alexwengg commited on
Commit
1c0b7f9
·
verified ·
1 Parent(s): c6e4872

Document ANE placement across all FP16 extraction stages

Browse files
Files changed (2) hide show
  1. README.md +2 -0
  2. reports/extraction-w8-validation.json +106 -1
README.md CHANGED
@@ -58,6 +58,8 @@ model = CoreMLAdaptiveBoundaryExtractor(".")
58
  print(model.extract("Alice founded Acme in Toronto.", Schema().entities(["person", "organization", "location"])))
59
  ```
60
 
 
 
61
  Both W8 feature packages and the five FP16 head packages are included. `extraction-w8-assets.lock.json` records exact file hashes. `quantize-extraction-coreml.py` regenerates the W8 packages from the pinned feature exports; `verify-adaptive-extraction.py` checks the selected native-output manifest. Weight-only W8 does not imply int8 activation or matrix arithmetic.
62
 
63
  ## Classification
 
58
  print(model.extract("Alice founded Acme in Toronto.", Schema().entities(["person", "organization", "location"])))
59
  ```
60
 
61
+ All six FP16 stages were profiled. The anchorless and record-assignment heads were fully assigned to ANE; the relation head was mixed, while the small scorer and explicit heads ran on CPU under this scheduler. `reports/extraction-w8-validation.json` records both the public compute-plan percentages and private executable-operation counts, which use different counting methods.
62
+
63
  Both W8 feature packages and the five FP16 head packages are included. `extraction-w8-assets.lock.json` records exact file hashes. `quantize-extraction-coreml.py` regenerates the W8 packages from the pinned feature exports; `verify-adaptive-extraction.py` checks the selected native-output manifest. Weight-only W8 does not imply int8 activation or matrix arithmetic.
64
 
65
  ## Classification
reports/extraction-w8-validation.json CHANGED
@@ -171,5 +171,110 @@
171
  "machine": "arm64",
172
  "macos": "27.0"
173
  }
174
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
175
  }
 
171
  "machine": "arm64",
172
  "macos": "27.0"
173
  }
174
+ },
175
+ "fp16_stage_profile": [
176
+ {
177
+ "package": "gliner2_multi_explicit_fp16_W64_Q8_S64.mlpackage",
178
+ "profiled_package": "uncompressed FP16 stage; W8 release changes only feature weights",
179
+ "private_executable_ops": {
180
+ "ane": 0,
181
+ "cpu": 204,
182
+ "total": 204
183
+ },
184
+ "public_compute_plan_ane_percent": 0.0,
185
+ "principal_cpu_fallbacks": [
186
+ {
187
+ "reason": "ANE supported but scheduler chose CPU",
188
+ "count": 159
189
+ },
190
+ {
191
+ "reason": "Unsupported tensor data type: int32",
192
+ "count": 32
193
+ }
194
+ ]
195
+ },
196
+ {
197
+ "package": "gliner2_multi_extraction_features_fp16_L128_W64_Q8.mlpackage",
198
+ "profiled_package": "uncompressed FP16 stage; W8 release changes only feature weights",
199
+ "private_executable_ops": {
200
+ "ane": 620,
201
+ "cpu": 28,
202
+ "total": 648
203
+ },
204
+ "public_compute_plan_ane_percent": 36.86,
205
+ "principal_cpu_fallbacks": [
206
+ {
207
+ "reason": "Unsupported tensor data type: int32",
208
+ "count": 24
209
+ },
210
+ {
211
+ "reason": "Unable to resolve operation input \"x\".",
212
+ "count": 2
213
+ }
214
+ ]
215
+ },
216
+ {
217
+ "package": "gliner2_multi_extraction_scorer_fp16_L128_W64_Q8.mlpackage",
218
+ "profiled_package": "uncompressed FP16 stage; W8 release changes only feature weights",
219
+ "private_executable_ops": {
220
+ "ane": 0,
221
+ "cpu": 137,
222
+ "total": 137
223
+ },
224
+ "public_compute_plan_ane_percent": 0.0,
225
+ "principal_cpu_fallbacks": [
226
+ {
227
+ "reason": "ANE supported but scheduler chose CPU",
228
+ "count": 90
229
+ },
230
+ {
231
+ "reason": "Unsupported tensor data type: int32",
232
+ "count": 36
233
+ }
234
+ ]
235
+ },
236
+ {
237
+ "package": "gliner2_multi_record_anchorless_fp16_F8_C192_I1536.mlpackage",
238
+ "profiled_package": "uncompressed FP16 stage; W8 release changes only feature weights",
239
+ "private_executable_ops": {
240
+ "ane": 17,
241
+ "cpu": 0,
242
+ "total": 17
243
+ },
244
+ "public_compute_plan_ane_percent": 100.0,
245
+ "principal_cpu_fallbacks": []
246
+ },
247
+ {
248
+ "package": "gliner2_multi_record_assignment_fp16_F8_C192_I1536.mlpackage",
249
+ "profiled_package": "uncompressed FP16 stage; W8 release changes only feature weights",
250
+ "private_executable_ops": {
251
+ "ane": 24,
252
+ "cpu": 0,
253
+ "total": 24
254
+ },
255
+ "public_compute_plan_ane_percent": 100.0,
256
+ "principal_cpu_fallbacks": []
257
+ },
258
+ {
259
+ "package": "gliner2_multi_relation_fp16_W64_R4_P256.mlpackage",
260
+ "profiled_package": "uncompressed FP16 stage; W8 release changes only feature weights",
261
+ "private_executable_ops": {
262
+ "ane": 49,
263
+ "cpu": 66,
264
+ "total": 115
265
+ },
266
+ "public_compute_plan_ane_percent": 93.49,
267
+ "principal_cpu_fallbacks": [
268
+ {
269
+ "reason": "Unsupported tensor data type: int32",
270
+ "count": 33
271
+ },
272
+ {
273
+ "reason": "Unable to resolve operation input \"y\".",
274
+ "count": 30
275
+ }
276
+ ]
277
+ }
278
+ ],
279
+ "profile_count_note": "Private executable-op counts and public compute-plan percentages use different counting methods; compute placement is hardware- and OS-specific."
280
  }