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 env_flags;
56mod kernel_registry;
57mod mem_ranges;
58mod residency;
59pub mod gguf;
60pub mod kernel_profile;
61pub mod graph;
62pub mod metal_capture;
63pub mod ops;
64pub mod turboquant;
65pub mod tq_oracle;
66pub mod weight;
67
68// ---- public re-exports ----
69pub use buffer::MlxBuffer;
70pub use buffer_pool::MlxBufferPool;
71pub use device::MlxDevice;
72pub use dtypes::DType;
73pub use encoder::{
74    auto_barrier_concurrent_count, auto_barrier_count, barrier_count, barrier_total_ns,
75    cmd_buf_count, dispatch_count, pipeline_dispatch_buckets, reset_counters,
76    reset_pipeline_dispatch_buckets, sync_count, CapturedNode, CapturedOpKind,
77    CommandEncoder, DispatchKind, KernelArg, RecordedBinding,
78};
79pub use encoder_session::EncoderSession;
80pub use mem_ranges::{BufferRange, MemRangeRole, MemRanges};
81pub use error::{MlxError, Result};
82pub use graph::{ComputeGraph, GraphExecutor, GraphSession, OpKind};
83pub use kernel_registry::KernelRegistry;
84// Test-only counters and gate-reset helpers.  Marked #[doc(hidden)] so
85// they don't appear in published rustdoc; consumers should not depend
86// on them outside test code.  Not feature-gated because integration
87// tests in tests/ are a separate crate and cannot rely on the lib's
88// `test` cfg flag.
89#[doc(hidden)]
90pub use residency::{
91    macos_15_or_newer_for_test, reset_residency_env_cache_for_test,
92    reset_residency_test_counters, residency_allocation_count_for_test,
93    residency_commit_call_count_for_test,
94};
95
96// Re-export GGUF parser.
97pub use gguf::{GgufFile, MetadataValue, TensorInfo};
98
99// Re-export ops.
100pub use ops::dense_mm_bf16::{dense_matmul_bf16_f32_tensor, DenseMmBf16F32Params};
101pub use ops::dense_mm_f16::{dense_matmul_f16_f32_tensor, DenseMmF16F32Params};
102pub use ops::dense_mm_f32_f32::{dense_matmul_f32_f32_tensor, DenseMmF32F32Params};
103pub use ops::quantized_matmul::{quantized_matmul, quantized_matmul_simd, QuantizedMatmulParams};
104pub use ops::quantized_matmul_ggml::{
105    dispatch_mm_for_test, quantized_matmul_ggml, quantized_matmul_mm_tensor_perm021,
106    GgmlQuantizedMatmulParams, GgmlQuantizedMatmulPerm021Params, GgmlType,
107    MM_ROUTING_THRESHOLD,
108};
109pub use ops::mul_mv_ext::{mul_mv_ext_dispatch, MulMvExtParams};
110pub use ops::quantized_matmul_id::{
111    quantized_matmul_id, quantized_matmul_id_into, QuantizedMatmulIdParams,
112};
113pub use ops::quantized_matmul_id_ggml::{
114    dispatch_id_mm_for_test, quantized_matmul_id_ggml, quantized_matmul_id_ggml_pooled,
115    quantized_matmul_id_swiglu_q4_0,
116    GgmlIdMmDispatchParams, GgmlQuantizedMatmulIdParams, IdMmScratch,
117    MM_ID_ROUTING_THRESHOLD,
118};
119
120// Re-export weight loading utilities.
121pub use weight::{
122    load_quantized_weights, safetensors_to_metal_buffer, QuantizationConfig, QuantizedWeight,
123    SafetensorsFile, TensorQuantConfig,
124};
125
126// Re-export metal types that appear in the public API.
