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

mlx_native/
lib.rs
