Add 14ch EPOC X Core ML profile (fp16) + parity artifacts

.gitattributes

@@ -4,5 +4,17 @@
 profiles/14ch/fp16/ZunaDecoderStep.mlpackage/Data/com.apple.CoreML/model.mlmodel filter=lfs diff=lfs merge=lfs -text
 profiles/14ch/fp16/ZunaDecoderStepUpdate.mlpackage/Data/com.apple.CoreML/model.mlmodel filter=lfs diff=lfs merge=lfs -text
 profiles/14ch/fp16/ZunaEncoder.mlpackage/Data/com.apple.CoreML/model.mlmodel filter=lfs diff=lfs merge=lfs -text
+profiles/16ch/fp16/ZunaDecoderStep.mlpackage/Data/com.apple.CoreML/model.mlmodel filter=lfs diff=lfs merge=lfs -text
+profiles/16ch/fp16/ZunaDecoderStepUpdate.mlpackage/Data/com.apple.CoreML/model.mlmodel filter=lfs diff=lfs merge=lfs -text
+profiles/16ch/fp16/ZunaEncoder.mlpackage/Data/com.apple.CoreML/model.mlmodel filter=lfs diff=lfs merge=lfs -text
+profiles/32ch/fp16/ZunaDecoderStep.mlpackage/Data/com.apple.CoreML/model.mlmodel filter=lfs diff=lfs merge=lfs -text
+profiles/32ch/fp16/ZunaDecoderStepUpdate.mlpackage/Data/com.apple.CoreML/model.mlmodel filter=lfs diff=lfs merge=lfs -text
+profiles/32ch/fp16/ZunaEncoder.mlpackage/Data/com.apple.CoreML/model.mlmodel filter=lfs diff=lfs merge=lfs -text
+profiles/64ch/fp16/ZunaDecoderStep.mlpackage/Data/com.apple.CoreML/model.mlmodel filter=lfs diff=lfs merge=lfs -text
+profiles/64ch/fp16/ZunaDecoderStepUpdate.mlpackage/Data/com.apple.CoreML/model.mlmodel filter=lfs diff=lfs merge=lfs -text
+profiles/64ch/fp16/ZunaEncoder.mlpackage/Data/com.apple.CoreML/model.mlmodel filter=lfs diff=lfs merge=lfs -text
+profiles/64ch/fp32/ZunaDecoderStep.mlpackage/Data/com.apple.CoreML/model.mlmodel filter=lfs diff=lfs merge=lfs -text
+profiles/64ch/fp32/ZunaDecoderStepUpdate.mlpackage/Data/com.apple.CoreML/model.mlmodel filter=lfs diff=lfs merge=lfs -text
+profiles/64ch/fp32/ZunaEncoder.mlpackage/Data/com.apple.CoreML/model.mlmodel filter=lfs diff=lfs merge=lfs -text
 profiles/validation/parity_summary.png filter=lfs diff=lfs merge=lfs -text
 profiles/validation/waveform_overlay_residual.png filter=lfs diff=lfs merge=lfs -text

README.md

@@ -61,6 +61,10 @@ These assumptions are what the released pretrained weights were optimized for.
 Profile artifacts are organized as:
 
 - `profiles/14ch/fp16/...`
+- `profiles/16ch/fp16/...`
+- `profiles/32ch/fp16/...`
+- `profiles/64ch/fp16/...`
+- `profiles/64ch/fp32/...`
 
 Each profile contains:
 
@@ -82,6 +86,10 @@ Use `DecoderStepUpdate` when you want a minimal host-side loop and fewer host te
 | Profile | Channels | Precision | Token Count | Encoder/Decoder Tensor Shape | final-z rel_l2 vs PyTorch |
 |---|---:|---|---:|---|---:|
 | `14ch-fp16` | 14 | `fp16` | 560 | `[1, 560, 32]` | 0.011380 |
+| `16ch-fp16` | 16 | `fp16` | 640 | `[1, 640, 32]` | 0.006580 |
+| `32ch-fp16` | 32 | `fp16` | 1280 | `[1, 1280, 32]` | 0.005629 |
+| `64ch-fp16` | 64 | `fp16` | 2560 | `[1, 2560, 32]` | 0.004366 |
+| `64ch-fp32` | 64 | `fp32` | 2560 | `[1, 2560, 32]` | 0.000002 |
 
 See `profiles/index.json` for machine-readable profile discovery.
 
