Test builds: Neural Engine-compilable conv-cache slices for M2 (FluidAudio #1141)

test-1141/README.md

@@ -0,0 +1,30 @@
+# Test builds for FluidAudio #1141 (not for production)
+
+https://github.com/FluidInference/FluidAudio/issues/1141
+
+On M2-generation Neural Engines, the compiler refuses the shipped multilingual encoders. The trigger is the conv-cache
+update in each Conformer layer, `concat(cache_time[l], x)[:, :, T:T+8]`, and Core ML then runs the whole encoder on the CPU.
+
+Each model here is the shipped `.mlmodelc` from `main` with only those slices rewritten. Each slice of the concat becomes
+a slice of the concat's inputs: `x[:, :, T-8:T]` for 1120/2240 ms, and `concat(cache[:, :, 7:], x)` for 560 ms.
+`weight.bin` is byte-identical to `main`. `patch_mlmodelc.py` is the script that made them
+(`python3 patch_mlmodelc.py <shipped.mlmodelc> <out.mlmodelc>`).
+
+| Path | Replaces |
+|---|---|
+| `multilingual/560ms/encoder.mlmodelc` | `multilingual/560ms/encoder.mlmodelc` (24 slices) |
+| `multilingual/1120ms/encoder.mlmodelc` | `multilingual/1120ms/encoder.mlmodelc` (24 slices) |
+| `multilingual/2240ms/encoder.mlmodelc` | `multilingual/2240ms/encoder.mlmodelc` (24 slices) |
+| `encoder-split/2240ms/encoder_body_0_11.mlmodelc` | same path on `main` (12 slices) |
+| `encoder-split/2240ms/encoder_body_12_23.mlmodelc` | same path on `main` (12 slices) |
+
+The split front end has no such slice; use `encoder-split/2240ms/encoder_frontend.mlmodelc` from `main`.
+Inputs, outputs and the rest of each tier folder (preprocessor, decoder, joint, tokenizer, metadata) are unchanged.
+
+Checked on M5 Pro, macOS 27.0, over 86 s of FLEURS en+es streamed with the caches fed back:
+
+- Outputs are byte-identical to `main` on both `.cpuAndNeuralEngine` and `.cpuOnly`, for all tiers and the split chain.
+- Latency: 560 ms goes from 8.1 to 8.5 ms per step (one extra concat per layer); 1120 ms, 2240 ms and the split are unchanged.
+- Load and memory footprint are unchanged.
+
+The Neural Engine compile has not been checked on M1, M2 or M3 yet. That is what these builds are for.

test-1141/encoder-split/2240ms/encoder_body_0_11.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c6a9212467d1fead00b74855bcbea0a197eead90e7f4100816c04aaad69e9c11
+size 243

test-1141/encoder-split/2240ms/encoder_body_0_11.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:780807c03fe6b8f858e6d20450e5ca411a85d4b4ddd57be32a2d52ff1948711c
+size 651

test-1141/encoder-split/2240ms/encoder_body_0_11.mlmodelc/metadata.json

@@ -0,0 +1,147 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "shortDescription" : "Nemotron-3.5 ASR streaming multilingual 2240ms encoder [body_0_11], LAYERPOS [42,13]",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 28 × 1024)",
+        "shortDescription" : "",
+        "shape" : "[1, 28, 1024]",
+        "name" : "hidden",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 12 × 42 × 1024)",
+        "shortDescription" : "",
+        "shape" : "[1, 12, 42, 1024]",
+        "name" : "cache_channel_out",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 12 × 1024 × 8)",
+        "shortDescription" : "",
+        "shape" : "[1, 12, 1024, 8]",
+        "name" : "cache_time_out",
+        "type" : "MultiArray"
+      }
+    ],
+    "storagePrecision" : "Mixed (Float16, Int8, Palettized (6 bits))",
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios17.reshape" : 72,
+      "Ios17.logicalAnd" : 3,
+      "Ios16.softmax" : 12,
+      "Ios17.matmul" : 36,
+      "Ios17.transpose" : 113,
+      "Split" : 12,
+      "Select" : 36,
+      "Ios17.expandDims" : 5,
+      "Ios17.add" : 86,
+      "Tile" : 1,
+      "Ios17.sliceByIndex" : 74,
+      "Ios16.sigmoid" : 12,
+      "Pad" : 12,
+      "Ios17.logicalNot" : 2,
+      "Ios17.layerNorm" : 72,
+      "Ios16.constexprAffineDequantize" : 69,
+      "Ios17.less" : 1,
+      "Ios17.cast" : 6,
+      "Ios16.silu" : 36,
+      "Ios17.conv" : 36,
+      "Ios16.constexprLutToDense" : 75,
+      "Ios17.linear" : 96,
+      "Ios17.concat" : 36,
+      "Ios17.greaterEqual" : 1,
+      "Stack" : 2,
+      "Ios17.mul" : 49
+    },
+    "computePrecision" : "Mixed (Float16, Float32, Int32)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 28 × 1024)",
+        "shortDescription" : "",
+        "shape" : "[1, 28, 1024]",
+        "name" : "pre_encoded",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "pre_encoded_length",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 12 × 42 × 1024)",
+        "shortDescription" : "",
+        "shape" : "[1, 12, 42, 1024]",
+        "name" : "cache_channel",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 12 × 1024 × 8)",
+        "shortDescription" : "",
+        "shape" : "[1, 12, 1024, 8]",
+        "name" : "cache_time",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "cache_len",
+        "type" : "MultiArray"
+      }
+    ],
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.conversion_date" : "2026-11-08",
+      "com.github.apple.coremltools.source" : "torch==2.7.0",
+      "com.github.apple.coremltools.version" : "9.0",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "generatedClassName" : "body_0_11",
+    "method" : "predict"
+  }
+]

