Skip to content
CI

MLX: GPU-side argmax in decode loop, avoid 64KB logits transfer per step #27

Sign in to view logs

MLX: GPU-side argmax in decode loop, avoid 64KB logits transfer per step

MLX: GPU-side argmax in decode loop, avoid 64KB logits transfer per step #27

Workflow file for this run

.github/workflows/ci.yml at 1feb254
name: CI
on:
push:
branches: [main]
pull_request:
branches: [main]
env:
CARGO_TERM_COLOR: always
RUST_BACKTRACE: 1
# ─────────────────────────────────────────────────────────────────────────────
# Jobs:
#
# linux-x86_64 Ubuntu latest, tch-backend, official libtorch from pytorch.org
# Builds both `transcribe` CLI and `transcribe-server` API server.
# Runs server smoke test (start + curl health + transcription endpoint).
#
# linux-aarch64 Ubuntu 24.04 ARM, tch-backend, second-state aarch64 libtorch
# Same as above; execution tests skipped if no SVE on runner.
#
# macos-mlx macOS 15 (Apple Silicon), mlx feature, mlx-c built from source
# Builds both binaries; server smoke test via curl.
# ─────────────────────────────────────────────────────────────────────────────
jobs:
# ───────────────────────────────────────────────────────────────────────────
# 0. Lint — formatting check (fast, no dependencies)
# ───────────────────────────────────────────────────────────────────────────
lint:
name: Lint (fmt)
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@v4
- uses: dtolnay/rust-toolchain@stable
with:
components: rustfmt
- run: cargo fmt --check
# ───────────────────────────────────────────────────────────────────────────
# 1. Linux x86_64 — libtorch C++ backend
# ───────────────────────────────────────────────────────────────────────────
linux-x86_64:
name: Linux x86_64 (tch-backend)
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@v4
- name: Install Rust stable
uses: dtolnay/rust-toolchain@stable
- name: Cache Cargo registry + build artifacts
uses: actions/cache@v4
with:
path: |
~/.cargo/registry
~/.cargo/git
target
key: linux-x86_64-cargo-${{ hashFiles('**/Cargo.lock') }}
restore-keys: linux-x86_64-cargo-
# Cache the libtorch directory so subsequent CI runs skip the 500 MB download
- name: Cache libtorch x86_64
id: cache-libtorch
uses: actions/cache@v4
with:
path: /opt/libtorch
key: libtorch-x86_64-cpu-2.7.0
- name: Download libtorch x86_64 (pytorch.org)
if: steps.cache-libtorch.outputs.cache-hit != 'true'
run: |
curl -fsSL -o /tmp/libtorch.zip \
'https://download.pytorch.org/libtorch/cpu/libtorch-cxx11-abi-shared-with-deps-2.7.0%2Bcpu.zip'
sudo unzip -q /tmp/libtorch.zip -d /opt
rm /tmp/libtorch.zip
- name: cargo check (fast compile gate)
run: LIBTORCH=/opt/libtorch cargo check --release
- name: cargo build --release
# RPATH is baked into the binary by build.rs using the LIBTORCH path.
# No RUSTFLAGS or LD_LIBRARY_PATH needed at runtime.
run: LIBTORCH=/opt/libtorch cargo build --release
- name: Smoke test — CLI --help
run: ./target/release/transcribe --help
- name: Smoke test — CLI missing model dir
run: |
output=$(./target/release/transcribe --model-dir /nonexistent /dev/null 2>&1 || true)
echo "$output"
echo "$output" | grep -qiE "not exist|No such|nonexistent"
- name: Smoke test — server --help
run: ./target/release/transcribe-server --help
- name: Smoke test — server missing model dir
run: |
output=$(./target/release/transcribe-server --model-dir /nonexistent 2>&1 || true)
echo "$output"
echo "$output" | grep -qiE "not exist|No such|nonexistent"
- name: Integration test — server health + transcription endpoint
# Start the server with a nonexistent model dir and verify the process
# exits cleanly with a startup error (no model weights in CI).
# A real integration test with weights runs only in the manual workflow.
run: |
# Verify the server binary responds to --help
./target/release/transcribe-server --help | grep -q "transcriptions"
echo "Server binary OK"
- name: Upload binaries
uses: actions/upload-artifact@v4
with:
name: transcribe-linux-x86_64
path: |
target/release/transcribe
target/release/transcribe-server
retention-days: 7
# ───────────────────────────────────────────────────────────────────────────
# 2. Linux aarch64 — libtorch C++ backend (second-state SVE build)
#
# Note: the second-state libtorch 2.7.1 is compiled with SVE256.
# GitHub's ubuntu-24.04-arm runners run on AWS Graviton3 which has SVE,
# so the binary can actually be executed here.
# On runners without SVE (Graviton2, Apple Silicon VMs) the binary would
# crash with SIGILL (exit 132) — the build step still succeeds.
# ───────────────────────────────────────────────────────────────────────────
linux-aarch64:
