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::{
109    quantized_matmul_id, quantized_matmul_id_into, QuantizedMatmulIdParams,
110};
111pub use ops::quantized_matmul_id_ggml::{
112    dispatch_id_mm_for_test, quantized_matmul_id_ggml, quantized_matmul_id_ggml_pooled,
113    quantized_matmul_id_swiglu_q4_0,
114    GgmlIdMmDispatchParams, GgmlQuantizedMatmulIdParams, IdMmScratch,
115    MM_ID_ROUTING_THRESHOLD,
116};
117
118// Re-export weight loading utilities.
119pub use weight::{
120    load_quantized_weights, safetensors_to_metal_buffer, QuantizationConfig, QuantizedWeight,
121    SafetensorsFile, TensorQuantConfig,
122};
123
124// Re-export metal types that appear in the public API.
125pub use metal::MTLSize;
126pub use metal;
127
128#[cfg(test)]
129#[allow(clippy::expect_used, clippy::unwrap_used, clippy::panic)]
130mod tests {
131    use super::*;
132
133    // ---- T10.7: compile-time Send + Sync assertions ----
134    fn _assert_send<T: Send>() {}
135    fn _assert_sync<T: Sync>() {}
136
137    #[allow(dead_code)]
138    fn assert_send_sync() {
139        _assert_send::<MlxDevice>();
140        _assert_sync::<MlxDevice>();
141        _assert_send::<MlxBuffer>();
142        _assert_sync::<MlxBuffer>();
143        _assert_send::<MlxError>();
144        _assert_sync::<MlxError>();
145    }
146
147    // ---- T10.1: device initialization ----
148    #[test]
149    fn test_device_init() {
150        let device = MlxDevice::new().expect("MlxDevice::new() should succeed on Apple Silicon");
151        let name = device.name();
152        assert!(!name.is_empty(), "Device name should not be empty");
153        println!("Metal device: {name}");
154    }
155
156    // ---- T10.2: buffer allocation ----
157    #[test]
158    fn test_buffer_alloc() {
159        let device = MlxDevice::new().expect("device");
160        let shape = vec![2, 3, 4];
161        let byte_len = 2 * 3 * 4 * DType::F32.size_of(); // 96 bytes
162        let buf = device
163            .alloc_buffer(byte_len, DType::F32, shape.clone())
164            .expect("alloc_buffer");
165
166        assert_eq!(buf.dtype(), DType::F32);
167        assert_eq!(buf.shape(), &shape);
168        assert_eq!(buf.byte_len(), byte_len);
169        assert_eq!(buf.element_count(), 24);
170    }
171
172    // ---- T10.3: buffer read/write round-trip ----
173    #[test]
174    fn test_buffer_readwrite() {
175        let device = MlxDevice::new().expect("device");
176        let n = 64;
177        let byte_len = n * std::mem::size_of::<f32>();
178        let mut buf = device
179            .alloc_buffer(byte_len, DType::F32, vec![n])
180            .expect("alloc_buffer");
181
182        // Write known data.
183        {
184            let slice: &mut [f32] = buf.as_mut_slice().expect("as_mut_slice");
185            assert_eq!(slice.len(), n);
186            for (i, val) in slice.iter_mut().enumerate() {
187                *val = i as f32 * 1.5;
188            }
189        }
190
191        // Read back and verify.
192        {
193            let slice: &[f32] = buf.as_slice().expect("as_slice");
194            for (i, &val) in slice.iter().enumerate() {
195                let expected = i as f32 * 1.5;
196                assert!(
197                    (val - expected).abs() < f32::EPSILON,
198                    "Mismatch at index {i}: got {val}, expected {expected}"
199                );
200            }
201        }
202    }
203
204    // ---- T10.4: encoder lifecycle ----
205    #[test]
206    fn test_encoder_lifecycle() {
207        let device = MlxDevice::new().expect("device");
208        let mut enc = device.command_encoder().expect("command_encoder");
209        // Commit an empty command buffer — should succeed (no-op on GPU).
210        enc.commit_and_wait()
211            .expect("commit_and_wait on empty encoder");
212    }
213
214    // ---- T10.5: buffer pool reuse ----
215    #[test]
216    fn test_buffer_pool_reuse() {
217        let device = MlxDevice::new().expect("device");
218        let mut pool = MlxBufferPool::new();
219
220        // Allocate a buffer.
221        let buf1 = pool
222            .alloc(&device, 1024, DType::F32, vec![256])
223            .expect("pool alloc 1");
224        let buf1_ptr = buf1.contents_ptr();
225        let buf1_byte_len = buf1.byte_len();
226
227        // Release it back to the pool.
228        pool.release(buf1);
229        assert_eq!(pool.free_count(), 1);
230
231        // Allocate again — should reuse the same Metal buffer.
232        let buf2 = pool
233            .alloc(&device, 1024, DType::F32, vec![256])
234            .expect("pool alloc 2");
235        let buf2_ptr = buf2.contents_ptr();
236        let buf2_byte_len = buf2.byte_len();
237
238        assert_eq!(buf1_ptr, buf2_ptr, "Pool should reuse the same Metal buffer");
239        assert_eq!(buf1_byte_len, buf2_byte_len, "Byte lengths should match");
240        assert_eq!(pool.free_count(), 0, "Free list should be empty after reuse");
241    }
242
243    // ---- T10.6: kernel registry caching ----
244    #[test]
245    fn test_kernel_registry_caching() {
246        let device = MlxDevice::new().expect("device");
247        let mut registry = KernelRegistry::new();
248
249        // Register a minimal test kernel.
