mlboydaisuke commited on
Commit
3b8cb71
·
verified ·
1 Parent(s): a3d78f6

Plain wording in the litertlm_manifest.json notes

Browse files
Files changed (1) hide show
  1. litertlm_manifest.json +8 -7
litertlm_manifest.json CHANGED
@@ -1,7 +1,7 @@
1
  {
2
  "manifest_schema": "0.1.3",
3
  "repo": "mlboydaisuke/S1-mini-LiteRT",
4
- "generated": "2026-10-07",
5
  "generator": "make_manifest.py",
6
  "model": {
7
  "display_name": "S1-mini (LiteRT-LM)",
@@ -70,7 +70,7 @@
70
  "platform": "android",
71
  "device_class": "flagship",
72
  "backend": "gpu",
73
- "reason": "measured on one Galaxy S26 with one binary (S7 Android backfill 2026-09-05, 205-token prompt, 2 cold runs per backend, cooled before each run): gpu decodes 25.13-25.8 tok/s against cpu's 12.29 (2.1x), prefill 940.7-1027.5 vs 166.7-175.3 tok/s, peak 1323 vs 2129 MB, engine init 6.3 vs 2.2 s. Same-device CPU control taken - a measured win for gpu on decode. On the Pixel 8a's Mali GPU this model deterministically dropped a date separator on number-dense input ('March 3rd, 2026.' -> 'March 32026.'); the same date vector run on this Adreno GPU came back intact ('The invoice came to $23,450 and it's due on March 3, 2026.', S7 probe 2026-09-05, same binary), so that Mali defect does not reproduce here. The full 12-case exactness check was not repeated on Adreno; where byte-exactness is the job, cpu remains the conservative choice"
74
  }
75
  ],
76
  "requirements": {
@@ -145,7 +145,7 @@
145
  "device": "iPhone 17 Pro",
146
  "os": "iOS",
147
  "backend": "gpu",
148
- "runtime": "litert-lm (Swift, yardstick harness)",
149
  "prompt_tokens": 180,
150
  "prefill_tps": 801,
151
  "decode_tps": 32.0,
@@ -159,7 +159,7 @@
159
  "device": "iPhone 17 Pro",
160
  "os": "iOS",
161
  "backend": "cpu",
162
- "runtime": "litert-lm (Swift, yardstick harness)",
163
  "prompt_tokens": 180,
164
  "prefill_tps": 281,
165
  "decode_tps": 15.5,
@@ -184,7 +184,7 @@
184
  "cache": "no",
185
  "runs": 2,
186
  "date": "2026-09-05",
187
- "source": "S7 Android backfill: 205-token prompt with --benchmark, 2 cold runs (compile/weight caches deleted between runs, device cooled below 42 C before each run); GPU gate full LITERT_CL delegation (15474 ops across 12 subgraphs), generation verified on the gate prompt; the cpu row of the same date is the same-device control"
188
  },
189
  {
190
  "device": "Galaxy S26 (SM-S942Q, Qualcomm SM8850, Adreno)",
@@ -200,14 +200,15 @@
200
  "cache": "no",
201
  "runs": 2,
202
  "date": "2026-09-05",
203
- "source": "S7 Android backfill: same-device CPU control - identical binary, prompt file and flags, only --backend differs; 205-token prompt with --benchmark, 2 cold runs (compile/weight caches deleted between runs, device cooled below 42 C before each run); same-device control for the gpu row of the same date"
204
  }
205
  ],
206
  "known_issues": [
207
  "Pixel 8a GPU: on number-dense input a date separator can be lost ('March 3rd, 2026.' -> 'March 32026.'), deterministic on repeat runs; the CPU backend on the same phone is byte-identical to macOS across all 12 test cases. Prefer CPU on this phone when exactness matters, GPU when time-to-first-token matters.",
208
  "The released checkpoint itself drifts from its own upstream card's example table (it keeps 'Hmm,' / 'So' discourse markers the card drops); parity here is stated against HF fp32, not against those published examples.",
209
  "All iPhone 17 Pro rows were measured with the device reporting a 'serious' thermal state; they are floors, not peaks."
210
- ]
 
