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 tq_oracle;
65pub mod weight;
66
67// ---- public re-exports ----
68pub use buffer::MlxBuffer;
69pub use buffer_pool::MlxBufferPool;
70pub use device::MlxDevice;
71pub use dtypes::DType;
72pub use encoder::{
73    auto_barrier_concurrent_count, auto_barrier_count, barrier_count, barrier_total_ns,
74    cmd_buf_count, dispatch_count, reset_counters, sync_count, CapturedNode, CapturedOpKind,
75    CommandEncoder, DispatchKind, KernelArg, RecordedBinding,
76};
77pub use encoder_session::EncoderSession;
78pub use mem_ranges::{BufferRange, MemRangeRole, MemRanges};
79pub use error::{MlxError, Result};
80pub use graph::{ComputeGraph, GraphExecutor, GraphSession, OpKind};
81pub use kernel_registry::KernelRegistry;
82// Test-only counters and gate-reset helpers.  Marked #[doc(hidden)] so
83// they don't appear in published rustdoc; consumers should not depend
84// on them outside test code.  Not feature-gated because integration
85// tests in tests/ are a separate crate and cannot rely on the lib's
86// `test` cfg flag.
87#[doc(hidden)]
88pub use residency::{
89    macos_15_or_newer_for_test, reset_residency_env_cache_for_test,
90    reset_residency_test_counters, residency_allocation_count_for_test,
91    residency_commit_call_count_for_test,
92};
93
94// Re-export GGUF parser.
95pub use gguf::{GgufFile, MetadataValue, TensorInfo};
96
97// Re-export ops.
98pub use ops::dense_mm_bf16::{dense_matmul_bf16_f32_tensor, DenseMmBf16F32Params};
99pub use ops::dense_mm_f16::{dense_matmul_f16_f32_tensor, DenseMmF16F32Params};
100pub use ops::dense_mm_f32_f32::{dense_matmul_f32_f32_tensor, DenseMmF32F32Params};
101pub use ops::quantized_matmul::{quantized_matmul, quantized_matmul_simd, QuantizedMatmulParams};
102pub use ops::quantized_matmul_ggml::{
103    dispatch_mm_for_test, quantized_matmul_ggml, quantized_matmul_mm_tensor_perm021,
104    GgmlQuantizedMatmulParams, GgmlQuantizedMatmulPerm021Params, GgmlType,
105    MM_ROUTING_THRESHOLD,
106};
107pub use ops::quantized_matmul_id::{quantized_matmul_id, QuantizedMatmulIdParams};
108pub use ops::quantized_matmul_id_ggml::{
109    dispatch_id_mm_for_test, quantized_matmul_id_ggml, quantized_matmul_id_ggml_pooled,
110    quantized_matmul_id_swiglu_q4_0,
111    GgmlIdMmDispatchParams, GgmlQuantizedMatmulIdParams, IdMmScratch,
112    MM_ID_ROUTING_THRESHOLD,
113};
114
115// Re-export weight loading utilities.
116pub use weight::{
117    load_quantized_weights, safetensors_to_metal_buffer, QuantizationConfig, QuantizedWeight,
118    SafetensorsFile, TensorQuantConfig,
119};
120
121// Re-export metal types that appear in the public API.
122pub use metal::MTLSize;
123pub use metal;
124
125#[cfg(test)]
126#[allow(clippy::expect_used, clippy::unwrap_used, clippy::panic)]
127mod tests {
128    use super::*;
129
130    // ---- T10.7: compile-time Send + Sync assertions ----
131    fn _assert_send<T: Send>() {}
132    fn _assert_sync<T: Sync>() {}
133
134    #[allow(dead_code)]
135    fn assert_send_sync() {
136        _assert_send::<MlxDevice>();
137        _assert_sync::<MlxDevice>();
138        _assert_send::<MlxBuffer>();
139        _assert_sync::<MlxBuffer>();
140        _assert_send::<MlxError>();
141        _assert_sync::<MlxError>();
142    }
143
144    // ---- T10.1: device initialization ----
145    #[test]
146    fn test_device_init() {
147        let device = MlxDevice::new().expect("MlxDevice::new() should succeed on Apple Silicon");
148        let name = device.name();
149        assert!(!name.is_empty(), "Device name should not be empty");
150        println!("Metal device: {name}");
151    }
152
153    // ---- T10.2: buffer allocation ----
154    #[test]
155    fn test_buffer_alloc() {
156        let device = MlxDevice::new().expect("device");
157        let shape = vec![2, 3, 4];
158        let byte_len = 2 * 3 * 4 * DType::F32.size_of(); // 96 bytes
159        let buf = device
160            .alloc_buffer(byte_len, DType::F32, shape.clone())
161            .expect("alloc_buffer");
162
163        assert_eq!(buf.dtype(), DType::F32);
164        assert_eq!(buf.shape(), &shape);
165        assert_eq!(buf.byte_len(), byte_len);
166        assert_eq!(buf.element_count(), 24);
167    }
168
169    // ---- T10.3: buffer read/write round-trip ----
170    #[test]
171    fn test_buffer_readwrite() {
172        let device = MlxDevice::new().expect("device");
173        let n = 64;
174        let byte_len = n * std::mem::size_of::<f32>();
175        let mut buf = device
176            .alloc_buffer(byte_len, DType::F32, vec![n])
177            .expect("alloc_buffer");
178
179        // Write known data.
