.79
null
null
null
null
null
null
null
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
round5
r5_zc_hybrid_r14_draft15b.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r5_zc_hybrid_r14_draft15b
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
qwen2.5-1.5b-instruct-q4_k_m.gguf
prompt_heavy
null
null
null
null
null
null
null
null
null
340
32
7.065
7.763
null
null
83.656
91.419
0.371
null
null
0.35
null
null
16
103
16
14
34
48
3,244.4
166.4
null
85,314.73
null
null
null
null
null
null
null
null
false
null
true
5,865.6
5,865.6
null
null
null
null
-41.6
-289.5
null
10.33
null
29.51
null
null
null
null
null
null
null
Layer-streaming inference manages memory by allocating a small, fixed pool of buffers for active computation, keeping only essential layers in memory, and asyn
round5
r5_zc_hybrid_r14_draft15b.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r5_zc_hybrid_r14_draft15b
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
qwen2.5-1.5b-instruct-q4_k_m.gguf
code_edit
null
null
null
null
null
null
null
null
null
106
48
7.065
5.808
null
null
22.176
27.984
2.119
null
null
1.715
null
null
5
44
36
14
34
48
3,244.4
166.4
null
26,608.85
null
null
null
null
null
null
null
null
false
null
true
5,870.3
5,870.3
null
null
null
null
4.3
-285.2
null
9.81
null
5.72
null
null
null
null
null
null
null
#[derive(Debug, Clone)] pub struct StreamConfig { pub model_path: PathBuf, pub n_ctx: u32, pub n_batch: u32, pub resident_layers: Option<u32>, pub vram_budget_b
round5
r5_zc_hybrid_r14_draft15b.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r5_zc_hybrid_r14_draft15b
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
qwen2.5-1.5b-instruct-q4_k_m.gguf
open_ended
null
null
null
null
null
null
null
null
null
31
48
7.065
5.726
null
null
50.56
56.285
0.93
null
null
0.853
null
null
10
94
35
14
34
48
3,244.4
166.4
null
53,051.28
null
null
null
null
null
null
null
null
false
null
true
5,870.3
5,870.3
null
null
null
null
-14.2
-299.4
null
10.43
null
12.23
null
null
null
null
null
null
null
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
round5
r5_zc_hybrid_r16_draft05b.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r5_zc_hybrid_r16_draft05b
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
qwen2.5-0.5b-instruct-q4_k_m.gguf
prompt_heavy
null
null
null
null
null
null
null
null
null
340
32
6.533
5.432
null
null
50.131
55.563
0.618
null
null
0.576
null
null
16
107
16
16
32
48
3,558.5
166.4
null
80,269.73
null
null
null
null
null
null
null
null
false
null
true
5,542.5
5,542.5
null
null
null
null
-51.2
-1,137.8
null
11.45
null
19.88
null
null
null
null
null
null
null
Layer-streaming inference manages memory by allocating a small, fixed pool of buffers for active computation, keeping only essential layers in memory, and asyn
round5
r5_zc_hybrid_r16_draft05b.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r5_zc_hybrid_r16_draft05b
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
qwen2.5-0.5b-instruct-q4_k_m.gguf
code_edit
null
null
null
null
null
null
null
null
null
106
48
6.533
3.664
null
null
13.386
17.05
3.511
null
null
2.815
null
null
5
44
36
16
32
48
3,558.5
166.4
null
25,038.13
null
null
null
null
null
null
null
null
false
null
true
5,547.2
5,547.2
null
null
null
null
-13
-1,150.8
null
10.73
null
3.81
null
null
null
null
null
null
null
#[derive(Debug, Clone)] pub struct StreamConfig { pub model_path: PathBuf, pub n_ctx: u32, pub n_batch: u32, pub resident_layers: Option<u32>, pub vram_budget_b
round5
r5_zc_hybrid_r16_draft05b.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r5_zc_hybrid_r16_draft05b
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
qwen2.5-0.5b-instruct-q4_k_m.gguf
open_ended
null
null
null
null
null
null
null
null
null
31
48
6.533
3.452
null
null
33.828
37.281
1.389
null
null
1.288
null
null
11
115
34
16
32
48
3,558.5
166.4
null
55,083.88
null
null
null
null
null
null
null
null
false
null
true
5,547.2
5,547.2
null
null
null
null
6.8
-1,144
null
10.97
null
8.52
null
null
null
null
null
null
null
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
round5
r5_zc_hybrid_r16_draft05b_temp07.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r5_zc_hybrid_r16_draft05b_temp07
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
qwen2.5-0.5b-instruct-q4_k_m.gguf
prompt_heavy
null
null
null
null
null
null
null
null
null
340
32
6.341
7.225
null
null
52.134
59.359
0.595
null
null
0.539
null
null
11
75
21
16
32
48
3,558.5
166.4
null
55,083.88
null
null
null
null
null
null
null
null
false
null
true
5,575.7
5,575.7
null
null
null
null
-24.5
257.5
null
10.25
null
19.02
null
null
null
null
null
null
null
The core memory-management mechanism described in the document is layer-streaming inference, which manages to fit large language models into limited memory by
round5
r5_zc_hybrid_r16_draft05b_temp07.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r5_zc_hybrid_r16_draft05b_temp07
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
qwen2.5-0.5b-instruct-q4_k_m.gguf
code_edit
null
null
null
null
null
null
null
null
null
106
48
6.341
5.607
null
null
20.974
26.581
2.241
null
null
1.806
null
null
5
44
36
16
32
48
3,558.5
166.4
null
25,038.13
null
null
null
null
null
null
null
null
false
null
true
5,580.9
5,580.9
null
null
null
null
0.2
257.7
null
10.46
null
5.79
null
null
null
null
null
null
null
#[derive(Debug, Clone)] pub struct StreamConfig { pub model_path: PathBuf, pub n_ctx: u32, pub n_batch: u32, pub resident_layers: Option<u32>, pub vram_budget_b
round5
r5_zc_hybrid_r16_draft05b_temp07.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r5_zc_hybrid_r16_draft05b_temp07
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
qwen2.5-0.5b-instruct-q4_k_m.gguf
open_ended
null
null
null
null
null
null
null
null
null
31
48
6.341
5.345
null
null
52.259
57.603
0.899
null
null
0.833
null
null
11
107
32
16
32
48
3,558.5
166.4
null
55,083.88
null
null
null
null
null
null
null
null
false
null
true
5,580.9
5,580.9
null
null
null
null
0.2
257.9
null
9.78
null
11.73
null
null
null
null
null
null
null
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
round5
r5_zc_hybrid_r16_ngram.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r5_zc_hybrid_r16_ngram
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
null
prompt_heavy
null
null
null
null
null
null
null
null
null
340
32
5.227