@@ -92,6 +100,10 @@ All published profiles are checked against the original PyTorch weights using a
 | Profile | MAE | RMSE | max_abs | rel_l2 | Threshold | Gate |
 |---|---:|---:|---:|---:|---:|---|
 | `14ch-fp16` | 0.006258 | 0.010665 | 0.193008 | 0.011380 | 0.012000 | **PASS** |
+| `16ch-fp16` | 0.004843 | 0.006150 | 0.057458 | 0.006580 | 0.010000 | **PASS** |
+| `32ch-fp16` | 0.004189 | 0.005253 | 0.020710 | 0.005629 | 0.010000 | **PASS** |
+| `64ch-fp16` | 0.003265 | 0.004077 | 0.018174 | 0.004366 | 0.010000 | **PASS** |
+| `64ch-fp32` | 0.000001 | 0.000001 | 0.000011 | 0.000002 | 0.005000 | **PASS** |
 
 ### Parity Visualization
 

profiles/index.json

@@ -1,6 +1,4 @@
 {
-  "format_version": 1,
-  "layout": "profiles",
   "profiles": [
     {
       "profile_id": "14ch-fp16",
@@ -16,6 +14,66 @@
         "decoder_step_coreml": "ZunaDecoderStep.mlpackage",
         "decoder_step_update_coreml": "ZunaDecoderStepUpdate.mlpackage"
       }
+    },
+    {
+      "profile_id": "16ch-fp16",
+      "channel_count": 16,
+      "precision": "fp16",
+      "token_count": 640,
+      "profile_dir": "16ch/fp16",
+      "artifacts": {
+        "encoder_torchscript": "ZunaEncoder.pt",
+        "decoder_step_torchscript": "ZunaDecoderStep.pt",
+        "decoder_step_update_torchscript": "ZunaDecoderStepUpdate.pt",
+        "encoder_coreml": "ZunaEncoder.mlpackage",
+        "decoder_step_coreml": "ZunaDecoderStep.mlpackage",
+        "decoder_step_update_coreml": "ZunaDecoderStepUpdate.mlpackage"
+      }
+    },
+    {
+      "profile_id": "32ch-fp16",
+      "channel_count": 32,
+      "precision": "fp16",
+      "token_count": 1280,
+      "profile_dir": "32ch/fp16",
+      "artifacts": {
+        "encoder_torchscript": "ZunaEncoder.pt",
+        "decoder_step_torchscript": "ZunaDecoderStep.pt",
+        "decoder_step_update_torchscript": "ZunaDecoderStepUpdate.pt",
+        "encoder_coreml": "ZunaEncoder.mlpackage",
+        "decoder_step_coreml": "ZunaDecoderStep.mlpackage",
+        "decoder_step_update_coreml": "ZunaDecoderStepUpdate.mlpackage"
+      }
+    },
+    {
+      "profile_id": "64ch-fp16",
+      "channel_count": 64,
+      "precision": "fp16",
+      "token_count": 2560,
+      "profile_dir": "64ch/fp16",
+      "artifacts": {
+        "encoder_torchscript": "ZunaEncoder.pt",
+        "decoder_step_torchscript": "ZunaDecoderStep.pt",
+        "decoder_step_update_torchscript": "ZunaDecoderStepUpdate.pt",
+        "encoder_coreml": "ZunaEncoder.mlpackage",
+        "decoder_step_coreml": "ZunaDecoderStep.mlpackage",
+        "decoder_step_update_coreml": "ZunaDecoderStepUpdate.mlpackage"
+      }
+    },
+    {
+      "profile_id": "64ch-fp32",
+      "channel_count": 64,
+      "precision": "fp32",
+      "token_count": 2560,
+      "profile_dir": "64ch/fp32",
+      "artifacts": {
+        "encoder_torchscript": "ZunaEncoder.pt",
+        "decoder_step_torchscript": "ZunaDecoderStep.pt",
+        "decoder_step_update_torchscript": "ZunaDecoderStepUpdate.pt",
+        "encoder_coreml": "ZunaEncoder.mlpackage",
+        "decoder_step_coreml": "ZunaDecoderStep.mlpackage",
+        "decoder_step_update_coreml": "ZunaDecoderStepUpdate.mlpackage"
+      }
     }
   ]
-}
+}