180        {
181            let slice: &mut [f32] = buf.as_mut_slice().expect("as_mut_slice");
182            assert_eq!(slice.len(), n);
183            for (i, val) in slice.iter_mut().enumerate() {
184                *val = i as f32 * 1.5;
185            }
186        }
187
188        // Read back and verify.
189        {
190            let slice: &[f32] = buf.as_slice().expect("as_slice");
191            for (i, &val) in slice.iter().enumerate() {
192                let expected = i as f32 * 1.5;
193                assert!(
194                    (val - expected).abs() < f32::EPSILON,
195                    "Mismatch at index {i}: got {val}, expected {expected}"
196                );
197            }
198        }
199    }
200
201    // ---- T10.4: encoder lifecycle ----
202    #[test]
203    fn test_encoder_lifecycle() {
204        let device = MlxDevice::new().expect("device");
205        let mut enc = device.command_encoder().expect("command_encoder");
206        // Commit an empty command buffer — should succeed (no-op on GPU).
207        enc.commit_and_wait()
208            .expect("commit_and_wait on empty encoder");
209    }
210
211    // ---- T10.5: buffer pool reuse ----
212    #[test]
213    fn test_buffer_pool_reuse() {
214        let device = MlxDevice::new().expect("device");
215        let mut pool = MlxBufferPool::new();
216
217        // Allocate a buffer.
218        let buf1 = pool
219            .alloc(&device, 1024, DType::F32, vec![256])
220            .expect("pool alloc 1");
221        let buf1_ptr = buf1.contents_ptr();
222        let buf1_byte_len = buf1.byte_len();
223
224        // Release it back to the pool.
225        pool.release(buf1);
226        assert_eq!(pool.free_count(), 1);
227
228        // Allocate again — should reuse the same Metal buffer.
229        let buf2 = pool
230            .alloc(&device, 1024, DType::F32, vec![256])
231            .expect("pool alloc 2");
232        let buf2_ptr = buf2.contents_ptr();
233        let buf2_byte_len = buf2.byte_len();
234
235        assert_eq!(buf1_ptr, buf2_ptr, "Pool should reuse the same Metal buffer");
236        assert_eq!(buf1_byte_len, buf2_byte_len, "Byte lengths should match");
237        assert_eq!(pool.free_count(), 0, "Free list should be empty after reuse");
238    }
239
240    // ---- T10.6: kernel registry caching ----
241    #[test]
242    fn test_kernel_registry_caching() {
243        let device = MlxDevice::new().expect("device");
244        let mut registry = KernelRegistry::new();
245
246        // Register a minimal test kernel.
247        registry.register_source(
248            "test_add",
249            r#"
250            #include <metal_stdlib>
251            using namespace metal;
252            kernel void test_add(
253                device float *a [[buffer(0)]],
254                device float *b [[buffer(1)]],
255                device float *c [[buffer(2)]],
256                uint id [[thread_position_in_grid]]
257            ) {
258                c[id] = a[id] + b[id];
259            }
260            "#,
261        );
262
263        // First call — compiles the shader.
264        assert!(!registry.is_cached("test_add"));
265        let p1 = registry
266            .get_pipeline("test_add", device.metal_device())
267            .expect("get_pipeline first call");
268        let p1_ptr = p1 as *const _;
269        assert!(registry.is_cached("test_add"));
270
271        // Second call — returns cached pipeline.