test-1141/encoder-split/2240ms/encoder_body_0_11.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:48d03ce9838198de9d7929f9cfac6bee7206b7b868dbd6538a1632352d90d4c2
+size 242596992

test-1141/encoder-split/2240ms/encoder_body_12_23.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4573c654d466a4a77343d78b389a45e084a83f448f33936a3501d2ac7046fea4
+size 243

test-1141/encoder-split/2240ms/encoder_body_12_23.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e6bc3f11214d1c2d912180b3050071952e10a2784934ae9804c7efd9568e8d5b
+size 729

test-1141/encoder-split/2240ms/encoder_body_12_23.mlmodelc/metadata.json

@@ -0,0 +1,181 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "shortDescription" : "Nemotron-3.5 ASR streaming multilingual 2240ms encoder [body_12_23], LAYERPOS [42,13]",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1024 × 28)",
+        "shortDescription" : "",
+        "shape" : "[1, 1024, 28]",
+        "name" : "encoded",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "encoded_length",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 12 × 42 × 1024)",
+        "shortDescription" : "",
+        "shape" : "[1, 12, 42, 1024]",
+        "name" : "cache_channel_out",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 12 × 1024 × 8)",
+        "shortDescription" : "",
+        "shape" : "[1, 12, 1024, 8]",
+        "name" : "cache_time_out",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "cache_len_out",
+        "type" : "MultiArray"
+      }
+    ],
+    "storagePrecision" : "Mixed (Float16, Int8, Palettized (6 bits))",
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios17.reshape" : 72,
+      "Ios17.logicalAnd" : 3,
+      "Ios16.softmax" : 12,
+      "Identity" : 1,
+      "Ios17.matmul" : 36,
+      "Ios17.transpose" : 114,
+      "Split" : 12,
+      "Ios17.expandDims" : 6,
+      "Select" : 39,
+      "Ios17.add" : 90,
+      "Tile" : 2,
+      "Ios17.sliceByIndex" : 74,
+      "Ios16.sigmoid" : 12,
+      "Pad" : 12,
+      "Ios17.logicalNot" : 2,
+      "Ios17.gather" : 1,
+      "Ios17.layerNorm" : 72,
+      "Ios16.constexprAffineDequantize" : 72,
+      "Ios17.less" : 1,
+      "Ios17.cast" : 12,
+      "Ios16.constexprLutToDense" : 53,
+      "Ios16.silu" : 36,
+      "Ios17.conv" : 36,
+      "Ios17.linear" : 98,
+      "Ios17.concat" : 37,
+      "Ios17.greaterEqual" : 4,
+      "Stack" : 2,
+      "Ios17.clip" : 1,
+      "Ios16.relu" : 1,
+      "Ios17.mul" : 49
+    },
+    "computePrecision" : "Mixed (Float16, Float32, Int16, Int32)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 28 × 1024)",
+        "shortDescription" : "",
+        "shape" : "[1, 28, 1024]",
+        "name" : "hidden",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "pre_encoded_length",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 12 × 42 × 1024)",
+        "shortDescription" : "",
+        "shape" : "[1, 12, 42, 1024]",
+        "name" : "cache_channel",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 12 × 1024 × 8)",
+        "shortDescription" : "",
+        "shape" : "[1, 12, 1024, 8]",
+        "name" : "cache_time",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "cache_len",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "prompt_id",
+        "type" : "MultiArray"
+      }
+    ],
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.conversion_date" : "2026-11-08",
+      "com.github.apple.coremltools.source" : "torch==2.7.0",
+      "com.github.apple.coremltools.version" : "9.0",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "generatedClassName" : "body_12_23",
+    "method" : "predict"
+  }
+]

