Upload 25 files

iteration_3/README.md

@@ -0,0 +1,115 @@
+# StyleTTS2 → CoreML iteration_3
+
+Mixed-precision build on top of iteration_2: 7 stages flipped to fp16
+weight precision, 1 stage kept at fp32 to avoid an audible-quality
+regression. Disk halved, pipeline-stage sum cut 24–41 % cool.
+
+## Pipeline (8 stages, 8 dispatches)
+
+```
+text_encoder           → CPU_ONLY      fp16   11 MB
+bert                   → ALL           fp16   12 MB
+ref_encoder            → CPU_AND_GPU   fp16   53 MB
+fused_diffusion_sampler → ALL          fp16   47 MB   ← Trial 4
+duration_predictor     → CPU_ONLY      fp16   15 MB
+fused_f0n_har_source   → CPU_ONLY      fp32   32 MB   ← Trial 6 (kept fp32: cumsum drift)
+decoder_pre            → CPU_AND_NE    fp16   64 MB
+decoder_upsample       → CPU_ONLY      fp16   40 MB
+```
+
+Total: **274 MB**, 8 mlpackages, 8 dispatches per utterance.
+
+## Performance
+
+Warm pipeline-stage sum (sum of per-stage timings reported by
+`coreml.inference`), 3-iter sweep with 8 s cooldown, M-series Mac:
+
+| Build           | min  | avg  | max   |
+|-----------------|------|------|-------|
+| iteration_2 fp32| 782  | 898  | 1075  |
+| iteration_3     | **460** | **683** | 1110 (thermal) |
+
+Cool-run delta: **−322 ms (−41 %)** at min, **−215 ms (−24 %)** at avg.
+The max bucket bunches because pipeline-wide variance dominates any
+config — same pattern observed in Trial 8b benches.
+
+Per-stage savings observed end-to-end:
+
+| stage                   | fp32 ms | fp16 ms | Δ        |
+|-------------------------|---------|---------|----------|
+| fused_diffusion_sampler | 18.3    | 14.7    | −3.6 ms  |
+| decoder_pre             | 35      | 7       | −28 ms   |
+| decoder_upsample        | 593–638 | 284–325 | **−309 ms** |
+
+## Mixed precision rationale
+
+| Stage                   | fp16 verdict        | Why                                     |
+|-------------------------|---------------------|-----------------------------------------|
+| text_encoder            | adopt               | clean A/B                               |
+| bert                    | adopt               | clean A/B                               |
+| ref_encoder             | adopt               | clean A/B                               |
+| fused_diffusion_sampler | adopt               | parity 4.66e-3, A/B clean                |
+| duration_predictor      | adopt               | clean A/B                               |
+| fused_f0n_har_source    | **drop**            | har computes sin(2π·cumsum(f0)) over 88 200 samples; fp16 cumsum drifts ~10 bits, audible phase distortion in second half |
+| decoder_pre             | adopt               | parity tight, A/B clean                  |
+| decoder_upsample        | adopt               | A/B clean; previously feared "+240 ms" regression on `ALL` did not reproduce on `CPU_ONLY` placement (this is the 8b-winning placement) |
+
+Drift evidence comes from per-stage CoreML parity vs eager fp32 plus
+direct A/B listening of three configurations:
+
+```
+sanity_fp16_mixed.wav         (5 fp16 / 3 fp32)   — clean
+sanity_fp16_plus_decpre.wav   (6 fp16 / 2 fp32)   — clean
+sanity_fp16_plus_decup.wav    (7 fp16 / 1 fp32)   — clean   ← this build
+sanity_fp16_plus_f0n.wav      (8 fp16)            — degraded second half
+```
+
+## Storage
+
+| Artifact                             | iteration_2 | iteration_3 |
+|--------------------------------------|-------------|-------------|
+| Total                                | 514 MB      | **274 MB** (−47 %) |
+| largest stage                        | decoder_pre 128 MB | decoder_pre 64 MB |
+| smallest stage                       | text_encoder 21 MB | text_encoder 11 MB |
+
+## Usage
+
+Same wiring as iteration_2 — `_STAGE_PRECISION` in `coreml/inference.py`
+selects fp16 / fp32 per stage. No code changes, only the manifest values
+flip:
+
+```python
+_STAGE_PRECISION: dict[str, str] = {
+    "text_encoder":             "fp16",
+    "bert":                     "fp16",
+    "ref_encoder":              "fp16",
+    "fused_diffusion_sampler":  "fp16",
+    "diffusion_unet":           "fp32",  # legacy fallback
+    "duration_predictor":       "fp16",
+    "fused_f0n_har_source":     "fp32",  # cumsum drift
+    "f0n_predictor":            "fp32",  # legacy fallback
+    "har_source":               "fp32",  # legacy fallback
+    "decoder_pre":              "fp16",
+    "decoder_upsample":         "fp16",
+}
+```
+
+CLI overrides still work:
+
+```bash
+# Re-run any stage at fp32 to A/B
+python -m coreml.inference --fp32 decoder_upsample
+
+# Drop back to iteration_2 wholesale
+python -m coreml.inference --fp32
+```
+
+## Skipped trials this iteration
+
+| Stage                    | Reason for staying fp32                              |
+|--------------------------|------------------------------------------------------|
+| fused_f0n_har_source     | har_source cumsum drift over 88 200-sample window    |
+
+Other quantization tiers (int8 weight-only, int4 palettization) deferred
+to a future iteration — fp16 already pays for itself on disk and warm
+latency.