211
  }
212
  ]
213
  }
 
1
  {
2
  "manifest_schema": "0.1.3",
3
  "repo": "mlboydaisuke/S1-mini-LiteRT",
4
+ "generated": "2026-11-14",
5
  "generator": "make_manifest.py",
6
  "model": {
7
  "display_name": "S1-mini (LiteRT-LM)",
 
70
  "platform": "android",
71
  "device_class": "flagship",
72
  "backend": "gpu",
73
+ "reason": "measured on one Galaxy S26 with one binary (2026-09-05, 205-token prompt, 2 cold runs per backend, cooled before each run): gpu decodes 25.13-25.8 tok/s against cpu's 12.29 (2.1x), prefill 940.7-1027.5 vs 166.7-175.3 tok/s, peak 1323 vs 2129 MB, engine init 6.3 vs 2.2 s. Same-device CPU control taken - a measured win for gpu on decode. On the Pixel 8a's Mali GPU this model deterministically dropped a date separator on number-dense input ('March 3rd, 2026.' -> 'March 32026.'); the same date vector run on this Adreno GPU came back intact ('The invoice came to $23,450 and it's due on March 3, 2026.', probe 2026-09-05, same binary), so that Mali defect does not reproduce here. The full 12-case exactness check was not repeated on Adreno; where byte-exactness is the job, cpu remains the conservative choice"
74
  }
75
  ],
76
  "requirements": {
 
145
  "device": "iPhone 17 Pro",
146
  "os": "iOS",
147
  "backend": "gpu",
148
+ "runtime": "litert-lm (Swift)",
149
  "prompt_tokens": 180,
150
  "prefill_tps": 801,
151
  "decode_tps": 32.0,
 
159
  "device": "iPhone 17 Pro",
160
  "os": "iOS",
161
  "backend": "cpu",
162
+ "runtime": "litert-lm (Swift)",
163
  "prompt_tokens": 180,
164
  "prefill_tps": 281,
165
  "decode_tps": 15.5,
 
184
  "cache": "no",
185
  "runs": 2,
186
  "date": "2026-09-05",
187
+ "source": "205-token prompt with --benchmark, 2 cold runs (compile/weight caches deleted between runs, device cooled below 42 C before each run); GPU gate full LITERT_CL delegation (15474 ops across 12 subgraphs), generation verified on the gate prompt; the cpu row of the same date is the same-device control"
188
  },
189
  {
190
  "device": "Galaxy S26 (SM-S942Q, Qualcomm SM8850, Adreno)",
 
200
  "cache": "no",
201
  "runs": 2,
202
  "date": "2026-09-05",
203
+ "source": "same-device CPU control - identical binary, prompt file and flags, only --backend differs; 205-token prompt with --benchmark, 2 cold runs (compile/weight caches deleted between runs, device cooled below 42 C before each run); same-device control for the gpu row of the same date"
204
  }
205
  ],
206
  "known_issues": [
207
  "Pixel 8a GPU: on number-dense input a date separator can be lost ('March 3rd, 2026.' -> 'March 32026.'), deterministic on repeat runs; the CPU backend on the same phone is byte-identical to macOS across all 12 test cases. Prefer CPU on this phone when exactness matters, GPU when time-to-first-token matters.",
208
  "The released checkpoint itself drifts from its own upstream card's example table (it keeps 'Hmm,' / 'So' discourse markers the card drops); parity here is stated against HF fp32, not against those published examples.",
209
  "All iPhone 17 Pro rows were measured with the device reporting a 'serious' thermal state; they are floors, not peaks."
210
+ ],
211
+ "context_length": 4096
212
  }
213
  ]
214
  }