6.954
null
null
125.207
132.161
0.248
null
null
0.242
null
null
25
37
7
16
32
48
3,558.5
166.4
null
125,338.35
null
null
null
null
null
null
null
null
false
null
true
4,992.8
4,992.8
null
null
null
null
0.9
-11.4
null
9.98
null
41.23
null
null
null
null
null
null
null
Layer-streaming inference manages memory by allocating a small, fixed pool of buffers for active computation, keeping only essential layers in memory, and asyn
round5
r5_zc_hybrid_r16_ngram.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r5_zc_hybrid_r16_ngram
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
null
code_edit
null
null
null
null
null
null
null
null
null
106
48
5.227
5.484
null
null
26.205
31.689
1.794
null
null
1.515
null
null
6
50
31
16
32
48
3,558.5
166.4
null
29,898.02
null
null
null
null
null
null
null
null
false
null
true
4,995.1
4,995.1
null
null
null
null
0.7
-10.7
null
9.8
null
6.47
null
null
null
null
null
null
null
#[derive(Debug, Clone)] pub struct StreamConfig { pub model_path: PathBuf, pub n_ctx: u32, pub n_batch: u32, pub resident_layers: Option<u32>, pub vram_budget_b
round5
r5_zc_hybrid_r16_ngram.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r5_zc_hybrid_r16_ngram
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
null
open_ended
null
null
null
null
null
null
null
null
null
31
48
5.227
5.319
null
null
145.983
151.302
0.322
null
null
0.317
null
null
29
30
19
16
32
48
3,558.5
166.4
null
145,368.85
null
null
null
null
null
null
null
null
false
null
true
4,997.1
4,997.1
null
null
null
null
3.4
-7.2
null
9.8
null
30.88
null
null
null
null
null
null
null
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
round5
r5_zc_hybrid_r19_draft05b.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r5_zc_hybrid_r19_draft05b
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
qwen2.5-0.5b-instruct-q4_k_m.gguf
prompt_heavy
null
null
null
null
null
null
null
null
null
340
32
7.737
4.043
null
null
31.922
35.965
0.971
null
null
0.89
null
null
16
107
16
19
29
48
4,020.4
166.4
null
72,732.06
null
null
null
null
null
null
null
null
false
null
true
5,890
5,890
null
null
null
null
-77.4
-448.5
null
12.96
null
14.57
null
null
null
null
null
null
null
Layer-streaming inference manages memory by allocating a small, fixed pool of buffers for active computation, keeping only essential layers in memory, and asyn
round5
r5_zc_hybrid_r19_draft05b.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r5_zc_hybrid_r19_draft05b
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
qwen2.5-0.5b-instruct-q4_k_m.gguf
code_edit
null
null
null
null
null
null
null
null
null
106
48
7.737
2.45
null
null
8.27
10.72
5.684
null
null
4.478
null
null
5
44
36
19
29
48
4,020.4
166.4
null
22,728.77
null
null
null
null
null
null
null
null
false
null
true
5,894.9
5,894.9
null
null
null
null
0.4
-448.1
null
12.7
null
2.84
null
null
null
null
null
null
null
#[derive(Debug, Clone)] pub struct StreamConfig { pub model_path: PathBuf, pub n_ctx: u32, pub n_batch: u32, pub resident_layers: Option<u32>, pub vram_budget_b
round5
r5_zc_hybrid_r19_draft05b.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r5_zc_hybrid_r19_draft05b
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
qwen2.5-0.5b-instruct-q4_k_m.gguf
open_ended
null
null
null
null
null
null
null
null
null
31
48
7.737
2.262
null
null
21.636
23.898
2.172
null
null
2.009
null
null
11
115
34
19
29
48
4,020.4
166.4
null
50,003.29
null
null
null
null
null
null
null
null
false
null
true
5,895.1
5,895.1
null
null
null
null
-33.9
-482
null
12.48
null
6.21
null
null
null
null
null
null
null
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
round5
r5_zc_multiturn_r16_draft05b.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r5_zc_multiturn_r16_draft05b
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
qwen2.5-0.5b-instruct-q4_k_m.gguf
prompt_heavy
null
null
null
null
null
null
null
null
null
340
32
5.897
7.188
null
null
78.222
85.41
0.396
null
null
0.375
null
null
16
107
16
16
32
48
3,558.5
166.4
null
80,122
null
null
null
null
null
null
null
null
false
null
true
5,574.5
5,574.5
null
null
null
null
-74.1
16.3
null
10.19
null
27.19
null
null
null
null
null
null
null
Layer-streaming inference manages memory by allocating a small, fixed pool of buffers for active computation, keeping only essential layers in memory, and asyn
round5
r5_zc_multiturn_r16_draft05b.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r5_zc_multiturn_r16_draft05b_multiturn
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
null
multi_turn_prefix_reuse
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
340
5.174
377
5.4
0.96
null
O_DIRECT prevents page-cache thrashing by bypassing the operating system's page cache
round5
r5_zc_stream_r0.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r5_zc_stream_r0
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
null
prompt_heavy
null
null
null
null
null
null
null
null
null
340
32
1.99
6.96
null
null
163.279
170.239
0.19
null
null
0.188
null
null
32
0
0
0
48
48
1,026.8
166.4
null
241,429.79
null
null
null
null
null
null
null
null
false
null
true
2,492.4
2,492.4
null
null
null
null
0.8
1.3
null
10.71
null
56.97
null
null
null
null
null
null
null
Layer-streaming inference manages memory by allocating a small, fixed pool of buffers for active computation, keeping only essential layers in memory, and asyn
round5
r5_zc_stream_r0.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r5_zc_stream_r0
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
null
code_edit
null
null
null
null
null
null
null
null
null
106
48
1.99
5.618
null
null
247.709
253.327
0.19
null
null
0.189
null
null
48
0
0
0
48
48
1,026.8
166.4
null
361,728.64
null
null
null
null
null
null
null
null
false
null
true
2,492.4
2,492.4
null
null
null
null
36.1
37.4
null
10.67
null
56.33
null
null
null
null
null
null
null
#[derive(Debug, Clone)] pub struct StreamConfig { pub model_path: PathBuf, pub n_ctx: u32, pub n_batch: u32, pub resident_layers: Option<u32>, pub vram_budget_b
round5
r5_zc_stream_r0.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r5_zc_stream_r0
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
null
open_ended
null
null
null
null
null
null
null
null
null
31
48
1.99
5.368
null
null
247.761
253.129
0.19
null
null
0.19
null
null
48
0
0
0
48
48
1,026.8
166.4
null
362,061.48
null
null
null
