Download recipe/export/keep_fp32.py from KaleComputer/kale-tts-1-tiny: direct link, hf CLI and curl.
- Browser
- Download file 1.7 kB
-
https://huggingface.co/KaleComputer/kale-tts-1-tiny/resolve/main/recipe/export/keep_fp32.py
- Command line
-
hf download hf://KaleComputer/kale-tts-1-tiny/recipe/export/keep_fp32.py
-
curl -L -o keep_fp32.py https://huggingface.co/KaleComputer/kale-tts-1-tiny/resolve/main/recipe/export/keep_fp32.py
1.7 kB
| # SPDX-License-Identifier: Apache-2.0 | |
| # Copyright 2026 Kale | |
| """keep chosen layers of an int8 graph in float32 (16 Nov, nano): for every int8 weight (DequantizeLinear) whose consumer node name matches | |
| one of the patterns, use the float32 reference graph's value of that weight instead (same name; checked shape-equal). | |
| usage: python keep_fp32.py <int8 graph> <float32 reference graph (same steps)> <out> <regex> [<regex> ...]""" | |
| import re, sys, collections | |
| import numpy as np, onnx | |
| from onnx import numpy_helper as nh | |
| src, ref, dst, pats = sys.argv[1], sys.argv[2], sys.argv[3], sys.argv[4:] | |
| m = onnx.load(src); g = m.graph; RI = {i.name: i for i in onnx.load(ref).graph.initializer} | |
| cons = collections.defaultdict(list) | |
| for n in g.node: | |
| for x in n.input: cons[x].append(n) | |
| def users(x, depth=0): | |
| out = [] | |
| for c in cons[x]: | |
| out.append(c.name) | |
| if c.op_type == "Identity" and depth < 3: out += users(c.output[0], depth + 1) | |
| return out | |
| keep, add, kept = [], [], [] | |
| for n in g.node: | |
| if n.op_type == "DequantizeLinear" and any(re.search(p, u) for u in users(n.output[0]) for p in pats): | |
| w = nh.to_array(RI[n.output[0]]).astype(np.float32); add.append(nh.from_array(w, n.output[0])); kept.append((n.output[0], w.size)); continue | |
| keep.append(n) | |
| g.ClearField("node"); g.node.extend(keep); g.initializer.extend(add) | |
| sys.path.insert(0, __import__("os").path.dirname(__file__)); from toposort import toposort, prune_initializers | |
| toposort(g); prune_initializers(g); onnx.checker.check_model(m); onnx.save(m, dst) | |
| print("kept in float32:", len(kept), "weights,", sum(s for _, s in kept), "values (+", round(sum(s for _, s in kept) * 3 / 1e6, 2), "MB vs int8)") | |