mlx_native/
lib.rs

1//! # mlx-native
2//!
3//! Pure-Rust Metal GPU compute library for MLX-compatible inference on Apple
4//! Silicon.
5//!
6//! This crate provides a thin, safe wrapper around Apple's Metal framework
7//! focused on compute shader dispatch for neural network inference.  It is
8//! designed to be the GPU backend for the `hf2q` inference engine.
9//!
10//! ## Key Types
11//!
12//! | Type | Purpose |
13//! |------|---------|
14//! | [`MlxDevice`]       | Metal device + command queue (entry point) |
15//! | [`CommandEncoder`]   | Batched compute command submission |
16//! | [`MlxBuffer`]        | Typed Metal buffer with shape/dtype metadata |
17//! | [`MlxBufferPool`]    | Arena allocator with power-of-two bucketing |
18//! | [`KernelRegistry`]   | Lazy MSL compilation + pipeline cache |
19//! | [`DType`]            | Element data type enum |
20//! | [`MlxError`]         | Unified error type (never panics) |
21//!
22//! ## Quick Start
23//!
24//! ```ignore
25//! use mlx_native::{MlxDevice, DType};
26//!
27//! let device = MlxDevice::new()?;
28//! let buf = device.alloc_buffer(1024, DType::F32, vec![256])?;
29//! let encoder = device.command_encoder()?;
30//! ```
31//!
32//! ## Design Principles
33//!
34//! * **No panics** — all public APIs return `Result<T, MlxError>`.
35//! * **Zero-copy** — `StorageModeShared` buffers on Apple Silicon unified memory.
36//! * **Thread-safe** — `MlxDevice` and `MlxBuffer` are `Send + Sync`.
37//! * **Lazy compilation** — MSL shaders compiled on first use, then cached.
38
39// Enforce the no-panic policy at compile time.
40#![deny(clippy::panic, clippy::unwrap_used, clippy::expect_used)]
41// The `objc` crate's `msg_send!` macro internally checks `cfg(feature = "cargo-clippy")`
42// which triggers unexpected_cfgs warnings. Suppress at crate level since we can't
43// control the macro expansion site.
44#![allow(unexpected_cfgs)]
45
46// ---- internal modules ----
47#[macro_use]
48mod error;
49mod buffer;
50mod buffer_pool;
51mod device;
52mod dtypes;
53mod encoder;
54mod encoder_session;
55mod kernel_registry;
56mod mem_ranges;
57mod residency;
58pub mod gguf;
59pub mod kernel_profile;
60pub mod graph;
61pub mod metal_capture;
62pub mod ops;
63pub mod turboquant;
64pub mod weight;
65
66// ---- public re-exports ----
67pub use buffer::MlxBuffer;
68pub use buffer_pool::MlxBufferPool;
69pub use device::MlxDevice;
70pub use dtypes::DType;
71pub use encoder::{
72    auto_barrier_concurrent_count, auto_barrier_count, barrier_count, barrier_total_ns,
73    cmd_buf_count, dispatch_count, reset_counters, sync_count, CapturedNode, CapturedOpKind,
74    CommandEncoder, DispatchKind, KernelArg, RecordedBinding,
75};
76pub use encoder_session::EncoderSession;
77pub use mem_ranges::{BufferRange, MemRangeRole, MemRanges};
78pub use error::{MlxError, Result};
79pub use graph::{ComputeGraph, GraphExecutor, GraphSession, OpKind};
80pub use kernel_registry::KernelRegistry;
81// Test-only counters and gate-reset helpers.  Marked #[doc(hidden)] so
82// they don't appear in published rustdoc; consumers should not depend
83// on them outside test code.  Not feature-gated because integration
84// tests in tests/ are a separate crate and cannot rely on the lib's
85// `test` cfg flag.
