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

mlx_native/
lib.rs
