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::mul_mv_ext::{mul_mv_ext_dispatch, MulMvExtParams};
108pub use ops::quantized_matmul_id::{quantized_matmul_id, QuantizedMatmulIdParams};
109pub use ops::quantized_matmul_id_ggml::{
110    dispatch_id_mm_for_test, quantized_matmul_id_ggml, quantized_matmul_id_ggml_pooled,
111    quantized_matmul_id_swiglu_q4_0,
112    GgmlIdMmDispatchParams, GgmlQuantizedMatmulIdParams, IdMmScratch,
113    MM_ID_ROUTING_THRESHOLD,
114};
115
116// Re-export weight loading utilities.
117pub use weight::{
118    load_quantized_weights, safetensors_to_metal_buffer, QuantizationConfig, QuantizedWeight,
119    SafetensorsFile, TensorQuantConfig,
120};
121
122// Re-export metal types that appear in the public API.
123pub use metal::MTLSize;
124pub use metal;
125
126#[cfg(test)]
127#[allow(clippy::expect_used, clippy::unwrap_used, clippy::panic)]
128mod tests {
129    use super::*;
130
131    // ---- T10.7: compile-time Send + Sync assertions ----
132    fn _assert_send<T: Send>() {}
133    fn _assert_sync<T: Sync>() {}
134
135    #[allow(dead_code)]
136    fn assert_send_sync() {
137        _assert_send::<MlxDevice>();
138        _assert_sync::<MlxDevice>();
139        _assert_send::<MlxBuffer>();
140        _assert_sync::<MlxBuffer>();
141        _assert_send::<MlxError>();
142        _assert_sync::<MlxError>();
143    }
144
145    // ---- T10.1: device initialization ----
146    #[test]
147    fn test_device_init() {
148        let device = MlxDevice::new().expect("MlxDevice::new() should succeed on Apple Silicon");
149        let name = device.name();
150        assert!(!name.is_empty(), "Device name should not be empty");
151        println!("Metal device: {name}");
152    }
153
154    // ---- T10.2: buffer allocation ----
155    #[test]
156    fn test_buffer_alloc() {
157        let device = MlxDevice::new().expect("device");
158        let shape = vec![2, 3, 4];
159        let byte_len = 2 * 3 * 4 * DType::F32.size_of(); // 96 bytes
160        let buf = device
161            .alloc_buffer(byte_len, DType::F32, shape.clone())
162            .expect("alloc_buffer");
163
164        assert_eq!(buf.dtype(), DType::F32);
165        assert_eq!(buf.shape(), &shape);
166        assert_eq!(buf.byte_len(), byte_len);
167        assert_eq!(buf.element_count(), 24);
168    }
169
170    // ---- T10.3: buffer read/write round-trip ----
171    #[test]
172    fn test_buffer_readwrite() {
173        let device = MlxDevice::new().expect("device");
174        let n = 64;
175        let byte_len = n * std::mem::size_of::<f32>();
176        let mut buf = device
177            .alloc_buffer(byte_len, DType::F32, vec![n])
178            .expect("alloc_buffer");
179
180        // Write known data.
181        {
182            let slice: &mut [f32] = buf.as_mut_slice().expect("as_mut_slice");
183            assert_eq!(slice.len(), n);
184            for (i, val) in slice.iter_mut().enumerate() {
185                *val = i as f32 * 1.5;
186            }
187        }
188
189        // Read back and verify.
190        {
191            let slice: &[f32] = buf.as_slice().expect("as_slice");
192            for (i, &val) in slice.iter().enumerate() {
193                let expected = i as f32 * 1.5;
194                assert!(
195                    (val - expected).abs() < f32::EPSILON,
196                    "Mismatch at index {i}: got {val}, expected {expected}"
197                );
198            }
199        }
200    }
201
202    // ---- T10.4: encoder lifecycle ----
203    #[test]
204    fn test_encoder_lifecycle() {
205        let device = MlxDevice::new().expect("device");
206        let mut enc = device.command_encoder().expect("command_encoder");
207        // Commit an empty command buffer — should succeed (no-op on GPU).
208        enc.commit_and_wait()
209            .expect("commit_and_wait on empty encoder");
210    }
211
212    // ---- T10.5: buffer pool reuse ----
213    #[test]
214    fn test_buffer_pool_reuse() {
215        let device = MlxDevice::new().expect("device");
216        let mut pool = MlxBufferPool::new();
217
218        // Allocate a buffer.
219        let buf1 = pool
220            .alloc(&device, 1024, DType::F32, vec![256])
221            .expect("pool alloc 1");
222        let buf1_ptr = buf1.contents_ptr();
223        let buf1_byte_len = buf1.byte_len();
224
225        // Release it back to the pool.
226        pool.release(buf1);
227        assert_eq!(pool.free_count(), 1);
228
229        // Allocate again — should reuse the same Metal buffer.
230        let buf2 = pool
231            .alloc(&device, 1024, DType::F32, vec![256])
232            .expect("pool alloc 2");
233        let buf2_ptr = buf2.contents_ptr();
234        let buf2_byte_len = buf2.byte_len();
235
236        assert_eq!(buf1_ptr, buf2_ptr, "Pool should reuse the same Metal buffer");
237        assert_eq!(buf1_byte_len, buf2_byte_len, "Byte lengths should match");
238        assert_eq!(pool.free_count(), 0, "Free list should be empty after reuse");
239    }
240
241    // ---- T10.6: kernel registry caching ----
242    #[test]
243    fn test_kernel_registry_caching() {
244        let device = MlxDevice::new().expect("device");
245        let mut registry = KernelRegistry::new();
246
247        // Register a minimal test kernel.