test-1141/encoder-split/2240ms/encoder_body_12_23.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ab6283745c243e083a532b4ea960986747c23cafaa2d8ac7b750adb6e4070ba5
+size 318141952

test-1141/multilingual/1120ms/encoder.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5488ed81fa1e370dd5411218d9b0e0562d01ca33c45b9ea1935bab6c8e8dc943
+size 243

test-1141/multilingual/1120ms/encoder.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:93d7f1b6412b965b379f3359e38f9c4330099b5b41f9d30b85bf7286b098a78d
+size 572

test-1141/multilingual/1120ms/encoder.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e3dbd37659c7ef51d2dc691afac42fbdd1f04998fa62816ff0f2eebec5cc3d01
+size 564646848

test-1141/multilingual/2240ms/encoder.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:af569fd95237bdf8b91d38094691bfc990fcf3534316fa615bac5964c02809e4
+size 243

test-1141/multilingual/2240ms/encoder.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d5471a4edf55ce02dab51e5d01c86b419a7659bcdd6aac216ba6f0b795eae8bf
+size 573

test-1141/multilingual/2240ms/encoder.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2e00be98049a22e095452c020f183d2b23728e145cc814ba031436931b4f2e8f
+size 565336640

test-1141/multilingual/560ms/encoder.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3409887ed9eadc8689bd4c64754d95e4119152dc0b9ee1650664d2a6a689c783
+size 243

test-1141/multilingual/560ms/encoder.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:269d7188428298be3d0b2803c0e42bf6d44f7b4a9371030ae690beac443afbfa
+size 572

test-1141/multilingual/560ms/encoder.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0e994440ed6c936ac375d375eaab16eb445881ff2d9bc62384de991b411e234e
+size 564307200

