Darkwood · Symfony · PHP FFI · Flow

When AI Becomes a Library

The Nolife Local Experiment
Depth Anything 3 + Symfony + PHP FFI + Darkwood Flow

The default reflex

We use LLMs for everything.

problem

↓

large model

↓

prompt

↓

text

The question

What if the task already has a specialized model?

problem

↓

specialized model

↓

structured computation

Depth is not language

General LLM

"The mountains are far away."

Depth Anything 3

504 × 280
depth values

relative units · not meters

Depth Anything 3

RGB image

↓

Depth Anything 3

↓

depth

confidence

camera

3D / GLB

DA3-BASE · relative depth · not meters

Research → runtime

PyTorch checkpoint

↓

GGUF

↓

ggml

↓

Metal

↓

Depth · Camera · GLB

What “specialized” means

Not this project

fine-tuning

LoRA

training from scratch

This project

purpose-trained model

local native runtime

quantized representation

The unusual constraint

No Python service.
No cloud AI API.

Symfony

↓

PHP FFI

↓

native model

Architecture

Symfony hosts the boundary.

flowchart TB
  B["Browser"]
  S["Symfony"]
  F["Darkwood Flow"]
  N["NativeDepthBridge"]
  FFI["PHP FFI"]
  L["libdepthanything"]
  G["ggml / Metal"]
  D["Depth Anything 3"]
  B --> S --> F --> N --> FFI --> L --> G --> D

Ownership

PHP owns

HTTP · validation · Flow orchestration · GD colormap Twig · CLI · observability

Native owns

ViT · DPT · depth confidence · camera Metal · GLB export

PHP does not perform matrix multiplication.

PHP FFI

$ffi->da_capi_depth_dense(
    $ctx, $imagePath, …
);

native pointer

↓

PHP array

↓

da_capi_free_floats

Application pipeline

Upload

↓

Validate → Infer → Colormap

↓

GLB → Result

Darkwood Flow · explicit steps · per-step timings

Observability

Flow did not make inference faster.
It made the pipeline visible.

The benchmark that did not make sense

Step	ms
Validate	~3
InferDepth	~2639
Colormap	~86
ExportScene	~2535
Total	~5260

Why is ExportScene another 2.5 seconds?

Before

The hidden second inference

image
 ├── inference → depth
 └── inference → GLB

da_capi_depth_dense da_capi_export_glb → depth_pose_native() × 2

After

One inference, several outputs

image
  ↓
one inference
  ↓
native result
 ├── depth
 ├── confidence
 ├── camera
 └── GLB

da_capi_export_glb_cached · ABI v11

Measured impact

~5.26 s 2 native inferences

~3.27 s 1 native inference · GLB ~19 ms

504×336 · q4_k · warm · Apple M4 / Metal

Measurement discipline

Different tensor size ≠ regression

Source	Processed	Infer median
1280×720	504×280	~2495 ms
1024×680	504×336	~3.0–3.2 s

Compare processed resolution — not headline numbers alone.

Quantization

4× smaller.
Not 4× faster.

q4_k ~99 MB · ~2495 ms

f32 ~393 MB · ~2495 ms

Warm infer-only · 504×280 · Apple M4 / Metal

Local means local

No remote AI inference

No cloud API at runtime

GGUF on disk · Metal on device

Demo runtime works with networking disabled

AI becomes infrastructure

The model starts to look like a dependency.

php bin/console app:depth-check
php bin/console app:depth-bench
php bin/console app:depth-stress
php bin/console app:depth-cleanup

Flow / PHP overhead ~114 ms · infer dominates

Development loop

Early

idea
→ ask agent
→ inspect

Later

observe → evidence
→ issue → agent
→ implement → measure

Issue-first

The valuable part was the problem definition.

Observation:
Two native inference passes.

Evidence:
Infer ≈ 2.6 s · GLB ≈ 2.5 s

Constraint:
One inference per request.

Expected:
Reuse native dense result.

Non-goal:
No Messenger / concurrency layer.

Synthesis

Reduce unnecessary generality.

Runtime

general model
        ↓
specialized model

Development

general prompt
        ↓
well-defined issue

What I learned

PHP FFI was enough

Native infer dominates cost

Observability found the real bug

q4_k saved size, not latency here

Context > code volume

Evidence beats speculation

Conclusion

The task should determine the model, the architecture, and the prompt.

Depth Anything 3 · specialized computation

↓

Darkwood Flow · observable pipeline

↓

Issue-first brief · specialized context

Resources

Source

github.com/matyo91/nolife-local

Slides

github.com/matyo91/slidewire

Upstream

github.com/mudler/depth-anything.cpp