null
null
null
null
null
false
null
true
2,492.4
2,492.4
null
null
null
null
26.8
64.2
null
10.7
null
56.41
null
null
null
null
null
null
null
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
round5
r5_zc_stream_r0_ngram.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r5_zc_stream_r0_ngram
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
null
prompt_heavy
null
null
null
null
null
null
null
null
null
340
32
2.011
6.246
null
null
109.781
116.027
0.282
null
null
0.276
null
null
25
37
7
0
48
48
1,026.8
166.4
null
188,653.43
null
null
null
null
null
null
null
null
false
null
true
2,500.8
2,500.7
null
null
null
null
0.9
1.3
null
11.16
null
40.47
null
null
null
null
null
null
null
Layer-streaming inference manages memory by allocating a small, fixed pool of buffers for active computation, keeping only essential layers in memory, and asyn
round5
r5_zc_stream_r0_ngram.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r5_zc_stream_r0_ngram
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
null
code_edit
null
null
null
null
null
null
null
null
null
106
48
2.011
4.906
null
null
22.969
27.875
2.046
null
null
1.722
null
null
6
50
31
0
48
48
1,026.8
166.4
null
45,236.88
null
null
null
null
null
null
null
null
false
null
true
2,503.1
2,503.1
null
null
null
null
0.2
1.5
null
11.03
null
6.41
null
null
null
null
null
null
null
#[derive(Debug, Clone)] pub struct StreamConfig { pub model_path: PathBuf, pub n_ctx: u32, pub n_batch: u32, pub resident_layers: Option<u32>, pub vram_budget_b
round5
r5_zc_stream_r0_ngram.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r5_zc_stream_r0_ngram
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
null
open_ended
null
null
null
null
null
null
null
null
null
31
48
2.011
4.612
null
null
128.056
132.668
0.367
null
null
0.362
null
null
29
30
19
0
48
48
1,026.8
166.4
null
218,644.93
null
null
null
null
null
null
null
null
false
null
true
2,505.1
2,505.1
null
null
null
null
1.5
3
null
10.97
null
30.31
null
null
null
null
null
null
null
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
round6
r6_auto_budget_6000_draft05b.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r6_auto_budget_6000_draft05b
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
qwen2.5-0.5b-instruct-q4_k_m.gguf
prompt_heavy
null
null
null
null
null
null
null
null
null
340
32
6.615
7.057
null
null
77.232
84.289
0.401
null
null
0.38
null
null
16
107
16
17
31
48
3,706.3
166.4
null
77,758.38
null
null
null
null
null
null
null
null
false
null
true
5,715
5,715
null
null
null
null
-91.1
151.9
null
10.11
null
26.62
null
null
null
null
null
null
null
Layer-streaming inference manages memory by allocating a small, fixed pool of buffers for active computation, keeping only essential layers in memory, and asyn
round6
r6_auto_budget_6000_draft05b.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r6_auto_budget_6000_draft05b
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
qwen2.5-0.5b-instruct-q4_k_m.gguf
code_edit
null
null
null
null
null
null
null
null
null
106
48
6.615
5.474
null
null
20.38
25.854
2.306
null
null
1.857
null
null
5
44
36
17
31
48
3,706.3
166.4
null
24,299.49
null
null
null
null
null
null
null
null
false
null
true
5,719.7
5,719.7
null
null
null
null
0.4
152.4
null
10.14
null
5.46
null
null
null
null
null
null
null
#[derive(Debug, Clone)] pub struct StreamConfig { pub model_path: PathBuf, pub n_ctx: u32, pub n_batch: u32, pub resident_layers: Option<u32>, pub vram_budget_b
round6
r6_auto_budget_6000_draft05b.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r6_auto_budget_6000_draft05b
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
qwen2.5-0.5b-instruct-q4_k_m.gguf
open_ended
null
null
null
null
null
null
null
null
null
31
48
6.615
5.291
null
null
51.813
57.105
0.907
null
null
0.841
null
null
11
115
34
17
31
48
3,706.3
166.4
null
53,458.88
null
null
null
null
null
null
null
null
false
null
true
5,719.7
5,719.7
null
null
null
null
0.1
152.5
null
9.78
null
11.64
null
null
null
null
null
null
null
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
round6
r6_hybrid_r12_draft15b.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r6_hybrid_r12_draft15b
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
qwen2.5-1.5b-instruct-q4_k_m.gguf
prompt_heavy
null
null
null
null
null
null
null
null
null
340
32
6.343
8.156
null
null
88.085
96.241
0.352
null
null
0.332
null
null
16
103
16
12
36
48
2,949
166.4
null
90,023.29
null
null
null
null
null
null
null
null
false
null
true
5,580.4
5,580.4
null
null
null
null
-82.5
391.5
null
10.05
null
30.22
null
null
null
null
null
null
null
Layer-streaming inference manages memory by allocating a small, fixed pool of buffers for active computation, keeping only essential layers in memory, and asyn
round6
r6_hybrid_r12_draft15b.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r6_hybrid_r12_draft15b
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
qwen2.5-1.5b-instruct-q4_k_m.gguf
code_edit
null
null
null
null
null
null
null
null
null
106
48
6.343
6.099
null
null
23.446
29.545
2.005
null
null
1.625
null
null
5
44
36
12
36
48
2,949
166.4
null
28,086.11
null
null
null
null
null
null
null
null
false
null
true
5,585
5,585
null
null
null
null
-69.6
321.9
null
9.65
null
5.94
null
null
null
null
null
null
null
#[derive(Debug, Clone)] pub struct StreamConfig { pub model_path: PathBuf, pub n_ctx: u32, pub n_batch: u32, pub resident_layers: Option<u32>, pub vram_budget_b
round6
r6_hybrid_r12_draft15b.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r6_hybrid_r12_draft15b
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
qwen2.5-1.5b-instruct-q4_k_m.gguf
open_ended
null
null
null
null
null
null
null
null
null
31
48
6.343
6.014
null
null
63.741
69.755
0.737
null
null
0.688
null
null
10
94
35
12
36
48
2,949
166.4
null
56,172.23
null
null
null
null
null
null
null
null
false
null
true
5,585
5,585
null
null
null
null
2.4
324.3
null
9.78
null
14.22
null
null
null
null
null
null
null
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
round6
r6_hybrid_r19_draft05b_temp07.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r6_hybrid_r19_draft05b_temp07
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
qwen2.5-0.5b-instruct-q4_k_m.gguf
prompt_heavy
null
null
null
null
null
null
null
null
null
340
32
7.145
3.976
null
null
20.457
24.432
1.515