127pub use metal::MTLSize;
128pub use metal;
129
130#[cfg(test)]
131#[allow(clippy::expect_used, clippy::unwrap_used, clippy::panic)]
132mod tests {
133    use super::*;
134
135    // ---- T10.7: compile-time Send + Sync assertions ----
136    fn _assert_send<T: Send>() {}
137    fn _assert_sync<T: Sync>() {}
138
139    #[allow(dead_code)]
140    fn assert_send_sync() {
141        _assert_send::<MlxDevice>();
142        _assert_sync::<MlxDevice>();
143        _assert_send::<MlxBuffer>();
144        _assert_sync::<MlxBuffer>();
145        _assert_send::<MlxError>();
146        _assert_sync::<MlxError>();
147    }
148
149    // ---- T10.1: device initialization ----
150    #[test]
151    fn test_device_init() {
152        let device = MlxDevice::new().expect("MlxDevice::new() should succeed on Apple Silicon");
153        let name = device.name();
154        assert!(!name.is_empty(), "Device name should not be empty");
155        println!("Metal device: {name}");
156    }
157
158    // ---- T10.2: buffer allocation ----
159    #[test]
160    fn test_buffer_alloc() {
161        let device = MlxDevice::new().expect("device");
162        let shape = vec![2, 3, 4];
163        let byte_len = 2 * 3 * 4 * DType::F32.size_of(); // 96 bytes
164        let buf = device
165            .alloc_buffer(byte_len, DType::F32, shape.clone())
166            .expect("alloc_buffer");
167
168        assert_eq!(buf.dtype(), DType::F32);
169        assert_eq!(buf.shape(), &shape);
170        assert_eq!(buf.byte_len(), byte_len);
171        assert_eq!(buf.element_count(), 24);
172    }
173
174    // ---- T10.3: buffer read/write round-trip ----
175    #[test]
176    fn test_buffer_readwrite() {
177        let device = MlxDevice::new().expect("device");
178        let n = 64;
179        let byte_len = n * std::mem::size_of::<f32>();
180        let mut buf = device
181            .alloc_buffer(byte_len, DType::F32, vec![n])
182            .expect("alloc_buffer");
183
184        // Write known data.
185        {
186            let slice: &mut [f32] = buf.as_mut_slice().expect("as_mut_slice");
187            assert_eq!(slice.len(), n);
188            for (i, val) in slice.iter_mut().enumerate() {
189                *val = i as f32 * 1.5;
190            }
191        }
192
193        // Read back and verify.
194        {
195            let slice: &[f32] = buf.as_slice().expect("as_slice");
196            for (i, &val) in slice.iter().enumerate() {
197                let expected = i as f32 * 1.5;
198                assert!(
199                    (val - expected).abs() < f32::EPSILON,
200                    "Mismatch at index {i}: got {val}, expected {expected}"
201                );
202            }
203        }
204    }
205
206    // ---- T10.4: encoder lifecycle ----
207    #[test]
208    fn test_encoder_lifecycle() {
209        let device = MlxDevice::new().expect("device");
210        let mut enc = device.command_encoder().expect("command_encoder");
211        // Commit an empty command buffer — should succeed (no-op on GPU).
212        enc.commit_and_wait()
213            .expect("commit_and_wait on empty encoder");
214    }
215
216    // ---- T10.5: buffer pool reuse ----
217    #[test]
218    fn test_buffer_pool_reuse() {
219        let device = MlxDevice::new().expect("device");
220        let mut pool = MlxBufferPool::new();
221
222        // Allocate a buffer.
223        let buf1 = pool
224            .alloc(&device, 1024, DType::F32, vec![256])
225            .expect("pool alloc 1");
226        let buf1_ptr = buf1.contents_ptr();
227        let buf1_byte_len = buf1.byte_len();
228
229        // Release it back to the pool.
230        pool.release(buf1);
231        assert_eq!(pool.free_count(), 1);
232
233        // Allocate again — should reuse the same Metal buffer.
234        let buf2 = pool
235            .alloc(&device, 1024, DType::F32, vec![256])
236            .expect("pool alloc 2");
237        let buf2_ptr = buf2.contents_ptr();
238        let buf2_byte_len = buf2.byte_len();
239
240        assert_eq!(buf1_ptr, buf2_ptr, "Pool should reuse the same Metal buffer");
241        assert_eq!(buf1_byte_len, buf2_byte_len, "Byte lengths should match");
242        assert_eq!(pool.free_count(), 0, "Free list should be empty after reuse");
243    }
244
245    // ---- T10.6: kernel registry caching ----
246    #[test]
247    fn test_kernel_registry_caching() {
248        let device = MlxDevice::new().expect("device");
249        let mut registry = KernelRegistry::new();
250
251        // Register a minimal test kernel.
252        registry.register_source(
253            "test_add",
254            r#"
255            #include <metal_stdlib>
256            using namespace metal;
257            kernel void test_add(
258                device float *a [[buffer(0)]],
259                device float *b [[buffer(1)]],
260                device float *c [[buffer(2)]],
261                uint id [[thread_position_in_grid]]
262            ) {
263                c[id] = a[id] + b[id];
264            }
265            "#,
266        );
267
268        // First call — compiles the shader.