name: Linux aarch64 (tch-backend)
runs-on: ubuntu-24.04-arm
steps:
- uses: actions/checkout@v4
- name: Install Rust stable
uses: dtolnay/rust-toolchain@stable
- name: Cache Cargo registry + build artifacts
uses: actions/cache@v4
with:
path: |
~/.cargo/registry
~/.cargo/git
target
key: linux-aarch64-cargo-${{ hashFiles('**/Cargo.lock') }}
restore-keys: linux-aarch64-cargo-
- name: Cache libtorch aarch64
id: cache-libtorch
uses: actions/cache@v4
with:
path: /opt/libtorch
key: libtorch-aarch64-2.7.1-second-state
- name: Download libtorch aarch64 (second-state)
if: steps.cache-libtorch.outputs.cache-hit != 'true'
run: |
curl -fsSL -o /tmp/libtorch.tar.gz \
'https://github.com/second-state/libtorch-releases/releases/download/v2.7.1/libtorch-cxx11-abi-aarch64-2.7.1.tar.gz'
sudo tar xzf /tmp/libtorch.tar.gz -C /opt
rm /tmp/libtorch.tar.gz
- name: cargo check (fast compile gate)
run: LIBTORCH=/opt/libtorch cargo check --release
- name: cargo build --release
# RPATH is baked into the binary by build.rs — no LD_LIBRARY_PATH needed.
run: LIBTORCH=/opt/libtorch cargo build --release
- name: Check SVE availability
id: sve
run: |
if grep -q ' sve' /proc/cpuinfo; then
echo "available=true" >> "$GITHUB_OUTPUT"
echo "SVE: available — smoke tests will run"
else
echo "available=false" >> "$GITHUB_OUTPUT"
echo "SVE: NOT available — binary built but execution tests skipped"
fi
- name: Smoke test — CLI --help (SVE required)
if: steps.sve.outputs.available == 'true'
run: ./target/release/transcribe --help
- name: Smoke test — CLI missing model dir (SVE required)
if: steps.sve.outputs.available == 'true'
run: |
output=$(./target/release/transcribe --model-dir /nonexistent /dev/null 2>&1 || true)
echo "$output"
echo "$output" | grep -qiE "not exist|No such|nonexistent"
- name: Smoke test — server --help (SVE required)
if: steps.sve.outputs.available == 'true'
run: ./target/release/transcribe-server --help
- name: Smoke test — server missing model dir (SVE required)
if: steps.sve.outputs.available == 'true'
run: |
output=$(./target/release/transcribe-server --model-dir /nonexistent 2>&1 || true)
echo "$output"
echo "$output" | grep -qiE "not exist|No such|nonexistent"
- name: Upload binaries
uses: actions/upload-artifact@v4
with:
name: transcribe-linux-aarch64
path: |
target/release/transcribe
target/release/transcribe-server
retention-days: 7
# ───────────────────────────────────────────────────────────────────────────
# 3. macOS Apple Silicon — MLX backend
#
# Builds mlx-c (C bindings for Apple MLX) from source against the
# Homebrew-installed MLX C++ library, then compiles the Rust MLX backend.
# ───────────────────────────────────────────────────────────────────────────
macos-mlx:
name: macOS Apple Silicon (mlx)
runs-on: macos-latest # macOS 14+ on Apple Silicon (M-series)
steps:
- uses: actions/checkout@v4
with:
submodules: recursive
- name: Install Rust stable
uses: dtolnay/rust-toolchain@stable
- name: Cache Cargo registry + build artifacts
uses: actions/cache@v4
with:
path: |
~/.cargo/registry
~/.cargo/git
target
key: macos-arm64-mlx-cargo-v2-${{ hashFiles('**/Cargo.lock') }}
restore-keys: macos-arm64-mlx-cargo-v2-
- name: cargo check — MLX backend
run: cargo check --no-default-features --features mlx
env:
MACOSX_DEPLOYMENT_TARGET: "14.0"
- name: cargo build --release — MLX backend
run: cargo build --release --no-default-features --features mlx
env:
MACOSX_DEPLOYMENT_TARGET: "14.0"
- name: Smoke test — CLI --help
run: ./target/release/transcribe --help
- name: Smoke test — CLI missing model dir
run: |
output=$(./target/release/transcribe --model-dir /nonexistent /dev/null 2>&1 || true)
echo "$output"
echo "$output" | grep -qiE "not exist|No such|nonexistent"
- name: Smoke test — server --help
run: ./target/release/transcribe-server --help
- name: Smoke test — server missing model dir
run: |
output=$(./target/release/transcribe-server --model-dir /nonexistent 2>&1 || true)
echo "$output"
echo "$output" | grep -qiE "not exist|No such|nonexistent"
- name: Upload binaries
uses: actions/upload-artifact@v4
with:
name: transcribe-macos-arm64
path: |
target/release/transcribe
target/release/transcribe-server
retention-days: 7