272        let p2 = registry
273            .get_pipeline("test_add", device.metal_device())
274            .expect("get_pipeline second call");
275        let p2_ptr = p2 as *const _;
276
277        assert_eq!(
278            p1_ptr, p2_ptr,
279            "Second get_pipeline call should return the same cached pipeline"
280        );
281    }
282
283    // ---- Additional: test alloc_buffer with zero length returns error ----
284    #[test]
285    fn test_buffer_alloc_zero_len_error() {
286        let device = MlxDevice::new().expect("device");
287        let result = device.alloc_buffer(0, DType::F32, vec![]);
288        assert!(result.is_err(), "Zero-length allocation should fail");
289        match result {
290            Err(MlxError::InvalidArgument(_)) => {}
291            other => panic!("Expected InvalidArgument, got {:?}", other),
292        }
293    }
294
295    // ---- Additional: test kernel not found ----
296    #[test]
297    fn test_kernel_not_found() {
298        let device = MlxDevice::new().expect("device");
299        let mut registry = KernelRegistry::new();
300        let result = registry.get_pipeline("nonexistent_kernel", device.metal_device());
301        assert!(result.is_err());
302        match result {
303            Err(MlxError::KernelNotFound(name)) => {
304                assert_eq!(name, "nonexistent_kernel");
305            }
306            other => panic!("Expected KernelNotFound, got {:?}", other),
307        }
308    }
309
310    // ---- Additional: test DType properties ----
311    #[test]
312    fn test_dtype_sizes() {
313        assert_eq!(DType::F32.size_of(), 4);
314        assert_eq!(DType::F16.size_of(), 2);
315        assert_eq!(DType::BF16.size_of(), 2);
316        assert_eq!(DType::U8.size_of(), 1);
317        assert_eq!(DType::U16.size_of(), 2);
318        assert_eq!(DType::U32.size_of(), 4);
319        assert_eq!(DType::I32.size_of(), 4);
320    }
321
322    // ---- Additional: test MlxBuffer Debug ----
323    #[test]
324    fn test_buffer_debug() {
325        let device = MlxDevice::new().expect("device");
326        let buf = device
327            .alloc_buffer(64, DType::F16, vec![4, 8])
328            .expect("alloc_buffer");
329        let debug_str = format!("{:?}", buf);
330        assert!(debug_str.contains("MlxBuffer"));
331        assert!(debug_str.contains("F16"));
332        assert!(debug_str.contains("[4, 8]"));
333    }
334
335    // ---- Additional: test MlxError Display ----
336    #[test]
337    fn test_error_display() {
338        let e = MlxError::DeviceNotFound;
339        assert!(format!("{e}").contains("Metal GPU device"));
340
341        let e = MlxError::ShaderCompilationError {
342            name: "foo".into(),
343            message: "syntax error".into(),
344        };
345        assert!(format!("{e}").contains("foo"));
346        assert!(format!("{e}").contains("syntax error"));
347    }
348
349    // ---- Additional: test buffer pool with different sizes ----
350    #[test]
351    fn test_buffer_pool_size_buckets() {
352        let device = MlxDevice::new().expect("device");
353        let mut pool = MlxBufferPool::new();
354
355        // Allocate a 100-byte buffer (rounds to 128-byte bucket).
356        let buf_100 = pool.alloc(&device, 100, DType::U8, vec![100]).expect("alloc 100");
357        assert!(
358            buf_100.byte_len() >= 100,
359            "Buffer should be at least 100 bytes"
360        );
361        pool.release(buf_100);
362
363        // Allocate a 128-byte buffer — should reuse the same Metal buffer.
364        let buf_128 = pool.alloc(&device, 128, DType::U8, vec![128]).expect("alloc 128");
365        assert!(buf_128.byte_len() >= 128);
366        pool.release(buf_128);
367
368        // Allocate a 200-byte buffer — different bucket (256), fresh allocation.
369        let buf_200 = pool.alloc(&device, 200, DType::U8, vec![200]).expect("alloc 200");
370        assert!(buf_200.byte_len() >= 200);
371        pool.release(buf_200);
372
373        assert_eq!(pool.free_count(), 2, "Two different bucket sizes in pool");
374    }
375}
mlx_native/lib.rs

mlx_native/
lib.rs