86#[doc(hidden)]
87pub use residency::{
88    macos_15_or_newer_for_test, reset_residency_env_cache_for_test,
89    reset_residency_test_counters, residency_allocation_count_for_test,
90    residency_commit_call_count_for_test,
91};
92
93// Re-export GGUF parser.
94pub use gguf::{GgufFile, MetadataValue, TensorInfo};
95
96// Re-export ops.
97pub use ops::dense_mm_bf16::{dense_matmul_bf16_f32_tensor, DenseMmBf16F32Params};
98pub use ops::dense_mm_f16::{dense_matmul_f16_f32_tensor, DenseMmF16F32Params};
99pub use ops::dense_mm_f32_f32::{dense_matmul_f32_f32_tensor, DenseMmF32F32Params};
100pub use ops::quantized_matmul::{quantized_matmul, quantized_matmul_simd, QuantizedMatmulParams};
101pub use ops::quantized_matmul_ggml::{
102    dispatch_mm_for_test, quantized_matmul_ggml, quantized_matmul_mm_tensor_perm021,
103    GgmlQuantizedMatmulParams, GgmlQuantizedMatmulPerm021Params, GgmlType,
104    MM_ROUTING_THRESHOLD,
105};
106pub use ops::quantized_matmul_id::{quantized_matmul_id, QuantizedMatmulIdParams};
107pub use ops::quantized_matmul_id_ggml::{
108    dispatch_id_mm_for_test, quantized_matmul_id_ggml, quantized_matmul_id_ggml_pooled,
109    quantized_matmul_id_swiglu_q4_0,
110    GgmlIdMmDispatchParams, GgmlQuantizedMatmulIdParams, IdMmScratch,
111    MM_ID_ROUTING_THRESHOLD,
112};
113
114// Re-export weight loading utilities.
115pub use weight::{
116    load_quantized_weights, safetensors_to_metal_buffer, QuantizationConfig, QuantizedWeight,
117    SafetensorsFile, TensorQuantConfig,
118};
119
120// Re-export metal types that appear in the public API.
121pub use metal::MTLSize;
122pub use metal;
123
124#[cfg(test)]
125#[allow(clippy::expect_used, clippy::unwrap_used, clippy::panic)]
126mod tests {
127    use super::*;
128
129    // ---- T10.7: compile-time Send + Sync assertions ----
130    fn _assert_send<T: Send>() {}
131    fn _assert_sync<T: Sync>() {}
132
133    #[allow(dead_code)]
134    fn assert_send_sync() {
135        _assert_send::<MlxDevice>();
136        _assert_sync::<MlxDevice>();
137        _assert_send::<MlxBuffer>();
138        _assert_sync::<MlxBuffer>();
139        _assert_send::<MlxError>();
140        _assert_sync::<MlxError>();
141    }
142
143    // ---- T10.1: device initialization ----
144    #[test]
145    fn test_device_init() {
146        let device = MlxDevice::new().expect("MlxDevice::new() should succeed on Apple Silicon");
147        let name = device.name();
148        assert!(!name.is_empty(), "Device name should not be empty");
149        println!("Metal device: {name}");
150    }
151
152    // ---- T10.2: buffer allocation ----
153    #[test]
154    fn test_buffer_alloc() {
155        let device = MlxDevice::new().expect("device");
156        let shape = vec![2, 3, 4];
157        let byte_len = 2 * 3 * 4 * DType::F32.size_of(); // 96 bytes
158        let buf = device
159            .alloc_buffer(byte_len, DType::F32, shape.clone())
160            .expect("alloc_buffer");
161
162        assert_eq!(buf.dtype(), DType::F32);
163        assert_eq!(buf.shape(), &shape);
164        assert_eq!(buf.byte_len(), byte_len);
165        assert_eq!(buf.element_count(), 24);
166    }
167
168    // ---- T10.3: buffer read/write round-trip ----
169    #[test]
170    fn test_buffer_readwrite() {
171        let device = MlxDevice::new().expect("device");
172        let n = 64;
173        let byte_len = n * std::mem::size_of::<f32>();
174        let mut buf = device
175            .alloc_buffer(byte_len, DType::F32, vec![n])
176            .expect("alloc_buffer");
177
178        // Write known data.