null
null
1.31
null
null
11
75
21
19
29
48
4,020.4
166.4
null
50,003.29
null
null
null
null
null
null
null
null
false
null
true
6,037.1
6,037.1
null
null
null
null
-33.1
-382.9
null
13.05
null
9.97
null
null
null
null
null
null
null
The core memory-management mechanism described in the document is layer-streaming inference, which manages to fit large language models into limited memory by
round6
r6_hybrid_r19_draft05b_temp07.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r6_hybrid_r19_draft05b_temp07
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
qwen2.5-0.5b-instruct-q4_k_m.gguf
code_edit
null
null
null
null
null
null
null
null
null
106
48
7.145
2.358
null
null
7.928
10.286
5.928
null
null
4.666
null
null
5
44
36
19
29
48
4,020.4
166.4
null
22,728.77
null
null
null
null
null
null
null
null
false
null
true
6,042.3
6,042.3
null
null
null
null
-35.3
-418.2
null
13.81
null
2.96
null
null
null
null
null
null
null
#[derive(Debug, Clone)] pub struct StreamConfig { pub model_path: PathBuf, pub n_ctx: u32, pub n_batch: u32, pub resident_layers: Option<u32>, pub vram_budget_b
round6
r6_hybrid_r19_draft05b_temp07.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r6_hybrid_r19_draft05b_temp07
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
qwen2.5-0.5b-instruct-q4_k_m.gguf
open_ended
null
null
null
null
null
null
null
null
null
31
48
7.145
2.168
null
null
20.461
22.629
2.297
null
null
2.121
null
null
11
107
32
19
29
48
4,020.4
166.4
null
50,003.29
null
null
null
null
null
null
null
null
false
null
true
6,042.3
6,042.3
null
null
null
null
4.3
-413.9
null
13.11
null
6.18
null
null
null
null
null
null
null
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
round6
r6_hybrid_r19_draft05b_th2.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r6_hybrid_r19_draft05b_th2
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
qwen2.5-0.5b-instruct-q4_k_m.gguf
prompt_heavy
null
null
null
null
null
null
null
null
null
340
32
7.943
3.957
null
null
31.355
35.312
0.989
null
null
0.906
null
null
16
107
16
19
29
48
4,020.4
166.4
null
72,732.06
null
null
null
null
null
null
null
null
false
null
true
5,927.6
5,927.6
null
null
null
null
-76.5
-829.4
null
12.61
null
13.92
null
null
null
null
null
null
null
Layer-streaming inference manages memory by allocating a small, fixed pool of buffers for active computation, keeping only essential layers in memory, and asyn
round6
r6_hybrid_r19_draft05b_th2.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r6_hybrid_r19_draft05b_th2
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
qwen2.5-0.5b-instruct-q4_k_m.gguf
code_edit
null
null
null
null
null
null
null
null
null
106
48
7.943
2.355
null
null
7.874
10.23
5.969
null
null
4.692
null
null
5
44
36
19
29
48
4,020.4
166.4
null
22,728.77
null
null
null
null
null
null
null
null
false
null
true
5,932.5
5,932.5
null
null
null
null
-11.2
-840.6
null
13.02
null
2.78
null
null
null
null
null
null
null
#[derive(Debug, Clone)] pub struct StreamConfig { pub model_path: PathBuf, pub n_ctx: u32, pub n_batch: u32, pub resident_layers: Option<u32>, pub vram_budget_b
round6
r6_hybrid_r19_draft05b_th2.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r6_hybrid_r19_draft05b_th2
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
qwen2.5-0.5b-instruct-q4_k_m.gguf
open_ended
null
null
null
null
null
null
null
null
null
31
48
7.943
2.189
null
null
32.682
34.871
1.438
null
null
1.377
null
null
11
115
34
19
29
48
4,020.4
166.4
null
50,003.29
null
null
null
null
null
null
null
null
false
null
true
5,932.7
5,932.7
null
null
null
null
-2.4
-843
null
11.23
null
8.16
null
null
null
null
null
null
null
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
round6
r6_hybrid_r19_ngram.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r6_hybrid_r19_ngram
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
null
prompt_heavy
null
null
null
null
null
null
null
null
null
340
32
6.029
4.353
null
null
64.64
68.993
0.48
null
null
0.464
null
null
25
37
7
19
29
48
4,020.4
166.4
null
113,643.85
null
null
null
null
null
null
null
null
false
null
true
5,458.6
5,458.6
null
null
null
null
3.9
-45.7
null
11.94
null
25.74
null
null
null
null
null
null
null
Layer-streaming inference manages memory by allocating a small, fixed pool of buffers for active computation, keeping only essential layers in memory, and asyn
round6
r6_hybrid_r19_ngram.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r6_hybrid_r19_ngram
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
null
code_edit
null
null
null
null
null
null
null
null
null
106
48
6.029
3.152
null
null
13.547
16.699
3.469
null
null
2.874
null
null
6
50
31
19
29
48
4,020.4
166.4
null
27,274.52
null
null
null
null
null
null
null
null
false
null
true
5,460.9
5,460.9
null
null
null
null
0.7
-45
null
11.95
null
4.16
null
null
null
null
null
null
null
#[derive(Debug, Clone)] pub struct StreamConfig { pub model_path: PathBuf, pub n_ctx: u32, pub n_batch: u32, pub resident_layers: Option<u32>, pub vram_budget_b
round6
r6_hybrid_r19_ngram.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r6_hybrid_r19_ngram
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
null
open_ended
null
null
null
null
null
null
null
null
null
31
48
6.029
2.87
null
null
75.455
78.325
0.623
null
null
0.613
null
null
29
30
19
19
29
48
4,020.4
166.4
null
131,826.86
null
null
null
null
null
null
null
null
false
null
true
5,463
5,462.9
null
null
null
null
22.1
-23
null
11.6
null
18.93
null
null
null
null
null
null
null
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
round6
r6_hybrid_r20_draft05b.json
NVIDIA Jetson Orin Nano (8 GiB unified LPDDR5, CUDA sm_87, dm-crypt NVMe)
orin-nano-8gb-cuda
layerstream-plugin
0.1.0
null
r6_hybrid_r20_draft05b
ok
Qwen2.5-14B-Instruct-Q4_K_M.gguf
qwen2.5-0.5b-instruct-q4_k_m.gguf
prompt_heavy
null
null
null
null
null
null
null
null
null
340
32
6.749
4.078
null
null
30.016
34.094
1.033
null
null
0.939
null
null
16
107
16
20
28
48
4,168.1
166.4
null
70,534.86
null
null
null
null
null
null
null
null
false
null
true
6,152.5
6,069.9
null
null
null
null
-74.6
-171.5
null
12.89
null
13.73
null
null
null
null
null
null
null
Layer-streaming inference manages memory by allocating a small, fixed pool of buffers for active computation, keeping only essential layers in memory, and asyn
End of preview. Expand in Data Studio

