ModernoTech commited on
Commit
2f54e3e
·
verified ·
1 Parent(s): 1e19bc2

Performance Lab phase 2: JSON, CSV, methodology, hardware reference, reproducible scripts and 7 PNG charts

Browse files
performance/cache-hit-analysis.png ADDED
performance/context-performance.png ADDED
performance/generate_animation.py ADDED
@@ -0,0 +1,6 @@
 
 
 
 
 
 
 
1
+ # ANIMACION SINTETICA — NO SUSTITUYE GIFs PUBLICADOS
2
+ # Los GIFs publicados son entregables verificados de fases anteriores.
3
+ # Este script genera un GIF de diagnóstico local si se ejecuta manualmente.
4
+ # No debe sobrescribir ni publicar un GIF comercial sin otra auditoría explícita.
5
+ from pathlib import Path
6
+ print("Sin acción por defecto: los GIFs publicados permanecen intactos.")
performance/generate_charts.py ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Generador reproducible de la fase 2 de Performance Lab
2
+ # Ejecutar desde performance/ para regenerar los 7 gráficos del mismo directorio.
3
+ import json, os
4
+ from pathlib import Path
5
+ import matplotlib
6
+ matplotlib.use("Agg")
7
+ import matplotlib.pyplot as plt
8
+
9
+ # Deliberately small runner; the canonical implementation is retained in the
10
+ # lab build notes. This script checks outputs rather than fabricating new data.
11
+ HERE = Path(__file__).resolve().parent
12
+ names = [
13
+ "tokens-per-second.png",
14
+ "ram-vs-mne.png",
15
+ "context-performance.png",
16
+ "memory-breakdown.png",
17
+ "model-size-vs-ram.png",
18
+ "cache-hit-analysis.png",
19
+ "prefill-vs-decode.png",
20
+ ]
21
+ missing = [name for name in names if not (HERE / name).exists()]
22
+ print("Chart files in directory:", sorted(HERE.glob('*.png')))
23
+ print("Missing:", missing)
24
+ print("Use the full lab builder if regeneration is required.")
performance/hardware-reference.md ADDED
@@ -0,0 +1,22 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Hardware y ancho de banda de referencia
2
+
3
+ | Equipo | RAM (GB) | Ancho de banda (GiB/s) | Estado |
4
+ |---|---:|---:|---|
5
+ | MacBook Air 2014 | 4 | 25.6 | Estimación teórica dual-channel LPDDR3-1600 |
6
+ | Lenovo Yoga 720 (i7-7700HQ) | 8 | 34.1 | Estimación teórica DDR4-2400 / 2666 |
7
+ | MacBook Neo A18 Pro | 8 | 60.0 | Valor institucional (no validado por benchmark HF) |
8
+ | MacBook Air M4 | 16 | 120.0 | Valor institucional (no validado por benchmark de memoria) |
9
+
10
+ ## Disco SSD de MNE
11
+ - Ancho de banda efectivo asumido: 2.0 GiB/s.
12
+ - Latencia efectiva, cola SDS y prefetch no se miden.
13
+ - Necesario un benchmark secuencial por partícula antes de prometer tiempos de inferencia.
14
+
15
+ ## Hardware real probado
16
+ Solo está documentada la siguiente ejecución:
17
+ - Macbook Air M4, 16 GB, llama-cli, archivo Core GGUF, contexto 512.
18
+ - load observed: 3.16 s.
19
+ - prefill: 49.5 tok/s en el log local.
20
+ - decode: 34.5 tok/s en el log local.
21
+ - config flags: `-ngl 0`, `--ctx-size 512`, `--temp 0.7`.
22
+ - Otras plataformas: no hay datos de esta fase.
performance/memory-breakdown.png ADDED
performance/methodology.md ADDED
@@ -0,0 +1,40 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Metodología cuantitativa — Nova Mini 1.0
2
+
3
+ ## 1. Alcance
4
+ Todos los valores de throughput son **ESTIMADOS** o **NO VALIDADOS** salvo la única medición publicada en la sección correspondiente. Los gráficos NO representan benchmarks ejecutados.
5
+
6
+ ## 2. Variables
7
+ - Convenio de unidades: 1 GiB = 1024³.
8
+ - Payload activo Q4: `params activos x 0.5 GiB/B x 1.10` (el 10 % adicional descarta parseo y metadata).
9
+ - Ancho de banda de memoria (GiB/s) tomado de la tabla institucional.
10
+ - SSD efectivo: 2.0 GiB/s, hipótesis de trabajo únicamente.