269        assert!(!registry.is_cached("test_add"));
270        let p1 = registry
271            .get_pipeline("test_add", device.metal_device())
272            .expect("get_pipeline first call");
273        let p1_ptr = p1 as *const _;
274        assert!(registry.is_cached("test_add"));
275
276        // Second call — returns cached pipeline.
277        let p2 = registry
278            .get_pipeline("test_add", device.metal_device())
279            .expect("get_pipeline second call");
280        let p2_ptr = p2 as *const _;
281
282        assert_eq!(
283            p1_ptr, p2_ptr,
284            "Second get_pipeline call should return the same cached pipeline"
285        );
286    }
287
288    // ---- Additional: test alloc_buffer with zero length returns error ----
289    #[test]
290    fn test_buffer_alloc_zero_len_error() {
291        let device = MlxDevice::new().expect("device");
292        let result = device.alloc_buffer(0, DType::F32, vec![]);
293        assert!(result.is_err(), "Zero-length allocation should fail");
294        match result {
295            Err(MlxError::InvalidArgument(_)) => {}
296            other => panic!("Expected InvalidArgument, got {:?}", other),
297        }
298    }
299
300    // ---- Additional: test kernel not found ----
301    #[test]
302    fn test_kernel_not_found() {
303        let device = MlxDevice::new().expect("device");
304        let mut registry = KernelRegistry::new();
305        let result = registry.get_pipeline("nonexistent_kernel", device.metal_device());
306        assert!(result.is_err());
307        match result {
308            Err(MlxError::KernelNotFound(name)) => {
309                assert_eq!(name, "nonexistent_kernel");
310            }
311            other => panic!("Expected KernelNotFound, got {:?}", other),
312        }
313    }
314
315    // ---- Additional: test DType properties ----
316    #[test]
317    fn test_dtype_sizes() {
318        assert_eq!(DType::F32.size_of(), 4);
319        assert_eq!(DType::F16.size_of(), 2);
320        assert_eq!(DType::BF16.size_of(), 2);
321        assert_eq!(DType::U8.size_of(), 1);
322        assert_eq!(DType::U16.size_of(), 2);
323        assert_eq!(DType::U32.size_of(), 4);
324        assert_eq!(DType::I32.size_of(), 4);
325    }
326
327    // ---- Additional: test MlxBuffer Debug ----
328    #[test]
329    fn test_buffer_debug() {
330        let device = MlxDevice::new().expect("device");
331        let buf = device
332            .alloc_buffer(64, DType::F16, vec![4, 8])
333            .expect("alloc_buffer");
334        let debug_str = format!("{:?}", buf);
335        assert!(debug_str.contains("MlxBuffer"));
336        assert!(debug_str.contains("F16"));
337        assert!(debug_str.contains("[4, 8]"));
338    }
339
340    // ---- Additional: test MlxError Display ----
341    #[test]
342    fn test_error_display() {
343        let e = MlxError::DeviceNotFound;
344        assert!(format!("{e}").contains("Metal GPU device"));
345
346        let e = MlxError::ShaderCompilationError {
347            name: "foo".into(),
348            message: "syntax error".into(),
349        };
350        assert!(format!("{e}").contains("foo"));
351        assert!(format!("{e}").contains("syntax error"));
352    }
353
354    // ---- Additional: test buffer pool with different sizes ----
355    #[test]
356    fn test_buffer_pool_size_buckets() {
357        let device = MlxDevice::new().expect("device");
358        let mut pool = MlxBufferPool::new();
359
360        // Allocate a 100-byte buffer (rounds to 128-byte bucket).
361        let buf_100 = pool.alloc(&device, 100, DType::U8, vec![100]).expect("alloc 100");
362        assert!(
363            buf_100.byte_len() >= 100,
364            "Buffer should be at least 100 bytes"
365        );
366        pool.release(buf_100);
367
368        // Allocate a 128-byte buffer — should reuse the same Metal buffer.
369        let buf_128 = pool.alloc(&device, 128, DType::U8, vec![128]).expect("alloc 128");
370        assert!(buf_128.byte_len() >= 128);
371        pool.release(buf_128);
372
373        // Allocate a 200-byte buffer — different bucket (256), fresh allocation.
374        let buf_200 = pool.alloc(&device, 200, DType::U8, vec![200]).expect("alloc 200");
375        assert!(buf_200.byte_len() >= 200);
376        pool.release(buf_200);
377
378        assert_eq!(pool.free_count(), 2, "Two different bucket sizes in pool");
379    }
380}
mlx_native/lib.rs

mlx_native/
lib.rs
