Robbyant/lingbot-map PR #88 — Apple Silicon & matplotlib 3.9 support — PR #88
Robbyant/lingbot-map · pull request #88 ·
Transcript
PlainEnglish
We're looking at pull request eighty-eight — a compatibility upgrade that brings Apple Silicon support and updates the codebase for the latest matplotlib release.
PlainEnglish
Let's walk through the objectives. The headline here is adding Apple Silicon MPS backend support so the model can run on Mac hardware. Along the way, we're fixing matplotlib compatibility for version three point nine and later, making all autocast contexts device-aware instead of hardcoded to CUDA, and working around some MPS datatype limitations.
PlainEnglish
Here's the situation. The prior code assumed you had CUDA hardware or fell back to CPU. If you tried running on a Mac with Apple Silicon, it wouldn't even attempt to use the MPS backend. And if you upgraded matplotlib to version three point nine, the visualization code would break because the old get cmap function was removed. This PR fixes both issues — it brings MPS into the device hierarchy and updates the colormap calls.
Architecture
Let's look at the device selection logic in demo dot py. Before, it was a simple ternary — CUDA if available, otherwise CPU. Now we've expanded it to a three-tier waterfall: CUDA first, then MPS if we're on Apple Silicon, then CPU as the fallback. This is the core of the Apple Silicon support.
Architecture
The next cluster of changes makes all the torch autocast contexts device-aware. Instead of hardcoding the string 'cuda', we now read the device type from the actual tensor location. This happens in demo dot py for the warm streaming calls and the main inference loop, and in gct base dot py for all four prediction heads. We also added an enabled flag so autocast skips when the dtype is already float thirty-two. This keeps mixed-precision behavior consistent across CUDA, MPS, and CPU.
Architecture
Here's a quirk we had to handle. MPS doesn't support float sixty-four or complex one twenty-eight datatypes. In the RoPE layer, the frequency precomputation normally uses complex one twenty-eight. So before we transfer the frequencies to the MPS device, we check the backend and downcast to complex sixty-four if needed. The check is idempotent — after the first call, the dtype is already sixty-four, so it skips the conversion.
Architecture
The matplotlib fix is straightforward. The old cm dot get cmap function was removed in matplotlib three point nine. We're now calling matplotlib dot colormaps dot get cmap instead. This change appears in three places — point cloud viewer dot py and two spots in vis utils dot py. It's a one-line swap, but it's critical for compatibility with recent matplotlib versions.
PlainEnglish
So where does that leave us? The codebase now has three-tier device support. CUDA is still the preferred path, MPS kicks in on Apple Silicon, and CPU is the final fallback. Mixed-precision inference runs with bfloat sixteen on both CUDA H one hundred plus cards and MPS, while other CUDA hardware uses float sixteen and CPU stays at float thirty-two. The model can now run on Mac hardware with the use sdpa flag. FlashInfer remains CUDA-only.
CodeQuality
There's one thing to keep an eye on. The RoPE frequency downcast happens in-place every time forward is called. The condition is idempotent after the first pass, so it doesn't keep converting the same tensor. But if a model is moved between CPU and MPS repeatedly in the same session, this could get fragile. The check guards the mutation with both device type and dtype, so it's acceptable for normal use, just something to be aware of.
PlainEnglish
That wraps it up. This is a focused compatibility upgrade — it brings MPS support into the device hierarchy, updates the matplotlib API calls, and handles the MPS dtype constraints with a simple downcast. The changes are purely additive; CUDA and CPU paths are unchanged. The model can now run on Mac hardware, and the visualization code works with recent matplotlib versions.
How this was made
Lenzon read Robbyant/lingbot-map at pull request #88 and generated this walkthrough automatically. The narration above is the transcript of what it says.
Explain a pull request from your own repo
Point Lenzon at a repo or a pull request and get a narrated walkthrough like this one.
Try it