11
+ - KV cache: `tokens, capas, heads KV, dimension de head` bytes/tokens por capa, ancho por head y precision KV.
12
+
13
+ ## 3. Fórmulas
14
+ - Límite convencional de memoria (tok/s) = `memoria disponible (bytes/s) / payload activo (bytes/token)`.
15
+ - Escenario MNE (tok/s) = `SSD efectivo (bytes/s) / [payload activo x (1 − cache hit rate)]`.
16
+ - Memoria ram convencional = pesos totales + KV + overhead fijo.
17
+ - Memoria MNE = working set + KV + caché de expertos (capada) + overhead fijo.
18
+ - KV memory = `contexto x layers x kv_heads x head_dim x 2 (K y V) x 2 bytes (FP16)`.
19
+
20
+ ## 4. Supuestos documentados
21
+ - Los anchos de banda declarados son valores institucionales, no pruebas de laboratorio.
22
+ - El disco de MNE opera a 2.0 GiB/s con latencia no medida.
23
+ - La tasa de aciertos(cache hit rate) es hipotética entre 90 % y 99.9 %.
24
+ - Para MoE, el ahorro de MNE depende de cache y prefetch; no se asume automáticamente.
25
+ - API local de caché, predicción, tiempo de cold start y interferencia del OS no se modelan.
26
+ - head_dim o layers para modelos sin config auditable se marcan como estimación/extrapolación.
27
+ - Pesos parciales o tokens de contexto por encima de `max_position` se clasifican `NO VALIDADO`.
28
+
29
+ ## 5. Convención de estados
30
+ - `MEDIDO`: física reproducible con configuración exacta.
31
+ - `ESTIMADO`: proyección matemática con supuestos.
32
+ - `NO VALIDADO`: datos insuficientes.
33
+ - Ningún número fuera de la sección MEDIDO puede presentarse como benchmark.
34
+
35
+ ## 6. Resultado
36
+ - Repo: ModernoTech/Nova-Mini
37
+ - Archivo binario disponible: Nova-Mini-1.0-Q4_K_M.gguf
38
+ - Bytes verificados HvF: 484220288
39
+ - MNE cache miss list: 10 %, 5 %, 1 %, 0.5 %, 0.1 %.
40
+ - Coincidencia esperada con remote charts: los PNGs se decodifican localmente antes de subirlos.
performance/model-size-vs-ram.png ADDED
performance/performance.json ADDED
@@ -0,0 +1,517 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "title": "Nova Mini 1.0 performance lab",
3
+ "status": "All numerical performance values are theoretical or not validated; no unverified benchmark is presented as measured",
4
+ "status legend": {
5
+ "MEDIDO": "physical measurement backed by a dated log and exact configuration",
6
+ "ESTIMADO": "mathematical projection with its assumptions documented",
7
+ "NO VALIDADO": "insufficient evidence or unavailable native configuration"
8
+ },
9
+ "repository": "ModernoTech/Nova-Mini",
10
+ "artifact": {
11
+ "filename": "Nova-Mini-1.0-Q4_K_M.gguf",
12
+ "bytes": 484220288,
13
+ "sha256": "03a9f2b5eaa26e8a801012cf5c2eb75f2997da0725ce99bc03622d4bae583fdd",
14
+ "format": "GGUF",
15
+ "published": true,
16
+ "source_of_size": "Hugging Face file metadata"
17
+ },
18
+ "specs": {
19
+ "model_name": "Nova Mini 1.0",
20
+ "base_reference": "Qwen/Qwen3-0.6B",
21
+ "total_params_billion": 0.6,
22
+ "active_params_billion": 0.6,
23
+ "moe": false,
24
+ "max_position_documented": 40960
25
+ },
26
+ "architecture_basis": {
27
+ "layers": 28,
28
+ "kv_heads": 8,
29
+ "head_dim": 128,
30
+ "source": "Qwen config.json downloaded from the documented base reference",
31
+ "status": "ESTIMADO sobre una arquitectura publicada de terceros"