250        registry.register_source(
251            "test_add",
252            r#"
253            #include <metal_stdlib>
254            using namespace metal;
255            kernel void test_add(
256                device float *a [[buffer(0)]],
257                device float *b [[buffer(1)]],
258                device float *c [[buffer(2)]],
259                uint id [[thread_position_in_grid]]
260            ) {
261                c[id] = a[id] + b[id];
262            }
263            "#,
264        );
265
266        // First call — compiles the shader.
267        assert!(!registry.is_cached("test_add"));
268        let p1 = registry
269            .get_pipeline("test_add", device.metal_device())
270            .expect("get_pipeline first call");
271        let p1_ptr = p1 as *const _;
272        assert!(registry.is_cached("test_add"));
273
274        // Second call — returns cached pipeline.
275        let p2 = registry
276            .get_pipeline("test_add", device.metal_device())
277            .expect("get_pipeline second call");
278        let p2_ptr = p2 as *const _;
279
280        assert_eq!(
281            p1_ptr, p2_ptr,
282            "Second get_pipeline call should return the same cached pipeline"
283        );
284    }
285
286    // ---- Additional: test alloc_buffer with zero length returns error ----
287    #[test]
288    fn test_buffer_alloc_zero_len_error() {
289        let device = MlxDevice::new().expect("device");
290        let result = device.alloc_buffer(0, DType::F32, vec![]);
291        assert!(result.is_err(), "Zero-length allocation should fail");
292        match result {
293            Err(MlxError::InvalidArgument(_)) => {}
294            other => panic!("Expected InvalidArgument, got {:?}", other),
295        }
296    }
297
298    // ---- Additional: test kernel not found ----
299    #[test]
300    fn test_kernel_not_found() {
301        let device = MlxDevice::new().expect("device");
302        let mut registry = KernelRegistry::new();
303        let result = registry.get_pipeline("nonexistent_kernel", device.metal_device());
304        assert!(result.is_err());
305        match result {
306            Err(MlxError::KernelNotFound(name)) => {
307                assert_eq!(name, "nonexistent_kernel");
308            }
309            other => panic!("Expected KernelNotFound, got {:?}", other),
310        }
311    }
312
313    // ---- Additional: test DType properties ----
314    #[test]
315    fn test_dtype_sizes() {
316        assert_eq!(DType::F32.size_of(), 4);
317        assert_eq!(DType::F16.size_of(), 2);
318        assert_eq!(DType::BF16.size_of(), 2);
319        assert_eq!(DType::U8.size_of(), 1);
320        assert_eq!(DType::U16.size_of(), 2);
321        assert_eq!(DType::U32.size_of(), 4);
322        assert_eq!(DType::I32.size_of(), 4);
323    }
324
325    // ---- Additional: test MlxBuffer Debug ----
326    #[test]
327    fn test_buffer_debug() {
328        let device = MlxDevice::new().expect("device");
329        let buf = device
330            .alloc_buffer(64, DType::F16, vec![4, 8])
331            .expect("alloc_buffer");
332        let debug_str = format!("{:?}", buf);
333        assert!(debug_str.contains("MlxBuffer"));
334        assert!(debug_str.contains("F16"));
335        assert!(debug_str.contains("[4, 8]"));
336    }
337
338    // ---- Additional: test MlxError Display ----
339    #[test]
340    fn test_error_display() {
341        let e = MlxError::DeviceNotFound;
342        assert!(format!("{e}").contains("Metal GPU device"));
343
344        let e = MlxError::ShaderCompilationError {
345            name: "foo".into(),
346            message: "syntax error".into(),
347        };
348        assert!(format!("{e}").contains("foo"));
349        assert!(format!("{e}").contains("syntax error"));
350    }
351
352    // ---- Additional: test buffer pool with different sizes ----
353    #[test]
354    fn test_buffer_pool_size_buckets() {
355        let device = MlxDevice::new().expect("device");
356        let mut pool = MlxBufferPool::new();
357
358        // Allocate a 100-byte buffer (rounds to 128-byte bucket).
359        let buf_100 = pool.alloc(&device, 100, DType::U8, vec![100]).expect("alloc 100");
360        assert!(
361            buf_100.byte_len() >= 100,
362            "Buffer should be at least 100 bytes"
363        );
364        pool.release(buf_100);
365
366        // Allocate a 128-byte buffer — should reuse the same Metal buffer.
367        let buf_128 = pool.alloc(&device, 128, DType::U8, vec![128]).expect("alloc 128");
368        assert!(buf_128.byte_len() >= 128);
369        pool.release(buf_128);
370
371        // Allocate a 200-byte buffer — different bucket (256), fresh allocation.
372        let buf_200 = pool.alloc(&device, 200, DType::U8, vec![200]).expect("alloc 200");
373        assert!(buf_200.byte_len() >= 200);
374        pool.release(buf_200);
375
376        assert_eq!(pool.free_count(), 2, "Two different bucket sizes in pool");
377    }
378}
mlx_native/lib.rs

mlx_native/
lib.rs