179        {
180            let slice: &mut [f32] = buf.as_mut_slice().expect("as_mut_slice");
181            assert_eq!(slice.len(), n);
182            for (i, val) in slice.iter_mut().enumerate() {
183                *val = i as f32 * 1.5;
184            }
185        }
186
187        // Read back and verify.
188        {
189            let slice: &[f32] = buf.as_slice().expect("as_slice");
190            for (i, &val) in slice.iter().enumerate() {
191                let expected = i as f32 * 1.5;
192                assert!(
193                    (val - expected).abs() < f32::EPSILON,
194                    "Mismatch at index {i}: got {val}, expected {expected}"
195                );
196            }
197        }
198    }
199
200    // ---- T10.4: encoder lifecycle ----
201    #[test]
202    fn test_encoder_lifecycle() {
203        let device = MlxDevice::new().expect("device");
204        let mut enc = device.command_encoder().expect("command_encoder");
205        // Commit an empty command buffer — should succeed (no-op on GPU).
206        enc.commit_and_wait()
207            .expect("commit_and_wait on empty encoder");
208    }
209
210    // ---- T10.5: buffer pool reuse ----
211    #[test]
212    fn test_buffer_pool_reuse() {
213        let device = MlxDevice::new().expect("device");
214        let mut pool = MlxBufferPool::new();
215
216        // Allocate a buffer.
217        let buf1 = pool
218            .alloc(&device, 1024, DType::F32, vec![256])
219            .expect("pool alloc 1");
220        let buf1_ptr = buf1.contents_ptr();
221        let buf1_byte_len = buf1.byte_len();
222
223        // Release it back to the pool.
224        pool.release(buf1);
225        assert_eq!(pool.free_count(), 1);
226
227        // Allocate again — should reuse the same Metal buffer.
228        let buf2 = pool
229            .alloc(&device, 1024, DType::F32, vec![256])
230            .expect("pool alloc 2");
231        let buf2_ptr = buf2.contents_ptr();
232        let buf2_byte_len = buf2.byte_len();
233
234        assert_eq!(buf1_ptr, buf2_ptr, "Pool should reuse the same Metal buffer");
235        assert_eq!(buf1_byte_len, buf2_byte_len, "Byte lengths should match");
236        assert_eq!(pool.free_count(), 0, "Free list should be empty after reuse");
237    }
238
239    // ---- T10.6: kernel registry caching ----
240    #[test]
241    fn test_kernel_registry_caching() {
242        let device = MlxDevice::new().expect("device");
243        let mut registry = KernelRegistry::new();
244
245        // Register a minimal test kernel.
246        registry.register_source(
247            "test_add",
248            r#"
249            #include <metal_stdlib>
250            using namespace metal;
251            kernel void test_add(
252                device float *a [[buffer(0)]],
253                device float *b [[buffer(1)]],
254                device float *c [[buffer(2)]],
255                uint id [[thread_position_in_grid]]
256            ) {
257                c[id] = a[id] + b[id];
258            }
259            "#,
260        );
261
262        // First call — compiles the shader.
263        assert!(!registry.is_cached("test_add"));
264        let p1 = registry
265            .get_pipeline("test_add", device.metal_device())
266            .expect("get_pipeline first call");
267        let p1_ptr = p1 as *const _;
268        assert!(registry.is_cached("test_add"));
269
270        // Second call — returns cached pipeline.