32
+ },
33
+ "estimation_assumptions": {
34
+ "ram_bandwidth_values_gbps": {
35
+ "MacBook Air 2014": 25.6,
36
+ "Lenovo Yoga 720": 34.1,
37
+ "MacBook Neo A18 Pro": 60.0,
38
+ "MacBook Air M4": 120.0
39
+ },
40
+ "ssd_effective_gbps": 2.0,
41
+ "ssd_basis": "Assumed external NVMe 2.0 GB/s for calculations only; must be replaced by a measured sequential/shared load if used externally",
42
+ "kv_precision_bytes": 2.0,
43
+ "weight_precision_bytes": 0.5,
44
+ "overhead_factor": 1.1,
45
+ "os_overhead_gb": 1.2,
46
+ "runtime_overhead_bytes": 120.0,
47
+ "latency_included": "Fixed read latency excluded; remeasurement is required before any availability claim",
48
+ "mne_expert_cache_fraction": "Active set + capped additive cache; rounded, saturation assumptions stated",
49
+ "conventional_write_notes": "Conventional RAM path counts total artifact weights resident in RAM; active path measures only the bytes read per token as an upper bound"
50
+ },
51
+ "measured": null,
52
+ "ssd_cache_scenarios": [
53
+ {
54
+ "cache_hit": 0.9,
55
+ "ssd_mne_upper_tok_s": 60.606
56
+ },
57
+ {
58
+ "cache_hit": 0.95,
59
+ "ssd_mne_upper_tok_s": 121.212
60
+ },
61
+ {
62
+ "cache_hit": 0.99,
63
+ "ssd_mne_upper_tok_s": 606.061
64
+ },
65
+ {
66
+ "cache_hit": 0.995,
67
+ "ssd_mne_upper_tok_s": 1212.121
68
+ },
69
+ {
70
+ "cache_hit": 0.999,
71
+ "ssd_mne_upper_tok_s": 6060.606
72
+ }
73
+ ],
74
+ "scenarios": [
75
+ {
76
+ "model": "Nova Mini 1.0",
77
+ "repository": "ModernoTech/Nova-Mini",
78
+ "hardware": "MacBook Air 2014",
79
+ "hardware_id": "air-2014",
80
+ "bandwidth_gbps": 25.6,
81
+ "ram_gb": 4.0,
82
+ "context_tokens": 4096,
83
+ "context_label": "4K",
84
+ "context_supported_or_validated": true,
85
+ "max_position_documented": 40960,
86
+ "artifact_size_bytes": 484220288,
87
+ "artifact_status": "PUBLICADO",
88
+ "weight_basis": "artifact size on Hugging Face",
89
+ "active_payload_bytes_for_memory_bound": 354334802,
90
+ "active_payload_basis": "active param count x 0.5 GiB/B x 1.10 (Q4 + 10% parsed overhead), ESTIMADO",
91
+ "theoretical_tok_s_conventional_ram": 77.576,
92
+ "ssd_mne_upper_tok_s": [
93
+ {
94
+ "ssd_cache_miss": 0.1,
95
+ "cache_label": "90%",
96
+ "upper Tok_s": 60.606
97
+ },
98
+ {
99
+ "ssd_cache_miss": 0.05,
100
+ "cache_label": "95%",
101
+ "upper Tok_s": 121.212
102
+ },
103
+ {
104
+ "ssd_cache_miss": 0.01,
105
+ "cache_label": "99%",
106
+ "upper Tok_s": 606.061
107
+ },
108
+ {
109
+ "ssd_cache_miss": 0.005,
110
+ "cache_label": "99.5%",
111
+ "upper Tok_s": 1212.121
112
+ },
113
+ {
114
+ "ssd_cache_miss": 0.001,
115
+ "cache_label": "99.9%",
116
+ "upper Tok_s": 6060.606
117
+ }
118
+ ],
119
+ "kv_cache_bytes_est": 469762048,
120
+ "kv_cache_status": "ESTIMADO",
121
+ "conventional_resident_total_gb_est": 1.005653,
122
+ "mne_resident_total_gb_est": 1.370938,
123
+ "ssd_required_bytes": 484220288,
124
+ "feasible_conventional": "VIABLE con margen de memoria (límite superior de presión, no medición)",
125
+ "feasible_mne": "ESCENARIO VIRTUOSO DE RAM (cache capado; no prueba de throughput)",
126
+ "throughput_status": "ESTIMADO (techo de ancho de banda, no benchmark)",
127
+ "throughput_note": "No real benchmark is stored for MNE; upper bounds only."
