docs: rebuild the README around the Pi 5 promise, ship CLI binaries — v0.8.1 - #210
Merged
Conversation
… v0.8.1 The first screen is now the product claim and its proof, in order: "Run Microsoft BitNet 2B locally on a 4 GB Raspberry Pi 5. One binary. No Python, no model setup, no cloud." — then the install command, a run example, the demo, and three measured numbers, each of which the repo already backs: the reference runs come from a 4 GB Pi 5 Model B (ram_gb 4.0 in benchmark/reference_runs.json), decode measured 17.9 t/s short-context and 15.0 t/s at 512 tokens, and the whole download is ~1.2 GB. Private and offline after download. What makes the promise literally true: the release now ships two fully static musl CLI binaries for linux-arm64 — geist-linux-arm64 (<1 MB, any GGUF) and geist-bitnet-linux-arm64 (~1.2 GB, the MIT-licensed BitNet b1.58 2B-4T folded in via .incbin, aliased zero-copy out of .rodata and demand-paged like an mmap, so RAM behavior on a 4 GB board equals the file-based load). GNU ld handles the >1 GB section; ld64 does not (measured: registry corruption, then SIGSEGV), which is why macOS gets no embedded artifact and builds from source instead. The embedded smoke in release.yml decodes real tokens from the embedded model before anything is published. simple_generate stops at the model's EOS (geist_model_eos_token) and defaults to 256 new tokens — an untemplated completion can run to the end of the context without ever emitting EOS, so "unbounded" was a footgun as a default (measured: the demo prompt loops until the window fills). An explicit count still overrides. Everything the first screen displaced moved rather than died: the model table to docs/MODELS.md, Metal/Vulkan to docs/BACKENDS.md, "Why C?" to docs/ARCHITECTURE.md, the bench methodology and the full CPU table to benchmark/README.md. v0.8.1 across geist.h, README, CITATION.cff and the CHANGELOG; check-version.sh green.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
First screen = the promise plus its proof: install command (
curl+chmod), run example, demo GIF, and three numbers the repo already backs (4 GB Pi perreference_runs.json, 15–18 decode t/s measured, ~1.2 GB download). Everything else moved down or out (docs/MODELS.md,docs/BACKENDS.md,benchmark/README.md, Why-C →ARCHITECTURE.md).Makes it literally true:
release.ymlshipsgeist-linux-arm64(<1 MB static CLI) andgeist-bitnet-linux-arm64(~1.2 GB, BitNet folded in via.incbin, zero-copy/demand-paged; smoke decodes real tokens before publish).simple_generategains EOS stop + 256-token default (unbounded loops forever on untemplated prompts — measured) and the-DGEIST_EMBED_MODELpath.v0.8.1 everywhere,
check-version.shgreen. Plan: merge → tag v0.8.1 →releases/latestserves the first-screencurlcorrectly. Release workflow dry-run (workflow_dispatch on this branch, artifacts only) before merging.