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

mlx_native/
lib.rs