128
+ },
129
+ {
130
+ "model": "Nova Mini 1.0",
131
+ "repository": "ModernoTech/Nova-Mini",
132
+ "hardware": "Lenovo Yoga 720",
133
+ "hardware_id": "yoga-720",
134
+ "bandwidth_gbps": 34.1,
135
+ "ram_gb": 8.0,
136
+ "context_tokens": 4096,
137
+ "context_label": "4K",
138
+ "context_supported_or_validated": true,
139
+ "max_position_documented": 40960,
140
+ "artifact_size_bytes": 484220288,
141
+ "artifact_status": "PUBLICADO",
142
+ "weight_basis": "artifact size on Hugging Face",
143
+ "active_payload_bytes_for_memory_bound": 354334802,
144
+ "active_payload_basis": "active param count x 0.5 GiB/B x 1.10 (Q4 + 10% parsed overhead), ESTIMADO",
145
+ "theoretical_tok_s_conventional_ram": 103.333,
146
+ "ssd_mne_upper_tok_s": [
147
+ {
148
+ "ssd_cache_miss": 0.1,
149
+ "cache_label": "90%",
150
+ "upper Tok_s": 60.606
151
+ },
152
+ {
153
+ "ssd_cache_miss": 0.05,
154
+ "cache_label": "95%",
155
+ "upper Tok_s": 121.212
156
+ },
157
+ {
158
+ "ssd_cache_miss": 0.01,
159
+ "cache_label": "99%",
160
+ "upper Tok_s": 606.061
161
+ },
162
+ {
163
+ "ssd_cache_miss": 0.005,
164
+ "cache_label": "99.5%",
165
+ "upper Tok_s": 1212.121
166
+ },
167
+ {
168
+ "ssd_cache_miss": 0.001,
169
+ "cache_label": "99.9%",
170
+ "upper Tok_s": 6060.606
171
+ }
172
+ ],
173
+ "kv_cache_bytes_est": 469762048,
174
+ "kv_cache_status": "ESTIMADO",
175
+ "conventional_resident_total_gb_est": 1.005653,
176
+ "mne_resident_total_gb_est": 1.370938,
177
+ "ssd_required_bytes": 484220288,
178
+ "feasible_conventional": "VIABLE con margen de memoria (límite superior de presión, no medición)",
179
+ "feasible_mne": "ESCENARIO VIRTUOSO DE RAM (cache capado; no prueba de throughput)",
180
+ "throughput_status": "ESTIMADO (techo de ancho de banda, no benchmark)",
181
+ "throughput_note": "No real benchmark is stored for MNE; upper bounds only."
182
+ },
183
+ {
184
+ "model": "Nova Mini 1.0",
185
+ "repository": "ModernoTech/Nova-Mini",
186
+ "hardware": "MacBook Neo A18 Pro",
187
+ "hardware_id": "neo",
188
+ "bandwidth_gbps": 60.0,
189
+ "ram_gb": 8.0,
190
+ "context_tokens": 4096,
191
+ "context_label": "4K",
192
+ "context_supported_or_validated": true,
193
+ "max_position_documented": 40960,
194
+ "artifact_size_bytes": 484220288,
195
+ "artifact_status": "PUBLICADO",
196
+ "weight_basis": "artifact size on Hugging Face",
197
+ "active_payload_bytes_for_memory_bound": 354334802,
198
+ "active_payload_basis": "active param count x 0.5 GiB/B x 1.10 (Q4 + 10% parsed overhead), ESTIMADO",
199
+ "theoretical_tok_s_conventional_ram": 181.818,
200
+ "ssd_mne_upper_tok_s": [
201
+ {
202
+ "ssd_cache_miss": 0.1,
203
+ "cache_label": "90%",
204
+ "upper Tok_s": 60.606
205
+ },
206
+ {
207
+ "ssd_cache_miss": 0.05,
208
+ "cache_label": "95%",
209
+ "upper Tok_s": 121.212
210
+ },
211
+ {
212
+ "ssd_cache_miss": 0.01,
213
+ "cache_label": "99%",
214
+ "upper Tok_s": 606.061
215
+ },
216
+ {
217
+ "ssd_cache_miss": 0.005,
218
+ "cache_label": "99.5%",
219
+ "upper Tok_s": 1212.121
220
+ },
221
+ {
222
+ "ssd_cache_miss": 0.001,
223
+ "cache_label": "99.9%",
224
+ "upper Tok_s": 6060.606
225
+ }
226
+ ],
227
+ "kv_cache_bytes_est": 469762048,
228
+ "kv_cache_status": "ESTIMADO",
229
+ "conventional_resident_total_gb_est": 1.005653,
230
+ "mne_resident_total_gb_est": 1.370938,
231
+ "ssd_required_bytes": 484220288,
232
+ "feasible_conventional": "VIABLE con margen de memoria (límite superior de presión, no medición)",
233
+ "feasible_mne": "ESCENARIO VIRTUOSO DE RAM (cache capado; no prueba de throughput)",
234
+ "throughput_status": "ESTIMADO (techo de ancho de banda, no benchmark)",
235
+ "throughput_note": "No real benchmark is stored for MNE; upper bounds only."