Out-of-Core NVMe Layer Streaming on an 8 GB Jetson Orin Nano: 4-Engine Hardware Benchmark Dataset

Source code & v0.1.0 release: github.com/kleinbem/layerstream-plugin (v0.1.0 releases Β· engineering details) Β· Interactive Explorer: kleinbem/layerstream-benchmarks Space

Reproducible bare-metal benchmark dataset (130 normalized runs, 76 columns in Parquet + JSON + JSONL, plus reproduction scripts in scripts/) comparing four inference approaches on an 8 GB NVIDIA Jetson Orin Nano running oversized dense and sparse MoE models (7B MoE, 14B at 8.37 GiB, 30B-A3B MoE at 17.3 GiB, and 32B at 18.49 GiB β€” 2.53Γ— total system RAM):

  1. layerstream-plugin v0.1.0 (113 rows across iterative tuning rounds R2–R8) β€” Native Rust + stock unpatched llama.cpp (0693dbfb) with page-aligned O_DIRECT β†’ cudaHostAlloc async ping-pong GPU streaming, auto-sized hybrid partial-residency, sparse MoE active-expert streaming + scan-resistant TinyLFU hot-expert RAM cache, and speculative verification (greedy & temperature > 0).
  2. Stock llama.cpp (0693dbfb, CUDA sm_87, 12 rows) β€” Standard mmap + -ngl / -ot exps=CPU partial offload (also representing the underlying GGUF runtime used by Ollama and LM Studio).
  3. Hugging Face accelerate (transformers 5.23.0 + bitsandbytes 0.50.2, 1 row) β€” device_map="auto" + offload_folder + 4-bit NF4 disk offloading.
  4. AirLLM 4.0.0 (torch 2.14.1+cu126, bitsandbytes 0.50.2, transformers 5.23.0, 4 rows) β€” Python layer-by-layer sharded streaming (4bit & BF16).

