This is a submission for [DEV's Summer Bug Smash: Clear the Lineup](https://dev.to/bugsmash) powered by [Sentry](https://sentry.io/).

#devchallenge #bugsmash

Summer Bug Smash: Clear the Lineup 🐛🛹

This is a submission for DEV's Summer Bug Smash: Clear the Lineup powered by Sentry.

Project Overview

CODA OS is a fully local AI ecosystem — 228K+ lines of code, custom 2.2B transformer, 3D body reconstruction from phone videos, voice assistant, video calling, messaging. Zero cloud dependency.

The specific repos I fixed:

rotation-scan-3d — CPU-only 3D body reconstruction from phone rotation videos
coda-forge-3d — 18-stage photorealistic 3D reconstruction pipeline
bhaasm-transformer — Custom decoder-only transformer with MoE, GQA, RoPE

Bug Fix or Performance Improvement

Three issues across three repos:

1. Texture bug in GLB export (coda-forge-3d)

The _export_textured_glb method accepted a texture_path parameter but never used it. The texture was hardcoded to None, so every textured GLB export lost its texture. This meant users calling export_all(mesh_path, texture_path="texture.png") would get a GLB without any texture applied.

2. Tokenizer loaded 256 times per generation (bhaasm-transformer)

The _tokenize and _detokenize functions in sampler.py each independently loaded the BPE tokenizer from disk on every single call. During generation with max_new_tokens=256, this meant 256+ file reads and tokenizer initializations. The tokenizer was never cached.

3. Redundant computation (rotation-scan-3d)

In pipeline_adapter.py, estimate_body_volume() was called twice in succession — once without weight_kg (result discarded) and once with weight_kg. The first call was wasted computation that did nothing.

Code

Fix 1: Texture bug — coda-forge-3d/src/codaforge/export/export_manager.py

# Before (broken):
def _export_textured_glb(self, mesh, texture_path, path):
    material = trimesh.visual.material.PBRMaterial(
        baseColorTexture=None,  # <-- BUG: texture ignored
        baseColorFactor=[1.0, 1.0, 1.0, 1.0],
    )
    if mesh.visual.kind == "texture":
        mesh.visual.material = material
    mesh.export(path, file_type="glb")

# After (fixed):
def _export_textured_glb(self, mesh, texture_path, path):
    import PIL.Image as PILImage
    img = PILImage.open(texture_path)
    material = trimesh.visual.material.PBRMaterial(
        baseColorTexture=img,  # <-- Now actually uses the texture
        baseColorFactor=[1.0, 1.0, 1.0, 1.0],
    )
    mesh.visual.material = material
    mesh.export(path, file_type="glb")

Fix 2: Tokenizer caching — bhaasm-transformer/src/bhaasm/inference/sampler.py

# Before (slow):
def _tokenize(text, vocab_size):
    try:
        tok_path = Path("data/bpe_tokenizer.json")
        if tok_path.exists():
            tok = load_tokenizer(str(tok_path))  # Loaded every call
            return tok.encode(text)
    except Exception:
        pass
    return [min(ord(c), vocab_size - 1) for c in text] + [0]

# After (fast):
from functools import lru_cache

@lru_cache(maxsize=1)
def _get_tokenizer(tokenizer_path="data/bpe_tokenizer.json"):
    tok_path = Path(tokenizer_path)
    if tok_path.exists():
        return load_tokenizer(str(tok_path))
    return None

def _tokenize(text, vocab_size):
    tok = _get_tokenizer()
    if tok is not None:
        return tok.encode(text)
    return [min(ord(c), vocab_size - 1) for c in text] + [0]

Fix 3: Redundant call — rotation-scan-3d/src/rotationscan/pipeline_adapter.py

# Before (wasted):
if weight_kg is not None and weight_kg > 0:
    estimate_body_volume(ellipse_params, lengths_cm)           # Result discarded
    estimate_body_volume(ellipse_params, lengths_cm, weight_kg=weight_kg)

# After (clean):
if weight_kg is not None and weight_kg > 0:
    estimate_body_volume(ellipse_params, lengths_cm, weight_kg=weight_kg)

My Improvements

Texture fix: The bug was subtle — the method signature accepted texture_path and the caller checked if texture_path and os.path.isfile(texture_path) before calling it, so it looked like it should work. But inside the method, baseColorTexture=None meant the texture was silently dropped. The fix loads the image with PIL and passes it to PBRMaterial. Also added error handling around trimesh.load in the parent export_all method.

Tokenizer performance: The old code loaded the tokenizer from disk on every _tokenize and _detokenize call. During a 256-token generation, that's 512+ file reads. Using functools.lru_cache(maxsize=1) caches the tokenizer object after the first load. Same tokenizer, zero re-reads. The _detokenize function was also updated to use the cached tokenizer instead of independently loading its own copy.

Redundant computation: The first estimate_body_volume call computed uncalibrated volume but discarded the result. The second call recomputed with weight calibration. Removing the first call eliminates unnecessary work without changing behavior — the calibrated result is identical.

All three fixes are committed and CI-passing across Python 3.10, 3.11, and 3.12.