271        let p2 = registry
272            .get_pipeline("test_add", device.metal_device())
273            .expect("get_pipeline second call");
274        let p2_ptr = p2 as *const _;
275
276        assert_eq!(
277            p1_ptr, p2_ptr,
278            "Second get_pipeline call should return the same cached pipeline"
279        );
280    }
281
282    // ---- Additional: test alloc_buffer with zero length returns error ----
283    #[test]
284    fn test_buffer_alloc_zero_len_error() {
285        let device = MlxDevice::new().expect("device");
286        let result = device.alloc_buffer(0, DType::F32, vec![]);
287        assert!(result.is_err(), "Zero-length allocation should fail");
288        match result {
289            Err(MlxError::InvalidArgument(_)) => {}
290            other => panic!("Expected InvalidArgument, got {:?}", other),
291        }
292    }
293
294    // ---- Additional: test kernel not found ----
295    #[test]
296    fn test_kernel_not_found() {
297        let device = MlxDevice::new().expect("device");
298        let mut registry = KernelRegistry::new();
299        let result = registry.get_pipeline("nonexistent_kernel", device.metal_device());
300        assert!(result.is_err());
301        match result {
302            Err(MlxError::KernelNotFound(name)) => {
303                assert_eq!(name, "nonexistent_kernel");
304            }
305            other => panic!("Expected KernelNotFound, got {:?}", other),
306        }
307    }
308
309    // ---- Additional: test DType properties ----
310    #[test]
311    fn test_dtype_sizes() {
312        assert_eq!(DType::F32.size_of(), 4);
313        assert_eq!(DType::F16.size_of(), 2);
314        assert_eq!(DType::BF16.size_of(), 2);
315        assert_eq!(DType::U8.size_of(), 1);
316        assert_eq!(DType::U16.size_of(), 2);
317        assert_eq!(DType::U32.size_of(), 4);
318        assert_eq!(DType::I32.size_of(), 4);
319    }
320
321    // ---- Additional: test MlxBuffer Debug ----
322    #[test]
323    fn test_buffer_debug() {
324        let device = MlxDevice::new().expect("device");
325        let buf = device
326            .alloc_buffer(64, DType::F16, vec![4, 8])
327            .expect("alloc_buffer");
328        let debug_str = format!("{:?}", buf);
329        assert!(debug_str.contains("MlxBuffer"));
330        assert!(debug_str.contains("F16"));
331        assert!(debug_str.contains("[4, 8]"));
332    }
333
334    // ---- Additional: test MlxError Display ----
335    #[test]
336    fn test_error_display() {
337        let e = MlxError::DeviceNotFound;
338        assert!(format!("{e}").contains("Metal GPU device"));
339
340        let e = MlxError::ShaderCompilationError {
341            name: "foo".into(),
342            message: "syntax error".into(),
343        };
344        assert!(format!("{e}").contains("foo"));
345        assert!(format!("{e}").contains("syntax error"));
346    }
347
348    // ---- Additional: test buffer pool with different sizes ----
349    #[test]
350    fn test_buffer_pool_size_buckets() {
351        let device = MlxDevice::new().expect("device");
352        let mut pool = MlxBufferPool::new();
353
354        // Allocate a 100-byte buffer (rounds to 128-byte bucket).
355        let buf_100 = pool.alloc(&device, 100, DType::U8, vec![100]).expect("alloc 100");
356        assert!(
357            buf_100.byte_len() >= 100,
358            "Buffer should be at least 100 bytes"
359        );
360        pool.release(buf_100);
361
362        // Allocate a 128-byte buffer — should reuse the same Metal buffer.
363        let buf_128 = pool.alloc(&device, 128, DType::U8, vec![128]).expect("alloc 128");
364        assert!(buf_128.byte_len() >= 128);
365        pool.release(buf_128);
366
367        // Allocate a 200-byte buffer — different bucket (256), fresh allocation.
368        let buf_200 = pool.alloc(&device, 200, DType::U8, vec![200]).expect("alloc 200");
369        assert!(buf_200.byte_len() >= 200);
370        pool.release(buf_200);
371
372        assert_eq!(pool.free_count(), 2, "Two different bucket sizes in pool");
373    }
374}
mlx_native/lib.rs

mlx_native/
lib.rs