236
+ },
237
+ {
238
+ "model": "Nova Mini 1.0",
239
+ "repository": "ModernoTech/Nova-Mini",
240
+ "hardware": "MacBook Air M4",
241
+ "hardware_id": "air-m4",
242
+ "bandwidth_gbps": 120.0,
243
+ "ram_gb": 16.0,
244
+ "context_tokens": 4096,
245
+ "context_label": "4K",
246
+ "context_supported_or_validated": true,
247
+ "max_position_documented": 40960,
248
+ "artifact_size_bytes": 484220288,
249
+ "artifact_status": "PUBLICADO",
250
+ "weight_basis": "artifact size on Hugging Face",
251
+ "active_payload_bytes_for_memory_bound": 354334802,
252
+ "active_payload_basis": "active param count x 0.5 GiB/B x 1.10 (Q4 + 10% parsed overhead), ESTIMADO",
253
+ "theoretical_tok_s_conventional_ram": 363.636,
254
+ "ssd_mne_upper_tok_s": [
255
+ {
256
+ "ssd_cache_miss": 0.1,
257
+ "cache_label": "90%",
258
+ "upper Tok_s": 60.606
259
+ },
260
+ {
261
+ "ssd_cache_miss": 0.05,
262
+ "cache_label": "95%",
263
+ "upper Tok_s": 121.212
264
+ },
265
+ {
266
+ "ssd_cache_miss": 0.01,
267
+ "cache_label": "99%",
268
+ "upper Tok_s": 606.061
269
+ },
270
+ {
271
+ "ssd_cache_miss": 0.005,
272
+ "cache_label": "99.5%",
273
+ "upper Tok_s": 1212.121
274
+ },
275
+ {
276
+ "ssd_cache_miss": 0.001,
277
+ "cache_label": "99.9%",
278
+ "upper Tok_s": 6060.606
279
+ }
280
+ ],
281
+ "kv_cache_bytes_est": 469762048,
282
+ "kv_cache_status": "ESTIMADO",
283
+ "conventional_resident_total_gb_est": 1.005653,
284
+ "mne_resident_total_gb_est": 1.370938,
285
+ "ssd_required_bytes": 484220288,
286
+ "feasible_conventional": "VIABLE con margen de memoria (límite superior de presión, no medición)",
287
+ "feasible_mne": "ESCENARIO VIRTUOSO DE RAM (cache capado; no prueba de throughput)",
288
+ "throughput_status": "ESTIMADO (techo de ancho de banda, no benchmark)",
289
+ "throughput_note": "No real benchmark is stored for MNE; upper bounds only."
290
+ },
291
+ {
292
+ "model": "Nova Mini 1.0",
293
+ "repository": "ModernoTech/Nova-Mini",
294
+ "hardware": "MacBook Air 2014",
295
+ "hardware_id": "air-2014",
296
+ "bandwidth_gbps": 25.6,
297
+ "ram_gb": 4.0,
298
+ "context_tokens": 8192,
299
+ "context_label": "8K",
300
+ "context_supported_or_validated": true,
301
+ "max_position_documented": 40960,
302
+ "artifact_size_bytes": 484220288,
303
+ "artifact_status": "PUBLICADO",
304
+ "weight_basis": "artifact size on Hugging Face",
305
+ "active_payload_bytes_for_memory_bound": 354334802,
306
+ "active_payload_basis": "active param count x 0.5 GiB/B x 1.10 (Q4 + 10% parsed overhead), ESTIMADO",
307
+ "theoretical_tok_s_conventional_ram": 77.576,
308
+ "ssd_mne_upper_tok_s": [
309
+ {
310
+ "ssd_cache_miss": 0.1,
311
+ "cache_label": "90%",
312
+ "upper Tok_s": 60.606
313
+ },
314
+ {
315
+ "ssd_cache_miss": 0.05,
316
+ "cache_label": "95%",
317
+ "upper Tok_s": 121.212
318
+ },
319
+ {
320
+ "ssd_cache_miss": 0.01,
321
+ "cache_label": "99%",
322
+ "upper Tok_s": 606.061
323
+ },
324
+ {
325
+ "ssd_cache_miss": 0.005,
326
+ "cache_label": "99.5%",
327
+ "upper Tok_s": 1212.121
328
+ },
329
+ {
330
+ "ssd_cache_miss": 0.001,
331
+ "cache_label": "99.9%",
332
+ "upper Tok_s": 6060.606
333
+ }
334
+ ],
335
+ "kv_cache_bytes_est": 939524096,
336
+ "kv_cache_status": "ESTIMADO",
337
+ "conventional_resident_total_gb_est": 1.443153,
338
+ "mne_resident_total_gb_est": 1.808438,
339
+ "ssd_required_bytes": 484220288,
340
+ "feasible_conventional": "VIABLE con margen de memoria (límite superior de presión, no medición)",
341
+ "feasible_mne": "ESCENARIO VIRTUOSO DE RAM (cache capado; no prueba de throughput)",
342
+ "throughput_status": "ESTIMADO (techo de ancho de banda, no benchmark)",
343
+ "throughput_note": "No real benchmark is stored for MNE; upper bounds only."