Every row was measured directly on bare metal with VmHWM peak RSS, /proc/meminfo page-cache & swap deltas, kernel pread_mib_s I/O telemetry (Round 8), and Jetson INA3221 hardware power-rail telemetry (idle_watts, avg_watts, net_watts, gross & net J/tok). Both wins and honest losses/limitations are reported transparently.


Hardware & Environment

Property Value
Device NVIDIA Jetson Orin Nano (aarch64, 6Γ— Cortex-A78AE @ 1.728 GHz performance)
Memory 7.3 GiB usable unified LPDDR5 (68 GB/s), Ampere 1024-core GPU (sm_87 @ 1,020 MHz)
Storage Crucial T500 PCIe NVMe SSD behind LUKS2 dm-crypt (AES-XTS) (ext4 on LVM)
OS / Stack NixOS 26.11 (aarch64-linux), JetPack 6.2 / CUDA 12.6, INA3221 hwmon VDD_IN rail (~7.7–8.1 W idle)
Models Qwen2.5-14B-Instruct (Q4_K_M 8.37 GiB GGUF vs. HF safetensors 4bit/BF16), Qwen2.5-32B-Instruct-Q4_K_M (18.49 GiB GGUF), OLMoE-1B-7B-0125-Instruct-Q2_K (2.5 GiB, 64 experts), Qwen3-30B-A3B-Q4_K_M (17.3 GiB, 128 experts), plus Qwen2.5-0.5B / 1.5B & Qwen3-0.6B draft models

Standardized Workloads (Exact Same Prompts Across Engines)

  1. prompt_heavy (340 prompt tokens β†’ 32 generated tokens): Systems-architecture summarization.
  2. code_edit (106–113 prompt tokens β†’ 48 generated tokens): Verbatim Rust struct edit (high structural overlap).
  3. open_ended (31 prompt tokens β†’ 48 generated tokens): Technical explanation from scratch (unconstrained prose).
  4. multi_turn_prefix_reuse (1,377 Turn-1 tokens β†’ 1,414 Turn-2 tokens): Multi-turn KV prefix reuse latency test.

1. Headline 4-Engine Comparison (Qwen2.5-14B-Instruct, 8.37 GiB on 7.3 GiB Orin Nano)

Round 8 (v0.1.0) rows report median across --reps 3 with --warmup, exact pread_mib_s, live swap_delta_mib, and gross / net J/tok (subtracting measured idle board watts).

Engine & Configuration Workload R Load TTFT Decode (tok/s) pread BW Peak RSS Swap $\Delta$ Power (Idle) Gross / Net J/tok
Stock llama.cpp (-ngl 12/18, -b 512)* all 3 β€” β€” OOM error_rc_1 β€” 5,978 MiB +3,711 MiB 9.9 W (8.0 W) β€”
Stock llama.cpp (-ngl 0 -dev none -b 512)* open_ended β€” β€” >300 s timeout β€” 6,110 MiB +363 MiB 10.6 W (7.9 W) β€”
HF accelerate (4-bit NF4 disk offload)† open_ended β€” 144.0 s >156 s timeout_300s β€” β€” β€” β€” β€”
AirLLM 4.0.0 (4bit, cold first-run)‑ prompt_heavy β€” 365.53 s 32.09 s 0.032 (32 tok) β€” 5,391 MiB β€” β€” β€”
AirLLM 4.0.0 (4bit, warm shards)‑ code_edit β€” 2.55 s 33.37 s 0.033 (8 tok) β€” 3,536 MiB β€” 9.30 W 288.2 / β€”
AirLLM 4.0.0 (4bit, warm shards)‑ open_ended β€” 2.57 s 32.86 s 0.033 (8 tok) β€” 3,531 MiB β€” 9.33 W 287.1 / β€”
AirLLM 4.0.0 (BF16, prefetching=True)‑ open_ended β€” 81.81 s 45.41 s 0.023 (4 tok) β€” 5,218 MiB β€” 9.31 W 412.4 / β€”
layerstream R8 (R=0 pure stream, reps=2) open_ended 0/48 2.99 s 5.97 s 0.174 (16 tok) 1,427 MiB/s 2,458 MiB -3.0 MiB 10.37 W (7.71 W) 59.75 / 15.33
layerstream R8 (auto + 0.5B, reps=3) prompt_heavy 4/48 4.94 s 4.47 s 0.760 (0.755–0.760) 2,796 MiB/s 3,709 MiB -0.3 MiB 13.21 W (7.77 W) 18.69 / 7.69
layerstream R8 (auto + 0.5B, reps=3) open_ended 4/48 4.94 s 2.95 s 1.578 (1.564–1.582) 2,794 MiB/s 3,716 MiB 0.0 MiB 12.84 W (7.77 W) 8.75 / 3.45
layerstream R8 (auto + 0.5B, reps=3) code_edit 4/48 4.94 s 3.26 s 3.158 (2.244–3.159) 2,800 MiB/s 3,714 MiB -0.3 MiB 12.88 W (7.77 W) 4.87 / 1.93
layerstream R8 (b5400 + 0.5B, reps=3) prompt_heavy 7/48 4.58 s 4.24 s 0.824 (0.817–0.827) 2,924 MiB/s 4,171 MiB 0.0 MiB 13.29 W (7.70 W) 17.51 / 7.37
layerstream R8 (b5400 + 0.5B, reps=3) open_ended 7/48 4.58 s 2.74 s 1.699 (1.685–1.713) 2,932 MiB/s 4,178 MiB 0.0 MiB 13.29 W (7.70 W) 8.42 / 3.54
layerstream R8 (b5400 + 0.5B, reps=3) code_edit 7/48 4.58 s 3.01 s 3.405 (3.284–3.414) 2,961 MiB/s 4,176 MiB 0.0 MiB 13.16 W (7.70 W) 4.61 / 1.91
layerstream R6 (R=20 + 0.5B, max R) code_edit 20/48 6.75 s 2.31 s 5.974 (open: 2.109) β€” 6,153 MiB β€” 12.89 W 2.73 / β€”
  • *Stock llama.cpp also represents the underlying GGUF runtime used by Ollama and LM Studio.
  • †HF accelerate and ‑AirLLM 4.0.0 ran in the same torch 2.14.1+cu126 Python venv (which emits a Jetson CC 8.7 unsupported-build warning). AirLLM warm runs generated 8 tokens (BF16: 4 tokens) vs. 48 tokens for layerstream-plugin, and AirLLM uses bnb 4-bit (36.8 GiB total disk footprint including HF cache + shards) vs. GGUF Q4_K_M (8.37 GiB single file, zero conversion).
  • Multi-turn KV prefix reuse (r8_14b_auto_draft05b_multiturn): Turn 1 (1,377 tokens) 14.877 s β†’ Turn 2 (1,414 tokens) 2.993 s (4.97Γ— faster).

