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 kernel_registry;
55pub mod gguf;
56pub mod graph;
57pub mod ops;
58pub mod turboquant;
59pub mod weight;
60
61// ---- public re-exports ----
62pub use buffer::MlxBuffer;
63pub use buffer_pool::MlxBufferPool;
64pub use device::MlxDevice;
65pub use dtypes::DType;
66pub use encoder::{
67    dispatch_count, reset_counters, sync_count, CapturedNode, CommandEncoder, DispatchKind,
68    RecordedBinding,
69};
70pub use error::{MlxError, Result};
71pub use graph::{ComputeGraph, GraphExecutor, GraphSession, OpKind};
72pub use kernel_registry::KernelRegistry;
73
74// Re-export GGUF parser.
75pub use gguf::{GgufFile, MetadataValue, TensorInfo};
76
77// Re-export ops.
78pub use ops::dense_mm_bf16::{dense_matmul_bf16_f32_tensor, DenseMmBf16F32Params};
79pub use ops::dense_mm_f16::{dense_matmul_f16_f32_tensor, DenseMmF16F32Params};
80pub use ops::dense_mm_f32_f32::{dense_matmul_f32_f32_tensor, DenseMmF32F32Params};
81pub use ops::quantized_matmul::{quantized_matmul, quantized_matmul_simd, QuantizedMatmulParams};
82pub use ops::quantized_matmul_ggml::{
83    dispatch_mm_for_test, quantized_matmul_ggml, quantized_matmul_mm_tensor_perm021,
84    GgmlQuantizedMatmulParams, GgmlQuantizedMatmulPerm021Params, GgmlType,
85    MM_ROUTING_THRESHOLD,
86};
87pub use ops::quantized_matmul_id::{quantized_matmul_id, QuantizedMatmulIdParams};
88pub use ops::quantized_matmul_id_ggml::{
89    dispatch_id_mm_for_test, quantized_matmul_id_ggml, quantized_matmul_id_ggml_pooled,
90    GgmlIdMmDispatchParams, GgmlQuantizedMatmulIdParams, IdMmScratch,
91    MM_ID_ROUTING_THRESHOLD,
92};
93
94// Re-export weight loading utilities.
95pub use weight::{
96    load_quantized_weights, safetensors_to_metal_buffer, QuantizationConfig, QuantizedWeight,
97    SafetensorsFile, TensorQuantConfig,
98};
99
100// Re-export metal types that appear in the public API.
101pub use metal::MTLSize;
102pub use metal;
103
104#[cfg(test)]
105#[allow(clippy::expect_used, clippy::unwrap_used, clippy::panic)]
106mod tests {
107    use super::*;
108
109    // ---- T10.7: compile-time Send + Sync assertions ----
110    fn _assert_send<T: Send>() {}
111    fn _assert_sync<T: Sync>() {}
112
113    #[allow(dead_code)]
114    fn assert_send_sync() {
115        _assert_send::<MlxDevice>();
116        _assert_sync::<MlxDevice>();
117        _assert_send::<MlxBuffer>();
118        _assert_sync::<MlxBuffer>();
119        _assert_send::<MlxError>();
120        _assert_sync::<MlxError>();
121    }
122
123    // ---- T10.1: device initialization ----
124    #[test]
125    fn test_device_init() {
126        let device = MlxDevice::new().expect("MlxDevice::new() should succeed on Apple Silicon");
127        let name = device.name();
128        assert!(!name.is_empty(), "Device name should not be empty");
129        println!("Metal device: {name}");
130    }
131
132    // ---- T10.2: buffer allocation ----
133    #[test]
134    fn test_buffer_alloc() {
135        let device = MlxDevice::new().expect("device");
136        let shape = vec![2, 3, 4];
137        let byte_len = 2 * 3 * 4 * DType::F32.size_of(); // 96 bytes
138        let buf = device
139            .alloc_buffer(byte_len, DType::F32, shape.clone())
140            .expect("alloc_buffer");
141
142        assert_eq!(buf.dtype(), DType::F32);
143        assert_eq!(buf.shape(), &shape);
144        assert_eq!(buf.byte_len(), byte_len);
145        assert_eq!(buf.element_count(), 24);
146    }
147
148    // ---- T10.3: buffer read/write round-trip ----
149    #[test]
150    fn test_buffer_readwrite() {
151        let device = MlxDevice::new().expect("device");
152        let n = 64;
153        let byte_len = n * std::mem::size_of::<f32>();
154        let mut buf = device
155            .alloc_buffer(byte_len, DType::F32, vec![n])
156            .expect("alloc_buffer");
157
158        // Write known data.