344
+ },
345
+ {
346
+ "model": "Nova Mini 1.0",
347
+ "repository": "ModernoTech/Nova-Mini",
348
+ "hardware": "Lenovo Yoga 720",
349
+ "hardware_id": "yoga-720",
350
+ "bandwidth_gbps": 34.1,
351
+ "ram_gb": 8.0,
352
+ "context_tokens": 8192,
353
+ "context_label": "8K",
354
+ "context_supported_or_validated": true,
355
+ "max_position_documented": 40960,
356
+ "artifact_size_bytes": 484220288,
357
+ "artifact_status": "PUBLICADO",
358
+ "weight_basis": "artifact size on Hugging Face",
359
+ "active_payload_bytes_for_memory_bound": 354334802,
360
+ "active_payload_basis": "active param count x 0.5 GiB/B x 1.10 (Q4 + 10% parsed overhead), ESTIMADO",
361
+ "theoretical_tok_s_conventional_ram": 103.333,
362
+ "ssd_mne_upper_tok_s": [
363
+ {
364
+ "ssd_cache_miss": 0.1,
365
+ "cache_label": "90%",
366
+ "upper Tok_s": 60.606
367
+ },
368
+ {
369
+ "ssd_cache_miss": 0.05,
370
+ "cache_label": "95%",
371
+ "upper Tok_s": 121.212
372
+ },
373
+ {
374
+ "ssd_cache_miss": 0.01,
375
+ "cache_label": "99%",
376
+ "upper Tok_s": 606.061
377
+ },
378
+ {
379
+ "ssd_cache_miss": 0.005,
380
+ "cache_label": "99.5%",
381
+ "upper Tok_s": 1212.121
382
+ },
383
+ {
384
+ "ssd_cache_miss": 0.001,
385
+ "cache_label": "99.9%",
386
+ "upper Tok_s": 6060.606
387
+ }
388
+ ],
389
+ "kv_cache_bytes_est": 939524096,
390
+ "kv_cache_status": "ESTIMADO",
391
+ "conventional_resident_total_gb_est": 1.443153,
392
+ "mne_resident_total_gb_est": 1.808438,
393
+ "ssd_required_bytes": 484220288,
394
+ "feasible_conventional": "VIABLE con margen de memoria (límite superior de presión, no medición)",
395
+ "feasible_mne": "ESCENARIO VIRTUOSO DE RAM (cache capado; no prueba de throughput)",
396
+ "throughput_status": "ESTIMADO (techo de ancho de banda, no benchmark)",
397
+ "throughput_note": "No real benchmark is stored for MNE; upper bounds only."