2. Running a 32B Model (18.49 GiB, 2.53Γ— Total System RAM) on 7.3 GiB Orin Nano

Engine & Mode (Qwen2.5-32B-Instruct-Q4_K_M) Workload R Load TTFT Decode (tok/s) pread BW Peak RSS Swap $\Delta$ Gross / Net J/tok
Stock llama.cpp (-ngl 10) open_ended β€” β€” >300 s timeout β€” 5,715 MiB +17,243 MiB β€”
layerstream R6 (R=0, 0 draft) prompt_heavy 0/64 2.14 s 14.48 s 0.094 (559.7 GiB read) β€” 2,957 MiB β€” 119.1 / β€”
layerstream R8 (b5400 + 0.5B, reps=2) prompt_heavy 2/64 3.76 s 18.01 s 0.171 (0.156–0.185) 1,229 MiB/s 4,109 MiB 0.0 MiB 64.13 / 14.21
layerstream R8 (b5400 + 0.5B, reps=2) open_ended 2/64 3.76 s 16.56 s 0.304 (0.304–0.304) 1,311 MiB/s 4,115 MiB 0.0 MiB 36.61 / 8.30
layerstream R8 (b5400 + 0.5B, reps=2) code_edit 2/64 3.76 s 14.70 s 0.840 (0.838–0.841) 1,338 MiB/s 4,113 MiB -0.5 MiB 15.51 / 3.82
layerstream R6 (b5600 + 0.5B, clean RAM) code_edit 6/64 6.84 s 8.42 s 0.944 (open: 0.506) β€” 5,294 MiB β€” 14.3 / β€”

*(Note: 573,082 MiB read in r6_32b_stream_r0 equals 559.7 GiB.)*


3. Sparse MoE Benchmarks (OLMoE-1B-7B & Qwen3-30B-A3B)

Model & Engine Workload R Load TTFT Decode (tok/s) Cache Hit pread BW Peak RSS Swap $\Delta$ Gross / Net J/tok
OLMoE-1B-7B (Q2_K) β€” layerstream full-layer open_ended 0/16 0.51 s 1.45 s 0.763 β€” 1,943 MiB/s 1,312 MiB -2.0 MiB 14.34 / 4.01
OLMoE-1B-7B (Q2_K) β€” layerstream sparse + TinyLFU prompt_heavy 0/16 0.68 s 2.08 s 13.667 (13.65–13.92) 100% 1,405 MiB/s 2,921 MiB 0.0 MiB 1.67 / 0.57
OLMoE-1B-7B (Q2_K) β€” layerstream sparse + TinyLFU code_edit 0/16 0.68 s 1.87 s 13.054 (12.45–13.24) 99.3% 1,269 MiB/s 2,953 MiB 0.0 MiB 1.44 / 0.51
OLMoE-1B-7B (Q2_K) β€” layerstream sparse + TinyLFU open_ended 0/16 0.68 s 1.72 s 12.047 (11.45–12.51) 98.0% 1,168 MiB/s 2,971 MiB 0.0 MiB 1.40 / 0.45
Qwen3-30B-A3B (17.3 GiB) β€” Stock llama.cpp GPU MoE open_ended β€” β€” OOM error_rc_1 β€” β€” 6,014 MiB +12,342 MiB β€”
Qwen3-30B-A3B (17.3 GiB) β€” Stock llama.cpp CPU mmap open_ended β€” 25.34 s 17.37 s 1.270 (wins!) OS cache β€” 6,071 MiB +248 MiB 15.26 / β€”
Qwen3-30B-A3B (17.3 GiB) β€” layerstream sparse + TinyLFU + 0.6B open_ended 0/48 3.26 s 11.24 s 0.317 (0.233–0.400) 23.3% 708 MiB/s 3,969 MiB 0.0 MiB 32.98 / 3.21
Qwen3-30B-A3B (17.3 GiB) β€” layerstream sparse + TinyLFU + 0.6B code_edit 0/48 33.06 s 13.60 s 0.113 (0.066–0.159) 16.0% 267 MiB/s 3,974 MiB -2.0 MiB 89.98 / 5.21