159        {
160            let slice: &mut [f32] = buf.as_mut_slice().expect("as_mut_slice");
161            assert_eq!(slice.len(), n);
162            for (i, val) in slice.iter_mut().enumerate() {
163                *val = i as f32 * 1.5;
164            }
165        }
166
167        // Read back and verify.
168        {
169            let slice: &[f32] = buf.as_slice().expect("as_slice");
170            for (i, &val) in slice.iter().enumerate() {
171                let expected = i as f32 * 1.5;
172                assert!(
173                    (val - expected).abs() < f32::EPSILON,
174                    "Mismatch at index {i}: got {val}, expected {expected}"
175                );
176            }
177        }
178    }
179
180    // ---- T10.4: encoder lifecycle ----
181    #[test]
182    fn test_encoder_lifecycle() {
183        let device = MlxDevice::new().expect("device");
184        let mut enc = device.command_encoder().expect("command_encoder");
185        // Commit an empty command buffer — should succeed (no-op on GPU).
186        enc.commit_and_wait()
187            .expect("commit_and_wait on empty encoder");
188    }
189
190    // ---- T10.5: buffer pool reuse ----
191    #[test]
192    fn test_buffer_pool_reuse() {
193        let device = MlxDevice::new().expect("device");
194        let mut pool = MlxBufferPool::new();
195
196        // Allocate a buffer.
197        let buf1 = pool
198            .alloc(&device, 1024, DType::F32, vec![256])
199            .expect("pool alloc 1");
200        let buf1_ptr = buf1.contents_ptr();
201        let buf1_byte_len = buf1.byte_len();
202
203        // Release it back to the pool.
204        pool.release(buf1);
205        assert_eq!(pool.free_count(), 1);
206
207        // Allocate again — should reuse the same Metal buffer.
208        let buf2 = pool
209            .alloc(&device, 1024, DType::F32, vec![256])
210            .expect("pool alloc 2");
211        let buf2_ptr = buf2.contents_ptr();
212        let buf2_byte_len = buf2.byte_len();
213
214        assert_eq!(buf1_ptr, buf2_ptr, "Pool should reuse the same Metal buffer");
215        assert_eq!(buf1_byte_len, buf2_byte_len, "Byte lengths should match");
216        assert_eq!(pool.free_count(), 0, "Free list should be empty after reuse");
217    }
218
219    // ---- T10.6: kernel registry caching ----
220    #[test]
221    fn test_kernel_registry_caching() {
222        let device = MlxDevice::new().expect("device");
223        let mut registry = KernelRegistry::new();
224
225        // Register a minimal test kernel.
226        registry.register_source(
227            "test_add",
228            r#"
229            #include <metal_stdlib>
230            using namespace metal;
231            kernel void test_add(
232                device float *a [[buffer(0)]],
233                device float *b [[buffer(1)]],
234                device float *c [[buffer(2)]],
235                uint id [[thread_position_in_grid]]
236            ) {
237                c[id] = a[id] + b[id];
238            }
239            "#,
240        );
241
242        // First call — compiles the shader.
243        assert!(!registry.is_cached("test_add"));
244        let p1 = registry
245            .get_pipeline("test_add", device.metal_device())
246            .expect("get_pipeline first call");
247        let p1_ptr = p1 as *const _;
248        assert!(registry.is_cached("test_add"));
249
250        // Second call — returns cached pipeline.