iteration_3/packages/bert_fp16.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:92c06d63856f46e8788c54fb2f2e7228d7da9798e2192c3078fb96a5f1de4074
+size 85458

iteration_3/packages/bert_fp16.mlpackage/Data/com.apple.CoreML/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fc4a9fb3870729f9572b0830993351524b04b99eba6cab982cef2a17507d9ba0
+size 12090496

iteration_3/packages/bert_fp16.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "514C9E67-3E15-43D6-AE2B-6179B9113D2E": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "BED7A6A1-56C6-4FB3-AB4B-06ADAD7C844E": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "BED7A6A1-56C6-4FB3-AB4B-06ADAD7C844E"
+}

iteration_3/packages/decoder_pre_fp16.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:753dbab37d3232a69b52d48f5d0732632e9307d388ed5224736e9c585db6029c
+size 55933

iteration_3/packages/decoder_pre_fp16.mlpackage/Data/com.apple.CoreML/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:db81849a38ce1959ea345219332051947f22f00dc2445cb9b7a119673ca4bf93
+size 67190976

iteration_3/packages/decoder_pre_fp16.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "211DC47B-E839-4B47-B64D-EE04F9C081B9": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        },
+        "BE7D7840-FCB4-4491-B2ED-0D81B5FD33AA": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        }
+    },
+    "rootModelIdentifier": "211DC47B-E839-4B47-B64D-EE04F9C081B9"
+}

iteration_3/packages/decoder_upsample_fp16.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ff29829e3c92a4208ef07d307293fd576c4484c6048e519b90cd32ee80180038
+size 491796

iteration_3/packages/decoder_upsample_fp16.mlpackage/Data/com.apple.CoreML/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:43161151f001bb951c34952465adfc3c4f5fb8ab2845f31903be09ea9f1a6bc5
+size 41400320

iteration_3/packages/decoder_upsample_fp16.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "26E8FCA8-BD9B-4185-B59E-00453487B2B3": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        },
+        "BFFB197D-D576-4F27-85E5-48F5438F08C2": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        }
+    },
+    "rootModelIdentifier": "26E8FCA8-BD9B-4185-B59E-00453487B2B3"
+}

iteration_3/packages/duration_predictor_fp16.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9b60eccf1aff0c09069d4eeebb5611c11caee89788229d0780ef606ac8fa1384
+size 29886

iteration_3/packages/duration_predictor_fp16.mlpackage/Data/com.apple.CoreML/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:75ba0b7b2f7dc6a687e9ec01d226c300b09f07832d8e4aac2705a16b5079910c
+size 15543524

iteration_3/packages/duration_predictor_fp16.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "EA4FC14C-8DE2-414B-A6C4-B93190F89ED0": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "EAAA83DE-C745-4884-AE8D-1ED5C06BC490": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "EAAA83DE-C745-4884-AE8D-1ED5C06BC490"
+}

iteration_3/packages/fused_diffusion_sampler_fp16.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f7b29b69d4fcddc08e802c02e60db2a8d515669b6b8cec7ad18f5a5e3a5054df
+size 311675

iteration_3/packages/fused_diffusion_sampler_fp16.mlpackage/Data/com.apple.CoreML/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f823b5c638d2eb2fd91bf8e4efe4a90b2e1d3d9e2f5ab40e7e93cb03cd212aca
+size 49361856

iteration_3/packages/fused_diffusion_sampler_fp16.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "2513097D-94F1-4E53-A221-C56C469859E3": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "A740A1E5-A518-4F32-B0B3-878148DC3920": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "A740A1E5-A518-4F32-B0B3-878148DC3920"
+}

iteration_3/packages/fused_f0n_har_source.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c98d56382c336dbea5a864e594022c11ced07372687dc66e56c7c91795e45c63
+size 61645

iteration_3/packages/fused_f0n_har_source.mlpackage/Data/com.apple.CoreML/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ce01b0523d8e925108b5f9ba113d5ff1721bb3e5e25f3eb6d8a2dfaa56876c59
+size 33640448

iteration_3/packages/fused_f0n_har_source.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "696D6238-15F9-4A4A-AB5B-3957A847EB37": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "8111F480-F3EF-48DE-91B8-5309A11BF779": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "8111F480-F3EF-48DE-91B8-5309A11BF779"
+}

iteration_3/packages/ref_encoder_fp16.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5cbd0cf223b874ed6b2de35606a5690bc6355b4890ea32ec30119db5dc00497e
+size 68843

iteration_3/packages/ref_encoder_fp16.mlpackage/Data/com.apple.CoreML/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:011d14fdb46589dfb79efb619d63846430be4e4ac86372f8819f35f5e0157391
+size 55386048

iteration_3/packages/ref_encoder_fp16.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "32FE6195-7355-4635-AECB-58D9F49F1E17": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "343F4722-A338-4705-8547-09E9A93DE8EC": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "343F4722-A338-4705-8547-09E9A93DE8EC"
+}

iteration_3/packages/text_encoder_fp16.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d88b74cb84892f7ff1e4d013517dd3d4dab56688b0a0fb4d920f72d0caf9e961
+size 16587

iteration_3/packages/text_encoder_fp16.mlpackage/Data/com.apple.CoreML/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0d7f6e5869bb9d523956183e0facdff160c301d28113290efa329ae7bf72d3ce
+size 11208000

iteration_3/packages/text_encoder_fp16.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "7F3243AB-2AFC-40E5-A6DE-069619301D63": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        },
+        "F67A2205-52AD-4B8E-A19F-A7FB9AEB48F9": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        }
+    },
+    "rootModelIdentifier": "7F3243AB-2AFC-40E5-A6DE-069619301D63"
+}