profiles/parity_report.json

@@ -1,8 +1,6 @@
 {
-  "thresholds": {
-    "fp16_rel_l2": 0.012,
-    "fp32_rel_l2": 0.005
-  },
+  "all_passed": true,
+  "failed_profiles": [],
   "reports": [
     {
       "profile_id": "14ch-fp16",
@@ -36,9 +34,138 @@
       },
       "threshold_rel_l2": 0.012,
       "pass": true
+    },
+    {
+      "profile_id": "16ch-fp16",
+      "channel_count": 16,
+      "precision": "fp16",
+      "token_count": 640,
+      "sample_steps": 20,
+      "encoder": {
+        "mae": 0.006940255872905254,
+        "rmse": 0.008843358606100082,
+        "max_abs": 0.03045654296875,
+        "rel_l2": 0.01402380783110857
+      },
+      "final_z_step_loop": {
+        "mae": 0.0048432364128530025,
+        "rmse": 0.006149778142571449,
+        "max_abs": 0.05745832622051239,
+        "rel_l2": 0.006579744163900614
+      },
+      "final_z_update_loop": {
+        "mae": 0.0049058133736252785,
+        "rmse": 0.006234212778508663,
+        "max_abs": 0.057101115584373474,
+        "rel_l2": 0.006670081987977028
+      },
+      "coreml_consistency": {
+        "mae": 0.0006364885484799743,
+        "rmse": 0.0010369644733145833,
+        "max_abs": 0.009854316711425781,
+        "rel_l2": 0.001108274213038385
+      },
+      "threshold_rel_l2": 0.01,
+      "pass": true
+    },
+    {
+      "profile_id": "32ch-fp16",
+      "channel_count": 32,
+      "precision": "fp16",
+      "token_count": 1280,
+      "sample_steps": 20,
+      "encoder": {
+        "mae": 0.0068974122405052185,
+        "rmse": 0.00882513914257288,
+        "max_abs": 0.03042382001876831,
+        "rel_l2": 0.014106319285929203
+      },
+      "final_z_step_loop": {
+        "mae": 0.004189030732959509,
+        "rmse": 0.005252954084426165,
+        "max_abs": 0.020709753036499023,
+        "rel_l2": 0.0056286221370100975
+      },
+      "final_z_update_loop": {
+        "mae": 0.004258580505847931,
+        "rmse": 0.005357986781746149,
+        "max_abs": 0.025191545486450195,
+        "rel_l2": 0.005741165950894356
+      },
+      "coreml_consistency": {
+        "mae": 0.0006323954439722002,
+        "rmse": 0.0010173922637477517,
+        "max_abs": 0.014103889465332031,
+        "rel_l2": 0.001088618068024516
+      },
+      "threshold_rel_l2": 0.01,
+      "pass": true
+    },
+    {
+      "profile_id": "64ch-fp16",
+      "channel_count": 64,
+      "precision": "fp16",
+      "token_count": 2560,
+      "sample_steps": 20,
+      "encoder": {
+        "mae": 0.00757429888471961,
+        "rmse": 0.009730321355164051,
+        "max_abs": 0.03688855469226837,
+        "rel_l2": 0.015860911458730698
+      },
+      "final_z_step_loop": {
+        "mae": 0.003264558967202902,
+        "rmse": 0.004077223129570484,
+        "max_abs": 0.018174409866333008,
+        "rel_l2": 0.004365743603557348
+      },
+      "final_z_update_loop": {
+        "mae": 0.0033559747971594334,
+        "rmse": 0.004213370848447084,
+        "max_abs": 0.02053523063659668,
+        "rel_l2": 0.004511525854468346
+      },
+      "coreml_consistency": {
+        "mae": 0.0006285925628617406,
+        "rmse": 0.0010274079395458102,
+        "max_abs": 0.012227773666381836,
+        "rel_l2": 0.0010984123218804598
+      },
+      "threshold_rel_l2": 0.01,
+      "pass": true
+    },
+    {
+      "profile_id": "64ch-fp32",
+      "channel_count": 64,
+      "precision": "fp32",
+      "token_count": 2560,
+      "sample_steps": 20,
+      "encoder": {
+        "mae": 3.0412613796215737e-06,
+        "rmse": 4.264039034751477e-06,
+        "max_abs": 2.333521842956543e-05,
+        "rel_l2": 6.950593615329126e-06
+      },
+      "final_z_step_loop": {
+        "mae": 1.130579448727076e-06,
+        "rmse": 1.4336205822473858e-06,
+        "max_abs": 1.0848045349121094e-05,
+        "rel_l2": 1.5350690318882698e-06
+      },
+      "final_z_update_loop": {
+        "mae": 1.130579448727076e-06,
+        "rmse": 1.4336205822473858e-06,
+        "max_abs": 1.0848045349121094e-05,
+        "rel_l2": 1.5350690318882698e-06
+      },
+      "coreml_consistency": {
+        "mae": 0.0,
+        "rmse": 0.0,
+        "max_abs": 0.0,
+        "rel_l2": 0.0
+      },
+      "threshold_rel_l2": 0.005,
+      "pass": true
     }
-  ],
-  "all_passed": true,
-  "failed_profiles": [],
-  "waveform_preview_file": "parity_waveform_preview.json"
-}
+  ]
+}