251        let p2 = registry
252            .get_pipeline("test_add", device.metal_device())
253            .expect("get_pipeline second call");
254        let p2_ptr = p2 as *const _;
255
256        assert_eq!(
257            p1_ptr, p2_ptr,
258            "Second get_pipeline call should return the same cached pipeline"
259        );
260    }
261
262    // ---- Additional: test alloc_buffer with zero length returns error ----
263    #[test]
264    fn test_buffer_alloc_zero_len_error() {
265        let device = MlxDevice::new().expect("device");
266        let result = device.alloc_buffer(0, DType::F32, vec![]);
267        assert!(result.is_err(), "Zero-length allocation should fail");
268        match result {
269            Err(MlxError::InvalidArgument(_)) => {}
270            other => panic!("Expected InvalidArgument, got {:?}", other),
271        }
272    }
273
274    // ---- Additional: test kernel not found ----
275    #[test]
276    fn test_kernel_not_found() {
277        let device = MlxDevice::new().expect("device");
278        let mut registry = KernelRegistry::new();
279        let result = registry.get_pipeline("nonexistent_kernel", device.metal_device());
280        assert!(result.is_err());
281        match result {
282            Err(MlxError::KernelNotFound(name)) => {
283                assert_eq!(name, "nonexistent_kernel");
284            }
285            other => panic!("Expected KernelNotFound, got {:?}", other),
286        }
287    }
288
289    // ---- Additional: test DType properties ----
290    #[test]
291    fn test_dtype_sizes() {
292        assert_eq!(DType::F32.size_of(), 4);
293        assert_eq!(DType::F16.size_of(), 2);
294        assert_eq!(DType::BF16.size_of(), 2);
295        assert_eq!(DType::U8.size_of(), 1);
296        assert_eq!(DType::U16.size_of(), 2);
297        assert_eq!(DType::U32.size_of(), 4);
298        assert_eq!(DType::I32.size_of(), 4);
299    }
300
301    // ---- Additional: test MlxBuffer Debug ----
302    #[test]
303    fn test_buffer_debug() {
304        let device = MlxDevice::new().expect("device");
305        let buf = device
306            .alloc_buffer(64, DType::F16, vec![4, 8])
307            .expect("alloc_buffer");
308        let debug_str = format!("{:?}", buf);
309        assert!(debug_str.contains("MlxBuffer"));
310        assert!(debug_str.contains("F16"));
311        assert!(debug_str.contains("[4, 8]"));
312    }
313
314    // ---- Additional: test MlxError Display ----
315    #[test]
316    fn test_error_display() {
317        let e = MlxError::DeviceNotFound;
318        assert!(format!("{e}").contains("Metal GPU device"));
319
320        let e = MlxError::ShaderCompilationError {
321            name: "foo".into(),
322            message: "syntax error".into(),
323        };
324        assert!(format!("{e}").contains("foo"));
325        assert!(format!("{e}").contains("syntax error"));
326    }
327
328    // ---- Additional: test buffer pool with different sizes ----
329    #[test]
330    fn test_buffer_pool_size_buckets() {
331        let device = MlxDevice::new().expect("device");
332        let mut pool = MlxBufferPool::new();
333
334        // Allocate a 100-byte buffer (rounds to 128-byte bucket).
335        let buf_100 = pool.alloc(&device, 100, DType::U8, vec![100]).expect("alloc 100");
336        assert!(
337            buf_100.byte_len() >= 100,
338            "Buffer should be at least 100 bytes"
339        );
340        pool.release(buf_100);
341
342        // Allocate a 128-byte buffer — should reuse the same Metal buffer.
343        let buf_128 = pool.alloc(&device, 128, DType::U8, vec![128]).expect("alloc 128");
344        assert!(buf_128.byte_len() >= 128);
345        pool.release(buf_128);
346
347        // Allocate a 200-byte buffer — different bucket (256), fresh allocation.
348        let buf_200 = pool.alloc(&device, 200, DType::U8, vec![200]).expect("alloc 200");
349        assert!(buf_200.byte_len() >= 200);
350        pool.release(buf_200);
351
352        assert_eq!(pool.free_count(), 2, "Two different bucket sizes in pool");
353    }
354}
mlx_native/lib.rs

mlx_native/
lib.rs