248        registry.register_source(
249            "test_add",
250            r#"
251            #include <metal_stdlib>
252            using namespace metal;
253            kernel void test_add(
254                device float *a [[buffer(0)]],
255                device float *b [[buffer(1)]],
256                device float *c [[buffer(2)]],
257                uint id [[thread_position_in_grid]]
258            ) {
259                c[id] = a[id] + b[id];
260            }
261            "#,
262        );
263
264        // First call — compiles the shader.
265        assert!(!registry.is_cached("test_add"));
266        let p1 = registry
267            .get_pipeline("test_add", device.metal_device())
268            .expect("get_pipeline first call");
269        let p1_ptr = p1 as *const _;
270        assert!(registry.is_cached("test_add"));
271
272        // Second call — returns cached pipeline.
273        let p2 = registry
274            .get_pipeline("test_add", device.metal_device())
275            .expect("get_pipeline second call");
276        let p2_ptr = p2 as *const _;
277
278        assert_eq!(
279            p1_ptr, p2_ptr,
280            "Second get_pipeline call should return the same cached pipeline"
281        );
282    }
283
284    // ---- Additional: test alloc_buffer with zero length returns error ----
285    #[test]
286    fn test_buffer_alloc_zero_len_error() {
287        let device = MlxDevice::new().expect("device");
288        let result = device.alloc_buffer(0, DType::F32, vec![]);
289        assert!(result.is_err(), "Zero-length allocation should fail");
290        match result {
291            Err(MlxError::InvalidArgument(_)) => {}
292            other => panic!("Expected InvalidArgument, got {:?}", other),
293        }
294    }
295
296    // ---- Additional: test kernel not found ----
297    #[test]
298    fn test_kernel_not_found() {
299        let device = MlxDevice::new().expect("device");
300        let mut registry = KernelRegistry::new();
301        let result = registry.get_pipeline("nonexistent_kernel", device.metal_device());
302        assert!(result.is_err());
303        match result {
304            Err(MlxError::KernelNotFound(name)) => {
305                assert_eq!(name, "nonexistent_kernel");
306            }
307            other => panic!("Expected KernelNotFound, got {:?}", other),
308        }
309    }
310
311    // ---- Additional: test DType properties ----
312    #[test]
313    fn test_dtype_sizes() {
314        assert_eq!(DType::F32.size_of(), 4);
315        assert_eq!(DType::F16.size_of(), 2);
316        assert_eq!(DType::BF16.size_of(), 2);
317        assert_eq!(DType::U8.size_of(), 1);
318        assert_eq!(DType::U16.size_of(), 2);
319        assert_eq!(DType::U32.size_of(), 4);
320        assert_eq!(DType::I32.size_of(), 4);
321    }
322
323    // ---- Additional: test MlxBuffer Debug ----
324    #[test]
325    fn test_buffer_debug() {
326        let device = MlxDevice::new().expect("device");
327        let buf = device
328            .alloc_buffer(64, DType::F16, vec![4, 8])
329            .expect("alloc_buffer");
330        let debug_str = format!("{:?}", buf);
331        assert!(debug_str.contains("MlxBuffer"));
332        assert!(debug_str.contains("F16"));
333        assert!(debug_str.contains("[4, 8]"));
334    }
335
336    // ---- Additional: test MlxError Display ----
337    #[test]
338    fn test_error_display() {
339        let e = MlxError::DeviceNotFound;
340        assert!(format!("{e}").contains("Metal GPU device"));
341
342        let e = MlxError::ShaderCompilationError {
343            name: "foo".into(),
344            message: "syntax error".into(),
345        };
346        assert!(format!("{e}").contains("foo"));
347        assert!(format!("{e}").contains("syntax error"));
348    }
349
350    // ---- Additional: test buffer pool with different sizes ----
351    #[test]
352    fn test_buffer_pool_size_buckets() {
353        let device = MlxDevice::new().expect("device");
354        let mut pool = MlxBufferPool::new();
355
356        // Allocate a 100-byte buffer (rounds to 128-byte bucket).
357        let buf_100 = pool.alloc(&device, 100, DType::U8, vec![100]).expect("alloc 100");
358        assert!(
359            buf_100.byte_len() >= 100,
360            "Buffer should be at least 100 bytes"
361        );
362        pool.release(buf_100);
363
364        // Allocate a 128-byte buffer — should reuse the same Metal buffer.
365        let buf_128 = pool.alloc(&device, 128, DType::U8, vec![128]).expect("alloc 128");
366        assert!(buf_128.byte_len() >= 128);
367        pool.release(buf_128);
368
369        // Allocate a 200-byte buffer — different bucket (256), fresh allocation.
370        let buf_200 = pool.alloc(&device, 200, DType::U8, vec![200]).expect("alloc 200");
371        assert!(buf_200.byte_len() >= 200);
372        pool.release(buf_200);
373
374        assert_eq!(pool.free_count(), 2, "Two different bucket sizes in pool");
375    }
376}
mlx_native/lib.rs

mlx_native/
lib.rs