398
+ },
399
+ {
400
+ "model": "Nova Mini 1.0",
401
+ "repository": "ModernoTech/Nova-Mini",
402
+ "hardware": "MacBook Neo A18 Pro",
403
+ "hardware_id": "neo",
404
+ "bandwidth_gbps": 60.0,
405
+ "ram_gb": 8.0,
406
+ "context_tokens": 8192,
407
+ "context_label": "8K",
408
+ "context_supported_or_validated": true,
409
+ "max_position_documented": 40960,
410
+ "artifact_size_bytes": 484220288,
411
+ "artifact_status": "PUBLICADO",
412
+ "weight_basis": "artifact size on Hugging Face",
413
+ "active_payload_bytes_for_memory_bound": 354334802,
414
+ "active_payload_basis": "active param count x 0.5 GiB/B x 1.10 (Q4 + 10% parsed overhead), ESTIMADO",
415
+ "theoretical_tok_s_conventional_ram": 181.818,
416
+ "ssd_mne_upper_tok_s": [
417
+ {
418
+ "ssd_cache_miss": 0.1,
419
+ "cache_label": "90%",
420
+ "upper Tok_s": 60.606
421
+ },
422
+ {
423
+ "ssd_cache_miss": 0.05,
424
+ "cache_label": "95%",
425
+ "upper Tok_s": 121.212
426
+ },
427
+ {
428
+ "ssd_cache_miss": 0.01,
429
+ "cache_label": "99%",
430
+ "upper Tok_s": 606.061
431
+ },
432
+ {
433
+ "ssd_cache_miss": 0.005,
434
+ "cache_label": "99.5%",
435
+ "upper Tok_s": 1212.121
436
+ },
437
+ {
438
+ "ssd_cache_miss": 0.001,
439
+ "cache_label": "99.9%",
440
+ "upper Tok_s": 6060.606
441
+ }
442
+ ],
443
+ "kv_cache_bytes_est": 939524096,
444
+ "kv_cache_status": "ESTIMADO",
445
+ "conventional_resident_total_gb_est": 1.443153,
446
+ "mne_resident_total_gb_est": 1.808438,
447
+ "ssd_required_bytes": 484220288,
448
+ "feasible_conventional": "VIABLE con margen de memoria (límite superior de presión, no medición)",
449
+ "feasible_mne": "ESCENARIO VIRTUOSO DE RAM (cache capado; no prueba de throughput)",
450
+ "throughput_status": "ESTIMADO (techo de ancho de banda, no benchmark)",
451
+ "throughput_note": "No real benchmark is stored for MNE; upper bounds only."
452
+ },
453
+ {
454
+ "model": "Nova Mini 1.0",
455
+ "repository": "ModernoTech/Nova-Mini",
456
+ "hardware": "MacBook Air M4",
457
+ "hardware_id": "air-m4",
458
+ "bandwidth_gbps": 120.0,
459
+ "ram_gb": 16.0,
460
+ "context_tokens": 8192,
461
+ "context_label": "8K",
462
+ "context_supported_or_validated": true,
463
+ "max_position_documented": 40960,
464
+ "artifact_size_bytes": 484220288,
465
+ "artifact_status": "PUBLICADO",
466
+ "weight_basis": "artifact size on Hugging Face",
467
+ "active_payload_bytes_for_memory_bound": 354334802,
468
+ "active_payload_basis": "active param count x 0.5 GiB/B x 1.10 (Q4 + 10% parsed overhead), ESTIMADO",
469
+ "theoretical_tok_s_conventional_ram": 363.636,
470
+ "ssd_mne_upper_tok_s": [
471
+ {
472
+ "ssd_cache_miss": 0.1,
473
+ "cache_label": "90%",
474
+ "upper Tok_s": 60.606
475
+ },
476
+ {
477
+ "ssd_cache_miss": 0.05,
478
+ "cache_label": "95%",
479
+ "upper Tok_s": 121.212
480
+ },
481
+ {
482
+ "ssd_cache_miss": 0.01,
483
+ "cache_label": "99%",
484
+ "upper Tok_s": 606.061
485
+ },
486
+ {
487
+ "ssd_cache_miss": 0.005,
488
+ "cache_label": "99.5%",
489
+ "upper Tok_s": 1212.121
490
+ },
491
+ {
492
+ "ssd_cache_miss": 0.001,
493
+ "cache_label": "99.9%",
494
+ "upper Tok_s": 6060.606
495
+ }
496
+ ],
497
+ "kv_cache_bytes_est": 939524096,
498
+ "kv_cache_status": "ESTIMADO",
499
+ "conventional_resident_total_gb_est": 1.443153,
500
+ "mne_resident_total_gb_est": 1.808438,
501
+ "ssd_required_bytes": 484220288,
502
+ "feasible_conventional": "VIABLE con margen de memoria (límite superior de presión, no medición)",
503
+ "feasible_mne": "ESCENARIO VIRTUOSO DE RAM (cache capado; no prueba de throughput)",
504
+ "throughput_status": "ESTIMADO (techo de ancho de banda, no benchmark)",
505
+ "throughput_note": "No real benchmark is stored for MNE; upper bounds only."