test-1141/patch_mlmodelc.py

@@ -0,0 +1,116 @@
+#!/usr/bin/env python3
+"""FluidAudio #1141 path 2b: apply the uncut-concat rewrite to a compiled .mlmodelc's model.mil text.
+
+  python3 patch_mlmodelc.py <src.mlmodelc> <dst.mlmodelc> [--dry-run]
+Same rewrite as the reporter's uncut_concat.py (slice of a concat -> concat of slices of its inputs), but
+on the shipped compiled model, so it needs no .mlpackage and no coremltools. The weights are hard-linked,
+not rewritten. Only slices along the concat's axis, with unit stride and no squeeze, are touched.
+"""
+import os
+import re
+import shutil
+import sys
+from pathlib import Path
+
+LINE = re.compile(r"^(\s*)tensor<(\w+), \[([^\]]*)\]> (\w+) = (\w+)\((.*)\)\[name = tensor<string, \[\]>\(\"([^\"]*)\"\)\];$")
+CONST = re.compile(r"^\s*tensor<\w+, \[[^\]]*\]> (\w+) = const\(\)")
+VAL = re.compile(r"val = tensor<\w+, \[[^\]]*\]>\((.*)\)\]")
+
+
+def ints(s):
+    return [int(v) for v in s.strip("[]").split(", ")] if s.strip("[]") else []
+
+
+def patch(text):
+    lines = text.split("\n")
+    defs, consts = {}, {}
+    for i, l in enumerate(lines):
+        if (c := CONST.match(l)) and (v := VAL.search(l)):
+            consts[c[1]] = v[1]
+            continue
+        m = LINE.match(l)
+        if not m:
+            continue
+        ind, dt, shp, var, op, args, name = m.groups()
+        defs[var] = dict(i=i, ind=ind, dt=dt, shape=ints(shp), op=op, args=dict(re.findall(r"(\w+) = (\([^)]*\)|\w+)", args)),
+                         name=name)
+    out_lines, n = {}, 0
+    for var, d in defs.items():
+        if d["op"] != "slice_by_index" or defs.get(d["args"].get("x"), {}).get("op") != "concat":
+            continue
+        a, cat = d["args"], defs[d["args"]["x"]]
+        if "stride" in a and set(consts.get(a["stride"], "?").strip("[]").split(", ")) != {"1"}:
+            continue
+        if "squeeze_mask" in a and "true" in consts.get(a["squeeze_mask"], "true"):
+            continue
+        if consts.get(cat["args"]["interleave"]) != "false":
+            continue
+        rank, ishape = len(cat["shape"]), cat["shape"]
+        axis = int(consts[cat["args"]["axis"]]) % rank
+        begin, end = ints(consts[a["begin"]]), ints(consts[a["end"]])
+        bm = consts.get(a.get("begin_mask"), "[" + ", ".join(["false"] * rank) + "]").strip("[]").split(", ")
+        em = consts.get(a.get("end_mask"), "[" + ", ".join(["false"] * rank) + "]").strip("[]").split(", ")
+        span, ok = None, True
+        for k in range(rank):
+            b = 0 if bm[k] == "true" else min(max(begin[k] + ishape[k] * (begin[k] < 0), 0), ishape[k])
+            e = ishape[k] if em[k] == "true" else min(max(end[k] + ishape[k] * (end[k] < 0), 0), ishape[k])
+            if k == axis:
+                span = (b, e)
+            elif (b, e) != (0, ishape[k]):
+                ok = False
+        if not ok or span[0] >= span[1]:
+            continue
+        lo, hi = span
+        parts, at = [], 0
+        for v in cat["args"]["values"].strip("()").split(", "):
+            w = defs[v]["shape"][axis]
+            s, z = max(lo, at), min(hi, at + w)
+            if s < z:
+                parts.append((v, s - at, z - at, w))
+            at += w
+        ind, dt = d["ind"], d["dt"]
+        new, pieces = [], []
+        for j, (v, s, z, w) in enumerate(parts):
+            if (s, z) == (0, w):
+                pieces.append(v)
+                continue
+            vs = defs[v]["shape"]
+            b = [0] * rank; e = list(vs); b[axis], e[axis] = s, z
+            pv, shp = f"{var}_uncut{j}", list(vs); shp[axis] = z - s
+            new += [f'{ind}tensor<int32, [{rank}]> {pv}_begin = const()[name = tensor<string, []>("{pv}_begin"), val = tensor<int32, [{rank}]>([{", ".join(map(str, b))}])];',
+                    f'{ind}tensor<int32, [{rank}]> {pv}_end = const()[name = tensor<string, []>("{pv}_end"), val = tensor<int32, [{rank}]>([{", ".join(map(str, e))}])];',
+                    f'{ind}tensor<{dt}, [{", ".join(map(str, shp))}]> {pv} = slice_by_index(begin = {pv}_begin, end = {pv}_end, x = {v})[name = tensor<string, []>("{pv}")];']
+            pieces.append(pv)
+        oshape = ", ".join(map(str, d["shape"]))
+        if len(pieces) == 1 and pieces[0].startswith(f"{var}_uncut"):
+            # rename the single slice to the original var/name
+            new[-1] = new[-1].replace(f"> {pieces[0]} = ", f"> {var} = ").replace(f'("{pieces[0]}")]', f'("{d["name"]}")]')
+        elif len(pieces) == 1:
+            new.append(f'{ind}tensor<{dt}, [{oshape}]> {var} = identity(x = {pieces[0]})[name = tensor<string, []>("{d["name"]}")];')
+        else:
+            new.append(f'{ind}tensor<{dt}, [{oshape}]> {var} = concat(axis = {cat["args"]["axis"]}, interleave = {cat["args"]["interleave"]}, '
+                       f'values = ({", ".join(pieces)}))[name = tensor<string, []>("{d["name"]}")];')
+        out_lines[d["i"]] = new
+        n += 1
+        print(f"  {var} = {cat['args']['values']}{ishape} [{lo}:{hi}] axis {axis} -> {', '.join(f'{v}[{s}:{z}]' for v, s, z, _ in parts)}")
+    res = []
+    for i, l in enumerate(lines):
+        res.extend(out_lines.get(i, [l]))
+    return "\n".join(res), n
+
+
+def main():
+    src, dst = Path(sys.argv[1]), Path(sys.argv[2])
+    text, n = patch((src / "model.mil").read_text())
+    print(f"{n} slices rewritten")
+    if "--dry-run" in sys.argv:
+        return
+    if dst.exists():
+        shutil.rmtree(dst)
+    shutil.copytree(src, dst, copy_function=os.link)  # hard links: weights untouched
+    (dst / "model.mil").unlink()
+    (dst / "model.mil").write_text(text)
+
+
+if __name__ == "__main__":
+    main()