4. Honest Trade-Offs & Limitations

  • Where stock llama.cpp CPU mmap beats layerstream-plugin: On Qwen3-30B-A3B (128 experts/layer) on orin-nano, stock llama.cpp CPU mmap (-ngl 0 -dev none -b 512) reached 1.270 tok/s (6,071 MiB RSS, +248 MiB swap) because the Linux page cache absorbs random hot expert slices in RAM, whereas layerstream-plugin's small random O_DIRECT bursts trigger 30-second Tegra PCIe/NVMe IRQ stalls (nvme1: I/O tag ... timeout, completion polled) on orin-nano.
  • Jetson CUDA uma_zero_copy=false: Upstream llama.cpp currently disables CUDA_Host unified memory on Jetson (upstream: ggml-org/llama.cpp#15034). All Jetson CUDA rows in this dataset therefore have uma_zero_copy: false and is_unified_memory: true, using pinned cudaHostAlloc + cudaMemcpyAsync instead of host-mapped device memory.
  • Post-eviction CUDA driver context: Cold startup uses 0 MB GPU memory; after serving and idle eviction, all model/KV/slot buffers are freed, leaving ~230–325 MiB held by CUDA's primary runtime context.
  • Host dm-crypt bandwidth variance: On orin-nano, LUKS2 dm-crypt sequential pread measured 2,794–2,961 MiB/s under clean memory conditions and dropped to 1,025–1,338 MiB/s under zram pressure (r8_14b_*_temp07, r8_32b_*); every Round 8 row records exact pread_mib_s alongside throughput so bandwidth effects are never hidden.
  • Single-sweep latency floor on dense R=0: Without hybrid residency or speculation, dense NVMe streaming is strictly bounded by model_size / nvme_bw (~0.17–0.19 tok/s on 14B, ~0.094 tok/s on 32B).

5. Dataset Schema (76 Columns) & Reproduction

Each of the 130 rows in data/train-00000-of-00001.parquet (benchmarks.json, benchmarks.jsonl) includes: round, source_file, hardware, hardware_id, engine, engine_version, git_rev, mode, status, model, draft_model, workload, chat_template_applied, warmup, reps, temperature, seed, threads, kv_cache_type, vram_budget_auto, vram_budget_mb, prompt_tokens, completion_tokens, load_time_s, ttft_s, ttft_s_min, ttft_s_max, decode_time_s, total_time_s, decode_tok_per_s, decode_tok_per_s_min, decode_tok_per_s_max, total_tok_per_s, total_tok_per_s_min, total_tok_per_s_max, sweeps, drafted_tokens, accepted_tokens, resident_layers, streamed_layers, n_layers, resident_mib, slot_mib, expert_slot_mib, io_mib_read, pread_mib_s, effective_io_mib_s, expert_cache_hits, expert_cache_misses, expert_cache_hit_rate, expert_cache_mib_saved, moe_layers, moe_active_streaming, uma_zero_copy, is_unified_memory, direct_io_active, peak_rss_mib, current_rss_mib, mem_avail_before_mib, mem_avail_after_mib, swap_delta_mib, cuda_max_allocated_mib, page_cache_delta_mib, page_cache_delta_since_start_mib, idle_watts, avg_watts, net_watts, joules_per_tok, net_joules_per_tok, turn1_prompt_tokens, turn1_ttft_s, turn2_prompt_tokens, turn2_ttft_s, ttft_speedup_x, notes, output_preview.

# Reproduce layerstream-plugin (CUDA sm_87)
cargo run --release -p layerstream-core --example bench --features cuda -- \
  --model Qwen2.5-14B-Instruct-Q4_K_M.gguf \
  --draft-model qwen2.5-0.5b-instruct-q4_k_m.gguf \
  --vram-budget-mb auto --kv-cache-type q8_0 --threads 4 --warmup --reps 3 --multi-turn

# Reproduce stock llama.cpp competitor baseline
python3 benchmarks/bench_llama_cpp.py \
  --bin ./llama-completion --model Qwen2.5-14B-Instruct-Q4_K_M.gguf \
  --ngl 12 --extra-arg -b --extra-arg 512 --extra-arg -ub --extra-arg 512

# Reproduce HF accelerate baseline
python3 benchmarks/bench_accelerate.py \
  --model Qwen/Qwen2.5-14B-Instruct --offload-dir ./accel-offload \
  --gpu-max-mem 3GiB --cpu-max-mem 1GiB

# Reproduce AirLLM 4.0.0 baseline
python3 benchmarks/bench_airllm.py \
  --model Qwen/Qwen2.5-14B-Instruct --compression 4bit \
  --layer-shards-dir ./airllm-shards

# Regenerate normalized JSON / JSONL / Parquet dataset
python3 benchmarks/export_hf_dataset.py \
  --results-dir benchmarks/results --out-dir benchmarks/hf-dataset
Downloads last month
-

Space using kleinbem/layerstream-orin-nano-benchmarks 1

Collection including kleinbem/layerstream-orin-nano-benchmarks