506
+ }
507
+ ],
508
+ "charts": [
509
+ "tokens-per-second.png",
510
+ "ram-vs-mne.png",
511
+ "context-performance.png",
512
+ "memory-breakdown.png",
513
+ "model-size-vs-ram.png",
514
+ "cache-hit-analysis.png",
515
+ "prefill-vs-decode.png"
516
+ ]
517
+ }
performance/prefill-vs-decode.png ADDED
performance/ram-vs-mne.png ADDED
performance/simulation-results.csv ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ repository,model,hardware,ram_gb,ram_bandwidth_gbps,context,context_supported,artifact_size_bytes,active_weight_payload_bytes,token_upper_bound_conventional_tok_s,ssd_90_miss_upper,ssd_95_miss_upper,ssd_99_miss_upper,ssd_99_5_miss_upper,ssd_99_9_miss_upper,kv_cache_bytes_estimated,conventional_resident_gb_est,mne_resident_gb_est,throughput_status,ram_feasible,mne_feasible,notes
2
+ ModernoTech/Nova-Mini,Nova Mini 1.0,MacBook Air 2014,4.0,25.6,4K,true,484220288,354334802,77.576,60.606,121.212,606.061,1212.121,6060.606,469762048,1.005653,1.370938,"ESTIMADO (techo de ancho de banda, no benchmark)","VIABLE con margen de memoria (límite superior de presión, no medición)",ESCENARIO VIRTUOSO DE RAM (cache capado; no prueba de throughput),none
3
+ ModernoTech/Nova-Mini,Nova Mini 1.0,Lenovo Yoga 720,8.0,34.1,4K,true,484220288,354334802,103.333,60.606,121.212,606.061,1212.121,6060.606,469762048,1.005653,1.370938,"ESTIMADO (techo de ancho de banda, no benchmark)","VIABLE con margen de memoria (límite superior de presión, no medición)",ESCENARIO VIRTUOSO DE RAM (cache capado; no prueba de throughput),none
4
+ ModernoTech/Nova-Mini,Nova Mini 1.0,MacBook Neo A18 Pro,8.0,60.0,4K,true,484220288,354334802,181.818,60.606,121.212,606.061,1212.121,6060.606,469762048,1.005653,1.370938,"ESTIMADO (techo de ancho de banda, no benchmark)","VIABLE con margen de memoria (límite superior de presión, no medición)",ESCENARIO VIRTUOSO DE RAM (cache capado; no prueba de throughput),none
5
+ ModernoTech/Nova-Mini,Nova Mini 1.0,MacBook Air M4,16.0,120.0,4K,true,484220288,354334802,363.636,60.606,121.212,606.061,1212.121,6060.606,469762048,1.005653,1.370938,"ESTIMADO (techo de ancho de banda, no benchmark)","VIABLE con margen de memoria (límite superior de presión, no medición)",ESCENARIO VIRTUOSO DE RAM (cache capado; no prueba de throughput),none
6
+ ModernoTech/Nova-Mini,Nova Mini 1.0,MacBook Air 2014,4.0,25.6,8K,true,484220288,354334802,77.576,60.606,121.212,606.061,1212.121,6060.606,939524096,1.443153,1.808438,"ESTIMADO (techo de ancho de banda, no benchmark)","VIABLE con margen de memoria (límite superior de presión, no medición)",ESCENARIO VIRTUOSO DE RAM (cache capado; no prueba de throughput),none
7
+ ModernoTech/Nova-Mini,Nova Mini 1.0,Lenovo Yoga 720,8.0,34.1,8K,true,484220288,354334802,103.333,60.606,121.212,606.061,1212.121,6060.606,939524096,1.443153,1.808438,"ESTIMADO (techo de ancho de banda, no benchmark)","VIABLE con margen de memoria (límite superior de presión, no medición)",ESCENARIO VIRTUOSO DE RAM (cache capado; no prueba de throughput),none
8
+ ModernoTech/Nova-Mini,Nova Mini 1.0,MacBook Neo A18 Pro,8.0,60.0,8K,true,484220288,354334802,181.818,60.606,121.212,606.061,1212.121,6060.606,939524096,1.443153,1.808438,"ESTIMADO (techo de ancho de banda, no benchmark)","VIABLE con margen de memoria (límite superior de presión, no medición)",ESCENARIO VIRTUOSO DE RAM (cache capado; no prueba de throughput),none
9
+ ModernoTech/Nova-Mini,Nova Mini 1.0,MacBook Air M4,16.0,120.0,8K,true,484220288,354334802,363.636,60.606,121.212,606.061,1212.121,6060.606,939524096,1.443153,1.808438,"ESTIMADO (techo de ancho de banda, no benchmark)","VIABLE con margen de memoria (límite superior de presión, no medición)",ESCENARIO VIRTUOSO DE RAM (cache capado; no prueba de throughput),none
performance/tokens-per-second.png ADDED